3party/mimalloc
0
0
Fork 0
mirror of https://github.com/microsoft/mimalloc.git synced 2024-12-25 20:14:12 +08:00
Code Issues Projects Releases Wiki Activity

merge from dev-slice v2.1.4

Browse Source
This commit is contained in:
Daan 2024-04-22 11:10:56 -07:00
commit 229ec9cbdc
55 changed files with 2475 additions and 1461 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
.gitattributes vendored
View File

@ -10,3 +10,4 @@
*.dll binary
*.lib binary
*.exe binary
bin export-ignore

48
CMakeLists.txt
View File

@ -19,6 +19,7 @@ option(MI_OSX_INTERPOSE "Use interpose to override standard malloc on macOS"
option(MI_OSX_ZONE "Use malloc zone to override standard malloc on macOS" ON)
option(MI_WIN_REDIRECT "Use redirection module ('mimalloc-redirect') on Windows if compiling mimalloc as a DLL" ON)
option(MI_LOCAL_DYNAMIC_TLS "Use slightly slower, dlopen-compatible TLS mechanism (Unix)" OFF)
option(MI_LIBC_MUSL "Set this when linking with musl libc" OFF)
option(MI_BUILD_SHARED "Build shared library" ON)
option(MI_BUILD_STATIC "Build static library" ON)
option(MI_BUILD_OBJECT "Build object library" ON)
@ -27,10 +28,11 @@ option(MI_DEBUG_TSAN "Build with thread sanitizer (needs clang)" OFF)
option(MI_DEBUG_UBSAN "Build with undefined-behavior sanitizer (needs clang++)" OFF)
option(MI_SKIP_COLLECT_ON_EXIT "Skip collecting memory on program exit" OFF)
option(MI_NO_PADDING "Force no use of padding even in DEBUG mode etc." OFF)
option(MI_INSTALL_TOPLEVEL "Install directly into $CMAKE_INSTALL_PREFIX instead of PREFIX/lib/mimalloc-version" OFF)
option(MI_NO_THP "Disable transparent huge pages support on Linux/Android for the mimalloc process only" OFF)
# deprecated options
option(MI_CHECK_FULL "Use full internal invariant checking in DEBUG mode (deprecated, use MI_DEBUG_FULL instead)" OFF)
option(MI_INSTALL_TOPLEVEL "Install directly into $CMAKE_INSTALL_PREFIX instead of PREFIX/lib/mimalloc-version (deprecated)" OFF)
option(MI_USE_LIBATOMIC "Explicitly link with -latomic (on older systems) (deprecated and detected automatically)" OFF)
include(CheckIncludeFiles)
@ -45,6 +47,7 @@ set(mi_sources
src/bitmap.c
src/heap.c
src/init.c
src/libc.c
src/options.c
src/os.c
src/page.c
@ -82,6 +85,17 @@ endif()
# Process options
# -----------------------------------------------------------------------------
# put -Wall early so other warnings can be disabled selectively
if(CMAKE_C_COMPILER_ID MATCHES "AppleClang|Clang")
list(APPEND mi_cflags -Wall -Wextra -Wpedantic)
endif()
if(CMAKE_C_COMPILER_ID MATCHES "GNU")
list(APPEND mi_cflags -Wall -Wextra)
endif()
if(CMAKE_C_COMPILER_ID MATCHES "Intel")
list(APPEND mi_cflags -Wall)
endif()
if(CMAKE_C_COMPILER_ID MATCHES "MSVC|Intel")
set(MI_USE_CXX "ON")
endif()
@ -184,6 +198,10 @@ endif()
if(MI_SEE_ASM)
message(STATUS "Generate assembly listings (MI_SEE_ASM=ON)")
list(APPEND mi_cflags -save-temps)
if(CMAKE_C_COMPILER_ID MATCHES "AppleClang|Clang")
message(STATUS "No GNU Line marker")
list(APPEND mi_cflags -Wno-gnu-line-marker)
endif()
endif()
if(MI_CHECK_FULL)
@ -246,7 +264,7 @@ if(MI_DEBUG_UBSAN)
message(WARNING "Can only use undefined-behavior sanitizer with clang++ (MI_DEBUG_UBSAN=ON but ignored)")
endif()
else()
message(WARNING "Can only use thread sanitizer with a debug build (CMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE})")
message(WARNING "Can only use undefined-behavior sanitizer with a debug build (CMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE})")
endif()
endif()
@ -262,24 +280,38 @@ if(MI_USE_CXX)
endif()
endif()
if(CMAKE_SYSTEM_NAME MATCHES "Haiku")
SET(CMAKE_INSTALL_LIBDIR ~/config/non-packaged/lib)
SET(CMAKE_INSTALL_INCLUDEDIR ~/config/non-packaged/headers)
if(CMAKE_SYSTEM_NAME MATCHES "Linux|Android")
if(MI_NO_THP)
message(STATUS "Disable transparent huge pages support (MI_NO_THP=ON)")
list(APPEND mi_defines MI_NO_THP=1)
endif()
endif()
if(MI_LIBC_MUSL)
message(STATUS "Assume using musl libc (MI_LIBC_MUSL=ON) (this implies MI_LOCAL_DYNAMIC_TLS=ON)")
set(MI_LOCAL_DYNAMIC_TLS "ON")
list(APPEND mi_defines MI_LIBC_MUSL=1)
endif()
# On Haiku use `-DCMAKE_INSTALL_PREFIX` instead, issue #788
# if(CMAKE_SYSTEM_NAME MATCHES "Haiku")
# SET(CMAKE_INSTALL_LIBDIR ~/config/non-packaged/lib)
# SET(CMAKE_INSTALL_INCLUDEDIR ~/config/non-packaged/headers)
# endif()
# Compiler flags
if(CMAKE_C_COMPILER_ID MATCHES "AppleClang|Clang|GNU")
list(APPEND mi_cflags -Wall -Wextra -Wno-unknown-pragmas -fvisibility=hidden)
list(APPEND mi_cflags -Wno-unknown-pragmas -fvisibility=hidden)
if(NOT MI_USE_CXX)
list(APPEND mi_cflags -Wstrict-prototypes)
endif()
if(CMAKE_C_COMPILER_ID MATCHES "AppleClang|Clang")
list(APPEND mi_cflags -Wpedantic -Wno-static-in-inline)
list(APPEND mi_cflags -Wno-static-in-inline)
endif()
endif()
if(CMAKE_C_COMPILER_ID MATCHES "Intel")
list(APPEND mi_cflags -Wall -fvisibility=hidden)
list(APPEND mi_cflags -fvisibility=hidden)
endif()
if(CMAKE_C_COMPILER_ID MATCHES "AppleClang|Clang|GNU|Intel" AND NOT CMAKE_SYSTEM_NAME MATCHES "Haiku")

18
SECURITY.md
View File

@ -1,20 +1,20 @@
<!-- BEGIN MICROSOFT SECURITY.MD V0.0.8 BLOCK -->
<!-- BEGIN MICROSOFT SECURITY.MD V0.0.9 BLOCK -->
## Security
Microsoft takes the security of our software products and services seriously, which includes all source code repositories managed through our GitHub organizations, which include [Microsoft](https://github.com/microsoft), [Azure](https://github.com/Azure), [DotNet](https://github.com/dotnet), [AspNet](https://github.com/aspnet), [Xamarin](https://github.com/xamarin), and [our GitHub organizations](https://opensource.microsoft.com/).
Microsoft takes the security of our software products and services seriously, which includes all source code repositories managed through our GitHub organizations, which include [Microsoft](https://github.com/Microsoft), [Azure](https://github.com/Azure), [DotNet](https://github.com/dotnet), [AspNet](https://github.com/aspnet) and [Xamarin](https://github.com/xamarin).
If you believe you have found a security vulnerability in any Microsoft-owned repository that meets [Microsoft's definition of a security vulnerability](https://aka.ms/opensource/security/definition), please report it to us as described below.
If you believe you have found a security vulnerability in any Microsoft-owned repository that meets [Microsoft's definition of a security vulnerability](https://aka.ms/security.md/definition), please report it to us as described below.
## Reporting Security Issues
**Please do not report security vulnerabilities through public GitHub issues.**
Instead, please report them to the Microsoft Security Response Center (MSRC) at [https://msrc.microsoft.com/create-report](https://aka.ms/opensource/security/create-report).
Instead, please report them to the Microsoft Security Response Center (MSRC) at [https://msrc.microsoft.com/create-report](https://aka.ms/security.md/msrc/create-report).
If you prefer to submit without logging in, send email to [secure@microsoft.com](mailto:secure@microsoft.com). If possible, encrypt your message with our PGP key; please download it from the [Microsoft Security Response Center PGP Key page](https://aka.ms/opensource/security/pgpkey).
If you prefer to submit without logging in, send email to [secure@microsoft.com](mailto:secure@microsoft.com). If possible, encrypt your message with our PGP key; please download it from the [Microsoft Security Response Center PGP Key page](https://aka.ms/security.md/msrc/pgp).
You should receive a response within 24 hours. If for some reason you do not, please follow up via email to ensure we received your original message. Additional information can be found at [microsoft.com/msrc](https://aka.ms/opensource/security/msrc).
You should receive a response within 24 hours. If for some reason you do not, please follow up via email to ensure we received your original message. Additional information can be found at [microsoft.com/msrc](https://www.microsoft.com/msrc).
Please include the requested information listed below (as much as you can provide) to help us better understand the nature and scope of the possible issue:
@ -28,7 +28,11 @@ Please include the requested information listed below (as much as you can provid
This information will help us triage your report more quickly.
<<<<<<< HEAD
If you are reporting for a bug bounty, more complete reports can contribute to a higher bounty award. Please visit our [Microsoft Bug Bounty Program](https://aka.ms/opensource/security/bounty) page for more details about our active programs.
=======
If you are reporting for a bug bounty, more complete reports can contribute to a higher bounty award. Please visit our [Microsoft Bug Bounty Program](https://aka.ms/security.md/msrc/bounty) page for more details about our active programs.
>>>>>>> dev-slice
## Preferred Languages
@ -36,6 +40,6 @@ We prefer all communications to be in English.
## Policy
Microsoft follows the principle of [Coordinated Vulnerability Disclosure](https://aka.ms/opensource/security/cvd).
Microsoft follows the principle of [Coordinated Vulnerability Disclosure](https://aka.ms/security.md/cvd).
<!-- END MICROSOFT SECURITY.MD BLOCK -->

BIN
bin/minject.exe
View File

Binary file not shown.

BIN
bin/minject32.exe
View File

Binary file not shown.

71
bin/readme.md Normal file
View File

@ -0,0 +1,71 @@
# Windows Override
<span id="override_on_windows">Dynamically overriding on mimalloc on Windows</span>
is robust and has the particular advantage to be able to redirect all malloc/free calls that go through
the (dynamic) C runtime allocator, including those from other DLL's or libraries.
As it intercepts all allocation calls on a low level, it can be used reliably
on large programs that include other 3rd party components.
There are four requirements to make the overriding work robustly:
1. Use the C-runtime library as a DLL (using the `/MD` or `/MDd` switch).
2. Link your program explicitly with `mimalloc-override.dll` library.
To ensure the `mimalloc-override.dll` is loaded at run-time it is easiest to insert some
call to the mimalloc API in the `main` function, like `mi_version()`
(or use the `/INCLUDE:mi_version` switch on the linker). See the `mimalloc-override-test` project
for an example on how to use this.
3. The `mimalloc-redirect.dll` (or `mimalloc-redirect32.dll`) must be put
in the same folder as the main `mimalloc-override.dll` at runtime (as it is a dependency of that DLL).
The redirection DLL ensures that all calls to the C runtime malloc API get redirected to
mimalloc functions (which reside in `mimalloc-override.dll`).
4. Ensure the `mimalloc-override.dll` comes as early as possible in the import
list of the final executable (so it can intercept all potential allocations).
For best performance on Windows with C++, it
is also recommended to also override the `new`/`delete` operations (by including
[`mimalloc-new-delete.h`](../include/mimalloc-new-delete.h)
a single(!) source file in your project).
The environment variable `MIMALLOC_DISABLE_REDIRECT=1` can be used to disable dynamic
overriding at run-time. Use `MIMALLOC_VERBOSE=1` to check if mimalloc was successfully redirected.
## Minject
We cannot always re-link an executable with `mimalloc-override.dll`, and similarly, we cannot always
ensure the the DLL comes first in the import table of the final executable.
In many cases though we can patch existing executables without any recompilation
if they are linked with the dynamic C runtime (`ucrtbase.dll`) -- just put the `mimalloc-override.dll`
into the import table (and put `mimalloc-redirect.dll` in the same folder)
Such patching can be done for example with [CFF Explorer](https://ntcore.com/?page_id=388).
The `minject` program can also do this from the command line, use `minject --help` for options:
```
> minject --help
minject:
Injects the mimalloc dll into the import table of a 64-bit executable,
and/or ensures that it comes first in het import table.
usage:
> minject [options] <exe>
options:
-h --help show this help
-v --verbose be verbose
-l --list only list imported modules
-i --inplace update the exe in-place (make sure there is a backup!)
-f --force always overwrite without prompting
--postfix=<p> use <p> as a postfix to the mimalloc dll (default is 'override')
e.g. use --postfix=override-debug to link with mimalloc-override-debug.dll
notes:
Without '--inplace' an injected <exe> is generated with the same name ending in '-mi'.
Ensure 'mimalloc-redirect.dll' is in the same folder as the mimalloc dll.
examples:
> minject --list myprogram.exe
> minject --force --inplace myprogram.exe
```

2
cmake/mimalloc-config-version.cmake
View File

@ -1,6 +1,6 @@
set(mi_version_major 2)
set(mi_version_minor 1)
set(mi_version_patch 2)
set(mi_version_patch 4)
set(mi_version ${mi_version_major}.${mi_version_minor})
set(PACKAGE_VERSION ${mi_version})

2
doc/doxyfile
View File

@ -466,7 +466,7 @@ LOOKUP_CACHE_SIZE = 0
# than 0 to get more control over the balance between CPU load and processing
# speed. At this moment only the input processing can be done using multiple
# threads. Since this is still an experimental feature the default is set to 1,
# which efficively disables parallel processing. Please report any issues you
# which effectively disables parallel processing. Please report any issues you
# encounter. Generating dot graphs in parallel is controlled by the
# DOT_NUM_THREADS setting.
# Minimum value: 0, maximum value: 32, default value: 1.

16
doc/mimalloc-doc.h
View File

@ -168,7 +168,7 @@ void* mi_expand(void* p, size_t newsize);
/// @returns A pointer to a block of \a count * \a size bytes, or \a NULL
/// if out of memory or if \a count * \a size overflows.
///
/// If there is no overflow, it behaves exactly like `mi_malloc(p,count*size)`.
/// If there is no overflow, it behaves exactly like `mi_malloc(count*size)`.
/// @see mi_calloc()
/// @see mi_zallocn()
void* mi_mallocn(size_t count, size_t size);
@ -441,7 +441,7 @@ bool mi_manage_os_memory(void* start, size_t size, bool is_committed, bool is_la
/// @param pages The number of 1GiB pages to reserve.
/// @param numa_nodes The number of nodes do evenly divide the pages over, or 0 for using the actual number of NUMA nodes.
/// @param timeout_msecs Maximum number of milli-seconds to try reserving, or 0 for no timeout.
/// @returns 0 if successfull, \a ENOMEM if running out of memory, or \a ETIMEDOUT if timed out.
/// @returns 0 if successful, \a ENOMEM if running out of memory, or \a ETIMEDOUT if timed out.
///
/// The reserved memory is used by mimalloc to satisfy allocations.
/// May quit before \a timeout_msecs are expired if it estimates it will take more than
@ -455,7 +455,7 @@ int mi_reserve_huge_os_pages_interleave(size_t pages, size_t numa_nodes, size_t
/// @param pages The number of 1GiB pages to reserve.
/// @param numa_node The NUMA node where the memory is reserved (start at 0).
/// @param timeout_msecs Maximum number of milli-seconds to try reserving, or 0 for no timeout.
/// @returns 0 if successfull, \a ENOMEM if running out of memory, or \a ETIMEDOUT if timed out.
/// @returns 0 if successful, \a ENOMEM if running out of memory, or \a ETIMEDOUT if timed out.
///
/// The reserved memory is used by mimalloc to satisfy allocations.
/// May quit before \a timeout_msecs are expired if it estimates it will take more than
@ -468,7 +468,7 @@ int mi_reserve_huge_os_pages_at(size_t pages, int numa_node, size_t timeout_msec
/// Is the C runtime \a malloc API redirected?
/// @returns \a true if all malloc API calls are redirected to mimalloc.
///
/// Currenty only used on Windows.
/// Currently only used on Windows.
bool mi_is_redirected();
/// Return process information (time and memory usage).
@ -499,11 +499,11 @@ void mi_process_info(size_t* elapsed_msecs, size_t* user_msecs, size_t* system_m
/// \{
/// The maximum supported alignment size (currently 1MiB).
#define MI_ALIGNMENT_MAX (1024*1024UL)
#define MI_BLOCK_ALIGNMENT_MAX (1024*1024UL)
/// Allocate \a size bytes aligned by \a alignment.
/// @param size number of bytes to allocate.
/// @param alignment the minimal alignment of the allocated memory. Must be less than #MI_ALIGNMENT_MAX.
/// @param alignment the minimal alignment of the allocated memory. Must be less than #MI_BLOCK_ALIGNMENT_MAX.
/// @returns pointer to the allocated memory or \a NULL if out of memory.
/// The returned pointer is aligned by \a alignment, i.e.
/// `(uintptr_t)p % alignment == 0`.
@ -558,7 +558,7 @@ mi_heap_t* mi_heap_new();
/// Delete a previously allocated heap.
/// This will release resources and migrate any
/// still allocated blocks in this heap (efficienty)
/// still allocated blocks in this heap (efficiently)
/// to the default heap.
///
/// If \a heap is the default heap, the default
@ -888,7 +888,7 @@ void mi_free_aligned(void* p, size_t alignment);
///
/// Note: use the `mimalloc-new-delete.h` header to override the \a new
/// and \a delete operators globally. The wrappers here are mostly
/// for convience for library writers that need to interface with
/// for convenience for library writers that need to interface with
/// mimalloc from C++.
///
/// \{

2
docs/bench.html
View File

@ -100,7 +100,7 @@ $(document).ready(function(){initNavTree('bench.html',''); initResizable(); });
<div class="contents">
<div class="textblock"><p>We tested <em>mimalloc</em> against many other top allocators over a wide range of benchmarks, ranging from various real world programs to synthetic benchmarks that see how the allocator behaves under more extreme circumstances.</p>
<p>In our benchmarks, <em>mimalloc</em> always outperforms all other leading allocators (<em>jemalloc</em>, <em>tcmalloc</em>, <em>Hoard</em>, etc) (Jan 2021), and usually uses less memory (up to 25% more in the worst case). A nice property is that it does <em>consistently</em> well over the wide range of benchmarks.</p>
<p>See the <a href="https://github.com/microsoft/mimalloc#Performance">Performance</a> section in the <em>mimalloc</em> repository for benchmark results, or the the technical report for detailed benchmark results. </p>
<p>See the <a href="https://github.com/microsoft/mimalloc#Performance">Performance</a> section in the <em>mimalloc</em> repository for benchmark results, or the technical report for detailed benchmark results. </p>
</div></div><!-- contents -->
</div><!-- PageDoc -->
</div><!-- doc-content -->

1
ide/vs2017/mimalloc-override.vcxproj
View File

@ -238,6 +238,7 @@
<ClCompile Include="..\..\src\bitmap.c" />
<ClCompile Include="..\..\src\heap.c" />
<ClCompile Include="..\..\src\init.c" />
<ClCompile Include="..\..\src\libc.c" />
<ClCompile Include="..\..\src\prim\prim.c" />
<ClCompile Include="..\..\src\options.c" />
<ClCompile Include="..\..\src\os.c" />

3
ide/vs2017/mimalloc-override.vcxproj.filters
View File

@ -94,5 +94,8 @@
<ClCompile Include="..\..\src\segment-map.c">
<Filter>Source Files</Filter>
</ClCompile>
<ClCompile Include="..\..\src\libc.c">
<Filter>Source Files</Filter>
</ClCompile>
</ItemGroup>
</Project>

1
ide/vs2017/mimalloc.vcxproj
View File

@ -227,6 +227,7 @@
<ClCompile Include="..\..\src\bitmap.c" />
<ClCompile Include="..\..\src\heap.c" />
<ClCompile Include="..\..\src\init.c" />
<ClCompile Include="..\..\src\libc.c" />
<ClCompile Include="..\..\src\prim\prim.c" />
<ClCompile Include="..\..\src\options.c" />
<ClCompile Include="..\..\src\page-queue.c">

3
ide/vs2017/mimalloc.vcxproj.filters
View File

@ -65,6 +65,9 @@
<ClCompile Include="..\..\src\segment-map.c">
<Filter>Source Files</Filter>
</ClCompile>
<ClCompile Include="..\..\src\libc.c">
<Filter>Source Files</Filter>
</ClCompile>
</ItemGroup>
<ItemGroup>
<ClInclude Include="$(ProjectDir)..\..\include\mimalloc.h">

1
ide/vs2019/mimalloc-override.vcxproj
View File

@ -238,6 +238,7 @@
<ClCompile Include="..\..\src\bitmap.c" />
<ClCompile Include="..\..\src\heap.c" />
<ClCompile Include="..\..\src\init.c" />
<ClCompile Include="..\..\src\libc.c" />
<ClCompile Include="..\..\src\prim\prim.c" />
<ClCompile Include="..\..\src\options.c" />
<ClCompile Include="..\..\src\os.c" />

3
ide/vs2019/mimalloc-override.vcxproj.filters
View File

@ -55,6 +55,9 @@
<ClCompile Include="..\..\src\segment-map.c">
<Filter>Source Files</Filter>
</ClCompile>
<ClCompile Include="..\..\src\libc.c">
<Filter>Source Files</Filter>
</ClCompile>
</ItemGroup>
<ItemGroup>
<ClInclude Include="$(ProjectDir)..\..\include\mimalloc.h">

1
ide/vs2019/mimalloc.vcxproj
View File

@ -219,6 +219,7 @@
</ClCompile>
<ClCompile Include="..\..\src\heap.c" />
<ClCompile Include="..\..\src\init.c" />
<ClCompile Include="..\..\src\libc.c" />
<ClCompile Include="..\..\src\prim\prim.c" />
<ClCompile Include="..\..\src\prim\windows\prim.c">
<ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">true</ExcludedFromBuild>

3
ide/vs2019/mimalloc.vcxproj.filters
View File

@ -58,6 +58,9 @@
<ClCompile Include="..\..\src\segment-map.c">
<Filter>Source Files</Filter>
</ClCompile>
<ClCompile Include="..\..\src\libc.c">
<Filter>Source Files</Filter>
</ClCompile>
</ItemGroup>
<ItemGroup>
<ClInclude Include="$(ProjectDir)..\..\include\mimalloc.h">

1
ide/vs2022/mimalloc-override.vcxproj
View File

@ -240,6 +240,7 @@
<ClCompile Include="..\..\src\bitmap.c" />
<ClCompile Include="..\..\src\heap.c" />
<ClCompile Include="..\..\src\init.c" />
<ClCompile Include="..\..\src\libc.c" />
<ClCompile Include="..\..\src\prim\prim.c" />
<ClCompile Include="..\..\src\prim\windows\prim.c">
<ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">true</ExcludedFromBuild>

3
ide/vs2022/mimalloc-override.vcxproj.filters
View File

@ -55,6 +55,9 @@
<ClCompile Include="..\..\src\stats.c">
<Filter>Sources</Filter>
</ClCompile>
<ClCompile Include="..\..\src\libc.c">
<Filter>Sources</Filter>
</ClCompile>
</ItemGroup>
<ItemGroup>
<ClInclude Include="..\..\include\mimalloc\atomic.h">

7
ide/vs2022/mimalloc.vcxproj
View File

@ -217,8 +217,15 @@
<ClCompile Include="..\..\src\bitmap.c">
<ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">false</ExcludedFromBuild>
</ClCompile>
<ClCompile Include="..\..\src\free.c">
<ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|x64'">true</ExcludedFromBuild>
<ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">true</ExcludedFromBuild>
<ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|Win32'">true</ExcludedFromBuild>
<ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Release|x64'">true</ExcludedFromBuild>
</ClCompile>
<ClCompile Include="..\..\src\heap.c" />
<ClCompile Include="..\..\src\init.c" />
<ClCompile Include="..\..\src\libc.c" />
<ClCompile Include="..\..\src\prim\prim.c" />
<ClCompile Include="..\..\src\prim\windows\prim.c">
<ExcludedFromBuild Condition="'$(Configuration)|$(Platform)'=='Debug|Win32'">true</ExcludedFromBuild>

6
ide/vs2022/mimalloc.vcxproj.filters
View File

@ -55,6 +55,12 @@
<ClCompile Include="..\..\src\stats.c">
<Filter>Sources</Filter>
</ClCompile>
<ClCompile Include="..\..\src\libc.c">
<Filter>Sources</Filter>
</ClCompile>
<ClCompile Include="..\..\src\free.c">
<Filter>Sources</Filter>
</ClCompile>
</ItemGroup>
<ItemGroup>
<ClInclude Include="..\..\src\bitmap.h">

35
include/mimalloc.h
View File

@ -8,7 +8,7 @@ terms of the MIT license. A copy of the license can be found in the file
#ifndef MIMALLOC_H
#define MIMALLOC_H
#define MI_MALLOC_VERSION 212 // major + 2 digits minor
#define MI_MALLOC_VERSION 214 // major + 2 digits minor
// ------------------------------------------------------
// Compiler specific attributes
@ -275,7 +275,7 @@ mi_decl_export int mi_reserve_huge_os_pages_at(size_t pages, int numa_node, size
mi_decl_export int mi_reserve_os_memory(size_t size, bool commit, bool allow_large) mi_attr_noexcept;
mi_decl_export bool mi_manage_os_memory(void* start, size_t size, bool is_committed, bool is_large, bool is_zero, int numa_node) mi_attr_noexcept;
mi_decl_export void mi_debug_show_arenas(void) mi_attr_noexcept;
mi_decl_export void mi_debug_show_arenas(bool show_inuse, bool show_abandoned, bool show_purge) mi_attr_noexcept;
// Experimental: heaps associated with specific memory arena's
typedef int mi_arena_id_t;
@ -321,37 +321,40 @@ typedef enum mi_option_e {
mi_option_show_errors, // print error messages
mi_option_show_stats, // print statistics on termination
mi_option_verbose, // print verbose messages
// the following options are experimental (see src/options.h)
// advanced options
mi_option_eager_commit, // eager commit segments? (after `eager_commit_delay` segments) (=1)
mi_option_arena_eager_commit, // eager commit arenas? Use 2 to enable just on overcommit systems (=2)
mi_option_purge_decommits, // should a memory purge decommit (or only reset) (=1)
mi_option_allow_large_os_pages, // allow large (2MiB) OS pages, implies eager commit
mi_option_reserve_huge_os_pages, // reserve N huge OS pages (1GiB/page) at startup
mi_option_purge_decommits, // should a memory purge decommit? (=1). Set to 0 to use memory reset on a purge (instead of decommit)
mi_option_allow_large_os_pages, // allow large (2 or 4 MiB) OS pages, implies eager commit. If false, also disables THP for the process.
mi_option_reserve_huge_os_pages, // reserve N huge OS pages (1GiB pages) at startup
mi_option_reserve_huge_os_pages_at, // reserve huge OS pages at a specific NUMA node
mi_option_reserve_os_memory, // reserve specified amount of OS memory in an arena at startup
mi_option_deprecated_segment_cache,
mi_option_deprecated_page_reset,
mi_option_abandoned_page_purge, // immediately purge delayed purges on thread termination
mi_option_deprecated_segment_reset,
mi_option_eager_commit_delay,
mi_option_purge_delay, // memory purging is delayed by N milli seconds; use 0 for immediate purging or -1 for no purging at all.
mi_option_eager_commit_delay, // the first N segments per thread are not eagerly committed (but per page in the segment on demand)
mi_option_purge_delay, // memory purging is delayed by N milli seconds; use 0 for immediate purging or -1 for no purging at all. (=10)
mi_option_use_numa_nodes, // 0 = use all available numa nodes, otherwise use at most N nodes.
mi_option_limit_os_alloc, // 1 = do not use OS memory for allocation (but only programmatically reserved arenas)
mi_option_os_tag, // tag used for OS logging (macOS only for now)
mi_option_disallow_os_alloc, // 1 = do not use OS memory for allocation (but only programmatically reserved arenas)
mi_option_os_tag, // tag used for OS logging (macOS only for now) (=100)
mi_option_max_errors, // issue at most N error messages
mi_option_max_warnings, // issue at most N warning messages
mi_option_max_segment_reclaim,
mi_option_destroy_on_exit, // if set, release all memory on exit; sometimes used for dynamic unloading but can be unsafe.
mi_option_arena_reserve, // initial memory size in KiB for arena reservation (1GiB on 64-bit)
mi_option_arena_purge_mult,
mi_option_max_segment_reclaim, // max. percentage of the abandoned segments can be reclaimed per try (=10%)
mi_option_destroy_on_exit, // if set, release all memory on exit; sometimes used for dynamic unloading but can be unsafe
mi_option_arena_reserve, // initial memory size in KiB for arena reservation (= 1 GiB on 64-bit)
mi_option_arena_purge_mult, // multiplier for `purge_delay` for the purging delay for arenas (=10)
mi_option_purge_extend_delay,
mi_option_abandoned_reclaim_on_free, // allow to reclaim an abandoned segment on a free (=1)
mi_option_disallow_arena_alloc, // 1 = do not use arena's for allocation (except if using specific arena id's)
_mi_option_last,
// legacy option names
mi_option_large_os_pages = mi_option_allow_large_os_pages,
mi_option_eager_region_commit = mi_option_arena_eager_commit,
mi_option_reset_decommits = mi_option_purge_decommits,
mi_option_reset_delay = mi_option_purge_delay,
mi_option_abandoned_page_reset = mi_option_abandoned_page_purge
mi_option_abandoned_page_reset = mi_option_abandoned_page_purge,
mi_option_limit_os_alloc = mi_option_disallow_os_alloc
} mi_option_t;
@ -494,7 +497,7 @@ template<class T, bool _mi_destroy> struct _mi_heap_stl_allocator_common : publi
using typename _mi_stl_allocator_common<T>::value_type;
using typename _mi_stl_allocator_common<T>::pointer;
_mi_heap_stl_allocator_common(mi_heap_t* hp) : heap(hp) { } /* will not delete nor destroy the passed in heap */
_mi_heap_stl_allocator_common(mi_heap_t* hp) : heap(hp, [](mi_heap_t*) {}) {} /* will not delete nor destroy the passed in heap */
#if (__cplusplus >= 201703L) // C++17
mi_decl_nodiscard T* allocate(size_type count) { return static_cast<T*>(mi_heap_alloc_new_n(this->heap.get(), count, sizeof(T))); }

12
include/mimalloc/atomic.h
View File

@ -23,7 +23,9 @@ terms of the MIT license. A copy of the license can be found in the file
#define _Atomic(tp) std::atomic<tp>
#define mi_atomic(name) std::atomic_##name
#define mi_memory_order(name) std::memory_order_##name
#if !defined(ATOMIC_VAR_INIT) || (__cplusplus >= 202002L) // c++20, see issue #571
#if (__cplusplus >= 202002L) // c++20, see issue #571
#define MI_ATOMIC_VAR_INIT(x) x
#elif !defined(ATOMIC_VAR_INIT)
#define MI_ATOMIC_VAR_INIT(x) x
#else
#define MI_ATOMIC_VAR_INIT(x) ATOMIC_VAR_INIT(x)
@ -39,7 +41,9 @@ terms of the MIT license. A copy of the license can be found in the file
#include <stdatomic.h>
#define mi_atomic(name) atomic_##name
#define mi_memory_order(name) memory_order_##name
#if !defined(ATOMIC_VAR_INIT) || (__STDC_VERSION__ >= 201710L) // c17, see issue #735
#if (__STDC_VERSION__ >= 201710L) // c17, see issue #735
#define MI_ATOMIC_VAR_INIT(x) x
#elif !defined(ATOMIC_VAR_INIT)
#define MI_ATOMIC_VAR_INIT(x) x
#else
#define MI_ATOMIC_VAR_INIT(x) ATOMIC_VAR_INIT(x)
@ -129,7 +133,9 @@ static inline void mi_atomic_maxi64_relaxed(volatile int64_t* p, int64_t x) {
#elif defined(_MSC_VER)
// MSVC C compilation wrapper that uses Interlocked operations to model C11 atomics.
#ifndef WIN32_LEAN_AND_MEAN
#define WIN32_LEAN_AND_MEAN
#endif
#include <windows.h>
#include <intrin.h>
#ifdef _WIN64
@ -323,7 +329,9 @@ static inline void mi_atomic_yield(void) {
std::this_thread::yield();
}
#elif defined(_WIN32)
#ifndef WIN32_LEAN_AND_MEAN
#define WIN32_LEAN_AND_MEAN
#endif
#include <windows.h>
static inline void mi_atomic_yield(void) {
YieldProcessor();

75
include/mimalloc/internal.h
View File

@ -30,14 +30,17 @@ terms of the MIT license. A copy of the license can be found in the file
#define mi_decl_noinline __declspec(noinline)
#define mi_decl_thread __declspec(thread)
#define mi_decl_cache_align __declspec(align(MI_CACHE_LINE))
#define mi_decl_weak
#elif (defined(__GNUC__) && (__GNUC__ >= 3)) || defined(__clang__) // includes clang and icc
#define mi_decl_noinline __attribute__((noinline))
#define mi_decl_thread __thread
#define mi_decl_cache_align __attribute__((aligned(MI_CACHE_LINE)))
#define mi_decl_weak __attribute__((weak))
#else
#define mi_decl_noinline
#define mi_decl_thread __thread // hope for the best :-)
#define mi_decl_cache_align
#define mi_decl_weak
#endif
#if defined(__EMSCRIPTEN__) && !defined(__wasi__)
@ -122,9 +125,21 @@ void* _mi_arena_alloc(size_t size, bool commit, bool allow_large, mi_arena_
void* _mi_arena_alloc_aligned(size_t size, size_t alignment, size_t align_offset, bool commit, bool allow_large, mi_arena_id_t req_arena_id, mi_memid_t* memid, mi_os_tld_t* tld);
bool _mi_arena_memid_is_suitable(mi_memid_t memid, mi_arena_id_t request_arena_id);
bool _mi_arena_contains(const void* p);
void _mi_arena_collect(bool force_purge, mi_stats_t* stats);
void _mi_arenas_collect(bool force_purge, mi_stats_t* stats);
void _mi_arena_unsafe_destroy_all(mi_stats_t* stats);
bool _mi_arena_segment_clear_abandoned(mi_segment_t* segment);
void _mi_arena_segment_mark_abandoned(mi_segment_t* segment);
size_t _mi_arena_segment_abandoned_count(void);
typedef struct mi_arena_field_cursor_s { // abstract
mi_arena_id_t start;
int count;
size_t bitmap_idx;
} mi_arena_field_cursor_t;
void _mi_arena_field_cursor_init(mi_heap_t* heap, mi_arena_field_cursor_t* current);
mi_segment_t* _mi_arena_segment_clear_abandoned_next(mi_arena_field_cursor_t* previous);
// "segment-map.c"
void _mi_segment_map_allocated_at(const mi_segment_t* segment);
void _mi_segment_map_freed_at(const mi_segment_t* segment);
@ -134,7 +149,7 @@ mi_page_t* _mi_segment_page_alloc(mi_heap_t* heap, size_t block_size, size_t pag
void _mi_segment_page_free(mi_page_t* page, bool force, mi_segments_tld_t* tld);
void _mi_segment_page_abandon(mi_page_t* page, mi_segments_tld_t* tld);
bool _mi_segment_try_reclaim_abandoned( mi_heap_t* heap, bool try_all, mi_segments_tld_t* tld);
void _mi_segment_thread_collect(mi_segments_tld_t* tld);
void _mi_segment_collect(mi_segment_t* segment, bool force, mi_segments_tld_t* tld);
#if MI_HUGE_PAGE_ABANDON
void _mi_segment_huge_page_free(mi_segment_t* segment, mi_page_t* page, mi_block_t* block);
@ -146,6 +161,7 @@ uint8_t* _mi_segment_page_start(const mi_segment_t* segment, const mi_page_t*
void _mi_abandoned_reclaim_all(mi_heap_t* heap, mi_segments_tld_t* tld);
void _mi_abandoned_await_readers(void);
void _mi_abandoned_collect(mi_heap_t* heap, bool force, mi_segments_tld_t* tld);
bool _mi_segment_attempt_reclaim(mi_heap_t* heap, mi_segment_t* segment);
// "page.c"
void* _mi_malloc_generic(mi_heap_t* heap, size_t size, bool zero, size_t huge_alignment) mi_attr_noexcept mi_attr_malloc;
@ -187,19 +203,22 @@ void* _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t size, bool
void* _mi_heap_malloc_zero(mi_heap_t* heap, size_t size, bool zero) mi_attr_noexcept;
void* _mi_heap_malloc_zero_ex(mi_heap_t* heap, size_t size, bool zero, size_t huge_alignment) mi_attr_noexcept; // called from `_mi_heap_malloc_aligned`
void* _mi_heap_realloc_zero(mi_heap_t* heap, void* p, size_t newsize, bool zero) mi_attr_noexcept;
mi_block_t* _mi_page_ptr_unalign(const mi_segment_t* segment, const mi_page_t* page, const void* p);
mi_block_t* _mi_page_ptr_unalign(const mi_page_t* page, const void* p);
bool _mi_free_delayed_block(mi_block_t* block);
void _mi_free_generic(const mi_segment_t* segment, mi_page_t* page, bool is_local, void* p) mi_attr_noexcept; // for runtime integration
void _mi_free_generic(mi_segment_t* segment, mi_page_t* page, bool is_local, void* p) mi_attr_noexcept; // for runtime integration
void _mi_padding_shrink(const mi_page_t* page, const mi_block_t* block, const size_t min_size);
// option.c, c primitives
// "libc.c"
#include <stdarg.h>
void _mi_vsnprintf(char* buf, size_t bufsize, const char* fmt, va_list args);
void _mi_snprintf(char* buf, size_t buflen, const char* fmt, ...);
char _mi_toupper(char c);
int _mi_strnicmp(const char* s, const char* t, size_t n);
void _mi_strlcpy(char* dest, const char* src, size_t dest_size);
void _mi_strlcat(char* dest, const char* src, size_t dest_size);
size_t _mi_strlen(const char* s);
size_t _mi_strnlen(const char* s, size_t max_len);
bool _mi_getenv(const char* name, char* result, size_t result_size);
#if MI_DEBUG>1
bool _mi_page_is_valid(mi_page_t* page);
@ -308,6 +327,17 @@ static inline uintptr_t _mi_align_down(uintptr_t sz, size_t alignment) {
}
}
// Align a pointer upwards
static inline void* mi_align_up_ptr(void* p, size_t alignment) {
return (void*)_mi_align_up((uintptr_t)p, alignment);
}
// Align a pointer downwards
static inline void* mi_align_down_ptr(void* p, size_t alignment) {
return (void*)_mi_align_down((uintptr_t)p, alignment);
}
// Divide upwards: `s <= _mi_divide_up(s,d)*d < s+d`.
static inline uintptr_t _mi_divide_up(uintptr_t size, size_t divider) {
mi_assert_internal(divider != 0);
@ -407,9 +437,14 @@ static inline mi_page_t* _mi_heap_get_free_small_page(mi_heap_t* heap, size_t si
// Large aligned blocks may be aligned at N*MI_SEGMENT_SIZE (inside a huge segment > MI_SEGMENT_SIZE),
// and we need align "down" to the segment info which is `MI_SEGMENT_SIZE` bytes before it;
// therefore we align one byte before `p`.
// We check for NULL afterwards on 64-bit systems to improve codegen for `mi_free`.
static inline mi_segment_t* _mi_ptr_segment(const void* p) {
mi_assert_internal(p != NULL);
return (mi_segment_t*)(((uintptr_t)p - 1) & ~MI_SEGMENT_MASK);
mi_segment_t* const segment = (mi_segment_t*)(((uintptr_t)p - 1) & ~MI_SEGMENT_MASK);
#if MI_INTPTR_SIZE <= 4
return (p==NULL ? NULL : segment);
#else
return ((intptr_t)segment <= 0 ? NULL : segment);
#endif
}
static inline mi_page_t* mi_slice_to_page(mi_slice_t* s) {
@ -424,6 +459,7 @@ static inline mi_slice_t* mi_page_to_slice(mi_page_t* p) {
// Segment belonging to a page
static inline mi_segment_t* _mi_page_segment(const mi_page_t* page) {
mi_assert_internal(page!=NULL);
mi_segment_t* segment = _mi_ptr_segment(page);
mi_assert_internal(segment == NULL || ((mi_slice_t*)page >= segment->slices && (mi_slice_t*)page < segment->slices + segment->slice_entries));
return segment;
@ -452,31 +488,28 @@ static inline mi_page_t* _mi_segment_page_of(const mi_segment_t* segment, const
}
// Quick page start for initialized pages
static inline uint8_t* _mi_page_start(const mi_segment_t* segment, const mi_page_t* page, size_t* page_size) {
return _mi_segment_page_start(segment, page, page_size);
static inline uint8_t* mi_page_start(const mi_page_t* page) {
mi_assert_internal(page->page_start != NULL);
mi_assert_expensive(_mi_segment_page_start(_mi_page_segment(page),page,NULL) == page->page_start);
return page->page_start;
}
// Get the page containing the pointer
static inline mi_page_t* _mi_ptr_page(void* p) {
mi_assert_internal(p!=NULL);
return _mi_segment_page_of(_mi_ptr_segment(p), p);
}
// Get the block size of a page (special case for huge objects)
static inline size_t mi_page_block_size(const mi_page_t* page) {
const size_t bsize = page->xblock_size;
mi_assert_internal(bsize > 0);
if mi_likely(bsize < MI_HUGE_BLOCK_SIZE) {
return bsize;
}
else {
size_t psize;
_mi_segment_page_start(_mi_page_segment(page), page, &psize);
return psize;
}
mi_assert_internal(page->block_size > 0);
return page->block_size;
}
static inline bool mi_page_is_huge(const mi_page_t* page) {
return (_mi_page_segment(page)->kind == MI_SEGMENT_HUGE);
mi_assert_internal((page->is_huge && _mi_page_segment(page)->kind == MI_SEGMENT_HUGE) ||
(!page->is_huge && _mi_page_segment(page)->kind != MI_SEGMENT_HUGE));
return page->is_huge;
}
// Get the usable block size of a page without fixed padding.

79
include/mimalloc/prim.h
View File

@ -125,33 +125,23 @@ void _mi_prim_thread_associate_default_heap(mi_heap_t* heap);
// for each thread (unequal to zero).
//-------------------------------------------------------------------
// defined in `init.c`; do not use these directly
extern mi_decl_thread mi_heap_t* _mi_heap_default; // default heap to allocate from
extern bool _mi_process_is_initialized; // has mi_process_init been called?
static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept;
#if defined(_WIN32)
#define WIN32_LEAN_AND_MEAN
#include <windows.h>
static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept {
// Windows: works on Intel and ARM in both 32- and 64-bit
return (uintptr_t)NtCurrentTeb();
}
// We use assembly for a fast thread id on the main platforms. The TLS layout depends on
// both the OS and libc implementation so we use specific tests for each main platform.
// On some libc + platform combinations we can directly access a thread-local storage (TLS) slot.
// The TLS layout depends on both the OS and libc implementation so we use specific tests for each main platform.
// If you test on another platform and it works please send a PR :-)
// see also https://akkadia.org/drepper/tls.pdf for more info on the TLS register.
#elif defined(__GNUC__) && ( \
//
// Note: on most platforms this is not actually used anymore as we prefer `__builtin_thread_pointer()` nowadays.
// However, we do still use it with older clang compilers and Apple OS (as we use TLS slot for the default heap there).
#if defined(__GNUC__) && ( \
(defined(__GLIBC__) && (defined(__x86_64__) || defined(__i386__) || defined(__arm__) || defined(__aarch64__))) \
|| (defined(__APPLE__) && (defined(__x86_64__) || defined(__aarch64__))) \
|| (defined(__APPLE__) && (defined(__x86_64__) || defined(__aarch64__) || defined(__POWERPC__))) \
|| (defined(__BIONIC__) && (defined(__x86_64__) || defined(__i386__) || defined(__arm__) || defined(__aarch64__))) \
|| (defined(__FreeBSD__) && (defined(__x86_64__) || defined(__i386__) || defined(__aarch64__))) \
|| (defined(__OpenBSD__) && (defined(__x86_64__) || defined(__i386__) || defined(__aarch64__))) \
)
#define MI_HAS_TLS_SLOT
static inline void* mi_prim_tls_slot(size_t slot) mi_attr_noexcept {
void* res;
const size_t ofs = (slot*sizeof(void*));
@ -175,6 +165,9 @@ static inline void* mi_prim_tls_slot(size_t slot) mi_attr_noexcept {
__asm__ volatile ("mrs %0, tpidr_el0" : "=r" (tcb));
#endif
res = tcb[slot];
#elif defined(__APPLE__) && defined(__POWERPC__) // ppc, issue #781
MI_UNUSED(ofs);
res = pthread_getspecific(slot);
#endif
return res;
}
@ -202,9 +195,53 @@ static inline void mi_prim_tls_slot_set(size_t slot, void* value) mi_attr_noexce
__asm__ volatile ("mrs %0, tpidr_el0" : "=r" (tcb));
#endif
tcb[slot] = value;
#elif defined(__APPLE__) && defined(__POWERPC__) // ppc, issue #781
MI_UNUSED(ofs);
pthread_setspecific(slot, value);
#endif
}
#endif
// Do we have __builtin_thread_pointer? (do not make this a compound test as it fails on older gcc's, see issue #851)
#if defined(__has_builtin)
#if __has_builtin(__builtin_thread_pointer)
#define MI_HAS_BUILTIN_THREAD_POINTER 1
#endif
#elif defined(__GNUC__) && (__GNUC__ >= 7) && defined(__aarch64__) // special case aarch64 for older gcc versions (issue #851)
#define MI_HAS_BUILTIN_THREAD_POINTER 1
#endif
// defined in `init.c`; do not use these directly
extern mi_decl_thread mi_heap_t* _mi_heap_default; // default heap to allocate from
extern bool _mi_process_is_initialized; // has mi_process_init been called?
static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept;
// Get a unique id for the current thread.
#if defined(_WIN32)
#ifndef WIN32_LEAN_AND_MEAN
#define WIN32_LEAN_AND_MEAN
#endif
#include <windows.h>
static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept {
// Windows: works on Intel and ARM in both 32- and 64-bit
return (uintptr_t)NtCurrentTeb();
}
#elif MI_HAS_BUILTIN_THREAD_POINTER && \
(!defined(__APPLE__)) && /* on apple (M1) the wrong register is read (tpidr_el0 instead of tpidrro_el0) so fall back to TLS slot assembly (<https://github.com/microsoft/mimalloc/issues/343#issuecomment-763272369>)*/ \
(!defined(__clang_major__) || __clang_major__ >= 14) // older clang versions emit bad code; fall back to using the TLS slot (<https://lore.kernel.org/linux-arm-kernel/202110280952.352F66D8@keescook/T/>)
static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept {
// Works on most Unix based platforms
return (uintptr_t)__builtin_thread_pointer();
}
#elif defined(MI_HAS_TLS_SLOT)
static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept {
#if defined(__BIONIC__)
// issue #384, #495: on the Bionic libc (Android), slot 1 is the thread id
@ -251,7 +288,6 @@ static inline mi_heap_t* mi_prim_get_default_heap(void);
#if defined(MI_MALLOC_OVERRIDE)
#if defined(__APPLE__) // macOS
#define MI_TLS_SLOT 89 // seems unused?
// #define MI_TLS_RECURSE_GUARD 1
// other possible unused ones are 9, 29, __PTK_FRAMEWORK_JAVASCRIPTCORE_KEY4 (94), __PTK_FRAMEWORK_GC_KEY9 (112) and __PTK_FRAMEWORK_OLDGC_KEY9 (89)
// see <https://github.com/rweichler/substrate/blob/master/include/pthread_machdep.h>
#elif defined(__OpenBSD__)
@ -269,6 +305,9 @@ static inline mi_heap_t* mi_prim_get_default_heap(void);
#if defined(MI_TLS_SLOT)
# if !defined(MI_HAS_TLS_SLOT)
# error "trying to use a TLS slot for the default heap, but the mi_prim_tls_slot primitives are not defined"
# endif
static inline mi_heap_t* mi_prim_get_default_heap(void) {
mi_heap_t* heap = (mi_heap_t*)mi_prim_tls_slot(MI_TLS_SLOT);

2
include/mimalloc/track.h
View File

@ -82,7 +82,9 @@ defined, undefined, or not accessible at all:
#define MI_TRACK_HEAP_DESTROY 1
#define MI_TRACK_TOOL "ETW"
#ifndef WIN32_LEAN_AND_MEAN
#define WIN32_LEAN_AND_MEAN
#endif
#include <windows.h>
#include "../src/prim/windows/etw.h"

145
include/mimalloc/types.h
View File

@ -1,5 +1,5 @@
/* ----------------------------------------------------------------------------
Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
Copyright (c) 2018-2024, Microsoft Research, Daan Leijen
This is free software; you can redistribute it and/or modify it under the
terms of the MIT license. A copy of the license can be found in the file
"LICENSE" at the root of this distribution.
@ -13,9 +13,12 @@ terms of the MIT license. A copy of the license can be found in the file
// mi_heap_t : all data for a thread-local heap, contains
// lists of all managed heap pages.
// mi_segment_t : a larger chunk of memory (32GiB) from where pages
// are allocated.
// mi_page_t : a mimalloc page (usually 64KiB or 512KiB) from
// are allocated. A segment is divided in slices (64KiB) from
// which pages are allocated.
// mi_page_t : a "mimalloc" page (usually 64KiB or 512KiB) from
// where objects are allocated.
// Note: we write "OS page" for OS memory pages while
// using plain "page" for mimalloc pages (`mi_page_t`).
// --------------------------------------------------------------------------
@ -89,10 +92,11 @@ terms of the MIT license. A copy of the license can be found in the file
#endif
// We used to abandon huge pages but to eagerly deallocate if freed from another thread,
// but that makes it not possible to visit them during a heap walk or include them in a
// `mi_heap_destroy`. We therefore instead reset/decommit the huge blocks if freed from
// another thread so most memory is available until it gets properly freed by the owning thread.
// We used to abandon huge pages in order to eagerly deallocate it if freed from another thread.
// Unfortunately, that makes it not possible to visit them during a heap walk or include them in a
// `mi_heap_destroy`. We therefore instead reset/decommit the huge blocks nowadays if freed from
// another thread so the memory becomes "virtually" available (and eventually gets properly freed by
// the owning thread).
// #define MI_HUGE_PAGE_ABANDON 1
@ -157,17 +161,24 @@ typedef int32_t mi_ssize_t;
// Main tuning parameters for segment and page sizes
// Sizes for 64-bit (usually divide by two for 32-bit)
#ifndef MI_SEGMENT_SLICE_SHIFT
#define MI_SEGMENT_SLICE_SHIFT (13 + MI_INTPTR_SHIFT) // 64KiB (32KiB on 32-bit)
#endif
#ifndef MI_SEGMENT_SHIFT
#if MI_INTPTR_SIZE > 4
#define MI_SEGMENT_SHIFT ( 9 + MI_SEGMENT_SLICE_SHIFT) // 32MiB
#else
#define MI_SEGMENT_SHIFT ( 7 + MI_SEGMENT_SLICE_SHIFT) // 4MiB on 32-bit
#endif
#endif
#ifndef MI_SMALL_PAGE_SHIFT
#define MI_SMALL_PAGE_SHIFT (MI_SEGMENT_SLICE_SHIFT) // 64KiB
#endif
#ifndef MI_MEDIUM_PAGE_SHIFT
#define MI_MEDIUM_PAGE_SHIFT ( 3 + MI_SMALL_PAGE_SHIFT) // 512KiB
#endif
// Derived constants
#define MI_SEGMENT_SIZE (MI_ZU(1)<<MI_SEGMENT_SHIFT)
@ -192,17 +203,22 @@ typedef int32_t mi_ssize_t;
#error "mimalloc internal: define more bins"
#endif
// Maximum slice offset (15)
#define MI_MAX_SLICE_OFFSET ((MI_ALIGNMENT_MAX / MI_SEGMENT_SLICE_SIZE) - 1)
// Used as a special value to encode block sizes in 32 bits.
#define MI_HUGE_BLOCK_SIZE ((uint32_t)(2*MI_GiB))
// blocks up to this size are always allocated aligned
#define MI_MAX_ALIGN_GUARANTEE (8*MI_MAX_ALIGN_SIZE)
// Alignments over MI_ALIGNMENT_MAX are allocated in dedicated huge page segments
#define MI_ALIGNMENT_MAX (MI_SEGMENT_SIZE >> 1)
// Alignments over MI_BLOCK_ALIGNMENT_MAX are allocated in dedicated huge page segments
#define MI_BLOCK_ALIGNMENT_MAX (MI_SEGMENT_SIZE >> 1)
// Maximum slice count (255) for which we can find the page for interior pointers
#define MI_MAX_SLICE_OFFSET_COUNT ((MI_BLOCK_ALIGNMENT_MAX / MI_SEGMENT_SLICE_SIZE) - 1)
// we never allocate more than PTRDIFF_MAX (see also <https://sourceware.org/ml/libc-announce/2019/msg00001.html>)
// on 64-bit+ systems we also limit the maximum allocation size such that the slice count fits in 32-bits. (issue #877)
#if PTRDIFF_MAX >= (MI_SEGMENT_SLIZE_SIZE * UINT32_MAX)
#define MI_MAX_ALLOC_SIZE (MI_SEGMENT_SLICE_SIZE * (UINT32_MAX-1))
#else
#define MI_MAX_ALLOC_SIZE PTRDIFF_MAX
#endif
// ------------------------------------------------------
@ -227,7 +243,7 @@ typedef enum mi_delayed_e {
MI_USE_DELAYED_FREE = 0, // push on the owning heap thread delayed list
MI_DELAYED_FREEING = 1, // temporary: another thread is accessing the owning heap
MI_NO_DELAYED_FREE = 2, // optimize: push on page local thread free queue if another block is already in the heap thread delayed free list
MI_NEVER_DELAYED_FREE = 3 // sticky, only resets on page reclaim
MI_NEVER_DELAYED_FREE = 3 // sticky: used for abondoned pages without a owning heap; this only resets on page reclaim
} mi_delayed_t;
@ -266,7 +282,6 @@ typedef uintptr_t mi_thread_free_t;
// implement a monotonic heartbeat. The `thread_free` list is needed for
// avoiding atomic operations in the common case.
//
//
// `used - |thread_free|` == actual blocks that are in use (alive)
// `used - |thread_free| + |free| + |local_free| == capacity`
//
@ -274,16 +289,13 @@ typedef uintptr_t mi_thread_free_t;
// the number of memory accesses in the `mi_page_all_free` function(s).
//
// Notes:
// - Access is optimized for `mi_free` and `mi_page_alloc` (in `alloc.c`)
// - Access is optimized for `free.c:mi_free` and `alloc.c:mi_page_alloc`
// - Using `uint16_t` does not seem to slow things down
// - The size is 8 words on 64-bit which helps the page index calculations
// (and 10 words on 32-bit, and encoded free lists add 2 words. Sizes 10
// and 12 are still good for address calculation)
// - To limit the structure size, the `xblock_size` is 32-bits only; for
// blocks > MI_HUGE_BLOCK_SIZE the size is determined from the segment page size
// - `thread_free` uses the bottom bits as a delayed-free flags to optimize
// - The size is 12 words on 64-bit which helps the page index calculations
// (and 14 words on 32-bit, and encoded free lists add 2 words)
// - `xthread_free` uses the bottom bits as a delayed-free flags to optimize
// concurrent frees where only the first concurrent free adds to the owning
// heap `thread_delayed_free` list (see `alloc.c:mi_free_block_mt`).
// heap `thread_delayed_free` list (see `free.c:mi_free_block_mt`).
// The invariant is that no-delayed-free is only set if there is
// at least one block that will be added, or as already been added, to
// the owning heap `thread_delayed_free` list. This guarantees that pages
@ -294,7 +306,8 @@ typedef struct mi_page_s {
uint32_t slice_offset; // distance from the actual page data slice (0 if a page)
uint8_t is_committed:1; // `true` if the page virtual memory is committed
uint8_t is_zero_init:1; // `true` if the page was initially zero initialized
uint8_t is_huge:1; // `true` if the page is in a huge segment (`segment->kind == MI_SEGMENT_HUGE`)
// padding
// layout like this to optimize access in `mi_malloc` and `mi_free`
uint16_t capacity; // number of blocks committed, must be the first field, see `segment.c:page_clear`
uint16_t reserved; // number of blocks reserved in memory
@ -303,9 +316,12 @@ typedef struct mi_page_s {
uint8_t retire_expire:7; // expiration count for retired blocks
mi_block_t* free; // list of available free blocks (`malloc` allocates from this list)
uint32_t used; // number of blocks in use (including blocks in `local_free` and `thread_free`)
uint32_t xblock_size; // size available in each block (always `>0`)
mi_block_t* local_free; // list of deferred free blocks by this thread (migrates to `free`)
uint16_t used; // number of blocks in use (including blocks in `thread_free`)
uint8_t block_size_shift; // if not zero, then `(1 << block_size_shift) == block_size` (only used for fast path in `free.c:_mi_page_ptr_unalign`)
// padding
size_t block_size; // size available in each block (always `>0`)
uint8_t* page_start; // start of the page area containing the blocks
#if (MI_ENCODE_FREELIST || MI_PADDING)
uintptr_t keys[2]; // two random keys to encode the free lists (see `_mi_block_next`) or padding canary
@ -317,10 +333,8 @@ typedef struct mi_page_s {
struct mi_page_s* next; // next page owned by this thread with the same `block_size`
struct mi_page_s* prev; // previous page owned by this thread with the same `block_size`
// 64-bit 9 words, 32-bit 12 words, (+2 for secure)
#if MI_INTPTR_SIZE==8
uintptr_t padding[1];
#endif
// 64-bit 11 words, 32-bit 13 words, (+2 for secure)
void* padding[1];
} mi_page_t;
@ -331,14 +345,15 @@ typedef struct mi_page_s {
typedef enum mi_page_kind_e {
MI_PAGE_SMALL, // small blocks go into 64KiB pages inside a segment
MI_PAGE_MEDIUM, // medium blocks go into medium pages inside a segment
MI_PAGE_LARGE, // larger blocks go into a page of just one block
MI_PAGE_HUGE, // huge blocks (> 16 MiB) are put into a single page in a single segment.
MI_PAGE_MEDIUM, // medium blocks go into 512KiB pages inside a segment
MI_PAGE_LARGE, // larger blocks go into a single page spanning a whole segment
MI_PAGE_HUGE // a huge page is a single page in a segment of variable size
// used for blocks `> MI_LARGE_OBJ_SIZE_MAX` or an aligment `> MI_BLOCK_ALIGNMENT_MAX`.
} mi_page_kind_t;
typedef enum mi_segment_kind_e {
MI_SEGMENT_NORMAL, // MI_SEGMENT_SIZE size with pages inside.
MI_SEGMENT_HUGE, // > MI_LARGE_SIZE_MAX segment with just one huge page inside.
MI_SEGMENT_HUGE, // segment with just one huge page inside.
} mi_segment_kind_t;
// ------------------------------------------------------
@ -371,13 +386,17 @@ typedef mi_page_t mi_slice_t;
typedef int64_t mi_msecs_t;
// ---------------------------------------------------------------
// a memory id tracks the provenance of arena/OS allocated memory
// ---------------------------------------------------------------
// Memory can reside in arena's, direct OS allocated, or statically allocated. The memid keeps track of this.
typedef enum mi_memkind_e {
MI_MEM_NONE, // not allocated
MI_MEM_EXTERNAL, // not owned by mimalloc but provided externally (via `mi_manage_os_memory` for example)
MI_MEM_STATIC, // allocated in a static area and should not be freed (for arena meta data for example)
MI_MEM_OS, // allocated from the OS
MI_MEM_OS_HUGE, // allocated as huge os pages
MI_MEM_OS_HUGE, // allocated as huge OS pages (usually 1GiB, pinned to physical memory)
MI_MEM_OS_REMAP, // allocated in a remapable area (i.e. using `mremap`)
MI_MEM_ARENA // allocated from an arena (the usual case)
} mi_memkind_t;
@ -394,7 +413,7 @@ typedef struct mi_memid_os_info {
typedef struct mi_memid_arena_info {
size_t block_index; // index in the arena
mi_arena_id_t id; // arena id (>= 1)
bool is_exclusive; // the arena can only be used for specific arena allocations
bool is_exclusive; // this arena can only be used for specific arena allocations
} mi_memid_arena_info_t;
typedef struct mi_memid_s {
@ -402,47 +421,56 @@ typedef struct mi_memid_s {
mi_memid_os_info_t os; // only used for MI_MEM_OS
mi_memid_arena_info_t arena; // only used for MI_MEM_ARENA
} mem;
bool is_pinned; // `true` if we cannot decommit/reset/protect in this memory (e.g. when allocated using large OS pages)
bool is_pinned; // `true` if we cannot decommit/reset/protect in this memory (e.g. when allocated using large (2Mib) or huge (1GiB) OS pages)
bool initially_committed;// `true` if the memory was originally allocated as committed
bool initially_zero; // `true` if the memory was originally zero initialized
mi_memkind_t memkind;
} mi_memid_t;
// Segments are large allocated memory blocks (8mb on 64 bit) from
// the OS. Inside segments we allocated fixed size _pages_ that
// contain blocks.
// -----------------------------------------------------------------------------------------
// Segments are large allocated memory blocks (8mb on 64 bit) from arenas or the OS.
//
// Inside segments we allocated fixed size mimalloc pages (`mi_page_t`) that contain blocks.
// The start of a segment is this structure with a fixed number of slice entries (`slices`)
// usually followed by a guard OS page and the actual allocation area with pages.
// While a page is not allocated, we view it's data as a `mi_slice_t` (instead of a `mi_page_t`).
// Of any free area, the first slice has the info and `slice_offset == 0`; for any subsequent
// slices part of the area, the `slice_offset` is the byte offset back to the first slice
// (so we can quickly find the page info on a free, `internal.h:_mi_segment_page_of`).
// For slices, the `block_size` field is repurposed to signify if a slice is used (`1`) or not (`0`).
// Small and medium pages use a fixed amount of slices to reduce slice fragmentation, while
// large and huge pages span a variable amount of slices.
typedef struct mi_segment_s {
// constant fields
mi_memid_t memid; // memory id for arena allocation
bool allow_decommit;
bool allow_purge;
mi_memid_t memid; // memory id for arena/OS allocation
bool allow_decommit; // can we decommmit the memory
bool allow_purge; // can we purge the memory (reset or decommit)
size_t segment_size;
// segment fields
mi_msecs_t purge_expire;
mi_commit_mask_t purge_mask;
mi_commit_mask_t commit_mask;
_Atomic(struct mi_segment_s*) abandoned_next;
mi_msecs_t purge_expire; // purge slices in the `purge_mask` after this time
mi_commit_mask_t purge_mask; // slices that can be purged
mi_commit_mask_t commit_mask; // slices that are currently committed
// from here is zero initialized
struct mi_segment_s* next; // the list of freed segments in the cache (must be first field, see `segment.c:mi_segment_init`)
bool was_reclaimed; // true if it was reclaimed (used to limit on-free reclamation)
size_t abandoned; // abandoned pages (i.e. the original owning thread stopped) (`abandoned <= used`)
size_t abandoned_visits; // count how often this segment is visited in the abandoned list (to force reclaim it it is too long)
size_t abandoned_visits; // count how often this segment is visited during abondoned reclamation (to force reclaim if it takes too long)
size_t used; // count of pages in use
uintptr_t cookie; // verify addresses in debug mode: `mi_ptr_cookie(segment) == segment->cookie`
size_t segment_slices; // for huge segments this may be different from `MI_SLICES_PER_SEGMENT`
size_t segment_info_slices; // initial slices we are using segment info and possible guard pages.
size_t segment_info_slices; // initial count of slices that we are using for segment info and possible guard pages.
// layout like this to optimize access in `mi_free`
mi_segment_kind_t kind;
size_t slice_entries; // entries in the `slices` array, at most `MI_SLICES_PER_SEGMENT`
_Atomic(mi_threadid_t) thread_id; // unique id of the thread owning this segment
mi_slice_t slices[MI_SLICES_PER_SEGMENT+1]; // one more for huge blocks with large alignment
mi_slice_t slices[MI_SLICES_PER_SEGMENT+1]; // one extra final entry for huge blocks with large alignment
} mi_segment_t;
@ -499,8 +527,6 @@ typedef struct mi_padding_s {
// A heap owns a set of pages.
struct mi_heap_s {
mi_tld_t* tld;
mi_page_t* pages_free_direct[MI_PAGES_DIRECT]; // optimize: array where every entry points a page with possibly free blocks in the corresponding queue for that size.
mi_page_queue_t pages[MI_BIN_FULL + 1]; // queue of pages for each size class (or "bin")
_Atomic(mi_block_t*) thread_delayed_free;
mi_threadid_t thread_id; // thread this heap belongs too
mi_arena_id_t arena_id; // arena id if the heap belongs to a specific arena (or 0)
@ -512,6 +538,8 @@ struct mi_heap_s {
size_t page_retired_max; // largest retired index into the `pages` array.
mi_heap_t* next; // list of heaps per thread
bool no_reclaim; // `true` if this heap should not reclaim abandoned pages
mi_page_t* pages_free_direct[MI_PAGES_DIRECT]; // optimize: array where every entry points a page with possibly free blocks in the corresponding queue for that size.
mi_page_queue_t pages[MI_BIN_FULL + 1]; // queue of pages for each size class (or "bin")
};
@ -600,6 +628,9 @@ typedef struct mi_stats_s {
mi_stat_counter_t normal_count;
mi_stat_counter_t huge_count;
mi_stat_counter_t large_count;
mi_stat_counter_t arena_count;
mi_stat_counter_t arena_crossover_count;
mi_stat_counter_t arena_rollback_count;
#if MI_STAT>1
mi_stat_count_t normal_bins[MI_BIN_HUGE+1];
#endif
@ -624,6 +655,7 @@ void _mi_stat_counter_increase(mi_stat_counter_t* stat, size_t amount);
#define mi_heap_stat_increase(heap,stat,amount) mi_stat_increase( (heap)->tld->stats.stat, amount)
#define mi_heap_stat_decrease(heap,stat,amount) mi_stat_decrease( (heap)->tld->stats.stat, amount)
// ------------------------------------------------------
// Thread Local data
// ------------------------------------------------------
@ -652,6 +684,7 @@ typedef struct mi_segments_tld_s {
size_t peak_count; // peak number of segments
size_t current_size; // current size of all segments
size_t peak_size; // peak size of all segments
size_t reclaim_count;// number of reclaimed (abandoned) segments
mi_stats_t* stats; // points to tld stats
mi_os_tld_t* os; // points to os stats
} mi_segments_tld_t;

97
readme.md
View File

@ -12,8 +12,8 @@ is a general purpose allocator with excellent [performance](#performance) charac
Initially developed by Daan Leijen for the runtime systems of the
[Koka](https://koka-lang.github.io) and [Lean](https://github.com/leanprover/lean) languages.
Latest release tag: `v2.1.2` (2023-04-24).
Latest stable tag: `v1.8.2` (2023-04-24).
Latest release tag: `v2.1.4` (2024-04-22).
Latest stable tag: `v1.8.4` (2024-04-22).
mimalloc is a drop-in replacement for `malloc` and can be used in other programs
without code changes, for example, on dynamically linked ELF-based systems (Linux, BSD, etc.) you can use it as:
@ -29,6 +29,8 @@ It also includes a robust way to override the default allocator in [Windows](#ov
bounded worst-case times with reference counting).
Partly due to its simplicity, mimalloc has been ported to many systems (Windows, macOS,
Linux, WASM, various BSD's, Haiku, MUSL, etc) and has excellent support for dynamic overriding.
At the same time, it is an industrial strength allocator that runs (very) large scale
distributed services on thousands of machines with excellent worst case latencies.
- __free list sharding__: instead of one big free list (per size class) we have
many smaller lists per "mimalloc page" which reduces fragmentation and
increases locality --
@ -78,6 +80,16 @@ Note: the `v2.x` version has a new algorithm for managing internal mimalloc page
and fragmentation compared to mimalloc `v1.x` (especially for large workloads). Should otherwise have similar performance
(see [below](#performance)); please report if you observe any significant performance regression.
* 2024-04-22, `v1.8.4`, `v2.1.4`: Fixes various bugs and build issues. Add `MI_LIBC_MUSL` cmake flag for musl builds.
Free-ing code is refactored into a separate module (`free.c`). Mimalloc page info is simplified with the block size
directly available (and new `block_size_shift` to improve aligned block free-ing).
New approach to collection of abandoned segments: When
a thread terminates the segments it owns are abandoned (containing still live objects) and these can be
reclaimed by other threads. We no longer use a list of abandoned segments but this is now done using bitmaps in arena's
which is more concurrent (and more aggressive). Abandoned memory can now also be reclaimed if a thread frees an object in
an abandoned page (which can be disabled using `mi_option_abandoned_reclaim_on_free`). The option `mi_option_max_segment_reclaim`
gives a maximum percentage of abandoned segments that can be reclaimed per try (=10%).
* 2023-04-24, `v1.8.2`, `v2.1.2`: Fixes build issues on freeBSD, musl, and C17 (UE 5.1.1). Reduce code size/complexity
by removing regions and segment-cache's and only use arenas with improved memory purging -- this may improve memory
usage as well for larger services. Renamed options for consistency. Improved Valgrind and ASAN checking.
@ -89,7 +101,7 @@ Note: the `v2.x` version has a new algorithm for managing internal mimalloc page
abstraction layer to make it easier to port and separate platform dependent code (in `src/prim`). Fixed C++ STL compilation on older Microsoft C++ compilers, and various small bug fixes.
* 2022-12-23, `v1.7.9`, `v2.0.9`: Supports building with [asan](#asan) and improved [Valgrind](#valgrind) support.
Support abitrary large alignments (in particular for `std::pmr` pools).
Support arbitrary large alignments (in particular for `std::pmr` pools).
Added C++ STL allocators attached to a specific heap (thanks @vmarkovtsev).
Heap walks now visit all object (including huge objects). Support Windows nano server containers (by Johannes Schindelin,@dscho). Various small bug fixes.
@ -141,7 +153,7 @@ mimalloc is used in various large scale low-latency services and programs, for e
## Windows
Open `ide/vs2019/mimalloc.sln` in Visual Studio 2019 and build.
Open `ide/vs2022/mimalloc.sln` in Visual Studio 2022 and build.
The `mimalloc` project builds a static library (in `out/msvc-x64`), while the
`mimalloc-override` project builds a DLL for overriding malloc
in the entire program.
@ -221,7 +233,7 @@ target_link_libraries(myapp PUBLIC mimalloc-static)
to link with the static library. See `test\CMakeLists.txt` for an example.
For best performance in C++ programs, it is also recommended to override the
global `new` and `delete` operators. For convience, mimalloc provides
global `new` and `delete` operators. For convenience, mimalloc provides
[`mimalloc-new-delete.h`](https://github.com/microsoft/mimalloc/blob/master/include/mimalloc-new-delete.h) which does this for you -- just include it in a single(!) source file in your project.
In C++, mimalloc also provides the `mi_stl_allocator` struct which implements the `std::allocator`
interface.
@ -277,17 +289,23 @@ You can set further options either programmatically (using [`mi_option_set`](htt
Advanced options:
- `MIMALLOC_ARENA_EAGER_COMMIT=2`: turns on eager commit for the large arenas (usually 1GiB) from which mimalloc
allocates segments and pages. Set this to 2 (default) to
only enable this on overcommit systems (e.g. Linux). Set this to 1 to enable explicitly on other systems
as well (like Windows or macOS) which may improve performance (as the whole arena is committed at once).
Note that eager commit only increases the commit but not the actual the peak resident set
(rss) so it is generally ok to enable this.
- `MIMALLOC_PURGE_DELAY=N`: the delay in `N` milli-seconds (by default `10`) after which mimalloc will purge
OS pages that are not in use. This signals to the OS that the underlying physical memory can be reused which
can reduce memory fragmentation especially in long running (server) programs. Setting `N` to `0` purges immediately when
a page becomes unused which can improve memory usage but also decreases performance. Setting `N` to a higher
value like `100` can improve performance (sometimes by a lot) at the cost of potentially using more memory at times.
Setting it to `-1` disables purging completely.
- `MIMALLOC_ARENA_EAGER_COMMIT=1`: turns on eager commit for the large arenas (usually 1GiB) from which mimalloc
allocates segments and pages. This is by default
only enabled on overcommit systems (e.g. Linux) but enabling it explicitly on other systems (like Windows or macOS)
may improve performance. Note that eager commit only increases the commit but not the actual the peak resident set
(rss) so it is generally ok to enable this.
- `MIMALLOC_PURGE_DECOMMITS=1`: By default "purging" memory means unused memory is decommitted (`MEM_DECOMMIT` on Windows,
`MADV_DONTNEED` (which decresease rss immediately) on `mmap` systems). Set this to 0 to instead "reset" unused
memory on a purge (`MEM_RESET` on Windows, generally `MADV_FREE` (which does not decrease rss immediately) on `mmap` systems).
Mimalloc generally does not "free" OS memory but only "purges" OS memory, in other words, it tries to keep virtual
address ranges and decommits within those ranges (to make the underlying physical memory available to other processes).
Further options for large workloads and services:
@ -295,9 +313,10 @@ Further options for large workloads and services:
at runtime. Setting `N` to 1 may avoid problems in some virtual environments. Also, setting it to a lower number than
the actual NUMA nodes is fine and will only cause threads to potentially allocate more memory across actual NUMA
nodes (but this can happen in any case as NUMA local allocation is always a best effort but not guaranteed).
- `MIMALLOC_ALLOW_LARGE_OS_PAGES=1`: use large OS pages (2MiB) when available; for some workloads this can significantly
improve performance. Use `MIMALLOC_VERBOSE` to check if the large OS pages are enabled -- usually one needs
to explicitly allow large OS pages (as on [Windows][windows-huge] and [Linux][linux-huge]). However, sometimes
- `MIMALLOC_ALLOW_LARGE_OS_PAGES=1`: use large OS pages (2 or 4MiB) when available; for some workloads this can significantly
improve performance. When this option is disabled, it also disables transparent huge pages (THP) for the process
(on Linux and Android). Use `MIMALLOC_VERBOSE` to check if the large OS pages are enabled -- usually one needs
to explicitly give permissions for large OS pages (as on [Windows][windows-huge] and [Linux][linux-huge]). However, sometimes
the OS is very slow to reserve contiguous physical memory for large OS pages so use with care on systems that
can have fragmented memory (for that reason, we generally recommend to use `MIMALLOC_RESERVE_HUGE_OS_PAGES` instead whenever possible).
- `MIMALLOC_RESERVE_HUGE_OS_PAGES=N`: where `N` is the number of 1GiB _huge_ OS pages. This reserves the huge pages at
@ -306,11 +325,12 @@ Further options for large workloads and services:
OS pages, use with care as reserving
contiguous physical memory can take a long time when memory is fragmented (but reserving the huge pages is done at
startup only once).
Note that we usually need to explicitly enable huge OS pages (as on [Windows][windows-huge] and [Linux][linux-huge])).
Note that we usually need to explicitly give permission for huge OS pages (as on [Windows][windows-huge] and [Linux][linux-huge])).
With huge OS pages, it may be beneficial to set the setting
`MIMALLOC_EAGER_COMMIT_DELAY=N` (`N` is 1 by default) to delay the initial `N` segments (of 4MiB)
of a thread to not allocate in the huge OS pages; this prevents threads that are short lived
and allocate just a little to take up space in the huge OS page area (which cannot be purged).
and allocate just a little to take up space in the huge OS page area (which cannot be purged as huge OS pages are pinned
to physical memory).
The huge pages are usually allocated evenly among NUMA nodes.
We can use `MIMALLOC_RESERVE_HUGE_OS_PAGES_AT=N` where `N` is the numa node (starting at 0) to allocate all
the huge pages at a specific numa node instead.
@ -391,32 +411,41 @@ the [shell](https://stackoverflow.com/questions/43941322/dyld-insert-libraries-i
### Dynamic Override on Windows
<span id="override_on_windows">Overriding on Windows</span> is robust and has the
particular advantage to be able to redirect all malloc/free calls that go through
<span id="override_on_windows">Dynamically overriding on mimalloc on Windows</span>
is robust and has the particular advantage to be able to redirect all malloc/free calls that go through
the (dynamic) C runtime allocator, including those from other DLL's or libraries.
As it intercepts all allocation calls on a low level, it can be used reliably
on large programs that include other 3rd party components.
There are four requirements to make the overriding work robustly:
The overriding on Windows requires that you link your program explicitly with
the mimalloc DLL and use the C-runtime library as a DLL (using the `/MD` or `/MDd` switch).
Also, the `mimalloc-redirect.dll` (or `mimalloc-redirect32.dll`) must be put
in the same folder as the main `mimalloc-override.dll` at runtime (as it is a dependency).
The redirection DLL ensures that all calls to the C runtime malloc API get redirected to
mimalloc (in `mimalloc-override.dll`).
To ensure the mimalloc DLL is loaded at run-time it is easiest to insert some
1. Use the C-runtime library as a DLL (using the `/MD` or `/MDd` switch).
2. Link your program explicitly with `mimalloc-override.dll` library.
To ensure the `mimalloc-override.dll` is loaded at run-time it is easiest to insert some
call to the mimalloc API in the `main` function, like `mi_version()`
(or use the `/INCLUDE:mi_version` switch on the linker). See the `mimalloc-override-test` project
for an example on how to use this. For best performance on Windows with C++, it
for an example on how to use this.
3. The [`mimalloc-redirect.dll`](bin) (or `mimalloc-redirect32.dll`) must be put
in the same folder as the main `mimalloc-override.dll` at runtime (as it is a dependency of that DLL).
The redirection DLL ensures that all calls to the C runtime malloc API get redirected to
mimalloc functions (which reside in `mimalloc-override.dll`).
4. Ensure the `mimalloc-override.dll` comes as early as possible in the import
list of the final executable (so it can intercept all potential allocations).
For best performance on Windows with C++, it
is also recommended to also override the `new`/`delete` operations (by including
[`mimalloc-new-delete.h`](https://github.com/microsoft/mimalloc/blob/master/include/mimalloc-new-delete.h) a single(!) source file in your project).
[`mimalloc-new-delete.h`](include/mimalloc-new-delete.h)
a single(!) source file in your project).
The environment variable `MIMALLOC_DISABLE_REDIRECT=1` can be used to disable dynamic
overriding at run-time. Use `MIMALLOC_VERBOSE=1` to check if mimalloc was successfully redirected.
(Note: in principle, it is possible to even patch existing executables without any recompilation
We cannot always re-link an executable with `mimalloc-override.dll`, and similarly, we cannot always
ensure the the DLL comes first in the import table of the final executable.
In many cases though we can patch existing executables without any recompilation
if they are linked with the dynamic C runtime (`ucrtbase.dll`) -- just put the `mimalloc-override.dll`
into the import table (and put `mimalloc-redirect.dll` in the same folder)
Such patching can be done for example with [CFF Explorer](https://ntcore.com/?page_id=388)).
Such patching can be done for example with [CFF Explorer](https://ntcore.com/?page_id=388) or
the [`minject`](bin) program.
## Static override
@ -438,7 +467,7 @@ This is provided by [`mimalloc-override.h`](https://github.com/microsoft/mimallo
under your control or otherwise mixing of pointers from different heaps may occur!
## Tools
# Tools
Generally, we recommend using the standard allocator with memory tracking tools, but mimalloc
can also be build to support the [address sanitizer][asan] or the excellent [Valgrind] tool.
@ -446,7 +475,7 @@ Moreover, it can be build to support Windows event tracing ([ETW]).
This has a small performance overhead but does allow detecting memory leaks and byte-precise
buffer overflows directly on final executables. See also the `test/test-wrong.c` file to test with various tools.
### Valgrind
## Valgrind
To build with [valgrind] support, use the `MI_TRACK_VALGRIND=ON` cmake option:
@ -480,7 +509,7 @@ Valgrind support is in its initial development -- please report any issues.
[Valgrind]: https://valgrind.org/
[valgrind-soname]: https://valgrind.org/docs/manual/manual-core.html#opt.soname-synonyms
### ASAN
## ASAN
To build with the address sanitizer, use the `-DMI_TRACK_ASAN=ON` cmake option:
@ -509,7 +538,7 @@ Adress sanitizer support is in its initial development -- please report any issu
[asan]: https://github.com/google/sanitizers/wiki/AddressSanitizer
### ETW
## ETW
Event tracing for Windows ([ETW]) provides a high performance way to capture all allocations though
mimalloc and analyze them later. To build with ETW support, use the `-DMI_TRACK_ETW=ON` cmake option.

8
src/alloc-aligned.c
View File

@ -33,7 +33,7 @@ static mi_decl_noinline void* mi_heap_malloc_zero_aligned_at_fallback(mi_heap_t*
void* p;
size_t oversize;
if mi_unlikely(alignment > MI_ALIGNMENT_MAX) {
if mi_unlikely(alignment > MI_BLOCK_ALIGNMENT_MAX) {
// use OS allocation for very large alignment and allocate inside a huge page (dedicated segment with 1 page)
// This can support alignments >= MI_SEGMENT_SIZE by ensuring the object can be aligned at a point in the
// first (and single) page such that the segment info is `MI_SEGMENT_SIZE` bytes before it (so it can be found by aligning the pointer down)
@ -69,13 +69,13 @@ static mi_decl_noinline void* mi_heap_malloc_zero_aligned_at_fallback(mi_heap_t*
// todo: expand padding if overallocated ?
mi_assert_internal(mi_page_usable_block_size(_mi_ptr_page(p)) >= adjust + size);
mi_assert_internal(p == _mi_page_ptr_unalign(_mi_ptr_segment(aligned_p), _mi_ptr_page(aligned_p), aligned_p));
mi_assert_internal(p == _mi_page_ptr_unalign(_mi_ptr_page(aligned_p), aligned_p));
mi_assert_internal(((uintptr_t)aligned_p + offset) % alignment == 0);
mi_assert_internal(mi_usable_size(aligned_p)>=size);
mi_assert_internal(mi_usable_size(p) == mi_usable_size(aligned_p)+adjust);
// now zero the block if needed
if (alignment > MI_ALIGNMENT_MAX) {
if (alignment > MI_BLOCK_ALIGNMENT_MAX) {
// for the tracker, on huge aligned allocations only from the start of the large block is defined
mi_track_mem_undefined(aligned_p, size);
if (zero) {
@ -159,7 +159,7 @@ mi_decl_nodiscard mi_decl_restrict void* mi_heap_malloc_aligned(mi_heap_t* heap,
// ensure a definition is emitted
#if defined(__cplusplus)
static void* _mi_heap_malloc_aligned = (void*)&mi_heap_malloc_aligned;
void* _mi_extern_heap_malloc_aligned = (void*)&mi_heap_malloc_aligned;
#endif
// ------------------------------------------------------

22
src/alloc-override.c
View File

@ -23,7 +23,7 @@ mi_decl_externc size_t malloc_good_size(size_t size);
#endif
// helper definition for C override of C++ new
typedef struct mi_nothrow_s { int _tag; } mi_nothrow_t;
typedef void* mi_nothrow_t;
// ------------------------------------------------------
// Override system malloc
@ -77,7 +77,9 @@ typedef struct mi_nothrow_s { int _tag; } mi_nothrow_t;
MI_INTERPOSE_MI(calloc),
MI_INTERPOSE_MI(realloc),
MI_INTERPOSE_MI(strdup),
#if defined(MAC_OS_X_VERSION_10_7) && MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_7
MI_INTERPOSE_MI(strndup),
#endif
MI_INTERPOSE_MI(realpath),
MI_INTERPOSE_MI(posix_memalign),
MI_INTERPOSE_MI(reallocf),
@ -128,11 +130,16 @@ typedef struct mi_nothrow_s { int _tag; } mi_nothrow_t;
// cannot override malloc unless using a dll.
// we just override new/delete which does work in a static library.
#else
// On all other systems forward to our API
// On all other systems forward allocation primitives to our API
mi_decl_export void* malloc(size_t size) MI_FORWARD1(mi_malloc, size)
mi_decl_export void* calloc(size_t size, size_t n) MI_FORWARD2(mi_calloc, size, n)
mi_decl_export void* realloc(void* p, size_t newsize) MI_FORWARD2(mi_realloc, p, newsize)
mi_decl_export void free(void* p) MI_FORWARD0(mi_free, p)
// In principle we do not need to forward `strdup`/`strndup` but on some systems these do not use `malloc` internally (but a more primitive call)
mi_decl_export char* strdup(const char* str) MI_FORWARD1(mi_strdup, str)
#if !defined(__APPLE__) || (defined(MAC_OS_X_VERSION_10_7) && MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_7)
mi_decl_export char* strndup(const char* str, size_t n) MI_FORWARD2(mi_strndup, str, n)
#endif
#endif
#if (defined(__GNUC__) || defined(__clang__)) && !defined(__APPLE__)
@ -192,11 +199,17 @@ typedef struct mi_nothrow_s { int _tag; } mi_nothrow_t;
void _ZdaPv(void* p) MI_FORWARD0(mi_free,p) // delete[]
void _ZdlPvm(void* p, size_t n) MI_FORWARD02(mi_free_size,p,n)
void _ZdaPvm(void* p, size_t n) MI_FORWARD02(mi_free_size,p,n)
void _ZdlPvSt11align_val_t(void* p, size_t al) { mi_free_aligned(p,al); }
void _ZdaPvSt11align_val_t(void* p, size_t al) { mi_free_aligned(p,al); }
void _ZdlPvmSt11align_val_t(void* p, size_t n, size_t al) { mi_free_size_aligned(p,n,al); }
void _ZdaPvmSt11align_val_t(void* p, size_t n, size_t al) { mi_free_size_aligned(p,n,al); }
void _ZdlPvRKSt9nothrow_t(void* p, mi_nothrow_t tag) { MI_UNUSED(tag); mi_free(p); } // operator delete(void*, std::nothrow_t const&)
void _ZdaPvRKSt9nothrow_t(void* p, mi_nothrow_t tag) { MI_UNUSED(tag); mi_free(p); } // operator delete[](void*, std::nothrow_t const&)
void _ZdlPvSt11align_val_tRKSt9nothrow_t(void* p, size_t al, mi_nothrow_t tag) { MI_UNUSED(tag); mi_free_aligned(p,al); } // operator delete(void*, std::align_val_t, std::nothrow_t const&)
void _ZdaPvSt11align_val_tRKSt9nothrow_t(void* p, size_t al, mi_nothrow_t tag) { MI_UNUSED(tag); mi_free_aligned(p,al); } // operator delete[](void*, std::align_val_t, std::nothrow_t const&)
#if (MI_INTPTR_SIZE==8)
void* _Znwm(size_t n) MI_FORWARD1(mi_new,n) // new 64-bit
void* _Znam(size_t n) MI_FORWARD1(mi_new,n) // new[] 64-bit
@ -259,10 +272,11 @@ extern "C" {
// no forwarding here due to aliasing/name mangling issues
void cfree(void* p) { mi_free(p); }
void* pvalloc(size_t size) { return mi_pvalloc(size); }
void* reallocarray(void* p, size_t count, size_t size) { return mi_reallocarray(p, count, size); }
int reallocarr(void* p, size_t count, size_t size) { return mi_reallocarr(p, count, size); }
void* memalign(size_t alignment, size_t size) { return mi_memalign(alignment, size); }
void* _aligned_malloc(size_t alignment, size_t size) { return mi_aligned_alloc(alignment, size); }
void* reallocarray(void* p, size_t count, size_t size) { return mi_reallocarray(p, count, size); }
// some systems define reallocarr so mark it as a weak symbol (#751)
mi_decl_weak int reallocarr(void* p, size_t count, size_t size) { return mi_reallocarr(p, count, size); }
#if defined(__wasi__)
// forward __libc interface (see PR #667)

506
src/alloc.c
View File

@ -1,5 +1,5 @@
/* ----------------------------------------------------------------------------
Copyright (c) 2018-2022, Microsoft Research, Daan Leijen
Copyright (c) 2018-2024, Microsoft Research, Daan Leijen
This is free software; you can redistribute it and/or modify it under the
terms of the MIT license. A copy of the license can be found in the file
"LICENSE" at the root of this distribution.
@ -18,6 +18,7 @@ terms of the MIT license. A copy of the license can be found in the file
#define MI_IN_ALLOC_C
#include "alloc-override.c"
#include "free.c"
#undef MI_IN_ALLOC_C
// ------------------------------------------------------
@ -26,16 +27,18 @@ terms of the MIT license. A copy of the license can be found in the file
// Fast allocation in a page: just pop from the free list.
// Fall back to generic allocation only if the list is empty.
extern inline void* _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t size, bool zero) mi_attr_noexcept {
mi_assert_internal(page->xblock_size==0||mi_page_block_size(page) >= size);
// Note: in release mode the (inlined) routine is about 7 instructions with a single test.
extern inline void* _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t size, bool zero) mi_attr_noexcept
{
mi_assert_internal(page->block_size == 0 /* empty heap */ || mi_page_block_size(page) >= size);
mi_block_t* const block = page->free;
if mi_unlikely(block == NULL) {
return _mi_malloc_generic(heap, size, zero, 0);
}
mi_assert_internal(block != NULL && _mi_ptr_page(block) == page);
// pop from the free list
page->used++;
page->free = mi_block_next(page, block);
page->used++;
mi_assert_internal(page->free == NULL || _mi_ptr_page(page->free) == page);
#if MI_DEBUG>3
if (page->free_is_zero) {
@ -50,14 +53,14 @@ extern inline void* _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t siz
// zero the block? note: we need to zero the full block size (issue #63)
if mi_unlikely(zero) {
mi_assert_internal(page->xblock_size != 0); // do not call with zero'ing for huge blocks (see _mi_malloc_generic)
mi_assert_internal(page->xblock_size >= MI_PADDING_SIZE);
mi_assert_internal(page->block_size != 0); // do not call with zero'ing for huge blocks (see _mi_malloc_generic)
mi_assert_internal(page->block_size >= MI_PADDING_SIZE);
if (page->free_is_zero) {
block->next = 0;
mi_track_mem_defined(block, page->xblock_size - MI_PADDING_SIZE);
mi_track_mem_defined(block, page->block_size - MI_PADDING_SIZE);
}
else {
_mi_memzero_aligned(block, page->xblock_size - MI_PADDING_SIZE);
_mi_memzero_aligned(block, page->block_size - MI_PADDING_SIZE);
}
}
@ -112,9 +115,11 @@ static inline mi_decl_restrict void* mi_heap_malloc_small_zero(mi_heap_t* heap,
#if (MI_PADDING)
if (size == 0) { size = sizeof(void*); }
#endif
mi_page_t* page = _mi_heap_get_free_small_page(heap, size + MI_PADDING_SIZE);
void* const p = _mi_page_malloc(heap, page, size + MI_PADDING_SIZE, zero);
mi_track_malloc(p,size,zero);
#if MI_STAT>1
if (p != NULL) {
if (!mi_heap_is_initialized(heap)) { heap = mi_prim_get_default_heap(); }
@ -190,484 +195,6 @@ mi_decl_nodiscard mi_decl_restrict void* mi_zalloc(size_t size) mi_attr_noexcept
}
// ------------------------------------------------------
// Check for double free in secure and debug mode
// This is somewhat expensive so only enabled for secure mode 4
// ------------------------------------------------------
#if (MI_ENCODE_FREELIST && (MI_SECURE>=4 || MI_DEBUG!=0))
// linear check if the free list contains a specific element
static bool mi_list_contains(const mi_page_t* page, const mi_block_t* list, const mi_block_t* elem) {
while (list != NULL) {
if (elem==list) return true;
list = mi_block_next(page, list);
}
return false;
}
static mi_decl_noinline bool mi_check_is_double_freex(const mi_page_t* page, const mi_block_t* block) {
// The decoded value is in the same page (or NULL).
// Walk the free lists to verify positively if it is already freed
if (mi_list_contains(page, page->free, block) ||
mi_list_contains(page, page->local_free, block) ||
mi_list_contains(page, mi_page_thread_free(page), block))
{
_mi_error_message(EAGAIN, "double free detected of block %p with size %zu\n", block, mi_page_block_size(page));
return true;
}
return false;
}
#define mi_track_page(page,access) { size_t psize; void* pstart = _mi_page_start(_mi_page_segment(page),page,&psize); mi_track_mem_##access( pstart, psize); }
static inline bool mi_check_is_double_free(const mi_page_t* page, const mi_block_t* block) {
bool is_double_free = false;
mi_block_t* n = mi_block_nextx(page, block, page->keys); // pretend it is freed, and get the decoded first field
if (((uintptr_t)n & (MI_INTPTR_SIZE-1))==0 && // quick check: aligned pointer?
(n==NULL || mi_is_in_same_page(block, n))) // quick check: in same page or NULL?
{
// Suspicous: decoded value a in block is in the same page (or NULL) -- maybe a double free?
// (continue in separate function to improve code generation)
is_double_free = mi_check_is_double_freex(page, block);
}
return is_double_free;
}
#else
static inline bool mi_check_is_double_free(const mi_page_t* page, const mi_block_t* block) {
MI_UNUSED(page);
MI_UNUSED(block);
return false;
}
#endif
// ---------------------------------------------------------------------------
// Check for heap block overflow by setting up padding at the end of the block
// ---------------------------------------------------------------------------
#if MI_PADDING // && !MI_TRACK_ENABLED
static bool mi_page_decode_padding(const mi_page_t* page, const mi_block_t* block, size_t* delta, size_t* bsize) {
*bsize = mi_page_usable_block_size(page);
const mi_padding_t* const padding = (mi_padding_t*)((uint8_t*)block + *bsize);
mi_track_mem_defined(padding,sizeof(mi_padding_t));
*delta = padding->delta;
uint32_t canary = padding->canary;
uintptr_t keys[2];
keys[0] = page->keys[0];
keys[1] = page->keys[1];
bool ok = ((uint32_t)mi_ptr_encode(page,block,keys) == canary && *delta <= *bsize);
mi_track_mem_noaccess(padding,sizeof(mi_padding_t));
return ok;
}
// Return the exact usable size of a block.
static size_t mi_page_usable_size_of(const mi_page_t* page, const mi_block_t* block) {
size_t bsize;
size_t delta;
bool ok = mi_page_decode_padding(page, block, &delta, &bsize);
mi_assert_internal(ok); mi_assert_internal(delta <= bsize);
return (ok ? bsize - delta : 0);
}
// When a non-thread-local block is freed, it becomes part of the thread delayed free
// list that is freed later by the owning heap. If the exact usable size is too small to
// contain the pointer for the delayed list, then shrink the padding (by decreasing delta)
// so it will later not trigger an overflow error in `mi_free_block`.
void _mi_padding_shrink(const mi_page_t* page, const mi_block_t* block, const size_t min_size) {
size_t bsize;
size_t delta;
bool ok = mi_page_decode_padding(page, block, &delta, &bsize);
mi_assert_internal(ok);
if (!ok || (bsize - delta) >= min_size) return; // usually already enough space
mi_assert_internal(bsize >= min_size);
if (bsize < min_size) return; // should never happen
size_t new_delta = (bsize - min_size);
mi_assert_internal(new_delta < bsize);
mi_padding_t* padding = (mi_padding_t*)((uint8_t*)block + bsize);
mi_track_mem_defined(padding,sizeof(mi_padding_t));
padding->delta = (uint32_t)new_delta;
mi_track_mem_noaccess(padding,sizeof(mi_padding_t));
}
#else
static size_t mi_page_usable_size_of(const mi_page_t* page, const mi_block_t* block) {
MI_UNUSED(block);
return mi_page_usable_block_size(page);
}
void _mi_padding_shrink(const mi_page_t* page, const mi_block_t* block, const size_t min_size) {
MI_UNUSED(page);
MI_UNUSED(block);
MI_UNUSED(min_size);
}
#endif
#if MI_PADDING && MI_PADDING_CHECK
static bool mi_verify_padding(const mi_page_t* page, const mi_block_t* block, size_t* size, size_t* wrong) {
size_t bsize;
size_t delta;
bool ok = mi_page_decode_padding(page, block, &delta, &bsize);
*size = *wrong = bsize;
if (!ok) return false;
mi_assert_internal(bsize >= delta);
*size = bsize - delta;
if (!mi_page_is_huge(page)) {
uint8_t* fill = (uint8_t*)block + bsize - delta;
const size_t maxpad = (delta > MI_MAX_ALIGN_SIZE ? MI_MAX_ALIGN_SIZE : delta); // check at most the first N padding bytes
mi_track_mem_defined(fill, maxpad);
for (size_t i = 0; i < maxpad; i++) {
if (fill[i] != MI_DEBUG_PADDING) {
*wrong = bsize - delta + i;
ok = false;
break;
}
}
mi_track_mem_noaccess(fill, maxpad);
}
return ok;
}
static void mi_check_padding(const mi_page_t* page, const mi_block_t* block) {
size_t size;
size_t wrong;
if (!mi_verify_padding(page,block,&size,&wrong)) {
_mi_error_message(EFAULT, "buffer overflow in heap block %p of size %zu: write after %zu bytes\n", block, size, wrong );
}
}
#else
static void mi_check_padding(const mi_page_t* page, const mi_block_t* block) {
MI_UNUSED(page);
MI_UNUSED(block);
}
#endif
// only maintain stats for smaller objects if requested
#if (MI_STAT>0)
static void mi_stat_free(const mi_page_t* page, const mi_block_t* block) {
#if (MI_STAT < 2)
MI_UNUSED(block);
#endif
mi_heap_t* const heap = mi_heap_get_default();
const size_t bsize = mi_page_usable_block_size(page);
#if (MI_STAT>1)
const size_t usize = mi_page_usable_size_of(page, block);
mi_heap_stat_decrease(heap, malloc, usize);
#endif
if (bsize <= MI_MEDIUM_OBJ_SIZE_MAX) {
mi_heap_stat_decrease(heap, normal, bsize);
#if (MI_STAT > 1)
mi_heap_stat_decrease(heap, normal_bins[_mi_bin(bsize)], 1);
#endif
}
else if (bsize <= MI_LARGE_OBJ_SIZE_MAX) {
mi_heap_stat_decrease(heap, large, bsize);
}
else {
mi_heap_stat_decrease(heap, huge, bsize);
}
}
#else
static void mi_stat_free(const mi_page_t* page, const mi_block_t* block) {
MI_UNUSED(page); MI_UNUSED(block);
}
#endif
#if MI_HUGE_PAGE_ABANDON
#if (MI_STAT>0)
// maintain stats for huge objects
static void mi_stat_huge_free(const mi_page_t* page) {
mi_heap_t* const heap = mi_heap_get_default();
const size_t bsize = mi_page_block_size(page); // to match stats in `page.c:mi_page_huge_alloc`
if (bsize <= MI_LARGE_OBJ_SIZE_MAX) {
mi_heap_stat_decrease(heap, large, bsize);
}
else {
mi_heap_stat_decrease(heap, huge, bsize);
}
}
#else
static void mi_stat_huge_free(const mi_page_t* page) {
MI_UNUSED(page);
}
#endif
#endif
// ------------------------------------------------------
// Free
// ------------------------------------------------------
// multi-threaded free (or free in huge block if compiled with MI_HUGE_PAGE_ABANDON)
static mi_decl_noinline void _mi_free_block_mt(mi_page_t* page, mi_block_t* block)
{
// The padding check may access the non-thread-owned page for the key values.
// that is safe as these are constant and the page won't be freed (as the block is not freed yet).
mi_check_padding(page, block);
_mi_padding_shrink(page, block, sizeof(mi_block_t)); // for small size, ensure we can fit the delayed thread pointers without triggering overflow detection
// huge page segments are always abandoned and can be freed immediately
mi_segment_t* segment = _mi_page_segment(page);
if (segment->kind == MI_SEGMENT_HUGE) {
#if MI_HUGE_PAGE_ABANDON
// huge page segments are always abandoned and can be freed immediately
mi_stat_huge_free(page);
_mi_segment_huge_page_free(segment, page, block);
return;
#else
// huge pages are special as they occupy the entire segment
// as these are large we reset the memory occupied by the page so it is available to other threads
// (as the owning thread needs to actually free the memory later).
_mi_segment_huge_page_reset(segment, page, block);
#endif
}
#if (MI_DEBUG>0) && !MI_TRACK_ENABLED && !MI_TSAN // note: when tracking, cannot use mi_usable_size with multi-threading
if (segment->kind != MI_SEGMENT_HUGE) { // not for huge segments as we just reset the content
memset(block, MI_DEBUG_FREED, mi_usable_size(block));
}
#endif
// Try to put the block on either the page-local thread free list, or the heap delayed free list.
mi_thread_free_t tfreex;
bool use_delayed;
mi_thread_free_t tfree = mi_atomic_load_relaxed(&page->xthread_free);
do {
use_delayed = (mi_tf_delayed(tfree) == MI_USE_DELAYED_FREE);
if mi_unlikely(use_delayed) {
// unlikely: this only happens on the first concurrent free in a page that is in the full list
tfreex = mi_tf_set_delayed(tfree,MI_DELAYED_FREEING);
}
else {
// usual: directly add to page thread_free list
mi_block_set_next(page, block, mi_tf_block(tfree));
tfreex = mi_tf_set_block(tfree,block);
}
} while (!mi_atomic_cas_weak_release(&page->xthread_free, &tfree, tfreex));
if mi_unlikely(use_delayed) {
// racy read on `heap`, but ok because MI_DELAYED_FREEING is set (see `mi_heap_delete` and `mi_heap_collect_abandon`)
mi_heap_t* const heap = (mi_heap_t*)(mi_atomic_load_acquire(&page->xheap)); //mi_page_heap(page);
mi_assert_internal(heap != NULL);
if (heap != NULL) {
// add to the delayed free list of this heap. (do this atomically as the lock only protects heap memory validity)
mi_block_t* dfree = mi_atomic_load_ptr_relaxed(mi_block_t, &heap->thread_delayed_free);
do {
mi_block_set_nextx(heap,block,dfree, heap->keys);
} while (!mi_atomic_cas_ptr_weak_release(mi_block_t,&heap->thread_delayed_free, &dfree, block));
}
// and reset the MI_DELAYED_FREEING flag
tfree = mi_atomic_load_relaxed(&page->xthread_free);
do {
tfreex = tfree;
mi_assert_internal(mi_tf_delayed(tfree) == MI_DELAYED_FREEING);
tfreex = mi_tf_set_delayed(tfree,MI_NO_DELAYED_FREE);
} while (!mi_atomic_cas_weak_release(&page->xthread_free, &tfree, tfreex));
}
}
// regular free
static inline void _mi_free_block(mi_page_t* page, bool local, mi_block_t* block)
{
// and push it on the free list
//const size_t bsize = mi_page_block_size(page);
if mi_likely(local) {
// owning thread can free a block directly
if mi_unlikely(mi_check_is_double_free(page, block)) return;
mi_check_padding(page, block);
#if (MI_DEBUG>0) && !MI_TRACK_ENABLED && !MI_TSAN
if (!mi_page_is_huge(page)) { // huge page content may be already decommitted
memset(block, MI_DEBUG_FREED, mi_page_block_size(page));
}
#endif
mi_block_set_next(page, block, page->local_free);
page->local_free = block;
page->used--;
if mi_unlikely(mi_page_all_free(page)) {
_mi_page_retire(page);
}
else if mi_unlikely(mi_page_is_in_full(page)) {
_mi_page_unfull(page);
}
}
else {
_mi_free_block_mt(page,block);
}
}
// Adjust a block that was allocated aligned, to the actual start of the block in the page.
mi_block_t* _mi_page_ptr_unalign(const mi_segment_t* segment, const mi_page_t* page, const void* p) {
mi_assert_internal(page!=NULL && p!=NULL);
const size_t diff = (uint8_t*)p - _mi_page_start(segment, page, NULL);
const size_t adjust = (diff % mi_page_block_size(page));
return (mi_block_t*)((uintptr_t)p - adjust);
}
void mi_decl_noinline _mi_free_generic(const mi_segment_t* segment, mi_page_t* page, bool is_local, void* p) mi_attr_noexcept {
mi_block_t* const block = (mi_page_has_aligned(page) ? _mi_page_ptr_unalign(segment, page, p) : (mi_block_t*)p);
mi_stat_free(page, block); // stat_free may access the padding
mi_track_free_size(block, mi_page_usable_size_of(page,block));
_mi_free_block(page, is_local, block);
}
// Get the segment data belonging to a pointer
// This is just a single `and` in assembly but does further checks in debug mode
// (and secure mode) if this was a valid pointer.
static inline mi_segment_t* mi_checked_ptr_segment(const void* p, const char* msg)
{
MI_UNUSED(msg);
mi_assert(p != NULL);
#if (MI_DEBUG>0)
if mi_unlikely(((uintptr_t)p & (MI_INTPTR_SIZE - 1)) != 0) {
_mi_error_message(EINVAL, "%s: invalid (unaligned) pointer: %p\n", msg, p);
return NULL;
}
#endif
mi_segment_t* const segment = _mi_ptr_segment(p);
mi_assert_internal(segment != NULL);
#if (MI_DEBUG>0)
if mi_unlikely(!mi_is_in_heap_region(p)) {
#if (MI_INTPTR_SIZE == 8 && defined(__linux__))
if (((uintptr_t)p >> 40) != 0x7F) { // linux tends to align large blocks above 0x7F000000000 (issue #640)
#else
{
#endif
_mi_warning_message("%s: pointer might not point to a valid heap region: %p\n"
"(this may still be a valid very large allocation (over 64MiB))\n", msg, p);
if mi_likely(_mi_ptr_cookie(segment) == segment->cookie) {
_mi_warning_message("(yes, the previous pointer %p was valid after all)\n", p);
}
}
}
#endif
#if (MI_DEBUG>0 || MI_SECURE>=4)
if mi_unlikely(_mi_ptr_cookie(segment) != segment->cookie) {
_mi_error_message(EINVAL, "%s: pointer does not point to a valid heap space: %p\n", msg, p);
return NULL;
}
#endif
return segment;
}
// Free a block
// fast path written carefully to prevent spilling on the stack
void mi_free(void* p) mi_attr_noexcept
{
if mi_unlikely(p == NULL) return;
mi_segment_t* const segment = mi_checked_ptr_segment(p,"mi_free");
const bool is_local= (_mi_prim_thread_id() == mi_atomic_load_relaxed(&segment->thread_id));
mi_page_t* const page = _mi_segment_page_of(segment, p);
if mi_likely(is_local) { // thread-local free?
if mi_likely(page->flags.full_aligned == 0) // and it is not a full page (full pages need to move from the full bin), nor has aligned blocks (aligned blocks need to be unaligned)
{
mi_block_t* const block = (mi_block_t*)p;
if mi_unlikely(mi_check_is_double_free(page, block)) return;
mi_check_padding(page, block);
mi_stat_free(page, block);
#if (MI_DEBUG>0) && !MI_TRACK_ENABLED && !MI_TSAN
memset(block, MI_DEBUG_FREED, mi_page_block_size(page));
#endif
mi_track_free_size(p, mi_page_usable_size_of(page,block)); // faster then mi_usable_size as we already know the page and that p is unaligned
mi_block_set_next(page, block, page->local_free);
page->local_free = block;
if mi_unlikely(--page->used == 0) { // using this expression generates better code than: page->used--; if (mi_page_all_free(page))
_mi_page_retire(page);
}
}
else {
// page is full or contains (inner) aligned blocks; use generic path
_mi_free_generic(segment, page, true, p);
}
}
else {
// not thread-local; use generic path
_mi_free_generic(segment, page, false, p);
}
}
// return true if successful
bool _mi_free_delayed_block(mi_block_t* block) {
// get segment and page
const mi_segment_t* const segment = _mi_ptr_segment(block);
mi_assert_internal(_mi_ptr_cookie(segment) == segment->cookie);
mi_assert_internal(_mi_thread_id() == segment->thread_id);
mi_page_t* const page = _mi_segment_page_of(segment, block);
// Clear the no-delayed flag so delayed freeing is used again for this page.
// This must be done before collecting the free lists on this page -- otherwise
// some blocks may end up in the page `thread_free` list with no blocks in the
// heap `thread_delayed_free` list which may cause the page to be never freed!
// (it would only be freed if we happen to scan it in `mi_page_queue_find_free_ex`)
if (!_mi_page_try_use_delayed_free(page, MI_USE_DELAYED_FREE, false /* dont overwrite never delayed */)) {
return false;
}
// collect all other non-local frees to ensure up-to-date `used` count
_mi_page_free_collect(page, false);
// and free the block (possibly freeing the page as well since used is updated)
_mi_free_block(page, true, block);
return true;
}
// Bytes available in a block
mi_decl_noinline static size_t mi_page_usable_aligned_size_of(const mi_segment_t* segment, const mi_page_t* page, const void* p) mi_attr_noexcept {
const mi_block_t* block = _mi_page_ptr_unalign(segment, page, p);
const size_t size = mi_page_usable_size_of(page, block);
const ptrdiff_t adjust = (uint8_t*)p - (uint8_t*)block;
mi_assert_internal(adjust >= 0 && (size_t)adjust <= size);
return (size - adjust);
}
static inline size_t _mi_usable_size(const void* p, const char* msg) mi_attr_noexcept {
if (p == NULL) return 0;
const mi_segment_t* const segment = mi_checked_ptr_segment(p, msg);
const mi_page_t* const page = _mi_segment_page_of(segment, p);
if mi_likely(!mi_page_has_aligned(page)) {
const mi_block_t* block = (const mi_block_t*)p;
return mi_page_usable_size_of(page, block);
}
else {
// split out to separate routine for improved code generation
return mi_page_usable_aligned_size_of(segment, page, p);
}
}
mi_decl_nodiscard size_t mi_usable_size(const void* p) mi_attr_noexcept {
return _mi_usable_size(p, "mi_usable_size");
}
// ------------------------------------------------------
// Allocation extensions
// ------------------------------------------------------
void mi_free_size(void* p, size_t size) mi_attr_noexcept {
MI_UNUSED_RELEASE(size);
mi_assert(p == NULL || size <= _mi_usable_size(p,"mi_free_size"));
mi_free(p);
}
void mi_free_size_aligned(void* p, size_t size, size_t alignment) mi_attr_noexcept {
MI_UNUSED_RELEASE(alignment);
mi_assert(((uintptr_t)p % alignment) == 0);
mi_free_size(p,size);
}
void mi_free_aligned(void* p, size_t alignment) mi_attr_noexcept {
MI_UNUSED_RELEASE(alignment);
mi_assert(((uintptr_t)p % alignment) == 0);
mi_free(p);
}
mi_decl_nodiscard extern inline mi_decl_restrict void* mi_heap_calloc(mi_heap_t* heap, size_t count, size_t size) mi_attr_noexcept {
size_t total;
if (mi_count_size_overflow(count,size,&total)) return NULL;
@ -869,7 +396,8 @@ char* mi_heap_realpath(mi_heap_t* heap, const char* fname, char* resolved_name)
char* rname = realpath(fname, NULL);
if (rname == NULL) return NULL;
char* result = mi_heap_strdup(heap, rname);
free(rname); // use regular free! (which may be redirected to our free but that's ok)
mi_cfree(rname); // use checked free (which may be redirected to our free but that's ok)
// note: with ASAN realpath is intercepted and mi_cfree may leak the returned pointer :-(
return result;
}
/*
@ -913,9 +441,13 @@ static bool mi_try_new_handler(bool nothrow) {
#endif
if (h==NULL) {
_mi_error_message(ENOMEM, "out of memory in 'new'");
#if defined(_CPPUNWIND) || defined(__cpp_exceptions) // exceptions are not always enabled
if (!nothrow) {
throw std::bad_alloc();
}
#else
MI_UNUSED(nothrow);
#endif
return false;
}
else {

238
src/arena.c
View File

@ -55,6 +55,7 @@ typedef struct mi_arena_s {
mi_bitmap_field_t* blocks_dirty; // are the blocks potentially non-zero?
mi_bitmap_field_t* blocks_committed; // are the blocks committed? (can be NULL for memory that cannot be decommitted)
mi_bitmap_field_t* blocks_purge; // blocks that can be (reset) decommitted. (can be NULL for memory that cannot be (reset) decommitted)
mi_bitmap_field_t* blocks_abandoned; // blocks that start with an abandoned segment. (This crosses API's but it is convenient to have here)
mi_bitmap_field_t blocks_inuse[1]; // in-place bitmap of in-use blocks (of size `field_count`)
} mi_arena_t;
@ -94,7 +95,7 @@ bool _mi_arena_memid_is_suitable(mi_memid_t memid, mi_arena_id_t request_arena_i
return mi_arena_id_is_suitable(memid.mem.arena.id, memid.mem.arena.is_exclusive, request_arena_id);
}
else {
return mi_arena_id_is_suitable(0, false, request_arena_id);
return mi_arena_id_is_suitable(_mi_arena_id_none(), false, request_arena_id);
}
}
@ -165,6 +166,7 @@ static void* mi_arena_static_zalloc(size_t size, size_t alignment, mi_memid_t* m
// success
*memid = _mi_memid_create(MI_MEM_STATIC);
memid->initially_zero = true;
const size_t start = _mi_align_up(oldtop, alignment);
uint8_t* const p = &mi_arena_static[start];
_mi_memzero(p, size);
@ -175,11 +177,19 @@ static void* mi_arena_meta_zalloc(size_t size, mi_memid_t* memid, mi_stats_t* st
*memid = _mi_memid_none();
// try static
void* p = mi_arena_static_zalloc(size, MI_ALIGNMENT_MAX, memid);
void* p = mi_arena_static_zalloc(size, MI_MAX_ALIGN_SIZE, memid);
if (p != NULL) return p;
// or fall back to the OS
return _mi_os_alloc(size, memid, stats);
p = _mi_os_alloc(size, memid, stats);
if (p == NULL) return NULL;
// zero the OS memory if needed
if (!memid->initially_zero) {
_mi_memzero_aligned(p, size);
memid->initially_zero = true;
}
return p;
}
static void mi_arena_meta_free(void* p, mi_memid_t memid, size_t size, mi_stats_t* stats) {
@ -201,10 +211,10 @@ static void* mi_arena_block_start(mi_arena_t* arena, mi_bitmap_index_t bindex) {
----------------------------------------------------------- */
// claim the `blocks_inuse` bits
static bool mi_arena_try_claim(mi_arena_t* arena, size_t blocks, mi_bitmap_index_t* bitmap_idx)
static bool mi_arena_try_claim(mi_arena_t* arena, size_t blocks, mi_bitmap_index_t* bitmap_idx, mi_stats_t* stats)
{
size_t idx = 0; // mi_atomic_load_relaxed(&arena->search_idx); // start from last search; ok to be relaxed as the exact start does not matter
if (_mi_bitmap_try_find_from_claim_across(arena->blocks_inuse, arena->field_count, idx, blocks, bitmap_idx)) {
if (_mi_bitmap_try_find_from_claim_across(arena->blocks_inuse, arena->field_count, idx, blocks, bitmap_idx, stats)) {
mi_atomic_store_relaxed(&arena->search_idx, mi_bitmap_index_field(*bitmap_idx)); // start search from found location next time around
return true;
};
@ -223,7 +233,7 @@ static mi_decl_noinline void* mi_arena_try_alloc_at(mi_arena_t* arena, size_t ar
mi_assert_internal(mi_arena_id_index(arena->id) == arena_index);
mi_bitmap_index_t bitmap_index;
if (!mi_arena_try_claim(arena, needed_bcount, &bitmap_index)) return NULL;
if (!mi_arena_try_claim(arena, needed_bcount, &bitmap_index, tld->stats)) return NULL;
// claimed it!
void* p = mi_arena_block_start(arena, bitmap_index);
@ -346,7 +356,7 @@ static bool mi_arena_reserve(size_t req_size, bool allow_large, mi_arena_id_t re
if (arena_reserve == 0) return false;
if (!_mi_os_has_virtual_reserve()) {
arena_reserve = arena_reserve/4; // be conservative if virtual reserve is not supported (for some embedded systems for example)
arena_reserve = arena_reserve/4; // be conservative if virtual reserve is not supported (for WASM for example)
}
arena_reserve = _mi_align_up(arena_reserve, MI_ARENA_BLOCK_SIZE);
if (arena_count >= 8 && arena_count <= 128) {
@ -359,7 +369,7 @@ static bool mi_arena_reserve(size_t req_size, bool allow_large, mi_arena_id_t re
if (mi_option_get(mi_option_arena_eager_commit) == 2) { arena_commit = _mi_os_has_overcommit(); }
else if (mi_option_get(mi_option_arena_eager_commit) == 1) { arena_commit = true; }
return (mi_reserve_os_memory_ex(arena_reserve, arena_commit, allow_large, false /* exclusive */, arena_id) == 0);
return (mi_reserve_os_memory_ex(arena_reserve, arena_commit, allow_large, false /* exclusive? */, arena_id) == 0);
}
@ -373,6 +383,7 @@ void* _mi_arena_alloc_aligned(size_t size, size_t alignment, size_t align_offset
const int numa_node = _mi_os_numa_node(tld); // current numa node
// try to allocate in an arena if the alignment is small enough and the object is not too small (as for heap meta data)
if (!mi_option_is_enabled(mi_option_disallow_arena_alloc) || req_arena_id != _mi_arena_id_none()) { // is arena allocation allowed?
if (size >= MI_ARENA_MIN_OBJ_SIZE && alignment <= MI_SEGMENT_ALIGN && align_offset == 0) {
void* p = mi_arena_try_alloc(numa_node, size, alignment, commit, allow_large, req_arena_id, memid, tld);
if (p != NULL) return p;
@ -388,9 +399,10 @@ void* _mi_arena_alloc_aligned(size_t size, size_t alignment, size_t align_offset
}
}
}
}
// if we cannot use OS allocation, return NULL
if (mi_option_is_enabled(mi_option_limit_os_alloc) || req_arena_id != _mi_arena_id_none()) {
if (mi_option_is_enabled(mi_option_disallow_os_alloc) || req_arena_id != _mi_arena_id_none()) {
errno = ENOMEM;
return NULL;
}
@ -450,7 +462,7 @@ static void mi_arena_purge(mi_arena_t* arena, size_t bitmap_idx, size_t blocks,
// and also undo the decommit stats (as it was already adjusted)
mi_assert_internal(mi_option_is_enabled(mi_option_purge_decommits));
needs_recommit = _mi_os_purge_ex(p, size, false /* allow reset? */, stats);
_mi_stat_increase(&stats->committed, size);
if (needs_recommit) { _mi_stat_increase(&_mi_stats_main.committed, size); }
}
// clear the purged blocks
@ -476,7 +488,7 @@ static void mi_arena_schedule_purge(mi_arena_t* arena, size_t bitmap_idx, size_t
// schedule decommit
mi_msecs_t expire = mi_atomic_loadi64_relaxed(&arena->purge_expire);
if (expire != 0) {
mi_atomic_addi64_acq_rel(&arena->purge_expire, delay/10); // add smallish extra delay
mi_atomic_addi64_acq_rel(&arena->purge_expire, (mi_msecs_t)(delay/10)); // add smallish extra delay
}
else {
mi_atomic_storei64_release(&arena->purge_expire, _mi_clock_now() + delay);
@ -520,7 +532,7 @@ static bool mi_arena_try_purge(mi_arena_t* arena, mi_msecs_t now, bool force, mi
if (!force && expire > now) return false;
// reset expire (if not already set concurrently)
mi_atomic_casi64_strong_acq_rel(&arena->purge_expire, &expire, 0);
mi_atomic_casi64_strong_acq_rel(&arena->purge_expire, &expire, (mi_msecs_t)0);
// potential purges scheduled, walk through the bitmap
bool any_purged = false;
@ -607,7 +619,7 @@ void _mi_arena_free(void* p, size_t size, size_t committed_size, mi_memid_t memi
// was a direct OS allocation, pass through
if (!all_committed && committed_size > 0) {
// if partially committed, adjust the committed stats (as `_mi_os_free` will increase decommit by the full size)
_mi_stat_decrease(&stats->committed, committed_size);
_mi_stat_decrease(&_mi_stats_main.committed, committed_size);
}
_mi_os_free(p, size, memid, stats);
}
@ -623,12 +635,12 @@ void _mi_arena_free(void* p, size_t size, size_t committed_size, mi_memid_t memi
// checks
if (arena == NULL) {
_mi_error_message(EINVAL, "trying to free from non-existent arena: %p, size %zu, memid: 0x%zx\n", p, size, memid);
_mi_error_message(EINVAL, "trying to free from an invalid arena: %p, size %zu, memid: 0x%zx\n", p, size, memid);
return;
}
mi_assert_internal(arena->field_count > mi_bitmap_index_field(bitmap_idx));
if (arena->field_count <= mi_bitmap_index_field(bitmap_idx)) {
_mi_error_message(EINVAL, "trying to free from non-existent arena block: %p, size %zu, memid: 0x%zx\n", p, size, memid);
_mi_error_message(EINVAL, "trying to free from an invalid arena block: %p, size %zu, memid: 0x%zx\n", p, size, memid);
return;
}
@ -650,7 +662,7 @@ void _mi_arena_free(void* p, size_t size, size_t committed_size, mi_memid_t memi
if (committed_size > 0) {
// if partially committed, adjust the committed stats (is it will be recommitted when re-using)
// in the delayed purge, we now need to not count a decommit if the range is not marked as committed.
_mi_stat_decrease(&stats->committed, committed_size);
_mi_stat_decrease(&_mi_stats_main.committed, committed_size);
}
// note: if not all committed, it may be that the purge will reset/decommit the entire range
// that contains already decommitted parts. Since purge consistently uses reset or decommit that
@ -701,15 +713,15 @@ static void mi_arenas_unsafe_destroy(void) {
}
// Purge the arenas; if `force_purge` is true, amenable parts are purged even if not yet expired
void _mi_arena_collect(bool force_purge, mi_stats_t* stats) {
mi_arenas_try_purge(force_purge, true /* visit all */, stats);
void _mi_arenas_collect(bool force_purge, mi_stats_t* stats) {
mi_arenas_try_purge(force_purge, force_purge /* visit all? */, stats);
}
// destroy owned arenas; this is unsafe and should only be done using `mi_option_destroy_on_exit`
// for dynamic libraries that are unloaded and need to release all their allocated memory.
void _mi_arena_unsafe_destroy_all(mi_stats_t* stats) {
mi_arenas_unsafe_destroy();
_mi_arena_collect(true /* force purge */, stats); // purge non-owned arenas
_mi_arenas_collect(true /* force purge */, stats); // purge non-owned arenas
}
// Is a pointer inside any of our arenas?
@ -724,12 +736,144 @@ bool _mi_arena_contains(const void* p) {
return false;
}
/* -----------------------------------------------------------
Abandoned blocks/segments.
This is used to atomically abandon/reclaim segments
(and crosses the arena API but it is convenient to have here).
Abandoned segments still have live blocks; they get reclaimed
when a thread frees a block in it, or when a thread needs a fresh
segment; these threads scan the abandoned segments through
the arena bitmaps.
----------------------------------------------------------- */
// Maintain a count of all abandoned segments
static mi_decl_cache_align _Atomic(size_t)abandoned_count;
size_t _mi_arena_segment_abandoned_count(void) {
return mi_atomic_load_relaxed(&abandoned_count);
}
// reclaim a specific abandoned segment; `true` on success.
// sets the thread_id.
bool _mi_arena_segment_clear_abandoned(mi_segment_t* segment )
{
if (segment->memid.memkind != MI_MEM_ARENA) {
// not in an arena, consider it un-abandoned now.
// but we need to still claim it atomically -- we use the thread_id for that.
size_t expected = 0;
if (mi_atomic_cas_strong_acq_rel(&segment->thread_id, &expected, _mi_thread_id())) {
mi_atomic_decrement_relaxed(&abandoned_count);
return true;
}
else {
return false;
}
}
// arena segment: use the blocks_abandoned bitmap.
size_t arena_idx;
size_t bitmap_idx;
mi_arena_memid_indices(segment->memid, &arena_idx, &bitmap_idx);
mi_assert_internal(arena_idx < MI_MAX_ARENAS);
mi_arena_t* arena = mi_atomic_load_ptr_acquire(mi_arena_t, &mi_arenas[arena_idx]);
mi_assert_internal(arena != NULL);
bool was_marked = _mi_bitmap_unclaim(arena->blocks_abandoned, arena->field_count, 1, bitmap_idx);
if (was_marked) {
mi_assert_internal(mi_atomic_load_relaxed(&segment->thread_id) == 0);
mi_atomic_decrement_relaxed(&abandoned_count);
mi_atomic_store_release(&segment->thread_id, _mi_thread_id());
}
// mi_assert_internal(was_marked);
mi_assert_internal(!was_marked || _mi_bitmap_is_claimed(arena->blocks_inuse, arena->field_count, 1, bitmap_idx));
//mi_assert_internal(arena->blocks_committed == NULL || _mi_bitmap_is_claimed(arena->blocks_committed, arena->field_count, 1, bitmap_idx));
return was_marked;
}
// mark a specific segment as abandoned
// clears the thread_id.
void _mi_arena_segment_mark_abandoned(mi_segment_t* segment)
{
mi_atomic_store_release(&segment->thread_id, 0);
mi_assert_internal(segment->used == segment->abandoned);
if (segment->memid.memkind != MI_MEM_ARENA) {
// not in an arena; count it as abandoned and return
mi_atomic_increment_relaxed(&abandoned_count);
return;
}
size_t arena_idx;
size_t bitmap_idx;
mi_arena_memid_indices(segment->memid, &arena_idx, &bitmap_idx);
mi_assert_internal(arena_idx < MI_MAX_ARENAS);
mi_arena_t* arena = mi_atomic_load_ptr_acquire(mi_arena_t, &mi_arenas[arena_idx]);
mi_assert_internal(arena != NULL);
const bool was_unmarked = _mi_bitmap_claim(arena->blocks_abandoned, arena->field_count, 1, bitmap_idx, NULL);
if (was_unmarked) { mi_atomic_increment_relaxed(&abandoned_count); }
mi_assert_internal(was_unmarked);
mi_assert_internal(_mi_bitmap_is_claimed(arena->blocks_inuse, arena->field_count, 1, bitmap_idx));
}
// start a cursor at a randomized arena
void _mi_arena_field_cursor_init(mi_heap_t* heap, mi_arena_field_cursor_t* current) {
const size_t max_arena = mi_atomic_load_relaxed(&mi_arena_count);
current->start = (max_arena == 0 ? 0 : (mi_arena_id_t)( _mi_heap_random_next(heap) % max_arena));
current->count = 0;
current->bitmap_idx = 0;
}
// reclaim abandoned segments
// this does not set the thread id (so it appears as still abandoned)
mi_segment_t* _mi_arena_segment_clear_abandoned_next(mi_arena_field_cursor_t* previous )
{
const int max_arena = (int)mi_atomic_load_relaxed(&mi_arena_count);
if (max_arena <= 0 || mi_atomic_load_relaxed(&abandoned_count) == 0) return NULL;
int count = previous->count;
size_t field_idx = mi_bitmap_index_field(previous->bitmap_idx);
size_t bit_idx = mi_bitmap_index_bit_in_field(previous->bitmap_idx) + 1;
// visit arena's (from previous)
for (; count < max_arena; count++, field_idx = 0, bit_idx = 0) {
mi_arena_id_t arena_idx = previous->start + count;
if (arena_idx >= max_arena) { arena_idx = arena_idx % max_arena; } // wrap around
mi_arena_t* arena = mi_atomic_load_ptr_acquire(mi_arena_t, &mi_arenas[arena_idx]);
if (arena != NULL) {
// visit the abandoned fields (starting at previous_idx)
for ( ; field_idx < arena->field_count; field_idx++, bit_idx = 0) {
size_t field = mi_atomic_load_relaxed(&arena->blocks_abandoned[field_idx]);
if mi_unlikely(field != 0) { // skip zero fields quickly
// visit each set bit in the field (todo: maybe use `ctz` here?)
for ( ; bit_idx < MI_BITMAP_FIELD_BITS; bit_idx++) {
// pre-check if the bit is set
size_t mask = ((size_t)1 << bit_idx);
if mi_unlikely((field & mask) == mask) {
mi_bitmap_index_t bitmap_idx = mi_bitmap_index_create(field_idx, bit_idx);
// try to reclaim it atomically
if (_mi_bitmap_unclaim(arena->blocks_abandoned, arena->field_count, 1, bitmap_idx)) {
mi_atomic_decrement_relaxed(&abandoned_count);
previous->bitmap_idx = bitmap_idx;
previous->count = count;
mi_assert_internal(_mi_bitmap_is_claimed(arena->blocks_inuse, arena->field_count, 1, bitmap_idx));
mi_segment_t* segment = (mi_segment_t*)mi_arena_block_start(arena, bitmap_idx);
mi_assert_internal(mi_atomic_load_relaxed(&segment->thread_id) == 0);
//mi_assert_internal(arena->blocks_committed == NULL || _mi_bitmap_is_claimed(arena->blocks_committed, arena->field_count, 1, bitmap_idx));
return segment;
}
}
}
}
}
}
}
// no more found
previous->bitmap_idx = 0;
previous->count = 0;
return NULL;
}
/* -----------------------------------------------------------
Add an arena.
----------------------------------------------------------- */
static bool mi_arena_add(mi_arena_t* arena, mi_arena_id_t* arena_id) {
static bool mi_arena_add(mi_arena_t* arena, mi_arena_id_t* arena_id, mi_stats_t* stats) {
mi_assert_internal(arena != NULL);
mi_assert_internal((uintptr_t)mi_atomic_load_ptr_relaxed(uint8_t,&arena->start) % MI_SEGMENT_ALIGN == 0);
mi_assert_internal(arena->block_count > 0);
@ -740,6 +884,7 @@ static bool mi_arena_add(mi_arena_t* arena, mi_arena_id_t* arena_id) {
mi_atomic_decrement_acq_rel(&mi_arena_count);
return false;
}
_mi_stat_counter_increase(&stats->arena_count,1);
arena->id = mi_arena_id_create(i);
mi_atomic_store_ptr_release(mi_arena_t,&mi_arenas[i], arena);
if (arena_id != NULL) { *arena_id = arena->id; }
@ -757,13 +902,13 @@ static bool mi_manage_os_memory_ex2(void* start, size_t size, bool is_large, int
const size_t bcount = size / MI_ARENA_BLOCK_SIZE;
const size_t fields = _mi_divide_up(bcount, MI_BITMAP_FIELD_BITS);
const size_t bitmaps = (memid.is_pinned ? 2 : 4);
const size_t bitmaps = (memid.is_pinned ? 3 : 5);
const size_t asize = sizeof(mi_arena_t) + (bitmaps*fields*sizeof(mi_bitmap_field_t));
mi_memid_t meta_memid;
mi_arena_t* arena = (mi_arena_t*)mi_arena_meta_zalloc(asize, &meta_memid, &_mi_stats_main); // TODO: can we avoid allocating from the OS?
if (arena == NULL) return false;
// already zero'd due to os_alloc
// already zero'd due to zalloc
// _mi_memzero(arena, asize);
arena->id = _mi_arena_id_none();
arena->memid = memid;
@ -777,9 +922,11 @@ static bool mi_manage_os_memory_ex2(void* start, size_t size, bool is_large, int
arena->is_large = is_large;
arena->purge_expire = 0;
arena->search_idx = 0;
// consequetive bitmaps
arena->blocks_dirty = &arena->blocks_inuse[fields]; // just after inuse bitmap
arena->blocks_committed = (arena->memid.is_pinned ? NULL : &arena->blocks_inuse[2*fields]); // just after dirty bitmap
arena->blocks_purge = (arena->memid.is_pinned ? NULL : &arena->blocks_inuse[3*fields]); // just after committed bitmap
arena->blocks_abandoned = &arena->blocks_inuse[2 * fields]; // just after dirty bitmap
arena->blocks_committed = (arena->memid.is_pinned ? NULL : &arena->blocks_inuse[3*fields]); // just after abandoned bitmap
arena->blocks_purge = (arena->memid.is_pinned ? NULL : &arena->blocks_inuse[4*fields]); // just after committed bitmap
// initialize committed bitmap?
if (arena->blocks_committed != NULL && arena->memid.initially_committed) {
memset((void*)arena->blocks_committed, 0xFF, fields*sizeof(mi_bitmap_field_t)); // cast to void* to avoid atomic warning
@ -793,7 +940,7 @@ static bool mi_manage_os_memory_ex2(void* start, size_t size, bool is_large, int
mi_bitmap_index_t postidx = mi_bitmap_index_create(fields - 1, MI_BITMAP_FIELD_BITS - post);
_mi_bitmap_claim(arena->blocks_inuse, fields, post, postidx, NULL);
}
return mi_arena_add(arena, arena_id);
return mi_arena_add(arena, arena_id, &_mi_stats_main);
}
@ -815,7 +962,7 @@ int mi_reserve_os_memory_ex(size_t size, bool commit, bool allow_large, bool exc
const bool is_large = memid.is_pinned; // todo: use separate is_large field?
if (!mi_manage_os_memory_ex2(start, size, is_large, -1 /* numa node */, exclusive, memid, arena_id)) {
_mi_os_free_ex(start, size, commit, memid, &_mi_stats_main);
_mi_verbose_message("failed to reserve %zu k memory\n", _mi_divide_up(size, 1024));
_mi_verbose_message("failed to reserve %zu KiB memory\n", _mi_divide_up(size, 1024));
return ENOMEM;
}
_mi_verbose_message("reserved %zu KiB memory%s\n", _mi_divide_up(size, 1024), is_large ? " (in large os pages)" : "");
@ -838,32 +985,55 @@ int mi_reserve_os_memory(size_t size, bool commit, bool allow_large) mi_attr_noe
Debugging
----------------------------------------------------------- */
static size_t mi_debug_show_bitmap(const char* prefix, mi_bitmap_field_t* fields, size_t field_count ) {
static size_t mi_debug_show_bitmap(const char* prefix, const char* header, size_t block_count, mi_bitmap_field_t* fields, size_t field_count ) {
_mi_verbose_message("%s%s:\n", prefix, header);
size_t bcount = 0;
size_t inuse_count = 0;
for (size_t i = 0; i < field_count; i++) {
char buf[MI_BITMAP_FIELD_BITS + 1];
uintptr_t field = mi_atomic_load_relaxed(&fields[i]);
for (size_t bit = 0; bit < MI_BITMAP_FIELD_BITS; bit++) {
for (size_t bit = 0; bit < MI_BITMAP_FIELD_BITS; bit++, bcount++) {
if (bcount < block_count) {
bool inuse = ((((uintptr_t)1 << bit) & field) != 0);
if (inuse) inuse_count++;
buf[MI_BITMAP_FIELD_BITS - 1 - bit] = (inuse ? 'x' : '.');
buf[bit] = (inuse ? 'x' : '.');
}
else {
buf[bit] = ' ';
}
}
buf[MI_BITMAP_FIELD_BITS] = 0;
_mi_verbose_message("%s %s\n", prefix, buf);
}
_mi_verbose_message("%s total ('x'): %zu\n", prefix, inuse_count);
return inuse_count;
}
void mi_debug_show_arenas(void) mi_attr_noexcept {
void mi_debug_show_arenas(bool show_inuse, bool show_abandoned, bool show_purge) mi_attr_noexcept {
size_t max_arenas = mi_atomic_load_relaxed(&mi_arena_count);
size_t inuse_total = 0;
size_t abandoned_total = 0;
size_t purge_total = 0;
for (size_t i = 0; i < max_arenas; i++) {
mi_arena_t* arena = mi_atomic_load_ptr_relaxed(mi_arena_t, &mi_arenas[i]);
if (arena == NULL) break;
size_t inuse_count = 0;
_mi_verbose_message("arena %zu: %zu blocks with %zu fields\n", i, arena->block_count, arena->field_count);
inuse_count += mi_debug_show_bitmap(" ", arena->blocks_inuse, arena->field_count);
_mi_verbose_message(" blocks in use ('x'): %zu\n", inuse_count);
_mi_verbose_message("arena %zu: %zu blocks of size %zuMiB (in %zu fields) %s\n", i, arena->block_count, MI_ARENA_BLOCK_SIZE / MI_MiB, arena->field_count, (arena->memid.is_pinned ? ", pinned" : ""));
if (show_inuse) {
inuse_total += mi_debug_show_bitmap(" ", "inuse blocks", arena->block_count, arena->blocks_inuse, arena->field_count);
}
if (arena->blocks_committed != NULL) {
mi_debug_show_bitmap(" ", "committed blocks", arena->block_count, arena->blocks_committed, arena->field_count);
}
if (show_abandoned) {
abandoned_total += mi_debug_show_bitmap(" ", "abandoned blocks", arena->block_count, arena->blocks_abandoned, arena->field_count);
}
if (show_purge && arena->blocks_purge != NULL) {
purge_total += mi_debug_show_bitmap(" ", "purgeable blocks", arena->block_count, arena->blocks_purge, arena->field_count);
}
}
if (show_inuse) _mi_verbose_message("total inuse blocks : %zu\n", inuse_total);
if (show_abandoned) _mi_verbose_message("total abandoned blocks: %zu\n", abandoned_total);
if (show_purge) _mi_verbose_message("total purgeable blocks: %zu\n", purge_total);
}

14
src/bitmap.c
View File

@ -7,7 +7,7 @@ terms of the MIT license. A copy of the license can be found in the file
/* ----------------------------------------------------------------------------
Concurrent bitmap that can set/reset sequences of bits atomically,
represeted as an array of fields where each field is a machine word (`size_t`)
represented as an array of fields where each field is a machine word (`size_t`)
There are two api's; the standard one cannot have sequences that cross
between the bitmap fields (and a sequence must be <= MI_BITMAP_FIELD_BITS).
@ -200,7 +200,7 @@ bool _mi_bitmap_is_any_claimed(mi_bitmap_t bitmap, size_t bitmap_fields, size_t
// Try to atomically claim a sequence of `count` bits starting from the field
// at `idx` in `bitmap` and crossing into subsequent fields. Returns `true` on success.
// Only needs to consider crossing into the next fields (see `mi_bitmap_try_find_from_claim_across`)
static bool mi_bitmap_try_find_claim_field_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t idx, const size_t count, const size_t retries, mi_bitmap_index_t* bitmap_idx)
static bool mi_bitmap_try_find_claim_field_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t idx, const size_t count, const size_t retries, mi_bitmap_index_t* bitmap_idx, mi_stats_t* stats)
{
mi_assert_internal(bitmap_idx != NULL);
@ -260,6 +260,7 @@ static bool mi_bitmap_try_find_claim_field_across(mi_bitmap_t bitmap, size_t bit
} while (!mi_atomic_cas_strong_acq_rel(field, &map, newmap));
// claimed!
mi_stat_counter_increase(stats->arena_crossover_count,1);
*bitmap_idx = mi_bitmap_index_create(idx, initial_idx);
return true;
@ -279,9 +280,10 @@ rollback:
newmap = (map & ~initial_mask);
} while (!mi_atomic_cas_strong_acq_rel(field, &map, newmap));
}
mi_stat_counter_increase(stats->arena_rollback_count,1);
// retry? (we make a recursive call instead of goto to be able to use const declarations)
if (retries <= 2) {
return mi_bitmap_try_find_claim_field_across(bitmap, bitmap_fields, idx, count, retries+1, bitmap_idx);
return mi_bitmap_try_find_claim_field_across(bitmap, bitmap_fields, idx, count, retries+1, bitmap_idx, stats);
}
else {
return false;
@ -291,7 +293,7 @@ rollback:
// Find `count` bits of zeros and set them to 1 atomically; returns `true` on success.
// Starts at idx, and wraps around to search in all `bitmap_fields` fields.
bool _mi_bitmap_try_find_from_claim_across(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx) {
bool _mi_bitmap_try_find_from_claim_across(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx, mi_stats_t* stats) {
mi_assert_internal(count > 0);
if (count <= 2) {
// we don't bother with crossover fields for small counts
@ -303,13 +305,15 @@ bool _mi_bitmap_try_find_from_claim_across(mi_bitmap_t bitmap, const size_t bitm
for (size_t visited = 0; visited < bitmap_fields; visited++, idx++) {
if (idx >= bitmap_fields) { idx = 0; } // wrap
// first try to claim inside a field
/*
if (count <= MI_BITMAP_FIELD_BITS) {
if (_mi_bitmap_try_find_claim_field(bitmap, idx, count, bitmap_idx)) {
return true;
}
}
*/
// if that fails, then try to claim across fields
if (mi_bitmap_try_find_claim_field_across(bitmap, bitmap_fields, idx, count, 0, bitmap_idx)) {
if (mi_bitmap_try_find_claim_field_across(bitmap, bitmap_fields, idx, count, 0, bitmap_idx, stats)) {
return true;
}
}

4
src/bitmap.h
View File

@ -7,7 +7,7 @@ terms of the MIT license. A copy of the license can be found in the file
/* ----------------------------------------------------------------------------
Concurrent bitmap that can set/reset sequences of bits atomically,
represeted as an array of fields where each field is a machine word (`size_t`)
represented as an array of fields where each field is a machine word (`size_t`)
There are two api's; the standard one cannot have sequences that cross
between the bitmap fields (and a sequence must be <= MI_BITMAP_FIELD_BITS).
@ -99,7 +99,7 @@ bool _mi_bitmap_is_any_claimed(mi_bitmap_t bitmap, size_t bitmap_fields, size_t
// Find `count` bits of zeros and set them to 1 atomically; returns `true` on success.
// Starts at idx, and wraps around to search in all `bitmap_fields` fields.
bool _mi_bitmap_try_find_from_claim_across(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx);
bool _mi_bitmap_try_find_from_claim_across(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx, mi_stats_t* stats);
// Set `count` bits at `bitmap_idx` to 0 atomically
// Returns `true` if all `count` bits were 1 previously.

530
src/free.c Normal file
View File

@ -0,0 +1,530 @@
/* ----------------------------------------------------------------------------
Copyright (c) 2018-2024, Microsoft Research, Daan Leijen
This is free software; you can redistribute it and/or modify it under the
terms of the MIT license. A copy of the license can be found in the file
"LICENSE" at the root of this distribution.
-----------------------------------------------------------------------------*/
#if !defined(MI_IN_ALLOC_C)
#error "this file should be included from 'alloc.c' (so aliases can work from alloc-override)"
// add includes help an IDE
#include "mimalloc.h"
#include "mimalloc/internal.h"
#include "mimalloc/atomic.h"
#include "mimalloc/prim.h" // _mi_prim_thread_id()
#endif
// forward declarations
static void mi_check_padding(const mi_page_t* page, const mi_block_t* block);
static bool mi_check_is_double_free(const mi_page_t* page, const mi_block_t* block);
static size_t mi_page_usable_size_of(const mi_page_t* page, const mi_block_t* block);
static void mi_stat_free(const mi_page_t* page, const mi_block_t* block);
// ------------------------------------------------------
// Free
// ------------------------------------------------------
// forward declaration of multi-threaded free (`_mt`) (or free in huge block if compiled with MI_HUGE_PAGE_ABANDON)
static mi_decl_noinline void mi_free_block_mt(mi_page_t* page, mi_segment_t* segment, mi_block_t* block);
// regular free of a (thread local) block pointer
// fast path written carefully to prevent spilling on the stack
static inline void mi_free_block_local(mi_page_t* page, mi_block_t* block, bool track_stats, bool check_full)
{
// checks
if mi_unlikely(mi_check_is_double_free(page, block)) return;
mi_check_padding(page, block);
if (track_stats) { mi_stat_free(page, block); }
#if (MI_DEBUG>0) && !MI_TRACK_ENABLED && !MI_TSAN
if (!mi_page_is_huge(page)) { // huge page content may be already decommitted
memset(block, MI_DEBUG_FREED, mi_page_block_size(page));
}
#endif
if (track_stats) { mi_track_free_size(block, mi_page_usable_size_of(page, block)); } // faster then mi_usable_size as we already know the page and that p is unaligned
// actual free: push on the local free list
mi_block_set_next(page, block, page->local_free);
page->local_free = block;
if mi_unlikely(--page->used == 0) {
_mi_page_retire(page);
}
else if mi_unlikely(check_full && mi_page_is_in_full(page)) {
_mi_page_unfull(page);
}
}
// Adjust a block that was allocated aligned, to the actual start of the block in the page.
// note: this can be called from `mi_free_generic_mt` where a non-owning thread accesses the
// `page_start` and `block_size` fields; however these are constant and the page won't be
// deallocated (as the block we are freeing keeps it alive) and thus safe to read concurrently.
mi_block_t* _mi_page_ptr_unalign(const mi_page_t* page, const void* p) {
mi_assert_internal(page!=NULL && p!=NULL);
size_t diff = (uint8_t*)p - page->page_start;
size_t adjust;
if mi_likely(page->block_size_shift != 0) {
adjust = diff & (((size_t)1 << page->block_size_shift) - 1);
}
else {
adjust = diff % mi_page_block_size(page);
}
return (mi_block_t*)((uintptr_t)p - adjust);
}
// free a local pointer (page parameter comes first for better codegen)
static void mi_decl_noinline mi_free_generic_local(mi_page_t* page, mi_segment_t* segment, void* p) mi_attr_noexcept {
MI_UNUSED(segment);
mi_block_t* const block = (mi_page_has_aligned(page) ? _mi_page_ptr_unalign(page, p) : (mi_block_t*)p);
mi_free_block_local(page, block, true /* track stats */, true /* check for a full page */);
}
// free a pointer owned by another thread (page parameter comes first for better codegen)
static void mi_decl_noinline mi_free_generic_mt(mi_page_t* page, mi_segment_t* segment, void* p) mi_attr_noexcept {
mi_block_t* const block = _mi_page_ptr_unalign(page, p); // don't check `has_aligned` flag to avoid a race (issue #865)
mi_free_block_mt(page, segment, block);
}
// generic free (for runtime integration)
void mi_decl_noinline _mi_free_generic(mi_segment_t* segment, mi_page_t* page, bool is_local, void* p) mi_attr_noexcept {
if (is_local) mi_free_generic_local(page,segment,p);
else mi_free_generic_mt(page,segment,p);
}
// Get the segment data belonging to a pointer
// This is just a single `and` in release mode but does further checks in debug mode
// (and secure mode) to see if this was a valid pointer.
static inline mi_segment_t* mi_checked_ptr_segment(const void* p, const char* msg)
{
MI_UNUSED(msg);
#if (MI_DEBUG>0)
if mi_unlikely(((uintptr_t)p & (MI_INTPTR_SIZE - 1)) != 0) {
_mi_error_message(EINVAL, "%s: invalid (unaligned) pointer: %p\n", msg, p);
return NULL;
}
#endif
mi_segment_t* const segment = _mi_ptr_segment(p);
if mi_unlikely(segment==NULL) return segment;
#if (MI_DEBUG>0)
if mi_unlikely(!mi_is_in_heap_region(p)) {
#if (MI_INTPTR_SIZE == 8 && defined(__linux__))
if (((uintptr_t)p >> 40) != 0x7F) { // linux tends to align large blocks above 0x7F000000000 (issue #640)
#else
{
#endif
_mi_warning_message("%s: pointer might not point to a valid heap region: %p\n"
"(this may still be a valid very large allocation (over 64MiB))\n", msg, p);
if mi_likely(_mi_ptr_cookie(segment) == segment->cookie) {
_mi_warning_message("(yes, the previous pointer %p was valid after all)\n", p);
}
}
}
#endif
#if (MI_DEBUG>0 || MI_SECURE>=4)
if mi_unlikely(_mi_ptr_cookie(segment) != segment->cookie) {
_mi_error_message(EINVAL, "%s: pointer does not point to a valid heap space: %p\n", msg, p);
return NULL;
}
#endif
return segment;
}
// Free a block
// Fast path written carefully to prevent register spilling on the stack
void mi_free(void* p) mi_attr_noexcept
{
mi_segment_t* const segment = mi_checked_ptr_segment(p,"mi_free");
if mi_unlikely(segment==NULL) return;
const bool is_local = (_mi_prim_thread_id() == mi_atomic_load_relaxed(&segment->thread_id));
mi_page_t* const page = _mi_segment_page_of(segment, p);
if mi_likely(is_local) { // thread-local free?
if mi_likely(page->flags.full_aligned == 0) { // and it is not a full page (full pages need to move from the full bin), nor has aligned blocks (aligned blocks need to be unaligned)
// thread-local, aligned, and not a full page
mi_block_t* const block = (mi_block_t*)p;
mi_free_block_local(page, block, true /* track stats */, false /* no need to check if the page is full */);
}
else {
// page is full or contains (inner) aligned blocks; use generic path
mi_free_generic_local(page, segment, p);
}
}
else {
// not thread-local; use generic path
mi_free_generic_mt(page, segment, p);
}
}
// return true if successful
bool _mi_free_delayed_block(mi_block_t* block) {
// get segment and page
mi_assert_internal(block!=NULL);
const mi_segment_t* const segment = _mi_ptr_segment(block);
mi_assert_internal(_mi_ptr_cookie(segment) == segment->cookie);
mi_assert_internal(_mi_thread_id() == segment->thread_id);
mi_page_t* const page = _mi_segment_page_of(segment, block);
// Clear the no-delayed flag so delayed freeing is used again for this page.
// This must be done before collecting the free lists on this page -- otherwise
// some blocks may end up in the page `thread_free` list with no blocks in the
// heap `thread_delayed_free` list which may cause the page to be never freed!
// (it would only be freed if we happen to scan it in `mi_page_queue_find_free_ex`)
if (!_mi_page_try_use_delayed_free(page, MI_USE_DELAYED_FREE, false /* dont overwrite never delayed */)) {
return false;
}
// collect all other non-local frees (move from `thread_free` to `free`) to ensure up-to-date `used` count
_mi_page_free_collect(page, false);
// and free the block (possibly freeing the page as well since `used` is updated)
mi_free_block_local(page, block, false /* stats have already been adjusted */, true /* check for a full page */);
return true;
}
// ------------------------------------------------------
// Multi-threaded Free (`_mt`)
// ------------------------------------------------------
// Push a block that is owned by another thread on its page-local thread free
// list or it's heap delayed free list. Such blocks are later collected by
// the owning thread in `_mi_free_delayed_block`.
static void mi_decl_noinline mi_free_block_delayed_mt( mi_page_t* page, mi_block_t* block )
{
// Try to put the block on either the page-local thread free list,
// or the heap delayed free list (if this is the first non-local free in that page)
mi_thread_free_t tfreex;
bool use_delayed;
mi_thread_free_t tfree = mi_atomic_load_relaxed(&page->xthread_free);
do {
use_delayed = (mi_tf_delayed(tfree) == MI_USE_DELAYED_FREE);
if mi_unlikely(use_delayed) {
// unlikely: this only happens on the first concurrent free in a page that is in the full list
tfreex = mi_tf_set_delayed(tfree,MI_DELAYED_FREEING);
}
else {
// usual: directly add to page thread_free list
mi_block_set_next(page, block, mi_tf_block(tfree));
tfreex = mi_tf_set_block(tfree,block);
}
} while (!mi_atomic_cas_weak_release(&page->xthread_free, &tfree, tfreex));
// If this was the first non-local free, we need to push it on the heap delayed free list instead
if mi_unlikely(use_delayed) {
// racy read on `heap`, but ok because MI_DELAYED_FREEING is set (see `mi_heap_delete` and `mi_heap_collect_abandon`)
mi_heap_t* const heap = (mi_heap_t*)(mi_atomic_load_acquire(&page->xheap)); //mi_page_heap(page);
mi_assert_internal(heap != NULL);
if (heap != NULL) {
// add to the delayed free list of this heap. (do this atomically as the lock only protects heap memory validity)
mi_block_t* dfree = mi_atomic_load_ptr_relaxed(mi_block_t, &heap->thread_delayed_free);
do {
mi_block_set_nextx(heap,block,dfree, heap->keys);
} while (!mi_atomic_cas_ptr_weak_release(mi_block_t,&heap->thread_delayed_free, &dfree, block));
}
// and reset the MI_DELAYED_FREEING flag
tfree = mi_atomic_load_relaxed(&page->xthread_free);
do {
tfreex = tfree;
mi_assert_internal(mi_tf_delayed(tfree) == MI_DELAYED_FREEING);
tfreex = mi_tf_set_delayed(tfree,MI_NO_DELAYED_FREE);
} while (!mi_atomic_cas_weak_release(&page->xthread_free, &tfree, tfreex));
}
}
// Multi-threaded free (`_mt`) (or free in huge block if compiled with MI_HUGE_PAGE_ABANDON)
static void mi_decl_noinline mi_free_block_mt(mi_page_t* page, mi_segment_t* segment, mi_block_t* block)
{
// first see if the segment was abandoned and if we can reclaim it into our thread
if (mi_option_is_enabled(mi_option_abandoned_reclaim_on_free) &&
#if MI_HUGE_PAGE_ABANDON
segment->page_kind != MI_PAGE_HUGE &&
#endif
mi_atomic_load_relaxed(&segment->thread_id) == 0)
{
// the segment is abandoned, try to reclaim it into our heap
if (_mi_segment_attempt_reclaim(mi_heap_get_default(), segment)) {
mi_assert_internal(_mi_prim_thread_id() == mi_atomic_load_relaxed(&segment->thread_id));
mi_free(block); // recursively free as now it will be a local free in our heap
return;
}
}
// The padding check may access the non-thread-owned page for the key values.
// that is safe as these are constant and the page won't be freed (as the block is not freed yet).
mi_check_padding(page, block);
// adjust stats (after padding check and potentially recursive `mi_free` above)
mi_stat_free(page, block); // stat_free may access the padding
mi_track_free_size(block, mi_page_usable_size_of(page,block));
// for small size, ensure we can fit the delayed thread pointers without triggering overflow detection
_mi_padding_shrink(page, block, sizeof(mi_block_t));
if (segment->kind == MI_SEGMENT_HUGE) {
#if MI_HUGE_PAGE_ABANDON
// huge page segments are always abandoned and can be freed immediately
_mi_segment_huge_page_free(segment, page, block);
return;
#else
// huge pages are special as they occupy the entire segment
// as these are large we reset the memory occupied by the page so it is available to other threads
// (as the owning thread needs to actually free the memory later).
_mi_segment_huge_page_reset(segment, page, block);
#endif
}
else {
#if (MI_DEBUG>0) && !MI_TRACK_ENABLED && !MI_TSAN // note: when tracking, cannot use mi_usable_size with multi-threading
memset(block, MI_DEBUG_FREED, mi_usable_size(block));
#endif
}
// and finally free the actual block by pushing it on the owning heap
// thread_delayed free list (or heap delayed free list)
mi_free_block_delayed_mt(page,block);
}
// ------------------------------------------------------
// Usable size
// ------------------------------------------------------
// Bytes available in a block
static size_t mi_decl_noinline mi_page_usable_aligned_size_of(const mi_page_t* page, const void* p) mi_attr_noexcept {
const mi_block_t* block = _mi_page_ptr_unalign(page, p);
const size_t size = mi_page_usable_size_of(page, block);
const ptrdiff_t adjust = (uint8_t*)p - (uint8_t*)block;
mi_assert_internal(adjust >= 0 && (size_t)adjust <= size);
return (size - adjust);
}
static inline size_t _mi_usable_size(const void* p, const char* msg) mi_attr_noexcept {
const mi_segment_t* const segment = mi_checked_ptr_segment(p, msg);
if mi_unlikely(segment==NULL) return 0;
const mi_page_t* const page = _mi_segment_page_of(segment, p);
if mi_likely(!mi_page_has_aligned(page)) {
const mi_block_t* block = (const mi_block_t*)p;
return mi_page_usable_size_of(page, block);
}
else {
// split out to separate routine for improved code generation
return mi_page_usable_aligned_size_of(page, p);
}
}
mi_decl_nodiscard size_t mi_usable_size(const void* p) mi_attr_noexcept {
return _mi_usable_size(p, "mi_usable_size");
}
// ------------------------------------------------------
// Free variants
// ------------------------------------------------------
void mi_free_size(void* p, size_t size) mi_attr_noexcept {
MI_UNUSED_RELEASE(size);
mi_assert(p == NULL || size <= _mi_usable_size(p,"mi_free_size"));
mi_free(p);
}
void mi_free_size_aligned(void* p, size_t size, size_t alignment) mi_attr_noexcept {
MI_UNUSED_RELEASE(alignment);
mi_assert(((uintptr_t)p % alignment) == 0);
mi_free_size(p,size);
}
void mi_free_aligned(void* p, size_t alignment) mi_attr_noexcept {
MI_UNUSED_RELEASE(alignment);
mi_assert(((uintptr_t)p % alignment) == 0);
mi_free(p);
}
// ------------------------------------------------------
// Check for double free in secure and debug mode
// This is somewhat expensive so only enabled for secure mode 4
// ------------------------------------------------------
#if (MI_ENCODE_FREELIST && (MI_SECURE>=4 || MI_DEBUG!=0))
// linear check if the free list contains a specific element
static bool mi_list_contains(const mi_page_t* page, const mi_block_t* list, const mi_block_t* elem) {
while (list != NULL) {
if (elem==list) return true;
list = mi_block_next(page, list);
}
return false;
}
static mi_decl_noinline bool mi_check_is_double_freex(const mi_page_t* page, const mi_block_t* block) {
// The decoded value is in the same page (or NULL).
// Walk the free lists to verify positively if it is already freed
if (mi_list_contains(page, page->free, block) ||
mi_list_contains(page, page->local_free, block) ||
mi_list_contains(page, mi_page_thread_free(page), block))
{
_mi_error_message(EAGAIN, "double free detected of block %p with size %zu\n", block, mi_page_block_size(page));
return true;
}
return false;
}
#define mi_track_page(page,access) { size_t psize; void* pstart = _mi_page_start(_mi_page_segment(page),page,&psize); mi_track_mem_##access( pstart, psize); }
static inline bool mi_check_is_double_free(const mi_page_t* page, const mi_block_t* block) {
bool is_double_free = false;
mi_block_t* n = mi_block_nextx(page, block, page->keys); // pretend it is freed, and get the decoded first field
if (((uintptr_t)n & (MI_INTPTR_SIZE-1))==0 && // quick check: aligned pointer?
(n==NULL || mi_is_in_same_page(block, n))) // quick check: in same page or NULL?
{
// Suspicious: decoded value a in block is in the same page (or NULL) -- maybe a double free?
// (continue in separate function to improve code generation)
is_double_free = mi_check_is_double_freex(page, block);
}
return is_double_free;
}
#else
static inline bool mi_check_is_double_free(const mi_page_t* page, const mi_block_t* block) {
MI_UNUSED(page);
MI_UNUSED(block);
return false;
}
#endif
// ---------------------------------------------------------------------------
// Check for heap block overflow by setting up padding at the end of the block
// ---------------------------------------------------------------------------
#if MI_PADDING // && !MI_TRACK_ENABLED
static bool mi_page_decode_padding(const mi_page_t* page, const mi_block_t* block, size_t* delta, size_t* bsize) {
*bsize = mi_page_usable_block_size(page);
const mi_padding_t* const padding = (mi_padding_t*)((uint8_t*)block + *bsize);
mi_track_mem_defined(padding,sizeof(mi_padding_t));
*delta = padding->delta;
uint32_t canary = padding->canary;
uintptr_t keys[2];
keys[0] = page->keys[0];
keys[1] = page->keys[1];
bool ok = ((uint32_t)mi_ptr_encode(page,block,keys) == canary && *delta <= *bsize);
mi_track_mem_noaccess(padding,sizeof(mi_padding_t));
return ok;
}
// Return the exact usable size of a block.
static size_t mi_page_usable_size_of(const mi_page_t* page, const mi_block_t* block) {
size_t bsize;
size_t delta;
bool ok = mi_page_decode_padding(page, block, &delta, &bsize);
mi_assert_internal(ok); mi_assert_internal(delta <= bsize);
return (ok ? bsize - delta : 0);
}
// When a non-thread-local block is freed, it becomes part of the thread delayed free
// list that is freed later by the owning heap. If the exact usable size is too small to
// contain the pointer for the delayed list, then shrink the padding (by decreasing delta)
// so it will later not trigger an overflow error in `mi_free_block`.
void _mi_padding_shrink(const mi_page_t* page, const mi_block_t* block, const size_t min_size) {
size_t bsize;
size_t delta;
bool ok = mi_page_decode_padding(page, block, &delta, &bsize);
mi_assert_internal(ok);
if (!ok || (bsize - delta) >= min_size) return; // usually already enough space
mi_assert_internal(bsize >= min_size);
if (bsize < min_size) return; // should never happen
size_t new_delta = (bsize - min_size);
mi_assert_internal(new_delta < bsize);
mi_padding_t* padding = (mi_padding_t*)((uint8_t*)block + bsize);
mi_track_mem_defined(padding,sizeof(mi_padding_t));
padding->delta = (uint32_t)new_delta;
mi_track_mem_noaccess(padding,sizeof(mi_padding_t));
}
#else
static size_t mi_page_usable_size_of(const mi_page_t* page, const mi_block_t* block) {
MI_UNUSED(block);
return mi_page_usable_block_size(page);
}
void _mi_padding_shrink(const mi_page_t* page, const mi_block_t* block, const size_t min_size) {
MI_UNUSED(page);
MI_UNUSED(block);
MI_UNUSED(min_size);
}
#endif
#if MI_PADDING && MI_PADDING_CHECK
static bool mi_verify_padding(const mi_page_t* page, const mi_block_t* block, size_t* size, size_t* wrong) {
size_t bsize;
size_t delta;
bool ok = mi_page_decode_padding(page, block, &delta, &bsize);
*size = *wrong = bsize;
if (!ok) return false;
mi_assert_internal(bsize >= delta);
*size = bsize - delta;
if (!mi_page_is_huge(page)) {
uint8_t* fill = (uint8_t*)block + bsize - delta;
const size_t maxpad = (delta > MI_MAX_ALIGN_SIZE ? MI_MAX_ALIGN_SIZE : delta); // check at most the first N padding bytes
mi_track_mem_defined(fill, maxpad);
for (size_t i = 0; i < maxpad; i++) {
if (fill[i] != MI_DEBUG_PADDING) {
*wrong = bsize - delta + i;
ok = false;
break;
}
}
mi_track_mem_noaccess(fill, maxpad);
}
return ok;
}
static void mi_check_padding(const mi_page_t* page, const mi_block_t* block) {
size_t size;
size_t wrong;
if (!mi_verify_padding(page,block,&size,&wrong)) {
_mi_error_message(EFAULT, "buffer overflow in heap block %p of size %zu: write after %zu bytes\n", block, size, wrong );
}
}
#else
static void mi_check_padding(const mi_page_t* page, const mi_block_t* block) {
MI_UNUSED(page);
MI_UNUSED(block);
}
#endif
// only maintain stats for smaller objects if requested
#if (MI_STAT>0)
static void mi_stat_free(const mi_page_t* page, const mi_block_t* block) {
#if (MI_STAT < 2)
MI_UNUSED(block);
#endif
mi_heap_t* const heap = mi_heap_get_default();
const size_t bsize = mi_page_usable_block_size(page);
#if (MI_STAT>1)
const size_t usize = mi_page_usable_size_of(page, block);
mi_heap_stat_decrease(heap, malloc, usize);
#endif
if (bsize <= MI_MEDIUM_OBJ_SIZE_MAX) {
mi_heap_stat_decrease(heap, normal, bsize);
#if (MI_STAT > 1)
mi_heap_stat_decrease(heap, normal_bins[_mi_bin(bsize)], 1);
#endif
}
else if (bsize <= MI_LARGE_OBJ_SIZE_MAX) {
mi_heap_stat_decrease(heap, large, bsize);
}
else {
mi_heap_stat_decrease(heap, huge, bsize);
}
}
#else
static void mi_stat_free(const mi_page_t* page, const mi_block_t* block) {
MI_UNUSED(page); MI_UNUSED(block);
}
#endif

26
src/heap.c
View File

@ -104,6 +104,10 @@ static bool mi_heap_page_collect(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t
// still used blocks but the thread is done; abandon the page
_mi_page_abandon(page, pq);
}
if (collect == MI_FORCE) {
mi_segment_t* segment = _mi_page_segment(page);
_mi_segment_collect(segment, true /* force? */, &heap->tld->segments);
}
return true; // don't break
}
@ -120,7 +124,7 @@ static void mi_heap_collect_ex(mi_heap_t* heap, mi_collect_t collect)
{
if (heap==NULL || !mi_heap_is_initialized(heap)) return;
const bool force = collect >= MI_FORCE;
const bool force = (collect >= MI_FORCE);
_mi_deferred_free(heap, force);
// note: never reclaim on collect but leave it to threads that need storage to reclaim
@ -158,16 +162,13 @@ static void mi_heap_collect_ex(mi_heap_t* heap, mi_collect_t collect)
// note: forced purge can be quite expensive if many threads are created/destroyed so we do not force on abandonment
_mi_abandoned_collect(heap, collect == MI_FORCE /* force? */, &heap->tld->segments);
// collect segment local caches
if (force) {
_mi_segment_thread_collect(&heap->tld->segments);
}
// collect regions on program-exit (or shared library unload)
// if forced, collect thread data cache on program-exit (or shared library unload)
if (force && _mi_is_main_thread() && mi_heap_is_backing(heap)) {
_mi_thread_data_collect(); // collect thread data cache
_mi_arena_collect(true /* force purge */, &heap->tld->stats);
}
// collect arenas (this is program wide so don't force purges on abandonment of threads)
_mi_arenas_collect(collect == MI_FORCE /* force purge? */, &heap->tld->stats);
}
void _mi_heap_collect_abandon(mi_heap_t* heap) {
@ -425,7 +426,7 @@ void mi_heap_delete(mi_heap_t* heap)
if (heap==NULL || !mi_heap_is_initialized(heap)) return;
if (!mi_heap_is_backing(heap)) {
// tranfer still used pages to the backing heap
// transfer still used pages to the backing heap
mi_heap_absorb(heap->tld->heap_backing, heap);
}
else {
@ -474,8 +475,7 @@ static bool mi_heap_page_check_owned(mi_heap_t* heap, mi_page_queue_t* pq, mi_pa
MI_UNUSED(heap);
MI_UNUSED(pq);
bool* found = (bool*)vfound;
mi_segment_t* segment = _mi_page_segment(page);
void* start = _mi_page_start(segment, page, NULL);
void* start = mi_page_start(page);
void* end = (uint8_t*)start + (page->capacity * mi_page_block_size(page));
*found = (p >= start && p < end);
return (!*found); // continue if not found
@ -521,7 +521,7 @@ static bool mi_heap_area_visit_blocks(const mi_heap_area_ex_t* xarea, mi_block_v
const size_t bsize = mi_page_block_size(page);
const size_t ubsize = mi_page_usable_block_size(page); // without padding
size_t psize;
uint8_t* pstart = _mi_page_start(_mi_page_segment(page), page, &psize);
uint8_t* pstart = _mi_segment_page_start(_mi_page_segment(page), page, &psize);
if (page->capacity == 1) {
// optimize page with one block
@ -588,7 +588,7 @@ static bool mi_heap_visit_areas_page(mi_heap_t* heap, mi_page_queue_t* pq, mi_pa
xarea.page = page;
xarea.area.reserved = page->reserved * bsize;
xarea.area.committed = page->capacity * bsize;
xarea.area.blocks = _mi_page_start(_mi_page_segment(page), page, NULL);
xarea.area.blocks = mi_page_start(page);
xarea.area.used = page->used; // number of blocks in use (#553)
xarea.area.block_size = ubsize;
xarea.area.full_block_size = bsize;

35
src/init.c
View File

@ -14,16 +14,19 @@ terms of the MIT license. A copy of the license can be found in the file
// Empty page used to initialize the small free pages array
const mi_page_t _mi_page_empty = {
0, false, false, false,
0,
false, false, false, false,
0, // capacity
0, // reserved capacity
{ 0 }, // flags
false, // is_zero
0, // retire_expire
NULL, // free
0, // used
0, // xblock_size
NULL, // local_free
0, // used
0, // block size shift
0, // block_size
NULL, // page_start
#if (MI_PADDING || MI_ENCODE_FREELIST)
{ 0, 0 },
#endif
@ -84,7 +87,9 @@ const mi_page_t _mi_page_empty = {
MI_STAT_COUNT_NULL(), MI_STAT_COUNT_NULL(), \
MI_STAT_COUNT_NULL(), \
{ 0, 0 }, { 0, 0 }, { 0, 0 }, { 0, 0 }, \
{ 0, 0 }, { 0, 0 }, { 0, 0 }, { 0, 0 }, { 0, 0 }, { 0, 0 } \
{ 0, 0 }, { 0, 0 }, { 0, 0 }, { 0, 0 }, \
{ 0, 0 }, { 0, 0 }, { 0, 0 }, { 0, 0 }, \
{ 0, 0 } \
MI_STAT_COUNT_END_NULL()
@ -110,8 +115,6 @@ const mi_page_t _mi_page_empty = {
mi_decl_cache_align const mi_heap_t _mi_heap_empty = {
NULL,
MI_SMALL_PAGES_EMPTY,
MI_PAGE_QUEUES_EMPTY,
MI_ATOMIC_VAR_INIT(NULL),
0, // tid
0, // cookie
@ -121,7 +124,9 @@ mi_decl_cache_align const mi_heap_t _mi_heap_empty = {
0, // page count
MI_BIN_FULL, 0, // page retired min/max
NULL, // next
false
false,
MI_SMALL_PAGES_EMPTY,
MI_PAGE_QUEUES_EMPTY
};
#define tld_empty_stats ((mi_stats_t*)((uint8_t*)&tld_empty + offsetof(mi_tld_t,stats)))
@ -131,7 +136,7 @@ mi_decl_cache_align static const mi_tld_t tld_empty = {
0,
false,
NULL, NULL,
{ MI_SEGMENT_SPAN_QUEUES_EMPTY, 0, 0, 0, 0, tld_empty_stats, tld_empty_os }, // segments
{ MI_SEGMENT_SPAN_QUEUES_EMPTY, 0, 0, 0, 0, 0, tld_empty_stats, tld_empty_os }, // segments
{ 0, tld_empty_stats }, // os
{ MI_STATS_NULL } // stats
};
@ -148,15 +153,13 @@ extern mi_heap_t _mi_heap_main;
static mi_tld_t tld_main = {
0, false,
&_mi_heap_main, & _mi_heap_main,
{ MI_SEGMENT_SPAN_QUEUES_EMPTY, 0, 0, 0, 0, &tld_main.stats, &tld_main.os }, // segments
{ MI_SEGMENT_SPAN_QUEUES_EMPTY, 0, 0, 0, 0, 0, &tld_main.stats, &tld_main.os }, // segments
{ 0, &tld_main.stats }, // os
{ MI_STATS_NULL } // stats
};
mi_heap_t _mi_heap_main = {
&tld_main,
MI_SMALL_PAGES_EMPTY,
MI_PAGE_QUEUES_EMPTY,
MI_ATOMIC_VAR_INIT(NULL),
0, // thread id
0, // initial cookie
@ -166,7 +169,9 @@ mi_heap_t _mi_heap_main = {
0, // page count
MI_BIN_FULL, 0, // page retired min/max
NULL, // next heap
false // can reclaim
false, // can reclaim
MI_SMALL_PAGES_EMPTY,
MI_PAGE_QUEUES_EMPTY
};
bool _mi_process_is_initialized = false; // set to `true` in `mi_process_init`.
@ -203,7 +208,7 @@ mi_heap_t* _mi_heap_main_get(void) {
typedef struct mi_thread_data_s {
mi_heap_t heap; // must come first due to cast in `_mi_heap_done`
mi_tld_t tld;
mi_memid_t memid;
mi_memid_t memid; // must come last due to zero'ing
} mi_thread_data_t;
@ -249,7 +254,7 @@ static mi_thread_data_t* mi_thread_data_zalloc(void) {
}
if (td != NULL && !is_zero) {
_mi_memzero_aligned(td, sizeof(*td));
_mi_memzero_aligned(td, offsetof(mi_thread_data_t,memid));
}
return td;
}
@ -455,7 +460,7 @@ void _mi_heap_set_default_direct(mi_heap_t* heap) {
#if defined(MI_TLS_SLOT)
mi_prim_tls_slot_set(MI_TLS_SLOT,heap);
#elif defined(MI_TLS_PTHREAD_SLOT_OFS)
*mi_tls_pthread_heap_slot() = heap;
*mi_prim_tls_pthread_heap_slot() = heap;
#elif defined(MI_TLS_PTHREAD)
// we use _mi_heap_default_key
#else

273
src/libc.c Normal file
View File

@ -0,0 +1,273 @@
/* ----------------------------------------------------------------------------
Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
This is free software; you can redistribute it and/or modify it under the
terms of the MIT license. A copy of the license can be found in the file
"LICENSE" at the root of this distribution.
-----------------------------------------------------------------------------*/
// --------------------------------------------------------
// This module defines various std libc functions to reduce
// the dependency on libc, and also prevent errors caused
// by some libc implementations when called before `main`
// executes (due to malloc redirection)
// --------------------------------------------------------
#include "mimalloc.h"
#include "mimalloc/internal.h"
#include "mimalloc/prim.h" // mi_prim_getenv
char _mi_toupper(char c) {
if (c >= 'a' && c <= 'z') return (c - 'a' + 'A');
else return c;
}
int _mi_strnicmp(const char* s, const char* t, size_t n) {
if (n == 0) return 0;
for (; *s != 0 && *t != 0 && n > 0; s++, t++, n--) {
if (_mi_toupper(*s) != _mi_toupper(*t)) break;
}
return (n == 0 ? 0 : *s - *t);
}
void _mi_strlcpy(char* dest, const char* src, size_t dest_size) {
if (dest==NULL || src==NULL || dest_size == 0) return;
// copy until end of src, or when dest is (almost) full
while (*src != 0 && dest_size > 1) {
*dest++ = *src++;
dest_size--;
}
// always zero terminate
*dest = 0;
}
void _mi_strlcat(char* dest, const char* src, size_t dest_size) {
if (dest==NULL || src==NULL || dest_size == 0) return;
// find end of string in the dest buffer
while (*dest != 0 && dest_size > 1) {
dest++;
dest_size--;
}
// and catenate
_mi_strlcpy(dest, src, dest_size);
}
size_t _mi_strlen(const char* s) {
if (s==NULL) return 0;
size_t len = 0;
while(s[len] != 0) { len++; }
return len;
}
size_t _mi_strnlen(const char* s, size_t max_len) {
if (s==NULL) return 0;
size_t len = 0;
while(s[len] != 0 && len < max_len) { len++; }
return len;
}
#ifdef MI_NO_GETENV
bool _mi_getenv(const char* name, char* result, size_t result_size) {
MI_UNUSED(name);
MI_UNUSED(result);
MI_UNUSED(result_size);
return false;
}
#else
bool _mi_getenv(const char* name, char* result, size_t result_size) {
if (name==NULL || result == NULL || result_size < 64) return false;
return _mi_prim_getenv(name,result,result_size);
}
#endif
// --------------------------------------------------------
// Define our own limited `_mi_vsnprintf` and `_mi_snprintf`
// This is mostly to avoid calling these when libc is not yet
// initialized (and to reduce dependencies)
//
// format: d i, p x u, s
// prec: z l ll L
// width: 10
// align-left: -
// fill: 0
// plus: +
// --------------------------------------------------------
static void mi_outc(char c, char** out, char* end) {
char* p = *out;
if (p >= end) return;
*p = c;
*out = p + 1;
}
static void mi_outs(const char* s, char** out, char* end) {
if (s == NULL) return;
char* p = *out;
while (*s != 0 && p < end) {
*p++ = *s++;
}
*out = p;
}
static void mi_out_fill(char fill, size_t len, char** out, char* end) {
char* p = *out;
for (size_t i = 0; i < len && p < end; i++) {
*p++ = fill;
}
*out = p;
}
static void mi_out_alignright(char fill, char* start, size_t len, size_t extra, char* end) {
if (len == 0 || extra == 0) return;
if (start + len + extra >= end) return;
// move `len` characters to the right (in reverse since it can overlap)
for (size_t i = 1; i <= len; i++) {
start[len + extra - i] = start[len - i];
}
// and fill the start
for (size_t i = 0; i < extra; i++) {
start[i] = fill;
}
}
static void mi_out_num(uintptr_t x, size_t base, char prefix, char** out, char* end)
{
if (x == 0 || base == 0 || base > 16) {
if (prefix != 0) { mi_outc(prefix, out, end); }
mi_outc('0',out,end);
}
else {
// output digits in reverse
char* start = *out;
while (x > 0) {
char digit = (char)(x % base);
mi_outc((digit <= 9 ? '0' + digit : 'A' + digit - 10),out,end);
x = x / base;
}
if (prefix != 0) {
mi_outc(prefix, out, end);
}
size_t len = *out - start;
// and reverse in-place
for (size_t i = 0; i < (len / 2); i++) {
char c = start[len - i - 1];
start[len - i - 1] = start[i];
start[i] = c;
}
}
}
#define MI_NEXTC() c = *in; if (c==0) break; in++;
void _mi_vsnprintf(char* buf, size_t bufsize, const char* fmt, va_list args) {
if (buf == NULL || bufsize == 0 || fmt == NULL) return;
buf[bufsize - 1] = 0;
char* const end = buf + (bufsize - 1);
const char* in = fmt;
char* out = buf;
while (true) {
if (out >= end) break;
char c;
MI_NEXTC();
if (c != '%') {
if ((c >= ' ' && c <= '~') || c=='\n' || c=='\r' || c=='\t') { // output visible ascii or standard control only
mi_outc(c, &out, end);
}
}
else {
MI_NEXTC();
char fill = ' ';
size_t width = 0;
char numtype = 'd';
char numplus = 0;
bool alignright = true;
if (c == '+' || c == ' ') { numplus = c; MI_NEXTC(); }
if (c == '-') { alignright = false; MI_NEXTC(); }
if (c == '0') { fill = '0'; MI_NEXTC(); }
if (c >= '1' && c <= '9') {
width = (c - '0'); MI_NEXTC();
while (c >= '0' && c <= '9') {
width = (10 * width) + (c - '0'); MI_NEXTC();
}
if (c == 0) break; // extra check due to while
}
if (c == 'z' || c == 't' || c == 'L') { numtype = c; MI_NEXTC(); }
else if (c == 'l') {
numtype = c; MI_NEXTC();
if (c == 'l') { numtype = 'L'; MI_NEXTC(); }
}
char* start = out;
if (c == 's') {
// string
const char* s = va_arg(args, const char*);
mi_outs(s, &out, end);
}
else if (c == 'p' || c == 'x' || c == 'u') {
// unsigned
uintptr_t x = 0;
if (c == 'x' || c == 'u') {
if (numtype == 'z') x = va_arg(args, size_t);
else if (numtype == 't') x = va_arg(args, uintptr_t); // unsigned ptrdiff_t
else if (numtype == 'L') x = (uintptr_t)va_arg(args, unsigned long long);
else x = va_arg(args, unsigned long);
}
else if (c == 'p') {
x = va_arg(args, uintptr_t);
mi_outs("0x", &out, end);
start = out;
width = (width >= 2 ? width - 2 : 0);
}
if (width == 0 && (c == 'x' || c == 'p')) {
if (c == 'p') { width = 2 * (x <= UINT32_MAX ? 4 : ((x >> 16) <= UINT32_MAX ? 6 : sizeof(void*))); }
if (width == 0) { width = 2; }
fill = '0';
}
mi_out_num(x, (c == 'x' || c == 'p' ? 16 : 10), numplus, &out, end);
}
else if (c == 'i' || c == 'd') {
// signed
intptr_t x = 0;
if (numtype == 'z') x = va_arg(args, intptr_t );
else if (numtype == 't') x = va_arg(args, ptrdiff_t);
else if (numtype == 'L') x = (intptr_t)va_arg(args, long long);
else x = va_arg(args, long);
char pre = 0;
if (x < 0) {
pre = '-';
if (x > INTPTR_MIN) { x = -x; }
}
else if (numplus != 0) {
pre = numplus;
}
mi_out_num((uintptr_t)x, 10, pre, &out, end);
}
else if (c >= ' ' && c <= '~') {
// unknown format
mi_outc('%', &out, end);
mi_outc(c, &out, end);
}
// fill & align
mi_assert_internal(out <= end);
mi_assert_internal(out >= start);
const size_t len = out - start;
if (len < width) {
mi_out_fill(fill, width - len, &out, end);
if (alignright && out <= end) {
mi_out_alignright(fill, start, len, width - len, end);
}
}
}
}
mi_assert_internal(out <= end);
*out = 0;
}
void _mi_snprintf(char* buf, size_t buflen, const char* fmt, ...) {
va_list args;
va_start(args, fmt);
_mi_vsnprintf(buf, buflen, fmt, args);
va_end(args);
}

92
src/options.c
View File

@ -9,9 +9,9 @@ terms of the MIT license. A copy of the license can be found in the file
#include "mimalloc/atomic.h"
#include "mimalloc/prim.h" // mi_prim_out_stderr
#include <stdio.h> // FILE
#include <stdio.h> // stdin/stdout
#include <stdlib.h> // abort
#include <stdarg.h>
static long mi_max_error_count = 16; // stop outputting errors after this (use < 0 for no limit)
@ -65,7 +65,7 @@ static mi_option_desc_t options[_mi_option_last] =
{ 0, UNINIT, MI_OPTION_LEGACY(allow_large_os_pages,large_os_pages) }, // use large OS pages, use only with eager commit to prevent fragmentation of VMA's
{ 0, UNINIT, MI_OPTION(reserve_huge_os_pages) }, // per 1GiB huge pages
{-1, UNINIT, MI_OPTION(reserve_huge_os_pages_at) }, // reserve huge pages at node N
{ 0, UNINIT, MI_OPTION(reserve_os_memory) },
{ 0, UNINIT, MI_OPTION(reserve_os_memory) }, // reserve OS memory in advance
{ 0, UNINIT, MI_OPTION(deprecated_segment_cache) }, // cache N segments per thread
{ 0, UNINIT, MI_OPTION(deprecated_page_reset) }, // reset page memory on free
{ 0, UNINIT, MI_OPTION_LEGACY(abandoned_page_purge,abandoned_page_reset) }, // reset free page memory when a thread terminates
@ -77,11 +77,11 @@ static mi_option_desc_t options[_mi_option_last] =
#endif
{ 10, UNINIT, MI_OPTION_LEGACY(purge_delay,reset_delay) }, // purge delay in milli-seconds
{ 0, UNINIT, MI_OPTION(use_numa_nodes) }, // 0 = use available numa nodes, otherwise use at most N nodes.
{ 0, UNINIT, MI_OPTION(limit_os_alloc) }, // 1 = do not use OS memory for allocation (but only reserved arenas)
{ 0, UNINIT, MI_OPTION_LEGACY(disallow_os_alloc,limit_os_alloc) }, // 1 = do not use OS memory for allocation (but only reserved arenas)
{ 100, UNINIT, MI_OPTION(os_tag) }, // only apple specific for now but might serve more or less related purpose
{ 16, UNINIT, MI_OPTION(max_errors) }, // maximum errors that are output
{ 16, UNINIT, MI_OPTION(max_warnings) }, // maximum warnings that are output
{ 8, UNINIT, MI_OPTION(max_segment_reclaim)}, // max. number of segment reclaims from the abandoned segments per try.
{ 10, UNINIT, MI_OPTION(max_segment_reclaim)}, // max. percentage of the abandoned segments per try.
{ 0, UNINIT, MI_OPTION(destroy_on_exit)}, // release all OS memory on process exit; careful with dangling pointer or after-exit frees!
#if (MI_INTPTR_SIZE>4)
{ 1024L * 1024L, UNINIT, MI_OPTION(arena_reserve) }, // reserve memory N KiB at a time
@ -90,10 +90,16 @@ static mi_option_desc_t options[_mi_option_last] =
#endif
{ 10, UNINIT, MI_OPTION(arena_purge_mult) }, // purge delay multiplier for arena's
{ 1, UNINIT, MI_OPTION_LEGACY(purge_extend_delay, decommit_extend_delay) },
{ 1, UNINIT, MI_OPTION(abandoned_reclaim_on_free) },// reclaim an abandoned segment on a free
{ 0, UNINIT, MI_OPTION(disallow_arena_alloc) }, // 1 = do not use arena's for allocation (except if using specific arena id's)
};
static void mi_option_init(mi_option_desc_t* desc);
static bool mi_option_has_size_in_kib(mi_option_t option) {
return (option == mi_option_reserve_os_memory || option == mi_option_arena_reserve);
}
void _mi_options_init(void) {
// called on process load; should not be called before the CRT is initialized!
// (e.g. do not call this from process_init as that may run before CRT initialization)
@ -104,7 +110,7 @@ void _mi_options_init(void) {
// if (option != mi_option_verbose)
{
mi_option_desc_t* desc = &options[option];
_mi_verbose_message("option '%s': %ld\n", desc->name, desc->value);
_mi_verbose_message("option '%s': %ld %s\n", desc->name, desc->value, (mi_option_has_size_in_kib(option) ? "KiB" : ""));
}
}
mi_max_error_count = mi_option_get(mi_option_max_errors);
@ -128,7 +134,7 @@ mi_decl_nodiscard long mi_option_get_clamp(mi_option_t option, long min, long ma
}
mi_decl_nodiscard size_t mi_option_get_size(mi_option_t option) {
mi_assert_internal(option == mi_option_reserve_os_memory || option == mi_option_arena_reserve);
mi_assert_internal(mi_option_has_size_in_kib(option));
long x = mi_option_get(option);
return (x < 0 ? 0 : (size_t)x * MI_KiB);
}
@ -311,12 +317,12 @@ void _mi_fputs(mi_output_fun* out, void* arg, const char* prefix, const char* me
}
// Define our own limited `fprintf` that avoids memory allocation.
// We do this using `snprintf` with a limited buffer.
// We do this using `_mi_vsnprintf` with a limited buffer.
static void mi_vfprintf( mi_output_fun* out, void* arg, const char* prefix, const char* fmt, va_list args ) {
char buf[512];
if (fmt==NULL) return;
if (!mi_recurse_enter()) return;
vsnprintf(buf,sizeof(buf)-1,fmt,args);
_mi_vsnprintf(buf, sizeof(buf)-1, fmt, args);
mi_recurse_exit();
_mi_fputs(out,arg,prefix,buf);
}
@ -331,7 +337,7 @@ void _mi_fprintf( mi_output_fun* out, void* arg, const char* fmt, ... ) {
static void mi_vfprintf_thread(mi_output_fun* out, void* arg, const char* prefix, const char* fmt, va_list args) {
if (prefix != NULL && _mi_strnlen(prefix,33) <= 32 && !_mi_is_main_thread()) {
char tprefix[64];
snprintf(tprefix, sizeof(tprefix), "%sthread 0x%llx: ", prefix, (unsigned long long)_mi_thread_id());
_mi_snprintf(tprefix, sizeof(tprefix), "%sthread 0x%tx: ", prefix, (uintptr_t)_mi_thread_id());
mi_vfprintf(out, arg, tprefix, fmt, args);
}
else {
@ -434,68 +440,6 @@ void _mi_error_message(int err, const char* fmt, ...) {
// --------------------------------------------------------
// Initialize options by checking the environment
// --------------------------------------------------------
char _mi_toupper(char c) {
if (c >= 'a' && c <= 'z') return (c - 'a' + 'A');
else return c;
}
int _mi_strnicmp(const char* s, const char* t, size_t n) {
if (n == 0) return 0;
for (; *s != 0 && *t != 0 && n > 0; s++, t++, n--) {
if (_mi_toupper(*s) != _mi_toupper(*t)) break;
}
return (n == 0 ? 0 : *s - *t);
}
void _mi_strlcpy(char* dest, const char* src, size_t dest_size) {
if (dest==NULL || src==NULL || dest_size == 0) return;
// copy until end of src, or when dest is (almost) full
while (*src != 0 && dest_size > 1) {
*dest++ = *src++;
dest_size--;
}
// always zero terminate
*dest = 0;
}
void _mi_strlcat(char* dest, const char* src, size_t dest_size) {
if (dest==NULL || src==NULL || dest_size == 0) return;
// find end of string in the dest buffer
while (*dest != 0 && dest_size > 1) {
dest++;
dest_size--;
}
// and catenate
_mi_strlcpy(dest, src, dest_size);
}
size_t _mi_strlen(const char* s) {
if (s==NULL) return 0;
size_t len = 0;
while(s[len] != 0) { len++; }
return len;
}
size_t _mi_strnlen(const char* s, size_t max_len) {
if (s==NULL) return 0;
size_t len = 0;
while(s[len] != 0 && len < max_len) { len++; }
return len;
}
#ifdef MI_NO_GETENV
static bool mi_getenv(const char* name, char* result, size_t result_size) {
MI_UNUSED(name);
MI_UNUSED(result);
MI_UNUSED(result_size);
return false;
}
#else
static bool mi_getenv(const char* name, char* result, size_t result_size) {
if (name==NULL || result == NULL || result_size < 64) return false;
return _mi_prim_getenv(name,result,result_size);
}
#endif
// TODO: implement ourselves to reduce dependencies on the C runtime
#include <stdlib.h> // strtol
@ -508,11 +452,11 @@ static void mi_option_init(mi_option_desc_t* desc) {
char buf[64+1];
_mi_strlcpy(buf, "mimalloc_", sizeof(buf));
_mi_strlcat(buf, desc->name, sizeof(buf));
bool found = mi_getenv(buf, s, sizeof(s));
bool found = _mi_getenv(buf, s, sizeof(s));
if (!found && desc->legacy_name != NULL) {
_mi_strlcpy(buf, "mimalloc_", sizeof(buf));
_mi_strlcat(buf, desc->legacy_name, sizeof(buf));
found = mi_getenv(buf, s, sizeof(s));
found = _mi_getenv(buf, s, sizeof(s));
if (found) {
_mi_warning_message("environment option \"mimalloc_%s\" is deprecated -- use \"mimalloc_%s\" instead.\n", desc->legacy_name, desc->name);
}

35
src/os.c
View File

@ -73,14 +73,6 @@ void _mi_os_init(void) {
bool _mi_os_decommit(void* addr, size_t size, mi_stats_t* stats);
bool _mi_os_commit(void* addr, size_t size, bool* is_zero, mi_stats_t* tld_stats);
static void* mi_align_up_ptr(void* p, size_t alignment) {
return (void*)_mi_align_up((uintptr_t)p, alignment);
}
static void* mi_align_down_ptr(void* p, size_t alignment) {
return (void*)_mi_align_down((uintptr_t)p, alignment);
}
/* -----------------------------------------------------------
aligned hinting
@ -141,13 +133,13 @@ static void mi_os_free_huge_os_pages(void* p, size_t size, mi_stats_t* stats);
static void mi_os_prim_free(void* addr, size_t size, bool still_committed, mi_stats_t* tld_stats) {
MI_UNUSED(tld_stats);
mi_stats_t* stats = &_mi_stats_main;
mi_assert_internal((size % _mi_os_page_size()) == 0);
if (addr == NULL || size == 0) return; // || _mi_os_is_huge_reserved(addr)
int err = _mi_prim_free(addr, size);
if (err != 0) {
_mi_warning_message("unable to free OS memory (error: %d (0x%x), size: 0x%zx bytes, address: %p)\n", err, err, size, addr);
}
mi_stats_t* stats = &_mi_stats_main;
if (still_committed) { _mi_stat_decrease(&stats->committed, size); }
_mi_stat_decrease(&stats->reserved, size);
}
@ -188,20 +180,22 @@ void _mi_os_free(void* p, size_t size, mi_memid_t memid, mi_stats_t* tld_stats)
-------------------------------------------------------------- */
// Note: the `try_alignment` is just a hint and the returned pointer is not guaranteed to be aligned.
static void* mi_os_prim_alloc(size_t size, size_t try_alignment, bool commit, bool allow_large, bool* is_large, bool* is_zero, mi_stats_t* stats) {
static void* mi_os_prim_alloc(size_t size, size_t try_alignment, bool commit, bool allow_large, bool* is_large, bool* is_zero, mi_stats_t* tld_stats) {
mi_assert_internal(size > 0 && (size % _mi_os_page_size()) == 0);
mi_assert_internal(is_zero != NULL);
mi_assert_internal(is_large != NULL);
if (size == 0) return NULL;
if (!commit) { allow_large = false; }
if (try_alignment == 0) { try_alignment = 1; } // avoid 0 to ensure there will be no divide by zero when aligning
*is_zero = false;
void* p = NULL;
int err = _mi_prim_alloc(size, try_alignment, commit, allow_large, is_large, is_zero, &p);
if (err != 0) {
_mi_warning_message("unable to allocate OS memory (error: %d (0x%x), size: 0x%zx bytes, align: 0x%zx, commit: %d, allow large: %d)\n", err, err, size, try_alignment, commit, allow_large);
}
MI_UNUSED(tld_stats);
mi_stats_t* stats = &_mi_stats_main;
mi_stat_counter_increase(stats->mmap_calls, 1);
if (p != NULL) {
_mi_stat_increase(&stats->reserved, size);
@ -289,10 +283,8 @@ static void* mi_os_prim_alloc_aligned(size_t size, size_t alignment, bool commit
OS API: alloc and alloc_aligned
----------------------------------------------------------- */
void* _mi_os_alloc(size_t size, mi_memid_t* memid, mi_stats_t* tld_stats) {
MI_UNUSED(tld_stats);
void* _mi_os_alloc(size_t size, mi_memid_t* memid, mi_stats_t* stats) {
*memid = _mi_memid_none();
mi_stats_t* stats = &_mi_stats_main;
if (size == 0) return NULL;
size = _mi_os_good_alloc_size(size);
bool os_is_large = false;
@ -304,10 +296,9 @@ void* _mi_os_alloc(size_t size, mi_memid_t* memid, mi_stats_t* tld_stats) {
return p;
}
void* _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool allow_large, mi_memid_t* memid, mi_stats_t* tld_stats)
void* _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool allow_large, mi_memid_t* memid, mi_stats_t* stats)
{
MI_UNUSED(&_mi_os_get_aligned_hint); // suppress unused warnings
MI_UNUSED(tld_stats);
*memid = _mi_memid_none();
if (size == 0) return NULL;
size = _mi_os_good_alloc_size(size);
@ -316,7 +307,7 @@ void* _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool allo
bool os_is_large = false;
bool os_is_zero = false;
void* os_base = NULL;
void* p = mi_os_prim_alloc_aligned(size, alignment, commit, allow_large, &os_is_large, &os_is_zero, &os_base, &_mi_stats_main /*tld->stats*/ );
void* p = mi_os_prim_alloc_aligned(size, alignment, commit, allow_large, &os_is_large, &os_is_zero, &os_base, stats );
if (p != NULL) {
*memid = _mi_memid_create_os(commit, os_is_zero, os_is_large);
memid->mem.os.base = os_base;
@ -327,13 +318,13 @@ void* _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool allo
/* -----------------------------------------------------------
OS aligned allocation with an offset. This is used
for large alignments > MI_ALIGNMENT_MAX. We use a large mimalloc
for large alignments > MI_BLOCK_ALIGNMENT_MAX. We use a large mimalloc
page where the object can be aligned at an offset from the start of the segment.
As we may need to overallocate, we need to free such pointers using `mi_free_aligned`
to use the actual start of the memory region.
----------------------------------------------------------- */
void* _mi_os_alloc_aligned_at_offset(size_t size, size_t alignment, size_t offset, bool commit, bool allow_large, mi_memid_t* memid, mi_stats_t* tld_stats) {
void* _mi_os_alloc_aligned_at_offset(size_t size, size_t alignment, size_t offset, bool commit, bool allow_large, mi_memid_t* memid, mi_stats_t* stats) {
mi_assert(offset <= MI_SEGMENT_SIZE);
mi_assert(offset <= size);
mi_assert((alignment % _mi_os_page_size()) == 0);
@ -341,20 +332,20 @@ void* _mi_os_alloc_aligned_at_offset(size_t size, size_t alignment, size_t offse
if (offset > MI_SEGMENT_SIZE) return NULL;
if (offset == 0) {
// regular aligned allocation
return _mi_os_alloc_aligned(size, alignment, commit, allow_large, memid, tld_stats);
return _mi_os_alloc_aligned(size, alignment, commit, allow_large, memid, stats);
}
else {
// overallocate to align at an offset
const size_t extra = _mi_align_up(offset, alignment) - offset;
const size_t oversize = size + extra;
void* const start = _mi_os_alloc_aligned(oversize, alignment, commit, allow_large, memid, tld_stats);
void* const start = _mi_os_alloc_aligned(oversize, alignment, commit, allow_large, memid, stats);
if (start == NULL) return NULL;
void* const p = (uint8_t*)start + extra;
mi_assert(_mi_is_aligned((uint8_t*)p + offset, alignment));
// decommit the overallocation at the start
if (commit && extra > _mi_os_page_size()) {
_mi_os_decommit(start, extra, tld_stats);
_mi_os_decommit(start, extra, stats);
}
return p;
}

51
src/page-queue.c
View File

@ -1,5 +1,5 @@
/*----------------------------------------------------------------------------
Copyright (c) 2018-2020, Microsoft Research, Daan Leijen
Copyright (c) 2018-2024, Microsoft Research, Daan Leijen
This is free software; you can redistribute it and/or modify it under the
terms of the MIT license. A copy of the license can be found in the file
"LICENSE" at the root of this distribution.
@ -11,6 +11,10 @@ terms of the MIT license. A copy of the license can be found in the file
#ifndef MI_IN_PAGE_C
#error "this file should be included from 'page.c'"
// include to help an IDE
#include "mimalloc.h"
#include "mimalloc/internal.h"
#include "mimalloc/atomic.h"
#endif
/* -----------------------------------------------------------
@ -137,21 +141,25 @@ static bool mi_heap_contains_queue(const mi_heap_t* heap, const mi_page_queue_t*
}
#endif
static mi_page_queue_t* mi_page_queue_of(const mi_page_t* page) {
uint8_t bin = (mi_page_is_in_full(page) ? MI_BIN_FULL : mi_bin(page->xblock_size));
mi_heap_t* heap = mi_page_heap(page);
mi_assert_internal(heap != NULL && bin <= MI_BIN_FULL);
mi_page_queue_t* pq = &heap->pages[bin];
mi_assert_internal(bin >= MI_BIN_HUGE || page->xblock_size == pq->block_size);
mi_assert_expensive(mi_page_queue_contains(pq, page));
return pq;
static inline bool mi_page_is_large_or_huge(const mi_page_t* page) {
return (mi_page_block_size(page) > MI_MEDIUM_OBJ_SIZE_MAX || mi_page_is_huge(page));
}
static mi_page_queue_t* mi_heap_page_queue_of(mi_heap_t* heap, const mi_page_t* page) {
uint8_t bin = (mi_page_is_in_full(page) ? MI_BIN_FULL : mi_bin(page->xblock_size));
mi_assert_internal(heap!=NULL);
uint8_t bin = (mi_page_is_in_full(page) ? MI_BIN_FULL : (mi_page_is_huge(page) ? MI_BIN_HUGE : mi_bin(mi_page_block_size(page))));
mi_assert_internal(bin <= MI_BIN_FULL);
mi_page_queue_t* pq = &heap->pages[bin];
mi_assert_internal(mi_page_is_in_full(page) || page->xblock_size == pq->block_size);
mi_assert_internal((mi_page_block_size(page) == pq->block_size) ||
(mi_page_is_large_or_huge(page) && mi_page_queue_is_huge(pq)) ||
(mi_page_is_in_full(page) && mi_page_queue_is_full(pq)));
return pq;
}
static mi_page_queue_t* mi_page_queue_of(const mi_page_t* page) {
mi_heap_t* heap = mi_page_heap(page);
mi_page_queue_t* pq = mi_heap_page_queue_of(heap, page);
mi_assert_expensive(mi_page_queue_contains(pq, page));
return pq;
}
@ -206,7 +214,9 @@ static bool mi_page_queue_is_empty(mi_page_queue_t* queue) {
static void mi_page_queue_remove(mi_page_queue_t* queue, mi_page_t* page) {
mi_assert_internal(page != NULL);
mi_assert_expensive(mi_page_queue_contains(queue, page));
mi_assert_internal(page->xblock_size == queue->block_size || (page->xblock_size > MI_MEDIUM_OBJ_SIZE_MAX && mi_page_queue_is_huge(queue)) || (mi_page_is_in_full(page) && mi_page_queue_is_full(queue)));
mi_assert_internal(mi_page_block_size(page) == queue->block_size ||
(mi_page_is_large_or_huge(page) && mi_page_queue_is_huge(queue)) ||
(mi_page_is_in_full(page) && mi_page_queue_is_full(queue)));
mi_heap_t* heap = mi_page_heap(page);
if (page->prev != NULL) page->prev->next = page->next;
@ -232,8 +242,8 @@ static void mi_page_queue_push(mi_heap_t* heap, mi_page_queue_t* queue, mi_page_
#if MI_HUGE_PAGE_ABANDON
mi_assert_internal(_mi_page_segment(page)->kind != MI_SEGMENT_HUGE);
#endif
mi_assert_internal(page->xblock_size == queue->block_size ||
(page->xblock_size > MI_MEDIUM_OBJ_SIZE_MAX) ||
mi_assert_internal(mi_page_block_size(page) == queue->block_size ||
(mi_page_is_large_or_huge(page) && mi_page_queue_is_huge(queue)) ||
(mi_page_is_in_full(page) && mi_page_queue_is_full(queue)));
mi_page_set_in_full(page, mi_page_queue_is_full(queue));
@ -259,12 +269,13 @@ static void mi_page_queue_enqueue_from(mi_page_queue_t* to, mi_page_queue_t* fro
mi_assert_internal(page != NULL);
mi_assert_expensive(mi_page_queue_contains(from, page));
mi_assert_expensive(!mi_page_queue_contains(to, page));
mi_assert_internal((page->xblock_size == to->block_size && page->xblock_size == from->block_size) ||
(page->xblock_size == to->block_size && mi_page_queue_is_full(from)) ||
(page->xblock_size == from->block_size && mi_page_queue_is_full(to)) ||
(page->xblock_size > MI_LARGE_OBJ_SIZE_MAX && mi_page_queue_is_huge(to)) ||
(page->xblock_size > MI_LARGE_OBJ_SIZE_MAX && mi_page_queue_is_full(to)));
const size_t bsize = mi_page_block_size(page);
MI_UNUSED(bsize);
mi_assert_internal((bsize == to->block_size && bsize == from->block_size) ||
(bsize == to->block_size && mi_page_queue_is_full(from)) ||
(bsize == from->block_size && mi_page_queue_is_full(to)) ||
(mi_page_is_large_or_huge(page) && mi_page_queue_is_huge(to)) ||
(mi_page_is_large_or_huge(page) && mi_page_queue_is_full(to)));
mi_heap_t* heap = mi_page_heap(page);
if (page->prev != NULL) page->prev->next = page->next;

82
src/page.c
View File

@ -1,5 +1,5 @@
/*----------------------------------------------------------------------------
Copyright (c) 2018-2020, Microsoft Research, Daan Leijen
Copyright (c) 2018-2024, Microsoft Research, Daan Leijen
This is free software; you can redistribute it and/or modify it under the
terms of the MIT license. A copy of the license can be found in the file
"LICENSE" at the root of this distribution.
@ -59,7 +59,7 @@ static inline uint8_t* mi_page_area(const mi_page_t* page) {
static bool mi_page_list_is_valid(mi_page_t* page, mi_block_t* p) {
size_t psize;
uint8_t* page_area = _mi_page_start(_mi_page_segment(page), page, &psize);
uint8_t* page_area = _mi_segment_page_start(_mi_page_segment(page), page, &psize);
mi_block_t* start = (mi_block_t*)page_area;
mi_block_t* end = (mi_block_t*)(page_area + psize);
while(p != NULL) {
@ -78,14 +78,13 @@ static bool mi_page_list_is_valid(mi_page_t* page, mi_block_t* p) {
}
static bool mi_page_is_valid_init(mi_page_t* page) {
mi_assert_internal(page->xblock_size > 0);
mi_assert_internal(mi_page_block_size(page) > 0);
mi_assert_internal(page->used <= page->capacity);
mi_assert_internal(page->capacity <= page->reserved);
mi_segment_t* segment = _mi_page_segment(page);
uint8_t* start = _mi_page_start(segment,page,NULL);
mi_assert_internal(start == _mi_segment_page_start(segment,page,NULL));
//const size_t bsize = mi_page_block_size(page);
uint8_t* start = mi_page_start(page);
mi_assert_internal(start == _mi_segment_page_start(_mi_page_segment(page), page, NULL));
mi_assert_internal(page->is_huge == (_mi_page_segment(page)->kind == MI_SEGMENT_HUGE));
//mi_assert_internal(start + page->capacity*page->block_size == page->top);
mi_assert_internal(mi_page_list_is_valid(page,page->free));
@ -193,8 +192,8 @@ static void _mi_page_thread_free_collect(mi_page_t* page)
if (head == NULL) return;
// find the tail -- also to get a proper count (without data races)
uint32_t max_count = page->capacity; // cannot collect more than capacity
uint32_t count = 1;
size_t max_count = page->capacity; // cannot collect more than capacity
size_t count = 1;
mi_block_t* tail = head;
mi_block_t* next;
while ((next = mi_block_next(page,tail)) != NULL && count <= max_count) {
@ -212,7 +211,7 @@ static void _mi_page_thread_free_collect(mi_page_t* page)
page->local_free = head;
// update counts now
page->used -= count;
page->used -= (uint16_t)count;
}
void _mi_page_free_collect(mi_page_t* page, bool force) {
@ -282,11 +281,13 @@ static mi_page_t* mi_page_fresh_alloc(mi_heap_t* heap, mi_page_queue_t* pq, size
// this may be out-of-memory, or an abandoned page was reclaimed (and in our queue)
return NULL;
}
#if MI_HUGE_PAGE_ABANDON
mi_assert_internal(pq==NULL || _mi_page_segment(page)->page_kind != MI_PAGE_HUGE);
#endif
mi_assert_internal(page_alignment >0 || block_size > MI_MEDIUM_OBJ_SIZE_MAX || _mi_page_segment(page)->kind != MI_SEGMENT_HUGE);
mi_assert_internal(pq!=NULL || page->xblock_size != 0);
mi_assert_internal(pq!=NULL || mi_page_block_size(page) >= block_size);
// a fresh page was found, initialize it
const size_t full_block_size = ((pq == NULL || mi_page_queue_is_huge(pq)) ? mi_page_block_size(page) : block_size); // see also: mi_segment_huge_page_alloc
const size_t full_block_size = (pq == NULL || mi_page_is_huge(page) ? mi_page_block_size(page) : block_size); // see also: mi_segment_huge_page_alloc
mi_assert_internal(full_block_size >= block_size);
mi_page_init(heap, page, full_block_size, heap->tld);
mi_heap_stat_increase(heap, pages, 1);
@ -427,8 +428,7 @@ void _mi_page_free(mi_page_t* page, mi_page_queue_t* pq, bool force) {
_mi_segment_page_free(page, force, segments_tld);
}
// Retire parameters
#define MI_MAX_RETIRE_SIZE (MI_MEDIUM_OBJ_SIZE_MAX)
#define MI_MAX_RETIRE_SIZE MI_MEDIUM_OBJ_SIZE_MAX // should be less than size for MI_BIN_HUGE
#define MI_RETIRE_CYCLES (16)
// Retire a page with no more used blocks
@ -451,10 +451,11 @@ void _mi_page_retire(mi_page_t* page) mi_attr_noexcept {
// how to check this efficiently though...
// for now, we don't retire if it is the only page left of this size class.
mi_page_queue_t* pq = mi_page_queue_of(page);
if mi_likely(page->xblock_size <= MI_MAX_RETIRE_SIZE && !mi_page_queue_is_special(pq)) { // not too large && not full or huge queue?
const size_t bsize = mi_page_block_size(page);
if mi_likely( /* bsize < MI_MAX_RETIRE_SIZE && */ !mi_page_queue_is_special(pq)) { // not full or huge queue?
if (pq->last==page && pq->first==page) { // the only page in the queue?
mi_stat_counter_increase(_mi_stats_main.page_no_retire,1);
page->retire_expire = 1 + (page->xblock_size <= MI_SMALL_OBJ_SIZE_MAX ? MI_RETIRE_CYCLES : MI_RETIRE_CYCLES/4);
page->retire_expire = (bsize <= MI_SMALL_OBJ_SIZE_MAX ? MI_RETIRE_CYCLES : MI_RETIRE_CYCLES/4);
mi_heap_t* heap = mi_page_heap(page);
mi_assert_internal(pq >= heap->pages);
const size_t index = pq - heap->pages;
@ -462,7 +463,7 @@ void _mi_page_retire(mi_page_t* page) mi_attr_noexcept {
if (index < heap->page_retired_min) heap->page_retired_min = index;
if (index > heap->page_retired_max) heap->page_retired_max = index;
mi_assert_internal(mi_page_all_free(page));
return; // dont't free after all
return; // don't free after all
}
}
_mi_page_free(page, pq, false);
@ -516,7 +517,7 @@ static void mi_page_free_list_extend_secure(mi_heap_t* const heap, mi_page_t* co
#endif
mi_assert_internal(page->capacity + extend <= page->reserved);
mi_assert_internal(bsize == mi_page_block_size(page));
void* const page_area = _mi_page_start(_mi_page_segment(page), page, NULL);
void* const page_area = mi_page_start(page);
// initialize a randomized free list
// set up `slice_count` slices to alternate between
@ -574,7 +575,7 @@ static mi_decl_noinline void mi_page_free_list_extend( mi_page_t* const page, co
#endif
mi_assert_internal(page->capacity + extend <= page->reserved);
mi_assert_internal(bsize == mi_page_block_size(page));
void* const page_area = _mi_page_start(_mi_page_segment(page), page, NULL );
void* const page_area = mi_page_start(page);
mi_block_t* const start = mi_page_block_at(page, page_area, bsize, page->capacity);
@ -617,16 +618,14 @@ static void mi_page_extend_free(mi_heap_t* heap, mi_page_t* page, mi_tld_t* tld)
#endif
if (page->capacity >= page->reserved) return;
size_t page_size;
_mi_page_start(_mi_page_segment(page), page, &page_size);
mi_stat_counter_increase(tld->stats.pages_extended, 1);
// calculate the extend count
const size_t bsize = (page->xblock_size < MI_HUGE_BLOCK_SIZE ? page->xblock_size : page_size);
const size_t bsize = mi_page_block_size(page);
size_t extend = page->reserved - page->capacity;
mi_assert_internal(extend > 0);
size_t max_extend = (bsize >= MI_MAX_EXTEND_SIZE ? MI_MIN_EXTEND : MI_MAX_EXTEND_SIZE/(uint32_t)bsize);
size_t max_extend = (bsize >= MI_MAX_EXTEND_SIZE ? MI_MIN_EXTEND : MI_MAX_EXTEND_SIZE/bsize);
if (max_extend < MI_MIN_EXTEND) { max_extend = MI_MIN_EXTEND; }
mi_assert_internal(max_extend > 0);
@ -660,11 +659,10 @@ static void mi_page_init(mi_heap_t* heap, mi_page_t* page, size_t block_size, mi
mi_assert_internal(block_size > 0);
// set fields
mi_page_set_heap(page, heap);
page->xblock_size = (block_size < MI_HUGE_BLOCK_SIZE ? (uint32_t)block_size : MI_HUGE_BLOCK_SIZE); // initialize before _mi_segment_page_start
page->block_size = block_size;
size_t page_size;
const void* page_start = _mi_segment_page_start(segment, page, &page_size);
MI_UNUSED(page_start);
mi_track_mem_noaccess(page_start,page_size);
page->page_start = _mi_segment_page_start(segment, page, &page_size);
mi_track_mem_noaccess(page->page_start,page_size);
mi_assert_internal(mi_page_block_size(page) <= page_size);
mi_assert_internal(page_size <= page->slice_count*MI_SEGMENT_SLICE_SIZE);
mi_assert_internal(page_size / block_size < (1L<<16));
@ -677,12 +675,18 @@ static void mi_page_init(mi_heap_t* heap, mi_page_t* page, size_t block_size, mi
page->free_is_zero = page->is_zero_init;
#if MI_DEBUG>2
if (page->is_zero_init) {
mi_track_mem_defined(page_start, page_size);
mi_assert_expensive(mi_mem_is_zero(page_start, page_size));
mi_track_mem_defined(page->page_start, page_size);
mi_assert_expensive(mi_mem_is_zero(page->page_start, page_size));
}
#endif
mi_assert_internal(page->is_committed);
if (block_size > 0 && _mi_is_power_of_two(block_size)) {
page->block_size_shift = (uint8_t)(mi_ctz((uintptr_t)block_size));
}
else {
page->block_size_shift = 0;
}
mi_assert_internal(page->capacity == 0);
mi_assert_internal(page->free == NULL);
mi_assert_internal(page->used == 0);
@ -695,6 +699,7 @@ static void mi_page_init(mi_heap_t* heap, mi_page_t* page, size_t block_size, mi
mi_assert_internal(page->keys[0] != 0);
mi_assert_internal(page->keys[1] != 0);
#endif
mi_assert_internal(page->block_size_shift == 0 || (block_size == ((size_t)1 << page->block_size_shift)));
mi_assert_expensive(mi_page_is_valid_init(page));
// initialize an initial free list
@ -820,11 +825,9 @@ void mi_register_deferred_free(mi_deferred_free_fun* fn, void* arg) mi_attr_noex
----------------------------------------------------------- */
// Large and huge page allocation.
// Huge pages are allocated directly without being in a queue.
// Because huge pages contain just one block, and the segment contains
// just that page, we always treat them as abandoned and any thread
// that frees the block can free the whole page and segment directly.
// Huge pages are also use if the requested alignment is very large (> MI_ALIGNMENT_MAX).
// Huge pages contain just one block, and the segment contains just that page (as `MI_SEGMENT_HUGE`).
// Huge pages are also use if the requested alignment is very large (> MI_BLOCK_ALIGNMENT_MAX)
// so their size is not always `> MI_LARGE_OBJ_SIZE_MAX`.
static mi_page_t* mi_large_huge_page_alloc(mi_heap_t* heap, size_t size, size_t page_alignment) {
size_t block_size = _mi_os_good_alloc_size(size);
mi_assert_internal(mi_bin(block_size) == MI_BIN_HUGE || page_alignment > 0);
@ -832,7 +835,7 @@ static mi_page_t* mi_large_huge_page_alloc(mi_heap_t* heap, size_t size, size_t
#if MI_HUGE_PAGE_ABANDON
mi_page_queue_t* pq = (is_huge ? NULL : mi_page_queue(heap, block_size));
#else
mi_page_queue_t* pq = mi_page_queue(heap, is_huge ? MI_HUGE_BLOCK_SIZE : block_size); // not block_size as that can be low if the page_alignment > 0
mi_page_queue_t* pq = mi_page_queue(heap, is_huge ? MI_LARGE_OBJ_SIZE_MAX+1 : block_size);
mi_assert_internal(!is_huge || mi_page_queue_is_huge(pq));
#endif
mi_page_t* page = mi_page_fresh_alloc(heap, pq, block_size, page_alignment);
@ -840,6 +843,7 @@ static mi_page_t* mi_large_huge_page_alloc(mi_heap_t* heap, size_t size, size_t
mi_assert_internal(mi_page_immediate_available(page));
if (is_huge) {
mi_assert_internal(mi_page_is_huge(page));
mi_assert_internal(_mi_page_segment(page)->kind == MI_SEGMENT_HUGE);
mi_assert_internal(_mi_page_segment(page)->used==1);
#if MI_HUGE_PAGE_ABANDON
@ -848,7 +852,7 @@ static mi_page_t* mi_large_huge_page_alloc(mi_heap_t* heap, size_t size, size_t
#endif
}
else {
mi_assert_internal(_mi_page_segment(page)->kind != MI_SEGMENT_HUGE);
mi_assert_internal(!mi_page_is_huge(page));
}
const size_t bsize = mi_page_usable_block_size(page); // note: not `mi_page_block_size` to account for padding
@ -871,7 +875,7 @@ static mi_page_t* mi_find_page(mi_heap_t* heap, size_t size, size_t huge_alignme
// huge allocation?
const size_t req_size = size - MI_PADDING_SIZE; // correct for padding_size in case of an overflow on `size`
if mi_unlikely(req_size > (MI_MEDIUM_OBJ_SIZE_MAX - MI_PADDING_SIZE) || huge_alignment > 0) {
if mi_unlikely(req_size > PTRDIFF_MAX) { // we don't allocate more than PTRDIFF_MAX (see <https://sourceware.org/ml/libc-announce/2019/msg00001.html>)
if mi_unlikely(req_size > MI_MAX_ALLOC_SIZE) {
_mi_error_message(EOVERFLOW, "allocation request is too large (%zu bytes)\n", req_size);
return NULL;
}
@ -926,7 +930,7 @@ void* _mi_malloc_generic(mi_heap_t* heap, size_t size, bool zero, size_t huge_al
mi_assert_internal(mi_page_block_size(page) >= size);
// and try again, this time succeeding! (i.e. this should never recurse through _mi_page_malloc)
if mi_unlikely(zero && page->xblock_size == 0) {
if mi_unlikely(zero && page->block_size == 0) {
// note: we cannot call _mi_page_malloc with zeroing for huge blocks; we zero it afterwards in that case.
void* p = _mi_page_malloc(heap, page, size, false);
mi_assert_internal(p != NULL);

251
src/prim/emscripten/prim.c Normal file
View File

@ -0,0 +1,251 @@
/* ----------------------------------------------------------------------------
Copyright (c) 2018-2023, Microsoft Research, Daan Leijen, Alon Zakai
This is free software; you can redistribute it and/or modify it under the
terms of the MIT license. A copy of the license can be found in the file
"LICENSE" at the root of this distribution.
-----------------------------------------------------------------------------*/
// This file is included in `src/prim/prim.c`
#include "mimalloc.h"
#include "mimalloc/internal.h"
#include "mimalloc/atomic.h"
#include "mimalloc/prim.h"
// Design
// ======
//
// mimalloc is built on top of emmalloc. emmalloc is a minimal allocator on top
// of sbrk. The reason for having three layers here is that we want mimalloc to
// be able to allocate and release system memory properly, the same way it would
// when using VirtualAlloc on Windows or mmap on POSIX, and sbrk is too limited.
// Specifically, sbrk can only go up and down, and not "skip" over regions, and
// so we end up either never freeing memory to the system, or we can get stuck
// with holes.
//
// Atm wasm generally does *not* free memory back the system: once grown, we do
// not shrink back down (https://github.com/WebAssembly/design/issues/1397).
// However, that is expected to improve
// (https://github.com/WebAssembly/memory-control/blob/main/proposals/memory-control/Overview.md)
// and so we do not want to bake those limitations in here.
//
// Even without that issue, we want our system allocator to handle holes, that
// is, it should merge freed regions and allow allocating new content there of
// the full size, etc., so that we do not waste space. That means that the
// system allocator really does need to handle the general problem of allocating
// and freeing variable-sized chunks of memory in a random order, like malloc/
// free do. And so it makes sense to layer mimalloc on top of such an
// implementation.
//
// emmalloc makes sense for the lower level because it is small and simple while
// still fully handling merging of holes etc. It is not the most efficient
// allocator, but our assumption is that mimalloc needs to be fast while the
// system allocator underneath it is called much less frequently.
//
//---------------------------------------------
// init
//---------------------------------------------
void _mi_prim_mem_init( mi_os_mem_config_t* config) {
config->page_size = 64*MI_KiB; // WebAssembly has a fixed page size: 64KiB
config->alloc_granularity = 16;
config->has_overcommit = false;
config->must_free_whole = true;
config->has_virtual_reserve = false;
}
extern void emmalloc_free(void*);
int _mi_prim_free(void* addr, size_t size) {
MI_UNUSED(size);
emmalloc_free(addr);
return 0;
}
//---------------------------------------------
// Allocation
//---------------------------------------------
extern void* emmalloc_memalign(size_t, size_t);
// Note: the `try_alignment` is just a hint and the returned pointer is not guaranteed to be aligned.
int _mi_prim_alloc(size_t size, size_t try_alignment, bool commit, bool allow_large, bool* is_large, bool* is_zero, void** addr) {
MI_UNUSED(try_alignment); MI_UNUSED(allow_large); MI_UNUSED(commit);
*is_large = false;
// TODO: Track the highest address ever seen; first uses of it are zeroes.
// That assumes no one else uses sbrk but us (they could go up,
// scribble, and then down), but we could assert on that perhaps.
*is_zero = false;
// emmalloc has some limitations on alignment size.
// TODO: Why does mimalloc ask for an align of 4MB? that ends up allocating
// 8, which wastes quite a lot for us in wasm. If that is unavoidable,
// we may want to improve emmalloc to support such alignment. See also
// https://github.com/emscripten-core/emscripten/issues/20645
#define MIN_EMMALLOC_ALIGN 8
#define MAX_EMMALLOC_ALIGN (1024*1024)
if (try_alignment < MIN_EMMALLOC_ALIGN) {
try_alignment = MIN_EMMALLOC_ALIGN;
} else if (try_alignment > MAX_EMMALLOC_ALIGN) {
try_alignment = MAX_EMMALLOC_ALIGN;
}
void* p = emmalloc_memalign(try_alignment, size);
*addr = p;
if (p == 0) {
return ENOMEM;
}
return 0;
}
//---------------------------------------------
// Commit/Reset
//---------------------------------------------
int _mi_prim_commit(void* addr, size_t size, bool* is_zero) {
MI_UNUSED(addr); MI_UNUSED(size);
// See TODO above.
*is_zero = false;
return 0;
}
int _mi_prim_decommit(void* addr, size_t size, bool* needs_recommit) {
MI_UNUSED(addr); MI_UNUSED(size);
*needs_recommit = false;
return 0;
}
int _mi_prim_reset(void* addr, size_t size) {
MI_UNUSED(addr); MI_UNUSED(size);
return 0;
}
int _mi_prim_protect(void* addr, size_t size, bool protect) {
MI_UNUSED(addr); MI_UNUSED(size); MI_UNUSED(protect);
return 0;
}
//---------------------------------------------
// Huge pages and NUMA nodes
//---------------------------------------------
int _mi_prim_alloc_huge_os_pages(void* hint_addr, size_t size, int numa_node, bool* is_zero, void** addr) {
MI_UNUSED(hint_addr); MI_UNUSED(size); MI_UNUSED(numa_node);
*is_zero = true;
*addr = NULL;
return ENOSYS;
}
size_t _mi_prim_numa_node(void) {
return 0;
}
size_t _mi_prim_numa_node_count(void) {
return 1;
}
//----------------------------------------------------------------
// Clock
//----------------------------------------------------------------
#include <emscripten/html5.h>
mi_msecs_t _mi_prim_clock_now(void) {
return emscripten_date_now();
}
//----------------------------------------------------------------
// Process info
//----------------------------------------------------------------
void _mi_prim_process_info(mi_process_info_t* pinfo)
{
// use defaults
MI_UNUSED(pinfo);
}
//----------------------------------------------------------------
// Output
//----------------------------------------------------------------
#include <emscripten/console.h>
void _mi_prim_out_stderr( const char* msg) {
emscripten_console_error(msg);
}
//----------------------------------------------------------------
// Environment
//----------------------------------------------------------------
bool _mi_prim_getenv(const char* name, char* result, size_t result_size) {
// For code size reasons, do not support environ customization for now.
MI_UNUSED(name);
MI_UNUSED(result);
MI_UNUSED(result_size);
return false;
}
//----------------------------------------------------------------
// Random
//----------------------------------------------------------------
bool _mi_prim_random_buf(void* buf, size_t buf_len) {
int err = getentropy(buf, buf_len);
return !err;
}
//----------------------------------------------------------------
// Thread init/done
//----------------------------------------------------------------
#ifdef __EMSCRIPTEN_SHARED_MEMORY__
// use pthread local storage keys to detect thread ending
// (and used with MI_TLS_PTHREADS for the default heap)
pthread_key_t _mi_heap_default_key = (pthread_key_t)(-1);
static void mi_pthread_done(void* value) {
if (value!=NULL) {
_mi_thread_done((mi_heap_t*)value);
}
}
void _mi_prim_thread_init_auto_done(void) {
mi_assert_internal(_mi_heap_default_key == (pthread_key_t)(-1));
pthread_key_create(&_mi_heap_default_key, &mi_pthread_done);
}
void _mi_prim_thread_done_auto_done(void) {
// nothing to do
}
void _mi_prim_thread_associate_default_heap(mi_heap_t* heap) {
if (_mi_heap_default_key != (pthread_key_t)(-1)) { // can happen during recursive invocation on freeBSD
pthread_setspecific(_mi_heap_default_key, heap);
}
}
#else
void _mi_prim_thread_init_auto_done(void) {
// nothing
}
void _mi_prim_thread_done_auto_done(void) {
// nothing
}
void _mi_prim_thread_associate_default_heap(mi_heap_t* heap) {
MI_UNUSED(heap);
}
#endif

2
src/prim/osx/alloc-override-zone.c
View File

@ -225,7 +225,9 @@ static malloc_zone_t mi_malloc_zone = {
// switch to version 9+ on OSX 10.6 to support memalign.
.memalign = &zone_memalign,
.free_definite_size = &zone_free_definite_size,
#if defined(MAC_OS_X_VERSION_10_7) && (MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_7)
.pressure_relief = &zone_pressure_relief,
#endif
#if defined(MAC_OS_X_VERSION_10_14) && (MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_14)
.claimed_address = &zone_claimed_address,
#endif

3
src/prim/prim.c
View File

@ -18,6 +18,9 @@ terms of the MIT license. A copy of the license can be found in the file
#define MI_USE_SBRK
#include "wasi/prim.c" // memory-grow or sbrk (Wasm)
#elif defined(__EMSCRIPTEN__)
#include "emscripten/prim.c" // emmalloc_*, + pthread support
#else
#include "unix/prim.c" // mmap() (Linux, macOSX, BSD, Illumnos, Haiku, DragonFly, etc.)

72
src/prim/unix/prim.c
View File

@ -27,19 +27,26 @@ terms of the MIT license. A copy of the license can be found in the file
#include <sys/mman.h> // mmap
#include <unistd.h> // sysconf
#include <fcntl.h> // open, close, read, access
#if defined(__linux__)
#include <features.h>
#include <fcntl.h>
#if defined(MI_NO_THP)
#include <sys/prctl.h>
#endif
#if defined(__GLIBC__)
#include <linux/mman.h> // linux mmap flags
#else
#include <sys/mman.h>
#endif
#elif defined(__APPLE__)
#include <AvailabilityMacros.h>
#include <TargetConditionals.h>
#if !TARGET_IOS_IPHONE && !TARGET_IOS_SIMULATOR
#include <mach/vm_statistics.h>
#if !defined(TARGET_OS_OSX) || TARGET_OS_OSX // see issue #879, used to be (!TARGET_IOS_IPHONE && !TARGET_IOS_SIMULATOR)
#include <mach/vm_statistics.h> // VM_MAKE_TAG, VM_FLAGS_SUPERPAGE_SIZE_2MB, etc.
#endif
#if !defined(MAC_OS_X_VERSION_10_7)
#define MAC_OS_X_VERSION_10_7 1070
#endif
#elif defined(__FreeBSD__) || defined(__DragonFly__)
#include <sys/param.h>
@ -50,17 +57,19 @@ terms of the MIT license. A copy of the license can be found in the file
#include <sys/sysctl.h>
#endif
#if !defined(__HAIKU__) && !defined(__APPLE__) && !defined(__CYGWIN__)
#if !defined(__HAIKU__) && !defined(__APPLE__) && !defined(__CYGWIN__) && !defined(__OpenBSD__) && !defined(__sun)
#define MI_HAS_SYSCALL_H
#include <sys/syscall.h>
#endif
//------------------------------------------------------------------------------------
// Use syscalls for some primitives to allow for libraries that override open/read/close etc.
// and do allocation themselves; using syscalls prevents recursion when mimalloc is
// still initializing (issue #713)
//------------------------------------------------------------------------------------
#if defined(MI_HAS_SYSCALL_H) && defined(SYS_open) && defined(SYS_close) && defined(SYS_read) && defined(SYS_access)
static int mi_prim_open(const char* fpath, int open_flags) {
@ -76,7 +85,8 @@ static int mi_prim_access(const char *fpath, int mode) {
return syscall(SYS_access,fpath,mode);
}
#elif !defined(__APPLE__) // avoid unused warnings
#elif !defined(__sun) && \
(!defined(__APPLE__) || (MAC_OS_X_VERSION_MIN_REQUIRED < MAC_OS_X_VERSION_10_7)) // avoid unused warnings on macOS and Solaris
static int mi_prim_open(const char* fpath, int open_flags) {
return open(fpath,open_flags);
@ -125,7 +135,8 @@ static bool unix_detect_overcommit(void) {
return os_overcommit;
}
void _mi_prim_mem_init( mi_os_mem_config_t* config ) {
void _mi_prim_mem_init( mi_os_mem_config_t* config )
{
long psize = sysconf(_SC_PAGESIZE);
if (psize > 0) {
config->page_size = (size_t)psize;
@ -135,6 +146,24 @@ void _mi_prim_mem_init( mi_os_mem_config_t* config ) {
config->has_overcommit = unix_detect_overcommit();
config->must_free_whole = false; // mmap can free in parts
config->has_virtual_reserve = true; // todo: check if this true for NetBSD? (for anonymous mmap with PROT_NONE)
// disable transparent huge pages for this process?
#if (defined(__linux__) || defined(__ANDROID__)) && defined(PR_GET_THP_DISABLE)
#if defined(MI_NO_THP)
if (true)
#else
if (!mi_option_is_enabled(mi_option_allow_large_os_pages)) // disable THP also if large OS pages are not allowed in the options
#endif
{
int val = 0;
if (prctl(PR_GET_THP_DISABLE, &val, 0, 0, 0) != 0) {
// Most likely since distros often come with always/madvise settings.
val = 1;
// Disabling only for mimalloc process rather than touching system wide settings
(void)prctl(PR_SET_THP_DISABLE, &val, 0, 0, 0);
}
}
#endif
}
@ -276,7 +305,7 @@ static void* unix_mmap(void* addr, size_t size, size_t try_alignment, int protec
*is_large = true;
p = unix_mmap_prim(addr, size, try_alignment, protect_flags, lflags, lfd);
#ifdef MAP_HUGE_1GB
if (p == NULL && (lflags & MAP_HUGE_1GB) != 0) {
if (p == NULL && (lflags & MAP_HUGE_1GB) == MAP_HUGE_1GB) {
mi_huge_pages_available = false; // don't try huge 1GiB pages again
_mi_warning_message("unable to allocate huge (1GiB) page, trying large (2MiB) pages instead (errno: %i)\n", errno);
lflags = ((lflags & ~MAP_HUGE_1GB) | MAP_HUGE_2MB);
@ -310,7 +339,7 @@ static void* unix_mmap(void* addr, size_t size, size_t try_alignment, int protec
#elif defined(__sun)
if (allow_large && _mi_os_use_large_page(size, try_alignment)) {
struct memcntl_mha cmd = {0};
cmd.mha_pagesize = large_os_page_size;
cmd.mha_pagesize = _mi_os_large_page_size();
cmd.mha_cmd = MHA_MAPSIZE_VA;
if (memcntl((caddr_t)p, size, MC_HAT_ADVISE, (caddr_t)&cmd, 0, 0) == 0) {
*is_large = true;
@ -472,8 +501,6 @@ int _mi_prim_alloc_huge_os_pages(void* hint_addr, size_t size, int numa_node, bo
#if defined(__linux__)
#include <stdio.h> // snprintf
size_t _mi_prim_numa_node(void) {
#if defined(MI_HAS_SYSCALL_H) && defined(SYS_getcpu)
unsigned long node = 0;
@ -491,7 +518,7 @@ size_t _mi_prim_numa_node_count(void) {
unsigned node = 0;
for(node = 0; node < 256; node++) {
// enumerate node entries -- todo: it there a more efficient way to do this? (but ensure there is no allocation)
snprintf(buf, 127, "/sys/devices/system/node/node%u", node + 1);
_mi_snprintf(buf, 127, "/sys/devices/system/node/node%u", node + 1);
if (mi_prim_access(buf,R_OK) != 0) break;
}
return (node+1);
@ -729,28 +756,20 @@ bool _mi_prim_getenv(const char* name, char* result, size_t result_size) {
// Random
//----------------------------------------------------------------
#if defined(__APPLE__)
#include <AvailabilityMacros.h>
#if defined(MAC_OS_X_VERSION_10_10) && MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_10
#if defined(__APPLE__) && defined(MAC_OS_X_VERSION_10_15) && (MAC_OS_X_VERSION_MIN_REQUIRED >= MAC_OS_X_VERSION_10_15)
#include <CommonCrypto/CommonCryptoError.h>
#include <CommonCrypto/CommonRandom.h>
#endif
bool _mi_prim_random_buf(void* buf, size_t buf_len) {
#if defined(MAC_OS_X_VERSION_10_15) && MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_15
// We prefere CCRandomGenerateBytes as it returns an error code while arc4random_buf
// may fail silently on macOS. See PR #390, and <https://opensource.apple.com/source/Libc/Libc-1439.40.11/gen/FreeBSD/arc4random.c.auto.html>
return (CCRandomGenerateBytes(buf, buf_len) == kCCSuccess);
#else
// fall back on older macOS
arc4random_buf(buf, buf_len);
return true;
#endif
}
#elif defined(__ANDROID__) || defined(__DragonFly__) || \
defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || \
defined(__sun)
defined(__sun) || \
(defined(__APPLE__) && (MAC_OS_X_VERSION_MIN_REQUIRED >= MAC_OS_X_VERSION_10_7))
#include <stdlib.h>
bool _mi_prim_random_buf(void* buf, size_t buf_len) {
@ -758,11 +777,10 @@ bool _mi_prim_random_buf(void* buf, size_t buf_len) {
return true;
}
#elif defined(__linux__) || defined(__HAIKU__)
#elif defined(__APPLE__) || defined(__linux__) || defined(__HAIKU__) // also for old apple versions < 10.7 (issue #829)
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <errno.h>
bool _mi_prim_random_buf(void* buf, size_t buf_len) {
@ -833,7 +851,9 @@ void _mi_prim_thread_init_auto_done(void) {
}
void _mi_prim_thread_done_auto_done(void) {
// nothing to do
if (_mi_heap_default_key != (pthread_key_t)(-1)) { // do not leak the key, see issue #809
pthread_key_delete(_mi_heap_default_key);
}
}
void _mi_prim_thread_associate_default_heap(mi_heap_t* heap) {

5
src/prim/wasi/prim.c
View File

@ -12,6 +12,9 @@ terms of the MIT license. A copy of the license can be found in the file
#include "mimalloc/atomic.h"
#include "mimalloc/prim.h"
#include <stdio.h> // fputs
#include <stdlib.h> // getenv
//---------------------------------------------
// Initialize
//---------------------------------------------
@ -40,6 +43,8 @@ int _mi_prim_free(void* addr, size_t size ) {
//---------------------------------------------
#if defined(MI_USE_SBRK)
#include <unistd.h> // for sbrk
static void* mi_memory_grow( size_t size ) {
void* p = sbrk(size);
if (p == (void*)(-1)) return NULL;

4
src/segment-map.c
View File

@ -29,6 +29,7 @@ terms of the MIT license. A copy of the license can be found in the file
static _Atomic(uintptr_t) mi_segment_map[MI_SEGMENT_MAP_WSIZE + 1]; // 2KiB per TB with 64MiB segments
static size_t mi_segment_map_index_of(const mi_segment_t* segment, size_t* bitidx) {
// note: segment can be invalid or NULL.
mi_assert_internal(_mi_ptr_segment(segment + 1) == segment); // is it aligned on MI_SEGMENT_SIZE?
if ((uintptr_t)segment >= MI_MAX_ADDRESS) {
*bitidx = 0;
@ -70,8 +71,7 @@ void _mi_segment_map_freed_at(const mi_segment_t* segment) {
// Determine the segment belonging to a pointer or NULL if it is not in a valid segment.
static mi_segment_t* _mi_segment_of(const void* p) {
if (p == NULL) return NULL;
mi_segment_t* segment = _mi_ptr_segment(p);
mi_assert_internal(segment != NULL);
mi_segment_t* segment = _mi_ptr_segment(p); // segment can be NULL
size_t bitidx;
size_t index = mi_segment_map_index_of(segment, &bitidx);
// fast path: for any pointer to valid small/medium/large object or first MI_SEGMENT_SIZE in huge

379
src/segment.c
View File

@ -1,5 +1,5 @@
/* ----------------------------------------------------------------------------
Copyright (c) 2018-2020, Microsoft Research, Daan Leijen
Copyright (c) 2018-2024, Microsoft Research, Daan Leijen
This is free software; you can redistribute it and/or modify it under the
terms of the MIT license. A copy of the license can be found in the file
"LICENSE" at the root of this distribution.
@ -11,7 +11,11 @@ terms of the MIT license. A copy of the license can be found in the file
#include <string.h> // memset
#include <stdio.h>
#define MI_PAGE_HUGE_ALIGN (256*1024)
// -------------------------------------------------------------------
// Segments
// mimalloc pages reside in segments. See `mi_segment_valid` for invariants.
// -------------------------------------------------------------------
static void mi_segment_try_purge(mi_segment_t* segment, bool force, mi_stats_t* stats);
@ -146,10 +150,6 @@ size_t _mi_commit_mask_next_run(const mi_commit_mask_t* cm, size_t* idx) {
/* --------------------------------------------------------------------------------
Segment allocation
If a thread ends, it "abandons" pages with used blocks
and there is an abandoned segment list whose segments can
be reclaimed by still running threads, much like work-stealing.
-------------------------------------------------------------------------------- */
@ -212,7 +212,7 @@ static void mi_span_queue_push(mi_span_queue_t* sq, mi_slice_t* slice) {
sq->first = slice;
if (slice->next != NULL) slice->next->prev = slice;
else sq->last = slice;
slice->xblock_size = 0; // free
slice->block_size = 0; // free
}
static mi_span_queue_t* mi_span_queue_for(size_t slice_count, mi_segments_tld_t* tld) {
@ -223,7 +223,7 @@ static mi_span_queue_t* mi_span_queue_for(size_t slice_count, mi_segments_tld_t*
}
static void mi_span_queue_delete(mi_span_queue_t* sq, mi_slice_t* slice) {
mi_assert_internal(slice->xblock_size==0 && slice->slice_count>0 && slice->slice_offset==0);
mi_assert_internal(slice->block_size==0 && slice->slice_count>0 && slice->slice_offset==0);
// should work too if the queue does not contain slice (which can happen during reclaim)
if (slice->prev != NULL) slice->prev->next = slice->next;
if (slice == sq->first) sq->first = slice->next;
@ -231,7 +231,7 @@ static void mi_span_queue_delete(mi_span_queue_t* sq, mi_slice_t* slice) {
if (slice == sq->last) sq->last = slice->prev;
slice->prev = NULL;
slice->next = NULL;
slice->xblock_size = 1; // no more free
slice->block_size = 1; // no more free
}
@ -240,7 +240,7 @@ static void mi_span_queue_delete(mi_span_queue_t* sq, mi_slice_t* slice) {
----------------------------------------------------------- */
static bool mi_slice_is_used(const mi_slice_t* slice) {
return (slice->xblock_size > 0);
return (slice->block_size > 0);
}
@ -268,19 +268,20 @@ static bool mi_segment_is_valid(mi_segment_t* segment, mi_segments_tld_t* tld) {
mi_assert_internal(slice->slice_offset == 0);
size_t index = mi_slice_index(slice);
size_t maxindex = (index + slice->slice_count >= segment->slice_entries ? segment->slice_entries : index + slice->slice_count) - 1;
if (mi_slice_is_used(slice)) { // a page in use, we need at least MAX_SLICE_OFFSET valid back offsets
if (mi_slice_is_used(slice)) { // a page in use, we need at least MAX_SLICE_OFFSET_COUNT valid back offsets
used_count++;
for (size_t i = 0; i <= MI_MAX_SLICE_OFFSET && index + i <= maxindex; i++) {
mi_assert_internal(slice->is_huge == (segment->kind == MI_SEGMENT_HUGE));
for (size_t i = 0; i <= MI_MAX_SLICE_OFFSET_COUNT && index + i <= maxindex; i++) {
mi_assert_internal(segment->slices[index + i].slice_offset == i*sizeof(mi_slice_t));
mi_assert_internal(i==0 || segment->slices[index + i].slice_count == 0);
mi_assert_internal(i==0 || segment->slices[index + i].xblock_size == 1);
mi_assert_internal(i==0 || segment->slices[index + i].block_size == 1);
}
// and the last entry as well (for coalescing)
const mi_slice_t* last = slice + slice->slice_count - 1;
if (last > slice && last < mi_segment_slices_end(segment)) {
mi_assert_internal(last->slice_offset == (slice->slice_count-1)*sizeof(mi_slice_t));
mi_assert_internal(last->slice_count == 0);
mi_assert_internal(last->xblock_size == 1);
mi_assert_internal(last->block_size == 1);
}
}
else { // free range of slices; only last slice needs a valid back offset
@ -289,7 +290,7 @@ static bool mi_segment_is_valid(mi_segment_t* segment, mi_segments_tld_t* tld) {
mi_assert_internal((uint8_t*)slice == (uint8_t*)last - last->slice_offset);
}
mi_assert_internal(slice == last || last->slice_count == 0 );
mi_assert_internal(last->xblock_size == 0 || (segment->kind==MI_SEGMENT_HUGE && last->xblock_size==1));
mi_assert_internal(last->block_size == 0 || (segment->kind==MI_SEGMENT_HUGE && last->block_size==1));
if (segment->kind != MI_SEGMENT_HUGE && segment->thread_id != 0) { // segment is not huge or abandoned
sq = mi_span_queue_for(slice->slice_count,tld);
mi_assert_internal(mi_span_queue_contains(sq,slice));
@ -331,8 +332,8 @@ static uint8_t* _mi_segment_page_start_from_slice(const mi_segment_t* segment, c
uint8_t* _mi_segment_page_start(const mi_segment_t* segment, const mi_page_t* page, size_t* page_size)
{
const mi_slice_t* slice = mi_page_to_slice((mi_page_t*)page);
uint8_t* p = _mi_segment_page_start_from_slice(segment, slice, page->xblock_size, page_size);
mi_assert_internal(page->xblock_size > 0 || _mi_ptr_page(p) == page);
uint8_t* p = _mi_segment_page_start_from_slice(segment, slice, mi_page_block_size(page), page_size);
mi_assert_internal(mi_page_block_size(page) > 0 || _mi_ptr_page(p) == page);
mi_assert_internal(_mi_ptr_segment(p) == segment);
return p;
}
@ -380,6 +381,10 @@ static void mi_segment_os_free(mi_segment_t* segment, mi_segments_tld_t* tld) {
segment->thread_id = 0;
_mi_segment_map_freed_at(segment);
mi_segments_track_size(-((long)mi_segment_size(segment)),tld);
if (segment->was_reclaimed) {
tld->reclaim_count--;
segment->was_reclaimed = false;
}
if (MI_SECURE>0) {
// _mi_os_unprotect(segment, mi_segment_size(segment)); // ensure no more guard pages are set
// unprotect the guard pages; we cannot just unprotect the whole segment size as part may be decommitted
@ -399,13 +404,6 @@ static void mi_segment_os_free(mi_segment_t* segment, mi_segments_tld_t* tld) {
_mi_arena_free(segment, mi_segment_size(segment), csize, segment->memid, tld->stats);
}
// called by threads that are terminating
void _mi_segment_thread_collect(mi_segments_tld_t* tld) {
MI_UNUSED(tld);
// nothing to do
}
/* -----------------------------------------------------------
Commit/Decommit ranges
----------------------------------------------------------- */
@ -568,7 +566,7 @@ static void mi_segment_schedule_purge(mi_segment_t* segment, uint8_t* p, size_t
}
static void mi_segment_try_purge(mi_segment_t* segment, bool force, mi_stats_t* stats) {
if (!segment->allow_purge || mi_commit_mask_is_empty(&segment->purge_mask)) return;
if (!segment->allow_purge || segment->purge_expire == 0 || mi_commit_mask_is_empty(&segment->purge_mask)) return;
mi_msecs_t now = _mi_clock_now();
if (!force && now < segment->purge_expire) return;
@ -590,13 +588,18 @@ static void mi_segment_try_purge(mi_segment_t* segment, bool force, mi_stats_t*
mi_assert_internal(mi_commit_mask_is_empty(&segment->purge_mask));
}
// called from `mi_heap_collect_ex`
// this can be called per-page so it is important that try_purge has fast exit path
void _mi_segment_collect(mi_segment_t* segment, bool force, mi_segments_tld_t* tld) {
mi_segment_try_purge(segment, force, tld->stats);
}
/* -----------------------------------------------------------
Span free
----------------------------------------------------------- */
static bool mi_segment_is_abandoned(mi_segment_t* segment) {
return (segment->thread_id == 0);
return (mi_atomic_load_relaxed(&segment->thread_id) == 0);
}
// note: can be called on abandoned segments
@ -616,7 +619,7 @@ static void mi_segment_span_free(mi_segment_t* segment, size_t slice_index, size
mi_slice_t* last = &segment->slices[slice_index + slice_count - 1];
last->slice_count = 0;
last->slice_offset = (uint32_t)(sizeof(mi_page_t)*(slice_count - 1));
last->xblock_size = 0;
last->block_size = 0;
}
// perhaps decommit
@ -626,7 +629,7 @@ static void mi_segment_span_free(mi_segment_t* segment, size_t slice_index, size
// and push it on the free page queue (if it was not a huge page)
if (sq != NULL) mi_span_queue_push( sq, slice );
else slice->xblock_size = 0; // mark huge page as free anyways
else slice->block_size = 0; // mark huge page as free anyways
}
/*
@ -640,7 +643,7 @@ static void mi_segment_span_add_free(mi_slice_t* slice, mi_segments_tld_t* tld)
*/
static void mi_segment_span_remove_from_queue(mi_slice_t* slice, mi_segments_tld_t* tld) {
mi_assert_internal(slice->slice_count > 0 && slice->slice_offset==0 && slice->xblock_size==0);
mi_assert_internal(slice->slice_count > 0 && slice->slice_offset==0 && slice->block_size==0);
mi_assert_internal(_mi_ptr_segment(slice)->kind != MI_SEGMENT_HUGE);
mi_span_queue_t* sq = mi_span_queue_for(slice->slice_count, tld);
mi_span_queue_delete(sq, slice);
@ -649,14 +652,14 @@ static void mi_segment_span_remove_from_queue(mi_slice_t* slice, mi_segments_tld
// note: can be called on abandoned segments
static mi_slice_t* mi_segment_span_free_coalesce(mi_slice_t* slice, mi_segments_tld_t* tld) {
mi_assert_internal(slice != NULL && slice->slice_count > 0 && slice->slice_offset == 0);
mi_segment_t* segment = _mi_ptr_segment(slice);
bool is_abandoned = mi_segment_is_abandoned(segment);
mi_segment_t* const segment = _mi_ptr_segment(slice);
const bool is_abandoned = (segment->thread_id == 0); // mi_segment_is_abandoned(segment);
// for huge pages, just mark as free but don't add to the queues
if (segment->kind == MI_SEGMENT_HUGE) {
// issue #691: segment->used can be 0 if the huge page block was freed while abandoned (reclaim will get here in that case)
mi_assert_internal((segment->used==0 && slice->xblock_size==0) || segment->used == 1); // decreased right after this call in `mi_segment_page_clear`
slice->xblock_size = 0; // mark as free anyways
mi_assert_internal((segment->used==0 && slice->block_size==0) || segment->used == 1); // decreased right after this call in `mi_segment_page_clear`
slice->block_size = 0; // mark as free anyways
// we should mark the last slice `xblock_size=0` now to maintain invariants but we skip it to
// avoid a possible cache miss (and the segment is about to be freed)
return slice;
@ -666,7 +669,7 @@ static mi_slice_t* mi_segment_span_free_coalesce(mi_slice_t* slice, mi_segments_
size_t slice_count = slice->slice_count;
mi_slice_t* next = slice + slice->slice_count;
mi_assert_internal(next <= mi_segment_slices_end(segment));
if (next < mi_segment_slices_end(segment) && next->xblock_size==0) {
if (next < mi_segment_slices_end(segment) && next->block_size==0) {
// free next block -- remove it from free and merge
mi_assert_internal(next->slice_count > 0 && next->slice_offset==0);
slice_count += next->slice_count; // extend
@ -675,7 +678,7 @@ static mi_slice_t* mi_segment_span_free_coalesce(mi_slice_t* slice, mi_segments_
if (slice > segment->slices) {
mi_slice_t* prev = mi_slice_first(slice - 1);
mi_assert_internal(prev >= segment->slices);
if (prev->xblock_size==0) {
if (prev->block_size==0) {
// free previous slice -- remove it from free and merge
mi_assert_internal(prev->slice_count > 0 && prev->slice_offset==0);
slice_count += prev->slice_count;
@ -699,7 +702,7 @@ static mi_slice_t* mi_segment_span_free_coalesce(mi_slice_t* slice, mi_segments_
static mi_page_t* mi_segment_span_allocate(mi_segment_t* segment, size_t slice_index, size_t slice_count, mi_segments_tld_t* tld) {
mi_assert_internal(slice_index < segment->slice_entries);
mi_slice_t* const slice = &segment->slices[slice_index];
mi_assert_internal(slice->xblock_size==0 || slice->xblock_size==1);
mi_assert_internal(slice->block_size==0 || slice->block_size==1);
// commit before changing the slice data
if (!mi_segment_ensure_committed(segment, _mi_segment_page_start_from_slice(segment, slice, 0, NULL), slice_count * MI_SEGMENT_SLICE_SIZE, tld->stats)) {
@ -711,20 +714,20 @@ static mi_page_t* mi_segment_span_allocate(mi_segment_t* segment, size_t slice_i
slice->slice_count = (uint32_t)slice_count;
mi_assert_internal(slice->slice_count == slice_count);
const size_t bsize = slice_count * MI_SEGMENT_SLICE_SIZE;
slice->xblock_size = (uint32_t)(bsize >= MI_HUGE_BLOCK_SIZE ? MI_HUGE_BLOCK_SIZE : bsize);
slice->block_size = bsize;
mi_page_t* page = mi_slice_to_page(slice);
mi_assert_internal(mi_page_block_size(page) == bsize);
// set slice back pointers for the first MI_MAX_SLICE_OFFSET entries
// set slice back pointers for the first MI_MAX_SLICE_OFFSET_COUNT entries
size_t extra = slice_count-1;
if (extra > MI_MAX_SLICE_OFFSET) extra = MI_MAX_SLICE_OFFSET;
if (extra > MI_MAX_SLICE_OFFSET_COUNT) extra = MI_MAX_SLICE_OFFSET_COUNT;
if (slice_index + extra >= segment->slice_entries) extra = segment->slice_entries - slice_index - 1; // huge objects may have more slices than avaiable entries in the segment->slices
mi_slice_t* slice_next = slice + 1;
for (size_t i = 1; i <= extra; i++, slice_next++) {
slice_next->slice_offset = (uint32_t)(sizeof(mi_slice_t)*i);
slice_next->slice_count = 0;
slice_next->xblock_size = 1;
slice_next->block_size = 1;
}
// and also for the last one (if not set already) (the last one is needed for coalescing and for large alignments)
@ -735,11 +738,12 @@ static mi_page_t* mi_segment_span_allocate(mi_segment_t* segment, size_t slice_i
if (last > slice) {
last->slice_offset = (uint32_t)(sizeof(mi_slice_t) * (last - slice));
last->slice_count = 0;
last->xblock_size = 1;
last->block_size = 1;
}
// and initialize the page
page->is_committed = true;
page->is_huge = (segment->kind == MI_SEGMENT_HUGE);
segment->used++;
return page;
}
@ -747,7 +751,7 @@ static mi_page_t* mi_segment_span_allocate(mi_segment_t* segment, size_t slice_i
static void mi_segment_slice_split(mi_segment_t* segment, mi_slice_t* slice, size_t slice_count, mi_segments_tld_t* tld) {
mi_assert_internal(_mi_ptr_segment(slice) == segment);
mi_assert_internal(slice->slice_count >= slice_count);
mi_assert_internal(slice->xblock_size > 0); // no more in free queue
mi_assert_internal(slice->block_size > 0); // no more in free queue
if (slice->slice_count <= slice_count) return;
mi_assert_internal(segment->kind != MI_SEGMENT_HUGE);
size_t next_index = mi_slice_index(slice) + slice_count;
@ -773,7 +777,7 @@ static mi_page_t* mi_segments_page_find_and_allocate(size_t slice_count, mi_aren
if (slice->slice_count > slice_count) {
mi_segment_slice_split(segment, slice, slice_count, tld);
}
mi_assert_internal(slice != NULL && slice->slice_count == slice_count && slice->xblock_size > 0);
mi_assert_internal(slice != NULL && slice->slice_count == slice_count && slice->block_size > 0);
mi_page_t* page = mi_segment_span_allocate(segment, mi_slice_index(slice), slice->slice_count, tld);
if (page == NULL) {
// commit failed; return NULL but first restore the slice
@ -814,6 +818,7 @@ static mi_segment_t* mi_segment_os_alloc( size_t required, size_t page_alignment
const size_t extra = align_offset - info_size;
// recalculate due to potential guard pages
*psegment_slices = mi_segment_calculate_slices(required + extra, ppre_size, pinfo_slices);
mi_assert_internal(*psegment_slices > 0 && *psegment_slices <= UINT32_MAX);
}
const size_t segment_size = (*psegment_slices) * MI_SEGMENT_SLICE_SIZE;
@ -847,7 +852,6 @@ static mi_segment_t* mi_segment_os_alloc( size_t required, size_t page_alignment
segment->commit_mask = commit_mask;
segment->purge_expire = 0;
mi_commit_mask_create_empty(&segment->purge_mask);
mi_atomic_store_ptr_release(mi_segment_t, &segment->abandoned_next, NULL); // tsan
mi_segments_track_size((long)(segment_size), tld);
_mi_segment_map_allocated_at(segment);
@ -864,6 +868,7 @@ static mi_segment_t* mi_segment_alloc(size_t required, size_t page_alignment, mi
size_t info_slices;
size_t pre_size;
size_t segment_slices = mi_segment_calculate_slices(required, &pre_size, &info_slices);
mi_assert_internal(segment_slices > 0 && segment_slices <= UINT32_MAX);
// Commit eagerly only if not the first N lazy segments (to reduce impact of many threads that allocate just a little)
const bool eager_delay = (// !_mi_os_has_overcommit() && // never delay on overcommit systems
@ -951,8 +956,8 @@ static void mi_segment_free(mi_segment_t* segment, bool force, mi_segments_tld_t
while (slice < end) {
mi_assert_internal(slice->slice_count > 0);
mi_assert_internal(slice->slice_offset == 0);
mi_assert_internal(mi_slice_index(slice)==0 || slice->xblock_size == 0); // no more used pages ..
if (slice->xblock_size == 0 && segment->kind != MI_SEGMENT_HUGE) {
mi_assert_internal(mi_slice_index(slice)==0 || slice->block_size == 0); // no more used pages ..
if (slice->block_size == 0 && segment->kind != MI_SEGMENT_HUGE) {
mi_segment_span_remove_from_queue(slice, tld);
}
#if MI_DEBUG>1
@ -978,7 +983,7 @@ static void mi_segment_abandon(mi_segment_t* segment, mi_segments_tld_t* tld);
// note: can be called on abandoned pages
static mi_slice_t* mi_segment_page_clear(mi_page_t* page, mi_segments_tld_t* tld) {
mi_assert_internal(page->xblock_size > 0);
mi_assert_internal(page->block_size > 0);
mi_assert_internal(mi_page_all_free(page));
mi_segment_t* segment = _mi_ptr_segment(page);
mi_assert_internal(segment->used > 0);
@ -990,7 +995,7 @@ static mi_slice_t* mi_segment_page_clear(mi_page_t* page, mi_segments_tld_t* tld
// reset the page memory to reduce memory pressure?
if (segment->allow_decommit && mi_option_is_enabled(mi_option_deprecated_page_reset)) {
size_t psize;
uint8_t* start = _mi_page_start(segment, page, &psize);
uint8_t* start = _mi_segment_page_start(segment, page, &psize);
_mi_os_reset(start, psize, tld->stats);
}
@ -998,7 +1003,7 @@ static mi_slice_t* mi_segment_page_clear(mi_page_t* page, mi_segments_tld_t* tld
page->is_zero_init = false;
ptrdiff_t ofs = offsetof(mi_page_t, capacity);
_mi_memzero((uint8_t*)page + ofs, sizeof(*page) - ofs);
page->xblock_size = 1;
page->block_size = 1;
// and free it
mi_slice_t* slice = mi_segment_span_free_coalesce(mi_page_to_slice(page), tld);
@ -1027,6 +1032,10 @@ void _mi_segment_page_free(mi_page_t* page, bool force, mi_segments_tld_t* tld)
// only abandoned pages; remove from free list and abandon
mi_segment_abandon(segment,tld);
}
else {
// perform delayed purges
mi_segment_try_purge(segment, false /* force? */, tld->stats);
}
}
@ -1036,172 +1045,21 @@ Abandonment
When threads terminate, they can leave segments with
live blocks (reachable through other threads). Such segments
are "abandoned" and will be reclaimed by other threads to
reuse their pages and/or free them eventually
reuse their pages and/or free them eventually. The
`thread_id` of such segments is 0.
We maintain a global list of abandoned segments that are
reclaimed on demand. Since this is shared among threads
the implementation needs to avoid the A-B-A problem on
popping abandoned segments: <https://en.wikipedia.org/wiki/ABA_problem>
We use tagged pointers to avoid accidentally identifying
reused segments, much like stamped references in Java.
Secondly, we maintain a reader counter to avoid resetting
or decommitting segments that have a pending read operation.
When a block is freed in an abandoned segment, the segment
is reclaimed into that thread.
Note: the current implementation is one possible design;
another way might be to keep track of abandoned segments
in the arenas/segment_cache's. This would have the advantage of keeping
all concurrent code in one place and not needing to deal
with ABA issues. The drawback is that it is unclear how to
scan abandoned segments efficiently in that case as they
would be spread among all other segments in the arenas.
Moreover, if threads are looking for a fresh segment, they
will first consider abondoned segments -- these can be found
by scanning the arena memory
(segments outside arena memoryare only reclaimed by a free).
----------------------------------------------------------- */
// Use the bottom 20-bits (on 64-bit) of the aligned segment pointers
// to put in a tag that increments on update to avoid the A-B-A problem.
#define MI_TAGGED_MASK MI_SEGMENT_MASK
typedef uintptr_t mi_tagged_segment_t;
static mi_segment_t* mi_tagged_segment_ptr(mi_tagged_segment_t ts) {
return (mi_segment_t*)(ts & ~MI_TAGGED_MASK);
}
static mi_tagged_segment_t mi_tagged_segment(mi_segment_t* segment, mi_tagged_segment_t ts) {
mi_assert_internal(((uintptr_t)segment & MI_TAGGED_MASK) == 0);
uintptr_t tag = ((ts & MI_TAGGED_MASK) + 1) & MI_TAGGED_MASK;
return ((uintptr_t)segment | tag);
}
// This is a list of visited abandoned pages that were full at the time.
// this list migrates to `abandoned` when that becomes NULL. The use of
// this list reduces contention and the rate at which segments are visited.
static mi_decl_cache_align _Atomic(mi_segment_t*) abandoned_visited; // = NULL
// The abandoned page list (tagged as it supports pop)
static mi_decl_cache_align _Atomic(mi_tagged_segment_t) abandoned; // = NULL
// Maintain these for debug purposes (these counts may be a bit off)
static mi_decl_cache_align _Atomic(size_t) abandoned_count;
static mi_decl_cache_align _Atomic(size_t) abandoned_visited_count;
// We also maintain a count of current readers of the abandoned list
// in order to prevent resetting/decommitting segment memory if it might
// still be read.
static mi_decl_cache_align _Atomic(size_t) abandoned_readers; // = 0
// Push on the visited list
static void mi_abandoned_visited_push(mi_segment_t* segment) {
mi_assert_internal(segment->thread_id == 0);
mi_assert_internal(mi_atomic_load_ptr_relaxed(mi_segment_t,&segment->abandoned_next) == NULL);
mi_assert_internal(segment->next == NULL);
mi_assert_internal(segment->used > 0);
mi_segment_t* anext = mi_atomic_load_ptr_relaxed(mi_segment_t, &abandoned_visited);
do {
mi_atomic_store_ptr_release(mi_segment_t, &segment->abandoned_next, anext);
} while (!mi_atomic_cas_ptr_weak_release(mi_segment_t, &abandoned_visited, &anext, segment));
mi_atomic_increment_relaxed(&abandoned_visited_count);
}
// Move the visited list to the abandoned list.
static bool mi_abandoned_visited_revisit(void)
{
// quick check if the visited list is empty
if (mi_atomic_load_ptr_relaxed(mi_segment_t, &abandoned_visited) == NULL) return false;
// grab the whole visited list
mi_segment_t* first = mi_atomic_exchange_ptr_acq_rel(mi_segment_t, &abandoned_visited, NULL);
if (first == NULL) return false;
// first try to swap directly if the abandoned list happens to be NULL
mi_tagged_segment_t afirst;
mi_tagged_segment_t ts = mi_atomic_load_relaxed(&abandoned);
if (mi_tagged_segment_ptr(ts)==NULL) {
size_t count = mi_atomic_load_relaxed(&abandoned_visited_count);
afirst = mi_tagged_segment(first, ts);
if (mi_atomic_cas_strong_acq_rel(&abandoned, &ts, afirst)) {
mi_atomic_add_relaxed(&abandoned_count, count);
mi_atomic_sub_relaxed(&abandoned_visited_count, count);
return true;
}
}
// find the last element of the visited list: O(n)
mi_segment_t* last = first;
mi_segment_t* next;
while ((next = mi_atomic_load_ptr_relaxed(mi_segment_t, &last->abandoned_next)) != NULL) {
last = next;
}
// and atomically prepend to the abandoned list
// (no need to increase the readers as we don't access the abandoned segments)
mi_tagged_segment_t anext = mi_atomic_load_relaxed(&abandoned);
size_t count;
do {
count = mi_atomic_load_relaxed(&abandoned_visited_count);
mi_atomic_store_ptr_release(mi_segment_t, &last->abandoned_next, mi_tagged_segment_ptr(anext));
afirst = mi_tagged_segment(first, anext);
} while (!mi_atomic_cas_weak_release(&abandoned, &anext, afirst));
mi_atomic_add_relaxed(&abandoned_count, count);
mi_atomic_sub_relaxed(&abandoned_visited_count, count);
return true;
}
// Push on the abandoned list.
static void mi_abandoned_push(mi_segment_t* segment) {
mi_assert_internal(segment->thread_id == 0);
mi_assert_internal(mi_atomic_load_ptr_relaxed(mi_segment_t, &segment->abandoned_next) == NULL);
mi_assert_internal(segment->next == NULL);
mi_assert_internal(segment->used > 0);
mi_tagged_segment_t next;
mi_tagged_segment_t ts = mi_atomic_load_relaxed(&abandoned);
do {
mi_atomic_store_ptr_release(mi_segment_t, &segment->abandoned_next, mi_tagged_segment_ptr(ts));
next = mi_tagged_segment(segment, ts);
} while (!mi_atomic_cas_weak_release(&abandoned, &ts, next));
mi_atomic_increment_relaxed(&abandoned_count);
}
// Wait until there are no more pending reads on segments that used to be in the abandoned list
// called for example from `arena.c` before decommitting
// legacy: Wait until there are no more pending reads on segments that used to be in the abandoned list
void _mi_abandoned_await_readers(void) {
size_t n;
do {
n = mi_atomic_load_acquire(&abandoned_readers);
if (n != 0) mi_atomic_yield();
} while (n != 0);
}
// Pop from the abandoned list
static mi_segment_t* mi_abandoned_pop(void) {
mi_segment_t* segment;
// Check efficiently if it is empty (or if the visited list needs to be moved)
mi_tagged_segment_t ts = mi_atomic_load_relaxed(&abandoned);
segment = mi_tagged_segment_ptr(ts);
if mi_likely(segment == NULL) {
if mi_likely(!mi_abandoned_visited_revisit()) { // try to swap in the visited list on NULL
return NULL;
}
}
// Do a pop. We use a reader count to prevent
// a segment to be decommitted while a read is still pending,
// and a tagged pointer to prevent A-B-A link corruption.
// (this is called from `region.c:_mi_mem_free` for example)
mi_atomic_increment_relaxed(&abandoned_readers); // ensure no segment gets decommitted
mi_tagged_segment_t next = 0;
ts = mi_atomic_load_acquire(&abandoned);
do {
segment = mi_tagged_segment_ptr(ts);
if (segment != NULL) {
mi_segment_t* anext = mi_atomic_load_ptr_relaxed(mi_segment_t, &segment->abandoned_next);
next = mi_tagged_segment(anext, ts); // note: reads the segment's `abandoned_next` field so should not be decommitted
}
} while (segment != NULL && !mi_atomic_cas_weak_acq_rel(&abandoned, &ts, next));
mi_atomic_decrement_relaxed(&abandoned_readers); // release reader lock
if (segment != NULL) {
mi_atomic_store_ptr_release(mi_segment_t, &segment->abandoned_next, NULL);
mi_atomic_decrement_relaxed(&abandoned_count);
}
return segment;
// nothing needed
}
/* -----------------------------------------------------------
@ -1211,7 +1069,6 @@ static mi_segment_t* mi_abandoned_pop(void) {
static void mi_segment_abandon(mi_segment_t* segment, mi_segments_tld_t* tld) {
mi_assert_internal(segment->used == segment->abandoned);
mi_assert_internal(segment->used > 0);
mi_assert_internal(mi_atomic_load_ptr_relaxed(mi_segment_t, &segment->abandoned_next) == NULL);
mi_assert_internal(segment->abandoned_visits == 0);
mi_assert_expensive(mi_segment_is_valid(segment,tld));
@ -1221,23 +1078,29 @@ static void mi_segment_abandon(mi_segment_t* segment, mi_segments_tld_t* tld) {
while (slice < end) {
mi_assert_internal(slice->slice_count > 0);
mi_assert_internal(slice->slice_offset == 0);
if (slice->xblock_size == 0) { // a free page
if (slice->block_size == 0) { // a free page
mi_segment_span_remove_from_queue(slice,tld);
slice->xblock_size = 0; // but keep it free
slice->block_size = 0; // but keep it free
}
slice = slice + slice->slice_count;
}
// perform delayed decommits (forcing is much slower on mstress)
mi_segment_try_purge(segment, mi_option_is_enabled(mi_option_abandoned_page_purge) /* force? */, tld->stats);
// Only abandoned segments in arena memory can be reclaimed without a free
// so if a segment is not from an arena we force purge here to be conservative.
const bool force_purge = (segment->memid.memkind != MI_MEM_ARENA) || mi_option_is_enabled(mi_option_abandoned_page_purge);
mi_segment_try_purge(segment, force_purge, tld->stats);
// all pages in the segment are abandoned; add it to the abandoned list
_mi_stat_increase(&tld->stats->segments_abandoned, 1);
mi_segments_track_size(-((long)mi_segment_size(segment)), tld);
segment->thread_id = 0;
mi_atomic_store_ptr_release(mi_segment_t, &segment->abandoned_next, NULL);
segment->abandoned_visits = 1; // from 0 to 1 to signify it is abandoned
mi_abandoned_push(segment);
if (segment->was_reclaimed) {
tld->reclaim_count--;
segment->was_reclaimed = false;
}
_mi_arena_segment_mark_abandoned(segment);
}
void _mi_segment_page_abandon(mi_page_t* page, mi_segments_tld_t* tld) {
@ -1264,7 +1127,7 @@ void _mi_segment_page_abandon(mi_page_t* page, mi_segments_tld_t* tld) {
static mi_slice_t* mi_slices_start_iterate(mi_segment_t* segment, const mi_slice_t** end) {
mi_slice_t* slice = &segment->slices[0];
*end = mi_segment_slices_end(segment);
mi_assert_internal(slice->slice_count>0 && slice->xblock_size>0); // segment allocated page
mi_assert_internal(slice->slice_count>0 && slice->block_size>0); // segment allocated page
slice = slice + slice->slice_count; // skip the first segment allocated page
return slice;
}
@ -1272,7 +1135,6 @@ static mi_slice_t* mi_slices_start_iterate(mi_segment_t* segment, const mi_slice
// Possibly free pages and check if free space is available
static bool mi_segment_check_free(mi_segment_t* segment, size_t slices_needed, size_t block_size, mi_segments_tld_t* tld)
{
mi_assert_internal(block_size < MI_HUGE_BLOCK_SIZE);
mi_assert_internal(mi_segment_is_abandoned(segment));
bool has_page = false;
@ -1297,13 +1159,11 @@ static bool mi_segment_check_free(mi_segment_t* segment, size_t slices_needed, s
has_page = true;
}
}
else {
if (page->xblock_size == block_size && mi_page_has_any_available(page)) {
else if (mi_page_block_size(page) == block_size && mi_page_has_any_available(page)) {
// a page has available free blocks of the right size
has_page = true;
}
}
}
else {
// empty span
if (slice->slice_count >= slices_needed) {
@ -1318,12 +1178,13 @@ static bool mi_segment_check_free(mi_segment_t* segment, size_t slices_needed, s
// Reclaim an abandoned segment; returns NULL if the segment was freed
// set `right_page_reclaimed` to `true` if it reclaimed a page of the right `block_size` that was not full.
static mi_segment_t* mi_segment_reclaim(mi_segment_t* segment, mi_heap_t* heap, size_t requested_block_size, bool* right_page_reclaimed, mi_segments_tld_t* tld) {
mi_assert_internal(mi_atomic_load_ptr_relaxed(mi_segment_t, &segment->abandoned_next) == NULL);
mi_assert_expensive(mi_segment_is_valid(segment, tld));
if (right_page_reclaimed != NULL) { *right_page_reclaimed = false; }
segment->thread_id = _mi_thread_id();
// can be 0 still with abandoned_next, or already a thread id for segments outside an arena that are reclaimed on a free.
mi_assert_internal(mi_atomic_load_relaxed(&segment->thread_id) == 0 || mi_atomic_load_relaxed(&segment->thread_id) == _mi_thread_id());
mi_atomic_store_release(&segment->thread_id, _mi_thread_id());
segment->abandoned_visits = 0;
segment->was_reclaimed = true;
tld->reclaim_count++;
mi_segments_track_size((long)mi_segment_size(segment), tld);
mi_assert_internal(segment->next == NULL);
_mi_stat_decrease(&tld->stats->segments_abandoned, 1);
@ -1354,7 +1215,7 @@ static mi_segment_t* mi_segment_reclaim(mi_segment_t* segment, mi_heap_t* heap,
else {
// otherwise reclaim it into the heap
_mi_page_reclaim(heap, page);
if (requested_block_size == page->xblock_size && mi_page_has_any_available(page)) {
if (requested_block_size == mi_page_block_size(page) && mi_page_has_any_available(page)) {
if (right_page_reclaimed != NULL) { *right_page_reclaimed = true; }
}
}
@ -1368,6 +1229,7 @@ static mi_segment_t* mi_segment_reclaim(mi_segment_t* segment, mi_heap_t* heap,
}
mi_assert(segment->abandoned == 0);
mi_assert_expensive(mi_segment_is_valid(segment, tld));
if (segment->used == 0) { // due to page_clear
mi_assert_internal(right_page_reclaimed == NULL || !(*right_page_reclaimed));
mi_segment_free(segment, false, tld);
@ -1378,23 +1240,54 @@ static mi_segment_t* mi_segment_reclaim(mi_segment_t* segment, mi_heap_t* heap,
}
}
// attempt to reclaim a particular segment (called from multi threaded free `alloc.c:mi_free_block_mt`)
bool _mi_segment_attempt_reclaim(mi_heap_t* heap, mi_segment_t* segment) {
if (mi_atomic_load_relaxed(&segment->thread_id) != 0) return false; // it is not abandoned
// don't reclaim more from a free than half the current segments
// this is to prevent a pure free-ing thread to start owning too many segments
if (heap->tld->segments.reclaim_count * 2 > heap->tld->segments.count) return false;
if (_mi_arena_segment_clear_abandoned(segment)) { // atomically unabandon
mi_segment_t* res = mi_segment_reclaim(segment, heap, 0, NULL, &heap->tld->segments);
mi_assert_internal(res == segment);
return (res != NULL);
}
return false;
}
void _mi_abandoned_reclaim_all(mi_heap_t* heap, mi_segments_tld_t* tld) {
mi_segment_t* segment;
while ((segment = mi_abandoned_pop()) != NULL) {
mi_arena_field_cursor_t current; _mi_arena_field_cursor_init(heap, &current);
while ((segment = _mi_arena_segment_clear_abandoned_next(&current)) != NULL) {
mi_segment_reclaim(segment, heap, 0, NULL, tld);
}
}
static long mi_segment_get_reclaim_tries(void) {
// limit the tries to 10% (default) of the abandoned segments with at least 8 and at most 1024 tries.
const size_t perc = (size_t)mi_option_get_clamp(mi_option_max_segment_reclaim, 0, 100);
if (perc <= 0) return 0;
const size_t total_count = _mi_arena_segment_abandoned_count();
if (total_count == 0) return 0;
const size_t relative_count = (total_count > 10000 ? (total_count / 100) * perc : (total_count * perc) / 100); // avoid overflow
long max_tries = (long)(relative_count <= 1 ? 1 : (relative_count > 1024 ? 1024 : relative_count));
if (max_tries < 8 && total_count > 8) { max_tries = 8; }
return max_tries;
}
static mi_segment_t* mi_segment_try_reclaim(mi_heap_t* heap, size_t needed_slices, size_t block_size, bool* reclaimed, mi_segments_tld_t* tld)
{
*reclaimed = false;
long max_tries = mi_segment_get_reclaim_tries();
if (max_tries <= 0) return NULL;
mi_segment_t* segment;
long max_tries = mi_option_get_clamp(mi_option_max_segment_reclaim, 8, 1024); // limit the work to bound allocation times
while ((max_tries-- > 0) && ((segment = mi_abandoned_pop()) != NULL)) {
mi_arena_field_cursor_t current; _mi_arena_field_cursor_init(heap, &current);
while ((max_tries-- > 0) && ((segment = _mi_arena_segment_clear_abandoned_next(&current)) != NULL))
{
segment->abandoned_visits++;
// todo: an arena exclusive heap will potentially visit many abandoned unsuitable segments
// and push them into the visited list and use many tries. Perhaps we can skip non-suitable ones in a better way?
// todo: should we respect numa affinity for abondoned reclaim? perhaps only for the first visit?
// todo: an arena exclusive heap will potentially visit many abandoned unsuitable segments and use many tries
// Perhaps we can skip non-suitable ones in a better way?
bool is_suitable = _mi_heap_memid_is_suitable(heap, segment->memid);
bool has_page = mi_segment_check_free(segment,needed_slices,block_size,tld); // try to free up pages (due to concurrent frees)
if (segment->used == 0) {
@ -1417,8 +1310,8 @@ static mi_segment_t* mi_segment_try_reclaim(mi_heap_t* heap, size_t needed_slice
}
else {
// otherwise, push on the visited list so it gets not looked at too quickly again
mi_segment_try_purge(segment, true /* force? */, tld->stats); // force purge if needed as we may not visit soon again
mi_abandoned_visited_push(segment);
mi_segment_try_purge(segment, false /* true force? */, tld->stats); // force purge if needed as we may not visit soon again
_mi_arena_segment_mark_abandoned(segment);
}
}
return NULL;
@ -1428,11 +1321,9 @@ static mi_segment_t* mi_segment_try_reclaim(mi_heap_t* heap, size_t needed_slice
void _mi_abandoned_collect(mi_heap_t* heap, bool force, mi_segments_tld_t* tld)
{
mi_segment_t* segment;
int max_tries = (force ? 16*1024 : 1024); // limit latency
if (force) {
mi_abandoned_visited_revisit();
}
while ((max_tries-- > 0) && ((segment = mi_abandoned_pop()) != NULL)) {
mi_arena_field_cursor_t current; _mi_arena_field_cursor_init(heap, &current);
long max_tries = (force ? (long)_mi_arena_segment_abandoned_count() : 1024); // limit latency
while ((max_tries-- > 0) && ((segment = _mi_arena_segment_clear_abandoned_next(&current)) != NULL)) {
mi_segment_check_free(segment,0,0,tld); // try to free up pages (due to concurrent frees)
if (segment->used == 0) {
// free the segment (by forced reclaim) to make it available to other threads.
@ -1444,7 +1335,7 @@ void _mi_abandoned_collect(mi_heap_t* heap, bool force, mi_segments_tld_t* tld)
// otherwise, purge if needed and push on the visited list
// note: forced purge can be expensive if many threads are destroyed/created as in mstress.
mi_segment_try_purge(segment, force, tld->stats);
mi_abandoned_visited_push(segment);
_mi_arena_segment_mark_abandoned(segment);
}
}
}
@ -1455,7 +1346,6 @@ void _mi_abandoned_collect(mi_heap_t* heap, bool force, mi_segments_tld_t* tld)
static mi_segment_t* mi_segment_reclaim_or_alloc(mi_heap_t* heap, size_t needed_slices, size_t block_size, mi_segments_tld_t* tld, mi_os_tld_t* os_tld)
{
mi_assert_internal(block_size < MI_HUGE_BLOCK_SIZE);
mi_assert_internal(block_size <= MI_LARGE_OBJ_SIZE_MAX);
// 1. try to reclaim an abandoned segment
@ -1522,11 +1412,12 @@ static mi_page_t* mi_segment_huge_page_alloc(size_t size, size_t page_alignment,
segment->thread_id = 0; // huge segments are immediately abandoned
#endif
// for huge pages we initialize the xblock_size as we may
// for huge pages we initialize the block_size as we may
// overallocate to accommodate large alignments.
size_t psize;
uint8_t* start = _mi_segment_page_start(segment, page, &psize);
page->xblock_size = (psize > MI_HUGE_BLOCK_SIZE ? MI_HUGE_BLOCK_SIZE : (uint32_t)psize);
page->block_size = psize;
mi_assert_internal(page->is_huge);
// decommit the part of the prefix of a page that will not be used; this can be quite large (close to MI_SEGMENT_SIZE)
if (page_alignment > 0 && segment->allow_decommit) {
@ -1557,7 +1448,7 @@ void _mi_segment_huge_page_free(mi_segment_t* segment, mi_page_t* page, mi_block
mi_block_set_next(page, block, page->free);
page->free = block;
page->used--;
page->is_zero = false;
page->is_zero_init = false;
mi_assert(page->used == 0);
mi_tld_t* tld = heap->tld;
_mi_segment_page_free(page, true, &tld->segments);
@ -1593,7 +1484,7 @@ void _mi_segment_huge_page_reset(mi_segment_t* segment, mi_page_t* page, mi_bloc
----------------------------------------------------------- */
mi_page_t* _mi_segment_page_alloc(mi_heap_t* heap, size_t block_size, size_t page_alignment, mi_segments_tld_t* tld, mi_os_tld_t* os_tld) {
mi_page_t* page;
if mi_unlikely(page_alignment > MI_ALIGNMENT_MAX) {
if mi_unlikely(page_alignment > MI_BLOCK_ALIGNMENT_MAX) {
mi_assert_internal(_mi_is_power_of_two(page_alignment));
mi_assert_internal(page_alignment >= MI_SEGMENT_SIZE);
if (page_alignment < MI_SEGMENT_SIZE) { page_alignment = MI_SEGMENT_SIZE; }

1
src/static.c
View File

@ -27,6 +27,7 @@ terms of the MIT license. A copy of the license can be found in the file
#include "bitmap.c"
#include "heap.c"
#include "init.c"
#include "libc.c"
#include "options.c"
#include "os.c"
#include "page.c" // includes page-queue.c

32
src/stats.c
View File

@ -9,7 +9,6 @@ terms of the MIT license. A copy of the license can be found in the file
#include "mimalloc/atomic.h"
#include "mimalloc/prim.h"
#include <stdio.h> // snprintf
#include <string.h> // memset
#if defined(_MSC_VER) && (_MSC_VER < 1920)
@ -146,7 +145,7 @@ static void mi_printf_amount(int64_t n, int64_t unit, mi_output_fun* out, void*
const int64_t pos = (n < 0 ? -n : n);
if (pos < base) {
if (n!=1 || suffix[0] != 'B') { // skip printing 1 B for the unit column
snprintf(buf, len, "%d %-3s", (int)n, (n==0 ? "" : suffix));
_mi_snprintf(buf, len, "%lld %-3s", (long long)n, (n==0 ? "" : suffix));
}
}
else {
@ -158,8 +157,8 @@ static void mi_printf_amount(int64_t n, int64_t unit, mi_output_fun* out, void*
const long whole = (long)(tens/10);
const long frac1 = (long)(tens%10);
char unitdesc[8];
snprintf(unitdesc, 8, "%s%s%s", magnitude, (base==1024 ? "i" : ""), suffix);
snprintf(buf, len, "%ld.%ld %-3s", whole, (frac1 < 0 ? -frac1 : frac1), unitdesc);
_mi_snprintf(unitdesc, 8, "%s%s%s", magnitude, (base==1024 ? "i" : ""), suffix);
_mi_snprintf(buf, len, "%ld.%ld %-3s", whole, (frac1 < 0 ? -frac1 : frac1), unitdesc);
}
_mi_fprintf(out, arg, (fmt==NULL ? "%12s" : fmt), buf);
}
@ -176,6 +175,7 @@ static void mi_print_count(int64_t n, int64_t unit, mi_output_fun* out, void* ar
static void mi_stat_print_ex(const mi_stat_count_t* stat, const char* msg, int64_t unit, mi_output_fun* out, void* arg, const char* notok ) {
_mi_fprintf(out, arg,"%10s:", msg);
if (unit != 0) {
if (unit > 0) {
mi_print_amount(stat->peak, unit, out, arg);
mi_print_amount(stat->allocated, unit, out, arg);
@ -183,16 +183,8 @@ static void mi_stat_print_ex(const mi_stat_count_t* stat, const char* msg, int64
mi_print_amount(stat->current, unit, out, arg);
mi_print_amount(unit, 1, out, arg);
mi_print_count(stat->allocated, unit, out, arg);
if (stat->allocated > stat->freed) {
_mi_fprintf(out, arg, " ");
_mi_fprintf(out, arg, (notok == NULL ? "not all freed" : notok));
_mi_fprintf(out, arg, "\n");
}
else {
_mi_fprintf(out, arg, " ok\n");
}
}
else if (unit<0) {
mi_print_amount(stat->peak, -1, out, arg);
mi_print_amount(stat->allocated, -1, out, arg);
mi_print_amount(stat->freed, -1, out, arg);
@ -204,11 +196,16 @@ static void mi_stat_print_ex(const mi_stat_count_t* stat, const char* msg, int64
mi_print_amount(-unit, 1, out, arg);
mi_print_count((stat->allocated / -unit), 0, out, arg);
}
if (stat->allocated > stat->freed)
_mi_fprintf(out, arg, " not all freed!\n");
else
}
if (stat->allocated > stat->freed) {
_mi_fprintf(out, arg, " ");
_mi_fprintf(out, arg, (notok == NULL ? "not all freed" : notok));
_mi_fprintf(out, arg, "\n");
}
else {
_mi_fprintf(out, arg, " ok\n");
}
}
else {
mi_print_amount(stat->peak, 1, out, arg);
mi_print_amount(stat->allocated, 1, out, arg);
@ -255,7 +252,7 @@ static void mi_stats_print_bins(const mi_stat_count_t* bins, size_t max, const c
if (bins[i].allocated > 0) {
found = true;
int64_t unit = _mi_bin_size((uint8_t)i);
snprintf(buf, 64, "%s %3lu", fmt, (long)i);
_mi_snprintf(buf, 64, "%s %3lu", fmt, (long)i);
mi_stat_print(&bins[i], buf, unit, out, arg);
}
}
@ -341,6 +338,9 @@ static void _mi_stats_print(mi_stats_t* stats, mi_output_fun* out0, void* arg0)
mi_stat_print(&stats->pages_abandoned, "-abandoned", -1, out, arg);
mi_stat_counter_print(&stats->pages_extended, "-extended", out, arg);
mi_stat_counter_print(&stats->page_no_retire, "-noretire", out, arg);
mi_stat_counter_print(&stats->arena_count, "arenas", out, arg);
mi_stat_counter_print(&stats->arena_crossover_count, "-crossover", out, arg);
mi_stat_counter_print(&stats->arena_rollback_count, "-rollback", out, arg);
mi_stat_counter_print(&stats->mmap_calls, "mmaps", out, arg);
mi_stat_counter_print(&stats->commit_calls, "commits", out, arg);
mi_stat_counter_print(&stats->reset_calls, "resets", out, arg);

4
test/main-override.cpp
View File

@ -108,7 +108,7 @@ static void various_tests() {
auto tbuf = new unsigned char[sizeof(Test)];
t = new (tbuf) Test(42);
t->~Test();
delete tbuf;
delete[] tbuf;
}
class Static {
@ -379,7 +379,7 @@ static void bench_alloc_large(void) {
static constexpr size_t kMaxBufferSize = 25 * 1024 * 1024;
std::unique_ptr<char[]> buffers[kNumBuffers];
std::random_device rd;
std::random_device rd; (void)rd;
std::mt19937 gen(42); //rd());
std::uniform_int_distribution<> size_distribution(kMinBufferSize, kMaxBufferSize);
std::uniform_int_distribution<> buf_number_distribution(0, kNumBuffers - 1);

84
test/test-api.c
View File

@ -34,7 +34,7 @@ we therefore test the API over various inputs. Please add more tests :-)
#include "mimalloc.h"
// #include "mimalloc/internal.h"
#include "mimalloc/types.h" // for MI_DEBUG and MI_ALIGNMENT_MAX
#include "mimalloc/types.h" // for MI_DEBUG and MI_BLOCK_ALIGNMENT_MAX
#include "testhelper.h"
@ -46,6 +46,11 @@ bool test_heap2(void);
bool test_stl_allocator1(void);
bool test_stl_allocator2(void);
bool test_stl_heap_allocator1(void);
bool test_stl_heap_allocator2(void);
bool test_stl_heap_allocator3(void);
bool test_stl_heap_allocator4(void);
bool mem_is_zero(uint8_t* p, size_t size) {
if (p==NULL) return false;
for (size_t i = 0; i < size; ++i) {
@ -154,7 +159,7 @@ int main(void) {
};
CHECK_BODY("malloc-aligned6") {
bool ok = true;
for (size_t align = 1; align <= MI_ALIGNMENT_MAX && ok; align *= 2) {
for (size_t align = 1; align <= MI_BLOCK_ALIGNMENT_MAX && ok; align *= 2) {
void* ps[8];
for (int i = 0; i < 8 && ok; i++) {
ps[i] = mi_malloc_aligned(align*13 // size
@ -170,16 +175,16 @@ int main(void) {
result = ok;
};
CHECK_BODY("malloc-aligned7") {
void* p = mi_malloc_aligned(1024,MI_ALIGNMENT_MAX);
void* p = mi_malloc_aligned(1024,MI_BLOCK_ALIGNMENT_MAX);
mi_free(p);
result = ((uintptr_t)p % MI_ALIGNMENT_MAX) == 0;
result = ((uintptr_t)p % MI_BLOCK_ALIGNMENT_MAX) == 0;
};
CHECK_BODY("malloc-aligned8") {
bool ok = true;
for (int i = 0; i < 5 && ok; i++) {
int n = (1 << i);
void* p = mi_malloc_aligned(1024, n * MI_ALIGNMENT_MAX);
ok = ((uintptr_t)p % (n*MI_ALIGNMENT_MAX)) == 0;
void* p = mi_malloc_aligned(1024, n * MI_BLOCK_ALIGNMENT_MAX);
ok = ((uintptr_t)p % (n*MI_BLOCK_ALIGNMENT_MAX)) == 0;
mi_free(p);
}
result = ok;
@ -187,7 +192,7 @@ int main(void) {
CHECK_BODY("malloc-aligned9") {
bool ok = true;
void* p[8];
size_t sizes[8] = { 8, 512, 1024 * 1024, MI_ALIGNMENT_MAX, MI_ALIGNMENT_MAX + 1, 2 * MI_ALIGNMENT_MAX, 8 * MI_ALIGNMENT_MAX, 0 };
size_t sizes[8] = { 8, 512, 1024 * 1024, MI_BLOCK_ALIGNMENT_MAX, MI_BLOCK_ALIGNMENT_MAX + 1, 2 * MI_BLOCK_ALIGNMENT_MAX, 8 * MI_BLOCK_ALIGNMENT_MAX, 0 };
for (int i = 0; i < 28 && ok; i++) {
int align = (1 << i);
for (int j = 0; j < 8 && ok; j++) {
@ -295,15 +300,22 @@ int main(void) {
// ---------------------------------------------------
// various
// ---------------------------------------------------
#if !defined(MI_TRACK_ASAN) // realpath may leak with ASAN enabled (as the ASAN allocator intercepts it)
CHECK_BODY("realpath") {
char* s = mi_realpath( ".", NULL );
// printf("realpath: %s\n",s);
mi_free(s);
};
#endif
CHECK("stl_allocator1", test_stl_allocator1());
CHECK("stl_allocator2", test_stl_allocator2());
CHECK("stl_heap_allocator1", test_stl_heap_allocator1());
CHECK("stl_heap_allocator2", test_stl_heap_allocator2());
CHECK("stl_heap_allocator3", test_stl_heap_allocator3());
CHECK("stl_heap_allocator4", test_stl_heap_allocator4());
// ---------------------------------------------------
// Done
// ---------------------------------------------------[]
@ -357,3 +369,61 @@ bool test_stl_allocator2(void) {
return true;
#endif
}
bool test_stl_heap_allocator1(void) {
#ifdef __cplusplus
std::vector<some_struct, mi_heap_stl_allocator<some_struct> > vec;
vec.push_back(some_struct());
vec.pop_back();
return vec.size() == 0;
#else
return true;
#endif
}
bool test_stl_heap_allocator2(void) {
#ifdef __cplusplus
std::vector<some_struct, mi_heap_destroy_stl_allocator<some_struct> > vec;
vec.push_back(some_struct());
vec.pop_back();
return vec.size() == 0;
#else
return true;
#endif
}
bool test_stl_heap_allocator3(void) {
#ifdef __cplusplus
mi_heap_t* heap = mi_heap_new();
bool good = false;
{
mi_heap_stl_allocator<some_struct> myAlloc(heap);
std::vector<some_struct, mi_heap_stl_allocator<some_struct> > vec(myAlloc);
vec.push_back(some_struct());
vec.pop_back();
good = vec.size() == 0;
}
mi_heap_delete(heap);
return good;
#else
return true;
#endif
}
bool test_stl_heap_allocator4(void) {
#ifdef __cplusplus
mi_heap_t* heap = mi_heap_new();
bool good = false;
{
mi_heap_destroy_stl_allocator<some_struct> myAlloc(heap);
std::vector<some_struct, mi_heap_destroy_stl_allocator<some_struct> > vec(myAlloc);
vec.push_back(some_struct());
vec.pop_back();
good = vec.size() == 0;
}
mi_heap_destroy(heap);
return good;
#else
return true;
#endif
}

19
test/test-stress.c
View File

@ -42,6 +42,7 @@ static int ITER = 50; // N full iterations destructing and re-creating a
static bool allow_large_objects = true; // allow very large objects? (set to `true` if SCALE>100)
static size_t use_one_size = 0; // use single object size of `N * sizeof(uintptr_t)`?
static bool main_participates = false; // main thread participates as a worker too
// #define USE_STD_MALLOC
#ifdef USE_STD_MALLOC
@ -276,8 +277,8 @@ int main(int argc, char** argv) {
#ifndef USE_STD_MALLOC
#ifndef NDEBUG
mi_collect(true);
//mi_debug_show_arenas();
// mi_collect(true);
mi_debug_show_arenas(true,true,true);
#endif
mi_stats_print(NULL);
#endif
@ -301,13 +302,15 @@ static void run_os_threads(size_t nthreads, void (*fun)(intptr_t)) {
thread_entry_fun = fun;
DWORD* tids = (DWORD*)custom_calloc(nthreads,sizeof(DWORD));
HANDLE* thandles = (HANDLE*)custom_calloc(nthreads,sizeof(HANDLE));
for (uintptr_t i = 0; i < nthreads; i++) {
const size_t start = (main_participates ? 1 : 0);
for (size_t i = start; i < nthreads; i++) {
thandles[i] = CreateThread(0, 8*1024, &thread_entry, (void*)(i), 0, &tids[i]);
}
for (size_t i = 0; i < nthreads; i++) {
if (main_participates) fun(0); // run the main thread as well
for (size_t i = start; i < nthreads; i++) {
WaitForSingleObject(thandles[i], INFINITE);
}
for (size_t i = 0; i < nthreads; i++) {
for (size_t i = start; i < nthreads; i++) {
CloseHandle(thandles[i]);
}
custom_free(tids);
@ -334,11 +337,13 @@ static void run_os_threads(size_t nthreads, void (*fun)(intptr_t)) {
thread_entry_fun = fun;
pthread_t* threads = (pthread_t*)custom_calloc(nthreads,sizeof(pthread_t));
memset(threads, 0, sizeof(pthread_t) * nthreads);
const size_t start = (main_participates ? 1 : 0);
//pthread_setconcurrency(nthreads);
for (size_t i = 0; i < nthreads; i++) {
for (size_t i = start; i < nthreads; i++) {
pthread_create(&threads[i], NULL, &thread_entry, (void*)i);
}
for (size_t i = 0; i < nthreads; i++) {
if (main_participates) fun(0); // run the main thread as well
for (size_t i = start; i < nthreads; i++) {
pthread_join(threads[i], NULL);
}
custom_free(threads);