Fixes GitHub issue #657.
This CL also makes the Windows CI green.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=237255887
This CL fixes the following issues:
* The Travis CI had the ctest invocation followed by a ";", so non-zero
exit codes (indicating test failures) did not cause the build to fail.
* The AppVeyor CI had the ctest invocation followed by a ";", causing an
error on Windows, where "&" plays the role of ";" [1].
The Windows CI (AppVeyor) will still be red after this CL, as some of
the tests are failing. However, this CL is a step forward, as it gets us
from failing to start tests to running tests and recording success/error
states.
[1] https://docs.microsoft.com/en-us/previous-versions/windows/it-pro/windows-xp/bb490954(v=technet.10)#using-multiple-commands-and-conditional-processing-symbols
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=236765633
This change adds a native Windows port (port_windows.h) and a
Windows Env (WindowsEnv).
Note1: "small" is defined when including <Windows.h> so some
parameters were renamed to avoid conflict.
Note2: leveldb::Env defines the method: "DeleteFile" which is
also a constant defined when including <Windows.h>. The solution
was to ensure this macro is defined in env.h which forces
the function, when compiled, to be either DeleteFileA or
DeleteFileW when building for MBCS or UNICODE respectively.
This resolves#519 on GitHub.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=236364778
This prevents file descriptors from leaking to child processes.
When compiled for older (pre-2.6.23) kernels which lack support for
O_CLOEXEC there is no change in behavior. With newer kernels, child
processes will no longer inherit leveldb's file handles, which
reduces the changes of accidentally corrupting the database.
Fixes https://github.com/google/leveldb/issues/623
Apple doesn't follow POSIX specifications for fsync(). Instead, fsync() guarantees to flush the buffer cache to the device, which means the data will survive kernel panics, but may not survive power outages. Applications that need stronger guarantees (like databases) need to use fcntl(F_FULLFSYNC).
This CL switches PosixWritableFile::Sync() to get the stronger guarantees on Apple systems. The improved implementation follows the same principles as SQLite [1] and node.js [2].
Research for the fcntl() to fsync() fallback strategy:
Apple's released source code at https://opensource.apple.com/ shows at least three different error codes being returned when a filesystem does not support F_FULLFSYNC.
fcntl() is implemented in xnu-4903.221.2 in bsd/kern/kern_descrip.c, where it delegates to fcntl_nocancel(). The documentation for fcntl_nocancel() mentions error codes for some operations, but does not include F_FULLFSYNC. The F_FULLSYNC branch in fcntl_nocancel() calls VNOP_IOCTL(_, F_FULLSYNC, NULL, 0, _), whose return value sets the error
code.
VNOP_IOCTL() is implemented in bsd/vfs/kpi_vfs.c and calls the ioctl function in the vnode's operation vector. The per-filesystem function names follow the pattern _vnop_ioctl() for all the instances in opensource code: {hfs,msdosfs,nfs,ntfs,smbfs,webdav,zfs}_vnop_ioctl().
hfs-407.30.1, msdosfs-229.200.3, and nfs in xnu-4903.221.2 handle F_FULLFSYNC. ntfs-94.200.1 and smb-759.40.1 do not handle F_FULLFSYNC, and the default branch returns ENOSUP. webdav-380.200.1 also does not handle F_FULLFSYNC, but the default branch returns EINVAL. zfs-59 also does not handle F_FULLSYNC, and its default branch returns ENOTTY.
From a different angle, Apple's ntfs-94.200.1 includes utility code that uses fcntl(F_FULLFSYNC) and falls back to fsync() just like we do, supporting the hypothesis that there is no good way to detect lack of F_FULLFSYNC support. Also, Apple's fcntl() man page [3] does not mention a way to detect lack of F_FULLFSYNC support.
[1] https://www.sqlite.org/src/doc/trunk/src/os_unix.c
[2] https://github.com/libuv/libuv/blob/master/src/unix/fs.c
[3] https://developer.apple.com/library/archive/documentatiVon/System/Conceptual/ManPages_iPhoneOS/man2/fcntl.2.html
Tested:
https://travis-ci.org/pwnall/leveldb/builds/477318498
TAP global presubmit
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=228593729
The CMake feature-detection code used check_symbol_exists(), which
invokes the C compiler. However, some glibc versions don't expose the
fdatasync() declaration when compiled with -std=c11, but do expose it
when compiled with -std=c++11. This most likely comes down to how
_POSIX_SOURCE is defined -- it needs to be >= 201112L for <unistd.h> to
expose fdatasync().
This CL switches to check_cxx_symbol_exists(), which uses the C++
compiler. Asides from fixing the problem above, this is the right thing
to do, because we use <unistd.h> in env_posix.cc, which is compiled with
the C++ compiler.
This CL also fixes a previously introduced inconsistency, where the
macro indicating the fdatasync() feature detection result was referred
to as HAVE_FDATASYNC and HAVE_FUNC_FDATASYNC. The former appears to be
used in other libraries, so this CL switches all our references to
HAVE_FDATASYNC.
Fixes https://github.com/google/leveldb/issues/629
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=228392612
The space in between the header and log message was mistakenly omitted
in a prior commit. Re-adding.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=228202737
General cleanup principles:
* Use override when applicable.
* Remove static when redundant (methods and globals in anonymous
namespaces).
* Use const on class members where possible.
* Standardize on "status" for Status local variables.
* Renames where clarity can be improved.
* Qualify standard library names with std:: when possible, to
distinguish from POSIX names.
* Qualify POSIX names with the global namespace (::) when possible, to
distinguish from standard library names.
This also refactors the background thread synchronization logic so that
it's statically analyzable.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=219212089
C++11 guarantees thread-safe initialization of static variables inside
functions. This is a more restricted form of std::call_once or
pthread_once_t (e.g., single call site), so the compiler might be able
to generate better code [1]. Equally important, having less
platform-dependent code in env_posix.cc makes it easier to port to other
platforms.
Due to the change above, this CL introduced a new approach for storing
the singleton PosixEnv instance returned by Env::Default(). The new
approach avoids a dynamic memory allocation, which eliminates the false
positive from LeakSanitizer reported in
https://github.com/google/leveldb/issues/539 and
https://github.com/google/leveldb/issues/113
[1] https://stackoverflow.com/a/27206650/
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=214293129
The version in the repository covers the Makefile build. The new version
is simpler and contains entries relevant to the CMake build.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=212661504
This commit replaces the use of pthreads in the POSIX port with std::thread
and port::Mutex + port::CondVar. This is intended to simplify porting
the env to a different platform.
The indirect use of pthreads in PosixLogger is replaced with
std:🧵:id(), based on an approach prototyped by @cmumfordx@.
The pthreads dependency in CMakeFiles is not removed, because some C++
standard library implementations must be linked against pthreads for
std::thread use. Figuring out this dependency is left for future work.
Switching away from pthreads also fixes
https://github.com/google/leveldb/issues/381
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=212478311
This is not an API-breaking change, because it reduces the API that the
leveldb embedder must implement. The project will build just fine
against ports that still implement InitOnce.
C++11 guarantees thread-safe initialization of static variables inside
functions. This is a more restricted form of std::call_once or
pthread_once_t (e.g., single call site), so the compiler might be able
to generate better code [1]. Equally important, having less code in
port_example.h makes it easier to port to other platforms.
Due to the change above, this CL introduces a new approach for storing
the singleton BytewiseComparatorImpl instance returned by
BytewiseComparator(). The new approach avoids a dynamic memory
allocation, which eliminates the false positive from LeakSanitizer
reported in https://github.com/google/leveldb/issues/200
[1] https://stackoverflow.com/a/27206650/
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=212348004
This is separated from the general cleanup because of the logic changes
in SyncDirIfManifest().
General cleanup principles:
* Use override when applicable.
* Remove static when redundant (methods and globals in anonymous
namespaces).
* Use const on class members where possible.
* Standardize on "status" for Status local variables.
* Renames where clarity can be improved.
* Qualify standard library names with std:: when possible, to
distinguish from POSIX names.
* Qualify POSIX names with the global namespace (::) when possible, to
distinguish from standard library names.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=211709673
General cleanup principles:
* Use override when applicable.
* Use const on class members where possible.
* Renames where clarity can be improved.
* Qualify standard library names with std:: when possible, to
distinguish from POSIX names.
* Qualify POSIX names with the global namespace (::) when possible, to
distinguish from standard library names.
This also revamps the logic for putting together a message into the
in-memory buffer before that is passed to fwrite(). While correct in
practice, the current implementation advances a char pointer past the
size of its buffer, which is technically undefined behavior.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=211472570
ssize_t is not standard C++. It is a POSIX extension. Therefore, it does
not belong in generic code.
This change tweaks the logic in DBIter to remove the need for signed
integers, so ssize_t can be replaced with size_t. The impacted method
and private member are renamed to better express their purpose.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=211471606
Now that we require C++11, we can use std::atomic<int>, which has
primitives for most of the logic we need. As a bonus, the happy path for
Limiter::Acquire() and Limiter::Release() only performs one atomic
operation.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=211469518
"Create a brand new [adjective] file" seems like the description for a
method that will create a new file, but is used for methods that open
existing files for read access.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=211468002
WriteBatchInternal has a method for efficiently concatenating two
WriteBatches. This commit exposes the method to the public API.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=208724311
This CL renames the private struct Iterator::Cleanup ->
Iterator::CleanupNode, to better reflect that it's a linked list node,
and extracts duplicated code from its user in IsEmpty() and Run()
methods.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=199175058
* Omit SnapshotImpl::list_ when assert() isn't on
* Make SnapshotImpl::number_ const and set it in the constructor
* Make SnapshotImpl::number_ private and access it via a getter
* Rename SnapshotImpl::number_ to SnapshotImpl::sequence_number_
* Rename SnapshotList::list_ to SnapshotList::head_
* Wrap casting from Snapshot* to SnapshotImpl* in ToSnapshotImpl()
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=194852828
The default size was changed in #f779e7a5 but the documentation was
never updated.
This fixes#566 reported on GitHub.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=194547959
The porting layer implements threading primitives: atomic pointers,
condition variables, mutexes, thread-safe initialization. These are all
specified in C++11, so the reference open source port implementation can
become platform-independent.
The porting layer will remain in place to allow the use of other
implementations with more features, such as the built-in deadlock
detection in abseil's Mutex.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=193245934
Commit a0008deb679480fd30e845d7e52421af72160c2c introduced
std::numeric_limits usage in logging.cc, but didn't #include <limits>
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=192840190
The old implementation caused odd crashes on ARM, which were fixed by
changing a local variable type. The main suspect is the use of a static
local variable. This CL replaces the static local variable with
constexpr, which still ensures the compiler sees the expressions as
constants.
The CL also replaces Slice operations in the functions' inner loop with
iterator-style pointer operations, which can help the compiler generate
less code.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=192832175
ConsumeDecimalNumber has fairly non-trivial logic, and a previous
version has crashed inexplicably on Android. Having some test coverage
will make it easier to tweak / simplify the function later on.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=192821751
This is an accidental leftover from the CMake migration. The macro has
been replaced with LEVELDB_IS_BIG_ENDIAN.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=192364918
The CMake-based build relies on __has_include, which is standardized in
C++17. Unfortunately, __has_include is available without requiring
--std=c++17 on all the compilers on CI, so this problem was not caught.
Fixes https://github.com/google/leveldb/issues/572
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=192208842
After this CL, all classes with Mutex members should be covered by annotations. Exceptions are atomic members, which shouldn't need locking, and DBImpl members that cause errors when annotated, which will be tackled separately.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=190260865
C++11 requires <atomic>. This lets us remove the header detection
(LEVELDB_ATOMIC_PRESENT) and simplify port/atomic_pointer.h.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=189919098
This CL switches the public headers to C++11 default and deleted constructors, and adds override to the relevant leveldb::EnvWrapper methods. This should be a good test for C++11 compiler support.
Once this CL settles, the rest of the codebase can be safely modernized to C++11.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=189873212
helpers/memenv/memenv.cc used SIZE_MAX without including <stdint.h>.
Since we're fixing this problem, replace SIZE_MAX with
std::numeric_limits<size_t>::max(), which is clearer.
Fixes https://github.com/google/leveldb/issues/562
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=189821707
This CL makes it easier to reason about thread safety by:
1) Adding Clang thread safety annotations according to comments.
2) Expanding a couple of variable names, without adding extra lines of code.
3) Adding const in a couple of places.
4) Replacing an always-non-null const pointer with a reference.
5) Fixing style warnings in the modified files.
This CL does not annotate the DBImpl members that claim to be protected
by the instance mutex, but are accessed without the mutex being held.
Those members (and their unprotected accesses) will be addressed in
future CLs.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=189354657
This CL removes unused headers included by util/testharness.h, adds
precise includes where the build breaks, and fixes style errors in the
edited files.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=189331061
External linkage is the default for function declarations in C++.
This also fixes ClangTidy errors generated by removing the "extern"
keyword as described above.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=188730416
This is a stopgap for removing warnings on Mac builds, so -Werror can be
turned on. C++11 will be required in the nearby future, which guarantees
<atomic> support. Once that happens, the simplified version of this will
match https://github.com/google/leveldb/pull/503
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=188553251
This aligns the Travis CI configuration with google/crc32c and
google/snappy, to simplify maintenance.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=188547648
This removes the use of the non-portable headers <sys/types.h> and <unistd.h> in c_test.c.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=188503102
When the max file size option was added in CL 134391640 the C API
was not modified to support this.
This change was contributed by GitHub user @olt and fixes issue #439.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=173466388
The C++ style guide URL was wrong.
This fixes issue #394. Reported by GitHub user @Loki-Astari.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=173188573
Deleted two unused assignments:
1. offset_in_block in Reader::SkipToInitialBlock().
2. in_fragmented_record in Reader::ReadRecord().
Reasons for the change:
1. offset_in_block is not read again after the if condition.
2. The kFullRecordType switch branch returns, so
in_fragmented_record isn't read again.
3. The kFirstType switch branch sets in_fragmented_record to
true after the if, so the write in the if is ignored.
Change contributed by @C0deAi on GitHub.
This fixes https://github.com/google/leveldb/issues/517
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=172763897
This follows the general naming convention for preprocessor macros used
to detect feature (library / header file / symbol) presence.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=171184641
CL 170738066 removed all instances of fread_unlocked, fwrite_unlocked
and fflush_unlocked calls from the codebase, so the feature detection
can be removed as well.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=171154269
gcc defaults to exporting all symbols, but other linkers do not. Adding
the LEVELDB_EXPORT macro allows a project to set LEVELDB_SHARED_LIBRARY
when building/linking with leveldb as a shared library.
This is to allow leveldb to be created as a shared library on all
platforms support by Chrome and enables a fix for
https://bugs.chromium.org/p/chromium/issues/detail?id=764810.
This also has the benefit of reducing the shared library size from
418863 to 380367 bytes (64-bit Linux).
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=171037148
CL 170738066 introduced std::min and std::max to env_test.cc. These
require the <algorithm> header.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=171024062
Deleting a PosixWritableFile without calling Close() leaks the file
descriptor. While the API description in include/leveldb/env.h does not
specify whether the caller is responsible for Close()ing the file before
deleting it, all other Env file implementations do release underlying
resources when destroyed, even if Close() is not called.
The leak shows up when running db_tests on Mac Travis, or on a vanilla
MacOS install.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=170906843
This also removes std::unique_ptr introduced in CL 170738066, because
it's C++11-only, and the open source version still supports older
versions at the moment.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=170876919
If the file already existed, we should have truncated it. This was not
detected by leveldb tests since leveldb code avoids reusing same files,
but there was code elsewhere that was directly using leveldb files and
relying on this behavior.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=170769101
If leveldb::Options::block_cache is set to a cache of zero capacity
then it is possible for LRUHandle::next to be used without having been
set.
Conditional jump or move depends on uninitialised value(s):
leveldb::(anonymous namespace)::LRUHandle::key() const (cache.cc:58)
leveldb::(anonymous namespace)::LRUCache::Unref(leveldb::(anonymous namespace)::LRUHandle*) (cache.cc:234)
leveldb::(anonymous namespace)::LRUCache::Release(leveldb::Cache::Handle*) (cache.cc:266)
leveldb::(anonymous namespace)::ShardedLRUCache::Release(leveldb::Cache::Handle*) (cache.cc:375)
leveldb::CacheTest::Insert(int, int, int) (cache_test.cc:59)
This bug forced a commit reversion in Chromium. For more information see
https://bugs.chromium.org/p/chromium/issues/detail?id=761398#c4
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=170749054
env_posix.cc and concurrent application calls to fflush(NULL).
The fix is to avoid using stdio in env_posix.cc but add our own
buffering where we need it.
Added a test to reproduce the bug.
Added a test for Env reads/writes.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=170738066
12 lines above, there is an "if (!s.ok()) { return s; }" block of code.
"s" is never modified between that block and the "if" removed by this
CL, so "s.ok()" must be true.
The code most likely intended to say "if (!builder->ok())", because the
builder->Add() call above can modify the TableBuilder's status, as a
side-effect. However, this approach would have required setting "s =
builder.status()" in the "else" branch, near the "builder.Abandon()"
call. So, removing the "if" outright is simpler than following that line
of thought.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=167326229
An Android test was occasionally crashing with a SEGV in ConsumeDecimalNumber
Switching a local variable from an int to uint64_t eliminated these crashes.
Speculating this is either a compiler, runtime library, or emulator issue.
Switching this type to uint64_t also eliminates a compiler warning
about comparing an int with a uint64_t.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=166399695
When faced with a pointer that is misaligned by K bytes (pointer % 8 ==
K), the code previously moved forward by K bytes. In order to end up
with an aligned pointer, the code must move by 8 - K bytes.
This lands https://github.com/google/leveldb/pull/488
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=166295921
The dead code has been in the codebase since the initial commit and is
generating a compiler warning when used in Xcode.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=164174594
Use __APPLE__ instead of OS_MACOS when testing for the Apple platform and
remove the latter symbol from the BUILD file. This fixes incompatibility issues
when using the library on an Apple device.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=162958094
BTRFS reorders rename and write operations, so it is possible that a filesystem crash and recovery results in a situation where the file pointed to by CURRENT does not exist. DB::Open currently reports an I/O error in this case. Reporting database corruption is a better hint to the caller, which can attempt to recover the database or erase it and start over.
This issue is not merely theoretical. It was reported as having showed up in the wild at https://github.com/google/leveldb/issues/195 and at https://crbug.com/738961. Also, asides from the BTRFS case described above, incorrect data in CURRENT seems like a possible corruption case that should be handled gracefully.
The Env API changes here can be considered backwards compatible, because an implementation that returns Status::IOError instead of Status::NotFound will still get the same functionality as before.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=161432630
env_test.cc defines EnvPosixTest which tests the Env implementation returned by Env::Default(). The naming is a bit unfortunate, as the tests in env_test.cc are written against the Env contract, and therefore are applicable to any Env implementation. An instance of the confusion caused by the naming is [] which added a dependency from env_test.cc to EnvPosixTestHelper, which is closely coupled to EnvPosix.
This change disentangles EnvPosix-specific test code into a env_posix_test.cc file. The code there uses EnvPosixTestHelper and specifically targets the EnvPosix implementation. env_test.cc now implements EnvTest, and contains tests that are also applicable to other ports, which may define their own Env implementation.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=148914642
LE_LOAD64 is only used when _mm_crc32_u64 is available, on 64-bit x86 processors.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=148906169
Prior index.html was using rawgit.com which doesn't process
Markdown and therefore only serves the markdown source.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=148902180
Markdown is more readable in a text editor and when hosted
on GitHub is more readable than HTML.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=148830423
This change authored by vadimskipin and submitted via:
https://github.com/google/leveldb/pull/309
Changes made to support iOS builds and other architectures
without support for SSE 4.2.
db_bench reports original crc32 speed at:
crc32c : 3.610 micros/op; 1082.0 MB/s (4K per op)
with this change performance has increased to:
crc32c : 0.843 micros/op; 4633.6 MB/s (4K per op)
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=148694935
Background compaction can create an unbounded number of
leveldb::RandomAccessFile instances. On 64-bit systems mmap is used and
file descriptors are only used beyond a certain number of mmap's.
32-bit systems to not use mmap at all. leveldb::RandomAccessFile does not
observe Options.max_open_files so compaction could exhaust the file
descriptor limit.
This change uses getrlimit to determine the maximum number of open
files and limits RandomAccessFile to approximately 20% of that value.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=143505556
Background:
LevelDB uses a cache (util/cache.h, util/cache.cc) of (key,value)
pairs for two purposes:
- a cache of (table, file handle) pairs
- a cache of blocks
The cache places the (key,value) pairs in a reference-counted
wrapper. When it returns a value, it returns a reference to this
wrapper. When the client has finished using the reference and
its enclosed (key,value), it calls Release() to decrement the
reference count.
Each (key,value) pair has an associated resource usage. The
cache maintains the sum of the usages of the elements it holds,
and removes values as needed to keep the sum below a capacity
threshold. It maintains an LRU list so that it will remove the
least-recently used elements first.
The max_open_files option to LevelDB sets the size of the cache
of (table, file handle) pairs. The option is not used in any
other way.
The observed behaviour:
If LevelDB at any time used more file handles concurrently than
the cache size set via max_open_files, it attempted to reduce the
number by evicting entries from the table cache. This could
happen most easily during compaction, and if max_open_files was
low. Because the handles were in use, their reference count did
not drop to zero, and so the usage sum in the cache was not
modified by the evictions. Subsequent Insert() calls returned
valid handles, but their entries were immediately evicted from
the cache, which though empty still acted as though full. As a
result, there was effectively no caching, and the number of open
file handles rose []ly until it hit system-imposed limits and
the process died.
If one set max_open_files lower, the cache was more likely to
exhibit this beahviour, and cause the process to run out of file
descriptors. That is, max_open_files acted in almost exactly the
opposite manner from what was intended.
The problems:
1. The cache kept all elements on its LRU list eligible for capacity
eviction---even those with outstanding references from clients. This was
ineffective in reducing resource consumption because there was an
outstanding reference, guaranteeing that the items remained. A secondary
issue was that there is no guarantee that these in-use items will be the
last things reached in the LRU chain, which actually recorded
"least-recently requested" rather than "least-recently used".
2. The sum of usages was decremented not when a (key,value) was evicted from
the cache, but when its reference count went to zero. Thus, when things
were removed from the cache, either by garbage collection or via Erase(),
the usage sum was not necessarily decreased. This allowed the cache to act
as though full when it was in fact not, reducing caching effectiveness, and
leading to more resources being consumed---the opposite of what the
evictions were intended to achieve.
3. (minor) The cache's clients insert items into it by first looking up the
key, and inserting only if no value is found. Although the cache has an
internal lock, the clients use no locking to ensure atomicity of the
Lookup/Insert pair. (see table/table.cc: block_cache->Insert() and
db/table_cache.cc: cache_->Insert()). Thus, if two threads Insert() at
about the same time, they can both Lookup(), find nothing, and both
Insert(). The second Insert() would evict the first value, leaving each
thread with a handle on its own version of the data, and with the second
version in the cache. It would be better if both threads ended up with a
handle on the same (key,value) pair, which implies it must be the first item
inserted. This suggests that Insert() should not replace an existing value.
This can be made safe with current usage inside LeveDB itself, but this is
not easy to change first because Cache is a public interface, so to change
the semantics of an existing call might break things, second because Cache
is an abstract virtual class, so adding a new abstract virtual method may
break other implementations, and third, the new method "insert without
replacing" cannot be implemented in terms of the existing methods, so cannot
be implemented with a non-abstract default. But fortunately, the effects
of this issue are minor, so this issue is not fixed by this change.
The changes:
The assumption in the fixes is that it is always better to cache
entries unless removal from the cache would lead to deallocation.
Cache entries now have an "in_cache" boolean indicating whether
the cache has a reference on the entry. The only ways that this can
become false without the entry being passed to its "deleter" are via
Erase(), via Insert() when an element with a duplicate key is inserted,
or on destruction of the cache.
The cache now keeps two linked lists instead of one. All items
in the cache are in one list or the other, and never both. Items
still referenced by clients but erased from the cache are in
neither list. The lists are:
- in-use: contains the items currently referenced by clients, in no particular
order. (This list is used for invariant checking. If we removed the check,
elements that would otherwise be on this list could be left as disconnected
singleton lists.)
- LRU: contains the items not currently referenced by clients, in LRU order
A new internal Ref() method increments the reference count. If
incrementing from 1 to 2 for an item in the cache, it is moved
from the LRU list to the in-use list.
The Unref() call now moves things from the in-use list to the LRU
list if the reference count falls to 1, and the item is in the
cache. It no longer adjusts the usage sum. The usage sum now
reflects only what is in the cache, rather than including
still-referenced items that have been evicted.
The LRU_Append() now takes a "list" parameter so that it can be
used to append either to the LRU list or the in-use list.
Lookup() is modified to use the new Ref() call, rather than
adjusting the reference count and LRU chain directly.
Insert() eviction code is also modified to adjust the usage sum and the
in_cache boolean of the evicted elements. Some LevelDB tests assume that there
will be no caching whatsoever if the cache size is set to zero, so this is
handled as a special case.
A new private method FinishErase() is factored out
with the common code from where items are removed from the cache.
Erase() is modified to adjust the usage sum and the in_cache
boolean of the erased elements, and to use FinishErase().
Prune() is modified to use FinishErase() also, and to make use of the fact that
the lru_ list now contains only items with reference count 1.
- EvictionPolicy is modified to test that an entry with an
outstanding handle is not evicted. This test fails with the old cache.cc.
- A new test case UseExceedsCacheSize verifies that even when the
cache is overfull of entries with outstanding handles, none are
evicted. This test fails with the old cache.cc, and is the key
issue that causes file descriptors to run out when the cache
size is set too small.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=123247237
