Add Mach exception server and fill out exceptions snapshot.
Note that:
- The 'capture' portion of this CL will be moved out of the snapshot
interface and into a separate in-process dump to disk location.
- All of the pointer dereferences need to be wrapped in vm_read.
- The read-fast-and-dump logic in exception_snapshot will end up in a
different file completely, but until we pick a
serialization/deserialization method, keep it as-is.
Bug: crashpad:31
Change-Id: I44203aa44036a341d6b4517fde7ab0cb9d7e94d7
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2160122
Commit-Queue: Justin Cohen <justincohen@chromium.org>
Reviewed-by: Mark Mentovai <mark@chromium.org>
This was previously used, but is no longer required. Delete the one
remaining include at the build rules.
Change-Id: If5083a4fb8a5562d3e40149976bd27fcec0fd302
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2165635
Reviewed-by: Mark Mentovai <mark@chromium.org>
Reviewed-by: Francois Rousseau <frousseau@google.com>
Commit-Queue: Scott Graham <scottmg@chromium.org>
This updates the way that the .defs files added in 1bfd7d06ed60 are
treated, by putting them in compat/ios, using compat as intended. The
.defs files in compat forward, via #include, to the ones in
third_party/xnu. Additionally, compat/mac is enabled for iOS, as
everything in compat/mac is sensible on iOS, and will have no effect
when rendered unnecessary by the iOS SDK.
This also changes util:mig_output to a static_library instead of a
source_set. I don’t think there was any reason for it to have been a
source_set to begin with. static_library is preferred for nearly
everything.
Bug: crashpad:31
Change-Id: I7c468d6d6785bf2bc825d45831ebb81e1c9ddfbc
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2160310
Reviewed-by: Justin Cohen <justincohen@chromium.org>
Commit-Queue: Mark Mentovai <mark@chromium.org>
This makes UniversalMachExcServer available on iOS.
UniversalMachExcServer is the foundation for a Mach exc and mach_exc
server.
Some code in UniversalMachExcServer needs to be evaluated to ensure that
portions that run in the same process that has sustained the exception
are safe to do so at that time. For example,
SimplifiedExcServer<ExcTraits>::Interface instantiates and appends to a
std::vector<>, which is generally unsafe in this context. However, that
code responds to exc requests. The mach_exc equivalent,
SimplifiedMachExcServer<MachExcTraits>::Interface, does not use a vector
at all.
This also enables support code in the form of CompositeMachMessageServer
and UniversalExceptionRaise, all of the tests for
CompositeMachMessageServer, and most of the test for
exc_server_variants.cc. The multiprocess-based exc_server_variants tests
remain disabled on iOS.
Bug: crashpad:31
Change-Id: I838ed770a33ca29c37383c32245eb340fb3ad2fb
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2159287
Reviewed-by: Justin Cohen <justincohen@chromium.org>
Commit-Queue: Mark Mentovai <mark@chromium.org>
mig was being invoked without any -arch argument, causing it to assume
the build system’s native architecture, which would be x86_64. This is
not correct for iOS device builds, which use arm64. The -arch argument
must be plumbed to mig for correct behavior.
When building for iOS, mig was being invoked without any -isysroot
argument, causing it to use the root for the build system, which runs
macOS and not iOS. The macOS SDK doesn’t include the ARM definitions
needed for iOS device builds.
<mach/exc.defs> and <mach/mach_exc.defs> depend on a small number of
other .defs files to provide definitions of standard types. All .defs
files are absent from the iOS SDK. These .defs files are borrowed from
xnu and placed in third_party/xnu. An additional --include argument is
added to allow mig to locate these files.
Bug: crashpad:31
Change-Id: I27154310352939ebe2fb6329bbbfda701c369289
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2159291
Reviewed-by: Justin Cohen <justincohen@chromium.org>
Commit-Queue: Mark Mentovai <mark@chromium.org>
This updates (and corrects) 8dbbaff2e1a5, which added exc.defs, by
adding mach_exc.defs too.
The difference betwen the exc and mach_exc subsystems is that the |code|
parameter is int[] in exc and int64_t[] in mach_exc. Many exceptions
carry the exception address in code[1], and a 32-bit int results in the
exception address being truncated in exc. No information is lost in
mach_exc, where a 64-bit int64_t is used.
In 8dbbaff2e1a5, I misremembered the type of the |code| parameter as a
type derived from uintptr_t, such as vm_address_t, an integer as wide as
a pointer. I was wrong, and mach_exc is necessary. I also noted that
Apple normally forbids mach_-prefixed interfaces in favor of the
prefix-less ones for the reasons I mentioned, and that, all else being
equal, it was desirable to adhere to the spirit of that convention.
Because neither exc nor mach_exc are available in the SDK, it’s moot
from a technical perspective, as we need to provide our own stubs either
way.
Bug: crashpad:31
Change-Id: Ied1be470e653b2bead1a283cb8b9283d210c328d
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2159286
Reviewed-by: Justin Cohen <justincohen@chromium.org>
Commit-Queue: Mark Mentovai <mark@chromium.org>
This enables the following code in util/mach on iOS:
- exception_behaviors.{cc,h}
- exception_ports.{cc,h}
- mach_message.{cc,h}
- mach_message_server.{cc,h}
Only the ExceptionBehaviors and MachMessage tests are built, because the
other two are tested by multiprocess tests that won’t run on iOS.
The AuditPIDFromMachMessageTrailer function from mach_message.h is
excluded on iOS because it relies on <bsm/libbsm.h>, which is broken on
iOS: it depends on <bsm/audit_record.h>, which is missing from the SDK.
Additionally, the BSM function that Crashpad uses, audit_token_to_au32,
is marked as unavailable on iOS. Crashpad uses it on macOS to
authenticate Mach messages sent by other processes, but this is moot on
iOS.
Bug: crashpad:31
Change-Id: I5ebc4b80543989b9cd0b85b82eb4b3ff98c44e6c
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2155086
Reviewed-by: Justin Cohen <justincohen@chromium.org>
Commit-Queue: Mark Mentovai <mark@chromium.org>
The iOS SDK doesn’t include a copy of <mach/exc.defs>. It only provides
<mach/exc.h>, which is just the user-side header. To obtain declarations
and implementations of the server-side stubs, a current copy of
<mach/exc.defs> is added to third_party, and the mig action in util is
updated to use it on iOS.
The three other mig subsystems that Crashpad uses are not brought to
iOS:
- mach_exc is identical to exc except it always uses 64-bit quantities
for addresses in place of exc’s use of quantiies sized for native
pointers. Because all iOS work is limited to a single process, there
is no need to consider cross-process operation with variable bitness,
so mach_exc is unnecessary. We’re also only targeting 64-bit for iOS,
so exc will always suffice. This follows the spirit of other
mach_-prefixed routines on iOS, where Apple forbids mach_vm_read to
user applications but permits vm_read.
- notify is primarily used on macOS in the Crashpad handler process to
receive a no-senders notification, which is used to trigger handler
shutdown when it has no more clients. This is not believed to be
useful to Crashpad on iOS, which is restricted to single-process
operation.
- child_port is a Crashpad-specific subsystem used to pass Mach rights
between processes, but is similarly useless when restricted to
single-process operation as on iOS.
Bug: crashpad:31
Change-Id: Id4cb3cdd529814438d378c20702c82c1e89dd2be
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2154530
Commit-Queue: Mark Mentovai <mark@chromium.org>
Reviewed-by: Justin Cohen <justincohen@chromium.org>
mach_extensions is sensible on iOS, but bootstrap is not available
outside of macOS. To allow mach_extensions to be used cleanly on iOS,
the bootstrap code is moved into its own macOS-specific file.
Bug: crashpad:31
Change-Id: I7bf9d5194253b563954a1e55fbf67a16f686e8ff
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2154529
Reviewed-by: Justin Cohen <justincohen@chromium.org>
Commit-Queue: Mark Mentovai <mark@chromium.org>
This builds some code in the util and test libraries on iOS that was
previously excluded. It also enables tests for this code, and other
tests that it was possible to enable either previously or as a result of
this change.
Previously, crashpad_util_test ran 178 tests from 46 test suites, and
crashpad_test_test ran 2 tests from 2 test suites. Now,
crashpad_util_test runs 284 tests from 62 test suites, and
crashpad_test_test runs 6 tests from 4 test suites.
The related .gn files also suffered through a slight cleanup.
Bug: crashpad:31
Change-Id: I84cdda5631f0ea4888ada902a8462776ac46fd2a
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2154526
Reviewed-by: Justin Cohen <justincohen@chromium.org>
Commit-Queue: Mark Mentovai <mark@chromium.org>
When code raises an Objective-C exception, unwind the stack looking for
any exception handlers. If an exception handler is encountered, test to
see if it is a function known to be a catch-and-rethrow 'sinkhole'
exception handler. Various routines in UIKit and elsewhere do this, and
they obscure the exception stack, since the original throw location is
no longer present on the stack (just the re-throw) when Crashpad
captures the crash report. In the case of sinkholes, trigger an
immediate exception to capture the original stack.
The is an improvement over the alternative,
NSSetUncaughtExceptionHandler, which passes along the stack frames, but
not the stack memory contents and full exception context itself.
The details of what happens after a fatal exception is triggered are
unresolved in this CL. For now, simply call std::terminate.
This code was inspired by chromium/src/chrome/browser/mac/
exception_processor.mm.
Bug: crashpad:31
Change-Id: Ieebc6476a0507c466c8219c10f790ec0a624e58c
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2125254
Commit-Queue: Justin Cohen <justincohen@chromium.org>
Reviewed-by: Robert Sesek <rsesek@chromium.org>
Reviewed-by: Mark Mentovai <mark@chromium.org>
Bionic installs signal handlers which request crash
dumps from Android's debuggerd, but there are errors
in how signals which aren't automatically re-raised
are handled on Marshmallow (API 23).
Before requesting a dump, Bionic acquires a lock to
communicate with debuggerd and expecting imminent
death, never releases it.
While handling the dump request, debuggerd allows
the dying process to continue before
ptrace-detaching it. So, when Bionic manually
re-raises a signal, it is intercepted by debuggerd
and the dying process is allowed to live.
Bionic restores SIG_DFL for the signal it's just
handled, but if a different crash signal is later
recieved, Bionic attempts to reacquire the lock to
communicate with debuggerd and blocks forever.
Disable Bionic's signal handlers for these signals
on Marshmallow.
Bug: chromium:1050178
Change-Id: Ia1fc5a24161a95931684d092ba8fee2f0dfbbdbb
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2134513
Reviewed-by: Mark Mentovai <mark@chromium.org>
Fixes the chromium presubmit error for UTF8String:
The use of -[NSString UTF8String] is dangerous as it can return null
even if |canBeConvertedToEncoding:NSUTF8StringEncoding| returns YES.
Please use |SysNSStringToUTF8| instead.
Bug: crashpad:31
Change-Id: Iaf939012ea9d342f6a01af58119cef962319aefe
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2119613
Reviewed-by: Mark Mentovai <mark@chromium.org>
Commit-Queue: Justin Cohen <justincohen@chromium.org>
Gather most of the necessary information for the system snapshot.
Note that:
- The 'capture' portion of this CL will be moved out of the snapshot
interface and into a separate in-process dump to disk location.
- All of the pointer dereferences need to be wrapped in vm_read.
- The read-fast-and-dump logic in thread_snapshot may end up in a
different file completely, but until we pick a
serialization/deserialization method, keep it as-is.
Bug: crashpad:31
Change-Id: Iac82491fdb4a823163f02149f52a1e18e26fa9de
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2090173
Commit-Queue: Justin Cohen <justincohen@chromium.org>
Reviewed-by: Mark Mentovai <mark@chromium.org>
This may help us debug some issues where the thread is failing to
suspend.
Bug: b/151318587
Change-Id: I0d2d539f769ebb1cdd71606e1d23d8fa66673879
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2103411
Commit-Queue: John Bauman <jbauman@chromium.org>
Reviewed-by: Scott Graham <scottmg@chromium.org>
These should be the remaining instances of UB errors we see.
Bug: fuchsia:46805
Change-Id: Id8285386fd6cb52518f6076ddb79ac60025f9f87
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2067754
Reviewed-by: Scott Graham <scottmg@chromium.org>
Commit-Queue: Scott Graham <scottmg@chromium.org>
First steps at bringing up the crashpad_client on iOS. Also updates
the XCUITest to trigger various crashes, with some swizzling
necessary to allow crashes.
Change-Id: I87dd36bed1c052b509d14bfa29679ed81e58a377
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2039470
Commit-Queue: Justin Cohen <justincohen@chromium.org>
Reviewed-by: Mark Mentovai <mark@chromium.org>
Reviewed-by: Rohit Rao <rohitrao@chromium.org>
lld wants to remove bl/blx substitution for non-function symbols
(https://reviews.llvm.org/D73542). GNU ld apparently already doesn't
do it.
Since _ZN8crashpad14CaptureContextEP10ucontext_t wasn't marked as a function,
chromium's thumb code would then branch without mode transition into
crashpads non-thumb assembly (in arm32).
So mark the symbol as function, so that things work even if that patch
relands. This should also make things work with GNU ld, though I haven't
verified that it was broken before and works now.
I also did this for aarch64 since it seems like The Right Thing To Do
(assuming the assembler accepts it, which I also haven't checked --
the CQ will hopefully check that).
Bug: chromium:1049649
Change-Id: I3452c16f0d52a2dc0397fd3d60d06b5c39a4b524
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/2044144
Reviewed-by: Mark Mentovai <mark@chromium.org>
Commit-Queue: Nico Weber <thakis@chromium.org>
- Add option to log minidump in handler_main, also add option to
disable to dump minidump and generate report.
- Implement log minidump in CrashReportExceptionHandler.
Bug: crashpad:308
Change-Id: I8d2f7e118912011a8416f1ec36c9ee9d561d06e6
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/1995825
Commit-Queue: Tao Bai <michaelbai@chromium.org>
Reviewed-by: Joshua Peraza <jperaza@chromium.org>
- This tool could compress/encode or decode/decompress the minidump
log file, will be used by script to symbolize the crash.
- Added FileOutputStream and FileEncoder.
Bug: crashpad:308
Change-Id: I15c3e4908882a09983ec81a90e38249967c29fc4
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/1968059
Commit-Queue: Tao Bai <michaelbai@chromium.org>
Reviewed-by: Joshua Peraza <jperaza@chromium.org>
Emit the received data to Android logcat in Android, and noop for
other platforms.
Bug: crashpad:308
Change-Id: I6e46e2fa8bd61f93f614ad0bfb6441a79139b04b
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/1958711
Reviewed-by: Joshua Peraza <jperaza@chromium.org>
This CL modifies the creation of the Named Pipe Security Descriptor to
allow access from AppContainer processes. The DACL only allows access for
the current user and SYSTEM which matches up with the auto-assigned DACL
used previously (the read-only logon SID ACE has been removed). As this
new code uses APIs from ADVAPI32 a check is made to ensure it's not being
called while the loader lock is held to avoid hitting previous similar
issues.
Bug: crashpad: 318
Change-Id: I3f9cf5c788dbadacad21c8a2d57a0188f690ac32
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/1955982
Commit-Queue: James Forshaw <forshaw@chromium.org>
Reviewed-by: Mark Mentovai <mark@chromium.org>
"\r\n" was used instead of "\n" on four new files.
No other "\r" appears in any text file, repository-wide.
Bug: crashpad:316
Change-Id: I94f5d20cd2498e76efdee6062382669362e6e53d
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/1954713
Reviewed-by: Joshua Peraza <jperaza@chromium.org>
Commit-Queue: Mark Mentovai <mark@chromium.org>
This CL adds code to check if the current thread holds the DLL loader
lock. This code can be used to enforce the requirement that certain
parts of crashpad, such as process creation are not done during calls
to DllMain which can lead to deadlocks and crashes. Only one check is
current enforced, in client process creation, and only in debug builds.
Bug: crashpad: 316
Change-Id: I5757a264bbf28ce2ab88a0cd7ac9481e46428c17
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/1945993
Reviewed-by: Mark Mentovai <mark@chromium.org>
Reviewed-by: Scott Graham <scottmg@chromium.org>
Commit-Queue: James Forshaw <forshaw@chromium.org>
ProcStatReader.Threads is flaky because it relies on an internal,
imprecise measurement of boot time. The flaky test asserts that a
thread started after the main thread should have a start time >= the
main thread. The start time is returned in a timeval, with microsecond
precision, but the measurement of boot time requires two system calls
and the time between those system calls can be approximately a
microsecond. An unlucky event such as a change in system time could
make this imprecision arbitrarily bad.
This patch lets the caller of ProcStatReader.StartTime() inject the
boot time, allowing ProcStatReader to guarantee that threads have
correctly ordered time, given the same input boot time.
Bug: 1016765
Change-Id: I6e4a944a1d58c3916090bab6a4b99573e71a89fc
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/1891588
Reviewed-by: Mark Mentovai <mark@chromium.org>
Commit-Queue: Joshua Peraza <jperaza@chromium.org>
* this renders as "HTTPTransport/HTTPTransport.ValidFormData_Gzip/http"
instead of the default "HTTPTransport/HTTPTransport.ValidFormData_Gzip/0"
* switch the parameter type from a base::FilePath::StringType to a string
Change-Id: I19743966406f92176c566827d74a79aef5a87bb5
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/1900324
Reviewed-by: Scott Graham <scottmg@chromium.org>
Commit-Queue: Francois Rousseau <frousseau@google.com>
Chromium requires build artifacts to be generated deterministically so
commit a long-lived (10 years) test certificate to the repository.
Change-Id: I7a6e2441f506196ca58fbbf757648fa0ac70bc9a
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/1872188
Reviewed-by: Mark Mentovai <mark@chromium.org>
Reviewed-by: Scott Graham <scottmg@chromium.org>
Or else the uploader will check fail when uploading to https://...
Change-Id: I88a765215cc7bff5809b8effd92f4e39bebd1e5b
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/1860940
Reviewed-by: Mark Mentovai <mark@chromium.org>
Commit-Queue: Joshua Peraza <jperaza@chromium.org>
AT_SYSINFO_EHDR may be defined even if no vdso is used. e.g. 32-bit ARM
processes have a vdso on 64-bit ARM cpus, but not on 32-bit cpus.
Change-Id: I4d9ce029bb47efc33ea16cb4c5c2055c1b9330c9
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/1860935
Reviewed-by: Mark Mentovai <mark@chromium.org>
Commit-Queue: Joshua Peraza <jperaza@chromium.org>
Since gtest 00938b2b228f, gtest has built-in first-class support for
skipping tests, which is functionally identical (at least in Crashpad’s
usage) to the home-grown support for run-time dynamically disabled tests
introduced in Crashpad 5e9ed4cb9f69.
Use the new standard pattern, and remove all vestiges of the custom
local one.
This was done previously in 79f4a3970a64, but was reverted in
bba9d0819c12 because Chromium’s test launcher did not support
GTEST_SKIP() at the time. The deficiency is on file as
https://crbug.com/912138.
While that bug was never specifically marked as “fixed” and I haven’t
found what changed in Chromium, I do now see some use of GTEST_SKIP() in
Chromium. I also prototyped this change in Chromium at
https://chromium-review.googlesource.com/c/1854691/ and found that
GTEST_SKIP() does indeed now appear to work.
Change-Id: I13fef8fe8bfd9854a40dfa5910a3282d1a85bc45
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/1855380
Reviewed-by: Joshua Peraza <jperaza@chromium.org>
Commit-Queue: Mark Mentovai <mark@chromium.org>
When all Crashpad clients have closed their crash handling sockets,
the handler's recvmsg() returns 0 and doesn't include any credentials.
Silence error logs for this normally occurring case.
Change-Id: I56acf3b38c8e95a9bbaa9bff04e0a6859a194e66
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/1816286
Reviewed-by: Mark Mentovai <mark@chromium.org>
Commit-Queue: Joshua Peraza <jperaza@chromium.org>
This change stops IsRegularFile and IsDirectory from logging
an error in the instance that a file or directory cannot be found.
Change-Id: I9f3c409933245708db775f566a27f5e49b2c71f3
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/1795924
Commit-Queue: Francois Rousseau <frousseau@google.com>
Reviewed-by: Scott Graham <scottmg@chromium.org>
This patch updates gyp_crashpad_android.py to function with NDK r20,
removes the requirement to generate a standalone toolchain, and updates
documentation on building for Android.
Also some gyp build fixes.
Change-Id: Ide338417ab2a21eca7a4bf42c1fb834e5639c186
Reviewed-on: https://chromium-review.googlesource.com/c/crashpad/crashpad/+/1798746
Reviewed-by: Mark Mentovai <mark@chromium.org>
Commit-Queue: Joshua Peraza <jperaza@chromium.org>
