This PR partially reverts 83e180c and instead opts to not define the GPU entry
point on Darwin platforms. Marking `__llvm_write_custom_profile` as used
was causing issues on embedded platforms.
In the caching part of the secondary path, when about to try to release
memory to the OS, we always wait while acquiring the lock. However, if
multiple threads are attempting this at the same time, all other threads
will likely do nothing when the release call is made. Change the
algorithm to skip the release if there is another release in process.
Also, pull the lock outside the releaseOlderThan function. This is so
that in the store path, we use the tryLock and skip if another thread is
releasing. But in the path where a forced release call is being made,
that call will wait for release to complete which guarantees that all
entries are released when requested.
Introduced a cmake option that is disabled by default that suppresses
searching via the PATH variable for a symbolizer. The option will be
enabled for downstream builds where the user will need to specify the
symbolizer path more explicitly, e.g., by using ASAN_SYMBOLIZER_PATH.
guard_acquire is a helper function used to implement TSan's
__cxa_guard_acquire and pthread_once interceptors.
https://reviews.llvm.org/D54664 introduced optional hooks to support
cooperative multi-threading. It worked by marking the entire
guard_acquire call as a potentially blocking region.
In principle, only the contended case needs to be a potentially blocking
region. This didn't matter for __cxa_guard_acquire because the compiler
emits an inline fast path before calling __cxa_guard_acquire. That is,
once we call __cxa_guard_acquire at all, we know we're in the contended
case.
https://reviews.llvm.org/D107359 then unified the __cxa_guard_acquire
and pthread_once interceptors, adding the hooks to pthread_once.
However, unlike __cxa_guard_acquire, pthread_once callers are not
expected to have an inline fast path. The fast path is inside the
function.
As a result, TSan unnecessarily calls into the cooperative
multi-threading engine on every pthread_once call, despite applications
generally expecting pthread_once to be fast after initialization. Fix
this by deferring the hooks to the contended case inside guard_acquire.
This generalizes https://github.com/llvm/llvm-project/pull/131975 to non-32-bit Linux (i.e., 64-bit Linux).
This works around an edge case in 64-bit Linux, whereby the memory layout is incompatible if the stack size is unlimited AND ASLR entropy is 31+ bits (see https://github.com/google/sanitizers/issues/856#issuecomment-2747076811).
More generally, this "re-exec without ASLR if layout is incompatible" is a hammer that can work around most shadow mapping issues, without incurring the overhead of using a dynamic shadow.
Function definition added in #131975 was missing 'Try' (do or do not ...). My guess is buildbots
mostly didn't trip because usage was gated by 'if (sizeof(uptr) ==
32)', which is rare among buildbots.
A bare-bones version of LLVM's unique_function: this behaves like a
std::unique_function, except that it supports move only callable types.
This will be used in upcoming improvements to the ORC runtime.
The "extensible_" prefix on these files was inherited from LLVM, where it
distinguished the dynamic RTTI APIs from the LLVM's custom static RTTI APIs.
In the ORC runtime these files will be used to hold all of our RTTI APIs
(the current dynamic ones, and any static ones added in the future), so we
shouldn't use this prefix.
Observed in Wine when trying to intercept `ExitThread`, which forwards
to `ntdll.RtlExitUserThread`.
`gdb` interprets it as `xchg %ax,%ax`.
`llvm-mc` outputs simply `nop`.
```
==Asan-i386-calls-Dynamic-Test.exe==964==interception_win: unhandled instruction at 0x7be27cf0: 66 90 55 89 e5 56 50 8b
```
```
Wine-gdb> bt
#0 0x789a1766 in __interception::GetInstructionSize (address=<optimized out>, rel_offset=<optimized out>) at C:/llvm-mingw/llvm-mingw/llvm-project/compiler-rt/lib/interception/interception_win.cpp:983
#1 0x789ab480 in __sanitizer::SharedPrintfCode(bool, char const*, char*) () at C:/llvm-mingw/llvm-mingw/llvm-project/compiler-rt/lib/sanitizer_common/sanitizer_printf.cpp:311
#2 0x789a18e7 in __interception::OverrideFunctionWithHotPatch (old_func=2078440688, new_func=2023702608, orig_old_func=warning: (Internal error: pc 0x792f1a2c in read in CU, but not in symtab.)warning: (Error: pc 0x792f1a2c in address map, but not in symtab.)0x792f1a2c) at C:/llvm-mingw/llvm-mingw/llvm-project/compiler-rt/lib/interception/interception_win.cpp:1118
#3 0x789a1f34 in __interception::OverrideFunction (old_func=2078440688, new_func=2023702608, orig_old_func=warning: (Internal error: pc 0x792f1a2c in read in CU, but not in symtab.)warning: (Error: pc 0x792f1a2c in address map, but not in symtab.)0x792f1a2c) at C:/llvm-mingw/llvm-mingw/llvm-project/compiler-rt/lib/interception/interception_win.cpp:1224
#4 0x789a24ce in __interception::OverrideFunction (func_name=0x78a0bc43 <vtable for __asan::AsanThreadContext+1163> "ExitThread", new_func=2023702608, orig_old_func=warning: (Internal error: pc 0x792f1a2c in read in CU, but not in symtab.)warning: (Error: pc 0x792f1a2c in address map, but not in symtab.)0x792f1a2c) at C:/llvm-mingw/llvm-mingw/llvm-project/compiler-rt/lib/interception/interception_win.cpp:1369
#5 0x789f40ef in __asan::InitializePlatformInterceptors () at C:/llvm-mingw/llvm-mingw/llvm-project/compiler-rt/lib/asan/asan_win.cpp:190
#6 0x789e0c3c in __asan::InitializeAsanInterceptors () at C:/llvm-mingw/llvm-mingw/llvm-project/compiler-rt/lib/asan/asan_interceptors.cpp:802
#7 0x789ee6b5 in __asan::AsanInitInternal () at C:/llvm-mingw/llvm-mingw/llvm-project/compiler-rt/lib/asan/asan_rtl.cpp:442
#8 0x789eefb0 in __asan::AsanInitFromRtl () at C:/llvm-mingw/llvm-mingw/llvm-project/compiler-rt/lib/asan/asan_rtl.cpp:522
#9 __asan::AsanInitializer::AsanInitializer (this=<optimized out>) at C:/llvm-mingw/llvm-mingw/llvm-project/compiler-rt/lib/asan/asan_rtl.cpp:542
#10 __cxx_global_var_init () at C:/llvm-mingw/llvm-mingw/llvm-project/compiler-rt/lib/asan/asan_rtl.cpp:546
...
Wine-gdb> disassemble /r 2078440688,2078440688+20
Dump of assembler code from 0x7be27cf0 to 0x7be27d04:
0x7be27cf0 <_RtlExitUserThread@4+0>: 66 90 xchg %ax,%ax
...
```
Mechanism to keep the compiler-rt and llvm view of `FunctionData` in sync. Since CtxInstrContextNode.h is exactly the same on both sides (there's an existing test, `compiler-rt/test/ctx_profile/TestCases/check-same-ctx-node.test`, checking that), we capture the structure in a macro that is then generated as `struct` fields on the compiler-rt side, and as `Type` objects on the llvm side. The macro needs to be told how to render a few kinds of fields.
If we add more fields to FunctionData that can be described by the current known types of fields, then the llvm side would automatically be updated. If we need to add more kinds of fields, which we do by adding parameters to the macro, the llvm side (if not updated) would trigger a compilation error.
On AIX, when accessing mmap'ed memory associated to a file on NFS, a
SIGBUS might be raised at random.
The problem is still in open state with the OS team.
This PR teaches the profile runtime, under certain conditions, to avoid
the mmap when reading the profile file during online merging.
This PR has no effect on any platform other than AIX because I'm not
aware of this problem on other platforms.
Other platforms can easily opt-in to this functionality in the future.
The logic in function `is_local_filesystem` was copied from
[llvm/lib/Support/Unix/Path.inc](f388ca3d9d/llvm/lib/Support/Unix/Path.inc (L515))
(https://reviews.llvm.org/D58801), because it seems that the
compiler-rt/profile cannot reuse code from llvm except through
`InstrProfData.inc`.
Thanks to @hubert-reinterpretcast for substantial feedback downstream.
---------
Co-authored-by: Wael Yehia <wyehia@ca.ibm.com>
Co-authored-by: Hubert Tong <hubert.reinterpretcast@gmail.com>
High-entropy ASLR allows up to 16-bits of entropy (2**16 4KB pages ==
256MB; a bit more in practice because of implementation details), which
is a significant chunk of the user address space on 32-bit systems (4GB
or less). This, combined with ASan's shadow (512MB) and ASan's fixed
shadow offset (512MB), makes it possible for large binaries to fail to
map the shadow.
This patch changes ASan to do a one-time re-exec without ASLR if it
cannot map the shadow, thus reclaiming the ~256MB of address space.
Alternatives considered:
1) We don't lower ASan's fixed shadow offset, because that would limit
non-PIE binaries.
2) We don't switch to a dynamic shadow offset, because ASan for 32-bit
Linux relies on the compile-time constant offset to optimize its
instrumentation and compiler-rt.
This is loosely inspired by
https://github.com/llvm/llvm-project/pull/78351,
https://github.com/llvm/llvm-project/pull/85142, and
https://github.com/llvm/llvm-project/pull/85674, though those were
required because there were no static shadow mappings that could fully
shadow the range of user mappings; this is not the case for ASan.
This pull request is the third part of an ongoing effort to extends PGO
instrumentation to GPU device code and depends on
https://github.com/llvm/llvm-project/pull/93365. This PR makes the
following changes:
- Allows PGO flags to be supplied to GPU targets
- Pulls version global from device
- Modifies `__llvm_write_custom_profile` and `lprofWriteDataImpl` to
allow the PGO version to be overridden
Collect profiles for functions we encounter when collecting a contextual profile, that are not associated with a call site. This is expected to happen for signal handlers, but it also - problematically - currently happens for mem{memset|copy|move|set}, which are currently inserted after profile instrumentation.
Collecting a "regular" flat profile in these cases would hide the problem - that we loose better profile opportunities.
`ContextRoot` `FunctionData` are currently known by the llvm side, which has to instantiate and zero-initialize them.
This patch makes `FunctionData` the only global value that needs to be known and instantiated by the compiler. On the compiler-rt side, `ContextRoot`s are hung off `FunctionData`, when applicable.
This is for two reasons. First, it is a step towards root autodetection (in a subsequent patch). An autodetection mechanism would instantiate the `ContextRoot` for the detected roots, and then `__llvm_ctx_profile_get_context` would detect that and route to `__llvm_ctx_profile_start_context`.
The second reason is that we will hang off `ContextRoot` more complex datatypes (next patch), and we want to avoid too deep of a coupling between llvm and compiler-rt. Acting as a place to hang related data, `FunctionData` can stay simple - pointers and an (atomic) int (the mutex).
On AIX, when accessing mmap'ed memory associated to a file on NFS, a
SIGBUS might be raised at random.
The problem is still in open state with the OS team.
This PR teaches the profile runtime, under certain conditions, to avoid
the mmap when reading the profile file during online merging.
This PR has no effect on any platform other than AIX because I'm not
aware of this problem on other platforms.
Other platforms can easily opt-in to this functionality in the future.
The logic in function `is_local_filesystem` was copied from
[llvm/lib/Support/Unix/Path.inc](f388ca3d9d/llvm/lib/Support/Unix/Path.inc (L515))
(https://reviews.llvm.org/D58801), because it seems that the
compiler-rt/profile cannot reuse code from llvm except through
`InstrProfData.inc`.
Thanks to @hubert-reinterpretcast for substantial feedback downstream.
---------
Co-authored-by: Wael Yehia <wyehia@ca.ibm.com>
Co-authored-by: Hubert Tong <hubert.reinterpretcast@gmail.com>
When we collect a contextual profile, we sample the threads entering its root and only collect on one at a time (see `ContextRoot::Taken`). If we want to compare profiles between contextual profiles, and/or flat profiles, we have a problem: we don't know how to compare the counter values relative to each other. To that end, we add `ContextRoot::TotalEntries`, which is incremented every time a root is entered and serves as multiplier for the counter values collected under that root.
We expose this in the profile and leave the normalization to the user of the profile, for a few reasons:
* it's only needed if reasoning about all profiles in aggregate.
* the goal, in compiler_rt, is to flush out the profile as quickly as possible, and performing multiplications adds an overhead that may not even be necessary if the consumer of the profile doesn't care about combining profiles
* the information itself may be interesting as an indication of relative sampling of various contexts.
Currently, the code model check is always performed even if there are no
globals, because:
1) the HWASan compiler pass always leaves a note
2) the HWASan runtime always performs the check if there is a HWASan
globals note.
This unnecessarily adds a 2**32 byte size limit.
This patch elides the check if the globals note doesn't actually contain
globals, thus allowing larger libraries to be successfully instrumented
without globals.
Sent from my iPhone
Follows the discussion here:
https://github.com/llvm/llvm-project/pull/129309
Recently, the test
`TestRtsan.AccessingALargeAtomicVariableDiesWhenRealtime` has been
failing on newer MacOS versions, because the internal locking mechanism
in `std::atomic<T>::load` (for types `T` that are larger than the
hardware lock-free limit), has changed to a function that wasn't being
intercepted by rtsan.
This PR introduces an interceptor for `_os_nospin_lock_lock`, which is
the new internal locking mechanism.
_Note: we'd probably do well to introduce interceptors for
`_os_nospin_lock_unlock` (and `os_unfair_lock_unlock`) too, which also
appear to have blocking implementations. This can follow in a separate
PR._
Collect flat profiles. We only do this for function activations that aren't otherwise collectible under a context root are encountered.
This allows us to reason about the full profile without concerning ourselves wether we are double-counting. For example we can combine (during profile use) flattened contextual profiles with flat profiles.
The condition to stop iterating so far was to look for load command cmd
field == 0. The iteration would continue past the commands area, and
would finally find lc->cmd ==0, if lucky. Or crash with bus error, if
out of luck.
Correcting this by limiting the number of iterations to the count
specified in mach_header(_64) ncmds field.
rdar://143903403
---------
Co-authored-by: Mariusz Borsa <m_borsa@apple.com>
The profile format has now a separate section called "Contexts" - there will be a corresponding one for flat profiles. The root has a separate tag because, in addition to not having a callsite ID as all the other context nodes have under it, it will have additional fields in subsequent patches.
The rest of this patch amounts to a bit of refactorings in the reader/writer (for better reuse later) and tests fixups.
Introduce a `ProfileWriter` abstraction to replace the callback passed to `__llvm_ctx_profile_fetch`. Subsequent changes will add support for flat profile collection (as in, collection of non-contextual profile for those functions not under a contextual root), which require also a change in the profile format. The abstraction makes it easy to add "write flat" - related capabilities without constantly complicating the signature of `__llvm_ctx_profile_fetch`.
Fix#126541
Since ```t->Destroy``` cannot be called after ```start_routine```(When
calling standard thread_start in crt)
Intercept `ExitThread` and free the memory created by `VirtualAlloc'
