The creation mutex could have been held if the parent thread was in the
middle of creating a thread when forking. This would result in a
deadlock once the fork child attempted to create another thread.
Forcefully dropping the lock in the fork child works around this
deadlock. This comes at the expense of potentially leaving resources
guarded by the thread creation mutex in an invalid state. Crashes caused
by this are easier to reason about than a delayed deadlock, though.
Moves the CTX LockBeforeFork in to the Syscallhandler's LockBeforeFork.
This lets the syscall handler just call its own LockBeforeFork and
UnlockAfterFork functions rather than two on each call site.
Also moves the CTX->UnlockAfterFork in to the SyscallHandler's to be
consistent with the LockBeforeFork half.
No functional change.
When code invalidation is happening we currently have the issue that a
thread can acquire the code invalidation mutex in the middle of
invalidation. This is due to us acquiring and releasing the mutex
between each thread's code invalidation.
We need to hold the mutex for the entire duration for all thread's code
invalidation.
This fixes a rare hang on proton startup and resolves a consistent hang
on Proton application shutdown.
This now puts us on par with FEX-2312.1 with hanging.
This does not fix a relatively rare hang on fork (which also existed with FEX-2312.1).
This also does not fix the issue that the intersection of our mutexes
between frontend and backend are very convoluted. In part of the work
that is going to fix the rare fork mutex hang will change more of this.
If the thread object is added to the tracking vector immediately then
there ends up being a race condition before the thread manages to fill
out the thread-specific data that only occurs at the start of the new
thread.
This manifests in a crash when a thread is allocating memory while
another thread is getting constructed. Easy fix is to defer the tracking
until the thread has setup its state.
This is used for instcountci to ensure instruction counts don't change
when a compiler supports this feature or not. Always runtime disable
when running in instcountci.
CMake option from #3394 can still be useful so leaving that in place.
We are required in our syscall emulation to handle cases where pointers
are invalid. This means we need to pessimistically assume a memcpy will
fault when reading application memory.
This implements a signal handler based approach to catching the SIGSEGV
on memcpy and returning an EFAULT if it faults.
This behaves exactly like pidof but only searches for FEX applications.
This fixes a long standing annoyance of mine that pidof doesn't work for
FEX. This behaves exactly like pidof but knows how to decode the command
line options to pull out the program data.
If the Linux kernel ever accepts the patches for binfmt_misc to change
how the interpreter is handled then this will become redundant, but
until that happens here is a utility that I want.
This unit test hasn't really served any purpose for a while now and
mostly just causes pain when reworking things in the IR.
Just remove the IRLoader, its unit tests, the github action steps and
the public FEXCore interface to it. Since it isn't used by anything
other than Thunks.
Also moves some IR definitions from the public API to the backend.
We only used this so that our Xavier CI system which were running old
kernels could run unit tests. We have now removed the Xaviers from CI
and this is no longer necessary.
Stop pretending that we support kernels older than 5.0 and allowing this
fallback.
The 32-bit allocator is still used for the MAP_32BIT mmap flag, so the
load bearing code can't be fully removed. Just remove the config and the
frontend things using it.
Currently no functional change but public API breaks should come early.
The thread state object will be used for looking up thread specific
codebuffers in the future when we support MDWE with code mirrors.
Previously we were only storing the 32-bit base address which isn't
actually how segment descriptors work.
In reality segment descriptors are 64-bit descriptors that are laid out
in a particular layout depending on the 4-bit type value. In reality we
only care about code and data segment layouts since the rest are
bonkers.
Describe these descriptors correctly and setup a default code descriptor
for the operating mode that FEX is starting in.
Lots going on here.
This moves OS thread object lifetime management and internal thread
state lifetime management to the frontend. This causes a bunch of thread
handling to move from the FEXCore Context to the frontend.
Looking at `FEXCore/include/FEXCore/Core/Context.h` really shows how
much of the API has moved to the frontend that FEXCore no longer needs
to manage. Primarily this makes FEXCore itself no longer need to care
about most of the management of the emulation state.
A large amount of the behaviour moved wholesale from Core.cpp to
LinuxEmulation's ThreadManager.cpp. Which this manages the lifetimes of
both the OS threads and the FEXCore thread state objects.
One feature lost was the instruction capability, but this was already
buggy and is going to be rewritten/fixed when gdbserver work continues.
Now that all of this management is moved to the frontend, the gdbserver
can start improving since it can start managing all thread state
directly.
FEX was incorrectly invalidating RBX and RBP when a new
thread/clone/fork was created. Only RAX should be set to zero to
signify that it is the child process.
I tried to find some reference as to why this was being done in FEX, and
I seem to recall some code somewhere that was setting up this state.
Sadly I can't seem to find any reference to this being required now by
the linux kernel when it creates new processes, so maybe I was just
gaslighting myself.
Fixes crash in bubblewrap where it was expecting a valid RBP to exist
after it does a clone syscall.
I have unit tests coming for this but it is uncovering some more bugs in
FEX's thread handling. Some of which are getting fixed with the current
PRs that are moving the thread handling to the frontend.
But since proton is broken now, this needs to get in.
This was left over and was using the wrong argument.
This would cause the stack for the child thread/process to be at the
start instead of the end.
Fixes a crash in bubblewrap.
Similar to #3284 but works around some of the bugs that one introduced.
This is the minimal amount of changes to move the ownership from FEXCore
to the frontend. Since the frontends don't yet have a full thread state
tracking, there is an opaque pointer that needs to be managed.
In the followup commits this will be changed to have the syscall handler
to be the thread object manager.
Now that only FEXLoader is under the FEXLoader folder, the mingw build
check can be moved up one level.
No functional change, just moving the check and re-aligning. Might be
good to view without whitespace changes.
Which requires moving LinuxEmulation to its own independent folder as
well. Since both IRLoader and FEXLoader rely on it.
No functional change, just moves the the code around.
When originally written I was testing symbol loading on an x86 host
assuming that the vdso symbol names were the same between x86 and
aarch64.
Turns out this is not true and aarch64 uses `__kernel_` prefix instead
of x86's `__vdso_`.
AArch64 VDSO only provides us with four symbols right now
- __kernel_rt_sigreturn
- __kernel_gettimeofday
- __kernel_clock_gettime
- __kernel_clock_getres
This means that currently AArch64 is missing getcpu and time.
time is unlikely to ever be implemented so the vdso scanning for that is
unlikely to ever get a pointer and will need to go down the glibc path.
getcpu is likely to get an implementation at some point. There's even an
old patch series that implemented it[1]. Currently the patch has been
dropped on the floor for whatever reason but hopefully they will come
back to it.
Benchmarks on Cortex-X1C:
- Native AArch64 clock_gettime bench
- VDSO: 33.3ns per call
- glibc (which uses vdso): 33.3ns per call
- Roughly equivalent
- Emulated x86_64 clock_gettime bench
- Prior to fix:
- 49ns per call for both VDSO and glibc
- After fix:
- VDSO: ~46ns per call
- glibc: ~49ns per call
- Emulated i386 clock_gettime bench
- Only tested after fix
- VDSO: 50ns per call
- glibc: 65ns per call
The only way to improve this further would be to optimize thunks to push
certain function signature arguments in registers instead of stack.
Which is a future optimization as time goes on.
[1] https://patchwork.kernel.org/project/linux-kselftest/list/?series=335203
Ryan Houdek