When I added %m to async_safe_* too, we never followed up and cleaned up
callers.
Test: treehugger
Change-Id: If81943c4c45de49f0fb4bc29cfbd3fc53d4a47fe
Actually, we don't want to reuse the kernel struct ucontext because its
uc_mcontext has the wrong type, which means the fields within that end
up with the wrong names. Add the call site that made that evident, and
update <sys/ucontext.h> appropriately.
Signed-off-by: Mao Han <han_mao@linux.alibaba.com>
Signed-off-by: Xia Lifang <lifang_xia@linux.alibaba.com>
Signed-off-by: Chen Guoyin <chenguoyin.cgy@linux.alibaba.com>
Signed-off-by: Wang Chen <wangchen20@iscas.ac.cn>
Signed-off-by: Lu Xufan <luxufan@iscas.ac.cn>
Test: treehugger
Change-Id: If1d079afef0d5953aa22d9b0e049cfb0119c7718
For the perf profiling signal handler to succeed in opening
/proc/self/mem, the process needs to be marked as dumpable in posix
terms. This patch addresses a scenario since Android S where the process
is considered profileable, but is not dumpable on "user" builds. The
solution is to mark the process as dumpable while opening the procfs
descriptors, restoring the original value afterwards. This is the same
approach as the heapprofd heap profiler, which performs the override
within the loaded client library [1].
The particular scenario being addressed is:
* user build
* app does not explicitly opt into being profiled by shell
* app does not explicitly opt out of all profiling
In this case, the app is considered profileable by the platform (but NOT
shell). Therefore ActivityThread marks the process as profileable [2],
but the zygote keeps the process as undumpable as it considers the
profileability from the shell domain [3]. We could change the logic in
the zygote to leave such processes in the dumpable state, but the
override within the signal handler is considered to be more contained as
the dumpability is only needed temporarily.
This override would also apply for any non-dumpable native services that
are signalled for profiling, which is also desireable for profiling
coverage.
This change does not elide any of the existing profileability
checks by the signal handler's preamble and the profiler itself.
[1]
https://cs.android.com/android/platform/superproject/+/master:external/perfetto/src/profiling/memory/client.cc;l=184;drc=78cd82ba31233ce810618e07d349fd34efdb861d
[2]
https://cs.android.com/android/platform/superproject/+/master:frameworks/base/core/java/android/app/ActivityThread.java;l=6610;drc=de9cf3392d7872c2bee69b65a614e77bb166b26e
[3]
https://cs.android.com/android/platform/superproject/+/master:frameworks/base/core/jni/com_android_internal_os_Zygote.cpp;l=1680;drc=master
Tested: clock app on barbet-user succeeds in opening the procfs
descriptors within the signal handler.
Tested: systemwide profiling on sargo-userdebug works as before.
Bug: 196810669
BYPASS_INCLUSIVE_LANGUAGE_REASON=referencing the name of a cmdline utility
Change-Id: Id621d4312418ff0736c97065e9ee577ff67f40da
Strictly this still isn't quite the same, because they won't actually be
profiled, but at least they won't *crash* now if they're sent a
profiling signal.
Bug: http://b/201497662
Test: treehugger
Change-Id: I0728492eed77584cd850d28897056996387e6671
These solves the issue that targetting non-profilable apps crashed them.
There is still a race condition between starting the app and the SIG_IGN
being installed, but that will be fixed in follow-ups.
This also does not cover programs that statically link libc, but those are
rare.
This might be reverted if we find a more general solution to
b/151835887.
Bug: 151328035
Test: java profile sysui, doesn't crash
Test: atest CtsBionicTestCases
Change-Id: I6b99352ed50afe15a609f7ddb85312c2676ddf11
This is a best-effort mitigation for potential crashes when the
profiling signal handler is triggered within certain secomp'd processes.
In particular, we're working around cases where the seccomp policy
doesn't allow some of the handler syscalls, and has a crashing
disposition towards violations via SECCOMP_RET_TRAP, plus a crashing
SIGSYS handler. While not general, this covers the configurations seen
in practice on Android (which are all using minijail in the same way).
By overriding the SIGSYS handling for the duration of the profiling
handler, we can instead receive such SIGSYS signals, and instead recover
from them in a non-crashing manner (the handler is responsible for
filling the syscall return register, since the syscall itself was
skipped).
For simplicity, we're swallowing all SIGSYS signals during this window,
without trying to figure out whether they're something that could be
caused by the profiling signal handler. I've quite convinced myself that
-ENOSYS seems to be safe to return to all of bionic's syscall wrappers
across the four architectures (looking at gensyscalls + the
special-cases like vfork and clone). It is theoretically possible for
all kinds of conflicting (ab)uses of SIGSYS to exist, but I'm assuming
it's not a realistic concern until proven otherwise.
Tested: manually sigqueue'd configstore on crosshatch, confirmed that
the SIGSYS override log was printed, and the process did not
crash (as it does on master).
Bug: 149328505
Change-Id: Iab8f09e51169807c9d3e1e0bcfd042f09f7df6a4
This patch adds a case for the profiling signal handler (previously just
for native heapprofd profiling) when si_value == 1, corresponding to
traced_perf being the requesting party.
The handler opens /proc/self/{maps,mem}, connects to (init-created)
/dev/socket/traced_perf, and then sends the fds over the socket.
Everything happens synchronously within the signal handler. Socket is
made non-blocking, and we do not retry.
Bug: 144281346
Change-Id: Iea904694caeefe317ed8818e5b150e8819af91c2
This patch refactors heapprofd_malloc to make it easier to reuse the
reserved signal for multiple purposes. We define a new generic signal
handler for profilers, which dispatches to more specific logic based on
the signal's payload (si_value).
The profiler signal handler is installed during libc preinit, after
malloc initialization (so races against synchronous heapprofd
initialization need not be considered). In terms of code organization, I
copied the existing approach with a loosely referenced function in
bionic_globals.h. Do tell if you'd rather a different approach here.
The profileability of a process is quite tied to the malloc
files/interfaces in bionic - in particular, it's set through
android_mallopt. I do not change that, but instead introduce a new
android_mallopt option to be able to query profileability of the
process (which is now used by the new profiler signal handler). As part
of that, gZygoteChildProfileable is moved from heapprofd_malloc to
common (alongside gZygoteChild).
I've removed the masking and reraising of the heapprofd signal when
racing against malloc_limit init. We're ok with taking a simpler
approach and dropping the heapprofd signal in such an unlikely race.
Note: this requires a corresponding change in heapprofd to use sigqueue()
instead of kill(), as the latter leaves the si_value uninitialized(?) on
the receiving side.
Bug: 144281346
Change-Id: I93bb2e82cff5870e5ca499cf86439860aca9dfa5
