Skip to content

[WIP] Implement remaining APIs for runtime issue 123819#1

Closed
Copilot wants to merge 1 commit intoissue-75348from
copilot/implement-remaining-apis
Closed

[WIP] Implement remaining APIs for runtime issue 123819#1
Copilot wants to merge 1 commit intoissue-75348from
copilot/implement-remaining-apis

Conversation

Copy link

Copilot AI commented Jan 31, 2026

No description provided.

Copilot AI requested a review from jgh07 January 31, 2026 13:58
Copilot stopped work on behalf of jgh07 due to an error January 31, 2026 13:58
@jgh07 jgh07 closed this Jan 31, 2026
@jgh07 jgh07 deleted the copilot/implement-remaining-apis branch January 31, 2026 14:07
jgh07 pushed a commit that referenced this pull request Feb 15, 2026
…er (dotnet#123735)

From discussion, opting into enabling the crash chaining is more
correct.

<s>The previously registered signal action/handler aren't guaranteed to
return, so we lose out on notifying shutdown and creating a dump in
those cases. Specifically, PROCCreateCrashDumpIfEnabled would be the
last chance to provide the managed context for the thread that crashed.

e.g. On Android CoreCLR, it seems that, by default, signal handlers are
already registered by Android's runtime
(/apex/com.android.runtime/bin/linker64 +
/system/lib64/libandroid_runtime.so). Whenever an unhandled synchronous
fault occurs, the previously registered handler will not return back to
invoke_previous_action and aborts the thread itself, so
PROCCreateCrashDumpIfEnabled will not be hit.</s>

## Sigsegv behavior Android CoreCLR vs other platforms

### Android CoreCLR
When intentionally writing to NULL (sigsegv) on Android CoreCLR, the
previously registered signal handler goes down this path
https://github.com/dotnet/runtime/blob/40e8c73b8f3b5f478a9bf03cf55c71d0608a8855/src/coreclr/pal/src/exception/signal.cpp#L454,
and the thread aborts before hitting PROCNotifyProcessShutdown and
PROCCreateCrashDumpIfEnabled.

### MacOS/Linux/NativeAOT(linux)
On MacOS, Linux, NativeAOT (Only checked linux at time of writing), the
same intentional SIGSEGV will hit
https://github.com/dotnet/runtime/blob/40e8c73b8f3b5f478a9bf03cf55c71d0608a8855/src/coreclr/pal/src/exception/signal.cpp#L431-L448
instead because there is no previously registered signal handler. In
those cases, PROCCreateCrashDumpIfEnabled is hit and managed callstacks
are captured in the dump.

## History investigation

From a github history dive, I didn't spot anything in particular
requiring the previous signal handler to be invoked before
PROCNotifyProcessShutdown + PROCCreateCrashDumpIfEnabled.

PROCNotifyProcessShutdown was first introduced in
dotnet@1433c3f.
It doesn't seem to state a particular reason for invoking it after the
previous signal handler.

PROCCreateCrashDumpIfEnabled was added to signal.cpp in
dotnet@7f9bd2c
because the PROCNotifyProcessShutdown didn't create a crash dump. It
doesn't state any particular reason for being invoked after the
previously registered signal handler, and was probably just placed next
to PROCNotifyProcessShutdown.

`invoke_previous_action` was introduced in
dotnet@a740f65
and was refactoring while maintaining the order.

## Android CoreCLR behavior after swapping order

Locally, I have POC changes to emit managed callstacks in Android's
PROCCreateCrashDumpIfEnabled.
```
01-28 17:26:40.951  2416  2440 F DOTNET  : Native crash detected; attempting managed stack trace.
01-28 17:26:40.951  2416  2440 F DOTNET  : {"stack":[
01-28 17:26:40.951  2416  2440 F DOTNET  : {"ip":"0x0","module":"0x0","offset":"0x0","name":"Program.MemSet(Void*, Int32, UIntPtr)"},
01-28 17:26:40.951  2416  2440 F DOTNET  : {"ip":"0x78d981145973","module":"0x0","offset":"0x0","name":"Program.MemSet(Void*, Int32, UIntPtr)"},
01-28 17:26:40.951  2416  2440 F DOTNET  : {"ip":"0x78d981145973","module":"0x0","offset":"0x73","name":"Program.ForceNativeSegv()"},
01-28 17:26:40.951  2416  2440 F DOTNET  : {"ip":"0x78d981141b60","module":"0x0","offset":"0x70","name":"Program.Main(System.String[])"}
01-28 17:26:40.951  2416  2440 F DOTNET  : ]}
01-28 17:26:40.952  2416  2440 F DOTNET  : Crash dump hook completed.
--------- beginning of crash
01-28 17:26:40.952  2416  2440 F libc    : Fatal signal 11 (SIGSEGV), code 1 (SEGV_MAPERR), fault addr 0x0 in tid 2440 (.dot.MonoRunner), pid 2416 (ulator.JIT.Test)
.....
01-28 17:26:46.882  2921  2921 F DEBUG   : *** *** *** *** *** *** *** *** *** *** *** *** *** *** *** ***
01-28 17:26:46.882  2921  2921 F DEBUG   : Build fingerprint: 'google/sdk_gphone64_x86_64/emu64xa:16/BE2A.250530.026.D1/13818094:user/release-keys'
01-28 17:26:46.882  2921  2921 F DEBUG   : Revision: '0'
01-28 17:26:46.882  2921  2921 F DEBUG   : ABI: 'x86_64'
01-28 17:26:46.882  2921  2921 F DEBUG   : Timestamp: 2026-01-28 17:26:41.492831700-0500
01-28 17:26:46.882  2921  2921 F DEBUG   : Process uptime: 20s
01-28 17:26:46.883  2921  2921 F DEBUG   : Cmdline: net.dot.Android.Device_Emulator.JIT.Test
01-28 17:26:46.883  2921  2921 F DEBUG   : pid: 2416, tid: 2440, name: .dot.MonoRunner  >>> net.dot.Android.Device_Emulator.JIT.Test <<<
01-28 17:26:46.883  2921  2921 F DEBUG   : uid: 10219
01-28 17:26:46.883  2921  2921 F DEBUG   : signal 11 (SIGSEGV), code 1 (SEGV_MAPERR), fault addr 0x0000000000000000
01-28 17:26:46.883  2921  2921 F DEBUG   : Cause: null pointer dereference
01-28 17:26:46.883  2921  2921 F DEBUG   : Abort message: 'CoreCLR: previous handler for '
01-28 17:26:46.883  2921  2921 F DEBUG   :     rax 0000000000000000  rbx 000078da87ffade0  rcx 0000000000000000  rdx 0000000000000001
01-28 17:26:46.884  1237  1297 I s.nexuslauncher: AssetManager2(0x78dd08cd9178) locale list changing from [] to [en-US]
01-28 17:26:46.903  2447  2594 I BugleNotifications: Creating notification input ids [CONTEXT im_entry_input="" im_notification_input="" im_settings_store_input="" im_final_input="" ]
01-28 17:26:46.905  2921  2921 F DEBUG   :     r8  00007ffcde5a8080  r9  34d9bb0e67871eb0  r10 000078ddb4111870  r11 0000000000000293
01-28 17:26:46.906  2921  2921 F DEBUG   :     r12 0000000000000001  r13 000078da87ffafa0  r14 0000000000000000  r15 000078da87ffaf18
01-28 17:26:46.906  2921  2921 F DEBUG   :     rdi 0000000000000000  rsi 0000000000000000
01-28 17:26:46.906  2921  2921 F DEBUG   :     rbp 000078da87ffac40  rsp 000078da87ffabc8  rip 000078ddb41118a2
01-28 17:26:46.906  2921  2921 F DEBUG   : 2 total frames
01-28 17:26:46.906  2921  2921 F DEBUG   : backtrace:
01-28 17:26:46.906  2921  2921 F DEBUG   :       #00 pc 000000000008f8a2  /apex/com.android.runtime/lib64/bionic/libc.so (memset_avx2+50) (BuildId: fcb82240218d1473de1e3d2137c0be35)
01-28 17:26:46.906  2921  2921 F DEBUG   :       #1 pc 0000000000049972  /memfd:doublemapper (deleted) (offset 0x111000)
```
Now theres a window to log managed callstacks before the original signal
handler aborts and triggers a tombstone.

## Android Mono behavior

Mono provides two embeddings APIs to configure signal and crash chaining
https://github.com/dotnet/runtime/blob/61d3943de41e948bb0ecf871b92eb456d2dd74d8/src/mono/mono/mini/driver.c#L2864-L2894
that determine whether synchronous faults would chain
https://github.com/dotnet/runtime/blob/61d3943de41e948bb0ecf871b92eb456d2dd74d8/src/mono/mono/mini/mini-runtime.c#L3892-L3903
They would only chain to the previous signal handler
https://github.com/dotnet/runtime/blob/61d3943de41e948bb0ecf871b92eb456d2dd74d8/src/mono/mono/mini/mini-posix.c#L193-L210
only after attempting to walk native and managed stacks
https://github.com/dotnet/runtime/blob/61d3943de41e948bb0ecf871b92eb456d2dd74d8/src/mono/mono/mini/mini-exceptions.c#L2992-L3012

## Alternatives

If there is any particular reason to preserve the order of
sa_sigaction/sa_handler with respect to PROCNotifyProcessShutdown and
PROCCreateCrashDumpIfEnabled for CoreCLR, a config knob can be added to
allow Android CoreCLR to opt into the swapped ordering behavior. This
may be in the form of config property key/values
https://github.com/dotnet/runtime/blob/54ca569eb62800cdb725d776e3dd2e564028594d/src/coreclr/dlls/mscoree/exports.cpp#L237-L238
or `clrconfigvalues`. That way AndroidSDK/AndroidAppBuilder may opt-in
at build-time.

Given that the history of the ordering didn't reveal any problems with
swapping the order, we can fallback to this behavior if the order swap
causes problems down the line.

The other way around is more restrictive. Should we first introduce all
the overhead to enable an opt-in/opt-out config knob, and later discover
that no platforms need to invoke their previous handlers before
PROCNotifyProcessShutdown/PROCCreateCrashDumpIfEnabled, it seems harder
to justify removing the knob.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants