NGX provider: a call waiting on the lock can hang after another call faults

Open
#75 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 1 day

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
30/100
Issue type
Bug
Clarity
Mostly clear
Activity status
Active
Tech stack
cpp

Research direction

Start in clean/rr_f/proxy/src/hook_dll.cpp at the __except handler (around line 1365, ngx_fault) and the trampolines (1401-1427), then read the mu_ lock sites in clean/rr_f/proxy/src/ngx_rr.cpp (245, 363, 613, 630). The test goes in test_hook_dll with the fake runtime, which is not yet reproduced. Done means a second thread blocked on mu_ returns within a timeout after a faulting evaluate, with a test that asserts this.

Written by the indexing model from the issue text.

Description

area: ray-reconstruction bug

This was found by reading the code. It has not been reproduced.

What happens

create, evaluate, release and release_all of the NGX provider each take a std::lock_guard<std::recursive_mutex> on mu_ (clean/rr_f/proxy/src/ngx_rr.cpp:245, 363, 613, 630).

The trampolines in clean/rr_f/proxy/src/hook_dll.cpp check g_ngx_faulted only before entering (hook_dll.cpp:1401-1427). The __except handler calls ngx_fault (hook_dll.cpp:1365), which closes the command-list bracket and sets the flag. Nothing else. The code's own comment (hook_dll.cpp:1356-1361) says the lock may stay held by the faulting thread after a structured exception, so nothing takes it any more. That holds only for calls that arrive after the flag is set.

The plugin is built with /EHsc (clean/rr_f/proxy/tools/build.py). With /EHsc, destructors of automatic objects are not guaranteed to run when an asynchronous (structured) exception is caught. So the lock_guard may not unlock. A fault inside a runtime function called through a C function-pointer table is the likely case.

A second thread already blocked on mu_ when the fault happens may then wait forever. If unwinding does release the lock, the waiter enters possibly corrupt provider state without rechecking the flag. That second fault would be caught by its own trampoline.

No other path takes mu_ after init. set_cmd_list_bracket is only called from OptiPluginInit, and the route guard uses the lock-free live().

When it can happen

Needs two things: a fault inside the provider or the runtime (itself a bug), and a second NGX call from another thread already waiting on mu_ at that moment. For example, a release on a loading or shutdown thread while the render thread evaluates. It turns a fault the design says is recoverable (the game falls back) into a game hang. Medium severity: it only follows another defect, but a hang is worse than the fallback.

To do

Preferred, lower risk:

  • In the trampolines, replace the blocking lock with a try_lock loop with Sleep(0) that exits when g_ngx_faulted becomes true.

Also:

  • In NgxRrProvider::create, evaluate and release, right after taking the lock, return the failure value if a provider-level faulted_ atomic is set (shared with g_ngx_faulted).

Alternative: in each __except handler (it runs on the faulting thread, the owner), set the flag first, then release the mutex as many times as that thread holds it. Track the depth in a thread_local.

Caution: unlocking from a handler must use the exact recursion depth, or the mutex is left inconsistent. The try_lock loop is simpler.

Test

CPU-only, in test_hook_dll (release build) with the fake runtime. Make the fake runtime's evaluate block on an event and then fault. Start a second thread that calls release through the NGX table, so it waits on mu_. Signal the event. Assert the second thread returns within a timeout.

Dominant language
C++
Stars
12
Forks
2
Avg merge
3h 10m
Merged PRs (30d)
171

Getting set up

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

More from phantoslayer21/wronggpu

All issues in phantoslayer21/wronggpu

Similar issues

More C++ issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.