-
-
Notifications
You must be signed in to change notification settings - Fork 5.5k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
free more thread state in jl_delete_thread and GC #52198
Conversation
fixed a few more things:
I also attempted to move thread-local free pages to the global page pool so they can be reused without that thread around, but I'm not sure I did it right. Help requested on that. @d-netto I also still see about 20kb of RSS growth per thread; not sure where it's coming from. On a thread whose ptls is about to be freed, I see its The goal is to get to ~0 growth for churning threads so long-lived processes are possible. |
67600e0
to
06cae4b
Compare
bump? |
1 similar comment
bump? |
Might need a rebase. |
fe737c4
to
c71f18a
Compare
c71f18a
to
a125bc2
Compare
This prevents most memory growth in workloads that start many foreign threads. In the future, we could do even better by moving pages in the heap of an exited thread (and also maybe pooled stacks) elsewhere so they can be reused, and then also free the TLS object itself.
Prior to this, especially on macOS, the gc-safepoint here would cause the process to segfault as we had already freed the current_task state. Rearrange this code so that the GC interactions (except for the atomic store to current_task) are all handled before entering GC safe, and then signaling the thread is deleted (via setting current_task = NULL, published by jl_unlock_profile_wr to other threads) is last. ``` ERROR: Exception handler triggered on unmanaged thread. Process 53827 stopped * thread #5, stop reason = EXC_BAD_ACCESS (code=2, address=0x100018008) frame #0: 0x0000000100b74344 libjulia-internal.1.12.0.dylib`jl_delete_thread [inlined] jl_gc_state_set(ptls=0x000000011f8b3200, state='\x02', old_state=<unavailable>) at julia_threads.h:272:9 [opt] 269 assert(old_state != JL_GC_CONCURRENT_COLLECTOR_THREAD); 270 jl_atomic_store_release(&ptls->gc_state, state); 271 if (state == JL_GC_STATE_UNSAFE || old_state == JL_GC_STATE_UNSAFE) -> 272 jl_gc_safepoint_(ptls); 273 return old_state; 274 } 275 STATIC_INLINE int8_t jl_gc_state_save_and_set(jl_ptls_t ptls, Target 0: (julia) stopped. (lldb) up frame #1: 0x0000000100b74320 libjulia-internal.1.12.0.dylib`jl_delete_thread [inlined] jl_gc_state_save_and_set(ptls=0x000000011f8b3200, state='\x02') at julia_threads.h:278:12 [opt] 275 STATIC_INLINE int8_t jl_gc_state_save_and_set(jl_ptls_t ptls, 276 int8_t state) 277 { -> 278 return jl_gc_state_set(ptls, state, jl_atomic_load_relaxed(&ptls->gc_state)); 279 } 280 #ifdef __clang_gcanalyzer__ 281 // these might not be a safepoint (if they are no-op safe=>safe transitions), but we have to assume it could be (statically) (lldb) frame #2: 0x0000000100b7431c libjulia-internal.1.12.0.dylib`jl_delete_thread(value=0x000000011f8b3200) at threading.c:537:11 [opt] 534 ptls->root_task = NULL; 535 jl_free_thread_gc_state(ptls); 536 // then park in safe-region -> 537 (void)jl_gc_safe_enter(ptls); 538 } ``` (test incorporated into #55793) (cherry picked from commit 0d09f3d, resolving conflicts from not having backported #52198)
This prevents most memory growth in workloads that start many foreign threads. In the future, we could do even better by moving pages in the heap of an exited thread (and also maybe pooled stacks) elsewhere so they can be reused, and then also free the TLS object itself.