X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=rts%2Fsm%2FGC.c;h=0d2ba85860fd36a996af6e69dceff1a820c8e84f;hb=06f9b7c365fb9e9b53723f892b4d63b4f7a56e9a;hp=eed2da702079bd5397b6c72892464bfc12c5c3b7;hpb=f2ca6deece1ed1724efdd6d2293dc7642059b2f2;p=ghc-hetmet.git diff --git a/rts/sm/GC.c b/rts/sm/GC.c index eed2da7..0d2ba85 100644 --- a/rts/sm/GC.c +++ b/rts/sm/GC.c @@ -40,6 +40,7 @@ #include "RetainerProfile.h" #include "RaiseAsync.h" #include "Sparks.h" +#include "Papi.h" #include "GC.h" #include "Compact.h" @@ -120,9 +121,12 @@ nat mutlist_MUTVARS, gc_thread *gc_threads = NULL; // gc_thread *gct = NULL; // this thread's gct TODO: make thread-local +// Number of threads running in *this* GC. Affects how many +// step->todos[] lists we have to look in to find work. +nat n_gc_threads; + // For stats: long copied; // *words* copied & scavenged during this GC -long scavd_copied; // *words* copied only during this GC #ifdef THREADED_RTS SpinLock recordMutableGen_sync; @@ -182,10 +186,12 @@ GarbageCollect ( rtsBool force_major_gc ) step *stp; lnat live, allocated; lnat oldgen_saved_blocks = 0; - nat n_threads; // number of threads participating in GC - + gc_thread *saved_gct; nat g, s, t; + // necessary if we stole a callee-saves register for gct: + saved_gct = gct; + #ifdef PROFILING CostCentreStack *prev_CCS; #endif @@ -201,16 +207,11 @@ GarbageCollect ( rtsBool force_major_gc ) } #endif - // tell the STM to discard any cached closures it's hoping to re-use - stmPreGCHook(); - // tell the stats department that we've started a GC stat_startGC(); -#ifdef DEBUG - // check for memory leaks if DEBUG is on - memInventory(); -#endif + // tell the STM to discard any cached closures it's hoping to re-use + stmPreGCHook(); #ifdef DEBUG mutlist_MUTVARS = 0; @@ -246,12 +247,12 @@ GarbageCollect ( rtsBool force_major_gc ) */ #if defined(THREADED_RTS) if (N == 0) { - n_threads = 1; + n_gc_threads = 1; } else { - n_threads = RtsFlags.ParFlags.gcThreads; + n_gc_threads = RtsFlags.ParFlags.gcThreads; } #else - n_threads = 1; + n_gc_threads = 1; #endif #ifdef RTS_GTK_FRONTPANEL @@ -260,6 +261,11 @@ GarbageCollect ( rtsBool force_major_gc ) } #endif +#ifdef DEBUG + // check for memory leaks if DEBUG is on + memInventory(traceClass(DEBUG_gc)); +#endif + // check stack sanity *before* GC (ToDo: check all threads) IF_DEBUG(sanity, checkFreeListSanity()); @@ -276,12 +282,12 @@ GarbageCollect ( rtsBool force_major_gc ) // Initialise all the generations/steps that we're collecting. for (g = 0; g <= N; g++) { - init_collected_gen(g,n_threads); + init_collected_gen(g,n_gc_threads); } // Initialise all the generations/steps that we're *not* collecting. for (g = N+1; g < RtsFlags.GcFlags.generations; g++) { - init_uncollected_gen(g,n_threads); + init_uncollected_gen(g,n_gc_threads); } /* Allocate a mark stack if we're doing a major collection. @@ -296,18 +302,17 @@ GarbageCollect ( rtsBool force_major_gc ) } // Initialise all our gc_thread structures - for (t = 0; t < n_threads; t++) { + for (t = 0; t < n_gc_threads; t++) { init_gc_thread(&gc_threads[t]); } // the main thread is running: this prevents any other threads from // exiting prematurely, so we can start them now. inc_running(); - wakeup_gc_threads(n_threads); + wakeup_gc_threads(n_gc_threads); // Initialise stats copied = 0; - scavd_copied = 0; // this is the main thread gct = &gc_threads[0]; @@ -332,13 +337,18 @@ GarbageCollect ( rtsBool force_major_gc ) } // follow roots from the CAF list (used by GHCi) - gct->evac_gen = 0; + gct->evac_step = 0; markCAFs(mark_root); // follow all the roots that the application knows about. - gct->evac_gen = 0; + gct->evac_step = 0; GetRoots(mark_root); +#if defined(RTS_USER_SIGNALS) + // mark the signal handlers (signals should be already blocked) + markSignalHandlers(mark_root); +#endif + // Mark the weak pointer list, and prepare to detect dead weak pointers. markWeakPtrList(); initWeakForGC(); @@ -416,7 +426,7 @@ GarbageCollect ( rtsBool force_major_gc ) step_workspace *ws; bdescr *prev; - for (t = 0; t < n_threads; t++) { + for (t = 0; t < n_gc_threads; t++) { thr = &gc_threads[t]; for (g = 0; g < RtsFlags.GcFlags.generations; g++) { @@ -431,10 +441,6 @@ GarbageCollect ( rtsBool force_major_gc ) // Push the final block if (ws->scan_bd) { push_scan_block(ws->scan_bd, ws); } - // update stats: we haven't counted the block at the - // front of the scavd_list yet. - scavd_copied += ws->scavd_list->free - ws->scavd_list->start; - ASSERT(countBlocks(ws->scavd_list) == ws->n_scavd_blocks); prev = ws->scavd_list; @@ -574,7 +580,9 @@ GarbageCollect ( rtsBool force_major_gc ) resize_generations(); // Guess the amount of live data for stats. - live = calcLive(); + live = calcLiveBlocks() * BLOCK_SIZE_W; + debugTrace(DEBUG_gc, "Slop: %ldKB", + (live - calcLiveWords()) / (1024/sizeof(W_))); // Free the small objects allocated via allocate(), since this will // all have been copied into G0S1 now. @@ -659,7 +667,7 @@ GarbageCollect ( rtsBool force_major_gc ) #ifdef DEBUG // check for memory leaks if DEBUG is on - memInventory(); + memInventory(traceClass(DEBUG_gc)); #endif #ifdef RTS_GTK_FRONTPANEL @@ -669,7 +677,7 @@ GarbageCollect ( rtsBool force_major_gc ) #endif // ok, GC over: tell the stats department what happened. - stat_endGC(allocated, live, copied, scavd_copied, N); + stat_endGC(allocated, live, copied, N); #if defined(RTS_USER_SIGNALS) if (RtsFlags.MiscFlags.install_signal_handlers) { @@ -679,7 +687,79 @@ GarbageCollect ( rtsBool force_major_gc ) #endif RELEASE_SM_LOCK; + + gct = saved_gct; +} + +/* ----------------------------------------------------------------------------- + * Mark all nodes pointed to by sparks in the spark queues (for GC) Does an + * implicit slide i.e. after marking all sparks are at the beginning of the + * spark pool and the spark pool only contains sparkable closures + * -------------------------------------------------------------------------- */ + +#ifdef THREADED_RTS +static void +markSparkQueue (evac_fn evac, Capability *cap) +{ + StgClosure **sparkp, **to_sparkp; + nat n, pruned_sparks; // stats only + StgSparkPool *pool; + + PAR_TICKY_MARK_SPARK_QUEUE_START(); + + n = 0; + pruned_sparks = 0; + + pool = &(cap->r.rSparks); + + ASSERT_SPARK_POOL_INVARIANTS(pool); + +#if defined(PARALLEL_HASKELL) + // stats only + n = 0; + pruned_sparks = 0; +#endif + + sparkp = pool->hd; + to_sparkp = pool->hd; + while (sparkp != pool->tl) { + ASSERT(*sparkp!=NULL); + ASSERT(LOOKS_LIKE_CLOSURE_PTR(((StgClosure *)*sparkp))); + // ToDo?: statistics gathering here (also for GUM!) + if (closure_SHOULD_SPARK(*sparkp)) { + evac(sparkp); + *to_sparkp++ = *sparkp; + if (to_sparkp == pool->lim) { + to_sparkp = pool->base; + } + n++; + } else { + pruned_sparks++; + } + sparkp++; + if (sparkp == pool->lim) { + sparkp = pool->base; + } + } + pool->tl = to_sparkp; + + PAR_TICKY_MARK_SPARK_QUEUE_END(n); + +#if defined(PARALLEL_HASKELL) + debugTrace(DEBUG_sched, + "marked %d sparks and pruned %d sparks on [%x]", + n, pruned_sparks, mytid); +#else + debugTrace(DEBUG_sched, + "marked %d sparks and pruned %d sparks", + n, pruned_sparks); +#endif + + debugTrace(DEBUG_sched, + "new spark queue len=%d; (hd=%p; tl=%p)\n", + sparkPoolSize(pool), pool->hd, pool->tl); } +#endif /* --------------------------------------------------------------------------- Where are the roots that we know about? @@ -692,11 +772,6 @@ GarbageCollect ( rtsBool force_major_gc ) ------------------------------------------------------------------------ */ -/* This has to be protected either by the scheduler monitor, or by the - garbage collection monitor (probably the latter). - KH @ 25/10/99 -*/ - void GetRoots( evac_fn evac ) { @@ -704,7 +779,12 @@ GetRoots( evac_fn evac ) Capability *cap; Task *task; - for (i = 0; i < n_capabilities; i++) { + // Each GC thread is responsible for following roots from the + // Capability of the same number. There will usually be the same + // or fewer Capabilities as GC threads, but just in case there + // are more, we mark every Capability whose number is the GC + // thread's index plus a multiple of the number of GC threads. + for (i = gct->thread_index; i < n_capabilities; i += n_gc_threads) { cap = &capabilities[i]; evac((StgClosure **)(void *)&cap->run_queue_hd); evac((StgClosure **)(void *)&cap->run_queue_tl); @@ -719,6 +799,9 @@ GetRoots( evac_fn evac ) evac((StgClosure **)(void *)&task->suspended_tso); } +#if defined(THREADED_RTS) + markSparkQueue(evac,cap); +#endif } #if !defined(THREADED_RTS) @@ -726,17 +809,6 @@ GetRoots( evac_fn evac ) evac((StgClosure **)(void *)&blocked_queue_tl); evac((StgClosure **)(void *)&sleeping_queue); #endif - - // evac((StgClosure **)&blackhole_queue); - -#if defined(THREADED_RTS) - markSparkQueue(evac); -#endif - -#if defined(RTS_USER_SIGNALS) - // mark the signal handlers (signals should be already blocked) - markSignalHandlers(evac); -#endif } /* ----------------------------------------------------------------------------- @@ -874,6 +946,10 @@ alloc_gc_thread (gc_thread *t, int n) init_gc_thread(t); +#ifdef USE_PAPI + t->papi_events = -1; +#endif + t->steps = stgMallocBytes(RtsFlags.GcFlags.generations * sizeof(step_workspace *), "initialise_gc_thread"); @@ -971,6 +1047,10 @@ gc_thread_work (void) // GarbageCollect(), or this is a worker thread and the main // thread bumped gc_running_threads before waking us up. + // Every thread evacuates some roots. + gct->evac_step = 0; + GetRoots(mark_root); + loop: scavenge_loop(); // scavenge_loop() only exits when there's no work to do @@ -1012,7 +1092,20 @@ gc_thread_mainloop (void) gct->wakeup = rtsFalse; if (gct->exit) break; +#ifdef USE_PAPI + // start performance counters in this thread... + if (gct->papi_events == -1) { + papi_init_eventset(&gct->papi_events); + } + papi_thread_start_gc1_count(gct->papi_events); +#endif + gc_thread_work(); + +#ifdef USE_PAPI + // count events in this thread towards the GC totals + papi_thread_stop_gc1_count(gct->papi_events); +#endif } } #endif @@ -1177,12 +1270,8 @@ init_collected_gen (nat g, nat n_threads) ws->buffer_todo_bd = NULL; gc_alloc_todo_block(ws); - // allocate a block for "already scavenged" objects. This goes - // on the front of the stp->blocks list, so it won't be - // traversed by the scavenging sweep. ws->scavd_list = NULL; ws->n_scavd_blocks = 0; - gc_alloc_scavd_block(ws); } } } @@ -1215,11 +1304,16 @@ init_uncollected_gen (nat g, nat threads) ws->buffer_todo_bd = NULL; ws->todo_large_objects = NULL; + ws->scavd_list = NULL; + ws->n_scavd_blocks = 0; + // If the block at the head of the list in this generation // is less than 3/4 full, then use it as a todo block. - if (isPartiallyFull(stp->blocks)) + if (stp->blocks && isPartiallyFull(stp->blocks)) { ws->todo_bd = stp->blocks; + ws->todo_free = ws->todo_bd->free; + ws->todo_lim = ws->todo_bd->start + BLOCK_SIZE_W; stp->blocks = stp->blocks->link; stp->n_blocks -= 1; ws->todo_bd->link = NULL; @@ -1240,25 +1334,6 @@ init_uncollected_gen (nat g, nat threads) ws->todo_bd = NULL; gc_alloc_todo_block(ws); } - - // Do the same trick for the scavd block - if (isPartiallyFull(stp->blocks)) - { - ws->scavd_list = stp->blocks; - stp->blocks = stp->blocks->link; - stp->n_blocks -= 1; - ws->scavd_list->link = NULL; - ws->n_scavd_blocks = 1; - // subtract the contents of this block from the stats, - // because we'll count the whole block later. - scavd_copied -= ws->scavd_list->free - ws->scavd_list->start; - } - else - { - ws->scavd_list = NULL; - ws->n_scavd_blocks = 0; - gc_alloc_scavd_block(ws); - } } } @@ -1282,7 +1357,7 @@ init_uncollected_gen (nat g, nat threads) static void init_gc_thread (gc_thread *t) { - t->evac_gen = 0; + t->evac_step = 0; t->failed_to_evac = rtsFalse; t->eager_promotion = rtsTrue; t->thunk_selector_depth = 0; @@ -1295,7 +1370,7 @@ init_gc_thread (gc_thread *t) static void mark_root(StgClosure **root) { - *root = evacuate(*root); + evacuate(root); } /* -----------------------------------------------------------------------------