diff options
Diffstat (limited to 'erts/emulator/sys')
-rw-r--r-- | erts/emulator/sys/common/erl_mseg.c | 675 | ||||
-rw-r--r-- | erts/emulator/sys/common/erl_poll.c | 285 | ||||
-rw-r--r-- | erts/emulator/sys/unix/sys.c | 67 | ||||
-rw-r--r-- | erts/emulator/sys/unix/sys_float.c | 7 | ||||
-rw-r--r-- | erts/emulator/sys/win32/erl_poll.c | 386 | ||||
-rw-r--r-- | erts/emulator/sys/win32/sys_interrupt.c | 8 |
6 files changed, 790 insertions, 638 deletions
diff --git a/erts/emulator/sys/common/erl_mseg.c b/erts/emulator/sys/common/erl_mseg.c index 010d60eb63..ceb290b644 100644 --- a/erts/emulator/sys/common/erl_mseg.c +++ b/erts/emulator/sys/common/erl_mseg.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2002-2010. All Rights Reserved. + * Copyright Ericsson AB 2002-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -77,8 +77,10 @@ static int atoms_initialized; static Uint cache_check_interval; +typedef struct mem_kind_t MemKind; + static void check_cache(void *unused); -static void mseg_clear_cache(void); +static void mseg_clear_cache(MemKind*); static int is_cache_check_scheduled; #ifdef ERTS_THREADS_NO_SMP static int is_cache_check_requested; @@ -123,6 +125,9 @@ static int mmap_fd; #error "Not supported" #endif /* #if HAVE_MMAP */ +#if defined(ERTS_MSEG_FAKE_SEGMENTS) && HALFWORD_HEAP +# warning "ERTS_MSEG_FAKE_SEGMENTS will only be used for high memory segments" +#endif #if defined(ERTS_MSEG_FAKE_SEGMENTS) #undef CAN_PARTLY_DESTROY @@ -160,14 +165,43 @@ static struct { CallCounter check_cache; } calls; -static cache_desc_t cache_descs[MAX_CACHE_SIZE]; -static cache_desc_t *free_cache_descs; -static cache_desc_t *cache; -static cache_desc_t *cache_end; -static Uint cache_hits; -static Uint cache_size; -static Uint min_cached_seg_size; -static Uint max_cached_seg_size; +struct mem_kind_t { + cache_desc_t cache_descs[MAX_CACHE_SIZE]; + cache_desc_t *free_cache_descs; + cache_desc_t *cache; + cache_desc_t *cache_end; + + Uint cache_size; + Uint min_cached_seg_size; + Uint max_cached_seg_size; + Uint cache_hits; + + struct { + struct { + Uint watermark; + Uint no; + Uint sz; + } current; + struct { + Uint no; + Uint sz; + } max; + struct { + Uint no; + Uint sz; + } max_ever; + } segments; + + const char* name; + MemKind* next; +};/*MemKind*/ + +#if HALFWORD_HEAP +static MemKind low_mem, hi_mem; +#else +static MemKind the_mem; +#endif +static MemKind* mk_list = NULL; static Uint max_cache_size; static Uint abs_max_cache_bad_fit; @@ -177,47 +211,32 @@ static Uint rel_max_cache_bad_fit; static Uint min_seg_size; #endif -struct { - struct { - Uint watermark; - Uint no; - Uint sz; - } current; - struct { - Uint no; - Uint sz; - } max; - struct { - Uint no; - Uint sz; - } max_ever; -} segments; -#define ERTS_MSEG_ALLOC_STAT(SZ) \ +#define ERTS_MSEG_ALLOC_STAT(C,SZ) \ do { \ - segments.current.no++; \ - if (segments.max.no < segments.current.no) \ - segments.max.no = segments.current.no; \ - if (segments.current.watermark < segments.current.no) \ - segments.current.watermark = segments.current.no; \ - segments.current.sz += (SZ); \ - if (segments.max.sz < segments.current.sz) \ - segments.max.sz = segments.current.sz; \ + C->segments.current.no++; \ + if (C->segments.max.no < C->segments.current.no) \ + C->segments.max.no = C->segments.current.no; \ + if (C->segments.current.watermark < C->segments.current.no) \ + C->segments.current.watermark = C->segments.current.no; \ + C->segments.current.sz += (SZ); \ + if (C->segments.max.sz < C->segments.current.sz) \ + C->segments.max.sz = C->segments.current.sz; \ } while (0) -#define ERTS_MSEG_DEALLOC_STAT(SZ) \ +#define ERTS_MSEG_DEALLOC_STAT(C,SZ) \ do { \ - ASSERT(segments.current.no > 0); \ - segments.current.no--; \ - ASSERT(segments.current.sz >= (SZ)); \ - segments.current.sz -= (SZ); \ + ASSERT(C->segments.current.no > 0); \ + C->segments.current.no--; \ + ASSERT(C->segments.current.sz >= (SZ)); \ + C->segments.current.sz -= (SZ); \ } while (0) -#define ERTS_MSEG_REALLOC_STAT(OSZ, NSZ) \ +#define ERTS_MSEG_REALLOC_STAT(C,OSZ, NSZ) \ do { \ - ASSERT(segments.current.sz >= (OSZ)); \ - segments.current.sz -= (OSZ); \ - segments.current.sz += (NSZ); \ + ASSERT(C->segments.current.sz >= (OSZ)); \ + C->segments.current.sz -= (OSZ); \ + C->segments.current.sz += (NSZ); \ } while (0) #define ONE_GIGA (1000000000) @@ -303,38 +322,45 @@ check_schedule_cache_check(void) static void mseg_shutdown(void) { + MemKind* mk; erts_mtx_lock(&mseg_mutex); - mseg_clear_cache(); + for (mk=mk_list; mk; mk=mk->next) { + mseg_clear_cache(mk); + } erts_mtx_unlock(&mseg_mutex); } static ERTS_INLINE void * -mseg_create(Uint size) +mseg_create(MemKind* mk, Uint size) { void *seg; ASSERT(size % page_size == 0); -#if defined(ERTS_MSEG_FAKE_SEGMENTS) - seg = erts_sys_alloc(ERTS_ALC_N_INVALID, NULL, size); -#elif HAVE_MMAP #if HALFWORD_HEAP - seg = pmmap(size); -#else - seg = (void *) mmap((void *) 0, (size_t) size, - MMAP_PROT, MMAP_FLAGS, MMAP_FD, 0); - if (seg == (void *) MAP_FAILED) - seg = NULL; -#endif -#if HALFWORD_HEAP - if ((unsigned long) seg & CHECK_POINTER_MASK) { - erts_fprintf(stderr,"Pointer mask failure (0x%08lx)\n",(unsigned long) seg); - return NULL; + if (mk == &low_mem) { + seg = pmmap(size); + if ((unsigned long) seg & CHECK_POINTER_MASK) { + erts_fprintf(stderr,"Pointer mask failure (0x%08lx)\n",(unsigned long) seg); + return NULL; + } } + else #endif + { +#if defined(ERTS_MSEG_FAKE_SEGMENTS) + seg = erts_sys_alloc(ERTS_ALC_N_INVALID, NULL, size); +#elif HAVE_MMAP + { + seg = (void *) mmap((void *) 0, (size_t) size, + MMAP_PROT, MMAP_FLAGS, MMAP_FD, 0); + if (seg == (void *) MAP_FAILED) + seg = NULL; + } #else -#error "Missing mseg_create() implementation" +# error "Missing mseg_create() implementation" #endif + } INC_CC(create); @@ -342,25 +368,29 @@ mseg_create(Uint size) } static ERTS_INLINE void -mseg_destroy(void *seg, Uint size) +mseg_destroy(MemKind* mk, void *seg, Uint size) { -#if defined(ERTS_MSEG_FAKE_SEGMENTS) - erts_sys_free(ERTS_ALC_N_INVALID, NULL, seg); -#elif HAVE_MMAP + int res; -#ifdef DEBUG - int res = -#endif #if HALFWORD_HEAP - pmunmap((void *) seg, size); + if (mk == &low_mem) { + res = pmunmap((void *) seg, size); + } + else +#endif + { +#ifdef ERTS_MSEG_FAKE_SEGMENTS + erts_sys_free(ERTS_ALC_N_INVALID, NULL, seg); + res = 0; +#elif HAVE_MMAP + res = munmap((void *) seg, size); #else - munmap((void *) seg, size); +# error "Missing mseg_destroy() implementation" #endif + } + ASSERT(size % page_size == 0); ASSERT(res == 0); -#else -#error "Missing mseg_destroy() implementation" -#endif INC_CC(destroy); @@ -369,39 +399,44 @@ mseg_destroy(void *seg, Uint size) #if HAVE_MSEG_RECREATE static ERTS_INLINE void * -mseg_recreate(void *old_seg, Uint old_size, Uint new_size) +mseg_recreate(MemKind* mk, void *old_seg, Uint old_size, Uint new_size) { void *new_seg; ASSERT(old_size % page_size == 0); ASSERT(new_size % page_size == 0); -#if defined(ERTS_MSEG_FAKE_SEGMENTS) - new_seg = erts_sys_realloc(ERTS_ALC_N_INVALID, NULL, old_seg, new_size); -#elif HAVE_MREMAP #if HALFWORD_HEAP - new_seg = (void *) pmremap((void *) old_seg, - (size_t) old_size, - (size_t) new_size); -#elif defined(__NetBSD__) - new_seg = (void *) mremap((void *) old_seg, - (size_t) old_size, - NULL, - (size_t) new_size, - 0); - if (new_seg == (void *) MAP_FAILED) - new_seg = NULL; -#else - new_seg = (void *) mremap((void *) old_seg, - (size_t) old_size, - (size_t) new_size, - MREMAP_MAYMOVE); - if (new_seg == (void *) MAP_FAILED) - new_seg = NULL; + if (mk == &low_mem) { + new_seg = (void *) pmremap((void *) old_seg, + (size_t) old_size, + (size_t) new_size); + } + else #endif + { +#if defined(ERTS_MSEG_FAKE_SEGMENTS) + new_seg = erts_sys_realloc(ERTS_ALC_N_INVALID, NULL, old_seg, new_size); +#elif HAVE_MREMAP + + #if defined(__NetBSD__) + new_seg = (void *) mremap((void *) old_seg, + (size_t) old_size, + NULL, + (size_t) new_size, + 0); + #else + new_seg = (void *) mremap((void *) old_seg, + (size_t) old_size, + (size_t) new_size, + MREMAP_MAYMOVE); + #endif + if (new_seg == (void *) MAP_FAILED) + new_seg = NULL; #else #error "Missing mseg_recreate() implementation" #endif + } INC_CC(recreate); @@ -412,134 +447,142 @@ mseg_recreate(void *old_seg, Uint old_size, Uint new_size) static ERTS_INLINE cache_desc_t * -alloc_cd(void) +alloc_cd(MemKind* mk) { - cache_desc_t *cd = free_cache_descs; + cache_desc_t *cd = mk->free_cache_descs; ERTS_LC_ASSERT(erts_lc_mtx_is_locked(&mseg_mutex)); if (cd) - free_cache_descs = cd->next; + mk->free_cache_descs = cd->next; return cd; } static ERTS_INLINE void -free_cd(cache_desc_t *cd) +free_cd(MemKind* mk, cache_desc_t *cd) { ERTS_LC_ASSERT(erts_lc_mtx_is_locked(&mseg_mutex)); - cd->next = free_cache_descs; - free_cache_descs = cd; + cd->next = mk->free_cache_descs; + mk->free_cache_descs = cd; } static ERTS_INLINE void -link_cd(cache_desc_t *cd) +link_cd(MemKind* mk, cache_desc_t *cd) { ERTS_LC_ASSERT(erts_lc_mtx_is_locked(&mseg_mutex)); - if (cache) - cache->prev = cd; - cd->next = cache; + if (mk->cache) + mk->cache->prev = cd; + cd->next = mk->cache; cd->prev = NULL; - cache = cd; + mk->cache = cd; - if (!cache_end) { + if (!mk->cache_end) { ASSERT(!cd->next); - cache_end = cd; + mk->cache_end = cd; } - cache_size++; + mk->cache_size++; } +#if CAN_PARTLY_DESTROY static ERTS_INLINE void -end_link_cd(cache_desc_t *cd) +end_link_cd(MemKind* mk, cache_desc_t *cd) { ERTS_LC_ASSERT(erts_lc_mtx_is_locked(&mseg_mutex)); - if (cache_end) - cache_end->next = cd; + if (mk->cache_end) + mk->cache_end->next = cd; cd->next = NULL; - cd->prev = cache_end; - cache_end = cd; + cd->prev = mk->cache_end; + mk->cache_end = cd; - if (!cache) { + if (!mk->cache) { ASSERT(!cd->prev); - cache = cd; + mk->cache = cd; } - cache_size++; + mk->cache_size++; } +#endif static ERTS_INLINE void -unlink_cd(cache_desc_t *cd) +unlink_cd(MemKind* mk, cache_desc_t *cd) { ERTS_LC_ASSERT(erts_lc_mtx_is_locked(&mseg_mutex)); if (cd->next) cd->next->prev = cd->prev; else - cache_end = cd->prev; + mk->cache_end = cd->prev; if (cd->prev) cd->prev->next = cd->next; else - cache = cd->next; - ASSERT(cache_size > 0); - cache_size--; + mk->cache = cd->next; + ASSERT(mk->cache_size > 0); + mk->cache_size--; } static ERTS_INLINE void -check_cache_limits(void) +check_cache_limits(MemKind* mk) { cache_desc_t *cd; ERTS_LC_ASSERT(erts_lc_mtx_is_locked(&mseg_mutex)); - max_cached_seg_size = 0; - min_cached_seg_size = ~((Uint) 0); - for (cd = cache; cd; cd = cd->next) { - if (cd->size < min_cached_seg_size) - min_cached_seg_size = cd->size; - if (cd->size > max_cached_seg_size) - max_cached_seg_size = cd->size; + mk->max_cached_seg_size = 0; + mk->min_cached_seg_size = ~((Uint) 0); + for (cd = mk->cache; cd; cd = cd->next) { + if (cd->size < mk->min_cached_seg_size) + mk->min_cached_seg_size = cd->size; + if (cd->size > mk->max_cached_seg_size) + mk->max_cached_seg_size = cd->size; } - } static ERTS_INLINE void -adjust_cache_size(int force_check_limits) +adjust_cache_size(MemKind* mk, int force_check_limits) { cache_desc_t *cd; int check_limits = force_check_limits; - Sint max_cached = ((Sint) segments.current.watermark - - (Sint) segments.current.no); + Sint max_cached = ((Sint) mk->segments.current.watermark + - (Sint) mk->segments.current.no); ERTS_LC_ASSERT(erts_lc_mtx_is_locked(&mseg_mutex)); - while (((Sint) cache_size) > max_cached && ((Sint) cache_size) > 0) { - ASSERT(cache_end); - cd = cache_end; + while (((Sint) mk->cache_size) > max_cached && ((Sint) mk->cache_size) > 0) { + ASSERT(mk->cache_end); + cd = mk->cache_end; if (!check_limits && - !(min_cached_seg_size < cd->size - && cd->size < max_cached_seg_size)) { + !(mk->min_cached_seg_size < cd->size + && cd->size < mk->max_cached_seg_size)) { check_limits = 1; } if (erts_mtrace_enabled) erts_mtrace_crr_free(SEGTYPE, SEGTYPE, cd->seg); - mseg_destroy(cd->seg, cd->size); - unlink_cd(cd); - free_cd(cd); + mseg_destroy(mk, cd->seg, cd->size); + unlink_cd(mk,cd); + free_cd(mk,cd); } if (check_limits) - check_cache_limits(); - + check_cache_limits(mk); } static void -check_cache(void *unused) +check_one_cache(MemKind* mk) { + if (mk->segments.current.watermark > mk->segments.current.no) + mk->segments.current.watermark--; + adjust_cache_size(mk, 0); + + if (mk->cache_size) + schedule_cache_check(); +} + +static void check_cache(void* unused) +{ + MemKind* mk; erts_mtx_lock(&mseg_mutex); is_cache_check_scheduled = 0; - if (segments.current.watermark > segments.current.no) - segments.current.watermark--; - adjust_cache_size(0); - - if (cache_size) - schedule_cache_check(); + for (mk=mk_list; mk; mk=mk->next) { + check_one_cache(mk); + } INC_CC(check_cache); @@ -547,28 +590,45 @@ check_cache(void *unused) } static void -mseg_clear_cache(void) +mseg_clear_cache(MemKind* mk) { - segments.current.watermark = 0; + mk->segments.current.watermark = 0; - adjust_cache_size(1); + adjust_cache_size(mk, 1); - ASSERT(!cache); - ASSERT(!cache_end); - ASSERT(!cache_size); + ASSERT(!mk->cache); + ASSERT(!mk->cache_end); + ASSERT(!mk->cache_size); - segments.current.watermark = segments.current.no; + mk->segments.current.watermark = mk->segments.current.no; INC_CC(clear_cache); } +static ERTS_INLINE MemKind* type2mk(ErtsAlcType_t atype) +{ +#if HALFWORD_HEAP + switch (atype) { + case ERTS_ALC_A_ETS: + case ERTS_ALC_A_BINARY: + case ERTS_ALC_A_FIXED_SIZE: + case ERTS_ALC_A_DRIVER: + return &hi_mem; + default: + return &low_mem; + } +#else + return &the_mem; +#endif +} + static void * mseg_alloc(ErtsAlcType_t atype, Uint *size_p, const ErtsMsegOpt_t *opt) { - Uint max, min, diff_size, size; cache_desc_t *cd, *cand_cd; void *seg; + MemKind* mk = type2mk(atype); INC_CC(alloc); @@ -581,11 +641,11 @@ mseg_alloc(ErtsAlcType_t atype, Uint *size_p, const ErtsMsegOpt_t *opt) if (!opt->cache) { create_seg: - adjust_cache_size(0); - seg = mseg_create(size); + adjust_cache_size(mk,0); + seg = mseg_create(mk, size); if (!seg) { - mseg_clear_cache(); - seg = mseg_create(size); + mseg_clear_cache(mk); + seg = mseg_create(mk, size); if (!seg) size = 0; } @@ -594,17 +654,17 @@ mseg_alloc(ErtsAlcType_t atype, Uint *size_p, const ErtsMsegOpt_t *opt) if (seg) { if (erts_mtrace_enabled) erts_mtrace_crr_alloc(seg, atype, ERTS_MTRACE_SEGMENT_ID, size); - ERTS_MSEG_ALLOC_STAT(size); + ERTS_MSEG_ALLOC_STAT(mk,size); } return seg; } - if (size > max_cached_seg_size) + if (size > mk->max_cached_seg_size) goto create_seg; - if (size < min_cached_seg_size) { + if (size < mk->min_cached_seg_size) { - diff_size = min_cached_seg_size - size; + diff_size = mk->min_cached_seg_size - size; if (diff_size > abs_max_cache_bad_fit) goto create_seg; @@ -618,7 +678,7 @@ mseg_alloc(ErtsAlcType_t atype, Uint *size_p, const ErtsMsegOpt_t *opt) min = ~((Uint) 0); cand_cd = NULL; - for (cd = cache; cd; cd = cd->next) { + for (cd = mk->cache; cd; cd = cd->next) { if (cd->size >= size) { if (!cand_cd) { cand_cd = cd; @@ -639,8 +699,8 @@ mseg_alloc(ErtsAlcType_t atype, Uint *size_p, const ErtsMsegOpt_t *opt) min = cd->size; } - min_cached_seg_size = min; - max_cached_seg_size = max; + mk->min_cached_seg_size = min; + mk->max_cached_seg_size = max; if (!cand_cd) goto create_seg; @@ -649,20 +709,20 @@ mseg_alloc(ErtsAlcType_t atype, Uint *size_p, const ErtsMsegOpt_t *opt) if (diff_size > abs_max_cache_bad_fit || 100*PAGES(diff_size) > rel_max_cache_bad_fit*PAGES(size)) { - if (max_cached_seg_size < cand_cd->size) - max_cached_seg_size = cand_cd->size; - if (min_cached_seg_size > cand_cd->size) - min_cached_seg_size = cand_cd->size; + if (mk->max_cached_seg_size < cand_cd->size) + mk->max_cached_seg_size = cand_cd->size; + if (mk->min_cached_seg_size > cand_cd->size) + mk->min_cached_seg_size = cand_cd->size; goto create_seg; } - cache_hits++; + mk->cache_hits++; size = cand_cd->size; seg = cand_cd->seg; - unlink_cd(cand_cd); - free_cd(cand_cd); + unlink_cd(mk,cand_cd); + free_cd(mk,cand_cd); *size_p = size; @@ -672,7 +732,8 @@ mseg_alloc(ErtsAlcType_t atype, Uint *size_p, const ErtsMsegOpt_t *opt) } if (seg) - ERTS_MSEG_ALLOC_STAT(size); + ERTS_MSEG_ALLOC_STAT(mk,size); + return seg; } @@ -681,41 +742,42 @@ static void mseg_dealloc(ErtsAlcType_t atype, void *seg, Uint size, const ErtsMsegOpt_t *opt) { + MemKind* mk = type2mk(atype); cache_desc_t *cd; - ERTS_MSEG_DEALLOC_STAT(size); + ERTS_MSEG_DEALLOC_STAT(mk,size); if (!opt->cache || max_cache_size == 0) { if (erts_mtrace_enabled) erts_mtrace_crr_free(atype, SEGTYPE, seg); - mseg_destroy(seg, size); + mseg_destroy(mk, seg, size); } else { int check_limits = 0; - if (size < min_cached_seg_size) - min_cached_seg_size = size; - if (size > max_cached_seg_size) - max_cached_seg_size = size; - - if (!free_cache_descs) { - cd = cache_end; - if (!(min_cached_seg_size < cd->size - && cd->size < max_cached_seg_size)) { + if (size < mk->min_cached_seg_size) + mk->min_cached_seg_size = size; + if (size > mk->max_cached_seg_size) + mk->max_cached_seg_size = size; + + if (!mk->free_cache_descs) { + cd = mk->cache_end; + if (!(mk->min_cached_seg_size < cd->size + && cd->size < mk->max_cached_seg_size)) { check_limits = 1; } if (erts_mtrace_enabled) erts_mtrace_crr_free(SEGTYPE, SEGTYPE, cd->seg); - mseg_destroy(cd->seg, cd->size); - unlink_cd(cd); - free_cd(cd); + mseg_destroy(mk, cd->seg, cd->size); + unlink_cd(mk,cd); + free_cd(mk,cd); } - cd = alloc_cd(); + cd = alloc_cd(mk); ASSERT(cd); cd->seg = seg; cd->size = size; - link_cd(cd); + link_cd(mk,cd); if (erts_mtrace_enabled) { erts_mtrace_crr_free(atype, SEGTYPE, seg); @@ -725,7 +787,7 @@ mseg_dealloc(ErtsAlcType_t atype, void *seg, Uint size, /* ASSERT(segments.current.watermark >= segments.current.no + cache_size); */ if (check_limits) - check_cache_limits(); + check_cache_limits(mk); schedule_cache_check(); @@ -738,6 +800,7 @@ static void * mseg_realloc(ErtsAlcType_t atype, void *seg, Uint old_size, Uint *new_size_p, const ErtsMsegOpt_t *opt) { + MemKind* mk = type2mk(atype); void *new_seg; Uint new_size; @@ -775,15 +838,15 @@ mseg_realloc(ErtsAlcType_t atype, void *seg, Uint old_size, Uint *new_size_p, #if CAN_PARTLY_DESTROY if (shrink_sz > min_seg_size - && free_cache_descs + && mk->free_cache_descs && opt->cache) { cache_desc_t *cd; - cd = alloc_cd(); + cd = alloc_cd(mk); ASSERT(cd); cd->seg = ((char *) seg) + new_size; cd->size = shrink_sz; - end_link_cd(cd); + end_link_cd(mk,cd); if (erts_mtrace_enabled) { erts_mtrace_crr_realloc(new_seg, @@ -802,7 +865,7 @@ mseg_realloc(ErtsAlcType_t atype, void *seg, Uint old_size, Uint *new_size_p, SEGTYPE, seg, new_size); - mseg_destroy(((char *) seg) + new_size, shrink_sz); + mseg_destroy(mk, ((char *) seg) + new_size, shrink_sz); } #elif HAVE_MSEG_RECREATE @@ -836,7 +899,7 @@ mseg_realloc(ErtsAlcType_t atype, void *seg, Uint old_size, Uint *new_size_p, #if !CAN_PARTLY_DESTROY do_recreate: #endif - new_seg = mseg_recreate((void *) seg, old_size, new_size); + new_seg = mseg_recreate(mk, (void *) seg, old_size, new_size); if (erts_mtrace_enabled) erts_mtrace_crr_realloc(new_seg, atype, SEGTYPE, seg, new_size); if (!new_seg) @@ -859,7 +922,7 @@ mseg_realloc(ErtsAlcType_t atype, void *seg, Uint old_size, Uint *new_size_p, *new_size_p = new_size; - ERTS_MSEG_REALLOC_STAT(old_size, new_size); + ERTS_MSEG_REALLOC_STAT(mk, old_size, new_size); return new_seg; } @@ -875,6 +938,8 @@ static struct { Eterm mcs; Eterm cci; + Eterm memkind; + Eterm name; Eterm status; Eterm cached_segments; Eterm cache_hits; @@ -924,6 +989,8 @@ init_atoms(void) #endif AM_INIT(version); + AM_INIT(memkind); + AM_INIT(name); AM_INIT(options); AM_INIT(amcbf); @@ -1134,65 +1201,88 @@ info_calls(int *print_to_p, void *print_to_arg, Uint **hpp, Uint *szp) } static Eterm -info_status(int *print_to_p, - void *print_to_arg, - int begin_new_max_period, - Uint **hpp, - Uint *szp) +info_status(MemKind* mk, int *print_to_p, void *print_to_arg, + int begin_new_max_period, Uint **hpp, Uint *szp) { Eterm res = THE_NON_VALUE; - if (segments.max_ever.no < segments.max.no) - segments.max_ever.no = segments.max.no; - if (segments.max_ever.sz < segments.max.sz) - segments.max_ever.sz = segments.max.sz; + if (mk->segments.max_ever.no < mk->segments.max.no) + mk->segments.max_ever.no = mk->segments.max.no; + if (mk->segments.max_ever.sz < mk->segments.max.sz) + mk->segments.max_ever.sz = mk->segments.max.sz; if (print_to_p) { int to = *print_to_p; void *arg = print_to_arg; - erts_print(to, arg, "cached_segments: %bpu\n", cache_size); - erts_print(to, arg, "cache_hits: %bpu\n", cache_hits); + erts_print(to, arg, "cached_segments: %bpu\n", mk->cache_size); + erts_print(to, arg, "cache_hits: %bpu\n", mk->cache_hits); erts_print(to, arg, "segments: %bpu %bpu %bpu\n", - segments.current.no, segments.max.no, segments.max_ever.no); + mk->segments.current.no, mk->segments.max.no, mk->segments.max_ever.no); erts_print(to, arg, "segments_size: %bpu %bpu %bpu\n", - segments.current.sz, segments.max.sz, segments.max_ever.sz); + mk->segments.current.sz, mk->segments.max.sz, mk->segments.max_ever.sz); erts_print(to, arg, "segments_watermark: %bpu\n", - segments.current.watermark); + mk->segments.current.watermark); } if (hpp || szp) { res = NIL; add_2tup(hpp, szp, &res, am.segments_watermark, - bld_unstable_uint(hpp, szp, segments.current.watermark)); + bld_unstable_uint(hpp, szp, mk->segments.current.watermark)); add_4tup(hpp, szp, &res, am.segments_size, - bld_unstable_uint(hpp, szp, segments.current.sz), - bld_unstable_uint(hpp, szp, segments.max.sz), - bld_unstable_uint(hpp, szp, segments.max_ever.sz)); + bld_unstable_uint(hpp, szp, mk->segments.current.sz), + bld_unstable_uint(hpp, szp, mk->segments.max.sz), + bld_unstable_uint(hpp, szp, mk->segments.max_ever.sz)); add_4tup(hpp, szp, &res, am.segments, - bld_unstable_uint(hpp, szp, segments.current.no), - bld_unstable_uint(hpp, szp, segments.max.no), - bld_unstable_uint(hpp, szp, segments.max_ever.no)); + bld_unstable_uint(hpp, szp, mk->segments.current.no), + bld_unstable_uint(hpp, szp, mk->segments.max.no), + bld_unstable_uint(hpp, szp, mk->segments.max_ever.no)); add_2tup(hpp, szp, &res, am.cache_hits, - bld_unstable_uint(hpp, szp, cache_hits)); + bld_unstable_uint(hpp, szp, mk->cache_hits)); add_2tup(hpp, szp, &res, am.cached_segments, - bld_unstable_uint(hpp, szp, cache_size)); + bld_unstable_uint(hpp, szp, mk->cache_size)); } if (begin_new_max_period) { - segments.max.no = segments.current.no; - segments.max.sz = segments.current.sz; + mk->segments.max.no = mk->segments.current.no; + mk->segments.max.sz = mk->segments.current.sz; } return res; } +static Eterm info_memkind(MemKind* mk, int *print_to_p, void *print_to_arg, + int begin_max_per, Uint **hpp, Uint *szp) +{ + Eterm res = THE_NON_VALUE; + Eterm atoms[3]; + Eterm values[3]; + + if (print_to_p) { + erts_print(*print_to_p, print_to_arg, "memory kind: %s\n", mk->name); + } + if (hpp || szp) { + atoms[0] = am.name; + atoms[1] = am.status; + atoms[2] = am.calls; + values[0] = erts_bld_string(hpp, szp, mk->name); + } + values[1] = info_status(mk, print_to_p, print_to_arg, begin_max_per, hpp, szp); + values[2] = info_calls(print_to_p, print_to_arg, hpp, szp); + + if (hpp || szp) + res = bld_2tup_list(hpp, szp, 3, atoms, values); + + return res; +} + + static Eterm info_version(int *print_to_p, void *print_to_arg, Uint **hpp, Uint *szp) { @@ -1239,6 +1329,7 @@ erts_mseg_info(int *print_to_p, Eterm res = THE_NON_VALUE; Eterm atoms[4]; Eterm values[4]; + Uint n = 0; erts_mtx_lock(&mseg_mutex); @@ -1249,17 +1340,19 @@ erts_mseg_info(int *print_to_p, atoms[0] = am.version; atoms[1] = am.options; - atoms[2] = am.status; - atoms[3] = am.calls; + atoms[2] = am.memkind; + atoms[3] = am.memkind; } - - values[0] = info_version(print_to_p, print_to_arg, hpp, szp); - values[1] = info_options("option ", print_to_p, print_to_arg, hpp, szp); - values[2] = info_status(print_to_p, print_to_arg, begin_max_per, hpp, szp); - values[3] = info_calls(print_to_p, print_to_arg, hpp, szp); - + values[n++] = info_version(print_to_p, print_to_arg, hpp, szp); + values[n++] = info_options("option ", print_to_p, print_to_arg, hpp, szp); +#if HALFWORD_HEAP + values[n++] = info_memkind(&low_mem, print_to_p, print_to_arg, begin_max_per, hpp, szp); + values[n++] = info_memkind(&hi_mem, print_to_p, print_to_arg, begin_max_per, hpp, szp); +#else + values[n++] = info_memkind(&the_mem, print_to_p, print_to_arg, begin_max_per, hpp, szp); +#endif if (hpp || szp) - res = bld_2tup_list(hpp, szp, 4, atoms, values); + res = bld_2tup_list(hpp, szp, n, atoms, values); erts_mtx_unlock(&mseg_mutex); @@ -1318,17 +1411,23 @@ erts_mseg_realloc(ErtsAlcType_t atype, void *seg, Uint old_size, void erts_mseg_clear_cache(void) { + MemKind* mk; erts_mtx_lock(&mseg_mutex); - mseg_clear_cache(); + for (mk=mk_list; mk; mk=mk->next) { + mseg_clear_cache(mk); + } erts_mtx_unlock(&mseg_mutex); } Uint erts_mseg_no(void) { - Uint n; + MemKind* mk; + Uint n = 0; erts_mtx_lock(&mseg_mutex); - n = segments.current.no; + for (mk=mk_list; mk; mk=mk->next) { + n += mk->segments.current.no; + } erts_mtx_unlock(&mseg_mutex); return n; } @@ -1339,11 +1438,43 @@ erts_mseg_unit_size(void) return page_size; } -void -erts_mseg_init(ErtsMsegInit_t *init) +static void mem_kind_init(MemKind* mk, const char* name) { unsigned i; + mk->cache = NULL; + mk->cache_end = NULL; + mk->max_cached_seg_size = 0; + mk->min_cached_seg_size = ~((Uint) 0); + mk->cache_size = 0; + mk->cache_hits = 0; + + if (max_cache_size > 0) { + for (i = 0; i < max_cache_size - 1; i++) + mk->cache_descs[i].next = &mk->cache_descs[i + 1]; + mk->cache_descs[max_cache_size - 1].next = NULL; + mk->free_cache_descs = &mk->cache_descs[0]; + } + else + mk->free_cache_descs = NULL; + + mk->segments.current.watermark = 0; + mk->segments.current.no = 0; + mk->segments.current.sz = 0; + mk->segments.max.no = 0; + mk->segments.max.sz = 0; + mk->segments.max_ever.no = 0; + mk->segments.max_ever.sz = 0; + + mk->name = name; + mk->next = mk_list; + mk_list = mk; +} + + +void +erts_mseg_init(ErtsMsegInit_t *init) +{ atoms_initialized = 0; is_init_done = 0; @@ -1386,40 +1517,33 @@ erts_mseg_init(ErtsMsegInit_t *init) min_seg_size = ~((Uint) 0); #endif - cache = NULL; - cache_end = NULL; - cache_hits = 0; - max_cached_seg_size = 0; - min_cached_seg_size = ~((Uint) 0); - cache_size = 0; + if (max_cache_size > MAX_CACHE_SIZE) + max_cache_size = MAX_CACHE_SIZE; + +#if HALFWORD_HEAP + mem_kind_init(&low_mem, "low memory"); + mem_kind_init(&hi_mem, "high memory"); +#else + mem_kind_init(&the_mem, "all memory"); +#endif is_cache_check_scheduled = 0; #ifdef ERTS_THREADS_NO_SMP is_cache_check_requested = 0; #endif +} - if (max_cache_size > MAX_CACHE_SIZE) - max_cache_size = MAX_CACHE_SIZE; - if (max_cache_size > 0) { - for (i = 0; i < max_cache_size - 1; i++) - cache_descs[i].next = &cache_descs[i + 1]; - cache_descs[max_cache_size - 1].next = NULL; - free_cache_descs = &cache_descs[0]; +static ERTS_INLINE Uint tot_cache_size(void) +{ + MemKind* mk; + Uint sz = 0; + for (mk=mk_list; mk; mk=mk->next) { + sz += mk->cache_size; } - else - free_cache_descs = NULL; - - segments.current.watermark = 0; - segments.current.no = 0; - segments.current.sz = 0; - segments.max.no = 0; - segments.max.sz = 0; - segments.max_ever.no = 0; - segments.max_ever.sz = 0; + return sz; } - /* * erts_mseg_late_init() have to be called after all allocators, * threads and timers have been initialized. @@ -1437,7 +1561,7 @@ erts_mseg_late_init(void) #ifdef ERTS_THREADS_NO_SMP async_handle = handle; #endif - if (cache_size) + if (tot_cache_size()) schedule_cache_check(); erts_mtx_unlock(&mseg_mutex); } @@ -1478,7 +1602,7 @@ erts_mseg_test(unsigned long op, case 0x406: { unsigned long res; erts_mtx_lock(&mseg_mutex); - res = (unsigned long) cache_size; + res = (unsigned long) tot_cache_size(); erts_mtx_unlock(&mseg_mutex); return res; } @@ -1569,11 +1693,14 @@ static void *do_map(void *ptr, size_t sz) return NULL; } - +#if HAVE_MMAP res = mmap(ptr, sz, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS | MAP_FIXED, -1 , 0); +#else +# error "Missing mmap support" +#endif if (res == MAP_FAILED) { #ifdef HARDDEBUG @@ -1673,10 +1800,19 @@ static int initialize_pmmap(void) MAP_NORESERVE | EXTRA_MAP_FLAGS, -1 , 0); #ifdef HARDDEBUG - printf("rsz = %ld, pages = %ld, rptr = %p\r\n", - (unsigned long) rsz, (unsigned long) (rsz / pagsz), - (void *) rptr); + printf("p=%p, rsz = %ld, pages = %ld, got range = %p -> %p\r\n", + p, (unsigned long) rsz, (unsigned long) (rsz / pagsz), + (void *) rptr, (void*)(rptr + rsz)); #endif + if ((UWord)(rptr + rsz) > RANGE_MAX) { + size_t rsz_trunc = RANGE_MAX - (UWord)rptr; +#ifdef HARDDEBUG + printf("Reducing mmap'ed memory from %lu to %lu Mb, reduced range = %p -> %p\r\n", + rsz/(1024*1024), rsz_trunc/(1024*1024), rptr, rptr+rsz_trunc); +#endif + munmap((void*)RANGE_MAX, rsz - rsz_trunc); + rsz = rsz_trunc; + } if (!do_map(rptr,pagsz)) { erl_exit(1,"Could not actually mmap first page for halfword emulator...\n"); } @@ -1757,6 +1893,7 @@ static int pmunmap(void *p, size_t size) FreeBlock *last; FreeBlock *nb = (FreeBlock *) p; + ASSERT(((unsigned long)p & CHECK_POINTER_MASK)==0); if (real_size > pagsz) { if (do_unmap(((char *) p) + pagsz,real_size - pagsz)) { return 1; diff --git a/erts/emulator/sys/common/erl_poll.c b/erts/emulator/sys/common/erl_poll.c index 4d0ca97889..3ae5b8d747 100644 --- a/erts/emulator/sys/common/erl_poll.c +++ b/erts/emulator/sys/common/erl_poll.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2006-2010. All Rights Reserved. + * Copyright Ericsson AB 2006-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -124,25 +124,11 @@ erts_smp_mtx_unlock(&(PS)->mtx) #define ERTS_POLLSET_SET_POLLED_CHK(PS) \ - ((int) erts_smp_atomic_xchg(&(PS)->polled, (erts_aint_t) 1)) + ((int) erts_atomic32_xchg(&(PS)->polled, (erts_aint32_t) 1)) #define ERTS_POLLSET_UNSET_POLLED(PS) \ - erts_smp_atomic_set(&(PS)->polled, (erts_aint_t) 0) + erts_atomic32_set(&(PS)->polled, (erts_aint32_t) 0) #define ERTS_POLLSET_IS_POLLED(PS) \ - ((int) erts_smp_atomic_read(&(PS)->polled)) - -#define ERTS_POLLSET_SET_POLLER_WOKEN_CHK(PS) set_poller_woken_chk((PS)) -#define ERTS_POLLSET_SET_POLLER_WOKEN(PS) \ -do { \ - ERTS_THR_MEMORY_BARRIER; \ - erts_smp_atomic_set(&(PS)->woken, (erts_aint_t) 1); \ -} while (0) -#define ERTS_POLLSET_UNSET_POLLER_WOKEN(PS) \ -do { \ - erts_smp_atomic_set(&(PS)->woken, (erts_aint_t) 0); \ - ERTS_THR_MEMORY_BARRIER; \ -} while (0) -#define ERTS_POLLSET_IS_POLLER_WOKEN(PS) \ - ((int) erts_smp_atomic_read(&(PS)->woken)) + ((int) erts_atomic32_read(&(PS)->polled)) #else @@ -152,69 +138,21 @@ do { \ #define ERTS_POLLSET_UNSET_POLLED(PS) #define ERTS_POLLSET_IS_POLLED(PS) 0 -#if ERTS_POLL_ASYNC_INTERRUPT_SUPPORT - -/* - * Ideally, the ERTS_POLLSET_SET_POLLER_WOKEN_CHK(PS) operation would - * be atomic. This operation isn't, but we will do okay anyway. The - * "woken check" is only an optimization. The only requirement we have: - * If (PS)->woken is set to a value != 0 when interrupting, we have to - * write on the the wakeup pipe at least once. Multiple writes are okay. - */ -#define ERTS_POLLSET_SET_POLLER_WOKEN_CHK(PS) ((PS)->woken++) -#define ERTS_POLLSET_SET_POLLER_WOKEN(PS) ((PS)->woken = 1, (void) 0) -#define ERTS_POLLSET_UNSET_POLLER_WOKEN(PS) ((PS)->woken = 0, (void) 0) -#define ERTS_POLLSET_IS_POLLER_WOKEN(PS) ((PS)->woken) - -#else - -#define ERTS_POLLSET_SET_POLLER_WOKEN_CHK(PS) 1 -#define ERTS_POLLSET_SET_POLLER_WOKEN(PS) -#define ERTS_POLLSET_UNSET_POLLER_WOKEN(PS) -#define ERTS_POLLSET_IS_POLLER_WOKEN(PS) 1 - -#endif - #endif #if ERTS_POLL_USE_UPDATE_REQUESTS_QUEUE #define ERTS_POLLSET_SET_HAVE_UPDATE_REQUESTS(PS) \ - erts_smp_atomic_set(&(PS)->have_update_requests, (erts_aint_t) 1) + erts_smp_atomic32_set(&(PS)->have_update_requests, (erts_aint32_t) 1) #define ERTS_POLLSET_UNSET_HAVE_UPDATE_REQUESTS(PS) \ - erts_smp_atomic_set(&(PS)->have_update_requests, (erts_aint_t) 0) + erts_smp_atomic32_set(&(PS)->have_update_requests, (erts_aint32_t) 0) #define ERTS_POLLSET_HAVE_UPDATE_REQUESTS(PS) \ - ((int) erts_smp_atomic_read(&(PS)->have_update_requests)) + ((int) erts_smp_atomic32_read(&(PS)->have_update_requests)) #else #define ERTS_POLLSET_SET_HAVE_UPDATE_REQUESTS(PS) #define ERTS_POLLSET_UNSET_HAVE_UPDATE_REQUESTS(PS) #define ERTS_POLLSET_HAVE_UPDATE_REQUESTS(PS) 0 #endif -#if ERTS_POLL_ASYNC_INTERRUPT_SUPPORT && !defined(ERTS_SMP) - -#define ERTS_POLLSET_UNSET_INTERRUPTED_CHK(PS) unset_interrupted_chk((PS)) -#define ERTS_POLLSET_UNSET_INTERRUPTED(PS) ((PS)->interrupt = 0, (void) 0) -#define ERTS_POLLSET_SET_INTERRUPTED(PS) ((PS)->interrupt = 1, (void) 0) -#define ERTS_POLLSET_IS_INTERRUPTED(PS) ((PS)->interrupt) - -#else - -#define ERTS_POLLSET_UNSET_INTERRUPTED_CHK(PS) unset_interrupted_chk((PS)) -#define ERTS_POLLSET_UNSET_INTERRUPTED(PS) \ -do { \ - erts_smp_atomic_set(&(PS)->interrupt, (erts_aint_t) 0); \ - ERTS_THR_MEMORY_BARRIER; \ -} while (0) -#define ERTS_POLLSET_SET_INTERRUPTED(PS) \ -do { \ - ERTS_THR_MEMORY_BARRIER; \ - erts_smp_atomic_set(&(PS)->interrupt, (erts_aint_t) 1); \ -} while (0) -#define ERTS_POLLSET_IS_INTERRUPTED(PS) \ - ((int) erts_smp_atomic_read(&(PS)->interrupt)) - -#endif - #if ERTS_POLL_USE_FALLBACK # if ERTS_POLL_USE_POLL # define ERTS_POLL_NEED_FALLBACK(PS) ((PS)->no_poll_fds > 1) @@ -318,14 +256,12 @@ struct ErtsPollSet_ { #if ERTS_POLL_USE_UPDATE_REQUESTS_QUEUE ErtsPollSetUpdateRequestsBlock update_requests; ErtsPollSetUpdateRequestsBlock *curr_upd_req_block; - erts_smp_atomic_t have_update_requests; + erts_smp_atomic32_t have_update_requests; #endif #ifdef ERTS_SMP - erts_smp_atomic_t polled; - erts_smp_atomic_t woken; + erts_atomic32_t polled; erts_smp_mtx_t mtx; #elif ERTS_POLL_ASYNC_INTERRUPT_SUPPORT - volatile int woken; #endif #if ERTS_POLL_USE_WAKEUP_PIPE int wake_fds[2]; @@ -333,12 +269,12 @@ struct ErtsPollSet_ { #if ERTS_POLL_USE_FALLBACK int fallback_used; #endif -#if ERTS_POLL_ASYNC_INTERRUPT_SUPPORT && !defined(ERTS_SMP) - volatile int interrupt; -#else - erts_smp_atomic_t interrupt; +#ifdef ERTS_SMP + erts_atomic32_t wakeup_state; +#elif ERTS_POLL_ASYNC_INTERRUPT_SUPPORT + volatile int wakeup_state; #endif - erts_smp_atomic_t timeout; + erts_smp_atomic32_t timeout; #ifdef ERTS_POLL_COUNT_AVOIDED_WAKEUPS erts_smp_atomic_t no_avoided_wakeups; erts_smp_atomic_t no_avoided_interrupts; @@ -346,34 +282,6 @@ struct ErtsPollSet_ { #endif }; -static ERTS_INLINE int -unset_interrupted_chk(ErtsPollSet ps) -{ - int res; -#if ERTS_POLL_ASYNC_INTERRUPT_SUPPORT && !defined(ERTS_SMP) - /* This operation isn't atomic, but we have no need at all for an - atomic operation here... */ - res = ps->interrupt; - ps->interrupt = 0; -#else - res = (int) erts_smp_atomic_xchg(&ps->interrupt, (erts_aint_t) 0); - ERTS_THR_MEMORY_BARRIER; -#endif - return res; - -} - -#ifdef ERTS_SMP - -static ERTS_INLINE int -set_poller_woken_chk(ErtsPollSet ps) -{ - ERTS_THR_MEMORY_BARRIER; - return (int) erts_smp_atomic_xchg(&ps->woken, (erts_aint_t) 1); -} - -#endif - void erts_silence_warn_unused_result(long unused); static void fatal_error(char *format, ...); static void fatal_error_async_signal_safe(char *error_str); @@ -430,6 +338,64 @@ static void check_poll_status(ErtsPollSet ps); static void print_misc_debug_info(void); #endif +#define ERTS_POLL_NOT_WOKEN 0 +#define ERTS_POLL_WOKEN -1 +#define ERTS_POLL_WOKEN_INTR 1 + +static ERTS_INLINE void +reset_wakeup_state(ErtsPollSet ps) +{ +#ifdef ERTS_SMP + erts_atomic32_set(&ps->wakeup_state, ERTS_POLL_NOT_WOKEN); + ERTS_THR_MEMORY_BARRIER; +#elif ERTS_POLL_ASYNC_INTERRUPT_SUPPORT + ps->wakeup_state = 0; +#endif +} + +static ERTS_INLINE int +is_woken(ErtsPollSet ps) +{ +#ifdef ERTS_SMP + return erts_atomic32_read_acqb(&ps->wakeup_state) != ERTS_POLL_NOT_WOKEN; +#elif ERTS_POLL_ASYNC_INTERRUPT_SUPPORT + return ps->wakeup_state != ERTS_POLL_NOT_WOKEN; +#else + return 0; +#endif +} + +static ERTS_INLINE int +is_interrupted_reset(ErtsPollSet ps) +{ +#ifdef ERTS_SMP + return (erts_atomic32_xchg(&ps->wakeup_state, ERTS_POLL_NOT_WOKEN) + == ERTS_POLL_WOKEN_INTR); +#elif ERTS_POLL_ASYNC_INTERRUPT_SUPPORT + int res = ps->wakeup_state == ERTS_POLL_WOKEN_INTR; + ps->wakeup_state = ERTS_POLL_NOT_WOKEN; + return res; +#else + return 0; +#endif +} + +static ERTS_INLINE void +woke_up(ErtsPollSet ps) +{ +#ifdef ERTS_SMP + erts_aint32_t wakeup_state = erts_atomic32_read(&ps->wakeup_state); + if (wakeup_state == ERTS_POLL_NOT_WOKEN) + (void) erts_atomic32_cmpxchg(&ps->wakeup_state, + ERTS_POLL_WOKEN, + ERTS_POLL_NOT_WOKEN); + ASSERT(erts_atomic32_read(&ps->wakeup_state) != ERTS_POLL_NOT_WOKEN); +#elif ERTS_POLL_ASYNC_INTERRUPT_SUPPORT + if (ps->wakeup_state == ERTS_POLL_NOT_WOKEN) + ps->wakeup_state = ERTS_POLL_WOKEN; +#endif +} + /* * --- Wakeup pipe ----------------------------------------------------------- */ @@ -437,14 +403,34 @@ static void print_misc_debug_info(void); #if ERTS_POLL_USE_WAKEUP_PIPE static ERTS_INLINE void -wake_poller(ErtsPollSet ps) +wake_poller(ErtsPollSet ps, int interrupted) { + int wake; +#ifdef ERTS_SMP + erts_aint32_t wakeup_state; + if (!interrupted) + wakeup_state = erts_atomic32_cmpxchg_relb(&ps->wakeup_state, + ERTS_POLL_WOKEN, + ERTS_POLL_NOT_WOKEN); + else { + /* + * We might unnecessarily write to the pipe, however, + * that isn't problematic. + */ + wakeup_state = erts_atomic32_read(&ps->wakeup_state); + erts_atomic32_set_relb(&ps->wakeup_state, ERTS_POLL_WOKEN_INTR); + } + wake = wakeup_state == ERTS_POLL_NOT_WOKEN; +#elif ERTS_POLL_ASYNC_INTERRUPT_SUPPORT + wake = ps->wakeup_state == ERTS_POLL_NOT_WOKEN; + ps->wakeup_state = interrupted ? ERTS_POLL_WOKEN_INTR : ERTS_POLL_NOT_WOKEN; +#endif /* * NOTE: This function might be called from signal handlers in the * non-smp case; therefore, it has to be async-signal safe in * the non-smp case. */ - if (!ERTS_POLLSET_SET_POLLER_WOKEN_CHK(ps)) { + if (wake) { ssize_t res; if (ps->wake_fds[1] < 0) return; /* Not initialized yet */ @@ -1387,9 +1373,7 @@ handle_update_requests(ErtsPollSet ps) #endif /* ERTS_POLL_USE_UPDATE_REQUESTS_QUEUE */ static ERTS_INLINE ErtsPollEvents -poll_control(ErtsPollSet ps, int fd, ErtsPollEvents events, int on, - int *have_set_have_update_requests, - int *do_wake) +poll_control(ErtsPollSet ps, int fd, ErtsPollEvents events, int on, int *do_wake) { ErtsPollEvents new_events; @@ -1493,7 +1477,6 @@ ERTS_POLL_EXPORT(erts_poll_controlv)(ErtsPollSet ps, int len) { int i; - int hshur = 0; int do_wake; int final_do_wake = 0; @@ -1505,17 +1488,17 @@ ERTS_POLL_EXPORT(erts_poll_controlv)(ErtsPollSet ps, pcev[i].fd, pcev[i].events, pcev[i].on, - &hshur, &do_wake); final_do_wake |= do_wake; } + ERTS_POLLSET_UNLOCK(ps); + #ifdef ERTS_SMP if (final_do_wake) - wake_poller(ps); + wake_poller(ps, 0); #endif /* ERTS_SMP */ - ERTS_POLLSET_UNLOCK(ps); } ErtsPollEvents @@ -1526,20 +1509,20 @@ ERTS_POLL_EXPORT(erts_poll_control)(ErtsPollSet ps, int* do_wake) /* In: Wake up polling thread */ /* Out: Poller is woken */ { - int hshur = 0; ErtsPollEvents res; ERTS_POLLSET_LOCK(ps); - res = poll_control(ps, fd, events, on, &hshur, do_wake); + res = poll_control(ps, fd, events, on, do_wake); + + ERTS_POLLSET_UNLOCK(ps); #ifdef ERTS_SMP if (*do_wake) { - wake_poller(ps); + wake_poller(ps, 0); } #endif /* ERTS_SMP */ - ERTS_POLLSET_UNLOCK(ps); return res; } @@ -1918,9 +1901,11 @@ check_fd_events(ErtsPollSet ps, SysTimeval *tv, int max_res, int *ps_locked) return 0; } else { - erts_aint_t timeout = tv->tv_sec*1000 + tv->tv_usec/1000; + long timeout = tv->tv_sec*1000 + tv->tv_usec/1000; + if (timeout > ERTS_AINT32_T_MAX) + timeout = ERTS_AINT32_T_MAX; ASSERT(timeout >= 0); - erts_smp_atomic_set(&ps->timeout, timeout); + erts_smp_atomic32_set_relb(&ps->timeout, (erts_aint32_t) timeout); #if ERTS_POLL_USE_FALLBACK if (!(ps->fallback_used = ERTS_POLL_NEED_FALLBACK(ps))) { @@ -2042,15 +2027,14 @@ ERTS_POLL_EXPORT(erts_poll_wait)(ErtsPollSet ps, (int) tv->tv_sec*1000 + tv->tv_usec/1000); #endif - ERTS_POLLSET_UNSET_POLLER_WOKEN(ps); if (ERTS_POLLSET_SET_POLLED_CHK(ps)) { res = EINVAL; /* Another thread is in erts_poll_wait() on this pollset... */ goto done; } - if (ERTS_POLLSET_IS_INTERRUPTED(ps)) { - /* Interrupt use zero timeout */ + if (is_woken(ps)) { + /* Use zero timeout */ itv.tv_sec = 0; itv.tv_usec = 0; tvp = &itv; @@ -2067,7 +2051,7 @@ ERTS_POLL_EXPORT(erts_poll_wait)(ErtsPollSet ps, ps_locked = 0; res = check_fd_events(ps, tvp, no_fds, &ps_locked); - ERTS_POLLSET_SET_POLLER_WOKEN(ps); + woke_up(ps); if (res == 0) { res = ETIMEDOUT; @@ -2099,9 +2083,7 @@ ERTS_POLL_EXPORT(erts_poll_wait)(ErtsPollSet ps, check_poll_result(pr, no_fds); #endif - res = (no_fds == 0 - ? (ERTS_POLLSET_UNSET_INTERRUPTED_CHK(ps) ? EINTR : EAGAIN) - : 0); + res = (no_fds == 0 ? (is_interrupted_reset(ps) ? EINTR : EAGAIN) : 0); *len = no_fds; } @@ -2112,7 +2094,7 @@ ERTS_POLL_EXPORT(erts_poll_wait)(ErtsPollSet ps, #endif done: - erts_smp_atomic_set(&ps->timeout, ERTS_AINT_T_MAX); + erts_smp_atomic32_set_relb(&ps->timeout, ERTS_AINT32_T_MAX); #ifdef ERTS_POLL_DEBUG_PRINT erts_printf("Leaving %s = erts_poll_wait()\n", res == 0 ? "0" : erl_errno_id(res)); @@ -2128,20 +2110,17 @@ ERTS_POLL_EXPORT(erts_poll_wait)(ErtsPollSet ps, void ERTS_POLL_EXPORT(erts_poll_interrupt)(ErtsPollSet ps, int set) { +#if ERTS_POLL_ASYNC_INTERRUPT_SUPPORT || defined(ERTS_SMP) /* * NOTE: This function might be called from signal handlers in the * non-smp case; therefore, it has to be async-signal safe in * the non-smp case. */ - if (set) { - ERTS_POLLSET_SET_INTERRUPTED(ps); -#if ERTS_POLL_ASYNC_INTERRUPT_SUPPORT || defined(ERTS_SMP) - wake_poller(ps); + if (!set) + reset_wakeup_state(ps); + else + wake_poller(ps, 1); #endif - } - else { - ERTS_POLLSET_UNSET_INTERRUPTED(ps); - } } /* @@ -2154,13 +2133,12 @@ ERTS_POLL_EXPORT(erts_poll_interrupt_timed)(ErtsPollSet ps, int set, long msec) { - if (set) { - if (erts_smp_atomic_read(&ps->timeout) > (erts_aint_t) msec) { - ERTS_POLLSET_SET_INTERRUPTED(ps); #if ERTS_POLL_ASYNC_INTERRUPT_SUPPORT || defined(ERTS_SMP) - wake_poller(ps); -#endif - } + if (!set) + reset_wakeup_state(ps); + else { + if (erts_smp_atomic32_read_acqb(&ps->timeout) > (erts_aint32_t) msec) + wake_poller(ps, 1); #ifdef ERTS_POLL_COUNT_AVOIDED_WAKEUPS else { if (ERTS_POLLSET_IS_POLLED(ps)) @@ -2170,9 +2148,7 @@ ERTS_POLL_EXPORT(erts_poll_interrupt_timed)(ErtsPollSet ps, erts_smp_atomic_inc(&ps->no_interrupt_timed); #endif } - else { - ERTS_POLLSET_UNSET_INTERRUPTED(ps); - } +#endif } int @@ -2283,14 +2259,16 @@ ERTS_POLL_EXPORT(erts_poll_create_pollset)(void) ps->update_requests.next = NULL; ps->update_requests.len = 0; ps->curr_upd_req_block = &ps->update_requests; - erts_smp_atomic_init(&ps->have_update_requests, 0); + erts_smp_atomic32_init(&ps->have_update_requests, 0); #endif #ifdef ERTS_SMP - erts_smp_atomic_init(&ps->polled, 0); - erts_smp_atomic_init(&ps->woken, 0); + erts_atomic32_init(&ps->polled, 0); erts_smp_mtx_init(&ps->mtx, "pollset"); +#endif +#ifdef ERTS_SMP + erts_atomic32_init(&ps->wakeup_state, (erts_aint32_t) 0); #elif ERTS_POLL_ASYNC_INTERRUPT_SUPPORT - ps->woken = 0; + ps->wakeup_state = 0; #endif #if ERTS_POLL_USE_WAKEUP_PIPE create_wakeup_pipe(ps); @@ -2312,12 +2290,7 @@ ERTS_POLL_EXPORT(erts_poll_create_pollset)(void) ps->internal_fd_limit = kp_fd + 1; ps->kp_fd = kp_fd; #endif -#if ERTS_POLL_ASYNC_INTERRUPT_SUPPORT && !defined(ERTS_SMP) - ps->interrupt = 0; -#else - erts_smp_atomic_init(&ps->interrupt, 0); -#endif - erts_smp_atomic_init(&ps->timeout, ERTS_AINT_T_MAX); + erts_smp_atomic32_init(&ps->timeout, ERTS_AINT32_T_MAX); #ifdef ERTS_POLL_COUNT_AVOIDED_WAKEUPS erts_smp_atomic_init(&ps->no_avoided_wakeups, 0); erts_smp_atomic_init(&ps->no_avoided_interrupts, 0); diff --git a/erts/emulator/sys/unix/sys.c b/erts/emulator/sys/unix/sys.c index bfc04faa45..bafbbb0f6c 100644 --- a/erts/emulator/sys/unix/sys.c +++ b/erts/emulator/sys/unix/sys.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 1996-2010. All Rights Reserved. + * Copyright Ericsson AB 1996-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -53,6 +53,11 @@ #define WANT_NONBLOCKING /* must define this to pull in defs from sys.h */ #include "sys.h" +#if defined(__APPLE__) && defined(__MACH__) && !defined(__DARWIN__) +#define __DARWIN__ 1 +#endif + + #ifdef USE_THREADS #include "erl_threads.h" #endif @@ -160,14 +165,14 @@ static int debug_log = 0; #endif #ifdef ERTS_SMP -erts_smp_atomic_t erts_got_sigusr1; +erts_smp_atomic32_t erts_got_sigusr1; #define ERTS_SET_GOT_SIGUSR1 \ - erts_smp_atomic_set(&erts_got_sigusr1, 1) + erts_smp_atomic32_set(&erts_got_sigusr1, 1) #define ERTS_UNSET_GOT_SIGUSR1 \ - erts_smp_atomic_set(&erts_got_sigusr1, 0) -static erts_smp_atomic_t have_prepared_crash_dump; + erts_smp_atomic32_set(&erts_got_sigusr1, 0) +static erts_smp_atomic32_t have_prepared_crash_dump; #define ERTS_PREPARED_CRASH_DUMP \ - ((int) erts_smp_atomic_xchg(&have_prepared_crash_dump, 1)) + ((int) erts_smp_atomic32_xchg(&have_prepared_crash_dump, 1)) #else volatile int erts_got_sigusr1; #define ERTS_SET_GOT_SIGUSR1 (erts_got_sigusr1 = 1) @@ -235,11 +240,11 @@ static int max_files = -1; * a few variables used by the break handler */ #ifdef ERTS_SMP -erts_smp_atomic_t erts_break_requested; +erts_smp_atomic32_t erts_break_requested; #define ERTS_SET_BREAK_REQUESTED \ - erts_smp_atomic_set(&erts_break_requested, (erts_aint_t) 1) + erts_smp_atomic32_set(&erts_break_requested, (erts_aint32_t) 1) #define ERTS_UNSET_BREAK_REQUESTED \ - erts_smp_atomic_set(&erts_break_requested, (erts_aint_t) 0) + erts_smp_atomic32_set(&erts_break_requested, (erts_aint32_t) 0) #else volatile int erts_break_requested = 0; #define ERTS_SET_BREAK_REQUESTED (erts_break_requested = 1) @@ -504,9 +509,9 @@ erts_sys_pre_init(void) #endif } #ifdef ERTS_SMP - erts_smp_atomic_init(&erts_break_requested, 0); - erts_smp_atomic_init(&erts_got_sigusr1, 0); - erts_smp_atomic_init(&have_prepared_crash_dump, 0); + erts_smp_atomic32_init(&erts_break_requested, 0); + erts_smp_atomic32_init(&erts_got_sigusr1, 0); + erts_smp_atomic32_init(&have_prepared_crash_dump, 0); #else erts_break_requested = 0; erts_got_sigusr1 = 0; @@ -2989,11 +2994,27 @@ init_smp_sig_notify(void) NULL, &thr_opts); } +#ifdef __DARWIN__ + +int erts_darwin_main_thread_pipe[2]; +int erts_darwin_main_thread_result_pipe[2]; + +static void initialize_darwin_main_thread_pipes(void) +{ + if (pipe(erts_darwin_main_thread_pipe) < 0 || + pipe(erts_darwin_main_thread_result_pipe) < 0) { + erl_exit(1,"Fatal error initializing Darwin main thread stealing"); + } +} +#endif void erts_sys_main_thread(void) { erts_thread_disable_fpe(); +#ifdef __DARWIN__ + initialize_darwin_main_thread_pipes(); +#endif /* Become signal receiver thread... */ #ifdef ERTS_ENABLE_LOCK_CHECK erts_lc_set_thread_name("signal_receiver"); @@ -3002,6 +3023,27 @@ erts_sys_main_thread(void) smp_sig_notify(0); /* Notify initialized */ while (1) { /* Wait for a signal to arrive... */ +#ifdef __DARWIN__ + /* + * The wx driver needs to be able to steal the main thread for Cocoa to + * work properly. + */ + fd_set readfds; + int res; + + FD_ZERO(&readfds); + FD_SET(erts_darwin_main_thread_pipe[0], &readfds); + res = select(erts_darwin_main_thread_pipe[0] + 1, &readfds, NULL, NULL, NULL); + if (res > 0 && FD_ISSET(erts_darwin_main_thread_pipe[0],&readfds)) { + void* (*func)(void*); + void* arg; + void *resp; + read(erts_darwin_main_thread_pipe[0],&func,sizeof(void* (*)(void*))); + read(erts_darwin_main_thread_pipe[0],&arg, sizeof(void*)); + resp = (*func)(arg); + write(erts_darwin_main_thread_result_pipe[1],&resp,sizeof(void *)); + } +#else #ifdef DEBUG int res = #else @@ -3010,6 +3052,7 @@ erts_sys_main_thread(void) select(0, NULL, NULL, NULL, NULL); ASSERT(res < 0); ASSERT(errno == EINTR); +#endif } } diff --git a/erts/emulator/sys/unix/sys_float.c b/erts/emulator/sys/unix/sys_float.c index 6e9376b0f3..8ec7b31ce0 100644 --- a/erts/emulator/sys/unix/sys_float.c +++ b/erts/emulator/sys/unix/sys_float.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2001-2010. All Rights Reserved. + * Copyright Ericsson AB 2001-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -36,11 +36,6 @@ erts_sys_init_float(void) # endif } -static ERTS_INLINE void set_current_fp_exception(unsigned long pc) -{ - /* nothing to do */ -} - #else /* !NO_FPE_SIGNALS */ #ifdef ERTS_SMP diff --git a/erts/emulator/sys/win32/erl_poll.c b/erts/emulator/sys/win32/erl_poll.c index d84ae2ede2..7662f190ef 100644 --- a/erts/emulator/sys/win32/erl_poll.c +++ b/erts/emulator/sys/win32/erl_poll.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 2007-2010. All Rights Reserved. + * Copyright Ericsson AB 2007-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -274,7 +274,6 @@ struct ErtsPollSet_ { Waiter** waiter; int allocated_waiters; /* Size ow waiter array */ int num_waiters; /* Number of waiter threads. */ - erts_atomic_t sys_io_ready; /* Tells us there is I/O ready (already). */ int restore_events; /* Tells us to restore waiters events next time around */ HANDLE event_io_ready; /* To be used when waiting for io */ @@ -282,12 +281,11 @@ struct ErtsPollSet_ { volatile int standby_wait_counter; /* Number of threads to wait for */ CRITICAL_SECTION standby_crit; /* CS to guard the counter */ HANDLE standby_wait_event; /* Event signalled when counte == 0 */ + erts_atomic32_t wakeup_state; #ifdef ERTS_SMP - erts_smp_atomic_t woken; erts_smp_mtx_t mtx; - erts_smp_atomic_t interrupt; #endif - erts_smp_atomic_t timeout; + erts_smp_atomic32_t timeout; }; #ifdef ERTS_SMP @@ -296,126 +294,24 @@ struct ErtsPollSet_ { erts_smp_mtx_lock(&(PS)->mtx) #define ERTS_POLLSET_UNLOCK(PS) \ erts_smp_mtx_unlock(&(PS)->mtx) -#define ERTS_POLLSET_SET_POLLED_CHK(PS) \ - ((int) erts_smp_atomic_xchg(&(PS)->polled, (erts_aint_t) 1)) -#define ERTS_POLLSET_SET_POLLED(PS) \ - erts_smp_atomic_set(&(PS)->polled, (erts_aint_t) 1) -#define ERTS_POLLSET_UNSET_POLLED(PS) \ - erts_smp_atomic_set(&(PS)->polled, (erts_aint_t) 0) -#define ERTS_POLLSET_IS_POLLED(PS) \ - ((int) erts_smp_atomic_read(&(PS)->polled)) - -#define ERTS_POLLSET_SET_POLLER_WOKEN_CHK(PS) set_poller_woken_chk((PS)) -#define ERTS_POLLSET_SET_POLLER_WOKEN(PS) \ -do { \ - ERTS_THR_MEMORY_BARRIER; \ - erts_smp_atomic_set(&(PS)->woken, (erts_aint_t) 1); \ -} while (0) -#define ERTS_POLLSET_UNSET_POLLER_WOKEN(PS) \ -do { \ - erts_smp_atomic_set(&(PS)->woken, (erts_aint_t) 0); \ - ERTS_THR_MEMORY_BARRIER; \ -} while (0) -#define ERTS_POLLSET_IS_POLLER_WOKEN(PS) \ - ((int) erts_smp_atomic_read(&(PS)->woken)) - -#define ERTS_POLLSET_UNSET_INTERRUPTED_CHK(PS) unset_interrupted_chk((PS)) -#define ERTS_POLLSET_UNSET_INTERRUPTED(PS) \ -do { \ - erts_smp_atomic_set(&(PS)->interrupt, (erts_aint_t) 0); \ - ERTS_THR_MEMORY_BARRIER; \ -} while (0) -#define ERTS_POLLSET_SET_INTERRUPTED(PS) \ -do { \ - ERTS_THR_MEMORY_BARRIER; \ - erts_smp_atomic_set(&(PS)->interrupt, (erts_aint_t) 1); \ -} while (0) -#define ERTS_POLLSET_IS_INTERRUPTED(PS) \ - ((int) erts_smp_atomic_read(&(PS)->interrupt)) - -static ERTS_INLINE int -unset_interrupted_chk(ErtsPollSet ps) -{ - int res = (int) erts_smp_atomic_xchg(&ps->interrupt, (erts_aint_t) 0); - ERTS_THR_MEMORY_BARRIER; - return res; - -} - -static ERTS_INLINE int -set_poller_woken_chk(ErtsPollSet ps) -{ - ERTS_THR_MEMORY_BARRIER; - return (int) erts_smp_atomic_xchg(&ps->woken, (erts_aint_t) 1); -} #else #define ERTS_POLLSET_LOCK(PS) #define ERTS_POLLSET_UNLOCK(PS) -#define ERTS_POLLSET_SET_POLLED_CHK(PS) 0 -#define ERTS_POLLSET_UNSET_POLLED(PS) -#define ERTS_POLLSET_IS_POLLED(PS) 0 -#define ERTS_POLLSET_SET_POLLER_WOKEN_CHK(PS) 1 -#define ERTS_POLLSET_SET_POLLER_WOKEN(PS) -#define ERTS_POLLSET_UNSET_POLLER_WOKEN(PS) -#define ERTS_POLLSET_IS_POLLER_WOKEN(PS) 1 - #endif /* - * While atomics are not yet implemented for windows in the common library... - * - * MSDN doc states that SMP machines and old compilers require - * InterLockedExchange to properly read and write interlocked - * variables, otherwise the processors might reschedule - * the access and order of atomics access is destroyed... - * While they only mention it in white-papers, the problem - * in VS2003 is due to the IA64 arch, so we can still count - * on the CPU not rescheduling the access to volatile in X86 arch using - * even the slightly older compiler... - * - * So here's (hopefully) a subset of the generally working atomic - * variable access... - */ - -#if defined(__GNUC__) -# if defined(__i386__) || defined(__x86_64__) -# define VOLATILE_IN_SEQUENCE 1 -# else -# define VOLATILE_IN_SEQUENCE 0 -# endif -#elif defined(_MSC_VER) -# if _MSC_VER < 1300 -# define VOLATILE_IN_SEQUENCE 0 /* Dont trust really old compilers */ -# else -# if defined(_M_IX86) -# define VOLATILE_IN_SEQUENCE 1 -# else /* I.e. IA64 */ -# if _MSC_VER >= 1400 -# define VOLATILE_IN_SEQUENCE 1 -# else -# define VOLATILE_IN_SEQUENCE 0 -# endif -# endif -# endif -#else -# define VOLATILE_IN_SEQUENCE 0 -#endif - - - -/* * Communication with sys_interrupt */ #ifdef ERTS_SMP -extern erts_smp_atomic_t erts_break_requested; +extern erts_smp_atomic32_t erts_break_requested; #define ERTS_SET_BREAK_REQUESTED \ - erts_smp_atomic_set(&erts_break_requested, (erts_aint_t) 1) + erts_smp_atomic32_set(&erts_break_requested, (erts_aint32_t) 1) #define ERTS_UNSET_BREAK_REQUESTED \ - erts_smp_atomic_set(&erts_break_requested, (erts_aint_t) 0) + erts_smp_atomic32_set(&erts_break_requested, (erts_aint32_t) 0) #else extern volatile int erts_break_requested; #define ERTS_SET_BREAK_REQUESTED (erts_break_requested = 1) @@ -424,7 +320,7 @@ extern volatile int erts_break_requested; static erts_mtx_t break_waiter_lock; static HANDLE break_happened_event; -static erts_atomic_t break_waiter_state; +static erts_atomic32_t break_waiter_state; #define BREAK_WAITER_GOT_BREAK 1 #define BREAK_WAITER_GOT_HALT 2 @@ -467,29 +363,172 @@ do { \ wait_standby(PS); \ } while(0) -#if ERTS_POLL_ASYNC_INTERRUPT_SUPPORT && !defined(ERTS_SMP) +#define ERTS_POLL_NOT_WOKEN ((erts_aint32_t) 0) +#define ERTS_POLL_WOKEN_IO_READY ((erts_aint32_t) 1) +#define ERTS_POLL_WOKEN_INTR ((erts_aint32_t) 2) +#define ERTS_POLL_WOKEN_TIMEDOUT ((erts_aint32_t) 3) static ERTS_INLINE int -unset_interrupted_chk(ErtsPollSet ps) +is_io_ready(ErtsPollSet ps) { - /* This operation isn't atomic, but we have no need at all for an - atomic operation here... */ - int res = ps->interrupt; - ps->interrupt = 0; - return res; + return erts_atomic32_read(&ps->wakeup_state) == ERTS_POLL_WOKEN_IO_READY; } +static ERTS_INLINE void +woke_up(ErtsPollSet ps) +{ + if (erts_atomic32_read(&ps->wakeup_state) == ERTS_POLL_NOT_WOKEN) + erts_atomic32_cmpxchg(&ps->wakeup_state, + ERTS_POLL_WOKEN_TIMEDOUT, + ERTS_POLL_NOT_WOKEN); +#ifdef DEBUG + { + erts_aint32_t wakeup_state = erts_atomic32_read(&ps->wakeup_state); + switch (wakeup_state) { + case ERTS_POLL_WOKEN_IO_READY: + case ERTS_POLL_WOKEN_INTR: + case ERTS_POLL_WOKEN_TIMEDOUT: + break; + default: + ASSERT(0); + break; + } + } #endif +} + +static ERTS_INLINE int +wakeup_cause(ErtsPollSet ps) +{ + int res; + erts_aint32_t wakeup_state = erts_atomic32_read(&ps->wakeup_state); + switch (wakeup_state) { + case ERTS_POLL_WOKEN_IO_READY: + res = 0; + break; + case ERTS_POLL_WOKEN_INTR: + res = EINTR; + break; + case ERTS_POLL_WOKEN_TIMEDOUT: + res = ETIMEDOUT; + break; + default: + res = 0; + erl_exit(ERTS_ABORT_EXIT, + "%s:%d: Internal error: Invalid wakeup_state=%d\n", + __FILE__, __LINE__, (int) wakeup_state); + } + return res; +} + +static ERTS_INLINE DWORD +poll_wait_timeout(ErtsPollSet ps, SysTimeval *tvp) +{ + time_t timeout = tvp->tv_sec * 1000 + tvp->tv_usec / 1000; + + if (timeout <= 0) { + woke_up(ps); + return (DWORD) 0; + } + + ResetEvent(ps->event_io_ready); + /* + * Since we don't know the internals of ResetEvent() we issue + * a memory barrier as a safety precaution ensuring that + * the load of wakeup_state wont be reordered with stores made + * by ResetEvent(). + */ + ERTS_THR_MEMORY_BARRIER; + if (erts_atomic32_read(&ps->wakeup_state) != ERTS_POLL_NOT_WOKEN) + return (DWORD) 0; + + if (timeout > ERTS_AINT32_T_MAX) /* Also prevents DWORD overflow */ + timeout = ERTS_AINT32_T_MAX; + + erts_smp_atomic32_set_relb(&ps->timeout, (erts_aint32_t) timeout); + return (DWORD) timeout; +} -#ifdef ERTS_SMP static ERTS_INLINE void -wake_poller(ErtsPollSet ps) +wake_poller(ErtsPollSet ps, int io_ready) { - if (!ERTS_POLLSET_SET_POLLER_WOKEN_CHK(ps)) { + erts_aint32_t wakeup_state; + if (io_ready) { + /* We may set the event multiple times. This is, however, harmless. */ + wakeup_state = erts_atomic32_read(&ps->wakeup_state); + erts_atomic32_set_relb(&ps->wakeup_state, ERTS_POLL_WOKEN_IO_READY); + } + else { + ERTS_THR_MEMORY_BARRIER; + wakeup_state = erts_atomic32_read(&ps->wakeup_state); + while (wakeup_state != ERTS_POLL_WOKEN_IO_READY + && wakeup_state != ERTS_POLL_WOKEN_INTR) { + erts_aint32_t act = erts_atomic32_cmpxchg(&ps->wakeup_state, + ERTS_POLL_WOKEN_INTR, + wakeup_state); + if (act == wakeup_state) { + wakeup_state = act; + break; + } + wakeup_state = act; + } + } + if (wakeup_state == ERTS_POLL_NOT_WOKEN) { + /* + * Since we don't know the internals of SetEvent() we issue + * a memory barrier as a safety precaution ensuring that + * the store we just made to wakeup_state wont be reordered + * with loads in SetEvent(). + */ + ERTS_THR_MEMORY_BARRIER; SetEvent(ps->event_io_ready); } } -#endif + +static ERTS_INLINE void +reset_io_ready(ErtsPollSet ps) +{ + erts_atomic32_set(&ps->wakeup_state, ERTS_POLL_NOT_WOKEN); +} + +static ERTS_INLINE void +restore_io_ready(ErtsPollSet ps) +{ + erts_atomic32_set(&ps->wakeup_state, ERTS_POLL_WOKEN_IO_READY); +} + +/* + * notify_io_ready() is used by threads waiting for events, when + * notifying a poller thread about I/O ready. + */ +static ERTS_INLINE void +notify_io_ready(ErtsPollSet ps) +{ + wake_poller(ps, 1); +} + +static ERTS_INLINE void +reset_interrupt(ErtsPollSet ps) +{ + /* We need to keep io-ready if set */ + erts_aint32_t wakeup_state = erts_atomic32_read(&ps->wakeup_state); + while (wakeup_state != ERTS_POLL_WOKEN_IO_READY + && wakeup_state != ERTS_POLL_NOT_WOKEN) { + erts_aint32_t act = erts_atomic32_cmpxchg(&ps->wakeup_state, + ERTS_POLL_NOT_WOKEN, + wakeup_state); + if (wakeup_state == act) + break; + wakeup_state = act; + } + ERTS_THR_MEMORY_BARRIER; +} + +static ERTS_INLINE void +set_interrupt(ErtsPollSet ps) +{ + wake_poller(ps, 0); +} static void setup_standby_wait(ErtsPollSet ps, int num_threads) { @@ -653,14 +692,14 @@ static void *break_waiter(void *param) case WAIT_OBJECT_0: ResetEvent(harr[0]); erts_mtx_lock(&break_waiter_lock); - erts_atomic_set(&break_waiter_state,BREAK_WAITER_GOT_BREAK); + erts_atomic32_set(&break_waiter_state,BREAK_WAITER_GOT_BREAK); SetEvent(break_happened_event); erts_mtx_unlock(&break_waiter_lock); break; case (WAIT_OBJECT_0+1): ResetEvent(harr[1]); erts_mtx_lock(&break_waiter_lock); - erts_atomic_set(&break_waiter_state,BREAK_WAITER_GOT_HALT); + erts_atomic32_set(&break_waiter_state,BREAK_WAITER_GOT_HALT); SetEvent(break_happened_event); erts_mtx_unlock(&break_waiter_lock); break; @@ -767,12 +806,7 @@ event_happened: consistency_check(w); #endif ASSERT(WAIT_OBJECT_0 < i && i < WAIT_OBJECT_0+w->active_events); - if (!erts_atomic_xchg(&ps->sys_io_ready,1)) { - HARDDEBUGF(("SET EventIoReady (%d)",erts_atomic_read(&ps->sys_io_ready))); - SetEvent(ps->event_io_ready); - } else { - HARDDEBUGF(("DONT SET EventIoReady")); - } + notify_io_ready(ps); /* * The main thread wont start working on our arrays untill we're @@ -967,15 +1001,10 @@ static int cancel_driver_select(ErtsPollSet ps, HANDLE event) void erts_poll_interrupt(ErtsPollSet ps, int set /* bool */) { HARDTRACEF(("In erts_poll_interrupt(%d)",set)); -#ifdef ERTS_SMP - if (set) { - ERTS_POLLSET_SET_INTERRUPTED(ps); - wake_poller(ps); - } - else { - ERTS_POLLSET_UNSET_INTERRUPTED(ps); - } -#endif + if (!set) + reset_interrupt(ps); + else + set_interrupt(ps); HARDTRACEF(("Out erts_poll_interrupt(%d)",set)); } @@ -984,17 +1013,10 @@ void erts_poll_interrupt_timed(ErtsPollSet ps, long msec) { HARDTRACEF(("In erts_poll_interrupt_timed(%d,%ld)",set,msec)); -#ifdef ERTS_SMP - if (set) { - if (erts_smp_atomic_read(&ps->timeout) > (erts_aint_t) msec) { - ERTS_POLLSET_SET_INTERRUPTED(ps); - wake_poller(ps); - } - } - else { - ERTS_POLLSET_UNSET_INTERRUPTED(ps); - } -#endif + if (!set) + reset_interrupt(ps); + else if (erts_smp_atomic32_read_acqb(&ps->timeout) > (erts_aint32_t) msec) + set_interrupt(ps); HARDTRACEF(("Out erts_poll_interrupt_timed")); } @@ -1068,10 +1090,8 @@ void erts_poll_controlv(ErtsPollSet ps, int erts_poll_wait(ErtsPollSet ps, ErtsPollResFd pr[], int *len, - SysTimeval *utvp) + SysTimeval *tvp) { - SysTimeval *tvp = utvp; - SysTimeval itv; int no_fds; DWORD timeout; EventData* ev; @@ -1084,7 +1104,7 @@ int erts_poll_wait(ErtsPollSet ps, HARDTRACEF(("In erts_poll_wait")); ERTS_POLLSET_LOCK(ps); - if (!erts_atomic_read(&ps->sys_io_ready) && ps->restore_events) { + if (!is_io_ready(ps) && ps->restore_events) { HARDDEBUGF(("Restore events: %d",ps->num_waiters)); ps->restore_events = 0; for (i = 0; i < ps->num_waiters; ++i) { @@ -1102,7 +1122,7 @@ int erts_poll_wait(ErtsPollSet ps, if (w->highwater != w->active_events) { HARDDEBUGF(("Oups!")); /* Oups, got signalled before we took the lock, can't reset */ - if(erts_atomic_read(&ps->sys_io_ready) == 0) { + if(!is_io_ready(ps)) { erl_exit(1,"Internal error: " "Inconsistent io structures in erl_poll.\n"); } @@ -1127,39 +1147,27 @@ int erts_poll_wait(ErtsPollSet ps, no_fds = ERTS_POLL_MAX_RES; #endif + timeout = poll_wait_timeout(ps, tvp); - ResetEvent(ps->event_io_ready); - ERTS_POLLSET_UNSET_POLLER_WOKEN(ps); - -#ifdef ERTS_SMP - if (ERTS_POLLSET_IS_INTERRUPTED(ps)) { - /* Interrupt use zero timeout */ - itv.tv_sec = 0; - itv.tv_usec = 0; - tvp = &itv; - } -#endif - - timeout = tvp->tv_sec * 1000 + tvp->tv_usec / 1000; /*HARDDEBUGF(("timeout = %ld",(long) timeout));*/ - erts_smp_atomic_set(&ps->timeout, timeout); - if (timeout > 0 && ! erts_atomic_read(&ps->sys_io_ready) && ! erts_atomic_read(&break_waiter_state)) { + if (timeout > 0 && !erts_atomic32_read(&break_waiter_state)) { HANDLE harr[2] = {ps->event_io_ready, break_happened_event}; int num_h = 2; - HARDDEBUGF(("Start waiting %d [%d]",num_h, (long) timeout)); + HARDDEBUGF(("Start waiting %d [%d]",num_h, (int) timeout)); ERTS_POLLSET_UNLOCK(ps); WaitForMultipleObjects(num_h, harr, FALSE, timeout); ERTS_POLLSET_LOCK(ps); - HARDDEBUGF(("Stop waiting %d [%d]",num_h, (long) timeout)); + HARDDEBUGF(("Stop waiting %d [%d]",num_h, (int) timeout)); + woke_up(ps); } ERTS_UNSET_BREAK_REQUESTED; - if(erts_atomic_read(&break_waiter_state)) { + if(erts_atomic32_read(&break_waiter_state)) { erts_mtx_lock(&break_waiter_lock); - break_state = erts_atomic_read(&break_waiter_state); - erts_atomic_set(&break_waiter_state,0); + break_state = erts_atomic32_read(&break_waiter_state); + erts_atomic32_set(&break_waiter_state,0); ResetEvent(break_happened_event); erts_mtx_unlock(&break_waiter_lock); switch (break_state) { @@ -1174,15 +1182,13 @@ int erts_poll_wait(ErtsPollSet ps, } } - ERTS_POLLSET_SET_POLLER_WOKEN(ps); - - if (!erts_atomic_read(&ps->sys_io_ready)) { - res = EINTR; - HARDDEBUGF(("EINTR!")); - goto done; + res = wakeup_cause(ps); + if (res != 0) { + HARDDEBUGF(("%s!", res == EINTR ? "EINTR" : "ETIMEDOUT")); + goto done; } - erts_atomic_set(&ps->sys_io_ready,0); + reset_io_ready(ps); n = ps->num_waiters; @@ -1204,9 +1210,9 @@ int erts_poll_wait(ErtsPollSet ps, if (num >= no_fds) { w->highwater=j+1; erts_mtx_unlock(&w->mtx); - /* This might mean we still have data to report, set - back the global flag! */ - erts_atomic_set(&ps->sys_io_ready,1); + /* This might mean we still have data to report, + restore flag indicating I/O ready! */ + restore_io_ready(ps); HARDDEBUGF(("To many FD's to report!")); goto done; } @@ -1228,7 +1234,7 @@ int erts_poll_wait(ErtsPollSet ps, erts_mtx_unlock(&w->mtx); } done: - erts_smp_atomic_set(&ps->timeout, ERTS_AINT_T_MAX); + erts_smp_atomic32_set(&ps->timeout, ERTS_AINT32_T_MAX); *len = num; ERTS_POLLSET_UNLOCK(ps); HARDTRACEF(("Out erts_poll_wait")); @@ -1306,15 +1312,13 @@ ErtsPollSet erts_poll_create_pollset(void) ps->standby_wait_counter = 0; ps->event_io_ready = CreateManualEvent(FALSE); ps->standby_wait_event = CreateManualEvent(FALSE); - erts_atomic_init(&ps->sys_io_ready,0); ps->restore_events = 0; + erts_atomic32_init(&ps->wakeup_state, ERTS_POLL_NOT_WOKEN); #ifdef ERTS_SMP - erts_smp_atomic_init(&ps->woken, 0); erts_smp_mtx_init(&ps->mtx, "pollset"); - erts_smp_atomic_init(&ps->interrupt, 0); #endif - erts_smp_atomic_init(&ps->timeout, ERTS_AINT_T_MAX); + erts_smp_atomic32_init(&ps->timeout, ERTS_AINT32_T_MAX); HARDTRACEF(("Out erts_poll_create_pollset")); return ps; @@ -1366,7 +1370,7 @@ void erts_poll_init(void) erts_mtx_init(&break_waiter_lock,"break_waiter_lock"); break_happened_event = CreateManualEvent(FALSE); - erts_atomic_init(&break_waiter_state, 0); + erts_atomic32_init(&break_waiter_state, 0); erts_thr_create(&thread, &break_waiter, NULL, NULL); ERTS_UNSET_BREAK_REQUESTED; diff --git a/erts/emulator/sys/win32/sys_interrupt.c b/erts/emulator/sys/win32/sys_interrupt.c index 262f84babc..943c338794 100644 --- a/erts/emulator/sys/win32/sys_interrupt.c +++ b/erts/emulator/sys/win32/sys_interrupt.c @@ -1,7 +1,7 @@ /* * %CopyrightBegin% * - * Copyright Ericsson AB 1997-2010. All Rights Reserved. + * Copyright Ericsson AB 1997-2011. All Rights Reserved. * * The contents of this file are subject to the Erlang Public License, * Version 1.1, (the "License"); you may not use this file except in @@ -31,11 +31,11 @@ #endif #ifdef ERTS_SMP -erts_smp_atomic_t erts_break_requested; +erts_smp_atomic32_t erts_break_requested; #define ERTS_SET_BREAK_REQUESTED \ - erts_smp_atomic_set(&erts_break_requested, (erts_aint_t) 1) + erts_smp_atomic32_set(&erts_break_requested, (erts_aint32_t) 1) #define ERTS_UNSET_BREAK_REQUESTED \ - erts_smp_atomic_set(&erts_break_requested, (erts_aint_t) 0) + erts_smp_atomic32_set(&erts_break_requested, (erts_aint32_t) 0) #else volatile int erts_break_requested = 0; #define ERTS_SET_BREAK_REQUESTED (erts_break_requested = 1) |