From b95268d46cbbdc2790d9a643c4e19b4de028567f Mon Sep 17 00:00:00 2001 From: Daan Date: Tue, 10 Jun 2025 16:11:30 -0700 Subject: [PATCH 1/3] only adjust commit stats on success --- src/os.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/src/os.c b/src/os.c index ad847ea7..5a1d9ecd 100644 --- a/src/os.c +++ b/src/os.c @@ -431,7 +431,6 @@ static void* mi_os_page_align_area_conservative(void* addr, size_t size, size_t* bool _mi_os_commit_ex(void* addr, size_t size, bool* is_zero, size_t stat_size) { if (is_zero != NULL) { *is_zero = false; } - mi_os_stat_increase(committed, stat_size); // use size for precise commit vs. decommit mi_os_stat_counter_increase(commit_calls, 1); // page align range @@ -455,6 +454,8 @@ bool _mi_os_commit_ex(void* addr, size_t size, bool* is_zero, size_t stat_size) if (os_is_zero) { mi_track_mem_defined(start,csize); } else { mi_track_mem_undefined(start,csize); } #endif + + mi_os_stat_increase(committed, stat_size); // use size for precise commit vs. decommit return true; } From 088e32e4efc4a4045d589121956ccd8c5aefd615 Mon Sep 17 00:00:00 2001 From: Daan Date: Tue, 10 Jun 2025 16:11:56 -0700 Subject: [PATCH 2/3] add mi_bitmap_popcountN --- src/bitmap.c | 51 +++++++++++++++++++++++++++++++++++++++++++++++++++ src/bitmap.h | 3 +++ 2 files changed, 54 insertions(+) diff --git a/src/bitmap.c b/src/bitmap.c index 2640c6ff..c4cb2fd7 100644 --- a/src/bitmap.c +++ b/src/bitmap.c @@ -256,6 +256,11 @@ static inline bool mi_bfield_atomic_is_xset_mask(mi_xset_t set, _Atomic(mi_bfiel else return mi_bfield_atomic_is_clear_mask(b, mask); } +// Count bits in a mask +static inline size_t mi_bfield_atomic_popcount_mask(_Atomic(mi_bfield_t)*b, mi_bfield_t mask) { + const mi_bfield_t x = mi_atomic_load_relaxed(b); + return mi_bfield_popcount(x & mask); +} /* -------------------------------------------------------------------------------- @@ -366,6 +371,38 @@ static inline bool mi_bchunk_clearN(mi_bchunk_t* chunk, size_t cidx, size_t n, b return mi_bchunk_xsetN_(MI_BIT_CLEAR, chunk, cidx, n, NULL, maybe_all_clear); } +// Check if a sequence of `n` bits within a chunk are all set/cleared. +// This can cross bfield's +mi_decl_noinline static size_t mi_bchunk_popcountN_(mi_bchunk_t* chunk, size_t field_idx, size_t idx, size_t n) { + mi_assert_internal((field_idx*MI_BFIELD_BITS) + idx + n <= MI_BCHUNK_BITS); + size_t count = 0; + while (n > 0) { + size_t m = MI_BFIELD_BITS - idx; // m is the bits to xset in this field + if (m > n) { m = n; } + mi_assert_internal(idx + m <= MI_BFIELD_BITS); + mi_assert_internal(field_idx < MI_BCHUNK_FIELDS); + const size_t mask = mi_bfield_mask(m, idx); + count += mi_bfield_atomic_popcount_mask(&chunk->bfields[field_idx], mask); + // next field + field_idx++; + idx = 0; + n -= m; + } + return count; +} + +// Count set bits a sequence of `n` bits. +static inline size_t mi_bchunk_popcountN(mi_bchunk_t* chunk, size_t cidx, size_t n) { + mi_assert_internal(cidx + n <= MI_BCHUNK_BITS); + mi_assert_internal(n>0); + if (n==0) return 0; + const size_t i = cidx / MI_BFIELD_BITS; + const size_t idx = cidx % MI_BFIELD_BITS; + if (n==1) { return (mi_bfield_atomic_is_set(&chunk->bfields[i], idx) ? 1 : 0); } + if (idx + n <= MI_BFIELD_BITS) { return mi_bfield_atomic_popcount_mask(&chunk->bfields[i], mi_bfield_mask(n, idx)); } + return mi_bchunk_popcountN_(chunk, i, idx, n); +} + // ------- mi_bchunk_is_xset --------------------------------------- @@ -1100,6 +1137,20 @@ bool mi_bitmap_clearN(mi_bitmap_t* bitmap, size_t idx, size_t n) { return were_allset; } +// Count bits set in a range of `n` bits. +// `n` cannot cross chunk boundaries (and `n <= MI_BCHUNK_BITS`)! +size_t mi_bitmap_popcountN( mi_bitmap_t* bitmap, size_t idx, size_t n) { + mi_assert_internal(n>0); + mi_assert_internal(n<=MI_BCHUNK_BITS); + + const size_t chunk_idx = idx / MI_BCHUNK_BITS; + const size_t cidx = idx % MI_BCHUNK_BITS; + mi_assert_internal(cidx + n <= MI_BCHUNK_BITS); // don't cross chunks (for now) + mi_assert_internal(chunk_idx < mi_bitmap_chunk_count(bitmap)); + if (cidx + n > MI_BCHUNK_BITS) { n = MI_BCHUNK_BITS - cidx; } // paranoia + return mi_bchunk_popcountN(&bitmap->chunks[chunk_idx], cidx, n); +} + // Set/clear a bit in the bitmap; returns `true` if atomically transitioned from 0 to 1 (or 1 to 0) bool mi_bitmap_set(mi_bitmap_t* bitmap, size_t idx) { diff --git a/src/bitmap.h b/src/bitmap.h index 5dec8a89..bad1bd74 100644 --- a/src/bitmap.h +++ b/src/bitmap.h @@ -206,6 +206,9 @@ bool _mi_bitmap_forall_set(mi_bitmap_t* bitmap, mi_forall_set_fun_t* visit, mi_a bool _mi_bitmap_forall_setc_ranges(mi_bitmap_t* bitmap, mi_forall_set_fun_t* visit, mi_arena_t* arena, void* arg); +// Count all set bits in given range in the bitmap. (cannot cross chunks) +size_t mi_bitmap_popcountN( mi_bitmap_t* bitmap, size_t idx, size_t n); + /* ---------------------------------------------------------------------------- Binned concurrent bitmap Assigns a size class to each chunk such that small blocks don't cause too From f5ce881fdf0a1087df00fb6fd810c16322221d0f Mon Sep 17 00:00:00 2001 From: Daan Date: Tue, 10 Jun 2025 16:13:39 -0700 Subject: [PATCH 3/3] fix state of commit bits on commit failure --- src/arena.c | 35 ++++++++++++++++++----------------- 1 file changed, 18 insertions(+), 17 deletions(-) diff --git a/src/arena.c b/src/arena.c index d0e60db7..b34f3170 100644 --- a/src/arena.c +++ b/src/arena.c @@ -200,29 +200,26 @@ static mi_decl_noinline void* mi_arena_try_alloc_at( } // set commit state - if (commit) { - memid->initially_committed = true; - + if (commit) { // commit requested, but the range may not be committed as a whole: ensure it is committed now - if (!mi_bitmap_is_setN(arena->slices_committed, slice_index, slice_count)) { - // not fully committed: commit the full range and set the commit bits - // we set the bits first since we own these slices (they are no longer free) - size_t already_committed_count = 0; - mi_bitmap_setN(arena->slices_committed, slice_index, slice_count, &already_committed_count); - // adjust the stats so we don't double count the commits - //if (already_committed_count > 0) { - // mi_subproc_stat_adjust_decrease(arena->subproc, committed, mi_size_of_slices(already_committed_count), true /* on alloc */); - //} - // now actually commit + const size_t already_committed = mi_bitmap_popcountN(arena->slices_committed, slice_index, slice_count); + if (already_committed < slice_count) { + // not all committed, try to commit now bool commit_zero = false; - if (!_mi_os_commit_ex(p, mi_size_of_slices(slice_count), &commit_zero, mi_size_of_slices(slice_count - already_committed_count))) { - // if the commit fails, roll back and return NULL - _mi_arenas_free(p, mi_size_of_slices(slice_count), *memid); // this will decommit as well (if partially committed) + if (!_mi_os_commit_ex(p, mi_size_of_slices(slice_count), &commit_zero, mi_size_of_slices(slice_count - already_committed))) { + // if the commit fails, release ownership, and return NULL; + // note: this does not roll back dirty bits but that is ok. + mi_bbitmap_setN(arena->slices_free, slice_index, slice_count); return NULL; } + if (commit_zero) { + memid->initially_zero = true; + } + + // set the commit bits + mi_bitmap_setN(arena->slices_committed, slice_index, slice_count, NULL); // committed - if (commit_zero) { memid->initially_zero = true; } #if MI_DEBUG > 1 if (memid->initially_zero) { if (!mi_mem_is_zero(p, mi_size_of_slices(slice_count))) { @@ -240,6 +237,10 @@ static mi_decl_noinline void* mi_arena_try_alloc_at( mi_subproc_stat_increase( arena->subproc, committed, mi_size_of_slices(touched_slices)); } } + + mi_assert_internal(mi_bitmap_is_setN(arena->slices_committed, slice_index, slice_count)); + memid->initially_committed = true; + // tool support if (memid->initially_zero) { mi_track_mem_defined(p, slice_count * MI_ARENA_SLICE_SIZE);