Revert "Remove BITMAP_USE_TREE."

Some systems use a native 64 KiB page size, which means that the bitmap
for the smallest size class can be 8192 bits, not just 512 bits as when
the page size is 4 KiB.  Linear search in bitmap_{sfu,ffu}() is
unacceptably slow for such large bitmaps.

This reverts commit 7c00f04ff4.
This commit is contained in:
Jason Evans 2017-04-16 09:25:56 -07:00
parent 38e847c1c5
commit 45f087eb03
5 changed files with 307 additions and 0 deletions

View File

@ -16,6 +16,12 @@ void bitmap_unset(bitmap_t *bitmap, const bitmap_info_t *binfo, size_t bit);
#if (defined(JEMALLOC_ENABLE_INLINE) || defined(JEMALLOC_BITMAP_C_))
JEMALLOC_INLINE bool
bitmap_full(bitmap_t *bitmap, const bitmap_info_t *binfo) {
#ifdef BITMAP_USE_TREE
size_t rgoff = binfo->levels[binfo->nlevels].group_offset - 1;
bitmap_t rg = bitmap[rgoff];
/* The bitmap is full iff the root group is 0. */
return (rg == 0);
#else
size_t i;
for (i = 0; i < binfo->ngroups; i++) {
@ -24,6 +30,7 @@ bitmap_full(bitmap_t *bitmap, const bitmap_info_t *binfo) {
}
}
return true;
#endif
}
JEMALLOC_INLINE bool
@ -52,6 +59,24 @@ bitmap_set(bitmap_t *bitmap, const bitmap_info_t *binfo, size_t bit) {
g ^= ZU(1) << (bit & BITMAP_GROUP_NBITS_MASK);
*gp = g;
assert(bitmap_get(bitmap, binfo, bit));
#ifdef BITMAP_USE_TREE
/* Propagate group state transitions up the tree. */
if (g == 0) {
unsigned i;
for (i = 1; i < binfo->nlevels; i++) {
bit = goff;
goff = bit >> LG_BITMAP_GROUP_NBITS;
gp = &bitmap[binfo->levels[i].group_offset + goff];
g = *gp;
assert(g & (ZU(1) << (bit & BITMAP_GROUP_NBITS_MASK)));
g ^= ZU(1) << (bit & BITMAP_GROUP_NBITS_MASK);
*gp = g;
if (g != 0) {
break;
}
}
}
#endif
}
/* ffu: find first unset >= bit. */
@ -59,6 +84,44 @@ JEMALLOC_INLINE size_t
bitmap_ffu(const bitmap_t *bitmap, const bitmap_info_t *binfo, size_t min_bit) {
assert(min_bit < binfo->nbits);
#ifdef BITMAP_USE_TREE
size_t bit = 0;
for (unsigned level = binfo->nlevels; level--;) {
size_t lg_bits_per_group = (LG_BITMAP_GROUP_NBITS * (level +
1));
bitmap_t group = bitmap[binfo->levels[level].group_offset + (bit
>> lg_bits_per_group)];
unsigned group_nmask = ((min_bit > bit) ? (min_bit - bit) : 0)
>> (lg_bits_per_group - LG_BITMAP_GROUP_NBITS);
assert(group_nmask <= BITMAP_GROUP_NBITS);
bitmap_t group_mask = ~((1LU << group_nmask) - 1);
bitmap_t group_masked = group & group_mask;
if (group_masked == 0LU) {
if (group == 0LU) {
return binfo->nbits;
}
/*
* min_bit was preceded by one or more unset bits in
* this group, but there are no other unset bits in this
* group. Try again starting at the first bit of the
* next sibling. This will recurse at most once per
* non-root level.
*/
size_t sib_base = bit + (1U << lg_bits_per_group);
assert(sib_base > min_bit);
assert(sib_base > bit);
if (sib_base >= binfo->nbits) {
return binfo->nbits;
}
return bitmap_ffu(bitmap, binfo, sib_base);
}
bit += (ffs_lu(group_masked) - 1) << (lg_bits_per_group -
LG_BITMAP_GROUP_NBITS);
}
assert(bit >= min_bit);
assert(bit < binfo->nbits);
return bit;
#else
size_t i = min_bit >> LG_BITMAP_GROUP_NBITS;
bitmap_t g = bitmap[i] & ~((1LU << (min_bit & BITMAP_GROUP_NBITS_MASK))
- 1);
@ -72,6 +135,7 @@ bitmap_ffu(const bitmap_t *bitmap, const bitmap_info_t *binfo, size_t min_bit) {
g = bitmap[i];
} while (i < binfo->ngroups);
return binfo->nbits;
#endif
}
/* sfu: set first unset. */
@ -83,6 +147,16 @@ bitmap_sfu(bitmap_t *bitmap, const bitmap_info_t *binfo) {
assert(!bitmap_full(bitmap, binfo));
#ifdef BITMAP_USE_TREE
i = binfo->nlevels - 1;
g = bitmap[binfo->levels[i].group_offset];
bit = ffs_lu(g) - 1;
while (i > 0) {
i--;
g = bitmap[binfo->levels[i].group_offset + bit];
bit = (bit << LG_BITMAP_GROUP_NBITS) + (ffs_lu(g) - 1);
}
#else
i = 0;
g = bitmap[0];
while ((bit = ffs_lu(g)) == 0) {
@ -90,6 +164,7 @@ bitmap_sfu(bitmap_t *bitmap, const bitmap_info_t *binfo) {
g = bitmap[i];
}
bit = (i << LG_BITMAP_GROUP_NBITS) + (bit - 1);
#endif
bitmap_set(bitmap, binfo, bit);
return bit;
}
@ -111,6 +186,26 @@ bitmap_unset(bitmap_t *bitmap, const bitmap_info_t *binfo, size_t bit) {
g ^= ZU(1) << (bit & BITMAP_GROUP_NBITS_MASK);
*gp = g;
assert(!bitmap_get(bitmap, binfo, bit));
#ifdef BITMAP_USE_TREE
/* Propagate group state transitions up the tree. */
if (propagate) {
unsigned i;
for (i = 1; i < binfo->nlevels; i++) {
bit = goff;
goff = bit >> LG_BITMAP_GROUP_NBITS;
gp = &bitmap[binfo->levels[i].group_offset + goff];
g = *gp;
propagate = (g == 0);
assert((g & (ZU(1) << (bit & BITMAP_GROUP_NBITS_MASK)))
== 0);
g ^= ZU(1) << (bit & BITMAP_GROUP_NBITS_MASK);
*gp = g;
if (!propagate) {
break;
}
}
}
#endif /* BITMAP_USE_TREE */
}
#endif

View File

@ -10,8 +10,19 @@ struct bitmap_info_s {
/* Logical number of bits in bitmap (stored at bottom level). */
size_t nbits;
#ifdef BITMAP_USE_TREE
/* Number of levels necessary for nbits. */
unsigned nlevels;
/*
* Only the first (nlevels+1) elements are used, and levels are ordered
* bottom to top (e.g. the bottom level is stored in levels[0]).
*/
bitmap_level_t levels[BITMAP_MAX_LEVELS+1];
#else /* BITMAP_USE_TREE */
/* Number of groups necessary for nbits. */
size_t ngroups;
#endif /* BITMAP_USE_TREE */
};
#endif /* JEMALLOC_INTERNAL_BITMAP_STRUCTS_H */

View File

@ -21,10 +21,115 @@ typedef unsigned long bitmap_t;
#define BITMAP_GROUP_NBITS (1U << LG_BITMAP_GROUP_NBITS)
#define BITMAP_GROUP_NBITS_MASK (BITMAP_GROUP_NBITS-1)
/*
* Do some analysis on how big the bitmap is before we use a tree. For a brute
* force linear search, if we would have to call ffs_lu() more than 2^3 times,
* use a tree instead.
*/
#if LG_BITMAP_MAXBITS - LG_BITMAP_GROUP_NBITS > 3
# define BITMAP_USE_TREE
#endif
/* Number of groups required to store a given number of bits. */
#define BITMAP_BITS2GROUPS(nbits) \
(((nbits) + BITMAP_GROUP_NBITS_MASK) >> LG_BITMAP_GROUP_NBITS)
/*
* Number of groups required at a particular level for a given number of bits.
*/
#define BITMAP_GROUPS_L0(nbits) \
BITMAP_BITS2GROUPS(nbits)
#define BITMAP_GROUPS_L1(nbits) \
BITMAP_BITS2GROUPS(BITMAP_BITS2GROUPS(nbits))
#define BITMAP_GROUPS_L2(nbits) \
BITMAP_BITS2GROUPS(BITMAP_BITS2GROUPS(BITMAP_BITS2GROUPS((nbits))))
#define BITMAP_GROUPS_L3(nbits) \
BITMAP_BITS2GROUPS(BITMAP_BITS2GROUPS(BITMAP_BITS2GROUPS( \
BITMAP_BITS2GROUPS((nbits)))))
#define BITMAP_GROUPS_L4(nbits) \
BITMAP_BITS2GROUPS(BITMAP_BITS2GROUPS(BITMAP_BITS2GROUPS( \
BITMAP_BITS2GROUPS(BITMAP_BITS2GROUPS((nbits))))))
/*
* Assuming the number of levels, number of groups required for a given number
* of bits.
*/
#define BITMAP_GROUPS_1_LEVEL(nbits) \
BITMAP_GROUPS_L0(nbits)
#define BITMAP_GROUPS_2_LEVEL(nbits) \
(BITMAP_GROUPS_1_LEVEL(nbits) + BITMAP_GROUPS_L1(nbits))
#define BITMAP_GROUPS_3_LEVEL(nbits) \
(BITMAP_GROUPS_2_LEVEL(nbits) + BITMAP_GROUPS_L2(nbits))
#define BITMAP_GROUPS_4_LEVEL(nbits) \
(BITMAP_GROUPS_3_LEVEL(nbits) + BITMAP_GROUPS_L3(nbits))
#define BITMAP_GROUPS_5_LEVEL(nbits) \
(BITMAP_GROUPS_4_LEVEL(nbits) + BITMAP_GROUPS_L4(nbits))
/*
* Maximum number of groups required to support LG_BITMAP_MAXBITS.
*/
#ifdef BITMAP_USE_TREE
#if LG_BITMAP_MAXBITS <= LG_BITMAP_GROUP_NBITS
# define BITMAP_GROUPS(nbits) BITMAP_GROUPS_1_LEVEL(nbits)
# define BITMAP_GROUPS_MAX BITMAP_GROUPS_1_LEVEL(BITMAP_MAXBITS)
#elif LG_BITMAP_MAXBITS <= LG_BITMAP_GROUP_NBITS * 2
# define BITMAP_GROUPS(nbits) BITMAP_GROUPS_2_LEVEL(nbits)
# define BITMAP_GROUPS_MAX BITMAP_GROUPS_2_LEVEL(BITMAP_MAXBITS)
#elif LG_BITMAP_MAXBITS <= LG_BITMAP_GROUP_NBITS * 3
# define BITMAP_GROUPS(nbits) BITMAP_GROUPS_3_LEVEL(nbits)
# define BITMAP_GROUPS_MAX BITMAP_GROUPS_3_LEVEL(BITMAP_MAXBITS)
#elif LG_BITMAP_MAXBITS <= LG_BITMAP_GROUP_NBITS * 4
# define BITMAP_GROUPS(nbits) BITMAP_GROUPS_4_LEVEL(nbits)
# define BITMAP_GROUPS_MAX BITMAP_GROUPS_4_LEVEL(BITMAP_MAXBITS)
#elif LG_BITMAP_MAXBITS <= LG_BITMAP_GROUP_NBITS * 5
# define BITMAP_GROUPS(nbits) BITMAP_GROUPS_5_LEVEL(nbits)
# define BITMAP_GROUPS_MAX BITMAP_GROUPS_5_LEVEL(BITMAP_MAXBITS)
#else
# error "Unsupported bitmap size"
#endif
/*
* Maximum number of levels possible. This could be statically computed based
* on LG_BITMAP_MAXBITS:
*
* #define BITMAP_MAX_LEVELS \
* (LG_BITMAP_MAXBITS / LG_SIZEOF_BITMAP) \
* + !!(LG_BITMAP_MAXBITS % LG_SIZEOF_BITMAP)
*
* However, that would not allow the generic BITMAP_INFO_INITIALIZER() macro, so
* instead hardcode BITMAP_MAX_LEVELS to the largest number supported by the
* various cascading macros. The only additional cost this incurs is some
* unused trailing entries in bitmap_info_t structures; the bitmaps themselves
* are not impacted.
*/
#define BITMAP_MAX_LEVELS 5
#define BITMAP_INFO_INITIALIZER(nbits) { \
/* nbits. */ \
nbits, \
/* nlevels. */ \
(BITMAP_GROUPS_L0(nbits) > BITMAP_GROUPS_L1(nbits)) + \
(BITMAP_GROUPS_L1(nbits) > BITMAP_GROUPS_L2(nbits)) + \
(BITMAP_GROUPS_L2(nbits) > BITMAP_GROUPS_L3(nbits)) + \
(BITMAP_GROUPS_L3(nbits) > BITMAP_GROUPS_L4(nbits)) + 1, \
/* levels. */ \
{ \
{0}, \
{BITMAP_GROUPS_L0(nbits)}, \
{BITMAP_GROUPS_L1(nbits) + BITMAP_GROUPS_L0(nbits)}, \
{BITMAP_GROUPS_L2(nbits) + BITMAP_GROUPS_L1(nbits) + \
BITMAP_GROUPS_L0(nbits)}, \
{BITMAP_GROUPS_L3(nbits) + BITMAP_GROUPS_L2(nbits) + \
BITMAP_GROUPS_L1(nbits) + BITMAP_GROUPS_L0(nbits)}, \
{BITMAP_GROUPS_L4(nbits) + BITMAP_GROUPS_L3(nbits) + \
BITMAP_GROUPS_L2(nbits) + BITMAP_GROUPS_L1(nbits) \
+ BITMAP_GROUPS_L0(nbits)} \
} \
}
#else /* BITMAP_USE_TREE */
#define BITMAP_GROUPS(nbits) BITMAP_BITS2GROUPS(nbits)
#define BITMAP_GROUPS_MAX BITMAP_BITS2GROUPS(BITMAP_MAXBITS)
@ -35,4 +140,6 @@ typedef unsigned long bitmap_t;
BITMAP_BITS2GROUPS(nbits) \
}
#endif /* BITMAP_USE_TREE */
#endif /* JEMALLOC_INTERNAL_BITMAP_TYPES_H */

View File

@ -6,6 +6,82 @@
/******************************************************************************/
#ifdef BITMAP_USE_TREE
void
bitmap_info_init(bitmap_info_t *binfo, size_t nbits) {
unsigned i;
size_t group_count;
assert(nbits > 0);
assert(nbits <= (ZU(1) << LG_BITMAP_MAXBITS));
/*
* Compute the number of groups necessary to store nbits bits, and
* progressively work upward through the levels until reaching a level
* that requires only one group.
*/
binfo->levels[0].group_offset = 0;
group_count = BITMAP_BITS2GROUPS(nbits);
for (i = 1; group_count > 1; i++) {
assert(i < BITMAP_MAX_LEVELS);
binfo->levels[i].group_offset = binfo->levels[i-1].group_offset
+ group_count;
group_count = BITMAP_BITS2GROUPS(group_count);
}
binfo->levels[i].group_offset = binfo->levels[i-1].group_offset
+ group_count;
assert(binfo->levels[i].group_offset <= BITMAP_GROUPS_MAX);
binfo->nlevels = i;
binfo->nbits = nbits;
}
static size_t
bitmap_info_ngroups(const bitmap_info_t *binfo) {
return binfo->levels[binfo->nlevels].group_offset;
}
void
bitmap_init(bitmap_t *bitmap, const bitmap_info_t *binfo, bool fill) {
size_t extra;
unsigned i;
/*
* Bits are actually inverted with regard to the external bitmap
* interface.
*/
if (fill) {
/* The "filled" bitmap starts out with all 0 bits. */
memset(bitmap, 0, bitmap_size(binfo));
return;
}
/*
* The "empty" bitmap starts out with all 1 bits, except for trailing
* unused bits (if any). Note that each group uses bit 0 to correspond
* to the first logical bit in the group, so extra bits are the most
* significant bits of the last group.
*/
memset(bitmap, 0xffU, bitmap_size(binfo));
extra = (BITMAP_GROUP_NBITS - (binfo->nbits & BITMAP_GROUP_NBITS_MASK))
& BITMAP_GROUP_NBITS_MASK;
if (extra != 0) {
bitmap[binfo->levels[1].group_offset - 1] >>= extra;
}
for (i = 1; i < binfo->nlevels; i++) {
size_t group_count = binfo->levels[i].group_offset -
binfo->levels[i-1].group_offset;
extra = (BITMAP_GROUP_NBITS - (group_count &
BITMAP_GROUP_NBITS_MASK)) & BITMAP_GROUP_NBITS_MASK;
if (extra != 0) {
bitmap[binfo->levels[i+1].group_offset - 1] >>= extra;
}
}
}
#else /* BITMAP_USE_TREE */
void
bitmap_info_init(bitmap_info_t *binfo, size_t nbits) {
assert(nbits > 0);
@ -37,6 +113,8 @@ bitmap_init(bitmap_t *bitmap, const bitmap_info_t *binfo, bool fill) {
}
}
#endif /* BITMAP_USE_TREE */
size_t
bitmap_size(const bitmap_info_t *binfo) {
return (bitmap_info_ngroups(binfo) << LG_SIZEOF_BITMAP);

View File

@ -103,8 +103,24 @@ test_bitmap_initializer_body(const bitmap_info_t *binfo, size_t nbits) {
assert_zu_eq(binfo->nbits, binfo_dyn.nbits,
"Unexpected difference between static and dynamic initialization, "
"nbits=%zu", nbits);
#ifdef BITMAP_USE_TREE
assert_u_eq(binfo->nlevels, binfo_dyn.nlevels,
"Unexpected difference between static and dynamic initialization, "
"nbits=%zu", nbits);
{
unsigned i;
for (i = 0; i < binfo->nlevels; i++) {
assert_zu_eq(binfo->levels[i].group_offset,
binfo_dyn.levels[i].group_offset,
"Unexpected difference between static and dynamic "
"initialization, nbits=%zu, level=%u", nbits, i);
}
}
#else
assert_zu_eq(binfo->ngroups, binfo_dyn.ngroups,
"Unexpected difference between static and dynamic initialization");
#endif
}
TEST_BEGIN(test_bitmap_initializer) {