Files
jemalloc/include/jemalloc/internal/sc.h
guangli-dai c067a55c79 Introducing a new usize calculation policy
Converting size to usize is what jemalloc has been done by ceiling
size to the closest size class. However, this causes lots of memory
wastes with HPA enabled.  This commit changes how usize is calculated so
that the gap between two contiguous usize is no larger than a page.
Specifically, this commit includes the following changes:

1. Adding a build-time config option (--enable-limit-usize-gap) and a
runtime one (limit_usize_gap) to guard the changes.
When build-time
config is enabled, some minor CPU overhead is expected because usize
will be stored and accessed apart from index.  When runtime option is
also enabled (it can only be enabled with the build-time config
enabled). a new usize calculation approach wil be employed.  This new
calculation will ceil size to the closest multiple of PAGE for all sizes
larger than USIZE_GROW_SLOW_THRESHOLD instead of using the size classes.
Note when the build-time config is enabled, the runtime option is
default on.

2. Prepare tcache for size to grow by PAGE over GROUP*PAGE.
To prepare for the upcoming changes where size class grows by PAGE when
larger than NGROUP * PAGE, disable the tcache when it is larger than 2 *
NGROUP * PAGE. The threshold for tcache is set higher to prevent perf
regression as much as possible while usizes between NGROUP * PAGE and 2 *
NGROUP * PAGE happen to grow by PAGE.

3. Prepare pac and hpa psset for size to grow by PAGE over GROUP*PAGE
For PAC, to avoid having too many bins, arena bins still have the same
layout.  This means some extra search is needed for a page-level request that
is not aligned with the orginal size class: it should also search the heap
before the current index since the previous heap might also be able to
have some allocations satisfying it.  The same changes apply to HPA's
psset.
This search relies on the enumeration of the heap because not all allocs in
the previous heap are guaranteed to satisfy the request.  To balance the
memory and CPU overhead, we currently enumerate at most a fixed number
of nodes before concluding none can satisfy the request during an
enumeration.

4. Add bytes counter to arena large stats.
To prepare for the upcoming usize changes, stats collected by
multiplying alive allocations and the bin size is no longer accurate.
Thus, add separate counters to record the bytes malloced and dalloced.

5. Change structs use when freeing to avoid using index2size for large sizes.
  - Change the definition of emap_alloc_ctx_t
  - Change the read of both from edata_t.
  - Change the assignment and usage of emap_alloc_ctx_t.
  - Change other callsites of index2size.
Note for the changes in the data structure, i.e., emap_alloc_ctx_t,
will be used when the build-time config (--enable-limit-usize-gap) is
enabled but they will store the same value as index2size(szind) if the
runtime option (opt_limit_usize_gap) is not enabled.

6. Adapt hpa to the usize changes.
Change the settings in sec to limit is usage for sizes larger than
USIZE_GROW_SLOW_THRESHOLD and modify corresponding tests.

7. Modify usize calculation and corresponding tests.
Change the sz_s2u_compute. Note sz_index2size is not always safe now
while sz_size2index still works as expected.
2025-03-06 15:08:13 -08:00

377 lines
15 KiB
C

#ifndef JEMALLOC_INTERNAL_SC_H
#define JEMALLOC_INTERNAL_SC_H
#include "jemalloc/internal/jemalloc_preamble.h"
#include "jemalloc/internal/jemalloc_internal_types.h"
/*
* Size class computations:
*
* These are a little tricky; we'll first start by describing how things
* generally work, and then describe some of the details.
*
* Ignore the first few size classes for a moment. We can then split all the
* remaining size classes into groups. The size classes in a group are spaced
* such that they cover allocation request sizes in a power-of-2 range. The
* power of two is called the base of the group, and the size classes in it
* satisfy allocations in the half-open range (base, base * 2]. There are
* SC_NGROUP size classes in each group, equally spaced in the range, so that
* each one covers allocations for base / SC_NGROUP possible allocation sizes.
* We call that value (base / SC_NGROUP) the delta of the group. Each size class
* is delta larger than the one before it (including the initial size class in a
* group, which is delta larger than base, the largest size class in the
* previous group).
* To make the math all work out nicely, we require that SC_NGROUP is a power of
* two, and define it in terms of SC_LG_NGROUP. We'll often talk in terms of
* lg_base and lg_delta. For each of these groups then, we have that
* lg_delta == lg_base - SC_LG_NGROUP.
* The size classes in a group with a given lg_base and lg_delta (which, recall,
* can be computed from lg_base for these groups) are therefore:
* base + 1 * delta
* which covers allocations in (base, base + 1 * delta]
* base + 2 * delta
* which covers allocations in (base + 1 * delta, base + 2 * delta].
* base + 3 * delta
* which covers allocations in (base + 2 * delta, base + 3 * delta].
* ...
* base + SC_NGROUP * delta ( == 2 * base)
* which covers allocations in (base + (SC_NGROUP - 1) * delta, 2 * base].
* (Note that currently SC_NGROUP is always 4, so the "..." is empty in
* practice.)
* Note that the last size class in the group is the next power of two (after
* base), so that we've set up the induction correctly for the next group's
* selection of delta.
*
* Now, let's start considering the first few size classes. Two extra constants
* come into play here: LG_QUANTUM and SC_LG_TINY_MIN. LG_QUANTUM ensures
* correct platform alignment; all objects of size (1 << LG_QUANTUM) or larger
* are at least (1 << LG_QUANTUM) aligned; this can be used to ensure that we
* never return improperly aligned memory, by making (1 << LG_QUANTUM) equal the
* highest required alignment of a platform. For allocation sizes smaller than
* (1 << LG_QUANTUM) though, we can be more relaxed (since we don't support
* platforms with types with alignment larger than their size). To allow such
* allocations (without wasting space unnecessarily), we introduce tiny size
* classes; one per power of two, up until we hit the quantum size. There are
* therefore LG_QUANTUM - SC_LG_TINY_MIN such size classes.
*
* Next, we have a size class of size (1 << LG_QUANTUM). This can't be the
* start of a group in the sense we described above (covering a power of two
* range) since, if we divided into it to pick a value of delta, we'd get a
* delta smaller than (1 << LG_QUANTUM) for sizes >= (1 << LG_QUANTUM), which
* is against the rules.
*
* The first base we can divide by SC_NGROUP while still being at least
* (1 << LG_QUANTUM) is SC_NGROUP * (1 << LG_QUANTUM). We can get there by
* having SC_NGROUP size classes, spaced (1 << LG_QUANTUM) apart. These size
* classes are:
* 1 * (1 << LG_QUANTUM)
* 2 * (1 << LG_QUANTUM)
* 3 * (1 << LG_QUANTUM)
* ... (although, as above, this "..." is empty in practice)
* SC_NGROUP * (1 << LG_QUANTUM).
*
* There are SC_NGROUP of these size classes, so we can regard it as a sort of
* pseudo-group, even though it spans multiple powers of 2, is divided
* differently, and both starts and ends on a power of 2 (as opposed to just
* ending). SC_NGROUP is itself a power of two, so the first group after the
* pseudo-group has the power-of-two base SC_NGROUP * (1 << LG_QUANTUM), for a
* lg_base of LG_QUANTUM + SC_LG_NGROUP. We can divide this base into SC_NGROUP
* sizes without violating our LG_QUANTUM requirements, so we can safely set
* lg_delta = lg_base - SC_LG_GROUP (== LG_QUANTUM).
*
* So, in order, the size classes are:
*
* Tiny size classes:
* - Count: LG_QUANTUM - SC_LG_TINY_MIN.
* - Sizes:
* 1 << SC_LG_TINY_MIN
* 1 << (SC_LG_TINY_MIN + 1)
* 1 << (SC_LG_TINY_MIN + 2)
* ...
* 1 << (LG_QUANTUM - 1)
*
* Initial pseudo-group:
* - Count: SC_NGROUP
* - Sizes:
* 1 * (1 << LG_QUANTUM)
* 2 * (1 << LG_QUANTUM)
* 3 * (1 << LG_QUANTUM)
* ...
* SC_NGROUP * (1 << LG_QUANTUM)
*
* Regular group 0:
* - Count: SC_NGROUP
* - Sizes:
* (relative to lg_base of LG_QUANTUM + SC_LG_NGROUP and lg_delta of
* lg_base - SC_LG_NGROUP)
* (1 << lg_base) + 1 * (1 << lg_delta)
* (1 << lg_base) + 2 * (1 << lg_delta)
* (1 << lg_base) + 3 * (1 << lg_delta)
* ...
* (1 << lg_base) + SC_NGROUP * (1 << lg_delta) [ == (1 << (lg_base + 1)) ]
*
* Regular group 1:
* - Count: SC_NGROUP
* - Sizes:
* (relative to lg_base of LG_QUANTUM + SC_LG_NGROUP + 1 and lg_delta of
* lg_base - SC_LG_NGROUP)
* (1 << lg_base) + 1 * (1 << lg_delta)
* (1 << lg_base) + 2 * (1 << lg_delta)
* (1 << lg_base) + 3 * (1 << lg_delta)
* ...
* (1 << lg_base) + SC_NGROUP * (1 << lg_delta) [ == (1 << (lg_base + 1)) ]
*
* ...
*
* Regular group N:
* - Count: SC_NGROUP
* - Sizes:
* (relative to lg_base of LG_QUANTUM + SC_LG_NGROUP + N and lg_delta of
* lg_base - SC_LG_NGROUP)
* (1 << lg_base) + 1 * (1 << lg_delta)
* (1 << lg_base) + 2 * (1 << lg_delta)
* (1 << lg_base) + 3 * (1 << lg_delta)
* ...
* (1 << lg_base) + SC_NGROUP * (1 << lg_delta) [ == (1 << (lg_base + 1)) ]
*
*
* Representation of metadata:
* To make the math easy, we'll mostly work in lg quantities. We record lg_base,
* lg_delta, and ndelta (i.e. number of deltas above the base) on a
* per-size-class basis, and maintain the invariant that, across all size
* classes, size == (1 << lg_base) + ndelta * (1 << lg_delta).
*
* For regular groups (i.e. those with lg_base >= LG_QUANTUM + SC_LG_NGROUP),
* lg_delta is lg_base - SC_LG_NGROUP, and ndelta goes from 1 to SC_NGROUP.
*
* For the initial tiny size classes (if any), lg_base is lg(size class size).
* lg_delta is lg_base for the first size class, and lg_base - 1 for all
* subsequent ones. ndelta is always 0.
*
* For the pseudo-group, if there are no tiny size classes, then we set
* lg_base == LG_QUANTUM, lg_delta == LG_QUANTUM, and have ndelta range from 0
* to SC_NGROUP - 1. (Note that delta == base, so base + (SC_NGROUP - 1) * delta
* is just SC_NGROUP * base, or (1 << (SC_LG_NGROUP + LG_QUANTUM)), so we do
* indeed get a power of two that way). If there *are* tiny size classes, then
* the first size class needs to have lg_delta relative to the largest tiny size
* class. We therefore set lg_base == LG_QUANTUM - 1,
* lg_delta == LG_QUANTUM - 1, and ndelta == 1, keeping the rest of the
* pseudo-group the same.
*
*
* Other terminology:
* "Small" size classes mean those that are allocated out of bins, which is the
* same as those that are slab allocated.
* "Large" size classes are those that are not small. The cutoff for counting as
* large is page size * group size.
*/
/*
* Size class N + (1 << SC_LG_NGROUP) twice the size of size class N.
*/
#define SC_LG_NGROUP 2
#define SC_LG_TINY_MIN 3
#if SC_LG_TINY_MIN == 0
/* The div module doesn't support division by 1, which this would require. */
#error "Unsupported LG_TINY_MIN"
#endif
/*
* The definitions below are all determined by the above settings and system
* characteristics.
*/
#define SC_NGROUP (1ULL << SC_LG_NGROUP)
#define SC_PTR_BITS ((1ULL << LG_SIZEOF_PTR) * 8)
#define SC_NTINY (LG_QUANTUM - SC_LG_TINY_MIN)
#define SC_LG_TINY_MAXCLASS (LG_QUANTUM > SC_LG_TINY_MIN ? LG_QUANTUM - 1 : -1)
#define SC_NPSEUDO SC_NGROUP
#define SC_LG_FIRST_REGULAR_BASE (LG_QUANTUM + SC_LG_NGROUP)
/*
* We cap allocations to be less than 2 ** (ptr_bits - 1), so the highest base
* we need is 2 ** (ptr_bits - 2). (This also means that the last group is 1
* size class shorter than the others).
* We could probably save some space in arenas by capping this at LG_VADDR size.
*/
#define SC_LG_BASE_MAX (SC_PTR_BITS - 2)
#define SC_NREGULAR (SC_NGROUP * \
(SC_LG_BASE_MAX - SC_LG_FIRST_REGULAR_BASE + 1) - 1)
#define SC_NSIZES (SC_NTINY + SC_NPSEUDO + SC_NREGULAR)
/*
* The number of size classes that are a multiple of the page size.
*
* Here are the first few bases that have a page-sized SC.
*
* lg(base) | base | highest SC | page-multiple SCs
* --------------|------------------------------------------
* LG_PAGE - 1 | PAGE / 2 | PAGE | 1
* LG_PAGE | PAGE | 2 * PAGE | 1
* LG_PAGE + 1 | 2 * PAGE | 4 * PAGE | 2
* LG_PAGE + 2 | 4 * PAGE | 8 * PAGE | 4
*
* The number of page-multiple SCs continues to grow in powers of two, up until
* lg_delta == lg_page, which corresponds to setting lg_base to lg_page +
* SC_LG_NGROUP. So, then, the number of size classes that are multiples of the
* page size whose lg_delta is less than the page size are
* is 1 + (2**0 + 2**1 + ... + 2**(lg_ngroup - 1) == 2**lg_ngroup.
*
* For each base with lg_base in [lg_page + lg_ngroup, lg_base_max), there are
* NGROUP page-sized size classes, and when lg_base == lg_base_max, there are
* NGROUP - 1.
*
* This gives us the quantity we seek.
*/
#define SC_NPSIZES ( \
SC_NGROUP \
+ (SC_LG_BASE_MAX - (LG_PAGE + SC_LG_NGROUP)) * SC_NGROUP \
+ SC_NGROUP - 1)
/*
* We declare a size class is binnable if size < page size * group. Or, in other
* words, lg(size) < lg(page size) + lg(group size).
*/
#define SC_NBINS ( \
/* Sub-regular size classes. */ \
SC_NTINY + SC_NPSEUDO \
/* Groups with lg_regular_min_base <= lg_base <= lg_base_max */ \
+ SC_NGROUP * (LG_PAGE + SC_LG_NGROUP - SC_LG_FIRST_REGULAR_BASE) \
/* Last SC of the last group hits the bound exactly; exclude it. */ \
- 1)
/*
* The size2index_tab lookup table uses uint8_t to encode each bin index, so we
* cannot support more than 256 small size classes.
*/
#if (SC_NBINS > 256)
# error "Too many small size classes"
#endif
/* The largest size class in the lookup table, and its binary log. */
#define SC_LG_MAX_LOOKUP 12
#define SC_LOOKUP_MAXCLASS (1 << SC_LG_MAX_LOOKUP)
/* Internal, only used for the definition of SC_SMALL_MAXCLASS. */
#define SC_SMALL_MAX_BASE (1 << (LG_PAGE + SC_LG_NGROUP - 1))
#define SC_SMALL_MAX_DELTA (1 << (LG_PAGE - 1))
/* The largest size class allocated out of a slab. */
#define SC_SMALL_MAXCLASS (SC_SMALL_MAX_BASE \
+ (SC_NGROUP - 1) * SC_SMALL_MAX_DELTA)
/* The fastpath assumes all lookup-able sizes are small. */
#if (SC_SMALL_MAXCLASS < SC_LOOKUP_MAXCLASS)
# error "Lookup table sizes must be small"
#endif
/* The smallest size class not allocated out of a slab. */
#define SC_LARGE_MINCLASS ((size_t)1ULL << (LG_PAGE + SC_LG_NGROUP))
#define SC_LG_LARGE_MINCLASS (LG_PAGE + SC_LG_NGROUP)
/* Internal; only used for the definition of SC_LARGE_MAXCLASS. */
#define SC_MAX_BASE ((size_t)1 << (SC_PTR_BITS - 2))
#define SC_MAX_DELTA ((size_t)1 << (SC_PTR_BITS - 2 - SC_LG_NGROUP))
/* The largest size class supported. */
#define SC_LARGE_MAXCLASS (SC_MAX_BASE + (SC_NGROUP - 1) * SC_MAX_DELTA)
/* Maximum number of regions in one slab. */
#ifndef CONFIG_LG_SLAB_MAXREGS
# define SC_LG_SLAB_MAXREGS (LG_PAGE - SC_LG_TINY_MIN)
#else
# if CONFIG_LG_SLAB_MAXREGS < (LG_PAGE - SC_LG_TINY_MIN)
# error "Unsupported SC_LG_SLAB_MAXREGS"
# else
# define SC_LG_SLAB_MAXREGS CONFIG_LG_SLAB_MAXREGS
# endif
#endif
/*
* When config_limit_usize_gap is enabled, the gaps between two contiguous
* size classes should not exceed PAGE. This means there should be no concept
* of size classes for sizes > SC_SMALL_MAXCLASS (or >= SC_LARGE_MINCLASS).
* However, between SC_LARGE_MINCLASS (SC_NGROUP * PAGE) and
* 2 * SC_NGROUP * PAGE, the size class also happens to be aligned with PAGE.
* Since tcache relies on size classes to work and it greatly increases the
* perf of allocs & deallocs, we extend the existence of size class to
* 2 * SC_NGROUP * PAGE ONLY for the tcache module. This means for all other
* modules, there is no size class for sizes >= SC_LARGE_MINCLASS. Yet for
* tcache, the threshold is moved up to 2 * SC_NGROUP * PAGE, which is
* USIZE_GROW_SLOW_THRESHOLD defined below. With the default SC_NGROUP being
* 2, and PAGE being 4KB, the threshold for tcache (USIZE_GROW_SLOW_THRESHOLD)
* is 32KB.
*/
#define LG_USIZE_GROW_SLOW_THRESHOLD (SC_LG_NGROUP + LG_PAGE + 1)
#define USIZE_GROW_SLOW_THRESHOLD (1U << LG_USIZE_GROW_SLOW_THRESHOLD)
#define SC_SLAB_MAXREGS (1U << SC_LG_SLAB_MAXREGS)
typedef struct sc_s sc_t;
struct sc_s {
/* Size class index, or -1 if not a valid size class. */
int index;
/* Lg group base size (no deltas added). */
int lg_base;
/* Lg delta to previous size class. */
int lg_delta;
/* Delta multiplier. size == 1<<lg_base + ndelta<<lg_delta */
int ndelta;
/*
* True if the size class is a multiple of the page size, false
* otherwise.
*/
bool psz;
/*
* True if the size class is a small, bin, size class. False otherwise.
*/
bool bin;
/* The slab page count if a small bin size class, 0 otherwise. */
int pgs;
/* Same as lg_delta if a lookup table size class, 0 otherwise. */
int lg_delta_lookup;
};
typedef struct sc_data_s sc_data_t;
struct sc_data_s {
/* Number of tiny size classes. */
unsigned ntiny;
/* Number of bins supported by the lookup table. */
int nlbins;
/* Number of small size class bins. */
int nbins;
/* Number of size classes. */
int nsizes;
/* Number of bits required to store NSIZES. */
int lg_ceil_nsizes;
/* Number of size classes that are a multiple of (1U << LG_PAGE). */
unsigned npsizes;
/* Lg of maximum tiny size class (or -1, if none). */
int lg_tiny_maxclass;
/* Maximum size class included in lookup table. */
size_t lookup_maxclass;
/* Maximum small size class. */
size_t small_maxclass;
/* Lg of minimum large size class. */
int lg_large_minclass;
/* The minimum large size class. */
size_t large_minclass;
/* Maximum (large) size class. */
size_t large_maxclass;
/* True if the sc_data_t has been initialized (for debugging only). */
bool initialized;
sc_t sc[SC_NSIZES];
};
size_t reg_size_compute(int lg_base, int lg_delta, int ndelta);
void sc_data_init(sc_data_t *data);
/*
* Updates slab sizes in [begin, end] to be pgs pages in length, if possible.
* Otherwise, does its best to accommodate the request.
*/
void sc_data_update_slab_size(sc_data_t *data, size_t begin, size_t end,
int pgs);
void sc_boot(sc_data_t *data);
#endif /* JEMALLOC_INTERNAL_SC_H */