mimalloc/src/segment-cache.c

/* ----------------------------------------------------------------------------
Copyright (c) 2020, Microsoft Research, Daan Leijen
This is free software; you can redistribute it and/or modify it under the
terms of the MIT license. A copy of the license can be found in the file
"LICENSE" at the root of this distribution.
-----------------------------------------------------------------------------*/

/* ----------------------------------------------------------------------------
  Implements a cache of segments to avoid expensive OS calls and to reuse
  the commit_mask to optimize the commit/decommit calls.
  The full memory map of all segments is also implemented here.
-----------------------------------------------------------------------------*/
#include "mimalloc.h"
#include "mimalloc/internal.h"
#include "mimalloc/atomic.h"

#include "./bitmap.h"  // atomic bitmap

// #define MI_CACHE_DISABLE 1    // define to completely disable the segment cache

#define MI_CACHE_FIELDS     (16)
#define MI_CACHE_MAX        (MI_BITMAP_FIELD_BITS*MI_CACHE_FIELDS)       // 1024 on 64-bit

#define BITS_SET()          MI_ATOMIC_VAR_INIT(UINTPTR_MAX)
#define MI_CACHE_BITS_SET   MI_INIT16(BITS_SET)                          // note: update if MI_CACHE_FIELDS changes

typedef struct mi_cache_slot_s {
  void*               p;
  mi_memid_t          memid;
  bool                is_pinned;
  mi_commit_mask_t    commit_mask;
  mi_commit_mask_t    purge_mask;
  _Atomic(mi_msecs_t) expire;
} mi_cache_slot_t;

static mi_decl_cache_align mi_cache_slot_t cache[MI_CACHE_MAX];    // = 0

static mi_decl_cache_align mi_bitmap_field_t cache_unavailable[MI_CACHE_FIELDS] = { MI_CACHE_BITS_SET };        // zero bit = available!
static mi_decl_cache_align mi_bitmap_field_t cache_unavailable_large[MI_CACHE_FIELDS] = { MI_CACHE_BITS_SET };
static mi_decl_cache_align mi_bitmap_field_t cache_inuse[MI_CACHE_FIELDS];   // zero bit = free

static bool mi_cdecl mi_segment_cache_is_suitable(mi_bitmap_index_t bitidx, void* arg) {
  mi_arena_id_t req_arena_id = *((mi_arena_id_t*)arg);
  mi_cache_slot_t* slot = &cache[mi_bitmap_index_bit(bitidx)];
  return _mi_arena_memid_is_suitable(slot->memid, req_arena_id);
}

mi_decl_noinline static void* mi_segment_cache_pop_ex(
                              bool all_suitable,
                              size_t size, mi_commit_mask_t* commit_mask, 
                              mi_commit_mask_t* purge_mask, bool large_allowed,
                              bool* large, bool* is_pinned, bool* is_zero, 
                              mi_arena_id_t _req_arena_id, mi_memid_t* memid, mi_os_tld_t* tld)
{
#ifdef MI_CACHE_DISABLE
  return NULL;
#else

  // only segment blocks
  if (size != MI_SEGMENT_SIZE) return NULL;

  // numa node determines start field
  const int numa_node = _mi_os_numa_node(tld);
  size_t start_field = 0;
  if (numa_node > 0) {
    start_field = (MI_CACHE_FIELDS / _mi_os_numa_node_count())*numa_node;
    if (start_field >= MI_CACHE_FIELDS) start_field = 0;
  }

  // find an available slot and make it unavailable
  mi_bitmap_index_t bitidx = 0;
  bool claimed = false;
  mi_arena_id_t req_arena_id = _req_arena_id;
  mi_bitmap_pred_fun_t pred_fun = (all_suitable ? NULL : &mi_segment_cache_is_suitable);  // cannot pass NULL as the arena may be exclusive itself; todo: do not put exclusive arenas in the cache?

  if (large_allowed) {  // large allowed?
    claimed = _mi_bitmap_try_find_from_claim_pred(cache_unavailable_large, MI_CACHE_FIELDS, start_field, 1, pred_fun, &req_arena_id, &bitidx);
    if (claimed) *large = true;
  }
  if (!claimed) {
    claimed = _mi_bitmap_try_find_from_claim_pred (cache_unavailable, MI_CACHE_FIELDS, start_field, 1, pred_fun, &req_arena_id, &bitidx);
    if (claimed) *large = false;
  }

  if (!claimed) return NULL;

  // no longer available but still in-use
  mi_assert_internal(_mi_bitmap_is_claimed(cache_unavailable, MI_CACHE_FIELDS, 1, bitidx));
  mi_assert_internal(_mi_bitmap_is_claimed(cache_unavailable_large, MI_CACHE_FIELDS, 1, bitidx));
  mi_assert_internal(_mi_bitmap_is_claimed(cache_inuse, MI_CACHE_FIELDS, 1, bitidx));

  // found a slot
  mi_cache_slot_t* slot = &cache[mi_bitmap_index_bit(bitidx)];
  void* p = slot->p;
  *memid = slot->memid;
  *is_pinned = slot->is_pinned;
  *is_zero = false;
  *commit_mask = slot->commit_mask;     
  *purge_mask = slot->purge_mask;
  slot->p = NULL;
  mi_atomic_storei64_release(&slot->expire,(mi_msecs_t)0);
  
  // mark the slot as free again
  _mi_bitmap_unclaim(cache_inuse, MI_CACHE_FIELDS, 1, bitidx);
  return p;
#endif
}


mi_decl_noinline void* _mi_segment_cache_pop(size_t size, mi_commit_mask_t* commit_mask, mi_commit_mask_t* purge_mask, bool large_allowed, bool* large, bool* is_pinned, bool* is_zero, mi_arena_id_t _req_arena_id, mi_memid_t* memid, mi_os_tld_t* tld)
{
  return mi_segment_cache_pop_ex(false, size, commit_mask, purge_mask, large_allowed, large, is_pinned, is_zero, _req_arena_id, memid, tld);
}

static mi_decl_noinline void mi_commit_mask_decommit(mi_commit_mask_t* cmask, void* p, size_t total, mi_stats_t* stats)
{
  if (mi_commit_mask_is_empty(cmask)) {
    // nothing
  }
  else if (mi_commit_mask_is_full(cmask)) {
    // decommit the whole in one call
    _mi_os_decommit(p, total, stats);
  }
  else {
    // decommit parts
    mi_assert_internal((total%MI_COMMIT_MASK_BITS)==0);
    size_t part = total/MI_COMMIT_MASK_BITS;
    size_t idx;
    size_t count;    
    mi_commit_mask_foreach(cmask, idx, count) {
      void*  start = (uint8_t*)p + (idx*part);
      size_t size = count*part;
      _mi_os_decommit(start, size, stats);
    }
    mi_commit_mask_foreach_end()
  }
  mi_commit_mask_create_empty(cmask);
}

#define MI_MAX_PURGE_PER_PUSH  (4)

static mi_decl_noinline void mi_segment_cache_purge(bool visit_all, bool force, mi_os_tld_t* tld)
{
  MI_UNUSED(tld);
  if (!mi_option_is_enabled(mi_option_allow_purge)) return;
  mi_msecs_t now = _mi_clock_now();
  size_t purged = 0;
  const size_t max_visits = (visit_all ? MI_CACHE_MAX /* visit all */ : MI_CACHE_FIELDS /* probe at most N (=16) slots */);
  size_t idx              = (visit_all ? 0 : _mi_random_shuffle((uintptr_t)now) % MI_CACHE_MAX /* random start */ );
  for (size_t visited = 0; visited < max_visits; visited++,idx++) {  // visit N slots
    if (idx >= MI_CACHE_MAX) idx = 0; // wrap
    mi_cache_slot_t* slot = &cache[idx];
    mi_msecs_t expire = mi_atomic_loadi64_relaxed(&slot->expire);
    if (expire != 0 && (force || now >= expire)) {  // racy read
      // seems expired, first claim it from available
      purged++;
      mi_bitmap_index_t bitidx = mi_bitmap_index_create_from_bit(idx);      
      if (_mi_bitmap_claim(cache_unavailable, MI_CACHE_FIELDS, 1, bitidx, NULL)) {  // no need to check large as those cannot be decommitted anyways
        // it was available, we claimed it (and made it unavailable)
        mi_assert_internal(_mi_bitmap_is_claimed(cache_unavailable, MI_CACHE_FIELDS, 1, bitidx));
        mi_assert_internal(_mi_bitmap_is_claimed(cache_unavailable_large, MI_CACHE_FIELDS, 1, bitidx));
        // we can now access it safely
        expire = mi_atomic_loadi64_acquire(&slot->expire);
        if (expire != 0 && (force || now >= expire)) {  // safe read
          mi_assert_internal(_mi_bitmap_is_claimed(cache_inuse, MI_CACHE_FIELDS, 1, bitidx));
          // still expired, decommit it
          mi_atomic_storei64_relaxed(&slot->expire,(mi_msecs_t)0);
          mi_assert_internal(!mi_commit_mask_is_empty(&slot->commit_mask));
          _mi_abandoned_await_readers();  // wait until safe to decommit
          // decommit committed parts
          // TODO: instead of decommit, we could also free to the OS?
          mi_commit_mask_decommit(&slot->commit_mask, slot->p, MI_SEGMENT_SIZE, tld->stats);
          mi_commit_mask_create_empty(&slot->purge_mask);
        }
        _mi_bitmap_unclaim(cache_unavailable, MI_CACHE_FIELDS, 1, bitidx); // make it available again for a pop
      }
      if (!visit_all && purged > MI_MAX_PURGE_PER_PUSH) break;  // bound to no more than N purge tries per push
    }
  }
}

void _mi_segment_cache_collect(bool force, mi_os_tld_t* tld) {
  if (force) {
    // called on `mi_collect(true)` but not on thread termination    
    _mi_segment_cache_free_all(tld);
  }
  else {
    mi_segment_cache_purge(true /* visit all */, false /* don't force unexpired */, tld);
  }
}

void _mi_segment_cache_free_all(mi_os_tld_t* tld) {
  mi_commit_mask_t commit_mask;
  mi_commit_mask_t purge_mask;
  bool is_pinned;
  bool is_zero;
  bool is_large;
  mi_memid_t memid;
  const size_t size = MI_SEGMENT_SIZE;
  void* p;
  do {
    // keep popping and freeing the memory
    p = mi_segment_cache_pop_ex(true /* all */, size, &commit_mask, &purge_mask,
                                true /* allow large */, &is_large, &is_pinned, &is_zero, _mi_arena_id_none(), &memid, tld);
    if (p != NULL) {
      size_t csize = _mi_commit_mask_committed_size(&commit_mask, size);
      if (csize > 0 && !is_pinned) { _mi_stat_decrease(&_mi_stats_main.committed, csize); }
      _mi_arena_free(p, size, MI_SEGMENT_ALIGN, 0, memid, is_pinned /* pretend not committed to not double count decommits */, tld->stats);
    }
  } while (p != NULL);  
}

mi_decl_noinline bool _mi_segment_cache_push(void* start, size_t size, mi_memid_t memid, const mi_commit_mask_t* commit_mask, const mi_commit_mask_t* purge_mask, bool is_large, bool is_pinned, mi_os_tld_t* tld)
{
#ifdef MI_CACHE_DISABLE
  return false;
#else

  // purge expired entries
  mi_segment_cache_purge(false /* limit purges to a constant N */, false /* don't force unexpired */, tld);

  // only cache normal segment blocks
  if (size != MI_SEGMENT_SIZE || ((uintptr_t)start % MI_SEGMENT_ALIGN) != 0) return false;

  // Also do not cache arena allocated segments that cannot be decommitted. (as arena allocation is fast)
  // This is a common case with reserved huge OS pages.
  // 
  // (note: we could also allow segments that are already fully decommitted but that never happens
  //  as the first slice is always committed (for the segment metadata))
  if (!_mi_arena_memid_is_os_allocated(memid) && is_pinned) return false;

  // numa node determines start field
  int numa_node = _mi_os_numa_node(NULL);
  size_t start_field = 0;
  if (numa_node > 0) {
    start_field = (MI_CACHE_FIELDS / _mi_os_numa_node_count()) * numa_node;
    if (start_field >= MI_CACHE_FIELDS) start_field = 0;
  }

  // find an available slot
  mi_bitmap_index_t bitidx;
  bool claimed = _mi_bitmap_try_find_from_claim(cache_inuse, MI_CACHE_FIELDS, start_field, 1, &bitidx);
  if (!claimed) return false;

  mi_assert_internal(_mi_bitmap_is_claimed(cache_unavailable, MI_CACHE_FIELDS, 1, bitidx));
  mi_assert_internal(_mi_bitmap_is_claimed(cache_unavailable_large, MI_CACHE_FIELDS, 1, bitidx));
#if MI_DEBUG>1
  if (is_pinned || is_large) {
    mi_assert_internal(mi_commit_mask_is_full(commit_mask));
  }
#endif

  // set the slot
  mi_cache_slot_t* slot = &cache[mi_bitmap_index_bit(bitidx)];
  slot->p = start;
  slot->memid = memid;
  slot->is_pinned = is_pinned;
  mi_atomic_storei64_relaxed(&slot->expire,(mi_msecs_t)0);
  slot->commit_mask = *commit_mask;
  slot->purge_mask = *purge_mask;
  if (!mi_commit_mask_is_empty(commit_mask) && !is_large && !is_pinned && mi_option_is_enabled(mi_option_allow_purge)) {
    long delay = mi_option_get(mi_option_purge_delay) * mi_option_get(mi_option_arena_purge_mult);
    if (delay == 0) {
      _mi_abandoned_await_readers(); // wait until safe to decommit
      mi_commit_mask_decommit(&slot->commit_mask, start, MI_SEGMENT_SIZE, tld->stats);
      mi_commit_mask_create_empty(&slot->purge_mask);
    }
    else {
      mi_atomic_storei64_release(&slot->expire, _mi_clock_now() + delay);
    }
  }

  // make it available
  _mi_bitmap_unclaim((is_large ? cache_unavailable_large : cache_unavailable), MI_CACHE_FIELDS, 1, bitidx);
  return true;
#endif
}