src/freedreno/drm/freedreno_bo_cache.c

   1 /*
   2  * Copyright (C) 2012-2018 Rob Clark <robclark@freedesktop.org>
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  21  * SOFTWARE.
  22  *
  23  * Authors:
  24  *    Rob Clark <robclark@freedesktop.org>
  25  */
  26
  27 #include "freedreno_drmif.h"
  28 #include "freedreno_priv.h"
  29
  30 void bo_del(struct fd_bo *bo);
  31 extern pthread_mutex_t table_lock;
  32
  33 static void
  34 add_bucket(struct fd_bo_cache *cache, int size)
  35 {
  36         unsigned int i = cache->num_buckets;
  37
  38         assert(i < ARRAY_SIZE(cache->cache_bucket));
  39
  40         list_inithead(&cache->cache_bucket[i].list);
  41         cache->cache_bucket[i].size = size;
  42         cache->num_buckets++;
  43 }
  44
  45 /**
  46  * @coarse: if true, only power-of-two bucket sizes, otherwise
  47  *    fill in for a bit smoother size curve..
  48  */
  49 void
  50 fd_bo_cache_init(struct fd_bo_cache *cache, int coarse)
  51 {
  52         unsigned long size, cache_max_size = 64 * 1024 * 1024;
  53
  54         /* OK, so power of two buckets was too wasteful of memory.
  55          * Give 3 other sizes between each power of two, to hopefully
  56          * cover things accurately enough.  (The alternative is
  57          * probably to just go for exact matching of sizes, and assume
  58          * that for things like composited window resize the tiled
  59          * width/height alignment and rounding of sizes to pages will
  60          * get us useful cache hit rates anyway)
  61          */
  62         add_bucket(cache, 4096);
  63         add_bucket(cache, 4096 * 2);
  64         if (!coarse)
  65                 add_bucket(cache, 4096 * 3);
  66
  67         /* Initialize the linked lists for BO reuse cache. */
  68         for (size = 4 * 4096; size <= cache_max_size; size *= 2) {
  69                 add_bucket(cache, size);
  70                 if (!coarse) {
  71                         add_bucket(cache, size + size * 1 / 4);
  72                         add_bucket(cache, size + size * 2 / 4);
  73                         add_bucket(cache, size + size * 3 / 4);
  74                 }
  75         }
  76 }
  77
  78 /* Frees older cached buffers.  Called under table_lock */
  79 void
  80 fd_bo_cache_cleanup(struct fd_bo_cache *cache, time_t time)
  81 {
  82         int i;
  83
  84         if (cache->time == time)
  85                 return;
  86
  87         for (i = 0; i < cache->num_buckets; i++) {
  88                 struct fd_bo_bucket *bucket = &cache->cache_bucket[i];
  89                 struct fd_bo *bo;
  90
  91                 while (!list_is_empty(&bucket->list)) {
  92                         bo = LIST_ENTRY(struct fd_bo, bucket->list.next, list);
  93
  94                         /* keep things in cache for at least 1 second: */
  95                         if (time && ((time - bo->free_time) <= 1))
  96                                 break;
  97
  98                         VG_BO_OBTAIN(bo);
  99                         list_del(&bo->list);
 100                         bo_del(bo);
 101                 }
 102         }
 103
 104         cache->time = time;
 105 }
 106
 107 static struct fd_bo_bucket * get_bucket(struct fd_bo_cache *cache, uint32_t size)
 108 {
 109         int i;
 110
 111         /* hmm, this is what intel does, but I suppose we could calculate our
 112          * way to the correct bucket size rather than looping..
 113          */
 114         for (i = 0; i < cache->num_buckets; i++) {
 115                 struct fd_bo_bucket *bucket = &cache->cache_bucket[i];
 116                 if (bucket->size >= size) {
 117                         return bucket;
 118                 }
 119         }
 120
 121         return NULL;
 122 }
 123
 124 static int is_idle(struct fd_bo *bo)
 125 {
 126         return fd_bo_cpu_prep(bo, NULL,
 127                         DRM_FREEDRENO_PREP_READ |
 128                         DRM_FREEDRENO_PREP_WRITE |
 129                         DRM_FREEDRENO_PREP_NOSYNC) == 0;
 130 }
 131
 132 static struct fd_bo *find_in_bucket(struct fd_bo_bucket *bucket, uint32_t flags)
 133 {
 134         struct fd_bo *bo = NULL;
 135
 136         /* TODO .. if we had an ALLOC_FOR_RENDER flag like intel, we could
 137          * skip the busy check.. if it is only going to be a render target
 138          * then we probably don't need to stall..
 139          *
 140          * NOTE that intel takes ALLOC_FOR_RENDER bo's from the list tail
 141          * (MRU, since likely to be in GPU cache), rather than head (LRU)..
 142          */
 143         pthread_mutex_lock(&table_lock);
 144         if (!list_is_empty(&bucket->list)) {
 145                 bo = LIST_ENTRY(struct fd_bo, bucket->list.next, list);
 146                 /* TODO check for compatible flags? */
 147                 if (is_idle(bo)) {
 148                         list_del(&bo->list);
 149                 } else {
 150                         bo = NULL;
 151                 }
 152         }
 153         pthread_mutex_unlock(&table_lock);
 154
 155         return bo;
 156 }
 157
 158 /* NOTE: size is potentially rounded up to bucket size: */
 159 struct fd_bo *
 160 fd_bo_cache_alloc(struct fd_bo_cache *cache, uint32_t *size, uint32_t flags)
 161 {
 162         struct fd_bo *bo = NULL;
 163         struct fd_bo_bucket *bucket;
 164
 165         *size = align(*size, 4096);
 166         bucket = get_bucket(cache, *size);
 167
 168         /* see if we can be green and recycle: */
 169 retry:
 170         if (bucket) {
 171                 *size = bucket->size;
 172                 bo = find_in_bucket(bucket, flags);
 173                 if (bo) {
 174                         VG_BO_OBTAIN(bo);
 175                         if (bo->funcs->madvise(bo, true) <= 0) {
 176                                 /* we've lost the backing pages, delete and try again: */
 177                                 pthread_mutex_lock(&table_lock);
 178                                 bo_del(bo);
 179                                 pthread_mutex_unlock(&table_lock);
 180                                 goto retry;
 181                         }
 182                         p_atomic_set(&bo->refcnt, 1);
 183                         fd_device_ref(bo->dev);
 184                         return bo;
 185                 }
 186         }
 187
 188         return NULL;
 189 }
 190
 191 int
 192 fd_bo_cache_free(struct fd_bo_cache *cache, struct fd_bo *bo)
 193 {
 194         struct fd_bo_bucket *bucket = get_bucket(cache, bo->size);
 195
 196         /* see if we can be green and recycle: */
 197         if (bucket) {
 198                 struct timespec time;
 199
 200                 bo->funcs->madvise(bo, false);
 201
 202                 clock_gettime(CLOCK_MONOTONIC, &time);
 203
 204                 bo->free_time = time.tv_sec;
 205                 VG_BO_RELEASE(bo);
 206                 list_addtail(&bo->list, &bucket->list);
 207                 fd_bo_cache_cleanup(cache, time.tv_sec);
 208
 209                 /* bo's in the bucket cache don't have a ref and
 210                  * don't hold a ref to the dev:
 211                  */
 212                 fd_device_del_locked(bo->dev);
 213
 214                 return 0;
 215         }
 216
 217         return -1;
 218 }