1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */ 2 3 /* 4 * Copyright (C) 2016 Rob Clark <robclark@freedesktop.org> 5 * 6 * Permission is hereby granted, free of charge, to any person obtaining a 7 * copy of this software and associated documentation files (the "Software"), 8 * to deal in the Software without restriction, including without limitation 9 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 10 * and/or sell copies of the Software, and to permit persons to whom the 11 * Software is furnished to do so, subject to the following conditions: 12 * 13 * The above copyright notice and this permission notice (including the next 14 * paragraph) shall be included in all copies or substantial portions of the 15 * Software. 16 * 17 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 18 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 19 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 20 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 21 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 22 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 23 * SOFTWARE. 24 * 25 * Authors: 26 * Rob Clark <robclark@freedesktop.org> 27 */ 28 29 #include "freedreno_drmif.h" 30 #include "freedreno_priv.h" 31 32 drm_private void bo_del(struct fd_bo *bo); 33 drm_private extern pthread_mutex_t table_lock; 34 35 static void 36 add_bucket(struct fd_bo_cache *cache, int size) 37 { 38 unsigned int i = cache->num_buckets; 39 40 assert(i < ARRAY_SIZE(cache->cache_bucket)); 41 42 list_inithead(&cache->cache_bucket[i].list); 43 cache->cache_bucket[i].size = size; 44 cache->num_buckets++; 45 } 46 47 /** 48 * @coarse: if true, only power-of-two bucket sizes, otherwise 49 * fill in for a bit smoother size curve.. 50 */ 51 drm_private void 52 fd_bo_cache_init(struct fd_bo_cache *cache, int coarse) 53 { 54 unsigned long size, cache_max_size = 64 * 1024 * 1024; 55 56 /* OK, so power of two buckets was too wasteful of memory. 57 * Give 3 other sizes between each power of two, to hopefully 58 * cover things accurately enough. (The alternative is 59 * probably to just go for exact matching of sizes, and assume 60 * that for things like composited window resize the tiled 61 * width/height alignment and rounding of sizes to pages will 62 * get us useful cache hit rates anyway) 63 */ 64 add_bucket(cache, 4096); 65 add_bucket(cache, 4096 * 2); 66 if (!coarse) 67 add_bucket(cache, 4096 * 3); 68 69 /* Initialize the linked lists for BO reuse cache. */ 70 for (size = 4 * 4096; size <= cache_max_size; size *= 2) { 71 add_bucket(cache, size); 72 if (!coarse) { 73 add_bucket(cache, size + size * 1 / 4); 74 add_bucket(cache, size + size * 2 / 4); 75 add_bucket(cache, size + size * 3 / 4); 76 } 77 } 78 } 79 80 /* Frees older cached buffers. Called under table_lock */ 81 drm_private void 82 fd_bo_cache_cleanup(struct fd_bo_cache *cache, time_t time) 83 { 84 int i; 85 86 if (cache->time == time) 87 return; 88 89 for (i = 0; i < cache->num_buckets; i++) { 90 struct fd_bo_bucket *bucket = &cache->cache_bucket[i]; 91 struct fd_bo *bo; 92 93 while (!LIST_IS_EMPTY(&bucket->list)) { 94 bo = LIST_ENTRY(struct fd_bo, bucket->list.next, list); 95 96 /* keep things in cache for at least 1 second: */ 97 if (time && ((time - bo->free_time) <= 1)) 98 break; 99 100 VG_BO_OBTAIN(bo); 101 list_del(&bo->list); 102 bo_del(bo); 103 } 104 } 105 106 cache->time = time; 107 } 108 109 static struct fd_bo_bucket * get_bucket(struct fd_bo_cache *cache, uint32_t size) 110 { 111 int i; 112 113 /* hmm, this is what intel does, but I suppose we could calculate our 114 * way to the correct bucket size rather than looping.. 115 */ 116 for (i = 0; i < cache->num_buckets; i++) { 117 struct fd_bo_bucket *bucket = &cache->cache_bucket[i]; 118 if (bucket->size >= size) { 119 return bucket; 120 } 121 } 122 123 return NULL; 124 } 125 126 static int is_idle(struct fd_bo *bo) 127 { 128 return fd_bo_cpu_prep(bo, NULL, 129 DRM_FREEDRENO_PREP_READ | 130 DRM_FREEDRENO_PREP_WRITE | 131 DRM_FREEDRENO_PREP_NOSYNC) == 0; 132 } 133 134 static struct fd_bo *find_in_bucket(struct fd_bo_bucket *bucket, uint32_t flags) 135 { 136 struct fd_bo *bo = NULL; 137 138 /* TODO .. if we had an ALLOC_FOR_RENDER flag like intel, we could 139 * skip the busy check.. if it is only going to be a render target 140 * then we probably don't need to stall.. 141 * 142 * NOTE that intel takes ALLOC_FOR_RENDER bo's from the list tail 143 * (MRU, since likely to be in GPU cache), rather than head (LRU).. 144 */ 145 pthread_mutex_lock(&table_lock); 146 if (!LIST_IS_EMPTY(&bucket->list)) { 147 bo = LIST_ENTRY(struct fd_bo, bucket->list.next, list); 148 /* TODO check for compatible flags? */ 149 if (is_idle(bo)) { 150 list_del(&bo->list); 151 } else { 152 bo = NULL; 153 } 154 } 155 pthread_mutex_unlock(&table_lock); 156 157 return bo; 158 } 159 160 /* NOTE: size is potentially rounded up to bucket size: */ 161 drm_private struct fd_bo * 162 fd_bo_cache_alloc(struct fd_bo_cache *cache, uint32_t *size, uint32_t flags) 163 { 164 struct fd_bo *bo = NULL; 165 struct fd_bo_bucket *bucket; 166 167 *size = ALIGN(*size, 4096); 168 bucket = get_bucket(cache, *size); 169 170 /* see if we can be green and recycle: */ 171 retry: 172 if (bucket) { 173 *size = bucket->size; 174 bo = find_in_bucket(bucket, flags); 175 if (bo) { 176 VG_BO_OBTAIN(bo); 177 if (bo->funcs->madvise(bo, TRUE) <= 0) { 178 /* we've lost the backing pages, delete and try again: */ 179 pthread_mutex_lock(&table_lock); 180 bo_del(bo); 181 pthread_mutex_unlock(&table_lock); 182 goto retry; 183 } 184 atomic_set(&bo->refcnt, 1); 185 fd_device_ref(bo->dev); 186 return bo; 187 } 188 } 189 190 return NULL; 191 } 192 193 drm_private int 194 fd_bo_cache_free(struct fd_bo_cache *cache, struct fd_bo *bo) 195 { 196 struct fd_bo_bucket *bucket = get_bucket(cache, bo->size); 197 198 /* see if we can be green and recycle: */ 199 if (bucket) { 200 struct timespec time; 201 202 bo->funcs->madvise(bo, FALSE); 203 204 clock_gettime(CLOCK_MONOTONIC, &time); 205 206 bo->free_time = time.tv_sec; 207 VG_BO_RELEASE(bo); 208 list_addtail(&bo->list, &bucket->list); 209 fd_bo_cache_cleanup(cache, time.tv_sec); 210 211 /* bo's in the bucket cache don't have a ref and 212 * don't hold a ref to the dev: 213 */ 214 fd_device_del_locked(bo->dev); 215 216 return 0; 217 } 218 219 return -1; 220 } 221