Home | History | Annotate | Download | only in freedreno
      1 /* -*- mode: C; c-file-style: "k&r"; tab-width 4; indent-tabs-mode: t; -*- */
      2 
      3 /*
      4  * Copyright (C) 2016 Rob Clark <robclark (at) 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 (at) freedesktop.org>
     27  */
     28 
     29 #ifdef HAVE_CONFIG_H
     30 # include <config.h>
     31 #endif
     32 
     33 #include "freedreno_drmif.h"
     34 #include "freedreno_priv.h"
     35 
     36 
     37 drm_private void bo_del(struct fd_bo *bo);
     38 drm_private extern pthread_mutex_t table_lock;
     39 
     40 static void
     41 add_bucket(struct fd_bo_cache *cache, int size)
     42 {
     43 	unsigned int i = cache->num_buckets;
     44 
     45 	assert(i < ARRAY_SIZE(cache->cache_bucket));
     46 
     47 	list_inithead(&cache->cache_bucket[i].list);
     48 	cache->cache_bucket[i].size = size;
     49 	cache->num_buckets++;
     50 }
     51 
     52 /**
     53  * @coarse: if true, only power-of-two bucket sizes, otherwise
     54  *    fill in for a bit smoother size curve..
     55  */
     56 drm_private void
     57 fd_bo_cache_init(struct fd_bo_cache *cache, int course)
     58 {
     59 	unsigned long size, cache_max_size = 64 * 1024 * 1024;
     60 
     61 	/* OK, so power of two buckets was too wasteful of memory.
     62 	 * Give 3 other sizes between each power of two, to hopefully
     63 	 * cover things accurately enough.  (The alternative is
     64 	 * probably to just go for exact matching of sizes, and assume
     65 	 * that for things like composited window resize the tiled
     66 	 * width/height alignment and rounding of sizes to pages will
     67 	 * get us useful cache hit rates anyway)
     68 	 */
     69 	add_bucket(cache, 4096);
     70 	add_bucket(cache, 4096 * 2);
     71 	if (!course)
     72 		add_bucket(cache, 4096 * 3);
     73 
     74 	/* Initialize the linked lists for BO reuse cache. */
     75 	for (size = 4 * 4096; size <= cache_max_size; size *= 2) {
     76 		add_bucket(cache, size);
     77 		if (!course) {
     78 			add_bucket(cache, size + size * 1 / 4);
     79 			add_bucket(cache, size + size * 2 / 4);
     80 			add_bucket(cache, size + size * 3 / 4);
     81 		}
     82 	}
     83 }
     84 
     85 /* Frees older cached buffers.  Called under table_lock */
     86 drm_private void
     87 fd_bo_cache_cleanup(struct fd_bo_cache *cache, time_t time)
     88 {
     89 	int i;
     90 
     91 	if (cache->time == time)
     92 		return;
     93 
     94 	for (i = 0; i < cache->num_buckets; i++) {
     95 		struct fd_bo_bucket *bucket = &cache->cache_bucket[i];
     96 		struct fd_bo *bo;
     97 
     98 		while (!LIST_IS_EMPTY(&bucket->list)) {
     99 			bo = LIST_ENTRY(struct fd_bo, bucket->list.next, list);
    100 
    101 			/* keep things in cache for at least 1 second: */
    102 			if (time && ((time - bo->free_time) <= 1))
    103 				break;
    104 
    105 			list_del(&bo->list);
    106 			bo_del(bo);
    107 		}
    108 	}
    109 
    110 	cache->time = time;
    111 }
    112 
    113 static struct fd_bo_bucket * get_bucket(struct fd_bo_cache *cache, uint32_t size)
    114 {
    115 	int i;
    116 
    117 	/* hmm, this is what intel does, but I suppose we could calculate our
    118 	 * way to the correct bucket size rather than looping..
    119 	 */
    120 	for (i = 0; i < cache->num_buckets; i++) {
    121 		struct fd_bo_bucket *bucket = &cache->cache_bucket[i];
    122 		if (bucket->size >= size) {
    123 			return bucket;
    124 		}
    125 	}
    126 
    127 	return NULL;
    128 }
    129 
    130 static int is_idle(struct fd_bo *bo)
    131 {
    132 	return fd_bo_cpu_prep(bo, NULL,
    133 			DRM_FREEDRENO_PREP_READ |
    134 			DRM_FREEDRENO_PREP_WRITE |
    135 			DRM_FREEDRENO_PREP_NOSYNC) == 0;
    136 }
    137 
    138 static struct fd_bo *find_in_bucket(struct fd_bo_bucket *bucket, uint32_t flags)
    139 {
    140 	struct fd_bo *bo = NULL;
    141 
    142 	/* TODO .. if we had an ALLOC_FOR_RENDER flag like intel, we could
    143 	 * skip the busy check.. if it is only going to be a render target
    144 	 * then we probably don't need to stall..
    145 	 *
    146 	 * NOTE that intel takes ALLOC_FOR_RENDER bo's from the list tail
    147 	 * (MRU, since likely to be in GPU cache), rather than head (LRU)..
    148 	 */
    149 	pthread_mutex_lock(&table_lock);
    150 	if (!LIST_IS_EMPTY(&bucket->list)) {
    151 		bo = LIST_ENTRY(struct fd_bo, bucket->list.next, list);
    152 		/* TODO check for compatible flags? */
    153 		if (is_idle(bo)) {
    154 			list_del(&bo->list);
    155 		} else {
    156 			bo = NULL;
    157 		}
    158 	}
    159 	pthread_mutex_unlock(&table_lock);
    160 
    161 	return bo;
    162 }
    163 
    164 /* NOTE: size is potentially rounded up to bucket size: */
    165 drm_private struct fd_bo *
    166 fd_bo_cache_alloc(struct fd_bo_cache *cache, uint32_t *size, uint32_t flags)
    167 {
    168 	struct fd_bo *bo = NULL;
    169 	struct fd_bo_bucket *bucket;
    170 
    171 	*size = ALIGN(*size, 4096);
    172 	bucket = get_bucket(cache, *size);
    173 
    174 	/* see if we can be green and recycle: */
    175 retry:
    176 	if (bucket) {
    177 		*size = bucket->size;
    178 		bo = find_in_bucket(bucket, flags);
    179 		if (bo) {
    180 			if (bo->funcs->madvise(bo, TRUE) <= 0) {
    181 				/* we've lost the backing pages, delete and try again: */
    182 				pthread_mutex_lock(&table_lock);
    183 				bo_del(bo);
    184 				pthread_mutex_unlock(&table_lock);
    185 				goto retry;
    186 			}
    187 			atomic_set(&bo->refcnt, 1);
    188 			fd_device_ref(bo->dev);
    189 			return bo;
    190 		}
    191 	}
    192 
    193 	return NULL;
    194 }
    195 
    196 drm_private int
    197 fd_bo_cache_free(struct fd_bo_cache *cache, struct fd_bo *bo)
    198 {
    199 	struct fd_bo_bucket *bucket = get_bucket(cache, bo->size);
    200 
    201 	/* see if we can be green and recycle: */
    202 	if (bucket) {
    203 		struct timespec time;
    204 
    205 		bo->funcs->madvise(bo, FALSE);
    206 
    207 		clock_gettime(CLOCK_MONOTONIC, &time);
    208 
    209 		bo->free_time = time.tv_sec;
    210 		list_addtail(&bo->list, &bucket->list);
    211 		fd_bo_cache_cleanup(cache, time.tv_sec);
    212 
    213 		/* bo's in the bucket cache don't have a ref and
    214 		 * don't hold a ref to the dev:
    215 		 */
    216 		fd_device_del_locked(bo->dev);
    217 
    218 		return 0;
    219 	}
    220 
    221 	return -1;
    222 }
    223