1/* 2 * Copyright 2015 Advanced Micro Devices, Inc. 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 shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 * 22 */ 23 24#include <linux/kthread.h> 25#include <linux/module.h> 26#include <linux/sched.h> 27#include <linux/slab.h> 28#include <linux/wait.h> 29 30#include <drm/gpu_scheduler.h> 31 32#include <sys/pool.h> 33 34static struct pool sched_fence_slab; 35 36int __init drm_sched_fence_slab_init(void) 37{ 38#ifdef __linux__ 39 sched_fence_slab = kmem_cache_create( 40 "drm_sched_fence", sizeof(struct drm_sched_fence), 0, 41 SLAB_HWCACHE_ALIGN, NULL); 42 if (!sched_fence_slab) 43 return -ENOMEM; 44#else 45 pool_init(&sched_fence_slab, sizeof(struct drm_sched_fence), 46 CACHELINESIZE, IPL_TTY, 0, "drm_sched_fence", NULL); 47#endif 48 49 return 0; 50} 51 52void __exit drm_sched_fence_slab_fini(void) 53{ 54 rcu_barrier(); 55#ifdef __linux__ 56 kmem_cache_destroy(sched_fence_slab); 57#else 58 pool_destroy(&sched_fence_slab); 59#endif 60} 61 62static void drm_sched_fence_set_parent(struct drm_sched_fence *s_fence, 63 struct dma_fence *fence) 64{ 65 /* 66 * smp_store_release() to ensure another thread racing us 67 * in drm_sched_fence_set_deadline_finished() sees the 68 * fence's parent set before test_bit() 69 */ 70 smp_store_release(&s_fence->parent, dma_fence_get(fence)); 71 if (test_bit(DRM_SCHED_FENCE_FLAG_HAS_DEADLINE_BIT, 72 &s_fence->finished.flags)) 73 dma_fence_set_deadline(fence, s_fence->deadline); 74} 75 76void drm_sched_fence_scheduled(struct drm_sched_fence *fence, 77 struct dma_fence *parent) 78{ 79 /* Set the parent before signaling the scheduled fence, such that, 80 * any waiter expecting the parent to be filled after the job has 81 * been scheduled (which is the case for drivers delegating waits 82 * to some firmware) doesn't have to busy wait for parent to show 83 * up. 84 */ 85 if (!IS_ERR_OR_NULL(parent)) 86 drm_sched_fence_set_parent(fence, parent); 87 88 dma_fence_signal(&fence->scheduled); 89} 90 91void drm_sched_fence_finished(struct drm_sched_fence *fence, int result) 92{ 93 if (result) 94 dma_fence_set_error(&fence->finished, result); 95 dma_fence_signal(&fence->finished); 96} 97 98static const char *drm_sched_fence_get_driver_name(struct dma_fence *fence) 99{ 100 return "drm_sched"; 101} 102 103static const char *drm_sched_fence_get_timeline_name(struct dma_fence *f) 104{ 105 struct drm_sched_fence *fence = to_drm_sched_fence(f); 106 return (const char *)fence->sched->name; 107} 108 109static void drm_sched_fence_free_rcu(struct rcu_head *rcu) 110{ 111 struct dma_fence *f = container_of(rcu, struct dma_fence, rcu); 112 struct drm_sched_fence *fence = to_drm_sched_fence(f); 113 114 if (!WARN_ON_ONCE(!fence)) { 115#ifdef __linux__ 116 kmem_cache_free(sched_fence_slab, fence); 117#else 118 pool_put(&sched_fence_slab, fence); 119#endif 120 } 121} 122 123/** 124 * drm_sched_fence_free - free up an uninitialized fence 125 * 126 * @fence: fence to free 127 * 128 * Free up the fence memory. Should only be used if drm_sched_fence_init() 129 * has not been called yet. 130 */ 131void drm_sched_fence_free(struct drm_sched_fence *fence) 132{ 133 /* This function should not be called if the fence has been initialized. */ 134 if (!WARN_ON_ONCE(fence->sched)) { 135#ifdef __linux__ 136 kmem_cache_free(sched_fence_slab, fence); 137#else 138 pool_put(&sched_fence_slab, fence); 139#endif 140 } 141} 142 143/** 144 * drm_sched_fence_release_scheduled - callback that fence can be freed 145 * 146 * @f: fence 147 * 148 * This function is called when the reference count becomes zero. 149 * It just RCU schedules freeing up the fence. 150 */ 151static void drm_sched_fence_release_scheduled(struct dma_fence *f) 152{ 153 struct drm_sched_fence *fence = to_drm_sched_fence(f); 154 155 dma_fence_put(fence->parent); 156 call_rcu(&fence->finished.rcu, drm_sched_fence_free_rcu); 157} 158 159/** 160 * drm_sched_fence_release_finished - drop extra reference 161 * 162 * @f: fence 163 * 164 * Drop the extra reference from the scheduled fence to the base fence. 165 */ 166static void drm_sched_fence_release_finished(struct dma_fence *f) 167{ 168 struct drm_sched_fence *fence = to_drm_sched_fence(f); 169 170 dma_fence_put(&fence->scheduled); 171} 172 173static void drm_sched_fence_set_deadline_finished(struct dma_fence *f, 174 ktime_t deadline) 175{ 176 struct drm_sched_fence *fence = to_drm_sched_fence(f); 177 struct dma_fence *parent; 178 unsigned long flags; 179 180 spin_lock_irqsave(&fence->lock, flags); 181 182 /* If we already have an earlier deadline, keep it: */ 183 if (test_bit(DRM_SCHED_FENCE_FLAG_HAS_DEADLINE_BIT, &f->flags) && 184 ktime_before(fence->deadline, deadline)) { 185 spin_unlock_irqrestore(&fence->lock, flags); 186 return; 187 } 188 189 fence->deadline = deadline; 190 set_bit(DRM_SCHED_FENCE_FLAG_HAS_DEADLINE_BIT, &f->flags); 191 192 spin_unlock_irqrestore(&fence->lock, flags); 193 194 /* 195 * smp_load_aquire() to ensure that if we are racing another 196 * thread calling drm_sched_fence_set_parent(), that we see 197 * the parent set before it calls test_bit(HAS_DEADLINE_BIT) 198 */ 199 parent = smp_load_acquire(&fence->parent); 200 if (parent) 201 dma_fence_set_deadline(parent, deadline); 202} 203 204static const struct dma_fence_ops drm_sched_fence_ops_scheduled = { 205 .get_driver_name = drm_sched_fence_get_driver_name, 206 .get_timeline_name = drm_sched_fence_get_timeline_name, 207 .release = drm_sched_fence_release_scheduled, 208}; 209 210static const struct dma_fence_ops drm_sched_fence_ops_finished = { 211 .get_driver_name = drm_sched_fence_get_driver_name, 212 .get_timeline_name = drm_sched_fence_get_timeline_name, 213 .release = drm_sched_fence_release_finished, 214 .set_deadline = drm_sched_fence_set_deadline_finished, 215}; 216 217struct drm_sched_fence *to_drm_sched_fence(struct dma_fence *f) 218{ 219 if (f->ops == &drm_sched_fence_ops_scheduled) 220 return container_of(f, struct drm_sched_fence, scheduled); 221 222 if (f->ops == &drm_sched_fence_ops_finished) 223 return container_of(f, struct drm_sched_fence, finished); 224 225 return NULL; 226} 227EXPORT_SYMBOL(to_drm_sched_fence); 228 229struct drm_sched_fence *drm_sched_fence_alloc(struct drm_sched_entity *entity, 230 void *owner) 231{ 232 struct drm_sched_fence *fence = NULL; 233 234#ifdef __linux__ 235 fence = kmem_cache_zalloc(sched_fence_slab, GFP_KERNEL); 236#else 237 fence = pool_get(&sched_fence_slab, PR_WAITOK | PR_ZERO); 238#endif 239 if (fence == NULL) 240 return NULL; 241 242 fence->owner = owner; 243 mtx_init(&fence->lock, IPL_TTY); 244 245 return fence; 246} 247 248void drm_sched_fence_init(struct drm_sched_fence *fence, 249 struct drm_sched_entity *entity) 250{ 251 unsigned seq; 252 253 fence->sched = entity->rq->sched; 254 seq = atomic_inc_return(&entity->fence_seq); 255 dma_fence_init(&fence->scheduled, &drm_sched_fence_ops_scheduled, 256 &fence->lock, entity->fence_context, seq); 257 dma_fence_init(&fence->finished, &drm_sched_fence_ops_finished, 258 &fence->lock, entity->fence_context + 1, seq); 259} 260 261module_init(drm_sched_fence_slab_init); 262module_exit(drm_sched_fence_slab_fini); 263 264MODULE_DESCRIPTION("DRM GPU scheduler"); 265MODULE_LICENSE("GPL and additional rights"); 266