1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2021 Intel Corporation 4 */ 5 6 #include "xe_sched_job.h" 7 8 #include <uapi/drm/xe_drm.h> 9 #include <linux/dma-fence-chain.h> 10 #include <linux/slab.h> 11 12 #include "xe_device.h" 13 #include "xe_exec_queue.h" 14 #include "xe_gt_types.h" 15 #include "xe_hw_engine_types.h" 16 #include "xe_hw_fence.h" 17 #include "xe_lrc.h" 18 #include "xe_macros.h" 19 #include "xe_pm.h" 20 #include "xe_sync_types.h" 21 #include "xe_trace.h" 22 #include "xe_vm.h" 23 24 static struct kmem_cache *xe_sched_job_slab; 25 static struct kmem_cache *xe_sched_job_parallel_slab; 26 27 int __init xe_sched_job_module_init(void) 28 { 29 xe_sched_job_slab = 30 kmem_cache_create("xe_sched_job", 31 sizeof(struct xe_sched_job) + 32 sizeof(struct xe_job_ptrs), 0, 33 SLAB_HWCACHE_ALIGN, NULL); 34 if (!xe_sched_job_slab) 35 return -ENOMEM; 36 37 xe_sched_job_parallel_slab = 38 kmem_cache_create("xe_sched_job_parallel", 39 sizeof(struct xe_sched_job) + 40 sizeof(struct xe_job_ptrs) * 41 XE_HW_ENGINE_MAX_INSTANCE, 0, 42 SLAB_HWCACHE_ALIGN, NULL); 43 if (!xe_sched_job_parallel_slab) { 44 kmem_cache_destroy(xe_sched_job_slab); 45 return -ENOMEM; 46 } 47 48 return 0; 49 } 50 51 void xe_sched_job_module_exit(void) 52 { 53 kmem_cache_destroy(xe_sched_job_slab); 54 kmem_cache_destroy(xe_sched_job_parallel_slab); 55 } 56 57 static struct xe_sched_job *job_alloc(bool parallel) 58 { 59 return kmem_cache_zalloc(parallel ? xe_sched_job_parallel_slab : 60 xe_sched_job_slab, GFP_KERNEL); 61 } 62 63 bool xe_sched_job_is_migration(struct xe_exec_queue *q) 64 { 65 return q->vm && (q->vm->flags & XE_VM_FLAG_MIGRATION); 66 } 67 68 static void job_free(struct xe_sched_job *job) 69 { 70 struct xe_exec_queue *q = job->q; 71 bool is_migration = xe_sched_job_is_migration(q); 72 73 kmem_cache_free(xe_exec_queue_is_parallel(job->q) || is_migration ? 74 xe_sched_job_parallel_slab : xe_sched_job_slab, job); 75 } 76 77 static struct xe_device *job_to_xe(struct xe_sched_job *job) 78 { 79 return gt_to_xe(job->q->gt); 80 } 81 82 /* Free unused pre-allocated fences */ 83 static void xe_sched_job_free_fences(struct xe_sched_job *job) 84 { 85 int i; 86 87 for (i = 0; i < job->q->width; ++i) { 88 struct xe_job_ptrs *ptrs = &job->ptrs[i]; 89 90 if (ptrs->lrc_fence) 91 xe_lrc_free_seqno_fence(ptrs->lrc_fence); 92 dma_fence_chain_free(ptrs->chain_fence); 93 } 94 } 95 96 struct xe_sched_job *xe_sched_job_create(struct xe_exec_queue *q, 97 u64 *batch_addr) 98 { 99 bool is_migration = xe_sched_job_is_migration(q); 100 struct xe_sched_job *job; 101 int err; 102 int i; 103 u32 width; 104 105 /* only a kernel context can submit a vm-less job */ 106 XE_WARN_ON(!q->vm && !(q->flags & EXEC_QUEUE_FLAG_KERNEL)); 107 108 job = job_alloc(xe_exec_queue_is_parallel(q) || is_migration); 109 if (!job) 110 return ERR_PTR(-ENOMEM); 111 112 job->q = q; 113 job->sample_timestamp = U64_MAX; 114 kref_init(&job->refcount); 115 xe_exec_queue_get(job->q); 116 117 err = drm_sched_job_init(&job->drm, q->entity, 1, NULL, 118 q->xef ? q->xef->drm->client_id : 0); 119 if (err) 120 goto err_free; 121 122 for (i = 0; i < q->width; ++i) { 123 struct dma_fence *fence = xe_lrc_alloc_seqno_fence(); 124 struct dma_fence_chain *chain; 125 126 if (IS_ERR(fence)) { 127 err = PTR_ERR(fence); 128 goto err_sched_job; 129 } 130 job->ptrs[i].lrc_fence = fence; 131 132 if (i + 1 == q->width) 133 continue; 134 135 chain = dma_fence_chain_alloc(); 136 if (!chain) { 137 err = -ENOMEM; 138 goto err_sched_job; 139 } 140 job->ptrs[i].chain_fence = chain; 141 } 142 143 width = q->width; 144 if (is_migration) 145 width = 2; 146 147 for (i = 0; i < width; ++i) 148 job->ptrs[i].batch_addr = batch_addr[i]; 149 150 atomic_inc(&q->job_cnt); 151 xe_pm_runtime_get_noresume(job_to_xe(job)); 152 trace_xe_sched_job_create(job); 153 return job; 154 155 err_sched_job: 156 xe_sched_job_free_fences(job); 157 drm_sched_job_cleanup(&job->drm); 158 err_free: 159 xe_exec_queue_put(q); 160 job_free(job); 161 return ERR_PTR(err); 162 } 163 164 /** 165 * xe_sched_job_destroy - Destroy Xe schedule job 166 * @ref: reference to Xe schedule job 167 * 168 * Called when ref == 0, drop a reference to job's xe_engine + fence, cleanup 169 * base DRM schedule job, and free memory for Xe schedule job. 170 */ 171 void xe_sched_job_destroy(struct kref *ref) 172 { 173 struct xe_sched_job *job = 174 container_of(ref, struct xe_sched_job, refcount); 175 struct xe_device *xe = job_to_xe(job); 176 struct xe_exec_queue *q = job->q; 177 178 xe_sched_job_free_fences(job); 179 dma_fence_put(job->fence); 180 drm_sched_job_cleanup(&job->drm); 181 job_free(job); 182 atomic_dec(&q->job_cnt); 183 xe_exec_queue_put(q); 184 xe_pm_runtime_put(xe); 185 } 186 187 /* Set the error status under the fence to avoid racing with signaling */ 188 static bool xe_fence_set_error(struct dma_fence *fence, int error) 189 { 190 unsigned long irq_flags; 191 bool signaled; 192 193 spin_lock_irqsave(fence->lock, irq_flags); 194 signaled = test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags); 195 if (!signaled) 196 dma_fence_set_error(fence, error); 197 spin_unlock_irqrestore(fence->lock, irq_flags); 198 199 return signaled; 200 } 201 202 void xe_sched_job_set_error(struct xe_sched_job *job, int error) 203 { 204 if (xe_fence_set_error(job->fence, error)) 205 return; 206 207 if (dma_fence_is_chain(job->fence)) { 208 struct dma_fence *iter; 209 210 dma_fence_chain_for_each(iter, job->fence) 211 xe_fence_set_error(dma_fence_chain_contained(iter), 212 error); 213 } 214 215 trace_xe_sched_job_set_error(job); 216 217 dma_fence_enable_sw_signaling(job->fence); 218 xe_hw_fence_irq_run(job->q->fence_irq); 219 } 220 221 bool xe_sched_job_started(struct xe_sched_job *job) 222 { 223 struct dma_fence *fence = dma_fence_chain_contained(job->fence); 224 struct xe_lrc *lrc = job->q->lrc[0]; 225 226 return !__dma_fence_is_later(fence, 227 xe_sched_job_lrc_seqno(job), 228 xe_lrc_start_seqno(lrc)); 229 } 230 231 bool xe_sched_job_completed(struct xe_sched_job *job) 232 { 233 struct dma_fence *fence = dma_fence_chain_contained(job->fence); 234 struct xe_lrc *lrc = job->q->lrc[0]; 235 236 /* 237 * Can safely check just LRC[0] seqno as that is last seqno written when 238 * parallel handshake is done. 239 */ 240 241 return !__dma_fence_is_later(fence, 242 xe_sched_job_lrc_seqno(job), 243 xe_lrc_seqno(lrc)); 244 } 245 246 void xe_sched_job_arm(struct xe_sched_job *job) 247 { 248 struct xe_exec_queue *q = job->q; 249 struct dma_fence *fence, *prev; 250 struct xe_vm *vm = q->vm; 251 u64 seqno = 0; 252 int i; 253 254 /* Migration and kernel engines have their own locking */ 255 if (IS_ENABLED(CONFIG_LOCKDEP) && 256 !(q->flags & (EXEC_QUEUE_FLAG_KERNEL | EXEC_QUEUE_FLAG_VM))) { 257 lockdep_assert_held(&q->vm->lock); 258 if (!xe_vm_in_lr_mode(q->vm)) 259 xe_vm_assert_held(q->vm); 260 } 261 262 if (vm && !xe_sched_job_is_migration(q) && !xe_vm_in_lr_mode(vm) && 263 (vm->batch_invalidate_tlb || vm->tlb_flush_seqno != q->tlb_flush_seqno)) { 264 xe_vm_assert_held(vm); 265 q->tlb_flush_seqno = vm->tlb_flush_seqno; 266 job->ring_ops_flush_tlb = true; 267 } 268 269 /* Arm the pre-allocated fences */ 270 for (i = 0; i < q->width; prev = fence, ++i) { 271 struct dma_fence_chain *chain; 272 273 fence = job->ptrs[i].lrc_fence; 274 xe_lrc_init_seqno_fence(q->lrc[i], fence); 275 job->ptrs[i].lrc_fence = NULL; 276 if (!i) { 277 job->lrc_seqno = fence->seqno; 278 continue; 279 } else { 280 xe_assert(gt_to_xe(q->gt), job->lrc_seqno == fence->seqno); 281 } 282 283 chain = job->ptrs[i - 1].chain_fence; 284 dma_fence_chain_init(chain, prev, fence, seqno++); 285 job->ptrs[i - 1].chain_fence = NULL; 286 fence = &chain->base; 287 } 288 289 job->fence = dma_fence_get(fence); /* Pairs with put in scheduler */ 290 drm_sched_job_arm(&job->drm); 291 } 292 293 void xe_sched_job_push(struct xe_sched_job *job) 294 { 295 xe_sched_job_get(job); 296 trace_xe_sched_job_exec(job); 297 drm_sched_entity_push_job(&job->drm); 298 xe_sched_job_put(job); 299 } 300 301 /** 302 * xe_sched_job_init_user_fence - Initialize user_fence for the job 303 * @job: job whose user_fence needs an init 304 * @sync: sync to be use to init user_fence 305 */ 306 void xe_sched_job_init_user_fence(struct xe_sched_job *job, 307 struct xe_sync_entry *sync) 308 { 309 if (sync->type != DRM_XE_SYNC_TYPE_USER_FENCE) 310 return; 311 312 job->user_fence.used = true; 313 job->user_fence.addr = sync->addr; 314 job->user_fence.value = sync->timeline_value; 315 } 316 317 struct xe_sched_job_snapshot * 318 xe_sched_job_snapshot_capture(struct xe_sched_job *job) 319 { 320 struct xe_exec_queue *q = job->q; 321 struct xe_device *xe = q->gt->tile->xe; 322 struct xe_sched_job_snapshot *snapshot; 323 size_t len = sizeof(*snapshot) + (sizeof(u64) * q->width); 324 u16 i; 325 326 snapshot = kzalloc(len, GFP_ATOMIC); 327 if (!snapshot) 328 return NULL; 329 330 snapshot->batch_addr_len = q->width; 331 for (i = 0; i < q->width; i++) 332 snapshot->batch_addr[i] = 333 xe_device_uncanonicalize_addr(xe, job->ptrs[i].batch_addr); 334 335 return snapshot; 336 } 337 338 void xe_sched_job_snapshot_free(struct xe_sched_job_snapshot *snapshot) 339 { 340 kfree(snapshot); 341 } 342 343 void 344 xe_sched_job_snapshot_print(struct xe_sched_job_snapshot *snapshot, 345 struct drm_printer *p) 346 { 347 u16 i; 348 349 if (!snapshot) 350 return; 351 352 for (i = 0; i < snapshot->batch_addr_len; i++) 353 drm_printf(p, "batch_addr[%u]: 0x%016llx\n", i, snapshot->batch_addr[i]); 354 } 355 356 int xe_sched_job_add_deps(struct xe_sched_job *job, struct dma_resv *resv, 357 enum dma_resv_usage usage) 358 { 359 return drm_sched_job_add_resv_dependencies(&job->drm, resv, usage); 360 } 361