1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2021 Intel Corporation 4 */ 5 6 #include "xe_sync.h" 7 8 #include <linux/dma-fence-array.h> 9 #include <linux/kthread.h> 10 #include <linux/sched/mm.h> 11 #include <linux/uaccess.h> 12 13 #include <drm/drm_print.h> 14 #include <drm/drm_syncobj.h> 15 #include <uapi/drm/xe_drm.h> 16 17 #include "xe_device_types.h" 18 #include "xe_exec_queue.h" 19 #include "xe_macros.h" 20 #include "xe_sched_job_types.h" 21 22 struct xe_user_fence { 23 struct xe_device *xe; 24 struct kref refcount; 25 struct dma_fence_cb cb; 26 struct work_struct worker; 27 struct mm_struct *mm; 28 u64 __user *addr; 29 u64 value; 30 int signalled; 31 }; 32 33 static void user_fence_destroy(struct kref *kref) 34 { 35 struct xe_user_fence *ufence = container_of(kref, struct xe_user_fence, 36 refcount); 37 38 mmdrop(ufence->mm); 39 kfree(ufence); 40 } 41 42 static void user_fence_get(struct xe_user_fence *ufence) 43 { 44 kref_get(&ufence->refcount); 45 } 46 47 static void user_fence_put(struct xe_user_fence *ufence) 48 { 49 kref_put(&ufence->refcount, user_fence_destroy); 50 } 51 52 static struct xe_user_fence *user_fence_create(struct xe_device *xe, u64 addr, 53 u64 value) 54 { 55 struct xe_user_fence *ufence; 56 u64 __user *ptr = u64_to_user_ptr(addr); 57 u64 __maybe_unused prefetch_val; 58 59 if (get_user(prefetch_val, ptr)) 60 return ERR_PTR(-EFAULT); 61 62 ufence = kzalloc(sizeof(*ufence), GFP_KERNEL); 63 if (!ufence) 64 return ERR_PTR(-ENOMEM); 65 66 ufence->xe = xe; 67 kref_init(&ufence->refcount); 68 ufence->addr = ptr; 69 ufence->value = value; 70 ufence->mm = current->mm; 71 mmgrab(ufence->mm); 72 73 return ufence; 74 } 75 76 static void user_fence_worker(struct work_struct *w) 77 { 78 struct xe_user_fence *ufence = container_of(w, struct xe_user_fence, worker); 79 80 if (mmget_not_zero(ufence->mm)) { 81 kthread_use_mm(ufence->mm); 82 if (copy_to_user(ufence->addr, &ufence->value, sizeof(ufence->value))) 83 XE_WARN_ON("Copy to user failed"); 84 kthread_unuse_mm(ufence->mm); 85 mmput(ufence->mm); 86 } else { 87 drm_dbg(&ufence->xe->drm, "mmget_not_zero() failed, ufence wasn't signaled\n"); 88 } 89 90 wake_up_all(&ufence->xe->ufence_wq); 91 WRITE_ONCE(ufence->signalled, 1); 92 user_fence_put(ufence); 93 } 94 95 static void kick_ufence(struct xe_user_fence *ufence, struct dma_fence *fence) 96 { 97 INIT_WORK(&ufence->worker, user_fence_worker); 98 queue_work(ufence->xe->ordered_wq, &ufence->worker); 99 dma_fence_put(fence); 100 } 101 102 static void user_fence_cb(struct dma_fence *fence, struct dma_fence_cb *cb) 103 { 104 struct xe_user_fence *ufence = container_of(cb, struct xe_user_fence, cb); 105 106 kick_ufence(ufence, fence); 107 } 108 109 int xe_sync_entry_parse(struct xe_device *xe, struct xe_file *xef, 110 struct xe_sync_entry *sync, 111 struct drm_xe_sync __user *sync_user, 112 unsigned int flags) 113 { 114 struct drm_xe_sync sync_in; 115 int err; 116 bool exec = flags & SYNC_PARSE_FLAG_EXEC; 117 bool in_lr_mode = flags & SYNC_PARSE_FLAG_LR_MODE; 118 bool disallow_user_fence = flags & SYNC_PARSE_FLAG_DISALLOW_USER_FENCE; 119 bool signal; 120 121 if (copy_from_user(&sync_in, sync_user, sizeof(*sync_user))) 122 return -EFAULT; 123 124 if (XE_IOCTL_DBG(xe, sync_in.flags & ~DRM_XE_SYNC_FLAG_SIGNAL) || 125 XE_IOCTL_DBG(xe, sync_in.reserved[0] || sync_in.reserved[1])) 126 return -EINVAL; 127 128 signal = sync_in.flags & DRM_XE_SYNC_FLAG_SIGNAL; 129 switch (sync_in.type) { 130 case DRM_XE_SYNC_TYPE_SYNCOBJ: 131 if (XE_IOCTL_DBG(xe, in_lr_mode && signal)) 132 return -EOPNOTSUPP; 133 134 if (XE_IOCTL_DBG(xe, upper_32_bits(sync_in.addr))) 135 return -EINVAL; 136 137 sync->syncobj = drm_syncobj_find(xef->drm, sync_in.handle); 138 if (XE_IOCTL_DBG(xe, !sync->syncobj)) 139 return -ENOENT; 140 141 if (!signal) { 142 sync->fence = drm_syncobj_fence_get(sync->syncobj); 143 if (XE_IOCTL_DBG(xe, !sync->fence)) 144 return -EINVAL; 145 } 146 break; 147 148 case DRM_XE_SYNC_TYPE_TIMELINE_SYNCOBJ: 149 if (XE_IOCTL_DBG(xe, in_lr_mode && signal)) 150 return -EOPNOTSUPP; 151 152 if (XE_IOCTL_DBG(xe, upper_32_bits(sync_in.addr))) 153 return -EINVAL; 154 155 if (XE_IOCTL_DBG(xe, sync_in.timeline_value == 0)) 156 return -EINVAL; 157 158 sync->syncobj = drm_syncobj_find(xef->drm, sync_in.handle); 159 if (XE_IOCTL_DBG(xe, !sync->syncobj)) 160 return -ENOENT; 161 162 if (signal) { 163 sync->chain_fence = dma_fence_chain_alloc(); 164 if (!sync->chain_fence) 165 return -ENOMEM; 166 } else { 167 sync->fence = drm_syncobj_fence_get(sync->syncobj); 168 if (XE_IOCTL_DBG(xe, !sync->fence)) 169 return -EINVAL; 170 171 err = dma_fence_chain_find_seqno(&sync->fence, 172 sync_in.timeline_value); 173 if (err) 174 return err; 175 } 176 break; 177 178 case DRM_XE_SYNC_TYPE_USER_FENCE: 179 if (XE_IOCTL_DBG(xe, disallow_user_fence)) 180 return -EOPNOTSUPP; 181 182 if (XE_IOCTL_DBG(xe, !signal)) 183 return -EOPNOTSUPP; 184 185 if (XE_IOCTL_DBG(xe, sync_in.addr & 0x7)) 186 return -EINVAL; 187 188 if (exec) { 189 sync->addr = sync_in.addr; 190 } else { 191 sync->ufence = user_fence_create(xe, sync_in.addr, 192 sync_in.timeline_value); 193 if (XE_IOCTL_DBG(xe, IS_ERR(sync->ufence))) 194 return PTR_ERR(sync->ufence); 195 } 196 197 break; 198 199 default: 200 return -EINVAL; 201 } 202 203 sync->type = sync_in.type; 204 sync->flags = sync_in.flags; 205 sync->timeline_value = sync_in.timeline_value; 206 207 return 0; 208 } 209 210 int xe_sync_entry_add_deps(struct xe_sync_entry *sync, struct xe_sched_job *job) 211 { 212 if (sync->fence) 213 return drm_sched_job_add_dependency(&job->drm, 214 dma_fence_get(sync->fence)); 215 216 return 0; 217 } 218 219 void xe_sync_entry_signal(struct xe_sync_entry *sync, struct dma_fence *fence) 220 { 221 if (!(sync->flags & DRM_XE_SYNC_FLAG_SIGNAL)) 222 return; 223 224 if (sync->chain_fence) { 225 drm_syncobj_add_point(sync->syncobj, sync->chain_fence, 226 fence, sync->timeline_value); 227 /* 228 * The chain's ownership is transferred to the 229 * timeline. 230 */ 231 sync->chain_fence = NULL; 232 } else if (sync->syncobj) { 233 drm_syncobj_replace_fence(sync->syncobj, fence); 234 } else if (sync->ufence) { 235 int err; 236 237 dma_fence_get(fence); 238 user_fence_get(sync->ufence); 239 err = dma_fence_add_callback(fence, &sync->ufence->cb, 240 user_fence_cb); 241 if (err == -ENOENT) { 242 kick_ufence(sync->ufence, fence); 243 } else if (err) { 244 XE_WARN_ON("failed to add user fence"); 245 user_fence_put(sync->ufence); 246 dma_fence_put(fence); 247 } 248 } 249 } 250 251 void xe_sync_entry_cleanup(struct xe_sync_entry *sync) 252 { 253 if (sync->syncobj) 254 drm_syncobj_put(sync->syncobj); 255 dma_fence_put(sync->fence); 256 dma_fence_chain_free(sync->chain_fence); 257 if (sync->ufence) 258 user_fence_put(sync->ufence); 259 } 260 261 /** 262 * xe_sync_in_fence_get() - Get a fence from syncs, exec queue, and VM 263 * @sync: input syncs 264 * @num_sync: number of syncs 265 * @q: exec queue 266 * @vm: VM 267 * 268 * Get a fence from syncs, exec queue, and VM. If syncs contain in-fences create 269 * and return a composite fence of all in-fences + last fence. If no in-fences 270 * return last fence on input exec queue. Caller must drop reference to 271 * returned fence. 272 * 273 * Return: fence on success, ERR_PTR(-ENOMEM) on failure 274 */ 275 struct dma_fence * 276 xe_sync_in_fence_get(struct xe_sync_entry *sync, int num_sync, 277 struct xe_exec_queue *q, struct xe_vm *vm) 278 { 279 struct dma_fence **fences = NULL; 280 struct dma_fence_array *cf = NULL; 281 struct dma_fence *fence; 282 int i, num_in_fence = 0, current_fence = 0; 283 284 lockdep_assert_held(&vm->lock); 285 286 /* Count in-fences */ 287 for (i = 0; i < num_sync; ++i) { 288 if (sync[i].fence) { 289 ++num_in_fence; 290 fence = sync[i].fence; 291 } 292 } 293 294 /* Easy case... */ 295 if (!num_in_fence) { 296 fence = xe_exec_queue_last_fence_get(q, vm); 297 return fence; 298 } 299 300 /* Create composite fence */ 301 fences = kmalloc_array(num_in_fence + 1, sizeof(*fences), GFP_KERNEL); 302 if (!fences) 303 return ERR_PTR(-ENOMEM); 304 for (i = 0; i < num_sync; ++i) { 305 if (sync[i].fence) { 306 dma_fence_get(sync[i].fence); 307 fences[current_fence++] = sync[i].fence; 308 } 309 } 310 fences[current_fence++] = xe_exec_queue_last_fence_get(q, vm); 311 cf = dma_fence_array_create(num_in_fence, fences, 312 vm->composite_fence_ctx, 313 vm->composite_fence_seqno++, 314 false); 315 if (!cf) { 316 --vm->composite_fence_seqno; 317 goto err_out; 318 } 319 320 return &cf->base; 321 322 err_out: 323 while (current_fence) 324 dma_fence_put(fences[--current_fence]); 325 kfree(fences); 326 kfree(cf); 327 328 return ERR_PTR(-ENOMEM); 329 } 330 331 /** 332 * __xe_sync_ufence_get() - Get user fence from user fence 333 * @ufence: input user fence 334 * 335 * Get a user fence reference from user fence 336 * 337 * Return: xe_user_fence pointer with reference 338 */ 339 struct xe_user_fence *__xe_sync_ufence_get(struct xe_user_fence *ufence) 340 { 341 user_fence_get(ufence); 342 343 return ufence; 344 } 345 346 /** 347 * xe_sync_ufence_get() - Get user fence from sync 348 * @sync: input sync 349 * 350 * Get a user fence reference from sync. 351 * 352 * Return: xe_user_fence pointer with reference 353 */ 354 struct xe_user_fence *xe_sync_ufence_get(struct xe_sync_entry *sync) 355 { 356 user_fence_get(sync->ufence); 357 358 return sync->ufence; 359 } 360 361 /** 362 * xe_sync_ufence_put() - Put user fence reference 363 * @ufence: user fence reference 364 * 365 */ 366 void xe_sync_ufence_put(struct xe_user_fence *ufence) 367 { 368 user_fence_put(ufence); 369 } 370 371 /** 372 * xe_sync_ufence_get_status() - Get user fence status 373 * @ufence: user fence 374 * 375 * Return: 1 if signalled, 0 not signalled, <0 on error 376 */ 377 int xe_sync_ufence_get_status(struct xe_user_fence *ufence) 378 { 379 return READ_ONCE(ufence->signalled); 380 } 381