drivers/gpu/drm/amd/amdkfd/kfd_process.c

   1 /*
   2  * Copyright 2014 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 #include <linux/mutex.h>
  24 #include <linux/log2.h>
  25 #include <linux/sched.h>
  26 #include <linux/slab.h>
  27 #include <linux/amd-iommu.h>
  28 #include <linux/notifier.h>
  29 #include <linux/compat.h>
  30
  31 struct mm_struct;
  32
  33 #include "kfd_priv.h"
  34 #include "kfd_dbgmgr.h"
  35
  36 /*
  37  * Initial size for the array of queues.
  38  * The allocated size is doubled each time
  39  * it is exceeded up to MAX_PROCESS_QUEUES.
  40  */
  41 #define INITIAL_QUEUE_ARRAY_SIZE 16
  42
  43 /*
  44  * List of struct kfd_process (field kfd_process).
  45  * Unique/indexed by mm_struct*
  46  */
  47 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */
  48 static DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
  49 static DEFINE_MUTEX(kfd_processes_mutex);
  50
  51 DEFINE_STATIC_SRCU(kfd_processes_srcu);
  52
  53 static struct workqueue_struct *kfd_process_wq;
  54
  55 struct kfd_process_release_work {
  56         struct work_struct kfd_work;
  57         struct kfd_process *p;
  58 };
  59
  60 static struct kfd_process *find_process(const struct task_struct *thread);
  61 static struct kfd_process *create_process(const struct task_struct *thread);
  62
  63 void kfd_process_create_wq(void)
  64 {
  65         if (!kfd_process_wq)
  66                 kfd_process_wq = create_workqueue("kfd_process_wq");
  67 }
  68
  69 void kfd_process_destroy_wq(void)
  70 {
  71         if (kfd_process_wq) {
  72                 flush_workqueue(kfd_process_wq);
  73                 destroy_workqueue(kfd_process_wq);
  74                 kfd_process_wq = NULL;
  75         }
  76 }
  77
  78 struct kfd_process *kfd_create_process(const struct task_struct *thread)
  79 {
  80         struct kfd_process *process;
  81
  82         BUG_ON(!kfd_process_wq);
  83
  84         if (thread->mm == NULL)
  85                 return ERR_PTR(-EINVAL);
  86
  87         /* Only the pthreads threading model is supported. */
  88         if (thread->group_leader->mm != thread->mm)
  89                 return ERR_PTR(-EINVAL);
  90
  91         /* Take mmap_sem because we call __mmu_notifier_register inside */
  92         down_write(&thread->mm->mmap_sem);
  93
  94         /*
  95          * take kfd processes mutex before starting of process creation
  96          * so there won't be a case where two threads of the same process
  97          * create two kfd_process structures
  98          */
  99         mutex_lock(&kfd_processes_mutex);
 100
 101         /* A prior open of /dev/kfd could have already created the process. */
 102         process = find_process(thread);
 103         if (process)
 104                 pr_debug("kfd: process already found\n");
 105
 106         if (!process)
 107                 process = create_process(thread);
 108
 109         mutex_unlock(&kfd_processes_mutex);
 110
 111         up_write(&thread->mm->mmap_sem);
 112
 113         return process;
 114 }
 115
 116 struct kfd_process *kfd_get_process(const struct task_struct *thread)
 117 {
 118         struct kfd_process *process;
 119
 120         if (thread->mm == NULL)
 121                 return ERR_PTR(-EINVAL);
 122
 123         /* Only the pthreads threading model is supported. */
 124         if (thread->group_leader->mm != thread->mm)
 125                 return ERR_PTR(-EINVAL);
 126
 127         process = find_process(thread);
 128         if (!process)
 129                 return ERR_PTR(-EINVAL);
 130
 131         return process;
 132 }
 133
 134 static struct kfd_process *find_process_by_mm(const struct mm_struct *mm)
 135 {
 136         struct kfd_process *process;
 137
 138         hash_for_each_possible_rcu(kfd_processes_table, process,
 139                                         kfd_processes, (uintptr_t)mm)
 140                 if (process->mm == mm)
 141                         return process;
 142
 143         return NULL;
 144 }
 145
 146 static struct kfd_process *find_process(const struct task_struct *thread)
 147 {
 148         struct kfd_process *p;
 149         int idx;
 150
 151         idx = srcu_read_lock(&kfd_processes_srcu);
 152         p = find_process_by_mm(thread->mm);
 153         srcu_read_unlock(&kfd_processes_srcu, idx);
 154
 155         return p;
 156 }
 157
 158 static void kfd_process_wq_release(struct work_struct *work)
 159 {
 160         struct kfd_process_release_work *my_work;
 161         struct kfd_process_device *pdd, *temp;
 162         struct kfd_process *p;
 163
 164         my_work = (struct kfd_process_release_work *) work;
 165
 166         p = my_work->p;
 167
 168         pr_debug("Releasing process (pasid %d) in workqueue\n",
 169                         p->pasid);
 170
 171         mutex_lock(&p->mutex);
 172
 173         list_for_each_entry_safe(pdd, temp, &p->per_device_data,
 174                                                         per_device_list) {
 175                 pr_debug("Releasing pdd (topology id %d) for process (pasid %d) in workqueue\n",
 176                                 pdd->dev->id, p->pasid);
 177
 178                 if (pdd->reset_wavefronts)
 179                         dbgdev_wave_reset_wavefronts(pdd->dev, p);
 180
 181                 amd_iommu_unbind_pasid(pdd->dev->pdev, p->pasid);
 182                 list_del(&pdd->per_device_list);
 183
 184                 kfree(pdd);
 185         }
 186
 187         kfd_event_free_process(p);
 188
 189         kfd_pasid_free(p->pasid);
 190
 191         mutex_unlock(&p->mutex);
 192
 193         mutex_destroy(&p->mutex);
 194
 195         kfree(p->queues);
 196
 197         kfree(p);
 198
 199         kfree((void *)work);
 200 }
 201
 202 static void kfd_process_destroy_delayed(struct rcu_head *rcu)
 203 {
 204         struct kfd_process_release_work *work;
 205         struct kfd_process *p;
 206
 207         BUG_ON(!kfd_process_wq);
 208
 209         p = container_of(rcu, struct kfd_process, rcu);
 210         BUG_ON(atomic_read(&p->mm->mm_count) <= 0);
 211
 212         mmdrop(p->mm);
 213
 214         work = kmalloc(sizeof(struct kfd_process_release_work), GFP_ATOMIC);
 215
 216         if (work) {
 217                 INIT_WORK((struct work_struct *) work, kfd_process_wq_release);
 218                 work->p = p;
 219                 queue_work(kfd_process_wq, (struct work_struct *) work);
 220         }
 221 }
 222
 223 static void kfd_process_notifier_release(struct mmu_notifier *mn,
 224                                         struct mm_struct *mm)
 225 {
 226         struct kfd_process *p;
 227         struct kfd_process_device *pdd = NULL;
 228
 229         /*
 230          * The kfd_process structure can not be free because the
 231          * mmu_notifier srcu is read locked
 232          */
 233         p = container_of(mn, struct kfd_process, mmu_notifier);
 234         BUG_ON(p->mm != mm);
 235
 236         mutex_lock(&kfd_processes_mutex);
 237         hash_del_rcu(&p->kfd_processes);
 238         mutex_unlock(&kfd_processes_mutex);
 239         synchronize_srcu(&kfd_processes_srcu);
 240
 241         mutex_lock(&p->mutex);
 242
 243         /* In case our notifier is called before IOMMU notifier */
 244         pqm_uninit(&p->pqm);
 245
 246         /* Iterate over all process device data structure and check
 247          * if we should delete debug managers and reset all wavefronts
 248          */
 249         list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
 250                 if ((pdd->dev->dbgmgr) &&
 251                                 (pdd->dev->dbgmgr->pasid == p->pasid))
 252                         kfd_dbgmgr_destroy(pdd->dev->dbgmgr);
 253
 254                 if (pdd->reset_wavefronts) {
 255                         pr_warn("amdkfd: Resetting all wave fronts\n");
 256                         dbgdev_wave_reset_wavefronts(pdd->dev, p);
 257                         pdd->reset_wavefronts = false;
 258                 }
 259         }
 260
 261         mutex_unlock(&p->mutex);
 262
 263         /*
 264          * Because we drop mm_count inside kfd_process_destroy_delayed
 265          * and because the mmu_notifier_unregister function also drop
 266          * mm_count we need to take an extra count here.
 267          */
 268         atomic_inc(&p->mm->mm_count);
 269         mmu_notifier_unregister_no_release(&p->mmu_notifier, p->mm);
 270         mmu_notifier_call_srcu(&p->rcu, &kfd_process_destroy_delayed);
 271 }
 272
 273 static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = {
 274         .release = kfd_process_notifier_release,
 275 };
 276
 277 static struct kfd_process *create_process(const struct task_struct *thread)
 278 {
 279         struct kfd_process *process;
 280         int err = -ENOMEM;
 281
 282         process = kzalloc(sizeof(*process), GFP_KERNEL);
 283
 284         if (!process)
 285                 goto err_alloc_process;
 286
 287         process->queues = kmalloc_array(INITIAL_QUEUE_ARRAY_SIZE,
 288                                         sizeof(process->queues[0]), GFP_KERNEL);
 289         if (!process->queues)
 290                 goto err_alloc_queues;
 291
 292         process->pasid = kfd_pasid_alloc();
 293         if (process->pasid == 0)
 294                 goto err_alloc_pasid;
 295
 296         mutex_init(&process->mutex);
 297
 298         process->mm = thread->mm;
 299
 300         /* register notifier */
 301         process->mmu_notifier.ops = &kfd_process_mmu_notifier_ops;
 302         err = __mmu_notifier_register(&process->mmu_notifier, process->mm);
 303         if (err)
 304                 goto err_mmu_notifier;
 305
 306         hash_add_rcu(kfd_processes_table, &process->kfd_processes,
 307                         (uintptr_t)process->mm);
 308
 309         process->lead_thread = thread->group_leader;
 310
 311         process->queue_array_size = INITIAL_QUEUE_ARRAY_SIZE;
 312
 313         INIT_LIST_HEAD(&process->per_device_data);
 314
 315         kfd_event_init_process(process);
 316
 317         err = pqm_init(&process->pqm, process);
 318         if (err != 0)
 319                 goto err_process_pqm_init;
 320
 321         /* init process apertures*/
 322         process->is_32bit_user_mode = is_compat_task();
 323         if (kfd_init_apertures(process) != 0)
 324                 goto err_init_apretures;
 325
 326         return process;
 327
 328 err_init_apretures:
 329         pqm_uninit(&process->pqm);
 330 err_process_pqm_init:
 331         hash_del_rcu(&process->kfd_processes);
 332         synchronize_rcu();
 333         mmu_notifier_unregister_no_release(&process->mmu_notifier, process->mm);
 334 err_mmu_notifier:
 335         kfd_pasid_free(process->pasid);
 336 err_alloc_pasid:
 337         kfree(process->queues);
 338 err_alloc_queues:
 339         kfree(process);
 340 err_alloc_process:
 341         return ERR_PTR(err);
 342 }
 343
 344 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
 345                                                         struct kfd_process *p)
 346 {
 347         struct kfd_process_device *pdd = NULL;
 348
 349         list_for_each_entry(pdd, &p->per_device_data, per_device_list)
 350                 if (pdd->dev == dev)
 351                         break;
 352
 353         return pdd;
 354 }
 355
 356 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
 357                                                         struct kfd_process *p)
 358 {
 359         struct kfd_process_device *pdd = NULL;
 360
 361         pdd = kzalloc(sizeof(*pdd), GFP_KERNEL);
 362         if (pdd != NULL) {
 363                 pdd->dev = dev;
 364                 INIT_LIST_HEAD(&pdd->qpd.queues_list);
 365                 INIT_LIST_HEAD(&pdd->qpd.priv_queue_list);
 366                 pdd->qpd.dqm = dev->dqm;
 367                 pdd->reset_wavefronts = false;
 368                 list_add(&pdd->per_device_list, &p->per_device_data);
 369         }
 370
 371         return pdd;
 372 }
 373
 374 /*
 375  * Direct the IOMMU to bind the process (specifically the pasid->mm)
 376  * to the device.
 377  * Unbinding occurs when the process dies or the device is removed.
 378  *
 379  * Assumes that the process lock is held.
 380  */
 381 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
 382                                                         struct kfd_process *p)
 383 {
 384         struct kfd_process_device *pdd;
 385         int err;
 386
 387         pdd = kfd_get_process_device_data(dev, p);
 388         if (!pdd) {
 389                 pr_err("Process device data doesn't exist\n");
 390                 return ERR_PTR(-ENOMEM);
 391         }
 392
 393         if (pdd->bound)
 394                 return pdd;
 395
 396         err = amd_iommu_bind_pasid(dev->pdev, p->pasid, p->lead_thread);
 397         if (err < 0)
 398                 return ERR_PTR(err);
 399
 400         pdd->bound = true;
 401
 402         return pdd;
 403 }
 404
 405 void kfd_unbind_process_from_device(struct kfd_dev *dev, unsigned int pasid)
 406 {
 407         struct kfd_process *p;
 408         struct kfd_process_device *pdd;
 409         int idx, i;
 410
 411         BUG_ON(dev == NULL);
 412
 413         idx = srcu_read_lock(&kfd_processes_srcu);
 414
 415         /*
 416          * Look for the process that matches the pasid. If there is no such
 417          * process, we either released it in amdkfd's own notifier, or there
 418          * is a bug. Unfortunately, there is no way to tell...
 419          */
 420         hash_for_each_rcu(kfd_processes_table, i, p, kfd_processes)
 421                 if (p->pasid == pasid) {
 422
 423                         srcu_read_unlock(&kfd_processes_srcu, idx);
 424
 425                         pr_debug("Unbinding process %d from IOMMU\n", pasid);
 426
 427                         mutex_lock(&p->mutex);
 428
 429                         if ((dev->dbgmgr) && (dev->dbgmgr->pasid == p->pasid))
 430                                 kfd_dbgmgr_destroy(dev->dbgmgr);
 431
 432                         pqm_uninit(&p->pqm);
 433
 434                         pdd = kfd_get_process_device_data(dev, p);
 435
 436                         if (!pdd) {
 437                                 mutex_unlock(&p->mutex);
 438                                 return;
 439                         }
 440
 441                         if (pdd->reset_wavefronts) {
 442                                 dbgdev_wave_reset_wavefronts(pdd->dev, p);
 443                                 pdd->reset_wavefronts = false;
 444                         }
 445
 446                         /*
 447                          * Just mark pdd as unbound, because we still need it
 448                          * to call amd_iommu_unbind_pasid() in when the
 449                          * process exits.
 450                          * We don't call amd_iommu_unbind_pasid() here
 451                          * because the IOMMU called us.
 452                          */
 453                         pdd->bound = false;
 454
 455                         mutex_unlock(&p->mutex);
 456
 457                         return;
 458                 }
 459
 460         srcu_read_unlock(&kfd_processes_srcu, idx);
 461 }
 462
 463 struct kfd_process_device *kfd_get_first_process_device_data(struct kfd_process *p)
 464 {
 465         return list_first_entry(&p->per_device_data,
 466                                 struct kfd_process_device,
 467                                 per_device_list);
 468 }
 469
 470 struct kfd_process_device *kfd_get_next_process_device_data(struct kfd_process *p,
 471                                                 struct kfd_process_device *pdd)
 472 {
 473         if (list_is_last(&pdd->per_device_list, &p->per_device_data))
 474                 return NULL;
 475         return list_next_entry(pdd, per_device_list);
 476 }
 477
 478 bool kfd_has_process_device_data(struct kfd_process *p)
 479 {
 480         return !(list_empty(&p->per_device_data));
 481 }
 482
 483 /* This returns with process->mutex locked. */
 484 struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid)
 485 {
 486         struct kfd_process *p;
 487         unsigned int temp;
 488
 489         int idx = srcu_read_lock(&kfd_processes_srcu);
 490
 491         hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
 492                 if (p->pasid == pasid) {
 493                         mutex_lock(&p->mutex);
 494                         break;
 495                 }
 496         }
 497
 498         srcu_read_unlock(&kfd_processes_srcu, idx);
 499
 500         return p;
 501 }