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

   1 // SPDX-License-Identifier: GPL-2.0 OR MIT
   2 /*
   3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice shall be included in
  13  * all copies or substantial portions of the Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  21  * OTHER DEALINGS IN THE SOFTWARE.
  22  */
  23
  24 #include <linux/mutex.h>
  25 #include <linux/log2.h>
  26 #include <linux/sched.h>
  27 #include <linux/sched/mm.h>
  28 #include <linux/sched/task.h>
  29 #include <linux/mmu_context.h>
  30 #include <linux/slab.h>
  31 #include <linux/notifier.h>
  32 #include <linux/compat.h>
  33 #include <linux/mman.h>
  34 #include <linux/file.h>
  35 #include <linux/pm_runtime.h>
  36 #include "amdgpu_amdkfd.h"
  37 #include "amdgpu.h"
  38
  39 struct mm_struct;
  40
  41 #include "kfd_priv.h"
  42 #include "kfd_device_queue_manager.h"
  43 #include "kfd_svm.h"
  44 #include "kfd_smi_events.h"
  45 #include "kfd_debug.h"
  46
  47 /*
  48  * List of struct kfd_process (field kfd_process).
  49  * Unique/indexed by mm_struct*
  50  */
  51 DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
  52 DEFINE_MUTEX(kfd_processes_mutex);
  53
  54 DEFINE_SRCU(kfd_processes_srcu);
  55
  56 /* For process termination handling */
  57 static struct workqueue_struct *kfd_process_wq;
  58
  59 /* Ordered, single-threaded workqueue for restoring evicted
  60  * processes. Restoring multiple processes concurrently under memory
  61  * pressure can lead to processes blocking each other from validating
  62  * their BOs and result in a live-lock situation where processes
  63  * remain evicted indefinitely.
  64  */
  65 static struct workqueue_struct *kfd_restore_wq;
  66
  67 static struct kfd_process *find_process(const struct task_struct *thread,
  68                                         bool ref);
  69 static void kfd_process_ref_release(struct kref *ref);
  70 static struct kfd_process *create_process(const struct task_struct *thread);
  71
  72 static void evict_process_worker(struct work_struct *work);
  73 static void restore_process_worker(struct work_struct *work);
  74
  75 static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd);
  76
  77 struct kfd_procfs_tree {
  78         struct kobject *kobj;
  79 };
  80
  81 static struct kfd_procfs_tree procfs;
  82
  83 /*
  84  * Structure for SDMA activity tracking
  85  */
  86 struct kfd_sdma_activity_handler_workarea {
  87         struct work_struct sdma_activity_work;
  88         struct kfd_process_device *pdd;
  89         uint64_t sdma_activity_counter;
  90 };
  91
  92 struct temp_sdma_queue_list {
  93         uint64_t __user *rptr;
  94         uint64_t sdma_val;
  95         unsigned int queue_id;
  96         struct list_head list;
  97 };
  98
  99 static void kfd_sdma_activity_worker(struct work_struct *work)
 100 {
 101         struct kfd_sdma_activity_handler_workarea *workarea;
 102         struct kfd_process_device *pdd;
 103         uint64_t val;
 104         struct mm_struct *mm;
 105         struct queue *q;
 106         struct qcm_process_device *qpd;
 107         struct device_queue_manager *dqm;
 108         int ret = 0;
 109         struct temp_sdma_queue_list sdma_q_list;
 110         struct temp_sdma_queue_list *sdma_q, *next;
 111
 112         workarea = container_of(work, struct kfd_sdma_activity_handler_workarea,
 113                                 sdma_activity_work);
 114
 115         pdd = workarea->pdd;
 116         if (!pdd)
 117                 return;
 118         dqm = pdd->dev->dqm;
 119         qpd = &pdd->qpd;
 120         if (!dqm || !qpd)
 121                 return;
 122         /*
 123          * Total SDMA activity is current SDMA activity + past SDMA activity
 124          * Past SDMA count is stored in pdd.
 125          * To get the current activity counters for all active SDMA queues,
 126          * we loop over all SDMA queues and get their counts from user-space.
 127          *
 128          * We cannot call get_user() with dqm_lock held as it can cause
 129          * a circular lock dependency situation. To read the SDMA stats,
 130          * we need to do the following:
 131          *
 132          * 1. Create a temporary list of SDMA queue nodes from the qpd->queues_list,
 133          *    with dqm_lock/dqm_unlock().
 134          * 2. Call get_user() for each node in temporary list without dqm_lock.
 135          *    Save the SDMA count for each node and also add the count to the total
 136          *    SDMA count counter.
 137          *    Its possible, during this step, a few SDMA queue nodes got deleted
 138          *    from the qpd->queues_list.
 139          * 3. Do a second pass over qpd->queues_list to check if any nodes got deleted.
 140          *    If any node got deleted, its SDMA count would be captured in the sdma
 141          *    past activity counter. So subtract the SDMA counter stored in step 2
 142          *    for this node from the total SDMA count.
 143          */
 144         INIT_LIST_HEAD(&sdma_q_list.list);
 145
 146         /*
 147          * Create the temp list of all SDMA queues
 148          */
 149         dqm_lock(dqm);
 150
 151         list_for_each_entry(q, &qpd->queues_list, list) {
 152                 if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
 153                     (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
 154                         continue;
 155
 156                 sdma_q = kzalloc(sizeof(struct temp_sdma_queue_list), GFP_KERNEL);
 157                 if (!sdma_q) {
 158                         dqm_unlock(dqm);
 159                         goto cleanup;
 160                 }
 161
 162                 INIT_LIST_HEAD(&sdma_q->list);
 163                 sdma_q->rptr = (uint64_t __user *)q->properties.read_ptr;
 164                 sdma_q->queue_id = q->properties.queue_id;
 165                 list_add_tail(&sdma_q->list, &sdma_q_list.list);
 166         }
 167
 168         /*
 169          * If the temp list is empty, then no SDMA queues nodes were found in
 170          * qpd->queues_list. Return the past activity count as the total sdma
 171          * count
 172          */
 173         if (list_empty(&sdma_q_list.list)) {
 174                 workarea->sdma_activity_counter = pdd->sdma_past_activity_counter;
 175                 dqm_unlock(dqm);
 176                 return;
 177         }
 178
 179         dqm_unlock(dqm);
 180
 181         /*
 182          * Get the usage count for each SDMA queue in temp_list.
 183          */
 184         mm = get_task_mm(pdd->process->lead_thread);
 185         if (!mm)
 186                 goto cleanup;
 187
 188         kthread_use_mm(mm);
 189
 190         list_for_each_entry(sdma_q, &sdma_q_list.list, list) {
 191                 val = 0;
 192                 ret = read_sdma_queue_counter(sdma_q->rptr, &val);
 193                 if (ret) {
 194                         pr_debug("Failed to read SDMA queue active counter for queue id: %d",
 195                                  sdma_q->queue_id);
 196                 } else {
 197                         sdma_q->sdma_val = val;
 198                         workarea->sdma_activity_counter += val;
 199                 }
 200         }
 201
 202         kthread_unuse_mm(mm);
 203         mmput(mm);
 204
 205         /*
 206          * Do a second iteration over qpd_queues_list to check if any SDMA
 207          * nodes got deleted while fetching SDMA counter.
 208          */
 209         dqm_lock(dqm);
 210
 211         workarea->sdma_activity_counter += pdd->sdma_past_activity_counter;
 212
 213         list_for_each_entry(q, &qpd->queues_list, list) {
 214                 if (list_empty(&sdma_q_list.list))
 215                         break;
 216
 217                 if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
 218                     (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
 219                         continue;
 220
 221                 list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
 222                         if (((uint64_t __user *)q->properties.read_ptr == sdma_q->rptr) &&
 223                              (sdma_q->queue_id == q->properties.queue_id)) {
 224                                 list_del(&sdma_q->list);
 225                                 kfree(sdma_q);
 226                                 break;
 227                         }
 228                 }
 229         }
 230
 231         dqm_unlock(dqm);
 232
 233         /*
 234          * If temp list is not empty, it implies some queues got deleted
 235          * from qpd->queues_list during SDMA usage read. Subtract the SDMA
 236          * count for each node from the total SDMA count.
 237          */
 238         list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
 239                 workarea->sdma_activity_counter -= sdma_q->sdma_val;
 240                 list_del(&sdma_q->list);
 241                 kfree(sdma_q);
 242         }
 243
 244         return;
 245
 246 cleanup:
 247         list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
 248                 list_del(&sdma_q->list);
 249                 kfree(sdma_q);
 250         }
 251 }
 252
 253 /**
 254  * kfd_get_cu_occupancy - Collect number of waves in-flight on this device
 255  * by current process. Translates acquired wave count into number of compute units
 256  * that are occupied.
 257  *
 258  * @attr: Handle of attribute that allows reporting of wave count. The attribute
 259  * handle encapsulates GPU device it is associated with, thereby allowing collection
 260  * of waves in flight, etc
 261  * @buffer: Handle of user provided buffer updated with wave count
 262  *
 263  * Return: Number of bytes written to user buffer or an error value
 264  */
 265 static int kfd_get_cu_occupancy(struct attribute *attr, char *buffer)
 266 {
 267         int cu_cnt;
 268         int wave_cnt;
 269         int max_waves_per_cu;
 270         struct kfd_node *dev = NULL;
 271         struct kfd_process *proc = NULL;
 272         struct kfd_process_device *pdd = NULL;
 273
 274         pdd = container_of(attr, struct kfd_process_device, attr_cu_occupancy);
 275         dev = pdd->dev;
 276         if (dev->kfd2kgd->get_cu_occupancy == NULL)
 277                 return -EINVAL;
 278
 279         cu_cnt = 0;
 280         proc = pdd->process;
 281         if (pdd->qpd.queue_count == 0) {
 282                 pr_debug("Gpu-Id: %d has no active queues for process %d\n",
 283                          dev->id, proc->pasid);
 284                 return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt);
 285         }
 286
 287         /* Collect wave count from device if it supports */
 288         wave_cnt = 0;
 289         max_waves_per_cu = 0;
 290         dev->kfd2kgd->get_cu_occupancy(dev->adev, proc->pasid, &wave_cnt,
 291                         &max_waves_per_cu, 0);
 292
 293         /* Translate wave count to number of compute units */
 294         cu_cnt = (wave_cnt + (max_waves_per_cu - 1)) / max_waves_per_cu;
 295         return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt);
 296 }
 297
 298 static ssize_t kfd_procfs_show(struct kobject *kobj, struct attribute *attr,
 299                                char *buffer)
 300 {
 301         if (strcmp(attr->name, "pasid") == 0) {
 302                 struct kfd_process *p = container_of(attr, struct kfd_process,
 303                                                      attr_pasid);
 304
 305                 return snprintf(buffer, PAGE_SIZE, "%d\n", p->pasid);
 306         } else if (strncmp(attr->name, "vram_", 5) == 0) {
 307                 struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
 308                                                               attr_vram);
 309                 return snprintf(buffer, PAGE_SIZE, "%llu\n", READ_ONCE(pdd->vram_usage));
 310         } else if (strncmp(attr->name, "sdma_", 5) == 0) {
 311                 struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
 312                                                               attr_sdma);
 313                 struct kfd_sdma_activity_handler_workarea sdma_activity_work_handler;
 314
 315                 INIT_WORK(&sdma_activity_work_handler.sdma_activity_work,
 316                                         kfd_sdma_activity_worker);
 317
 318                 sdma_activity_work_handler.pdd = pdd;
 319                 sdma_activity_work_handler.sdma_activity_counter = 0;
 320
 321                 schedule_work(&sdma_activity_work_handler.sdma_activity_work);
 322
 323                 flush_work(&sdma_activity_work_handler.sdma_activity_work);
 324
 325                 return snprintf(buffer, PAGE_SIZE, "%llu\n",
 326                                 (sdma_activity_work_handler.sdma_activity_counter)/
 327                                  SDMA_ACTIVITY_DIVISOR);
 328         } else {
 329                 pr_err("Invalid attribute");
 330                 return -EINVAL;
 331         }
 332
 333         return 0;
 334 }
 335
 336 static void kfd_procfs_kobj_release(struct kobject *kobj)
 337 {
 338         kfree(kobj);
 339 }
 340
 341 static const struct sysfs_ops kfd_procfs_ops = {
 342         .show = kfd_procfs_show,
 343 };
 344
 345 static const struct kobj_type procfs_type = {
 346         .release = kfd_procfs_kobj_release,
 347         .sysfs_ops = &kfd_procfs_ops,
 348 };
 349
 350 void kfd_procfs_init(void)
 351 {
 352         int ret = 0;
 353
 354         procfs.kobj = kfd_alloc_struct(procfs.kobj);
 355         if (!procfs.kobj)
 356                 return;
 357
 358         ret = kobject_init_and_add(procfs.kobj, &procfs_type,
 359                                    &kfd_device->kobj, "proc");
 360         if (ret) {
 361                 pr_warn("Could not create procfs proc folder");
 362                 /* If we fail to create the procfs, clean up */
 363                 kfd_procfs_shutdown();
 364         }
 365 }
 366
 367 void kfd_procfs_shutdown(void)
 368 {
 369         if (procfs.kobj) {
 370                 kobject_del(procfs.kobj);
 371                 kobject_put(procfs.kobj);
 372                 procfs.kobj = NULL;
 373         }
 374 }
 375
 376 static ssize_t kfd_procfs_queue_show(struct kobject *kobj,
 377                                      struct attribute *attr, char *buffer)
 378 {
 379         struct queue *q = container_of(kobj, struct queue, kobj);
 380
 381         if (!strcmp(attr->name, "size"))
 382                 return snprintf(buffer, PAGE_SIZE, "%llu",
 383                                 q->properties.queue_size);
 384         else if (!strcmp(attr->name, "type"))
 385                 return snprintf(buffer, PAGE_SIZE, "%d", q->properties.type);
 386         else if (!strcmp(attr->name, "gpuid"))
 387                 return snprintf(buffer, PAGE_SIZE, "%u", q->device->id);
 388         else
 389                 pr_err("Invalid attribute");
 390
 391         return 0;
 392 }
 393
 394 static ssize_t kfd_procfs_stats_show(struct kobject *kobj,
 395                                      struct attribute *attr, char *buffer)
 396 {
 397         if (strcmp(attr->name, "evicted_ms") == 0) {
 398                 struct kfd_process_device *pdd = container_of(attr,
 399                                 struct kfd_process_device,
 400                                 attr_evict);
 401                 uint64_t evict_jiffies;
 402
 403                 evict_jiffies = atomic64_read(&pdd->evict_duration_counter);
 404
 405                 return snprintf(buffer,
 406                                 PAGE_SIZE,
 407                                 "%llu\n",
 408                                 jiffies64_to_msecs(evict_jiffies));
 409
 410         /* Sysfs handle that gets CU occupancy is per device */
 411         } else if (strcmp(attr->name, "cu_occupancy") == 0) {
 412                 return kfd_get_cu_occupancy(attr, buffer);
 413         } else {
 414                 pr_err("Invalid attribute");
 415         }
 416
 417         return 0;
 418 }
 419
 420 static ssize_t kfd_sysfs_counters_show(struct kobject *kobj,
 421                                        struct attribute *attr, char *buf)
 422 {
 423         struct kfd_process_device *pdd;
 424
 425         if (!strcmp(attr->name, "faults")) {
 426                 pdd = container_of(attr, struct kfd_process_device,
 427                                    attr_faults);
 428                 return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->faults));
 429         }
 430         if (!strcmp(attr->name, "page_in")) {
 431                 pdd = container_of(attr, struct kfd_process_device,
 432                                    attr_page_in);
 433                 return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->page_in));
 434         }
 435         if (!strcmp(attr->name, "page_out")) {
 436                 pdd = container_of(attr, struct kfd_process_device,
 437                                    attr_page_out);
 438                 return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->page_out));
 439         }
 440         return 0;
 441 }
 442
 443 static struct attribute attr_queue_size = {
 444         .name = "size",
 445         .mode = KFD_SYSFS_FILE_MODE
 446 };
 447
 448 static struct attribute attr_queue_type = {
 449         .name = "type",
 450         .mode = KFD_SYSFS_FILE_MODE
 451 };
 452
 453 static struct attribute attr_queue_gpuid = {
 454         .name = "gpuid",
 455         .mode = KFD_SYSFS_FILE_MODE
 456 };
 457
 458 static struct attribute *procfs_queue_attrs[] = {
 459         &attr_queue_size,
 460         &attr_queue_type,
 461         &attr_queue_gpuid,
 462         NULL
 463 };
 464 ATTRIBUTE_GROUPS(procfs_queue);
 465
 466 static const struct sysfs_ops procfs_queue_ops = {
 467         .show = kfd_procfs_queue_show,
 468 };
 469
 470 static const struct kobj_type procfs_queue_type = {
 471         .sysfs_ops = &procfs_queue_ops,
 472         .default_groups = procfs_queue_groups,
 473 };
 474
 475 static const struct sysfs_ops procfs_stats_ops = {
 476         .show = kfd_procfs_stats_show,
 477 };
 478
 479 static const struct kobj_type procfs_stats_type = {
 480         .sysfs_ops = &procfs_stats_ops,
 481         .release = kfd_procfs_kobj_release,
 482 };
 483
 484 static const struct sysfs_ops sysfs_counters_ops = {
 485         .show = kfd_sysfs_counters_show,
 486 };
 487
 488 static const struct kobj_type sysfs_counters_type = {
 489         .sysfs_ops = &sysfs_counters_ops,
 490         .release = kfd_procfs_kobj_release,
 491 };
 492
 493 int kfd_procfs_add_queue(struct queue *q)
 494 {
 495         struct kfd_process *proc;
 496         int ret;
 497
 498         if (!q || !q->process)
 499                 return -EINVAL;
 500         proc = q->process;
 501
 502         /* Create proc/<pid>/queues/<queue id> folder */
 503         if (!proc->kobj_queues)
 504                 return -EFAULT;
 505         ret = kobject_init_and_add(&q->kobj, &procfs_queue_type,
 506                         proc->kobj_queues, "%u", q->properties.queue_id);
 507         if (ret < 0) {
 508                 pr_warn("Creating proc/<pid>/queues/%u failed",
 509                         q->properties.queue_id);
 510                 kobject_put(&q->kobj);
 511                 return ret;
 512         }
 513
 514         return 0;
 515 }
 516
 517 static void kfd_sysfs_create_file(struct kobject *kobj, struct attribute *attr,
 518                                  char *name)
 519 {
 520         int ret;
 521
 522         if (!kobj || !attr || !name)
 523                 return;
 524
 525         attr->name = name;
 526         attr->mode = KFD_SYSFS_FILE_MODE;
 527         sysfs_attr_init(attr);
 528
 529         ret = sysfs_create_file(kobj, attr);
 530         if (ret)
 531                 pr_warn("Create sysfs %s/%s failed %d", kobj->name, name, ret);
 532 }
 533
 534 static void kfd_procfs_add_sysfs_stats(struct kfd_process *p)
 535 {
 536         int ret;
 537         int i;
 538         char stats_dir_filename[MAX_SYSFS_FILENAME_LEN];
 539
 540         if (!p || !p->kobj)
 541                 return;
 542
 543         /*
 544          * Create sysfs files for each GPU:
 545          * - proc/<pid>/stats_<gpuid>/
 546          * - proc/<pid>/stats_<gpuid>/evicted_ms
 547          * - proc/<pid>/stats_<gpuid>/cu_occupancy
 548          */
 549         for (i = 0; i < p->n_pdds; i++) {
 550                 struct kfd_process_device *pdd = p->pdds[i];
 551
 552                 snprintf(stats_dir_filename, MAX_SYSFS_FILENAME_LEN,
 553                                 "stats_%u", pdd->dev->id);
 554                 pdd->kobj_stats = kfd_alloc_struct(pdd->kobj_stats);
 555                 if (!pdd->kobj_stats)
 556                         return;
 557
 558                 ret = kobject_init_and_add(pdd->kobj_stats,
 559                                            &procfs_stats_type,
 560                                            p->kobj,
 561                                            stats_dir_filename);
 562
 563                 if (ret) {
 564                         pr_warn("Creating KFD proc/stats_%s folder failed",
 565                                 stats_dir_filename);
 566                         kobject_put(pdd->kobj_stats);
 567                         pdd->kobj_stats = NULL;
 568                         return;
 569                 }
 570
 571                 kfd_sysfs_create_file(pdd->kobj_stats, &pdd->attr_evict,
 572                                       "evicted_ms");
 573                 /* Add sysfs file to report compute unit occupancy */
 574                 if (pdd->dev->kfd2kgd->get_cu_occupancy)
 575                         kfd_sysfs_create_file(pdd->kobj_stats,
 576                                               &pdd->attr_cu_occupancy,
 577                                               "cu_occupancy");
 578         }
 579 }
 580
 581 static void kfd_procfs_add_sysfs_counters(struct kfd_process *p)
 582 {
 583         int ret = 0;
 584         int i;
 585         char counters_dir_filename[MAX_SYSFS_FILENAME_LEN];
 586
 587         if (!p || !p->kobj)
 588                 return;
 589
 590         /*
 591          * Create sysfs files for each GPU which supports SVM
 592          * - proc/<pid>/counters_<gpuid>/
 593          * - proc/<pid>/counters_<gpuid>/faults
 594          * - proc/<pid>/counters_<gpuid>/page_in
 595          * - proc/<pid>/counters_<gpuid>/page_out
 596          */
 597         for_each_set_bit(i, p->svms.bitmap_supported, p->n_pdds) {
 598                 struct kfd_process_device *pdd = p->pdds[i];
 599                 struct kobject *kobj_counters;
 600
 601                 snprintf(counters_dir_filename, MAX_SYSFS_FILENAME_LEN,
 602                         "counters_%u", pdd->dev->id);
 603                 kobj_counters = kfd_alloc_struct(kobj_counters);
 604                 if (!kobj_counters)
 605                         return;
 606
 607                 ret = kobject_init_and_add(kobj_counters, &sysfs_counters_type,
 608                                            p->kobj, counters_dir_filename);
 609                 if (ret) {
 610                         pr_warn("Creating KFD proc/%s folder failed",
 611                                 counters_dir_filename);
 612                         kobject_put(kobj_counters);
 613                         return;
 614                 }
 615
 616                 pdd->kobj_counters = kobj_counters;
 617                 kfd_sysfs_create_file(kobj_counters, &pdd->attr_faults,
 618                                       "faults");
 619                 kfd_sysfs_create_file(kobj_counters, &pdd->attr_page_in,
 620                                       "page_in");
 621                 kfd_sysfs_create_file(kobj_counters, &pdd->attr_page_out,
 622                                       "page_out");
 623         }
 624 }
 625
 626 static void kfd_procfs_add_sysfs_files(struct kfd_process *p)
 627 {
 628         int i;
 629
 630         if (!p || !p->kobj)
 631                 return;
 632
 633         /*
 634          * Create sysfs files for each GPU:
 635          * - proc/<pid>/vram_<gpuid>
 636          * - proc/<pid>/sdma_<gpuid>
 637          */
 638         for (i = 0; i < p->n_pdds; i++) {
 639                 struct kfd_process_device *pdd = p->pdds[i];
 640
 641                 snprintf(pdd->vram_filename, MAX_SYSFS_FILENAME_LEN, "vram_%u",
 642                          pdd->dev->id);
 643                 kfd_sysfs_create_file(p->kobj, &pdd->attr_vram,
 644                                       pdd->vram_filename);
 645
 646                 snprintf(pdd->sdma_filename, MAX_SYSFS_FILENAME_LEN, "sdma_%u",
 647                          pdd->dev->id);
 648                 kfd_sysfs_create_file(p->kobj, &pdd->attr_sdma,
 649                                             pdd->sdma_filename);
 650         }
 651 }
 652
 653 void kfd_procfs_del_queue(struct queue *q)
 654 {
 655         if (!q)
 656                 return;
 657
 658         kobject_del(&q->kobj);
 659         kobject_put(&q->kobj);
 660 }
 661
 662 int kfd_process_create_wq(void)
 663 {
 664         if (!kfd_process_wq)
 665                 kfd_process_wq = alloc_workqueue("kfd_process_wq", 0, 0);
 666         if (!kfd_restore_wq)
 667                 kfd_restore_wq = alloc_ordered_workqueue("kfd_restore_wq",
 668                                                          WQ_FREEZABLE);
 669
 670         if (!kfd_process_wq || !kfd_restore_wq) {
 671                 kfd_process_destroy_wq();
 672                 return -ENOMEM;
 673         }
 674
 675         return 0;
 676 }
 677
 678 void kfd_process_destroy_wq(void)
 679 {
 680         if (kfd_process_wq) {
 681                 destroy_workqueue(kfd_process_wq);
 682                 kfd_process_wq = NULL;
 683         }
 684         if (kfd_restore_wq) {
 685                 destroy_workqueue(kfd_restore_wq);
 686                 kfd_restore_wq = NULL;
 687         }
 688 }
 689
 690 static void kfd_process_free_gpuvm(struct kgd_mem *mem,
 691                         struct kfd_process_device *pdd, void **kptr)
 692 {
 693         struct kfd_node *dev = pdd->dev;
 694
 695         if (kptr && *kptr) {
 696                 amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem);
 697                 *kptr = NULL;
 698         }
 699
 700         amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(dev->adev, mem, pdd->drm_priv);
 701         amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, mem, pdd->drm_priv,
 702                                                NULL);
 703 }
 704
 705 /* kfd_process_alloc_gpuvm - Allocate GPU VM for the KFD process
 706  *      This function should be only called right after the process
 707  *      is created and when kfd_processes_mutex is still being held
 708  *      to avoid concurrency. Because of that exclusiveness, we do
 709  *      not need to take p->mutex.
 710  */
 711 static int kfd_process_alloc_gpuvm(struct kfd_process_device *pdd,
 712                                    uint64_t gpu_va, uint32_t size,
 713                                    uint32_t flags, struct kgd_mem **mem, void **kptr)
 714 {
 715         struct kfd_node *kdev = pdd->dev;
 716         int err;
 717
 718         err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(kdev->adev, gpu_va, size,
 719                                                  pdd->drm_priv, mem, NULL,
 720                                                  flags, false);
 721         if (err)
 722                 goto err_alloc_mem;
 723
 724         err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(kdev->adev, *mem,
 725                         pdd->drm_priv);
 726         if (err)
 727                 goto err_map_mem;
 728
 729         err = amdgpu_amdkfd_gpuvm_sync_memory(kdev->adev, *mem, true);
 730         if (err) {
 731                 pr_debug("Sync memory failed, wait interrupted by user signal\n");
 732                 goto sync_memory_failed;
 733         }
 734
 735         if (kptr) {
 736                 err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(
 737                                 (struct kgd_mem *)*mem, kptr, NULL);
 738                 if (err) {
 739                         pr_debug("Map GTT BO to kernel failed\n");
 740                         goto sync_memory_failed;
 741                 }
 742         }
 743
 744         return err;
 745
 746 sync_memory_failed:
 747         amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(kdev->adev, *mem, pdd->drm_priv);
 748
 749 err_map_mem:
 750         amdgpu_amdkfd_gpuvm_free_memory_of_gpu(kdev->adev, *mem, pdd->drm_priv,
 751                                                NULL);
 752 err_alloc_mem:
 753         *mem = NULL;
 754         *kptr = NULL;
 755         return err;
 756 }
 757
 758 /* kfd_process_device_reserve_ib_mem - Reserve memory inside the
 759  *      process for IB usage The memory reserved is for KFD to submit
 760  *      IB to AMDGPU from kernel.  If the memory is reserved
 761  *      successfully, ib_kaddr will have the CPU/kernel
 762  *      address. Check ib_kaddr before accessing the memory.
 763  */
 764 static int kfd_process_device_reserve_ib_mem(struct kfd_process_device *pdd)
 765 {
 766         struct qcm_process_device *qpd = &pdd->qpd;
 767         uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT |
 768                         KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE |
 769                         KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE |
 770                         KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
 771         struct kgd_mem *mem;
 772         void *kaddr;
 773         int ret;
 774
 775         if (qpd->ib_kaddr || !qpd->ib_base)
 776                 return 0;
 777
 778         /* ib_base is only set for dGPU */
 779         ret = kfd_process_alloc_gpuvm(pdd, qpd->ib_base, PAGE_SIZE, flags,
 780                                       &mem, &kaddr);
 781         if (ret)
 782                 return ret;
 783
 784         qpd->ib_mem = mem;
 785         qpd->ib_kaddr = kaddr;
 786
 787         return 0;
 788 }
 789
 790 static void kfd_process_device_destroy_ib_mem(struct kfd_process_device *pdd)
 791 {
 792         struct qcm_process_device *qpd = &pdd->qpd;
 793
 794         if (!qpd->ib_kaddr || !qpd->ib_base)
 795                 return;
 796
 797         kfd_process_free_gpuvm(qpd->ib_mem, pdd, &qpd->ib_kaddr);
 798 }
 799
 800 struct kfd_process *kfd_create_process(struct task_struct *thread)
 801 {
 802         struct kfd_process *process;
 803         int ret;
 804
 805         if (!(thread->mm && mmget_not_zero(thread->mm)))
 806                 return ERR_PTR(-EINVAL);
 807
 808         /* Only the pthreads threading model is supported. */
 809         if (thread->group_leader->mm != thread->mm) {
 810                 mmput(thread->mm);
 811                 return ERR_PTR(-EINVAL);
 812         }
 813
 814         /*
 815          * take kfd processes mutex before starting of process creation
 816          * so there won't be a case where two threads of the same process
 817          * create two kfd_process structures
 818          */
 819         mutex_lock(&kfd_processes_mutex);
 820
 821         if (kfd_is_locked()) {
 822                 mutex_unlock(&kfd_processes_mutex);
 823                 pr_debug("KFD is locked! Cannot create process");
 824                 return ERR_PTR(-EINVAL);
 825         }
 826
 827         /* A prior open of /dev/kfd could have already created the process. */
 828         process = find_process(thread, false);
 829         if (process) {
 830                 pr_debug("Process already found\n");
 831         } else {
 832                 process = create_process(thread);
 833                 if (IS_ERR(process))
 834                         goto out;
 835
 836                 if (!procfs.kobj)
 837                         goto out;
 838
 839                 process->kobj = kfd_alloc_struct(process->kobj);
 840                 if (!process->kobj) {
 841                         pr_warn("Creating procfs kobject failed");
 842                         goto out;
 843                 }
 844                 ret = kobject_init_and_add(process->kobj, &procfs_type,
 845                                            procfs.kobj, "%d",
 846                                            (int)process->lead_thread->pid);
 847                 if (ret) {
 848                         pr_warn("Creating procfs pid directory failed");
 849                         kobject_put(process->kobj);
 850                         goto out;
 851                 }
 852
 853                 kfd_sysfs_create_file(process->kobj, &process->attr_pasid,
 854                                       "pasid");
 855
 856                 process->kobj_queues = kobject_create_and_add("queues",
 857                                                         process->kobj);
 858                 if (!process->kobj_queues)
 859                         pr_warn("Creating KFD proc/queues folder failed");
 860
 861                 kfd_procfs_add_sysfs_stats(process);
 862                 kfd_procfs_add_sysfs_files(process);
 863                 kfd_procfs_add_sysfs_counters(process);
 864
 865                 init_waitqueue_head(&process->wait_irq_drain);
 866         }
 867 out:
 868         if (!IS_ERR(process))
 869                 kref_get(&process->ref);
 870         mutex_unlock(&kfd_processes_mutex);
 871         mmput(thread->mm);
 872
 873         return process;
 874 }
 875
 876 struct kfd_process *kfd_get_process(const struct task_struct *thread)
 877 {
 878         struct kfd_process *process;
 879
 880         if (!thread->mm)
 881                 return ERR_PTR(-EINVAL);
 882
 883         /* Only the pthreads threading model is supported. */
 884         if (thread->group_leader->mm != thread->mm)
 885                 return ERR_PTR(-EINVAL);
 886
 887         process = find_process(thread, false);
 888         if (!process)
 889                 return ERR_PTR(-EINVAL);
 890
 891         return process;
 892 }
 893
 894 static struct kfd_process *find_process_by_mm(const struct mm_struct *mm)
 895 {
 896         struct kfd_process *process;
 897
 898         hash_for_each_possible_rcu(kfd_processes_table, process,
 899                                         kfd_processes, (uintptr_t)mm)
 900                 if (process->mm == mm)
 901                         return process;
 902
 903         return NULL;
 904 }
 905
 906 static struct kfd_process *find_process(const struct task_struct *thread,
 907                                         bool ref)
 908 {
 909         struct kfd_process *p;
 910         int idx;
 911
 912         idx = srcu_read_lock(&kfd_processes_srcu);
 913         p = find_process_by_mm(thread->mm);
 914         if (p && ref)
 915                 kref_get(&p->ref);
 916         srcu_read_unlock(&kfd_processes_srcu, idx);
 917
 918         return p;
 919 }
 920
 921 void kfd_unref_process(struct kfd_process *p)
 922 {
 923         kref_put(&p->ref, kfd_process_ref_release);
 924 }
 925
 926 /* This increments the process->ref counter. */
 927 struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid)
 928 {
 929         struct task_struct *task = NULL;
 930         struct kfd_process *p    = NULL;
 931
 932         if (!pid) {
 933                 task = current;
 934                 get_task_struct(task);
 935         } else {
 936                 task = get_pid_task(pid, PIDTYPE_PID);
 937         }
 938
 939         if (task) {
 940                 p = find_process(task, true);
 941                 put_task_struct(task);
 942         }
 943
 944         return p;
 945 }
 946
 947 static void kfd_process_device_free_bos(struct kfd_process_device *pdd)
 948 {
 949         struct kfd_process *p = pdd->process;
 950         void *mem;
 951         int id;
 952         int i;
 953
 954         /*
 955          * Remove all handles from idr and release appropriate
 956          * local memory object
 957          */
 958         idr_for_each_entry(&pdd->alloc_idr, mem, id) {
 959
 960                 for (i = 0; i < p->n_pdds; i++) {
 961                         struct kfd_process_device *peer_pdd = p->pdds[i];
 962
 963                         if (!peer_pdd->drm_priv)
 964                                 continue;
 965                         amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
 966                                 peer_pdd->dev->adev, mem, peer_pdd->drm_priv);
 967                 }
 968
 969                 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, mem,
 970                                                        pdd->drm_priv, NULL);
 971                 kfd_process_device_remove_obj_handle(pdd, id);
 972         }
 973 }
 974
 975 /*
 976  * Just kunmap and unpin signal BO here. It will be freed in
 977  * kfd_process_free_outstanding_kfd_bos()
 978  */
 979 static void kfd_process_kunmap_signal_bo(struct kfd_process *p)
 980 {
 981         struct kfd_process_device *pdd;
 982         struct kfd_node *kdev;
 983         void *mem;
 984
 985         kdev = kfd_device_by_id(GET_GPU_ID(p->signal_handle));
 986         if (!kdev)
 987                 return;
 988
 989         mutex_lock(&p->mutex);
 990
 991         pdd = kfd_get_process_device_data(kdev, p);
 992         if (!pdd)
 993                 goto out;
 994
 995         mem = kfd_process_device_translate_handle(
 996                 pdd, GET_IDR_HANDLE(p->signal_handle));
 997         if (!mem)
 998                 goto out;
 999
1000         amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem);
1001
1002 out:
1003         mutex_unlock(&p->mutex);
1004 }
1005
1006 static void kfd_process_free_outstanding_kfd_bos(struct kfd_process *p)
1007 {
1008         int i;
1009
1010         for (i = 0; i < p->n_pdds; i++)
1011                 kfd_process_device_free_bos(p->pdds[i]);
1012 }
1013
1014 static void kfd_process_destroy_pdds(struct kfd_process *p)
1015 {
1016         int i;
1017
1018         for (i = 0; i < p->n_pdds; i++) {
1019                 struct kfd_process_device *pdd = p->pdds[i];
1020
1021                 pr_debug("Releasing pdd (topology id %d) for process (pasid 0x%x)\n",
1022                                 pdd->dev->id, p->pasid);
1023
1024                 kfd_process_device_destroy_cwsr_dgpu(pdd);
1025                 kfd_process_device_destroy_ib_mem(pdd);
1026
1027                 if (pdd->drm_file) {
1028                         amdgpu_amdkfd_gpuvm_release_process_vm(
1029                                         pdd->dev->adev, pdd->drm_priv);
1030                         fput(pdd->drm_file);
1031                 }
1032
1033                 if (pdd->qpd.cwsr_kaddr && !pdd->qpd.cwsr_base)
1034                         free_pages((unsigned long)pdd->qpd.cwsr_kaddr,
1035                                 get_order(KFD_CWSR_TBA_TMA_SIZE));
1036
1037                 idr_destroy(&pdd->alloc_idr);
1038
1039                 kfd_free_process_doorbells(pdd->dev->kfd, pdd);
1040
1041                 if (pdd->dev->kfd->shared_resources.enable_mes)
1042                         amdgpu_amdkfd_free_gtt_mem(pdd->dev->adev,
1043                                                    pdd->proc_ctx_bo);
1044                 /*
1045                  * before destroying pdd, make sure to report availability
1046                  * for auto suspend
1047                  */
1048                 if (pdd->runtime_inuse) {
1049                         pm_runtime_mark_last_busy(adev_to_drm(pdd->dev->adev)->dev);
1050                         pm_runtime_put_autosuspend(adev_to_drm(pdd->dev->adev)->dev);
1051                         pdd->runtime_inuse = false;
1052                 }
1053
1054                 kfree(pdd);
1055                 p->pdds[i] = NULL;
1056         }
1057         p->n_pdds = 0;
1058 }
1059
1060 static void kfd_process_remove_sysfs(struct kfd_process *p)
1061 {
1062         struct kfd_process_device *pdd;
1063         int i;
1064
1065         if (!p->kobj)
1066                 return;
1067
1068         sysfs_remove_file(p->kobj, &p->attr_pasid);
1069         kobject_del(p->kobj_queues);
1070         kobject_put(p->kobj_queues);
1071         p->kobj_queues = NULL;
1072
1073         for (i = 0; i < p->n_pdds; i++) {
1074                 pdd = p->pdds[i];
1075
1076                 sysfs_remove_file(p->kobj, &pdd->attr_vram);
1077                 sysfs_remove_file(p->kobj, &pdd->attr_sdma);
1078
1079                 sysfs_remove_file(pdd->kobj_stats, &pdd->attr_evict);
1080                 if (pdd->dev->kfd2kgd->get_cu_occupancy)
1081                         sysfs_remove_file(pdd->kobj_stats,
1082                                           &pdd->attr_cu_occupancy);
1083                 kobject_del(pdd->kobj_stats);
1084                 kobject_put(pdd->kobj_stats);
1085                 pdd->kobj_stats = NULL;
1086         }
1087
1088         for_each_set_bit(i, p->svms.bitmap_supported, p->n_pdds) {
1089                 pdd = p->pdds[i];
1090
1091                 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_faults);
1092                 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_page_in);
1093                 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_page_out);
1094                 kobject_del(pdd->kobj_counters);
1095                 kobject_put(pdd->kobj_counters);
1096                 pdd->kobj_counters = NULL;
1097         }
1098
1099         kobject_del(p->kobj);
1100         kobject_put(p->kobj);
1101         p->kobj = NULL;
1102 }
1103
1104 /* No process locking is needed in this function, because the process
1105  * is not findable any more. We must assume that no other thread is
1106  * using it any more, otherwise we couldn't safely free the process
1107  * structure in the end.
1108  */
1109 static void kfd_process_wq_release(struct work_struct *work)
1110 {
1111         struct kfd_process *p = container_of(work, struct kfd_process,
1112                                              release_work);
1113
1114         kfd_process_dequeue_from_all_devices(p);
1115         pqm_uninit(&p->pqm);
1116
1117         /* Signal the eviction fence after user mode queues are
1118          * destroyed. This allows any BOs to be freed without
1119          * triggering pointless evictions or waiting for fences.
1120          */
1121         dma_fence_signal(p->ef);
1122
1123         kfd_process_remove_sysfs(p);
1124
1125         kfd_process_kunmap_signal_bo(p);
1126         kfd_process_free_outstanding_kfd_bos(p);
1127         svm_range_list_fini(p);
1128
1129         kfd_process_destroy_pdds(p);
1130         dma_fence_put(p->ef);
1131
1132         kfd_event_free_process(p);
1133
1134         kfd_pasid_free(p->pasid);
1135         mutex_destroy(&p->mutex);
1136
1137         put_task_struct(p->lead_thread);
1138
1139         kfree(p);
1140 }
1141
1142 static void kfd_process_ref_release(struct kref *ref)
1143 {
1144         struct kfd_process *p = container_of(ref, struct kfd_process, ref);
1145
1146         INIT_WORK(&p->release_work, kfd_process_wq_release);
1147         queue_work(kfd_process_wq, &p->release_work);
1148 }
1149
1150 static struct mmu_notifier *kfd_process_alloc_notifier(struct mm_struct *mm)
1151 {
1152         int idx = srcu_read_lock(&kfd_processes_srcu);
1153         struct kfd_process *p = find_process_by_mm(mm);
1154
1155         srcu_read_unlock(&kfd_processes_srcu, idx);
1156
1157         return p ? &p->mmu_notifier : ERR_PTR(-ESRCH);
1158 }
1159
1160 static void kfd_process_free_notifier(struct mmu_notifier *mn)
1161 {
1162         kfd_unref_process(container_of(mn, struct kfd_process, mmu_notifier));
1163 }
1164
1165 static void kfd_process_notifier_release_internal(struct kfd_process *p)
1166 {
1167         int i;
1168
1169         cancel_delayed_work_sync(&p->eviction_work);
1170         cancel_delayed_work_sync(&p->restore_work);
1171
1172         for (i = 0; i < p->n_pdds; i++) {
1173                 struct kfd_process_device *pdd = p->pdds[i];
1174
1175                 /* re-enable GFX OFF since runtime enable with ttmp setup disabled it. */
1176                 if (!kfd_dbg_is_rlc_restore_supported(pdd->dev) && p->runtime_info.ttmp_setup)
1177                         amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
1178         }
1179
1180         /* Indicate to other users that MM is no longer valid */
1181         p->mm = NULL;
1182         kfd_dbg_trap_disable(p);
1183
1184         if (atomic_read(&p->debugged_process_count) > 0) {
1185                 struct kfd_process *target;
1186                 unsigned int temp;
1187                 int idx = srcu_read_lock(&kfd_processes_srcu);
1188
1189                 hash_for_each_rcu(kfd_processes_table, temp, target, kfd_processes) {
1190                         if (target->debugger_process && target->debugger_process == p) {
1191                                 mutex_lock_nested(&target->mutex, 1);
1192                                 kfd_dbg_trap_disable(target);
1193                                 mutex_unlock(&target->mutex);
1194                                 if (atomic_read(&p->debugged_process_count) == 0)
1195                                         break;
1196                         }
1197                 }
1198
1199                 srcu_read_unlock(&kfd_processes_srcu, idx);
1200         }
1201
1202         mmu_notifier_put(&p->mmu_notifier);
1203 }
1204
1205 static void kfd_process_notifier_release(struct mmu_notifier *mn,
1206                                         struct mm_struct *mm)
1207 {
1208         struct kfd_process *p;
1209
1210         /*
1211          * The kfd_process structure can not be free because the
1212          * mmu_notifier srcu is read locked
1213          */
1214         p = container_of(mn, struct kfd_process, mmu_notifier);
1215         if (WARN_ON(p->mm != mm))
1216                 return;
1217
1218         mutex_lock(&kfd_processes_mutex);
1219         /*
1220          * Do early return if table is empty.
1221          *
1222          * This could potentially happen if this function is called concurrently
1223          * by mmu_notifier and by kfd_cleanup_pocesses.
1224          *
1225          */
1226         if (hash_empty(kfd_processes_table)) {
1227                 mutex_unlock(&kfd_processes_mutex);
1228                 return;
1229         }
1230         hash_del_rcu(&p->kfd_processes);
1231         mutex_unlock(&kfd_processes_mutex);
1232         synchronize_srcu(&kfd_processes_srcu);
1233
1234         kfd_process_notifier_release_internal(p);
1235 }
1236
1237 static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = {
1238         .release = kfd_process_notifier_release,
1239         .alloc_notifier = kfd_process_alloc_notifier,
1240         .free_notifier = kfd_process_free_notifier,
1241 };
1242
1243 /*
1244  * This code handles the case when driver is being unloaded before all
1245  * mm_struct are released.  We need to safely free the kfd_process and
1246  * avoid race conditions with mmu_notifier that might try to free them.
1247  *
1248  */
1249 void kfd_cleanup_processes(void)
1250 {
1251         struct kfd_process *p;
1252         struct hlist_node *p_temp;
1253         unsigned int temp;
1254         HLIST_HEAD(cleanup_list);
1255
1256         /*
1257          * Move all remaining kfd_process from the process table to a
1258          * temp list for processing.   Once done, callback from mmu_notifier
1259          * release will not see the kfd_process in the table and do early return,
1260          * avoiding double free issues.
1261          */
1262         mutex_lock(&kfd_processes_mutex);
1263         hash_for_each_safe(kfd_processes_table, temp, p_temp, p, kfd_processes) {
1264                 hash_del_rcu(&p->kfd_processes);
1265                 synchronize_srcu(&kfd_processes_srcu);
1266                 hlist_add_head(&p->kfd_processes, &cleanup_list);
1267         }
1268         mutex_unlock(&kfd_processes_mutex);
1269
1270         hlist_for_each_entry_safe(p, p_temp, &cleanup_list, kfd_processes)
1271                 kfd_process_notifier_release_internal(p);
1272
1273         /*
1274          * Ensures that all outstanding free_notifier get called, triggering
1275          * the release of the kfd_process struct.
1276          */
1277         mmu_notifier_synchronize();
1278 }
1279
1280 int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep)
1281 {
1282         unsigned long  offset;
1283         int i;
1284
1285         if (p->has_cwsr)
1286                 return 0;
1287
1288         for (i = 0; i < p->n_pdds; i++) {
1289                 struct kfd_node *dev = p->pdds[i]->dev;
1290                 struct qcm_process_device *qpd = &p->pdds[i]->qpd;
1291
1292                 if (!dev->kfd->cwsr_enabled || qpd->cwsr_kaddr || qpd->cwsr_base)
1293                         continue;
1294
1295                 offset = KFD_MMAP_TYPE_RESERVED_MEM | KFD_MMAP_GPU_ID(dev->id);
1296                 qpd->tba_addr = (int64_t)vm_mmap(filep, 0,
1297                         KFD_CWSR_TBA_TMA_SIZE, PROT_READ | PROT_EXEC,
1298                         MAP_SHARED, offset);
1299
1300                 if (IS_ERR_VALUE(qpd->tba_addr)) {
1301                         int err = qpd->tba_addr;
1302
1303                         pr_err("Failure to set tba address. error %d.\n", err);
1304                         qpd->tba_addr = 0;
1305                         qpd->cwsr_kaddr = NULL;
1306                         return err;
1307                 }
1308
1309                 memcpy(qpd->cwsr_kaddr, dev->kfd->cwsr_isa, dev->kfd->cwsr_isa_size);
1310
1311                 kfd_process_set_trap_debug_flag(qpd, p->debug_trap_enabled);
1312
1313                 qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1314                 pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1315                         qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1316         }
1317
1318         p->has_cwsr = true;
1319
1320         return 0;
1321 }
1322
1323 static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd)
1324 {
1325         struct kfd_node *dev = pdd->dev;
1326         struct qcm_process_device *qpd = &pdd->qpd;
1327         uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT
1328                         | KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE
1329                         | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
1330         struct kgd_mem *mem;
1331         void *kaddr;
1332         int ret;
1333
1334         if (!dev->kfd->cwsr_enabled || qpd->cwsr_kaddr || !qpd->cwsr_base)
1335                 return 0;
1336
1337         /* cwsr_base is only set for dGPU */
1338         ret = kfd_process_alloc_gpuvm(pdd, qpd->cwsr_base,
1339                                       KFD_CWSR_TBA_TMA_SIZE, flags, &mem, &kaddr);
1340         if (ret)
1341                 return ret;
1342
1343         qpd->cwsr_mem = mem;
1344         qpd->cwsr_kaddr = kaddr;
1345         qpd->tba_addr = qpd->cwsr_base;
1346
1347         memcpy(qpd->cwsr_kaddr, dev->kfd->cwsr_isa, dev->kfd->cwsr_isa_size);
1348
1349         kfd_process_set_trap_debug_flag(&pdd->qpd,
1350                                         pdd->process->debug_trap_enabled);
1351
1352         qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1353         pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1354                  qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1355
1356         return 0;
1357 }
1358
1359 static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd)
1360 {
1361         struct kfd_node *dev = pdd->dev;
1362         struct qcm_process_device *qpd = &pdd->qpd;
1363
1364         if (!dev->kfd->cwsr_enabled || !qpd->cwsr_kaddr || !qpd->cwsr_base)
1365                 return;
1366
1367         kfd_process_free_gpuvm(qpd->cwsr_mem, pdd, &qpd->cwsr_kaddr);
1368 }
1369
1370 void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
1371                                   uint64_t tba_addr,
1372                                   uint64_t tma_addr)
1373 {
1374         if (qpd->cwsr_kaddr) {
1375                 /* KFD trap handler is bound, record as second-level TBA/TMA
1376                  * in first-level TMA. First-level trap will jump to second.
1377                  */
1378                 uint64_t *tma =
1379                         (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1380                 tma[0] = tba_addr;
1381                 tma[1] = tma_addr;
1382         } else {
1383                 /* No trap handler bound, bind as first-level TBA/TMA. */
1384                 qpd->tba_addr = tba_addr;
1385                 qpd->tma_addr = tma_addr;
1386         }
1387 }
1388
1389 bool kfd_process_xnack_mode(struct kfd_process *p, bool supported)
1390 {
1391         int i;
1392
1393         /* On most GFXv9 GPUs, the retry mode in the SQ must match the
1394          * boot time retry setting. Mixing processes with different
1395          * XNACK/retry settings can hang the GPU.
1396          *
1397          * Different GPUs can have different noretry settings depending
1398          * on HW bugs or limitations. We need to find at least one
1399          * XNACK mode for this process that's compatible with all GPUs.
1400          * Fortunately GPUs with retry enabled (noretry=0) can run code
1401          * built for XNACK-off. On GFXv9 it may perform slower.
1402          *
1403          * Therefore applications built for XNACK-off can always be
1404          * supported and will be our fallback if any GPU does not
1405          * support retry.
1406          */
1407         for (i = 0; i < p->n_pdds; i++) {
1408                 struct kfd_node *dev = p->pdds[i]->dev;
1409
1410                 /* Only consider GFXv9 and higher GPUs. Older GPUs don't
1411                  * support the SVM APIs and don't need to be considered
1412                  * for the XNACK mode selection.
1413                  */
1414                 if (!KFD_IS_SOC15(dev))
1415                         continue;
1416                 /* Aldebaran can always support XNACK because it can support
1417                  * per-process XNACK mode selection. But let the dev->noretry
1418                  * setting still influence the default XNACK mode.
1419                  */
1420                 if (supported && KFD_SUPPORT_XNACK_PER_PROCESS(dev)) {
1421                         if (!amdgpu_sriov_xnack_support(dev->kfd->adev)) {
1422                                 pr_debug("SRIOV platform xnack not supported\n");
1423                                 return false;
1424                         }
1425                         continue;
1426                 }
1427
1428                 /* GFXv10 and later GPUs do not support shader preemption
1429                  * during page faults. This can lead to poor QoS for queue
1430                  * management and memory-manager-related preemptions or
1431                  * even deadlocks.
1432                  */
1433                 if (KFD_GC_VERSION(dev) >= IP_VERSION(10, 1, 1))
1434                         return false;
1435
1436                 if (dev->kfd->noretry)
1437                         return false;
1438         }
1439
1440         return true;
1441 }
1442
1443 void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd,
1444                                      bool enabled)
1445 {
1446         if (qpd->cwsr_kaddr) {
1447                 uint64_t *tma =
1448                         (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1449                 tma[2] = enabled;
1450         }
1451 }
1452
1453 /*
1454  * On return the kfd_process is fully operational and will be freed when the
1455  * mm is released
1456  */
1457 static struct kfd_process *create_process(const struct task_struct *thread)
1458 {
1459         struct kfd_process *process;
1460         struct mmu_notifier *mn;
1461         int err = -ENOMEM;
1462
1463         process = kzalloc(sizeof(*process), GFP_KERNEL);
1464         if (!process)
1465                 goto err_alloc_process;
1466
1467         kref_init(&process->ref);
1468         mutex_init(&process->mutex);
1469         process->mm = thread->mm;
1470         process->lead_thread = thread->group_leader;
1471         process->n_pdds = 0;
1472         process->queues_paused = false;
1473         INIT_DELAYED_WORK(&process->eviction_work, evict_process_worker);
1474         INIT_DELAYED_WORK(&process->restore_work, restore_process_worker);
1475         process->last_restore_timestamp = get_jiffies_64();
1476         err = kfd_event_init_process(process);
1477         if (err)
1478                 goto err_event_init;
1479         process->is_32bit_user_mode = in_compat_syscall();
1480         process->debug_trap_enabled = false;
1481         process->debugger_process = NULL;
1482         process->exception_enable_mask = 0;
1483         atomic_set(&process->debugged_process_count, 0);
1484         sema_init(&process->runtime_enable_sema, 0);
1485
1486         process->pasid = kfd_pasid_alloc();
1487         if (process->pasid == 0) {
1488                 err = -ENOSPC;
1489                 goto err_alloc_pasid;
1490         }
1491
1492         err = pqm_init(&process->pqm, process);
1493         if (err != 0)
1494                 goto err_process_pqm_init;
1495
1496         /* init process apertures*/
1497         err = kfd_init_apertures(process);
1498         if (err != 0)
1499                 goto err_init_apertures;
1500
1501         /* Check XNACK support after PDDs are created in kfd_init_apertures */
1502         process->xnack_enabled = kfd_process_xnack_mode(process, false);
1503
1504         err = svm_range_list_init(process);
1505         if (err)
1506                 goto err_init_svm_range_list;
1507
1508         /* alloc_notifier needs to find the process in the hash table */
1509         hash_add_rcu(kfd_processes_table, &process->kfd_processes,
1510                         (uintptr_t)process->mm);
1511
1512         /* Avoid free_notifier to start kfd_process_wq_release if
1513          * mmu_notifier_get failed because of pending signal.
1514          */
1515         kref_get(&process->ref);
1516
1517         /* MMU notifier registration must be the last call that can fail
1518          * because after this point we cannot unwind the process creation.
1519          * After this point, mmu_notifier_put will trigger the cleanup by
1520          * dropping the last process reference in the free_notifier.
1521          */
1522         mn = mmu_notifier_get(&kfd_process_mmu_notifier_ops, process->mm);
1523         if (IS_ERR(mn)) {
1524                 err = PTR_ERR(mn);
1525                 goto err_register_notifier;
1526         }
1527         BUG_ON(mn != &process->mmu_notifier);
1528
1529         kfd_unref_process(process);
1530         get_task_struct(process->lead_thread);
1531
1532         INIT_WORK(&process->debug_event_workarea, debug_event_write_work_handler);
1533
1534         return process;
1535
1536 err_register_notifier:
1537         hash_del_rcu(&process->kfd_processes);
1538         svm_range_list_fini(process);
1539 err_init_svm_range_list:
1540         kfd_process_free_outstanding_kfd_bos(process);
1541         kfd_process_destroy_pdds(process);
1542 err_init_apertures:
1543         pqm_uninit(&process->pqm);
1544 err_process_pqm_init:
1545         kfd_pasid_free(process->pasid);
1546 err_alloc_pasid:
1547         kfd_event_free_process(process);
1548 err_event_init:
1549         mutex_destroy(&process->mutex);
1550         kfree(process);
1551 err_alloc_process:
1552         return ERR_PTR(err);
1553 }
1554
1555 struct kfd_process_device *kfd_get_process_device_data(struct kfd_node *dev,
1556                                                         struct kfd_process *p)
1557 {
1558         int i;
1559
1560         for (i = 0; i < p->n_pdds; i++)
1561                 if (p->pdds[i]->dev == dev)
1562                         return p->pdds[i];
1563
1564         return NULL;
1565 }
1566
1567 struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev,
1568                                                         struct kfd_process *p)
1569 {
1570         struct kfd_process_device *pdd = NULL;
1571         int retval = 0;
1572
1573         if (WARN_ON_ONCE(p->n_pdds >= MAX_GPU_INSTANCE))
1574                 return NULL;
1575         pdd = kzalloc(sizeof(*pdd), GFP_KERNEL);
1576         if (!pdd)
1577                 return NULL;
1578
1579         pdd->dev = dev;
1580         INIT_LIST_HEAD(&pdd->qpd.queues_list);
1581         INIT_LIST_HEAD(&pdd->qpd.priv_queue_list);
1582         pdd->qpd.dqm = dev->dqm;
1583         pdd->qpd.pqm = &p->pqm;
1584         pdd->qpd.evicted = 0;
1585         pdd->qpd.mapped_gws_queue = false;
1586         pdd->process = p;
1587         pdd->bound = PDD_UNBOUND;
1588         pdd->already_dequeued = false;
1589         pdd->runtime_inuse = false;
1590         pdd->vram_usage = 0;
1591         pdd->sdma_past_activity_counter = 0;
1592         pdd->user_gpu_id = dev->id;
1593         atomic64_set(&pdd->evict_duration_counter, 0);
1594
1595         if (dev->kfd->shared_resources.enable_mes) {
1596                 retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev,
1597                                                 AMDGPU_MES_PROC_CTX_SIZE,
1598                                                 &pdd->proc_ctx_bo,
1599                                                 &pdd->proc_ctx_gpu_addr,
1600                                                 &pdd->proc_ctx_cpu_ptr,
1601                                                 false);
1602                 if (retval) {
1603                         pr_err("failed to allocate process context bo\n");
1604                         goto err_free_pdd;
1605                 }
1606                 memset(pdd->proc_ctx_cpu_ptr, 0, AMDGPU_MES_PROC_CTX_SIZE);
1607         }
1608
1609         p->pdds[p->n_pdds++] = pdd;
1610         if (kfd_dbg_is_per_vmid_supported(pdd->dev))
1611                 pdd->spi_dbg_override = pdd->dev->kfd2kgd->disable_debug_trap(
1612                                                         pdd->dev->adev,
1613                                                         false,
1614                                                         0);
1615
1616         /* Init idr used for memory handle translation */
1617         idr_init(&pdd->alloc_idr);
1618
1619         return pdd;
1620
1621 err_free_pdd:
1622         kfree(pdd);
1623         return NULL;
1624 }
1625
1626 /**
1627  * kfd_process_device_init_vm - Initialize a VM for a process-device
1628  *
1629  * @pdd: The process-device
1630  * @drm_file: Optional pointer to a DRM file descriptor
1631  *
1632  * If @drm_file is specified, it will be used to acquire the VM from
1633  * that file descriptor. If successful, the @pdd takes ownership of
1634  * the file descriptor.
1635  *
1636  * If @drm_file is NULL, a new VM is created.
1637  *
1638  * Returns 0 on success, -errno on failure.
1639  */
1640 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
1641                                struct file *drm_file)
1642 {
1643         struct amdgpu_fpriv *drv_priv;
1644         struct amdgpu_vm *avm;
1645         struct kfd_process *p;
1646         struct dma_fence *ef;
1647         struct kfd_node *dev;
1648         int ret;
1649
1650         if (!drm_file)
1651                 return -EINVAL;
1652
1653         if (pdd->drm_priv)
1654                 return -EBUSY;
1655
1656         ret = amdgpu_file_to_fpriv(drm_file, &drv_priv);
1657         if (ret)
1658                 return ret;
1659         avm = &drv_priv->vm;
1660
1661         p = pdd->process;
1662         dev = pdd->dev;
1663
1664         ret = amdgpu_amdkfd_gpuvm_acquire_process_vm(dev->adev, avm,
1665                                                      &p->kgd_process_info,
1666                                                      &ef);
1667         if (ret) {
1668                 pr_err("Failed to create process VM object\n");
1669                 return ret;
1670         }
1671         RCU_INIT_POINTER(p->ef, ef);
1672         pdd->drm_priv = drm_file->private_data;
1673
1674         ret = kfd_process_device_reserve_ib_mem(pdd);
1675         if (ret)
1676                 goto err_reserve_ib_mem;
1677         ret = kfd_process_device_init_cwsr_dgpu(pdd);
1678         if (ret)
1679                 goto err_init_cwsr;
1680
1681         ret = amdgpu_amdkfd_gpuvm_set_vm_pasid(dev->adev, avm, p->pasid);
1682         if (ret)
1683                 goto err_set_pasid;
1684
1685         pdd->drm_file = drm_file;
1686
1687         return 0;
1688
1689 err_set_pasid:
1690         kfd_process_device_destroy_cwsr_dgpu(pdd);
1691 err_init_cwsr:
1692         kfd_process_device_destroy_ib_mem(pdd);
1693 err_reserve_ib_mem:
1694         pdd->drm_priv = NULL;
1695         amdgpu_amdkfd_gpuvm_destroy_cb(dev->adev, avm);
1696
1697         return ret;
1698 }
1699
1700 /*
1701  * Direct the IOMMU to bind the process (specifically the pasid->mm)
1702  * to the device.
1703  * Unbinding occurs when the process dies or the device is removed.
1704  *
1705  * Assumes that the process lock is held.
1706  */
1707 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_node *dev,
1708                                                         struct kfd_process *p)
1709 {
1710         struct kfd_process_device *pdd;
1711         int err;
1712
1713         pdd = kfd_get_process_device_data(dev, p);
1714         if (!pdd) {
1715                 pr_err("Process device data doesn't exist\n");
1716                 return ERR_PTR(-ENOMEM);
1717         }
1718
1719         if (!pdd->drm_priv)
1720                 return ERR_PTR(-ENODEV);
1721
1722         /*
1723          * signal runtime-pm system to auto resume and prevent
1724          * further runtime suspend once device pdd is created until
1725          * pdd is destroyed.
1726          */
1727         if (!pdd->runtime_inuse) {
1728                 err = pm_runtime_get_sync(adev_to_drm(dev->adev)->dev);
1729                 if (err < 0) {
1730                         pm_runtime_put_autosuspend(adev_to_drm(dev->adev)->dev);
1731                         return ERR_PTR(err);
1732                 }
1733         }
1734
1735         /*
1736          * make sure that runtime_usage counter is incremented just once
1737          * per pdd
1738          */
1739         pdd->runtime_inuse = true;
1740
1741         return pdd;
1742 }
1743
1744 /* Create specific handle mapped to mem from process local memory idr
1745  * Assumes that the process lock is held.
1746  */
1747 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
1748                                         void *mem)
1749 {
1750         return idr_alloc(&pdd->alloc_idr, mem, 0, 0, GFP_KERNEL);
1751 }
1752
1753 /* Translate specific handle from process local memory idr
1754  * Assumes that the process lock is held.
1755  */
1756 void *kfd_process_device_translate_handle(struct kfd_process_device *pdd,
1757                                         int handle)
1758 {
1759         if (handle < 0)
1760                 return NULL;
1761
1762         return idr_find(&pdd->alloc_idr, handle);
1763 }
1764
1765 /* Remove specific handle from process local memory idr
1766  * Assumes that the process lock is held.
1767  */
1768 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
1769                                         int handle)
1770 {
1771         if (handle >= 0)
1772                 idr_remove(&pdd->alloc_idr, handle);
1773 }
1774
1775 /* This increments the process->ref counter. */
1776 struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid)
1777 {
1778         struct kfd_process *p, *ret_p = NULL;
1779         unsigned int temp;
1780
1781         int idx = srcu_read_lock(&kfd_processes_srcu);
1782
1783         hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1784                 if (p->pasid == pasid) {
1785                         kref_get(&p->ref);
1786                         ret_p = p;
1787                         break;
1788                 }
1789         }
1790
1791         srcu_read_unlock(&kfd_processes_srcu, idx);
1792
1793         return ret_p;
1794 }
1795
1796 /* This increments the process->ref counter. */
1797 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm)
1798 {
1799         struct kfd_process *p;
1800
1801         int idx = srcu_read_lock(&kfd_processes_srcu);
1802
1803         p = find_process_by_mm(mm);
1804         if (p)
1805                 kref_get(&p->ref);
1806
1807         srcu_read_unlock(&kfd_processes_srcu, idx);
1808
1809         return p;
1810 }
1811
1812 /* kfd_process_evict_queues - Evict all user queues of a process
1813  *
1814  * Eviction is reference-counted per process-device. This means multiple
1815  * evictions from different sources can be nested safely.
1816  */
1817 int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger)
1818 {
1819         int r = 0;
1820         int i;
1821         unsigned int n_evicted = 0;
1822
1823         for (i = 0; i < p->n_pdds; i++) {
1824                 struct kfd_process_device *pdd = p->pdds[i];
1825
1826                 kfd_smi_event_queue_eviction(pdd->dev, p->lead_thread->pid,
1827                                              trigger);
1828
1829                 r = pdd->dev->dqm->ops.evict_process_queues(pdd->dev->dqm,
1830                                                             &pdd->qpd);
1831                 /* evict return -EIO if HWS is hang or asic is resetting, in this case
1832                  * we would like to set all the queues to be in evicted state to prevent
1833                  * them been add back since they actually not be saved right now.
1834                  */
1835                 if (r && r != -EIO) {
1836                         pr_err("Failed to evict process queues\n");
1837                         goto fail;
1838                 }
1839                 n_evicted++;
1840         }
1841
1842         return r;
1843
1844 fail:
1845         /* To keep state consistent, roll back partial eviction by
1846          * restoring queues
1847          */
1848         for (i = 0; i < p->n_pdds; i++) {
1849                 struct kfd_process_device *pdd = p->pdds[i];
1850
1851                 if (n_evicted == 0)
1852                         break;
1853
1854                 kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid);
1855
1856                 if (pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1857                                                               &pdd->qpd))
1858                         pr_err("Failed to restore queues\n");
1859
1860                 n_evicted--;
1861         }
1862
1863         return r;
1864 }
1865
1866 /* kfd_process_restore_queues - Restore all user queues of a process */
1867 int kfd_process_restore_queues(struct kfd_process *p)
1868 {
1869         int r, ret = 0;
1870         int i;
1871
1872         for (i = 0; i < p->n_pdds; i++) {
1873                 struct kfd_process_device *pdd = p->pdds[i];
1874
1875                 kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid);
1876
1877                 r = pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1878                                                               &pdd->qpd);
1879                 if (r) {
1880                         pr_err("Failed to restore process queues\n");
1881                         if (!ret)
1882                                 ret = r;
1883                 }
1884         }
1885
1886         return ret;
1887 }
1888
1889 int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id)
1890 {
1891         int i;
1892
1893         for (i = 0; i < p->n_pdds; i++)
1894                 if (p->pdds[i] && gpu_id == p->pdds[i]->user_gpu_id)
1895                         return i;
1896         return -EINVAL;
1897 }
1898
1899 int
1900 kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node,
1901                             uint32_t *gpuid, uint32_t *gpuidx)
1902 {
1903         int i;
1904
1905         for (i = 0; i < p->n_pdds; i++)
1906                 if (p->pdds[i] && p->pdds[i]->dev == node) {
1907                         *gpuid = p->pdds[i]->user_gpu_id;
1908                         *gpuidx = i;
1909                         return 0;
1910                 }
1911         return -EINVAL;
1912 }
1913
1914 static int signal_eviction_fence(struct kfd_process *p)
1915 {
1916         struct dma_fence *ef;
1917         int ret;
1918
1919         rcu_read_lock();
1920         ef = dma_fence_get_rcu_safe(&p->ef);
1921         rcu_read_unlock();
1922
1923         ret = dma_fence_signal(ef);
1924         dma_fence_put(ef);
1925
1926         return ret;
1927 }
1928
1929 static void evict_process_worker(struct work_struct *work)
1930 {
1931         int ret;
1932         struct kfd_process *p;
1933         struct delayed_work *dwork;
1934
1935         dwork = to_delayed_work(work);
1936
1937         /* Process termination destroys this worker thread. So during the
1938          * lifetime of this thread, kfd_process p will be valid
1939          */
1940         p = container_of(dwork, struct kfd_process, eviction_work);
1941
1942         pr_debug("Started evicting pasid 0x%x\n", p->pasid);
1943         ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_TRIGGER_TTM);
1944         if (!ret) {
1945                 /* If another thread already signaled the eviction fence,
1946                  * they are responsible stopping the queues and scheduling
1947                  * the restore work.
1948                  */
1949                 if (!signal_eviction_fence(p))
1950                         queue_delayed_work(kfd_restore_wq, &p->restore_work,
1951                                 msecs_to_jiffies(PROCESS_RESTORE_TIME_MS));
1952                 else
1953                         kfd_process_restore_queues(p);
1954
1955                 pr_debug("Finished evicting pasid 0x%x\n", p->pasid);
1956         } else
1957                 pr_err("Failed to evict queues of pasid 0x%x\n", p->pasid);
1958 }
1959
1960 static int restore_process_helper(struct kfd_process *p)
1961 {
1962         int ret = 0;
1963
1964         /* VMs may not have been acquired yet during debugging. */
1965         if (p->kgd_process_info) {
1966                 ret = amdgpu_amdkfd_gpuvm_restore_process_bos(
1967                         p->kgd_process_info, &p->ef);
1968                 if (ret)
1969                         return ret;
1970         }
1971
1972         ret = kfd_process_restore_queues(p);
1973         if (!ret)
1974                 pr_debug("Finished restoring pasid 0x%x\n", p->pasid);
1975         else
1976                 pr_err("Failed to restore queues of pasid 0x%x\n", p->pasid);
1977
1978         return ret;
1979 }
1980
1981 static void restore_process_worker(struct work_struct *work)
1982 {
1983         struct delayed_work *dwork;
1984         struct kfd_process *p;
1985         int ret = 0;
1986
1987         dwork = to_delayed_work(work);
1988
1989         /* Process termination destroys this worker thread. So during the
1990          * lifetime of this thread, kfd_process p will be valid
1991          */
1992         p = container_of(dwork, struct kfd_process, restore_work);
1993         pr_debug("Started restoring pasid 0x%x\n", p->pasid);
1994
1995         /* Setting last_restore_timestamp before successful restoration.
1996          * Otherwise this would have to be set by KGD (restore_process_bos)
1997          * before KFD BOs are unreserved. If not, the process can be evicted
1998          * again before the timestamp is set.
1999          * If restore fails, the timestamp will be set again in the next
2000          * attempt. This would mean that the minimum GPU quanta would be
2001          * PROCESS_ACTIVE_TIME_MS - (time to execute the following two
2002          * functions)
2003          */
2004
2005         p->last_restore_timestamp = get_jiffies_64();
2006
2007         ret = restore_process_helper(p);
2008         if (ret) {
2009                 pr_debug("Failed to restore BOs of pasid 0x%x, retry after %d ms\n",
2010                          p->pasid, PROCESS_BACK_OFF_TIME_MS);
2011                 ret = queue_delayed_work(kfd_restore_wq, &p->restore_work,
2012                                 msecs_to_jiffies(PROCESS_BACK_OFF_TIME_MS));
2013                 WARN(!ret, "reschedule restore work failed\n");
2014         }
2015 }
2016
2017 void kfd_suspend_all_processes(void)
2018 {
2019         struct kfd_process *p;
2020         unsigned int temp;
2021         int idx = srcu_read_lock(&kfd_processes_srcu);
2022
2023         WARN(debug_evictions, "Evicting all processes");
2024         hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2025                 if (kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_TRIGGER_SUSPEND))
2026                         pr_err("Failed to suspend process 0x%x\n", p->pasid);
2027                 signal_eviction_fence(p);
2028         }
2029         srcu_read_unlock(&kfd_processes_srcu, idx);
2030 }
2031
2032 int kfd_resume_all_processes(void)
2033 {
2034         struct kfd_process *p;
2035         unsigned int temp;
2036         int ret = 0, idx = srcu_read_lock(&kfd_processes_srcu);
2037
2038         hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2039                 if (restore_process_helper(p)) {
2040                         pr_err("Restore process %d failed during resume\n",
2041                                p->pasid);
2042                         ret = -EFAULT;
2043                 }
2044         }
2045         srcu_read_unlock(&kfd_processes_srcu, idx);
2046         return ret;
2047 }
2048
2049 int kfd_reserved_mem_mmap(struct kfd_node *dev, struct kfd_process *process,
2050                           struct vm_area_struct *vma)
2051 {
2052         struct kfd_process_device *pdd;
2053         struct qcm_process_device *qpd;
2054
2055         if ((vma->vm_end - vma->vm_start) != KFD_CWSR_TBA_TMA_SIZE) {
2056                 pr_err("Incorrect CWSR mapping size.\n");
2057                 return -EINVAL;
2058         }
2059
2060         pdd = kfd_get_process_device_data(dev, process);
2061         if (!pdd)
2062                 return -EINVAL;
2063         qpd = &pdd->qpd;
2064
2065         qpd->cwsr_kaddr = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
2066                                         get_order(KFD_CWSR_TBA_TMA_SIZE));
2067         if (!qpd->cwsr_kaddr) {
2068                 pr_err("Error allocating per process CWSR buffer.\n");
2069                 return -ENOMEM;
2070         }
2071
2072         vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND
2073                 | VM_NORESERVE | VM_DONTDUMP | VM_PFNMAP);
2074         /* Mapping pages to user process */
2075         return remap_pfn_range(vma, vma->vm_start,
2076                                PFN_DOWN(__pa(qpd->cwsr_kaddr)),
2077                                KFD_CWSR_TBA_TMA_SIZE, vma->vm_page_prot);
2078 }
2079
2080 /* assumes caller holds process lock. */
2081 int kfd_process_drain_interrupts(struct kfd_process_device *pdd)
2082 {
2083         uint32_t irq_drain_fence[8];
2084         uint8_t node_id = 0;
2085         int r = 0;
2086
2087         if (!KFD_IS_SOC15(pdd->dev))
2088                 return 0;
2089
2090         pdd->process->irq_drain_is_open = true;
2091
2092         memset(irq_drain_fence, 0, sizeof(irq_drain_fence));
2093         irq_drain_fence[0] = (KFD_IRQ_FENCE_SOURCEID << 8) |
2094                                                         KFD_IRQ_FENCE_CLIENTID;
2095         irq_drain_fence[3] = pdd->process->pasid;
2096
2097         /*
2098          * For GFX 9.4.3, send the NodeId also in IH cookie DW[3]
2099          */
2100         if (KFD_GC_VERSION(pdd->dev->kfd) == IP_VERSION(9, 4, 3)) {
2101                 node_id = ffs(pdd->dev->interrupt_bitmap) - 1;
2102                 irq_drain_fence[3] |= node_id << 16;
2103         }
2104
2105         /* ensure stale irqs scheduled KFD interrupts and send drain fence. */
2106         if (amdgpu_amdkfd_send_close_event_drain_irq(pdd->dev->adev,
2107                                                      irq_drain_fence)) {
2108                 pdd->process->irq_drain_is_open = false;
2109                 return 0;
2110         }
2111
2112         r = wait_event_interruptible(pdd->process->wait_irq_drain,
2113                                      !READ_ONCE(pdd->process->irq_drain_is_open));
2114         if (r)
2115                 pdd->process->irq_drain_is_open = false;
2116
2117         return r;
2118 }
2119
2120 void kfd_process_close_interrupt_drain(unsigned int pasid)
2121 {
2122         struct kfd_process *p;
2123
2124         p = kfd_lookup_process_by_pasid(pasid);
2125
2126         if (!p)
2127                 return;
2128
2129         WRITE_ONCE(p->irq_drain_is_open, false);
2130         wake_up_all(&p->wait_irq_drain);
2131         kfd_unref_process(p);
2132 }
2133
2134 struct send_exception_work_handler_workarea {
2135         struct work_struct work;
2136         struct kfd_process *p;
2137         unsigned int queue_id;
2138         uint64_t error_reason;
2139 };
2140
2141 static void send_exception_work_handler(struct work_struct *work)
2142 {
2143         struct send_exception_work_handler_workarea *workarea;
2144         struct kfd_process *p;
2145         struct queue *q;
2146         struct mm_struct *mm;
2147         struct kfd_context_save_area_header __user *csa_header;
2148         uint64_t __user *err_payload_ptr;
2149         uint64_t cur_err;
2150         uint32_t ev_id;
2151
2152         workarea = container_of(work,
2153                                 struct send_exception_work_handler_workarea,
2154                                 work);
2155         p = workarea->p;
2156
2157         mm = get_task_mm(p->lead_thread);
2158
2159         if (!mm)
2160                 return;
2161
2162         kthread_use_mm(mm);
2163
2164         q = pqm_get_user_queue(&p->pqm, workarea->queue_id);
2165
2166         if (!q)
2167                 goto out;
2168
2169         csa_header = (void __user *)q->properties.ctx_save_restore_area_address;
2170
2171         get_user(err_payload_ptr, (uint64_t __user **)&csa_header->err_payload_addr);
2172         get_user(cur_err, err_payload_ptr);
2173         cur_err |= workarea->error_reason;
2174         put_user(cur_err, err_payload_ptr);
2175         get_user(ev_id, &csa_header->err_event_id);
2176
2177         kfd_set_event(p, ev_id);
2178
2179 out:
2180         kthread_unuse_mm(mm);
2181         mmput(mm);
2182 }
2183
2184 int kfd_send_exception_to_runtime(struct kfd_process *p,
2185                         unsigned int queue_id,
2186                         uint64_t error_reason)
2187 {
2188         struct send_exception_work_handler_workarea worker;
2189
2190         INIT_WORK_ONSTACK(&worker.work, send_exception_work_handler);
2191
2192         worker.p = p;
2193         worker.queue_id = queue_id;
2194         worker.error_reason = error_reason;
2195
2196         schedule_work(&worker.work);
2197         flush_work(&worker.work);
2198         destroy_work_on_stack(&worker.work);
2199
2200         return 0;
2201 }
2202
2203 struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *p, uint32_t gpu_id)
2204 {
2205         int i;
2206
2207         if (gpu_id) {
2208                 for (i = 0; i < p->n_pdds; i++) {
2209                         struct kfd_process_device *pdd = p->pdds[i];
2210
2211                         if (pdd->user_gpu_id == gpu_id)
2212                                 return pdd;
2213                 }
2214         }
2215         return NULL;
2216 }
2217
2218 int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id)
2219 {
2220         int i;
2221
2222         if (!actual_gpu_id)
2223                 return 0;
2224
2225         for (i = 0; i < p->n_pdds; i++) {
2226                 struct kfd_process_device *pdd = p->pdds[i];
2227
2228                 if (pdd->dev->id == actual_gpu_id)
2229                         return pdd->user_gpu_id;
2230         }
2231         return -EINVAL;
2232 }
2233
2234 #if defined(CONFIG_DEBUG_FS)
2235
2236 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data)
2237 {
2238         struct kfd_process *p;
2239         unsigned int temp;
2240         int r = 0;
2241
2242         int idx = srcu_read_lock(&kfd_processes_srcu);
2243
2244         hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2245                 seq_printf(m, "Process %d PASID 0x%x:\n",
2246                            p->lead_thread->tgid, p->pasid);
2247
2248                 mutex_lock(&p->mutex);
2249                 r = pqm_debugfs_mqds(m, &p->pqm);
2250                 mutex_unlock(&p->mutex);
2251
2252                 if (r)
2253                         break;
2254         }
2255
2256         srcu_read_unlock(&kfd_processes_srcu, idx);
2257
2258         return r;
2259 }
2260
2261 #endif