1 // SPDX-License-Identifier: GPL-2.0 OR MIT
3 * Copyright 2014-2022 Advanced Micro Devices, Inc.
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:
12 * The above copyright notice and this permission notice shall be included in
13 * all copies or substantial portions of the Software.
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.
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"
42 #include "kfd_device_queue_manager.h"
44 #include "kfd_smi_events.h"
45 #include "kfd_debug.h"
48 * List of struct kfd_process (field kfd_process).
49 * Unique/indexed by mm_struct*
51 DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
52 DEFINE_MUTEX(kfd_processes_mutex);
54 DEFINE_SRCU(kfd_processes_srcu);
56 /* For process termination handling */
57 static struct workqueue_struct *kfd_process_wq;
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.
65 static struct workqueue_struct *kfd_restore_wq;
67 static struct kfd_process *find_process(const struct task_struct *thread,
69 static void kfd_process_ref_release(struct kref *ref);
70 static struct kfd_process *create_process(const struct task_struct *thread);
72 static void evict_process_worker(struct work_struct *work);
73 static void restore_process_worker(struct work_struct *work);
75 static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd);
77 struct kfd_procfs_tree {
81 static struct kfd_procfs_tree procfs;
84 * Structure for SDMA activity tracking
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;
92 struct temp_sdma_queue_list {
93 uint64_t __user *rptr;
95 unsigned int queue_id;
96 struct list_head list;
99 static void kfd_sdma_activity_worker(struct work_struct *work)
101 struct kfd_sdma_activity_handler_workarea *workarea;
102 struct kfd_process_device *pdd;
104 struct mm_struct *mm;
106 struct qcm_process_device *qpd;
107 struct device_queue_manager *dqm;
109 struct temp_sdma_queue_list sdma_q_list;
110 struct temp_sdma_queue_list *sdma_q, *next;
112 workarea = container_of(work, struct kfd_sdma_activity_handler_workarea,
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.
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:
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.
144 INIT_LIST_HEAD(&sdma_q_list.list);
147 * Create the temp list of all SDMA queues
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))
156 sdma_q = kzalloc(sizeof(struct temp_sdma_queue_list), GFP_KERNEL);
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);
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
173 if (list_empty(&sdma_q_list.list)) {
174 workarea->sdma_activity_counter = pdd->sdma_past_activity_counter;
182 * Get the usage count for each SDMA queue in temp_list.
184 mm = get_task_mm(pdd->process->lead_thread);
190 list_for_each_entry(sdma_q, &sdma_q_list.list, list) {
192 ret = read_sdma_queue_counter(sdma_q->rptr, &val);
194 pr_debug("Failed to read SDMA queue active counter for queue id: %d",
197 sdma_q->sdma_val = val;
198 workarea->sdma_activity_counter += val;
202 kthread_unuse_mm(mm);
206 * Do a second iteration over qpd_queues_list to check if any SDMA
207 * nodes got deleted while fetching SDMA counter.
211 workarea->sdma_activity_counter += pdd->sdma_past_activity_counter;
213 list_for_each_entry(q, &qpd->queues_list, list) {
214 if (list_empty(&sdma_q_list.list))
217 if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
218 (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
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);
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.
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);
247 list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
248 list_del(&sdma_q->list);
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
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
263 * Return: Number of bytes written to user buffer or an error value
265 static int kfd_get_cu_occupancy(struct attribute *attr, char *buffer)
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;
274 struct kfd_cu_occupancy *cu_occupancy;
277 pdd = container_of(attr, struct kfd_process_device, attr_cu_occupancy);
279 if (dev->kfd2kgd->get_cu_occupancy == NULL)
284 if (pdd->qpd.queue_count == 0) {
285 pr_debug("Gpu-Id: %d has no active queues for process %d\n",
286 dev->id, proc->pasid);
287 return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt);
290 /* Collect wave count from device if it supports */
292 max_waves_per_cu = 0;
294 cu_occupancy = kcalloc(AMDGPU_MAX_QUEUES, sizeof(*cu_occupancy), GFP_KERNEL);
299 * For GFX 9.4.3, fetch the CU occupancy from the first XCC in the partition.
300 * For AQL queues, because of cooperative dispatch we multiply the wave count
301 * by number of XCCs in the partition to get the total wave counts across all
302 * XCCs in the partition.
303 * For PM4 queues, there is no cooperative dispatch so wave_cnt stay as it is.
305 dev->kfd2kgd->get_cu_occupancy(dev->adev, cu_occupancy,
306 &max_waves_per_cu, ffs(dev->xcc_mask) - 1);
308 for (i = 0; i < AMDGPU_MAX_QUEUES; i++) {
309 if (cu_occupancy[i].wave_cnt != 0 &&
310 kfd_dqm_is_queue_in_process(dev->dqm, &pdd->qpd,
311 cu_occupancy[i].doorbell_off,
313 if (unlikely(queue_format == KFD_QUEUE_FORMAT_PM4))
314 wave_cnt += cu_occupancy[i].wave_cnt;
316 wave_cnt += (NUM_XCC(dev->xcc_mask) *
317 cu_occupancy[i].wave_cnt);
321 /* Translate wave count to number of compute units */
322 cu_cnt = (wave_cnt + (max_waves_per_cu - 1)) / max_waves_per_cu;
324 return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt);
327 static ssize_t kfd_procfs_show(struct kobject *kobj, struct attribute *attr,
330 if (strcmp(attr->name, "pasid") == 0) {
331 struct kfd_process *p = container_of(attr, struct kfd_process,
334 return snprintf(buffer, PAGE_SIZE, "%d\n", p->pasid);
335 } else if (strncmp(attr->name, "vram_", 5) == 0) {
336 struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
338 return snprintf(buffer, PAGE_SIZE, "%llu\n", atomic64_read(&pdd->vram_usage));
339 } else if (strncmp(attr->name, "sdma_", 5) == 0) {
340 struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
342 struct kfd_sdma_activity_handler_workarea sdma_activity_work_handler;
344 INIT_WORK_ONSTACK(&sdma_activity_work_handler.sdma_activity_work,
345 kfd_sdma_activity_worker);
347 sdma_activity_work_handler.pdd = pdd;
348 sdma_activity_work_handler.sdma_activity_counter = 0;
350 schedule_work(&sdma_activity_work_handler.sdma_activity_work);
352 flush_work(&sdma_activity_work_handler.sdma_activity_work);
353 destroy_work_on_stack(&sdma_activity_work_handler.sdma_activity_work);
355 return snprintf(buffer, PAGE_SIZE, "%llu\n",
356 (sdma_activity_work_handler.sdma_activity_counter)/
357 SDMA_ACTIVITY_DIVISOR);
359 pr_err("Invalid attribute");
366 static void kfd_procfs_kobj_release(struct kobject *kobj)
371 static const struct sysfs_ops kfd_procfs_ops = {
372 .show = kfd_procfs_show,
375 static const struct kobj_type procfs_type = {
376 .release = kfd_procfs_kobj_release,
377 .sysfs_ops = &kfd_procfs_ops,
380 void kfd_procfs_init(void)
384 procfs.kobj = kfd_alloc_struct(procfs.kobj);
388 ret = kobject_init_and_add(procfs.kobj, &procfs_type,
389 &kfd_device->kobj, "proc");
391 pr_warn("Could not create procfs proc folder");
392 /* If we fail to create the procfs, clean up */
393 kfd_procfs_shutdown();
397 void kfd_procfs_shutdown(void)
400 kobject_del(procfs.kobj);
401 kobject_put(procfs.kobj);
406 static ssize_t kfd_procfs_queue_show(struct kobject *kobj,
407 struct attribute *attr, char *buffer)
409 struct queue *q = container_of(kobj, struct queue, kobj);
411 if (!strcmp(attr->name, "size"))
412 return snprintf(buffer, PAGE_SIZE, "%llu",
413 q->properties.queue_size);
414 else if (!strcmp(attr->name, "type"))
415 return snprintf(buffer, PAGE_SIZE, "%d", q->properties.type);
416 else if (!strcmp(attr->name, "gpuid"))
417 return snprintf(buffer, PAGE_SIZE, "%u", q->device->id);
419 pr_err("Invalid attribute");
424 static ssize_t kfd_procfs_stats_show(struct kobject *kobj,
425 struct attribute *attr, char *buffer)
427 if (strcmp(attr->name, "evicted_ms") == 0) {
428 struct kfd_process_device *pdd = container_of(attr,
429 struct kfd_process_device,
431 uint64_t evict_jiffies;
433 evict_jiffies = atomic64_read(&pdd->evict_duration_counter);
435 return snprintf(buffer,
438 jiffies64_to_msecs(evict_jiffies));
440 /* Sysfs handle that gets CU occupancy is per device */
441 } else if (strcmp(attr->name, "cu_occupancy") == 0) {
442 return kfd_get_cu_occupancy(attr, buffer);
444 pr_err("Invalid attribute");
450 static ssize_t kfd_sysfs_counters_show(struct kobject *kobj,
451 struct attribute *attr, char *buf)
453 struct kfd_process_device *pdd;
455 if (!strcmp(attr->name, "faults")) {
456 pdd = container_of(attr, struct kfd_process_device,
458 return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->faults));
460 if (!strcmp(attr->name, "page_in")) {
461 pdd = container_of(attr, struct kfd_process_device,
463 return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->page_in));
465 if (!strcmp(attr->name, "page_out")) {
466 pdd = container_of(attr, struct kfd_process_device,
468 return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->page_out));
473 static struct attribute attr_queue_size = {
475 .mode = KFD_SYSFS_FILE_MODE
478 static struct attribute attr_queue_type = {
480 .mode = KFD_SYSFS_FILE_MODE
483 static struct attribute attr_queue_gpuid = {
485 .mode = KFD_SYSFS_FILE_MODE
488 static struct attribute *procfs_queue_attrs[] = {
494 ATTRIBUTE_GROUPS(procfs_queue);
496 static const struct sysfs_ops procfs_queue_ops = {
497 .show = kfd_procfs_queue_show,
500 static const struct kobj_type procfs_queue_type = {
501 .sysfs_ops = &procfs_queue_ops,
502 .default_groups = procfs_queue_groups,
505 static const struct sysfs_ops procfs_stats_ops = {
506 .show = kfd_procfs_stats_show,
509 static const struct kobj_type procfs_stats_type = {
510 .sysfs_ops = &procfs_stats_ops,
511 .release = kfd_procfs_kobj_release,
514 static const struct sysfs_ops sysfs_counters_ops = {
515 .show = kfd_sysfs_counters_show,
518 static const struct kobj_type sysfs_counters_type = {
519 .sysfs_ops = &sysfs_counters_ops,
520 .release = kfd_procfs_kobj_release,
523 int kfd_procfs_add_queue(struct queue *q)
525 struct kfd_process *proc;
528 if (!q || !q->process)
532 /* Create proc/<pid>/queues/<queue id> folder */
533 if (!proc->kobj_queues)
535 ret = kobject_init_and_add(&q->kobj, &procfs_queue_type,
536 proc->kobj_queues, "%u", q->properties.queue_id);
538 pr_warn("Creating proc/<pid>/queues/%u failed",
539 q->properties.queue_id);
540 kobject_put(&q->kobj);
547 static void kfd_sysfs_create_file(struct kobject *kobj, struct attribute *attr,
552 if (!kobj || !attr || !name)
556 attr->mode = KFD_SYSFS_FILE_MODE;
557 sysfs_attr_init(attr);
559 ret = sysfs_create_file(kobj, attr);
561 pr_warn("Create sysfs %s/%s failed %d", kobj->name, name, ret);
564 static void kfd_procfs_add_sysfs_stats(struct kfd_process *p)
568 char stats_dir_filename[MAX_SYSFS_FILENAME_LEN];
574 * Create sysfs files for each GPU:
575 * - proc/<pid>/stats_<gpuid>/
576 * - proc/<pid>/stats_<gpuid>/evicted_ms
577 * - proc/<pid>/stats_<gpuid>/cu_occupancy
579 for (i = 0; i < p->n_pdds; i++) {
580 struct kfd_process_device *pdd = p->pdds[i];
582 snprintf(stats_dir_filename, MAX_SYSFS_FILENAME_LEN,
583 "stats_%u", pdd->dev->id);
584 pdd->kobj_stats = kfd_alloc_struct(pdd->kobj_stats);
585 if (!pdd->kobj_stats)
588 ret = kobject_init_and_add(pdd->kobj_stats,
594 pr_warn("Creating KFD proc/stats_%s folder failed",
596 kobject_put(pdd->kobj_stats);
597 pdd->kobj_stats = NULL;
601 kfd_sysfs_create_file(pdd->kobj_stats, &pdd->attr_evict,
603 /* Add sysfs file to report compute unit occupancy */
604 if (pdd->dev->kfd2kgd->get_cu_occupancy)
605 kfd_sysfs_create_file(pdd->kobj_stats,
606 &pdd->attr_cu_occupancy,
611 static void kfd_procfs_add_sysfs_counters(struct kfd_process *p)
615 char counters_dir_filename[MAX_SYSFS_FILENAME_LEN];
621 * Create sysfs files for each GPU which supports SVM
622 * - proc/<pid>/counters_<gpuid>/
623 * - proc/<pid>/counters_<gpuid>/faults
624 * - proc/<pid>/counters_<gpuid>/page_in
625 * - proc/<pid>/counters_<gpuid>/page_out
627 for_each_set_bit(i, p->svms.bitmap_supported, p->n_pdds) {
628 struct kfd_process_device *pdd = p->pdds[i];
629 struct kobject *kobj_counters;
631 snprintf(counters_dir_filename, MAX_SYSFS_FILENAME_LEN,
632 "counters_%u", pdd->dev->id);
633 kobj_counters = kfd_alloc_struct(kobj_counters);
637 ret = kobject_init_and_add(kobj_counters, &sysfs_counters_type,
638 p->kobj, counters_dir_filename);
640 pr_warn("Creating KFD proc/%s folder failed",
641 counters_dir_filename);
642 kobject_put(kobj_counters);
646 pdd->kobj_counters = kobj_counters;
647 kfd_sysfs_create_file(kobj_counters, &pdd->attr_faults,
649 kfd_sysfs_create_file(kobj_counters, &pdd->attr_page_in,
651 kfd_sysfs_create_file(kobj_counters, &pdd->attr_page_out,
656 static void kfd_procfs_add_sysfs_files(struct kfd_process *p)
664 * Create sysfs files for each GPU:
665 * - proc/<pid>/vram_<gpuid>
666 * - proc/<pid>/sdma_<gpuid>
668 for (i = 0; i < p->n_pdds; i++) {
669 struct kfd_process_device *pdd = p->pdds[i];
671 snprintf(pdd->vram_filename, MAX_SYSFS_FILENAME_LEN, "vram_%u",
673 kfd_sysfs_create_file(p->kobj, &pdd->attr_vram,
676 snprintf(pdd->sdma_filename, MAX_SYSFS_FILENAME_LEN, "sdma_%u",
678 kfd_sysfs_create_file(p->kobj, &pdd->attr_sdma,
683 void kfd_procfs_del_queue(struct queue *q)
688 kobject_del(&q->kobj);
689 kobject_put(&q->kobj);
692 int kfd_process_create_wq(void)
695 kfd_process_wq = alloc_workqueue("kfd_process_wq", 0, 0);
697 kfd_restore_wq = alloc_ordered_workqueue("kfd_restore_wq",
700 if (!kfd_process_wq || !kfd_restore_wq) {
701 kfd_process_destroy_wq();
708 void kfd_process_destroy_wq(void)
710 if (kfd_process_wq) {
711 destroy_workqueue(kfd_process_wq);
712 kfd_process_wq = NULL;
714 if (kfd_restore_wq) {
715 destroy_workqueue(kfd_restore_wq);
716 kfd_restore_wq = NULL;
720 static void kfd_process_free_gpuvm(struct kgd_mem *mem,
721 struct kfd_process_device *pdd, void **kptr)
723 struct kfd_node *dev = pdd->dev;
726 amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem);
730 amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(dev->adev, mem, pdd->drm_priv);
731 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, mem, pdd->drm_priv,
735 /* kfd_process_alloc_gpuvm - Allocate GPU VM for the KFD process
736 * This function should be only called right after the process
737 * is created and when kfd_processes_mutex is still being held
738 * to avoid concurrency. Because of that exclusiveness, we do
739 * not need to take p->mutex.
741 static int kfd_process_alloc_gpuvm(struct kfd_process_device *pdd,
742 uint64_t gpu_va, uint32_t size,
743 uint32_t flags, struct kgd_mem **mem, void **kptr)
745 struct kfd_node *kdev = pdd->dev;
748 err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(kdev->adev, gpu_va, size,
749 pdd->drm_priv, mem, NULL,
754 err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(kdev->adev, *mem,
759 err = amdgpu_amdkfd_gpuvm_sync_memory(kdev->adev, *mem, true);
761 pr_debug("Sync memory failed, wait interrupted by user signal\n");
762 goto sync_memory_failed;
766 err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(
767 (struct kgd_mem *)*mem, kptr, NULL);
769 pr_debug("Map GTT BO to kernel failed\n");
770 goto sync_memory_failed;
777 amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(kdev->adev, *mem, pdd->drm_priv);
780 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(kdev->adev, *mem, pdd->drm_priv,
788 /* kfd_process_device_reserve_ib_mem - Reserve memory inside the
789 * process for IB usage The memory reserved is for KFD to submit
790 * IB to AMDGPU from kernel. If the memory is reserved
791 * successfully, ib_kaddr will have the CPU/kernel
792 * address. Check ib_kaddr before accessing the memory.
794 static int kfd_process_device_reserve_ib_mem(struct kfd_process_device *pdd)
796 struct qcm_process_device *qpd = &pdd->qpd;
797 uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT |
798 KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE |
799 KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE |
800 KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
805 if (qpd->ib_kaddr || !qpd->ib_base)
808 /* ib_base is only set for dGPU */
809 ret = kfd_process_alloc_gpuvm(pdd, qpd->ib_base, PAGE_SIZE, flags,
815 qpd->ib_kaddr = kaddr;
820 static void kfd_process_device_destroy_ib_mem(struct kfd_process_device *pdd)
822 struct qcm_process_device *qpd = &pdd->qpd;
824 if (!qpd->ib_kaddr || !qpd->ib_base)
827 kfd_process_free_gpuvm(qpd->ib_mem, pdd, &qpd->ib_kaddr);
830 struct kfd_process *kfd_create_process(struct task_struct *thread)
832 struct kfd_process *process;
835 if (!(thread->mm && mmget_not_zero(thread->mm)))
836 return ERR_PTR(-EINVAL);
838 /* Only the pthreads threading model is supported. */
839 if (thread->group_leader->mm != thread->mm) {
841 return ERR_PTR(-EINVAL);
845 * take kfd processes mutex before starting of process creation
846 * so there won't be a case where two threads of the same process
847 * create two kfd_process structures
849 mutex_lock(&kfd_processes_mutex);
851 if (kfd_is_locked()) {
852 pr_debug("KFD is locked! Cannot create process");
853 process = ERR_PTR(-EINVAL);
857 /* A prior open of /dev/kfd could have already created the process.
858 * find_process will increase process kref in this case
860 process = find_process(thread, true);
862 pr_debug("Process already found\n");
864 /* If the process just called exec(3), it is possible that the
865 * cleanup of the kfd_process (following the release of the mm
866 * of the old process image) is still in the cleanup work queue.
867 * Make sure to drain any job before trying to recreate any
868 * resource for this process.
870 flush_workqueue(kfd_process_wq);
872 process = create_process(thread);
879 process->kobj = kfd_alloc_struct(process->kobj);
880 if (!process->kobj) {
881 pr_warn("Creating procfs kobject failed");
884 ret = kobject_init_and_add(process->kobj, &procfs_type,
886 (int)process->lead_thread->pid);
888 pr_warn("Creating procfs pid directory failed");
889 kobject_put(process->kobj);
893 kfd_sysfs_create_file(process->kobj, &process->attr_pasid,
896 process->kobj_queues = kobject_create_and_add("queues",
898 if (!process->kobj_queues)
899 pr_warn("Creating KFD proc/queues folder failed");
901 kfd_procfs_add_sysfs_stats(process);
902 kfd_procfs_add_sysfs_files(process);
903 kfd_procfs_add_sysfs_counters(process);
905 init_waitqueue_head(&process->wait_irq_drain);
908 mutex_unlock(&kfd_processes_mutex);
914 struct kfd_process *kfd_get_process(const struct task_struct *thread)
916 struct kfd_process *process;
919 return ERR_PTR(-EINVAL);
921 /* Only the pthreads threading model is supported. */
922 if (thread->group_leader->mm != thread->mm)
923 return ERR_PTR(-EINVAL);
925 process = find_process(thread, false);
927 return ERR_PTR(-EINVAL);
932 static struct kfd_process *find_process_by_mm(const struct mm_struct *mm)
934 struct kfd_process *process;
936 hash_for_each_possible_rcu(kfd_processes_table, process,
937 kfd_processes, (uintptr_t)mm)
938 if (process->mm == mm)
944 static struct kfd_process *find_process(const struct task_struct *thread,
947 struct kfd_process *p;
950 idx = srcu_read_lock(&kfd_processes_srcu);
951 p = find_process_by_mm(thread->mm);
954 srcu_read_unlock(&kfd_processes_srcu, idx);
959 void kfd_unref_process(struct kfd_process *p)
961 kref_put(&p->ref, kfd_process_ref_release);
964 /* This increments the process->ref counter. */
965 struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid)
967 struct task_struct *task = NULL;
968 struct kfd_process *p = NULL;
972 get_task_struct(task);
974 task = get_pid_task(pid, PIDTYPE_PID);
978 p = find_process(task, true);
979 put_task_struct(task);
985 static void kfd_process_device_free_bos(struct kfd_process_device *pdd)
987 struct kfd_process *p = pdd->process;
993 * Remove all handles from idr and release appropriate
994 * local memory object
996 idr_for_each_entry(&pdd->alloc_idr, mem, id) {
998 for (i = 0; i < p->n_pdds; i++) {
999 struct kfd_process_device *peer_pdd = p->pdds[i];
1001 if (!peer_pdd->drm_priv)
1003 amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1004 peer_pdd->dev->adev, mem, peer_pdd->drm_priv);
1007 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, mem,
1008 pdd->drm_priv, NULL);
1009 kfd_process_device_remove_obj_handle(pdd, id);
1014 * Just kunmap and unpin signal BO here. It will be freed in
1015 * kfd_process_free_outstanding_kfd_bos()
1017 static void kfd_process_kunmap_signal_bo(struct kfd_process *p)
1019 struct kfd_process_device *pdd;
1020 struct kfd_node *kdev;
1023 kdev = kfd_device_by_id(GET_GPU_ID(p->signal_handle));
1027 mutex_lock(&p->mutex);
1029 pdd = kfd_get_process_device_data(kdev, p);
1033 mem = kfd_process_device_translate_handle(
1034 pdd, GET_IDR_HANDLE(p->signal_handle));
1038 amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem);
1041 mutex_unlock(&p->mutex);
1044 static void kfd_process_free_outstanding_kfd_bos(struct kfd_process *p)
1048 for (i = 0; i < p->n_pdds; i++)
1049 kfd_process_device_free_bos(p->pdds[i]);
1052 static void kfd_process_destroy_pdds(struct kfd_process *p)
1056 for (i = 0; i < p->n_pdds; i++) {
1057 struct kfd_process_device *pdd = p->pdds[i];
1059 pr_debug("Releasing pdd (topology id %d) for process (pasid 0x%x)\n",
1060 pdd->dev->id, p->pasid);
1062 kfd_process_device_destroy_cwsr_dgpu(pdd);
1063 kfd_process_device_destroy_ib_mem(pdd);
1065 if (pdd->drm_file) {
1066 amdgpu_amdkfd_gpuvm_release_process_vm(
1067 pdd->dev->adev, pdd->drm_priv);
1068 fput(pdd->drm_file);
1071 if (pdd->qpd.cwsr_kaddr && !pdd->qpd.cwsr_base)
1072 free_pages((unsigned long)pdd->qpd.cwsr_kaddr,
1073 get_order(KFD_CWSR_TBA_TMA_SIZE));
1075 idr_destroy(&pdd->alloc_idr);
1077 kfd_free_process_doorbells(pdd->dev->kfd, pdd);
1079 if (pdd->dev->kfd->shared_resources.enable_mes &&
1080 pdd->proc_ctx_cpu_ptr)
1081 amdgpu_amdkfd_free_gtt_mem(pdd->dev->adev,
1084 * before destroying pdd, make sure to report availability
1087 if (pdd->runtime_inuse) {
1088 pm_runtime_mark_last_busy(adev_to_drm(pdd->dev->adev)->dev);
1089 pm_runtime_put_autosuspend(adev_to_drm(pdd->dev->adev)->dev);
1090 pdd->runtime_inuse = false;
1099 static void kfd_process_remove_sysfs(struct kfd_process *p)
1101 struct kfd_process_device *pdd;
1107 sysfs_remove_file(p->kobj, &p->attr_pasid);
1108 kobject_del(p->kobj_queues);
1109 kobject_put(p->kobj_queues);
1110 p->kobj_queues = NULL;
1112 for (i = 0; i < p->n_pdds; i++) {
1115 sysfs_remove_file(p->kobj, &pdd->attr_vram);
1116 sysfs_remove_file(p->kobj, &pdd->attr_sdma);
1118 sysfs_remove_file(pdd->kobj_stats, &pdd->attr_evict);
1119 if (pdd->dev->kfd2kgd->get_cu_occupancy)
1120 sysfs_remove_file(pdd->kobj_stats,
1121 &pdd->attr_cu_occupancy);
1122 kobject_del(pdd->kobj_stats);
1123 kobject_put(pdd->kobj_stats);
1124 pdd->kobj_stats = NULL;
1127 for_each_set_bit(i, p->svms.bitmap_supported, p->n_pdds) {
1130 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_faults);
1131 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_page_in);
1132 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_page_out);
1133 kobject_del(pdd->kobj_counters);
1134 kobject_put(pdd->kobj_counters);
1135 pdd->kobj_counters = NULL;
1138 kobject_del(p->kobj);
1139 kobject_put(p->kobj);
1143 /* No process locking is needed in this function, because the process
1144 * is not findable any more. We must assume that no other thread is
1145 * using it any more, otherwise we couldn't safely free the process
1146 * structure in the end.
1148 static void kfd_process_wq_release(struct work_struct *work)
1150 struct kfd_process *p = container_of(work, struct kfd_process,
1152 struct dma_fence *ef;
1154 kfd_process_dequeue_from_all_devices(p);
1155 pqm_uninit(&p->pqm);
1157 /* Signal the eviction fence after user mode queues are
1158 * destroyed. This allows any BOs to be freed without
1159 * triggering pointless evictions or waiting for fences.
1162 ef = rcu_access_pointer(p->ef);
1164 dma_fence_signal(ef);
1166 kfd_process_remove_sysfs(p);
1168 kfd_process_kunmap_signal_bo(p);
1169 kfd_process_free_outstanding_kfd_bos(p);
1170 svm_range_list_fini(p);
1172 kfd_process_destroy_pdds(p);
1175 kfd_event_free_process(p);
1177 kfd_pasid_free(p->pasid);
1178 mutex_destroy(&p->mutex);
1180 put_task_struct(p->lead_thread);
1185 static void kfd_process_ref_release(struct kref *ref)
1187 struct kfd_process *p = container_of(ref, struct kfd_process, ref);
1189 INIT_WORK(&p->release_work, kfd_process_wq_release);
1190 queue_work(kfd_process_wq, &p->release_work);
1193 static struct mmu_notifier *kfd_process_alloc_notifier(struct mm_struct *mm)
1195 /* This increments p->ref counter if kfd process p exists */
1196 struct kfd_process *p = kfd_lookup_process_by_mm(mm);
1198 return p ? &p->mmu_notifier : ERR_PTR(-ESRCH);
1201 static void kfd_process_free_notifier(struct mmu_notifier *mn)
1203 kfd_unref_process(container_of(mn, struct kfd_process, mmu_notifier));
1206 static void kfd_process_notifier_release_internal(struct kfd_process *p)
1210 cancel_delayed_work_sync(&p->eviction_work);
1211 cancel_delayed_work_sync(&p->restore_work);
1213 for (i = 0; i < p->n_pdds; i++) {
1214 struct kfd_process_device *pdd = p->pdds[i];
1216 /* re-enable GFX OFF since runtime enable with ttmp setup disabled it. */
1217 if (!kfd_dbg_is_rlc_restore_supported(pdd->dev) && p->runtime_info.ttmp_setup)
1218 amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
1221 /* Indicate to other users that MM is no longer valid */
1223 kfd_dbg_trap_disable(p);
1225 if (atomic_read(&p->debugged_process_count) > 0) {
1226 struct kfd_process *target;
1228 int idx = srcu_read_lock(&kfd_processes_srcu);
1230 hash_for_each_rcu(kfd_processes_table, temp, target, kfd_processes) {
1231 if (target->debugger_process && target->debugger_process == p) {
1232 mutex_lock_nested(&target->mutex, 1);
1233 kfd_dbg_trap_disable(target);
1234 mutex_unlock(&target->mutex);
1235 if (atomic_read(&p->debugged_process_count) == 0)
1240 srcu_read_unlock(&kfd_processes_srcu, idx);
1243 mmu_notifier_put(&p->mmu_notifier);
1246 static void kfd_process_notifier_release(struct mmu_notifier *mn,
1247 struct mm_struct *mm)
1249 struct kfd_process *p;
1252 * The kfd_process structure can not be free because the
1253 * mmu_notifier srcu is read locked
1255 p = container_of(mn, struct kfd_process, mmu_notifier);
1256 if (WARN_ON(p->mm != mm))
1259 mutex_lock(&kfd_processes_mutex);
1261 * Do early return if table is empty.
1263 * This could potentially happen if this function is called concurrently
1264 * by mmu_notifier and by kfd_cleanup_pocesses.
1267 if (hash_empty(kfd_processes_table)) {
1268 mutex_unlock(&kfd_processes_mutex);
1271 hash_del_rcu(&p->kfd_processes);
1272 mutex_unlock(&kfd_processes_mutex);
1273 synchronize_srcu(&kfd_processes_srcu);
1275 kfd_process_notifier_release_internal(p);
1278 static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = {
1279 .release = kfd_process_notifier_release,
1280 .alloc_notifier = kfd_process_alloc_notifier,
1281 .free_notifier = kfd_process_free_notifier,
1285 * This code handles the case when driver is being unloaded before all
1286 * mm_struct are released. We need to safely free the kfd_process and
1287 * avoid race conditions with mmu_notifier that might try to free them.
1290 void kfd_cleanup_processes(void)
1292 struct kfd_process *p;
1293 struct hlist_node *p_temp;
1295 HLIST_HEAD(cleanup_list);
1298 * Move all remaining kfd_process from the process table to a
1299 * temp list for processing. Once done, callback from mmu_notifier
1300 * release will not see the kfd_process in the table and do early return,
1301 * avoiding double free issues.
1303 mutex_lock(&kfd_processes_mutex);
1304 hash_for_each_safe(kfd_processes_table, temp, p_temp, p, kfd_processes) {
1305 hash_del_rcu(&p->kfd_processes);
1306 synchronize_srcu(&kfd_processes_srcu);
1307 hlist_add_head(&p->kfd_processes, &cleanup_list);
1309 mutex_unlock(&kfd_processes_mutex);
1311 hlist_for_each_entry_safe(p, p_temp, &cleanup_list, kfd_processes)
1312 kfd_process_notifier_release_internal(p);
1315 * Ensures that all outstanding free_notifier get called, triggering
1316 * the release of the kfd_process struct.
1318 mmu_notifier_synchronize();
1321 int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep)
1323 unsigned long offset;
1329 for (i = 0; i < p->n_pdds; i++) {
1330 struct kfd_node *dev = p->pdds[i]->dev;
1331 struct qcm_process_device *qpd = &p->pdds[i]->qpd;
1333 if (!dev->kfd->cwsr_enabled || qpd->cwsr_kaddr || qpd->cwsr_base)
1336 offset = KFD_MMAP_TYPE_RESERVED_MEM | KFD_MMAP_GPU_ID(dev->id);
1337 qpd->tba_addr = (int64_t)vm_mmap(filep, 0,
1338 KFD_CWSR_TBA_TMA_SIZE, PROT_READ | PROT_EXEC,
1339 MAP_SHARED, offset);
1341 if (IS_ERR_VALUE(qpd->tba_addr)) {
1342 int err = qpd->tba_addr;
1344 dev_err(dev->adev->dev,
1345 "Failure to set tba address. error %d.\n", err);
1347 qpd->cwsr_kaddr = NULL;
1351 memcpy(qpd->cwsr_kaddr, dev->kfd->cwsr_isa, dev->kfd->cwsr_isa_size);
1353 kfd_process_set_trap_debug_flag(qpd, p->debug_trap_enabled);
1355 qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1356 pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1357 qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1365 static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd)
1367 struct kfd_node *dev = pdd->dev;
1368 struct qcm_process_device *qpd = &pdd->qpd;
1369 uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT
1370 | KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE
1371 | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
1372 struct kgd_mem *mem;
1376 if (!dev->kfd->cwsr_enabled || qpd->cwsr_kaddr || !qpd->cwsr_base)
1379 /* cwsr_base is only set for dGPU */
1380 ret = kfd_process_alloc_gpuvm(pdd, qpd->cwsr_base,
1381 KFD_CWSR_TBA_TMA_SIZE, flags, &mem, &kaddr);
1385 qpd->cwsr_mem = mem;
1386 qpd->cwsr_kaddr = kaddr;
1387 qpd->tba_addr = qpd->cwsr_base;
1389 memcpy(qpd->cwsr_kaddr, dev->kfd->cwsr_isa, dev->kfd->cwsr_isa_size);
1391 kfd_process_set_trap_debug_flag(&pdd->qpd,
1392 pdd->process->debug_trap_enabled);
1394 qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1395 pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1396 qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1401 static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd)
1403 struct kfd_node *dev = pdd->dev;
1404 struct qcm_process_device *qpd = &pdd->qpd;
1406 if (!dev->kfd->cwsr_enabled || !qpd->cwsr_kaddr || !qpd->cwsr_base)
1409 kfd_process_free_gpuvm(qpd->cwsr_mem, pdd, &qpd->cwsr_kaddr);
1412 void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
1416 if (qpd->cwsr_kaddr) {
1417 /* KFD trap handler is bound, record as second-level TBA/TMA
1418 * in first-level TMA. First-level trap will jump to second.
1421 (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1425 /* No trap handler bound, bind as first-level TBA/TMA. */
1426 qpd->tba_addr = tba_addr;
1427 qpd->tma_addr = tma_addr;
1431 bool kfd_process_xnack_mode(struct kfd_process *p, bool supported)
1435 /* On most GFXv9 GPUs, the retry mode in the SQ must match the
1436 * boot time retry setting. Mixing processes with different
1437 * XNACK/retry settings can hang the GPU.
1439 * Different GPUs can have different noretry settings depending
1440 * on HW bugs or limitations. We need to find at least one
1441 * XNACK mode for this process that's compatible with all GPUs.
1442 * Fortunately GPUs with retry enabled (noretry=0) can run code
1443 * built for XNACK-off. On GFXv9 it may perform slower.
1445 * Therefore applications built for XNACK-off can always be
1446 * supported and will be our fallback if any GPU does not
1449 for (i = 0; i < p->n_pdds; i++) {
1450 struct kfd_node *dev = p->pdds[i]->dev;
1452 /* Only consider GFXv9 and higher GPUs. Older GPUs don't
1453 * support the SVM APIs and don't need to be considered
1454 * for the XNACK mode selection.
1456 if (!KFD_IS_SOC15(dev))
1458 /* Aldebaran can always support XNACK because it can support
1459 * per-process XNACK mode selection. But let the dev->noretry
1460 * setting still influence the default XNACK mode.
1462 if (supported && KFD_SUPPORT_XNACK_PER_PROCESS(dev)) {
1463 if (!amdgpu_sriov_xnack_support(dev->kfd->adev)) {
1464 pr_debug("SRIOV platform xnack not supported\n");
1470 /* GFXv10 and later GPUs do not support shader preemption
1471 * during page faults. This can lead to poor QoS for queue
1472 * management and memory-manager-related preemptions or
1475 if (KFD_GC_VERSION(dev) >= IP_VERSION(10, 1, 1))
1478 if (dev->kfd->noretry)
1485 void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd,
1488 if (qpd->cwsr_kaddr) {
1490 (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1496 * On return the kfd_process is fully operational and will be freed when the
1499 static struct kfd_process *create_process(const struct task_struct *thread)
1501 struct kfd_process *process;
1502 struct mmu_notifier *mn;
1505 process = kzalloc(sizeof(*process), GFP_KERNEL);
1507 goto err_alloc_process;
1509 kref_init(&process->ref);
1510 mutex_init(&process->mutex);
1511 process->mm = thread->mm;
1512 process->lead_thread = thread->group_leader;
1513 process->n_pdds = 0;
1514 process->queues_paused = false;
1515 INIT_DELAYED_WORK(&process->eviction_work, evict_process_worker);
1516 INIT_DELAYED_WORK(&process->restore_work, restore_process_worker);
1517 process->last_restore_timestamp = get_jiffies_64();
1518 err = kfd_event_init_process(process);
1520 goto err_event_init;
1521 process->is_32bit_user_mode = in_compat_syscall();
1522 process->debug_trap_enabled = false;
1523 process->debugger_process = NULL;
1524 process->exception_enable_mask = 0;
1525 atomic_set(&process->debugged_process_count, 0);
1526 sema_init(&process->runtime_enable_sema, 0);
1528 process->pasid = kfd_pasid_alloc();
1529 if (process->pasid == 0) {
1531 goto err_alloc_pasid;
1534 err = pqm_init(&process->pqm, process);
1536 goto err_process_pqm_init;
1538 /* init process apertures*/
1539 err = kfd_init_apertures(process);
1541 goto err_init_apertures;
1543 /* Check XNACK support after PDDs are created in kfd_init_apertures */
1544 process->xnack_enabled = kfd_process_xnack_mode(process, false);
1546 err = svm_range_list_init(process);
1548 goto err_init_svm_range_list;
1550 /* alloc_notifier needs to find the process in the hash table */
1551 hash_add_rcu(kfd_processes_table, &process->kfd_processes,
1552 (uintptr_t)process->mm);
1554 /* Avoid free_notifier to start kfd_process_wq_release if
1555 * mmu_notifier_get failed because of pending signal.
1557 kref_get(&process->ref);
1559 /* MMU notifier registration must be the last call that can fail
1560 * because after this point we cannot unwind the process creation.
1561 * After this point, mmu_notifier_put will trigger the cleanup by
1562 * dropping the last process reference in the free_notifier.
1564 mn = mmu_notifier_get(&kfd_process_mmu_notifier_ops, process->mm);
1567 goto err_register_notifier;
1569 BUG_ON(mn != &process->mmu_notifier);
1571 kfd_unref_process(process);
1572 get_task_struct(process->lead_thread);
1574 INIT_WORK(&process->debug_event_workarea, debug_event_write_work_handler);
1578 err_register_notifier:
1579 hash_del_rcu(&process->kfd_processes);
1580 svm_range_list_fini(process);
1581 err_init_svm_range_list:
1582 kfd_process_free_outstanding_kfd_bos(process);
1583 kfd_process_destroy_pdds(process);
1585 pqm_uninit(&process->pqm);
1586 err_process_pqm_init:
1587 kfd_pasid_free(process->pasid);
1589 kfd_event_free_process(process);
1591 mutex_destroy(&process->mutex);
1594 return ERR_PTR(err);
1597 struct kfd_process_device *kfd_get_process_device_data(struct kfd_node *dev,
1598 struct kfd_process *p)
1602 for (i = 0; i < p->n_pdds; i++)
1603 if (p->pdds[i]->dev == dev)
1609 struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev,
1610 struct kfd_process *p)
1612 struct kfd_process_device *pdd = NULL;
1614 if (WARN_ON_ONCE(p->n_pdds >= MAX_GPU_INSTANCE))
1616 pdd = kzalloc(sizeof(*pdd), GFP_KERNEL);
1621 INIT_LIST_HEAD(&pdd->qpd.queues_list);
1622 INIT_LIST_HEAD(&pdd->qpd.priv_queue_list);
1623 pdd->qpd.dqm = dev->dqm;
1624 pdd->qpd.pqm = &p->pqm;
1625 pdd->qpd.evicted = 0;
1626 pdd->qpd.mapped_gws_queue = false;
1628 pdd->bound = PDD_UNBOUND;
1629 pdd->already_dequeued = false;
1630 pdd->runtime_inuse = false;
1631 atomic64_set(&pdd->vram_usage, 0);
1632 pdd->sdma_past_activity_counter = 0;
1633 pdd->user_gpu_id = dev->id;
1634 atomic64_set(&pdd->evict_duration_counter, 0);
1636 p->pdds[p->n_pdds++] = pdd;
1637 if (kfd_dbg_is_per_vmid_supported(pdd->dev))
1638 pdd->spi_dbg_override = pdd->dev->kfd2kgd->disable_debug_trap(
1643 /* Init idr used for memory handle translation */
1644 idr_init(&pdd->alloc_idr);
1650 * kfd_process_device_init_vm - Initialize a VM for a process-device
1652 * @pdd: The process-device
1653 * @drm_file: Optional pointer to a DRM file descriptor
1655 * If @drm_file is specified, it will be used to acquire the VM from
1656 * that file descriptor. If successful, the @pdd takes ownership of
1657 * the file descriptor.
1659 * If @drm_file is NULL, a new VM is created.
1661 * Returns 0 on success, -errno on failure.
1663 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
1664 struct file *drm_file)
1666 struct amdgpu_fpriv *drv_priv;
1667 struct amdgpu_vm *avm;
1668 struct kfd_process *p;
1669 struct dma_fence *ef;
1670 struct kfd_node *dev;
1679 ret = amdgpu_file_to_fpriv(drm_file, &drv_priv);
1682 avm = &drv_priv->vm;
1687 ret = amdgpu_amdkfd_gpuvm_acquire_process_vm(dev->adev, avm,
1688 &p->kgd_process_info,
1689 p->ef ? NULL : &ef);
1691 dev_err(dev->adev->dev, "Failed to create process VM object\n");
1696 RCU_INIT_POINTER(p->ef, ef);
1698 pdd->drm_priv = drm_file->private_data;
1700 ret = kfd_process_device_reserve_ib_mem(pdd);
1702 goto err_reserve_ib_mem;
1703 ret = kfd_process_device_init_cwsr_dgpu(pdd);
1707 ret = amdgpu_amdkfd_gpuvm_set_vm_pasid(dev->adev, avm, p->pasid);
1711 pdd->drm_file = drm_file;
1716 kfd_process_device_destroy_cwsr_dgpu(pdd);
1718 kfd_process_device_destroy_ib_mem(pdd);
1720 pdd->drm_priv = NULL;
1721 amdgpu_amdkfd_gpuvm_destroy_cb(dev->adev, avm);
1727 * Direct the IOMMU to bind the process (specifically the pasid->mm)
1729 * Unbinding occurs when the process dies or the device is removed.
1731 * Assumes that the process lock is held.
1733 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_node *dev,
1734 struct kfd_process *p)
1736 struct kfd_process_device *pdd;
1739 pdd = kfd_get_process_device_data(dev, p);
1741 dev_err(dev->adev->dev, "Process device data doesn't exist\n");
1742 return ERR_PTR(-ENOMEM);
1746 return ERR_PTR(-ENODEV);
1749 * signal runtime-pm system to auto resume and prevent
1750 * further runtime suspend once device pdd is created until
1753 if (!pdd->runtime_inuse) {
1754 err = pm_runtime_get_sync(adev_to_drm(dev->adev)->dev);
1756 pm_runtime_put_autosuspend(adev_to_drm(dev->adev)->dev);
1757 return ERR_PTR(err);
1762 * make sure that runtime_usage counter is incremented just once
1765 pdd->runtime_inuse = true;
1770 /* Create specific handle mapped to mem from process local memory idr
1771 * Assumes that the process lock is held.
1773 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
1776 return idr_alloc(&pdd->alloc_idr, mem, 0, 0, GFP_KERNEL);
1779 /* Translate specific handle from process local memory idr
1780 * Assumes that the process lock is held.
1782 void *kfd_process_device_translate_handle(struct kfd_process_device *pdd,
1788 return idr_find(&pdd->alloc_idr, handle);
1791 /* Remove specific handle from process local memory idr
1792 * Assumes that the process lock is held.
1794 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
1798 idr_remove(&pdd->alloc_idr, handle);
1801 /* This increments the process->ref counter. */
1802 struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid)
1804 struct kfd_process *p, *ret_p = NULL;
1807 int idx = srcu_read_lock(&kfd_processes_srcu);
1809 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1810 if (p->pasid == pasid) {
1817 srcu_read_unlock(&kfd_processes_srcu, idx);
1822 /* This increments the process->ref counter. */
1823 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm)
1825 struct kfd_process *p;
1827 int idx = srcu_read_lock(&kfd_processes_srcu);
1829 p = find_process_by_mm(mm);
1833 srcu_read_unlock(&kfd_processes_srcu, idx);
1838 /* kfd_process_evict_queues - Evict all user queues of a process
1840 * Eviction is reference-counted per process-device. This means multiple
1841 * evictions from different sources can be nested safely.
1843 int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger)
1847 unsigned int n_evicted = 0;
1849 for (i = 0; i < p->n_pdds; i++) {
1850 struct kfd_process_device *pdd = p->pdds[i];
1851 struct device *dev = pdd->dev->adev->dev;
1853 kfd_smi_event_queue_eviction(pdd->dev, p->lead_thread->pid,
1856 r = pdd->dev->dqm->ops.evict_process_queues(pdd->dev->dqm,
1858 /* evict return -EIO if HWS is hang or asic is resetting, in this case
1859 * we would like to set all the queues to be in evicted state to prevent
1860 * them been add back since they actually not be saved right now.
1862 if (r && r != -EIO) {
1863 dev_err(dev, "Failed to evict process queues\n");
1868 pdd->dev->dqm->is_hws_hang = false;
1874 /* To keep state consistent, roll back partial eviction by
1877 for (i = 0; i < p->n_pdds; i++) {
1878 struct kfd_process_device *pdd = p->pdds[i];
1883 kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid);
1885 if (pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1887 dev_err(pdd->dev->adev->dev,
1888 "Failed to restore queues\n");
1896 /* kfd_process_restore_queues - Restore all user queues of a process */
1897 int kfd_process_restore_queues(struct kfd_process *p)
1902 for (i = 0; i < p->n_pdds; i++) {
1903 struct kfd_process_device *pdd = p->pdds[i];
1904 struct device *dev = pdd->dev->adev->dev;
1906 kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid);
1908 r = pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1911 dev_err(dev, "Failed to restore process queues\n");
1920 int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id)
1924 for (i = 0; i < p->n_pdds; i++)
1925 if (p->pdds[i] && gpu_id == p->pdds[i]->user_gpu_id)
1931 kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node,
1932 uint32_t *gpuid, uint32_t *gpuidx)
1936 for (i = 0; i < p->n_pdds; i++)
1937 if (p->pdds[i] && p->pdds[i]->dev == node) {
1938 *gpuid = p->pdds[i]->user_gpu_id;
1945 static int signal_eviction_fence(struct kfd_process *p)
1947 struct dma_fence *ef;
1951 ef = dma_fence_get_rcu_safe(&p->ef);
1956 ret = dma_fence_signal(ef);
1962 static void evict_process_worker(struct work_struct *work)
1965 struct kfd_process *p;
1966 struct delayed_work *dwork;
1968 dwork = to_delayed_work(work);
1970 /* Process termination destroys this worker thread. So during the
1971 * lifetime of this thread, kfd_process p will be valid
1973 p = container_of(dwork, struct kfd_process, eviction_work);
1975 pr_debug("Started evicting pasid 0x%x\n", p->pasid);
1976 ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_TRIGGER_TTM);
1978 /* If another thread already signaled the eviction fence,
1979 * they are responsible stopping the queues and scheduling
1982 if (signal_eviction_fence(p) ||
1983 mod_delayed_work(kfd_restore_wq, &p->restore_work,
1984 msecs_to_jiffies(PROCESS_RESTORE_TIME_MS)))
1985 kfd_process_restore_queues(p);
1987 pr_debug("Finished evicting pasid 0x%x\n", p->pasid);
1989 pr_err("Failed to evict queues of pasid 0x%x\n", p->pasid);
1992 static int restore_process_helper(struct kfd_process *p)
1996 /* VMs may not have been acquired yet during debugging. */
1997 if (p->kgd_process_info) {
1998 ret = amdgpu_amdkfd_gpuvm_restore_process_bos(
1999 p->kgd_process_info, &p->ef);
2004 ret = kfd_process_restore_queues(p);
2006 pr_debug("Finished restoring pasid 0x%x\n", p->pasid);
2008 pr_err("Failed to restore queues of pasid 0x%x\n", p->pasid);
2013 static void restore_process_worker(struct work_struct *work)
2015 struct delayed_work *dwork;
2016 struct kfd_process *p;
2019 dwork = to_delayed_work(work);
2021 /* Process termination destroys this worker thread. So during the
2022 * lifetime of this thread, kfd_process p will be valid
2024 p = container_of(dwork, struct kfd_process, restore_work);
2025 pr_debug("Started restoring pasid 0x%x\n", p->pasid);
2027 /* Setting last_restore_timestamp before successful restoration.
2028 * Otherwise this would have to be set by KGD (restore_process_bos)
2029 * before KFD BOs are unreserved. If not, the process can be evicted
2030 * again before the timestamp is set.
2031 * If restore fails, the timestamp will be set again in the next
2032 * attempt. This would mean that the minimum GPU quanta would be
2033 * PROCESS_ACTIVE_TIME_MS - (time to execute the following two
2037 p->last_restore_timestamp = get_jiffies_64();
2039 ret = restore_process_helper(p);
2041 pr_debug("Failed to restore BOs of pasid 0x%x, retry after %d ms\n",
2042 p->pasid, PROCESS_BACK_OFF_TIME_MS);
2043 if (mod_delayed_work(kfd_restore_wq, &p->restore_work,
2044 msecs_to_jiffies(PROCESS_RESTORE_TIME_MS)))
2045 kfd_process_restore_queues(p);
2049 void kfd_suspend_all_processes(void)
2051 struct kfd_process *p;
2053 int idx = srcu_read_lock(&kfd_processes_srcu);
2055 WARN(debug_evictions, "Evicting all processes");
2056 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2057 if (kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_TRIGGER_SUSPEND))
2058 pr_err("Failed to suspend process 0x%x\n", p->pasid);
2059 signal_eviction_fence(p);
2061 srcu_read_unlock(&kfd_processes_srcu, idx);
2064 int kfd_resume_all_processes(void)
2066 struct kfd_process *p;
2068 int ret = 0, idx = srcu_read_lock(&kfd_processes_srcu);
2070 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2071 if (restore_process_helper(p)) {
2072 pr_err("Restore process %d failed during resume\n",
2077 srcu_read_unlock(&kfd_processes_srcu, idx);
2081 int kfd_reserved_mem_mmap(struct kfd_node *dev, struct kfd_process *process,
2082 struct vm_area_struct *vma)
2084 struct kfd_process_device *pdd;
2085 struct qcm_process_device *qpd;
2087 if ((vma->vm_end - vma->vm_start) != KFD_CWSR_TBA_TMA_SIZE) {
2088 dev_err(dev->adev->dev, "Incorrect CWSR mapping size.\n");
2092 pdd = kfd_get_process_device_data(dev, process);
2097 qpd->cwsr_kaddr = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
2098 get_order(KFD_CWSR_TBA_TMA_SIZE));
2099 if (!qpd->cwsr_kaddr) {
2100 dev_err(dev->adev->dev,
2101 "Error allocating per process CWSR buffer.\n");
2105 vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND
2106 | VM_NORESERVE | VM_DONTDUMP | VM_PFNMAP);
2107 /* Mapping pages to user process */
2108 return remap_pfn_range(vma, vma->vm_start,
2109 PFN_DOWN(__pa(qpd->cwsr_kaddr)),
2110 KFD_CWSR_TBA_TMA_SIZE, vma->vm_page_prot);
2113 /* assumes caller holds process lock. */
2114 int kfd_process_drain_interrupts(struct kfd_process_device *pdd)
2116 uint32_t irq_drain_fence[8];
2117 uint8_t node_id = 0;
2120 if (!KFD_IS_SOC15(pdd->dev))
2123 pdd->process->irq_drain_is_open = true;
2125 memset(irq_drain_fence, 0, sizeof(irq_drain_fence));
2126 irq_drain_fence[0] = (KFD_IRQ_FENCE_SOURCEID << 8) |
2127 KFD_IRQ_FENCE_CLIENTID;
2128 irq_drain_fence[3] = pdd->process->pasid;
2131 * For GFX 9.4.3/9.5.0, send the NodeId also in IH cookie DW[3]
2133 if (KFD_GC_VERSION(pdd->dev->kfd) == IP_VERSION(9, 4, 3) ||
2134 KFD_GC_VERSION(pdd->dev->kfd) == IP_VERSION(9, 4, 4) ||
2135 KFD_GC_VERSION(pdd->dev->kfd) == IP_VERSION(9, 5, 0)) {
2136 node_id = ffs(pdd->dev->interrupt_bitmap) - 1;
2137 irq_drain_fence[3] |= node_id << 16;
2140 /* ensure stale irqs scheduled KFD interrupts and send drain fence. */
2141 if (amdgpu_amdkfd_send_close_event_drain_irq(pdd->dev->adev,
2143 pdd->process->irq_drain_is_open = false;
2147 r = wait_event_interruptible(pdd->process->wait_irq_drain,
2148 !READ_ONCE(pdd->process->irq_drain_is_open));
2150 pdd->process->irq_drain_is_open = false;
2155 void kfd_process_close_interrupt_drain(unsigned int pasid)
2157 struct kfd_process *p;
2159 p = kfd_lookup_process_by_pasid(pasid);
2164 WRITE_ONCE(p->irq_drain_is_open, false);
2165 wake_up_all(&p->wait_irq_drain);
2166 kfd_unref_process(p);
2169 struct send_exception_work_handler_workarea {
2170 struct work_struct work;
2171 struct kfd_process *p;
2172 unsigned int queue_id;
2173 uint64_t error_reason;
2176 static void send_exception_work_handler(struct work_struct *work)
2178 struct send_exception_work_handler_workarea *workarea;
2179 struct kfd_process *p;
2181 struct mm_struct *mm;
2182 struct kfd_context_save_area_header __user *csa_header;
2183 uint64_t __user *err_payload_ptr;
2187 workarea = container_of(work,
2188 struct send_exception_work_handler_workarea,
2192 mm = get_task_mm(p->lead_thread);
2199 q = pqm_get_user_queue(&p->pqm, workarea->queue_id);
2204 csa_header = (void __user *)q->properties.ctx_save_restore_area_address;
2206 get_user(err_payload_ptr, (uint64_t __user **)&csa_header->err_payload_addr);
2207 get_user(cur_err, err_payload_ptr);
2208 cur_err |= workarea->error_reason;
2209 put_user(cur_err, err_payload_ptr);
2210 get_user(ev_id, &csa_header->err_event_id);
2212 kfd_set_event(p, ev_id);
2215 kthread_unuse_mm(mm);
2219 int kfd_send_exception_to_runtime(struct kfd_process *p,
2220 unsigned int queue_id,
2221 uint64_t error_reason)
2223 struct send_exception_work_handler_workarea worker;
2225 INIT_WORK_ONSTACK(&worker.work, send_exception_work_handler);
2228 worker.queue_id = queue_id;
2229 worker.error_reason = error_reason;
2231 schedule_work(&worker.work);
2232 flush_work(&worker.work);
2233 destroy_work_on_stack(&worker.work);
2238 struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *p, uint32_t gpu_id)
2243 for (i = 0; i < p->n_pdds; i++) {
2244 struct kfd_process_device *pdd = p->pdds[i];
2246 if (pdd->user_gpu_id == gpu_id)
2253 int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id)
2260 for (i = 0; i < p->n_pdds; i++) {
2261 struct kfd_process_device *pdd = p->pdds[i];
2263 if (pdd->dev->id == actual_gpu_id)
2264 return pdd->user_gpu_id;
2269 #if defined(CONFIG_DEBUG_FS)
2271 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data)
2273 struct kfd_process *p;
2277 int idx = srcu_read_lock(&kfd_processes_srcu);
2279 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2280 seq_printf(m, "Process %d PASID 0x%x:\n",
2281 p->lead_thread->tgid, p->pasid);
2283 mutex_lock(&p->mutex);
2284 r = pqm_debugfs_mqds(m, &p->pqm);
2285 mutex_unlock(&p->mutex);
2291 srcu_read_unlock(&kfd_processes_srcu, idx);