2 * Copyright 2014 Advanced Micro Devices, Inc.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20 * OTHER DEALINGS IN THE SOFTWARE.
23 #include <linux/mutex.h>
24 #include <linux/log2.h>
25 #include <linux/sched.h>
26 #include <linux/sched/mm.h>
27 #include <linux/sched/task.h>
28 #include <linux/mmu_context.h>
29 #include <linux/slab.h>
30 #include <linux/amd-iommu.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"
43 #include "kfd_dbgmgr.h"
44 #include "kfd_iommu.h"
47 * List of struct kfd_process (field kfd_process).
48 * Unique/indexed by mm_struct*
50 DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
51 static DEFINE_MUTEX(kfd_processes_mutex);
53 DEFINE_SRCU(kfd_processes_srcu);
55 /* For process termination handling */
56 static struct workqueue_struct *kfd_process_wq;
58 /* Ordered, single-threaded workqueue for restoring evicted
59 * processes. Restoring multiple processes concurrently under memory
60 * pressure can lead to processes blocking each other from validating
61 * their BOs and result in a live-lock situation where processes
62 * remain evicted indefinitely.
64 static struct workqueue_struct *kfd_restore_wq;
66 static struct kfd_process *find_process(const struct task_struct *thread);
67 static void kfd_process_ref_release(struct kref *ref);
68 static struct kfd_process *create_process(const struct task_struct *thread);
69 static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep);
71 static void evict_process_worker(struct work_struct *work);
72 static void restore_process_worker(struct work_struct *work);
74 struct kfd_procfs_tree {
78 static struct kfd_procfs_tree procfs;
81 * Structure for SDMA activity tracking
83 struct kfd_sdma_activity_handler_workarea {
84 struct work_struct sdma_activity_work;
85 struct kfd_process_device *pdd;
86 uint64_t sdma_activity_counter;
89 struct temp_sdma_queue_list {
92 unsigned int queue_id;
93 struct list_head list;
96 static void kfd_sdma_activity_worker(struct work_struct *work)
98 struct kfd_sdma_activity_handler_workarea *workarea;
99 struct kfd_process_device *pdd;
101 struct mm_struct *mm;
103 struct qcm_process_device *qpd;
104 struct device_queue_manager *dqm;
106 struct temp_sdma_queue_list sdma_q_list;
107 struct temp_sdma_queue_list *sdma_q, *next;
109 workarea = container_of(work, struct kfd_sdma_activity_handler_workarea,
122 * Total SDMA activity is current SDMA activity + past SDMA activity
123 * Past SDMA count is stored in pdd.
124 * To get the current activity counters for all active SDMA queues,
125 * we loop over all SDMA queues and get their counts from user-space.
127 * We cannot call get_user() with dqm_lock held as it can cause
128 * a circular lock dependency situation. To read the SDMA stats,
129 * we need to do the following:
131 * 1. Create a temporary list of SDMA queue nodes from the qpd->queues_list,
132 * with dqm_lock/dqm_unlock().
133 * 2. Call get_user() for each node in temporary list without dqm_lock.
134 * Save the SDMA count for each node and also add the count to the total
135 * SDMA count counter.
136 * Its possible, during this step, a few SDMA queue nodes got deleted
137 * from the qpd->queues_list.
138 * 3. Do a second pass over qpd->queues_list to check if any nodes got deleted.
139 * If any node got deleted, its SDMA count would be captured in the sdma
140 * past activity counter. So subtract the SDMA counter stored in step 2
141 * for this node from the total SDMA count.
143 INIT_LIST_HEAD(&sdma_q_list.list);
146 * Create the temp list of all SDMA queues
150 list_for_each_entry(q, &qpd->queues_list, list) {
151 if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
152 (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
155 sdma_q = kzalloc(sizeof(struct temp_sdma_queue_list), GFP_KERNEL);
161 INIT_LIST_HEAD(&sdma_q->list);
162 sdma_q->rptr = (uint64_t)q->properties.read_ptr;
163 sdma_q->queue_id = q->properties.queue_id;
164 list_add_tail(&sdma_q->list, &sdma_q_list.list);
168 * If the temp list is empty, then no SDMA queues nodes were found in
169 * qpd->queues_list. Return the past activity count as the total sdma
172 if (list_empty(&sdma_q_list.list)) {
173 workarea->sdma_activity_counter = pdd->sdma_past_activity_counter;
181 * Get the usage count for each SDMA queue in temp_list.
183 mm = get_task_mm(pdd->process->lead_thread);
189 list_for_each_entry(sdma_q, &sdma_q_list.list, list) {
191 ret = read_sdma_queue_counter(sdma_q->rptr, &val);
193 pr_debug("Failed to read SDMA queue active counter for queue id: %d",
196 sdma_q->sdma_val = val;
197 workarea->sdma_activity_counter += val;
201 kthread_unuse_mm(mm);
205 * Do a second iteration over qpd_queues_list to check if any SDMA
206 * nodes got deleted while fetching SDMA counter.
210 workarea->sdma_activity_counter += pdd->sdma_past_activity_counter;
212 list_for_each_entry(q, &qpd->queues_list, list) {
213 if (list_empty(&sdma_q_list.list))
216 if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
217 (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
220 list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
221 if (((uint64_t)q->properties.read_ptr == sdma_q->rptr) &&
222 (sdma_q->queue_id == q->properties.queue_id)) {
223 list_del(&sdma_q->list);
233 * If temp list is not empty, it implies some queues got deleted
234 * from qpd->queues_list during SDMA usage read. Subtract the SDMA
235 * count for each node from the total SDMA count.
237 list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
238 workarea->sdma_activity_counter -= sdma_q->sdma_val;
239 list_del(&sdma_q->list);
246 list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
247 list_del(&sdma_q->list);
252 static ssize_t kfd_procfs_show(struct kobject *kobj, struct attribute *attr,
255 if (strcmp(attr->name, "pasid") == 0) {
256 struct kfd_process *p = container_of(attr, struct kfd_process,
259 return snprintf(buffer, PAGE_SIZE, "%d\n", p->pasid);
260 } else if (strncmp(attr->name, "vram_", 5) == 0) {
261 struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
263 return snprintf(buffer, PAGE_SIZE, "%llu\n", READ_ONCE(pdd->vram_usage));
264 } else if (strncmp(attr->name, "sdma_", 5) == 0) {
265 struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
267 struct kfd_sdma_activity_handler_workarea sdma_activity_work_handler;
269 INIT_WORK(&sdma_activity_work_handler.sdma_activity_work,
270 kfd_sdma_activity_worker);
272 sdma_activity_work_handler.pdd = pdd;
274 schedule_work(&sdma_activity_work_handler.sdma_activity_work);
276 flush_work(&sdma_activity_work_handler.sdma_activity_work);
278 return snprintf(buffer, PAGE_SIZE, "%llu\n",
279 (sdma_activity_work_handler.sdma_activity_counter)/
280 SDMA_ACTIVITY_DIVISOR);
282 pr_err("Invalid attribute");
289 static void kfd_procfs_kobj_release(struct kobject *kobj)
294 static const struct sysfs_ops kfd_procfs_ops = {
295 .show = kfd_procfs_show,
298 static struct kobj_type procfs_type = {
299 .release = kfd_procfs_kobj_release,
300 .sysfs_ops = &kfd_procfs_ops,
303 void kfd_procfs_init(void)
307 procfs.kobj = kfd_alloc_struct(procfs.kobj);
311 ret = kobject_init_and_add(procfs.kobj, &procfs_type,
312 &kfd_device->kobj, "proc");
314 pr_warn("Could not create procfs proc folder");
315 /* If we fail to create the procfs, clean up */
316 kfd_procfs_shutdown();
320 void kfd_procfs_shutdown(void)
323 kobject_del(procfs.kobj);
324 kobject_put(procfs.kobj);
329 static ssize_t kfd_procfs_queue_show(struct kobject *kobj,
330 struct attribute *attr, char *buffer)
332 struct queue *q = container_of(kobj, struct queue, kobj);
334 if (!strcmp(attr->name, "size"))
335 return snprintf(buffer, PAGE_SIZE, "%llu",
336 q->properties.queue_size);
337 else if (!strcmp(attr->name, "type"))
338 return snprintf(buffer, PAGE_SIZE, "%d", q->properties.type);
339 else if (!strcmp(attr->name, "gpuid"))
340 return snprintf(buffer, PAGE_SIZE, "%u", q->device->id);
342 pr_err("Invalid attribute");
347 static struct attribute attr_queue_size = {
349 .mode = KFD_SYSFS_FILE_MODE
352 static struct attribute attr_queue_type = {
354 .mode = KFD_SYSFS_FILE_MODE
357 static struct attribute attr_queue_gpuid = {
359 .mode = KFD_SYSFS_FILE_MODE
362 static struct attribute *procfs_queue_attrs[] = {
369 static const struct sysfs_ops procfs_queue_ops = {
370 .show = kfd_procfs_queue_show,
373 static struct kobj_type procfs_queue_type = {
374 .sysfs_ops = &procfs_queue_ops,
375 .default_attrs = procfs_queue_attrs,
378 int kfd_procfs_add_queue(struct queue *q)
380 struct kfd_process *proc;
383 if (!q || !q->process)
387 /* Create proc/<pid>/queues/<queue id> folder */
388 if (!proc->kobj_queues)
390 ret = kobject_init_and_add(&q->kobj, &procfs_queue_type,
391 proc->kobj_queues, "%u", q->properties.queue_id);
393 pr_warn("Creating proc/<pid>/queues/%u failed",
394 q->properties.queue_id);
395 kobject_put(&q->kobj);
402 static int kfd_sysfs_create_file(struct kfd_process *p, struct attribute *attr,
407 if (!p || !attr || !name)
411 attr->mode = KFD_SYSFS_FILE_MODE;
412 sysfs_attr_init(attr);
414 ret = sysfs_create_file(p->kobj, attr);
419 static int kfd_procfs_add_sysfs_files(struct kfd_process *p)
422 struct kfd_process_device *pdd;
431 * Create sysfs files for each GPU:
432 * - proc/<pid>/vram_<gpuid>
433 * - proc/<pid>/sdma_<gpuid>
435 list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
436 snprintf(pdd->vram_filename, MAX_SYSFS_FILENAME_LEN, "vram_%u",
438 ret = kfd_sysfs_create_file(p, &pdd->attr_vram, pdd->vram_filename);
440 pr_warn("Creating vram usage for gpu id %d failed",
443 snprintf(pdd->sdma_filename, MAX_SYSFS_FILENAME_LEN, "sdma_%u",
445 ret = kfd_sysfs_create_file(p, &pdd->attr_sdma, pdd->sdma_filename);
447 pr_warn("Creating sdma usage for gpu id %d failed",
455 void kfd_procfs_del_queue(struct queue *q)
460 kobject_del(&q->kobj);
461 kobject_put(&q->kobj);
464 int kfd_process_create_wq(void)
467 kfd_process_wq = alloc_workqueue("kfd_process_wq", 0, 0);
469 kfd_restore_wq = alloc_ordered_workqueue("kfd_restore_wq", 0);
471 if (!kfd_process_wq || !kfd_restore_wq) {
472 kfd_process_destroy_wq();
479 void kfd_process_destroy_wq(void)
481 if (kfd_process_wq) {
482 destroy_workqueue(kfd_process_wq);
483 kfd_process_wq = NULL;
485 if (kfd_restore_wq) {
486 destroy_workqueue(kfd_restore_wq);
487 kfd_restore_wq = NULL;
491 static void kfd_process_free_gpuvm(struct kgd_mem *mem,
492 struct kfd_process_device *pdd)
494 struct kfd_dev *dev = pdd->dev;
496 amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(dev->kgd, mem, pdd->vm);
497 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, mem, NULL);
500 /* kfd_process_alloc_gpuvm - Allocate GPU VM for the KFD process
501 * This function should be only called right after the process
502 * is created and when kfd_processes_mutex is still being held
503 * to avoid concurrency. Because of that exclusiveness, we do
504 * not need to take p->mutex.
506 static int kfd_process_alloc_gpuvm(struct kfd_process_device *pdd,
507 uint64_t gpu_va, uint32_t size,
508 uint32_t flags, void **kptr)
510 struct kfd_dev *kdev = pdd->dev;
511 struct kgd_mem *mem = NULL;
515 err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(kdev->kgd, gpu_va, size,
516 pdd->vm, &mem, NULL, flags);
520 err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(kdev->kgd, mem, pdd->vm);
524 err = amdgpu_amdkfd_gpuvm_sync_memory(kdev->kgd, mem, true);
526 pr_debug("Sync memory failed, wait interrupted by user signal\n");
527 goto sync_memory_failed;
530 /* Create an obj handle so kfd_process_device_remove_obj_handle
531 * will take care of the bo removal when the process finishes.
532 * We do not need to take p->mutex, because the process is just
533 * created and the ioctls have not had the chance to run.
535 handle = kfd_process_device_create_obj_handle(pdd, mem);
543 err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(kdev->kgd,
544 (struct kgd_mem *)mem, kptr, NULL);
546 pr_debug("Map GTT BO to kernel failed\n");
547 goto free_obj_handle;
554 kfd_process_device_remove_obj_handle(pdd, handle);
557 kfd_process_free_gpuvm(mem, pdd);
561 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(kdev->kgd, mem, NULL);
567 /* kfd_process_device_reserve_ib_mem - Reserve memory inside the
568 * process for IB usage The memory reserved is for KFD to submit
569 * IB to AMDGPU from kernel. If the memory is reserved
570 * successfully, ib_kaddr will have the CPU/kernel
571 * address. Check ib_kaddr before accessing the memory.
573 static int kfd_process_device_reserve_ib_mem(struct kfd_process_device *pdd)
575 struct qcm_process_device *qpd = &pdd->qpd;
576 uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT |
577 KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE |
578 KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE |
579 KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
583 if (qpd->ib_kaddr || !qpd->ib_base)
586 /* ib_base is only set for dGPU */
587 ret = kfd_process_alloc_gpuvm(pdd, qpd->ib_base, PAGE_SIZE, flags,
592 qpd->ib_kaddr = kaddr;
597 struct kfd_process *kfd_create_process(struct file *filep)
599 struct kfd_process *process;
600 struct task_struct *thread = current;
604 return ERR_PTR(-EINVAL);
606 /* Only the pthreads threading model is supported. */
607 if (thread->group_leader->mm != thread->mm)
608 return ERR_PTR(-EINVAL);
611 * take kfd processes mutex before starting of process creation
612 * so there won't be a case where two threads of the same process
613 * create two kfd_process structures
615 mutex_lock(&kfd_processes_mutex);
617 /* A prior open of /dev/kfd could have already created the process. */
618 process = find_process(thread);
620 pr_debug("Process already found\n");
622 process = create_process(thread);
626 ret = kfd_process_init_cwsr_apu(process, filep);
628 process = ERR_PTR(ret);
635 process->kobj = kfd_alloc_struct(process->kobj);
636 if (!process->kobj) {
637 pr_warn("Creating procfs kobject failed");
640 ret = kobject_init_and_add(process->kobj, &procfs_type,
642 (int)process->lead_thread->pid);
644 pr_warn("Creating procfs pid directory failed");
645 kobject_put(process->kobj);
649 process->attr_pasid.name = "pasid";
650 process->attr_pasid.mode = KFD_SYSFS_FILE_MODE;
651 sysfs_attr_init(&process->attr_pasid);
652 ret = sysfs_create_file(process->kobj, &process->attr_pasid);
654 pr_warn("Creating pasid for pid %d failed",
655 (int)process->lead_thread->pid);
657 process->kobj_queues = kobject_create_and_add("queues",
659 if (!process->kobj_queues)
660 pr_warn("Creating KFD proc/queues folder failed");
662 ret = kfd_procfs_add_sysfs_files(process);
664 pr_warn("Creating sysfs usage file for pid %d failed",
665 (int)process->lead_thread->pid);
668 if (!IS_ERR(process))
669 kref_get(&process->ref);
670 mutex_unlock(&kfd_processes_mutex);
675 struct kfd_process *kfd_get_process(const struct task_struct *thread)
677 struct kfd_process *process;
680 return ERR_PTR(-EINVAL);
682 /* Only the pthreads threading model is supported. */
683 if (thread->group_leader->mm != thread->mm)
684 return ERR_PTR(-EINVAL);
686 process = find_process(thread);
688 return ERR_PTR(-EINVAL);
693 static struct kfd_process *find_process_by_mm(const struct mm_struct *mm)
695 struct kfd_process *process;
697 hash_for_each_possible_rcu(kfd_processes_table, process,
698 kfd_processes, (uintptr_t)mm)
699 if (process->mm == mm)
705 static struct kfd_process *find_process(const struct task_struct *thread)
707 struct kfd_process *p;
710 idx = srcu_read_lock(&kfd_processes_srcu);
711 p = find_process_by_mm(thread->mm);
712 srcu_read_unlock(&kfd_processes_srcu, idx);
717 void kfd_unref_process(struct kfd_process *p)
719 kref_put(&p->ref, kfd_process_ref_release);
722 static void kfd_process_device_free_bos(struct kfd_process_device *pdd)
724 struct kfd_process *p = pdd->process;
729 * Remove all handles from idr and release appropriate
730 * local memory object
732 idr_for_each_entry(&pdd->alloc_idr, mem, id) {
733 struct kfd_process_device *peer_pdd;
735 list_for_each_entry(peer_pdd, &p->per_device_data,
739 amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
740 peer_pdd->dev->kgd, mem, peer_pdd->vm);
743 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->kgd, mem, NULL);
744 kfd_process_device_remove_obj_handle(pdd, id);
748 static void kfd_process_free_outstanding_kfd_bos(struct kfd_process *p)
750 struct kfd_process_device *pdd;
752 list_for_each_entry(pdd, &p->per_device_data, per_device_list)
753 kfd_process_device_free_bos(pdd);
756 static void kfd_process_destroy_pdds(struct kfd_process *p)
758 struct kfd_process_device *pdd, *temp;
760 list_for_each_entry_safe(pdd, temp, &p->per_device_data,
762 pr_debug("Releasing pdd (topology id %d) for process (pasid 0x%x)\n",
763 pdd->dev->id, p->pasid);
766 amdgpu_amdkfd_gpuvm_release_process_vm(
767 pdd->dev->kgd, pdd->vm);
771 amdgpu_amdkfd_gpuvm_destroy_process_vm(
772 pdd->dev->kgd, pdd->vm);
774 list_del(&pdd->per_device_list);
776 if (pdd->qpd.cwsr_kaddr && !pdd->qpd.cwsr_base)
777 free_pages((unsigned long)pdd->qpd.cwsr_kaddr,
778 get_order(KFD_CWSR_TBA_TMA_SIZE));
780 kfree(pdd->qpd.doorbell_bitmap);
781 idr_destroy(&pdd->alloc_idr);
784 * before destroying pdd, make sure to report availability
787 if (pdd->runtime_inuse) {
788 pm_runtime_mark_last_busy(pdd->dev->ddev->dev);
789 pm_runtime_put_autosuspend(pdd->dev->ddev->dev);
790 pdd->runtime_inuse = false;
797 /* No process locking is needed in this function, because the process
798 * is not findable any more. We must assume that no other thread is
799 * using it any more, otherwise we couldn't safely free the process
800 * structure in the end.
802 static void kfd_process_wq_release(struct work_struct *work)
804 struct kfd_process *p = container_of(work, struct kfd_process,
806 struct kfd_process_device *pdd;
808 /* Remove the procfs files */
810 sysfs_remove_file(p->kobj, &p->attr_pasid);
811 kobject_del(p->kobj_queues);
812 kobject_put(p->kobj_queues);
813 p->kobj_queues = NULL;
815 list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
816 sysfs_remove_file(p->kobj, &pdd->attr_vram);
817 sysfs_remove_file(p->kobj, &pdd->attr_sdma);
820 kobject_del(p->kobj);
821 kobject_put(p->kobj);
825 kfd_iommu_unbind_process(p);
827 kfd_process_free_outstanding_kfd_bos(p);
829 kfd_process_destroy_pdds(p);
830 dma_fence_put(p->ef);
832 kfd_event_free_process(p);
834 kfd_pasid_free(p->pasid);
835 kfd_free_process_doorbells(p);
837 mutex_destroy(&p->mutex);
839 put_task_struct(p->lead_thread);
844 static void kfd_process_ref_release(struct kref *ref)
846 struct kfd_process *p = container_of(ref, struct kfd_process, ref);
848 INIT_WORK(&p->release_work, kfd_process_wq_release);
849 queue_work(kfd_process_wq, &p->release_work);
852 static void kfd_process_free_notifier(struct mmu_notifier *mn)
854 kfd_unref_process(container_of(mn, struct kfd_process, mmu_notifier));
857 static void kfd_process_notifier_release(struct mmu_notifier *mn,
858 struct mm_struct *mm)
860 struct kfd_process *p;
861 struct kfd_process_device *pdd = NULL;
864 * The kfd_process structure can not be free because the
865 * mmu_notifier srcu is read locked
867 p = container_of(mn, struct kfd_process, mmu_notifier);
868 if (WARN_ON(p->mm != mm))
871 mutex_lock(&kfd_processes_mutex);
872 hash_del_rcu(&p->kfd_processes);
873 mutex_unlock(&kfd_processes_mutex);
874 synchronize_srcu(&kfd_processes_srcu);
876 cancel_delayed_work_sync(&p->eviction_work);
877 cancel_delayed_work_sync(&p->restore_work);
879 mutex_lock(&p->mutex);
881 /* Iterate over all process device data structures and if the
882 * pdd is in debug mode, we should first force unregistration,
883 * then we will be able to destroy the queues
885 list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
886 struct kfd_dev *dev = pdd->dev;
888 mutex_lock(kfd_get_dbgmgr_mutex());
889 if (dev && dev->dbgmgr && dev->dbgmgr->pasid == p->pasid) {
890 if (!kfd_dbgmgr_unregister(dev->dbgmgr, p)) {
891 kfd_dbgmgr_destroy(dev->dbgmgr);
895 mutex_unlock(kfd_get_dbgmgr_mutex());
898 kfd_process_dequeue_from_all_devices(p);
901 /* Indicate to other users that MM is no longer valid */
903 /* Signal the eviction fence after user mode queues are
904 * destroyed. This allows any BOs to be freed without
905 * triggering pointless evictions or waiting for fences.
907 dma_fence_signal(p->ef);
909 mutex_unlock(&p->mutex);
911 mmu_notifier_put(&p->mmu_notifier);
914 static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = {
915 .release = kfd_process_notifier_release,
916 .free_notifier = kfd_process_free_notifier,
919 static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep)
921 unsigned long offset;
922 struct kfd_process_device *pdd;
924 list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
925 struct kfd_dev *dev = pdd->dev;
926 struct qcm_process_device *qpd = &pdd->qpd;
928 if (!dev->cwsr_enabled || qpd->cwsr_kaddr || qpd->cwsr_base)
931 offset = KFD_MMAP_TYPE_RESERVED_MEM | KFD_MMAP_GPU_ID(dev->id);
932 qpd->tba_addr = (int64_t)vm_mmap(filep, 0,
933 KFD_CWSR_TBA_TMA_SIZE, PROT_READ | PROT_EXEC,
936 if (IS_ERR_VALUE(qpd->tba_addr)) {
937 int err = qpd->tba_addr;
939 pr_err("Failure to set tba address. error %d.\n", err);
941 qpd->cwsr_kaddr = NULL;
945 memcpy(qpd->cwsr_kaddr, dev->cwsr_isa, dev->cwsr_isa_size);
947 qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
948 pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
949 qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
955 static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd)
957 struct kfd_dev *dev = pdd->dev;
958 struct qcm_process_device *qpd = &pdd->qpd;
959 uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT
960 | KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE
961 | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
965 if (!dev->cwsr_enabled || qpd->cwsr_kaddr || !qpd->cwsr_base)
968 /* cwsr_base is only set for dGPU */
969 ret = kfd_process_alloc_gpuvm(pdd, qpd->cwsr_base,
970 KFD_CWSR_TBA_TMA_SIZE, flags, &kaddr);
974 qpd->cwsr_kaddr = kaddr;
975 qpd->tba_addr = qpd->cwsr_base;
977 memcpy(qpd->cwsr_kaddr, dev->cwsr_isa, dev->cwsr_isa_size);
979 qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
980 pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
981 qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
987 * On return the kfd_process is fully operational and will be freed when the
990 static struct kfd_process *create_process(const struct task_struct *thread)
992 struct kfd_process *process;
995 process = kzalloc(sizeof(*process), GFP_KERNEL);
997 goto err_alloc_process;
999 kref_init(&process->ref);
1000 mutex_init(&process->mutex);
1001 process->mm = thread->mm;
1002 process->lead_thread = thread->group_leader;
1003 INIT_LIST_HEAD(&process->per_device_data);
1004 INIT_DELAYED_WORK(&process->eviction_work, evict_process_worker);
1005 INIT_DELAYED_WORK(&process->restore_work, restore_process_worker);
1006 process->last_restore_timestamp = get_jiffies_64();
1007 kfd_event_init_process(process);
1008 process->is_32bit_user_mode = in_compat_syscall();
1010 process->pasid = kfd_pasid_alloc();
1011 if (process->pasid == 0)
1012 goto err_alloc_pasid;
1014 if (kfd_alloc_process_doorbells(process) < 0)
1015 goto err_alloc_doorbells;
1017 err = pqm_init(&process->pqm, process);
1019 goto err_process_pqm_init;
1021 /* init process apertures*/
1022 err = kfd_init_apertures(process);
1024 goto err_init_apertures;
1026 /* Must be last, have to use release destruction after this */
1027 process->mmu_notifier.ops = &kfd_process_mmu_notifier_ops;
1028 err = mmu_notifier_register(&process->mmu_notifier, process->mm);
1030 goto err_register_notifier;
1032 get_task_struct(process->lead_thread);
1033 hash_add_rcu(kfd_processes_table, &process->kfd_processes,
1034 (uintptr_t)process->mm);
1038 err_register_notifier:
1039 kfd_process_free_outstanding_kfd_bos(process);
1040 kfd_process_destroy_pdds(process);
1042 pqm_uninit(&process->pqm);
1043 err_process_pqm_init:
1044 kfd_free_process_doorbells(process);
1045 err_alloc_doorbells:
1046 kfd_pasid_free(process->pasid);
1048 mutex_destroy(&process->mutex);
1051 return ERR_PTR(err);
1054 static int init_doorbell_bitmap(struct qcm_process_device *qpd,
1055 struct kfd_dev *dev)
1058 int range_start = dev->shared_resources.non_cp_doorbells_start;
1059 int range_end = dev->shared_resources.non_cp_doorbells_end;
1061 if (!KFD_IS_SOC15(dev->device_info->asic_family))
1064 qpd->doorbell_bitmap =
1065 kzalloc(DIV_ROUND_UP(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
1066 BITS_PER_BYTE), GFP_KERNEL);
1067 if (!qpd->doorbell_bitmap)
1070 /* Mask out doorbells reserved for SDMA, IH, and VCN on SOC15. */
1071 pr_debug("reserved doorbell 0x%03x - 0x%03x\n", range_start, range_end);
1072 pr_debug("reserved doorbell 0x%03x - 0x%03x\n",
1073 range_start + KFD_QUEUE_DOORBELL_MIRROR_OFFSET,
1074 range_end + KFD_QUEUE_DOORBELL_MIRROR_OFFSET);
1076 for (i = 0; i < KFD_MAX_NUM_OF_QUEUES_PER_PROCESS / 2; i++) {
1077 if (i >= range_start && i <= range_end) {
1078 set_bit(i, qpd->doorbell_bitmap);
1079 set_bit(i + KFD_QUEUE_DOORBELL_MIRROR_OFFSET,
1080 qpd->doorbell_bitmap);
1087 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
1088 struct kfd_process *p)
1090 struct kfd_process_device *pdd = NULL;
1092 list_for_each_entry(pdd, &p->per_device_data, per_device_list)
1093 if (pdd->dev == dev)
1099 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
1100 struct kfd_process *p)
1102 struct kfd_process_device *pdd = NULL;
1104 pdd = kzalloc(sizeof(*pdd), GFP_KERNEL);
1108 if (init_doorbell_bitmap(&pdd->qpd, dev)) {
1109 pr_err("Failed to init doorbell for process\n");
1115 INIT_LIST_HEAD(&pdd->qpd.queues_list);
1116 INIT_LIST_HEAD(&pdd->qpd.priv_queue_list);
1117 pdd->qpd.dqm = dev->dqm;
1118 pdd->qpd.pqm = &p->pqm;
1119 pdd->qpd.evicted = 0;
1120 pdd->qpd.mapped_gws_queue = false;
1122 pdd->bound = PDD_UNBOUND;
1123 pdd->already_dequeued = false;
1124 pdd->runtime_inuse = false;
1125 pdd->vram_usage = 0;
1126 pdd->sdma_past_activity_counter = 0;
1127 list_add(&pdd->per_device_list, &p->per_device_data);
1129 /* Init idr used for memory handle translation */
1130 idr_init(&pdd->alloc_idr);
1136 * kfd_process_device_init_vm - Initialize a VM for a process-device
1138 * @pdd: The process-device
1139 * @drm_file: Optional pointer to a DRM file descriptor
1141 * If @drm_file is specified, it will be used to acquire the VM from
1142 * that file descriptor. If successful, the @pdd takes ownership of
1143 * the file descriptor.
1145 * If @drm_file is NULL, a new VM is created.
1147 * Returns 0 on success, -errno on failure.
1149 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
1150 struct file *drm_file)
1152 struct kfd_process *p;
1153 struct kfd_dev *dev;
1157 return drm_file ? -EBUSY : 0;
1163 ret = amdgpu_amdkfd_gpuvm_acquire_process_vm(
1164 dev->kgd, drm_file, p->pasid,
1165 &pdd->vm, &p->kgd_process_info, &p->ef);
1167 ret = amdgpu_amdkfd_gpuvm_create_process_vm(dev->kgd, p->pasid,
1168 &pdd->vm, &p->kgd_process_info, &p->ef);
1170 pr_err("Failed to create process VM object\n");
1174 amdgpu_vm_set_task_info(pdd->vm);
1176 ret = kfd_process_device_reserve_ib_mem(pdd);
1178 goto err_reserve_ib_mem;
1179 ret = kfd_process_device_init_cwsr_dgpu(pdd);
1183 pdd->drm_file = drm_file;
1189 kfd_process_device_free_bos(pdd);
1191 amdgpu_amdkfd_gpuvm_destroy_process_vm(dev->kgd, pdd->vm);
1198 * Direct the IOMMU to bind the process (specifically the pasid->mm)
1200 * Unbinding occurs when the process dies or the device is removed.
1202 * Assumes that the process lock is held.
1204 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
1205 struct kfd_process *p)
1207 struct kfd_process_device *pdd;
1210 pdd = kfd_get_process_device_data(dev, p);
1212 pr_err("Process device data doesn't exist\n");
1213 return ERR_PTR(-ENOMEM);
1217 * signal runtime-pm system to auto resume and prevent
1218 * further runtime suspend once device pdd is created until
1221 if (!pdd->runtime_inuse) {
1222 err = pm_runtime_get_sync(dev->ddev->dev);
1224 pm_runtime_put_autosuspend(dev->ddev->dev);
1225 return ERR_PTR(err);
1229 err = kfd_iommu_bind_process_to_device(pdd);
1233 err = kfd_process_device_init_vm(pdd, NULL);
1238 * make sure that runtime_usage counter is incremented just once
1241 pdd->runtime_inuse = true;
1246 /* balance runpm reference count and exit with error */
1247 if (!pdd->runtime_inuse) {
1248 pm_runtime_mark_last_busy(dev->ddev->dev);
1249 pm_runtime_put_autosuspend(dev->ddev->dev);
1252 return ERR_PTR(err);
1255 struct kfd_process_device *kfd_get_first_process_device_data(
1256 struct kfd_process *p)
1258 return list_first_entry(&p->per_device_data,
1259 struct kfd_process_device,
1263 struct kfd_process_device *kfd_get_next_process_device_data(
1264 struct kfd_process *p,
1265 struct kfd_process_device *pdd)
1267 if (list_is_last(&pdd->per_device_list, &p->per_device_data))
1269 return list_next_entry(pdd, per_device_list);
1272 bool kfd_has_process_device_data(struct kfd_process *p)
1274 return !(list_empty(&p->per_device_data));
1277 /* Create specific handle mapped to mem from process local memory idr
1278 * Assumes that the process lock is held.
1280 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
1283 return idr_alloc(&pdd->alloc_idr, mem, 0, 0, GFP_KERNEL);
1286 /* Translate specific handle from process local memory idr
1287 * Assumes that the process lock is held.
1289 void *kfd_process_device_translate_handle(struct kfd_process_device *pdd,
1295 return idr_find(&pdd->alloc_idr, handle);
1298 /* Remove specific handle from process local memory idr
1299 * Assumes that the process lock is held.
1301 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
1305 idr_remove(&pdd->alloc_idr, handle);
1308 /* This increments the process->ref counter. */
1309 struct kfd_process *kfd_lookup_process_by_pasid(unsigned int pasid)
1311 struct kfd_process *p, *ret_p = NULL;
1314 int idx = srcu_read_lock(&kfd_processes_srcu);
1316 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1317 if (p->pasid == pasid) {
1324 srcu_read_unlock(&kfd_processes_srcu, idx);
1329 /* This increments the process->ref counter. */
1330 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm)
1332 struct kfd_process *p;
1334 int idx = srcu_read_lock(&kfd_processes_srcu);
1336 p = find_process_by_mm(mm);
1340 srcu_read_unlock(&kfd_processes_srcu, idx);
1345 /* kfd_process_evict_queues - Evict all user queues of a process
1347 * Eviction is reference-counted per process-device. This means multiple
1348 * evictions from different sources can be nested safely.
1350 int kfd_process_evict_queues(struct kfd_process *p)
1352 struct kfd_process_device *pdd;
1354 unsigned int n_evicted = 0;
1356 list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
1357 r = pdd->dev->dqm->ops.evict_process_queues(pdd->dev->dqm,
1360 pr_err("Failed to evict process queues\n");
1369 /* To keep state consistent, roll back partial eviction by
1372 list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
1375 if (pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1377 pr_err("Failed to restore queues\n");
1385 /* kfd_process_restore_queues - Restore all user queues of a process */
1386 int kfd_process_restore_queues(struct kfd_process *p)
1388 struct kfd_process_device *pdd;
1391 list_for_each_entry(pdd, &p->per_device_data, per_device_list) {
1392 r = pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1395 pr_err("Failed to restore process queues\n");
1404 static void evict_process_worker(struct work_struct *work)
1407 struct kfd_process *p;
1408 struct delayed_work *dwork;
1410 dwork = to_delayed_work(work);
1412 /* Process termination destroys this worker thread. So during the
1413 * lifetime of this thread, kfd_process p will be valid
1415 p = container_of(dwork, struct kfd_process, eviction_work);
1416 WARN_ONCE(p->last_eviction_seqno != p->ef->seqno,
1417 "Eviction fence mismatch\n");
1419 /* Narrow window of overlap between restore and evict work
1420 * item is possible. Once amdgpu_amdkfd_gpuvm_restore_process_bos
1421 * unreserves KFD BOs, it is possible to evicted again. But
1422 * restore has few more steps of finish. So lets wait for any
1423 * previous restore work to complete
1425 flush_delayed_work(&p->restore_work);
1427 pr_debug("Started evicting pasid 0x%x\n", p->pasid);
1428 ret = kfd_process_evict_queues(p);
1430 dma_fence_signal(p->ef);
1431 dma_fence_put(p->ef);
1433 queue_delayed_work(kfd_restore_wq, &p->restore_work,
1434 msecs_to_jiffies(PROCESS_RESTORE_TIME_MS));
1436 pr_debug("Finished evicting pasid 0x%x\n", p->pasid);
1438 pr_err("Failed to evict queues of pasid 0x%x\n", p->pasid);
1441 static void restore_process_worker(struct work_struct *work)
1443 struct delayed_work *dwork;
1444 struct kfd_process *p;
1447 dwork = to_delayed_work(work);
1449 /* Process termination destroys this worker thread. So during the
1450 * lifetime of this thread, kfd_process p will be valid
1452 p = container_of(dwork, struct kfd_process, restore_work);
1453 pr_debug("Started restoring pasid 0x%x\n", p->pasid);
1455 /* Setting last_restore_timestamp before successful restoration.
1456 * Otherwise this would have to be set by KGD (restore_process_bos)
1457 * before KFD BOs are unreserved. If not, the process can be evicted
1458 * again before the timestamp is set.
1459 * If restore fails, the timestamp will be set again in the next
1460 * attempt. This would mean that the minimum GPU quanta would be
1461 * PROCESS_ACTIVE_TIME_MS - (time to execute the following two
1465 p->last_restore_timestamp = get_jiffies_64();
1466 ret = amdgpu_amdkfd_gpuvm_restore_process_bos(p->kgd_process_info,
1469 pr_debug("Failed to restore BOs of pasid 0x%x, retry after %d ms\n",
1470 p->pasid, PROCESS_BACK_OFF_TIME_MS);
1471 ret = queue_delayed_work(kfd_restore_wq, &p->restore_work,
1472 msecs_to_jiffies(PROCESS_BACK_OFF_TIME_MS));
1473 WARN(!ret, "reschedule restore work failed\n");
1477 ret = kfd_process_restore_queues(p);
1479 pr_debug("Finished restoring pasid 0x%x\n", p->pasid);
1481 pr_err("Failed to restore queues of pasid 0x%x\n", p->pasid);
1484 void kfd_suspend_all_processes(void)
1486 struct kfd_process *p;
1488 int idx = srcu_read_lock(&kfd_processes_srcu);
1490 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1491 cancel_delayed_work_sync(&p->eviction_work);
1492 cancel_delayed_work_sync(&p->restore_work);
1494 if (kfd_process_evict_queues(p))
1495 pr_err("Failed to suspend process 0x%x\n", p->pasid);
1496 dma_fence_signal(p->ef);
1497 dma_fence_put(p->ef);
1500 srcu_read_unlock(&kfd_processes_srcu, idx);
1503 int kfd_resume_all_processes(void)
1505 struct kfd_process *p;
1507 int ret = 0, idx = srcu_read_lock(&kfd_processes_srcu);
1509 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1510 if (!queue_delayed_work(kfd_restore_wq, &p->restore_work, 0)) {
1511 pr_err("Restore process %d failed during resume\n",
1516 srcu_read_unlock(&kfd_processes_srcu, idx);
1520 int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process,
1521 struct vm_area_struct *vma)
1523 struct kfd_process_device *pdd;
1524 struct qcm_process_device *qpd;
1526 if ((vma->vm_end - vma->vm_start) != KFD_CWSR_TBA_TMA_SIZE) {
1527 pr_err("Incorrect CWSR mapping size.\n");
1531 pdd = kfd_get_process_device_data(dev, process);
1536 qpd->cwsr_kaddr = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
1537 get_order(KFD_CWSR_TBA_TMA_SIZE));
1538 if (!qpd->cwsr_kaddr) {
1539 pr_err("Error allocating per process CWSR buffer.\n");
1543 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND
1544 | VM_NORESERVE | VM_DONTDUMP | VM_PFNMAP;
1545 /* Mapping pages to user process */
1546 return remap_pfn_range(vma, vma->vm_start,
1547 PFN_DOWN(__pa(qpd->cwsr_kaddr)),
1548 KFD_CWSR_TBA_TMA_SIZE, vma->vm_page_prot);
1551 void kfd_flush_tlb(struct kfd_process_device *pdd)
1553 struct kfd_dev *dev = pdd->dev;
1555 if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1556 /* Nothing to flush until a VMID is assigned, which
1557 * only happens when the first queue is created.
1560 amdgpu_amdkfd_flush_gpu_tlb_vmid(dev->kgd,
1563 amdgpu_amdkfd_flush_gpu_tlb_pasid(dev->kgd,
1564 pdd->process->pasid);
1568 #if defined(CONFIG_DEBUG_FS)
1570 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data)
1572 struct kfd_process *p;
1576 int idx = srcu_read_lock(&kfd_processes_srcu);
1578 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1579 seq_printf(m, "Process %d PASID 0x%x:\n",
1580 p->lead_thread->tgid, p->pasid);
1582 mutex_lock(&p->mutex);
1583 r = pqm_debugfs_mqds(m, &p->pqm);
1584 mutex_unlock(&p->mutex);
1590 srcu_read_unlock(&kfd_processes_srcu, idx);