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.
25 #include <linux/slab.h>
26 #include <linux/list.h>
27 #include "kfd_device_queue_manager.h"
29 #include "kfd_kernel_queue.h"
30 #include "amdgpu_amdkfd.h"
32 static inline struct process_queue_node *get_queue_by_qid(
33 struct process_queue_manager *pqm, unsigned int qid)
35 struct process_queue_node *pqn;
37 list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
38 if ((pqn->q && pqn->q->properties.queue_id == qid) ||
39 (pqn->kq && pqn->kq->queue->properties.queue_id == qid))
46 static int assign_queue_slot_by_qid(struct process_queue_manager *pqm,
49 if (qid >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS)
52 if (__test_and_set_bit(qid, pqm->queue_slot_bitmap)) {
53 pr_err("Cannot create new queue because requested qid(%u) is in use\n", qid);
60 static int find_available_queue_slot(struct process_queue_manager *pqm,
65 found = find_first_zero_bit(pqm->queue_slot_bitmap,
66 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
68 pr_debug("The new slot id %lu\n", found);
70 if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) {
71 pr_info("Cannot open more queues for process with pasid 0x%x\n",
76 set_bit(found, pqm->queue_slot_bitmap);
82 void kfd_process_dequeue_from_device(struct kfd_process_device *pdd)
84 struct kfd_node *dev = pdd->dev;
86 if (pdd->already_dequeued)
89 dev->dqm->ops.process_termination(dev->dqm, &pdd->qpd);
90 pdd->already_dequeued = true;
93 int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
96 struct kfd_node *dev = NULL;
97 struct process_queue_node *pqn;
98 struct kfd_process_device *pdd;
99 struct kgd_mem *mem = NULL;
102 pqn = get_queue_by_qid(pqm, qid);
104 pr_err("Queue id does not match any known queue\n");
109 dev = pqn->q->device;
113 pdd = kfd_get_process_device_data(dev, pqm->process);
115 pr_err("Process device data doesn't exist\n");
119 /* Only allow one queue per process can have GWS assigned */
120 if (gws && pdd->qpd.num_gws)
123 if (!gws && pdd->qpd.num_gws == 0)
126 if (KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 3)) {
128 ret = amdgpu_amdkfd_add_gws_to_process(pdd->process->kgd_process_info,
131 ret = amdgpu_amdkfd_remove_gws_from_process(pdd->process->kgd_process_info,
138 * Intentionally set GWS to a non-NULL value
141 pqn->q->gws = gws ? ERR_PTR(-ENOMEM) : NULL;
144 pdd->qpd.num_gws = gws ? dev->adev->gds.gws_size : 0;
146 return pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
150 void kfd_process_dequeue_from_all_devices(struct kfd_process *p)
154 for (i = 0; i < p->n_pdds; i++)
155 kfd_process_dequeue_from_device(p->pdds[i]);
158 int pqm_init(struct process_queue_manager *pqm, struct kfd_process *p)
160 INIT_LIST_HEAD(&pqm->queues);
161 pqm->queue_slot_bitmap = bitmap_zalloc(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
163 if (!pqm->queue_slot_bitmap)
170 void pqm_uninit(struct process_queue_manager *pqm)
172 struct process_queue_node *pqn, *next;
174 list_for_each_entry_safe(pqn, next, &pqm->queues, process_queue_list) {
175 if (pqn->q && pqn->q->gws &&
176 KFD_GC_VERSION(pqn->q->device) != IP_VERSION(9, 4, 3))
177 amdgpu_amdkfd_remove_gws_from_process(pqm->process->kgd_process_info,
179 kfd_procfs_del_queue(pqn->q);
180 uninit_queue(pqn->q);
181 list_del(&pqn->process_queue_list);
185 bitmap_free(pqm->queue_slot_bitmap);
186 pqm->queue_slot_bitmap = NULL;
189 static int init_user_queue(struct process_queue_manager *pqm,
190 struct kfd_node *dev, struct queue **q,
191 struct queue_properties *q_properties,
192 struct file *f, struct amdgpu_bo *wptr_bo,
197 /* Doorbell initialized in user space*/
198 q_properties->doorbell_ptr = NULL;
199 q_properties->exception_status = KFD_EC_MASK(EC_QUEUE_NEW);
201 /* let DQM handle it*/
202 q_properties->vmid = 0;
203 q_properties->queue_id = qid;
205 retval = init_queue(q, q_properties);
210 (*q)->process = pqm->process;
212 if (dev->kfd->shared_resources.enable_mes) {
213 retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev,
214 AMDGPU_MES_GANG_CTX_SIZE,
216 &(*q)->gang_ctx_gpu_addr,
217 &(*q)->gang_ctx_cpu_ptr,
220 pr_err("failed to allocate gang context bo\n");
223 memset((*q)->gang_ctx_cpu_ptr, 0, AMDGPU_MES_GANG_CTX_SIZE);
224 (*q)->wptr_bo = wptr_bo;
227 pr_debug("PQM After init queue");
236 int pqm_create_queue(struct process_queue_manager *pqm,
237 struct kfd_node *dev,
239 struct queue_properties *properties,
241 struct amdgpu_bo *wptr_bo,
242 const struct kfd_criu_queue_priv_data *q_data,
243 const void *restore_mqd,
244 const void *restore_ctl_stack,
245 uint32_t *p_doorbell_offset_in_process)
248 struct kfd_process_device *pdd;
250 struct process_queue_node *pqn;
251 struct kernel_queue *kq;
252 enum kfd_queue_type type = properties->type;
253 unsigned int max_queues = 127; /* HWS limit */
256 * On GFX 9.4.3, increase the number of queues that
257 * can be created to 255. No HWS limit on GFX 9.4.3.
259 if (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3))
265 pdd = kfd_get_process_device_data(dev, pqm->process);
267 pr_err("Process device data doesn't exist\n");
272 * for debug process, verify that it is within the static queues limit
273 * currently limit is set to half of the total avail HQD slots
274 * If we are just about to create DIQ, the is_debug flag is not set yet
275 * Hence we also check the type as well
277 if ((pdd->qpd.is_debug) || (type == KFD_QUEUE_TYPE_DIQ))
278 max_queues = dev->kfd->device_info.max_no_of_hqd/2;
280 if (pdd->qpd.queue_count >= max_queues)
284 retval = assign_queue_slot_by_qid(pqm, q_data->q_id);
287 retval = find_available_queue_slot(pqm, qid);
292 if (list_empty(&pdd->qpd.queues_list) &&
293 list_empty(&pdd->qpd.priv_queue_list))
294 dev->dqm->ops.register_process(dev->dqm, &pdd->qpd);
296 pqn = kzalloc(sizeof(*pqn), GFP_KERNEL);
299 goto err_allocate_pqn;
303 case KFD_QUEUE_TYPE_SDMA:
304 case KFD_QUEUE_TYPE_SDMA_XGMI:
305 /* SDMA queues are always allocated statically no matter
306 * which scheduler mode is used. We also do not need to
307 * check whether a SDMA queue can be allocated here, because
308 * allocate_sdma_queue() in create_queue() has the
309 * corresponding check logic.
311 retval = init_user_queue(pqm, dev, &q, properties, f, wptr_bo, *qid);
313 goto err_create_queue;
316 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd, q_data,
317 restore_mqd, restore_ctl_stack);
321 case KFD_QUEUE_TYPE_COMPUTE:
322 /* check if there is over subscription */
323 if ((dev->dqm->sched_policy ==
324 KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION) &&
325 ((dev->dqm->processes_count >= dev->vm_info.vmid_num_kfd) ||
326 (dev->dqm->active_queue_count >= get_cp_queues_num(dev->dqm)))) {
327 pr_debug("Over-subscription is not allowed when amdkfd.sched_policy == 1\n");
329 goto err_create_queue;
332 retval = init_user_queue(pqm, dev, &q, properties, f, wptr_bo, *qid);
334 goto err_create_queue;
337 retval = dev->dqm->ops.create_queue(dev->dqm, q, &pdd->qpd, q_data,
338 restore_mqd, restore_ctl_stack);
341 case KFD_QUEUE_TYPE_DIQ:
342 kq = kernel_queue_init(dev, KFD_QUEUE_TYPE_DIQ);
345 goto err_create_queue;
347 kq->queue->properties.queue_id = *qid;
350 retval = kfd_process_drain_interrupts(pdd);
354 retval = dev->dqm->ops.create_kernel_queue(dev->dqm,
358 WARN(1, "Invalid queue type %d", type);
363 pr_err("Pasid 0x%x DQM create queue type %d failed. ret %d\n",
364 pqm->process->pasid, type, retval);
365 goto err_create_queue;
368 if (q && p_doorbell_offset_in_process)
369 /* Return the doorbell offset within the doorbell page
370 * to the caller so it can be passed up to user mode
372 * There are always 1024 doorbells per process, so in case
373 * of 8-byte doorbells, there are two doorbell pages per
376 *p_doorbell_offset_in_process =
377 (q->properties.doorbell_off * sizeof(uint32_t)) &
378 (kfd_doorbell_process_slice(dev->kfd) - 1);
380 pr_debug("PQM After DQM create queue\n");
382 list_add(&pqn->process_queue_list, &pqm->queues);
385 pr_debug("PQM done creating queue\n");
386 kfd_procfs_add_queue(q);
387 print_queue_properties(&q->properties);
395 kernel_queue_uninit(kq, false);
398 /* check if queues list is empty unregister process from device */
399 clear_bit(*qid, pqm->queue_slot_bitmap);
400 if (list_empty(&pdd->qpd.queues_list) &&
401 list_empty(&pdd->qpd.priv_queue_list))
402 dev->dqm->ops.unregister_process(dev->dqm, &pdd->qpd);
406 int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
408 struct process_queue_node *pqn;
409 struct kfd_process_device *pdd;
410 struct device_queue_manager *dqm;
411 struct kfd_node *dev;
418 pqn = get_queue_by_qid(pqm, qid);
420 pr_err("Queue id does not match any known queue\n");
428 dev = pqn->q->device;
432 pdd = kfd_get_process_device_data(dev, pqm->process);
434 pr_err("Process device data doesn't exist\n");
439 /* destroy kernel queue (DIQ) */
440 dqm = pqn->kq->dev->dqm;
441 dqm->ops.destroy_kernel_queue(dqm, pqn->kq, &pdd->qpd);
442 kernel_queue_uninit(pqn->kq, false);
446 kfd_procfs_del_queue(pqn->q);
447 dqm = pqn->q->device->dqm;
448 retval = dqm->ops.destroy_queue(dqm, &pdd->qpd, pqn->q);
450 pr_err("Pasid 0x%x destroy queue %d failed, ret %d\n",
452 pqn->q->properties.queue_id, retval);
453 if (retval != -ETIME)
454 goto err_destroy_queue;
458 if (KFD_GC_VERSION(pqn->q->device) != IP_VERSION(9, 4, 3))
459 amdgpu_amdkfd_remove_gws_from_process(
460 pqm->process->kgd_process_info,
462 pdd->qpd.num_gws = 0;
465 if (dev->kfd->shared_resources.enable_mes) {
466 amdgpu_amdkfd_free_gtt_mem(dev->adev,
467 pqn->q->gang_ctx_bo);
469 amdgpu_amdkfd_free_gtt_mem(dev->adev, pqn->q->wptr_bo);
472 uninit_queue(pqn->q);
475 list_del(&pqn->process_queue_list);
477 clear_bit(qid, pqm->queue_slot_bitmap);
479 if (list_empty(&pdd->qpd.queues_list) &&
480 list_empty(&pdd->qpd.priv_queue_list))
481 dqm->ops.unregister_process(dqm, &pdd->qpd);
487 int pqm_update_queue_properties(struct process_queue_manager *pqm,
488 unsigned int qid, struct queue_properties *p)
491 struct process_queue_node *pqn;
493 pqn = get_queue_by_qid(pqm, qid);
495 pr_debug("No queue %d exists for update operation\n", qid);
499 pqn->q->properties.queue_address = p->queue_address;
500 pqn->q->properties.queue_size = p->queue_size;
501 pqn->q->properties.queue_percent = p->queue_percent;
502 pqn->q->properties.priority = p->priority;
503 pqn->q->properties.pm4_target_xcc = p->pm4_target_xcc;
505 retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
513 int pqm_update_mqd(struct process_queue_manager *pqm,
514 unsigned int qid, struct mqd_update_info *minfo)
517 struct process_queue_node *pqn;
519 pqn = get_queue_by_qid(pqm, qid);
521 pr_debug("No queue %d exists for update operation\n", qid);
525 /* CUs are masked for debugger requirements so deny user mask */
526 if (pqn->q->properties.is_dbg_wa && minfo && minfo->cu_mask.ptr)
529 /* ASICs that have WGPs must enforce pairwise enabled mask checks. */
530 if (minfo && minfo->cu_mask.ptr &&
531 KFD_GC_VERSION(pqn->q->device) >= IP_VERSION(10, 0, 0)) {
534 for (i = 0; i < minfo->cu_mask.count; i += 2) {
535 uint32_t cu_pair = (minfo->cu_mask.ptr[i / 32] >> (i % 32)) & 0x3;
537 if (cu_pair && cu_pair != 0x3) {
538 pr_debug("CUs must be adjacent pairwise enabled.\n");
544 retval = pqn->q->device->dqm->ops.update_queue(pqn->q->device->dqm,
549 if (minfo && minfo->cu_mask.ptr)
550 pqn->q->properties.is_user_cu_masked = true;
555 struct kernel_queue *pqm_get_kernel_queue(
556 struct process_queue_manager *pqm,
559 struct process_queue_node *pqn;
561 pqn = get_queue_by_qid(pqm, qid);
568 struct queue *pqm_get_user_queue(struct process_queue_manager *pqm,
571 struct process_queue_node *pqn;
573 pqn = get_queue_by_qid(pqm, qid);
574 return pqn ? pqn->q : NULL;
577 int pqm_get_wave_state(struct process_queue_manager *pqm,
579 void __user *ctl_stack,
580 u32 *ctl_stack_used_size,
581 u32 *save_area_used_size)
583 struct process_queue_node *pqn;
585 pqn = get_queue_by_qid(pqm, qid);
587 pr_debug("amdkfd: No queue %d exists for operation\n",
592 return pqn->q->device->dqm->ops.get_wave_state(pqn->q->device->dqm,
596 save_area_used_size);
599 int pqm_get_queue_snapshot(struct process_queue_manager *pqm,
600 uint64_t exception_clear_mask,
602 int *num_qss_entries,
603 uint32_t *entry_size)
605 struct process_queue_node *pqn;
606 struct kfd_queue_snapshot_entry src;
607 uint32_t tmp_entry_size = *entry_size, tmp_qss_entries = *num_qss_entries;
610 *num_qss_entries = 0;
614 *entry_size = min_t(size_t, *entry_size, sizeof(struct kfd_queue_snapshot_entry));
615 mutex_lock(&pqm->process->event_mutex);
617 memset(&src, 0, sizeof(src));
619 list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
623 if (*num_qss_entries < tmp_qss_entries) {
624 set_queue_snapshot_entry(pqn->q, exception_clear_mask, &src);
626 if (copy_to_user(buf, &src, *entry_size)) {
630 buf += tmp_entry_size;
632 *num_qss_entries += 1;
635 mutex_unlock(&pqm->process->event_mutex);
639 static int get_queue_data_sizes(struct kfd_process_device *pdd,
642 uint32_t *ctl_stack_size)
646 ret = pqm_get_queue_checkpoint_info(&pdd->process->pqm,
647 q->properties.queue_id,
651 pr_err("Failed to get queue dump info (%d)\n", ret);
656 int kfd_process_get_queue_info(struct kfd_process *p,
657 uint32_t *num_queues,
658 uint64_t *priv_data_sizes)
660 uint32_t extra_data_sizes = 0;
667 /* Run over all PDDs of the process */
668 for (i = 0; i < p->n_pdds; i++) {
669 struct kfd_process_device *pdd = p->pdds[i];
671 list_for_each_entry(q, &pdd->qpd.queues_list, list) {
672 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
673 q->properties.type == KFD_QUEUE_TYPE_SDMA ||
674 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
675 uint32_t mqd_size, ctl_stack_size;
677 *num_queues = *num_queues + 1;
679 ret = get_queue_data_sizes(pdd, q, &mqd_size, &ctl_stack_size);
683 extra_data_sizes += mqd_size + ctl_stack_size;
685 pr_err("Unsupported queue type (%d)\n", q->properties.type);
690 *priv_data_sizes = extra_data_sizes +
691 (*num_queues * sizeof(struct kfd_criu_queue_priv_data));
696 static int pqm_checkpoint_mqd(struct process_queue_manager *pqm,
701 struct process_queue_node *pqn;
703 pqn = get_queue_by_qid(pqm, qid);
705 pr_debug("amdkfd: No queue %d exists for operation\n", qid);
709 if (!pqn->q->device->dqm->ops.checkpoint_mqd) {
710 pr_err("amdkfd: queue dumping not supported on this device\n");
714 return pqn->q->device->dqm->ops.checkpoint_mqd(pqn->q->device->dqm,
715 pqn->q, mqd, ctl_stack);
718 static int criu_checkpoint_queue(struct kfd_process_device *pdd,
720 struct kfd_criu_queue_priv_data *q_data)
722 uint8_t *mqd, *ctl_stack;
725 mqd = (void *)(q_data + 1);
726 ctl_stack = mqd + q_data->mqd_size;
728 q_data->gpu_id = pdd->user_gpu_id;
729 q_data->type = q->properties.type;
730 q_data->format = q->properties.format;
731 q_data->q_id = q->properties.queue_id;
732 q_data->q_address = q->properties.queue_address;
733 q_data->q_size = q->properties.queue_size;
734 q_data->priority = q->properties.priority;
735 q_data->q_percent = q->properties.queue_percent;
736 q_data->read_ptr_addr = (uint64_t)q->properties.read_ptr;
737 q_data->write_ptr_addr = (uint64_t)q->properties.write_ptr;
738 q_data->doorbell_id = q->doorbell_id;
740 q_data->sdma_id = q->sdma_id;
742 q_data->eop_ring_buffer_address =
743 q->properties.eop_ring_buffer_address;
745 q_data->eop_ring_buffer_size = q->properties.eop_ring_buffer_size;
747 q_data->ctx_save_restore_area_address =
748 q->properties.ctx_save_restore_area_address;
750 q_data->ctx_save_restore_area_size =
751 q->properties.ctx_save_restore_area_size;
753 q_data->gws = !!q->gws;
755 ret = pqm_checkpoint_mqd(&pdd->process->pqm, q->properties.queue_id, mqd, ctl_stack);
757 pr_err("Failed checkpoint queue_mqd (%d)\n", ret);
761 pr_debug("Dumping Queue: gpu_id:%x queue_id:%u\n", q_data->gpu_id, q_data->q_id);
765 static int criu_checkpoint_queues_device(struct kfd_process_device *pdd,
766 uint8_t __user *user_priv,
767 unsigned int *q_index,
768 uint64_t *queues_priv_data_offset)
770 unsigned int q_private_data_size = 0;
771 uint8_t *q_private_data = NULL; /* Local buffer to store individual queue private data */
775 list_for_each_entry(q, &pdd->qpd.queues_list, list) {
776 struct kfd_criu_queue_priv_data *q_data;
777 uint64_t q_data_size;
779 uint32_t ctl_stack_size;
781 if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE &&
782 q->properties.type != KFD_QUEUE_TYPE_SDMA &&
783 q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI) {
785 pr_err("Unsupported queue type (%d)\n", q->properties.type);
790 ret = get_queue_data_sizes(pdd, q, &mqd_size, &ctl_stack_size);
794 q_data_size = sizeof(*q_data) + mqd_size + ctl_stack_size;
796 /* Increase local buffer space if needed */
797 if (q_private_data_size < q_data_size) {
798 kfree(q_private_data);
800 q_private_data = kzalloc(q_data_size, GFP_KERNEL);
801 if (!q_private_data) {
805 q_private_data_size = q_data_size;
808 q_data = (struct kfd_criu_queue_priv_data *)q_private_data;
810 /* data stored in this order: priv_data, mqd, ctl_stack */
811 q_data->mqd_size = mqd_size;
812 q_data->ctl_stack_size = ctl_stack_size;
814 ret = criu_checkpoint_queue(pdd, q, q_data);
818 q_data->object_type = KFD_CRIU_OBJECT_TYPE_QUEUE;
820 ret = copy_to_user(user_priv + *queues_priv_data_offset,
821 q_data, q_data_size);
826 *queues_priv_data_offset += q_data_size;
827 *q_index = *q_index + 1;
830 kfree(q_private_data);
835 int kfd_criu_checkpoint_queues(struct kfd_process *p,
836 uint8_t __user *user_priv_data,
837 uint64_t *priv_data_offset)
839 int ret = 0, pdd_index, q_index = 0;
841 for (pdd_index = 0; pdd_index < p->n_pdds; pdd_index++) {
842 struct kfd_process_device *pdd = p->pdds[pdd_index];
845 * criu_checkpoint_queues_device will copy data to user and update q_index and
846 * queues_priv_data_offset
848 ret = criu_checkpoint_queues_device(pdd, user_priv_data, &q_index,
858 static void set_queue_properties_from_criu(struct queue_properties *qp,
859 struct kfd_criu_queue_priv_data *q_data)
861 qp->is_interop = false;
862 qp->queue_percent = q_data->q_percent;
863 qp->priority = q_data->priority;
864 qp->queue_address = q_data->q_address;
865 qp->queue_size = q_data->q_size;
866 qp->read_ptr = (uint32_t *) q_data->read_ptr_addr;
867 qp->write_ptr = (uint32_t *) q_data->write_ptr_addr;
868 qp->eop_ring_buffer_address = q_data->eop_ring_buffer_address;
869 qp->eop_ring_buffer_size = q_data->eop_ring_buffer_size;
870 qp->ctx_save_restore_area_address = q_data->ctx_save_restore_area_address;
871 qp->ctx_save_restore_area_size = q_data->ctx_save_restore_area_size;
872 qp->ctl_stack_size = q_data->ctl_stack_size;
873 qp->type = q_data->type;
874 qp->format = q_data->format;
877 int kfd_criu_restore_queue(struct kfd_process *p,
878 uint8_t __user *user_priv_ptr,
879 uint64_t *priv_data_offset,
880 uint64_t max_priv_data_size)
882 uint8_t *mqd, *ctl_stack, *q_extra_data = NULL;
883 struct kfd_criu_queue_priv_data *q_data;
884 struct kfd_process_device *pdd;
885 uint64_t q_extra_data_size;
886 struct queue_properties qp;
887 unsigned int queue_id;
890 if (*priv_data_offset + sizeof(*q_data) > max_priv_data_size)
893 q_data = kmalloc(sizeof(*q_data), GFP_KERNEL);
897 ret = copy_from_user(q_data, user_priv_ptr + *priv_data_offset, sizeof(*q_data));
903 *priv_data_offset += sizeof(*q_data);
904 q_extra_data_size = (uint64_t)q_data->ctl_stack_size + q_data->mqd_size;
906 if (*priv_data_offset + q_extra_data_size > max_priv_data_size) {
911 q_extra_data = kmalloc(q_extra_data_size, GFP_KERNEL);
917 ret = copy_from_user(q_extra_data, user_priv_ptr + *priv_data_offset, q_extra_data_size);
923 *priv_data_offset += q_extra_data_size;
925 pdd = kfd_process_device_data_by_id(p, q_data->gpu_id);
927 pr_err("Failed to get pdd\n");
932 if (!pdd->doorbell_index &&
933 kfd_alloc_process_doorbells(pdd->dev->kfd, &pdd->doorbell_index) < 0) {
938 /* data stored in this order: mqd, ctl_stack */
940 ctl_stack = mqd + q_data->mqd_size;
942 memset(&qp, 0, sizeof(qp));
943 set_queue_properties_from_criu(&qp, q_data);
945 print_queue_properties(&qp);
947 ret = pqm_create_queue(&p->pqm, pdd->dev, NULL, &qp, &queue_id, NULL, q_data, mqd, ctl_stack,
950 pr_err("Failed to create new queue err:%d\n", ret);
955 ret = pqm_set_gws(&p->pqm, q_data->q_id, pdd->dev->gws);
959 pr_err("Failed to restore queue (%d)\n", ret);
961 pr_debug("Queue id %d was restored successfully\n", queue_id);
968 int pqm_get_queue_checkpoint_info(struct process_queue_manager *pqm,
971 uint32_t *ctl_stack_size)
973 struct process_queue_node *pqn;
975 pqn = get_queue_by_qid(pqm, qid);
977 pr_debug("amdkfd: No queue %d exists for operation\n", qid);
981 if (!pqn->q->device->dqm->ops.get_queue_checkpoint_info) {
982 pr_err("amdkfd: queue dumping not supported on this device\n");
986 pqn->q->device->dqm->ops.get_queue_checkpoint_info(pqn->q->device->dqm,
992 #if defined(CONFIG_DEBUG_FS)
994 int pqm_debugfs_mqds(struct seq_file *m, void *data)
996 struct process_queue_manager *pqm = data;
997 struct process_queue_node *pqn;
999 enum KFD_MQD_TYPE mqd_type;
1000 struct mqd_manager *mqd_mgr;
1001 int r = 0, xcc, num_xccs = 1;
1005 list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
1008 switch (q->properties.type) {
1009 case KFD_QUEUE_TYPE_SDMA:
1010 case KFD_QUEUE_TYPE_SDMA_XGMI:
1011 seq_printf(m, " SDMA queue on device %x\n",
1013 mqd_type = KFD_MQD_TYPE_SDMA;
1015 case KFD_QUEUE_TYPE_COMPUTE:
1016 seq_printf(m, " Compute queue on device %x\n",
1018 mqd_type = KFD_MQD_TYPE_CP;
1019 num_xccs = NUM_XCC(q->device->xcc_mask);
1023 " Bad user queue type %d on device %x\n",
1024 q->properties.type, q->device->id);
1027 mqd_mgr = q->device->dqm->mqd_mgrs[mqd_type];
1028 size = mqd_mgr->mqd_stride(mqd_mgr,
1030 } else if (pqn->kq) {
1032 mqd_mgr = pqn->kq->mqd_mgr;
1033 switch (q->properties.type) {
1034 case KFD_QUEUE_TYPE_DIQ:
1035 seq_printf(m, " DIQ on device %x\n",
1040 " Bad kernel queue type %d on device %x\n",
1047 " Weird: Queue node with neither kernel nor user queue\n");
1051 for (xcc = 0; xcc < num_xccs; xcc++) {
1052 mqd = q->mqd + size * xcc;
1053 r = mqd_mgr->debugfs_show_mqd(m, mqd);