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.
24 #include <linux/slab.h>
25 #include <linux/list.h>
26 #include <linux/types.h>
27 #include <linux/printk.h>
28 #include <linux/bitops.h>
29 #include <linux/sched.h>
31 #include "kfd_device_queue_manager.h"
32 #include "kfd_mqd_manager.h"
34 #include "kfd_kernel_queue.h"
36 /* Size of the per-pipe EOP queue */
37 #define CIK_HPD_EOP_BYTES_LOG2 11
38 #define CIK_HPD_EOP_BYTES (1U << CIK_HPD_EOP_BYTES_LOG2)
40 static int set_pasid_vmid_mapping(struct device_queue_manager *dqm,
41 unsigned int pasid, unsigned int vmid);
43 static int create_compute_queue_nocpsch(struct device_queue_manager *dqm,
45 struct qcm_process_device *qpd);
47 static int execute_queues_cpsch(struct device_queue_manager *dqm, bool lock);
48 static int destroy_queues_cpsch(struct device_queue_manager *dqm,
49 bool preempt_static_queues, bool lock);
51 static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
53 struct qcm_process_device *qpd);
55 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
56 unsigned int sdma_queue_id);
59 enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type)
61 if (type == KFD_QUEUE_TYPE_SDMA)
62 return KFD_MQD_TYPE_SDMA;
63 return KFD_MQD_TYPE_CP;
66 static bool is_pipe_enabled(struct device_queue_manager *dqm, int mec, int pipe)
69 int pipe_offset = mec * dqm->dev->shared_resources.num_pipe_per_mec
70 + pipe * dqm->dev->shared_resources.num_queue_per_pipe;
72 /* queue is available for KFD usage if bit is 1 */
73 for (i = 0; i < dqm->dev->shared_resources.num_queue_per_pipe; ++i)
74 if (test_bit(pipe_offset + i,
75 dqm->dev->shared_resources.queue_bitmap))
80 unsigned int get_mec_num(struct device_queue_manager *dqm)
82 BUG_ON(!dqm || !dqm->dev);
84 return dqm->dev->shared_resources.num_mec;
87 unsigned int get_queues_num(struct device_queue_manager *dqm)
89 BUG_ON(!dqm || !dqm->dev);
90 return bitmap_weight(dqm->dev->shared_resources.queue_bitmap,
94 unsigned int get_queues_per_pipe(struct device_queue_manager *dqm)
96 BUG_ON(!dqm || !dqm->dev);
97 return dqm->dev->shared_resources.num_queue_per_pipe;
100 unsigned int get_pipes_per_mec(struct device_queue_manager *dqm)
102 BUG_ON(!dqm || !dqm->dev);
103 return dqm->dev->shared_resources.num_pipe_per_mec;
106 void program_sh_mem_settings(struct device_queue_manager *dqm,
107 struct qcm_process_device *qpd)
109 return dqm->dev->kfd2kgd->program_sh_mem_settings(
110 dqm->dev->kgd, qpd->vmid,
112 qpd->sh_mem_ape1_base,
113 qpd->sh_mem_ape1_limit,
117 static int allocate_vmid(struct device_queue_manager *dqm,
118 struct qcm_process_device *qpd,
121 int bit, allocated_vmid;
123 if (dqm->vmid_bitmap == 0)
126 bit = find_first_bit((unsigned long *)&dqm->vmid_bitmap, CIK_VMID_NUM);
127 clear_bit(bit, (unsigned long *)&dqm->vmid_bitmap);
129 /* Kaveri kfd vmid's starts from vmid 8 */
130 allocated_vmid = bit + KFD_VMID_START_OFFSET;
131 pr_debug("kfd: vmid allocation %d\n", allocated_vmid);
132 qpd->vmid = allocated_vmid;
133 q->properties.vmid = allocated_vmid;
135 set_pasid_vmid_mapping(dqm, q->process->pasid, q->properties.vmid);
136 program_sh_mem_settings(dqm, qpd);
141 static void deallocate_vmid(struct device_queue_manager *dqm,
142 struct qcm_process_device *qpd,
145 int bit = qpd->vmid - KFD_VMID_START_OFFSET;
147 /* Release the vmid mapping */
148 set_pasid_vmid_mapping(dqm, 0, qpd->vmid);
150 set_bit(bit, (unsigned long *)&dqm->vmid_bitmap);
152 q->properties.vmid = 0;
155 static int create_queue_nocpsch(struct device_queue_manager *dqm,
157 struct qcm_process_device *qpd,
162 BUG_ON(!dqm || !q || !qpd || !allocated_vmid);
164 pr_debug("kfd: In func %s\n", __func__);
167 mutex_lock(&dqm->lock);
169 if (dqm->total_queue_count >= max_num_of_queues_per_device) {
170 pr_warn("amdkfd: Can't create new usermode queue because %d queues were already created\n",
171 dqm->total_queue_count);
172 mutex_unlock(&dqm->lock);
176 if (list_empty(&qpd->queues_list)) {
177 retval = allocate_vmid(dqm, qpd, q);
179 mutex_unlock(&dqm->lock);
183 *allocated_vmid = qpd->vmid;
184 q->properties.vmid = qpd->vmid;
186 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
187 retval = create_compute_queue_nocpsch(dqm, q, qpd);
188 if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
189 retval = create_sdma_queue_nocpsch(dqm, q, qpd);
192 if (list_empty(&qpd->queues_list)) {
193 deallocate_vmid(dqm, qpd, q);
196 mutex_unlock(&dqm->lock);
200 list_add(&q->list, &qpd->queues_list);
201 if (q->properties.is_active)
204 if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
205 dqm->sdma_queue_count++;
208 * Unconditionally increment this counter, regardless of the queue's
209 * type or whether the queue is active.
211 dqm->total_queue_count++;
212 pr_debug("Total of %d queues are accountable so far\n",
213 dqm->total_queue_count);
215 mutex_unlock(&dqm->lock);
219 static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q)
226 for (pipe = dqm->next_pipe_to_allocate, i = 0; i < get_pipes_per_mec(dqm);
227 pipe = ((pipe + 1) % get_pipes_per_mec(dqm)), ++i) {
229 if (!is_pipe_enabled(dqm, 0, pipe))
232 if (dqm->allocated_queues[pipe] != 0) {
233 bit = find_first_bit(
234 (unsigned long *)&dqm->allocated_queues[pipe],
235 get_queues_per_pipe(dqm));
238 (unsigned long *)&dqm->allocated_queues[pipe]);
249 pr_debug("kfd: DQM %s hqd slot - pipe (%d) queue(%d)\n",
250 __func__, q->pipe, q->queue);
251 /* horizontal hqd allocation */
252 dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_per_mec(dqm);
257 static inline void deallocate_hqd(struct device_queue_manager *dqm,
260 set_bit(q->queue, (unsigned long *)&dqm->allocated_queues[q->pipe]);
263 static int create_compute_queue_nocpsch(struct device_queue_manager *dqm,
265 struct qcm_process_device *qpd)
268 struct mqd_manager *mqd;
270 BUG_ON(!dqm || !q || !qpd);
272 mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_COMPUTE);
276 retval = allocate_hqd(dqm, q);
280 retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
281 &q->gart_mqd_addr, &q->properties);
283 deallocate_hqd(dqm, q);
287 pr_debug("kfd: loading mqd to hqd on pipe (%d) queue (%d)\n",
291 retval = mqd->load_mqd(mqd, q->mqd, q->pipe,
292 q->queue, (uint32_t __user *) q->properties.write_ptr);
294 deallocate_hqd(dqm, q);
295 mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
302 static int destroy_queue_nocpsch(struct device_queue_manager *dqm,
303 struct qcm_process_device *qpd,
307 struct mqd_manager *mqd;
309 BUG_ON(!dqm || !q || !q->mqd || !qpd);
313 pr_debug("kfd: In Func %s\n", __func__);
315 mutex_lock(&dqm->lock);
317 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) {
318 mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_COMPUTE);
323 deallocate_hqd(dqm, q);
324 } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
325 mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_SDMA);
330 dqm->sdma_queue_count--;
331 deallocate_sdma_queue(dqm, q->sdma_id);
333 pr_debug("q->properties.type is invalid (%d)\n",
339 retval = mqd->destroy_mqd(mqd, q->mqd,
340 KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
341 QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS,
347 mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
350 if (list_empty(&qpd->queues_list))
351 deallocate_vmid(dqm, qpd, q);
352 if (q->properties.is_active)
356 * Unconditionally decrement this counter, regardless of the queue's
359 dqm->total_queue_count--;
360 pr_debug("Total of %d queues are accountable so far\n",
361 dqm->total_queue_count);
364 mutex_unlock(&dqm->lock);
368 static int update_queue(struct device_queue_manager *dqm, struct queue *q)
371 struct mqd_manager *mqd;
372 bool prev_active = false;
374 BUG_ON(!dqm || !q || !q->mqd);
376 mutex_lock(&dqm->lock);
377 mqd = dqm->ops.get_mqd_manager(dqm,
378 get_mqd_type_from_queue_type(q->properties.type));
380 mutex_unlock(&dqm->lock);
384 if (q->properties.is_active)
389 * check active state vs. the previous state
390 * and modify counter accordingly
392 retval = mqd->update_mqd(mqd, q->mqd, &q->properties);
393 if ((q->properties.is_active) && (!prev_active))
395 else if ((!q->properties.is_active) && (prev_active))
398 if (sched_policy != KFD_SCHED_POLICY_NO_HWS)
399 retval = execute_queues_cpsch(dqm, false);
401 mutex_unlock(&dqm->lock);
405 static struct mqd_manager *get_mqd_manager_nocpsch(
406 struct device_queue_manager *dqm, enum KFD_MQD_TYPE type)
408 struct mqd_manager *mqd;
410 BUG_ON(!dqm || type >= KFD_MQD_TYPE_MAX);
412 pr_debug("kfd: In func %s mqd type %d\n", __func__, type);
414 mqd = dqm->mqds[type];
416 mqd = mqd_manager_init(type, dqm->dev);
418 pr_err("kfd: mqd manager is NULL");
419 dqm->mqds[type] = mqd;
425 static int register_process_nocpsch(struct device_queue_manager *dqm,
426 struct qcm_process_device *qpd)
428 struct device_process_node *n;
431 BUG_ON(!dqm || !qpd);
433 pr_debug("kfd: In func %s\n", __func__);
435 n = kzalloc(sizeof(struct device_process_node), GFP_KERNEL);
441 mutex_lock(&dqm->lock);
442 list_add(&n->list, &dqm->queues);
444 retval = dqm->ops_asic_specific.register_process(dqm, qpd);
446 dqm->processes_count++;
448 mutex_unlock(&dqm->lock);
453 static int unregister_process_nocpsch(struct device_queue_manager *dqm,
454 struct qcm_process_device *qpd)
457 struct device_process_node *cur, *next;
459 BUG_ON(!dqm || !qpd);
461 pr_debug("In func %s\n", __func__);
463 pr_debug("qpd->queues_list is %s\n",
464 list_empty(&qpd->queues_list) ? "empty" : "not empty");
467 mutex_lock(&dqm->lock);
469 list_for_each_entry_safe(cur, next, &dqm->queues, list) {
470 if (qpd == cur->qpd) {
471 list_del(&cur->list);
473 dqm->processes_count--;
477 /* qpd not found in dqm list */
480 mutex_unlock(&dqm->lock);
485 set_pasid_vmid_mapping(struct device_queue_manager *dqm, unsigned int pasid,
488 uint32_t pasid_mapping;
490 pasid_mapping = (pasid == 0) ? 0 :
492 ATC_VMID_PASID_MAPPING_VALID;
494 return dqm->dev->kfd2kgd->set_pasid_vmid_mapping(
495 dqm->dev->kgd, pasid_mapping,
499 static void init_interrupts(struct device_queue_manager *dqm)
505 for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++)
506 if (is_pipe_enabled(dqm, 0, i))
507 dqm->dev->kfd2kgd->init_interrupts(dqm->dev->kgd, i);
510 static int init_scheduler(struct device_queue_manager *dqm)
516 pr_debug("kfd: In %s\n", __func__);
521 static int initialize_nocpsch(struct device_queue_manager *dqm)
527 pr_debug("kfd: In func %s num of pipes: %d\n",
528 __func__, get_pipes_per_mec(dqm));
530 mutex_init(&dqm->lock);
531 INIT_LIST_HEAD(&dqm->queues);
532 dqm->queue_count = dqm->next_pipe_to_allocate = 0;
533 dqm->sdma_queue_count = 0;
534 dqm->allocated_queues = kcalloc(get_pipes_per_mec(dqm),
535 sizeof(unsigned int), GFP_KERNEL);
536 if (!dqm->allocated_queues) {
537 mutex_destroy(&dqm->lock);
541 for (i = 0; i < get_pipes_per_mec(dqm); i++)
542 dqm->allocated_queues[i] = (1 << get_queues_per_pipe(dqm)) - 1;
544 dqm->vmid_bitmap = (1 << VMID_PER_DEVICE) - 1;
545 dqm->sdma_bitmap = (1 << CIK_SDMA_QUEUES) - 1;
551 static void uninitialize_nocpsch(struct device_queue_manager *dqm)
557 BUG_ON(dqm->queue_count > 0 || dqm->processes_count > 0);
559 kfree(dqm->allocated_queues);
560 for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++)
562 mutex_destroy(&dqm->lock);
563 kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem);
566 static int start_nocpsch(struct device_queue_manager *dqm)
568 init_interrupts(dqm);
572 static int stop_nocpsch(struct device_queue_manager *dqm)
577 static int allocate_sdma_queue(struct device_queue_manager *dqm,
578 unsigned int *sdma_queue_id)
582 if (dqm->sdma_bitmap == 0)
585 bit = find_first_bit((unsigned long *)&dqm->sdma_bitmap,
588 clear_bit(bit, (unsigned long *)&dqm->sdma_bitmap);
589 *sdma_queue_id = bit;
594 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
595 unsigned int sdma_queue_id)
597 if (sdma_queue_id >= CIK_SDMA_QUEUES)
599 set_bit(sdma_queue_id, (unsigned long *)&dqm->sdma_bitmap);
602 static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
604 struct qcm_process_device *qpd)
606 struct mqd_manager *mqd;
609 mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_SDMA);
613 retval = allocate_sdma_queue(dqm, &q->sdma_id);
617 q->properties.sdma_queue_id = q->sdma_id % CIK_SDMA_QUEUES_PER_ENGINE;
618 q->properties.sdma_engine_id = q->sdma_id / CIK_SDMA_ENGINE_NUM;
620 pr_debug("kfd: sdma id is: %d\n", q->sdma_id);
621 pr_debug(" sdma queue id: %d\n", q->properties.sdma_queue_id);
622 pr_debug(" sdma engine id: %d\n", q->properties.sdma_engine_id);
624 dqm->ops_asic_specific.init_sdma_vm(dqm, q, qpd);
625 retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
626 &q->gart_mqd_addr, &q->properties);
628 deallocate_sdma_queue(dqm, q->sdma_id);
632 retval = mqd->load_mqd(mqd, q->mqd, 0,
635 deallocate_sdma_queue(dqm, q->sdma_id);
636 mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
644 * Device Queue Manager implementation for cp scheduler
647 static int set_sched_resources(struct device_queue_manager *dqm)
650 struct scheduling_resources res;
654 pr_debug("kfd: In func %s\n", __func__);
656 res.vmid_mask = (1 << VMID_PER_DEVICE) - 1;
657 res.vmid_mask <<= KFD_VMID_START_OFFSET;
660 for (i = 0; i < KGD_MAX_QUEUES; ++i) {
661 mec = (i / dqm->dev->shared_resources.num_queue_per_pipe)
662 / dqm->dev->shared_resources.num_pipe_per_mec;
664 if (!test_bit(i, dqm->dev->shared_resources.queue_bitmap))
667 /* only acquire queues from the first MEC */
671 /* This situation may be hit in the future if a new HW
672 * generation exposes more than 64 queues. If so, the
673 * definition of res.queue_mask needs updating */
674 if (WARN_ON(i > (sizeof(res.queue_mask)*8))) {
675 pr_err("Invalid queue enabled by amdgpu: %d\n", i);
679 res.queue_mask |= (1ull << i);
681 res.gws_mask = res.oac_mask = res.gds_heap_base =
682 res.gds_heap_size = 0;
684 pr_debug("kfd: scheduling resources:\n"
685 " vmid mask: 0x%8X\n"
686 " queue mask: 0x%8llX\n",
687 res.vmid_mask, res.queue_mask);
689 return pm_send_set_resources(&dqm->packets, &res);
692 static int initialize_cpsch(struct device_queue_manager *dqm)
698 pr_debug("kfd: In func %s num of pipes: %d\n",
699 __func__, get_pipes_per_mec(dqm));
701 mutex_init(&dqm->lock);
702 INIT_LIST_HEAD(&dqm->queues);
703 dqm->queue_count = dqm->processes_count = 0;
704 dqm->sdma_queue_count = 0;
705 dqm->active_runlist = false;
706 retval = dqm->ops_asic_specific.initialize(dqm);
708 goto fail_init_pipelines;
713 mutex_destroy(&dqm->lock);
717 static int start_cpsch(struct device_queue_manager *dqm)
719 struct device_process_node *node;
726 retval = pm_init(&dqm->packets, dqm);
728 goto fail_packet_manager_init;
730 retval = set_sched_resources(dqm);
732 goto fail_set_sched_resources;
734 pr_debug("kfd: allocating fence memory\n");
736 /* allocate fence memory on the gart */
737 retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr),
741 goto fail_allocate_vidmem;
743 dqm->fence_addr = dqm->fence_mem->cpu_ptr;
744 dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr;
746 init_interrupts(dqm);
748 list_for_each_entry(node, &dqm->queues, list)
749 if (node->qpd->pqm->process && dqm->dev)
750 kfd_bind_process_to_device(dqm->dev,
751 node->qpd->pqm->process);
753 execute_queues_cpsch(dqm, true);
756 fail_allocate_vidmem:
757 fail_set_sched_resources:
758 pm_uninit(&dqm->packets);
759 fail_packet_manager_init:
763 static int stop_cpsch(struct device_queue_manager *dqm)
765 struct device_process_node *node;
766 struct kfd_process_device *pdd;
770 destroy_queues_cpsch(dqm, true, true);
772 list_for_each_entry(node, &dqm->queues, list) {
773 pdd = qpd_to_pdd(node->qpd);
776 kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
777 pm_uninit(&dqm->packets);
782 static int create_kernel_queue_cpsch(struct device_queue_manager *dqm,
783 struct kernel_queue *kq,
784 struct qcm_process_device *qpd)
786 BUG_ON(!dqm || !kq || !qpd);
788 pr_debug("kfd: In func %s\n", __func__);
790 mutex_lock(&dqm->lock);
791 if (dqm->total_queue_count >= max_num_of_queues_per_device) {
792 pr_warn("amdkfd: Can't create new kernel queue because %d queues were already created\n",
793 dqm->total_queue_count);
794 mutex_unlock(&dqm->lock);
799 * Unconditionally increment this counter, regardless of the queue's
800 * type or whether the queue is active.
802 dqm->total_queue_count++;
803 pr_debug("Total of %d queues are accountable so far\n",
804 dqm->total_queue_count);
806 list_add(&kq->list, &qpd->priv_queue_list);
808 qpd->is_debug = true;
809 execute_queues_cpsch(dqm, false);
810 mutex_unlock(&dqm->lock);
815 static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm,
816 struct kernel_queue *kq,
817 struct qcm_process_device *qpd)
821 pr_debug("kfd: In %s\n", __func__);
823 mutex_lock(&dqm->lock);
824 /* here we actually preempt the DIQ */
825 destroy_queues_cpsch(dqm, true, false);
828 qpd->is_debug = false;
829 execute_queues_cpsch(dqm, false);
831 * Unconditionally decrement this counter, regardless of the queue's
834 dqm->total_queue_count--;
835 pr_debug("Total of %d queues are accountable so far\n",
836 dqm->total_queue_count);
837 mutex_unlock(&dqm->lock);
840 static void select_sdma_engine_id(struct queue *q)
844 q->sdma_id = sdma_id;
845 sdma_id = (sdma_id + 1) % 2;
848 static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
849 struct qcm_process_device *qpd, int *allocate_vmid)
852 struct mqd_manager *mqd;
854 BUG_ON(!dqm || !q || !qpd);
861 mutex_lock(&dqm->lock);
863 if (dqm->total_queue_count >= max_num_of_queues_per_device) {
864 pr_warn("amdkfd: Can't create new usermode queue because %d queues were already created\n",
865 dqm->total_queue_count);
870 if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
871 select_sdma_engine_id(q);
873 mqd = dqm->ops.get_mqd_manager(dqm,
874 get_mqd_type_from_queue_type(q->properties.type));
877 mutex_unlock(&dqm->lock);
881 dqm->ops_asic_specific.init_sdma_vm(dqm, q, qpd);
882 retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
883 &q->gart_mqd_addr, &q->properties);
887 list_add(&q->list, &qpd->queues_list);
888 if (q->properties.is_active) {
890 retval = execute_queues_cpsch(dqm, false);
893 if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
894 dqm->sdma_queue_count++;
896 * Unconditionally increment this counter, regardless of the queue's
897 * type or whether the queue is active.
899 dqm->total_queue_count++;
901 pr_debug("Total of %d queues are accountable so far\n",
902 dqm->total_queue_count);
905 mutex_unlock(&dqm->lock);
909 int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
910 unsigned int fence_value,
911 unsigned long timeout)
916 while (*fence_addr != fence_value) {
917 if (time_after(jiffies, timeout)) {
918 pr_err("kfd: qcm fence wait loop timeout expired\n");
927 static int destroy_sdma_queues(struct device_queue_manager *dqm,
928 unsigned int sdma_engine)
930 return pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_SDMA,
931 KFD_PREEMPT_TYPE_FILTER_DYNAMIC_QUEUES, 0, false,
935 static int destroy_queues_cpsch(struct device_queue_manager *dqm,
936 bool preempt_static_queues, bool lock)
939 enum kfd_preempt_type_filter preempt_type;
940 struct kfd_process_device *pdd;
947 mutex_lock(&dqm->lock);
948 if (!dqm->active_runlist)
951 pr_debug("kfd: Before destroying queues, sdma queue count is : %u\n",
952 dqm->sdma_queue_count);
954 if (dqm->sdma_queue_count > 0) {
955 destroy_sdma_queues(dqm, 0);
956 destroy_sdma_queues(dqm, 1);
959 preempt_type = preempt_static_queues ?
960 KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES :
961 KFD_PREEMPT_TYPE_FILTER_DYNAMIC_QUEUES;
963 retval = pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_COMPUTE,
964 preempt_type, 0, false, 0);
968 *dqm->fence_addr = KFD_FENCE_INIT;
969 pm_send_query_status(&dqm->packets, dqm->fence_gpu_addr,
970 KFD_FENCE_COMPLETED);
971 /* should be timed out */
972 retval = amdkfd_fence_wait_timeout(dqm->fence_addr, KFD_FENCE_COMPLETED,
973 QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS);
975 pdd = kfd_get_process_device_data(dqm->dev,
976 kfd_get_process(current));
977 pdd->reset_wavefronts = true;
980 pm_release_ib(&dqm->packets);
981 dqm->active_runlist = false;
985 mutex_unlock(&dqm->lock);
989 static int execute_queues_cpsch(struct device_queue_manager *dqm, bool lock)
996 mutex_lock(&dqm->lock);
998 retval = destroy_queues_cpsch(dqm, false, false);
1000 pr_err("kfd: the cp might be in an unrecoverable state due to an unsuccessful queues preemption");
1004 if (dqm->queue_count <= 0 || dqm->processes_count <= 0) {
1009 if (dqm->active_runlist) {
1014 retval = pm_send_runlist(&dqm->packets, &dqm->queues);
1016 pr_err("kfd: failed to execute runlist");
1019 dqm->active_runlist = true;
1023 mutex_unlock(&dqm->lock);
1027 static int destroy_queue_cpsch(struct device_queue_manager *dqm,
1028 struct qcm_process_device *qpd,
1032 struct mqd_manager *mqd;
1033 bool preempt_all_queues;
1035 BUG_ON(!dqm || !qpd || !q);
1037 preempt_all_queues = false;
1041 /* remove queue from list to prevent rescheduling after preemption */
1042 mutex_lock(&dqm->lock);
1044 if (qpd->is_debug) {
1046 * error, currently we do not allow to destroy a queue
1047 * of a currently debugged process
1050 goto failed_try_destroy_debugged_queue;
1054 mqd = dqm->ops.get_mqd_manager(dqm,
1055 get_mqd_type_from_queue_type(q->properties.type));
1061 if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
1062 dqm->sdma_queue_count--;
1065 if (q->properties.is_active)
1068 execute_queues_cpsch(dqm, false);
1070 mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
1073 * Unconditionally decrement this counter, regardless of the queue's
1076 dqm->total_queue_count--;
1077 pr_debug("Total of %d queues are accountable so far\n",
1078 dqm->total_queue_count);
1080 mutex_unlock(&dqm->lock);
1085 failed_try_destroy_debugged_queue:
1087 mutex_unlock(&dqm->lock);
1092 * Low bits must be 0000/FFFF as required by HW, high bits must be 0 to
1093 * stay in user mode.
1095 #define APE1_FIXED_BITS_MASK 0xFFFF80000000FFFFULL
1096 /* APE1 limit is inclusive and 64K aligned. */
1097 #define APE1_LIMIT_ALIGNMENT 0xFFFF
1099 static bool set_cache_memory_policy(struct device_queue_manager *dqm,
1100 struct qcm_process_device *qpd,
1101 enum cache_policy default_policy,
1102 enum cache_policy alternate_policy,
1103 void __user *alternate_aperture_base,
1104 uint64_t alternate_aperture_size)
1108 pr_debug("kfd: In func %s\n", __func__);
1110 mutex_lock(&dqm->lock);
1112 if (alternate_aperture_size == 0) {
1113 /* base > limit disables APE1 */
1114 qpd->sh_mem_ape1_base = 1;
1115 qpd->sh_mem_ape1_limit = 0;
1118 * In FSA64, APE1_Base[63:0] = { 16{SH_MEM_APE1_BASE[31]},
1119 * SH_MEM_APE1_BASE[31:0], 0x0000 }
1120 * APE1_Limit[63:0] = { 16{SH_MEM_APE1_LIMIT[31]},
1121 * SH_MEM_APE1_LIMIT[31:0], 0xFFFF }
1122 * Verify that the base and size parameters can be
1123 * represented in this format and convert them.
1124 * Additionally restrict APE1 to user-mode addresses.
1127 uint64_t base = (uintptr_t)alternate_aperture_base;
1128 uint64_t limit = base + alternate_aperture_size - 1;
1133 if ((base & APE1_FIXED_BITS_MASK) != 0)
1136 if ((limit & APE1_FIXED_BITS_MASK) != APE1_LIMIT_ALIGNMENT)
1139 qpd->sh_mem_ape1_base = base >> 16;
1140 qpd->sh_mem_ape1_limit = limit >> 16;
1143 retval = dqm->ops_asic_specific.set_cache_memory_policy(
1148 alternate_aperture_base,
1149 alternate_aperture_size);
1151 if ((sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0))
1152 program_sh_mem_settings(dqm, qpd);
1154 pr_debug("kfd: sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n",
1155 qpd->sh_mem_config, qpd->sh_mem_ape1_base,
1156 qpd->sh_mem_ape1_limit);
1158 mutex_unlock(&dqm->lock);
1162 mutex_unlock(&dqm->lock);
1166 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
1168 struct device_queue_manager *dqm;
1172 pr_debug("kfd: loading device queue manager\n");
1174 dqm = kzalloc(sizeof(struct device_queue_manager), GFP_KERNEL);
1179 switch (sched_policy) {
1180 case KFD_SCHED_POLICY_HWS:
1181 case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION:
1182 /* initialize dqm for cp scheduling */
1183 dqm->ops.create_queue = create_queue_cpsch;
1184 dqm->ops.initialize = initialize_cpsch;
1185 dqm->ops.start = start_cpsch;
1186 dqm->ops.stop = stop_cpsch;
1187 dqm->ops.destroy_queue = destroy_queue_cpsch;
1188 dqm->ops.update_queue = update_queue;
1189 dqm->ops.get_mqd_manager = get_mqd_manager_nocpsch;
1190 dqm->ops.register_process = register_process_nocpsch;
1191 dqm->ops.unregister_process = unregister_process_nocpsch;
1192 dqm->ops.uninitialize = uninitialize_nocpsch;
1193 dqm->ops.create_kernel_queue = create_kernel_queue_cpsch;
1194 dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch;
1195 dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1197 case KFD_SCHED_POLICY_NO_HWS:
1198 /* initialize dqm for no cp scheduling */
1199 dqm->ops.start = start_nocpsch;
1200 dqm->ops.stop = stop_nocpsch;
1201 dqm->ops.create_queue = create_queue_nocpsch;
1202 dqm->ops.destroy_queue = destroy_queue_nocpsch;
1203 dqm->ops.update_queue = update_queue;
1204 dqm->ops.get_mqd_manager = get_mqd_manager_nocpsch;
1205 dqm->ops.register_process = register_process_nocpsch;
1206 dqm->ops.unregister_process = unregister_process_nocpsch;
1207 dqm->ops.initialize = initialize_nocpsch;
1208 dqm->ops.uninitialize = uninitialize_nocpsch;
1209 dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1216 switch (dev->device_info->asic_family) {
1218 device_queue_manager_init_vi(&dqm->ops_asic_specific);
1222 device_queue_manager_init_cik(&dqm->ops_asic_specific);
1226 if (dqm->ops.initialize(dqm) != 0) {
1234 void device_queue_manager_uninit(struct device_queue_manager *dqm)
1238 dqm->ops.uninitialize(dqm);