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 unsigned int get_first_pipe(struct device_queue_manager *dqm)
68 BUG_ON(!dqm || !dqm->dev);
69 return dqm->dev->shared_resources.first_compute_pipe;
72 unsigned int get_pipes_num(struct device_queue_manager *dqm)
74 BUG_ON(!dqm || !dqm->dev);
75 return dqm->dev->shared_resources.compute_pipe_count;
78 static inline unsigned int get_pipes_num_cpsch(void)
80 return PIPE_PER_ME_CP_SCHEDULING;
83 void program_sh_mem_settings(struct device_queue_manager *dqm,
84 struct qcm_process_device *qpd)
86 return dqm->dev->kfd2kgd->program_sh_mem_settings(
87 dqm->dev->kgd, qpd->vmid,
89 qpd->sh_mem_ape1_base,
90 qpd->sh_mem_ape1_limit,
94 static int allocate_vmid(struct device_queue_manager *dqm,
95 struct qcm_process_device *qpd,
98 int bit, allocated_vmid;
100 if (dqm->vmid_bitmap == 0)
103 bit = find_first_bit((unsigned long *)&dqm->vmid_bitmap, CIK_VMID_NUM);
104 clear_bit(bit, (unsigned long *)&dqm->vmid_bitmap);
106 /* Kaveri kfd vmid's starts from vmid 8 */
107 allocated_vmid = bit + KFD_VMID_START_OFFSET;
108 pr_debug("kfd: vmid allocation %d\n", allocated_vmid);
109 qpd->vmid = allocated_vmid;
110 q->properties.vmid = allocated_vmid;
112 set_pasid_vmid_mapping(dqm, q->process->pasid, q->properties.vmid);
113 program_sh_mem_settings(dqm, qpd);
118 static void deallocate_vmid(struct device_queue_manager *dqm,
119 struct qcm_process_device *qpd,
122 int bit = qpd->vmid - KFD_VMID_START_OFFSET;
124 /* Release the vmid mapping */
125 set_pasid_vmid_mapping(dqm, 0, qpd->vmid);
127 set_bit(bit, (unsigned long *)&dqm->vmid_bitmap);
129 q->properties.vmid = 0;
132 static int create_queue_nocpsch(struct device_queue_manager *dqm,
134 struct qcm_process_device *qpd,
139 BUG_ON(!dqm || !q || !qpd || !allocated_vmid);
141 pr_debug("kfd: In func %s\n", __func__);
144 mutex_lock(&dqm->lock);
146 if (dqm->total_queue_count >= max_num_of_queues_per_device) {
147 pr_warn("amdkfd: Can't create new usermode queue because %d queues were already created\n",
148 dqm->total_queue_count);
149 mutex_unlock(&dqm->lock);
153 if (list_empty(&qpd->queues_list)) {
154 retval = allocate_vmid(dqm, qpd, q);
156 mutex_unlock(&dqm->lock);
160 *allocated_vmid = qpd->vmid;
161 q->properties.vmid = qpd->vmid;
163 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
164 retval = create_compute_queue_nocpsch(dqm, q, qpd);
165 if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
166 retval = create_sdma_queue_nocpsch(dqm, q, qpd);
169 if (list_empty(&qpd->queues_list)) {
170 deallocate_vmid(dqm, qpd, q);
173 mutex_unlock(&dqm->lock);
177 list_add(&q->list, &qpd->queues_list);
178 if (q->properties.is_active)
181 if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
182 dqm->sdma_queue_count++;
185 * Unconditionally increment this counter, regardless of the queue's
186 * type or whether the queue is active.
188 dqm->total_queue_count++;
189 pr_debug("Total of %d queues are accountable so far\n",
190 dqm->total_queue_count);
192 mutex_unlock(&dqm->lock);
196 static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q)
203 for (pipe = dqm->next_pipe_to_allocate, i = 0; i < get_pipes_num(dqm);
204 pipe = ((pipe + 1) % get_pipes_num(dqm)), ++i) {
205 if (dqm->allocated_queues[pipe] != 0) {
206 bit = find_first_bit(
207 (unsigned long *)&dqm->allocated_queues[pipe],
211 (unsigned long *)&dqm->allocated_queues[pipe]);
222 pr_debug("kfd: DQM %s hqd slot - pipe (%d) queue(%d)\n",
223 __func__, q->pipe, q->queue);
224 /* horizontal hqd allocation */
225 dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_num(dqm);
230 static inline void deallocate_hqd(struct device_queue_manager *dqm,
233 set_bit(q->queue, (unsigned long *)&dqm->allocated_queues[q->pipe]);
236 static int create_compute_queue_nocpsch(struct device_queue_manager *dqm,
238 struct qcm_process_device *qpd)
241 struct mqd_manager *mqd;
243 BUG_ON(!dqm || !q || !qpd);
245 mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_COMPUTE);
249 retval = allocate_hqd(dqm, q);
253 retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
254 &q->gart_mqd_addr, &q->properties);
256 deallocate_hqd(dqm, q);
260 pr_debug("kfd: loading mqd to hqd on pipe (%d) queue (%d)\n",
264 retval = mqd->load_mqd(mqd, q->mqd, q->pipe,
265 q->queue, (uint32_t __user *) q->properties.write_ptr);
267 deallocate_hqd(dqm, q);
268 mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
275 static int destroy_queue_nocpsch(struct device_queue_manager *dqm,
276 struct qcm_process_device *qpd,
280 struct mqd_manager *mqd;
282 BUG_ON(!dqm || !q || !q->mqd || !qpd);
286 pr_debug("kfd: In Func %s\n", __func__);
288 mutex_lock(&dqm->lock);
290 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) {
291 mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_COMPUTE);
296 deallocate_hqd(dqm, q);
297 } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
298 mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_SDMA);
303 dqm->sdma_queue_count--;
304 deallocate_sdma_queue(dqm, q->sdma_id);
306 pr_debug("q->properties.type is invalid (%d)\n",
312 retval = mqd->destroy_mqd(mqd, q->mqd,
313 KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
314 QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS,
320 mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
323 if (list_empty(&qpd->queues_list))
324 deallocate_vmid(dqm, qpd, q);
325 if (q->properties.is_active)
329 * Unconditionally decrement this counter, regardless of the queue's
332 dqm->total_queue_count--;
333 pr_debug("Total of %d queues are accountable so far\n",
334 dqm->total_queue_count);
337 mutex_unlock(&dqm->lock);
341 static int update_queue(struct device_queue_manager *dqm, struct queue *q)
344 struct mqd_manager *mqd;
345 bool prev_active = false;
347 BUG_ON(!dqm || !q || !q->mqd);
349 mutex_lock(&dqm->lock);
350 mqd = dqm->ops.get_mqd_manager(dqm,
351 get_mqd_type_from_queue_type(q->properties.type));
353 mutex_unlock(&dqm->lock);
357 if (q->properties.is_active)
362 * check active state vs. the previous state
363 * and modify counter accordingly
365 retval = mqd->update_mqd(mqd, q->mqd, &q->properties);
366 if ((q->properties.is_active) && (!prev_active))
368 else if ((!q->properties.is_active) && (prev_active))
371 if (sched_policy != KFD_SCHED_POLICY_NO_HWS)
372 retval = execute_queues_cpsch(dqm, false);
374 mutex_unlock(&dqm->lock);
378 static struct mqd_manager *get_mqd_manager_nocpsch(
379 struct device_queue_manager *dqm, enum KFD_MQD_TYPE type)
381 struct mqd_manager *mqd;
383 BUG_ON(!dqm || type >= KFD_MQD_TYPE_MAX);
385 pr_debug("kfd: In func %s mqd type %d\n", __func__, type);
387 mqd = dqm->mqds[type];
389 mqd = mqd_manager_init(type, dqm->dev);
391 pr_err("kfd: mqd manager is NULL");
392 dqm->mqds[type] = mqd;
398 static int register_process_nocpsch(struct device_queue_manager *dqm,
399 struct qcm_process_device *qpd)
401 struct device_process_node *n;
404 BUG_ON(!dqm || !qpd);
406 pr_debug("kfd: In func %s\n", __func__);
408 n = kzalloc(sizeof(struct device_process_node), GFP_KERNEL);
414 mutex_lock(&dqm->lock);
415 list_add(&n->list, &dqm->queues);
417 retval = dqm->ops_asic_specific.register_process(dqm, qpd);
419 dqm->processes_count++;
421 mutex_unlock(&dqm->lock);
426 static int unregister_process_nocpsch(struct device_queue_manager *dqm,
427 struct qcm_process_device *qpd)
430 struct device_process_node *cur, *next;
432 BUG_ON(!dqm || !qpd);
434 pr_debug("In func %s\n", __func__);
436 pr_debug("qpd->queues_list is %s\n",
437 list_empty(&qpd->queues_list) ? "empty" : "not empty");
440 mutex_lock(&dqm->lock);
442 list_for_each_entry_safe(cur, next, &dqm->queues, list) {
443 if (qpd == cur->qpd) {
444 list_del(&cur->list);
446 dqm->processes_count--;
450 /* qpd not found in dqm list */
453 mutex_unlock(&dqm->lock);
458 set_pasid_vmid_mapping(struct device_queue_manager *dqm, unsigned int pasid,
461 uint32_t pasid_mapping;
463 pasid_mapping = (pasid == 0) ? 0 :
465 ATC_VMID_PASID_MAPPING_VALID;
467 return dqm->dev->kfd2kgd->set_pasid_vmid_mapping(
468 dqm->dev->kgd, pasid_mapping,
472 int init_pipelines(struct device_queue_manager *dqm,
473 unsigned int pipes_num, unsigned int first_pipe)
476 struct mqd_manager *mqd;
477 unsigned int i, err, inx;
478 uint64_t pipe_hpd_addr;
480 BUG_ON(!dqm || !dqm->dev);
482 pr_debug("kfd: In func %s\n", __func__);
485 * Allocate memory for the HPDs. This is hardware-owned per-pipe data.
486 * The driver never accesses this memory after zeroing it.
487 * It doesn't even have to be saved/restored on suspend/resume
488 * because it contains no data when there are no active queues.
491 err = kfd_gtt_sa_allocate(dqm->dev, CIK_HPD_EOP_BYTES * pipes_num,
495 pr_err("kfd: error allocate vidmem num pipes: %d\n",
500 hpdptr = dqm->pipeline_mem->cpu_ptr;
501 dqm->pipelines_addr = dqm->pipeline_mem->gpu_addr;
503 memset(hpdptr, 0, CIK_HPD_EOP_BYTES * pipes_num);
505 mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_COMPUTE);
507 kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem);
511 for (i = 0; i < pipes_num; i++) {
512 inx = i + first_pipe;
514 * HPD buffer on GTT is allocated by amdkfd, no need to waste
515 * space in GTT for pipelines we don't initialize
517 pipe_hpd_addr = dqm->pipelines_addr + i * CIK_HPD_EOP_BYTES;
518 pr_debug("kfd: pipeline address %llX\n", pipe_hpd_addr);
519 /* = log2(bytes/4)-1 */
520 dqm->dev->kfd2kgd->init_pipeline(dqm->dev->kgd, inx,
521 CIK_HPD_EOP_BYTES_LOG2 - 3, pipe_hpd_addr);
527 static void init_interrupts(struct device_queue_manager *dqm)
533 for (i = 0 ; i < get_pipes_num(dqm) ; i++)
534 dqm->dev->kfd2kgd->init_interrupts(dqm->dev->kgd,
535 i + get_first_pipe(dqm));
538 static int init_scheduler(struct device_queue_manager *dqm)
544 pr_debug("kfd: In %s\n", __func__);
546 retval = init_pipelines(dqm, get_pipes_num(dqm), get_first_pipe(dqm));
550 static int initialize_nocpsch(struct device_queue_manager *dqm)
556 pr_debug("kfd: In func %s num of pipes: %d\n",
557 __func__, get_pipes_num(dqm));
559 mutex_init(&dqm->lock);
560 INIT_LIST_HEAD(&dqm->queues);
561 dqm->queue_count = dqm->next_pipe_to_allocate = 0;
562 dqm->sdma_queue_count = 0;
563 dqm->allocated_queues = kcalloc(get_pipes_num(dqm),
564 sizeof(unsigned int), GFP_KERNEL);
565 if (!dqm->allocated_queues) {
566 mutex_destroy(&dqm->lock);
570 for (i = 0; i < get_pipes_num(dqm); i++)
571 dqm->allocated_queues[i] = (1 << QUEUES_PER_PIPE) - 1;
573 dqm->vmid_bitmap = (1 << VMID_PER_DEVICE) - 1;
574 dqm->sdma_bitmap = (1 << CIK_SDMA_QUEUES) - 1;
580 static void uninitialize_nocpsch(struct device_queue_manager *dqm)
586 BUG_ON(dqm->queue_count > 0 || dqm->processes_count > 0);
588 kfree(dqm->allocated_queues);
589 for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++)
591 mutex_destroy(&dqm->lock);
592 kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem);
595 static int start_nocpsch(struct device_queue_manager *dqm)
597 init_interrupts(dqm);
601 static int stop_nocpsch(struct device_queue_manager *dqm)
606 static int allocate_sdma_queue(struct device_queue_manager *dqm,
607 unsigned int *sdma_queue_id)
611 if (dqm->sdma_bitmap == 0)
614 bit = find_first_bit((unsigned long *)&dqm->sdma_bitmap,
617 clear_bit(bit, (unsigned long *)&dqm->sdma_bitmap);
618 *sdma_queue_id = bit;
623 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
624 unsigned int sdma_queue_id)
626 if (sdma_queue_id >= CIK_SDMA_QUEUES)
628 set_bit(sdma_queue_id, (unsigned long *)&dqm->sdma_bitmap);
631 static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
633 struct qcm_process_device *qpd)
635 struct mqd_manager *mqd;
638 mqd = dqm->ops.get_mqd_manager(dqm, KFD_MQD_TYPE_SDMA);
642 retval = allocate_sdma_queue(dqm, &q->sdma_id);
646 q->properties.sdma_queue_id = q->sdma_id % CIK_SDMA_QUEUES_PER_ENGINE;
647 q->properties.sdma_engine_id = q->sdma_id / CIK_SDMA_ENGINE_NUM;
649 pr_debug("kfd: sdma id is: %d\n", q->sdma_id);
650 pr_debug(" sdma queue id: %d\n", q->properties.sdma_queue_id);
651 pr_debug(" sdma engine id: %d\n", q->properties.sdma_engine_id);
653 dqm->ops_asic_specific.init_sdma_vm(dqm, q, qpd);
654 retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
655 &q->gart_mqd_addr, &q->properties);
657 deallocate_sdma_queue(dqm, q->sdma_id);
661 retval = mqd->load_mqd(mqd, q->mqd, 0,
664 deallocate_sdma_queue(dqm, q->sdma_id);
665 mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
673 * Device Queue Manager implementation for cp scheduler
676 static int set_sched_resources(struct device_queue_manager *dqm)
678 struct scheduling_resources res;
679 unsigned int queue_num, queue_mask;
683 pr_debug("kfd: In func %s\n", __func__);
685 queue_num = get_pipes_num_cpsch() * QUEUES_PER_PIPE;
686 queue_mask = (1 << queue_num) - 1;
687 res.vmid_mask = (1 << VMID_PER_DEVICE) - 1;
688 res.vmid_mask <<= KFD_VMID_START_OFFSET;
689 res.queue_mask = queue_mask << (get_first_pipe(dqm) * QUEUES_PER_PIPE);
690 res.gws_mask = res.oac_mask = res.gds_heap_base =
691 res.gds_heap_size = 0;
693 pr_debug("kfd: scheduling resources:\n"
694 " vmid mask: 0x%8X\n"
695 " queue mask: 0x%8llX\n",
696 res.vmid_mask, res.queue_mask);
698 return pm_send_set_resources(&dqm->packets, &res);
701 static int initialize_cpsch(struct device_queue_manager *dqm)
707 pr_debug("kfd: In func %s num of pipes: %d\n",
708 __func__, get_pipes_num_cpsch());
710 mutex_init(&dqm->lock);
711 INIT_LIST_HEAD(&dqm->queues);
712 dqm->queue_count = dqm->processes_count = 0;
713 dqm->sdma_queue_count = 0;
714 dqm->active_runlist = false;
715 retval = dqm->ops_asic_specific.initialize(dqm);
717 goto fail_init_pipelines;
722 mutex_destroy(&dqm->lock);
726 static int start_cpsch(struct device_queue_manager *dqm)
728 struct device_process_node *node;
735 retval = pm_init(&dqm->packets, dqm);
737 goto fail_packet_manager_init;
739 retval = set_sched_resources(dqm);
741 goto fail_set_sched_resources;
743 pr_debug("kfd: allocating fence memory\n");
745 /* allocate fence memory on the gart */
746 retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr),
750 goto fail_allocate_vidmem;
752 dqm->fence_addr = dqm->fence_mem->cpu_ptr;
753 dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr;
755 init_interrupts(dqm);
757 list_for_each_entry(node, &dqm->queues, list)
758 if (node->qpd->pqm->process && dqm->dev)
759 kfd_bind_process_to_device(dqm->dev,
760 node->qpd->pqm->process);
762 execute_queues_cpsch(dqm, true);
765 fail_allocate_vidmem:
766 fail_set_sched_resources:
767 pm_uninit(&dqm->packets);
768 fail_packet_manager_init:
772 static int stop_cpsch(struct device_queue_manager *dqm)
774 struct device_process_node *node;
775 struct kfd_process_device *pdd;
779 destroy_queues_cpsch(dqm, true, true);
781 list_for_each_entry(node, &dqm->queues, list) {
782 pdd = qpd_to_pdd(node->qpd);
785 kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
786 pm_uninit(&dqm->packets);
791 static int create_kernel_queue_cpsch(struct device_queue_manager *dqm,
792 struct kernel_queue *kq,
793 struct qcm_process_device *qpd)
795 BUG_ON(!dqm || !kq || !qpd);
797 pr_debug("kfd: In func %s\n", __func__);
799 mutex_lock(&dqm->lock);
800 if (dqm->total_queue_count >= max_num_of_queues_per_device) {
801 pr_warn("amdkfd: Can't create new kernel queue because %d queues were already created\n",
802 dqm->total_queue_count);
803 mutex_unlock(&dqm->lock);
808 * Unconditionally increment this counter, regardless of the queue's
809 * type or whether the queue is active.
811 dqm->total_queue_count++;
812 pr_debug("Total of %d queues are accountable so far\n",
813 dqm->total_queue_count);
815 list_add(&kq->list, &qpd->priv_queue_list);
817 qpd->is_debug = true;
818 execute_queues_cpsch(dqm, false);
819 mutex_unlock(&dqm->lock);
824 static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm,
825 struct kernel_queue *kq,
826 struct qcm_process_device *qpd)
830 pr_debug("kfd: In %s\n", __func__);
832 mutex_lock(&dqm->lock);
833 /* here we actually preempt the DIQ */
834 destroy_queues_cpsch(dqm, true, false);
837 qpd->is_debug = false;
838 execute_queues_cpsch(dqm, false);
840 * Unconditionally decrement this counter, regardless of the queue's
843 dqm->total_queue_count--;
844 pr_debug("Total of %d queues are accountable so far\n",
845 dqm->total_queue_count);
846 mutex_unlock(&dqm->lock);
849 static void select_sdma_engine_id(struct queue *q)
853 q->sdma_id = sdma_id;
854 sdma_id = (sdma_id + 1) % 2;
857 static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
858 struct qcm_process_device *qpd, int *allocate_vmid)
861 struct mqd_manager *mqd;
863 BUG_ON(!dqm || !q || !qpd);
870 mutex_lock(&dqm->lock);
872 if (dqm->total_queue_count >= max_num_of_queues_per_device) {
873 pr_warn("amdkfd: Can't create new usermode queue because %d queues were already created\n",
874 dqm->total_queue_count);
879 if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
880 select_sdma_engine_id(q);
882 mqd = dqm->ops.get_mqd_manager(dqm,
883 get_mqd_type_from_queue_type(q->properties.type));
886 mutex_unlock(&dqm->lock);
890 dqm->ops_asic_specific.init_sdma_vm(dqm, q, qpd);
891 retval = mqd->init_mqd(mqd, &q->mqd, &q->mqd_mem_obj,
892 &q->gart_mqd_addr, &q->properties);
896 list_add(&q->list, &qpd->queues_list);
897 if (q->properties.is_active) {
899 retval = execute_queues_cpsch(dqm, false);
902 if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
903 dqm->sdma_queue_count++;
905 * Unconditionally increment this counter, regardless of the queue's
906 * type or whether the queue is active.
908 dqm->total_queue_count++;
910 pr_debug("Total of %d queues are accountable so far\n",
911 dqm->total_queue_count);
914 mutex_unlock(&dqm->lock);
918 int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
919 unsigned int fence_value,
920 unsigned long timeout)
925 while (*fence_addr != fence_value) {
926 if (time_after(jiffies, timeout)) {
927 pr_err("kfd: qcm fence wait loop timeout expired\n");
936 static int destroy_sdma_queues(struct device_queue_manager *dqm,
937 unsigned int sdma_engine)
939 return pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_SDMA,
940 KFD_PREEMPT_TYPE_FILTER_DYNAMIC_QUEUES, 0, false,
944 static int destroy_queues_cpsch(struct device_queue_manager *dqm,
945 bool preempt_static_queues, bool lock)
948 enum kfd_preempt_type_filter preempt_type;
949 struct kfd_process_device *pdd;
956 mutex_lock(&dqm->lock);
957 if (!dqm->active_runlist)
960 pr_debug("kfd: Before destroying queues, sdma queue count is : %u\n",
961 dqm->sdma_queue_count);
963 if (dqm->sdma_queue_count > 0) {
964 destroy_sdma_queues(dqm, 0);
965 destroy_sdma_queues(dqm, 1);
968 preempt_type = preempt_static_queues ?
969 KFD_PREEMPT_TYPE_FILTER_ALL_QUEUES :
970 KFD_PREEMPT_TYPE_FILTER_DYNAMIC_QUEUES;
972 retval = pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_COMPUTE,
973 preempt_type, 0, false, 0);
977 *dqm->fence_addr = KFD_FENCE_INIT;
978 pm_send_query_status(&dqm->packets, dqm->fence_gpu_addr,
979 KFD_FENCE_COMPLETED);
980 /* should be timed out */
981 retval = amdkfd_fence_wait_timeout(dqm->fence_addr, KFD_FENCE_COMPLETED,
982 QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS);
984 pdd = kfd_get_process_device_data(dqm->dev,
985 kfd_get_process(current));
986 pdd->reset_wavefronts = true;
989 pm_release_ib(&dqm->packets);
990 dqm->active_runlist = false;
994 mutex_unlock(&dqm->lock);
998 static int execute_queues_cpsch(struct device_queue_manager *dqm, bool lock)
1005 mutex_lock(&dqm->lock);
1007 retval = destroy_queues_cpsch(dqm, false, false);
1009 pr_err("kfd: the cp might be in an unrecoverable state due to an unsuccessful queues preemption");
1013 if (dqm->queue_count <= 0 || dqm->processes_count <= 0) {
1018 if (dqm->active_runlist) {
1023 retval = pm_send_runlist(&dqm->packets, &dqm->queues);
1025 pr_err("kfd: failed to execute runlist");
1028 dqm->active_runlist = true;
1032 mutex_unlock(&dqm->lock);
1036 static int destroy_queue_cpsch(struct device_queue_manager *dqm,
1037 struct qcm_process_device *qpd,
1041 struct mqd_manager *mqd;
1042 bool preempt_all_queues;
1044 BUG_ON(!dqm || !qpd || !q);
1046 preempt_all_queues = false;
1050 /* remove queue from list to prevent rescheduling after preemption */
1051 mutex_lock(&dqm->lock);
1053 if (qpd->is_debug) {
1055 * error, currently we do not allow to destroy a queue
1056 * of a currently debugged process
1059 goto failed_try_destroy_debugged_queue;
1063 mqd = dqm->ops.get_mqd_manager(dqm,
1064 get_mqd_type_from_queue_type(q->properties.type));
1070 if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
1071 dqm->sdma_queue_count--;
1074 if (q->properties.is_active)
1077 execute_queues_cpsch(dqm, false);
1079 mqd->uninit_mqd(mqd, q->mqd, q->mqd_mem_obj);
1082 * Unconditionally decrement this counter, regardless of the queue's
1085 dqm->total_queue_count--;
1086 pr_debug("Total of %d queues are accountable so far\n",
1087 dqm->total_queue_count);
1089 mutex_unlock(&dqm->lock);
1094 failed_try_destroy_debugged_queue:
1096 mutex_unlock(&dqm->lock);
1101 * Low bits must be 0000/FFFF as required by HW, high bits must be 0 to
1102 * stay in user mode.
1104 #define APE1_FIXED_BITS_MASK 0xFFFF80000000FFFFULL
1105 /* APE1 limit is inclusive and 64K aligned. */
1106 #define APE1_LIMIT_ALIGNMENT 0xFFFF
1108 static bool set_cache_memory_policy(struct device_queue_manager *dqm,
1109 struct qcm_process_device *qpd,
1110 enum cache_policy default_policy,
1111 enum cache_policy alternate_policy,
1112 void __user *alternate_aperture_base,
1113 uint64_t alternate_aperture_size)
1117 pr_debug("kfd: In func %s\n", __func__);
1119 mutex_lock(&dqm->lock);
1121 if (alternate_aperture_size == 0) {
1122 /* base > limit disables APE1 */
1123 qpd->sh_mem_ape1_base = 1;
1124 qpd->sh_mem_ape1_limit = 0;
1127 * In FSA64, APE1_Base[63:0] = { 16{SH_MEM_APE1_BASE[31]},
1128 * SH_MEM_APE1_BASE[31:0], 0x0000 }
1129 * APE1_Limit[63:0] = { 16{SH_MEM_APE1_LIMIT[31]},
1130 * SH_MEM_APE1_LIMIT[31:0], 0xFFFF }
1131 * Verify that the base and size parameters can be
1132 * represented in this format and convert them.
1133 * Additionally restrict APE1 to user-mode addresses.
1136 uint64_t base = (uintptr_t)alternate_aperture_base;
1137 uint64_t limit = base + alternate_aperture_size - 1;
1142 if ((base & APE1_FIXED_BITS_MASK) != 0)
1145 if ((limit & APE1_FIXED_BITS_MASK) != APE1_LIMIT_ALIGNMENT)
1148 qpd->sh_mem_ape1_base = base >> 16;
1149 qpd->sh_mem_ape1_limit = limit >> 16;
1152 retval = dqm->ops_asic_specific.set_cache_memory_policy(
1157 alternate_aperture_base,
1158 alternate_aperture_size);
1160 if ((sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0))
1161 program_sh_mem_settings(dqm, qpd);
1163 pr_debug("kfd: sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n",
1164 qpd->sh_mem_config, qpd->sh_mem_ape1_base,
1165 qpd->sh_mem_ape1_limit);
1167 mutex_unlock(&dqm->lock);
1171 mutex_unlock(&dqm->lock);
1175 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
1177 struct device_queue_manager *dqm;
1181 pr_debug("kfd: loading device queue manager\n");
1183 dqm = kzalloc(sizeof(struct device_queue_manager), GFP_KERNEL);
1188 switch (sched_policy) {
1189 case KFD_SCHED_POLICY_HWS:
1190 case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION:
1191 /* initialize dqm for cp scheduling */
1192 dqm->ops.create_queue = create_queue_cpsch;
1193 dqm->ops.initialize = initialize_cpsch;
1194 dqm->ops.start = start_cpsch;
1195 dqm->ops.stop = stop_cpsch;
1196 dqm->ops.destroy_queue = destroy_queue_cpsch;
1197 dqm->ops.update_queue = update_queue;
1198 dqm->ops.get_mqd_manager = get_mqd_manager_nocpsch;
1199 dqm->ops.register_process = register_process_nocpsch;
1200 dqm->ops.unregister_process = unregister_process_nocpsch;
1201 dqm->ops.uninitialize = uninitialize_nocpsch;
1202 dqm->ops.create_kernel_queue = create_kernel_queue_cpsch;
1203 dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch;
1204 dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1206 case KFD_SCHED_POLICY_NO_HWS:
1207 /* initialize dqm for no cp scheduling */
1208 dqm->ops.start = start_nocpsch;
1209 dqm->ops.stop = stop_nocpsch;
1210 dqm->ops.create_queue = create_queue_nocpsch;
1211 dqm->ops.destroy_queue = destroy_queue_nocpsch;
1212 dqm->ops.update_queue = update_queue;
1213 dqm->ops.get_mqd_manager = get_mqd_manager_nocpsch;
1214 dqm->ops.register_process = register_process_nocpsch;
1215 dqm->ops.unregister_process = unregister_process_nocpsch;
1216 dqm->ops.initialize = initialize_nocpsch;
1217 dqm->ops.uninitialize = uninitialize_nocpsch;
1218 dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1225 switch (dev->device_info->asic_family) {
1227 device_queue_manager_init_vi(&dqm->ops_asic_specific);
1231 device_queue_manager_init_cik(&dqm->ops_asic_specific);
1235 if (dqm->ops.initialize(dqm) != 0) {
1243 void device_queue_manager_uninit(struct device_queue_manager *dqm)
1247 dqm->ops.uninitialize(dqm);