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/ratelimit.h>
26 #include <linux/printk.h>
27 #include <linux/slab.h>
28 #include <linux/list.h>
29 #include <linux/types.h>
30 #include <linux/bitops.h>
31 #include <linux/sched.h>
33 #include "kfd_device_queue_manager.h"
34 #include "kfd_mqd_manager.h"
36 #include "kfd_kernel_queue.h"
37 #include "amdgpu_amdkfd.h"
38 #include "mes_api_def.h"
40 /* Size of the per-pipe EOP queue */
41 #define CIK_HPD_EOP_BYTES_LOG2 11
42 #define CIK_HPD_EOP_BYTES (1U << CIK_HPD_EOP_BYTES_LOG2)
44 static int set_pasid_vmid_mapping(struct device_queue_manager *dqm,
45 u32 pasid, unsigned int vmid);
47 static int execute_queues_cpsch(struct device_queue_manager *dqm,
48 enum kfd_unmap_queues_filter filter,
49 uint32_t filter_param);
50 static int unmap_queues_cpsch(struct device_queue_manager *dqm,
51 enum kfd_unmap_queues_filter filter,
52 uint32_t filter_param, bool reset);
54 static int map_queues_cpsch(struct device_queue_manager *dqm);
56 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
59 static inline void deallocate_hqd(struct device_queue_manager *dqm,
61 static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q);
62 static int allocate_sdma_queue(struct device_queue_manager *dqm,
63 struct queue *q, const uint32_t *restore_sdma_id);
64 static void kfd_process_hw_exception(struct work_struct *work);
67 enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type)
69 if (type == KFD_QUEUE_TYPE_SDMA || type == KFD_QUEUE_TYPE_SDMA_XGMI)
70 return KFD_MQD_TYPE_SDMA;
71 return KFD_MQD_TYPE_CP;
74 static bool is_pipe_enabled(struct device_queue_manager *dqm, int mec, int pipe)
77 int pipe_offset = (mec * dqm->dev->shared_resources.num_pipe_per_mec
78 + pipe) * dqm->dev->shared_resources.num_queue_per_pipe;
80 /* queue is available for KFD usage if bit is 1 */
81 for (i = 0; i < dqm->dev->shared_resources.num_queue_per_pipe; ++i)
82 if (test_bit(pipe_offset + i,
83 dqm->dev->shared_resources.cp_queue_bitmap))
88 unsigned int get_cp_queues_num(struct device_queue_manager *dqm)
90 return bitmap_weight(dqm->dev->shared_resources.cp_queue_bitmap,
94 unsigned int get_queues_per_pipe(struct device_queue_manager *dqm)
96 return dqm->dev->shared_resources.num_queue_per_pipe;
99 unsigned int get_pipes_per_mec(struct device_queue_manager *dqm)
101 return dqm->dev->shared_resources.num_pipe_per_mec;
104 static unsigned int get_num_all_sdma_engines(struct device_queue_manager *dqm)
106 return kfd_get_num_sdma_engines(dqm->dev) +
107 kfd_get_num_xgmi_sdma_engines(dqm->dev);
110 unsigned int get_num_sdma_queues(struct device_queue_manager *dqm)
112 return kfd_get_num_sdma_engines(dqm->dev) *
113 dqm->dev->device_info.num_sdma_queues_per_engine;
116 unsigned int get_num_xgmi_sdma_queues(struct device_queue_manager *dqm)
118 return kfd_get_num_xgmi_sdma_engines(dqm->dev) *
119 dqm->dev->device_info.num_sdma_queues_per_engine;
122 static inline uint64_t get_reserved_sdma_queues_bitmap(struct device_queue_manager *dqm)
124 return dqm->dev->device_info.reserved_sdma_queues_bitmap;
127 void program_sh_mem_settings(struct device_queue_manager *dqm,
128 struct qcm_process_device *qpd)
130 return dqm->dev->kfd2kgd->program_sh_mem_settings(
131 dqm->dev->adev, qpd->vmid,
133 qpd->sh_mem_ape1_base,
134 qpd->sh_mem_ape1_limit,
138 static void kfd_hws_hang(struct device_queue_manager *dqm)
141 * Issue a GPU reset if HWS is unresponsive
143 dqm->is_hws_hang = true;
145 /* It's possible we're detecting a HWS hang in the
146 * middle of a GPU reset. No need to schedule another
147 * reset in this case.
149 if (!dqm->is_resetting)
150 schedule_work(&dqm->hw_exception_work);
153 static int convert_to_mes_queue_type(int queue_type)
157 switch (queue_type) {
158 case KFD_QUEUE_TYPE_COMPUTE:
159 mes_queue_type = MES_QUEUE_TYPE_COMPUTE;
161 case KFD_QUEUE_TYPE_SDMA:
162 mes_queue_type = MES_QUEUE_TYPE_SDMA;
165 WARN(1, "Invalid queue type %d", queue_type);
166 mes_queue_type = -EINVAL;
170 return mes_queue_type;
173 static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
174 struct qcm_process_device *qpd)
176 struct amdgpu_device *adev = (struct amdgpu_device *)dqm->dev->adev;
177 struct kfd_process_device *pdd = qpd_to_pdd(qpd);
178 struct mes_add_queue_input queue_input;
180 uint64_t wptr_addr_off;
182 if (dqm->is_hws_hang)
185 memset(&queue_input, 0x0, sizeof(struct mes_add_queue_input));
186 queue_input.process_id = qpd->pqm->process->pasid;
187 queue_input.page_table_base_addr = qpd->page_table_base;
188 queue_input.process_va_start = 0;
189 queue_input.process_va_end = adev->vm_manager.max_pfn - 1;
190 /* MES unit for quantum is 100ns */
191 queue_input.process_quantum = KFD_MES_PROCESS_QUANTUM; /* Equivalent to 10ms. */
192 queue_input.process_context_addr = pdd->proc_ctx_gpu_addr;
193 queue_input.gang_quantum = KFD_MES_GANG_QUANTUM; /* Equivalent to 1ms */
194 queue_input.gang_context_addr = q->gang_ctx_gpu_addr;
195 queue_input.inprocess_gang_priority = q->properties.priority;
196 queue_input.gang_global_priority_level =
197 AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
198 queue_input.doorbell_offset = q->properties.doorbell_off;
199 queue_input.mqd_addr = q->gart_mqd_addr;
200 queue_input.wptr_addr = (uint64_t)q->properties.write_ptr;
203 wptr_addr_off = (uint64_t)q->properties.write_ptr & (PAGE_SIZE - 1);
204 queue_input.wptr_mc_addr = ((uint64_t)q->wptr_bo->tbo.resource->start << PAGE_SHIFT) + wptr_addr_off;
207 queue_input.is_kfd_process = 1;
208 queue_input.is_aql_queue = (q->properties.format == KFD_QUEUE_FORMAT_AQL);
209 queue_input.queue_size = q->properties.queue_size >> 2;
211 queue_input.paging = false;
212 queue_input.tba_addr = qpd->tba_addr;
213 queue_input.tma_addr = qpd->tma_addr;
215 queue_type = convert_to_mes_queue_type(q->properties.type);
216 if (queue_type < 0) {
217 pr_err("Queue type not supported with MES, queue:%d\n",
221 queue_input.queue_type = (uint32_t)queue_type;
224 queue_input.gws_base = 0;
225 queue_input.gws_size = qpd->num_gws;
228 amdgpu_mes_lock(&adev->mes);
229 r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input);
230 amdgpu_mes_unlock(&adev->mes);
232 pr_err("failed to add hardware queue to MES, doorbell=0x%x\n",
233 q->properties.doorbell_off);
234 pr_err("MES might be in unrecoverable state, issue a GPU reset\n");
241 static int remove_queue_mes(struct device_queue_manager *dqm, struct queue *q,
242 struct qcm_process_device *qpd)
244 struct amdgpu_device *adev = (struct amdgpu_device *)dqm->dev->adev;
246 struct mes_remove_queue_input queue_input;
248 if (dqm->is_hws_hang)
251 memset(&queue_input, 0x0, sizeof(struct mes_remove_queue_input));
252 queue_input.doorbell_offset = q->properties.doorbell_off;
253 queue_input.gang_context_addr = q->gang_ctx_gpu_addr;
255 amdgpu_mes_lock(&adev->mes);
256 r = adev->mes.funcs->remove_hw_queue(&adev->mes, &queue_input);
257 amdgpu_mes_unlock(&adev->mes);
260 pr_err("failed to remove hardware queue from MES, doorbell=0x%x\n",
261 q->properties.doorbell_off);
262 pr_err("MES might be in unrecoverable state, issue a GPU reset\n");
269 static int remove_all_queues_mes(struct device_queue_manager *dqm)
271 struct device_process_node *cur;
272 struct qcm_process_device *qpd;
276 list_for_each_entry(cur, &dqm->queues, list) {
278 list_for_each_entry(q, &qpd->queues_list, list) {
279 if (q->properties.is_active) {
280 retval = remove_queue_mes(dqm, q, qpd);
282 pr_err("%s: Failed to remove queue %d for dev %d",
284 q->properties.queue_id,
295 static void increment_queue_count(struct device_queue_manager *dqm,
296 struct qcm_process_device *qpd,
299 dqm->active_queue_count++;
300 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
301 q->properties.type == KFD_QUEUE_TYPE_DIQ)
302 dqm->active_cp_queue_count++;
304 if (q->properties.is_gws) {
305 dqm->gws_queue_count++;
306 qpd->mapped_gws_queue = true;
310 static void decrement_queue_count(struct device_queue_manager *dqm,
311 struct qcm_process_device *qpd,
314 dqm->active_queue_count--;
315 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
316 q->properties.type == KFD_QUEUE_TYPE_DIQ)
317 dqm->active_cp_queue_count--;
319 if (q->properties.is_gws) {
320 dqm->gws_queue_count--;
321 qpd->mapped_gws_queue = false;
326 * Allocate a doorbell ID to this queue.
327 * If doorbell_id is passed in, make sure requested ID is valid then allocate it.
329 static int allocate_doorbell(struct qcm_process_device *qpd,
331 uint32_t const *restore_id)
333 struct kfd_dev *dev = qpd->dqm->dev;
335 if (!KFD_IS_SOC15(dev)) {
336 /* On pre-SOC15 chips we need to use the queue ID to
337 * preserve the user mode ABI.
340 if (restore_id && *restore_id != q->properties.queue_id)
343 q->doorbell_id = q->properties.queue_id;
344 } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
345 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
346 /* For SDMA queues on SOC15 with 8-byte doorbell, use static
347 * doorbell assignments based on the engine and queue id.
348 * The doobell index distance between RLC (2*i) and (2*i+1)
349 * for a SDMA engine is 512.
352 uint32_t *idx_offset = dev->shared_resources.sdma_doorbell_idx;
353 uint32_t valid_id = idx_offset[q->properties.sdma_engine_id]
354 + (q->properties.sdma_queue_id & 1)
355 * KFD_QUEUE_DOORBELL_MIRROR_OFFSET
356 + (q->properties.sdma_queue_id >> 1);
358 if (restore_id && *restore_id != valid_id)
360 q->doorbell_id = valid_id;
362 /* For CP queues on SOC15 */
364 /* make sure that ID is free */
365 if (__test_and_set_bit(*restore_id, qpd->doorbell_bitmap))
368 q->doorbell_id = *restore_id;
370 /* or reserve a free doorbell ID */
373 found = find_first_zero_bit(qpd->doorbell_bitmap,
374 KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
375 if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) {
376 pr_debug("No doorbells available");
379 set_bit(found, qpd->doorbell_bitmap);
380 q->doorbell_id = found;
384 q->properties.doorbell_off =
385 kfd_get_doorbell_dw_offset_in_bar(dev, qpd_to_pdd(qpd),
390 static void deallocate_doorbell(struct qcm_process_device *qpd,
394 struct kfd_dev *dev = qpd->dqm->dev;
396 if (!KFD_IS_SOC15(dev) ||
397 q->properties.type == KFD_QUEUE_TYPE_SDMA ||
398 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
401 old = test_and_clear_bit(q->doorbell_id, qpd->doorbell_bitmap);
405 static void program_trap_handler_settings(struct device_queue_manager *dqm,
406 struct qcm_process_device *qpd)
408 if (dqm->dev->kfd2kgd->program_trap_handler_settings)
409 dqm->dev->kfd2kgd->program_trap_handler_settings(
410 dqm->dev->adev, qpd->vmid,
411 qpd->tba_addr, qpd->tma_addr);
414 static int allocate_vmid(struct device_queue_manager *dqm,
415 struct qcm_process_device *qpd,
418 int allocated_vmid = -1, i;
420 for (i = dqm->dev->vm_info.first_vmid_kfd;
421 i <= dqm->dev->vm_info.last_vmid_kfd; i++) {
422 if (!dqm->vmid_pasid[i]) {
428 if (allocated_vmid < 0) {
429 pr_err("no more vmid to allocate\n");
433 pr_debug("vmid allocated: %d\n", allocated_vmid);
435 dqm->vmid_pasid[allocated_vmid] = q->process->pasid;
437 set_pasid_vmid_mapping(dqm, q->process->pasid, allocated_vmid);
439 qpd->vmid = allocated_vmid;
440 q->properties.vmid = allocated_vmid;
442 program_sh_mem_settings(dqm, qpd);
444 if (KFD_IS_SOC15(dqm->dev) && dqm->dev->cwsr_enabled)
445 program_trap_handler_settings(dqm, qpd);
447 /* qpd->page_table_base is set earlier when register_process()
448 * is called, i.e. when the first queue is created.
450 dqm->dev->kfd2kgd->set_vm_context_page_table_base(dqm->dev->adev,
452 qpd->page_table_base);
453 /* invalidate the VM context after pasid and vmid mapping is set up */
454 kfd_flush_tlb(qpd_to_pdd(qpd), TLB_FLUSH_LEGACY);
456 if (dqm->dev->kfd2kgd->set_scratch_backing_va)
457 dqm->dev->kfd2kgd->set_scratch_backing_va(dqm->dev->adev,
458 qpd->sh_hidden_private_base, qpd->vmid);
463 static int flush_texture_cache_nocpsch(struct kfd_dev *kdev,
464 struct qcm_process_device *qpd)
466 const struct packet_manager_funcs *pmf = qpd->dqm->packet_mgr.pmf;
472 ret = pmf->release_mem(qpd->ib_base, (uint32_t *)qpd->ib_kaddr);
476 return amdgpu_amdkfd_submit_ib(kdev->adev, KGD_ENGINE_MEC1, qpd->vmid,
477 qpd->ib_base, (uint32_t *)qpd->ib_kaddr,
478 pmf->release_mem_size / sizeof(uint32_t));
481 static void deallocate_vmid(struct device_queue_manager *dqm,
482 struct qcm_process_device *qpd,
485 /* On GFX v7, CP doesn't flush TC at dequeue */
486 if (q->device->adev->asic_type == CHIP_HAWAII)
487 if (flush_texture_cache_nocpsch(q->device, qpd))
488 pr_err("Failed to flush TC\n");
490 kfd_flush_tlb(qpd_to_pdd(qpd), TLB_FLUSH_LEGACY);
492 /* Release the vmid mapping */
493 set_pasid_vmid_mapping(dqm, 0, qpd->vmid);
494 dqm->vmid_pasid[qpd->vmid] = 0;
497 q->properties.vmid = 0;
500 static int create_queue_nocpsch(struct device_queue_manager *dqm,
502 struct qcm_process_device *qpd,
503 const struct kfd_criu_queue_priv_data *qd,
504 const void *restore_mqd, const void *restore_ctl_stack)
506 struct mqd_manager *mqd_mgr;
511 if (dqm->total_queue_count >= max_num_of_queues_per_device) {
512 pr_warn("Can't create new usermode queue because %d queues were already created\n",
513 dqm->total_queue_count);
518 if (list_empty(&qpd->queues_list)) {
519 retval = allocate_vmid(dqm, qpd, q);
523 q->properties.vmid = qpd->vmid;
525 * Eviction state logic: mark all queues as evicted, even ones
526 * not currently active. Restoring inactive queues later only
527 * updates the is_evicted flag but is a no-op otherwise.
529 q->properties.is_evicted = !!qpd->evicted;
531 q->properties.tba_addr = qpd->tba_addr;
532 q->properties.tma_addr = qpd->tma_addr;
534 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
535 q->properties.type)];
536 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) {
537 retval = allocate_hqd(dqm, q);
539 goto deallocate_vmid;
540 pr_debug("Loading mqd to hqd on pipe %d, queue %d\n",
542 } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
543 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
544 retval = allocate_sdma_queue(dqm, q, qd ? &qd->sdma_id : NULL);
546 goto deallocate_vmid;
547 dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
550 retval = allocate_doorbell(qpd, q, qd ? &qd->doorbell_id : NULL);
552 goto out_deallocate_hqd;
554 /* Temporarily release dqm lock to avoid a circular lock dependency */
556 q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties);
559 if (!q->mqd_mem_obj) {
561 goto out_deallocate_doorbell;
565 mqd_mgr->restore_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, &q->gart_mqd_addr,
566 &q->properties, restore_mqd, restore_ctl_stack,
569 mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj,
570 &q->gart_mqd_addr, &q->properties);
572 if (q->properties.is_active) {
573 if (!dqm->sched_running) {
574 WARN_ONCE(1, "Load non-HWS mqd while stopped\n");
575 goto add_queue_to_list;
578 if (WARN(q->process->mm != current->mm,
579 "should only run in user thread"))
582 retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe,
583 q->queue, &q->properties, current->mm);
589 list_add(&q->list, &qpd->queues_list);
591 if (q->properties.is_active)
592 increment_queue_count(dqm, qpd, q);
595 * Unconditionally increment this counter, regardless of the queue's
596 * type or whether the queue is active.
598 dqm->total_queue_count++;
599 pr_debug("Total of %d queues are accountable so far\n",
600 dqm->total_queue_count);
604 mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
605 out_deallocate_doorbell:
606 deallocate_doorbell(qpd, q);
608 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
609 deallocate_hqd(dqm, q);
610 else if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
611 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
612 deallocate_sdma_queue(dqm, q);
614 if (list_empty(&qpd->queues_list))
615 deallocate_vmid(dqm, qpd, q);
621 static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q)
628 for (pipe = dqm->next_pipe_to_allocate, i = 0;
629 i < get_pipes_per_mec(dqm);
630 pipe = ((pipe + 1) % get_pipes_per_mec(dqm)), ++i) {
632 if (!is_pipe_enabled(dqm, 0, pipe))
635 if (dqm->allocated_queues[pipe] != 0) {
636 bit = ffs(dqm->allocated_queues[pipe]) - 1;
637 dqm->allocated_queues[pipe] &= ~(1 << bit);
648 pr_debug("hqd slot - pipe %d, queue %d\n", q->pipe, q->queue);
649 /* horizontal hqd allocation */
650 dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_per_mec(dqm);
655 static inline void deallocate_hqd(struct device_queue_manager *dqm,
658 dqm->allocated_queues[q->pipe] |= (1 << q->queue);
661 #define SQ_IND_CMD_CMD_KILL 0x00000003
662 #define SQ_IND_CMD_MODE_BROADCAST 0x00000001
664 static int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p)
668 uint16_t queried_pasid;
669 union SQ_CMD_BITS reg_sq_cmd;
670 union GRBM_GFX_INDEX_BITS reg_gfx_index;
671 struct kfd_process_device *pdd;
672 int first_vmid_to_scan = dev->vm_info.first_vmid_kfd;
673 int last_vmid_to_scan = dev->vm_info.last_vmid_kfd;
675 reg_sq_cmd.u32All = 0;
676 reg_gfx_index.u32All = 0;
678 pr_debug("Killing all process wavefronts\n");
680 if (!dev->kfd2kgd->get_atc_vmid_pasid_mapping_info) {
681 pr_err("no vmid pasid mapping supported \n");
685 /* Scan all registers in the range ATC_VMID8_PASID_MAPPING ..
686 * ATC_VMID15_PASID_MAPPING
687 * to check which VMID the current process is mapped to.
690 for (vmid = first_vmid_to_scan; vmid <= last_vmid_to_scan; vmid++) {
691 status = dev->kfd2kgd->get_atc_vmid_pasid_mapping_info
692 (dev->adev, vmid, &queried_pasid);
694 if (status && queried_pasid == p->pasid) {
695 pr_debug("Killing wave fronts of vmid %d and pasid 0x%x\n",
701 if (vmid > last_vmid_to_scan) {
702 pr_err("Didn't find vmid for pasid 0x%x\n", p->pasid);
706 /* taking the VMID for that process on the safe way using PDD */
707 pdd = kfd_get_process_device_data(dev, p);
711 reg_gfx_index.bits.sh_broadcast_writes = 1;
712 reg_gfx_index.bits.se_broadcast_writes = 1;
713 reg_gfx_index.bits.instance_broadcast_writes = 1;
714 reg_sq_cmd.bits.mode = SQ_IND_CMD_MODE_BROADCAST;
715 reg_sq_cmd.bits.cmd = SQ_IND_CMD_CMD_KILL;
716 reg_sq_cmd.bits.vm_id = vmid;
718 dev->kfd2kgd->wave_control_execute(dev->adev,
719 reg_gfx_index.u32All,
725 /* Access to DQM has to be locked before calling destroy_queue_nocpsch_locked
726 * to avoid asynchronized access
728 static int destroy_queue_nocpsch_locked(struct device_queue_manager *dqm,
729 struct qcm_process_device *qpd,
733 struct mqd_manager *mqd_mgr;
735 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
736 q->properties.type)];
738 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
739 deallocate_hqd(dqm, q);
740 else if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
741 deallocate_sdma_queue(dqm, q);
742 else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
743 deallocate_sdma_queue(dqm, q);
745 pr_debug("q->properties.type %d is invalid\n",
749 dqm->total_queue_count--;
751 deallocate_doorbell(qpd, q);
753 if (!dqm->sched_running) {
754 WARN_ONCE(1, "Destroy non-HWS queue while stopped\n");
758 retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
759 KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
760 KFD_UNMAP_LATENCY_MS,
762 if (retval == -ETIME)
763 qpd->reset_wavefronts = true;
766 if (list_empty(&qpd->queues_list)) {
767 if (qpd->reset_wavefronts) {
768 pr_warn("Resetting wave fronts (nocpsch) on dev %p\n",
770 /* dbgdev_wave_reset_wavefronts has to be called before
771 * deallocate_vmid(), i.e. when vmid is still in use.
773 dbgdev_wave_reset_wavefronts(dqm->dev,
775 qpd->reset_wavefronts = false;
778 deallocate_vmid(dqm, qpd, q);
781 if (q->properties.is_active)
782 decrement_queue_count(dqm, qpd, q);
787 static int destroy_queue_nocpsch(struct device_queue_manager *dqm,
788 struct qcm_process_device *qpd,
792 uint64_t sdma_val = 0;
793 struct kfd_process_device *pdd = qpd_to_pdd(qpd);
794 struct mqd_manager *mqd_mgr =
795 dqm->mqd_mgrs[get_mqd_type_from_queue_type(q->properties.type)];
797 /* Get the SDMA queue stats */
798 if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) ||
799 (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
800 retval = read_sdma_queue_counter((uint64_t __user *)q->properties.read_ptr,
803 pr_err("Failed to read SDMA queue counter for queue: %d\n",
804 q->properties.queue_id);
808 retval = destroy_queue_nocpsch_locked(dqm, qpd, q);
810 pdd->sdma_past_activity_counter += sdma_val;
813 mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
818 static int update_queue(struct device_queue_manager *dqm, struct queue *q,
819 struct mqd_update_info *minfo)
822 struct mqd_manager *mqd_mgr;
823 struct kfd_process_device *pdd;
824 bool prev_active = false;
827 pdd = kfd_get_process_device_data(q->device, q->process);
832 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
833 q->properties.type)];
835 /* Save previous activity state for counters */
836 prev_active = q->properties.is_active;
838 /* Make sure the queue is unmapped before updating the MQD */
839 if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) {
840 if (!dqm->dev->shared_resources.enable_mes)
841 retval = unmap_queues_cpsch(dqm,
842 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, false);
843 else if (prev_active)
844 retval = remove_queue_mes(dqm, q, &pdd->qpd);
847 pr_err("unmap queue failed\n");
850 } else if (prev_active &&
851 (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
852 q->properties.type == KFD_QUEUE_TYPE_SDMA ||
853 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
855 if (!dqm->sched_running) {
856 WARN_ONCE(1, "Update non-HWS queue while stopped\n");
860 retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
861 (dqm->dev->cwsr_enabled ?
862 KFD_PREEMPT_TYPE_WAVEFRONT_SAVE :
863 KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN),
864 KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
866 pr_err("destroy mqd failed\n");
871 mqd_mgr->update_mqd(mqd_mgr, q->mqd, &q->properties, minfo);
874 * check active state vs. the previous state and modify
875 * counter accordingly. map_queues_cpsch uses the
876 * dqm->active_queue_count to determine whether a new runlist must be
879 if (q->properties.is_active && !prev_active) {
880 increment_queue_count(dqm, &pdd->qpd, q);
881 } else if (!q->properties.is_active && prev_active) {
882 decrement_queue_count(dqm, &pdd->qpd, q);
883 } else if (q->gws && !q->properties.is_gws) {
884 if (q->properties.is_active) {
885 dqm->gws_queue_count++;
886 pdd->qpd.mapped_gws_queue = true;
888 q->properties.is_gws = true;
889 } else if (!q->gws && q->properties.is_gws) {
890 if (q->properties.is_active) {
891 dqm->gws_queue_count--;
892 pdd->qpd.mapped_gws_queue = false;
894 q->properties.is_gws = false;
897 if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) {
898 if (!dqm->dev->shared_resources.enable_mes)
899 retval = map_queues_cpsch(dqm);
900 else if (q->properties.is_active)
901 retval = add_queue_mes(dqm, q, &pdd->qpd);
902 } else if (q->properties.is_active &&
903 (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
904 q->properties.type == KFD_QUEUE_TYPE_SDMA ||
905 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
906 if (WARN(q->process->mm != current->mm,
907 "should only run in user thread"))
910 retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd,
912 &q->properties, current->mm);
920 static int evict_process_queues_nocpsch(struct device_queue_manager *dqm,
921 struct qcm_process_device *qpd)
924 struct mqd_manager *mqd_mgr;
925 struct kfd_process_device *pdd;
929 if (qpd->evicted++ > 0) /* already evicted, do nothing */
932 pdd = qpd_to_pdd(qpd);
933 pr_debug_ratelimited("Evicting PASID 0x%x queues\n",
934 pdd->process->pasid);
936 pdd->last_evict_timestamp = get_jiffies_64();
937 /* Mark all queues as evicted. Deactivate all active queues on
940 list_for_each_entry(q, &qpd->queues_list, list) {
941 q->properties.is_evicted = true;
942 if (!q->properties.is_active)
945 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
946 q->properties.type)];
947 q->properties.is_active = false;
948 decrement_queue_count(dqm, qpd, q);
950 if (WARN_ONCE(!dqm->sched_running, "Evict when stopped\n"))
953 retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
954 (dqm->dev->cwsr_enabled ?
955 KFD_PREEMPT_TYPE_WAVEFRONT_SAVE :
956 KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN),
957 KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
959 /* Return the first error, but keep going to
960 * maintain a consistent eviction state
970 static int evict_process_queues_cpsch(struct device_queue_manager *dqm,
971 struct qcm_process_device *qpd)
974 struct kfd_process_device *pdd;
978 if (qpd->evicted++ > 0) /* already evicted, do nothing */
981 pdd = qpd_to_pdd(qpd);
982 pr_debug_ratelimited("Evicting PASID 0x%x queues\n",
983 pdd->process->pasid);
985 /* Mark all queues as evicted. Deactivate all active queues on
988 list_for_each_entry(q, &qpd->queues_list, list) {
989 q->properties.is_evicted = true;
990 if (!q->properties.is_active)
993 q->properties.is_active = false;
994 decrement_queue_count(dqm, qpd, q);
996 if (dqm->dev->shared_resources.enable_mes) {
997 retval = remove_queue_mes(dqm, q, qpd);
999 pr_err("Failed to evict queue %d\n",
1000 q->properties.queue_id);
1005 pdd->last_evict_timestamp = get_jiffies_64();
1006 if (!dqm->dev->shared_resources.enable_mes)
1007 retval = execute_queues_cpsch(dqm,
1009 KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES :
1010 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1017 static int restore_process_queues_nocpsch(struct device_queue_manager *dqm,
1018 struct qcm_process_device *qpd)
1020 struct mm_struct *mm = NULL;
1022 struct mqd_manager *mqd_mgr;
1023 struct kfd_process_device *pdd;
1025 uint64_t eviction_duration;
1026 int retval, ret = 0;
1028 pdd = qpd_to_pdd(qpd);
1029 /* Retrieve PD base */
1030 pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv);
1033 if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */
1035 if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
1040 pr_debug_ratelimited("Restoring PASID 0x%x queues\n",
1041 pdd->process->pasid);
1043 /* Update PD Base in QPD */
1044 qpd->page_table_base = pd_base;
1045 pr_debug("Updated PD address to 0x%llx\n", pd_base);
1047 if (!list_empty(&qpd->queues_list)) {
1048 dqm->dev->kfd2kgd->set_vm_context_page_table_base(
1051 qpd->page_table_base);
1052 kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY);
1055 /* Take a safe reference to the mm_struct, which may otherwise
1056 * disappear even while the kfd_process is still referenced.
1058 mm = get_task_mm(pdd->process->lead_thread);
1064 /* Remove the eviction flags. Activate queues that are not
1065 * inactive for other reasons.
1067 list_for_each_entry(q, &qpd->queues_list, list) {
1068 q->properties.is_evicted = false;
1069 if (!QUEUE_IS_ACTIVE(q->properties))
1072 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1073 q->properties.type)];
1074 q->properties.is_active = true;
1075 increment_queue_count(dqm, qpd, q);
1077 if (WARN_ONCE(!dqm->sched_running, "Restore when stopped\n"))
1080 retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe,
1081 q->queue, &q->properties, mm);
1083 /* Return the first error, but keep going to
1084 * maintain a consistent eviction state
1089 eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp;
1090 atomic64_add(eviction_duration, &pdd->evict_duration_counter);
1098 static int restore_process_queues_cpsch(struct device_queue_manager *dqm,
1099 struct qcm_process_device *qpd)
1102 struct kfd_process_device *pdd;
1104 uint64_t eviction_duration;
1107 pdd = qpd_to_pdd(qpd);
1108 /* Retrieve PD base */
1109 pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv);
1112 if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */
1114 if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
1119 pr_debug_ratelimited("Restoring PASID 0x%x queues\n",
1120 pdd->process->pasid);
1122 /* Update PD Base in QPD */
1123 qpd->page_table_base = pd_base;
1124 pr_debug("Updated PD address to 0x%llx\n", pd_base);
1126 /* activate all active queues on the qpd */
1127 list_for_each_entry(q, &qpd->queues_list, list) {
1128 q->properties.is_evicted = false;
1129 if (!QUEUE_IS_ACTIVE(q->properties))
1132 q->properties.is_active = true;
1133 increment_queue_count(dqm, &pdd->qpd, q);
1135 if (dqm->dev->shared_resources.enable_mes) {
1136 retval = add_queue_mes(dqm, q, qpd);
1138 pr_err("Failed to restore queue %d\n",
1139 q->properties.queue_id);
1144 if (!dqm->dev->shared_resources.enable_mes)
1145 retval = execute_queues_cpsch(dqm,
1146 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1148 eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp;
1149 atomic64_add(eviction_duration, &pdd->evict_duration_counter);
1155 static int register_process(struct device_queue_manager *dqm,
1156 struct qcm_process_device *qpd)
1158 struct device_process_node *n;
1159 struct kfd_process_device *pdd;
1163 n = kzalloc(sizeof(*n), GFP_KERNEL);
1169 pdd = qpd_to_pdd(qpd);
1170 /* Retrieve PD base */
1171 pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv);
1174 list_add(&n->list, &dqm->queues);
1176 /* Update PD Base in QPD */
1177 qpd->page_table_base = pd_base;
1178 pr_debug("Updated PD address to 0x%llx\n", pd_base);
1180 retval = dqm->asic_ops.update_qpd(dqm, qpd);
1182 dqm->processes_count++;
1186 /* Outside the DQM lock because under the DQM lock we can't do
1187 * reclaim or take other locks that others hold while reclaiming.
1189 kfd_inc_compute_active(dqm->dev);
1194 static int unregister_process(struct device_queue_manager *dqm,
1195 struct qcm_process_device *qpd)
1198 struct device_process_node *cur, *next;
1200 pr_debug("qpd->queues_list is %s\n",
1201 list_empty(&qpd->queues_list) ? "empty" : "not empty");
1206 list_for_each_entry_safe(cur, next, &dqm->queues, list) {
1207 if (qpd == cur->qpd) {
1208 list_del(&cur->list);
1210 dqm->processes_count--;
1214 /* qpd not found in dqm list */
1219 /* Outside the DQM lock because under the DQM lock we can't do
1220 * reclaim or take other locks that others hold while reclaiming.
1223 kfd_dec_compute_active(dqm->dev);
1229 set_pasid_vmid_mapping(struct device_queue_manager *dqm, u32 pasid,
1232 return dqm->dev->kfd2kgd->set_pasid_vmid_mapping(
1233 dqm->dev->adev, pasid, vmid);
1236 static void init_interrupts(struct device_queue_manager *dqm)
1240 for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++)
1241 if (is_pipe_enabled(dqm, 0, i))
1242 dqm->dev->kfd2kgd->init_interrupts(dqm->dev->adev, i);
1245 static void init_sdma_bitmaps(struct device_queue_manager *dqm)
1247 unsigned int num_sdma_queues =
1248 min_t(unsigned int, sizeof(dqm->sdma_bitmap)*8,
1249 get_num_sdma_queues(dqm));
1250 unsigned int num_xgmi_sdma_queues =
1251 min_t(unsigned int, sizeof(dqm->xgmi_sdma_bitmap)*8,
1252 get_num_xgmi_sdma_queues(dqm));
1254 if (num_sdma_queues)
1255 dqm->sdma_bitmap = GENMASK_ULL(num_sdma_queues-1, 0);
1256 if (num_xgmi_sdma_queues)
1257 dqm->xgmi_sdma_bitmap = GENMASK_ULL(num_xgmi_sdma_queues-1, 0);
1259 dqm->sdma_bitmap &= ~get_reserved_sdma_queues_bitmap(dqm);
1260 pr_info("sdma_bitmap: %llx\n", dqm->sdma_bitmap);
1263 static int initialize_nocpsch(struct device_queue_manager *dqm)
1267 pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm));
1269 dqm->allocated_queues = kcalloc(get_pipes_per_mec(dqm),
1270 sizeof(unsigned int), GFP_KERNEL);
1271 if (!dqm->allocated_queues)
1274 mutex_init(&dqm->lock_hidden);
1275 INIT_LIST_HEAD(&dqm->queues);
1276 dqm->active_queue_count = dqm->next_pipe_to_allocate = 0;
1277 dqm->active_cp_queue_count = 0;
1278 dqm->gws_queue_count = 0;
1280 for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) {
1281 int pipe_offset = pipe * get_queues_per_pipe(dqm);
1283 for (queue = 0; queue < get_queues_per_pipe(dqm); queue++)
1284 if (test_bit(pipe_offset + queue,
1285 dqm->dev->shared_resources.cp_queue_bitmap))
1286 dqm->allocated_queues[pipe] |= 1 << queue;
1289 memset(dqm->vmid_pasid, 0, sizeof(dqm->vmid_pasid));
1291 init_sdma_bitmaps(dqm);
1296 static void uninitialize(struct device_queue_manager *dqm)
1300 WARN_ON(dqm->active_queue_count > 0 || dqm->processes_count > 0);
1302 kfree(dqm->allocated_queues);
1303 for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++)
1304 kfree(dqm->mqd_mgrs[i]);
1305 mutex_destroy(&dqm->lock_hidden);
1308 static int start_nocpsch(struct device_queue_manager *dqm)
1312 pr_info("SW scheduler is used");
1313 init_interrupts(dqm);
1315 if (dqm->dev->adev->asic_type == CHIP_HAWAII)
1316 r = pm_init(&dqm->packet_mgr, dqm);
1318 dqm->sched_running = true;
1323 static int stop_nocpsch(struct device_queue_manager *dqm)
1325 if (dqm->dev->adev->asic_type == CHIP_HAWAII)
1326 pm_uninit(&dqm->packet_mgr, false);
1327 dqm->sched_running = false;
1332 static void pre_reset(struct device_queue_manager *dqm)
1335 dqm->is_resetting = true;
1339 static int allocate_sdma_queue(struct device_queue_manager *dqm,
1340 struct queue *q, const uint32_t *restore_sdma_id)
1344 if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
1345 if (dqm->sdma_bitmap == 0) {
1346 pr_err("No more SDMA queue to allocate\n");
1350 if (restore_sdma_id) {
1351 /* Re-use existing sdma_id */
1352 if (!(dqm->sdma_bitmap & (1ULL << *restore_sdma_id))) {
1353 pr_err("SDMA queue already in use\n");
1356 dqm->sdma_bitmap &= ~(1ULL << *restore_sdma_id);
1357 q->sdma_id = *restore_sdma_id;
1359 /* Find first available sdma_id */
1360 bit = __ffs64(dqm->sdma_bitmap);
1361 dqm->sdma_bitmap &= ~(1ULL << bit);
1365 q->properties.sdma_engine_id = q->sdma_id %
1366 kfd_get_num_sdma_engines(dqm->dev);
1367 q->properties.sdma_queue_id = q->sdma_id /
1368 kfd_get_num_sdma_engines(dqm->dev);
1369 } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1370 if (dqm->xgmi_sdma_bitmap == 0) {
1371 pr_err("No more XGMI SDMA queue to allocate\n");
1374 if (restore_sdma_id) {
1375 /* Re-use existing sdma_id */
1376 if (!(dqm->xgmi_sdma_bitmap & (1ULL << *restore_sdma_id))) {
1377 pr_err("SDMA queue already in use\n");
1380 dqm->xgmi_sdma_bitmap &= ~(1ULL << *restore_sdma_id);
1381 q->sdma_id = *restore_sdma_id;
1383 bit = __ffs64(dqm->xgmi_sdma_bitmap);
1384 dqm->xgmi_sdma_bitmap &= ~(1ULL << bit);
1387 /* sdma_engine_id is sdma id including
1388 * both PCIe-optimized SDMAs and XGMI-
1389 * optimized SDMAs. The calculation below
1390 * assumes the first N engines are always
1391 * PCIe-optimized ones
1393 q->properties.sdma_engine_id =
1394 kfd_get_num_sdma_engines(dqm->dev) +
1395 q->sdma_id % kfd_get_num_xgmi_sdma_engines(dqm->dev);
1396 q->properties.sdma_queue_id = q->sdma_id /
1397 kfd_get_num_xgmi_sdma_engines(dqm->dev);
1400 pr_debug("SDMA engine id: %d\n", q->properties.sdma_engine_id);
1401 pr_debug("SDMA queue id: %d\n", q->properties.sdma_queue_id);
1406 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
1409 if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
1410 if (q->sdma_id >= get_num_sdma_queues(dqm))
1412 dqm->sdma_bitmap |= (1ULL << q->sdma_id);
1413 } else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1414 if (q->sdma_id >= get_num_xgmi_sdma_queues(dqm))
1416 dqm->xgmi_sdma_bitmap |= (1ULL << q->sdma_id);
1421 * Device Queue Manager implementation for cp scheduler
1424 static int set_sched_resources(struct device_queue_manager *dqm)
1427 struct scheduling_resources res;
1429 res.vmid_mask = dqm->dev->shared_resources.compute_vmid_bitmap;
1432 for (i = 0; i < KGD_MAX_QUEUES; ++i) {
1433 mec = (i / dqm->dev->shared_resources.num_queue_per_pipe)
1434 / dqm->dev->shared_resources.num_pipe_per_mec;
1436 if (!test_bit(i, dqm->dev->shared_resources.cp_queue_bitmap))
1439 /* only acquire queues from the first MEC */
1443 /* This situation may be hit in the future if a new HW
1444 * generation exposes more than 64 queues. If so, the
1445 * definition of res.queue_mask needs updating
1447 if (WARN_ON(i >= (sizeof(res.queue_mask)*8))) {
1448 pr_err("Invalid queue enabled by amdgpu: %d\n", i);
1452 res.queue_mask |= 1ull
1453 << amdgpu_queue_mask_bit_to_set_resource_bit(
1456 res.gws_mask = ~0ull;
1457 res.oac_mask = res.gds_heap_base = res.gds_heap_size = 0;
1459 pr_debug("Scheduling resources:\n"
1460 "vmid mask: 0x%8X\n"
1461 "queue mask: 0x%8llX\n",
1462 res.vmid_mask, res.queue_mask);
1464 return pm_send_set_resources(&dqm->packet_mgr, &res);
1467 static int initialize_cpsch(struct device_queue_manager *dqm)
1469 pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm));
1471 mutex_init(&dqm->lock_hidden);
1472 INIT_LIST_HEAD(&dqm->queues);
1473 dqm->active_queue_count = dqm->processes_count = 0;
1474 dqm->active_cp_queue_count = 0;
1475 dqm->gws_queue_count = 0;
1476 dqm->active_runlist = false;
1477 INIT_WORK(&dqm->hw_exception_work, kfd_process_hw_exception);
1479 init_sdma_bitmaps(dqm);
1484 static int start_cpsch(struct device_queue_manager *dqm)
1492 if (!dqm->dev->shared_resources.enable_mes) {
1493 retval = pm_init(&dqm->packet_mgr, dqm);
1495 goto fail_packet_manager_init;
1497 retval = set_sched_resources(dqm);
1499 goto fail_set_sched_resources;
1501 pr_debug("Allocating fence memory\n");
1503 /* allocate fence memory on the gart */
1504 retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr),
1508 goto fail_allocate_vidmem;
1510 dqm->fence_addr = (uint64_t *)dqm->fence_mem->cpu_ptr;
1511 dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr;
1513 init_interrupts(dqm);
1515 /* clear hang status when driver try to start the hw scheduler */
1516 dqm->is_hws_hang = false;
1517 dqm->is_resetting = false;
1518 dqm->sched_running = true;
1519 if (!dqm->dev->shared_resources.enable_mes)
1520 execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1524 fail_allocate_vidmem:
1525 fail_set_sched_resources:
1526 if (!dqm->dev->shared_resources.enable_mes)
1527 pm_uninit(&dqm->packet_mgr, false);
1528 fail_packet_manager_init:
1533 static int stop_cpsch(struct device_queue_manager *dqm)
1538 if (!dqm->sched_running) {
1543 if (!dqm->is_hws_hang) {
1544 if (!dqm->dev->shared_resources.enable_mes)
1545 unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, false);
1547 remove_all_queues_mes(dqm);
1550 hanging = dqm->is_hws_hang || dqm->is_resetting;
1551 dqm->sched_running = false;
1553 if (!dqm->dev->shared_resources.enable_mes)
1554 pm_release_ib(&dqm->packet_mgr);
1556 kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
1557 if (!dqm->dev->shared_resources.enable_mes)
1558 pm_uninit(&dqm->packet_mgr, hanging);
1564 static int create_kernel_queue_cpsch(struct device_queue_manager *dqm,
1565 struct kernel_queue *kq,
1566 struct qcm_process_device *qpd)
1569 if (dqm->total_queue_count >= max_num_of_queues_per_device) {
1570 pr_warn("Can't create new kernel queue because %d queues were already created\n",
1571 dqm->total_queue_count);
1577 * Unconditionally increment this counter, regardless of the queue's
1578 * type or whether the queue is active.
1580 dqm->total_queue_count++;
1581 pr_debug("Total of %d queues are accountable so far\n",
1582 dqm->total_queue_count);
1584 list_add(&kq->list, &qpd->priv_queue_list);
1585 increment_queue_count(dqm, qpd, kq->queue);
1586 qpd->is_debug = true;
1587 execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1593 static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm,
1594 struct kernel_queue *kq,
1595 struct qcm_process_device *qpd)
1598 list_del(&kq->list);
1599 decrement_queue_count(dqm, qpd, kq->queue);
1600 qpd->is_debug = false;
1601 execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
1603 * Unconditionally decrement this counter, regardless of the queue's
1606 dqm->total_queue_count--;
1607 pr_debug("Total of %d queues are accountable so far\n",
1608 dqm->total_queue_count);
1612 static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
1613 struct qcm_process_device *qpd,
1614 const struct kfd_criu_queue_priv_data *qd,
1615 const void *restore_mqd, const void *restore_ctl_stack)
1618 struct mqd_manager *mqd_mgr;
1620 if (dqm->total_queue_count >= max_num_of_queues_per_device) {
1621 pr_warn("Can't create new usermode queue because %d queues were already created\n",
1622 dqm->total_queue_count);
1627 if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1628 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1630 retval = allocate_sdma_queue(dqm, q, qd ? &qd->sdma_id : NULL);
1636 retval = allocate_doorbell(qpd, q, qd ? &qd->doorbell_id : NULL);
1638 goto out_deallocate_sdma_queue;
1640 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1641 q->properties.type)];
1643 if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1644 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
1645 dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
1646 q->properties.tba_addr = qpd->tba_addr;
1647 q->properties.tma_addr = qpd->tma_addr;
1648 q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties);
1649 if (!q->mqd_mem_obj) {
1651 goto out_deallocate_doorbell;
1656 * Eviction state logic: mark all queues as evicted, even ones
1657 * not currently active. Restoring inactive queues later only
1658 * updates the is_evicted flag but is a no-op otherwise.
1660 q->properties.is_evicted = !!qpd->evicted;
1663 mqd_mgr->restore_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, &q->gart_mqd_addr,
1664 &q->properties, restore_mqd, restore_ctl_stack,
1665 qd->ctl_stack_size);
1667 mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj,
1668 &q->gart_mqd_addr, &q->properties);
1670 list_add(&q->list, &qpd->queues_list);
1673 if (q->properties.is_active) {
1674 increment_queue_count(dqm, qpd, q);
1676 if (!dqm->dev->shared_resources.enable_mes)
1677 retval = execute_queues_cpsch(dqm,
1678 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1680 retval = add_queue_mes(dqm, q, qpd);
1686 * Unconditionally increment this counter, regardless of the queue's
1687 * type or whether the queue is active.
1689 dqm->total_queue_count++;
1691 pr_debug("Total of %d queues are accountable so far\n",
1692 dqm->total_queue_count);
1700 if (q->properties.is_active)
1701 decrement_queue_count(dqm, qpd, q);
1702 mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
1704 out_deallocate_doorbell:
1705 deallocate_doorbell(qpd, q);
1706 out_deallocate_sdma_queue:
1707 if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1708 q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1710 deallocate_sdma_queue(dqm, q);
1717 int amdkfd_fence_wait_timeout(uint64_t *fence_addr,
1718 uint64_t fence_value,
1719 unsigned int timeout_ms)
1721 unsigned long end_jiffies = msecs_to_jiffies(timeout_ms) + jiffies;
1723 while (*fence_addr != fence_value) {
1724 if (time_after(jiffies, end_jiffies)) {
1725 pr_err("qcm fence wait loop timeout expired\n");
1726 /* In HWS case, this is used to halt the driver thread
1727 * in order not to mess up CP states before doing
1728 * scandumps for FW debugging.
1730 while (halt_if_hws_hang)
1741 /* dqm->lock mutex has to be locked before calling this function */
1742 static int map_queues_cpsch(struct device_queue_manager *dqm)
1746 if (!dqm->sched_running)
1748 if (dqm->active_queue_count <= 0 || dqm->processes_count <= 0)
1750 if (dqm->active_runlist)
1753 retval = pm_send_runlist(&dqm->packet_mgr, &dqm->queues);
1754 pr_debug("%s sent runlist\n", __func__);
1756 pr_err("failed to execute runlist\n");
1759 dqm->active_runlist = true;
1764 /* dqm->lock mutex has to be locked before calling this function */
1765 static int unmap_queues_cpsch(struct device_queue_manager *dqm,
1766 enum kfd_unmap_queues_filter filter,
1767 uint32_t filter_param, bool reset)
1770 struct mqd_manager *mqd_mgr;
1772 if (!dqm->sched_running)
1774 if (dqm->is_hws_hang || dqm->is_resetting)
1776 if (!dqm->active_runlist)
1779 retval = pm_send_unmap_queue(&dqm->packet_mgr, filter, filter_param, reset);
1783 *dqm->fence_addr = KFD_FENCE_INIT;
1784 pm_send_query_status(&dqm->packet_mgr, dqm->fence_gpu_addr,
1785 KFD_FENCE_COMPLETED);
1786 /* should be timed out */
1787 retval = amdkfd_fence_wait_timeout(dqm->fence_addr, KFD_FENCE_COMPLETED,
1788 queue_preemption_timeout_ms);
1790 pr_err("The cp might be in an unrecoverable state due to an unsuccessful queues preemption\n");
1795 /* In the current MEC firmware implementation, if compute queue
1796 * doesn't response to the preemption request in time, HIQ will
1797 * abandon the unmap request without returning any timeout error
1798 * to driver. Instead, MEC firmware will log the doorbell of the
1799 * unresponding compute queue to HIQ.MQD.queue_doorbell_id fields.
1800 * To make sure the queue unmap was successful, driver need to
1801 * check those fields
1803 mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ];
1804 if (mqd_mgr->read_doorbell_id(dqm->packet_mgr.priv_queue->queue->mqd)) {
1805 pr_err("HIQ MQD's queue_doorbell_id0 is not 0, Queue preemption time out\n");
1806 while (halt_if_hws_hang)
1811 pm_release_ib(&dqm->packet_mgr);
1812 dqm->active_runlist = false;
1817 /* only for compute queue */
1818 static int reset_queues_cpsch(struct device_queue_manager *dqm,
1825 retval = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_BY_PASID,
1832 /* dqm->lock mutex has to be locked before calling this function */
1833 static int execute_queues_cpsch(struct device_queue_manager *dqm,
1834 enum kfd_unmap_queues_filter filter,
1835 uint32_t filter_param)
1839 if (dqm->is_hws_hang)
1841 retval = unmap_queues_cpsch(dqm, filter, filter_param, false);
1845 return map_queues_cpsch(dqm);
1848 static int destroy_queue_cpsch(struct device_queue_manager *dqm,
1849 struct qcm_process_device *qpd,
1853 struct mqd_manager *mqd_mgr;
1854 uint64_t sdma_val = 0;
1855 struct kfd_process_device *pdd = qpd_to_pdd(qpd);
1857 /* Get the SDMA queue stats */
1858 if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) ||
1859 (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
1860 retval = read_sdma_queue_counter((uint64_t __user *)q->properties.read_ptr,
1863 pr_err("Failed to read SDMA queue counter for queue: %d\n",
1864 q->properties.queue_id);
1869 /* remove queue from list to prevent rescheduling after preemption */
1872 if (qpd->is_debug) {
1874 * error, currently we do not allow to destroy a queue
1875 * of a currently debugged process
1878 goto failed_try_destroy_debugged_queue;
1882 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1883 q->properties.type)];
1885 deallocate_doorbell(qpd, q);
1887 if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) ||
1888 (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
1889 deallocate_sdma_queue(dqm, q);
1890 pdd->sdma_past_activity_counter += sdma_val;
1895 if (q->properties.is_active) {
1896 if (!dqm->dev->shared_resources.enable_mes) {
1897 decrement_queue_count(dqm, qpd, q);
1898 retval = execute_queues_cpsch(dqm,
1899 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1900 if (retval == -ETIME)
1901 qpd->reset_wavefronts = true;
1903 retval = remove_queue_mes(dqm, q, qpd);
1908 * Unconditionally decrement this counter, regardless of the queue's
1911 dqm->total_queue_count--;
1912 pr_debug("Total of %d queues are accountable so far\n",
1913 dqm->total_queue_count);
1917 /* Do free_mqd after dqm_unlock(dqm) to avoid circular locking */
1918 mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
1922 failed_try_destroy_debugged_queue:
1929 * Low bits must be 0000/FFFF as required by HW, high bits must be 0 to
1930 * stay in user mode.
1932 #define APE1_FIXED_BITS_MASK 0xFFFF80000000FFFFULL
1933 /* APE1 limit is inclusive and 64K aligned. */
1934 #define APE1_LIMIT_ALIGNMENT 0xFFFF
1936 static bool set_cache_memory_policy(struct device_queue_manager *dqm,
1937 struct qcm_process_device *qpd,
1938 enum cache_policy default_policy,
1939 enum cache_policy alternate_policy,
1940 void __user *alternate_aperture_base,
1941 uint64_t alternate_aperture_size)
1945 if (!dqm->asic_ops.set_cache_memory_policy)
1950 if (alternate_aperture_size == 0) {
1951 /* base > limit disables APE1 */
1952 qpd->sh_mem_ape1_base = 1;
1953 qpd->sh_mem_ape1_limit = 0;
1956 * In FSA64, APE1_Base[63:0] = { 16{SH_MEM_APE1_BASE[31]},
1957 * SH_MEM_APE1_BASE[31:0], 0x0000 }
1958 * APE1_Limit[63:0] = { 16{SH_MEM_APE1_LIMIT[31]},
1959 * SH_MEM_APE1_LIMIT[31:0], 0xFFFF }
1960 * Verify that the base and size parameters can be
1961 * represented in this format and convert them.
1962 * Additionally restrict APE1 to user-mode addresses.
1965 uint64_t base = (uintptr_t)alternate_aperture_base;
1966 uint64_t limit = base + alternate_aperture_size - 1;
1968 if (limit <= base || (base & APE1_FIXED_BITS_MASK) != 0 ||
1969 (limit & APE1_FIXED_BITS_MASK) != APE1_LIMIT_ALIGNMENT) {
1974 qpd->sh_mem_ape1_base = base >> 16;
1975 qpd->sh_mem_ape1_limit = limit >> 16;
1978 retval = dqm->asic_ops.set_cache_memory_policy(
1983 alternate_aperture_base,
1984 alternate_aperture_size);
1986 if ((dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0))
1987 program_sh_mem_settings(dqm, qpd);
1989 pr_debug("sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n",
1990 qpd->sh_mem_config, qpd->sh_mem_ape1_base,
1991 qpd->sh_mem_ape1_limit);
1998 static int process_termination_nocpsch(struct device_queue_manager *dqm,
1999 struct qcm_process_device *qpd)
2002 struct device_process_node *cur, *next_dpn;
2008 /* Clear all user mode queues */
2009 while (!list_empty(&qpd->queues_list)) {
2010 struct mqd_manager *mqd_mgr;
2013 q = list_first_entry(&qpd->queues_list, struct queue, list);
2014 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
2015 q->properties.type)];
2016 ret = destroy_queue_nocpsch_locked(dqm, qpd, q);
2020 mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
2024 /* Unregister process */
2025 list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
2026 if (qpd == cur->qpd) {
2027 list_del(&cur->list);
2029 dqm->processes_count--;
2037 /* Outside the DQM lock because under the DQM lock we can't do
2038 * reclaim or take other locks that others hold while reclaiming.
2041 kfd_dec_compute_active(dqm->dev);
2046 static int get_wave_state(struct device_queue_manager *dqm,
2048 void __user *ctl_stack,
2049 u32 *ctl_stack_used_size,
2050 u32 *save_area_used_size)
2052 struct mqd_manager *mqd_mgr;
2056 mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_CP];
2058 if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE ||
2059 q->properties.is_active || !q->device->cwsr_enabled ||
2060 !mqd_mgr->get_wave_state) {
2068 * get_wave_state is outside the dqm lock to prevent circular locking
2069 * and the queue should be protected against destruction by the process
2072 return mqd_mgr->get_wave_state(mqd_mgr, q->mqd, ctl_stack,
2073 ctl_stack_used_size, save_area_used_size);
2076 static void get_queue_checkpoint_info(struct device_queue_manager *dqm,
2077 const struct queue *q,
2079 u32 *ctl_stack_size)
2081 struct mqd_manager *mqd_mgr;
2082 enum KFD_MQD_TYPE mqd_type =
2083 get_mqd_type_from_queue_type(q->properties.type);
2086 mqd_mgr = dqm->mqd_mgrs[mqd_type];
2087 *mqd_size = mqd_mgr->mqd_size;
2088 *ctl_stack_size = 0;
2090 if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE && mqd_mgr->get_checkpoint_info)
2091 mqd_mgr->get_checkpoint_info(mqd_mgr, q->mqd, ctl_stack_size);
2096 static int checkpoint_mqd(struct device_queue_manager *dqm,
2097 const struct queue *q,
2101 struct mqd_manager *mqd_mgr;
2103 enum KFD_MQD_TYPE mqd_type =
2104 get_mqd_type_from_queue_type(q->properties.type);
2108 if (q->properties.is_active || !q->device->cwsr_enabled) {
2113 mqd_mgr = dqm->mqd_mgrs[mqd_type];
2114 if (!mqd_mgr->checkpoint_mqd) {
2119 mqd_mgr->checkpoint_mqd(mqd_mgr, q->mqd, mqd, ctl_stack);
2126 static int process_termination_cpsch(struct device_queue_manager *dqm,
2127 struct qcm_process_device *qpd)
2131 struct kernel_queue *kq, *kq_next;
2132 struct mqd_manager *mqd_mgr;
2133 struct device_process_node *cur, *next_dpn;
2134 enum kfd_unmap_queues_filter filter =
2135 KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES;
2142 /* Clean all kernel queues */
2143 list_for_each_entry_safe(kq, kq_next, &qpd->priv_queue_list, list) {
2144 list_del(&kq->list);
2145 decrement_queue_count(dqm, qpd, kq->queue);
2146 qpd->is_debug = false;
2147 dqm->total_queue_count--;
2148 filter = KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES;
2151 /* Clear all user mode queues */
2152 list_for_each_entry(q, &qpd->queues_list, list) {
2153 if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
2154 deallocate_sdma_queue(dqm, q);
2155 else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
2156 deallocate_sdma_queue(dqm, q);
2158 if (q->properties.is_active) {
2159 decrement_queue_count(dqm, qpd, q);
2161 if (dqm->dev->shared_resources.enable_mes) {
2162 retval = remove_queue_mes(dqm, q, qpd);
2164 pr_err("Failed to remove queue %d\n",
2165 q->properties.queue_id);
2169 dqm->total_queue_count--;
2172 /* Unregister process */
2173 list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
2174 if (qpd == cur->qpd) {
2175 list_del(&cur->list);
2177 dqm->processes_count--;
2183 if (!dqm->dev->shared_resources.enable_mes)
2184 retval = execute_queues_cpsch(dqm, filter, 0);
2186 if ((!dqm->is_hws_hang) && (retval || qpd->reset_wavefronts)) {
2187 pr_warn("Resetting wave fronts (cpsch) on dev %p\n", dqm->dev);
2188 dbgdev_wave_reset_wavefronts(dqm->dev, qpd->pqm->process);
2189 qpd->reset_wavefronts = false;
2192 /* Lastly, free mqd resources.
2193 * Do free_mqd() after dqm_unlock to avoid circular locking.
2195 while (!list_empty(&qpd->queues_list)) {
2196 q = list_first_entry(&qpd->queues_list, struct queue, list);
2197 mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
2198 q->properties.type)];
2202 mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
2207 /* Outside the DQM lock because under the DQM lock we can't do
2208 * reclaim or take other locks that others hold while reclaiming.
2211 kfd_dec_compute_active(dqm->dev);
2216 static int init_mqd_managers(struct device_queue_manager *dqm)
2219 struct mqd_manager *mqd_mgr;
2221 for (i = 0; i < KFD_MQD_TYPE_MAX; i++) {
2222 mqd_mgr = dqm->asic_ops.mqd_manager_init(i, dqm->dev);
2224 pr_err("mqd manager [%d] initialization failed\n", i);
2227 dqm->mqd_mgrs[i] = mqd_mgr;
2233 for (j = 0; j < i; j++) {
2234 kfree(dqm->mqd_mgrs[j]);
2235 dqm->mqd_mgrs[j] = NULL;
2241 /* Allocate one hiq mqd (HWS) and all SDMA mqd in a continuous trunk*/
2242 static int allocate_hiq_sdma_mqd(struct device_queue_manager *dqm)
2245 struct kfd_dev *dev = dqm->dev;
2246 struct kfd_mem_obj *mem_obj = &dqm->hiq_sdma_mqd;
2247 uint32_t size = dqm->mqd_mgrs[KFD_MQD_TYPE_SDMA]->mqd_size *
2248 get_num_all_sdma_engines(dqm) *
2249 dev->device_info.num_sdma_queues_per_engine +
2250 dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size;
2252 retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev, size,
2253 &(mem_obj->gtt_mem), &(mem_obj->gpu_addr),
2254 (void *)&(mem_obj->cpu_ptr), false);
2259 struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
2261 struct device_queue_manager *dqm;
2263 pr_debug("Loading device queue manager\n");
2265 dqm = kzalloc(sizeof(*dqm), GFP_KERNEL);
2269 switch (dev->adev->asic_type) {
2270 /* HWS is not available on Hawaii. */
2272 /* HWS depends on CWSR for timely dequeue. CWSR is not
2273 * available on Tonga.
2275 * FIXME: This argument also applies to Kaveri.
2278 dqm->sched_policy = KFD_SCHED_POLICY_NO_HWS;
2281 dqm->sched_policy = sched_policy;
2286 switch (dqm->sched_policy) {
2287 case KFD_SCHED_POLICY_HWS:
2288 case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION:
2289 /* initialize dqm for cp scheduling */
2290 dqm->ops.create_queue = create_queue_cpsch;
2291 dqm->ops.initialize = initialize_cpsch;
2292 dqm->ops.start = start_cpsch;
2293 dqm->ops.stop = stop_cpsch;
2294 dqm->ops.pre_reset = pre_reset;
2295 dqm->ops.destroy_queue = destroy_queue_cpsch;
2296 dqm->ops.update_queue = update_queue;
2297 dqm->ops.register_process = register_process;
2298 dqm->ops.unregister_process = unregister_process;
2299 dqm->ops.uninitialize = uninitialize;
2300 dqm->ops.create_kernel_queue = create_kernel_queue_cpsch;
2301 dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch;
2302 dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
2303 dqm->ops.process_termination = process_termination_cpsch;
2304 dqm->ops.evict_process_queues = evict_process_queues_cpsch;
2305 dqm->ops.restore_process_queues = restore_process_queues_cpsch;
2306 dqm->ops.get_wave_state = get_wave_state;
2307 dqm->ops.reset_queues = reset_queues_cpsch;
2308 dqm->ops.get_queue_checkpoint_info = get_queue_checkpoint_info;
2309 dqm->ops.checkpoint_mqd = checkpoint_mqd;
2311 case KFD_SCHED_POLICY_NO_HWS:
2312 /* initialize dqm for no cp scheduling */
2313 dqm->ops.start = start_nocpsch;
2314 dqm->ops.stop = stop_nocpsch;
2315 dqm->ops.pre_reset = pre_reset;
2316 dqm->ops.create_queue = create_queue_nocpsch;
2317 dqm->ops.destroy_queue = destroy_queue_nocpsch;
2318 dqm->ops.update_queue = update_queue;
2319 dqm->ops.register_process = register_process;
2320 dqm->ops.unregister_process = unregister_process;
2321 dqm->ops.initialize = initialize_nocpsch;
2322 dqm->ops.uninitialize = uninitialize;
2323 dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
2324 dqm->ops.process_termination = process_termination_nocpsch;
2325 dqm->ops.evict_process_queues = evict_process_queues_nocpsch;
2326 dqm->ops.restore_process_queues =
2327 restore_process_queues_nocpsch;
2328 dqm->ops.get_wave_state = get_wave_state;
2329 dqm->ops.get_queue_checkpoint_info = get_queue_checkpoint_info;
2330 dqm->ops.checkpoint_mqd = checkpoint_mqd;
2333 pr_err("Invalid scheduling policy %d\n", dqm->sched_policy);
2337 switch (dev->adev->asic_type) {
2339 device_queue_manager_init_vi(&dqm->asic_ops);
2343 device_queue_manager_init_cik(&dqm->asic_ops);
2347 device_queue_manager_init_cik_hawaii(&dqm->asic_ops);
2352 case CHIP_POLARIS10:
2353 case CHIP_POLARIS11:
2354 case CHIP_POLARIS12:
2356 device_queue_manager_init_vi_tonga(&dqm->asic_ops);
2360 if (KFD_GC_VERSION(dev) >= IP_VERSION(11, 0, 0))
2361 device_queue_manager_init_v11(&dqm->asic_ops);
2362 else if (KFD_GC_VERSION(dev) >= IP_VERSION(10, 1, 1))
2363 device_queue_manager_init_v10_navi10(&dqm->asic_ops);
2364 else if (KFD_GC_VERSION(dev) >= IP_VERSION(9, 0, 1))
2365 device_queue_manager_init_v9(&dqm->asic_ops);
2367 WARN(1, "Unexpected ASIC family %u",
2368 dev->adev->asic_type);
2373 if (init_mqd_managers(dqm))
2376 if (allocate_hiq_sdma_mqd(dqm)) {
2377 pr_err("Failed to allocate hiq sdma mqd trunk buffer\n");
2381 if (!dqm->ops.initialize(dqm))
2389 static void deallocate_hiq_sdma_mqd(struct kfd_dev *dev,
2390 struct kfd_mem_obj *mqd)
2392 WARN(!mqd, "No hiq sdma mqd trunk to free");
2394 amdgpu_amdkfd_free_gtt_mem(dev->adev, mqd->gtt_mem);
2397 void device_queue_manager_uninit(struct device_queue_manager *dqm)
2399 dqm->ops.uninitialize(dqm);
2400 deallocate_hiq_sdma_mqd(dqm->dev, &dqm->hiq_sdma_mqd);
2404 int kfd_dqm_evict_pasid(struct device_queue_manager *dqm, u32 pasid)
2406 struct kfd_process_device *pdd;
2407 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
2412 WARN(debug_evictions, "Evicting pid %d", p->lead_thread->pid);
2413 pdd = kfd_get_process_device_data(dqm->dev, p);
2415 ret = dqm->ops.evict_process_queues(dqm, &pdd->qpd);
2416 kfd_unref_process(p);
2421 static void kfd_process_hw_exception(struct work_struct *work)
2423 struct device_queue_manager *dqm = container_of(work,
2424 struct device_queue_manager, hw_exception_work);
2425 amdgpu_amdkfd_gpu_reset(dqm->dev->adev);
2428 #if defined(CONFIG_DEBUG_FS)
2430 static void seq_reg_dump(struct seq_file *m,
2431 uint32_t (*dump)[2], uint32_t n_regs)
2435 for (i = 0, count = 0; i < n_regs; i++) {
2437 dump[i-1][0] + sizeof(uint32_t) != dump[i][0]) {
2438 seq_printf(m, "%s %08x: %08x",
2440 dump[i][0], dump[i][1]);
2443 seq_printf(m, " %08x", dump[i][1]);
2451 int dqm_debugfs_hqds(struct seq_file *m, void *data)
2453 struct device_queue_manager *dqm = data;
2454 uint32_t (*dump)[2], n_regs;
2458 if (!dqm->sched_running) {
2459 seq_puts(m, " Device is stopped\n");
2463 r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->adev,
2464 KFD_CIK_HIQ_PIPE, KFD_CIK_HIQ_QUEUE,
2467 seq_printf(m, " HIQ on MEC %d Pipe %d Queue %d\n",
2468 KFD_CIK_HIQ_PIPE/get_pipes_per_mec(dqm)+1,
2469 KFD_CIK_HIQ_PIPE%get_pipes_per_mec(dqm),
2471 seq_reg_dump(m, dump, n_regs);
2476 for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) {
2477 int pipe_offset = pipe * get_queues_per_pipe(dqm);
2479 for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) {
2480 if (!test_bit(pipe_offset + queue,
2481 dqm->dev->shared_resources.cp_queue_bitmap))
2484 r = dqm->dev->kfd2kgd->hqd_dump(
2485 dqm->dev->adev, pipe, queue, &dump, &n_regs);
2489 seq_printf(m, " CP Pipe %d, Queue %d\n",
2491 seq_reg_dump(m, dump, n_regs);
2497 for (pipe = 0; pipe < get_num_all_sdma_engines(dqm); pipe++) {
2499 queue < dqm->dev->device_info.num_sdma_queues_per_engine;
2501 r = dqm->dev->kfd2kgd->hqd_sdma_dump(
2502 dqm->dev->adev, pipe, queue, &dump, &n_regs);
2506 seq_printf(m, " SDMA Engine %d, RLC %d\n",
2508 seq_reg_dump(m, dump, n_regs);
2517 int dqm_debugfs_hang_hws(struct device_queue_manager *dqm)
2522 r = pm_debugfs_hang_hws(&dqm->packet_mgr);
2527 dqm->active_runlist = true;
2528 r = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);