Linux Audio

Check our new training course

Loading...
v6.9.4
   1// SPDX-License-Identifier: GPL-2.0 OR MIT
   2/*
   3 * Copyright 2014-2022 Advanced Micro Devices, Inc.
   4 *
   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:
  11 *
  12 * The above copyright notice and this permission notice shall be included in
  13 * all copies or substantial portions of the Software.
  14 *
  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.
  22 *
  23 */
  24
  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>
  32#include "kfd_priv.h"
  33#include "kfd_device_queue_manager.h"
  34#include "kfd_mqd_manager.h"
  35#include "cik_regs.h"
  36#include "kfd_kernel_queue.h"
  37#include "amdgpu_amdkfd.h"
  38#include "mes_api_def.h"
  39#include "kfd_debug.h"
  40
  41/* Size of the per-pipe EOP queue */
  42#define CIK_HPD_EOP_BYTES_LOG2 11
  43#define CIK_HPD_EOP_BYTES (1U << CIK_HPD_EOP_BYTES_LOG2)
  44
  45static int set_pasid_vmid_mapping(struct device_queue_manager *dqm,
  46				  u32 pasid, unsigned int vmid);
  47
  48static int execute_queues_cpsch(struct device_queue_manager *dqm,
  49				enum kfd_unmap_queues_filter filter,
  50				uint32_t filter_param,
  51				uint32_t grace_period);
  52static int unmap_queues_cpsch(struct device_queue_manager *dqm,
  53				enum kfd_unmap_queues_filter filter,
  54				uint32_t filter_param,
  55				uint32_t grace_period,
  56				bool reset);
  57
  58static int map_queues_cpsch(struct device_queue_manager *dqm);
  59
  60static void deallocate_sdma_queue(struct device_queue_manager *dqm,
  61				struct queue *q);
  62
  63static inline void deallocate_hqd(struct device_queue_manager *dqm,
  64				struct queue *q);
  65static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q);
  66static int allocate_sdma_queue(struct device_queue_manager *dqm,
  67				struct queue *q, const uint32_t *restore_sdma_id);
  68static void kfd_process_hw_exception(struct work_struct *work);
  69
  70static inline
  71enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type)
  72{
  73	if (type == KFD_QUEUE_TYPE_SDMA || type == KFD_QUEUE_TYPE_SDMA_XGMI)
  74		return KFD_MQD_TYPE_SDMA;
  75	return KFD_MQD_TYPE_CP;
  76}
  77
  78static bool is_pipe_enabled(struct device_queue_manager *dqm, int mec, int pipe)
  79{
  80	int i;
  81	int pipe_offset = (mec * dqm->dev->kfd->shared_resources.num_pipe_per_mec
  82		+ pipe) * dqm->dev->kfd->shared_resources.num_queue_per_pipe;
  83
  84	/* queue is available for KFD usage if bit is 1 */
  85	for (i = 0; i <  dqm->dev->kfd->shared_resources.num_queue_per_pipe; ++i)
  86		if (test_bit(pipe_offset + i,
  87			      dqm->dev->kfd->shared_resources.cp_queue_bitmap))
  88			return true;
  89	return false;
  90}
  91
  92unsigned int get_cp_queues_num(struct device_queue_manager *dqm)
  93{
  94	return bitmap_weight(dqm->dev->kfd->shared_resources.cp_queue_bitmap,
  95				AMDGPU_MAX_QUEUES);
  96}
  97
  98unsigned int get_queues_per_pipe(struct device_queue_manager *dqm)
  99{
 100	return dqm->dev->kfd->shared_resources.num_queue_per_pipe;
 101}
 102
 103unsigned int get_pipes_per_mec(struct device_queue_manager *dqm)
 104{
 105	return dqm->dev->kfd->shared_resources.num_pipe_per_mec;
 106}
 107
 108static unsigned int get_num_all_sdma_engines(struct device_queue_manager *dqm)
 109{
 110	return kfd_get_num_sdma_engines(dqm->dev) +
 111		kfd_get_num_xgmi_sdma_engines(dqm->dev);
 112}
 113
 114unsigned int get_num_sdma_queues(struct device_queue_manager *dqm)
 115{
 116	return kfd_get_num_sdma_engines(dqm->dev) *
 117		dqm->dev->kfd->device_info.num_sdma_queues_per_engine;
 118}
 119
 120unsigned int get_num_xgmi_sdma_queues(struct device_queue_manager *dqm)
 121{
 122	return kfd_get_num_xgmi_sdma_engines(dqm->dev) *
 123		dqm->dev->kfd->device_info.num_sdma_queues_per_engine;
 124}
 125
 126static void init_sdma_bitmaps(struct device_queue_manager *dqm)
 127{
 128	bitmap_zero(dqm->sdma_bitmap, KFD_MAX_SDMA_QUEUES);
 129	bitmap_set(dqm->sdma_bitmap, 0, get_num_sdma_queues(dqm));
 130
 131	bitmap_zero(dqm->xgmi_sdma_bitmap, KFD_MAX_SDMA_QUEUES);
 132	bitmap_set(dqm->xgmi_sdma_bitmap, 0, get_num_xgmi_sdma_queues(dqm));
 133
 134	/* Mask out the reserved queues */
 135	bitmap_andnot(dqm->sdma_bitmap, dqm->sdma_bitmap,
 136		      dqm->dev->kfd->device_info.reserved_sdma_queues_bitmap,
 137		      KFD_MAX_SDMA_QUEUES);
 138}
 139
 140void program_sh_mem_settings(struct device_queue_manager *dqm,
 141					struct qcm_process_device *qpd)
 142{
 143	uint32_t xcc_mask = dqm->dev->xcc_mask;
 144	int xcc_id;
 145
 146	for_each_inst(xcc_id, xcc_mask)
 147		dqm->dev->kfd2kgd->program_sh_mem_settings(
 148			dqm->dev->adev, qpd->vmid, qpd->sh_mem_config,
 149			qpd->sh_mem_ape1_base, qpd->sh_mem_ape1_limit,
 150			qpd->sh_mem_bases, xcc_id);
 151}
 152
 153static void kfd_hws_hang(struct device_queue_manager *dqm)
 154{
 155	/*
 156	 * Issue a GPU reset if HWS is unresponsive
 157	 */
 158	dqm->is_hws_hang = true;
 159
 160	/* It's possible we're detecting a HWS hang in the
 161	 * middle of a GPU reset. No need to schedule another
 162	 * reset in this case.
 163	 */
 164	if (!dqm->is_resetting)
 165		schedule_work(&dqm->hw_exception_work);
 166}
 167
 168static int convert_to_mes_queue_type(int queue_type)
 169{
 170	int mes_queue_type;
 171
 172	switch (queue_type) {
 173	case KFD_QUEUE_TYPE_COMPUTE:
 174		mes_queue_type = MES_QUEUE_TYPE_COMPUTE;
 175		break;
 176	case KFD_QUEUE_TYPE_SDMA:
 177		mes_queue_type = MES_QUEUE_TYPE_SDMA;
 178		break;
 179	default:
 180		WARN(1, "Invalid queue type %d", queue_type);
 181		mes_queue_type = -EINVAL;
 182		break;
 183	}
 184
 185	return mes_queue_type;
 186}
 187
 188static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
 189			 struct qcm_process_device *qpd)
 190{
 191	struct amdgpu_device *adev = (struct amdgpu_device *)dqm->dev->adev;
 192	struct kfd_process_device *pdd = qpd_to_pdd(qpd);
 193	struct mes_add_queue_input queue_input;
 194	int r, queue_type;
 195	uint64_t wptr_addr_off;
 196
 197	if (dqm->is_hws_hang)
 198		return -EIO;
 199
 200	memset(&queue_input, 0x0, sizeof(struct mes_add_queue_input));
 201	queue_input.process_id = qpd->pqm->process->pasid;
 202	queue_input.page_table_base_addr =  qpd->page_table_base;
 203	queue_input.process_va_start = 0;
 204	queue_input.process_va_end = adev->vm_manager.max_pfn - 1;
 205	/* MES unit for quantum is 100ns */
 206	queue_input.process_quantum = KFD_MES_PROCESS_QUANTUM;  /* Equivalent to 10ms. */
 207	queue_input.process_context_addr = pdd->proc_ctx_gpu_addr;
 208	queue_input.gang_quantum = KFD_MES_GANG_QUANTUM; /* Equivalent to 1ms */
 209	queue_input.gang_context_addr = q->gang_ctx_gpu_addr;
 210	queue_input.inprocess_gang_priority = q->properties.priority;
 211	queue_input.gang_global_priority_level =
 212					AMDGPU_MES_PRIORITY_LEVEL_NORMAL;
 213	queue_input.doorbell_offset = q->properties.doorbell_off;
 214	queue_input.mqd_addr = q->gart_mqd_addr;
 215	queue_input.wptr_addr = (uint64_t)q->properties.write_ptr;
 216
 217	if (q->wptr_bo) {
 218		wptr_addr_off = (uint64_t)q->properties.write_ptr & (PAGE_SIZE - 1);
 219		queue_input.wptr_mc_addr = amdgpu_bo_gpu_offset(q->wptr_bo) + wptr_addr_off;
 220	}
 221
 222	queue_input.is_kfd_process = 1;
 223	queue_input.is_aql_queue = (q->properties.format == KFD_QUEUE_FORMAT_AQL);
 224	queue_input.queue_size = q->properties.queue_size >> 2;
 225
 226	queue_input.paging = false;
 227	queue_input.tba_addr = qpd->tba_addr;
 228	queue_input.tma_addr = qpd->tma_addr;
 229	queue_input.trap_en = !kfd_dbg_has_cwsr_workaround(q->device);
 230	queue_input.skip_process_ctx_clear =
 231		qpd->pqm->process->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED &&
 232						(qpd->pqm->process->debug_trap_enabled ||
 233						 kfd_dbg_has_ttmps_always_setup(q->device));
 234
 235	queue_type = convert_to_mes_queue_type(q->properties.type);
 236	if (queue_type < 0) {
 237		dev_err(adev->dev, "Queue type not supported with MES, queue:%d\n",
 238			q->properties.type);
 239		return -EINVAL;
 240	}
 241	queue_input.queue_type = (uint32_t)queue_type;
 242
 243	queue_input.exclusively_scheduled = q->properties.is_gws;
 
 
 
 244
 245	amdgpu_mes_lock(&adev->mes);
 246	r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input);
 247	amdgpu_mes_unlock(&adev->mes);
 248	if (r) {
 249		dev_err(adev->dev, "failed to add hardware queue to MES, doorbell=0x%x\n",
 250			q->properties.doorbell_off);
 251		dev_err(adev->dev, "MES might be in unrecoverable state, issue a GPU reset\n");
 252		kfd_hws_hang(dqm);
 253	}
 254
 255	return r;
 256}
 257
 258static int remove_queue_mes(struct device_queue_manager *dqm, struct queue *q,
 259			struct qcm_process_device *qpd)
 260{
 261	struct amdgpu_device *adev = (struct amdgpu_device *)dqm->dev->adev;
 262	int r;
 263	struct mes_remove_queue_input queue_input;
 264
 265	if (dqm->is_hws_hang)
 266		return -EIO;
 267
 268	memset(&queue_input, 0x0, sizeof(struct mes_remove_queue_input));
 269	queue_input.doorbell_offset = q->properties.doorbell_off;
 270	queue_input.gang_context_addr = q->gang_ctx_gpu_addr;
 271
 272	amdgpu_mes_lock(&adev->mes);
 273	r = adev->mes.funcs->remove_hw_queue(&adev->mes, &queue_input);
 274	amdgpu_mes_unlock(&adev->mes);
 275
 276	if (r) {
 277		dev_err(adev->dev, "failed to remove hardware queue from MES, doorbell=0x%x\n",
 278			q->properties.doorbell_off);
 279		dev_err(adev->dev, "MES might be in unrecoverable state, issue a GPU reset\n");
 280		kfd_hws_hang(dqm);
 281	}
 282
 283	return r;
 284}
 285
 286static int remove_all_queues_mes(struct device_queue_manager *dqm)
 287{
 288	struct device_process_node *cur;
 289	struct device *dev = dqm->dev->adev->dev;
 290	struct qcm_process_device *qpd;
 291	struct queue *q;
 292	int retval = 0;
 293
 294	list_for_each_entry(cur, &dqm->queues, list) {
 295		qpd = cur->qpd;
 296		list_for_each_entry(q, &qpd->queues_list, list) {
 297			if (q->properties.is_active) {
 298				retval = remove_queue_mes(dqm, q, qpd);
 299				if (retval) {
 300					dev_err(dev, "%s: Failed to remove queue %d for dev %d",
 301						__func__,
 302						q->properties.queue_id,
 303						dqm->dev->id);
 304					return retval;
 305				}
 306			}
 307		}
 308	}
 309
 310	return retval;
 311}
 312
 313static void increment_queue_count(struct device_queue_manager *dqm,
 314				  struct qcm_process_device *qpd,
 315				  struct queue *q)
 316{
 317	dqm->active_queue_count++;
 318	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
 319	    q->properties.type == KFD_QUEUE_TYPE_DIQ)
 320		dqm->active_cp_queue_count++;
 321
 322	if (q->properties.is_gws) {
 323		dqm->gws_queue_count++;
 324		qpd->mapped_gws_queue = true;
 325	}
 326}
 327
 328static void decrement_queue_count(struct device_queue_manager *dqm,
 329				  struct qcm_process_device *qpd,
 330				  struct queue *q)
 331{
 332	dqm->active_queue_count--;
 333	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
 334	    q->properties.type == KFD_QUEUE_TYPE_DIQ)
 335		dqm->active_cp_queue_count--;
 336
 337	if (q->properties.is_gws) {
 338		dqm->gws_queue_count--;
 339		qpd->mapped_gws_queue = false;
 340	}
 341}
 342
 343/*
 344 * Allocate a doorbell ID to this queue.
 345 * If doorbell_id is passed in, make sure requested ID is valid then allocate it.
 346 */
 347static int allocate_doorbell(struct qcm_process_device *qpd,
 348			     struct queue *q,
 349			     uint32_t const *restore_id)
 350{
 351	struct kfd_node *dev = qpd->dqm->dev;
 352
 353	if (!KFD_IS_SOC15(dev)) {
 354		/* On pre-SOC15 chips we need to use the queue ID to
 355		 * preserve the user mode ABI.
 356		 */
 357
 358		if (restore_id && *restore_id != q->properties.queue_id)
 359			return -EINVAL;
 360
 361		q->doorbell_id = q->properties.queue_id;
 362	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
 363			q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
 364		/* For SDMA queues on SOC15 with 8-byte doorbell, use static
 365		 * doorbell assignments based on the engine and queue id.
 366		 * The doobell index distance between RLC (2*i) and (2*i+1)
 367		 * for a SDMA engine is 512.
 368		 */
 369
 370		uint32_t *idx_offset = dev->kfd->shared_resources.sdma_doorbell_idx;
 371
 372		/*
 373		 * q->properties.sdma_engine_id corresponds to the virtual
 374		 * sdma engine number. However, for doorbell allocation,
 375		 * we need the physical sdma engine id in order to get the
 376		 * correct doorbell offset.
 377		 */
 378		uint32_t valid_id = idx_offset[qpd->dqm->dev->node_id *
 379					       get_num_all_sdma_engines(qpd->dqm) +
 380					       q->properties.sdma_engine_id]
 381						+ (q->properties.sdma_queue_id & 1)
 382						* KFD_QUEUE_DOORBELL_MIRROR_OFFSET
 383						+ (q->properties.sdma_queue_id >> 1);
 384
 385		if (restore_id && *restore_id != valid_id)
 386			return -EINVAL;
 387		q->doorbell_id = valid_id;
 388	} else {
 389		/* For CP queues on SOC15 */
 390		if (restore_id) {
 391			/* make sure that ID is free  */
 392			if (__test_and_set_bit(*restore_id, qpd->doorbell_bitmap))
 393				return -EINVAL;
 394
 395			q->doorbell_id = *restore_id;
 396		} else {
 397			/* or reserve a free doorbell ID */
 398			unsigned int found;
 399
 400			found = find_first_zero_bit(qpd->doorbell_bitmap,
 401						    KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
 402			if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) {
 403				pr_debug("No doorbells available");
 404				return -EBUSY;
 405			}
 406			set_bit(found, qpd->doorbell_bitmap);
 407			q->doorbell_id = found;
 408		}
 409	}
 410
 411	q->properties.doorbell_off = amdgpu_doorbell_index_on_bar(dev->adev,
 412								  qpd->proc_doorbells,
 413								  q->doorbell_id,
 414								  dev->kfd->device_info.doorbell_size);
 415	return 0;
 416}
 417
 418static void deallocate_doorbell(struct qcm_process_device *qpd,
 419				struct queue *q)
 420{
 421	unsigned int old;
 422	struct kfd_node *dev = qpd->dqm->dev;
 423
 424	if (!KFD_IS_SOC15(dev) ||
 425	    q->properties.type == KFD_QUEUE_TYPE_SDMA ||
 426	    q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
 427		return;
 428
 429	old = test_and_clear_bit(q->doorbell_id, qpd->doorbell_bitmap);
 430	WARN_ON(!old);
 431}
 432
 433static void program_trap_handler_settings(struct device_queue_manager *dqm,
 434				struct qcm_process_device *qpd)
 435{
 436	uint32_t xcc_mask = dqm->dev->xcc_mask;
 437	int xcc_id;
 438
 439	if (dqm->dev->kfd2kgd->program_trap_handler_settings)
 440		for_each_inst(xcc_id, xcc_mask)
 441			dqm->dev->kfd2kgd->program_trap_handler_settings(
 442				dqm->dev->adev, qpd->vmid, qpd->tba_addr,
 443				qpd->tma_addr, xcc_id);
 444}
 445
 446static int allocate_vmid(struct device_queue_manager *dqm,
 447			struct qcm_process_device *qpd,
 448			struct queue *q)
 449{
 450	struct device *dev = dqm->dev->adev->dev;
 451	int allocated_vmid = -1, i;
 452
 453	for (i = dqm->dev->vm_info.first_vmid_kfd;
 454			i <= dqm->dev->vm_info.last_vmid_kfd; i++) {
 455		if (!dqm->vmid_pasid[i]) {
 456			allocated_vmid = i;
 457			break;
 458		}
 459	}
 460
 461	if (allocated_vmid < 0) {
 462		dev_err(dev, "no more vmid to allocate\n");
 463		return -ENOSPC;
 464	}
 465
 466	pr_debug("vmid allocated: %d\n", allocated_vmid);
 467
 468	dqm->vmid_pasid[allocated_vmid] = q->process->pasid;
 469
 470	set_pasid_vmid_mapping(dqm, q->process->pasid, allocated_vmid);
 471
 472	qpd->vmid = allocated_vmid;
 473	q->properties.vmid = allocated_vmid;
 474
 475	program_sh_mem_settings(dqm, qpd);
 476
 477	if (KFD_IS_SOC15(dqm->dev) && dqm->dev->kfd->cwsr_enabled)
 478		program_trap_handler_settings(dqm, qpd);
 479
 480	/* qpd->page_table_base is set earlier when register_process()
 481	 * is called, i.e. when the first queue is created.
 482	 */
 483	dqm->dev->kfd2kgd->set_vm_context_page_table_base(dqm->dev->adev,
 484			qpd->vmid,
 485			qpd->page_table_base);
 486	/* invalidate the VM context after pasid and vmid mapping is set up */
 487	kfd_flush_tlb(qpd_to_pdd(qpd), TLB_FLUSH_LEGACY);
 488
 489	if (dqm->dev->kfd2kgd->set_scratch_backing_va)
 490		dqm->dev->kfd2kgd->set_scratch_backing_va(dqm->dev->adev,
 491				qpd->sh_hidden_private_base, qpd->vmid);
 492
 493	return 0;
 494}
 495
 496static int flush_texture_cache_nocpsch(struct kfd_node *kdev,
 497				struct qcm_process_device *qpd)
 498{
 499	const struct packet_manager_funcs *pmf = qpd->dqm->packet_mgr.pmf;
 500	int ret;
 501
 502	if (!qpd->ib_kaddr)
 503		return -ENOMEM;
 504
 505	ret = pmf->release_mem(qpd->ib_base, (uint32_t *)qpd->ib_kaddr);
 506	if (ret)
 507		return ret;
 508
 509	return amdgpu_amdkfd_submit_ib(kdev->adev, KGD_ENGINE_MEC1, qpd->vmid,
 510				qpd->ib_base, (uint32_t *)qpd->ib_kaddr,
 511				pmf->release_mem_size / sizeof(uint32_t));
 512}
 513
 514static void deallocate_vmid(struct device_queue_manager *dqm,
 515				struct qcm_process_device *qpd,
 516				struct queue *q)
 517{
 518	struct device *dev = dqm->dev->adev->dev;
 519
 520	/* On GFX v7, CP doesn't flush TC at dequeue */
 521	if (q->device->adev->asic_type == CHIP_HAWAII)
 522		if (flush_texture_cache_nocpsch(q->device, qpd))
 523			dev_err(dev, "Failed to flush TC\n");
 524
 525	kfd_flush_tlb(qpd_to_pdd(qpd), TLB_FLUSH_LEGACY);
 526
 527	/* Release the vmid mapping */
 528	set_pasid_vmid_mapping(dqm, 0, qpd->vmid);
 529	dqm->vmid_pasid[qpd->vmid] = 0;
 530
 531	qpd->vmid = 0;
 532	q->properties.vmid = 0;
 533}
 534
 535static int create_queue_nocpsch(struct device_queue_manager *dqm,
 536				struct queue *q,
 537				struct qcm_process_device *qpd,
 538				const struct kfd_criu_queue_priv_data *qd,
 539				const void *restore_mqd, const void *restore_ctl_stack)
 540{
 541	struct mqd_manager *mqd_mgr;
 542	int retval;
 543
 544	dqm_lock(dqm);
 545
 546	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
 547		pr_warn("Can't create new usermode queue because %d queues were already created\n",
 548				dqm->total_queue_count);
 549		retval = -EPERM;
 550		goto out_unlock;
 551	}
 552
 553	if (list_empty(&qpd->queues_list)) {
 554		retval = allocate_vmid(dqm, qpd, q);
 555		if (retval)
 556			goto out_unlock;
 557	}
 558	q->properties.vmid = qpd->vmid;
 559	/*
 560	 * Eviction state logic: mark all queues as evicted, even ones
 561	 * not currently active. Restoring inactive queues later only
 562	 * updates the is_evicted flag but is a no-op otherwise.
 563	 */
 564	q->properties.is_evicted = !!qpd->evicted;
 565
 566	q->properties.tba_addr = qpd->tba_addr;
 567	q->properties.tma_addr = qpd->tma_addr;
 568
 569	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
 570			q->properties.type)];
 571	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) {
 572		retval = allocate_hqd(dqm, q);
 573		if (retval)
 574			goto deallocate_vmid;
 575		pr_debug("Loading mqd to hqd on pipe %d, queue %d\n",
 576			q->pipe, q->queue);
 577	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
 578		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
 579		retval = allocate_sdma_queue(dqm, q, qd ? &qd->sdma_id : NULL);
 580		if (retval)
 581			goto deallocate_vmid;
 582		dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
 583	}
 584
 585	retval = allocate_doorbell(qpd, q, qd ? &qd->doorbell_id : NULL);
 586	if (retval)
 587		goto out_deallocate_hqd;
 588
 589	/* Temporarily release dqm lock to avoid a circular lock dependency */
 590	dqm_unlock(dqm);
 591	q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties);
 592	dqm_lock(dqm);
 593
 594	if (!q->mqd_mem_obj) {
 595		retval = -ENOMEM;
 596		goto out_deallocate_doorbell;
 597	}
 598
 599	if (qd)
 600		mqd_mgr->restore_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, &q->gart_mqd_addr,
 601				     &q->properties, restore_mqd, restore_ctl_stack,
 602				     qd->ctl_stack_size);
 603	else
 604		mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj,
 605					&q->gart_mqd_addr, &q->properties);
 606
 607	if (q->properties.is_active) {
 608		if (!dqm->sched_running) {
 609			WARN_ONCE(1, "Load non-HWS mqd while stopped\n");
 610			goto add_queue_to_list;
 611		}
 612
 613		if (WARN(q->process->mm != current->mm,
 614					"should only run in user thread"))
 615			retval = -EFAULT;
 616		else
 617			retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe,
 618					q->queue, &q->properties, current->mm);
 619		if (retval)
 620			goto out_free_mqd;
 621	}
 622
 623add_queue_to_list:
 624	list_add(&q->list, &qpd->queues_list);
 625	qpd->queue_count++;
 626	if (q->properties.is_active)
 627		increment_queue_count(dqm, qpd, q);
 628
 629	/*
 630	 * Unconditionally increment this counter, regardless of the queue's
 631	 * type or whether the queue is active.
 632	 */
 633	dqm->total_queue_count++;
 634	pr_debug("Total of %d queues are accountable so far\n",
 635			dqm->total_queue_count);
 636	goto out_unlock;
 637
 638out_free_mqd:
 639	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
 640out_deallocate_doorbell:
 641	deallocate_doorbell(qpd, q);
 642out_deallocate_hqd:
 643	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
 644		deallocate_hqd(dqm, q);
 645	else if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
 646		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
 647		deallocate_sdma_queue(dqm, q);
 648deallocate_vmid:
 649	if (list_empty(&qpd->queues_list))
 650		deallocate_vmid(dqm, qpd, q);
 651out_unlock:
 652	dqm_unlock(dqm);
 653	return retval;
 654}
 655
 656static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q)
 657{
 658	bool set;
 659	int pipe, bit, i;
 660
 661	set = false;
 662
 663	for (pipe = dqm->next_pipe_to_allocate, i = 0;
 664			i < get_pipes_per_mec(dqm);
 665			pipe = ((pipe + 1) % get_pipes_per_mec(dqm)), ++i) {
 666
 667		if (!is_pipe_enabled(dqm, 0, pipe))
 668			continue;
 669
 670		if (dqm->allocated_queues[pipe] != 0) {
 671			bit = ffs(dqm->allocated_queues[pipe]) - 1;
 672			dqm->allocated_queues[pipe] &= ~(1 << bit);
 673			q->pipe = pipe;
 674			q->queue = bit;
 675			set = true;
 676			break;
 677		}
 678	}
 679
 680	if (!set)
 681		return -EBUSY;
 682
 683	pr_debug("hqd slot - pipe %d, queue %d\n", q->pipe, q->queue);
 684	/* horizontal hqd allocation */
 685	dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_per_mec(dqm);
 686
 687	return 0;
 688}
 689
 690static inline void deallocate_hqd(struct device_queue_manager *dqm,
 691				struct queue *q)
 692{
 693	dqm->allocated_queues[q->pipe] |= (1 << q->queue);
 694}
 695
 696#define SQ_IND_CMD_CMD_KILL		0x00000003
 697#define SQ_IND_CMD_MODE_BROADCAST	0x00000001
 698
 699static int dbgdev_wave_reset_wavefronts(struct kfd_node *dev, struct kfd_process *p)
 700{
 701	int status = 0;
 702	unsigned int vmid;
 703	uint16_t queried_pasid;
 704	union SQ_CMD_BITS reg_sq_cmd;
 705	union GRBM_GFX_INDEX_BITS reg_gfx_index;
 706	struct kfd_process_device *pdd;
 707	int first_vmid_to_scan = dev->vm_info.first_vmid_kfd;
 708	int last_vmid_to_scan = dev->vm_info.last_vmid_kfd;
 709	uint32_t xcc_mask = dev->xcc_mask;
 710	int xcc_id;
 711
 712	reg_sq_cmd.u32All = 0;
 713	reg_gfx_index.u32All = 0;
 714
 715	pr_debug("Killing all process wavefronts\n");
 716
 717	if (!dev->kfd2kgd->get_atc_vmid_pasid_mapping_info) {
 718		dev_err(dev->adev->dev, "no vmid pasid mapping supported\n");
 719		return -EOPNOTSUPP;
 720	}
 721
 722	/* Scan all registers in the range ATC_VMID8_PASID_MAPPING ..
 723	 * ATC_VMID15_PASID_MAPPING
 724	 * to check which VMID the current process is mapped to.
 725	 */
 726
 727	for (vmid = first_vmid_to_scan; vmid <= last_vmid_to_scan; vmid++) {
 728		status = dev->kfd2kgd->get_atc_vmid_pasid_mapping_info
 729				(dev->adev, vmid, &queried_pasid);
 730
 731		if (status && queried_pasid == p->pasid) {
 732			pr_debug("Killing wave fronts of vmid %d and pasid 0x%x\n",
 733					vmid, p->pasid);
 734			break;
 735		}
 736	}
 737
 738	if (vmid > last_vmid_to_scan) {
 739		dev_err(dev->adev->dev, "Didn't find vmid for pasid 0x%x\n", p->pasid);
 740		return -EFAULT;
 741	}
 742
 743	/* taking the VMID for that process on the safe way using PDD */
 744	pdd = kfd_get_process_device_data(dev, p);
 745	if (!pdd)
 746		return -EFAULT;
 747
 748	reg_gfx_index.bits.sh_broadcast_writes = 1;
 749	reg_gfx_index.bits.se_broadcast_writes = 1;
 750	reg_gfx_index.bits.instance_broadcast_writes = 1;
 751	reg_sq_cmd.bits.mode = SQ_IND_CMD_MODE_BROADCAST;
 752	reg_sq_cmd.bits.cmd = SQ_IND_CMD_CMD_KILL;
 753	reg_sq_cmd.bits.vm_id = vmid;
 754
 755	for_each_inst(xcc_id, xcc_mask)
 756		dev->kfd2kgd->wave_control_execute(
 757			dev->adev, reg_gfx_index.u32All,
 758			reg_sq_cmd.u32All, xcc_id);
 759
 760	return 0;
 761}
 762
 763/* Access to DQM has to be locked before calling destroy_queue_nocpsch_locked
 764 * to avoid asynchronized access
 765 */
 766static int destroy_queue_nocpsch_locked(struct device_queue_manager *dqm,
 767				struct qcm_process_device *qpd,
 768				struct queue *q)
 769{
 770	int retval;
 771	struct mqd_manager *mqd_mgr;
 772
 773	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
 774			q->properties.type)];
 775
 776	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
 777		deallocate_hqd(dqm, q);
 778	else if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
 779		deallocate_sdma_queue(dqm, q);
 780	else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
 781		deallocate_sdma_queue(dqm, q);
 782	else {
 783		pr_debug("q->properties.type %d is invalid\n",
 784				q->properties.type);
 785		return -EINVAL;
 786	}
 787	dqm->total_queue_count--;
 788
 789	deallocate_doorbell(qpd, q);
 790
 791	if (!dqm->sched_running) {
 792		WARN_ONCE(1, "Destroy non-HWS queue while stopped\n");
 793		return 0;
 794	}
 795
 796	retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
 797				KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
 798				KFD_UNMAP_LATENCY_MS,
 799				q->pipe, q->queue);
 800	if (retval == -ETIME)
 801		qpd->reset_wavefronts = true;
 802
 803	list_del(&q->list);
 804	if (list_empty(&qpd->queues_list)) {
 805		if (qpd->reset_wavefronts) {
 806			pr_warn("Resetting wave fronts (nocpsch) on dev %p\n",
 807					dqm->dev);
 808			/* dbgdev_wave_reset_wavefronts has to be called before
 809			 * deallocate_vmid(), i.e. when vmid is still in use.
 810			 */
 811			dbgdev_wave_reset_wavefronts(dqm->dev,
 812					qpd->pqm->process);
 813			qpd->reset_wavefronts = false;
 814		}
 815
 816		deallocate_vmid(dqm, qpd, q);
 817	}
 818	qpd->queue_count--;
 819	if (q->properties.is_active)
 820		decrement_queue_count(dqm, qpd, q);
 821
 822	return retval;
 823}
 824
 825static int destroy_queue_nocpsch(struct device_queue_manager *dqm,
 826				struct qcm_process_device *qpd,
 827				struct queue *q)
 828{
 829	int retval;
 830	uint64_t sdma_val = 0;
 831	struct device *dev = dqm->dev->adev->dev;
 832	struct kfd_process_device *pdd = qpd_to_pdd(qpd);
 833	struct mqd_manager *mqd_mgr =
 834		dqm->mqd_mgrs[get_mqd_type_from_queue_type(q->properties.type)];
 835
 836	/* Get the SDMA queue stats */
 837	if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) ||
 838	    (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
 839		retval = read_sdma_queue_counter((uint64_t __user *)q->properties.read_ptr,
 840							&sdma_val);
 841		if (retval)
 842			dev_err(dev, "Failed to read SDMA queue counter for queue: %d\n",
 843				q->properties.queue_id);
 844	}
 845
 846	dqm_lock(dqm);
 847	retval = destroy_queue_nocpsch_locked(dqm, qpd, q);
 848	if (!retval)
 849		pdd->sdma_past_activity_counter += sdma_val;
 850	dqm_unlock(dqm);
 851
 852	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
 853
 854	return retval;
 855}
 856
 857static int update_queue(struct device_queue_manager *dqm, struct queue *q,
 858			struct mqd_update_info *minfo)
 859{
 860	int retval = 0;
 861	struct device *dev = dqm->dev->adev->dev;
 862	struct mqd_manager *mqd_mgr;
 863	struct kfd_process_device *pdd;
 864	bool prev_active = false;
 865
 866	dqm_lock(dqm);
 867	pdd = kfd_get_process_device_data(q->device, q->process);
 868	if (!pdd) {
 869		retval = -ENODEV;
 870		goto out_unlock;
 871	}
 872	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
 873			q->properties.type)];
 874
 875	/* Save previous activity state for counters */
 876	prev_active = q->properties.is_active;
 877
 878	/* Make sure the queue is unmapped before updating the MQD */
 879	if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) {
 880		if (!dqm->dev->kfd->shared_resources.enable_mes)
 881			retval = unmap_queues_cpsch(dqm,
 882						    KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD, false);
 883		else if (prev_active)
 884			retval = remove_queue_mes(dqm, q, &pdd->qpd);
 885
 886		if (retval) {
 887			dev_err(dev, "unmap queue failed\n");
 888			goto out_unlock;
 889		}
 890	} else if (prev_active &&
 891		   (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
 892		    q->properties.type == KFD_QUEUE_TYPE_SDMA ||
 893		    q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
 894
 895		if (!dqm->sched_running) {
 896			WARN_ONCE(1, "Update non-HWS queue while stopped\n");
 897			goto out_unlock;
 898		}
 899
 900		retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
 901				(dqm->dev->kfd->cwsr_enabled ?
 902				 KFD_PREEMPT_TYPE_WAVEFRONT_SAVE :
 903				 KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN),
 904				KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
 905		if (retval) {
 906			dev_err(dev, "destroy mqd failed\n");
 907			goto out_unlock;
 908		}
 909	}
 910
 911	mqd_mgr->update_mqd(mqd_mgr, q->mqd, &q->properties, minfo);
 912
 913	/*
 914	 * check active state vs. the previous state and modify
 915	 * counter accordingly. map_queues_cpsch uses the
 916	 * dqm->active_queue_count to determine whether a new runlist must be
 917	 * uploaded.
 918	 */
 919	if (q->properties.is_active && !prev_active) {
 920		increment_queue_count(dqm, &pdd->qpd, q);
 921	} else if (!q->properties.is_active && prev_active) {
 922		decrement_queue_count(dqm, &pdd->qpd, q);
 923	} else if (q->gws && !q->properties.is_gws) {
 924		if (q->properties.is_active) {
 925			dqm->gws_queue_count++;
 926			pdd->qpd.mapped_gws_queue = true;
 927		}
 928		q->properties.is_gws = true;
 929	} else if (!q->gws && q->properties.is_gws) {
 930		if (q->properties.is_active) {
 931			dqm->gws_queue_count--;
 932			pdd->qpd.mapped_gws_queue = false;
 933		}
 934		q->properties.is_gws = false;
 935	}
 936
 937	if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) {
 938		if (!dqm->dev->kfd->shared_resources.enable_mes)
 939			retval = map_queues_cpsch(dqm);
 940		else if (q->properties.is_active)
 941			retval = add_queue_mes(dqm, q, &pdd->qpd);
 942	} else if (q->properties.is_active &&
 943		 (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
 944		  q->properties.type == KFD_QUEUE_TYPE_SDMA ||
 945		  q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
 946		if (WARN(q->process->mm != current->mm,
 947			 "should only run in user thread"))
 948			retval = -EFAULT;
 949		else
 950			retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd,
 951						   q->pipe, q->queue,
 952						   &q->properties, current->mm);
 953	}
 954
 955out_unlock:
 956	dqm_unlock(dqm);
 957	return retval;
 958}
 959
 960/* suspend_single_queue does not lock the dqm like the
 961 * evict_process_queues_cpsch or evict_process_queues_nocpsch. You should
 962 * lock the dqm before calling, and unlock after calling.
 963 *
 964 * The reason we don't lock the dqm is because this function may be
 965 * called on multiple queues in a loop, so rather than locking/unlocking
 966 * multiple times, we will just keep the dqm locked for all of the calls.
 967 */
 968static int suspend_single_queue(struct device_queue_manager *dqm,
 969				      struct kfd_process_device *pdd,
 970				      struct queue *q)
 971{
 972	bool is_new;
 973
 974	if (q->properties.is_suspended)
 975		return 0;
 976
 977	pr_debug("Suspending PASID %u queue [%i]\n",
 978			pdd->process->pasid,
 979			q->properties.queue_id);
 980
 981	is_new = q->properties.exception_status & KFD_EC_MASK(EC_QUEUE_NEW);
 982
 983	if (is_new || q->properties.is_being_destroyed) {
 984		pr_debug("Suspend: skip %s queue id %i\n",
 985				is_new ? "new" : "destroyed",
 986				q->properties.queue_id);
 987		return -EBUSY;
 988	}
 989
 990	q->properties.is_suspended = true;
 991	if (q->properties.is_active) {
 992		if (dqm->dev->kfd->shared_resources.enable_mes) {
 993			int r = remove_queue_mes(dqm, q, &pdd->qpd);
 994
 995			if (r)
 996				return r;
 997		}
 998
 999		decrement_queue_count(dqm, &pdd->qpd, q);
1000		q->properties.is_active = false;
1001	}
1002
1003	return 0;
1004}
1005
1006/* resume_single_queue does not lock the dqm like the functions
1007 * restore_process_queues_cpsch or restore_process_queues_nocpsch. You should
1008 * lock the dqm before calling, and unlock after calling.
1009 *
1010 * The reason we don't lock the dqm is because this function may be
1011 * called on multiple queues in a loop, so rather than locking/unlocking
1012 * multiple times, we will just keep the dqm locked for all of the calls.
1013 */
1014static int resume_single_queue(struct device_queue_manager *dqm,
1015				      struct qcm_process_device *qpd,
1016				      struct queue *q)
1017{
1018	struct kfd_process_device *pdd;
1019
1020	if (!q->properties.is_suspended)
1021		return 0;
1022
1023	pdd = qpd_to_pdd(qpd);
1024
1025	pr_debug("Restoring from suspend PASID %u queue [%i]\n",
1026			    pdd->process->pasid,
1027			    q->properties.queue_id);
1028
1029	q->properties.is_suspended = false;
1030
1031	if (QUEUE_IS_ACTIVE(q->properties)) {
1032		if (dqm->dev->kfd->shared_resources.enable_mes) {
1033			int r = add_queue_mes(dqm, q, &pdd->qpd);
1034
1035			if (r)
1036				return r;
1037		}
1038
1039		q->properties.is_active = true;
1040		increment_queue_count(dqm, qpd, q);
1041	}
1042
1043	return 0;
1044}
1045
1046static int evict_process_queues_nocpsch(struct device_queue_manager *dqm,
1047					struct qcm_process_device *qpd)
1048{
1049	struct queue *q;
1050	struct mqd_manager *mqd_mgr;
1051	struct kfd_process_device *pdd;
1052	int retval, ret = 0;
1053
1054	dqm_lock(dqm);
1055	if (qpd->evicted++ > 0) /* already evicted, do nothing */
1056		goto out;
1057
1058	pdd = qpd_to_pdd(qpd);
1059	pr_debug_ratelimited("Evicting PASID 0x%x queues\n",
1060			    pdd->process->pasid);
1061
1062	pdd->last_evict_timestamp = get_jiffies_64();
1063	/* Mark all queues as evicted. Deactivate all active queues on
1064	 * the qpd.
1065	 */
1066	list_for_each_entry(q, &qpd->queues_list, list) {
1067		q->properties.is_evicted = true;
1068		if (!q->properties.is_active)
1069			continue;
1070
1071		mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1072				q->properties.type)];
1073		q->properties.is_active = false;
1074		decrement_queue_count(dqm, qpd, q);
1075
1076		if (WARN_ONCE(!dqm->sched_running, "Evict when stopped\n"))
1077			continue;
1078
1079		retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
1080				(dqm->dev->kfd->cwsr_enabled ?
1081				 KFD_PREEMPT_TYPE_WAVEFRONT_SAVE :
1082				 KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN),
1083				KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
1084		if (retval && !ret)
1085			/* Return the first error, but keep going to
1086			 * maintain a consistent eviction state
1087			 */
1088			ret = retval;
1089	}
1090
1091out:
1092	dqm_unlock(dqm);
1093	return ret;
1094}
1095
1096static int evict_process_queues_cpsch(struct device_queue_manager *dqm,
1097				      struct qcm_process_device *qpd)
1098{
1099	struct queue *q;
1100	struct device *dev = dqm->dev->adev->dev;
1101	struct kfd_process_device *pdd;
1102	int retval = 0;
1103
1104	dqm_lock(dqm);
1105	if (qpd->evicted++ > 0) /* already evicted, do nothing */
1106		goto out;
1107
1108	pdd = qpd_to_pdd(qpd);
1109
1110	/* The debugger creates processes that temporarily have not acquired
1111	 * all VMs for all devices and has no VMs itself.
1112	 * Skip queue eviction on process eviction.
1113	 */
1114	if (!pdd->drm_priv)
1115		goto out;
1116
1117	pr_debug_ratelimited("Evicting PASID 0x%x queues\n",
1118			    pdd->process->pasid);
1119
1120	/* Mark all queues as evicted. Deactivate all active queues on
1121	 * the qpd.
1122	 */
1123	list_for_each_entry(q, &qpd->queues_list, list) {
1124		q->properties.is_evicted = true;
1125		if (!q->properties.is_active)
1126			continue;
1127
1128		q->properties.is_active = false;
1129		decrement_queue_count(dqm, qpd, q);
1130
1131		if (dqm->dev->kfd->shared_resources.enable_mes) {
1132			retval = remove_queue_mes(dqm, q, qpd);
1133			if (retval) {
1134				dev_err(dev, "Failed to evict queue %d\n",
1135					q->properties.queue_id);
1136				goto out;
1137			}
1138		}
1139	}
1140	pdd->last_evict_timestamp = get_jiffies_64();
1141	if (!dqm->dev->kfd->shared_resources.enable_mes)
1142		retval = execute_queues_cpsch(dqm,
1143					      qpd->is_debug ?
1144					      KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES :
1145					      KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0,
1146					      USE_DEFAULT_GRACE_PERIOD);
1147
1148out:
1149	dqm_unlock(dqm);
1150	return retval;
1151}
1152
1153static int restore_process_queues_nocpsch(struct device_queue_manager *dqm,
1154					  struct qcm_process_device *qpd)
1155{
1156	struct mm_struct *mm = NULL;
1157	struct queue *q;
1158	struct mqd_manager *mqd_mgr;
1159	struct kfd_process_device *pdd;
1160	uint64_t pd_base;
1161	uint64_t eviction_duration;
1162	int retval, ret = 0;
1163
1164	pdd = qpd_to_pdd(qpd);
1165	/* Retrieve PD base */
1166	pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv);
1167
1168	dqm_lock(dqm);
1169	if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */
1170		goto out;
1171	if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
1172		qpd->evicted--;
1173		goto out;
1174	}
1175
1176	pr_debug_ratelimited("Restoring PASID 0x%x queues\n",
1177			    pdd->process->pasid);
1178
1179	/* Update PD Base in QPD */
1180	qpd->page_table_base = pd_base;
1181	pr_debug("Updated PD address to 0x%llx\n", pd_base);
1182
1183	if (!list_empty(&qpd->queues_list)) {
1184		dqm->dev->kfd2kgd->set_vm_context_page_table_base(
1185				dqm->dev->adev,
1186				qpd->vmid,
1187				qpd->page_table_base);
1188		kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY);
1189	}
1190
1191	/* Take a safe reference to the mm_struct, which may otherwise
1192	 * disappear even while the kfd_process is still referenced.
1193	 */
1194	mm = get_task_mm(pdd->process->lead_thread);
1195	if (!mm) {
1196		ret = -EFAULT;
1197		goto out;
1198	}
1199
1200	/* Remove the eviction flags. Activate queues that are not
1201	 * inactive for other reasons.
1202	 */
1203	list_for_each_entry(q, &qpd->queues_list, list) {
1204		q->properties.is_evicted = false;
1205		if (!QUEUE_IS_ACTIVE(q->properties))
1206			continue;
1207
1208		mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1209				q->properties.type)];
1210		q->properties.is_active = true;
1211		increment_queue_count(dqm, qpd, q);
1212
1213		if (WARN_ONCE(!dqm->sched_running, "Restore when stopped\n"))
1214			continue;
1215
1216		retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe,
1217				       q->queue, &q->properties, mm);
1218		if (retval && !ret)
1219			/* Return the first error, but keep going to
1220			 * maintain a consistent eviction state
1221			 */
1222			ret = retval;
1223	}
1224	qpd->evicted = 0;
1225	eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp;
1226	atomic64_add(eviction_duration, &pdd->evict_duration_counter);
1227out:
1228	if (mm)
1229		mmput(mm);
1230	dqm_unlock(dqm);
1231	return ret;
1232}
1233
1234static int restore_process_queues_cpsch(struct device_queue_manager *dqm,
1235					struct qcm_process_device *qpd)
1236{
1237	struct queue *q;
1238	struct device *dev = dqm->dev->adev->dev;
1239	struct kfd_process_device *pdd;
 
1240	uint64_t eviction_duration;
1241	int retval = 0;
1242
1243	pdd = qpd_to_pdd(qpd);
 
 
1244
1245	dqm_lock(dqm);
1246	if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */
1247		goto out;
1248	if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
1249		qpd->evicted--;
1250		goto out;
1251	}
1252
1253	/* The debugger creates processes that temporarily have not acquired
1254	 * all VMs for all devices and has no VMs itself.
1255	 * Skip queue restore on process restore.
1256	 */
1257	if (!pdd->drm_priv)
1258		goto vm_not_acquired;
1259
1260	pr_debug_ratelimited("Restoring PASID 0x%x queues\n",
1261			    pdd->process->pasid);
1262
1263	/* Update PD Base in QPD */
1264	qpd->page_table_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv);
1265	pr_debug("Updated PD address to 0x%llx\n", qpd->page_table_base);
1266
1267	/* activate all active queues on the qpd */
1268	list_for_each_entry(q, &qpd->queues_list, list) {
1269		q->properties.is_evicted = false;
1270		if (!QUEUE_IS_ACTIVE(q->properties))
1271			continue;
1272
1273		q->properties.is_active = true;
1274		increment_queue_count(dqm, &pdd->qpd, q);
1275
1276		if (dqm->dev->kfd->shared_resources.enable_mes) {
1277			retval = add_queue_mes(dqm, q, qpd);
1278			if (retval) {
1279				dev_err(dev, "Failed to restore queue %d\n",
1280					q->properties.queue_id);
1281				goto out;
1282			}
1283		}
1284	}
1285	if (!dqm->dev->kfd->shared_resources.enable_mes)
1286		retval = execute_queues_cpsch(dqm,
1287					      KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD);
 
1288	eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp;
1289	atomic64_add(eviction_duration, &pdd->evict_duration_counter);
1290vm_not_acquired:
1291	qpd->evicted = 0;
1292out:
1293	dqm_unlock(dqm);
1294	return retval;
1295}
1296
1297static int register_process(struct device_queue_manager *dqm,
1298					struct qcm_process_device *qpd)
1299{
1300	struct device_process_node *n;
1301	struct kfd_process_device *pdd;
1302	uint64_t pd_base;
1303	int retval;
1304
1305	n = kzalloc(sizeof(*n), GFP_KERNEL);
1306	if (!n)
1307		return -ENOMEM;
1308
1309	n->qpd = qpd;
1310
1311	pdd = qpd_to_pdd(qpd);
1312	/* Retrieve PD base */
1313	pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv);
1314
1315	dqm_lock(dqm);
1316	list_add(&n->list, &dqm->queues);
1317
1318	/* Update PD Base in QPD */
1319	qpd->page_table_base = pd_base;
1320	pr_debug("Updated PD address to 0x%llx\n", pd_base);
1321
1322	retval = dqm->asic_ops.update_qpd(dqm, qpd);
1323
1324	dqm->processes_count++;
1325
1326	dqm_unlock(dqm);
1327
1328	/* Outside the DQM lock because under the DQM lock we can't do
1329	 * reclaim or take other locks that others hold while reclaiming.
1330	 */
1331	kfd_inc_compute_active(dqm->dev);
1332
1333	return retval;
1334}
1335
1336static int unregister_process(struct device_queue_manager *dqm,
1337					struct qcm_process_device *qpd)
1338{
1339	int retval;
1340	struct device_process_node *cur, *next;
1341
1342	pr_debug("qpd->queues_list is %s\n",
1343			list_empty(&qpd->queues_list) ? "empty" : "not empty");
1344
1345	retval = 0;
1346	dqm_lock(dqm);
1347
1348	list_for_each_entry_safe(cur, next, &dqm->queues, list) {
1349		if (qpd == cur->qpd) {
1350			list_del(&cur->list);
1351			kfree(cur);
1352			dqm->processes_count--;
1353			goto out;
1354		}
1355	}
1356	/* qpd not found in dqm list */
1357	retval = 1;
1358out:
1359	dqm_unlock(dqm);
1360
1361	/* Outside the DQM lock because under the DQM lock we can't do
1362	 * reclaim or take other locks that others hold while reclaiming.
1363	 */
1364	if (!retval)
1365		kfd_dec_compute_active(dqm->dev);
1366
1367	return retval;
1368}
1369
1370static int
1371set_pasid_vmid_mapping(struct device_queue_manager *dqm, u32 pasid,
1372			unsigned int vmid)
1373{
1374	uint32_t xcc_mask = dqm->dev->xcc_mask;
1375	int xcc_id, ret;
 
1376
1377	for_each_inst(xcc_id, xcc_mask) {
1378		ret = dqm->dev->kfd2kgd->set_pasid_vmid_mapping(
1379			dqm->dev->adev, pasid, vmid, xcc_id);
1380		if (ret)
1381			break;
1382	}
1383
1384	return ret;
 
 
1385}
1386
1387static void init_interrupts(struct device_queue_manager *dqm)
1388{
1389	uint32_t xcc_mask = dqm->dev->xcc_mask;
1390	unsigned int i, xcc_id;
 
 
 
 
 
 
 
 
 
1391
1392	for_each_inst(xcc_id, xcc_mask) {
1393		for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++) {
1394			if (is_pipe_enabled(dqm, 0, i)) {
1395				dqm->dev->kfd2kgd->init_interrupts(
1396					dqm->dev->adev, i, xcc_id);
1397			}
1398		}
1399	}
1400}
1401
1402static int initialize_nocpsch(struct device_queue_manager *dqm)
1403{
1404	int pipe, queue;
1405
1406	pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm));
1407
1408	dqm->allocated_queues = kcalloc(get_pipes_per_mec(dqm),
1409					sizeof(unsigned int), GFP_KERNEL);
1410	if (!dqm->allocated_queues)
1411		return -ENOMEM;
1412
1413	mutex_init(&dqm->lock_hidden);
1414	INIT_LIST_HEAD(&dqm->queues);
1415	dqm->active_queue_count = dqm->next_pipe_to_allocate = 0;
1416	dqm->active_cp_queue_count = 0;
1417	dqm->gws_queue_count = 0;
1418
1419	for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) {
1420		int pipe_offset = pipe * get_queues_per_pipe(dqm);
1421
1422		for (queue = 0; queue < get_queues_per_pipe(dqm); queue++)
1423			if (test_bit(pipe_offset + queue,
1424				     dqm->dev->kfd->shared_resources.cp_queue_bitmap))
1425				dqm->allocated_queues[pipe] |= 1 << queue;
1426	}
1427
1428	memset(dqm->vmid_pasid, 0, sizeof(dqm->vmid_pasid));
1429
1430	init_sdma_bitmaps(dqm);
1431
1432	return 0;
1433}
1434
1435static void uninitialize(struct device_queue_manager *dqm)
1436{
1437	int i;
1438
1439	WARN_ON(dqm->active_queue_count > 0 || dqm->processes_count > 0);
1440
1441	kfree(dqm->allocated_queues);
1442	for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++)
1443		kfree(dqm->mqd_mgrs[i]);
1444	mutex_destroy(&dqm->lock_hidden);
1445}
1446
1447static int start_nocpsch(struct device_queue_manager *dqm)
1448{
1449	int r = 0;
1450
1451	pr_info("SW scheduler is used");
1452	init_interrupts(dqm);
1453
1454	if (dqm->dev->adev->asic_type == CHIP_HAWAII)
1455		r = pm_init(&dqm->packet_mgr, dqm);
1456	if (!r)
1457		dqm->sched_running = true;
1458
1459	return r;
1460}
1461
1462static int stop_nocpsch(struct device_queue_manager *dqm)
1463{
1464	dqm_lock(dqm);
1465	if (!dqm->sched_running) {
1466		dqm_unlock(dqm);
1467		return 0;
1468	}
1469
1470	if (dqm->dev->adev->asic_type == CHIP_HAWAII)
1471		pm_uninit(&dqm->packet_mgr, false);
1472	dqm->sched_running = false;
1473	dqm_unlock(dqm);
1474
1475	return 0;
1476}
1477
1478static void pre_reset(struct device_queue_manager *dqm)
1479{
1480	dqm_lock(dqm);
1481	dqm->is_resetting = true;
1482	dqm_unlock(dqm);
1483}
1484
1485static int allocate_sdma_queue(struct device_queue_manager *dqm,
1486				struct queue *q, const uint32_t *restore_sdma_id)
1487{
1488	struct device *dev = dqm->dev->adev->dev;
1489	int bit;
1490
1491	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
1492		if (bitmap_empty(dqm->sdma_bitmap, KFD_MAX_SDMA_QUEUES)) {
1493			dev_err(dev, "No more SDMA queue to allocate\n");
1494			return -ENOMEM;
1495		}
1496
1497		if (restore_sdma_id) {
1498			/* Re-use existing sdma_id */
1499			if (!test_bit(*restore_sdma_id, dqm->sdma_bitmap)) {
1500				dev_err(dev, "SDMA queue already in use\n");
1501				return -EBUSY;
1502			}
1503			clear_bit(*restore_sdma_id, dqm->sdma_bitmap);
1504			q->sdma_id = *restore_sdma_id;
1505		} else {
1506			/* Find first available sdma_id */
1507			bit = find_first_bit(dqm->sdma_bitmap,
1508					     get_num_sdma_queues(dqm));
1509			clear_bit(bit, dqm->sdma_bitmap);
1510			q->sdma_id = bit;
1511		}
1512
1513		q->properties.sdma_engine_id =
1514			q->sdma_id % kfd_get_num_sdma_engines(dqm->dev);
1515		q->properties.sdma_queue_id = q->sdma_id /
1516				kfd_get_num_sdma_engines(dqm->dev);
1517	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1518		if (bitmap_empty(dqm->xgmi_sdma_bitmap, KFD_MAX_SDMA_QUEUES)) {
1519			dev_err(dev, "No more XGMI SDMA queue to allocate\n");
1520			return -ENOMEM;
1521		}
1522		if (restore_sdma_id) {
1523			/* Re-use existing sdma_id */
1524			if (!test_bit(*restore_sdma_id, dqm->xgmi_sdma_bitmap)) {
1525				dev_err(dev, "SDMA queue already in use\n");
1526				return -EBUSY;
1527			}
1528			clear_bit(*restore_sdma_id, dqm->xgmi_sdma_bitmap);
1529			q->sdma_id = *restore_sdma_id;
1530		} else {
1531			bit = find_first_bit(dqm->xgmi_sdma_bitmap,
1532					     get_num_xgmi_sdma_queues(dqm));
1533			clear_bit(bit, dqm->xgmi_sdma_bitmap);
1534			q->sdma_id = bit;
1535		}
1536		/* sdma_engine_id is sdma id including
1537		 * both PCIe-optimized SDMAs and XGMI-
1538		 * optimized SDMAs. The calculation below
1539		 * assumes the first N engines are always
1540		 * PCIe-optimized ones
1541		 */
1542		q->properties.sdma_engine_id =
1543			kfd_get_num_sdma_engines(dqm->dev) +
1544			q->sdma_id % kfd_get_num_xgmi_sdma_engines(dqm->dev);
1545		q->properties.sdma_queue_id = q->sdma_id /
1546			kfd_get_num_xgmi_sdma_engines(dqm->dev);
1547	}
1548
1549	pr_debug("SDMA engine id: %d\n", q->properties.sdma_engine_id);
1550	pr_debug("SDMA queue id: %d\n", q->properties.sdma_queue_id);
1551
1552	return 0;
1553}
1554
1555static void deallocate_sdma_queue(struct device_queue_manager *dqm,
1556				struct queue *q)
1557{
1558	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
1559		if (q->sdma_id >= get_num_sdma_queues(dqm))
1560			return;
1561		set_bit(q->sdma_id, dqm->sdma_bitmap);
1562	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1563		if (q->sdma_id >= get_num_xgmi_sdma_queues(dqm))
1564			return;
1565		set_bit(q->sdma_id, dqm->xgmi_sdma_bitmap);
1566	}
1567}
1568
1569/*
1570 * Device Queue Manager implementation for cp scheduler
1571 */
1572
1573static int set_sched_resources(struct device_queue_manager *dqm)
1574{
1575	int i, mec;
1576	struct scheduling_resources res;
1577	struct device *dev = dqm->dev->adev->dev;
1578
1579	res.vmid_mask = dqm->dev->compute_vmid_bitmap;
1580
1581	res.queue_mask = 0;
1582	for (i = 0; i < AMDGPU_MAX_QUEUES; ++i) {
1583		mec = (i / dqm->dev->kfd->shared_resources.num_queue_per_pipe)
1584			/ dqm->dev->kfd->shared_resources.num_pipe_per_mec;
1585
1586		if (!test_bit(i, dqm->dev->kfd->shared_resources.cp_queue_bitmap))
1587			continue;
1588
1589		/* only acquire queues from the first MEC */
1590		if (mec > 0)
1591			continue;
1592
1593		/* This situation may be hit in the future if a new HW
1594		 * generation exposes more than 64 queues. If so, the
1595		 * definition of res.queue_mask needs updating
1596		 */
1597		if (WARN_ON(i >= (sizeof(res.queue_mask)*8))) {
1598			dev_err(dev, "Invalid queue enabled by amdgpu: %d\n", i);
1599			break;
1600		}
1601
1602		res.queue_mask |= 1ull
1603			<< amdgpu_queue_mask_bit_to_set_resource_bit(
1604				dqm->dev->adev, i);
1605	}
1606	res.gws_mask = ~0ull;
1607	res.oac_mask = res.gds_heap_base = res.gds_heap_size = 0;
1608
1609	pr_debug("Scheduling resources:\n"
1610			"vmid mask: 0x%8X\n"
1611			"queue mask: 0x%8llX\n",
1612			res.vmid_mask, res.queue_mask);
1613
1614	return pm_send_set_resources(&dqm->packet_mgr, &res);
1615}
1616
1617static int initialize_cpsch(struct device_queue_manager *dqm)
1618{
1619	pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm));
1620
1621	mutex_init(&dqm->lock_hidden);
1622	INIT_LIST_HEAD(&dqm->queues);
1623	dqm->active_queue_count = dqm->processes_count = 0;
1624	dqm->active_cp_queue_count = 0;
1625	dqm->gws_queue_count = 0;
1626	dqm->active_runlist = false;
1627	INIT_WORK(&dqm->hw_exception_work, kfd_process_hw_exception);
1628	dqm->trap_debug_vmid = 0;
1629
1630	init_sdma_bitmaps(dqm);
1631
1632	if (dqm->dev->kfd2kgd->get_iq_wait_times)
1633		dqm->dev->kfd2kgd->get_iq_wait_times(dqm->dev->adev,
1634					&dqm->wait_times,
1635					ffs(dqm->dev->xcc_mask) - 1);
1636	return 0;
1637}
1638
1639static int start_cpsch(struct device_queue_manager *dqm)
1640{
1641	struct device *dev = dqm->dev->adev->dev;
1642	int retval;
1643
1644	retval = 0;
1645
1646	dqm_lock(dqm);
1647
1648	if (!dqm->dev->kfd->shared_resources.enable_mes) {
1649		retval = pm_init(&dqm->packet_mgr, dqm);
1650		if (retval)
1651			goto fail_packet_manager_init;
1652
1653		retval = set_sched_resources(dqm);
1654		if (retval)
1655			goto fail_set_sched_resources;
1656	}
1657	pr_debug("Allocating fence memory\n");
1658
1659	/* allocate fence memory on the gart */
1660	retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr),
1661					&dqm->fence_mem);
1662
1663	if (retval)
1664		goto fail_allocate_vidmem;
1665
1666	dqm->fence_addr = (uint64_t *)dqm->fence_mem->cpu_ptr;
1667	dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr;
1668
1669	init_interrupts(dqm);
1670
1671	/* clear hang status when driver try to start the hw scheduler */
1672	dqm->is_hws_hang = false;
1673	dqm->is_resetting = false;
1674	dqm->sched_running = true;
1675
1676	if (!dqm->dev->kfd->shared_resources.enable_mes)
1677		execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD);
1678
1679	/* Set CWSR grace period to 1x1000 cycle for GFX9.4.3 APU */
1680	if (amdgpu_emu_mode == 0 && dqm->dev->adev->gmc.is_app_apu &&
1681	    (KFD_GC_VERSION(dqm->dev) == IP_VERSION(9, 4, 3))) {
1682		uint32_t reg_offset = 0;
1683		uint32_t grace_period = 1;
1684
1685		retval = pm_update_grace_period(&dqm->packet_mgr,
1686						grace_period);
1687		if (retval)
1688			dev_err(dev, "Setting grace timeout failed\n");
1689		else if (dqm->dev->kfd2kgd->build_grace_period_packet_info)
1690			/* Update dqm->wait_times maintained in software */
1691			dqm->dev->kfd2kgd->build_grace_period_packet_info(
1692					dqm->dev->adev,	dqm->wait_times,
1693					grace_period, &reg_offset,
1694					&dqm->wait_times);
1695	}
1696
1697	dqm_unlock(dqm);
1698
1699	return 0;
1700fail_allocate_vidmem:
1701fail_set_sched_resources:
1702	if (!dqm->dev->kfd->shared_resources.enable_mes)
1703		pm_uninit(&dqm->packet_mgr, false);
1704fail_packet_manager_init:
1705	dqm_unlock(dqm);
1706	return retval;
1707}
1708
1709static int stop_cpsch(struct device_queue_manager *dqm)
1710{
1711	bool hanging;
1712
1713	dqm_lock(dqm);
1714	if (!dqm->sched_running) {
1715		dqm_unlock(dqm);
1716		return 0;
1717	}
1718
1719	if (!dqm->is_hws_hang) {
1720		if (!dqm->dev->kfd->shared_resources.enable_mes)
1721			unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD, false);
1722		else
1723			remove_all_queues_mes(dqm);
1724	}
1725
1726	hanging = dqm->is_hws_hang || dqm->is_resetting;
1727	dqm->sched_running = false;
1728
1729	if (!dqm->dev->kfd->shared_resources.enable_mes)
1730		pm_release_ib(&dqm->packet_mgr);
1731
1732	kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
1733	if (!dqm->dev->kfd->shared_resources.enable_mes)
1734		pm_uninit(&dqm->packet_mgr, hanging);
1735	dqm_unlock(dqm);
1736
1737	return 0;
1738}
1739
1740static int create_kernel_queue_cpsch(struct device_queue_manager *dqm,
1741					struct kernel_queue *kq,
1742					struct qcm_process_device *qpd)
1743{
1744	dqm_lock(dqm);
1745	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
1746		pr_warn("Can't create new kernel queue because %d queues were already created\n",
1747				dqm->total_queue_count);
1748		dqm_unlock(dqm);
1749		return -EPERM;
1750	}
1751
1752	/*
1753	 * Unconditionally increment this counter, regardless of the queue's
1754	 * type or whether the queue is active.
1755	 */
1756	dqm->total_queue_count++;
1757	pr_debug("Total of %d queues are accountable so far\n",
1758			dqm->total_queue_count);
1759
1760	list_add(&kq->list, &qpd->priv_queue_list);
1761	increment_queue_count(dqm, qpd, kq->queue);
1762	qpd->is_debug = true;
1763	execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0,
1764			USE_DEFAULT_GRACE_PERIOD);
1765	dqm_unlock(dqm);
1766
1767	return 0;
1768}
1769
1770static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm,
1771					struct kernel_queue *kq,
1772					struct qcm_process_device *qpd)
1773{
1774	dqm_lock(dqm);
1775	list_del(&kq->list);
1776	decrement_queue_count(dqm, qpd, kq->queue);
1777	qpd->is_debug = false;
1778	execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0,
1779			USE_DEFAULT_GRACE_PERIOD);
1780	/*
1781	 * Unconditionally decrement this counter, regardless of the queue's
1782	 * type.
1783	 */
1784	dqm->total_queue_count--;
1785	pr_debug("Total of %d queues are accountable so far\n",
1786			dqm->total_queue_count);
1787	dqm_unlock(dqm);
1788}
1789
1790static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
1791			struct qcm_process_device *qpd,
1792			const struct kfd_criu_queue_priv_data *qd,
1793			const void *restore_mqd, const void *restore_ctl_stack)
1794{
1795	int retval;
1796	struct mqd_manager *mqd_mgr;
1797
1798	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
1799		pr_warn("Can't create new usermode queue because %d queues were already created\n",
1800				dqm->total_queue_count);
1801		retval = -EPERM;
1802		goto out;
1803	}
1804
1805	if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1806		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1807		dqm_lock(dqm);
1808		retval = allocate_sdma_queue(dqm, q, qd ? &qd->sdma_id : NULL);
1809		dqm_unlock(dqm);
1810		if (retval)
1811			goto out;
1812	}
1813
1814	retval = allocate_doorbell(qpd, q, qd ? &qd->doorbell_id : NULL);
1815	if (retval)
1816		goto out_deallocate_sdma_queue;
1817
1818	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1819			q->properties.type)];
1820
1821	if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1822		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
1823		dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
1824	q->properties.tba_addr = qpd->tba_addr;
1825	q->properties.tma_addr = qpd->tma_addr;
1826	q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties);
1827	if (!q->mqd_mem_obj) {
1828		retval = -ENOMEM;
1829		goto out_deallocate_doorbell;
1830	}
1831
1832	dqm_lock(dqm);
1833	/*
1834	 * Eviction state logic: mark all queues as evicted, even ones
1835	 * not currently active. Restoring inactive queues later only
1836	 * updates the is_evicted flag but is a no-op otherwise.
1837	 */
1838	q->properties.is_evicted = !!qpd->evicted;
1839	q->properties.is_dbg_wa = qpd->pqm->process->debug_trap_enabled &&
1840				  kfd_dbg_has_cwsr_workaround(q->device);
1841
1842	if (qd)
1843		mqd_mgr->restore_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, &q->gart_mqd_addr,
1844				     &q->properties, restore_mqd, restore_ctl_stack,
1845				     qd->ctl_stack_size);
1846	else
1847		mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj,
1848					&q->gart_mqd_addr, &q->properties);
1849
1850	list_add(&q->list, &qpd->queues_list);
1851	qpd->queue_count++;
1852
1853	if (q->properties.is_active) {
1854		increment_queue_count(dqm, qpd, q);
1855
1856		if (!dqm->dev->kfd->shared_resources.enable_mes)
1857			retval = execute_queues_cpsch(dqm,
1858					KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, USE_DEFAULT_GRACE_PERIOD);
1859		else
1860			retval = add_queue_mes(dqm, q, qpd);
1861		if (retval)
1862			goto cleanup_queue;
1863	}
1864
1865	/*
1866	 * Unconditionally increment this counter, regardless of the queue's
1867	 * type or whether the queue is active.
1868	 */
1869	dqm->total_queue_count++;
1870
1871	pr_debug("Total of %d queues are accountable so far\n",
1872			dqm->total_queue_count);
1873
1874	dqm_unlock(dqm);
1875	return retval;
1876
1877cleanup_queue:
1878	qpd->queue_count--;
1879	list_del(&q->list);
1880	if (q->properties.is_active)
1881		decrement_queue_count(dqm, qpd, q);
1882	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
1883	dqm_unlock(dqm);
1884out_deallocate_doorbell:
1885	deallocate_doorbell(qpd, q);
1886out_deallocate_sdma_queue:
1887	if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1888		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1889		dqm_lock(dqm);
1890		deallocate_sdma_queue(dqm, q);
1891		dqm_unlock(dqm);
1892	}
1893out:
1894	return retval;
1895}
1896
1897int amdkfd_fence_wait_timeout(struct device_queue_manager *dqm,
1898			      uint64_t fence_value,
1899			      unsigned int timeout_ms)
1900{
1901	unsigned long end_jiffies = msecs_to_jiffies(timeout_ms) + jiffies;
1902	struct device *dev = dqm->dev->adev->dev;
1903	uint64_t *fence_addr =  dqm->fence_addr;
1904
1905	while (*fence_addr != fence_value) {
1906		/* Fatal err detected, this response won't come */
1907		if (amdgpu_amdkfd_is_fed(dqm->dev->adev))
1908			return -EIO;
1909
1910		if (time_after(jiffies, end_jiffies)) {
1911			dev_err(dev, "qcm fence wait loop timeout expired\n");
1912			/* In HWS case, this is used to halt the driver thread
1913			 * in order not to mess up CP states before doing
1914			 * scandumps for FW debugging.
1915			 */
1916			while (halt_if_hws_hang)
1917				schedule();
1918
1919			return -ETIME;
1920		}
1921		schedule();
1922	}
1923
1924	return 0;
1925}
1926
1927/* dqm->lock mutex has to be locked before calling this function */
1928static int map_queues_cpsch(struct device_queue_manager *dqm)
1929{
1930	struct device *dev = dqm->dev->adev->dev;
1931	int retval;
1932
1933	if (!dqm->sched_running)
1934		return 0;
1935	if (dqm->active_queue_count <= 0 || dqm->processes_count <= 0)
1936		return 0;
1937	if (dqm->active_runlist)
1938		return 0;
1939
1940	retval = pm_send_runlist(&dqm->packet_mgr, &dqm->queues);
1941	pr_debug("%s sent runlist\n", __func__);
1942	if (retval) {
1943		dev_err(dev, "failed to execute runlist\n");
1944		return retval;
1945	}
1946	dqm->active_runlist = true;
1947
1948	return retval;
1949}
1950
1951/* dqm->lock mutex has to be locked before calling this function */
1952static int unmap_queues_cpsch(struct device_queue_manager *dqm,
1953				enum kfd_unmap_queues_filter filter,
1954				uint32_t filter_param,
1955				uint32_t grace_period,
1956				bool reset)
1957{
1958	struct device *dev = dqm->dev->adev->dev;
1959	struct mqd_manager *mqd_mgr;
1960	int retval = 0;
 
1961
1962	if (!dqm->sched_running)
1963		return 0;
1964	if (dqm->is_hws_hang || dqm->is_resetting)
1965		return -EIO;
1966	if (!dqm->active_runlist)
1967		return retval;
1968
1969	if (grace_period != USE_DEFAULT_GRACE_PERIOD) {
1970		retval = pm_update_grace_period(&dqm->packet_mgr, grace_period);
1971		if (retval)
1972			return retval;
1973	}
1974
1975	retval = pm_send_unmap_queue(&dqm->packet_mgr, filter, filter_param, reset);
1976	if (retval)
1977		return retval;
1978
1979	*dqm->fence_addr = KFD_FENCE_INIT;
1980	pm_send_query_status(&dqm->packet_mgr, dqm->fence_gpu_addr,
1981				KFD_FENCE_COMPLETED);
1982	/* should be timed out */
1983	retval = amdkfd_fence_wait_timeout(dqm, KFD_FENCE_COMPLETED,
1984					   queue_preemption_timeout_ms);
1985	if (retval) {
1986		dev_err(dev, "The cp might be in an unrecoverable state due to an unsuccessful queues preemption\n");
1987		kfd_hws_hang(dqm);
1988		return retval;
1989	}
1990
1991	/* In the current MEC firmware implementation, if compute queue
1992	 * doesn't response to the preemption request in time, HIQ will
1993	 * abandon the unmap request without returning any timeout error
1994	 * to driver. Instead, MEC firmware will log the doorbell of the
1995	 * unresponding compute queue to HIQ.MQD.queue_doorbell_id fields.
1996	 * To make sure the queue unmap was successful, driver need to
1997	 * check those fields
1998	 */
1999	mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ];
2000	if (mqd_mgr->read_doorbell_id(dqm->packet_mgr.priv_queue->queue->mqd)) {
2001		dev_err(dev, "HIQ MQD's queue_doorbell_id0 is not 0, Queue preemption time out\n");
2002		while (halt_if_hws_hang)
2003			schedule();
2004		kfd_hws_hang(dqm);
2005		return -ETIME;
2006	}
2007
2008	/* We need to reset the grace period value for this device */
2009	if (grace_period != USE_DEFAULT_GRACE_PERIOD) {
2010		if (pm_update_grace_period(&dqm->packet_mgr,
2011					USE_DEFAULT_GRACE_PERIOD))
2012			dev_err(dev, "Failed to reset grace period\n");
2013	}
2014
2015	pm_release_ib(&dqm->packet_mgr);
2016	dqm->active_runlist = false;
2017
2018	return retval;
2019}
2020
2021/* only for compute queue */
2022static int reset_queues_cpsch(struct device_queue_manager *dqm,
2023			uint16_t pasid)
2024{
2025	int retval;
2026
2027	dqm_lock(dqm);
2028
2029	retval = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_BY_PASID,
2030			pasid, USE_DEFAULT_GRACE_PERIOD, true);
2031
2032	dqm_unlock(dqm);
2033	return retval;
2034}
2035
2036/* dqm->lock mutex has to be locked before calling this function */
2037static int execute_queues_cpsch(struct device_queue_manager *dqm,
2038				enum kfd_unmap_queues_filter filter,
2039				uint32_t filter_param,
2040				uint32_t grace_period)
2041{
2042	int retval;
2043
2044	if (dqm->is_hws_hang)
2045		return -EIO;
2046	retval = unmap_queues_cpsch(dqm, filter, filter_param, grace_period, false);
2047	if (retval)
2048		return retval;
2049
2050	return map_queues_cpsch(dqm);
2051}
2052
2053static int wait_on_destroy_queue(struct device_queue_manager *dqm,
2054				 struct queue *q)
2055{
2056	struct kfd_process_device *pdd = kfd_get_process_device_data(q->device,
2057								q->process);
2058	int ret = 0;
2059
2060	if (pdd->qpd.is_debug)
2061		return ret;
2062
2063	q->properties.is_being_destroyed = true;
2064
2065	if (pdd->process->debug_trap_enabled && q->properties.is_suspended) {
2066		dqm_unlock(dqm);
2067		mutex_unlock(&q->process->mutex);
2068		ret = wait_event_interruptible(dqm->destroy_wait,
2069						!q->properties.is_suspended);
2070
2071		mutex_lock(&q->process->mutex);
2072		dqm_lock(dqm);
2073	}
2074
2075	return ret;
2076}
2077
2078static int destroy_queue_cpsch(struct device_queue_manager *dqm,
2079				struct qcm_process_device *qpd,
2080				struct queue *q)
2081{
2082	int retval;
2083	struct mqd_manager *mqd_mgr;
2084	uint64_t sdma_val = 0;
2085	struct kfd_process_device *pdd = qpd_to_pdd(qpd);
2086	struct device *dev = dqm->dev->adev->dev;
2087
2088	/* Get the SDMA queue stats */
2089	if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) ||
2090	    (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
2091		retval = read_sdma_queue_counter((uint64_t __user *)q->properties.read_ptr,
2092							&sdma_val);
2093		if (retval)
2094			dev_err(dev, "Failed to read SDMA queue counter for queue: %d\n",
2095				q->properties.queue_id);
2096	}
2097
 
 
2098	/* remove queue from list to prevent rescheduling after preemption */
2099	dqm_lock(dqm);
2100
2101	retval = wait_on_destroy_queue(dqm, q);
2102
2103	if (retval) {
2104		dqm_unlock(dqm);
2105		return retval;
2106	}
2107
2108	if (qpd->is_debug) {
2109		/*
2110		 * error, currently we do not allow to destroy a queue
2111		 * of a currently debugged process
2112		 */
2113		retval = -EBUSY;
2114		goto failed_try_destroy_debugged_queue;
2115
2116	}
2117
2118	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
2119			q->properties.type)];
2120
2121	deallocate_doorbell(qpd, q);
2122
2123	if ((q->properties.type == KFD_QUEUE_TYPE_SDMA) ||
2124	    (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
2125		deallocate_sdma_queue(dqm, q);
2126		pdd->sdma_past_activity_counter += sdma_val;
2127	}
2128
2129	list_del(&q->list);
2130	qpd->queue_count--;
2131	if (q->properties.is_active) {
2132		decrement_queue_count(dqm, qpd, q);
2133		if (!dqm->dev->kfd->shared_resources.enable_mes) {
2134			retval = execute_queues_cpsch(dqm,
2135						      KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0,
2136						      USE_DEFAULT_GRACE_PERIOD);
2137			if (retval == -ETIME)
2138				qpd->reset_wavefronts = true;
2139		} else {
2140			retval = remove_queue_mes(dqm, q, qpd);
2141		}
2142	}
2143
2144	/*
2145	 * Unconditionally decrement this counter, regardless of the queue's
2146	 * type
2147	 */
2148	dqm->total_queue_count--;
2149	pr_debug("Total of %d queues are accountable so far\n",
2150			dqm->total_queue_count);
2151
2152	dqm_unlock(dqm);
2153
2154	/*
2155	 * Do free_mqd and raise delete event after dqm_unlock(dqm) to avoid
2156	 * circular locking
2157	 */
2158	kfd_dbg_ev_raise(KFD_EC_MASK(EC_DEVICE_QUEUE_DELETE),
2159				qpd->pqm->process, q->device,
2160				-1, false, NULL, 0);
2161
2162	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
2163
2164	return retval;
2165
2166failed_try_destroy_debugged_queue:
2167
2168	dqm_unlock(dqm);
2169	return retval;
2170}
2171
2172/*
2173 * Low bits must be 0000/FFFF as required by HW, high bits must be 0 to
2174 * stay in user mode.
2175 */
2176#define APE1_FIXED_BITS_MASK 0xFFFF80000000FFFFULL
2177/* APE1 limit is inclusive and 64K aligned. */
2178#define APE1_LIMIT_ALIGNMENT 0xFFFF
2179
2180static bool set_cache_memory_policy(struct device_queue_manager *dqm,
2181				   struct qcm_process_device *qpd,
2182				   enum cache_policy default_policy,
2183				   enum cache_policy alternate_policy,
2184				   void __user *alternate_aperture_base,
2185				   uint64_t alternate_aperture_size)
2186{
2187	bool retval = true;
2188
2189	if (!dqm->asic_ops.set_cache_memory_policy)
2190		return retval;
2191
2192	dqm_lock(dqm);
2193
2194	if (alternate_aperture_size == 0) {
2195		/* base > limit disables APE1 */
2196		qpd->sh_mem_ape1_base = 1;
2197		qpd->sh_mem_ape1_limit = 0;
2198	} else {
2199		/*
2200		 * In FSA64, APE1_Base[63:0] = { 16{SH_MEM_APE1_BASE[31]},
2201		 *			SH_MEM_APE1_BASE[31:0], 0x0000 }
2202		 * APE1_Limit[63:0] = { 16{SH_MEM_APE1_LIMIT[31]},
2203		 *			SH_MEM_APE1_LIMIT[31:0], 0xFFFF }
2204		 * Verify that the base and size parameters can be
2205		 * represented in this format and convert them.
2206		 * Additionally restrict APE1 to user-mode addresses.
2207		 */
2208
2209		uint64_t base = (uintptr_t)alternate_aperture_base;
2210		uint64_t limit = base + alternate_aperture_size - 1;
2211
2212		if (limit <= base || (base & APE1_FIXED_BITS_MASK) != 0 ||
2213		   (limit & APE1_FIXED_BITS_MASK) != APE1_LIMIT_ALIGNMENT) {
2214			retval = false;
2215			goto out;
2216		}
2217
2218		qpd->sh_mem_ape1_base = base >> 16;
2219		qpd->sh_mem_ape1_limit = limit >> 16;
2220	}
2221
2222	retval = dqm->asic_ops.set_cache_memory_policy(
2223			dqm,
2224			qpd,
2225			default_policy,
2226			alternate_policy,
2227			alternate_aperture_base,
2228			alternate_aperture_size);
2229
2230	if ((dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0))
2231		program_sh_mem_settings(dqm, qpd);
2232
2233	pr_debug("sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n",
2234		qpd->sh_mem_config, qpd->sh_mem_ape1_base,
2235		qpd->sh_mem_ape1_limit);
2236
2237out:
2238	dqm_unlock(dqm);
2239	return retval;
2240}
2241
2242static int process_termination_nocpsch(struct device_queue_manager *dqm,
2243		struct qcm_process_device *qpd)
2244{
2245	struct queue *q;
2246	struct device_process_node *cur, *next_dpn;
2247	int retval = 0;
2248	bool found = false;
2249
2250	dqm_lock(dqm);
2251
2252	/* Clear all user mode queues */
2253	while (!list_empty(&qpd->queues_list)) {
2254		struct mqd_manager *mqd_mgr;
2255		int ret;
2256
2257		q = list_first_entry(&qpd->queues_list, struct queue, list);
2258		mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
2259				q->properties.type)];
2260		ret = destroy_queue_nocpsch_locked(dqm, qpd, q);
2261		if (ret)
2262			retval = ret;
2263		dqm_unlock(dqm);
2264		mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
2265		dqm_lock(dqm);
2266	}
2267
2268	/* Unregister process */
2269	list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
2270		if (qpd == cur->qpd) {
2271			list_del(&cur->list);
2272			kfree(cur);
2273			dqm->processes_count--;
2274			found = true;
2275			break;
2276		}
2277	}
2278
2279	dqm_unlock(dqm);
2280
2281	/* Outside the DQM lock because under the DQM lock we can't do
2282	 * reclaim or take other locks that others hold while reclaiming.
2283	 */
2284	if (found)
2285		kfd_dec_compute_active(dqm->dev);
2286
2287	return retval;
2288}
2289
2290static int get_wave_state(struct device_queue_manager *dqm,
2291			  struct queue *q,
2292			  void __user *ctl_stack,
2293			  u32 *ctl_stack_used_size,
2294			  u32 *save_area_used_size)
2295{
2296	struct mqd_manager *mqd_mgr;
2297
2298	dqm_lock(dqm);
2299
2300	mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_CP];
2301
2302	if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE ||
2303	    q->properties.is_active || !q->device->kfd->cwsr_enabled ||
2304	    !mqd_mgr->get_wave_state) {
2305		dqm_unlock(dqm);
2306		return -EINVAL;
2307	}
2308
2309	dqm_unlock(dqm);
2310
2311	/*
2312	 * get_wave_state is outside the dqm lock to prevent circular locking
2313	 * and the queue should be protected against destruction by the process
2314	 * lock.
2315	 */
2316	return mqd_mgr->get_wave_state(mqd_mgr, q->mqd, &q->properties,
2317			ctl_stack, ctl_stack_used_size, save_area_used_size);
2318}
2319
2320static void get_queue_checkpoint_info(struct device_queue_manager *dqm,
2321			const struct queue *q,
2322			u32 *mqd_size,
2323			u32 *ctl_stack_size)
2324{
2325	struct mqd_manager *mqd_mgr;
2326	enum KFD_MQD_TYPE mqd_type =
2327			get_mqd_type_from_queue_type(q->properties.type);
2328
2329	dqm_lock(dqm);
2330	mqd_mgr = dqm->mqd_mgrs[mqd_type];
2331	*mqd_size = mqd_mgr->mqd_size;
2332	*ctl_stack_size = 0;
2333
2334	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE && mqd_mgr->get_checkpoint_info)
2335		mqd_mgr->get_checkpoint_info(mqd_mgr, q->mqd, ctl_stack_size);
2336
2337	dqm_unlock(dqm);
2338}
2339
2340static int checkpoint_mqd(struct device_queue_manager *dqm,
2341			  const struct queue *q,
2342			  void *mqd,
2343			  void *ctl_stack)
2344{
2345	struct mqd_manager *mqd_mgr;
2346	int r = 0;
2347	enum KFD_MQD_TYPE mqd_type =
2348			get_mqd_type_from_queue_type(q->properties.type);
2349
2350	dqm_lock(dqm);
2351
2352	if (q->properties.is_active || !q->device->kfd->cwsr_enabled) {
2353		r = -EINVAL;
2354		goto dqm_unlock;
2355	}
2356
2357	mqd_mgr = dqm->mqd_mgrs[mqd_type];
2358	if (!mqd_mgr->checkpoint_mqd) {
2359		r = -EOPNOTSUPP;
2360		goto dqm_unlock;
2361	}
2362
2363	mqd_mgr->checkpoint_mqd(mqd_mgr, q->mqd, mqd, ctl_stack);
2364
2365dqm_unlock:
2366	dqm_unlock(dqm);
2367	return r;
2368}
2369
2370static int process_termination_cpsch(struct device_queue_manager *dqm,
2371		struct qcm_process_device *qpd)
2372{
2373	int retval;
2374	struct queue *q;
2375	struct device *dev = dqm->dev->adev->dev;
2376	struct kernel_queue *kq, *kq_next;
2377	struct mqd_manager *mqd_mgr;
2378	struct device_process_node *cur, *next_dpn;
2379	enum kfd_unmap_queues_filter filter =
2380		KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES;
2381	bool found = false;
2382
2383	retval = 0;
2384
2385	dqm_lock(dqm);
2386
2387	/* Clean all kernel queues */
2388	list_for_each_entry_safe(kq, kq_next, &qpd->priv_queue_list, list) {
2389		list_del(&kq->list);
2390		decrement_queue_count(dqm, qpd, kq->queue);
2391		qpd->is_debug = false;
2392		dqm->total_queue_count--;
2393		filter = KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES;
2394	}
2395
2396	/* Clear all user mode queues */
2397	list_for_each_entry(q, &qpd->queues_list, list) {
2398		if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
2399			deallocate_sdma_queue(dqm, q);
2400		else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
2401			deallocate_sdma_queue(dqm, q);
2402
2403		if (q->properties.is_active) {
2404			decrement_queue_count(dqm, qpd, q);
2405
2406			if (dqm->dev->kfd->shared_resources.enable_mes) {
2407				retval = remove_queue_mes(dqm, q, qpd);
2408				if (retval)
2409					dev_err(dev, "Failed to remove queue %d\n",
2410						q->properties.queue_id);
2411			}
2412		}
2413
2414		dqm->total_queue_count--;
2415	}
2416
2417	/* Unregister process */
2418	list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
2419		if (qpd == cur->qpd) {
2420			list_del(&cur->list);
2421			kfree(cur);
2422			dqm->processes_count--;
2423			found = true;
2424			break;
2425		}
2426	}
2427
2428	if (!dqm->dev->kfd->shared_resources.enable_mes)
2429		retval = execute_queues_cpsch(dqm, filter, 0, USE_DEFAULT_GRACE_PERIOD);
2430
2431	if ((!dqm->is_hws_hang) && (retval || qpd->reset_wavefronts)) {
2432		pr_warn("Resetting wave fronts (cpsch) on dev %p\n", dqm->dev);
2433		dbgdev_wave_reset_wavefronts(dqm->dev, qpd->pqm->process);
2434		qpd->reset_wavefronts = false;
2435	}
2436
2437	/* Lastly, free mqd resources.
2438	 * Do free_mqd() after dqm_unlock to avoid circular locking.
2439	 */
2440	while (!list_empty(&qpd->queues_list)) {
2441		q = list_first_entry(&qpd->queues_list, struct queue, list);
2442		mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
2443				q->properties.type)];
2444		list_del(&q->list);
2445		qpd->queue_count--;
2446		dqm_unlock(dqm);
2447		mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
2448		dqm_lock(dqm);
2449	}
2450	dqm_unlock(dqm);
2451
2452	/* Outside the DQM lock because under the DQM lock we can't do
2453	 * reclaim or take other locks that others hold while reclaiming.
2454	 */
2455	if (found)
2456		kfd_dec_compute_active(dqm->dev);
2457
2458	return retval;
2459}
2460
2461static int init_mqd_managers(struct device_queue_manager *dqm)
2462{
2463	int i, j;
2464	struct device *dev = dqm->dev->adev->dev;
2465	struct mqd_manager *mqd_mgr;
2466
2467	for (i = 0; i < KFD_MQD_TYPE_MAX; i++) {
2468		mqd_mgr = dqm->asic_ops.mqd_manager_init(i, dqm->dev);
2469		if (!mqd_mgr) {
2470			dev_err(dev, "mqd manager [%d] initialization failed\n", i);
2471			goto out_free;
2472		}
2473		dqm->mqd_mgrs[i] = mqd_mgr;
2474	}
2475
2476	return 0;
2477
2478out_free:
2479	for (j = 0; j < i; j++) {
2480		kfree(dqm->mqd_mgrs[j]);
2481		dqm->mqd_mgrs[j] = NULL;
2482	}
2483
2484	return -ENOMEM;
2485}
2486
2487/* Allocate one hiq mqd (HWS) and all SDMA mqd in a continuous trunk*/
2488static int allocate_hiq_sdma_mqd(struct device_queue_manager *dqm)
2489{
2490	int retval;
2491	struct kfd_node *dev = dqm->dev;
2492	struct kfd_mem_obj *mem_obj = &dqm->hiq_sdma_mqd;
2493	uint32_t size = dqm->mqd_mgrs[KFD_MQD_TYPE_SDMA]->mqd_size *
2494		get_num_all_sdma_engines(dqm) *
2495		dev->kfd->device_info.num_sdma_queues_per_engine +
2496		(dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size *
2497		NUM_XCC(dqm->dev->xcc_mask));
2498
2499	retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev, size,
2500		&(mem_obj->gtt_mem), &(mem_obj->gpu_addr),
2501		(void *)&(mem_obj->cpu_ptr), false);
2502
2503	return retval;
2504}
2505
2506struct device_queue_manager *device_queue_manager_init(struct kfd_node *dev)
2507{
2508	struct device_queue_manager *dqm;
2509
2510	pr_debug("Loading device queue manager\n");
2511
2512	dqm = kzalloc(sizeof(*dqm), GFP_KERNEL);
2513	if (!dqm)
2514		return NULL;
2515
2516	switch (dev->adev->asic_type) {
2517	/* HWS is not available on Hawaii. */
2518	case CHIP_HAWAII:
2519	/* HWS depends on CWSR for timely dequeue. CWSR is not
2520	 * available on Tonga.
2521	 *
2522	 * FIXME: This argument also applies to Kaveri.
2523	 */
2524	case CHIP_TONGA:
2525		dqm->sched_policy = KFD_SCHED_POLICY_NO_HWS;
2526		break;
2527	default:
2528		dqm->sched_policy = sched_policy;
2529		break;
2530	}
2531
2532	dqm->dev = dev;
2533	switch (dqm->sched_policy) {
2534	case KFD_SCHED_POLICY_HWS:
2535	case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION:
2536		/* initialize dqm for cp scheduling */
2537		dqm->ops.create_queue = create_queue_cpsch;
2538		dqm->ops.initialize = initialize_cpsch;
2539		dqm->ops.start = start_cpsch;
2540		dqm->ops.stop = stop_cpsch;
2541		dqm->ops.pre_reset = pre_reset;
2542		dqm->ops.destroy_queue = destroy_queue_cpsch;
2543		dqm->ops.update_queue = update_queue;
2544		dqm->ops.register_process = register_process;
2545		dqm->ops.unregister_process = unregister_process;
2546		dqm->ops.uninitialize = uninitialize;
2547		dqm->ops.create_kernel_queue = create_kernel_queue_cpsch;
2548		dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch;
2549		dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
2550		dqm->ops.process_termination = process_termination_cpsch;
2551		dqm->ops.evict_process_queues = evict_process_queues_cpsch;
2552		dqm->ops.restore_process_queues = restore_process_queues_cpsch;
2553		dqm->ops.get_wave_state = get_wave_state;
2554		dqm->ops.reset_queues = reset_queues_cpsch;
2555		dqm->ops.get_queue_checkpoint_info = get_queue_checkpoint_info;
2556		dqm->ops.checkpoint_mqd = checkpoint_mqd;
2557		break;
2558	case KFD_SCHED_POLICY_NO_HWS:
2559		/* initialize dqm for no cp scheduling */
2560		dqm->ops.start = start_nocpsch;
2561		dqm->ops.stop = stop_nocpsch;
2562		dqm->ops.pre_reset = pre_reset;
2563		dqm->ops.create_queue = create_queue_nocpsch;
2564		dqm->ops.destroy_queue = destroy_queue_nocpsch;
2565		dqm->ops.update_queue = update_queue;
2566		dqm->ops.register_process = register_process;
2567		dqm->ops.unregister_process = unregister_process;
2568		dqm->ops.initialize = initialize_nocpsch;
2569		dqm->ops.uninitialize = uninitialize;
2570		dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
2571		dqm->ops.process_termination = process_termination_nocpsch;
2572		dqm->ops.evict_process_queues = evict_process_queues_nocpsch;
2573		dqm->ops.restore_process_queues =
2574			restore_process_queues_nocpsch;
2575		dqm->ops.get_wave_state = get_wave_state;
2576		dqm->ops.get_queue_checkpoint_info = get_queue_checkpoint_info;
2577		dqm->ops.checkpoint_mqd = checkpoint_mqd;
2578		break;
2579	default:
2580		dev_err(dev->adev->dev, "Invalid scheduling policy %d\n", dqm->sched_policy);
2581		goto out_free;
2582	}
2583
2584	switch (dev->adev->asic_type) {
 
 
 
 
2585	case CHIP_KAVERI:
2586	case CHIP_HAWAII:
2587		device_queue_manager_init_cik(&dqm->asic_ops);
2588		break;
2589
2590	case CHIP_CARRIZO:
 
 
 
2591	case CHIP_TONGA:
2592	case CHIP_FIJI:
2593	case CHIP_POLARIS10:
2594	case CHIP_POLARIS11:
2595	case CHIP_POLARIS12:
2596	case CHIP_VEGAM:
2597		device_queue_manager_init_vi(&dqm->asic_ops);
2598		break;
2599
2600	default:
2601		if (KFD_GC_VERSION(dev) >= IP_VERSION(11, 0, 0))
2602			device_queue_manager_init_v11(&dqm->asic_ops);
2603		else if (KFD_GC_VERSION(dev) >= IP_VERSION(10, 1, 1))
2604			device_queue_manager_init_v10(&dqm->asic_ops);
2605		else if (KFD_GC_VERSION(dev) >= IP_VERSION(9, 0, 1))
2606			device_queue_manager_init_v9(&dqm->asic_ops);
2607		else {
2608			WARN(1, "Unexpected ASIC family %u",
2609			     dev->adev->asic_type);
2610			goto out_free;
2611		}
2612	}
2613
2614	if (init_mqd_managers(dqm))
2615		goto out_free;
2616
2617	if (!dev->kfd->shared_resources.enable_mes && allocate_hiq_sdma_mqd(dqm)) {
2618		dev_err(dev->adev->dev, "Failed to allocate hiq sdma mqd trunk buffer\n");
2619		goto out_free;
2620	}
2621
2622	if (!dqm->ops.initialize(dqm)) {
2623		init_waitqueue_head(&dqm->destroy_wait);
2624		return dqm;
2625	}
2626
2627out_free:
2628	kfree(dqm);
2629	return NULL;
2630}
2631
2632static void deallocate_hiq_sdma_mqd(struct kfd_node *dev,
2633				    struct kfd_mem_obj *mqd)
2634{
2635	WARN(!mqd, "No hiq sdma mqd trunk to free");
2636
2637	amdgpu_amdkfd_free_gtt_mem(dev->adev, mqd->gtt_mem);
2638}
2639
2640void device_queue_manager_uninit(struct device_queue_manager *dqm)
2641{
2642	dqm->ops.stop(dqm);
2643	dqm->ops.uninitialize(dqm);
2644	if (!dqm->dev->kfd->shared_resources.enable_mes)
2645		deallocate_hiq_sdma_mqd(dqm->dev, &dqm->hiq_sdma_mqd);
2646	kfree(dqm);
2647}
2648
2649int kfd_dqm_evict_pasid(struct device_queue_manager *dqm, u32 pasid)
2650{
2651	struct kfd_process_device *pdd;
2652	struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
2653	int ret = 0;
2654
2655	if (!p)
2656		return -EINVAL;
2657	WARN(debug_evictions, "Evicting pid %d", p->lead_thread->pid);
2658	pdd = kfd_get_process_device_data(dqm->dev, p);
2659	if (pdd)
2660		ret = dqm->ops.evict_process_queues(dqm, &pdd->qpd);
2661	kfd_unref_process(p);
2662
2663	return ret;
2664}
2665
2666static void kfd_process_hw_exception(struct work_struct *work)
2667{
2668	struct device_queue_manager *dqm = container_of(work,
2669			struct device_queue_manager, hw_exception_work);
2670	amdgpu_amdkfd_gpu_reset(dqm->dev->adev);
2671}
2672
2673int reserve_debug_trap_vmid(struct device_queue_manager *dqm,
2674				struct qcm_process_device *qpd)
2675{
2676	int r;
2677	struct device *dev = dqm->dev->adev->dev;
2678	int updated_vmid_mask;
2679
2680	if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
2681		dev_err(dev, "Unsupported on sched_policy: %i\n", dqm->sched_policy);
2682		return -EINVAL;
2683	}
2684
2685	dqm_lock(dqm);
2686
2687	if (dqm->trap_debug_vmid != 0) {
2688		dev_err(dev, "Trap debug id already reserved\n");
2689		r = -EBUSY;
2690		goto out_unlock;
2691	}
2692
2693	r = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0,
2694			USE_DEFAULT_GRACE_PERIOD, false);
2695	if (r)
2696		goto out_unlock;
2697
2698	updated_vmid_mask = dqm->dev->kfd->shared_resources.compute_vmid_bitmap;
2699	updated_vmid_mask &= ~(1 << dqm->dev->vm_info.last_vmid_kfd);
2700
2701	dqm->dev->kfd->shared_resources.compute_vmid_bitmap = updated_vmid_mask;
2702	dqm->trap_debug_vmid = dqm->dev->vm_info.last_vmid_kfd;
2703	r = set_sched_resources(dqm);
2704	if (r)
2705		goto out_unlock;
2706
2707	r = map_queues_cpsch(dqm);
2708	if (r)
2709		goto out_unlock;
2710
2711	pr_debug("Reserved VMID for trap debug: %i\n", dqm->trap_debug_vmid);
2712
2713out_unlock:
2714	dqm_unlock(dqm);
2715	return r;
2716}
2717
2718/*
2719 * Releases vmid for the trap debugger
2720 */
2721int release_debug_trap_vmid(struct device_queue_manager *dqm,
2722			struct qcm_process_device *qpd)
2723{
2724	struct device *dev = dqm->dev->adev->dev;
2725	int r;
2726	int updated_vmid_mask;
2727	uint32_t trap_debug_vmid;
2728
2729	if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
2730		dev_err(dev, "Unsupported on sched_policy: %i\n", dqm->sched_policy);
2731		return -EINVAL;
2732	}
2733
2734	dqm_lock(dqm);
2735	trap_debug_vmid = dqm->trap_debug_vmid;
2736	if (dqm->trap_debug_vmid == 0) {
2737		dev_err(dev, "Trap debug id is not reserved\n");
2738		r = -EINVAL;
2739		goto out_unlock;
2740	}
2741
2742	r = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0,
2743			USE_DEFAULT_GRACE_PERIOD, false);
2744	if (r)
2745		goto out_unlock;
2746
2747	updated_vmid_mask = dqm->dev->kfd->shared_resources.compute_vmid_bitmap;
2748	updated_vmid_mask |= (1 << dqm->dev->vm_info.last_vmid_kfd);
2749
2750	dqm->dev->kfd->shared_resources.compute_vmid_bitmap = updated_vmid_mask;
2751	dqm->trap_debug_vmid = 0;
2752	r = set_sched_resources(dqm);
2753	if (r)
2754		goto out_unlock;
2755
2756	r = map_queues_cpsch(dqm);
2757	if (r)
2758		goto out_unlock;
2759
2760	pr_debug("Released VMID for trap debug: %i\n", trap_debug_vmid);
2761
2762out_unlock:
2763	dqm_unlock(dqm);
2764	return r;
2765}
2766
2767#define QUEUE_NOT_FOUND		-1
2768/* invalidate queue operation in array */
2769static void q_array_invalidate(uint32_t num_queues, uint32_t *queue_ids)
2770{
2771	int i;
2772
2773	for (i = 0; i < num_queues; i++)
2774		queue_ids[i] |= KFD_DBG_QUEUE_INVALID_MASK;
2775}
2776
2777/* find queue index in array */
2778static int q_array_get_index(unsigned int queue_id,
2779		uint32_t num_queues,
2780		uint32_t *queue_ids)
2781{
2782	int i;
2783
2784	for (i = 0; i < num_queues; i++)
2785		if (queue_id == (queue_ids[i] & ~KFD_DBG_QUEUE_INVALID_MASK))
2786			return i;
2787
2788	return QUEUE_NOT_FOUND;
2789}
2790
2791struct copy_context_work_handler_workarea {
2792	struct work_struct copy_context_work;
2793	struct kfd_process *p;
2794};
2795
2796static void copy_context_work_handler (struct work_struct *work)
2797{
2798	struct copy_context_work_handler_workarea *workarea;
2799	struct mqd_manager *mqd_mgr;
2800	struct queue *q;
2801	struct mm_struct *mm;
2802	struct kfd_process *p;
2803	uint32_t tmp_ctl_stack_used_size, tmp_save_area_used_size;
2804	int i;
2805
2806	workarea = container_of(work,
2807			struct copy_context_work_handler_workarea,
2808			copy_context_work);
2809
2810	p = workarea->p;
2811	mm = get_task_mm(p->lead_thread);
2812
2813	if (!mm)
2814		return;
2815
2816	kthread_use_mm(mm);
2817	for (i = 0; i < p->n_pdds; i++) {
2818		struct kfd_process_device *pdd = p->pdds[i];
2819		struct device_queue_manager *dqm = pdd->dev->dqm;
2820		struct qcm_process_device *qpd = &pdd->qpd;
2821
2822		list_for_each_entry(q, &qpd->queues_list, list) {
2823			mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_CP];
2824
2825			/* We ignore the return value from get_wave_state
2826			 * because
2827			 * i) right now, it always returns 0, and
2828			 * ii) if we hit an error, we would continue to the
2829			 *      next queue anyway.
2830			 */
2831			mqd_mgr->get_wave_state(mqd_mgr,
2832					q->mqd,
2833					&q->properties,
2834					(void __user *)	q->properties.ctx_save_restore_area_address,
2835					&tmp_ctl_stack_used_size,
2836					&tmp_save_area_used_size);
2837		}
2838	}
2839	kthread_unuse_mm(mm);
2840	mmput(mm);
2841}
2842
2843static uint32_t *get_queue_ids(uint32_t num_queues, uint32_t *usr_queue_id_array)
2844{
2845	size_t array_size = num_queues * sizeof(uint32_t);
2846
2847	if (!usr_queue_id_array)
2848		return NULL;
2849
2850	return memdup_user(usr_queue_id_array, array_size);
2851}
2852
2853int resume_queues(struct kfd_process *p,
2854		uint32_t num_queues,
2855		uint32_t *usr_queue_id_array)
2856{
2857	uint32_t *queue_ids = NULL;
2858	int total_resumed = 0;
2859	int i;
2860
2861	if (usr_queue_id_array) {
2862		queue_ids = get_queue_ids(num_queues, usr_queue_id_array);
2863
2864		if (IS_ERR(queue_ids))
2865			return PTR_ERR(queue_ids);
2866
2867		/* mask all queues as invalid.  unmask per successful request */
2868		q_array_invalidate(num_queues, queue_ids);
2869	}
2870
2871	for (i = 0; i < p->n_pdds; i++) {
2872		struct kfd_process_device *pdd = p->pdds[i];
2873		struct device_queue_manager *dqm = pdd->dev->dqm;
2874		struct device *dev = dqm->dev->adev->dev;
2875		struct qcm_process_device *qpd = &pdd->qpd;
2876		struct queue *q;
2877		int r, per_device_resumed = 0;
2878
2879		dqm_lock(dqm);
2880
2881		/* unmask queues that resume or already resumed as valid */
2882		list_for_each_entry(q, &qpd->queues_list, list) {
2883			int q_idx = QUEUE_NOT_FOUND;
2884
2885			if (queue_ids)
2886				q_idx = q_array_get_index(
2887						q->properties.queue_id,
2888						num_queues,
2889						queue_ids);
2890
2891			if (!queue_ids || q_idx != QUEUE_NOT_FOUND) {
2892				int err = resume_single_queue(dqm, &pdd->qpd, q);
2893
2894				if (queue_ids) {
2895					if (!err) {
2896						queue_ids[q_idx] &=
2897							~KFD_DBG_QUEUE_INVALID_MASK;
2898					} else {
2899						queue_ids[q_idx] |=
2900							KFD_DBG_QUEUE_ERROR_MASK;
2901						break;
2902					}
2903				}
2904
2905				if (dqm->dev->kfd->shared_resources.enable_mes) {
2906					wake_up_all(&dqm->destroy_wait);
2907					if (!err)
2908						total_resumed++;
2909				} else {
2910					per_device_resumed++;
2911				}
2912			}
2913		}
2914
2915		if (!per_device_resumed) {
2916			dqm_unlock(dqm);
2917			continue;
2918		}
2919
2920		r = execute_queues_cpsch(dqm,
2921					KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES,
2922					0,
2923					USE_DEFAULT_GRACE_PERIOD);
2924		if (r) {
2925			dev_err(dev, "Failed to resume process queues\n");
2926			if (queue_ids) {
2927				list_for_each_entry(q, &qpd->queues_list, list) {
2928					int q_idx = q_array_get_index(
2929							q->properties.queue_id,
2930							num_queues,
2931							queue_ids);
2932
2933					/* mask queue as error on resume fail */
2934					if (q_idx != QUEUE_NOT_FOUND)
2935						queue_ids[q_idx] |=
2936							KFD_DBG_QUEUE_ERROR_MASK;
2937				}
2938			}
2939		} else {
2940			wake_up_all(&dqm->destroy_wait);
2941			total_resumed += per_device_resumed;
2942		}
2943
2944		dqm_unlock(dqm);
2945	}
2946
2947	if (queue_ids) {
2948		if (copy_to_user((void __user *)usr_queue_id_array, queue_ids,
2949				num_queues * sizeof(uint32_t)))
2950			pr_err("copy_to_user failed on queue resume\n");
2951
2952		kfree(queue_ids);
2953	}
2954
2955	return total_resumed;
2956}
2957
2958int suspend_queues(struct kfd_process *p,
2959			uint32_t num_queues,
2960			uint32_t grace_period,
2961			uint64_t exception_clear_mask,
2962			uint32_t *usr_queue_id_array)
2963{
2964	uint32_t *queue_ids = get_queue_ids(num_queues, usr_queue_id_array);
2965	int total_suspended = 0;
2966	int i;
2967
2968	if (IS_ERR(queue_ids))
2969		return PTR_ERR(queue_ids);
2970
2971	/* mask all queues as invalid.  umask on successful request */
2972	q_array_invalidate(num_queues, queue_ids);
2973
2974	for (i = 0; i < p->n_pdds; i++) {
2975		struct kfd_process_device *pdd = p->pdds[i];
2976		struct device_queue_manager *dqm = pdd->dev->dqm;
2977		struct device *dev = dqm->dev->adev->dev;
2978		struct qcm_process_device *qpd = &pdd->qpd;
2979		struct queue *q;
2980		int r, per_device_suspended = 0;
2981
2982		mutex_lock(&p->event_mutex);
2983		dqm_lock(dqm);
2984
2985		/* unmask queues that suspend or already suspended */
2986		list_for_each_entry(q, &qpd->queues_list, list) {
2987			int q_idx = q_array_get_index(q->properties.queue_id,
2988							num_queues,
2989							queue_ids);
2990
2991			if (q_idx != QUEUE_NOT_FOUND) {
2992				int err = suspend_single_queue(dqm, pdd, q);
2993				bool is_mes = dqm->dev->kfd->shared_resources.enable_mes;
2994
2995				if (!err) {
2996					queue_ids[q_idx] &= ~KFD_DBG_QUEUE_INVALID_MASK;
2997					if (exception_clear_mask && is_mes)
2998						q->properties.exception_status &=
2999							~exception_clear_mask;
3000
3001					if (is_mes)
3002						total_suspended++;
3003					else
3004						per_device_suspended++;
3005				} else if (err != -EBUSY) {
3006					r = err;
3007					queue_ids[q_idx] |= KFD_DBG_QUEUE_ERROR_MASK;
3008					break;
3009				}
3010			}
3011		}
3012
3013		if (!per_device_suspended) {
3014			dqm_unlock(dqm);
3015			mutex_unlock(&p->event_mutex);
3016			if (total_suspended)
3017				amdgpu_amdkfd_debug_mem_fence(dqm->dev->adev);
3018			continue;
3019		}
3020
3021		r = execute_queues_cpsch(dqm,
3022			KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0,
3023			grace_period);
3024
3025		if (r)
3026			dev_err(dev, "Failed to suspend process queues.\n");
3027		else
3028			total_suspended += per_device_suspended;
3029
3030		list_for_each_entry(q, &qpd->queues_list, list) {
3031			int q_idx = q_array_get_index(q->properties.queue_id,
3032						num_queues, queue_ids);
3033
3034			if (q_idx == QUEUE_NOT_FOUND)
3035				continue;
3036
3037			/* mask queue as error on suspend fail */
3038			if (r)
3039				queue_ids[q_idx] |= KFD_DBG_QUEUE_ERROR_MASK;
3040			else if (exception_clear_mask)
3041				q->properties.exception_status &=
3042							~exception_clear_mask;
3043		}
3044
3045		dqm_unlock(dqm);
3046		mutex_unlock(&p->event_mutex);
3047		amdgpu_device_flush_hdp(dqm->dev->adev, NULL);
3048	}
3049
3050	if (total_suspended) {
3051		struct copy_context_work_handler_workarea copy_context_worker;
3052
3053		INIT_WORK_ONSTACK(
3054				&copy_context_worker.copy_context_work,
3055				copy_context_work_handler);
3056
3057		copy_context_worker.p = p;
3058
3059		schedule_work(&copy_context_worker.copy_context_work);
3060
3061
3062		flush_work(&copy_context_worker.copy_context_work);
3063		destroy_work_on_stack(&copy_context_worker.copy_context_work);
3064	}
3065
3066	if (copy_to_user((void __user *)usr_queue_id_array, queue_ids,
3067			num_queues * sizeof(uint32_t)))
3068		pr_err("copy_to_user failed on queue suspend\n");
3069
3070	kfree(queue_ids);
3071
3072	return total_suspended;
3073}
3074
3075static uint32_t set_queue_type_for_user(struct queue_properties *q_props)
3076{
3077	switch (q_props->type) {
3078	case KFD_QUEUE_TYPE_COMPUTE:
3079		return q_props->format == KFD_QUEUE_FORMAT_PM4
3080					? KFD_IOC_QUEUE_TYPE_COMPUTE
3081					: KFD_IOC_QUEUE_TYPE_COMPUTE_AQL;
3082	case KFD_QUEUE_TYPE_SDMA:
3083		return KFD_IOC_QUEUE_TYPE_SDMA;
3084	case KFD_QUEUE_TYPE_SDMA_XGMI:
3085		return KFD_IOC_QUEUE_TYPE_SDMA_XGMI;
3086	default:
3087		WARN_ONCE(true, "queue type not recognized!");
3088		return 0xffffffff;
3089	};
3090}
3091
3092void set_queue_snapshot_entry(struct queue *q,
3093			      uint64_t exception_clear_mask,
3094			      struct kfd_queue_snapshot_entry *qss_entry)
3095{
3096	qss_entry->ring_base_address = q->properties.queue_address;
3097	qss_entry->write_pointer_address = (uint64_t)q->properties.write_ptr;
3098	qss_entry->read_pointer_address = (uint64_t)q->properties.read_ptr;
3099	qss_entry->ctx_save_restore_address =
3100				q->properties.ctx_save_restore_area_address;
3101	qss_entry->ctx_save_restore_area_size =
3102				q->properties.ctx_save_restore_area_size;
3103	qss_entry->exception_status = q->properties.exception_status;
3104	qss_entry->queue_id = q->properties.queue_id;
3105	qss_entry->gpu_id = q->device->id;
3106	qss_entry->ring_size = (uint32_t)q->properties.queue_size;
3107	qss_entry->queue_type = set_queue_type_for_user(&q->properties);
3108	q->properties.exception_status &= ~exception_clear_mask;
3109}
3110
3111int debug_lock_and_unmap(struct device_queue_manager *dqm)
3112{
3113	struct device *dev = dqm->dev->adev->dev;
3114	int r;
3115
3116	if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
3117		dev_err(dev, "Unsupported on sched_policy: %i\n", dqm->sched_policy);
3118		return -EINVAL;
3119	}
3120
3121	if (!kfd_dbg_is_per_vmid_supported(dqm->dev))
3122		return 0;
3123
3124	dqm_lock(dqm);
3125
3126	r = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0, 0, false);
3127	if (r)
3128		dqm_unlock(dqm);
3129
3130	return r;
3131}
3132
3133int debug_map_and_unlock(struct device_queue_manager *dqm)
3134{
3135	struct device *dev = dqm->dev->adev->dev;
3136	int r;
3137
3138	if (dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
3139		dev_err(dev, "Unsupported on sched_policy: %i\n", dqm->sched_policy);
3140		return -EINVAL;
3141	}
3142
3143	if (!kfd_dbg_is_per_vmid_supported(dqm->dev))
3144		return 0;
3145
3146	r = map_queues_cpsch(dqm);
3147
3148	dqm_unlock(dqm);
3149
3150	return r;
3151}
3152
3153int debug_refresh_runlist(struct device_queue_manager *dqm)
3154{
3155	int r = debug_lock_and_unmap(dqm);
3156
3157	if (r)
3158		return r;
3159
3160	return debug_map_and_unlock(dqm);
3161}
3162
3163#if defined(CONFIG_DEBUG_FS)
3164
3165static void seq_reg_dump(struct seq_file *m,
3166			 uint32_t (*dump)[2], uint32_t n_regs)
3167{
3168	uint32_t i, count;
3169
3170	for (i = 0, count = 0; i < n_regs; i++) {
3171		if (count == 0 ||
3172		    dump[i-1][0] + sizeof(uint32_t) != dump[i][0]) {
3173			seq_printf(m, "%s    %08x: %08x",
3174				   i ? "\n" : "",
3175				   dump[i][0], dump[i][1]);
3176			count = 7;
3177		} else {
3178			seq_printf(m, " %08x", dump[i][1]);
3179			count--;
3180		}
3181	}
3182
3183	seq_puts(m, "\n");
3184}
3185
3186int dqm_debugfs_hqds(struct seq_file *m, void *data)
3187{
3188	struct device_queue_manager *dqm = data;
3189	uint32_t xcc_mask = dqm->dev->xcc_mask;
3190	uint32_t (*dump)[2], n_regs;
3191	int pipe, queue;
3192	int r = 0, xcc_id;
3193	uint32_t sdma_engine_start;
3194
3195	if (!dqm->sched_running) {
3196		seq_puts(m, " Device is stopped\n");
3197		return 0;
3198	}
3199
3200	for_each_inst(xcc_id, xcc_mask) {
3201		r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->adev,
3202						KFD_CIK_HIQ_PIPE,
3203						KFD_CIK_HIQ_QUEUE, &dump,
3204						&n_regs, xcc_id);
3205		if (!r) {
3206			seq_printf(
3207				m,
3208				"   Inst %d, HIQ on MEC %d Pipe %d Queue %d\n",
3209				xcc_id,
3210				KFD_CIK_HIQ_PIPE / get_pipes_per_mec(dqm) + 1,
3211				KFD_CIK_HIQ_PIPE % get_pipes_per_mec(dqm),
3212				KFD_CIK_HIQ_QUEUE);
3213			seq_reg_dump(m, dump, n_regs);
3214
3215			kfree(dump);
3216		}
 
 
 
3217
3218		for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) {
3219			int pipe_offset = pipe * get_queues_per_pipe(dqm);
 
 
3220
3221			for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) {
3222				if (!test_bit(pipe_offset + queue,
3223				      dqm->dev->kfd->shared_resources.cp_queue_bitmap))
3224					continue;
3225
3226				r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->adev,
3227								pipe, queue,
3228								&dump, &n_regs,
3229								xcc_id);
3230				if (r)
3231					break;
3232
3233				seq_printf(m,
3234					   " Inst %d,  CP Pipe %d, Queue %d\n",
3235					   xcc_id, pipe, queue);
3236				seq_reg_dump(m, dump, n_regs);
3237
3238				kfree(dump);
3239			}
 
 
 
3240		}
3241	}
3242
3243	sdma_engine_start = dqm->dev->node_id * get_num_all_sdma_engines(dqm);
3244	for (pipe = sdma_engine_start;
3245	     pipe < (sdma_engine_start + get_num_all_sdma_engines(dqm));
3246	     pipe++) {
3247		for (queue = 0;
3248		     queue < dqm->dev->kfd->device_info.num_sdma_queues_per_engine;
3249		     queue++) {
3250			r = dqm->dev->kfd2kgd->hqd_sdma_dump(
3251				dqm->dev->adev, pipe, queue, &dump, &n_regs);
3252			if (r)
3253				break;
3254
3255			seq_printf(m, "  SDMA Engine %d, RLC %d\n",
3256				  pipe, queue);
3257			seq_reg_dump(m, dump, n_regs);
3258
3259			kfree(dump);
3260		}
3261	}
3262
3263	return r;
3264}
3265
3266int dqm_debugfs_hang_hws(struct device_queue_manager *dqm)
3267{
3268	int r = 0;
3269
3270	dqm_lock(dqm);
3271	r = pm_debugfs_hang_hws(&dqm->packet_mgr);
3272	if (r) {
3273		dqm_unlock(dqm);
3274		return r;
3275	}
3276	dqm->active_runlist = true;
3277	r = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES,
3278				0, USE_DEFAULT_GRACE_PERIOD);
3279	dqm_unlock(dqm);
3280
3281	return r;
3282}
3283
3284#endif
v6.2
   1// SPDX-License-Identifier: GPL-2.0 OR MIT
   2/*
   3 * Copyright 2014-2022 Advanced Micro Devices, Inc.
   4 *
   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:
  11 *
  12 * The above copyright notice and this permission notice shall be included in
  13 * all copies or substantial portions of the Software.
  14 *
  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.
  22 *
  23 */
  24
  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>
  32#include "kfd_priv.h"
  33#include "kfd_device_queue_manager.h"
  34#include "kfd_mqd_manager.h"
  35#include "cik_regs.h"
  36#include "kfd_kernel_queue.h"
  37#include "amdgpu_amdkfd.h"
  38#include "mes_api_def.h"
 
  39
  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)
  43
  44static int set_pasid_vmid_mapping(struct device_queue_manager *dqm,
  45				  u32 pasid, unsigned int vmid);
  46
  47static int execute_queues_cpsch(struct device_queue_manager *dqm,
  48				enum kfd_unmap_queues_filter filter,
  49				uint32_t filter_param);
 
  50static int unmap_queues_cpsch(struct device_queue_manager *dqm,
  51				enum kfd_unmap_queues_filter filter,
  52				uint32_t filter_param, bool reset);
 
 
  53
  54static int map_queues_cpsch(struct device_queue_manager *dqm);
  55
  56static void deallocate_sdma_queue(struct device_queue_manager *dqm,
  57				struct queue *q);
  58
  59static inline void deallocate_hqd(struct device_queue_manager *dqm,
  60				struct queue *q);
  61static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q);
  62static int allocate_sdma_queue(struct device_queue_manager *dqm,
  63				struct queue *q, const uint32_t *restore_sdma_id);
  64static void kfd_process_hw_exception(struct work_struct *work);
  65
  66static inline
  67enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type)
  68{
  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;
  72}
  73
  74static bool is_pipe_enabled(struct device_queue_manager *dqm, int mec, int pipe)
  75{
  76	int i;
  77	int pipe_offset = (mec * dqm->dev->shared_resources.num_pipe_per_mec
  78		+ pipe) * dqm->dev->shared_resources.num_queue_per_pipe;
  79
  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))
  84			return true;
  85	return false;
  86}
  87
  88unsigned int get_cp_queues_num(struct device_queue_manager *dqm)
  89{
  90	return bitmap_weight(dqm->dev->shared_resources.cp_queue_bitmap,
  91				KGD_MAX_QUEUES);
  92}
  93
  94unsigned int get_queues_per_pipe(struct device_queue_manager *dqm)
  95{
  96	return dqm->dev->shared_resources.num_queue_per_pipe;
  97}
  98
  99unsigned int get_pipes_per_mec(struct device_queue_manager *dqm)
 100{
 101	return dqm->dev->shared_resources.num_pipe_per_mec;
 102}
 103
 104static unsigned int get_num_all_sdma_engines(struct device_queue_manager *dqm)
 105{
 106	return kfd_get_num_sdma_engines(dqm->dev) +
 107		kfd_get_num_xgmi_sdma_engines(dqm->dev);
 108}
 109
 110unsigned int get_num_sdma_queues(struct device_queue_manager *dqm)
 111{
 112	return kfd_get_num_sdma_engines(dqm->dev) *
 113		dqm->dev->device_info.num_sdma_queues_per_engine;
 114}
 115
 116unsigned int get_num_xgmi_sdma_queues(struct device_queue_manager *dqm)
 117{
 118	return kfd_get_num_xgmi_sdma_engines(dqm->dev) *
 119		dqm->dev->device_info.num_sdma_queues_per_engine;
 120}
 121
 122static inline uint64_t get_reserved_sdma_queues_bitmap(struct device_queue_manager *dqm)
 123{
 124	return dqm->dev->device_info.reserved_sdma_queues_bitmap;
 
 
 
 
 
 
 
 
 
 125}
 126
 127void program_sh_mem_settings(struct device_queue_manager *dqm,
 128					struct qcm_process_device *qpd)
 129{
 130	return dqm->dev->kfd2kgd->program_sh_mem_settings(
 131						dqm->dev->adev, qpd->vmid,
 132						qpd->sh_mem_config,
 133						qpd->sh_mem_ape1_base,
 134						qpd->sh_mem_ape1_limit,
 135						qpd->sh_mem_bases);
 
 
 136}
 137
 138static void kfd_hws_hang(struct device_queue_manager *dqm)
 139{
 140	/*
 141	 * Issue a GPU reset if HWS is unresponsive
 142	 */
 143	dqm->is_hws_hang = true;
 144
 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.
 148	 */
 149	if (!dqm->is_resetting)
 150		schedule_work(&dqm->hw_exception_work);
 151}
 152
 153static int convert_to_mes_queue_type(int queue_type)
 154{
 155	int mes_queue_type;
 156
 157	switch (queue_type) {
 158	case KFD_QUEUE_TYPE_COMPUTE:
 159		mes_queue_type = MES_QUEUE_TYPE_COMPUTE;
 160		break;
 161	case KFD_QUEUE_TYPE_SDMA:
 162		mes_queue_type = MES_QUEUE_TYPE_SDMA;
 163		break;
 164	default:
 165		WARN(1, "Invalid queue type %d", queue_type);
 166		mes_queue_type = -EINVAL;
 167		break;
 168	}
 169
 170	return mes_queue_type;
 171}
 172
 173static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
 174			 struct qcm_process_device *qpd)
 175{
 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;
 179	int r, queue_type;
 180	uint64_t wptr_addr_off;
 181
 182	if (dqm->is_hws_hang)
 183		return -EIO;
 184
 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;
 201
 202	if (q->wptr_bo) {
 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;
 205	}
 206
 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;
 210
 211	queue_input.paging = false;
 212	queue_input.tba_addr = qpd->tba_addr;
 213	queue_input.tma_addr = qpd->tma_addr;
 
 
 
 
 
 214
 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",
 218				q->properties.type);
 219		return -EINVAL;
 220	}
 221	queue_input.queue_type = (uint32_t)queue_type;
 222
 223	if (q->gws) {
 224		queue_input.gws_base = 0;
 225		queue_input.gws_size = qpd->num_gws;
 226	}
 227
 228	amdgpu_mes_lock(&adev->mes);
 229	r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input);
 230	amdgpu_mes_unlock(&adev->mes);
 231	if (r) {
 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");
 235		kfd_hws_hang(dqm);
 236}
 237
 238	return r;
 239}
 240
 241static int remove_queue_mes(struct device_queue_manager *dqm, struct queue *q,
 242			struct qcm_process_device *qpd)
 243{
 244	struct amdgpu_device *adev = (struct amdgpu_device *)dqm->dev->adev;
 245	int r;
 246	struct mes_remove_queue_input queue_input;
 247
 248	if (dqm->is_hws_hang)
 249		return -EIO;
 250
 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;
 254
 255	amdgpu_mes_lock(&adev->mes);
 256	r = adev->mes.funcs->remove_hw_queue(&adev->mes, &queue_input);
 257	amdgpu_mes_unlock(&adev->mes);
 258
 259	if (r) {
 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");
 263		kfd_hws_hang(dqm);
 264	}
 265
 266	return r;
 267}
 268
 269static int remove_all_queues_mes(struct device_queue_manager *dqm)
 270{
 271	struct device_process_node *cur;
 
 272	struct qcm_process_device *qpd;
 273	struct queue *q;
 274	int retval = 0;
 275
 276	list_for_each_entry(cur, &dqm->queues, list) {
 277		qpd = cur->qpd;
 278		list_for_each_entry(q, &qpd->queues_list, list) {
 279			if (q->properties.is_active) {
 280				retval = remove_queue_mes(dqm, q, qpd);
 281				if (retval) {
 282					pr_err("%s: Failed to remove queue %d for dev %d",
 283						__func__,
 284						q->properties.queue_id,
 285						dqm->dev->id);
 286					return retval;
 287				}
 288			}
 289		}
 290	}
 291
 292	return retval;
 293}
 294
 295static void increment_queue_count(struct device_queue_manager *dqm,
 296				  struct qcm_process_device *qpd,
 297				  struct queue *q)
 298{
 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++;
 303
 304	if (q->properties.is_gws) {
 305		dqm->gws_queue_count++;
 306		qpd->mapped_gws_queue = true;
 307	}
 308}
 309
 310static void decrement_queue_count(struct device_queue_manager *dqm,
 311				  struct qcm_process_device *qpd,
 312				  struct queue *q)
 313{
 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--;
 318
 319	if (q->properties.is_gws) {
 320		dqm->gws_queue_count--;
 321		qpd->mapped_gws_queue = false;
 322	}
 323}
 324
 325/*
 326 * Allocate a doorbell ID to this queue.
 327 * If doorbell_id is passed in, make sure requested ID is valid then allocate it.
 328 */
 329static int allocate_doorbell(struct qcm_process_device *qpd,
 330			     struct queue *q,
 331			     uint32_t const *restore_id)
 332{
 333	struct kfd_dev *dev = qpd->dqm->dev;
 334
 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.
 338		 */
 339
 340		if (restore_id && *restore_id != q->properties.queue_id)
 341			return -EINVAL;
 342
 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.
 350		 */
 351
 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);
 357
 358		if (restore_id && *restore_id != valid_id)
 359			return -EINVAL;
 360		q->doorbell_id = valid_id;
 361	} else {
 362		/* For CP queues on SOC15 */
 363		if (restore_id) {
 364			/* make sure that ID is free  */
 365			if (__test_and_set_bit(*restore_id, qpd->doorbell_bitmap))
 366				return -EINVAL;
 367
 368			q->doorbell_id = *restore_id;
 369		} else {
 370			/* or reserve a free doorbell ID */
 371			unsigned int found;
 372
 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");
 377				return -EBUSY;
 378			}
 379			set_bit(found, qpd->doorbell_bitmap);
 380			q->doorbell_id = found;
 381		}
 382	}
 383
 384	q->properties.doorbell_off =
 385		kfd_get_doorbell_dw_offset_in_bar(dev, qpd_to_pdd(qpd),
 386					  q->doorbell_id);
 
 387	return 0;
 388}
 389
 390static void deallocate_doorbell(struct qcm_process_device *qpd,
 391				struct queue *q)
 392{
 393	unsigned int old;
 394	struct kfd_dev *dev = qpd->dqm->dev;
 395
 396	if (!KFD_IS_SOC15(dev) ||
 397	    q->properties.type == KFD_QUEUE_TYPE_SDMA ||
 398	    q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
 399		return;
 400
 401	old = test_and_clear_bit(q->doorbell_id, qpd->doorbell_bitmap);
 402	WARN_ON(!old);
 403}
 404
 405static void program_trap_handler_settings(struct device_queue_manager *dqm,
 406				struct qcm_process_device *qpd)
 407{
 
 
 
 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);
 
 412}
 413
 414static int allocate_vmid(struct device_queue_manager *dqm,
 415			struct qcm_process_device *qpd,
 416			struct queue *q)
 417{
 
 418	int allocated_vmid = -1, i;
 419
 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]) {
 423			allocated_vmid = i;
 424			break;
 425		}
 426	}
 427
 428	if (allocated_vmid < 0) {
 429		pr_err("no more vmid to allocate\n");
 430		return -ENOSPC;
 431	}
 432
 433	pr_debug("vmid allocated: %d\n", allocated_vmid);
 434
 435	dqm->vmid_pasid[allocated_vmid] = q->process->pasid;
 436
 437	set_pasid_vmid_mapping(dqm, q->process->pasid, allocated_vmid);
 438
 439	qpd->vmid = allocated_vmid;
 440	q->properties.vmid = allocated_vmid;
 441
 442	program_sh_mem_settings(dqm, qpd);
 443
 444	if (KFD_IS_SOC15(dqm->dev) && dqm->dev->cwsr_enabled)
 445		program_trap_handler_settings(dqm, qpd);
 446
 447	/* qpd->page_table_base is set earlier when register_process()
 448	 * is called, i.e. when the first queue is created.
 449	 */
 450	dqm->dev->kfd2kgd->set_vm_context_page_table_base(dqm->dev->adev,
 451			qpd->vmid,
 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);
 455
 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);
 459
 460	return 0;
 461}
 462
 463static int flush_texture_cache_nocpsch(struct kfd_dev *kdev,
 464				struct qcm_process_device *qpd)
 465{
 466	const struct packet_manager_funcs *pmf = qpd->dqm->packet_mgr.pmf;
 467	int ret;
 468
 469	if (!qpd->ib_kaddr)
 470		return -ENOMEM;
 471
 472	ret = pmf->release_mem(qpd->ib_base, (uint32_t *)qpd->ib_kaddr);
 473	if (ret)
 474		return ret;
 475
 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));
 479}
 480
 481static void deallocate_vmid(struct device_queue_manager *dqm,
 482				struct qcm_process_device *qpd,
 483				struct queue *q)
 484{
 
 
 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");
 489
 490	kfd_flush_tlb(qpd_to_pdd(qpd), TLB_FLUSH_LEGACY);
 491
 492	/* Release the vmid mapping */
 493	set_pasid_vmid_mapping(dqm, 0, qpd->vmid);
 494	dqm->vmid_pasid[qpd->vmid] = 0;
 495
 496	qpd->vmid = 0;
 497	q->properties.vmid = 0;
 498}
 499
 500static int create_queue_nocpsch(struct device_queue_manager *dqm,
 501				struct queue *q,
 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)
 505{
 506	struct mqd_manager *mqd_mgr;
 507	int retval;
 508
 509	dqm_lock(dqm);
 510
 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);
 514		retval = -EPERM;
 515		goto out_unlock;
 516	}
 517
 518	if (list_empty(&qpd->queues_list)) {
 519		retval = allocate_vmid(dqm, qpd, q);
 520		if (retval)
 521			goto out_unlock;
 522	}
 523	q->properties.vmid = qpd->vmid;
 524	/*
 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.
 528	 */
 529	q->properties.is_evicted = !!qpd->evicted;
 530
 531	q->properties.tba_addr = qpd->tba_addr;
 532	q->properties.tma_addr = qpd->tma_addr;
 533
 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);
 538		if (retval)
 539			goto deallocate_vmid;
 540		pr_debug("Loading mqd to hqd on pipe %d, queue %d\n",
 541			q->pipe, q->queue);
 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);
 545		if (retval)
 546			goto deallocate_vmid;
 547		dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
 548	}
 549
 550	retval = allocate_doorbell(qpd, q, qd ? &qd->doorbell_id : NULL);
 551	if (retval)
 552		goto out_deallocate_hqd;
 553
 554	/* Temporarily release dqm lock to avoid a circular lock dependency */
 555	dqm_unlock(dqm);
 556	q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties);
 557	dqm_lock(dqm);
 558
 559	if (!q->mqd_mem_obj) {
 560		retval = -ENOMEM;
 561		goto out_deallocate_doorbell;
 562	}
 563
 564	if (qd)
 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,
 567				     qd->ctl_stack_size);
 568	else
 569		mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj,
 570					&q->gart_mqd_addr, &q->properties);
 571
 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;
 576		}
 577
 578		if (WARN(q->process->mm != current->mm,
 579					"should only run in user thread"))
 580			retval = -EFAULT;
 581		else
 582			retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe,
 583					q->queue, &q->properties, current->mm);
 584		if (retval)
 585			goto out_free_mqd;
 586	}
 587
 588add_queue_to_list:
 589	list_add(&q->list, &qpd->queues_list);
 590	qpd->queue_count++;
 591	if (q->properties.is_active)
 592		increment_queue_count(dqm, qpd, q);
 593
 594	/*
 595	 * Unconditionally increment this counter, regardless of the queue's
 596	 * type or whether the queue is active.
 597	 */
 598	dqm->total_queue_count++;
 599	pr_debug("Total of %d queues are accountable so far\n",
 600			dqm->total_queue_count);
 601	goto out_unlock;
 602
 603out_free_mqd:
 604	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
 605out_deallocate_doorbell:
 606	deallocate_doorbell(qpd, q);
 607out_deallocate_hqd:
 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);
 613deallocate_vmid:
 614	if (list_empty(&qpd->queues_list))
 615		deallocate_vmid(dqm, qpd, q);
 616out_unlock:
 617	dqm_unlock(dqm);
 618	return retval;
 619}
 620
 621static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q)
 622{
 623	bool set;
 624	int pipe, bit, i;
 625
 626	set = false;
 627
 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) {
 631
 632		if (!is_pipe_enabled(dqm, 0, pipe))
 633			continue;
 634
 635		if (dqm->allocated_queues[pipe] != 0) {
 636			bit = ffs(dqm->allocated_queues[pipe]) - 1;
 637			dqm->allocated_queues[pipe] &= ~(1 << bit);
 638			q->pipe = pipe;
 639			q->queue = bit;
 640			set = true;
 641			break;
 642		}
 643	}
 644
 645	if (!set)
 646		return -EBUSY;
 647
 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);
 651
 652	return 0;
 653}
 654
 655static inline void deallocate_hqd(struct device_queue_manager *dqm,
 656				struct queue *q)
 657{
 658	dqm->allocated_queues[q->pipe] |= (1 << q->queue);
 659}
 660
 661#define SQ_IND_CMD_CMD_KILL		0x00000003
 662#define SQ_IND_CMD_MODE_BROADCAST	0x00000001
 663
 664static int dbgdev_wave_reset_wavefronts(struct kfd_dev *dev, struct kfd_process *p)
 665{
 666	int status = 0;
 667	unsigned int vmid;
 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;
 
 
 674
 675	reg_sq_cmd.u32All = 0;
 676	reg_gfx_index.u32All = 0;
 677
 678	pr_debug("Killing all process wavefronts\n");
 679
 680	if (!dev->kfd2kgd->get_atc_vmid_pasid_mapping_info) {
 681		pr_err("no vmid pasid mapping supported \n");
 682		return -EOPNOTSUPP;
 683	}
 684
 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.
 688	 */
 689
 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);
 693
 694		if (status && queried_pasid == p->pasid) {
 695			pr_debug("Killing wave fronts of vmid %d and pasid 0x%x\n",
 696					vmid, p->pasid);
 697			break;
 698		}
 699	}
 700
 701	if (vmid > last_vmid_to_scan) {
 702		pr_err("Didn't find vmid for pasid 0x%x\n", p->pasid);
 703		return -EFAULT;
 704	}
 705
 706	/* taking the VMID for that process on the safe way using PDD */
 707	pdd = kfd_get_process_device_data(dev, p);
 708	if (!pdd)
 709		return -EFAULT;
 710
 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;
 717
 718	dev->kfd2kgd->wave_control_execute(dev->adev,
 719					reg_gfx_index.u32All,
 720					reg_sq_cmd.u32All);
 
 721
 722	return 0;
 723}
 724
 725/* Access to DQM has to be locked before calling destroy_queue_nocpsch_locked
 726 * to avoid asynchronized access
 727 */
 728static int destroy_queue_nocpsch_locked(struct device_queue_manager *dqm,
 729				struct qcm_process_device *qpd,
 730				struct queue *q)
 731{
 732	int retval;
 733	struct mqd_manager *mqd_mgr;
 734
 735	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
 736			q->properties.type)];
 737
 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);
 744	else {
 745		pr_debug("q->properties.type %d is invalid\n",
 746				q->properties.type);
 747		return -EINVAL;
 748	}
 749	dqm->total_queue_count--;
 750
 751	deallocate_doorbell(qpd, q);
 752
 753	if (!dqm->sched_running) {
 754		WARN_ONCE(1, "Destroy non-HWS queue while stopped\n");
 755		return 0;
 756	}
 757
 758	retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
 759				KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
 760				KFD_UNMAP_LATENCY_MS,
 761				q->pipe, q->queue);
 762	if (retval == -ETIME)
 763		qpd->reset_wavefronts = true;
 764
 765	list_del(&q->list);
 766	if (list_empty(&qpd->queues_list)) {
 767		if (qpd->reset_wavefronts) {
 768			pr_warn("Resetting wave fronts (nocpsch) on dev %p\n",
 769					dqm->dev);
 770			/* dbgdev_wave_reset_wavefronts has to be called before
 771			 * deallocate_vmid(), i.e. when vmid is still in use.
 772			 */
 773			dbgdev_wave_reset_wavefronts(dqm->dev,
 774					qpd->pqm->process);
 775			qpd->reset_wavefronts = false;
 776		}
 777
 778		deallocate_vmid(dqm, qpd, q);
 779	}
 780	qpd->queue_count--;
 781	if (q->properties.is_active)
 782		decrement_queue_count(dqm, qpd, q);
 783
 784	return retval;
 785}
 786
 787static int destroy_queue_nocpsch(struct device_queue_manager *dqm,
 788				struct qcm_process_device *qpd,
 789				struct queue *q)
 790{
 791	int retval;
 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)];
 796
 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,
 801							&sdma_val);
 802		if (retval)
 803			pr_err("Failed to read SDMA queue counter for queue: %d\n",
 804				q->properties.queue_id);
 805	}
 806
 807	dqm_lock(dqm);
 808	retval = destroy_queue_nocpsch_locked(dqm, qpd, q);
 809	if (!retval)
 810		pdd->sdma_past_activity_counter += sdma_val;
 811	dqm_unlock(dqm);
 812
 813	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
 814
 815	return retval;
 816}
 817
 818static int update_queue(struct device_queue_manager *dqm, struct queue *q,
 819			struct mqd_update_info *minfo)
 820{
 821	int retval = 0;
 
 822	struct mqd_manager *mqd_mgr;
 823	struct kfd_process_device *pdd;
 824	bool prev_active = false;
 825
 826	dqm_lock(dqm);
 827	pdd = kfd_get_process_device_data(q->device, q->process);
 828	if (!pdd) {
 829		retval = -ENODEV;
 830		goto out_unlock;
 831	}
 832	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
 833			q->properties.type)];
 834
 835	/* Save previous activity state for counters */
 836	prev_active = q->properties.is_active;
 837
 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);
 845
 846		if (retval) {
 847			pr_err("unmap queue failed\n");
 848			goto out_unlock;
 849		}
 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)) {
 854
 855		if (!dqm->sched_running) {
 856			WARN_ONCE(1, "Update non-HWS queue while stopped\n");
 857			goto out_unlock;
 858		}
 859
 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);
 865		if (retval) {
 866			pr_err("destroy mqd failed\n");
 867			goto out_unlock;
 868		}
 869	}
 870
 871	mqd_mgr->update_mqd(mqd_mgr, q->mqd, &q->properties, minfo);
 872
 873	/*
 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
 877	 * uploaded.
 878	 */
 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;
 887		}
 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;
 893		}
 894		q->properties.is_gws = false;
 895	}
 896
 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"))
 908			retval = -EFAULT;
 909		else
 910			retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd,
 911						   q->pipe, q->queue,
 912						   &q->properties, current->mm);
 913	}
 914
 915out_unlock:
 916	dqm_unlock(dqm);
 917	return retval;
 918}
 919
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 920static int evict_process_queues_nocpsch(struct device_queue_manager *dqm,
 921					struct qcm_process_device *qpd)
 922{
 923	struct queue *q;
 924	struct mqd_manager *mqd_mgr;
 925	struct kfd_process_device *pdd;
 926	int retval, ret = 0;
 927
 928	dqm_lock(dqm);
 929	if (qpd->evicted++ > 0) /* already evicted, do nothing */
 930		goto out;
 931
 932	pdd = qpd_to_pdd(qpd);
 933	pr_debug_ratelimited("Evicting PASID 0x%x queues\n",
 934			    pdd->process->pasid);
 935
 936	pdd->last_evict_timestamp = get_jiffies_64();
 937	/* Mark all queues as evicted. Deactivate all active queues on
 938	 * the qpd.
 939	 */
 940	list_for_each_entry(q, &qpd->queues_list, list) {
 941		q->properties.is_evicted = true;
 942		if (!q->properties.is_active)
 943			continue;
 944
 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);
 949
 950		if (WARN_ONCE(!dqm->sched_running, "Evict when stopped\n"))
 951			continue;
 952
 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);
 958		if (retval && !ret)
 959			/* Return the first error, but keep going to
 960			 * maintain a consistent eviction state
 961			 */
 962			ret = retval;
 963	}
 964
 965out:
 966	dqm_unlock(dqm);
 967	return ret;
 968}
 969
 970static int evict_process_queues_cpsch(struct device_queue_manager *dqm,
 971				      struct qcm_process_device *qpd)
 972{
 973	struct queue *q;
 
 974	struct kfd_process_device *pdd;
 975	int retval = 0;
 976
 977	dqm_lock(dqm);
 978	if (qpd->evicted++ > 0) /* already evicted, do nothing */
 979		goto out;
 980
 981	pdd = qpd_to_pdd(qpd);
 
 
 
 
 
 
 
 
 982	pr_debug_ratelimited("Evicting PASID 0x%x queues\n",
 983			    pdd->process->pasid);
 984
 985	/* Mark all queues as evicted. Deactivate all active queues on
 986	 * the qpd.
 987	 */
 988	list_for_each_entry(q, &qpd->queues_list, list) {
 989		q->properties.is_evicted = true;
 990		if (!q->properties.is_active)
 991			continue;
 992
 993		q->properties.is_active = false;
 994		decrement_queue_count(dqm, qpd, q);
 995
 996		if (dqm->dev->shared_resources.enable_mes) {
 997			retval = remove_queue_mes(dqm, q, qpd);
 998			if (retval) {
 999				pr_err("Failed to evict queue %d\n",
1000					q->properties.queue_id);
1001				goto out;
1002			}
1003		}
1004	}
1005	pdd->last_evict_timestamp = get_jiffies_64();
1006	if (!dqm->dev->shared_resources.enable_mes)
1007		retval = execute_queues_cpsch(dqm,
1008					      qpd->is_debug ?
1009					      KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES :
1010					      KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
 
1011
1012out:
1013	dqm_unlock(dqm);
1014	return retval;
1015}
1016
1017static int restore_process_queues_nocpsch(struct device_queue_manager *dqm,
1018					  struct qcm_process_device *qpd)
1019{
1020	struct mm_struct *mm = NULL;
1021	struct queue *q;
1022	struct mqd_manager *mqd_mgr;
1023	struct kfd_process_device *pdd;
1024	uint64_t pd_base;
1025	uint64_t eviction_duration;
1026	int retval, ret = 0;
1027
1028	pdd = qpd_to_pdd(qpd);
1029	/* Retrieve PD base */
1030	pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv);
1031
1032	dqm_lock(dqm);
1033	if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */
1034		goto out;
1035	if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
1036		qpd->evicted--;
1037		goto out;
1038	}
1039
1040	pr_debug_ratelimited("Restoring PASID 0x%x queues\n",
1041			    pdd->process->pasid);
1042
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);
1046
1047	if (!list_empty(&qpd->queues_list)) {
1048		dqm->dev->kfd2kgd->set_vm_context_page_table_base(
1049				dqm->dev->adev,
1050				qpd->vmid,
1051				qpd->page_table_base);
1052		kfd_flush_tlb(pdd, TLB_FLUSH_LEGACY);
1053	}
1054
1055	/* Take a safe reference to the mm_struct, which may otherwise
1056	 * disappear even while the kfd_process is still referenced.
1057	 */
1058	mm = get_task_mm(pdd->process->lead_thread);
1059	if (!mm) {
1060		ret = -EFAULT;
1061		goto out;
1062	}
1063
1064	/* Remove the eviction flags. Activate queues that are not
1065	 * inactive for other reasons.
1066	 */
1067	list_for_each_entry(q, &qpd->queues_list, list) {
1068		q->properties.is_evicted = false;
1069		if (!QUEUE_IS_ACTIVE(q->properties))
1070			continue;
1071
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);
1076
1077		if (WARN_ONCE(!dqm->sched_running, "Restore when stopped\n"))
1078			continue;
1079
1080		retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe,
1081				       q->queue, &q->properties, mm);
1082		if (retval && !ret)
1083			/* Return the first error, but keep going to
1084			 * maintain a consistent eviction state
1085			 */
1086			ret = retval;
1087	}
1088	qpd->evicted = 0;
1089	eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp;
1090	atomic64_add(eviction_duration, &pdd->evict_duration_counter);
1091out:
1092	if (mm)
1093		mmput(mm);
1094	dqm_unlock(dqm);
1095	return ret;
1096}
1097
1098static int restore_process_queues_cpsch(struct device_queue_manager *dqm,
1099					struct qcm_process_device *qpd)
1100{
1101	struct queue *q;
 
1102	struct kfd_process_device *pdd;
1103	uint64_t pd_base;
1104	uint64_t eviction_duration;
1105	int retval = 0;
1106
1107	pdd = qpd_to_pdd(qpd);
1108	/* Retrieve PD base */
1109	pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv);
1110
1111	dqm_lock(dqm);
1112	if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */
1113		goto out;
1114	if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
1115		qpd->evicted--;
1116		goto out;
1117	}
1118
 
 
 
 
 
 
 
1119	pr_debug_ratelimited("Restoring PASID 0x%x queues\n",
1120			    pdd->process->pasid);
1121
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);
1125
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))
1130			continue;
1131
1132		q->properties.is_active = true;
1133		increment_queue_count(dqm, &pdd->qpd, q);
1134
1135		if (dqm->dev->shared_resources.enable_mes) {
1136			retval = add_queue_mes(dqm, q, qpd);
1137			if (retval) {
1138				pr_err("Failed to restore queue %d\n",
1139					q->properties.queue_id);
1140				goto out;
1141			}
1142		}
1143	}
1144	if (!dqm->dev->shared_resources.enable_mes)
1145		retval = execute_queues_cpsch(dqm,
1146					      KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1147	qpd->evicted = 0;
1148	eviction_duration = get_jiffies_64() - pdd->last_evict_timestamp;
1149	atomic64_add(eviction_duration, &pdd->evict_duration_counter);
 
 
1150out:
1151	dqm_unlock(dqm);
1152	return retval;
1153}
1154
1155static int register_process(struct device_queue_manager *dqm,
1156					struct qcm_process_device *qpd)
1157{
1158	struct device_process_node *n;
1159	struct kfd_process_device *pdd;
1160	uint64_t pd_base;
1161	int retval;
1162
1163	n = kzalloc(sizeof(*n), GFP_KERNEL);
1164	if (!n)
1165		return -ENOMEM;
1166
1167	n->qpd = qpd;
1168
1169	pdd = qpd_to_pdd(qpd);
1170	/* Retrieve PD base */
1171	pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->drm_priv);
1172
1173	dqm_lock(dqm);
1174	list_add(&n->list, &dqm->queues);
1175
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);
1179
1180	retval = dqm->asic_ops.update_qpd(dqm, qpd);
1181
1182	dqm->processes_count++;
1183
1184	dqm_unlock(dqm);
1185
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.
1188	 */
1189	kfd_inc_compute_active(dqm->dev);
1190
1191	return retval;
1192}
1193
1194static int unregister_process(struct device_queue_manager *dqm,
1195					struct qcm_process_device *qpd)
1196{
1197	int retval;
1198	struct device_process_node *cur, *next;
1199
1200	pr_debug("qpd->queues_list is %s\n",
1201			list_empty(&qpd->queues_list) ? "empty" : "not empty");
1202
1203	retval = 0;
1204	dqm_lock(dqm);
1205
1206	list_for_each_entry_safe(cur, next, &dqm->queues, list) {
1207		if (qpd == cur->qpd) {
1208			list_del(&cur->list);
1209			kfree(cur);
1210			dqm->processes_count--;
1211			goto out;
1212		}
1213	}
1214	/* qpd not found in dqm list */
1215	retval = 1;
1216out:
1217	dqm_unlock(dqm);
1218
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.
1221	 */
1222	if (!retval)
1223		kfd_dec_compute_active(dqm->dev);
1224
1225	return retval;
1226}
1227
1228static int
1229set_pasid_vmid_mapping(struct device_queue_manager *dqm, u32 pasid,
1230			unsigned int vmid)
1231{
1232	return dqm->dev->kfd2kgd->set_pasid_vmid_mapping(
1233						dqm->dev->adev, pasid, vmid);
1234}
1235
1236static void init_interrupts(struct device_queue_manager *dqm)
1237{
1238	unsigned int i;
 
 
 
1239
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);
1243}
1244
1245static void init_sdma_bitmaps(struct device_queue_manager *dqm)
1246{
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));
1253
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);
1258
1259	dqm->sdma_bitmap &= ~get_reserved_sdma_queues_bitmap(dqm);
1260	pr_info("sdma_bitmap: %llx\n", dqm->sdma_bitmap);
 
 
 
 
 
 
1261}
1262
1263static int initialize_nocpsch(struct device_queue_manager *dqm)
1264{
1265	int pipe, queue;
1266
1267	pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm));
1268
1269	dqm->allocated_queues = kcalloc(get_pipes_per_mec(dqm),
1270					sizeof(unsigned int), GFP_KERNEL);
1271	if (!dqm->allocated_queues)
1272		return -ENOMEM;
1273
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;
1279
1280	for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) {
1281		int pipe_offset = pipe * get_queues_per_pipe(dqm);
1282
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;
1287	}
1288
1289	memset(dqm->vmid_pasid, 0, sizeof(dqm->vmid_pasid));
1290
1291	init_sdma_bitmaps(dqm);
1292
1293	return 0;
1294}
1295
1296static void uninitialize(struct device_queue_manager *dqm)
1297{
1298	int i;
1299
1300	WARN_ON(dqm->active_queue_count > 0 || dqm->processes_count > 0);
1301
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);
1306}
1307
1308static int start_nocpsch(struct device_queue_manager *dqm)
1309{
1310	int r = 0;
1311
1312	pr_info("SW scheduler is used");
1313	init_interrupts(dqm);
1314
1315	if (dqm->dev->adev->asic_type == CHIP_HAWAII)
1316		r = pm_init(&dqm->packet_mgr, dqm);
1317	if (!r)
1318		dqm->sched_running = true;
1319
1320	return r;
1321}
1322
1323static int stop_nocpsch(struct device_queue_manager *dqm)
1324{
 
 
 
 
 
 
1325	if (dqm->dev->adev->asic_type == CHIP_HAWAII)
1326		pm_uninit(&dqm->packet_mgr, false);
1327	dqm->sched_running = false;
 
1328
1329	return 0;
1330}
1331
1332static void pre_reset(struct device_queue_manager *dqm)
1333{
1334	dqm_lock(dqm);
1335	dqm->is_resetting = true;
1336	dqm_unlock(dqm);
1337}
1338
1339static int allocate_sdma_queue(struct device_queue_manager *dqm,
1340				struct queue *q, const uint32_t *restore_sdma_id)
1341{
 
1342	int bit;
1343
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");
1347			return -ENOMEM;
1348		}
1349
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");
1354				return -EBUSY;
1355			}
1356			dqm->sdma_bitmap &= ~(1ULL << *restore_sdma_id);
1357			q->sdma_id = *restore_sdma_id;
1358		} else {
1359			/* Find first available sdma_id */
1360			bit = __ffs64(dqm->sdma_bitmap);
1361			dqm->sdma_bitmap &= ~(1ULL << bit);
 
1362			q->sdma_id = bit;
1363		}
1364
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");
1372			return -ENOMEM;
1373		}
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");
1378				return -EBUSY;
1379			}
1380			dqm->xgmi_sdma_bitmap &= ~(1ULL << *restore_sdma_id);
1381			q->sdma_id = *restore_sdma_id;
1382		} else {
1383			bit = __ffs64(dqm->xgmi_sdma_bitmap);
1384			dqm->xgmi_sdma_bitmap &= ~(1ULL << bit);
 
1385			q->sdma_id = bit;
1386		}
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
1392		 */
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);
1398	}
1399
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);
1402
1403	return 0;
1404}
1405
1406static void deallocate_sdma_queue(struct device_queue_manager *dqm,
1407				struct queue *q)
1408{
1409	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
1410		if (q->sdma_id >= get_num_sdma_queues(dqm))
1411			return;
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))
1415			return;
1416		dqm->xgmi_sdma_bitmap |= (1ULL << q->sdma_id);
1417	}
1418}
1419
1420/*
1421 * Device Queue Manager implementation for cp scheduler
1422 */
1423
1424static int set_sched_resources(struct device_queue_manager *dqm)
1425{
1426	int i, mec;
1427	struct scheduling_resources res;
 
1428
1429	res.vmid_mask = dqm->dev->shared_resources.compute_vmid_bitmap;
1430
1431	res.queue_mask = 0;
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;
1435
1436		if (!test_bit(i, dqm->dev->shared_resources.cp_queue_bitmap))
1437			continue;
1438
1439		/* only acquire queues from the first MEC */
1440		if (mec > 0)
1441			continue;
1442
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
1446		 */
1447		if (WARN_ON(i >= (sizeof(res.queue_mask)*8))) {
1448			pr_err("Invalid queue enabled by amdgpu: %d\n", i);
1449			break;
1450		}
1451
1452		res.queue_mask |= 1ull
1453			<< amdgpu_queue_mask_bit_to_set_resource_bit(
1454				dqm->dev->adev, i);
1455	}
1456	res.gws_mask = ~0ull;
1457	res.oac_mask = res.gds_heap_base = res.gds_heap_size = 0;
1458
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);
1463
1464	return pm_send_set_resources(&dqm->packet_mgr, &res);
1465}
1466
1467static int initialize_cpsch(struct device_queue_manager *dqm)
1468{
1469	pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm));
1470
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);
 
1478
1479	init_sdma_bitmaps(dqm);
1480
 
 
 
 
1481	return 0;
1482}
1483
1484static int start_cpsch(struct device_queue_manager *dqm)
1485{
 
1486	int retval;
1487
1488	retval = 0;
1489
1490	dqm_lock(dqm);
1491
1492	if (!dqm->dev->shared_resources.enable_mes) {
1493		retval = pm_init(&dqm->packet_mgr, dqm);
1494		if (retval)
1495			goto fail_packet_manager_init;
1496
1497		retval = set_sched_resources(dqm);
1498		if (retval)
1499			goto fail_set_sched_resources;
1500	}
1501	pr_debug("Allocating fence memory\n");
1502
1503	/* allocate fence memory on the gart */
1504	retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr),
1505					&dqm->fence_mem);
1506
1507	if (retval)
1508		goto fail_allocate_vidmem;
1509
1510	dqm->fence_addr = (uint64_t *)dqm->fence_mem->cpu_ptr;
1511	dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr;
1512
1513	init_interrupts(dqm);
1514
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);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1521	dqm_unlock(dqm);
1522
1523	return 0;
1524fail_allocate_vidmem:
1525fail_set_sched_resources:
1526	if (!dqm->dev->shared_resources.enable_mes)
1527		pm_uninit(&dqm->packet_mgr, false);
1528fail_packet_manager_init:
1529	dqm_unlock(dqm);
1530	return retval;
1531}
1532
1533static int stop_cpsch(struct device_queue_manager *dqm)
1534{
1535	bool hanging;
1536
1537	dqm_lock(dqm);
1538	if (!dqm->sched_running) {
1539		dqm_unlock(dqm);
1540		return 0;
1541	}
1542
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);
1546		else
1547			remove_all_queues_mes(dqm);
1548	}
1549
1550	hanging = dqm->is_hws_hang || dqm->is_resetting;
1551	dqm->sched_running = false;
1552
1553	if (!dqm->dev->shared_resources.enable_mes)
1554		pm_release_ib(&dqm->packet_mgr);
1555
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);
1559	dqm_unlock(dqm);
1560
1561	return 0;
1562}
1563
1564static int create_kernel_queue_cpsch(struct device_queue_manager *dqm,
1565					struct kernel_queue *kq,
1566					struct qcm_process_device *qpd)
1567{
1568	dqm_lock(dqm);
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);
1572		dqm_unlock(dqm);
1573		return -EPERM;
1574	}
1575
1576	/*
1577	 * Unconditionally increment this counter, regardless of the queue's
1578	 * type or whether the queue is active.
1579	 */
1580	dqm->total_queue_count++;
1581	pr_debug("Total of %d queues are accountable so far\n",
1582			dqm->total_queue_count);
1583
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);
 
1588	dqm_unlock(dqm);
1589
1590	return 0;
1591}
1592
1593static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm,
1594					struct kernel_queue *kq,
1595					struct qcm_process_device *qpd)
1596{
1597	dqm_lock(dqm);
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);
 
1602	/*
1603	 * Unconditionally decrement this counter, regardless of the queue's
1604	 * type.
1605	 */
1606	dqm->total_queue_count--;
1607	pr_debug("Total of %d queues are accountable so far\n",
1608			dqm->total_queue_count);
1609	dqm_unlock(dqm);
1610}
1611
1612static 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)
1616{
1617	int retval;
1618	struct mqd_manager *mqd_mgr;
1619
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);
1623		retval = -EPERM;
1624		goto out;
1625	}
1626
1627	if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1628		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1629		dqm_lock(dqm);
1630		retval = allocate_sdma_queue(dqm, q, qd ? &qd->sdma_id : NULL);
1631		dqm_unlock(dqm);
1632		if (retval)
1633			goto out;
1634	}
1635
1636	retval = allocate_doorbell(qpd, q, qd ? &qd->doorbell_id : NULL);
1637	if (retval)
1638		goto out_deallocate_sdma_queue;
1639
1640	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1641			q->properties.type)];
1642
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) {
1650		retval = -ENOMEM;
1651		goto out_deallocate_doorbell;
1652	}
1653
1654	dqm_lock(dqm);
1655	/*
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.
1659	 */
1660	q->properties.is_evicted = !!qpd->evicted;
 
 
1661
1662	if (qd)
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);
1666	else
1667		mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj,
1668					&q->gart_mqd_addr, &q->properties);
1669
1670	list_add(&q->list, &qpd->queues_list);
1671	qpd->queue_count++;
1672
1673	if (q->properties.is_active) {
1674		increment_queue_count(dqm, qpd, q);
1675
1676		if (!dqm->dev->shared_resources.enable_mes)
1677			retval = execute_queues_cpsch(dqm,
1678					KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1679		else
1680			retval = add_queue_mes(dqm, q, qpd);
1681		if (retval)
1682			goto cleanup_queue;
1683	}
1684
1685	/*
1686	 * Unconditionally increment this counter, regardless of the queue's
1687	 * type or whether the queue is active.
1688	 */
1689	dqm->total_queue_count++;
1690
1691	pr_debug("Total of %d queues are accountable so far\n",
1692			dqm->total_queue_count);
1693
1694	dqm_unlock(dqm);
1695	return retval;
1696
1697cleanup_queue:
1698	qpd->queue_count--;
1699	list_del(&q->list);
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);
1703	dqm_unlock(dqm);
1704out_deallocate_doorbell:
1705	deallocate_doorbell(qpd, q);
1706out_deallocate_sdma_queue:
1707	if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1708		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1709		dqm_lock(dqm);
1710		deallocate_sdma_queue(dqm, q);
1711		dqm_unlock(dqm);
1712	}
1713out:
1714	return retval;
1715}
1716
1717int amdkfd_fence_wait_timeout(uint64_t *fence_addr,
1718				uint64_t fence_value,
1719				unsigned int timeout_ms)
1720{
1721	unsigned long end_jiffies = msecs_to_jiffies(timeout_ms) + jiffies;
 
 
1722
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.
1729			 */
1730			while (halt_if_hws_hang)
1731				schedule();
1732
1733			return -ETIME;
1734		}
1735		schedule();
1736	}
1737
1738	return 0;
1739}
1740
1741/* dqm->lock mutex has to be locked before calling this function */
1742static int map_queues_cpsch(struct device_queue_manager *dqm)
1743{
 
1744	int retval;
1745
1746	if (!dqm->sched_running)
1747		return 0;
1748	if (dqm->active_queue_count <= 0 || dqm->processes_count <= 0)
1749		return 0;
1750	if (dqm->active_runlist)
1751		return 0;
1752
1753	retval = pm_send_runlist(&dqm->packet_mgr, &dqm->queues);
1754	pr_debug("%s sent runlist\n", __func__);
1755	if (retval) {
1756		pr_err("failed to execute runlist\n");
1757		return retval;
1758	}
1759	dqm->active_runlist = true;
1760
1761	return retval;
1762}
1763
1764/* dqm->lock mutex has to be locked before calling this function */
1765static int unmap_queues_cpsch(struct device_queue_manager *dqm,
1766				enum kfd_unmap_queues_filter filter,
1767				uint32_t filter_param, bool reset)
 
 
1768{
 
 
1769	int retval = 0;
1770	struct mqd_manager *mqd_mgr;
1771
1772	if (!dqm->sched_running)
1773		return 0;
1774	if (dqm->is_hws_hang || dqm->is_resetting)
1775		return -EIO;
1776	if (!dqm->active_runlist)
1777		return retval;
1778
 
 
 
 
 
 
1779	retval = pm_send_unmap_queue(&dqm->packet_mgr, filter, filter_param, reset);
1780	if (retval)
1781		return retval;
1782
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);
1789	if (retval) {
1790		pr_err("The cp might be in an unrecoverable state due to an unsuccessful queues preemption\n");
1791		kfd_hws_hang(dqm);
1792		return retval;
1793	}
1794
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
1802	 */
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)
1807			schedule();
 
1808		return -ETIME;
1809	}
1810
 
 
 
 
 
 
 
1811	pm_release_ib(&dqm->packet_mgr);
1812	dqm->active_runlist = false;
1813
1814	return retval;
1815}
1816
1817/* only for compute queue */
1818static int reset_queues_cpsch(struct device_queue_manager *dqm,
1819			uint16_t pasid)
1820{
1821	int retval;
1822
1823	dqm_lock(dqm);
1824
1825	retval = unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_BY_PASID,
1826			pasid, true);
1827
1828	dqm_unlock(dqm);
1829	return retval;
1830}
1831
1832/* dqm->lock mutex has to be locked before calling this function */
1833static int execute_queues_cpsch(struct device_queue_manager *dqm,
1834				enum kfd_unmap_queues_filter filter,
1835				uint32_t filter_param)
 
1836{
1837	int retval;
1838
1839	if (dqm->is_hws_hang)
1840		return -EIO;
1841	retval = unmap_queues_cpsch(dqm, filter, filter_param, false);
1842	if (retval)
1843		return retval;
1844
1845	return map_queues_cpsch(dqm);
1846}
1847
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1848static int destroy_queue_cpsch(struct device_queue_manager *dqm,
1849				struct qcm_process_device *qpd,
1850				struct queue *q)
1851{
1852	int retval;
1853	struct mqd_manager *mqd_mgr;
1854	uint64_t sdma_val = 0;
1855	struct kfd_process_device *pdd = qpd_to_pdd(qpd);
 
1856
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,
1861							&sdma_val);
1862		if (retval)
1863			pr_err("Failed to read SDMA queue counter for queue: %d\n",
1864				q->properties.queue_id);
1865	}
1866
1867	retval = 0;
1868
1869	/* remove queue from list to prevent rescheduling after preemption */
1870	dqm_lock(dqm);
1871
 
 
 
 
 
 
 
1872	if (qpd->is_debug) {
1873		/*
1874		 * error, currently we do not allow to destroy a queue
1875		 * of a currently debugged process
1876		 */
1877		retval = -EBUSY;
1878		goto failed_try_destroy_debugged_queue;
1879
1880	}
1881
1882	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1883			q->properties.type)];
1884
1885	deallocate_doorbell(qpd, q);
1886
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;
1891	}
1892
1893	list_del(&q->list);
1894	qpd->queue_count--;
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;
1902		} else {
1903			retval = remove_queue_mes(dqm, q, qpd);
1904		}
1905	}
1906
1907	/*
1908	 * Unconditionally decrement this counter, regardless of the queue's
1909	 * type
1910	 */
1911	dqm->total_queue_count--;
1912	pr_debug("Total of %d queues are accountable so far\n",
1913			dqm->total_queue_count);
1914
1915	dqm_unlock(dqm);
1916
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);
1919
1920	return retval;
1921
1922failed_try_destroy_debugged_queue:
1923
1924	dqm_unlock(dqm);
1925	return retval;
1926}
1927
1928/*
1929 * Low bits must be 0000/FFFF as required by HW, high bits must be 0 to
1930 * stay in user mode.
1931 */
1932#define APE1_FIXED_BITS_MASK 0xFFFF80000000FFFFULL
1933/* APE1 limit is inclusive and 64K aligned. */
1934#define APE1_LIMIT_ALIGNMENT 0xFFFF
1935
1936static 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)
1942{
1943	bool retval = true;
1944
1945	if (!dqm->asic_ops.set_cache_memory_policy)
1946		return retval;
1947
1948	dqm_lock(dqm);
1949
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;
1954	} else {
1955		/*
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.
1963		 */
1964
1965		uint64_t base = (uintptr_t)alternate_aperture_base;
1966		uint64_t limit = base + alternate_aperture_size - 1;
1967
1968		if (limit <= base || (base & APE1_FIXED_BITS_MASK) != 0 ||
1969		   (limit & APE1_FIXED_BITS_MASK) != APE1_LIMIT_ALIGNMENT) {
1970			retval = false;
1971			goto out;
1972		}
1973
1974		qpd->sh_mem_ape1_base = base >> 16;
1975		qpd->sh_mem_ape1_limit = limit >> 16;
1976	}
1977
1978	retval = dqm->asic_ops.set_cache_memory_policy(
1979			dqm,
1980			qpd,
1981			default_policy,
1982			alternate_policy,
1983			alternate_aperture_base,
1984			alternate_aperture_size);
1985
1986	if ((dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0))
1987		program_sh_mem_settings(dqm, qpd);
1988
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);
1992
1993out:
1994	dqm_unlock(dqm);
1995	return retval;
1996}
1997
1998static int process_termination_nocpsch(struct device_queue_manager *dqm,
1999		struct qcm_process_device *qpd)
2000{
2001	struct queue *q;
2002	struct device_process_node *cur, *next_dpn;
2003	int retval = 0;
2004	bool found = false;
2005
2006	dqm_lock(dqm);
2007
2008	/* Clear all user mode queues */
2009	while (!list_empty(&qpd->queues_list)) {
2010		struct mqd_manager *mqd_mgr;
2011		int ret;
2012
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);
2017		if (ret)
2018			retval = ret;
2019		dqm_unlock(dqm);
2020		mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
2021		dqm_lock(dqm);
2022	}
2023
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);
2028			kfree(cur);
2029			dqm->processes_count--;
2030			found = true;
2031			break;
2032		}
2033	}
2034
2035	dqm_unlock(dqm);
2036
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.
2039	 */
2040	if (found)
2041		kfd_dec_compute_active(dqm->dev);
2042
2043	return retval;
2044}
2045
2046static int get_wave_state(struct device_queue_manager *dqm,
2047			  struct queue *q,
2048			  void __user *ctl_stack,
2049			  u32 *ctl_stack_used_size,
2050			  u32 *save_area_used_size)
2051{
2052	struct mqd_manager *mqd_mgr;
2053
2054	dqm_lock(dqm);
2055
2056	mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_CP];
2057
2058	if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE ||
2059	    q->properties.is_active || !q->device->cwsr_enabled ||
2060	    !mqd_mgr->get_wave_state) {
2061		dqm_unlock(dqm);
2062		return -EINVAL;
2063	}
2064
2065	dqm_unlock(dqm);
2066
2067	/*
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
2070	 * lock.
2071	 */
2072	return mqd_mgr->get_wave_state(mqd_mgr, q->mqd, ctl_stack,
2073			ctl_stack_used_size, save_area_used_size);
2074}
2075
2076static void get_queue_checkpoint_info(struct device_queue_manager *dqm,
2077			const struct queue *q,
2078			u32 *mqd_size,
2079			u32 *ctl_stack_size)
2080{
2081	struct mqd_manager *mqd_mgr;
2082	enum KFD_MQD_TYPE mqd_type =
2083			get_mqd_type_from_queue_type(q->properties.type);
2084
2085	dqm_lock(dqm);
2086	mqd_mgr = dqm->mqd_mgrs[mqd_type];
2087	*mqd_size = mqd_mgr->mqd_size;
2088	*ctl_stack_size = 0;
2089
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);
2092
2093	dqm_unlock(dqm);
2094}
2095
2096static int checkpoint_mqd(struct device_queue_manager *dqm,
2097			  const struct queue *q,
2098			  void *mqd,
2099			  void *ctl_stack)
2100{
2101	struct mqd_manager *mqd_mgr;
2102	int r = 0;
2103	enum KFD_MQD_TYPE mqd_type =
2104			get_mqd_type_from_queue_type(q->properties.type);
2105
2106	dqm_lock(dqm);
2107
2108	if (q->properties.is_active || !q->device->cwsr_enabled) {
2109		r = -EINVAL;
2110		goto dqm_unlock;
2111	}
2112
2113	mqd_mgr = dqm->mqd_mgrs[mqd_type];
2114	if (!mqd_mgr->checkpoint_mqd) {
2115		r = -EOPNOTSUPP;
2116		goto dqm_unlock;
2117	}
2118
2119	mqd_mgr->checkpoint_mqd(mqd_mgr, q->mqd, mqd, ctl_stack);
2120
2121dqm_unlock:
2122	dqm_unlock(dqm);
2123	return r;
2124}
2125
2126static int process_termination_cpsch(struct device_queue_manager *dqm,
2127		struct qcm_process_device *qpd)
2128{
2129	int retval;
2130	struct queue *q;
 
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;
2136	bool found = false;
2137
2138	retval = 0;
2139
2140	dqm_lock(dqm);
2141
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;
2149	}
2150
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);
2157
2158		if (q->properties.is_active) {
2159			decrement_queue_count(dqm, qpd, q);
2160
2161			if (dqm->dev->shared_resources.enable_mes) {
2162				retval = remove_queue_mes(dqm, q, qpd);
2163				if (retval)
2164					pr_err("Failed to remove queue %d\n",
2165						q->properties.queue_id);
2166			}
2167		}
2168
2169		dqm->total_queue_count--;
2170	}
2171
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);
2176			kfree(cur);
2177			dqm->processes_count--;
2178			found = true;
2179			break;
2180		}
2181	}
2182
2183	if (!dqm->dev->shared_resources.enable_mes)
2184		retval = execute_queues_cpsch(dqm, filter, 0);
2185
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;
2190	}
2191
2192	/* Lastly, free mqd resources.
2193	 * Do free_mqd() after dqm_unlock to avoid circular locking.
2194	 */
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)];
2199		list_del(&q->list);
2200		qpd->queue_count--;
2201		dqm_unlock(dqm);
2202		mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
2203		dqm_lock(dqm);
2204	}
2205	dqm_unlock(dqm);
2206
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.
2209	 */
2210	if (found)
2211		kfd_dec_compute_active(dqm->dev);
2212
2213	return retval;
2214}
2215
2216static int init_mqd_managers(struct device_queue_manager *dqm)
2217{
2218	int i, j;
 
2219	struct mqd_manager *mqd_mgr;
2220
2221	for (i = 0; i < KFD_MQD_TYPE_MAX; i++) {
2222		mqd_mgr = dqm->asic_ops.mqd_manager_init(i, dqm->dev);
2223		if (!mqd_mgr) {
2224			pr_err("mqd manager [%d] initialization failed\n", i);
2225			goto out_free;
2226		}
2227		dqm->mqd_mgrs[i] = mqd_mgr;
2228	}
2229
2230	return 0;
2231
2232out_free:
2233	for (j = 0; j < i; j++) {
2234		kfree(dqm->mqd_mgrs[j]);
2235		dqm->mqd_mgrs[j] = NULL;
2236	}
2237
2238	return -ENOMEM;
2239}
2240
2241/* Allocate one hiq mqd (HWS) and all SDMA mqd in a continuous trunk*/
2242static int allocate_hiq_sdma_mqd(struct device_queue_manager *dqm)
2243{
2244	int retval;
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;
 
2251
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);
2255
2256	return retval;
2257}
2258
2259struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
2260{
2261	struct device_queue_manager *dqm;
2262
2263	pr_debug("Loading device queue manager\n");
2264
2265	dqm = kzalloc(sizeof(*dqm), GFP_KERNEL);
2266	if (!dqm)
2267		return NULL;
2268
2269	switch (dev->adev->asic_type) {
2270	/* HWS is not available on Hawaii. */
2271	case CHIP_HAWAII:
2272	/* HWS depends on CWSR for timely dequeue. CWSR is not
2273	 * available on Tonga.
2274	 *
2275	 * FIXME: This argument also applies to Kaveri.
2276	 */
2277	case CHIP_TONGA:
2278		dqm->sched_policy = KFD_SCHED_POLICY_NO_HWS;
2279		break;
2280	default:
2281		dqm->sched_policy = sched_policy;
2282		break;
2283	}
2284
2285	dqm->dev = dev;
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;
2310		break;
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;
2331		break;
2332	default:
2333		pr_err("Invalid scheduling policy %d\n", dqm->sched_policy);
2334		goto out_free;
2335	}
2336
2337	switch (dev->adev->asic_type) {
2338	case CHIP_CARRIZO:
2339		device_queue_manager_init_vi(&dqm->asic_ops);
2340		break;
2341
2342	case CHIP_KAVERI:
 
2343		device_queue_manager_init_cik(&dqm->asic_ops);
2344		break;
2345
2346	case CHIP_HAWAII:
2347		device_queue_manager_init_cik_hawaii(&dqm->asic_ops);
2348		break;
2349
2350	case CHIP_TONGA:
2351	case CHIP_FIJI:
2352	case CHIP_POLARIS10:
2353	case CHIP_POLARIS11:
2354	case CHIP_POLARIS12:
2355	case CHIP_VEGAM:
2356		device_queue_manager_init_vi_tonga(&dqm->asic_ops);
2357		break;
2358
2359	default:
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);
2366		else {
2367			WARN(1, "Unexpected ASIC family %u",
2368			     dev->adev->asic_type);
2369			goto out_free;
2370		}
2371	}
2372
2373	if (init_mqd_managers(dqm))
2374		goto out_free;
2375
2376	if (allocate_hiq_sdma_mqd(dqm)) {
2377		pr_err("Failed to allocate hiq sdma mqd trunk buffer\n");
2378		goto out_free;
2379	}
2380
2381	if (!dqm->ops.initialize(dqm))
 
2382		return dqm;
 
2383
2384out_free:
2385	kfree(dqm);
2386	return NULL;
2387}
2388
2389static void deallocate_hiq_sdma_mqd(struct kfd_dev *dev,
2390				    struct kfd_mem_obj *mqd)
2391{
2392	WARN(!mqd, "No hiq sdma mqd trunk to free");
2393
2394	amdgpu_amdkfd_free_gtt_mem(dev->adev, mqd->gtt_mem);
2395}
2396
2397void device_queue_manager_uninit(struct device_queue_manager *dqm)
2398{
 
2399	dqm->ops.uninitialize(dqm);
2400	deallocate_hiq_sdma_mqd(dqm->dev, &dqm->hiq_sdma_mqd);
 
2401	kfree(dqm);
2402}
2403
2404int kfd_dqm_evict_pasid(struct device_queue_manager *dqm, u32 pasid)
2405{
2406	struct kfd_process_device *pdd;
2407	struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
2408	int ret = 0;
2409
2410	if (!p)
2411		return -EINVAL;
2412	WARN(debug_evictions, "Evicting pid %d", p->lead_thread->pid);
2413	pdd = kfd_get_process_device_data(dqm->dev, p);
2414	if (pdd)
2415		ret = dqm->ops.evict_process_queues(dqm, &pdd->qpd);
2416	kfd_unref_process(p);
2417
2418	return ret;
2419}
2420
2421static void kfd_process_hw_exception(struct work_struct *work)
2422{
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);
2426}
2427
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2428#if defined(CONFIG_DEBUG_FS)
2429
2430static void seq_reg_dump(struct seq_file *m,
2431			 uint32_t (*dump)[2], uint32_t n_regs)
2432{
2433	uint32_t i, count;
2434
2435	for (i = 0, count = 0; i < n_regs; i++) {
2436		if (count == 0 ||
2437		    dump[i-1][0] + sizeof(uint32_t) != dump[i][0]) {
2438			seq_printf(m, "%s    %08x: %08x",
2439				   i ? "\n" : "",
2440				   dump[i][0], dump[i][1]);
2441			count = 7;
2442		} else {
2443			seq_printf(m, " %08x", dump[i][1]);
2444			count--;
2445		}
2446	}
2447
2448	seq_puts(m, "\n");
2449}
2450
2451int dqm_debugfs_hqds(struct seq_file *m, void *data)
2452{
2453	struct device_queue_manager *dqm = data;
 
2454	uint32_t (*dump)[2], n_regs;
2455	int pipe, queue;
2456	int r = 0;
 
2457
2458	if (!dqm->sched_running) {
2459		seq_puts(m, " Device is stopped\n");
2460		return 0;
2461	}
2462
2463	r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->adev,
2464					KFD_CIK_HIQ_PIPE, KFD_CIK_HIQ_QUEUE,
2465					&dump, &n_regs);
2466	if (!r) {
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),
2470			   KFD_CIK_HIQ_QUEUE);
2471		seq_reg_dump(m, dump, n_regs);
 
 
 
 
 
2472
2473		kfree(dump);
2474	}
2475
2476	for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) {
2477		int pipe_offset = pipe * get_queues_per_pipe(dqm);
2478
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))
2482				continue;
2483
2484			r = dqm->dev->kfd2kgd->hqd_dump(
2485				dqm->dev->adev, pipe, queue, &dump, &n_regs);
2486			if (r)
2487				break;
 
 
 
 
 
 
 
 
 
 
 
 
2488
2489			seq_printf(m, "  CP Pipe %d, Queue %d\n",
2490				  pipe, queue);
2491			seq_reg_dump(m, dump, n_regs);
2492
2493			kfree(dump);
2494		}
2495	}
2496
2497	for (pipe = 0; pipe < get_num_all_sdma_engines(dqm); pipe++) {
 
 
 
2498		for (queue = 0;
2499		     queue < dqm->dev->device_info.num_sdma_queues_per_engine;
2500		     queue++) {
2501			r = dqm->dev->kfd2kgd->hqd_sdma_dump(
2502				dqm->dev->adev, pipe, queue, &dump, &n_regs);
2503			if (r)
2504				break;
2505
2506			seq_printf(m, "  SDMA Engine %d, RLC %d\n",
2507				  pipe, queue);
2508			seq_reg_dump(m, dump, n_regs);
2509
2510			kfree(dump);
2511		}
2512	}
2513
2514	return r;
2515}
2516
2517int dqm_debugfs_hang_hws(struct device_queue_manager *dqm)
2518{
2519	int r = 0;
2520
2521	dqm_lock(dqm);
2522	r = pm_debugfs_hang_hws(&dqm->packet_mgr);
2523	if (r) {
2524		dqm_unlock(dqm);
2525		return r;
2526	}
2527	dqm->active_runlist = true;
2528	r = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
 
2529	dqm_unlock(dqm);
2530
2531	return r;
2532}
2533
2534#endif