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
v5.4
 
   1/*
   2 * Copyright 2014 Advanced Micro Devices, Inc.
   3 *
   4 * Permission is hereby granted, free of charge, to any person obtaining a
   5 * copy of this software and associated documentation files (the "Software"),
   6 * to deal in the Software without restriction, including without limitation
   7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8 * and/or sell copies of the Software, and to permit persons to whom the
   9 * Software is furnished to do so, subject to the following conditions:
  10 *
  11 * The above copyright notice and this permission notice shall be included in
  12 * all copies or substantial portions of the Software.
  13 *
  14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20 * OTHER DEALINGS IN THE SOFTWARE.
  21 *
  22 */
  23
  24#include <linux/ratelimit.h>
  25#include <linux/printk.h>
  26#include <linux/slab.h>
  27#include <linux/list.h>
  28#include <linux/types.h>
  29#include <linux/bitops.h>
  30#include <linux/sched.h>
  31#include "kfd_priv.h"
  32#include "kfd_device_queue_manager.h"
  33#include "kfd_mqd_manager.h"
  34#include "cik_regs.h"
  35#include "kfd_kernel_queue.h"
  36#include "amdgpu_amdkfd.h"
 
 
  37
  38/* Size of the per-pipe EOP queue */
  39#define CIK_HPD_EOP_BYTES_LOG2 11
  40#define CIK_HPD_EOP_BYTES (1U << CIK_HPD_EOP_BYTES_LOG2)
  41
  42static int set_pasid_vmid_mapping(struct device_queue_manager *dqm,
  43					unsigned int pasid, unsigned int vmid);
  44
  45static int execute_queues_cpsch(struct device_queue_manager *dqm,
  46				enum kfd_unmap_queues_filter filter,
  47				uint32_t filter_param);
 
  48static int unmap_queues_cpsch(struct device_queue_manager *dqm,
  49				enum kfd_unmap_queues_filter filter,
  50				uint32_t filter_param);
 
 
  51
  52static int map_queues_cpsch(struct device_queue_manager *dqm);
  53
  54static void deallocate_sdma_queue(struct device_queue_manager *dqm,
  55				struct queue *q);
  56
  57static inline void deallocate_hqd(struct device_queue_manager *dqm,
  58				struct queue *q);
  59static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q);
  60static int allocate_sdma_queue(struct device_queue_manager *dqm,
  61				struct queue *q);
  62static void kfd_process_hw_exception(struct work_struct *work);
  63
  64static inline
  65enum KFD_MQD_TYPE get_mqd_type_from_queue_type(enum kfd_queue_type type)
  66{
  67	if (type == KFD_QUEUE_TYPE_SDMA || type == KFD_QUEUE_TYPE_SDMA_XGMI)
  68		return KFD_MQD_TYPE_SDMA;
  69	return KFD_MQD_TYPE_CP;
  70}
  71
  72static bool is_pipe_enabled(struct device_queue_manager *dqm, int mec, int pipe)
  73{
  74	int i;
  75	int pipe_offset = mec * dqm->dev->shared_resources.num_pipe_per_mec
  76		+ pipe * dqm->dev->shared_resources.num_queue_per_pipe;
  77
  78	/* queue is available for KFD usage if bit is 1 */
  79	for (i = 0; i <  dqm->dev->shared_resources.num_queue_per_pipe; ++i)
  80		if (test_bit(pipe_offset + i,
  81			      dqm->dev->shared_resources.queue_bitmap))
  82			return true;
  83	return false;
  84}
  85
  86unsigned int get_queues_num(struct device_queue_manager *dqm)
  87{
  88	return bitmap_weight(dqm->dev->shared_resources.queue_bitmap,
  89				KGD_MAX_QUEUES);
  90}
  91
  92unsigned int get_queues_per_pipe(struct device_queue_manager *dqm)
  93{
  94	return dqm->dev->shared_resources.num_queue_per_pipe;
  95}
  96
  97unsigned int get_pipes_per_mec(struct device_queue_manager *dqm)
  98{
  99	return dqm->dev->shared_resources.num_pipe_per_mec;
 100}
 101
 102static unsigned int get_num_sdma_engines(struct device_queue_manager *dqm)
 103{
 104	return dqm->dev->device_info->num_sdma_engines;
 
 105}
 106
 107static unsigned int get_num_xgmi_sdma_engines(struct device_queue_manager *dqm)
 108{
 109	return dqm->dev->device_info->num_xgmi_sdma_engines;
 
 110}
 111
 112unsigned int get_num_sdma_queues(struct device_queue_manager *dqm)
 113{
 114	return dqm->dev->device_info->num_sdma_engines
 115			* dqm->dev->device_info->num_sdma_queues_per_engine;
 116}
 117
 118unsigned int get_num_xgmi_sdma_queues(struct device_queue_manager *dqm)
 119{
 120	return dqm->dev->device_info->num_xgmi_sdma_engines
 121			* dqm->dev->device_info->num_sdma_queues_per_engine;
 
 
 
 
 
 
 
 
 122}
 123
 124void program_sh_mem_settings(struct device_queue_manager *dqm,
 125					struct qcm_process_device *qpd)
 126{
 127	return dqm->dev->kfd2kgd->program_sh_mem_settings(
 128						dqm->dev->kgd, qpd->vmid,
 129						qpd->sh_mem_config,
 130						qpd->sh_mem_ape1_base,
 131						qpd->sh_mem_ape1_limit,
 132						qpd->sh_mem_bases);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 133}
 134
 135static int allocate_doorbell(struct qcm_process_device *qpd, struct queue *q)
 
 
 
 
 
 
 136{
 137	struct kfd_dev *dev = qpd->dqm->dev;
 138
 139	if (!KFD_IS_SOC15(dev->device_info->asic_family)) {
 140		/* On pre-SOC15 chips we need to use the queue ID to
 141		 * preserve the user mode ABI.
 142		 */
 
 
 
 
 143		q->doorbell_id = q->properties.queue_id;
 144	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
 145			q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
 146		/* For SDMA queues on SOC15 with 8-byte doorbell, use static
 147		 * doorbell assignments based on the engine and queue id.
 148		 * The doobell index distance between RLC (2*i) and (2*i+1)
 149		 * for a SDMA engine is 512.
 150		 */
 151		uint32_t *idx_offset =
 152				dev->shared_resources.sdma_doorbell_idx;
 153
 154		q->doorbell_id = idx_offset[q->properties.sdma_engine_id]
 155			+ (q->properties.sdma_queue_id & 1)
 156			* KFD_QUEUE_DOORBELL_MIRROR_OFFSET
 157			+ (q->properties.sdma_queue_id >> 1);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 158	} else {
 159		/* For CP queues on SOC15 reserve a free doorbell ID */
 160		unsigned int found;
 
 
 
 
 
 
 
 
 161
 162		found = find_first_zero_bit(qpd->doorbell_bitmap,
 163					    KFD_MAX_NUM_OF_QUEUES_PER_PROCESS);
 164		if (found >= KFD_MAX_NUM_OF_QUEUES_PER_PROCESS) {
 165			pr_debug("No doorbells available");
 166			return -EBUSY;
 167		}
 168		set_bit(found, qpd->doorbell_bitmap);
 169		q->doorbell_id = found;
 170	}
 171
 172	q->properties.doorbell_off =
 173		kfd_doorbell_id_to_offset(dev, q->process,
 174					  q->doorbell_id);
 175
 
 
 
 
 176	return 0;
 177}
 178
 179static void deallocate_doorbell(struct qcm_process_device *qpd,
 180				struct queue *q)
 181{
 182	unsigned int old;
 183	struct kfd_dev *dev = qpd->dqm->dev;
 184
 185	if (!KFD_IS_SOC15(dev->device_info->asic_family) ||
 186	    q->properties.type == KFD_QUEUE_TYPE_SDMA ||
 187	    q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
 188		return;
 189
 190	old = test_and_clear_bit(q->doorbell_id, qpd->doorbell_bitmap);
 191	WARN_ON(!old);
 192}
 193
 
 
 
 
 
 
 
 
 
 
 
 
 
 194static int allocate_vmid(struct device_queue_manager *dqm,
 195			struct qcm_process_device *qpd,
 196			struct queue *q)
 197{
 198	int bit, allocated_vmid;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 199
 200	if (dqm->vmid_bitmap == 0)
 201		return -ENOMEM;
 202
 203	bit = ffs(dqm->vmid_bitmap) - 1;
 204	dqm->vmid_bitmap &= ~(1 << bit);
 205
 206	allocated_vmid = bit + dqm->dev->vm_info.first_vmid_kfd;
 207	pr_debug("vmid allocation %d\n", allocated_vmid);
 208	qpd->vmid = allocated_vmid;
 209	q->properties.vmid = allocated_vmid;
 210
 211	set_pasid_vmid_mapping(dqm, q->process->pasid, q->properties.vmid);
 212	program_sh_mem_settings(dqm, qpd);
 213
 
 
 
 214	/* qpd->page_table_base is set earlier when register_process()
 215	 * is called, i.e. when the first queue is created.
 216	 */
 217	dqm->dev->kfd2kgd->set_vm_context_page_table_base(dqm->dev->kgd,
 218			qpd->vmid,
 219			qpd->page_table_base);
 220	/* invalidate the VM context after pasid and vmid mapping is set up */
 221	kfd_flush_tlb(qpd_to_pdd(qpd));
 222
 223	dqm->dev->kfd2kgd->set_scratch_backing_va(
 224		dqm->dev->kgd, qpd->sh_hidden_private_base, qpd->vmid);
 
 225
 226	return 0;
 227}
 228
 229static int flush_texture_cache_nocpsch(struct kfd_dev *kdev,
 230				struct qcm_process_device *qpd)
 231{
 232	const struct packet_manager_funcs *pmf = qpd->dqm->packets.pmf;
 233	int ret;
 234
 235	if (!qpd->ib_kaddr)
 236		return -ENOMEM;
 237
 238	ret = pmf->release_mem(qpd->ib_base, (uint32_t *)qpd->ib_kaddr);
 239	if (ret)
 240		return ret;
 241
 242	return amdgpu_amdkfd_submit_ib(kdev->kgd, KGD_ENGINE_MEC1, qpd->vmid,
 243				qpd->ib_base, (uint32_t *)qpd->ib_kaddr,
 244				pmf->release_mem_size / sizeof(uint32_t));
 245}
 246
 247static void deallocate_vmid(struct device_queue_manager *dqm,
 248				struct qcm_process_device *qpd,
 249				struct queue *q)
 250{
 251	int bit = qpd->vmid - dqm->dev->vm_info.first_vmid_kfd;
 252
 253	/* On GFX v7, CP doesn't flush TC at dequeue */
 254	if (q->device->device_info->asic_family == CHIP_HAWAII)
 255		if (flush_texture_cache_nocpsch(q->device, qpd))
 256			pr_err("Failed to flush TC\n");
 257
 258	kfd_flush_tlb(qpd_to_pdd(qpd));
 259
 260	/* Release the vmid mapping */
 261	set_pasid_vmid_mapping(dqm, 0, qpd->vmid);
 
 262
 263	dqm->vmid_bitmap |= (1 << bit);
 264	qpd->vmid = 0;
 265	q->properties.vmid = 0;
 266}
 267
 268static int create_queue_nocpsch(struct device_queue_manager *dqm,
 269				struct queue *q,
 270				struct qcm_process_device *qpd)
 
 
 271{
 272	struct mqd_manager *mqd_mgr;
 273	int retval;
 274
 275	print_queue(q);
 276
 277	dqm_lock(dqm);
 278
 279	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
 280		pr_warn("Can't create new usermode queue because %d queues were already created\n",
 281				dqm->total_queue_count);
 282		retval = -EPERM;
 283		goto out_unlock;
 284	}
 285
 286	if (list_empty(&qpd->queues_list)) {
 287		retval = allocate_vmid(dqm, qpd, q);
 288		if (retval)
 289			goto out_unlock;
 290	}
 291	q->properties.vmid = qpd->vmid;
 292	/*
 293	 * Eviction state logic: mark all queues as evicted, even ones
 294	 * not currently active. Restoring inactive queues later only
 295	 * updates the is_evicted flag but is a no-op otherwise.
 296	 */
 297	q->properties.is_evicted = !!qpd->evicted;
 298
 299	q->properties.tba_addr = qpd->tba_addr;
 300	q->properties.tma_addr = qpd->tma_addr;
 301
 302	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
 303			q->properties.type)];
 304	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) {
 305		retval = allocate_hqd(dqm, q);
 306		if (retval)
 307			goto deallocate_vmid;
 308		pr_debug("Loading mqd to hqd on pipe %d, queue %d\n",
 309			q->pipe, q->queue);
 310	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
 311		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
 312		retval = allocate_sdma_queue(dqm, q);
 313		if (retval)
 314			goto deallocate_vmid;
 315		dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
 316	}
 317
 318	retval = allocate_doorbell(qpd, q);
 319	if (retval)
 320		goto out_deallocate_hqd;
 321
 322	/* Temporarily release dqm lock to avoid a circular lock dependency */
 323	dqm_unlock(dqm);
 324	q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties);
 325	dqm_lock(dqm);
 326
 327	if (!q->mqd_mem_obj) {
 328		retval = -ENOMEM;
 329		goto out_deallocate_doorbell;
 330	}
 331	mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj,
 332				&q->gart_mqd_addr, &q->properties);
 
 
 
 
 
 
 
 333	if (q->properties.is_active) {
 
 
 
 
 334
 335		if (WARN(q->process->mm != current->mm,
 336					"should only run in user thread"))
 337			retval = -EFAULT;
 338		else
 339			retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe,
 340					q->queue, &q->properties, current->mm);
 341		if (retval)
 342			goto out_free_mqd;
 343	}
 344
 
 345	list_add(&q->list, &qpd->queues_list);
 346	qpd->queue_count++;
 347	if (q->properties.is_active)
 348		dqm->queue_count++;
 349
 350	if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
 351		dqm->sdma_queue_count++;
 352	else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
 353		dqm->xgmi_sdma_queue_count++;
 354
 355	/*
 356	 * Unconditionally increment this counter, regardless of the queue's
 357	 * type or whether the queue is active.
 358	 */
 359	dqm->total_queue_count++;
 360	pr_debug("Total of %d queues are accountable so far\n",
 361			dqm->total_queue_count);
 362	goto out_unlock;
 363
 364out_free_mqd:
 365	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
 366out_deallocate_doorbell:
 367	deallocate_doorbell(qpd, q);
 368out_deallocate_hqd:
 369	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE)
 370		deallocate_hqd(dqm, q);
 371	else if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
 372		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
 373		deallocate_sdma_queue(dqm, q);
 374deallocate_vmid:
 375	if (list_empty(&qpd->queues_list))
 376		deallocate_vmid(dqm, qpd, q);
 377out_unlock:
 378	dqm_unlock(dqm);
 379	return retval;
 380}
 381
 382static int allocate_hqd(struct device_queue_manager *dqm, struct queue *q)
 383{
 384	bool set;
 385	int pipe, bit, i;
 386
 387	set = false;
 388
 389	for (pipe = dqm->next_pipe_to_allocate, i = 0;
 390			i < get_pipes_per_mec(dqm);
 391			pipe = ((pipe + 1) % get_pipes_per_mec(dqm)), ++i) {
 392
 393		if (!is_pipe_enabled(dqm, 0, pipe))
 394			continue;
 395
 396		if (dqm->allocated_queues[pipe] != 0) {
 397			bit = ffs(dqm->allocated_queues[pipe]) - 1;
 398			dqm->allocated_queues[pipe] &= ~(1 << bit);
 399			q->pipe = pipe;
 400			q->queue = bit;
 401			set = true;
 402			break;
 403		}
 404	}
 405
 406	if (!set)
 407		return -EBUSY;
 408
 409	pr_debug("hqd slot - pipe %d, queue %d\n", q->pipe, q->queue);
 410	/* horizontal hqd allocation */
 411	dqm->next_pipe_to_allocate = (pipe + 1) % get_pipes_per_mec(dqm);
 412
 413	return 0;
 414}
 415
 416static inline void deallocate_hqd(struct device_queue_manager *dqm,
 417				struct queue *q)
 418{
 419	dqm->allocated_queues[q->pipe] |= (1 << q->queue);
 420}
 421
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 422/* Access to DQM has to be locked before calling destroy_queue_nocpsch_locked
 423 * to avoid asynchronized access
 424 */
 425static int destroy_queue_nocpsch_locked(struct device_queue_manager *dqm,
 426				struct qcm_process_device *qpd,
 427				struct queue *q)
 428{
 429	int retval;
 430	struct mqd_manager *mqd_mgr;
 431
 432	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
 433			q->properties.type)];
 434
 435	if (q->properties.type == KFD_QUEUE_TYPE_COMPUTE) {
 436		deallocate_hqd(dqm, q);
 437	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
 438		dqm->sdma_queue_count--;
 439		deallocate_sdma_queue(dqm, q);
 440	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
 441		dqm->xgmi_sdma_queue_count--;
 442		deallocate_sdma_queue(dqm, q);
 443	} else {
 444		pr_debug("q->properties.type %d is invalid\n",
 445				q->properties.type);
 446		return -EINVAL;
 447	}
 448	dqm->total_queue_count--;
 449
 450	deallocate_doorbell(qpd, q);
 451
 
 
 
 
 
 452	retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
 453				KFD_PREEMPT_TYPE_WAVEFRONT_RESET,
 454				KFD_UNMAP_LATENCY_MS,
 455				q->pipe, q->queue);
 456	if (retval == -ETIME)
 457		qpd->reset_wavefronts = true;
 458
 459	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
 460
 461	list_del(&q->list);
 462	if (list_empty(&qpd->queues_list)) {
 463		if (qpd->reset_wavefronts) {
 464			pr_warn("Resetting wave fronts (nocpsch) on dev %p\n",
 465					dqm->dev);
 466			/* dbgdev_wave_reset_wavefronts has to be called before
 467			 * deallocate_vmid(), i.e. when vmid is still in use.
 468			 */
 469			dbgdev_wave_reset_wavefronts(dqm->dev,
 470					qpd->pqm->process);
 471			qpd->reset_wavefronts = false;
 472		}
 473
 474		deallocate_vmid(dqm, qpd, q);
 475	}
 476	qpd->queue_count--;
 477	if (q->properties.is_active)
 478		dqm->queue_count--;
 479
 480	return retval;
 481}
 482
 483static int destroy_queue_nocpsch(struct device_queue_manager *dqm,
 484				struct qcm_process_device *qpd,
 485				struct queue *q)
 486{
 487	int retval;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 488
 489	dqm_lock(dqm);
 490	retval = destroy_queue_nocpsch_locked(dqm, qpd, q);
 
 
 491	dqm_unlock(dqm);
 492
 
 
 493	return retval;
 494}
 495
 496static int update_queue(struct device_queue_manager *dqm, struct queue *q)
 
 497{
 498	int retval = 0;
 
 499	struct mqd_manager *mqd_mgr;
 500	struct kfd_process_device *pdd;
 501	bool prev_active = false;
 502
 503	dqm_lock(dqm);
 504	pdd = kfd_get_process_device_data(q->device, q->process);
 505	if (!pdd) {
 506		retval = -ENODEV;
 507		goto out_unlock;
 508	}
 509	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
 510			q->properties.type)];
 511
 512	/* Save previous activity state for counters */
 513	prev_active = q->properties.is_active;
 514
 515	/* Make sure the queue is unmapped before updating the MQD */
 516	if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS) {
 517		retval = unmap_queues_cpsch(dqm,
 518				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
 
 
 
 
 519		if (retval) {
 520			pr_err("unmap queue failed\n");
 521			goto out_unlock;
 522		}
 523	} else if (prev_active &&
 524		   (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
 525		    q->properties.type == KFD_QUEUE_TYPE_SDMA ||
 526		    q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
 
 
 
 
 
 
 527		retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
 528				KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN,
 
 
 529				KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
 530		if (retval) {
 531			pr_err("destroy mqd failed\n");
 532			goto out_unlock;
 533		}
 534	}
 535
 536	mqd_mgr->update_mqd(mqd_mgr, q->mqd, &q->properties);
 537
 538	/*
 539	 * check active state vs. the previous state and modify
 540	 * counter accordingly. map_queues_cpsch uses the
 541	 * dqm->queue_count to determine whether a new runlist must be
 542	 * uploaded.
 543	 */
 544	if (q->properties.is_active && !prev_active)
 545		dqm->queue_count++;
 546	else if (!q->properties.is_active && prev_active)
 547		dqm->queue_count--;
 548
 549	if (dqm->sched_policy != KFD_SCHED_POLICY_NO_HWS)
 550		retval = map_queues_cpsch(dqm);
 551	else if (q->properties.is_active &&
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 552		 (q->properties.type == KFD_QUEUE_TYPE_COMPUTE ||
 553		  q->properties.type == KFD_QUEUE_TYPE_SDMA ||
 554		  q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)) {
 555		if (WARN(q->process->mm != current->mm,
 556			 "should only run in user thread"))
 557			retval = -EFAULT;
 558		else
 559			retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd,
 560						   q->pipe, q->queue,
 561						   &q->properties, current->mm);
 562	}
 563
 564out_unlock:
 565	dqm_unlock(dqm);
 566	return retval;
 567}
 568
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 569static int evict_process_queues_nocpsch(struct device_queue_manager *dqm,
 570					struct qcm_process_device *qpd)
 571{
 572	struct queue *q;
 573	struct mqd_manager *mqd_mgr;
 574	struct kfd_process_device *pdd;
 575	int retval, ret = 0;
 576
 577	dqm_lock(dqm);
 578	if (qpd->evicted++ > 0) /* already evicted, do nothing */
 579		goto out;
 580
 581	pdd = qpd_to_pdd(qpd);
 582	pr_info_ratelimited("Evicting PASID %u queues\n",
 583			    pdd->process->pasid);
 584
 
 585	/* Mark all queues as evicted. Deactivate all active queues on
 586	 * the qpd.
 587	 */
 588	list_for_each_entry(q, &qpd->queues_list, list) {
 589		q->properties.is_evicted = true;
 590		if (!q->properties.is_active)
 591			continue;
 592
 593		mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
 594				q->properties.type)];
 595		q->properties.is_active = false;
 
 
 
 
 
 596		retval = mqd_mgr->destroy_mqd(mqd_mgr, q->mqd,
 597				KFD_PREEMPT_TYPE_WAVEFRONT_DRAIN,
 
 
 598				KFD_UNMAP_LATENCY_MS, q->pipe, q->queue);
 599		if (retval && !ret)
 600			/* Return the first error, but keep going to
 601			 * maintain a consistent eviction state
 602			 */
 603			ret = retval;
 604		dqm->queue_count--;
 605	}
 606
 607out:
 608	dqm_unlock(dqm);
 609	return ret;
 610}
 611
 612static int evict_process_queues_cpsch(struct device_queue_manager *dqm,
 613				      struct qcm_process_device *qpd)
 614{
 615	struct queue *q;
 
 616	struct kfd_process_device *pdd;
 617	int retval = 0;
 618
 619	dqm_lock(dqm);
 620	if (qpd->evicted++ > 0) /* already evicted, do nothing */
 621		goto out;
 622
 623	pdd = qpd_to_pdd(qpd);
 624	pr_info_ratelimited("Evicting PASID %u queues\n",
 
 
 
 
 
 
 
 
 625			    pdd->process->pasid);
 626
 627	/* Mark all queues as evicted. Deactivate all active queues on
 628	 * the qpd.
 629	 */
 630	list_for_each_entry(q, &qpd->queues_list, list) {
 631		q->properties.is_evicted = true;
 632		if (!q->properties.is_active)
 633			continue;
 634
 635		q->properties.is_active = false;
 636		dqm->queue_count--;
 
 
 
 
 
 
 
 
 
 637	}
 638	retval = execute_queues_cpsch(dqm,
 639				qpd->is_debug ?
 640				KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES :
 641				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
 
 
 
 642
 643out:
 644	dqm_unlock(dqm);
 645	return retval;
 646}
 647
 648static int restore_process_queues_nocpsch(struct device_queue_manager *dqm,
 649					  struct qcm_process_device *qpd)
 650{
 651	struct mm_struct *mm = NULL;
 652	struct queue *q;
 653	struct mqd_manager *mqd_mgr;
 654	struct kfd_process_device *pdd;
 655	uint64_t pd_base;
 
 656	int retval, ret = 0;
 657
 658	pdd = qpd_to_pdd(qpd);
 659	/* Retrieve PD base */
 660	pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->vm);
 661
 662	dqm_lock(dqm);
 663	if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */
 664		goto out;
 665	if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
 666		qpd->evicted--;
 667		goto out;
 668	}
 669
 670	pr_info_ratelimited("Restoring PASID %u queues\n",
 671			    pdd->process->pasid);
 672
 673	/* Update PD Base in QPD */
 674	qpd->page_table_base = pd_base;
 675	pr_debug("Updated PD address to 0x%llx\n", pd_base);
 676
 677	if (!list_empty(&qpd->queues_list)) {
 678		dqm->dev->kfd2kgd->set_vm_context_page_table_base(
 679				dqm->dev->kgd,
 680				qpd->vmid,
 681				qpd->page_table_base);
 682		kfd_flush_tlb(pdd);
 683	}
 684
 685	/* Take a safe reference to the mm_struct, which may otherwise
 686	 * disappear even while the kfd_process is still referenced.
 687	 */
 688	mm = get_task_mm(pdd->process->lead_thread);
 689	if (!mm) {
 690		ret = -EFAULT;
 691		goto out;
 692	}
 693
 694	/* Remove the eviction flags. Activate queues that are not
 695	 * inactive for other reasons.
 696	 */
 697	list_for_each_entry(q, &qpd->queues_list, list) {
 698		q->properties.is_evicted = false;
 699		if (!QUEUE_IS_ACTIVE(q->properties))
 700			continue;
 701
 702		mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
 703				q->properties.type)];
 704		q->properties.is_active = true;
 
 
 
 
 
 705		retval = mqd_mgr->load_mqd(mqd_mgr, q->mqd, q->pipe,
 706				       q->queue, &q->properties, mm);
 707		if (retval && !ret)
 708			/* Return the first error, but keep going to
 709			 * maintain a consistent eviction state
 710			 */
 711			ret = retval;
 712		dqm->queue_count++;
 713	}
 714	qpd->evicted = 0;
 
 
 715out:
 716	if (mm)
 717		mmput(mm);
 718	dqm_unlock(dqm);
 719	return ret;
 720}
 721
 722static int restore_process_queues_cpsch(struct device_queue_manager *dqm,
 723					struct qcm_process_device *qpd)
 724{
 725	struct queue *q;
 
 726	struct kfd_process_device *pdd;
 727	uint64_t pd_base;
 728	int retval = 0;
 729
 730	pdd = qpd_to_pdd(qpd);
 731	/* Retrieve PD base */
 732	pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->vm);
 733
 734	dqm_lock(dqm);
 735	if (WARN_ON_ONCE(!qpd->evicted)) /* already restored, do nothing */
 736		goto out;
 737	if (qpd->evicted > 1) { /* ref count still > 0, decrement & quit */
 738		qpd->evicted--;
 739		goto out;
 740	}
 741
 742	pr_info_ratelimited("Restoring PASID %u queues\n",
 
 
 
 
 
 
 
 743			    pdd->process->pasid);
 744
 745	/* Update PD Base in QPD */
 746	qpd->page_table_base = pd_base;
 747	pr_debug("Updated PD address to 0x%llx\n", pd_base);
 748
 749	/* activate all active queues on the qpd */
 750	list_for_each_entry(q, &qpd->queues_list, list) {
 751		q->properties.is_evicted = false;
 752		if (!QUEUE_IS_ACTIVE(q->properties))
 753			continue;
 754
 755		q->properties.is_active = true;
 756		dqm->queue_count++;
 
 
 
 
 
 
 
 
 
 757	}
 758	retval = execute_queues_cpsch(dqm,
 759				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
 
 
 
 
 760	qpd->evicted = 0;
 761out:
 762	dqm_unlock(dqm);
 763	return retval;
 764}
 765
 766static int register_process(struct device_queue_manager *dqm,
 767					struct qcm_process_device *qpd)
 768{
 769	struct device_process_node *n;
 770	struct kfd_process_device *pdd;
 771	uint64_t pd_base;
 772	int retval;
 773
 774	n = kzalloc(sizeof(*n), GFP_KERNEL);
 775	if (!n)
 776		return -ENOMEM;
 777
 778	n->qpd = qpd;
 779
 780	pdd = qpd_to_pdd(qpd);
 781	/* Retrieve PD base */
 782	pd_base = amdgpu_amdkfd_gpuvm_get_process_page_dir(pdd->vm);
 783
 784	dqm_lock(dqm);
 785	list_add(&n->list, &dqm->queues);
 786
 787	/* Update PD Base in QPD */
 788	qpd->page_table_base = pd_base;
 789	pr_debug("Updated PD address to 0x%llx\n", pd_base);
 790
 791	retval = dqm->asic_ops.update_qpd(dqm, qpd);
 792
 793	dqm->processes_count++;
 794
 795	dqm_unlock(dqm);
 796
 797	/* Outside the DQM lock because under the DQM lock we can't do
 798	 * reclaim or take other locks that others hold while reclaiming.
 799	 */
 800	kfd_inc_compute_active(dqm->dev);
 801
 802	return retval;
 803}
 804
 805static int unregister_process(struct device_queue_manager *dqm,
 806					struct qcm_process_device *qpd)
 807{
 808	int retval;
 809	struct device_process_node *cur, *next;
 810
 811	pr_debug("qpd->queues_list is %s\n",
 812			list_empty(&qpd->queues_list) ? "empty" : "not empty");
 813
 814	retval = 0;
 815	dqm_lock(dqm);
 816
 817	list_for_each_entry_safe(cur, next, &dqm->queues, list) {
 818		if (qpd == cur->qpd) {
 819			list_del(&cur->list);
 820			kfree(cur);
 821			dqm->processes_count--;
 822			goto out;
 823		}
 824	}
 825	/* qpd not found in dqm list */
 826	retval = 1;
 827out:
 828	dqm_unlock(dqm);
 829
 830	/* Outside the DQM lock because under the DQM lock we can't do
 831	 * reclaim or take other locks that others hold while reclaiming.
 832	 */
 833	if (!retval)
 834		kfd_dec_compute_active(dqm->dev);
 835
 836	return retval;
 837}
 838
 839static int
 840set_pasid_vmid_mapping(struct device_queue_manager *dqm, unsigned int pasid,
 841			unsigned int vmid)
 842{
 843	return dqm->dev->kfd2kgd->set_pasid_vmid_mapping(
 844						dqm->dev->kgd, pasid, vmid);
 
 
 
 
 
 
 
 
 
 845}
 846
 847static void init_interrupts(struct device_queue_manager *dqm)
 848{
 849	unsigned int i;
 
 850
 851	for (i = 0 ; i < get_pipes_per_mec(dqm) ; i++)
 852		if (is_pipe_enabled(dqm, 0, i))
 853			dqm->dev->kfd2kgd->init_interrupts(dqm->dev->kgd, i);
 
 
 
 
 
 854}
 855
 856static int initialize_nocpsch(struct device_queue_manager *dqm)
 857{
 858	int pipe, queue;
 859
 860	pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm));
 861
 862	dqm->allocated_queues = kcalloc(get_pipes_per_mec(dqm),
 863					sizeof(unsigned int), GFP_KERNEL);
 864	if (!dqm->allocated_queues)
 865		return -ENOMEM;
 866
 867	mutex_init(&dqm->lock_hidden);
 868	INIT_LIST_HEAD(&dqm->queues);
 869	dqm->queue_count = dqm->next_pipe_to_allocate = 0;
 870	dqm->sdma_queue_count = 0;
 871	dqm->xgmi_sdma_queue_count = 0;
 872
 873	for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) {
 874		int pipe_offset = pipe * get_queues_per_pipe(dqm);
 875
 876		for (queue = 0; queue < get_queues_per_pipe(dqm); queue++)
 877			if (test_bit(pipe_offset + queue,
 878				     dqm->dev->shared_resources.queue_bitmap))
 879				dqm->allocated_queues[pipe] |= 1 << queue;
 880	}
 881
 882	dqm->vmid_bitmap = (1 << dqm->dev->vm_info.vmid_num_kfd) - 1;
 883	dqm->sdma_bitmap = ~0ULL >> (64 - get_num_sdma_queues(dqm));
 884	dqm->xgmi_sdma_bitmap = ~0ULL >> (64 - get_num_xgmi_sdma_queues(dqm));
 885
 886	return 0;
 887}
 888
 889static void uninitialize(struct device_queue_manager *dqm)
 890{
 891	int i;
 892
 893	WARN_ON(dqm->queue_count > 0 || dqm->processes_count > 0);
 894
 895	kfree(dqm->allocated_queues);
 896	for (i = 0 ; i < KFD_MQD_TYPE_MAX ; i++)
 897		kfree(dqm->mqd_mgrs[i]);
 898	mutex_destroy(&dqm->lock_hidden);
 899	kfd_gtt_sa_free(dqm->dev, dqm->pipeline_mem);
 900}
 901
 902static int start_nocpsch(struct device_queue_manager *dqm)
 903{
 
 
 
 904	init_interrupts(dqm);
 905	return pm_init(&dqm->packets, dqm);
 
 
 
 
 
 
 906}
 907
 908static int stop_nocpsch(struct device_queue_manager *dqm)
 909{
 910	pm_uninit(&dqm->packets);
 
 
 
 
 
 
 
 
 
 
 911	return 0;
 912}
 913
 
 
 
 
 
 
 
 914static int allocate_sdma_queue(struct device_queue_manager *dqm,
 915				struct queue *q)
 916{
 
 917	int bit;
 918
 919	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
 920		if (dqm->sdma_bitmap == 0)
 
 921			return -ENOMEM;
 922		bit = __ffs64(dqm->sdma_bitmap);
 923		dqm->sdma_bitmap &= ~(1ULL << bit);
 924		q->sdma_id = bit;
 925		q->properties.sdma_engine_id = q->sdma_id %
 926				get_num_sdma_engines(dqm);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 927		q->properties.sdma_queue_id = q->sdma_id /
 928				get_num_sdma_engines(dqm);
 929	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
 930		if (dqm->xgmi_sdma_bitmap == 0)
 
 931			return -ENOMEM;
 932		bit = __ffs64(dqm->xgmi_sdma_bitmap);
 933		dqm->xgmi_sdma_bitmap &= ~(1ULL << bit);
 934		q->sdma_id = bit;
 
 
 
 
 
 
 
 
 
 
 
 
 935		/* sdma_engine_id is sdma id including
 936		 * both PCIe-optimized SDMAs and XGMI-
 937		 * optimized SDMAs. The calculation below
 938		 * assumes the first N engines are always
 939		 * PCIe-optimized ones
 940		 */
 941		q->properties.sdma_engine_id = get_num_sdma_engines(dqm) +
 942				q->sdma_id % get_num_xgmi_sdma_engines(dqm);
 
 943		q->properties.sdma_queue_id = q->sdma_id /
 944				get_num_xgmi_sdma_engines(dqm);
 945	}
 946
 947	pr_debug("SDMA engine id: %d\n", q->properties.sdma_engine_id);
 948	pr_debug("SDMA queue id: %d\n", q->properties.sdma_queue_id);
 949
 950	return 0;
 951}
 952
 953static void deallocate_sdma_queue(struct device_queue_manager *dqm,
 954				struct queue *q)
 955{
 956	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
 957		if (q->sdma_id >= get_num_sdma_queues(dqm))
 958			return;
 959		dqm->sdma_bitmap |= (1ULL << q->sdma_id);
 960	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
 961		if (q->sdma_id >= get_num_xgmi_sdma_queues(dqm))
 962			return;
 963		dqm->xgmi_sdma_bitmap |= (1ULL << q->sdma_id);
 964	}
 965}
 966
 967/*
 968 * Device Queue Manager implementation for cp scheduler
 969 */
 970
 971static int set_sched_resources(struct device_queue_manager *dqm)
 972{
 973	int i, mec;
 974	struct scheduling_resources res;
 
 975
 976	res.vmid_mask = dqm->dev->shared_resources.compute_vmid_bitmap;
 977
 978	res.queue_mask = 0;
 979	for (i = 0; i < KGD_MAX_QUEUES; ++i) {
 980		mec = (i / dqm->dev->shared_resources.num_queue_per_pipe)
 981			/ dqm->dev->shared_resources.num_pipe_per_mec;
 982
 983		if (!test_bit(i, dqm->dev->shared_resources.queue_bitmap))
 984			continue;
 985
 986		/* only acquire queues from the first MEC */
 987		if (mec > 0)
 988			continue;
 989
 990		/* This situation may be hit in the future if a new HW
 991		 * generation exposes more than 64 queues. If so, the
 992		 * definition of res.queue_mask needs updating
 993		 */
 994		if (WARN_ON(i >= (sizeof(res.queue_mask)*8))) {
 995			pr_err("Invalid queue enabled by amdgpu: %d\n", i);
 996			break;
 997		}
 998
 999		res.queue_mask |= (1ull << i);
 
 
1000	}
1001	res.gws_mask = ~0ull;
1002	res.oac_mask = res.gds_heap_base = res.gds_heap_size = 0;
1003
1004	pr_debug("Scheduling resources:\n"
1005			"vmid mask: 0x%8X\n"
1006			"queue mask: 0x%8llX\n",
1007			res.vmid_mask, res.queue_mask);
1008
1009	return pm_send_set_resources(&dqm->packets, &res);
1010}
1011
1012static int initialize_cpsch(struct device_queue_manager *dqm)
1013{
1014	pr_debug("num of pipes: %d\n", get_pipes_per_mec(dqm));
1015
1016	mutex_init(&dqm->lock_hidden);
1017	INIT_LIST_HEAD(&dqm->queues);
1018	dqm->queue_count = dqm->processes_count = 0;
1019	dqm->sdma_queue_count = 0;
1020	dqm->xgmi_sdma_queue_count = 0;
1021	dqm->active_runlist = false;
1022	dqm->sdma_bitmap = ~0ULL >> (64 - get_num_sdma_queues(dqm));
1023	dqm->xgmi_sdma_bitmap = ~0ULL >> (64 - get_num_xgmi_sdma_queues(dqm));
1024
1025	INIT_WORK(&dqm->hw_exception_work, kfd_process_hw_exception);
1026
 
 
 
 
1027	return 0;
1028}
1029
1030static int start_cpsch(struct device_queue_manager *dqm)
1031{
 
1032	int retval;
1033
1034	retval = 0;
1035
1036	retval = pm_init(&dqm->packets, dqm);
1037	if (retval)
1038		goto fail_packet_manager_init;
1039
1040	retval = set_sched_resources(dqm);
1041	if (retval)
1042		goto fail_set_sched_resources;
 
1043
 
 
 
 
1044	pr_debug("Allocating fence memory\n");
1045
1046	/* allocate fence memory on the gart */
1047	retval = kfd_gtt_sa_allocate(dqm->dev, sizeof(*dqm->fence_addr),
1048					&dqm->fence_mem);
1049
1050	if (retval)
1051		goto fail_allocate_vidmem;
1052
1053	dqm->fence_addr = dqm->fence_mem->cpu_ptr;
1054	dqm->fence_gpu_addr = dqm->fence_mem->gpu_addr;
1055
1056	init_interrupts(dqm);
1057
1058	dqm_lock(dqm);
1059	/* clear hang status when driver try to start the hw scheduler */
1060	dqm->is_hws_hang = false;
1061	execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1062	dqm_unlock(dqm);
1063
1064	return 0;
1065fail_allocate_vidmem:
1066fail_set_sched_resources:
1067	pm_uninit(&dqm->packets);
 
1068fail_packet_manager_init:
 
1069	return retval;
1070}
1071
1072static int stop_cpsch(struct device_queue_manager *dqm)
1073{
 
 
1074	dqm_lock(dqm);
1075	unmap_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
1076	dqm_unlock(dqm);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1077
1078	kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
1079	pm_uninit(&dqm->packets);
 
 
1080
1081	return 0;
1082}
1083
1084static int create_kernel_queue_cpsch(struct device_queue_manager *dqm,
1085					struct kernel_queue *kq,
1086					struct qcm_process_device *qpd)
1087{
1088	dqm_lock(dqm);
1089	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
1090		pr_warn("Can't create new kernel queue because %d queues were already created\n",
1091				dqm->total_queue_count);
1092		dqm_unlock(dqm);
1093		return -EPERM;
1094	}
1095
1096	/*
1097	 * Unconditionally increment this counter, regardless of the queue's
1098	 * type or whether the queue is active.
1099	 */
1100	dqm->total_queue_count++;
1101	pr_debug("Total of %d queues are accountable so far\n",
1102			dqm->total_queue_count);
1103
1104	list_add(&kq->list, &qpd->priv_queue_list);
1105	dqm->queue_count++;
1106	qpd->is_debug = true;
1107	execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
 
1108	dqm_unlock(dqm);
1109
1110	return 0;
1111}
1112
1113static void destroy_kernel_queue_cpsch(struct device_queue_manager *dqm,
1114					struct kernel_queue *kq,
1115					struct qcm_process_device *qpd)
1116{
1117	dqm_lock(dqm);
1118	list_del(&kq->list);
1119	dqm->queue_count--;
1120	qpd->is_debug = false;
1121	execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
 
1122	/*
1123	 * Unconditionally decrement this counter, regardless of the queue's
1124	 * type.
1125	 */
1126	dqm->total_queue_count--;
1127	pr_debug("Total of %d queues are accountable so far\n",
1128			dqm->total_queue_count);
1129	dqm_unlock(dqm);
1130}
1131
1132static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
1133			struct qcm_process_device *qpd)
 
 
1134{
1135	int retval;
1136	struct mqd_manager *mqd_mgr;
1137
1138	if (dqm->total_queue_count >= max_num_of_queues_per_device) {
1139		pr_warn("Can't create new usermode queue because %d queues were already created\n",
1140				dqm->total_queue_count);
1141		retval = -EPERM;
1142		goto out;
1143	}
1144
1145	if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1146		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1147		dqm_lock(dqm);
1148		retval = allocate_sdma_queue(dqm, q);
1149		dqm_unlock(dqm);
1150		if (retval)
1151			goto out;
1152	}
1153
1154	retval = allocate_doorbell(qpd, q);
1155	if (retval)
1156		goto out_deallocate_sdma_queue;
1157
1158	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1159			q->properties.type)];
1160
1161	if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1162		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
1163		dqm->asic_ops.init_sdma_vm(dqm, q, qpd);
1164	q->properties.tba_addr = qpd->tba_addr;
1165	q->properties.tma_addr = qpd->tma_addr;
1166	q->mqd_mem_obj = mqd_mgr->allocate_mqd(mqd_mgr->dev, &q->properties);
1167	if (!q->mqd_mem_obj) {
1168		retval = -ENOMEM;
1169		goto out_deallocate_doorbell;
1170	}
1171
1172	dqm_lock(dqm);
1173	/*
1174	 * Eviction state logic: mark all queues as evicted, even ones
1175	 * not currently active. Restoring inactive queues later only
1176	 * updates the is_evicted flag but is a no-op otherwise.
1177	 */
1178	q->properties.is_evicted = !!qpd->evicted;
1179	mqd_mgr->init_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj,
1180				&q->gart_mqd_addr, &q->properties);
 
 
 
 
 
 
 
 
1181
1182	list_add(&q->list, &qpd->queues_list);
1183	qpd->queue_count++;
 
1184	if (q->properties.is_active) {
1185		dqm->queue_count++;
1186		retval = execute_queues_cpsch(dqm,
1187				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
 
 
 
 
 
 
1188	}
1189
1190	if (q->properties.type == KFD_QUEUE_TYPE_SDMA)
1191		dqm->sdma_queue_count++;
1192	else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI)
1193		dqm->xgmi_sdma_queue_count++;
1194	/*
1195	 * Unconditionally increment this counter, regardless of the queue's
1196	 * type or whether the queue is active.
1197	 */
1198	dqm->total_queue_count++;
1199
1200	pr_debug("Total of %d queues are accountable so far\n",
1201			dqm->total_queue_count);
1202
1203	dqm_unlock(dqm);
1204	return retval;
1205
 
 
 
 
 
 
 
1206out_deallocate_doorbell:
1207	deallocate_doorbell(qpd, q);
1208out_deallocate_sdma_queue:
1209	if (q->properties.type == KFD_QUEUE_TYPE_SDMA ||
1210		q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1211		dqm_lock(dqm);
1212		deallocate_sdma_queue(dqm, q);
1213		dqm_unlock(dqm);
1214	}
1215out:
1216	return retval;
1217}
1218
1219int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
1220				unsigned int fence_value,
1221				unsigned int timeout_ms)
1222{
1223	unsigned long end_jiffies = msecs_to_jiffies(timeout_ms) + jiffies;
 
 
1224
1225	while (*fence_addr != fence_value) {
 
 
 
 
1226		if (time_after(jiffies, end_jiffies)) {
1227			pr_err("qcm fence wait loop timeout expired\n");
1228			/* In HWS case, this is used to halt the driver thread
1229			 * in order not to mess up CP states before doing
1230			 * scandumps for FW debugging.
1231			 */
1232			while (halt_if_hws_hang)
1233				schedule();
1234
1235			return -ETIME;
1236		}
1237		schedule();
1238	}
1239
1240	return 0;
1241}
1242
1243static int unmap_sdma_queues(struct device_queue_manager *dqm)
1244{
1245	int i, retval = 0;
1246
1247	for (i = 0; i < dqm->dev->device_info->num_sdma_engines +
1248		dqm->dev->device_info->num_xgmi_sdma_engines; i++) {
1249		retval = pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_SDMA,
1250			KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0, false, i);
1251		if (retval)
1252			return retval;
1253	}
1254	return retval;
1255}
1256
1257/* dqm->lock mutex has to be locked before calling this function */
1258static int map_queues_cpsch(struct device_queue_manager *dqm)
1259{
 
1260	int retval;
1261
1262	if (dqm->queue_count <= 0 || dqm->processes_count <= 0)
 
 
1263		return 0;
1264
1265	if (dqm->active_runlist)
1266		return 0;
1267
1268	retval = pm_send_runlist(&dqm->packets, &dqm->queues);
1269	pr_debug("%s sent runlist\n", __func__);
1270	if (retval) {
1271		pr_err("failed to execute runlist\n");
1272		return retval;
1273	}
1274	dqm->active_runlist = true;
1275
1276	return retval;
1277}
1278
1279/* dqm->lock mutex has to be locked before calling this function */
1280static int unmap_queues_cpsch(struct device_queue_manager *dqm,
1281				enum kfd_unmap_queues_filter filter,
1282				uint32_t filter_param)
 
 
1283{
 
 
1284	int retval = 0;
1285
1286	if (dqm->is_hws_hang)
 
 
1287		return -EIO;
1288	if (!dqm->active_runlist)
1289		return retval;
1290
1291	pr_debug("Before destroying queues, sdma queue count is : %u, xgmi sdma queue count is : %u\n",
1292		dqm->sdma_queue_count, dqm->xgmi_sdma_queue_count);
1293
1294	if (dqm->sdma_queue_count > 0 || dqm->xgmi_sdma_queue_count)
1295		unmap_sdma_queues(dqm);
1296
1297	retval = pm_send_unmap_queue(&dqm->packets, KFD_QUEUE_TYPE_COMPUTE,
1298			filter, filter_param, false, 0);
1299	if (retval)
1300		return retval;
1301
1302	*dqm->fence_addr = KFD_FENCE_INIT;
1303	pm_send_query_status(&dqm->packets, dqm->fence_gpu_addr,
1304				KFD_FENCE_COMPLETED);
1305	/* should be timed out */
1306	retval = amdkfd_fence_wait_timeout(dqm->fence_addr, KFD_FENCE_COMPLETED,
1307				queue_preemption_timeout_ms);
1308	if (retval)
 
 
1309		return retval;
 
1310
1311	pm_release_ib(&dqm->packets);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1312	dqm->active_runlist = false;
1313
1314	return retval;
1315}
1316
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1317/* dqm->lock mutex has to be locked before calling this function */
1318static int execute_queues_cpsch(struct device_queue_manager *dqm,
1319				enum kfd_unmap_queues_filter filter,
1320				uint32_t filter_param)
 
1321{
1322	int retval;
1323
1324	if (dqm->is_hws_hang)
1325		return -EIO;
1326	retval = unmap_queues_cpsch(dqm, filter, filter_param);
1327	if (retval) {
1328		pr_err("The cp might be in an unrecoverable state due to an unsuccessful queues preemption\n");
1329		dqm->is_hws_hang = true;
1330		schedule_work(&dqm->hw_exception_work);
1331		return retval;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1332	}
1333
1334	return map_queues_cpsch(dqm);
1335}
1336
1337static int destroy_queue_cpsch(struct device_queue_manager *dqm,
1338				struct qcm_process_device *qpd,
1339				struct queue *q)
1340{
1341	int retval;
1342	struct mqd_manager *mqd_mgr;
1343
1344	retval = 0;
 
 
 
 
 
 
 
 
 
 
 
1345
1346	/* remove queue from list to prevent rescheduling after preemption */
1347	dqm_lock(dqm);
1348
 
 
 
 
 
 
 
1349	if (qpd->is_debug) {
1350		/*
1351		 * error, currently we do not allow to destroy a queue
1352		 * of a currently debugged process
1353		 */
1354		retval = -EBUSY;
1355		goto failed_try_destroy_debugged_queue;
1356
1357	}
1358
1359	mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1360			q->properties.type)];
1361
1362	deallocate_doorbell(qpd, q);
1363
1364	if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
1365		dqm->sdma_queue_count--;
1366		deallocate_sdma_queue(dqm, q);
1367	} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1368		dqm->xgmi_sdma_queue_count--;
1369		deallocate_sdma_queue(dqm, q);
 
1370	}
1371
1372	list_del(&q->list);
1373	qpd->queue_count--;
1374	if (q->properties.is_active) {
1375		dqm->queue_count--;
1376		retval = execute_queues_cpsch(dqm,
1377				KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
1378		if (retval == -ETIME)
1379			qpd->reset_wavefronts = true;
 
 
 
 
 
1380	}
1381
1382	/*
1383	 * Unconditionally decrement this counter, regardless of the queue's
1384	 * type
1385	 */
1386	dqm->total_queue_count--;
1387	pr_debug("Total of %d queues are accountable so far\n",
1388			dqm->total_queue_count);
1389
1390	dqm_unlock(dqm);
1391
1392	/* Do free_mqd after dqm_unlock(dqm) to avoid circular locking */
 
 
 
 
 
 
 
1393	mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
1394
1395	return retval;
1396
1397failed_try_destroy_debugged_queue:
1398
1399	dqm_unlock(dqm);
1400	return retval;
1401}
1402
1403/*
1404 * Low bits must be 0000/FFFF as required by HW, high bits must be 0 to
1405 * stay in user mode.
1406 */
1407#define APE1_FIXED_BITS_MASK 0xFFFF80000000FFFFULL
1408/* APE1 limit is inclusive and 64K aligned. */
1409#define APE1_LIMIT_ALIGNMENT 0xFFFF
1410
1411static bool set_cache_memory_policy(struct device_queue_manager *dqm,
1412				   struct qcm_process_device *qpd,
1413				   enum cache_policy default_policy,
1414				   enum cache_policy alternate_policy,
1415				   void __user *alternate_aperture_base,
1416				   uint64_t alternate_aperture_size)
1417{
1418	bool retval = true;
1419
1420	if (!dqm->asic_ops.set_cache_memory_policy)
1421		return retval;
1422
1423	dqm_lock(dqm);
1424
1425	if (alternate_aperture_size == 0) {
1426		/* base > limit disables APE1 */
1427		qpd->sh_mem_ape1_base = 1;
1428		qpd->sh_mem_ape1_limit = 0;
1429	} else {
1430		/*
1431		 * In FSA64, APE1_Base[63:0] = { 16{SH_MEM_APE1_BASE[31]},
1432		 *			SH_MEM_APE1_BASE[31:0], 0x0000 }
1433		 * APE1_Limit[63:0] = { 16{SH_MEM_APE1_LIMIT[31]},
1434		 *			SH_MEM_APE1_LIMIT[31:0], 0xFFFF }
1435		 * Verify that the base and size parameters can be
1436		 * represented in this format and convert them.
1437		 * Additionally restrict APE1 to user-mode addresses.
1438		 */
1439
1440		uint64_t base = (uintptr_t)alternate_aperture_base;
1441		uint64_t limit = base + alternate_aperture_size - 1;
1442
1443		if (limit <= base || (base & APE1_FIXED_BITS_MASK) != 0 ||
1444		   (limit & APE1_FIXED_BITS_MASK) != APE1_LIMIT_ALIGNMENT) {
1445			retval = false;
1446			goto out;
1447		}
1448
1449		qpd->sh_mem_ape1_base = base >> 16;
1450		qpd->sh_mem_ape1_limit = limit >> 16;
1451	}
1452
1453	retval = dqm->asic_ops.set_cache_memory_policy(
1454			dqm,
1455			qpd,
1456			default_policy,
1457			alternate_policy,
1458			alternate_aperture_base,
1459			alternate_aperture_size);
1460
1461	if ((dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) && (qpd->vmid != 0))
1462		program_sh_mem_settings(dqm, qpd);
1463
1464	pr_debug("sh_mem_config: 0x%x, ape1_base: 0x%x, ape1_limit: 0x%x\n",
1465		qpd->sh_mem_config, qpd->sh_mem_ape1_base,
1466		qpd->sh_mem_ape1_limit);
1467
1468out:
1469	dqm_unlock(dqm);
1470	return retval;
1471}
1472
1473static int set_trap_handler(struct device_queue_manager *dqm,
1474				struct qcm_process_device *qpd,
1475				uint64_t tba_addr,
1476				uint64_t tma_addr)
1477{
1478	uint64_t *tma;
1479
1480	if (dqm->dev->cwsr_enabled) {
1481		/* Jump from CWSR trap handler to user trap */
1482		tma = (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1483		tma[0] = tba_addr;
1484		tma[1] = tma_addr;
1485	} else {
1486		qpd->tba_addr = tba_addr;
1487		qpd->tma_addr = tma_addr;
1488	}
1489
1490	return 0;
1491}
1492
1493static int process_termination_nocpsch(struct device_queue_manager *dqm,
1494		struct qcm_process_device *qpd)
1495{
1496	struct queue *q, *next;
1497	struct device_process_node *cur, *next_dpn;
1498	int retval = 0;
1499	bool found = false;
1500
1501	dqm_lock(dqm);
1502
1503	/* Clear all user mode queues */
1504	list_for_each_entry_safe(q, next, &qpd->queues_list, list) {
 
1505		int ret;
1506
 
 
 
1507		ret = destroy_queue_nocpsch_locked(dqm, qpd, q);
1508		if (ret)
1509			retval = ret;
 
 
 
1510	}
1511
1512	/* Unregister process */
1513	list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
1514		if (qpd == cur->qpd) {
1515			list_del(&cur->list);
1516			kfree(cur);
1517			dqm->processes_count--;
1518			found = true;
1519			break;
1520		}
1521	}
1522
1523	dqm_unlock(dqm);
1524
1525	/* Outside the DQM lock because under the DQM lock we can't do
1526	 * reclaim or take other locks that others hold while reclaiming.
1527	 */
1528	if (found)
1529		kfd_dec_compute_active(dqm->dev);
1530
1531	return retval;
1532}
1533
1534static int get_wave_state(struct device_queue_manager *dqm,
1535			  struct queue *q,
1536			  void __user *ctl_stack,
1537			  u32 *ctl_stack_used_size,
1538			  u32 *save_area_used_size)
1539{
1540	struct mqd_manager *mqd_mgr;
1541	int r;
1542
1543	dqm_lock(dqm);
1544
 
 
1545	if (q->properties.type != KFD_QUEUE_TYPE_COMPUTE ||
1546	    q->properties.is_active || !q->device->cwsr_enabled) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1547		r = -EINVAL;
1548		goto dqm_unlock;
1549	}
1550
1551	mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_COMPUTE];
1552
1553	if (!mqd_mgr->get_wave_state) {
1554		r = -EINVAL;
1555		goto dqm_unlock;
1556	}
1557
1558	r = mqd_mgr->get_wave_state(mqd_mgr, q->mqd, ctl_stack,
1559			ctl_stack_used_size, save_area_used_size);
1560
1561dqm_unlock:
1562	dqm_unlock(dqm);
1563	return r;
1564}
1565
1566static int process_termination_cpsch(struct device_queue_manager *dqm,
1567		struct qcm_process_device *qpd)
1568{
1569	int retval;
1570	struct queue *q, *next;
 
1571	struct kernel_queue *kq, *kq_next;
1572	struct mqd_manager *mqd_mgr;
1573	struct device_process_node *cur, *next_dpn;
1574	enum kfd_unmap_queues_filter filter =
1575		KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES;
1576	bool found = false;
1577
1578	retval = 0;
1579
1580	dqm_lock(dqm);
1581
1582	/* Clean all kernel queues */
1583	list_for_each_entry_safe(kq, kq_next, &qpd->priv_queue_list, list) {
1584		list_del(&kq->list);
1585		dqm->queue_count--;
1586		qpd->is_debug = false;
1587		dqm->total_queue_count--;
1588		filter = KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES;
1589	}
1590
1591	/* Clear all user mode queues */
1592	list_for_each_entry(q, &qpd->queues_list, list) {
1593		if (q->properties.type == KFD_QUEUE_TYPE_SDMA) {
1594			dqm->sdma_queue_count--;
1595			deallocate_sdma_queue(dqm, q);
1596		} else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) {
1597			dqm->xgmi_sdma_queue_count--;
1598			deallocate_sdma_queue(dqm, q);
 
 
 
 
 
 
 
 
 
 
1599		}
1600
1601		if (q->properties.is_active)
1602			dqm->queue_count--;
1603
1604		dqm->total_queue_count--;
1605	}
1606
1607	/* Unregister process */
1608	list_for_each_entry_safe(cur, next_dpn, &dqm->queues, list) {
1609		if (qpd == cur->qpd) {
1610			list_del(&cur->list);
1611			kfree(cur);
1612			dqm->processes_count--;
1613			found = true;
1614			break;
1615		}
1616	}
1617
1618	retval = execute_queues_cpsch(dqm, filter, 0);
 
 
1619	if ((!dqm->is_hws_hang) && (retval || qpd->reset_wavefronts)) {
1620		pr_warn("Resetting wave fronts (cpsch) on dev %p\n", dqm->dev);
1621		dbgdev_wave_reset_wavefronts(dqm->dev, qpd->pqm->process);
1622		qpd->reset_wavefronts = false;
1623	}
1624
1625	dqm_unlock(dqm);
1626
1627	/* Outside the DQM lock because under the DQM lock we can't do
1628	 * reclaim or take other locks that others hold while reclaiming.
1629	 */
1630	if (found)
1631		kfd_dec_compute_active(dqm->dev);
1632
1633	/* Lastly, free mqd resources.
1634	 * Do free_mqd() after dqm_unlock to avoid circular locking.
1635	 */
1636	list_for_each_entry_safe(q, next, &qpd->queues_list, list) {
 
1637		mqd_mgr = dqm->mqd_mgrs[get_mqd_type_from_queue_type(
1638				q->properties.type)];
1639		list_del(&q->list);
1640		qpd->queue_count--;
 
1641		mqd_mgr->free_mqd(mqd_mgr, q->mqd, q->mqd_mem_obj);
 
1642	}
 
 
 
 
 
 
 
1643
1644	return retval;
1645}
1646
1647static int init_mqd_managers(struct device_queue_manager *dqm)
1648{
1649	int i, j;
 
1650	struct mqd_manager *mqd_mgr;
1651
1652	for (i = 0; i < KFD_MQD_TYPE_MAX; i++) {
1653		mqd_mgr = dqm->asic_ops.mqd_manager_init(i, dqm->dev);
1654		if (!mqd_mgr) {
1655			pr_err("mqd manager [%d] initialization failed\n", i);
1656			goto out_free;
1657		}
1658		dqm->mqd_mgrs[i] = mqd_mgr;
1659	}
1660
1661	return 0;
1662
1663out_free:
1664	for (j = 0; j < i; j++) {
1665		kfree(dqm->mqd_mgrs[j]);
1666		dqm->mqd_mgrs[j] = NULL;
1667	}
1668
1669	return -ENOMEM;
1670}
1671
1672/* Allocate one hiq mqd (HWS) and all SDMA mqd in a continuous trunk*/
1673static int allocate_hiq_sdma_mqd(struct device_queue_manager *dqm)
1674{
1675	int retval;
1676	struct kfd_dev *dev = dqm->dev;
1677	struct kfd_mem_obj *mem_obj = &dqm->hiq_sdma_mqd;
1678	uint32_t size = dqm->mqd_mgrs[KFD_MQD_TYPE_SDMA]->mqd_size *
1679		dev->device_info->num_sdma_engines *
1680		dev->device_info->num_sdma_queues_per_engine +
1681		dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ]->mqd_size;
 
1682
1683	retval = amdgpu_amdkfd_alloc_gtt_mem(dev->kgd, size,
1684		&(mem_obj->gtt_mem), &(mem_obj->gpu_addr),
1685		(void *)&(mem_obj->cpu_ptr), true);
1686
1687	return retval;
1688}
1689
1690struct device_queue_manager *device_queue_manager_init(struct kfd_dev *dev)
1691{
1692	struct device_queue_manager *dqm;
1693
1694	pr_debug("Loading device queue manager\n");
1695
1696	dqm = kzalloc(sizeof(*dqm), GFP_KERNEL);
1697	if (!dqm)
1698		return NULL;
1699
1700	switch (dev->device_info->asic_family) {
1701	/* HWS is not available on Hawaii. */
1702	case CHIP_HAWAII:
1703	/* HWS depends on CWSR for timely dequeue. CWSR is not
1704	 * available on Tonga.
1705	 *
1706	 * FIXME: This argument also applies to Kaveri.
1707	 */
1708	case CHIP_TONGA:
1709		dqm->sched_policy = KFD_SCHED_POLICY_NO_HWS;
1710		break;
1711	default:
1712		dqm->sched_policy = sched_policy;
1713		break;
1714	}
1715
1716	dqm->dev = dev;
1717	switch (dqm->sched_policy) {
1718	case KFD_SCHED_POLICY_HWS:
1719	case KFD_SCHED_POLICY_HWS_NO_OVERSUBSCRIPTION:
1720		/* initialize dqm for cp scheduling */
1721		dqm->ops.create_queue = create_queue_cpsch;
1722		dqm->ops.initialize = initialize_cpsch;
1723		dqm->ops.start = start_cpsch;
1724		dqm->ops.stop = stop_cpsch;
 
1725		dqm->ops.destroy_queue = destroy_queue_cpsch;
1726		dqm->ops.update_queue = update_queue;
1727		dqm->ops.register_process = register_process;
1728		dqm->ops.unregister_process = unregister_process;
1729		dqm->ops.uninitialize = uninitialize;
1730		dqm->ops.create_kernel_queue = create_kernel_queue_cpsch;
1731		dqm->ops.destroy_kernel_queue = destroy_kernel_queue_cpsch;
1732		dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1733		dqm->ops.set_trap_handler = set_trap_handler;
1734		dqm->ops.process_termination = process_termination_cpsch;
1735		dqm->ops.evict_process_queues = evict_process_queues_cpsch;
1736		dqm->ops.restore_process_queues = restore_process_queues_cpsch;
1737		dqm->ops.get_wave_state = get_wave_state;
 
 
 
1738		break;
1739	case KFD_SCHED_POLICY_NO_HWS:
1740		/* initialize dqm for no cp scheduling */
1741		dqm->ops.start = start_nocpsch;
1742		dqm->ops.stop = stop_nocpsch;
 
1743		dqm->ops.create_queue = create_queue_nocpsch;
1744		dqm->ops.destroy_queue = destroy_queue_nocpsch;
1745		dqm->ops.update_queue = update_queue;
1746		dqm->ops.register_process = register_process;
1747		dqm->ops.unregister_process = unregister_process;
1748		dqm->ops.initialize = initialize_nocpsch;
1749		dqm->ops.uninitialize = uninitialize;
1750		dqm->ops.set_cache_memory_policy = set_cache_memory_policy;
1751		dqm->ops.set_trap_handler = set_trap_handler;
1752		dqm->ops.process_termination = process_termination_nocpsch;
1753		dqm->ops.evict_process_queues = evict_process_queues_nocpsch;
1754		dqm->ops.restore_process_queues =
1755			restore_process_queues_nocpsch;
1756		dqm->ops.get_wave_state = get_wave_state;
 
 
1757		break;
1758	default:
1759		pr_err("Invalid scheduling policy %d\n", dqm->sched_policy);
1760		goto out_free;
1761	}
1762
1763	switch (dev->device_info->asic_family) {
1764	case CHIP_CARRIZO:
1765		device_queue_manager_init_vi(&dqm->asic_ops);
1766		break;
1767
1768	case CHIP_KAVERI:
 
1769		device_queue_manager_init_cik(&dqm->asic_ops);
1770		break;
1771
1772	case CHIP_HAWAII:
1773		device_queue_manager_init_cik_hawaii(&dqm->asic_ops);
1774		break;
1775
1776	case CHIP_TONGA:
1777	case CHIP_FIJI:
1778	case CHIP_POLARIS10:
1779	case CHIP_POLARIS11:
1780	case CHIP_POLARIS12:
1781	case CHIP_VEGAM:
1782		device_queue_manager_init_vi_tonga(&dqm->asic_ops);
1783		break;
1784
1785	case CHIP_VEGA10:
1786	case CHIP_VEGA12:
1787	case CHIP_VEGA20:
1788	case CHIP_RAVEN:
1789	case CHIP_ARCTURUS:
1790		device_queue_manager_init_v9(&dqm->asic_ops);
1791		break;
1792	case CHIP_NAVI10:
1793		device_queue_manager_init_v10_navi10(&dqm->asic_ops);
1794		break;
1795	default:
1796		WARN(1, "Unexpected ASIC family %u",
1797		     dev->device_info->asic_family);
1798		goto out_free;
 
 
 
 
 
 
 
 
1799	}
1800
1801	if (init_mqd_managers(dqm))
1802		goto out_free;
1803
1804	if (allocate_hiq_sdma_mqd(dqm)) {
1805		pr_err("Failed to allocate hiq sdma mqd trunk buffer\n");
1806		goto out_free;
1807	}
1808
1809	if (!dqm->ops.initialize(dqm))
 
1810		return dqm;
 
1811
1812out_free:
1813	kfree(dqm);
1814	return NULL;
1815}
1816
1817static void deallocate_hiq_sdma_mqd(struct kfd_dev *dev,
1818				    struct kfd_mem_obj *mqd)
1819{
1820	WARN(!mqd, "No hiq sdma mqd trunk to free");
1821
1822	amdgpu_amdkfd_free_gtt_mem(dev->kgd, mqd->gtt_mem);
1823}
1824
1825void device_queue_manager_uninit(struct device_queue_manager *dqm)
1826{
 
1827	dqm->ops.uninitialize(dqm);
1828	deallocate_hiq_sdma_mqd(dqm->dev, &dqm->hiq_sdma_mqd);
 
1829	kfree(dqm);
1830}
1831
1832int kfd_process_vm_fault(struct device_queue_manager *dqm,
1833			 unsigned int pasid)
1834{
1835	struct kfd_process_device *pdd;
1836	struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
1837	int ret = 0;
1838
1839	if (!p)
1840		return -EINVAL;
 
1841	pdd = kfd_get_process_device_data(dqm->dev, p);
1842	if (pdd)
1843		ret = dqm->ops.evict_process_queues(dqm, &pdd->qpd);
1844	kfd_unref_process(p);
1845
1846	return ret;
1847}
1848
1849static void kfd_process_hw_exception(struct work_struct *work)
1850{
1851	struct device_queue_manager *dqm = container_of(work,
1852			struct device_queue_manager, hw_exception_work);
1853	amdgpu_amdkfd_gpu_reset(dqm->dev->kgd);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1854}
1855
1856#if defined(CONFIG_DEBUG_FS)
1857
1858static void seq_reg_dump(struct seq_file *m,
1859			 uint32_t (*dump)[2], uint32_t n_regs)
1860{
1861	uint32_t i, count;
1862
1863	for (i = 0, count = 0; i < n_regs; i++) {
1864		if (count == 0 ||
1865		    dump[i-1][0] + sizeof(uint32_t) != dump[i][0]) {
1866			seq_printf(m, "%s    %08x: %08x",
1867				   i ? "\n" : "",
1868				   dump[i][0], dump[i][1]);
1869			count = 7;
1870		} else {
1871			seq_printf(m, " %08x", dump[i][1]);
1872			count--;
1873		}
1874	}
1875
1876	seq_puts(m, "\n");
1877}
1878
1879int dqm_debugfs_hqds(struct seq_file *m, void *data)
1880{
1881	struct device_queue_manager *dqm = data;
 
1882	uint32_t (*dump)[2], n_regs;
1883	int pipe, queue;
1884	int r = 0;
 
1885
1886	r = dqm->dev->kfd2kgd->hqd_dump(dqm->dev->kgd,
1887					KFD_CIK_HIQ_PIPE, KFD_CIK_HIQ_QUEUE,
1888					&dump, &n_regs);
1889	if (!r) {
1890		seq_printf(m, "  HIQ on MEC %d Pipe %d Queue %d\n",
1891			   KFD_CIK_HIQ_PIPE/get_pipes_per_mec(dqm)+1,
1892			   KFD_CIK_HIQ_PIPE%get_pipes_per_mec(dqm),
1893			   KFD_CIK_HIQ_QUEUE);
1894		seq_reg_dump(m, dump, n_regs);
1895
1896		kfree(dump);
1897	}
1898
1899	for (pipe = 0; pipe < get_pipes_per_mec(dqm); pipe++) {
1900		int pipe_offset = pipe * get_queues_per_pipe(dqm);
 
 
 
 
 
 
 
 
 
 
 
 
1901
1902		for (queue = 0; queue < get_queues_per_pipe(dqm); queue++) {
1903			if (!test_bit(pipe_offset + queue,
1904				      dqm->dev->shared_resources.queue_bitmap))
1905				continue;
1906
1907			r = dqm->dev->kfd2kgd->hqd_dump(
1908				dqm->dev->kgd, pipe, queue, &dump, &n_regs);
1909			if (r)
1910				break;
1911
1912			seq_printf(m, "  CP Pipe %d, Queue %d\n",
1913				  pipe, queue);
1914			seq_reg_dump(m, dump, n_regs);
 
 
 
 
 
 
 
 
 
 
 
 
 
1915
1916			kfree(dump);
 
1917		}
1918	}
1919
1920	for (pipe = 0; pipe < get_num_sdma_engines(dqm); pipe++) {
 
 
 
1921		for (queue = 0;
1922		     queue < dqm->dev->device_info->num_sdma_queues_per_engine;
1923		     queue++) {
1924			r = dqm->dev->kfd2kgd->hqd_sdma_dump(
1925				dqm->dev->kgd, pipe, queue, &dump, &n_regs);
1926			if (r)
1927				break;
1928
1929			seq_printf(m, "  SDMA Engine %d, RLC %d\n",
1930				  pipe, queue);
1931			seq_reg_dump(m, dump, n_regs);
1932
1933			kfree(dump);
1934		}
1935	}
1936
1937	return r;
1938}
1939
1940int dqm_debugfs_execute_queues(struct device_queue_manager *dqm)
1941{
1942	int r = 0;
1943
1944	dqm_lock(dqm);
 
 
 
 
 
1945	dqm->active_runlist = true;
1946	r = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_ALL_QUEUES, 0);
 
1947	dqm_unlock(dqm);
1948
1949	return r;
1950}
1951
1952#endif