kfd_events.c

   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#include <linux/mm_types.h>
  25#include <linux/slab.h>
  26#include <linux/types.h>
  27#include <linux/sched/signal.h>
  28#include <linux/sched/mm.h>
  29#include <linux/uaccess.h>
  30#include <linux/mman.h>
  31#include <linux/memory.h>
  32#include "kfd_priv.h"
  33#include "kfd_events.h"
  34#include "kfd_device_queue_manager.h"
  35#include <linux/device.h>
  36
  37/*
  38 * Wrapper around wait_queue_entry_t
  39 */
  40struct kfd_event_waiter {
  41	wait_queue_entry_t wait;
  42	struct kfd_event *event; /* Event to wait for */
  43	bool activated;		 /* Becomes true when event is signaled */
  44	bool event_age_enabled;  /* set to true when last_event_age is non-zero */
  45};
  46
  47/*
  48 * Each signal event needs a 64-bit signal slot where the signaler will write
  49 * a 1 before sending an interrupt. (This is needed because some interrupts
  50 * do not contain enough spare data bits to identify an event.)
  51 * We get whole pages and map them to the process VA.
  52 * Individual signal events use their event_id as slot index.
  53 */
  54struct kfd_signal_page {
  55	uint64_t *kernel_address;
  56	uint64_t __user *user_address;
  57	bool need_to_free_pages;
  58};
  59
  60static uint64_t *page_slots(struct kfd_signal_page *page)
  61{
  62	return page->kernel_address;
  63}
  64
  65static struct kfd_signal_page *allocate_signal_page(struct kfd_process *p)
  66{
  67	void *backing_store;
  68	struct kfd_signal_page *page;
  69
  70	page = kzalloc(sizeof(*page), GFP_KERNEL);
  71	if (!page)
  72		return NULL;
  73
  74	backing_store = (void *) __get_free_pages(GFP_KERNEL,
  75					get_order(KFD_SIGNAL_EVENT_LIMIT * 8));
  76	if (!backing_store)
  77		goto fail_alloc_signal_store;
  78
  79	/* Initialize all events to unsignaled */
  80	memset(backing_store, (uint8_t) UNSIGNALED_EVENT_SLOT,
  81	       KFD_SIGNAL_EVENT_LIMIT * 8);
  82
  83	page->kernel_address = backing_store;
  84	page->need_to_free_pages = true;
  85	pr_debug("Allocated new event signal page at %p, for process %p\n",
  86			page, p);
  87
  88	return page;
  89
  90fail_alloc_signal_store:
  91	kfree(page);
  92	return NULL;
  93}
  94
  95static int allocate_event_notification_slot(struct kfd_process *p,
  96					    struct kfd_event *ev,
  97					    const int *restore_id)
  98{
  99	int id;
 100
 101	if (!p->signal_page) {
 102		p->signal_page = allocate_signal_page(p);
 103		if (!p->signal_page)
 104			return -ENOMEM;
 105		/* Oldest user mode expects 256 event slots */
 106		p->signal_mapped_size = 256*8;
 107	}
 108
 109	if (restore_id) {
 110		id = idr_alloc(&p->event_idr, ev, *restore_id, *restore_id + 1,
 111				GFP_KERNEL);
 112	} else {
 113		/*
 114		 * Compatibility with old user mode: Only use signal slots
 115		 * user mode has mapped, may be less than
 116		 * KFD_SIGNAL_EVENT_LIMIT. This also allows future increase
 117		 * of the event limit without breaking user mode.
 118		 */
 119		id = idr_alloc(&p->event_idr, ev, 0, p->signal_mapped_size / 8,
 120				GFP_KERNEL);
 121	}
 122	if (id < 0)
 123		return id;
 124
 125	ev->event_id = id;
 126	page_slots(p->signal_page)[id] = UNSIGNALED_EVENT_SLOT;
 127
 128	return 0;
 129}
 130
 131/*
 132 * Assumes that p->event_mutex or rcu_readlock is held and of course that p is
 133 * not going away.
 134 */
 135static struct kfd_event *lookup_event_by_id(struct kfd_process *p, uint32_t id)
 136{
 137	return idr_find(&p->event_idr, id);
 138}
 139
 140/**
 141 * lookup_signaled_event_by_partial_id - Lookup signaled event from partial ID
 142 * @p:     Pointer to struct kfd_process
 143 * @id:    ID to look up
 144 * @bits:  Number of valid bits in @id
 145 *
 146 * Finds the first signaled event with a matching partial ID. If no
 147 * matching signaled event is found, returns NULL. In that case the
 148 * caller should assume that the partial ID is invalid and do an
 149 * exhaustive search of all siglaned events.
 150 *
 151 * If multiple events with the same partial ID signal at the same
 152 * time, they will be found one interrupt at a time, not necessarily
 153 * in the same order the interrupts occurred. As long as the number of
 154 * interrupts is correct, all signaled events will be seen by the
 155 * driver.
 156 */
 157static struct kfd_event *lookup_signaled_event_by_partial_id(
 158	struct kfd_process *p, uint32_t id, uint32_t bits)
 159{
 160	struct kfd_event *ev;
 161
 162	if (!p->signal_page || id >= KFD_SIGNAL_EVENT_LIMIT)
 163		return NULL;
 164
 165	/* Fast path for the common case that @id is not a partial ID
 166	 * and we only need a single lookup.
 167	 */
 168	if (bits > 31 || (1U << bits) >= KFD_SIGNAL_EVENT_LIMIT) {
 169		if (page_slots(p->signal_page)[id] == UNSIGNALED_EVENT_SLOT)
 170			return NULL;
 171
 172		return idr_find(&p->event_idr, id);
 173	}
 174
 175	/* General case for partial IDs: Iterate over all matching IDs
 176	 * and find the first one that has signaled.
 177	 */
 178	for (ev = NULL; id < KFD_SIGNAL_EVENT_LIMIT && !ev; id += 1U << bits) {
 179		if (page_slots(p->signal_page)[id] == UNSIGNALED_EVENT_SLOT)
 180			continue;
 181
 182		ev = idr_find(&p->event_idr, id);
 183	}
 184
 185	return ev;
 186}
 187
 188static int create_signal_event(struct file *devkfd, struct kfd_process *p,
 189				struct kfd_event *ev, const int *restore_id)
 190{
 191	int ret;
 192
 193	if (p->signal_mapped_size &&
 194	    p->signal_event_count == p->signal_mapped_size / 8) {
 195		if (!p->signal_event_limit_reached) {
 196			pr_debug("Signal event wasn't created because limit was reached\n");
 197			p->signal_event_limit_reached = true;
 198		}
 199		return -ENOSPC;
 200	}
 201
 202	ret = allocate_event_notification_slot(p, ev, restore_id);
 203	if (ret) {
 204		pr_warn("Signal event wasn't created because out of kernel memory\n");
 205		return ret;
 206	}
 207
 208	p->signal_event_count++;
 209
 210	ev->user_signal_address = &p->signal_page->user_address[ev->event_id];
 211	pr_debug("Signal event number %zu created with id %d, address %p\n",
 212			p->signal_event_count, ev->event_id,
 213			ev->user_signal_address);
 214
 215	return 0;
 216}
 217
 218static int create_other_event(struct kfd_process *p, struct kfd_event *ev, const int *restore_id)
 219{
 220	int id;
 221
 222	if (restore_id)
 223		id = idr_alloc(&p->event_idr, ev, *restore_id, *restore_id + 1,
 224			GFP_KERNEL);
 225	else
 226		/* Cast KFD_LAST_NONSIGNAL_EVENT to uint32_t. This allows an
 227		 * intentional integer overflow to -1 without a compiler
 228		 * warning. idr_alloc treats a negative value as "maximum
 229		 * signed integer".
 230		 */
 231		id = idr_alloc(&p->event_idr, ev, KFD_FIRST_NONSIGNAL_EVENT_ID,
 232				(uint32_t)KFD_LAST_NONSIGNAL_EVENT_ID + 1,
 233				GFP_KERNEL);
 234
 235	if (id < 0)
 236		return id;
 237	ev->event_id = id;
 238
 239	return 0;
 240}
 241
 242int kfd_event_init_process(struct kfd_process *p)
 243{
 244	int id;
 245
 246	mutex_init(&p->event_mutex);
 247	idr_init(&p->event_idr);
 248	p->signal_page = NULL;
 249	p->signal_event_count = 1;
 250	/* Allocate event ID 0. It is used for a fast path to ignore bogus events
 251	 * that are sent by the CP without a context ID
 252	 */
 253	id = idr_alloc(&p->event_idr, NULL, 0, 1, GFP_KERNEL);
 254	if (id < 0) {
 255		idr_destroy(&p->event_idr);
 256		mutex_destroy(&p->event_mutex);
 257		return id;
 258	}
 259	return 0;
 260}
 261
 262static void destroy_event(struct kfd_process *p, struct kfd_event *ev)
 263{
 264	struct kfd_event_waiter *waiter;
 265
 266	/* Wake up pending waiters. They will return failure */
 267	spin_lock(&ev->lock);
 268	list_for_each_entry(waiter, &ev->wq.head, wait.entry)
 269		WRITE_ONCE(waiter->event, NULL);
 270	wake_up_all(&ev->wq);
 271	spin_unlock(&ev->lock);
 272
 273	if (ev->type == KFD_EVENT_TYPE_SIGNAL ||
 274	    ev->type == KFD_EVENT_TYPE_DEBUG)
 275		p->signal_event_count--;
 276
 277	idr_remove(&p->event_idr, ev->event_id);
 278	kfree_rcu(ev, rcu);
 279}
 280
 281static void destroy_events(struct kfd_process *p)
 282{
 283	struct kfd_event *ev;
 284	uint32_t id;
 285
 286	idr_for_each_entry(&p->event_idr, ev, id)
 287		if (ev)
 288			destroy_event(p, ev);
 289	idr_destroy(&p->event_idr);
 290	mutex_destroy(&p->event_mutex);
 291}
 292
 293/*
 294 * We assume that the process is being destroyed and there is no need to
 295 * unmap the pages or keep bookkeeping data in order.
 296 */
 297static void shutdown_signal_page(struct kfd_process *p)
 298{
 299	struct kfd_signal_page *page = p->signal_page;
 300
 301	if (page) {
 302		if (page->need_to_free_pages)
 303			free_pages((unsigned long)page->kernel_address,
 304				   get_order(KFD_SIGNAL_EVENT_LIMIT * 8));
 305		kfree(page);
 306	}
 307}
 308
 309void kfd_event_free_process(struct kfd_process *p)
 310{
 311	destroy_events(p);
 312	shutdown_signal_page(p);
 313}
 314
 315static bool event_can_be_gpu_signaled(const struct kfd_event *ev)
 316{
 317	return ev->type == KFD_EVENT_TYPE_SIGNAL ||
 318					ev->type == KFD_EVENT_TYPE_DEBUG;
 319}
 320
 321static bool event_can_be_cpu_signaled(const struct kfd_event *ev)
 322{
 323	return ev->type == KFD_EVENT_TYPE_SIGNAL;
 324}
 325
 326static int kfd_event_page_set(struct kfd_process *p, void *kernel_address,
 327		       uint64_t size, uint64_t user_handle)
 328{
 329	struct kfd_signal_page *page;
 330
 331	if (p->signal_page)
 332		return -EBUSY;
 333
 334	if (size < KFD_SIGNAL_EVENT_LIMIT * 8) {
 335		pr_err("Event page size %llu is too small, need at least %lu bytes\n",
 336				size, (unsigned long)(KFD_SIGNAL_EVENT_LIMIT * 8));
 337		return -EINVAL;
 338	}
 339
 340	page = kzalloc(sizeof(*page), GFP_KERNEL);
 341	if (!page)
 342		return -ENOMEM;
 343
 344	/* Initialize all events to unsignaled */
 345	memset(kernel_address, (uint8_t) UNSIGNALED_EVENT_SLOT,
 346	       KFD_SIGNAL_EVENT_LIMIT * 8);
 347
 348	page->kernel_address = kernel_address;
 349
 350	p->signal_page = page;
 351	p->signal_mapped_size = size;
 352	p->signal_handle = user_handle;
 353	return 0;
 354}
 355
 356int kfd_kmap_event_page(struct kfd_process *p, uint64_t event_page_offset)
 357{
 358	struct kfd_node *kfd;
 359	struct kfd_process_device *pdd;
 360	void *mem, *kern_addr;
 361	uint64_t size;
 362	int err = 0;
 363
 364	if (p->signal_page) {
 365		pr_err("Event page is already set\n");
 366		return -EINVAL;
 367	}
 368
 369	pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(event_page_offset));
 370	if (!pdd) {
 371		pr_err("Getting device by id failed in %s\n", __func__);
 372		return -EINVAL;
 373	}
 374	kfd = pdd->dev;
 375
 376	pdd = kfd_bind_process_to_device(kfd, p);
 377	if (IS_ERR(pdd))
 378		return PTR_ERR(pdd);
 379
 380	mem = kfd_process_device_translate_handle(pdd,
 381			GET_IDR_HANDLE(event_page_offset));
 382	if (!mem) {
 383		pr_err("Can't find BO, offset is 0x%llx\n", event_page_offset);
 384		return -EINVAL;
 385	}
 386
 387	err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(mem, &kern_addr, &size);
 388	if (err) {
 389		pr_err("Failed to map event page to kernel\n");
 390		return err;
 391	}
 392
 393	err = kfd_event_page_set(p, kern_addr, size, event_page_offset);
 394	if (err) {
 395		pr_err("Failed to set event page\n");
 396		amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem);
 397		return err;
 398	}
 399	return err;
 400}
 401
 402int kfd_event_create(struct file *devkfd, struct kfd_process *p,
 403		     uint32_t event_type, bool auto_reset, uint32_t node_id,
 404		     uint32_t *event_id, uint32_t *event_trigger_data,
 405		     uint64_t *event_page_offset, uint32_t *event_slot_index)
 406{
 407	int ret = 0;
 408	struct kfd_event *ev = kzalloc(sizeof(*ev), GFP_KERNEL);
 409
 410	if (!ev)
 411		return -ENOMEM;
 412
 413	ev->type = event_type;
 414	ev->auto_reset = auto_reset;
 415	ev->signaled = false;
 416
 417	spin_lock_init(&ev->lock);
 418	init_waitqueue_head(&ev->wq);
 419
 420	*event_page_offset = 0;
 421
 422	mutex_lock(&p->event_mutex);
 423
 424	switch (event_type) {
 425	case KFD_EVENT_TYPE_SIGNAL:
 426	case KFD_EVENT_TYPE_DEBUG:
 427		ret = create_signal_event(devkfd, p, ev, NULL);
 428		if (!ret) {
 429			*event_page_offset = KFD_MMAP_TYPE_EVENTS;
 430			*event_slot_index = ev->event_id;
 431		}
 432		break;
 433	default:
 434		ret = create_other_event(p, ev, NULL);
 435		break;
 436	}
 437
 438	if (!ret) {
 439		*event_id = ev->event_id;
 440		*event_trigger_data = ev->event_id;
 441		ev->event_age = 1;
 442	} else {
 443		kfree(ev);
 444	}
 445
 446	mutex_unlock(&p->event_mutex);
 447
 448	return ret;
 449}
 450
 451int kfd_criu_restore_event(struct file *devkfd,
 452			   struct kfd_process *p,
 453			   uint8_t __user *user_priv_ptr,
 454			   uint64_t *priv_data_offset,
 455			   uint64_t max_priv_data_size)
 456{
 457	struct kfd_criu_event_priv_data *ev_priv;
 458	struct kfd_event *ev = NULL;
 459	int ret = 0;
 460
 461	ev_priv = kmalloc(sizeof(*ev_priv), GFP_KERNEL);
 462	if (!ev_priv)
 463		return -ENOMEM;
 464
 465	ev = kzalloc(sizeof(*ev), GFP_KERNEL);
 466	if (!ev) {
 467		ret = -ENOMEM;
 468		goto exit;
 469	}
 470
 471	if (*priv_data_offset + sizeof(*ev_priv) > max_priv_data_size) {
 472		ret = -EINVAL;
 473		goto exit;
 474	}
 475
 476	ret = copy_from_user(ev_priv, user_priv_ptr + *priv_data_offset, sizeof(*ev_priv));
 477	if (ret) {
 478		ret = -EFAULT;
 479		goto exit;
 480	}
 481	*priv_data_offset += sizeof(*ev_priv);
 482
 483	if (ev_priv->user_handle) {
 484		ret = kfd_kmap_event_page(p, ev_priv->user_handle);
 485		if (ret)
 486			goto exit;
 487	}
 488
 489	ev->type = ev_priv->type;
 490	ev->auto_reset = ev_priv->auto_reset;
 491	ev->signaled = ev_priv->signaled;
 492
 493	spin_lock_init(&ev->lock);
 494	init_waitqueue_head(&ev->wq);
 495
 496	mutex_lock(&p->event_mutex);
 497	switch (ev->type) {
 498	case KFD_EVENT_TYPE_SIGNAL:
 499	case KFD_EVENT_TYPE_DEBUG:
 500		ret = create_signal_event(devkfd, p, ev, &ev_priv->event_id);
 501		break;
 502	case KFD_EVENT_TYPE_MEMORY:
 503		memcpy(&ev->memory_exception_data,
 504			&ev_priv->memory_exception_data,
 505			sizeof(struct kfd_hsa_memory_exception_data));
 506
 507		ret = create_other_event(p, ev, &ev_priv->event_id);
 508		break;
 509	case KFD_EVENT_TYPE_HW_EXCEPTION:
 510		memcpy(&ev->hw_exception_data,
 511			&ev_priv->hw_exception_data,
 512			sizeof(struct kfd_hsa_hw_exception_data));
 513
 514		ret = create_other_event(p, ev, &ev_priv->event_id);
 515		break;
 516	}
 517	mutex_unlock(&p->event_mutex);
 518
 519exit:
 520	if (ret)
 521		kfree(ev);
 522
 523	kfree(ev_priv);
 524
 525	return ret;
 526}
 527
 528int kfd_criu_checkpoint_events(struct kfd_process *p,
 529			 uint8_t __user *user_priv_data,
 530			 uint64_t *priv_data_offset)
 531{
 532	struct kfd_criu_event_priv_data *ev_privs;
 533	int i = 0;
 534	int ret =  0;
 535	struct kfd_event *ev;
 536	uint32_t ev_id;
 537
 538	uint32_t num_events = kfd_get_num_events(p);
 539
 540	if (!num_events)
 541		return 0;
 542
 543	ev_privs = kvzalloc(num_events * sizeof(*ev_privs), GFP_KERNEL);
 544	if (!ev_privs)
 545		return -ENOMEM;
 546
 547
 548	idr_for_each_entry(&p->event_idr, ev, ev_id) {
 549		struct kfd_criu_event_priv_data *ev_priv;
 550
 551		/*
 552		 * Currently, all events have same size of private_data, but the current ioctl's
 553		 * and CRIU plugin supports private_data of variable sizes
 554		 */
 555		ev_priv = &ev_privs[i];
 556
 557		ev_priv->object_type = KFD_CRIU_OBJECT_TYPE_EVENT;
 558
 559		/* We store the user_handle with the first event */
 560		if (i == 0 && p->signal_page)
 561			ev_priv->user_handle = p->signal_handle;
 562
 563		ev_priv->event_id = ev->event_id;
 564		ev_priv->auto_reset = ev->auto_reset;
 565		ev_priv->type = ev->type;
 566		ev_priv->signaled = ev->signaled;
 567
 568		if (ev_priv->type == KFD_EVENT_TYPE_MEMORY)
 569			memcpy(&ev_priv->memory_exception_data,
 570				&ev->memory_exception_data,
 571				sizeof(struct kfd_hsa_memory_exception_data));
 572		else if (ev_priv->type == KFD_EVENT_TYPE_HW_EXCEPTION)
 573			memcpy(&ev_priv->hw_exception_data,
 574				&ev->hw_exception_data,
 575				sizeof(struct kfd_hsa_hw_exception_data));
 576
 577		pr_debug("Checkpointed event[%d] id = 0x%08x auto_reset = %x type = %x signaled = %x\n",
 578			  i,
 579			  ev_priv->event_id,
 580			  ev_priv->auto_reset,
 581			  ev_priv->type,
 582			  ev_priv->signaled);
 583		i++;
 584	}
 585
 586	ret = copy_to_user(user_priv_data + *priv_data_offset,
 587			   ev_privs, num_events * sizeof(*ev_privs));
 588	if (ret) {
 589		pr_err("Failed to copy events priv to user\n");
 590		ret = -EFAULT;
 591	}
 592
 593	*priv_data_offset += num_events * sizeof(*ev_privs);
 594
 595	kvfree(ev_privs);
 596	return ret;
 597}
 598
 599int kfd_get_num_events(struct kfd_process *p)
 600{
 601	struct kfd_event *ev;
 602	uint32_t id;
 603	u32 num_events = 0;
 604
 605	idr_for_each_entry(&p->event_idr, ev, id)
 606		num_events++;
 607
 608	return num_events;
 609}
 610
 611/* Assumes that p is current. */
 612int kfd_event_destroy(struct kfd_process *p, uint32_t event_id)
 613{
 614	struct kfd_event *ev;
 615	int ret = 0;
 616
 617	mutex_lock(&p->event_mutex);
 618
 619	ev = lookup_event_by_id(p, event_id);
 620
 621	if (ev)
 622		destroy_event(p, ev);
 623	else
 624		ret = -EINVAL;
 625
 626	mutex_unlock(&p->event_mutex);
 627	return ret;
 628}
 629
 630static void set_event(struct kfd_event *ev)
 631{
 632	struct kfd_event_waiter *waiter;
 633
 634	/* Auto reset if the list is non-empty and we're waking
 635	 * someone. waitqueue_active is safe here because we're
 636	 * protected by the ev->lock, which is also held when
 637	 * updating the wait queues in kfd_wait_on_events.
 638	 */
 639	ev->signaled = !ev->auto_reset || !waitqueue_active(&ev->wq);
 640	if (!(++ev->event_age)) {
 641		/* Never wrap back to reserved/default event age 0/1 */
 642		ev->event_age = 2;
 643		WARN_ONCE(1, "event_age wrap back!");
 644	}
 645
 646	list_for_each_entry(waiter, &ev->wq.head, wait.entry)
 647		WRITE_ONCE(waiter->activated, true);
 648
 649	wake_up_all(&ev->wq);
 650}
 651
 652/* Assumes that p is current. */
 653int kfd_set_event(struct kfd_process *p, uint32_t event_id)
 654{
 655	int ret = 0;
 656	struct kfd_event *ev;
 657
 658	rcu_read_lock();
 659
 660	ev = lookup_event_by_id(p, event_id);
 661	if (!ev) {
 662		ret = -EINVAL;
 663		goto unlock_rcu;
 664	}
 665	spin_lock(&ev->lock);
 666
 667	if (event_can_be_cpu_signaled(ev))
 668		set_event(ev);
 669	else
 670		ret = -EINVAL;
 671
 672	spin_unlock(&ev->lock);
 673unlock_rcu:
 674	rcu_read_unlock();
 675	return ret;
 676}
 677
 678static void reset_event(struct kfd_event *ev)
 679{
 680	ev->signaled = false;
 681}
 682
 683/* Assumes that p is current. */
 684int kfd_reset_event(struct kfd_process *p, uint32_t event_id)
 685{
 686	int ret = 0;
 687	struct kfd_event *ev;
 688
 689	rcu_read_lock();
 690
 691	ev = lookup_event_by_id(p, event_id);
 692	if (!ev) {
 693		ret = -EINVAL;
 694		goto unlock_rcu;
 695	}
 696	spin_lock(&ev->lock);
 697
 698	if (event_can_be_cpu_signaled(ev))
 699		reset_event(ev);
 700	else
 701		ret = -EINVAL;
 702
 703	spin_unlock(&ev->lock);
 704unlock_rcu:
 705	rcu_read_unlock();
 706	return ret;
 707
 708}
 709
 710static void acknowledge_signal(struct kfd_process *p, struct kfd_event *ev)
 711{
 712	WRITE_ONCE(page_slots(p->signal_page)[ev->event_id], UNSIGNALED_EVENT_SLOT);
 713}
 714
 715static void set_event_from_interrupt(struct kfd_process *p,
 716					struct kfd_event *ev)
 717{
 718	if (ev && event_can_be_gpu_signaled(ev)) {
 719		acknowledge_signal(p, ev);
 720		spin_lock(&ev->lock);
 721		set_event(ev);
 722		spin_unlock(&ev->lock);
 723	}
 724}
 725
 726void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
 727				uint32_t valid_id_bits)
 728{
 729	struct kfd_event *ev = NULL;
 730
 731	/*
 732	 * Because we are called from arbitrary context (workqueue) as opposed
 733	 * to process context, kfd_process could attempt to exit while we are
 734	 * running so the lookup function increments the process ref count.
 735	 */
 736	struct kfd_process *p = kfd_lookup_process_by_pasid(pasid, NULL);
 737
 738	if (!p)
 739		return; /* Presumably process exited. */
 740
 741	rcu_read_lock();
 742
 743	if (valid_id_bits)
 744		ev = lookup_signaled_event_by_partial_id(p, partial_id,
 745							 valid_id_bits);
 746	if (ev) {
 747		set_event_from_interrupt(p, ev);
 748	} else if (p->signal_page) {
 749		/*
 750		 * Partial ID lookup failed. Assume that the event ID
 751		 * in the interrupt payload was invalid and do an
 752		 * exhaustive search of signaled events.
 753		 */
 754		uint64_t *slots = page_slots(p->signal_page);
 755		uint32_t id;
 756
 757		if (valid_id_bits)
 758			pr_debug_ratelimited("Partial ID invalid: %u (%u valid bits)\n",
 759					     partial_id, valid_id_bits);
 760
 761		if (p->signal_event_count < KFD_SIGNAL_EVENT_LIMIT / 64) {
 762			/* With relatively few events, it's faster to
 763			 * iterate over the event IDR
 764			 */
 765			idr_for_each_entry(&p->event_idr, ev, id) {
 766				if (id >= KFD_SIGNAL_EVENT_LIMIT)
 767					break;
 768
 769				if (READ_ONCE(slots[id]) != UNSIGNALED_EVENT_SLOT)
 770					set_event_from_interrupt(p, ev);
 771			}
 772		} else {
 773			/* With relatively many events, it's faster to
 774			 * iterate over the signal slots and lookup
 775			 * only signaled events from the IDR.
 776			 */
 777			for (id = 1; id < KFD_SIGNAL_EVENT_LIMIT; id++)
 778				if (READ_ONCE(slots[id]) != UNSIGNALED_EVENT_SLOT) {
 779					ev = lookup_event_by_id(p, id);
 780					set_event_from_interrupt(p, ev);
 781				}
 782		}
 783	}
 784
 785	rcu_read_unlock();
 786	kfd_unref_process(p);
 787}
 788
 789static struct kfd_event_waiter *alloc_event_waiters(uint32_t num_events)
 790{
 791	struct kfd_event_waiter *event_waiters;
 792	uint32_t i;
 793
 794	event_waiters = kcalloc(num_events, sizeof(struct kfd_event_waiter),
 795				GFP_KERNEL);
 796	if (!event_waiters)
 797		return NULL;
 798
 799	for (i = 0; i < num_events; i++)
 800		init_wait(&event_waiters[i].wait);
 801
 802	return event_waiters;
 803}
 804
 805static int init_event_waiter(struct kfd_process *p,
 806		struct kfd_event_waiter *waiter,
 807		struct kfd_event_data *event_data)
 808{
 809	struct kfd_event *ev = lookup_event_by_id(p, event_data->event_id);
 810
 811	if (!ev)
 812		return -EINVAL;
 813
 814	spin_lock(&ev->lock);
 815	waiter->event = ev;
 816	waiter->activated = ev->signaled;
 817	ev->signaled = ev->signaled && !ev->auto_reset;
 818
 819	/* last_event_age = 0 reserved for backward compatible */
 820	if (waiter->event->type == KFD_EVENT_TYPE_SIGNAL &&
 821		event_data->signal_event_data.last_event_age) {
 822		waiter->event_age_enabled = true;
 823		if (ev->event_age != event_data->signal_event_data.last_event_age)
 824			waiter->activated = true;
 825	}
 826
 827	if (!waiter->activated)
 828		add_wait_queue(&ev->wq, &waiter->wait);
 829	spin_unlock(&ev->lock);
 830
 831	return 0;
 832}
 833
 834/* test_event_condition - Test condition of events being waited for
 835 * @all:           Return completion only if all events have signaled
 836 * @num_events:    Number of events to wait for
 837 * @event_waiters: Array of event waiters, one per event
 838 *
 839 * Returns KFD_IOC_WAIT_RESULT_COMPLETE if all (or one) event(s) have
 840 * signaled. Returns KFD_IOC_WAIT_RESULT_TIMEOUT if no (or not all)
 841 * events have signaled. Returns KFD_IOC_WAIT_RESULT_FAIL if any of
 842 * the events have been destroyed.
 843 */
 844static uint32_t test_event_condition(bool all, uint32_t num_events,
 845				struct kfd_event_waiter *event_waiters)
 846{
 847	uint32_t i;
 848	uint32_t activated_count = 0;
 849
 850	for (i = 0; i < num_events; i++) {
 851		if (!READ_ONCE(event_waiters[i].event))
 852			return KFD_IOC_WAIT_RESULT_FAIL;
 853
 854		if (READ_ONCE(event_waiters[i].activated)) {
 855			if (!all)
 856				return KFD_IOC_WAIT_RESULT_COMPLETE;
 857
 858			activated_count++;
 859		}
 860	}
 861
 862	return activated_count == num_events ?
 863		KFD_IOC_WAIT_RESULT_COMPLETE : KFD_IOC_WAIT_RESULT_TIMEOUT;
 864}
 865
 866/*
 867 * Copy event specific data, if defined.
 868 * Currently only memory exception events have additional data to copy to user
 869 */
 870static int copy_signaled_event_data(uint32_t num_events,
 871		struct kfd_event_waiter *event_waiters,
 872		struct kfd_event_data __user *data)
 873{
 874	void *src;
 875	void __user *dst;
 876	struct kfd_event_waiter *waiter;
 877	struct kfd_event *event;
 878	uint32_t i, size = 0;
 879
 880	for (i = 0; i < num_events; i++) {
 881		waiter = &event_waiters[i];
 882		event = waiter->event;
 883		if (!event)
 884			return -EINVAL; /* event was destroyed */
 885		if (waiter->activated) {
 886			if (event->type == KFD_EVENT_TYPE_MEMORY) {
 887				dst = &data[i].memory_exception_data;
 888				src = &event->memory_exception_data;
 889				size = sizeof(struct kfd_hsa_memory_exception_data);
 890			} else if (event->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
 891				dst = &data[i].memory_exception_data;
 892				src = &event->hw_exception_data;
 893				size = sizeof(struct kfd_hsa_hw_exception_data);
 894			} else if (event->type == KFD_EVENT_TYPE_SIGNAL &&
 895				waiter->event_age_enabled) {
 896				dst = &data[i].signal_event_data.last_event_age;
 897				src = &event->event_age;
 898				size = sizeof(u64);
 899			}
 900			if (size && copy_to_user(dst, src, size))
 901				return -EFAULT;
 902		}
 903	}
 904
 905	return 0;
 906}
 907
 908static long user_timeout_to_jiffies(uint32_t user_timeout_ms)
 909{
 910	if (user_timeout_ms == KFD_EVENT_TIMEOUT_IMMEDIATE)
 911		return 0;
 912
 913	if (user_timeout_ms == KFD_EVENT_TIMEOUT_INFINITE)
 914		return MAX_SCHEDULE_TIMEOUT;
 915
 916	/*
 917	 * msecs_to_jiffies interprets all values above 2^31-1 as infinite,
 918	 * but we consider them finite.
 919	 * This hack is wrong, but nobody is likely to notice.
 920	 */
 921	user_timeout_ms = min_t(uint32_t, user_timeout_ms, 0x7FFFFFFF);
 922
 923	return msecs_to_jiffies(user_timeout_ms) + 1;
 924}
 925
 926static void free_waiters(uint32_t num_events, struct kfd_event_waiter *waiters,
 927			 bool undo_auto_reset)
 928{
 929	uint32_t i;
 930
 931	for (i = 0; i < num_events; i++)
 932		if (waiters[i].event) {
 933			spin_lock(&waiters[i].event->lock);
 934			remove_wait_queue(&waiters[i].event->wq,
 935					  &waiters[i].wait);
 936			if (undo_auto_reset && waiters[i].activated &&
 937			    waiters[i].event && waiters[i].event->auto_reset)
 938				set_event(waiters[i].event);
 939			spin_unlock(&waiters[i].event->lock);
 940		}
 941
 942	kfree(waiters);
 943}
 944
 945int kfd_wait_on_events(struct kfd_process *p,
 946		       uint32_t num_events, void __user *data,
 947		       bool all, uint32_t *user_timeout_ms,
 948		       uint32_t *wait_result)
 949{
 950	struct kfd_event_data __user *events =
 951			(struct kfd_event_data __user *) data;
 952	uint32_t i;
 953	int ret = 0;
 954
 955	struct kfd_event_waiter *event_waiters = NULL;
 956	long timeout = user_timeout_to_jiffies(*user_timeout_ms);
 957
 958	event_waiters = alloc_event_waiters(num_events);
 959	if (!event_waiters) {
 960		ret = -ENOMEM;
 961		goto out;
 962	}
 963
 964	/* Use p->event_mutex here to protect against concurrent creation and
 965	 * destruction of events while we initialize event_waiters.
 966	 */
 967	mutex_lock(&p->event_mutex);
 968
 969	for (i = 0; i < num_events; i++) {
 970		struct kfd_event_data event_data;
 971
 972		if (copy_from_user(&event_data, &events[i],
 973				sizeof(struct kfd_event_data))) {
 974			ret = -EFAULT;
 975			goto out_unlock;
 976		}
 977
 978		ret = init_event_waiter(p, &event_waiters[i], &event_data);
 979		if (ret)
 980			goto out_unlock;
 981	}
 982
 983	/* Check condition once. */
 984	*wait_result = test_event_condition(all, num_events, event_waiters);
 985	if (*wait_result == KFD_IOC_WAIT_RESULT_COMPLETE) {
 986		ret = copy_signaled_event_data(num_events,
 987					       event_waiters, events);
 988		goto out_unlock;
 989	} else if (WARN_ON(*wait_result == KFD_IOC_WAIT_RESULT_FAIL)) {
 990		/* This should not happen. Events shouldn't be
 991		 * destroyed while we're holding the event_mutex
 992		 */
 993		goto out_unlock;
 994	}
 995
 996	mutex_unlock(&p->event_mutex);
 997
 998	while (true) {
 999		if (fatal_signal_pending(current)) {
1000			ret = -EINTR;
1001			break;
1002		}
1003
1004		if (signal_pending(current)) {
1005			ret = -ERESTARTSYS;
1006			if (*user_timeout_ms != KFD_EVENT_TIMEOUT_IMMEDIATE &&
1007			    *user_timeout_ms != KFD_EVENT_TIMEOUT_INFINITE)
1008				*user_timeout_ms = jiffies_to_msecs(
1009					max(0l, timeout-1));
1010			break;
1011		}
1012
1013		/* Set task state to interruptible sleep before
1014		 * checking wake-up conditions. A concurrent wake-up
1015		 * will put the task back into runnable state. In that
1016		 * case schedule_timeout will not put the task to
1017		 * sleep and we'll get a chance to re-check the
1018		 * updated conditions almost immediately. Otherwise,
1019		 * this race condition would lead to a soft hang or a
1020		 * very long sleep.
1021		 */
1022		set_current_state(TASK_INTERRUPTIBLE);
1023
1024		*wait_result = test_event_condition(all, num_events,
1025						    event_waiters);
1026		if (*wait_result != KFD_IOC_WAIT_RESULT_TIMEOUT)
1027			break;
1028
1029		if (timeout <= 0)
1030			break;
1031
1032		timeout = schedule_timeout(timeout);
1033	}
1034	__set_current_state(TASK_RUNNING);
1035
1036	mutex_lock(&p->event_mutex);
1037	/* copy_signaled_event_data may sleep. So this has to happen
1038	 * after the task state is set back to RUNNING.
1039	 *
1040	 * The event may also have been destroyed after signaling. So
1041	 * copy_signaled_event_data also must confirm that the event
1042	 * still exists. Therefore this must be under the p->event_mutex
1043	 * which is also held when events are destroyed.
1044	 */
1045	if (!ret && *wait_result == KFD_IOC_WAIT_RESULT_COMPLETE)
1046		ret = copy_signaled_event_data(num_events,
1047					       event_waiters, events);
1048
1049out_unlock:
1050	free_waiters(num_events, event_waiters, ret == -ERESTARTSYS);
1051	mutex_unlock(&p->event_mutex);
1052out:
1053	if (ret)
1054		*wait_result = KFD_IOC_WAIT_RESULT_FAIL;
1055	else if (*wait_result == KFD_IOC_WAIT_RESULT_FAIL)
1056		ret = -EIO;
1057
1058	return ret;
1059}
1060
1061int kfd_event_mmap(struct kfd_process *p, struct vm_area_struct *vma)
1062{
1063	unsigned long pfn;
1064	struct kfd_signal_page *page;
1065	int ret;
1066
1067	/* check required size doesn't exceed the allocated size */
1068	if (get_order(KFD_SIGNAL_EVENT_LIMIT * 8) <
1069			get_order(vma->vm_end - vma->vm_start)) {
1070		pr_err("Event page mmap requested illegal size\n");
1071		return -EINVAL;
1072	}
1073
1074	page = p->signal_page;
1075	if (!page) {
1076		/* Probably KFD bug, but mmap is user-accessible. */
1077		pr_debug("Signal page could not be found\n");
1078		return -EINVAL;
1079	}
1080
1081	pfn = __pa(page->kernel_address);
1082	pfn >>= PAGE_SHIFT;
1083
1084	vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE
1085		       | VM_DONTDUMP | VM_PFNMAP);
1086
1087	pr_debug("Mapping signal page\n");
1088	pr_debug("     start user address  == 0x%08lx\n", vma->vm_start);
1089	pr_debug("     end user address    == 0x%08lx\n", vma->vm_end);
1090	pr_debug("     pfn                 == 0x%016lX\n", pfn);
1091	pr_debug("     vm_flags            == 0x%08lX\n", vma->vm_flags);
1092	pr_debug("     size                == 0x%08lX\n",
1093			vma->vm_end - vma->vm_start);
1094
1095	page->user_address = (uint64_t __user *)vma->vm_start;
1096
1097	/* mapping the page to user process */
1098	ret = remap_pfn_range(vma, vma->vm_start, pfn,
1099			vma->vm_end - vma->vm_start, vma->vm_page_prot);
1100	if (!ret)
1101		p->signal_mapped_size = vma->vm_end - vma->vm_start;
1102
1103	return ret;
1104}
1105
1106/*
1107 * Assumes that p is not going away.
1108 */
1109static void lookup_events_by_type_and_signal(struct kfd_process *p,
1110		int type, void *event_data)
1111{
1112	struct kfd_hsa_memory_exception_data *ev_data;
1113	struct kfd_event *ev;
1114	uint32_t id;
1115	bool send_signal = true;
1116
1117	ev_data = (struct kfd_hsa_memory_exception_data *) event_data;
1118
1119	rcu_read_lock();
1120
1121	id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1122	idr_for_each_entry_continue(&p->event_idr, ev, id)
1123		if (ev->type == type) {
1124			send_signal = false;
1125			dev_dbg(kfd_device,
1126					"Event found: id %X type %d",
1127					ev->event_id, ev->type);
1128			spin_lock(&ev->lock);
1129			set_event(ev);
1130			if (ev->type == KFD_EVENT_TYPE_MEMORY && ev_data)
1131				ev->memory_exception_data = *ev_data;
1132			spin_unlock(&ev->lock);
1133		}
1134
1135	if (type == KFD_EVENT_TYPE_MEMORY) {
1136		dev_warn(kfd_device,
1137			"Sending SIGSEGV to process pid %d",
1138				p->lead_thread->pid);
1139		send_sig(SIGSEGV, p->lead_thread, 0);
1140	}
1141
1142	/* Send SIGTERM no event of type "type" has been found*/
1143	if (send_signal) {
1144		if (send_sigterm) {
1145			dev_warn(kfd_device,
1146				"Sending SIGTERM to process pid %d",
1147					p->lead_thread->pid);
1148			send_sig(SIGTERM, p->lead_thread, 0);
1149		} else {
1150			dev_err(kfd_device,
1151				"Process pid %d got unhandled exception",
1152				p->lead_thread->pid);
1153		}
1154	}
1155
1156	rcu_read_unlock();
1157}
1158
1159void kfd_signal_hw_exception_event(u32 pasid)
1160{
1161	/*
1162	 * Because we are called from arbitrary context (workqueue) as opposed
1163	 * to process context, kfd_process could attempt to exit while we are
1164	 * running so the lookup function increments the process ref count.
1165	 */
1166	struct kfd_process *p = kfd_lookup_process_by_pasid(pasid, NULL);
1167
1168	if (!p)
1169		return; /* Presumably process exited. */
1170
1171	lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_HW_EXCEPTION, NULL);
1172	kfd_unref_process(p);
1173}
1174
1175void kfd_signal_vm_fault_event_with_userptr(struct kfd_process *p, uint64_t gpu_va)
1176{
1177	struct kfd_process_device *pdd;
1178	struct kfd_hsa_memory_exception_data exception_data;
1179	int i;
1180
1181	memset(&exception_data, 0, sizeof(exception_data));
1182	exception_data.va = gpu_va;
1183	exception_data.failure.NotPresent = 1;
1184
1185	// Send VM seg fault to all kfd process device
1186	for (i = 0; i < p->n_pdds; i++) {
1187		pdd = p->pdds[i];
1188		exception_data.gpu_id = pdd->user_gpu_id;
1189		kfd_evict_process_device(pdd);
1190		kfd_signal_vm_fault_event(pdd, NULL, &exception_data);
1191	}
1192}
1193
1194void kfd_signal_vm_fault_event(struct kfd_process_device *pdd,
1195				struct kfd_vm_fault_info *info,
1196				struct kfd_hsa_memory_exception_data *data)
1197{
1198	struct kfd_event *ev;
1199	uint32_t id;
1200	struct kfd_process *p = pdd->process;
1201	struct kfd_hsa_memory_exception_data memory_exception_data;
1202	int user_gpu_id;
1203
1204	user_gpu_id = kfd_process_get_user_gpu_id(p, pdd->dev->id);
1205	if (unlikely(user_gpu_id == -EINVAL)) {
1206		WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n",
1207			  pdd->dev->id);
1208		return;
1209	}
1210
1211	/* SoC15 chips and onwards will pass in data from now on. */
1212	if (!data) {
1213		memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1214		memory_exception_data.gpu_id = user_gpu_id;
1215		memory_exception_data.failure.imprecise = true;
1216
1217		/* Set failure reason */
1218		if (info) {
1219			memory_exception_data.va = (info->page_addr) <<
1220								PAGE_SHIFT;
1221			memory_exception_data.failure.NotPresent =
1222				info->prot_valid ? 1 : 0;
1223			memory_exception_data.failure.NoExecute =
1224				info->prot_exec ? 1 : 0;
1225			memory_exception_data.failure.ReadOnly =
1226				info->prot_write ? 1 : 0;
1227			memory_exception_data.failure.imprecise = 0;
1228		}
1229	}
1230
1231	rcu_read_lock();
1232
1233	id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1234	idr_for_each_entry_continue(&p->event_idr, ev, id)
1235		if (ev->type == KFD_EVENT_TYPE_MEMORY) {
1236			spin_lock(&ev->lock);
1237			ev->memory_exception_data = data ? *data :
1238							memory_exception_data;
1239			set_event(ev);
1240			spin_unlock(&ev->lock);
1241		}
1242
1243	rcu_read_unlock();
1244}
1245
1246void kfd_signal_reset_event(struct kfd_node *dev)
1247{
1248	struct kfd_hsa_hw_exception_data hw_exception_data;
1249	struct kfd_hsa_memory_exception_data memory_exception_data;
1250	struct kfd_process *p;
1251	struct kfd_event *ev;
1252	unsigned int temp;
1253	uint32_t id, idx;
1254	int reset_cause = atomic_read(&dev->sram_ecc_flag) ?
1255			KFD_HW_EXCEPTION_ECC :
1256			KFD_HW_EXCEPTION_GPU_HANG;
1257
1258	/* Whole gpu reset caused by GPU hang and memory is lost */
1259	memset(&hw_exception_data, 0, sizeof(hw_exception_data));
1260	hw_exception_data.memory_lost = 1;
1261	hw_exception_data.reset_cause = reset_cause;
1262
1263	memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1264	memory_exception_data.ErrorType = KFD_MEM_ERR_SRAM_ECC;
1265	memory_exception_data.failure.imprecise = true;
1266
1267	idx = srcu_read_lock(&kfd_processes_srcu);
1268	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1269		int user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
1270		struct kfd_process_device *pdd = kfd_get_process_device_data(dev, p);
1271
1272		if (unlikely(user_gpu_id == -EINVAL)) {
1273			WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
1274			continue;
1275		}
1276
1277		if (unlikely(!pdd)) {
1278			WARN_ONCE(1, "Could not get device data from process pid:%d\n",
1279				  p->lead_thread->pid);
1280			continue;
1281		}
1282
1283		if (dev->dqm->detect_hang_count && !pdd->has_reset_queue)
1284			continue;
1285
1286		if (dev->dqm->detect_hang_count) {
1287			struct amdgpu_task_info *ti;
1288			struct amdgpu_fpriv *drv_priv;
1289
1290			if (unlikely(amdgpu_file_to_fpriv(pdd->drm_file, &drv_priv))) {
1291				WARN_ONCE(1, "Could not get vm for device %x from pid:%d\n",
1292					  dev->id, p->lead_thread->pid);
1293				continue;
1294			}
1295
1296			ti = amdgpu_vm_get_task_info_vm(&drv_priv->vm);
1297			if (ti) {
1298				dev_err(dev->adev->dev,
1299					"Queues reset on process %s tid %d thread %s pid %d\n",
1300					ti->process_name, ti->tgid, ti->task.comm, ti->task.pid);
1301				amdgpu_vm_put_task_info(ti);
1302			}
1303		}
1304
1305		rcu_read_lock();
1306
1307		id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1308		idr_for_each_entry_continue(&p->event_idr, ev, id) {
1309			if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
1310				spin_lock(&ev->lock);
1311				ev->hw_exception_data = hw_exception_data;
1312				ev->hw_exception_data.gpu_id = user_gpu_id;
1313				set_event(ev);
1314				spin_unlock(&ev->lock);
1315			}
1316			if (ev->type == KFD_EVENT_TYPE_MEMORY &&
1317			    reset_cause == KFD_HW_EXCEPTION_ECC) {
1318				spin_lock(&ev->lock);
1319				ev->memory_exception_data = memory_exception_data;
1320				ev->memory_exception_data.gpu_id = user_gpu_id;
1321				set_event(ev);
1322				spin_unlock(&ev->lock);
1323			}
1324		}
1325
1326		rcu_read_unlock();
1327	}
1328	srcu_read_unlock(&kfd_processes_srcu, idx);
1329}
1330
1331void kfd_signal_poison_consumed_event(struct kfd_node *dev, u32 pasid)
1332{
1333	struct kfd_process *p = kfd_lookup_process_by_pasid(pasid, NULL);
1334	struct kfd_hsa_memory_exception_data memory_exception_data;
1335	struct kfd_hsa_hw_exception_data hw_exception_data;
1336	struct kfd_event *ev;
1337	uint32_t id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1338	int user_gpu_id;
1339
1340	if (!p) {
1341		dev_warn(dev->adev->dev, "Not find process with pasid:%d\n", pasid);
1342		return; /* Presumably process exited. */
1343	}
1344
1345	user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
1346	if (unlikely(user_gpu_id == -EINVAL)) {
1347		WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
1348		kfd_unref_process(p);
1349		return;
1350	}
1351
1352	memset(&hw_exception_data, 0, sizeof(hw_exception_data));
1353	hw_exception_data.gpu_id = user_gpu_id;
1354	hw_exception_data.memory_lost = 1;
1355	hw_exception_data.reset_cause = KFD_HW_EXCEPTION_ECC;
1356
1357	memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1358	memory_exception_data.ErrorType = KFD_MEM_ERR_POISON_CONSUMED;
1359	memory_exception_data.gpu_id = user_gpu_id;
1360	memory_exception_data.failure.imprecise = true;
1361
1362	rcu_read_lock();
1363
1364	idr_for_each_entry_continue(&p->event_idr, ev, id) {
1365		if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
1366			spin_lock(&ev->lock);
1367			ev->hw_exception_data = hw_exception_data;
1368			set_event(ev);
1369			spin_unlock(&ev->lock);
1370		}
1371
1372		if (ev->type == KFD_EVENT_TYPE_MEMORY) {
1373			spin_lock(&ev->lock);
1374			ev->memory_exception_data = memory_exception_data;
1375			set_event(ev);
1376			spin_unlock(&ev->lock);
1377		}
1378	}
1379
1380	dev_warn(dev->adev->dev, "Send SIGBUS to process %s(pasid:%d)\n",
1381		p->lead_thread->comm, pasid);
1382	rcu_read_unlock();
1383
1384	/* user application will handle SIGBUS signal */
1385	send_sig(SIGBUS, p->lead_thread, 0);
1386
1387	kfd_unref_process(p);
1388}