linux/drivers/gpu/drm/amd/amdkfd/kfd_events.c
<<
>>
Prefs
   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#include <linux/mm_types.h>
  24#include <linux/slab.h>
  25#include <linux/types.h>
  26#include <linux/sched/signal.h>
  27#include <linux/sched/mm.h>
  28#include <linux/uaccess.h>
  29#include <linux/mman.h>
  30#include <linux/memory.h>
  31#include "kfd_priv.h"
  32#include "kfd_events.h"
  33#include "kfd_iommu.h"
  34#include <linux/device.h>
  35
  36/*
  37 * Wrapper around wait_queue_entry_t
  38 */
  39struct kfd_event_waiter {
  40        wait_queue_entry_t wait;
  41        struct kfd_event *event; /* Event to wait for */
  42        bool activated;          /* Becomes true when event is signaled */
  43};
  44
  45/*
  46 * Each signal event needs a 64-bit signal slot where the signaler will write
  47 * a 1 before sending an interrupt. (This is needed because some interrupts
  48 * do not contain enough spare data bits to identify an event.)
  49 * We get whole pages and map them to the process VA.
  50 * Individual signal events use their event_id as slot index.
  51 */
  52struct kfd_signal_page {
  53        uint64_t *kernel_address;
  54        uint64_t __user *user_address;
  55        bool need_to_free_pages;
  56};
  57
  58
  59static uint64_t *page_slots(struct kfd_signal_page *page)
  60{
  61        return page->kernel_address;
  62}
  63
  64static struct kfd_signal_page *allocate_signal_page(struct kfd_process *p)
  65{
  66        void *backing_store;
  67        struct kfd_signal_page *page;
  68
  69        page = kzalloc(sizeof(*page), GFP_KERNEL);
  70        if (!page)
  71                return NULL;
  72
  73        backing_store = (void *) __get_free_pages(GFP_KERNEL,
  74                                        get_order(KFD_SIGNAL_EVENT_LIMIT * 8));
  75        if (!backing_store)
  76                goto fail_alloc_signal_store;
  77
  78        /* Initialize all events to unsignaled */
  79        memset(backing_store, (uint8_t) UNSIGNALED_EVENT_SLOT,
  80               KFD_SIGNAL_EVENT_LIMIT * 8);
  81
  82        page->kernel_address = backing_store;
  83        page->need_to_free_pages = true;
  84        pr_debug("Allocated new event signal page at %p, for process %p\n",
  85                        page, p);
  86
  87        return page;
  88
  89fail_alloc_signal_store:
  90        kfree(page);
  91        return NULL;
  92}
  93
  94static int allocate_event_notification_slot(struct kfd_process *p,
  95                                            struct kfd_event *ev)
  96{
  97        int id;
  98
  99        if (!p->signal_page) {
 100                p->signal_page = allocate_signal_page(p);
 101                if (!p->signal_page)
 102                        return -ENOMEM;
 103                /* Oldest user mode expects 256 event slots */
 104                p->signal_mapped_size = 256*8;
 105        }
 106
 107        /*
 108         * Compatibility with old user mode: Only use signal slots
 109         * user mode has mapped, may be less than
 110         * KFD_SIGNAL_EVENT_LIMIT. This also allows future increase
 111         * of the event limit without breaking user mode.
 112         */
 113        id = idr_alloc(&p->event_idr, ev, 0, p->signal_mapped_size / 8,
 114                       GFP_KERNEL);
 115        if (id < 0)
 116                return id;
 117
 118        ev->event_id = id;
 119        page_slots(p->signal_page)[id] = UNSIGNALED_EVENT_SLOT;
 120
 121        return 0;
 122}
 123
 124/*
 125 * Assumes that p->event_mutex is held and of course that p is not going
 126 * away (current or locked).
 127 */
 128static struct kfd_event *lookup_event_by_id(struct kfd_process *p, uint32_t id)
 129{
 130        return idr_find(&p->event_idr, id);
 131}
 132
 133/**
 134 * lookup_signaled_event_by_partial_id - Lookup signaled event from partial ID
 135 * @p:     Pointer to struct kfd_process
 136 * @id:    ID to look up
 137 * @bits:  Number of valid bits in @id
 138 *
 139 * Finds the first signaled event with a matching partial ID. If no
 140 * matching signaled event is found, returns NULL. In that case the
 141 * caller should assume that the partial ID is invalid and do an
 142 * exhaustive search of all siglaned events.
 143 *
 144 * If multiple events with the same partial ID signal at the same
 145 * time, they will be found one interrupt at a time, not necessarily
 146 * in the same order the interrupts occurred. As long as the number of
 147 * interrupts is correct, all signaled events will be seen by the
 148 * driver.
 149 */
 150static struct kfd_event *lookup_signaled_event_by_partial_id(
 151        struct kfd_process *p, uint32_t id, uint32_t bits)
 152{
 153        struct kfd_event *ev;
 154
 155        if (!p->signal_page || id >= KFD_SIGNAL_EVENT_LIMIT)
 156                return NULL;
 157
 158        /* Fast path for the common case that @id is not a partial ID
 159         * and we only need a single lookup.
 160         */
 161        if (bits > 31 || (1U << bits) >= KFD_SIGNAL_EVENT_LIMIT) {
 162                if (page_slots(p->signal_page)[id] == UNSIGNALED_EVENT_SLOT)
 163                        return NULL;
 164
 165                return idr_find(&p->event_idr, id);
 166        }
 167
 168        /* General case for partial IDs: Iterate over all matching IDs
 169         * and find the first one that has signaled.
 170         */
 171        for (ev = NULL; id < KFD_SIGNAL_EVENT_LIMIT && !ev; id += 1U << bits) {
 172                if (page_slots(p->signal_page)[id] == UNSIGNALED_EVENT_SLOT)
 173                        continue;
 174
 175                ev = idr_find(&p->event_idr, id);
 176        }
 177
 178        return ev;
 179}
 180
 181static int create_signal_event(struct file *devkfd,
 182                                struct kfd_process *p,
 183                                struct kfd_event *ev)
 184{
 185        int ret;
 186
 187        if (p->signal_mapped_size &&
 188            p->signal_event_count == p->signal_mapped_size / 8) {
 189                if (!p->signal_event_limit_reached) {
 190                        pr_debug("Signal event wasn't created because limit was reached\n");
 191                        p->signal_event_limit_reached = true;
 192                }
 193                return -ENOSPC;
 194        }
 195
 196        ret = allocate_event_notification_slot(p, ev);
 197        if (ret) {
 198                pr_warn("Signal event wasn't created because out of kernel memory\n");
 199                return ret;
 200        }
 201
 202        p->signal_event_count++;
 203
 204        ev->user_signal_address = &p->signal_page->user_address[ev->event_id];
 205        pr_debug("Signal event number %zu created with id %d, address %p\n",
 206                        p->signal_event_count, ev->event_id,
 207                        ev->user_signal_address);
 208
 209        return 0;
 210}
 211
 212static int create_other_event(struct kfd_process *p, struct kfd_event *ev)
 213{
 214        /* Cast KFD_LAST_NONSIGNAL_EVENT to uint32_t. This allows an
 215         * intentional integer overflow to -1 without a compiler
 216         * warning. idr_alloc treats a negative value as "maximum
 217         * signed integer".
 218         */
 219        int id = idr_alloc(&p->event_idr, ev, KFD_FIRST_NONSIGNAL_EVENT_ID,
 220                           (uint32_t)KFD_LAST_NONSIGNAL_EVENT_ID + 1,
 221                           GFP_KERNEL);
 222
 223        if (id < 0)
 224                return id;
 225        ev->event_id = id;
 226
 227        return 0;
 228}
 229
 230void kfd_event_init_process(struct kfd_process *p)
 231{
 232        mutex_init(&p->event_mutex);
 233        idr_init(&p->event_idr);
 234        p->signal_page = NULL;
 235        p->signal_event_count = 0;
 236}
 237
 238static void destroy_event(struct kfd_process *p, struct kfd_event *ev)
 239{
 240        struct kfd_event_waiter *waiter;
 241
 242        /* Wake up pending waiters. They will return failure */
 243        list_for_each_entry(waiter, &ev->wq.head, wait.entry)
 244                waiter->event = NULL;
 245        wake_up_all(&ev->wq);
 246
 247        if (ev->type == KFD_EVENT_TYPE_SIGNAL ||
 248            ev->type == KFD_EVENT_TYPE_DEBUG)
 249                p->signal_event_count--;
 250
 251        idr_remove(&p->event_idr, ev->event_id);
 252        kfree(ev);
 253}
 254
 255static void destroy_events(struct kfd_process *p)
 256{
 257        struct kfd_event *ev;
 258        uint32_t id;
 259
 260        idr_for_each_entry(&p->event_idr, ev, id)
 261                destroy_event(p, ev);
 262        idr_destroy(&p->event_idr);
 263}
 264
 265/*
 266 * We assume that the process is being destroyed and there is no need to
 267 * unmap the pages or keep bookkeeping data in order.
 268 */
 269static void shutdown_signal_page(struct kfd_process *p)
 270{
 271        struct kfd_signal_page *page = p->signal_page;
 272
 273        if (page) {
 274                if (page->need_to_free_pages)
 275                        free_pages((unsigned long)page->kernel_address,
 276                                   get_order(KFD_SIGNAL_EVENT_LIMIT * 8));
 277                kfree(page);
 278        }
 279}
 280
 281void kfd_event_free_process(struct kfd_process *p)
 282{
 283        destroy_events(p);
 284        shutdown_signal_page(p);
 285}
 286
 287static bool event_can_be_gpu_signaled(const struct kfd_event *ev)
 288{
 289        return ev->type == KFD_EVENT_TYPE_SIGNAL ||
 290                                        ev->type == KFD_EVENT_TYPE_DEBUG;
 291}
 292
 293static bool event_can_be_cpu_signaled(const struct kfd_event *ev)
 294{
 295        return ev->type == KFD_EVENT_TYPE_SIGNAL;
 296}
 297
 298int kfd_event_page_set(struct kfd_process *p, void *kernel_address,
 299                       uint64_t size)
 300{
 301        struct kfd_signal_page *page;
 302
 303        if (p->signal_page)
 304                return -EBUSY;
 305
 306        page = kzalloc(sizeof(*page), GFP_KERNEL);
 307        if (!page)
 308                return -ENOMEM;
 309
 310        /* Initialize all events to unsignaled */
 311        memset(kernel_address, (uint8_t) UNSIGNALED_EVENT_SLOT,
 312               KFD_SIGNAL_EVENT_LIMIT * 8);
 313
 314        page->kernel_address = kernel_address;
 315
 316        p->signal_page = page;
 317        p->signal_mapped_size = size;
 318
 319        return 0;
 320}
 321
 322int kfd_event_create(struct file *devkfd, struct kfd_process *p,
 323                     uint32_t event_type, bool auto_reset, uint32_t node_id,
 324                     uint32_t *event_id, uint32_t *event_trigger_data,
 325                     uint64_t *event_page_offset, uint32_t *event_slot_index)
 326{
 327        int ret = 0;
 328        struct kfd_event *ev = kzalloc(sizeof(*ev), GFP_KERNEL);
 329
 330        if (!ev)
 331                return -ENOMEM;
 332
 333        ev->type = event_type;
 334        ev->auto_reset = auto_reset;
 335        ev->signaled = false;
 336
 337        init_waitqueue_head(&ev->wq);
 338
 339        *event_page_offset = 0;
 340
 341        mutex_lock(&p->event_mutex);
 342
 343        switch (event_type) {
 344        case KFD_EVENT_TYPE_SIGNAL:
 345        case KFD_EVENT_TYPE_DEBUG:
 346                ret = create_signal_event(devkfd, p, ev);
 347                if (!ret) {
 348                        *event_page_offset = KFD_MMAP_TYPE_EVENTS;
 349                        *event_slot_index = ev->event_id;
 350                }
 351                break;
 352        default:
 353                ret = create_other_event(p, ev);
 354                break;
 355        }
 356
 357        if (!ret) {
 358                *event_id = ev->event_id;
 359                *event_trigger_data = ev->event_id;
 360        } else {
 361                kfree(ev);
 362        }
 363
 364        mutex_unlock(&p->event_mutex);
 365
 366        return ret;
 367}
 368
 369/* Assumes that p is current. */
 370int kfd_event_destroy(struct kfd_process *p, uint32_t event_id)
 371{
 372        struct kfd_event *ev;
 373        int ret = 0;
 374
 375        mutex_lock(&p->event_mutex);
 376
 377        ev = lookup_event_by_id(p, event_id);
 378
 379        if (ev)
 380                destroy_event(p, ev);
 381        else
 382                ret = -EINVAL;
 383
 384        mutex_unlock(&p->event_mutex);
 385        return ret;
 386}
 387
 388static void set_event(struct kfd_event *ev)
 389{
 390        struct kfd_event_waiter *waiter;
 391
 392        /* Auto reset if the list is non-empty and we're waking
 393         * someone. waitqueue_active is safe here because we're
 394         * protected by the p->event_mutex, which is also held when
 395         * updating the wait queues in kfd_wait_on_events.
 396         */
 397        ev->signaled = !ev->auto_reset || !waitqueue_active(&ev->wq);
 398
 399        list_for_each_entry(waiter, &ev->wq.head, wait.entry)
 400                waiter->activated = true;
 401
 402        wake_up_all(&ev->wq);
 403}
 404
 405/* Assumes that p is current. */
 406int kfd_set_event(struct kfd_process *p, uint32_t event_id)
 407{
 408        int ret = 0;
 409        struct kfd_event *ev;
 410
 411        mutex_lock(&p->event_mutex);
 412
 413        ev = lookup_event_by_id(p, event_id);
 414
 415        if (ev && event_can_be_cpu_signaled(ev))
 416                set_event(ev);
 417        else
 418                ret = -EINVAL;
 419
 420        mutex_unlock(&p->event_mutex);
 421        return ret;
 422}
 423
 424static void reset_event(struct kfd_event *ev)
 425{
 426        ev->signaled = false;
 427}
 428
 429/* Assumes that p is current. */
 430int kfd_reset_event(struct kfd_process *p, uint32_t event_id)
 431{
 432        int ret = 0;
 433        struct kfd_event *ev;
 434
 435        mutex_lock(&p->event_mutex);
 436
 437        ev = lookup_event_by_id(p, event_id);
 438
 439        if (ev && event_can_be_cpu_signaled(ev))
 440                reset_event(ev);
 441        else
 442                ret = -EINVAL;
 443
 444        mutex_unlock(&p->event_mutex);
 445        return ret;
 446
 447}
 448
 449static void acknowledge_signal(struct kfd_process *p, struct kfd_event *ev)
 450{
 451        page_slots(p->signal_page)[ev->event_id] = UNSIGNALED_EVENT_SLOT;
 452}
 453
 454static void set_event_from_interrupt(struct kfd_process *p,
 455                                        struct kfd_event *ev)
 456{
 457        if (ev && event_can_be_gpu_signaled(ev)) {
 458                acknowledge_signal(p, ev);
 459                set_event(ev);
 460        }
 461}
 462
 463void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
 464                                uint32_t valid_id_bits)
 465{
 466        struct kfd_event *ev = NULL;
 467
 468        /*
 469         * Because we are called from arbitrary context (workqueue) as opposed
 470         * to process context, kfd_process could attempt to exit while we are
 471         * running so the lookup function increments the process ref count.
 472         */
 473        struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
 474
 475        if (!p)
 476                return; /* Presumably process exited. */
 477
 478        mutex_lock(&p->event_mutex);
 479
 480        if (valid_id_bits)
 481                ev = lookup_signaled_event_by_partial_id(p, partial_id,
 482                                                         valid_id_bits);
 483        if (ev) {
 484                set_event_from_interrupt(p, ev);
 485        } else if (p->signal_page) {
 486                /*
 487                 * Partial ID lookup failed. Assume that the event ID
 488                 * in the interrupt payload was invalid and do an
 489                 * exhaustive search of signaled events.
 490                 */
 491                uint64_t *slots = page_slots(p->signal_page);
 492                uint32_t id;
 493
 494                if (valid_id_bits)
 495                        pr_debug_ratelimited("Partial ID invalid: %u (%u valid bits)\n",
 496                                             partial_id, valid_id_bits);
 497
 498                if (p->signal_event_count < KFD_SIGNAL_EVENT_LIMIT / 64) {
 499                        /* With relatively few events, it's faster to
 500                         * iterate over the event IDR
 501                         */
 502                        idr_for_each_entry(&p->event_idr, ev, id) {
 503                                if (id >= KFD_SIGNAL_EVENT_LIMIT)
 504                                        break;
 505
 506                                if (slots[id] != UNSIGNALED_EVENT_SLOT)
 507                                        set_event_from_interrupt(p, ev);
 508                        }
 509                } else {
 510                        /* With relatively many events, it's faster to
 511                         * iterate over the signal slots and lookup
 512                         * only signaled events from the IDR.
 513                         */
 514                        for (id = 0; id < KFD_SIGNAL_EVENT_LIMIT; id++)
 515                                if (slots[id] != UNSIGNALED_EVENT_SLOT) {
 516                                        ev = lookup_event_by_id(p, id);
 517                                        set_event_from_interrupt(p, ev);
 518                                }
 519                }
 520        }
 521
 522        mutex_unlock(&p->event_mutex);
 523        kfd_unref_process(p);
 524}
 525
 526static struct kfd_event_waiter *alloc_event_waiters(uint32_t num_events)
 527{
 528        struct kfd_event_waiter *event_waiters;
 529        uint32_t i;
 530
 531        event_waiters = kmalloc_array(num_events,
 532                                        sizeof(struct kfd_event_waiter),
 533                                        GFP_KERNEL);
 534
 535        for (i = 0; (event_waiters) && (i < num_events) ; i++) {
 536                init_wait(&event_waiters[i].wait);
 537                event_waiters[i].activated = false;
 538        }
 539
 540        return event_waiters;
 541}
 542
 543static int init_event_waiter_get_status(struct kfd_process *p,
 544                struct kfd_event_waiter *waiter,
 545                uint32_t event_id)
 546{
 547        struct kfd_event *ev = lookup_event_by_id(p, event_id);
 548
 549        if (!ev)
 550                return -EINVAL;
 551
 552        waiter->event = ev;
 553        waiter->activated = ev->signaled;
 554        ev->signaled = ev->signaled && !ev->auto_reset;
 555
 556        return 0;
 557}
 558
 559static void init_event_waiter_add_to_waitlist(struct kfd_event_waiter *waiter)
 560{
 561        struct kfd_event *ev = waiter->event;
 562
 563        /* Only add to the wait list if we actually need to
 564         * wait on this event.
 565         */
 566        if (!waiter->activated)
 567                add_wait_queue(&ev->wq, &waiter->wait);
 568}
 569
 570/* test_event_condition - Test condition of events being waited for
 571 * @all:           Return completion only if all events have signaled
 572 * @num_events:    Number of events to wait for
 573 * @event_waiters: Array of event waiters, one per event
 574 *
 575 * Returns KFD_IOC_WAIT_RESULT_COMPLETE if all (or one) event(s) have
 576 * signaled. Returns KFD_IOC_WAIT_RESULT_TIMEOUT if no (or not all)
 577 * events have signaled. Returns KFD_IOC_WAIT_RESULT_FAIL if any of
 578 * the events have been destroyed.
 579 */
 580static uint32_t test_event_condition(bool all, uint32_t num_events,
 581                                struct kfd_event_waiter *event_waiters)
 582{
 583        uint32_t i;
 584        uint32_t activated_count = 0;
 585
 586        for (i = 0; i < num_events; i++) {
 587                if (!event_waiters[i].event)
 588                        return KFD_IOC_WAIT_RESULT_FAIL;
 589
 590                if (event_waiters[i].activated) {
 591                        if (!all)
 592                                return KFD_IOC_WAIT_RESULT_COMPLETE;
 593
 594                        activated_count++;
 595                }
 596        }
 597
 598        return activated_count == num_events ?
 599                KFD_IOC_WAIT_RESULT_COMPLETE : KFD_IOC_WAIT_RESULT_TIMEOUT;
 600}
 601
 602/*
 603 * Copy event specific data, if defined.
 604 * Currently only memory exception events have additional data to copy to user
 605 */
 606static int copy_signaled_event_data(uint32_t num_events,
 607                struct kfd_event_waiter *event_waiters,
 608                struct kfd_event_data __user *data)
 609{
 610        struct kfd_hsa_memory_exception_data *src;
 611        struct kfd_hsa_memory_exception_data __user *dst;
 612        struct kfd_event_waiter *waiter;
 613        struct kfd_event *event;
 614        uint32_t i;
 615
 616        for (i = 0; i < num_events; i++) {
 617                waiter = &event_waiters[i];
 618                event = waiter->event;
 619                if (waiter->activated && event->type == KFD_EVENT_TYPE_MEMORY) {
 620                        dst = &data[i].memory_exception_data;
 621                        src = &event->memory_exception_data;
 622                        if (copy_to_user(dst, src,
 623                                sizeof(struct kfd_hsa_memory_exception_data)))
 624                                return -EFAULT;
 625                }
 626        }
 627
 628        return 0;
 629
 630}
 631
 632
 633
 634static long user_timeout_to_jiffies(uint32_t user_timeout_ms)
 635{
 636        if (user_timeout_ms == KFD_EVENT_TIMEOUT_IMMEDIATE)
 637                return 0;
 638
 639        if (user_timeout_ms == KFD_EVENT_TIMEOUT_INFINITE)
 640                return MAX_SCHEDULE_TIMEOUT;
 641
 642        /*
 643         * msecs_to_jiffies interprets all values above 2^31-1 as infinite,
 644         * but we consider them finite.
 645         * This hack is wrong, but nobody is likely to notice.
 646         */
 647        user_timeout_ms = min_t(uint32_t, user_timeout_ms, 0x7FFFFFFF);
 648
 649        return msecs_to_jiffies(user_timeout_ms) + 1;
 650}
 651
 652static void free_waiters(uint32_t num_events, struct kfd_event_waiter *waiters)
 653{
 654        uint32_t i;
 655
 656        for (i = 0; i < num_events; i++)
 657                if (waiters[i].event)
 658                        remove_wait_queue(&waiters[i].event->wq,
 659                                          &waiters[i].wait);
 660
 661        kfree(waiters);
 662}
 663
 664int kfd_wait_on_events(struct kfd_process *p,
 665                       uint32_t num_events, void __user *data,
 666                       bool all, uint32_t user_timeout_ms,
 667                       uint32_t *wait_result)
 668{
 669        struct kfd_event_data __user *events =
 670                        (struct kfd_event_data __user *) data;
 671        uint32_t i;
 672        int ret = 0;
 673
 674        struct kfd_event_waiter *event_waiters = NULL;
 675        long timeout = user_timeout_to_jiffies(user_timeout_ms);
 676
 677        event_waiters = alloc_event_waiters(num_events);
 678        if (!event_waiters) {
 679                ret = -ENOMEM;
 680                goto out;
 681        }
 682
 683        mutex_lock(&p->event_mutex);
 684
 685        for (i = 0; i < num_events; i++) {
 686                struct kfd_event_data event_data;
 687
 688                if (copy_from_user(&event_data, &events[i],
 689                                sizeof(struct kfd_event_data))) {
 690                        ret = -EFAULT;
 691                        goto out_unlock;
 692                }
 693
 694                ret = init_event_waiter_get_status(p, &event_waiters[i],
 695                                event_data.event_id);
 696                if (ret)
 697                        goto out_unlock;
 698        }
 699
 700        /* Check condition once. */
 701        *wait_result = test_event_condition(all, num_events, event_waiters);
 702        if (*wait_result == KFD_IOC_WAIT_RESULT_COMPLETE) {
 703                ret = copy_signaled_event_data(num_events,
 704                                               event_waiters, events);
 705                goto out_unlock;
 706        } else if (WARN_ON(*wait_result == KFD_IOC_WAIT_RESULT_FAIL)) {
 707                /* This should not happen. Events shouldn't be
 708                 * destroyed while we're holding the event_mutex
 709                 */
 710                goto out_unlock;
 711        }
 712
 713        /* Add to wait lists if we need to wait. */
 714        for (i = 0; i < num_events; i++)
 715                init_event_waiter_add_to_waitlist(&event_waiters[i]);
 716
 717        mutex_unlock(&p->event_mutex);
 718
 719        while (true) {
 720                if (fatal_signal_pending(current)) {
 721                        ret = -EINTR;
 722                        break;
 723                }
 724
 725                if (signal_pending(current)) {
 726                        /*
 727                         * This is wrong when a nonzero, non-infinite timeout
 728                         * is specified. We need to use
 729                         * ERESTARTSYS_RESTARTBLOCK, but struct restart_block
 730                         * contains a union with data for each user and it's
 731                         * in generic kernel code that I don't want to
 732                         * touch yet.
 733                         */
 734                        ret = -ERESTARTSYS;
 735                        break;
 736                }
 737
 738                /* Set task state to interruptible sleep before
 739                 * checking wake-up conditions. A concurrent wake-up
 740                 * will put the task back into runnable state. In that
 741                 * case schedule_timeout will not put the task to
 742                 * sleep and we'll get a chance to re-check the
 743                 * updated conditions almost immediately. Otherwise,
 744                 * this race condition would lead to a soft hang or a
 745                 * very long sleep.
 746                 */
 747                set_current_state(TASK_INTERRUPTIBLE);
 748
 749                *wait_result = test_event_condition(all, num_events,
 750                                                    event_waiters);
 751                if (*wait_result != KFD_IOC_WAIT_RESULT_TIMEOUT)
 752                        break;
 753
 754                if (timeout <= 0)
 755                        break;
 756
 757                timeout = schedule_timeout(timeout);
 758        }
 759        __set_current_state(TASK_RUNNING);
 760
 761        /* copy_signaled_event_data may sleep. So this has to happen
 762         * after the task state is set back to RUNNING.
 763         */
 764        if (!ret && *wait_result == KFD_IOC_WAIT_RESULT_COMPLETE)
 765                ret = copy_signaled_event_data(num_events,
 766                                               event_waiters, events);
 767
 768        mutex_lock(&p->event_mutex);
 769out_unlock:
 770        free_waiters(num_events, event_waiters);
 771        mutex_unlock(&p->event_mutex);
 772out:
 773        if (ret)
 774                *wait_result = KFD_IOC_WAIT_RESULT_FAIL;
 775        else if (*wait_result == KFD_IOC_WAIT_RESULT_FAIL)
 776                ret = -EIO;
 777
 778        return ret;
 779}
 780
 781int kfd_event_mmap(struct kfd_process *p, struct vm_area_struct *vma)
 782{
 783        unsigned long pfn;
 784        struct kfd_signal_page *page;
 785        int ret;
 786
 787        /* check required size doesn't exceed the allocated size */
 788        if (get_order(KFD_SIGNAL_EVENT_LIMIT * 8) <
 789                        get_order(vma->vm_end - vma->vm_start)) {
 790                pr_err("Event page mmap requested illegal size\n");
 791                return -EINVAL;
 792        }
 793
 794        page = p->signal_page;
 795        if (!page) {
 796                /* Probably KFD bug, but mmap is user-accessible. */
 797                pr_debug("Signal page could not be found\n");
 798                return -EINVAL;
 799        }
 800
 801        pfn = __pa(page->kernel_address);
 802        pfn >>= PAGE_SHIFT;
 803
 804        vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE
 805                       | VM_DONTDUMP | VM_PFNMAP;
 806
 807        pr_debug("Mapping signal page\n");
 808        pr_debug("     start user address  == 0x%08lx\n", vma->vm_start);
 809        pr_debug("     end user address    == 0x%08lx\n", vma->vm_end);
 810        pr_debug("     pfn                 == 0x%016lX\n", pfn);
 811        pr_debug("     vm_flags            == 0x%08lX\n", vma->vm_flags);
 812        pr_debug("     size                == 0x%08lX\n",
 813                        vma->vm_end - vma->vm_start);
 814
 815        page->user_address = (uint64_t __user *)vma->vm_start;
 816
 817        /* mapping the page to user process */
 818        ret = remap_pfn_range(vma, vma->vm_start, pfn,
 819                        vma->vm_end - vma->vm_start, vma->vm_page_prot);
 820        if (!ret)
 821                p->signal_mapped_size = vma->vm_end - vma->vm_start;
 822
 823        return ret;
 824}
 825
 826/*
 827 * Assumes that p->event_mutex is held and of course
 828 * that p is not going away (current or locked).
 829 */
 830static void lookup_events_by_type_and_signal(struct kfd_process *p,
 831                int type, void *event_data)
 832{
 833        struct kfd_hsa_memory_exception_data *ev_data;
 834        struct kfd_event *ev;
 835        uint32_t id;
 836        bool send_signal = true;
 837
 838        ev_data = (struct kfd_hsa_memory_exception_data *) event_data;
 839
 840        id = KFD_FIRST_NONSIGNAL_EVENT_ID;
 841        idr_for_each_entry_continue(&p->event_idr, ev, id)
 842                if (ev->type == type) {
 843                        send_signal = false;
 844                        dev_dbg(kfd_device,
 845                                        "Event found: id %X type %d",
 846                                        ev->event_id, ev->type);
 847                        set_event(ev);
 848                        if (ev->type == KFD_EVENT_TYPE_MEMORY && ev_data)
 849                                ev->memory_exception_data = *ev_data;
 850                }
 851
 852        if (type == KFD_EVENT_TYPE_MEMORY) {
 853                dev_warn(kfd_device,
 854                        "Sending SIGSEGV to process %d (pasid 0x%x)",
 855                                p->lead_thread->pid, p->pasid);
 856                send_sig(SIGSEGV, p->lead_thread, 0);
 857        }
 858
 859        /* Send SIGTERM no event of type "type" has been found*/
 860        if (send_signal) {
 861                if (send_sigterm) {
 862                        dev_warn(kfd_device,
 863                                "Sending SIGTERM to process %d (pasid 0x%x)",
 864                                        p->lead_thread->pid, p->pasid);
 865                        send_sig(SIGTERM, p->lead_thread, 0);
 866                } else {
 867                        dev_err(kfd_device,
 868                                "Process %d (pasid 0x%x) got unhandled exception",
 869                                p->lead_thread->pid, p->pasid);
 870                }
 871        }
 872}
 873
 874#ifdef KFD_SUPPORT_IOMMU_V2
 875void kfd_signal_iommu_event(struct kfd_dev *dev, u32 pasid,
 876                unsigned long address, bool is_write_requested,
 877                bool is_execute_requested)
 878{
 879        struct kfd_hsa_memory_exception_data memory_exception_data;
 880        struct vm_area_struct *vma;
 881
 882        /*
 883         * Because we are called from arbitrary context (workqueue) as opposed
 884         * to process context, kfd_process could attempt to exit while we are
 885         * running so the lookup function increments the process ref count.
 886         */
 887        struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
 888        struct mm_struct *mm;
 889
 890        if (!p)
 891                return; /* Presumably process exited. */
 892
 893        /* Take a safe reference to the mm_struct, which may otherwise
 894         * disappear even while the kfd_process is still referenced.
 895         */
 896        mm = get_task_mm(p->lead_thread);
 897        if (!mm) {
 898                kfd_unref_process(p);
 899                return; /* Process is exiting */
 900        }
 901
 902        memset(&memory_exception_data, 0, sizeof(memory_exception_data));
 903
 904        mmap_read_lock(mm);
 905        vma = find_vma(mm, address);
 906
 907        memory_exception_data.gpu_id = dev->id;
 908        memory_exception_data.va = address;
 909        /* Set failure reason */
 910        memory_exception_data.failure.NotPresent = 1;
 911        memory_exception_data.failure.NoExecute = 0;
 912        memory_exception_data.failure.ReadOnly = 0;
 913        if (vma && address >= vma->vm_start) {
 914                memory_exception_data.failure.NotPresent = 0;
 915
 916                if (is_write_requested && !(vma->vm_flags & VM_WRITE))
 917                        memory_exception_data.failure.ReadOnly = 1;
 918                else
 919                        memory_exception_data.failure.ReadOnly = 0;
 920
 921                if (is_execute_requested && !(vma->vm_flags & VM_EXEC))
 922                        memory_exception_data.failure.NoExecute = 1;
 923                else
 924                        memory_exception_data.failure.NoExecute = 0;
 925        }
 926
 927        mmap_read_unlock(mm);
 928        mmput(mm);
 929
 930        pr_debug("notpresent %d, noexecute %d, readonly %d\n",
 931                        memory_exception_data.failure.NotPresent,
 932                        memory_exception_data.failure.NoExecute,
 933                        memory_exception_data.failure.ReadOnly);
 934
 935        /* Workaround on Raven to not kill the process when memory is freed
 936         * before IOMMU is able to finish processing all the excessive PPRs
 937         */
 938        if (dev->device_info->asic_family != CHIP_RAVEN &&
 939            dev->device_info->asic_family != CHIP_RENOIR) {
 940                mutex_lock(&p->event_mutex);
 941
 942                /* Lookup events by type and signal them */
 943                lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_MEMORY,
 944                                &memory_exception_data);
 945
 946                mutex_unlock(&p->event_mutex);
 947        }
 948
 949        kfd_unref_process(p);
 950}
 951#endif /* KFD_SUPPORT_IOMMU_V2 */
 952
 953void kfd_signal_hw_exception_event(u32 pasid)
 954{
 955        /*
 956         * Because we are called from arbitrary context (workqueue) as opposed
 957         * to process context, kfd_process could attempt to exit while we are
 958         * running so the lookup function increments the process ref count.
 959         */
 960        struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
 961
 962        if (!p)
 963                return; /* Presumably process exited. */
 964
 965        mutex_lock(&p->event_mutex);
 966
 967        /* Lookup events by type and signal them */
 968        lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_HW_EXCEPTION, NULL);
 969
 970        mutex_unlock(&p->event_mutex);
 971        kfd_unref_process(p);
 972}
 973
 974void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid,
 975                                struct kfd_vm_fault_info *info)
 976{
 977        struct kfd_event *ev;
 978        uint32_t id;
 979        struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
 980        struct kfd_hsa_memory_exception_data memory_exception_data;
 981
 982        if (!p)
 983                return; /* Presumably process exited. */
 984        memset(&memory_exception_data, 0, sizeof(memory_exception_data));
 985        memory_exception_data.gpu_id = dev->id;
 986        memory_exception_data.failure.imprecise = true;
 987        /* Set failure reason */
 988        if (info) {
 989                memory_exception_data.va = (info->page_addr) << PAGE_SHIFT;
 990                memory_exception_data.failure.NotPresent =
 991                        info->prot_valid ? 1 : 0;
 992                memory_exception_data.failure.NoExecute =
 993                        info->prot_exec ? 1 : 0;
 994                memory_exception_data.failure.ReadOnly =
 995                        info->prot_write ? 1 : 0;
 996                memory_exception_data.failure.imprecise = 0;
 997        }
 998        mutex_lock(&p->event_mutex);
 999
1000        id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1001        idr_for_each_entry_continue(&p->event_idr, ev, id)
1002                if (ev->type == KFD_EVENT_TYPE_MEMORY) {
1003                        ev->memory_exception_data = memory_exception_data;
1004                        set_event(ev);
1005                }
1006
1007        mutex_unlock(&p->event_mutex);
1008        kfd_unref_process(p);
1009}
1010
1011void kfd_signal_reset_event(struct kfd_dev *dev)
1012{
1013        struct kfd_hsa_hw_exception_data hw_exception_data;
1014        struct kfd_hsa_memory_exception_data memory_exception_data;
1015        struct kfd_process *p;
1016        struct kfd_event *ev;
1017        unsigned int temp;
1018        uint32_t id, idx;
1019        int reset_cause = atomic_read(&dev->sram_ecc_flag) ?
1020                        KFD_HW_EXCEPTION_ECC :
1021                        KFD_HW_EXCEPTION_GPU_HANG;
1022
1023        /* Whole gpu reset caused by GPU hang and memory is lost */
1024        memset(&hw_exception_data, 0, sizeof(hw_exception_data));
1025        hw_exception_data.gpu_id = dev->id;
1026        hw_exception_data.memory_lost = 1;
1027        hw_exception_data.reset_cause = reset_cause;
1028
1029        memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1030        memory_exception_data.ErrorType = KFD_MEM_ERR_SRAM_ECC;
1031        memory_exception_data.gpu_id = dev->id;
1032        memory_exception_data.failure.imprecise = true;
1033
1034        idx = srcu_read_lock(&kfd_processes_srcu);
1035        hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1036                mutex_lock(&p->event_mutex);
1037                id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1038                idr_for_each_entry_continue(&p->event_idr, ev, id) {
1039                        if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
1040                                ev->hw_exception_data = hw_exception_data;
1041                                set_event(ev);
1042                        }
1043                        if (ev->type == KFD_EVENT_TYPE_MEMORY &&
1044                            reset_cause == KFD_HW_EXCEPTION_ECC) {
1045                                ev->memory_exception_data = memory_exception_data;
1046                                set_event(ev);
1047                        }
1048                }
1049                mutex_unlock(&p->event_mutex);
1050        }
1051        srcu_read_unlock(&kfd_processes_srcu, idx);
1052}
1053