1 // SPDX-License-Identifier: GPL-2.0 OR MIT
3 * Copyright 2014-2022 Advanced Micro Devices, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
12 * The above copyright notice and this permission notice shall be included in
13 * all copies or substantial portions of the Software.
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21 * OTHER DEALINGS IN THE SOFTWARE.
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>
33 #include "kfd_events.h"
34 #include "kfd_iommu.h"
35 #include <linux/device.h>
38 * Wrapper around wait_queue_entry_t
40 struct 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 */
47 * Each signal event needs a 64-bit signal slot where the signaler will write
48 * a 1 before sending an interrupt. (This is needed because some interrupts
49 * do not contain enough spare data bits to identify an event.)
50 * We get whole pages and map them to the process VA.
51 * Individual signal events use their event_id as slot index.
53 struct kfd_signal_page {
54 uint64_t *kernel_address;
55 uint64_t __user *user_address;
56 bool need_to_free_pages;
59 static uint64_t *page_slots(struct kfd_signal_page *page)
61 return page->kernel_address;
64 static struct kfd_signal_page *allocate_signal_page(struct kfd_process *p)
67 struct kfd_signal_page *page;
69 page = kzalloc(sizeof(*page), GFP_KERNEL);
73 backing_store = (void *) __get_free_pages(GFP_KERNEL,
74 get_order(KFD_SIGNAL_EVENT_LIMIT * 8));
76 goto fail_alloc_signal_store;
78 /* Initialize all events to unsignaled */
79 memset(backing_store, (uint8_t) UNSIGNALED_EVENT_SLOT,
80 KFD_SIGNAL_EVENT_LIMIT * 8);
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",
89 fail_alloc_signal_store:
94 static int allocate_event_notification_slot(struct kfd_process *p,
96 const int *restore_id)
100 if (!p->signal_page) {
101 p->signal_page = allocate_signal_page(p);
104 /* Oldest user mode expects 256 event slots */
105 p->signal_mapped_size = 256*8;
109 id = idr_alloc(&p->event_idr, ev, *restore_id, *restore_id + 1,
113 * Compatibility with old user mode: Only use signal slots
114 * user mode has mapped, may be less than
115 * KFD_SIGNAL_EVENT_LIMIT. This also allows future increase
116 * of the event limit without breaking user mode.
118 id = idr_alloc(&p->event_idr, ev, 0, p->signal_mapped_size / 8,
125 page_slots(p->signal_page)[id] = UNSIGNALED_EVENT_SLOT;
131 * Assumes that p->event_mutex or rcu_readlock is held and of course that p is
134 static struct kfd_event *lookup_event_by_id(struct kfd_process *p, uint32_t id)
136 return idr_find(&p->event_idr, id);
140 * lookup_signaled_event_by_partial_id - Lookup signaled event from partial ID
141 * @p: Pointer to struct kfd_process
143 * @bits: Number of valid bits in @id
145 * Finds the first signaled event with a matching partial ID. If no
146 * matching signaled event is found, returns NULL. In that case the
147 * caller should assume that the partial ID is invalid and do an
148 * exhaustive search of all siglaned events.
150 * If multiple events with the same partial ID signal at the same
151 * time, they will be found one interrupt at a time, not necessarily
152 * in the same order the interrupts occurred. As long as the number of
153 * interrupts is correct, all signaled events will be seen by the
156 static struct kfd_event *lookup_signaled_event_by_partial_id(
157 struct kfd_process *p, uint32_t id, uint32_t bits)
159 struct kfd_event *ev;
161 if (!p->signal_page || id >= KFD_SIGNAL_EVENT_LIMIT)
164 /* Fast path for the common case that @id is not a partial ID
165 * and we only need a single lookup.
167 if (bits > 31 || (1U << bits) >= KFD_SIGNAL_EVENT_LIMIT) {
168 if (page_slots(p->signal_page)[id] == UNSIGNALED_EVENT_SLOT)
171 return idr_find(&p->event_idr, id);
174 /* General case for partial IDs: Iterate over all matching IDs
175 * and find the first one that has signaled.
177 for (ev = NULL; id < KFD_SIGNAL_EVENT_LIMIT && !ev; id += 1U << bits) {
178 if (page_slots(p->signal_page)[id] == UNSIGNALED_EVENT_SLOT)
181 ev = idr_find(&p->event_idr, id);
187 static int create_signal_event(struct file *devkfd, struct kfd_process *p,
188 struct kfd_event *ev, const int *restore_id)
192 if (p->signal_mapped_size &&
193 p->signal_event_count == p->signal_mapped_size / 8) {
194 if (!p->signal_event_limit_reached) {
195 pr_debug("Signal event wasn't created because limit was reached\n");
196 p->signal_event_limit_reached = true;
201 ret = allocate_event_notification_slot(p, ev, restore_id);
203 pr_warn("Signal event wasn't created because out of kernel memory\n");
207 p->signal_event_count++;
209 ev->user_signal_address = &p->signal_page->user_address[ev->event_id];
210 pr_debug("Signal event number %zu created with id %d, address %p\n",
211 p->signal_event_count, ev->event_id,
212 ev->user_signal_address);
217 static int create_other_event(struct kfd_process *p, struct kfd_event *ev, const int *restore_id)
222 id = idr_alloc(&p->event_idr, ev, *restore_id, *restore_id + 1,
225 /* Cast KFD_LAST_NONSIGNAL_EVENT to uint32_t. This allows an
226 * intentional integer overflow to -1 without a compiler
227 * warning. idr_alloc treats a negative value as "maximum
230 id = idr_alloc(&p->event_idr, ev, KFD_FIRST_NONSIGNAL_EVENT_ID,
231 (uint32_t)KFD_LAST_NONSIGNAL_EVENT_ID + 1,
241 int kfd_event_init_process(struct kfd_process *p)
245 mutex_init(&p->event_mutex);
246 idr_init(&p->event_idr);
247 p->signal_page = NULL;
248 p->signal_event_count = 1;
249 /* Allocate event ID 0. It is used for a fast path to ignore bogus events
250 * that are sent by the CP without a context ID
252 id = idr_alloc(&p->event_idr, NULL, 0, 1, GFP_KERNEL);
254 idr_destroy(&p->event_idr);
255 mutex_destroy(&p->event_mutex);
261 static void destroy_event(struct kfd_process *p, struct kfd_event *ev)
263 struct kfd_event_waiter *waiter;
265 /* Wake up pending waiters. They will return failure */
266 spin_lock(&ev->lock);
267 list_for_each_entry(waiter, &ev->wq.head, wait.entry)
268 WRITE_ONCE(waiter->event, NULL);
269 wake_up_all(&ev->wq);
270 spin_unlock(&ev->lock);
272 if (ev->type == KFD_EVENT_TYPE_SIGNAL ||
273 ev->type == KFD_EVENT_TYPE_DEBUG)
274 p->signal_event_count--;
276 idr_remove(&p->event_idr, ev->event_id);
280 static void destroy_events(struct kfd_process *p)
282 struct kfd_event *ev;
285 idr_for_each_entry(&p->event_idr, ev, id)
287 destroy_event(p, ev);
288 idr_destroy(&p->event_idr);
289 mutex_destroy(&p->event_mutex);
293 * We assume that the process is being destroyed and there is no need to
294 * unmap the pages or keep bookkeeping data in order.
296 static void shutdown_signal_page(struct kfd_process *p)
298 struct kfd_signal_page *page = p->signal_page;
301 if (page->need_to_free_pages)
302 free_pages((unsigned long)page->kernel_address,
303 get_order(KFD_SIGNAL_EVENT_LIMIT * 8));
308 void kfd_event_free_process(struct kfd_process *p)
311 shutdown_signal_page(p);
314 static bool event_can_be_gpu_signaled(const struct kfd_event *ev)
316 return ev->type == KFD_EVENT_TYPE_SIGNAL ||
317 ev->type == KFD_EVENT_TYPE_DEBUG;
320 static bool event_can_be_cpu_signaled(const struct kfd_event *ev)
322 return ev->type == KFD_EVENT_TYPE_SIGNAL;
325 static int kfd_event_page_set(struct kfd_process *p, void *kernel_address,
326 uint64_t size, uint64_t user_handle)
328 struct kfd_signal_page *page;
333 page = kzalloc(sizeof(*page), GFP_KERNEL);
337 /* Initialize all events to unsignaled */
338 memset(kernel_address, (uint8_t) UNSIGNALED_EVENT_SLOT,
339 KFD_SIGNAL_EVENT_LIMIT * 8);
341 page->kernel_address = kernel_address;
343 p->signal_page = page;
344 p->signal_mapped_size = size;
345 p->signal_handle = user_handle;
349 int kfd_kmap_event_page(struct kfd_process *p, uint64_t event_page_offset)
352 struct kfd_process_device *pdd;
353 void *mem, *kern_addr;
357 if (p->signal_page) {
358 pr_err("Event page is already set\n");
362 pdd = kfd_process_device_data_by_id(p, GET_GPU_ID(event_page_offset));
364 pr_err("Getting device by id failed in %s\n", __func__);
369 pdd = kfd_bind_process_to_device(kfd, p);
373 mem = kfd_process_device_translate_handle(pdd,
374 GET_IDR_HANDLE(event_page_offset));
376 pr_err("Can't find BO, offset is 0x%llx\n", event_page_offset);
380 err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(mem, &kern_addr, &size);
382 pr_err("Failed to map event page to kernel\n");
386 err = kfd_event_page_set(p, kern_addr, size, event_page_offset);
388 pr_err("Failed to set event page\n");
389 amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem);
395 int kfd_event_create(struct file *devkfd, struct kfd_process *p,
396 uint32_t event_type, bool auto_reset, uint32_t node_id,
397 uint32_t *event_id, uint32_t *event_trigger_data,
398 uint64_t *event_page_offset, uint32_t *event_slot_index)
401 struct kfd_event *ev = kzalloc(sizeof(*ev), GFP_KERNEL);
406 ev->type = event_type;
407 ev->auto_reset = auto_reset;
408 ev->signaled = false;
410 spin_lock_init(&ev->lock);
411 init_waitqueue_head(&ev->wq);
413 *event_page_offset = 0;
415 mutex_lock(&p->event_mutex);
417 switch (event_type) {
418 case KFD_EVENT_TYPE_SIGNAL:
419 case KFD_EVENT_TYPE_DEBUG:
420 ret = create_signal_event(devkfd, p, ev, NULL);
422 *event_page_offset = KFD_MMAP_TYPE_EVENTS;
423 *event_slot_index = ev->event_id;
427 ret = create_other_event(p, ev, NULL);
432 *event_id = ev->event_id;
433 *event_trigger_data = ev->event_id;
438 mutex_unlock(&p->event_mutex);
443 int kfd_criu_restore_event(struct file *devkfd,
444 struct kfd_process *p,
445 uint8_t __user *user_priv_ptr,
446 uint64_t *priv_data_offset,
447 uint64_t max_priv_data_size)
449 struct kfd_criu_event_priv_data *ev_priv;
450 struct kfd_event *ev = NULL;
453 ev_priv = kmalloc(sizeof(*ev_priv), GFP_KERNEL);
457 ev = kzalloc(sizeof(*ev), GFP_KERNEL);
463 if (*priv_data_offset + sizeof(*ev_priv) > max_priv_data_size) {
468 ret = copy_from_user(ev_priv, user_priv_ptr + *priv_data_offset, sizeof(*ev_priv));
473 *priv_data_offset += sizeof(*ev_priv);
475 if (ev_priv->user_handle) {
476 ret = kfd_kmap_event_page(p, ev_priv->user_handle);
481 ev->type = ev_priv->type;
482 ev->auto_reset = ev_priv->auto_reset;
483 ev->signaled = ev_priv->signaled;
485 spin_lock_init(&ev->lock);
486 init_waitqueue_head(&ev->wq);
488 mutex_lock(&p->event_mutex);
490 case KFD_EVENT_TYPE_SIGNAL:
491 case KFD_EVENT_TYPE_DEBUG:
492 ret = create_signal_event(devkfd, p, ev, &ev_priv->event_id);
494 case KFD_EVENT_TYPE_MEMORY:
495 memcpy(&ev->memory_exception_data,
496 &ev_priv->memory_exception_data,
497 sizeof(struct kfd_hsa_memory_exception_data));
499 ret = create_other_event(p, ev, &ev_priv->event_id);
501 case KFD_EVENT_TYPE_HW_EXCEPTION:
502 memcpy(&ev->hw_exception_data,
503 &ev_priv->hw_exception_data,
504 sizeof(struct kfd_hsa_hw_exception_data));
506 ret = create_other_event(p, ev, &ev_priv->event_id);
516 mutex_unlock(&p->event_mutex);
521 int kfd_criu_checkpoint_events(struct kfd_process *p,
522 uint8_t __user *user_priv_data,
523 uint64_t *priv_data_offset)
525 struct kfd_criu_event_priv_data *ev_privs;
528 struct kfd_event *ev;
531 uint32_t num_events = kfd_get_num_events(p);
536 ev_privs = kvzalloc(num_events * sizeof(*ev_privs), GFP_KERNEL);
541 idr_for_each_entry(&p->event_idr, ev, ev_id) {
542 struct kfd_criu_event_priv_data *ev_priv;
545 * Currently, all events have same size of private_data, but the current ioctl's
546 * and CRIU plugin supports private_data of variable sizes
548 ev_priv = &ev_privs[i];
550 ev_priv->object_type = KFD_CRIU_OBJECT_TYPE_EVENT;
552 /* We store the user_handle with the first event */
553 if (i == 0 && p->signal_page)
554 ev_priv->user_handle = p->signal_handle;
556 ev_priv->event_id = ev->event_id;
557 ev_priv->auto_reset = ev->auto_reset;
558 ev_priv->type = ev->type;
559 ev_priv->signaled = ev->signaled;
561 if (ev_priv->type == KFD_EVENT_TYPE_MEMORY)
562 memcpy(&ev_priv->memory_exception_data,
563 &ev->memory_exception_data,
564 sizeof(struct kfd_hsa_memory_exception_data));
565 else if (ev_priv->type == KFD_EVENT_TYPE_HW_EXCEPTION)
566 memcpy(&ev_priv->hw_exception_data,
567 &ev->hw_exception_data,
568 sizeof(struct kfd_hsa_hw_exception_data));
570 pr_debug("Checkpointed event[%d] id = 0x%08x auto_reset = %x type = %x signaled = %x\n",
579 ret = copy_to_user(user_priv_data + *priv_data_offset,
580 ev_privs, num_events * sizeof(*ev_privs));
582 pr_err("Failed to copy events priv to user\n");
586 *priv_data_offset += num_events * sizeof(*ev_privs);
592 int kfd_get_num_events(struct kfd_process *p)
594 struct kfd_event *ev;
598 idr_for_each_entry(&p->event_idr, ev, id)
604 /* Assumes that p is current. */
605 int kfd_event_destroy(struct kfd_process *p, uint32_t event_id)
607 struct kfd_event *ev;
610 mutex_lock(&p->event_mutex);
612 ev = lookup_event_by_id(p, event_id);
615 destroy_event(p, ev);
619 mutex_unlock(&p->event_mutex);
623 static void set_event(struct kfd_event *ev)
625 struct kfd_event_waiter *waiter;
627 /* Auto reset if the list is non-empty and we're waking
628 * someone. waitqueue_active is safe here because we're
629 * protected by the ev->lock, which is also held when
630 * updating the wait queues in kfd_wait_on_events.
632 ev->signaled = !ev->auto_reset || !waitqueue_active(&ev->wq);
634 list_for_each_entry(waiter, &ev->wq.head, wait.entry)
635 WRITE_ONCE(waiter->activated, true);
637 wake_up_all(&ev->wq);
640 /* Assumes that p is current. */
641 int kfd_set_event(struct kfd_process *p, uint32_t event_id)
644 struct kfd_event *ev;
648 ev = lookup_event_by_id(p, event_id);
653 spin_lock(&ev->lock);
655 if (event_can_be_cpu_signaled(ev))
660 spin_unlock(&ev->lock);
666 static void reset_event(struct kfd_event *ev)
668 ev->signaled = false;
671 /* Assumes that p is current. */
672 int kfd_reset_event(struct kfd_process *p, uint32_t event_id)
675 struct kfd_event *ev;
679 ev = lookup_event_by_id(p, event_id);
684 spin_lock(&ev->lock);
686 if (event_can_be_cpu_signaled(ev))
691 spin_unlock(&ev->lock);
698 static void acknowledge_signal(struct kfd_process *p, struct kfd_event *ev)
700 WRITE_ONCE(page_slots(p->signal_page)[ev->event_id], UNSIGNALED_EVENT_SLOT);
703 static void set_event_from_interrupt(struct kfd_process *p,
704 struct kfd_event *ev)
706 if (ev && event_can_be_gpu_signaled(ev)) {
707 acknowledge_signal(p, ev);
708 spin_lock(&ev->lock);
710 spin_unlock(&ev->lock);
714 void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
715 uint32_t valid_id_bits)
717 struct kfd_event *ev = NULL;
720 * Because we are called from arbitrary context (workqueue) as opposed
721 * to process context, kfd_process could attempt to exit while we are
722 * running so the lookup function increments the process ref count.
724 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
727 return; /* Presumably process exited. */
732 ev = lookup_signaled_event_by_partial_id(p, partial_id,
735 set_event_from_interrupt(p, ev);
736 } else if (p->signal_page) {
738 * Partial ID lookup failed. Assume that the event ID
739 * in the interrupt payload was invalid and do an
740 * exhaustive search of signaled events.
742 uint64_t *slots = page_slots(p->signal_page);
746 pr_debug_ratelimited("Partial ID invalid: %u (%u valid bits)\n",
747 partial_id, valid_id_bits);
749 if (p->signal_event_count < KFD_SIGNAL_EVENT_LIMIT / 64) {
750 /* With relatively few events, it's faster to
751 * iterate over the event IDR
753 idr_for_each_entry(&p->event_idr, ev, id) {
754 if (id >= KFD_SIGNAL_EVENT_LIMIT)
757 if (READ_ONCE(slots[id]) != UNSIGNALED_EVENT_SLOT)
758 set_event_from_interrupt(p, ev);
761 /* With relatively many events, it's faster to
762 * iterate over the signal slots and lookup
763 * only signaled events from the IDR.
765 for (id = 1; id < KFD_SIGNAL_EVENT_LIMIT; id++)
766 if (READ_ONCE(slots[id]) != UNSIGNALED_EVENT_SLOT) {
767 ev = lookup_event_by_id(p, id);
768 set_event_from_interrupt(p, ev);
774 kfd_unref_process(p);
777 static struct kfd_event_waiter *alloc_event_waiters(uint32_t num_events)
779 struct kfd_event_waiter *event_waiters;
782 event_waiters = kmalloc_array(num_events,
783 sizeof(struct kfd_event_waiter),
788 for (i = 0; (event_waiters) && (i < num_events) ; i++) {
789 init_wait(&event_waiters[i].wait);
790 event_waiters[i].activated = false;
793 return event_waiters;
796 static int init_event_waiter(struct kfd_process *p,
797 struct kfd_event_waiter *waiter,
800 struct kfd_event *ev = lookup_event_by_id(p, event_id);
805 spin_lock(&ev->lock);
807 waiter->activated = ev->signaled;
808 ev->signaled = ev->signaled && !ev->auto_reset;
809 if (!waiter->activated)
810 add_wait_queue(&ev->wq, &waiter->wait);
811 spin_unlock(&ev->lock);
816 /* test_event_condition - Test condition of events being waited for
817 * @all: Return completion only if all events have signaled
818 * @num_events: Number of events to wait for
819 * @event_waiters: Array of event waiters, one per event
821 * Returns KFD_IOC_WAIT_RESULT_COMPLETE if all (or one) event(s) have
822 * signaled. Returns KFD_IOC_WAIT_RESULT_TIMEOUT if no (or not all)
823 * events have signaled. Returns KFD_IOC_WAIT_RESULT_FAIL if any of
824 * the events have been destroyed.
826 static uint32_t test_event_condition(bool all, uint32_t num_events,
827 struct kfd_event_waiter *event_waiters)
830 uint32_t activated_count = 0;
832 for (i = 0; i < num_events; i++) {
833 if (!READ_ONCE(event_waiters[i].event))
834 return KFD_IOC_WAIT_RESULT_FAIL;
836 if (READ_ONCE(event_waiters[i].activated)) {
838 return KFD_IOC_WAIT_RESULT_COMPLETE;
844 return activated_count == num_events ?
845 KFD_IOC_WAIT_RESULT_COMPLETE : KFD_IOC_WAIT_RESULT_TIMEOUT;
849 * Copy event specific data, if defined.
850 * Currently only memory exception events have additional data to copy to user
852 static int copy_signaled_event_data(uint32_t num_events,
853 struct kfd_event_waiter *event_waiters,
854 struct kfd_event_data __user *data)
856 struct kfd_hsa_memory_exception_data *src;
857 struct kfd_hsa_memory_exception_data __user *dst;
858 struct kfd_event_waiter *waiter;
859 struct kfd_event *event;
862 for (i = 0; i < num_events; i++) {
863 waiter = &event_waiters[i];
864 event = waiter->event;
866 return -EINVAL; /* event was destroyed */
867 if (waiter->activated && event->type == KFD_EVENT_TYPE_MEMORY) {
868 dst = &data[i].memory_exception_data;
869 src = &event->memory_exception_data;
870 if (copy_to_user(dst, src,
871 sizeof(struct kfd_hsa_memory_exception_data)))
879 static long user_timeout_to_jiffies(uint32_t user_timeout_ms)
881 if (user_timeout_ms == KFD_EVENT_TIMEOUT_IMMEDIATE)
884 if (user_timeout_ms == KFD_EVENT_TIMEOUT_INFINITE)
885 return MAX_SCHEDULE_TIMEOUT;
888 * msecs_to_jiffies interprets all values above 2^31-1 as infinite,
889 * but we consider them finite.
890 * This hack is wrong, but nobody is likely to notice.
892 user_timeout_ms = min_t(uint32_t, user_timeout_ms, 0x7FFFFFFF);
894 return msecs_to_jiffies(user_timeout_ms) + 1;
897 static void free_waiters(uint32_t num_events, struct kfd_event_waiter *waiters,
898 bool undo_auto_reset)
902 for (i = 0; i < num_events; i++)
903 if (waiters[i].event) {
904 spin_lock(&waiters[i].event->lock);
905 remove_wait_queue(&waiters[i].event->wq,
907 if (undo_auto_reset && waiters[i].activated &&
908 waiters[i].event && waiters[i].event->auto_reset)
909 set_event(waiters[i].event);
910 spin_unlock(&waiters[i].event->lock);
916 int kfd_wait_on_events(struct kfd_process *p,
917 uint32_t num_events, void __user *data,
918 bool all, uint32_t *user_timeout_ms,
919 uint32_t *wait_result)
921 struct kfd_event_data __user *events =
922 (struct kfd_event_data __user *) data;
926 struct kfd_event_waiter *event_waiters = NULL;
927 long timeout = user_timeout_to_jiffies(*user_timeout_ms);
929 event_waiters = alloc_event_waiters(num_events);
930 if (!event_waiters) {
935 /* Use p->event_mutex here to protect against concurrent creation and
936 * destruction of events while we initialize event_waiters.
938 mutex_lock(&p->event_mutex);
940 for (i = 0; i < num_events; i++) {
941 struct kfd_event_data event_data;
943 if (copy_from_user(&event_data, &events[i],
944 sizeof(struct kfd_event_data))) {
949 ret = init_event_waiter(p, &event_waiters[i],
950 event_data.event_id);
955 /* Check condition once. */
956 *wait_result = test_event_condition(all, num_events, event_waiters);
957 if (*wait_result == KFD_IOC_WAIT_RESULT_COMPLETE) {
958 ret = copy_signaled_event_data(num_events,
959 event_waiters, events);
961 } else if (WARN_ON(*wait_result == KFD_IOC_WAIT_RESULT_FAIL)) {
962 /* This should not happen. Events shouldn't be
963 * destroyed while we're holding the event_mutex
968 mutex_unlock(&p->event_mutex);
971 if (fatal_signal_pending(current)) {
976 if (signal_pending(current)) {
978 if (*user_timeout_ms != KFD_EVENT_TIMEOUT_IMMEDIATE &&
979 *user_timeout_ms != KFD_EVENT_TIMEOUT_INFINITE)
980 *user_timeout_ms = jiffies_to_msecs(
985 /* Set task state to interruptible sleep before
986 * checking wake-up conditions. A concurrent wake-up
987 * will put the task back into runnable state. In that
988 * case schedule_timeout will not put the task to
989 * sleep and we'll get a chance to re-check the
990 * updated conditions almost immediately. Otherwise,
991 * this race condition would lead to a soft hang or a
994 set_current_state(TASK_INTERRUPTIBLE);
996 *wait_result = test_event_condition(all, num_events,
998 if (*wait_result != KFD_IOC_WAIT_RESULT_TIMEOUT)
1004 timeout = schedule_timeout(timeout);
1006 __set_current_state(TASK_RUNNING);
1008 mutex_lock(&p->event_mutex);
1009 /* copy_signaled_event_data may sleep. So this has to happen
1010 * after the task state is set back to RUNNING.
1012 * The event may also have been destroyed after signaling. So
1013 * copy_signaled_event_data also must confirm that the event
1014 * still exists. Therefore this must be under the p->event_mutex
1015 * which is also held when events are destroyed.
1017 if (!ret && *wait_result == KFD_IOC_WAIT_RESULT_COMPLETE)
1018 ret = copy_signaled_event_data(num_events,
1019 event_waiters, events);
1022 free_waiters(num_events, event_waiters, ret == -ERESTARTSYS);
1023 mutex_unlock(&p->event_mutex);
1026 *wait_result = KFD_IOC_WAIT_RESULT_FAIL;
1027 else if (*wait_result == KFD_IOC_WAIT_RESULT_FAIL)
1033 int kfd_event_mmap(struct kfd_process *p, struct vm_area_struct *vma)
1036 struct kfd_signal_page *page;
1039 /* check required size doesn't exceed the allocated size */
1040 if (get_order(KFD_SIGNAL_EVENT_LIMIT * 8) <
1041 get_order(vma->vm_end - vma->vm_start)) {
1042 pr_err("Event page mmap requested illegal size\n");
1046 page = p->signal_page;
1048 /* Probably KFD bug, but mmap is user-accessible. */
1049 pr_debug("Signal page could not be found\n");
1053 pfn = __pa(page->kernel_address);
1056 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND | VM_NORESERVE
1057 | VM_DONTDUMP | VM_PFNMAP;
1059 pr_debug("Mapping signal page\n");
1060 pr_debug(" start user address == 0x%08lx\n", vma->vm_start);
1061 pr_debug(" end user address == 0x%08lx\n", vma->vm_end);
1062 pr_debug(" pfn == 0x%016lX\n", pfn);
1063 pr_debug(" vm_flags == 0x%08lX\n", vma->vm_flags);
1064 pr_debug(" size == 0x%08lX\n",
1065 vma->vm_end - vma->vm_start);
1067 page->user_address = (uint64_t __user *)vma->vm_start;
1069 /* mapping the page to user process */
1070 ret = remap_pfn_range(vma, vma->vm_start, pfn,
1071 vma->vm_end - vma->vm_start, vma->vm_page_prot);
1073 p->signal_mapped_size = vma->vm_end - vma->vm_start;
1079 * Assumes that p is not going away.
1081 static void lookup_events_by_type_and_signal(struct kfd_process *p,
1082 int type, void *event_data)
1084 struct kfd_hsa_memory_exception_data *ev_data;
1085 struct kfd_event *ev;
1087 bool send_signal = true;
1089 ev_data = (struct kfd_hsa_memory_exception_data *) event_data;
1093 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1094 idr_for_each_entry_continue(&p->event_idr, ev, id)
1095 if (ev->type == type) {
1096 send_signal = false;
1098 "Event found: id %X type %d",
1099 ev->event_id, ev->type);
1100 spin_lock(&ev->lock);
1102 if (ev->type == KFD_EVENT_TYPE_MEMORY && ev_data)
1103 ev->memory_exception_data = *ev_data;
1104 spin_unlock(&ev->lock);
1107 if (type == KFD_EVENT_TYPE_MEMORY) {
1108 dev_warn(kfd_device,
1109 "Sending SIGSEGV to process %d (pasid 0x%x)",
1110 p->lead_thread->pid, p->pasid);
1111 send_sig(SIGSEGV, p->lead_thread, 0);
1114 /* Send SIGTERM no event of type "type" has been found*/
1117 dev_warn(kfd_device,
1118 "Sending SIGTERM to process %d (pasid 0x%x)",
1119 p->lead_thread->pid, p->pasid);
1120 send_sig(SIGTERM, p->lead_thread, 0);
1123 "Process %d (pasid 0x%x) got unhandled exception",
1124 p->lead_thread->pid, p->pasid);
1131 #ifdef KFD_SUPPORT_IOMMU_V2
1132 void kfd_signal_iommu_event(struct kfd_dev *dev, u32 pasid,
1133 unsigned long address, bool is_write_requested,
1134 bool is_execute_requested)
1136 struct kfd_hsa_memory_exception_data memory_exception_data;
1137 struct vm_area_struct *vma;
1141 * Because we are called from arbitrary context (workqueue) as opposed
1142 * to process context, kfd_process could attempt to exit while we are
1143 * running so the lookup function increments the process ref count.
1145 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
1146 struct mm_struct *mm;
1149 return; /* Presumably process exited. */
1151 /* Take a safe reference to the mm_struct, which may otherwise
1152 * disappear even while the kfd_process is still referenced.
1154 mm = get_task_mm(p->lead_thread);
1156 kfd_unref_process(p);
1157 return; /* Process is exiting */
1160 user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
1161 if (unlikely(user_gpu_id == -EINVAL)) {
1162 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
1165 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1168 vma = find_vma(mm, address);
1170 memory_exception_data.gpu_id = user_gpu_id;
1171 memory_exception_data.va = address;
1172 /* Set failure reason */
1173 memory_exception_data.failure.NotPresent = 1;
1174 memory_exception_data.failure.NoExecute = 0;
1175 memory_exception_data.failure.ReadOnly = 0;
1176 if (vma && address >= vma->vm_start) {
1177 memory_exception_data.failure.NotPresent = 0;
1179 if (is_write_requested && !(vma->vm_flags & VM_WRITE))
1180 memory_exception_data.failure.ReadOnly = 1;
1182 memory_exception_data.failure.ReadOnly = 0;
1184 if (is_execute_requested && !(vma->vm_flags & VM_EXEC))
1185 memory_exception_data.failure.NoExecute = 1;
1187 memory_exception_data.failure.NoExecute = 0;
1190 mmap_read_unlock(mm);
1193 pr_debug("notpresent %d, noexecute %d, readonly %d\n",
1194 memory_exception_data.failure.NotPresent,
1195 memory_exception_data.failure.NoExecute,
1196 memory_exception_data.failure.ReadOnly);
1198 /* Workaround on Raven to not kill the process when memory is freed
1199 * before IOMMU is able to finish processing all the excessive PPRs
1202 if (KFD_GC_VERSION(dev) != IP_VERSION(9, 1, 0) &&
1203 KFD_GC_VERSION(dev) != IP_VERSION(9, 2, 2) &&
1204 KFD_GC_VERSION(dev) != IP_VERSION(9, 3, 0))
1205 lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_MEMORY,
1206 &memory_exception_data);
1208 kfd_unref_process(p);
1210 #endif /* KFD_SUPPORT_IOMMU_V2 */
1212 void kfd_signal_hw_exception_event(u32 pasid)
1215 * Because we are called from arbitrary context (workqueue) as opposed
1216 * to process context, kfd_process could attempt to exit while we are
1217 * running so the lookup function increments the process ref count.
1219 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
1222 return; /* Presumably process exited. */
1224 lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_HW_EXCEPTION, NULL);
1225 kfd_unref_process(p);
1228 void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid,
1229 struct kfd_vm_fault_info *info)
1231 struct kfd_event *ev;
1233 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
1234 struct kfd_hsa_memory_exception_data memory_exception_data;
1238 return; /* Presumably process exited. */
1240 user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
1241 if (unlikely(user_gpu_id == -EINVAL)) {
1242 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
1246 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1247 memory_exception_data.gpu_id = user_gpu_id;
1248 memory_exception_data.failure.imprecise = true;
1249 /* Set failure reason */
1251 memory_exception_data.va = (info->page_addr) << PAGE_SHIFT;
1252 memory_exception_data.failure.NotPresent =
1253 info->prot_valid ? 1 : 0;
1254 memory_exception_data.failure.NoExecute =
1255 info->prot_exec ? 1 : 0;
1256 memory_exception_data.failure.ReadOnly =
1257 info->prot_write ? 1 : 0;
1258 memory_exception_data.failure.imprecise = 0;
1263 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1264 idr_for_each_entry_continue(&p->event_idr, ev, id)
1265 if (ev->type == KFD_EVENT_TYPE_MEMORY) {
1266 spin_lock(&ev->lock);
1267 ev->memory_exception_data = memory_exception_data;
1269 spin_unlock(&ev->lock);
1273 kfd_unref_process(p);
1276 void kfd_signal_reset_event(struct kfd_dev *dev)
1278 struct kfd_hsa_hw_exception_data hw_exception_data;
1279 struct kfd_hsa_memory_exception_data memory_exception_data;
1280 struct kfd_process *p;
1281 struct kfd_event *ev;
1284 int reset_cause = atomic_read(&dev->sram_ecc_flag) ?
1285 KFD_HW_EXCEPTION_ECC :
1286 KFD_HW_EXCEPTION_GPU_HANG;
1288 /* Whole gpu reset caused by GPU hang and memory is lost */
1289 memset(&hw_exception_data, 0, sizeof(hw_exception_data));
1290 hw_exception_data.memory_lost = 1;
1291 hw_exception_data.reset_cause = reset_cause;
1293 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1294 memory_exception_data.ErrorType = KFD_MEM_ERR_SRAM_ECC;
1295 memory_exception_data.failure.imprecise = true;
1297 idx = srcu_read_lock(&kfd_processes_srcu);
1298 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1299 int user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
1301 if (unlikely(user_gpu_id == -EINVAL)) {
1302 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
1308 id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1309 idr_for_each_entry_continue(&p->event_idr, ev, id) {
1310 if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
1311 spin_lock(&ev->lock);
1312 ev->hw_exception_data = hw_exception_data;
1313 ev->hw_exception_data.gpu_id = user_gpu_id;
1315 spin_unlock(&ev->lock);
1317 if (ev->type == KFD_EVENT_TYPE_MEMORY &&
1318 reset_cause == KFD_HW_EXCEPTION_ECC) {
1319 spin_lock(&ev->lock);
1320 ev->memory_exception_data = memory_exception_data;
1321 ev->memory_exception_data.gpu_id = user_gpu_id;
1323 spin_unlock(&ev->lock);
1329 srcu_read_unlock(&kfd_processes_srcu, idx);
1332 void kfd_signal_poison_consumed_event(struct kfd_dev *dev, u32 pasid)
1334 struct kfd_process *p = kfd_lookup_process_by_pasid(pasid);
1335 struct kfd_hsa_memory_exception_data memory_exception_data;
1336 struct kfd_hsa_hw_exception_data hw_exception_data;
1337 struct kfd_event *ev;
1338 uint32_t id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1342 return; /* Presumably process exited. */
1344 user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
1345 if (unlikely(user_gpu_id == -EINVAL)) {
1346 WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
1350 memset(&hw_exception_data, 0, sizeof(hw_exception_data));
1351 hw_exception_data.gpu_id = user_gpu_id;
1352 hw_exception_data.memory_lost = 1;
1353 hw_exception_data.reset_cause = KFD_HW_EXCEPTION_ECC;
1355 memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1356 memory_exception_data.ErrorType = KFD_MEM_ERR_POISON_CONSUMED;
1357 memory_exception_data.gpu_id = user_gpu_id;
1358 memory_exception_data.failure.imprecise = true;
1362 idr_for_each_entry_continue(&p->event_idr, ev, id) {
1363 if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
1364 spin_lock(&ev->lock);
1365 ev->hw_exception_data = hw_exception_data;
1367 spin_unlock(&ev->lock);
1370 if (ev->type == KFD_EVENT_TYPE_MEMORY) {
1371 spin_lock(&ev->lock);
1372 ev->memory_exception_data = memory_exception_data;
1374 spin_unlock(&ev->lock);
1380 /* user application will handle SIGBUS signal */
1381 send_sig(SIGBUS, p->lead_thread, 0);
1383 kfd_unref_process(p);