Commit | Line | Data |
---|---|---|
f3a39818 AL |
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> | |
3f07c014 | 26 | #include <linux/sched/signal.h> |
9b56bb11 | 27 | #include <linux/sched/mm.h> |
f3a39818 | 28 | #include <linux/uaccess.h> |
f3a39818 AL |
29 | #include <linux/mman.h> |
30 | #include <linux/memory.h> | |
31 | #include "kfd_priv.h" | |
32 | #include "kfd_events.h" | |
64d1c3a4 | 33 | #include "kfd_iommu.h" |
59d3e8be | 34 | #include <linux/device.h> |
f3a39818 AL |
35 | |
36 | /* | |
74e40716 | 37 | * Wrapper around wait_queue_entry_t |
f3a39818 AL |
38 | */ |
39 | struct kfd_event_waiter { | |
74e40716 FK |
40 | wait_queue_entry_t wait; |
41 | struct kfd_event *event; /* Event to wait for */ | |
42 | bool activated; /* Becomes true when event is signaled */ | |
f3a39818 AL |
43 | }; |
44 | ||
45 | /* | |
f3a39818 | 46 | * Each signal event needs a 64-bit signal slot where the signaler will write |
482f0777 | 47 | * a 1 before sending an interrupt. (This is needed because some interrupts |
f3a39818 | 48 | * do not contain enough spare data bits to identify an event.) |
482f0777 FK |
49 | * We get whole pages and map them to the process VA. |
50 | * Individual signal events use their event_id as slot index. | |
f3a39818 | 51 | */ |
50cb7dd9 | 52 | struct kfd_signal_page { |
f3a39818 AL |
53 | uint64_t *kernel_address; |
54 | uint64_t __user *user_address; | |
0fc8011f | 55 | bool need_to_free_pages; |
f3a39818 AL |
56 | }; |
57 | ||
f3a39818 | 58 | |
50cb7dd9 | 59 | static uint64_t *page_slots(struct kfd_signal_page *page) |
f3a39818 AL |
60 | { |
61 | return page->kernel_address; | |
62 | } | |
63 | ||
50cb7dd9 | 64 | static struct kfd_signal_page *allocate_signal_page(struct kfd_process *p) |
f3a39818 AL |
65 | { |
66 | void *backing_store; | |
50cb7dd9 | 67 | struct kfd_signal_page *page; |
f3a39818 | 68 | |
50cb7dd9 | 69 | page = kzalloc(sizeof(*page), GFP_KERNEL); |
f3a39818 | 70 | if (!page) |
50cb7dd9 | 71 | return NULL; |
f3a39818 | 72 | |
50cb7dd9 | 73 | backing_store = (void *) __get_free_pages(GFP_KERNEL, |
f3a39818 AL |
74 | get_order(KFD_SIGNAL_EVENT_LIMIT * 8)); |
75 | if (!backing_store) | |
76 | goto fail_alloc_signal_store; | |
77 | ||
50cb7dd9 | 78 | /* Initialize all events to unsignaled */ |
f3a39818 | 79 | memset(backing_store, (uint8_t) UNSIGNALED_EVENT_SLOT, |
50cb7dd9 | 80 | KFD_SIGNAL_EVENT_LIMIT * 8); |
f3a39818 AL |
81 | |
82 | page->kernel_address = backing_store; | |
0fc8011f | 83 | page->need_to_free_pages = true; |
79775b62 | 84 | pr_debug("Allocated new event signal page at %p, for process %p\n", |
f3a39818 | 85 | page, p); |
f3a39818 | 86 | |
50cb7dd9 | 87 | return page; |
f3a39818 AL |
88 | |
89 | fail_alloc_signal_store: | |
90 | kfree(page); | |
50cb7dd9 | 91 | return NULL; |
f3a39818 AL |
92 | } |
93 | ||
482f0777 FK |
94 | static int allocate_event_notification_slot(struct kfd_process *p, |
95 | struct kfd_event *ev) | |
f3a39818 | 96 | { |
482f0777 FK |
97 | int id; |
98 | ||
50cb7dd9 FK |
99 | if (!p->signal_page) { |
100 | p->signal_page = allocate_signal_page(p); | |
101 | if (!p->signal_page) | |
482f0777 | 102 | return -ENOMEM; |
b9a5d0a5 FK |
103 | /* Oldest user mode expects 256 event slots */ |
104 | p->signal_mapped_size = 256*8; | |
f3a39818 AL |
105 | } |
106 | ||
b9a5d0a5 FK |
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, | |
482f0777 FK |
114 | GFP_KERNEL); |
115 | if (id < 0) | |
116 | return id; | |
f3a39818 | 117 | |
482f0777 FK |
118 | ev->event_id = id; |
119 | page_slots(p->signal_page)[id] = UNSIGNALED_EVENT_SLOT; | |
f3a39818 | 120 | |
482f0777 | 121 | return 0; |
f3a39818 AL |
122 | } |
123 | ||
f3a39818 AL |
124 | /* |
125 | * Assumes that p->event_mutex is held and of course that p is not going | |
126 | * away (current or locked). | |
127 | */ | |
128 | static struct kfd_event *lookup_event_by_id(struct kfd_process *p, uint32_t id) | |
129 | { | |
482f0777 | 130 | return idr_find(&p->event_idr, id); |
f3a39818 AL |
131 | } |
132 | ||
3f04f961 FK |
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 | */ | |
150 | static 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 | ||
f3a39818 AL |
181 | static int create_signal_event(struct file *devkfd, |
182 | struct kfd_process *p, | |
183 | struct kfd_event *ev) | |
184 | { | |
482f0777 FK |
185 | int ret; |
186 | ||
b9a5d0a5 FK |
187 | if (p->signal_mapped_size && |
188 | p->signal_event_count == p->signal_mapped_size / 8) { | |
c986169f | 189 | if (!p->signal_event_limit_reached) { |
8f2e0c03 | 190 | pr_debug("Signal event wasn't created because limit was reached\n"); |
c986169f FK |
191 | p->signal_event_limit_reached = true; |
192 | } | |
482f0777 | 193 | return -ENOSPC; |
f3a39818 AL |
194 | } |
195 | ||
482f0777 FK |
196 | ret = allocate_event_notification_slot(p, ev); |
197 | if (ret) { | |
79775b62 | 198 | pr_warn("Signal event wasn't created because out of kernel memory\n"); |
482f0777 | 199 | return ret; |
f3a39818 AL |
200 | } |
201 | ||
202 | p->signal_event_count++; | |
203 | ||
482f0777 | 204 | ev->user_signal_address = &p->signal_page->user_address[ev->event_id]; |
79775b62 | 205 | pr_debug("Signal event number %zu created with id %d, address %p\n", |
6235e15e OG |
206 | p->signal_event_count, ev->event_id, |
207 | ev->user_signal_address); | |
208 | ||
f3a39818 AL |
209 | return 0; |
210 | } | |
211 | ||
f3a39818 AL |
212 | static int create_other_event(struct kfd_process *p, struct kfd_event *ev) |
213 | { | |
482f0777 FK |
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; | |
f3a39818 AL |
226 | |
227 | return 0; | |
228 | } | |
229 | ||
230 | void kfd_event_init_process(struct kfd_process *p) | |
231 | { | |
232 | mutex_init(&p->event_mutex); | |
482f0777 | 233 | idr_init(&p->event_idr); |
50cb7dd9 | 234 | p->signal_page = NULL; |
f3a39818 AL |
235 | p->signal_event_count = 0; |
236 | } | |
237 | ||
238 | static void destroy_event(struct kfd_process *p, struct kfd_event *ev) | |
239 | { | |
74e40716 | 240 | struct kfd_event_waiter *waiter; |
fe528c13 | 241 | |
74e40716 FK |
242 | /* Wake up pending waiters. They will return failure */ |
243 | list_for_each_entry(waiter, &ev->wq.head, wait.entry) | |
fe528c13 | 244 | waiter->event = NULL; |
74e40716 | 245 | wake_up_all(&ev->wq); |
fe528c13 | 246 | |
482f0777 FK |
247 | if (ev->type == KFD_EVENT_TYPE_SIGNAL || |
248 | ev->type == KFD_EVENT_TYPE_DEBUG) | |
f3a39818 | 249 | p->signal_event_count--; |
f3a39818 | 250 | |
482f0777 | 251 | idr_remove(&p->event_idr, ev->event_id); |
f3a39818 AL |
252 | kfree(ev); |
253 | } | |
254 | ||
255 | static void destroy_events(struct kfd_process *p) | |
256 | { | |
257 | struct kfd_event *ev; | |
482f0777 | 258 | uint32_t id; |
f3a39818 | 259 | |
482f0777 | 260 | idr_for_each_entry(&p->event_idr, ev, id) |
f3a39818 | 261 | destroy_event(p, ev); |
482f0777 | 262 | idr_destroy(&p->event_idr); |
f3a39818 AL |
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 | */ | |
50cb7dd9 | 269 | static void shutdown_signal_page(struct kfd_process *p) |
f3a39818 | 270 | { |
50cb7dd9 | 271 | struct kfd_signal_page *page = p->signal_page; |
f3a39818 | 272 | |
50cb7dd9 | 273 | if (page) { |
0fc8011f FK |
274 | if (page->need_to_free_pages) |
275 | free_pages((unsigned long)page->kernel_address, | |
276 | get_order(KFD_SIGNAL_EVENT_LIMIT * 8)); | |
f3a39818 AL |
277 | kfree(page); |
278 | } | |
279 | } | |
280 | ||
281 | void kfd_event_free_process(struct kfd_process *p) | |
282 | { | |
283 | destroy_events(p); | |
50cb7dd9 | 284 | shutdown_signal_page(p); |
f3a39818 AL |
285 | } |
286 | ||
287 | static 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 | ||
293 | static bool event_can_be_cpu_signaled(const struct kfd_event *ev) | |
294 | { | |
295 | return ev->type == KFD_EVENT_TYPE_SIGNAL; | |
296 | } | |
297 | ||
0fc8011f FK |
298 | int 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 | ||
f3a39818 AL |
322 | int 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 | ||
74e40716 | 337 | init_waitqueue_head(&ev->wq); |
f3a39818 AL |
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) { | |
df03ef93 | 348 | *event_page_offset = KFD_MMAP_TYPE_EVENTS; |
482f0777 | 349 | *event_slot_index = ev->event_id; |
f3a39818 AL |
350 | } |
351 | break; | |
352 | default: | |
353 | ret = create_other_event(p, ev); | |
354 | break; | |
355 | } | |
356 | ||
357 | if (!ret) { | |
f3a39818 AL |
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. */ | |
370 | int 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 | ||
388 | static void set_event(struct kfd_event *ev) | |
389 | { | |
390 | struct kfd_event_waiter *waiter; | |
f3a39818 | 391 | |
74e40716 FK |
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); | |
f3a39818 | 398 | |
74e40716 | 399 | list_for_each_entry(waiter, &ev->wq.head, wait.entry) |
f3a39818 AL |
400 | waiter->activated = true; |
401 | ||
74e40716 | 402 | wake_up_all(&ev->wq); |
f3a39818 AL |
403 | } |
404 | ||
405 | /* Assumes that p is current. */ | |
406 | int 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 | ||
424 | static void reset_event(struct kfd_event *ev) | |
425 | { | |
426 | ev->signaled = false; | |
427 | } | |
428 | ||
429 | /* Assumes that p is current. */ | |
430 | int 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 | ||
449 | static void acknowledge_signal(struct kfd_process *p, struct kfd_event *ev) | |
450 | { | |
482f0777 | 451 | page_slots(p->signal_page)[ev->event_id] = UNSIGNALED_EVENT_SLOT; |
f3a39818 AL |
452 | } |
453 | ||
454 | static 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 | ||
c7b6bac9 | 463 | void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id, |
f3a39818 AL |
464 | uint32_t valid_id_bits) |
465 | { | |
3f04f961 | 466 | struct kfd_event *ev = NULL; |
f3a39818 AL |
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 | |
abb208a8 | 471 | * running so the lookup function increments the process ref count. |
f3a39818 AL |
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 | ||
3f04f961 FK |
480 | if (valid_id_bits) |
481 | ev = lookup_signaled_event_by_partial_id(p, partial_id, | |
482 | valid_id_bits); | |
483 | if (ev) { | |
f3a39818 | 484 | set_event_from_interrupt(p, ev); |
50cb7dd9 | 485 | } else if (p->signal_page) { |
f3a39818 | 486 | /* |
3f04f961 FK |
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. | |
f3a39818 | 490 | */ |
482f0777 FK |
491 | uint64_t *slots = page_slots(p->signal_page); |
492 | uint32_t id; | |
f3a39818 | 493 | |
3f04f961 FK |
494 | if (valid_id_bits) |
495 | pr_debug_ratelimited("Partial ID invalid: %u (%u valid bits)\n", | |
496 | partial_id, valid_id_bits); | |
497 | ||
eeb27b7e | 498 | if (p->signal_event_count < KFD_SIGNAL_EVENT_LIMIT / 64) { |
482f0777 FK |
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); | |
50cb7dd9 | 508 | } |
482f0777 FK |
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 | } | |
f3a39818 AL |
520 | } |
521 | ||
522 | mutex_unlock(&p->event_mutex); | |
abb208a8 | 523 | kfd_unref_process(p); |
f3a39818 AL |
524 | } |
525 | ||
526 | static 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++) { | |
74e40716 | 536 | init_wait(&event_waiters[i].wait); |
f3a39818 AL |
537 | event_waiters[i].activated = false; |
538 | } | |
539 | ||
540 | return event_waiters; | |
541 | } | |
542 | ||
1f9d09be | 543 | static int init_event_waiter_get_status(struct kfd_process *p, |
59d3e8be | 544 | struct kfd_event_waiter *waiter, |
ebf947fe | 545 | uint32_t event_id) |
f3a39818 AL |
546 | { |
547 | struct kfd_event *ev = lookup_event_by_id(p, event_id); | |
548 | ||
549 | if (!ev) | |
550 | return -EINVAL; | |
551 | ||
59d3e8be | 552 | waiter->event = ev; |
f3a39818 AL |
553 | waiter->activated = ev->signaled; |
554 | ev->signaled = ev->signaled && !ev->auto_reset; | |
555 | ||
f3a39818 AL |
556 | return 0; |
557 | } | |
558 | ||
1f9d09be SK |
559 | static 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) | |
74e40716 | 567 | add_wait_queue(&ev->wq, &waiter->wait); |
1f9d09be SK |
568 | } |
569 | ||
fe528c13 FK |
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 | */ | |
580 | static uint32_t test_event_condition(bool all, uint32_t num_events, | |
f3a39818 AL |
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++) { | |
fe528c13 FK |
587 | if (!event_waiters[i].event) |
588 | return KFD_IOC_WAIT_RESULT_FAIL; | |
589 | ||
f3a39818 AL |
590 | if (event_waiters[i].activated) { |
591 | if (!all) | |
fe528c13 | 592 | return KFD_IOC_WAIT_RESULT_COMPLETE; |
f3a39818 AL |
593 | |
594 | activated_count++; | |
595 | } | |
596 | } | |
597 | ||
fe528c13 FK |
598 | return activated_count == num_events ? |
599 | KFD_IOC_WAIT_RESULT_COMPLETE : KFD_IOC_WAIT_RESULT_TIMEOUT; | |
f3a39818 AL |
600 | } |
601 | ||
59d3e8be AS |
602 | /* |
603 | * Copy event specific data, if defined. | |
604 | * Currently only memory exception events have additional data to copy to user | |
605 | */ | |
fdf0c833 | 606 | static int copy_signaled_event_data(uint32_t num_events, |
59d3e8be AS |
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) { | |
ebf947fe | 620 | dst = &data[i].memory_exception_data; |
59d3e8be AS |
621 | src = &event->memory_exception_data; |
622 | if (copy_to_user(dst, src, | |
623 | sizeof(struct kfd_hsa_memory_exception_data))) | |
fdf0c833 | 624 | return -EFAULT; |
59d3e8be AS |
625 | } |
626 | } | |
627 | ||
fdf0c833 | 628 | return 0; |
59d3e8be AS |
629 | |
630 | } | |
631 | ||
632 | ||
633 | ||
f3a39818 AL |
634 | static 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 | ||
652 | static 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++) | |
74e40716 FK |
657 | if (waiters[i].event) |
658 | remove_wait_queue(&waiters[i].event->wq, | |
659 | &waiters[i].wait); | |
f3a39818 AL |
660 | |
661 | kfree(waiters); | |
662 | } | |
663 | ||
664 | int kfd_wait_on_events(struct kfd_process *p, | |
59d3e8be | 665 | uint32_t num_events, void __user *data, |
f3a39818 | 666 | bool all, uint32_t user_timeout_ms, |
fdf0c833 | 667 | uint32_t *wait_result) |
f3a39818 | 668 | { |
59d3e8be AS |
669 | struct kfd_event_data __user *events = |
670 | (struct kfd_event_data __user *) data; | |
f3a39818 AL |
671 | uint32_t i; |
672 | int ret = 0; | |
1f9d09be | 673 | |
f3a39818 AL |
674 | struct kfd_event_waiter *event_waiters = NULL; |
675 | long timeout = user_timeout_to_jiffies(user_timeout_ms); | |
676 | ||
fdf0c833 FK |
677 | event_waiters = alloc_event_waiters(num_events); |
678 | if (!event_waiters) { | |
679 | ret = -ENOMEM; | |
680 | goto out; | |
681 | } | |
682 | ||
f3a39818 AL |
683 | mutex_lock(&p->event_mutex); |
684 | ||
f3a39818 | 685 | for (i = 0; i < num_events; i++) { |
59d3e8be | 686 | struct kfd_event_data event_data; |
f3a39818 | 687 | |
59d3e8be | 688 | if (copy_from_user(&event_data, &events[i], |
8bf79388 PB |
689 | sizeof(struct kfd_event_data))) { |
690 | ret = -EFAULT; | |
fdf0c833 | 691 | goto out_unlock; |
8bf79388 | 692 | } |
f3a39818 | 693 | |
1f9d09be | 694 | ret = init_event_waiter_get_status(p, &event_waiters[i], |
ebf947fe | 695 | event_data.event_id); |
f3a39818 | 696 | if (ret) |
fdf0c833 | 697 | goto out_unlock; |
f3a39818 AL |
698 | } |
699 | ||
1f9d09be | 700 | /* Check condition once. */ |
fe528c13 FK |
701 | *wait_result = test_event_condition(all, num_events, event_waiters); |
702 | if (*wait_result == KFD_IOC_WAIT_RESULT_COMPLETE) { | |
fdf0c833 FK |
703 | ret = copy_signaled_event_data(num_events, |
704 | event_waiters, events); | |
705 | goto out_unlock; | |
fe528c13 FK |
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; | |
1f9d09be SK |
711 | } |
712 | ||
fe528c13 FK |
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 | ||
f3a39818 AL |
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 | ||
d9aeec4c SK |
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 | ||
fe528c13 FK |
749 | *wait_result = test_event_condition(all, num_events, |
750 | event_waiters); | |
751 | if (*wait_result != KFD_IOC_WAIT_RESULT_TIMEOUT) | |
f3a39818 | 752 | break; |
f3a39818 | 753 | |
fe528c13 | 754 | if (timeout <= 0) |
f3a39818 | 755 | break; |
f3a39818 | 756 | |
d9aeec4c | 757 | timeout = schedule_timeout(timeout); |
f3a39818 AL |
758 | } |
759 | __set_current_state(TASK_RUNNING); | |
760 | ||
fdf0c833 FK |
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 | ||
f3a39818 | 768 | mutex_lock(&p->event_mutex); |
fdf0c833 | 769 | out_unlock: |
f3a39818 AL |
770 | free_waiters(num_events, event_waiters); |
771 | mutex_unlock(&p->event_mutex); | |
fdf0c833 FK |
772 | out: |
773 | if (ret) | |
774 | *wait_result = KFD_IOC_WAIT_RESULT_FAIL; | |
fe528c13 FK |
775 | else if (*wait_result == KFD_IOC_WAIT_RESULT_FAIL) |
776 | ret = -EIO; | |
f3a39818 AL |
777 | |
778 | return ret; | |
779 | } | |
780 | ||
781 | int kfd_event_mmap(struct kfd_process *p, struct vm_area_struct *vma) | |
782 | { | |
f3a39818 | 783 | unsigned long pfn; |
50cb7dd9 | 784 | struct kfd_signal_page *page; |
b9a5d0a5 | 785 | int ret; |
f3a39818 | 786 | |
b9a5d0a5 FK |
787 | /* check required size doesn't exceed the allocated size */ |
788 | if (get_order(KFD_SIGNAL_EVENT_LIMIT * 8) < | |
f3a39818 | 789 | get_order(vma->vm_end - vma->vm_start)) { |
79775b62 | 790 | pr_err("Event page mmap requested illegal size\n"); |
f3a39818 AL |
791 | return -EINVAL; |
792 | } | |
793 | ||
50cb7dd9 | 794 | page = p->signal_page; |
f3a39818 AL |
795 | if (!page) { |
796 | /* Probably KFD bug, but mmap is user-accessible. */ | |
50cb7dd9 | 797 | pr_debug("Signal page could not be found\n"); |
f3a39818 AL |
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 | ||
79775b62 | 807 | pr_debug("Mapping signal page\n"); |
f3a39818 AL |
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 */ | |
b9a5d0a5 | 818 | ret = remap_pfn_range(vma, vma->vm_start, pfn, |
f3a39818 | 819 | vma->vm_end - vma->vm_start, vma->vm_page_prot); |
b9a5d0a5 FK |
820 | if (!ret) |
821 | p->signal_mapped_size = vma->vm_end - vma->vm_start; | |
822 | ||
823 | return ret; | |
f3a39818 | 824 | } |
59d3e8be AS |
825 | |
826 | /* | |
827 | * Assumes that p->event_mutex is held and of course | |
828 | * that p is not going away (current or locked). | |
829 | */ | |
830 | static 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; | |
482f0777 | 835 | uint32_t id; |
59d3e8be AS |
836 | bool send_signal = true; |
837 | ||
838 | ev_data = (struct kfd_hsa_memory_exception_data *) event_data; | |
839 | ||
482f0777 FK |
840 | id = KFD_FIRST_NONSIGNAL_EVENT_ID; |
841 | idr_for_each_entry_continue(&p->event_idr, ev, id) | |
59d3e8be AS |
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 | ||
101fee63 MR |
852 | if (type == KFD_EVENT_TYPE_MEMORY) { |
853 | dev_warn(kfd_device, | |
6027b1bf YZ |
854 | "Sending SIGSEGV to process %d (pasid 0x%x)", |
855 | p->lead_thread->pid, p->pasid); | |
101fee63 MR |
856 | send_sig(SIGSEGV, p->lead_thread, 0); |
857 | } | |
858 | ||
59d3e8be AS |
859 | /* Send SIGTERM no event of type "type" has been found*/ |
860 | if (send_signal) { | |
81663016 OG |
861 | if (send_sigterm) { |
862 | dev_warn(kfd_device, | |
6027b1bf YZ |
863 | "Sending SIGTERM to process %d (pasid 0x%x)", |
864 | p->lead_thread->pid, p->pasid); | |
81663016 OG |
865 | send_sig(SIGTERM, p->lead_thread, 0); |
866 | } else { | |
867 | dev_err(kfd_device, | |
6027b1bf YZ |
868 | "Process %d (pasid 0x%x) got unhandled exception", |
869 | p->lead_thread->pid, p->pasid); | |
81663016 | 870 | } |
59d3e8be AS |
871 | } |
872 | } | |
873 | ||
64d1c3a4 | 874 | #ifdef KFD_SUPPORT_IOMMU_V2 |
c7b6bac9 | 875 | void kfd_signal_iommu_event(struct kfd_dev *dev, u32 pasid, |
59d3e8be AS |
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 | |
abb208a8 | 885 | * running so the lookup function increments the process ref count. |
59d3e8be AS |
886 | */ |
887 | struct kfd_process *p = kfd_lookup_process_by_pasid(pasid); | |
9b56bb11 | 888 | struct mm_struct *mm; |
59d3e8be AS |
889 | |
890 | if (!p) | |
891 | return; /* Presumably process exited. */ | |
892 | ||
9b56bb11 FK |
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) { | |
abb208a8 | 898 | kfd_unref_process(p); |
9b56bb11 FK |
899 | return; /* Process is exiting */ |
900 | } | |
901 | ||
59d3e8be AS |
902 | memset(&memory_exception_data, 0, sizeof(memory_exception_data)); |
903 | ||
d8ed45c5 | 904 | mmap_read_lock(mm); |
9b56bb11 | 905 | vma = find_vma(mm, address); |
59d3e8be AS |
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; | |
359cecdd YZ |
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 | |
59d3e8be | 919 | memory_exception_data.failure.ReadOnly = 0; |
359cecdd YZ |
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; | |
59d3e8be AS |
925 | } |
926 | ||
d8ed45c5 | 927 | mmap_read_unlock(mm); |
9b56bb11 | 928 | mmput(mm); |
59d3e8be | 929 | |
8725aeca YZ |
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); | |
59d3e8be | 934 | |
8725aeca YZ |
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 | */ | |
514e5e7e HR |
938 | if (dev->device_info->asic_family != CHIP_RAVEN && |
939 | dev->device_info->asic_family != CHIP_RENOIR) { | |
8725aeca YZ |
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 | } | |
59d3e8be | 948 | |
abb208a8 | 949 | kfd_unref_process(p); |
59d3e8be | 950 | } |
64d1c3a4 | 951 | #endif /* KFD_SUPPORT_IOMMU_V2 */ |
930c5ff4 | 952 | |
c7b6bac9 | 953 | void kfd_signal_hw_exception_event(u32 pasid) |
930c5ff4 AS |
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 | |
abb208a8 | 958 | * running so the lookup function increments the process ref count. |
930c5ff4 AS |
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); | |
abb208a8 | 971 | kfd_unref_process(p); |
930c5ff4 | 972 | } |
2640c3fa | 973 | |
c7b6bac9 | 974 | void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid, |
2640c3fa | 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; | |
0d87c9cf | 986 | memory_exception_data.failure.imprecise = true; |
2640c3fa | 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 | } | |
e42051d2 SL |
1010 | |
1011 | void kfd_signal_reset_event(struct kfd_dev *dev) | |
1012 | { | |
1013 | struct kfd_hsa_hw_exception_data hw_exception_data; | |
9b54d201 | 1014 | struct kfd_hsa_memory_exception_data memory_exception_data; |
e42051d2 SL |
1015 | struct kfd_process *p; |
1016 | struct kfd_event *ev; | |
1017 | unsigned int temp; | |
1018 | uint32_t id, idx; | |
9b54d201 EH |
1019 | int reset_cause = atomic_read(&dev->sram_ecc_flag) ? |
1020 | KFD_HW_EXCEPTION_ECC : | |
1021 | KFD_HW_EXCEPTION_GPU_HANG; | |
e42051d2 SL |
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; | |
9b54d201 EH |
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; | |
e42051d2 SL |
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; | |
9b54d201 | 1038 | idr_for_each_entry_continue(&p->event_idr, ev, id) { |
e42051d2 SL |
1039 | if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) { |
1040 | ev->hw_exception_data = hw_exception_data; | |
1041 | set_event(ev); | |
1042 | } | |
9b54d201 EH |
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 | } | |
e42051d2 SL |
1049 | mutex_unlock(&p->event_mutex); |
1050 | } | |
1051 | srcu_read_unlock(&kfd_processes_srcu, idx); | |
1052 | } |