Commit | Line | Data |
---|---|---|
d38ceaf9 AD |
1 | /* |
2 | * Copyright 2009 Jerome Glisse. | |
3 | * All Rights Reserved. | |
4 | * | |
5 | * Permission is hereby granted, free of charge, to any person obtaining a | |
6 | * copy of this software and associated documentation files (the | |
7 | * "Software"), to deal in the Software without restriction, including | |
8 | * without limitation the rights to use, copy, modify, merge, publish, | |
9 | * distribute, sub license, and/or sell copies of the Software, and to | |
10 | * permit persons to whom the Software is furnished to do so, subject to | |
11 | * the following conditions: | |
12 | * | |
13 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | |
14 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | |
15 | * FITNESS FOR A PARTICULAR PURPOSE AND NON-INFRINGEMENT. IN NO EVENT SHALL | |
16 | * THE COPYRIGHT HOLDERS, AUTHORS AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, | |
17 | * DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR | |
18 | * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE | |
19 | * USE OR OTHER DEALINGS IN THE SOFTWARE. | |
20 | * | |
21 | * The above copyright notice and this permission notice (including the | |
22 | * next paragraph) shall be included in all copies or substantial portions | |
23 | * of the Software. | |
24 | * | |
25 | */ | |
26 | /* | |
27 | * Authors: | |
28 | * Jerome Glisse <glisse@freedesktop.org> | |
29 | * Dave Airlie | |
30 | */ | |
31 | #include <linux/seq_file.h> | |
32 | #include <linux/atomic.h> | |
33 | #include <linux/wait.h> | |
34 | #include <linux/kref.h> | |
35 | #include <linux/slab.h> | |
36 | #include <linux/firmware.h> | |
37 | #include <drm/drmP.h> | |
38 | #include "amdgpu.h" | |
39 | #include "amdgpu_trace.h" | |
40 | ||
41 | /* | |
42 | * Fences | |
43 | * Fences mark an event in the GPUs pipeline and are used | |
44 | * for GPU/CPU synchronization. When the fence is written, | |
45 | * it is expected that all buffers associated with that fence | |
46 | * are no longer in use by the associated ring on the GPU and | |
47 | * that the the relevant GPU caches have been flushed. | |
48 | */ | |
49 | ||
b49c84a5 CZ |
50 | static struct kmem_cache *amdgpu_fence_slab; |
51 | static atomic_t amdgpu_fence_slab_ref = ATOMIC_INIT(0); | |
52 | ||
d38ceaf9 AD |
53 | /** |
54 | * amdgpu_fence_write - write a fence value | |
55 | * | |
56 | * @ring: ring the fence is associated with | |
57 | * @seq: sequence number to write | |
58 | * | |
59 | * Writes a fence value to memory (all asics). | |
60 | */ | |
61 | static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq) | |
62 | { | |
63 | struct amdgpu_fence_driver *drv = &ring->fence_drv; | |
64 | ||
65 | if (drv->cpu_addr) | |
66 | *drv->cpu_addr = cpu_to_le32(seq); | |
67 | } | |
68 | ||
69 | /** | |
70 | * amdgpu_fence_read - read a fence value | |
71 | * | |
72 | * @ring: ring the fence is associated with | |
73 | * | |
74 | * Reads a fence value from memory (all asics). | |
75 | * Returns the value of the fence read from memory. | |
76 | */ | |
77 | static u32 amdgpu_fence_read(struct amdgpu_ring *ring) | |
78 | { | |
79 | struct amdgpu_fence_driver *drv = &ring->fence_drv; | |
80 | u32 seq = 0; | |
81 | ||
82 | if (drv->cpu_addr) | |
83 | seq = le32_to_cpu(*drv->cpu_addr); | |
84 | else | |
85 | seq = lower_32_bits(atomic64_read(&drv->last_seq)); | |
86 | ||
87 | return seq; | |
88 | } | |
89 | ||
d38ceaf9 AD |
90 | /** |
91 | * amdgpu_fence_emit - emit a fence on the requested ring | |
92 | * | |
93 | * @ring: ring the fence is associated with | |
94 | * @owner: creator of the fence | |
95 | * @fence: amdgpu fence object | |
96 | * | |
97 | * Emits a fence command on the requested ring (all asics). | |
98 | * Returns 0 on success, -ENOMEM on failure. | |
99 | */ | |
100 | int amdgpu_fence_emit(struct amdgpu_ring *ring, void *owner, | |
101 | struct amdgpu_fence **fence) | |
102 | { | |
103 | struct amdgpu_device *adev = ring->adev; | |
104 | ||
105 | /* we are protected by the ring emission mutex */ | |
b49c84a5 | 106 | *fence = kmem_cache_alloc(amdgpu_fence_slab, GFP_KERNEL); |
d38ceaf9 AD |
107 | if ((*fence) == NULL) { |
108 | return -ENOMEM; | |
109 | } | |
110 | (*fence)->seq = ++ring->fence_drv.sync_seq[ring->idx]; | |
111 | (*fence)->ring = ring; | |
112 | (*fence)->owner = owner; | |
113 | fence_init(&(*fence)->base, &amdgpu_fence_ops, | |
7f06c236 | 114 | &ring->fence_drv.fence_queue.lock, |
115 | adev->fence_context + ring->idx, | |
d38ceaf9 | 116 | (*fence)->seq); |
890ee23f CZ |
117 | amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr, |
118 | (*fence)->seq, | |
119 | AMDGPU_FENCE_FLAG_INT); | |
d38ceaf9 AD |
120 | return 0; |
121 | } | |
122 | ||
c2776afe CK |
123 | /** |
124 | * amdgpu_fence_schedule_fallback - schedule fallback check | |
125 | * | |
126 | * @ring: pointer to struct amdgpu_ring | |
127 | * | |
128 | * Start a timer as fallback to our interrupts. | |
129 | */ | |
130 | static void amdgpu_fence_schedule_fallback(struct amdgpu_ring *ring) | |
131 | { | |
132 | mod_timer(&ring->fence_drv.fallback_timer, | |
133 | jiffies + AMDGPU_FENCE_JIFFIES_TIMEOUT); | |
134 | } | |
135 | ||
d38ceaf9 AD |
136 | /** |
137 | * amdgpu_fence_activity - check for fence activity | |
138 | * | |
139 | * @ring: pointer to struct amdgpu_ring | |
140 | * | |
141 | * Checks the current fence value and calculates the last | |
142 | * signalled fence value. Returns true if activity occured | |
143 | * on the ring, and the fence_queue should be waken up. | |
144 | */ | |
145 | static bool amdgpu_fence_activity(struct amdgpu_ring *ring) | |
146 | { | |
147 | uint64_t seq, last_seq, last_emitted; | |
148 | unsigned count_loop = 0; | |
149 | bool wake = false; | |
150 | ||
151 | /* Note there is a scenario here for an infinite loop but it's | |
152 | * very unlikely to happen. For it to happen, the current polling | |
153 | * process need to be interrupted by another process and another | |
154 | * process needs to update the last_seq btw the atomic read and | |
155 | * xchg of the current process. | |
156 | * | |
157 | * More over for this to go in infinite loop there need to be | |
86c2b790 | 158 | * continuously new fence signaled ie amdgpu_fence_read needs |
d38ceaf9 AD |
159 | * to return a different value each time for both the currently |
160 | * polling process and the other process that xchg the last_seq | |
161 | * btw atomic read and xchg of the current process. And the | |
162 | * value the other process set as last seq must be higher than | |
163 | * the seq value we just read. Which means that current process | |
86c2b790 | 164 | * need to be interrupted after amdgpu_fence_read and before |
d38ceaf9 AD |
165 | * atomic xchg. |
166 | * | |
167 | * To be even more safe we count the number of time we loop and | |
168 | * we bail after 10 loop just accepting the fact that we might | |
169 | * have temporarly set the last_seq not to the true real last | |
170 | * seq but to an older one. | |
171 | */ | |
172 | last_seq = atomic64_read(&ring->fence_drv.last_seq); | |
173 | do { | |
174 | last_emitted = ring->fence_drv.sync_seq[ring->idx]; | |
175 | seq = amdgpu_fence_read(ring); | |
176 | seq |= last_seq & 0xffffffff00000000LL; | |
177 | if (seq < last_seq) { | |
178 | seq &= 0xffffffff; | |
179 | seq |= last_emitted & 0xffffffff00000000LL; | |
180 | } | |
181 | ||
182 | if (seq <= last_seq || seq > last_emitted) { | |
183 | break; | |
184 | } | |
185 | /* If we loop over we don't want to return without | |
186 | * checking if a fence is signaled as it means that the | |
187 | * seq we just read is different from the previous on. | |
188 | */ | |
189 | wake = true; | |
190 | last_seq = seq; | |
191 | if ((count_loop++) > 10) { | |
192 | /* We looped over too many time leave with the | |
193 | * fact that we might have set an older fence | |
194 | * seq then the current real last seq as signaled | |
195 | * by the hw. | |
196 | */ | |
197 | break; | |
198 | } | |
199 | } while (atomic64_xchg(&ring->fence_drv.last_seq, seq) > seq); | |
200 | ||
201 | if (seq < last_emitted) | |
c2776afe | 202 | amdgpu_fence_schedule_fallback(ring); |
d38ceaf9 AD |
203 | |
204 | return wake; | |
205 | } | |
206 | ||
207 | /** | |
c2776afe | 208 | * amdgpu_fence_process - process a fence |
d38ceaf9 | 209 | * |
c2776afe CK |
210 | * @adev: amdgpu_device pointer |
211 | * @ring: ring index the fence is associated with | |
d38ceaf9 | 212 | * |
c2776afe CK |
213 | * Checks the current fence value and wakes the fence queue |
214 | * if the sequence number has increased (all asics). | |
d38ceaf9 | 215 | */ |
c2776afe | 216 | void amdgpu_fence_process(struct amdgpu_ring *ring) |
d38ceaf9 | 217 | { |
0c418f10 | 218 | if (amdgpu_fence_activity(ring)) |
7f06c236 | 219 | wake_up_all(&ring->fence_drv.fence_queue); |
d38ceaf9 AD |
220 | } |
221 | ||
222 | /** | |
c2776afe | 223 | * amdgpu_fence_fallback - fallback for hardware interrupts |
d38ceaf9 | 224 | * |
c2776afe | 225 | * @work: delayed work item |
d38ceaf9 | 226 | * |
c2776afe | 227 | * Checks for fence activity. |
d38ceaf9 | 228 | */ |
c2776afe | 229 | static void amdgpu_fence_fallback(unsigned long arg) |
d38ceaf9 | 230 | { |
c2776afe CK |
231 | struct amdgpu_ring *ring = (void *)arg; |
232 | ||
233 | amdgpu_fence_process(ring); | |
d38ceaf9 AD |
234 | } |
235 | ||
236 | /** | |
237 | * amdgpu_fence_seq_signaled - check if a fence sequence number has signaled | |
238 | * | |
239 | * @ring: ring the fence is associated with | |
240 | * @seq: sequence number | |
241 | * | |
242 | * Check if the last signaled fence sequnce number is >= the requested | |
243 | * sequence number (all asics). | |
244 | * Returns true if the fence has signaled (current fence value | |
245 | * is >= requested value) or false if it has not (current fence | |
246 | * value is < the requested value. Helper function for | |
247 | * amdgpu_fence_signaled(). | |
248 | */ | |
249 | static bool amdgpu_fence_seq_signaled(struct amdgpu_ring *ring, u64 seq) | |
250 | { | |
251 | if (atomic64_read(&ring->fence_drv.last_seq) >= seq) | |
252 | return true; | |
253 | ||
254 | /* poll new last sequence at least once */ | |
255 | amdgpu_fence_process(ring); | |
256 | if (atomic64_read(&ring->fence_drv.last_seq) >= seq) | |
257 | return true; | |
258 | ||
259 | return false; | |
260 | } | |
261 | ||
7f06c236 | 262 | /* |
263 | * amdgpu_ring_wait_seq_timeout - wait for seq of the specific ring to signal | |
264 | * @ring: ring to wait on for the seq number | |
265 | * @seq: seq number wait for | |
7f06c236 | 266 | * |
267 | * return value: | |
00d2a2b2 CK |
268 | * 0: seq signaled, and gpu not hang |
269 | * -EDEADL: GPU hang detected | |
7f06c236 | 270 | * -EINVAL: some paramter is not valid |
d38ceaf9 | 271 | */ |
00d2a2b2 | 272 | static int amdgpu_fence_ring_wait_seq(struct amdgpu_ring *ring, uint64_t seq) |
d38ceaf9 | 273 | { |
7f06c236 | 274 | bool signaled = false; |
d38ceaf9 | 275 | |
7f06c236 | 276 | BUG_ON(!ring); |
277 | if (seq > ring->fence_drv.sync_seq[ring->idx]) | |
278 | return -EINVAL; | |
d38ceaf9 | 279 | |
7f06c236 | 280 | if (atomic64_read(&ring->fence_drv.last_seq) >= seq) |
00d2a2b2 | 281 | return 0; |
d38ceaf9 | 282 | |
c2776afe | 283 | amdgpu_fence_schedule_fallback(ring); |
00d2a2b2 | 284 | wait_event(ring->fence_drv.fence_queue, ( |
b7e4dad3 | 285 | (signaled = amdgpu_fence_seq_signaled(ring, seq)))); |
d38ceaf9 | 286 | |
00d2a2b2 CK |
287 | if (signaled) |
288 | return 0; | |
289 | else | |
290 | return -EDEADLK; | |
d38ceaf9 AD |
291 | } |
292 | ||
d38ceaf9 AD |
293 | /** |
294 | * amdgpu_fence_wait_next - wait for the next fence to signal | |
295 | * | |
296 | * @adev: amdgpu device pointer | |
297 | * @ring: ring index the fence is associated with | |
298 | * | |
299 | * Wait for the next fence on the requested ring to signal (all asics). | |
300 | * Returns 0 if the next fence has passed, error for all other cases. | |
301 | * Caller must hold ring lock. | |
302 | */ | |
303 | int amdgpu_fence_wait_next(struct amdgpu_ring *ring) | |
304 | { | |
7f06c236 | 305 | uint64_t seq = atomic64_read(&ring->fence_drv.last_seq) + 1ULL; |
00d2a2b2 | 306 | |
7f06c236 | 307 | if (seq >= ring->fence_drv.sync_seq[ring->idx]) |
d38ceaf9 | 308 | return -ENOENT; |
7f06c236 | 309 | |
00d2a2b2 | 310 | return amdgpu_fence_ring_wait_seq(ring, seq); |
d38ceaf9 AD |
311 | } |
312 | ||
313 | /** | |
314 | * amdgpu_fence_wait_empty - wait for all fences to signal | |
315 | * | |
316 | * @adev: amdgpu device pointer | |
317 | * @ring: ring index the fence is associated with | |
318 | * | |
319 | * Wait for all fences on the requested ring to signal (all asics). | |
320 | * Returns 0 if the fences have passed, error for all other cases. | |
321 | * Caller must hold ring lock. | |
322 | */ | |
323 | int amdgpu_fence_wait_empty(struct amdgpu_ring *ring) | |
324 | { | |
7f06c236 | 325 | uint64_t seq = ring->fence_drv.sync_seq[ring->idx]; |
00d2a2b2 | 326 | |
7f06c236 | 327 | if (!seq) |
d38ceaf9 AD |
328 | return 0; |
329 | ||
00d2a2b2 | 330 | return amdgpu_fence_ring_wait_seq(ring, seq); |
d38ceaf9 AD |
331 | } |
332 | ||
d38ceaf9 AD |
333 | /** |
334 | * amdgpu_fence_count_emitted - get the count of emitted fences | |
335 | * | |
336 | * @ring: ring the fence is associated with | |
337 | * | |
338 | * Get the number of fences emitted on the requested ring (all asics). | |
339 | * Returns the number of emitted fences on the ring. Used by the | |
340 | * dynpm code to ring track activity. | |
341 | */ | |
342 | unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring) | |
343 | { | |
344 | uint64_t emitted; | |
345 | ||
346 | /* We are not protected by ring lock when reading the last sequence | |
347 | * but it's ok to report slightly wrong fence count here. | |
348 | */ | |
349 | amdgpu_fence_process(ring); | |
350 | emitted = ring->fence_drv.sync_seq[ring->idx] | |
351 | - atomic64_read(&ring->fence_drv.last_seq); | |
352 | /* to avoid 32bits warp around */ | |
353 | if (emitted > 0x10000000) | |
354 | emitted = 0x10000000; | |
355 | ||
356 | return (unsigned)emitted; | |
357 | } | |
358 | ||
359 | /** | |
360 | * amdgpu_fence_need_sync - do we need a semaphore | |
361 | * | |
362 | * @fence: amdgpu fence object | |
363 | * @dst_ring: which ring to check against | |
364 | * | |
365 | * Check if the fence needs to be synced against another ring | |
366 | * (all asics). If so, we need to emit a semaphore. | |
367 | * Returns true if we need to sync with another ring, false if | |
368 | * not. | |
369 | */ | |
370 | bool amdgpu_fence_need_sync(struct amdgpu_fence *fence, | |
371 | struct amdgpu_ring *dst_ring) | |
372 | { | |
373 | struct amdgpu_fence_driver *fdrv; | |
374 | ||
375 | if (!fence) | |
376 | return false; | |
377 | ||
378 | if (fence->ring == dst_ring) | |
379 | return false; | |
380 | ||
381 | /* we are protected by the ring mutex */ | |
382 | fdrv = &dst_ring->fence_drv; | |
383 | if (fence->seq <= fdrv->sync_seq[fence->ring->idx]) | |
384 | return false; | |
385 | ||
386 | return true; | |
387 | } | |
388 | ||
389 | /** | |
390 | * amdgpu_fence_note_sync - record the sync point | |
391 | * | |
392 | * @fence: amdgpu fence object | |
393 | * @dst_ring: which ring to check against | |
394 | * | |
395 | * Note the sequence number at which point the fence will | |
396 | * be synced with the requested ring (all asics). | |
397 | */ | |
398 | void amdgpu_fence_note_sync(struct amdgpu_fence *fence, | |
399 | struct amdgpu_ring *dst_ring) | |
400 | { | |
401 | struct amdgpu_fence_driver *dst, *src; | |
402 | unsigned i; | |
403 | ||
404 | if (!fence) | |
405 | return; | |
406 | ||
407 | if (fence->ring == dst_ring) | |
408 | return; | |
409 | ||
410 | /* we are protected by the ring mutex */ | |
411 | src = &fence->ring->fence_drv; | |
412 | dst = &dst_ring->fence_drv; | |
413 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { | |
414 | if (i == dst_ring->idx) | |
415 | continue; | |
416 | ||
417 | dst->sync_seq[i] = max(dst->sync_seq[i], src->sync_seq[i]); | |
418 | } | |
419 | } | |
420 | ||
421 | /** | |
422 | * amdgpu_fence_driver_start_ring - make the fence driver | |
423 | * ready for use on the requested ring. | |
424 | * | |
425 | * @ring: ring to start the fence driver on | |
426 | * @irq_src: interrupt source to use for this ring | |
427 | * @irq_type: interrupt type to use for this ring | |
428 | * | |
429 | * Make the fence driver ready for processing (all asics). | |
430 | * Not all asics have all rings, so each asic will only | |
431 | * start the fence driver on the rings it has. | |
432 | * Returns 0 for success, errors for failure. | |
433 | */ | |
434 | int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring, | |
435 | struct amdgpu_irq_src *irq_src, | |
436 | unsigned irq_type) | |
437 | { | |
438 | struct amdgpu_device *adev = ring->adev; | |
439 | uint64_t index; | |
440 | ||
441 | if (ring != &adev->uvd.ring) { | |
442 | ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs]; | |
443 | ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4); | |
444 | } else { | |
445 | /* put fence directly behind firmware */ | |
446 | index = ALIGN(adev->uvd.fw->size, 8); | |
447 | ring->fence_drv.cpu_addr = adev->uvd.cpu_addr + index; | |
448 | ring->fence_drv.gpu_addr = adev->uvd.gpu_addr + index; | |
449 | } | |
450 | amdgpu_fence_write(ring, atomic64_read(&ring->fence_drv.last_seq)); | |
c6a4079b CZ |
451 | amdgpu_irq_get(adev, irq_src, irq_type); |
452 | ||
d38ceaf9 AD |
453 | ring->fence_drv.irq_src = irq_src; |
454 | ring->fence_drv.irq_type = irq_type; | |
c6a4079b CZ |
455 | ring->fence_drv.initialized = true; |
456 | ||
d38ceaf9 AD |
457 | dev_info(adev->dev, "fence driver on ring %d use gpu addr 0x%016llx, " |
458 | "cpu addr 0x%p\n", ring->idx, | |
459 | ring->fence_drv.gpu_addr, ring->fence_drv.cpu_addr); | |
460 | return 0; | |
461 | } | |
462 | ||
463 | /** | |
464 | * amdgpu_fence_driver_init_ring - init the fence driver | |
465 | * for the requested ring. | |
466 | * | |
467 | * @ring: ring to init the fence driver on | |
468 | * | |
469 | * Init the fence driver for the requested ring (all asics). | |
470 | * Helper function for amdgpu_fence_driver_init(). | |
471 | */ | |
4f839a24 | 472 | int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring) |
d38ceaf9 | 473 | { |
4f839a24 | 474 | int i, r; |
d38ceaf9 AD |
475 | |
476 | ring->fence_drv.cpu_addr = NULL; | |
477 | ring->fence_drv.gpu_addr = 0; | |
478 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) | |
479 | ring->fence_drv.sync_seq[i] = 0; | |
480 | ||
481 | atomic64_set(&ring->fence_drv.last_seq, 0); | |
482 | ring->fence_drv.initialized = false; | |
483 | ||
c2776afe CK |
484 | setup_timer(&ring->fence_drv.fallback_timer, amdgpu_fence_fallback, |
485 | (unsigned long)ring); | |
b80d8475 | 486 | |
5ec92a76 CK |
487 | init_waitqueue_head(&ring->fence_drv.fence_queue); |
488 | ||
b80d8475 | 489 | if (amdgpu_enable_scheduler) { |
2440ff2c JZ |
490 | long timeout = msecs_to_jiffies(amdgpu_lockup_timeout); |
491 | if (timeout == 0) { | |
492 | /* | |
493 | * FIXME: | |
494 | * Delayed workqueue cannot use it directly, | |
495 | * so the scheduler will not use delayed workqueue if | |
496 | * MAX_SCHEDULE_TIMEOUT is set. | |
497 | * Currently keep it simple and silly. | |
498 | */ | |
499 | timeout = MAX_SCHEDULE_TIMEOUT; | |
500 | } | |
4f839a24 | 501 | r = amd_sched_init(&ring->sched, &amdgpu_sched_ops, |
2440ff2c JZ |
502 | amdgpu_sched_hw_submission, |
503 | timeout, ring->name); | |
4f839a24 CK |
504 | if (r) { |
505 | DRM_ERROR("Failed to create scheduler on ring %s.\n", | |
506 | ring->name); | |
507 | return r; | |
508 | } | |
b80d8475 | 509 | } |
4f839a24 CK |
510 | |
511 | return 0; | |
d38ceaf9 AD |
512 | } |
513 | ||
514 | /** | |
515 | * amdgpu_fence_driver_init - init the fence driver | |
516 | * for all possible rings. | |
517 | * | |
518 | * @adev: amdgpu device pointer | |
519 | * | |
520 | * Init the fence driver for all possible rings (all asics). | |
521 | * Not all asics have all rings, so each asic will only | |
522 | * start the fence driver on the rings it has using | |
523 | * amdgpu_fence_driver_start_ring(). | |
524 | * Returns 0 for success. | |
525 | */ | |
526 | int amdgpu_fence_driver_init(struct amdgpu_device *adev) | |
527 | { | |
b49c84a5 CZ |
528 | if (atomic_inc_return(&amdgpu_fence_slab_ref) == 1) { |
529 | amdgpu_fence_slab = kmem_cache_create( | |
530 | "amdgpu_fence", sizeof(struct amdgpu_fence), 0, | |
531 | SLAB_HWCACHE_ALIGN, NULL); | |
532 | if (!amdgpu_fence_slab) | |
533 | return -ENOMEM; | |
534 | } | |
d38ceaf9 AD |
535 | if (amdgpu_debugfs_fence_init(adev)) |
536 | dev_err(adev->dev, "fence debugfs file creation failed\n"); | |
537 | ||
538 | return 0; | |
539 | } | |
540 | ||
541 | /** | |
542 | * amdgpu_fence_driver_fini - tear down the fence driver | |
543 | * for all possible rings. | |
544 | * | |
545 | * @adev: amdgpu device pointer | |
546 | * | |
547 | * Tear down the fence driver for all possible rings (all asics). | |
548 | */ | |
549 | void amdgpu_fence_driver_fini(struct amdgpu_device *adev) | |
550 | { | |
551 | int i, r; | |
552 | ||
b49c84a5 CZ |
553 | if (atomic_dec_and_test(&amdgpu_fence_slab_ref)) |
554 | kmem_cache_destroy(amdgpu_fence_slab); | |
d38ceaf9 AD |
555 | mutex_lock(&adev->ring_lock); |
556 | for (i = 0; i < AMDGPU_MAX_RINGS; i++) { | |
557 | struct amdgpu_ring *ring = adev->rings[i]; | |
c2776afe | 558 | |
d38ceaf9 AD |
559 | if (!ring || !ring->fence_drv.initialized) |
560 | continue; | |
561 | r = amdgpu_fence_wait_empty(ring); | |
562 | if (r) { | |
563 | /* no need to trigger GPU reset as we are unloading */ | |
564 | amdgpu_fence_driver_force_completion(adev); | |
565 | } | |
7f06c236 | 566 | wake_up_all(&ring->fence_drv.fence_queue); |
c6a4079b CZ |
567 | amdgpu_irq_put(adev, ring->fence_drv.irq_src, |
568 | ring->fence_drv.irq_type); | |
4f839a24 | 569 | amd_sched_fini(&ring->sched); |
c2776afe | 570 | del_timer_sync(&ring->fence_drv.fallback_timer); |
d38ceaf9 AD |
571 | ring->fence_drv.initialized = false; |
572 | } | |
573 | mutex_unlock(&adev->ring_lock); | |
574 | } | |
575 | ||
5ceb54c6 AD |
576 | /** |
577 | * amdgpu_fence_driver_suspend - suspend the fence driver | |
578 | * for all possible rings. | |
579 | * | |
580 | * @adev: amdgpu device pointer | |
581 | * | |
582 | * Suspend the fence driver for all possible rings (all asics). | |
583 | */ | |
584 | void amdgpu_fence_driver_suspend(struct amdgpu_device *adev) | |
585 | { | |
586 | int i, r; | |
587 | ||
588 | mutex_lock(&adev->ring_lock); | |
589 | for (i = 0; i < AMDGPU_MAX_RINGS; i++) { | |
590 | struct amdgpu_ring *ring = adev->rings[i]; | |
591 | if (!ring || !ring->fence_drv.initialized) | |
592 | continue; | |
593 | ||
594 | /* wait for gpu to finish processing current batch */ | |
595 | r = amdgpu_fence_wait_empty(ring); | |
596 | if (r) { | |
597 | /* delay GPU reset to resume */ | |
598 | amdgpu_fence_driver_force_completion(adev); | |
599 | } | |
600 | ||
601 | /* disable the interrupt */ | |
602 | amdgpu_irq_put(adev, ring->fence_drv.irq_src, | |
603 | ring->fence_drv.irq_type); | |
604 | } | |
605 | mutex_unlock(&adev->ring_lock); | |
606 | } | |
607 | ||
608 | /** | |
609 | * amdgpu_fence_driver_resume - resume the fence driver | |
610 | * for all possible rings. | |
611 | * | |
612 | * @adev: amdgpu device pointer | |
613 | * | |
614 | * Resume the fence driver for all possible rings (all asics). | |
615 | * Not all asics have all rings, so each asic will only | |
616 | * start the fence driver on the rings it has using | |
617 | * amdgpu_fence_driver_start_ring(). | |
618 | * Returns 0 for success. | |
619 | */ | |
620 | void amdgpu_fence_driver_resume(struct amdgpu_device *adev) | |
621 | { | |
622 | int i; | |
623 | ||
624 | mutex_lock(&adev->ring_lock); | |
625 | for (i = 0; i < AMDGPU_MAX_RINGS; i++) { | |
626 | struct amdgpu_ring *ring = adev->rings[i]; | |
627 | if (!ring || !ring->fence_drv.initialized) | |
628 | continue; | |
629 | ||
630 | /* enable the interrupt */ | |
631 | amdgpu_irq_get(adev, ring->fence_drv.irq_src, | |
632 | ring->fence_drv.irq_type); | |
633 | } | |
634 | mutex_unlock(&adev->ring_lock); | |
635 | } | |
636 | ||
d38ceaf9 AD |
637 | /** |
638 | * amdgpu_fence_driver_force_completion - force all fence waiter to complete | |
639 | * | |
640 | * @adev: amdgpu device pointer | |
641 | * | |
642 | * In case of GPU reset failure make sure no process keep waiting on fence | |
643 | * that will never complete. | |
644 | */ | |
645 | void amdgpu_fence_driver_force_completion(struct amdgpu_device *adev) | |
646 | { | |
647 | int i; | |
648 | ||
649 | for (i = 0; i < AMDGPU_MAX_RINGS; i++) { | |
650 | struct amdgpu_ring *ring = adev->rings[i]; | |
651 | if (!ring || !ring->fence_drv.initialized) | |
652 | continue; | |
653 | ||
654 | amdgpu_fence_write(ring, ring->fence_drv.sync_seq[i]); | |
655 | } | |
656 | } | |
657 | ||
a95e2642 CK |
658 | /* |
659 | * Common fence implementation | |
660 | */ | |
661 | ||
662 | static const char *amdgpu_fence_get_driver_name(struct fence *fence) | |
663 | { | |
664 | return "amdgpu"; | |
665 | } | |
666 | ||
667 | static const char *amdgpu_fence_get_timeline_name(struct fence *f) | |
668 | { | |
669 | struct amdgpu_fence *fence = to_amdgpu_fence(f); | |
670 | return (const char *)fence->ring->name; | |
671 | } | |
672 | ||
673 | /** | |
674 | * amdgpu_fence_is_signaled - test if fence is signaled | |
675 | * | |
676 | * @f: fence to test | |
677 | * | |
678 | * Test the fence sequence number if it is already signaled. If it isn't | |
679 | * signaled start fence processing. Returns True if the fence is signaled. | |
680 | */ | |
681 | static bool amdgpu_fence_is_signaled(struct fence *f) | |
682 | { | |
683 | struct amdgpu_fence *fence = to_amdgpu_fence(f); | |
684 | struct amdgpu_ring *ring = fence->ring; | |
685 | ||
686 | if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq) | |
687 | return true; | |
688 | ||
689 | amdgpu_fence_process(ring); | |
690 | ||
691 | if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq) | |
692 | return true; | |
693 | ||
694 | return false; | |
695 | } | |
696 | ||
697 | /** | |
698 | * amdgpu_fence_check_signaled - callback from fence_queue | |
699 | * | |
700 | * this function is called with fence_queue lock held, which is also used | |
701 | * for the fence locking itself, so unlocked variants are used for | |
702 | * fence_signal, and remove_wait_queue. | |
703 | */ | |
704 | static int amdgpu_fence_check_signaled(wait_queue_t *wait, unsigned mode, int flags, void *key) | |
705 | { | |
706 | struct amdgpu_fence *fence; | |
707 | struct amdgpu_device *adev; | |
708 | u64 seq; | |
709 | int ret; | |
710 | ||
711 | fence = container_of(wait, struct amdgpu_fence, fence_wake); | |
712 | adev = fence->ring->adev; | |
713 | ||
714 | /* | |
715 | * We cannot use amdgpu_fence_process here because we're already | |
716 | * in the waitqueue, in a call from wake_up_all. | |
717 | */ | |
718 | seq = atomic64_read(&fence->ring->fence_drv.last_seq); | |
719 | if (seq >= fence->seq) { | |
720 | ret = fence_signal_locked(&fence->base); | |
721 | if (!ret) | |
722 | FENCE_TRACE(&fence->base, "signaled from irq context\n"); | |
723 | else | |
724 | FENCE_TRACE(&fence->base, "was already signaled\n"); | |
725 | ||
726 | __remove_wait_queue(&fence->ring->fence_drv.fence_queue, &fence->fence_wake); | |
727 | fence_put(&fence->base); | |
728 | } else | |
729 | FENCE_TRACE(&fence->base, "pending\n"); | |
730 | return 0; | |
731 | } | |
732 | ||
733 | /** | |
734 | * amdgpu_fence_enable_signaling - enable signalling on fence | |
735 | * @fence: fence | |
736 | * | |
737 | * This function is called with fence_queue lock held, and adds a callback | |
738 | * to fence_queue that checks if this fence is signaled, and if so it | |
739 | * signals the fence and removes itself. | |
740 | */ | |
741 | static bool amdgpu_fence_enable_signaling(struct fence *f) | |
742 | { | |
743 | struct amdgpu_fence *fence = to_amdgpu_fence(f); | |
744 | struct amdgpu_ring *ring = fence->ring; | |
745 | ||
746 | if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq) | |
747 | return false; | |
748 | ||
749 | fence->fence_wake.flags = 0; | |
750 | fence->fence_wake.private = NULL; | |
751 | fence->fence_wake.func = amdgpu_fence_check_signaled; | |
752 | __add_wait_queue(&ring->fence_drv.fence_queue, &fence->fence_wake); | |
753 | fence_get(f); | |
c2776afe CK |
754 | if (!timer_pending(&ring->fence_drv.fallback_timer)) |
755 | amdgpu_fence_schedule_fallback(ring); | |
a95e2642 CK |
756 | FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx); |
757 | return true; | |
758 | } | |
759 | ||
b49c84a5 CZ |
760 | static void amdgpu_fence_release(struct fence *f) |
761 | { | |
762 | struct amdgpu_fence *fence = to_amdgpu_fence(f); | |
763 | kmem_cache_free(amdgpu_fence_slab, fence); | |
764 | } | |
765 | ||
a95e2642 CK |
766 | const struct fence_ops amdgpu_fence_ops = { |
767 | .get_driver_name = amdgpu_fence_get_driver_name, | |
768 | .get_timeline_name = amdgpu_fence_get_timeline_name, | |
769 | .enable_signaling = amdgpu_fence_enable_signaling, | |
770 | .signaled = amdgpu_fence_is_signaled, | |
771 | .wait = fence_default_wait, | |
b49c84a5 | 772 | .release = amdgpu_fence_release, |
a95e2642 | 773 | }; |
d38ceaf9 AD |
774 | |
775 | /* | |
776 | * Fence debugfs | |
777 | */ | |
778 | #if defined(CONFIG_DEBUG_FS) | |
779 | static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data) | |
780 | { | |
781 | struct drm_info_node *node = (struct drm_info_node *)m->private; | |
782 | struct drm_device *dev = node->minor->dev; | |
783 | struct amdgpu_device *adev = dev->dev_private; | |
784 | int i, j; | |
785 | ||
786 | for (i = 0; i < AMDGPU_MAX_RINGS; ++i) { | |
787 | struct amdgpu_ring *ring = adev->rings[i]; | |
788 | if (!ring || !ring->fence_drv.initialized) | |
789 | continue; | |
790 | ||
791 | amdgpu_fence_process(ring); | |
792 | ||
344c19f9 | 793 | seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name); |
d38ceaf9 AD |
794 | seq_printf(m, "Last signaled fence 0x%016llx\n", |
795 | (unsigned long long)atomic64_read(&ring->fence_drv.last_seq)); | |
796 | seq_printf(m, "Last emitted 0x%016llx\n", | |
797 | ring->fence_drv.sync_seq[i]); | |
798 | ||
799 | for (j = 0; j < AMDGPU_MAX_RINGS; ++j) { | |
800 | struct amdgpu_ring *other = adev->rings[j]; | |
344c19f9 CK |
801 | if (i != j && other && other->fence_drv.initialized && |
802 | ring->fence_drv.sync_seq[j]) | |
d38ceaf9 AD |
803 | seq_printf(m, "Last sync to ring %d 0x%016llx\n", |
804 | j, ring->fence_drv.sync_seq[j]); | |
805 | } | |
806 | } | |
807 | return 0; | |
808 | } | |
809 | ||
810 | static struct drm_info_list amdgpu_debugfs_fence_list[] = { | |
811 | {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL}, | |
812 | }; | |
813 | #endif | |
814 | ||
815 | int amdgpu_debugfs_fence_init(struct amdgpu_device *adev) | |
816 | { | |
817 | #if defined(CONFIG_DEBUG_FS) | |
818 | return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 1); | |
819 | #else | |
820 | return 0; | |
821 | #endif | |
822 | } | |
823 |