blob: 2c485a22e14b035355b84da5da5c57299207db70 [file] [log] [blame]
Alex Deucherd38ceaf2015-04-20 16:55:21 -04001/*
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
50/**
51 * amdgpu_fence_write - write a fence value
52 *
53 * @ring: ring the fence is associated with
54 * @seq: sequence number to write
55 *
56 * Writes a fence value to memory (all asics).
57 */
58static void amdgpu_fence_write(struct amdgpu_ring *ring, u32 seq)
59{
60 struct amdgpu_fence_driver *drv = &ring->fence_drv;
61
62 if (drv->cpu_addr)
63 *drv->cpu_addr = cpu_to_le32(seq);
64}
65
66/**
67 * amdgpu_fence_read - read a fence value
68 *
69 * @ring: ring the fence is associated with
70 *
71 * Reads a fence value from memory (all asics).
72 * Returns the value of the fence read from memory.
73 */
74static u32 amdgpu_fence_read(struct amdgpu_ring *ring)
75{
76 struct amdgpu_fence_driver *drv = &ring->fence_drv;
77 u32 seq = 0;
78
79 if (drv->cpu_addr)
80 seq = le32_to_cpu(*drv->cpu_addr);
81 else
82 seq = lower_32_bits(atomic64_read(&drv->last_seq));
83
84 return seq;
85}
86
87/**
88 * amdgpu_fence_schedule_check - schedule lockup check
89 *
90 * @ring: pointer to struct amdgpu_ring
91 *
92 * Queues a delayed work item to check for lockups.
93 */
94static void amdgpu_fence_schedule_check(struct amdgpu_ring *ring)
95{
96 /*
97 * Do not reset the timer here with mod_delayed_work,
98 * this can livelock in an interaction with TTM delayed destroy.
99 */
100 queue_delayed_work(system_power_efficient_wq,
101 &ring->fence_drv.lockup_work,
102 AMDGPU_FENCE_JIFFIES_TIMEOUT);
103}
104
105/**
106 * amdgpu_fence_emit - emit a fence on the requested ring
107 *
108 * @ring: ring the fence is associated with
109 * @owner: creator of the fence
110 * @fence: amdgpu fence object
111 *
112 * Emits a fence command on the requested ring (all asics).
113 * Returns 0 on success, -ENOMEM on failure.
114 */
115int amdgpu_fence_emit(struct amdgpu_ring *ring, void *owner,
116 struct amdgpu_fence **fence)
117{
118 struct amdgpu_device *adev = ring->adev;
119
120 /* we are protected by the ring emission mutex */
121 *fence = kmalloc(sizeof(struct amdgpu_fence), GFP_KERNEL);
122 if ((*fence) == NULL) {
123 return -ENOMEM;
124 }
125 (*fence)->seq = ++ring->fence_drv.sync_seq[ring->idx];
126 (*fence)->ring = ring;
127 (*fence)->owner = owner;
128 fence_init(&(*fence)->base, &amdgpu_fence_ops,
monk.liu7f06c232015-07-30 18:28:12 +0800129 &ring->fence_drv.fence_queue.lock,
130 adev->fence_context + ring->idx,
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400131 (*fence)->seq);
Chunming Zhou890ee232015-06-01 14:35:03 +0800132 amdgpu_ring_emit_fence(ring, ring->fence_drv.gpu_addr,
133 (*fence)->seq,
134 AMDGPU_FENCE_FLAG_INT);
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400135 trace_amdgpu_fence_emit(ring->adev->ddev, ring->idx, (*fence)->seq);
136 return 0;
137}
138
139/**
140 * amdgpu_fence_check_signaled - callback from fence_queue
141 *
142 * this function is called with fence_queue lock held, which is also used
143 * for the fence locking itself, so unlocked variants are used for
144 * fence_signal, and remove_wait_queue.
145 */
146static int amdgpu_fence_check_signaled(wait_queue_t *wait, unsigned mode, int flags, void *key)
147{
148 struct amdgpu_fence *fence;
149 struct amdgpu_device *adev;
150 u64 seq;
151 int ret;
152
153 fence = container_of(wait, struct amdgpu_fence, fence_wake);
154 adev = fence->ring->adev;
155
156 /*
157 * We cannot use amdgpu_fence_process here because we're already
158 * in the waitqueue, in a call from wake_up_all.
159 */
160 seq = atomic64_read(&fence->ring->fence_drv.last_seq);
161 if (seq >= fence->seq) {
162 ret = fence_signal_locked(&fence->base);
163 if (!ret)
164 FENCE_TRACE(&fence->base, "signaled from irq context\n");
165 else
166 FENCE_TRACE(&fence->base, "was already signaled\n");
167
monk.liu7f06c232015-07-30 18:28:12 +0800168 __remove_wait_queue(&fence->ring->fence_drv.fence_queue, &fence->fence_wake);
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400169 fence_put(&fence->base);
170 } else
171 FENCE_TRACE(&fence->base, "pending\n");
172 return 0;
173}
174
175/**
176 * amdgpu_fence_activity - check for fence activity
177 *
178 * @ring: pointer to struct amdgpu_ring
179 *
180 * Checks the current fence value and calculates the last
181 * signalled fence value. Returns true if activity occured
182 * on the ring, and the fence_queue should be waken up.
183 */
184static bool amdgpu_fence_activity(struct amdgpu_ring *ring)
185{
186 uint64_t seq, last_seq, last_emitted;
187 unsigned count_loop = 0;
188 bool wake = false;
189
190 /* Note there is a scenario here for an infinite loop but it's
191 * very unlikely to happen. For it to happen, the current polling
192 * process need to be interrupted by another process and another
193 * process needs to update the last_seq btw the atomic read and
194 * xchg of the current process.
195 *
196 * More over for this to go in infinite loop there need to be
Jammy Zhou86c2b792015-05-13 22:52:42 +0800197 * continuously new fence signaled ie amdgpu_fence_read needs
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400198 * to return a different value each time for both the currently
199 * polling process and the other process that xchg the last_seq
200 * btw atomic read and xchg of the current process. And the
201 * value the other process set as last seq must be higher than
202 * the seq value we just read. Which means that current process
Jammy Zhou86c2b792015-05-13 22:52:42 +0800203 * need to be interrupted after amdgpu_fence_read and before
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400204 * atomic xchg.
205 *
206 * To be even more safe we count the number of time we loop and
207 * we bail after 10 loop just accepting the fact that we might
208 * have temporarly set the last_seq not to the true real last
209 * seq but to an older one.
210 */
211 last_seq = atomic64_read(&ring->fence_drv.last_seq);
212 do {
213 last_emitted = ring->fence_drv.sync_seq[ring->idx];
214 seq = amdgpu_fence_read(ring);
215 seq |= last_seq & 0xffffffff00000000LL;
216 if (seq < last_seq) {
217 seq &= 0xffffffff;
218 seq |= last_emitted & 0xffffffff00000000LL;
219 }
220
221 if (seq <= last_seq || seq > last_emitted) {
222 break;
223 }
224 /* If we loop over we don't want to return without
225 * checking if a fence is signaled as it means that the
226 * seq we just read is different from the previous on.
227 */
228 wake = true;
229 last_seq = seq;
230 if ((count_loop++) > 10) {
231 /* We looped over too many time leave with the
232 * fact that we might have set an older fence
233 * seq then the current real last seq as signaled
234 * by the hw.
235 */
236 break;
237 }
238 } while (atomic64_xchg(&ring->fence_drv.last_seq, seq) > seq);
239
240 if (seq < last_emitted)
241 amdgpu_fence_schedule_check(ring);
242
243 return wake;
244}
245
246/**
247 * amdgpu_fence_check_lockup - check for hardware lockup
248 *
249 * @work: delayed work item
250 *
251 * Checks for fence activity and if there is none probe
252 * the hardware if a lockup occured.
253 */
254static void amdgpu_fence_check_lockup(struct work_struct *work)
255{
256 struct amdgpu_fence_driver *fence_drv;
257 struct amdgpu_ring *ring;
258
259 fence_drv = container_of(work, struct amdgpu_fence_driver,
260 lockup_work.work);
261 ring = fence_drv->ring;
262
Christian König0c418f12015-09-01 15:13:53 +0200263 if (amdgpu_fence_activity(ring))
monk.liu7f06c232015-07-30 18:28:12 +0800264 wake_up_all(&ring->fence_drv.fence_queue);
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400265}
266
267/**
268 * amdgpu_fence_process - process a fence
269 *
270 * @adev: amdgpu_device pointer
271 * @ring: ring index the fence is associated with
272 *
273 * Checks the current fence value and wakes the fence queue
274 * if the sequence number has increased (all asics).
275 */
276void amdgpu_fence_process(struct amdgpu_ring *ring)
277{
Christian König68ed3de2015-08-07 15:57:21 +0200278 if (amdgpu_fence_activity(ring))
monk.liu7f06c232015-07-30 18:28:12 +0800279 wake_up_all(&ring->fence_drv.fence_queue);
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400280}
281
282/**
283 * amdgpu_fence_seq_signaled - check if a fence sequence number has signaled
284 *
285 * @ring: ring the fence is associated with
286 * @seq: sequence number
287 *
288 * Check if the last signaled fence sequnce number is >= the requested
289 * sequence number (all asics).
290 * Returns true if the fence has signaled (current fence value
291 * is >= requested value) or false if it has not (current fence
292 * value is < the requested value. Helper function for
293 * amdgpu_fence_signaled().
294 */
295static bool amdgpu_fence_seq_signaled(struct amdgpu_ring *ring, u64 seq)
296{
297 if (atomic64_read(&ring->fence_drv.last_seq) >= seq)
298 return true;
299
300 /* poll new last sequence at least once */
301 amdgpu_fence_process(ring);
302 if (atomic64_read(&ring->fence_drv.last_seq) >= seq)
303 return true;
304
305 return false;
306}
307
308static bool amdgpu_fence_is_signaled(struct fence *f)
309{
310 struct amdgpu_fence *fence = to_amdgpu_fence(f);
311 struct amdgpu_ring *ring = fence->ring;
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400312
313 if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq)
314 return true;
315
Christian König0c418f12015-09-01 15:13:53 +0200316 amdgpu_fence_process(ring);
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400317
Christian König0c418f12015-09-01 15:13:53 +0200318 if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq)
319 return true;
320
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400321 return false;
322}
323
324/**
325 * amdgpu_fence_enable_signaling - enable signalling on fence
326 * @fence: fence
327 *
328 * This function is called with fence_queue lock held, and adds a callback
329 * to fence_queue that checks if this fence is signaled, and if so it
330 * signals the fence and removes itself.
331 */
332static bool amdgpu_fence_enable_signaling(struct fence *f)
333{
334 struct amdgpu_fence *fence = to_amdgpu_fence(f);
335 struct amdgpu_ring *ring = fence->ring;
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400336
337 if (atomic64_read(&ring->fence_drv.last_seq) >= fence->seq)
338 return false;
339
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400340 fence->fence_wake.flags = 0;
341 fence->fence_wake.private = NULL;
342 fence->fence_wake.func = amdgpu_fence_check_signaled;
monk.liu7f06c232015-07-30 18:28:12 +0800343 __add_wait_queue(&ring->fence_drv.fence_queue, &fence->fence_wake);
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400344 fence_get(f);
Christian König243724472015-11-03 11:26:42 +0100345 amdgpu_fence_schedule_check(ring);
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400346 FENCE_TRACE(&fence->base, "armed on ring %i!\n", ring->idx);
347 return true;
348}
349
monk.liu7f06c232015-07-30 18:28:12 +0800350/*
351 * amdgpu_ring_wait_seq_timeout - wait for seq of the specific ring to signal
352 * @ring: ring to wait on for the seq number
353 * @seq: seq number wait for
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400354 *
monk.liu7f06c232015-07-30 18:28:12 +0800355 * return value:
Christian König00d2a2b2015-08-07 16:15:36 +0200356 * 0: seq signaled, and gpu not hang
357 * -EDEADL: GPU hang detected
monk.liu7f06c232015-07-30 18:28:12 +0800358 * -EINVAL: some paramter is not valid
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400359 */
Christian König00d2a2b2015-08-07 16:15:36 +0200360static int amdgpu_fence_ring_wait_seq(struct amdgpu_ring *ring, uint64_t seq)
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400361{
monk.liu7f06c232015-07-30 18:28:12 +0800362 bool signaled = false;
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400363
monk.liu7f06c232015-07-30 18:28:12 +0800364 BUG_ON(!ring);
365 if (seq > ring->fence_drv.sync_seq[ring->idx])
366 return -EINVAL;
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400367
monk.liu7f06c232015-07-30 18:28:12 +0800368 if (atomic64_read(&ring->fence_drv.last_seq) >= seq)
Christian König00d2a2b2015-08-07 16:15:36 +0200369 return 0;
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400370
Christian König243724472015-11-03 11:26:42 +0100371 amdgpu_fence_schedule_check(ring);
Christian König00d2a2b2015-08-07 16:15:36 +0200372 wait_event(ring->fence_drv.fence_queue, (
Christian Königb7e4dad2015-09-01 10:50:26 +0200373 (signaled = amdgpu_fence_seq_signaled(ring, seq))));
monk.liu7f06c232015-07-30 18:28:12 +0800374
Christian König00d2a2b2015-08-07 16:15:36 +0200375 if (signaled)
376 return 0;
377 else
378 return -EDEADLK;
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400379}
380
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400381/**
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400382 * amdgpu_fence_wait_next - wait for the next fence to signal
383 *
384 * @adev: amdgpu device pointer
385 * @ring: ring index the fence is associated with
386 *
387 * Wait for the next fence on the requested ring to signal (all asics).
388 * Returns 0 if the next fence has passed, error for all other cases.
389 * Caller must hold ring lock.
390 */
391int amdgpu_fence_wait_next(struct amdgpu_ring *ring)
392{
monk.liu7f06c232015-07-30 18:28:12 +0800393 uint64_t seq = atomic64_read(&ring->fence_drv.last_seq) + 1ULL;
Christian König00d2a2b2015-08-07 16:15:36 +0200394
monk.liu7f06c232015-07-30 18:28:12 +0800395 if (seq >= ring->fence_drv.sync_seq[ring->idx])
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400396 return -ENOENT;
monk.liu7f06c232015-07-30 18:28:12 +0800397
Christian König00d2a2b2015-08-07 16:15:36 +0200398 return amdgpu_fence_ring_wait_seq(ring, seq);
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400399}
400
401/**
402 * amdgpu_fence_wait_empty - wait for all fences to signal
403 *
404 * @adev: amdgpu device pointer
405 * @ring: ring index the fence is associated with
406 *
407 * Wait for all fences on the requested ring to signal (all asics).
408 * Returns 0 if the fences have passed, error for all other cases.
409 * Caller must hold ring lock.
410 */
411int amdgpu_fence_wait_empty(struct amdgpu_ring *ring)
412{
monk.liu7f06c232015-07-30 18:28:12 +0800413 uint64_t seq = ring->fence_drv.sync_seq[ring->idx];
Christian König00d2a2b2015-08-07 16:15:36 +0200414
monk.liu7f06c232015-07-30 18:28:12 +0800415 if (!seq)
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400416 return 0;
417
Christian König00d2a2b2015-08-07 16:15:36 +0200418 return amdgpu_fence_ring_wait_seq(ring, seq);
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400419}
420
421/**
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400422 * amdgpu_fence_count_emitted - get the count of emitted fences
423 *
424 * @ring: ring the fence is associated with
425 *
426 * Get the number of fences emitted on the requested ring (all asics).
427 * Returns the number of emitted fences on the ring. Used by the
428 * dynpm code to ring track activity.
429 */
430unsigned amdgpu_fence_count_emitted(struct amdgpu_ring *ring)
431{
432 uint64_t emitted;
433
434 /* We are not protected by ring lock when reading the last sequence
435 * but it's ok to report slightly wrong fence count here.
436 */
437 amdgpu_fence_process(ring);
438 emitted = ring->fence_drv.sync_seq[ring->idx]
439 - atomic64_read(&ring->fence_drv.last_seq);
440 /* to avoid 32bits warp around */
441 if (emitted > 0x10000000)
442 emitted = 0x10000000;
443
444 return (unsigned)emitted;
445}
446
447/**
448 * amdgpu_fence_need_sync - do we need a semaphore
449 *
450 * @fence: amdgpu fence object
451 * @dst_ring: which ring to check against
452 *
453 * Check if the fence needs to be synced against another ring
454 * (all asics). If so, we need to emit a semaphore.
455 * Returns true if we need to sync with another ring, false if
456 * not.
457 */
458bool amdgpu_fence_need_sync(struct amdgpu_fence *fence,
459 struct amdgpu_ring *dst_ring)
460{
461 struct amdgpu_fence_driver *fdrv;
462
463 if (!fence)
464 return false;
465
466 if (fence->ring == dst_ring)
467 return false;
468
469 /* we are protected by the ring mutex */
470 fdrv = &dst_ring->fence_drv;
471 if (fence->seq <= fdrv->sync_seq[fence->ring->idx])
472 return false;
473
474 return true;
475}
476
477/**
478 * amdgpu_fence_note_sync - record the sync point
479 *
480 * @fence: amdgpu fence object
481 * @dst_ring: which ring to check against
482 *
483 * Note the sequence number at which point the fence will
484 * be synced with the requested ring (all asics).
485 */
486void amdgpu_fence_note_sync(struct amdgpu_fence *fence,
487 struct amdgpu_ring *dst_ring)
488{
489 struct amdgpu_fence_driver *dst, *src;
490 unsigned i;
491
492 if (!fence)
493 return;
494
495 if (fence->ring == dst_ring)
496 return;
497
498 /* we are protected by the ring mutex */
499 src = &fence->ring->fence_drv;
500 dst = &dst_ring->fence_drv;
501 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
502 if (i == dst_ring->idx)
503 continue;
504
505 dst->sync_seq[i] = max(dst->sync_seq[i], src->sync_seq[i]);
506 }
507}
508
509/**
510 * amdgpu_fence_driver_start_ring - make the fence driver
511 * ready for use on the requested ring.
512 *
513 * @ring: ring to start the fence driver on
514 * @irq_src: interrupt source to use for this ring
515 * @irq_type: interrupt type to use for this ring
516 *
517 * Make the fence driver ready for processing (all asics).
518 * Not all asics have all rings, so each asic will only
519 * start the fence driver on the rings it has.
520 * Returns 0 for success, errors for failure.
521 */
522int amdgpu_fence_driver_start_ring(struct amdgpu_ring *ring,
523 struct amdgpu_irq_src *irq_src,
524 unsigned irq_type)
525{
526 struct amdgpu_device *adev = ring->adev;
527 uint64_t index;
528
529 if (ring != &adev->uvd.ring) {
530 ring->fence_drv.cpu_addr = &adev->wb.wb[ring->fence_offs];
531 ring->fence_drv.gpu_addr = adev->wb.gpu_addr + (ring->fence_offs * 4);
532 } else {
533 /* put fence directly behind firmware */
534 index = ALIGN(adev->uvd.fw->size, 8);
535 ring->fence_drv.cpu_addr = adev->uvd.cpu_addr + index;
536 ring->fence_drv.gpu_addr = adev->uvd.gpu_addr + index;
537 }
538 amdgpu_fence_write(ring, atomic64_read(&ring->fence_drv.last_seq));
Chunming Zhouc6a40792015-06-01 14:14:32 +0800539 amdgpu_irq_get(adev, irq_src, irq_type);
540
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400541 ring->fence_drv.irq_src = irq_src;
542 ring->fence_drv.irq_type = irq_type;
Chunming Zhouc6a40792015-06-01 14:14:32 +0800543 ring->fence_drv.initialized = true;
544
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400545 dev_info(adev->dev, "fence driver on ring %d use gpu addr 0x%016llx, "
546 "cpu addr 0x%p\n", ring->idx,
547 ring->fence_drv.gpu_addr, ring->fence_drv.cpu_addr);
548 return 0;
549}
550
551/**
552 * amdgpu_fence_driver_init_ring - init the fence driver
553 * for the requested ring.
554 *
555 * @ring: ring to init the fence driver on
556 *
557 * Init the fence driver for the requested ring (all asics).
558 * Helper function for amdgpu_fence_driver_init().
559 */
Christian König4f839a22015-09-08 20:22:31 +0200560int amdgpu_fence_driver_init_ring(struct amdgpu_ring *ring)
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400561{
Christian König4f839a22015-09-08 20:22:31 +0200562 int i, r;
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400563
564 ring->fence_drv.cpu_addr = NULL;
565 ring->fence_drv.gpu_addr = 0;
566 for (i = 0; i < AMDGPU_MAX_RINGS; ++i)
567 ring->fence_drv.sync_seq[i] = 0;
568
569 atomic64_set(&ring->fence_drv.last_seq, 0);
570 ring->fence_drv.initialized = false;
571
572 INIT_DELAYED_WORK(&ring->fence_drv.lockup_work,
573 amdgpu_fence_check_lockup);
574 ring->fence_drv.ring = ring;
Alex Deucherb80d8472015-08-16 22:55:02 -0400575
Christian König5ec92a72015-09-07 18:43:02 +0200576 init_waitqueue_head(&ring->fence_drv.fence_queue);
577
Alex Deucherb80d8472015-08-16 22:55:02 -0400578 if (amdgpu_enable_scheduler) {
Junwei Zhang2440ff22015-10-10 08:48:42 +0800579 long timeout = msecs_to_jiffies(amdgpu_lockup_timeout);
580 if (timeout == 0) {
581 /*
582 * FIXME:
583 * Delayed workqueue cannot use it directly,
584 * so the scheduler will not use delayed workqueue if
585 * MAX_SCHEDULE_TIMEOUT is set.
586 * Currently keep it simple and silly.
587 */
588 timeout = MAX_SCHEDULE_TIMEOUT;
589 }
Christian König4f839a22015-09-08 20:22:31 +0200590 r = amd_sched_init(&ring->sched, &amdgpu_sched_ops,
Junwei Zhang2440ff22015-10-10 08:48:42 +0800591 amdgpu_sched_hw_submission,
592 timeout, ring->name);
Christian König4f839a22015-09-08 20:22:31 +0200593 if (r) {
594 DRM_ERROR("Failed to create scheduler on ring %s.\n",
595 ring->name);
596 return r;
597 }
Alex Deucherb80d8472015-08-16 22:55:02 -0400598 }
Christian König4f839a22015-09-08 20:22:31 +0200599
600 return 0;
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400601}
602
603/**
604 * amdgpu_fence_driver_init - init the fence driver
605 * for all possible rings.
606 *
607 * @adev: amdgpu device pointer
608 *
609 * Init the fence driver for all possible rings (all asics).
610 * Not all asics have all rings, so each asic will only
611 * start the fence driver on the rings it has using
612 * amdgpu_fence_driver_start_ring().
613 * Returns 0 for success.
614 */
615int amdgpu_fence_driver_init(struct amdgpu_device *adev)
616{
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400617 if (amdgpu_debugfs_fence_init(adev))
618 dev_err(adev->dev, "fence debugfs file creation failed\n");
619
620 return 0;
621}
622
623/**
624 * amdgpu_fence_driver_fini - tear down the fence driver
625 * for all possible rings.
626 *
627 * @adev: amdgpu device pointer
628 *
629 * Tear down the fence driver for all possible rings (all asics).
630 */
631void amdgpu_fence_driver_fini(struct amdgpu_device *adev)
632{
633 int i, r;
634
635 mutex_lock(&adev->ring_lock);
636 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
637 struct amdgpu_ring *ring = adev->rings[i];
638 if (!ring || !ring->fence_drv.initialized)
639 continue;
640 r = amdgpu_fence_wait_empty(ring);
641 if (r) {
642 /* no need to trigger GPU reset as we are unloading */
643 amdgpu_fence_driver_force_completion(adev);
644 }
monk.liu7f06c232015-07-30 18:28:12 +0800645 wake_up_all(&ring->fence_drv.fence_queue);
Chunming Zhouc6a40792015-06-01 14:14:32 +0800646 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
647 ring->fence_drv.irq_type);
Christian König4f839a22015-09-08 20:22:31 +0200648 amd_sched_fini(&ring->sched);
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400649 ring->fence_drv.initialized = false;
650 }
651 mutex_unlock(&adev->ring_lock);
652}
653
654/**
Alex Deucher5ceb54c2015-08-05 12:41:48 -0400655 * amdgpu_fence_driver_suspend - suspend the fence driver
656 * for all possible rings.
657 *
658 * @adev: amdgpu device pointer
659 *
660 * Suspend the fence driver for all possible rings (all asics).
661 */
662void amdgpu_fence_driver_suspend(struct amdgpu_device *adev)
663{
664 int i, r;
665
666 mutex_lock(&adev->ring_lock);
667 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
668 struct amdgpu_ring *ring = adev->rings[i];
669 if (!ring || !ring->fence_drv.initialized)
670 continue;
671
672 /* wait for gpu to finish processing current batch */
673 r = amdgpu_fence_wait_empty(ring);
674 if (r) {
675 /* delay GPU reset to resume */
676 amdgpu_fence_driver_force_completion(adev);
677 }
678
679 /* disable the interrupt */
680 amdgpu_irq_put(adev, ring->fence_drv.irq_src,
681 ring->fence_drv.irq_type);
682 }
683 mutex_unlock(&adev->ring_lock);
684}
685
686/**
687 * amdgpu_fence_driver_resume - resume the fence driver
688 * for all possible rings.
689 *
690 * @adev: amdgpu device pointer
691 *
692 * Resume the fence driver for all possible rings (all asics).
693 * Not all asics have all rings, so each asic will only
694 * start the fence driver on the rings it has using
695 * amdgpu_fence_driver_start_ring().
696 * Returns 0 for success.
697 */
698void amdgpu_fence_driver_resume(struct amdgpu_device *adev)
699{
700 int i;
701
702 mutex_lock(&adev->ring_lock);
703 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
704 struct amdgpu_ring *ring = adev->rings[i];
705 if (!ring || !ring->fence_drv.initialized)
706 continue;
707
708 /* enable the interrupt */
709 amdgpu_irq_get(adev, ring->fence_drv.irq_src,
710 ring->fence_drv.irq_type);
711 }
712 mutex_unlock(&adev->ring_lock);
713}
714
715/**
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400716 * amdgpu_fence_driver_force_completion - force all fence waiter to complete
717 *
718 * @adev: amdgpu device pointer
719 *
720 * In case of GPU reset failure make sure no process keep waiting on fence
721 * that will never complete.
722 */
723void amdgpu_fence_driver_force_completion(struct amdgpu_device *adev)
724{
725 int i;
726
727 for (i = 0; i < AMDGPU_MAX_RINGS; i++) {
728 struct amdgpu_ring *ring = adev->rings[i];
729 if (!ring || !ring->fence_drv.initialized)
730 continue;
731
732 amdgpu_fence_write(ring, ring->fence_drv.sync_seq[i]);
733 }
734}
735
736
737/*
738 * Fence debugfs
739 */
740#if defined(CONFIG_DEBUG_FS)
741static int amdgpu_debugfs_fence_info(struct seq_file *m, void *data)
742{
743 struct drm_info_node *node = (struct drm_info_node *)m->private;
744 struct drm_device *dev = node->minor->dev;
745 struct amdgpu_device *adev = dev->dev_private;
746 int i, j;
747
748 for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
749 struct amdgpu_ring *ring = adev->rings[i];
750 if (!ring || !ring->fence_drv.initialized)
751 continue;
752
753 amdgpu_fence_process(ring);
754
Christian König344c19f2015-06-02 15:47:16 +0200755 seq_printf(m, "--- ring %d (%s) ---\n", i, ring->name);
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400756 seq_printf(m, "Last signaled fence 0x%016llx\n",
757 (unsigned long long)atomic64_read(&ring->fence_drv.last_seq));
758 seq_printf(m, "Last emitted 0x%016llx\n",
759 ring->fence_drv.sync_seq[i]);
760
761 for (j = 0; j < AMDGPU_MAX_RINGS; ++j) {
762 struct amdgpu_ring *other = adev->rings[j];
Christian König344c19f2015-06-02 15:47:16 +0200763 if (i != j && other && other->fence_drv.initialized &&
764 ring->fence_drv.sync_seq[j])
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400765 seq_printf(m, "Last sync to ring %d 0x%016llx\n",
766 j, ring->fence_drv.sync_seq[j]);
767 }
768 }
769 return 0;
770}
771
772static struct drm_info_list amdgpu_debugfs_fence_list[] = {
773 {"amdgpu_fence_info", &amdgpu_debugfs_fence_info, 0, NULL},
774};
775#endif
776
777int amdgpu_debugfs_fence_init(struct amdgpu_device *adev)
778{
779#if defined(CONFIG_DEBUG_FS)
780 return amdgpu_debugfs_add_files(adev, amdgpu_debugfs_fence_list, 1);
781#else
782 return 0;
783#endif
784}
785
786static const char *amdgpu_fence_get_driver_name(struct fence *fence)
787{
788 return "amdgpu";
789}
790
791static const char *amdgpu_fence_get_timeline_name(struct fence *f)
792{
793 struct amdgpu_fence *fence = to_amdgpu_fence(f);
794 return (const char *)fence->ring->name;
795}
796
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400797const struct fence_ops amdgpu_fence_ops = {
798 .get_driver_name = amdgpu_fence_get_driver_name,
799 .get_timeline_name = amdgpu_fence_get_timeline_name,
800 .enable_signaling = amdgpu_fence_enable_signaling,
801 .signaled = amdgpu_fence_is_signaled,
Christian König318cd342015-10-15 17:58:09 +0200802 .wait = fence_default_wait,
Alex Deucherd38ceaf2015-04-20 16:55:21 -0400803 .release = NULL,
804};