Rev 3031 | Rev 3764 | Go to most recent revision | Details | Compare with Previous | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
1125 | serge | 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 |
||
29 | * Dave Airlie |
||
30 | */ |
||
31 | #include |
||
32 | #include |
||
3192 | Serge | 33 | #include |
1125 | serge | 34 | #include |
35 | #include |
||
1986 | serge | 36 | #include |
2997 | Serge | 37 | #include |
1125 | serge | 38 | #include "radeon_reg.h" |
39 | #include "radeon.h" |
||
40 | |||
2997 | Serge | 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. Whether |
||
48 | * we use a scratch register or memory location depends on the asic |
||
49 | * and whether writeback is enabled. |
||
50 | */ |
||
51 | |||
52 | /** |
||
53 | * radeon_fence_write - write a fence value |
||
54 | * |
||
55 | * @rdev: radeon_device pointer |
||
56 | * @seq: sequence number to write |
||
57 | * @ring: ring index the fence is associated with |
||
58 | * |
||
59 | * Writes a fence value to memory or a scratch register (all asics). |
||
60 | */ |
||
61 | static void radeon_fence_write(struct radeon_device *rdev, u32 seq, int ring) |
||
2004 | serge | 62 | { |
2997 | Serge | 63 | struct radeon_fence_driver *drv = &rdev->fence_drv[ring]; |
64 | if (likely(rdev->wb.enabled || !drv->scratch_reg)) { |
||
65 | *drv->cpu_addr = cpu_to_le32(seq); |
||
66 | } else { |
||
67 | WREG32(drv->scratch_reg, seq); |
||
68 | } |
||
2004 | serge | 69 | } |
70 | |||
2997 | Serge | 71 | /** |
72 | * radeon_fence_read - read a fence value |
||
73 | * |
||
74 | * @rdev: radeon_device pointer |
||
75 | * @ring: ring index the fence is associated with |
||
76 | * |
||
77 | * Reads a fence value from memory or a scratch register (all asics). |
||
78 | * Returns the value of the fence read from memory or register. |
||
79 | */ |
||
80 | static u32 radeon_fence_read(struct radeon_device *rdev, int ring) |
||
2004 | serge | 81 | { |
2997 | Serge | 82 | struct radeon_fence_driver *drv = &rdev->fence_drv[ring]; |
83 | u32 seq = 0; |
||
2004 | serge | 84 | |
2997 | Serge | 85 | if (likely(rdev->wb.enabled || !drv->scratch_reg)) { |
86 | seq = le32_to_cpu(*drv->cpu_addr); |
||
87 | } else { |
||
88 | seq = RREG32(drv->scratch_reg); |
||
89 | } |
||
2004 | serge | 90 | return seq; |
91 | } |
||
92 | |||
2997 | Serge | 93 | /** |
94 | * radeon_fence_emit - emit a fence on the requested ring |
||
95 | * |
||
96 | * @rdev: radeon_device pointer |
||
97 | * @fence: radeon fence object |
||
98 | * @ring: ring index the fence is associated with |
||
99 | * |
||
100 | * Emits a fence command on the requested ring (all asics). |
||
101 | * Returns 0 on success, -ENOMEM on failure. |
||
102 | */ |
||
103 | int radeon_fence_emit(struct radeon_device *rdev, |
||
104 | struct radeon_fence **fence, |
||
105 | int ring) |
||
1125 | serge | 106 | { |
2997 | Serge | 107 | /* we are protected by the ring emission mutex */ |
108 | *fence = kmalloc(sizeof(struct radeon_fence), GFP_KERNEL); |
||
109 | if ((*fence) == NULL) { |
||
110 | return -ENOMEM; |
||
1125 | serge | 111 | } |
2997 | Serge | 112 | kref_init(&((*fence)->kref)); |
113 | (*fence)->rdev = rdev; |
||
114 | (*fence)->seq = ++rdev->fence_drv[ring].sync_seq[ring]; |
||
115 | (*fence)->ring = ring; |
||
116 | radeon_fence_ring_emit(rdev, ring, *fence); |
||
117 | // trace_radeon_fence_emit(rdev->ddev, (*fence)->seq); |
||
1125 | serge | 118 | return 0; |
119 | } |
||
120 | |||
2997 | Serge | 121 | /** |
122 | * radeon_fence_process - process a fence |
||
123 | * |
||
124 | * @rdev: radeon_device pointer |
||
125 | * @ring: ring index the fence is associated with |
||
126 | * |
||
127 | * Checks the current fence value and wakes the fence queue |
||
128 | * if the sequence number has increased (all asics). |
||
129 | */ |
||
130 | void radeon_fence_process(struct radeon_device *rdev, int ring) |
||
1125 | serge | 131 | { |
2997 | Serge | 132 | uint64_t seq, last_seq, last_emitted; |
133 | unsigned count_loop = 0; |
||
1125 | serge | 134 | bool wake = false; |
135 | |||
2997 | Serge | 136 | /* Note there is a scenario here for an infinite loop but it's |
137 | * very unlikely to happen. For it to happen, the current polling |
||
138 | * process need to be interrupted by another process and another |
||
139 | * process needs to update the last_seq btw the atomic read and |
||
140 | * xchg of the current process. |
||
141 | * |
||
142 | * More over for this to go in infinite loop there need to be |
||
143 | * continuously new fence signaled ie radeon_fence_read needs |
||
144 | * to return a different value each time for both the currently |
||
145 | * polling process and the other process that xchg the last_seq |
||
146 | * btw atomic read and xchg of the current process. And the |
||
147 | * value the other process set as last seq must be higher than |
||
148 | * the seq value we just read. Which means that current process |
||
149 | * need to be interrupted after radeon_fence_read and before |
||
150 | * atomic xchg. |
||
151 | * |
||
152 | * To be even more safe we count the number of time we loop and |
||
153 | * we bail after 10 loop just accepting the fact that we might |
||
154 | * have temporarly set the last_seq not to the true real last |
||
155 | * seq but to an older one. |
||
156 | */ |
||
157 | last_seq = atomic64_read(&rdev->fence_drv[ring].last_seq); |
||
158 | do { |
||
159 | last_emitted = rdev->fence_drv[ring].sync_seq[ring]; |
||
160 | seq = radeon_fence_read(rdev, ring); |
||
161 | seq |= last_seq & 0xffffffff00000000LL; |
||
162 | if (seq < last_seq) { |
||
163 | seq &= 0xffffffff; |
||
164 | seq |= last_emitted & 0xffffffff00000000LL; |
||
165 | } |
||
166 | |||
167 | if (seq <= last_seq || seq > last_emitted) { |
||
168 | break; |
||
169 | } |
||
170 | /* If we loop over we don't want to return without |
||
171 | * checking if a fence is signaled as it means that the |
||
172 | * seq we just read is different from the previous on. |
||
173 | */ |
||
174 | wake = true; |
||
175 | last_seq = seq; |
||
176 | if ((count_loop++) > 10) { |
||
177 | /* We looped over too many time leave with the |
||
178 | * fact that we might have set an older fence |
||
179 | * seq then the current real last seq as signaled |
||
180 | * by the hw. |
||
1963 | serge | 181 | */ |
1125 | serge | 182 | break; |
183 | } |
||
2997 | Serge | 184 | } while (atomic64_xchg(&rdev->fence_drv[ring].last_seq, seq) > seq); |
185 | |||
186 | if (wake) { |
||
187 | rdev->fence_drv[ring].last_activity = GetTimerTicks(); |
||
188 | wake_up_all(&rdev->fence_queue); |
||
1125 | serge | 189 | } |
2997 | Serge | 190 | } |
191 | |||
192 | /** |
||
193 | * radeon_fence_destroy - destroy a fence |
||
194 | * |
||
195 | * @kref: fence kref |
||
196 | * |
||
197 | * Frees the fence object (all asics). |
||
198 | */ |
||
199 | static void radeon_fence_destroy(struct kref *kref) |
||
200 | { |
||
201 | struct radeon_fence *fence; |
||
202 | |||
203 | fence = container_of(kref, struct radeon_fence, kref); |
||
204 | kfree(fence); |
||
205 | } |
||
206 | |||
207 | /** |
||
208 | * radeon_fence_seq_signaled - check if a fence sequeuce number has signaled |
||
209 | * |
||
210 | * @rdev: radeon device pointer |
||
211 | * @seq: sequence number |
||
212 | * @ring: ring index the fence is associated with |
||
213 | * |
||
214 | * Check if the last singled fence sequnce number is >= the requested |
||
215 | * sequence number (all asics). |
||
216 | * Returns true if the fence has signaled (current fence value |
||
217 | * is >= requested value) or false if it has not (current fence |
||
218 | * value is < the requested value. Helper function for |
||
219 | * radeon_fence_signaled(). |
||
220 | */ |
||
221 | static bool radeon_fence_seq_signaled(struct radeon_device *rdev, |
||
222 | u64 seq, unsigned ring) |
||
223 | { |
||
224 | if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) { |
||
225 | return true; |
||
1125 | serge | 226 | } |
2997 | Serge | 227 | /* poll new last sequence at least once */ |
228 | radeon_fence_process(rdev, ring); |
||
229 | if (atomic64_read(&rdev->fence_drv[ring].last_seq) >= seq) { |
||
230 | return true; |
||
231 | } |
||
232 | return false; |
||
1125 | serge | 233 | } |
234 | |||
2997 | Serge | 235 | /** |
236 | * radeon_fence_signaled - check if a fence has signaled |
||
237 | * |
||
238 | * @fence: radeon fence object |
||
239 | * |
||
240 | * Check if the requested fence has signaled (all asics). |
||
241 | * Returns true if the fence has signaled or false if it has not. |
||
242 | */ |
||
243 | bool radeon_fence_signaled(struct radeon_fence *fence) |
||
244 | { |
||
245 | if (!fence) { |
||
246 | return true; |
||
247 | } |
||
248 | if (fence->seq == RADEON_FENCE_SIGNALED_SEQ) { |
||
249 | return true; |
||
250 | } |
||
251 | if (radeon_fence_seq_signaled(fence->rdev, fence->seq, fence->ring)) { |
||
252 | fence->seq = RADEON_FENCE_SIGNALED_SEQ; |
||
253 | return true; |
||
254 | } |
||
255 | return false; |
||
256 | } |
||
1125 | serge | 257 | |
2997 | Serge | 258 | /** |
259 | * radeon_fence_wait_seq - wait for a specific sequence number |
||
260 | * |
||
261 | * @rdev: radeon device pointer |
||
262 | * @target_seq: sequence number we want to wait for |
||
263 | * @ring: ring index the fence is associated with |
||
264 | * @intr: use interruptable sleep |
||
265 | * @lock_ring: whether the ring should be locked or not |
||
266 | * |
||
267 | * Wait for the requested sequence number to be written (all asics). |
||
268 | * @intr selects whether to use interruptable (true) or non-interruptable |
||
269 | * (false) sleep when waiting for the sequence number. Helper function |
||
270 | * for radeon_fence_wait(), et al. |
||
271 | * Returns 0 if the sequence number has passed, error for all other cases. |
||
272 | * -EDEADLK is returned when a GPU lockup has been detected and the ring is |
||
273 | * marked as not ready so no further jobs get scheduled until a successful |
||
274 | * reset. |
||
275 | */ |
||
276 | static int radeon_fence_wait_seq(struct radeon_device *rdev, u64 target_seq, |
||
277 | unsigned ring, bool intr, bool lock_ring) |
||
1125 | serge | 278 | { |
2997 | Serge | 279 | unsigned long timeout, last_activity; |
280 | uint64_t seq; |
||
281 | unsigned i; |
||
282 | bool signaled; |
||
283 | int r; |
||
1125 | serge | 284 | |
2997 | Serge | 285 | while (target_seq > atomic64_read(&rdev->fence_drv[ring].last_seq)) { |
286 | if (!rdev->ring[ring].ready) { |
||
287 | return -EBUSY; |
||
288 | } |
||
289 | |||
290 | timeout = GetTimerTicks() - RADEON_FENCE_JIFFIES_TIMEOUT; |
||
291 | if (time_after(rdev->fence_drv[ring].last_activity, timeout)) { |
||
292 | /* the normal case, timeout is somewhere before last_activity */ |
||
293 | timeout = rdev->fence_drv[ring].last_activity - timeout; |
||
294 | } else { |
||
295 | /* either jiffies wrapped around, or no fence was signaled in the last 500ms |
||
296 | * anyway we will just wait for the minimum amount and then check for a lockup |
||
297 | */ |
||
298 | timeout = 1; |
||
1125 | serge | 299 | } |
2997 | Serge | 300 | seq = atomic64_read(&rdev->fence_drv[ring].last_seq); |
301 | /* Save current last activity valuee, used to check for GPU lockups */ |
||
302 | last_activity = rdev->fence_drv[ring].last_activity; |
||
2005 | serge | 303 | |
2997 | Serge | 304 | // trace_radeon_fence_wait_begin(rdev->ddev, seq); |
305 | radeon_irq_kms_sw_irq_get(rdev, ring); |
||
3192 | Serge | 306 | if (intr) { |
307 | r = wait_event_interruptible_timeout(rdev->fence_queue, |
||
308 | (signaled = radeon_fence_seq_signaled(rdev, target_seq, ring)), |
||
309 | timeout); |
||
310 | } else { |
||
311 | r = wait_event_timeout(rdev->fence_queue, |
||
312 | (signaled = radeon_fence_seq_signaled(rdev, target_seq, ring)), |
||
313 | timeout); |
||
314 | } |
||
2997 | Serge | 315 | radeon_irq_kms_sw_irq_put(rdev, ring); |
3192 | Serge | 316 | if (unlikely(r < 0)) { |
317 | return r; |
||
318 | } |
||
2997 | Serge | 319 | // trace_radeon_fence_wait_end(rdev->ddev, seq); |
1125 | serge | 320 | |
2997 | Serge | 321 | if (unlikely(!signaled)) { |
322 | /* we were interrupted for some reason and fence |
||
323 | * isn't signaled yet, resume waiting */ |
||
324 | if (r) { |
||
325 | continue; |
||
326 | } |
||
1125 | serge | 327 | |
2997 | Serge | 328 | /* check if sequence value has changed since last_activity */ |
329 | if (seq != atomic64_read(&rdev->fence_drv[ring].last_seq)) { |
||
330 | continue; |
||
331 | } |
||
1125 | serge | 332 | |
2997 | Serge | 333 | if (lock_ring) { |
334 | mutex_lock(&rdev->ring_lock); |
||
335 | } |
||
1404 | serge | 336 | |
2997 | Serge | 337 | /* test if somebody else has already decided that this is a lockup */ |
338 | if (last_activity != rdev->fence_drv[ring].last_activity) { |
||
339 | if (lock_ring) { |
||
340 | mutex_unlock(&rdev->ring_lock); |
||
341 | } |
||
342 | continue; |
||
343 | } |
||
1404 | serge | 344 | |
2997 | Serge | 345 | if (radeon_ring_is_lockup(rdev, ring, &rdev->ring[ring])) { |
346 | /* good news we believe it's a lockup */ |
||
347 | dev_warn(rdev->dev, "GPU lockup (waiting for 0x%016llx last fence id 0x%016llx)\n", |
||
348 | target_seq, seq); |
||
349 | |||
350 | /* change last activity so nobody else think there is a lockup */ |
||
351 | for (i = 0; i < RADEON_NUM_RINGS; ++i) { |
||
3031 | serge | 352 | rdev->fence_drv[i].last_activity = GetTimerTicks(); |
2997 | Serge | 353 | } |
354 | |||
355 | /* mark the ring as not ready any more */ |
||
356 | rdev->ring[ring].ready = false; |
||
357 | if (lock_ring) { |
||
358 | mutex_unlock(&rdev->ring_lock); |
||
359 | } |
||
360 | return -EDEADLK; |
||
361 | } |
||
362 | |||
363 | if (lock_ring) { |
||
364 | mutex_unlock(&rdev->ring_lock); |
||
365 | } |
||
366 | } |
||
1125 | serge | 367 | } |
2997 | Serge | 368 | return 0; |
1125 | serge | 369 | } |
370 | |||
2997 | Serge | 371 | /** |
372 | * radeon_fence_wait - wait for a fence to signal |
||
373 | * |
||
374 | * @fence: radeon fence object |
||
375 | * @intr: use interruptable sleep |
||
376 | * |
||
377 | * Wait for the requested fence to signal (all asics). |
||
378 | * @intr selects whether to use interruptable (true) or non-interruptable |
||
379 | * (false) sleep when waiting for the fence. |
||
380 | * Returns 0 if the fence has passed, error for all other cases. |
||
381 | */ |
||
1179 | serge | 382 | int radeon_fence_wait(struct radeon_fence *fence, bool intr) |
383 | { |
||
1125 | serge | 384 | int r; |
385 | |||
386 | if (fence == NULL) { |
||
387 | WARN(1, "Querying an invalid fence : %p !\n", fence); |
||
2997 | Serge | 388 | return -EINVAL; |
1125 | serge | 389 | } |
2997 | Serge | 390 | |
391 | r = radeon_fence_wait_seq(fence->rdev, fence->seq, |
||
392 | fence->ring, intr, true); |
||
393 | if (r) { |
||
394 | return r; |
||
395 | } |
||
396 | fence->seq = RADEON_FENCE_SIGNALED_SEQ; |
||
1125 | serge | 397 | return 0; |
2997 | Serge | 398 | } |
399 | |||
400 | static bool radeon_fence_any_seq_signaled(struct radeon_device *rdev, u64 *seq) |
||
401 | { |
||
402 | unsigned i; |
||
403 | |||
404 | for (i = 0; i < RADEON_NUM_RINGS; ++i) { |
||
405 | if (seq[i] && radeon_fence_seq_signaled(rdev, seq[i], i)) { |
||
406 | return true; |
||
407 | } |
||
1125 | serge | 408 | } |
2997 | Serge | 409 | return false; |
410 | } |
||
2005 | serge | 411 | |
2997 | Serge | 412 | /** |
413 | * radeon_fence_wait_any_seq - wait for a sequence number on any ring |
||
414 | * |
||
415 | * @rdev: radeon device pointer |
||
416 | * @target_seq: sequence number(s) we want to wait for |
||
417 | * @intr: use interruptable sleep |
||
418 | * |
||
419 | * Wait for the requested sequence number(s) to be written by any ring |
||
420 | * (all asics). Sequnce number array is indexed by ring id. |
||
421 | * @intr selects whether to use interruptable (true) or non-interruptable |
||
422 | * (false) sleep when waiting for the sequence number. Helper function |
||
423 | * for radeon_fence_wait_any(), et al. |
||
424 | * Returns 0 if the sequence number has passed, error for all other cases. |
||
425 | */ |
||
426 | static int radeon_fence_wait_any_seq(struct radeon_device *rdev, |
||
427 | u64 *target_seq, bool intr) |
||
428 | { |
||
429 | unsigned long timeout, last_activity, tmp; |
||
430 | unsigned i, ring = RADEON_NUM_RINGS; |
||
431 | bool signaled; |
||
432 | int r; |
||
2005 | serge | 433 | |
2997 | Serge | 434 | for (i = 0, last_activity = 0; i < RADEON_NUM_RINGS; ++i) { |
435 | if (!target_seq[i]) { |
||
436 | continue; |
||
1963 | serge | 437 | } |
2005 | serge | 438 | |
2997 | Serge | 439 | /* use the most recent one as indicator */ |
440 | if (time_after(rdev->fence_drv[i].last_activity, last_activity)) { |
||
441 | last_activity = rdev->fence_drv[i].last_activity; |
||
442 | } |
||
2005 | serge | 443 | |
2997 | Serge | 444 | /* For lockup detection just pick the lowest ring we are |
445 | * actively waiting for |
||
446 | */ |
||
447 | if (i < ring) { |
||
448 | ring = i; |
||
449 | } |
||
1125 | serge | 450 | } |
2997 | Serge | 451 | |
452 | /* nothing to wait for ? */ |
||
453 | if (ring == RADEON_NUM_RINGS) { |
||
454 | return -ENOENT; |
||
455 | } |
||
456 | |||
457 | while (!radeon_fence_any_seq_signaled(rdev, target_seq)) { |
||
458 | timeout = GetTimerTicks() - RADEON_FENCE_JIFFIES_TIMEOUT; |
||
459 | if (time_after(last_activity, timeout)) { |
||
460 | /* the normal case, timeout is somewhere before last_activity */ |
||
461 | timeout = last_activity - timeout; |
||
462 | } else { |
||
463 | /* either jiffies wrapped around, or no fence was signaled in the last 500ms |
||
464 | * anyway we will just wait for the minimum amount and then check for a lockup |
||
465 | */ |
||
466 | timeout = 1; |
||
467 | } |
||
468 | |||
469 | // trace_radeon_fence_wait_begin(rdev->ddev, target_seq[ring]); |
||
470 | for (i = 0; i < RADEON_NUM_RINGS; ++i) { |
||
471 | if (target_seq[i]) { |
||
472 | radeon_irq_kms_sw_irq_get(rdev, i); |
||
473 | } |
||
474 | } |
||
3192 | Serge | 475 | if (intr) { |
476 | r = wait_event_interruptible_timeout(rdev->fence_queue, |
||
477 | (signaled = radeon_fence_any_seq_signaled(rdev, target_seq)), |
||
478 | timeout); |
||
479 | } else { |
||
480 | r = wait_event_timeout(rdev->fence_queue, |
||
481 | (signaled = radeon_fence_any_seq_signaled(rdev, target_seq)), |
||
482 | timeout); |
||
483 | } |
||
2997 | Serge | 484 | for (i = 0; i < RADEON_NUM_RINGS; ++i) { |
485 | if (target_seq[i]) { |
||
486 | radeon_irq_kms_sw_irq_put(rdev, i); |
||
487 | } |
||
488 | } |
||
489 | if (unlikely(r < 0)) { |
||
490 | return r; |
||
491 | } |
||
2004 | serge | 492 | // trace_radeon_fence_wait_end(rdev->ddev, seq); |
2997 | Serge | 493 | |
494 | if (unlikely(!signaled)) { |
||
495 | /* we were interrupted for some reason and fence |
||
496 | * isn't signaled yet, resume waiting */ |
||
1963 | serge | 497 | if (r) { |
2997 | Serge | 498 | continue; |
499 | } |
||
500 | |||
501 | mutex_lock(&rdev->ring_lock); |
||
502 | for (i = 0, tmp = 0; i < RADEON_NUM_RINGS; ++i) { |
||
503 | if (time_after(rdev->fence_drv[i].last_activity, tmp)) { |
||
504 | tmp = rdev->fence_drv[i].last_activity; |
||
505 | } |
||
506 | } |
||
507 | /* test if somebody else has already decided that this is a lockup */ |
||
508 | if (last_activity != tmp) { |
||
509 | last_activity = tmp; |
||
510 | mutex_unlock(&rdev->ring_lock); |
||
511 | continue; |
||
1125 | serge | 512 | } |
2997 | Serge | 513 | |
514 | if (radeon_ring_is_lockup(rdev, ring, &rdev->ring[ring])) { |
||
1963 | serge | 515 | /* good news we believe it's a lockup */ |
2997 | Serge | 516 | dev_warn(rdev->dev, "GPU lockup (waiting for 0x%016llx)\n", |
517 | target_seq[ring]); |
||
2005 | serge | 518 | |
2997 | Serge | 519 | /* change last activity so nobody else think there is a lockup */ |
520 | for (i = 0; i < RADEON_NUM_RINGS; ++i) { |
||
521 | rdev->fence_drv[i].last_activity = GetTimerTicks(); |
||
522 | } |
||
523 | |||
524 | /* mark the ring as not ready any more */ |
||
525 | rdev->ring[ring].ready = false; |
||
526 | mutex_unlock(&rdev->ring_lock); |
||
527 | return -EDEADLK; |
||
1125 | serge | 528 | } |
2997 | Serge | 529 | mutex_unlock(&rdev->ring_lock); |
530 | } |
||
1125 | serge | 531 | } |
2997 | Serge | 532 | return 0; |
1125 | serge | 533 | } |
534 | |||
2997 | Serge | 535 | /** |
536 | * radeon_fence_wait_any - wait for a fence to signal on any ring |
||
537 | * |
||
538 | * @rdev: radeon device pointer |
||
539 | * @fences: radeon fence object(s) |
||
540 | * @intr: use interruptable sleep |
||
541 | * |
||
542 | * Wait for any requested fence to signal (all asics). Fence |
||
543 | * array is indexed by ring id. @intr selects whether to use |
||
544 | * interruptable (true) or non-interruptable (false) sleep when |
||
545 | * waiting for the fences. Used by the suballocator. |
||
546 | * Returns 0 if any fence has passed, error for all other cases. |
||
547 | */ |
||
548 | int radeon_fence_wait_any(struct radeon_device *rdev, |
||
549 | struct radeon_fence **fences, |
||
550 | bool intr) |
||
1125 | serge | 551 | { |
2997 | Serge | 552 | uint64_t seq[RADEON_NUM_RINGS]; |
553 | unsigned i; |
||
1125 | serge | 554 | int r; |
555 | |||
2997 | Serge | 556 | for (i = 0; i < RADEON_NUM_RINGS; ++i) { |
557 | seq[i] = 0; |
||
558 | |||
559 | if (!fences[i]) { |
||
560 | continue; |
||
561 | } |
||
562 | |||
563 | if (fences[i]->seq == RADEON_FENCE_SIGNALED_SEQ) { |
||
564 | /* something was allready signaled */ |
||
1125 | serge | 565 | return 0; |
566 | } |
||
2997 | Serge | 567 | |
568 | seq[i] = fences[i]->seq; |
||
1125 | serge | 569 | } |
2997 | Serge | 570 | |
571 | r = radeon_fence_wait_any_seq(rdev, seq, intr); |
||
572 | if (r) { |
||
573 | return r; |
||
574 | } |
||
575 | return 0; |
||
1125 | serge | 576 | } |
577 | |||
2997 | Serge | 578 | /** |
579 | * radeon_fence_wait_next_locked - wait for the next fence to signal |
||
580 | * |
||
581 | * @rdev: radeon device pointer |
||
582 | * @ring: ring index the fence is associated with |
||
583 | * |
||
584 | * Wait for the next fence on the requested ring to signal (all asics). |
||
585 | * Returns 0 if the next fence has passed, error for all other cases. |
||
586 | * Caller must hold ring lock. |
||
587 | */ |
||
588 | int radeon_fence_wait_next_locked(struct radeon_device *rdev, int ring) |
||
1125 | serge | 589 | { |
2997 | Serge | 590 | uint64_t seq; |
1125 | serge | 591 | |
2997 | Serge | 592 | seq = atomic64_read(&rdev->fence_drv[ring].last_seq) + 1ULL; |
593 | if (seq >= rdev->fence_drv[ring].sync_seq[ring]) { |
||
594 | /* nothing to wait for, last_seq is |
||
595 | already the last emited fence */ |
||
596 | return -ENOENT; |
||
1125 | serge | 597 | } |
2997 | Serge | 598 | return radeon_fence_wait_seq(rdev, seq, ring, false, false); |
599 | } |
||
600 | |||
601 | /** |
||
602 | * radeon_fence_wait_empty_locked - wait for all fences to signal |
||
603 | * |
||
604 | * @rdev: radeon device pointer |
||
605 | * @ring: ring index the fence is associated with |
||
606 | * |
||
607 | * Wait for all fences on the requested ring to signal (all asics). |
||
608 | * Returns 0 if the fences have passed, error for all other cases. |
||
609 | * Caller must hold ring lock. |
||
610 | */ |
||
3192 | Serge | 611 | int radeon_fence_wait_empty_locked(struct radeon_device *rdev, int ring) |
2997 | Serge | 612 | { |
613 | uint64_t seq = rdev->fence_drv[ring].sync_seq[ring]; |
||
3192 | Serge | 614 | int r; |
2997 | Serge | 615 | |
616 | r = radeon_fence_wait_seq(rdev, seq, ring, false, false); |
||
3192 | Serge | 617 | if (r) { |
2997 | Serge | 618 | if (r == -EDEADLK) { |
3192 | Serge | 619 | return -EDEADLK; |
1125 | serge | 620 | } |
3192 | Serge | 621 | dev_err(rdev->dev, "error waiting for ring[%d] to become idle (%d)\n", |
622 | ring, r); |
||
2997 | Serge | 623 | } |
3192 | Serge | 624 | return 0; |
1125 | serge | 625 | } |
626 | |||
2997 | Serge | 627 | /** |
628 | * radeon_fence_ref - take a ref on a fence |
||
629 | * |
||
630 | * @fence: radeon fence object |
||
631 | * |
||
632 | * Take a reference on a fence (all asics). |
||
633 | * Returns the fence. |
||
634 | */ |
||
1125 | serge | 635 | struct radeon_fence *radeon_fence_ref(struct radeon_fence *fence) |
636 | { |
||
637 | kref_get(&fence->kref); |
||
638 | return fence; |
||
639 | } |
||
640 | |||
2997 | Serge | 641 | /** |
642 | * radeon_fence_unref - remove a ref on a fence |
||
643 | * |
||
644 | * @fence: radeon fence object |
||
645 | * |
||
646 | * Remove a reference on a fence (all asics). |
||
647 | */ |
||
1125 | serge | 648 | void radeon_fence_unref(struct radeon_fence **fence) |
649 | { |
||
2997 | Serge | 650 | struct radeon_fence *tmp = *fence; |
1125 | serge | 651 | |
2997 | Serge | 652 | *fence = NULL; |
653 | if (tmp) { |
||
654 | kref_put(&tmp->kref, radeon_fence_destroy); |
||
655 | } |
||
656 | } |
||
2005 | serge | 657 | |
2997 | Serge | 658 | /** |
659 | * radeon_fence_count_emitted - get the count of emitted fences |
||
660 | * |
||
661 | * @rdev: radeon device pointer |
||
662 | * @ring: ring index the fence is associated with |
||
663 | * |
||
664 | * Get the number of fences emitted on the requested ring (all asics). |
||
665 | * Returns the number of emitted fences on the ring. Used by the |
||
666 | * dynpm code to ring track activity. |
||
667 | */ |
||
668 | unsigned radeon_fence_count_emitted(struct radeon_device *rdev, int ring) |
||
669 | { |
||
670 | uint64_t emitted; |
||
671 | |||
672 | /* We are not protected by ring lock when reading the last sequence |
||
673 | * but it's ok to report slightly wrong fence count here. |
||
674 | */ |
||
675 | radeon_fence_process(rdev, ring); |
||
676 | emitted = rdev->fence_drv[ring].sync_seq[ring] |
||
677 | - atomic64_read(&rdev->fence_drv[ring].last_seq); |
||
678 | /* to avoid 32bits warp around */ |
||
679 | if (emitted > 0x10000000) { |
||
680 | emitted = 0x10000000; |
||
681 | } |
||
682 | return (unsigned)emitted; |
||
1125 | serge | 683 | } |
684 | |||
2997 | Serge | 685 | /** |
686 | * radeon_fence_need_sync - do we need a semaphore |
||
687 | * |
||
688 | * @fence: radeon fence object |
||
689 | * @dst_ring: which ring to check against |
||
690 | * |
||
691 | * Check if the fence needs to be synced against another ring |
||
692 | * (all asics). If so, we need to emit a semaphore. |
||
693 | * Returns true if we need to sync with another ring, false if |
||
694 | * not. |
||
695 | */ |
||
696 | bool radeon_fence_need_sync(struct radeon_fence *fence, int dst_ring) |
||
1125 | serge | 697 | { |
2997 | Serge | 698 | struct radeon_fence_driver *fdrv; |
1125 | serge | 699 | |
2997 | Serge | 700 | if (!fence) { |
701 | return false; |
||
702 | } |
||
703 | |||
704 | if (fence->ring == dst_ring) { |
||
705 | return false; |
||
706 | } |
||
707 | |||
708 | /* we are protected by the ring mutex */ |
||
709 | fdrv = &fence->rdev->fence_drv[dst_ring]; |
||
710 | if (fence->seq <= fdrv->sync_seq[fence->ring]) { |
||
711 | return false; |
||
712 | } |
||
713 | |||
714 | return true; |
||
1125 | serge | 715 | } |
716 | |||
2997 | Serge | 717 | /** |
718 | * radeon_fence_note_sync - record the sync point |
||
719 | * |
||
720 | * @fence: radeon fence object |
||
721 | * @dst_ring: which ring to check against |
||
722 | * |
||
723 | * Note the sequence number at which point the fence will |
||
724 | * be synced with the requested ring (all asics). |
||
725 | */ |
||
726 | void radeon_fence_note_sync(struct radeon_fence *fence, int dst_ring) |
||
1125 | serge | 727 | { |
2997 | Serge | 728 | struct radeon_fence_driver *dst, *src; |
729 | unsigned i; |
||
730 | |||
731 | if (!fence) { |
||
732 | return; |
||
733 | } |
||
734 | |||
735 | if (fence->ring == dst_ring) { |
||
736 | return; |
||
737 | } |
||
738 | |||
739 | /* we are protected by the ring mutex */ |
||
740 | src = &fence->rdev->fence_drv[fence->ring]; |
||
741 | dst = &fence->rdev->fence_drv[dst_ring]; |
||
742 | for (i = 0; i < RADEON_NUM_RINGS; ++i) { |
||
743 | if (i == dst_ring) { |
||
744 | continue; |
||
745 | } |
||
746 | dst->sync_seq[i] = max(dst->sync_seq[i], src->sync_seq[i]); |
||
747 | } |
||
748 | } |
||
749 | |||
750 | /** |
||
751 | * radeon_fence_driver_start_ring - make the fence driver |
||
752 | * ready for use on the requested ring. |
||
753 | * |
||
754 | * @rdev: radeon device pointer |
||
755 | * @ring: ring index to start the fence driver on |
||
756 | * |
||
757 | * Make the fence driver ready for processing (all asics). |
||
758 | * Not all asics have all rings, so each asic will only |
||
759 | * start the fence driver on the rings it has. |
||
760 | * Returns 0 for success, errors for failure. |
||
761 | */ |
||
762 | int radeon_fence_driver_start_ring(struct radeon_device *rdev, int ring) |
||
763 | { |
||
764 | uint64_t index; |
||
1125 | serge | 765 | int r; |
766 | |||
2997 | Serge | 767 | radeon_scratch_free(rdev, rdev->fence_drv[ring].scratch_reg); |
3192 | Serge | 768 | if (rdev->wb.use_event || !radeon_ring_supports_scratch_reg(rdev, &rdev->ring[ring])) { |
2997 | Serge | 769 | rdev->fence_drv[ring].scratch_reg = 0; |
770 | index = R600_WB_EVENT_OFFSET + ring * 4; |
||
771 | } else { |
||
772 | r = radeon_scratch_get(rdev, &rdev->fence_drv[ring].scratch_reg); |
||
1125 | serge | 773 | if (r) { |
1404 | serge | 774 | dev_err(rdev->dev, "fence failed to get scratch register\n"); |
1125 | serge | 775 | return r; |
776 | } |
||
2997 | Serge | 777 | index = RADEON_WB_SCRATCH_OFFSET + |
778 | rdev->fence_drv[ring].scratch_reg - |
||
779 | rdev->scratch.reg_base; |
||
780 | } |
||
781 | rdev->fence_drv[ring].cpu_addr = &rdev->wb.wb[index/4]; |
||
782 | rdev->fence_drv[ring].gpu_addr = rdev->wb.gpu_addr + index; |
||
783 | radeon_fence_write(rdev, atomic64_read(&rdev->fence_drv[ring].last_seq), ring); |
||
784 | rdev->fence_drv[ring].initialized = true; |
||
785 | dev_info(rdev->dev, "fence driver on ring %d use gpu addr 0x%016llx and cpu addr 0x%p\n", |
||
786 | ring, rdev->fence_drv[ring].gpu_addr, rdev->fence_drv[ring].cpu_addr); |
||
2004 | serge | 787 | return 0; |
1125 | serge | 788 | } |
789 | |||
2997 | Serge | 790 | /** |
791 | * radeon_fence_driver_init_ring - init the fence driver |
||
792 | * for the requested ring. |
||
793 | * |
||
794 | * @rdev: radeon device pointer |
||
795 | * @ring: ring index to start the fence driver on |
||
796 | * |
||
797 | * Init the fence driver for the requested ring (all asics). |
||
798 | * Helper function for radeon_fence_driver_init(). |
||
799 | */ |
||
800 | static void radeon_fence_driver_init_ring(struct radeon_device *rdev, int ring) |
||
801 | { |
||
802 | int i; |
||
1125 | serge | 803 | |
2997 | Serge | 804 | rdev->fence_drv[ring].scratch_reg = -1; |
805 | rdev->fence_drv[ring].cpu_addr = NULL; |
||
806 | rdev->fence_drv[ring].gpu_addr = 0; |
||
807 | for (i = 0; i < RADEON_NUM_RINGS; ++i) |
||
808 | rdev->fence_drv[ring].sync_seq[i] = 0; |
||
809 | atomic64_set(&rdev->fence_drv[ring].last_seq, 0); |
||
3031 | serge | 810 | rdev->fence_drv[ring].last_activity = GetTimerTicks(); |
2997 | Serge | 811 | rdev->fence_drv[ring].initialized = false; |
812 | } |
||
813 | |||
814 | /** |
||
815 | * radeon_fence_driver_init - init the fence driver |
||
816 | * for all possible rings. |
||
817 | * |
||
818 | * @rdev: radeon device pointer |
||
819 | * |
||
820 | * Init the fence driver for all possible rings (all asics). |
||
821 | * Not all asics have all rings, so each asic will only |
||
822 | * start the fence driver on the rings it has using |
||
823 | * radeon_fence_driver_start_ring(). |
||
824 | * Returns 0 for success. |
||
825 | */ |
||
826 | int radeon_fence_driver_init(struct radeon_device *rdev) |
||
827 | { |
||
828 | int ring; |
||
829 | |||
830 | init_waitqueue_head(&rdev->fence_queue); |
||
831 | for (ring = 0; ring < RADEON_NUM_RINGS; ring++) { |
||
832 | radeon_fence_driver_init_ring(rdev, ring); |
||
833 | } |
||
834 | if (radeon_debugfs_fence_init(rdev)) { |
||
835 | dev_err(rdev->dev, "fence debugfs file creation failed\n"); |
||
836 | } |
||
837 | return 0; |
||
838 | } |
||
839 | |||
840 | /** |
||
841 | * radeon_fence_driver_fini - tear down the fence driver |
||
842 | * for all possible rings. |
||
843 | * |
||
844 | * @rdev: radeon device pointer |
||
845 | * |
||
846 | * Tear down the fence driver for all possible rings (all asics). |
||
847 | */ |
||
848 | void radeon_fence_driver_fini(struct radeon_device *rdev) |
||
849 | { |
||
3192 | Serge | 850 | int ring, r; |
2997 | Serge | 851 | |
852 | mutex_lock(&rdev->ring_lock); |
||
853 | for (ring = 0; ring < RADEON_NUM_RINGS; ring++) { |
||
854 | if (!rdev->fence_drv[ring].initialized) |
||
855 | continue; |
||
3192 | Serge | 856 | r = radeon_fence_wait_empty_locked(rdev, ring); |
857 | if (r) { |
||
858 | /* no need to trigger GPU reset as we are unloading */ |
||
859 | radeon_fence_driver_force_completion(rdev); |
||
860 | } |
||
2997 | Serge | 861 | wake_up_all(&rdev->fence_queue); |
862 | radeon_scratch_free(rdev, rdev->fence_drv[ring].scratch_reg); |
||
863 | rdev->fence_drv[ring].initialized = false; |
||
864 | } |
||
865 | mutex_unlock(&rdev->ring_lock); |
||
866 | } |
||
867 | |||
3192 | Serge | 868 | /** |
869 | * radeon_fence_driver_force_completion - force all fence waiter to complete |
||
870 | * |
||
871 | * @rdev: radeon device pointer |
||
872 | * |
||
873 | * In case of GPU reset failure make sure no process keep waiting on fence |
||
874 | * that will never complete. |
||
875 | */ |
||
876 | void radeon_fence_driver_force_completion(struct radeon_device *rdev) |
||
877 | { |
||
878 | int ring; |
||
2997 | Serge | 879 | |
3192 | Serge | 880 | for (ring = 0; ring < RADEON_NUM_RINGS; ring++) { |
881 | if (!rdev->fence_drv[ring].initialized) |
||
882 | continue; |
||
883 | radeon_fence_write(rdev, rdev->fence_drv[ring].sync_seq[ring], ring); |
||
884 | } |
||
885 | } |
||
886 | |||
887 | |||
1125 | serge | 888 | /* |
889 | * Fence debugfs |
||
890 | */ |
||
891 | #if defined(CONFIG_DEBUG_FS) |
||
892 | static int radeon_debugfs_fence_info(struct seq_file *m, void *data) |
||
893 | { |
||
894 | struct drm_info_node *node = (struct drm_info_node *)m->private; |
||
895 | struct drm_device *dev = node->minor->dev; |
||
896 | struct radeon_device *rdev = dev->dev_private; |
||
2997 | Serge | 897 | int i, j; |
1125 | serge | 898 | |
2997 | Serge | 899 | for (i = 0; i < RADEON_NUM_RINGS; ++i) { |
900 | if (!rdev->fence_drv[i].initialized) |
||
901 | continue; |
||
902 | |||
903 | seq_printf(m, "--- ring %d ---\n", i); |
||
904 | seq_printf(m, "Last signaled fence 0x%016llx\n", |
||
905 | (unsigned long long)atomic64_read(&rdev->fence_drv[i].last_seq)); |
||
906 | seq_printf(m, "Last emitted 0x%016llx\n", |
||
907 | rdev->fence_drv[i].sync_seq[i]); |
||
908 | |||
909 | for (j = 0; j < RADEON_NUM_RINGS; ++j) { |
||
910 | if (i != j && rdev->fence_drv[j].initialized) |
||
911 | seq_printf(m, "Last sync to ring %d 0x%016llx\n", |
||
912 | j, rdev->fence_drv[i].sync_seq[j]); |
||
913 | } |
||
1125 | serge | 914 | } |
915 | return 0; |
||
916 | } |
||
917 | |||
918 | static struct drm_info_list radeon_debugfs_fence_list[] = { |
||
919 | {"radeon_fence_info", &radeon_debugfs_fence_info, 0, NULL}, |
||
920 | }; |
||
921 | #endif |
||
922 | |||
923 | int radeon_debugfs_fence_init(struct radeon_device *rdev) |
||
924 | { |
||
925 | #if defined(CONFIG_DEBUG_FS) |
||
926 | return radeon_debugfs_add_files(rdev, radeon_debugfs_fence_list, 1); |
||
927 | #else |
||
928 | return 0; |
||
929 | #endif |
||
930 | }>>>>>>>=>>>>>>>>>>>>>=>> |