xref: /linux/drivers/gpu/drm/radeon/radeon_ring.c (revision 33d5ae6c)
1 /*
2  * Copyright 2008 Advanced Micro Devices, Inc.
3  * Copyright 2008 Red Hat Inc.
4  * Copyright 2009 Jerome Glisse.
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a
7  * copy of this software and associated documentation files (the "Software"),
8  * to deal in the Software without restriction, including without limitation
9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
10  * and/or sell copies of the Software, and to permit persons to whom the
11  * Software is furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in
14  * all copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
19  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
20  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
21  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
22  * OTHER DEALINGS IN THE SOFTWARE.
23  *
24  * Authors: Dave Airlie
25  *          Alex Deucher
26  *          Jerome Glisse
27  *          Christian König
28  */
29 
30 #include <linux/debugfs.h>
31 
32 #include <drm/drm_device.h>
33 #include <drm/drm_file.h>
34 
35 #include "radeon.h"
36 
37 /*
38  * Rings
39  * Most engines on the GPU are fed via ring buffers.  Ring
40  * buffers are areas of GPU accessible memory that the host
41  * writes commands into and the GPU reads commands out of.
42  * There is a rptr (read pointer) that determines where the
43  * GPU is currently reading, and a wptr (write pointer)
44  * which determines where the host has written.  When the
45  * pointers are equal, the ring is idle.  When the host
46  * writes commands to the ring buffer, it increments the
47  * wptr.  The GPU then starts fetching commands and executes
48  * them until the pointers are equal again.
49  */
50 static void radeon_debugfs_ring_init(struct radeon_device *rdev, struct radeon_ring *ring);
51 
52 /**
53  * radeon_ring_supports_scratch_reg - check if the ring supports
54  * writing to scratch registers
55  *
56  * @rdev: radeon_device pointer
57  * @ring: radeon_ring structure holding ring information
58  *
59  * Check if a specific ring supports writing to scratch registers (all asics).
60  * Returns true if the ring supports writing to scratch regs, false if not.
61  */
radeon_ring_supports_scratch_reg(struct radeon_device * rdev,struct radeon_ring * ring)62 bool radeon_ring_supports_scratch_reg(struct radeon_device *rdev,
63 				      struct radeon_ring *ring)
64 {
65 	switch (ring->idx) {
66 	case RADEON_RING_TYPE_GFX_INDEX:
67 	case CAYMAN_RING_TYPE_CP1_INDEX:
68 	case CAYMAN_RING_TYPE_CP2_INDEX:
69 		return true;
70 	default:
71 		return false;
72 	}
73 }
74 
75 /**
76  * radeon_ring_free_size - update the free size
77  *
78  * @rdev: radeon_device pointer
79  * @ring: radeon_ring structure holding ring information
80  *
81  * Update the free dw slots in the ring buffer (all asics).
82  */
radeon_ring_free_size(struct radeon_device * rdev,struct radeon_ring * ring)83 void radeon_ring_free_size(struct radeon_device *rdev, struct radeon_ring *ring)
84 {
85 	uint32_t rptr = radeon_ring_get_rptr(rdev, ring);
86 
87 	/* This works because ring_size is a power of 2 */
88 	ring->ring_free_dw = rptr + (ring->ring_size / 4);
89 	ring->ring_free_dw -= ring->wptr;
90 	ring->ring_free_dw &= ring->ptr_mask;
91 	if (!ring->ring_free_dw) {
92 		/* this is an empty ring */
93 		ring->ring_free_dw = ring->ring_size / 4;
94 		/*  update lockup info to avoid false positive */
95 		radeon_ring_lockup_update(rdev, ring);
96 	}
97 }
98 
99 /**
100  * radeon_ring_alloc - allocate space on the ring buffer
101  *
102  * @rdev: radeon_device pointer
103  * @ring: radeon_ring structure holding ring information
104  * @ndw: number of dwords to allocate in the ring buffer
105  *
106  * Allocate @ndw dwords in the ring buffer (all asics).
107  * Returns 0 on success, error on failure.
108  */
radeon_ring_alloc(struct radeon_device * rdev,struct radeon_ring * ring,unsigned ndw)109 int radeon_ring_alloc(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ndw)
110 {
111 	int r;
112 
113 	/* make sure we aren't trying to allocate more space than there is on the ring */
114 	if (ndw > (ring->ring_size / 4))
115 		return -ENOMEM;
116 	/* Align requested size with padding so unlock_commit can
117 	 * pad safely */
118 	radeon_ring_free_size(rdev, ring);
119 	ndw = (ndw + ring->align_mask) & ~ring->align_mask;
120 	while (ndw > (ring->ring_free_dw - 1)) {
121 		radeon_ring_free_size(rdev, ring);
122 		if (ndw < ring->ring_free_dw) {
123 			break;
124 		}
125 		r = radeon_fence_wait_next(rdev, ring->idx);
126 		if (r)
127 			return r;
128 	}
129 	ring->count_dw = ndw;
130 	ring->wptr_old = ring->wptr;
131 	return 0;
132 }
133 
134 /**
135  * radeon_ring_lock - lock the ring and allocate space on it
136  *
137  * @rdev: radeon_device pointer
138  * @ring: radeon_ring structure holding ring information
139  * @ndw: number of dwords to allocate in the ring buffer
140  *
141  * Lock the ring and allocate @ndw dwords in the ring buffer
142  * (all asics).
143  * Returns 0 on success, error on failure.
144  */
radeon_ring_lock(struct radeon_device * rdev,struct radeon_ring * ring,unsigned ndw)145 int radeon_ring_lock(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ndw)
146 {
147 	int r;
148 
149 	mutex_lock(&rdev->ring_lock);
150 	r = radeon_ring_alloc(rdev, ring, ndw);
151 	if (r) {
152 		mutex_unlock(&rdev->ring_lock);
153 		return r;
154 	}
155 	return 0;
156 }
157 
158 /**
159  * radeon_ring_commit - tell the GPU to execute the new
160  * commands on the ring buffer
161  *
162  * @rdev: radeon_device pointer
163  * @ring: radeon_ring structure holding ring information
164  * @hdp_flush: Whether or not to perform an HDP cache flush
165  *
166  * Update the wptr (write pointer) to tell the GPU to
167  * execute new commands on the ring buffer (all asics).
168  */
radeon_ring_commit(struct radeon_device * rdev,struct radeon_ring * ring,bool hdp_flush)169 void radeon_ring_commit(struct radeon_device *rdev, struct radeon_ring *ring,
170 			bool hdp_flush)
171 {
172 	/* If we are emitting the HDP flush via the ring buffer, we need to
173 	 * do it before padding.
174 	 */
175 	if (hdp_flush && rdev->asic->ring[ring->idx]->hdp_flush)
176 		rdev->asic->ring[ring->idx]->hdp_flush(rdev, ring);
177 	/* We pad to match fetch size */
178 	while (ring->wptr & ring->align_mask) {
179 		radeon_ring_write(ring, ring->nop);
180 	}
181 	mb();
182 	/* If we are emitting the HDP flush via MMIO, we need to do it after
183 	 * all CPU writes to VRAM finished.
184 	 */
185 	if (hdp_flush && rdev->asic->mmio_hdp_flush)
186 		rdev->asic->mmio_hdp_flush(rdev);
187 	radeon_ring_set_wptr(rdev, ring);
188 }
189 
190 /**
191  * radeon_ring_unlock_commit - tell the GPU to execute the new
192  * commands on the ring buffer and unlock it
193  *
194  * @rdev: radeon_device pointer
195  * @ring: radeon_ring structure holding ring information
196  * @hdp_flush: Whether or not to perform an HDP cache flush
197  *
198  * Call radeon_ring_commit() then unlock the ring (all asics).
199  */
radeon_ring_unlock_commit(struct radeon_device * rdev,struct radeon_ring * ring,bool hdp_flush)200 void radeon_ring_unlock_commit(struct radeon_device *rdev, struct radeon_ring *ring,
201 			       bool hdp_flush)
202 {
203 	radeon_ring_commit(rdev, ring, hdp_flush);
204 	mutex_unlock(&rdev->ring_lock);
205 }
206 
207 /**
208  * radeon_ring_undo - reset the wptr
209  *
210  * @ring: radeon_ring structure holding ring information
211  *
212  * Reset the driver's copy of the wptr (all asics).
213  */
radeon_ring_undo(struct radeon_ring * ring)214 void radeon_ring_undo(struct radeon_ring *ring)
215 {
216 	ring->wptr = ring->wptr_old;
217 }
218 
219 /**
220  * radeon_ring_unlock_undo - reset the wptr and unlock the ring
221  *
222  * @rdev:       radeon device structure
223  * @ring: radeon_ring structure holding ring information
224  *
225  * Call radeon_ring_undo() then unlock the ring (all asics).
226  */
radeon_ring_unlock_undo(struct radeon_device * rdev,struct radeon_ring * ring)227 void radeon_ring_unlock_undo(struct radeon_device *rdev, struct radeon_ring *ring)
228 {
229 	radeon_ring_undo(ring);
230 	mutex_unlock(&rdev->ring_lock);
231 }
232 
233 /**
234  * radeon_ring_lockup_update - update lockup variables
235  *
236  * @rdev:       radeon device structure
237  * @ring: radeon_ring structure holding ring information
238  *
239  * Update the last rptr value and timestamp (all asics).
240  */
radeon_ring_lockup_update(struct radeon_device * rdev,struct radeon_ring * ring)241 void radeon_ring_lockup_update(struct radeon_device *rdev,
242 			       struct radeon_ring *ring)
243 {
244 	atomic_set(&ring->last_rptr, radeon_ring_get_rptr(rdev, ring));
245 	atomic64_set(&ring->last_activity, jiffies_64);
246 }
247 
248 /**
249  * radeon_ring_test_lockup() - check if ring is lockedup by recording information
250  * @rdev:       radeon device structure
251  * @ring:       radeon_ring structure holding ring information
252  *
253  */
radeon_ring_test_lockup(struct radeon_device * rdev,struct radeon_ring * ring)254 bool radeon_ring_test_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
255 {
256 	uint32_t rptr = radeon_ring_get_rptr(rdev, ring);
257 	uint64_t last = atomic64_read(&ring->last_activity);
258 	uint64_t elapsed;
259 
260 	if (rptr != atomic_read(&ring->last_rptr)) {
261 		/* ring is still working, no lockup */
262 		radeon_ring_lockup_update(rdev, ring);
263 		return false;
264 	}
265 
266 	elapsed = jiffies_to_msecs(jiffies_64 - last);
267 	if (radeon_lockup_timeout && elapsed >= radeon_lockup_timeout) {
268 		dev_err(rdev->dev, "ring %d stalled for more than %llumsec\n",
269 			ring->idx, elapsed);
270 		return true;
271 	}
272 	/* give a chance to the GPU ... */
273 	return false;
274 }
275 
276 /**
277  * radeon_ring_backup - Back up the content of a ring
278  *
279  * @rdev: radeon_device pointer
280  * @ring: the ring we want to back up
281  * @data: placeholder for returned commit data
282  *
283  * Saves all unprocessed commits from a ring, returns the number of dwords saved.
284  */
radeon_ring_backup(struct radeon_device * rdev,struct radeon_ring * ring,uint32_t ** data)285 unsigned radeon_ring_backup(struct radeon_device *rdev, struct radeon_ring *ring,
286 			    uint32_t **data)
287 {
288 	unsigned size, ptr, i;
289 
290 	/* just in case lock the ring */
291 	mutex_lock(&rdev->ring_lock);
292 	*data = NULL;
293 
294 	if (ring->ring_obj == NULL) {
295 		mutex_unlock(&rdev->ring_lock);
296 		return 0;
297 	}
298 
299 	/* it doesn't make sense to save anything if all fences are signaled */
300 	if (!radeon_fence_count_emitted(rdev, ring->idx)) {
301 		mutex_unlock(&rdev->ring_lock);
302 		return 0;
303 	}
304 
305 	/* calculate the number of dw on the ring */
306 	if (ring->rptr_save_reg)
307 		ptr = RREG32(ring->rptr_save_reg);
308 	else if (rdev->wb.enabled)
309 		ptr = le32_to_cpu(*ring->next_rptr_cpu_addr);
310 	else {
311 		/* no way to read back the next rptr */
312 		mutex_unlock(&rdev->ring_lock);
313 		return 0;
314 	}
315 
316 	size = ring->wptr + (ring->ring_size / 4);
317 	size -= ptr;
318 	size &= ring->ptr_mask;
319 	if (size == 0) {
320 		mutex_unlock(&rdev->ring_lock);
321 		return 0;
322 	}
323 
324 	/* and then save the content of the ring */
325 	*data = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
326 	if (!*data) {
327 		mutex_unlock(&rdev->ring_lock);
328 		return 0;
329 	}
330 	for (i = 0; i < size; ++i) {
331 		(*data)[i] = ring->ring[ptr++];
332 		ptr &= ring->ptr_mask;
333 	}
334 
335 	mutex_unlock(&rdev->ring_lock);
336 	return size;
337 }
338 
339 /**
340  * radeon_ring_restore - append saved commands to the ring again
341  *
342  * @rdev: radeon_device pointer
343  * @ring: ring to append commands to
344  * @size: number of dwords we want to write
345  * @data: saved commands
346  *
347  * Allocates space on the ring and restore the previously saved commands.
348  */
radeon_ring_restore(struct radeon_device * rdev,struct radeon_ring * ring,unsigned size,uint32_t * data)349 int radeon_ring_restore(struct radeon_device *rdev, struct radeon_ring *ring,
350 			unsigned size, uint32_t *data)
351 {
352 	int i, r;
353 
354 	if (!size || !data)
355 		return 0;
356 
357 	/* restore the saved ring content */
358 	r = radeon_ring_lock(rdev, ring, size);
359 	if (r)
360 		return r;
361 
362 	for (i = 0; i < size; ++i) {
363 		radeon_ring_write(ring, data[i]);
364 	}
365 
366 	radeon_ring_unlock_commit(rdev, ring, false);
367 	kvfree(data);
368 	return 0;
369 }
370 
371 /**
372  * radeon_ring_init - init driver ring struct.
373  *
374  * @rdev: radeon_device pointer
375  * @ring: radeon_ring structure holding ring information
376  * @ring_size: size of the ring
377  * @rptr_offs: offset of the rptr writeback location in the WB buffer
378  * @nop: nop packet for this ring
379  *
380  * Initialize the driver information for the selected ring (all asics).
381  * Returns 0 on success, error on failure.
382  */
radeon_ring_init(struct radeon_device * rdev,struct radeon_ring * ring,unsigned ring_size,unsigned rptr_offs,u32 nop)383 int radeon_ring_init(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ring_size,
384 		     unsigned rptr_offs, u32 nop)
385 {
386 	int r;
387 
388 	ring->ring_size = ring_size;
389 	ring->rptr_offs = rptr_offs;
390 	ring->nop = nop;
391 	ring->rdev = rdev;
392 	/* Allocate ring buffer */
393 	if (ring->ring_obj == NULL) {
394 		r = radeon_bo_create(rdev, ring->ring_size, PAGE_SIZE, true,
395 				     RADEON_GEM_DOMAIN_GTT, 0, NULL,
396 				     NULL, &ring->ring_obj);
397 		if (r) {
398 			dev_err(rdev->dev, "(%d) ring create failed\n", r);
399 			return r;
400 		}
401 		r = radeon_bo_reserve(ring->ring_obj, false);
402 		if (unlikely(r != 0))
403 			return r;
404 		r = radeon_bo_pin(ring->ring_obj, RADEON_GEM_DOMAIN_GTT,
405 					&ring->gpu_addr);
406 		if (r) {
407 			radeon_bo_unreserve(ring->ring_obj);
408 			dev_err(rdev->dev, "(%d) ring pin failed\n", r);
409 			return r;
410 		}
411 		r = radeon_bo_kmap(ring->ring_obj,
412 				       (void **)&ring->ring);
413 		radeon_bo_unreserve(ring->ring_obj);
414 		if (r) {
415 			dev_err(rdev->dev, "(%d) ring map failed\n", r);
416 			return r;
417 		}
418 		radeon_debugfs_ring_init(rdev, ring);
419 	}
420 	ring->ptr_mask = (ring->ring_size / 4) - 1;
421 	ring->ring_free_dw = ring->ring_size / 4;
422 	if (rdev->wb.enabled) {
423 		u32 index = RADEON_WB_RING0_NEXT_RPTR + (ring->idx * 4);
424 		ring->next_rptr_gpu_addr = rdev->wb.gpu_addr + index;
425 		ring->next_rptr_cpu_addr = &rdev->wb.wb[index/4];
426 	}
427 	radeon_ring_lockup_update(rdev, ring);
428 	return 0;
429 }
430 
431 /**
432  * radeon_ring_fini - tear down the driver ring struct.
433  *
434  * @rdev: radeon_device pointer
435  * @ring: radeon_ring structure holding ring information
436  *
437  * Tear down the driver information for the selected ring (all asics).
438  */
radeon_ring_fini(struct radeon_device * rdev,struct radeon_ring * ring)439 void radeon_ring_fini(struct radeon_device *rdev, struct radeon_ring *ring)
440 {
441 	int r;
442 	struct radeon_bo *ring_obj;
443 
444 	mutex_lock(&rdev->ring_lock);
445 	ring_obj = ring->ring_obj;
446 	ring->ready = false;
447 	ring->ring = NULL;
448 	ring->ring_obj = NULL;
449 	mutex_unlock(&rdev->ring_lock);
450 
451 	if (ring_obj) {
452 		r = radeon_bo_reserve(ring_obj, false);
453 		if (likely(r == 0)) {
454 			radeon_bo_kunmap(ring_obj);
455 			radeon_bo_unpin(ring_obj);
456 			radeon_bo_unreserve(ring_obj);
457 		}
458 		radeon_bo_unref(&ring_obj);
459 	}
460 }
461 
462 /*
463  * Debugfs info
464  */
465 #if defined(CONFIG_DEBUG_FS)
466 
radeon_debugfs_ring_info_show(struct seq_file * m,void * unused)467 static int radeon_debugfs_ring_info_show(struct seq_file *m, void *unused)
468 {
469 	struct radeon_ring *ring = m->private;
470 	struct radeon_device *rdev = ring->rdev;
471 
472 	uint32_t rptr, wptr, rptr_next;
473 	unsigned count, i, j;
474 
475 	radeon_ring_free_size(rdev, ring);
476 	count = (ring->ring_size / 4) - ring->ring_free_dw;
477 
478 	wptr = radeon_ring_get_wptr(rdev, ring);
479 	seq_printf(m, "wptr: 0x%08x [%5d]\n",
480 		   wptr, wptr);
481 
482 	rptr = radeon_ring_get_rptr(rdev, ring);
483 	seq_printf(m, "rptr: 0x%08x [%5d]\n",
484 		   rptr, rptr);
485 
486 	if (ring->rptr_save_reg) {
487 		rptr_next = RREG32(ring->rptr_save_reg);
488 		seq_printf(m, "rptr next(0x%04x): 0x%08x [%5d]\n",
489 			   ring->rptr_save_reg, rptr_next, rptr_next);
490 	} else
491 		rptr_next = ~0;
492 
493 	seq_printf(m, "driver's copy of the wptr: 0x%08x [%5d]\n",
494 		   ring->wptr, ring->wptr);
495 	seq_printf(m, "last semaphore signal addr : 0x%016llx\n",
496 		   ring->last_semaphore_signal_addr);
497 	seq_printf(m, "last semaphore wait addr   : 0x%016llx\n",
498 		   ring->last_semaphore_wait_addr);
499 	seq_printf(m, "%u free dwords in ring\n", ring->ring_free_dw);
500 	seq_printf(m, "%u dwords in ring\n", count);
501 
502 	if (!ring->ring)
503 		return 0;
504 
505 	/* print 8 dw before current rptr as often it's the last executed
506 	 * packet that is the root issue
507 	 */
508 	i = (rptr + ring->ptr_mask + 1 - 32) & ring->ptr_mask;
509 	for (j = 0; j <= (count + 32); j++) {
510 		seq_printf(m, "r[%5d]=0x%08x", i, ring->ring[i]);
511 		if (rptr == i)
512 			seq_puts(m, " *");
513 		if (rptr_next == i)
514 			seq_puts(m, " #");
515 		seq_puts(m, "\n");
516 		i = (i + 1) & ring->ptr_mask;
517 	}
518 	return 0;
519 }
520 
521 DEFINE_SHOW_ATTRIBUTE(radeon_debugfs_ring_info);
522 
radeon_debugfs_ring_idx_to_name(uint32_t ridx)523 static const char *radeon_debugfs_ring_idx_to_name(uint32_t ridx)
524 {
525 	switch (ridx) {
526 	case RADEON_RING_TYPE_GFX_INDEX:
527 		return "radeon_ring_gfx";
528 	case CAYMAN_RING_TYPE_CP1_INDEX:
529 		return "radeon_ring_cp1";
530 	case CAYMAN_RING_TYPE_CP2_INDEX:
531 		return "radeon_ring_cp2";
532 	case R600_RING_TYPE_DMA_INDEX:
533 		return "radeon_ring_dma1";
534 	case CAYMAN_RING_TYPE_DMA1_INDEX:
535 		return "radeon_ring_dma2";
536 	case R600_RING_TYPE_UVD_INDEX:
537 		return "radeon_ring_uvd";
538 	case TN_RING_TYPE_VCE1_INDEX:
539 		return "radeon_ring_vce1";
540 	case TN_RING_TYPE_VCE2_INDEX:
541 		return "radeon_ring_vce2";
542 	default:
543 		return NULL;
544 
545 	}
546 }
547 #endif
548 
radeon_debugfs_ring_init(struct radeon_device * rdev,struct radeon_ring * ring)549 static void radeon_debugfs_ring_init(struct radeon_device *rdev, struct radeon_ring *ring)
550 {
551 #if defined(CONFIG_DEBUG_FS)
552 	const char *ring_name = radeon_debugfs_ring_idx_to_name(ring->idx);
553 	struct dentry *root = rdev->ddev->primary->debugfs_root;
554 
555 	if (ring_name)
556 		debugfs_create_file(ring_name, 0444, root, ring,
557 				    &radeon_debugfs_ring_info_fops);
558 
559 #endif
560 }
561