1 /*	$NetBSD: radeon_ring.c,v 1.5 2021/12/18 23:45:43 riastradh Exp $	*/
2 
3 /*
4  * Copyright 2008 Advanced Micro Devices, Inc.
5  * Copyright 2008 Red Hat Inc.
6  * Copyright 2009 Jerome Glisse.
7  *
8  * Permission is hereby granted, free of charge, to any person obtaining a
9  * copy of this software and associated documentation files (the "Software"),
10  * to deal in the Software without restriction, including without limitation
11  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
12  * and/or sell copies of the Software, and to permit persons to whom the
13  * Software is furnished to do so, subject to the following conditions:
14  *
15  * The above copyright notice and this permission notice shall be included in
16  * all copies or substantial portions of the Software.
17  *
18  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
19  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
20  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
21  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
22  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
23  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
24  * OTHER DEALINGS IN THE SOFTWARE.
25  *
26  * Authors: Dave Airlie
27  *          Alex Deucher
28  *          Jerome Glisse
29  *          Christian König
30  */
31 
32 #include <sys/cdefs.h>
33 __KERNEL_RCSID(0, "$NetBSD: radeon_ring.c,v 1.5 2021/12/18 23:45:43 riastradh Exp $");
34 
35 #include <drm/drm_debugfs.h>
36 #include <drm/drm_device.h>
37 #include <drm/drm_file.h>
38 
39 #include "radeon.h"
40 
41 /*
42  * Rings
43  * Most engines on the GPU are fed via ring buffers.  Ring
44  * buffers are areas of GPU accessible memory that the host
45  * writes commands into and the GPU reads commands out of.
46  * There is a rptr (read pointer) that determines where the
47  * GPU is currently reading, and a wptr (write pointer)
48  * which determines where the host has written.  When the
49  * pointers are equal, the ring is idle.  When the host
50  * writes commands to the ring buffer, it increments the
51  * wptr.  The GPU then starts fetching commands and executes
52  * them until the pointers are equal again.
53  */
54 static int radeon_debugfs_ring_init(struct radeon_device *rdev, struct radeon_ring *ring);
55 
56 /**
57  * radeon_ring_supports_scratch_reg - check if the ring supports
58  * writing to scratch registers
59  *
60  * @rdev: radeon_device pointer
61  * @ring: radeon_ring structure holding ring information
62  *
63  * Check if a specific ring supports writing to scratch registers (all asics).
64  * Returns true if the ring supports writing to scratch regs, false if not.
65  */
radeon_ring_supports_scratch_reg(struct radeon_device * rdev,struct radeon_ring * ring)66 bool radeon_ring_supports_scratch_reg(struct radeon_device *rdev,
67 				      struct radeon_ring *ring)
68 {
69 	switch (ring->idx) {
70 	case RADEON_RING_TYPE_GFX_INDEX:
71 	case CAYMAN_RING_TYPE_CP1_INDEX:
72 	case CAYMAN_RING_TYPE_CP2_INDEX:
73 		return true;
74 	default:
75 		return false;
76 	}
77 }
78 
79 /**
80  * radeon_ring_free_size - update the free size
81  *
82  * @rdev: radeon_device pointer
83  * @ring: radeon_ring structure holding ring information
84  *
85  * Update the free dw slots in the ring buffer (all asics).
86  */
radeon_ring_free_size(struct radeon_device * rdev,struct radeon_ring * ring)87 void radeon_ring_free_size(struct radeon_device *rdev, struct radeon_ring *ring)
88 {
89 	uint32_t rptr = radeon_ring_get_rptr(rdev, ring);
90 
91 	/* This works because ring_size is a power of 2 */
92 	ring->ring_free_dw = rptr + (ring->ring_size / 4);
93 	ring->ring_free_dw -= ring->wptr;
94 	ring->ring_free_dw &= ring->ptr_mask;
95 	if (!ring->ring_free_dw) {
96 		/* this is an empty ring */
97 		ring->ring_free_dw = ring->ring_size / 4;
98 		/*  update lockup info to avoid false positive */
99 		radeon_ring_lockup_update(rdev, ring);
100 	}
101 }
102 
103 /**
104  * radeon_ring_alloc - allocate space on the ring buffer
105  *
106  * @rdev: radeon_device pointer
107  * @ring: radeon_ring structure holding ring information
108  * @ndw: number of dwords to allocate in the ring buffer
109  *
110  * Allocate @ndw dwords in the ring buffer (all asics).
111  * Returns 0 on success, error on failure.
112  */
radeon_ring_alloc(struct radeon_device * rdev,struct radeon_ring * ring,unsigned ndw)113 int radeon_ring_alloc(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ndw)
114 {
115 	int r;
116 
117 	/* make sure we aren't trying to allocate more space than there is on the ring */
118 	if (ndw > (ring->ring_size / 4))
119 		return -ENOMEM;
120 	/* Align requested size with padding so unlock_commit can
121 	 * pad safely */
122 	radeon_ring_free_size(rdev, ring);
123 	ndw = (ndw + ring->align_mask) & ~ring->align_mask;
124 	while (ndw > (ring->ring_free_dw - 1)) {
125 		radeon_ring_free_size(rdev, ring);
126 		if (ndw < ring->ring_free_dw) {
127 			break;
128 		}
129 		r = radeon_fence_wait_next(rdev, ring->idx);
130 		if (r)
131 			return r;
132 	}
133 	ring->count_dw = ndw;
134 	ring->wptr_old = ring->wptr;
135 	return 0;
136 }
137 
138 /**
139  * radeon_ring_lock - lock the ring and allocate space on it
140  *
141  * @rdev: radeon_device pointer
142  * @ring: radeon_ring structure holding ring information
143  * @ndw: number of dwords to allocate in the ring buffer
144  *
145  * Lock the ring and allocate @ndw dwords in the ring buffer
146  * (all asics).
147  * Returns 0 on success, error on failure.
148  */
radeon_ring_lock(struct radeon_device * rdev,struct radeon_ring * ring,unsigned ndw)149 int radeon_ring_lock(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ndw)
150 {
151 	int r;
152 
153 	mutex_lock(&rdev->ring_lock);
154 	r = radeon_ring_alloc(rdev, ring, ndw);
155 	if (r) {
156 		mutex_unlock(&rdev->ring_lock);
157 		return r;
158 	}
159 	return 0;
160 }
161 
162 /**
163  * radeon_ring_commit - tell the GPU to execute the new
164  * commands on the ring buffer
165  *
166  * @rdev: radeon_device pointer
167  * @ring: radeon_ring structure holding ring information
168  * @hdp_flush: Whether or not to perform an HDP cache flush
169  *
170  * Update the wptr (write pointer) to tell the GPU to
171  * execute new commands on the ring buffer (all asics).
172  */
radeon_ring_commit(struct radeon_device * rdev,struct radeon_ring * ring,bool hdp_flush)173 void radeon_ring_commit(struct radeon_device *rdev, struct radeon_ring *ring,
174 			bool hdp_flush)
175 {
176 	/* If we are emitting the HDP flush via the ring buffer, we need to
177 	 * do it before padding.
178 	 */
179 	if (hdp_flush && rdev->asic->ring[ring->idx]->hdp_flush)
180 		rdev->asic->ring[ring->idx]->hdp_flush(rdev, ring);
181 	/* We pad to match fetch size */
182 	while (ring->wptr & ring->align_mask) {
183 		radeon_ring_write(ring, ring->nop);
184 	}
185 	mb();
186 	/* If we are emitting the HDP flush via MMIO, we need to do it after
187 	 * all CPU writes to VRAM finished.
188 	 */
189 	if (hdp_flush && rdev->asic->mmio_hdp_flush)
190 		rdev->asic->mmio_hdp_flush(rdev);
191 	radeon_ring_set_wptr(rdev, ring);
192 }
193 
194 /**
195  * radeon_ring_unlock_commit - tell the GPU to execute the new
196  * commands on the ring buffer and unlock it
197  *
198  * @rdev: radeon_device pointer
199  * @ring: radeon_ring structure holding ring information
200  * @hdp_flush: Whether or not to perform an HDP cache flush
201  *
202  * Call radeon_ring_commit() then unlock the ring (all asics).
203  */
radeon_ring_unlock_commit(struct radeon_device * rdev,struct radeon_ring * ring,bool hdp_flush)204 void radeon_ring_unlock_commit(struct radeon_device *rdev, struct radeon_ring *ring,
205 			       bool hdp_flush)
206 {
207 	radeon_ring_commit(rdev, ring, hdp_flush);
208 	mutex_unlock(&rdev->ring_lock);
209 }
210 
211 /**
212  * radeon_ring_undo - reset the wptr
213  *
214  * @ring: radeon_ring structure holding ring information
215  *
216  * Reset the driver's copy of the wptr (all asics).
217  */
radeon_ring_undo(struct radeon_ring * ring)218 void radeon_ring_undo(struct radeon_ring *ring)
219 {
220 	ring->wptr = ring->wptr_old;
221 }
222 
223 /**
224  * radeon_ring_unlock_undo - reset the wptr and unlock the ring
225  *
226  * @ring: radeon_ring structure holding ring information
227  *
228  * Call radeon_ring_undo() then unlock the ring (all asics).
229  */
radeon_ring_unlock_undo(struct radeon_device * rdev,struct radeon_ring * ring)230 void radeon_ring_unlock_undo(struct radeon_device *rdev, struct radeon_ring *ring)
231 {
232 	radeon_ring_undo(ring);
233 	mutex_unlock(&rdev->ring_lock);
234 }
235 
236 /**
237  * radeon_ring_lockup_update - update lockup variables
238  *
239  * @ring: radeon_ring structure holding ring information
240  *
241  * Update the last rptr value and timestamp (all asics).
242  */
radeon_ring_lockup_update(struct radeon_device * rdev,struct radeon_ring * ring)243 void radeon_ring_lockup_update(struct radeon_device *rdev,
244 			       struct radeon_ring *ring)
245 {
246 	atomic_set(&ring->last_rptr, radeon_ring_get_rptr(rdev, ring));
247 	atomic64_set(&ring->last_activity, jiffies_64);
248 }
249 
250 /**
251  * radeon_ring_test_lockup() - check if ring is lockedup by recording information
252  * @rdev:       radeon device structure
253  * @ring:       radeon_ring structure holding ring information
254  *
255  */
radeon_ring_test_lockup(struct radeon_device * rdev,struct radeon_ring * ring)256 bool radeon_ring_test_lockup(struct radeon_device *rdev, struct radeon_ring *ring)
257 {
258 	uint32_t rptr = radeon_ring_get_rptr(rdev, ring);
259 	uint64_t last = atomic64_read(&ring->last_activity);
260 	uint64_t elapsed;
261 
262 	if (rptr != atomic_read(&ring->last_rptr)) {
263 		/* ring is still working, no lockup */
264 		radeon_ring_lockup_update(rdev, ring);
265 		return false;
266 	}
267 
268 	elapsed = jiffies_to_msecs(jiffies_64 - last);
269 	if (radeon_lockup_timeout && elapsed >= radeon_lockup_timeout) {
270 		dev_err(rdev->dev, "ring %d stalled for more than %"PRIu64"msec\n",
271 			ring->idx, elapsed);
272 		return true;
273 	}
274 	/* give a chance to the GPU ... */
275 	return false;
276 }
277 
278 /**
279  * radeon_ring_backup - Back up the content of a ring
280  *
281  * @rdev: radeon_device pointer
282  * @ring: the ring we want to back up
283  *
284  * Saves all unprocessed commits from a ring, returns the number of dwords saved.
285  */
radeon_ring_backup(struct radeon_device * rdev,struct radeon_ring * ring,uint32_t ** data)286 unsigned radeon_ring_backup(struct radeon_device *rdev, struct radeon_ring *ring,
287 			    uint32_t **data)
288 {
289 	unsigned size, ptr, i;
290 
291 	/* just in case lock the ring */
292 	mutex_lock(&rdev->ring_lock);
293 	*data = NULL;
294 
295 	if (ring->ring_obj == NULL) {
296 		mutex_unlock(&rdev->ring_lock);
297 		return 0;
298 	}
299 
300 	/* it doesn't make sense to save anything if all fences are signaled */
301 	if (!radeon_fence_count_emitted(rdev, ring->idx)) {
302 		mutex_unlock(&rdev->ring_lock);
303 		return 0;
304 	}
305 
306 	/* calculate the number of dw on the ring */
307 	if (ring->rptr_save_reg)
308 		ptr = RREG32(ring->rptr_save_reg);
309 	else if (rdev->wb.enabled)
310 		ptr = le32_to_cpu(*ring->next_rptr_cpu_addr);
311 	else {
312 		/* no way to read back the next rptr */
313 		mutex_unlock(&rdev->ring_lock);
314 		return 0;
315 	}
316 
317 	size = ring->wptr + (ring->ring_size / 4);
318 	size -= ptr;
319 	size &= ring->ptr_mask;
320 	if (size == 0) {
321 		mutex_unlock(&rdev->ring_lock);
322 		return 0;
323 	}
324 
325 	/* and then save the content of the ring */
326 	*data = kvmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
327 	if (!*data) {
328 		mutex_unlock(&rdev->ring_lock);
329 		return 0;
330 	}
331 	for (i = 0; i < size; ++i) {
332 		(*data)[i] = ring->ring[ptr++];
333 		ptr &= ring->ptr_mask;
334 	}
335 
336 	mutex_unlock(&rdev->ring_lock);
337 	return size;
338 }
339 
340 /**
341  * radeon_ring_restore - append saved commands to the ring again
342  *
343  * @rdev: radeon_device pointer
344  * @ring: ring to append commands to
345  * @size: number of dwords we want to write
346  * @data: saved commands
347  *
348  * Allocates space on the ring and restore the previously saved commands.
349  */
radeon_ring_restore(struct radeon_device * rdev,struct radeon_ring * ring,unsigned size,uint32_t * data)350 int radeon_ring_restore(struct radeon_device *rdev, struct radeon_ring *ring,
351 			unsigned size, uint32_t *data)
352 {
353 	int i, r;
354 
355 	if (!size || !data)
356 		return 0;
357 
358 	/* restore the saved ring content */
359 	r = radeon_ring_lock(rdev, ring, size);
360 	if (r)
361 		return r;
362 
363 	for (i = 0; i < size; ++i) {
364 		radeon_ring_write(ring, data[i]);
365 	}
366 
367 	radeon_ring_unlock_commit(rdev, ring, false);
368 	kvfree(data);
369 	return 0;
370 }
371 
372 /**
373  * radeon_ring_init - init driver ring struct.
374  *
375  * @rdev: radeon_device pointer
376  * @ring: radeon_ring structure holding ring information
377  * @ring_size: size of the ring
378  * @rptr_offs: offset of the rptr writeback location in the WB buffer
379  * @nop: nop packet for this ring
380  *
381  * Initialize the driver information for the selected ring (all asics).
382  * Returns 0 on success, error on failure.
383  */
radeon_ring_init(struct radeon_device * rdev,struct radeon_ring * ring,unsigned ring_size,unsigned rptr_offs,u32 nop)384 int radeon_ring_init(struct radeon_device *rdev, struct radeon_ring *ring, unsigned ring_size,
385 		     unsigned rptr_offs, u32 nop)
386 {
387 	int r;
388 
389 	ring->ring_size = ring_size;
390 	ring->rptr_offs = rptr_offs;
391 	ring->nop = nop;
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 **)__UNVOLATILE(&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 	}
419 	ring->ptr_mask = (ring->ring_size / 4) - 1;
420 	ring->ring_free_dw = ring->ring_size / 4;
421 	if (rdev->wb.enabled) {
422 		u32 index = RADEON_WB_RING0_NEXT_RPTR + (ring->idx * 4);
423 		ring->next_rptr_gpu_addr = rdev->wb.gpu_addr + index;
424 		ring->next_rptr_cpu_addr = &rdev->wb.wb[index/4];
425 	}
426 	if (radeon_debugfs_ring_init(rdev, ring)) {
427 		DRM_ERROR("Failed to register debugfs file for rings !\n");
428 	}
429 	radeon_ring_lockup_update(rdev, ring);
430 	return 0;
431 }
432 
433 /**
434  * radeon_ring_fini - tear down the driver ring struct.
435  *
436  * @rdev: radeon_device pointer
437  * @ring: radeon_ring structure holding ring information
438  *
439  * Tear down the driver information for the selected ring (all asics).
440  */
radeon_ring_fini(struct radeon_device * rdev,struct radeon_ring * ring)441 void radeon_ring_fini(struct radeon_device *rdev, struct radeon_ring *ring)
442 {
443 	int r;
444 	struct radeon_bo *ring_obj;
445 
446 	mutex_lock(&rdev->ring_lock);
447 	ring_obj = ring->ring_obj;
448 	ring->ready = false;
449 	ring->ring = NULL;
450 	ring->ring_obj = NULL;
451 	mutex_unlock(&rdev->ring_lock);
452 
453 	if (ring_obj) {
454 		r = radeon_bo_reserve(ring_obj, false);
455 		if (likely(r == 0)) {
456 			radeon_bo_kunmap(ring_obj);
457 			radeon_bo_unpin(ring_obj);
458 			radeon_bo_unreserve(ring_obj);
459 		}
460 		radeon_bo_unref(&ring_obj);
461 	}
462 }
463 
464 /*
465  * Debugfs info
466  */
467 #if defined(CONFIG_DEBUG_FS)
468 
radeon_debugfs_ring_info(struct seq_file * m,void * data)469 static int radeon_debugfs_ring_info(struct seq_file *m, void *data)
470 {
471 	struct drm_info_node *node = (struct drm_info_node *) m->private;
472 	struct drm_device *dev = node->minor->dev;
473 	struct radeon_device *rdev = dev->dev_private;
474 	int ridx = *(int*)node->info_ent->data;
475 	struct radeon_ring *ring = &rdev->ring[ridx];
476 
477 	uint32_t rptr, wptr, rptr_next;
478 	unsigned count, i, j;
479 
480 	radeon_ring_free_size(rdev, ring);
481 	count = (ring->ring_size / 4) - ring->ring_free_dw;
482 
483 	wptr = radeon_ring_get_wptr(rdev, ring);
484 	seq_printf(m, "wptr: 0x%08x [%5d]\n",
485 		   wptr, wptr);
486 
487 	rptr = radeon_ring_get_rptr(rdev, ring);
488 	seq_printf(m, "rptr: 0x%08x [%5d]\n",
489 		   rptr, rptr);
490 
491 	if (ring->rptr_save_reg) {
492 		rptr_next = RREG32(ring->rptr_save_reg);
493 		seq_printf(m, "rptr next(0x%04x): 0x%08x [%5d]\n",
494 			   ring->rptr_save_reg, rptr_next, rptr_next);
495 	} else
496 		rptr_next = ~0;
497 
498 	seq_printf(m, "driver's copy of the wptr: 0x%08x [%5d]\n",
499 		   ring->wptr, ring->wptr);
500 	seq_printf(m, "last semaphore signal addr : 0x%016llx\n",
501 		   ring->last_semaphore_signal_addr);
502 	seq_printf(m, "last semaphore wait addr   : 0x%016llx\n",
503 		   ring->last_semaphore_wait_addr);
504 	seq_printf(m, "%u free dwords in ring\n", ring->ring_free_dw);
505 	seq_printf(m, "%u dwords in ring\n", count);
506 
507 	if (!ring->ring)
508 		return 0;
509 
510 	/* print 8 dw before current rptr as often it's the last executed
511 	 * packet that is the root issue
512 	 */
513 	i = (rptr + ring->ptr_mask + 1 - 32) & ring->ptr_mask;
514 	for (j = 0; j <= (count + 32); j++) {
515 		seq_printf(m, "r[%5d]=0x%08x", i, ring->ring[i]);
516 		if (rptr == i)
517 			seq_puts(m, " *");
518 		if (rptr_next == i)
519 			seq_puts(m, " #");
520 		seq_puts(m, "\n");
521 		i = (i + 1) & ring->ptr_mask;
522 	}
523 	return 0;
524 }
525 
526 static int radeon_gfx_index = RADEON_RING_TYPE_GFX_INDEX;
527 static int cayman_cp1_index = CAYMAN_RING_TYPE_CP1_INDEX;
528 static int cayman_cp2_index = CAYMAN_RING_TYPE_CP2_INDEX;
529 static int radeon_dma1_index = R600_RING_TYPE_DMA_INDEX;
530 static int radeon_dma2_index = CAYMAN_RING_TYPE_DMA1_INDEX;
531 static int r600_uvd_index = R600_RING_TYPE_UVD_INDEX;
532 static int si_vce1_index = TN_RING_TYPE_VCE1_INDEX;
533 static int si_vce2_index = TN_RING_TYPE_VCE2_INDEX;
534 
535 static struct drm_info_list radeon_debugfs_ring_info_list[] = {
536 	{"radeon_ring_gfx", radeon_debugfs_ring_info, 0, &radeon_gfx_index},
537 	{"radeon_ring_cp1", radeon_debugfs_ring_info, 0, &cayman_cp1_index},
538 	{"radeon_ring_cp2", radeon_debugfs_ring_info, 0, &cayman_cp2_index},
539 	{"radeon_ring_dma1", radeon_debugfs_ring_info, 0, &radeon_dma1_index},
540 	{"radeon_ring_dma2", radeon_debugfs_ring_info, 0, &radeon_dma2_index},
541 	{"radeon_ring_uvd", radeon_debugfs_ring_info, 0, &r600_uvd_index},
542 	{"radeon_ring_vce1", radeon_debugfs_ring_info, 0, &si_vce1_index},
543 	{"radeon_ring_vce2", radeon_debugfs_ring_info, 0, &si_vce2_index},
544 };
545 
546 #endif
547 
radeon_debugfs_ring_init(struct radeon_device * rdev,struct radeon_ring * ring)548 static int radeon_debugfs_ring_init(struct radeon_device *rdev, struct radeon_ring *ring)
549 {
550 #if defined(CONFIG_DEBUG_FS)
551 	unsigned i;
552 	for (i = 0; i < ARRAY_SIZE(radeon_debugfs_ring_info_list); ++i) {
553 		struct drm_info_list *info = &radeon_debugfs_ring_info_list[i];
554 		int ridx = *(int*)radeon_debugfs_ring_info_list[i].data;
555 		unsigned r;
556 
557 		if (&rdev->ring[ridx] != ring)
558 			continue;
559 
560 		r = radeon_debugfs_add_files(rdev, info, 1);
561 		if (r)
562 			return r;
563 	}
564 #endif
565 	return 0;
566 }
567