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 #include <linux/seq_file.h>
30 #include <linux/slab.h>
31 #include <linux/debugfs.h>
32 #include <drm/drmP.h>
33 #include <drm/amdgpu_drm.h>
34 #include "amdgpu.h"
35 #include "atom.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 int amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
51 				    struct amdgpu_ring *ring);
52 static void amdgpu_debugfs_ring_fini(struct amdgpu_ring *ring);
53 
54 /**
55  * amdgpu_ring_alloc - allocate space on the ring buffer
56  *
57  * @adev: amdgpu_device pointer
58  * @ring: amdgpu_ring structure holding ring information
59  * @ndw: number of dwords to allocate in the ring buffer
60  *
61  * Allocate @ndw dwords in the ring buffer (all asics).
62  * Returns 0 on success, error on failure.
63  */
64 int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw)
65 {
66 	/* Align requested size with padding so unlock_commit can
67 	 * pad safely */
68 	ndw = (ndw + ring->funcs->align_mask) & ~ring->funcs->align_mask;
69 
70 	/* Make sure we aren't trying to allocate more space
71 	 * than the maximum for one submission
72 	 */
73 	if (WARN_ON_ONCE(ndw > ring->max_dw))
74 		return -ENOMEM;
75 
76 	ring->count_dw = ndw;
77 	ring->wptr_old = ring->wptr;
78 
79 	if (ring->funcs->begin_use)
80 		ring->funcs->begin_use(ring);
81 
82 	return 0;
83 }
84 
85 /** amdgpu_ring_insert_nop - insert NOP packets
86  *
87  * @ring: amdgpu_ring structure holding ring information
88  * @count: the number of NOP packets to insert
89  *
90  * This is the generic insert_nop function for rings except SDMA
91  */
92 void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
93 {
94 	int i;
95 
96 	for (i = 0; i < count; i++)
97 		amdgpu_ring_write(ring, ring->funcs->nop);
98 }
99 
100 /** amdgpu_ring_generic_pad_ib - pad IB with NOP packets
101  *
102  * @ring: amdgpu_ring structure holding ring information
103  * @ib: IB to add NOP packets to
104  *
105  * This is the generic pad_ib function for rings except SDMA
106  */
107 void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
108 {
109 	while (ib->length_dw & ring->funcs->align_mask)
110 		ib->ptr[ib->length_dw++] = ring->funcs->nop;
111 }
112 
113 /**
114  * amdgpu_ring_commit - tell the GPU to execute the new
115  * commands on the ring buffer
116  *
117  * @adev: amdgpu_device pointer
118  * @ring: amdgpu_ring structure holding ring information
119  *
120  * Update the wptr (write pointer) to tell the GPU to
121  * execute new commands on the ring buffer (all asics).
122  */
123 void amdgpu_ring_commit(struct amdgpu_ring *ring)
124 {
125 	uint32_t count;
126 
127 	/* We pad to match fetch size */
128 	count = ring->funcs->align_mask + 1 -
129 		(ring->wptr & ring->funcs->align_mask);
130 	count %= ring->funcs->align_mask + 1;
131 	ring->funcs->insert_nop(ring, count);
132 
133 	mb();
134 	amdgpu_ring_set_wptr(ring);
135 
136 	if (ring->funcs->end_use)
137 		ring->funcs->end_use(ring);
138 
139 	if (ring->funcs->type != AMDGPU_RING_TYPE_KIQ)
140 		amdgpu_ring_lru_touch(ring->adev, ring);
141 }
142 
143 /**
144  * amdgpu_ring_undo - reset the wptr
145  *
146  * @ring: amdgpu_ring structure holding ring information
147  *
148  * Reset the driver's copy of the wptr (all asics).
149  */
150 void amdgpu_ring_undo(struct amdgpu_ring *ring)
151 {
152 	ring->wptr = ring->wptr_old;
153 
154 	if (ring->funcs->end_use)
155 		ring->funcs->end_use(ring);
156 }
157 
158 /**
159  * amdgpu_ring_priority_put - restore a ring's priority
160  *
161  * @ring: amdgpu_ring structure holding the information
162  * @priority: target priority
163  *
164  * Release a request for executing at @priority
165  */
166 void amdgpu_ring_priority_put(struct amdgpu_ring *ring,
167 			      enum drm_sched_priority priority)
168 {
169 	int i;
170 
171 	if (!ring->funcs->set_priority)
172 		return;
173 
174 	if (atomic_dec_return(&ring->num_jobs[priority]) > 0)
175 		return;
176 
177 	/* no need to restore if the job is already at the lowest priority */
178 	if (priority == DRM_SCHED_PRIORITY_NORMAL)
179 		return;
180 
181 	mutex_lock(&ring->priority_mutex);
182 	/* something higher prio is executing, no need to decay */
183 	if (ring->priority > priority)
184 		goto out_unlock;
185 
186 	/* decay priority to the next level with a job available */
187 	for (i = priority; i >= DRM_SCHED_PRIORITY_MIN; i--) {
188 		if (i == DRM_SCHED_PRIORITY_NORMAL
189 				|| atomic_read(&ring->num_jobs[i])) {
190 			ring->priority = i;
191 			ring->funcs->set_priority(ring, i);
192 			break;
193 		}
194 	}
195 
196 out_unlock:
197 	mutex_unlock(&ring->priority_mutex);
198 }
199 
200 /**
201  * amdgpu_ring_priority_get - change the ring's priority
202  *
203  * @ring: amdgpu_ring structure holding the information
204  * @priority: target priority
205  *
206  * Request a ring's priority to be raised to @priority (refcounted).
207  */
208 void amdgpu_ring_priority_get(struct amdgpu_ring *ring,
209 			      enum drm_sched_priority priority)
210 {
211 	if (!ring->funcs->set_priority)
212 		return;
213 
214 	atomic_inc(&ring->num_jobs[priority]);
215 
216 	mutex_lock(&ring->priority_mutex);
217 	if (priority <= ring->priority)
218 		goto out_unlock;
219 
220 	ring->priority = priority;
221 	ring->funcs->set_priority(ring, priority);
222 
223 out_unlock:
224 	mutex_unlock(&ring->priority_mutex);
225 }
226 
227 /**
228  * amdgpu_ring_init - init driver ring struct.
229  *
230  * @adev: amdgpu_device pointer
231  * @ring: amdgpu_ring structure holding ring information
232  * @max_ndw: maximum number of dw for ring alloc
233  * @nop: nop packet for this ring
234  *
235  * Initialize the driver information for the selected ring (all asics).
236  * Returns 0 on success, error on failure.
237  */
238 int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
239 		     unsigned max_dw, struct amdgpu_irq_src *irq_src,
240 		     unsigned irq_type)
241 {
242 	int r, i;
243 	int sched_hw_submission = amdgpu_sched_hw_submission;
244 
245 	/* Set the hw submission limit higher for KIQ because
246 	 * it's used for a number of gfx/compute tasks by both
247 	 * KFD and KGD which may have outstanding fences and
248 	 * it doesn't really use the gpu scheduler anyway;
249 	 * KIQ tasks get submitted directly to the ring.
250 	 */
251 	if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
252 		sched_hw_submission = max(sched_hw_submission, 256);
253 
254 	if (ring->adev == NULL) {
255 		if (adev->num_rings >= AMDGPU_MAX_RINGS)
256 			return -EINVAL;
257 
258 		ring->adev = adev;
259 		ring->idx = adev->num_rings++;
260 		adev->rings[ring->idx] = ring;
261 		r = amdgpu_fence_driver_init_ring(ring, sched_hw_submission);
262 		if (r)
263 			return r;
264 	}
265 
266 	r = amdgpu_device_wb_get(adev, &ring->rptr_offs);
267 	if (r) {
268 		dev_err(adev->dev, "(%d) ring rptr_offs wb alloc failed\n", r);
269 		return r;
270 	}
271 
272 	r = amdgpu_device_wb_get(adev, &ring->wptr_offs);
273 	if (r) {
274 		dev_err(adev->dev, "(%d) ring wptr_offs wb alloc failed\n", r);
275 		return r;
276 	}
277 
278 	r = amdgpu_device_wb_get(adev, &ring->fence_offs);
279 	if (r) {
280 		dev_err(adev->dev, "(%d) ring fence_offs wb alloc failed\n", r);
281 		return r;
282 	}
283 
284 	r = amdgpu_device_wb_get(adev, &ring->cond_exe_offs);
285 	if (r) {
286 		dev_err(adev->dev, "(%d) ring cond_exec_polling wb alloc failed\n", r);
287 		return r;
288 	}
289 	ring->cond_exe_gpu_addr = adev->wb.gpu_addr + (ring->cond_exe_offs * 4);
290 	ring->cond_exe_cpu_addr = &adev->wb.wb[ring->cond_exe_offs];
291 	/* always set cond_exec_polling to CONTINUE */
292 	*ring->cond_exe_cpu_addr = 1;
293 
294 	r = amdgpu_fence_driver_start_ring(ring, irq_src, irq_type);
295 	if (r) {
296 		dev_err(adev->dev, "failed initializing fences (%d).\n", r);
297 		return r;
298 	}
299 
300 	ring->ring_size = roundup_pow_of_two(max_dw * 4 * sched_hw_submission);
301 
302 	ring->buf_mask = (ring->ring_size / 4) - 1;
303 	ring->ptr_mask = ring->funcs->support_64bit_ptrs ?
304 		0xffffffffffffffff : ring->buf_mask;
305 	/* Allocate ring buffer */
306 	if (ring->ring_obj == NULL) {
307 		r = amdgpu_bo_create_kernel(adev, ring->ring_size, PAGE_SIZE,
308 					    AMDGPU_GEM_DOMAIN_GTT,
309 					    &ring->ring_obj,
310 					    &ring->gpu_addr,
311 					    (void **)&ring->ring);
312 		if (r) {
313 			dev_err(adev->dev, "(%d) ring create failed\n", r);
314 			return r;
315 		}
316 		amdgpu_ring_clear_ring(ring);
317 	}
318 
319 	ring->max_dw = max_dw;
320 	ring->priority = DRM_SCHED_PRIORITY_NORMAL;
321 	mutex_init(&ring->priority_mutex);
322 	INIT_LIST_HEAD(&ring->lru_list);
323 	amdgpu_ring_lru_touch(adev, ring);
324 
325 	for (i = 0; i < DRM_SCHED_PRIORITY_MAX; ++i)
326 		atomic_set(&ring->num_jobs[i], 0);
327 
328 	if (amdgpu_debugfs_ring_init(adev, ring)) {
329 		DRM_ERROR("Failed to register debugfs file for rings !\n");
330 	}
331 
332 	return 0;
333 }
334 
335 /**
336  * amdgpu_ring_fini - tear down the driver ring struct.
337  *
338  * @adev: amdgpu_device pointer
339  * @ring: amdgpu_ring structure holding ring information
340  *
341  * Tear down the driver information for the selected ring (all asics).
342  */
343 void amdgpu_ring_fini(struct amdgpu_ring *ring)
344 {
345 	ring->ready = false;
346 
347 	/* Not to finish a ring which is not initialized */
348 	if (!(ring->adev) || !(ring->adev->rings[ring->idx]))
349 		return;
350 
351 	amdgpu_device_wb_free(ring->adev, ring->rptr_offs);
352 	amdgpu_device_wb_free(ring->adev, ring->wptr_offs);
353 
354 	amdgpu_device_wb_free(ring->adev, ring->cond_exe_offs);
355 	amdgpu_device_wb_free(ring->adev, ring->fence_offs);
356 
357 	amdgpu_bo_free_kernel(&ring->ring_obj,
358 			      &ring->gpu_addr,
359 			      (void **)&ring->ring);
360 
361 	amdgpu_debugfs_ring_fini(ring);
362 
363 	ring->adev->rings[ring->idx] = NULL;
364 }
365 
366 static void amdgpu_ring_lru_touch_locked(struct amdgpu_device *adev,
367 					 struct amdgpu_ring *ring)
368 {
369 	/* list_move_tail handles the case where ring isn't part of the list */
370 	list_move_tail(&ring->lru_list, &adev->ring_lru_list);
371 }
372 
373 static bool amdgpu_ring_is_blacklisted(struct amdgpu_ring *ring,
374 				       int *blacklist, int num_blacklist)
375 {
376 	int i;
377 
378 	for (i = 0; i < num_blacklist; i++) {
379 		if (ring->idx == blacklist[i])
380 			return true;
381 	}
382 
383 	return false;
384 }
385 
386 /**
387  * amdgpu_ring_lru_get - get the least recently used ring for a HW IP block
388  *
389  * @adev: amdgpu_device pointer
390  * @type: amdgpu_ring_type enum
391  * @blacklist: blacklisted ring ids array
392  * @num_blacklist: number of entries in @blacklist
393  * @lru_pipe_order: find a ring from the least recently used pipe
394  * @ring: output ring
395  *
396  * Retrieve the amdgpu_ring structure for the least recently used ring of
397  * a specific IP block (all asics).
398  * Returns 0 on success, error on failure.
399  */
400 int amdgpu_ring_lru_get(struct amdgpu_device *adev, int type,
401 			int *blacklist,	int num_blacklist,
402 			bool lru_pipe_order, struct amdgpu_ring **ring)
403 {
404 	struct amdgpu_ring *entry;
405 
406 	/* List is sorted in LRU order, find first entry corresponding
407 	 * to the desired HW IP */
408 	*ring = NULL;
409 	spin_lock(&adev->ring_lru_list_lock);
410 	list_for_each_entry(entry, &adev->ring_lru_list, lru_list) {
411 		if (entry->funcs->type != type)
412 			continue;
413 
414 		if (amdgpu_ring_is_blacklisted(entry, blacklist, num_blacklist))
415 			continue;
416 
417 		if (!*ring) {
418 			*ring = entry;
419 
420 			/* We are done for ring LRU */
421 			if (!lru_pipe_order)
422 				break;
423 		}
424 
425 		/* Move all rings on the same pipe to the end of the list */
426 		if (entry->pipe == (*ring)->pipe)
427 			amdgpu_ring_lru_touch_locked(adev, entry);
428 	}
429 
430 	/* Move the ring we found to the end of the list */
431 	if (*ring)
432 		amdgpu_ring_lru_touch_locked(adev, *ring);
433 
434 	spin_unlock(&adev->ring_lru_list_lock);
435 
436 	if (!*ring) {
437 		DRM_ERROR("Ring LRU contains no entries for ring type:%d\n", type);
438 		return -EINVAL;
439 	}
440 
441 	return 0;
442 }
443 
444 /**
445  * amdgpu_ring_lru_touch - mark a ring as recently being used
446  *
447  * @adev: amdgpu_device pointer
448  * @ring: ring to touch
449  *
450  * Move @ring to the tail of the lru list
451  */
452 void amdgpu_ring_lru_touch(struct amdgpu_device *adev, struct amdgpu_ring *ring)
453 {
454 	spin_lock(&adev->ring_lru_list_lock);
455 	amdgpu_ring_lru_touch_locked(adev, ring);
456 	spin_unlock(&adev->ring_lru_list_lock);
457 }
458 
459 /*
460  * Debugfs info
461  */
462 #if defined(CONFIG_DEBUG_FS)
463 
464 /* Layout of file is 12 bytes consisting of
465  * - rptr
466  * - wptr
467  * - driver's copy of wptr
468  *
469  * followed by n-words of ring data
470  */
471 static ssize_t amdgpu_debugfs_ring_read(struct file *f, char __user *buf,
472 					size_t size, loff_t *pos)
473 {
474 	struct amdgpu_ring *ring = file_inode(f)->i_private;
475 	int r, i;
476 	uint32_t value, result, early[3];
477 
478 	if (*pos & 3 || size & 3)
479 		return -EINVAL;
480 
481 	result = 0;
482 
483 	if (*pos < 12) {
484 		early[0] = amdgpu_ring_get_rptr(ring) & ring->buf_mask;
485 		early[1] = amdgpu_ring_get_wptr(ring) & ring->buf_mask;
486 		early[2] = ring->wptr & ring->buf_mask;
487 		for (i = *pos / 4; i < 3 && size; i++) {
488 			r = put_user(early[i], (uint32_t *)buf);
489 			if (r)
490 				return r;
491 			buf += 4;
492 			result += 4;
493 			size -= 4;
494 			*pos += 4;
495 		}
496 	}
497 
498 	while (size) {
499 		if (*pos >= (ring->ring_size + 12))
500 			return result;
501 
502 		value = ring->ring[(*pos - 12)/4];
503 		r = put_user(value, (uint32_t*)buf);
504 		if (r)
505 			return r;
506 		buf += 4;
507 		result += 4;
508 		size -= 4;
509 		*pos += 4;
510 	}
511 
512 	return result;
513 }
514 
515 static const struct file_operations amdgpu_debugfs_ring_fops = {
516 	.owner = THIS_MODULE,
517 	.read = amdgpu_debugfs_ring_read,
518 	.llseek = default_llseek
519 };
520 
521 #endif
522 
523 static int amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
524 				    struct amdgpu_ring *ring)
525 {
526 #if defined(CONFIG_DEBUG_FS)
527 	struct drm_minor *minor = adev->ddev->primary;
528 	struct dentry *ent, *root = minor->debugfs_root;
529 	char name[32];
530 
531 	sprintf(name, "amdgpu_ring_%s", ring->name);
532 
533 	ent = debugfs_create_file(name,
534 				  S_IFREG | S_IRUGO, root,
535 				  ring, &amdgpu_debugfs_ring_fops);
536 	if (!ent)
537 		return -ENOMEM;
538 
539 	i_size_write(ent->d_inode, ring->ring_size + 12);
540 	ring->ent = ent;
541 #endif
542 	return 0;
543 }
544 
545 static void amdgpu_debugfs_ring_fini(struct amdgpu_ring *ring)
546 {
547 #if defined(CONFIG_DEBUG_FS)
548 	debugfs_remove(ring->ent);
549 #endif
550 }
551