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 
140 /**
141  * amdgpu_ring_undo - reset the wptr
142  *
143  * @ring: amdgpu_ring structure holding ring information
144  *
145  * Reset the driver's copy of the wptr (all asics).
146  */
147 void amdgpu_ring_undo(struct amdgpu_ring *ring)
148 {
149 	ring->wptr = ring->wptr_old;
150 
151 	if (ring->funcs->end_use)
152 		ring->funcs->end_use(ring);
153 }
154 
155 /**
156  * amdgpu_ring_priority_put - restore a ring's priority
157  *
158  * @ring: amdgpu_ring structure holding the information
159  * @priority: target priority
160  *
161  * Release a request for executing at @priority
162  */
163 void amdgpu_ring_priority_put(struct amdgpu_ring *ring,
164 			      enum drm_sched_priority priority)
165 {
166 	int i;
167 
168 	if (!ring->funcs->set_priority)
169 		return;
170 
171 	if (atomic_dec_return(&ring->num_jobs[priority]) > 0)
172 		return;
173 
174 	/* no need to restore if the job is already at the lowest priority */
175 	if (priority == DRM_SCHED_PRIORITY_NORMAL)
176 		return;
177 
178 	mutex_lock(&ring->priority_mutex);
179 	/* something higher prio is executing, no need to decay */
180 	if (ring->priority > priority)
181 		goto out_unlock;
182 
183 	/* decay priority to the next level with a job available */
184 	for (i = priority; i >= DRM_SCHED_PRIORITY_MIN; i--) {
185 		if (i == DRM_SCHED_PRIORITY_NORMAL
186 				|| atomic_read(&ring->num_jobs[i])) {
187 			ring->priority = i;
188 			ring->funcs->set_priority(ring, i);
189 			break;
190 		}
191 	}
192 
193 out_unlock:
194 	mutex_unlock(&ring->priority_mutex);
195 }
196 
197 /**
198  * amdgpu_ring_priority_get - change the ring's priority
199  *
200  * @ring: amdgpu_ring structure holding the information
201  * @priority: target priority
202  *
203  * Request a ring's priority to be raised to @priority (refcounted).
204  */
205 void amdgpu_ring_priority_get(struct amdgpu_ring *ring,
206 			      enum drm_sched_priority priority)
207 {
208 	if (!ring->funcs->set_priority)
209 		return;
210 
211 	if (atomic_inc_return(&ring->num_jobs[priority]) <= 0)
212 		return;
213 
214 	mutex_lock(&ring->priority_mutex);
215 	if (priority <= ring->priority)
216 		goto out_unlock;
217 
218 	ring->priority = priority;
219 	ring->funcs->set_priority(ring, priority);
220 
221 out_unlock:
222 	mutex_unlock(&ring->priority_mutex);
223 }
224 
225 /**
226  * amdgpu_ring_init - init driver ring struct.
227  *
228  * @adev: amdgpu_device pointer
229  * @ring: amdgpu_ring structure holding ring information
230  * @max_ndw: maximum number of dw for ring alloc
231  * @nop: nop packet for this ring
232  *
233  * Initialize the driver information for the selected ring (all asics).
234  * Returns 0 on success, error on failure.
235  */
236 int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
237 		     unsigned max_dw, struct amdgpu_irq_src *irq_src,
238 		     unsigned irq_type)
239 {
240 	int r, i;
241 	int sched_hw_submission = amdgpu_sched_hw_submission;
242 
243 	/* Set the hw submission limit higher for KIQ because
244 	 * it's used for a number of gfx/compute tasks by both
245 	 * KFD and KGD which may have outstanding fences and
246 	 * it doesn't really use the gpu scheduler anyway;
247 	 * KIQ tasks get submitted directly to the ring.
248 	 */
249 	if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
250 		sched_hw_submission = max(sched_hw_submission, 256);
251 	else if (ring == &adev->sdma.instance[0].page)
252 		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 + ring->funcs->extra_dw, 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 
323 	for (i = 0; i < DRM_SCHED_PRIORITY_MAX; ++i)
324 		atomic_set(&ring->num_jobs[i], 0);
325 
326 	if (amdgpu_debugfs_ring_init(adev, ring)) {
327 		DRM_ERROR("Failed to register debugfs file for rings !\n");
328 	}
329 
330 	return 0;
331 }
332 
333 /**
334  * amdgpu_ring_fini - tear down the driver ring struct.
335  *
336  * @adev: amdgpu_device pointer
337  * @ring: amdgpu_ring structure holding ring information
338  *
339  * Tear down the driver information for the selected ring (all asics).
340  */
341 void amdgpu_ring_fini(struct amdgpu_ring *ring)
342 {
343 	ring->sched.ready = false;
344 
345 	/* Not to finish a ring which is not initialized */
346 	if (!(ring->adev) || !(ring->adev->rings[ring->idx]))
347 		return;
348 
349 	amdgpu_device_wb_free(ring->adev, ring->rptr_offs);
350 	amdgpu_device_wb_free(ring->adev, ring->wptr_offs);
351 
352 	amdgpu_device_wb_free(ring->adev, ring->cond_exe_offs);
353 	amdgpu_device_wb_free(ring->adev, ring->fence_offs);
354 
355 	amdgpu_bo_free_kernel(&ring->ring_obj,
356 			      &ring->gpu_addr,
357 			      (void **)&ring->ring);
358 
359 	amdgpu_debugfs_ring_fini(ring);
360 
361 	dma_fence_put(ring->vmid_wait);
362 	ring->vmid_wait = NULL;
363 	ring->me = 0;
364 
365 	ring->adev->rings[ring->idx] = NULL;
366 }
367 
368 /**
369  * amdgpu_ring_emit_reg_write_reg_wait_helper - ring helper
370  *
371  * @adev: amdgpu_device pointer
372  * @reg0: register to write
373  * @reg1: register to wait on
374  * @ref: reference value to write/wait on
375  * @mask: mask to wait on
376  *
377  * Helper for rings that don't support write and wait in a
378  * single oneshot packet.
379  */
380 void amdgpu_ring_emit_reg_write_reg_wait_helper(struct amdgpu_ring *ring,
381 						uint32_t reg0, uint32_t reg1,
382 						uint32_t ref, uint32_t mask)
383 {
384 	amdgpu_ring_emit_wreg(ring, reg0, ref);
385 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
386 }
387 
388 /**
389  * amdgpu_ring_soft_recovery - try to soft recover a ring lockup
390  *
391  * @ring: ring to try the recovery on
392  * @vmid: VMID we try to get going again
393  * @fence: timedout fence
394  *
395  * Tries to get a ring proceeding again when it is stuck.
396  */
397 bool amdgpu_ring_soft_recovery(struct amdgpu_ring *ring, unsigned int vmid,
398 			       struct dma_fence *fence)
399 {
400 	ktime_t deadline = ktime_add_us(ktime_get(), 10000);
401 
402 	if (!ring->funcs->soft_recovery || !fence)
403 		return false;
404 
405 	atomic_inc(&ring->adev->gpu_reset_counter);
406 	while (!dma_fence_is_signaled(fence) &&
407 	       ktime_to_ns(ktime_sub(deadline, ktime_get())) > 0)
408 		ring->funcs->soft_recovery(ring, vmid);
409 
410 	return dma_fence_is_signaled(fence);
411 }
412 
413 /*
414  * Debugfs info
415  */
416 #if defined(CONFIG_DEBUG_FS)
417 
418 /* Layout of file is 12 bytes consisting of
419  * - rptr
420  * - wptr
421  * - driver's copy of wptr
422  *
423  * followed by n-words of ring data
424  */
425 static ssize_t amdgpu_debugfs_ring_read(struct file *f, char __user *buf,
426 					size_t size, loff_t *pos)
427 {
428 	struct amdgpu_ring *ring = file_inode(f)->i_private;
429 	int r, i;
430 	uint32_t value, result, early[3];
431 
432 	if (*pos & 3 || size & 3)
433 		return -EINVAL;
434 
435 	result = 0;
436 
437 	if (*pos < 12) {
438 		early[0] = amdgpu_ring_get_rptr(ring) & ring->buf_mask;
439 		early[1] = amdgpu_ring_get_wptr(ring) & ring->buf_mask;
440 		early[2] = ring->wptr & ring->buf_mask;
441 		for (i = *pos / 4; i < 3 && size; i++) {
442 			r = put_user(early[i], (uint32_t *)buf);
443 			if (r)
444 				return r;
445 			buf += 4;
446 			result += 4;
447 			size -= 4;
448 			*pos += 4;
449 		}
450 	}
451 
452 	while (size) {
453 		if (*pos >= (ring->ring_size + 12))
454 			return result;
455 
456 		value = ring->ring[(*pos - 12)/4];
457 		r = put_user(value, (uint32_t*)buf);
458 		if (r)
459 			return r;
460 		buf += 4;
461 		result += 4;
462 		size -= 4;
463 		*pos += 4;
464 	}
465 
466 	return result;
467 }
468 
469 static const struct file_operations amdgpu_debugfs_ring_fops = {
470 	.owner = THIS_MODULE,
471 	.read = amdgpu_debugfs_ring_read,
472 	.llseek = default_llseek
473 };
474 
475 #endif
476 
477 static int amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
478 				    struct amdgpu_ring *ring)
479 {
480 #if defined(CONFIG_DEBUG_FS)
481 	struct drm_minor *minor = adev->ddev->primary;
482 	struct dentry *ent, *root = minor->debugfs_root;
483 	char name[32];
484 
485 	sprintf(name, "amdgpu_ring_%s", ring->name);
486 
487 	ent = debugfs_create_file(name,
488 				  S_IFREG | S_IRUGO, root,
489 				  ring, &amdgpu_debugfs_ring_fops);
490 	if (!ent)
491 		return -ENOMEM;
492 
493 	i_size_write(ent->d_inode, ring->ring_size + 12);
494 	ring->ent = ent;
495 #endif
496 	return 0;
497 }
498 
499 static void amdgpu_debugfs_ring_fini(struct amdgpu_ring *ring)
500 {
501 #if defined(CONFIG_DEBUG_FS)
502 	debugfs_remove(ring->ent);
503 #endif
504 }
505 
506 /**
507  * amdgpu_ring_test_helper - tests ring and set sched readiness status
508  *
509  * @ring: ring to try the recovery on
510  *
511  * Tests ring and set sched readiness status
512  *
513  * Returns 0 on success, error on failure.
514  */
515 int amdgpu_ring_test_helper(struct amdgpu_ring *ring)
516 {
517 	struct amdgpu_device *adev = ring->adev;
518 	int r;
519 
520 	r = amdgpu_ring_test_ring(ring);
521 	if (r)
522 		DRM_DEV_ERROR(adev->dev, "ring %s test failed (%d)\n",
523 			      ring->name, r);
524 	else
525 		DRM_DEV_DEBUG(adev->dev, "ring test on %s succeeded\n",
526 			      ring->name);
527 
528 	ring->sched.ready = !r;
529 	return r;
530 }
531