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/uaccess.h>
32 #include <linux/debugfs.h>
33 
34 #include <drm/amdgpu_drm.h>
35 #include "amdgpu.h"
36 #include "atom.h"
37 
38 /*
39  * Rings
40  * Most engines on the GPU are fed via ring buffers.  Ring
41  * buffers are areas of GPU accessible memory that the host
42  * writes commands into and the GPU reads commands out of.
43  * There is a rptr (read pointer) that determines where the
44  * GPU is currently reading, and a wptr (write pointer)
45  * which determines where the host has written.  When the
46  * pointers are equal, the ring is idle.  When the host
47  * writes commands to the ring buffer, it increments the
48  * wptr.  The GPU then starts fetching commands and executes
49  * them until the pointers are equal again.
50  */
51 
52 /**
53  * amdgpu_ring_alloc - allocate space on the ring buffer
54  *
55  * @ring: amdgpu_ring structure holding ring information
56  * @ndw: number of dwords to allocate in the ring buffer
57  *
58  * Allocate @ndw dwords in the ring buffer (all asics).
59  * Returns 0 on success, error on failure.
60  */
61 int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw)
62 {
63 	/* Align requested size with padding so unlock_commit can
64 	 * pad safely */
65 	ndw = (ndw + ring->funcs->align_mask) & ~ring->funcs->align_mask;
66 
67 	/* Make sure we aren't trying to allocate more space
68 	 * than the maximum for one submission
69 	 */
70 	if (WARN_ON_ONCE(ndw > ring->max_dw))
71 		return -ENOMEM;
72 
73 	ring->count_dw = ndw;
74 	ring->wptr_old = ring->wptr;
75 
76 	if (ring->funcs->begin_use)
77 		ring->funcs->begin_use(ring);
78 
79 	return 0;
80 }
81 
82 /** amdgpu_ring_insert_nop - insert NOP packets
83  *
84  * @ring: amdgpu_ring structure holding ring information
85  * @count: the number of NOP packets to insert
86  *
87  * This is the generic insert_nop function for rings except SDMA
88  */
89 void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
90 {
91 	int i;
92 
93 	for (i = 0; i < count; i++)
94 		amdgpu_ring_write(ring, ring->funcs->nop);
95 }
96 
97 /**
98  * amdgpu_ring_generic_pad_ib - pad IB with NOP packets
99  *
100  * @ring: amdgpu_ring structure holding ring information
101  * @ib: IB to add NOP packets to
102  *
103  * This is the generic pad_ib function for rings except SDMA
104  */
105 void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
106 {
107 	while (ib->length_dw & ring->funcs->align_mask)
108 		ib->ptr[ib->length_dw++] = ring->funcs->nop;
109 }
110 
111 /**
112  * amdgpu_ring_commit - tell the GPU to execute the new
113  * commands on the ring buffer
114  *
115  * @ring: amdgpu_ring structure holding ring information
116  *
117  * Update the wptr (write pointer) to tell the GPU to
118  * execute new commands on the ring buffer (all asics).
119  */
120 void amdgpu_ring_commit(struct amdgpu_ring *ring)
121 {
122 	uint32_t count;
123 
124 	/* We pad to match fetch size */
125 	count = ring->funcs->align_mask + 1 -
126 		(ring->wptr & ring->funcs->align_mask);
127 	count %= ring->funcs->align_mask + 1;
128 	ring->funcs->insert_nop(ring, count);
129 
130 	mb();
131 	amdgpu_ring_set_wptr(ring);
132 
133 	if (ring->funcs->end_use)
134 		ring->funcs->end_use(ring);
135 }
136 
137 /**
138  * amdgpu_ring_undo - reset the wptr
139  *
140  * @ring: amdgpu_ring structure holding ring information
141  *
142  * Reset the driver's copy of the wptr (all asics).
143  */
144 void amdgpu_ring_undo(struct amdgpu_ring *ring)
145 {
146 	ring->wptr = ring->wptr_old;
147 
148 	if (ring->funcs->end_use)
149 		ring->funcs->end_use(ring);
150 }
151 
152 /**
153  * amdgpu_ring_init - init driver ring struct.
154  *
155  * @adev: amdgpu_device pointer
156  * @ring: amdgpu_ring structure holding ring information
157  * @max_dw: maximum number of dw for ring alloc
158  * @irq_src: interrupt source to use for this ring
159  * @irq_type: interrupt type to use for this ring
160  * @hw_prio: ring priority (NORMAL/HIGH)
161  *
162  * Initialize the driver information for the selected ring (all asics).
163  * Returns 0 on success, error on failure.
164  */
165 int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
166 		     unsigned int max_dw, struct amdgpu_irq_src *irq_src,
167 		     unsigned int irq_type, unsigned int hw_prio,
168 		     atomic_t *sched_score)
169 {
170 	int r;
171 	int sched_hw_submission = amdgpu_sched_hw_submission;
172 	u32 *num_sched;
173 	u32 hw_ip;
174 
175 	/* Set the hw submission limit higher for KIQ because
176 	 * it's used for a number of gfx/compute tasks by both
177 	 * KFD and KGD which may have outstanding fences and
178 	 * it doesn't really use the gpu scheduler anyway;
179 	 * KIQ tasks get submitted directly to the ring.
180 	 */
181 	if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
182 		sched_hw_submission = max(sched_hw_submission, 256);
183 	else if (ring == &adev->sdma.instance[0].page)
184 		sched_hw_submission = 256;
185 
186 	if (ring->adev == NULL) {
187 		if (adev->num_rings >= AMDGPU_MAX_RINGS)
188 			return -EINVAL;
189 
190 		ring->adev = adev;
191 		ring->idx = adev->num_rings++;
192 		adev->rings[ring->idx] = ring;
193 		r = amdgpu_fence_driver_init_ring(ring, sched_hw_submission,
194 						  sched_score);
195 		if (r)
196 			return r;
197 	}
198 
199 	r = amdgpu_device_wb_get(adev, &ring->rptr_offs);
200 	if (r) {
201 		dev_err(adev->dev, "(%d) ring rptr_offs wb alloc failed\n", r);
202 		return r;
203 	}
204 
205 	r = amdgpu_device_wb_get(adev, &ring->wptr_offs);
206 	if (r) {
207 		dev_err(adev->dev, "(%d) ring wptr_offs wb alloc failed\n", r);
208 		return r;
209 	}
210 
211 	r = amdgpu_device_wb_get(adev, &ring->fence_offs);
212 	if (r) {
213 		dev_err(adev->dev, "(%d) ring fence_offs wb alloc failed\n", r);
214 		return r;
215 	}
216 
217 	r = amdgpu_device_wb_get(adev, &ring->trail_fence_offs);
218 	if (r) {
219 		dev_err(adev->dev,
220 			"(%d) ring trail_fence_offs wb alloc failed\n", r);
221 		return r;
222 	}
223 	ring->trail_fence_gpu_addr =
224 		adev->wb.gpu_addr + (ring->trail_fence_offs * 4);
225 	ring->trail_fence_cpu_addr = &adev->wb.wb[ring->trail_fence_offs];
226 
227 	r = amdgpu_device_wb_get(adev, &ring->cond_exe_offs);
228 	if (r) {
229 		dev_err(adev->dev, "(%d) ring cond_exec_polling wb alloc failed\n", r);
230 		return r;
231 	}
232 	ring->cond_exe_gpu_addr = adev->wb.gpu_addr + (ring->cond_exe_offs * 4);
233 	ring->cond_exe_cpu_addr = &adev->wb.wb[ring->cond_exe_offs];
234 	/* always set cond_exec_polling to CONTINUE */
235 	*ring->cond_exe_cpu_addr = 1;
236 
237 	r = amdgpu_fence_driver_start_ring(ring, irq_src, irq_type);
238 	if (r) {
239 		dev_err(adev->dev, "failed initializing fences (%d).\n", r);
240 		return r;
241 	}
242 
243 	ring->ring_size = roundup_pow_of_two(max_dw * 4 * sched_hw_submission);
244 
245 	ring->buf_mask = (ring->ring_size / 4) - 1;
246 	ring->ptr_mask = ring->funcs->support_64bit_ptrs ?
247 		0xffffffffffffffff : ring->buf_mask;
248 	/* Allocate ring buffer */
249 	if (ring->ring_obj == NULL) {
250 		r = amdgpu_bo_create_kernel(adev, ring->ring_size + ring->funcs->extra_dw, PAGE_SIZE,
251 					    AMDGPU_GEM_DOMAIN_GTT,
252 					    &ring->ring_obj,
253 					    &ring->gpu_addr,
254 					    (void **)&ring->ring);
255 		if (r) {
256 			dev_err(adev->dev, "(%d) ring create failed\n", r);
257 			return r;
258 		}
259 		amdgpu_ring_clear_ring(ring);
260 	}
261 
262 	ring->max_dw = max_dw;
263 	ring->hw_prio = hw_prio;
264 
265 	if (!ring->no_scheduler) {
266 		hw_ip = ring->funcs->type;
267 		num_sched = &adev->gpu_sched[hw_ip][hw_prio].num_scheds;
268 		adev->gpu_sched[hw_ip][hw_prio].sched[(*num_sched)++] =
269 			&ring->sched;
270 	}
271 
272 	return 0;
273 }
274 
275 /**
276  * amdgpu_ring_fini - tear down the driver ring struct.
277  *
278  * @ring: amdgpu_ring structure holding ring information
279  *
280  * Tear down the driver information for the selected ring (all asics).
281  */
282 void amdgpu_ring_fini(struct amdgpu_ring *ring)
283 {
284 
285 	/* Not to finish a ring which is not initialized */
286 	if (!(ring->adev) || !(ring->adev->rings[ring->idx]))
287 		return;
288 
289 	ring->sched.ready = false;
290 
291 	amdgpu_device_wb_free(ring->adev, ring->rptr_offs);
292 	amdgpu_device_wb_free(ring->adev, ring->wptr_offs);
293 
294 	amdgpu_device_wb_free(ring->adev, ring->cond_exe_offs);
295 	amdgpu_device_wb_free(ring->adev, ring->fence_offs);
296 
297 	amdgpu_bo_free_kernel(&ring->ring_obj,
298 			      &ring->gpu_addr,
299 			      (void **)&ring->ring);
300 
301 	dma_fence_put(ring->vmid_wait);
302 	ring->vmid_wait = NULL;
303 	ring->me = 0;
304 
305 	ring->adev->rings[ring->idx] = NULL;
306 }
307 
308 /**
309  * amdgpu_ring_emit_reg_write_reg_wait_helper - ring helper
310  *
311  * @ring: ring to write to
312  * @reg0: register to write
313  * @reg1: register to wait on
314  * @ref: reference value to write/wait on
315  * @mask: mask to wait on
316  *
317  * Helper for rings that don't support write and wait in a
318  * single oneshot packet.
319  */
320 void amdgpu_ring_emit_reg_write_reg_wait_helper(struct amdgpu_ring *ring,
321 						uint32_t reg0, uint32_t reg1,
322 						uint32_t ref, uint32_t mask)
323 {
324 	amdgpu_ring_emit_wreg(ring, reg0, ref);
325 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
326 }
327 
328 /**
329  * amdgpu_ring_soft_recovery - try to soft recover a ring lockup
330  *
331  * @ring: ring to try the recovery on
332  * @vmid: VMID we try to get going again
333  * @fence: timedout fence
334  *
335  * Tries to get a ring proceeding again when it is stuck.
336  */
337 bool amdgpu_ring_soft_recovery(struct amdgpu_ring *ring, unsigned int vmid,
338 			       struct dma_fence *fence)
339 {
340 	ktime_t deadline = ktime_add_us(ktime_get(), 10000);
341 
342 	if (amdgpu_sriov_vf(ring->adev) || !ring->funcs->soft_recovery || !fence)
343 		return false;
344 
345 	atomic_inc(&ring->adev->gpu_reset_counter);
346 	while (!dma_fence_is_signaled(fence) &&
347 	       ktime_to_ns(ktime_sub(deadline, ktime_get())) > 0)
348 		ring->funcs->soft_recovery(ring, vmid);
349 
350 	return dma_fence_is_signaled(fence);
351 }
352 
353 /*
354  * Debugfs info
355  */
356 #if defined(CONFIG_DEBUG_FS)
357 
358 /* Layout of file is 12 bytes consisting of
359  * - rptr
360  * - wptr
361  * - driver's copy of wptr
362  *
363  * followed by n-words of ring data
364  */
365 static ssize_t amdgpu_debugfs_ring_read(struct file *f, char __user *buf,
366 					size_t size, loff_t *pos)
367 {
368 	struct amdgpu_ring *ring = file_inode(f)->i_private;
369 	int r, i;
370 	uint32_t value, result, early[3];
371 
372 	if (*pos & 3 || size & 3)
373 		return -EINVAL;
374 
375 	result = 0;
376 
377 	if (*pos < 12) {
378 		early[0] = amdgpu_ring_get_rptr(ring) & ring->buf_mask;
379 		early[1] = amdgpu_ring_get_wptr(ring) & ring->buf_mask;
380 		early[2] = ring->wptr & ring->buf_mask;
381 		for (i = *pos / 4; i < 3 && size; i++) {
382 			r = put_user(early[i], (uint32_t *)buf);
383 			if (r)
384 				return r;
385 			buf += 4;
386 			result += 4;
387 			size -= 4;
388 			*pos += 4;
389 		}
390 	}
391 
392 	while (size) {
393 		if (*pos >= (ring->ring_size + 12))
394 			return result;
395 
396 		value = ring->ring[(*pos - 12)/4];
397 		r = put_user(value, (uint32_t *)buf);
398 		if (r)
399 			return r;
400 		buf += 4;
401 		result += 4;
402 		size -= 4;
403 		*pos += 4;
404 	}
405 
406 	return result;
407 }
408 
409 static const struct file_operations amdgpu_debugfs_ring_fops = {
410 	.owner = THIS_MODULE,
411 	.read = amdgpu_debugfs_ring_read,
412 	.llseek = default_llseek
413 };
414 
415 #endif
416 
417 int amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
418 			     struct amdgpu_ring *ring)
419 {
420 #if defined(CONFIG_DEBUG_FS)
421 	struct drm_minor *minor = adev_to_drm(adev)->primary;
422 	struct dentry *ent, *root = minor->debugfs_root;
423 	char name[32];
424 
425 	sprintf(name, "amdgpu_ring_%s", ring->name);
426 
427 	ent = debugfs_create_file(name,
428 				  S_IFREG | S_IRUGO, root,
429 				  ring, &amdgpu_debugfs_ring_fops);
430 	if (!ent)
431 		return -ENOMEM;
432 
433 	i_size_write(ent->d_inode, ring->ring_size + 12);
434 	ring->ent = ent;
435 #endif
436 	return 0;
437 }
438 
439 /**
440  * amdgpu_ring_test_helper - tests ring and set sched readiness status
441  *
442  * @ring: ring to try the recovery on
443  *
444  * Tests ring and set sched readiness status
445  *
446  * Returns 0 on success, error on failure.
447  */
448 int amdgpu_ring_test_helper(struct amdgpu_ring *ring)
449 {
450 	struct amdgpu_device *adev = ring->adev;
451 	int r;
452 
453 	r = amdgpu_ring_test_ring(ring);
454 	if (r)
455 		DRM_DEV_ERROR(adev->dev, "ring %s test failed (%d)\n",
456 			      ring->name, r);
457 	else
458 		DRM_DEV_DEBUG(adev->dev, "ring test on %s succeeded\n",
459 			      ring->name);
460 
461 	ring->sched.ready = !r;
462 	return r;
463 }
464