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  * @sched_score: optional score atomic shared with other schedulers
162  *
163  * Initialize the driver information for the selected ring (all asics).
164  * Returns 0 on success, error on failure.
165  */
166 int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
167 		     unsigned int max_dw, struct amdgpu_irq_src *irq_src,
168 		     unsigned int irq_type, unsigned int hw_prio,
169 		     atomic_t *sched_score)
170 {
171 	int r;
172 	int sched_hw_submission = amdgpu_sched_hw_submission;
173 	u32 *num_sched;
174 	u32 hw_ip;
175 
176 	/* Set the hw submission limit higher for KIQ because
177 	 * it's used for a number of gfx/compute tasks by both
178 	 * KFD and KGD which may have outstanding fences and
179 	 * it doesn't really use the gpu scheduler anyway;
180 	 * KIQ tasks get submitted directly to the ring.
181 	 */
182 	if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
183 		sched_hw_submission = max(sched_hw_submission, 256);
184 	else if (ring == &adev->sdma.instance[0].page)
185 		sched_hw_submission = 256;
186 
187 	if (ring->adev == NULL) {
188 		if (adev->num_rings >= AMDGPU_MAX_RINGS)
189 			return -EINVAL;
190 
191 		ring->adev = adev;
192 		ring->idx = adev->num_rings++;
193 		adev->rings[ring->idx] = ring;
194 		r = amdgpu_fence_driver_init_ring(ring, sched_hw_submission,
195 						  sched_score);
196 		if (r)
197 			return r;
198 	}
199 
200 	r = amdgpu_device_wb_get(adev, &ring->rptr_offs);
201 	if (r) {
202 		dev_err(adev->dev, "(%d) ring rptr_offs wb alloc failed\n", r);
203 		return r;
204 	}
205 
206 	r = amdgpu_device_wb_get(adev, &ring->wptr_offs);
207 	if (r) {
208 		dev_err(adev->dev, "(%d) ring wptr_offs wb alloc failed\n", r);
209 		return r;
210 	}
211 
212 	r = amdgpu_device_wb_get(adev, &ring->fence_offs);
213 	if (r) {
214 		dev_err(adev->dev, "(%d) ring fence_offs wb alloc failed\n", r);
215 		return r;
216 	}
217 
218 	r = amdgpu_device_wb_get(adev, &ring->trail_fence_offs);
219 	if (r) {
220 		dev_err(adev->dev,
221 			"(%d) ring trail_fence_offs wb alloc failed\n", r);
222 		return r;
223 	}
224 	ring->trail_fence_gpu_addr =
225 		adev->wb.gpu_addr + (ring->trail_fence_offs * 4);
226 	ring->trail_fence_cpu_addr = &adev->wb.wb[ring->trail_fence_offs];
227 
228 	r = amdgpu_device_wb_get(adev, &ring->cond_exe_offs);
229 	if (r) {
230 		dev_err(adev->dev, "(%d) ring cond_exec_polling wb alloc failed\n", r);
231 		return r;
232 	}
233 	ring->cond_exe_gpu_addr = adev->wb.gpu_addr + (ring->cond_exe_offs * 4);
234 	ring->cond_exe_cpu_addr = &adev->wb.wb[ring->cond_exe_offs];
235 	/* always set cond_exec_polling to CONTINUE */
236 	*ring->cond_exe_cpu_addr = 1;
237 
238 	r = amdgpu_fence_driver_start_ring(ring, irq_src, irq_type);
239 	if (r) {
240 		dev_err(adev->dev, "failed initializing fences (%d).\n", r);
241 		return r;
242 	}
243 
244 	ring->ring_size = roundup_pow_of_two(max_dw * 4 * sched_hw_submission);
245 
246 	ring->buf_mask = (ring->ring_size / 4) - 1;
247 	ring->ptr_mask = ring->funcs->support_64bit_ptrs ?
248 		0xffffffffffffffff : ring->buf_mask;
249 	/* Allocate ring buffer */
250 	if (ring->ring_obj == NULL) {
251 		r = amdgpu_bo_create_kernel(adev, ring->ring_size + ring->funcs->extra_dw, PAGE_SIZE,
252 					    AMDGPU_GEM_DOMAIN_GTT,
253 					    &ring->ring_obj,
254 					    &ring->gpu_addr,
255 					    (void **)&ring->ring);
256 		if (r) {
257 			dev_err(adev->dev, "(%d) ring create failed\n", r);
258 			return r;
259 		}
260 		amdgpu_ring_clear_ring(ring);
261 	}
262 
263 	ring->max_dw = max_dw;
264 	ring->hw_prio = hw_prio;
265 
266 	if (!ring->no_scheduler) {
267 		hw_ip = ring->funcs->type;
268 		num_sched = &adev->gpu_sched[hw_ip][hw_prio].num_scheds;
269 		adev->gpu_sched[hw_ip][hw_prio].sched[(*num_sched)++] =
270 			&ring->sched;
271 	}
272 
273 	return 0;
274 }
275 
276 /**
277  * amdgpu_ring_fini - tear down the driver ring struct.
278  *
279  * @ring: amdgpu_ring structure holding ring information
280  *
281  * Tear down the driver information for the selected ring (all asics).
282  */
283 void amdgpu_ring_fini(struct amdgpu_ring *ring)
284 {
285 
286 	/* Not to finish a ring which is not initialized */
287 	if (!(ring->adev) || !(ring->adev->rings[ring->idx]))
288 		return;
289 
290 	ring->sched.ready = false;
291 
292 	amdgpu_device_wb_free(ring->adev, ring->rptr_offs);
293 	amdgpu_device_wb_free(ring->adev, ring->wptr_offs);
294 
295 	amdgpu_device_wb_free(ring->adev, ring->cond_exe_offs);
296 	amdgpu_device_wb_free(ring->adev, ring->fence_offs);
297 
298 	amdgpu_bo_free_kernel(&ring->ring_obj,
299 			      &ring->gpu_addr,
300 			      (void **)&ring->ring);
301 
302 	dma_fence_put(ring->vmid_wait);
303 	ring->vmid_wait = NULL;
304 	ring->me = 0;
305 
306 	ring->adev->rings[ring->idx] = NULL;
307 }
308 
309 /**
310  * amdgpu_ring_emit_reg_write_reg_wait_helper - ring helper
311  *
312  * @ring: ring to write to
313  * @reg0: register to write
314  * @reg1: register to wait on
315  * @ref: reference value to write/wait on
316  * @mask: mask to wait on
317  *
318  * Helper for rings that don't support write and wait in a
319  * single oneshot packet.
320  */
321 void amdgpu_ring_emit_reg_write_reg_wait_helper(struct amdgpu_ring *ring,
322 						uint32_t reg0, uint32_t reg1,
323 						uint32_t ref, uint32_t mask)
324 {
325 	amdgpu_ring_emit_wreg(ring, reg0, ref);
326 	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
327 }
328 
329 /**
330  * amdgpu_ring_soft_recovery - try to soft recover a ring lockup
331  *
332  * @ring: ring to try the recovery on
333  * @vmid: VMID we try to get going again
334  * @fence: timedout fence
335  *
336  * Tries to get a ring proceeding again when it is stuck.
337  */
338 bool amdgpu_ring_soft_recovery(struct amdgpu_ring *ring, unsigned int vmid,
339 			       struct dma_fence *fence)
340 {
341 	ktime_t deadline = ktime_add_us(ktime_get(), 10000);
342 
343 	if (amdgpu_sriov_vf(ring->adev) || !ring->funcs->soft_recovery || !fence)
344 		return false;
345 
346 	atomic_inc(&ring->adev->gpu_reset_counter);
347 	while (!dma_fence_is_signaled(fence) &&
348 	       ktime_to_ns(ktime_sub(deadline, ktime_get())) > 0)
349 		ring->funcs->soft_recovery(ring, vmid);
350 
351 	return dma_fence_is_signaled(fence);
352 }
353 
354 /*
355  * Debugfs info
356  */
357 #if defined(CONFIG_DEBUG_FS)
358 
359 /* Layout of file is 12 bytes consisting of
360  * - rptr
361  * - wptr
362  * - driver's copy of wptr
363  *
364  * followed by n-words of ring data
365  */
366 static ssize_t amdgpu_debugfs_ring_read(struct file *f, char __user *buf,
367 					size_t size, loff_t *pos)
368 {
369 	struct amdgpu_ring *ring = file_inode(f)->i_private;
370 	int r, i;
371 	uint32_t value, result, early[3];
372 
373 	if (*pos & 3 || size & 3)
374 		return -EINVAL;
375 
376 	result = 0;
377 
378 	if (*pos < 12) {
379 		early[0] = amdgpu_ring_get_rptr(ring) & ring->buf_mask;
380 		early[1] = amdgpu_ring_get_wptr(ring) & ring->buf_mask;
381 		early[2] = ring->wptr & ring->buf_mask;
382 		for (i = *pos / 4; i < 3 && size; i++) {
383 			r = put_user(early[i], (uint32_t *)buf);
384 			if (r)
385 				return r;
386 			buf += 4;
387 			result += 4;
388 			size -= 4;
389 			*pos += 4;
390 		}
391 	}
392 
393 	while (size) {
394 		if (*pos >= (ring->ring_size + 12))
395 			return result;
396 
397 		value = ring->ring[(*pos - 12)/4];
398 		r = put_user(value, (uint32_t *)buf);
399 		if (r)
400 			return r;
401 		buf += 4;
402 		result += 4;
403 		size -= 4;
404 		*pos += 4;
405 	}
406 
407 	return result;
408 }
409 
410 static const struct file_operations amdgpu_debugfs_ring_fops = {
411 	.owner = THIS_MODULE,
412 	.read = amdgpu_debugfs_ring_read,
413 	.llseek = default_llseek
414 };
415 
416 #endif
417 
418 void amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
419 			      struct amdgpu_ring *ring)
420 {
421 #if defined(CONFIG_DEBUG_FS)
422 	struct drm_minor *minor = adev_to_drm(adev)->primary;
423 	struct dentry *root = minor->debugfs_root;
424 	char name[32];
425 
426 	sprintf(name, "amdgpu_ring_%s", ring->name);
427 	debugfs_create_file_size(name, S_IFREG | S_IRUGO, root, ring,
428 				 &amdgpu_debugfs_ring_fops,
429 				 ring->ring_size + 12);
430 
431 #endif
432 }
433 
434 /**
435  * amdgpu_ring_test_helper - tests ring and set sched readiness status
436  *
437  * @ring: ring to try the recovery on
438  *
439  * Tests ring and set sched readiness status
440  *
441  * Returns 0 on success, error on failure.
442  */
443 int amdgpu_ring_test_helper(struct amdgpu_ring *ring)
444 {
445 	struct amdgpu_device *adev = ring->adev;
446 	int r;
447 
448 	r = amdgpu_ring_test_ring(ring);
449 	if (r)
450 		DRM_DEV_ERROR(adev->dev, "ring %s test failed (%d)\n",
451 			      ring->name, r);
452 	else
453 		DRM_DEV_DEBUG(adev->dev, "ring test on %s succeeded\n",
454 			      ring->name);
455 
456 	ring->sched.ready = !r;
457 	return r;
458 }
459