Loading...
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 <drm/drmP.h>
32#include <drm/amdgpu_drm.h>
33#include "amdgpu.h"
34#include "atom.h"
35
36/*
37 * Rings
38 * Most engines on the GPU are fed via ring buffers. Ring
39 * buffers are areas of GPU accessible memory that the host
40 * writes commands into and the GPU reads commands out of.
41 * There is a rptr (read pointer) that determines where the
42 * GPU is currently reading, and a wptr (write pointer)
43 * which determines where the host has written. When the
44 * pointers are equal, the ring is idle. When the host
45 * writes commands to the ring buffer, it increments the
46 * wptr. The GPU then starts fetching commands and executes
47 * them until the pointers are equal again.
48 */
49static int amdgpu_debugfs_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring);
50
51/**
52 * amdgpu_ring_alloc - allocate space on the ring buffer
53 *
54 * @adev: amdgpu_device pointer
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 */
61int 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->align_mask) & ~ring->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 return 0;
76}
77
78/** amdgpu_ring_insert_nop - insert NOP packets
79 *
80 * @ring: amdgpu_ring structure holding ring information
81 * @count: the number of NOP packets to insert
82 *
83 * This is the generic insert_nop function for rings except SDMA
84 */
85void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
86{
87 int i;
88
89 for (i = 0; i < count; i++)
90 amdgpu_ring_write(ring, ring->nop);
91}
92
93/** amdgpu_ring_generic_pad_ib - pad IB with NOP packets
94 *
95 * @ring: amdgpu_ring structure holding ring information
96 * @ib: IB to add NOP packets to
97 *
98 * This is the generic pad_ib function for rings except SDMA
99 */
100void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
101{
102 while (ib->length_dw & ring->align_mask)
103 ib->ptr[ib->length_dw++] = ring->nop;
104}
105
106/**
107 * amdgpu_ring_commit - tell the GPU to execute the new
108 * commands on the ring buffer
109 *
110 * @adev: amdgpu_device pointer
111 * @ring: amdgpu_ring structure holding ring information
112 *
113 * Update the wptr (write pointer) to tell the GPU to
114 * execute new commands on the ring buffer (all asics).
115 */
116void amdgpu_ring_commit(struct amdgpu_ring *ring)
117{
118 uint32_t count;
119
120 /* We pad to match fetch size */
121 count = ring->align_mask + 1 - (ring->wptr & ring->align_mask);
122 count %= ring->align_mask + 1;
123 ring->funcs->insert_nop(ring, count);
124
125 mb();
126 amdgpu_ring_set_wptr(ring);
127}
128
129/**
130 * amdgpu_ring_undo - reset the wptr
131 *
132 * @ring: amdgpu_ring structure holding ring information
133 *
134 * Reset the driver's copy of the wptr (all asics).
135 */
136void amdgpu_ring_undo(struct amdgpu_ring *ring)
137{
138 ring->wptr = ring->wptr_old;
139}
140
141/**
142 * amdgpu_ring_backup - Back up the content of a ring
143 *
144 * @ring: the ring we want to back up
145 *
146 * Saves all unprocessed commits from a ring, returns the number of dwords saved.
147 */
148unsigned amdgpu_ring_backup(struct amdgpu_ring *ring,
149 uint32_t **data)
150{
151 unsigned size, ptr, i;
152
153 *data = NULL;
154
155 if (ring->ring_obj == NULL)
156 return 0;
157
158 /* it doesn't make sense to save anything if all fences are signaled */
159 if (!amdgpu_fence_count_emitted(ring))
160 return 0;
161
162 ptr = le32_to_cpu(*ring->next_rptr_cpu_addr);
163
164 size = ring->wptr + (ring->ring_size / 4);
165 size -= ptr;
166 size &= ring->ptr_mask;
167 if (size == 0)
168 return 0;
169
170 /* and then save the content of the ring */
171 *data = kmalloc_array(size, sizeof(uint32_t), GFP_KERNEL);
172 if (!*data)
173 return 0;
174 for (i = 0; i < size; ++i) {
175 (*data)[i] = ring->ring[ptr++];
176 ptr &= ring->ptr_mask;
177 }
178
179 return size;
180}
181
182/**
183 * amdgpu_ring_restore - append saved commands to the ring again
184 *
185 * @ring: ring to append commands to
186 * @size: number of dwords we want to write
187 * @data: saved commands
188 *
189 * Allocates space on the ring and restore the previously saved commands.
190 */
191int amdgpu_ring_restore(struct amdgpu_ring *ring,
192 unsigned size, uint32_t *data)
193{
194 int i, r;
195
196 if (!size || !data)
197 return 0;
198
199 /* restore the saved ring content */
200 r = amdgpu_ring_alloc(ring, size);
201 if (r)
202 return r;
203
204 for (i = 0; i < size; ++i) {
205 amdgpu_ring_write(ring, data[i]);
206 }
207
208 amdgpu_ring_commit(ring);
209 kfree(data);
210 return 0;
211}
212
213/**
214 * amdgpu_ring_init - init driver ring struct.
215 *
216 * @adev: amdgpu_device pointer
217 * @ring: amdgpu_ring structure holding ring information
218 * @ring_size: size of the ring
219 * @nop: nop packet for this ring
220 *
221 * Initialize the driver information for the selected ring (all asics).
222 * Returns 0 on success, error on failure.
223 */
224int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
225 unsigned ring_size, u32 nop, u32 align_mask,
226 struct amdgpu_irq_src *irq_src, unsigned irq_type,
227 enum amdgpu_ring_type ring_type)
228{
229 u32 rb_bufsz;
230 int r;
231
232 if (ring->adev == NULL) {
233 if (adev->num_rings >= AMDGPU_MAX_RINGS)
234 return -EINVAL;
235
236 ring->adev = adev;
237 ring->idx = adev->num_rings++;
238 adev->rings[ring->idx] = ring;
239 r = amdgpu_fence_driver_init_ring(ring,
240 amdgpu_sched_hw_submission);
241 if (r)
242 return r;
243 }
244
245 r = amdgpu_wb_get(adev, &ring->rptr_offs);
246 if (r) {
247 dev_err(adev->dev, "(%d) ring rptr_offs wb alloc failed\n", r);
248 return r;
249 }
250
251 r = amdgpu_wb_get(adev, &ring->wptr_offs);
252 if (r) {
253 dev_err(adev->dev, "(%d) ring wptr_offs wb alloc failed\n", r);
254 return r;
255 }
256
257 r = amdgpu_wb_get(adev, &ring->fence_offs);
258 if (r) {
259 dev_err(adev->dev, "(%d) ring fence_offs wb alloc failed\n", r);
260 return r;
261 }
262
263 r = amdgpu_wb_get(adev, &ring->next_rptr_offs);
264 if (r) {
265 dev_err(adev->dev, "(%d) ring next_rptr wb alloc failed\n", r);
266 return r;
267 }
268 ring->next_rptr_gpu_addr = adev->wb.gpu_addr + (ring->next_rptr_offs * 4);
269 ring->next_rptr_cpu_addr = &adev->wb.wb[ring->next_rptr_offs];
270 spin_lock_init(&ring->fence_lock);
271 r = amdgpu_fence_driver_start_ring(ring, irq_src, irq_type);
272 if (r) {
273 dev_err(adev->dev, "failed initializing fences (%d).\n", r);
274 return r;
275 }
276
277 /* Align ring size */
278 rb_bufsz = order_base_2(ring_size / 8);
279 ring_size = (1 << (rb_bufsz + 1)) * 4;
280 ring->ring_size = ring_size;
281 ring->align_mask = align_mask;
282 ring->nop = nop;
283 ring->type = ring_type;
284
285 /* Allocate ring buffer */
286 if (ring->ring_obj == NULL) {
287 r = amdgpu_bo_create(adev, ring->ring_size, PAGE_SIZE, true,
288 AMDGPU_GEM_DOMAIN_GTT, 0,
289 NULL, NULL, &ring->ring_obj);
290 if (r) {
291 dev_err(adev->dev, "(%d) ring create failed\n", r);
292 return r;
293 }
294 r = amdgpu_bo_reserve(ring->ring_obj, false);
295 if (unlikely(r != 0))
296 return r;
297 r = amdgpu_bo_pin(ring->ring_obj, AMDGPU_GEM_DOMAIN_GTT,
298 &ring->gpu_addr);
299 if (r) {
300 amdgpu_bo_unreserve(ring->ring_obj);
301 dev_err(adev->dev, "(%d) ring pin failed\n", r);
302 return r;
303 }
304 r = amdgpu_bo_kmap(ring->ring_obj,
305 (void **)&ring->ring);
306 amdgpu_bo_unreserve(ring->ring_obj);
307 if (r) {
308 dev_err(adev->dev, "(%d) ring map failed\n", r);
309 return r;
310 }
311 }
312 ring->ptr_mask = (ring->ring_size / 4) - 1;
313 ring->max_dw = DIV_ROUND_UP(ring->ring_size / 4,
314 amdgpu_sched_hw_submission);
315
316 if (amdgpu_debugfs_ring_init(adev, ring)) {
317 DRM_ERROR("Failed to register debugfs file for rings !\n");
318 }
319 return 0;
320}
321
322/**
323 * amdgpu_ring_fini - tear down the driver ring struct.
324 *
325 * @adev: amdgpu_device pointer
326 * @ring: amdgpu_ring structure holding ring information
327 *
328 * Tear down the driver information for the selected ring (all asics).
329 */
330void amdgpu_ring_fini(struct amdgpu_ring *ring)
331{
332 int r;
333 struct amdgpu_bo *ring_obj;
334
335 ring_obj = ring->ring_obj;
336 ring->ready = false;
337 ring->ring = NULL;
338 ring->ring_obj = NULL;
339
340 amdgpu_wb_free(ring->adev, ring->fence_offs);
341 amdgpu_wb_free(ring->adev, ring->rptr_offs);
342 amdgpu_wb_free(ring->adev, ring->wptr_offs);
343 amdgpu_wb_free(ring->adev, ring->next_rptr_offs);
344
345 if (ring_obj) {
346 r = amdgpu_bo_reserve(ring_obj, false);
347 if (likely(r == 0)) {
348 amdgpu_bo_kunmap(ring_obj);
349 amdgpu_bo_unpin(ring_obj);
350 amdgpu_bo_unreserve(ring_obj);
351 }
352 amdgpu_bo_unref(&ring_obj);
353 }
354}
355
356/*
357 * Debugfs info
358 */
359#if defined(CONFIG_DEBUG_FS)
360
361static int amdgpu_debugfs_ring_info(struct seq_file *m, void *data)
362{
363 struct drm_info_node *node = (struct drm_info_node *) m->private;
364 struct drm_device *dev = node->minor->dev;
365 struct amdgpu_device *adev = dev->dev_private;
366 int roffset = *(int*)node->info_ent->data;
367 struct amdgpu_ring *ring = (void *)(((uint8_t*)adev) + roffset);
368
369 uint32_t rptr, wptr, rptr_next;
370 unsigned i;
371
372 wptr = amdgpu_ring_get_wptr(ring);
373 seq_printf(m, "wptr: 0x%08x [%5d]\n", wptr, wptr);
374
375 rptr = amdgpu_ring_get_rptr(ring);
376 rptr_next = le32_to_cpu(*ring->next_rptr_cpu_addr);
377
378 seq_printf(m, "rptr: 0x%08x [%5d]\n", rptr, rptr);
379
380 seq_printf(m, "driver's copy of the wptr: 0x%08x [%5d]\n",
381 ring->wptr, ring->wptr);
382
383 if (!ring->ready)
384 return 0;
385
386 /* print 8 dw before current rptr as often it's the last executed
387 * packet that is the root issue
388 */
389 i = (rptr + ring->ptr_mask + 1 - 32) & ring->ptr_mask;
390 while (i != rptr) {
391 seq_printf(m, "r[%5d]=0x%08x", i, ring->ring[i]);
392 if (i == rptr)
393 seq_puts(m, " *");
394 if (i == rptr_next)
395 seq_puts(m, " #");
396 seq_puts(m, "\n");
397 i = (i + 1) & ring->ptr_mask;
398 }
399 while (i != wptr) {
400 seq_printf(m, "r[%5d]=0x%08x", i, ring->ring[i]);
401 if (i == rptr)
402 seq_puts(m, " *");
403 if (i == rptr_next)
404 seq_puts(m, " #");
405 seq_puts(m, "\n");
406 i = (i + 1) & ring->ptr_mask;
407 }
408 return 0;
409}
410
411/* TODO: clean this up !*/
412static int amdgpu_gfx_index = offsetof(struct amdgpu_device, gfx.gfx_ring[0]);
413static int cayman_cp1_index = offsetof(struct amdgpu_device, gfx.compute_ring[0]);
414static int cayman_cp2_index = offsetof(struct amdgpu_device, gfx.compute_ring[1]);
415static int amdgpu_dma1_index = offsetof(struct amdgpu_device, sdma.instance[0].ring);
416static int amdgpu_dma2_index = offsetof(struct amdgpu_device, sdma.instance[1].ring);
417static int r600_uvd_index = offsetof(struct amdgpu_device, uvd.ring);
418static int si_vce1_index = offsetof(struct amdgpu_device, vce.ring[0]);
419static int si_vce2_index = offsetof(struct amdgpu_device, vce.ring[1]);
420
421static struct drm_info_list amdgpu_debugfs_ring_info_list[] = {
422 {"amdgpu_ring_gfx", amdgpu_debugfs_ring_info, 0, &amdgpu_gfx_index},
423 {"amdgpu_ring_cp1", amdgpu_debugfs_ring_info, 0, &cayman_cp1_index},
424 {"amdgpu_ring_cp2", amdgpu_debugfs_ring_info, 0, &cayman_cp2_index},
425 {"amdgpu_ring_dma1", amdgpu_debugfs_ring_info, 0, &amdgpu_dma1_index},
426 {"amdgpu_ring_dma2", amdgpu_debugfs_ring_info, 0, &amdgpu_dma2_index},
427 {"amdgpu_ring_uvd", amdgpu_debugfs_ring_info, 0, &r600_uvd_index},
428 {"amdgpu_ring_vce1", amdgpu_debugfs_ring_info, 0, &si_vce1_index},
429 {"amdgpu_ring_vce2", amdgpu_debugfs_ring_info, 0, &si_vce2_index},
430};
431
432#endif
433
434static int amdgpu_debugfs_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring)
435{
436#if defined(CONFIG_DEBUG_FS)
437 unsigned i;
438 for (i = 0; i < ARRAY_SIZE(amdgpu_debugfs_ring_info_list); ++i) {
439 struct drm_info_list *info = &amdgpu_debugfs_ring_info_list[i];
440 int roffset = *(int*)amdgpu_debugfs_ring_info_list[i].data;
441 struct amdgpu_ring *other = (void *)(((uint8_t*)adev) + roffset);
442 unsigned r;
443
444 if (other != ring)
445 continue;
446
447 r = amdgpu_debugfs_add_files(adev, info, 1);
448 if (r)
449 return r;
450 }
451#endif
452 return 0;
453}
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 */
50static int amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
51 struct amdgpu_ring *ring);
52static 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 */
64int 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 */
92void 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 */
107void 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 */
123void 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 */
150void 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 */
166void 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
196out_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 */
208void 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
223out_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 */
238int 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 */
343void 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 dma_fence_put(ring->vmid_wait);
364 ring->vmid_wait = NULL;
365
366 ring->adev->rings[ring->idx] = NULL;
367}
368
369static void amdgpu_ring_lru_touch_locked(struct amdgpu_device *adev,
370 struct amdgpu_ring *ring)
371{
372 /* list_move_tail handles the case where ring isn't part of the list */
373 list_move_tail(&ring->lru_list, &adev->ring_lru_list);
374}
375
376static bool amdgpu_ring_is_blacklisted(struct amdgpu_ring *ring,
377 int *blacklist, int num_blacklist)
378{
379 int i;
380
381 for (i = 0; i < num_blacklist; i++) {
382 if (ring->idx == blacklist[i])
383 return true;
384 }
385
386 return false;
387}
388
389/**
390 * amdgpu_ring_lru_get - get the least recently used ring for a HW IP block
391 *
392 * @adev: amdgpu_device pointer
393 * @type: amdgpu_ring_type enum
394 * @blacklist: blacklisted ring ids array
395 * @num_blacklist: number of entries in @blacklist
396 * @lru_pipe_order: find a ring from the least recently used pipe
397 * @ring: output ring
398 *
399 * Retrieve the amdgpu_ring structure for the least recently used ring of
400 * a specific IP block (all asics).
401 * Returns 0 on success, error on failure.
402 */
403int amdgpu_ring_lru_get(struct amdgpu_device *adev, int type,
404 int *blacklist, int num_blacklist,
405 bool lru_pipe_order, struct amdgpu_ring **ring)
406{
407 struct amdgpu_ring *entry;
408
409 /* List is sorted in LRU order, find first entry corresponding
410 * to the desired HW IP */
411 *ring = NULL;
412 spin_lock(&adev->ring_lru_list_lock);
413 list_for_each_entry(entry, &adev->ring_lru_list, lru_list) {
414 if (entry->funcs->type != type)
415 continue;
416
417 if (amdgpu_ring_is_blacklisted(entry, blacklist, num_blacklist))
418 continue;
419
420 if (!*ring) {
421 *ring = entry;
422
423 /* We are done for ring LRU */
424 if (!lru_pipe_order)
425 break;
426 }
427
428 /* Move all rings on the same pipe to the end of the list */
429 if (entry->pipe == (*ring)->pipe)
430 amdgpu_ring_lru_touch_locked(adev, entry);
431 }
432
433 /* Move the ring we found to the end of the list */
434 if (*ring)
435 amdgpu_ring_lru_touch_locked(adev, *ring);
436
437 spin_unlock(&adev->ring_lru_list_lock);
438
439 if (!*ring) {
440 DRM_ERROR("Ring LRU contains no entries for ring type:%d\n", type);
441 return -EINVAL;
442 }
443
444 return 0;
445}
446
447/**
448 * amdgpu_ring_lru_touch - mark a ring as recently being used
449 *
450 * @adev: amdgpu_device pointer
451 * @ring: ring to touch
452 *
453 * Move @ring to the tail of the lru list
454 */
455void amdgpu_ring_lru_touch(struct amdgpu_device *adev, struct amdgpu_ring *ring)
456{
457 spin_lock(&adev->ring_lru_list_lock);
458 amdgpu_ring_lru_touch_locked(adev, ring);
459 spin_unlock(&adev->ring_lru_list_lock);
460}
461
462/*
463 * Debugfs info
464 */
465#if defined(CONFIG_DEBUG_FS)
466
467/* Layout of file is 12 bytes consisting of
468 * - rptr
469 * - wptr
470 * - driver's copy of wptr
471 *
472 * followed by n-words of ring data
473 */
474static ssize_t amdgpu_debugfs_ring_read(struct file *f, char __user *buf,
475 size_t size, loff_t *pos)
476{
477 struct amdgpu_ring *ring = file_inode(f)->i_private;
478 int r, i;
479 uint32_t value, result, early[3];
480
481 if (*pos & 3 || size & 3)
482 return -EINVAL;
483
484 result = 0;
485
486 if (*pos < 12) {
487 early[0] = amdgpu_ring_get_rptr(ring) & ring->buf_mask;
488 early[1] = amdgpu_ring_get_wptr(ring) & ring->buf_mask;
489 early[2] = ring->wptr & ring->buf_mask;
490 for (i = *pos / 4; i < 3 && size; i++) {
491 r = put_user(early[i], (uint32_t *)buf);
492 if (r)
493 return r;
494 buf += 4;
495 result += 4;
496 size -= 4;
497 *pos += 4;
498 }
499 }
500
501 while (size) {
502 if (*pos >= (ring->ring_size + 12))
503 return result;
504
505 value = ring->ring[(*pos - 12)/4];
506 r = put_user(value, (uint32_t*)buf);
507 if (r)
508 return r;
509 buf += 4;
510 result += 4;
511 size -= 4;
512 *pos += 4;
513 }
514
515 return result;
516}
517
518static const struct file_operations amdgpu_debugfs_ring_fops = {
519 .owner = THIS_MODULE,
520 .read = amdgpu_debugfs_ring_read,
521 .llseek = default_llseek
522};
523
524#endif
525
526static int amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
527 struct amdgpu_ring *ring)
528{
529#if defined(CONFIG_DEBUG_FS)
530 struct drm_minor *minor = adev->ddev->primary;
531 struct dentry *ent, *root = minor->debugfs_root;
532 char name[32];
533
534 sprintf(name, "amdgpu_ring_%s", ring->name);
535
536 ent = debugfs_create_file(name,
537 S_IFREG | S_IRUGO, root,
538 ring, &amdgpu_debugfs_ring_fops);
539 if (!ent)
540 return -ENOMEM;
541
542 i_size_write(ent->d_inode, ring->ring_size + 12);
543 ring->ent = ent;
544#endif
545 return 0;
546}
547
548static void amdgpu_debugfs_ring_fini(struct amdgpu_ring *ring)
549{
550#if defined(CONFIG_DEBUG_FS)
551 debugfs_remove(ring->ent);
552#endif
553}