]> Git Repo - linux.git/blob - drivers/gpu/drm/amd/amdgpu/amdgpu_ring.c
Merge branch 'drm-next-4.15' of git://people.freedesktop.org/~agd5f/linux into drm...
[linux.git] / drivers / gpu / drm / amd / amdgpu / amdgpu_ring.c
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         amdgpu_ring_lru_touch(ring->adev, ring);
140 }
141
142 /**
143  * amdgpu_ring_undo - reset the wptr
144  *
145  * @ring: amdgpu_ring structure holding ring information
146  *
147  * Reset the driver's copy of the wptr (all asics).
148  */
149 void amdgpu_ring_undo(struct amdgpu_ring *ring)
150 {
151         ring->wptr = ring->wptr_old;
152
153         if (ring->funcs->end_use)
154                 ring->funcs->end_use(ring);
155 }
156
157 /**
158  * amdgpu_ring_priority_put - restore a ring's priority
159  *
160  * @ring: amdgpu_ring structure holding the information
161  * @priority: target priority
162  *
163  * Release a request for executing at @priority
164  */
165 void amdgpu_ring_priority_put(struct amdgpu_ring *ring,
166                               enum amd_sched_priority priority)
167 {
168         int i;
169
170         if (!ring->funcs->set_priority)
171                 return;
172
173         if (atomic_dec_return(&ring->num_jobs[priority]) > 0)
174                 return;
175
176         /* no need to restore if the job is already at the lowest priority */
177         if (priority == AMD_SCHED_PRIORITY_NORMAL)
178                 return;
179
180         mutex_lock(&ring->priority_mutex);
181         /* something higher prio is executing, no need to decay */
182         if (ring->priority > priority)
183                 goto out_unlock;
184
185         /* decay priority to the next level with a job available */
186         for (i = priority; i >= AMD_SCHED_PRIORITY_MIN; i--) {
187                 if (i == AMD_SCHED_PRIORITY_NORMAL
188                                 || atomic_read(&ring->num_jobs[i])) {
189                         ring->priority = i;
190                         ring->funcs->set_priority(ring, i);
191                         break;
192                 }
193         }
194
195 out_unlock:
196         mutex_unlock(&ring->priority_mutex);
197 }
198
199 /**
200  * amdgpu_ring_priority_get - change the ring's priority
201  *
202  * @ring: amdgpu_ring structure holding the information
203  * @priority: target priority
204  *
205  * Request a ring's priority to be raised to @priority (refcounted).
206  */
207 void amdgpu_ring_priority_get(struct amdgpu_ring *ring,
208                               enum amd_sched_priority priority)
209 {
210         if (!ring->funcs->set_priority)
211                 return;
212
213         atomic_inc(&ring->num_jobs[priority]);
214
215         mutex_lock(&ring->priority_mutex);
216         if (priority <= ring->priority)
217                 goto out_unlock;
218
219         ring->priority = priority;
220         ring->funcs->set_priority(ring, priority);
221
222 out_unlock:
223         mutex_unlock(&ring->priority_mutex);
224 }
225
226 /**
227  * amdgpu_ring_init - init driver ring struct.
228  *
229  * @adev: amdgpu_device pointer
230  * @ring: amdgpu_ring structure holding ring information
231  * @max_ndw: maximum number of dw for ring alloc
232  * @nop: nop packet for this ring
233  *
234  * Initialize the driver information for the selected ring (all asics).
235  * Returns 0 on success, error on failure.
236  */
237 int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
238                      unsigned max_dw, struct amdgpu_irq_src *irq_src,
239                      unsigned irq_type)
240 {
241         int r, i;
242         int sched_hw_submission = amdgpu_sched_hw_submission;
243
244         /* Set the hw submission limit higher for KIQ because
245          * it's used for a number of gfx/compute tasks by both
246          * KFD and KGD which may have outstanding fences and
247          * it doesn't really use the gpu scheduler anyway;
248          * KIQ tasks get submitted directly to the ring.
249          */
250         if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
251                 sched_hw_submission = max(sched_hw_submission, 256);
252
253         if (ring->adev == NULL) {
254                 if (adev->num_rings >= AMDGPU_MAX_RINGS)
255                         return -EINVAL;
256
257                 ring->adev = adev;
258                 ring->idx = adev->num_rings++;
259                 adev->rings[ring->idx] = ring;
260                 r = amdgpu_fence_driver_init_ring(ring, sched_hw_submission);
261                 if (r)
262                         return r;
263         }
264
265         r = amdgpu_wb_get(adev, &ring->rptr_offs);
266         if (r) {
267                 dev_err(adev->dev, "(%d) ring rptr_offs wb alloc failed\n", r);
268                 return r;
269         }
270
271         r = amdgpu_wb_get(adev, &ring->wptr_offs);
272         if (r) {
273                 dev_err(adev->dev, "(%d) ring wptr_offs wb alloc failed\n", r);
274                 return r;
275         }
276
277         r = amdgpu_wb_get(adev, &ring->fence_offs);
278         if (r) {
279                 dev_err(adev->dev, "(%d) ring fence_offs wb alloc failed\n", r);
280                 return r;
281         }
282
283         r = amdgpu_wb_get(adev, &ring->cond_exe_offs);
284         if (r) {
285                 dev_err(adev->dev, "(%d) ring cond_exec_polling wb alloc failed\n", r);
286                 return r;
287         }
288         ring->cond_exe_gpu_addr = adev->wb.gpu_addr + (ring->cond_exe_offs * 4);
289         ring->cond_exe_cpu_addr = &adev->wb.wb[ring->cond_exe_offs];
290         /* always set cond_exec_polling to CONTINUE */
291         *ring->cond_exe_cpu_addr = 1;
292
293         r = amdgpu_fence_driver_start_ring(ring, irq_src, irq_type);
294         if (r) {
295                 dev_err(adev->dev, "failed initializing fences (%d).\n", r);
296                 return r;
297         }
298
299         ring->ring_size = roundup_pow_of_two(max_dw * 4 * sched_hw_submission);
300
301         ring->buf_mask = (ring->ring_size / 4) - 1;
302         ring->ptr_mask = ring->funcs->support_64bit_ptrs ?
303                 0xffffffffffffffff : ring->buf_mask;
304         /* Allocate ring buffer */
305         if (ring->ring_obj == NULL) {
306                 r = amdgpu_bo_create_kernel(adev, ring->ring_size, PAGE_SIZE,
307                                             AMDGPU_GEM_DOMAIN_GTT,
308                                             &ring->ring_obj,
309                                             &ring->gpu_addr,
310                                             (void **)&ring->ring);
311                 if (r) {
312                         dev_err(adev->dev, "(%d) ring create failed\n", r);
313                         return r;
314                 }
315                 amdgpu_ring_clear_ring(ring);
316         }
317
318         ring->max_dw = max_dw;
319         ring->priority = AMD_SCHED_PRIORITY_NORMAL;
320         mutex_init(&ring->priority_mutex);
321         INIT_LIST_HEAD(&ring->lru_list);
322         amdgpu_ring_lru_touch(adev, ring);
323
324         for (i = 0; i < AMD_SCHED_PRIORITY_MAX; ++i)
325                 atomic_set(&ring->num_jobs[i], 0);
326
327         if (amdgpu_debugfs_ring_init(adev, ring)) {
328                 DRM_ERROR("Failed to register debugfs file for rings !\n");
329         }
330
331         return 0;
332 }
333
334 /**
335  * amdgpu_ring_fini - tear down the driver ring struct.
336  *
337  * @adev: amdgpu_device pointer
338  * @ring: amdgpu_ring structure holding ring information
339  *
340  * Tear down the driver information for the selected ring (all asics).
341  */
342 void amdgpu_ring_fini(struct amdgpu_ring *ring)
343 {
344         ring->ready = false;
345
346         /* Not to finish a ring which is not initialized */
347         if (!(ring->adev) || !(ring->adev->rings[ring->idx]))
348                 return;
349
350         amdgpu_wb_free(ring->adev, ring->rptr_offs);
351         amdgpu_wb_free(ring->adev, ring->wptr_offs);
352
353         amdgpu_wb_free(ring->adev, ring->cond_exe_offs);
354         amdgpu_wb_free(ring->adev, ring->fence_offs);
355
356         amdgpu_bo_free_kernel(&ring->ring_obj,
357                               &ring->gpu_addr,
358                               (void **)&ring->ring);
359
360         amdgpu_debugfs_ring_fini(ring);
361
362         ring->adev->rings[ring->idx] = NULL;
363 }
364
365 static void amdgpu_ring_lru_touch_locked(struct amdgpu_device *adev,
366                                          struct amdgpu_ring *ring)
367 {
368         /* list_move_tail handles the case where ring isn't part of the list */
369         list_move_tail(&ring->lru_list, &adev->ring_lru_list);
370 }
371
372 static bool amdgpu_ring_is_blacklisted(struct amdgpu_ring *ring,
373                                        int *blacklist, int num_blacklist)
374 {
375         int i;
376
377         for (i = 0; i < num_blacklist; i++) {
378                 if (ring->idx == blacklist[i])
379                         return true;
380         }
381
382         return false;
383 }
384
385 /**
386  * amdgpu_ring_lru_get - get the least recently used ring for a HW IP block
387  *
388  * @adev: amdgpu_device pointer
389  * @type: amdgpu_ring_type enum
390  * @blacklist: blacklisted ring ids array
391  * @num_blacklist: number of entries in @blacklist
392  * @lru_pipe_order: find a ring from the least recently used pipe
393  * @ring: output ring
394  *
395  * Retrieve the amdgpu_ring structure for the least recently used ring of
396  * a specific IP block (all asics).
397  * Returns 0 on success, error on failure.
398  */
399 int amdgpu_ring_lru_get(struct amdgpu_device *adev, int type,
400                         int *blacklist, int num_blacklist,
401                         bool lru_pipe_order, struct amdgpu_ring **ring)
402 {
403         struct amdgpu_ring *entry;
404
405         /* List is sorted in LRU order, find first entry corresponding
406          * to the desired HW IP */
407         *ring = NULL;
408         spin_lock(&adev->ring_lru_list_lock);
409         list_for_each_entry(entry, &adev->ring_lru_list, lru_list) {
410                 if (entry->funcs->type != type)
411                         continue;
412
413                 if (amdgpu_ring_is_blacklisted(entry, blacklist, num_blacklist))
414                         continue;
415
416                 if (!*ring) {
417                         *ring = entry;
418
419                         /* We are done for ring LRU */
420                         if (!lru_pipe_order)
421                                 break;
422                 }
423
424                 /* Move all rings on the same pipe to the end of the list */
425                 if (entry->pipe == (*ring)->pipe)
426                         amdgpu_ring_lru_touch_locked(adev, entry);
427         }
428
429         /* Move the ring we found to the end of the list */
430         if (*ring)
431                 amdgpu_ring_lru_touch_locked(adev, *ring);
432
433         spin_unlock(&adev->ring_lru_list_lock);
434
435         if (!*ring) {
436                 DRM_ERROR("Ring LRU contains no entries for ring type:%d\n", type);
437                 return -EINVAL;
438         }
439
440         return 0;
441 }
442
443 /**
444  * amdgpu_ring_lru_touch - mark a ring as recently being used
445  *
446  * @adev: amdgpu_device pointer
447  * @ring: ring to touch
448  *
449  * Move @ring to the tail of the lru list
450  */
451 void amdgpu_ring_lru_touch(struct amdgpu_device *adev, struct amdgpu_ring *ring)
452 {
453         spin_lock(&adev->ring_lru_list_lock);
454         amdgpu_ring_lru_touch_locked(adev, ring);
455         spin_unlock(&adev->ring_lru_list_lock);
456 }
457
458 /*
459  * Debugfs info
460  */
461 #if defined(CONFIG_DEBUG_FS)
462
463 /* Layout of file is 12 bytes consisting of
464  * - rptr
465  * - wptr
466  * - driver's copy of wptr
467  *
468  * followed by n-words of ring data
469  */
470 static ssize_t amdgpu_debugfs_ring_read(struct file *f, char __user *buf,
471                                         size_t size, loff_t *pos)
472 {
473         struct amdgpu_ring *ring = file_inode(f)->i_private;
474         int r, i;
475         uint32_t value, result, early[3];
476
477         if (*pos & 3 || size & 3)
478                 return -EINVAL;
479
480         result = 0;
481
482         if (*pos < 12) {
483                 early[0] = amdgpu_ring_get_rptr(ring);
484                 early[1] = amdgpu_ring_get_wptr(ring) & ring->buf_mask;
485                 early[2] = ring->wptr & ring->buf_mask;
486                 for (i = *pos / 4; i < 3 && size; i++) {
487                         r = put_user(early[i], (uint32_t *)buf);
488                         if (r)
489                                 return r;
490                         buf += 4;
491                         result += 4;
492                         size -= 4;
493                         *pos += 4;
494                 }
495         }
496
497         while (size) {
498                 if (*pos >= (ring->ring_size + 12))
499                         return result;
500
501                 value = ring->ring[(*pos - 12)/4];
502                 r = put_user(value, (uint32_t*)buf);
503                 if (r)
504                         return r;
505                 buf += 4;
506                 result += 4;
507                 size -= 4;
508                 *pos += 4;
509         }
510
511         return result;
512 }
513
514 static const struct file_operations amdgpu_debugfs_ring_fops = {
515         .owner = THIS_MODULE,
516         .read = amdgpu_debugfs_ring_read,
517         .llseek = default_llseek
518 };
519
520 #endif
521
522 static int amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
523                                     struct amdgpu_ring *ring)
524 {
525 #if defined(CONFIG_DEBUG_FS)
526         struct drm_minor *minor = adev->ddev->primary;
527         struct dentry *ent, *root = minor->debugfs_root;
528         char name[32];
529
530         sprintf(name, "amdgpu_ring_%s", ring->name);
531
532         ent = debugfs_create_file(name,
533                                   S_IFREG | S_IRUGO, root,
534                                   ring, &amdgpu_debugfs_ring_fops);
535         if (!ent)
536                 return -ENOMEM;
537
538         i_size_write(ent->d_inode, ring->ring_size + 12);
539         ring->ent = ent;
540 #endif
541         return 0;
542 }
543
544 static void amdgpu_debugfs_ring_fini(struct amdgpu_ring *ring)
545 {
546 #if defined(CONFIG_DEBUG_FS)
547         debugfs_remove(ring->ent);
548 #endif
549 }
This page took 0.06502 seconds and 4 git commands to generate.