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 <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 * @adev: amdgpu_device pointer
56 * @ring: amdgpu_ring structure holding ring information
57 * @ndw: number of dwords to allocate in the ring buffer
58 *
59 * Allocate @ndw dwords in the ring buffer (all asics).
60 * Returns 0 on success, error on failure.
61 */
62int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw)
63{
64 /* Align requested size with padding so unlock_commit can
65 * pad safely */
66 ndw = (ndw + ring->funcs->align_mask) & ~ring->funcs->align_mask;
67
68 /* Make sure we aren't trying to allocate more space
69 * than the maximum for one submission
70 */
71 if (WARN_ON_ONCE(ndw > ring->max_dw))
72 return -ENOMEM;
73
74 ring->count_dw = ndw;
75 ring->wptr_old = ring->wptr;
76
77 if (ring->funcs->begin_use)
78 ring->funcs->begin_use(ring);
79
80 return 0;
81}
82
83/** amdgpu_ring_insert_nop - insert NOP packets
84 *
85 * @ring: amdgpu_ring structure holding ring information
86 * @count: the number of NOP packets to insert
87 *
88 * This is the generic insert_nop function for rings except SDMA
89 */
90void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
91{
92 int i;
93
94 for (i = 0; i < count; i++)
95 amdgpu_ring_write(ring, ring->funcs->nop);
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 */
105void 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 * @adev: amdgpu_device pointer
116 * @ring: amdgpu_ring structure holding ring information
117 *
118 * Update the wptr (write pointer) to tell the GPU to
119 * execute new commands on the ring buffer (all asics).
120 */
121void amdgpu_ring_commit(struct amdgpu_ring *ring)
122{
123 uint32_t count;
124
125 /* We pad to match fetch size */
126 count = ring->funcs->align_mask + 1 -
127 (ring->wptr & ring->funcs->align_mask);
128 count %= ring->funcs->align_mask + 1;
129 ring->funcs->insert_nop(ring, count);
130
131 mb();
132 amdgpu_ring_set_wptr(ring);
133
134 if (ring->funcs->end_use)
135 ring->funcs->end_use(ring);
136}
137
138/**
139 * amdgpu_ring_undo - reset the wptr
140 *
141 * @ring: amdgpu_ring structure holding ring information
142 *
143 * Reset the driver's copy of the wptr (all asics).
144 */
145void amdgpu_ring_undo(struct amdgpu_ring *ring)
146{
147 ring->wptr = ring->wptr_old;
148
149 if (ring->funcs->end_use)
150 ring->funcs->end_use(ring);
151}
152
153/**
154 * amdgpu_ring_init - init driver ring struct.
155 *
156 * @adev: amdgpu_device pointer
157 * @ring: amdgpu_ring structure holding ring information
158 * @max_ndw: maximum number of dw for ring alloc
159 * @nop: nop packet for this ring
160 *
161 * Initialize the driver information for the selected ring (all asics).
162 * Returns 0 on success, error on failure.
163 */
164int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
165 unsigned int max_dw, struct amdgpu_irq_src *irq_src,
166 unsigned int irq_type, unsigned int hw_prio)
167{
168 int r, i;
169 int sched_hw_submission = amdgpu_sched_hw_submission;
170 u32 *num_sched;
171 u32 hw_ip;
172
173 /* Set the hw submission limit higher for KIQ because
174 * it's used for a number of gfx/compute tasks by both
175 * KFD and KGD which may have outstanding fences and
176 * it doesn't really use the gpu scheduler anyway;
177 * KIQ tasks get submitted directly to the ring.
178 */
179 if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
180 sched_hw_submission = max(sched_hw_submission, 256);
181 else if (ring == &adev->sdma.instance[0].page)
182 sched_hw_submission = 256;
183
184 if (ring->adev == NULL) {
185 if (adev->num_rings >= AMDGPU_MAX_RINGS)
186 return -EINVAL;
187
188 ring->adev = adev;
189 ring->idx = adev->num_rings++;
190 adev->rings[ring->idx] = ring;
191 r = amdgpu_fence_driver_init_ring(ring, sched_hw_submission);
192 if (r)
193 return r;
194 }
195
196 r = amdgpu_device_wb_get(adev, &ring->rptr_offs);
197 if (r) {
198 dev_err(adev->dev, "(%d) ring rptr_offs wb alloc failed\n", r);
199 return r;
200 }
201
202 r = amdgpu_device_wb_get(adev, &ring->wptr_offs);
203 if (r) {
204 dev_err(adev->dev, "(%d) ring wptr_offs wb alloc failed\n", r);
205 return r;
206 }
207
208 r = amdgpu_device_wb_get(adev, &ring->fence_offs);
209 if (r) {
210 dev_err(adev->dev, "(%d) ring fence_offs wb alloc failed\n", r);
211 return r;
212 }
213
214 r = amdgpu_device_wb_get(adev, &ring->trail_fence_offs);
215 if (r) {
216 dev_err(adev->dev,
217 "(%d) ring trail_fence_offs wb alloc failed\n", r);
218 return r;
219 }
220 ring->trail_fence_gpu_addr =
221 adev->wb.gpu_addr + (ring->trail_fence_offs * 4);
222 ring->trail_fence_cpu_addr = &adev->wb.wb[ring->trail_fence_offs];
223
224 r = amdgpu_device_wb_get(adev, &ring->cond_exe_offs);
225 if (r) {
226 dev_err(adev->dev, "(%d) ring cond_exec_polling wb alloc failed\n", r);
227 return r;
228 }
229 ring->cond_exe_gpu_addr = adev->wb.gpu_addr + (ring->cond_exe_offs * 4);
230 ring->cond_exe_cpu_addr = &adev->wb.wb[ring->cond_exe_offs];
231 /* always set cond_exec_polling to CONTINUE */
232 *ring->cond_exe_cpu_addr = 1;
233
234 r = amdgpu_fence_driver_start_ring(ring, irq_src, irq_type);
235 if (r) {
236 dev_err(adev->dev, "failed initializing fences (%d).\n", r);
237 return r;
238 }
239
240 ring->ring_size = roundup_pow_of_two(max_dw * 4 * sched_hw_submission);
241
242 ring->buf_mask = (ring->ring_size / 4) - 1;
243 ring->ptr_mask = ring->funcs->support_64bit_ptrs ?
244 0xffffffffffffffff : ring->buf_mask;
245 /* Allocate ring buffer */
246 if (ring->ring_obj == NULL) {
247 r = amdgpu_bo_create_kernel(adev, ring->ring_size + ring->funcs->extra_dw, PAGE_SIZE,
248 AMDGPU_GEM_DOMAIN_GTT,
249 &ring->ring_obj,
250 &ring->gpu_addr,
251 (void **)&ring->ring);
252 if (r) {
253 dev_err(adev->dev, "(%d) ring create failed\n", r);
254 return r;
255 }
256 amdgpu_ring_clear_ring(ring);
257 }
258
259 ring->max_dw = max_dw;
260 ring->priority = DRM_SCHED_PRIORITY_NORMAL;
261 mutex_init(&ring->priority_mutex);
262
263 if (!ring->no_scheduler) {
264 hw_ip = ring->funcs->type;
265 num_sched = &adev->gpu_sched[hw_ip][hw_prio].num_scheds;
266 adev->gpu_sched[hw_ip][hw_prio].sched[(*num_sched)++] =
267 &ring->sched;
268 }
269
270 for (i = 0; i < DRM_SCHED_PRIORITY_MAX; ++i)
271 atomic_set(&ring->num_jobs[i], 0);
272
273 return 0;
274}
275
276/**
277 * amdgpu_ring_fini - tear down the driver ring struct.
278 *
279 * @adev: amdgpu_device pointer
280 * @ring: amdgpu_ring structure holding ring information
281 *
282 * Tear down the driver information for the selected ring (all asics).
283 */
284void amdgpu_ring_fini(struct amdgpu_ring *ring)
285{
286
287 /* Not to finish a ring which is not initialized */
288 if (!(ring->adev) || !(ring->adev->rings[ring->idx]))
289 return;
290
291 ring->sched.ready = false;
292
293 amdgpu_device_wb_free(ring->adev, ring->rptr_offs);
294 amdgpu_device_wb_free(ring->adev, ring->wptr_offs);
295
296 amdgpu_device_wb_free(ring->adev, ring->cond_exe_offs);
297 amdgpu_device_wb_free(ring->adev, ring->fence_offs);
298
299 amdgpu_bo_free_kernel(&ring->ring_obj,
300 &ring->gpu_addr,
301 (void **)&ring->ring);
302
303 dma_fence_put(ring->vmid_wait);
304 ring->vmid_wait = NULL;
305 ring->me = 0;
306
307 ring->adev->rings[ring->idx] = NULL;
308}
309
310/**
311 * amdgpu_ring_emit_reg_write_reg_wait_helper - ring helper
312 *
313 * @adev: amdgpu_device pointer
314 * @reg0: register to write
315 * @reg1: register to wait on
316 * @ref: reference value to write/wait on
317 * @mask: mask to wait on
318 *
319 * Helper for rings that don't support write and wait in a
320 * single oneshot packet.
321 */
322void amdgpu_ring_emit_reg_write_reg_wait_helper(struct amdgpu_ring *ring,
323 uint32_t reg0, uint32_t reg1,
324 uint32_t ref, uint32_t mask)
325{
326 amdgpu_ring_emit_wreg(ring, reg0, ref);
327 amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
328}
329
330/**
331 * amdgpu_ring_soft_recovery - try to soft recover a ring lockup
332 *
333 * @ring: ring to try the recovery on
334 * @vmid: VMID we try to get going again
335 * @fence: timedout fence
336 *
337 * Tries to get a ring proceeding again when it is stuck.
338 */
339bool amdgpu_ring_soft_recovery(struct amdgpu_ring *ring, unsigned int vmid,
340 struct dma_fence *fence)
341{
342 ktime_t deadline = ktime_add_us(ktime_get(), 10000);
343
344 if (amdgpu_sriov_vf(ring->adev) || !ring->funcs->soft_recovery || !fence)
345 return false;
346
347 atomic_inc(&ring->adev->gpu_reset_counter);
348 while (!dma_fence_is_signaled(fence) &&
349 ktime_to_ns(ktime_sub(deadline, ktime_get())) > 0)
350 ring->funcs->soft_recovery(ring, vmid);
351
352 return dma_fence_is_signaled(fence);
353}
354
355/*
356 * Debugfs info
357 */
358#if defined(CONFIG_DEBUG_FS)
359
360/* Layout of file is 12 bytes consisting of
361 * - rptr
362 * - wptr
363 * - driver's copy of wptr
364 *
365 * followed by n-words of ring data
366 */
367static ssize_t amdgpu_debugfs_ring_read(struct file *f, char __user *buf,
368 size_t size, loff_t *pos)
369{
370 struct amdgpu_ring *ring = file_inode(f)->i_private;
371 int r, i;
372 uint32_t value, result, early[3];
373
374 if (*pos & 3 || size & 3)
375 return -EINVAL;
376
377 result = 0;
378
379 if (*pos < 12) {
380 early[0] = amdgpu_ring_get_rptr(ring) & ring->buf_mask;
381 early[1] = amdgpu_ring_get_wptr(ring) & ring->buf_mask;
382 early[2] = ring->wptr & ring->buf_mask;
383 for (i = *pos / 4; i < 3 && size; i++) {
384 r = put_user(early[i], (uint32_t *)buf);
385 if (r)
386 return r;
387 buf += 4;
388 result += 4;
389 size -= 4;
390 *pos += 4;
391 }
392 }
393
394 while (size) {
395 if (*pos >= (ring->ring_size + 12))
396 return result;
397
398 value = ring->ring[(*pos - 12)/4];
399 r = put_user(value, (uint32_t*)buf);
400 if (r)
401 return r;
402 buf += 4;
403 result += 4;
404 size -= 4;
405 *pos += 4;
406 }
407
408 return result;
409}
410
411static const struct file_operations amdgpu_debugfs_ring_fops = {
412 .owner = THIS_MODULE,
413 .read = amdgpu_debugfs_ring_read,
414 .llseek = default_llseek
415};
416
417#endif
418
419int amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
420 struct amdgpu_ring *ring)
421{
422#if defined(CONFIG_DEBUG_FS)
423 struct drm_minor *minor = adev->ddev->primary;
424 struct dentry *ent, *root = minor->debugfs_root;
425 char name[32];
426
427 sprintf(name, "amdgpu_ring_%s", ring->name);
428
429 ent = debugfs_create_file(name,
430 S_IFREG | S_IRUGO, root,
431 ring, &amdgpu_debugfs_ring_fops);
432 if (!ent)
433 return -ENOMEM;
434
435 i_size_write(ent->d_inode, ring->ring_size + 12);
436 ring->ent = ent;
437#endif
438 return 0;
439}
440
441/**
442 * amdgpu_ring_test_helper - tests ring and set sched readiness status
443 *
444 * @ring: ring to try the recovery on
445 *
446 * Tests ring and set sched readiness status
447 *
448 * Returns 0 on success, error on failure.
449 */
450int amdgpu_ring_test_helper(struct amdgpu_ring *ring)
451{
452 struct amdgpu_device *adev = ring->adev;
453 int r;
454
455 r = amdgpu_ring_test_ring(ring);
456 if (r)
457 DRM_DEV_ERROR(adev->dev, "ring %s test failed (%d)\n",
458 ring->name, r);
459 else
460 DRM_DEV_DEBUG(adev->dev, "ring test on %s succeeded\n",
461 ring->name);
462
463 ring->sched.ready = !r;
464 return r;
465}
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_max_ibs - Return max IBs that fit in a single submission.
54 *
55 * @type: ring type for which to return the limit.
56 */
57unsigned int amdgpu_ring_max_ibs(enum amdgpu_ring_type type)
58{
59 switch (type) {
60 case AMDGPU_RING_TYPE_GFX:
61 /* Need to keep at least 192 on GFX7+ for old radv. */
62 return 192;
63 case AMDGPU_RING_TYPE_COMPUTE:
64 return 125;
65 case AMDGPU_RING_TYPE_VCN_JPEG:
66 return 16;
67 default:
68 return 49;
69 }
70}
71
72/**
73 * amdgpu_ring_alloc - allocate space on the ring buffer
74 *
75 * @ring: amdgpu_ring structure holding ring information
76 * @ndw: number of dwords to allocate in the ring buffer
77 *
78 * Allocate @ndw dwords in the ring buffer (all asics).
79 * Returns 0 on success, error on failure.
80 */
81int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned int ndw)
82{
83 /* Align requested size with padding so unlock_commit can
84 * pad safely */
85 ndw = (ndw + ring->funcs->align_mask) & ~ring->funcs->align_mask;
86
87 /* Make sure we aren't trying to allocate more space
88 * than the maximum for one submission
89 */
90 if (WARN_ON_ONCE(ndw > ring->max_dw))
91 return -ENOMEM;
92
93 ring->count_dw = ndw;
94 ring->wptr_old = ring->wptr;
95
96 if (ring->funcs->begin_use)
97 ring->funcs->begin_use(ring);
98
99 return 0;
100}
101
102/** amdgpu_ring_insert_nop - insert NOP packets
103 *
104 * @ring: amdgpu_ring structure holding ring information
105 * @count: the number of NOP packets to insert
106 *
107 * This is the generic insert_nop function for rings except SDMA
108 */
109void amdgpu_ring_insert_nop(struct amdgpu_ring *ring, uint32_t count)
110{
111 uint32_t occupied, chunk1, chunk2;
112
113 occupied = ring->wptr & ring->buf_mask;
114 chunk1 = ring->buf_mask + 1 - occupied;
115 chunk1 = (chunk1 >= count) ? count : chunk1;
116 chunk2 = count - chunk1;
117
118 if (chunk1)
119 memset32(&ring->ring[occupied], ring->funcs->nop, chunk1);
120
121 if (chunk2)
122 memset32(ring->ring, ring->funcs->nop, chunk2);
123
124 ring->wptr += count;
125 ring->wptr &= ring->ptr_mask;
126 ring->count_dw -= count;
127}
128
129/**
130 * amdgpu_ring_generic_pad_ib - pad IB with NOP packets
131 *
132 * @ring: amdgpu_ring structure holding ring information
133 * @ib: IB to add NOP packets to
134 *
135 * This is the generic pad_ib function for rings except SDMA
136 */
137void amdgpu_ring_generic_pad_ib(struct amdgpu_ring *ring, struct amdgpu_ib *ib)
138{
139 while (ib->length_dw & ring->funcs->align_mask)
140 ib->ptr[ib->length_dw++] = ring->funcs->nop;
141}
142
143/**
144 * amdgpu_ring_commit - tell the GPU to execute the new
145 * commands on the ring buffer
146 *
147 * @ring: amdgpu_ring structure holding ring information
148 *
149 * Update the wptr (write pointer) to tell the GPU to
150 * execute new commands on the ring buffer (all asics).
151 */
152void amdgpu_ring_commit(struct amdgpu_ring *ring)
153{
154 uint32_t count;
155
156 if (ring->count_dw < 0)
157 DRM_ERROR("amdgpu: writing more dwords to the ring than expected!\n");
158
159 /* We pad to match fetch size */
160 count = ring->funcs->align_mask + 1 -
161 (ring->wptr & ring->funcs->align_mask);
162 count &= ring->funcs->align_mask;
163
164 if (count != 0)
165 ring->funcs->insert_nop(ring, count);
166
167 mb();
168 amdgpu_ring_set_wptr(ring);
169
170 if (ring->funcs->end_use)
171 ring->funcs->end_use(ring);
172}
173
174/**
175 * amdgpu_ring_undo - reset the wptr
176 *
177 * @ring: amdgpu_ring structure holding ring information
178 *
179 * Reset the driver's copy of the wptr (all asics).
180 */
181void amdgpu_ring_undo(struct amdgpu_ring *ring)
182{
183 ring->wptr = ring->wptr_old;
184
185 if (ring->funcs->end_use)
186 ring->funcs->end_use(ring);
187}
188
189#define amdgpu_ring_get_gpu_addr(ring, offset) \
190 (ring->is_mes_queue ? \
191 (ring->mes_ctx->meta_data_gpu_addr + offset) : \
192 (ring->adev->wb.gpu_addr + offset * 4))
193
194#define amdgpu_ring_get_cpu_addr(ring, offset) \
195 (ring->is_mes_queue ? \
196 (void *)((uint8_t *)(ring->mes_ctx->meta_data_ptr) + offset) : \
197 (&ring->adev->wb.wb[offset]))
198
199/**
200 * amdgpu_ring_init - init driver ring struct.
201 *
202 * @adev: amdgpu_device pointer
203 * @ring: amdgpu_ring structure holding ring information
204 * @max_dw: maximum number of dw for ring alloc
205 * @irq_src: interrupt source to use for this ring
206 * @irq_type: interrupt type to use for this ring
207 * @hw_prio: ring priority (NORMAL/HIGH)
208 * @sched_score: optional score atomic shared with other schedulers
209 *
210 * Initialize the driver information for the selected ring (all asics).
211 * Returns 0 on success, error on failure.
212 */
213int amdgpu_ring_init(struct amdgpu_device *adev, struct amdgpu_ring *ring,
214 unsigned int max_dw, struct amdgpu_irq_src *irq_src,
215 unsigned int irq_type, unsigned int hw_prio,
216 atomic_t *sched_score)
217{
218 int r;
219 int sched_hw_submission = amdgpu_sched_hw_submission;
220 u32 *num_sched;
221 u32 hw_ip;
222 unsigned int max_ibs_dw;
223
224 /* Set the hw submission limit higher for KIQ because
225 * it's used for a number of gfx/compute tasks by both
226 * KFD and KGD which may have outstanding fences and
227 * it doesn't really use the gpu scheduler anyway;
228 * KIQ tasks get submitted directly to the ring.
229 */
230 if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
231 sched_hw_submission = max(sched_hw_submission, 256);
232 if (ring->funcs->type == AMDGPU_RING_TYPE_MES)
233 sched_hw_submission = 8;
234 else if (ring == &adev->sdma.instance[0].page)
235 sched_hw_submission = 256;
236
237 if (ring->adev == NULL) {
238 if (adev->num_rings >= AMDGPU_MAX_RINGS)
239 return -EINVAL;
240
241 ring->adev = adev;
242 ring->num_hw_submission = sched_hw_submission;
243 ring->sched_score = sched_score;
244 ring->vmid_wait = dma_fence_get_stub();
245
246 if (!ring->is_mes_queue) {
247 ring->idx = adev->num_rings++;
248 adev->rings[ring->idx] = ring;
249 }
250
251 r = amdgpu_fence_driver_init_ring(ring);
252 if (r)
253 return r;
254 }
255
256 if (ring->is_mes_queue) {
257 ring->rptr_offs = amdgpu_mes_ctx_get_offs(ring,
258 AMDGPU_MES_CTX_RPTR_OFFS);
259 ring->wptr_offs = amdgpu_mes_ctx_get_offs(ring,
260 AMDGPU_MES_CTX_WPTR_OFFS);
261 ring->fence_offs = amdgpu_mes_ctx_get_offs(ring,
262 AMDGPU_MES_CTX_FENCE_OFFS);
263 ring->trail_fence_offs = amdgpu_mes_ctx_get_offs(ring,
264 AMDGPU_MES_CTX_TRAIL_FENCE_OFFS);
265 ring->cond_exe_offs = amdgpu_mes_ctx_get_offs(ring,
266 AMDGPU_MES_CTX_COND_EXE_OFFS);
267 } else {
268 r = amdgpu_device_wb_get(adev, &ring->rptr_offs);
269 if (r) {
270 dev_err(adev->dev, "(%d) ring rptr_offs wb alloc failed\n", r);
271 return r;
272 }
273
274 r = amdgpu_device_wb_get(adev, &ring->wptr_offs);
275 if (r) {
276 dev_err(adev->dev, "(%d) ring wptr_offs wb alloc failed\n", r);
277 return r;
278 }
279
280 r = amdgpu_device_wb_get(adev, &ring->fence_offs);
281 if (r) {
282 dev_err(adev->dev, "(%d) ring fence_offs wb alloc failed\n", r);
283 return r;
284 }
285
286 r = amdgpu_device_wb_get(adev, &ring->trail_fence_offs);
287 if (r) {
288 dev_err(adev->dev, "(%d) ring trail_fence_offs wb alloc failed\n", r);
289 return r;
290 }
291
292 r = amdgpu_device_wb_get(adev, &ring->cond_exe_offs);
293 if (r) {
294 dev_err(adev->dev, "(%d) ring cond_exec_polling wb alloc failed\n", r);
295 return r;
296 }
297 }
298
299 ring->fence_gpu_addr =
300 amdgpu_ring_get_gpu_addr(ring, ring->fence_offs);
301 ring->fence_cpu_addr =
302 amdgpu_ring_get_cpu_addr(ring, ring->fence_offs);
303
304 ring->rptr_gpu_addr =
305 amdgpu_ring_get_gpu_addr(ring, ring->rptr_offs);
306 ring->rptr_cpu_addr =
307 amdgpu_ring_get_cpu_addr(ring, ring->rptr_offs);
308
309 ring->wptr_gpu_addr =
310 amdgpu_ring_get_gpu_addr(ring, ring->wptr_offs);
311 ring->wptr_cpu_addr =
312 amdgpu_ring_get_cpu_addr(ring, ring->wptr_offs);
313
314 ring->trail_fence_gpu_addr =
315 amdgpu_ring_get_gpu_addr(ring, ring->trail_fence_offs);
316 ring->trail_fence_cpu_addr =
317 amdgpu_ring_get_cpu_addr(ring, ring->trail_fence_offs);
318
319 ring->cond_exe_gpu_addr =
320 amdgpu_ring_get_gpu_addr(ring, ring->cond_exe_offs);
321 ring->cond_exe_cpu_addr =
322 amdgpu_ring_get_cpu_addr(ring, ring->cond_exe_offs);
323
324 /* always set cond_exec_polling to CONTINUE */
325 *ring->cond_exe_cpu_addr = 1;
326
327 r = amdgpu_fence_driver_start_ring(ring, irq_src, irq_type);
328 if (r) {
329 dev_err(adev->dev, "failed initializing fences (%d).\n", r);
330 return r;
331 }
332
333 max_ibs_dw = ring->funcs->emit_frame_size +
334 amdgpu_ring_max_ibs(ring->funcs->type) * ring->funcs->emit_ib_size;
335 max_ibs_dw = (max_ibs_dw + ring->funcs->align_mask) & ~ring->funcs->align_mask;
336
337 if (WARN_ON(max_ibs_dw > max_dw))
338 max_dw = max_ibs_dw;
339
340 ring->ring_size = roundup_pow_of_two(max_dw * 4 * sched_hw_submission);
341
342 ring->buf_mask = (ring->ring_size / 4) - 1;
343 ring->ptr_mask = ring->funcs->support_64bit_ptrs ?
344 0xffffffffffffffff : ring->buf_mask;
345
346 /* Allocate ring buffer */
347 if (ring->is_mes_queue) {
348 int offset = 0;
349
350 BUG_ON(ring->ring_size > PAGE_SIZE*4);
351
352 offset = amdgpu_mes_ctx_get_offs(ring,
353 AMDGPU_MES_CTX_RING_OFFS);
354 ring->gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset);
355 ring->ring = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset);
356 amdgpu_ring_clear_ring(ring);
357
358 } else if (ring->ring_obj == NULL) {
359 r = amdgpu_bo_create_kernel(adev, ring->ring_size + ring->funcs->extra_dw, PAGE_SIZE,
360 AMDGPU_GEM_DOMAIN_GTT,
361 &ring->ring_obj,
362 &ring->gpu_addr,
363 (void **)&ring->ring);
364 if (r) {
365 dev_err(adev->dev, "(%d) ring create failed\n", r);
366 return r;
367 }
368 amdgpu_ring_clear_ring(ring);
369 }
370
371 ring->max_dw = max_dw;
372 ring->hw_prio = hw_prio;
373
374 if (!ring->no_scheduler && ring->funcs->type < AMDGPU_HW_IP_NUM) {
375 hw_ip = ring->funcs->type;
376 num_sched = &adev->gpu_sched[hw_ip][hw_prio].num_scheds;
377 adev->gpu_sched[hw_ip][hw_prio].sched[(*num_sched)++] =
378 &ring->sched;
379 }
380
381 return 0;
382}
383
384/**
385 * amdgpu_ring_fini - tear down the driver ring struct.
386 *
387 * @ring: amdgpu_ring structure holding ring information
388 *
389 * Tear down the driver information for the selected ring (all asics).
390 */
391void amdgpu_ring_fini(struct amdgpu_ring *ring)
392{
393
394 /* Not to finish a ring which is not initialized */
395 if (!(ring->adev) ||
396 (!ring->is_mes_queue && !(ring->adev->rings[ring->idx])))
397 return;
398
399 ring->sched.ready = false;
400
401 if (!ring->is_mes_queue) {
402 amdgpu_device_wb_free(ring->adev, ring->rptr_offs);
403 amdgpu_device_wb_free(ring->adev, ring->wptr_offs);
404
405 amdgpu_device_wb_free(ring->adev, ring->cond_exe_offs);
406 amdgpu_device_wb_free(ring->adev, ring->fence_offs);
407
408 amdgpu_bo_free_kernel(&ring->ring_obj,
409 &ring->gpu_addr,
410 (void **)&ring->ring);
411 } else {
412 kfree(ring->fence_drv.fences);
413 }
414
415 dma_fence_put(ring->vmid_wait);
416 ring->vmid_wait = NULL;
417 ring->me = 0;
418
419 if (!ring->is_mes_queue)
420 ring->adev->rings[ring->idx] = NULL;
421}
422
423/**
424 * amdgpu_ring_emit_reg_write_reg_wait_helper - ring helper
425 *
426 * @ring: ring to write to
427 * @reg0: register to write
428 * @reg1: register to wait on
429 * @ref: reference value to write/wait on
430 * @mask: mask to wait on
431 *
432 * Helper for rings that don't support write and wait in a
433 * single oneshot packet.
434 */
435void amdgpu_ring_emit_reg_write_reg_wait_helper(struct amdgpu_ring *ring,
436 uint32_t reg0, uint32_t reg1,
437 uint32_t ref, uint32_t mask)
438{
439 amdgpu_ring_emit_wreg(ring, reg0, ref);
440 amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
441}
442
443/**
444 * amdgpu_ring_soft_recovery - try to soft recover a ring lockup
445 *
446 * @ring: ring to try the recovery on
447 * @vmid: VMID we try to get going again
448 * @fence: timedout fence
449 *
450 * Tries to get a ring proceeding again when it is stuck.
451 */
452bool amdgpu_ring_soft_recovery(struct amdgpu_ring *ring, unsigned int vmid,
453 struct dma_fence *fence)
454{
455 unsigned long flags;
456 ktime_t deadline;
457
458 if (unlikely(ring->adev->debug_disable_soft_recovery))
459 return false;
460
461 deadline = ktime_add_us(ktime_get(), 10000);
462
463 if (amdgpu_sriov_vf(ring->adev) || !ring->funcs->soft_recovery || !fence)
464 return false;
465
466 spin_lock_irqsave(fence->lock, flags);
467 if (!dma_fence_is_signaled_locked(fence))
468 dma_fence_set_error(fence, -ENODATA);
469 spin_unlock_irqrestore(fence->lock, flags);
470
471 atomic_inc(&ring->adev->gpu_reset_counter);
472 while (!dma_fence_is_signaled(fence) &&
473 ktime_to_ns(ktime_sub(deadline, ktime_get())) > 0)
474 ring->funcs->soft_recovery(ring, vmid);
475
476 return dma_fence_is_signaled(fence);
477}
478
479/*
480 * Debugfs info
481 */
482#if defined(CONFIG_DEBUG_FS)
483
484/* Layout of file is 12 bytes consisting of
485 * - rptr
486 * - wptr
487 * - driver's copy of wptr
488 *
489 * followed by n-words of ring data
490 */
491static ssize_t amdgpu_debugfs_ring_read(struct file *f, char __user *buf,
492 size_t size, loff_t *pos)
493{
494 struct amdgpu_ring *ring = file_inode(f)->i_private;
495 uint32_t value, result, early[3];
496 loff_t i;
497 int r;
498
499 if (*pos & 3 || size & 3)
500 return -EINVAL;
501
502 result = 0;
503
504 if (*pos < 12) {
505 early[0] = amdgpu_ring_get_rptr(ring) & ring->buf_mask;
506 early[1] = amdgpu_ring_get_wptr(ring) & ring->buf_mask;
507 early[2] = ring->wptr & ring->buf_mask;
508 for (i = *pos / 4; i < 3 && size; i++) {
509 r = put_user(early[i], (uint32_t *)buf);
510 if (r)
511 return r;
512 buf += 4;
513 result += 4;
514 size -= 4;
515 *pos += 4;
516 }
517 }
518
519 while (size) {
520 if (*pos >= (ring->ring_size + 12))
521 return result;
522
523 value = ring->ring[(*pos - 12)/4];
524 r = put_user(value, (uint32_t *)buf);
525 if (r)
526 return r;
527 buf += 4;
528 result += 4;
529 size -= 4;
530 *pos += 4;
531 }
532
533 return result;
534}
535
536static const struct file_operations amdgpu_debugfs_ring_fops = {
537 .owner = THIS_MODULE,
538 .read = amdgpu_debugfs_ring_read,
539 .llseek = default_llseek
540};
541
542static ssize_t amdgpu_debugfs_mqd_read(struct file *f, char __user *buf,
543 size_t size, loff_t *pos)
544{
545 struct amdgpu_ring *ring = file_inode(f)->i_private;
546 volatile u32 *mqd;
547 u32 *kbuf;
548 int r, i;
549 uint32_t value, result;
550
551 if (*pos & 3 || size & 3)
552 return -EINVAL;
553
554 kbuf = kmalloc(ring->mqd_size, GFP_KERNEL);
555 if (!kbuf)
556 return -ENOMEM;
557
558 r = amdgpu_bo_reserve(ring->mqd_obj, false);
559 if (unlikely(r != 0))
560 goto err_free;
561
562 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&mqd);
563 if (r)
564 goto err_unreserve;
565
566 /*
567 * Copy to local buffer to avoid put_user(), which might fault
568 * and acquire mmap_sem, under reservation_ww_class_mutex.
569 */
570 for (i = 0; i < ring->mqd_size/sizeof(u32); i++)
571 kbuf[i] = mqd[i];
572
573 amdgpu_bo_kunmap(ring->mqd_obj);
574 amdgpu_bo_unreserve(ring->mqd_obj);
575
576 result = 0;
577 while (size) {
578 if (*pos >= ring->mqd_size)
579 break;
580
581 value = kbuf[*pos/4];
582 r = put_user(value, (uint32_t *)buf);
583 if (r)
584 goto err_free;
585 buf += 4;
586 result += 4;
587 size -= 4;
588 *pos += 4;
589 }
590
591 kfree(kbuf);
592 return result;
593
594err_unreserve:
595 amdgpu_bo_unreserve(ring->mqd_obj);
596err_free:
597 kfree(kbuf);
598 return r;
599}
600
601static const struct file_operations amdgpu_debugfs_mqd_fops = {
602 .owner = THIS_MODULE,
603 .read = amdgpu_debugfs_mqd_read,
604 .llseek = default_llseek
605};
606
607static int amdgpu_debugfs_ring_error(void *data, u64 val)
608{
609 struct amdgpu_ring *ring = data;
610
611 amdgpu_fence_driver_set_error(ring, val);
612 return 0;
613}
614
615DEFINE_DEBUGFS_ATTRIBUTE_SIGNED(amdgpu_debugfs_error_fops, NULL,
616 amdgpu_debugfs_ring_error, "%lld\n");
617
618#endif
619
620void amdgpu_debugfs_ring_init(struct amdgpu_device *adev,
621 struct amdgpu_ring *ring)
622{
623#if defined(CONFIG_DEBUG_FS)
624 struct drm_minor *minor = adev_to_drm(adev)->primary;
625 struct dentry *root = minor->debugfs_root;
626 char name[32];
627
628 sprintf(name, "amdgpu_ring_%s", ring->name);
629 debugfs_create_file_size(name, S_IFREG | 0444, root, ring,
630 &amdgpu_debugfs_ring_fops,
631 ring->ring_size + 12);
632
633 if (ring->mqd_obj) {
634 sprintf(name, "amdgpu_mqd_%s", ring->name);
635 debugfs_create_file_size(name, S_IFREG | 0444, root, ring,
636 &amdgpu_debugfs_mqd_fops,
637 ring->mqd_size);
638 }
639
640 sprintf(name, "amdgpu_error_%s", ring->name);
641 debugfs_create_file(name, 0200, root, ring,
642 &amdgpu_debugfs_error_fops);
643
644#endif
645}
646
647/**
648 * amdgpu_ring_test_helper - tests ring and set sched readiness status
649 *
650 * @ring: ring to try the recovery on
651 *
652 * Tests ring and set sched readiness status
653 *
654 * Returns 0 on success, error on failure.
655 */
656int amdgpu_ring_test_helper(struct amdgpu_ring *ring)
657{
658 struct amdgpu_device *adev = ring->adev;
659 int r;
660
661 r = amdgpu_ring_test_ring(ring);
662 if (r)
663 DRM_DEV_ERROR(adev->dev, "ring %s test failed (%d)\n",
664 ring->name, r);
665 else
666 DRM_DEV_DEBUG(adev->dev, "ring test on %s succeeded\n",
667 ring->name);
668
669 ring->sched.ready = !r;
670
671 return r;
672}
673
674static void amdgpu_ring_to_mqd_prop(struct amdgpu_ring *ring,
675 struct amdgpu_mqd_prop *prop)
676{
677 struct amdgpu_device *adev = ring->adev;
678 bool is_high_prio_compute = ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE &&
679 amdgpu_gfx_is_high_priority_compute_queue(adev, ring);
680 bool is_high_prio_gfx = ring->funcs->type == AMDGPU_RING_TYPE_GFX &&
681 amdgpu_gfx_is_high_priority_graphics_queue(adev, ring);
682
683 memset(prop, 0, sizeof(*prop));
684
685 prop->mqd_gpu_addr = ring->mqd_gpu_addr;
686 prop->hqd_base_gpu_addr = ring->gpu_addr;
687 prop->rptr_gpu_addr = ring->rptr_gpu_addr;
688 prop->wptr_gpu_addr = ring->wptr_gpu_addr;
689 prop->queue_size = ring->ring_size;
690 prop->eop_gpu_addr = ring->eop_gpu_addr;
691 prop->use_doorbell = ring->use_doorbell;
692 prop->doorbell_index = ring->doorbell_index;
693
694 /* map_queues packet doesn't need activate the queue,
695 * so only kiq need set this field.
696 */
697 prop->hqd_active = ring->funcs->type == AMDGPU_RING_TYPE_KIQ;
698
699 prop->allow_tunneling = is_high_prio_compute;
700 if (is_high_prio_compute || is_high_prio_gfx) {
701 prop->hqd_pipe_priority = AMDGPU_GFX_PIPE_PRIO_HIGH;
702 prop->hqd_queue_priority = AMDGPU_GFX_QUEUE_PRIORITY_MAXIMUM;
703 }
704}
705
706int amdgpu_ring_init_mqd(struct amdgpu_ring *ring)
707{
708 struct amdgpu_device *adev = ring->adev;
709 struct amdgpu_mqd *mqd_mgr;
710 struct amdgpu_mqd_prop prop;
711
712 amdgpu_ring_to_mqd_prop(ring, &prop);
713
714 ring->wptr = 0;
715
716 if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ)
717 mqd_mgr = &adev->mqds[AMDGPU_HW_IP_COMPUTE];
718 else
719 mqd_mgr = &adev->mqds[ring->funcs->type];
720
721 return mqd_mgr->init_mqd(adev, ring->mqd_ptr, &prop);
722}
723
724void amdgpu_ring_ib_begin(struct amdgpu_ring *ring)
725{
726 if (ring->is_sw_ring)
727 amdgpu_sw_ring_ib_begin(ring);
728}
729
730void amdgpu_ring_ib_end(struct amdgpu_ring *ring)
731{
732 if (ring->is_sw_ring)
733 amdgpu_sw_ring_ib_end(ring);
734}
735
736void amdgpu_ring_ib_on_emit_cntl(struct amdgpu_ring *ring)
737{
738 if (ring->is_sw_ring)
739 amdgpu_sw_ring_ib_mark_offset(ring, AMDGPU_MUX_OFFSET_TYPE_CONTROL);
740}
741
742void amdgpu_ring_ib_on_emit_ce(struct amdgpu_ring *ring)
743{
744 if (ring->is_sw_ring)
745 amdgpu_sw_ring_ib_mark_offset(ring, AMDGPU_MUX_OFFSET_TYPE_CE);
746}
747
748void amdgpu_ring_ib_on_emit_de(struct amdgpu_ring *ring)
749{
750 if (ring->is_sw_ring)
751 amdgpu_sw_ring_ib_mark_offset(ring, AMDGPU_MUX_OFFSET_TYPE_DE);
752}
753
754bool amdgpu_ring_sched_ready(struct amdgpu_ring *ring)
755{
756 if (!ring)
757 return false;
758
759 if (ring->no_scheduler || !drm_sched_wqueue_ready(&ring->sched))
760 return false;
761
762 return true;
763}