]> Git Repo - linux.git/blob - drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
Merge branch 'core-rcu-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[linux.git] / drivers / gpu / drm / amd / amdgpu / amdgpu_gfx.c
1 /*
2  * Copyright 2014 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  */
25 #include <drm/drmP.h>
26 #include "amdgpu.h"
27 #include "amdgpu_gfx.h"
28 #include "amdgpu_rlc.h"
29
30 /* delay 0.1 second to enable gfx off feature */
31 #define GFX_OFF_DELAY_ENABLE         msecs_to_jiffies(100)
32
33 /*
34  * GPU GFX IP block helpers function.
35  */
36
37 int amdgpu_gfx_queue_to_bit(struct amdgpu_device *adev, int mec,
38                             int pipe, int queue)
39 {
40         int bit = 0;
41
42         bit += mec * adev->gfx.mec.num_pipe_per_mec
43                 * adev->gfx.mec.num_queue_per_pipe;
44         bit += pipe * adev->gfx.mec.num_queue_per_pipe;
45         bit += queue;
46
47         return bit;
48 }
49
50 void amdgpu_gfx_bit_to_queue(struct amdgpu_device *adev, int bit,
51                              int *mec, int *pipe, int *queue)
52 {
53         *queue = bit % adev->gfx.mec.num_queue_per_pipe;
54         *pipe = (bit / adev->gfx.mec.num_queue_per_pipe)
55                 % adev->gfx.mec.num_pipe_per_mec;
56         *mec = (bit / adev->gfx.mec.num_queue_per_pipe)
57                / adev->gfx.mec.num_pipe_per_mec;
58
59 }
60
61 bool amdgpu_gfx_is_mec_queue_enabled(struct amdgpu_device *adev,
62                                      int mec, int pipe, int queue)
63 {
64         return test_bit(amdgpu_gfx_queue_to_bit(adev, mec, pipe, queue),
65                         adev->gfx.mec.queue_bitmap);
66 }
67
68 /**
69  * amdgpu_gfx_scratch_get - Allocate a scratch register
70  *
71  * @adev: amdgpu_device pointer
72  * @reg: scratch register mmio offset
73  *
74  * Allocate a CP scratch register for use by the driver (all asics).
75  * Returns 0 on success or -EINVAL on failure.
76  */
77 int amdgpu_gfx_scratch_get(struct amdgpu_device *adev, uint32_t *reg)
78 {
79         int i;
80
81         i = ffs(adev->gfx.scratch.free_mask);
82         if (i != 0 && i <= adev->gfx.scratch.num_reg) {
83                 i--;
84                 adev->gfx.scratch.free_mask &= ~(1u << i);
85                 *reg = adev->gfx.scratch.reg_base + i;
86                 return 0;
87         }
88         return -EINVAL;
89 }
90
91 /**
92  * amdgpu_gfx_scratch_free - Free a scratch register
93  *
94  * @adev: amdgpu_device pointer
95  * @reg: scratch register mmio offset
96  *
97  * Free a CP scratch register allocated for use by the driver (all asics)
98  */
99 void amdgpu_gfx_scratch_free(struct amdgpu_device *adev, uint32_t reg)
100 {
101         adev->gfx.scratch.free_mask |= 1u << (reg - adev->gfx.scratch.reg_base);
102 }
103
104 /**
105  * amdgpu_gfx_parse_disable_cu - Parse the disable_cu module parameter
106  *
107  * @mask: array in which the per-shader array disable masks will be stored
108  * @max_se: number of SEs
109  * @max_sh: number of SHs
110  *
111  * The bitmask of CUs to be disabled in the shader array determined by se and
112  * sh is stored in mask[se * max_sh + sh].
113  */
114 void amdgpu_gfx_parse_disable_cu(unsigned *mask, unsigned max_se, unsigned max_sh)
115 {
116         unsigned se, sh, cu;
117         const char *p;
118
119         memset(mask, 0, sizeof(*mask) * max_se * max_sh);
120
121         if (!amdgpu_disable_cu || !*amdgpu_disable_cu)
122                 return;
123
124         p = amdgpu_disable_cu;
125         for (;;) {
126                 char *next;
127                 int ret = sscanf(p, "%u.%u.%u", &se, &sh, &cu);
128                 if (ret < 3) {
129                         DRM_ERROR("amdgpu: could not parse disable_cu\n");
130                         return;
131                 }
132
133                 if (se < max_se && sh < max_sh && cu < 16) {
134                         DRM_INFO("amdgpu: disabling CU %u.%u.%u\n", se, sh, cu);
135                         mask[se * max_sh + sh] |= 1u << cu;
136                 } else {
137                         DRM_ERROR("amdgpu: disable_cu %u.%u.%u is out of range\n",
138                                   se, sh, cu);
139                 }
140
141                 next = strchr(p, ',');
142                 if (!next)
143                         break;
144                 p = next + 1;
145         }
146 }
147
148 static bool amdgpu_gfx_is_multipipe_capable(struct amdgpu_device *adev)
149 {
150         if (amdgpu_compute_multipipe != -1) {
151                 DRM_INFO("amdgpu: forcing compute pipe policy %d\n",
152                          amdgpu_compute_multipipe);
153                 return amdgpu_compute_multipipe == 1;
154         }
155
156         /* FIXME: spreading the queues across pipes causes perf regressions
157          * on POLARIS11 compute workloads */
158         if (adev->asic_type == CHIP_POLARIS11)
159                 return false;
160
161         return adev->gfx.mec.num_mec > 1;
162 }
163
164 void amdgpu_gfx_compute_queue_acquire(struct amdgpu_device *adev)
165 {
166         int i, queue, pipe, mec;
167         bool multipipe_policy = amdgpu_gfx_is_multipipe_capable(adev);
168
169         /* policy for amdgpu compute queue ownership */
170         for (i = 0; i < AMDGPU_MAX_COMPUTE_QUEUES; ++i) {
171                 queue = i % adev->gfx.mec.num_queue_per_pipe;
172                 pipe = (i / adev->gfx.mec.num_queue_per_pipe)
173                         % adev->gfx.mec.num_pipe_per_mec;
174                 mec = (i / adev->gfx.mec.num_queue_per_pipe)
175                         / adev->gfx.mec.num_pipe_per_mec;
176
177                 /* we've run out of HW */
178                 if (mec >= adev->gfx.mec.num_mec)
179                         break;
180
181                 if (multipipe_policy) {
182                         /* policy: amdgpu owns the first two queues of the first MEC */
183                         if (mec == 0 && queue < 2)
184                                 set_bit(i, adev->gfx.mec.queue_bitmap);
185                 } else {
186                         /* policy: amdgpu owns all queues in the first pipe */
187                         if (mec == 0 && pipe == 0)
188                                 set_bit(i, adev->gfx.mec.queue_bitmap);
189                 }
190         }
191
192         /* update the number of active compute rings */
193         adev->gfx.num_compute_rings =
194                 bitmap_weight(adev->gfx.mec.queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES);
195
196         /* If you hit this case and edited the policy, you probably just
197          * need to increase AMDGPU_MAX_COMPUTE_RINGS */
198         if (WARN_ON(adev->gfx.num_compute_rings > AMDGPU_MAX_COMPUTE_RINGS))
199                 adev->gfx.num_compute_rings = AMDGPU_MAX_COMPUTE_RINGS;
200 }
201
202 static int amdgpu_gfx_kiq_acquire(struct amdgpu_device *adev,
203                                   struct amdgpu_ring *ring)
204 {
205         int queue_bit;
206         int mec, pipe, queue;
207
208         queue_bit = adev->gfx.mec.num_mec
209                     * adev->gfx.mec.num_pipe_per_mec
210                     * adev->gfx.mec.num_queue_per_pipe;
211
212         while (queue_bit-- >= 0) {
213                 if (test_bit(queue_bit, adev->gfx.mec.queue_bitmap))
214                         continue;
215
216                 amdgpu_gfx_bit_to_queue(adev, queue_bit, &mec, &pipe, &queue);
217
218                 /*
219                  * 1. Using pipes 2/3 from MEC 2 seems cause problems.
220                  * 2. It must use queue id 0, because CGPG_IDLE/SAVE/LOAD/RUN
221                  * only can be issued on queue 0.
222                  */
223                 if ((mec == 1 && pipe > 1) || queue != 0)
224                         continue;
225
226                 ring->me = mec + 1;
227                 ring->pipe = pipe;
228                 ring->queue = queue;
229
230                 return 0;
231         }
232
233         dev_err(adev->dev, "Failed to find a queue for KIQ\n");
234         return -EINVAL;
235 }
236
237 int amdgpu_gfx_kiq_init_ring(struct amdgpu_device *adev,
238                              struct amdgpu_ring *ring,
239                              struct amdgpu_irq_src *irq)
240 {
241         struct amdgpu_kiq *kiq = &adev->gfx.kiq;
242         int r = 0;
243
244         spin_lock_init(&kiq->ring_lock);
245
246         r = amdgpu_device_wb_get(adev, &adev->virt.reg_val_offs);
247         if (r)
248                 return r;
249
250         ring->adev = NULL;
251         ring->ring_obj = NULL;
252         ring->use_doorbell = true;
253         ring->doorbell_index = adev->doorbell_index.kiq;
254
255         r = amdgpu_gfx_kiq_acquire(adev, ring);
256         if (r)
257                 return r;
258
259         ring->eop_gpu_addr = kiq->eop_gpu_addr;
260         sprintf(ring->name, "kiq_%d.%d.%d", ring->me, ring->pipe, ring->queue);
261         r = amdgpu_ring_init(adev, ring, 1024,
262                              irq, AMDGPU_CP_KIQ_IRQ_DRIVER0);
263         if (r)
264                 dev_warn(adev->dev, "(%d) failed to init kiq ring\n", r);
265
266         return r;
267 }
268
269 void amdgpu_gfx_kiq_free_ring(struct amdgpu_ring *ring,
270                               struct amdgpu_irq_src *irq)
271 {
272         amdgpu_device_wb_free(ring->adev, ring->adev->virt.reg_val_offs);
273         amdgpu_ring_fini(ring);
274 }
275
276 void amdgpu_gfx_kiq_fini(struct amdgpu_device *adev)
277 {
278         struct amdgpu_kiq *kiq = &adev->gfx.kiq;
279
280         amdgpu_bo_free_kernel(&kiq->eop_obj, &kiq->eop_gpu_addr, NULL);
281 }
282
283 int amdgpu_gfx_kiq_init(struct amdgpu_device *adev,
284                         unsigned hpd_size)
285 {
286         int r;
287         u32 *hpd;
288         struct amdgpu_kiq *kiq = &adev->gfx.kiq;
289
290         r = amdgpu_bo_create_kernel(adev, hpd_size, PAGE_SIZE,
291                                     AMDGPU_GEM_DOMAIN_GTT, &kiq->eop_obj,
292                                     &kiq->eop_gpu_addr, (void **)&hpd);
293         if (r) {
294                 dev_warn(adev->dev, "failed to create KIQ bo (%d).\n", r);
295                 return r;
296         }
297
298         memset(hpd, 0, hpd_size);
299
300         r = amdgpu_bo_reserve(kiq->eop_obj, true);
301         if (unlikely(r != 0))
302                 dev_warn(adev->dev, "(%d) reserve kiq eop bo failed\n", r);
303         amdgpu_bo_kunmap(kiq->eop_obj);
304         amdgpu_bo_unreserve(kiq->eop_obj);
305
306         return 0;
307 }
308
309 /* create MQD for each compute queue */
310 int amdgpu_gfx_compute_mqd_sw_init(struct amdgpu_device *adev,
311                                    unsigned mqd_size)
312 {
313         struct amdgpu_ring *ring = NULL;
314         int r, i;
315
316         /* create MQD for KIQ */
317         ring = &adev->gfx.kiq.ring;
318         if (!ring->mqd_obj) {
319                 /* originaly the KIQ MQD is put in GTT domain, but for SRIOV VRAM domain is a must
320                  * otherwise hypervisor trigger SAVE_VF fail after driver unloaded which mean MQD
321                  * deallocated and gart_unbind, to strict diverage we decide to use VRAM domain for
322                  * KIQ MQD no matter SRIOV or Bare-metal
323                  */
324                 r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
325                                             AMDGPU_GEM_DOMAIN_VRAM, &ring->mqd_obj,
326                                             &ring->mqd_gpu_addr, &ring->mqd_ptr);
327                 if (r) {
328                         dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
329                         return r;
330                 }
331
332                 /* prepare MQD backup */
333                 adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS] = kmalloc(mqd_size, GFP_KERNEL);
334                 if (!adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS])
335                                 dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
336         }
337
338         /* create MQD for each KCQ */
339         for (i = 0; i < adev->gfx.num_compute_rings; i++) {
340                 ring = &adev->gfx.compute_ring[i];
341                 if (!ring->mqd_obj) {
342                         r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
343                                                     AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
344                                                     &ring->mqd_gpu_addr, &ring->mqd_ptr);
345                         if (r) {
346                                 dev_warn(adev->dev, "failed to create ring mqd ob (%d)", r);
347                                 return r;
348                         }
349
350                         /* prepare MQD backup */
351                         adev->gfx.mec.mqd_backup[i] = kmalloc(mqd_size, GFP_KERNEL);
352                         if (!adev->gfx.mec.mqd_backup[i])
353                                 dev_warn(adev->dev, "no memory to create MQD backup for ring %s\n", ring->name);
354                 }
355         }
356
357         return 0;
358 }
359
360 void amdgpu_gfx_compute_mqd_sw_fini(struct amdgpu_device *adev)
361 {
362         struct amdgpu_ring *ring = NULL;
363         int i;
364
365         for (i = 0; i < adev->gfx.num_compute_rings; i++) {
366                 ring = &adev->gfx.compute_ring[i];
367                 kfree(adev->gfx.mec.mqd_backup[i]);
368                 amdgpu_bo_free_kernel(&ring->mqd_obj,
369                                       &ring->mqd_gpu_addr,
370                                       &ring->mqd_ptr);
371         }
372
373         ring = &adev->gfx.kiq.ring;
374         kfree(adev->gfx.mec.mqd_backup[AMDGPU_MAX_COMPUTE_RINGS]);
375         amdgpu_bo_free_kernel(&ring->mqd_obj,
376                               &ring->mqd_gpu_addr,
377                               &ring->mqd_ptr);
378 }
379
380 /* amdgpu_gfx_off_ctrl - Handle gfx off feature enable/disable
381  *
382  * @adev: amdgpu_device pointer
383  * @bool enable true: enable gfx off feature, false: disable gfx off feature
384  *
385  * 1. gfx off feature will be enabled by gfx ip after gfx cg gp enabled.
386  * 2. other client can send request to disable gfx off feature, the request should be honored.
387  * 3. other client can cancel their request of disable gfx off feature
388  * 4. other client should not send request to enable gfx off feature before disable gfx off feature.
389  */
390
391 void amdgpu_gfx_off_ctrl(struct amdgpu_device *adev, bool enable)
392 {
393         if (!(adev->powerplay.pp_feature & PP_GFXOFF_MASK))
394                 return;
395
396         if (!adev->powerplay.pp_funcs || !adev->powerplay.pp_funcs->set_powergating_by_smu)
397                 return;
398
399
400         mutex_lock(&adev->gfx.gfx_off_mutex);
401
402         if (!enable)
403                 adev->gfx.gfx_off_req_count++;
404         else if (adev->gfx.gfx_off_req_count > 0)
405                 adev->gfx.gfx_off_req_count--;
406
407         if (enable && !adev->gfx.gfx_off_state && !adev->gfx.gfx_off_req_count) {
408                 schedule_delayed_work(&adev->gfx.gfx_off_delay_work, GFX_OFF_DELAY_ENABLE);
409         } else if (!enable && adev->gfx.gfx_off_state) {
410                 if (!amdgpu_dpm_set_powergating_by_smu(adev, AMD_IP_BLOCK_TYPE_GFX, false))
411                         adev->gfx.gfx_off_state = false;
412         }
413
414         mutex_unlock(&adev->gfx.gfx_off_mutex);
415 }
This page took 0.059755 seconds and 4 git commands to generate.