drm/i915: Remove impossible tests for dev->dev_private
[deliverable/linux.git] / drivers / gpu / drm / i915 / i915_guc_submission.c
1 /*
2 * Copyright © 2014 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 */
24 #include <linux/firmware.h>
25 #include <linux/circ_buf.h>
26 #include "i915_drv.h"
27 #include "intel_guc.h"
28
29 /**
30 * DOC: GuC-based command submission
31 *
32 * i915_guc_client:
33 * We use the term client to avoid confusion with contexts. A i915_guc_client is
34 * equivalent to GuC object guc_context_desc. This context descriptor is
35 * allocated from a pool of 1024 entries. Kernel driver will allocate doorbell
36 * and workqueue for it. Also the process descriptor (guc_process_desc), which
37 * is mapped to client space. So the client can write Work Item then ring the
38 * doorbell.
39 *
40 * To simplify the implementation, we allocate one gem object that contains all
41 * pages for doorbell, process descriptor and workqueue.
42 *
43 * The Scratch registers:
44 * There are 16 MMIO-based registers start from 0xC180. The kernel driver writes
45 * a value to the action register (SOFT_SCRATCH_0) along with any data. It then
46 * triggers an interrupt on the GuC via another register write (0xC4C8).
47 * Firmware writes a success/fail code back to the action register after
48 * processes the request. The kernel driver polls waiting for this update and
49 * then proceeds.
50 * See host2guc_action()
51 *
52 * Doorbells:
53 * Doorbells are interrupts to uKernel. A doorbell is a single cache line (QW)
54 * mapped into process space.
55 *
56 * Work Items:
57 * There are several types of work items that the host may place into a
58 * workqueue, each with its own requirements and limitations. Currently only
59 * WQ_TYPE_INORDER is needed to support legacy submission via GuC, which
60 * represents in-order queue. The kernel driver packs ring tail pointer and an
61 * ELSP context descriptor dword into Work Item.
62 * See guc_add_workqueue_item()
63 *
64 */
65
66 /*
67 * Read GuC command/status register (SOFT_SCRATCH_0)
68 * Return true if it contains a response rather than a command
69 */
70 static inline bool host2guc_action_response(struct drm_i915_private *dev_priv,
71 u32 *status)
72 {
73 u32 val = I915_READ(SOFT_SCRATCH(0));
74 *status = val;
75 return GUC2HOST_IS_RESPONSE(val);
76 }
77
78 static int host2guc_action(struct intel_guc *guc, u32 *data, u32 len)
79 {
80 struct drm_i915_private *dev_priv = guc_to_i915(guc);
81 u32 status;
82 int i;
83 int ret;
84
85 if (WARN_ON(len < 1 || len > 15))
86 return -EINVAL;
87
88 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
89
90 dev_priv->guc.action_count += 1;
91 dev_priv->guc.action_cmd = data[0];
92
93 for (i = 0; i < len; i++)
94 I915_WRITE(SOFT_SCRATCH(i), data[i]);
95
96 POSTING_READ(SOFT_SCRATCH(i - 1));
97
98 I915_WRITE(HOST2GUC_INTERRUPT, HOST2GUC_TRIGGER);
99
100 /* No HOST2GUC command should take longer than 10ms */
101 ret = wait_for_atomic(host2guc_action_response(dev_priv, &status), 10);
102 if (status != GUC2HOST_STATUS_SUCCESS) {
103 /*
104 * Either the GuC explicitly returned an error (which
105 * we convert to -EIO here) or no response at all was
106 * received within the timeout limit (-ETIMEDOUT)
107 */
108 if (ret != -ETIMEDOUT)
109 ret = -EIO;
110
111 DRM_ERROR("GUC: host2guc action 0x%X failed. ret=%d "
112 "status=0x%08X response=0x%08X\n",
113 data[0], ret, status,
114 I915_READ(SOFT_SCRATCH(15)));
115
116 dev_priv->guc.action_fail += 1;
117 dev_priv->guc.action_err = ret;
118 }
119 dev_priv->guc.action_status = status;
120
121 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
122
123 return ret;
124 }
125
126 /*
127 * Tell the GuC to allocate or deallocate a specific doorbell
128 */
129
130 static int host2guc_allocate_doorbell(struct intel_guc *guc,
131 struct i915_guc_client *client)
132 {
133 u32 data[2];
134
135 data[0] = HOST2GUC_ACTION_ALLOCATE_DOORBELL;
136 data[1] = client->ctx_index;
137
138 return host2guc_action(guc, data, 2);
139 }
140
141 static int host2guc_release_doorbell(struct intel_guc *guc,
142 struct i915_guc_client *client)
143 {
144 u32 data[2];
145
146 data[0] = HOST2GUC_ACTION_DEALLOCATE_DOORBELL;
147 data[1] = client->ctx_index;
148
149 return host2guc_action(guc, data, 2);
150 }
151
152 static int host2guc_sample_forcewake(struct intel_guc *guc,
153 struct i915_guc_client *client)
154 {
155 struct drm_i915_private *dev_priv = guc_to_i915(guc);
156 u32 data[2];
157
158 data[0] = HOST2GUC_ACTION_SAMPLE_FORCEWAKE;
159 /* WaRsDisableCoarsePowerGating:skl,bxt */
160 if (!intel_enable_rc6() || NEEDS_WaRsDisableCoarsePowerGating(dev_priv))
161 data[1] = 0;
162 else
163 /* bit 0 and 1 are for Render and Media domain separately */
164 data[1] = GUC_FORCEWAKE_RENDER | GUC_FORCEWAKE_MEDIA;
165
166 return host2guc_action(guc, data, ARRAY_SIZE(data));
167 }
168
169 /*
170 * Initialise, update, or clear doorbell data shared with the GuC
171 *
172 * These functions modify shared data and so need access to the mapped
173 * client object which contains the page being used for the doorbell
174 */
175
176 static int guc_update_doorbell_id(struct intel_guc *guc,
177 struct i915_guc_client *client,
178 u16 new_id)
179 {
180 struct sg_table *sg = guc->ctx_pool_obj->pages;
181 void *doorbell_bitmap = guc->doorbell_bitmap;
182 struct guc_doorbell_info *doorbell;
183 struct guc_context_desc desc;
184 size_t len;
185
186 doorbell = client->client_base + client->doorbell_offset;
187
188 if (client->doorbell_id != GUC_INVALID_DOORBELL_ID &&
189 test_bit(client->doorbell_id, doorbell_bitmap)) {
190 /* Deactivate the old doorbell */
191 doorbell->db_status = GUC_DOORBELL_DISABLED;
192 (void)host2guc_release_doorbell(guc, client);
193 __clear_bit(client->doorbell_id, doorbell_bitmap);
194 }
195
196 /* Update the GuC's idea of the doorbell ID */
197 len = sg_pcopy_to_buffer(sg->sgl, sg->nents, &desc, sizeof(desc),
198 sizeof(desc) * client->ctx_index);
199 if (len != sizeof(desc))
200 return -EFAULT;
201 desc.db_id = new_id;
202 len = sg_pcopy_from_buffer(sg->sgl, sg->nents, &desc, sizeof(desc),
203 sizeof(desc) * client->ctx_index);
204 if (len != sizeof(desc))
205 return -EFAULT;
206
207 client->doorbell_id = new_id;
208 if (new_id == GUC_INVALID_DOORBELL_ID)
209 return 0;
210
211 /* Activate the new doorbell */
212 __set_bit(new_id, doorbell_bitmap);
213 doorbell->cookie = 0;
214 doorbell->db_status = GUC_DOORBELL_ENABLED;
215 return host2guc_allocate_doorbell(guc, client);
216 }
217
218 static int guc_init_doorbell(struct intel_guc *guc,
219 struct i915_guc_client *client,
220 uint16_t db_id)
221 {
222 return guc_update_doorbell_id(guc, client, db_id);
223 }
224
225 static void guc_disable_doorbell(struct intel_guc *guc,
226 struct i915_guc_client *client)
227 {
228 (void)guc_update_doorbell_id(guc, client, GUC_INVALID_DOORBELL_ID);
229
230 /* XXX: wait for any interrupts */
231 /* XXX: wait for workqueue to drain */
232 }
233
234 static uint16_t
235 select_doorbell_register(struct intel_guc *guc, uint32_t priority)
236 {
237 /*
238 * The bitmap tracks which doorbell registers are currently in use.
239 * It is split into two halves; the first half is used for normal
240 * priority contexts, the second half for high-priority ones.
241 * Note that logically higher priorities are numerically less than
242 * normal ones, so the test below means "is it high-priority?"
243 */
244 const bool hi_pri = (priority <= GUC_CTX_PRIORITY_HIGH);
245 const uint16_t half = GUC_MAX_DOORBELLS / 2;
246 const uint16_t start = hi_pri ? half : 0;
247 const uint16_t end = start + half;
248 uint16_t id;
249
250 id = find_next_zero_bit(guc->doorbell_bitmap, end, start);
251 if (id == end)
252 id = GUC_INVALID_DOORBELL_ID;
253
254 DRM_DEBUG_DRIVER("assigned %s priority doorbell id 0x%x\n",
255 hi_pri ? "high" : "normal", id);
256
257 return id;
258 }
259
260 /*
261 * Select, assign and relase doorbell cachelines
262 *
263 * These functions track which doorbell cachelines are in use.
264 * The data they manipulate is protected by the host2guc lock.
265 */
266
267 static uint32_t select_doorbell_cacheline(struct intel_guc *guc)
268 {
269 const uint32_t cacheline_size = cache_line_size();
270 uint32_t offset;
271
272 /* Doorbell uses a single cache line within a page */
273 offset = offset_in_page(guc->db_cacheline);
274
275 /* Moving to next cache line to reduce contention */
276 guc->db_cacheline += cacheline_size;
277
278 DRM_DEBUG_DRIVER("selected doorbell cacheline 0x%x, next 0x%x, linesize %u\n",
279 offset, guc->db_cacheline, cacheline_size);
280
281 return offset;
282 }
283
284 /*
285 * Initialise the process descriptor shared with the GuC firmware.
286 */
287 static void guc_init_proc_desc(struct intel_guc *guc,
288 struct i915_guc_client *client)
289 {
290 struct guc_process_desc *desc;
291
292 desc = client->client_base + client->proc_desc_offset;
293
294 memset(desc, 0, sizeof(*desc));
295
296 /*
297 * XXX: pDoorbell and WQVBaseAddress are pointers in process address
298 * space for ring3 clients (set them as in mmap_ioctl) or kernel
299 * space for kernel clients (map on demand instead? May make debug
300 * easier to have it mapped).
301 */
302 desc->wq_base_addr = 0;
303 desc->db_base_addr = 0;
304
305 desc->context_id = client->ctx_index;
306 desc->wq_size_bytes = client->wq_size;
307 desc->wq_status = WQ_STATUS_ACTIVE;
308 desc->priority = client->priority;
309 }
310
311 /*
312 * Initialise/clear the context descriptor shared with the GuC firmware.
313 *
314 * This descriptor tells the GuC where (in GGTT space) to find the important
315 * data structures relating to this client (doorbell, process descriptor,
316 * write queue, etc).
317 */
318
319 static void guc_init_ctx_desc(struct intel_guc *guc,
320 struct i915_guc_client *client)
321 {
322 struct drm_i915_gem_object *client_obj = client->client_obj;
323 struct drm_i915_private *dev_priv = guc_to_i915(guc);
324 struct intel_engine_cs *engine;
325 struct i915_gem_context *ctx = client->owner;
326 struct guc_context_desc desc;
327 struct sg_table *sg;
328 u32 gfx_addr;
329
330 memset(&desc, 0, sizeof(desc));
331
332 desc.attribute = GUC_CTX_DESC_ATTR_ACTIVE | GUC_CTX_DESC_ATTR_KERNEL;
333 desc.context_id = client->ctx_index;
334 desc.priority = client->priority;
335 desc.db_id = client->doorbell_id;
336
337 for_each_engine(engine, dev_priv) {
338 struct intel_context *ce = &ctx->engine[engine->id];
339 struct guc_execlist_context *lrc = &desc.lrc[engine->guc_id];
340 struct drm_i915_gem_object *obj;
341
342 /* TODO: We have a design issue to be solved here. Only when we
343 * receive the first batch, we know which engine is used by the
344 * user. But here GuC expects the lrc and ring to be pinned. It
345 * is not an issue for default context, which is the only one
346 * for now who owns a GuC client. But for future owner of GuC
347 * client, need to make sure lrc is pinned prior to enter here.
348 */
349 if (!ce->state)
350 break; /* XXX: continue? */
351
352 lrc->context_desc = lower_32_bits(ce->lrc_desc);
353
354 /* The state page is after PPHWSP */
355 gfx_addr = i915_gem_obj_ggtt_offset(ce->state);
356 lrc->ring_lcra = gfx_addr + LRC_STATE_PN * PAGE_SIZE;
357 lrc->context_id = (client->ctx_index << GUC_ELC_CTXID_OFFSET) |
358 (engine->guc_id << GUC_ELC_ENGINE_OFFSET);
359
360 obj = ce->ringbuf->obj;
361 gfx_addr = i915_gem_obj_ggtt_offset(obj);
362
363 lrc->ring_begin = gfx_addr;
364 lrc->ring_end = gfx_addr + obj->base.size - 1;
365 lrc->ring_next_free_location = gfx_addr;
366 lrc->ring_current_tail_pointer_value = 0;
367
368 desc.engines_used |= (1 << engine->guc_id);
369 }
370
371 WARN_ON(desc.engines_used == 0);
372
373 /*
374 * The doorbell, process descriptor, and workqueue are all parts
375 * of the client object, which the GuC will reference via the GGTT
376 */
377 gfx_addr = i915_gem_obj_ggtt_offset(client_obj);
378 desc.db_trigger_phy = sg_dma_address(client_obj->pages->sgl) +
379 client->doorbell_offset;
380 desc.db_trigger_cpu = (uintptr_t)client->client_base +
381 client->doorbell_offset;
382 desc.db_trigger_uk = gfx_addr + client->doorbell_offset;
383 desc.process_desc = gfx_addr + client->proc_desc_offset;
384 desc.wq_addr = gfx_addr + client->wq_offset;
385 desc.wq_size = client->wq_size;
386
387 /*
388 * XXX: Take LRCs from an existing context if this is not an
389 * IsKMDCreatedContext client
390 */
391 desc.desc_private = (uintptr_t)client;
392
393 /* Pool context is pinned already */
394 sg = guc->ctx_pool_obj->pages;
395 sg_pcopy_from_buffer(sg->sgl, sg->nents, &desc, sizeof(desc),
396 sizeof(desc) * client->ctx_index);
397 }
398
399 static void guc_fini_ctx_desc(struct intel_guc *guc,
400 struct i915_guc_client *client)
401 {
402 struct guc_context_desc desc;
403 struct sg_table *sg;
404
405 memset(&desc, 0, sizeof(desc));
406
407 sg = guc->ctx_pool_obj->pages;
408 sg_pcopy_from_buffer(sg->sgl, sg->nents, &desc, sizeof(desc),
409 sizeof(desc) * client->ctx_index);
410 }
411
412 /**
413 * i915_guc_wq_check_space() - check that the GuC can accept a request
414 * @request: request associated with the commands
415 *
416 * Return: 0 if space is available
417 * -EAGAIN if space is not currently available
418 *
419 * This function must be called (and must return 0) before a request
420 * is submitted to the GuC via i915_guc_submit() below. Once a result
421 * of 0 has been returned, it remains valid until (but only until)
422 * the next call to submit().
423 *
424 * This precheck allows the caller to determine in advance that space
425 * will be available for the next submission before committing resources
426 * to it, and helps avoid late failures with complicated recovery paths.
427 */
428 int i915_guc_wq_check_space(struct drm_i915_gem_request *request)
429 {
430 const size_t wqi_size = sizeof(struct guc_wq_item);
431 struct i915_guc_client *gc = request->i915->guc.execbuf_client;
432 struct guc_process_desc *desc;
433 u32 freespace;
434
435 GEM_BUG_ON(gc == NULL);
436
437 desc = gc->client_base + gc->proc_desc_offset;
438
439 freespace = CIRC_SPACE(gc->wq_tail, desc->head, gc->wq_size);
440 if (likely(freespace >= wqi_size))
441 return 0;
442
443 gc->no_wq_space += 1;
444
445 return -EAGAIN;
446 }
447
448 static void guc_add_workqueue_item(struct i915_guc_client *gc,
449 struct drm_i915_gem_request *rq)
450 {
451 /* wqi_len is in DWords, and does not include the one-word header */
452 const size_t wqi_size = sizeof(struct guc_wq_item);
453 const u32 wqi_len = wqi_size/sizeof(u32) - 1;
454 struct guc_process_desc *desc;
455 struct guc_wq_item *wqi;
456 void *base;
457 u32 freespace, tail, wq_off, wq_page;
458
459 desc = gc->client_base + gc->proc_desc_offset;
460
461 /* Free space is guaranteed, see i915_guc_wq_check_space() above */
462 freespace = CIRC_SPACE(gc->wq_tail, desc->head, gc->wq_size);
463 GEM_BUG_ON(freespace < wqi_size);
464
465 /* The GuC firmware wants the tail index in QWords, not bytes */
466 tail = rq->tail;
467 GEM_BUG_ON(tail & 7);
468 tail >>= 3;
469 GEM_BUG_ON(tail > WQ_RING_TAIL_MAX);
470
471 /* For now workqueue item is 4 DWs; workqueue buffer is 2 pages. So we
472 * should not have the case where structure wqi is across page, neither
473 * wrapped to the beginning. This simplifies the implementation below.
474 *
475 * XXX: if not the case, we need save data to a temp wqi and copy it to
476 * workqueue buffer dw by dw.
477 */
478 BUILD_BUG_ON(wqi_size != 16);
479
480 /* postincrement WQ tail for next time */
481 wq_off = gc->wq_tail;
482 gc->wq_tail += wqi_size;
483 gc->wq_tail &= gc->wq_size - 1;
484 GEM_BUG_ON(wq_off & (wqi_size - 1));
485
486 /* WQ starts from the page after doorbell / process_desc */
487 wq_page = (wq_off + GUC_DB_SIZE) >> PAGE_SHIFT;
488 wq_off &= PAGE_SIZE - 1;
489 base = kmap_atomic(i915_gem_object_get_page(gc->client_obj, wq_page));
490 wqi = (struct guc_wq_item *)((char *)base + wq_off);
491
492 /* Now fill in the 4-word work queue item */
493 wqi->header = WQ_TYPE_INORDER |
494 (wqi_len << WQ_LEN_SHIFT) |
495 (rq->engine->guc_id << WQ_TARGET_SHIFT) |
496 WQ_NO_WCFLUSH_WAIT;
497
498 /* The GuC wants only the low-order word of the context descriptor */
499 wqi->context_desc = (u32)intel_lr_context_descriptor(rq->ctx,
500 rq->engine);
501
502 wqi->ring_tail = tail << WQ_RING_TAIL_SHIFT;
503 wqi->fence_id = rq->seqno;
504
505 kunmap_atomic(base);
506 }
507
508 static int guc_ring_doorbell(struct i915_guc_client *gc)
509 {
510 struct guc_process_desc *desc;
511 union guc_doorbell_qw db_cmp, db_exc, db_ret;
512 union guc_doorbell_qw *db;
513 int attempt = 2, ret = -EAGAIN;
514
515 desc = gc->client_base + gc->proc_desc_offset;
516
517 /* Update the tail so it is visible to GuC */
518 desc->tail = gc->wq_tail;
519
520 /* current cookie */
521 db_cmp.db_status = GUC_DOORBELL_ENABLED;
522 db_cmp.cookie = gc->cookie;
523
524 /* cookie to be updated */
525 db_exc.db_status = GUC_DOORBELL_ENABLED;
526 db_exc.cookie = gc->cookie + 1;
527 if (db_exc.cookie == 0)
528 db_exc.cookie = 1;
529
530 /* pointer of current doorbell cacheline */
531 db = gc->client_base + gc->doorbell_offset;
532
533 while (attempt--) {
534 /* lets ring the doorbell */
535 db_ret.value_qw = atomic64_cmpxchg((atomic64_t *)db,
536 db_cmp.value_qw, db_exc.value_qw);
537
538 /* if the exchange was successfully executed */
539 if (db_ret.value_qw == db_cmp.value_qw) {
540 /* db was successfully rung */
541 gc->cookie = db_exc.cookie;
542 ret = 0;
543 break;
544 }
545
546 /* XXX: doorbell was lost and need to acquire it again */
547 if (db_ret.db_status == GUC_DOORBELL_DISABLED)
548 break;
549
550 DRM_ERROR("Cookie mismatch. Expected %d, returned %d\n",
551 db_cmp.cookie, db_ret.cookie);
552
553 /* update the cookie to newly read cookie from GuC */
554 db_cmp.cookie = db_ret.cookie;
555 db_exc.cookie = db_ret.cookie + 1;
556 if (db_exc.cookie == 0)
557 db_exc.cookie = 1;
558 }
559
560 return ret;
561 }
562
563 /**
564 * i915_guc_submit() - Submit commands through GuC
565 * @rq: request associated with the commands
566 *
567 * Return: 0 on success, otherwise an errno.
568 * (Note: nonzero really shouldn't happen!)
569 *
570 * The caller must have already called i915_guc_wq_check_space() above
571 * with a result of 0 (success) since the last request submission. This
572 * guarantees that there is space in the work queue for the new request,
573 * so enqueuing the item cannot fail.
574 *
575 * Bad Things Will Happen if the caller violates this protocol e.g. calls
576 * submit() when check() says there's no space, or calls submit() multiple
577 * times with no intervening check().
578 *
579 * The only error here arises if the doorbell hardware isn't functioning
580 * as expected, which really shouln't happen.
581 */
582 int i915_guc_submit(struct drm_i915_gem_request *rq)
583 {
584 unsigned int engine_id = rq->engine->id;
585 struct intel_guc *guc = &rq->i915->guc;
586 struct i915_guc_client *client = guc->execbuf_client;
587 int b_ret;
588
589 guc_add_workqueue_item(client, rq);
590 b_ret = guc_ring_doorbell(client);
591
592 client->submissions[engine_id] += 1;
593 client->retcode = b_ret;
594 if (b_ret)
595 client->b_fail += 1;
596
597 guc->submissions[engine_id] += 1;
598 guc->last_seqno[engine_id] = rq->seqno;
599
600 return b_ret;
601 }
602
603 /*
604 * Everything below here is concerned with setup & teardown, and is
605 * therefore not part of the somewhat time-critical batch-submission
606 * path of i915_guc_submit() above.
607 */
608
609 /**
610 * gem_allocate_guc_obj() - Allocate gem object for GuC usage
611 * @dev_priv: driver private data structure
612 * @size: size of object
613 *
614 * This is a wrapper to create a gem obj. In order to use it inside GuC, the
615 * object needs to be pinned lifetime. Also we must pin it to gtt space other
616 * than [0, GUC_WOPCM_TOP) because this range is reserved inside GuC.
617 *
618 * Return: A drm_i915_gem_object if successful, otherwise NULL.
619 */
620 static struct drm_i915_gem_object *
621 gem_allocate_guc_obj(struct drm_i915_private *dev_priv, u32 size)
622 {
623 struct drm_i915_gem_object *obj;
624
625 obj = i915_gem_object_create(dev_priv->dev, size);
626 if (IS_ERR(obj))
627 return NULL;
628
629 if (i915_gem_object_get_pages(obj)) {
630 drm_gem_object_unreference(&obj->base);
631 return NULL;
632 }
633
634 if (i915_gem_obj_ggtt_pin(obj, PAGE_SIZE,
635 PIN_OFFSET_BIAS | GUC_WOPCM_TOP)) {
636 drm_gem_object_unreference(&obj->base);
637 return NULL;
638 }
639
640 /* Invalidate GuC TLB to let GuC take the latest updates to GTT. */
641 I915_WRITE(GEN8_GTCR, GEN8_GTCR_INVALIDATE);
642
643 return obj;
644 }
645
646 /**
647 * gem_release_guc_obj() - Release gem object allocated for GuC usage
648 * @obj: gem obj to be released
649 */
650 static void gem_release_guc_obj(struct drm_i915_gem_object *obj)
651 {
652 if (!obj)
653 return;
654
655 if (i915_gem_obj_is_pinned(obj))
656 i915_gem_object_ggtt_unpin(obj);
657
658 drm_gem_object_unreference(&obj->base);
659 }
660
661 static void
662 guc_client_free(struct drm_i915_private *dev_priv,
663 struct i915_guc_client *client)
664 {
665 struct intel_guc *guc = &dev_priv->guc;
666
667 if (!client)
668 return;
669
670 /*
671 * XXX: wait for any outstanding submissions before freeing memory.
672 * Be sure to drop any locks
673 */
674
675 if (client->client_base) {
676 /*
677 * If we got as far as setting up a doorbell, make sure we
678 * shut it down before unmapping & deallocating the memory.
679 */
680 guc_disable_doorbell(guc, client);
681
682 kunmap(kmap_to_page(client->client_base));
683 }
684
685 gem_release_guc_obj(client->client_obj);
686
687 if (client->ctx_index != GUC_INVALID_CTX_ID) {
688 guc_fini_ctx_desc(guc, client);
689 ida_simple_remove(&guc->ctx_ids, client->ctx_index);
690 }
691
692 kfree(client);
693 }
694
695 /*
696 * Borrow the first client to set up & tear down every doorbell
697 * in turn, to ensure that all doorbell h/w is (re)initialised.
698 */
699 static void guc_init_doorbell_hw(struct intel_guc *guc)
700 {
701 struct drm_i915_private *dev_priv = guc_to_i915(guc);
702 struct i915_guc_client *client = guc->execbuf_client;
703 uint16_t db_id, i;
704 int err;
705
706 db_id = client->doorbell_id;
707
708 for (i = 0; i < GUC_MAX_DOORBELLS; ++i) {
709 i915_reg_t drbreg = GEN8_DRBREGL(i);
710 u32 value = I915_READ(drbreg);
711
712 err = guc_update_doorbell_id(guc, client, i);
713
714 /* Report update failure or unexpectedly active doorbell */
715 if (err || (i != db_id && (value & GUC_DOORBELL_ENABLED)))
716 DRM_DEBUG_DRIVER("Doorbell %d (reg 0x%x) was 0x%x, err %d\n",
717 i, drbreg.reg, value, err);
718 }
719
720 /* Restore to original value */
721 err = guc_update_doorbell_id(guc, client, db_id);
722 if (err)
723 DRM_ERROR("Failed to restore doorbell to %d, err %d\n",
724 db_id, err);
725
726 for (i = 0; i < GUC_MAX_DOORBELLS; ++i) {
727 i915_reg_t drbreg = GEN8_DRBREGL(i);
728 u32 value = I915_READ(drbreg);
729
730 if (i != db_id && (value & GUC_DOORBELL_ENABLED))
731 DRM_DEBUG_DRIVER("Doorbell %d (reg 0x%x) finally 0x%x\n",
732 i, drbreg.reg, value);
733
734 }
735 }
736
737 /**
738 * guc_client_alloc() - Allocate an i915_guc_client
739 * @dev_priv: driver private data structure
740 * @priority: four levels priority _CRITICAL, _HIGH, _NORMAL and _LOW
741 * The kernel client to replace ExecList submission is created with
742 * NORMAL priority. Priority of a client for scheduler can be HIGH,
743 * while a preemption context can use CRITICAL.
744 * @ctx: the context that owns the client (we use the default render
745 * context)
746 *
747 * Return: An i915_guc_client object if success, else NULL.
748 */
749 static struct i915_guc_client *
750 guc_client_alloc(struct drm_i915_private *dev_priv,
751 uint32_t priority,
752 struct i915_gem_context *ctx)
753 {
754 struct i915_guc_client *client;
755 struct intel_guc *guc = &dev_priv->guc;
756 struct drm_i915_gem_object *obj;
757 uint16_t db_id;
758
759 client = kzalloc(sizeof(*client), GFP_KERNEL);
760 if (!client)
761 return NULL;
762
763 client->doorbell_id = GUC_INVALID_DOORBELL_ID;
764 client->priority = priority;
765 client->owner = ctx;
766 client->guc = guc;
767
768 client->ctx_index = (uint32_t)ida_simple_get(&guc->ctx_ids, 0,
769 GUC_MAX_GPU_CONTEXTS, GFP_KERNEL);
770 if (client->ctx_index >= GUC_MAX_GPU_CONTEXTS) {
771 client->ctx_index = GUC_INVALID_CTX_ID;
772 goto err;
773 }
774
775 /* The first page is doorbell/proc_desc. Two followed pages are wq. */
776 obj = gem_allocate_guc_obj(dev_priv, GUC_DB_SIZE + GUC_WQ_SIZE);
777 if (!obj)
778 goto err;
779
780 /* We'll keep just the first (doorbell/proc) page permanently kmap'd. */
781 client->client_obj = obj;
782 client->client_base = kmap(i915_gem_object_get_page(obj, 0));
783 client->wq_offset = GUC_DB_SIZE;
784 client->wq_size = GUC_WQ_SIZE;
785
786 db_id = select_doorbell_register(guc, client->priority);
787 if (db_id == GUC_INVALID_DOORBELL_ID)
788 /* XXX: evict a doorbell instead? */
789 goto err;
790
791 client->doorbell_offset = select_doorbell_cacheline(guc);
792
793 /*
794 * Since the doorbell only requires a single cacheline, we can save
795 * space by putting the application process descriptor in the same
796 * page. Use the half of the page that doesn't include the doorbell.
797 */
798 if (client->doorbell_offset >= (GUC_DB_SIZE / 2))
799 client->proc_desc_offset = 0;
800 else
801 client->proc_desc_offset = (GUC_DB_SIZE / 2);
802
803 guc_init_proc_desc(guc, client);
804 guc_init_ctx_desc(guc, client);
805 if (guc_init_doorbell(guc, client, db_id))
806 goto err;
807
808 DRM_DEBUG_DRIVER("new priority %u client %p: ctx_index %u\n",
809 priority, client, client->ctx_index);
810 DRM_DEBUG_DRIVER("doorbell id %u, cacheline offset 0x%x\n",
811 client->doorbell_id, client->doorbell_offset);
812
813 return client;
814
815 err:
816 DRM_ERROR("FAILED to create priority %u GuC client!\n", priority);
817
818 guc_client_free(dev_priv, client);
819 return NULL;
820 }
821
822 static void guc_create_log(struct intel_guc *guc)
823 {
824 struct drm_i915_private *dev_priv = guc_to_i915(guc);
825 struct drm_i915_gem_object *obj;
826 unsigned long offset;
827 uint32_t size, flags;
828
829 if (i915.guc_log_level < GUC_LOG_VERBOSITY_MIN)
830 return;
831
832 if (i915.guc_log_level > GUC_LOG_VERBOSITY_MAX)
833 i915.guc_log_level = GUC_LOG_VERBOSITY_MAX;
834
835 /* The first page is to save log buffer state. Allocate one
836 * extra page for others in case for overlap */
837 size = (1 + GUC_LOG_DPC_PAGES + 1 +
838 GUC_LOG_ISR_PAGES + 1 +
839 GUC_LOG_CRASH_PAGES + 1) << PAGE_SHIFT;
840
841 obj = guc->log_obj;
842 if (!obj) {
843 obj = gem_allocate_guc_obj(dev_priv, size);
844 if (!obj) {
845 /* logging will be off */
846 i915.guc_log_level = -1;
847 return;
848 }
849
850 guc->log_obj = obj;
851 }
852
853 /* each allocated unit is a page */
854 flags = GUC_LOG_VALID | GUC_LOG_NOTIFY_ON_HALF_FULL |
855 (GUC_LOG_DPC_PAGES << GUC_LOG_DPC_SHIFT) |
856 (GUC_LOG_ISR_PAGES << GUC_LOG_ISR_SHIFT) |
857 (GUC_LOG_CRASH_PAGES << GUC_LOG_CRASH_SHIFT);
858
859 offset = i915_gem_obj_ggtt_offset(obj) >> PAGE_SHIFT; /* in pages */
860 guc->log_flags = (offset << GUC_LOG_BUF_ADDR_SHIFT) | flags;
861 }
862
863 static void init_guc_policies(struct guc_policies *policies)
864 {
865 struct guc_policy *policy;
866 u32 p, i;
867
868 policies->dpc_promote_time = 500000;
869 policies->max_num_work_items = POLICY_MAX_NUM_WI;
870
871 for (p = 0; p < GUC_CTX_PRIORITY_NUM; p++) {
872 for (i = GUC_RENDER_ENGINE; i < GUC_MAX_ENGINES_NUM; i++) {
873 policy = &policies->policy[p][i];
874
875 policy->execution_quantum = 1000000;
876 policy->preemption_time = 500000;
877 policy->fault_time = 250000;
878 policy->policy_flags = 0;
879 }
880 }
881
882 policies->is_valid = 1;
883 }
884
885 static void guc_create_ads(struct intel_guc *guc)
886 {
887 struct drm_i915_private *dev_priv = guc_to_i915(guc);
888 struct drm_i915_gem_object *obj;
889 struct guc_ads *ads;
890 struct guc_policies *policies;
891 struct guc_mmio_reg_state *reg_state;
892 struct intel_engine_cs *engine;
893 struct page *page;
894 u32 size;
895
896 /* The ads obj includes the struct itself and buffers passed to GuC */
897 size = sizeof(struct guc_ads) + sizeof(struct guc_policies) +
898 sizeof(struct guc_mmio_reg_state) +
899 GUC_S3_SAVE_SPACE_PAGES * PAGE_SIZE;
900
901 obj = guc->ads_obj;
902 if (!obj) {
903 obj = gem_allocate_guc_obj(dev_priv, PAGE_ALIGN(size));
904 if (!obj)
905 return;
906
907 guc->ads_obj = obj;
908 }
909
910 page = i915_gem_object_get_page(obj, 0);
911 ads = kmap(page);
912
913 /*
914 * The GuC requires a "Golden Context" when it reinitialises
915 * engines after a reset. Here we use the Render ring default
916 * context, which must already exist and be pinned in the GGTT,
917 * so its address won't change after we've told the GuC where
918 * to find it.
919 */
920 engine = &dev_priv->engine[RCS];
921 ads->golden_context_lrca = engine->status_page.gfx_addr;
922
923 for_each_engine(engine, dev_priv)
924 ads->eng_state_size[engine->guc_id] = intel_lr_context_size(engine);
925
926 /* GuC scheduling policies */
927 policies = (void *)ads + sizeof(struct guc_ads);
928 init_guc_policies(policies);
929
930 ads->scheduler_policies = i915_gem_obj_ggtt_offset(obj) +
931 sizeof(struct guc_ads);
932
933 /* MMIO reg state */
934 reg_state = (void *)policies + sizeof(struct guc_policies);
935
936 for_each_engine(engine, dev_priv) {
937 reg_state->mmio_white_list[engine->guc_id].mmio_start =
938 engine->mmio_base + GUC_MMIO_WHITE_LIST_START;
939
940 /* Nothing to be saved or restored for now. */
941 reg_state->mmio_white_list[engine->guc_id].count = 0;
942 }
943
944 ads->reg_state_addr = ads->scheduler_policies +
945 sizeof(struct guc_policies);
946
947 ads->reg_state_buffer = ads->reg_state_addr +
948 sizeof(struct guc_mmio_reg_state);
949
950 kunmap(page);
951 }
952
953 /*
954 * Set up the memory resources to be shared with the GuC. At this point,
955 * we require just one object that can be mapped through the GGTT.
956 */
957 int i915_guc_submission_init(struct drm_i915_private *dev_priv)
958 {
959 const size_t ctxsize = sizeof(struct guc_context_desc);
960 const size_t poolsize = GUC_MAX_GPU_CONTEXTS * ctxsize;
961 const size_t gemsize = round_up(poolsize, PAGE_SIZE);
962 struct intel_guc *guc = &dev_priv->guc;
963
964 /* Wipe bitmap & delete client in case of reinitialisation */
965 bitmap_clear(guc->doorbell_bitmap, 0, GUC_MAX_DOORBELLS);
966 i915_guc_submission_disable(dev_priv);
967
968 if (!i915.enable_guc_submission)
969 return 0; /* not enabled */
970
971 if (guc->ctx_pool_obj)
972 return 0; /* already allocated */
973
974 guc->ctx_pool_obj = gem_allocate_guc_obj(dev_priv, gemsize);
975 if (!guc->ctx_pool_obj)
976 return -ENOMEM;
977
978 ida_init(&guc->ctx_ids);
979 guc_create_log(guc);
980 guc_create_ads(guc);
981
982 return 0;
983 }
984
985 int i915_guc_submission_enable(struct drm_i915_private *dev_priv)
986 {
987 struct intel_guc *guc = &dev_priv->guc;
988 struct i915_guc_client *client;
989
990 /* client for execbuf submission */
991 client = guc_client_alloc(dev_priv,
992 GUC_CTX_PRIORITY_KMD_NORMAL,
993 dev_priv->kernel_context);
994 if (!client) {
995 DRM_ERROR("Failed to create execbuf guc_client\n");
996 return -ENOMEM;
997 }
998
999 guc->execbuf_client = client;
1000 host2guc_sample_forcewake(guc, client);
1001 guc_init_doorbell_hw(guc);
1002
1003 return 0;
1004 }
1005
1006 void i915_guc_submission_disable(struct drm_i915_private *dev_priv)
1007 {
1008 struct intel_guc *guc = &dev_priv->guc;
1009
1010 guc_client_free(dev_priv, guc->execbuf_client);
1011 guc->execbuf_client = NULL;
1012 }
1013
1014 void i915_guc_submission_fini(struct drm_i915_private *dev_priv)
1015 {
1016 struct intel_guc *guc = &dev_priv->guc;
1017
1018 gem_release_guc_obj(dev_priv->guc.ads_obj);
1019 guc->ads_obj = NULL;
1020
1021 gem_release_guc_obj(dev_priv->guc.log_obj);
1022 guc->log_obj = NULL;
1023
1024 if (guc->ctx_pool_obj)
1025 ida_destroy(&guc->ctx_ids);
1026 gem_release_guc_obj(guc->ctx_pool_obj);
1027 guc->ctx_pool_obj = NULL;
1028 }
1029
1030 /**
1031 * intel_guc_suspend() - notify GuC entering suspend state
1032 * @dev: drm device
1033 */
1034 int intel_guc_suspend(struct drm_device *dev)
1035 {
1036 struct drm_i915_private *dev_priv = to_i915(dev);
1037 struct intel_guc *guc = &dev_priv->guc;
1038 struct i915_gem_context *ctx;
1039 u32 data[3];
1040
1041 if (guc->guc_fw.guc_fw_load_status != GUC_FIRMWARE_SUCCESS)
1042 return 0;
1043
1044 ctx = dev_priv->kernel_context;
1045
1046 data[0] = HOST2GUC_ACTION_ENTER_S_STATE;
1047 /* any value greater than GUC_POWER_D0 */
1048 data[1] = GUC_POWER_D1;
1049 /* first page is shared data with GuC */
1050 data[2] = i915_gem_obj_ggtt_offset(ctx->engine[RCS].state);
1051
1052 return host2guc_action(guc, data, ARRAY_SIZE(data));
1053 }
1054
1055
1056 /**
1057 * intel_guc_resume() - notify GuC resuming from suspend state
1058 * @dev: drm device
1059 */
1060 int intel_guc_resume(struct drm_device *dev)
1061 {
1062 struct drm_i915_private *dev_priv = to_i915(dev);
1063 struct intel_guc *guc = &dev_priv->guc;
1064 struct i915_gem_context *ctx;
1065 u32 data[3];
1066
1067 if (guc->guc_fw.guc_fw_load_status != GUC_FIRMWARE_SUCCESS)
1068 return 0;
1069
1070 ctx = dev_priv->kernel_context;
1071
1072 data[0] = HOST2GUC_ACTION_EXIT_S_STATE;
1073 data[1] = GUC_POWER_D0;
1074 /* first page is shared data with GuC */
1075 data[2] = i915_gem_obj_ggtt_offset(ctx->engine[RCS].state);
1076
1077 return host2guc_action(guc, data, ARRAY_SIZE(data));
1078 }
This page took 0.052686 seconds and 6 git commands to generate.