Branch data Line data Source code
1 : : // SPDX-License-Identifier: MIT
2 : : /*
3 : : * Copyright © 2014 Intel Corporation
4 : : */
5 : :
6 : : #include <linux/circ_buf.h>
7 : :
8 : : #include "gem/i915_gem_context.h"
9 : : #include "gt/intel_context.h"
10 : : #include "gt/intel_engine_pm.h"
11 : : #include "gt/intel_gt.h"
12 : : #include "gt/intel_gt_pm.h"
13 : : #include "gt/intel_lrc_reg.h"
14 : : #include "gt/intel_ring.h"
15 : :
16 : : #include "intel_guc_submission.h"
17 : :
18 : : #include "i915_drv.h"
19 : : #include "i915_trace.h"
20 : :
21 : : /**
22 : : * DOC: GuC-based command submission
23 : : *
24 : : * IMPORTANT NOTE: GuC submission is currently not supported in i915. The GuC
25 : : * firmware is moving to an updated submission interface and we plan to
26 : : * turn submission back on when that lands. The below documentation (and related
27 : : * code) matches the old submission model and will be updated as part of the
28 : : * upgrade to the new flow.
29 : : *
30 : : * GuC stage descriptor:
31 : : * During initialization, the driver allocates a static pool of 1024 such
32 : : * descriptors, and shares them with the GuC. Currently, we only use one
33 : : * descriptor. This stage descriptor lets the GuC know about the workqueue and
34 : : * process descriptor. Theoretically, it also lets the GuC know about our HW
35 : : * contexts (context ID, etc...), but we actually employ a kind of submission
36 : : * where the GuC uses the LRCA sent via the work item instead. This is called
37 : : * a "proxy" submission.
38 : : *
39 : : * The Scratch registers:
40 : : * There are 16 MMIO-based registers start from 0xC180. The kernel driver writes
41 : : * a value to the action register (SOFT_SCRATCH_0) along with any data. It then
42 : : * triggers an interrupt on the GuC via another register write (0xC4C8).
43 : : * Firmware writes a success/fail code back to the action register after
44 : : * processes the request. The kernel driver polls waiting for this update and
45 : : * then proceeds.
46 : : *
47 : : * Work Items:
48 : : * There are several types of work items that the host may place into a
49 : : * workqueue, each with its own requirements and limitations. Currently only
50 : : * WQ_TYPE_INORDER is needed to support legacy submission via GuC, which
51 : : * represents in-order queue. The kernel driver packs ring tail pointer and an
52 : : * ELSP context descriptor dword into Work Item.
53 : : * See guc_add_request()
54 : : *
55 : : */
56 : :
57 : 0 : static inline struct i915_priolist *to_priolist(struct rb_node *rb)
58 : : {
59 : 0 : return rb_entry(rb, struct i915_priolist, node);
60 : : }
61 : :
62 : 0 : static struct guc_stage_desc *__get_stage_desc(struct intel_guc *guc, u32 id)
63 : : {
64 : 0 : struct guc_stage_desc *base = guc->stage_desc_pool_vaddr;
65 : :
66 : 0 : return &base[id];
67 : : }
68 : :
69 : 0 : static int guc_workqueue_create(struct intel_guc *guc)
70 : : {
71 : 0 : return intel_guc_allocate_and_map_vma(guc, GUC_WQ_SIZE, &guc->workqueue,
72 : : &guc->workqueue_vaddr);
73 : : }
74 : :
75 : 0 : static void guc_workqueue_destroy(struct intel_guc *guc)
76 : : {
77 : 0 : i915_vma_unpin_and_release(&guc->workqueue, I915_VMA_RELEASE_MAP);
78 : 0 : }
79 : :
80 : : /*
81 : : * Initialise the process descriptor shared with the GuC firmware.
82 : : */
83 : 0 : static int guc_proc_desc_create(struct intel_guc *guc)
84 : : {
85 : 0 : const u32 size = PAGE_ALIGN(sizeof(struct guc_process_desc));
86 : :
87 : 0 : return intel_guc_allocate_and_map_vma(guc, size, &guc->proc_desc,
88 : : &guc->proc_desc_vaddr);
89 : : }
90 : :
91 : 0 : static void guc_proc_desc_destroy(struct intel_guc *guc)
92 : : {
93 : 0 : i915_vma_unpin_and_release(&guc->proc_desc, I915_VMA_RELEASE_MAP);
94 : : }
95 : :
96 : : static void guc_proc_desc_init(struct intel_guc *guc)
97 : : {
98 : : struct guc_process_desc *desc;
99 : :
100 : : desc = memset(guc->proc_desc_vaddr, 0, sizeof(*desc));
101 : :
102 : : /*
103 : : * XXX: pDoorbell and WQVBaseAddress are pointers in process address
104 : : * space for ring3 clients (set them as in mmap_ioctl) or kernel
105 : : * space for kernel clients (map on demand instead? May make debug
106 : : * easier to have it mapped).
107 : : */
108 : : desc->wq_base_addr = 0;
109 : : desc->db_base_addr = 0;
110 : :
111 : : desc->wq_size_bytes = GUC_WQ_SIZE;
112 : : desc->wq_status = WQ_STATUS_ACTIVE;
113 : : desc->priority = GUC_CLIENT_PRIORITY_KMD_NORMAL;
114 : : }
115 : :
116 : 0 : static void guc_proc_desc_fini(struct intel_guc *guc)
117 : : {
118 : 0 : memset(guc->proc_desc_vaddr, 0, sizeof(struct guc_process_desc));
119 : : }
120 : :
121 : 0 : static int guc_stage_desc_pool_create(struct intel_guc *guc)
122 : : {
123 : 0 : u32 size = PAGE_ALIGN(sizeof(struct guc_stage_desc) *
124 : : GUC_MAX_STAGE_DESCRIPTORS);
125 : :
126 : 0 : return intel_guc_allocate_and_map_vma(guc, size, &guc->stage_desc_pool,
127 : : &guc->stage_desc_pool_vaddr);
128 : : }
129 : :
130 : 0 : static void guc_stage_desc_pool_destroy(struct intel_guc *guc)
131 : : {
132 : 0 : i915_vma_unpin_and_release(&guc->stage_desc_pool, I915_VMA_RELEASE_MAP);
133 : 0 : }
134 : :
135 : : /*
136 : : * Initialise/clear the stage descriptor shared with the GuC firmware.
137 : : *
138 : : * This descriptor tells the GuC where (in GGTT space) to find the important
139 : : * data structures related to work submission (process descriptor, write queue,
140 : : * etc).
141 : : */
142 : 0 : static void guc_stage_desc_init(struct intel_guc *guc)
143 : : {
144 : 0 : struct guc_stage_desc *desc;
145 : :
146 : : /* we only use 1 stage desc, so hardcode it to 0 */
147 : 0 : desc = __get_stage_desc(guc, 0);
148 : 0 : memset(desc, 0, sizeof(*desc));
149 : :
150 : 0 : desc->attribute = GUC_STAGE_DESC_ATTR_ACTIVE |
151 : : GUC_STAGE_DESC_ATTR_KERNEL;
152 : :
153 : 0 : desc->stage_id = 0;
154 : 0 : desc->priority = GUC_CLIENT_PRIORITY_KMD_NORMAL;
155 : :
156 : 0 : desc->process_desc = intel_guc_ggtt_offset(guc, guc->proc_desc);
157 : 0 : desc->wq_addr = intel_guc_ggtt_offset(guc, guc->workqueue);
158 : 0 : desc->wq_size = GUC_WQ_SIZE;
159 : 0 : }
160 : :
161 : 0 : static void guc_stage_desc_fini(struct intel_guc *guc)
162 : : {
163 : 0 : struct guc_stage_desc *desc;
164 : :
165 : 0 : desc = __get_stage_desc(guc, 0);
166 : 0 : memset(desc, 0, sizeof(*desc));
167 : : }
168 : :
169 : : /* Construct a Work Item and append it to the GuC's Work Queue */
170 : 0 : static void guc_wq_item_append(struct intel_guc *guc,
171 : : u32 target_engine, u32 context_desc,
172 : : u32 ring_tail, u32 fence_id)
173 : : {
174 : : /* wqi_len is in DWords, and does not include the one-word header */
175 : 0 : const size_t wqi_size = sizeof(struct guc_wq_item);
176 : 0 : const u32 wqi_len = wqi_size / sizeof(u32) - 1;
177 : 0 : struct guc_process_desc *desc = guc->proc_desc_vaddr;
178 : 0 : struct guc_wq_item *wqi;
179 : 0 : u32 wq_off;
180 : :
181 : 0 : lockdep_assert_held(&guc->wq_lock);
182 : :
183 : : /* For now workqueue item is 4 DWs; workqueue buffer is 2 pages. So we
184 : : * should not have the case where structure wqi is across page, neither
185 : : * wrapped to the beginning. This simplifies the implementation below.
186 : : *
187 : : * XXX: if not the case, we need save data to a temp wqi and copy it to
188 : : * workqueue buffer dw by dw.
189 : : */
190 : 0 : BUILD_BUG_ON(wqi_size != 16);
191 : :
192 : : /* We expect the WQ to be active if we're appending items to it */
193 : 0 : GEM_BUG_ON(desc->wq_status != WQ_STATUS_ACTIVE);
194 : :
195 : : /* Free space is guaranteed. */
196 : 0 : wq_off = READ_ONCE(desc->tail);
197 : 0 : GEM_BUG_ON(CIRC_SPACE(wq_off, READ_ONCE(desc->head),
198 : : GUC_WQ_SIZE) < wqi_size);
199 : 0 : GEM_BUG_ON(wq_off & (wqi_size - 1));
200 : :
201 : 0 : wqi = guc->workqueue_vaddr + wq_off;
202 : :
203 : : /* Now fill in the 4-word work queue item */
204 : 0 : wqi->header = WQ_TYPE_INORDER |
205 : : (wqi_len << WQ_LEN_SHIFT) |
206 : 0 : (target_engine << WQ_TARGET_SHIFT) |
207 : : WQ_NO_WCFLUSH_WAIT;
208 : 0 : wqi->context_desc = context_desc;
209 : 0 : wqi->submit_element_info = ring_tail << WQ_RING_TAIL_SHIFT;
210 : 0 : GEM_BUG_ON(ring_tail > WQ_RING_TAIL_MAX);
211 : 0 : wqi->fence_id = fence_id;
212 : :
213 : : /* Make the update visible to GuC */
214 : 0 : WRITE_ONCE(desc->tail, (wq_off + wqi_size) & (GUC_WQ_SIZE - 1));
215 : : }
216 : :
217 : 0 : static void guc_add_request(struct intel_guc *guc, struct i915_request *rq)
218 : : {
219 : 0 : struct intel_engine_cs *engine = rq->engine;
220 : 0 : u32 ctx_desc = lower_32_bits(rq->context->lrc_desc);
221 : 0 : u32 ring_tail = intel_ring_set_tail(rq->ring, rq->tail) / sizeof(u64);
222 : :
223 : 0 : guc_wq_item_append(guc, engine->guc_id, ctx_desc,
224 : 0 : ring_tail, rq->fence.seqno);
225 : : }
226 : :
227 : : /*
228 : : * When we're doing submissions using regular execlists backend, writing to
229 : : * ELSP from CPU side is enough to make sure that writes to ringbuffer pages
230 : : * pinned in mappable aperture portion of GGTT are visible to command streamer.
231 : : * Writes done by GuC on our behalf are not guaranteeing such ordering,
232 : : * therefore, to ensure the flush, we're issuing a POSTING READ.
233 : : */
234 : 0 : static void flush_ggtt_writes(struct i915_vma *vma)
235 : : {
236 [ # # ]: 0 : if (i915_vma_is_map_and_fenceable(vma))
237 : 0 : intel_uncore_posting_read_fw(vma->vm->gt->uncore,
238 : : GUC_STATUS);
239 : 0 : }
240 : :
241 : : static void guc_submit(struct intel_engine_cs *engine,
242 : : struct i915_request **out,
243 : : struct i915_request **end)
244 : : {
245 : : struct intel_guc *guc = &engine->gt->uc.guc;
246 : :
247 : : spin_lock(&guc->wq_lock);
248 : :
249 : : do {
250 : : struct i915_request *rq = *out++;
251 : :
252 : : flush_ggtt_writes(rq->ring->vma);
253 : : guc_add_request(guc, rq);
254 : : } while (out != end);
255 : :
256 : : spin_unlock(&guc->wq_lock);
257 : : }
258 : :
259 : : static inline int rq_prio(const struct i915_request *rq)
260 : : {
261 : : return rq->sched.attr.priority | __NO_PREEMPTION;
262 : : }
263 : :
264 : : static struct i915_request *schedule_in(struct i915_request *rq, int idx)
265 : : {
266 : : trace_i915_request_in(rq, idx);
267 : :
268 : : /*
269 : : * Currently we are not tracking the rq->context being inflight
270 : : * (ce->inflight = rq->engine). It is only used by the execlists
271 : : * backend at the moment, a similar counting strategy would be
272 : : * required if we generalise the inflight tracking.
273 : : */
274 : :
275 : : __intel_gt_pm_get(rq->engine->gt);
276 : : return i915_request_get(rq);
277 : : }
278 : :
279 : 0 : static void schedule_out(struct i915_request *rq)
280 : : {
281 : 0 : trace_i915_request_out(rq);
282 : :
283 : 0 : intel_gt_pm_put_async(rq->engine->gt);
284 : 0 : i915_request_put(rq);
285 : 0 : }
286 : :
287 : 0 : static void __guc_dequeue(struct intel_engine_cs *engine)
288 : : {
289 : 0 : struct intel_engine_execlists * const execlists = &engine->execlists;
290 : 0 : struct i915_request **first = execlists->inflight;
291 : 0 : struct i915_request ** const last_port = first + execlists->port_mask;
292 : 0 : struct i915_request *last = first[0];
293 : 0 : struct i915_request **port;
294 : 0 : bool submit = false;
295 : 0 : struct rb_node *rb;
296 : :
297 : 0 : lockdep_assert_held(&engine->active.lock);
298 : :
299 [ # # ]: 0 : if (last) {
300 [ # # ]: 0 : if (*++first)
301 : : return;
302 : :
303 : : last = NULL;
304 : : }
305 : :
306 : : /*
307 : : * We write directly into the execlists->inflight queue and don't use
308 : : * the execlists->pending queue, as we don't have a distinct switch
309 : : * event.
310 : : */
311 : 0 : port = first;
312 [ # # ]: 0 : while ((rb = rb_first_cached(&execlists->queue))) {
313 : 0 : struct i915_priolist *p = to_priolist(rb);
314 : 0 : struct i915_request *rq, *rn;
315 : 0 : int i;
316 : :
317 [ # # # # ]: 0 : priolist_for_each_request_consume(rq, rn, p, i) {
318 [ # # # # ]: 0 : if (last && rq->context != last->context) {
319 [ # # ]: 0 : if (port == last_port)
320 : 0 : goto done;
321 : :
322 : 0 : *port = schedule_in(last,
323 : : port - execlists->inflight);
324 : 0 : port++;
325 : : }
326 : :
327 : 0 : list_del_init(&rq->sched.link);
328 : 0 : __i915_request_submit(rq);
329 : 0 : submit = true;
330 : 0 : last = rq;
331 : : }
332 : :
333 : 0 : rb_erase_cached(&p->node, &execlists->queue);
334 [ # # ]: 0 : i915_priolist_free(p);
335 : : }
336 : 0 : done:
337 : 0 : execlists->queue_priority_hint =
338 [ # # ]: 0 : rb ? to_priolist(rb)->priority : INT_MIN;
339 [ # # ]: 0 : if (submit) {
340 : 0 : *port = schedule_in(last, port - execlists->inflight);
341 : 0 : *++port = NULL;
342 : 0 : guc_submit(engine, first, port);
343 : : }
344 : 0 : execlists->active = execlists->inflight;
345 : : }
346 : :
347 : 0 : static void guc_submission_tasklet(unsigned long data)
348 : : {
349 : 0 : struct intel_engine_cs * const engine = (struct intel_engine_cs *)data;
350 : 0 : struct intel_engine_execlists * const execlists = &engine->execlists;
351 : 0 : struct i915_request **port, *rq;
352 : 0 : unsigned long flags;
353 : :
354 : 0 : spin_lock_irqsave(&engine->active.lock, flags);
355 : :
356 [ # # ]: 0 : for (port = execlists->inflight; (rq = *port); port++) {
357 [ # # ]: 0 : if (!i915_request_completed(rq))
358 : : break;
359 : :
360 : 0 : schedule_out(rq);
361 : : }
362 [ # # ]: 0 : if (port != execlists->inflight) {
363 : 0 : int idx = port - execlists->inflight;
364 : 0 : int rem = ARRAY_SIZE(execlists->inflight) - idx;
365 : 0 : memmove(execlists->inflight, port, rem * sizeof(*port));
366 : : }
367 : :
368 : 0 : __guc_dequeue(engine);
369 : :
370 : 0 : spin_unlock_irqrestore(&engine->active.lock, flags);
371 : 0 : }
372 : :
373 : 0 : static void guc_reset_prepare(struct intel_engine_cs *engine)
374 : : {
375 : 0 : struct intel_engine_execlists * const execlists = &engine->execlists;
376 : :
377 : 0 : ENGINE_TRACE(engine, "\n");
378 : :
379 : : /*
380 : : * Prevent request submission to the hardware until we have
381 : : * completed the reset in i915_gem_reset_finish(). If a request
382 : : * is completed by one engine, it may then queue a request
383 : : * to a second via its execlists->tasklet *just* as we are
384 : : * calling engine->init_hw() and also writing the ELSP.
385 : : * Turning off the execlists->tasklet until the reset is over
386 : : * prevents the race.
387 : : */
388 : 0 : __tasklet_disable_sync_once(&execlists->tasklet);
389 : 0 : }
390 : :
391 : : static void
392 : 0 : cancel_port_requests(struct intel_engine_execlists * const execlists)
393 : : {
394 : 0 : struct i915_request * const *port, *rq;
395 : :
396 : : /* Note we are only using the inflight and not the pending queue */
397 : :
398 [ # # ]: 0 : for (port = execlists->active; (rq = *port); port++)
399 : 0 : schedule_out(rq);
400 : 0 : execlists->active =
401 : 0 : memset(execlists->inflight, 0, sizeof(execlists->inflight));
402 : 0 : }
403 : :
404 : 0 : static void guc_reset_rewind(struct intel_engine_cs *engine, bool stalled)
405 : : {
406 : 0 : struct intel_engine_execlists * const execlists = &engine->execlists;
407 : 0 : struct i915_request *rq;
408 : 0 : unsigned long flags;
409 : :
410 : 0 : spin_lock_irqsave(&engine->active.lock, flags);
411 : :
412 : 0 : cancel_port_requests(execlists);
413 : :
414 : : /* Push back any incomplete requests for replay after the reset. */
415 : 0 : rq = execlists_unwind_incomplete_requests(execlists);
416 [ # # ]: 0 : if (!rq)
417 : 0 : goto out_unlock;
418 : :
419 [ # # ]: 0 : if (!i915_request_started(rq))
420 : 0 : stalled = false;
421 : :
422 : 0 : __i915_request_reset(rq, stalled);
423 : 0 : intel_lr_context_reset(engine, rq->context, rq->head, stalled);
424 : :
425 : 0 : out_unlock:
426 : 0 : spin_unlock_irqrestore(&engine->active.lock, flags);
427 : 0 : }
428 : :
429 : 0 : static void guc_reset_cancel(struct intel_engine_cs *engine)
430 : : {
431 : 0 : struct intel_engine_execlists * const execlists = &engine->execlists;
432 : 0 : struct i915_request *rq, *rn;
433 : 0 : struct rb_node *rb;
434 : 0 : unsigned long flags;
435 : :
436 : 0 : ENGINE_TRACE(engine, "\n");
437 : :
438 : : /*
439 : : * Before we call engine->cancel_requests(), we should have exclusive
440 : : * access to the submission state. This is arranged for us by the
441 : : * caller disabling the interrupt generation, the tasklet and other
442 : : * threads that may then access the same state, giving us a free hand
443 : : * to reset state. However, we still need to let lockdep be aware that
444 : : * we know this state may be accessed in hardirq context, so we
445 : : * disable the irq around this manipulation and we want to keep
446 : : * the spinlock focused on its duties and not accidentally conflate
447 : : * coverage to the submission's irq state. (Similarly, although we
448 : : * shouldn't need to disable irq around the manipulation of the
449 : : * submission's irq state, we also wish to remind ourselves that
450 : : * it is irq state.)
451 : : */
452 : 0 : spin_lock_irqsave(&engine->active.lock, flags);
453 : :
454 : : /* Cancel the requests on the HW and clear the ELSP tracker. */
455 : 0 : cancel_port_requests(execlists);
456 : :
457 : : /* Mark all executing requests as skipped. */
458 [ # # ]: 0 : list_for_each_entry(rq, &engine->active.requests, sched.link) {
459 [ # # ]: 0 : if (!i915_request_signaled(rq))
460 : 0 : dma_fence_set_error(&rq->fence, -EIO);
461 : :
462 : 0 : i915_request_mark_complete(rq);
463 : : }
464 : :
465 : : /* Flush the queued requests to the timeline list (for retiring). */
466 [ # # ]: 0 : while ((rb = rb_first_cached(&execlists->queue))) {
467 : 0 : struct i915_priolist *p = to_priolist(rb);
468 : 0 : int i;
469 : :
470 [ # # # # ]: 0 : priolist_for_each_request_consume(rq, rn, p, i) {
471 : 0 : list_del_init(&rq->sched.link);
472 : 0 : __i915_request_submit(rq);
473 : 0 : dma_fence_set_error(&rq->fence, -EIO);
474 : 0 : i915_request_mark_complete(rq);
475 : : }
476 : :
477 : 0 : rb_erase_cached(&p->node, &execlists->queue);
478 [ # # ]: 0 : i915_priolist_free(p);
479 : : }
480 : :
481 : : /* Remaining _unready_ requests will be nop'ed when submitted */
482 : :
483 : 0 : execlists->queue_priority_hint = INT_MIN;
484 : 0 : execlists->queue = RB_ROOT_CACHED;
485 : :
486 : 0 : spin_unlock_irqrestore(&engine->active.lock, flags);
487 : 0 : }
488 : :
489 : 0 : static void guc_reset_finish(struct intel_engine_cs *engine)
490 : : {
491 : 0 : struct intel_engine_execlists * const execlists = &engine->execlists;
492 : :
493 [ # # ]: 0 : if (__tasklet_enable(&execlists->tasklet))
494 : : /* And kick in case we missed a new request submission. */
495 : 0 : tasklet_hi_schedule(&execlists->tasklet);
496 : :
497 : 0 : ENGINE_TRACE(engine, "depth->%d\n",
498 : : atomic_read(&execlists->tasklet.count));
499 : 0 : }
500 : :
501 : : /*
502 : : * Everything below here is concerned with setup & teardown, and is
503 : : * therefore not part of the somewhat time-critical batch-submission
504 : : * path of guc_submit() above.
505 : : */
506 : :
507 : : /*
508 : : * Set up the memory resources to be shared with the GuC (via the GGTT)
509 : : * at firmware loading time.
510 : : */
511 : 0 : int intel_guc_submission_init(struct intel_guc *guc)
512 : : {
513 : 0 : int ret;
514 : :
515 [ # # ]: 0 : if (guc->stage_desc_pool)
516 : : return 0;
517 : :
518 : 0 : ret = guc_stage_desc_pool_create(guc);
519 [ # # ]: 0 : if (ret)
520 : : return ret;
521 : : /*
522 : : * Keep static analysers happy, let them know that we allocated the
523 : : * vma after testing that it didn't exist earlier.
524 : : */
525 : 0 : GEM_BUG_ON(!guc->stage_desc_pool);
526 : :
527 : 0 : ret = guc_workqueue_create(guc);
528 [ # # ]: 0 : if (ret)
529 : 0 : goto err_pool;
530 : :
531 : 0 : ret = guc_proc_desc_create(guc);
532 [ # # ]: 0 : if (ret)
533 : 0 : goto err_workqueue;
534 : :
535 : 0 : spin_lock_init(&guc->wq_lock);
536 : :
537 : 0 : return 0;
538 : :
539 : : err_workqueue:
540 : 0 : guc_workqueue_destroy(guc);
541 : 0 : err_pool:
542 : 0 : guc_stage_desc_pool_destroy(guc);
543 : 0 : return ret;
544 : : }
545 : :
546 : 0 : void intel_guc_submission_fini(struct intel_guc *guc)
547 : : {
548 [ # # ]: 0 : if (guc->stage_desc_pool) {
549 : 0 : guc_proc_desc_destroy(guc);
550 : 0 : guc_workqueue_destroy(guc);
551 : 0 : guc_stage_desc_pool_destroy(guc);
552 : : }
553 : 0 : }
554 : :
555 : : static void guc_interrupts_capture(struct intel_gt *gt)
556 : : {
557 : : struct intel_uncore *uncore = gt->uncore;
558 : : u32 irqs = GT_CONTEXT_SWITCH_INTERRUPT;
559 : : u32 dmask = irqs << 16 | irqs;
560 : :
561 : : GEM_BUG_ON(INTEL_GEN(gt->i915) < 11);
562 : :
563 : : /* Don't handle the ctx switch interrupt in GuC submission mode */
564 : : intel_uncore_rmw(uncore, GEN11_RENDER_COPY_INTR_ENABLE, dmask, 0);
565 : : intel_uncore_rmw(uncore, GEN11_VCS_VECS_INTR_ENABLE, dmask, 0);
566 : : }
567 : :
568 : : static void guc_interrupts_release(struct intel_gt *gt)
569 : : {
570 : : struct intel_uncore *uncore = gt->uncore;
571 : : u32 irqs = GT_CONTEXT_SWITCH_INTERRUPT;
572 : : u32 dmask = irqs << 16 | irqs;
573 : :
574 : : GEM_BUG_ON(INTEL_GEN(gt->i915) < 11);
575 : :
576 : : /* Handle ctx switch interrupts again */
577 : : intel_uncore_rmw(uncore, GEN11_RENDER_COPY_INTR_ENABLE, 0, dmask);
578 : : intel_uncore_rmw(uncore, GEN11_VCS_VECS_INTR_ENABLE, 0, dmask);
579 : : }
580 : :
581 : 0 : static void guc_set_default_submission(struct intel_engine_cs *engine)
582 : : {
583 : : /*
584 : : * We inherit a bunch of functions from execlists that we'd like
585 : : * to keep using:
586 : : *
587 : : * engine->submit_request = execlists_submit_request;
588 : : * engine->cancel_requests = execlists_cancel_requests;
589 : : * engine->schedule = execlists_schedule;
590 : : *
591 : : * But we need to override the actual submission backend in order
592 : : * to talk to the GuC.
593 : : */
594 : 0 : intel_execlists_set_default_submission(engine);
595 : :
596 : 0 : engine->execlists.tasklet.func = guc_submission_tasklet;
597 : :
598 : : /* do not use execlists park/unpark */
599 : 0 : engine->park = engine->unpark = NULL;
600 : :
601 : 0 : engine->reset.prepare = guc_reset_prepare;
602 : 0 : engine->reset.rewind = guc_reset_rewind;
603 : 0 : engine->reset.cancel = guc_reset_cancel;
604 : 0 : engine->reset.finish = guc_reset_finish;
605 : :
606 : 0 : engine->flags &= ~I915_ENGINE_SUPPORTS_STATS;
607 : 0 : engine->flags |= I915_ENGINE_NEEDS_BREADCRUMB_TASKLET;
608 : :
609 : : /*
610 : : * For the breadcrumb irq to work we need the interrupts to stay
611 : : * enabled. However, on all platforms on which we'll have support for
612 : : * GuC submission we don't allow disabling the interrupts at runtime, so
613 : : * we're always safe with the current flow.
614 : : */
615 : 0 : GEM_BUG_ON(engine->irq_enable || engine->irq_disable);
616 : 0 : }
617 : :
618 : 0 : void intel_guc_submission_enable(struct intel_guc *guc)
619 : : {
620 : 0 : struct intel_gt *gt = guc_to_gt(guc);
621 : 0 : struct intel_engine_cs *engine;
622 : 0 : enum intel_engine_id id;
623 : :
624 : : /*
625 : : * We're using GuC work items for submitting work through GuC. Since
626 : : * we're coalescing multiple requests from a single context into a
627 : : * single work item prior to assigning it to execlist_port, we can
628 : : * never have more work items than the total number of ports (for all
629 : : * engines). The GuC firmware is controlling the HEAD of work queue,
630 : : * and it is guaranteed that it will remove the work item from the
631 : : * queue before our request is completed.
632 : : */
633 : 0 : BUILD_BUG_ON(ARRAY_SIZE(engine->execlists.inflight) *
634 : : sizeof(struct guc_wq_item) *
635 : : I915_NUM_ENGINES > GUC_WQ_SIZE);
636 : :
637 : 0 : guc_proc_desc_init(guc);
638 : 0 : guc_stage_desc_init(guc);
639 : :
640 : : /* Take over from manual control of ELSP (execlists) */
641 : 0 : guc_interrupts_capture(gt);
642 : :
643 [ # # # # ]: 0 : for_each_engine(engine, gt, id) {
644 : 0 : engine->set_default_submission = guc_set_default_submission;
645 : 0 : engine->set_default_submission(engine);
646 : : }
647 : 0 : }
648 : :
649 : 0 : void intel_guc_submission_disable(struct intel_guc *guc)
650 : : {
651 : 0 : struct intel_gt *gt = guc_to_gt(guc);
652 : :
653 : 0 : GEM_BUG_ON(gt->awake); /* GT should be parked first */
654 : :
655 : : /* Note: By the time we're here, GuC may have already been reset */
656 : :
657 : 0 : guc_interrupts_release(gt);
658 : :
659 : 0 : guc_stage_desc_fini(guc);
660 : 0 : guc_proc_desc_fini(guc);
661 : 0 : }
662 : :
663 : 0 : static bool __guc_submission_support(struct intel_guc *guc)
664 : : {
665 : : /* XXX: GuC submission is unavailable for now */
666 : 0 : return false;
667 : :
668 : : if (!intel_guc_is_supported(guc))
669 : : return false;
670 : :
671 : : return i915_modparams.enable_guc & ENABLE_GUC_SUBMISSION;
672 : : }
673 : :
674 : 0 : void intel_guc_submission_init_early(struct intel_guc *guc)
675 : : {
676 : 0 : guc->submission_supported = __guc_submission_support(guc);
677 : 0 : }
678 : :
679 : 0 : bool intel_engine_in_guc_submission_mode(const struct intel_engine_cs *engine)
680 : : {
681 : 0 : return engine->set_default_submission == guc_set_default_submission;
682 : : }
|