2 * Copyright © 2014 Intel Corporation
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:
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
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
25 #include <linux/circ_buf.h>
26 #include <trace/events/dma_fence.h>
28 #include "intel_guc_submission.h"
29 #include "intel_lrc_reg.h"
32 #define GUC_PREEMPT_FINISHED 0x1
33 #define GUC_PREEMPT_BREADCRUMB_DWORDS 0x8
34 #define GUC_PREEMPT_BREADCRUMB_BYTES \
35 (sizeof(u32) * GUC_PREEMPT_BREADCRUMB_DWORDS)
38 * DOC: GuC-based command submission
41 * A intel_guc_client refers to a submission path through GuC. Currently, there
42 * are two clients. One of them (the execbuf_client) is charged with all
43 * submissions to the GuC, the other one (preempt_client) is responsible for
44 * preempting the execbuf_client. This struct is the owner of a doorbell, a
45 * process descriptor and a workqueue (all of them inside a single gem object
46 * that contains all required pages for these elements).
48 * GuC stage descriptor:
49 * During initialization, the driver allocates a static pool of 1024 such
50 * descriptors, and shares them with the GuC.
51 * Currently, there exists a 1:1 mapping between a intel_guc_client and a
52 * guc_stage_desc (via the client's stage_id), so effectively only one
53 * gets used. This stage descriptor lets the GuC know about the doorbell,
54 * workqueue and process descriptor. Theoretically, it also lets the GuC
55 * know about our HW contexts (context ID, etc...), but we actually
56 * employ a kind of submission where the GuC uses the LRCA sent via the work
57 * item instead (the single guc_stage_desc associated to execbuf client
58 * contains information about the default kernel context only, but this is
59 * essentially unused). This is called a "proxy" submission.
61 * The Scratch registers:
62 * There are 16 MMIO-based registers start from 0xC180. The kernel driver writes
63 * a value to the action register (SOFT_SCRATCH_0) along with any data. It then
64 * triggers an interrupt on the GuC via another register write (0xC4C8).
65 * Firmware writes a success/fail code back to the action register after
66 * processes the request. The kernel driver polls waiting for this update and
68 * See intel_guc_send()
71 * Doorbells are interrupts to uKernel. A doorbell is a single cache line (QW)
72 * mapped into process space.
75 * There are several types of work items that the host may place into a
76 * workqueue, each with its own requirements and limitations. Currently only
77 * WQ_TYPE_INORDER is needed to support legacy submission via GuC, which
78 * represents in-order queue. The kernel driver packs ring tail pointer and an
79 * ELSP context descriptor dword into Work Item.
80 * See guc_add_request()
84 static inline u32 intel_hws_preempt_done_address(struct intel_engine_cs *engine)
86 return (i915_ggtt_offset(engine->status_page.vma) +
87 I915_GEM_HWS_PREEMPT_ADDR);
90 static inline struct i915_priolist *to_priolist(struct rb_node *rb)
92 return rb_entry(rb, struct i915_priolist, node);
95 static inline bool is_high_priority(struct intel_guc_client *client)
97 return (client->priority == GUC_CLIENT_PRIORITY_KMD_HIGH ||
98 client->priority == GUC_CLIENT_PRIORITY_HIGH);
101 static int reserve_doorbell(struct intel_guc_client *client)
103 unsigned long offset;
107 GEM_BUG_ON(client->doorbell_id != GUC_DOORBELL_INVALID);
110 * The bitmap tracks which doorbell registers are currently in use.
111 * It is split into two halves; the first half is used for normal
112 * priority contexts, the second half for high-priority ones.
115 end = GUC_NUM_DOORBELLS / 2;
116 if (is_high_priority(client)) {
121 id = find_next_zero_bit(client->guc->doorbell_bitmap, end, offset);
125 __set_bit(id, client->guc->doorbell_bitmap);
126 client->doorbell_id = id;
127 DRM_DEBUG_DRIVER("client %u (high prio=%s) reserved doorbell: %d\n",
128 client->stage_id, yesno(is_high_priority(client)),
133 static bool has_doorbell(struct intel_guc_client *client)
135 if (client->doorbell_id == GUC_DOORBELL_INVALID)
138 return test_bit(client->doorbell_id, client->guc->doorbell_bitmap);
141 static void unreserve_doorbell(struct intel_guc_client *client)
143 GEM_BUG_ON(!has_doorbell(client));
145 __clear_bit(client->doorbell_id, client->guc->doorbell_bitmap);
146 client->doorbell_id = GUC_DOORBELL_INVALID;
150 * Tell the GuC to allocate or deallocate a specific doorbell
153 static int __guc_allocate_doorbell(struct intel_guc *guc, u32 stage_id)
156 INTEL_GUC_ACTION_ALLOCATE_DOORBELL,
160 return intel_guc_send(guc, action, ARRAY_SIZE(action));
163 static int __guc_deallocate_doorbell(struct intel_guc *guc, u32 stage_id)
166 INTEL_GUC_ACTION_DEALLOCATE_DOORBELL,
170 return intel_guc_send(guc, action, ARRAY_SIZE(action));
173 static struct guc_stage_desc *__get_stage_desc(struct intel_guc_client *client)
175 struct guc_stage_desc *base = client->guc->stage_desc_pool_vaddr;
177 return &base[client->stage_id];
181 * Initialise, update, or clear doorbell data shared with the GuC
183 * These functions modify shared data and so need access to the mapped
184 * client object which contains the page being used for the doorbell
187 static void __update_doorbell_desc(struct intel_guc_client *client, u16 new_id)
189 struct guc_stage_desc *desc;
191 /* Update the GuC's idea of the doorbell ID */
192 desc = __get_stage_desc(client);
193 desc->db_id = new_id;
196 static struct guc_doorbell_info *__get_doorbell(struct intel_guc_client *client)
198 return client->vaddr + client->doorbell_offset;
201 static bool __doorbell_valid(struct intel_guc *guc, u16 db_id)
203 struct drm_i915_private *dev_priv = guc_to_i915(guc);
205 GEM_BUG_ON(db_id >= GUC_NUM_DOORBELLS);
206 return I915_READ(GEN8_DRBREGL(db_id)) & GEN8_DRB_VALID;
209 static void __init_doorbell(struct intel_guc_client *client)
211 struct guc_doorbell_info *doorbell;
213 doorbell = __get_doorbell(client);
214 doorbell->db_status = GUC_DOORBELL_ENABLED;
215 doorbell->cookie = 0;
218 static void __fini_doorbell(struct intel_guc_client *client)
220 struct guc_doorbell_info *doorbell;
221 u16 db_id = client->doorbell_id;
223 doorbell = __get_doorbell(client);
224 doorbell->db_status = GUC_DOORBELL_DISABLED;
226 /* Doorbell release flow requires that we wait for GEN8_DRB_VALID bit
227 * to go to zero after updating db_status before we call the GuC to
228 * release the doorbell
230 if (wait_for_us(!__doorbell_valid(client->guc, db_id), 10))
231 WARN_ONCE(true, "Doorbell never became invalid after disable\n");
234 static int create_doorbell(struct intel_guc_client *client)
238 if (WARN_ON(!has_doorbell(client)))
239 return -ENODEV; /* internal setup error, should never happen */
241 __update_doorbell_desc(client, client->doorbell_id);
242 __init_doorbell(client);
244 ret = __guc_allocate_doorbell(client->guc, client->stage_id);
246 __fini_doorbell(client);
247 __update_doorbell_desc(client, GUC_DOORBELL_INVALID);
248 DRM_DEBUG_DRIVER("Couldn't create client %u doorbell: %d\n",
249 client->stage_id, ret);
256 static int destroy_doorbell(struct intel_guc_client *client)
260 GEM_BUG_ON(!has_doorbell(client));
262 __fini_doorbell(client);
263 ret = __guc_deallocate_doorbell(client->guc, client->stage_id);
265 DRM_ERROR("Couldn't destroy client %u doorbell: %d\n",
266 client->stage_id, ret);
268 __update_doorbell_desc(client, GUC_DOORBELL_INVALID);
273 static unsigned long __select_cacheline(struct intel_guc *guc)
275 unsigned long offset;
277 /* Doorbell uses a single cache line within a page */
278 offset = offset_in_page(guc->db_cacheline);
280 /* Moving to next cache line to reduce contention */
281 guc->db_cacheline += cache_line_size();
283 DRM_DEBUG_DRIVER("reserved cacheline 0x%lx, next 0x%x, linesize %u\n",
284 offset, guc->db_cacheline, cache_line_size());
288 static inline struct guc_process_desc *
289 __get_process_desc(struct intel_guc_client *client)
291 return client->vaddr + client->proc_desc_offset;
295 * Initialise the process descriptor shared with the GuC firmware.
297 static void guc_proc_desc_init(struct intel_guc_client *client)
299 struct guc_process_desc *desc;
301 desc = memset(__get_process_desc(client), 0, sizeof(*desc));
304 * XXX: pDoorbell and WQVBaseAddress are pointers in process address
305 * space for ring3 clients (set them as in mmap_ioctl) or kernel
306 * space for kernel clients (map on demand instead? May make debug
307 * easier to have it mapped).
309 desc->wq_base_addr = 0;
310 desc->db_base_addr = 0;
312 desc->stage_id = client->stage_id;
313 desc->wq_size_bytes = GUC_WQ_SIZE;
314 desc->wq_status = WQ_STATUS_ACTIVE;
315 desc->priority = client->priority;
318 static void guc_proc_desc_fini(struct intel_guc_client *client)
320 struct guc_process_desc *desc;
322 desc = __get_process_desc(client);
323 memset(desc, 0, sizeof(*desc));
326 static int guc_stage_desc_pool_create(struct intel_guc *guc)
328 struct i915_vma *vma;
331 vma = intel_guc_allocate_vma(guc,
332 PAGE_ALIGN(sizeof(struct guc_stage_desc) *
333 GUC_MAX_STAGE_DESCRIPTORS));
337 vaddr = i915_gem_object_pin_map(vma->obj, I915_MAP_WB);
339 i915_vma_unpin_and_release(&vma, 0);
340 return PTR_ERR(vaddr);
343 guc->stage_desc_pool = vma;
344 guc->stage_desc_pool_vaddr = vaddr;
345 ida_init(&guc->stage_ids);
350 static void guc_stage_desc_pool_destroy(struct intel_guc *guc)
352 ida_destroy(&guc->stage_ids);
353 i915_vma_unpin_and_release(&guc->stage_desc_pool, I915_VMA_RELEASE_MAP);
357 * Initialise/clear the stage descriptor shared with the GuC firmware.
359 * This descriptor tells the GuC where (in GGTT space) to find the important
360 * data structures relating to this client (doorbell, process descriptor,
363 static void guc_stage_desc_init(struct intel_guc_client *client)
365 struct intel_guc *guc = client->guc;
366 struct drm_i915_private *dev_priv = guc_to_i915(guc);
367 struct intel_engine_cs *engine;
368 struct i915_gem_context *ctx = client->owner;
369 struct guc_stage_desc *desc;
373 desc = __get_stage_desc(client);
374 memset(desc, 0, sizeof(*desc));
376 desc->attribute = GUC_STAGE_DESC_ATTR_ACTIVE |
377 GUC_STAGE_DESC_ATTR_KERNEL;
378 if (is_high_priority(client))
379 desc->attribute |= GUC_STAGE_DESC_ATTR_PREEMPT;
380 desc->stage_id = client->stage_id;
381 desc->priority = client->priority;
382 desc->db_id = client->doorbell_id;
384 for_each_engine_masked(engine, dev_priv, client->engines, tmp) {
385 struct intel_context *ce = to_intel_context(ctx, engine);
386 u32 guc_engine_id = engine->guc_id;
387 struct guc_execlist_context *lrc = &desc->lrc[guc_engine_id];
389 /* TODO: We have a design issue to be solved here. Only when we
390 * receive the first batch, we know which engine is used by the
391 * user. But here GuC expects the lrc and ring to be pinned. It
392 * is not an issue for default context, which is the only one
393 * for now who owns a GuC client. But for future owner of GuC
394 * client, need to make sure lrc is pinned prior to enter here.
397 break; /* XXX: continue? */
400 * XXX: When this is a GUC_STAGE_DESC_ATTR_KERNEL client (proxy
401 * submission or, in other words, not using a direct submission
402 * model) the KMD's LRCA is not used for any work submission.
403 * Instead, the GuC uses the LRCA of the user mode context (see
404 * guc_add_request below).
406 lrc->context_desc = lower_32_bits(ce->lrc_desc);
408 /* The state page is after PPHWSP */
409 lrc->ring_lrca = intel_guc_ggtt_offset(guc, ce->state) +
410 LRC_STATE_PN * PAGE_SIZE;
412 /* XXX: In direct submission, the GuC wants the HW context id
413 * here. In proxy submission, it wants the stage id
415 lrc->context_id = (client->stage_id << GUC_ELC_CTXID_OFFSET) |
416 (guc_engine_id << GUC_ELC_ENGINE_OFFSET);
418 lrc->ring_begin = intel_guc_ggtt_offset(guc, ce->ring->vma);
419 lrc->ring_end = lrc->ring_begin + ce->ring->size - 1;
420 lrc->ring_next_free_location = lrc->ring_begin;
421 lrc->ring_current_tail_pointer_value = 0;
423 desc->engines_used |= (1 << guc_engine_id);
426 DRM_DEBUG_DRIVER("Host engines 0x%x => GuC engines used 0x%x\n",
427 client->engines, desc->engines_used);
428 WARN_ON(desc->engines_used == 0);
431 * The doorbell, process descriptor, and workqueue are all parts
432 * of the client object, which the GuC will reference via the GGTT
434 gfx_addr = intel_guc_ggtt_offset(guc, client->vma);
435 desc->db_trigger_phy = sg_dma_address(client->vma->pages->sgl) +
436 client->doorbell_offset;
437 desc->db_trigger_cpu = ptr_to_u64(__get_doorbell(client));
438 desc->db_trigger_uk = gfx_addr + client->doorbell_offset;
439 desc->process_desc = gfx_addr + client->proc_desc_offset;
440 desc->wq_addr = gfx_addr + GUC_DB_SIZE;
441 desc->wq_size = GUC_WQ_SIZE;
443 desc->desc_private = ptr_to_u64(client);
446 static void guc_stage_desc_fini(struct intel_guc_client *client)
448 struct guc_stage_desc *desc;
450 desc = __get_stage_desc(client);
451 memset(desc, 0, sizeof(*desc));
454 /* Construct a Work Item and append it to the GuC's Work Queue */
455 static void guc_wq_item_append(struct intel_guc_client *client,
456 u32 target_engine, u32 context_desc,
457 u32 ring_tail, u32 fence_id)
459 /* wqi_len is in DWords, and does not include the one-word header */
460 const size_t wqi_size = sizeof(struct guc_wq_item);
461 const u32 wqi_len = wqi_size / sizeof(u32) - 1;
462 struct guc_process_desc *desc = __get_process_desc(client);
463 struct guc_wq_item *wqi;
466 lockdep_assert_held(&client->wq_lock);
468 /* For now workqueue item is 4 DWs; workqueue buffer is 2 pages. So we
469 * should not have the case where structure wqi is across page, neither
470 * wrapped to the beginning. This simplifies the implementation below.
472 * XXX: if not the case, we need save data to a temp wqi and copy it to
473 * workqueue buffer dw by dw.
475 BUILD_BUG_ON(wqi_size != 16);
477 /* We expect the WQ to be active if we're appending items to it */
478 GEM_BUG_ON(desc->wq_status != WQ_STATUS_ACTIVE);
480 /* Free space is guaranteed. */
481 wq_off = READ_ONCE(desc->tail);
482 GEM_BUG_ON(CIRC_SPACE(wq_off, READ_ONCE(desc->head),
483 GUC_WQ_SIZE) < wqi_size);
484 GEM_BUG_ON(wq_off & (wqi_size - 1));
486 /* WQ starts from the page after doorbell / process_desc */
487 wqi = client->vaddr + wq_off + GUC_DB_SIZE;
489 if (I915_SELFTEST_ONLY(client->use_nop_wqi)) {
490 wqi->header = WQ_TYPE_NOOP | (wqi_len << WQ_LEN_SHIFT);
492 /* Now fill in the 4-word work queue item */
493 wqi->header = WQ_TYPE_INORDER |
494 (wqi_len << WQ_LEN_SHIFT) |
495 (target_engine << WQ_TARGET_SHIFT) |
497 wqi->context_desc = context_desc;
498 wqi->submit_element_info = ring_tail << WQ_RING_TAIL_SHIFT;
499 GEM_BUG_ON(ring_tail > WQ_RING_TAIL_MAX);
500 wqi->fence_id = fence_id;
503 /* Make the update visible to GuC */
504 WRITE_ONCE(desc->tail, (wq_off + wqi_size) & (GUC_WQ_SIZE - 1));
507 static void guc_ring_doorbell(struct intel_guc_client *client)
509 struct guc_doorbell_info *db;
512 lockdep_assert_held(&client->wq_lock);
514 /* pointer of current doorbell cacheline */
515 db = __get_doorbell(client);
518 * We're not expecting the doorbell cookie to change behind our back,
519 * we also need to treat 0 as a reserved value.
521 cookie = READ_ONCE(db->cookie);
522 WARN_ON_ONCE(xchg(&db->cookie, cookie + 1 ?: cookie + 2) != cookie);
524 /* XXX: doorbell was lost and need to acquire it again */
525 GEM_BUG_ON(db->db_status != GUC_DOORBELL_ENABLED);
528 static void guc_add_request(struct intel_guc *guc, struct i915_request *rq)
530 struct intel_guc_client *client = guc->execbuf_client;
531 struct intel_engine_cs *engine = rq->engine;
532 u32 ctx_desc = lower_32_bits(rq->hw_context->lrc_desc);
533 u32 ring_tail = intel_ring_set_tail(rq->ring, rq->tail) / sizeof(u64);
535 spin_lock(&client->wq_lock);
537 guc_wq_item_append(client, engine->guc_id, ctx_desc,
538 ring_tail, rq->global_seqno);
539 guc_ring_doorbell(client);
541 client->submissions[engine->id] += 1;
543 spin_unlock(&client->wq_lock);
547 * When we're doing submissions using regular execlists backend, writing to
548 * ELSP from CPU side is enough to make sure that writes to ringbuffer pages
549 * pinned in mappable aperture portion of GGTT are visible to command streamer.
550 * Writes done by GuC on our behalf are not guaranteeing such ordering,
551 * therefore, to ensure the flush, we're issuing a POSTING READ.
553 static void flush_ggtt_writes(struct i915_vma *vma)
555 struct drm_i915_private *dev_priv = vma->vm->i915;
557 if (i915_vma_is_map_and_fenceable(vma))
558 POSTING_READ_FW(GUC_STATUS);
561 static void inject_preempt_context(struct work_struct *work)
563 struct guc_preempt_work *preempt_work =
564 container_of(work, typeof(*preempt_work), work);
565 struct intel_engine_cs *engine = preempt_work->engine;
566 struct intel_guc *guc = container_of(preempt_work, typeof(*guc),
567 preempt_work[engine->id]);
568 struct intel_guc_client *client = guc->preempt_client;
569 struct guc_stage_desc *stage_desc = __get_stage_desc(client);
570 struct intel_context *ce = to_intel_context(client->owner, engine);
573 if (!ce->ring->emit) { /* recreate upon load/resume */
574 u32 addr = intel_hws_preempt_done_address(engine);
577 cs = ce->ring->vaddr;
578 if (engine->id == RCS) {
579 cs = gen8_emit_ggtt_write_rcs(cs,
580 GUC_PREEMPT_FINISHED,
582 PIPE_CONTROL_CS_STALL);
584 cs = gen8_emit_ggtt_write(cs,
585 GUC_PREEMPT_FINISHED,
590 *cs++ = MI_USER_INTERRUPT;
593 ce->ring->emit = GUC_PREEMPT_BREADCRUMB_BYTES;
594 GEM_BUG_ON((void *)cs - ce->ring->vaddr != ce->ring->emit);
596 flush_ggtt_writes(ce->ring->vma);
599 spin_lock_irq(&client->wq_lock);
600 guc_wq_item_append(client, engine->guc_id, lower_32_bits(ce->lrc_desc),
601 GUC_PREEMPT_BREADCRUMB_BYTES / sizeof(u64), 0);
602 spin_unlock_irq(&client->wq_lock);
605 * If GuC firmware performs an engine reset while that engine had
606 * a preemption pending, it will set the terminated attribute bit
607 * on our preemption stage descriptor. GuC firmware retains all
608 * pending work items for a high-priority GuC client, unlike the
609 * normal-priority GuC client where work items are dropped. It
610 * wants to make sure the preempt-to-idle work doesn't run when
611 * scheduling resumes, and uses this bit to inform its scheduler
612 * and presumably us as well. Our job is to clear it for the next
613 * preemption after reset, otherwise that and future preemptions
614 * will never complete. We'll just clear it every time.
616 stage_desc->attribute &= ~GUC_STAGE_DESC_ATTR_TERMINATED;
618 data[0] = INTEL_GUC_ACTION_REQUEST_PREEMPTION;
619 data[1] = client->stage_id;
620 data[2] = INTEL_GUC_PREEMPT_OPTION_DROP_WORK_Q |
621 INTEL_GUC_PREEMPT_OPTION_DROP_SUBMIT_Q;
622 data[3] = engine->guc_id;
623 data[4] = guc->execbuf_client->priority;
624 data[5] = guc->execbuf_client->stage_id;
625 data[6] = intel_guc_ggtt_offset(guc, guc->shared_data);
627 if (WARN_ON(intel_guc_send(guc, data, ARRAY_SIZE(data)))) {
628 execlists_clear_active(&engine->execlists,
629 EXECLISTS_ACTIVE_PREEMPT);
630 tasklet_schedule(&engine->execlists.tasklet);
633 (void)I915_SELFTEST_ONLY(engine->execlists.preempt_hang.count++);
637 * We're using user interrupt and HWSP value to mark that preemption has
638 * finished and GPU is idle. Normally, we could unwind and continue similar to
639 * execlists submission path. Unfortunately, with GuC we also need to wait for
640 * it to finish its own postprocessing, before attempting to submit. Otherwise
641 * GuC may silently ignore our submissions, and thus we risk losing request at
642 * best, executing out-of-order and causing kernel panic at worst.
644 #define GUC_PREEMPT_POSTPROCESS_DELAY_MS 10
645 static void wait_for_guc_preempt_report(struct intel_engine_cs *engine)
647 struct intel_guc *guc = &engine->i915->guc;
648 struct guc_shared_ctx_data *data = guc->shared_data_vaddr;
649 struct guc_ctx_report *report =
650 &data->preempt_ctx_report[engine->guc_id];
652 WARN_ON(wait_for_atomic(report->report_return_status ==
653 INTEL_GUC_REPORT_STATUS_COMPLETE,
654 GUC_PREEMPT_POSTPROCESS_DELAY_MS));
656 * GuC is expecting that we're also going to clear the affected context
657 * counter, let's also reset the return status to not depend on GuC
658 * resetting it after recieving another preempt action
660 report->affected_count = 0;
661 report->report_return_status = INTEL_GUC_REPORT_STATUS_UNKNOWN;
664 static void complete_preempt_context(struct intel_engine_cs *engine)
666 struct intel_engine_execlists *execlists = &engine->execlists;
668 GEM_BUG_ON(!execlists_is_active(execlists, EXECLISTS_ACTIVE_PREEMPT));
670 if (inject_preempt_hang(execlists))
673 execlists_cancel_port_requests(execlists);
674 execlists_unwind_incomplete_requests(execlists);
676 wait_for_guc_preempt_report(engine);
677 intel_write_status_page(engine, I915_GEM_HWS_PREEMPT, 0);
681 * guc_submit() - Submit commands through GuC
682 * @engine: engine associated with the commands
684 * The only error here arises if the doorbell hardware isn't functioning
685 * as expected, which really shouln't happen.
687 static void guc_submit(struct intel_engine_cs *engine)
689 struct intel_guc *guc = &engine->i915->guc;
690 struct intel_engine_execlists * const execlists = &engine->execlists;
691 struct execlist_port *port = execlists->port;
694 for (n = 0; n < execlists_num_ports(execlists); n++) {
695 struct i915_request *rq;
698 rq = port_unpack(&port[n], &count);
699 if (rq && count == 0) {
700 port_set(&port[n], port_pack(rq, ++count));
702 flush_ggtt_writes(rq->ring->vma);
704 guc_add_request(guc, rq);
709 static void port_assign(struct execlist_port *port, struct i915_request *rq)
711 GEM_BUG_ON(port_isset(port));
713 port_set(port, i915_request_get(rq));
716 static inline int rq_prio(const struct i915_request *rq)
718 return rq->sched.attr.priority;
721 static inline int port_prio(const struct execlist_port *port)
723 return rq_prio(port_request(port));
726 static bool __guc_dequeue(struct intel_engine_cs *engine)
728 struct intel_engine_execlists * const execlists = &engine->execlists;
729 struct execlist_port *port = execlists->port;
730 struct i915_request *last = NULL;
731 const struct execlist_port * const last_port =
732 &execlists->port[execlists->port_mask];
736 lockdep_assert_held(&engine->timeline.lock);
738 if (port_isset(port)) {
739 if (intel_engine_has_preemption(engine)) {
740 struct guc_preempt_work *preempt_work =
741 &engine->i915->guc.preempt_work[engine->id];
742 int prio = execlists->queue_priority_hint;
744 if (__execlists_need_preempt(prio, port_prio(port))) {
745 execlists_set_active(execlists,
746 EXECLISTS_ACTIVE_PREEMPT);
747 queue_work(engine->i915->guc.preempt_wq,
748 &preempt_work->work);
754 if (port_isset(port))
757 GEM_BUG_ON(port_isset(port));
759 while ((rb = rb_first_cached(&execlists->queue))) {
760 struct i915_priolist *p = to_priolist(rb);
761 struct i915_request *rq, *rn;
764 priolist_for_each_request_consume(rq, rn, p, i) {
765 if (last && rq->hw_context != last->hw_context) {
766 if (port == last_port)
770 port_assign(port, last);
774 list_del_init(&rq->sched.link);
776 __i915_request_submit(rq);
777 trace_i915_request_in(rq, port_index(port, execlists));
783 rb_erase_cached(&p->node, &execlists->queue);
784 if (p->priority != I915_PRIORITY_NORMAL)
785 kmem_cache_free(engine->i915->priorities, p);
788 execlists->queue_priority_hint =
789 rb ? to_priolist(rb)->priority : INT_MIN;
791 port_assign(port, last);
793 execlists_user_begin(execlists, execlists->port);
795 /* We must always keep the beast fed if we have work piled up */
796 GEM_BUG_ON(port_isset(execlists->port) &&
797 !execlists_is_active(execlists, EXECLISTS_ACTIVE_USER));
798 GEM_BUG_ON(rb_first_cached(&execlists->queue) &&
799 !port_isset(execlists->port));
804 static void guc_dequeue(struct intel_engine_cs *engine)
806 if (__guc_dequeue(engine))
810 static void guc_submission_tasklet(unsigned long data)
812 struct intel_engine_cs * const engine = (struct intel_engine_cs *)data;
813 struct intel_engine_execlists * const execlists = &engine->execlists;
814 struct execlist_port *port = execlists->port;
815 struct i915_request *rq;
818 spin_lock_irqsave(&engine->timeline.lock, flags);
820 rq = port_request(port);
821 while (rq && i915_request_completed(rq)) {
822 trace_i915_request_out(rq);
823 i915_request_put(rq);
825 port = execlists_port_complete(execlists, port);
826 if (port_isset(port)) {
827 execlists_user_begin(execlists, port);
828 rq = port_request(port);
830 execlists_user_end(execlists);
835 if (execlists_is_active(execlists, EXECLISTS_ACTIVE_PREEMPT) &&
836 intel_read_status_page(engine, I915_GEM_HWS_PREEMPT) ==
837 GUC_PREEMPT_FINISHED)
838 complete_preempt_context(engine);
840 if (!execlists_is_active(execlists, EXECLISTS_ACTIVE_PREEMPT))
843 spin_unlock_irqrestore(&engine->timeline.lock, flags);
846 static void guc_reset_prepare(struct intel_engine_cs *engine)
848 struct intel_engine_execlists * const execlists = &engine->execlists;
850 GEM_TRACE("%s\n", engine->name);
853 * Prevent request submission to the hardware until we have
854 * completed the reset in i915_gem_reset_finish(). If a request
855 * is completed by one engine, it may then queue a request
856 * to a second via its execlists->tasklet *just* as we are
857 * calling engine->init_hw() and also writing the ELSP.
858 * Turning off the execlists->tasklet until the reset is over
861 __tasklet_disable_sync_once(&execlists->tasklet);
864 * We're using worker to queue preemption requests from the tasklet in
865 * GuC submission mode.
866 * Even though tasklet was disabled, we may still have a worker queued.
867 * Let's make sure that all workers scheduled before disabling the
868 * tasklet are completed before continuing with the reset.
870 if (engine->i915->guc.preempt_wq)
871 flush_workqueue(engine->i915->guc.preempt_wq);
875 * Everything below here is concerned with setup & teardown, and is
876 * therefore not part of the somewhat time-critical batch-submission
877 * path of guc_submit() above.
880 /* Check that a doorbell register is in the expected state */
881 static bool doorbell_ok(struct intel_guc *guc, u16 db_id)
885 GEM_BUG_ON(db_id >= GUC_NUM_DOORBELLS);
887 valid = __doorbell_valid(guc, db_id);
889 if (test_bit(db_id, guc->doorbell_bitmap) == valid)
892 DRM_DEBUG_DRIVER("Doorbell %u has unexpected state: valid=%s\n",
893 db_id, yesno(valid));
898 static bool guc_verify_doorbells(struct intel_guc *guc)
900 bool doorbells_ok = true;
903 for (db_id = 0; db_id < GUC_NUM_DOORBELLS; ++db_id)
904 if (!doorbell_ok(guc, db_id))
905 doorbells_ok = false;
911 * guc_client_alloc() - Allocate an intel_guc_client
912 * @dev_priv: driver private data structure
913 * @engines: The set of engines to enable for this client
914 * @priority: four levels priority _CRITICAL, _HIGH, _NORMAL and _LOW
915 * The kernel client to replace ExecList submission is created with
916 * NORMAL priority. Priority of a client for scheduler can be HIGH,
917 * while a preemption context can use CRITICAL.
918 * @ctx: the context that owns the client (we use the default render
921 * Return: An intel_guc_client object if success, else NULL.
923 static struct intel_guc_client *
924 guc_client_alloc(struct drm_i915_private *dev_priv,
927 struct i915_gem_context *ctx)
929 struct intel_guc_client *client;
930 struct intel_guc *guc = &dev_priv->guc;
931 struct i915_vma *vma;
935 client = kzalloc(sizeof(*client), GFP_KERNEL);
937 return ERR_PTR(-ENOMEM);
941 client->engines = engines;
942 client->priority = priority;
943 client->doorbell_id = GUC_DOORBELL_INVALID;
944 spin_lock_init(&client->wq_lock);
946 ret = ida_simple_get(&guc->stage_ids, 0, GUC_MAX_STAGE_DESCRIPTORS,
951 client->stage_id = ret;
953 /* The first page is doorbell/proc_desc. Two followed pages are wq. */
954 vma = intel_guc_allocate_vma(guc, GUC_DB_SIZE + GUC_WQ_SIZE);
960 /* We'll keep just the first (doorbell/proc) page permanently kmap'd. */
963 vaddr = i915_gem_object_pin_map(vma->obj, I915_MAP_WB);
965 ret = PTR_ERR(vaddr);
968 client->vaddr = vaddr;
970 ret = reserve_doorbell(client);
974 client->doorbell_offset = __select_cacheline(guc);
977 * Since the doorbell only requires a single cacheline, we can save
978 * space by putting the application process descriptor in the same
979 * page. Use the half of the page that doesn't include the doorbell.
981 if (client->doorbell_offset >= (GUC_DB_SIZE / 2))
982 client->proc_desc_offset = 0;
984 client->proc_desc_offset = (GUC_DB_SIZE / 2);
986 DRM_DEBUG_DRIVER("new priority %u client %p for engine(s) 0x%x: stage_id %u\n",
987 priority, client, client->engines, client->stage_id);
988 DRM_DEBUG_DRIVER("doorbell id %u, cacheline offset 0x%lx\n",
989 client->doorbell_id, client->doorbell_offset);
994 i915_gem_object_unpin_map(client->vma->obj);
996 i915_vma_unpin_and_release(&client->vma, 0);
998 ida_simple_remove(&guc->stage_ids, client->stage_id);
1001 return ERR_PTR(ret);
1004 static void guc_client_free(struct intel_guc_client *client)
1006 unreserve_doorbell(client);
1007 i915_vma_unpin_and_release(&client->vma, I915_VMA_RELEASE_MAP);
1008 ida_simple_remove(&client->guc->stage_ids, client->stage_id);
1012 static inline bool ctx_save_restore_disabled(struct intel_context *ce)
1014 u32 sr = ce->lrc_reg_state[CTX_CONTEXT_CONTROL + 1];
1016 #define SR_DISABLED \
1017 _MASKED_BIT_ENABLE(CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT | \
1018 CTX_CTRL_ENGINE_CTX_SAVE_INHIBIT)
1020 return (sr & SR_DISABLED) == SR_DISABLED;
1025 static int guc_clients_create(struct intel_guc *guc)
1027 struct drm_i915_private *dev_priv = guc_to_i915(guc);
1028 struct intel_guc_client *client;
1030 GEM_BUG_ON(guc->execbuf_client);
1031 GEM_BUG_ON(guc->preempt_client);
1033 client = guc_client_alloc(dev_priv,
1034 INTEL_INFO(dev_priv)->ring_mask,
1035 GUC_CLIENT_PRIORITY_KMD_NORMAL,
1036 dev_priv->kernel_context);
1037 if (IS_ERR(client)) {
1038 DRM_ERROR("Failed to create GuC client for submission!\n");
1039 return PTR_ERR(client);
1041 guc->execbuf_client = client;
1043 if (dev_priv->preempt_context) {
1044 client = guc_client_alloc(dev_priv,
1045 INTEL_INFO(dev_priv)->ring_mask,
1046 GUC_CLIENT_PRIORITY_KMD_HIGH,
1047 dev_priv->preempt_context);
1048 if (IS_ERR(client)) {
1049 DRM_ERROR("Failed to create GuC client for preemption!\n");
1050 guc_client_free(guc->execbuf_client);
1051 guc->execbuf_client = NULL;
1052 return PTR_ERR(client);
1054 guc->preempt_client = client;
1060 static void guc_clients_destroy(struct intel_guc *guc)
1062 struct intel_guc_client *client;
1064 client = fetch_and_zero(&guc->preempt_client);
1066 guc_client_free(client);
1068 client = fetch_and_zero(&guc->execbuf_client);
1070 guc_client_free(client);
1073 static int __guc_client_enable(struct intel_guc_client *client)
1077 guc_proc_desc_init(client);
1078 guc_stage_desc_init(client);
1080 ret = create_doorbell(client);
1087 guc_stage_desc_fini(client);
1088 guc_proc_desc_fini(client);
1092 static void __guc_client_disable(struct intel_guc_client *client)
1095 * By the time we're here, GuC may have already been reset. if that is
1096 * the case, instead of trying (in vain) to communicate with it, let's
1097 * just cleanup the doorbell HW and our internal state.
1099 if (intel_guc_is_alive(client->guc))
1100 destroy_doorbell(client);
1102 __fini_doorbell(client);
1104 guc_stage_desc_fini(client);
1105 guc_proc_desc_fini(client);
1108 static int guc_clients_enable(struct intel_guc *guc)
1112 ret = __guc_client_enable(guc->execbuf_client);
1116 if (guc->preempt_client) {
1117 ret = __guc_client_enable(guc->preempt_client);
1119 __guc_client_disable(guc->execbuf_client);
1127 static void guc_clients_disable(struct intel_guc *guc)
1129 if (guc->preempt_client)
1130 __guc_client_disable(guc->preempt_client);
1132 if (guc->execbuf_client)
1133 __guc_client_disable(guc->execbuf_client);
1137 * Set up the memory resources to be shared with the GuC (via the GGTT)
1138 * at firmware loading time.
1140 int intel_guc_submission_init(struct intel_guc *guc)
1142 struct drm_i915_private *dev_priv = guc_to_i915(guc);
1143 struct intel_engine_cs *engine;
1144 enum intel_engine_id id;
1147 if (guc->stage_desc_pool)
1150 ret = guc_stage_desc_pool_create(guc);
1154 * Keep static analysers happy, let them know that we allocated the
1155 * vma after testing that it didn't exist earlier.
1157 GEM_BUG_ON(!guc->stage_desc_pool);
1159 WARN_ON(!guc_verify_doorbells(guc));
1160 ret = guc_clients_create(guc);
1164 for_each_engine(engine, dev_priv, id) {
1165 guc->preempt_work[id].engine = engine;
1166 INIT_WORK(&guc->preempt_work[id].work, inject_preempt_context);
1172 guc_stage_desc_pool_destroy(guc);
1176 void intel_guc_submission_fini(struct intel_guc *guc)
1178 struct drm_i915_private *dev_priv = guc_to_i915(guc);
1179 struct intel_engine_cs *engine;
1180 enum intel_engine_id id;
1182 for_each_engine(engine, dev_priv, id)
1183 cancel_work_sync(&guc->preempt_work[id].work);
1185 guc_clients_destroy(guc);
1186 WARN_ON(!guc_verify_doorbells(guc));
1188 if (guc->stage_desc_pool)
1189 guc_stage_desc_pool_destroy(guc);
1192 static void guc_interrupts_capture(struct drm_i915_private *dev_priv)
1194 struct intel_rps *rps = &dev_priv->gt_pm.rps;
1195 struct intel_engine_cs *engine;
1196 enum intel_engine_id id;
1199 /* tell all command streamers to forward interrupts (but not vblank)
1202 irqs = _MASKED_BIT_ENABLE(GFX_INTERRUPT_STEERING);
1203 for_each_engine(engine, dev_priv, id)
1204 I915_WRITE(RING_MODE_GEN7(engine), irqs);
1206 /* route USER_INTERRUPT to Host, all others are sent to GuC. */
1207 irqs = GT_RENDER_USER_INTERRUPT << GEN8_RCS_IRQ_SHIFT |
1208 GT_RENDER_USER_INTERRUPT << GEN8_BCS_IRQ_SHIFT;
1209 /* These three registers have the same bit definitions */
1210 I915_WRITE(GUC_BCS_RCS_IER, ~irqs);
1211 I915_WRITE(GUC_VCS2_VCS1_IER, ~irqs);
1212 I915_WRITE(GUC_WD_VECS_IER, ~irqs);
1215 * The REDIRECT_TO_GUC bit of the PMINTRMSK register directs all
1216 * (unmasked) PM interrupts to the GuC. All other bits of this
1217 * register *disable* generation of a specific interrupt.
1219 * 'pm_intrmsk_mbz' indicates bits that are NOT to be set when
1220 * writing to the PM interrupt mask register, i.e. interrupts
1221 * that must not be disabled.
1223 * If the GuC is handling these interrupts, then we must not let
1224 * the PM code disable ANY interrupt that the GuC is expecting.
1225 * So for each ENABLED (0) bit in this register, we must SET the
1226 * bit in pm_intrmsk_mbz so that it's left enabled for the GuC.
1227 * GuC needs ARAT expired interrupt unmasked hence it is set in
1230 * Here we CLEAR REDIRECT_TO_GUC bit in pm_intrmsk_mbz, which will
1231 * result in the register bit being left SET!
1233 rps->pm_intrmsk_mbz |= ARAT_EXPIRED_INTRMSK;
1234 rps->pm_intrmsk_mbz &= ~GEN8_PMINTR_DISABLE_REDIRECT_TO_GUC;
1237 static void guc_interrupts_release(struct drm_i915_private *dev_priv)
1239 struct intel_rps *rps = &dev_priv->gt_pm.rps;
1240 struct intel_engine_cs *engine;
1241 enum intel_engine_id id;
1245 * tell all command streamers NOT to forward interrupts or vblank
1248 irqs = _MASKED_FIELD(GFX_FORWARD_VBLANK_MASK, GFX_FORWARD_VBLANK_NEVER);
1249 irqs |= _MASKED_BIT_DISABLE(GFX_INTERRUPT_STEERING);
1250 for_each_engine(engine, dev_priv, id)
1251 I915_WRITE(RING_MODE_GEN7(engine), irqs);
1253 /* route all GT interrupts to the host */
1254 I915_WRITE(GUC_BCS_RCS_IER, 0);
1255 I915_WRITE(GUC_VCS2_VCS1_IER, 0);
1256 I915_WRITE(GUC_WD_VECS_IER, 0);
1258 rps->pm_intrmsk_mbz |= GEN8_PMINTR_DISABLE_REDIRECT_TO_GUC;
1259 rps->pm_intrmsk_mbz &= ~ARAT_EXPIRED_INTRMSK;
1262 static void guc_submission_park(struct intel_engine_cs *engine)
1264 intel_engine_unpin_breadcrumbs_irq(engine);
1267 static void guc_submission_unpark(struct intel_engine_cs *engine)
1269 intel_engine_pin_breadcrumbs_irq(engine);
1272 static void guc_set_default_submission(struct intel_engine_cs *engine)
1275 * We inherit a bunch of functions from execlists that we'd like
1278 * engine->submit_request = execlists_submit_request;
1279 * engine->cancel_requests = execlists_cancel_requests;
1280 * engine->schedule = execlists_schedule;
1282 * But we need to override the actual submission backend in order
1283 * to talk to the GuC.
1285 intel_execlists_set_default_submission(engine);
1287 engine->execlists.tasklet.func = guc_submission_tasklet;
1289 engine->park = guc_submission_park;
1290 engine->unpark = guc_submission_unpark;
1292 engine->reset.prepare = guc_reset_prepare;
1294 engine->flags &= ~I915_ENGINE_SUPPORTS_STATS;
1297 int intel_guc_submission_enable(struct intel_guc *guc)
1299 struct drm_i915_private *dev_priv = guc_to_i915(guc);
1300 struct intel_engine_cs *engine;
1301 enum intel_engine_id id;
1305 * We're using GuC work items for submitting work through GuC. Since
1306 * we're coalescing multiple requests from a single context into a
1307 * single work item prior to assigning it to execlist_port, we can
1308 * never have more work items than the total number of ports (for all
1309 * engines). The GuC firmware is controlling the HEAD of work queue,
1310 * and it is guaranteed that it will remove the work item from the
1311 * queue before our request is completed.
1313 BUILD_BUG_ON(ARRAY_SIZE(engine->execlists.port) *
1314 sizeof(struct guc_wq_item) *
1315 I915_NUM_ENGINES > GUC_WQ_SIZE);
1317 GEM_BUG_ON(!guc->execbuf_client);
1319 err = intel_guc_sample_forcewake(guc);
1323 err = guc_clients_enable(guc);
1327 /* Take over from manual control of ELSP (execlists) */
1328 guc_interrupts_capture(dev_priv);
1330 for_each_engine(engine, dev_priv, id) {
1331 engine->set_default_submission = guc_set_default_submission;
1332 engine->set_default_submission(engine);
1338 void intel_guc_submission_disable(struct intel_guc *guc)
1340 struct drm_i915_private *dev_priv = guc_to_i915(guc);
1342 GEM_BUG_ON(dev_priv->gt.awake); /* GT should be parked first */
1344 guc_interrupts_release(dev_priv);
1345 guc_clients_disable(guc);
1348 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1349 #include "selftests/intel_guc.c"