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
24 * Ben Widawsky <ben@bwidawsk.net>
25 * Michel Thierry <michel.thierry@intel.com>
26 * Thomas Daniel <thomas.daniel@intel.com>
27 * Oscar Mateo <oscar.mateo@intel.com>
32 * DOC: Logical Rings, Logical Ring Contexts and Execlists
35 * GEN8 brings an expansion of the HW contexts: "Logical Ring Contexts".
36 * These expanded contexts enable a number of new abilities, especially
37 * "Execlists" (also implemented in this file).
39 * One of the main differences with the legacy HW contexts is that logical
40 * ring contexts incorporate many more things to the context's state, like
41 * PDPs or ringbuffer control registers:
43 * The reason why PDPs are included in the context is straightforward: as
44 * PPGTTs (per-process GTTs) are actually per-context, having the PDPs
45 * contained there mean you don't need to do a ppgtt->switch_mm yourself,
46 * instead, the GPU will do it for you on the context switch.
48 * But, what about the ringbuffer control registers (head, tail, etc..)?
49 * shouldn't we just need a set of those per engine command streamer? This is
50 * where the name "Logical Rings" starts to make sense: by virtualizing the
51 * rings, the engine cs shifts to a new "ring buffer" with every context
52 * switch. When you want to submit a workload to the GPU you: A) choose your
53 * context, B) find its appropriate virtualized ring, C) write commands to it
54 * and then, finally, D) tell the GPU to switch to that context.
56 * Instead of the legacy MI_SET_CONTEXT, the way you tell the GPU to switch
57 * to a contexts is via a context execution list, ergo "Execlists".
60 * Regarding the creation of contexts, we have:
62 * - One global default context.
63 * - One local default context for each opened fd.
64 * - One local extra context for each context create ioctl call.
66 * Now that ringbuffers belong per-context (and not per-engine, like before)
67 * and that contexts are uniquely tied to a given engine (and not reusable,
68 * like before) we need:
70 * - One ringbuffer per-engine inside each context.
71 * - One backing object per-engine inside each context.
73 * The global default context starts its life with these new objects fully
74 * allocated and populated. The local default context for each opened fd is
75 * more complex, because we don't know at creation time which engine is going
76 * to use them. To handle this, we have implemented a deferred creation of LR
79 * The local context starts its life as a hollow or blank holder, that only
80 * gets populated for a given engine once we receive an execbuffer. If later
81 * on we receive another execbuffer ioctl for the same context but a different
82 * engine, we allocate/populate a new ringbuffer and context backing object and
85 * Finally, regarding local contexts created using the ioctl call: as they are
86 * only allowed with the render ring, we can allocate & populate them right
87 * away (no need to defer anything, at least for now).
89 * Execlists implementation:
90 * Execlists are the new method by which, on gen8+ hardware, workloads are
91 * submitted for execution (as opposed to the legacy, ringbuffer-based, method).
92 * This method works as follows:
94 * When a request is committed, its commands (the BB start and any leading or
95 * trailing commands, like the seqno breadcrumbs) are placed in the ringbuffer
96 * for the appropriate context. The tail pointer in the hardware context is not
97 * updated at this time, but instead, kept by the driver in the ringbuffer
98 * structure. A structure representing this request is added to a request queue
99 * for the appropriate engine: this structure contains a copy of the context's
100 * tail after the request was written to the ring buffer and a pointer to the
103 * If the engine's request queue was empty before the request was added, the
104 * queue is processed immediately. Otherwise the queue will be processed during
105 * a context switch interrupt. In any case, elements on the queue will get sent
106 * (in pairs) to the GPU's ExecLists Submit Port (ELSP, for short) with a
107 * globally unique 20-bits submission ID.
109 * When execution of a request completes, the GPU updates the context status
110 * buffer with a context complete event and generates a context switch interrupt.
111 * During the interrupt handling, the driver examines the events in the buffer:
112 * for each context complete event, if the announced ID matches that on the head
113 * of the request queue, then that request is retired and removed from the queue.
115 * After processing, if any requests were retired and the queue is not empty
116 * then a new execution list can be submitted. The two requests at the front of
117 * the queue are next to be submitted but since a context may not occur twice in
118 * an execution list, if subsequent requests have the same ID as the first then
119 * the two requests must be combined. This is done simply by discarding requests
120 * at the head of the queue until either only one requests is left (in which case
121 * we use a NULL second context) or the first two requests have unique IDs.
123 * By always executing the first two requests in the queue the driver ensures
124 * that the GPU is kept as busy as possible. In the case where a single context
125 * completes but a second context is still executing, the request for this second
126 * context will be at the head of the queue when we remove the first one. This
127 * request will then be resubmitted along with a new request for a different context,
128 * which will cause the hardware to continue executing the second request and queue
129 * the new request (the GPU detects the condition of a context getting preempted
130 * with the same context and optimizes the context switch flow by not doing
131 * preemption, but just sampling the new tail pointer).
134 #include <linux/interrupt.h>
136 #include "i915_drv.h"
137 #include "i915_perf.h"
138 #include "i915_trace.h"
139 #include "i915_vgpu.h"
140 #include "intel_breadcrumbs.h"
141 #include "intel_context.h"
142 #include "intel_engine_pm.h"
143 #include "intel_gt.h"
144 #include "intel_gt_pm.h"
145 #include "intel_gt_requests.h"
146 #include "intel_lrc_reg.h"
147 #include "intel_mocs.h"
148 #include "intel_reset.h"
149 #include "intel_ring.h"
150 #include "intel_workarounds.h"
151 #include "shmem_utils.h"
153 #define RING_EXECLIST_QFULL (1 << 0x2)
154 #define RING_EXECLIST1_VALID (1 << 0x3)
155 #define RING_EXECLIST0_VALID (1 << 0x4)
156 #define RING_EXECLIST_ACTIVE_STATUS (3 << 0xE)
157 #define RING_EXECLIST1_ACTIVE (1 << 0x11)
158 #define RING_EXECLIST0_ACTIVE (1 << 0x12)
160 #define GEN8_CTX_STATUS_IDLE_ACTIVE (1 << 0)
161 #define GEN8_CTX_STATUS_PREEMPTED (1 << 1)
162 #define GEN8_CTX_STATUS_ELEMENT_SWITCH (1 << 2)
163 #define GEN8_CTX_STATUS_ACTIVE_IDLE (1 << 3)
164 #define GEN8_CTX_STATUS_COMPLETE (1 << 4)
165 #define GEN8_CTX_STATUS_LITE_RESTORE (1 << 15)
167 #define GEN8_CTX_STATUS_COMPLETED_MASK \
168 (GEN8_CTX_STATUS_COMPLETE | GEN8_CTX_STATUS_PREEMPTED)
170 #define CTX_DESC_FORCE_RESTORE BIT_ULL(2)
172 #define GEN12_CTX_STATUS_SWITCHED_TO_NEW_QUEUE (0x1) /* lower csb dword */
173 #define GEN12_CTX_SWITCH_DETAIL(csb_dw) ((csb_dw) & 0xF) /* upper csb dword */
174 #define GEN12_CSB_SW_CTX_ID_MASK GENMASK(25, 15)
175 #define GEN12_IDLE_CTX_ID 0x7FF
176 #define GEN12_CSB_CTX_VALID(csb_dw) \
177 (FIELD_GET(GEN12_CSB_SW_CTX_ID_MASK, csb_dw) != GEN12_IDLE_CTX_ID)
179 /* Typical size of the average request (2 pipecontrols and a MI_BB) */
180 #define EXECLISTS_REQUEST_SIZE 64 /* bytes */
182 struct virtual_engine {
183 struct intel_engine_cs base;
184 struct intel_context context;
187 * We allow only a single request through the virtual engine at a time
188 * (each request in the timeline waits for the completion fence of
189 * the previous before being submitted). By restricting ourselves to
190 * only submitting a single request, each request is placed on to a
191 * physical to maximise load spreading (by virtue of the late greedy
192 * scheduling -- each real engine takes the next available request
195 struct i915_request *request;
198 * We keep a rbtree of available virtual engines inside each physical
199 * engine, sorted by priority. Here we preallocate the nodes we need
200 * for the virtual engine, indexed by physical_engine->id.
205 } nodes[I915_NUM_ENGINES];
208 * Keep track of bonded pairs -- restrictions upon on our selection
209 * of physical engines any particular request may be submitted to.
210 * If we receive a submit-fence from a master engine, we will only
211 * use one of sibling_mask physical engines.
214 const struct intel_engine_cs *master;
215 intel_engine_mask_t sibling_mask;
217 unsigned int num_bonds;
219 /* And finally, which physical engines this virtual engine maps onto. */
220 unsigned int num_siblings;
221 struct intel_engine_cs *siblings[];
224 static struct virtual_engine *to_virtual_engine(struct intel_engine_cs *engine)
226 GEM_BUG_ON(!intel_engine_is_virtual(engine));
227 return container_of(engine, struct virtual_engine, base);
230 static int __execlists_context_alloc(struct intel_context *ce,
231 struct intel_engine_cs *engine);
233 static void execlists_init_reg_state(u32 *reg_state,
234 const struct intel_context *ce,
235 const struct intel_engine_cs *engine,
236 const struct intel_ring *ring,
239 __execlists_update_reg_state(const struct intel_context *ce,
240 const struct intel_engine_cs *engine,
243 static int lrc_ring_mi_mode(const struct intel_engine_cs *engine)
245 if (INTEL_GEN(engine->i915) >= 12)
247 else if (INTEL_GEN(engine->i915) >= 9)
249 else if (engine->class == RENDER_CLASS)
255 static int lrc_ring_gpr0(const struct intel_engine_cs *engine)
257 if (INTEL_GEN(engine->i915) >= 12)
259 else if (INTEL_GEN(engine->i915) >= 9)
261 else if (engine->class == RENDER_CLASS)
267 static int lrc_ring_wa_bb_per_ctx(const struct intel_engine_cs *engine)
269 if (INTEL_GEN(engine->i915) >= 12)
271 else if (INTEL_GEN(engine->i915) >= 9 || engine->class == RENDER_CLASS)
277 static int lrc_ring_indirect_ptr(const struct intel_engine_cs *engine)
281 x = lrc_ring_wa_bb_per_ctx(engine);
288 static int lrc_ring_indirect_offset(const struct intel_engine_cs *engine)
292 x = lrc_ring_indirect_ptr(engine);
299 static int lrc_ring_cmd_buf_cctl(const struct intel_engine_cs *engine)
301 if (engine->class != RENDER_CLASS)
304 if (INTEL_GEN(engine->i915) >= 12)
306 else if (INTEL_GEN(engine->i915) >= 11)
313 lrc_ring_indirect_offset_default(const struct intel_engine_cs *engine)
315 switch (INTEL_GEN(engine->i915)) {
317 MISSING_CASE(INTEL_GEN(engine->i915));
320 return GEN12_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
322 return GEN11_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
324 return GEN10_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
326 return GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
328 return GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
333 lrc_ring_setup_indirect_ctx(u32 *regs,
334 const struct intel_engine_cs *engine,
335 u32 ctx_bb_ggtt_addr,
339 GEM_BUG_ON(!IS_ALIGNED(size, CACHELINE_BYTES));
340 GEM_BUG_ON(lrc_ring_indirect_ptr(engine) == -1);
341 regs[lrc_ring_indirect_ptr(engine) + 1] =
342 ctx_bb_ggtt_addr | (size / CACHELINE_BYTES);
344 GEM_BUG_ON(lrc_ring_indirect_offset(engine) == -1);
345 regs[lrc_ring_indirect_offset(engine) + 1] =
346 lrc_ring_indirect_offset_default(engine) << 6;
349 static u32 intel_context_get_runtime(const struct intel_context *ce)
352 * We can use either ppHWSP[16] which is recorded before the context
353 * switch (and so excludes the cost of context switches) or use the
354 * value from the context image itself, which is saved/restored earlier
355 * and so includes the cost of the save.
357 return READ_ONCE(ce->lrc_reg_state[CTX_TIMESTAMP]);
360 static void mark_eio(struct i915_request *rq)
362 if (i915_request_completed(rq))
365 GEM_BUG_ON(i915_request_signaled(rq));
367 i915_request_set_error_once(rq, -EIO);
368 i915_request_mark_complete(rq);
371 static struct i915_request *
372 active_request(const struct intel_timeline * const tl, struct i915_request *rq)
374 struct i915_request *active = rq;
377 list_for_each_entry_continue_reverse(rq, &tl->requests, link) {
378 if (i915_request_completed(rq))
388 static inline u32 intel_hws_preempt_address(struct intel_engine_cs *engine)
390 return (i915_ggtt_offset(engine->status_page.vma) +
391 I915_GEM_HWS_PREEMPT_ADDR);
395 ring_set_paused(const struct intel_engine_cs *engine, int state)
398 * We inspect HWS_PREEMPT with a semaphore inside
399 * engine->emit_fini_breadcrumb. If the dword is true,
400 * the ring is paused as the semaphore will busywait
401 * until the dword is false.
403 engine->status_page.addr[I915_GEM_HWS_PREEMPT] = state;
408 static inline struct i915_priolist *to_priolist(struct rb_node *rb)
410 return rb_entry(rb, struct i915_priolist, node);
413 static inline int rq_prio(const struct i915_request *rq)
415 return READ_ONCE(rq->sched.attr.priority);
418 static int effective_prio(const struct i915_request *rq)
420 int prio = rq_prio(rq);
423 * If this request is special and must not be interrupted at any
424 * cost, so be it. Note we are only checking the most recent request
425 * in the context and so may be masking an earlier vip request. It
426 * is hoped that under the conditions where nopreempt is used, this
427 * will not matter (i.e. all requests to that context will be
428 * nopreempt for as long as desired).
430 if (i915_request_has_nopreempt(rq))
431 prio = I915_PRIORITY_UNPREEMPTABLE;
436 static int queue_prio(const struct intel_engine_execlists *execlists)
438 struct i915_priolist *p;
441 rb = rb_first_cached(&execlists->queue);
446 * As the priolist[] are inverted, with the highest priority in [0],
447 * we have to flip the index value to become priority.
450 if (!I915_USER_PRIORITY_SHIFT)
453 return ((p->priority + 1) << I915_USER_PRIORITY_SHIFT) - ffs(p->used);
456 static inline bool need_preempt(const struct intel_engine_cs *engine,
457 const struct i915_request *rq,
462 if (!intel_engine_has_semaphores(engine))
466 * Check if the current priority hint merits a preemption attempt.
468 * We record the highest value priority we saw during rescheduling
469 * prior to this dequeue, therefore we know that if it is strictly
470 * less than the current tail of ESLP[0], we do not need to force
471 * a preempt-to-idle cycle.
473 * However, the priority hint is a mere hint that we may need to
474 * preempt. If that hint is stale or we may be trying to preempt
475 * ourselves, ignore the request.
477 * More naturally we would write
478 * prio >= max(0, last);
479 * except that we wish to prevent triggering preemption at the same
480 * priority level: the task that is running should remain running
481 * to preserve FIFO ordering of dependencies.
483 last_prio = max(effective_prio(rq), I915_PRIORITY_NORMAL - 1);
484 if (engine->execlists.queue_priority_hint <= last_prio)
488 * Check against the first request in ELSP[1], it will, thanks to the
489 * power of PI, be the highest priority of that context.
491 if (!list_is_last(&rq->sched.link, &engine->active.requests) &&
492 rq_prio(list_next_entry(rq, sched.link)) > last_prio)
496 struct virtual_engine *ve =
497 rb_entry(rb, typeof(*ve), nodes[engine->id].rb);
498 bool preempt = false;
500 if (engine == ve->siblings[0]) { /* only preempt one sibling */
501 struct i915_request *next;
504 next = READ_ONCE(ve->request);
506 preempt = rq_prio(next) > last_prio;
515 * If the inflight context did not trigger the preemption, then maybe
516 * it was the set of queued requests? Pick the highest priority in
517 * the queue (the first active priolist) and see if it deserves to be
518 * running instead of ELSP[0].
520 * The highest priority request in the queue can not be either
521 * ELSP[0] or ELSP[1] as, thanks again to PI, if it was the same
522 * context, it's priority would not exceed ELSP[0] aka last_prio.
524 return queue_prio(&engine->execlists) > last_prio;
527 __maybe_unused static inline bool
528 assert_priority_queue(const struct i915_request *prev,
529 const struct i915_request *next)
532 * Without preemption, the prev may refer to the still active element
533 * which we refuse to let go.
535 * Even with preemption, there are times when we think it is better not
536 * to preempt and leave an ostensibly lower priority request in flight.
538 if (i915_request_is_active(prev))
541 return rq_prio(prev) >= rq_prio(next);
545 * The context descriptor encodes various attributes of a context,
546 * including its GTT address and some flags. Because it's fairly
547 * expensive to calculate, we'll just do it once and cache the result,
548 * which remains valid until the context is unpinned.
550 * This is what a descriptor looks like, from LSB to MSB::
552 * bits 0-11: flags, GEN8_CTX_* (cached in ctx->desc_template)
553 * bits 12-31: LRCA, GTT address of (the HWSP of) this context
554 * bits 32-52: ctx ID, a globally unique tag (highest bit used by GuC)
555 * bits 53-54: mbz, reserved for use by hardware
556 * bits 55-63: group ID, currently unused and set to 0
558 * Starting from Gen11, the upper dword of the descriptor has a new format:
560 * bits 32-36: reserved
561 * bits 37-47: SW context ID
562 * bits 48:53: engine instance
563 * bit 54: mbz, reserved for use by hardware
564 * bits 55-60: SW counter
565 * bits 61-63: engine class
567 * engine info, SW context ID and SW counter need to form a unique number
568 * (Context ID) per lrc.
571 lrc_descriptor(struct intel_context *ce, struct intel_engine_cs *engine)
575 desc = INTEL_LEGACY_32B_CONTEXT;
576 if (i915_vm_is_4lvl(ce->vm))
577 desc = INTEL_LEGACY_64B_CONTEXT;
578 desc <<= GEN8_CTX_ADDRESSING_MODE_SHIFT;
580 desc |= GEN8_CTX_VALID | GEN8_CTX_PRIVILEGE;
581 if (IS_GEN(engine->i915, 8))
582 desc |= GEN8_CTX_L3LLC_COHERENT;
584 return i915_ggtt_offset(ce->state) | desc;
587 static inline unsigned int dword_in_page(void *addr)
589 return offset_in_page(addr) / sizeof(u32);
592 static void set_offsets(u32 *regs,
594 const struct intel_engine_cs *engine,
596 #define NOP(x) (BIT(7) | (x))
597 #define LRI(count, flags) ((flags) << 6 | (count) | BUILD_BUG_ON_ZERO(count >= BIT(6)))
598 #define POSTED BIT(0)
599 #define REG(x) (((x) >> 2) | BUILD_BUG_ON_ZERO(x >= 0x200))
601 (((x) >> 9) | BIT(7) | BUILD_BUG_ON_ZERO(x >= 0x10000)), \
603 #define END(total_state_size) 0, (total_state_size)
605 const u32 base = engine->mmio_base;
610 if (*data & BIT(7)) { /* skip */
611 count = *data++ & ~BIT(7);
613 memset32(regs, MI_NOOP, count);
618 count = *data & 0x3f;
622 *regs = MI_LOAD_REGISTER_IMM(count);
624 *regs |= MI_LRI_FORCE_POSTED;
625 if (INTEL_GEN(engine->i915) >= 11)
626 *regs |= MI_LRI_LRM_CS_MMIO;
637 offset |= v & ~BIT(7);
638 } while (v & BIT(7));
640 regs[0] = base + (offset << 2);
650 /* Clear past the tail for HW access */
651 GEM_BUG_ON(dword_in_page(regs) > count);
652 memset32(regs, MI_NOOP, count - dword_in_page(regs));
654 /* Close the batch; used mainly by live_lrc_layout() */
655 *regs = MI_BATCH_BUFFER_END;
656 if (INTEL_GEN(engine->i915) >= 10)
661 static const u8 gen8_xcs_offsets[] = {
696 static const u8 gen9_xcs_offsets[] = {
780 static const u8 gen12_xcs_offsets[] = {
812 static const u8 gen8_rcs_offsets[] = {
849 static const u8 gen9_rcs_offsets[] = {
933 static const u8 gen11_rcs_offsets[] = {
974 static const u8 gen12_rcs_offsets[] = {
1076 static const u8 *reg_offsets(const struct intel_engine_cs *engine)
1079 * The gen12+ lists only have the registers we program in the basic
1080 * default state. We rely on the context image using relative
1081 * addressing to automatic fixup the register state between the
1082 * physical engines for virtual engine.
1084 GEM_BUG_ON(INTEL_GEN(engine->i915) >= 12 &&
1085 !intel_engine_has_relative_mmio(engine));
1087 if (engine->class == RENDER_CLASS) {
1088 if (INTEL_GEN(engine->i915) >= 12)
1089 return gen12_rcs_offsets;
1090 else if (INTEL_GEN(engine->i915) >= 11)
1091 return gen11_rcs_offsets;
1092 else if (INTEL_GEN(engine->i915) >= 9)
1093 return gen9_rcs_offsets;
1095 return gen8_rcs_offsets;
1097 if (INTEL_GEN(engine->i915) >= 12)
1098 return gen12_xcs_offsets;
1099 else if (INTEL_GEN(engine->i915) >= 9)
1100 return gen9_xcs_offsets;
1102 return gen8_xcs_offsets;
1106 static struct i915_request *
1107 __unwind_incomplete_requests(struct intel_engine_cs *engine)
1109 struct i915_request *rq, *rn, *active = NULL;
1110 struct list_head *pl;
1111 int prio = I915_PRIORITY_INVALID;
1113 lockdep_assert_held(&engine->active.lock);
1115 list_for_each_entry_safe_reverse(rq, rn,
1116 &engine->active.requests,
1118 if (i915_request_completed(rq))
1121 __i915_request_unsubmit(rq);
1124 * Push the request back into the queue for later resubmission.
1125 * If this request is not native to this physical engine (i.e.
1126 * it came from a virtual source), push it back onto the virtual
1127 * engine so that it can be moved across onto another physical
1128 * engine as load dictates.
1130 if (likely(rq->execution_mask == engine->mask)) {
1131 GEM_BUG_ON(rq_prio(rq) == I915_PRIORITY_INVALID);
1132 if (rq_prio(rq) != prio) {
1134 pl = i915_sched_lookup_priolist(engine, prio);
1136 GEM_BUG_ON(RB_EMPTY_ROOT(&engine->execlists.queue.rb_root));
1138 list_move(&rq->sched.link, pl);
1139 set_bit(I915_FENCE_FLAG_PQUEUE, &rq->fence.flags);
1141 /* Check in case we rollback so far we wrap [size/2] */
1142 if (intel_ring_direction(rq->ring,
1144 rq->ring->tail + 8) > 0)
1145 rq->context->lrc.desc |= CTX_DESC_FORCE_RESTORE;
1149 struct intel_engine_cs *owner = rq->context->engine;
1151 WRITE_ONCE(rq->engine, owner);
1152 owner->submit_request(rq);
1160 struct i915_request *
1161 execlists_unwind_incomplete_requests(struct intel_engine_execlists *execlists)
1163 struct intel_engine_cs *engine =
1164 container_of(execlists, typeof(*engine), execlists);
1166 return __unwind_incomplete_requests(engine);
1170 execlists_context_status_change(struct i915_request *rq, unsigned long status)
1173 * Only used when GVT-g is enabled now. When GVT-g is disabled,
1174 * The compiler should eliminate this function as dead-code.
1176 if (!IS_ENABLED(CONFIG_DRM_I915_GVT))
1179 atomic_notifier_call_chain(&rq->engine->context_status_notifier,
1183 static void intel_engine_context_in(struct intel_engine_cs *engine)
1185 unsigned long flags;
1187 if (atomic_add_unless(&engine->stats.active, 1, 0))
1190 write_seqlock_irqsave(&engine->stats.lock, flags);
1191 if (!atomic_add_unless(&engine->stats.active, 1, 0)) {
1192 engine->stats.start = ktime_get();
1193 atomic_inc(&engine->stats.active);
1195 write_sequnlock_irqrestore(&engine->stats.lock, flags);
1198 static void intel_engine_context_out(struct intel_engine_cs *engine)
1200 unsigned long flags;
1202 GEM_BUG_ON(!atomic_read(&engine->stats.active));
1204 if (atomic_add_unless(&engine->stats.active, -1, 1))
1207 write_seqlock_irqsave(&engine->stats.lock, flags);
1208 if (atomic_dec_and_test(&engine->stats.active)) {
1209 engine->stats.total =
1210 ktime_add(engine->stats.total,
1211 ktime_sub(ktime_get(), engine->stats.start));
1213 write_sequnlock_irqrestore(&engine->stats.lock, flags);
1217 execlists_check_context(const struct intel_context *ce,
1218 const struct intel_engine_cs *engine)
1220 const struct intel_ring *ring = ce->ring;
1221 u32 *regs = ce->lrc_reg_state;
1225 if (regs[CTX_RING_START] != i915_ggtt_offset(ring->vma)) {
1226 pr_err("%s: context submitted with incorrect RING_START [%08x], expected %08x\n",
1228 regs[CTX_RING_START],
1229 i915_ggtt_offset(ring->vma));
1230 regs[CTX_RING_START] = i915_ggtt_offset(ring->vma);
1234 if ((regs[CTX_RING_CTL] & ~(RING_WAIT | RING_WAIT_SEMAPHORE)) !=
1235 (RING_CTL_SIZE(ring->size) | RING_VALID)) {
1236 pr_err("%s: context submitted with incorrect RING_CTL [%08x], expected %08x\n",
1239 (u32)(RING_CTL_SIZE(ring->size) | RING_VALID));
1240 regs[CTX_RING_CTL] = RING_CTL_SIZE(ring->size) | RING_VALID;
1244 x = lrc_ring_mi_mode(engine);
1245 if (x != -1 && regs[x + 1] & (regs[x + 1] >> 16) & STOP_RING) {
1246 pr_err("%s: context submitted with STOP_RING [%08x] in RING_MI_MODE\n",
1247 engine->name, regs[x + 1]);
1248 regs[x + 1] &= ~STOP_RING;
1249 regs[x + 1] |= STOP_RING << 16;
1253 WARN_ONCE(!valid, "Invalid lrc state found before submission\n");
1256 static void restore_default_state(struct intel_context *ce,
1257 struct intel_engine_cs *engine)
1261 regs = memset(ce->lrc_reg_state, 0, engine->context_size - PAGE_SIZE);
1262 execlists_init_reg_state(regs, ce, engine, ce->ring, true);
1264 ce->runtime.last = intel_context_get_runtime(ce);
1267 static void reset_active(struct i915_request *rq,
1268 struct intel_engine_cs *engine)
1270 struct intel_context * const ce = rq->context;
1274 * The executing context has been cancelled. We want to prevent
1275 * further execution along this context and propagate the error on
1276 * to anything depending on its results.
1278 * In __i915_request_submit(), we apply the -EIO and remove the
1279 * requests' payloads for any banned requests. But first, we must
1280 * rewind the context back to the start of the incomplete request so
1281 * that we do not jump back into the middle of the batch.
1283 * We preserve the breadcrumbs and semaphores of the incomplete
1284 * requests so that inter-timeline dependencies (i.e other timelines)
1285 * remain correctly ordered. And we defer to __i915_request_submit()
1286 * so that all asynchronous waits are correctly handled.
1288 ENGINE_TRACE(engine, "{ rq=%llx:%lld }\n",
1289 rq->fence.context, rq->fence.seqno);
1291 /* On resubmission of the active request, payload will be scrubbed */
1292 if (i915_request_completed(rq))
1295 head = active_request(ce->timeline, rq)->head;
1296 head = intel_ring_wrap(ce->ring, head);
1298 /* Scrub the context image to prevent replaying the previous batch */
1299 restore_default_state(ce, engine);
1300 __execlists_update_reg_state(ce, engine, head);
1302 /* We've switched away, so this should be a no-op, but intent matters */
1303 ce->lrc.desc |= CTX_DESC_FORCE_RESTORE;
1306 static void st_update_runtime_underflow(struct intel_context *ce, s32 dt)
1308 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1309 ce->runtime.num_underflow += dt < 0;
1310 ce->runtime.max_underflow = max_t(u32, ce->runtime.max_underflow, -dt);
1314 static void intel_context_update_runtime(struct intel_context *ce)
1319 if (intel_context_is_barrier(ce))
1322 old = ce->runtime.last;
1323 ce->runtime.last = intel_context_get_runtime(ce);
1324 dt = ce->runtime.last - old;
1326 if (unlikely(dt <= 0)) {
1327 CE_TRACE(ce, "runtime underflow: last=%u, new=%u, delta=%d\n",
1328 old, ce->runtime.last, dt);
1329 st_update_runtime_underflow(ce, dt);
1333 ewma_runtime_add(&ce->runtime.avg, dt);
1334 ce->runtime.total += dt;
1337 static inline struct intel_engine_cs *
1338 __execlists_schedule_in(struct i915_request *rq)
1340 struct intel_engine_cs * const engine = rq->engine;
1341 struct intel_context * const ce = rq->context;
1343 intel_context_get(ce);
1345 if (unlikely(intel_context_is_banned(ce)))
1346 reset_active(rq, engine);
1348 if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
1349 execlists_check_context(ce, engine);
1352 /* Use a fixed tag for OA and friends */
1353 GEM_BUG_ON(ce->tag <= BITS_PER_LONG);
1354 ce->lrc.ccid = ce->tag;
1356 /* We don't need a strict matching tag, just different values */
1357 unsigned int tag = ffs(READ_ONCE(engine->context_tag));
1359 GEM_BUG_ON(tag == 0 || tag >= BITS_PER_LONG);
1360 clear_bit(tag - 1, &engine->context_tag);
1361 ce->lrc.ccid = tag << (GEN11_SW_CTX_ID_SHIFT - 32);
1363 BUILD_BUG_ON(BITS_PER_LONG > GEN12_MAX_CONTEXT_HW_ID);
1366 ce->lrc.ccid |= engine->execlists.ccid;
1368 __intel_gt_pm_get(engine->gt);
1369 if (engine->fw_domain && !atomic_fetch_inc(&engine->fw_active))
1370 intel_uncore_forcewake_get(engine->uncore, engine->fw_domain);
1371 execlists_context_status_change(rq, INTEL_CONTEXT_SCHEDULE_IN);
1372 intel_engine_context_in(engine);
1377 static inline struct i915_request *
1378 execlists_schedule_in(struct i915_request *rq, int idx)
1380 struct intel_context * const ce = rq->context;
1381 struct intel_engine_cs *old;
1383 GEM_BUG_ON(!intel_engine_pm_is_awake(rq->engine));
1384 trace_i915_request_in(rq, idx);
1386 old = READ_ONCE(ce->inflight);
1389 WRITE_ONCE(ce->inflight, __execlists_schedule_in(rq));
1392 } while (!try_cmpxchg(&ce->inflight, &old, ptr_inc(old)));
1394 GEM_BUG_ON(intel_context_inflight(ce) != rq->engine);
1395 return i915_request_get(rq);
1398 static void kick_siblings(struct i915_request *rq, struct intel_context *ce)
1400 struct virtual_engine *ve = container_of(ce, typeof(*ve), context);
1401 struct i915_request *next = READ_ONCE(ve->request);
1403 if (next == rq || (next && next->execution_mask & ~rq->execution_mask))
1404 tasklet_hi_schedule(&ve->base.execlists.tasklet);
1408 __execlists_schedule_out(struct i915_request *rq,
1409 struct intel_engine_cs * const engine,
1412 struct intel_context * const ce = rq->context;
1415 * NB process_csb() is not under the engine->active.lock and hence
1416 * schedule_out can race with schedule_in meaning that we should
1417 * refrain from doing non-trivial work here.
1421 * If we have just completed this context, the engine may now be
1422 * idle and we want to re-enter powersaving.
1424 if (list_is_last_rcu(&rq->link, &ce->timeline->requests) &&
1425 i915_request_completed(rq))
1426 intel_engine_add_retire(engine, ce->timeline);
1428 ccid >>= GEN11_SW_CTX_ID_SHIFT - 32;
1429 ccid &= GEN12_MAX_CONTEXT_HW_ID;
1430 if (ccid < BITS_PER_LONG) {
1431 GEM_BUG_ON(ccid == 0);
1432 GEM_BUG_ON(test_bit(ccid - 1, &engine->context_tag));
1433 set_bit(ccid - 1, &engine->context_tag);
1436 intel_context_update_runtime(ce);
1437 intel_engine_context_out(engine);
1438 execlists_context_status_change(rq, INTEL_CONTEXT_SCHEDULE_OUT);
1439 if (engine->fw_domain && !atomic_dec_return(&engine->fw_active))
1440 intel_uncore_forcewake_put(engine->uncore, engine->fw_domain);
1441 intel_gt_pm_put_async(engine->gt);
1444 * If this is part of a virtual engine, its next request may
1445 * have been blocked waiting for access to the active context.
1446 * We have to kick all the siblings again in case we need to
1447 * switch (e.g. the next request is not runnable on this
1448 * engine). Hopefully, we will already have submitted the next
1449 * request before the tasklet runs and do not need to rebuild
1450 * each virtual tree and kick everyone again.
1452 if (ce->engine != engine)
1453 kick_siblings(rq, ce);
1455 intel_context_put(ce);
1459 execlists_schedule_out(struct i915_request *rq)
1461 struct intel_context * const ce = rq->context;
1462 struct intel_engine_cs *cur, *old;
1465 trace_i915_request_out(rq);
1467 ccid = rq->context->lrc.ccid;
1468 old = READ_ONCE(ce->inflight);
1470 cur = ptr_unmask_bits(old, 2) ? ptr_dec(old) : NULL;
1471 while (!try_cmpxchg(&ce->inflight, &old, cur));
1473 __execlists_schedule_out(rq, old, ccid);
1475 i915_request_put(rq);
1478 static u64 execlists_update_context(struct i915_request *rq)
1480 struct intel_context *ce = rq->context;
1481 u64 desc = ce->lrc.desc;
1485 * WaIdleLiteRestore:bdw,skl
1487 * We should never submit the context with the same RING_TAIL twice
1488 * just in case we submit an empty ring, which confuses the HW.
1490 * We append a couple of NOOPs (gen8_emit_wa_tail) after the end of
1491 * the normal request to be able to always advance the RING_TAIL on
1492 * subsequent resubmissions (for lite restore). Should that fail us,
1493 * and we try and submit the same tail again, force the context
1496 * If we need to return to a preempted context, we need to skip the
1497 * lite-restore and force it to reload the RING_TAIL. Otherwise, the
1498 * HW has a tendency to ignore us rewinding the TAIL to the end of
1499 * an earlier request.
1501 GEM_BUG_ON(ce->lrc_reg_state[CTX_RING_TAIL] != rq->ring->tail);
1502 prev = rq->ring->tail;
1503 tail = intel_ring_set_tail(rq->ring, rq->tail);
1504 if (unlikely(intel_ring_direction(rq->ring, tail, prev) <= 0))
1505 desc |= CTX_DESC_FORCE_RESTORE;
1506 ce->lrc_reg_state[CTX_RING_TAIL] = tail;
1507 rq->tail = rq->wa_tail;
1510 * Make sure the context image is complete before we submit it to HW.
1512 * Ostensibly, writes (including the WCB) should be flushed prior to
1513 * an uncached write such as our mmio register access, the empirical
1514 * evidence (esp. on Braswell) suggests that the WC write into memory
1515 * may not be visible to the HW prior to the completion of the UC
1516 * register write and that we may begin execution from the context
1517 * before its image is complete leading to invalid PD chasing.
1521 ce->lrc.desc &= ~CTX_DESC_FORCE_RESTORE;
1525 static inline void write_desc(struct intel_engine_execlists *execlists, u64 desc, u32 port)
1527 if (execlists->ctrl_reg) {
1528 writel(lower_32_bits(desc), execlists->submit_reg + port * 2);
1529 writel(upper_32_bits(desc), execlists->submit_reg + port * 2 + 1);
1531 writel(upper_32_bits(desc), execlists->submit_reg);
1532 writel(lower_32_bits(desc), execlists->submit_reg);
1536 static __maybe_unused char *
1537 dump_port(char *buf, int buflen, const char *prefix, struct i915_request *rq)
1542 snprintf(buf, buflen, "%sccid:%x %llx:%lld%s prio %d",
1544 rq->context->lrc.ccid,
1545 rq->fence.context, rq->fence.seqno,
1546 i915_request_completed(rq) ? "!" :
1547 i915_request_started(rq) ? "*" :
1554 static __maybe_unused void
1555 trace_ports(const struct intel_engine_execlists *execlists,
1557 struct i915_request * const *ports)
1559 const struct intel_engine_cs *engine =
1560 container_of(execlists, typeof(*engine), execlists);
1561 char __maybe_unused p0[40], p1[40];
1566 ENGINE_TRACE(engine, "%s { %s%s }\n", msg,
1567 dump_port(p0, sizeof(p0), "", ports[0]),
1568 dump_port(p1, sizeof(p1), ", ", ports[1]));
1572 reset_in_progress(const struct intel_engine_execlists *execlists)
1574 return unlikely(!__tasklet_is_enabled(&execlists->tasklet));
1577 static __maybe_unused bool
1578 assert_pending_valid(const struct intel_engine_execlists *execlists,
1581 struct intel_engine_cs *engine =
1582 container_of(execlists, typeof(*engine), execlists);
1583 struct i915_request * const *port, *rq;
1584 struct intel_context *ce = NULL;
1585 bool sentinel = false;
1588 trace_ports(execlists, msg, execlists->pending);
1590 /* We may be messing around with the lists during reset, lalala */
1591 if (reset_in_progress(execlists))
1594 if (!execlists->pending[0]) {
1595 GEM_TRACE_ERR("%s: Nothing pending for promotion!\n",
1600 if (execlists->pending[execlists_num_ports(execlists)]) {
1601 GEM_TRACE_ERR("%s: Excess pending[%d] for promotion!\n",
1602 engine->name, execlists_num_ports(execlists));
1606 for (port = execlists->pending; (rq = *port); port++) {
1607 unsigned long flags;
1610 GEM_BUG_ON(!kref_read(&rq->fence.refcount));
1611 GEM_BUG_ON(!i915_request_is_active(rq));
1613 if (ce == rq->context) {
1614 GEM_TRACE_ERR("%s: Dup context:%llx in pending[%zd]\n",
1616 ce->timeline->fence_context,
1617 port - execlists->pending);
1622 if (ccid == ce->lrc.ccid) {
1623 GEM_TRACE_ERR("%s: Dup ccid:%x context:%llx in pending[%zd]\n",
1625 ccid, ce->timeline->fence_context,
1626 port - execlists->pending);
1629 ccid = ce->lrc.ccid;
1632 * Sentinels are supposed to be the last request so they flush
1633 * the current execution off the HW. Check that they are the only
1634 * request in the pending submission.
1637 GEM_TRACE_ERR("%s: context:%llx after sentinel in pending[%zd]\n",
1639 ce->timeline->fence_context,
1640 port - execlists->pending);
1643 sentinel = i915_request_has_sentinel(rq);
1645 /* Hold tightly onto the lock to prevent concurrent retires! */
1646 if (!spin_trylock_irqsave(&rq->lock, flags))
1649 if (i915_request_completed(rq))
1652 if (i915_active_is_idle(&ce->active) &&
1653 !intel_context_is_barrier(ce)) {
1654 GEM_TRACE_ERR("%s: Inactive context:%llx in pending[%zd]\n",
1656 ce->timeline->fence_context,
1657 port - execlists->pending);
1662 if (!i915_vma_is_pinned(ce->state)) {
1663 GEM_TRACE_ERR("%s: Unpinned context:%llx in pending[%zd]\n",
1665 ce->timeline->fence_context,
1666 port - execlists->pending);
1671 if (!i915_vma_is_pinned(ce->ring->vma)) {
1672 GEM_TRACE_ERR("%s: Unpinned ring:%llx in pending[%zd]\n",
1674 ce->timeline->fence_context,
1675 port - execlists->pending);
1681 spin_unlock_irqrestore(&rq->lock, flags);
1689 static void execlists_submit_ports(struct intel_engine_cs *engine)
1691 struct intel_engine_execlists *execlists = &engine->execlists;
1694 GEM_BUG_ON(!assert_pending_valid(execlists, "submit"));
1697 * We can skip acquiring intel_runtime_pm_get() here as it was taken
1698 * on our behalf by the request (see i915_gem_mark_busy()) and it will
1699 * not be relinquished until the device is idle (see
1700 * i915_gem_idle_work_handler()). As a precaution, we make sure
1701 * that all ELSP are drained i.e. we have processed the CSB,
1702 * before allowing ourselves to idle and calling intel_runtime_pm_put().
1704 GEM_BUG_ON(!intel_engine_pm_is_awake(engine));
1707 * ELSQ note: the submit queue is not cleared after being submitted
1708 * to the HW so we need to make sure we always clean it up. This is
1709 * currently ensured by the fact that we always write the same number
1710 * of elsq entries, keep this in mind before changing the loop below.
1712 for (n = execlists_num_ports(execlists); n--; ) {
1713 struct i915_request *rq = execlists->pending[n];
1715 write_desc(execlists,
1716 rq ? execlists_update_context(rq) : 0,
1720 /* we need to manually load the submit queue */
1721 if (execlists->ctrl_reg)
1722 writel(EL_CTRL_LOAD, execlists->ctrl_reg);
1725 static bool ctx_single_port_submission(const struct intel_context *ce)
1727 return (IS_ENABLED(CONFIG_DRM_I915_GVT) &&
1728 intel_context_force_single_submission(ce));
1731 static bool can_merge_ctx(const struct intel_context *prev,
1732 const struct intel_context *next)
1737 if (ctx_single_port_submission(prev))
1743 static unsigned long i915_request_flags(const struct i915_request *rq)
1745 return READ_ONCE(rq->fence.flags);
1748 static bool can_merge_rq(const struct i915_request *prev,
1749 const struct i915_request *next)
1751 GEM_BUG_ON(prev == next);
1752 GEM_BUG_ON(!assert_priority_queue(prev, next));
1755 * We do not submit known completed requests. Therefore if the next
1756 * request is already completed, we can pretend to merge it in
1757 * with the previous context (and we will skip updating the ELSP
1758 * and tracking). Thus hopefully keeping the ELSP full with active
1759 * contexts, despite the best efforts of preempt-to-busy to confuse
1762 if (i915_request_completed(next))
1765 if (unlikely((i915_request_flags(prev) ^ i915_request_flags(next)) &
1766 (BIT(I915_FENCE_FLAG_NOPREEMPT) |
1767 BIT(I915_FENCE_FLAG_SENTINEL))))
1770 if (!can_merge_ctx(prev->context, next->context))
1773 GEM_BUG_ON(i915_seqno_passed(prev->fence.seqno, next->fence.seqno));
1777 static void virtual_update_register_offsets(u32 *regs,
1778 struct intel_engine_cs *engine)
1780 set_offsets(regs, reg_offsets(engine), engine, false);
1783 static bool virtual_matches(const struct virtual_engine *ve,
1784 const struct i915_request *rq,
1785 const struct intel_engine_cs *engine)
1787 const struct intel_engine_cs *inflight;
1789 if (!(rq->execution_mask & engine->mask)) /* We peeked too soon! */
1793 * We track when the HW has completed saving the context image
1794 * (i.e. when we have seen the final CS event switching out of
1795 * the context) and must not overwrite the context image before
1796 * then. This restricts us to only using the active engine
1797 * while the previous virtualized request is inflight (so
1798 * we reuse the register offsets). This is a very small
1799 * hystersis on the greedy seelction algorithm.
1801 inflight = intel_context_inflight(&ve->context);
1802 if (inflight && inflight != engine)
1808 static void virtual_xfer_context(struct virtual_engine *ve,
1809 struct intel_engine_cs *engine)
1813 if (likely(engine == ve->siblings[0]))
1816 GEM_BUG_ON(READ_ONCE(ve->context.inflight));
1817 if (!intel_engine_has_relative_mmio(engine))
1818 virtual_update_register_offsets(ve->context.lrc_reg_state,
1822 * Move the bound engine to the top of the list for
1823 * future execution. We then kick this tasklet first
1824 * before checking others, so that we preferentially
1825 * reuse this set of bound registers.
1827 for (n = 1; n < ve->num_siblings; n++) {
1828 if (ve->siblings[n] == engine) {
1829 swap(ve->siblings[n], ve->siblings[0]);
1835 #define for_each_waiter(p__, rq__) \
1836 list_for_each_entry_lockless(p__, \
1837 &(rq__)->sched.waiters_list, \
1840 #define for_each_signaler(p__, rq__) \
1841 list_for_each_entry_rcu(p__, \
1842 &(rq__)->sched.signalers_list, \
1845 static void defer_request(struct i915_request *rq, struct list_head * const pl)
1850 * We want to move the interrupted request to the back of
1851 * the round-robin list (i.e. its priority level), but
1852 * in doing so, we must then move all requests that were in
1853 * flight and were waiting for the interrupted request to
1854 * be run after it again.
1857 struct i915_dependency *p;
1859 GEM_BUG_ON(i915_request_is_active(rq));
1860 list_move_tail(&rq->sched.link, pl);
1862 for_each_waiter(p, rq) {
1863 struct i915_request *w =
1864 container_of(p->waiter, typeof(*w), sched);
1866 if (p->flags & I915_DEPENDENCY_WEAK)
1869 /* Leave semaphores spinning on the other engines */
1870 if (w->engine != rq->engine)
1873 /* No waiter should start before its signaler */
1874 GEM_BUG_ON(i915_request_has_initial_breadcrumb(w) &&
1875 i915_request_started(w) &&
1876 !i915_request_completed(rq));
1878 GEM_BUG_ON(i915_request_is_active(w));
1879 if (!i915_request_is_ready(w))
1882 if (rq_prio(w) < rq_prio(rq))
1885 GEM_BUG_ON(rq_prio(w) > rq_prio(rq));
1886 list_move_tail(&w->sched.link, &list);
1889 rq = list_first_entry_or_null(&list, typeof(*rq), sched.link);
1893 static void defer_active(struct intel_engine_cs *engine)
1895 struct i915_request *rq;
1897 rq = __unwind_incomplete_requests(engine);
1901 defer_request(rq, i915_sched_lookup_priolist(engine, rq_prio(rq)));
1905 need_timeslice(const struct intel_engine_cs *engine,
1906 const struct i915_request *rq,
1907 const struct rb_node *rb)
1911 if (!intel_engine_has_timeslices(engine))
1914 hint = engine->execlists.queue_priority_hint;
1917 const struct virtual_engine *ve =
1918 rb_entry(rb, typeof(*ve), nodes[engine->id].rb);
1919 const struct intel_engine_cs *inflight =
1920 intel_context_inflight(&ve->context);
1922 if (!inflight || inflight == engine) {
1923 struct i915_request *next;
1926 next = READ_ONCE(ve->request);
1928 hint = max(hint, rq_prio(next));
1933 if (!list_is_last(&rq->sched.link, &engine->active.requests))
1934 hint = max(hint, rq_prio(list_next_entry(rq, sched.link)));
1936 GEM_BUG_ON(hint >= I915_PRIORITY_UNPREEMPTABLE);
1937 return hint >= effective_prio(rq);
1941 timeslice_yield(const struct intel_engine_execlists *el,
1942 const struct i915_request *rq)
1945 * Once bitten, forever smitten!
1947 * If the active context ever busy-waited on a semaphore,
1948 * it will be treated as a hog until the end of its timeslice (i.e.
1949 * until it is scheduled out and replaced by a new submission,
1950 * possibly even its own lite-restore). The HW only sends an interrupt
1951 * on the first miss, and we do know if that semaphore has been
1952 * signaled, or even if it is now stuck on another semaphore. Play
1953 * safe, yield if it might be stuck -- it will be given a fresh
1954 * timeslice in the near future.
1956 return rq->context->lrc.ccid == READ_ONCE(el->yield);
1960 timeslice_expired(const struct intel_engine_execlists *el,
1961 const struct i915_request *rq)
1963 return timer_expired(&el->timer) || timeslice_yield(el, rq);
1967 switch_prio(struct intel_engine_cs *engine, const struct i915_request *rq)
1969 if (list_is_last(&rq->sched.link, &engine->active.requests))
1970 return engine->execlists.queue_priority_hint;
1972 return rq_prio(list_next_entry(rq, sched.link));
1975 static inline unsigned long
1976 timeslice(const struct intel_engine_cs *engine)
1978 return READ_ONCE(engine->props.timeslice_duration_ms);
1981 static unsigned long active_timeslice(const struct intel_engine_cs *engine)
1983 const struct intel_engine_execlists *execlists = &engine->execlists;
1984 const struct i915_request *rq = *execlists->active;
1986 if (!rq || i915_request_completed(rq))
1989 if (READ_ONCE(execlists->switch_priority_hint) < effective_prio(rq))
1992 return timeslice(engine);
1995 static void set_timeslice(struct intel_engine_cs *engine)
1997 unsigned long duration;
1999 if (!intel_engine_has_timeslices(engine))
2002 duration = active_timeslice(engine);
2003 ENGINE_TRACE(engine, "bump timeslicing, interval:%lu", duration);
2005 set_timer_ms(&engine->execlists.timer, duration);
2008 static void start_timeslice(struct intel_engine_cs *engine, int prio)
2010 struct intel_engine_execlists *execlists = &engine->execlists;
2011 unsigned long duration;
2013 if (!intel_engine_has_timeslices(engine))
2016 WRITE_ONCE(execlists->switch_priority_hint, prio);
2017 if (prio == INT_MIN)
2020 if (timer_pending(&execlists->timer))
2023 duration = timeslice(engine);
2024 ENGINE_TRACE(engine,
2025 "start timeslicing, prio:%d, interval:%lu",
2028 set_timer_ms(&execlists->timer, duration);
2031 static void record_preemption(struct intel_engine_execlists *execlists)
2033 (void)I915_SELFTEST_ONLY(execlists->preempt_hang.count++);
2036 static unsigned long active_preempt_timeout(struct intel_engine_cs *engine,
2037 const struct i915_request *rq)
2042 /* Force a fast reset for terminated contexts (ignoring sysfs!) */
2043 if (unlikely(intel_context_is_banned(rq->context)))
2046 return READ_ONCE(engine->props.preempt_timeout_ms);
2049 static void set_preempt_timeout(struct intel_engine_cs *engine,
2050 const struct i915_request *rq)
2052 if (!intel_engine_has_preempt_reset(engine))
2055 set_timer_ms(&engine->execlists.preempt,
2056 active_preempt_timeout(engine, rq));
2059 static inline void clear_ports(struct i915_request **ports, int count)
2061 memset_p((void **)ports, NULL, count);
2065 copy_ports(struct i915_request **dst, struct i915_request **src, int count)
2067 /* A memcpy_p() would be very useful here! */
2069 WRITE_ONCE(*dst++, *src++); /* avoid write tearing */
2072 static void execlists_dequeue(struct intel_engine_cs *engine)
2074 struct intel_engine_execlists * const execlists = &engine->execlists;
2075 struct i915_request **port = execlists->pending;
2076 struct i915_request ** const last_port = port + execlists->port_mask;
2077 struct i915_request * const *active;
2078 struct i915_request *last;
2080 bool submit = false;
2083 * Hardware submission is through 2 ports. Conceptually each port
2084 * has a (RING_START, RING_HEAD, RING_TAIL) tuple. RING_START is
2085 * static for a context, and unique to each, so we only execute
2086 * requests belonging to a single context from each ring. RING_HEAD
2087 * is maintained by the CS in the context image, it marks the place
2088 * where it got up to last time, and through RING_TAIL we tell the CS
2089 * where we want to execute up to this time.
2091 * In this list the requests are in order of execution. Consecutive
2092 * requests from the same context are adjacent in the ringbuffer. We
2093 * can combine these requests into a single RING_TAIL update:
2095 * RING_HEAD...req1...req2
2097 * since to execute req2 the CS must first execute req1.
2099 * Our goal then is to point each port to the end of a consecutive
2100 * sequence of requests as being the most optimal (fewest wake ups
2101 * and context switches) submission.
2104 for (rb = rb_first_cached(&execlists->virtual); rb; ) {
2105 struct virtual_engine *ve =
2106 rb_entry(rb, typeof(*ve), nodes[engine->id].rb);
2107 struct i915_request *rq = READ_ONCE(ve->request);
2109 if (!rq) { /* lazily cleanup after another engine handled rq */
2110 rb_erase_cached(rb, &execlists->virtual);
2112 rb = rb_first_cached(&execlists->virtual);
2116 if (!virtual_matches(ve, rq, engine)) {
2125 * If the queue is higher priority than the last
2126 * request in the currently active context, submit afresh.
2127 * We will resubmit again afterwards in case we need to split
2128 * the active context to interject the preemption request,
2129 * i.e. we will retrigger preemption following the ack in case
2132 active = READ_ONCE(execlists->active);
2135 * In theory we can skip over completed contexts that have not
2136 * yet been processed by events (as those events are in flight):
2138 * while ((last = *active) && i915_request_completed(last))
2141 * However, the GPU cannot handle this as it will ultimately
2142 * find itself trying to jump back into a context it has just
2143 * completed and barf.
2146 if ((last = *active)) {
2147 if (need_preempt(engine, last, rb)) {
2148 if (i915_request_completed(last)) {
2149 tasklet_hi_schedule(&execlists->tasklet);
2153 ENGINE_TRACE(engine,
2154 "preempting last=%llx:%lld, prio=%d, hint=%d\n",
2155 last->fence.context,
2157 last->sched.attr.priority,
2158 execlists->queue_priority_hint);
2159 record_preemption(execlists);
2162 * Don't let the RING_HEAD advance past the breadcrumb
2163 * as we unwind (and until we resubmit) so that we do
2164 * not accidentally tell it to go backwards.
2166 ring_set_paused(engine, 1);
2169 * Note that we have not stopped the GPU at this point,
2170 * so we are unwinding the incomplete requests as they
2171 * remain inflight and so by the time we do complete
2172 * the preemption, some of the unwound requests may
2175 __unwind_incomplete_requests(engine);
2178 } else if (need_timeslice(engine, last, rb) &&
2179 timeslice_expired(execlists, last)) {
2180 if (i915_request_completed(last)) {
2181 tasklet_hi_schedule(&execlists->tasklet);
2185 ENGINE_TRACE(engine,
2186 "expired last=%llx:%lld, prio=%d, hint=%d, yield?=%s\n",
2187 last->fence.context,
2189 last->sched.attr.priority,
2190 execlists->queue_priority_hint,
2191 yesno(timeslice_yield(execlists, last)));
2193 ring_set_paused(engine, 1);
2194 defer_active(engine);
2197 * Unlike for preemption, if we rewind and continue
2198 * executing the same context as previously active,
2199 * the order of execution will remain the same and
2200 * the tail will only advance. We do not need to
2201 * force a full context restore, as a lite-restore
2202 * is sufficient to resample the monotonic TAIL.
2204 * If we switch to any other context, similarly we
2205 * will not rewind TAIL of current context, and
2206 * normal save/restore will preserve state and allow
2207 * us to later continue executing the same request.
2212 * Otherwise if we already have a request pending
2213 * for execution after the current one, we can
2214 * just wait until the next CS event before
2215 * queuing more. In either case we will force a
2216 * lite-restore preemption event, but if we wait
2217 * we hopefully coalesce several updates into a single
2220 if (!list_is_last(&last->sched.link,
2221 &engine->active.requests)) {
2223 * Even if ELSP[1] is occupied and not worthy
2224 * of timeslices, our queue might be.
2226 start_timeslice(engine, queue_prio(execlists));
2232 while (rb) { /* XXX virtual is always taking precedence */
2233 struct virtual_engine *ve =
2234 rb_entry(rb, typeof(*ve), nodes[engine->id].rb);
2235 struct i915_request *rq;
2237 spin_lock(&ve->base.active.lock);
2240 if (unlikely(!rq)) { /* lost the race to a sibling */
2241 spin_unlock(&ve->base.active.lock);
2242 rb_erase_cached(rb, &execlists->virtual);
2244 rb = rb_first_cached(&execlists->virtual);
2248 GEM_BUG_ON(rq != ve->request);
2249 GEM_BUG_ON(rq->engine != &ve->base);
2250 GEM_BUG_ON(rq->context != &ve->context);
2252 if (rq_prio(rq) >= queue_prio(execlists)) {
2253 if (!virtual_matches(ve, rq, engine)) {
2254 spin_unlock(&ve->base.active.lock);
2259 if (last && !can_merge_rq(last, rq)) {
2260 spin_unlock(&ve->base.active.lock);
2261 start_timeslice(engine, rq_prio(rq));
2262 return; /* leave this for another sibling */
2265 ENGINE_TRACE(engine,
2266 "virtual rq=%llx:%lld%s, new engine? %s\n",
2269 i915_request_completed(rq) ? "!" :
2270 i915_request_started(rq) ? "*" :
2272 yesno(engine != ve->siblings[0]));
2274 WRITE_ONCE(ve->request, NULL);
2275 WRITE_ONCE(ve->base.execlists.queue_priority_hint,
2277 rb_erase_cached(rb, &execlists->virtual);
2280 GEM_BUG_ON(!(rq->execution_mask & engine->mask));
2281 WRITE_ONCE(rq->engine, engine);
2283 if (__i915_request_submit(rq)) {
2285 * Only after we confirm that we will submit
2286 * this request (i.e. it has not already
2287 * completed), do we want to update the context.
2289 * This serves two purposes. It avoids
2290 * unnecessary work if we are resubmitting an
2291 * already completed request after timeslicing.
2292 * But more importantly, it prevents us altering
2293 * ve->siblings[] on an idle context, where
2294 * we may be using ve->siblings[] in
2295 * virtual_context_enter / virtual_context_exit.
2297 virtual_xfer_context(ve, engine);
2298 GEM_BUG_ON(ve->siblings[0] != engine);
2303 i915_request_put(rq);
2306 * Hmm, we have a bunch of virtual engine requests,
2307 * but the first one was already completed (thanks
2308 * preempt-to-busy!). Keep looking at the veng queue
2309 * until we have no more relevant requests (i.e.
2310 * the normal submit queue has higher priority).
2313 spin_unlock(&ve->base.active.lock);
2314 rb = rb_first_cached(&execlists->virtual);
2319 spin_unlock(&ve->base.active.lock);
2323 while ((rb = rb_first_cached(&execlists->queue))) {
2324 struct i915_priolist *p = to_priolist(rb);
2325 struct i915_request *rq, *rn;
2328 priolist_for_each_request_consume(rq, rn, p, i) {
2332 * Can we combine this request with the current port?
2333 * It has to be the same context/ringbuffer and not
2334 * have any exceptions (e.g. GVT saying never to
2335 * combine contexts).
2337 * If we can combine the requests, we can execute both
2338 * by updating the RING_TAIL to point to the end of the
2339 * second request, and so we never need to tell the
2340 * hardware about the first.
2342 if (last && !can_merge_rq(last, rq)) {
2344 * If we are on the second port and cannot
2345 * combine this request with the last, then we
2348 if (port == last_port)
2352 * We must not populate both ELSP[] with the
2353 * same LRCA, i.e. we must submit 2 different
2354 * contexts if we submit 2 ELSP.
2356 if (last->context == rq->context)
2359 if (i915_request_has_sentinel(last))
2363 * If GVT overrides us we only ever submit
2364 * port[0], leaving port[1] empty. Note that we
2365 * also have to be careful that we don't queue
2366 * the same context (even though a different
2367 * request) to the second port.
2369 if (ctx_single_port_submission(last->context) ||
2370 ctx_single_port_submission(rq->context))
2376 if (__i915_request_submit(rq)) {
2378 *port = execlists_schedule_in(last, port - execlists->pending);
2384 !can_merge_ctx(last->context,
2387 i915_seqno_passed(last->fence.seqno,
2395 rb_erase_cached(&p->node, &execlists->queue);
2396 i915_priolist_free(p);
2401 * Here be a bit of magic! Or sleight-of-hand, whichever you prefer.
2403 * We choose the priority hint such that if we add a request of greater
2404 * priority than this, we kick the submission tasklet to decide on
2405 * the right order of submitting the requests to hardware. We must
2406 * also be prepared to reorder requests as they are in-flight on the
2407 * HW. We derive the priority hint then as the first "hole" in
2408 * the HW submission ports and if there are no available slots,
2409 * the priority of the lowest executing request, i.e. last.
2411 * When we do receive a higher priority request ready to run from the
2412 * user, see queue_request(), the priority hint is bumped to that
2413 * request triggering preemption on the next dequeue (or subsequent
2414 * interrupt for secondary ports).
2416 execlists->queue_priority_hint = queue_prio(execlists);
2419 *port = execlists_schedule_in(last, port - execlists->pending);
2420 execlists->switch_priority_hint =
2421 switch_prio(engine, *execlists->pending);
2424 * Skip if we ended up with exactly the same set of requests,
2425 * e.g. trying to timeslice a pair of ordered contexts
2427 if (!memcmp(active, execlists->pending,
2428 (port - execlists->pending + 1) * sizeof(*port))) {
2430 execlists_schedule_out(fetch_and_zero(port));
2431 while (port-- != execlists->pending);
2435 clear_ports(port + 1, last_port - port);
2437 WRITE_ONCE(execlists->yield, -1);
2438 set_preempt_timeout(engine, *active);
2439 execlists_submit_ports(engine);
2441 start_timeslice(engine, execlists->queue_priority_hint);
2443 ring_set_paused(engine, 0);
2448 cancel_port_requests(struct intel_engine_execlists * const execlists)
2450 struct i915_request * const *port;
2452 for (port = execlists->pending; *port; port++)
2453 execlists_schedule_out(*port);
2454 clear_ports(execlists->pending, ARRAY_SIZE(execlists->pending));
2456 /* Mark the end of active before we overwrite *active */
2457 for (port = xchg(&execlists->active, execlists->pending); *port; port++)
2458 execlists_schedule_out(*port);
2459 clear_ports(execlists->inflight, ARRAY_SIZE(execlists->inflight));
2461 smp_wmb(); /* complete the seqlock for execlists_active() */
2462 WRITE_ONCE(execlists->active, execlists->inflight);
2466 invalidate_csb_entries(const u64 *first, const u64 *last)
2468 clflush((void *)first);
2469 clflush((void *)last);
2473 * Starting with Gen12, the status has a new format:
2475 * bit 0: switched to new queue
2477 * bit 2: semaphore wait mode (poll or signal), only valid when
2478 * switch detail is set to "wait on semaphore"
2479 * bits 3-5: engine class
2480 * bits 6-11: engine instance
2481 * bits 12-14: reserved
2482 * bits 15-25: sw context id of the lrc the GT switched to
2483 * bits 26-31: sw counter of the lrc the GT switched to
2484 * bits 32-35: context switch detail
2486 * - 1: wait on sync flip
2487 * - 2: wait on vblank
2488 * - 3: wait on scanline
2489 * - 4: wait on semaphore
2490 * - 5: context preempted (not on SEMAPHORE_WAIT or
2493 * bits 37-43: wait detail (for switch detail 1 to 4)
2494 * bits 44-46: reserved
2495 * bits 47-57: sw context id of the lrc the GT switched away from
2496 * bits 58-63: sw counter of the lrc the GT switched away from
2498 static inline bool gen12_csb_parse(const u64 *csb)
2500 bool ctx_away_valid;
2504 /* HSD#22011248461 */
2505 entry = READ_ONCE(*csb);
2506 if (unlikely(entry == -1)) {
2508 if (wait_for_atomic_us((entry = READ_ONCE(*csb)) != -1, 50))
2509 GEM_WARN_ON("50us CSB timeout");
2512 WRITE_ONCE(*(u64 *)csb, -1);
2514 ctx_away_valid = GEN12_CSB_CTX_VALID(upper_32_bits(entry));
2516 lower_32_bits(entry) & GEN12_CTX_STATUS_SWITCHED_TO_NEW_QUEUE;
2519 * The context switch detail is not guaranteed to be 5 when a preemption
2520 * occurs, so we can't just check for that. The check below works for
2521 * all the cases we care about, including preemptions of WAIT
2522 * instructions and lite-restore. Preempt-to-idle via the CTRL register
2523 * would require some extra handling, but we don't support that.
2525 if (!ctx_away_valid || new_queue) {
2526 GEM_BUG_ON(!GEN12_CSB_CTX_VALID(lower_32_bits(entry)));
2531 * switch detail = 5 is covered by the case above and we do not expect a
2532 * context switch on an unsuccessful wait instruction since we always
2535 GEM_BUG_ON(GEN12_CTX_SWITCH_DETAIL(upper_32_bits(entry)));
2539 static inline bool gen8_csb_parse(const u64 *csb)
2541 return *csb & (GEN8_CTX_STATUS_IDLE_ACTIVE | GEN8_CTX_STATUS_PREEMPTED);
2544 static void process_csb(struct intel_engine_cs *engine)
2546 struct intel_engine_execlists * const execlists = &engine->execlists;
2547 const u64 * const buf = execlists->csb_status;
2548 const u8 num_entries = execlists->csb_size;
2552 * As we modify our execlists state tracking we require exclusive
2553 * access. Either we are inside the tasklet, or the tasklet is disabled
2554 * and we assume that is only inside the reset paths and so serialised.
2556 GEM_BUG_ON(!tasklet_is_locked(&execlists->tasklet) &&
2557 !reset_in_progress(execlists));
2558 GEM_BUG_ON(!intel_engine_in_execlists_submission_mode(engine));
2561 * Note that csb_write, csb_status may be either in HWSP or mmio.
2562 * When reading from the csb_write mmio register, we have to be
2563 * careful to only use the GEN8_CSB_WRITE_PTR portion, which is
2564 * the low 4bits. As it happens we know the next 4bits are always
2565 * zero and so we can simply masked off the low u8 of the register
2566 * and treat it identically to reading from the HWSP (without having
2567 * to use explicit shifting and masking, and probably bifurcating
2568 * the code to handle the legacy mmio read).
2570 head = execlists->csb_head;
2571 tail = READ_ONCE(*execlists->csb_write);
2572 if (unlikely(head == tail))
2576 * We will consume all events from HW, or at least pretend to.
2578 * The sequence of events from the HW is deterministic, and derived
2579 * from our writes to the ELSP, with a smidgen of variability for
2580 * the arrival of the asynchronous requests wrt to the inflight
2581 * execution. If the HW sends an event that does not correspond with
2582 * the one we are expecting, we have to abandon all hope as we lose
2583 * all tracking of what the engine is actually executing. We will
2584 * only detect we are out of sequence with the HW when we get an
2585 * 'impossible' event because we have already drained our own
2586 * preemption/promotion queue. If this occurs, we know that we likely
2587 * lost track of execution earlier and must unwind and restart, the
2588 * simplest way is by stop processing the event queue and force the
2591 execlists->csb_head = tail;
2592 ENGINE_TRACE(engine, "cs-irq head=%d, tail=%d\n", head, tail);
2595 * Hopefully paired with a wmb() in HW!
2597 * We must complete the read of the write pointer before any reads
2598 * from the CSB, so that we do not see stale values. Without an rmb
2599 * (lfence) the HW may speculatively perform the CSB[] reads *before*
2600 * we perform the READ_ONCE(*csb_write).
2606 if (++head == num_entries)
2610 * We are flying near dragons again.
2612 * We hold a reference to the request in execlist_port[]
2613 * but no more than that. We are operating in softirq
2614 * context and so cannot hold any mutex or sleep. That
2615 * prevents us stopping the requests we are processing
2616 * in port[] from being retired simultaneously (the
2617 * breadcrumb will be complete before we see the
2618 * context-switch). As we only hold the reference to the
2619 * request, any pointer chasing underneath the request
2620 * is subject to a potential use-after-free. Thus we
2621 * store all of the bookkeeping within port[] as
2622 * required, and avoid using unguarded pointers beneath
2623 * request itself. The same applies to the atomic
2627 ENGINE_TRACE(engine, "csb[%d]: status=0x%08x:0x%08x\n",
2629 upper_32_bits(buf[head]),
2630 lower_32_bits(buf[head]));
2632 if (INTEL_GEN(engine->i915) >= 12)
2633 promote = gen12_csb_parse(buf + head);
2635 promote = gen8_csb_parse(buf + head);
2637 struct i915_request * const *old = execlists->active;
2639 if (GEM_WARN_ON(!*execlists->pending)) {
2640 execlists->error_interrupt |= ERROR_CSB;
2644 ring_set_paused(engine, 0);
2646 /* Point active to the new ELSP; prevent overwriting */
2647 WRITE_ONCE(execlists->active, execlists->pending);
2648 smp_wmb(); /* notify execlists_active() */
2650 /* cancel old inflight, prepare for switch */
2651 trace_ports(execlists, "preempted", old);
2653 execlists_schedule_out(*old++);
2655 /* switch pending to inflight */
2656 GEM_BUG_ON(!assert_pending_valid(execlists, "promote"));
2657 copy_ports(execlists->inflight,
2659 execlists_num_ports(execlists));
2660 smp_wmb(); /* complete the seqlock */
2661 WRITE_ONCE(execlists->active, execlists->inflight);
2663 /* XXX Magic delay for tgl */
2664 ENGINE_POSTING_READ(engine, RING_CONTEXT_STATUS_PTR);
2666 WRITE_ONCE(execlists->pending[0], NULL);
2668 if (GEM_WARN_ON(!*execlists->active)) {
2669 execlists->error_interrupt |= ERROR_CSB;
2673 /* port0 completed, advanced to port1 */
2674 trace_ports(execlists, "completed", execlists->active);
2677 * We rely on the hardware being strongly
2678 * ordered, that the breadcrumb write is
2679 * coherent (visible from the CPU) before the
2680 * user interrupt is processed. One might assume
2681 * that the breadcrumb write being before the
2682 * user interrupt and the CS event for the context
2683 * switch would therefore be before the CS event
2686 if (GEM_SHOW_DEBUG() &&
2687 !i915_request_completed(*execlists->active)) {
2688 struct i915_request *rq = *execlists->active;
2689 const u32 *regs __maybe_unused =
2690 rq->context->lrc_reg_state;
2692 ENGINE_TRACE(engine,
2693 "context completed before request!\n");
2694 ENGINE_TRACE(engine,
2695 "ring:{start:0x%08x, head:%04x, tail:%04x, ctl:%08x, mode:%08x}\n",
2696 ENGINE_READ(engine, RING_START),
2697 ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR,
2698 ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR,
2699 ENGINE_READ(engine, RING_CTL),
2700 ENGINE_READ(engine, RING_MI_MODE));
2701 ENGINE_TRACE(engine,
2702 "rq:{start:%08x, head:%04x, tail:%04x, seqno:%llx:%d, hwsp:%d}, ",
2703 i915_ggtt_offset(rq->ring->vma),
2706 lower_32_bits(rq->fence.seqno),
2708 ENGINE_TRACE(engine,
2709 "ctx:{start:%08x, head:%04x, tail:%04x}, ",
2710 regs[CTX_RING_START],
2711 regs[CTX_RING_HEAD],
2712 regs[CTX_RING_TAIL]);
2715 execlists_schedule_out(*execlists->active++);
2717 GEM_BUG_ON(execlists->active - execlists->inflight >
2718 execlists_num_ports(execlists));
2720 } while (head != tail);
2722 set_timeslice(engine);
2725 * Gen11 has proven to fail wrt global observation point between
2726 * entry and tail update, failing on the ordering and thus
2727 * we see an old entry in the context status buffer.
2729 * Forcibly evict out entries for the next gpu csb update,
2730 * to increase the odds that we get a fresh entries with non
2731 * working hardware. The cost for doing so comes out mostly with
2732 * the wash as hardware, working or not, will need to do the
2733 * invalidation before.
2735 invalidate_csb_entries(&buf[0], &buf[num_entries - 1]);
2738 static void __execlists_submission_tasklet(struct intel_engine_cs *const engine)
2740 lockdep_assert_held(&engine->active.lock);
2741 if (!READ_ONCE(engine->execlists.pending[0])) {
2742 rcu_read_lock(); /* protect peeking at execlists->active */
2743 execlists_dequeue(engine);
2748 static void __execlists_hold(struct i915_request *rq)
2753 struct i915_dependency *p;
2755 if (i915_request_is_active(rq))
2756 __i915_request_unsubmit(rq);
2758 clear_bit(I915_FENCE_FLAG_PQUEUE, &rq->fence.flags);
2759 list_move_tail(&rq->sched.link, &rq->engine->active.hold);
2760 i915_request_set_hold(rq);
2761 RQ_TRACE(rq, "on hold\n");
2763 for_each_waiter(p, rq) {
2764 struct i915_request *w =
2765 container_of(p->waiter, typeof(*w), sched);
2767 /* Leave semaphores spinning on the other engines */
2768 if (w->engine != rq->engine)
2771 if (!i915_request_is_ready(w))
2774 if (i915_request_completed(w))
2777 if (i915_request_on_hold(w))
2780 list_move_tail(&w->sched.link, &list);
2783 rq = list_first_entry_or_null(&list, typeof(*rq), sched.link);
2787 static bool execlists_hold(struct intel_engine_cs *engine,
2788 struct i915_request *rq)
2790 spin_lock_irq(&engine->active.lock);
2792 if (i915_request_completed(rq)) { /* too late! */
2797 if (rq->engine != engine) { /* preempted virtual engine */
2798 struct virtual_engine *ve = to_virtual_engine(rq->engine);
2801 * intel_context_inflight() is only protected by virtue
2802 * of process_csb() being called only by the tasklet (or
2803 * directly from inside reset while the tasklet is suspended).
2804 * Assert that neither of those are allowed to run while we
2805 * poke at the request queues.
2807 GEM_BUG_ON(!reset_in_progress(&engine->execlists));
2810 * An unsubmitted request along a virtual engine will
2811 * remain on the active (this) engine until we are able
2812 * to process the context switch away (and so mark the
2813 * context as no longer in flight). That cannot have happened
2814 * yet, otherwise we would not be hanging!
2816 spin_lock(&ve->base.active.lock);
2817 GEM_BUG_ON(intel_context_inflight(rq->context) != engine);
2818 GEM_BUG_ON(ve->request != rq);
2820 spin_unlock(&ve->base.active.lock);
2821 i915_request_put(rq);
2823 rq->engine = engine;
2827 * Transfer this request onto the hold queue to prevent it
2828 * being resumbitted to HW (and potentially completed) before we have
2829 * released it. Since we may have already submitted following
2830 * requests, we need to remove those as well.
2832 GEM_BUG_ON(i915_request_on_hold(rq));
2833 GEM_BUG_ON(rq->engine != engine);
2834 __execlists_hold(rq);
2835 GEM_BUG_ON(list_empty(&engine->active.hold));
2838 spin_unlock_irq(&engine->active.lock);
2842 static bool hold_request(const struct i915_request *rq)
2844 struct i915_dependency *p;
2845 bool result = false;
2848 * If one of our ancestors is on hold, we must also be on hold,
2849 * otherwise we will bypass it and execute before it.
2852 for_each_signaler(p, rq) {
2853 const struct i915_request *s =
2854 container_of(p->signaler, typeof(*s), sched);
2856 if (s->engine != rq->engine)
2859 result = i915_request_on_hold(s);
2868 static void __execlists_unhold(struct i915_request *rq)
2873 struct i915_dependency *p;
2875 RQ_TRACE(rq, "hold release\n");
2877 GEM_BUG_ON(!i915_request_on_hold(rq));
2878 GEM_BUG_ON(!i915_sw_fence_signaled(&rq->submit));
2880 i915_request_clear_hold(rq);
2881 list_move_tail(&rq->sched.link,
2882 i915_sched_lookup_priolist(rq->engine,
2884 set_bit(I915_FENCE_FLAG_PQUEUE, &rq->fence.flags);
2886 /* Also release any children on this engine that are ready */
2887 for_each_waiter(p, rq) {
2888 struct i915_request *w =
2889 container_of(p->waiter, typeof(*w), sched);
2891 /* Propagate any change in error status */
2892 if (rq->fence.error)
2893 i915_request_set_error_once(w, rq->fence.error);
2895 if (w->engine != rq->engine)
2898 if (!i915_request_on_hold(w))
2901 /* Check that no other parents are also on hold */
2902 if (hold_request(w))
2905 list_move_tail(&w->sched.link, &list);
2908 rq = list_first_entry_or_null(&list, typeof(*rq), sched.link);
2912 static void execlists_unhold(struct intel_engine_cs *engine,
2913 struct i915_request *rq)
2915 spin_lock_irq(&engine->active.lock);
2918 * Move this request back to the priority queue, and all of its
2919 * children and grandchildren that were suspended along with it.
2921 __execlists_unhold(rq);
2923 if (rq_prio(rq) > engine->execlists.queue_priority_hint) {
2924 engine->execlists.queue_priority_hint = rq_prio(rq);
2925 tasklet_hi_schedule(&engine->execlists.tasklet);
2928 spin_unlock_irq(&engine->active.lock);
2931 struct execlists_capture {
2932 struct work_struct work;
2933 struct i915_request *rq;
2934 struct i915_gpu_coredump *error;
2937 static void execlists_capture_work(struct work_struct *work)
2939 struct execlists_capture *cap = container_of(work, typeof(*cap), work);
2940 const gfp_t gfp = GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN;
2941 struct intel_engine_cs *engine = cap->rq->engine;
2942 struct intel_gt_coredump *gt = cap->error->gt;
2943 struct intel_engine_capture_vma *vma;
2945 /* Compress all the objects attached to the request, slow! */
2946 vma = intel_engine_coredump_add_request(gt->engine, cap->rq, gfp);
2948 struct i915_vma_compress *compress =
2949 i915_vma_capture_prepare(gt);
2951 intel_engine_coredump_add_vma(gt->engine, vma, compress);
2952 i915_vma_capture_finish(gt, compress);
2955 gt->simulated = gt->engine->simulated;
2956 cap->error->simulated = gt->simulated;
2958 /* Publish the error state, and announce it to the world */
2959 i915_error_state_store(cap->error);
2960 i915_gpu_coredump_put(cap->error);
2962 /* Return this request and all that depend upon it for signaling */
2963 execlists_unhold(engine, cap->rq);
2964 i915_request_put(cap->rq);
2969 static struct execlists_capture *capture_regs(struct intel_engine_cs *engine)
2971 const gfp_t gfp = GFP_ATOMIC | __GFP_NOWARN;
2972 struct execlists_capture *cap;
2974 cap = kmalloc(sizeof(*cap), gfp);
2978 cap->error = i915_gpu_coredump_alloc(engine->i915, gfp);
2982 cap->error->gt = intel_gt_coredump_alloc(engine->gt, gfp);
2983 if (!cap->error->gt)
2986 cap->error->gt->engine = intel_engine_coredump_alloc(engine, gfp);
2987 if (!cap->error->gt->engine)
2993 kfree(cap->error->gt);
3001 static struct i915_request *
3002 active_context(struct intel_engine_cs *engine, u32 ccid)
3004 const struct intel_engine_execlists * const el = &engine->execlists;
3005 struct i915_request * const *port, *rq;
3008 * Use the most recent result from process_csb(), but just in case
3009 * we trigger an error (via interrupt) before the first CS event has
3010 * been written, peek at the next submission.
3013 for (port = el->active; (rq = *port); port++) {
3014 if (rq->context->lrc.ccid == ccid) {
3015 ENGINE_TRACE(engine,
3016 "ccid found at active:%zd\n",
3022 for (port = el->pending; (rq = *port); port++) {
3023 if (rq->context->lrc.ccid == ccid) {
3024 ENGINE_TRACE(engine,
3025 "ccid found at pending:%zd\n",
3026 port - el->pending);
3031 ENGINE_TRACE(engine, "ccid:%x not found\n", ccid);
3035 static u32 active_ccid(struct intel_engine_cs *engine)
3037 return ENGINE_READ_FW(engine, RING_EXECLIST_STATUS_HI);
3040 static void execlists_capture(struct intel_engine_cs *engine)
3042 struct execlists_capture *cap;
3044 if (!IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR))
3048 * We need to _quickly_ capture the engine state before we reset.
3049 * We are inside an atomic section (softirq) here and we are delaying
3050 * the forced preemption event.
3052 cap = capture_regs(engine);
3056 spin_lock_irq(&engine->active.lock);
3057 cap->rq = active_context(engine, active_ccid(engine));
3059 cap->rq = active_request(cap->rq->context->timeline, cap->rq);
3060 cap->rq = i915_request_get_rcu(cap->rq);
3062 spin_unlock_irq(&engine->active.lock);
3067 * Remove the request from the execlists queue, and take ownership
3068 * of the request. We pass it to our worker who will _slowly_ compress
3069 * all the pages the _user_ requested for debugging their batch, after
3070 * which we return it to the queue for signaling.
3072 * By removing them from the execlists queue, we also remove the
3073 * requests from being processed by __unwind_incomplete_requests()
3074 * during the intel_engine_reset(), and so they will *not* be replayed
3077 * Note that because we have not yet reset the engine at this point,
3078 * it is possible for the request that we have identified as being
3079 * guilty, did in fact complete and we will then hit an arbitration
3080 * point allowing the outstanding preemption to succeed. The likelihood
3081 * of that is very low (as capturing of the engine registers should be
3082 * fast enough to run inside an irq-off atomic section!), so we will
3083 * simply hold that request accountable for being non-preemptible
3084 * long enough to force the reset.
3086 if (!execlists_hold(engine, cap->rq))
3089 INIT_WORK(&cap->work, execlists_capture_work);
3090 schedule_work(&cap->work);
3094 i915_request_put(cap->rq);
3096 i915_gpu_coredump_put(cap->error);
3100 static void execlists_reset(struct intel_engine_cs *engine, const char *msg)
3102 const unsigned int bit = I915_RESET_ENGINE + engine->id;
3103 unsigned long *lock = &engine->gt->reset.flags;
3105 if (!intel_has_reset_engine(engine->gt))
3108 if (test_and_set_bit(bit, lock))
3111 ENGINE_TRACE(engine, "reset for %s\n", msg);
3113 /* Mark this tasklet as disabled to avoid waiting for it to complete */
3114 tasklet_disable_nosync(&engine->execlists.tasklet);
3116 ring_set_paused(engine, 1); /* Freeze the current request in place */
3117 execlists_capture(engine);
3118 intel_engine_reset(engine, msg);
3120 tasklet_enable(&engine->execlists.tasklet);
3121 clear_and_wake_up_bit(bit, lock);
3124 static bool preempt_timeout(const struct intel_engine_cs *const engine)
3126 const struct timer_list *t = &engine->execlists.preempt;
3128 if (!CONFIG_DRM_I915_PREEMPT_TIMEOUT)
3131 if (!timer_expired(t))
3134 return READ_ONCE(engine->execlists.pending[0]);
3138 * Check the unread Context Status Buffers and manage the submission of new
3139 * contexts to the ELSP accordingly.
3141 static void execlists_submission_tasklet(unsigned long data)
3143 struct intel_engine_cs * const engine = (struct intel_engine_cs *)data;
3144 bool timeout = preempt_timeout(engine);
3146 process_csb(engine);
3148 if (unlikely(READ_ONCE(engine->execlists.error_interrupt))) {
3151 /* Generate the error message in priority wrt to the user! */
3152 if (engine->execlists.error_interrupt & GENMASK(15, 0))
3153 msg = "CS error"; /* thrown by a user payload */
3154 else if (engine->execlists.error_interrupt & ERROR_CSB)
3155 msg = "invalid CSB event";
3157 msg = "internal error";
3159 engine->execlists.error_interrupt = 0;
3160 execlists_reset(engine, msg);
3163 if (!READ_ONCE(engine->execlists.pending[0]) || timeout) {
3164 unsigned long flags;
3166 spin_lock_irqsave(&engine->active.lock, flags);
3167 __execlists_submission_tasklet(engine);
3168 spin_unlock_irqrestore(&engine->active.lock, flags);
3170 /* Recheck after serialising with direct-submission */
3171 if (unlikely(timeout && preempt_timeout(engine)))
3172 execlists_reset(engine, "preemption time out");
3176 static void __execlists_kick(struct intel_engine_execlists *execlists)
3178 /* Kick the tasklet for some interrupt coalescing and reset handling */
3179 tasklet_hi_schedule(&execlists->tasklet);
3182 #define execlists_kick(t, member) \
3183 __execlists_kick(container_of(t, struct intel_engine_execlists, member))
3185 static void execlists_timeslice(struct timer_list *timer)
3187 execlists_kick(timer, timer);
3190 static void execlists_preempt(struct timer_list *timer)
3192 execlists_kick(timer, preempt);
3195 static void queue_request(struct intel_engine_cs *engine,
3196 struct i915_request *rq)
3198 GEM_BUG_ON(!list_empty(&rq->sched.link));
3199 list_add_tail(&rq->sched.link,
3200 i915_sched_lookup_priolist(engine, rq_prio(rq)));
3201 set_bit(I915_FENCE_FLAG_PQUEUE, &rq->fence.flags);
3204 static void __submit_queue_imm(struct intel_engine_cs *engine)
3206 struct intel_engine_execlists * const execlists = &engine->execlists;
3208 if (reset_in_progress(execlists))
3209 return; /* defer until we restart the engine following reset */
3211 __execlists_submission_tasklet(engine);
3214 static void submit_queue(struct intel_engine_cs *engine,
3215 const struct i915_request *rq)
3217 struct intel_engine_execlists *execlists = &engine->execlists;
3219 if (rq_prio(rq) <= execlists->queue_priority_hint)
3222 execlists->queue_priority_hint = rq_prio(rq);
3223 __submit_queue_imm(engine);
3226 static bool ancestor_on_hold(const struct intel_engine_cs *engine,
3227 const struct i915_request *rq)
3229 GEM_BUG_ON(i915_request_on_hold(rq));
3230 return !list_empty(&engine->active.hold) && hold_request(rq);
3233 static void flush_csb(struct intel_engine_cs *engine)
3235 struct intel_engine_execlists *el = &engine->execlists;
3237 if (READ_ONCE(el->pending[0]) && tasklet_trylock(&el->tasklet)) {
3238 if (!reset_in_progress(el))
3239 process_csb(engine);
3240 tasklet_unlock(&el->tasklet);
3244 static void execlists_submit_request(struct i915_request *request)
3246 struct intel_engine_cs *engine = request->engine;
3247 unsigned long flags;
3249 /* Hopefully we clear execlists->pending[] to let us through */
3252 /* Will be called from irq-context when using foreign fences. */
3253 spin_lock_irqsave(&engine->active.lock, flags);
3255 if (unlikely(ancestor_on_hold(engine, request))) {
3256 RQ_TRACE(request, "ancestor on hold\n");
3257 list_add_tail(&request->sched.link, &engine->active.hold);
3258 i915_request_set_hold(request);
3260 queue_request(engine, request);
3262 GEM_BUG_ON(RB_EMPTY_ROOT(&engine->execlists.queue.rb_root));
3263 GEM_BUG_ON(list_empty(&request->sched.link));
3265 submit_queue(engine, request);
3268 spin_unlock_irqrestore(&engine->active.lock, flags);
3271 static void __execlists_context_fini(struct intel_context *ce)
3273 intel_ring_put(ce->ring);
3274 i915_vma_put(ce->state);
3277 static void execlists_context_destroy(struct kref *kref)
3279 struct intel_context *ce = container_of(kref, typeof(*ce), ref);
3281 GEM_BUG_ON(!i915_active_is_idle(&ce->active));
3282 GEM_BUG_ON(intel_context_is_pinned(ce));
3285 __execlists_context_fini(ce);
3287 intel_context_fini(ce);
3288 intel_context_free(ce);
3292 set_redzone(void *vaddr, const struct intel_engine_cs *engine)
3294 if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
3297 vaddr += engine->context_size;
3299 memset(vaddr, CONTEXT_REDZONE, I915_GTT_PAGE_SIZE);
3303 check_redzone(const void *vaddr, const struct intel_engine_cs *engine)
3305 if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
3308 vaddr += engine->context_size;
3310 if (memchr_inv(vaddr, CONTEXT_REDZONE, I915_GTT_PAGE_SIZE))
3311 drm_err_once(&engine->i915->drm,
3312 "%s context redzone overwritten!\n",
3316 static void execlists_context_unpin(struct intel_context *ce)
3318 check_redzone((void *)ce->lrc_reg_state - LRC_STATE_OFFSET,
3322 static void execlists_context_post_unpin(struct intel_context *ce)
3324 i915_gem_object_unpin_map(ce->state->obj);
3328 gen12_emit_timestamp_wa(const struct intel_context *ce, u32 *cs)
3330 *cs++ = MI_LOAD_REGISTER_MEM_GEN8 |
3331 MI_SRM_LRM_GLOBAL_GTT |
3333 *cs++ = i915_mmio_reg_offset(GEN8_RING_CS_GPR(0, 0));
3334 *cs++ = i915_ggtt_offset(ce->state) + LRC_STATE_OFFSET +
3335 CTX_TIMESTAMP * sizeof(u32);
3338 *cs++ = MI_LOAD_REGISTER_REG |
3339 MI_LRR_SOURCE_CS_MMIO |
3341 *cs++ = i915_mmio_reg_offset(GEN8_RING_CS_GPR(0, 0));
3342 *cs++ = i915_mmio_reg_offset(RING_CTX_TIMESTAMP(0));
3344 *cs++ = MI_LOAD_REGISTER_REG |
3345 MI_LRR_SOURCE_CS_MMIO |
3347 *cs++ = i915_mmio_reg_offset(GEN8_RING_CS_GPR(0, 0));
3348 *cs++ = i915_mmio_reg_offset(RING_CTX_TIMESTAMP(0));
3354 gen12_emit_restore_scratch(const struct intel_context *ce, u32 *cs)
3356 GEM_BUG_ON(lrc_ring_gpr0(ce->engine) == -1);
3358 *cs++ = MI_LOAD_REGISTER_MEM_GEN8 |
3359 MI_SRM_LRM_GLOBAL_GTT |
3361 *cs++ = i915_mmio_reg_offset(GEN8_RING_CS_GPR(0, 0));
3362 *cs++ = i915_ggtt_offset(ce->state) + LRC_STATE_OFFSET +
3363 (lrc_ring_gpr0(ce->engine) + 1) * sizeof(u32);
3370 gen12_emit_cmd_buf_wa(const struct intel_context *ce, u32 *cs)
3372 GEM_BUG_ON(lrc_ring_cmd_buf_cctl(ce->engine) == -1);
3374 *cs++ = MI_LOAD_REGISTER_MEM_GEN8 |
3375 MI_SRM_LRM_GLOBAL_GTT |
3377 *cs++ = i915_mmio_reg_offset(GEN8_RING_CS_GPR(0, 0));
3378 *cs++ = i915_ggtt_offset(ce->state) + LRC_STATE_OFFSET +
3379 (lrc_ring_cmd_buf_cctl(ce->engine) + 1) * sizeof(u32);
3382 *cs++ = MI_LOAD_REGISTER_REG |
3383 MI_LRR_SOURCE_CS_MMIO |
3385 *cs++ = i915_mmio_reg_offset(GEN8_RING_CS_GPR(0, 0));
3386 *cs++ = i915_mmio_reg_offset(RING_CMD_BUF_CCTL(0));
3392 gen12_emit_indirect_ctx_rcs(const struct intel_context *ce, u32 *cs)
3394 cs = gen12_emit_timestamp_wa(ce, cs);
3395 cs = gen12_emit_cmd_buf_wa(ce, cs);
3396 cs = gen12_emit_restore_scratch(ce, cs);
3402 gen12_emit_indirect_ctx_xcs(const struct intel_context *ce, u32 *cs)
3404 cs = gen12_emit_timestamp_wa(ce, cs);
3405 cs = gen12_emit_restore_scratch(ce, cs);
3410 static inline u32 context_wa_bb_offset(const struct intel_context *ce)
3412 return PAGE_SIZE * ce->wa_bb_page;
3415 static u32 *context_indirect_bb(const struct intel_context *ce)
3419 GEM_BUG_ON(!ce->wa_bb_page);
3421 ptr = ce->lrc_reg_state;
3422 ptr -= LRC_STATE_OFFSET; /* back to start of context image */
3423 ptr += context_wa_bb_offset(ce);
3429 setup_indirect_ctx_bb(const struct intel_context *ce,
3430 const struct intel_engine_cs *engine,
3431 u32 *(*emit)(const struct intel_context *, u32 *))
3433 u32 * const start = context_indirect_bb(ce);
3436 cs = emit(ce, start);
3437 GEM_BUG_ON(cs - start > I915_GTT_PAGE_SIZE / sizeof(*cs));
3438 while ((unsigned long)cs % CACHELINE_BYTES)
3441 lrc_ring_setup_indirect_ctx(ce->lrc_reg_state, engine,
3442 i915_ggtt_offset(ce->state) +
3443 context_wa_bb_offset(ce),
3444 (cs - start) * sizeof(*cs));
3448 __execlists_update_reg_state(const struct intel_context *ce,
3449 const struct intel_engine_cs *engine,
3452 struct intel_ring *ring = ce->ring;
3453 u32 *regs = ce->lrc_reg_state;
3455 GEM_BUG_ON(!intel_ring_offset_valid(ring, head));
3456 GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->tail));
3458 regs[CTX_RING_START] = i915_ggtt_offset(ring->vma);
3459 regs[CTX_RING_HEAD] = head;
3460 regs[CTX_RING_TAIL] = ring->tail;
3461 regs[CTX_RING_CTL] = RING_CTL_SIZE(ring->size) | RING_VALID;
3464 if (engine->class == RENDER_CLASS) {
3465 regs[CTX_R_PWR_CLK_STATE] =
3466 intel_sseu_make_rpcs(engine->gt, &ce->sseu);
3468 i915_oa_init_reg_state(ce, engine);
3471 if (ce->wa_bb_page) {
3472 u32 *(*fn)(const struct intel_context *ce, u32 *cs);
3474 fn = gen12_emit_indirect_ctx_xcs;
3475 if (ce->engine->class == RENDER_CLASS)
3476 fn = gen12_emit_indirect_ctx_rcs;
3478 /* Mutually exclusive wrt to global indirect bb */
3479 GEM_BUG_ON(engine->wa_ctx.indirect_ctx.size);
3480 setup_indirect_ctx_bb(ce, engine, fn);
3485 execlists_context_pre_pin(struct intel_context *ce,
3486 struct i915_gem_ww_ctx *ww, void **vaddr)
3488 GEM_BUG_ON(!ce->state);
3489 GEM_BUG_ON(!i915_vma_is_pinned(ce->state));
3491 *vaddr = i915_gem_object_pin_map(ce->state->obj,
3492 i915_coherent_map_type(ce->engine->i915) |
3495 return PTR_ERR_OR_ZERO(*vaddr);
3499 __execlists_context_pin(struct intel_context *ce,
3500 struct intel_engine_cs *engine,
3503 ce->lrc.lrca = lrc_descriptor(ce, engine) | CTX_DESC_FORCE_RESTORE;
3504 ce->lrc_reg_state = vaddr + LRC_STATE_OFFSET;
3505 __execlists_update_reg_state(ce, engine, ce->ring->tail);
3510 static int execlists_context_pin(struct intel_context *ce, void *vaddr)
3512 return __execlists_context_pin(ce, ce->engine, vaddr);
3515 static int execlists_context_alloc(struct intel_context *ce)
3517 return __execlists_context_alloc(ce, ce->engine);
3520 static void execlists_context_reset(struct intel_context *ce)
3522 CE_TRACE(ce, "reset\n");
3523 GEM_BUG_ON(!intel_context_is_pinned(ce));
3525 intel_ring_reset(ce->ring, ce->ring->emit);
3527 /* Scrub away the garbage */
3528 execlists_init_reg_state(ce->lrc_reg_state,
3529 ce, ce->engine, ce->ring, true);
3530 __execlists_update_reg_state(ce, ce->engine, ce->ring->tail);
3532 ce->lrc.desc |= CTX_DESC_FORCE_RESTORE;
3535 static const struct intel_context_ops execlists_context_ops = {
3536 .alloc = execlists_context_alloc,
3538 .pre_pin = execlists_context_pre_pin,
3539 .pin = execlists_context_pin,
3540 .unpin = execlists_context_unpin,
3541 .post_unpin = execlists_context_post_unpin,
3543 .enter = intel_context_enter_engine,
3544 .exit = intel_context_exit_engine,
3546 .reset = execlists_context_reset,
3547 .destroy = execlists_context_destroy,
3550 static u32 hwsp_offset(const struct i915_request *rq)
3552 const struct intel_timeline_cacheline *cl;
3554 /* Before the request is executed, the timeline/cachline is fixed */
3556 cl = rcu_dereference_protected(rq->hwsp_cacheline, 1);
3558 return cl->ggtt_offset;
3560 return rcu_dereference_protected(rq->timeline, 1)->hwsp_offset;
3563 static int gen8_emit_init_breadcrumb(struct i915_request *rq)
3567 GEM_BUG_ON(i915_request_has_initial_breadcrumb(rq));
3568 if (!i915_request_timeline(rq)->has_initial_breadcrumb)
3571 cs = intel_ring_begin(rq, 6);
3576 * Check if we have been preempted before we even get started.
3578 * After this point i915_request_started() reports true, even if
3579 * we get preempted and so are no longer running.
3581 *cs++ = MI_ARB_CHECK;
3584 *cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
3585 *cs++ = hwsp_offset(rq);
3587 *cs++ = rq->fence.seqno - 1;
3589 intel_ring_advance(rq, cs);
3591 /* Record the updated position of the request's payload */
3592 rq->infix = intel_ring_offset(rq, cs);
3594 __set_bit(I915_FENCE_FLAG_INITIAL_BREADCRUMB, &rq->fence.flags);
3599 static int emit_pdps(struct i915_request *rq)
3601 const struct intel_engine_cs * const engine = rq->engine;
3602 struct i915_ppgtt * const ppgtt = i915_vm_to_ppgtt(rq->context->vm);
3606 GEM_BUG_ON(intel_vgpu_active(rq->engine->i915));
3609 * Beware ye of the dragons, this sequence is magic!
3611 * Small changes to this sequence can cause anything from
3612 * GPU hangs to forcewake errors and machine lockups!
3615 /* Flush any residual operations from the context load */
3616 err = engine->emit_flush(rq, EMIT_FLUSH);
3620 /* Magic required to prevent forcewake errors! */
3621 err = engine->emit_flush(rq, EMIT_INVALIDATE);
3625 cs = intel_ring_begin(rq, 4 * GEN8_3LVL_PDPES + 2);
3629 /* Ensure the LRI have landed before we invalidate & continue */
3630 *cs++ = MI_LOAD_REGISTER_IMM(2 * GEN8_3LVL_PDPES) | MI_LRI_FORCE_POSTED;
3631 for (i = GEN8_3LVL_PDPES; i--; ) {
3632 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
3633 u32 base = engine->mmio_base;
3635 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(base, i));
3636 *cs++ = upper_32_bits(pd_daddr);
3637 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(base, i));
3638 *cs++ = lower_32_bits(pd_daddr);
3642 intel_ring_advance(rq, cs);
3647 static int execlists_request_alloc(struct i915_request *request)
3651 GEM_BUG_ON(!intel_context_is_pinned(request->context));
3654 * Flush enough space to reduce the likelihood of waiting after
3655 * we start building the request - in which case we will just
3656 * have to repeat work.
3658 request->reserved_space += EXECLISTS_REQUEST_SIZE;
3661 * Note that after this point, we have committed to using
3662 * this request as it is being used to both track the
3663 * state of engine initialisation and liveness of the
3664 * golden renderstate above. Think twice before you try
3665 * to cancel/unwind this request now.
3668 if (!i915_vm_is_4lvl(request->context->vm)) {
3669 ret = emit_pdps(request);
3674 /* Unconditionally invalidate GPU caches and TLBs. */
3675 ret = request->engine->emit_flush(request, EMIT_INVALIDATE);
3679 request->reserved_space -= EXECLISTS_REQUEST_SIZE;
3684 * In this WA we need to set GEN8_L3SQCREG4[21:21] and reset it after
3685 * PIPE_CONTROL instruction. This is required for the flush to happen correctly
3686 * but there is a slight complication as this is applied in WA batch where the
3687 * values are only initialized once so we cannot take register value at the
3688 * beginning and reuse it further; hence we save its value to memory, upload a
3689 * constant value with bit21 set and then we restore it back with the saved value.
3690 * To simplify the WA, a constant value is formed by using the default value
3691 * of this register. This shouldn't be a problem because we are only modifying
3692 * it for a short period and this batch in non-premptible. We can ofcourse
3693 * use additional instructions that read the actual value of the register
3694 * at that time and set our bit of interest but it makes the WA complicated.
3696 * This WA is also required for Gen9 so extracting as a function avoids
3700 gen8_emit_flush_coherentl3_wa(struct intel_engine_cs *engine, u32 *batch)
3702 /* NB no one else is allowed to scribble over scratch + 256! */
3703 *batch++ = MI_STORE_REGISTER_MEM_GEN8 | MI_SRM_LRM_GLOBAL_GTT;
3704 *batch++ = i915_mmio_reg_offset(GEN8_L3SQCREG4);
3705 *batch++ = intel_gt_scratch_offset(engine->gt,
3706 INTEL_GT_SCRATCH_FIELD_COHERENTL3_WA);
3709 *batch++ = MI_LOAD_REGISTER_IMM(1);
3710 *batch++ = i915_mmio_reg_offset(GEN8_L3SQCREG4);
3711 *batch++ = 0x40400000 | GEN8_LQSC_FLUSH_COHERENT_LINES;
3713 batch = gen8_emit_pipe_control(batch,
3714 PIPE_CONTROL_CS_STALL |
3715 PIPE_CONTROL_DC_FLUSH_ENABLE,
3718 *batch++ = MI_LOAD_REGISTER_MEM_GEN8 | MI_SRM_LRM_GLOBAL_GTT;
3719 *batch++ = i915_mmio_reg_offset(GEN8_L3SQCREG4);
3720 *batch++ = intel_gt_scratch_offset(engine->gt,
3721 INTEL_GT_SCRATCH_FIELD_COHERENTL3_WA);
3728 * Typically we only have one indirect_ctx and per_ctx batch buffer which are
3729 * initialized at the beginning and shared across all contexts but this field
3730 * helps us to have multiple batches at different offsets and select them based
3731 * on a criteria. At the moment this batch always start at the beginning of the page
3732 * and at this point we don't have multiple wa_ctx batch buffers.
3734 * The number of WA applied are not known at the beginning; we use this field
3735 * to return the no of DWORDS written.
3737 * It is to be noted that this batch does not contain MI_BATCH_BUFFER_END
3738 * so it adds NOOPs as padding to make it cacheline aligned.
3739 * MI_BATCH_BUFFER_END will be added to perctx batch and both of them together
3740 * makes a complete batch buffer.
3742 static u32 *gen8_init_indirectctx_bb(struct intel_engine_cs *engine, u32 *batch)
3744 /* WaDisableCtxRestoreArbitration:bdw,chv */
3745 *batch++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
3747 /* WaFlushCoherentL3CacheLinesAtContextSwitch:bdw */
3748 if (IS_BROADWELL(engine->i915))
3749 batch = gen8_emit_flush_coherentl3_wa(engine, batch);
3751 /* WaClearSlmSpaceAtContextSwitch:bdw,chv */
3752 /* Actual scratch location is at 128 bytes offset */
3753 batch = gen8_emit_pipe_control(batch,
3754 PIPE_CONTROL_FLUSH_L3 |
3755 PIPE_CONTROL_STORE_DATA_INDEX |
3756 PIPE_CONTROL_CS_STALL |
3757 PIPE_CONTROL_QW_WRITE,
3758 LRC_PPHWSP_SCRATCH_ADDR);
3760 *batch++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
3762 /* Pad to end of cacheline */
3763 while ((unsigned long)batch % CACHELINE_BYTES)
3767 * MI_BATCH_BUFFER_END is not required in Indirect ctx BB because
3768 * execution depends on the length specified in terms of cache lines
3769 * in the register CTX_RCS_INDIRECT_CTX
3780 static u32 *emit_lri(u32 *batch, const struct lri *lri, unsigned int count)
3782 GEM_BUG_ON(!count || count > 63);
3784 *batch++ = MI_LOAD_REGISTER_IMM(count);
3786 *batch++ = i915_mmio_reg_offset(lri->reg);
3787 *batch++ = lri->value;
3788 } while (lri++, --count);
3794 static u32 *gen9_init_indirectctx_bb(struct intel_engine_cs *engine, u32 *batch)
3796 static const struct lri lri[] = {
3797 /* WaDisableGatherAtSetShaderCommonSlice:skl,bxt,kbl,glk */
3799 COMMON_SLICE_CHICKEN2,
3800 __MASKED_FIELD(GEN9_DISABLE_GATHER_AT_SET_SHADER_COMMON_SLICE,
3807 __MASKED_FIELD(FF_SLICE_CHICKEN_CL_PROVOKING_VERTEX_FIX,
3808 FF_SLICE_CHICKEN_CL_PROVOKING_VERTEX_FIX),
3814 __MASKED_FIELD(_3D_CHICKEN_SF_PROVOKING_VERTEX_FIX,
3815 _3D_CHICKEN_SF_PROVOKING_VERTEX_FIX),
3819 *batch++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
3821 /* WaFlushCoherentL3CacheLinesAtContextSwitch:skl,bxt,glk */
3822 batch = gen8_emit_flush_coherentl3_wa(engine, batch);
3824 /* WaClearSlmSpaceAtContextSwitch:skl,bxt,kbl,glk,cfl */
3825 batch = gen8_emit_pipe_control(batch,
3826 PIPE_CONTROL_FLUSH_L3 |
3827 PIPE_CONTROL_STORE_DATA_INDEX |
3828 PIPE_CONTROL_CS_STALL |
3829 PIPE_CONTROL_QW_WRITE,
3830 LRC_PPHWSP_SCRATCH_ADDR);
3832 batch = emit_lri(batch, lri, ARRAY_SIZE(lri));
3834 /* WaMediaPoolStateCmdInWABB:bxt,glk */
3835 if (HAS_POOLED_EU(engine->i915)) {
3837 * EU pool configuration is setup along with golden context
3838 * during context initialization. This value depends on
3839 * device type (2x6 or 3x6) and needs to be updated based
3840 * on which subslice is disabled especially for 2x6
3841 * devices, however it is safe to load default
3842 * configuration of 3x6 device instead of masking off
3843 * corresponding bits because HW ignores bits of a disabled
3844 * subslice and drops down to appropriate config. Please
3845 * see render_state_setup() in i915_gem_render_state.c for
3846 * possible configurations, to avoid duplication they are
3847 * not shown here again.
3849 *batch++ = GEN9_MEDIA_POOL_STATE;
3850 *batch++ = GEN9_MEDIA_POOL_ENABLE;
3851 *batch++ = 0x00777000;
3857 *batch++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
3859 /* Pad to end of cacheline */
3860 while ((unsigned long)batch % CACHELINE_BYTES)
3867 gen10_init_indirectctx_bb(struct intel_engine_cs *engine, u32 *batch)
3872 * WaPipeControlBefore3DStateSamplePattern: cnl
3874 * Ensure the engine is idle prior to programming a
3875 * 3DSTATE_SAMPLE_PATTERN during a context restore.
3877 batch = gen8_emit_pipe_control(batch,
3878 PIPE_CONTROL_CS_STALL,
3881 * WaPipeControlBefore3DStateSamplePattern says we need 4 dwords for
3882 * the PIPE_CONTROL followed by 12 dwords of 0x0, so 16 dwords in
3883 * total. However, a PIPE_CONTROL is 6 dwords long, not 4, which is
3884 * confusing. Since gen8_emit_pipe_control() already advances the
3885 * batch by 6 dwords, we advance the other 10 here, completing a
3886 * cacheline. It's not clear if the workaround requires this padding
3887 * before other commands, or if it's just the regular padding we would
3888 * already have for the workaround bb, so leave it here for now.
3890 for (i = 0; i < 10; i++)
3893 /* Pad to end of cacheline */
3894 while ((unsigned long)batch % CACHELINE_BYTES)
3900 #define CTX_WA_BB_OBJ_SIZE (PAGE_SIZE)
3902 static int lrc_setup_wa_ctx(struct intel_engine_cs *engine)
3904 struct drm_i915_gem_object *obj;
3905 struct i915_vma *vma;
3908 obj = i915_gem_object_create_shmem(engine->i915, CTX_WA_BB_OBJ_SIZE);
3910 return PTR_ERR(obj);
3912 vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
3918 err = i915_ggtt_pin(vma, NULL, 0, PIN_HIGH);
3922 engine->wa_ctx.vma = vma;
3926 i915_gem_object_put(obj);
3930 static void lrc_destroy_wa_ctx(struct intel_engine_cs *engine)
3932 i915_vma_unpin_and_release(&engine->wa_ctx.vma, 0);
3935 typedef u32 *(*wa_bb_func_t)(struct intel_engine_cs *engine, u32 *batch);
3937 static int intel_init_workaround_bb(struct intel_engine_cs *engine)
3939 struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx;
3940 struct i915_wa_ctx_bb *wa_bb[2] = { &wa_ctx->indirect_ctx,
3942 wa_bb_func_t wa_bb_fn[2];
3943 void *batch, *batch_ptr;
3947 if (engine->class != RENDER_CLASS)
3950 switch (INTEL_GEN(engine->i915)) {
3955 wa_bb_fn[0] = gen10_init_indirectctx_bb;
3959 wa_bb_fn[0] = gen9_init_indirectctx_bb;
3963 wa_bb_fn[0] = gen8_init_indirectctx_bb;
3967 MISSING_CASE(INTEL_GEN(engine->i915));
3971 ret = lrc_setup_wa_ctx(engine);
3973 drm_dbg(&engine->i915->drm,
3974 "Failed to setup context WA page: %d\n", ret);
3978 batch = i915_gem_object_pin_map(wa_ctx->vma->obj, I915_MAP_WB);
3981 * Emit the two workaround batch buffers, recording the offset from the
3982 * start of the workaround batch buffer object for each and their
3986 for (i = 0; i < ARRAY_SIZE(wa_bb_fn); i++) {
3987 wa_bb[i]->offset = batch_ptr - batch;
3988 if (GEM_DEBUG_WARN_ON(!IS_ALIGNED(wa_bb[i]->offset,
3989 CACHELINE_BYTES))) {
3994 batch_ptr = wa_bb_fn[i](engine, batch_ptr);
3995 wa_bb[i]->size = batch_ptr - (batch + wa_bb[i]->offset);
3997 GEM_BUG_ON(batch_ptr - batch > CTX_WA_BB_OBJ_SIZE);
3999 __i915_gem_object_flush_map(wa_ctx->vma->obj, 0, batch_ptr - batch);
4000 __i915_gem_object_release_map(wa_ctx->vma->obj);
4002 lrc_destroy_wa_ctx(engine);
4007 static void reset_csb_pointers(struct intel_engine_cs *engine)
4009 struct intel_engine_execlists * const execlists = &engine->execlists;
4010 const unsigned int reset_value = execlists->csb_size - 1;
4012 ring_set_paused(engine, 0);
4015 * Sometimes Icelake forgets to reset its pointers on a GPU reset.
4016 * Bludgeon them with a mmio update to be sure.
4018 ENGINE_WRITE(engine, RING_CONTEXT_STATUS_PTR,
4019 0xffff << 16 | reset_value << 8 | reset_value);
4020 ENGINE_POSTING_READ(engine, RING_CONTEXT_STATUS_PTR);
4023 * After a reset, the HW starts writing into CSB entry [0]. We
4024 * therefore have to set our HEAD pointer back one entry so that
4025 * the *first* entry we check is entry 0. To complicate this further,
4026 * as we don't wait for the first interrupt after reset, we have to
4027 * fake the HW write to point back to the last entry so that our
4028 * inline comparison of our cached head position against the last HW
4029 * write works even before the first interrupt.
4031 execlists->csb_head = reset_value;
4032 WRITE_ONCE(*execlists->csb_write, reset_value);
4033 wmb(); /* Make sure this is visible to HW (paranoia?) */
4035 /* Check that the GPU does indeed update the CSB entries! */
4036 memset(execlists->csb_status, -1, (reset_value + 1) * sizeof(u64));
4037 invalidate_csb_entries(&execlists->csb_status[0],
4038 &execlists->csb_status[reset_value]);
4040 /* Once more for luck and our trusty paranoia */
4041 ENGINE_WRITE(engine, RING_CONTEXT_STATUS_PTR,
4042 0xffff << 16 | reset_value << 8 | reset_value);
4043 ENGINE_POSTING_READ(engine, RING_CONTEXT_STATUS_PTR);
4045 GEM_BUG_ON(READ_ONCE(*execlists->csb_write) != reset_value);
4048 static void execlists_sanitize(struct intel_engine_cs *engine)
4051 * Poison residual state on resume, in case the suspend didn't!
4053 * We have to assume that across suspend/resume (or other loss
4054 * of control) that the contents of our pinned buffers has been
4055 * lost, replaced by garbage. Since this doesn't always happen,
4056 * let's poison such state so that we more quickly spot when
4057 * we falsely assume it has been preserved.
4059 if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
4060 memset(engine->status_page.addr, POISON_INUSE, PAGE_SIZE);
4062 reset_csb_pointers(engine);
4065 * The kernel_context HWSP is stored in the status_page. As above,
4066 * that may be lost on resume/initialisation, and so we need to
4067 * reset the value in the HWSP.
4069 intel_timeline_reset_seqno(engine->kernel_context->timeline);
4071 /* And scrub the dirty cachelines for the HWSP */
4072 clflush_cache_range(engine->status_page.addr, PAGE_SIZE);
4075 static void enable_error_interrupt(struct intel_engine_cs *engine)
4079 engine->execlists.error_interrupt = 0;
4080 ENGINE_WRITE(engine, RING_EMR, ~0u);
4081 ENGINE_WRITE(engine, RING_EIR, ~0u); /* clear all existing errors */
4083 status = ENGINE_READ(engine, RING_ESR);
4084 if (unlikely(status)) {
4085 drm_err(&engine->i915->drm,
4086 "engine '%s' resumed still in error: %08x\n",
4087 engine->name, status);
4088 __intel_gt_reset(engine->gt, engine->mask);
4092 * On current gen8+, we have 2 signals to play with
4094 * - I915_ERROR_INSTUCTION (bit 0)
4096 * Generate an error if the command parser encounters an invalid
4099 * This is a fatal error.
4103 * Generate an error on privilege violation (where the CP replaces
4104 * the instruction with a no-op). This also fires for writes into
4105 * read-only scratch pages.
4107 * This is a non-fatal error, parsing continues.
4109 * * there are a few others defined for odd HW that we do not use
4111 * Since CP_PRIV fires for cases where we have chosen to ignore the
4112 * error (as the HW is validating and suppressing the mistakes), we
4113 * only unmask the instruction error bit.
4115 ENGINE_WRITE(engine, RING_EMR, ~I915_ERROR_INSTRUCTION);
4118 static void enable_execlists(struct intel_engine_cs *engine)
4122 assert_forcewakes_active(engine->uncore, FORCEWAKE_ALL);
4124 intel_engine_set_hwsp_writemask(engine, ~0u); /* HWSTAM */
4126 if (INTEL_GEN(engine->i915) >= 11)
4127 mode = _MASKED_BIT_ENABLE(GEN11_GFX_DISABLE_LEGACY_MODE);
4129 mode = _MASKED_BIT_ENABLE(GFX_RUN_LIST_ENABLE);
4130 ENGINE_WRITE_FW(engine, RING_MODE_GEN7, mode);
4132 ENGINE_WRITE_FW(engine, RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
4134 ENGINE_WRITE_FW(engine,
4136 i915_ggtt_offset(engine->status_page.vma));
4137 ENGINE_POSTING_READ(engine, RING_HWS_PGA);
4139 enable_error_interrupt(engine);
4141 engine->context_tag = GENMASK(BITS_PER_LONG - 2, 0);
4144 static bool unexpected_starting_state(struct intel_engine_cs *engine)
4146 bool unexpected = false;
4148 if (ENGINE_READ_FW(engine, RING_MI_MODE) & STOP_RING) {
4149 drm_dbg(&engine->i915->drm,
4150 "STOP_RING still set in RING_MI_MODE\n");
4157 static int execlists_resume(struct intel_engine_cs *engine)
4159 intel_mocs_init_engine(engine);
4161 intel_breadcrumbs_reset(engine->breadcrumbs);
4163 if (GEM_SHOW_DEBUG() && unexpected_starting_state(engine)) {
4164 struct drm_printer p = drm_debug_printer(__func__);
4166 intel_engine_dump(engine, &p, NULL);
4169 enable_execlists(engine);
4174 static void execlists_reset_prepare(struct intel_engine_cs *engine)
4176 struct intel_engine_execlists * const execlists = &engine->execlists;
4177 unsigned long flags;
4179 ENGINE_TRACE(engine, "depth<-%d\n",
4180 atomic_read(&execlists->tasklet.count));
4183 * Prevent request submission to the hardware until we have
4184 * completed the reset in i915_gem_reset_finish(). If a request
4185 * is completed by one engine, it may then queue a request
4186 * to a second via its execlists->tasklet *just* as we are
4187 * calling engine->resume() and also writing the ELSP.
4188 * Turning off the execlists->tasklet until the reset is over
4189 * prevents the race.
4191 __tasklet_disable_sync_once(&execlists->tasklet);
4192 GEM_BUG_ON(!reset_in_progress(execlists));
4194 /* And flush any current direct submission. */
4195 spin_lock_irqsave(&engine->active.lock, flags);
4196 spin_unlock_irqrestore(&engine->active.lock, flags);
4199 * We stop engines, otherwise we might get failed reset and a
4200 * dead gpu (on elk). Also as modern gpu as kbl can suffer
4201 * from system hang if batchbuffer is progressing when
4202 * the reset is issued, regardless of READY_TO_RESET ack.
4203 * Thus assume it is best to stop engines on all gens
4204 * where we have a gpu reset.
4206 * WaKBLVECSSemaphoreWaitPoll:kbl (on ALL_ENGINES)
4208 * FIXME: Wa for more modern gens needs to be validated
4210 ring_set_paused(engine, 1);
4211 intel_engine_stop_cs(engine);
4213 engine->execlists.reset_ccid = active_ccid(engine);
4216 static void __reset_stop_ring(u32 *regs, const struct intel_engine_cs *engine)
4220 x = lrc_ring_mi_mode(engine);
4222 regs[x + 1] &= ~STOP_RING;
4223 regs[x + 1] |= STOP_RING << 16;
4227 static void __execlists_reset_reg_state(const struct intel_context *ce,
4228 const struct intel_engine_cs *engine)
4230 u32 *regs = ce->lrc_reg_state;
4232 __reset_stop_ring(regs, engine);
4235 static void __execlists_reset(struct intel_engine_cs *engine, bool stalled)
4237 struct intel_engine_execlists * const execlists = &engine->execlists;
4238 struct intel_context *ce;
4239 struct i915_request *rq;
4242 mb(); /* paranoia: read the CSB pointers from after the reset */
4243 clflush(execlists->csb_write);
4246 process_csb(engine); /* drain preemption events */
4248 /* Following the reset, we need to reload the CSB read/write pointers */
4249 reset_csb_pointers(engine);
4252 * Save the currently executing context, even if we completed
4253 * its request, it was still running at the time of the
4254 * reset and will have been clobbered.
4256 rq = active_context(engine, engine->execlists.reset_ccid);
4261 GEM_BUG_ON(!i915_vma_is_pinned(ce->state));
4263 if (i915_request_completed(rq)) {
4264 /* Idle context; tidy up the ring so we can restart afresh */
4265 head = intel_ring_wrap(ce->ring, rq->tail);
4269 /* We still have requests in-flight; the engine should be active */
4270 GEM_BUG_ON(!intel_engine_pm_is_awake(engine));
4272 /* Context has requests still in-flight; it should not be idle! */
4273 GEM_BUG_ON(i915_active_is_idle(&ce->active));
4275 rq = active_request(ce->timeline, rq);
4276 head = intel_ring_wrap(ce->ring, rq->head);
4277 GEM_BUG_ON(head == ce->ring->tail);
4280 * If this request hasn't started yet, e.g. it is waiting on a
4281 * semaphore, we need to avoid skipping the request or else we
4282 * break the signaling chain. However, if the context is corrupt
4283 * the request will not restart and we will be stuck with a wedged
4284 * device. It is quite often the case that if we issue a reset
4285 * while the GPU is loading the context image, that the context
4286 * image becomes corrupt.
4288 * Otherwise, if we have not started yet, the request should replay
4289 * perfectly and we do not need to flag the result as being erroneous.
4291 if (!i915_request_started(rq))
4295 * If the request was innocent, we leave the request in the ELSP
4296 * and will try to replay it on restarting. The context image may
4297 * have been corrupted by the reset, in which case we may have
4298 * to service a new GPU hang, but more likely we can continue on
4301 * If the request was guilty, we presume the context is corrupt
4302 * and have to at least restore the RING register in the context
4303 * image back to the expected values to skip over the guilty request.
4305 __i915_request_reset(rq, stalled);
4308 * We want a simple context + ring to execute the breadcrumb update.
4309 * We cannot rely on the context being intact across the GPU hang,
4310 * so clear it and rebuild just what we need for the breadcrumb.
4311 * All pending requests for this context will be zapped, and any
4312 * future request will be after userspace has had the opportunity
4313 * to recreate its own state.
4316 ENGINE_TRACE(engine, "replay {head:%04x, tail:%04x}\n",
4317 head, ce->ring->tail);
4318 __execlists_reset_reg_state(ce, engine);
4319 __execlists_update_reg_state(ce, engine, head);
4320 ce->lrc.desc |= CTX_DESC_FORCE_RESTORE; /* paranoid: GPU was reset! */
4323 /* Push back any incomplete requests for replay after the reset. */
4324 cancel_port_requests(execlists);
4325 __unwind_incomplete_requests(engine);
4328 static void execlists_reset_rewind(struct intel_engine_cs *engine, bool stalled)
4330 unsigned long flags;
4332 ENGINE_TRACE(engine, "\n");
4334 spin_lock_irqsave(&engine->active.lock, flags);
4336 __execlists_reset(engine, stalled);
4338 spin_unlock_irqrestore(&engine->active.lock, flags);
4341 static void nop_submission_tasklet(unsigned long data)
4343 struct intel_engine_cs * const engine = (struct intel_engine_cs *)data;
4345 /* The driver is wedged; don't process any more events. */
4346 WRITE_ONCE(engine->execlists.queue_priority_hint, INT_MIN);
4349 static void execlists_reset_cancel(struct intel_engine_cs *engine)
4351 struct intel_engine_execlists * const execlists = &engine->execlists;
4352 struct i915_request *rq, *rn;
4354 unsigned long flags;
4356 ENGINE_TRACE(engine, "\n");
4359 * Before we call engine->cancel_requests(), we should have exclusive
4360 * access to the submission state. This is arranged for us by the
4361 * caller disabling the interrupt generation, the tasklet and other
4362 * threads that may then access the same state, giving us a free hand
4363 * to reset state. However, we still need to let lockdep be aware that
4364 * we know this state may be accessed in hardirq context, so we
4365 * disable the irq around this manipulation and we want to keep
4366 * the spinlock focused on its duties and not accidentally conflate
4367 * coverage to the submission's irq state. (Similarly, although we
4368 * shouldn't need to disable irq around the manipulation of the
4369 * submission's irq state, we also wish to remind ourselves that
4372 spin_lock_irqsave(&engine->active.lock, flags);
4374 __execlists_reset(engine, true);
4376 /* Mark all executing requests as skipped. */
4377 list_for_each_entry(rq, &engine->active.requests, sched.link)
4380 /* Flush the queued requests to the timeline list (for retiring). */
4381 while ((rb = rb_first_cached(&execlists->queue))) {
4382 struct i915_priolist *p = to_priolist(rb);
4385 priolist_for_each_request_consume(rq, rn, p, i) {
4387 __i915_request_submit(rq);
4390 rb_erase_cached(&p->node, &execlists->queue);
4391 i915_priolist_free(p);
4394 /* On-hold requests will be flushed to timeline upon their release */
4395 list_for_each_entry(rq, &engine->active.hold, sched.link)
4398 /* Cancel all attached virtual engines */
4399 while ((rb = rb_first_cached(&execlists->virtual))) {
4400 struct virtual_engine *ve =
4401 rb_entry(rb, typeof(*ve), nodes[engine->id].rb);
4403 rb_erase_cached(rb, &execlists->virtual);
4406 spin_lock(&ve->base.active.lock);
4407 rq = fetch_and_zero(&ve->request);
4411 rq->engine = engine;
4412 __i915_request_submit(rq);
4413 i915_request_put(rq);
4415 ve->base.execlists.queue_priority_hint = INT_MIN;
4417 spin_unlock(&ve->base.active.lock);
4420 /* Remaining _unready_ requests will be nop'ed when submitted */
4422 execlists->queue_priority_hint = INT_MIN;
4423 execlists->queue = RB_ROOT_CACHED;
4425 GEM_BUG_ON(__tasklet_is_enabled(&execlists->tasklet));
4426 execlists->tasklet.func = nop_submission_tasklet;
4428 spin_unlock_irqrestore(&engine->active.lock, flags);
4431 static void execlists_reset_finish(struct intel_engine_cs *engine)
4433 struct intel_engine_execlists * const execlists = &engine->execlists;
4436 * After a GPU reset, we may have requests to replay. Do so now while
4437 * we still have the forcewake to be sure that the GPU is not allowed
4438 * to sleep before we restart and reload a context.
4440 GEM_BUG_ON(!reset_in_progress(execlists));
4441 if (!RB_EMPTY_ROOT(&execlists->queue.rb_root))
4442 execlists->tasklet.func(execlists->tasklet.data);
4444 if (__tasklet_enable(&execlists->tasklet))
4445 /* And kick in case we missed a new request submission. */
4446 tasklet_hi_schedule(&execlists->tasklet);
4447 ENGINE_TRACE(engine, "depth->%d\n",
4448 atomic_read(&execlists->tasklet.count));
4451 static int gen8_emit_bb_start_noarb(struct i915_request *rq,
4452 u64 offset, u32 len,
4453 const unsigned int flags)
4457 cs = intel_ring_begin(rq, 4);
4462 * WaDisableCtxRestoreArbitration:bdw,chv
4464 * We don't need to perform MI_ARB_ENABLE as often as we do (in
4465 * particular all the gen that do not need the w/a at all!), if we
4466 * took care to make sure that on every switch into this context
4467 * (both ordinary and for preemption) that arbitrartion was enabled
4468 * we would be fine. However, for gen8 there is another w/a that
4469 * requires us to not preempt inside GPGPU execution, so we keep
4470 * arbitration disabled for gen8 batches. Arbitration will be
4471 * re-enabled before we close the request
4472 * (engine->emit_fini_breadcrumb).
4474 *cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
4476 /* FIXME(BDW+): Address space and security selectors. */
4477 *cs++ = MI_BATCH_BUFFER_START_GEN8 |
4478 (flags & I915_DISPATCH_SECURE ? 0 : BIT(8));
4479 *cs++ = lower_32_bits(offset);
4480 *cs++ = upper_32_bits(offset);
4482 intel_ring_advance(rq, cs);
4487 static int gen8_emit_bb_start(struct i915_request *rq,
4488 u64 offset, u32 len,
4489 const unsigned int flags)
4493 cs = intel_ring_begin(rq, 6);
4497 *cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
4499 *cs++ = MI_BATCH_BUFFER_START_GEN8 |
4500 (flags & I915_DISPATCH_SECURE ? 0 : BIT(8));
4501 *cs++ = lower_32_bits(offset);
4502 *cs++ = upper_32_bits(offset);
4504 *cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
4507 intel_ring_advance(rq, cs);
4512 static void gen8_logical_ring_enable_irq(struct intel_engine_cs *engine)
4514 ENGINE_WRITE(engine, RING_IMR,
4515 ~(engine->irq_enable_mask | engine->irq_keep_mask));
4516 ENGINE_POSTING_READ(engine, RING_IMR);
4519 static void gen8_logical_ring_disable_irq(struct intel_engine_cs *engine)
4521 ENGINE_WRITE(engine, RING_IMR, ~engine->irq_keep_mask);
4524 static int gen8_emit_flush(struct i915_request *request, u32 mode)
4528 cs = intel_ring_begin(request, 4);
4532 cmd = MI_FLUSH_DW + 1;
4534 /* We always require a command barrier so that subsequent
4535 * commands, such as breadcrumb interrupts, are strictly ordered
4536 * wrt the contents of the write cache being flushed to memory
4537 * (and thus being coherent from the CPU).
4539 cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
4541 if (mode & EMIT_INVALIDATE) {
4542 cmd |= MI_INVALIDATE_TLB;
4543 if (request->engine->class == VIDEO_DECODE_CLASS)
4544 cmd |= MI_INVALIDATE_BSD;
4548 *cs++ = LRC_PPHWSP_SCRATCH_ADDR;
4549 *cs++ = 0; /* upper addr */
4550 *cs++ = 0; /* value */
4551 intel_ring_advance(request, cs);
4556 static int gen8_emit_flush_render(struct i915_request *request,
4559 bool vf_flush_wa = false, dc_flush_wa = false;
4563 flags |= PIPE_CONTROL_CS_STALL;
4565 if (mode & EMIT_FLUSH) {
4566 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
4567 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
4568 flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
4569 flags |= PIPE_CONTROL_FLUSH_ENABLE;
4572 if (mode & EMIT_INVALIDATE) {
4573 flags |= PIPE_CONTROL_TLB_INVALIDATE;
4574 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
4575 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
4576 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
4577 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
4578 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
4579 flags |= PIPE_CONTROL_QW_WRITE;
4580 flags |= PIPE_CONTROL_STORE_DATA_INDEX;
4583 * On GEN9: before VF_CACHE_INVALIDATE we need to emit a NULL
4586 if (IS_GEN(request->engine->i915, 9))
4589 /* WaForGAMHang:kbl */
4590 if (IS_KBL_GT_REVID(request->engine->i915, 0, KBL_REVID_B0))
4602 cs = intel_ring_begin(request, len);
4607 cs = gen8_emit_pipe_control(cs, 0, 0);
4610 cs = gen8_emit_pipe_control(cs, PIPE_CONTROL_DC_FLUSH_ENABLE,
4613 cs = gen8_emit_pipe_control(cs, flags, LRC_PPHWSP_SCRATCH_ADDR);
4616 cs = gen8_emit_pipe_control(cs, PIPE_CONTROL_CS_STALL, 0);
4618 intel_ring_advance(request, cs);
4623 static int gen11_emit_flush_render(struct i915_request *request,
4626 if (mode & EMIT_FLUSH) {
4630 flags |= PIPE_CONTROL_CS_STALL;
4632 flags |= PIPE_CONTROL_TILE_CACHE_FLUSH;
4633 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
4634 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
4635 flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
4636 flags |= PIPE_CONTROL_FLUSH_ENABLE;
4637 flags |= PIPE_CONTROL_QW_WRITE;
4638 flags |= PIPE_CONTROL_STORE_DATA_INDEX;
4640 cs = intel_ring_begin(request, 6);
4644 cs = gen8_emit_pipe_control(cs, flags, LRC_PPHWSP_SCRATCH_ADDR);
4645 intel_ring_advance(request, cs);
4648 if (mode & EMIT_INVALIDATE) {
4652 flags |= PIPE_CONTROL_CS_STALL;
4654 flags |= PIPE_CONTROL_COMMAND_CACHE_INVALIDATE;
4655 flags |= PIPE_CONTROL_TLB_INVALIDATE;
4656 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
4657 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
4658 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
4659 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
4660 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
4661 flags |= PIPE_CONTROL_QW_WRITE;
4662 flags |= PIPE_CONTROL_STORE_DATA_INDEX;
4664 cs = intel_ring_begin(request, 6);
4668 cs = gen8_emit_pipe_control(cs, flags, LRC_PPHWSP_SCRATCH_ADDR);
4669 intel_ring_advance(request, cs);
4675 static u32 preparser_disable(bool state)
4677 return MI_ARB_CHECK | 1 << 8 | state;
4680 static i915_reg_t aux_inv_reg(const struct intel_engine_cs *engine)
4682 static const i915_reg_t vd[] = {
4689 static const i915_reg_t ve[] = {
4694 if (engine->class == VIDEO_DECODE_CLASS)
4695 return vd[engine->instance];
4697 if (engine->class == VIDEO_ENHANCEMENT_CLASS)
4698 return ve[engine->instance];
4700 GEM_BUG_ON("unknown aux_inv_reg\n");
4702 return INVALID_MMIO_REG;
4706 gen12_emit_aux_table_inv(const i915_reg_t inv_reg, u32 *cs)
4708 *cs++ = MI_LOAD_REGISTER_IMM(1);
4709 *cs++ = i915_mmio_reg_offset(inv_reg);
4716 static int gen12_emit_flush_render(struct i915_request *request,
4719 if (mode & EMIT_FLUSH) {
4723 flags |= PIPE_CONTROL_TILE_CACHE_FLUSH;
4724 flags |= PIPE_CONTROL_FLUSH_L3;
4725 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
4726 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
4727 /* Wa_1409600907:tgl */
4728 flags |= PIPE_CONTROL_DEPTH_STALL;
4729 flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
4730 flags |= PIPE_CONTROL_FLUSH_ENABLE;
4732 flags |= PIPE_CONTROL_STORE_DATA_INDEX;
4733 flags |= PIPE_CONTROL_QW_WRITE;
4735 flags |= PIPE_CONTROL_CS_STALL;
4737 cs = intel_ring_begin(request, 6);
4741 cs = gen12_emit_pipe_control(cs,
4742 PIPE_CONTROL0_HDC_PIPELINE_FLUSH,
4743 flags, LRC_PPHWSP_SCRATCH_ADDR);
4744 intel_ring_advance(request, cs);
4747 if (mode & EMIT_INVALIDATE) {
4751 flags |= PIPE_CONTROL_COMMAND_CACHE_INVALIDATE;
4752 flags |= PIPE_CONTROL_TLB_INVALIDATE;
4753 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
4754 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
4755 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
4756 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
4757 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
4759 flags |= PIPE_CONTROL_STORE_DATA_INDEX;
4760 flags |= PIPE_CONTROL_QW_WRITE;
4762 flags |= PIPE_CONTROL_CS_STALL;
4764 cs = intel_ring_begin(request, 8 + 4);
4769 * Prevent the pre-parser from skipping past the TLB
4770 * invalidate and loading a stale page for the batch
4771 * buffer / request payload.
4773 *cs++ = preparser_disable(true);
4775 cs = gen8_emit_pipe_control(cs, flags, LRC_PPHWSP_SCRATCH_ADDR);
4777 /* hsdes: 1809175790 */
4778 cs = gen12_emit_aux_table_inv(GEN12_GFX_CCS_AUX_NV, cs);
4780 *cs++ = preparser_disable(false);
4781 intel_ring_advance(request, cs);
4787 static int gen12_emit_flush(struct i915_request *request, u32 mode)
4789 intel_engine_mask_t aux_inv = 0;
4793 if (mode & EMIT_INVALIDATE)
4795 if (mode & EMIT_INVALIDATE)
4796 aux_inv = request->engine->mask & ~BIT(BCS0);
4798 cmd += 2 * hweight8(aux_inv) + 2;
4800 cs = intel_ring_begin(request, cmd);
4804 if (mode & EMIT_INVALIDATE)
4805 *cs++ = preparser_disable(true);
4807 cmd = MI_FLUSH_DW + 1;
4809 /* We always require a command barrier so that subsequent
4810 * commands, such as breadcrumb interrupts, are strictly ordered
4811 * wrt the contents of the write cache being flushed to memory
4812 * (and thus being coherent from the CPU).
4814 cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
4816 if (mode & EMIT_INVALIDATE) {
4817 cmd |= MI_INVALIDATE_TLB;
4818 if (request->engine->class == VIDEO_DECODE_CLASS)
4819 cmd |= MI_INVALIDATE_BSD;
4823 *cs++ = LRC_PPHWSP_SCRATCH_ADDR;
4824 *cs++ = 0; /* upper addr */
4825 *cs++ = 0; /* value */
4827 if (aux_inv) { /* hsdes: 1809175790 */
4828 struct intel_engine_cs *engine;
4831 *cs++ = MI_LOAD_REGISTER_IMM(hweight8(aux_inv));
4832 for_each_engine_masked(engine, request->engine->gt,
4834 *cs++ = i915_mmio_reg_offset(aux_inv_reg(engine));
4840 if (mode & EMIT_INVALIDATE)
4841 *cs++ = preparser_disable(false);
4843 intel_ring_advance(request, cs);
4848 static void assert_request_valid(struct i915_request *rq)
4850 struct intel_ring *ring __maybe_unused = rq->ring;
4852 /* Can we unwind this request without appearing to go forwards? */
4853 GEM_BUG_ON(intel_ring_direction(ring, rq->wa_tail, rq->head) <= 0);
4857 * Reserve space for 2 NOOPs at the end of each request to be
4858 * used as a workaround for not being allowed to do lite
4859 * restore with HEAD==TAIL (WaIdleLiteRestore).
4861 static u32 *gen8_emit_wa_tail(struct i915_request *request, u32 *cs)
4863 /* Ensure there's always at least one preemption point per-request. */
4864 *cs++ = MI_ARB_CHECK;
4866 request->wa_tail = intel_ring_offset(request, cs);
4868 /* Check that entire request is less than half the ring */
4869 assert_request_valid(request);
4874 static u32 *emit_preempt_busywait(struct i915_request *request, u32 *cs)
4876 *cs++ = MI_SEMAPHORE_WAIT |
4877 MI_SEMAPHORE_GLOBAL_GTT |
4879 MI_SEMAPHORE_SAD_EQ_SDD;
4881 *cs++ = intel_hws_preempt_address(request->engine);
4887 static __always_inline u32*
4888 gen8_emit_fini_breadcrumb_tail(struct i915_request *request, u32 *cs)
4890 *cs++ = MI_USER_INTERRUPT;
4892 *cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
4893 if (intel_engine_has_semaphores(request->engine))
4894 cs = emit_preempt_busywait(request, cs);
4896 request->tail = intel_ring_offset(request, cs);
4897 assert_ring_tail_valid(request->ring, request->tail);
4899 return gen8_emit_wa_tail(request, cs);
4902 static u32 *emit_xcs_breadcrumb(struct i915_request *rq, u32 *cs)
4904 return gen8_emit_ggtt_write(cs, rq->fence.seqno, hwsp_offset(rq), 0);
4907 static u32 *gen8_emit_fini_breadcrumb(struct i915_request *rq, u32 *cs)
4909 return gen8_emit_fini_breadcrumb_tail(rq, emit_xcs_breadcrumb(rq, cs));
4912 static u32 *gen8_emit_fini_breadcrumb_rcs(struct i915_request *request, u32 *cs)
4914 cs = gen8_emit_pipe_control(cs,
4915 PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
4916 PIPE_CONTROL_DEPTH_CACHE_FLUSH |
4917 PIPE_CONTROL_DC_FLUSH_ENABLE,
4920 /* XXX flush+write+CS_STALL all in one upsets gem_concurrent_blt:kbl */
4921 cs = gen8_emit_ggtt_write_rcs(cs,
4922 request->fence.seqno,
4923 hwsp_offset(request),
4924 PIPE_CONTROL_FLUSH_ENABLE |
4925 PIPE_CONTROL_CS_STALL);
4927 return gen8_emit_fini_breadcrumb_tail(request, cs);
4931 gen11_emit_fini_breadcrumb_rcs(struct i915_request *request, u32 *cs)
4933 cs = gen8_emit_ggtt_write_rcs(cs,
4934 request->fence.seqno,
4935 hwsp_offset(request),
4936 PIPE_CONTROL_CS_STALL |
4937 PIPE_CONTROL_TILE_CACHE_FLUSH |
4938 PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
4939 PIPE_CONTROL_DEPTH_CACHE_FLUSH |
4940 PIPE_CONTROL_DC_FLUSH_ENABLE |
4941 PIPE_CONTROL_FLUSH_ENABLE);
4943 return gen8_emit_fini_breadcrumb_tail(request, cs);
4947 * Note that the CS instruction pre-parser will not stall on the breadcrumb
4948 * flush and will continue pre-fetching the instructions after it before the
4949 * memory sync is completed. On pre-gen12 HW, the pre-parser will stop at
4950 * BB_START/END instructions, so, even though we might pre-fetch the pre-amble
4951 * of the next request before the memory has been flushed, we're guaranteed that
4952 * we won't access the batch itself too early.
4953 * However, on gen12+ the parser can pre-fetch across the BB_START/END commands,
4954 * so, if the current request is modifying an instruction in the next request on
4955 * the same intel_context, we might pre-fetch and then execute the pre-update
4956 * instruction. To avoid this, the users of self-modifying code should either
4957 * disable the parser around the code emitting the memory writes, via a new flag
4958 * added to MI_ARB_CHECK, or emit the writes from a different intel_context. For
4959 * the in-kernel use-cases we've opted to use a separate context, see
4960 * reloc_gpu() as an example.
4961 * All the above applies only to the instructions themselves. Non-inline data
4962 * used by the instructions is not pre-fetched.
4965 static u32 *gen12_emit_preempt_busywait(struct i915_request *request, u32 *cs)
4967 *cs++ = MI_SEMAPHORE_WAIT_TOKEN |
4968 MI_SEMAPHORE_GLOBAL_GTT |
4970 MI_SEMAPHORE_SAD_EQ_SDD;
4972 *cs++ = intel_hws_preempt_address(request->engine);
4980 static __always_inline u32*
4981 gen12_emit_fini_breadcrumb_tail(struct i915_request *request, u32 *cs)
4983 *cs++ = MI_USER_INTERRUPT;
4985 *cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
4986 if (intel_engine_has_semaphores(request->engine))
4987 cs = gen12_emit_preempt_busywait(request, cs);
4989 request->tail = intel_ring_offset(request, cs);
4990 assert_ring_tail_valid(request->ring, request->tail);
4992 return gen8_emit_wa_tail(request, cs);
4995 static u32 *gen12_emit_fini_breadcrumb(struct i915_request *rq, u32 *cs)
4997 /* XXX Stalling flush before seqno write; post-sync not */
4998 cs = emit_xcs_breadcrumb(rq, __gen8_emit_flush_dw(cs, 0, 0, 0));
4999 return gen12_emit_fini_breadcrumb_tail(rq, cs);
5003 gen12_emit_fini_breadcrumb_rcs(struct i915_request *request, u32 *cs)
5005 cs = gen12_emit_ggtt_write_rcs(cs,
5006 request->fence.seqno,
5007 hwsp_offset(request),
5008 PIPE_CONTROL0_HDC_PIPELINE_FLUSH,
5009 PIPE_CONTROL_CS_STALL |
5010 PIPE_CONTROL_TILE_CACHE_FLUSH |
5011 PIPE_CONTROL_FLUSH_L3 |
5012 PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
5013 PIPE_CONTROL_DEPTH_CACHE_FLUSH |
5014 /* Wa_1409600907:tgl */
5015 PIPE_CONTROL_DEPTH_STALL |
5016 PIPE_CONTROL_DC_FLUSH_ENABLE |
5017 PIPE_CONTROL_FLUSH_ENABLE);
5019 return gen12_emit_fini_breadcrumb_tail(request, cs);
5022 static void execlists_park(struct intel_engine_cs *engine)
5024 cancel_timer(&engine->execlists.timer);
5025 cancel_timer(&engine->execlists.preempt);
5028 void intel_execlists_set_default_submission(struct intel_engine_cs *engine)
5030 engine->submit_request = execlists_submit_request;
5031 engine->schedule = i915_schedule;
5032 engine->execlists.tasklet.func = execlists_submission_tasklet;
5034 engine->reset.prepare = execlists_reset_prepare;
5035 engine->reset.rewind = execlists_reset_rewind;
5036 engine->reset.cancel = execlists_reset_cancel;
5037 engine->reset.finish = execlists_reset_finish;
5039 engine->park = execlists_park;
5040 engine->unpark = NULL;
5042 engine->flags |= I915_ENGINE_SUPPORTS_STATS;
5043 if (!intel_vgpu_active(engine->i915)) {
5044 engine->flags |= I915_ENGINE_HAS_SEMAPHORES;
5045 if (HAS_LOGICAL_RING_PREEMPTION(engine->i915)) {
5046 engine->flags |= I915_ENGINE_HAS_PREEMPTION;
5047 if (IS_ACTIVE(CONFIG_DRM_I915_TIMESLICE_DURATION))
5048 engine->flags |= I915_ENGINE_HAS_TIMESLICES;
5052 if (INTEL_GEN(engine->i915) >= 12)
5053 engine->flags |= I915_ENGINE_HAS_RELATIVE_MMIO;
5055 if (intel_engine_has_preemption(engine))
5056 engine->emit_bb_start = gen8_emit_bb_start;
5058 engine->emit_bb_start = gen8_emit_bb_start_noarb;
5061 static void execlists_shutdown(struct intel_engine_cs *engine)
5063 /* Synchronise with residual timers and any softirq they raise */
5064 del_timer_sync(&engine->execlists.timer);
5065 del_timer_sync(&engine->execlists.preempt);
5066 tasklet_kill(&engine->execlists.tasklet);
5069 static void execlists_release(struct intel_engine_cs *engine)
5071 engine->sanitize = NULL; /* no longer in control, nothing to sanitize */
5073 execlists_shutdown(engine);
5075 intel_engine_cleanup_common(engine);
5076 lrc_destroy_wa_ctx(engine);
5080 logical_ring_default_vfuncs(struct intel_engine_cs *engine)
5082 /* Default vfuncs which can be overriden by each engine. */
5084 engine->resume = execlists_resume;
5086 engine->cops = &execlists_context_ops;
5087 engine->request_alloc = execlists_request_alloc;
5089 engine->emit_flush = gen8_emit_flush;
5090 engine->emit_init_breadcrumb = gen8_emit_init_breadcrumb;
5091 engine->emit_fini_breadcrumb = gen8_emit_fini_breadcrumb;
5092 if (INTEL_GEN(engine->i915) >= 12) {
5093 engine->emit_fini_breadcrumb = gen12_emit_fini_breadcrumb;
5094 engine->emit_flush = gen12_emit_flush;
5096 engine->set_default_submission = intel_execlists_set_default_submission;
5098 if (INTEL_GEN(engine->i915) < 11) {
5099 engine->irq_enable = gen8_logical_ring_enable_irq;
5100 engine->irq_disable = gen8_logical_ring_disable_irq;
5103 * TODO: On Gen11 interrupt masks need to be clear
5104 * to allow C6 entry. Keep interrupts enabled at
5105 * and take the hit of generating extra interrupts
5106 * until a more refined solution exists.
5112 logical_ring_default_irqs(struct intel_engine_cs *engine)
5114 unsigned int shift = 0;
5116 if (INTEL_GEN(engine->i915) < 11) {
5117 const u8 irq_shifts[] = {
5118 [RCS0] = GEN8_RCS_IRQ_SHIFT,
5119 [BCS0] = GEN8_BCS_IRQ_SHIFT,
5120 [VCS0] = GEN8_VCS0_IRQ_SHIFT,
5121 [VCS1] = GEN8_VCS1_IRQ_SHIFT,
5122 [VECS0] = GEN8_VECS_IRQ_SHIFT,
5125 shift = irq_shifts[engine->id];
5128 engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT << shift;
5129 engine->irq_keep_mask = GT_CONTEXT_SWITCH_INTERRUPT << shift;
5130 engine->irq_keep_mask |= GT_CS_MASTER_ERROR_INTERRUPT << shift;
5131 engine->irq_keep_mask |= GT_WAIT_SEMAPHORE_INTERRUPT << shift;
5134 static void rcs_submission_override(struct intel_engine_cs *engine)
5136 switch (INTEL_GEN(engine->i915)) {
5138 engine->emit_flush = gen12_emit_flush_render;
5139 engine->emit_fini_breadcrumb = gen12_emit_fini_breadcrumb_rcs;
5142 engine->emit_flush = gen11_emit_flush_render;
5143 engine->emit_fini_breadcrumb = gen11_emit_fini_breadcrumb_rcs;
5146 engine->emit_flush = gen8_emit_flush_render;
5147 engine->emit_fini_breadcrumb = gen8_emit_fini_breadcrumb_rcs;
5152 int intel_execlists_submission_setup(struct intel_engine_cs *engine)
5154 struct intel_engine_execlists * const execlists = &engine->execlists;
5155 struct drm_i915_private *i915 = engine->i915;
5156 struct intel_uncore *uncore = engine->uncore;
5157 u32 base = engine->mmio_base;
5159 tasklet_init(&engine->execlists.tasklet,
5160 execlists_submission_tasklet, (unsigned long)engine);
5161 timer_setup(&engine->execlists.timer, execlists_timeslice, 0);
5162 timer_setup(&engine->execlists.preempt, execlists_preempt, 0);
5164 logical_ring_default_vfuncs(engine);
5165 logical_ring_default_irqs(engine);
5167 if (engine->class == RENDER_CLASS)
5168 rcs_submission_override(engine);
5170 if (intel_init_workaround_bb(engine))
5172 * We continue even if we fail to initialize WA batch
5173 * because we only expect rare glitches but nothing
5174 * critical to prevent us from using GPU
5176 drm_err(&i915->drm, "WA batch buffer initialization failed\n");
5178 if (HAS_LOGICAL_RING_ELSQ(i915)) {
5179 execlists->submit_reg = uncore->regs +
5180 i915_mmio_reg_offset(RING_EXECLIST_SQ_CONTENTS(base));
5181 execlists->ctrl_reg = uncore->regs +
5182 i915_mmio_reg_offset(RING_EXECLIST_CONTROL(base));
5184 execlists->submit_reg = uncore->regs +
5185 i915_mmio_reg_offset(RING_ELSP(base));
5188 execlists->csb_status =
5189 (u64 *)&engine->status_page.addr[I915_HWS_CSB_BUF0_INDEX];
5191 execlists->csb_write =
5192 &engine->status_page.addr[intel_hws_csb_write_index(i915)];
5194 if (INTEL_GEN(i915) < 11)
5195 execlists->csb_size = GEN8_CSB_ENTRIES;
5197 execlists->csb_size = GEN11_CSB_ENTRIES;
5199 if (INTEL_GEN(engine->i915) >= 11) {
5200 execlists->ccid |= engine->instance << (GEN11_ENGINE_INSTANCE_SHIFT - 32);
5201 execlists->ccid |= engine->class << (GEN11_ENGINE_CLASS_SHIFT - 32);
5204 /* Finally, take ownership and responsibility for cleanup! */
5205 engine->sanitize = execlists_sanitize;
5206 engine->release = execlists_release;
5211 static void init_common_reg_state(u32 * const regs,
5212 const struct intel_engine_cs *engine,
5213 const struct intel_ring *ring,
5218 ctl = _MASKED_BIT_ENABLE(CTX_CTRL_INHIBIT_SYN_CTX_SWITCH);
5219 ctl |= _MASKED_BIT_DISABLE(CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT);
5221 ctl |= CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT;
5222 if (INTEL_GEN(engine->i915) < 11)
5223 ctl |= _MASKED_BIT_DISABLE(CTX_CTRL_ENGINE_CTX_SAVE_INHIBIT |
5224 CTX_CTRL_RS_CTX_ENABLE);
5225 regs[CTX_CONTEXT_CONTROL] = ctl;
5227 regs[CTX_RING_CTL] = RING_CTL_SIZE(ring->size) | RING_VALID;
5228 regs[CTX_TIMESTAMP] = 0;
5231 static void init_wa_bb_reg_state(u32 * const regs,
5232 const struct intel_engine_cs *engine)
5234 const struct i915_ctx_workarounds * const wa_ctx = &engine->wa_ctx;
5236 if (wa_ctx->per_ctx.size) {
5237 const u32 ggtt_offset = i915_ggtt_offset(wa_ctx->vma);
5239 GEM_BUG_ON(lrc_ring_wa_bb_per_ctx(engine) == -1);
5240 regs[lrc_ring_wa_bb_per_ctx(engine) + 1] =
5241 (ggtt_offset + wa_ctx->per_ctx.offset) | 0x01;
5244 if (wa_ctx->indirect_ctx.size) {
5245 lrc_ring_setup_indirect_ctx(regs, engine,
5246 i915_ggtt_offset(wa_ctx->vma) +
5247 wa_ctx->indirect_ctx.offset,
5248 wa_ctx->indirect_ctx.size);
5252 static void init_ppgtt_reg_state(u32 *regs, const struct i915_ppgtt *ppgtt)
5254 if (i915_vm_is_4lvl(&ppgtt->vm)) {
5255 /* 64b PPGTT (48bit canonical)
5256 * PDP0_DESCRIPTOR contains the base address to PML4 and
5257 * other PDP Descriptors are ignored.
5259 ASSIGN_CTX_PML4(ppgtt, regs);
5261 ASSIGN_CTX_PDP(ppgtt, regs, 3);
5262 ASSIGN_CTX_PDP(ppgtt, regs, 2);
5263 ASSIGN_CTX_PDP(ppgtt, regs, 1);
5264 ASSIGN_CTX_PDP(ppgtt, regs, 0);
5268 static struct i915_ppgtt *vm_alias(struct i915_address_space *vm)
5270 if (i915_is_ggtt(vm))
5271 return i915_vm_to_ggtt(vm)->alias;
5273 return i915_vm_to_ppgtt(vm);
5276 static void execlists_init_reg_state(u32 *regs,
5277 const struct intel_context *ce,
5278 const struct intel_engine_cs *engine,
5279 const struct intel_ring *ring,
5283 * A context is actually a big batch buffer with several
5284 * MI_LOAD_REGISTER_IMM commands followed by (reg, value) pairs. The
5285 * values we are setting here are only for the first context restore:
5286 * on a subsequent save, the GPU will recreate this batchbuffer with new
5287 * values (including all the missing MI_LOAD_REGISTER_IMM commands that
5288 * we are not initializing here).
5290 * Must keep consistent with virtual_update_register_offsets().
5292 set_offsets(regs, reg_offsets(engine), engine, inhibit);
5294 init_common_reg_state(regs, engine, ring, inhibit);
5295 init_ppgtt_reg_state(regs, vm_alias(ce->vm));
5297 init_wa_bb_reg_state(regs, engine);
5299 __reset_stop_ring(regs, engine);
5303 populate_lr_context(struct intel_context *ce,
5304 struct drm_i915_gem_object *ctx_obj,
5305 struct intel_engine_cs *engine,
5306 struct intel_ring *ring)
5308 bool inhibit = true;
5311 vaddr = i915_gem_object_pin_map(ctx_obj, I915_MAP_WB);
5312 if (IS_ERR(vaddr)) {
5313 drm_dbg(&engine->i915->drm, "Could not map object pages!\n");
5314 return PTR_ERR(vaddr);
5317 set_redzone(vaddr, engine);
5319 if (engine->default_state) {
5320 shmem_read(engine->default_state, 0,
5321 vaddr, engine->context_size);
5322 __set_bit(CONTEXT_VALID_BIT, &ce->flags);
5326 /* Clear the ppHWSP (inc. per-context counters) */
5327 memset(vaddr, 0, PAGE_SIZE);
5330 * The second page of the context object contains some registers which
5331 * must be set up prior to the first execution.
5333 execlists_init_reg_state(vaddr + LRC_STATE_OFFSET,
5334 ce, engine, ring, inhibit);
5336 __i915_gem_object_flush_map(ctx_obj, 0, engine->context_size);
5337 i915_gem_object_unpin_map(ctx_obj);
5341 static struct intel_timeline *pinned_timeline(struct intel_context *ce)
5343 struct intel_timeline *tl = fetch_and_zero(&ce->timeline);
5345 return intel_timeline_create_from_engine(ce->engine,
5346 page_unmask_bits(tl));
5349 static int __execlists_context_alloc(struct intel_context *ce,
5350 struct intel_engine_cs *engine)
5352 struct drm_i915_gem_object *ctx_obj;
5353 struct intel_ring *ring;
5354 struct i915_vma *vma;
5358 GEM_BUG_ON(ce->state);
5359 context_size = round_up(engine->context_size, I915_GTT_PAGE_SIZE);
5361 if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
5362 context_size += I915_GTT_PAGE_SIZE; /* for redzone */
5364 if (INTEL_GEN(engine->i915) == 12) {
5365 ce->wa_bb_page = context_size / PAGE_SIZE;
5366 context_size += PAGE_SIZE;
5369 ctx_obj = i915_gem_object_create_shmem(engine->i915, context_size);
5370 if (IS_ERR(ctx_obj))
5371 return PTR_ERR(ctx_obj);
5373 vma = i915_vma_instance(ctx_obj, &engine->gt->ggtt->vm, NULL);
5376 goto error_deref_obj;
5379 if (!page_mask_bits(ce->timeline)) {
5380 struct intel_timeline *tl;
5383 * Use the static global HWSP for the kernel context, and
5384 * a dynamically allocated cacheline for everyone else.
5386 if (unlikely(ce->timeline))
5387 tl = pinned_timeline(ce);
5389 tl = intel_timeline_create(engine->gt);
5392 goto error_deref_obj;
5398 ring = intel_engine_create_ring(engine, (unsigned long)ce->ring);
5400 ret = PTR_ERR(ring);
5401 goto error_deref_obj;
5404 ret = populate_lr_context(ce, ctx_obj, engine, ring);
5406 drm_dbg(&engine->i915->drm,
5407 "Failed to populate LRC: %d\n", ret);
5408 goto error_ring_free;
5417 intel_ring_put(ring);
5419 i915_gem_object_put(ctx_obj);
5423 static struct list_head *virtual_queue(struct virtual_engine *ve)
5425 return &ve->base.execlists.default_priolist.requests[0];
5428 static void virtual_context_destroy(struct kref *kref)
5430 struct virtual_engine *ve =
5431 container_of(kref, typeof(*ve), context.ref);
5434 GEM_BUG_ON(!list_empty(virtual_queue(ve)));
5435 GEM_BUG_ON(ve->request);
5436 GEM_BUG_ON(ve->context.inflight);
5438 for (n = 0; n < ve->num_siblings; n++) {
5439 struct intel_engine_cs *sibling = ve->siblings[n];
5440 struct rb_node *node = &ve->nodes[sibling->id].rb;
5441 unsigned long flags;
5443 if (RB_EMPTY_NODE(node))
5446 spin_lock_irqsave(&sibling->active.lock, flags);
5448 /* Detachment is lazily performed in the execlists tasklet */
5449 if (!RB_EMPTY_NODE(node))
5450 rb_erase_cached(node, &sibling->execlists.virtual);
5452 spin_unlock_irqrestore(&sibling->active.lock, flags);
5454 GEM_BUG_ON(__tasklet_is_scheduled(&ve->base.execlists.tasklet));
5456 if (ve->context.state)
5457 __execlists_context_fini(&ve->context);
5458 intel_context_fini(&ve->context);
5460 intel_engine_free_request_pool(&ve->base);
5466 static void virtual_engine_initial_hint(struct virtual_engine *ve)
5471 * Pick a random sibling on starting to help spread the load around.
5473 * New contexts are typically created with exactly the same order
5474 * of siblings, and often started in batches. Due to the way we iterate
5475 * the array of sibling when submitting requests, sibling[0] is
5476 * prioritised for dequeuing. If we make sure that sibling[0] is fairly
5477 * randomised across the system, we also help spread the load by the
5478 * first engine we inspect being different each time.
5480 * NB This does not force us to execute on this engine, it will just
5481 * typically be the first we inspect for submission.
5483 swp = prandom_u32_max(ve->num_siblings);
5485 swap(ve->siblings[swp], ve->siblings[0]);
5488 static int virtual_context_alloc(struct intel_context *ce)
5490 struct virtual_engine *ve = container_of(ce, typeof(*ve), context);
5492 return __execlists_context_alloc(ce, ve->siblings[0]);
5495 static int virtual_context_pin(struct intel_context *ce, void *vaddr)
5497 struct virtual_engine *ve = container_of(ce, typeof(*ve), context);
5499 /* Note: we must use a real engine class for setting up reg state */
5500 return __execlists_context_pin(ce, ve->siblings[0], vaddr);
5503 static void virtual_context_enter(struct intel_context *ce)
5505 struct virtual_engine *ve = container_of(ce, typeof(*ve), context);
5508 for (n = 0; n < ve->num_siblings; n++)
5509 intel_engine_pm_get(ve->siblings[n]);
5511 intel_timeline_enter(ce->timeline);
5514 static void virtual_context_exit(struct intel_context *ce)
5516 struct virtual_engine *ve = container_of(ce, typeof(*ve), context);
5519 intel_timeline_exit(ce->timeline);
5521 for (n = 0; n < ve->num_siblings; n++)
5522 intel_engine_pm_put(ve->siblings[n]);
5525 static const struct intel_context_ops virtual_context_ops = {
5526 .alloc = virtual_context_alloc,
5528 .pre_pin = execlists_context_pre_pin,
5529 .pin = virtual_context_pin,
5530 .unpin = execlists_context_unpin,
5531 .post_unpin = execlists_context_post_unpin,
5533 .enter = virtual_context_enter,
5534 .exit = virtual_context_exit,
5536 .destroy = virtual_context_destroy,
5539 static intel_engine_mask_t virtual_submission_mask(struct virtual_engine *ve)
5541 struct i915_request *rq;
5542 intel_engine_mask_t mask;
5544 rq = READ_ONCE(ve->request);
5548 /* The rq is ready for submission; rq->execution_mask is now stable. */
5549 mask = rq->execution_mask;
5550 if (unlikely(!mask)) {
5551 /* Invalid selection, submit to a random engine in error */
5552 i915_request_set_error_once(rq, -ENODEV);
5553 mask = ve->siblings[0]->mask;
5556 ENGINE_TRACE(&ve->base, "rq=%llx:%lld, mask=%x, prio=%d\n",
5557 rq->fence.context, rq->fence.seqno,
5558 mask, ve->base.execlists.queue_priority_hint);
5563 static void virtual_submission_tasklet(unsigned long data)
5565 struct virtual_engine * const ve = (struct virtual_engine *)data;
5566 const int prio = READ_ONCE(ve->base.execlists.queue_priority_hint);
5567 intel_engine_mask_t mask;
5571 mask = virtual_submission_mask(ve);
5573 if (unlikely(!mask))
5576 local_irq_disable();
5577 for (n = 0; n < ve->num_siblings; n++) {
5578 struct intel_engine_cs *sibling = READ_ONCE(ve->siblings[n]);
5579 struct ve_node * const node = &ve->nodes[sibling->id];
5580 struct rb_node **parent, *rb;
5583 if (!READ_ONCE(ve->request))
5584 break; /* already handled by a sibling's tasklet */
5586 if (unlikely(!(mask & sibling->mask))) {
5587 if (!RB_EMPTY_NODE(&node->rb)) {
5588 spin_lock(&sibling->active.lock);
5589 rb_erase_cached(&node->rb,
5590 &sibling->execlists.virtual);
5591 RB_CLEAR_NODE(&node->rb);
5592 spin_unlock(&sibling->active.lock);
5597 spin_lock(&sibling->active.lock);
5599 if (!RB_EMPTY_NODE(&node->rb)) {
5601 * Cheat and avoid rebalancing the tree if we can
5602 * reuse this node in situ.
5604 first = rb_first_cached(&sibling->execlists.virtual) ==
5606 if (prio == node->prio || (prio > node->prio && first))
5609 rb_erase_cached(&node->rb, &sibling->execlists.virtual);
5614 parent = &sibling->execlists.virtual.rb_root.rb_node;
5616 struct ve_node *other;
5619 other = rb_entry(rb, typeof(*other), rb);
5620 if (prio > other->prio) {
5621 parent = &rb->rb_left;
5623 parent = &rb->rb_right;
5628 rb_link_node(&node->rb, rb, parent);
5629 rb_insert_color_cached(&node->rb,
5630 &sibling->execlists.virtual,
5634 GEM_BUG_ON(RB_EMPTY_NODE(&node->rb));
5636 if (first && prio > sibling->execlists.queue_priority_hint)
5637 tasklet_hi_schedule(&sibling->execlists.tasklet);
5639 spin_unlock(&sibling->active.lock);
5644 static void virtual_submit_request(struct i915_request *rq)
5646 struct virtual_engine *ve = to_virtual_engine(rq->engine);
5647 struct i915_request *old;
5648 unsigned long flags;
5650 ENGINE_TRACE(&ve->base, "rq=%llx:%lld\n",
5654 GEM_BUG_ON(ve->base.submit_request != virtual_submit_request);
5656 spin_lock_irqsave(&ve->base.active.lock, flags);
5659 if (old) { /* background completion event from preempt-to-busy */
5660 GEM_BUG_ON(!i915_request_completed(old));
5661 __i915_request_submit(old);
5662 i915_request_put(old);
5665 if (i915_request_completed(rq)) {
5666 __i915_request_submit(rq);
5668 ve->base.execlists.queue_priority_hint = INT_MIN;
5671 ve->base.execlists.queue_priority_hint = rq_prio(rq);
5672 ve->request = i915_request_get(rq);
5674 GEM_BUG_ON(!list_empty(virtual_queue(ve)));
5675 list_move_tail(&rq->sched.link, virtual_queue(ve));
5677 tasklet_hi_schedule(&ve->base.execlists.tasklet);
5680 spin_unlock_irqrestore(&ve->base.active.lock, flags);
5683 static struct ve_bond *
5684 virtual_find_bond(struct virtual_engine *ve,
5685 const struct intel_engine_cs *master)
5689 for (i = 0; i < ve->num_bonds; i++) {
5690 if (ve->bonds[i].master == master)
5691 return &ve->bonds[i];
5698 virtual_bond_execute(struct i915_request *rq, struct dma_fence *signal)
5700 struct virtual_engine *ve = to_virtual_engine(rq->engine);
5701 intel_engine_mask_t allowed, exec;
5702 struct ve_bond *bond;
5704 allowed = ~to_request(signal)->engine->mask;
5706 bond = virtual_find_bond(ve, to_request(signal)->engine);
5708 allowed &= bond->sibling_mask;
5710 /* Restrict the bonded request to run on only the available engines */
5711 exec = READ_ONCE(rq->execution_mask);
5712 while (!try_cmpxchg(&rq->execution_mask, &exec, exec & allowed))
5715 /* Prevent the master from being re-run on the bonded engines */
5716 to_request(signal)->execution_mask &= ~allowed;
5719 struct intel_context *
5720 intel_execlists_create_virtual(struct intel_engine_cs **siblings,
5723 struct virtual_engine *ve;
5728 return ERR_PTR(-EINVAL);
5731 return intel_context_create(siblings[0]);
5733 ve = kzalloc(struct_size(ve, siblings, count), GFP_KERNEL);
5735 return ERR_PTR(-ENOMEM);
5737 ve->base.i915 = siblings[0]->i915;
5738 ve->base.gt = siblings[0]->gt;
5739 ve->base.uncore = siblings[0]->uncore;
5742 ve->base.class = OTHER_CLASS;
5743 ve->base.uabi_class = I915_ENGINE_CLASS_INVALID;
5744 ve->base.instance = I915_ENGINE_CLASS_INVALID_VIRTUAL;
5745 ve->base.uabi_instance = I915_ENGINE_CLASS_INVALID_VIRTUAL;
5748 * The decision on whether to submit a request using semaphores
5749 * depends on the saturated state of the engine. We only compute
5750 * this during HW submission of the request, and we need for this
5751 * state to be globally applied to all requests being submitted
5752 * to this engine. Virtual engines encompass more than one physical
5753 * engine and so we cannot accurately tell in advance if one of those
5754 * engines is already saturated and so cannot afford to use a semaphore
5755 * and be pessimized in priority for doing so -- if we are the only
5756 * context using semaphores after all other clients have stopped, we
5757 * will be starved on the saturated system. Such a global switch for
5758 * semaphores is less than ideal, but alas is the current compromise.
5760 ve->base.saturated = ALL_ENGINES;
5762 snprintf(ve->base.name, sizeof(ve->base.name), "virtual");
5764 intel_engine_init_active(&ve->base, ENGINE_VIRTUAL);
5765 intel_engine_init_execlists(&ve->base);
5767 ve->base.cops = &virtual_context_ops;
5768 ve->base.request_alloc = execlists_request_alloc;
5770 ve->base.schedule = i915_schedule;
5771 ve->base.submit_request = virtual_submit_request;
5772 ve->base.bond_execute = virtual_bond_execute;
5774 INIT_LIST_HEAD(virtual_queue(ve));
5775 ve->base.execlists.queue_priority_hint = INT_MIN;
5776 tasklet_init(&ve->base.execlists.tasklet,
5777 virtual_submission_tasklet,
5780 intel_context_init(&ve->context, &ve->base);
5782 ve->base.breadcrumbs = intel_breadcrumbs_create(NULL);
5783 if (!ve->base.breadcrumbs) {
5788 for (n = 0; n < count; n++) {
5789 struct intel_engine_cs *sibling = siblings[n];
5791 GEM_BUG_ON(!is_power_of_2(sibling->mask));
5792 if (sibling->mask & ve->base.mask) {
5793 DRM_DEBUG("duplicate %s entry in load balancer\n",
5800 * The virtual engine implementation is tightly coupled to
5801 * the execlists backend -- we push out request directly
5802 * into a tree inside each physical engine. We could support
5803 * layering if we handle cloning of the requests and
5804 * submitting a copy into each backend.
5806 if (sibling->execlists.tasklet.func !=
5807 execlists_submission_tasklet) {
5812 GEM_BUG_ON(RB_EMPTY_NODE(&ve->nodes[sibling->id].rb));
5813 RB_CLEAR_NODE(&ve->nodes[sibling->id].rb);
5815 ve->siblings[ve->num_siblings++] = sibling;
5816 ve->base.mask |= sibling->mask;
5819 * All physical engines must be compatible for their emission
5820 * functions (as we build the instructions during request
5821 * construction and do not alter them before submission
5822 * on the physical engine). We use the engine class as a guide
5823 * here, although that could be refined.
5825 if (ve->base.class != OTHER_CLASS) {
5826 if (ve->base.class != sibling->class) {
5827 DRM_DEBUG("invalid mixing of engine class, sibling %d, already %d\n",
5828 sibling->class, ve->base.class);
5835 ve->base.class = sibling->class;
5836 ve->base.uabi_class = sibling->uabi_class;
5837 snprintf(ve->base.name, sizeof(ve->base.name),
5838 "v%dx%d", ve->base.class, count);
5839 ve->base.context_size = sibling->context_size;
5841 ve->base.emit_bb_start = sibling->emit_bb_start;
5842 ve->base.emit_flush = sibling->emit_flush;
5843 ve->base.emit_init_breadcrumb = sibling->emit_init_breadcrumb;
5844 ve->base.emit_fini_breadcrumb = sibling->emit_fini_breadcrumb;
5845 ve->base.emit_fini_breadcrumb_dw =
5846 sibling->emit_fini_breadcrumb_dw;
5848 ve->base.flags = sibling->flags;
5851 ve->base.flags |= I915_ENGINE_IS_VIRTUAL;
5853 virtual_engine_initial_hint(ve);
5854 return &ve->context;
5857 intel_context_put(&ve->context);
5858 return ERR_PTR(err);
5861 struct intel_context *
5862 intel_execlists_clone_virtual(struct intel_engine_cs *src)
5864 struct virtual_engine *se = to_virtual_engine(src);
5865 struct intel_context *dst;
5867 dst = intel_execlists_create_virtual(se->siblings,
5872 if (se->num_bonds) {
5873 struct virtual_engine *de = to_virtual_engine(dst->engine);
5875 de->bonds = kmemdup(se->bonds,
5876 sizeof(*se->bonds) * se->num_bonds,
5879 intel_context_put(dst);
5880 return ERR_PTR(-ENOMEM);
5883 de->num_bonds = se->num_bonds;
5889 int intel_virtual_engine_attach_bond(struct intel_engine_cs *engine,
5890 const struct intel_engine_cs *master,
5891 const struct intel_engine_cs *sibling)
5893 struct virtual_engine *ve = to_virtual_engine(engine);
5894 struct ve_bond *bond;
5897 /* Sanity check the sibling is part of the virtual engine */
5898 for (n = 0; n < ve->num_siblings; n++)
5899 if (sibling == ve->siblings[n])
5901 if (n == ve->num_siblings)
5904 bond = virtual_find_bond(ve, master);
5906 bond->sibling_mask |= sibling->mask;
5910 bond = krealloc(ve->bonds,
5911 sizeof(*bond) * (ve->num_bonds + 1),
5916 bond[ve->num_bonds].master = master;
5917 bond[ve->num_bonds].sibling_mask = sibling->mask;
5925 struct intel_engine_cs *
5926 intel_virtual_engine_get_sibling(struct intel_engine_cs *engine,
5927 unsigned int sibling)
5929 struct virtual_engine *ve = to_virtual_engine(engine);
5931 if (sibling >= ve->num_siblings)
5934 return ve->siblings[sibling];
5937 void intel_execlists_show_requests(struct intel_engine_cs *engine,
5938 struct drm_printer *m,
5939 void (*show_request)(struct drm_printer *m,
5940 struct i915_request *rq,
5941 const char *prefix),
5944 const struct intel_engine_execlists *execlists = &engine->execlists;
5945 struct i915_request *rq, *last;
5946 unsigned long flags;
5950 spin_lock_irqsave(&engine->active.lock, flags);
5954 list_for_each_entry(rq, &engine->active.requests, sched.link) {
5955 if (count++ < max - 1)
5956 show_request(m, rq, "\t\tE ");
5963 "\t\t...skipping %d executing requests...\n",
5966 show_request(m, last, "\t\tE ");
5969 if (execlists->switch_priority_hint != INT_MIN)
5970 drm_printf(m, "\t\tSwitch priority hint: %d\n",
5971 READ_ONCE(execlists->switch_priority_hint));
5972 if (execlists->queue_priority_hint != INT_MIN)
5973 drm_printf(m, "\t\tQueue priority hint: %d\n",
5974 READ_ONCE(execlists->queue_priority_hint));
5978 for (rb = rb_first_cached(&execlists->queue); rb; rb = rb_next(rb)) {
5979 struct i915_priolist *p = rb_entry(rb, typeof(*p), node);
5982 priolist_for_each_request(rq, p, i) {
5983 if (count++ < max - 1)
5984 show_request(m, rq, "\t\tQ ");
5992 "\t\t...skipping %d queued requests...\n",
5995 show_request(m, last, "\t\tQ ");
6000 for (rb = rb_first_cached(&execlists->virtual); rb; rb = rb_next(rb)) {
6001 struct virtual_engine *ve =
6002 rb_entry(rb, typeof(*ve), nodes[engine->id].rb);
6003 struct i915_request *rq = READ_ONCE(ve->request);
6006 if (count++ < max - 1)
6007 show_request(m, rq, "\t\tV ");
6015 "\t\t...skipping %d virtual requests...\n",
6018 show_request(m, last, "\t\tV ");
6021 spin_unlock_irqrestore(&engine->active.lock, flags);
6024 void intel_lr_context_reset(struct intel_engine_cs *engine,
6025 struct intel_context *ce,
6029 GEM_BUG_ON(!intel_context_is_pinned(ce));
6032 * We want a simple context + ring to execute the breadcrumb update.
6033 * We cannot rely on the context being intact across the GPU hang,
6034 * so clear it and rebuild just what we need for the breadcrumb.
6035 * All pending requests for this context will be zapped, and any
6036 * future request will be after userspace has had the opportunity
6037 * to recreate its own state.
6040 restore_default_state(ce, engine);
6042 /* Rerun the request; its payload has been neutered (if guilty). */
6043 __execlists_update_reg_state(ce, engine, head);
6047 intel_engine_in_execlists_submission_mode(const struct intel_engine_cs *engine)
6049 return engine->set_default_submission ==
6050 intel_execlists_set_default_submission;
6053 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
6054 #include "selftest_lrc.c"