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 <drm/drmP.h>
137 #include <drm/i915_drm.h>
138 #include "i915_drv.h"
139 #include "intel_mocs.h"
141 #define GEN9_LR_CONTEXT_RENDER_SIZE (22 * PAGE_SIZE)
142 #define GEN8_LR_CONTEXT_RENDER_SIZE (20 * PAGE_SIZE)
143 #define GEN8_LR_CONTEXT_OTHER_SIZE (2 * PAGE_SIZE)
145 #define RING_EXECLIST_QFULL (1 << 0x2)
146 #define RING_EXECLIST1_VALID (1 << 0x3)
147 #define RING_EXECLIST0_VALID (1 << 0x4)
148 #define RING_EXECLIST_ACTIVE_STATUS (3 << 0xE)
149 #define RING_EXECLIST1_ACTIVE (1 << 0x11)
150 #define RING_EXECLIST0_ACTIVE (1 << 0x12)
152 #define GEN8_CTX_STATUS_IDLE_ACTIVE (1 << 0)
153 #define GEN8_CTX_STATUS_PREEMPTED (1 << 1)
154 #define GEN8_CTX_STATUS_ELEMENT_SWITCH (1 << 2)
155 #define GEN8_CTX_STATUS_ACTIVE_IDLE (1 << 3)
156 #define GEN8_CTX_STATUS_COMPLETE (1 << 4)
157 #define GEN8_CTX_STATUS_LITE_RESTORE (1 << 15)
159 #define CTX_LRI_HEADER_0 0x01
160 #define CTX_CONTEXT_CONTROL 0x02
161 #define CTX_RING_HEAD 0x04
162 #define CTX_RING_TAIL 0x06
163 #define CTX_RING_BUFFER_START 0x08
164 #define CTX_RING_BUFFER_CONTROL 0x0a
165 #define CTX_BB_HEAD_U 0x0c
166 #define CTX_BB_HEAD_L 0x0e
167 #define CTX_BB_STATE 0x10
168 #define CTX_SECOND_BB_HEAD_U 0x12
169 #define CTX_SECOND_BB_HEAD_L 0x14
170 #define CTX_SECOND_BB_STATE 0x16
171 #define CTX_BB_PER_CTX_PTR 0x18
172 #define CTX_RCS_INDIRECT_CTX 0x1a
173 #define CTX_RCS_INDIRECT_CTX_OFFSET 0x1c
174 #define CTX_LRI_HEADER_1 0x21
175 #define CTX_CTX_TIMESTAMP 0x22
176 #define CTX_PDP3_UDW 0x24
177 #define CTX_PDP3_LDW 0x26
178 #define CTX_PDP2_UDW 0x28
179 #define CTX_PDP2_LDW 0x2a
180 #define CTX_PDP1_UDW 0x2c
181 #define CTX_PDP1_LDW 0x2e
182 #define CTX_PDP0_UDW 0x30
183 #define CTX_PDP0_LDW 0x32
184 #define CTX_LRI_HEADER_2 0x41
185 #define CTX_R_PWR_CLK_STATE 0x42
186 #define CTX_GPGPU_CSR_BASE_ADDRESS 0x44
188 #define GEN8_CTX_VALID (1<<0)
189 #define GEN8_CTX_FORCE_PD_RESTORE (1<<1)
190 #define GEN8_CTX_FORCE_RESTORE (1<<2)
191 #define GEN8_CTX_L3LLC_COHERENT (1<<5)
192 #define GEN8_CTX_PRIVILEGE (1<<8)
194 #define ASSIGN_CTX_REG(reg_state, pos, reg, val) do { \
195 (reg_state)[(pos)+0] = i915_mmio_reg_offset(reg); \
196 (reg_state)[(pos)+1] = (val); \
199 #define ASSIGN_CTX_PDP(ppgtt, reg_state, n) do { \
200 const u64 _addr = i915_page_dir_dma_addr((ppgtt), (n)); \
201 reg_state[CTX_PDP ## n ## _UDW+1] = upper_32_bits(_addr); \
202 reg_state[CTX_PDP ## n ## _LDW+1] = lower_32_bits(_addr); \
205 #define ASSIGN_CTX_PML4(ppgtt, reg_state) do { \
206 reg_state[CTX_PDP0_UDW + 1] = upper_32_bits(px_dma(&ppgtt->pml4)); \
207 reg_state[CTX_PDP0_LDW + 1] = lower_32_bits(px_dma(&ppgtt->pml4)); \
212 FAULT_AND_HALT, /* Debug only */
214 FAULT_AND_CONTINUE /* Unsupported */
216 #define GEN8_CTX_ID_SHIFT 32
217 #define GEN8_CTX_ID_WIDTH 21
218 #define GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT 0x17
219 #define GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT 0x26
221 /* Typical size of the average request (2 pipecontrols and a MI_BB) */
222 #define EXECLISTS_REQUEST_SIZE 64 /* bytes */
224 static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
225 struct intel_engine_cs *engine);
226 static int intel_lr_context_pin(struct i915_gem_context *ctx,
227 struct intel_engine_cs *engine);
230 * intel_sanitize_enable_execlists() - sanitize i915.enable_execlists
231 * @dev_priv: i915 device private
232 * @enable_execlists: value of i915.enable_execlists module parameter.
234 * Only certain platforms support Execlists (the prerequisites being
235 * support for Logical Ring Contexts and Aliasing PPGTT or better).
237 * Return: 1 if Execlists is supported and has to be enabled.
239 int intel_sanitize_enable_execlists(struct drm_i915_private *dev_priv, int enable_execlists)
241 /* On platforms with execlist available, vGPU will only
242 * support execlist mode, no ring buffer mode.
244 if (HAS_LOGICAL_RING_CONTEXTS(dev_priv) && intel_vgpu_active(dev_priv))
247 if (INTEL_GEN(dev_priv) >= 9)
250 if (enable_execlists == 0)
253 if (HAS_LOGICAL_RING_CONTEXTS(dev_priv) &&
254 USES_PPGTT(dev_priv) &&
255 i915.use_mmio_flip >= 0)
262 logical_ring_init_platform_invariants(struct intel_engine_cs *engine)
264 struct drm_i915_private *dev_priv = engine->i915;
266 if (IS_GEN8(dev_priv) || IS_GEN9(dev_priv))
267 engine->idle_lite_restore_wa = ~0;
269 engine->disable_lite_restore_wa = (IS_SKL_REVID(dev_priv, 0, SKL_REVID_B0) ||
270 IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1)) &&
271 (engine->id == VCS || engine->id == VCS2);
273 engine->ctx_desc_template = GEN8_CTX_VALID;
274 if (IS_GEN8(dev_priv))
275 engine->ctx_desc_template |= GEN8_CTX_L3LLC_COHERENT;
276 engine->ctx_desc_template |= GEN8_CTX_PRIVILEGE;
278 /* TODO: WaDisableLiteRestore when we start using semaphore
279 * signalling between Command Streamers */
280 /* ring->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE; */
282 /* WaEnableForceRestoreInCtxtDescForVCS:skl */
283 /* WaEnableForceRestoreInCtxtDescForVCS:bxt */
284 if (engine->disable_lite_restore_wa)
285 engine->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE;
289 * intel_lr_context_descriptor_update() - calculate & cache the descriptor
290 * descriptor for a pinned context
292 * @ctx: Context to work on
293 * @engine: Engine the descriptor will be used with
295 * The context descriptor encodes various attributes of a context,
296 * including its GTT address and some flags. Because it's fairly
297 * expensive to calculate, we'll just do it once and cache the result,
298 * which remains valid until the context is unpinned.
300 * This is what a descriptor looks like, from LSB to MSB:
301 * bits 0-11: flags, GEN8_CTX_* (cached in ctx_desc_template)
302 * bits 12-31: LRCA, GTT address of (the HWSP of) this context
303 * bits 32-52: ctx ID, a globally unique tag
304 * bits 53-54: mbz, reserved for use by hardware
305 * bits 55-63: group ID, currently unused and set to 0
308 intel_lr_context_descriptor_update(struct i915_gem_context *ctx,
309 struct intel_engine_cs *engine)
311 struct intel_context *ce = &ctx->engine[engine->id];
314 BUILD_BUG_ON(MAX_CONTEXT_HW_ID > (1<<GEN8_CTX_ID_WIDTH));
316 desc = ctx->desc_template; /* bits 3-4 */
317 desc |= engine->ctx_desc_template; /* bits 0-11 */
318 desc |= ce->lrc_vma->node.start + LRC_PPHWSP_PN * PAGE_SIZE;
320 desc |= (u64)ctx->hw_id << GEN8_CTX_ID_SHIFT; /* bits 32-52 */
325 uint64_t intel_lr_context_descriptor(struct i915_gem_context *ctx,
326 struct intel_engine_cs *engine)
328 return ctx->engine[engine->id].lrc_desc;
331 static void execlists_elsp_write(struct drm_i915_gem_request *rq0,
332 struct drm_i915_gem_request *rq1)
335 struct intel_engine_cs *engine = rq0->engine;
336 struct drm_i915_private *dev_priv = rq0->i915;
340 desc[1] = intel_lr_context_descriptor(rq1->ctx, rq1->engine);
341 rq1->elsp_submitted++;
346 desc[0] = intel_lr_context_descriptor(rq0->ctx, rq0->engine);
347 rq0->elsp_submitted++;
349 /* You must always write both descriptors in the order below. */
350 I915_WRITE_FW(RING_ELSP(engine), upper_32_bits(desc[1]));
351 I915_WRITE_FW(RING_ELSP(engine), lower_32_bits(desc[1]));
353 I915_WRITE_FW(RING_ELSP(engine), upper_32_bits(desc[0]));
354 /* The context is automatically loaded after the following */
355 I915_WRITE_FW(RING_ELSP(engine), lower_32_bits(desc[0]));
357 /* ELSP is a wo register, use another nearby reg for posting */
358 POSTING_READ_FW(RING_EXECLIST_STATUS_LO(engine));
362 execlists_update_context_pdps(struct i915_hw_ppgtt *ppgtt, u32 *reg_state)
364 ASSIGN_CTX_PDP(ppgtt, reg_state, 3);
365 ASSIGN_CTX_PDP(ppgtt, reg_state, 2);
366 ASSIGN_CTX_PDP(ppgtt, reg_state, 1);
367 ASSIGN_CTX_PDP(ppgtt, reg_state, 0);
370 static void execlists_update_context(struct drm_i915_gem_request *rq)
372 struct intel_engine_cs *engine = rq->engine;
373 struct i915_hw_ppgtt *ppgtt = rq->ctx->ppgtt;
374 uint32_t *reg_state = rq->ctx->engine[engine->id].lrc_reg_state;
376 reg_state[CTX_RING_TAIL+1] = rq->tail;
378 /* True 32b PPGTT with dynamic page allocation: update PDP
379 * registers and point the unallocated PDPs to scratch page.
380 * PML4 is allocated during ppgtt init, so this is not needed
383 if (ppgtt && !USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
384 execlists_update_context_pdps(ppgtt, reg_state);
387 static void execlists_submit_requests(struct drm_i915_gem_request *rq0,
388 struct drm_i915_gem_request *rq1)
390 struct drm_i915_private *dev_priv = rq0->i915;
391 unsigned int fw_domains = rq0->engine->fw_domains;
393 execlists_update_context(rq0);
396 execlists_update_context(rq1);
398 spin_lock_irq(&dev_priv->uncore.lock);
399 intel_uncore_forcewake_get__locked(dev_priv, fw_domains);
401 execlists_elsp_write(rq0, rq1);
403 intel_uncore_forcewake_put__locked(dev_priv, fw_domains);
404 spin_unlock_irq(&dev_priv->uncore.lock);
407 static inline void execlists_context_status_change(
408 struct drm_i915_gem_request *rq,
409 unsigned long status)
412 * Only used when GVT-g is enabled now. When GVT-g is disabled,
413 * The compiler should eliminate this function as dead-code.
415 if (!IS_ENABLED(CONFIG_DRM_I915_GVT))
418 atomic_notifier_call_chain(&rq->ctx->status_notifier, status, rq);
421 static void execlists_context_unqueue(struct intel_engine_cs *engine)
423 struct drm_i915_gem_request *req0 = NULL, *req1 = NULL;
424 struct drm_i915_gem_request *cursor, *tmp;
426 assert_spin_locked(&engine->execlist_lock);
429 * If irqs are not active generate a warning as batches that finish
430 * without the irqs may get lost and a GPU Hang may occur.
432 WARN_ON(!intel_irqs_enabled(engine->i915));
434 /* Try to read in pairs */
435 list_for_each_entry_safe(cursor, tmp, &engine->execlist_queue,
439 } else if (req0->ctx == cursor->ctx) {
440 /* Same ctx: ignore first request, as second request
441 * will update tail past first request's workload */
442 cursor->elsp_submitted = req0->elsp_submitted;
443 list_del(&req0->execlist_link);
444 i915_gem_request_unreference(req0);
447 if (IS_ENABLED(CONFIG_DRM_I915_GVT)) {
449 * req0 (after merged) ctx requires single
450 * submission, stop picking
452 if (req0->ctx->execlists_force_single_submission)
455 * req0 ctx doesn't require single submission,
456 * but next req ctx requires, stop picking
458 if (cursor->ctx->execlists_force_single_submission)
462 WARN_ON(req1->elsp_submitted);
470 execlists_context_status_change(req0, INTEL_CONTEXT_SCHEDULE_IN);
473 execlists_context_status_change(req1,
474 INTEL_CONTEXT_SCHEDULE_IN);
476 if (req0->elsp_submitted & engine->idle_lite_restore_wa) {
478 * WaIdleLiteRestore: make sure we never cause a lite restore
481 * Apply the wa NOOPS to prevent ring:HEAD == req:TAIL as we
482 * resubmit the request. See gen8_emit_request() for where we
483 * prepare the padding after the end of the request.
485 struct intel_ringbuffer *ringbuf;
487 ringbuf = req0->ctx->engine[engine->id].ringbuf;
489 req0->tail &= ringbuf->size - 1;
492 execlists_submit_requests(req0, req1);
496 execlists_check_remove_request(struct intel_engine_cs *engine, u32 ctx_id)
498 struct drm_i915_gem_request *head_req;
500 assert_spin_locked(&engine->execlist_lock);
502 head_req = list_first_entry_or_null(&engine->execlist_queue,
503 struct drm_i915_gem_request,
506 if (WARN_ON(!head_req || (head_req->ctx_hw_id != ctx_id)))
509 WARN(head_req->elsp_submitted == 0, "Never submitted head request\n");
511 if (--head_req->elsp_submitted > 0)
514 execlists_context_status_change(head_req, INTEL_CONTEXT_SCHEDULE_OUT);
516 list_del(&head_req->execlist_link);
517 i915_gem_request_unreference(head_req);
523 get_context_status(struct intel_engine_cs *engine, unsigned int read_pointer,
526 struct drm_i915_private *dev_priv = engine->i915;
529 read_pointer %= GEN8_CSB_ENTRIES;
531 status = I915_READ_FW(RING_CONTEXT_STATUS_BUF_LO(engine, read_pointer));
533 if (status & GEN8_CTX_STATUS_IDLE_ACTIVE)
536 *context_id = I915_READ_FW(RING_CONTEXT_STATUS_BUF_HI(engine,
543 * intel_lrc_irq_handler() - handle Context Switch interrupts
544 * @data: tasklet handler passed in unsigned long
546 * Check the unread Context Status Buffers and manage the submission of new
547 * contexts to the ELSP accordingly.
549 static void intel_lrc_irq_handler(unsigned long data)
551 struct intel_engine_cs *engine = (struct intel_engine_cs *)data;
552 struct drm_i915_private *dev_priv = engine->i915;
554 unsigned int read_pointer, write_pointer;
555 u32 csb[GEN8_CSB_ENTRIES][2];
556 unsigned int csb_read = 0, i;
557 unsigned int submit_contexts = 0;
559 intel_uncore_forcewake_get(dev_priv, engine->fw_domains);
561 status_pointer = I915_READ_FW(RING_CONTEXT_STATUS_PTR(engine));
563 read_pointer = engine->next_context_status_buffer;
564 write_pointer = GEN8_CSB_WRITE_PTR(status_pointer);
565 if (read_pointer > write_pointer)
566 write_pointer += GEN8_CSB_ENTRIES;
568 while (read_pointer < write_pointer) {
569 if (WARN_ON_ONCE(csb_read == GEN8_CSB_ENTRIES))
571 csb[csb_read][0] = get_context_status(engine, ++read_pointer,
576 engine->next_context_status_buffer = write_pointer % GEN8_CSB_ENTRIES;
578 /* Update the read pointer to the old write pointer. Manual ringbuffer
579 * management ftw </sarcasm> */
580 I915_WRITE_FW(RING_CONTEXT_STATUS_PTR(engine),
581 _MASKED_FIELD(GEN8_CSB_READ_PTR_MASK,
582 engine->next_context_status_buffer << 8));
584 intel_uncore_forcewake_put(dev_priv, engine->fw_domains);
586 spin_lock(&engine->execlist_lock);
588 for (i = 0; i < csb_read; i++) {
589 if (unlikely(csb[i][0] & GEN8_CTX_STATUS_PREEMPTED)) {
590 if (csb[i][0] & GEN8_CTX_STATUS_LITE_RESTORE) {
591 if (execlists_check_remove_request(engine, csb[i][1]))
592 WARN(1, "Lite Restored request removed from queue\n");
594 WARN(1, "Preemption without Lite Restore\n");
597 if (csb[i][0] & (GEN8_CTX_STATUS_ACTIVE_IDLE |
598 GEN8_CTX_STATUS_ELEMENT_SWITCH))
600 execlists_check_remove_request(engine, csb[i][1]);
603 if (submit_contexts) {
604 if (!engine->disable_lite_restore_wa ||
605 (csb[i][0] & GEN8_CTX_STATUS_ACTIVE_IDLE))
606 execlists_context_unqueue(engine);
609 spin_unlock(&engine->execlist_lock);
611 if (unlikely(submit_contexts > 2))
612 DRM_ERROR("More than two context complete events?\n");
615 static void execlists_context_queue(struct drm_i915_gem_request *request)
617 struct intel_engine_cs *engine = request->engine;
618 struct drm_i915_gem_request *cursor;
619 int num_elements = 0;
621 spin_lock_bh(&engine->execlist_lock);
623 list_for_each_entry(cursor, &engine->execlist_queue, execlist_link)
624 if (++num_elements > 2)
627 if (num_elements > 2) {
628 struct drm_i915_gem_request *tail_req;
630 tail_req = list_last_entry(&engine->execlist_queue,
631 struct drm_i915_gem_request,
634 if (request->ctx == tail_req->ctx) {
635 WARN(tail_req->elsp_submitted != 0,
636 "More than 2 already-submitted reqs queued\n");
637 list_del(&tail_req->execlist_link);
638 i915_gem_request_unreference(tail_req);
642 i915_gem_request_reference(request);
643 list_add_tail(&request->execlist_link, &engine->execlist_queue);
644 request->ctx_hw_id = request->ctx->hw_id;
645 if (num_elements == 0)
646 execlists_context_unqueue(engine);
648 spin_unlock_bh(&engine->execlist_lock);
651 static int logical_ring_invalidate_all_caches(struct drm_i915_gem_request *req)
653 struct intel_engine_cs *engine = req->engine;
654 uint32_t flush_domains;
658 if (engine->gpu_caches_dirty)
659 flush_domains = I915_GEM_GPU_DOMAINS;
661 ret = engine->emit_flush(req, I915_GEM_GPU_DOMAINS, flush_domains);
665 engine->gpu_caches_dirty = false;
669 static int execlists_move_to_gpu(struct drm_i915_gem_request *req,
670 struct list_head *vmas)
672 const unsigned other_rings = ~intel_engine_flag(req->engine);
673 struct i915_vma *vma;
674 uint32_t flush_domains = 0;
675 bool flush_chipset = false;
678 list_for_each_entry(vma, vmas, exec_list) {
679 struct drm_i915_gem_object *obj = vma->obj;
681 if (obj->active & other_rings) {
682 ret = i915_gem_object_sync(obj, req->engine, &req);
687 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
688 flush_chipset |= i915_gem_clflush_object(obj, false);
690 flush_domains |= obj->base.write_domain;
693 if (flush_domains & I915_GEM_DOMAIN_GTT)
696 /* Unconditionally invalidate gpu caches and ensure that we do flush
697 * any residual writes from the previous batch.
699 return logical_ring_invalidate_all_caches(req);
702 int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *request)
704 struct intel_engine_cs *engine = request->engine;
705 struct intel_context *ce = &request->ctx->engine[engine->id];
708 /* Flush enough space to reduce the likelihood of waiting after
709 * we start building the request - in which case we will just
710 * have to repeat work.
712 request->reserved_space += EXECLISTS_REQUEST_SIZE;
715 ret = execlists_context_deferred_alloc(request->ctx, engine);
720 request->ringbuf = ce->ringbuf;
722 if (i915.enable_guc_submission) {
724 * Check that the GuC has space for the request before
725 * going any further, as the i915_add_request() call
726 * later on mustn't fail ...
728 ret = i915_guc_wq_check_space(request);
733 ret = intel_lr_context_pin(request->ctx, engine);
737 ret = intel_ring_begin(request, 0);
741 if (!ce->initialised) {
742 ret = engine->init_context(request);
746 ce->initialised = true;
749 /* Note that after this point, we have committed to using
750 * this request as it is being used to both track the
751 * state of engine initialisation and liveness of the
752 * golden renderstate above. Think twice before you try
753 * to cancel/unwind this request now.
756 request->reserved_space -= EXECLISTS_REQUEST_SIZE;
760 intel_lr_context_unpin(request->ctx, engine);
765 * intel_logical_ring_advance_and_submit() - advance the tail and submit the workload
766 * @request: Request to advance the logical ringbuffer of.
768 * The tail is updated in our logical ringbuffer struct, not in the actual context. What
769 * really happens during submission is that the context and current tail will be placed
770 * on a queue waiting for the ELSP to be ready to accept a new context submission. At that
771 * point, the tail *inside* the context is updated and the ELSP written to.
774 intel_logical_ring_advance_and_submit(struct drm_i915_gem_request *request)
776 struct intel_ringbuffer *ringbuf = request->ringbuf;
777 struct intel_engine_cs *engine = request->engine;
779 intel_logical_ring_advance(ringbuf);
780 request->tail = ringbuf->tail;
783 * Here we add two extra NOOPs as padding to avoid
784 * lite restore of a context with HEAD==TAIL.
786 * Caller must reserve WA_TAIL_DWORDS for us!
788 intel_logical_ring_emit(ringbuf, MI_NOOP);
789 intel_logical_ring_emit(ringbuf, MI_NOOP);
790 intel_logical_ring_advance(ringbuf);
792 if (intel_engine_stopped(engine))
795 /* We keep the previous context alive until we retire the following
796 * request. This ensures that any the context object is still pinned
797 * for any residual writes the HW makes into it on the context switch
798 * into the next object following the breadcrumb. Otherwise, we may
799 * retire the context too early.
801 request->previous_context = engine->last_context;
802 engine->last_context = request->ctx;
804 if (i915.enable_guc_submission)
805 i915_guc_submit(request);
807 execlists_context_queue(request);
813 * execlists_submission() - submit a batchbuffer for execution, Execlists style
814 * @params: execbuffer call parameters.
815 * @args: execbuffer call arguments.
816 * @vmas: list of vmas.
818 * This is the evil twin version of i915_gem_ringbuffer_submission. It abstracts
819 * away the submission details of the execbuffer ioctl call.
821 * Return: non-zero if the submission fails.
823 int intel_execlists_submission(struct i915_execbuffer_params *params,
824 struct drm_i915_gem_execbuffer2 *args,
825 struct list_head *vmas)
827 struct drm_device *dev = params->dev;
828 struct intel_engine_cs *engine = params->engine;
829 struct drm_i915_private *dev_priv = dev->dev_private;
830 struct intel_ringbuffer *ringbuf = params->ctx->engine[engine->id].ringbuf;
836 instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
837 instp_mask = I915_EXEC_CONSTANTS_MASK;
838 switch (instp_mode) {
839 case I915_EXEC_CONSTANTS_REL_GENERAL:
840 case I915_EXEC_CONSTANTS_ABSOLUTE:
841 case I915_EXEC_CONSTANTS_REL_SURFACE:
842 if (instp_mode != 0 && engine != &dev_priv->engine[RCS]) {
843 DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
847 if (instp_mode != dev_priv->relative_constants_mode) {
848 if (instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
849 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
853 /* The HW changed the meaning on this bit on gen6 */
854 instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
858 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode);
862 if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
863 DRM_DEBUG("sol reset is gen7 only\n");
867 ret = execlists_move_to_gpu(params->request, vmas);
871 if (engine == &dev_priv->engine[RCS] &&
872 instp_mode != dev_priv->relative_constants_mode) {
873 ret = intel_ring_begin(params->request, 4);
877 intel_logical_ring_emit(ringbuf, MI_NOOP);
878 intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(1));
879 intel_logical_ring_emit_reg(ringbuf, INSTPM);
880 intel_logical_ring_emit(ringbuf, instp_mask << 16 | instp_mode);
881 intel_logical_ring_advance(ringbuf);
883 dev_priv->relative_constants_mode = instp_mode;
886 exec_start = params->batch_obj_vm_offset +
887 args->batch_start_offset;
889 ret = engine->emit_bb_start(params->request, exec_start, params->dispatch_flags);
893 trace_i915_gem_ring_dispatch(params->request, params->dispatch_flags);
895 i915_gem_execbuffer_move_to_active(vmas, params->request);
900 void intel_execlists_cancel_requests(struct intel_engine_cs *engine)
902 struct drm_i915_gem_request *req, *tmp;
903 LIST_HEAD(cancel_list);
905 WARN_ON(!mutex_is_locked(&engine->i915->dev->struct_mutex));
907 spin_lock_bh(&engine->execlist_lock);
908 list_replace_init(&engine->execlist_queue, &cancel_list);
909 spin_unlock_bh(&engine->execlist_lock);
911 list_for_each_entry_safe(req, tmp, &cancel_list, execlist_link) {
912 list_del(&req->execlist_link);
913 i915_gem_request_unreference(req);
917 void intel_logical_ring_stop(struct intel_engine_cs *engine)
919 struct drm_i915_private *dev_priv = engine->i915;
922 if (!intel_engine_initialized(engine))
925 ret = intel_engine_idle(engine);
927 DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n",
930 /* TODO: Is this correct with Execlists enabled? */
931 I915_WRITE_MODE(engine, _MASKED_BIT_ENABLE(STOP_RING));
932 if (intel_wait_for_register(dev_priv,
933 RING_MI_MODE(engine->mmio_base),
934 MODE_IDLE, MODE_IDLE,
936 DRM_ERROR("%s :timed out trying to stop ring\n", engine->name);
939 I915_WRITE_MODE(engine, _MASKED_BIT_DISABLE(STOP_RING));
942 int logical_ring_flush_all_caches(struct drm_i915_gem_request *req)
944 struct intel_engine_cs *engine = req->engine;
947 if (!engine->gpu_caches_dirty)
950 ret = engine->emit_flush(req, 0, I915_GEM_GPU_DOMAINS);
954 engine->gpu_caches_dirty = false;
958 static int intel_lr_context_pin(struct i915_gem_context *ctx,
959 struct intel_engine_cs *engine)
961 struct drm_i915_private *dev_priv = ctx->i915;
962 struct intel_context *ce = &ctx->engine[engine->id];
967 lockdep_assert_held(&ctx->i915->dev->struct_mutex);
972 ret = i915_gem_obj_ggtt_pin(ce->state, GEN8_LR_CONTEXT_ALIGN,
973 PIN_OFFSET_BIAS | GUC_WOPCM_TOP);
977 vaddr = i915_gem_object_pin_map(ce->state);
979 ret = PTR_ERR(vaddr);
983 lrc_reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
985 ret = intel_pin_and_map_ringbuffer_obj(dev_priv, ce->ringbuf);
989 i915_gem_context_reference(ctx);
990 ce->lrc_vma = i915_gem_obj_to_ggtt(ce->state);
991 intel_lr_context_descriptor_update(ctx, engine);
993 lrc_reg_state[CTX_RING_BUFFER_START+1] = ce->ringbuf->vma->node.start;
994 ce->lrc_reg_state = lrc_reg_state;
995 ce->state->dirty = true;
997 /* Invalidate GuC TLB. */
998 if (i915.enable_guc_submission)
999 I915_WRITE(GEN8_GTCR, GEN8_GTCR_INVALIDATE);
1004 i915_gem_object_unpin_map(ce->state);
1006 i915_gem_object_ggtt_unpin(ce->state);
1012 void intel_lr_context_unpin(struct i915_gem_context *ctx,
1013 struct intel_engine_cs *engine)
1015 struct intel_context *ce = &ctx->engine[engine->id];
1017 lockdep_assert_held(&ctx->i915->dev->struct_mutex);
1018 GEM_BUG_ON(ce->pin_count == 0);
1020 if (--ce->pin_count)
1023 intel_unpin_ringbuffer_obj(ce->ringbuf);
1025 i915_gem_object_unpin_map(ce->state);
1026 i915_gem_object_ggtt_unpin(ce->state);
1030 ce->lrc_reg_state = NULL;
1032 i915_gem_context_unreference(ctx);
1035 static int intel_logical_ring_workarounds_emit(struct drm_i915_gem_request *req)
1038 struct intel_engine_cs *engine = req->engine;
1039 struct intel_ringbuffer *ringbuf = req->ringbuf;
1040 struct i915_workarounds *w = &req->i915->workarounds;
1045 engine->gpu_caches_dirty = true;
1046 ret = logical_ring_flush_all_caches(req);
1050 ret = intel_ring_begin(req, w->count * 2 + 2);
1054 intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(w->count));
1055 for (i = 0; i < w->count; i++) {
1056 intel_logical_ring_emit_reg(ringbuf, w->reg[i].addr);
1057 intel_logical_ring_emit(ringbuf, w->reg[i].value);
1059 intel_logical_ring_emit(ringbuf, MI_NOOP);
1061 intel_logical_ring_advance(ringbuf);
1063 engine->gpu_caches_dirty = true;
1064 ret = logical_ring_flush_all_caches(req);
1071 #define wa_ctx_emit(batch, index, cmd) \
1073 int __index = (index)++; \
1074 if (WARN_ON(__index >= (PAGE_SIZE / sizeof(uint32_t)))) { \
1077 batch[__index] = (cmd); \
1080 #define wa_ctx_emit_reg(batch, index, reg) \
1081 wa_ctx_emit((batch), (index), i915_mmio_reg_offset(reg))
1084 * In this WA we need to set GEN8_L3SQCREG4[21:21] and reset it after
1085 * PIPE_CONTROL instruction. This is required for the flush to happen correctly
1086 * but there is a slight complication as this is applied in WA batch where the
1087 * values are only initialized once so we cannot take register value at the
1088 * beginning and reuse it further; hence we save its value to memory, upload a
1089 * constant value with bit21 set and then we restore it back with the saved value.
1090 * To simplify the WA, a constant value is formed by using the default value
1091 * of this register. This shouldn't be a problem because we are only modifying
1092 * it for a short period and this batch in non-premptible. We can ofcourse
1093 * use additional instructions that read the actual value of the register
1094 * at that time and set our bit of interest but it makes the WA complicated.
1096 * This WA is also required for Gen9 so extracting as a function avoids
1099 static inline int gen8_emit_flush_coherentl3_wa(struct intel_engine_cs *engine,
1100 uint32_t *const batch,
1103 uint32_t l3sqc4_flush = (0x40400000 | GEN8_LQSC_FLUSH_COHERENT_LINES);
1106 * WaDisableLSQCROPERFforOCL:skl,kbl
1107 * This WA is implemented in skl_init_clock_gating() but since
1108 * this batch updates GEN8_L3SQCREG4 with default value we need to
1109 * set this bit here to retain the WA during flush.
1111 if (IS_SKL_REVID(engine->i915, 0, SKL_REVID_E0) ||
1112 IS_KBL_REVID(engine->i915, 0, KBL_REVID_E0))
1113 l3sqc4_flush |= GEN8_LQSC_RO_PERF_DIS;
1115 wa_ctx_emit(batch, index, (MI_STORE_REGISTER_MEM_GEN8 |
1116 MI_SRM_LRM_GLOBAL_GTT));
1117 wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
1118 wa_ctx_emit(batch, index, engine->scratch.gtt_offset + 256);
1119 wa_ctx_emit(batch, index, 0);
1121 wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1));
1122 wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
1123 wa_ctx_emit(batch, index, l3sqc4_flush);
1125 wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
1126 wa_ctx_emit(batch, index, (PIPE_CONTROL_CS_STALL |
1127 PIPE_CONTROL_DC_FLUSH_ENABLE));
1128 wa_ctx_emit(batch, index, 0);
1129 wa_ctx_emit(batch, index, 0);
1130 wa_ctx_emit(batch, index, 0);
1131 wa_ctx_emit(batch, index, 0);
1133 wa_ctx_emit(batch, index, (MI_LOAD_REGISTER_MEM_GEN8 |
1134 MI_SRM_LRM_GLOBAL_GTT));
1135 wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
1136 wa_ctx_emit(batch, index, engine->scratch.gtt_offset + 256);
1137 wa_ctx_emit(batch, index, 0);
1142 static inline uint32_t wa_ctx_start(struct i915_wa_ctx_bb *wa_ctx,
1144 uint32_t start_alignment)
1146 return wa_ctx->offset = ALIGN(offset, start_alignment);
1149 static inline int wa_ctx_end(struct i915_wa_ctx_bb *wa_ctx,
1151 uint32_t size_alignment)
1153 wa_ctx->size = offset - wa_ctx->offset;
1155 WARN(wa_ctx->size % size_alignment,
1156 "wa_ctx_bb failed sanity checks: size %d is not aligned to %d\n",
1157 wa_ctx->size, size_alignment);
1162 * gen8_init_indirectctx_bb() - initialize indirect ctx batch with WA
1164 * @engine: only applicable for RCS
1165 * @wa_ctx: structure representing wa_ctx
1166 * offset: specifies start of the batch, should be cache-aligned. This is updated
1167 * with the offset value received as input.
1168 * size: size of the batch in DWORDS but HW expects in terms of cachelines
1169 * @batch: page in which WA are loaded
1170 * @offset: This field specifies the start of the batch, it should be
1171 * cache-aligned otherwise it is adjusted accordingly.
1172 * Typically we only have one indirect_ctx and per_ctx batch buffer which are
1173 * initialized at the beginning and shared across all contexts but this field
1174 * helps us to have multiple batches at different offsets and select them based
1175 * on a criteria. At the moment this batch always start at the beginning of the page
1176 * and at this point we don't have multiple wa_ctx batch buffers.
1178 * The number of WA applied are not known at the beginning; we use this field
1179 * to return the no of DWORDS written.
1181 * It is to be noted that this batch does not contain MI_BATCH_BUFFER_END
1182 * so it adds NOOPs as padding to make it cacheline aligned.
1183 * MI_BATCH_BUFFER_END will be added to perctx batch and both of them together
1184 * makes a complete batch buffer.
1186 * Return: non-zero if we exceed the PAGE_SIZE limit.
1189 static int gen8_init_indirectctx_bb(struct intel_engine_cs *engine,
1190 struct i915_wa_ctx_bb *wa_ctx,
1191 uint32_t *const batch,
1194 uint32_t scratch_addr;
1195 uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1197 /* WaDisableCtxRestoreArbitration:bdw,chv */
1198 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_DISABLE);
1200 /* WaFlushCoherentL3CacheLinesAtContextSwitch:bdw */
1201 if (IS_BROADWELL(engine->i915)) {
1202 int rc = gen8_emit_flush_coherentl3_wa(engine, batch, index);
1208 /* WaClearSlmSpaceAtContextSwitch:bdw,chv */
1209 /* Actual scratch location is at 128 bytes offset */
1210 scratch_addr = engine->scratch.gtt_offset + 2*CACHELINE_BYTES;
1212 wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
1213 wa_ctx_emit(batch, index, (PIPE_CONTROL_FLUSH_L3 |
1214 PIPE_CONTROL_GLOBAL_GTT_IVB |
1215 PIPE_CONTROL_CS_STALL |
1216 PIPE_CONTROL_QW_WRITE));
1217 wa_ctx_emit(batch, index, scratch_addr);
1218 wa_ctx_emit(batch, index, 0);
1219 wa_ctx_emit(batch, index, 0);
1220 wa_ctx_emit(batch, index, 0);
1222 /* Pad to end of cacheline */
1223 while (index % CACHELINE_DWORDS)
1224 wa_ctx_emit(batch, index, MI_NOOP);
1227 * MI_BATCH_BUFFER_END is not required in Indirect ctx BB because
1228 * execution depends on the length specified in terms of cache lines
1229 * in the register CTX_RCS_INDIRECT_CTX
1232 return wa_ctx_end(wa_ctx, *offset = index, CACHELINE_DWORDS);
1236 * gen8_init_perctx_bb() - initialize per ctx batch with WA
1238 * @engine: only applicable for RCS
1239 * @wa_ctx: structure representing wa_ctx
1240 * offset: specifies start of the batch, should be cache-aligned.
1241 * size: size of the batch in DWORDS but HW expects in terms of cachelines
1242 * @batch: page in which WA are loaded
1243 * @offset: This field specifies the start of this batch.
1244 * This batch is started immediately after indirect_ctx batch. Since we ensure
1245 * that indirect_ctx ends on a cacheline this batch is aligned automatically.
1247 * The number of DWORDS written are returned using this field.
1249 * This batch is terminated with MI_BATCH_BUFFER_END and so we need not add padding
1250 * to align it with cacheline as padding after MI_BATCH_BUFFER_END is redundant.
1252 static int gen8_init_perctx_bb(struct intel_engine_cs *engine,
1253 struct i915_wa_ctx_bb *wa_ctx,
1254 uint32_t *const batch,
1257 uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1259 /* WaDisableCtxRestoreArbitration:bdw,chv */
1260 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_ENABLE);
1262 wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END);
1264 return wa_ctx_end(wa_ctx, *offset = index, 1);
1267 static int gen9_init_indirectctx_bb(struct intel_engine_cs *engine,
1268 struct i915_wa_ctx_bb *wa_ctx,
1269 uint32_t *const batch,
1273 uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1275 /* WaDisableCtxRestoreArbitration:skl,bxt */
1276 if (IS_SKL_REVID(engine->i915, 0, SKL_REVID_D0) ||
1277 IS_BXT_REVID(engine->i915, 0, BXT_REVID_A1))
1278 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_DISABLE);
1280 /* WaFlushCoherentL3CacheLinesAtContextSwitch:skl,bxt */
1281 ret = gen8_emit_flush_coherentl3_wa(engine, batch, index);
1286 /* WaClearSlmSpaceAtContextSwitch:kbl */
1287 /* Actual scratch location is at 128 bytes offset */
1288 if (IS_KBL_REVID(engine->i915, 0, KBL_REVID_A0)) {
1289 uint32_t scratch_addr
1290 = engine->scratch.gtt_offset + 2*CACHELINE_BYTES;
1292 wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
1293 wa_ctx_emit(batch, index, (PIPE_CONTROL_FLUSH_L3 |
1294 PIPE_CONTROL_GLOBAL_GTT_IVB |
1295 PIPE_CONTROL_CS_STALL |
1296 PIPE_CONTROL_QW_WRITE));
1297 wa_ctx_emit(batch, index, scratch_addr);
1298 wa_ctx_emit(batch, index, 0);
1299 wa_ctx_emit(batch, index, 0);
1300 wa_ctx_emit(batch, index, 0);
1302 /* Pad to end of cacheline */
1303 while (index % CACHELINE_DWORDS)
1304 wa_ctx_emit(batch, index, MI_NOOP);
1306 return wa_ctx_end(wa_ctx, *offset = index, CACHELINE_DWORDS);
1309 static int gen9_init_perctx_bb(struct intel_engine_cs *engine,
1310 struct i915_wa_ctx_bb *wa_ctx,
1311 uint32_t *const batch,
1314 uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1316 /* WaSetDisablePixMaskCammingAndRhwoInCommonSliceChicken:skl,bxt */
1317 if (IS_SKL_REVID(engine->i915, 0, SKL_REVID_B0) ||
1318 IS_BXT_REVID(engine->i915, 0, BXT_REVID_A1)) {
1319 wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1));
1320 wa_ctx_emit_reg(batch, index, GEN9_SLICE_COMMON_ECO_CHICKEN0);
1321 wa_ctx_emit(batch, index,
1322 _MASKED_BIT_ENABLE(DISABLE_PIXEL_MASK_CAMMING));
1323 wa_ctx_emit(batch, index, MI_NOOP);
1326 /* WaClearTdlStateAckDirtyBits:bxt */
1327 if (IS_BXT_REVID(engine->i915, 0, BXT_REVID_B0)) {
1328 wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(4));
1330 wa_ctx_emit_reg(batch, index, GEN8_STATE_ACK);
1331 wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS));
1333 wa_ctx_emit_reg(batch, index, GEN9_STATE_ACK_SLICE1);
1334 wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS));
1336 wa_ctx_emit_reg(batch, index, GEN9_STATE_ACK_SLICE2);
1337 wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS));
1339 wa_ctx_emit_reg(batch, index, GEN7_ROW_CHICKEN2);
1340 /* dummy write to CS, mask bits are 0 to ensure the register is not modified */
1341 wa_ctx_emit(batch, index, 0x0);
1342 wa_ctx_emit(batch, index, MI_NOOP);
1345 /* WaDisableCtxRestoreArbitration:skl,bxt */
1346 if (IS_SKL_REVID(engine->i915, 0, SKL_REVID_D0) ||
1347 IS_BXT_REVID(engine->i915, 0, BXT_REVID_A1))
1348 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_ENABLE);
1350 wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END);
1352 return wa_ctx_end(wa_ctx, *offset = index, 1);
1355 static int lrc_setup_wa_ctx_obj(struct intel_engine_cs *engine, u32 size)
1359 engine->wa_ctx.obj = i915_gem_object_create(engine->i915->dev,
1361 if (IS_ERR(engine->wa_ctx.obj)) {
1362 DRM_DEBUG_DRIVER("alloc LRC WA ctx backing obj failed.\n");
1363 ret = PTR_ERR(engine->wa_ctx.obj);
1364 engine->wa_ctx.obj = NULL;
1368 ret = i915_gem_obj_ggtt_pin(engine->wa_ctx.obj, PAGE_SIZE, 0);
1370 DRM_DEBUG_DRIVER("pin LRC WA ctx backing obj failed: %d\n",
1372 drm_gem_object_unreference(&engine->wa_ctx.obj->base);
1379 static void lrc_destroy_wa_ctx_obj(struct intel_engine_cs *engine)
1381 if (engine->wa_ctx.obj) {
1382 i915_gem_object_ggtt_unpin(engine->wa_ctx.obj);
1383 drm_gem_object_unreference(&engine->wa_ctx.obj->base);
1384 engine->wa_ctx.obj = NULL;
1388 static int intel_init_workaround_bb(struct intel_engine_cs *engine)
1394 struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx;
1396 WARN_ON(engine->id != RCS);
1398 /* update this when WA for higher Gen are added */
1399 if (INTEL_GEN(engine->i915) > 9) {
1400 DRM_ERROR("WA batch buffer is not initialized for Gen%d\n",
1401 INTEL_GEN(engine->i915));
1405 /* some WA perform writes to scratch page, ensure it is valid */
1406 if (engine->scratch.obj == NULL) {
1407 DRM_ERROR("scratch page not allocated for %s\n", engine->name);
1411 ret = lrc_setup_wa_ctx_obj(engine, PAGE_SIZE);
1413 DRM_DEBUG_DRIVER("Failed to setup context WA page: %d\n", ret);
1417 page = i915_gem_object_get_dirty_page(wa_ctx->obj, 0);
1418 batch = kmap_atomic(page);
1421 if (IS_GEN8(engine->i915)) {
1422 ret = gen8_init_indirectctx_bb(engine,
1423 &wa_ctx->indirect_ctx,
1429 ret = gen8_init_perctx_bb(engine,
1435 } else if (IS_GEN9(engine->i915)) {
1436 ret = gen9_init_indirectctx_bb(engine,
1437 &wa_ctx->indirect_ctx,
1443 ret = gen9_init_perctx_bb(engine,
1452 kunmap_atomic(batch);
1454 lrc_destroy_wa_ctx_obj(engine);
1459 static void lrc_init_hws(struct intel_engine_cs *engine)
1461 struct drm_i915_private *dev_priv = engine->i915;
1463 I915_WRITE(RING_HWS_PGA(engine->mmio_base),
1464 (u32)engine->status_page.gfx_addr);
1465 POSTING_READ(RING_HWS_PGA(engine->mmio_base));
1468 static int gen8_init_common_ring(struct intel_engine_cs *engine)
1470 struct drm_i915_private *dev_priv = engine->i915;
1471 unsigned int next_context_status_buffer_hw;
1473 lrc_init_hws(engine);
1475 I915_WRITE_IMR(engine,
1476 ~(engine->irq_enable_mask | engine->irq_keep_mask));
1477 I915_WRITE(RING_HWSTAM(engine->mmio_base), 0xffffffff);
1479 I915_WRITE(RING_MODE_GEN7(engine),
1480 _MASKED_BIT_DISABLE(GFX_REPLAY_MODE) |
1481 _MASKED_BIT_ENABLE(GFX_RUN_LIST_ENABLE));
1482 POSTING_READ(RING_MODE_GEN7(engine));
1485 * Instead of resetting the Context Status Buffer (CSB) read pointer to
1486 * zero, we need to read the write pointer from hardware and use its
1487 * value because "this register is power context save restored".
1488 * Effectively, these states have been observed:
1490 * | Suspend-to-idle (freeze) | Suspend-to-RAM (mem) |
1491 * BDW | CSB regs not reset | CSB regs reset |
1492 * CHT | CSB regs not reset | CSB regs not reset |
1496 next_context_status_buffer_hw =
1497 GEN8_CSB_WRITE_PTR(I915_READ(RING_CONTEXT_STATUS_PTR(engine)));
1500 * When the CSB registers are reset (also after power-up / gpu reset),
1501 * CSB write pointer is set to all 1's, which is not valid, use '5' in
1502 * this special case, so the first element read is CSB[0].
1504 if (next_context_status_buffer_hw == GEN8_CSB_PTR_MASK)
1505 next_context_status_buffer_hw = (GEN8_CSB_ENTRIES - 1);
1507 engine->next_context_status_buffer = next_context_status_buffer_hw;
1508 DRM_DEBUG_DRIVER("Execlists enabled for %s\n", engine->name);
1510 intel_engine_init_hangcheck(engine);
1512 return intel_mocs_init_engine(engine);
1515 static int gen8_init_render_ring(struct intel_engine_cs *engine)
1517 struct drm_i915_private *dev_priv = engine->i915;
1520 ret = gen8_init_common_ring(engine);
1524 /* We need to disable the AsyncFlip performance optimisations in order
1525 * to use MI_WAIT_FOR_EVENT within the CS. It should already be
1526 * programmed to '1' on all products.
1528 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw,chv
1530 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
1532 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
1534 return init_workarounds_ring(engine);
1537 static int gen9_init_render_ring(struct intel_engine_cs *engine)
1541 ret = gen8_init_common_ring(engine);
1545 return init_workarounds_ring(engine);
1548 static int intel_logical_ring_emit_pdps(struct drm_i915_gem_request *req)
1550 struct i915_hw_ppgtt *ppgtt = req->ctx->ppgtt;
1551 struct intel_engine_cs *engine = req->engine;
1552 struct intel_ringbuffer *ringbuf = req->ringbuf;
1553 const int num_lri_cmds = GEN8_LEGACY_PDPES * 2;
1556 ret = intel_ring_begin(req, num_lri_cmds * 2 + 2);
1560 intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(num_lri_cmds));
1561 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
1562 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
1564 intel_logical_ring_emit_reg(ringbuf,
1565 GEN8_RING_PDP_UDW(engine, i));
1566 intel_logical_ring_emit(ringbuf, upper_32_bits(pd_daddr));
1567 intel_logical_ring_emit_reg(ringbuf,
1568 GEN8_RING_PDP_LDW(engine, i));
1569 intel_logical_ring_emit(ringbuf, lower_32_bits(pd_daddr));
1572 intel_logical_ring_emit(ringbuf, MI_NOOP);
1573 intel_logical_ring_advance(ringbuf);
1578 static int gen8_emit_bb_start(struct drm_i915_gem_request *req,
1579 u64 offset, unsigned dispatch_flags)
1581 struct intel_ringbuffer *ringbuf = req->ringbuf;
1582 bool ppgtt = !(dispatch_flags & I915_DISPATCH_SECURE);
1585 /* Don't rely in hw updating PDPs, specially in lite-restore.
1586 * Ideally, we should set Force PD Restore in ctx descriptor,
1587 * but we can't. Force Restore would be a second option, but
1588 * it is unsafe in case of lite-restore (because the ctx is
1589 * not idle). PML4 is allocated during ppgtt init so this is
1590 * not needed in 48-bit.*/
1591 if (req->ctx->ppgtt &&
1592 (intel_engine_flag(req->engine) & req->ctx->ppgtt->pd_dirty_rings)) {
1593 if (!USES_FULL_48BIT_PPGTT(req->i915) &&
1594 !intel_vgpu_active(req->i915)) {
1595 ret = intel_logical_ring_emit_pdps(req);
1600 req->ctx->ppgtt->pd_dirty_rings &= ~intel_engine_flag(req->engine);
1603 ret = intel_ring_begin(req, 4);
1607 /* FIXME(BDW): Address space and security selectors. */
1608 intel_logical_ring_emit(ringbuf, MI_BATCH_BUFFER_START_GEN8 |
1610 (dispatch_flags & I915_DISPATCH_RS ?
1611 MI_BATCH_RESOURCE_STREAMER : 0));
1612 intel_logical_ring_emit(ringbuf, lower_32_bits(offset));
1613 intel_logical_ring_emit(ringbuf, upper_32_bits(offset));
1614 intel_logical_ring_emit(ringbuf, MI_NOOP);
1615 intel_logical_ring_advance(ringbuf);
1620 static bool gen8_logical_ring_get_irq(struct intel_engine_cs *engine)
1622 struct drm_i915_private *dev_priv = engine->i915;
1623 unsigned long flags;
1625 if (WARN_ON(!intel_irqs_enabled(dev_priv)))
1628 spin_lock_irqsave(&dev_priv->irq_lock, flags);
1629 if (engine->irq_refcount++ == 0) {
1630 I915_WRITE_IMR(engine,
1631 ~(engine->irq_enable_mask | engine->irq_keep_mask));
1632 POSTING_READ(RING_IMR(engine->mmio_base));
1634 spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1639 static void gen8_logical_ring_put_irq(struct intel_engine_cs *engine)
1641 struct drm_i915_private *dev_priv = engine->i915;
1642 unsigned long flags;
1644 spin_lock_irqsave(&dev_priv->irq_lock, flags);
1645 if (--engine->irq_refcount == 0) {
1646 I915_WRITE_IMR(engine, ~engine->irq_keep_mask);
1647 POSTING_READ(RING_IMR(engine->mmio_base));
1649 spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1652 static int gen8_emit_flush(struct drm_i915_gem_request *request,
1653 u32 invalidate_domains,
1656 struct intel_ringbuffer *ringbuf = request->ringbuf;
1657 struct intel_engine_cs *engine = ringbuf->engine;
1658 struct drm_i915_private *dev_priv = request->i915;
1662 ret = intel_ring_begin(request, 4);
1666 cmd = MI_FLUSH_DW + 1;
1668 /* We always require a command barrier so that subsequent
1669 * commands, such as breadcrumb interrupts, are strictly ordered
1670 * wrt the contents of the write cache being flushed to memory
1671 * (and thus being coherent from the CPU).
1673 cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1675 if (invalidate_domains & I915_GEM_GPU_DOMAINS) {
1676 cmd |= MI_INVALIDATE_TLB;
1677 if (engine == &dev_priv->engine[VCS])
1678 cmd |= MI_INVALIDATE_BSD;
1681 intel_logical_ring_emit(ringbuf, cmd);
1682 intel_logical_ring_emit(ringbuf,
1683 I915_GEM_HWS_SCRATCH_ADDR |
1684 MI_FLUSH_DW_USE_GTT);
1685 intel_logical_ring_emit(ringbuf, 0); /* upper addr */
1686 intel_logical_ring_emit(ringbuf, 0); /* value */
1687 intel_logical_ring_advance(ringbuf);
1692 static int gen8_emit_flush_render(struct drm_i915_gem_request *request,
1693 u32 invalidate_domains,
1696 struct intel_ringbuffer *ringbuf = request->ringbuf;
1697 struct intel_engine_cs *engine = ringbuf->engine;
1698 u32 scratch_addr = engine->scratch.gtt_offset + 2 * CACHELINE_BYTES;
1699 bool vf_flush_wa = false, dc_flush_wa = false;
1704 flags |= PIPE_CONTROL_CS_STALL;
1706 if (flush_domains) {
1707 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
1708 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
1709 flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
1710 flags |= PIPE_CONTROL_FLUSH_ENABLE;
1713 if (invalidate_domains) {
1714 flags |= PIPE_CONTROL_TLB_INVALIDATE;
1715 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
1716 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1717 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
1718 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
1719 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
1720 flags |= PIPE_CONTROL_QW_WRITE;
1721 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
1724 * On GEN9: before VF_CACHE_INVALIDATE we need to emit a NULL
1727 if (IS_GEN9(request->i915))
1730 /* WaForGAMHang:kbl */
1731 if (IS_KBL_REVID(request->i915, 0, KBL_REVID_B0))
1743 ret = intel_ring_begin(request, len);
1748 intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1749 intel_logical_ring_emit(ringbuf, 0);
1750 intel_logical_ring_emit(ringbuf, 0);
1751 intel_logical_ring_emit(ringbuf, 0);
1752 intel_logical_ring_emit(ringbuf, 0);
1753 intel_logical_ring_emit(ringbuf, 0);
1757 intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1758 intel_logical_ring_emit(ringbuf, PIPE_CONTROL_DC_FLUSH_ENABLE);
1759 intel_logical_ring_emit(ringbuf, 0);
1760 intel_logical_ring_emit(ringbuf, 0);
1761 intel_logical_ring_emit(ringbuf, 0);
1762 intel_logical_ring_emit(ringbuf, 0);
1765 intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1766 intel_logical_ring_emit(ringbuf, flags);
1767 intel_logical_ring_emit(ringbuf, scratch_addr);
1768 intel_logical_ring_emit(ringbuf, 0);
1769 intel_logical_ring_emit(ringbuf, 0);
1770 intel_logical_ring_emit(ringbuf, 0);
1773 intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1774 intel_logical_ring_emit(ringbuf, PIPE_CONTROL_CS_STALL);
1775 intel_logical_ring_emit(ringbuf, 0);
1776 intel_logical_ring_emit(ringbuf, 0);
1777 intel_logical_ring_emit(ringbuf, 0);
1778 intel_logical_ring_emit(ringbuf, 0);
1781 intel_logical_ring_advance(ringbuf);
1786 static void bxt_a_seqno_barrier(struct intel_engine_cs *engine)
1789 * On BXT A steppings there is a HW coherency issue whereby the
1790 * MI_STORE_DATA_IMM storing the completed request's seqno
1791 * occasionally doesn't invalidate the CPU cache. Work around this by
1792 * clflushing the corresponding cacheline whenever the caller wants
1793 * the coherency to be guaranteed. Note that this cacheline is known
1794 * to be clean at this point, since we only write it in
1795 * bxt_a_set_seqno(), where we also do a clflush after the write. So
1796 * this clflush in practice becomes an invalidate operation.
1798 intel_flush_status_page(engine, I915_GEM_HWS_INDEX);
1802 * Reserve space for 2 NOOPs at the end of each request to be
1803 * used as a workaround for not being allowed to do lite
1804 * restore with HEAD==TAIL (WaIdleLiteRestore).
1806 #define WA_TAIL_DWORDS 2
1808 static int gen8_emit_request(struct drm_i915_gem_request *request)
1810 struct intel_ringbuffer *ringbuf = request->ringbuf;
1813 ret = intel_ring_begin(request, 6 + WA_TAIL_DWORDS);
1817 /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */
1818 BUILD_BUG_ON(I915_GEM_HWS_INDEX_ADDR & (1 << 5));
1820 intel_logical_ring_emit(ringbuf,
1821 (MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW);
1822 intel_logical_ring_emit(ringbuf,
1823 intel_hws_seqno_address(request->engine) |
1824 MI_FLUSH_DW_USE_GTT);
1825 intel_logical_ring_emit(ringbuf, 0);
1826 intel_logical_ring_emit(ringbuf, request->seqno);
1827 intel_logical_ring_emit(ringbuf, MI_USER_INTERRUPT);
1828 intel_logical_ring_emit(ringbuf, MI_NOOP);
1829 return intel_logical_ring_advance_and_submit(request);
1832 static int gen8_emit_request_render(struct drm_i915_gem_request *request)
1834 struct intel_ringbuffer *ringbuf = request->ringbuf;
1837 ret = intel_ring_begin(request, 8 + WA_TAIL_DWORDS);
1841 /* We're using qword write, seqno should be aligned to 8 bytes. */
1842 BUILD_BUG_ON(I915_GEM_HWS_INDEX & 1);
1844 /* w/a for post sync ops following a GPGPU operation we
1845 * need a prior CS_STALL, which is emitted by the flush
1846 * following the batch.
1848 intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1849 intel_logical_ring_emit(ringbuf,
1850 (PIPE_CONTROL_GLOBAL_GTT_IVB |
1851 PIPE_CONTROL_CS_STALL |
1852 PIPE_CONTROL_QW_WRITE));
1853 intel_logical_ring_emit(ringbuf,
1854 intel_hws_seqno_address(request->engine));
1855 intel_logical_ring_emit(ringbuf, 0);
1856 intel_logical_ring_emit(ringbuf, i915_gem_request_get_seqno(request));
1857 /* We're thrashing one dword of HWS. */
1858 intel_logical_ring_emit(ringbuf, 0);
1859 intel_logical_ring_emit(ringbuf, MI_USER_INTERRUPT);
1860 intel_logical_ring_emit(ringbuf, MI_NOOP);
1861 return intel_logical_ring_advance_and_submit(request);
1864 static int intel_lr_context_render_state_init(struct drm_i915_gem_request *req)
1866 struct render_state so;
1869 ret = i915_gem_render_state_prepare(req->engine, &so);
1873 if (so.rodata == NULL)
1876 ret = req->engine->emit_bb_start(req, so.ggtt_offset,
1877 I915_DISPATCH_SECURE);
1881 ret = req->engine->emit_bb_start(req,
1882 (so.ggtt_offset + so.aux_batch_offset),
1883 I915_DISPATCH_SECURE);
1887 i915_vma_move_to_active(i915_gem_obj_to_ggtt(so.obj), req);
1890 i915_gem_render_state_fini(&so);
1894 static int gen8_init_rcs_context(struct drm_i915_gem_request *req)
1898 ret = intel_logical_ring_workarounds_emit(req);
1902 ret = intel_rcs_context_init_mocs(req);
1904 * Failing to program the MOCS is non-fatal.The system will not
1905 * run at peak performance. So generate an error and carry on.
1908 DRM_ERROR("MOCS failed to program: expect performance issues.\n");
1910 return intel_lr_context_render_state_init(req);
1914 * intel_logical_ring_cleanup() - deallocate the Engine Command Streamer
1916 * @engine: Engine Command Streamer.
1919 void intel_logical_ring_cleanup(struct intel_engine_cs *engine)
1921 struct drm_i915_private *dev_priv;
1923 if (!intel_engine_initialized(engine))
1927 * Tasklet cannot be active at this point due intel_mark_active/idle
1928 * so this is just for documentation.
1930 if (WARN_ON(test_bit(TASKLET_STATE_SCHED, &engine->irq_tasklet.state)))
1931 tasklet_kill(&engine->irq_tasklet);
1933 dev_priv = engine->i915;
1935 if (engine->buffer) {
1936 intel_logical_ring_stop(engine);
1937 WARN_ON((I915_READ_MODE(engine) & MODE_IDLE) == 0);
1940 if (engine->cleanup)
1941 engine->cleanup(engine);
1943 i915_cmd_parser_fini_ring(engine);
1944 i915_gem_batch_pool_fini(&engine->batch_pool);
1946 intel_engine_fini_breadcrumbs(engine);
1948 if (engine->status_page.obj) {
1949 i915_gem_object_unpin_map(engine->status_page.obj);
1950 engine->status_page.obj = NULL;
1952 intel_lr_context_unpin(dev_priv->kernel_context, engine);
1954 engine->idle_lite_restore_wa = 0;
1955 engine->disable_lite_restore_wa = false;
1956 engine->ctx_desc_template = 0;
1958 lrc_destroy_wa_ctx_obj(engine);
1959 engine->i915 = NULL;
1963 logical_ring_default_vfuncs(struct intel_engine_cs *engine)
1965 /* Default vfuncs which can be overriden by each engine. */
1966 engine->init_hw = gen8_init_common_ring;
1967 engine->emit_request = gen8_emit_request;
1968 engine->emit_flush = gen8_emit_flush;
1969 engine->irq_get = gen8_logical_ring_get_irq;
1970 engine->irq_put = gen8_logical_ring_put_irq;
1971 engine->emit_bb_start = gen8_emit_bb_start;
1972 if (IS_BXT_REVID(engine->i915, 0, BXT_REVID_A1))
1973 engine->irq_seqno_barrier = bxt_a_seqno_barrier;
1977 logical_ring_default_irqs(struct intel_engine_cs *engine, unsigned shift)
1979 engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT << shift;
1980 engine->irq_keep_mask = GT_CONTEXT_SWITCH_INTERRUPT << shift;
1984 lrc_setup_hws(struct intel_engine_cs *engine,
1985 struct drm_i915_gem_object *dctx_obj)
1989 /* The HWSP is part of the default context object in LRC mode. */
1990 engine->status_page.gfx_addr = i915_gem_obj_ggtt_offset(dctx_obj) +
1991 LRC_PPHWSP_PN * PAGE_SIZE;
1992 hws = i915_gem_object_pin_map(dctx_obj);
1994 return PTR_ERR(hws);
1995 engine->status_page.page_addr = hws + LRC_PPHWSP_PN * PAGE_SIZE;
1996 engine->status_page.obj = dctx_obj;
2002 logical_ring_init(struct intel_engine_cs *engine)
2004 struct i915_gem_context *dctx = engine->i915->kernel_context;
2007 ret = intel_engine_init_breadcrumbs(engine);
2011 ret = i915_cmd_parser_init_ring(engine);
2015 ret = execlists_context_deferred_alloc(dctx, engine);
2019 /* As this is the default context, always pin it */
2020 ret = intel_lr_context_pin(dctx, engine);
2022 DRM_ERROR("Failed to pin context for %s: %d\n",
2027 /* And setup the hardware status page. */
2028 ret = lrc_setup_hws(engine, dctx->engine[engine->id].state);
2030 DRM_ERROR("Failed to set up hws %s: %d\n", engine->name, ret);
2037 intel_logical_ring_cleanup(engine);
2041 static int logical_render_ring_init(struct intel_engine_cs *engine)
2043 struct drm_i915_private *dev_priv = engine->i915;
2046 if (HAS_L3_DPF(dev_priv))
2047 engine->irq_keep_mask |= GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
2049 /* Override some for render ring. */
2050 if (INTEL_GEN(dev_priv) >= 9)
2051 engine->init_hw = gen9_init_render_ring;
2053 engine->init_hw = gen8_init_render_ring;
2054 engine->init_context = gen8_init_rcs_context;
2055 engine->cleanup = intel_fini_pipe_control;
2056 engine->emit_flush = gen8_emit_flush_render;
2057 engine->emit_request = gen8_emit_request_render;
2059 ret = intel_init_pipe_control(engine);
2063 ret = intel_init_workaround_bb(engine);
2066 * We continue even if we fail to initialize WA batch
2067 * because we only expect rare glitches but nothing
2068 * critical to prevent us from using GPU
2070 DRM_ERROR("WA batch buffer initialization failed: %d\n",
2074 ret = logical_ring_init(engine);
2076 lrc_destroy_wa_ctx_obj(engine);
2082 static const struct logical_ring_info {
2088 int (*init)(struct intel_engine_cs *engine);
2089 } logical_rings[] = {
2091 .name = "render ring",
2092 .exec_id = I915_EXEC_RENDER,
2093 .guc_id = GUC_RENDER_ENGINE,
2094 .mmio_base = RENDER_RING_BASE,
2095 .irq_shift = GEN8_RCS_IRQ_SHIFT,
2096 .init = logical_render_ring_init,
2099 .name = "blitter ring",
2100 .exec_id = I915_EXEC_BLT,
2101 .guc_id = GUC_BLITTER_ENGINE,
2102 .mmio_base = BLT_RING_BASE,
2103 .irq_shift = GEN8_BCS_IRQ_SHIFT,
2104 .init = logical_ring_init,
2108 .exec_id = I915_EXEC_BSD,
2109 .guc_id = GUC_VIDEO_ENGINE,
2110 .mmio_base = GEN6_BSD_RING_BASE,
2111 .irq_shift = GEN8_VCS1_IRQ_SHIFT,
2112 .init = logical_ring_init,
2115 .name = "bsd2 ring",
2116 .exec_id = I915_EXEC_BSD,
2117 .guc_id = GUC_VIDEO_ENGINE2,
2118 .mmio_base = GEN8_BSD2_RING_BASE,
2119 .irq_shift = GEN8_VCS2_IRQ_SHIFT,
2120 .init = logical_ring_init,
2123 .name = "video enhancement ring",
2124 .exec_id = I915_EXEC_VEBOX,
2125 .guc_id = GUC_VIDEOENHANCE_ENGINE,
2126 .mmio_base = VEBOX_RING_BASE,
2127 .irq_shift = GEN8_VECS_IRQ_SHIFT,
2128 .init = logical_ring_init,
2132 static struct intel_engine_cs *
2133 logical_ring_setup(struct drm_i915_private *dev_priv, enum intel_engine_id id)
2135 const struct logical_ring_info *info = &logical_rings[id];
2136 struct intel_engine_cs *engine = &dev_priv->engine[id];
2137 enum forcewake_domains fw_domains;
2140 engine->name = info->name;
2141 engine->exec_id = info->exec_id;
2142 engine->guc_id = info->guc_id;
2143 engine->mmio_base = info->mmio_base;
2145 engine->i915 = dev_priv;
2147 /* Intentionally left blank. */
2148 engine->buffer = NULL;
2150 fw_domains = intel_uncore_forcewake_for_reg(dev_priv,
2154 fw_domains |= intel_uncore_forcewake_for_reg(dev_priv,
2155 RING_CONTEXT_STATUS_PTR(engine),
2156 FW_REG_READ | FW_REG_WRITE);
2158 fw_domains |= intel_uncore_forcewake_for_reg(dev_priv,
2159 RING_CONTEXT_STATUS_BUF_BASE(engine),
2162 engine->fw_domains = fw_domains;
2164 INIT_LIST_HEAD(&engine->active_list);
2165 INIT_LIST_HEAD(&engine->request_list);
2166 INIT_LIST_HEAD(&engine->buffers);
2167 INIT_LIST_HEAD(&engine->execlist_queue);
2168 spin_lock_init(&engine->execlist_lock);
2170 tasklet_init(&engine->irq_tasklet,
2171 intel_lrc_irq_handler, (unsigned long)engine);
2173 logical_ring_init_platform_invariants(engine);
2174 logical_ring_default_vfuncs(engine);
2175 logical_ring_default_irqs(engine, info->irq_shift);
2177 intel_engine_init_hangcheck(engine);
2178 i915_gem_batch_pool_init(dev_priv->dev, &engine->batch_pool);
2184 * intel_logical_rings_init() - allocate, populate and init the Engine Command Streamers
2187 * This function inits the engines for an Execlists submission style (the
2188 * equivalent in the legacy ringbuffer submission world would be
2189 * i915_gem_init_engines). It does it only for those engines that are present in
2192 * Return: non-zero if the initialization failed.
2194 int intel_logical_rings_init(struct drm_device *dev)
2196 struct drm_i915_private *dev_priv = dev->dev_private;
2197 unsigned int mask = 0;
2201 WARN_ON(INTEL_INFO(dev_priv)->ring_mask &
2202 GENMASK(sizeof(mask) * BITS_PER_BYTE - 1, I915_NUM_ENGINES));
2204 for (i = 0; i < ARRAY_SIZE(logical_rings); i++) {
2205 if (!HAS_ENGINE(dev_priv, i))
2208 if (!logical_rings[i].init)
2211 ret = logical_rings[i].init(logical_ring_setup(dev_priv, i));
2215 mask |= ENGINE_MASK(i);
2219 * Catch failures to update logical_rings table when the new engines
2220 * are added to the driver by a warning and disabling the forgotten
2223 if (WARN_ON(mask != INTEL_INFO(dev_priv)->ring_mask)) {
2224 struct intel_device_info *info =
2225 (struct intel_device_info *)&dev_priv->info;
2226 info->ring_mask = mask;
2232 for (i = 0; i < I915_NUM_ENGINES; i++)
2233 intel_logical_ring_cleanup(&dev_priv->engine[i]);
2239 make_rpcs(struct drm_i915_private *dev_priv)
2244 * No explicit RPCS request is needed to ensure full
2245 * slice/subslice/EU enablement prior to Gen9.
2247 if (INTEL_GEN(dev_priv) < 9)
2251 * Starting in Gen9, render power gating can leave
2252 * slice/subslice/EU in a partially enabled state. We
2253 * must make an explicit request through RPCS for full
2256 if (INTEL_INFO(dev_priv)->has_slice_pg) {
2257 rpcs |= GEN8_RPCS_S_CNT_ENABLE;
2258 rpcs |= INTEL_INFO(dev_priv)->slice_total <<
2259 GEN8_RPCS_S_CNT_SHIFT;
2260 rpcs |= GEN8_RPCS_ENABLE;
2263 if (INTEL_INFO(dev_priv)->has_subslice_pg) {
2264 rpcs |= GEN8_RPCS_SS_CNT_ENABLE;
2265 rpcs |= INTEL_INFO(dev_priv)->subslice_per_slice <<
2266 GEN8_RPCS_SS_CNT_SHIFT;
2267 rpcs |= GEN8_RPCS_ENABLE;
2270 if (INTEL_INFO(dev_priv)->has_eu_pg) {
2271 rpcs |= INTEL_INFO(dev_priv)->eu_per_subslice <<
2272 GEN8_RPCS_EU_MIN_SHIFT;
2273 rpcs |= INTEL_INFO(dev_priv)->eu_per_subslice <<
2274 GEN8_RPCS_EU_MAX_SHIFT;
2275 rpcs |= GEN8_RPCS_ENABLE;
2281 static u32 intel_lr_indirect_ctx_offset(struct intel_engine_cs *engine)
2283 u32 indirect_ctx_offset;
2285 switch (INTEL_GEN(engine->i915)) {
2287 MISSING_CASE(INTEL_GEN(engine->i915));
2290 indirect_ctx_offset =
2291 GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
2294 indirect_ctx_offset =
2295 GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
2299 return indirect_ctx_offset;
2303 populate_lr_context(struct i915_gem_context *ctx,
2304 struct drm_i915_gem_object *ctx_obj,
2305 struct intel_engine_cs *engine,
2306 struct intel_ringbuffer *ringbuf)
2308 struct drm_i915_private *dev_priv = ctx->i915;
2309 struct i915_hw_ppgtt *ppgtt = ctx->ppgtt;
2315 ppgtt = dev_priv->mm.aliasing_ppgtt;
2317 ret = i915_gem_object_set_to_cpu_domain(ctx_obj, true);
2319 DRM_DEBUG_DRIVER("Could not set to CPU domain\n");
2323 vaddr = i915_gem_object_pin_map(ctx_obj);
2324 if (IS_ERR(vaddr)) {
2325 ret = PTR_ERR(vaddr);
2326 DRM_DEBUG_DRIVER("Could not map object pages! (%d)\n", ret);
2329 ctx_obj->dirty = true;
2331 /* The second page of the context object contains some fields which must
2332 * be set up prior to the first execution. */
2333 reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
2335 /* A context is actually a big batch buffer with several MI_LOAD_REGISTER_IMM
2336 * commands followed by (reg, value) pairs. The values we are setting here are
2337 * only for the first context restore: on a subsequent save, the GPU will
2338 * recreate this batchbuffer with new values (including all the missing
2339 * MI_LOAD_REGISTER_IMM commands that we are not initializing here). */
2340 reg_state[CTX_LRI_HEADER_0] =
2341 MI_LOAD_REGISTER_IMM(engine->id == RCS ? 14 : 11) | MI_LRI_FORCE_POSTED;
2342 ASSIGN_CTX_REG(reg_state, CTX_CONTEXT_CONTROL,
2343 RING_CONTEXT_CONTROL(engine),
2344 _MASKED_BIT_ENABLE(CTX_CTRL_INHIBIT_SYN_CTX_SWITCH |
2345 CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT |
2346 (HAS_RESOURCE_STREAMER(dev_priv) ?
2347 CTX_CTRL_RS_CTX_ENABLE : 0)));
2348 ASSIGN_CTX_REG(reg_state, CTX_RING_HEAD, RING_HEAD(engine->mmio_base),
2350 ASSIGN_CTX_REG(reg_state, CTX_RING_TAIL, RING_TAIL(engine->mmio_base),
2352 /* Ring buffer start address is not known until the buffer is pinned.
2353 * It is written to the context image in execlists_update_context()
2355 ASSIGN_CTX_REG(reg_state, CTX_RING_BUFFER_START,
2356 RING_START(engine->mmio_base), 0);
2357 ASSIGN_CTX_REG(reg_state, CTX_RING_BUFFER_CONTROL,
2358 RING_CTL(engine->mmio_base),
2359 ((ringbuf->size - PAGE_SIZE) & RING_NR_PAGES) | RING_VALID);
2360 ASSIGN_CTX_REG(reg_state, CTX_BB_HEAD_U,
2361 RING_BBADDR_UDW(engine->mmio_base), 0);
2362 ASSIGN_CTX_REG(reg_state, CTX_BB_HEAD_L,
2363 RING_BBADDR(engine->mmio_base), 0);
2364 ASSIGN_CTX_REG(reg_state, CTX_BB_STATE,
2365 RING_BBSTATE(engine->mmio_base),
2367 ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_HEAD_U,
2368 RING_SBBADDR_UDW(engine->mmio_base), 0);
2369 ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_HEAD_L,
2370 RING_SBBADDR(engine->mmio_base), 0);
2371 ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_STATE,
2372 RING_SBBSTATE(engine->mmio_base), 0);
2373 if (engine->id == RCS) {
2374 ASSIGN_CTX_REG(reg_state, CTX_BB_PER_CTX_PTR,
2375 RING_BB_PER_CTX_PTR(engine->mmio_base), 0);
2376 ASSIGN_CTX_REG(reg_state, CTX_RCS_INDIRECT_CTX,
2377 RING_INDIRECT_CTX(engine->mmio_base), 0);
2378 ASSIGN_CTX_REG(reg_state, CTX_RCS_INDIRECT_CTX_OFFSET,
2379 RING_INDIRECT_CTX_OFFSET(engine->mmio_base), 0);
2380 if (engine->wa_ctx.obj) {
2381 struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx;
2382 uint32_t ggtt_offset = i915_gem_obj_ggtt_offset(wa_ctx->obj);
2384 reg_state[CTX_RCS_INDIRECT_CTX+1] =
2385 (ggtt_offset + wa_ctx->indirect_ctx.offset * sizeof(uint32_t)) |
2386 (wa_ctx->indirect_ctx.size / CACHELINE_DWORDS);
2388 reg_state[CTX_RCS_INDIRECT_CTX_OFFSET+1] =
2389 intel_lr_indirect_ctx_offset(engine) << 6;
2391 reg_state[CTX_BB_PER_CTX_PTR+1] =
2392 (ggtt_offset + wa_ctx->per_ctx.offset * sizeof(uint32_t)) |
2396 reg_state[CTX_LRI_HEADER_1] = MI_LOAD_REGISTER_IMM(9) | MI_LRI_FORCE_POSTED;
2397 ASSIGN_CTX_REG(reg_state, CTX_CTX_TIMESTAMP,
2398 RING_CTX_TIMESTAMP(engine->mmio_base), 0);
2399 /* PDP values well be assigned later if needed */
2400 ASSIGN_CTX_REG(reg_state, CTX_PDP3_UDW, GEN8_RING_PDP_UDW(engine, 3),
2402 ASSIGN_CTX_REG(reg_state, CTX_PDP3_LDW, GEN8_RING_PDP_LDW(engine, 3),
2404 ASSIGN_CTX_REG(reg_state, CTX_PDP2_UDW, GEN8_RING_PDP_UDW(engine, 2),
2406 ASSIGN_CTX_REG(reg_state, CTX_PDP2_LDW, GEN8_RING_PDP_LDW(engine, 2),
2408 ASSIGN_CTX_REG(reg_state, CTX_PDP1_UDW, GEN8_RING_PDP_UDW(engine, 1),
2410 ASSIGN_CTX_REG(reg_state, CTX_PDP1_LDW, GEN8_RING_PDP_LDW(engine, 1),
2412 ASSIGN_CTX_REG(reg_state, CTX_PDP0_UDW, GEN8_RING_PDP_UDW(engine, 0),
2414 ASSIGN_CTX_REG(reg_state, CTX_PDP0_LDW, GEN8_RING_PDP_LDW(engine, 0),
2417 if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
2418 /* 64b PPGTT (48bit canonical)
2419 * PDP0_DESCRIPTOR contains the base address to PML4 and
2420 * other PDP Descriptors are ignored.
2422 ASSIGN_CTX_PML4(ppgtt, reg_state);
2425 * PDP*_DESCRIPTOR contains the base address of space supported.
2426 * With dynamic page allocation, PDPs may not be allocated at
2427 * this point. Point the unallocated PDPs to the scratch page
2429 execlists_update_context_pdps(ppgtt, reg_state);
2432 if (engine->id == RCS) {
2433 reg_state[CTX_LRI_HEADER_2] = MI_LOAD_REGISTER_IMM(1);
2434 ASSIGN_CTX_REG(reg_state, CTX_R_PWR_CLK_STATE, GEN8_R_PWR_CLK_STATE,
2435 make_rpcs(dev_priv));
2438 i915_gem_object_unpin_map(ctx_obj);
2444 * intel_lr_context_size() - return the size of the context for an engine
2445 * @engine: which engine to find the context size for
2447 * Each engine may require a different amount of space for a context image,
2448 * so when allocating (or copying) an image, this function can be used to
2449 * find the right size for the specific engine.
2451 * Return: size (in bytes) of an engine-specific context image
2453 * Note: this size includes the HWSP, which is part of the context image
2454 * in LRC mode, but does not include the "shared data page" used with
2455 * GuC submission. The caller should account for this if using the GuC.
2457 uint32_t intel_lr_context_size(struct intel_engine_cs *engine)
2461 WARN_ON(INTEL_GEN(engine->i915) < 8);
2463 switch (engine->id) {
2465 if (INTEL_GEN(engine->i915) >= 9)
2466 ret = GEN9_LR_CONTEXT_RENDER_SIZE;
2468 ret = GEN8_LR_CONTEXT_RENDER_SIZE;
2474 ret = GEN8_LR_CONTEXT_OTHER_SIZE;
2482 * execlists_context_deferred_alloc() - create the LRC specific bits of a context
2483 * @ctx: LR context to create.
2484 * @engine: engine to be used with the context.
2486 * This function can be called more than once, with different engines, if we plan
2487 * to use the context with them. The context backing objects and the ringbuffers
2488 * (specially the ringbuffer backing objects) suck a lot of memory up, and that's why
2489 * the creation is a deferred call: it's better to make sure first that we need to use
2490 * a given ring with the context.
2492 * Return: non-zero on error.
2494 static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
2495 struct intel_engine_cs *engine)
2497 struct drm_i915_gem_object *ctx_obj;
2498 struct intel_context *ce = &ctx->engine[engine->id];
2499 uint32_t context_size;
2500 struct intel_ringbuffer *ringbuf;
2505 context_size = round_up(intel_lr_context_size(engine), 4096);
2507 /* One extra page as the sharing data between driver and GuC */
2508 context_size += PAGE_SIZE * LRC_PPHWSP_PN;
2510 ctx_obj = i915_gem_object_create(ctx->i915->dev, context_size);
2511 if (IS_ERR(ctx_obj)) {
2512 DRM_DEBUG_DRIVER("Alloc LRC backing obj failed.\n");
2513 return PTR_ERR(ctx_obj);
2516 ringbuf = intel_engine_create_ringbuffer(engine, ctx->ring_size);
2517 if (IS_ERR(ringbuf)) {
2518 ret = PTR_ERR(ringbuf);
2519 goto error_deref_obj;
2522 ret = populate_lr_context(ctx, ctx_obj, engine, ringbuf);
2524 DRM_DEBUG_DRIVER("Failed to populate LRC: %d\n", ret);
2528 ce->ringbuf = ringbuf;
2529 ce->state = ctx_obj;
2530 ce->initialised = engine->init_context == NULL;
2535 intel_ringbuffer_free(ringbuf);
2537 drm_gem_object_unreference(&ctx_obj->base);
2543 void intel_lr_context_reset(struct drm_i915_private *dev_priv,
2544 struct i915_gem_context *ctx)
2546 struct intel_engine_cs *engine;
2548 for_each_engine(engine, dev_priv) {
2549 struct intel_context *ce = &ctx->engine[engine->id];
2550 struct drm_i915_gem_object *ctx_obj = ce->state;
2552 uint32_t *reg_state;
2557 vaddr = i915_gem_object_pin_map(ctx_obj);
2558 if (WARN_ON(IS_ERR(vaddr)))
2561 reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
2562 ctx_obj->dirty = true;
2564 reg_state[CTX_RING_HEAD+1] = 0;
2565 reg_state[CTX_RING_TAIL+1] = 0;
2567 i915_gem_object_unpin_map(ctx_obj);
2569 ce->ringbuf->head = 0;
2570 ce->ringbuf->tail = 0;