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
291 * @ctx: Context to work on
292 * @engine: Engine the descriptor will be used with
294 * The context descriptor encodes various attributes of a context,
295 * including its GTT address and some flags. Because it's fairly
296 * expensive to calculate, we'll just do it once and cache the result,
297 * which remains valid until the context is unpinned.
299 * 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] = intel_ring_offset(rq->ring, 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_elsp_submit_contexts(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_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_put(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.
486 req0->tail &= req0->ring->size - 1;
489 execlists_elsp_submit_contexts(req0, req1);
493 execlists_check_remove_request(struct intel_engine_cs *engine, u32 ctx_id)
495 struct drm_i915_gem_request *head_req;
497 assert_spin_locked(&engine->execlist_lock);
499 head_req = list_first_entry_or_null(&engine->execlist_queue,
500 struct drm_i915_gem_request,
503 if (WARN_ON(!head_req || (head_req->ctx_hw_id != ctx_id)))
506 WARN(head_req->elsp_submitted == 0, "Never submitted head request\n");
508 if (--head_req->elsp_submitted > 0)
511 execlists_context_status_change(head_req, INTEL_CONTEXT_SCHEDULE_OUT);
513 list_del(&head_req->execlist_link);
514 i915_gem_request_put(head_req);
520 get_context_status(struct intel_engine_cs *engine, unsigned int read_pointer,
523 struct drm_i915_private *dev_priv = engine->i915;
526 read_pointer %= GEN8_CSB_ENTRIES;
528 status = I915_READ_FW(RING_CONTEXT_STATUS_BUF_LO(engine, read_pointer));
530 if (status & GEN8_CTX_STATUS_IDLE_ACTIVE)
533 *context_id = I915_READ_FW(RING_CONTEXT_STATUS_BUF_HI(engine,
540 * Check the unread Context Status Buffers and manage the submission of new
541 * contexts to the ELSP accordingly.
543 static void intel_lrc_irq_handler(unsigned long data)
545 struct intel_engine_cs *engine = (struct intel_engine_cs *)data;
546 struct drm_i915_private *dev_priv = engine->i915;
548 unsigned int read_pointer, write_pointer;
549 u32 csb[GEN8_CSB_ENTRIES][2];
550 unsigned int csb_read = 0, i;
551 unsigned int submit_contexts = 0;
553 intel_uncore_forcewake_get(dev_priv, engine->fw_domains);
555 status_pointer = I915_READ_FW(RING_CONTEXT_STATUS_PTR(engine));
557 read_pointer = engine->next_context_status_buffer;
558 write_pointer = GEN8_CSB_WRITE_PTR(status_pointer);
559 if (read_pointer > write_pointer)
560 write_pointer += GEN8_CSB_ENTRIES;
562 while (read_pointer < write_pointer) {
563 if (WARN_ON_ONCE(csb_read == GEN8_CSB_ENTRIES))
565 csb[csb_read][0] = get_context_status(engine, ++read_pointer,
570 engine->next_context_status_buffer = write_pointer % GEN8_CSB_ENTRIES;
572 /* Update the read pointer to the old write pointer. Manual ringbuffer
573 * management ftw </sarcasm> */
574 I915_WRITE_FW(RING_CONTEXT_STATUS_PTR(engine),
575 _MASKED_FIELD(GEN8_CSB_READ_PTR_MASK,
576 engine->next_context_status_buffer << 8));
578 intel_uncore_forcewake_put(dev_priv, engine->fw_domains);
580 spin_lock(&engine->execlist_lock);
582 for (i = 0; i < csb_read; i++) {
583 if (unlikely(csb[i][0] & GEN8_CTX_STATUS_PREEMPTED)) {
584 if (csb[i][0] & GEN8_CTX_STATUS_LITE_RESTORE) {
585 if (execlists_check_remove_request(engine, csb[i][1]))
586 WARN(1, "Lite Restored request removed from queue\n");
588 WARN(1, "Preemption without Lite Restore\n");
591 if (csb[i][0] & (GEN8_CTX_STATUS_ACTIVE_IDLE |
592 GEN8_CTX_STATUS_ELEMENT_SWITCH))
594 execlists_check_remove_request(engine, csb[i][1]);
597 if (submit_contexts) {
598 if (!engine->disable_lite_restore_wa ||
599 (csb[i][0] & GEN8_CTX_STATUS_ACTIVE_IDLE))
600 execlists_unqueue(engine);
603 spin_unlock(&engine->execlist_lock);
605 if (unlikely(submit_contexts > 2))
606 DRM_ERROR("More than two context complete events?\n");
609 static void execlists_submit_request(struct drm_i915_gem_request *request)
611 struct intel_engine_cs *engine = request->engine;
612 struct drm_i915_gem_request *cursor;
613 int num_elements = 0;
615 spin_lock_bh(&engine->execlist_lock);
617 list_for_each_entry(cursor, &engine->execlist_queue, execlist_link)
618 if (++num_elements > 2)
621 if (num_elements > 2) {
622 struct drm_i915_gem_request *tail_req;
624 tail_req = list_last_entry(&engine->execlist_queue,
625 struct drm_i915_gem_request,
628 if (request->ctx == tail_req->ctx) {
629 WARN(tail_req->elsp_submitted != 0,
630 "More than 2 already-submitted reqs queued\n");
631 list_del(&tail_req->execlist_link);
632 i915_gem_request_put(tail_req);
636 i915_gem_request_get(request);
637 list_add_tail(&request->execlist_link, &engine->execlist_queue);
638 request->ctx_hw_id = request->ctx->hw_id;
639 if (num_elements == 0)
640 execlists_unqueue(engine);
642 spin_unlock_bh(&engine->execlist_lock);
645 int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *request)
647 struct intel_engine_cs *engine = request->engine;
648 struct intel_context *ce = &request->ctx->engine[engine->id];
651 /* Flush enough space to reduce the likelihood of waiting after
652 * we start building the request - in which case we will just
653 * have to repeat work.
655 request->reserved_space += EXECLISTS_REQUEST_SIZE;
658 ret = execlists_context_deferred_alloc(request->ctx, engine);
663 request->ring = ce->ring;
665 if (i915.enable_guc_submission) {
667 * Check that the GuC has space for the request before
668 * going any further, as the i915_add_request() call
669 * later on mustn't fail ...
671 ret = i915_guc_wq_check_space(request);
676 ret = intel_lr_context_pin(request->ctx, engine);
680 ret = intel_ring_begin(request, 0);
684 if (!ce->initialised) {
685 ret = engine->init_context(request);
689 ce->initialised = true;
692 /* Note that after this point, we have committed to using
693 * this request as it is being used to both track the
694 * state of engine initialisation and liveness of the
695 * golden renderstate above. Think twice before you try
696 * to cancel/unwind this request now.
699 request->reserved_space -= EXECLISTS_REQUEST_SIZE;
703 intel_lr_context_unpin(request->ctx, engine);
708 * intel_logical_ring_advance() - advance the tail and prepare for submission
709 * @request: Request to advance the logical ringbuffer of.
711 * The tail is updated in our logical ringbuffer struct, not in the actual context. What
712 * really happens during submission is that the context and current tail will be placed
713 * on a queue waiting for the ELSP to be ready to accept a new context submission. At that
714 * point, the tail *inside* the context is updated and the ELSP written to.
717 intel_logical_ring_advance(struct drm_i915_gem_request *request)
719 struct intel_ring *ring = request->ring;
720 struct intel_engine_cs *engine = request->engine;
722 intel_ring_advance(ring);
723 request->tail = ring->tail;
726 * Here we add two extra NOOPs as padding to avoid
727 * lite restore of a context with HEAD==TAIL.
729 * Caller must reserve WA_TAIL_DWORDS for us!
731 intel_ring_emit(ring, MI_NOOP);
732 intel_ring_emit(ring, MI_NOOP);
733 intel_ring_advance(ring);
735 /* We keep the previous context alive until we retire the following
736 * request. This ensures that any the context object is still pinned
737 * for any residual writes the HW makes into it on the context switch
738 * into the next object following the breadcrumb. Otherwise, we may
739 * retire the context too early.
741 request->previous_context = engine->last_context;
742 engine->last_context = request->ctx;
746 void intel_execlists_cancel_requests(struct intel_engine_cs *engine)
748 struct drm_i915_gem_request *req, *tmp;
749 LIST_HEAD(cancel_list);
751 WARN_ON(!mutex_is_locked(&engine->i915->drm.struct_mutex));
753 spin_lock_bh(&engine->execlist_lock);
754 list_replace_init(&engine->execlist_queue, &cancel_list);
755 spin_unlock_bh(&engine->execlist_lock);
757 list_for_each_entry_safe(req, tmp, &cancel_list, execlist_link) {
758 list_del(&req->execlist_link);
759 i915_gem_request_put(req);
763 void intel_logical_ring_stop(struct intel_engine_cs *engine)
765 struct drm_i915_private *dev_priv = engine->i915;
768 if (!intel_engine_initialized(engine))
771 ret = intel_engine_idle(engine);
773 DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n",
776 /* TODO: Is this correct with Execlists enabled? */
777 I915_WRITE_MODE(engine, _MASKED_BIT_ENABLE(STOP_RING));
778 if (intel_wait_for_register(dev_priv,
779 RING_MI_MODE(engine->mmio_base),
780 MODE_IDLE, MODE_IDLE,
782 DRM_ERROR("%s :timed out trying to stop ring\n", engine->name);
785 I915_WRITE_MODE(engine, _MASKED_BIT_DISABLE(STOP_RING));
788 static int intel_lr_context_pin(struct i915_gem_context *ctx,
789 struct intel_engine_cs *engine)
791 struct drm_i915_private *dev_priv = ctx->i915;
792 struct intel_context *ce = &ctx->engine[engine->id];
797 lockdep_assert_held(&ctx->i915->drm.struct_mutex);
802 ret = i915_gem_object_ggtt_pin(ce->state, NULL,
803 0, GEN8_LR_CONTEXT_ALIGN,
804 PIN_OFFSET_BIAS | GUC_WOPCM_TOP);
808 vaddr = i915_gem_object_pin_map(ce->state);
810 ret = PTR_ERR(vaddr);
814 lrc_reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
816 ret = intel_ring_pin(ce->ring);
820 ce->lrc_vma = i915_gem_obj_to_ggtt(ce->state);
821 intel_lr_context_descriptor_update(ctx, engine);
823 lrc_reg_state[CTX_RING_BUFFER_START+1] = ce->ring->vma->node.start;
824 ce->lrc_reg_state = lrc_reg_state;
825 ce->state->dirty = true;
827 /* Invalidate GuC TLB. */
828 if (i915.enable_guc_submission)
829 I915_WRITE(GEN8_GTCR, GEN8_GTCR_INVALIDATE);
831 i915_gem_context_get(ctx);
835 i915_gem_object_unpin_map(ce->state);
837 i915_gem_object_ggtt_unpin(ce->state);
843 void intel_lr_context_unpin(struct i915_gem_context *ctx,
844 struct intel_engine_cs *engine)
846 struct intel_context *ce = &ctx->engine[engine->id];
848 lockdep_assert_held(&ctx->i915->drm.struct_mutex);
849 GEM_BUG_ON(ce->pin_count == 0);
854 intel_ring_unpin(ce->ring);
856 i915_gem_object_unpin_map(ce->state);
857 i915_gem_object_ggtt_unpin(ce->state);
861 ce->lrc_reg_state = NULL;
863 i915_gem_context_put(ctx);
866 static int intel_logical_ring_workarounds_emit(struct drm_i915_gem_request *req)
869 struct intel_ring *ring = req->ring;
870 struct i915_workarounds *w = &req->i915->workarounds;
875 ret = req->engine->emit_flush(req, EMIT_BARRIER);
879 ret = intel_ring_begin(req, w->count * 2 + 2);
883 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(w->count));
884 for (i = 0; i < w->count; i++) {
885 intel_ring_emit_reg(ring, w->reg[i].addr);
886 intel_ring_emit(ring, w->reg[i].value);
888 intel_ring_emit(ring, MI_NOOP);
890 intel_ring_advance(ring);
892 ret = req->engine->emit_flush(req, EMIT_BARRIER);
899 #define wa_ctx_emit(batch, index, cmd) \
901 int __index = (index)++; \
902 if (WARN_ON(__index >= (PAGE_SIZE / sizeof(uint32_t)))) { \
905 batch[__index] = (cmd); \
908 #define wa_ctx_emit_reg(batch, index, reg) \
909 wa_ctx_emit((batch), (index), i915_mmio_reg_offset(reg))
912 * In this WA we need to set GEN8_L3SQCREG4[21:21] and reset it after
913 * PIPE_CONTROL instruction. This is required for the flush to happen correctly
914 * but there is a slight complication as this is applied in WA batch where the
915 * values are only initialized once so we cannot take register value at the
916 * beginning and reuse it further; hence we save its value to memory, upload a
917 * constant value with bit21 set and then we restore it back with the saved value.
918 * To simplify the WA, a constant value is formed by using the default value
919 * of this register. This shouldn't be a problem because we are only modifying
920 * it for a short period and this batch in non-premptible. We can ofcourse
921 * use additional instructions that read the actual value of the register
922 * at that time and set our bit of interest but it makes the WA complicated.
924 * This WA is also required for Gen9 so extracting as a function avoids
927 static inline int gen8_emit_flush_coherentl3_wa(struct intel_engine_cs *engine,
931 struct drm_i915_private *dev_priv = engine->i915;
932 uint32_t l3sqc4_flush = (0x40400000 | GEN8_LQSC_FLUSH_COHERENT_LINES);
935 * WaDisableLSQCROPERFforOCL:skl,kbl
936 * This WA is implemented in skl_init_clock_gating() but since
937 * this batch updates GEN8_L3SQCREG4 with default value we need to
938 * set this bit here to retain the WA during flush.
940 if (IS_SKL_REVID(dev_priv, 0, SKL_REVID_E0) ||
941 IS_KBL_REVID(dev_priv, 0, KBL_REVID_E0))
942 l3sqc4_flush |= GEN8_LQSC_RO_PERF_DIS;
944 wa_ctx_emit(batch, index, (MI_STORE_REGISTER_MEM_GEN8 |
945 MI_SRM_LRM_GLOBAL_GTT));
946 wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
947 wa_ctx_emit(batch, index, engine->scratch.gtt_offset + 256);
948 wa_ctx_emit(batch, index, 0);
950 wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1));
951 wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
952 wa_ctx_emit(batch, index, l3sqc4_flush);
954 wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
955 wa_ctx_emit(batch, index, (PIPE_CONTROL_CS_STALL |
956 PIPE_CONTROL_DC_FLUSH_ENABLE));
957 wa_ctx_emit(batch, index, 0);
958 wa_ctx_emit(batch, index, 0);
959 wa_ctx_emit(batch, index, 0);
960 wa_ctx_emit(batch, index, 0);
962 wa_ctx_emit(batch, index, (MI_LOAD_REGISTER_MEM_GEN8 |
963 MI_SRM_LRM_GLOBAL_GTT));
964 wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
965 wa_ctx_emit(batch, index, engine->scratch.gtt_offset + 256);
966 wa_ctx_emit(batch, index, 0);
971 static inline uint32_t wa_ctx_start(struct i915_wa_ctx_bb *wa_ctx,
973 uint32_t start_alignment)
975 return wa_ctx->offset = ALIGN(offset, start_alignment);
978 static inline int wa_ctx_end(struct i915_wa_ctx_bb *wa_ctx,
980 uint32_t size_alignment)
982 wa_ctx->size = offset - wa_ctx->offset;
984 WARN(wa_ctx->size % size_alignment,
985 "wa_ctx_bb failed sanity checks: size %d is not aligned to %d\n",
986 wa_ctx->size, size_alignment);
991 * Typically we only have one indirect_ctx and per_ctx batch buffer which are
992 * initialized at the beginning and shared across all contexts but this field
993 * helps us to have multiple batches at different offsets and select them based
994 * on a criteria. At the moment this batch always start at the beginning of the page
995 * and at this point we don't have multiple wa_ctx batch buffers.
997 * The number of WA applied are not known at the beginning; we use this field
998 * to return the no of DWORDS written.
1000 * It is to be noted that this batch does not contain MI_BATCH_BUFFER_END
1001 * so it adds NOOPs as padding to make it cacheline aligned.
1002 * MI_BATCH_BUFFER_END will be added to perctx batch and both of them together
1003 * makes a complete batch buffer.
1005 static int gen8_init_indirectctx_bb(struct intel_engine_cs *engine,
1006 struct i915_wa_ctx_bb *wa_ctx,
1010 uint32_t scratch_addr;
1011 uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1013 /* WaDisableCtxRestoreArbitration:bdw,chv */
1014 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_DISABLE);
1016 /* WaFlushCoherentL3CacheLinesAtContextSwitch:bdw */
1017 if (IS_BROADWELL(engine->i915)) {
1018 int rc = gen8_emit_flush_coherentl3_wa(engine, batch, index);
1024 /* WaClearSlmSpaceAtContextSwitch:bdw,chv */
1025 /* Actual scratch location is at 128 bytes offset */
1026 scratch_addr = engine->scratch.gtt_offset + 2*CACHELINE_BYTES;
1028 wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
1029 wa_ctx_emit(batch, index, (PIPE_CONTROL_FLUSH_L3 |
1030 PIPE_CONTROL_GLOBAL_GTT_IVB |
1031 PIPE_CONTROL_CS_STALL |
1032 PIPE_CONTROL_QW_WRITE));
1033 wa_ctx_emit(batch, index, scratch_addr);
1034 wa_ctx_emit(batch, index, 0);
1035 wa_ctx_emit(batch, index, 0);
1036 wa_ctx_emit(batch, index, 0);
1038 /* Pad to end of cacheline */
1039 while (index % CACHELINE_DWORDS)
1040 wa_ctx_emit(batch, index, MI_NOOP);
1043 * MI_BATCH_BUFFER_END is not required in Indirect ctx BB because
1044 * execution depends on the length specified in terms of cache lines
1045 * in the register CTX_RCS_INDIRECT_CTX
1048 return wa_ctx_end(wa_ctx, *offset = index, CACHELINE_DWORDS);
1052 * This batch is started immediately after indirect_ctx batch. Since we ensure
1053 * that indirect_ctx ends on a cacheline this batch is aligned automatically.
1055 * The number of DWORDS written are returned using this field.
1057 * This batch is terminated with MI_BATCH_BUFFER_END and so we need not add padding
1058 * to align it with cacheline as padding after MI_BATCH_BUFFER_END is redundant.
1060 static int gen8_init_perctx_bb(struct intel_engine_cs *engine,
1061 struct i915_wa_ctx_bb *wa_ctx,
1065 uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1067 /* WaDisableCtxRestoreArbitration:bdw,chv */
1068 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_ENABLE);
1070 wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END);
1072 return wa_ctx_end(wa_ctx, *offset = index, 1);
1075 static int gen9_init_indirectctx_bb(struct intel_engine_cs *engine,
1076 struct i915_wa_ctx_bb *wa_ctx,
1081 struct drm_i915_private *dev_priv = engine->i915;
1082 uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1084 /* WaDisableCtxRestoreArbitration:skl,bxt */
1085 if (IS_SKL_REVID(dev_priv, 0, SKL_REVID_D0) ||
1086 IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1))
1087 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_DISABLE);
1089 /* WaFlushCoherentL3CacheLinesAtContextSwitch:skl,bxt */
1090 ret = gen8_emit_flush_coherentl3_wa(engine, batch, index);
1095 /* WaDisableGatherAtSetShaderCommonSlice:skl,bxt,kbl */
1096 wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1));
1097 wa_ctx_emit_reg(batch, index, COMMON_SLICE_CHICKEN2);
1098 wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(
1099 GEN9_DISABLE_GATHER_AT_SET_SHADER_COMMON_SLICE));
1100 wa_ctx_emit(batch, index, MI_NOOP);
1102 /* WaClearSlmSpaceAtContextSwitch:kbl */
1103 /* Actual scratch location is at 128 bytes offset */
1104 if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_A0)) {
1105 uint32_t scratch_addr
1106 = engine->scratch.gtt_offset + 2*CACHELINE_BYTES;
1108 wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
1109 wa_ctx_emit(batch, index, (PIPE_CONTROL_FLUSH_L3 |
1110 PIPE_CONTROL_GLOBAL_GTT_IVB |
1111 PIPE_CONTROL_CS_STALL |
1112 PIPE_CONTROL_QW_WRITE));
1113 wa_ctx_emit(batch, index, scratch_addr);
1114 wa_ctx_emit(batch, index, 0);
1115 wa_ctx_emit(batch, index, 0);
1116 wa_ctx_emit(batch, index, 0);
1119 /* WaMediaPoolStateCmdInWABB:bxt */
1120 if (HAS_POOLED_EU(engine->i915)) {
1122 * EU pool configuration is setup along with golden context
1123 * during context initialization. This value depends on
1124 * device type (2x6 or 3x6) and needs to be updated based
1125 * on which subslice is disabled especially for 2x6
1126 * devices, however it is safe to load default
1127 * configuration of 3x6 device instead of masking off
1128 * corresponding bits because HW ignores bits of a disabled
1129 * subslice and drops down to appropriate config. Please
1130 * see render_state_setup() in i915_gem_render_state.c for
1131 * possible configurations, to avoid duplication they are
1132 * not shown here again.
1134 u32 eu_pool_config = 0x00777000;
1135 wa_ctx_emit(batch, index, GEN9_MEDIA_POOL_STATE);
1136 wa_ctx_emit(batch, index, GEN9_MEDIA_POOL_ENABLE);
1137 wa_ctx_emit(batch, index, eu_pool_config);
1138 wa_ctx_emit(batch, index, 0);
1139 wa_ctx_emit(batch, index, 0);
1140 wa_ctx_emit(batch, index, 0);
1143 /* Pad to end of cacheline */
1144 while (index % CACHELINE_DWORDS)
1145 wa_ctx_emit(batch, index, MI_NOOP);
1147 return wa_ctx_end(wa_ctx, *offset = index, CACHELINE_DWORDS);
1150 static int gen9_init_perctx_bb(struct intel_engine_cs *engine,
1151 struct i915_wa_ctx_bb *wa_ctx,
1155 uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1157 /* WaSetDisablePixMaskCammingAndRhwoInCommonSliceChicken:skl,bxt */
1158 if (IS_SKL_REVID(engine->i915, 0, SKL_REVID_B0) ||
1159 IS_BXT_REVID(engine->i915, 0, BXT_REVID_A1)) {
1160 wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1));
1161 wa_ctx_emit_reg(batch, index, GEN9_SLICE_COMMON_ECO_CHICKEN0);
1162 wa_ctx_emit(batch, index,
1163 _MASKED_BIT_ENABLE(DISABLE_PIXEL_MASK_CAMMING));
1164 wa_ctx_emit(batch, index, MI_NOOP);
1167 /* WaClearTdlStateAckDirtyBits:bxt */
1168 if (IS_BXT_REVID(engine->i915, 0, BXT_REVID_B0)) {
1169 wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(4));
1171 wa_ctx_emit_reg(batch, index, GEN8_STATE_ACK);
1172 wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS));
1174 wa_ctx_emit_reg(batch, index, GEN9_STATE_ACK_SLICE1);
1175 wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS));
1177 wa_ctx_emit_reg(batch, index, GEN9_STATE_ACK_SLICE2);
1178 wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS));
1180 wa_ctx_emit_reg(batch, index, GEN7_ROW_CHICKEN2);
1181 /* dummy write to CS, mask bits are 0 to ensure the register is not modified */
1182 wa_ctx_emit(batch, index, 0x0);
1183 wa_ctx_emit(batch, index, MI_NOOP);
1186 /* WaDisableCtxRestoreArbitration:skl,bxt */
1187 if (IS_SKL_REVID(engine->i915, 0, SKL_REVID_D0) ||
1188 IS_BXT_REVID(engine->i915, 0, BXT_REVID_A1))
1189 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_ENABLE);
1191 wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END);
1193 return wa_ctx_end(wa_ctx, *offset = index, 1);
1196 static int lrc_setup_wa_ctx_obj(struct intel_engine_cs *engine, u32 size)
1200 engine->wa_ctx.obj = i915_gem_object_create(&engine->i915->drm,
1202 if (IS_ERR(engine->wa_ctx.obj)) {
1203 DRM_DEBUG_DRIVER("alloc LRC WA ctx backing obj failed.\n");
1204 ret = PTR_ERR(engine->wa_ctx.obj);
1205 engine->wa_ctx.obj = NULL;
1209 ret = i915_gem_object_ggtt_pin(engine->wa_ctx.obj, NULL,
1212 DRM_DEBUG_DRIVER("pin LRC WA ctx backing obj failed: %d\n",
1214 i915_gem_object_put(engine->wa_ctx.obj);
1221 static void lrc_destroy_wa_ctx_obj(struct intel_engine_cs *engine)
1223 if (engine->wa_ctx.obj) {
1224 i915_gem_object_ggtt_unpin(engine->wa_ctx.obj);
1225 i915_gem_object_put(engine->wa_ctx.obj);
1226 engine->wa_ctx.obj = NULL;
1230 static int intel_init_workaround_bb(struct intel_engine_cs *engine)
1236 struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx;
1238 WARN_ON(engine->id != RCS);
1240 /* update this when WA for higher Gen are added */
1241 if (INTEL_GEN(engine->i915) > 9) {
1242 DRM_ERROR("WA batch buffer is not initialized for Gen%d\n",
1243 INTEL_GEN(engine->i915));
1247 /* some WA perform writes to scratch page, ensure it is valid */
1248 if (engine->scratch.obj == NULL) {
1249 DRM_ERROR("scratch page not allocated for %s\n", engine->name);
1253 ret = lrc_setup_wa_ctx_obj(engine, PAGE_SIZE);
1255 DRM_DEBUG_DRIVER("Failed to setup context WA page: %d\n", ret);
1259 page = i915_gem_object_get_dirty_page(wa_ctx->obj, 0);
1260 batch = kmap_atomic(page);
1263 if (IS_GEN8(engine->i915)) {
1264 ret = gen8_init_indirectctx_bb(engine,
1265 &wa_ctx->indirect_ctx,
1271 ret = gen8_init_perctx_bb(engine,
1277 } else if (IS_GEN9(engine->i915)) {
1278 ret = gen9_init_indirectctx_bb(engine,
1279 &wa_ctx->indirect_ctx,
1285 ret = gen9_init_perctx_bb(engine,
1294 kunmap_atomic(batch);
1296 lrc_destroy_wa_ctx_obj(engine);
1301 static void lrc_init_hws(struct intel_engine_cs *engine)
1303 struct drm_i915_private *dev_priv = engine->i915;
1305 I915_WRITE(RING_HWS_PGA(engine->mmio_base),
1306 (u32)engine->status_page.gfx_addr);
1307 POSTING_READ(RING_HWS_PGA(engine->mmio_base));
1310 static int gen8_init_common_ring(struct intel_engine_cs *engine)
1312 struct drm_i915_private *dev_priv = engine->i915;
1313 unsigned int next_context_status_buffer_hw;
1315 lrc_init_hws(engine);
1317 I915_WRITE_IMR(engine,
1318 ~(engine->irq_enable_mask | engine->irq_keep_mask));
1319 I915_WRITE(RING_HWSTAM(engine->mmio_base), 0xffffffff);
1321 I915_WRITE(RING_MODE_GEN7(engine),
1322 _MASKED_BIT_DISABLE(GFX_REPLAY_MODE) |
1323 _MASKED_BIT_ENABLE(GFX_RUN_LIST_ENABLE));
1324 POSTING_READ(RING_MODE_GEN7(engine));
1327 * Instead of resetting the Context Status Buffer (CSB) read pointer to
1328 * zero, we need to read the write pointer from hardware and use its
1329 * value because "this register is power context save restored".
1330 * Effectively, these states have been observed:
1332 * | Suspend-to-idle (freeze) | Suspend-to-RAM (mem) |
1333 * BDW | CSB regs not reset | CSB regs reset |
1334 * CHT | CSB regs not reset | CSB regs not reset |
1338 next_context_status_buffer_hw =
1339 GEN8_CSB_WRITE_PTR(I915_READ(RING_CONTEXT_STATUS_PTR(engine)));
1342 * When the CSB registers are reset (also after power-up / gpu reset),
1343 * CSB write pointer is set to all 1's, which is not valid, use '5' in
1344 * this special case, so the first element read is CSB[0].
1346 if (next_context_status_buffer_hw == GEN8_CSB_PTR_MASK)
1347 next_context_status_buffer_hw = (GEN8_CSB_ENTRIES - 1);
1349 engine->next_context_status_buffer = next_context_status_buffer_hw;
1350 DRM_DEBUG_DRIVER("Execlists enabled for %s\n", engine->name);
1352 intel_engine_init_hangcheck(engine);
1354 return intel_mocs_init_engine(engine);
1357 static int gen8_init_render_ring(struct intel_engine_cs *engine)
1359 struct drm_i915_private *dev_priv = engine->i915;
1362 ret = gen8_init_common_ring(engine);
1366 /* We need to disable the AsyncFlip performance optimisations in order
1367 * to use MI_WAIT_FOR_EVENT within the CS. It should already be
1368 * programmed to '1' on all products.
1370 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw,chv
1372 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
1374 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
1376 return init_workarounds_ring(engine);
1379 static int gen9_init_render_ring(struct intel_engine_cs *engine)
1383 ret = gen8_init_common_ring(engine);
1387 return init_workarounds_ring(engine);
1390 static int intel_logical_ring_emit_pdps(struct drm_i915_gem_request *req)
1392 struct i915_hw_ppgtt *ppgtt = req->ctx->ppgtt;
1393 struct intel_ring *ring = req->ring;
1394 struct intel_engine_cs *engine = req->engine;
1395 const int num_lri_cmds = GEN8_LEGACY_PDPES * 2;
1398 ret = intel_ring_begin(req, num_lri_cmds * 2 + 2);
1402 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_lri_cmds));
1403 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
1404 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
1406 intel_ring_emit_reg(ring, GEN8_RING_PDP_UDW(engine, i));
1407 intel_ring_emit(ring, upper_32_bits(pd_daddr));
1408 intel_ring_emit_reg(ring, GEN8_RING_PDP_LDW(engine, i));
1409 intel_ring_emit(ring, lower_32_bits(pd_daddr));
1412 intel_ring_emit(ring, MI_NOOP);
1413 intel_ring_advance(ring);
1418 static int gen8_emit_bb_start(struct drm_i915_gem_request *req,
1419 u64 offset, u32 len,
1420 unsigned int dispatch_flags)
1422 struct intel_ring *ring = req->ring;
1423 bool ppgtt = !(dispatch_flags & I915_DISPATCH_SECURE);
1426 /* Don't rely in hw updating PDPs, specially in lite-restore.
1427 * Ideally, we should set Force PD Restore in ctx descriptor,
1428 * but we can't. Force Restore would be a second option, but
1429 * it is unsafe in case of lite-restore (because the ctx is
1430 * not idle). PML4 is allocated during ppgtt init so this is
1431 * not needed in 48-bit.*/
1432 if (req->ctx->ppgtt &&
1433 (intel_engine_flag(req->engine) & req->ctx->ppgtt->pd_dirty_rings)) {
1434 if (!USES_FULL_48BIT_PPGTT(req->i915) &&
1435 !intel_vgpu_active(req->i915)) {
1436 ret = intel_logical_ring_emit_pdps(req);
1441 req->ctx->ppgtt->pd_dirty_rings &= ~intel_engine_flag(req->engine);
1444 ret = intel_ring_begin(req, 4);
1448 /* FIXME(BDW): Address space and security selectors. */
1449 intel_ring_emit(ring, MI_BATCH_BUFFER_START_GEN8 |
1451 (dispatch_flags & I915_DISPATCH_RS ?
1452 MI_BATCH_RESOURCE_STREAMER : 0));
1453 intel_ring_emit(ring, lower_32_bits(offset));
1454 intel_ring_emit(ring, upper_32_bits(offset));
1455 intel_ring_emit(ring, MI_NOOP);
1456 intel_ring_advance(ring);
1461 static void gen8_logical_ring_enable_irq(struct intel_engine_cs *engine)
1463 struct drm_i915_private *dev_priv = engine->i915;
1464 I915_WRITE_IMR(engine,
1465 ~(engine->irq_enable_mask | engine->irq_keep_mask));
1466 POSTING_READ_FW(RING_IMR(engine->mmio_base));
1469 static void gen8_logical_ring_disable_irq(struct intel_engine_cs *engine)
1471 struct drm_i915_private *dev_priv = engine->i915;
1472 I915_WRITE_IMR(engine, ~engine->irq_keep_mask);
1475 static int gen8_emit_flush(struct drm_i915_gem_request *request, u32 mode)
1477 struct intel_ring *ring = request->ring;
1481 ret = intel_ring_begin(request, 4);
1485 cmd = MI_FLUSH_DW + 1;
1487 /* We always require a command barrier so that subsequent
1488 * commands, such as breadcrumb interrupts, are strictly ordered
1489 * wrt the contents of the write cache being flushed to memory
1490 * (and thus being coherent from the CPU).
1492 cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1494 if (mode & EMIT_INVALIDATE) {
1495 cmd |= MI_INVALIDATE_TLB;
1496 if (request->engine->id == VCS)
1497 cmd |= MI_INVALIDATE_BSD;
1500 intel_ring_emit(ring, cmd);
1501 intel_ring_emit(ring,
1502 I915_GEM_HWS_SCRATCH_ADDR |
1503 MI_FLUSH_DW_USE_GTT);
1504 intel_ring_emit(ring, 0); /* upper addr */
1505 intel_ring_emit(ring, 0); /* value */
1506 intel_ring_advance(ring);
1511 static int gen8_emit_flush_render(struct drm_i915_gem_request *request,
1514 struct intel_ring *ring = request->ring;
1515 struct intel_engine_cs *engine = request->engine;
1516 u32 scratch_addr = engine->scratch.gtt_offset + 2 * CACHELINE_BYTES;
1517 bool vf_flush_wa = false, dc_flush_wa = false;
1522 flags |= PIPE_CONTROL_CS_STALL;
1524 if (mode & EMIT_FLUSH) {
1525 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
1526 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
1527 flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
1528 flags |= PIPE_CONTROL_FLUSH_ENABLE;
1531 if (mode & EMIT_INVALIDATE) {
1532 flags |= PIPE_CONTROL_TLB_INVALIDATE;
1533 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
1534 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1535 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
1536 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
1537 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
1538 flags |= PIPE_CONTROL_QW_WRITE;
1539 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
1542 * On GEN9: before VF_CACHE_INVALIDATE we need to emit a NULL
1545 if (IS_GEN9(request->i915))
1548 /* WaForGAMHang:kbl */
1549 if (IS_KBL_REVID(request->i915, 0, KBL_REVID_B0))
1561 ret = intel_ring_begin(request, len);
1566 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6));
1567 intel_ring_emit(ring, 0);
1568 intel_ring_emit(ring, 0);
1569 intel_ring_emit(ring, 0);
1570 intel_ring_emit(ring, 0);
1571 intel_ring_emit(ring, 0);
1575 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6));
1576 intel_ring_emit(ring, PIPE_CONTROL_DC_FLUSH_ENABLE);
1577 intel_ring_emit(ring, 0);
1578 intel_ring_emit(ring, 0);
1579 intel_ring_emit(ring, 0);
1580 intel_ring_emit(ring, 0);
1583 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6));
1584 intel_ring_emit(ring, flags);
1585 intel_ring_emit(ring, scratch_addr);
1586 intel_ring_emit(ring, 0);
1587 intel_ring_emit(ring, 0);
1588 intel_ring_emit(ring, 0);
1591 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6));
1592 intel_ring_emit(ring, PIPE_CONTROL_CS_STALL);
1593 intel_ring_emit(ring, 0);
1594 intel_ring_emit(ring, 0);
1595 intel_ring_emit(ring, 0);
1596 intel_ring_emit(ring, 0);
1599 intel_ring_advance(ring);
1604 static void bxt_a_seqno_barrier(struct intel_engine_cs *engine)
1607 * On BXT A steppings there is a HW coherency issue whereby the
1608 * MI_STORE_DATA_IMM storing the completed request's seqno
1609 * occasionally doesn't invalidate the CPU cache. Work around this by
1610 * clflushing the corresponding cacheline whenever the caller wants
1611 * the coherency to be guaranteed. Note that this cacheline is known
1612 * to be clean at this point, since we only write it in
1613 * bxt_a_set_seqno(), where we also do a clflush after the write. So
1614 * this clflush in practice becomes an invalidate operation.
1616 intel_flush_status_page(engine, I915_GEM_HWS_INDEX);
1620 * Reserve space for 2 NOOPs at the end of each request to be
1621 * used as a workaround for not being allowed to do lite
1622 * restore with HEAD==TAIL (WaIdleLiteRestore).
1624 #define WA_TAIL_DWORDS 2
1626 static int gen8_emit_request(struct drm_i915_gem_request *request)
1628 struct intel_ring *ring = request->ring;
1631 ret = intel_ring_begin(request, 6 + WA_TAIL_DWORDS);
1635 /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */
1636 BUILD_BUG_ON(I915_GEM_HWS_INDEX_ADDR & (1 << 5));
1638 intel_ring_emit(ring, (MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW);
1639 intel_ring_emit(ring,
1640 intel_hws_seqno_address(request->engine) |
1641 MI_FLUSH_DW_USE_GTT);
1642 intel_ring_emit(ring, 0);
1643 intel_ring_emit(ring, request->fence.seqno);
1644 intel_ring_emit(ring, MI_USER_INTERRUPT);
1645 intel_ring_emit(ring, MI_NOOP);
1646 return intel_logical_ring_advance(request);
1649 static int gen8_emit_request_render(struct drm_i915_gem_request *request)
1651 struct intel_ring *ring = request->ring;
1654 ret = intel_ring_begin(request, 8 + WA_TAIL_DWORDS);
1658 /* We're using qword write, seqno should be aligned to 8 bytes. */
1659 BUILD_BUG_ON(I915_GEM_HWS_INDEX & 1);
1661 /* w/a for post sync ops following a GPGPU operation we
1662 * need a prior CS_STALL, which is emitted by the flush
1663 * following the batch.
1665 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6));
1666 intel_ring_emit(ring,
1667 (PIPE_CONTROL_GLOBAL_GTT_IVB |
1668 PIPE_CONTROL_CS_STALL |
1669 PIPE_CONTROL_QW_WRITE));
1670 intel_ring_emit(ring, intel_hws_seqno_address(request->engine));
1671 intel_ring_emit(ring, 0);
1672 intel_ring_emit(ring, i915_gem_request_get_seqno(request));
1673 /* We're thrashing one dword of HWS. */
1674 intel_ring_emit(ring, 0);
1675 intel_ring_emit(ring, MI_USER_INTERRUPT);
1676 intel_ring_emit(ring, MI_NOOP);
1677 return intel_logical_ring_advance(request);
1680 static int gen8_init_rcs_context(struct drm_i915_gem_request *req)
1684 ret = intel_logical_ring_workarounds_emit(req);
1688 ret = intel_rcs_context_init_mocs(req);
1690 * Failing to program the MOCS is non-fatal.The system will not
1691 * run at peak performance. So generate an error and carry on.
1694 DRM_ERROR("MOCS failed to program: expect performance issues.\n");
1696 return i915_gem_render_state_init(req);
1700 * intel_logical_ring_cleanup() - deallocate the Engine Command Streamer
1701 * @engine: Engine Command Streamer.
1703 void intel_logical_ring_cleanup(struct intel_engine_cs *engine)
1705 struct drm_i915_private *dev_priv;
1707 if (!intel_engine_initialized(engine))
1711 * Tasklet cannot be active at this point due intel_mark_active/idle
1712 * so this is just for documentation.
1714 if (WARN_ON(test_bit(TASKLET_STATE_SCHED, &engine->irq_tasklet.state)))
1715 tasklet_kill(&engine->irq_tasklet);
1717 dev_priv = engine->i915;
1719 if (engine->buffer) {
1720 intel_logical_ring_stop(engine);
1721 WARN_ON((I915_READ_MODE(engine) & MODE_IDLE) == 0);
1724 if (engine->cleanup)
1725 engine->cleanup(engine);
1727 intel_engine_cleanup_common(engine);
1729 if (engine->status_page.obj) {
1730 i915_gem_object_unpin_map(engine->status_page.obj);
1731 engine->status_page.obj = NULL;
1733 intel_lr_context_unpin(dev_priv->kernel_context, engine);
1735 engine->idle_lite_restore_wa = 0;
1736 engine->disable_lite_restore_wa = false;
1737 engine->ctx_desc_template = 0;
1739 lrc_destroy_wa_ctx_obj(engine);
1740 engine->i915 = NULL;
1743 void intel_execlists_enable_submission(struct drm_i915_private *dev_priv)
1745 struct intel_engine_cs *engine;
1747 for_each_engine(engine, dev_priv)
1748 engine->submit_request = execlists_submit_request;
1752 logical_ring_default_vfuncs(struct intel_engine_cs *engine)
1754 /* Default vfuncs which can be overriden by each engine. */
1755 engine->init_hw = gen8_init_common_ring;
1756 engine->emit_flush = gen8_emit_flush;
1757 engine->emit_request = gen8_emit_request;
1758 engine->submit_request = execlists_submit_request;
1760 engine->irq_enable = gen8_logical_ring_enable_irq;
1761 engine->irq_disable = gen8_logical_ring_disable_irq;
1762 engine->emit_bb_start = gen8_emit_bb_start;
1763 if (IS_BXT_REVID(engine->i915, 0, BXT_REVID_A1))
1764 engine->irq_seqno_barrier = bxt_a_seqno_barrier;
1768 logical_ring_default_irqs(struct intel_engine_cs *engine)
1770 unsigned shift = engine->irq_shift;
1771 engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT << shift;
1772 engine->irq_keep_mask = GT_CONTEXT_SWITCH_INTERRUPT << shift;
1776 lrc_setup_hws(struct intel_engine_cs *engine,
1777 struct drm_i915_gem_object *dctx_obj)
1781 /* The HWSP is part of the default context object in LRC mode. */
1782 engine->status_page.gfx_addr = i915_gem_obj_ggtt_offset(dctx_obj) +
1783 LRC_PPHWSP_PN * PAGE_SIZE;
1784 hws = i915_gem_object_pin_map(dctx_obj);
1786 return PTR_ERR(hws);
1787 engine->status_page.page_addr = hws + LRC_PPHWSP_PN * PAGE_SIZE;
1788 engine->status_page.obj = dctx_obj;
1794 logical_ring_setup(struct intel_engine_cs *engine)
1796 struct drm_i915_private *dev_priv = engine->i915;
1797 enum forcewake_domains fw_domains;
1799 intel_engine_setup_common(engine);
1801 /* Intentionally left blank. */
1802 engine->buffer = NULL;
1804 fw_domains = intel_uncore_forcewake_for_reg(dev_priv,
1808 fw_domains |= intel_uncore_forcewake_for_reg(dev_priv,
1809 RING_CONTEXT_STATUS_PTR(engine),
1810 FW_REG_READ | FW_REG_WRITE);
1812 fw_domains |= intel_uncore_forcewake_for_reg(dev_priv,
1813 RING_CONTEXT_STATUS_BUF_BASE(engine),
1816 engine->fw_domains = fw_domains;
1818 tasklet_init(&engine->irq_tasklet,
1819 intel_lrc_irq_handler, (unsigned long)engine);
1821 logical_ring_init_platform_invariants(engine);
1822 logical_ring_default_vfuncs(engine);
1823 logical_ring_default_irqs(engine);
1827 logical_ring_init(struct intel_engine_cs *engine)
1829 struct i915_gem_context *dctx = engine->i915->kernel_context;
1832 ret = intel_engine_init_common(engine);
1836 ret = execlists_context_deferred_alloc(dctx, engine);
1840 /* As this is the default context, always pin it */
1841 ret = intel_lr_context_pin(dctx, engine);
1843 DRM_ERROR("Failed to pin context for %s: %d\n",
1848 /* And setup the hardware status page. */
1849 ret = lrc_setup_hws(engine, dctx->engine[engine->id].state);
1851 DRM_ERROR("Failed to set up hws %s: %d\n", engine->name, ret);
1858 intel_logical_ring_cleanup(engine);
1862 int logical_render_ring_init(struct intel_engine_cs *engine)
1864 struct drm_i915_private *dev_priv = engine->i915;
1867 logical_ring_setup(engine);
1869 if (HAS_L3_DPF(dev_priv))
1870 engine->irq_keep_mask |= GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
1872 /* Override some for render ring. */
1873 if (INTEL_GEN(dev_priv) >= 9)
1874 engine->init_hw = gen9_init_render_ring;
1876 engine->init_hw = gen8_init_render_ring;
1877 engine->init_context = gen8_init_rcs_context;
1878 engine->cleanup = intel_fini_pipe_control;
1879 engine->emit_flush = gen8_emit_flush_render;
1880 engine->emit_request = gen8_emit_request_render;
1882 ret = intel_init_pipe_control(engine, 4096);
1886 ret = intel_init_workaround_bb(engine);
1889 * We continue even if we fail to initialize WA batch
1890 * because we only expect rare glitches but nothing
1891 * critical to prevent us from using GPU
1893 DRM_ERROR("WA batch buffer initialization failed: %d\n",
1897 ret = logical_ring_init(engine);
1899 lrc_destroy_wa_ctx_obj(engine);
1905 int logical_xcs_ring_init(struct intel_engine_cs *engine)
1907 logical_ring_setup(engine);
1909 return logical_ring_init(engine);
1913 make_rpcs(struct drm_i915_private *dev_priv)
1918 * No explicit RPCS request is needed to ensure full
1919 * slice/subslice/EU enablement prior to Gen9.
1921 if (INTEL_GEN(dev_priv) < 9)
1925 * Starting in Gen9, render power gating can leave
1926 * slice/subslice/EU in a partially enabled state. We
1927 * must make an explicit request through RPCS for full
1930 if (INTEL_INFO(dev_priv)->has_slice_pg) {
1931 rpcs |= GEN8_RPCS_S_CNT_ENABLE;
1932 rpcs |= INTEL_INFO(dev_priv)->slice_total <<
1933 GEN8_RPCS_S_CNT_SHIFT;
1934 rpcs |= GEN8_RPCS_ENABLE;
1937 if (INTEL_INFO(dev_priv)->has_subslice_pg) {
1938 rpcs |= GEN8_RPCS_SS_CNT_ENABLE;
1939 rpcs |= INTEL_INFO(dev_priv)->subslice_per_slice <<
1940 GEN8_RPCS_SS_CNT_SHIFT;
1941 rpcs |= GEN8_RPCS_ENABLE;
1944 if (INTEL_INFO(dev_priv)->has_eu_pg) {
1945 rpcs |= INTEL_INFO(dev_priv)->eu_per_subslice <<
1946 GEN8_RPCS_EU_MIN_SHIFT;
1947 rpcs |= INTEL_INFO(dev_priv)->eu_per_subslice <<
1948 GEN8_RPCS_EU_MAX_SHIFT;
1949 rpcs |= GEN8_RPCS_ENABLE;
1955 static u32 intel_lr_indirect_ctx_offset(struct intel_engine_cs *engine)
1957 u32 indirect_ctx_offset;
1959 switch (INTEL_GEN(engine->i915)) {
1961 MISSING_CASE(INTEL_GEN(engine->i915));
1964 indirect_ctx_offset =
1965 GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
1968 indirect_ctx_offset =
1969 GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
1973 return indirect_ctx_offset;
1977 populate_lr_context(struct i915_gem_context *ctx,
1978 struct drm_i915_gem_object *ctx_obj,
1979 struct intel_engine_cs *engine,
1980 struct intel_ring *ring)
1982 struct drm_i915_private *dev_priv = ctx->i915;
1983 struct i915_hw_ppgtt *ppgtt = ctx->ppgtt;
1989 ppgtt = dev_priv->mm.aliasing_ppgtt;
1991 ret = i915_gem_object_set_to_cpu_domain(ctx_obj, true);
1993 DRM_DEBUG_DRIVER("Could not set to CPU domain\n");
1997 vaddr = i915_gem_object_pin_map(ctx_obj);
1998 if (IS_ERR(vaddr)) {
1999 ret = PTR_ERR(vaddr);
2000 DRM_DEBUG_DRIVER("Could not map object pages! (%d)\n", ret);
2003 ctx_obj->dirty = true;
2005 /* The second page of the context object contains some fields which must
2006 * be set up prior to the first execution. */
2007 reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
2009 /* A context is actually a big batch buffer with several MI_LOAD_REGISTER_IMM
2010 * commands followed by (reg, value) pairs. The values we are setting here are
2011 * only for the first context restore: on a subsequent save, the GPU will
2012 * recreate this batchbuffer with new values (including all the missing
2013 * MI_LOAD_REGISTER_IMM commands that we are not initializing here). */
2014 reg_state[CTX_LRI_HEADER_0] =
2015 MI_LOAD_REGISTER_IMM(engine->id == RCS ? 14 : 11) | MI_LRI_FORCE_POSTED;
2016 ASSIGN_CTX_REG(reg_state, CTX_CONTEXT_CONTROL,
2017 RING_CONTEXT_CONTROL(engine),
2018 _MASKED_BIT_ENABLE(CTX_CTRL_INHIBIT_SYN_CTX_SWITCH |
2019 CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT |
2020 (HAS_RESOURCE_STREAMER(dev_priv) ?
2021 CTX_CTRL_RS_CTX_ENABLE : 0)));
2022 ASSIGN_CTX_REG(reg_state, CTX_RING_HEAD, RING_HEAD(engine->mmio_base),
2024 ASSIGN_CTX_REG(reg_state, CTX_RING_TAIL, RING_TAIL(engine->mmio_base),
2026 /* Ring buffer start address is not known until the buffer is pinned.
2027 * It is written to the context image in execlists_update_context()
2029 ASSIGN_CTX_REG(reg_state, CTX_RING_BUFFER_START,
2030 RING_START(engine->mmio_base), 0);
2031 ASSIGN_CTX_REG(reg_state, CTX_RING_BUFFER_CONTROL,
2032 RING_CTL(engine->mmio_base),
2033 ((ring->size - PAGE_SIZE) & RING_NR_PAGES) | RING_VALID);
2034 ASSIGN_CTX_REG(reg_state, CTX_BB_HEAD_U,
2035 RING_BBADDR_UDW(engine->mmio_base), 0);
2036 ASSIGN_CTX_REG(reg_state, CTX_BB_HEAD_L,
2037 RING_BBADDR(engine->mmio_base), 0);
2038 ASSIGN_CTX_REG(reg_state, CTX_BB_STATE,
2039 RING_BBSTATE(engine->mmio_base),
2041 ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_HEAD_U,
2042 RING_SBBADDR_UDW(engine->mmio_base), 0);
2043 ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_HEAD_L,
2044 RING_SBBADDR(engine->mmio_base), 0);
2045 ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_STATE,
2046 RING_SBBSTATE(engine->mmio_base), 0);
2047 if (engine->id == RCS) {
2048 ASSIGN_CTX_REG(reg_state, CTX_BB_PER_CTX_PTR,
2049 RING_BB_PER_CTX_PTR(engine->mmio_base), 0);
2050 ASSIGN_CTX_REG(reg_state, CTX_RCS_INDIRECT_CTX,
2051 RING_INDIRECT_CTX(engine->mmio_base), 0);
2052 ASSIGN_CTX_REG(reg_state, CTX_RCS_INDIRECT_CTX_OFFSET,
2053 RING_INDIRECT_CTX_OFFSET(engine->mmio_base), 0);
2054 if (engine->wa_ctx.obj) {
2055 struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx;
2056 uint32_t ggtt_offset = i915_gem_obj_ggtt_offset(wa_ctx->obj);
2058 reg_state[CTX_RCS_INDIRECT_CTX+1] =
2059 (ggtt_offset + wa_ctx->indirect_ctx.offset * sizeof(uint32_t)) |
2060 (wa_ctx->indirect_ctx.size / CACHELINE_DWORDS);
2062 reg_state[CTX_RCS_INDIRECT_CTX_OFFSET+1] =
2063 intel_lr_indirect_ctx_offset(engine) << 6;
2065 reg_state[CTX_BB_PER_CTX_PTR+1] =
2066 (ggtt_offset + wa_ctx->per_ctx.offset * sizeof(uint32_t)) |
2070 reg_state[CTX_LRI_HEADER_1] = MI_LOAD_REGISTER_IMM(9) | MI_LRI_FORCE_POSTED;
2071 ASSIGN_CTX_REG(reg_state, CTX_CTX_TIMESTAMP,
2072 RING_CTX_TIMESTAMP(engine->mmio_base), 0);
2073 /* PDP values well be assigned later if needed */
2074 ASSIGN_CTX_REG(reg_state, CTX_PDP3_UDW, GEN8_RING_PDP_UDW(engine, 3),
2076 ASSIGN_CTX_REG(reg_state, CTX_PDP3_LDW, GEN8_RING_PDP_LDW(engine, 3),
2078 ASSIGN_CTX_REG(reg_state, CTX_PDP2_UDW, GEN8_RING_PDP_UDW(engine, 2),
2080 ASSIGN_CTX_REG(reg_state, CTX_PDP2_LDW, GEN8_RING_PDP_LDW(engine, 2),
2082 ASSIGN_CTX_REG(reg_state, CTX_PDP1_UDW, GEN8_RING_PDP_UDW(engine, 1),
2084 ASSIGN_CTX_REG(reg_state, CTX_PDP1_LDW, GEN8_RING_PDP_LDW(engine, 1),
2086 ASSIGN_CTX_REG(reg_state, CTX_PDP0_UDW, GEN8_RING_PDP_UDW(engine, 0),
2088 ASSIGN_CTX_REG(reg_state, CTX_PDP0_LDW, GEN8_RING_PDP_LDW(engine, 0),
2091 if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
2092 /* 64b PPGTT (48bit canonical)
2093 * PDP0_DESCRIPTOR contains the base address to PML4 and
2094 * other PDP Descriptors are ignored.
2096 ASSIGN_CTX_PML4(ppgtt, reg_state);
2099 * PDP*_DESCRIPTOR contains the base address of space supported.
2100 * With dynamic page allocation, PDPs may not be allocated at
2101 * this point. Point the unallocated PDPs to the scratch page
2103 execlists_update_context_pdps(ppgtt, reg_state);
2106 if (engine->id == RCS) {
2107 reg_state[CTX_LRI_HEADER_2] = MI_LOAD_REGISTER_IMM(1);
2108 ASSIGN_CTX_REG(reg_state, CTX_R_PWR_CLK_STATE, GEN8_R_PWR_CLK_STATE,
2109 make_rpcs(dev_priv));
2112 i915_gem_object_unpin_map(ctx_obj);
2118 * intel_lr_context_size() - return the size of the context for an engine
2119 * @engine: which engine to find the context size for
2121 * Each engine may require a different amount of space for a context image,
2122 * so when allocating (or copying) an image, this function can be used to
2123 * find the right size for the specific engine.
2125 * Return: size (in bytes) of an engine-specific context image
2127 * Note: this size includes the HWSP, which is part of the context image
2128 * in LRC mode, but does not include the "shared data page" used with
2129 * GuC submission. The caller should account for this if using the GuC.
2131 uint32_t intel_lr_context_size(struct intel_engine_cs *engine)
2135 WARN_ON(INTEL_GEN(engine->i915) < 8);
2137 switch (engine->id) {
2139 if (INTEL_GEN(engine->i915) >= 9)
2140 ret = GEN9_LR_CONTEXT_RENDER_SIZE;
2142 ret = GEN8_LR_CONTEXT_RENDER_SIZE;
2148 ret = GEN8_LR_CONTEXT_OTHER_SIZE;
2155 static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
2156 struct intel_engine_cs *engine)
2158 struct drm_i915_gem_object *ctx_obj;
2159 struct intel_context *ce = &ctx->engine[engine->id];
2160 uint32_t context_size;
2161 struct intel_ring *ring;
2166 context_size = round_up(intel_lr_context_size(engine), 4096);
2168 /* One extra page as the sharing data between driver and GuC */
2169 context_size += PAGE_SIZE * LRC_PPHWSP_PN;
2171 ctx_obj = i915_gem_object_create(&ctx->i915->drm, context_size);
2172 if (IS_ERR(ctx_obj)) {
2173 DRM_DEBUG_DRIVER("Alloc LRC backing obj failed.\n");
2174 return PTR_ERR(ctx_obj);
2177 ring = intel_engine_create_ring(engine, ctx->ring_size);
2179 ret = PTR_ERR(ring);
2180 goto error_deref_obj;
2183 ret = populate_lr_context(ctx, ctx_obj, engine, ring);
2185 DRM_DEBUG_DRIVER("Failed to populate LRC: %d\n", ret);
2186 goto error_ring_free;
2190 ce->state = ctx_obj;
2191 ce->initialised = engine->init_context == NULL;
2196 intel_ring_free(ring);
2198 i915_gem_object_put(ctx_obj);
2204 void intel_lr_context_reset(struct drm_i915_private *dev_priv,
2205 struct i915_gem_context *ctx)
2207 struct intel_engine_cs *engine;
2209 for_each_engine(engine, dev_priv) {
2210 struct intel_context *ce = &ctx->engine[engine->id];
2211 struct drm_i915_gem_object *ctx_obj = ce->state;
2213 uint32_t *reg_state;
2218 vaddr = i915_gem_object_pin_map(ctx_obj);
2219 if (WARN_ON(IS_ERR(vaddr)))
2222 reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
2223 ctx_obj->dirty = true;
2225 reg_state[CTX_RING_HEAD+1] = 0;
2226 reg_state[CTX_RING_TAIL+1] = 0;
2228 i915_gem_object_unpin_map(ctx_obj);