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 GEN8_CTX_STATUS_COMPLETED_MASK \
160 (GEN8_CTX_STATUS_ACTIVE_IDLE | \
161 GEN8_CTX_STATUS_PREEMPTED | \
162 GEN8_CTX_STATUS_ELEMENT_SWITCH)
164 #define CTX_LRI_HEADER_0 0x01
165 #define CTX_CONTEXT_CONTROL 0x02
166 #define CTX_RING_HEAD 0x04
167 #define CTX_RING_TAIL 0x06
168 #define CTX_RING_BUFFER_START 0x08
169 #define CTX_RING_BUFFER_CONTROL 0x0a
170 #define CTX_BB_HEAD_U 0x0c
171 #define CTX_BB_HEAD_L 0x0e
172 #define CTX_BB_STATE 0x10
173 #define CTX_SECOND_BB_HEAD_U 0x12
174 #define CTX_SECOND_BB_HEAD_L 0x14
175 #define CTX_SECOND_BB_STATE 0x16
176 #define CTX_BB_PER_CTX_PTR 0x18
177 #define CTX_RCS_INDIRECT_CTX 0x1a
178 #define CTX_RCS_INDIRECT_CTX_OFFSET 0x1c
179 #define CTX_LRI_HEADER_1 0x21
180 #define CTX_CTX_TIMESTAMP 0x22
181 #define CTX_PDP3_UDW 0x24
182 #define CTX_PDP3_LDW 0x26
183 #define CTX_PDP2_UDW 0x28
184 #define CTX_PDP2_LDW 0x2a
185 #define CTX_PDP1_UDW 0x2c
186 #define CTX_PDP1_LDW 0x2e
187 #define CTX_PDP0_UDW 0x30
188 #define CTX_PDP0_LDW 0x32
189 #define CTX_LRI_HEADER_2 0x41
190 #define CTX_R_PWR_CLK_STATE 0x42
191 #define CTX_GPGPU_CSR_BASE_ADDRESS 0x44
193 #define GEN8_CTX_VALID (1<<0)
194 #define GEN8_CTX_FORCE_PD_RESTORE (1<<1)
195 #define GEN8_CTX_FORCE_RESTORE (1<<2)
196 #define GEN8_CTX_L3LLC_COHERENT (1<<5)
197 #define GEN8_CTX_PRIVILEGE (1<<8)
199 #define ASSIGN_CTX_REG(reg_state, pos, reg, val) do { \
200 (reg_state)[(pos)+0] = i915_mmio_reg_offset(reg); \
201 (reg_state)[(pos)+1] = (val); \
204 #define ASSIGN_CTX_PDP(ppgtt, reg_state, n) do { \
205 const u64 _addr = i915_page_dir_dma_addr((ppgtt), (n)); \
206 reg_state[CTX_PDP ## n ## _UDW+1] = upper_32_bits(_addr); \
207 reg_state[CTX_PDP ## n ## _LDW+1] = lower_32_bits(_addr); \
210 #define ASSIGN_CTX_PML4(ppgtt, reg_state) do { \
211 reg_state[CTX_PDP0_UDW + 1] = upper_32_bits(px_dma(&ppgtt->pml4)); \
212 reg_state[CTX_PDP0_LDW + 1] = lower_32_bits(px_dma(&ppgtt->pml4)); \
217 FAULT_AND_HALT, /* Debug only */
219 FAULT_AND_CONTINUE /* Unsupported */
221 #define GEN8_CTX_ID_SHIFT 32
222 #define GEN8_CTX_ID_WIDTH 21
223 #define GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT 0x17
224 #define GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT 0x26
226 /* Typical size of the average request (2 pipecontrols and a MI_BB) */
227 #define EXECLISTS_REQUEST_SIZE 64 /* bytes */
229 #define WA_TAIL_DWORDS 2
231 static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
232 struct intel_engine_cs *engine);
233 static int intel_lr_context_pin(struct i915_gem_context *ctx,
234 struct intel_engine_cs *engine);
235 static void execlists_init_reg_state(u32 *reg_state,
236 struct i915_gem_context *ctx,
237 struct intel_engine_cs *engine,
238 struct intel_ring *ring);
241 * intel_sanitize_enable_execlists() - sanitize i915.enable_execlists
242 * @dev_priv: i915 device private
243 * @enable_execlists: value of i915.enable_execlists module parameter.
245 * Only certain platforms support Execlists (the prerequisites being
246 * support for Logical Ring Contexts and Aliasing PPGTT or better).
248 * Return: 1 if Execlists is supported and has to be enabled.
250 int intel_sanitize_enable_execlists(struct drm_i915_private *dev_priv, int enable_execlists)
252 /* On platforms with execlist available, vGPU will only
253 * support execlist mode, no ring buffer mode.
255 if (HAS_LOGICAL_RING_CONTEXTS(dev_priv) && intel_vgpu_active(dev_priv))
258 if (INTEL_GEN(dev_priv) >= 9)
261 if (enable_execlists == 0)
264 if (HAS_LOGICAL_RING_CONTEXTS(dev_priv) &&
265 USES_PPGTT(dev_priv) &&
266 i915.use_mmio_flip >= 0)
273 logical_ring_init_platform_invariants(struct intel_engine_cs *engine)
275 struct drm_i915_private *dev_priv = engine->i915;
277 engine->disable_lite_restore_wa =
278 (IS_SKL_REVID(dev_priv, 0, SKL_REVID_B0) ||
279 IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1)) &&
280 (engine->id == VCS || engine->id == VCS2);
282 engine->ctx_desc_template = GEN8_CTX_VALID;
283 if (IS_GEN8(dev_priv))
284 engine->ctx_desc_template |= GEN8_CTX_L3LLC_COHERENT;
285 engine->ctx_desc_template |= GEN8_CTX_PRIVILEGE;
287 /* TODO: WaDisableLiteRestore when we start using semaphore
288 * signalling between Command Streamers */
289 /* ring->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE; */
291 /* WaEnableForceRestoreInCtxtDescForVCS:skl */
292 /* WaEnableForceRestoreInCtxtDescForVCS:bxt */
293 if (engine->disable_lite_restore_wa)
294 engine->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE;
298 * intel_lr_context_descriptor_update() - calculate & cache the descriptor
299 * descriptor for a pinned context
300 * @ctx: Context to work on
301 * @engine: Engine the descriptor will be used with
303 * The context descriptor encodes various attributes of a context,
304 * including its GTT address and some flags. Because it's fairly
305 * expensive to calculate, we'll just do it once and cache the result,
306 * which remains valid until the context is unpinned.
308 * This is what a descriptor looks like, from LSB to MSB::
310 * bits 0-11: flags, GEN8_CTX_* (cached in ctx_desc_template)
311 * bits 12-31: LRCA, GTT address of (the HWSP of) this context
312 * bits 32-52: ctx ID, a globally unique tag
313 * bits 53-54: mbz, reserved for use by hardware
314 * bits 55-63: group ID, currently unused and set to 0
317 intel_lr_context_descriptor_update(struct i915_gem_context *ctx,
318 struct intel_engine_cs *engine)
320 struct intel_context *ce = &ctx->engine[engine->id];
323 BUILD_BUG_ON(MAX_CONTEXT_HW_ID > (1<<GEN8_CTX_ID_WIDTH));
325 desc = ctx->desc_template; /* bits 3-4 */
326 desc |= engine->ctx_desc_template; /* bits 0-11 */
327 desc |= i915_ggtt_offset(ce->state) + LRC_PPHWSP_PN * PAGE_SIZE;
329 desc |= (u64)ctx->hw_id << GEN8_CTX_ID_SHIFT; /* bits 32-52 */
334 uint64_t intel_lr_context_descriptor(struct i915_gem_context *ctx,
335 struct intel_engine_cs *engine)
337 return ctx->engine[engine->id].lrc_desc;
341 execlists_context_status_change(struct drm_i915_gem_request *rq,
342 unsigned long status)
345 * Only used when GVT-g is enabled now. When GVT-g is disabled,
346 * The compiler should eliminate this function as dead-code.
348 if (!IS_ENABLED(CONFIG_DRM_I915_GVT))
351 atomic_notifier_call_chain(&rq->ctx->status_notifier, status, rq);
355 execlists_update_context_pdps(struct i915_hw_ppgtt *ppgtt, u32 *reg_state)
357 ASSIGN_CTX_PDP(ppgtt, reg_state, 3);
358 ASSIGN_CTX_PDP(ppgtt, reg_state, 2);
359 ASSIGN_CTX_PDP(ppgtt, reg_state, 1);
360 ASSIGN_CTX_PDP(ppgtt, reg_state, 0);
363 static u64 execlists_update_context(struct drm_i915_gem_request *rq)
365 struct intel_context *ce = &rq->ctx->engine[rq->engine->id];
366 struct i915_hw_ppgtt *ppgtt = rq->ctx->ppgtt;
367 u32 *reg_state = ce->lrc_reg_state;
369 reg_state[CTX_RING_TAIL+1] = intel_ring_offset(rq->ring, rq->tail);
371 /* True 32b PPGTT with dynamic page allocation: update PDP
372 * registers and point the unallocated PDPs to scratch page.
373 * PML4 is allocated during ppgtt init, so this is not needed
376 if (ppgtt && !USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
377 execlists_update_context_pdps(ppgtt, reg_state);
382 static void execlists_submit_ports(struct intel_engine_cs *engine)
384 struct drm_i915_private *dev_priv = engine->i915;
385 struct execlist_port *port = engine->execlist_port;
387 dev_priv->regs + i915_mmio_reg_offset(RING_ELSP(engine));
391 execlists_context_status_change(port[0].request,
392 INTEL_CONTEXT_SCHEDULE_IN);
393 desc[0] = execlists_update_context(port[0].request);
394 engine->preempt_wa = port[0].count++; /* bdw only? fixed on skl? */
396 if (port[1].request) {
397 GEM_BUG_ON(port[1].count);
398 execlists_context_status_change(port[1].request,
399 INTEL_CONTEXT_SCHEDULE_IN);
400 desc[1] = execlists_update_context(port[1].request);
405 GEM_BUG_ON(desc[0] == desc[1]);
407 /* You must always write both descriptors in the order below. */
408 writel(upper_32_bits(desc[1]), elsp);
409 writel(lower_32_bits(desc[1]), elsp);
411 writel(upper_32_bits(desc[0]), elsp);
412 /* The context is automatically loaded after the following */
413 writel(lower_32_bits(desc[0]), elsp);
416 static bool ctx_single_port_submission(const struct i915_gem_context *ctx)
418 return (IS_ENABLED(CONFIG_DRM_I915_GVT) &&
419 ctx->execlists_force_single_submission);
422 static bool can_merge_ctx(const struct i915_gem_context *prev,
423 const struct i915_gem_context *next)
428 if (ctx_single_port_submission(prev))
434 static void execlists_dequeue(struct intel_engine_cs *engine)
436 struct drm_i915_gem_request *cursor, *last;
437 struct execlist_port *port = engine->execlist_port;
440 last = port->request;
442 /* WaIdleLiteRestore:bdw,skl
443 * Apply the wa NOOPs to prevent ring:HEAD == req:TAIL
444 * as we resubmit the request. See gen8_emit_request()
445 * for where we prepare the padding after the end of the
448 last->tail = last->wa_tail;
450 GEM_BUG_ON(port[1].request);
452 /* Hardware submission is through 2 ports. Conceptually each port
453 * has a (RING_START, RING_HEAD, RING_TAIL) tuple. RING_START is
454 * static for a context, and unique to each, so we only execute
455 * requests belonging to a single context from each ring. RING_HEAD
456 * is maintained by the CS in the context image, it marks the place
457 * where it got up to last time, and through RING_TAIL we tell the CS
458 * where we want to execute up to this time.
460 * In this list the requests are in order of execution. Consecutive
461 * requests from the same context are adjacent in the ringbuffer. We
462 * can combine these requests into a single RING_TAIL update:
464 * RING_HEAD...req1...req2
466 * since to execute req2 the CS must first execute req1.
468 * Our goal then is to point each port to the end of a consecutive
469 * sequence of requests as being the most optimal (fewest wake ups
470 * and context switches) submission.
473 spin_lock(&engine->execlist_lock);
474 list_for_each_entry(cursor, &engine->execlist_queue, execlist_link) {
475 /* Can we combine this request with the current port? It has to
476 * be the same context/ringbuffer and not have any exceptions
477 * (e.g. GVT saying never to combine contexts).
479 * If we can combine the requests, we can execute both by
480 * updating the RING_TAIL to point to the end of the second
481 * request, and so we never need to tell the hardware about
484 if (last && !can_merge_ctx(cursor->ctx, last->ctx)) {
485 /* If we are on the second port and cannot combine
486 * this request with the last, then we are done.
488 if (port != engine->execlist_port)
491 /* If GVT overrides us we only ever submit port[0],
492 * leaving port[1] empty. Note that we also have
493 * to be careful that we don't queue the same
494 * context (even though a different request) to
497 if (ctx_single_port_submission(cursor->ctx))
500 GEM_BUG_ON(last->ctx == cursor->ctx);
502 i915_gem_request_assign(&port->request, last);
509 /* Decouple all the requests submitted from the queue */
510 engine->execlist_queue.next = &cursor->execlist_link;
511 cursor->execlist_link.prev = &engine->execlist_queue;
513 i915_gem_request_assign(&port->request, last);
515 spin_unlock(&engine->execlist_lock);
518 execlists_submit_ports(engine);
521 static bool execlists_elsp_idle(struct intel_engine_cs *engine)
523 return !engine->execlist_port[0].request;
526 static bool execlists_elsp_ready(struct intel_engine_cs *engine)
530 port = 1; /* wait for a free slot */
531 if (engine->disable_lite_restore_wa || engine->preempt_wa)
532 port = 0; /* wait for GPU to be idle before continuing */
534 return !engine->execlist_port[port].request;
538 * Check the unread Context Status Buffers and manage the submission of new
539 * contexts to the ELSP accordingly.
541 static void intel_lrc_irq_handler(unsigned long data)
543 struct intel_engine_cs *engine = (struct intel_engine_cs *)data;
544 struct execlist_port *port = engine->execlist_port;
545 struct drm_i915_private *dev_priv = engine->i915;
547 intel_uncore_forcewake_get(dev_priv, engine->fw_domains);
549 if (!execlists_elsp_idle(engine)) {
550 u32 __iomem *csb_mmio =
551 dev_priv->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_PTR(engine));
553 dev_priv->regs + i915_mmio_reg_offset(RING_CONTEXT_STATUS_BUF_LO(engine, 0));
554 unsigned int csb, head, tail;
556 csb = readl(csb_mmio);
557 head = GEN8_CSB_READ_PTR(csb);
558 tail = GEN8_CSB_WRITE_PTR(csb);
560 tail += GEN8_CSB_ENTRIES;
561 while (head < tail) {
562 unsigned int idx = ++head % GEN8_CSB_ENTRIES;
563 unsigned int status = readl(buf + 2 * idx);
565 if (!(status & GEN8_CTX_STATUS_COMPLETED_MASK))
568 GEM_BUG_ON(port[0].count == 0);
569 if (--port[0].count == 0) {
570 GEM_BUG_ON(status & GEN8_CTX_STATUS_PREEMPTED);
571 execlists_context_status_change(port[0].request,
572 INTEL_CONTEXT_SCHEDULE_OUT);
574 i915_gem_request_put(port[0].request);
576 memset(&port[1], 0, sizeof(port[1]));
578 engine->preempt_wa = false;
581 GEM_BUG_ON(port[0].count == 0 &&
582 !(status & GEN8_CTX_STATUS_ACTIVE_IDLE));
585 writel(_MASKED_FIELD(GEN8_CSB_READ_PTR_MASK,
586 GEN8_CSB_WRITE_PTR(csb) << 8),
590 if (execlists_elsp_ready(engine))
591 execlists_dequeue(engine);
593 intel_uncore_forcewake_put(dev_priv, engine->fw_domains);
596 static void execlists_submit_request(struct drm_i915_gem_request *request)
598 struct intel_engine_cs *engine = request->engine;
601 spin_lock_irqsave(&engine->execlist_lock, flags);
603 list_add_tail(&request->execlist_link, &engine->execlist_queue);
604 if (execlists_elsp_idle(engine))
605 tasklet_hi_schedule(&engine->irq_tasklet);
607 spin_unlock_irqrestore(&engine->execlist_lock, flags);
610 int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *request)
612 struct intel_engine_cs *engine = request->engine;
613 struct intel_context *ce = &request->ctx->engine[engine->id];
616 /* Flush enough space to reduce the likelihood of waiting after
617 * we start building the request - in which case we will just
618 * have to repeat work.
620 request->reserved_space += EXECLISTS_REQUEST_SIZE;
623 ret = execlists_context_deferred_alloc(request->ctx, engine);
628 request->ring = ce->ring;
630 if (i915.enable_guc_submission) {
632 * Check that the GuC has space for the request before
633 * going any further, as the i915_add_request() call
634 * later on mustn't fail ...
636 ret = i915_guc_wq_reserve(request);
641 ret = intel_lr_context_pin(request->ctx, engine);
645 ret = intel_ring_begin(request, 0);
649 if (!ce->initialised) {
650 ret = engine->init_context(request);
654 ce->initialised = true;
657 /* Note that after this point, we have committed to using
658 * this request as it is being used to both track the
659 * state of engine initialisation and liveness of the
660 * golden renderstate above. Think twice before you try
661 * to cancel/unwind this request now.
664 request->reserved_space -= EXECLISTS_REQUEST_SIZE;
668 intel_lr_context_unpin(request->ctx, engine);
673 * intel_logical_ring_advance() - advance the tail and prepare for submission
674 * @request: Request to advance the logical ringbuffer of.
676 * The tail is updated in our logical ringbuffer struct, not in the actual context. What
677 * really happens during submission is that the context and current tail will be placed
678 * on a queue waiting for the ELSP to be ready to accept a new context submission. At that
679 * point, the tail *inside* the context is updated and the ELSP written to.
682 intel_logical_ring_advance(struct drm_i915_gem_request *request)
684 struct intel_ring *ring = request->ring;
685 struct intel_engine_cs *engine = request->engine;
687 intel_ring_advance(ring);
688 request->tail = ring->tail;
691 * Here we add two extra NOOPs as padding to avoid
692 * lite restore of a context with HEAD==TAIL.
694 * Caller must reserve WA_TAIL_DWORDS for us!
696 intel_ring_emit(ring, MI_NOOP);
697 intel_ring_emit(ring, MI_NOOP);
698 intel_ring_advance(ring);
699 request->wa_tail = ring->tail;
701 /* We keep the previous context alive until we retire the following
702 * request. This ensures that any the context object is still pinned
703 * for any residual writes the HW makes into it on the context switch
704 * into the next object following the breadcrumb. Otherwise, we may
705 * retire the context too early.
707 request->previous_context = engine->last_context;
708 engine->last_context = request->ctx;
712 static int intel_lr_context_pin(struct i915_gem_context *ctx,
713 struct intel_engine_cs *engine)
715 struct intel_context *ce = &ctx->engine[engine->id];
719 lockdep_assert_held(&ctx->i915->drm.struct_mutex);
724 ret = i915_vma_pin(ce->state, 0, GEN8_LR_CONTEXT_ALIGN,
725 PIN_OFFSET_BIAS | GUC_WOPCM_TOP | PIN_GLOBAL);
729 vaddr = i915_gem_object_pin_map(ce->state->obj, I915_MAP_WB);
731 ret = PTR_ERR(vaddr);
735 ret = intel_ring_pin(ce->ring);
739 intel_lr_context_descriptor_update(ctx, engine);
741 ce->lrc_reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
742 ce->lrc_reg_state[CTX_RING_BUFFER_START+1] =
743 i915_ggtt_offset(ce->ring->vma);
745 ce->state->obj->dirty = true;
747 /* Invalidate GuC TLB. */
748 if (i915.enable_guc_submission) {
749 struct drm_i915_private *dev_priv = ctx->i915;
750 I915_WRITE(GEN8_GTCR, GEN8_GTCR_INVALIDATE);
753 i915_gem_context_get(ctx);
757 i915_gem_object_unpin_map(ce->state->obj);
759 __i915_vma_unpin(ce->state);
765 void intel_lr_context_unpin(struct i915_gem_context *ctx,
766 struct intel_engine_cs *engine)
768 struct intel_context *ce = &ctx->engine[engine->id];
770 lockdep_assert_held(&ctx->i915->drm.struct_mutex);
771 GEM_BUG_ON(ce->pin_count == 0);
776 intel_ring_unpin(ce->ring);
778 i915_gem_object_unpin_map(ce->state->obj);
779 i915_vma_unpin(ce->state);
781 i915_gem_context_put(ctx);
784 static int intel_logical_ring_workarounds_emit(struct drm_i915_gem_request *req)
787 struct intel_ring *ring = req->ring;
788 struct i915_workarounds *w = &req->i915->workarounds;
793 ret = req->engine->emit_flush(req, EMIT_BARRIER);
797 ret = intel_ring_begin(req, w->count * 2 + 2);
801 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(w->count));
802 for (i = 0; i < w->count; i++) {
803 intel_ring_emit_reg(ring, w->reg[i].addr);
804 intel_ring_emit(ring, w->reg[i].value);
806 intel_ring_emit(ring, MI_NOOP);
808 intel_ring_advance(ring);
810 ret = req->engine->emit_flush(req, EMIT_BARRIER);
817 #define wa_ctx_emit(batch, index, cmd) \
819 int __index = (index)++; \
820 if (WARN_ON(__index >= (PAGE_SIZE / sizeof(uint32_t)))) { \
823 batch[__index] = (cmd); \
826 #define wa_ctx_emit_reg(batch, index, reg) \
827 wa_ctx_emit((batch), (index), i915_mmio_reg_offset(reg))
830 * In this WA we need to set GEN8_L3SQCREG4[21:21] and reset it after
831 * PIPE_CONTROL instruction. This is required for the flush to happen correctly
832 * but there is a slight complication as this is applied in WA batch where the
833 * values are only initialized once so we cannot take register value at the
834 * beginning and reuse it further; hence we save its value to memory, upload a
835 * constant value with bit21 set and then we restore it back with the saved value.
836 * To simplify the WA, a constant value is formed by using the default value
837 * of this register. This shouldn't be a problem because we are only modifying
838 * it for a short period and this batch in non-premptible. We can ofcourse
839 * use additional instructions that read the actual value of the register
840 * at that time and set our bit of interest but it makes the WA complicated.
842 * This WA is also required for Gen9 so extracting as a function avoids
845 static inline int gen8_emit_flush_coherentl3_wa(struct intel_engine_cs *engine,
849 struct drm_i915_private *dev_priv = engine->i915;
850 uint32_t l3sqc4_flush = (0x40400000 | GEN8_LQSC_FLUSH_COHERENT_LINES);
853 * WaDisableLSQCROPERFforOCL:skl,kbl
854 * This WA is implemented in skl_init_clock_gating() but since
855 * this batch updates GEN8_L3SQCREG4 with default value we need to
856 * set this bit here to retain the WA during flush.
858 if (IS_SKL_REVID(dev_priv, 0, SKL_REVID_E0) ||
859 IS_KBL_REVID(dev_priv, 0, KBL_REVID_E0))
860 l3sqc4_flush |= GEN8_LQSC_RO_PERF_DIS;
862 wa_ctx_emit(batch, index, (MI_STORE_REGISTER_MEM_GEN8 |
863 MI_SRM_LRM_GLOBAL_GTT));
864 wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
865 wa_ctx_emit(batch, index, i915_ggtt_offset(engine->scratch) + 256);
866 wa_ctx_emit(batch, index, 0);
868 wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1));
869 wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
870 wa_ctx_emit(batch, index, l3sqc4_flush);
872 wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
873 wa_ctx_emit(batch, index, (PIPE_CONTROL_CS_STALL |
874 PIPE_CONTROL_DC_FLUSH_ENABLE));
875 wa_ctx_emit(batch, index, 0);
876 wa_ctx_emit(batch, index, 0);
877 wa_ctx_emit(batch, index, 0);
878 wa_ctx_emit(batch, index, 0);
880 wa_ctx_emit(batch, index, (MI_LOAD_REGISTER_MEM_GEN8 |
881 MI_SRM_LRM_GLOBAL_GTT));
882 wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
883 wa_ctx_emit(batch, index, i915_ggtt_offset(engine->scratch) + 256);
884 wa_ctx_emit(batch, index, 0);
889 static inline uint32_t wa_ctx_start(struct i915_wa_ctx_bb *wa_ctx,
891 uint32_t start_alignment)
893 return wa_ctx->offset = ALIGN(offset, start_alignment);
896 static inline int wa_ctx_end(struct i915_wa_ctx_bb *wa_ctx,
898 uint32_t size_alignment)
900 wa_ctx->size = offset - wa_ctx->offset;
902 WARN(wa_ctx->size % size_alignment,
903 "wa_ctx_bb failed sanity checks: size %d is not aligned to %d\n",
904 wa_ctx->size, size_alignment);
909 * Typically we only have one indirect_ctx and per_ctx batch buffer which are
910 * initialized at the beginning and shared across all contexts but this field
911 * helps us to have multiple batches at different offsets and select them based
912 * on a criteria. At the moment this batch always start at the beginning of the page
913 * and at this point we don't have multiple wa_ctx batch buffers.
915 * The number of WA applied are not known at the beginning; we use this field
916 * to return the no of DWORDS written.
918 * It is to be noted that this batch does not contain MI_BATCH_BUFFER_END
919 * so it adds NOOPs as padding to make it cacheline aligned.
920 * MI_BATCH_BUFFER_END will be added to perctx batch and both of them together
921 * makes a complete batch buffer.
923 static int gen8_init_indirectctx_bb(struct intel_engine_cs *engine,
924 struct i915_wa_ctx_bb *wa_ctx,
928 uint32_t scratch_addr;
929 uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
931 /* WaDisableCtxRestoreArbitration:bdw,chv */
932 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_DISABLE);
934 /* WaFlushCoherentL3CacheLinesAtContextSwitch:bdw */
935 if (IS_BROADWELL(engine->i915)) {
936 int rc = gen8_emit_flush_coherentl3_wa(engine, batch, index);
942 /* WaClearSlmSpaceAtContextSwitch:bdw,chv */
943 /* Actual scratch location is at 128 bytes offset */
944 scratch_addr = i915_ggtt_offset(engine->scratch) + 2 * CACHELINE_BYTES;
946 wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
947 wa_ctx_emit(batch, index, (PIPE_CONTROL_FLUSH_L3 |
948 PIPE_CONTROL_GLOBAL_GTT_IVB |
949 PIPE_CONTROL_CS_STALL |
950 PIPE_CONTROL_QW_WRITE));
951 wa_ctx_emit(batch, index, scratch_addr);
952 wa_ctx_emit(batch, index, 0);
953 wa_ctx_emit(batch, index, 0);
954 wa_ctx_emit(batch, index, 0);
956 /* Pad to end of cacheline */
957 while (index % CACHELINE_DWORDS)
958 wa_ctx_emit(batch, index, MI_NOOP);
961 * MI_BATCH_BUFFER_END is not required in Indirect ctx BB because
962 * execution depends on the length specified in terms of cache lines
963 * in the register CTX_RCS_INDIRECT_CTX
966 return wa_ctx_end(wa_ctx, *offset = index, CACHELINE_DWORDS);
970 * This batch is started immediately after indirect_ctx batch. Since we ensure
971 * that indirect_ctx ends on a cacheline this batch is aligned automatically.
973 * The number of DWORDS written are returned using this field.
975 * This batch is terminated with MI_BATCH_BUFFER_END and so we need not add padding
976 * to align it with cacheline as padding after MI_BATCH_BUFFER_END is redundant.
978 static int gen8_init_perctx_bb(struct intel_engine_cs *engine,
979 struct i915_wa_ctx_bb *wa_ctx,
983 uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
985 /* WaDisableCtxRestoreArbitration:bdw,chv */
986 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_ENABLE);
988 wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END);
990 return wa_ctx_end(wa_ctx, *offset = index, 1);
993 static int gen9_init_indirectctx_bb(struct intel_engine_cs *engine,
994 struct i915_wa_ctx_bb *wa_ctx,
999 struct drm_i915_private *dev_priv = engine->i915;
1000 uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1002 /* WaDisableCtxRestoreArbitration:skl,bxt */
1003 if (IS_SKL_REVID(dev_priv, 0, SKL_REVID_D0) ||
1004 IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1))
1005 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_DISABLE);
1007 /* WaFlushCoherentL3CacheLinesAtContextSwitch:skl,bxt */
1008 ret = gen8_emit_flush_coherentl3_wa(engine, batch, index);
1013 /* WaDisableGatherAtSetShaderCommonSlice:skl,bxt,kbl */
1014 wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1));
1015 wa_ctx_emit_reg(batch, index, COMMON_SLICE_CHICKEN2);
1016 wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(
1017 GEN9_DISABLE_GATHER_AT_SET_SHADER_COMMON_SLICE));
1018 wa_ctx_emit(batch, index, MI_NOOP);
1020 /* WaClearSlmSpaceAtContextSwitch:kbl */
1021 /* Actual scratch location is at 128 bytes offset */
1022 if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_A0)) {
1024 i915_ggtt_offset(engine->scratch) + 2 * CACHELINE_BYTES;
1026 wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
1027 wa_ctx_emit(batch, index, (PIPE_CONTROL_FLUSH_L3 |
1028 PIPE_CONTROL_GLOBAL_GTT_IVB |
1029 PIPE_CONTROL_CS_STALL |
1030 PIPE_CONTROL_QW_WRITE));
1031 wa_ctx_emit(batch, index, scratch_addr);
1032 wa_ctx_emit(batch, index, 0);
1033 wa_ctx_emit(batch, index, 0);
1034 wa_ctx_emit(batch, index, 0);
1037 /* WaMediaPoolStateCmdInWABB:bxt */
1038 if (HAS_POOLED_EU(engine->i915)) {
1040 * EU pool configuration is setup along with golden context
1041 * during context initialization. This value depends on
1042 * device type (2x6 or 3x6) and needs to be updated based
1043 * on which subslice is disabled especially for 2x6
1044 * devices, however it is safe to load default
1045 * configuration of 3x6 device instead of masking off
1046 * corresponding bits because HW ignores bits of a disabled
1047 * subslice and drops down to appropriate config. Please
1048 * see render_state_setup() in i915_gem_render_state.c for
1049 * possible configurations, to avoid duplication they are
1050 * not shown here again.
1052 u32 eu_pool_config = 0x00777000;
1053 wa_ctx_emit(batch, index, GEN9_MEDIA_POOL_STATE);
1054 wa_ctx_emit(batch, index, GEN9_MEDIA_POOL_ENABLE);
1055 wa_ctx_emit(batch, index, eu_pool_config);
1056 wa_ctx_emit(batch, index, 0);
1057 wa_ctx_emit(batch, index, 0);
1058 wa_ctx_emit(batch, index, 0);
1061 /* Pad to end of cacheline */
1062 while (index % CACHELINE_DWORDS)
1063 wa_ctx_emit(batch, index, MI_NOOP);
1065 return wa_ctx_end(wa_ctx, *offset = index, CACHELINE_DWORDS);
1068 static int gen9_init_perctx_bb(struct intel_engine_cs *engine,
1069 struct i915_wa_ctx_bb *wa_ctx,
1073 uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1075 /* WaSetDisablePixMaskCammingAndRhwoInCommonSliceChicken:skl,bxt */
1076 if (IS_SKL_REVID(engine->i915, 0, SKL_REVID_B0) ||
1077 IS_BXT_REVID(engine->i915, 0, BXT_REVID_A1)) {
1078 wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1));
1079 wa_ctx_emit_reg(batch, index, GEN9_SLICE_COMMON_ECO_CHICKEN0);
1080 wa_ctx_emit(batch, index,
1081 _MASKED_BIT_ENABLE(DISABLE_PIXEL_MASK_CAMMING));
1082 wa_ctx_emit(batch, index, MI_NOOP);
1085 /* WaClearTdlStateAckDirtyBits:bxt */
1086 if (IS_BXT_REVID(engine->i915, 0, BXT_REVID_B0)) {
1087 wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(4));
1089 wa_ctx_emit_reg(batch, index, GEN8_STATE_ACK);
1090 wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS));
1092 wa_ctx_emit_reg(batch, index, GEN9_STATE_ACK_SLICE1);
1093 wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS));
1095 wa_ctx_emit_reg(batch, index, GEN9_STATE_ACK_SLICE2);
1096 wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS));
1098 wa_ctx_emit_reg(batch, index, GEN7_ROW_CHICKEN2);
1099 /* dummy write to CS, mask bits are 0 to ensure the register is not modified */
1100 wa_ctx_emit(batch, index, 0x0);
1101 wa_ctx_emit(batch, index, MI_NOOP);
1104 /* WaDisableCtxRestoreArbitration:skl,bxt */
1105 if (IS_SKL_REVID(engine->i915, 0, SKL_REVID_D0) ||
1106 IS_BXT_REVID(engine->i915, 0, BXT_REVID_A1))
1107 wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_ENABLE);
1109 wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END);
1111 return wa_ctx_end(wa_ctx, *offset = index, 1);
1114 static int lrc_setup_wa_ctx_obj(struct intel_engine_cs *engine, u32 size)
1116 struct drm_i915_gem_object *obj;
1117 struct i915_vma *vma;
1120 obj = i915_gem_object_create(&engine->i915->drm, PAGE_ALIGN(size));
1122 return PTR_ERR(obj);
1124 vma = i915_vma_create(obj, &engine->i915->ggtt.base, NULL);
1130 err = i915_vma_pin(vma, 0, PAGE_SIZE, PIN_GLOBAL | PIN_HIGH);
1134 engine->wa_ctx.vma = vma;
1138 i915_gem_object_put(obj);
1142 static void lrc_destroy_wa_ctx_obj(struct intel_engine_cs *engine)
1144 i915_vma_unpin_and_release(&engine->wa_ctx.vma);
1147 static int intel_init_workaround_bb(struct intel_engine_cs *engine)
1149 struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx;
1155 WARN_ON(engine->id != RCS);
1157 /* update this when WA for higher Gen are added */
1158 if (INTEL_GEN(engine->i915) > 9) {
1159 DRM_ERROR("WA batch buffer is not initialized for Gen%d\n",
1160 INTEL_GEN(engine->i915));
1164 /* some WA perform writes to scratch page, ensure it is valid */
1165 if (!engine->scratch) {
1166 DRM_ERROR("scratch page not allocated for %s\n", engine->name);
1170 ret = lrc_setup_wa_ctx_obj(engine, PAGE_SIZE);
1172 DRM_DEBUG_DRIVER("Failed to setup context WA page: %d\n", ret);
1176 page = i915_gem_object_get_dirty_page(wa_ctx->vma->obj, 0);
1177 batch = kmap_atomic(page);
1180 if (IS_GEN8(engine->i915)) {
1181 ret = gen8_init_indirectctx_bb(engine,
1182 &wa_ctx->indirect_ctx,
1188 ret = gen8_init_perctx_bb(engine,
1194 } else if (IS_GEN9(engine->i915)) {
1195 ret = gen9_init_indirectctx_bb(engine,
1196 &wa_ctx->indirect_ctx,
1202 ret = gen9_init_perctx_bb(engine,
1211 kunmap_atomic(batch);
1213 lrc_destroy_wa_ctx_obj(engine);
1218 static void lrc_init_hws(struct intel_engine_cs *engine)
1220 struct drm_i915_private *dev_priv = engine->i915;
1222 I915_WRITE(RING_HWS_PGA(engine->mmio_base),
1223 engine->status_page.ggtt_offset);
1224 POSTING_READ(RING_HWS_PGA(engine->mmio_base));
1227 static int gen8_init_common_ring(struct intel_engine_cs *engine)
1229 struct drm_i915_private *dev_priv = engine->i915;
1232 ret = intel_mocs_init_engine(engine);
1236 lrc_init_hws(engine);
1238 intel_engine_reset_breadcrumbs(engine);
1240 I915_WRITE(RING_HWSTAM(engine->mmio_base), 0xffffffff);
1242 I915_WRITE(RING_MODE_GEN7(engine),
1243 _MASKED_BIT_DISABLE(GFX_REPLAY_MODE) |
1244 _MASKED_BIT_ENABLE(GFX_RUN_LIST_ENABLE));
1246 DRM_DEBUG_DRIVER("Execlists enabled for %s\n", engine->name);
1248 intel_engine_init_hangcheck(engine);
1250 if (!execlists_elsp_idle(engine))
1251 execlists_submit_ports(engine);
1256 static int gen8_init_render_ring(struct intel_engine_cs *engine)
1258 struct drm_i915_private *dev_priv = engine->i915;
1261 ret = gen8_init_common_ring(engine);
1265 /* We need to disable the AsyncFlip performance optimisations in order
1266 * to use MI_WAIT_FOR_EVENT within the CS. It should already be
1267 * programmed to '1' on all products.
1269 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw,chv
1271 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
1273 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
1275 return init_workarounds_ring(engine);
1278 static int gen9_init_render_ring(struct intel_engine_cs *engine)
1282 ret = gen8_init_common_ring(engine);
1286 return init_workarounds_ring(engine);
1289 static void reset_common_ring(struct intel_engine_cs *engine,
1290 struct drm_i915_gem_request *request)
1292 struct drm_i915_private *dev_priv = engine->i915;
1293 struct execlist_port *port = engine->execlist_port;
1294 struct intel_context *ce = &request->ctx->engine[engine->id];
1296 /* We want a simple context + ring to execute the breadcrumb update.
1297 * We cannot rely on the context being intact across the GPU hang,
1298 * so clear it and rebuild just what we need for the breadcrumb.
1299 * All pending requests for this context will be zapped, and any
1300 * future request will be after userspace has had the opportunity
1301 * to recreate its own state.
1303 execlists_init_reg_state(ce->lrc_reg_state,
1304 request->ctx, engine, ce->ring);
1306 /* Move the RING_HEAD onto the breadcrumb, past the hanging batch */
1307 ce->lrc_reg_state[CTX_RING_BUFFER_START+1] =
1308 i915_ggtt_offset(ce->ring->vma);
1309 ce->lrc_reg_state[CTX_RING_HEAD+1] = request->postfix;
1311 request->ring->head = request->postfix;
1312 request->ring->last_retired_head = -1;
1313 intel_ring_update_space(request->ring);
1315 if (i915.enable_guc_submission)
1318 /* Catch up with any missed context-switch interrupts */
1319 I915_WRITE(RING_CONTEXT_STATUS_PTR(engine), _MASKED_FIELD(0xffff, 0));
1320 if (request->ctx != port[0].request->ctx) {
1321 i915_gem_request_put(port[0].request);
1323 memset(&port[1], 0, sizeof(port[1]));
1326 /* CS is stopped, and we will resubmit both ports on resume */
1327 GEM_BUG_ON(request->ctx != port[0].request->ctx);
1331 /* Reset WaIdleLiteRestore:bdw,skl as well */
1332 request->tail = request->wa_tail - WA_TAIL_DWORDS * sizeof(u32);
1335 static int intel_logical_ring_emit_pdps(struct drm_i915_gem_request *req)
1337 struct i915_hw_ppgtt *ppgtt = req->ctx->ppgtt;
1338 struct intel_ring *ring = req->ring;
1339 struct intel_engine_cs *engine = req->engine;
1340 const int num_lri_cmds = GEN8_LEGACY_PDPES * 2;
1343 ret = intel_ring_begin(req, num_lri_cmds * 2 + 2);
1347 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_lri_cmds));
1348 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
1349 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
1351 intel_ring_emit_reg(ring, GEN8_RING_PDP_UDW(engine, i));
1352 intel_ring_emit(ring, upper_32_bits(pd_daddr));
1353 intel_ring_emit_reg(ring, GEN8_RING_PDP_LDW(engine, i));
1354 intel_ring_emit(ring, lower_32_bits(pd_daddr));
1357 intel_ring_emit(ring, MI_NOOP);
1358 intel_ring_advance(ring);
1363 static int gen8_emit_bb_start(struct drm_i915_gem_request *req,
1364 u64 offset, u32 len,
1365 unsigned int dispatch_flags)
1367 struct intel_ring *ring = req->ring;
1368 bool ppgtt = !(dispatch_flags & I915_DISPATCH_SECURE);
1371 /* Don't rely in hw updating PDPs, specially in lite-restore.
1372 * Ideally, we should set Force PD Restore in ctx descriptor,
1373 * but we can't. Force Restore would be a second option, but
1374 * it is unsafe in case of lite-restore (because the ctx is
1375 * not idle). PML4 is allocated during ppgtt init so this is
1376 * not needed in 48-bit.*/
1377 if (req->ctx->ppgtt &&
1378 (intel_engine_flag(req->engine) & req->ctx->ppgtt->pd_dirty_rings)) {
1379 if (!USES_FULL_48BIT_PPGTT(req->i915) &&
1380 !intel_vgpu_active(req->i915)) {
1381 ret = intel_logical_ring_emit_pdps(req);
1386 req->ctx->ppgtt->pd_dirty_rings &= ~intel_engine_flag(req->engine);
1389 ret = intel_ring_begin(req, 4);
1393 /* FIXME(BDW): Address space and security selectors. */
1394 intel_ring_emit(ring, MI_BATCH_BUFFER_START_GEN8 |
1396 (dispatch_flags & I915_DISPATCH_RS ?
1397 MI_BATCH_RESOURCE_STREAMER : 0));
1398 intel_ring_emit(ring, lower_32_bits(offset));
1399 intel_ring_emit(ring, upper_32_bits(offset));
1400 intel_ring_emit(ring, MI_NOOP);
1401 intel_ring_advance(ring);
1406 static void gen8_logical_ring_enable_irq(struct intel_engine_cs *engine)
1408 struct drm_i915_private *dev_priv = engine->i915;
1409 I915_WRITE_IMR(engine,
1410 ~(engine->irq_enable_mask | engine->irq_keep_mask));
1411 POSTING_READ_FW(RING_IMR(engine->mmio_base));
1414 static void gen8_logical_ring_disable_irq(struct intel_engine_cs *engine)
1416 struct drm_i915_private *dev_priv = engine->i915;
1417 I915_WRITE_IMR(engine, ~engine->irq_keep_mask);
1420 static int gen8_emit_flush(struct drm_i915_gem_request *request, u32 mode)
1422 struct intel_ring *ring = request->ring;
1426 ret = intel_ring_begin(request, 4);
1430 cmd = MI_FLUSH_DW + 1;
1432 /* We always require a command barrier so that subsequent
1433 * commands, such as breadcrumb interrupts, are strictly ordered
1434 * wrt the contents of the write cache being flushed to memory
1435 * (and thus being coherent from the CPU).
1437 cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1439 if (mode & EMIT_INVALIDATE) {
1440 cmd |= MI_INVALIDATE_TLB;
1441 if (request->engine->id == VCS)
1442 cmd |= MI_INVALIDATE_BSD;
1445 intel_ring_emit(ring, cmd);
1446 intel_ring_emit(ring,
1447 I915_GEM_HWS_SCRATCH_ADDR |
1448 MI_FLUSH_DW_USE_GTT);
1449 intel_ring_emit(ring, 0); /* upper addr */
1450 intel_ring_emit(ring, 0); /* value */
1451 intel_ring_advance(ring);
1456 static int gen8_emit_flush_render(struct drm_i915_gem_request *request,
1459 struct intel_ring *ring = request->ring;
1460 struct intel_engine_cs *engine = request->engine;
1462 i915_ggtt_offset(engine->scratch) + 2 * CACHELINE_BYTES;
1463 bool vf_flush_wa = false, dc_flush_wa = false;
1468 flags |= PIPE_CONTROL_CS_STALL;
1470 if (mode & EMIT_FLUSH) {
1471 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
1472 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
1473 flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
1474 flags |= PIPE_CONTROL_FLUSH_ENABLE;
1477 if (mode & EMIT_INVALIDATE) {
1478 flags |= PIPE_CONTROL_TLB_INVALIDATE;
1479 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
1480 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1481 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
1482 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
1483 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
1484 flags |= PIPE_CONTROL_QW_WRITE;
1485 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
1488 * On GEN9: before VF_CACHE_INVALIDATE we need to emit a NULL
1491 if (IS_GEN9(request->i915))
1494 /* WaForGAMHang:kbl */
1495 if (IS_KBL_REVID(request->i915, 0, KBL_REVID_B0))
1507 ret = intel_ring_begin(request, len);
1512 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6));
1513 intel_ring_emit(ring, 0);
1514 intel_ring_emit(ring, 0);
1515 intel_ring_emit(ring, 0);
1516 intel_ring_emit(ring, 0);
1517 intel_ring_emit(ring, 0);
1521 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6));
1522 intel_ring_emit(ring, PIPE_CONTROL_DC_FLUSH_ENABLE);
1523 intel_ring_emit(ring, 0);
1524 intel_ring_emit(ring, 0);
1525 intel_ring_emit(ring, 0);
1526 intel_ring_emit(ring, 0);
1529 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6));
1530 intel_ring_emit(ring, flags);
1531 intel_ring_emit(ring, scratch_addr);
1532 intel_ring_emit(ring, 0);
1533 intel_ring_emit(ring, 0);
1534 intel_ring_emit(ring, 0);
1537 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6));
1538 intel_ring_emit(ring, PIPE_CONTROL_CS_STALL);
1539 intel_ring_emit(ring, 0);
1540 intel_ring_emit(ring, 0);
1541 intel_ring_emit(ring, 0);
1542 intel_ring_emit(ring, 0);
1545 intel_ring_advance(ring);
1550 static void bxt_a_seqno_barrier(struct intel_engine_cs *engine)
1553 * On BXT A steppings there is a HW coherency issue whereby the
1554 * MI_STORE_DATA_IMM storing the completed request's seqno
1555 * occasionally doesn't invalidate the CPU cache. Work around this by
1556 * clflushing the corresponding cacheline whenever the caller wants
1557 * the coherency to be guaranteed. Note that this cacheline is known
1558 * to be clean at this point, since we only write it in
1559 * bxt_a_set_seqno(), where we also do a clflush after the write. So
1560 * this clflush in practice becomes an invalidate operation.
1562 intel_flush_status_page(engine, I915_GEM_HWS_INDEX);
1566 * Reserve space for 2 NOOPs at the end of each request to be
1567 * used as a workaround for not being allowed to do lite
1568 * restore with HEAD==TAIL (WaIdleLiteRestore).
1571 static int gen8_emit_request(struct drm_i915_gem_request *request)
1573 struct intel_ring *ring = request->ring;
1576 ret = intel_ring_begin(request, 6 + WA_TAIL_DWORDS);
1580 /* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */
1581 BUILD_BUG_ON(I915_GEM_HWS_INDEX_ADDR & (1 << 5));
1583 intel_ring_emit(ring, (MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW);
1584 intel_ring_emit(ring,
1585 intel_hws_seqno_address(request->engine) |
1586 MI_FLUSH_DW_USE_GTT);
1587 intel_ring_emit(ring, 0);
1588 intel_ring_emit(ring, request->fence.seqno);
1589 intel_ring_emit(ring, MI_USER_INTERRUPT);
1590 intel_ring_emit(ring, MI_NOOP);
1591 return intel_logical_ring_advance(request);
1594 static int gen8_emit_request_render(struct drm_i915_gem_request *request)
1596 struct intel_ring *ring = request->ring;
1599 ret = intel_ring_begin(request, 8 + WA_TAIL_DWORDS);
1603 /* We're using qword write, seqno should be aligned to 8 bytes. */
1604 BUILD_BUG_ON(I915_GEM_HWS_INDEX & 1);
1606 /* w/a for post sync ops following a GPGPU operation we
1607 * need a prior CS_STALL, which is emitted by the flush
1608 * following the batch.
1610 intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6));
1611 intel_ring_emit(ring,
1612 (PIPE_CONTROL_GLOBAL_GTT_IVB |
1613 PIPE_CONTROL_CS_STALL |
1614 PIPE_CONTROL_QW_WRITE));
1615 intel_ring_emit(ring, intel_hws_seqno_address(request->engine));
1616 intel_ring_emit(ring, 0);
1617 intel_ring_emit(ring, i915_gem_request_get_seqno(request));
1618 /* We're thrashing one dword of HWS. */
1619 intel_ring_emit(ring, 0);
1620 intel_ring_emit(ring, MI_USER_INTERRUPT);
1621 intel_ring_emit(ring, MI_NOOP);
1622 return intel_logical_ring_advance(request);
1625 static int gen8_init_rcs_context(struct drm_i915_gem_request *req)
1629 ret = intel_logical_ring_workarounds_emit(req);
1633 ret = intel_rcs_context_init_mocs(req);
1635 * Failing to program the MOCS is non-fatal.The system will not
1636 * run at peak performance. So generate an error and carry on.
1639 DRM_ERROR("MOCS failed to program: expect performance issues.\n");
1641 return i915_gem_render_state_init(req);
1645 * intel_logical_ring_cleanup() - deallocate the Engine Command Streamer
1646 * @engine: Engine Command Streamer.
1648 void intel_logical_ring_cleanup(struct intel_engine_cs *engine)
1650 struct drm_i915_private *dev_priv;
1652 if (!intel_engine_initialized(engine))
1656 * Tasklet cannot be active at this point due intel_mark_active/idle
1657 * so this is just for documentation.
1659 if (WARN_ON(test_bit(TASKLET_STATE_SCHED, &engine->irq_tasklet.state)))
1660 tasklet_kill(&engine->irq_tasklet);
1662 dev_priv = engine->i915;
1664 if (engine->buffer) {
1665 WARN_ON((I915_READ_MODE(engine) & MODE_IDLE) == 0);
1668 if (engine->cleanup)
1669 engine->cleanup(engine);
1671 intel_engine_cleanup_common(engine);
1673 if (engine->status_page.vma) {
1674 i915_gem_object_unpin_map(engine->status_page.vma->obj);
1675 engine->status_page.vma = NULL;
1677 intel_lr_context_unpin(dev_priv->kernel_context, engine);
1679 lrc_destroy_wa_ctx_obj(engine);
1680 engine->i915 = NULL;
1683 void intel_execlists_enable_submission(struct drm_i915_private *dev_priv)
1685 struct intel_engine_cs *engine;
1687 for_each_engine(engine, dev_priv)
1688 engine->submit_request = execlists_submit_request;
1692 logical_ring_default_vfuncs(struct intel_engine_cs *engine)
1694 /* Default vfuncs which can be overriden by each engine. */
1695 engine->init_hw = gen8_init_common_ring;
1696 engine->reset_hw = reset_common_ring;
1697 engine->emit_flush = gen8_emit_flush;
1698 engine->emit_request = gen8_emit_request;
1699 engine->submit_request = execlists_submit_request;
1701 engine->irq_enable = gen8_logical_ring_enable_irq;
1702 engine->irq_disable = gen8_logical_ring_disable_irq;
1703 engine->emit_bb_start = gen8_emit_bb_start;
1704 if (IS_BXT_REVID(engine->i915, 0, BXT_REVID_A1))
1705 engine->irq_seqno_barrier = bxt_a_seqno_barrier;
1709 logical_ring_default_irqs(struct intel_engine_cs *engine)
1711 unsigned shift = engine->irq_shift;
1712 engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT << shift;
1713 engine->irq_keep_mask = GT_CONTEXT_SWITCH_INTERRUPT << shift;
1717 lrc_setup_hws(struct intel_engine_cs *engine, struct i915_vma *vma)
1719 const int hws_offset = LRC_PPHWSP_PN * PAGE_SIZE;
1722 /* The HWSP is part of the default context object in LRC mode. */
1723 hws = i915_gem_object_pin_map(vma->obj, I915_MAP_WB);
1725 return PTR_ERR(hws);
1727 engine->status_page.page_addr = hws + hws_offset;
1728 engine->status_page.ggtt_offset = i915_ggtt_offset(vma) + hws_offset;
1729 engine->status_page.vma = vma;
1735 logical_ring_setup(struct intel_engine_cs *engine)
1737 struct drm_i915_private *dev_priv = engine->i915;
1738 enum forcewake_domains fw_domains;
1740 intel_engine_setup_common(engine);
1742 /* Intentionally left blank. */
1743 engine->buffer = NULL;
1745 fw_domains = intel_uncore_forcewake_for_reg(dev_priv,
1749 fw_domains |= intel_uncore_forcewake_for_reg(dev_priv,
1750 RING_CONTEXT_STATUS_PTR(engine),
1751 FW_REG_READ | FW_REG_WRITE);
1753 fw_domains |= intel_uncore_forcewake_for_reg(dev_priv,
1754 RING_CONTEXT_STATUS_BUF_BASE(engine),
1757 engine->fw_domains = fw_domains;
1759 tasklet_init(&engine->irq_tasklet,
1760 intel_lrc_irq_handler, (unsigned long)engine);
1762 logical_ring_init_platform_invariants(engine);
1763 logical_ring_default_vfuncs(engine);
1764 logical_ring_default_irqs(engine);
1768 logical_ring_init(struct intel_engine_cs *engine)
1770 struct i915_gem_context *dctx = engine->i915->kernel_context;
1773 ret = intel_engine_init_common(engine);
1777 ret = execlists_context_deferred_alloc(dctx, engine);
1781 /* As this is the default context, always pin it */
1782 ret = intel_lr_context_pin(dctx, engine);
1784 DRM_ERROR("Failed to pin context for %s: %d\n",
1789 /* And setup the hardware status page. */
1790 ret = lrc_setup_hws(engine, dctx->engine[engine->id].state);
1792 DRM_ERROR("Failed to set up hws %s: %d\n", engine->name, ret);
1799 intel_logical_ring_cleanup(engine);
1803 int logical_render_ring_init(struct intel_engine_cs *engine)
1805 struct drm_i915_private *dev_priv = engine->i915;
1808 logical_ring_setup(engine);
1810 if (HAS_L3_DPF(dev_priv))
1811 engine->irq_keep_mask |= GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
1813 /* Override some for render ring. */
1814 if (INTEL_GEN(dev_priv) >= 9)
1815 engine->init_hw = gen9_init_render_ring;
1817 engine->init_hw = gen8_init_render_ring;
1818 engine->init_context = gen8_init_rcs_context;
1819 engine->emit_flush = gen8_emit_flush_render;
1820 engine->emit_request = gen8_emit_request_render;
1822 ret = intel_engine_create_scratch(engine, 4096);
1826 ret = intel_init_workaround_bb(engine);
1829 * We continue even if we fail to initialize WA batch
1830 * because we only expect rare glitches but nothing
1831 * critical to prevent us from using GPU
1833 DRM_ERROR("WA batch buffer initialization failed: %d\n",
1837 ret = logical_ring_init(engine);
1839 lrc_destroy_wa_ctx_obj(engine);
1845 int logical_xcs_ring_init(struct intel_engine_cs *engine)
1847 logical_ring_setup(engine);
1849 return logical_ring_init(engine);
1853 make_rpcs(struct drm_i915_private *dev_priv)
1858 * No explicit RPCS request is needed to ensure full
1859 * slice/subslice/EU enablement prior to Gen9.
1861 if (INTEL_GEN(dev_priv) < 9)
1865 * Starting in Gen9, render power gating can leave
1866 * slice/subslice/EU in a partially enabled state. We
1867 * must make an explicit request through RPCS for full
1870 if (INTEL_INFO(dev_priv)->sseu.has_slice_pg) {
1871 rpcs |= GEN8_RPCS_S_CNT_ENABLE;
1872 rpcs |= hweight8(INTEL_INFO(dev_priv)->sseu.slice_mask) <<
1873 GEN8_RPCS_S_CNT_SHIFT;
1874 rpcs |= GEN8_RPCS_ENABLE;
1877 if (INTEL_INFO(dev_priv)->sseu.has_subslice_pg) {
1878 rpcs |= GEN8_RPCS_SS_CNT_ENABLE;
1879 rpcs |= hweight8(INTEL_INFO(dev_priv)->sseu.subslice_mask) <<
1880 GEN8_RPCS_SS_CNT_SHIFT;
1881 rpcs |= GEN8_RPCS_ENABLE;
1884 if (INTEL_INFO(dev_priv)->sseu.has_eu_pg) {
1885 rpcs |= INTEL_INFO(dev_priv)->sseu.eu_per_subslice <<
1886 GEN8_RPCS_EU_MIN_SHIFT;
1887 rpcs |= INTEL_INFO(dev_priv)->sseu.eu_per_subslice <<
1888 GEN8_RPCS_EU_MAX_SHIFT;
1889 rpcs |= GEN8_RPCS_ENABLE;
1895 static u32 intel_lr_indirect_ctx_offset(struct intel_engine_cs *engine)
1897 u32 indirect_ctx_offset;
1899 switch (INTEL_GEN(engine->i915)) {
1901 MISSING_CASE(INTEL_GEN(engine->i915));
1904 indirect_ctx_offset =
1905 GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
1908 indirect_ctx_offset =
1909 GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
1913 return indirect_ctx_offset;
1916 static void execlists_init_reg_state(u32 *reg_state,
1917 struct i915_gem_context *ctx,
1918 struct intel_engine_cs *engine,
1919 struct intel_ring *ring)
1921 struct drm_i915_private *dev_priv = engine->i915;
1922 struct i915_hw_ppgtt *ppgtt = ctx->ppgtt ?: dev_priv->mm.aliasing_ppgtt;
1924 /* A context is actually a big batch buffer with several MI_LOAD_REGISTER_IMM
1925 * commands followed by (reg, value) pairs. The values we are setting here are
1926 * only for the first context restore: on a subsequent save, the GPU will
1927 * recreate this batchbuffer with new values (including all the missing
1928 * MI_LOAD_REGISTER_IMM commands that we are not initializing here). */
1929 reg_state[CTX_LRI_HEADER_0] =
1930 MI_LOAD_REGISTER_IMM(engine->id == RCS ? 14 : 11) | MI_LRI_FORCE_POSTED;
1931 ASSIGN_CTX_REG(reg_state, CTX_CONTEXT_CONTROL,
1932 RING_CONTEXT_CONTROL(engine),
1933 _MASKED_BIT_ENABLE(CTX_CTRL_INHIBIT_SYN_CTX_SWITCH |
1934 CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT |
1935 (HAS_RESOURCE_STREAMER(dev_priv) ?
1936 CTX_CTRL_RS_CTX_ENABLE : 0)));
1937 ASSIGN_CTX_REG(reg_state, CTX_RING_HEAD, RING_HEAD(engine->mmio_base),
1939 ASSIGN_CTX_REG(reg_state, CTX_RING_TAIL, RING_TAIL(engine->mmio_base),
1941 ASSIGN_CTX_REG(reg_state, CTX_RING_BUFFER_START,
1942 RING_START(engine->mmio_base), 0);
1943 ASSIGN_CTX_REG(reg_state, CTX_RING_BUFFER_CONTROL,
1944 RING_CTL(engine->mmio_base),
1945 ((ring->size - PAGE_SIZE) & RING_NR_PAGES) | RING_VALID);
1946 ASSIGN_CTX_REG(reg_state, CTX_BB_HEAD_U,
1947 RING_BBADDR_UDW(engine->mmio_base), 0);
1948 ASSIGN_CTX_REG(reg_state, CTX_BB_HEAD_L,
1949 RING_BBADDR(engine->mmio_base), 0);
1950 ASSIGN_CTX_REG(reg_state, CTX_BB_STATE,
1951 RING_BBSTATE(engine->mmio_base),
1953 ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_HEAD_U,
1954 RING_SBBADDR_UDW(engine->mmio_base), 0);
1955 ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_HEAD_L,
1956 RING_SBBADDR(engine->mmio_base), 0);
1957 ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_STATE,
1958 RING_SBBSTATE(engine->mmio_base), 0);
1959 if (engine->id == RCS) {
1960 ASSIGN_CTX_REG(reg_state, CTX_BB_PER_CTX_PTR,
1961 RING_BB_PER_CTX_PTR(engine->mmio_base), 0);
1962 ASSIGN_CTX_REG(reg_state, CTX_RCS_INDIRECT_CTX,
1963 RING_INDIRECT_CTX(engine->mmio_base), 0);
1964 ASSIGN_CTX_REG(reg_state, CTX_RCS_INDIRECT_CTX_OFFSET,
1965 RING_INDIRECT_CTX_OFFSET(engine->mmio_base), 0);
1966 if (engine->wa_ctx.vma) {
1967 struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx;
1968 u32 ggtt_offset = i915_ggtt_offset(wa_ctx->vma);
1970 reg_state[CTX_RCS_INDIRECT_CTX+1] =
1971 (ggtt_offset + wa_ctx->indirect_ctx.offset * sizeof(uint32_t)) |
1972 (wa_ctx->indirect_ctx.size / CACHELINE_DWORDS);
1974 reg_state[CTX_RCS_INDIRECT_CTX_OFFSET+1] =
1975 intel_lr_indirect_ctx_offset(engine) << 6;
1977 reg_state[CTX_BB_PER_CTX_PTR+1] =
1978 (ggtt_offset + wa_ctx->per_ctx.offset * sizeof(uint32_t)) |
1982 reg_state[CTX_LRI_HEADER_1] = MI_LOAD_REGISTER_IMM(9) | MI_LRI_FORCE_POSTED;
1983 ASSIGN_CTX_REG(reg_state, CTX_CTX_TIMESTAMP,
1984 RING_CTX_TIMESTAMP(engine->mmio_base), 0);
1985 /* PDP values well be assigned later if needed */
1986 ASSIGN_CTX_REG(reg_state, CTX_PDP3_UDW, GEN8_RING_PDP_UDW(engine, 3),
1988 ASSIGN_CTX_REG(reg_state, CTX_PDP3_LDW, GEN8_RING_PDP_LDW(engine, 3),
1990 ASSIGN_CTX_REG(reg_state, CTX_PDP2_UDW, GEN8_RING_PDP_UDW(engine, 2),
1992 ASSIGN_CTX_REG(reg_state, CTX_PDP2_LDW, GEN8_RING_PDP_LDW(engine, 2),
1994 ASSIGN_CTX_REG(reg_state, CTX_PDP1_UDW, GEN8_RING_PDP_UDW(engine, 1),
1996 ASSIGN_CTX_REG(reg_state, CTX_PDP1_LDW, GEN8_RING_PDP_LDW(engine, 1),
1998 ASSIGN_CTX_REG(reg_state, CTX_PDP0_UDW, GEN8_RING_PDP_UDW(engine, 0),
2000 ASSIGN_CTX_REG(reg_state, CTX_PDP0_LDW, GEN8_RING_PDP_LDW(engine, 0),
2003 if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
2004 /* 64b PPGTT (48bit canonical)
2005 * PDP0_DESCRIPTOR contains the base address to PML4 and
2006 * other PDP Descriptors are ignored.
2008 ASSIGN_CTX_PML4(ppgtt, reg_state);
2011 * PDP*_DESCRIPTOR contains the base address of space supported.
2012 * With dynamic page allocation, PDPs may not be allocated at
2013 * this point. Point the unallocated PDPs to the scratch page
2015 execlists_update_context_pdps(ppgtt, reg_state);
2018 if (engine->id == RCS) {
2019 reg_state[CTX_LRI_HEADER_2] = MI_LOAD_REGISTER_IMM(1);
2020 ASSIGN_CTX_REG(reg_state, CTX_R_PWR_CLK_STATE, GEN8_R_PWR_CLK_STATE,
2021 make_rpcs(dev_priv));
2026 populate_lr_context(struct i915_gem_context *ctx,
2027 struct drm_i915_gem_object *ctx_obj,
2028 struct intel_engine_cs *engine,
2029 struct intel_ring *ring)
2034 ret = i915_gem_object_set_to_cpu_domain(ctx_obj, true);
2036 DRM_DEBUG_DRIVER("Could not set to CPU domain\n");
2040 vaddr = i915_gem_object_pin_map(ctx_obj, I915_MAP_WB);
2041 if (IS_ERR(vaddr)) {
2042 ret = PTR_ERR(vaddr);
2043 DRM_DEBUG_DRIVER("Could not map object pages! (%d)\n", ret);
2046 ctx_obj->dirty = true;
2048 /* The second page of the context object contains some fields which must
2049 * be set up prior to the first execution. */
2051 execlists_init_reg_state(vaddr + LRC_STATE_PN * PAGE_SIZE,
2054 i915_gem_object_unpin_map(ctx_obj);
2060 * intel_lr_context_size() - return the size of the context for an engine
2061 * @engine: which engine to find the context size for
2063 * Each engine may require a different amount of space for a context image,
2064 * so when allocating (or copying) an image, this function can be used to
2065 * find the right size for the specific engine.
2067 * Return: size (in bytes) of an engine-specific context image
2069 * Note: this size includes the HWSP, which is part of the context image
2070 * in LRC mode, but does not include the "shared data page" used with
2071 * GuC submission. The caller should account for this if using the GuC.
2073 uint32_t intel_lr_context_size(struct intel_engine_cs *engine)
2077 WARN_ON(INTEL_GEN(engine->i915) < 8);
2079 switch (engine->id) {
2081 if (INTEL_GEN(engine->i915) >= 9)
2082 ret = GEN9_LR_CONTEXT_RENDER_SIZE;
2084 ret = GEN8_LR_CONTEXT_RENDER_SIZE;
2090 ret = GEN8_LR_CONTEXT_OTHER_SIZE;
2097 static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
2098 struct intel_engine_cs *engine)
2100 struct drm_i915_gem_object *ctx_obj;
2101 struct intel_context *ce = &ctx->engine[engine->id];
2102 struct i915_vma *vma;
2103 uint32_t context_size;
2104 struct intel_ring *ring;
2109 context_size = round_up(intel_lr_context_size(engine), 4096);
2111 /* One extra page as the sharing data between driver and GuC */
2112 context_size += PAGE_SIZE * LRC_PPHWSP_PN;
2114 ctx_obj = i915_gem_object_create(&ctx->i915->drm, context_size);
2115 if (IS_ERR(ctx_obj)) {
2116 DRM_DEBUG_DRIVER("Alloc LRC backing obj failed.\n");
2117 return PTR_ERR(ctx_obj);
2120 vma = i915_vma_create(ctx_obj, &ctx->i915->ggtt.base, NULL);
2123 goto error_deref_obj;
2126 ring = intel_engine_create_ring(engine, ctx->ring_size);
2128 ret = PTR_ERR(ring);
2129 goto error_deref_obj;
2132 ret = populate_lr_context(ctx, ctx_obj, engine, ring);
2134 DRM_DEBUG_DRIVER("Failed to populate LRC: %d\n", ret);
2135 goto error_ring_free;
2140 ce->initialised = engine->init_context == NULL;
2145 intel_ring_free(ring);
2147 i915_gem_object_put(ctx_obj);
2151 void intel_lr_context_resume(struct drm_i915_private *dev_priv)
2153 struct i915_gem_context *ctx = dev_priv->kernel_context;
2154 struct intel_engine_cs *engine;
2156 for_each_engine(engine, dev_priv) {
2157 struct intel_context *ce = &ctx->engine[engine->id];
2159 uint32_t *reg_state;
2164 vaddr = i915_gem_object_pin_map(ce->state->obj, I915_MAP_WB);
2165 if (WARN_ON(IS_ERR(vaddr)))
2168 reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
2170 reg_state[CTX_RING_HEAD+1] = 0;
2171 reg_state[CTX_RING_TAIL+1] = 0;
2173 ce->state->obj->dirty = true;
2174 i915_gem_object_unpin_map(ce->state->obj);