xref: /dflybsd-src/sys/dev/drm/i915/intel_lrc.c (revision 2c3b1d1bc3a233e4bf80218452e719882a1bb011)
1 /*
2  * Copyright © 2014 Intel Corporation
3  *
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:
10  *
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
13  * Software.
14  *
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
21  * IN THE SOFTWARE.
22  *
23  * Authors:
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>
28  *
29  */
30 
31 /**
32  * DOC: Logical Rings, Logical Ring Contexts and Execlists
33  *
34  * Motivation:
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).
38  *
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:
42  *
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.
47  *
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.
55  *
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".
58  *
59  * LRC implementation:
60  * Regarding the creation of contexts, we have:
61  *
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.
65  *
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:
69  *
70  * - One ringbuffer per-engine inside each context.
71  * - One backing object per-engine inside each context.
72  *
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
77  * contexts:
78  *
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
83  * so on.
84  *
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).
88  *
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:
93  *
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
101  * context itself.
102  *
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.
108  *
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.
114  *
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.
122  *
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).
132  *
133  */
134 #include <linux/interrupt.h>
135 
136 #include <drm/drmP.h>
137 #include <drm/i915_drm.h>
138 #include "i915_drv.h"
139 #include "intel_mocs.h"
140 
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)
144 
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)
151 
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)
158 
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
187 
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)
193 
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); \
197 } while (0)
198 
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); \
203 } while (0)
204 
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)); \
208 } while (0)
209 
210 enum {
211 	FAULT_AND_HANG = 0,
212 	FAULT_AND_HALT, /* Debug only */
213 	FAULT_AND_STREAM,
214 	FAULT_AND_CONTINUE /* Unsupported */
215 };
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
220 
221 /* Typical size of the average request (2 pipecontrols and a MI_BB) */
222 #define EXECLISTS_REQUEST_SIZE 64 /* bytes */
223 
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);
228 
229 /**
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.
233  *
234  * Only certain platforms support Execlists (the prerequisites being
235  * support for Logical Ring Contexts and Aliasing PPGTT or better).
236  *
237  * Return: 1 if Execlists is supported and has to be enabled.
238  */
239 int intel_sanitize_enable_execlists(struct drm_i915_private *dev_priv, int enable_execlists)
240 {
241 	/* On platforms with execlist available, vGPU will only
242 	 * support execlist mode, no ring buffer mode.
243 	 */
244 	if (HAS_LOGICAL_RING_CONTEXTS(dev_priv) && intel_vgpu_active(dev_priv))
245 		return 1;
246 
247 	if (INTEL_GEN(dev_priv) >= 9)
248 		return 1;
249 
250 	if (enable_execlists == 0)
251 		return 0;
252 
253 	if (HAS_LOGICAL_RING_CONTEXTS(dev_priv) &&
254 	    USES_PPGTT(dev_priv) &&
255 	    i915.use_mmio_flip >= 0)
256 		return 1;
257 
258 	return 0;
259 }
260 
261 static void
262 logical_ring_init_platform_invariants(struct intel_engine_cs *engine)
263 {
264 	struct drm_i915_private *dev_priv = engine->i915;
265 
266 	if (IS_GEN8(dev_priv) || IS_GEN9(dev_priv))
267 		engine->idle_lite_restore_wa = ~0;
268 
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);
272 
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;
277 
278 	/* TODO: WaDisableLiteRestore when we start using semaphore
279 	 * signalling between Command Streamers */
280 	/* ring->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE; */
281 
282 	/* WaEnableForceRestoreInCtxtDescForVCS:skl */
283 	/* WaEnableForceRestoreInCtxtDescForVCS:bxt */
284 	if (engine->disable_lite_restore_wa)
285 		engine->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE;
286 }
287 
288 /**
289  * intel_lr_context_descriptor_update() - calculate & cache the descriptor
290  * 					  descriptor for a pinned context
291  *
292  * @ctx: Context to work on
293  * @engine: Engine the descriptor will be used with
294  *
295  * The context descriptor encodes various attributes of a context,
296  * including its GTT address and some flags. Because it's fairly
297  * expensive to calculate, we'll just do it once and cache the result,
298  * which remains valid until the context is unpinned.
299  *
300  * This is what a descriptor looks like, from LSB to MSB:
301  *    bits  0-11:    flags, GEN8_CTX_* (cached in ctx_desc_template)
302  *    bits 12-31:    LRCA, GTT address of (the HWSP of) this context
303  *    bits 32-52:    ctx ID, a globally unique tag
304  *    bits 53-54:    mbz, reserved for use by hardware
305  *    bits 55-63:    group ID, currently unused and set to 0
306  */
307 static void
308 intel_lr_context_descriptor_update(struct i915_gem_context *ctx,
309 				   struct intel_engine_cs *engine)
310 {
311 	struct intel_context *ce = &ctx->engine[engine->id];
312 	u64 desc;
313 
314 	BUILD_BUG_ON(MAX_CONTEXT_HW_ID > (1<<GEN8_CTX_ID_WIDTH));
315 
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;
319 								/* bits 12-31 */
320 	desc |= (u64)ctx->hw_id << GEN8_CTX_ID_SHIFT;		/* bits 32-52 */
321 
322 	ce->lrc_desc = desc;
323 }
324 
325 uint64_t intel_lr_context_descriptor(struct i915_gem_context *ctx,
326 				     struct intel_engine_cs *engine)
327 {
328 	return ctx->engine[engine->id].lrc_desc;
329 }
330 
331 static void execlists_elsp_write(struct drm_i915_gem_request *rq0,
332 				 struct drm_i915_gem_request *rq1)
333 {
334 
335 	struct intel_engine_cs *engine = rq0->engine;
336 	struct drm_i915_private *dev_priv = rq0->i915;
337 	uint64_t desc[2];
338 
339 	if (rq1) {
340 		desc[1] = intel_lr_context_descriptor(rq1->ctx, rq1->engine);
341 		rq1->elsp_submitted++;
342 	} else {
343 		desc[1] = 0;
344 	}
345 
346 	desc[0] = intel_lr_context_descriptor(rq0->ctx, rq0->engine);
347 	rq0->elsp_submitted++;
348 
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]));
352 
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]));
356 
357 	/* ELSP is a wo register, use another nearby reg for posting */
358 	POSTING_READ_FW(RING_EXECLIST_STATUS_LO(engine));
359 }
360 
361 static void
362 execlists_update_context_pdps(struct i915_hw_ppgtt *ppgtt, u32 *reg_state)
363 {
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);
368 }
369 
370 static void execlists_update_context(struct drm_i915_gem_request *rq)
371 {
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;
375 
376 	reg_state[CTX_RING_TAIL+1] = rq->tail;
377 
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
381 	 * in 48-bit mode.
382 	 */
383 	if (ppgtt && !USES_FULL_48BIT_PPGTT(ppgtt->base.dev))
384 		execlists_update_context_pdps(ppgtt, reg_state);
385 }
386 
387 static void execlists_submit_requests(struct drm_i915_gem_request *rq0,
388 				      struct drm_i915_gem_request *rq1)
389 {
390 	struct drm_i915_private *dev_priv = rq0->i915;
391 	unsigned int fw_domains = rq0->engine->fw_domains;
392 
393 	execlists_update_context(rq0);
394 
395 	if (rq1)
396 		execlists_update_context(rq1);
397 
398 	spin_lock_irq(&dev_priv->uncore.lock);
399 	intel_uncore_forcewake_get__locked(dev_priv, fw_domains);
400 
401 	execlists_elsp_write(rq0, rq1);
402 
403 	intel_uncore_forcewake_put__locked(dev_priv, fw_domains);
404 	spin_unlock_irq(&dev_priv->uncore.lock);
405 }
406 
407 static inline void execlists_context_status_change(
408 		struct drm_i915_gem_request *rq,
409 		unsigned long status)
410 {
411 	/*
412 	 * Only used when GVT-g is enabled now. When GVT-g is disabled,
413 	 * The compiler should eliminate this function as dead-code.
414 	 */
415 	if (!IS_ENABLED(CONFIG_DRM_I915_GVT))
416 		return;
417 
418 	atomic_notifier_call_chain(&rq->ctx->status_notifier, status, rq);
419 }
420 
421 static void execlists_context_unqueue(struct intel_engine_cs *engine)
422 {
423 	struct drm_i915_gem_request *req0 = NULL, *req1 = NULL;
424 	struct drm_i915_gem_request *cursor, *tmp;
425 
426 	assert_spin_locked(&engine->execlist_lock);
427 
428 	/*
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.
431 	 */
432 	WARN_ON(!intel_irqs_enabled(engine->i915));
433 
434 	/* Try to read in pairs */
435 	list_for_each_entry_safe(cursor, tmp, &engine->execlist_queue,
436 				 execlist_link) {
437 		if (!req0) {
438 			req0 = cursor;
439 		} else if (req0->ctx == cursor->ctx) {
440 			/* Same ctx: ignore first request, as second request
441 			 * will update tail past first request's workload */
442 			cursor->elsp_submitted = req0->elsp_submitted;
443 			list_del(&req0->execlist_link);
444 			i915_gem_request_unreference(req0);
445 			req0 = cursor;
446 		} else {
447 			if (IS_ENABLED(CONFIG_DRM_I915_GVT)) {
448 				/*
449 				 * req0 (after merged) ctx requires single
450 				 * submission, stop picking
451 				 */
452 				if (req0->ctx->execlists_force_single_submission)
453 					break;
454 				/*
455 				 * req0 ctx doesn't require single submission,
456 				 * but next req ctx requires, stop picking
457 				 */
458 				if (cursor->ctx->execlists_force_single_submission)
459 					break;
460 			}
461 			req1 = cursor;
462 			WARN_ON(req1->elsp_submitted);
463 			break;
464 		}
465 	}
466 
467 	if (unlikely(!req0))
468 		return;
469 
470 	execlists_context_status_change(req0, INTEL_CONTEXT_SCHEDULE_IN);
471 
472 	if (req1)
473 		execlists_context_status_change(req1,
474 						INTEL_CONTEXT_SCHEDULE_IN);
475 
476 	if (req0->elsp_submitted & engine->idle_lite_restore_wa) {
477 		/*
478 		 * WaIdleLiteRestore: make sure we never cause a lite restore
479 		 * with HEAD==TAIL.
480 		 *
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.
484 		 */
485 		struct intel_ringbuffer *ringbuf;
486 
487 		ringbuf = req0->ctx->engine[engine->id].ringbuf;
488 		req0->tail += 8;
489 		req0->tail &= ringbuf->size - 1;
490 	}
491 
492 	execlists_submit_requests(req0, req1);
493 }
494 
495 static unsigned int
496 execlists_check_remove_request(struct intel_engine_cs *engine, u32 ctx_id)
497 {
498 	struct drm_i915_gem_request *head_req;
499 
500 	assert_spin_locked(&engine->execlist_lock);
501 
502 	head_req = list_first_entry_or_null(&engine->execlist_queue,
503 					    struct drm_i915_gem_request,
504 					    execlist_link);
505 
506 	if (WARN_ON(!head_req || (head_req->ctx_hw_id != ctx_id)))
507                return 0;
508 
509 	WARN(head_req->elsp_submitted == 0, "Never submitted head request\n");
510 
511 	if (--head_req->elsp_submitted > 0)
512 		return 0;
513 
514 	execlists_context_status_change(head_req, INTEL_CONTEXT_SCHEDULE_OUT);
515 
516 	list_del(&head_req->execlist_link);
517 	i915_gem_request_unreference(head_req);
518 
519 	return 1;
520 }
521 
522 static u32
523 get_context_status(struct intel_engine_cs *engine, unsigned int read_pointer,
524 		   u32 *context_id)
525 {
526 	struct drm_i915_private *dev_priv = engine->i915;
527 	u32 status;
528 
529 	read_pointer %= GEN8_CSB_ENTRIES;
530 
531 	status = I915_READ_FW(RING_CONTEXT_STATUS_BUF_LO(engine, read_pointer));
532 
533 	if (status & GEN8_CTX_STATUS_IDLE_ACTIVE)
534 		return 0;
535 
536 	*context_id = I915_READ_FW(RING_CONTEXT_STATUS_BUF_HI(engine,
537 							      read_pointer));
538 
539 	return status;
540 }
541 
542 /**
543  * intel_lrc_irq_handler() - handle Context Switch interrupts
544  * @data: tasklet handler passed in unsigned long
545  *
546  * Check the unread Context Status Buffers and manage the submission of new
547  * contexts to the ELSP accordingly.
548  */
549 static void intel_lrc_irq_handler(unsigned long data)
550 {
551 	struct intel_engine_cs *engine = (struct intel_engine_cs *)data;
552 	struct drm_i915_private *dev_priv = engine->i915;
553 	u32 status_pointer;
554 	unsigned int read_pointer, write_pointer;
555 	u32 csb[GEN8_CSB_ENTRIES][2];
556 	unsigned int csb_read = 0, i;
557 	unsigned int submit_contexts = 0;
558 
559 	intel_uncore_forcewake_get(dev_priv, engine->fw_domains);
560 
561 	status_pointer = I915_READ_FW(RING_CONTEXT_STATUS_PTR(engine));
562 
563 	read_pointer = engine->next_context_status_buffer;
564 	write_pointer = GEN8_CSB_WRITE_PTR(status_pointer);
565 	if (read_pointer > write_pointer)
566 		write_pointer += GEN8_CSB_ENTRIES;
567 
568 	while (read_pointer < write_pointer) {
569 		if (WARN_ON_ONCE(csb_read == GEN8_CSB_ENTRIES))
570 			break;
571 		csb[csb_read][0] = get_context_status(engine, ++read_pointer,
572 						      &csb[csb_read][1]);
573 		csb_read++;
574 	}
575 
576 	engine->next_context_status_buffer = write_pointer % GEN8_CSB_ENTRIES;
577 
578 	/* Update the read pointer to the old write pointer. Manual ringbuffer
579 	 * management ftw </sarcasm> */
580 	I915_WRITE_FW(RING_CONTEXT_STATUS_PTR(engine),
581 		      _MASKED_FIELD(GEN8_CSB_READ_PTR_MASK,
582 				    engine->next_context_status_buffer << 8));
583 
584 	intel_uncore_forcewake_put(dev_priv, engine->fw_domains);
585 
586 	lockmgr(&engine->execlist_lock, LK_EXCLUSIVE);
587 
588 	for (i = 0; i < csb_read; i++) {
589 		if (unlikely(csb[i][0] & GEN8_CTX_STATUS_PREEMPTED)) {
590 			if (csb[i][0] & GEN8_CTX_STATUS_LITE_RESTORE) {
591 				if (execlists_check_remove_request(engine, csb[i][1]))
592 					WARN(1, "Lite Restored request removed from queue\n");
593 			} else
594 				WARN(1, "Preemption without Lite Restore\n");
595 		}
596 
597 		if (csb[i][0] & (GEN8_CTX_STATUS_ACTIVE_IDLE |
598 		    GEN8_CTX_STATUS_ELEMENT_SWITCH))
599 			submit_contexts +=
600 				execlists_check_remove_request(engine, csb[i][1]);
601 	}
602 
603 	if (submit_contexts) {
604 		if (!engine->disable_lite_restore_wa ||
605 		    (csb[i][0] & GEN8_CTX_STATUS_ACTIVE_IDLE))
606 			execlists_context_unqueue(engine);
607 	}
608 
609 	lockmgr(&engine->execlist_lock, LK_RELEASE);
610 
611 	if (unlikely(submit_contexts > 2))
612 		DRM_ERROR("More than two context complete events?\n");
613 }
614 
615 static void execlists_context_queue(struct drm_i915_gem_request *request)
616 {
617 	struct intel_engine_cs *engine = request->engine;
618 	struct drm_i915_gem_request *cursor;
619 	int num_elements = 0;
620 
621 	spin_lock_bh(&engine->execlist_lock);
622 
623 	list_for_each_entry(cursor, &engine->execlist_queue, execlist_link)
624 		if (++num_elements > 2)
625 			break;
626 
627 	if (num_elements > 2) {
628 		struct drm_i915_gem_request *tail_req;
629 
630 		tail_req = list_last_entry(&engine->execlist_queue,
631 					   struct drm_i915_gem_request,
632 					   execlist_link);
633 
634 		if (request->ctx == tail_req->ctx) {
635 			WARN(tail_req->elsp_submitted != 0,
636 				"More than 2 already-submitted reqs queued\n");
637 			list_del(&tail_req->execlist_link);
638 			i915_gem_request_unreference(tail_req);
639 		}
640 	}
641 
642 	i915_gem_request_reference(request);
643 	list_add_tail(&request->execlist_link, &engine->execlist_queue);
644 	request->ctx_hw_id = request->ctx->hw_id;
645 	if (num_elements == 0)
646 		execlists_context_unqueue(engine);
647 
648 	spin_unlock_bh(&engine->execlist_lock);
649 }
650 
651 static int logical_ring_invalidate_all_caches(struct drm_i915_gem_request *req)
652 {
653 	struct intel_engine_cs *engine = req->engine;
654 	uint32_t flush_domains;
655 	int ret;
656 
657 	flush_domains = 0;
658 	if (engine->gpu_caches_dirty)
659 		flush_domains = I915_GEM_GPU_DOMAINS;
660 
661 	ret = engine->emit_flush(req, I915_GEM_GPU_DOMAINS, flush_domains);
662 	if (ret)
663 		return ret;
664 
665 	engine->gpu_caches_dirty = false;
666 	return 0;
667 }
668 
669 static int execlists_move_to_gpu(struct drm_i915_gem_request *req,
670 				 struct list_head *vmas)
671 {
672 	const unsigned other_rings = ~intel_engine_flag(req->engine);
673 	struct i915_vma *vma;
674 	uint32_t flush_domains = 0;
675 	bool flush_chipset = false;
676 	int ret;
677 
678 	list_for_each_entry(vma, vmas, exec_list) {
679 		struct drm_i915_gem_object *obj = vma->obj;
680 
681 		if (obj->active & other_rings) {
682 			ret = i915_gem_object_sync(obj, req->engine, &req);
683 			if (ret)
684 				return ret;
685 		}
686 
687 		if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
688 			flush_chipset |= i915_gem_clflush_object(obj, false);
689 
690 		flush_domains |= obj->base.write_domain;
691 	}
692 
693 	if (flush_domains & I915_GEM_DOMAIN_GTT)
694 		wmb();
695 
696 	/* Unconditionally invalidate gpu caches and ensure that we do flush
697 	 * any residual writes from the previous batch.
698 	 */
699 	return logical_ring_invalidate_all_caches(req);
700 }
701 
702 int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *request)
703 {
704 	struct intel_engine_cs *engine = request->engine;
705 	struct intel_context *ce = &request->ctx->engine[engine->id];
706 	int ret;
707 
708 	/* Flush enough space to reduce the likelihood of waiting after
709 	 * we start building the request - in which case we will just
710 	 * have to repeat work.
711 	 */
712 	request->reserved_space += EXECLISTS_REQUEST_SIZE;
713 
714 	if (!ce->state) {
715 		ret = execlists_context_deferred_alloc(request->ctx, engine);
716 		if (ret)
717 			return ret;
718 	}
719 
720 	request->ringbuf = ce->ringbuf;
721 
722 	if (i915.enable_guc_submission) {
723 		/*
724 		 * Check that the GuC has space for the request before
725 		 * going any further, as the i915_add_request() call
726 		 * later on mustn't fail ...
727 		 */
728 		ret = i915_guc_wq_check_space(request);
729 		if (ret)
730 			return ret;
731 	}
732 
733 	ret = intel_lr_context_pin(request->ctx, engine);
734 	if (ret)
735 		return ret;
736 
737 	ret = intel_ring_begin(request, 0);
738 	if (ret)
739 		goto err_unpin;
740 
741 	if (!ce->initialised) {
742 		ret = engine->init_context(request);
743 		if (ret)
744 			goto err_unpin;
745 
746 		ce->initialised = true;
747 	}
748 
749 	/* Note that after this point, we have committed to using
750 	 * this request as it is being used to both track the
751 	 * state of engine initialisation and liveness of the
752 	 * golden renderstate above. Think twice before you try
753 	 * to cancel/unwind this request now.
754 	 */
755 
756 	request->reserved_space -= EXECLISTS_REQUEST_SIZE;
757 	return 0;
758 
759 err_unpin:
760 	intel_lr_context_unpin(request->ctx, engine);
761 	return ret;
762 }
763 
764 /*
765  * intel_logical_ring_advance_and_submit() - advance the tail and submit the workload
766  * @request: Request to advance the logical ringbuffer of.
767  *
768  * The tail is updated in our logical ringbuffer struct, not in the actual context. What
769  * really happens during submission is that the context and current tail will be placed
770  * on a queue waiting for the ELSP to be ready to accept a new context submission. At that
771  * point, the tail *inside* the context is updated and the ELSP written to.
772  */
773 static int
774 intel_logical_ring_advance_and_submit(struct drm_i915_gem_request *request)
775 {
776 	struct intel_ringbuffer *ringbuf = request->ringbuf;
777 	struct intel_engine_cs *engine = request->engine;
778 
779 	intel_logical_ring_advance(ringbuf);
780 	request->tail = ringbuf->tail;
781 
782 	/*
783 	 * Here we add two extra NOOPs as padding to avoid
784 	 * lite restore of a context with HEAD==TAIL.
785 	 *
786 	 * Caller must reserve WA_TAIL_DWORDS for us!
787 	 */
788 	intel_logical_ring_emit(ringbuf, MI_NOOP);
789 	intel_logical_ring_emit(ringbuf, MI_NOOP);
790 	intel_logical_ring_advance(ringbuf);
791 
792 	/* We keep the previous context alive until we retire the following
793 	 * request. This ensures that any the context object is still pinned
794 	 * for any residual writes the HW makes into it on the context switch
795 	 * into the next object following the breadcrumb. Otherwise, we may
796 	 * retire the context too early.
797 	 */
798 	request->previous_context = engine->last_context;
799 	engine->last_context = request->ctx;
800 
801 	if (i915.enable_guc_submission)
802 		i915_guc_submit(request);
803 	else
804 		execlists_context_queue(request);
805 
806 	return 0;
807 }
808 
809 /**
810  * execlists_submission() - submit a batchbuffer for execution, Execlists style
811  * @params: execbuffer call parameters.
812  * @args: execbuffer call arguments.
813  * @vmas: list of vmas.
814  *
815  * This is the evil twin version of i915_gem_ringbuffer_submission. It abstracts
816  * away the submission details of the execbuffer ioctl call.
817  *
818  * Return: non-zero if the submission fails.
819  */
820 int intel_execlists_submission(struct i915_execbuffer_params *params,
821 			       struct drm_i915_gem_execbuffer2 *args,
822 			       struct list_head *vmas)
823 {
824 	struct drm_device       *dev = params->dev;
825 	struct intel_engine_cs *engine = params->engine;
826 	struct drm_i915_private *dev_priv = to_i915(dev);
827 	struct intel_ringbuffer *ringbuf = params->ctx->engine[engine->id].ringbuf;
828 	u64 exec_start;
829 	int instp_mode;
830 	u32 instp_mask;
831 	int ret;
832 
833 	instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
834 	instp_mask = I915_EXEC_CONSTANTS_MASK;
835 	switch (instp_mode) {
836 	case I915_EXEC_CONSTANTS_REL_GENERAL:
837 	case I915_EXEC_CONSTANTS_ABSOLUTE:
838 	case I915_EXEC_CONSTANTS_REL_SURFACE:
839 		if (instp_mode != 0 && engine != &dev_priv->engine[RCS]) {
840 			DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
841 			return -EINVAL;
842 		}
843 
844 		if (instp_mode != dev_priv->relative_constants_mode) {
845 			if (instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
846 				DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
847 				return -EINVAL;
848 			}
849 
850 			/* The HW changed the meaning on this bit on gen6 */
851 			instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
852 		}
853 		break;
854 	default:
855 		DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode);
856 		return -EINVAL;
857 	}
858 
859 	if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
860 		DRM_DEBUG("sol reset is gen7 only\n");
861 		return -EINVAL;
862 	}
863 
864 	ret = execlists_move_to_gpu(params->request, vmas);
865 	if (ret)
866 		return ret;
867 
868 	if (engine == &dev_priv->engine[RCS] &&
869 	    instp_mode != dev_priv->relative_constants_mode) {
870 		ret = intel_ring_begin(params->request, 4);
871 		if (ret)
872 			return ret;
873 
874 		intel_logical_ring_emit(ringbuf, MI_NOOP);
875 		intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(1));
876 		intel_logical_ring_emit_reg(ringbuf, INSTPM);
877 		intel_logical_ring_emit(ringbuf, instp_mask << 16 | instp_mode);
878 		intel_logical_ring_advance(ringbuf);
879 
880 		dev_priv->relative_constants_mode = instp_mode;
881 	}
882 
883 	exec_start = params->batch_obj_vm_offset +
884 		     args->batch_start_offset;
885 
886 	ret = engine->emit_bb_start(params->request, exec_start, params->dispatch_flags);
887 	if (ret)
888 		return ret;
889 
890 	trace_i915_gem_ring_dispatch(params->request, params->dispatch_flags);
891 
892 	i915_gem_execbuffer_move_to_active(vmas, params->request);
893 
894 	return 0;
895 }
896 
897 void intel_execlists_cancel_requests(struct intel_engine_cs *engine)
898 {
899 	struct drm_i915_gem_request *req, *tmp;
900 	LINUX_LIST_HEAD(cancel_list);
901 
902 	WARN_ON(!mutex_is_locked(&engine->i915->drm.struct_mutex));
903 
904 	spin_lock_bh(&engine->execlist_lock);
905 	list_replace_init(&engine->execlist_queue, &cancel_list);
906 	spin_unlock_bh(&engine->execlist_lock);
907 
908 	list_for_each_entry_safe(req, tmp, &cancel_list, execlist_link) {
909 		list_del(&req->execlist_link);
910 		i915_gem_request_unreference(req);
911 	}
912 }
913 
914 void intel_logical_ring_stop(struct intel_engine_cs *engine)
915 {
916 	struct drm_i915_private *dev_priv = engine->i915;
917 	int ret;
918 
919 	if (!intel_engine_initialized(engine))
920 		return;
921 
922 	ret = intel_engine_idle(engine);
923 	if (ret)
924 		DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n",
925 			  engine->name, ret);
926 
927 	/* TODO: Is this correct with Execlists enabled? */
928 	I915_WRITE_MODE(engine, _MASKED_BIT_ENABLE(STOP_RING));
929 	if (intel_wait_for_register(dev_priv,
930 				    RING_MI_MODE(engine->mmio_base),
931 				    MODE_IDLE, MODE_IDLE,
932 				    1000)) {
933 		DRM_ERROR("%s :timed out trying to stop ring\n", engine->name);
934 		return;
935 	}
936 	I915_WRITE_MODE(engine, _MASKED_BIT_DISABLE(STOP_RING));
937 }
938 
939 int logical_ring_flush_all_caches(struct drm_i915_gem_request *req)
940 {
941 	struct intel_engine_cs *engine = req->engine;
942 	int ret;
943 
944 	if (!engine->gpu_caches_dirty)
945 		return 0;
946 
947 	ret = engine->emit_flush(req, 0, I915_GEM_GPU_DOMAINS);
948 	if (ret)
949 		return ret;
950 
951 	engine->gpu_caches_dirty = false;
952 	return 0;
953 }
954 
955 static int intel_lr_context_pin(struct i915_gem_context *ctx,
956 				struct intel_engine_cs *engine)
957 {
958 	struct drm_i915_private *dev_priv = ctx->i915;
959 	struct intel_context *ce = &ctx->engine[engine->id];
960 	void *vaddr;
961 	u32 *lrc_reg_state;
962 	int ret;
963 
964 	lockdep_assert_held(&ctx->i915->drm.struct_mutex);
965 
966 	if (ce->pin_count++)
967 		return 0;
968 
969 	ret = i915_gem_obj_ggtt_pin(ce->state, GEN8_LR_CONTEXT_ALIGN,
970 				    PIN_OFFSET_BIAS | GUC_WOPCM_TOP);
971 	if (ret)
972 		goto err;
973 
974 	vaddr = i915_gem_object_pin_map(ce->state);
975 	if (IS_ERR(vaddr)) {
976 		ret = PTR_ERR(vaddr);
977 		goto unpin_ctx_obj;
978 	}
979 
980 	lrc_reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
981 
982 	ret = intel_pin_and_map_ringbuffer_obj(dev_priv, ce->ringbuf);
983 	if (ret)
984 		goto unpin_map;
985 
986 	i915_gem_context_reference(ctx);
987 	ce->lrc_vma = i915_gem_obj_to_ggtt(ce->state);
988 	intel_lr_context_descriptor_update(ctx, engine);
989 
990 	lrc_reg_state[CTX_RING_BUFFER_START+1] = ce->ringbuf->vma->node.start;
991 	ce->lrc_reg_state = lrc_reg_state;
992 	ce->state->dirty = true;
993 
994 	/* Invalidate GuC TLB. */
995 	if (i915.enable_guc_submission)
996 		I915_WRITE(GEN8_GTCR, GEN8_GTCR_INVALIDATE);
997 
998 	return 0;
999 
1000 unpin_map:
1001 	i915_gem_object_unpin_map(ce->state);
1002 unpin_ctx_obj:
1003 	i915_gem_object_ggtt_unpin(ce->state);
1004 err:
1005 	ce->pin_count = 0;
1006 	return ret;
1007 }
1008 
1009 void intel_lr_context_unpin(struct i915_gem_context *ctx,
1010 			    struct intel_engine_cs *engine)
1011 {
1012 	struct intel_context *ce = &ctx->engine[engine->id];
1013 
1014 	lockdep_assert_held(&ctx->i915->drm.struct_mutex);
1015 	GEM_BUG_ON(ce->pin_count == 0);
1016 
1017 	if (--ce->pin_count)
1018 		return;
1019 
1020 	intel_unpin_ringbuffer_obj(ce->ringbuf);
1021 
1022 	i915_gem_object_unpin_map(ce->state);
1023 	i915_gem_object_ggtt_unpin(ce->state);
1024 
1025 	ce->lrc_vma = NULL;
1026 	ce->lrc_desc = 0;
1027 	ce->lrc_reg_state = NULL;
1028 
1029 	i915_gem_context_unreference(ctx);
1030 }
1031 
1032 static int intel_logical_ring_workarounds_emit(struct drm_i915_gem_request *req)
1033 {
1034 	int ret, i;
1035 	struct intel_engine_cs *engine = req->engine;
1036 	struct intel_ringbuffer *ringbuf = req->ringbuf;
1037 	struct i915_workarounds *w = &req->i915->workarounds;
1038 
1039 	if (w->count == 0)
1040 		return 0;
1041 
1042 	engine->gpu_caches_dirty = true;
1043 	ret = logical_ring_flush_all_caches(req);
1044 	if (ret)
1045 		return ret;
1046 
1047 	ret = intel_ring_begin(req, w->count * 2 + 2);
1048 	if (ret)
1049 		return ret;
1050 
1051 	intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(w->count));
1052 	for (i = 0; i < w->count; i++) {
1053 		intel_logical_ring_emit_reg(ringbuf, w->reg[i].addr);
1054 		intel_logical_ring_emit(ringbuf, w->reg[i].value);
1055 	}
1056 	intel_logical_ring_emit(ringbuf, MI_NOOP);
1057 
1058 	intel_logical_ring_advance(ringbuf);
1059 
1060 	engine->gpu_caches_dirty = true;
1061 	ret = logical_ring_flush_all_caches(req);
1062 	if (ret)
1063 		return ret;
1064 
1065 	return 0;
1066 }
1067 
1068 #define wa_ctx_emit(batch, index, cmd)					\
1069 	do {								\
1070 		int __index = (index)++;				\
1071 		if (WARN_ON(__index >= (PAGE_SIZE / sizeof(uint32_t)))) { \
1072 			return -ENOSPC;					\
1073 		}							\
1074 		batch[__index] = (cmd);					\
1075 	} while (0)
1076 
1077 #define wa_ctx_emit_reg(batch, index, reg) \
1078 	wa_ctx_emit((batch), (index), i915_mmio_reg_offset(reg))
1079 
1080 /*
1081  * In this WA we need to set GEN8_L3SQCREG4[21:21] and reset it after
1082  * PIPE_CONTROL instruction. This is required for the flush to happen correctly
1083  * but there is a slight complication as this is applied in WA batch where the
1084  * values are only initialized once so we cannot take register value at the
1085  * beginning and reuse it further; hence we save its value to memory, upload a
1086  * constant value with bit21 set and then we restore it back with the saved value.
1087  * To simplify the WA, a constant value is formed by using the default value
1088  * of this register. This shouldn't be a problem because we are only modifying
1089  * it for a short period and this batch in non-premptible. We can ofcourse
1090  * use additional instructions that read the actual value of the register
1091  * at that time and set our bit of interest but it makes the WA complicated.
1092  *
1093  * This WA is also required for Gen9 so extracting as a function avoids
1094  * code duplication.
1095  */
1096 static inline int gen8_emit_flush_coherentl3_wa(struct intel_engine_cs *engine,
1097 						uint32_t *const batch,
1098 						uint32_t index)
1099 {
1100 	struct drm_i915_private *dev_priv = engine->i915;
1101 	uint32_t l3sqc4_flush = (0x40400000 | GEN8_LQSC_FLUSH_COHERENT_LINES);
1102 
1103 	/*
1104 	 * WaDisableLSQCROPERFforOCL:skl,kbl
1105 	 * This WA is implemented in skl_init_clock_gating() but since
1106 	 * this batch updates GEN8_L3SQCREG4 with default value we need to
1107 	 * set this bit here to retain the WA during flush.
1108 	 */
1109 	if (IS_SKL_REVID(dev_priv, 0, SKL_REVID_E0) ||
1110 	    IS_KBL_REVID(dev_priv, 0, KBL_REVID_E0))
1111 		l3sqc4_flush |= GEN8_LQSC_RO_PERF_DIS;
1112 
1113 	wa_ctx_emit(batch, index, (MI_STORE_REGISTER_MEM_GEN8 |
1114 				   MI_SRM_LRM_GLOBAL_GTT));
1115 	wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
1116 	wa_ctx_emit(batch, index, engine->scratch.gtt_offset + 256);
1117 	wa_ctx_emit(batch, index, 0);
1118 
1119 	wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1));
1120 	wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
1121 	wa_ctx_emit(batch, index, l3sqc4_flush);
1122 
1123 	wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
1124 	wa_ctx_emit(batch, index, (PIPE_CONTROL_CS_STALL |
1125 				   PIPE_CONTROL_DC_FLUSH_ENABLE));
1126 	wa_ctx_emit(batch, index, 0);
1127 	wa_ctx_emit(batch, index, 0);
1128 	wa_ctx_emit(batch, index, 0);
1129 	wa_ctx_emit(batch, index, 0);
1130 
1131 	wa_ctx_emit(batch, index, (MI_LOAD_REGISTER_MEM_GEN8 |
1132 				   MI_SRM_LRM_GLOBAL_GTT));
1133 	wa_ctx_emit_reg(batch, index, GEN8_L3SQCREG4);
1134 	wa_ctx_emit(batch, index, engine->scratch.gtt_offset + 256);
1135 	wa_ctx_emit(batch, index, 0);
1136 
1137 	return index;
1138 }
1139 
1140 static inline uint32_t wa_ctx_start(struct i915_wa_ctx_bb *wa_ctx,
1141 				    uint32_t offset,
1142 				    uint32_t start_alignment)
1143 {
1144 	return wa_ctx->offset = ALIGN(offset, start_alignment);
1145 }
1146 
1147 static inline int wa_ctx_end(struct i915_wa_ctx_bb *wa_ctx,
1148 			     uint32_t offset,
1149 			     uint32_t size_alignment)
1150 {
1151 	wa_ctx->size = offset - wa_ctx->offset;
1152 
1153 	WARN(wa_ctx->size % size_alignment,
1154 	     "wa_ctx_bb failed sanity checks: size %d is not aligned to %d\n",
1155 	     wa_ctx->size, size_alignment);
1156 	return 0;
1157 }
1158 
1159 /**
1160  * gen8_init_indirectctx_bb() - initialize indirect ctx batch with WA
1161  *
1162  * @engine: only applicable for RCS
1163  * @wa_ctx: structure representing wa_ctx
1164  *  offset: specifies start of the batch, should be cache-aligned. This is updated
1165  *    with the offset value received as input.
1166  *  size: size of the batch in DWORDS but HW expects in terms of cachelines
1167  * @batch: page in which WA are loaded
1168  * @offset: This field specifies the start of the batch, it should be
1169  *  cache-aligned otherwise it is adjusted accordingly.
1170  *  Typically we only have one indirect_ctx and per_ctx batch buffer which are
1171  *  initialized at the beginning and shared across all contexts but this field
1172  *  helps us to have multiple batches at different offsets and select them based
1173  *  on a criteria. At the moment this batch always start at the beginning of the page
1174  *  and at this point we don't have multiple wa_ctx batch buffers.
1175  *
1176  *  The number of WA applied are not known at the beginning; we use this field
1177  *  to return the no of DWORDS written.
1178  *
1179  *  It is to be noted that this batch does not contain MI_BATCH_BUFFER_END
1180  *  so it adds NOOPs as padding to make it cacheline aligned.
1181  *  MI_BATCH_BUFFER_END will be added to perctx batch and both of them together
1182  *  makes a complete batch buffer.
1183  *
1184  * Return: non-zero if we exceed the PAGE_SIZE limit.
1185  */
1186 
1187 static int gen8_init_indirectctx_bb(struct intel_engine_cs *engine,
1188 				    struct i915_wa_ctx_bb *wa_ctx,
1189 				    uint32_t *const batch,
1190 				    uint32_t *offset)
1191 {
1192 	uint32_t scratch_addr;
1193 	uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1194 
1195 	/* WaDisableCtxRestoreArbitration:bdw,chv */
1196 	wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_DISABLE);
1197 
1198 	/* WaFlushCoherentL3CacheLinesAtContextSwitch:bdw */
1199 	if (IS_BROADWELL(engine->i915)) {
1200 		int rc = gen8_emit_flush_coherentl3_wa(engine, batch, index);
1201 		if (rc < 0)
1202 			return rc;
1203 		index = rc;
1204 	}
1205 
1206 	/* WaClearSlmSpaceAtContextSwitch:bdw,chv */
1207 	/* Actual scratch location is at 128 bytes offset */
1208 	scratch_addr = engine->scratch.gtt_offset + 2*CACHELINE_BYTES;
1209 
1210 	wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
1211 	wa_ctx_emit(batch, index, (PIPE_CONTROL_FLUSH_L3 |
1212 				   PIPE_CONTROL_GLOBAL_GTT_IVB |
1213 				   PIPE_CONTROL_CS_STALL |
1214 				   PIPE_CONTROL_QW_WRITE));
1215 	wa_ctx_emit(batch, index, scratch_addr);
1216 	wa_ctx_emit(batch, index, 0);
1217 	wa_ctx_emit(batch, index, 0);
1218 	wa_ctx_emit(batch, index, 0);
1219 
1220 	/* Pad to end of cacheline */
1221 	while (index % CACHELINE_DWORDS)
1222 		wa_ctx_emit(batch, index, MI_NOOP);
1223 
1224 	/*
1225 	 * MI_BATCH_BUFFER_END is not required in Indirect ctx BB because
1226 	 * execution depends on the length specified in terms of cache lines
1227 	 * in the register CTX_RCS_INDIRECT_CTX
1228 	 */
1229 
1230 	return wa_ctx_end(wa_ctx, *offset = index, CACHELINE_DWORDS);
1231 }
1232 
1233 /**
1234  * gen8_init_perctx_bb() - initialize per ctx batch with WA
1235  *
1236  * @engine: only applicable for RCS
1237  * @wa_ctx: structure representing wa_ctx
1238  *  offset: specifies start of the batch, should be cache-aligned.
1239  *  size: size of the batch in DWORDS but HW expects in terms of cachelines
1240  * @batch: page in which WA are loaded
1241  * @offset: This field specifies the start of this batch.
1242  *   This batch is started immediately after indirect_ctx batch. Since we ensure
1243  *   that indirect_ctx ends on a cacheline this batch is aligned automatically.
1244  *
1245  *   The number of DWORDS written are returned using this field.
1246  *
1247  *  This batch is terminated with MI_BATCH_BUFFER_END and so we need not add padding
1248  *  to align it with cacheline as padding after MI_BATCH_BUFFER_END is redundant.
1249  */
1250 static int gen8_init_perctx_bb(struct intel_engine_cs *engine,
1251 			       struct i915_wa_ctx_bb *wa_ctx,
1252 			       uint32_t *const batch,
1253 			       uint32_t *offset)
1254 {
1255 	uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1256 
1257 	/* WaDisableCtxRestoreArbitration:bdw,chv */
1258 	wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_ENABLE);
1259 
1260 	wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END);
1261 
1262 	return wa_ctx_end(wa_ctx, *offset = index, 1);
1263 }
1264 
1265 static int gen9_init_indirectctx_bb(struct intel_engine_cs *engine,
1266 				    struct i915_wa_ctx_bb *wa_ctx,
1267 				    uint32_t *const batch,
1268 				    uint32_t *offset)
1269 {
1270 	int ret;
1271 	struct drm_i915_private *dev_priv = engine->i915;
1272 	uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1273 
1274 	/* WaDisableCtxRestoreArbitration:skl,bxt */
1275 	if (IS_SKL_REVID(dev_priv, 0, SKL_REVID_D0) ||
1276 	    IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1))
1277 		wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_DISABLE);
1278 
1279 	/* WaFlushCoherentL3CacheLinesAtContextSwitch:skl,bxt */
1280 	ret = gen8_emit_flush_coherentl3_wa(engine, batch, index);
1281 	if (ret < 0)
1282 		return ret;
1283 	index = ret;
1284 
1285 	/* WaClearSlmSpaceAtContextSwitch:kbl */
1286 	/* Actual scratch location is at 128 bytes offset */
1287 	if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_A0)) {
1288 		uint32_t scratch_addr
1289 			= engine->scratch.gtt_offset + 2*CACHELINE_BYTES;
1290 
1291 		wa_ctx_emit(batch, index, GFX_OP_PIPE_CONTROL(6));
1292 		wa_ctx_emit(batch, index, (PIPE_CONTROL_FLUSH_L3 |
1293 					   PIPE_CONTROL_GLOBAL_GTT_IVB |
1294 					   PIPE_CONTROL_CS_STALL |
1295 					   PIPE_CONTROL_QW_WRITE));
1296 		wa_ctx_emit(batch, index, scratch_addr);
1297 		wa_ctx_emit(batch, index, 0);
1298 		wa_ctx_emit(batch, index, 0);
1299 		wa_ctx_emit(batch, index, 0);
1300 	}
1301 
1302 	/* WaMediaPoolStateCmdInWABB:bxt */
1303 	if (HAS_POOLED_EU(engine->i915)) {
1304 		/*
1305 		 * EU pool configuration is setup along with golden context
1306 		 * during context initialization. This value depends on
1307 		 * device type (2x6 or 3x6) and needs to be updated based
1308 		 * on which subslice is disabled especially for 2x6
1309 		 * devices, however it is safe to load default
1310 		 * configuration of 3x6 device instead of masking off
1311 		 * corresponding bits because HW ignores bits of a disabled
1312 		 * subslice and drops down to appropriate config. Please
1313 		 * see render_state_setup() in i915_gem_render_state.c for
1314 		 * possible configurations, to avoid duplication they are
1315 		 * not shown here again.
1316 		 */
1317 		u32 eu_pool_config = 0x00777000;
1318 		wa_ctx_emit(batch, index, GEN9_MEDIA_POOL_STATE);
1319 		wa_ctx_emit(batch, index, GEN9_MEDIA_POOL_ENABLE);
1320 		wa_ctx_emit(batch, index, eu_pool_config);
1321 		wa_ctx_emit(batch, index, 0);
1322 		wa_ctx_emit(batch, index, 0);
1323 		wa_ctx_emit(batch, index, 0);
1324 	}
1325 
1326 	/* Pad to end of cacheline */
1327 	while (index % CACHELINE_DWORDS)
1328 		wa_ctx_emit(batch, index, MI_NOOP);
1329 
1330 	return wa_ctx_end(wa_ctx, *offset = index, CACHELINE_DWORDS);
1331 }
1332 
1333 static int gen9_init_perctx_bb(struct intel_engine_cs *engine,
1334 			       struct i915_wa_ctx_bb *wa_ctx,
1335 			       uint32_t *const batch,
1336 			       uint32_t *offset)
1337 {
1338 	uint32_t index = wa_ctx_start(wa_ctx, *offset, CACHELINE_DWORDS);
1339 
1340 	/* WaSetDisablePixMaskCammingAndRhwoInCommonSliceChicken:skl,bxt */
1341 	if (IS_SKL_REVID(engine->i915, 0, SKL_REVID_B0) ||
1342 	    IS_BXT_REVID(engine->i915, 0, BXT_REVID_A1)) {
1343 		wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(1));
1344 		wa_ctx_emit_reg(batch, index, GEN9_SLICE_COMMON_ECO_CHICKEN0);
1345 		wa_ctx_emit(batch, index,
1346 			    _MASKED_BIT_ENABLE(DISABLE_PIXEL_MASK_CAMMING));
1347 		wa_ctx_emit(batch, index, MI_NOOP);
1348 	}
1349 
1350 	/* WaClearTdlStateAckDirtyBits:bxt */
1351 	if (IS_BXT_REVID(engine->i915, 0, BXT_REVID_B0)) {
1352 		wa_ctx_emit(batch, index, MI_LOAD_REGISTER_IMM(4));
1353 
1354 		wa_ctx_emit_reg(batch, index, GEN8_STATE_ACK);
1355 		wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS));
1356 
1357 		wa_ctx_emit_reg(batch, index, GEN9_STATE_ACK_SLICE1);
1358 		wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS));
1359 
1360 		wa_ctx_emit_reg(batch, index, GEN9_STATE_ACK_SLICE2);
1361 		wa_ctx_emit(batch, index, _MASKED_BIT_DISABLE(GEN9_SUBSLICE_TDL_ACK_BITS));
1362 
1363 		wa_ctx_emit_reg(batch, index, GEN7_ROW_CHICKEN2);
1364 		/* dummy write to CS, mask bits are 0 to ensure the register is not modified */
1365 		wa_ctx_emit(batch, index, 0x0);
1366 		wa_ctx_emit(batch, index, MI_NOOP);
1367 	}
1368 
1369 	/* WaDisableCtxRestoreArbitration:skl,bxt */
1370 	if (IS_SKL_REVID(engine->i915, 0, SKL_REVID_D0) ||
1371 	    IS_BXT_REVID(engine->i915, 0, BXT_REVID_A1))
1372 		wa_ctx_emit(batch, index, MI_ARB_ON_OFF | MI_ARB_ENABLE);
1373 
1374 	wa_ctx_emit(batch, index, MI_BATCH_BUFFER_END);
1375 
1376 	return wa_ctx_end(wa_ctx, *offset = index, 1);
1377 }
1378 
1379 static int lrc_setup_wa_ctx_obj(struct intel_engine_cs *engine, u32 size)
1380 {
1381 	int ret;
1382 
1383 	engine->wa_ctx.obj = i915_gem_object_create(&engine->i915->drm,
1384 						    PAGE_ALIGN(size));
1385 	if (IS_ERR(engine->wa_ctx.obj)) {
1386 		DRM_DEBUG_DRIVER("alloc LRC WA ctx backing obj failed.\n");
1387 		ret = PTR_ERR(engine->wa_ctx.obj);
1388 		engine->wa_ctx.obj = NULL;
1389 		return ret;
1390 	}
1391 
1392 	ret = i915_gem_obj_ggtt_pin(engine->wa_ctx.obj, PAGE_SIZE, 0);
1393 	if (ret) {
1394 		DRM_DEBUG_DRIVER("pin LRC WA ctx backing obj failed: %d\n",
1395 				 ret);
1396 		drm_gem_object_unreference(&engine->wa_ctx.obj->base);
1397 		return ret;
1398 	}
1399 
1400 	return 0;
1401 }
1402 
1403 static void lrc_destroy_wa_ctx_obj(struct intel_engine_cs *engine)
1404 {
1405 	if (engine->wa_ctx.obj) {
1406 		i915_gem_object_ggtt_unpin(engine->wa_ctx.obj);
1407 		drm_gem_object_unreference(&engine->wa_ctx.obj->base);
1408 		engine->wa_ctx.obj = NULL;
1409 	}
1410 }
1411 
1412 static int intel_init_workaround_bb(struct intel_engine_cs *engine)
1413 {
1414 	int ret;
1415 	uint32_t *batch;
1416 	uint32_t offset;
1417 	struct page *page;
1418 	struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx;
1419 
1420 	WARN_ON(engine->id != RCS);
1421 
1422 	/* update this when WA for higher Gen are added */
1423 	if (INTEL_GEN(engine->i915) > 9) {
1424 		DRM_ERROR("WA batch buffer is not initialized for Gen%d\n",
1425 			  INTEL_GEN(engine->i915));
1426 		return 0;
1427 	}
1428 
1429 	/* some WA perform writes to scratch page, ensure it is valid */
1430 	if (engine->scratch.obj == NULL) {
1431 		DRM_ERROR("scratch page not allocated for %s\n", engine->name);
1432 		return -EINVAL;
1433 	}
1434 
1435 	ret = lrc_setup_wa_ctx_obj(engine, PAGE_SIZE);
1436 	if (ret) {
1437 		DRM_DEBUG_DRIVER("Failed to setup context WA page: %d\n", ret);
1438 		return ret;
1439 	}
1440 
1441 	page = i915_gem_object_get_dirty_page(wa_ctx->obj, 0);
1442 	batch = kmap_atomic(page);
1443 	offset = 0;
1444 
1445 	if (IS_GEN8(engine->i915)) {
1446 		ret = gen8_init_indirectctx_bb(engine,
1447 					       &wa_ctx->indirect_ctx,
1448 					       batch,
1449 					       &offset);
1450 		if (ret)
1451 			goto out;
1452 
1453 		ret = gen8_init_perctx_bb(engine,
1454 					  &wa_ctx->per_ctx,
1455 					  batch,
1456 					  &offset);
1457 		if (ret)
1458 			goto out;
1459 	} else if (IS_GEN9(engine->i915)) {
1460 		ret = gen9_init_indirectctx_bb(engine,
1461 					       &wa_ctx->indirect_ctx,
1462 					       batch,
1463 					       &offset);
1464 		if (ret)
1465 			goto out;
1466 
1467 		ret = gen9_init_perctx_bb(engine,
1468 					  &wa_ctx->per_ctx,
1469 					  batch,
1470 					  &offset);
1471 		if (ret)
1472 			goto out;
1473 	}
1474 
1475 out:
1476 	kunmap_atomic(batch);
1477 	if (ret)
1478 		lrc_destroy_wa_ctx_obj(engine);
1479 
1480 	return ret;
1481 }
1482 
1483 static void lrc_init_hws(struct intel_engine_cs *engine)
1484 {
1485 	struct drm_i915_private *dev_priv = engine->i915;
1486 
1487 	I915_WRITE(RING_HWS_PGA(engine->mmio_base),
1488 		   (u32)engine->status_page.gfx_addr);
1489 	POSTING_READ(RING_HWS_PGA(engine->mmio_base));
1490 }
1491 
1492 static int gen8_init_common_ring(struct intel_engine_cs *engine)
1493 {
1494 	struct drm_i915_private *dev_priv = engine->i915;
1495 	unsigned int next_context_status_buffer_hw;
1496 
1497 	lrc_init_hws(engine);
1498 
1499 	I915_WRITE_IMR(engine,
1500 		       ~(engine->irq_enable_mask | engine->irq_keep_mask));
1501 	I915_WRITE(RING_HWSTAM(engine->mmio_base), 0xffffffff);
1502 
1503 	I915_WRITE(RING_MODE_GEN7(engine),
1504 		   _MASKED_BIT_DISABLE(GFX_REPLAY_MODE) |
1505 		   _MASKED_BIT_ENABLE(GFX_RUN_LIST_ENABLE));
1506 	POSTING_READ(RING_MODE_GEN7(engine));
1507 
1508 	/*
1509 	 * Instead of resetting the Context Status Buffer (CSB) read pointer to
1510 	 * zero, we need to read the write pointer from hardware and use its
1511 	 * value because "this register is power context save restored".
1512 	 * Effectively, these states have been observed:
1513 	 *
1514 	 *      | Suspend-to-idle (freeze) | Suspend-to-RAM (mem) |
1515 	 * BDW  | CSB regs not reset       | CSB regs reset       |
1516 	 * CHT  | CSB regs not reset       | CSB regs not reset   |
1517 	 * SKL  |         ?                |         ?            |
1518 	 * BXT  |         ?                |         ?            |
1519 	 */
1520 	next_context_status_buffer_hw =
1521 		GEN8_CSB_WRITE_PTR(I915_READ(RING_CONTEXT_STATUS_PTR(engine)));
1522 
1523 	/*
1524 	 * When the CSB registers are reset (also after power-up / gpu reset),
1525 	 * CSB write pointer is set to all 1's, which is not valid, use '5' in
1526 	 * this special case, so the first element read is CSB[0].
1527 	 */
1528 	if (next_context_status_buffer_hw == GEN8_CSB_PTR_MASK)
1529 		next_context_status_buffer_hw = (GEN8_CSB_ENTRIES - 1);
1530 
1531 	engine->next_context_status_buffer = next_context_status_buffer_hw;
1532 	DRM_DEBUG_DRIVER("Execlists enabled for %s\n", engine->name);
1533 
1534 	intel_engine_init_hangcheck(engine);
1535 
1536 	return intel_mocs_init_engine(engine);
1537 }
1538 
1539 static int gen8_init_render_ring(struct intel_engine_cs *engine)
1540 {
1541 	struct drm_i915_private *dev_priv = engine->i915;
1542 	int ret;
1543 
1544 	ret = gen8_init_common_ring(engine);
1545 	if (ret)
1546 		return ret;
1547 
1548 	/* We need to disable the AsyncFlip performance optimisations in order
1549 	 * to use MI_WAIT_FOR_EVENT within the CS. It should already be
1550 	 * programmed to '1' on all products.
1551 	 *
1552 	 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw,chv
1553 	 */
1554 	I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
1555 
1556 	I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
1557 
1558 	return init_workarounds_ring(engine);
1559 }
1560 
1561 static int gen9_init_render_ring(struct intel_engine_cs *engine)
1562 {
1563 	int ret;
1564 
1565 	ret = gen8_init_common_ring(engine);
1566 	if (ret)
1567 		return ret;
1568 
1569 	return init_workarounds_ring(engine);
1570 }
1571 
1572 static int intel_logical_ring_emit_pdps(struct drm_i915_gem_request *req)
1573 {
1574 	struct i915_hw_ppgtt *ppgtt = req->ctx->ppgtt;
1575 	struct intel_engine_cs *engine = req->engine;
1576 	struct intel_ringbuffer *ringbuf = req->ringbuf;
1577 	const int num_lri_cmds = GEN8_LEGACY_PDPES * 2;
1578 	int i, ret;
1579 
1580 	ret = intel_ring_begin(req, num_lri_cmds * 2 + 2);
1581 	if (ret)
1582 		return ret;
1583 
1584 	intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(num_lri_cmds));
1585 	for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) {
1586 		const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
1587 
1588 		intel_logical_ring_emit_reg(ringbuf,
1589 					    GEN8_RING_PDP_UDW(engine, i));
1590 		intel_logical_ring_emit(ringbuf, upper_32_bits(pd_daddr));
1591 		intel_logical_ring_emit_reg(ringbuf,
1592 					    GEN8_RING_PDP_LDW(engine, i));
1593 		intel_logical_ring_emit(ringbuf, lower_32_bits(pd_daddr));
1594 	}
1595 
1596 	intel_logical_ring_emit(ringbuf, MI_NOOP);
1597 	intel_logical_ring_advance(ringbuf);
1598 
1599 	return 0;
1600 }
1601 
1602 static int gen8_emit_bb_start(struct drm_i915_gem_request *req,
1603 			      u64 offset, unsigned dispatch_flags)
1604 {
1605 	struct intel_ringbuffer *ringbuf = req->ringbuf;
1606 	bool ppgtt = !(dispatch_flags & I915_DISPATCH_SECURE);
1607 	int ret;
1608 
1609 	/* Don't rely in hw updating PDPs, specially in lite-restore.
1610 	 * Ideally, we should set Force PD Restore in ctx descriptor,
1611 	 * but we can't. Force Restore would be a second option, but
1612 	 * it is unsafe in case of lite-restore (because the ctx is
1613 	 * not idle). PML4 is allocated during ppgtt init so this is
1614 	 * not needed in 48-bit.*/
1615 	if (req->ctx->ppgtt &&
1616 	    (intel_engine_flag(req->engine) & req->ctx->ppgtt->pd_dirty_rings)) {
1617 		if (!USES_FULL_48BIT_PPGTT(req->i915) &&
1618 		    !intel_vgpu_active(req->i915)) {
1619 			ret = intel_logical_ring_emit_pdps(req);
1620 			if (ret)
1621 				return ret;
1622 		}
1623 
1624 		req->ctx->ppgtt->pd_dirty_rings &= ~intel_engine_flag(req->engine);
1625 	}
1626 
1627 	ret = intel_ring_begin(req, 4);
1628 	if (ret)
1629 		return ret;
1630 
1631 	/* FIXME(BDW): Address space and security selectors. */
1632 	intel_logical_ring_emit(ringbuf, MI_BATCH_BUFFER_START_GEN8 |
1633 				(ppgtt<<8) |
1634 				(dispatch_flags & I915_DISPATCH_RS ?
1635 				 MI_BATCH_RESOURCE_STREAMER : 0));
1636 	intel_logical_ring_emit(ringbuf, lower_32_bits(offset));
1637 	intel_logical_ring_emit(ringbuf, upper_32_bits(offset));
1638 	intel_logical_ring_emit(ringbuf, MI_NOOP);
1639 	intel_logical_ring_advance(ringbuf);
1640 
1641 	return 0;
1642 }
1643 
1644 static void gen8_logical_ring_enable_irq(struct intel_engine_cs *engine)
1645 {
1646 	struct drm_i915_private *dev_priv = engine->i915;
1647 	I915_WRITE_IMR(engine,
1648 		       ~(engine->irq_enable_mask | engine->irq_keep_mask));
1649 	POSTING_READ_FW(RING_IMR(engine->mmio_base));
1650 }
1651 
1652 static void gen8_logical_ring_disable_irq(struct intel_engine_cs *engine)
1653 {
1654 	struct drm_i915_private *dev_priv = engine->i915;
1655 	I915_WRITE_IMR(engine, ~engine->irq_keep_mask);
1656 }
1657 
1658 static int gen8_emit_flush(struct drm_i915_gem_request *request,
1659 			   u32 invalidate_domains,
1660 			   u32 unused)
1661 {
1662 	struct intel_ringbuffer *ringbuf = request->ringbuf;
1663 	struct intel_engine_cs *engine = ringbuf->engine;
1664 	struct drm_i915_private *dev_priv = request->i915;
1665 	uint32_t cmd;
1666 	int ret;
1667 
1668 	ret = intel_ring_begin(request, 4);
1669 	if (ret)
1670 		return ret;
1671 
1672 	cmd = MI_FLUSH_DW + 1;
1673 
1674 	/* We always require a command barrier so that subsequent
1675 	 * commands, such as breadcrumb interrupts, are strictly ordered
1676 	 * wrt the contents of the write cache being flushed to memory
1677 	 * (and thus being coherent from the CPU).
1678 	 */
1679 	cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1680 
1681 	if (invalidate_domains & I915_GEM_GPU_DOMAINS) {
1682 		cmd |= MI_INVALIDATE_TLB;
1683 		if (engine == &dev_priv->engine[VCS])
1684 			cmd |= MI_INVALIDATE_BSD;
1685 	}
1686 
1687 	intel_logical_ring_emit(ringbuf, cmd);
1688 	intel_logical_ring_emit(ringbuf,
1689 				I915_GEM_HWS_SCRATCH_ADDR |
1690 				MI_FLUSH_DW_USE_GTT);
1691 	intel_logical_ring_emit(ringbuf, 0); /* upper addr */
1692 	intel_logical_ring_emit(ringbuf, 0); /* value */
1693 	intel_logical_ring_advance(ringbuf);
1694 
1695 	return 0;
1696 }
1697 
1698 static int gen8_emit_flush_render(struct drm_i915_gem_request *request,
1699 				  u32 invalidate_domains,
1700 				  u32 flush_domains)
1701 {
1702 	struct intel_ringbuffer *ringbuf = request->ringbuf;
1703 	struct intel_engine_cs *engine = ringbuf->engine;
1704 	u32 scratch_addr = engine->scratch.gtt_offset + 2 * CACHELINE_BYTES;
1705 	bool vf_flush_wa = false, dc_flush_wa = false;
1706 	u32 flags = 0;
1707 	int ret;
1708 	int len;
1709 
1710 	flags |= PIPE_CONTROL_CS_STALL;
1711 
1712 	if (flush_domains) {
1713 		flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
1714 		flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
1715 		flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
1716 		flags |= PIPE_CONTROL_FLUSH_ENABLE;
1717 	}
1718 
1719 	if (invalidate_domains) {
1720 		flags |= PIPE_CONTROL_TLB_INVALIDATE;
1721 		flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
1722 		flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1723 		flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
1724 		flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
1725 		flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
1726 		flags |= PIPE_CONTROL_QW_WRITE;
1727 		flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
1728 
1729 		/*
1730 		 * On GEN9: before VF_CACHE_INVALIDATE we need to emit a NULL
1731 		 * pipe control.
1732 		 */
1733 		if (IS_GEN9(request->i915))
1734 			vf_flush_wa = true;
1735 
1736 		/* WaForGAMHang:kbl */
1737 		if (IS_KBL_REVID(request->i915, 0, KBL_REVID_B0))
1738 			dc_flush_wa = true;
1739 	}
1740 
1741 	len = 6;
1742 
1743 	if (vf_flush_wa)
1744 		len += 6;
1745 
1746 	if (dc_flush_wa)
1747 		len += 12;
1748 
1749 	ret = intel_ring_begin(request, len);
1750 	if (ret)
1751 		return ret;
1752 
1753 	if (vf_flush_wa) {
1754 		intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1755 		intel_logical_ring_emit(ringbuf, 0);
1756 		intel_logical_ring_emit(ringbuf, 0);
1757 		intel_logical_ring_emit(ringbuf, 0);
1758 		intel_logical_ring_emit(ringbuf, 0);
1759 		intel_logical_ring_emit(ringbuf, 0);
1760 	}
1761 
1762 	if (dc_flush_wa) {
1763 		intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1764 		intel_logical_ring_emit(ringbuf, PIPE_CONTROL_DC_FLUSH_ENABLE);
1765 		intel_logical_ring_emit(ringbuf, 0);
1766 		intel_logical_ring_emit(ringbuf, 0);
1767 		intel_logical_ring_emit(ringbuf, 0);
1768 		intel_logical_ring_emit(ringbuf, 0);
1769 	}
1770 
1771 	intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1772 	intel_logical_ring_emit(ringbuf, flags);
1773 	intel_logical_ring_emit(ringbuf, scratch_addr);
1774 	intel_logical_ring_emit(ringbuf, 0);
1775 	intel_logical_ring_emit(ringbuf, 0);
1776 	intel_logical_ring_emit(ringbuf, 0);
1777 
1778 	if (dc_flush_wa) {
1779 		intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1780 		intel_logical_ring_emit(ringbuf, PIPE_CONTROL_CS_STALL);
1781 		intel_logical_ring_emit(ringbuf, 0);
1782 		intel_logical_ring_emit(ringbuf, 0);
1783 		intel_logical_ring_emit(ringbuf, 0);
1784 		intel_logical_ring_emit(ringbuf, 0);
1785 	}
1786 
1787 	intel_logical_ring_advance(ringbuf);
1788 
1789 	return 0;
1790 }
1791 
1792 static void bxt_a_seqno_barrier(struct intel_engine_cs *engine)
1793 {
1794 	/*
1795 	 * On BXT A steppings there is a HW coherency issue whereby the
1796 	 * MI_STORE_DATA_IMM storing the completed request's seqno
1797 	 * occasionally doesn't invalidate the CPU cache. Work around this by
1798 	 * clflushing the corresponding cacheline whenever the caller wants
1799 	 * the coherency to be guaranteed. Note that this cacheline is known
1800 	 * to be clean at this point, since we only write it in
1801 	 * bxt_a_set_seqno(), where we also do a clflush after the write. So
1802 	 * this clflush in practice becomes an invalidate operation.
1803 	 */
1804 	intel_flush_status_page(engine, I915_GEM_HWS_INDEX);
1805 }
1806 
1807 /*
1808  * Reserve space for 2 NOOPs at the end of each request to be
1809  * used as a workaround for not being allowed to do lite
1810  * restore with HEAD==TAIL (WaIdleLiteRestore).
1811  */
1812 #define WA_TAIL_DWORDS 2
1813 
1814 static int gen8_emit_request(struct drm_i915_gem_request *request)
1815 {
1816 	struct intel_ringbuffer *ringbuf = request->ringbuf;
1817 	int ret;
1818 
1819 	ret = intel_ring_begin(request, 6 + WA_TAIL_DWORDS);
1820 	if (ret)
1821 		return ret;
1822 
1823 	/* w/a: bit 5 needs to be zero for MI_FLUSH_DW address. */
1824 	BUILD_BUG_ON(I915_GEM_HWS_INDEX_ADDR & (1 << 5));
1825 
1826 	intel_logical_ring_emit(ringbuf,
1827 				(MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW);
1828 	intel_logical_ring_emit(ringbuf,
1829 				intel_hws_seqno_address(request->engine) |
1830 				MI_FLUSH_DW_USE_GTT);
1831 	intel_logical_ring_emit(ringbuf, 0);
1832 	intel_logical_ring_emit(ringbuf, request->seqno);
1833 	intel_logical_ring_emit(ringbuf, MI_USER_INTERRUPT);
1834 	intel_logical_ring_emit(ringbuf, MI_NOOP);
1835 	return intel_logical_ring_advance_and_submit(request);
1836 }
1837 
1838 static int gen8_emit_request_render(struct drm_i915_gem_request *request)
1839 {
1840 	struct intel_ringbuffer *ringbuf = request->ringbuf;
1841 	int ret;
1842 
1843 	ret = intel_ring_begin(request, 8 + WA_TAIL_DWORDS);
1844 	if (ret)
1845 		return ret;
1846 
1847 	/* We're using qword write, seqno should be aligned to 8 bytes. */
1848 	BUILD_BUG_ON(I915_GEM_HWS_INDEX & 1);
1849 
1850 	/* w/a for post sync ops following a GPGPU operation we
1851 	 * need a prior CS_STALL, which is emitted by the flush
1852 	 * following the batch.
1853 	 */
1854 	intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1855 	intel_logical_ring_emit(ringbuf,
1856 				(PIPE_CONTROL_GLOBAL_GTT_IVB |
1857 				 PIPE_CONTROL_CS_STALL |
1858 				 PIPE_CONTROL_QW_WRITE));
1859 	intel_logical_ring_emit(ringbuf,
1860 				intel_hws_seqno_address(request->engine));
1861 	intel_logical_ring_emit(ringbuf, 0);
1862 	intel_logical_ring_emit(ringbuf, i915_gem_request_get_seqno(request));
1863 	/* We're thrashing one dword of HWS. */
1864 	intel_logical_ring_emit(ringbuf, 0);
1865 	intel_logical_ring_emit(ringbuf, MI_USER_INTERRUPT);
1866 	intel_logical_ring_emit(ringbuf, MI_NOOP);
1867 	return intel_logical_ring_advance_and_submit(request);
1868 }
1869 
1870 static int intel_lr_context_render_state_init(struct drm_i915_gem_request *req)
1871 {
1872 	struct render_state so;
1873 	int ret;
1874 
1875 	ret = i915_gem_render_state_prepare(req->engine, &so);
1876 	if (ret)
1877 		return ret;
1878 
1879 	if (so.rodata == NULL)
1880 		return 0;
1881 
1882 	ret = req->engine->emit_bb_start(req, so.ggtt_offset,
1883 				       I915_DISPATCH_SECURE);
1884 	if (ret)
1885 		goto out;
1886 
1887 	ret = req->engine->emit_bb_start(req,
1888 				       (so.ggtt_offset + so.aux_batch_offset),
1889 				       I915_DISPATCH_SECURE);
1890 	if (ret)
1891 		goto out;
1892 
1893 	i915_vma_move_to_active(i915_gem_obj_to_ggtt(so.obj), req);
1894 
1895 out:
1896 	i915_gem_render_state_fini(&so);
1897 	return ret;
1898 }
1899 
1900 static int gen8_init_rcs_context(struct drm_i915_gem_request *req)
1901 {
1902 	int ret;
1903 
1904 	ret = intel_logical_ring_workarounds_emit(req);
1905 	if (ret)
1906 		return ret;
1907 
1908 	ret = intel_rcs_context_init_mocs(req);
1909 	/*
1910 	 * Failing to program the MOCS is non-fatal.The system will not
1911 	 * run at peak performance. So generate an error and carry on.
1912 	 */
1913 	if (ret)
1914 		DRM_ERROR("MOCS failed to program: expect performance issues.\n");
1915 
1916 	return intel_lr_context_render_state_init(req);
1917 }
1918 
1919 /**
1920  * intel_logical_ring_cleanup() - deallocate the Engine Command Streamer
1921  *
1922  * @engine: Engine Command Streamer.
1923  *
1924  */
1925 void intel_logical_ring_cleanup(struct intel_engine_cs *engine)
1926 {
1927 	struct drm_i915_private *dev_priv;
1928 
1929 	if (!intel_engine_initialized(engine))
1930 		return;
1931 
1932 	/*
1933 	 * Tasklet cannot be active at this point due intel_mark_active/idle
1934 	 * so this is just for documentation.
1935 	 */
1936 	if (WARN_ON(test_bit(TASKLET_STATE_SCHED, &engine->irq_tasklet.state)))
1937 		tasklet_kill(&engine->irq_tasklet);
1938 
1939 	dev_priv = engine->i915;
1940 
1941 	if (engine->buffer) {
1942 		intel_logical_ring_stop(engine);
1943 		WARN_ON((I915_READ_MODE(engine) & MODE_IDLE) == 0);
1944 	}
1945 
1946 	if (engine->cleanup)
1947 		engine->cleanup(engine);
1948 
1949 	i915_cmd_parser_fini_ring(engine);
1950 	i915_gem_batch_pool_fini(&engine->batch_pool);
1951 
1952 	intel_engine_fini_breadcrumbs(engine);
1953 
1954 	if (engine->status_page.obj) {
1955 		i915_gem_object_unpin_map(engine->status_page.obj);
1956 		engine->status_page.obj = NULL;
1957 	}
1958 	intel_lr_context_unpin(dev_priv->kernel_context, engine);
1959 
1960 	engine->idle_lite_restore_wa = 0;
1961 	engine->disable_lite_restore_wa = false;
1962 	engine->ctx_desc_template = 0;
1963 
1964 	lrc_destroy_wa_ctx_obj(engine);
1965 	engine->i915 = NULL;
1966 }
1967 
1968 static void
1969 logical_ring_default_vfuncs(struct intel_engine_cs *engine)
1970 {
1971 	/* Default vfuncs which can be overriden by each engine. */
1972 	engine->init_hw = gen8_init_common_ring;
1973 	engine->emit_request = gen8_emit_request;
1974 	engine->emit_flush = gen8_emit_flush;
1975 	engine->irq_enable = gen8_logical_ring_enable_irq;
1976 	engine->irq_disable = gen8_logical_ring_disable_irq;
1977 	engine->emit_bb_start = gen8_emit_bb_start;
1978 	if (IS_BXT_REVID(engine->i915, 0, BXT_REVID_A1))
1979 		engine->irq_seqno_barrier = bxt_a_seqno_barrier;
1980 }
1981 
1982 static inline void
1983 logical_ring_default_irqs(struct intel_engine_cs *engine, unsigned shift)
1984 {
1985 	engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT << shift;
1986 	engine->irq_keep_mask = GT_CONTEXT_SWITCH_INTERRUPT << shift;
1987 }
1988 
1989 static int
1990 lrc_setup_hws(struct intel_engine_cs *engine,
1991 	      struct drm_i915_gem_object *dctx_obj)
1992 {
1993 	void *hws;
1994 
1995 	/* The HWSP is part of the default context object in LRC mode. */
1996 	engine->status_page.gfx_addr = i915_gem_obj_ggtt_offset(dctx_obj) +
1997 				       LRC_PPHWSP_PN * PAGE_SIZE;
1998 	hws = i915_gem_object_pin_map(dctx_obj);
1999 	if (IS_ERR(hws))
2000 		return PTR_ERR(hws);
2001 	engine->status_page.page_addr = hws + LRC_PPHWSP_PN * PAGE_SIZE;
2002 	engine->status_page.obj = dctx_obj;
2003 
2004 	return 0;
2005 }
2006 
2007 static int
2008 logical_ring_init(struct intel_engine_cs *engine)
2009 {
2010 	struct i915_gem_context *dctx = engine->i915->kernel_context;
2011 	int ret;
2012 
2013 	ret = intel_engine_init_breadcrumbs(engine);
2014 	if (ret)
2015 		goto error;
2016 
2017 	ret = i915_cmd_parser_init_ring(engine);
2018 	if (ret)
2019 		goto error;
2020 
2021 	ret = execlists_context_deferred_alloc(dctx, engine);
2022 	if (ret)
2023 		goto error;
2024 
2025 	/* As this is the default context, always pin it */
2026 	ret = intel_lr_context_pin(dctx, engine);
2027 	if (ret) {
2028 		DRM_ERROR("Failed to pin context for %s: %d\n",
2029 			  engine->name, ret);
2030 		goto error;
2031 	}
2032 
2033 	/* And setup the hardware status page. */
2034 	ret = lrc_setup_hws(engine, dctx->engine[engine->id].state);
2035 	if (ret) {
2036 		DRM_ERROR("Failed to set up hws %s: %d\n", engine->name, ret);
2037 		goto error;
2038 	}
2039 
2040 	return 0;
2041 
2042 error:
2043 	intel_logical_ring_cleanup(engine);
2044 	return ret;
2045 }
2046 
2047 static int logical_render_ring_init(struct intel_engine_cs *engine)
2048 {
2049 	struct drm_i915_private *dev_priv = engine->i915;
2050 	int ret;
2051 
2052 	if (HAS_L3_DPF(dev_priv))
2053 		engine->irq_keep_mask |= GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
2054 
2055 	/* Override some for render ring. */
2056 	if (INTEL_GEN(dev_priv) >= 9)
2057 		engine->init_hw = gen9_init_render_ring;
2058 	else
2059 		engine->init_hw = gen8_init_render_ring;
2060 	engine->init_context = gen8_init_rcs_context;
2061 	engine->cleanup = intel_fini_pipe_control;
2062 	engine->emit_flush = gen8_emit_flush_render;
2063 	engine->emit_request = gen8_emit_request_render;
2064 
2065 	ret = intel_init_pipe_control(engine, 4096);
2066 	if (ret)
2067 		return ret;
2068 
2069 	ret = intel_init_workaround_bb(engine);
2070 	if (ret) {
2071 		/*
2072 		 * We continue even if we fail to initialize WA batch
2073 		 * because we only expect rare glitches but nothing
2074 		 * critical to prevent us from using GPU
2075 		 */
2076 		DRM_ERROR("WA batch buffer initialization failed: %d\n",
2077 			  ret);
2078 	}
2079 
2080 	ret = logical_ring_init(engine);
2081 	if (ret) {
2082 		lrc_destroy_wa_ctx_obj(engine);
2083 	}
2084 
2085 	return ret;
2086 }
2087 
2088 static const struct logical_ring_info {
2089 	const char *name;
2090 	unsigned exec_id;
2091 	unsigned guc_id;
2092 	u32 mmio_base;
2093 	unsigned irq_shift;
2094 	int (*init)(struct intel_engine_cs *engine);
2095 } logical_rings[] = {
2096 	[RCS] = {
2097 		.name = "render ring",
2098 		.exec_id = I915_EXEC_RENDER,
2099 		.guc_id = GUC_RENDER_ENGINE,
2100 		.mmio_base = RENDER_RING_BASE,
2101 		.irq_shift = GEN8_RCS_IRQ_SHIFT,
2102 		.init = logical_render_ring_init,
2103 	},
2104 	[BCS] = {
2105 		.name = "blitter ring",
2106 		.exec_id = I915_EXEC_BLT,
2107 		.guc_id = GUC_BLITTER_ENGINE,
2108 		.mmio_base = BLT_RING_BASE,
2109 		.irq_shift = GEN8_BCS_IRQ_SHIFT,
2110 		.init = logical_ring_init,
2111 	},
2112 	[VCS] = {
2113 		.name = "bsd ring",
2114 		.exec_id = I915_EXEC_BSD,
2115 		.guc_id = GUC_VIDEO_ENGINE,
2116 		.mmio_base = GEN6_BSD_RING_BASE,
2117 		.irq_shift = GEN8_VCS1_IRQ_SHIFT,
2118 		.init = logical_ring_init,
2119 	},
2120 	[VCS2] = {
2121 		.name = "bsd2 ring",
2122 		.exec_id = I915_EXEC_BSD,
2123 		.guc_id = GUC_VIDEO_ENGINE2,
2124 		.mmio_base = GEN8_BSD2_RING_BASE,
2125 		.irq_shift = GEN8_VCS2_IRQ_SHIFT,
2126 		.init = logical_ring_init,
2127 	},
2128 	[VECS] = {
2129 		.name = "video enhancement ring",
2130 		.exec_id = I915_EXEC_VEBOX,
2131 		.guc_id = GUC_VIDEOENHANCE_ENGINE,
2132 		.mmio_base = VEBOX_RING_BASE,
2133 		.irq_shift = GEN8_VECS_IRQ_SHIFT,
2134 		.init = logical_ring_init,
2135 	},
2136 };
2137 
2138 static struct intel_engine_cs *
2139 logical_ring_setup(struct drm_i915_private *dev_priv, enum intel_engine_id id)
2140 {
2141 	const struct logical_ring_info *info = &logical_rings[id];
2142 	struct intel_engine_cs *engine = &dev_priv->engine[id];
2143 	enum forcewake_domains fw_domains;
2144 
2145 	engine->id = id;
2146 	engine->name = info->name;
2147 	engine->exec_id = info->exec_id;
2148 	engine->guc_id = info->guc_id;
2149 	engine->mmio_base = info->mmio_base;
2150 
2151 	engine->i915 = dev_priv;
2152 
2153 	/* Intentionally left blank. */
2154 	engine->buffer = NULL;
2155 
2156 	fw_domains = intel_uncore_forcewake_for_reg(dev_priv,
2157 						    RING_ELSP(engine),
2158 						    FW_REG_WRITE);
2159 
2160 	fw_domains |= intel_uncore_forcewake_for_reg(dev_priv,
2161 						     RING_CONTEXT_STATUS_PTR(engine),
2162 						     FW_REG_READ | FW_REG_WRITE);
2163 
2164 	fw_domains |= intel_uncore_forcewake_for_reg(dev_priv,
2165 						     RING_CONTEXT_STATUS_BUF_BASE(engine),
2166 						     FW_REG_READ);
2167 
2168 	engine->fw_domains = fw_domains;
2169 
2170 	INIT_LIST_HEAD(&engine->active_list);
2171 	INIT_LIST_HEAD(&engine->request_list);
2172 	INIT_LIST_HEAD(&engine->buffers);
2173 	INIT_LIST_HEAD(&engine->execlist_queue);
2174 	lockinit(&engine->execlist_lock, "i915el", 0, LK_CANRECURSE);
2175 
2176 	tasklet_init(&engine->irq_tasklet,
2177 		     intel_lrc_irq_handler, (unsigned long)engine);
2178 
2179 	logical_ring_init_platform_invariants(engine);
2180 	logical_ring_default_vfuncs(engine);
2181 	logical_ring_default_irqs(engine, info->irq_shift);
2182 
2183 	intel_engine_init_hangcheck(engine);
2184 	i915_gem_batch_pool_init(&dev_priv->drm, &engine->batch_pool);
2185 
2186 	return engine;
2187 }
2188 
2189 /**
2190  * intel_logical_rings_init() - allocate, populate and init the Engine Command Streamers
2191  * @dev: DRM device.
2192  *
2193  * This function inits the engines for an Execlists submission style (the
2194  * equivalent in the legacy ringbuffer submission world would be
2195  * i915_gem_init_engines). It does it only for those engines that are present in
2196  * the hardware.
2197  *
2198  * Return: non-zero if the initialization failed.
2199  */
2200 int intel_logical_rings_init(struct drm_device *dev)
2201 {
2202 	struct drm_i915_private *dev_priv = to_i915(dev);
2203 	unsigned int mask = 0;
2204 	unsigned int i;
2205 	int ret;
2206 
2207 	WARN_ON(INTEL_INFO(dev_priv)->ring_mask &
2208 		GENMASK(sizeof(mask) * BITS_PER_BYTE - 1, I915_NUM_ENGINES));
2209 
2210 	for (i = 0; i < ARRAY_SIZE(logical_rings); i++) {
2211 		if (!HAS_ENGINE(dev_priv, i))
2212 			continue;
2213 
2214 		if (!logical_rings[i].init)
2215 			continue;
2216 
2217 		ret = logical_rings[i].init(logical_ring_setup(dev_priv, i));
2218 		if (ret)
2219 			goto cleanup;
2220 
2221 		mask |= ENGINE_MASK(i);
2222 	}
2223 
2224 	/*
2225 	 * Catch failures to update logical_rings table when the new engines
2226 	 * are added to the driver by a warning and disabling the forgotten
2227 	 * engines.
2228 	 */
2229 	if (WARN_ON(mask != INTEL_INFO(dev_priv)->ring_mask)) {
2230 		struct intel_device_info *info =
2231 			(struct intel_device_info *)&dev_priv->info;
2232 		info->ring_mask = mask;
2233 	}
2234 
2235 	return 0;
2236 
2237 cleanup:
2238 	for (i = 0; i < I915_NUM_ENGINES; i++)
2239 		intel_logical_ring_cleanup(&dev_priv->engine[i]);
2240 
2241 	return ret;
2242 }
2243 
2244 static u32
2245 make_rpcs(struct drm_i915_private *dev_priv)
2246 {
2247 	u32 rpcs = 0;
2248 
2249 	/*
2250 	 * No explicit RPCS request is needed to ensure full
2251 	 * slice/subslice/EU enablement prior to Gen9.
2252 	*/
2253 	if (INTEL_GEN(dev_priv) < 9)
2254 		return 0;
2255 
2256 	/*
2257 	 * Starting in Gen9, render power gating can leave
2258 	 * slice/subslice/EU in a partially enabled state. We
2259 	 * must make an explicit request through RPCS for full
2260 	 * enablement.
2261 	*/
2262 	if (INTEL_INFO(dev_priv)->has_slice_pg) {
2263 		rpcs |= GEN8_RPCS_S_CNT_ENABLE;
2264 		rpcs |= INTEL_INFO(dev_priv)->slice_total <<
2265 			GEN8_RPCS_S_CNT_SHIFT;
2266 		rpcs |= GEN8_RPCS_ENABLE;
2267 	}
2268 
2269 	if (INTEL_INFO(dev_priv)->has_subslice_pg) {
2270 		rpcs |= GEN8_RPCS_SS_CNT_ENABLE;
2271 		rpcs |= INTEL_INFO(dev_priv)->subslice_per_slice <<
2272 			GEN8_RPCS_SS_CNT_SHIFT;
2273 		rpcs |= GEN8_RPCS_ENABLE;
2274 	}
2275 
2276 	if (INTEL_INFO(dev_priv)->has_eu_pg) {
2277 		rpcs |= INTEL_INFO(dev_priv)->eu_per_subslice <<
2278 			GEN8_RPCS_EU_MIN_SHIFT;
2279 		rpcs |= INTEL_INFO(dev_priv)->eu_per_subslice <<
2280 			GEN8_RPCS_EU_MAX_SHIFT;
2281 		rpcs |= GEN8_RPCS_ENABLE;
2282 	}
2283 
2284 	return rpcs;
2285 }
2286 
2287 static u32 intel_lr_indirect_ctx_offset(struct intel_engine_cs *engine)
2288 {
2289 	u32 indirect_ctx_offset;
2290 
2291 	switch (INTEL_GEN(engine->i915)) {
2292 	default:
2293 		MISSING_CASE(INTEL_GEN(engine->i915));
2294 		/* fall through */
2295 	case 9:
2296 		indirect_ctx_offset =
2297 			GEN9_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
2298 		break;
2299 	case 8:
2300 		indirect_ctx_offset =
2301 			GEN8_CTX_RCS_INDIRECT_CTX_OFFSET_DEFAULT;
2302 		break;
2303 	}
2304 
2305 	return indirect_ctx_offset;
2306 }
2307 
2308 static int
2309 populate_lr_context(struct i915_gem_context *ctx,
2310 		    struct drm_i915_gem_object *ctx_obj,
2311 		    struct intel_engine_cs *engine,
2312 		    struct intel_ringbuffer *ringbuf)
2313 {
2314 	struct drm_i915_private *dev_priv = ctx->i915;
2315 	struct i915_hw_ppgtt *ppgtt = ctx->ppgtt;
2316 	void *vaddr;
2317 	u32 *reg_state;
2318 	int ret;
2319 
2320 	if (!ppgtt)
2321 		ppgtt = dev_priv->mm.aliasing_ppgtt;
2322 
2323 	ret = i915_gem_object_set_to_cpu_domain(ctx_obj, true);
2324 	if (ret) {
2325 		DRM_DEBUG_DRIVER("Could not set to CPU domain\n");
2326 		return ret;
2327 	}
2328 
2329 	vaddr = i915_gem_object_pin_map(ctx_obj);
2330 	if (IS_ERR(vaddr)) {
2331 		ret = PTR_ERR(vaddr);
2332 		DRM_DEBUG_DRIVER("Could not map object pages! (%d)\n", ret);
2333 		return ret;
2334 	}
2335 	ctx_obj->dirty = true;
2336 
2337 	/* The second page of the context object contains some fields which must
2338 	 * be set up prior to the first execution. */
2339 	reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
2340 
2341 	/* A context is actually a big batch buffer with several MI_LOAD_REGISTER_IMM
2342 	 * commands followed by (reg, value) pairs. The values we are setting here are
2343 	 * only for the first context restore: on a subsequent save, the GPU will
2344 	 * recreate this batchbuffer with new values (including all the missing
2345 	 * MI_LOAD_REGISTER_IMM commands that we are not initializing here). */
2346 	reg_state[CTX_LRI_HEADER_0] =
2347 		MI_LOAD_REGISTER_IMM(engine->id == RCS ? 14 : 11) | MI_LRI_FORCE_POSTED;
2348 	ASSIGN_CTX_REG(reg_state, CTX_CONTEXT_CONTROL,
2349 		       RING_CONTEXT_CONTROL(engine),
2350 		       _MASKED_BIT_ENABLE(CTX_CTRL_INHIBIT_SYN_CTX_SWITCH |
2351 					  CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT |
2352 					  (HAS_RESOURCE_STREAMER(dev_priv) ?
2353 					    CTX_CTRL_RS_CTX_ENABLE : 0)));
2354 	ASSIGN_CTX_REG(reg_state, CTX_RING_HEAD, RING_HEAD(engine->mmio_base),
2355 		       0);
2356 	ASSIGN_CTX_REG(reg_state, CTX_RING_TAIL, RING_TAIL(engine->mmio_base),
2357 		       0);
2358 	/* Ring buffer start address is not known until the buffer is pinned.
2359 	 * It is written to the context image in execlists_update_context()
2360 	 */
2361 	ASSIGN_CTX_REG(reg_state, CTX_RING_BUFFER_START,
2362 		       RING_START(engine->mmio_base), 0);
2363 	ASSIGN_CTX_REG(reg_state, CTX_RING_BUFFER_CONTROL,
2364 		       RING_CTL(engine->mmio_base),
2365 		       ((ringbuf->size - PAGE_SIZE) & RING_NR_PAGES) | RING_VALID);
2366 	ASSIGN_CTX_REG(reg_state, CTX_BB_HEAD_U,
2367 		       RING_BBADDR_UDW(engine->mmio_base), 0);
2368 	ASSIGN_CTX_REG(reg_state, CTX_BB_HEAD_L,
2369 		       RING_BBADDR(engine->mmio_base), 0);
2370 	ASSIGN_CTX_REG(reg_state, CTX_BB_STATE,
2371 		       RING_BBSTATE(engine->mmio_base),
2372 		       RING_BB_PPGTT);
2373 	ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_HEAD_U,
2374 		       RING_SBBADDR_UDW(engine->mmio_base), 0);
2375 	ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_HEAD_L,
2376 		       RING_SBBADDR(engine->mmio_base), 0);
2377 	ASSIGN_CTX_REG(reg_state, CTX_SECOND_BB_STATE,
2378 		       RING_SBBSTATE(engine->mmio_base), 0);
2379 	if (engine->id == RCS) {
2380 		ASSIGN_CTX_REG(reg_state, CTX_BB_PER_CTX_PTR,
2381 			       RING_BB_PER_CTX_PTR(engine->mmio_base), 0);
2382 		ASSIGN_CTX_REG(reg_state, CTX_RCS_INDIRECT_CTX,
2383 			       RING_INDIRECT_CTX(engine->mmio_base), 0);
2384 		ASSIGN_CTX_REG(reg_state, CTX_RCS_INDIRECT_CTX_OFFSET,
2385 			       RING_INDIRECT_CTX_OFFSET(engine->mmio_base), 0);
2386 		if (engine->wa_ctx.obj) {
2387 			struct i915_ctx_workarounds *wa_ctx = &engine->wa_ctx;
2388 			uint32_t ggtt_offset = i915_gem_obj_ggtt_offset(wa_ctx->obj);
2389 
2390 			reg_state[CTX_RCS_INDIRECT_CTX+1] =
2391 				(ggtt_offset + wa_ctx->indirect_ctx.offset * sizeof(uint32_t)) |
2392 				(wa_ctx->indirect_ctx.size / CACHELINE_DWORDS);
2393 
2394 			reg_state[CTX_RCS_INDIRECT_CTX_OFFSET+1] =
2395 				intel_lr_indirect_ctx_offset(engine) << 6;
2396 
2397 			reg_state[CTX_BB_PER_CTX_PTR+1] =
2398 				(ggtt_offset + wa_ctx->per_ctx.offset * sizeof(uint32_t)) |
2399 				0x01;
2400 		}
2401 	}
2402 	reg_state[CTX_LRI_HEADER_1] = MI_LOAD_REGISTER_IMM(9) | MI_LRI_FORCE_POSTED;
2403 	ASSIGN_CTX_REG(reg_state, CTX_CTX_TIMESTAMP,
2404 		       RING_CTX_TIMESTAMP(engine->mmio_base), 0);
2405 	/* PDP values well be assigned later if needed */
2406 	ASSIGN_CTX_REG(reg_state, CTX_PDP3_UDW, GEN8_RING_PDP_UDW(engine, 3),
2407 		       0);
2408 	ASSIGN_CTX_REG(reg_state, CTX_PDP3_LDW, GEN8_RING_PDP_LDW(engine, 3),
2409 		       0);
2410 	ASSIGN_CTX_REG(reg_state, CTX_PDP2_UDW, GEN8_RING_PDP_UDW(engine, 2),
2411 		       0);
2412 	ASSIGN_CTX_REG(reg_state, CTX_PDP2_LDW, GEN8_RING_PDP_LDW(engine, 2),
2413 		       0);
2414 	ASSIGN_CTX_REG(reg_state, CTX_PDP1_UDW, GEN8_RING_PDP_UDW(engine, 1),
2415 		       0);
2416 	ASSIGN_CTX_REG(reg_state, CTX_PDP1_LDW, GEN8_RING_PDP_LDW(engine, 1),
2417 		       0);
2418 	ASSIGN_CTX_REG(reg_state, CTX_PDP0_UDW, GEN8_RING_PDP_UDW(engine, 0),
2419 		       0);
2420 	ASSIGN_CTX_REG(reg_state, CTX_PDP0_LDW, GEN8_RING_PDP_LDW(engine, 0),
2421 		       0);
2422 
2423 	if (USES_FULL_48BIT_PPGTT(ppgtt->base.dev)) {
2424 		/* 64b PPGTT (48bit canonical)
2425 		 * PDP0_DESCRIPTOR contains the base address to PML4 and
2426 		 * other PDP Descriptors are ignored.
2427 		 */
2428 		ASSIGN_CTX_PML4(ppgtt, reg_state);
2429 	} else {
2430 		/* 32b PPGTT
2431 		 * PDP*_DESCRIPTOR contains the base address of space supported.
2432 		 * With dynamic page allocation, PDPs may not be allocated at
2433 		 * this point. Point the unallocated PDPs to the scratch page
2434 		 */
2435 		execlists_update_context_pdps(ppgtt, reg_state);
2436 	}
2437 
2438 	if (engine->id == RCS) {
2439 		reg_state[CTX_LRI_HEADER_2] = MI_LOAD_REGISTER_IMM(1);
2440 		ASSIGN_CTX_REG(reg_state, CTX_R_PWR_CLK_STATE, GEN8_R_PWR_CLK_STATE,
2441 			       make_rpcs(dev_priv));
2442 	}
2443 
2444 	i915_gem_object_unpin_map(ctx_obj);
2445 
2446 	return 0;
2447 }
2448 
2449 /**
2450  * intel_lr_context_size() - return the size of the context for an engine
2451  * @engine: which engine to find the context size for
2452  *
2453  * Each engine may require a different amount of space for a context image,
2454  * so when allocating (or copying) an image, this function can be used to
2455  * find the right size for the specific engine.
2456  *
2457  * Return: size (in bytes) of an engine-specific context image
2458  *
2459  * Note: this size includes the HWSP, which is part of the context image
2460  * in LRC mode, but does not include the "shared data page" used with
2461  * GuC submission. The caller should account for this if using the GuC.
2462  */
2463 uint32_t intel_lr_context_size(struct intel_engine_cs *engine)
2464 {
2465 	int ret = 0;
2466 
2467 	WARN_ON(INTEL_GEN(engine->i915) < 8);
2468 
2469 	switch (engine->id) {
2470 	case RCS:
2471 		if (INTEL_GEN(engine->i915) >= 9)
2472 			ret = GEN9_LR_CONTEXT_RENDER_SIZE;
2473 		else
2474 			ret = GEN8_LR_CONTEXT_RENDER_SIZE;
2475 		break;
2476 	case VCS:
2477 	case BCS:
2478 	case VECS:
2479 	case VCS2:
2480 		ret = GEN8_LR_CONTEXT_OTHER_SIZE;
2481 		break;
2482 	}
2483 
2484 	return ret;
2485 }
2486 
2487 /**
2488  * execlists_context_deferred_alloc() - create the LRC specific bits of a context
2489  * @ctx: LR context to create.
2490  * @engine: engine to be used with the context.
2491  *
2492  * This function can be called more than once, with different engines, if we plan
2493  * to use the context with them. The context backing objects and the ringbuffers
2494  * (specially the ringbuffer backing objects) suck a lot of memory up, and that's why
2495  * the creation is a deferred call: it's better to make sure first that we need to use
2496  * a given ring with the context.
2497  *
2498  * Return: non-zero on error.
2499  */
2500 static int execlists_context_deferred_alloc(struct i915_gem_context *ctx,
2501 					    struct intel_engine_cs *engine)
2502 {
2503 	struct drm_i915_gem_object *ctx_obj;
2504 	struct intel_context *ce = &ctx->engine[engine->id];
2505 	uint32_t context_size;
2506 	struct intel_ringbuffer *ringbuf;
2507 	int ret;
2508 
2509 	WARN_ON(ce->state);
2510 
2511 	context_size = round_up(intel_lr_context_size(engine), 4096);
2512 
2513 	/* One extra page as the sharing data between driver and GuC */
2514 	context_size += PAGE_SIZE * LRC_PPHWSP_PN;
2515 
2516 	ctx_obj = i915_gem_object_create(&ctx->i915->drm, context_size);
2517 	if (IS_ERR(ctx_obj)) {
2518 		DRM_DEBUG_DRIVER("Alloc LRC backing obj failed.\n");
2519 		return PTR_ERR(ctx_obj);
2520 	}
2521 
2522 	ringbuf = intel_engine_create_ringbuffer(engine, ctx->ring_size);
2523 	if (IS_ERR(ringbuf)) {
2524 		ret = PTR_ERR(ringbuf);
2525 		goto error_deref_obj;
2526 	}
2527 
2528 	ret = populate_lr_context(ctx, ctx_obj, engine, ringbuf);
2529 	if (ret) {
2530 		DRM_DEBUG_DRIVER("Failed to populate LRC: %d\n", ret);
2531 		goto error_ringbuf;
2532 	}
2533 
2534 	ce->ringbuf = ringbuf;
2535 	ce->state = ctx_obj;
2536 	ce->initialised = engine->init_context == NULL;
2537 
2538 	return 0;
2539 
2540 error_ringbuf:
2541 	intel_ringbuffer_free(ringbuf);
2542 error_deref_obj:
2543 	drm_gem_object_unreference(&ctx_obj->base);
2544 	ce->ringbuf = NULL;
2545 	ce->state = NULL;
2546 	return ret;
2547 }
2548 
2549 void intel_lr_context_reset(struct drm_i915_private *dev_priv,
2550 			    struct i915_gem_context *ctx)
2551 {
2552 	struct intel_engine_cs *engine;
2553 
2554 	for_each_engine(engine, dev_priv) {
2555 		struct intel_context *ce = &ctx->engine[engine->id];
2556 		struct drm_i915_gem_object *ctx_obj = ce->state;
2557 		void *vaddr;
2558 		uint32_t *reg_state;
2559 
2560 		if (!ctx_obj)
2561 			continue;
2562 
2563 		vaddr = i915_gem_object_pin_map(ctx_obj);
2564 		if (WARN_ON(IS_ERR(vaddr)))
2565 			continue;
2566 
2567 		reg_state = vaddr + LRC_STATE_PN * PAGE_SIZE;
2568 		ctx_obj->dirty = true;
2569 
2570 		reg_state[CTX_RING_HEAD+1] = 0;
2571 		reg_state[CTX_RING_TAIL+1] = 0;
2572 
2573 		i915_gem_object_unpin_map(ctx_obj);
2574 
2575 		ce->ringbuf->head = 0;
2576 		ce->ringbuf->tail = 0;
2577 	}
2578 }
2579