xref: /dflybsd-src/sys/dev/drm/i915/intel_lrc.c (revision 0d27ae553036e6073e8793cf6f2fb4583c33f3dc)
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 
135 #include <drm/drmP.h>
136 #include <drm/i915_drm.h>
137 #include "i915_drv.h"
138 #include "intel_drv.h"
139 
140 #define GEN9_LR_CONTEXT_RENDER_SIZE (22 * PAGE_SIZE)
141 #define GEN8_LR_CONTEXT_RENDER_SIZE (20 * PAGE_SIZE)
142 #define GEN8_LR_CONTEXT_OTHER_SIZE (2 * PAGE_SIZE)
143 
144 #define RING_EXECLIST_QFULL		(1 << 0x2)
145 #define RING_EXECLIST1_VALID		(1 << 0x3)
146 #define RING_EXECLIST0_VALID		(1 << 0x4)
147 #define RING_EXECLIST_ACTIVE_STATUS	(3 << 0xE)
148 #define RING_EXECLIST1_ACTIVE		(1 << 0x11)
149 #define RING_EXECLIST0_ACTIVE		(1 << 0x12)
150 
151 #define GEN8_CTX_STATUS_IDLE_ACTIVE	(1 << 0)
152 #define GEN8_CTX_STATUS_PREEMPTED	(1 << 1)
153 #define GEN8_CTX_STATUS_ELEMENT_SWITCH	(1 << 2)
154 #define GEN8_CTX_STATUS_ACTIVE_IDLE	(1 << 3)
155 #define GEN8_CTX_STATUS_COMPLETE	(1 << 4)
156 #define GEN8_CTX_STATUS_LITE_RESTORE	(1 << 15)
157 
158 #define CTX_LRI_HEADER_0		0x01
159 #define CTX_CONTEXT_CONTROL		0x02
160 #define CTX_RING_HEAD			0x04
161 #define CTX_RING_TAIL			0x06
162 #define CTX_RING_BUFFER_START		0x08
163 #define CTX_RING_BUFFER_CONTROL		0x0a
164 #define CTX_BB_HEAD_U			0x0c
165 #define CTX_BB_HEAD_L			0x0e
166 #define CTX_BB_STATE			0x10
167 #define CTX_SECOND_BB_HEAD_U		0x12
168 #define CTX_SECOND_BB_HEAD_L		0x14
169 #define CTX_SECOND_BB_STATE		0x16
170 #define CTX_BB_PER_CTX_PTR		0x18
171 #define CTX_RCS_INDIRECT_CTX		0x1a
172 #define CTX_RCS_INDIRECT_CTX_OFFSET	0x1c
173 #define CTX_LRI_HEADER_1		0x21
174 #define CTX_CTX_TIMESTAMP		0x22
175 #define CTX_PDP3_UDW			0x24
176 #define CTX_PDP3_LDW			0x26
177 #define CTX_PDP2_UDW			0x28
178 #define CTX_PDP2_LDW			0x2a
179 #define CTX_PDP1_UDW			0x2c
180 #define CTX_PDP1_LDW			0x2e
181 #define CTX_PDP0_UDW			0x30
182 #define CTX_PDP0_LDW			0x32
183 #define CTX_LRI_HEADER_2		0x41
184 #define CTX_R_PWR_CLK_STATE		0x42
185 #define CTX_GPGPU_CSR_BASE_ADDRESS	0x44
186 
187 #define GEN8_CTX_VALID (1<<0)
188 #define GEN8_CTX_FORCE_PD_RESTORE (1<<1)
189 #define GEN8_CTX_FORCE_RESTORE (1<<2)
190 #define GEN8_CTX_L3LLC_COHERENT (1<<5)
191 #define GEN8_CTX_PRIVILEGE (1<<8)
192 
193 #define ASSIGN_CTX_PDP(ppgtt, reg_state, n) { \
194 	const u64 _addr = test_bit(n, ppgtt->pdp.used_pdpes) ? \
195 		ppgtt->pdp.page_directory[n]->daddr : \
196 		ppgtt->scratch_pd->daddr; \
197 	reg_state[CTX_PDP ## n ## _UDW+1] = upper_32_bits(_addr); \
198 	reg_state[CTX_PDP ## n ## _LDW+1] = lower_32_bits(_addr); \
199 }
200 
201 enum {
202 	ADVANCED_CONTEXT = 0,
203 	LEGACY_CONTEXT,
204 	ADVANCED_AD_CONTEXT,
205 	LEGACY_64B_CONTEXT
206 };
207 #define GEN8_CTX_MODE_SHIFT 3
208 enum {
209 	FAULT_AND_HANG = 0,
210 	FAULT_AND_HALT, /* Debug only */
211 	FAULT_AND_STREAM,
212 	FAULT_AND_CONTINUE /* Unsupported */
213 };
214 #define GEN8_CTX_ID_SHIFT 32
215 
216 static int intel_lr_context_pin(struct intel_engine_cs *ring,
217 		struct intel_context *ctx);
218 
219 /**
220  * intel_sanitize_enable_execlists() - sanitize i915.enable_execlists
221  * @dev: DRM device.
222  * @enable_execlists: value of i915.enable_execlists module parameter.
223  *
224  * Only certain platforms support Execlists (the prerequisites being
225  * support for Logical Ring Contexts and Aliasing PPGTT or better).
226  *
227  * Return: 1 if Execlists is supported and has to be enabled.
228  */
229 int intel_sanitize_enable_execlists(struct drm_device *dev, int enable_execlists)
230 {
231 	WARN_ON(i915.enable_ppgtt == -1);
232 
233 	if (INTEL_INFO(dev)->gen >= 9)
234 		return 1;
235 
236 	if (enable_execlists == 0)
237 		return 0;
238 
239 	if (HAS_LOGICAL_RING_CONTEXTS(dev) && USES_PPGTT(dev) &&
240 	    i915.use_mmio_flip >= 0)
241 		return 1;
242 
243 	return 0;
244 }
245 
246 /**
247  * intel_execlists_ctx_id() - get the Execlists Context ID
248  * @ctx_obj: Logical Ring Context backing object.
249  *
250  * Do not confuse with ctx->id! Unfortunately we have a name overload
251  * here: the old context ID we pass to userspace as a handler so that
252  * they can refer to a context, and the new context ID we pass to the
253  * ELSP so that the GPU can inform us of the context status via
254  * interrupts.
255  *
256  * Return: 20-bits globally unique context ID.
257  */
258 u32 intel_execlists_ctx_id(struct drm_i915_gem_object *ctx_obj)
259 {
260 	u32 lrca = i915_gem_obj_ggtt_offset(ctx_obj);
261 
262 	/* LRCA is required to be 4K aligned so the more significant 20 bits
263 	 * are globally unique */
264 	return lrca >> 12;
265 }
266 
267 static uint64_t execlists_ctx_descriptor(struct intel_engine_cs *ring,
268 					 struct drm_i915_gem_object *ctx_obj)
269 {
270 	struct drm_device *dev = ring->dev;
271 	uint64_t desc;
272 	uint64_t lrca = i915_gem_obj_ggtt_offset(ctx_obj);
273 
274 	WARN_ON(lrca & 0xFFFFFFFF00000FFFULL);
275 
276 	desc = GEN8_CTX_VALID;
277 	desc |= LEGACY_CONTEXT << GEN8_CTX_MODE_SHIFT;
278 	if (IS_GEN8(ctx_obj->base.dev))
279 		desc |= GEN8_CTX_L3LLC_COHERENT;
280 	desc |= GEN8_CTX_PRIVILEGE;
281 	desc |= lrca;
282 	desc |= (u64)intel_execlists_ctx_id(ctx_obj) << GEN8_CTX_ID_SHIFT;
283 
284 	/* TODO: WaDisableLiteRestore when we start using semaphore
285 	 * signalling between Command Streamers */
286 	/* desc |= GEN8_CTX_FORCE_RESTORE; */
287 
288 	/* WaEnableForceRestoreInCtxtDescForVCS:skl */
289 	if (IS_GEN9(dev) &&
290 	    INTEL_REVID(dev) <= SKL_REVID_B0 &&
291 	    (ring->id == BCS || ring->id == VCS ||
292 	    ring->id == VECS || ring->id == VCS2))
293 		desc |= GEN8_CTX_FORCE_RESTORE;
294 
295 	return desc;
296 }
297 
298 static void execlists_elsp_write(struct intel_engine_cs *ring,
299 				 struct drm_i915_gem_object *ctx_obj0,
300 				 struct drm_i915_gem_object *ctx_obj1)
301 {
302 	struct drm_device *dev = ring->dev;
303 	struct drm_i915_private *dev_priv = dev->dev_private;
304 	uint64_t temp = 0;
305 	uint32_t desc[4];
306 
307 	/* XXX: You must always write both descriptors in the order below. */
308 	if (ctx_obj1)
309 		temp = execlists_ctx_descriptor(ring, ctx_obj1);
310 	else
311 		temp = 0;
312 	desc[1] = (u32)(temp >> 32);
313 	desc[0] = (u32)temp;
314 
315 	temp = execlists_ctx_descriptor(ring, ctx_obj0);
316 	desc[3] = (u32)(temp >> 32);
317 	desc[2] = (u32)temp;
318 
319 	lockmgr(&dev_priv->uncore.lock, LK_EXCLUSIVE);
320 	intel_uncore_forcewake_get__locked(dev_priv, FORCEWAKE_ALL);
321 	I915_WRITE_FW(RING_ELSP(ring), desc[1]);
322 	I915_WRITE_FW(RING_ELSP(ring), desc[0]);
323 	I915_WRITE_FW(RING_ELSP(ring), desc[3]);
324 
325 	/* The context is automatically loaded after the following */
326 	I915_WRITE_FW(RING_ELSP(ring), desc[2]);
327 
328 	/* ELSP is a wo register, so use another nearby reg for posting instead */
329 	POSTING_READ_FW(RING_EXECLIST_STATUS(ring));
330 	intel_uncore_forcewake_put__locked(dev_priv, FORCEWAKE_ALL);
331 	lockmgr(&dev_priv->uncore.lock, LK_RELEASE);
332 }
333 
334 static int execlists_update_context(struct drm_i915_gem_object *ctx_obj,
335 				    struct drm_i915_gem_object *ring_obj,
336 				    struct i915_hw_ppgtt *ppgtt,
337 				    u32 tail)
338 {
339 	struct vm_page *page;
340 	uint32_t *reg_state;
341 
342 	page = i915_gem_object_get_page(ctx_obj, 1);
343 	reg_state = kmap_atomic(page);
344 
345 	reg_state[CTX_RING_TAIL+1] = tail;
346 	reg_state[CTX_RING_BUFFER_START+1] = i915_gem_obj_ggtt_offset(ring_obj);
347 
348 	/* True PPGTT with dynamic page allocation: update PDP registers and
349 	 * point the unallocated PDPs to the scratch page
350 	 */
351 	if (ppgtt) {
352 		ASSIGN_CTX_PDP(ppgtt, reg_state, 3);
353 		ASSIGN_CTX_PDP(ppgtt, reg_state, 2);
354 		ASSIGN_CTX_PDP(ppgtt, reg_state, 1);
355 		ASSIGN_CTX_PDP(ppgtt, reg_state, 0);
356 	}
357 
358 	kunmap_atomic(reg_state);
359 
360 	return 0;
361 }
362 
363 static void execlists_submit_contexts(struct intel_engine_cs *ring,
364 				      struct intel_context *to0, u32 tail0,
365 				      struct intel_context *to1, u32 tail1)
366 {
367 	struct drm_i915_gem_object *ctx_obj0 = to0->engine[ring->id].state;
368 	struct intel_ringbuffer *ringbuf0 = to0->engine[ring->id].ringbuf;
369 	struct drm_i915_gem_object *ctx_obj1 = NULL;
370 	struct intel_ringbuffer *ringbuf1 = NULL;
371 
372 	BUG_ON(!ctx_obj0);
373 	WARN_ON(!i915_gem_obj_is_pinned(ctx_obj0));
374 	WARN_ON(!i915_gem_obj_is_pinned(ringbuf0->obj));
375 
376 	execlists_update_context(ctx_obj0, ringbuf0->obj, to0->ppgtt, tail0);
377 
378 	if (to1) {
379 		ringbuf1 = to1->engine[ring->id].ringbuf;
380 		ctx_obj1 = to1->engine[ring->id].state;
381 		BUG_ON(!ctx_obj1);
382 		WARN_ON(!i915_gem_obj_is_pinned(ctx_obj1));
383 		WARN_ON(!i915_gem_obj_is_pinned(ringbuf1->obj));
384 
385 		execlists_update_context(ctx_obj1, ringbuf1->obj, to1->ppgtt, tail1);
386 	}
387 
388 	execlists_elsp_write(ring, ctx_obj0, ctx_obj1);
389 }
390 
391 static void execlists_context_unqueue(struct intel_engine_cs *ring)
392 {
393 	struct drm_i915_gem_request *req0 = NULL, *req1 = NULL;
394 	struct drm_i915_gem_request *cursor = NULL, *tmp = NULL;
395 
396 	assert_spin_locked(&ring->execlist_lock);
397 
398 	/*
399 	 * If irqs are not active generate a warning as batches that finish
400 	 * without the irqs may get lost and a GPU Hang may occur.
401 	 */
402 	WARN_ON(!intel_irqs_enabled(ring->dev->dev_private));
403 
404 	if (list_empty(&ring->execlist_queue))
405 		return;
406 
407 	/* Try to read in pairs */
408 	list_for_each_entry_safe(cursor, tmp, &ring->execlist_queue,
409 				 execlist_link) {
410 		if (!req0) {
411 			req0 = cursor;
412 		} else if (req0->ctx == cursor->ctx) {
413 			/* Same ctx: ignore first request, as second request
414 			 * will update tail past first request's workload */
415 			cursor->elsp_submitted = req0->elsp_submitted;
416 			list_del(&req0->execlist_link);
417 			list_add_tail(&req0->execlist_link,
418 				&ring->execlist_retired_req_list);
419 			req0 = cursor;
420 		} else {
421 			req1 = cursor;
422 			break;
423 		}
424 	}
425 
426 	if (IS_GEN8(ring->dev) || IS_GEN9(ring->dev)) {
427 		/*
428 		 * WaIdleLiteRestore: make sure we never cause a lite
429 		 * restore with HEAD==TAIL
430 		 */
431 		if (req0->elsp_submitted) {
432 			/*
433 			 * Apply the wa NOOPS to prevent ring:HEAD == req:TAIL
434 			 * as we resubmit the request. See gen8_emit_request()
435 			 * for where we prepare the padding after the end of the
436 			 * request.
437 			 */
438 			struct intel_ringbuffer *ringbuf;
439 
440 			ringbuf = req0->ctx->engine[ring->id].ringbuf;
441 			req0->tail += 8;
442 			req0->tail &= ringbuf->size - 1;
443 		}
444 	}
445 
446 	WARN_ON(req1 && req1->elsp_submitted);
447 
448 	execlists_submit_contexts(ring, req0->ctx, req0->tail,
449 				  req1 ? req1->ctx : NULL,
450 				  req1 ? req1->tail : 0);
451 
452 	req0->elsp_submitted++;
453 	if (req1)
454 		req1->elsp_submitted++;
455 }
456 
457 static bool execlists_check_remove_request(struct intel_engine_cs *ring,
458 					   u32 request_id)
459 {
460 	struct drm_i915_gem_request *head_req;
461 
462 	assert_spin_locked(&ring->execlist_lock);
463 
464 	head_req = list_first_entry_or_null(&ring->execlist_queue,
465 					    struct drm_i915_gem_request,
466 					    execlist_link);
467 
468 	if (head_req != NULL) {
469 		struct drm_i915_gem_object *ctx_obj =
470 				head_req->ctx->engine[ring->id].state;
471 		if (intel_execlists_ctx_id(ctx_obj) == request_id) {
472 			WARN(head_req->elsp_submitted == 0,
473 			     "Never submitted head request\n");
474 
475 			if (--head_req->elsp_submitted <= 0) {
476 				list_del(&head_req->execlist_link);
477 				list_add_tail(&head_req->execlist_link,
478 					&ring->execlist_retired_req_list);
479 				return true;
480 			}
481 		}
482 	}
483 
484 	return false;
485 }
486 
487 /**
488  * intel_lrc_irq_handler() - handle Context Switch interrupts
489  * @ring: Engine Command Streamer to handle.
490  *
491  * Check the unread Context Status Buffers and manage the submission of new
492  * contexts to the ELSP accordingly.
493  */
494 void intel_lrc_irq_handler(struct intel_engine_cs *ring)
495 {
496 	struct drm_i915_private *dev_priv = ring->dev->dev_private;
497 	u32 status_pointer;
498 	u8 read_pointer;
499 	u8 write_pointer;
500 	u32 status;
501 	u32 status_id;
502 	u32 submit_contexts = 0;
503 
504 	status_pointer = I915_READ(RING_CONTEXT_STATUS_PTR(ring));
505 
506 	read_pointer = ring->next_context_status_buffer;
507 	write_pointer = status_pointer & 0x07;
508 	if (read_pointer > write_pointer)
509 		write_pointer += 6;
510 
511 	lockmgr(&ring->execlist_lock, LK_EXCLUSIVE);
512 
513 	while (read_pointer < write_pointer) {
514 		read_pointer++;
515 		status = I915_READ(RING_CONTEXT_STATUS_BUF(ring) +
516 				(read_pointer % 6) * 8);
517 		status_id = I915_READ(RING_CONTEXT_STATUS_BUF(ring) +
518 				(read_pointer % 6) * 8 + 4);
519 
520 		if (status & GEN8_CTX_STATUS_PREEMPTED) {
521 			if (status & GEN8_CTX_STATUS_LITE_RESTORE) {
522 				if (execlists_check_remove_request(ring, status_id))
523 					WARN(1, "Lite Restored request removed from queue\n");
524 			} else
525 				WARN(1, "Preemption without Lite Restore\n");
526 		}
527 
528 		 if ((status & GEN8_CTX_STATUS_ACTIVE_IDLE) ||
529 		     (status & GEN8_CTX_STATUS_ELEMENT_SWITCH)) {
530 			if (execlists_check_remove_request(ring, status_id))
531 				submit_contexts++;
532 		}
533 	}
534 
535 	if (submit_contexts != 0)
536 		execlists_context_unqueue(ring);
537 
538 	lockmgr(&ring->execlist_lock, LK_RELEASE);
539 
540 	WARN(submit_contexts > 2, "More than two context complete events?\n");
541 	ring->next_context_status_buffer = write_pointer % 6;
542 
543 	I915_WRITE(RING_CONTEXT_STATUS_PTR(ring),
544 		   ((u32)ring->next_context_status_buffer & 0x07) << 8);
545 }
546 
547 static int execlists_context_queue(struct intel_engine_cs *ring,
548 				   struct intel_context *to,
549 				   u32 tail,
550 				   struct drm_i915_gem_request *request)
551 {
552 	struct drm_i915_gem_request *cursor;
553 	int num_elements = 0;
554 
555 	if (to != ring->default_context)
556 		intel_lr_context_pin(ring, to);
557 
558 	if (!request) {
559 		/*
560 		 * If there isn't a request associated with this submission,
561 		 * create one as a temporary holder.
562 		 */
563 		request = kzalloc(sizeof(*request), GFP_KERNEL);
564 		if (request == NULL)
565 			return -ENOMEM;
566 		request->ring = ring;
567 		request->ctx = to;
568 		kref_init(&request->ref);
569 		i915_gem_context_reference(request->ctx);
570 	} else {
571 		i915_gem_request_reference(request);
572 		WARN_ON(to != request->ctx);
573 	}
574 	request->tail = tail;
575 
576 	lockmgr(&ring->execlist_lock, LK_EXCLUSIVE);
577 
578 	list_for_each_entry(cursor, &ring->execlist_queue, execlist_link)
579 		if (++num_elements > 2)
580 			break;
581 
582 	if (num_elements > 2) {
583 		struct drm_i915_gem_request *tail_req;
584 
585 		tail_req = list_last_entry(&ring->execlist_queue,
586 					   struct drm_i915_gem_request,
587 					   execlist_link);
588 
589 		if (to == tail_req->ctx) {
590 			WARN(tail_req->elsp_submitted != 0,
591 				"More than 2 already-submitted reqs queued\n");
592 			list_del(&tail_req->execlist_link);
593 			list_add_tail(&tail_req->execlist_link,
594 				&ring->execlist_retired_req_list);
595 		}
596 	}
597 
598 	list_add_tail(&request->execlist_link, &ring->execlist_queue);
599 	if (num_elements == 0)
600 		execlists_context_unqueue(ring);
601 
602 	lockmgr(&ring->execlist_lock, LK_RELEASE);
603 
604 	return 0;
605 }
606 
607 static int logical_ring_invalidate_all_caches(struct intel_ringbuffer *ringbuf,
608 					      struct intel_context *ctx)
609 {
610 	struct intel_engine_cs *ring = ringbuf->ring;
611 	uint32_t flush_domains;
612 	int ret;
613 
614 	flush_domains = 0;
615 	if (ring->gpu_caches_dirty)
616 		flush_domains = I915_GEM_GPU_DOMAINS;
617 
618 	ret = ring->emit_flush(ringbuf, ctx,
619 			       I915_GEM_GPU_DOMAINS, flush_domains);
620 	if (ret)
621 		return ret;
622 
623 	ring->gpu_caches_dirty = false;
624 	return 0;
625 }
626 
627 static int execlists_move_to_gpu(struct intel_ringbuffer *ringbuf,
628 				 struct intel_context *ctx,
629 				 struct list_head *vmas)
630 {
631 	struct intel_engine_cs *ring = ringbuf->ring;
632 	const unsigned other_rings = ~intel_ring_flag(ring);
633 	struct i915_vma *vma;
634 	uint32_t flush_domains = 0;
635 	bool flush_chipset = false;
636 	int ret;
637 
638 	list_for_each_entry(vma, vmas, exec_list) {
639 		struct drm_i915_gem_object *obj = vma->obj;
640 
641 		if (obj->active & other_rings) {
642 			ret = i915_gem_object_sync(obj, ring);
643 			if (ret)
644 				return ret;
645 		}
646 
647 		if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
648 			flush_chipset |= i915_gem_clflush_object(obj, false);
649 
650 		flush_domains |= obj->base.write_domain;
651 	}
652 
653 	if (flush_domains & I915_GEM_DOMAIN_GTT)
654 		wmb();
655 
656 	/* Unconditionally invalidate gpu caches and ensure that we do flush
657 	 * any residual writes from the previous batch.
658 	 */
659 	return logical_ring_invalidate_all_caches(ringbuf, ctx);
660 }
661 
662 int intel_logical_ring_alloc_request_extras(struct drm_i915_gem_request *request,
663 					    struct intel_context *ctx)
664 {
665 	int ret;
666 
667 	if (ctx != request->ring->default_context) {
668 		ret = intel_lr_context_pin(request->ring, ctx);
669 		if (ret)
670 			return ret;
671 	}
672 
673 	request->ringbuf = ctx->engine[request->ring->id].ringbuf;
674 	request->ctx     = ctx;
675 	i915_gem_context_reference(request->ctx);
676 
677 	return 0;
678 }
679 
680 static int logical_ring_wait_for_space(struct intel_ringbuffer *ringbuf,
681 				       struct intel_context *ctx,
682 				       int bytes)
683 {
684 	struct intel_engine_cs *ring = ringbuf->ring;
685 	struct drm_i915_gem_request *request;
686 	unsigned space;
687 	int ret;
688 
689 	if (intel_ring_space(ringbuf) >= bytes)
690 		return 0;
691 
692 	list_for_each_entry(request, &ring->request_list, list) {
693 		/*
694 		 * The request queue is per-engine, so can contain requests
695 		 * from multiple ringbuffers. Here, we must ignore any that
696 		 * aren't from the ringbuffer we're considering.
697 		 */
698 		if (request->ringbuf != ringbuf)
699 			continue;
700 
701 		/* Would completion of this request free enough space? */
702 		space = __intel_ring_space(request->postfix, ringbuf->tail,
703 					   ringbuf->size);
704 		if (space >= bytes)
705 			break;
706 	}
707 
708 	if (WARN_ON(&request->list == &ring->request_list))
709 		return -ENOSPC;
710 
711 	ret = i915_wait_request(request);
712 	if (ret)
713 		return ret;
714 
715 	ringbuf->space = space;
716 	return 0;
717 }
718 
719 /*
720  * intel_logical_ring_advance_and_submit() - advance the tail and submit the workload
721  * @ringbuf: Logical Ringbuffer to advance.
722  *
723  * The tail is updated in our logical ringbuffer struct, not in the actual context. What
724  * really happens during submission is that the context and current tail will be placed
725  * on a queue waiting for the ELSP to be ready to accept a new context submission. At that
726  * point, the tail *inside* the context is updated and the ELSP written to.
727  */
728 static void
729 intel_logical_ring_advance_and_submit(struct intel_ringbuffer *ringbuf,
730 				      struct intel_context *ctx,
731 				      struct drm_i915_gem_request *request)
732 {
733 	struct intel_engine_cs *ring = ringbuf->ring;
734 
735 	intel_logical_ring_advance(ringbuf);
736 
737 	if (intel_ring_stopped(ring))
738 		return;
739 
740 	execlists_context_queue(ring, ctx, ringbuf->tail, request);
741 }
742 
743 static int logical_ring_wrap_buffer(struct intel_ringbuffer *ringbuf,
744 				    struct intel_context *ctx)
745 {
746 	uint32_t __iomem *virt;
747 	int rem = ringbuf->size - ringbuf->tail;
748 
749 	if (ringbuf->space < rem) {
750 		int ret = logical_ring_wait_for_space(ringbuf, ctx, rem);
751 
752 		if (ret)
753 			return ret;
754 	}
755 
756 	virt = (uint32_t *)(ringbuf->virtual_start + ringbuf->tail);
757 	rem /= 4;
758 	while (rem--)
759 		iowrite32(MI_NOOP, virt++);
760 
761 	ringbuf->tail = 0;
762 	intel_ring_update_space(ringbuf);
763 
764 	return 0;
765 }
766 
767 static int logical_ring_prepare(struct intel_ringbuffer *ringbuf,
768 				struct intel_context *ctx, int bytes)
769 {
770 	int ret;
771 
772 	if (unlikely(ringbuf->tail + bytes > ringbuf->effective_size)) {
773 		ret = logical_ring_wrap_buffer(ringbuf, ctx);
774 		if (unlikely(ret))
775 			return ret;
776 	}
777 
778 	if (unlikely(ringbuf->space < bytes)) {
779 		ret = logical_ring_wait_for_space(ringbuf, ctx, bytes);
780 		if (unlikely(ret))
781 			return ret;
782 	}
783 
784 	return 0;
785 }
786 
787 /**
788  * intel_logical_ring_begin() - prepare the logical ringbuffer to accept some commands
789  *
790  * @ringbuf: Logical ringbuffer.
791  * @num_dwords: number of DWORDs that we plan to write to the ringbuffer.
792  *
793  * The ringbuffer might not be ready to accept the commands right away (maybe it needs to
794  * be wrapped, or wait a bit for the tail to be updated). This function takes care of that
795  * and also preallocates a request (every workload submission is still mediated through
796  * requests, same as it did with legacy ringbuffer submission).
797  *
798  * Return: non-zero if the ringbuffer is not ready to be written to.
799  */
800 static int intel_logical_ring_begin(struct intel_ringbuffer *ringbuf,
801 				    struct intel_context *ctx, int num_dwords)
802 {
803 	struct intel_engine_cs *ring = ringbuf->ring;
804 	struct drm_device *dev = ring->dev;
805 	struct drm_i915_private *dev_priv = dev->dev_private;
806 	int ret;
807 
808 	ret = i915_gem_check_wedge(&dev_priv->gpu_error,
809 				   dev_priv->mm.interruptible);
810 	if (ret)
811 		return ret;
812 
813 	ret = logical_ring_prepare(ringbuf, ctx, num_dwords * sizeof(uint32_t));
814 	if (ret)
815 		return ret;
816 
817 	/* Preallocate the olr before touching the ring */
818 	ret = i915_gem_request_alloc(ring, ctx);
819 	if (ret)
820 		return ret;
821 
822 	ringbuf->space -= num_dwords * sizeof(uint32_t);
823 	return 0;
824 }
825 
826 /**
827  * execlists_submission() - submit a batchbuffer for execution, Execlists style
828  * @dev: DRM device.
829  * @file: DRM file.
830  * @ring: Engine Command Streamer to submit to.
831  * @ctx: Context to employ for this submission.
832  * @args: execbuffer call arguments.
833  * @vmas: list of vmas.
834  * @batch_obj: the batchbuffer to submit.
835  * @exec_start: batchbuffer start virtual address pointer.
836  * @dispatch_flags: translated execbuffer call flags.
837  *
838  * This is the evil twin version of i915_gem_ringbuffer_submission. It abstracts
839  * away the submission details of the execbuffer ioctl call.
840  *
841  * Return: non-zero if the submission fails.
842  */
843 int intel_execlists_submission(struct drm_device *dev, struct drm_file *file,
844 			       struct intel_engine_cs *ring,
845 			       struct intel_context *ctx,
846 			       struct drm_i915_gem_execbuffer2 *args,
847 			       struct list_head *vmas,
848 			       struct drm_i915_gem_object *batch_obj,
849 			       u64 exec_start, u32 dispatch_flags)
850 {
851 	struct drm_i915_private *dev_priv = dev->dev_private;
852 	struct intel_ringbuffer *ringbuf = ctx->engine[ring->id].ringbuf;
853 	int instp_mode;
854 	u32 instp_mask;
855 	int ret;
856 
857 	instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
858 	instp_mask = I915_EXEC_CONSTANTS_MASK;
859 	switch (instp_mode) {
860 	case I915_EXEC_CONSTANTS_REL_GENERAL:
861 	case I915_EXEC_CONSTANTS_ABSOLUTE:
862 	case I915_EXEC_CONSTANTS_REL_SURFACE:
863 		if (instp_mode != 0 && ring != &dev_priv->ring[RCS]) {
864 			DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
865 			return -EINVAL;
866 		}
867 
868 		if (instp_mode != dev_priv->relative_constants_mode) {
869 			if (instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
870 				DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
871 				return -EINVAL;
872 			}
873 
874 			/* The HW changed the meaning on this bit on gen6 */
875 			instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
876 		}
877 		break;
878 	default:
879 		DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode);
880 		return -EINVAL;
881 	}
882 
883 	if (args->num_cliprects != 0) {
884 		DRM_DEBUG("clip rectangles are only valid on pre-gen5\n");
885 		return -EINVAL;
886 	} else {
887 		if (args->DR4 == 0xffffffff) {
888 			DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
889 			args->DR4 = 0;
890 		}
891 
892 		if (args->DR1 || args->DR4 || args->cliprects_ptr) {
893 			DRM_DEBUG("0 cliprects but dirt in cliprects fields\n");
894 			return -EINVAL;
895 		}
896 	}
897 
898 	if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
899 		DRM_DEBUG("sol reset is gen7 only\n");
900 		return -EINVAL;
901 	}
902 
903 	ret = execlists_move_to_gpu(ringbuf, ctx, vmas);
904 	if (ret)
905 		return ret;
906 
907 	if (ring == &dev_priv->ring[RCS] &&
908 	    instp_mode != dev_priv->relative_constants_mode) {
909 		ret = intel_logical_ring_begin(ringbuf, ctx, 4);
910 		if (ret)
911 			return ret;
912 
913 		intel_logical_ring_emit(ringbuf, MI_NOOP);
914 		intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(1));
915 		intel_logical_ring_emit(ringbuf, INSTPM);
916 		intel_logical_ring_emit(ringbuf, instp_mask << 16 | instp_mode);
917 		intel_logical_ring_advance(ringbuf);
918 
919 		dev_priv->relative_constants_mode = instp_mode;
920 	}
921 
922 	ret = ring->emit_bb_start(ringbuf, ctx, exec_start, dispatch_flags);
923 	if (ret)
924 		return ret;
925 
926 	trace_i915_gem_ring_dispatch(intel_ring_get_request(ring), dispatch_flags);
927 
928 	i915_gem_execbuffer_move_to_active(vmas, ring);
929 	i915_gem_execbuffer_retire_commands(dev, file, ring, batch_obj);
930 
931 	return 0;
932 }
933 
934 void intel_execlists_retire_requests(struct intel_engine_cs *ring)
935 {
936 	struct drm_i915_gem_request *req, *tmp;
937 	struct list_head retired_list;
938 
939 	WARN_ON(!mutex_is_locked(&ring->dev->struct_mutex));
940 	if (list_empty(&ring->execlist_retired_req_list))
941 		return;
942 
943 	INIT_LIST_HEAD(&retired_list);
944 	lockmgr(&ring->execlist_lock, LK_EXCLUSIVE);
945 	list_replace_init(&ring->execlist_retired_req_list, &retired_list);
946 	lockmgr(&ring->execlist_lock, LK_RELEASE);
947 
948 	list_for_each_entry_safe(req, tmp, &retired_list, execlist_link) {
949 		struct intel_context *ctx = req->ctx;
950 		struct drm_i915_gem_object *ctx_obj =
951 				ctx->engine[ring->id].state;
952 
953 		if (ctx_obj && (ctx != ring->default_context))
954 			intel_lr_context_unpin(ring, ctx);
955 		list_del(&req->execlist_link);
956 		i915_gem_request_unreference(req);
957 	}
958 }
959 
960 void intel_logical_ring_stop(struct intel_engine_cs *ring)
961 {
962 	struct drm_i915_private *dev_priv = ring->dev->dev_private;
963 	int ret;
964 
965 	if (!intel_ring_initialized(ring))
966 		return;
967 
968 	ret = intel_ring_idle(ring);
969 	if (ret && !i915_reset_in_progress(&to_i915(ring->dev)->gpu_error))
970 		DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n",
971 			  ring->name, ret);
972 
973 	/* TODO: Is this correct with Execlists enabled? */
974 	I915_WRITE_MODE(ring, _MASKED_BIT_ENABLE(STOP_RING));
975 	if (wait_for_atomic((I915_READ_MODE(ring) & MODE_IDLE) != 0, 1000)) {
976 		DRM_ERROR("%s :timed out trying to stop ring\n", ring->name);
977 		return;
978 	}
979 	I915_WRITE_MODE(ring, _MASKED_BIT_DISABLE(STOP_RING));
980 }
981 
982 int logical_ring_flush_all_caches(struct intel_ringbuffer *ringbuf,
983 				  struct intel_context *ctx)
984 {
985 	struct intel_engine_cs *ring = ringbuf->ring;
986 	int ret;
987 
988 	if (!ring->gpu_caches_dirty)
989 		return 0;
990 
991 	ret = ring->emit_flush(ringbuf, ctx, 0, I915_GEM_GPU_DOMAINS);
992 	if (ret)
993 		return ret;
994 
995 	ring->gpu_caches_dirty = false;
996 	return 0;
997 }
998 
999 static int intel_lr_context_pin(struct intel_engine_cs *ring,
1000 		struct intel_context *ctx)
1001 {
1002 	struct drm_i915_gem_object *ctx_obj = ctx->engine[ring->id].state;
1003 	struct intel_ringbuffer *ringbuf = ctx->engine[ring->id].ringbuf;
1004 	int ret = 0;
1005 
1006 	WARN_ON(!mutex_is_locked(&ring->dev->struct_mutex));
1007 	if (ctx->engine[ring->id].pin_count++ == 0) {
1008 		ret = i915_gem_obj_ggtt_pin(ctx_obj,
1009 				GEN8_LR_CONTEXT_ALIGN, 0);
1010 		if (ret)
1011 			goto reset_pin_count;
1012 
1013 		ret = intel_pin_and_map_ringbuffer_obj(ring->dev, ringbuf);
1014 		if (ret)
1015 			goto unpin_ctx_obj;
1016 
1017 		ctx_obj->dirty = true;
1018 	}
1019 
1020 	return ret;
1021 
1022 unpin_ctx_obj:
1023 	i915_gem_object_ggtt_unpin(ctx_obj);
1024 reset_pin_count:
1025 	ctx->engine[ring->id].pin_count = 0;
1026 
1027 	return ret;
1028 }
1029 
1030 void intel_lr_context_unpin(struct intel_engine_cs *ring,
1031 		struct intel_context *ctx)
1032 {
1033 	struct drm_i915_gem_object *ctx_obj = ctx->engine[ring->id].state;
1034 	struct intel_ringbuffer *ringbuf = ctx->engine[ring->id].ringbuf;
1035 
1036 	if (ctx_obj) {
1037 		WARN_ON(!mutex_is_locked(&ring->dev->struct_mutex));
1038 		if (--ctx->engine[ring->id].pin_count == 0) {
1039 			intel_unpin_ringbuffer_obj(ringbuf);
1040 			i915_gem_object_ggtt_unpin(ctx_obj);
1041 		}
1042 	}
1043 }
1044 
1045 static int intel_logical_ring_workarounds_emit(struct intel_engine_cs *ring,
1046 					       struct intel_context *ctx)
1047 {
1048 	int ret, i;
1049 	struct intel_ringbuffer *ringbuf = ctx->engine[ring->id].ringbuf;
1050 	struct drm_device *dev = ring->dev;
1051 	struct drm_i915_private *dev_priv = dev->dev_private;
1052 	struct i915_workarounds *w = &dev_priv->workarounds;
1053 
1054 	if (WARN_ON_ONCE(w->count == 0))
1055 		return 0;
1056 
1057 	ring->gpu_caches_dirty = true;
1058 	ret = logical_ring_flush_all_caches(ringbuf, ctx);
1059 	if (ret)
1060 		return ret;
1061 
1062 	ret = intel_logical_ring_begin(ringbuf, ctx, w->count * 2 + 2);
1063 	if (ret)
1064 		return ret;
1065 
1066 	intel_logical_ring_emit(ringbuf, MI_LOAD_REGISTER_IMM(w->count));
1067 	for (i = 0; i < w->count; i++) {
1068 		intel_logical_ring_emit(ringbuf, w->reg[i].addr);
1069 		intel_logical_ring_emit(ringbuf, w->reg[i].value);
1070 	}
1071 	intel_logical_ring_emit(ringbuf, MI_NOOP);
1072 
1073 	intel_logical_ring_advance(ringbuf);
1074 
1075 	ring->gpu_caches_dirty = true;
1076 	ret = logical_ring_flush_all_caches(ringbuf, ctx);
1077 	if (ret)
1078 		return ret;
1079 
1080 	return 0;
1081 }
1082 
1083 static int gen8_init_common_ring(struct intel_engine_cs *ring)
1084 {
1085 	struct drm_device *dev = ring->dev;
1086 	struct drm_i915_private *dev_priv = dev->dev_private;
1087 
1088 	I915_WRITE_IMR(ring, ~(ring->irq_enable_mask | ring->irq_keep_mask));
1089 	I915_WRITE(RING_HWSTAM(ring->mmio_base), 0xffffffff);
1090 
1091 	if (ring->status_page.obj) {
1092 		I915_WRITE(RING_HWS_PGA(ring->mmio_base),
1093 			   (u32)ring->status_page.gfx_addr);
1094 		POSTING_READ(RING_HWS_PGA(ring->mmio_base));
1095 	}
1096 
1097 	I915_WRITE(RING_MODE_GEN7(ring),
1098 		   _MASKED_BIT_DISABLE(GFX_REPLAY_MODE) |
1099 		   _MASKED_BIT_ENABLE(GFX_RUN_LIST_ENABLE));
1100 	POSTING_READ(RING_MODE_GEN7(ring));
1101 	ring->next_context_status_buffer = 0;
1102 	DRM_DEBUG_DRIVER("Execlists enabled for %s\n", ring->name);
1103 
1104 	memset(&ring->hangcheck, 0, sizeof(ring->hangcheck));
1105 
1106 	return 0;
1107 }
1108 
1109 static int gen8_init_render_ring(struct intel_engine_cs *ring)
1110 {
1111 	struct drm_device *dev = ring->dev;
1112 	struct drm_i915_private *dev_priv = dev->dev_private;
1113 	int ret;
1114 
1115 	ret = gen8_init_common_ring(ring);
1116 	if (ret)
1117 		return ret;
1118 
1119 	/* We need to disable the AsyncFlip performance optimisations in order
1120 	 * to use MI_WAIT_FOR_EVENT within the CS. It should already be
1121 	 * programmed to '1' on all products.
1122 	 *
1123 	 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw,chv
1124 	 */
1125 	I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
1126 
1127 	I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
1128 
1129 	return init_workarounds_ring(ring);
1130 }
1131 
1132 static int gen9_init_render_ring(struct intel_engine_cs *ring)
1133 {
1134 	int ret;
1135 
1136 	ret = gen8_init_common_ring(ring);
1137 	if (ret)
1138 		return ret;
1139 
1140 	return init_workarounds_ring(ring);
1141 }
1142 
1143 static int gen8_emit_bb_start(struct intel_ringbuffer *ringbuf,
1144 			      struct intel_context *ctx,
1145 			      u64 offset, unsigned dispatch_flags)
1146 {
1147 	bool ppgtt = !(dispatch_flags & I915_DISPATCH_SECURE);
1148 	int ret;
1149 
1150 	ret = intel_logical_ring_begin(ringbuf, ctx, 4);
1151 	if (ret)
1152 		return ret;
1153 
1154 	/* FIXME(BDW): Address space and security selectors. */
1155 	intel_logical_ring_emit(ringbuf, MI_BATCH_BUFFER_START_GEN8 | (ppgtt<<8));
1156 	intel_logical_ring_emit(ringbuf, lower_32_bits(offset));
1157 	intel_logical_ring_emit(ringbuf, upper_32_bits(offset));
1158 	intel_logical_ring_emit(ringbuf, MI_NOOP);
1159 	intel_logical_ring_advance(ringbuf);
1160 
1161 	return 0;
1162 }
1163 
1164 static bool gen8_logical_ring_get_irq(struct intel_engine_cs *ring)
1165 {
1166 	struct drm_device *dev = ring->dev;
1167 	struct drm_i915_private *dev_priv = dev->dev_private;
1168 
1169 	if (WARN_ON(!intel_irqs_enabled(dev_priv)))
1170 		return false;
1171 
1172 	lockmgr(&dev_priv->irq_lock, LK_EXCLUSIVE);
1173 	if (ring->irq_refcount++ == 0) {
1174 		I915_WRITE_IMR(ring, ~(ring->irq_enable_mask | ring->irq_keep_mask));
1175 		POSTING_READ(RING_IMR(ring->mmio_base));
1176 	}
1177 	lockmgr(&dev_priv->irq_lock, LK_RELEASE);
1178 
1179 	return true;
1180 }
1181 
1182 static void gen8_logical_ring_put_irq(struct intel_engine_cs *ring)
1183 {
1184 	struct drm_device *dev = ring->dev;
1185 	struct drm_i915_private *dev_priv = dev->dev_private;
1186 
1187 	lockmgr(&dev_priv->irq_lock, LK_EXCLUSIVE);
1188 	if (--ring->irq_refcount == 0) {
1189 		I915_WRITE_IMR(ring, ~ring->irq_keep_mask);
1190 		POSTING_READ(RING_IMR(ring->mmio_base));
1191 	}
1192 	lockmgr(&dev_priv->irq_lock, LK_RELEASE);
1193 }
1194 
1195 static int gen8_emit_flush(struct intel_ringbuffer *ringbuf,
1196 			   struct intel_context *ctx,
1197 			   u32 invalidate_domains,
1198 			   u32 unused)
1199 {
1200 	struct intel_engine_cs *ring = ringbuf->ring;
1201 	struct drm_device *dev = ring->dev;
1202 	struct drm_i915_private *dev_priv = dev->dev_private;
1203 	uint32_t cmd;
1204 	int ret;
1205 
1206 	ret = intel_logical_ring_begin(ringbuf, ctx, 4);
1207 	if (ret)
1208 		return ret;
1209 
1210 	cmd = MI_FLUSH_DW + 1;
1211 
1212 	/* We always require a command barrier so that subsequent
1213 	 * commands, such as breadcrumb interrupts, are strictly ordered
1214 	 * wrt the contents of the write cache being flushed to memory
1215 	 * (and thus being coherent from the CPU).
1216 	 */
1217 	cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1218 
1219 	if (invalidate_domains & I915_GEM_GPU_DOMAINS) {
1220 		cmd |= MI_INVALIDATE_TLB;
1221 		if (ring == &dev_priv->ring[VCS])
1222 			cmd |= MI_INVALIDATE_BSD;
1223 	}
1224 
1225 	intel_logical_ring_emit(ringbuf, cmd);
1226 	intel_logical_ring_emit(ringbuf,
1227 				I915_GEM_HWS_SCRATCH_ADDR |
1228 				MI_FLUSH_DW_USE_GTT);
1229 	intel_logical_ring_emit(ringbuf, 0); /* upper addr */
1230 	intel_logical_ring_emit(ringbuf, 0); /* value */
1231 	intel_logical_ring_advance(ringbuf);
1232 
1233 	return 0;
1234 }
1235 
1236 static int gen8_emit_flush_render(struct intel_ringbuffer *ringbuf,
1237 				  struct intel_context *ctx,
1238 				  u32 invalidate_domains,
1239 				  u32 flush_domains)
1240 {
1241 	struct intel_engine_cs *ring = ringbuf->ring;
1242 	u32 scratch_addr = ring->scratch.gtt_offset + 2 * CACHELINE_BYTES;
1243 	bool vf_flush_wa;
1244 	u32 flags = 0;
1245 	int ret;
1246 
1247 	flags |= PIPE_CONTROL_CS_STALL;
1248 
1249 	if (flush_domains) {
1250 		flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
1251 		flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
1252 		flags |= PIPE_CONTROL_FLUSH_ENABLE;
1253 	}
1254 
1255 	if (invalidate_domains) {
1256 		flags |= PIPE_CONTROL_TLB_INVALIDATE;
1257 		flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
1258 		flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
1259 		flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
1260 		flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
1261 		flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
1262 		flags |= PIPE_CONTROL_QW_WRITE;
1263 		flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
1264 	}
1265 
1266 	/*
1267 	 * On GEN9+ Before VF_CACHE_INVALIDATE we need to emit a NULL pipe
1268 	 * control.
1269 	 */
1270 	vf_flush_wa = INTEL_INFO(ring->dev)->gen >= 9 &&
1271 		      flags & PIPE_CONTROL_VF_CACHE_INVALIDATE;
1272 
1273 	ret = intel_logical_ring_begin(ringbuf, ctx, vf_flush_wa ? 12 : 6);
1274 	if (ret)
1275 		return ret;
1276 
1277 	if (vf_flush_wa) {
1278 		intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1279 		intel_logical_ring_emit(ringbuf, 0);
1280 		intel_logical_ring_emit(ringbuf, 0);
1281 		intel_logical_ring_emit(ringbuf, 0);
1282 		intel_logical_ring_emit(ringbuf, 0);
1283 		intel_logical_ring_emit(ringbuf, 0);
1284 	}
1285 
1286 	intel_logical_ring_emit(ringbuf, GFX_OP_PIPE_CONTROL(6));
1287 	intel_logical_ring_emit(ringbuf, flags);
1288 	intel_logical_ring_emit(ringbuf, scratch_addr);
1289 	intel_logical_ring_emit(ringbuf, 0);
1290 	intel_logical_ring_emit(ringbuf, 0);
1291 	intel_logical_ring_emit(ringbuf, 0);
1292 	intel_logical_ring_advance(ringbuf);
1293 
1294 	return 0;
1295 }
1296 
1297 static u32 gen8_get_seqno(struct intel_engine_cs *ring, bool lazy_coherency)
1298 {
1299 	return intel_read_status_page(ring, I915_GEM_HWS_INDEX);
1300 }
1301 
1302 static void gen8_set_seqno(struct intel_engine_cs *ring, u32 seqno)
1303 {
1304 	intel_write_status_page(ring, I915_GEM_HWS_INDEX, seqno);
1305 }
1306 
1307 static int gen8_emit_request(struct intel_ringbuffer *ringbuf,
1308 			     struct drm_i915_gem_request *request)
1309 {
1310 	struct intel_engine_cs *ring = ringbuf->ring;
1311 	u32 cmd;
1312 	int ret;
1313 
1314 	/*
1315 	 * Reserve space for 2 NOOPs at the end of each request to be
1316 	 * used as a workaround for not being allowed to do lite
1317 	 * restore with HEAD==TAIL (WaIdleLiteRestore).
1318 	 */
1319 	ret = intel_logical_ring_begin(ringbuf, request->ctx, 8);
1320 	if (ret)
1321 		return ret;
1322 
1323 	cmd = MI_STORE_DWORD_IMM_GEN4;
1324 	cmd |= MI_GLOBAL_GTT;
1325 
1326 	intel_logical_ring_emit(ringbuf, cmd);
1327 	intel_logical_ring_emit(ringbuf,
1328 				(ring->status_page.gfx_addr +
1329 				(I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT)));
1330 	intel_logical_ring_emit(ringbuf, 0);
1331 	intel_logical_ring_emit(ringbuf,
1332 		i915_gem_request_get_seqno(ring->outstanding_lazy_request));
1333 	intel_logical_ring_emit(ringbuf, MI_USER_INTERRUPT);
1334 	intel_logical_ring_emit(ringbuf, MI_NOOP);
1335 	intel_logical_ring_advance_and_submit(ringbuf, request->ctx, request);
1336 
1337 	/*
1338 	 * Here we add two extra NOOPs as padding to avoid
1339 	 * lite restore of a context with HEAD==TAIL.
1340 	 */
1341 	intel_logical_ring_emit(ringbuf, MI_NOOP);
1342 	intel_logical_ring_emit(ringbuf, MI_NOOP);
1343 	intel_logical_ring_advance(ringbuf);
1344 
1345 	return 0;
1346 }
1347 
1348 static int intel_lr_context_render_state_init(struct intel_engine_cs *ring,
1349 					      struct intel_context *ctx)
1350 {
1351 	struct intel_ringbuffer *ringbuf = ctx->engine[ring->id].ringbuf;
1352 	struct render_state so;
1353 	struct drm_i915_file_private *file_priv = ctx->file_priv;
1354 	struct drm_file *file = file_priv ? file_priv->file : NULL;
1355 	int ret;
1356 
1357 	ret = i915_gem_render_state_prepare(ring, &so);
1358 	if (ret)
1359 		return ret;
1360 
1361 	if (so.rodata == NULL)
1362 		return 0;
1363 
1364 	ret = ring->emit_bb_start(ringbuf,
1365 			ctx,
1366 			so.ggtt_offset,
1367 			I915_DISPATCH_SECURE);
1368 	if (ret)
1369 		goto out;
1370 
1371 	i915_vma_move_to_active(i915_gem_obj_to_ggtt(so.obj), ring);
1372 
1373 	ret = __i915_add_request(ring, file, so.obj);
1374 	/* intel_logical_ring_add_request moves object to inactive if it
1375 	 * fails */
1376 out:
1377 	i915_gem_render_state_fini(&so);
1378 	return ret;
1379 }
1380 
1381 static int gen8_init_rcs_context(struct intel_engine_cs *ring,
1382 		       struct intel_context *ctx)
1383 {
1384 	int ret;
1385 
1386 	ret = intel_logical_ring_workarounds_emit(ring, ctx);
1387 	if (ret)
1388 		return ret;
1389 
1390 	return intel_lr_context_render_state_init(ring, ctx);
1391 }
1392 
1393 /**
1394  * intel_logical_ring_cleanup() - deallocate the Engine Command Streamer
1395  *
1396  * @ring: Engine Command Streamer.
1397  *
1398  */
1399 void intel_logical_ring_cleanup(struct intel_engine_cs *ring)
1400 {
1401 	struct drm_i915_private *dev_priv;
1402 
1403 	if (!intel_ring_initialized(ring))
1404 		return;
1405 
1406 	dev_priv = ring->dev->dev_private;
1407 
1408 	intel_logical_ring_stop(ring);
1409 	WARN_ON((I915_READ_MODE(ring) & MODE_IDLE) == 0);
1410 	i915_gem_request_assign(&ring->outstanding_lazy_request, NULL);
1411 
1412 	if (ring->cleanup)
1413 		ring->cleanup(ring);
1414 
1415 	i915_cmd_parser_fini_ring(ring);
1416 	i915_gem_batch_pool_fini(&ring->batch_pool);
1417 
1418 	if (ring->status_page.obj) {
1419 		kunmap(sg_page(ring->status_page.obj->pages->sgl));
1420 		ring->status_page.obj = NULL;
1421 	}
1422 }
1423 
1424 static int logical_ring_init(struct drm_device *dev, struct intel_engine_cs *ring)
1425 {
1426 	int ret;
1427 
1428 	/* Intentionally left blank. */
1429 	ring->buffer = NULL;
1430 
1431 	ring->dev = dev;
1432 	INIT_LIST_HEAD(&ring->active_list);
1433 	INIT_LIST_HEAD(&ring->request_list);
1434 	i915_gem_batch_pool_init(dev, &ring->batch_pool);
1435 	init_waitqueue_head(&ring->irq_queue);
1436 
1437 	INIT_LIST_HEAD(&ring->execlist_queue);
1438 	INIT_LIST_HEAD(&ring->execlist_retired_req_list);
1439 	lockinit(&ring->execlist_lock, "i915el", 0, LK_CANRECURSE);
1440 
1441 	ret = i915_cmd_parser_init_ring(ring);
1442 	if (ret)
1443 		return ret;
1444 
1445 	ret = intel_lr_context_deferred_create(ring->default_context, ring);
1446 
1447 	return ret;
1448 }
1449 
1450 static int logical_render_ring_init(struct drm_device *dev)
1451 {
1452 	struct drm_i915_private *dev_priv = dev->dev_private;
1453 	struct intel_engine_cs *ring = &dev_priv->ring[RCS];
1454 	int ret;
1455 
1456 	ring->name = "render ring";
1457 	ring->id = RCS;
1458 	ring->mmio_base = RENDER_RING_BASE;
1459 	ring->irq_enable_mask =
1460 		GT_RENDER_USER_INTERRUPT << GEN8_RCS_IRQ_SHIFT;
1461 	ring->irq_keep_mask =
1462 		GT_CONTEXT_SWITCH_INTERRUPT << GEN8_RCS_IRQ_SHIFT;
1463 	if (HAS_L3_DPF(dev))
1464 		ring->irq_keep_mask |= GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
1465 
1466 	if (INTEL_INFO(dev)->gen >= 9)
1467 		ring->init_hw = gen9_init_render_ring;
1468 	else
1469 		ring->init_hw = gen8_init_render_ring;
1470 	ring->init_context = gen8_init_rcs_context;
1471 	ring->cleanup = intel_fini_pipe_control;
1472 	ring->get_seqno = gen8_get_seqno;
1473 	ring->set_seqno = gen8_set_seqno;
1474 	ring->emit_request = gen8_emit_request;
1475 	ring->emit_flush = gen8_emit_flush_render;
1476 	ring->irq_get = gen8_logical_ring_get_irq;
1477 	ring->irq_put = gen8_logical_ring_put_irq;
1478 	ring->emit_bb_start = gen8_emit_bb_start;
1479 
1480 	ring->dev = dev;
1481 	ret = logical_ring_init(dev, ring);
1482 	if (ret)
1483 		return ret;
1484 
1485 	return intel_init_pipe_control(ring);
1486 }
1487 
1488 static int logical_bsd_ring_init(struct drm_device *dev)
1489 {
1490 	struct drm_i915_private *dev_priv = dev->dev_private;
1491 	struct intel_engine_cs *ring = &dev_priv->ring[VCS];
1492 
1493 	ring->name = "bsd ring";
1494 	ring->id = VCS;
1495 	ring->mmio_base = GEN6_BSD_RING_BASE;
1496 	ring->irq_enable_mask =
1497 		GT_RENDER_USER_INTERRUPT << GEN8_VCS1_IRQ_SHIFT;
1498 	ring->irq_keep_mask =
1499 		GT_CONTEXT_SWITCH_INTERRUPT << GEN8_VCS1_IRQ_SHIFT;
1500 
1501 	ring->init_hw = gen8_init_common_ring;
1502 	ring->get_seqno = gen8_get_seqno;
1503 	ring->set_seqno = gen8_set_seqno;
1504 	ring->emit_request = gen8_emit_request;
1505 	ring->emit_flush = gen8_emit_flush;
1506 	ring->irq_get = gen8_logical_ring_get_irq;
1507 	ring->irq_put = gen8_logical_ring_put_irq;
1508 	ring->emit_bb_start = gen8_emit_bb_start;
1509 
1510 	return logical_ring_init(dev, ring);
1511 }
1512 
1513 static int logical_bsd2_ring_init(struct drm_device *dev)
1514 {
1515 	struct drm_i915_private *dev_priv = dev->dev_private;
1516 	struct intel_engine_cs *ring = &dev_priv->ring[VCS2];
1517 
1518 	ring->name = "bds2 ring";
1519 	ring->id = VCS2;
1520 	ring->mmio_base = GEN8_BSD2_RING_BASE;
1521 	ring->irq_enable_mask =
1522 		GT_RENDER_USER_INTERRUPT << GEN8_VCS2_IRQ_SHIFT;
1523 	ring->irq_keep_mask =
1524 		GT_CONTEXT_SWITCH_INTERRUPT << GEN8_VCS2_IRQ_SHIFT;
1525 
1526 	ring->init_hw = gen8_init_common_ring;
1527 	ring->get_seqno = gen8_get_seqno;
1528 	ring->set_seqno = gen8_set_seqno;
1529 	ring->emit_request = gen8_emit_request;
1530 	ring->emit_flush = gen8_emit_flush;
1531 	ring->irq_get = gen8_logical_ring_get_irq;
1532 	ring->irq_put = gen8_logical_ring_put_irq;
1533 	ring->emit_bb_start = gen8_emit_bb_start;
1534 
1535 	return logical_ring_init(dev, ring);
1536 }
1537 
1538 static int logical_blt_ring_init(struct drm_device *dev)
1539 {
1540 	struct drm_i915_private *dev_priv = dev->dev_private;
1541 	struct intel_engine_cs *ring = &dev_priv->ring[BCS];
1542 
1543 	ring->name = "blitter ring";
1544 	ring->id = BCS;
1545 	ring->mmio_base = BLT_RING_BASE;
1546 	ring->irq_enable_mask =
1547 		GT_RENDER_USER_INTERRUPT << GEN8_BCS_IRQ_SHIFT;
1548 	ring->irq_keep_mask =
1549 		GT_CONTEXT_SWITCH_INTERRUPT << GEN8_BCS_IRQ_SHIFT;
1550 
1551 	ring->init_hw = gen8_init_common_ring;
1552 	ring->get_seqno = gen8_get_seqno;
1553 	ring->set_seqno = gen8_set_seqno;
1554 	ring->emit_request = gen8_emit_request;
1555 	ring->emit_flush = gen8_emit_flush;
1556 	ring->irq_get = gen8_logical_ring_get_irq;
1557 	ring->irq_put = gen8_logical_ring_put_irq;
1558 	ring->emit_bb_start = gen8_emit_bb_start;
1559 
1560 	return logical_ring_init(dev, ring);
1561 }
1562 
1563 static int logical_vebox_ring_init(struct drm_device *dev)
1564 {
1565 	struct drm_i915_private *dev_priv = dev->dev_private;
1566 	struct intel_engine_cs *ring = &dev_priv->ring[VECS];
1567 
1568 	ring->name = "video enhancement ring";
1569 	ring->id = VECS;
1570 	ring->mmio_base = VEBOX_RING_BASE;
1571 	ring->irq_enable_mask =
1572 		GT_RENDER_USER_INTERRUPT << GEN8_VECS_IRQ_SHIFT;
1573 	ring->irq_keep_mask =
1574 		GT_CONTEXT_SWITCH_INTERRUPT << GEN8_VECS_IRQ_SHIFT;
1575 
1576 	ring->init_hw = gen8_init_common_ring;
1577 	ring->get_seqno = gen8_get_seqno;
1578 	ring->set_seqno = gen8_set_seqno;
1579 	ring->emit_request = gen8_emit_request;
1580 	ring->emit_flush = gen8_emit_flush;
1581 	ring->irq_get = gen8_logical_ring_get_irq;
1582 	ring->irq_put = gen8_logical_ring_put_irq;
1583 	ring->emit_bb_start = gen8_emit_bb_start;
1584 
1585 	return logical_ring_init(dev, ring);
1586 }
1587 
1588 /**
1589  * intel_logical_rings_init() - allocate, populate and init the Engine Command Streamers
1590  * @dev: DRM device.
1591  *
1592  * This function inits the engines for an Execlists submission style (the equivalent in the
1593  * legacy ringbuffer submission world would be i915_gem_init_rings). It does it only for
1594  * those engines that are present in the hardware.
1595  *
1596  * Return: non-zero if the initialization failed.
1597  */
1598 int intel_logical_rings_init(struct drm_device *dev)
1599 {
1600 	struct drm_i915_private *dev_priv = dev->dev_private;
1601 	int ret;
1602 
1603 	ret = logical_render_ring_init(dev);
1604 	if (ret)
1605 		return ret;
1606 
1607 	if (HAS_BSD(dev)) {
1608 		ret = logical_bsd_ring_init(dev);
1609 		if (ret)
1610 			goto cleanup_render_ring;
1611 	}
1612 
1613 	if (HAS_BLT(dev)) {
1614 		ret = logical_blt_ring_init(dev);
1615 		if (ret)
1616 			goto cleanup_bsd_ring;
1617 	}
1618 
1619 	if (HAS_VEBOX(dev)) {
1620 		ret = logical_vebox_ring_init(dev);
1621 		if (ret)
1622 			goto cleanup_blt_ring;
1623 	}
1624 
1625 	if (HAS_BSD2(dev)) {
1626 		ret = logical_bsd2_ring_init(dev);
1627 		if (ret)
1628 			goto cleanup_vebox_ring;
1629 	}
1630 
1631 	ret = i915_gem_set_seqno(dev, ((u32)~0 - 0x1000));
1632 	if (ret)
1633 		goto cleanup_bsd2_ring;
1634 
1635 	return 0;
1636 
1637 cleanup_bsd2_ring:
1638 	intel_logical_ring_cleanup(&dev_priv->ring[VCS2]);
1639 cleanup_vebox_ring:
1640 	intel_logical_ring_cleanup(&dev_priv->ring[VECS]);
1641 cleanup_blt_ring:
1642 	intel_logical_ring_cleanup(&dev_priv->ring[BCS]);
1643 cleanup_bsd_ring:
1644 	intel_logical_ring_cleanup(&dev_priv->ring[VCS]);
1645 cleanup_render_ring:
1646 	intel_logical_ring_cleanup(&dev_priv->ring[RCS]);
1647 
1648 	return ret;
1649 }
1650 
1651 static u32
1652 make_rpcs(struct drm_device *dev)
1653 {
1654 	u32 rpcs = 0;
1655 
1656 	/*
1657 	 * No explicit RPCS request is needed to ensure full
1658 	 * slice/subslice/EU enablement prior to Gen9.
1659 	*/
1660 	if (INTEL_INFO(dev)->gen < 9)
1661 		return 0;
1662 
1663 	/*
1664 	 * Starting in Gen9, render power gating can leave
1665 	 * slice/subslice/EU in a partially enabled state. We
1666 	 * must make an explicit request through RPCS for full
1667 	 * enablement.
1668 	*/
1669 	if (INTEL_INFO(dev)->has_slice_pg) {
1670 		rpcs |= GEN8_RPCS_S_CNT_ENABLE;
1671 		rpcs |= INTEL_INFO(dev)->slice_total <<
1672 			GEN8_RPCS_S_CNT_SHIFT;
1673 		rpcs |= GEN8_RPCS_ENABLE;
1674 	}
1675 
1676 	if (INTEL_INFO(dev)->has_subslice_pg) {
1677 		rpcs |= GEN8_RPCS_SS_CNT_ENABLE;
1678 		rpcs |= INTEL_INFO(dev)->subslice_per_slice <<
1679 			GEN8_RPCS_SS_CNT_SHIFT;
1680 		rpcs |= GEN8_RPCS_ENABLE;
1681 	}
1682 
1683 	if (INTEL_INFO(dev)->has_eu_pg) {
1684 		rpcs |= INTEL_INFO(dev)->eu_per_subslice <<
1685 			GEN8_RPCS_EU_MIN_SHIFT;
1686 		rpcs |= INTEL_INFO(dev)->eu_per_subslice <<
1687 			GEN8_RPCS_EU_MAX_SHIFT;
1688 		rpcs |= GEN8_RPCS_ENABLE;
1689 	}
1690 
1691 	return rpcs;
1692 }
1693 
1694 static int
1695 populate_lr_context(struct intel_context *ctx, struct drm_i915_gem_object *ctx_obj,
1696 		    struct intel_engine_cs *ring, struct intel_ringbuffer *ringbuf)
1697 {
1698 	struct drm_device *dev = ring->dev;
1699 	struct drm_i915_private *dev_priv = dev->dev_private;
1700 	struct i915_hw_ppgtt *ppgtt = ctx->ppgtt;
1701 	struct vm_page *page;
1702 	uint32_t *reg_state;
1703 	int ret;
1704 
1705 	if (!ppgtt)
1706 		ppgtt = dev_priv->mm.aliasing_ppgtt;
1707 
1708 	ret = i915_gem_object_set_to_cpu_domain(ctx_obj, true);
1709 	if (ret) {
1710 		DRM_DEBUG_DRIVER("Could not set to CPU domain\n");
1711 		return ret;
1712 	}
1713 
1714 	ret = i915_gem_object_get_pages(ctx_obj);
1715 	if (ret) {
1716 		DRM_DEBUG_DRIVER("Could not get object pages\n");
1717 		return ret;
1718 	}
1719 
1720 	i915_gem_object_pin_pages(ctx_obj);
1721 
1722 	/* The second page of the context object contains some fields which must
1723 	 * be set up prior to the first execution. */
1724 	page = i915_gem_object_get_page(ctx_obj, 1);
1725 	reg_state = kmap_atomic(page);
1726 
1727 	/* A context is actually a big batch buffer with several MI_LOAD_REGISTER_IMM
1728 	 * commands followed by (reg, value) pairs. The values we are setting here are
1729 	 * only for the first context restore: on a subsequent save, the GPU will
1730 	 * recreate this batchbuffer with new values (including all the missing
1731 	 * MI_LOAD_REGISTER_IMM commands that we are not initializing here). */
1732 	if (ring->id == RCS)
1733 		reg_state[CTX_LRI_HEADER_0] = MI_LOAD_REGISTER_IMM(14);
1734 	else
1735 		reg_state[CTX_LRI_HEADER_0] = MI_LOAD_REGISTER_IMM(11);
1736 	reg_state[CTX_LRI_HEADER_0] |= MI_LRI_FORCE_POSTED;
1737 	reg_state[CTX_CONTEXT_CONTROL] = RING_CONTEXT_CONTROL(ring);
1738 	reg_state[CTX_CONTEXT_CONTROL+1] =
1739 		_MASKED_BIT_ENABLE(CTX_CTRL_INHIBIT_SYN_CTX_SWITCH |
1740 				CTX_CTRL_ENGINE_CTX_RESTORE_INHIBIT);
1741 	reg_state[CTX_RING_HEAD] = RING_HEAD(ring->mmio_base);
1742 	reg_state[CTX_RING_HEAD+1] = 0;
1743 	reg_state[CTX_RING_TAIL] = RING_TAIL(ring->mmio_base);
1744 	reg_state[CTX_RING_TAIL+1] = 0;
1745 	reg_state[CTX_RING_BUFFER_START] = RING_START(ring->mmio_base);
1746 	/* Ring buffer start address is not known until the buffer is pinned.
1747 	 * It is written to the context image in execlists_update_context()
1748 	 */
1749 	reg_state[CTX_RING_BUFFER_CONTROL] = RING_CTL(ring->mmio_base);
1750 	reg_state[CTX_RING_BUFFER_CONTROL+1] =
1751 			((ringbuf->size - PAGE_SIZE) & RING_NR_PAGES) | RING_VALID;
1752 	reg_state[CTX_BB_HEAD_U] = ring->mmio_base + 0x168;
1753 	reg_state[CTX_BB_HEAD_U+1] = 0;
1754 	reg_state[CTX_BB_HEAD_L] = ring->mmio_base + 0x140;
1755 	reg_state[CTX_BB_HEAD_L+1] = 0;
1756 	reg_state[CTX_BB_STATE] = ring->mmio_base + 0x110;
1757 	reg_state[CTX_BB_STATE+1] = (1<<5);
1758 	reg_state[CTX_SECOND_BB_HEAD_U] = ring->mmio_base + 0x11c;
1759 	reg_state[CTX_SECOND_BB_HEAD_U+1] = 0;
1760 	reg_state[CTX_SECOND_BB_HEAD_L] = ring->mmio_base + 0x114;
1761 	reg_state[CTX_SECOND_BB_HEAD_L+1] = 0;
1762 	reg_state[CTX_SECOND_BB_STATE] = ring->mmio_base + 0x118;
1763 	reg_state[CTX_SECOND_BB_STATE+1] = 0;
1764 	if (ring->id == RCS) {
1765 		/* TODO: according to BSpec, the register state context
1766 		 * for CHV does not have these. OTOH, these registers do
1767 		 * exist in CHV. I'm waiting for a clarification */
1768 		reg_state[CTX_BB_PER_CTX_PTR] = ring->mmio_base + 0x1c0;
1769 		reg_state[CTX_BB_PER_CTX_PTR+1] = 0;
1770 		reg_state[CTX_RCS_INDIRECT_CTX] = ring->mmio_base + 0x1c4;
1771 		reg_state[CTX_RCS_INDIRECT_CTX+1] = 0;
1772 		reg_state[CTX_RCS_INDIRECT_CTX_OFFSET] = ring->mmio_base + 0x1c8;
1773 		reg_state[CTX_RCS_INDIRECT_CTX_OFFSET+1] = 0;
1774 	}
1775 	reg_state[CTX_LRI_HEADER_1] = MI_LOAD_REGISTER_IMM(9);
1776 	reg_state[CTX_LRI_HEADER_1] |= MI_LRI_FORCE_POSTED;
1777 	reg_state[CTX_CTX_TIMESTAMP] = ring->mmio_base + 0x3a8;
1778 	reg_state[CTX_CTX_TIMESTAMP+1] = 0;
1779 	reg_state[CTX_PDP3_UDW] = GEN8_RING_PDP_UDW(ring, 3);
1780 	reg_state[CTX_PDP3_LDW] = GEN8_RING_PDP_LDW(ring, 3);
1781 	reg_state[CTX_PDP2_UDW] = GEN8_RING_PDP_UDW(ring, 2);
1782 	reg_state[CTX_PDP2_LDW] = GEN8_RING_PDP_LDW(ring, 2);
1783 	reg_state[CTX_PDP1_UDW] = GEN8_RING_PDP_UDW(ring, 1);
1784 	reg_state[CTX_PDP1_LDW] = GEN8_RING_PDP_LDW(ring, 1);
1785 	reg_state[CTX_PDP0_UDW] = GEN8_RING_PDP_UDW(ring, 0);
1786 	reg_state[CTX_PDP0_LDW] = GEN8_RING_PDP_LDW(ring, 0);
1787 
1788 	/* With dynamic page allocation, PDPs may not be allocated at this point,
1789 	 * Point the unallocated PDPs to the scratch page
1790 	 */
1791 	ASSIGN_CTX_PDP(ppgtt, reg_state, 3);
1792 	ASSIGN_CTX_PDP(ppgtt, reg_state, 2);
1793 	ASSIGN_CTX_PDP(ppgtt, reg_state, 1);
1794 	ASSIGN_CTX_PDP(ppgtt, reg_state, 0);
1795 	if (ring->id == RCS) {
1796 		reg_state[CTX_LRI_HEADER_2] = MI_LOAD_REGISTER_IMM(1);
1797 		reg_state[CTX_R_PWR_CLK_STATE] = GEN8_R_PWR_CLK_STATE;
1798 		reg_state[CTX_R_PWR_CLK_STATE+1] = make_rpcs(dev);
1799 	}
1800 
1801 	kunmap_atomic(reg_state);
1802 
1803 	ctx_obj->dirty = 1;
1804 	set_page_dirty(page);
1805 	i915_gem_object_unpin_pages(ctx_obj);
1806 
1807 	return 0;
1808 }
1809 
1810 /**
1811  * intel_lr_context_free() - free the LRC specific bits of a context
1812  * @ctx: the LR context to free.
1813  *
1814  * The real context freeing is done in i915_gem_context_free: this only
1815  * takes care of the bits that are LRC related: the per-engine backing
1816  * objects and the logical ringbuffer.
1817  */
1818 void intel_lr_context_free(struct intel_context *ctx)
1819 {
1820 	int i;
1821 
1822 	for (i = 0; i < I915_NUM_RINGS; i++) {
1823 		struct drm_i915_gem_object *ctx_obj = ctx->engine[i].state;
1824 
1825 		if (ctx_obj) {
1826 			struct intel_ringbuffer *ringbuf =
1827 					ctx->engine[i].ringbuf;
1828 			struct intel_engine_cs *ring = ringbuf->ring;
1829 
1830 			if (ctx == ring->default_context) {
1831 				intel_unpin_ringbuffer_obj(ringbuf);
1832 				i915_gem_object_ggtt_unpin(ctx_obj);
1833 			}
1834 			WARN_ON(ctx->engine[ring->id].pin_count);
1835 			intel_destroy_ringbuffer_obj(ringbuf);
1836 			kfree(ringbuf);
1837 			drm_gem_object_unreference(&ctx_obj->base);
1838 		}
1839 	}
1840 }
1841 
1842 static uint32_t get_lr_context_size(struct intel_engine_cs *ring)
1843 {
1844 	int ret = 0;
1845 
1846 	WARN_ON(INTEL_INFO(ring->dev)->gen < 8);
1847 
1848 	switch (ring->id) {
1849 	case RCS:
1850 		if (INTEL_INFO(ring->dev)->gen >= 9)
1851 			ret = GEN9_LR_CONTEXT_RENDER_SIZE;
1852 		else
1853 			ret = GEN8_LR_CONTEXT_RENDER_SIZE;
1854 		break;
1855 	case VCS:
1856 	case BCS:
1857 	case VECS:
1858 	case VCS2:
1859 		ret = GEN8_LR_CONTEXT_OTHER_SIZE;
1860 		break;
1861 	}
1862 
1863 	return ret;
1864 }
1865 
1866 static void lrc_setup_hardware_status_page(struct intel_engine_cs *ring,
1867 		struct drm_i915_gem_object *default_ctx_obj)
1868 {
1869 	struct drm_i915_private *dev_priv = ring->dev->dev_private;
1870 
1871 	/* The status page is offset 0 from the default context object
1872 	 * in LRC mode. */
1873 	ring->status_page.gfx_addr = i915_gem_obj_ggtt_offset(default_ctx_obj);
1874 	ring->status_page.page_addr =
1875 			kmap(sg_page(default_ctx_obj->pages->sgl));
1876 	ring->status_page.obj = default_ctx_obj;
1877 
1878 	I915_WRITE(RING_HWS_PGA(ring->mmio_base),
1879 			(u32)ring->status_page.gfx_addr);
1880 	POSTING_READ(RING_HWS_PGA(ring->mmio_base));
1881 }
1882 
1883 /**
1884  * intel_lr_context_deferred_create() - create the LRC specific bits of a context
1885  * @ctx: LR context to create.
1886  * @ring: engine to be used with the context.
1887  *
1888  * This function can be called more than once, with different engines, if we plan
1889  * to use the context with them. The context backing objects and the ringbuffers
1890  * (specially the ringbuffer backing objects) suck a lot of memory up, and that's why
1891  * the creation is a deferred call: it's better to make sure first that we need to use
1892  * a given ring with the context.
1893  *
1894  * Return: non-zero on error.
1895  */
1896 int intel_lr_context_deferred_create(struct intel_context *ctx,
1897 				     struct intel_engine_cs *ring)
1898 {
1899 	const bool is_global_default_ctx = (ctx == ring->default_context);
1900 	struct drm_device *dev = ring->dev;
1901 	struct drm_i915_gem_object *ctx_obj;
1902 	uint32_t context_size;
1903 	struct intel_ringbuffer *ringbuf;
1904 	int ret;
1905 
1906 	WARN_ON(ctx->legacy_hw_ctx.rcs_state != NULL);
1907 	WARN_ON(ctx->engine[ring->id].state);
1908 
1909 	context_size = round_up(get_lr_context_size(ring), 4096);
1910 
1911 	ctx_obj = i915_gem_alloc_object(dev, context_size);
1912 	if (!ctx_obj) {
1913 		DRM_DEBUG_DRIVER("Alloc LRC backing obj failed.\n");
1914 		return -ENOMEM;
1915 	}
1916 
1917 	if (is_global_default_ctx) {
1918 		ret = i915_gem_obj_ggtt_pin(ctx_obj, GEN8_LR_CONTEXT_ALIGN, 0);
1919 		if (ret) {
1920 			DRM_DEBUG_DRIVER("Pin LRC backing obj failed: %d\n",
1921 					ret);
1922 			drm_gem_object_unreference(&ctx_obj->base);
1923 			return ret;
1924 		}
1925 	}
1926 
1927 	ringbuf = kzalloc(sizeof(*ringbuf), GFP_KERNEL);
1928 	if (!ringbuf) {
1929 		DRM_DEBUG_DRIVER("Failed to allocate ringbuffer %s\n",
1930 				ring->name);
1931 		ret = -ENOMEM;
1932 		goto error_unpin_ctx;
1933 	}
1934 
1935 	ringbuf->ring = ring;
1936 
1937 	ringbuf->size = 32 * PAGE_SIZE;
1938 	ringbuf->effective_size = ringbuf->size;
1939 	ringbuf->head = 0;
1940 	ringbuf->tail = 0;
1941 	ringbuf->last_retired_head = -1;
1942 	intel_ring_update_space(ringbuf);
1943 
1944 	if (ringbuf->obj == NULL) {
1945 		ret = intel_alloc_ringbuffer_obj(dev, ringbuf);
1946 		if (ret) {
1947 			DRM_DEBUG_DRIVER(
1948 				"Failed to allocate ringbuffer obj %s: %d\n",
1949 				ring->name, ret);
1950 			goto error_free_rbuf;
1951 		}
1952 
1953 		if (is_global_default_ctx) {
1954 			ret = intel_pin_and_map_ringbuffer_obj(dev, ringbuf);
1955 			if (ret) {
1956 				DRM_ERROR(
1957 					"Failed to pin and map ringbuffer %s: %d\n",
1958 					ring->name, ret);
1959 				goto error_destroy_rbuf;
1960 			}
1961 		}
1962 
1963 	}
1964 
1965 	ret = populate_lr_context(ctx, ctx_obj, ring, ringbuf);
1966 	if (ret) {
1967 		DRM_DEBUG_DRIVER("Failed to populate LRC: %d\n", ret);
1968 		goto error;
1969 	}
1970 
1971 	ctx->engine[ring->id].ringbuf = ringbuf;
1972 	ctx->engine[ring->id].state = ctx_obj;
1973 
1974 	if (ctx == ring->default_context)
1975 		lrc_setup_hardware_status_page(ring, ctx_obj);
1976 	else if (ring->id == RCS && !ctx->rcs_initialized) {
1977 		if (ring->init_context) {
1978 			ret = ring->init_context(ring, ctx);
1979 			if (ret) {
1980 				DRM_ERROR("ring init context: %d\n", ret);
1981 				ctx->engine[ring->id].ringbuf = NULL;
1982 				ctx->engine[ring->id].state = NULL;
1983 				goto error;
1984 			}
1985 		}
1986 
1987 		ctx->rcs_initialized = true;
1988 	}
1989 
1990 	return 0;
1991 
1992 error:
1993 	if (is_global_default_ctx)
1994 		intel_unpin_ringbuffer_obj(ringbuf);
1995 error_destroy_rbuf:
1996 	intel_destroy_ringbuffer_obj(ringbuf);
1997 error_free_rbuf:
1998 	kfree(ringbuf);
1999 error_unpin_ctx:
2000 	if (is_global_default_ctx)
2001 		i915_gem_object_ggtt_unpin(ctx_obj);
2002 	drm_gem_object_unreference(&ctx_obj->base);
2003 	return ret;
2004 }
2005 
2006 void intel_lr_context_reset(struct drm_device *dev,
2007 			struct intel_context *ctx)
2008 {
2009 	struct drm_i915_private *dev_priv = dev->dev_private;
2010 	struct intel_engine_cs *ring;
2011 	int i;
2012 
2013 	for_each_ring(ring, dev_priv, i) {
2014 		struct drm_i915_gem_object *ctx_obj =
2015 				ctx->engine[ring->id].state;
2016 		struct intel_ringbuffer *ringbuf =
2017 				ctx->engine[ring->id].ringbuf;
2018 		uint32_t *reg_state;
2019 		struct vm_page *page;
2020 
2021 		if (!ctx_obj)
2022 			continue;
2023 
2024 		if (i915_gem_object_get_pages(ctx_obj)) {
2025 			WARN(1, "Failed get_pages for context obj\n");
2026 			continue;
2027 		}
2028 		page = i915_gem_object_get_page(ctx_obj, 1);
2029 		reg_state = kmap_atomic(page);
2030 
2031 		reg_state[CTX_RING_HEAD+1] = 0;
2032 		reg_state[CTX_RING_TAIL+1] = 0;
2033 
2034 		kunmap_atomic(reg_state);
2035 
2036 		ringbuf->head = 0;
2037 		ringbuf->tail = 0;
2038 	}
2039 }
2040