xref: /openbmc/linux/drivers/gpu/drm/i915/gvt/execlist.c (revision 8cb5d748)
1 /*
2  * Copyright(c) 2011-2016 Intel Corporation. All rights reserved.
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 FROM,
20  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21  * SOFTWARE.
22  *
23  * Authors:
24  *    Zhiyuan Lv <zhiyuan.lv@intel.com>
25  *    Zhi Wang <zhi.a.wang@intel.com>
26  *
27  * Contributors:
28  *    Min He <min.he@intel.com>
29  *    Bing Niu <bing.niu@intel.com>
30  *    Ping Gao <ping.a.gao@intel.com>
31  *    Tina Zhang <tina.zhang@intel.com>
32  *
33  */
34 
35 #include "i915_drv.h"
36 #include "gvt.h"
37 
38 #define _EL_OFFSET_STATUS       0x234
39 #define _EL_OFFSET_STATUS_BUF   0x370
40 #define _EL_OFFSET_STATUS_PTR   0x3A0
41 
42 #define execlist_ring_mmio(gvt, ring_id, offset) \
43 	(gvt->dev_priv->engine[ring_id]->mmio_base + (offset))
44 
45 #define valid_context(ctx) ((ctx)->valid)
46 #define same_context(a, b) (((a)->context_id == (b)->context_id) && \
47 		((a)->lrca == (b)->lrca))
48 
49 static void clean_workloads(struct intel_vgpu *vgpu, unsigned long engine_mask);
50 
51 static int context_switch_events[] = {
52 	[RCS] = RCS_AS_CONTEXT_SWITCH,
53 	[BCS] = BCS_AS_CONTEXT_SWITCH,
54 	[VCS] = VCS_AS_CONTEXT_SWITCH,
55 	[VCS2] = VCS2_AS_CONTEXT_SWITCH,
56 	[VECS] = VECS_AS_CONTEXT_SWITCH,
57 };
58 
59 static int ring_id_to_context_switch_event(int ring_id)
60 {
61 	if (WARN_ON(ring_id < RCS ||
62 		    ring_id >= ARRAY_SIZE(context_switch_events)))
63 		return -EINVAL;
64 
65 	return context_switch_events[ring_id];
66 }
67 
68 static void switch_virtual_execlist_slot(struct intel_vgpu_execlist *execlist)
69 {
70 	gvt_dbg_el("[before] running slot %d/context %x pending slot %d\n",
71 			execlist->running_slot ?
72 			execlist->running_slot->index : -1,
73 			execlist->running_context ?
74 			execlist->running_context->context_id : 0,
75 			execlist->pending_slot ?
76 			execlist->pending_slot->index : -1);
77 
78 	execlist->running_slot = execlist->pending_slot;
79 	execlist->pending_slot = NULL;
80 	execlist->running_context = execlist->running_context ?
81 		&execlist->running_slot->ctx[0] : NULL;
82 
83 	gvt_dbg_el("[after] running slot %d/context %x pending slot %d\n",
84 			execlist->running_slot ?
85 			execlist->running_slot->index : -1,
86 			execlist->running_context ?
87 			execlist->running_context->context_id : 0,
88 			execlist->pending_slot ?
89 			execlist->pending_slot->index : -1);
90 }
91 
92 static void emulate_execlist_status(struct intel_vgpu_execlist *execlist)
93 {
94 	struct intel_vgpu_execlist_slot *running = execlist->running_slot;
95 	struct intel_vgpu_execlist_slot *pending = execlist->pending_slot;
96 	struct execlist_ctx_descriptor_format *desc = execlist->running_context;
97 	struct intel_vgpu *vgpu = execlist->vgpu;
98 	struct execlist_status_format status;
99 	int ring_id = execlist->ring_id;
100 	u32 status_reg = execlist_ring_mmio(vgpu->gvt,
101 			ring_id, _EL_OFFSET_STATUS);
102 
103 	status.ldw = vgpu_vreg(vgpu, status_reg);
104 	status.udw = vgpu_vreg(vgpu, status_reg + 4);
105 
106 	if (running) {
107 		status.current_execlist_pointer = !!running->index;
108 		status.execlist_write_pointer = !!!running->index;
109 		status.execlist_0_active = status.execlist_0_valid =
110 			!!!(running->index);
111 		status.execlist_1_active = status.execlist_1_valid =
112 			!!(running->index);
113 	} else {
114 		status.context_id = 0;
115 		status.execlist_0_active = status.execlist_0_valid = 0;
116 		status.execlist_1_active = status.execlist_1_valid = 0;
117 	}
118 
119 	status.context_id = desc ? desc->context_id : 0;
120 	status.execlist_queue_full = !!(pending);
121 
122 	vgpu_vreg(vgpu, status_reg) = status.ldw;
123 	vgpu_vreg(vgpu, status_reg + 4) = status.udw;
124 
125 	gvt_dbg_el("vgpu%d: status reg offset %x ldw %x udw %x\n",
126 		vgpu->id, status_reg, status.ldw, status.udw);
127 }
128 
129 static void emulate_csb_update(struct intel_vgpu_execlist *execlist,
130 		struct execlist_context_status_format *status,
131 		bool trigger_interrupt_later)
132 {
133 	struct intel_vgpu *vgpu = execlist->vgpu;
134 	int ring_id = execlist->ring_id;
135 	struct execlist_context_status_pointer_format ctx_status_ptr;
136 	u32 write_pointer;
137 	u32 ctx_status_ptr_reg, ctx_status_buf_reg, offset;
138 
139 	ctx_status_ptr_reg = execlist_ring_mmio(vgpu->gvt, ring_id,
140 			_EL_OFFSET_STATUS_PTR);
141 	ctx_status_buf_reg = execlist_ring_mmio(vgpu->gvt, ring_id,
142 			_EL_OFFSET_STATUS_BUF);
143 
144 	ctx_status_ptr.dw = vgpu_vreg(vgpu, ctx_status_ptr_reg);
145 
146 	write_pointer = ctx_status_ptr.write_ptr;
147 
148 	if (write_pointer == 0x7)
149 		write_pointer = 0;
150 	else {
151 		++write_pointer;
152 		write_pointer %= 0x6;
153 	}
154 
155 	offset = ctx_status_buf_reg + write_pointer * 8;
156 
157 	vgpu_vreg(vgpu, offset) = status->ldw;
158 	vgpu_vreg(vgpu, offset + 4) = status->udw;
159 
160 	ctx_status_ptr.write_ptr = write_pointer;
161 	vgpu_vreg(vgpu, ctx_status_ptr_reg) = ctx_status_ptr.dw;
162 
163 	gvt_dbg_el("vgpu%d: w pointer %u reg %x csb l %x csb h %x\n",
164 		vgpu->id, write_pointer, offset, status->ldw, status->udw);
165 
166 	if (trigger_interrupt_later)
167 		return;
168 
169 	intel_vgpu_trigger_virtual_event(vgpu,
170 			ring_id_to_context_switch_event(execlist->ring_id));
171 }
172 
173 static int emulate_execlist_ctx_schedule_out(
174 		struct intel_vgpu_execlist *execlist,
175 		struct execlist_ctx_descriptor_format *ctx)
176 {
177 	struct intel_vgpu *vgpu = execlist->vgpu;
178 	struct intel_vgpu_execlist_slot *running = execlist->running_slot;
179 	struct intel_vgpu_execlist_slot *pending = execlist->pending_slot;
180 	struct execlist_ctx_descriptor_format *ctx0 = &running->ctx[0];
181 	struct execlist_ctx_descriptor_format *ctx1 = &running->ctx[1];
182 	struct execlist_context_status_format status;
183 
184 	memset(&status, 0, sizeof(status));
185 
186 	gvt_dbg_el("schedule out context id %x\n", ctx->context_id);
187 
188 	if (WARN_ON(!same_context(ctx, execlist->running_context))) {
189 		gvt_vgpu_err("schedule out context is not running context,"
190 				"ctx id %x running ctx id %x\n",
191 				ctx->context_id,
192 				execlist->running_context->context_id);
193 		return -EINVAL;
194 	}
195 
196 	/* ctx1 is valid, ctx0/ctx is scheduled-out -> element switch */
197 	if (valid_context(ctx1) && same_context(ctx0, ctx)) {
198 		gvt_dbg_el("ctx 1 valid, ctx/ctx 0 is scheduled-out\n");
199 
200 		execlist->running_context = ctx1;
201 
202 		emulate_execlist_status(execlist);
203 
204 		status.context_complete = status.element_switch = 1;
205 		status.context_id = ctx->context_id;
206 
207 		emulate_csb_update(execlist, &status, false);
208 		/*
209 		 * ctx1 is not valid, ctx == ctx0
210 		 * ctx1 is valid, ctx1 == ctx
211 		 *	--> last element is finished
212 		 * emulate:
213 		 *	active-to-idle if there is *no* pending execlist
214 		 *	context-complete if there *is* pending execlist
215 		 */
216 	} else if ((!valid_context(ctx1) && same_context(ctx0, ctx))
217 			|| (valid_context(ctx1) && same_context(ctx1, ctx))) {
218 		gvt_dbg_el("need to switch virtual execlist slot\n");
219 
220 		switch_virtual_execlist_slot(execlist);
221 
222 		emulate_execlist_status(execlist);
223 
224 		status.context_complete = status.active_to_idle = 1;
225 		status.context_id = ctx->context_id;
226 
227 		if (!pending) {
228 			emulate_csb_update(execlist, &status, false);
229 		} else {
230 			emulate_csb_update(execlist, &status, true);
231 
232 			memset(&status, 0, sizeof(status));
233 
234 			status.idle_to_active = 1;
235 			status.context_id = 0;
236 
237 			emulate_csb_update(execlist, &status, false);
238 		}
239 	} else {
240 		WARN_ON(1);
241 		return -EINVAL;
242 	}
243 
244 	return 0;
245 }
246 
247 static struct intel_vgpu_execlist_slot *get_next_execlist_slot(
248 		struct intel_vgpu_execlist *execlist)
249 {
250 	struct intel_vgpu *vgpu = execlist->vgpu;
251 	int ring_id = execlist->ring_id;
252 	u32 status_reg = execlist_ring_mmio(vgpu->gvt, ring_id,
253 			_EL_OFFSET_STATUS);
254 	struct execlist_status_format status;
255 
256 	status.ldw = vgpu_vreg(vgpu, status_reg);
257 	status.udw = vgpu_vreg(vgpu, status_reg + 4);
258 
259 	if (status.execlist_queue_full) {
260 		gvt_vgpu_err("virtual execlist slots are full\n");
261 		return NULL;
262 	}
263 
264 	return &execlist->slot[status.execlist_write_pointer];
265 }
266 
267 static int emulate_execlist_schedule_in(struct intel_vgpu_execlist *execlist,
268 		struct execlist_ctx_descriptor_format ctx[2])
269 {
270 	struct intel_vgpu_execlist_slot *running = execlist->running_slot;
271 	struct intel_vgpu_execlist_slot *slot =
272 		get_next_execlist_slot(execlist);
273 
274 	struct execlist_ctx_descriptor_format *ctx0, *ctx1;
275 	struct execlist_context_status_format status;
276 	struct intel_vgpu *vgpu = execlist->vgpu;
277 
278 	gvt_dbg_el("emulate schedule-in\n");
279 
280 	if (!slot) {
281 		gvt_vgpu_err("no available execlist slot\n");
282 		return -EINVAL;
283 	}
284 
285 	memset(&status, 0, sizeof(status));
286 	memset(slot->ctx, 0, sizeof(slot->ctx));
287 
288 	slot->ctx[0] = ctx[0];
289 	slot->ctx[1] = ctx[1];
290 
291 	gvt_dbg_el("alloc slot index %d ctx 0 %x ctx 1 %x\n",
292 			slot->index, ctx[0].context_id,
293 			ctx[1].context_id);
294 
295 	/*
296 	 * no running execlist, make this write bundle as running execlist
297 	 * -> idle-to-active
298 	 */
299 	if (!running) {
300 		gvt_dbg_el("no current running execlist\n");
301 
302 		execlist->running_slot = slot;
303 		execlist->pending_slot = NULL;
304 		execlist->running_context = &slot->ctx[0];
305 
306 		gvt_dbg_el("running slot index %d running context %x\n",
307 				execlist->running_slot->index,
308 				execlist->running_context->context_id);
309 
310 		emulate_execlist_status(execlist);
311 
312 		status.idle_to_active = 1;
313 		status.context_id = 0;
314 
315 		emulate_csb_update(execlist, &status, false);
316 		return 0;
317 	}
318 
319 	ctx0 = &running->ctx[0];
320 	ctx1 = &running->ctx[1];
321 
322 	gvt_dbg_el("current running slot index %d ctx 0 %x ctx 1 %x\n",
323 		running->index, ctx0->context_id, ctx1->context_id);
324 
325 	/*
326 	 * already has an running execlist
327 	 *	a. running ctx1 is valid,
328 	 *	   ctx0 is finished, and running ctx1 == new execlist ctx[0]
329 	 *	b. running ctx1 is not valid,
330 	 *	   ctx0 == new execlist ctx[0]
331 	 * ----> lite-restore + preempted
332 	 */
333 	if ((valid_context(ctx1) && same_context(ctx1, &slot->ctx[0]) &&
334 		/* condition a */
335 		(!same_context(ctx0, execlist->running_context))) ||
336 			(!valid_context(ctx1) &&
337 			 same_context(ctx0, &slot->ctx[0]))) { /* condition b */
338 		gvt_dbg_el("need to switch virtual execlist slot\n");
339 
340 		execlist->pending_slot = slot;
341 		switch_virtual_execlist_slot(execlist);
342 
343 		emulate_execlist_status(execlist);
344 
345 		status.lite_restore = status.preempted = 1;
346 		status.context_id = ctx[0].context_id;
347 
348 		emulate_csb_update(execlist, &status, false);
349 	} else {
350 		gvt_dbg_el("emulate as pending slot\n");
351 		/*
352 		 * otherwise
353 		 * --> emulate pending execlist exist + but no preemption case
354 		 */
355 		execlist->pending_slot = slot;
356 		emulate_execlist_status(execlist);
357 	}
358 	return 0;
359 }
360 
361 static void free_workload(struct intel_vgpu_workload *workload)
362 {
363 	intel_vgpu_unpin_mm(workload->shadow_mm);
364 	intel_gvt_mm_unreference(workload->shadow_mm);
365 	kmem_cache_free(workload->vgpu->workloads, workload);
366 }
367 
368 #define get_desc_from_elsp_dwords(ed, i) \
369 	((struct execlist_ctx_descriptor_format *)&((ed)->data[i * 2]))
370 
371 static void prepare_shadow_batch_buffer(struct intel_vgpu_workload *workload)
372 {
373 	const int gmadr_bytes = workload->vgpu->gvt->device_info.gmadr_bytes_in_cmd;
374 	struct intel_shadow_bb_entry *entry_obj;
375 
376 	/* pin the gem object to ggtt */
377 	list_for_each_entry(entry_obj, &workload->shadow_bb, list) {
378 		struct i915_vma *vma;
379 
380 		vma = i915_gem_object_ggtt_pin(entry_obj->obj, NULL, 0, 4, 0);
381 		if (IS_ERR(vma)) {
382 			return;
383 		}
384 
385 		/* FIXME: we are not tracking our pinned VMA leaving it
386 		 * up to the core to fix up the stray pin_count upon
387 		 * free.
388 		 */
389 
390 		/* update the relocate gma with shadow batch buffer*/
391 		entry_obj->bb_start_cmd_va[1] = i915_ggtt_offset(vma);
392 		if (gmadr_bytes == 8)
393 			entry_obj->bb_start_cmd_va[2] = 0;
394 	}
395 }
396 
397 static int update_wa_ctx_2_shadow_ctx(struct intel_shadow_wa_ctx *wa_ctx)
398 {
399 	struct intel_vgpu_workload *workload = container_of(wa_ctx,
400 					struct intel_vgpu_workload,
401 					wa_ctx);
402 	int ring_id = workload->ring_id;
403 	struct i915_gem_context *shadow_ctx = workload->vgpu->shadow_ctx;
404 	struct drm_i915_gem_object *ctx_obj =
405 		shadow_ctx->engine[ring_id].state->obj;
406 	struct execlist_ring_context *shadow_ring_context;
407 	struct page *page;
408 
409 	page = i915_gem_object_get_page(ctx_obj, LRC_STATE_PN);
410 	shadow_ring_context = kmap_atomic(page);
411 
412 	shadow_ring_context->bb_per_ctx_ptr.val =
413 		(shadow_ring_context->bb_per_ctx_ptr.val &
414 		(~PER_CTX_ADDR_MASK)) | wa_ctx->per_ctx.shadow_gma;
415 	shadow_ring_context->rcs_indirect_ctx.val =
416 		(shadow_ring_context->rcs_indirect_ctx.val &
417 		(~INDIRECT_CTX_ADDR_MASK)) | wa_ctx->indirect_ctx.shadow_gma;
418 
419 	kunmap_atomic(shadow_ring_context);
420 	return 0;
421 }
422 
423 static void prepare_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
424 {
425 	struct i915_vma *vma;
426 	unsigned char *per_ctx_va =
427 		(unsigned char *)wa_ctx->indirect_ctx.shadow_va +
428 		wa_ctx->indirect_ctx.size;
429 
430 	if (wa_ctx->indirect_ctx.size == 0)
431 		return;
432 
433 	vma = i915_gem_object_ggtt_pin(wa_ctx->indirect_ctx.obj, NULL,
434 				       0, CACHELINE_BYTES, 0);
435 	if (IS_ERR(vma)) {
436 		return;
437 	}
438 
439 	/* FIXME: we are not tracking our pinned VMA leaving it
440 	 * up to the core to fix up the stray pin_count upon
441 	 * free.
442 	 */
443 
444 	wa_ctx->indirect_ctx.shadow_gma = i915_ggtt_offset(vma);
445 
446 	wa_ctx->per_ctx.shadow_gma = *((unsigned int *)per_ctx_va + 1);
447 	memset(per_ctx_va, 0, CACHELINE_BYTES);
448 
449 	update_wa_ctx_2_shadow_ctx(wa_ctx);
450 }
451 
452 static int prepare_execlist_workload(struct intel_vgpu_workload *workload)
453 {
454 	struct intel_vgpu *vgpu = workload->vgpu;
455 	struct execlist_ctx_descriptor_format ctx[2];
456 	int ring_id = workload->ring_id;
457 
458 	intel_vgpu_pin_mm(workload->shadow_mm);
459 	intel_vgpu_sync_oos_pages(workload->vgpu);
460 	intel_vgpu_flush_post_shadow(workload->vgpu);
461 	prepare_shadow_batch_buffer(workload);
462 	prepare_shadow_wa_ctx(&workload->wa_ctx);
463 	if (!workload->emulate_schedule_in)
464 		return 0;
465 
466 	ctx[0] = *get_desc_from_elsp_dwords(&workload->elsp_dwords, 1);
467 	ctx[1] = *get_desc_from_elsp_dwords(&workload->elsp_dwords, 0);
468 
469 	return emulate_execlist_schedule_in(&vgpu->execlist[ring_id], ctx);
470 }
471 
472 static void release_shadow_batch_buffer(struct intel_vgpu_workload *workload)
473 {
474 	/* release all the shadow batch buffer */
475 	if (!list_empty(&workload->shadow_bb)) {
476 		struct intel_shadow_bb_entry *entry_obj =
477 			list_first_entry(&workload->shadow_bb,
478 					 struct intel_shadow_bb_entry,
479 					 list);
480 		struct intel_shadow_bb_entry *temp;
481 
482 		list_for_each_entry_safe(entry_obj, temp, &workload->shadow_bb,
483 					 list) {
484 			i915_gem_object_unpin_map(entry_obj->obj);
485 			i915_gem_object_put(entry_obj->obj);
486 			list_del(&entry_obj->list);
487 			kfree(entry_obj);
488 		}
489 	}
490 }
491 
492 static void release_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
493 {
494 	if (!wa_ctx->indirect_ctx.obj)
495 		return;
496 
497 	i915_gem_object_unpin_map(wa_ctx->indirect_ctx.obj);
498 	i915_gem_object_put(wa_ctx->indirect_ctx.obj);
499 }
500 
501 static int complete_execlist_workload(struct intel_vgpu_workload *workload)
502 {
503 	struct intel_vgpu *vgpu = workload->vgpu;
504 	int ring_id = workload->ring_id;
505 	struct intel_vgpu_execlist *execlist = &vgpu->execlist[ring_id];
506 	struct intel_vgpu_workload *next_workload;
507 	struct list_head *next = workload_q_head(vgpu, ring_id)->next;
508 	bool lite_restore = false;
509 	int ret;
510 
511 	gvt_dbg_el("complete workload %p status %d\n", workload,
512 			workload->status);
513 
514 	release_shadow_batch_buffer(workload);
515 	release_shadow_wa_ctx(&workload->wa_ctx);
516 
517 	if (workload->status || (vgpu->resetting_eng & ENGINE_MASK(ring_id))) {
518 		/* if workload->status is not successful means HW GPU
519 		 * has occurred GPU hang or something wrong with i915/GVT,
520 		 * and GVT won't inject context switch interrupt to guest.
521 		 * So this error is a vGPU hang actually to the guest.
522 		 * According to this we should emunlate a vGPU hang. If
523 		 * there are pending workloads which are already submitted
524 		 * from guest, we should clean them up like HW GPU does.
525 		 *
526 		 * if it is in middle of engine resetting, the pending
527 		 * workloads won't be submitted to HW GPU and will be
528 		 * cleaned up during the resetting process later, so doing
529 		 * the workload clean up here doesn't have any impact.
530 		 **/
531 		clean_workloads(vgpu, ENGINE_MASK(ring_id));
532 		goto out;
533 	}
534 
535 	if (!list_empty(workload_q_head(vgpu, ring_id))) {
536 		struct execlist_ctx_descriptor_format *this_desc, *next_desc;
537 
538 		next_workload = container_of(next,
539 				struct intel_vgpu_workload, list);
540 		this_desc = &workload->ctx_desc;
541 		next_desc = &next_workload->ctx_desc;
542 
543 		lite_restore = same_context(this_desc, next_desc);
544 	}
545 
546 	if (lite_restore) {
547 		gvt_dbg_el("next context == current - no schedule-out\n");
548 		free_workload(workload);
549 		return 0;
550 	}
551 
552 	ret = emulate_execlist_ctx_schedule_out(execlist, &workload->ctx_desc);
553 	if (ret)
554 		goto err;
555 out:
556 	free_workload(workload);
557 	return 0;
558 err:
559 	free_workload(workload);
560 	return ret;
561 }
562 
563 #define RING_CTX_OFF(x) \
564 	offsetof(struct execlist_ring_context, x)
565 
566 static void read_guest_pdps(struct intel_vgpu *vgpu,
567 		u64 ring_context_gpa, u32 pdp[8])
568 {
569 	u64 gpa;
570 	int i;
571 
572 	gpa = ring_context_gpa + RING_CTX_OFF(pdp3_UDW.val);
573 
574 	for (i = 0; i < 8; i++)
575 		intel_gvt_hypervisor_read_gpa(vgpu,
576 				gpa + i * 8, &pdp[7 - i], 4);
577 }
578 
579 static int prepare_mm(struct intel_vgpu_workload *workload)
580 {
581 	struct execlist_ctx_descriptor_format *desc = &workload->ctx_desc;
582 	struct intel_vgpu_mm *mm;
583 	struct intel_vgpu *vgpu = workload->vgpu;
584 	int page_table_level;
585 	u32 pdp[8];
586 
587 	if (desc->addressing_mode == 1) { /* legacy 32-bit */
588 		page_table_level = 3;
589 	} else if (desc->addressing_mode == 3) { /* legacy 64 bit */
590 		page_table_level = 4;
591 	} else {
592 		gvt_vgpu_err("Advanced Context mode(SVM) is not supported!\n");
593 		return -EINVAL;
594 	}
595 
596 	read_guest_pdps(workload->vgpu, workload->ring_context_gpa, pdp);
597 
598 	mm = intel_vgpu_find_ppgtt_mm(workload->vgpu, page_table_level, pdp);
599 	if (mm) {
600 		intel_gvt_mm_reference(mm);
601 	} else {
602 
603 		mm = intel_vgpu_create_mm(workload->vgpu, INTEL_GVT_MM_PPGTT,
604 				pdp, page_table_level, 0);
605 		if (IS_ERR(mm)) {
606 			gvt_vgpu_err("fail to create mm object.\n");
607 			return PTR_ERR(mm);
608 		}
609 	}
610 	workload->shadow_mm = mm;
611 	return 0;
612 }
613 
614 #define get_last_workload(q) \
615 	(list_empty(q) ? NULL : container_of(q->prev, \
616 	struct intel_vgpu_workload, list))
617 
618 static int submit_context(struct intel_vgpu *vgpu, int ring_id,
619 		struct execlist_ctx_descriptor_format *desc,
620 		bool emulate_schedule_in)
621 {
622 	struct list_head *q = workload_q_head(vgpu, ring_id);
623 	struct intel_vgpu_workload *last_workload = get_last_workload(q);
624 	struct intel_vgpu_workload *workload = NULL;
625 	struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
626 	u64 ring_context_gpa;
627 	u32 head, tail, start, ctl, ctx_ctl, per_ctx, indirect_ctx;
628 	int ret;
629 
630 	ring_context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
631 			(u32)((desc->lrca + 1) << GTT_PAGE_SHIFT));
632 	if (ring_context_gpa == INTEL_GVT_INVALID_ADDR) {
633 		gvt_vgpu_err("invalid guest context LRCA: %x\n", desc->lrca);
634 		return -EINVAL;
635 	}
636 
637 	intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
638 			RING_CTX_OFF(ring_header.val), &head, 4);
639 
640 	intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
641 			RING_CTX_OFF(ring_tail.val), &tail, 4);
642 
643 	head &= RB_HEAD_OFF_MASK;
644 	tail &= RB_TAIL_OFF_MASK;
645 
646 	if (last_workload && same_context(&last_workload->ctx_desc, desc)) {
647 		gvt_dbg_el("ring id %d cur workload == last\n", ring_id);
648 		gvt_dbg_el("ctx head %x real head %lx\n", head,
649 				last_workload->rb_tail);
650 		/*
651 		 * cannot use guest context head pointer here,
652 		 * as it might not be updated at this time
653 		 */
654 		head = last_workload->rb_tail;
655 	}
656 
657 	gvt_dbg_el("ring id %d begin a new workload\n", ring_id);
658 
659 	workload = kmem_cache_zalloc(vgpu->workloads, GFP_KERNEL);
660 	if (!workload)
661 		return -ENOMEM;
662 
663 	/* record some ring buffer register values for scan and shadow */
664 	intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
665 			RING_CTX_OFF(rb_start.val), &start, 4);
666 	intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
667 			RING_CTX_OFF(rb_ctrl.val), &ctl, 4);
668 	intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
669 			RING_CTX_OFF(ctx_ctrl.val), &ctx_ctl, 4);
670 
671 	INIT_LIST_HEAD(&workload->list);
672 	INIT_LIST_HEAD(&workload->shadow_bb);
673 
674 	init_waitqueue_head(&workload->shadow_ctx_status_wq);
675 	atomic_set(&workload->shadow_ctx_active, 0);
676 
677 	workload->vgpu = vgpu;
678 	workload->ring_id = ring_id;
679 	workload->ctx_desc = *desc;
680 	workload->ring_context_gpa = ring_context_gpa;
681 	workload->rb_head = head;
682 	workload->rb_tail = tail;
683 	workload->rb_start = start;
684 	workload->rb_ctl = ctl;
685 	workload->prepare = prepare_execlist_workload;
686 	workload->complete = complete_execlist_workload;
687 	workload->status = -EINPROGRESS;
688 	workload->emulate_schedule_in = emulate_schedule_in;
689 	workload->shadowed = false;
690 
691 	if (ring_id == RCS) {
692 		intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
693 			RING_CTX_OFF(bb_per_ctx_ptr.val), &per_ctx, 4);
694 		intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
695 			RING_CTX_OFF(rcs_indirect_ctx.val), &indirect_ctx, 4);
696 
697 		workload->wa_ctx.indirect_ctx.guest_gma =
698 			indirect_ctx & INDIRECT_CTX_ADDR_MASK;
699 		workload->wa_ctx.indirect_ctx.size =
700 			(indirect_ctx & INDIRECT_CTX_SIZE_MASK) *
701 			CACHELINE_BYTES;
702 		workload->wa_ctx.per_ctx.guest_gma =
703 			per_ctx & PER_CTX_ADDR_MASK;
704 
705 		WARN_ON(workload->wa_ctx.indirect_ctx.size && !(per_ctx & 0x1));
706 	}
707 
708 	if (emulate_schedule_in)
709 		workload->elsp_dwords = vgpu->execlist[ring_id].elsp_dwords;
710 
711 	gvt_dbg_el("workload %p ring id %d head %x tail %x start %x ctl %x\n",
712 			workload, ring_id, head, tail, start, ctl);
713 
714 	gvt_dbg_el("workload %p emulate schedule_in %d\n", workload,
715 			emulate_schedule_in);
716 
717 	ret = prepare_mm(workload);
718 	if (ret) {
719 		kmem_cache_free(vgpu->workloads, workload);
720 		return ret;
721 	}
722 
723 	/* Only scan and shadow the first workload in the queue
724 	 * as there is only one pre-allocated buf-obj for shadow.
725 	 */
726 	if (list_empty(workload_q_head(vgpu, ring_id))) {
727 		intel_runtime_pm_get(dev_priv);
728 		mutex_lock(&dev_priv->drm.struct_mutex);
729 		intel_gvt_scan_and_shadow_workload(workload);
730 		mutex_unlock(&dev_priv->drm.struct_mutex);
731 		intel_runtime_pm_put(dev_priv);
732 	}
733 
734 	queue_workload(workload);
735 	return 0;
736 }
737 
738 int intel_vgpu_submit_execlist(struct intel_vgpu *vgpu, int ring_id)
739 {
740 	struct intel_vgpu_execlist *execlist = &vgpu->execlist[ring_id];
741 	struct execlist_ctx_descriptor_format desc[2];
742 	int i, ret;
743 
744 	desc[0] = *get_desc_from_elsp_dwords(&execlist->elsp_dwords, 1);
745 	desc[1] = *get_desc_from_elsp_dwords(&execlist->elsp_dwords, 0);
746 
747 	if (!desc[0].valid) {
748 		gvt_vgpu_err("invalid elsp submission, desc0 is invalid\n");
749 		goto inv_desc;
750 	}
751 
752 	for (i = 0; i < ARRAY_SIZE(desc); i++) {
753 		if (!desc[i].valid)
754 			continue;
755 		if (!desc[i].privilege_access) {
756 			gvt_vgpu_err("unexpected GGTT elsp submission\n");
757 			goto inv_desc;
758 		}
759 	}
760 
761 	/* submit workload */
762 	for (i = 0; i < ARRAY_SIZE(desc); i++) {
763 		if (!desc[i].valid)
764 			continue;
765 		ret = submit_context(vgpu, ring_id, &desc[i], i == 0);
766 		if (ret) {
767 			gvt_vgpu_err("failed to submit desc %d\n", i);
768 			return ret;
769 		}
770 	}
771 
772 	return 0;
773 
774 inv_desc:
775 	gvt_vgpu_err("descriptors content: desc0 %08x %08x desc1 %08x %08x\n",
776 		     desc[0].udw, desc[0].ldw, desc[1].udw, desc[1].ldw);
777 	return -EINVAL;
778 }
779 
780 static void init_vgpu_execlist(struct intel_vgpu *vgpu, int ring_id)
781 {
782 	struct intel_vgpu_execlist *execlist = &vgpu->execlist[ring_id];
783 	struct execlist_context_status_pointer_format ctx_status_ptr;
784 	u32 ctx_status_ptr_reg;
785 
786 	memset(execlist, 0, sizeof(*execlist));
787 
788 	execlist->vgpu = vgpu;
789 	execlist->ring_id = ring_id;
790 	execlist->slot[0].index = 0;
791 	execlist->slot[1].index = 1;
792 
793 	ctx_status_ptr_reg = execlist_ring_mmio(vgpu->gvt, ring_id,
794 			_EL_OFFSET_STATUS_PTR);
795 
796 	ctx_status_ptr.dw = vgpu_vreg(vgpu, ctx_status_ptr_reg);
797 	ctx_status_ptr.read_ptr = 0;
798 	ctx_status_ptr.write_ptr = 0x7;
799 	vgpu_vreg(vgpu, ctx_status_ptr_reg) = ctx_status_ptr.dw;
800 }
801 
802 static void clean_workloads(struct intel_vgpu *vgpu, unsigned long engine_mask)
803 {
804 	struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
805 	struct intel_engine_cs *engine;
806 	struct intel_vgpu_workload *pos, *n;
807 	unsigned int tmp;
808 
809 	/* free the unsubmited workloads in the queues. */
810 	for_each_engine_masked(engine, dev_priv, engine_mask, tmp) {
811 		list_for_each_entry_safe(pos, n,
812 			&vgpu->workload_q_head[engine->id], list) {
813 			list_del_init(&pos->list);
814 			free_workload(pos);
815 		}
816 
817 		clear_bit(engine->id, vgpu->shadow_ctx_desc_updated);
818 	}
819 }
820 
821 void intel_vgpu_clean_execlist(struct intel_vgpu *vgpu)
822 {
823 	clean_workloads(vgpu, ALL_ENGINES);
824 	kmem_cache_destroy(vgpu->workloads);
825 }
826 
827 int intel_vgpu_init_execlist(struct intel_vgpu *vgpu)
828 {
829 	enum intel_engine_id i;
830 	struct intel_engine_cs *engine;
831 
832 	/* each ring has a virtual execlist engine */
833 	for_each_engine(engine, vgpu->gvt->dev_priv, i) {
834 		init_vgpu_execlist(vgpu, i);
835 		INIT_LIST_HEAD(&vgpu->workload_q_head[i]);
836 	}
837 
838 	vgpu->workloads = kmem_cache_create("gvt-g_vgpu_workload",
839 			sizeof(struct intel_vgpu_workload), 0,
840 			SLAB_HWCACHE_ALIGN,
841 			NULL);
842 
843 	if (!vgpu->workloads)
844 		return -ENOMEM;
845 
846 	return 0;
847 }
848 
849 void intel_vgpu_reset_execlist(struct intel_vgpu *vgpu,
850 		unsigned long engine_mask)
851 {
852 	struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
853 	struct intel_engine_cs *engine;
854 	unsigned int tmp;
855 
856 	clean_workloads(vgpu, engine_mask);
857 	for_each_engine_masked(engine, dev_priv, engine_mask, tmp)
858 		init_vgpu_execlist(vgpu, engine->id);
859 }
860