2 * Copyright(c) 2011-2016 Intel Corporation. All rights reserved.
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
24 * Zhiyuan Lv <zhiyuan.lv@intel.com>
25 * Zhi Wang <zhi.a.wang@intel.com>
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>
38 #define _EL_OFFSET_STATUS 0x234
39 #define _EL_OFFSET_STATUS_BUF 0x370
40 #define _EL_OFFSET_STATUS_PTR 0x3A0
42 #define execlist_ring_mmio(gvt, ring_id, offset) \
43 (gvt->dev_priv->engine[ring_id]->mmio_base + (offset))
45 #define valid_context(ctx) ((ctx)->valid)
46 #define same_context(a, b) (((a)->context_id == (b)->context_id) && \
47 ((a)->lrca == (b)->lrca))
49 static int context_switch_events[] = {
50 [RCS] = RCS_AS_CONTEXT_SWITCH,
51 [BCS] = BCS_AS_CONTEXT_SWITCH,
52 [VCS] = VCS_AS_CONTEXT_SWITCH,
53 [VCS2] = VCS2_AS_CONTEXT_SWITCH,
54 [VECS] = VECS_AS_CONTEXT_SWITCH,
57 static int ring_id_to_context_switch_event(int ring_id)
59 if (WARN_ON(ring_id < RCS && ring_id >
60 ARRAY_SIZE(context_switch_events)))
63 return context_switch_events[ring_id];
66 static void switch_virtual_execlist_slot(struct intel_vgpu_execlist *execlist)
68 gvt_dbg_el("[before] running slot %d/context %x pending slot %d\n",
69 execlist->running_slot ?
70 execlist->running_slot->index : -1,
71 execlist->running_context ?
72 execlist->running_context->context_id : 0,
73 execlist->pending_slot ?
74 execlist->pending_slot->index : -1);
76 execlist->running_slot = execlist->pending_slot;
77 execlist->pending_slot = NULL;
78 execlist->running_context = execlist->running_context ?
79 &execlist->running_slot->ctx[0] : NULL;
81 gvt_dbg_el("[after] running slot %d/context %x pending slot %d\n",
82 execlist->running_slot ?
83 execlist->running_slot->index : -1,
84 execlist->running_context ?
85 execlist->running_context->context_id : 0,
86 execlist->pending_slot ?
87 execlist->pending_slot->index : -1);
90 static void emulate_execlist_status(struct intel_vgpu_execlist *execlist)
92 struct intel_vgpu_execlist_slot *running = execlist->running_slot;
93 struct intel_vgpu_execlist_slot *pending = execlist->pending_slot;
94 struct execlist_ctx_descriptor_format *desc = execlist->running_context;
95 struct intel_vgpu *vgpu = execlist->vgpu;
96 struct execlist_status_format status;
97 int ring_id = execlist->ring_id;
98 u32 status_reg = execlist_ring_mmio(vgpu->gvt,
99 ring_id, _EL_OFFSET_STATUS);
101 status.ldw = vgpu_vreg(vgpu, status_reg);
102 status.udw = vgpu_vreg(vgpu, status_reg + 4);
105 status.current_execlist_pointer = !!running->index;
106 status.execlist_write_pointer = !!!running->index;
107 status.execlist_0_active = status.execlist_0_valid =
109 status.execlist_1_active = status.execlist_1_valid =
112 status.context_id = 0;
113 status.execlist_0_active = status.execlist_0_valid = 0;
114 status.execlist_1_active = status.execlist_1_valid = 0;
117 status.context_id = desc ? desc->context_id : 0;
118 status.execlist_queue_full = !!(pending);
120 vgpu_vreg(vgpu, status_reg) = status.ldw;
121 vgpu_vreg(vgpu, status_reg + 4) = status.udw;
123 gvt_dbg_el("vgpu%d: status reg offset %x ldw %x udw %x\n",
124 vgpu->id, status_reg, status.ldw, status.udw);
127 static void emulate_csb_update(struct intel_vgpu_execlist *execlist,
128 struct execlist_context_status_format *status,
129 bool trigger_interrupt_later)
131 struct intel_vgpu *vgpu = execlist->vgpu;
132 int ring_id = execlist->ring_id;
133 struct execlist_context_status_pointer_format ctx_status_ptr;
135 u32 ctx_status_ptr_reg, ctx_status_buf_reg, offset;
137 ctx_status_ptr_reg = execlist_ring_mmio(vgpu->gvt, ring_id,
138 _EL_OFFSET_STATUS_PTR);
139 ctx_status_buf_reg = execlist_ring_mmio(vgpu->gvt, ring_id,
140 _EL_OFFSET_STATUS_BUF);
142 ctx_status_ptr.dw = vgpu_vreg(vgpu, ctx_status_ptr_reg);
144 write_pointer = ctx_status_ptr.write_ptr;
146 if (write_pointer == 0x7)
150 write_pointer %= 0x6;
153 offset = ctx_status_buf_reg + write_pointer * 8;
155 vgpu_vreg(vgpu, offset) = status->ldw;
156 vgpu_vreg(vgpu, offset + 4) = status->udw;
158 ctx_status_ptr.write_ptr = write_pointer;
159 vgpu_vreg(vgpu, ctx_status_ptr_reg) = ctx_status_ptr.dw;
161 gvt_dbg_el("vgpu%d: w pointer %u reg %x csb l %x csb h %x\n",
162 vgpu->id, write_pointer, offset, status->ldw, status->udw);
164 if (trigger_interrupt_later)
167 intel_vgpu_trigger_virtual_event(vgpu,
168 ring_id_to_context_switch_event(execlist->ring_id));
171 static int emulate_execlist_ctx_schedule_out(
172 struct intel_vgpu_execlist *execlist,
173 struct execlist_ctx_descriptor_format *ctx)
175 struct intel_vgpu_execlist_slot *running = execlist->running_slot;
176 struct intel_vgpu_execlist_slot *pending = execlist->pending_slot;
177 struct execlist_ctx_descriptor_format *ctx0 = &running->ctx[0];
178 struct execlist_ctx_descriptor_format *ctx1 = &running->ctx[1];
179 struct execlist_context_status_format status;
181 memset(&status, 0, sizeof(status));
183 gvt_dbg_el("schedule out context id %x\n", ctx->context_id);
185 if (WARN_ON(!same_context(ctx, execlist->running_context))) {
186 gvt_err("schedule out context is not running context,"
187 "ctx id %x running ctx id %x\n",
189 execlist->running_context->context_id);
193 /* ctx1 is valid, ctx0/ctx is scheduled-out -> element switch */
194 if (valid_context(ctx1) && same_context(ctx0, ctx)) {
195 gvt_dbg_el("ctx 1 valid, ctx/ctx 0 is scheduled-out\n");
197 execlist->running_context = ctx1;
199 emulate_execlist_status(execlist);
201 status.context_complete = status.element_switch = 1;
202 status.context_id = ctx->context_id;
204 emulate_csb_update(execlist, &status, false);
206 * ctx1 is not valid, ctx == ctx0
207 * ctx1 is valid, ctx1 == ctx
208 * --> last element is finished
210 * active-to-idle if there is *no* pending execlist
211 * context-complete if there *is* pending execlist
213 } else if ((!valid_context(ctx1) && same_context(ctx0, ctx))
214 || (valid_context(ctx1) && same_context(ctx1, ctx))) {
215 gvt_dbg_el("need to switch virtual execlist slot\n");
217 switch_virtual_execlist_slot(execlist);
219 emulate_execlist_status(execlist);
221 status.context_complete = status.active_to_idle = 1;
222 status.context_id = ctx->context_id;
225 emulate_csb_update(execlist, &status, false);
227 emulate_csb_update(execlist, &status, true);
229 memset(&status, 0, sizeof(status));
231 status.idle_to_active = 1;
232 status.context_id = 0;
234 emulate_csb_update(execlist, &status, false);
244 static struct intel_vgpu_execlist_slot *get_next_execlist_slot(
245 struct intel_vgpu_execlist *execlist)
247 struct intel_vgpu *vgpu = execlist->vgpu;
248 int ring_id = execlist->ring_id;
249 u32 status_reg = execlist_ring_mmio(vgpu->gvt, ring_id,
251 struct execlist_status_format status;
253 status.ldw = vgpu_vreg(vgpu, status_reg);
254 status.udw = vgpu_vreg(vgpu, status_reg + 4);
256 if (status.execlist_queue_full) {
257 gvt_err("virtual execlist slots are full\n");
261 return &execlist->slot[status.execlist_write_pointer];
264 static int emulate_execlist_schedule_in(struct intel_vgpu_execlist *execlist,
265 struct execlist_ctx_descriptor_format ctx[2])
267 struct intel_vgpu_execlist_slot *running = execlist->running_slot;
268 struct intel_vgpu_execlist_slot *slot =
269 get_next_execlist_slot(execlist);
271 struct execlist_ctx_descriptor_format *ctx0, *ctx1;
272 struct execlist_context_status_format status;
274 gvt_dbg_el("emulate schedule-in\n");
277 gvt_err("no available execlist slot\n");
281 memset(&status, 0, sizeof(status));
282 memset(slot->ctx, 0, sizeof(slot->ctx));
284 slot->ctx[0] = ctx[0];
285 slot->ctx[1] = ctx[1];
287 gvt_dbg_el("alloc slot index %d ctx 0 %x ctx 1 %x\n",
288 slot->index, ctx[0].context_id,
292 * no running execlist, make this write bundle as running execlist
296 gvt_dbg_el("no current running execlist\n");
298 execlist->running_slot = slot;
299 execlist->pending_slot = NULL;
300 execlist->running_context = &slot->ctx[0];
302 gvt_dbg_el("running slot index %d running context %x\n",
303 execlist->running_slot->index,
304 execlist->running_context->context_id);
306 emulate_execlist_status(execlist);
308 status.idle_to_active = 1;
309 status.context_id = 0;
311 emulate_csb_update(execlist, &status, false);
315 ctx0 = &running->ctx[0];
316 ctx1 = &running->ctx[1];
318 gvt_dbg_el("current running slot index %d ctx 0 %x ctx 1 %x\n",
319 running->index, ctx0->context_id, ctx1->context_id);
322 * already has an running execlist
323 * a. running ctx1 is valid,
324 * ctx0 is finished, and running ctx1 == new execlist ctx[0]
325 * b. running ctx1 is not valid,
326 * ctx0 == new execlist ctx[0]
327 * ----> lite-restore + preempted
329 if ((valid_context(ctx1) && same_context(ctx1, &slot->ctx[0]) &&
331 (!same_context(ctx0, execlist->running_context))) ||
332 (!valid_context(ctx1) &&
333 same_context(ctx0, &slot->ctx[0]))) { /* condition b */
334 gvt_dbg_el("need to switch virtual execlist slot\n");
336 execlist->pending_slot = slot;
337 switch_virtual_execlist_slot(execlist);
339 emulate_execlist_status(execlist);
341 status.lite_restore = status.preempted = 1;
342 status.context_id = ctx[0].context_id;
344 emulate_csb_update(execlist, &status, false);
346 gvt_dbg_el("emulate as pending slot\n");
349 * --> emulate pending execlist exist + but no preemption case
351 execlist->pending_slot = slot;
352 emulate_execlist_status(execlist);
357 static void free_workload(struct intel_vgpu_workload *workload)
359 intel_vgpu_unpin_mm(workload->shadow_mm);
360 intel_gvt_mm_unreference(workload->shadow_mm);
361 kmem_cache_free(workload->vgpu->workloads, workload);
364 #define get_desc_from_elsp_dwords(ed, i) \
365 ((struct execlist_ctx_descriptor_format *)&((ed)->data[i * 2]))
368 #define BATCH_BUFFER_ADDR_MASK ((1UL << 32) - (1U << 2))
369 #define BATCH_BUFFER_ADDR_HIGH_MASK ((1UL << 16) - (1U))
370 static int set_gma_to_bb_cmd(struct intel_shadow_bb_entry *entry_obj,
371 unsigned long add, int gmadr_bytes)
373 if (WARN_ON(gmadr_bytes != 4 && gmadr_bytes != 8))
376 *((u32 *)(entry_obj->bb_start_cmd_va + (1 << 2))) = add &
377 BATCH_BUFFER_ADDR_MASK;
378 if (gmadr_bytes == 8) {
379 *((u32 *)(entry_obj->bb_start_cmd_va + (2 << 2))) =
380 add & BATCH_BUFFER_ADDR_HIGH_MASK;
386 static void prepare_shadow_batch_buffer(struct intel_vgpu_workload *workload)
388 int gmadr_bytes = workload->vgpu->gvt->device_info.gmadr_bytes_in_cmd;
389 struct i915_vma *vma;
392 /* pin the gem object to ggtt */
393 if (!list_empty(&workload->shadow_bb)) {
394 struct intel_shadow_bb_entry *entry_obj =
395 list_first_entry(&workload->shadow_bb,
396 struct intel_shadow_bb_entry,
398 struct intel_shadow_bb_entry *temp;
400 list_for_each_entry_safe(entry_obj, temp, &workload->shadow_bb,
402 vma = i915_gem_object_ggtt_pin(entry_obj->obj, NULL, 0,
405 gvt_err("Cannot pin\n");
408 i915_gem_object_unpin_pages(entry_obj->obj);
410 /* update the relocate gma with shadow batch buffer*/
411 gma = i915_gem_object_ggtt_offset(entry_obj->obj, NULL);
412 WARN_ON(!IS_ALIGNED(gma, 4));
413 set_gma_to_bb_cmd(entry_obj, gma, gmadr_bytes);
418 static int update_wa_ctx_2_shadow_ctx(struct intel_shadow_wa_ctx *wa_ctx)
420 int ring_id = wa_ctx->workload->ring_id;
421 struct i915_gem_context *shadow_ctx =
422 wa_ctx->workload->vgpu->shadow_ctx;
423 struct drm_i915_gem_object *ctx_obj =
424 shadow_ctx->engine[ring_id].state->obj;
425 struct execlist_ring_context *shadow_ring_context;
428 page = i915_gem_object_get_page(ctx_obj, LRC_STATE_PN);
429 shadow_ring_context = kmap_atomic(page);
431 shadow_ring_context->bb_per_ctx_ptr.val =
432 (shadow_ring_context->bb_per_ctx_ptr.val &
433 (~PER_CTX_ADDR_MASK)) | wa_ctx->per_ctx.shadow_gma;
434 shadow_ring_context->rcs_indirect_ctx.val =
435 (shadow_ring_context->rcs_indirect_ctx.val &
436 (~INDIRECT_CTX_ADDR_MASK)) | wa_ctx->indirect_ctx.shadow_gma;
438 kunmap_atomic(shadow_ring_context);
442 static void prepare_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
444 struct i915_vma *vma;
446 unsigned char *per_ctx_va =
447 (unsigned char *)wa_ctx->indirect_ctx.shadow_va +
448 wa_ctx->indirect_ctx.size;
450 if (wa_ctx->indirect_ctx.size == 0)
453 vma = i915_gem_object_ggtt_pin(wa_ctx->indirect_ctx.obj, NULL, 0, 0, 0);
455 gvt_err("Cannot pin indirect ctx obj\n");
458 i915_gem_object_unpin_pages(wa_ctx->indirect_ctx.obj);
460 gma = i915_gem_object_ggtt_offset(wa_ctx->indirect_ctx.obj, NULL);
461 WARN_ON(!IS_ALIGNED(gma, CACHELINE_BYTES));
462 wa_ctx->indirect_ctx.shadow_gma = gma;
464 wa_ctx->per_ctx.shadow_gma = *((unsigned int *)per_ctx_va + 1);
465 memset(per_ctx_va, 0, CACHELINE_BYTES);
467 update_wa_ctx_2_shadow_ctx(wa_ctx);
470 static int prepare_execlist_workload(struct intel_vgpu_workload *workload)
472 struct intel_vgpu *vgpu = workload->vgpu;
473 struct execlist_ctx_descriptor_format ctx[2];
474 int ring_id = workload->ring_id;
476 intel_vgpu_pin_mm(workload->shadow_mm);
477 intel_vgpu_sync_oos_pages(workload->vgpu);
478 intel_vgpu_flush_post_shadow(workload->vgpu);
479 prepare_shadow_batch_buffer(workload);
480 prepare_shadow_wa_ctx(&workload->wa_ctx);
481 if (!workload->emulate_schedule_in)
484 ctx[0] = *get_desc_from_elsp_dwords(&workload->elsp_dwords, 1);
485 ctx[1] = *get_desc_from_elsp_dwords(&workload->elsp_dwords, 0);
487 return emulate_execlist_schedule_in(&vgpu->execlist[ring_id], ctx);
490 static void release_shadow_batch_buffer(struct intel_vgpu_workload *workload)
492 /* release all the shadow batch buffer */
493 if (!list_empty(&workload->shadow_bb)) {
494 struct intel_shadow_bb_entry *entry_obj =
495 list_first_entry(&workload->shadow_bb,
496 struct intel_shadow_bb_entry,
498 struct intel_shadow_bb_entry *temp;
500 list_for_each_entry_safe(entry_obj, temp, &workload->shadow_bb,
502 drm_gem_object_unreference(&(entry_obj->obj->base));
503 kvfree(entry_obj->va);
504 list_del(&entry_obj->list);
510 static void release_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
512 if (wa_ctx->indirect_ctx.size == 0)
515 drm_gem_object_unreference(&(wa_ctx->indirect_ctx.obj->base));
516 kvfree(wa_ctx->indirect_ctx.shadow_va);
519 static int complete_execlist_workload(struct intel_vgpu_workload *workload)
521 struct intel_vgpu *vgpu = workload->vgpu;
522 struct intel_vgpu_execlist *execlist =
523 &vgpu->execlist[workload->ring_id];
524 struct intel_vgpu_workload *next_workload;
525 struct list_head *next = workload_q_head(vgpu, workload->ring_id)->next;
526 bool lite_restore = false;
529 gvt_dbg_el("complete workload %p status %d\n", workload,
532 release_shadow_batch_buffer(workload);
533 release_shadow_wa_ctx(&workload->wa_ctx);
535 if (workload->status || vgpu->resetting)
538 if (!list_empty(workload_q_head(vgpu, workload->ring_id))) {
539 struct execlist_ctx_descriptor_format *this_desc, *next_desc;
541 next_workload = container_of(next,
542 struct intel_vgpu_workload, list);
543 this_desc = &workload->ctx_desc;
544 next_desc = &next_workload->ctx_desc;
546 lite_restore = same_context(this_desc, next_desc);
550 gvt_dbg_el("next context == current - no schedule-out\n");
551 free_workload(workload);
555 ret = emulate_execlist_ctx_schedule_out(execlist, &workload->ctx_desc);
559 free_workload(workload);
562 free_workload(workload);
566 #define RING_CTX_OFF(x) \
567 offsetof(struct execlist_ring_context, x)
569 static void read_guest_pdps(struct intel_vgpu *vgpu,
570 u64 ring_context_gpa, u32 pdp[8])
575 gpa = ring_context_gpa + RING_CTX_OFF(pdp3_UDW.val);
577 for (i = 0; i < 8; i++)
578 intel_gvt_hypervisor_read_gpa(vgpu,
579 gpa + i * 8, &pdp[7 - i], 4);
582 static int prepare_mm(struct intel_vgpu_workload *workload)
584 struct execlist_ctx_descriptor_format *desc = &workload->ctx_desc;
585 struct intel_vgpu_mm *mm;
586 int page_table_level;
589 if (desc->addressing_mode == 1) { /* legacy 32-bit */
590 page_table_level = 3;
591 } else if (desc->addressing_mode == 3) { /* legacy 64 bit */
592 page_table_level = 4;
594 gvt_err("Advanced Context mode(SVM) is not supported!\n");
598 read_guest_pdps(workload->vgpu, workload->ring_context_gpa, pdp);
600 mm = intel_vgpu_find_ppgtt_mm(workload->vgpu, page_table_level, pdp);
602 intel_gvt_mm_reference(mm);
605 mm = intel_vgpu_create_mm(workload->vgpu, INTEL_GVT_MM_PPGTT,
606 pdp, page_table_level, 0);
608 gvt_err("fail to create mm object.\n");
612 workload->shadow_mm = mm;
616 #define get_last_workload(q) \
617 (list_empty(q) ? NULL : container_of(q->prev, \
618 struct intel_vgpu_workload, list))
620 bool submit_context(struct intel_vgpu *vgpu, int ring_id,
621 struct execlist_ctx_descriptor_format *desc,
622 bool emulate_schedule_in)
624 struct list_head *q = workload_q_head(vgpu, ring_id);
625 struct intel_vgpu_workload *last_workload = get_last_workload(q);
626 struct intel_vgpu_workload *workload = NULL;
627 u64 ring_context_gpa;
628 u32 head, tail, start, ctl, ctx_ctl, per_ctx, indirect_ctx;
631 ring_context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
632 (u32)((desc->lrca + 1) << GTT_PAGE_SHIFT));
633 if (ring_context_gpa == INTEL_GVT_INVALID_ADDR) {
634 gvt_err("invalid guest context LRCA: %x\n", desc->lrca);
638 intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
639 RING_CTX_OFF(ring_header.val), &head, 4);
641 intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
642 RING_CTX_OFF(ring_tail.val), &tail, 4);
644 head &= RB_HEAD_OFF_MASK;
645 tail &= RB_TAIL_OFF_MASK;
647 if (last_workload && same_context(&last_workload->ctx_desc, desc)) {
648 gvt_dbg_el("ring id %d cur workload == last\n", ring_id);
649 gvt_dbg_el("ctx head %x real head %lx\n", head,
650 last_workload->rb_tail);
652 * cannot use guest context head pointer here,
653 * as it might not be updated at this time
655 head = last_workload->rb_tail;
658 gvt_dbg_el("ring id %d begin a new workload\n", ring_id);
660 workload = kmem_cache_zalloc(vgpu->workloads, GFP_KERNEL);
664 /* record some ring buffer register values for scan and shadow */
665 intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
666 RING_CTX_OFF(rb_start.val), &start, 4);
667 intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
668 RING_CTX_OFF(rb_ctrl.val), &ctl, 4);
669 intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
670 RING_CTX_OFF(ctx_ctrl.val), &ctx_ctl, 4);
672 INIT_LIST_HEAD(&workload->list);
673 INIT_LIST_HEAD(&workload->shadow_bb);
675 init_waitqueue_head(&workload->shadow_ctx_status_wq);
676 atomic_set(&workload->shadow_ctx_active, 0);
678 workload->vgpu = vgpu;
679 workload->ring_id = ring_id;
680 workload->ctx_desc = *desc;
681 workload->ring_context_gpa = ring_context_gpa;
682 workload->rb_head = head;
683 workload->rb_tail = tail;
684 workload->rb_start = start;
685 workload->rb_ctl = ctl;
686 workload->prepare = prepare_execlist_workload;
687 workload->complete = complete_execlist_workload;
688 workload->status = -EINPROGRESS;
689 workload->emulate_schedule_in = emulate_schedule_in;
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);
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) *
702 workload->wa_ctx.per_ctx.guest_gma =
703 per_ctx & PER_CTX_ADDR_MASK;
704 workload->wa_ctx.workload = workload;
706 WARN_ON(workload->wa_ctx.indirect_ctx.size && !(per_ctx & 0x1));
709 if (emulate_schedule_in)
710 memcpy(&workload->elsp_dwords,
711 &vgpu->execlist[ring_id].elsp_dwords,
712 sizeof(workload->elsp_dwords));
714 gvt_dbg_el("workload %p ring id %d head %x tail %x start %x ctl %x\n",
715 workload, ring_id, head, tail, start, ctl);
717 gvt_dbg_el("workload %p emulate schedule_in %d\n", workload,
718 emulate_schedule_in);
720 ret = prepare_mm(workload);
722 kmem_cache_free(vgpu->workloads, workload);
726 queue_workload(workload);
730 int intel_vgpu_submit_execlist(struct intel_vgpu *vgpu, int ring_id)
732 struct intel_vgpu_execlist *execlist = &vgpu->execlist[ring_id];
733 struct execlist_ctx_descriptor_format *desc[2], valid_desc[2];
734 unsigned long valid_desc_bitmap = 0;
735 bool emulate_schedule_in = true;
739 memset(valid_desc, 0, sizeof(valid_desc));
741 desc[0] = get_desc_from_elsp_dwords(&execlist->elsp_dwords, 1);
742 desc[1] = get_desc_from_elsp_dwords(&execlist->elsp_dwords, 0);
744 for (i = 0; i < 2; i++) {
748 if (!desc[i]->privilege_access) {
749 gvt_err("vgpu%d: unexpected GGTT elsp submission\n",
754 /* TODO: add another guest context checks here. */
755 set_bit(i, &valid_desc_bitmap);
756 valid_desc[i] = *desc[i];
759 if (!valid_desc_bitmap) {
760 gvt_err("vgpu%d: no valid desc in a elsp submission\n",
765 if (!test_bit(0, (void *)&valid_desc_bitmap) &&
766 test_bit(1, (void *)&valid_desc_bitmap)) {
767 gvt_err("vgpu%d: weird elsp submission, desc 0 is not valid\n",
772 /* submit workload */
773 for_each_set_bit(i, (void *)&valid_desc_bitmap, 2) {
774 ret = submit_context(vgpu, ring_id, &valid_desc[i],
775 emulate_schedule_in);
777 gvt_err("vgpu%d: fail to schedule workload\n",
781 emulate_schedule_in = false;
786 static void init_vgpu_execlist(struct intel_vgpu *vgpu, int ring_id)
788 struct intel_vgpu_execlist *execlist = &vgpu->execlist[ring_id];
789 struct execlist_context_status_pointer_format ctx_status_ptr;
790 u32 ctx_status_ptr_reg;
792 memset(execlist, 0, sizeof(*execlist));
794 execlist->vgpu = vgpu;
795 execlist->ring_id = ring_id;
796 execlist->slot[0].index = 0;
797 execlist->slot[1].index = 1;
799 ctx_status_ptr_reg = execlist_ring_mmio(vgpu->gvt, ring_id,
800 _EL_OFFSET_STATUS_PTR);
802 ctx_status_ptr.dw = vgpu_vreg(vgpu, ctx_status_ptr_reg);
803 ctx_status_ptr.read_ptr = ctx_status_ptr.write_ptr = 0x7;
804 vgpu_vreg(vgpu, ctx_status_ptr_reg) = ctx_status_ptr.dw;
807 void intel_vgpu_clean_execlist(struct intel_vgpu *vgpu)
809 kmem_cache_destroy(vgpu->workloads);
812 int intel_vgpu_init_execlist(struct intel_vgpu *vgpu)
816 /* each ring has a virtual execlist engine */
817 for (i = 0; i < I915_NUM_ENGINES; i++) {
818 init_vgpu_execlist(vgpu, i);
819 INIT_LIST_HEAD(&vgpu->workload_q_head[i]);
822 vgpu->workloads = kmem_cache_create("gvt-g vgpu workload",
823 sizeof(struct intel_vgpu_workload), 0,
827 if (!vgpu->workloads)
833 void intel_vgpu_reset_execlist(struct intel_vgpu *vgpu,
834 unsigned long ring_bitmap)
837 struct list_head *pos, *n;
838 struct intel_vgpu_workload *workload = NULL;
840 for_each_set_bit(bit, &ring_bitmap, sizeof(ring_bitmap) * 8) {
841 if (bit >= I915_NUM_ENGINES)
843 /* free the unsubmited workload in the queue */
844 list_for_each_safe(pos, n, &vgpu->workload_q_head[bit]) {
845 workload = container_of(pos,
846 struct intel_vgpu_workload, list);
847 list_del_init(&workload->list);
848 free_workload(workload);
851 init_vgpu_execlist(vgpu, bit);