drivers/gpu/drm/i915/gvt/scheduler.c

   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  *    Zhi Wang <zhi.a.wang@intel.com>
  25  *
  26  * Contributors:
  27  *    Ping Gao <ping.a.gao@intel.com>
  28  *    Tina Zhang <tina.zhang@intel.com>
  29  *    Chanbin Du <changbin.du@intel.com>
  30  *    Min He <min.he@intel.com>
  31  *    Bing Niu <bing.niu@intel.com>
  32  *    Zhenyu Wang <zhenyuw@linux.intel.com>
  33  *
  34  */
  35
  36 #include <linux/kthread.h>
  37
  38 #include "i915_drv.h"
  39 #include "gvt.h"
  40
  41 #define RING_CTX_OFF(x) \
  42         offsetof(struct execlist_ring_context, x)
  43
  44 static void set_context_pdp_root_pointer(
  45                 struct execlist_ring_context *ring_context,
  46                 u32 pdp[8])
  47 {
  48         int i;
  49
  50         for (i = 0; i < 8; i++)
  51                 ring_context->pdps[i].val = pdp[7 - i];
  52 }
  53
  54 static void update_shadow_pdps(struct intel_vgpu_workload *workload)
  55 {
  56         struct drm_i915_gem_object *ctx_obj =
  57                 workload->req->hw_context->state->obj;
  58         struct execlist_ring_context *shadow_ring_context;
  59         struct page *page;
  60
  61         if (WARN_ON(!workload->shadow_mm))
  62                 return;
  63
  64         if (WARN_ON(!atomic_read(&workload->shadow_mm->pincount)))
  65                 return;
  66
  67         page = i915_gem_object_get_page(ctx_obj, LRC_STATE_PN);
  68         shadow_ring_context = kmap(page);
  69         set_context_pdp_root_pointer(shadow_ring_context,
  70                         (void *)workload->shadow_mm->ppgtt_mm.shadow_pdps);
  71         kunmap(page);
  72 }
  73
  74 /*
  75  * when populating shadow ctx from guest, we should not overrride oa related
  76  * registers, so that they will not be overlapped by guest oa configs. Thus
  77  * made it possible to capture oa data from host for both host and guests.
  78  */
  79 static void sr_oa_regs(struct intel_vgpu_workload *workload,
  80                 u32 *reg_state, bool save)
  81 {
  82         struct drm_i915_private *dev_priv = workload->vgpu->gvt->dev_priv;
  83         u32 ctx_oactxctrl = dev_priv->perf.oa.ctx_oactxctrl_offset;
  84         u32 ctx_flexeu0 = dev_priv->perf.oa.ctx_flexeu0_offset;
  85         int i = 0;
  86         u32 flex_mmio[] = {
  87                 i915_mmio_reg_offset(EU_PERF_CNTL0),
  88                 i915_mmio_reg_offset(EU_PERF_CNTL1),
  89                 i915_mmio_reg_offset(EU_PERF_CNTL2),
  90                 i915_mmio_reg_offset(EU_PERF_CNTL3),
  91                 i915_mmio_reg_offset(EU_PERF_CNTL4),
  92                 i915_mmio_reg_offset(EU_PERF_CNTL5),
  93                 i915_mmio_reg_offset(EU_PERF_CNTL6),
  94         };
  95
  96         if (workload->ring_id != RCS)
  97                 return;
  98
  99         if (save) {
 100                 workload->oactxctrl = reg_state[ctx_oactxctrl + 1];
 101
 102                 for (i = 0; i < ARRAY_SIZE(workload->flex_mmio); i++) {
 103                         u32 state_offset = ctx_flexeu0 + i * 2;
 104
 105                         workload->flex_mmio[i] = reg_state[state_offset + 1];
 106                 }
 107         } else {
 108                 reg_state[ctx_oactxctrl] =
 109                         i915_mmio_reg_offset(GEN8_OACTXCONTROL);
 110                 reg_state[ctx_oactxctrl + 1] = workload->oactxctrl;
 111
 112                 for (i = 0; i < ARRAY_SIZE(workload->flex_mmio); i++) {
 113                         u32 state_offset = ctx_flexeu0 + i * 2;
 114                         u32 mmio = flex_mmio[i];
 115
 116                         reg_state[state_offset] = mmio;
 117                         reg_state[state_offset + 1] = workload->flex_mmio[i];
 118                 }
 119         }
 120 }
 121
 122 static int populate_shadow_context(struct intel_vgpu_workload *workload)
 123 {
 124         struct intel_vgpu *vgpu = workload->vgpu;
 125         struct intel_gvt *gvt = vgpu->gvt;
 126         int ring_id = workload->ring_id;
 127         struct drm_i915_gem_object *ctx_obj =
 128                 workload->req->hw_context->state->obj;
 129         struct execlist_ring_context *shadow_ring_context;
 130         struct page *page;
 131         void *dst;
 132         unsigned long context_gpa, context_page_num;
 133         int i;
 134
 135         gvt_dbg_sched("ring id %d workload lrca %x", ring_id,
 136                         workload->ctx_desc.lrca);
 137
 138         context_page_num = gvt->dev_priv->engine[ring_id]->context_size;
 139
 140         context_page_num = context_page_num >> PAGE_SHIFT;
 141
 142         if (IS_BROADWELL(gvt->dev_priv) && ring_id == RCS)
 143                 context_page_num = 19;
 144
 145         i = 2;
 146
 147         while (i < context_page_num) {
 148                 context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
 149                                 (u32)((workload->ctx_desc.lrca + i) <<
 150                                 I915_GTT_PAGE_SHIFT));
 151                 if (context_gpa == INTEL_GVT_INVALID_ADDR) {
 152                         gvt_vgpu_err("Invalid guest context descriptor\n");
 153                         return -EFAULT;
 154                 }
 155
 156                 page = i915_gem_object_get_page(ctx_obj, LRC_HEADER_PAGES + i);
 157                 dst = kmap(page);
 158                 intel_gvt_hypervisor_read_gpa(vgpu, context_gpa, dst,
 159                                 I915_GTT_PAGE_SIZE);
 160                 kunmap(page);
 161                 i++;
 162         }
 163
 164         page = i915_gem_object_get_page(ctx_obj, LRC_STATE_PN);
 165         shadow_ring_context = kmap(page);
 166
 167         sr_oa_regs(workload, (u32 *)shadow_ring_context, true);
 168 #define COPY_REG(name) \
 169         intel_gvt_hypervisor_read_gpa(vgpu, workload->ring_context_gpa \
 170                 + RING_CTX_OFF(name.val), &shadow_ring_context->name.val, 4)
 171 #define COPY_REG_MASKED(name) {\
 172                 intel_gvt_hypervisor_read_gpa(vgpu, workload->ring_context_gpa \
 173                                               + RING_CTX_OFF(name.val),\
 174                                               &shadow_ring_context->name.val, 4);\
 175                 shadow_ring_context->name.val |= 0xffff << 16;\
 176         }
 177
 178         COPY_REG_MASKED(ctx_ctrl);
 179         COPY_REG(ctx_timestamp);
 180
 181         if (ring_id == RCS) {
 182                 COPY_REG(bb_per_ctx_ptr);
 183                 COPY_REG(rcs_indirect_ctx);
 184                 COPY_REG(rcs_indirect_ctx_offset);
 185         }
 186 #undef COPY_REG
 187 #undef COPY_REG_MASKED
 188
 189         intel_gvt_hypervisor_read_gpa(vgpu,
 190                         workload->ring_context_gpa +
 191                         sizeof(*shadow_ring_context),
 192                         (void *)shadow_ring_context +
 193                         sizeof(*shadow_ring_context),
 194                         I915_GTT_PAGE_SIZE - sizeof(*shadow_ring_context));
 195
 196         sr_oa_regs(workload, (u32 *)shadow_ring_context, false);
 197         kunmap(page);
 198         return 0;
 199 }
 200
 201 static inline bool is_gvt_request(struct i915_request *req)
 202 {
 203         return i915_gem_context_force_single_submission(req->gem_context);
 204 }
 205
 206 static void save_ring_hw_state(struct intel_vgpu *vgpu, int ring_id)
 207 {
 208         struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
 209         u32 ring_base = dev_priv->engine[ring_id]->mmio_base;
 210         i915_reg_t reg;
 211
 212         reg = RING_INSTDONE(ring_base);
 213         vgpu_vreg(vgpu, i915_mmio_reg_offset(reg)) = I915_READ_FW(reg);
 214         reg = RING_ACTHD(ring_base);
 215         vgpu_vreg(vgpu, i915_mmio_reg_offset(reg)) = I915_READ_FW(reg);
 216         reg = RING_ACTHD_UDW(ring_base);
 217         vgpu_vreg(vgpu, i915_mmio_reg_offset(reg)) = I915_READ_FW(reg);
 218 }
 219
 220 static int shadow_context_status_change(struct notifier_block *nb,
 221                 unsigned long action, void *data)
 222 {
 223         struct i915_request *req = data;
 224         struct intel_gvt *gvt = container_of(nb, struct intel_gvt,
 225                                 shadow_ctx_notifier_block[req->engine->id]);
 226         struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
 227         enum intel_engine_id ring_id = req->engine->id;
 228         struct intel_vgpu_workload *workload;
 229         unsigned long flags;
 230
 231         if (!is_gvt_request(req)) {
 232                 spin_lock_irqsave(&scheduler->mmio_context_lock, flags);
 233                 if (action == INTEL_CONTEXT_SCHEDULE_IN &&
 234                     scheduler->engine_owner[ring_id]) {
 235                         /* Switch ring from vGPU to host. */
 236                         intel_gvt_switch_mmio(scheduler->engine_owner[ring_id],
 237                                               NULL, ring_id);
 238                         scheduler->engine_owner[ring_id] = NULL;
 239                 }
 240                 spin_unlock_irqrestore(&scheduler->mmio_context_lock, flags);
 241
 242                 return NOTIFY_OK;
 243         }
 244
 245         workload = scheduler->current_workload[ring_id];
 246         if (unlikely(!workload))
 247                 return NOTIFY_OK;
 248
 249         switch (action) {
 250         case INTEL_CONTEXT_SCHEDULE_IN:
 251                 spin_lock_irqsave(&scheduler->mmio_context_lock, flags);
 252                 if (workload->vgpu != scheduler->engine_owner[ring_id]) {
 253                         /* Switch ring from host to vGPU or vGPU to vGPU. */
 254                         intel_gvt_switch_mmio(scheduler->engine_owner[ring_id],
 255                                               workload->vgpu, ring_id);
 256                         scheduler->engine_owner[ring_id] = workload->vgpu;
 257                 } else
 258                         gvt_dbg_sched("skip ring %d mmio switch for vgpu%d\n",
 259                                       ring_id, workload->vgpu->id);
 260                 spin_unlock_irqrestore(&scheduler->mmio_context_lock, flags);
 261                 atomic_set(&workload->shadow_ctx_active, 1);
 262                 break;
 263         case INTEL_CONTEXT_SCHEDULE_OUT:
 264                 save_ring_hw_state(workload->vgpu, ring_id);
 265                 atomic_set(&workload->shadow_ctx_active, 0);
 266                 break;
 267         case INTEL_CONTEXT_SCHEDULE_PREEMPTED:
 268                 save_ring_hw_state(workload->vgpu, ring_id);
 269                 break;
 270         default:
 271                 WARN_ON(1);
 272                 return NOTIFY_OK;
 273         }
 274         wake_up(&workload->shadow_ctx_status_wq);
 275         return NOTIFY_OK;
 276 }
 277
 278 static void shadow_context_descriptor_update(struct intel_context *ce)
 279 {
 280         u64 desc = 0;
 281
 282         desc = ce->lrc_desc;
 283
 284         /* Update bits 0-11 of the context descriptor which includes flags
 285          * like GEN8_CTX_* cached in desc_template
 286          */
 287         desc &= U64_MAX << 12;
 288         desc |= ce->gem_context->desc_template & ((1ULL << 12) - 1);
 289
 290         ce->lrc_desc = desc;
 291 }
 292
 293 static int copy_workload_to_ring_buffer(struct intel_vgpu_workload *workload)
 294 {
 295         struct intel_vgpu *vgpu = workload->vgpu;
 296         struct i915_request *req = workload->req;
 297         void *shadow_ring_buffer_va;
 298         u32 *cs;
 299
 300         if ((IS_KABYLAKE(req->i915) || IS_BROXTON(req->i915))
 301                 && is_inhibit_context(req->hw_context))
 302                 intel_vgpu_restore_inhibit_context(vgpu, req);
 303
 304         /* allocate shadow ring buffer */
 305         cs = intel_ring_begin(workload->req, workload->rb_len / sizeof(u32));
 306         if (IS_ERR(cs)) {
 307                 gvt_vgpu_err("fail to alloc size =%ld shadow  ring buffer\n",
 308                         workload->rb_len);
 309                 return PTR_ERR(cs);
 310         }
 311
 312         shadow_ring_buffer_va = workload->shadow_ring_buffer_va;
 313
 314         /* get shadow ring buffer va */
 315         workload->shadow_ring_buffer_va = cs;
 316
 317         memcpy(cs, shadow_ring_buffer_va,
 318                         workload->rb_len);
 319
 320         cs += workload->rb_len / sizeof(u32);
 321         intel_ring_advance(workload->req, cs);
 322
 323         return 0;
 324 }
 325
 326 static void release_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
 327 {
 328         if (!wa_ctx->indirect_ctx.obj)
 329                 return;
 330
 331         i915_gem_object_unpin_map(wa_ctx->indirect_ctx.obj);
 332         i915_gem_object_put(wa_ctx->indirect_ctx.obj);
 333 }
 334
 335 /**
 336  * intel_gvt_scan_and_shadow_workload - audit the workload by scanning and
 337  * shadow it as well, include ringbuffer,wa_ctx and ctx.
 338  * @workload: an abstract entity for each execlist submission.
 339  *
 340  * This function is called before the workload submitting to i915, to make
 341  * sure the content of the workload is valid.
 342  */
 343 int intel_gvt_scan_and_shadow_workload(struct intel_vgpu_workload *workload)
 344 {
 345         struct intel_vgpu *vgpu = workload->vgpu;
 346         struct intel_vgpu_submission *s = &vgpu->submission;
 347         struct i915_gem_context *shadow_ctx = s->shadow_ctx;
 348         struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
 349         struct intel_engine_cs *engine = dev_priv->engine[workload->ring_id];
 350         struct intel_context *ce;
 351         struct i915_request *rq;
 352         int ret;
 353
 354         lockdep_assert_held(&dev_priv->drm.struct_mutex);
 355
 356         if (workload->req)
 357                 return 0;
 358
 359         /* pin shadow context by gvt even the shadow context will be pinned
 360          * when i915 alloc request. That is because gvt will update the guest
 361          * context from shadow context when workload is completed, and at that
 362          * moment, i915 may already unpined the shadow context to make the
 363          * shadow_ctx pages invalid. So gvt need to pin itself. After update
 364          * the guest context, gvt can unpin the shadow_ctx safely.
 365          */
 366         ce = intel_context_pin(shadow_ctx, engine);
 367         if (IS_ERR(ce)) {
 368                 gvt_vgpu_err("fail to pin shadow context\n");
 369                 return PTR_ERR(ce);
 370         }
 371
 372         shadow_ctx->desc_template &= ~(0x3 << GEN8_CTX_ADDRESSING_MODE_SHIFT);
 373         shadow_ctx->desc_template |= workload->ctx_desc.addressing_mode <<
 374                                     GEN8_CTX_ADDRESSING_MODE_SHIFT;
 375
 376         if (!test_and_set_bit(workload->ring_id, s->shadow_ctx_desc_updated))
 377                 shadow_context_descriptor_update(ce);
 378
 379         ret = intel_gvt_scan_and_shadow_ringbuffer(workload);
 380         if (ret)
 381                 goto err_unpin;
 382
 383         if ((workload->ring_id == RCS) &&
 384             (workload->wa_ctx.indirect_ctx.size != 0)) {
 385                 ret = intel_gvt_scan_and_shadow_wa_ctx(&workload->wa_ctx);
 386                 if (ret)
 387                         goto err_shadow;
 388         }
 389
 390         rq = i915_request_alloc(engine, shadow_ctx);
 391         if (IS_ERR(rq)) {
 392                 gvt_vgpu_err("fail to allocate gem request\n");
 393                 ret = PTR_ERR(rq);
 394                 goto err_shadow;
 395         }
 396         workload->req = i915_request_get(rq);
 397
 398         ret = populate_shadow_context(workload);
 399         if (ret)
 400                 goto err_req;
 401
 402         return 0;
 403 err_req:
 404         rq = fetch_and_zero(&workload->req);
 405         i915_request_put(rq);
 406 err_shadow:
 407         release_shadow_wa_ctx(&workload->wa_ctx);
 408 err_unpin:
 409         intel_context_unpin(ce);
 410         return ret;
 411 }
 412
 413 static void release_shadow_batch_buffer(struct intel_vgpu_workload *workload);
 414
 415 static int prepare_shadow_batch_buffer(struct intel_vgpu_workload *workload)
 416 {
 417         struct intel_gvt *gvt = workload->vgpu->gvt;
 418         const int gmadr_bytes = gvt->device_info.gmadr_bytes_in_cmd;
 419         struct intel_vgpu_shadow_bb *bb;
 420         int ret;
 421
 422         list_for_each_entry(bb, &workload->shadow_bb, list) {
 423                 /* For privilge batch buffer and not wa_ctx, the bb_start_cmd_va
 424                  * is only updated into ring_scan_buffer, not real ring address
 425                  * allocated in later copy_workload_to_ring_buffer. pls be noted
 426                  * shadow_ring_buffer_va is now pointed to real ring buffer va
 427                  * in copy_workload_to_ring_buffer.
 428                  */
 429
 430                 if (bb->bb_offset)
 431                         bb->bb_start_cmd_va = workload->shadow_ring_buffer_va
 432                                 + bb->bb_offset;
 433
 434                 if (bb->ppgtt) {
 435                         /* for non-priv bb, scan&shadow is only for
 436                          * debugging purpose, so the content of shadow bb
 437                          * is the same as original bb. Therefore,
 438                          * here, rather than switch to shadow bb's gma
 439                          * address, we directly use original batch buffer's
 440                          * gma address, and send original bb to hardware
 441                          * directly
 442                          */
 443                         if (bb->clflush & CLFLUSH_AFTER) {
 444                                 drm_clflush_virt_range(bb->va,
 445                                                 bb->obj->base.size);
 446                                 bb->clflush &= ~CLFLUSH_AFTER;
 447                         }
 448                         i915_gem_obj_finish_shmem_access(bb->obj);
 449                         bb->accessing = false;
 450
 451                 } else {
 452                         bb->vma = i915_gem_object_ggtt_pin(bb->obj,
 453                                         NULL, 0, 0, 0);
 454                         if (IS_ERR(bb->vma)) {
 455                                 ret = PTR_ERR(bb->vma);
 456                                 goto err;
 457                         }
 458
 459                         /* relocate shadow batch buffer */
 460                         bb->bb_start_cmd_va[1] = i915_ggtt_offset(bb->vma);
 461                         if (gmadr_bytes == 8)
 462                                 bb->bb_start_cmd_va[2] = 0;
 463
 464                         /* No one is going to touch shadow bb from now on. */
 465                         if (bb->clflush & CLFLUSH_AFTER) {
 466                                 drm_clflush_virt_range(bb->va,
 467                                                 bb->obj->base.size);
 468                                 bb->clflush &= ~CLFLUSH_AFTER;
 469                         }
 470
 471                         ret = i915_gem_object_set_to_gtt_domain(bb->obj,
 472                                         false);
 473                         if (ret)
 474                                 goto err;
 475
 476                         i915_gem_obj_finish_shmem_access(bb->obj);
 477                         bb->accessing = false;
 478
 479                         ret = i915_vma_move_to_active(bb->vma,
 480                                                       workload->req,
 481                                                       0);
 482                         if (ret)
 483                                 goto err;
 484                 }
 485         }
 486         return 0;
 487 err:
 488         release_shadow_batch_buffer(workload);
 489         return ret;
 490 }
 491
 492 static void update_wa_ctx_2_shadow_ctx(struct intel_shadow_wa_ctx *wa_ctx)
 493 {
 494         struct intel_vgpu_workload *workload =
 495                 container_of(wa_ctx, struct intel_vgpu_workload, wa_ctx);
 496         struct i915_request *rq = workload->req;
 497         struct execlist_ring_context *shadow_ring_context =
 498                 (struct execlist_ring_context *)rq->hw_context->lrc_reg_state;
 499
 500         shadow_ring_context->bb_per_ctx_ptr.val =
 501                 (shadow_ring_context->bb_per_ctx_ptr.val &
 502                 (~PER_CTX_ADDR_MASK)) | wa_ctx->per_ctx.shadow_gma;
 503         shadow_ring_context->rcs_indirect_ctx.val =
 504                 (shadow_ring_context->rcs_indirect_ctx.val &
 505                 (~INDIRECT_CTX_ADDR_MASK)) | wa_ctx->indirect_ctx.shadow_gma;
 506 }
 507
 508 static int prepare_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
 509 {
 510         struct i915_vma *vma;
 511         unsigned char *per_ctx_va =
 512                 (unsigned char *)wa_ctx->indirect_ctx.shadow_va +
 513                 wa_ctx->indirect_ctx.size;
 514
 515         if (wa_ctx->indirect_ctx.size == 0)
 516                 return 0;
 517
 518         vma = i915_gem_object_ggtt_pin(wa_ctx->indirect_ctx.obj, NULL,
 519                                        0, CACHELINE_BYTES, 0);
 520         if (IS_ERR(vma))
 521                 return PTR_ERR(vma);
 522
 523         /* FIXME: we are not tracking our pinned VMA leaving it
 524          * up to the core to fix up the stray pin_count upon
 525          * free.
 526          */
 527
 528         wa_ctx->indirect_ctx.shadow_gma = i915_ggtt_offset(vma);
 529
 530         wa_ctx->per_ctx.shadow_gma = *((unsigned int *)per_ctx_va + 1);
 531         memset(per_ctx_va, 0, CACHELINE_BYTES);
 532
 533         update_wa_ctx_2_shadow_ctx(wa_ctx);
 534         return 0;
 535 }
 536
 537 static void release_shadow_batch_buffer(struct intel_vgpu_workload *workload)
 538 {
 539         struct intel_vgpu *vgpu = workload->vgpu;
 540         struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
 541         struct intel_vgpu_shadow_bb *bb, *pos;
 542
 543         if (list_empty(&workload->shadow_bb))
 544                 return;
 545
 546         bb = list_first_entry(&workload->shadow_bb,
 547                         struct intel_vgpu_shadow_bb, list);
 548
 549         mutex_lock(&dev_priv->drm.struct_mutex);
 550
 551         list_for_each_entry_safe(bb, pos, &workload->shadow_bb, list) {
 552                 if (bb->obj) {
 553                         if (bb->accessing)
 554                                 i915_gem_obj_finish_shmem_access(bb->obj);
 555
 556                         if (bb->va && !IS_ERR(bb->va))
 557                                 i915_gem_object_unpin_map(bb->obj);
 558
 559                         if (bb->vma && !IS_ERR(bb->vma)) {
 560                                 i915_vma_unpin(bb->vma);
 561                                 i915_vma_close(bb->vma);
 562                         }
 563                         __i915_gem_object_release_unless_active(bb->obj);
 564                 }
 565                 list_del(&bb->list);
 566                 kfree(bb);
 567         }
 568
 569         mutex_unlock(&dev_priv->drm.struct_mutex);
 570 }
 571
 572 static int prepare_workload(struct intel_vgpu_workload *workload)
 573 {
 574         struct intel_vgpu *vgpu = workload->vgpu;
 575         int ret = 0;
 576
 577         ret = intel_vgpu_pin_mm(workload->shadow_mm);
 578         if (ret) {
 579                 gvt_vgpu_err("fail to vgpu pin mm\n");
 580                 return ret;
 581         }
 582
 583         update_shadow_pdps(workload);
 584
 585         ret = intel_vgpu_sync_oos_pages(workload->vgpu);
 586         if (ret) {
 587                 gvt_vgpu_err("fail to vgpu sync oos pages\n");
 588                 goto err_unpin_mm;
 589         }
 590
 591         ret = intel_vgpu_flush_post_shadow(workload->vgpu);
 592         if (ret) {
 593                 gvt_vgpu_err("fail to flush post shadow\n");
 594                 goto err_unpin_mm;
 595         }
 596
 597         ret = copy_workload_to_ring_buffer(workload);
 598         if (ret) {
 599                 gvt_vgpu_err("fail to generate request\n");
 600                 goto err_unpin_mm;
 601         }
 602
 603         ret = prepare_shadow_batch_buffer(workload);
 604         if (ret) {
 605                 gvt_vgpu_err("fail to prepare_shadow_batch_buffer\n");
 606                 goto err_unpin_mm;
 607         }
 608
 609         ret = prepare_shadow_wa_ctx(&workload->wa_ctx);
 610         if (ret) {
 611                 gvt_vgpu_err("fail to prepare_shadow_wa_ctx\n");
 612                 goto err_shadow_batch;
 613         }
 614
 615         if (workload->prepare) {
 616                 ret = workload->prepare(workload);
 617                 if (ret)
 618                         goto err_shadow_wa_ctx;
 619         }
 620
 621         return 0;
 622 err_shadow_wa_ctx:
 623         release_shadow_wa_ctx(&workload->wa_ctx);
 624 err_shadow_batch:
 625         release_shadow_batch_buffer(workload);
 626 err_unpin_mm:
 627         intel_vgpu_unpin_mm(workload->shadow_mm);
 628         return ret;
 629 }
 630
 631 static int dispatch_workload(struct intel_vgpu_workload *workload)
 632 {
 633         struct intel_vgpu *vgpu = workload->vgpu;
 634         struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
 635         int ring_id = workload->ring_id;
 636         int ret;
 637
 638         gvt_dbg_sched("ring id %d prepare to dispatch workload %p\n",
 639                 ring_id, workload);
 640
 641         mutex_lock(&vgpu->vgpu_lock);
 642         mutex_lock(&dev_priv->drm.struct_mutex);
 643
 644         ret = intel_gvt_scan_and_shadow_workload(workload);
 645         if (ret)
 646                 goto out;
 647
 648         ret = prepare_workload(workload);
 649
 650 out:
 651         if (ret)
 652                 workload->status = ret;
 653
 654         if (!IS_ERR_OR_NULL(workload->req)) {
 655                 gvt_dbg_sched("ring id %d submit workload to i915 %p\n",
 656                                 ring_id, workload->req);
 657                 i915_request_add(workload->req);
 658                 workload->dispatched = true;
 659         }
 660
 661         mutex_unlock(&dev_priv->drm.struct_mutex);
 662         mutex_unlock(&vgpu->vgpu_lock);
 663         return ret;
 664 }
 665
 666 static struct intel_vgpu_workload *pick_next_workload(
 667                 struct intel_gvt *gvt, int ring_id)
 668 {
 669         struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
 670         struct intel_vgpu_workload *workload = NULL;
 671
 672         mutex_lock(&gvt->sched_lock);
 673
 674         /*
 675          * no current vgpu / will be scheduled out / no workload
 676          * bail out
 677          */
 678         if (!scheduler->current_vgpu) {
 679                 gvt_dbg_sched("ring id %d stop - no current vgpu\n", ring_id);
 680                 goto out;
 681         }
 682
 683         if (scheduler->need_reschedule) {
 684                 gvt_dbg_sched("ring id %d stop - will reschedule\n", ring_id);
 685                 goto out;
 686         }
 687
 688         if (list_empty(workload_q_head(scheduler->current_vgpu, ring_id)))
 689                 goto out;
 690
 691         /*
 692          * still have current workload, maybe the workload disptacher
 693          * fail to submit it for some reason, resubmit it.
 694          */
 695         if (scheduler->current_workload[ring_id]) {
 696                 workload = scheduler->current_workload[ring_id];
 697                 gvt_dbg_sched("ring id %d still have current workload %p\n",
 698                                 ring_id, workload);
 699                 goto out;
 700         }
 701
 702         /*
 703          * pick a workload as current workload
 704          * once current workload is set, schedule policy routines
 705          * will wait the current workload is finished when trying to
 706          * schedule out a vgpu.
 707          */
 708         scheduler->current_workload[ring_id] = container_of(
 709                         workload_q_head(scheduler->current_vgpu, ring_id)->next,
 710                         struct intel_vgpu_workload, list);
 711
 712         workload = scheduler->current_workload[ring_id];
 713
 714         gvt_dbg_sched("ring id %d pick new workload %p\n", ring_id, workload);
 715
 716         atomic_inc(&workload->vgpu->submission.running_workload_num);
 717 out:
 718         mutex_unlock(&gvt->sched_lock);
 719         return workload;
 720 }
 721
 722 static void update_guest_context(struct intel_vgpu_workload *workload)
 723 {
 724         struct i915_request *rq = workload->req;
 725         struct intel_vgpu *vgpu = workload->vgpu;
 726         struct intel_gvt *gvt = vgpu->gvt;
 727         struct drm_i915_gem_object *ctx_obj = rq->hw_context->state->obj;
 728         struct execlist_ring_context *shadow_ring_context;
 729         struct page *page;
 730         void *src;
 731         unsigned long context_gpa, context_page_num;
 732         int i;
 733
 734         gvt_dbg_sched("ring id %d workload lrca %x\n", rq->engine->id,
 735                       workload->ctx_desc.lrca);
 736
 737         context_page_num = rq->engine->context_size;
 738         context_page_num = context_page_num >> PAGE_SHIFT;
 739
 740         if (IS_BROADWELL(gvt->dev_priv) && rq->engine->id == RCS)
 741                 context_page_num = 19;
 742
 743         i = 2;
 744
 745         while (i < context_page_num) {
 746                 context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
 747                                 (u32)((workload->ctx_desc.lrca + i) <<
 748                                         I915_GTT_PAGE_SHIFT));
 749                 if (context_gpa == INTEL_GVT_INVALID_ADDR) {
 750                         gvt_vgpu_err("invalid guest context descriptor\n");
 751                         return;
 752                 }
 753
 754                 page = i915_gem_object_get_page(ctx_obj, LRC_HEADER_PAGES + i);
 755                 src = kmap(page);
 756                 intel_gvt_hypervisor_write_gpa(vgpu, context_gpa, src,
 757                                 I915_GTT_PAGE_SIZE);
 758                 kunmap(page);
 759                 i++;
 760         }
 761
 762         intel_gvt_hypervisor_write_gpa(vgpu, workload->ring_context_gpa +
 763                 RING_CTX_OFF(ring_header.val), &workload->rb_tail, 4);
 764
 765         page = i915_gem_object_get_page(ctx_obj, LRC_STATE_PN);
 766         shadow_ring_context = kmap(page);
 767
 768 #define COPY_REG(name) \
 769         intel_gvt_hypervisor_write_gpa(vgpu, workload->ring_context_gpa + \
 770                 RING_CTX_OFF(name.val), &shadow_ring_context->name.val, 4)
 771
 772         COPY_REG(ctx_ctrl);
 773         COPY_REG(ctx_timestamp);
 774
 775 #undef COPY_REG
 776
 777         intel_gvt_hypervisor_write_gpa(vgpu,
 778                         workload->ring_context_gpa +
 779                         sizeof(*shadow_ring_context),
 780                         (void *)shadow_ring_context +
 781                         sizeof(*shadow_ring_context),
 782                         I915_GTT_PAGE_SIZE - sizeof(*shadow_ring_context));
 783
 784         kunmap(page);
 785 }
 786
 787 static void clean_workloads(struct intel_vgpu *vgpu, unsigned long engine_mask)
 788 {
 789         struct intel_vgpu_submission *s = &vgpu->submission;
 790         struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
 791         struct intel_engine_cs *engine;
 792         struct intel_vgpu_workload *pos, *n;
 793         unsigned int tmp;
 794
 795         /* free the unsubmited workloads in the queues. */
 796         for_each_engine_masked(engine, dev_priv, engine_mask, tmp) {
 797                 list_for_each_entry_safe(pos, n,
 798                         &s->workload_q_head[engine->id], list) {
 799                         list_del_init(&pos->list);
 800                         intel_vgpu_destroy_workload(pos);
 801                 }
 802                 clear_bit(engine->id, s->shadow_ctx_desc_updated);
 803         }
 804 }
 805
 806 static void complete_current_workload(struct intel_gvt *gvt, int ring_id)
 807 {
 808         struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
 809         struct intel_vgpu_workload *workload =
 810                 scheduler->current_workload[ring_id];
 811         struct intel_vgpu *vgpu = workload->vgpu;
 812         struct intel_vgpu_submission *s = &vgpu->submission;
 813         struct i915_request *rq = workload->req;
 814         int event;
 815
 816         mutex_lock(&vgpu->vgpu_lock);
 817         mutex_lock(&gvt->sched_lock);
 818
 819         /* For the workload w/ request, needs to wait for the context
 820          * switch to make sure request is completed.
 821          * For the workload w/o request, directly complete the workload.
 822          */
 823         if (rq) {
 824                 wait_event(workload->shadow_ctx_status_wq,
 825                            !atomic_read(&workload->shadow_ctx_active));
 826
 827                 /* If this request caused GPU hang, req->fence.error will
 828                  * be set to -EIO. Use -EIO to set workload status so
 829                  * that when this request caused GPU hang, didn't trigger
 830                  * context switch interrupt to guest.
 831                  */
 832                 if (likely(workload->status == -EINPROGRESS)) {
 833                         if (workload->req->fence.error == -EIO)
 834                                 workload->status = -EIO;
 835                         else
 836                                 workload->status = 0;
 837                 }
 838
 839                 if (!workload->status && !(vgpu->resetting_eng &
 840                                            ENGINE_MASK(ring_id))) {
 841                         update_guest_context(workload);
 842
 843                         for_each_set_bit(event, workload->pending_events,
 844                                          INTEL_GVT_EVENT_MAX)
 845                                 intel_vgpu_trigger_virtual_event(vgpu, event);
 846                 }
 847
 848                 /* unpin shadow ctx as the shadow_ctx update is done */
 849                 mutex_lock(&rq->i915->drm.struct_mutex);
 850                 intel_context_unpin(rq->hw_context);
 851                 mutex_unlock(&rq->i915->drm.struct_mutex);
 852
 853                 i915_request_put(fetch_and_zero(&workload->req));
 854         }
 855
 856         gvt_dbg_sched("ring id %d complete workload %p status %d\n",
 857                         ring_id, workload, workload->status);
 858
 859         scheduler->current_workload[ring_id] = NULL;
 860
 861         list_del_init(&workload->list);
 862
 863         if (!workload->status) {
 864                 release_shadow_batch_buffer(workload);
 865                 release_shadow_wa_ctx(&workload->wa_ctx);
 866         }
 867
 868         if (workload->status || (vgpu->resetting_eng & ENGINE_MASK(ring_id))) {
 869                 /* if workload->status is not successful means HW GPU
 870                  * has occurred GPU hang or something wrong with i915/GVT,
 871                  * and GVT won't inject context switch interrupt to guest.
 872                  * So this error is a vGPU hang actually to the guest.
 873                  * According to this we should emunlate a vGPU hang. If
 874                  * there are pending workloads which are already submitted
 875                  * from guest, we should clean them up like HW GPU does.
 876                  *
 877                  * if it is in middle of engine resetting, the pending
 878                  * workloads won't be submitted to HW GPU and will be
 879                  * cleaned up during the resetting process later, so doing
 880                  * the workload clean up here doesn't have any impact.
 881                  **/
 882                 clean_workloads(vgpu, ENGINE_MASK(ring_id));
 883         }
 884
 885         workload->complete(workload);
 886
 887         atomic_dec(&s->running_workload_num);
 888         wake_up(&scheduler->workload_complete_wq);
 889
 890         if (gvt->scheduler.need_reschedule)
 891                 intel_gvt_request_service(gvt, INTEL_GVT_REQUEST_EVENT_SCHED);
 892
 893         mutex_unlock(&gvt->sched_lock);
 894         mutex_unlock(&vgpu->vgpu_lock);
 895 }
 896
 897 struct workload_thread_param {
 898         struct intel_gvt *gvt;
 899         int ring_id;
 900 };
 901
 902 static int workload_thread(void *priv)
 903 {
 904         struct workload_thread_param *p = (struct workload_thread_param *)priv;
 905         struct intel_gvt *gvt = p->gvt;
 906         int ring_id = p->ring_id;
 907         struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
 908         struct intel_vgpu_workload *workload = NULL;
 909         struct intel_vgpu *vgpu = NULL;
 910         int ret;
 911         bool need_force_wake = IS_SKYLAKE(gvt->dev_priv)
 912                         || IS_KABYLAKE(gvt->dev_priv)
 913                         || IS_BROXTON(gvt->dev_priv);
 914         DEFINE_WAIT_FUNC(wait, woken_wake_function);
 915
 916         kfree(p);
 917
 918         gvt_dbg_core("workload thread for ring %d started\n", ring_id);
 919
 920         while (!kthread_should_stop()) {
 921                 add_wait_queue(&scheduler->waitq[ring_id], &wait);
 922                 do {
 923                         workload = pick_next_workload(gvt, ring_id);
 924                         if (workload)
 925                                 break;
 926                         wait_woken(&wait, TASK_INTERRUPTIBLE,
 927                                    MAX_SCHEDULE_TIMEOUT);
 928                 } while (!kthread_should_stop());
 929                 remove_wait_queue(&scheduler->waitq[ring_id], &wait);
 930
 931                 if (!workload)
 932                         break;
 933
 934                 gvt_dbg_sched("ring id %d next workload %p vgpu %d\n",
 935                                 workload->ring_id, workload,
 936                                 workload->vgpu->id);
 937
 938                 intel_runtime_pm_get(gvt->dev_priv);
 939
 940                 gvt_dbg_sched("ring id %d will dispatch workload %p\n",
 941                                 workload->ring_id, workload);
 942
 943                 if (need_force_wake)
 944                         intel_uncore_forcewake_get(gvt->dev_priv,
 945                                         FORCEWAKE_ALL);
 946
 947                 ret = dispatch_workload(workload);
 948
 949                 if (ret) {
 950                         vgpu = workload->vgpu;
 951                         gvt_vgpu_err("fail to dispatch workload, skip\n");
 952                         goto complete;
 953                 }
 954
 955                 gvt_dbg_sched("ring id %d wait workload %p\n",
 956                                 workload->ring_id, workload);
 957                 i915_request_wait(workload->req, 0, MAX_SCHEDULE_TIMEOUT);
 958
 959 complete:
 960                 gvt_dbg_sched("will complete workload %p, status: %d\n",
 961                                 workload, workload->status);
 962
 963                 complete_current_workload(gvt, ring_id);
 964
 965                 if (need_force_wake)
 966                         intel_uncore_forcewake_put(gvt->dev_priv,
 967                                         FORCEWAKE_ALL);
 968
 969                 intel_runtime_pm_put(gvt->dev_priv);
 970                 if (ret && (vgpu_is_vm_unhealthy(ret)))
 971                         enter_failsafe_mode(vgpu, GVT_FAILSAFE_GUEST_ERR);
 972         }
 973         return 0;
 974 }
 975
 976 void intel_gvt_wait_vgpu_idle(struct intel_vgpu *vgpu)
 977 {
 978         struct intel_vgpu_submission *s = &vgpu->submission;
 979         struct intel_gvt *gvt = vgpu->gvt;
 980         struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
 981
 982         if (atomic_read(&s->running_workload_num)) {
 983                 gvt_dbg_sched("wait vgpu idle\n");
 984
 985                 wait_event(scheduler->workload_complete_wq,
 986                                 !atomic_read(&s->running_workload_num));
 987         }
 988 }
 989
 990 void intel_gvt_clean_workload_scheduler(struct intel_gvt *gvt)
 991 {
 992         struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
 993         struct intel_engine_cs *engine;
 994         enum intel_engine_id i;
 995
 996         gvt_dbg_core("clean workload scheduler\n");
 997
 998         for_each_engine(engine, gvt->dev_priv, i) {
 999                 atomic_notifier_chain_unregister(
1000                                         &engine->context_status_notifier,
1001                                         &gvt->shadow_ctx_notifier_block[i]);
1002                 kthread_stop(scheduler->thread[i]);
1003         }
1004 }
1005
1006 int intel_gvt_init_workload_scheduler(struct intel_gvt *gvt)
1007 {
1008         struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
1009         struct workload_thread_param *param = NULL;
1010         struct intel_engine_cs *engine;
1011         enum intel_engine_id i;
1012         int ret;
1013
1014         gvt_dbg_core("init workload scheduler\n");
1015
1016         init_waitqueue_head(&scheduler->workload_complete_wq);
1017
1018         for_each_engine(engine, gvt->dev_priv, i) {
1019                 init_waitqueue_head(&scheduler->waitq[i]);
1020
1021                 param = kzalloc(sizeof(*param), GFP_KERNEL);
1022                 if (!param) {
1023                         ret = -ENOMEM;
1024                         goto err;
1025                 }
1026
1027                 param->gvt = gvt;
1028                 param->ring_id = i;
1029
1030                 scheduler->thread[i] = kthread_run(workload_thread, param,
1031                         "gvt workload %d", i);
1032                 if (IS_ERR(scheduler->thread[i])) {
1033                         gvt_err("fail to create workload thread\n");
1034                         ret = PTR_ERR(scheduler->thread[i]);
1035                         goto err;
1036                 }
1037
1038                 gvt->shadow_ctx_notifier_block[i].notifier_call =
1039                                         shadow_context_status_change;
1040                 atomic_notifier_chain_register(&engine->context_status_notifier,
1041                                         &gvt->shadow_ctx_notifier_block[i]);
1042         }
1043         return 0;
1044 err:
1045         intel_gvt_clean_workload_scheduler(gvt);
1046         kfree(param);
1047         param = NULL;
1048         return ret;
1049 }
1050
1051 /**
1052  * intel_vgpu_clean_submission - free submission-related resource for vGPU
1053  * @vgpu: a vGPU
1054  *
1055  * This function is called when a vGPU is being destroyed.
1056  *
1057  */
1058 void intel_vgpu_clean_submission(struct intel_vgpu *vgpu)
1059 {
1060         struct intel_vgpu_submission *s = &vgpu->submission;
1061
1062         intel_vgpu_select_submission_ops(vgpu, ALL_ENGINES, 0);
1063         i915_gem_context_put(s->shadow_ctx);
1064         kmem_cache_destroy(s->workloads);
1065 }
1066
1067
1068 /**
1069  * intel_vgpu_reset_submission - reset submission-related resource for vGPU
1070  * @vgpu: a vGPU
1071  * @engine_mask: engines expected to be reset
1072  *
1073  * This function is called when a vGPU is being destroyed.
1074  *
1075  */
1076 void intel_vgpu_reset_submission(struct intel_vgpu *vgpu,
1077                 unsigned long engine_mask)
1078 {
1079         struct intel_vgpu_submission *s = &vgpu->submission;
1080
1081         if (!s->active)
1082                 return;
1083
1084         clean_workloads(vgpu, engine_mask);
1085         s->ops->reset(vgpu, engine_mask);
1086 }
1087
1088 /**
1089  * intel_vgpu_setup_submission - setup submission-related resource for vGPU
1090  * @vgpu: a vGPU
1091  *
1092  * This function is called when a vGPU is being created.
1093  *
1094  * Returns:
1095  * Zero on success, negative error code if failed.
1096  *
1097  */
1098 int intel_vgpu_setup_submission(struct intel_vgpu *vgpu)
1099 {
1100         struct intel_vgpu_submission *s = &vgpu->submission;
1101         enum intel_engine_id i;
1102         struct intel_engine_cs *engine;
1103         int ret;
1104
1105         s->shadow_ctx = i915_gem_context_create_gvt(
1106                         &vgpu->gvt->dev_priv->drm);
1107         if (IS_ERR(s->shadow_ctx))
1108                 return PTR_ERR(s->shadow_ctx);
1109
1110         bitmap_zero(s->shadow_ctx_desc_updated, I915_NUM_ENGINES);
1111
1112         s->workloads = kmem_cache_create_usercopy("gvt-g_vgpu_workload",
1113                                                   sizeof(struct intel_vgpu_workload), 0,
1114                                                   SLAB_HWCACHE_ALIGN,
1115                                                   offsetof(struct intel_vgpu_workload, rb_tail),
1116                                                   sizeof_field(struct intel_vgpu_workload, rb_tail),
1117                                                   NULL);
1118
1119         if (!s->workloads) {
1120                 ret = -ENOMEM;
1121                 goto out_shadow_ctx;
1122         }
1123
1124         for_each_engine(engine, vgpu->gvt->dev_priv, i)
1125                 INIT_LIST_HEAD(&s->workload_q_head[i]);
1126
1127         atomic_set(&s->running_workload_num, 0);
1128         bitmap_zero(s->tlb_handle_pending, I915_NUM_ENGINES);
1129
1130         return 0;
1131
1132 out_shadow_ctx:
1133         i915_gem_context_put(s->shadow_ctx);
1134         return ret;
1135 }
1136
1137 /**
1138  * intel_vgpu_select_submission_ops - select virtual submission interface
1139  * @vgpu: a vGPU
1140  * @interface: expected vGPU virtual submission interface
1141  *
1142  * This function is called when guest configures submission interface.
1143  *
1144  * Returns:
1145  * Zero on success, negative error code if failed.
1146  *
1147  */
1148 int intel_vgpu_select_submission_ops(struct intel_vgpu *vgpu,
1149                                      unsigned long engine_mask,
1150                                      unsigned int interface)
1151 {
1152         struct intel_vgpu_submission *s = &vgpu->submission;
1153         const struct intel_vgpu_submission_ops *ops[] = {
1154                 [INTEL_VGPU_EXECLIST_SUBMISSION] =
1155                         &intel_vgpu_execlist_submission_ops,
1156         };
1157         int ret;
1158
1159         if (WARN_ON(interface >= ARRAY_SIZE(ops)))
1160                 return -EINVAL;
1161
1162         if (WARN_ON(interface == 0 && engine_mask != ALL_ENGINES))
1163                 return -EINVAL;
1164
1165         if (s->active)
1166                 s->ops->clean(vgpu, engine_mask);
1167
1168         if (interface == 0) {
1169                 s->ops = NULL;
1170                 s->virtual_submission_interface = 0;
1171                 s->active = false;
1172                 gvt_dbg_core("vgpu%d: remove submission ops\n", vgpu->id);
1173                 return 0;
1174         }
1175
1176         ret = ops[interface]->init(vgpu, engine_mask);
1177         if (ret)
1178                 return ret;
1179
1180         s->ops = ops[interface];
1181         s->virtual_submission_interface = interface;
1182         s->active = true;
1183
1184         gvt_dbg_core("vgpu%d: activate ops [ %s ]\n",
1185                         vgpu->id, s->ops->name);
1186
1187         return 0;
1188 }
1189
1190 /**
1191  * intel_vgpu_destroy_workload - destroy a vGPU workload
1192  * @vgpu: a vGPU
1193  *
1194  * This function is called when destroy a vGPU workload.
1195  *
1196  */
1197 void intel_vgpu_destroy_workload(struct intel_vgpu_workload *workload)
1198 {
1199         struct intel_vgpu_submission *s = &workload->vgpu->submission;
1200
1201         if (workload->shadow_mm)
1202                 intel_vgpu_mm_put(workload->shadow_mm);
1203
1204         kmem_cache_free(s->workloads, workload);
1205 }
1206
1207 static struct intel_vgpu_workload *
1208 alloc_workload(struct intel_vgpu *vgpu)
1209 {
1210         struct intel_vgpu_submission *s = &vgpu->submission;
1211         struct intel_vgpu_workload *workload;
1212
1213         workload = kmem_cache_zalloc(s->workloads, GFP_KERNEL);
1214         if (!workload)
1215                 return ERR_PTR(-ENOMEM);
1216
1217         INIT_LIST_HEAD(&workload->list);
1218         INIT_LIST_HEAD(&workload->shadow_bb);
1219
1220         init_waitqueue_head(&workload->shadow_ctx_status_wq);
1221         atomic_set(&workload->shadow_ctx_active, 0);
1222
1223         workload->status = -EINPROGRESS;
1224         workload->vgpu = vgpu;
1225
1226         return workload;
1227 }
1228
1229 #define RING_CTX_OFF(x) \
1230         offsetof(struct execlist_ring_context, x)
1231
1232 static void read_guest_pdps(struct intel_vgpu *vgpu,
1233                 u64 ring_context_gpa, u32 pdp[8])
1234 {
1235         u64 gpa;
1236         int i;
1237
1238         gpa = ring_context_gpa + RING_CTX_OFF(pdps[0].val);
1239
1240         for (i = 0; i < 8; i++)
1241                 intel_gvt_hypervisor_read_gpa(vgpu,
1242                                 gpa + i * 8, &pdp[7 - i], 4);
1243 }
1244
1245 static int prepare_mm(struct intel_vgpu_workload *workload)
1246 {
1247         struct execlist_ctx_descriptor_format *desc = &workload->ctx_desc;
1248         struct intel_vgpu_mm *mm;
1249         struct intel_vgpu *vgpu = workload->vgpu;
1250         intel_gvt_gtt_type_t root_entry_type;
1251         u64 pdps[GVT_RING_CTX_NR_PDPS];
1252
1253         switch (desc->addressing_mode) {
1254         case 1: /* legacy 32-bit */
1255                 root_entry_type = GTT_TYPE_PPGTT_ROOT_L3_ENTRY;
1256                 break;
1257         case 3: /* legacy 64-bit */
1258                 root_entry_type = GTT_TYPE_PPGTT_ROOT_L4_ENTRY;
1259                 break;
1260         default:
1261                 gvt_vgpu_err("Advanced Context mode(SVM) is not supported!\n");
1262                 return -EINVAL;
1263         }
1264
1265         read_guest_pdps(workload->vgpu, workload->ring_context_gpa, (void *)pdps);
1266
1267         mm = intel_vgpu_get_ppgtt_mm(workload->vgpu, root_entry_type, pdps);
1268         if (IS_ERR(mm))
1269                 return PTR_ERR(mm);
1270
1271         workload->shadow_mm = mm;
1272         return 0;
1273 }
1274
1275 #define same_context(a, b) (((a)->context_id == (b)->context_id) && \
1276                 ((a)->lrca == (b)->lrca))
1277
1278 #define get_last_workload(q) \
1279         (list_empty(q) ? NULL : container_of(q->prev, \
1280         struct intel_vgpu_workload, list))
1281 /**
1282  * intel_vgpu_create_workload - create a vGPU workload
1283  * @vgpu: a vGPU
1284  * @desc: a guest context descriptor
1285  *
1286  * This function is called when creating a vGPU workload.
1287  *
1288  * Returns:
1289  * struct intel_vgpu_workload * on success, negative error code in
1290  * pointer if failed.
1291  *
1292  */
1293 struct intel_vgpu_workload *
1294 intel_vgpu_create_workload(struct intel_vgpu *vgpu, int ring_id,
1295                            struct execlist_ctx_descriptor_format *desc)
1296 {
1297         struct intel_vgpu_submission *s = &vgpu->submission;
1298         struct list_head *q = workload_q_head(vgpu, ring_id);
1299         struct intel_vgpu_workload *last_workload = get_last_workload(q);
1300         struct intel_vgpu_workload *workload = NULL;
1301         struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
1302         u64 ring_context_gpa;
1303         u32 head, tail, start, ctl, ctx_ctl, per_ctx, indirect_ctx;
1304         int ret;
1305
1306         ring_context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
1307                         (u32)((desc->lrca + 1) << I915_GTT_PAGE_SHIFT));
1308         if (ring_context_gpa == INTEL_GVT_INVALID_ADDR) {
1309                 gvt_vgpu_err("invalid guest context LRCA: %x\n", desc->lrca);
1310                 return ERR_PTR(-EINVAL);
1311         }
1312
1313         intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
1314                         RING_CTX_OFF(ring_header.val), &head, 4);
1315
1316         intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
1317                         RING_CTX_OFF(ring_tail.val), &tail, 4);
1318
1319         head &= RB_HEAD_OFF_MASK;
1320         tail &= RB_TAIL_OFF_MASK;
1321
1322         if (last_workload && same_context(&last_workload->ctx_desc, desc)) {
1323                 gvt_dbg_el("ring id %d cur workload == last\n", ring_id);
1324                 gvt_dbg_el("ctx head %x real head %lx\n", head,
1325                                 last_workload->rb_tail);
1326                 /*
1327                  * cannot use guest context head pointer here,
1328                  * as it might not be updated at this time
1329                  */
1330                 head = last_workload->rb_tail;
1331         }
1332
1333         gvt_dbg_el("ring id %d begin a new workload\n", ring_id);
1334
1335         /* record some ring buffer register values for scan and shadow */
1336         intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
1337                         RING_CTX_OFF(rb_start.val), &start, 4);
1338         intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
1339                         RING_CTX_OFF(rb_ctrl.val), &ctl, 4);
1340         intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
1341                         RING_CTX_OFF(ctx_ctrl.val), &ctx_ctl, 4);
1342
1343         workload = alloc_workload(vgpu);
1344         if (IS_ERR(workload))
1345                 return workload;
1346
1347         workload->ring_id = ring_id;
1348         workload->ctx_desc = *desc;
1349         workload->ring_context_gpa = ring_context_gpa;
1350         workload->rb_head = head;
1351         workload->rb_tail = tail;
1352         workload->rb_start = start;
1353         workload->rb_ctl = ctl;
1354
1355         if (ring_id == RCS) {
1356                 intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
1357                         RING_CTX_OFF(bb_per_ctx_ptr.val), &per_ctx, 4);
1358                 intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
1359                         RING_CTX_OFF(rcs_indirect_ctx.val), &indirect_ctx, 4);
1360
1361                 workload->wa_ctx.indirect_ctx.guest_gma =
1362                         indirect_ctx & INDIRECT_CTX_ADDR_MASK;
1363                 workload->wa_ctx.indirect_ctx.size =
1364                         (indirect_ctx & INDIRECT_CTX_SIZE_MASK) *
1365                         CACHELINE_BYTES;
1366                 workload->wa_ctx.per_ctx.guest_gma =
1367                         per_ctx & PER_CTX_ADDR_MASK;
1368                 workload->wa_ctx.per_ctx.valid = per_ctx & 1;
1369         }
1370
1371         gvt_dbg_el("workload %p ring id %d head %x tail %x start %x ctl %x\n",
1372                         workload, ring_id, head, tail, start, ctl);
1373
1374         ret = prepare_mm(workload);
1375         if (ret) {
1376                 kmem_cache_free(s->workloads, workload);
1377                 return ERR_PTR(ret);
1378         }
1379
1380         /* Only scan and shadow the first workload in the queue
1381          * as there is only one pre-allocated buf-obj for shadow.
1382          */
1383         if (list_empty(workload_q_head(vgpu, ring_id))) {
1384                 intel_runtime_pm_get(dev_priv);
1385                 mutex_lock(&dev_priv->drm.struct_mutex);
1386                 ret = intel_gvt_scan_and_shadow_workload(workload);
1387                 mutex_unlock(&dev_priv->drm.struct_mutex);
1388                 intel_runtime_pm_put(dev_priv);
1389         }
1390
1391         if (ret && (vgpu_is_vm_unhealthy(ret))) {
1392                 enter_failsafe_mode(vgpu, GVT_FAILSAFE_GUEST_ERR);
1393                 intel_vgpu_destroy_workload(workload);
1394                 return ERR_PTR(ret);
1395         }
1396
1397         return workload;
1398 }
1399
1400 /**
1401  * intel_vgpu_queue_workload - Qeue a vGPU workload
1402  * @workload: the workload to queue in
1403  */
1404 void intel_vgpu_queue_workload(struct intel_vgpu_workload *workload)
1405 {
1406         list_add_tail(&workload->list,
1407                 workload_q_head(workload->vgpu, workload->ring_id));
1408         intel_gvt_kick_schedule(workload->vgpu->gvt);
1409         wake_up(&workload->vgpu->gvt->scheduler.waitq[workload->ring_id]);
1410 }