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 "gem/i915_gem_context.h"
  39 #include "gem/i915_gem_pm.h"
  40 #include "gt/intel_context.h"
  41
  42 #include "i915_drv.h"
  43 #include "gvt.h"
  44
  45 #define RING_CTX_OFF(x) \
  46         offsetof(struct execlist_ring_context, x)
  47
  48 static void set_context_pdp_root_pointer(
  49                 struct execlist_ring_context *ring_context,
  50                 u32 pdp[8])
  51 {
  52         int i;
  53
  54         for (i = 0; i < 8; i++)
  55                 ring_context->pdps[i].val = pdp[7 - i];
  56 }
  57
  58 static void update_shadow_pdps(struct intel_vgpu_workload *workload)
  59 {
  60         struct drm_i915_gem_object *ctx_obj =
  61                 workload->req->hw_context->state->obj;
  62         struct execlist_ring_context *shadow_ring_context;
  63         struct page *page;
  64
  65         if (WARN_ON(!workload->shadow_mm))
  66                 return;
  67
  68         if (WARN_ON(!atomic_read(&workload->shadow_mm->pincount)))
  69                 return;
  70
  71         page = i915_gem_object_get_page(ctx_obj, LRC_STATE_PN);
  72         shadow_ring_context = kmap(page);
  73         set_context_pdp_root_pointer(shadow_ring_context,
  74                         (void *)workload->shadow_mm->ppgtt_mm.shadow_pdps);
  75         kunmap(page);
  76 }
  77
  78 /*
  79  * when populating shadow ctx from guest, we should not overrride oa related
  80  * registers, so that they will not be overlapped by guest oa configs. Thus
  81  * made it possible to capture oa data from host for both host and guests.
  82  */
  83 static void sr_oa_regs(struct intel_vgpu_workload *workload,
  84                 u32 *reg_state, bool save)
  85 {
  86         struct drm_i915_private *dev_priv = workload->vgpu->gvt->dev_priv;
  87         u32 ctx_oactxctrl = dev_priv->perf.ctx_oactxctrl_offset;
  88         u32 ctx_flexeu0 = dev_priv->perf.ctx_flexeu0_offset;
  89         int i = 0;
  90         u32 flex_mmio[] = {
  91                 i915_mmio_reg_offset(EU_PERF_CNTL0),
  92                 i915_mmio_reg_offset(EU_PERF_CNTL1),
  93                 i915_mmio_reg_offset(EU_PERF_CNTL2),
  94                 i915_mmio_reg_offset(EU_PERF_CNTL3),
  95                 i915_mmio_reg_offset(EU_PERF_CNTL4),
  96                 i915_mmio_reg_offset(EU_PERF_CNTL5),
  97                 i915_mmio_reg_offset(EU_PERF_CNTL6),
  98         };
  99
 100         if (workload->ring_id != RCS0)
 101                 return;
 102
 103         if (save) {
 104                 workload->oactxctrl = reg_state[ctx_oactxctrl + 1];
 105
 106                 for (i = 0; i < ARRAY_SIZE(workload->flex_mmio); i++) {
 107                         u32 state_offset = ctx_flexeu0 + i * 2;
 108
 109                         workload->flex_mmio[i] = reg_state[state_offset + 1];
 110                 }
 111         } else {
 112                 reg_state[ctx_oactxctrl] =
 113                         i915_mmio_reg_offset(GEN8_OACTXCONTROL);
 114                 reg_state[ctx_oactxctrl + 1] = workload->oactxctrl;
 115
 116                 for (i = 0; i < ARRAY_SIZE(workload->flex_mmio); i++) {
 117                         u32 state_offset = ctx_flexeu0 + i * 2;
 118                         u32 mmio = flex_mmio[i];
 119
 120                         reg_state[state_offset] = mmio;
 121                         reg_state[state_offset + 1] = workload->flex_mmio[i];
 122                 }
 123         }
 124 }
 125
 126 static int populate_shadow_context(struct intel_vgpu_workload *workload)
 127 {
 128         struct intel_vgpu *vgpu = workload->vgpu;
 129         struct intel_gvt *gvt = vgpu->gvt;
 130         int ring_id = workload->ring_id;
 131         struct drm_i915_gem_object *ctx_obj =
 132                 workload->req->hw_context->state->obj;
 133         struct execlist_ring_context *shadow_ring_context;
 134         struct page *page;
 135         void *dst;
 136         unsigned long context_gpa, context_page_num;
 137         int i;
 138
 139         page = i915_gem_object_get_page(ctx_obj, LRC_STATE_PN);
 140         shadow_ring_context = kmap(page);
 141
 142         sr_oa_regs(workload, (u32 *)shadow_ring_context, true);
 143 #define COPY_REG(name) \
 144         intel_gvt_hypervisor_read_gpa(vgpu, workload->ring_context_gpa \
 145                 + RING_CTX_OFF(name.val), &shadow_ring_context->name.val, 4)
 146 #define COPY_REG_MASKED(name) {\
 147                 intel_gvt_hypervisor_read_gpa(vgpu, workload->ring_context_gpa \
 148                                               + RING_CTX_OFF(name.val),\
 149                                               &shadow_ring_context->name.val, 4);\
 150                 shadow_ring_context->name.val |= 0xffff << 16;\
 151         }
 152
 153         COPY_REG_MASKED(ctx_ctrl);
 154         COPY_REG(ctx_timestamp);
 155
 156         if (ring_id == RCS0) {
 157                 COPY_REG(bb_per_ctx_ptr);
 158                 COPY_REG(rcs_indirect_ctx);
 159                 COPY_REG(rcs_indirect_ctx_offset);
 160         }
 161 #undef COPY_REG
 162 #undef COPY_REG_MASKED
 163
 164         intel_gvt_hypervisor_read_gpa(vgpu,
 165                         workload->ring_context_gpa +
 166                         sizeof(*shadow_ring_context),
 167                         (void *)shadow_ring_context +
 168                         sizeof(*shadow_ring_context),
 169                         I915_GTT_PAGE_SIZE - sizeof(*shadow_ring_context));
 170
 171         sr_oa_regs(workload, (u32 *)shadow_ring_context, false);
 172         kunmap(page);
 173
 174         if (IS_RESTORE_INHIBIT(shadow_ring_context->ctx_ctrl.val))
 175                 return 0;
 176
 177         gvt_dbg_sched("ring id %d workload lrca %x", ring_id,
 178                         workload->ctx_desc.lrca);
 179
 180         context_page_num = gvt->dev_priv->engine[ring_id]->context_size;
 181
 182         context_page_num = context_page_num >> PAGE_SHIFT;
 183
 184         if (IS_BROADWELL(gvt->dev_priv) && ring_id == RCS0)
 185                 context_page_num = 19;
 186
 187         i = 2;
 188         while (i < context_page_num) {
 189                 context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
 190                                 (u32)((workload->ctx_desc.lrca + i) <<
 191                                 I915_GTT_PAGE_SHIFT));
 192                 if (context_gpa == INTEL_GVT_INVALID_ADDR) {
 193                         gvt_vgpu_err("Invalid guest context descriptor\n");
 194                         return -EFAULT;
 195                 }
 196
 197                 page = i915_gem_object_get_page(ctx_obj, LRC_HEADER_PAGES + i);
 198                 dst = kmap(page);
 199                 intel_gvt_hypervisor_read_gpa(vgpu, context_gpa, dst,
 200                                 I915_GTT_PAGE_SIZE);
 201                 kunmap(page);
 202                 i++;
 203         }
 204         return 0;
 205 }
 206
 207 static inline bool is_gvt_request(struct i915_request *req)
 208 {
 209         return i915_gem_context_force_single_submission(req->gem_context);
 210 }
 211
 212 static void save_ring_hw_state(struct intel_vgpu *vgpu, int ring_id)
 213 {
 214         struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
 215         u32 ring_base = dev_priv->engine[ring_id]->mmio_base;
 216         i915_reg_t reg;
 217
 218         reg = RING_INSTDONE(ring_base);
 219         vgpu_vreg(vgpu, i915_mmio_reg_offset(reg)) = I915_READ_FW(reg);
 220         reg = RING_ACTHD(ring_base);
 221         vgpu_vreg(vgpu, i915_mmio_reg_offset(reg)) = I915_READ_FW(reg);
 222         reg = RING_ACTHD_UDW(ring_base);
 223         vgpu_vreg(vgpu, i915_mmio_reg_offset(reg)) = I915_READ_FW(reg);
 224 }
 225
 226 static int shadow_context_status_change(struct notifier_block *nb,
 227                 unsigned long action, void *data)
 228 {
 229         struct i915_request *req = data;
 230         struct intel_gvt *gvt = container_of(nb, struct intel_gvt,
 231                                 shadow_ctx_notifier_block[req->engine->id]);
 232         struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
 233         enum intel_engine_id ring_id = req->engine->id;
 234         struct intel_vgpu_workload *workload;
 235         unsigned long flags;
 236
 237         if (!is_gvt_request(req)) {
 238                 spin_lock_irqsave(&scheduler->mmio_context_lock, flags);
 239                 if (action == INTEL_CONTEXT_SCHEDULE_IN &&
 240                     scheduler->engine_owner[ring_id]) {
 241                         /* Switch ring from vGPU to host. */
 242                         intel_gvt_switch_mmio(scheduler->engine_owner[ring_id],
 243                                               NULL, ring_id);
 244                         scheduler->engine_owner[ring_id] = NULL;
 245                 }
 246                 spin_unlock_irqrestore(&scheduler->mmio_context_lock, flags);
 247
 248                 return NOTIFY_OK;
 249         }
 250
 251         workload = scheduler->current_workload[ring_id];
 252         if (unlikely(!workload))
 253                 return NOTIFY_OK;
 254
 255         switch (action) {
 256         case INTEL_CONTEXT_SCHEDULE_IN:
 257                 spin_lock_irqsave(&scheduler->mmio_context_lock, flags);
 258                 if (workload->vgpu != scheduler->engine_owner[ring_id]) {
 259                         /* Switch ring from host to vGPU or vGPU to vGPU. */
 260                         intel_gvt_switch_mmio(scheduler->engine_owner[ring_id],
 261                                               workload->vgpu, ring_id);
 262                         scheduler->engine_owner[ring_id] = workload->vgpu;
 263                 } else
 264                         gvt_dbg_sched("skip ring %d mmio switch for vgpu%d\n",
 265                                       ring_id, workload->vgpu->id);
 266                 spin_unlock_irqrestore(&scheduler->mmio_context_lock, flags);
 267                 atomic_set(&workload->shadow_ctx_active, 1);
 268                 break;
 269         case INTEL_CONTEXT_SCHEDULE_OUT:
 270                 save_ring_hw_state(workload->vgpu, ring_id);
 271                 atomic_set(&workload->shadow_ctx_active, 0);
 272                 break;
 273         case INTEL_CONTEXT_SCHEDULE_PREEMPTED:
 274                 save_ring_hw_state(workload->vgpu, ring_id);
 275                 break;
 276         default:
 277                 WARN_ON(1);
 278                 return NOTIFY_OK;
 279         }
 280         wake_up(&workload->shadow_ctx_status_wq);
 281         return NOTIFY_OK;
 282 }
 283
 284 static void
 285 shadow_context_descriptor_update(struct intel_context *ce,
 286                                  struct intel_vgpu_workload *workload)
 287 {
 288         u64 desc = ce->lrc_desc;
 289
 290         /*
 291          * Update bits 0-11 of the context descriptor which includes flags
 292          * like GEN8_CTX_* cached in desc_template
 293          */
 294         desc &= ~(0x3 << GEN8_CTX_ADDRESSING_MODE_SHIFT);
 295         desc |= workload->ctx_desc.addressing_mode <<
 296                 GEN8_CTX_ADDRESSING_MODE_SHIFT;
 297
 298         ce->lrc_desc = desc;
 299 }
 300
 301 static int copy_workload_to_ring_buffer(struct intel_vgpu_workload *workload)
 302 {
 303         struct intel_vgpu *vgpu = workload->vgpu;
 304         struct i915_request *req = workload->req;
 305         void *shadow_ring_buffer_va;
 306         u32 *cs;
 307         int err;
 308
 309         if (IS_GEN(req->i915, 9) && is_inhibit_context(req->hw_context))
 310                 intel_vgpu_restore_inhibit_context(vgpu, req);
 311
 312         /*
 313          * To track whether a request has started on HW, we can emit a
 314          * breadcrumb at the beginning of the request and check its
 315          * timeline's HWSP to see if the breadcrumb has advanced past the
 316          * start of this request. Actually, the request must have the
 317          * init_breadcrumb if its timeline set has_init_bread_crumb, or the
 318          * scheduler might get a wrong state of it during reset. Since the
 319          * requests from gvt always set the has_init_breadcrumb flag, here
 320          * need to do the emit_init_breadcrumb for all the requests.
 321          */
 322         if (req->engine->emit_init_breadcrumb) {
 323                 err = req->engine->emit_init_breadcrumb(req);
 324                 if (err) {
 325                         gvt_vgpu_err("fail to emit init breadcrumb\n");
 326                         return err;
 327                 }
 328         }
 329
 330         /* allocate shadow ring buffer */
 331         cs = intel_ring_begin(workload->req, workload->rb_len / sizeof(u32));
 332         if (IS_ERR(cs)) {
 333                 gvt_vgpu_err("fail to alloc size =%ld shadow  ring buffer\n",
 334                         workload->rb_len);
 335                 return PTR_ERR(cs);
 336         }
 337
 338         shadow_ring_buffer_va = workload->shadow_ring_buffer_va;
 339
 340         /* get shadow ring buffer va */
 341         workload->shadow_ring_buffer_va = cs;
 342
 343         memcpy(cs, shadow_ring_buffer_va,
 344                         workload->rb_len);
 345
 346         cs += workload->rb_len / sizeof(u32);
 347         intel_ring_advance(workload->req, cs);
 348
 349         return 0;
 350 }
 351
 352 static void release_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
 353 {
 354         if (!wa_ctx->indirect_ctx.obj)
 355                 return;
 356
 357         i915_gem_object_unpin_map(wa_ctx->indirect_ctx.obj);
 358         i915_gem_object_put(wa_ctx->indirect_ctx.obj);
 359
 360         wa_ctx->indirect_ctx.obj = NULL;
 361         wa_ctx->indirect_ctx.shadow_va = NULL;
 362 }
 363
 364 static void set_context_ppgtt_from_shadow(struct intel_vgpu_workload *workload,
 365                                           struct i915_gem_context *ctx)
 366 {
 367         struct intel_vgpu_mm *mm = workload->shadow_mm;
 368         struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(ctx->vm);
 369         int i = 0;
 370
 371         if (mm->ppgtt_mm.root_entry_type == GTT_TYPE_PPGTT_ROOT_L4_ENTRY) {
 372                 px_dma(ppgtt->pd) = mm->ppgtt_mm.shadow_pdps[0];
 373         } else {
 374                 for (i = 0; i < GVT_RING_CTX_NR_PDPS; i++) {
 375                         struct i915_page_directory * const pd =
 376                                 i915_pd_entry(ppgtt->pd, i);
 377                         /* skip now as current i915 ppgtt alloc won't allocate
 378                            top level pdp for non 4-level table, won't impact
 379                            shadow ppgtt. */
 380                         if (!pd)
 381                                 break;
 382                         px_dma(pd) = mm->ppgtt_mm.shadow_pdps[i];
 383                 }
 384         }
 385 }
 386
 387 static int
 388 intel_gvt_workload_req_alloc(struct intel_vgpu_workload *workload)
 389 {
 390         struct intel_vgpu *vgpu = workload->vgpu;
 391         struct intel_vgpu_submission *s = &vgpu->submission;
 392         struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
 393         struct i915_request *rq;
 394
 395         lockdep_assert_held(&dev_priv->drm.struct_mutex);
 396
 397         if (workload->req)
 398                 return 0;
 399
 400         rq = i915_request_create(s->shadow[workload->ring_id]);
 401         if (IS_ERR(rq)) {
 402                 gvt_vgpu_err("fail to allocate gem request\n");
 403                 return PTR_ERR(rq);
 404         }
 405
 406         workload->req = i915_request_get(rq);
 407         return 0;
 408 }
 409
 410 /**
 411  * intel_gvt_scan_and_shadow_workload - audit the workload by scanning and
 412  * shadow it as well, include ringbuffer,wa_ctx and ctx.
 413  * @workload: an abstract entity for each execlist submission.
 414  *
 415  * This function is called before the workload submitting to i915, to make
 416  * sure the content of the workload is valid.
 417  */
 418 int intel_gvt_scan_and_shadow_workload(struct intel_vgpu_workload *workload)
 419 {
 420         struct intel_vgpu *vgpu = workload->vgpu;
 421         struct intel_vgpu_submission *s = &vgpu->submission;
 422         struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
 423         int ret;
 424
 425         lockdep_assert_held(&dev_priv->drm.struct_mutex);
 426
 427         if (workload->shadow)
 428                 return 0;
 429
 430         if (!test_and_set_bit(workload->ring_id, s->shadow_ctx_desc_updated))
 431                 shadow_context_descriptor_update(s->shadow[workload->ring_id],
 432                                                  workload);
 433
 434         ret = intel_gvt_scan_and_shadow_ringbuffer(workload);
 435         if (ret)
 436                 return ret;
 437
 438         if (workload->ring_id == RCS0 && workload->wa_ctx.indirect_ctx.size) {
 439                 ret = intel_gvt_scan_and_shadow_wa_ctx(&workload->wa_ctx);
 440                 if (ret)
 441                         goto err_shadow;
 442         }
 443
 444         workload->shadow = true;
 445         return 0;
 446 err_shadow:
 447         release_shadow_wa_ctx(&workload->wa_ctx);
 448         return ret;
 449 }
 450
 451 static void release_shadow_batch_buffer(struct intel_vgpu_workload *workload);
 452
 453 static int prepare_shadow_batch_buffer(struct intel_vgpu_workload *workload)
 454 {
 455         struct intel_gvt *gvt = workload->vgpu->gvt;
 456         const int gmadr_bytes = gvt->device_info.gmadr_bytes_in_cmd;
 457         struct intel_vgpu_shadow_bb *bb;
 458         int ret;
 459
 460         list_for_each_entry(bb, &workload->shadow_bb, list) {
 461                 /* For privilge batch buffer and not wa_ctx, the bb_start_cmd_va
 462                  * is only updated into ring_scan_buffer, not real ring address
 463                  * allocated in later copy_workload_to_ring_buffer. pls be noted
 464                  * shadow_ring_buffer_va is now pointed to real ring buffer va
 465                  * in copy_workload_to_ring_buffer.
 466                  */
 467
 468                 if (bb->bb_offset)
 469                         bb->bb_start_cmd_va = workload->shadow_ring_buffer_va
 470                                 + bb->bb_offset;
 471
 472                 if (bb->ppgtt) {
 473                         /* for non-priv bb, scan&shadow is only for
 474                          * debugging purpose, so the content of shadow bb
 475                          * is the same as original bb. Therefore,
 476                          * here, rather than switch to shadow bb's gma
 477                          * address, we directly use original batch buffer's
 478                          * gma address, and send original bb to hardware
 479                          * directly
 480                          */
 481                         if (bb->clflush & CLFLUSH_AFTER) {
 482                                 drm_clflush_virt_range(bb->va,
 483                                                 bb->obj->base.size);
 484                                 bb->clflush &= ~CLFLUSH_AFTER;
 485                         }
 486                         i915_gem_object_finish_access(bb->obj);
 487                         bb->accessing = false;
 488
 489                 } else {
 490                         bb->vma = i915_gem_object_ggtt_pin(bb->obj,
 491                                         NULL, 0, 0, 0);
 492                         if (IS_ERR(bb->vma)) {
 493                                 ret = PTR_ERR(bb->vma);
 494                                 goto err;
 495                         }
 496
 497                         /* relocate shadow batch buffer */
 498                         bb->bb_start_cmd_va[1] = i915_ggtt_offset(bb->vma);
 499                         if (gmadr_bytes == 8)
 500                                 bb->bb_start_cmd_va[2] = 0;
 501
 502                         /* No one is going to touch shadow bb from now on. */
 503                         if (bb->clflush & CLFLUSH_AFTER) {
 504                                 drm_clflush_virt_range(bb->va,
 505                                                 bb->obj->base.size);
 506                                 bb->clflush &= ~CLFLUSH_AFTER;
 507                         }
 508
 509                         ret = i915_gem_object_set_to_gtt_domain(bb->obj,
 510                                                                 false);
 511                         if (ret)
 512                                 goto err;
 513
 514                         ret = i915_vma_move_to_active(bb->vma,
 515                                                       workload->req,
 516                                                       0);
 517                         if (ret)
 518                                 goto err;
 519
 520                         i915_gem_object_finish_access(bb->obj);
 521                         bb->accessing = false;
 522                 }
 523         }
 524         return 0;
 525 err:
 526         release_shadow_batch_buffer(workload);
 527         return ret;
 528 }
 529
 530 static void update_wa_ctx_2_shadow_ctx(struct intel_shadow_wa_ctx *wa_ctx)
 531 {
 532         struct intel_vgpu_workload *workload =
 533                 container_of(wa_ctx, struct intel_vgpu_workload, wa_ctx);
 534         struct i915_request *rq = workload->req;
 535         struct execlist_ring_context *shadow_ring_context =
 536                 (struct execlist_ring_context *)rq->hw_context->lrc_reg_state;
 537
 538         shadow_ring_context->bb_per_ctx_ptr.val =
 539                 (shadow_ring_context->bb_per_ctx_ptr.val &
 540                 (~PER_CTX_ADDR_MASK)) | wa_ctx->per_ctx.shadow_gma;
 541         shadow_ring_context->rcs_indirect_ctx.val =
 542                 (shadow_ring_context->rcs_indirect_ctx.val &
 543                 (~INDIRECT_CTX_ADDR_MASK)) | wa_ctx->indirect_ctx.shadow_gma;
 544 }
 545
 546 static int prepare_shadow_wa_ctx(struct intel_shadow_wa_ctx *wa_ctx)
 547 {
 548         struct i915_vma *vma;
 549         unsigned char *per_ctx_va =
 550                 (unsigned char *)wa_ctx->indirect_ctx.shadow_va +
 551                 wa_ctx->indirect_ctx.size;
 552
 553         if (wa_ctx->indirect_ctx.size == 0)
 554                 return 0;
 555
 556         vma = i915_gem_object_ggtt_pin(wa_ctx->indirect_ctx.obj, NULL,
 557                                        0, CACHELINE_BYTES, 0);
 558         if (IS_ERR(vma))
 559                 return PTR_ERR(vma);
 560
 561         /* FIXME: we are not tracking our pinned VMA leaving it
 562          * up to the core to fix up the stray pin_count upon
 563          * free.
 564          */
 565
 566         wa_ctx->indirect_ctx.shadow_gma = i915_ggtt_offset(vma);
 567
 568         wa_ctx->per_ctx.shadow_gma = *((unsigned int *)per_ctx_va + 1);
 569         memset(per_ctx_va, 0, CACHELINE_BYTES);
 570
 571         update_wa_ctx_2_shadow_ctx(wa_ctx);
 572         return 0;
 573 }
 574
 575 static void update_vreg_in_ctx(struct intel_vgpu_workload *workload)
 576 {
 577         struct intel_vgpu *vgpu = workload->vgpu;
 578         struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
 579         u32 ring_base;
 580
 581         ring_base = dev_priv->engine[workload->ring_id]->mmio_base;
 582         vgpu_vreg_t(vgpu, RING_START(ring_base)) = workload->rb_start;
 583 }
 584
 585 static void release_shadow_batch_buffer(struct intel_vgpu_workload *workload)
 586 {
 587         struct intel_vgpu *vgpu = workload->vgpu;
 588         struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
 589         struct intel_vgpu_shadow_bb *bb, *pos;
 590
 591         if (list_empty(&workload->shadow_bb))
 592                 return;
 593
 594         bb = list_first_entry(&workload->shadow_bb,
 595                         struct intel_vgpu_shadow_bb, list);
 596
 597         mutex_lock(&dev_priv->drm.struct_mutex);
 598
 599         list_for_each_entry_safe(bb, pos, &workload->shadow_bb, list) {
 600                 if (bb->obj) {
 601                         if (bb->accessing)
 602                                 i915_gem_object_finish_access(bb->obj);
 603
 604                         if (bb->va && !IS_ERR(bb->va))
 605                                 i915_gem_object_unpin_map(bb->obj);
 606
 607                         if (bb->vma && !IS_ERR(bb->vma)) {
 608                                 i915_vma_unpin(bb->vma);
 609                                 i915_vma_close(bb->vma);
 610                         }
 611                         i915_gem_object_put(bb->obj);
 612                 }
 613                 list_del(&bb->list);
 614                 kfree(bb);
 615         }
 616
 617         mutex_unlock(&dev_priv->drm.struct_mutex);
 618 }
 619
 620 static int prepare_workload(struct intel_vgpu_workload *workload)
 621 {
 622         struct intel_vgpu *vgpu = workload->vgpu;
 623         struct intel_vgpu_submission *s = &vgpu->submission;
 624         int ring = workload->ring_id;
 625         int ret = 0;
 626
 627         ret = intel_vgpu_pin_mm(workload->shadow_mm);
 628         if (ret) {
 629                 gvt_vgpu_err("fail to vgpu pin mm\n");
 630                 return ret;
 631         }
 632
 633         if (workload->shadow_mm->type != INTEL_GVT_MM_PPGTT ||
 634             !workload->shadow_mm->ppgtt_mm.shadowed) {
 635                 gvt_vgpu_err("workload shadow ppgtt isn't ready\n");
 636                 return -EINVAL;
 637         }
 638
 639         update_shadow_pdps(workload);
 640
 641         set_context_ppgtt_from_shadow(workload, s->shadow[ring]->gem_context);
 642
 643         ret = intel_vgpu_sync_oos_pages(workload->vgpu);
 644         if (ret) {
 645                 gvt_vgpu_err("fail to vgpu sync oos pages\n");
 646                 goto err_unpin_mm;
 647         }
 648
 649         ret = intel_vgpu_flush_post_shadow(workload->vgpu);
 650         if (ret) {
 651                 gvt_vgpu_err("fail to flush post shadow\n");
 652                 goto err_unpin_mm;
 653         }
 654
 655         ret = copy_workload_to_ring_buffer(workload);
 656         if (ret) {
 657                 gvt_vgpu_err("fail to generate request\n");
 658                 goto err_unpin_mm;
 659         }
 660
 661         ret = prepare_shadow_batch_buffer(workload);
 662         if (ret) {
 663                 gvt_vgpu_err("fail to prepare_shadow_batch_buffer\n");
 664                 goto err_unpin_mm;
 665         }
 666
 667         ret = prepare_shadow_wa_ctx(&workload->wa_ctx);
 668         if (ret) {
 669                 gvt_vgpu_err("fail to prepare_shadow_wa_ctx\n");
 670                 goto err_shadow_batch;
 671         }
 672
 673         if (workload->prepare) {
 674                 ret = workload->prepare(workload);
 675                 if (ret)
 676                         goto err_shadow_wa_ctx;
 677         }
 678
 679         return 0;
 680 err_shadow_wa_ctx:
 681         release_shadow_wa_ctx(&workload->wa_ctx);
 682 err_shadow_batch:
 683         release_shadow_batch_buffer(workload);
 684 err_unpin_mm:
 685         intel_vgpu_unpin_mm(workload->shadow_mm);
 686         return ret;
 687 }
 688
 689 static int dispatch_workload(struct intel_vgpu_workload *workload)
 690 {
 691         struct intel_vgpu *vgpu = workload->vgpu;
 692         struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
 693         struct i915_request *rq;
 694         int ring_id = workload->ring_id;
 695         int ret;
 696
 697         gvt_dbg_sched("ring id %d prepare to dispatch workload %p\n",
 698                 ring_id, workload);
 699
 700         mutex_lock(&vgpu->vgpu_lock);
 701         mutex_lock(&dev_priv->drm.struct_mutex);
 702
 703         ret = intel_gvt_workload_req_alloc(workload);
 704         if (ret)
 705                 goto err_req;
 706
 707         ret = intel_gvt_scan_and_shadow_workload(workload);
 708         if (ret)
 709                 goto out;
 710
 711         ret = populate_shadow_context(workload);
 712         if (ret) {
 713                 release_shadow_wa_ctx(&workload->wa_ctx);
 714                 goto out;
 715         }
 716
 717         ret = prepare_workload(workload);
 718 out:
 719         if (ret) {
 720                 /* We might still need to add request with
 721                  * clean ctx to retire it properly..
 722                  */
 723                 rq = fetch_and_zero(&workload->req);
 724                 i915_request_put(rq);
 725         }
 726
 727         if (!IS_ERR_OR_NULL(workload->req)) {
 728                 gvt_dbg_sched("ring id %d submit workload to i915 %p\n",
 729                                 ring_id, workload->req);
 730                 i915_request_add(workload->req);
 731                 workload->dispatched = true;
 732         }
 733 err_req:
 734         if (ret)
 735                 workload->status = ret;
 736         mutex_unlock(&dev_priv->drm.struct_mutex);
 737         mutex_unlock(&vgpu->vgpu_lock);
 738         return ret;
 739 }
 740
 741 static struct intel_vgpu_workload *pick_next_workload(
 742                 struct intel_gvt *gvt, int ring_id)
 743 {
 744         struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
 745         struct intel_vgpu_workload *workload = NULL;
 746
 747         mutex_lock(&gvt->sched_lock);
 748
 749         /*
 750          * no current vgpu / will be scheduled out / no workload
 751          * bail out
 752          */
 753         if (!scheduler->current_vgpu) {
 754                 gvt_dbg_sched("ring id %d stop - no current vgpu\n", ring_id);
 755                 goto out;
 756         }
 757
 758         if (scheduler->need_reschedule) {
 759                 gvt_dbg_sched("ring id %d stop - will reschedule\n", ring_id);
 760                 goto out;
 761         }
 762
 763         if (!scheduler->current_vgpu->active ||
 764             list_empty(workload_q_head(scheduler->current_vgpu, ring_id)))
 765                 goto out;
 766
 767         /*
 768          * still have current workload, maybe the workload disptacher
 769          * fail to submit it for some reason, resubmit it.
 770          */
 771         if (scheduler->current_workload[ring_id]) {
 772                 workload = scheduler->current_workload[ring_id];
 773                 gvt_dbg_sched("ring id %d still have current workload %p\n",
 774                                 ring_id, workload);
 775                 goto out;
 776         }
 777
 778         /*
 779          * pick a workload as current workload
 780          * once current workload is set, schedule policy routines
 781          * will wait the current workload is finished when trying to
 782          * schedule out a vgpu.
 783          */
 784         scheduler->current_workload[ring_id] = container_of(
 785                         workload_q_head(scheduler->current_vgpu, ring_id)->next,
 786                         struct intel_vgpu_workload, list);
 787
 788         workload = scheduler->current_workload[ring_id];
 789
 790         gvt_dbg_sched("ring id %d pick new workload %p\n", ring_id, workload);
 791
 792         atomic_inc(&workload->vgpu->submission.running_workload_num);
 793 out:
 794         mutex_unlock(&gvt->sched_lock);
 795         return workload;
 796 }
 797
 798 static void update_guest_context(struct intel_vgpu_workload *workload)
 799 {
 800         struct i915_request *rq = workload->req;
 801         struct intel_vgpu *vgpu = workload->vgpu;
 802         struct intel_gvt *gvt = vgpu->gvt;
 803         struct drm_i915_gem_object *ctx_obj = rq->hw_context->state->obj;
 804         struct execlist_ring_context *shadow_ring_context;
 805         struct page *page;
 806         void *src;
 807         unsigned long context_gpa, context_page_num;
 808         int i;
 809         struct drm_i915_private *dev_priv = gvt->dev_priv;
 810         u32 ring_base;
 811         u32 head, tail;
 812         u16 wrap_count;
 813
 814         gvt_dbg_sched("ring id %d workload lrca %x\n", rq->engine->id,
 815                       workload->ctx_desc.lrca);
 816
 817         head = workload->rb_head;
 818         tail = workload->rb_tail;
 819         wrap_count = workload->guest_rb_head >> RB_HEAD_WRAP_CNT_OFF;
 820
 821         if (tail < head) {
 822                 if (wrap_count == RB_HEAD_WRAP_CNT_MAX)
 823                         wrap_count = 0;
 824                 else
 825                         wrap_count += 1;
 826         }
 827
 828         head = (wrap_count << RB_HEAD_WRAP_CNT_OFF) | tail;
 829
 830         ring_base = dev_priv->engine[workload->ring_id]->mmio_base;
 831         vgpu_vreg_t(vgpu, RING_TAIL(ring_base)) = tail;
 832         vgpu_vreg_t(vgpu, RING_HEAD(ring_base)) = head;
 833
 834         context_page_num = rq->engine->context_size;
 835         context_page_num = context_page_num >> PAGE_SHIFT;
 836
 837         if (IS_BROADWELL(gvt->dev_priv) && rq->engine->id == RCS0)
 838                 context_page_num = 19;
 839
 840         i = 2;
 841
 842         while (i < context_page_num) {
 843                 context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
 844                                 (u32)((workload->ctx_desc.lrca + i) <<
 845                                         I915_GTT_PAGE_SHIFT));
 846                 if (context_gpa == INTEL_GVT_INVALID_ADDR) {
 847                         gvt_vgpu_err("invalid guest context descriptor\n");
 848                         return;
 849                 }
 850
 851                 page = i915_gem_object_get_page(ctx_obj, LRC_HEADER_PAGES + i);
 852                 src = kmap(page);
 853                 intel_gvt_hypervisor_write_gpa(vgpu, context_gpa, src,
 854                                 I915_GTT_PAGE_SIZE);
 855                 kunmap(page);
 856                 i++;
 857         }
 858
 859         intel_gvt_hypervisor_write_gpa(vgpu, workload->ring_context_gpa +
 860                 RING_CTX_OFF(ring_header.val), &workload->rb_tail, 4);
 861
 862         page = i915_gem_object_get_page(ctx_obj, LRC_STATE_PN);
 863         shadow_ring_context = kmap(page);
 864
 865 #define COPY_REG(name) \
 866         intel_gvt_hypervisor_write_gpa(vgpu, workload->ring_context_gpa + \
 867                 RING_CTX_OFF(name.val), &shadow_ring_context->name.val, 4)
 868
 869         COPY_REG(ctx_ctrl);
 870         COPY_REG(ctx_timestamp);
 871
 872 #undef COPY_REG
 873
 874         intel_gvt_hypervisor_write_gpa(vgpu,
 875                         workload->ring_context_gpa +
 876                         sizeof(*shadow_ring_context),
 877                         (void *)shadow_ring_context +
 878                         sizeof(*shadow_ring_context),
 879                         I915_GTT_PAGE_SIZE - sizeof(*shadow_ring_context));
 880
 881         kunmap(page);
 882 }
 883
 884 void intel_vgpu_clean_workloads(struct intel_vgpu *vgpu,
 885                                 intel_engine_mask_t engine_mask)
 886 {
 887         struct intel_vgpu_submission *s = &vgpu->submission;
 888         struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
 889         struct intel_engine_cs *engine;
 890         struct intel_vgpu_workload *pos, *n;
 891         intel_engine_mask_t tmp;
 892
 893         /* free the unsubmited workloads in the queues. */
 894         for_each_engine_masked(engine, dev_priv, engine_mask, tmp) {
 895                 list_for_each_entry_safe(pos, n,
 896                         &s->workload_q_head[engine->id], list) {
 897                         list_del_init(&pos->list);
 898                         intel_vgpu_destroy_workload(pos);
 899                 }
 900                 clear_bit(engine->id, s->shadow_ctx_desc_updated);
 901         }
 902 }
 903
 904 static void complete_current_workload(struct intel_gvt *gvt, int ring_id)
 905 {
 906         struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
 907         struct intel_vgpu_workload *workload =
 908                 scheduler->current_workload[ring_id];
 909         struct intel_vgpu *vgpu = workload->vgpu;
 910         struct intel_vgpu_submission *s = &vgpu->submission;
 911         struct i915_request *rq = workload->req;
 912         int event;
 913
 914         mutex_lock(&vgpu->vgpu_lock);
 915         mutex_lock(&gvt->sched_lock);
 916
 917         /* For the workload w/ request, needs to wait for the context
 918          * switch to make sure request is completed.
 919          * For the workload w/o request, directly complete the workload.
 920          */
 921         if (rq) {
 922                 wait_event(workload->shadow_ctx_status_wq,
 923                            !atomic_read(&workload->shadow_ctx_active));
 924
 925                 /* If this request caused GPU hang, req->fence.error will
 926                  * be set to -EIO. Use -EIO to set workload status so
 927                  * that when this request caused GPU hang, didn't trigger
 928                  * context switch interrupt to guest.
 929                  */
 930                 if (likely(workload->status == -EINPROGRESS)) {
 931                         if (workload->req->fence.error == -EIO)
 932                                 workload->status = -EIO;
 933                         else
 934                                 workload->status = 0;
 935                 }
 936
 937                 if (!workload->status &&
 938                     !(vgpu->resetting_eng & BIT(ring_id))) {
 939                         update_guest_context(workload);
 940
 941                         for_each_set_bit(event, workload->pending_events,
 942                                          INTEL_GVT_EVENT_MAX)
 943                                 intel_vgpu_trigger_virtual_event(vgpu, event);
 944                 }
 945
 946                 i915_request_put(fetch_and_zero(&workload->req));
 947         }
 948
 949         gvt_dbg_sched("ring id %d complete workload %p status %d\n",
 950                         ring_id, workload, workload->status);
 951
 952         scheduler->current_workload[ring_id] = NULL;
 953
 954         list_del_init(&workload->list);
 955
 956         if (workload->status || vgpu->resetting_eng & BIT(ring_id)) {
 957                 /* if workload->status is not successful means HW GPU
 958                  * has occurred GPU hang or something wrong with i915/GVT,
 959                  * and GVT won't inject context switch interrupt to guest.
 960                  * So this error is a vGPU hang actually to the guest.
 961                  * According to this we should emunlate a vGPU hang. If
 962                  * there are pending workloads which are already submitted
 963                  * from guest, we should clean them up like HW GPU does.
 964                  *
 965                  * if it is in middle of engine resetting, the pending
 966                  * workloads won't be submitted to HW GPU and will be
 967                  * cleaned up during the resetting process later, so doing
 968                  * the workload clean up here doesn't have any impact.
 969                  **/
 970                 intel_vgpu_clean_workloads(vgpu, BIT(ring_id));
 971         }
 972
 973         workload->complete(workload);
 974
 975         atomic_dec(&s->running_workload_num);
 976         wake_up(&scheduler->workload_complete_wq);
 977
 978         if (gvt->scheduler.need_reschedule)
 979                 intel_gvt_request_service(gvt, INTEL_GVT_REQUEST_EVENT_SCHED);
 980
 981         mutex_unlock(&gvt->sched_lock);
 982         mutex_unlock(&vgpu->vgpu_lock);
 983 }
 984
 985 struct workload_thread_param {
 986         struct intel_gvt *gvt;
 987         int ring_id;
 988 };
 989
 990 static int workload_thread(void *priv)
 991 {
 992         struct workload_thread_param *p = (struct workload_thread_param *)priv;
 993         struct intel_gvt *gvt = p->gvt;
 994         int ring_id = p->ring_id;
 995         struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
 996         struct intel_vgpu_workload *workload = NULL;
 997         struct intel_vgpu *vgpu = NULL;
 998         int ret;
 999         bool need_force_wake = (INTEL_GEN(gvt->dev_priv) >= 9);
1000         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1001         struct intel_runtime_pm *rpm = &gvt->dev_priv->runtime_pm;
1002
1003         kfree(p);
1004
1005         gvt_dbg_core("workload thread for ring %d started\n", ring_id);
1006
1007         while (!kthread_should_stop()) {
1008                 add_wait_queue(&scheduler->waitq[ring_id], &wait);
1009                 do {
1010                         workload = pick_next_workload(gvt, ring_id);
1011                         if (workload)
1012                                 break;
1013                         wait_woken(&wait, TASK_INTERRUPTIBLE,
1014                                    MAX_SCHEDULE_TIMEOUT);
1015                 } while (!kthread_should_stop());
1016                 remove_wait_queue(&scheduler->waitq[ring_id], &wait);
1017
1018                 if (!workload)
1019                         break;
1020
1021                 gvt_dbg_sched("ring id %d next workload %p vgpu %d\n",
1022                                 workload->ring_id, workload,
1023                                 workload->vgpu->id);
1024
1025                 intel_runtime_pm_get(rpm);
1026
1027                 gvt_dbg_sched("ring id %d will dispatch workload %p\n",
1028                                 workload->ring_id, workload);
1029
1030                 if (need_force_wake)
1031                         intel_uncore_forcewake_get(&gvt->dev_priv->uncore,
1032                                         FORCEWAKE_ALL);
1033                 /*
1034                  * Update the vReg of the vGPU which submitted this
1035                  * workload. The vGPU may use these registers for checking
1036                  * the context state. The value comes from GPU commands
1037                  * in this workload.
1038                  */
1039                 update_vreg_in_ctx(workload);
1040
1041                 ret = dispatch_workload(workload);
1042
1043                 if (ret) {
1044                         vgpu = workload->vgpu;
1045                         gvt_vgpu_err("fail to dispatch workload, skip\n");
1046                         goto complete;
1047                 }
1048
1049                 gvt_dbg_sched("ring id %d wait workload %p\n",
1050                                 workload->ring_id, workload);
1051                 i915_request_wait(workload->req, 0, MAX_SCHEDULE_TIMEOUT);
1052
1053 complete:
1054                 gvt_dbg_sched("will complete workload %p, status: %d\n",
1055                                 workload, workload->status);
1056
1057                 complete_current_workload(gvt, ring_id);
1058
1059                 if (need_force_wake)
1060                         intel_uncore_forcewake_put(&gvt->dev_priv->uncore,
1061                                         FORCEWAKE_ALL);
1062
1063                 intel_runtime_pm_put_unchecked(rpm);
1064                 if (ret && (vgpu_is_vm_unhealthy(ret)))
1065                         enter_failsafe_mode(vgpu, GVT_FAILSAFE_GUEST_ERR);
1066         }
1067         return 0;
1068 }
1069
1070 void intel_gvt_wait_vgpu_idle(struct intel_vgpu *vgpu)
1071 {
1072         struct intel_vgpu_submission *s = &vgpu->submission;
1073         struct intel_gvt *gvt = vgpu->gvt;
1074         struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
1075
1076         if (atomic_read(&s->running_workload_num)) {
1077                 gvt_dbg_sched("wait vgpu idle\n");
1078
1079                 wait_event(scheduler->workload_complete_wq,
1080                                 !atomic_read(&s->running_workload_num));
1081         }
1082 }
1083
1084 void intel_gvt_clean_workload_scheduler(struct intel_gvt *gvt)
1085 {
1086         struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
1087         struct intel_engine_cs *engine;
1088         enum intel_engine_id i;
1089
1090         gvt_dbg_core("clean workload scheduler\n");
1091
1092         for_each_engine(engine, gvt->dev_priv, i) {
1093                 atomic_notifier_chain_unregister(
1094                                         &engine->context_status_notifier,
1095                                         &gvt->shadow_ctx_notifier_block[i]);
1096                 kthread_stop(scheduler->thread[i]);
1097         }
1098 }
1099
1100 int intel_gvt_init_workload_scheduler(struct intel_gvt *gvt)
1101 {
1102         struct intel_gvt_workload_scheduler *scheduler = &gvt->scheduler;
1103         struct workload_thread_param *param = NULL;
1104         struct intel_engine_cs *engine;
1105         enum intel_engine_id i;
1106         int ret;
1107
1108         gvt_dbg_core("init workload scheduler\n");
1109
1110         init_waitqueue_head(&scheduler->workload_complete_wq);
1111
1112         for_each_engine(engine, gvt->dev_priv, i) {
1113                 init_waitqueue_head(&scheduler->waitq[i]);
1114
1115                 param = kzalloc(sizeof(*param), GFP_KERNEL);
1116                 if (!param) {
1117                         ret = -ENOMEM;
1118                         goto err;
1119                 }
1120
1121                 param->gvt = gvt;
1122                 param->ring_id = i;
1123
1124                 scheduler->thread[i] = kthread_run(workload_thread, param,
1125                         "gvt workload %d", i);
1126                 if (IS_ERR(scheduler->thread[i])) {
1127                         gvt_err("fail to create workload thread\n");
1128                         ret = PTR_ERR(scheduler->thread[i]);
1129                         goto err;
1130                 }
1131
1132                 gvt->shadow_ctx_notifier_block[i].notifier_call =
1133                                         shadow_context_status_change;
1134                 atomic_notifier_chain_register(&engine->context_status_notifier,
1135                                         &gvt->shadow_ctx_notifier_block[i]);
1136         }
1137         return 0;
1138 err:
1139         intel_gvt_clean_workload_scheduler(gvt);
1140         kfree(param);
1141         param = NULL;
1142         return ret;
1143 }
1144
1145 static void
1146 i915_context_ppgtt_root_restore(struct intel_vgpu_submission *s,
1147                                 struct i915_ppgtt *ppgtt)
1148 {
1149         int i;
1150
1151         if (i915_vm_is_4lvl(&ppgtt->vm)) {
1152                 px_dma(ppgtt->pd) = s->i915_context_pml4;
1153         } else {
1154                 for (i = 0; i < GEN8_3LVL_PDPES; i++) {
1155                         struct i915_page_directory * const pd =
1156                                 i915_pd_entry(ppgtt->pd, i);
1157
1158                         px_dma(pd) = s->i915_context_pdps[i];
1159                 }
1160         }
1161 }
1162
1163 /**
1164  * intel_vgpu_clean_submission - free submission-related resource for vGPU
1165  * @vgpu: a vGPU
1166  *
1167  * This function is called when a vGPU is being destroyed.
1168  *
1169  */
1170 void intel_vgpu_clean_submission(struct intel_vgpu *vgpu)
1171 {
1172         struct intel_vgpu_submission *s = &vgpu->submission;
1173         struct intel_engine_cs *engine;
1174         enum intel_engine_id id;
1175
1176         intel_vgpu_select_submission_ops(vgpu, ALL_ENGINES, 0);
1177
1178         i915_context_ppgtt_root_restore(s, i915_vm_to_ppgtt(s->shadow[0]->vm));
1179         for_each_engine(engine, vgpu->gvt->dev_priv, id)
1180                 intel_context_unpin(s->shadow[id]);
1181
1182         kmem_cache_destroy(s->workloads);
1183 }
1184
1185
1186 /**
1187  * intel_vgpu_reset_submission - reset submission-related resource for vGPU
1188  * @vgpu: a vGPU
1189  * @engine_mask: engines expected to be reset
1190  *
1191  * This function is called when a vGPU is being destroyed.
1192  *
1193  */
1194 void intel_vgpu_reset_submission(struct intel_vgpu *vgpu,
1195                                  intel_engine_mask_t engine_mask)
1196 {
1197         struct intel_vgpu_submission *s = &vgpu->submission;
1198
1199         if (!s->active)
1200                 return;
1201
1202         intel_vgpu_clean_workloads(vgpu, engine_mask);
1203         s->ops->reset(vgpu, engine_mask);
1204 }
1205
1206 static void
1207 i915_context_ppgtt_root_save(struct intel_vgpu_submission *s,
1208                              struct i915_ppgtt *ppgtt)
1209 {
1210         int i;
1211
1212         if (i915_vm_is_4lvl(&ppgtt->vm)) {
1213                 s->i915_context_pml4 = px_dma(ppgtt->pd);
1214         } else {
1215                 for (i = 0; i < GEN8_3LVL_PDPES; i++) {
1216                         struct i915_page_directory * const pd =
1217                                 i915_pd_entry(ppgtt->pd, i);
1218
1219                         s->i915_context_pdps[i] = px_dma(pd);
1220                 }
1221         }
1222 }
1223
1224 /**
1225  * intel_vgpu_setup_submission - setup submission-related resource for vGPU
1226  * @vgpu: a vGPU
1227  *
1228  * This function is called when a vGPU is being created.
1229  *
1230  * Returns:
1231  * Zero on success, negative error code if failed.
1232  *
1233  */
1234 int intel_vgpu_setup_submission(struct intel_vgpu *vgpu)
1235 {
1236         struct drm_i915_private *i915 = vgpu->gvt->dev_priv;
1237         struct intel_vgpu_submission *s = &vgpu->submission;
1238         struct intel_engine_cs *engine;
1239         struct i915_gem_context *ctx;
1240         enum intel_engine_id i;
1241         int ret;
1242
1243         mutex_lock(&i915->drm.struct_mutex);
1244
1245         ctx = i915_gem_context_create_kernel(i915, I915_PRIORITY_MAX);
1246         if (IS_ERR(ctx)) {
1247                 ret = PTR_ERR(ctx);
1248                 goto out_unlock;
1249         }
1250
1251         i915_gem_context_set_force_single_submission(ctx);
1252
1253         i915_context_ppgtt_root_save(s, i915_vm_to_ppgtt(ctx->vm));
1254
1255         for_each_engine(engine, i915, i) {
1256                 struct intel_context *ce;
1257
1258                 INIT_LIST_HEAD(&s->workload_q_head[i]);
1259                 s->shadow[i] = ERR_PTR(-EINVAL);
1260
1261                 ce = intel_context_create(ctx, engine);
1262                 if (IS_ERR(ce)) {
1263                         ret = PTR_ERR(ce);
1264                         goto out_shadow_ctx;
1265                 }
1266
1267                 if (!USES_GUC_SUBMISSION(i915)) { /* Max ring buffer size */
1268                         const unsigned int ring_size = 512 * SZ_4K;
1269
1270                         ce->ring = __intel_context_ring_size(ring_size);
1271                 }
1272
1273                 ret = intel_context_pin(ce);
1274                 intel_context_put(ce);
1275                 if (ret)
1276                         goto out_shadow_ctx;
1277
1278                 s->shadow[i] = ce;
1279         }
1280
1281         bitmap_zero(s->shadow_ctx_desc_updated, I915_NUM_ENGINES);
1282
1283         s->workloads = kmem_cache_create_usercopy("gvt-g_vgpu_workload",
1284                                                   sizeof(struct intel_vgpu_workload), 0,
1285                                                   SLAB_HWCACHE_ALIGN,
1286                                                   offsetof(struct intel_vgpu_workload, rb_tail),
1287                                                   sizeof_field(struct intel_vgpu_workload, rb_tail),
1288                                                   NULL);
1289
1290         if (!s->workloads) {
1291                 ret = -ENOMEM;
1292                 goto out_shadow_ctx;
1293         }
1294
1295         atomic_set(&s->running_workload_num, 0);
1296         bitmap_zero(s->tlb_handle_pending, I915_NUM_ENGINES);
1297
1298         i915_gem_context_put(ctx);
1299         mutex_unlock(&i915->drm.struct_mutex);
1300         return 0;
1301
1302 out_shadow_ctx:
1303         i915_context_ppgtt_root_restore(s, i915_vm_to_ppgtt(ctx->vm));
1304         for_each_engine(engine, i915, i) {
1305                 if (IS_ERR(s->shadow[i]))
1306                         break;
1307
1308                 intel_context_unpin(s->shadow[i]);
1309                 intel_context_put(s->shadow[i]);
1310         }
1311         i915_gem_context_put(ctx);
1312 out_unlock:
1313         mutex_unlock(&i915->drm.struct_mutex);
1314         return ret;
1315 }
1316
1317 /**
1318  * intel_vgpu_select_submission_ops - select virtual submission interface
1319  * @vgpu: a vGPU
1320  * @engine_mask: either ALL_ENGINES or target engine mask
1321  * @interface: expected vGPU virtual submission interface
1322  *
1323  * This function is called when guest configures submission interface.
1324  *
1325  * Returns:
1326  * Zero on success, negative error code if failed.
1327  *
1328  */
1329 int intel_vgpu_select_submission_ops(struct intel_vgpu *vgpu,
1330                                      intel_engine_mask_t engine_mask,
1331                                      unsigned int interface)
1332 {
1333         struct intel_vgpu_submission *s = &vgpu->submission;
1334         const struct intel_vgpu_submission_ops *ops[] = {
1335                 [INTEL_VGPU_EXECLIST_SUBMISSION] =
1336                         &intel_vgpu_execlist_submission_ops,
1337         };
1338         int ret;
1339
1340         if (WARN_ON(interface >= ARRAY_SIZE(ops)))
1341                 return -EINVAL;
1342
1343         if (WARN_ON(interface == 0 && engine_mask != ALL_ENGINES))
1344                 return -EINVAL;
1345
1346         if (s->active)
1347                 s->ops->clean(vgpu, engine_mask);
1348
1349         if (interface == 0) {
1350                 s->ops = NULL;
1351                 s->virtual_submission_interface = 0;
1352                 s->active = false;
1353                 gvt_dbg_core("vgpu%d: remove submission ops\n", vgpu->id);
1354                 return 0;
1355         }
1356
1357         ret = ops[interface]->init(vgpu, engine_mask);
1358         if (ret)
1359                 return ret;
1360
1361         s->ops = ops[interface];
1362         s->virtual_submission_interface = interface;
1363         s->active = true;
1364
1365         gvt_dbg_core("vgpu%d: activate ops [ %s ]\n",
1366                         vgpu->id, s->ops->name);
1367
1368         return 0;
1369 }
1370
1371 /**
1372  * intel_vgpu_destroy_workload - destroy a vGPU workload
1373  * @workload: workload to destroy
1374  *
1375  * This function is called when destroy a vGPU workload.
1376  *
1377  */
1378 void intel_vgpu_destroy_workload(struct intel_vgpu_workload *workload)
1379 {
1380         struct intel_vgpu_submission *s = &workload->vgpu->submission;
1381
1382         release_shadow_batch_buffer(workload);
1383         release_shadow_wa_ctx(&workload->wa_ctx);
1384
1385         if (workload->shadow_mm)
1386                 intel_vgpu_mm_put(workload->shadow_mm);
1387
1388         kmem_cache_free(s->workloads, workload);
1389 }
1390
1391 static struct intel_vgpu_workload *
1392 alloc_workload(struct intel_vgpu *vgpu)
1393 {
1394         struct intel_vgpu_submission *s = &vgpu->submission;
1395         struct intel_vgpu_workload *workload;
1396
1397         workload = kmem_cache_zalloc(s->workloads, GFP_KERNEL);
1398         if (!workload)
1399                 return ERR_PTR(-ENOMEM);
1400
1401         INIT_LIST_HEAD(&workload->list);
1402         INIT_LIST_HEAD(&workload->shadow_bb);
1403
1404         init_waitqueue_head(&workload->shadow_ctx_status_wq);
1405         atomic_set(&workload->shadow_ctx_active, 0);
1406
1407         workload->status = -EINPROGRESS;
1408         workload->vgpu = vgpu;
1409
1410         return workload;
1411 }
1412
1413 #define RING_CTX_OFF(x) \
1414         offsetof(struct execlist_ring_context, x)
1415
1416 static void read_guest_pdps(struct intel_vgpu *vgpu,
1417                 u64 ring_context_gpa, u32 pdp[8])
1418 {
1419         u64 gpa;
1420         int i;
1421
1422         gpa = ring_context_gpa + RING_CTX_OFF(pdps[0].val);
1423
1424         for (i = 0; i < 8; i++)
1425                 intel_gvt_hypervisor_read_gpa(vgpu,
1426                                 gpa + i * 8, &pdp[7 - i], 4);
1427 }
1428
1429 static int prepare_mm(struct intel_vgpu_workload *workload)
1430 {
1431         struct execlist_ctx_descriptor_format *desc = &workload->ctx_desc;
1432         struct intel_vgpu_mm *mm;
1433         struct intel_vgpu *vgpu = workload->vgpu;
1434         enum intel_gvt_gtt_type root_entry_type;
1435         u64 pdps[GVT_RING_CTX_NR_PDPS];
1436
1437         switch (desc->addressing_mode) {
1438         case 1: /* legacy 32-bit */
1439                 root_entry_type = GTT_TYPE_PPGTT_ROOT_L3_ENTRY;
1440                 break;
1441         case 3: /* legacy 64-bit */
1442                 root_entry_type = GTT_TYPE_PPGTT_ROOT_L4_ENTRY;
1443                 break;
1444         default:
1445                 gvt_vgpu_err("Advanced Context mode(SVM) is not supported!\n");
1446                 return -EINVAL;
1447         }
1448
1449         read_guest_pdps(workload->vgpu, workload->ring_context_gpa, (void *)pdps);
1450
1451         mm = intel_vgpu_get_ppgtt_mm(workload->vgpu, root_entry_type, pdps);
1452         if (IS_ERR(mm))
1453                 return PTR_ERR(mm);
1454
1455         workload->shadow_mm = mm;
1456         return 0;
1457 }
1458
1459 #define same_context(a, b) (((a)->context_id == (b)->context_id) && \
1460                 ((a)->lrca == (b)->lrca))
1461
1462 /**
1463  * intel_vgpu_create_workload - create a vGPU workload
1464  * @vgpu: a vGPU
1465  * @ring_id: ring index
1466  * @desc: a guest context descriptor
1467  *
1468  * This function is called when creating a vGPU workload.
1469  *
1470  * Returns:
1471  * struct intel_vgpu_workload * on success, negative error code in
1472  * pointer if failed.
1473  *
1474  */
1475 struct intel_vgpu_workload *
1476 intel_vgpu_create_workload(struct intel_vgpu *vgpu, int ring_id,
1477                            struct execlist_ctx_descriptor_format *desc)
1478 {
1479         struct intel_vgpu_submission *s = &vgpu->submission;
1480         struct list_head *q = workload_q_head(vgpu, ring_id);
1481         struct intel_vgpu_workload *last_workload = NULL;
1482         struct intel_vgpu_workload *workload = NULL;
1483         struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
1484         u64 ring_context_gpa;
1485         u32 head, tail, start, ctl, ctx_ctl, per_ctx, indirect_ctx;
1486         u32 guest_head;
1487         int ret;
1488
1489         ring_context_gpa = intel_vgpu_gma_to_gpa(vgpu->gtt.ggtt_mm,
1490                         (u32)((desc->lrca + 1) << I915_GTT_PAGE_SHIFT));
1491         if (ring_context_gpa == INTEL_GVT_INVALID_ADDR) {
1492                 gvt_vgpu_err("invalid guest context LRCA: %x\n", desc->lrca);
1493                 return ERR_PTR(-EINVAL);
1494         }
1495
1496         intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
1497                         RING_CTX_OFF(ring_header.val), &head, 4);
1498
1499         intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
1500                         RING_CTX_OFF(ring_tail.val), &tail, 4);
1501
1502         guest_head = head;
1503
1504         head &= RB_HEAD_OFF_MASK;
1505         tail &= RB_TAIL_OFF_MASK;
1506
1507         list_for_each_entry_reverse(last_workload, q, list) {
1508
1509                 if (same_context(&last_workload->ctx_desc, desc)) {
1510                         gvt_dbg_el("ring id %d cur workload == last\n",
1511                                         ring_id);
1512                         gvt_dbg_el("ctx head %x real head %lx\n", head,
1513                                         last_workload->rb_tail);
1514                         /*
1515                          * cannot use guest context head pointer here,
1516                          * as it might not be updated at this time
1517                          */
1518                         head = last_workload->rb_tail;
1519                         break;
1520                 }
1521         }
1522
1523         gvt_dbg_el("ring id %d begin a new workload\n", ring_id);
1524
1525         /* record some ring buffer register values for scan and shadow */
1526         intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
1527                         RING_CTX_OFF(rb_start.val), &start, 4);
1528         intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
1529                         RING_CTX_OFF(rb_ctrl.val), &ctl, 4);
1530         intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
1531                         RING_CTX_OFF(ctx_ctrl.val), &ctx_ctl, 4);
1532
1533         if (!intel_gvt_ggtt_validate_range(vgpu, start,
1534                                 _RING_CTL_BUF_SIZE(ctl))) {
1535                 gvt_vgpu_err("context contain invalid rb at: 0x%x\n", start);
1536                 return ERR_PTR(-EINVAL);
1537         }
1538
1539         workload = alloc_workload(vgpu);
1540         if (IS_ERR(workload))
1541                 return workload;
1542
1543         workload->ring_id = ring_id;
1544         workload->ctx_desc = *desc;
1545         workload->ring_context_gpa = ring_context_gpa;
1546         workload->rb_head = head;
1547         workload->guest_rb_head = guest_head;
1548         workload->rb_tail = tail;
1549         workload->rb_start = start;
1550         workload->rb_ctl = ctl;
1551
1552         if (ring_id == RCS0) {
1553                 intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
1554                         RING_CTX_OFF(bb_per_ctx_ptr.val), &per_ctx, 4);
1555                 intel_gvt_hypervisor_read_gpa(vgpu, ring_context_gpa +
1556                         RING_CTX_OFF(rcs_indirect_ctx.val), &indirect_ctx, 4);
1557
1558                 workload->wa_ctx.indirect_ctx.guest_gma =
1559                         indirect_ctx & INDIRECT_CTX_ADDR_MASK;
1560                 workload->wa_ctx.indirect_ctx.size =
1561                         (indirect_ctx & INDIRECT_CTX_SIZE_MASK) *
1562                         CACHELINE_BYTES;
1563
1564                 if (workload->wa_ctx.indirect_ctx.size != 0) {
1565                         if (!intel_gvt_ggtt_validate_range(vgpu,
1566                                 workload->wa_ctx.indirect_ctx.guest_gma,
1567                                 workload->wa_ctx.indirect_ctx.size)) {
1568                                 gvt_vgpu_err("invalid wa_ctx at: 0x%lx\n",
1569                                     workload->wa_ctx.indirect_ctx.guest_gma);
1570                                 kmem_cache_free(s->workloads, workload);
1571                                 return ERR_PTR(-EINVAL);
1572                         }
1573                 }
1574
1575                 workload->wa_ctx.per_ctx.guest_gma =
1576                         per_ctx & PER_CTX_ADDR_MASK;
1577                 workload->wa_ctx.per_ctx.valid = per_ctx & 1;
1578                 if (workload->wa_ctx.per_ctx.valid) {
1579                         if (!intel_gvt_ggtt_validate_range(vgpu,
1580                                 workload->wa_ctx.per_ctx.guest_gma,
1581                                 CACHELINE_BYTES)) {
1582                                 gvt_vgpu_err("invalid per_ctx at: 0x%lx\n",
1583                                         workload->wa_ctx.per_ctx.guest_gma);
1584                                 kmem_cache_free(s->workloads, workload);
1585                                 return ERR_PTR(-EINVAL);
1586                         }
1587                 }
1588         }
1589
1590         gvt_dbg_el("workload %p ring id %d head %x tail %x start %x ctl %x\n",
1591                         workload, ring_id, head, tail, start, ctl);
1592
1593         ret = prepare_mm(workload);
1594         if (ret) {
1595                 kmem_cache_free(s->workloads, workload);
1596                 return ERR_PTR(ret);
1597         }
1598
1599         /* Only scan and shadow the first workload in the queue
1600          * as there is only one pre-allocated buf-obj for shadow.
1601          */
1602         if (list_empty(workload_q_head(vgpu, ring_id))) {
1603                 intel_runtime_pm_get(&dev_priv->runtime_pm);
1604                 mutex_lock(&dev_priv->drm.struct_mutex);
1605                 ret = intel_gvt_scan_and_shadow_workload(workload);
1606                 mutex_unlock(&dev_priv->drm.struct_mutex);
1607                 intel_runtime_pm_put_unchecked(&dev_priv->runtime_pm);
1608         }
1609
1610         if (ret) {
1611                 if (vgpu_is_vm_unhealthy(ret))
1612                         enter_failsafe_mode(vgpu, GVT_FAILSAFE_GUEST_ERR);
1613                 intel_vgpu_destroy_workload(workload);
1614                 return ERR_PTR(ret);
1615         }
1616
1617         return workload;
1618 }
1619
1620 /**
1621  * intel_vgpu_queue_workload - Qeue a vGPU workload
1622  * @workload: the workload to queue in
1623  */
1624 void intel_vgpu_queue_workload(struct intel_vgpu_workload *workload)
1625 {
1626         list_add_tail(&workload->list,
1627                 workload_q_head(workload->vgpu, workload->ring_id));
1628         intel_gvt_kick_schedule(workload->vgpu->gvt);
1629         wake_up(&workload->vgpu->gvt->scheduler.waitq[workload->ring_id]);
1630 }