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

   1 /*
   2  * GTT virtualization
   3  *
   4  * Copyright(c) 2011-2016 Intel Corporation. All rights reserved.
   5  *
   6  * Permission is hereby granted, free of charge, to any person obtaining a
   7  * copy of this software and associated documentation files (the "Software"),
   8  * to deal in the Software without restriction, including without limitation
   9  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  10  * and/or sell copies of the Software, and to permit persons to whom the
  11  * Software is furnished to do so, subject to the following conditions:
  12  *
  13  * The above copyright notice and this permission notice (including the next
  14  * paragraph) shall be included in all copies or substantial portions of the
  15  * Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  23  * SOFTWARE.
  24  *
  25  * Authors:
  26  *    Zhi Wang <[email protected]>
  27  *    Zhenyu Wang <[email protected]>
  28  *    Xiao Zheng <[email protected]>
  29  *
  30  * Contributors:
  31  *    Min He <[email protected]>
  32  *    Bing Niu <[email protected]>
  33  *
  34  */
  35
  36 #include "i915_drv.h"
  37 #include "gvt.h"
  38 #include "i915_pvinfo.h"
  39 #include "trace.h"
  40
  41 #if defined(VERBOSE_DEBUG)
  42 #define gvt_vdbg_mm(fmt, args...) gvt_dbg_mm(fmt, ##args)
  43 #else
  44 #define gvt_vdbg_mm(fmt, args...)
  45 #endif
  46
  47 static bool enable_out_of_sync = false;
  48 static int preallocated_oos_pages = 8192;
  49
  50 /*
  51  * validate a gm address and related range size,
  52  * translate it to host gm address
  53  */
  54 bool intel_gvt_ggtt_validate_range(struct intel_vgpu *vgpu, u64 addr, u32 size)
  55 {
  56         if ((!vgpu_gmadr_is_valid(vgpu, addr)) || (size
  57                         && !vgpu_gmadr_is_valid(vgpu, addr + size - 1))) {
  58                 gvt_vgpu_err("invalid range gmadr 0x%llx size 0x%x\n",
  59                                 addr, size);
  60                 return false;
  61         }
  62         return true;
  63 }
  64
  65 /* translate a guest gmadr to host gmadr */
  66 int intel_gvt_ggtt_gmadr_g2h(struct intel_vgpu *vgpu, u64 g_addr, u64 *h_addr)
  67 {
  68         if (WARN(!vgpu_gmadr_is_valid(vgpu, g_addr),
  69                  "invalid guest gmadr %llx\n", g_addr))
  70                 return -EACCES;
  71
  72         if (vgpu_gmadr_is_aperture(vgpu, g_addr))
  73                 *h_addr = vgpu_aperture_gmadr_base(vgpu)
  74                           + (g_addr - vgpu_aperture_offset(vgpu));
  75         else
  76                 *h_addr = vgpu_hidden_gmadr_base(vgpu)
  77                           + (g_addr - vgpu_hidden_offset(vgpu));
  78         return 0;
  79 }
  80
  81 /* translate a host gmadr to guest gmadr */
  82 int intel_gvt_ggtt_gmadr_h2g(struct intel_vgpu *vgpu, u64 h_addr, u64 *g_addr)
  83 {
  84         if (WARN(!gvt_gmadr_is_valid(vgpu->gvt, h_addr),
  85                  "invalid host gmadr %llx\n", h_addr))
  86                 return -EACCES;
  87
  88         if (gvt_gmadr_is_aperture(vgpu->gvt, h_addr))
  89                 *g_addr = vgpu_aperture_gmadr_base(vgpu)
  90                         + (h_addr - gvt_aperture_gmadr_base(vgpu->gvt));
  91         else
  92                 *g_addr = vgpu_hidden_gmadr_base(vgpu)
  93                         + (h_addr - gvt_hidden_gmadr_base(vgpu->gvt));
  94         return 0;
  95 }
  96
  97 int intel_gvt_ggtt_index_g2h(struct intel_vgpu *vgpu, unsigned long g_index,
  98                              unsigned long *h_index)
  99 {
 100         u64 h_addr;
 101         int ret;
 102
 103         ret = intel_gvt_ggtt_gmadr_g2h(vgpu, g_index << I915_GTT_PAGE_SHIFT,
 104                                        &h_addr);
 105         if (ret)
 106                 return ret;
 107
 108         *h_index = h_addr >> I915_GTT_PAGE_SHIFT;
 109         return 0;
 110 }
 111
 112 int intel_gvt_ggtt_h2g_index(struct intel_vgpu *vgpu, unsigned long h_index,
 113                              unsigned long *g_index)
 114 {
 115         u64 g_addr;
 116         int ret;
 117
 118         ret = intel_gvt_ggtt_gmadr_h2g(vgpu, h_index << I915_GTT_PAGE_SHIFT,
 119                                        &g_addr);
 120         if (ret)
 121                 return ret;
 122
 123         *g_index = g_addr >> I915_GTT_PAGE_SHIFT;
 124         return 0;
 125 }
 126
 127 #define gtt_type_is_entry(type) \
 128         (type > GTT_TYPE_INVALID && type < GTT_TYPE_PPGTT_ENTRY \
 129          && type != GTT_TYPE_PPGTT_PTE_ENTRY \
 130          && type != GTT_TYPE_PPGTT_ROOT_ENTRY)
 131
 132 #define gtt_type_is_pt(type) \
 133         (type >= GTT_TYPE_PPGTT_PTE_PT && type < GTT_TYPE_MAX)
 134
 135 #define gtt_type_is_pte_pt(type) \
 136         (type == GTT_TYPE_PPGTT_PTE_PT)
 137
 138 #define gtt_type_is_root_pointer(type) \
 139         (gtt_type_is_entry(type) && type > GTT_TYPE_PPGTT_ROOT_ENTRY)
 140
 141 #define gtt_init_entry(e, t, p, v) do { \
 142         (e)->type = t; \
 143         (e)->pdev = p; \
 144         memcpy(&(e)->val64, &v, sizeof(v)); \
 145 } while (0)
 146
 147 /*
 148  * Mappings between GTT_TYPE* enumerations.
 149  * Following information can be found according to the given type:
 150  * - type of next level page table
 151  * - type of entry inside this level page table
 152  * - type of entry with PSE set
 153  *
 154  * If the given type doesn't have such a kind of information,
 155  * e.g. give a l4 root entry type, then request to get its PSE type,
 156  * give a PTE page table type, then request to get its next level page
 157  * table type, as we know l4 root entry doesn't have a PSE bit,
 158  * and a PTE page table doesn't have a next level page table type,
 159  * GTT_TYPE_INVALID will be returned. This is useful when traversing a
 160  * page table.
 161  */
 162
 163 struct gtt_type_table_entry {
 164         int entry_type;
 165         int pt_type;
 166         int next_pt_type;
 167         int pse_entry_type;
 168 };
 169
 170 #define GTT_TYPE_TABLE_ENTRY(type, e_type, cpt_type, npt_type, pse_type) \
 171         [type] = { \
 172                 .entry_type = e_type, \
 173                 .pt_type = cpt_type, \
 174                 .next_pt_type = npt_type, \
 175                 .pse_entry_type = pse_type, \
 176         }
 177
 178 static struct gtt_type_table_entry gtt_type_table[] = {
 179         GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_ROOT_L4_ENTRY,
 180                         GTT_TYPE_PPGTT_ROOT_L4_ENTRY,
 181                         GTT_TYPE_INVALID,
 182                         GTT_TYPE_PPGTT_PML4_PT,
 183                         GTT_TYPE_INVALID),
 184         GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PML4_PT,
 185                         GTT_TYPE_PPGTT_PML4_ENTRY,
 186                         GTT_TYPE_PPGTT_PML4_PT,
 187                         GTT_TYPE_PPGTT_PDP_PT,
 188                         GTT_TYPE_INVALID),
 189         GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PML4_ENTRY,
 190                         GTT_TYPE_PPGTT_PML4_ENTRY,
 191                         GTT_TYPE_PPGTT_PML4_PT,
 192                         GTT_TYPE_PPGTT_PDP_PT,
 193                         GTT_TYPE_INVALID),
 194         GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PDP_PT,
 195                         GTT_TYPE_PPGTT_PDP_ENTRY,
 196                         GTT_TYPE_PPGTT_PDP_PT,
 197                         GTT_TYPE_PPGTT_PDE_PT,
 198                         GTT_TYPE_PPGTT_PTE_1G_ENTRY),
 199         GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_ROOT_L3_ENTRY,
 200                         GTT_TYPE_PPGTT_ROOT_L3_ENTRY,
 201                         GTT_TYPE_INVALID,
 202                         GTT_TYPE_PPGTT_PDE_PT,
 203                         GTT_TYPE_PPGTT_PTE_1G_ENTRY),
 204         GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PDP_ENTRY,
 205                         GTT_TYPE_PPGTT_PDP_ENTRY,
 206                         GTT_TYPE_PPGTT_PDP_PT,
 207                         GTT_TYPE_PPGTT_PDE_PT,
 208                         GTT_TYPE_PPGTT_PTE_1G_ENTRY),
 209         GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PDE_PT,
 210                         GTT_TYPE_PPGTT_PDE_ENTRY,
 211                         GTT_TYPE_PPGTT_PDE_PT,
 212                         GTT_TYPE_PPGTT_PTE_PT,
 213                         GTT_TYPE_PPGTT_PTE_2M_ENTRY),
 214         GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PDE_ENTRY,
 215                         GTT_TYPE_PPGTT_PDE_ENTRY,
 216                         GTT_TYPE_PPGTT_PDE_PT,
 217                         GTT_TYPE_PPGTT_PTE_PT,
 218                         GTT_TYPE_PPGTT_PTE_2M_ENTRY),
 219         /* We take IPS bit as 'PSE' for PTE level. */
 220         GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PTE_PT,
 221                         GTT_TYPE_PPGTT_PTE_4K_ENTRY,
 222                         GTT_TYPE_PPGTT_PTE_PT,
 223                         GTT_TYPE_INVALID,
 224                         GTT_TYPE_PPGTT_PTE_64K_ENTRY),
 225         GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PTE_4K_ENTRY,
 226                         GTT_TYPE_PPGTT_PTE_4K_ENTRY,
 227                         GTT_TYPE_PPGTT_PTE_PT,
 228                         GTT_TYPE_INVALID,
 229                         GTT_TYPE_PPGTT_PTE_64K_ENTRY),
 230         GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PTE_64K_ENTRY,
 231                         GTT_TYPE_PPGTT_PTE_4K_ENTRY,
 232                         GTT_TYPE_PPGTT_PTE_PT,
 233                         GTT_TYPE_INVALID,
 234                         GTT_TYPE_PPGTT_PTE_64K_ENTRY),
 235         GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PTE_2M_ENTRY,
 236                         GTT_TYPE_PPGTT_PDE_ENTRY,
 237                         GTT_TYPE_PPGTT_PDE_PT,
 238                         GTT_TYPE_INVALID,
 239                         GTT_TYPE_PPGTT_PTE_2M_ENTRY),
 240         GTT_TYPE_TABLE_ENTRY(GTT_TYPE_PPGTT_PTE_1G_ENTRY,
 241                         GTT_TYPE_PPGTT_PDP_ENTRY,
 242                         GTT_TYPE_PPGTT_PDP_PT,
 243                         GTT_TYPE_INVALID,
 244                         GTT_TYPE_PPGTT_PTE_1G_ENTRY),
 245         GTT_TYPE_TABLE_ENTRY(GTT_TYPE_GGTT_PTE,
 246                         GTT_TYPE_GGTT_PTE,
 247                         GTT_TYPE_INVALID,
 248                         GTT_TYPE_INVALID,
 249                         GTT_TYPE_INVALID),
 250 };
 251
 252 static inline int get_next_pt_type(int type)
 253 {
 254         return gtt_type_table[type].next_pt_type;
 255 }
 256
 257 static inline int get_pt_type(int type)
 258 {
 259         return gtt_type_table[type].pt_type;
 260 }
 261
 262 static inline int get_entry_type(int type)
 263 {
 264         return gtt_type_table[type].entry_type;
 265 }
 266
 267 static inline int get_pse_type(int type)
 268 {
 269         return gtt_type_table[type].pse_entry_type;
 270 }
 271
 272 static u64 read_pte64(struct drm_i915_private *dev_priv, unsigned long index)
 273 {
 274         void __iomem *addr = (gen8_pte_t __iomem *)dev_priv->ggtt.gsm + index;
 275
 276         return readq(addr);
 277 }
 278
 279 static void ggtt_invalidate(struct drm_i915_private *dev_priv)
 280 {
 281         mmio_hw_access_pre(dev_priv);
 282         I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
 283         mmio_hw_access_post(dev_priv);
 284 }
 285
 286 static void write_pte64(struct drm_i915_private *dev_priv,
 287                 unsigned long index, u64 pte)
 288 {
 289         void __iomem *addr = (gen8_pte_t __iomem *)dev_priv->ggtt.gsm + index;
 290
 291         writeq(pte, addr);
 292 }
 293
 294 static inline int gtt_get_entry64(void *pt,
 295                 struct intel_gvt_gtt_entry *e,
 296                 unsigned long index, bool hypervisor_access, unsigned long gpa,
 297                 struct intel_vgpu *vgpu)
 298 {
 299         const struct intel_gvt_device_info *info = &vgpu->gvt->device_info;
 300         int ret;
 301
 302         if (WARN_ON(info->gtt_entry_size != 8))
 303                 return -EINVAL;
 304
 305         if (hypervisor_access) {
 306                 ret = intel_gvt_hypervisor_read_gpa(vgpu, gpa +
 307                                 (index << info->gtt_entry_size_shift),
 308                                 &e->val64, 8);
 309                 if (WARN_ON(ret))
 310                         return ret;
 311         } else if (!pt) {
 312                 e->val64 = read_pte64(vgpu->gvt->dev_priv, index);
 313         } else {
 314                 e->val64 = *((u64 *)pt + index);
 315         }
 316         return 0;
 317 }
 318
 319 static inline int gtt_set_entry64(void *pt,
 320                 struct intel_gvt_gtt_entry *e,
 321                 unsigned long index, bool hypervisor_access, unsigned long gpa,
 322                 struct intel_vgpu *vgpu)
 323 {
 324         const struct intel_gvt_device_info *info = &vgpu->gvt->device_info;
 325         int ret;
 326
 327         if (WARN_ON(info->gtt_entry_size != 8))
 328                 return -EINVAL;
 329
 330         if (hypervisor_access) {
 331                 ret = intel_gvt_hypervisor_write_gpa(vgpu, gpa +
 332                                 (index << info->gtt_entry_size_shift),
 333                                 &e->val64, 8);
 334                 if (WARN_ON(ret))
 335                         return ret;
 336         } else if (!pt) {
 337                 write_pte64(vgpu->gvt->dev_priv, index, e->val64);
 338         } else {
 339                 *((u64 *)pt + index) = e->val64;
 340         }
 341         return 0;
 342 }
 343
 344 #define GTT_HAW 46
 345
 346 #define ADDR_1G_MASK    GENMASK_ULL(GTT_HAW - 1, 30)
 347 #define ADDR_2M_MASK    GENMASK_ULL(GTT_HAW - 1, 21)
 348 #define ADDR_64K_MASK   GENMASK_ULL(GTT_HAW - 1, 16)
 349 #define ADDR_4K_MASK    GENMASK_ULL(GTT_HAW - 1, 12)
 350
 351 #define GTT_SPTE_FLAG_MASK GENMASK_ULL(62, 52)
 352 #define GTT_SPTE_FLAG_64K_SPLITED BIT(52) /* splited 64K gtt entry */
 353
 354 #define GTT_64K_PTE_STRIDE 16
 355
 356 static unsigned long gen8_gtt_get_pfn(struct intel_gvt_gtt_entry *e)
 357 {
 358         unsigned long pfn;
 359
 360         if (e->type == GTT_TYPE_PPGTT_PTE_1G_ENTRY)
 361                 pfn = (e->val64 & ADDR_1G_MASK) >> PAGE_SHIFT;
 362         else if (e->type == GTT_TYPE_PPGTT_PTE_2M_ENTRY)
 363                 pfn = (e->val64 & ADDR_2M_MASK) >> PAGE_SHIFT;
 364         else if (e->type == GTT_TYPE_PPGTT_PTE_64K_ENTRY)
 365                 pfn = (e->val64 & ADDR_64K_MASK) >> PAGE_SHIFT;
 366         else
 367                 pfn = (e->val64 & ADDR_4K_MASK) >> PAGE_SHIFT;
 368         return pfn;
 369 }
 370
 371 static void gen8_gtt_set_pfn(struct intel_gvt_gtt_entry *e, unsigned long pfn)
 372 {
 373         if (e->type == GTT_TYPE_PPGTT_PTE_1G_ENTRY) {
 374                 e->val64 &= ~ADDR_1G_MASK;
 375                 pfn &= (ADDR_1G_MASK >> PAGE_SHIFT);
 376         } else if (e->type == GTT_TYPE_PPGTT_PTE_2M_ENTRY) {
 377                 e->val64 &= ~ADDR_2M_MASK;
 378                 pfn &= (ADDR_2M_MASK >> PAGE_SHIFT);
 379         } else if (e->type == GTT_TYPE_PPGTT_PTE_64K_ENTRY) {
 380                 e->val64 &= ~ADDR_64K_MASK;
 381                 pfn &= (ADDR_64K_MASK >> PAGE_SHIFT);
 382         } else {
 383                 e->val64 &= ~ADDR_4K_MASK;
 384                 pfn &= (ADDR_4K_MASK >> PAGE_SHIFT);
 385         }
 386
 387         e->val64 |= (pfn << PAGE_SHIFT);
 388 }
 389
 390 static bool gen8_gtt_test_pse(struct intel_gvt_gtt_entry *e)
 391 {
 392         return !!(e->val64 & _PAGE_PSE);
 393 }
 394
 395 static void gen8_gtt_clear_pse(struct intel_gvt_gtt_entry *e)
 396 {
 397         if (gen8_gtt_test_pse(e)) {
 398                 switch (e->type) {
 399                 case GTT_TYPE_PPGTT_PTE_2M_ENTRY:
 400                         e->val64 &= ~_PAGE_PSE;
 401                         e->type = GTT_TYPE_PPGTT_PDE_ENTRY;
 402                         break;
 403                 case GTT_TYPE_PPGTT_PTE_1G_ENTRY:
 404                         e->type = GTT_TYPE_PPGTT_PDP_ENTRY;
 405                         e->val64 &= ~_PAGE_PSE;
 406                         break;
 407                 default:
 408                         WARN_ON(1);
 409                 }
 410         }
 411 }
 412
 413 static bool gen8_gtt_test_ips(struct intel_gvt_gtt_entry *e)
 414 {
 415         if (GEM_WARN_ON(e->type != GTT_TYPE_PPGTT_PDE_ENTRY))
 416                 return false;
 417
 418         return !!(e->val64 & GEN8_PDE_IPS_64K);
 419 }
 420
 421 static void gen8_gtt_clear_ips(struct intel_gvt_gtt_entry *e)
 422 {
 423         if (GEM_WARN_ON(e->type != GTT_TYPE_PPGTT_PDE_ENTRY))
 424                 return;
 425
 426         e->val64 &= ~GEN8_PDE_IPS_64K;
 427 }
 428
 429 static bool gen8_gtt_test_present(struct intel_gvt_gtt_entry *e)
 430 {
 431         /*
 432          * i915 writes PDP root pointer registers without present bit,
 433          * it also works, so we need to treat root pointer entry
 434          * specifically.
 435          */
 436         if (e->type == GTT_TYPE_PPGTT_ROOT_L3_ENTRY
 437                         || e->type == GTT_TYPE_PPGTT_ROOT_L4_ENTRY)
 438                 return (e->val64 != 0);
 439         else
 440                 return (e->val64 & _PAGE_PRESENT);
 441 }
 442
 443 static void gtt_entry_clear_present(struct intel_gvt_gtt_entry *e)
 444 {
 445         e->val64 &= ~_PAGE_PRESENT;
 446 }
 447
 448 static void gtt_entry_set_present(struct intel_gvt_gtt_entry *e)
 449 {
 450         e->val64 |= _PAGE_PRESENT;
 451 }
 452
 453 static bool gen8_gtt_test_64k_splited(struct intel_gvt_gtt_entry *e)
 454 {
 455         return !!(e->val64 & GTT_SPTE_FLAG_64K_SPLITED);
 456 }
 457
 458 static void gen8_gtt_set_64k_splited(struct intel_gvt_gtt_entry *e)
 459 {
 460         e->val64 |= GTT_SPTE_FLAG_64K_SPLITED;
 461 }
 462
 463 static void gen8_gtt_clear_64k_splited(struct intel_gvt_gtt_entry *e)
 464 {
 465         e->val64 &= ~GTT_SPTE_FLAG_64K_SPLITED;
 466 }
 467
 468 /*
 469  * Per-platform GMA routines.
 470  */
 471 static unsigned long gma_to_ggtt_pte_index(unsigned long gma)
 472 {
 473         unsigned long x = (gma >> I915_GTT_PAGE_SHIFT);
 474
 475         trace_gma_index(__func__, gma, x);
 476         return x;
 477 }
 478
 479 #define DEFINE_PPGTT_GMA_TO_INDEX(prefix, ename, exp) \
 480 static unsigned long prefix##_gma_to_##ename##_index(unsigned long gma) \
 481 { \
 482         unsigned long x = (exp); \
 483         trace_gma_index(__func__, gma, x); \
 484         return x; \
 485 }
 486
 487 DEFINE_PPGTT_GMA_TO_INDEX(gen8, pte, (gma >> 12 & 0x1ff));
 488 DEFINE_PPGTT_GMA_TO_INDEX(gen8, pde, (gma >> 21 & 0x1ff));
 489 DEFINE_PPGTT_GMA_TO_INDEX(gen8, l3_pdp, (gma >> 30 & 0x3));
 490 DEFINE_PPGTT_GMA_TO_INDEX(gen8, l4_pdp, (gma >> 30 & 0x1ff));
 491 DEFINE_PPGTT_GMA_TO_INDEX(gen8, pml4, (gma >> 39 & 0x1ff));
 492
 493 static struct intel_gvt_gtt_pte_ops gen8_gtt_pte_ops = {
 494         .get_entry = gtt_get_entry64,
 495         .set_entry = gtt_set_entry64,
 496         .clear_present = gtt_entry_clear_present,
 497         .set_present = gtt_entry_set_present,
 498         .test_present = gen8_gtt_test_present,
 499         .test_pse = gen8_gtt_test_pse,
 500         .clear_pse = gen8_gtt_clear_pse,
 501         .clear_ips = gen8_gtt_clear_ips,
 502         .test_ips = gen8_gtt_test_ips,
 503         .clear_64k_splited = gen8_gtt_clear_64k_splited,
 504         .set_64k_splited = gen8_gtt_set_64k_splited,
 505         .test_64k_splited = gen8_gtt_test_64k_splited,
 506         .get_pfn = gen8_gtt_get_pfn,
 507         .set_pfn = gen8_gtt_set_pfn,
 508 };
 509
 510 static struct intel_gvt_gtt_gma_ops gen8_gtt_gma_ops = {
 511         .gma_to_ggtt_pte_index = gma_to_ggtt_pte_index,
 512         .gma_to_pte_index = gen8_gma_to_pte_index,
 513         .gma_to_pde_index = gen8_gma_to_pde_index,
 514         .gma_to_l3_pdp_index = gen8_gma_to_l3_pdp_index,
 515         .gma_to_l4_pdp_index = gen8_gma_to_l4_pdp_index,
 516         .gma_to_pml4_index = gen8_gma_to_pml4_index,
 517 };
 518
 519 /* Update entry type per pse and ips bit. */
 520 static void update_entry_type_for_real(struct intel_gvt_gtt_pte_ops *pte_ops,
 521         struct intel_gvt_gtt_entry *entry, bool ips)
 522 {
 523         switch (entry->type) {
 524         case GTT_TYPE_PPGTT_PDE_ENTRY:
 525         case GTT_TYPE_PPGTT_PDP_ENTRY:
 526                 if (pte_ops->test_pse(entry))
 527                         entry->type = get_pse_type(entry->type);
 528                 break;
 529         case GTT_TYPE_PPGTT_PTE_4K_ENTRY:
 530                 if (ips)
 531                         entry->type = get_pse_type(entry->type);
 532                 break;
 533         default:
 534                 GEM_BUG_ON(!gtt_type_is_entry(entry->type));
 535         }
 536
 537         GEM_BUG_ON(entry->type == GTT_TYPE_INVALID);
 538 }
 539
 540 /*
 541  * MM helpers.
 542  */
 543 static void _ppgtt_get_root_entry(struct intel_vgpu_mm *mm,
 544                 struct intel_gvt_gtt_entry *entry, unsigned long index,
 545                 bool guest)
 546 {
 547         struct intel_gvt_gtt_pte_ops *pte_ops = mm->vgpu->gvt->gtt.pte_ops;
 548
 549         GEM_BUG_ON(mm->type != INTEL_GVT_MM_PPGTT);
 550
 551         entry->type = mm->ppgtt_mm.root_entry_type;
 552         pte_ops->get_entry(guest ? mm->ppgtt_mm.guest_pdps :
 553                            mm->ppgtt_mm.shadow_pdps,
 554                            entry, index, false, 0, mm->vgpu);
 555         update_entry_type_for_real(pte_ops, entry, false);
 556 }
 557
 558 static inline void ppgtt_get_guest_root_entry(struct intel_vgpu_mm *mm,
 559                 struct intel_gvt_gtt_entry *entry, unsigned long index)
 560 {
 561         _ppgtt_get_root_entry(mm, entry, index, true);
 562 }
 563
 564 static inline void ppgtt_get_shadow_root_entry(struct intel_vgpu_mm *mm,
 565                 struct intel_gvt_gtt_entry *entry, unsigned long index)
 566 {
 567         _ppgtt_get_root_entry(mm, entry, index, false);
 568 }
 569
 570 static void _ppgtt_set_root_entry(struct intel_vgpu_mm *mm,
 571                 struct intel_gvt_gtt_entry *entry, unsigned long index,
 572                 bool guest)
 573 {
 574         struct intel_gvt_gtt_pte_ops *pte_ops = mm->vgpu->gvt->gtt.pte_ops;
 575
 576         pte_ops->set_entry(guest ? mm->ppgtt_mm.guest_pdps :
 577                            mm->ppgtt_mm.shadow_pdps,
 578                            entry, index, false, 0, mm->vgpu);
 579 }
 580
 581 static inline void ppgtt_set_guest_root_entry(struct intel_vgpu_mm *mm,
 582                 struct intel_gvt_gtt_entry *entry, unsigned long index)
 583 {
 584         _ppgtt_set_root_entry(mm, entry, index, true);
 585 }
 586
 587 static inline void ppgtt_set_shadow_root_entry(struct intel_vgpu_mm *mm,
 588                 struct intel_gvt_gtt_entry *entry, unsigned long index)
 589 {
 590         _ppgtt_set_root_entry(mm, entry, index, false);
 591 }
 592
 593 static void ggtt_get_guest_entry(struct intel_vgpu_mm *mm,
 594                 struct intel_gvt_gtt_entry *entry, unsigned long index)
 595 {
 596         struct intel_gvt_gtt_pte_ops *pte_ops = mm->vgpu->gvt->gtt.pte_ops;
 597
 598         GEM_BUG_ON(mm->type != INTEL_GVT_MM_GGTT);
 599
 600         entry->type = GTT_TYPE_GGTT_PTE;
 601         pte_ops->get_entry(mm->ggtt_mm.virtual_ggtt, entry, index,
 602                            false, 0, mm->vgpu);
 603 }
 604
 605 static void ggtt_set_guest_entry(struct intel_vgpu_mm *mm,
 606                 struct intel_gvt_gtt_entry *entry, unsigned long index)
 607 {
 608         struct intel_gvt_gtt_pte_ops *pte_ops = mm->vgpu->gvt->gtt.pte_ops;
 609
 610         GEM_BUG_ON(mm->type != INTEL_GVT_MM_GGTT);
 611
 612         pte_ops->set_entry(mm->ggtt_mm.virtual_ggtt, entry, index,
 613                            false, 0, mm->vgpu);
 614 }
 615
 616 static void ggtt_get_host_entry(struct intel_vgpu_mm *mm,
 617                 struct intel_gvt_gtt_entry *entry, unsigned long index)
 618 {
 619         struct intel_gvt_gtt_pte_ops *pte_ops = mm->vgpu->gvt->gtt.pte_ops;
 620
 621         GEM_BUG_ON(mm->type != INTEL_GVT_MM_GGTT);
 622
 623         pte_ops->get_entry(NULL, entry, index, false, 0, mm->vgpu);
 624 }
 625
 626 static void ggtt_set_host_entry(struct intel_vgpu_mm *mm,
 627                 struct intel_gvt_gtt_entry *entry, unsigned long index)
 628 {
 629         struct intel_gvt_gtt_pte_ops *pte_ops = mm->vgpu->gvt->gtt.pte_ops;
 630
 631         GEM_BUG_ON(mm->type != INTEL_GVT_MM_GGTT);
 632
 633         pte_ops->set_entry(NULL, entry, index, false, 0, mm->vgpu);
 634 }
 635
 636 /*
 637  * PPGTT shadow page table helpers.
 638  */
 639 static inline int ppgtt_spt_get_entry(
 640                 struct intel_vgpu_ppgtt_spt *spt,
 641                 void *page_table, int type,
 642                 struct intel_gvt_gtt_entry *e, unsigned long index,
 643                 bool guest)
 644 {
 645         struct intel_gvt *gvt = spt->vgpu->gvt;
 646         struct intel_gvt_gtt_pte_ops *ops = gvt->gtt.pte_ops;
 647         int ret;
 648
 649         e->type = get_entry_type(type);
 650
 651         if (WARN(!gtt_type_is_entry(e->type), "invalid entry type\n"))
 652                 return -EINVAL;
 653
 654         ret = ops->get_entry(page_table, e, index, guest,
 655                         spt->guest_page.gfn << I915_GTT_PAGE_SHIFT,
 656                         spt->vgpu);
 657         if (ret)
 658                 return ret;
 659
 660         update_entry_type_for_real(ops, e, guest ?
 661                                    spt->guest_page.pde_ips : false);
 662
 663         gvt_vdbg_mm("read ppgtt entry, spt type %d, entry type %d, index %lu, value %llx\n",
 664                     type, e->type, index, e->val64);
 665         return 0;
 666 }
 667
 668 static inline int ppgtt_spt_set_entry(
 669                 struct intel_vgpu_ppgtt_spt *spt,
 670                 void *page_table, int type,
 671                 struct intel_gvt_gtt_entry *e, unsigned long index,
 672                 bool guest)
 673 {
 674         struct intel_gvt *gvt = spt->vgpu->gvt;
 675         struct intel_gvt_gtt_pte_ops *ops = gvt->gtt.pte_ops;
 676
 677         if (WARN(!gtt_type_is_entry(e->type), "invalid entry type\n"))
 678                 return -EINVAL;
 679
 680         gvt_vdbg_mm("set ppgtt entry, spt type %d, entry type %d, index %lu, value %llx\n",
 681                     type, e->type, index, e->val64);
 682
 683         return ops->set_entry(page_table, e, index, guest,
 684                         spt->guest_page.gfn << I915_GTT_PAGE_SHIFT,
 685                         spt->vgpu);
 686 }
 687
 688 #define ppgtt_get_guest_entry(spt, e, index) \
 689         ppgtt_spt_get_entry(spt, NULL, \
 690                 spt->guest_page.type, e, index, true)
 691
 692 #define ppgtt_set_guest_entry(spt, e, index) \
 693         ppgtt_spt_set_entry(spt, NULL, \
 694                 spt->guest_page.type, e, index, true)
 695
 696 #define ppgtt_get_shadow_entry(spt, e, index) \
 697         ppgtt_spt_get_entry(spt, spt->shadow_page.vaddr, \
 698                 spt->shadow_page.type, e, index, false)
 699
 700 #define ppgtt_set_shadow_entry(spt, e, index) \
 701         ppgtt_spt_set_entry(spt, spt->shadow_page.vaddr, \
 702                 spt->shadow_page.type, e, index, false)
 703
 704 static void *alloc_spt(gfp_t gfp_mask)
 705 {
 706         struct intel_vgpu_ppgtt_spt *spt;
 707
 708         spt = kzalloc(sizeof(*spt), gfp_mask);
 709         if (!spt)
 710                 return NULL;
 711
 712         spt->shadow_page.page = alloc_page(gfp_mask);
 713         if (!spt->shadow_page.page) {
 714                 kfree(spt);
 715                 return NULL;
 716         }
 717         return spt;
 718 }
 719
 720 static void free_spt(struct intel_vgpu_ppgtt_spt *spt)
 721 {
 722         __free_page(spt->shadow_page.page);
 723         kfree(spt);
 724 }
 725
 726 static int detach_oos_page(struct intel_vgpu *vgpu,
 727                 struct intel_vgpu_oos_page *oos_page);
 728
 729 static void ppgtt_free_spt(struct intel_vgpu_ppgtt_spt *spt)
 730 {
 731         struct device *kdev = &spt->vgpu->gvt->dev_priv->drm.pdev->dev;
 732
 733         trace_spt_free(spt->vgpu->id, spt, spt->guest_page.type);
 734
 735         dma_unmap_page(kdev, spt->shadow_page.mfn << I915_GTT_PAGE_SHIFT, 4096,
 736                        PCI_DMA_BIDIRECTIONAL);
 737
 738         radix_tree_delete(&spt->vgpu->gtt.spt_tree, spt->shadow_page.mfn);
 739
 740         if (spt->guest_page.gfn) {
 741                 if (spt->guest_page.oos_page)
 742                         detach_oos_page(spt->vgpu, spt->guest_page.oos_page);
 743
 744                 intel_vgpu_unregister_page_track(spt->vgpu, spt->guest_page.gfn);
 745         }
 746
 747         list_del_init(&spt->post_shadow_list);
 748         free_spt(spt);
 749 }
 750
 751 static void ppgtt_free_all_spt(struct intel_vgpu *vgpu)
 752 {
 753         struct intel_vgpu_ppgtt_spt *spt;
 754         struct radix_tree_iter iter;
 755         void **slot;
 756
 757         radix_tree_for_each_slot(slot, &vgpu->gtt.spt_tree, &iter, 0) {
 758                 spt = radix_tree_deref_slot(slot);
 759                 ppgtt_free_spt(spt);
 760         }
 761 }
 762
 763 static int ppgtt_handle_guest_write_page_table_bytes(
 764                 struct intel_vgpu_ppgtt_spt *spt,
 765                 u64 pa, void *p_data, int bytes);
 766
 767 static int ppgtt_write_protection_handler(
 768                 struct intel_vgpu_page_track *page_track,
 769                 u64 gpa, void *data, int bytes)
 770 {
 771         struct intel_vgpu_ppgtt_spt *spt = page_track->priv_data;
 772
 773         int ret;
 774
 775         if (bytes != 4 && bytes != 8)
 776                 return -EINVAL;
 777
 778         ret = ppgtt_handle_guest_write_page_table_bytes(spt, gpa, data, bytes);
 779         if (ret)
 780                 return ret;
 781         return ret;
 782 }
 783
 784 /* Find a spt by guest gfn. */
 785 static struct intel_vgpu_ppgtt_spt *intel_vgpu_find_spt_by_gfn(
 786                 struct intel_vgpu *vgpu, unsigned long gfn)
 787 {
 788         struct intel_vgpu_page_track *track;
 789
 790         track = intel_vgpu_find_page_track(vgpu, gfn);
 791         if (track && track->handler == ppgtt_write_protection_handler)
 792                 return track->priv_data;
 793
 794         return NULL;
 795 }
 796
 797 /* Find the spt by shadow page mfn. */
 798 static inline struct intel_vgpu_ppgtt_spt *intel_vgpu_find_spt_by_mfn(
 799                 struct intel_vgpu *vgpu, unsigned long mfn)
 800 {
 801         return radix_tree_lookup(&vgpu->gtt.spt_tree, mfn);
 802 }
 803
 804 static int reclaim_one_ppgtt_mm(struct intel_gvt *gvt);
 805
 806 /* Allocate shadow page table without guest page. */
 807 static struct intel_vgpu_ppgtt_spt *ppgtt_alloc_spt(
 808                 struct intel_vgpu *vgpu, intel_gvt_gtt_type_t type)
 809 {
 810         struct device *kdev = &vgpu->gvt->dev_priv->drm.pdev->dev;
 811         struct intel_vgpu_ppgtt_spt *spt = NULL;
 812         dma_addr_t daddr;
 813         int ret;
 814
 815 retry:
 816         spt = alloc_spt(GFP_KERNEL | __GFP_ZERO);
 817         if (!spt) {
 818                 if (reclaim_one_ppgtt_mm(vgpu->gvt))
 819                         goto retry;
 820
 821                 gvt_vgpu_err("fail to allocate ppgtt shadow page\n");
 822                 return ERR_PTR(-ENOMEM);
 823         }
 824
 825         spt->vgpu = vgpu;
 826         atomic_set(&spt->refcount, 1);
 827         INIT_LIST_HEAD(&spt->post_shadow_list);
 828
 829         /*
 830          * Init shadow_page.
 831          */
 832         spt->shadow_page.type = type;
 833         daddr = dma_map_page(kdev, spt->shadow_page.page,
 834                              0, 4096, PCI_DMA_BIDIRECTIONAL);
 835         if (dma_mapping_error(kdev, daddr)) {
 836                 gvt_vgpu_err("fail to map dma addr\n");
 837                 ret = -EINVAL;
 838                 goto err_free_spt;
 839         }
 840         spt->shadow_page.vaddr = page_address(spt->shadow_page.page);
 841         spt->shadow_page.mfn = daddr >> I915_GTT_PAGE_SHIFT;
 842
 843         ret = radix_tree_insert(&vgpu->gtt.spt_tree, spt->shadow_page.mfn, spt);
 844         if (ret)
 845                 goto err_unmap_dma;
 846
 847         return spt;
 848
 849 err_unmap_dma:
 850         dma_unmap_page(kdev, daddr, PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 851 err_free_spt:
 852         free_spt(spt);
 853         return ERR_PTR(ret);
 854 }
 855
 856 /* Allocate shadow page table associated with specific gfn. */
 857 static struct intel_vgpu_ppgtt_spt *ppgtt_alloc_spt_gfn(
 858                 struct intel_vgpu *vgpu, intel_gvt_gtt_type_t type,
 859                 unsigned long gfn, bool guest_pde_ips)
 860 {
 861         struct intel_vgpu_ppgtt_spt *spt;
 862         int ret;
 863
 864         spt = ppgtt_alloc_spt(vgpu, type);
 865         if (IS_ERR(spt))
 866                 return spt;
 867
 868         /*
 869          * Init guest_page.
 870          */
 871         ret = intel_vgpu_register_page_track(vgpu, gfn,
 872                         ppgtt_write_protection_handler, spt);
 873         if (ret) {
 874                 ppgtt_free_spt(spt);
 875                 return ERR_PTR(ret);
 876         }
 877
 878         spt->guest_page.type = type;
 879         spt->guest_page.gfn = gfn;
 880         spt->guest_page.pde_ips = guest_pde_ips;
 881
 882         trace_spt_alloc(vgpu->id, spt, type, spt->shadow_page.mfn, gfn);
 883
 884         return spt;
 885 }
 886
 887 #define pt_entry_size_shift(spt) \
 888         ((spt)->vgpu->gvt->device_info.gtt_entry_size_shift)
 889
 890 #define pt_entries(spt) \
 891         (I915_GTT_PAGE_SIZE >> pt_entry_size_shift(spt))
 892
 893 #define for_each_present_guest_entry(spt, e, i) \
 894         for (i = 0; i < pt_entries(spt); \
 895              i += spt->guest_page.pde_ips ? GTT_64K_PTE_STRIDE : 1) \
 896                 if (!ppgtt_get_guest_entry(spt, e, i) && \
 897                     spt->vgpu->gvt->gtt.pte_ops->test_present(e))
 898
 899 #define for_each_present_shadow_entry(spt, e, i) \
 900         for (i = 0; i < pt_entries(spt); \
 901              i += spt->shadow_page.pde_ips ? GTT_64K_PTE_STRIDE : 1) \
 902                 if (!ppgtt_get_shadow_entry(spt, e, i) && \
 903                     spt->vgpu->gvt->gtt.pte_ops->test_present(e))
 904
 905 #define for_each_shadow_entry(spt, e, i) \
 906         for (i = 0; i < pt_entries(spt); \
 907              i += (spt->shadow_page.pde_ips ? GTT_64K_PTE_STRIDE : 1)) \
 908                 if (!ppgtt_get_shadow_entry(spt, e, i))
 909
 910 static inline void ppgtt_get_spt(struct intel_vgpu_ppgtt_spt *spt)
 911 {
 912         int v = atomic_read(&spt->refcount);
 913
 914         trace_spt_refcount(spt->vgpu->id, "inc", spt, v, (v + 1));
 915         atomic_inc(&spt->refcount);
 916 }
 917
 918 static inline int ppgtt_put_spt(struct intel_vgpu_ppgtt_spt *spt)
 919 {
 920         int v = atomic_read(&spt->refcount);
 921
 922         trace_spt_refcount(spt->vgpu->id, "dec", spt, v, (v - 1));
 923         return atomic_dec_return(&spt->refcount);
 924 }
 925
 926 static int ppgtt_invalidate_spt(struct intel_vgpu_ppgtt_spt *spt);
 927
 928 static int ppgtt_invalidate_spt_by_shadow_entry(struct intel_vgpu *vgpu,
 929                 struct intel_gvt_gtt_entry *e)
 930 {
 931         struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
 932         struct intel_vgpu_ppgtt_spt *s;
 933         intel_gvt_gtt_type_t cur_pt_type;
 934
 935         GEM_BUG_ON(!gtt_type_is_pt(get_next_pt_type(e->type)));
 936
 937         if (e->type != GTT_TYPE_PPGTT_ROOT_L3_ENTRY
 938                 && e->type != GTT_TYPE_PPGTT_ROOT_L4_ENTRY) {
 939                 cur_pt_type = get_next_pt_type(e->type) + 1;
 940                 if (ops->get_pfn(e) ==
 941                         vgpu->gtt.scratch_pt[cur_pt_type].page_mfn)
 942                         return 0;
 943         }
 944         s = intel_vgpu_find_spt_by_mfn(vgpu, ops->get_pfn(e));
 945         if (!s) {
 946                 gvt_vgpu_err("fail to find shadow page: mfn: 0x%lx\n",
 947                                 ops->get_pfn(e));
 948                 return -ENXIO;
 949         }
 950         return ppgtt_invalidate_spt(s);
 951 }
 952
 953 static inline void ppgtt_invalidate_pte(struct intel_vgpu_ppgtt_spt *spt,
 954                 struct intel_gvt_gtt_entry *entry)
 955 {
 956         struct intel_vgpu *vgpu = spt->vgpu;
 957         struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
 958         unsigned long pfn;
 959         int type;
 960
 961         pfn = ops->get_pfn(entry);
 962         type = spt->shadow_page.type;
 963
 964         /* Uninitialized spte or unshadowed spte. */
 965         if (!pfn || pfn == vgpu->gtt.scratch_pt[type].page_mfn)
 966                 return;
 967
 968         intel_gvt_hypervisor_dma_unmap_guest_page(vgpu, pfn << PAGE_SHIFT);
 969 }
 970
 971 static int ppgtt_invalidate_spt(struct intel_vgpu_ppgtt_spt *spt)
 972 {
 973         struct intel_vgpu *vgpu = spt->vgpu;
 974         struct intel_gvt_gtt_entry e;
 975         unsigned long index;
 976         int ret;
 977
 978         trace_spt_change(spt->vgpu->id, "die", spt,
 979                         spt->guest_page.gfn, spt->shadow_page.type);
 980
 981         if (ppgtt_put_spt(spt) > 0)
 982                 return 0;
 983
 984         for_each_present_shadow_entry(spt, &e, index) {
 985                 switch (e.type) {
 986                 case GTT_TYPE_PPGTT_PTE_4K_ENTRY:
 987                         gvt_vdbg_mm("invalidate 4K entry\n");
 988                         ppgtt_invalidate_pte(spt, &e);
 989                         break;
 990                 case GTT_TYPE_PPGTT_PTE_64K_ENTRY:
 991                         /* We don't setup 64K shadow entry so far. */
 992                         WARN(1, "suspicious 64K gtt entry\n");
 993                         continue;
 994                 case GTT_TYPE_PPGTT_PTE_2M_ENTRY:
 995                         gvt_vdbg_mm("invalidate 2M entry\n");
 996                         continue;
 997                 case GTT_TYPE_PPGTT_PTE_1G_ENTRY:
 998                         WARN(1, "GVT doesn't support 1GB page\n");
 999                         continue;
1000                 case GTT_TYPE_PPGTT_PML4_ENTRY:
1001                 case GTT_TYPE_PPGTT_PDP_ENTRY:
1002                 case GTT_TYPE_PPGTT_PDE_ENTRY:
1003                         gvt_vdbg_mm("invalidate PMUL4/PDP/PDE entry\n");
1004                         ret = ppgtt_invalidate_spt_by_shadow_entry(
1005                                         spt->vgpu, &e);
1006                         if (ret)
1007                                 goto fail;
1008                         break;
1009                 default:
1010                         GEM_BUG_ON(1);
1011                 }
1012         }
1013
1014         trace_spt_change(spt->vgpu->id, "release", spt,
1015                          spt->guest_page.gfn, spt->shadow_page.type);
1016         ppgtt_free_spt(spt);
1017         return 0;
1018 fail:
1019         gvt_vgpu_err("fail: shadow page %p shadow entry 0x%llx type %d\n",
1020                         spt, e.val64, e.type);
1021         return ret;
1022 }
1023
1024 static bool vgpu_ips_enabled(struct intel_vgpu *vgpu)
1025 {
1026         struct drm_i915_private *dev_priv = vgpu->gvt->dev_priv;
1027
1028         if (INTEL_GEN(dev_priv) == 9 || INTEL_GEN(dev_priv) == 10) {
1029                 u32 ips = vgpu_vreg_t(vgpu, GEN8_GAMW_ECO_DEV_RW_IA) &
1030                         GAMW_ECO_ENABLE_64K_IPS_FIELD;
1031
1032                 return ips == GAMW_ECO_ENABLE_64K_IPS_FIELD;
1033         } else if (INTEL_GEN(dev_priv) >= 11) {
1034                 /* 64K paging only controlled by IPS bit in PTE now. */
1035                 return true;
1036         } else
1037                 return false;
1038 }
1039
1040 static int ppgtt_populate_spt(struct intel_vgpu_ppgtt_spt *spt);
1041
1042 static struct intel_vgpu_ppgtt_spt *ppgtt_populate_spt_by_guest_entry(
1043                 struct intel_vgpu *vgpu, struct intel_gvt_gtt_entry *we)
1044 {
1045         struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
1046         struct intel_vgpu_ppgtt_spt *spt = NULL;
1047         bool ips = false;
1048         int ret;
1049
1050         GEM_BUG_ON(!gtt_type_is_pt(get_next_pt_type(we->type)));
1051
1052         if (we->type == GTT_TYPE_PPGTT_PDE_ENTRY)
1053                 ips = vgpu_ips_enabled(vgpu) && ops->test_ips(we);
1054
1055         spt = intel_vgpu_find_spt_by_gfn(vgpu, ops->get_pfn(we));
1056         if (spt) {
1057                 ppgtt_get_spt(spt);
1058
1059                 if (ips != spt->guest_page.pde_ips) {
1060                         spt->guest_page.pde_ips = ips;
1061
1062                         gvt_dbg_mm("reshadow PDE since ips changed\n");
1063                         clear_page(spt->shadow_page.vaddr);
1064                         ret = ppgtt_populate_spt(spt);
1065                         if (ret) {
1066                                 ppgtt_put_spt(spt);
1067                                 goto err;
1068                         }
1069                 }
1070         } else {
1071                 int type = get_next_pt_type(we->type);
1072
1073                 spt = ppgtt_alloc_spt_gfn(vgpu, type, ops->get_pfn(we), ips);
1074                 if (IS_ERR(spt)) {
1075                         ret = PTR_ERR(spt);
1076                         goto err;
1077                 }
1078
1079                 ret = intel_vgpu_enable_page_track(vgpu, spt->guest_page.gfn);
1080                 if (ret)
1081                         goto err_free_spt;
1082
1083                 ret = ppgtt_populate_spt(spt);
1084                 if (ret)
1085                         goto err_free_spt;
1086
1087                 trace_spt_change(vgpu->id, "new", spt, spt->guest_page.gfn,
1088                                  spt->shadow_page.type);
1089         }
1090         return spt;
1091
1092 err_free_spt:
1093         ppgtt_free_spt(spt);
1094 err:
1095         gvt_vgpu_err("fail: shadow page %p guest entry 0x%llx type %d\n",
1096                      spt, we->val64, we->type);
1097         return ERR_PTR(ret);
1098 }
1099
1100 static inline void ppgtt_generate_shadow_entry(struct intel_gvt_gtt_entry *se,
1101                 struct intel_vgpu_ppgtt_spt *s, struct intel_gvt_gtt_entry *ge)
1102 {
1103         struct intel_gvt_gtt_pte_ops *ops = s->vgpu->gvt->gtt.pte_ops;
1104
1105         se->type = ge->type;
1106         se->val64 = ge->val64;
1107
1108         /* Because we always split 64KB pages, so clear IPS in shadow PDE. */
1109         if (se->type == GTT_TYPE_PPGTT_PDE_ENTRY)
1110                 ops->clear_ips(se);
1111
1112         ops->set_pfn(se, s->shadow_page.mfn);
1113 }
1114
1115 /**
1116  * Check if can do 2M page
1117  * @vgpu: target vgpu
1118  * @entry: target pfn's gtt entry
1119  *
1120  * Return 1 if 2MB huge gtt shadowing is possilbe, 0 if miscondition,
1121  * negtive if found err.
1122  */
1123 static int is_2MB_gtt_possible(struct intel_vgpu *vgpu,
1124         struct intel_gvt_gtt_entry *entry)
1125 {
1126         struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
1127         unsigned long pfn;
1128
1129         if (!HAS_PAGE_SIZES(vgpu->gvt->dev_priv, I915_GTT_PAGE_SIZE_2M))
1130                 return 0;
1131
1132         pfn = intel_gvt_hypervisor_gfn_to_mfn(vgpu, ops->get_pfn(entry));
1133         if (pfn == INTEL_GVT_INVALID_ADDR)
1134                 return -EINVAL;
1135
1136         return PageTransHuge(pfn_to_page(pfn));
1137 }
1138
1139 static int split_2MB_gtt_entry(struct intel_vgpu *vgpu,
1140         struct intel_vgpu_ppgtt_spt *spt, unsigned long index,
1141         struct intel_gvt_gtt_entry *se)
1142 {
1143         struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
1144         struct intel_vgpu_ppgtt_spt *sub_spt;
1145         struct intel_gvt_gtt_entry sub_se;
1146         unsigned long start_gfn;
1147         dma_addr_t dma_addr;
1148         unsigned long sub_index;
1149         int ret;
1150
1151         gvt_dbg_mm("Split 2M gtt entry, index %lu\n", index);
1152
1153         start_gfn = ops->get_pfn(se);
1154
1155         sub_spt = ppgtt_alloc_spt(vgpu, GTT_TYPE_PPGTT_PTE_PT);
1156         if (IS_ERR(sub_spt))
1157                 return PTR_ERR(sub_spt);
1158
1159         for_each_shadow_entry(sub_spt, &sub_se, sub_index) {
1160                 ret = intel_gvt_hypervisor_dma_map_guest_page(vgpu,
1161                                 start_gfn + sub_index, PAGE_SIZE, &dma_addr);
1162                 if (ret) {
1163                         ppgtt_invalidate_spt(spt);
1164                         return ret;
1165                 }
1166                 sub_se.val64 = se->val64;
1167
1168                 /* Copy the PAT field from PDE. */
1169                 sub_se.val64 &= ~_PAGE_PAT;
1170                 sub_se.val64 |= (se->val64 & _PAGE_PAT_LARGE) >> 5;
1171
1172                 ops->set_pfn(&sub_se, dma_addr >> PAGE_SHIFT);
1173                 ppgtt_set_shadow_entry(sub_spt, &sub_se, sub_index);
1174         }
1175
1176         /* Clear dirty field. */
1177         se->val64 &= ~_PAGE_DIRTY;
1178
1179         ops->clear_pse(se);
1180         ops->clear_ips(se);
1181         ops->set_pfn(se, sub_spt->shadow_page.mfn);
1182         ppgtt_set_shadow_entry(spt, se, index);
1183         return 0;
1184 }
1185
1186 static int split_64KB_gtt_entry(struct intel_vgpu *vgpu,
1187         struct intel_vgpu_ppgtt_spt *spt, unsigned long index,
1188         struct intel_gvt_gtt_entry *se)
1189 {
1190         struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
1191         struct intel_gvt_gtt_entry entry = *se;
1192         unsigned long start_gfn;
1193         dma_addr_t dma_addr;
1194         int i, ret;
1195
1196         gvt_vdbg_mm("Split 64K gtt entry, index %lu\n", index);
1197
1198         GEM_BUG_ON(index % GTT_64K_PTE_STRIDE);
1199
1200         start_gfn = ops->get_pfn(se);
1201
1202         entry.type = GTT_TYPE_PPGTT_PTE_4K_ENTRY;
1203         ops->set_64k_splited(&entry);
1204
1205         for (i = 0; i < GTT_64K_PTE_STRIDE; i++) {
1206                 ret = intel_gvt_hypervisor_dma_map_guest_page(vgpu,
1207                                         start_gfn + i, PAGE_SIZE, &dma_addr);
1208                 if (ret)
1209                         return ret;
1210
1211                 ops->set_pfn(&entry, dma_addr >> PAGE_SHIFT);
1212                 ppgtt_set_shadow_entry(spt, &entry, index + i);
1213         }
1214         return 0;
1215 }
1216
1217 static int ppgtt_populate_shadow_entry(struct intel_vgpu *vgpu,
1218         struct intel_vgpu_ppgtt_spt *spt, unsigned long index,
1219         struct intel_gvt_gtt_entry *ge)
1220 {
1221         struct intel_gvt_gtt_pte_ops *pte_ops = vgpu->gvt->gtt.pte_ops;
1222         struct intel_gvt_gtt_entry se = *ge;
1223         unsigned long gfn, page_size = PAGE_SIZE;
1224         dma_addr_t dma_addr;
1225         int ret;
1226
1227         if (!pte_ops->test_present(ge))
1228                 return 0;
1229
1230         gfn = pte_ops->get_pfn(ge);
1231
1232         switch (ge->type) {
1233         case GTT_TYPE_PPGTT_PTE_4K_ENTRY:
1234                 gvt_vdbg_mm("shadow 4K gtt entry\n");
1235                 break;
1236         case GTT_TYPE_PPGTT_PTE_64K_ENTRY:
1237                 gvt_vdbg_mm("shadow 64K gtt entry\n");
1238                 /*
1239                  * The layout of 64K page is special, the page size is
1240                  * controlled by uper PDE. To be simple, we always split
1241                  * 64K page to smaller 4K pages in shadow PT.
1242                  */
1243                 return split_64KB_gtt_entry(vgpu, spt, index, &se);
1244         case GTT_TYPE_PPGTT_PTE_2M_ENTRY:
1245                 gvt_vdbg_mm("shadow 2M gtt entry\n");
1246                 ret = is_2MB_gtt_possible(vgpu, ge);
1247                 if (ret == 0)
1248                         return split_2MB_gtt_entry(vgpu, spt, index, &se);
1249                 else if (ret < 0)
1250                         return ret;
1251                 page_size = I915_GTT_PAGE_SIZE_2M;
1252                 break;
1253         case GTT_TYPE_PPGTT_PTE_1G_ENTRY:
1254                 gvt_vgpu_err("GVT doesn't support 1GB entry\n");
1255                 return -EINVAL;
1256         default:
1257                 GEM_BUG_ON(1);
1258         };
1259
1260         /* direct shadow */
1261         ret = intel_gvt_hypervisor_dma_map_guest_page(vgpu, gfn, page_size,
1262                                                       &dma_addr);
1263         if (ret)
1264                 return -ENXIO;
1265
1266         pte_ops->set_pfn(&se, dma_addr >> PAGE_SHIFT);
1267         ppgtt_set_shadow_entry(spt, &se, index);
1268         return 0;
1269 }
1270
1271 static int ppgtt_populate_spt(struct intel_vgpu_ppgtt_spt *spt)
1272 {
1273         struct intel_vgpu *vgpu = spt->vgpu;
1274         struct intel_gvt *gvt = vgpu->gvt;
1275         struct intel_gvt_gtt_pte_ops *ops = gvt->gtt.pte_ops;
1276         struct intel_vgpu_ppgtt_spt *s;
1277         struct intel_gvt_gtt_entry se, ge;
1278         unsigned long gfn, i;
1279         int ret;
1280
1281         trace_spt_change(spt->vgpu->id, "born", spt,
1282                          spt->guest_page.gfn, spt->shadow_page.type);
1283
1284         for_each_present_guest_entry(spt, &ge, i) {
1285                 if (gtt_type_is_pt(get_next_pt_type(ge.type))) {
1286                         s = ppgtt_populate_spt_by_guest_entry(vgpu, &ge);
1287                         if (IS_ERR(s)) {
1288                                 ret = PTR_ERR(s);
1289                                 goto fail;
1290                         }
1291                         ppgtt_get_shadow_entry(spt, &se, i);
1292                         ppgtt_generate_shadow_entry(&se, s, &ge);
1293                         ppgtt_set_shadow_entry(spt, &se, i);
1294                 } else {
1295                         gfn = ops->get_pfn(&ge);
1296                         if (!intel_gvt_hypervisor_is_valid_gfn(vgpu, gfn)) {
1297                                 ops->set_pfn(&se, gvt->gtt.scratch_mfn);
1298                                 ppgtt_set_shadow_entry(spt, &se, i);
1299                                 continue;
1300                         }
1301
1302                         ret = ppgtt_populate_shadow_entry(vgpu, spt, i, &ge);
1303                         if (ret)
1304                                 goto fail;
1305                 }
1306         }
1307         return 0;
1308 fail:
1309         gvt_vgpu_err("fail: shadow page %p guest entry 0x%llx type %d\n",
1310                         spt, ge.val64, ge.type);
1311         return ret;
1312 }
1313
1314 static int ppgtt_handle_guest_entry_removal(struct intel_vgpu_ppgtt_spt *spt,
1315                 struct intel_gvt_gtt_entry *se, unsigned long index)
1316 {
1317         struct intel_vgpu *vgpu = spt->vgpu;
1318         struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
1319         int ret;
1320
1321         trace_spt_guest_change(spt->vgpu->id, "remove", spt,
1322                                spt->shadow_page.type, se->val64, index);
1323
1324         gvt_vdbg_mm("destroy old shadow entry, type %d, index %lu, value %llx\n",
1325                     se->type, index, se->val64);
1326
1327         if (!ops->test_present(se))
1328                 return 0;
1329
1330         if (ops->get_pfn(se) ==
1331             vgpu->gtt.scratch_pt[spt->shadow_page.type].page_mfn)
1332                 return 0;
1333
1334         if (gtt_type_is_pt(get_next_pt_type(se->type))) {
1335                 struct intel_vgpu_ppgtt_spt *s =
1336                         intel_vgpu_find_spt_by_mfn(vgpu, ops->get_pfn(se));
1337                 if (!s) {
1338                         gvt_vgpu_err("fail to find guest page\n");
1339                         ret = -ENXIO;
1340                         goto fail;
1341                 }
1342                 ret = ppgtt_invalidate_spt(s);
1343                 if (ret)
1344                         goto fail;
1345         } else {
1346                 /* We don't setup 64K shadow entry so far. */
1347                 WARN(se->type == GTT_TYPE_PPGTT_PTE_64K_ENTRY,
1348                      "suspicious 64K entry\n");
1349                 ppgtt_invalidate_pte(spt, se);
1350         }
1351
1352         return 0;
1353 fail:
1354         gvt_vgpu_err("fail: shadow page %p guest entry 0x%llx type %d\n",
1355                         spt, se->val64, se->type);
1356         return ret;
1357 }
1358
1359 static int ppgtt_handle_guest_entry_add(struct intel_vgpu_ppgtt_spt *spt,
1360                 struct intel_gvt_gtt_entry *we, unsigned long index)
1361 {
1362         struct intel_vgpu *vgpu = spt->vgpu;
1363         struct intel_gvt_gtt_entry m;
1364         struct intel_vgpu_ppgtt_spt *s;
1365         int ret;
1366
1367         trace_spt_guest_change(spt->vgpu->id, "add", spt, spt->shadow_page.type,
1368                                we->val64, index);
1369
1370         gvt_vdbg_mm("add shadow entry: type %d, index %lu, value %llx\n",
1371                     we->type, index, we->val64);
1372
1373         if (gtt_type_is_pt(get_next_pt_type(we->type))) {
1374                 s = ppgtt_populate_spt_by_guest_entry(vgpu, we);
1375                 if (IS_ERR(s)) {
1376                         ret = PTR_ERR(s);
1377                         goto fail;
1378                 }
1379                 ppgtt_get_shadow_entry(spt, &m, index);
1380                 ppgtt_generate_shadow_entry(&m, s, we);
1381                 ppgtt_set_shadow_entry(spt, &m, index);
1382         } else {
1383                 ret = ppgtt_populate_shadow_entry(vgpu, spt, index, we);
1384                 if (ret)
1385                         goto fail;
1386         }
1387         return 0;
1388 fail:
1389         gvt_vgpu_err("fail: spt %p guest entry 0x%llx type %d\n",
1390                 spt, we->val64, we->type);
1391         return ret;
1392 }
1393
1394 static int sync_oos_page(struct intel_vgpu *vgpu,
1395                 struct intel_vgpu_oos_page *oos_page)
1396 {
1397         const struct intel_gvt_device_info *info = &vgpu->gvt->device_info;
1398         struct intel_gvt *gvt = vgpu->gvt;
1399         struct intel_gvt_gtt_pte_ops *ops = gvt->gtt.pte_ops;
1400         struct intel_vgpu_ppgtt_spt *spt = oos_page->spt;
1401         struct intel_gvt_gtt_entry old, new;
1402         int index;
1403         int ret;
1404
1405         trace_oos_change(vgpu->id, "sync", oos_page->id,
1406                          spt, spt->guest_page.type);
1407
1408         old.type = new.type = get_entry_type(spt->guest_page.type);
1409         old.val64 = new.val64 = 0;
1410
1411         for (index = 0; index < (I915_GTT_PAGE_SIZE >>
1412                                 info->gtt_entry_size_shift); index++) {
1413                 ops->get_entry(oos_page->mem, &old, index, false, 0, vgpu);
1414                 ops->get_entry(NULL, &new, index, true,
1415                                spt->guest_page.gfn << PAGE_SHIFT, vgpu);
1416
1417                 if (old.val64 == new.val64
1418                         && !test_and_clear_bit(index, spt->post_shadow_bitmap))
1419                         continue;
1420
1421                 trace_oos_sync(vgpu->id, oos_page->id,
1422                                 spt, spt->guest_page.type,
1423                                 new.val64, index);
1424
1425                 ret = ppgtt_populate_shadow_entry(vgpu, spt, index, &new);
1426                 if (ret)
1427                         return ret;
1428
1429                 ops->set_entry(oos_page->mem, &new, index, false, 0, vgpu);
1430         }
1431
1432         spt->guest_page.write_cnt = 0;
1433         list_del_init(&spt->post_shadow_list);
1434         return 0;
1435 }
1436
1437 static int detach_oos_page(struct intel_vgpu *vgpu,
1438                 struct intel_vgpu_oos_page *oos_page)
1439 {
1440         struct intel_gvt *gvt = vgpu->gvt;
1441         struct intel_vgpu_ppgtt_spt *spt = oos_page->spt;
1442
1443         trace_oos_change(vgpu->id, "detach", oos_page->id,
1444                          spt, spt->guest_page.type);
1445
1446         spt->guest_page.write_cnt = 0;
1447         spt->guest_page.oos_page = NULL;
1448         oos_page->spt = NULL;
1449
1450         list_del_init(&oos_page->vm_list);
1451         list_move_tail(&oos_page->list, &gvt->gtt.oos_page_free_list_head);
1452
1453         return 0;
1454 }
1455
1456 static int attach_oos_page(struct intel_vgpu_oos_page *oos_page,
1457                 struct intel_vgpu_ppgtt_spt *spt)
1458 {
1459         struct intel_gvt *gvt = spt->vgpu->gvt;
1460         int ret;
1461
1462         ret = intel_gvt_hypervisor_read_gpa(spt->vgpu,
1463                         spt->guest_page.gfn << I915_GTT_PAGE_SHIFT,
1464                         oos_page->mem, I915_GTT_PAGE_SIZE);
1465         if (ret)
1466                 return ret;
1467
1468         oos_page->spt = spt;
1469         spt->guest_page.oos_page = oos_page;
1470
1471         list_move_tail(&oos_page->list, &gvt->gtt.oos_page_use_list_head);
1472
1473         trace_oos_change(spt->vgpu->id, "attach", oos_page->id,
1474                          spt, spt->guest_page.type);
1475         return 0;
1476 }
1477
1478 static int ppgtt_set_guest_page_sync(struct intel_vgpu_ppgtt_spt *spt)
1479 {
1480         struct intel_vgpu_oos_page *oos_page = spt->guest_page.oos_page;
1481         int ret;
1482
1483         ret = intel_vgpu_enable_page_track(spt->vgpu, spt->guest_page.gfn);
1484         if (ret)
1485                 return ret;
1486
1487         trace_oos_change(spt->vgpu->id, "set page sync", oos_page->id,
1488                          spt, spt->guest_page.type);
1489
1490         list_del_init(&oos_page->vm_list);
1491         return sync_oos_page(spt->vgpu, oos_page);
1492 }
1493
1494 static int ppgtt_allocate_oos_page(struct intel_vgpu_ppgtt_spt *spt)
1495 {
1496         struct intel_gvt *gvt = spt->vgpu->gvt;
1497         struct intel_gvt_gtt *gtt = &gvt->gtt;
1498         struct intel_vgpu_oos_page *oos_page = spt->guest_page.oos_page;
1499         int ret;
1500
1501         WARN(oos_page, "shadow PPGTT page has already has a oos page\n");
1502
1503         if (list_empty(&gtt->oos_page_free_list_head)) {
1504                 oos_page = container_of(gtt->oos_page_use_list_head.next,
1505                         struct intel_vgpu_oos_page, list);
1506                 ret = ppgtt_set_guest_page_sync(oos_page->spt);
1507                 if (ret)
1508                         return ret;
1509                 ret = detach_oos_page(spt->vgpu, oos_page);
1510                 if (ret)
1511                         return ret;
1512         } else
1513                 oos_page = container_of(gtt->oos_page_free_list_head.next,
1514                         struct intel_vgpu_oos_page, list);
1515         return attach_oos_page(oos_page, spt);
1516 }
1517
1518 static int ppgtt_set_guest_page_oos(struct intel_vgpu_ppgtt_spt *spt)
1519 {
1520         struct intel_vgpu_oos_page *oos_page = spt->guest_page.oos_page;
1521
1522         if (WARN(!oos_page, "shadow PPGTT page should have a oos page\n"))
1523                 return -EINVAL;
1524
1525         trace_oos_change(spt->vgpu->id, "set page out of sync", oos_page->id,
1526                          spt, spt->guest_page.type);
1527
1528         list_add_tail(&oos_page->vm_list, &spt->vgpu->gtt.oos_page_list_head);
1529         return intel_vgpu_disable_page_track(spt->vgpu, spt->guest_page.gfn);
1530 }
1531
1532 /**
1533  * intel_vgpu_sync_oos_pages - sync all the out-of-synced shadow for vGPU
1534  * @vgpu: a vGPU
1535  *
1536  * This function is called before submitting a guest workload to host,
1537  * to sync all the out-of-synced shadow for vGPU
1538  *
1539  * Returns:
1540  * Zero on success, negative error code if failed.
1541  */
1542 int intel_vgpu_sync_oos_pages(struct intel_vgpu *vgpu)
1543 {
1544         struct list_head *pos, *n;
1545         struct intel_vgpu_oos_page *oos_page;
1546         int ret;
1547
1548         if (!enable_out_of_sync)
1549                 return 0;
1550
1551         list_for_each_safe(pos, n, &vgpu->gtt.oos_page_list_head) {
1552                 oos_page = container_of(pos,
1553                                 struct intel_vgpu_oos_page, vm_list);
1554                 ret = ppgtt_set_guest_page_sync(oos_page->spt);
1555                 if (ret)
1556                         return ret;
1557         }
1558         return 0;
1559 }
1560
1561 /*
1562  * The heart of PPGTT shadow page table.
1563  */
1564 static int ppgtt_handle_guest_write_page_table(
1565                 struct intel_vgpu_ppgtt_spt *spt,
1566                 struct intel_gvt_gtt_entry *we, unsigned long index)
1567 {
1568         struct intel_vgpu *vgpu = spt->vgpu;
1569         int type = spt->shadow_page.type;
1570         struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
1571         struct intel_gvt_gtt_entry old_se;
1572         int new_present;
1573         int i, ret;
1574
1575         new_present = ops->test_present(we);
1576
1577         /*
1578          * Adding the new entry first and then removing the old one, that can
1579          * guarantee the ppgtt table is validated during the window between
1580          * adding and removal.
1581          */
1582         ppgtt_get_shadow_entry(spt, &old_se, index);
1583
1584         if (new_present) {
1585                 ret = ppgtt_handle_guest_entry_add(spt, we, index);
1586                 if (ret)
1587                         goto fail;
1588         }
1589
1590         ret = ppgtt_handle_guest_entry_removal(spt, &old_se, index);
1591         if (ret)
1592                 goto fail;
1593
1594         if (!new_present) {
1595                 /* For 64KB splited entries, we need clear them all. */
1596                 if (ops->test_64k_splited(&old_se) &&
1597                     !(index % GTT_64K_PTE_STRIDE)) {
1598                         gvt_vdbg_mm("remove splited 64K shadow entries\n");
1599                         for (i = 0; i < GTT_64K_PTE_STRIDE; i++) {
1600                                 ops->clear_64k_splited(&old_se);
1601                                 ops->set_pfn(&old_se,
1602                                         vgpu->gtt.scratch_pt[type].page_mfn);
1603                                 ppgtt_set_shadow_entry(spt, &old_se, index + i);
1604                         }
1605                 } else if (old_se.type == GTT_TYPE_PPGTT_PTE_2M_ENTRY ||
1606                            old_se.type == GTT_TYPE_PPGTT_PTE_1G_ENTRY) {
1607                         ops->clear_pse(&old_se);
1608                         ops->set_pfn(&old_se,
1609                                      vgpu->gtt.scratch_pt[type].page_mfn);
1610                         ppgtt_set_shadow_entry(spt, &old_se, index);
1611                 } else {
1612                         ops->set_pfn(&old_se,
1613                                      vgpu->gtt.scratch_pt[type].page_mfn);
1614                         ppgtt_set_shadow_entry(spt, &old_se, index);
1615                 }
1616         }
1617
1618         return 0;
1619 fail:
1620         gvt_vgpu_err("fail: shadow page %p guest entry 0x%llx type %d.\n",
1621                         spt, we->val64, we->type);
1622         return ret;
1623 }
1624
1625
1626
1627 static inline bool can_do_out_of_sync(struct intel_vgpu_ppgtt_spt *spt)
1628 {
1629         return enable_out_of_sync
1630                 && gtt_type_is_pte_pt(spt->guest_page.type)
1631                 && spt->guest_page.write_cnt >= 2;
1632 }
1633
1634 static void ppgtt_set_post_shadow(struct intel_vgpu_ppgtt_spt *spt,
1635                 unsigned long index)
1636 {
1637         set_bit(index, spt->post_shadow_bitmap);
1638         if (!list_empty(&spt->post_shadow_list))
1639                 return;
1640
1641         list_add_tail(&spt->post_shadow_list,
1642                         &spt->vgpu->gtt.post_shadow_list_head);
1643 }
1644
1645 /**
1646  * intel_vgpu_flush_post_shadow - flush the post shadow transactions
1647  * @vgpu: a vGPU
1648  *
1649  * This function is called before submitting a guest workload to host,
1650  * to flush all the post shadows for a vGPU.
1651  *
1652  * Returns:
1653  * Zero on success, negative error code if failed.
1654  */
1655 int intel_vgpu_flush_post_shadow(struct intel_vgpu *vgpu)
1656 {
1657         struct list_head *pos, *n;
1658         struct intel_vgpu_ppgtt_spt *spt;
1659         struct intel_gvt_gtt_entry ge;
1660         unsigned long index;
1661         int ret;
1662
1663         list_for_each_safe(pos, n, &vgpu->gtt.post_shadow_list_head) {
1664                 spt = container_of(pos, struct intel_vgpu_ppgtt_spt,
1665                                 post_shadow_list);
1666
1667                 for_each_set_bit(index, spt->post_shadow_bitmap,
1668                                 GTT_ENTRY_NUM_IN_ONE_PAGE) {
1669                         ppgtt_get_guest_entry(spt, &ge, index);
1670
1671                         ret = ppgtt_handle_guest_write_page_table(spt,
1672                                                         &ge, index);
1673                         if (ret)
1674                                 return ret;
1675                         clear_bit(index, spt->post_shadow_bitmap);
1676                 }
1677                 list_del_init(&spt->post_shadow_list);
1678         }
1679         return 0;
1680 }
1681
1682 static int ppgtt_handle_guest_write_page_table_bytes(
1683                 struct intel_vgpu_ppgtt_spt *spt,
1684                 u64 pa, void *p_data, int bytes)
1685 {
1686         struct intel_vgpu *vgpu = spt->vgpu;
1687         struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
1688         const struct intel_gvt_device_info *info = &vgpu->gvt->device_info;
1689         struct intel_gvt_gtt_entry we, se;
1690         unsigned long index;
1691         int ret;
1692
1693         index = (pa & (PAGE_SIZE - 1)) >> info->gtt_entry_size_shift;
1694
1695         ppgtt_get_guest_entry(spt, &we, index);
1696
1697         /*
1698          * For page table which has 64K gtt entry, only PTE#0, PTE#16,
1699          * PTE#32, ... PTE#496 are used. Unused PTEs update should be
1700          * ignored.
1701          */
1702         if (we.type == GTT_TYPE_PPGTT_PTE_64K_ENTRY &&
1703             (index % GTT_64K_PTE_STRIDE)) {
1704                 gvt_vdbg_mm("Ignore write to unused PTE entry, index %lu\n",
1705                             index);
1706                 return 0;
1707         }
1708
1709         if (bytes == info->gtt_entry_size) {
1710                 ret = ppgtt_handle_guest_write_page_table(spt, &we, index);
1711                 if (ret)
1712                         return ret;
1713         } else {
1714                 if (!test_bit(index, spt->post_shadow_bitmap)) {
1715                         int type = spt->shadow_page.type;
1716
1717                         ppgtt_get_shadow_entry(spt, &se, index);
1718                         ret = ppgtt_handle_guest_entry_removal(spt, &se, index);
1719                         if (ret)
1720                                 return ret;
1721                         ops->set_pfn(&se, vgpu->gtt.scratch_pt[type].page_mfn);
1722                         ppgtt_set_shadow_entry(spt, &se, index);
1723                 }
1724                 ppgtt_set_post_shadow(spt, index);
1725         }
1726
1727         if (!enable_out_of_sync)
1728                 return 0;
1729
1730         spt->guest_page.write_cnt++;
1731
1732         if (spt->guest_page.oos_page)
1733                 ops->set_entry(spt->guest_page.oos_page->mem, &we, index,
1734                                 false, 0, vgpu);
1735
1736         if (can_do_out_of_sync(spt)) {
1737                 if (!spt->guest_page.oos_page)
1738                         ppgtt_allocate_oos_page(spt);
1739
1740                 ret = ppgtt_set_guest_page_oos(spt);
1741                 if (ret < 0)
1742                         return ret;
1743         }
1744         return 0;
1745 }
1746
1747 static void invalidate_ppgtt_mm(struct intel_vgpu_mm *mm)
1748 {
1749         struct intel_vgpu *vgpu = mm->vgpu;
1750         struct intel_gvt *gvt = vgpu->gvt;
1751         struct intel_gvt_gtt *gtt = &gvt->gtt;
1752         struct intel_gvt_gtt_pte_ops *ops = gtt->pte_ops;
1753         struct intel_gvt_gtt_entry se;
1754         int index;
1755
1756         if (!mm->ppgtt_mm.shadowed)
1757                 return;
1758
1759         for (index = 0; index < ARRAY_SIZE(mm->ppgtt_mm.shadow_pdps); index++) {
1760                 ppgtt_get_shadow_root_entry(mm, &se, index);
1761
1762                 if (!ops->test_present(&se))
1763                         continue;
1764
1765                 ppgtt_invalidate_spt_by_shadow_entry(vgpu, &se);
1766                 se.val64 = 0;
1767                 ppgtt_set_shadow_root_entry(mm, &se, index);
1768
1769                 trace_spt_guest_change(vgpu->id, "destroy root pointer",
1770                                        NULL, se.type, se.val64, index);
1771         }
1772
1773         mm->ppgtt_mm.shadowed = false;
1774 }
1775
1776
1777 static int shadow_ppgtt_mm(struct intel_vgpu_mm *mm)
1778 {
1779         struct intel_vgpu *vgpu = mm->vgpu;
1780         struct intel_gvt *gvt = vgpu->gvt;
1781         struct intel_gvt_gtt *gtt = &gvt->gtt;
1782         struct intel_gvt_gtt_pte_ops *ops = gtt->pte_ops;
1783         struct intel_vgpu_ppgtt_spt *spt;
1784         struct intel_gvt_gtt_entry ge, se;
1785         int index, ret;
1786
1787         if (mm->ppgtt_mm.shadowed)
1788                 return 0;
1789
1790         mm->ppgtt_mm.shadowed = true;
1791
1792         for (index = 0; index < ARRAY_SIZE(mm->ppgtt_mm.guest_pdps); index++) {
1793                 ppgtt_get_guest_root_entry(mm, &ge, index);
1794
1795                 if (!ops->test_present(&ge))
1796                         continue;
1797
1798                 trace_spt_guest_change(vgpu->id, __func__, NULL,
1799                                        ge.type, ge.val64, index);
1800
1801                 spt = ppgtt_populate_spt_by_guest_entry(vgpu, &ge);
1802                 if (IS_ERR(spt)) {
1803                         gvt_vgpu_err("fail to populate guest root pointer\n");
1804                         ret = PTR_ERR(spt);
1805                         goto fail;
1806                 }
1807                 ppgtt_generate_shadow_entry(&se, spt, &ge);
1808                 ppgtt_set_shadow_root_entry(mm, &se, index);
1809
1810                 trace_spt_guest_change(vgpu->id, "populate root pointer",
1811                                        NULL, se.type, se.val64, index);
1812         }
1813
1814         return 0;
1815 fail:
1816         invalidate_ppgtt_mm(mm);
1817         return ret;
1818 }
1819
1820 static struct intel_vgpu_mm *vgpu_alloc_mm(struct intel_vgpu *vgpu)
1821 {
1822         struct intel_vgpu_mm *mm;
1823
1824         mm = kzalloc(sizeof(*mm), GFP_KERNEL);
1825         if (!mm)
1826                 return NULL;
1827
1828         mm->vgpu = vgpu;
1829         kref_init(&mm->ref);
1830         atomic_set(&mm->pincount, 0);
1831
1832         return mm;
1833 }
1834
1835 static void vgpu_free_mm(struct intel_vgpu_mm *mm)
1836 {
1837         kfree(mm);
1838 }
1839
1840 /**
1841  * intel_vgpu_create_ppgtt_mm - create a ppgtt mm object for a vGPU
1842  * @vgpu: a vGPU
1843  * @root_entry_type: ppgtt root entry type
1844  * @pdps: guest pdps.
1845  *
1846  * This function is used to create a ppgtt mm object for a vGPU.
1847  *
1848  * Returns:
1849  * Zero on success, negative error code in pointer if failed.
1850  */
1851 struct intel_vgpu_mm *intel_vgpu_create_ppgtt_mm(struct intel_vgpu *vgpu,
1852                 intel_gvt_gtt_type_t root_entry_type, u64 pdps[])
1853 {
1854         struct intel_gvt *gvt = vgpu->gvt;
1855         struct intel_vgpu_mm *mm;
1856         int ret;
1857
1858         mm = vgpu_alloc_mm(vgpu);
1859         if (!mm)
1860                 return ERR_PTR(-ENOMEM);
1861
1862         mm->type = INTEL_GVT_MM_PPGTT;
1863
1864         GEM_BUG_ON(root_entry_type != GTT_TYPE_PPGTT_ROOT_L3_ENTRY &&
1865                    root_entry_type != GTT_TYPE_PPGTT_ROOT_L4_ENTRY);
1866         mm->ppgtt_mm.root_entry_type = root_entry_type;
1867
1868         INIT_LIST_HEAD(&mm->ppgtt_mm.list);
1869         INIT_LIST_HEAD(&mm->ppgtt_mm.lru_list);
1870
1871         if (root_entry_type == GTT_TYPE_PPGTT_ROOT_L4_ENTRY)
1872                 mm->ppgtt_mm.guest_pdps[0] = pdps[0];
1873         else
1874                 memcpy(mm->ppgtt_mm.guest_pdps, pdps,
1875                        sizeof(mm->ppgtt_mm.guest_pdps));
1876
1877         ret = shadow_ppgtt_mm(mm);
1878         if (ret) {
1879                 gvt_vgpu_err("failed to shadow ppgtt mm\n");
1880                 vgpu_free_mm(mm);
1881                 return ERR_PTR(ret);
1882         }
1883
1884         list_add_tail(&mm->ppgtt_mm.list, &vgpu->gtt.ppgtt_mm_list_head);
1885         list_add_tail(&mm->ppgtt_mm.lru_list, &gvt->gtt.ppgtt_mm_lru_list_head);
1886         return mm;
1887 }
1888
1889 static struct intel_vgpu_mm *intel_vgpu_create_ggtt_mm(struct intel_vgpu *vgpu)
1890 {
1891         struct intel_vgpu_mm *mm;
1892         unsigned long nr_entries;
1893
1894         mm = vgpu_alloc_mm(vgpu);
1895         if (!mm)
1896                 return ERR_PTR(-ENOMEM);
1897
1898         mm->type = INTEL_GVT_MM_GGTT;
1899
1900         nr_entries = gvt_ggtt_gm_sz(vgpu->gvt) >> I915_GTT_PAGE_SHIFT;
1901         mm->ggtt_mm.virtual_ggtt =
1902                 vzalloc(array_size(nr_entries,
1903                                    vgpu->gvt->device_info.gtt_entry_size));
1904         if (!mm->ggtt_mm.virtual_ggtt) {
1905                 vgpu_free_mm(mm);
1906                 return ERR_PTR(-ENOMEM);
1907         }
1908
1909         return mm;
1910 }
1911
1912 /**
1913  * _intel_vgpu_mm_release - destroy a mm object
1914  * @mm_ref: a kref object
1915  *
1916  * This function is used to destroy a mm object for vGPU
1917  *
1918  */
1919 void _intel_vgpu_mm_release(struct kref *mm_ref)
1920 {
1921         struct intel_vgpu_mm *mm = container_of(mm_ref, typeof(*mm), ref);
1922
1923         if (GEM_WARN_ON(atomic_read(&mm->pincount)))
1924                 gvt_err("vgpu mm pin count bug detected\n");
1925
1926         if (mm->type == INTEL_GVT_MM_PPGTT) {
1927                 list_del(&mm->ppgtt_mm.list);
1928                 list_del(&mm->ppgtt_mm.lru_list);
1929                 invalidate_ppgtt_mm(mm);
1930         } else {
1931                 vfree(mm->ggtt_mm.virtual_ggtt);
1932         }
1933
1934         vgpu_free_mm(mm);
1935 }
1936
1937 /**
1938  * intel_vgpu_unpin_mm - decrease the pin count of a vGPU mm object
1939  * @mm: a vGPU mm object
1940  *
1941  * This function is called when user doesn't want to use a vGPU mm object
1942  */
1943 void intel_vgpu_unpin_mm(struct intel_vgpu_mm *mm)
1944 {
1945         atomic_dec(&mm->pincount);
1946 }
1947
1948 /**
1949  * intel_vgpu_pin_mm - increase the pin count of a vGPU mm object
1950  * @mm: target vgpu mm
1951  *
1952  * This function is called when user wants to use a vGPU mm object. If this
1953  * mm object hasn't been shadowed yet, the shadow will be populated at this
1954  * time.
1955  *
1956  * Returns:
1957  * Zero on success, negative error code if failed.
1958  */
1959 int intel_vgpu_pin_mm(struct intel_vgpu_mm *mm)
1960 {
1961         int ret;
1962
1963         atomic_inc(&mm->pincount);
1964
1965         if (mm->type == INTEL_GVT_MM_PPGTT) {
1966                 ret = shadow_ppgtt_mm(mm);
1967                 if (ret)
1968                         return ret;
1969
1970                 list_move_tail(&mm->ppgtt_mm.lru_list,
1971                                &mm->vgpu->gvt->gtt.ppgtt_mm_lru_list_head);
1972
1973         }
1974
1975         return 0;
1976 }
1977
1978 static int reclaim_one_ppgtt_mm(struct intel_gvt *gvt)
1979 {
1980         struct intel_vgpu_mm *mm;
1981         struct list_head *pos, *n;
1982
1983         list_for_each_safe(pos, n, &gvt->gtt.ppgtt_mm_lru_list_head) {
1984                 mm = container_of(pos, struct intel_vgpu_mm, ppgtt_mm.lru_list);
1985
1986                 if (atomic_read(&mm->pincount))
1987                         continue;
1988
1989                 list_del_init(&mm->ppgtt_mm.lru_list);
1990                 invalidate_ppgtt_mm(mm);
1991                 return 1;
1992         }
1993         return 0;
1994 }
1995
1996 /*
1997  * GMA translation APIs.
1998  */
1999 static inline int ppgtt_get_next_level_entry(struct intel_vgpu_mm *mm,
2000                 struct intel_gvt_gtt_entry *e, unsigned long index, bool guest)
2001 {
2002         struct intel_vgpu *vgpu = mm->vgpu;
2003         struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
2004         struct intel_vgpu_ppgtt_spt *s;
2005
2006         s = intel_vgpu_find_spt_by_mfn(vgpu, ops->get_pfn(e));
2007         if (!s)
2008                 return -ENXIO;
2009
2010         if (!guest)
2011                 ppgtt_get_shadow_entry(s, e, index);
2012         else
2013                 ppgtt_get_guest_entry(s, e, index);
2014         return 0;
2015 }
2016
2017 /**
2018  * intel_vgpu_gma_to_gpa - translate a gma to GPA
2019  * @mm: mm object. could be a PPGTT or GGTT mm object
2020  * @gma: graphics memory address in this mm object
2021  *
2022  * This function is used to translate a graphics memory address in specific
2023  * graphics memory space to guest physical address.
2024  *
2025  * Returns:
2026  * Guest physical address on success, INTEL_GVT_INVALID_ADDR if failed.
2027  */
2028 unsigned long intel_vgpu_gma_to_gpa(struct intel_vgpu_mm *mm, unsigned long gma)
2029 {
2030         struct intel_vgpu *vgpu = mm->vgpu;
2031         struct intel_gvt *gvt = vgpu->gvt;
2032         struct intel_gvt_gtt_pte_ops *pte_ops = gvt->gtt.pte_ops;
2033         struct intel_gvt_gtt_gma_ops *gma_ops = gvt->gtt.gma_ops;
2034         unsigned long gpa = INTEL_GVT_INVALID_ADDR;
2035         unsigned long gma_index[4];
2036         struct intel_gvt_gtt_entry e;
2037         int i, levels = 0;
2038         int ret;
2039
2040         GEM_BUG_ON(mm->type != INTEL_GVT_MM_GGTT &&
2041                    mm->type != INTEL_GVT_MM_PPGTT);
2042
2043         if (mm->type == INTEL_GVT_MM_GGTT) {
2044                 if (!vgpu_gmadr_is_valid(vgpu, gma))
2045                         goto err;
2046
2047                 ggtt_get_guest_entry(mm, &e,
2048                         gma_ops->gma_to_ggtt_pte_index(gma));
2049
2050                 gpa = (pte_ops->get_pfn(&e) << I915_GTT_PAGE_SHIFT)
2051                         + (gma & ~I915_GTT_PAGE_MASK);
2052
2053                 trace_gma_translate(vgpu->id, "ggtt", 0, 0, gma, gpa);
2054         } else {
2055                 switch (mm->ppgtt_mm.root_entry_type) {
2056                 case GTT_TYPE_PPGTT_ROOT_L4_ENTRY:
2057                         ppgtt_get_shadow_root_entry(mm, &e, 0);
2058
2059                         gma_index[0] = gma_ops->gma_to_pml4_index(gma);
2060                         gma_index[1] = gma_ops->gma_to_l4_pdp_index(gma);
2061                         gma_index[2] = gma_ops->gma_to_pde_index(gma);
2062                         gma_index[3] = gma_ops->gma_to_pte_index(gma);
2063                         levels = 4;
2064                         break;
2065                 case GTT_TYPE_PPGTT_ROOT_L3_ENTRY:
2066                         ppgtt_get_shadow_root_entry(mm, &e,
2067                                         gma_ops->gma_to_l3_pdp_index(gma));
2068
2069                         gma_index[0] = gma_ops->gma_to_pde_index(gma);
2070                         gma_index[1] = gma_ops->gma_to_pte_index(gma);
2071                         levels = 2;
2072                         break;
2073                 default:
2074                         GEM_BUG_ON(1);
2075                 }
2076
2077                 /* walk the shadow page table and get gpa from guest entry */
2078                 for (i = 0; i < levels; i++) {
2079                         ret = ppgtt_get_next_level_entry(mm, &e, gma_index[i],
2080                                 (i == levels - 1));
2081                         if (ret)
2082                                 goto err;
2083
2084                         if (!pte_ops->test_present(&e)) {
2085                                 gvt_dbg_core("GMA 0x%lx is not present\n", gma);
2086                                 goto err;
2087                         }
2088                 }
2089
2090                 gpa = (pte_ops->get_pfn(&e) << I915_GTT_PAGE_SHIFT) +
2091                                         (gma & ~I915_GTT_PAGE_MASK);
2092                 trace_gma_translate(vgpu->id, "ppgtt", 0,
2093                                     mm->ppgtt_mm.root_entry_type, gma, gpa);
2094         }
2095
2096         return gpa;
2097 err:
2098         gvt_vgpu_err("invalid mm type: %d gma %lx\n", mm->type, gma);
2099         return INTEL_GVT_INVALID_ADDR;
2100 }
2101
2102 static int emulate_ggtt_mmio_read(struct intel_vgpu *vgpu,
2103         unsigned int off, void *p_data, unsigned int bytes)
2104 {
2105         struct intel_vgpu_mm *ggtt_mm = vgpu->gtt.ggtt_mm;
2106         const struct intel_gvt_device_info *info = &vgpu->gvt->device_info;
2107         unsigned long index = off >> info->gtt_entry_size_shift;
2108         struct intel_gvt_gtt_entry e;
2109
2110         if (bytes != 4 && bytes != 8)
2111                 return -EINVAL;
2112
2113         ggtt_get_guest_entry(ggtt_mm, &e, index);
2114         memcpy(p_data, (void *)&e.val64 + (off & (info->gtt_entry_size - 1)),
2115                         bytes);
2116         return 0;
2117 }
2118
2119 /**
2120  * intel_vgpu_emulate_gtt_mmio_read - emulate GTT MMIO register read
2121  * @vgpu: a vGPU
2122  * @off: register offset
2123  * @p_data: data will be returned to guest
2124  * @bytes: data length
2125  *
2126  * This function is used to emulate the GTT MMIO register read
2127  *
2128  * Returns:
2129  * Zero on success, error code if failed.
2130  */
2131 int intel_vgpu_emulate_ggtt_mmio_read(struct intel_vgpu *vgpu, unsigned int off,
2132         void *p_data, unsigned int bytes)
2133 {
2134         const struct intel_gvt_device_info *info = &vgpu->gvt->device_info;
2135         int ret;
2136
2137         if (bytes != 4 && bytes != 8)
2138                 return -EINVAL;
2139
2140         off -= info->gtt_start_offset;
2141         ret = emulate_ggtt_mmio_read(vgpu, off, p_data, bytes);
2142         return ret;
2143 }
2144
2145 static void ggtt_invalidate_pte(struct intel_vgpu *vgpu,
2146                 struct intel_gvt_gtt_entry *entry)
2147 {
2148         struct intel_gvt_gtt_pte_ops *pte_ops = vgpu->gvt->gtt.pte_ops;
2149         unsigned long pfn;
2150
2151         pfn = pte_ops->get_pfn(entry);
2152         if (pfn != vgpu->gvt->gtt.scratch_mfn)
2153                 intel_gvt_hypervisor_dma_unmap_guest_page(vgpu,
2154                                                 pfn << PAGE_SHIFT);
2155 }
2156
2157 static int emulate_ggtt_mmio_write(struct intel_vgpu *vgpu, unsigned int off,
2158         void *p_data, unsigned int bytes)
2159 {
2160         struct intel_gvt *gvt = vgpu->gvt;
2161         const struct intel_gvt_device_info *info = &gvt->device_info;
2162         struct intel_vgpu_mm *ggtt_mm = vgpu->gtt.ggtt_mm;
2163         struct intel_gvt_gtt_pte_ops *ops = gvt->gtt.pte_ops;
2164         unsigned long g_gtt_index = off >> info->gtt_entry_size_shift;
2165         unsigned long gma, gfn;
2166         struct intel_gvt_gtt_entry e, m;
2167         dma_addr_t dma_addr;
2168         int ret;
2169         struct intel_gvt_partial_pte *partial_pte, *pos, *n;
2170         bool partial_update = false;
2171
2172         if (bytes != 4 && bytes != 8)
2173                 return -EINVAL;
2174
2175         gma = g_gtt_index << I915_GTT_PAGE_SHIFT;
2176
2177         /* the VM may configure the whole GM space when ballooning is used */
2178         if (!vgpu_gmadr_is_valid(vgpu, gma))
2179                 return 0;
2180
2181         e.type = GTT_TYPE_GGTT_PTE;
2182         memcpy((void *)&e.val64 + (off & (info->gtt_entry_size - 1)), p_data,
2183                         bytes);
2184
2185         /* If ggtt entry size is 8 bytes, and it's split into two 4 bytes
2186          * write, save the first 4 bytes in a list and update virtual
2187          * PTE. Only update shadow PTE when the second 4 bytes comes.
2188          */
2189         if (bytes < info->gtt_entry_size) {
2190                 bool found = false;
2191
2192                 list_for_each_entry_safe(pos, n,
2193                                 &ggtt_mm->ggtt_mm.partial_pte_list, list) {
2194                         if (g_gtt_index == pos->offset >>
2195                                         info->gtt_entry_size_shift) {
2196                                 if (off != pos->offset) {
2197                                         /* the second partial part*/
2198                                         int last_off = pos->offset &
2199                                                 (info->gtt_entry_size - 1);
2200
2201                                         memcpy((void *)&e.val64 + last_off,
2202                                                 (void *)&pos->data + last_off,
2203                                                 bytes);
2204
2205                                         list_del(&pos->list);
2206                                         kfree(pos);
2207                                         found = true;
2208                                         break;
2209                                 }
2210
2211                                 /* update of the first partial part */
2212                                 pos->data = e.val64;
2213                                 ggtt_set_guest_entry(ggtt_mm, &e, g_gtt_index);
2214                                 return 0;
2215                         }
2216                 }
2217
2218                 if (!found) {
2219                         /* the first partial part */
2220                         partial_pte = kzalloc(sizeof(*partial_pte), GFP_KERNEL);
2221                         if (!partial_pte)
2222                                 return -ENOMEM;
2223                         partial_pte->offset = off;
2224                         partial_pte->data = e.val64;
2225                         list_add_tail(&partial_pte->list,
2226                                 &ggtt_mm->ggtt_mm.partial_pte_list);
2227                         partial_update = true;
2228                 }
2229         }
2230
2231         if (!partial_update && (ops->test_present(&e))) {
2232                 gfn = ops->get_pfn(&e);
2233                 m = e;
2234
2235                 /* one PTE update may be issued in multiple writes and the
2236                  * first write may not construct a valid gfn
2237                  */
2238                 if (!intel_gvt_hypervisor_is_valid_gfn(vgpu, gfn)) {
2239                         ops->set_pfn(&m, gvt->gtt.scratch_mfn);
2240                         goto out;
2241                 }
2242
2243                 ret = intel_gvt_hypervisor_dma_map_guest_page(vgpu, gfn,
2244                                                         PAGE_SIZE, &dma_addr);
2245                 if (ret) {
2246                         gvt_vgpu_err("fail to populate guest ggtt entry\n");
2247                         /* guest driver may read/write the entry when partial
2248                          * update the entry in this situation p2m will fail
2249                          * settting the shadow entry to point to a scratch page
2250                          */
2251                         ops->set_pfn(&m, gvt->gtt.scratch_mfn);
2252                 } else
2253                         ops->set_pfn(&m, dma_addr >> PAGE_SHIFT);
2254         } else {
2255                 ops->set_pfn(&m, gvt->gtt.scratch_mfn);
2256                 ops->clear_present(&m);
2257         }
2258
2259 out:
2260         ggtt_set_guest_entry(ggtt_mm, &e, g_gtt_index);
2261
2262         ggtt_get_host_entry(ggtt_mm, &e, g_gtt_index);
2263         ggtt_invalidate_pte(vgpu, &e);
2264
2265         ggtt_set_host_entry(ggtt_mm, &m, g_gtt_index);
2266         ggtt_invalidate(gvt->dev_priv);
2267         return 0;
2268 }
2269
2270 /*
2271  * intel_vgpu_emulate_ggtt_mmio_write - emulate GTT MMIO register write
2272  * @vgpu: a vGPU
2273  * @off: register offset
2274  * @p_data: data from guest write
2275  * @bytes: data length
2276  *
2277  * This function is used to emulate the GTT MMIO register write
2278  *
2279  * Returns:
2280  * Zero on success, error code if failed.
2281  */
2282 int intel_vgpu_emulate_ggtt_mmio_write(struct intel_vgpu *vgpu,
2283                 unsigned int off, void *p_data, unsigned int bytes)
2284 {
2285         const struct intel_gvt_device_info *info = &vgpu->gvt->device_info;
2286         int ret;
2287
2288         if (bytes != 4 && bytes != 8)
2289                 return -EINVAL;
2290
2291         off -= info->gtt_start_offset;
2292         ret = emulate_ggtt_mmio_write(vgpu, off, p_data, bytes);
2293         return ret;
2294 }
2295
2296 static int alloc_scratch_pages(struct intel_vgpu *vgpu,
2297                 intel_gvt_gtt_type_t type)
2298 {
2299         struct intel_vgpu_gtt *gtt = &vgpu->gtt;
2300         struct intel_gvt_gtt_pte_ops *ops = vgpu->gvt->gtt.pte_ops;
2301         int page_entry_num = I915_GTT_PAGE_SIZE >>
2302                                 vgpu->gvt->device_info.gtt_entry_size_shift;
2303         void *scratch_pt;
2304         int i;
2305         struct device *dev = &vgpu->gvt->dev_priv->drm.pdev->dev;
2306         dma_addr_t daddr;
2307
2308         if (WARN_ON(type < GTT_TYPE_PPGTT_PTE_PT || type >= GTT_TYPE_MAX))
2309                 return -EINVAL;
2310
2311         scratch_pt = (void *)get_zeroed_page(GFP_KERNEL);
2312         if (!scratch_pt) {
2313                 gvt_vgpu_err("fail to allocate scratch page\n");
2314                 return -ENOMEM;
2315         }
2316
2317         daddr = dma_map_page(dev, virt_to_page(scratch_pt), 0,
2318                         4096, PCI_DMA_BIDIRECTIONAL);
2319         if (dma_mapping_error(dev, daddr)) {
2320                 gvt_vgpu_err("fail to dmamap scratch_pt\n");
2321                 __free_page(virt_to_page(scratch_pt));
2322                 return -ENOMEM;
2323         }
2324         gtt->scratch_pt[type].page_mfn =
2325                 (unsigned long)(daddr >> I915_GTT_PAGE_SHIFT);
2326         gtt->scratch_pt[type].page = virt_to_page(scratch_pt);
2327         gvt_dbg_mm("vgpu%d create scratch_pt: type %d mfn=0x%lx\n",
2328                         vgpu->id, type, gtt->scratch_pt[type].page_mfn);
2329
2330         /* Build the tree by full filled the scratch pt with the entries which
2331          * point to the next level scratch pt or scratch page. The
2332          * scratch_pt[type] indicate the scratch pt/scratch page used by the
2333          * 'type' pt.
2334          * e.g. scratch_pt[GTT_TYPE_PPGTT_PDE_PT] is used by
2335          * GTT_TYPE_PPGTT_PDE_PT level pt, that means this scratch_pt it self
2336          * is GTT_TYPE_PPGTT_PTE_PT, and full filled by scratch page mfn.
2337          */
2338         if (type > GTT_TYPE_PPGTT_PTE_PT) {
2339                 struct intel_gvt_gtt_entry se;
2340
2341                 memset(&se, 0, sizeof(struct intel_gvt_gtt_entry));
2342                 se.type = get_entry_type(type - 1);
2343                 ops->set_pfn(&se, gtt->scratch_pt[type - 1].page_mfn);
2344
2345                 /* The entry parameters like present/writeable/cache type
2346                  * set to the same as i915's scratch page tree.
2347                  */
2348                 se.val64 |= _PAGE_PRESENT | _PAGE_RW;
2349                 if (type == GTT_TYPE_PPGTT_PDE_PT)
2350                         se.val64 |= PPAT_CACHED;
2351
2352                 for (i = 0; i < page_entry_num; i++)
2353                         ops->set_entry(scratch_pt, &se, i, false, 0, vgpu);
2354         }
2355
2356         return 0;
2357 }
2358
2359 static int release_scratch_page_tree(struct intel_vgpu *vgpu)
2360 {
2361         int i;
2362         struct device *dev = &vgpu->gvt->dev_priv->drm.pdev->dev;
2363         dma_addr_t daddr;
2364
2365         for (i = GTT_TYPE_PPGTT_PTE_PT; i < GTT_TYPE_MAX; i++) {
2366                 if (vgpu->gtt.scratch_pt[i].page != NULL) {
2367                         daddr = (dma_addr_t)(vgpu->gtt.scratch_pt[i].page_mfn <<
2368                                         I915_GTT_PAGE_SHIFT);
2369                         dma_unmap_page(dev, daddr, 4096, PCI_DMA_BIDIRECTIONAL);
2370                         __free_page(vgpu->gtt.scratch_pt[i].page);
2371                         vgpu->gtt.scratch_pt[i].page = NULL;
2372                         vgpu->gtt.scratch_pt[i].page_mfn = 0;
2373                 }
2374         }
2375
2376         return 0;
2377 }
2378
2379 static int create_scratch_page_tree(struct intel_vgpu *vgpu)
2380 {
2381         int i, ret;
2382
2383         for (i = GTT_TYPE_PPGTT_PTE_PT; i < GTT_TYPE_MAX; i++) {
2384                 ret = alloc_scratch_pages(vgpu, i);
2385                 if (ret)
2386                         goto err;
2387         }
2388
2389         return 0;
2390
2391 err:
2392         release_scratch_page_tree(vgpu);
2393         return ret;
2394 }
2395
2396 /**
2397  * intel_vgpu_init_gtt - initialize per-vGPU graphics memory virulization
2398  * @vgpu: a vGPU
2399  *
2400  * This function is used to initialize per-vGPU graphics memory virtualization
2401  * components.
2402  *
2403  * Returns:
2404  * Zero on success, error code if failed.
2405  */
2406 int intel_vgpu_init_gtt(struct intel_vgpu *vgpu)
2407 {
2408         struct intel_vgpu_gtt *gtt = &vgpu->gtt;
2409
2410         INIT_RADIX_TREE(&gtt->spt_tree, GFP_KERNEL);
2411
2412         INIT_LIST_HEAD(&gtt->ppgtt_mm_list_head);
2413         INIT_LIST_HEAD(&gtt->oos_page_list_head);
2414         INIT_LIST_HEAD(&gtt->post_shadow_list_head);
2415
2416         gtt->ggtt_mm = intel_vgpu_create_ggtt_mm(vgpu);
2417         if (IS_ERR(gtt->ggtt_mm)) {
2418                 gvt_vgpu_err("fail to create mm for ggtt.\n");
2419                 return PTR_ERR(gtt->ggtt_mm);
2420         }
2421
2422         intel_vgpu_reset_ggtt(vgpu, false);
2423
2424         INIT_LIST_HEAD(&gtt->ggtt_mm->ggtt_mm.partial_pte_list);
2425
2426         return create_scratch_page_tree(vgpu);
2427 }
2428
2429 static void intel_vgpu_destroy_all_ppgtt_mm(struct intel_vgpu *vgpu)
2430 {
2431         struct list_head *pos, *n;
2432         struct intel_vgpu_mm *mm;
2433
2434         list_for_each_safe(pos, n, &vgpu->gtt.ppgtt_mm_list_head) {
2435                 mm = container_of(pos, struct intel_vgpu_mm, ppgtt_mm.list);
2436                 intel_vgpu_destroy_mm(mm);
2437         }
2438
2439         if (GEM_WARN_ON(!list_empty(&vgpu->gtt.ppgtt_mm_list_head)))
2440                 gvt_err("vgpu ppgtt mm is not fully destroyed\n");
2441
2442         if (GEM_WARN_ON(!radix_tree_empty(&vgpu->gtt.spt_tree))) {
2443                 gvt_err("Why we still has spt not freed?\n");
2444                 ppgtt_free_all_spt(vgpu);
2445         }
2446 }
2447
2448 static void intel_vgpu_destroy_ggtt_mm(struct intel_vgpu *vgpu)
2449 {
2450         struct intel_gvt_partial_pte *pos, *next;
2451
2452         list_for_each_entry_safe(pos, next,
2453                                  &vgpu->gtt.ggtt_mm->ggtt_mm.partial_pte_list,
2454                                  list) {
2455                 gvt_dbg_mm("partial PTE update on hold 0x%lx : 0x%llx\n",
2456                         pos->offset, pos->data);
2457                 kfree(pos);
2458         }
2459         intel_vgpu_destroy_mm(vgpu->gtt.ggtt_mm);
2460         vgpu->gtt.ggtt_mm = NULL;
2461 }
2462
2463 /**
2464  * intel_vgpu_clean_gtt - clean up per-vGPU graphics memory virulization
2465  * @vgpu: a vGPU
2466  *
2467  * This function is used to clean up per-vGPU graphics memory virtualization
2468  * components.
2469  *
2470  * Returns:
2471  * Zero on success, error code if failed.
2472  */
2473 void intel_vgpu_clean_gtt(struct intel_vgpu *vgpu)
2474 {
2475         intel_vgpu_destroy_all_ppgtt_mm(vgpu);
2476         intel_vgpu_destroy_ggtt_mm(vgpu);
2477         release_scratch_page_tree(vgpu);
2478 }
2479
2480 static void clean_spt_oos(struct intel_gvt *gvt)
2481 {
2482         struct intel_gvt_gtt *gtt = &gvt->gtt;
2483         struct list_head *pos, *n;
2484         struct intel_vgpu_oos_page *oos_page;
2485
2486         WARN(!list_empty(&gtt->oos_page_use_list_head),
2487                 "someone is still using oos page\n");
2488
2489         list_for_each_safe(pos, n, &gtt->oos_page_free_list_head) {
2490                 oos_page = container_of(pos, struct intel_vgpu_oos_page, list);
2491                 list_del(&oos_page->list);
2492                 kfree(oos_page);
2493         }
2494 }
2495
2496 static int setup_spt_oos(struct intel_gvt *gvt)
2497 {
2498         struct intel_gvt_gtt *gtt = &gvt->gtt;
2499         struct intel_vgpu_oos_page *oos_page;
2500         int i;
2501         int ret;
2502
2503         INIT_LIST_HEAD(&gtt->oos_page_free_list_head);
2504         INIT_LIST_HEAD(&gtt->oos_page_use_list_head);
2505
2506         for (i = 0; i < preallocated_oos_pages; i++) {
2507                 oos_page = kzalloc(sizeof(*oos_page), GFP_KERNEL);
2508                 if (!oos_page) {
2509                         ret = -ENOMEM;
2510                         goto fail;
2511                 }
2512
2513                 INIT_LIST_HEAD(&oos_page->list);
2514                 INIT_LIST_HEAD(&oos_page->vm_list);
2515                 oos_page->id = i;
2516                 list_add_tail(&oos_page->list, &gtt->oos_page_free_list_head);
2517         }
2518
2519         gvt_dbg_mm("%d oos pages preallocated\n", i);
2520
2521         return 0;
2522 fail:
2523         clean_spt_oos(gvt);
2524         return ret;
2525 }
2526
2527 /**
2528  * intel_vgpu_find_ppgtt_mm - find a PPGTT mm object
2529  * @vgpu: a vGPU
2530  * @pdps: pdp root array
2531  *
2532  * This function is used to find a PPGTT mm object from mm object pool
2533  *
2534  * Returns:
2535  * pointer to mm object on success, NULL if failed.
2536  */
2537 struct intel_vgpu_mm *intel_vgpu_find_ppgtt_mm(struct intel_vgpu *vgpu,
2538                 u64 pdps[])
2539 {
2540         struct intel_vgpu_mm *mm;
2541         struct list_head *pos;
2542
2543         list_for_each(pos, &vgpu->gtt.ppgtt_mm_list_head) {
2544                 mm = container_of(pos, struct intel_vgpu_mm, ppgtt_mm.list);
2545
2546                 switch (mm->ppgtt_mm.root_entry_type) {
2547                 case GTT_TYPE_PPGTT_ROOT_L4_ENTRY:
2548                         if (pdps[0] == mm->ppgtt_mm.guest_pdps[0])
2549                                 return mm;
2550                         break;
2551                 case GTT_TYPE_PPGTT_ROOT_L3_ENTRY:
2552                         if (!memcmp(pdps, mm->ppgtt_mm.guest_pdps,
2553                                     sizeof(mm->ppgtt_mm.guest_pdps)))
2554                                 return mm;
2555                         break;
2556                 default:
2557                         GEM_BUG_ON(1);
2558                 }
2559         }
2560         return NULL;
2561 }
2562
2563 /**
2564  * intel_vgpu_get_ppgtt_mm - get or create a PPGTT mm object.
2565  * @vgpu: a vGPU
2566  * @root_entry_type: ppgtt root entry type
2567  * @pdps: guest pdps
2568  *
2569  * This function is used to find or create a PPGTT mm object from a guest.
2570  *
2571  * Returns:
2572  * Zero on success, negative error code if failed.
2573  */
2574 struct intel_vgpu_mm *intel_vgpu_get_ppgtt_mm(struct intel_vgpu *vgpu,
2575                 intel_gvt_gtt_type_t root_entry_type, u64 pdps[])
2576 {
2577         struct intel_vgpu_mm *mm;
2578
2579         mm = intel_vgpu_find_ppgtt_mm(vgpu, pdps);
2580         if (mm) {
2581                 intel_vgpu_mm_get(mm);
2582         } else {
2583                 mm = intel_vgpu_create_ppgtt_mm(vgpu, root_entry_type, pdps);
2584                 if (IS_ERR(mm))
2585                         gvt_vgpu_err("fail to create mm\n");
2586         }
2587         return mm;
2588 }
2589
2590 /**
2591  * intel_vgpu_put_ppgtt_mm - find and put a PPGTT mm object.
2592  * @vgpu: a vGPU
2593  * @pdps: guest pdps
2594  *
2595  * This function is used to find a PPGTT mm object from a guest and destroy it.
2596  *
2597  * Returns:
2598  * Zero on success, negative error code if failed.
2599  */
2600 int intel_vgpu_put_ppgtt_mm(struct intel_vgpu *vgpu, u64 pdps[])
2601 {
2602         struct intel_vgpu_mm *mm;
2603
2604         mm = intel_vgpu_find_ppgtt_mm(vgpu, pdps);
2605         if (!mm) {
2606                 gvt_vgpu_err("fail to find ppgtt instance.\n");
2607                 return -EINVAL;
2608         }
2609         intel_vgpu_mm_put(mm);
2610         return 0;
2611 }
2612
2613 /**
2614  * intel_gvt_init_gtt - initialize mm components of a GVT device
2615  * @gvt: GVT device
2616  *
2617  * This function is called at the initialization stage, to initialize
2618  * the mm components of a GVT device.
2619  *
2620  * Returns:
2621  * zero on success, negative error code if failed.
2622  */
2623 int intel_gvt_init_gtt(struct intel_gvt *gvt)
2624 {
2625         int ret;
2626         void *page;
2627         struct device *dev = &gvt->dev_priv->drm.pdev->dev;
2628         dma_addr_t daddr;
2629
2630         gvt_dbg_core("init gtt\n");
2631
2632         gvt->gtt.pte_ops = &gen8_gtt_pte_ops;
2633         gvt->gtt.gma_ops = &gen8_gtt_gma_ops;
2634
2635         page = (void *)get_zeroed_page(GFP_KERNEL);
2636         if (!page) {
2637                 gvt_err("fail to allocate scratch ggtt page\n");
2638                 return -ENOMEM;
2639         }
2640
2641         daddr = dma_map_page(dev, virt_to_page(page), 0,
2642                         4096, PCI_DMA_BIDIRECTIONAL);
2643         if (dma_mapping_error(dev, daddr)) {
2644                 gvt_err("fail to dmamap scratch ggtt page\n");
2645                 __free_page(virt_to_page(page));
2646                 return -ENOMEM;
2647         }
2648
2649         gvt->gtt.scratch_page = virt_to_page(page);
2650         gvt->gtt.scratch_mfn = (unsigned long)(daddr >> I915_GTT_PAGE_SHIFT);
2651
2652         if (enable_out_of_sync) {
2653                 ret = setup_spt_oos(gvt);
2654                 if (ret) {
2655                         gvt_err("fail to initialize SPT oos\n");
2656                         dma_unmap_page(dev, daddr, 4096, PCI_DMA_BIDIRECTIONAL);
2657                         __free_page(gvt->gtt.scratch_page);
2658                         return ret;
2659                 }
2660         }
2661         INIT_LIST_HEAD(&gvt->gtt.ppgtt_mm_lru_list_head);
2662         return 0;
2663 }
2664
2665 /**
2666  * intel_gvt_clean_gtt - clean up mm components of a GVT device
2667  * @gvt: GVT device
2668  *
2669  * This function is called at the driver unloading stage, to clean up the
2670  * the mm components of a GVT device.
2671  *
2672  */
2673 void intel_gvt_clean_gtt(struct intel_gvt *gvt)
2674 {
2675         struct device *dev = &gvt->dev_priv->drm.pdev->dev;
2676         dma_addr_t daddr = (dma_addr_t)(gvt->gtt.scratch_mfn <<
2677                                         I915_GTT_PAGE_SHIFT);
2678
2679         dma_unmap_page(dev, daddr, 4096, PCI_DMA_BIDIRECTIONAL);
2680
2681         __free_page(gvt->gtt.scratch_page);
2682
2683         if (enable_out_of_sync)
2684                 clean_spt_oos(gvt);
2685 }
2686
2687 /**
2688  * intel_vgpu_invalidate_ppgtt - invalidate PPGTT instances
2689  * @vgpu: a vGPU
2690  *
2691  * This function is called when invalidate all PPGTT instances of a vGPU.
2692  *
2693  */
2694 void intel_vgpu_invalidate_ppgtt(struct intel_vgpu *vgpu)
2695 {
2696         struct list_head *pos, *n;
2697         struct intel_vgpu_mm *mm;
2698
2699         list_for_each_safe(pos, n, &vgpu->gtt.ppgtt_mm_list_head) {
2700                 mm = container_of(pos, struct intel_vgpu_mm, ppgtt_mm.list);
2701                 if (mm->type == INTEL_GVT_MM_PPGTT) {
2702                         list_del_init(&mm->ppgtt_mm.lru_list);
2703                         if (mm->ppgtt_mm.shadowed)
2704                                 invalidate_ppgtt_mm(mm);
2705                 }
2706         }
2707 }
2708
2709 /**
2710  * intel_vgpu_reset_ggtt - reset the GGTT entry
2711  * @vgpu: a vGPU
2712  * @invalidate_old: invalidate old entries
2713  *
2714  * This function is called at the vGPU create stage
2715  * to reset all the GGTT entries.
2716  *
2717  */
2718 void intel_vgpu_reset_ggtt(struct intel_vgpu *vgpu, bool invalidate_old)
2719 {
2720         struct intel_gvt *gvt = vgpu->gvt;
2721         struct drm_i915_private *dev_priv = gvt->dev_priv;
2722         struct intel_gvt_gtt_pte_ops *pte_ops = vgpu->gvt->gtt.pte_ops;
2723         struct intel_gvt_gtt_entry entry = {.type = GTT_TYPE_GGTT_PTE};
2724         struct intel_gvt_gtt_entry old_entry;
2725         u32 index;
2726         u32 num_entries;
2727
2728         pte_ops->set_pfn(&entry, gvt->gtt.scratch_mfn);
2729         pte_ops->set_present(&entry);
2730
2731         index = vgpu_aperture_gmadr_base(vgpu) >> PAGE_SHIFT;
2732         num_entries = vgpu_aperture_sz(vgpu) >> PAGE_SHIFT;
2733         while (num_entries--) {
2734                 if (invalidate_old) {
2735                         ggtt_get_host_entry(vgpu->gtt.ggtt_mm, &old_entry, index);
2736                         ggtt_invalidate_pte(vgpu, &old_entry);
2737                 }
2738                 ggtt_set_host_entry(vgpu->gtt.ggtt_mm, &entry, index++);
2739         }
2740
2741         index = vgpu_hidden_gmadr_base(vgpu) >> PAGE_SHIFT;
2742         num_entries = vgpu_hidden_sz(vgpu) >> PAGE_SHIFT;
2743         while (num_entries--) {
2744                 if (invalidate_old) {
2745                         ggtt_get_host_entry(vgpu->gtt.ggtt_mm, &old_entry, index);
2746                         ggtt_invalidate_pte(vgpu, &old_entry);
2747                 }
2748                 ggtt_set_host_entry(vgpu->gtt.ggtt_mm, &entry, index++);
2749         }
2750
2751         ggtt_invalidate(dev_priv);
2752 }
2753
2754 /**
2755  * intel_vgpu_reset_gtt - reset the all GTT related status
2756  * @vgpu: a vGPU
2757  *
2758  * This function is called from vfio core to reset reset all
2759  * GTT related status, including GGTT, PPGTT, scratch page.
2760  *
2761  */
2762 void intel_vgpu_reset_gtt(struct intel_vgpu *vgpu)
2763 {
2764         /* Shadow pages are only created when there is no page
2765          * table tracking data, so remove page tracking data after
2766          * removing the shadow pages.
2767          */
2768         intel_vgpu_destroy_all_ppgtt_mm(vgpu);
2769         intel_vgpu_reset_ggtt(vgpu, true);
2770 }