arch/arm/kvm/mmu.c

   1 /*
   2  * Copyright (C) 2012 - Virtual Open Systems and Columbia University
   3  * Author: Christoffer Dall <[email protected]>
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License, version 2, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write to the Free Software
  16  * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
  17  */
  18
  19 #include <linux/mman.h>
  20 #include <linux/kvm_host.h>
  21 #include <linux/io.h>
  22 #include <linux/hugetlb.h>
  23 #include <trace/events/kvm.h>
  24 #include <asm/pgalloc.h>
  25 #include <asm/cacheflush.h>
  26 #include <asm/kvm_arm.h>
  27 #include <asm/kvm_mmu.h>
  28 #include <asm/kvm_mmio.h>
  29 #include <asm/kvm_asm.h>
  30 #include <asm/kvm_emulate.h>
  31
  32 #include "trace.h"
  33
  34 extern char  __hyp_idmap_text_start[], __hyp_idmap_text_end[];
  35
  36 static pgd_t *boot_hyp_pgd;
  37 static pgd_t *hyp_pgd;
  38 static DEFINE_MUTEX(kvm_hyp_pgd_mutex);
  39
  40 static void *init_bounce_page;
  41 static unsigned long hyp_idmap_start;
  42 static unsigned long hyp_idmap_end;
  43 static phys_addr_t hyp_idmap_vector;
  44
  45 #define pgd_order get_order(PTRS_PER_PGD * sizeof(pgd_t))
  46
  47 #define kvm_pmd_huge(_x)        (pmd_huge(_x) || pmd_trans_huge(_x))
  48
  49 static void kvm_tlb_flush_vmid_ipa(struct kvm *kvm, phys_addr_t ipa)
  50 {
  51         /*
  52          * This function also gets called when dealing with HYP page
  53          * tables. As HYP doesn't have an associated struct kvm (and
  54          * the HYP page tables are fairly static), we don't do
  55          * anything there.
  56          */
  57         if (kvm)
  58                 kvm_call_hyp(__kvm_tlb_flush_vmid_ipa, kvm, ipa);
  59 }
  60
  61 static int mmu_topup_memory_cache(struct kvm_mmu_memory_cache *cache,
  62                                   int min, int max)
  63 {
  64         void *page;
  65
  66         BUG_ON(max > KVM_NR_MEM_OBJS);
  67         if (cache->nobjs >= min)
  68                 return 0;
  69         while (cache->nobjs < max) {
  70                 page = (void *)__get_free_page(PGALLOC_GFP);
  71                 if (!page)
  72                         return -ENOMEM;
  73                 cache->objects[cache->nobjs++] = page;
  74         }
  75         return 0;
  76 }
  77
  78 static void mmu_free_memory_cache(struct kvm_mmu_memory_cache *mc)
  79 {
  80         while (mc->nobjs)
  81                 free_page((unsigned long)mc->objects[--mc->nobjs]);
  82 }
  83
  84 static void *mmu_memory_cache_alloc(struct kvm_mmu_memory_cache *mc)
  85 {
  86         void *p;
  87
  88         BUG_ON(!mc || !mc->nobjs);
  89         p = mc->objects[--mc->nobjs];
  90         return p;
  91 }
  92
  93 static void clear_pgd_entry(struct kvm *kvm, pgd_t *pgd, phys_addr_t addr)
  94 {
  95         pud_t *pud_table __maybe_unused = pud_offset(pgd, 0);
  96         pgd_clear(pgd);
  97         kvm_tlb_flush_vmid_ipa(kvm, addr);
  98         pud_free(NULL, pud_table);
  99         put_page(virt_to_page(pgd));
 100 }
 101
 102 static void clear_pud_entry(struct kvm *kvm, pud_t *pud, phys_addr_t addr)
 103 {
 104         pmd_t *pmd_table = pmd_offset(pud, 0);
 105         VM_BUG_ON(pud_huge(*pud));
 106         pud_clear(pud);
 107         kvm_tlb_flush_vmid_ipa(kvm, addr);
 108         pmd_free(NULL, pmd_table);
 109         put_page(virt_to_page(pud));
 110 }
 111
 112 static void clear_pmd_entry(struct kvm *kvm, pmd_t *pmd, phys_addr_t addr)
 113 {
 114         pte_t *pte_table = pte_offset_kernel(pmd, 0);
 115         VM_BUG_ON(kvm_pmd_huge(*pmd));
 116         pmd_clear(pmd);
 117         kvm_tlb_flush_vmid_ipa(kvm, addr);
 118         pte_free_kernel(NULL, pte_table);
 119         put_page(virt_to_page(pmd));
 120 }
 121
 122 static void unmap_ptes(struct kvm *kvm, pmd_t *pmd,
 123                        phys_addr_t addr, phys_addr_t end)
 124 {
 125         phys_addr_t start_addr = addr;
 126         pte_t *pte, *start_pte;
 127
 128         start_pte = pte = pte_offset_kernel(pmd, addr);
 129         do {
 130                 if (!pte_none(*pte)) {
 131                         kvm_set_pte(pte, __pte(0));
 132                         put_page(virt_to_page(pte));
 133                         kvm_tlb_flush_vmid_ipa(kvm, addr);
 134                 }
 135         } while (pte++, addr += PAGE_SIZE, addr != end);
 136
 137         if (kvm_pte_table_empty(start_pte))
 138                 clear_pmd_entry(kvm, pmd, start_addr);
 139 }
 140
 141 static void unmap_pmds(struct kvm *kvm, pud_t *pud,
 142                        phys_addr_t addr, phys_addr_t end)
 143 {
 144         phys_addr_t next, start_addr = addr;
 145         pmd_t *pmd, *start_pmd;
 146
 147         start_pmd = pmd = pmd_offset(pud, addr);
 148         do {
 149                 next = kvm_pmd_addr_end(addr, end);
 150                 if (!pmd_none(*pmd)) {
 151                         if (kvm_pmd_huge(*pmd)) {
 152                                 pmd_clear(pmd);
 153                                 kvm_tlb_flush_vmid_ipa(kvm, addr);
 154                                 put_page(virt_to_page(pmd));
 155                         } else {
 156                                 unmap_ptes(kvm, pmd, addr, next);
 157                         }
 158                 }
 159         } while (pmd++, addr = next, addr != end);
 160
 161         if (kvm_pmd_table_empty(start_pmd))
 162                 clear_pud_entry(kvm, pud, start_addr);
 163 }
 164
 165 static void unmap_puds(struct kvm *kvm, pgd_t *pgd,
 166                        phys_addr_t addr, phys_addr_t end)
 167 {
 168         phys_addr_t next, start_addr = addr;
 169         pud_t *pud, *start_pud;
 170
 171         start_pud = pud = pud_offset(pgd, addr);
 172         do {
 173                 next = kvm_pud_addr_end(addr, end);
 174                 if (!pud_none(*pud)) {
 175                         if (pud_huge(*pud)) {
 176                                 pud_clear(pud);
 177                                 kvm_tlb_flush_vmid_ipa(kvm, addr);
 178                                 put_page(virt_to_page(pud));
 179                         } else {
 180                                 unmap_pmds(kvm, pud, addr, next);
 181                         }
 182                 }
 183         } while (pud++, addr = next, addr != end);
 184
 185         if (kvm_pud_table_empty(start_pud))
 186                 clear_pgd_entry(kvm, pgd, start_addr);
 187 }
 188
 189
 190 static void unmap_range(struct kvm *kvm, pgd_t *pgdp,
 191                         phys_addr_t start, u64 size)
 192 {
 193         pgd_t *pgd;
 194         phys_addr_t addr = start, end = start + size;
 195         phys_addr_t next;
 196
 197         pgd = pgdp + pgd_index(addr);
 198         do {
 199                 next = kvm_pgd_addr_end(addr, end);
 200                 unmap_puds(kvm, pgd, addr, next);
 201         } while (pgd++, addr = next, addr != end);
 202 }
 203
 204 static void stage2_flush_ptes(struct kvm *kvm, pmd_t *pmd,
 205                               phys_addr_t addr, phys_addr_t end)
 206 {
 207         pte_t *pte;
 208
 209         pte = pte_offset_kernel(pmd, addr);
 210         do {
 211                 if (!pte_none(*pte)) {
 212                         hva_t hva = gfn_to_hva(kvm, addr >> PAGE_SHIFT);
 213                         kvm_flush_dcache_to_poc((void*)hva, PAGE_SIZE);
 214                 }
 215         } while (pte++, addr += PAGE_SIZE, addr != end);
 216 }
 217
 218 static void stage2_flush_pmds(struct kvm *kvm, pud_t *pud,
 219                               phys_addr_t addr, phys_addr_t end)
 220 {
 221         pmd_t *pmd;
 222         phys_addr_t next;
 223
 224         pmd = pmd_offset(pud, addr);
 225         do {
 226                 next = kvm_pmd_addr_end(addr, end);
 227                 if (!pmd_none(*pmd)) {
 228                         if (kvm_pmd_huge(*pmd)) {
 229                                 hva_t hva = gfn_to_hva(kvm, addr >> PAGE_SHIFT);
 230                                 kvm_flush_dcache_to_poc((void*)hva, PMD_SIZE);
 231                         } else {
 232                                 stage2_flush_ptes(kvm, pmd, addr, next);
 233                         }
 234                 }
 235         } while (pmd++, addr = next, addr != end);
 236 }
 237
 238 static void stage2_flush_puds(struct kvm *kvm, pgd_t *pgd,
 239                               phys_addr_t addr, phys_addr_t end)
 240 {
 241         pud_t *pud;
 242         phys_addr_t next;
 243
 244         pud = pud_offset(pgd, addr);
 245         do {
 246                 next = kvm_pud_addr_end(addr, end);
 247                 if (!pud_none(*pud)) {
 248                         if (pud_huge(*pud)) {
 249                                 hva_t hva = gfn_to_hva(kvm, addr >> PAGE_SHIFT);
 250                                 kvm_flush_dcache_to_poc((void*)hva, PUD_SIZE);
 251                         } else {
 252                                 stage2_flush_pmds(kvm, pud, addr, next);
 253                         }
 254                 }
 255         } while (pud++, addr = next, addr != end);
 256 }
 257
 258 static void stage2_flush_memslot(struct kvm *kvm,
 259                                  struct kvm_memory_slot *memslot)
 260 {
 261         phys_addr_t addr = memslot->base_gfn << PAGE_SHIFT;
 262         phys_addr_t end = addr + PAGE_SIZE * memslot->npages;
 263         phys_addr_t next;
 264         pgd_t *pgd;
 265
 266         pgd = kvm->arch.pgd + pgd_index(addr);
 267         do {
 268                 next = kvm_pgd_addr_end(addr, end);
 269                 stage2_flush_puds(kvm, pgd, addr, next);
 270         } while (pgd++, addr = next, addr != end);
 271 }
 272
 273 /**
 274  * stage2_flush_vm - Invalidate cache for pages mapped in stage 2
 275  * @kvm: The struct kvm pointer
 276  *
 277  * Go through the stage 2 page tables and invalidate any cache lines
 278  * backing memory already mapped to the VM.
 279  */
 280 void stage2_flush_vm(struct kvm *kvm)
 281 {
 282         struct kvm_memslots *slots;
 283         struct kvm_memory_slot *memslot;
 284         int idx;
 285
 286         idx = srcu_read_lock(&kvm->srcu);
 287         spin_lock(&kvm->mmu_lock);
 288
 289         slots = kvm_memslots(kvm);
 290         kvm_for_each_memslot(memslot, slots)
 291                 stage2_flush_memslot(kvm, memslot);
 292
 293         spin_unlock(&kvm->mmu_lock);
 294         srcu_read_unlock(&kvm->srcu, idx);
 295 }
 296
 297 /**
 298  * free_boot_hyp_pgd - free HYP boot page tables
 299  *
 300  * Free the HYP boot page tables. The bounce page is also freed.
 301  */
 302 void free_boot_hyp_pgd(void)
 303 {
 304         mutex_lock(&kvm_hyp_pgd_mutex);
 305
 306         if (boot_hyp_pgd) {
 307                 unmap_range(NULL, boot_hyp_pgd, hyp_idmap_start, PAGE_SIZE);
 308                 unmap_range(NULL, boot_hyp_pgd, TRAMPOLINE_VA, PAGE_SIZE);
 309                 free_pages((unsigned long)boot_hyp_pgd, pgd_order);
 310                 boot_hyp_pgd = NULL;
 311         }
 312
 313         if (hyp_pgd)
 314                 unmap_range(NULL, hyp_pgd, TRAMPOLINE_VA, PAGE_SIZE);
 315
 316         free_page((unsigned long)init_bounce_page);
 317         init_bounce_page = NULL;
 318
 319         mutex_unlock(&kvm_hyp_pgd_mutex);
 320 }
 321
 322 /**
 323  * free_hyp_pgds - free Hyp-mode page tables
 324  *
 325  * Assumes hyp_pgd is a page table used strictly in Hyp-mode and
 326  * therefore contains either mappings in the kernel memory area (above
 327  * PAGE_OFFSET), or device mappings in the vmalloc range (from
 328  * VMALLOC_START to VMALLOC_END).
 329  *
 330  * boot_hyp_pgd should only map two pages for the init code.
 331  */
 332 void free_hyp_pgds(void)
 333 {
 334         unsigned long addr;
 335
 336         free_boot_hyp_pgd();
 337
 338         mutex_lock(&kvm_hyp_pgd_mutex);
 339
 340         if (hyp_pgd) {
 341                 for (addr = PAGE_OFFSET; virt_addr_valid(addr); addr += PGDIR_SIZE)
 342                         unmap_range(NULL, hyp_pgd, KERN_TO_HYP(addr), PGDIR_SIZE);
 343                 for (addr = VMALLOC_START; is_vmalloc_addr((void*)addr); addr += PGDIR_SIZE)
 344                         unmap_range(NULL, hyp_pgd, KERN_TO_HYP(addr), PGDIR_SIZE);
 345
 346                 free_pages((unsigned long)hyp_pgd, pgd_order);
 347                 hyp_pgd = NULL;
 348         }
 349
 350         mutex_unlock(&kvm_hyp_pgd_mutex);
 351 }
 352
 353 static void create_hyp_pte_mappings(pmd_t *pmd, unsigned long start,
 354                                     unsigned long end, unsigned long pfn,
 355                                     pgprot_t prot)
 356 {
 357         pte_t *pte;
 358         unsigned long addr;
 359
 360         addr = start;
 361         do {
 362                 pte = pte_offset_kernel(pmd, addr);
 363                 kvm_set_pte(pte, pfn_pte(pfn, prot));
 364                 get_page(virt_to_page(pte));
 365                 kvm_flush_dcache_to_poc(pte, sizeof(*pte));
 366                 pfn++;
 367         } while (addr += PAGE_SIZE, addr != end);
 368 }
 369
 370 static int create_hyp_pmd_mappings(pud_t *pud, unsigned long start,
 371                                    unsigned long end, unsigned long pfn,
 372                                    pgprot_t prot)
 373 {
 374         pmd_t *pmd;
 375         pte_t *pte;
 376         unsigned long addr, next;
 377
 378         addr = start;
 379         do {
 380                 pmd = pmd_offset(pud, addr);
 381
 382                 BUG_ON(pmd_sect(*pmd));
 383
 384                 if (pmd_none(*pmd)) {
 385                         pte = pte_alloc_one_kernel(NULL, addr);
 386                         if (!pte) {
 387                                 kvm_err("Cannot allocate Hyp pte\n");
 388                                 return -ENOMEM;
 389                         }
 390                         pmd_populate_kernel(NULL, pmd, pte);
 391                         get_page(virt_to_page(pmd));
 392                         kvm_flush_dcache_to_poc(pmd, sizeof(*pmd));
 393                 }
 394
 395                 next = pmd_addr_end(addr, end);
 396
 397                 create_hyp_pte_mappings(pmd, addr, next, pfn, prot);
 398                 pfn += (next - addr) >> PAGE_SHIFT;
 399         } while (addr = next, addr != end);
 400
 401         return 0;
 402 }
 403
 404 static int __create_hyp_mappings(pgd_t *pgdp,
 405                                  unsigned long start, unsigned long end,
 406                                  unsigned long pfn, pgprot_t prot)
 407 {
 408         pgd_t *pgd;
 409         pud_t *pud;
 410         pmd_t *pmd;
 411         unsigned long addr, next;
 412         int err = 0;
 413
 414         mutex_lock(&kvm_hyp_pgd_mutex);
 415         addr = start & PAGE_MASK;
 416         end = PAGE_ALIGN(end);
 417         do {
 418                 pgd = pgdp + pgd_index(addr);
 419                 pud = pud_offset(pgd, addr);
 420
 421                 if (pud_none_or_clear_bad(pud)) {
 422                         pmd = pmd_alloc_one(NULL, addr);
 423                         if (!pmd) {
 424                                 kvm_err("Cannot allocate Hyp pmd\n");
 425                                 err = -ENOMEM;
 426                                 goto out;
 427                         }
 428                         pud_populate(NULL, pud, pmd);
 429                         get_page(virt_to_page(pud));
 430                         kvm_flush_dcache_to_poc(pud, sizeof(*pud));
 431                 }
 432
 433                 next = pgd_addr_end(addr, end);
 434                 err = create_hyp_pmd_mappings(pud, addr, next, pfn, prot);
 435                 if (err)
 436                         goto out;
 437                 pfn += (next - addr) >> PAGE_SHIFT;
 438         } while (addr = next, addr != end);
 439 out:
 440         mutex_unlock(&kvm_hyp_pgd_mutex);
 441         return err;
 442 }
 443
 444 static phys_addr_t kvm_kaddr_to_phys(void *kaddr)
 445 {
 446         if (!is_vmalloc_addr(kaddr)) {
 447                 BUG_ON(!virt_addr_valid(kaddr));
 448                 return __pa(kaddr);
 449         } else {
 450                 return page_to_phys(vmalloc_to_page(kaddr)) +
 451                        offset_in_page(kaddr);
 452         }
 453 }
 454
 455 /**
 456  * create_hyp_mappings - duplicate a kernel virtual address range in Hyp mode
 457  * @from:       The virtual kernel start address of the range
 458  * @to:         The virtual kernel end address of the range (exclusive)
 459  *
 460  * The same virtual address as the kernel virtual address is also used
 461  * in Hyp-mode mapping (modulo HYP_PAGE_OFFSET) to the same underlying
 462  * physical pages.
 463  */
 464 int create_hyp_mappings(void *from, void *to)
 465 {
 466         phys_addr_t phys_addr;
 467         unsigned long virt_addr;
 468         unsigned long start = KERN_TO_HYP((unsigned long)from);
 469         unsigned long end = KERN_TO_HYP((unsigned long)to);
 470
 471         start = start & PAGE_MASK;
 472         end = PAGE_ALIGN(end);
 473
 474         for (virt_addr = start; virt_addr < end; virt_addr += PAGE_SIZE) {
 475                 int err;
 476
 477                 phys_addr = kvm_kaddr_to_phys(from + virt_addr - start);
 478                 err = __create_hyp_mappings(hyp_pgd, virt_addr,
 479                                             virt_addr + PAGE_SIZE,
 480                                             __phys_to_pfn(phys_addr),
 481                                             PAGE_HYP);
 482                 if (err)
 483                         return err;
 484         }
 485
 486         return 0;
 487 }
 488
 489 /**
 490  * create_hyp_io_mappings - duplicate a kernel IO mapping into Hyp mode
 491  * @from:       The kernel start VA of the range
 492  * @to:         The kernel end VA of the range (exclusive)
 493  * @phys_addr:  The physical start address which gets mapped
 494  *
 495  * The resulting HYP VA is the same as the kernel VA, modulo
 496  * HYP_PAGE_OFFSET.
 497  */
 498 int create_hyp_io_mappings(void *from, void *to, phys_addr_t phys_addr)
 499 {
 500         unsigned long start = KERN_TO_HYP((unsigned long)from);
 501         unsigned long end = KERN_TO_HYP((unsigned long)to);
 502
 503         /* Check for a valid kernel IO mapping */
 504         if (!is_vmalloc_addr(from) || !is_vmalloc_addr(to - 1))
 505                 return -EINVAL;
 506
 507         return __create_hyp_mappings(hyp_pgd, start, end,
 508                                      __phys_to_pfn(phys_addr), PAGE_HYP_DEVICE);
 509 }
 510
 511 /**
 512  * kvm_alloc_stage2_pgd - allocate level-1 table for stage-2 translation.
 513  * @kvm:        The KVM struct pointer for the VM.
 514  *
 515  * Allocates the 1st level table only of size defined by S2_PGD_ORDER (can
 516  * support either full 40-bit input addresses or limited to 32-bit input
 517  * addresses). Clears the allocated pages.
 518  *
 519  * Note we don't need locking here as this is only called when the VM is
 520  * created, which can only be done once.
 521  */
 522 int kvm_alloc_stage2_pgd(struct kvm *kvm)
 523 {
 524         pgd_t *pgd;
 525
 526         if (kvm->arch.pgd != NULL) {
 527                 kvm_err("kvm_arch already initialized?\n");
 528                 return -EINVAL;
 529         }
 530
 531         pgd = (pgd_t *)__get_free_pages(GFP_KERNEL, S2_PGD_ORDER);
 532         if (!pgd)
 533                 return -ENOMEM;
 534
 535         memset(pgd, 0, PTRS_PER_S2_PGD * sizeof(pgd_t));
 536         kvm_clean_pgd(pgd);
 537         kvm->arch.pgd = pgd;
 538
 539         return 0;
 540 }
 541
 542 /**
 543  * unmap_stage2_range -- Clear stage2 page table entries to unmap a range
 544  * @kvm:   The VM pointer
 545  * @start: The intermediate physical base address of the range to unmap
 546  * @size:  The size of the area to unmap
 547  *
 548  * Clear a range of stage-2 mappings, lowering the various ref-counts.  Must
 549  * be called while holding mmu_lock (unless for freeing the stage2 pgd before
 550  * destroying the VM), otherwise another faulting VCPU may come in and mess
 551  * with things behind our backs.
 552  */
 553 static void unmap_stage2_range(struct kvm *kvm, phys_addr_t start, u64 size)
 554 {
 555         unmap_range(kvm, kvm->arch.pgd, start, size);
 556 }
 557
 558 /**
 559  * kvm_free_stage2_pgd - free all stage-2 tables
 560  * @kvm:        The KVM struct pointer for the VM.
 561  *
 562  * Walks the level-1 page table pointed to by kvm->arch.pgd and frees all
 563  * underlying level-2 and level-3 tables before freeing the actual level-1 table
 564  * and setting the struct pointer to NULL.
 565  *
 566  * Note we don't need locking here as this is only called when the VM is
 567  * destroyed, which can only be done once.
 568  */
 569 void kvm_free_stage2_pgd(struct kvm *kvm)
 570 {
 571         if (kvm->arch.pgd == NULL)
 572                 return;
 573
 574         unmap_stage2_range(kvm, 0, KVM_PHYS_SIZE);
 575         free_pages((unsigned long)kvm->arch.pgd, S2_PGD_ORDER);
 576         kvm->arch.pgd = NULL;
 577 }
 578
 579 static pmd_t *stage2_get_pmd(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
 580                              phys_addr_t addr)
 581 {
 582         pgd_t *pgd;
 583         pud_t *pud;
 584         pmd_t *pmd;
 585
 586         pgd = kvm->arch.pgd + pgd_index(addr);
 587         pud = pud_offset(pgd, addr);
 588         if (pud_none(*pud)) {
 589                 if (!cache)
 590                         return NULL;
 591                 pmd = mmu_memory_cache_alloc(cache);
 592                 pud_populate(NULL, pud, pmd);
 593                 get_page(virt_to_page(pud));
 594         }
 595
 596         return pmd_offset(pud, addr);
 597 }
 598
 599 static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
 600                                *cache, phys_addr_t addr, const pmd_t *new_pmd)
 601 {
 602         pmd_t *pmd, old_pmd;
 603
 604         pmd = stage2_get_pmd(kvm, cache, addr);
 605         VM_BUG_ON(!pmd);
 606
 607         /*
 608          * Mapping in huge pages should only happen through a fault.  If a
 609          * page is merged into a transparent huge page, the individual
 610          * subpages of that huge page should be unmapped through MMU
 611          * notifiers before we get here.
 612          *
 613          * Merging of CompoundPages is not supported; they should become
 614          * splitting first, unmapped, merged, and mapped back in on-demand.
 615          */
 616         VM_BUG_ON(pmd_present(*pmd) && pmd_pfn(*pmd) != pmd_pfn(*new_pmd));
 617
 618         old_pmd = *pmd;
 619         kvm_set_pmd(pmd, *new_pmd);
 620         if (pmd_present(old_pmd))
 621                 kvm_tlb_flush_vmid_ipa(kvm, addr);
 622         else
 623                 get_page(virt_to_page(pmd));
 624         return 0;
 625 }
 626
 627 static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
 628                           phys_addr_t addr, const pte_t *new_pte, bool iomap)
 629 {
 630         pmd_t *pmd;
 631         pte_t *pte, old_pte;
 632
 633         /* Create stage-2 page table mapping - Level 1 */
 634         pmd = stage2_get_pmd(kvm, cache, addr);
 635         if (!pmd) {
 636                 /*
 637                  * Ignore calls from kvm_set_spte_hva for unallocated
 638                  * address ranges.
 639                  */
 640                 return 0;
 641         }
 642
 643         /* Create stage-2 page mappings - Level 2 */
 644         if (pmd_none(*pmd)) {
 645                 if (!cache)
 646                         return 0; /* ignore calls from kvm_set_spte_hva */
 647                 pte = mmu_memory_cache_alloc(cache);
 648                 kvm_clean_pte(pte);
 649                 pmd_populate_kernel(NULL, pmd, pte);
 650                 get_page(virt_to_page(pmd));
 651         }
 652
 653         pte = pte_offset_kernel(pmd, addr);
 654
 655         if (iomap && pte_present(*pte))
 656                 return -EFAULT;
 657
 658         /* Create 2nd stage page table mapping - Level 3 */
 659         old_pte = *pte;
 660         kvm_set_pte(pte, *new_pte);
 661         if (pte_present(old_pte))
 662                 kvm_tlb_flush_vmid_ipa(kvm, addr);
 663         else
 664                 get_page(virt_to_page(pte));
 665
 666         return 0;
 667 }
 668
 669 /**
 670  * kvm_phys_addr_ioremap - map a device range to guest IPA
 671  *
 672  * @kvm:        The KVM pointer
 673  * @guest_ipa:  The IPA at which to insert the mapping
 674  * @pa:         The physical address of the device
 675  * @size:       The size of the mapping
 676  */
 677 int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
 678                           phys_addr_t pa, unsigned long size)
 679 {
 680         phys_addr_t addr, end;
 681         int ret = 0;
 682         unsigned long pfn;
 683         struct kvm_mmu_memory_cache cache = { 0, };
 684
 685         end = (guest_ipa + size + PAGE_SIZE - 1) & PAGE_MASK;
 686         pfn = __phys_to_pfn(pa);
 687
 688         for (addr = guest_ipa; addr < end; addr += PAGE_SIZE) {
 689                 pte_t pte = pfn_pte(pfn, PAGE_S2_DEVICE);
 690
 691                 ret = mmu_topup_memory_cache(&cache, 2, 2);
 692                 if (ret)
 693                         goto out;
 694                 spin_lock(&kvm->mmu_lock);
 695                 ret = stage2_set_pte(kvm, &cache, addr, &pte, true);
 696                 spin_unlock(&kvm->mmu_lock);
 697                 if (ret)
 698                         goto out;
 699
 700                 pfn++;
 701         }
 702
 703 out:
 704         mmu_free_memory_cache(&cache);
 705         return ret;
 706 }
 707
 708 static bool transparent_hugepage_adjust(pfn_t *pfnp, phys_addr_t *ipap)
 709 {
 710         pfn_t pfn = *pfnp;
 711         gfn_t gfn = *ipap >> PAGE_SHIFT;
 712
 713         if (PageTransCompound(pfn_to_page(pfn))) {
 714                 unsigned long mask;
 715                 /*
 716                  * The address we faulted on is backed by a transparent huge
 717                  * page.  However, because we map the compound huge page and
 718                  * not the individual tail page, we need to transfer the
 719                  * refcount to the head page.  We have to be careful that the
 720                  * THP doesn't start to split while we are adjusting the
 721                  * refcounts.
 722                  *
 723                  * We are sure this doesn't happen, because mmu_notifier_retry
 724                  * was successful and we are holding the mmu_lock, so if this
 725                  * THP is trying to split, it will be blocked in the mmu
 726                  * notifier before touching any of the pages, specifically
 727                  * before being able to call __split_huge_page_refcount().
 728                  *
 729                  * We can therefore safely transfer the refcount from PG_tail
 730                  * to PG_head and switch the pfn from a tail page to the head
 731                  * page accordingly.
 732                  */
 733                 mask = PTRS_PER_PMD - 1;
 734                 VM_BUG_ON((gfn & mask) != (pfn & mask));
 735                 if (pfn & mask) {
 736                         *ipap &= PMD_MASK;
 737                         kvm_release_pfn_clean(pfn);
 738                         pfn &= ~mask;
 739                         kvm_get_pfn(pfn);
 740                         *pfnp = pfn;
 741                 }
 742
 743                 return true;
 744         }
 745
 746         return false;
 747 }
 748
 749 static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
 750                           struct kvm_memory_slot *memslot,
 751                           unsigned long fault_status)
 752 {
 753         int ret;
 754         bool write_fault, writable, hugetlb = false, force_pte = false;
 755         unsigned long mmu_seq;
 756         gfn_t gfn = fault_ipa >> PAGE_SHIFT;
 757         unsigned long hva = gfn_to_hva(vcpu->kvm, gfn);
 758         struct kvm *kvm = vcpu->kvm;
 759         struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
 760         struct vm_area_struct *vma;
 761         pfn_t pfn;
 762         pgprot_t mem_type = PAGE_S2;
 763
 764         write_fault = kvm_is_write_fault(kvm_vcpu_get_hsr(vcpu));
 765         if (fault_status == FSC_PERM && !write_fault) {
 766                 kvm_err("Unexpected L2 read permission error\n");
 767                 return -EFAULT;
 768         }
 769
 770         /* Let's check if we will get back a huge page backed by hugetlbfs */
 771         down_read(&current->mm->mmap_sem);
 772         vma = find_vma_intersection(current->mm, hva, hva + 1);
 773         if (is_vm_hugetlb_page(vma)) {
 774                 hugetlb = true;
 775                 gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
 776         } else {
 777                 /*
 778                  * Pages belonging to memslots that don't have the same
 779                  * alignment for userspace and IPA cannot be mapped using
 780                  * block descriptors even if the pages belong to a THP for
 781                  * the process, because the stage-2 block descriptor will
 782                  * cover more than a single THP and we loose atomicity for
 783                  * unmapping, updates, and splits of the THP or other pages
 784                  * in the stage-2 block range.
 785                  */
 786                 if ((memslot->userspace_addr & ~PMD_MASK) !=
 787                     ((memslot->base_gfn << PAGE_SHIFT) & ~PMD_MASK))
 788                         force_pte = true;
 789         }
 790         up_read(&current->mm->mmap_sem);
 791
 792         /* We need minimum second+third level pages */
 793         ret = mmu_topup_memory_cache(memcache, 2, KVM_NR_MEM_OBJS);
 794         if (ret)
 795                 return ret;
 796
 797         mmu_seq = vcpu->kvm->mmu_notifier_seq;
 798         /*
 799          * Ensure the read of mmu_notifier_seq happens before we call
 800          * gfn_to_pfn_prot (which calls get_user_pages), so that we don't risk
 801          * the page we just got a reference to gets unmapped before we have a
 802          * chance to grab the mmu_lock, which ensure that if the page gets
 803          * unmapped afterwards, the call to kvm_unmap_hva will take it away
 804          * from us again properly. This smp_rmb() interacts with the smp_wmb()
 805          * in kvm_mmu_notifier_invalidate_<page|range_end>.
 806          */
 807         smp_rmb();
 808
 809         pfn = gfn_to_pfn_prot(kvm, gfn, write_fault, &writable);
 810         if (is_error_pfn(pfn))
 811                 return -EFAULT;
 812
 813         if (kvm_is_mmio_pfn(pfn))
 814                 mem_type = PAGE_S2_DEVICE;
 815
 816         spin_lock(&kvm->mmu_lock);
 817         if (mmu_notifier_retry(kvm, mmu_seq))
 818                 goto out_unlock;
 819         if (!hugetlb && !force_pte)
 820                 hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
 821
 822         if (hugetlb) {
 823                 pmd_t new_pmd = pfn_pmd(pfn, mem_type);
 824                 new_pmd = pmd_mkhuge(new_pmd);
 825                 if (writable) {
 826                         kvm_set_s2pmd_writable(&new_pmd);
 827                         kvm_set_pfn_dirty(pfn);
 828                 }
 829                 coherent_cache_guest_page(vcpu, hva & PMD_MASK, PMD_SIZE);
 830                 ret = stage2_set_pmd_huge(kvm, memcache, fault_ipa, &new_pmd);
 831         } else {
 832                 pte_t new_pte = pfn_pte(pfn, mem_type);
 833                 if (writable) {
 834                         kvm_set_s2pte_writable(&new_pte);
 835                         kvm_set_pfn_dirty(pfn);
 836                 }
 837                 coherent_cache_guest_page(vcpu, hva, PAGE_SIZE);
 838                 ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte,
 839                                      mem_type == PAGE_S2_DEVICE);
 840         }
 841
 842
 843 out_unlock:
 844         spin_unlock(&kvm->mmu_lock);
 845         kvm_release_pfn_clean(pfn);
 846         return ret;
 847 }
 848
 849 /**
 850  * kvm_handle_guest_abort - handles all 2nd stage aborts
 851  * @vcpu:       the VCPU pointer
 852  * @run:        the kvm_run structure
 853  *
 854  * Any abort that gets to the host is almost guaranteed to be caused by a
 855  * missing second stage translation table entry, which can mean that either the
 856  * guest simply needs more memory and we must allocate an appropriate page or it
 857  * can mean that the guest tried to access I/O memory, which is emulated by user
 858  * space. The distinction is based on the IPA causing the fault and whether this
 859  * memory region has been registered as standard RAM by user space.
 860  */
 861 int kvm_handle_guest_abort(struct kvm_vcpu *vcpu, struct kvm_run *run)
 862 {
 863         unsigned long fault_status;
 864         phys_addr_t fault_ipa;
 865         struct kvm_memory_slot *memslot;
 866         bool is_iabt;
 867         gfn_t gfn;
 868         int ret, idx;
 869
 870         is_iabt = kvm_vcpu_trap_is_iabt(vcpu);
 871         fault_ipa = kvm_vcpu_get_fault_ipa(vcpu);
 872
 873         trace_kvm_guest_fault(*vcpu_pc(vcpu), kvm_vcpu_get_hsr(vcpu),
 874                               kvm_vcpu_get_hfar(vcpu), fault_ipa);
 875
 876         /* Check the stage-2 fault is trans. fault or write fault */
 877         fault_status = kvm_vcpu_trap_get_fault(vcpu);
 878         if (fault_status != FSC_FAULT && fault_status != FSC_PERM) {
 879                 kvm_err("Unsupported fault status: EC=%#x DFCS=%#lx\n",
 880                         kvm_vcpu_trap_get_class(vcpu), fault_status);
 881                 return -EFAULT;
 882         }
 883
 884         idx = srcu_read_lock(&vcpu->kvm->srcu);
 885
 886         gfn = fault_ipa >> PAGE_SHIFT;
 887         if (!kvm_is_visible_gfn(vcpu->kvm, gfn)) {
 888                 if (is_iabt) {
 889                         /* Prefetch Abort on I/O address */
 890                         kvm_inject_pabt(vcpu, kvm_vcpu_get_hfar(vcpu));
 891                         ret = 1;
 892                         goto out_unlock;
 893                 }
 894
 895                 if (fault_status != FSC_FAULT) {
 896                         kvm_err("Unsupported fault status on io memory: %#lx\n",
 897                                 fault_status);
 898                         ret = -EFAULT;
 899                         goto out_unlock;
 900                 }
 901
 902                 /*
 903                  * The IPA is reported as [MAX:12], so we need to
 904                  * complement it with the bottom 12 bits from the
 905                  * faulting VA. This is always 12 bits, irrespective
 906                  * of the page size.
 907                  */
 908                 fault_ipa |= kvm_vcpu_get_hfar(vcpu) & ((1 << 12) - 1);
 909                 ret = io_mem_abort(vcpu, run, fault_ipa);
 910                 goto out_unlock;
 911         }
 912
 913         memslot = gfn_to_memslot(vcpu->kvm, gfn);
 914
 915         ret = user_mem_abort(vcpu, fault_ipa, memslot, fault_status);
 916         if (ret == 0)
 917                 ret = 1;
 918 out_unlock:
 919         srcu_read_unlock(&vcpu->kvm->srcu, idx);
 920         return ret;
 921 }
 922
 923 static void handle_hva_to_gpa(struct kvm *kvm,
 924                               unsigned long start,
 925                               unsigned long end,
 926                               void (*handler)(struct kvm *kvm,
 927                                               gpa_t gpa, void *data),
 928                               void *data)
 929 {
 930         struct kvm_memslots *slots;
 931         struct kvm_memory_slot *memslot;
 932
 933         slots = kvm_memslots(kvm);
 934
 935         /* we only care about the pages that the guest sees */
 936         kvm_for_each_memslot(memslot, slots) {
 937                 unsigned long hva_start, hva_end;
 938                 gfn_t gfn, gfn_end;
 939
 940                 hva_start = max(start, memslot->userspace_addr);
 941                 hva_end = min(end, memslot->userspace_addr +
 942                                         (memslot->npages << PAGE_SHIFT));
 943                 if (hva_start >= hva_end)
 944                         continue;
 945
 946                 /*
 947                  * {gfn(page) | page intersects with [hva_start, hva_end)} =
 948                  * {gfn_start, gfn_start+1, ..., gfn_end-1}.
 949                  */
 950                 gfn = hva_to_gfn_memslot(hva_start, memslot);
 951                 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
 952
 953                 for (; gfn < gfn_end; ++gfn) {
 954                         gpa_t gpa = gfn << PAGE_SHIFT;
 955                         handler(kvm, gpa, data);
 956                 }
 957         }
 958 }
 959
 960 static void kvm_unmap_hva_handler(struct kvm *kvm, gpa_t gpa, void *data)
 961 {
 962         unmap_stage2_range(kvm, gpa, PAGE_SIZE);
 963 }
 964
 965 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 966 {
 967         unsigned long end = hva + PAGE_SIZE;
 968
 969         if (!kvm->arch.pgd)
 970                 return 0;
 971
 972         trace_kvm_unmap_hva(hva);
 973         handle_hva_to_gpa(kvm, hva, end, &kvm_unmap_hva_handler, NULL);
 974         return 0;
 975 }
 976
 977 int kvm_unmap_hva_range(struct kvm *kvm,
 978                         unsigned long start, unsigned long end)
 979 {
 980         if (!kvm->arch.pgd)
 981                 return 0;
 982
 983         trace_kvm_unmap_hva_range(start, end);
 984         handle_hva_to_gpa(kvm, start, end, &kvm_unmap_hva_handler, NULL);
 985         return 0;
 986 }
 987
 988 static void kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, void *data)
 989 {
 990         pte_t *pte = (pte_t *)data;
 991
 992         stage2_set_pte(kvm, NULL, gpa, pte, false);
 993 }
 994
 995
 996 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
 997 {
 998         unsigned long end = hva + PAGE_SIZE;
 999         pte_t stage2_pte;
1000
1001         if (!kvm->arch.pgd)
1002                 return;
1003
1004         trace_kvm_set_spte_hva(hva);
1005         stage2_pte = pfn_pte(pte_pfn(pte), PAGE_S2);
1006         handle_hva_to_gpa(kvm, hva, end, &kvm_set_spte_handler, &stage2_pte);
1007 }
1008
1009 void kvm_mmu_free_memory_caches(struct kvm_vcpu *vcpu)
1010 {
1011         mmu_free_memory_cache(&vcpu->arch.mmu_page_cache);
1012 }
1013
1014 phys_addr_t kvm_mmu_get_httbr(void)
1015 {
1016         return virt_to_phys(hyp_pgd);
1017 }
1018
1019 phys_addr_t kvm_mmu_get_boot_httbr(void)
1020 {
1021         return virt_to_phys(boot_hyp_pgd);
1022 }
1023
1024 phys_addr_t kvm_get_idmap_vector(void)
1025 {
1026         return hyp_idmap_vector;
1027 }
1028
1029 int kvm_mmu_init(void)
1030 {
1031         int err;
1032
1033         hyp_idmap_start = kvm_virt_to_phys(__hyp_idmap_text_start);
1034         hyp_idmap_end = kvm_virt_to_phys(__hyp_idmap_text_end);
1035         hyp_idmap_vector = kvm_virt_to_phys(__kvm_hyp_init);
1036
1037         if ((hyp_idmap_start ^ hyp_idmap_end) & PAGE_MASK) {
1038                 /*
1039                  * Our init code is crossing a page boundary. Allocate
1040                  * a bounce page, copy the code over and use that.
1041                  */
1042                 size_t len = __hyp_idmap_text_end - __hyp_idmap_text_start;
1043                 phys_addr_t phys_base;
1044
1045                 init_bounce_page = (void *)__get_free_page(GFP_KERNEL);
1046                 if (!init_bounce_page) {
1047                         kvm_err("Couldn't allocate HYP init bounce page\n");
1048                         err = -ENOMEM;
1049                         goto out;
1050                 }
1051
1052                 memcpy(init_bounce_page, __hyp_idmap_text_start, len);
1053                 /*
1054                  * Warning: the code we just copied to the bounce page
1055                  * must be flushed to the point of coherency.
1056                  * Otherwise, the data may be sitting in L2, and HYP
1057                  * mode won't be able to observe it as it runs with
1058                  * caches off at that point.
1059                  */
1060                 kvm_flush_dcache_to_poc(init_bounce_page, len);
1061
1062                 phys_base = kvm_virt_to_phys(init_bounce_page);
1063                 hyp_idmap_vector += phys_base - hyp_idmap_start;
1064                 hyp_idmap_start = phys_base;
1065                 hyp_idmap_end = phys_base + len;
1066
1067                 kvm_info("Using HYP init bounce page @%lx\n",
1068                          (unsigned long)phys_base);
1069         }
1070
1071         hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, pgd_order);
1072         boot_hyp_pgd = (pgd_t *)__get_free_pages(GFP_KERNEL | __GFP_ZERO, pgd_order);
1073
1074         if (!hyp_pgd || !boot_hyp_pgd) {
1075                 kvm_err("Hyp mode PGD not allocated\n");
1076                 err = -ENOMEM;
1077                 goto out;
1078         }
1079
1080         /* Create the idmap in the boot page tables */
1081         err =   __create_hyp_mappings(boot_hyp_pgd,
1082                                       hyp_idmap_start, hyp_idmap_end,
1083                                       __phys_to_pfn(hyp_idmap_start),
1084                                       PAGE_HYP);
1085
1086         if (err) {
1087                 kvm_err("Failed to idmap %lx-%lx\n",
1088                         hyp_idmap_start, hyp_idmap_end);
1089                 goto out;
1090         }
1091
1092         /* Map the very same page at the trampoline VA */
1093         err =   __create_hyp_mappings(boot_hyp_pgd,
1094                                       TRAMPOLINE_VA, TRAMPOLINE_VA + PAGE_SIZE,
1095                                       __phys_to_pfn(hyp_idmap_start),
1096                                       PAGE_HYP);
1097         if (err) {
1098                 kvm_err("Failed to map trampoline @%lx into boot HYP pgd\n",
1099                         TRAMPOLINE_VA);
1100                 goto out;
1101         }
1102
1103         /* Map the same page again into the runtime page tables */
1104         err =   __create_hyp_mappings(hyp_pgd,
1105                                       TRAMPOLINE_VA, TRAMPOLINE_VA + PAGE_SIZE,
1106                                       __phys_to_pfn(hyp_idmap_start),
1107                                       PAGE_HYP);
1108         if (err) {
1109                 kvm_err("Failed to map trampoline @%lx into runtime HYP pgd\n",
1110                         TRAMPOLINE_VA);
1111                 goto out;
1112         }
1113
1114         return 0;
1115 out:
1116         free_hyp_pgds();
1117         return err;
1118 }
1119
1120 void kvm_arch_commit_memory_region(struct kvm *kvm,
1121                                    struct kvm_userspace_memory_region *mem,
1122                                    const struct kvm_memory_slot *old,
1123                                    enum kvm_mr_change change)
1124 {
1125         gpa_t gpa = old->base_gfn << PAGE_SHIFT;
1126         phys_addr_t size = old->npages << PAGE_SHIFT;
1127         if (change == KVM_MR_DELETE || change == KVM_MR_MOVE) {
1128                 spin_lock(&kvm->mmu_lock);
1129                 unmap_stage2_range(kvm, gpa, size);
1130                 spin_unlock(&kvm->mmu_lock);
1131         }
1132 }
1133
1134 int kvm_arch_prepare_memory_region(struct kvm *kvm,
1135                                    struct kvm_memory_slot *memslot,
1136                                    struct kvm_userspace_memory_region *mem,
1137                                    enum kvm_mr_change change)
1138 {
1139         return 0;
1140 }
1141
1142 void kvm_arch_free_memslot(struct kvm *kvm, struct kvm_memory_slot *free,
1143                            struct kvm_memory_slot *dont)
1144 {
1145 }
1146
1147 int kvm_arch_create_memslot(struct kvm *kvm, struct kvm_memory_slot *slot,
1148                             unsigned long npages)
1149 {
1150         return 0;
1151 }
1152
1153 void kvm_arch_memslots_updated(struct kvm *kvm)
1154 {
1155 }
1156
1157 void kvm_arch_flush_shadow_all(struct kvm *kvm)
1158 {
1159 }
1160
1161 void kvm_arch_flush_shadow_memslot(struct kvm *kvm,
1162                                    struct kvm_memory_slot *slot)
1163 {
1164 }