arch/arm/kvm/mmu.c

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