arch/powerpc/kvm/book3s_64_mmu_hv.c

   1 /*
   2  * This program is free software; you can redistribute it and/or modify
   3  * it under the terms of the GNU General Public License, version 2, as
   4  * published by the Free Software Foundation.
   5  *
   6  * This program is distributed in the hope that it will be useful,
   7  * but WITHOUT ANY WARRANTY; without even the implied warranty of
   8  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
   9  * GNU General Public License for more details.
  10  *
  11  * You should have received a copy of the GNU General Public License
  12  * along with this program; if not, write to the Free Software
  13  * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
  14  *
  15  * Copyright 2010 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
  16  */
  17
  18 #include <linux/types.h>
  19 #include <linux/string.h>
  20 #include <linux/kvm.h>
  21 #include <linux/kvm_host.h>
  22 #include <linux/highmem.h>
  23 #include <linux/gfp.h>
  24 #include <linux/slab.h>
  25 #include <linux/hugetlb.h>
  26 #include <linux/vmalloc.h>
  27 #include <linux/srcu.h>
  28 #include <linux/anon_inodes.h>
  29 #include <linux/file.h>
  30
  31 #include <asm/tlbflush.h>
  32 #include <asm/kvm_ppc.h>
  33 #include <asm/kvm_book3s.h>
  34 #include <asm/mmu-hash64.h>
  35 #include <asm/hvcall.h>
  36 #include <asm/synch.h>
  37 #include <asm/ppc-opcode.h>
  38 #include <asm/cputable.h>
  39
  40 /* POWER7 has 10-bit LPIDs, PPC970 has 6-bit LPIDs */
  41 #define MAX_LPID_970    63
  42
  43 /* Power architecture requires HPT is at least 256kB */
  44 #define PPC_MIN_HPT_ORDER       18
  45
  46 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
  47                                 long pte_index, unsigned long pteh,
  48                                 unsigned long ptel, unsigned long *pte_idx_ret);
  49 static void kvmppc_rmap_reset(struct kvm *kvm);
  50
  51 long kvmppc_alloc_hpt(struct kvm *kvm, u32 *htab_orderp)
  52 {
  53         unsigned long hpt;
  54         struct revmap_entry *rev;
  55         struct kvmppc_linear_info *li;
  56         long order = kvm_hpt_order;
  57
  58         if (htab_orderp) {
  59                 order = *htab_orderp;
  60                 if (order < PPC_MIN_HPT_ORDER)
  61                         order = PPC_MIN_HPT_ORDER;
  62         }
  63
  64         /*
  65          * If the user wants a different size from default,
  66          * try first to allocate it from the kernel page allocator.
  67          */
  68         hpt = 0;
  69         if (order != kvm_hpt_order) {
  70                 hpt = __get_free_pages(GFP_KERNEL|__GFP_ZERO|__GFP_REPEAT|
  71                                        __GFP_NOWARN, order - PAGE_SHIFT);
  72                 if (!hpt)
  73                         --order;
  74         }
  75
  76         /* Next try to allocate from the preallocated pool */
  77         if (!hpt) {
  78                 li = kvm_alloc_hpt();
  79                 if (li) {
  80                         hpt = (ulong)li->base_virt;
  81                         kvm->arch.hpt_li = li;
  82                         order = kvm_hpt_order;
  83                 }
  84         }
  85
  86         /* Lastly try successively smaller sizes from the page allocator */
  87         while (!hpt && order > PPC_MIN_HPT_ORDER) {
  88                 hpt = __get_free_pages(GFP_KERNEL|__GFP_ZERO|__GFP_REPEAT|
  89                                        __GFP_NOWARN, order - PAGE_SHIFT);
  90                 if (!hpt)
  91                         --order;
  92         }
  93
  94         if (!hpt)
  95                 return -ENOMEM;
  96
  97         kvm->arch.hpt_virt = hpt;
  98         kvm->arch.hpt_order = order;
  99         /* HPTEs are 2**4 bytes long */
 100         kvm->arch.hpt_npte = 1ul << (order - 4);
 101         /* 128 (2**7) bytes in each HPTEG */
 102         kvm->arch.hpt_mask = (1ul << (order - 7)) - 1;
 103
 104         /* Allocate reverse map array */
 105         rev = vmalloc(sizeof(struct revmap_entry) * kvm->arch.hpt_npte);
 106         if (!rev) {
 107                 pr_err("kvmppc_alloc_hpt: Couldn't alloc reverse map array\n");
 108                 goto out_freehpt;
 109         }
 110         kvm->arch.revmap = rev;
 111         kvm->arch.sdr1 = __pa(hpt) | (order - 18);
 112
 113         pr_info("KVM guest htab at %lx (order %ld), LPID %x\n",
 114                 hpt, order, kvm->arch.lpid);
 115
 116         if (htab_orderp)
 117                 *htab_orderp = order;
 118         return 0;
 119
 120  out_freehpt:
 121         if (kvm->arch.hpt_li)
 122                 kvm_release_hpt(kvm->arch.hpt_li);
 123         else
 124                 free_pages(hpt, order - PAGE_SHIFT);
 125         return -ENOMEM;
 126 }
 127
 128 long kvmppc_alloc_reset_hpt(struct kvm *kvm, u32 *htab_orderp)
 129 {
 130         long err = -EBUSY;
 131         long order;
 132
 133         mutex_lock(&kvm->lock);
 134         if (kvm->arch.rma_setup_done) {
 135                 kvm->arch.rma_setup_done = 0;
 136                 /* order rma_setup_done vs. vcpus_running */
 137                 smp_mb();
 138                 if (atomic_read(&kvm->arch.vcpus_running)) {
 139                         kvm->arch.rma_setup_done = 1;
 140                         goto out;
 141                 }
 142         }
 143         if (kvm->arch.hpt_virt) {
 144                 order = kvm->arch.hpt_order;
 145                 /* Set the entire HPT to 0, i.e. invalid HPTEs */
 146                 memset((void *)kvm->arch.hpt_virt, 0, 1ul << order);
 147                 /*
 148                  * Reset all the reverse-mapping chains for all memslots
 149                  */
 150                 kvmppc_rmap_reset(kvm);
 151                 /* Ensure that each vcpu will flush its TLB on next entry. */
 152                 cpumask_setall(&kvm->arch.need_tlb_flush);
 153                 *htab_orderp = order;
 154                 err = 0;
 155         } else {
 156                 err = kvmppc_alloc_hpt(kvm, htab_orderp);
 157                 order = *htab_orderp;
 158         }
 159  out:
 160         mutex_unlock(&kvm->lock);
 161         return err;
 162 }
 163
 164 void kvmppc_free_hpt(struct kvm *kvm)
 165 {
 166         kvmppc_free_lpid(kvm->arch.lpid);
 167         vfree(kvm->arch.revmap);
 168         if (kvm->arch.hpt_li)
 169                 kvm_release_hpt(kvm->arch.hpt_li);
 170         else
 171                 free_pages(kvm->arch.hpt_virt,
 172                            kvm->arch.hpt_order - PAGE_SHIFT);
 173 }
 174
 175 /* Bits in first HPTE dword for pagesize 4k, 64k or 16M */
 176 static inline unsigned long hpte0_pgsize_encoding(unsigned long pgsize)
 177 {
 178         return (pgsize > 0x1000) ? HPTE_V_LARGE : 0;
 179 }
 180
 181 /* Bits in second HPTE dword for pagesize 4k, 64k or 16M */
 182 static inline unsigned long hpte1_pgsize_encoding(unsigned long pgsize)
 183 {
 184         return (pgsize == 0x10000) ? 0x1000 : 0;
 185 }
 186
 187 void kvmppc_map_vrma(struct kvm_vcpu *vcpu, struct kvm_memory_slot *memslot,
 188                      unsigned long porder)
 189 {
 190         unsigned long i;
 191         unsigned long npages;
 192         unsigned long hp_v, hp_r;
 193         unsigned long addr, hash;
 194         unsigned long psize;
 195         unsigned long hp0, hp1;
 196         unsigned long idx_ret;
 197         long ret;
 198         struct kvm *kvm = vcpu->kvm;
 199
 200         psize = 1ul << porder;
 201         npages = memslot->npages >> (porder - PAGE_SHIFT);
 202
 203         /* VRMA can't be > 1TB */
 204         if (npages > 1ul << (40 - porder))
 205                 npages = 1ul << (40 - porder);
 206         /* Can't use more than 1 HPTE per HPTEG */
 207         if (npages > kvm->arch.hpt_mask + 1)
 208                 npages = kvm->arch.hpt_mask + 1;
 209
 210         hp0 = HPTE_V_1TB_SEG | (VRMA_VSID << (40 - 16)) |
 211                 HPTE_V_BOLTED | hpte0_pgsize_encoding(psize);
 212         hp1 = hpte1_pgsize_encoding(psize) |
 213                 HPTE_R_R | HPTE_R_C | HPTE_R_M | PP_RWXX;
 214
 215         for (i = 0; i < npages; ++i) {
 216                 addr = i << porder;
 217                 /* can't use hpt_hash since va > 64 bits */
 218                 hash = (i ^ (VRMA_VSID ^ (VRMA_VSID << 25))) & kvm->arch.hpt_mask;
 219                 /*
 220                  * We assume that the hash table is empty and no
 221                  * vcpus are using it at this stage.  Since we create
 222                  * at most one HPTE per HPTEG, we just assume entry 7
 223                  * is available and use it.
 224                  */
 225                 hash = (hash << 3) + 7;
 226                 hp_v = hp0 | ((addr >> 16) & ~0x7fUL);
 227                 hp_r = hp1 | addr;
 228                 ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, hash, hp_v, hp_r,
 229                                                  &idx_ret);
 230                 if (ret != H_SUCCESS) {
 231                         pr_err("KVM: map_vrma at %lx failed, ret=%ld\n",
 232                                addr, ret);
 233                         break;
 234                 }
 235         }
 236 }
 237
 238 int kvmppc_mmu_hv_init(void)
 239 {
 240         unsigned long host_lpid, rsvd_lpid;
 241
 242         if (!cpu_has_feature(CPU_FTR_HVMODE))
 243                 return -EINVAL;
 244
 245         /* POWER7 has 10-bit LPIDs, PPC970 and e500mc have 6-bit LPIDs */
 246         if (cpu_has_feature(CPU_FTR_ARCH_206)) {
 247                 host_lpid = mfspr(SPRN_LPID);   /* POWER7 */
 248                 rsvd_lpid = LPID_RSVD;
 249         } else {
 250                 host_lpid = 0;                  /* PPC970 */
 251                 rsvd_lpid = MAX_LPID_970;
 252         }
 253
 254         kvmppc_init_lpid(rsvd_lpid + 1);
 255
 256         kvmppc_claim_lpid(host_lpid);
 257         /* rsvd_lpid is reserved for use in partition switching */
 258         kvmppc_claim_lpid(rsvd_lpid);
 259
 260         return 0;
 261 }
 262
 263 void kvmppc_mmu_destroy(struct kvm_vcpu *vcpu)
 264 {
 265 }
 266
 267 static void kvmppc_mmu_book3s_64_hv_reset_msr(struct kvm_vcpu *vcpu)
 268 {
 269         kvmppc_set_msr(vcpu, MSR_SF | MSR_ME);
 270 }
 271
 272 /*
 273  * This is called to get a reference to a guest page if there isn't
 274  * one already in the memslot->arch.slot_phys[] array.
 275  */
 276 static long kvmppc_get_guest_page(struct kvm *kvm, unsigned long gfn,
 277                                   struct kvm_memory_slot *memslot,
 278                                   unsigned long psize)
 279 {
 280         unsigned long start;
 281         long np, err;
 282         struct page *page, *hpage, *pages[1];
 283         unsigned long s, pgsize;
 284         unsigned long *physp;
 285         unsigned int is_io, got, pgorder;
 286         struct vm_area_struct *vma;
 287         unsigned long pfn, i, npages;
 288
 289         physp = memslot->arch.slot_phys;
 290         if (!physp)
 291                 return -EINVAL;
 292         if (physp[gfn - memslot->base_gfn])
 293                 return 0;
 294
 295         is_io = 0;
 296         got = 0;
 297         page = NULL;
 298         pgsize = psize;
 299         err = -EINVAL;
 300         start = gfn_to_hva_memslot(memslot, gfn);
 301
 302         /* Instantiate and get the page we want access to */
 303         np = get_user_pages_fast(start, 1, 1, pages);
 304         if (np != 1) {
 305                 /* Look up the vma for the page */
 306                 down_read(&current->mm->mmap_sem);
 307                 vma = find_vma(current->mm, start);
 308                 if (!vma || vma->vm_start > start ||
 309                     start + psize > vma->vm_end ||
 310                     !(vma->vm_flags & VM_PFNMAP))
 311                         goto up_err;
 312                 is_io = hpte_cache_bits(pgprot_val(vma->vm_page_prot));
 313                 pfn = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 314                 /* check alignment of pfn vs. requested page size */
 315                 if (psize > PAGE_SIZE && (pfn & ((psize >> PAGE_SHIFT) - 1)))
 316                         goto up_err;
 317                 up_read(&current->mm->mmap_sem);
 318
 319         } else {
 320                 page = pages[0];
 321                 got = KVMPPC_GOT_PAGE;
 322
 323                 /* See if this is a large page */
 324                 s = PAGE_SIZE;
 325                 if (PageHuge(page)) {
 326                         hpage = compound_head(page);
 327                         s <<= compound_order(hpage);
 328                         /* Get the whole large page if slot alignment is ok */
 329                         if (s > psize && slot_is_aligned(memslot, s) &&
 330                             !(memslot->userspace_addr & (s - 1))) {
 331                                 start &= ~(s - 1);
 332                                 pgsize = s;
 333                                 get_page(hpage);
 334                                 put_page(page);
 335                                 page = hpage;
 336                         }
 337                 }
 338                 if (s < psize)
 339                         goto out;
 340                 pfn = page_to_pfn(page);
 341         }
 342
 343         npages = pgsize >> PAGE_SHIFT;
 344         pgorder = __ilog2(npages);
 345         physp += (gfn - memslot->base_gfn) & ~(npages - 1);
 346         spin_lock(&kvm->arch.slot_phys_lock);
 347         for (i = 0; i < npages; ++i) {
 348                 if (!physp[i]) {
 349                         physp[i] = ((pfn + i) << PAGE_SHIFT) +
 350                                 got + is_io + pgorder;
 351                         got = 0;
 352                 }
 353         }
 354         spin_unlock(&kvm->arch.slot_phys_lock);
 355         err = 0;
 356
 357  out:
 358         if (got)
 359                 put_page(page);
 360         return err;
 361
 362  up_err:
 363         up_read(&current->mm->mmap_sem);
 364         return err;
 365 }
 366
 367 long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
 368                                 long pte_index, unsigned long pteh,
 369                                 unsigned long ptel, unsigned long *pte_idx_ret)
 370 {
 371         unsigned long psize, gpa, gfn;
 372         struct kvm_memory_slot *memslot;
 373         long ret;
 374
 375         if (kvm->arch.using_mmu_notifiers)
 376                 goto do_insert;
 377
 378         psize = hpte_page_size(pteh, ptel);
 379         if (!psize)
 380                 return H_PARAMETER;
 381
 382         pteh &= ~(HPTE_V_HVLOCK | HPTE_V_ABSENT | HPTE_V_VALID);
 383
 384         /* Find the memslot (if any) for this address */
 385         gpa = (ptel & HPTE_R_RPN) & ~(psize - 1);
 386         gfn = gpa >> PAGE_SHIFT;
 387         memslot = gfn_to_memslot(kvm, gfn);
 388         if (memslot && !(memslot->flags & KVM_MEMSLOT_INVALID)) {
 389                 if (!slot_is_aligned(memslot, psize))
 390                         return H_PARAMETER;
 391                 if (kvmppc_get_guest_page(kvm, gfn, memslot, psize) < 0)
 392                         return H_PARAMETER;
 393         }
 394
 395  do_insert:
 396         /* Protect linux PTE lookup from page table destruction */
 397         rcu_read_lock_sched();  /* this disables preemption too */
 398         ret = kvmppc_do_h_enter(kvm, flags, pte_index, pteh, ptel,
 399                                 current->mm->pgd, false, pte_idx_ret);
 400         rcu_read_unlock_sched();
 401         if (ret == H_TOO_HARD) {
 402                 /* this can't happen */
 403                 pr_err("KVM: Oops, kvmppc_h_enter returned too hard!\n");
 404                 ret = H_RESOURCE;       /* or something */
 405         }
 406         return ret;
 407
 408 }
 409
 410 /*
 411  * We come here on a H_ENTER call from the guest when we are not
 412  * using mmu notifiers and we don't have the requested page pinned
 413  * already.
 414  */
 415 long kvmppc_virtmode_h_enter(struct kvm_vcpu *vcpu, unsigned long flags,
 416                              long pte_index, unsigned long pteh,
 417                              unsigned long ptel)
 418 {
 419         return kvmppc_virtmode_do_h_enter(vcpu->kvm, flags, pte_index,
 420                                           pteh, ptel, &vcpu->arch.gpr[4]);
 421 }
 422
 423 static struct kvmppc_slb *kvmppc_mmu_book3s_hv_find_slbe(struct kvm_vcpu *vcpu,
 424                                                          gva_t eaddr)
 425 {
 426         u64 mask;
 427         int i;
 428
 429         for (i = 0; i < vcpu->arch.slb_nr; i++) {
 430                 if (!(vcpu->arch.slb[i].orige & SLB_ESID_V))
 431                         continue;
 432
 433                 if (vcpu->arch.slb[i].origv & SLB_VSID_B_1T)
 434                         mask = ESID_MASK_1T;
 435                 else
 436                         mask = ESID_MASK;
 437
 438                 if (((vcpu->arch.slb[i].orige ^ eaddr) & mask) == 0)
 439                         return &vcpu->arch.slb[i];
 440         }
 441         return NULL;
 442 }
 443
 444 static unsigned long kvmppc_mmu_get_real_addr(unsigned long v, unsigned long r,
 445                         unsigned long ea)
 446 {
 447         unsigned long ra_mask;
 448
 449         ra_mask = hpte_page_size(v, r) - 1;
 450         return (r & HPTE_R_RPN & ~ra_mask) | (ea & ra_mask);
 451 }
 452
 453 static int kvmppc_mmu_book3s_64_hv_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
 454                         struct kvmppc_pte *gpte, bool data)
 455 {
 456         struct kvm *kvm = vcpu->kvm;
 457         struct kvmppc_slb *slbe;
 458         unsigned long slb_v;
 459         unsigned long pp, key;
 460         unsigned long v, gr;
 461         unsigned long *hptep;
 462         int index;
 463         int virtmode = vcpu->arch.shregs.msr & (data ? MSR_DR : MSR_IR);
 464
 465         /* Get SLB entry */
 466         if (virtmode) {
 467                 slbe = kvmppc_mmu_book3s_hv_find_slbe(vcpu, eaddr);
 468                 if (!slbe)
 469                         return -EINVAL;
 470                 slb_v = slbe->origv;
 471         } else {
 472                 /* real mode access */
 473                 slb_v = vcpu->kvm->arch.vrma_slb_v;
 474         }
 475
 476         /* Find the HPTE in the hash table */
 477         index = kvmppc_hv_find_lock_hpte(kvm, eaddr, slb_v,
 478                                          HPTE_V_VALID | HPTE_V_ABSENT);
 479         if (index < 0)
 480                 return -ENOENT;
 481         hptep = (unsigned long *)(kvm->arch.hpt_virt + (index << 4));
 482         v = hptep[0] & ~HPTE_V_HVLOCK;
 483         gr = kvm->arch.revmap[index].guest_rpte;
 484
 485         /* Unlock the HPTE */
 486         asm volatile("lwsync" : : : "memory");
 487         hptep[0] = v;
 488
 489         gpte->eaddr = eaddr;
 490         gpte->vpage = ((v & HPTE_V_AVPN) << 4) | ((eaddr >> 12) & 0xfff);
 491
 492         /* Get PP bits and key for permission check */
 493         pp = gr & (HPTE_R_PP0 | HPTE_R_PP);
 494         key = (vcpu->arch.shregs.msr & MSR_PR) ? SLB_VSID_KP : SLB_VSID_KS;
 495         key &= slb_v;
 496
 497         /* Calculate permissions */
 498         gpte->may_read = hpte_read_permission(pp, key);
 499         gpte->may_write = hpte_write_permission(pp, key);
 500         gpte->may_execute = gpte->may_read && !(gr & (HPTE_R_N | HPTE_R_G));
 501
 502         /* Storage key permission check for POWER7 */
 503         if (data && virtmode && cpu_has_feature(CPU_FTR_ARCH_206)) {
 504                 int amrfield = hpte_get_skey_perm(gr, vcpu->arch.amr);
 505                 if (amrfield & 1)
 506                         gpte->may_read = 0;
 507                 if (amrfield & 2)
 508                         gpte->may_write = 0;
 509         }
 510
 511         /* Get the guest physical address */
 512         gpte->raddr = kvmppc_mmu_get_real_addr(v, gr, eaddr);
 513         return 0;
 514 }
 515
 516 /*
 517  * Quick test for whether an instruction is a load or a store.
 518  * If the instruction is a load or a store, then this will indicate
 519  * which it is, at least on server processors.  (Embedded processors
 520  * have some external PID instructions that don't follow the rule
 521  * embodied here.)  If the instruction isn't a load or store, then
 522  * this doesn't return anything useful.
 523  */
 524 static int instruction_is_store(unsigned int instr)
 525 {
 526         unsigned int mask;
 527
 528         mask = 0x10000000;
 529         if ((instr & 0xfc000000) == 0x7c000000)
 530                 mask = 0x100;           /* major opcode 31 */
 531         return (instr & mask) != 0;
 532 }
 533
 534 static int kvmppc_hv_emulate_mmio(struct kvm_run *run, struct kvm_vcpu *vcpu,
 535                                   unsigned long gpa, gva_t ea, int is_store)
 536 {
 537         int ret;
 538         u32 last_inst;
 539         unsigned long srr0 = kvmppc_get_pc(vcpu);
 540
 541         /* We try to load the last instruction.  We don't let
 542          * emulate_instruction do it as it doesn't check what
 543          * kvmppc_ld returns.
 544          * If we fail, we just return to the guest and try executing it again.
 545          */
 546         if (vcpu->arch.last_inst == KVM_INST_FETCH_FAILED) {
 547                 ret = kvmppc_ld(vcpu, &srr0, sizeof(u32), &last_inst, false);
 548                 if (ret != EMULATE_DONE || last_inst == KVM_INST_FETCH_FAILED)
 549                         return RESUME_GUEST;
 550                 vcpu->arch.last_inst = last_inst;
 551         }
 552
 553         /*
 554          * WARNING: We do not know for sure whether the instruction we just
 555          * read from memory is the same that caused the fault in the first
 556          * place.  If the instruction we read is neither an load or a store,
 557          * then it can't access memory, so we don't need to worry about
 558          * enforcing access permissions.  So, assuming it is a load or
 559          * store, we just check that its direction (load or store) is
 560          * consistent with the original fault, since that's what we
 561          * checked the access permissions against.  If there is a mismatch
 562          * we just return and retry the instruction.
 563          */
 564
 565         if (instruction_is_store(vcpu->arch.last_inst) != !!is_store)
 566                 return RESUME_GUEST;
 567
 568         /*
 569          * Emulated accesses are emulated by looking at the hash for
 570          * translation once, then performing the access later. The
 571          * translation could be invalidated in the meantime in which
 572          * point performing the subsequent memory access on the old
 573          * physical address could possibly be a security hole for the
 574          * guest (but not the host).
 575          *
 576          * This is less of an issue for MMIO stores since they aren't
 577          * globally visible. It could be an issue for MMIO loads to
 578          * a certain extent but we'll ignore it for now.
 579          */
 580
 581         vcpu->arch.paddr_accessed = gpa;
 582         vcpu->arch.vaddr_accessed = ea;
 583         return kvmppc_emulate_mmio(run, vcpu);
 584 }
 585
 586 int kvmppc_book3s_hv_page_fault(struct kvm_run *run, struct kvm_vcpu *vcpu,
 587                                 unsigned long ea, unsigned long dsisr)
 588 {
 589         struct kvm *kvm = vcpu->kvm;
 590         unsigned long *hptep, hpte[3], r;
 591         unsigned long mmu_seq, psize, pte_size;
 592         unsigned long gpa, gfn, hva, pfn;
 593         struct kvm_memory_slot *memslot;
 594         unsigned long *rmap;
 595         struct revmap_entry *rev;
 596         struct page *page, *pages[1];
 597         long index, ret, npages;
 598         unsigned long is_io;
 599         unsigned int writing, write_ok;
 600         struct vm_area_struct *vma;
 601         unsigned long rcbits;
 602
 603         /*
 604          * Real-mode code has already searched the HPT and found the
 605          * entry we're interested in.  Lock the entry and check that
 606          * it hasn't changed.  If it has, just return and re-execute the
 607          * instruction.
 608          */
 609         if (ea != vcpu->arch.pgfault_addr)
 610                 return RESUME_GUEST;
 611         index = vcpu->arch.pgfault_index;
 612         hptep = (unsigned long *)(kvm->arch.hpt_virt + (index << 4));
 613         rev = &kvm->arch.revmap[index];
 614         preempt_disable();
 615         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 616                 cpu_relax();
 617         hpte[0] = hptep[0] & ~HPTE_V_HVLOCK;
 618         hpte[1] = hptep[1];
 619         hpte[2] = r = rev->guest_rpte;
 620         asm volatile("lwsync" : : : "memory");
 621         hptep[0] = hpte[0];
 622         preempt_enable();
 623
 624         if (hpte[0] != vcpu->arch.pgfault_hpte[0] ||
 625             hpte[1] != vcpu->arch.pgfault_hpte[1])
 626                 return RESUME_GUEST;
 627
 628         /* Translate the logical address and get the page */
 629         psize = hpte_page_size(hpte[0], r);
 630         gpa = (r & HPTE_R_RPN & ~(psize - 1)) | (ea & (psize - 1));
 631         gfn = gpa >> PAGE_SHIFT;
 632         memslot = gfn_to_memslot(kvm, gfn);
 633
 634         /* No memslot means it's an emulated MMIO region */
 635         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
 636                 return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
 637                                               dsisr & DSISR_ISSTORE);
 638
 639         if (!kvm->arch.using_mmu_notifiers)
 640                 return -EFAULT;         /* should never get here */
 641
 642         /* used to check for invalidations in progress */
 643         mmu_seq = kvm->mmu_notifier_seq;
 644         smp_rmb();
 645
 646         is_io = 0;
 647         pfn = 0;
 648         page = NULL;
 649         pte_size = PAGE_SIZE;
 650         writing = (dsisr & DSISR_ISSTORE) != 0;
 651         /* If writing != 0, then the HPTE must allow writing, if we get here */
 652         write_ok = writing;
 653         hva = gfn_to_hva_memslot(memslot, gfn);
 654         npages = get_user_pages_fast(hva, 1, writing, pages);
 655         if (npages < 1) {
 656                 /* Check if it's an I/O mapping */
 657                 down_read(&current->mm->mmap_sem);
 658                 vma = find_vma(current->mm, hva);
 659                 if (vma && vma->vm_start <= hva && hva + psize <= vma->vm_end &&
 660                     (vma->vm_flags & VM_PFNMAP)) {
 661                         pfn = vma->vm_pgoff +
 662                                 ((hva - vma->vm_start) >> PAGE_SHIFT);
 663                         pte_size = psize;
 664                         is_io = hpte_cache_bits(pgprot_val(vma->vm_page_prot));
 665                         write_ok = vma->vm_flags & VM_WRITE;
 666                 }
 667                 up_read(&current->mm->mmap_sem);
 668                 if (!pfn)
 669                         return -EFAULT;
 670         } else {
 671                 page = pages[0];
 672                 if (PageHuge(page)) {
 673                         page = compound_head(page);
 674                         pte_size <<= compound_order(page);
 675                 }
 676                 /* if the guest wants write access, see if that is OK */
 677                 if (!writing && hpte_is_writable(r)) {
 678                         pte_t *ptep, pte;
 679
 680                         /*
 681                          * We need to protect against page table destruction
 682                          * while looking up and updating the pte.
 683                          */
 684                         rcu_read_lock_sched();
 685                         ptep = find_linux_pte_or_hugepte(current->mm->pgd,
 686                                                          hva, NULL);
 687                         if (ptep && pte_present(*ptep)) {
 688                                 pte = kvmppc_read_update_linux_pte(ptep, 1);
 689                                 if (pte_write(pte))
 690                                         write_ok = 1;
 691                         }
 692                         rcu_read_unlock_sched();
 693                 }
 694                 pfn = page_to_pfn(page);
 695         }
 696
 697         ret = -EFAULT;
 698         if (psize > pte_size)
 699                 goto out_put;
 700
 701         /* Check WIMG vs. the actual page we're accessing */
 702         if (!hpte_cache_flags_ok(r, is_io)) {
 703                 if (is_io)
 704                         return -EFAULT;
 705                 /*
 706                  * Allow guest to map emulated device memory as
 707                  * uncacheable, but actually make it cacheable.
 708                  */
 709                 r = (r & ~(HPTE_R_W|HPTE_R_I|HPTE_R_G)) | HPTE_R_M;
 710         }
 711
 712         /* Set the HPTE to point to pfn */
 713         r = (r & ~(HPTE_R_PP0 - pte_size)) | (pfn << PAGE_SHIFT);
 714         if (hpte_is_writable(r) && !write_ok)
 715                 r = hpte_make_readonly(r);
 716         ret = RESUME_GUEST;
 717         preempt_disable();
 718         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 719                 cpu_relax();
 720         if ((hptep[0] & ~HPTE_V_HVLOCK) != hpte[0] || hptep[1] != hpte[1] ||
 721             rev->guest_rpte != hpte[2])
 722                 /* HPTE has been changed under us; let the guest retry */
 723                 goto out_unlock;
 724         hpte[0] = (hpte[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
 725
 726         rmap = &memslot->arch.rmap[gfn - memslot->base_gfn];
 727         lock_rmap(rmap);
 728
 729         /* Check if we might have been invalidated; let the guest retry if so */
 730         ret = RESUME_GUEST;
 731         if (mmu_notifier_retry(vcpu->kvm, mmu_seq)) {
 732                 unlock_rmap(rmap);
 733                 goto out_unlock;
 734         }
 735
 736         /* Only set R/C in real HPTE if set in both *rmap and guest_rpte */
 737         rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
 738         r &= rcbits | ~(HPTE_R_R | HPTE_R_C);
 739
 740         if (hptep[0] & HPTE_V_VALID) {
 741                 /* HPTE was previously valid, so we need to invalidate it */
 742                 unlock_rmap(rmap);
 743                 hptep[0] |= HPTE_V_ABSENT;
 744                 kvmppc_invalidate_hpte(kvm, hptep, index);
 745                 /* don't lose previous R and C bits */
 746                 r |= hptep[1] & (HPTE_R_R | HPTE_R_C);
 747         } else {
 748                 kvmppc_add_revmap_chain(kvm, rev, rmap, index, 0);
 749         }
 750
 751         hptep[1] = r;
 752         eieio();
 753         hptep[0] = hpte[0];
 754         asm volatile("ptesync" : : : "memory");
 755         preempt_enable();
 756         if (page && hpte_is_writable(r))
 757                 SetPageDirty(page);
 758
 759  out_put:
 760         if (page) {
 761                 /*
 762                  * We drop pages[0] here, not page because page might
 763                  * have been set to the head page of a compound, but
 764                  * we have to drop the reference on the correct tail
 765                  * page to match the get inside gup()
 766                  */
 767                 put_page(pages[0]);
 768         }
 769         return ret;
 770
 771  out_unlock:
 772         hptep[0] &= ~HPTE_V_HVLOCK;
 773         preempt_enable();
 774         goto out_put;
 775 }
 776
 777 static void kvmppc_rmap_reset(struct kvm *kvm)
 778 {
 779         struct kvm_memslots *slots;
 780         struct kvm_memory_slot *memslot;
 781         int srcu_idx;
 782
 783         srcu_idx = srcu_read_lock(&kvm->srcu);
 784         slots = kvm->memslots;
 785         kvm_for_each_memslot(memslot, slots) {
 786                 /*
 787                  * This assumes it is acceptable to lose reference and
 788                  * change bits across a reset.
 789                  */
 790                 memset(memslot->arch.rmap, 0,
 791                        memslot->npages * sizeof(*memslot->arch.rmap));
 792         }
 793         srcu_read_unlock(&kvm->srcu, srcu_idx);
 794 }
 795
 796 static int kvm_handle_hva_range(struct kvm *kvm,
 797                                 unsigned long start,
 798                                 unsigned long end,
 799                                 int (*handler)(struct kvm *kvm,
 800                                                unsigned long *rmapp,
 801                                                unsigned long gfn))
 802 {
 803         int ret;
 804         int retval = 0;
 805         struct kvm_memslots *slots;
 806         struct kvm_memory_slot *memslot;
 807
 808         slots = kvm_memslots(kvm);
 809         kvm_for_each_memslot(memslot, slots) {
 810                 unsigned long hva_start, hva_end;
 811                 gfn_t gfn, gfn_end;
 812
 813                 hva_start = max(start, memslot->userspace_addr);
 814                 hva_end = min(end, memslot->userspace_addr +
 815                                         (memslot->npages << PAGE_SHIFT));
 816                 if (hva_start >= hva_end)
 817                         continue;
 818                 /*
 819                  * {gfn(page) | page intersects with [hva_start, hva_end)} =
 820                  * {gfn, gfn+1, ..., gfn_end-1}.
 821                  */
 822                 gfn = hva_to_gfn_memslot(hva_start, memslot);
 823                 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
 824
 825                 for (; gfn < gfn_end; ++gfn) {
 826                         gfn_t gfn_offset = gfn - memslot->base_gfn;
 827
 828                         ret = handler(kvm, &memslot->arch.rmap[gfn_offset], gfn);
 829                         retval |= ret;
 830                 }
 831         }
 832
 833         return retval;
 834 }
 835
 836 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 837                           int (*handler)(struct kvm *kvm, unsigned long *rmapp,
 838                                          unsigned long gfn))
 839 {
 840         return kvm_handle_hva_range(kvm, hva, hva + 1, handler);
 841 }
 842
 843 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
 844                            unsigned long gfn)
 845 {
 846         struct revmap_entry *rev = kvm->arch.revmap;
 847         unsigned long h, i, j;
 848         unsigned long *hptep;
 849         unsigned long ptel, psize, rcbits;
 850
 851         for (;;) {
 852                 lock_rmap(rmapp);
 853                 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 854                         unlock_rmap(rmapp);
 855                         break;
 856                 }
 857
 858                 /*
 859                  * To avoid an ABBA deadlock with the HPTE lock bit,
 860                  * we can't spin on the HPTE lock while holding the
 861                  * rmap chain lock.
 862                  */
 863                 i = *rmapp & KVMPPC_RMAP_INDEX;
 864                 hptep = (unsigned long *) (kvm->arch.hpt_virt + (i << 4));
 865                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 866                         /* unlock rmap before spinning on the HPTE lock */
 867                         unlock_rmap(rmapp);
 868                         while (hptep[0] & HPTE_V_HVLOCK)
 869                                 cpu_relax();
 870                         continue;
 871                 }
 872                 j = rev[i].forw;
 873                 if (j == i) {
 874                         /* chain is now empty */
 875                         *rmapp &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
 876                 } else {
 877                         /* remove i from chain */
 878                         h = rev[i].back;
 879                         rev[h].forw = j;
 880                         rev[j].back = h;
 881                         rev[i].forw = rev[i].back = i;
 882                         *rmapp = (*rmapp & ~KVMPPC_RMAP_INDEX) | j;
 883                 }
 884
 885                 /* Now check and modify the HPTE */
 886                 ptel = rev[i].guest_rpte;
 887                 psize = hpte_page_size(hptep[0], ptel);
 888                 if ((hptep[0] & HPTE_V_VALID) &&
 889                     hpte_rpn(ptel, psize) == gfn) {
 890                         if (kvm->arch.using_mmu_notifiers)
 891                                 hptep[0] |= HPTE_V_ABSENT;
 892                         kvmppc_invalidate_hpte(kvm, hptep, i);
 893                         /* Harvest R and C */
 894                         rcbits = hptep[1] & (HPTE_R_R | HPTE_R_C);
 895                         *rmapp |= rcbits << KVMPPC_RMAP_RC_SHIFT;
 896                         if (rcbits & ~rev[i].guest_rpte) {
 897                                 rev[i].guest_rpte = ptel | rcbits;
 898                                 note_hpte_modification(kvm, &rev[i]);
 899                         }
 900                 }
 901                 unlock_rmap(rmapp);
 902                 hptep[0] &= ~HPTE_V_HVLOCK;
 903         }
 904         return 0;
 905 }
 906
 907 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 908 {
 909         if (kvm->arch.using_mmu_notifiers)
 910                 kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 911         return 0;
 912 }
 913
 914 int kvm_unmap_hva_range(struct kvm *kvm, unsigned long start, unsigned long end)
 915 {
 916         if (kvm->arch.using_mmu_notifiers)
 917                 kvm_handle_hva_range(kvm, start, end, kvm_unmap_rmapp);
 918         return 0;
 919 }
 920
 921 void kvmppc_core_flush_memslot(struct kvm *kvm, struct kvm_memory_slot *memslot)
 922 {
 923         unsigned long *rmapp;
 924         unsigned long gfn;
 925         unsigned long n;
 926
 927         rmapp = memslot->arch.rmap;
 928         gfn = memslot->base_gfn;
 929         for (n = memslot->npages; n; --n) {
 930                 /*
 931                  * Testing the present bit without locking is OK because
 932                  * the memslot has been marked invalid already, and hence
 933                  * no new HPTEs referencing this page can be created,
 934                  * thus the present bit can't go from 0 to 1.
 935                  */
 936                 if (*rmapp & KVMPPC_RMAP_PRESENT)
 937                         kvm_unmap_rmapp(kvm, rmapp, gfn);
 938                 ++rmapp;
 939                 ++gfn;
 940         }
 941 }
 942
 943 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
 944                          unsigned long gfn)
 945 {
 946         struct revmap_entry *rev = kvm->arch.revmap;
 947         unsigned long head, i, j;
 948         unsigned long *hptep;
 949         int ret = 0;
 950
 951  retry:
 952         lock_rmap(rmapp);
 953         if (*rmapp & KVMPPC_RMAP_REFERENCED) {
 954                 *rmapp &= ~KVMPPC_RMAP_REFERENCED;
 955                 ret = 1;
 956         }
 957         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 958                 unlock_rmap(rmapp);
 959                 return ret;
 960         }
 961
 962         i = head = *rmapp & KVMPPC_RMAP_INDEX;
 963         do {
 964                 hptep = (unsigned long *) (kvm->arch.hpt_virt + (i << 4));
 965                 j = rev[i].forw;
 966
 967                 /* If this HPTE isn't referenced, ignore it */
 968                 if (!(hptep[1] & HPTE_R_R))
 969                         continue;
 970
 971                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 972                         /* unlock rmap before spinning on the HPTE lock */
 973                         unlock_rmap(rmapp);
 974                         while (hptep[0] & HPTE_V_HVLOCK)
 975                                 cpu_relax();
 976                         goto retry;
 977                 }
 978
 979                 /* Now check and modify the HPTE */
 980                 if ((hptep[0] & HPTE_V_VALID) && (hptep[1] & HPTE_R_R)) {
 981                         kvmppc_clear_ref_hpte(kvm, hptep, i);
 982                         if (!(rev[i].guest_rpte & HPTE_R_R)) {
 983                                 rev[i].guest_rpte |= HPTE_R_R;
 984                                 note_hpte_modification(kvm, &rev[i]);
 985                         }
 986                         ret = 1;
 987                 }
 988                 hptep[0] &= ~HPTE_V_HVLOCK;
 989         } while ((i = j) != head);
 990
 991         unlock_rmap(rmapp);
 992         return ret;
 993 }
 994
 995 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 996 {
 997         if (!kvm->arch.using_mmu_notifiers)
 998                 return 0;
 999         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
1000 }
1001
1002 static int kvm_test_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
1003                               unsigned long gfn)
1004 {
1005         struct revmap_entry *rev = kvm->arch.revmap;
1006         unsigned long head, i, j;
1007         unsigned long *hp;
1008         int ret = 1;
1009
1010         if (*rmapp & KVMPPC_RMAP_REFERENCED)
1011                 return 1;
1012
1013         lock_rmap(rmapp);
1014         if (*rmapp & KVMPPC_RMAP_REFERENCED)
1015                 goto out;
1016
1017         if (*rmapp & KVMPPC_RMAP_PRESENT) {
1018                 i = head = *rmapp & KVMPPC_RMAP_INDEX;
1019                 do {
1020                         hp = (unsigned long *)(kvm->arch.hpt_virt + (i << 4));
1021                         j = rev[i].forw;
1022                         if (hp[1] & HPTE_R_R)
1023                                 goto out;
1024                 } while ((i = j) != head);
1025         }
1026         ret = 0;
1027
1028  out:
1029         unlock_rmap(rmapp);
1030         return ret;
1031 }
1032
1033 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1034 {
1035         if (!kvm->arch.using_mmu_notifiers)
1036                 return 0;
1037         return kvm_handle_hva(kvm, hva, kvm_test_age_rmapp);
1038 }
1039
1040 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1041 {
1042         if (!kvm->arch.using_mmu_notifiers)
1043                 return;
1044         kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
1045 }
1046
1047 static int kvm_test_clear_dirty(struct kvm *kvm, unsigned long *rmapp)
1048 {
1049         struct revmap_entry *rev = kvm->arch.revmap;
1050         unsigned long head, i, j;
1051         unsigned long *hptep;
1052         int ret = 0;
1053
1054  retry:
1055         lock_rmap(rmapp);
1056         if (*rmapp & KVMPPC_RMAP_CHANGED) {
1057                 *rmapp &= ~KVMPPC_RMAP_CHANGED;
1058                 ret = 1;
1059         }
1060         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
1061                 unlock_rmap(rmapp);
1062                 return ret;
1063         }
1064
1065         i = head = *rmapp & KVMPPC_RMAP_INDEX;
1066         do {
1067                 hptep = (unsigned long *) (kvm->arch.hpt_virt + (i << 4));
1068                 j = rev[i].forw;
1069
1070                 if (!(hptep[1] & HPTE_R_C))
1071                         continue;
1072
1073                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
1074                         /* unlock rmap before spinning on the HPTE lock */
1075                         unlock_rmap(rmapp);
1076                         while (hptep[0] & HPTE_V_HVLOCK)
1077                                 cpu_relax();
1078                         goto retry;
1079                 }
1080
1081                 /* Now check and modify the HPTE */
1082                 if ((hptep[0] & HPTE_V_VALID) && (hptep[1] & HPTE_R_C)) {
1083                         /* need to make it temporarily absent to clear C */
1084                         hptep[0] |= HPTE_V_ABSENT;
1085                         kvmppc_invalidate_hpte(kvm, hptep, i);
1086                         hptep[1] &= ~HPTE_R_C;
1087                         eieio();
1088                         hptep[0] = (hptep[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
1089                         if (!(rev[i].guest_rpte & HPTE_R_C)) {
1090                                 rev[i].guest_rpte |= HPTE_R_C;
1091                                 note_hpte_modification(kvm, &rev[i]);
1092                         }
1093                         ret = 1;
1094                 }
1095                 hptep[0] &= ~HPTE_V_HVLOCK;
1096         } while ((i = j) != head);
1097
1098         unlock_rmap(rmapp);
1099         return ret;
1100 }
1101
1102 static void harvest_vpa_dirty(struct kvmppc_vpa *vpa,
1103                               struct kvm_memory_slot *memslot,
1104                               unsigned long *map)
1105 {
1106         unsigned long gfn;
1107
1108         if (!vpa->dirty || !vpa->pinned_addr)
1109                 return;
1110         gfn = vpa->gpa >> PAGE_SHIFT;
1111         if (gfn < memslot->base_gfn ||
1112             gfn >= memslot->base_gfn + memslot->npages)
1113                 return;
1114
1115         vpa->dirty = false;
1116         if (map)
1117                 __set_bit_le(gfn - memslot->base_gfn, map);
1118 }
1119
1120 long kvmppc_hv_get_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot,
1121                              unsigned long *map)
1122 {
1123         unsigned long i;
1124         unsigned long *rmapp;
1125         struct kvm_vcpu *vcpu;
1126
1127         preempt_disable();
1128         rmapp = memslot->arch.rmap;
1129         for (i = 0; i < memslot->npages; ++i) {
1130                 if (kvm_test_clear_dirty(kvm, rmapp) && map)
1131                         __set_bit_le(i, map);
1132                 ++rmapp;
1133         }
1134
1135         /* Harvest dirty bits from VPA and DTL updates */
1136         /* Note: we never modify the SLB shadow buffer areas */
1137         kvm_for_each_vcpu(i, vcpu, kvm) {
1138                 spin_lock(&vcpu->arch.vpa_update_lock);
1139                 harvest_vpa_dirty(&vcpu->arch.vpa, memslot, map);
1140                 harvest_vpa_dirty(&vcpu->arch.dtl, memslot, map);
1141                 spin_unlock(&vcpu->arch.vpa_update_lock);
1142         }
1143         preempt_enable();
1144         return 0;
1145 }
1146
1147 void *kvmppc_pin_guest_page(struct kvm *kvm, unsigned long gpa,
1148                             unsigned long *nb_ret)
1149 {
1150         struct kvm_memory_slot *memslot;
1151         unsigned long gfn = gpa >> PAGE_SHIFT;
1152         struct page *page, *pages[1];
1153         int npages;
1154         unsigned long hva, offset;
1155         unsigned long pa;
1156         unsigned long *physp;
1157         int srcu_idx;
1158
1159         srcu_idx = srcu_read_lock(&kvm->srcu);
1160         memslot = gfn_to_memslot(kvm, gfn);
1161         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
1162                 goto err;
1163         if (!kvm->arch.using_mmu_notifiers) {
1164                 physp = memslot->arch.slot_phys;
1165                 if (!physp)
1166                         goto err;
1167                 physp += gfn - memslot->base_gfn;
1168                 pa = *physp;
1169                 if (!pa) {
1170                         if (kvmppc_get_guest_page(kvm, gfn, memslot,
1171                                                   PAGE_SIZE) < 0)
1172                                 goto err;
1173                         pa = *physp;
1174                 }
1175                 page = pfn_to_page(pa >> PAGE_SHIFT);
1176                 get_page(page);
1177         } else {
1178                 hva = gfn_to_hva_memslot(memslot, gfn);
1179                 npages = get_user_pages_fast(hva, 1, 1, pages);
1180                 if (npages < 1)
1181                         goto err;
1182                 page = pages[0];
1183         }
1184         srcu_read_unlock(&kvm->srcu, srcu_idx);
1185
1186         offset = gpa & (PAGE_SIZE - 1);
1187         if (nb_ret)
1188                 *nb_ret = PAGE_SIZE - offset;
1189         return page_address(page) + offset;
1190
1191  err:
1192         srcu_read_unlock(&kvm->srcu, srcu_idx);
1193         return NULL;
1194 }
1195
1196 void kvmppc_unpin_guest_page(struct kvm *kvm, void *va, unsigned long gpa,
1197                              bool dirty)
1198 {
1199         struct page *page = virt_to_page(va);
1200         struct kvm_memory_slot *memslot;
1201         unsigned long gfn;
1202         unsigned long *rmap;
1203         int srcu_idx;
1204
1205         put_page(page);
1206
1207         if (!dirty || !kvm->arch.using_mmu_notifiers)
1208                 return;
1209
1210         /* We need to mark this page dirty in the rmap chain */
1211         gfn = gpa >> PAGE_SHIFT;
1212         srcu_idx = srcu_read_lock(&kvm->srcu);
1213         memslot = gfn_to_memslot(kvm, gfn);
1214         if (memslot) {
1215                 rmap = &memslot->arch.rmap[gfn - memslot->base_gfn];
1216                 lock_rmap(rmap);
1217                 *rmap |= KVMPPC_RMAP_CHANGED;
1218                 unlock_rmap(rmap);
1219         }
1220         srcu_read_unlock(&kvm->srcu, srcu_idx);
1221 }
1222
1223 /*
1224  * Functions for reading and writing the hash table via reads and
1225  * writes on a file descriptor.
1226  *
1227  * Reads return the guest view of the hash table, which has to be
1228  * pieced together from the real hash table and the guest_rpte
1229  * values in the revmap array.
1230  *
1231  * On writes, each HPTE written is considered in turn, and if it
1232  * is valid, it is written to the HPT as if an H_ENTER with the
1233  * exact flag set was done.  When the invalid count is non-zero
1234  * in the header written to the stream, the kernel will make
1235  * sure that that many HPTEs are invalid, and invalidate them
1236  * if not.
1237  */
1238
1239 struct kvm_htab_ctx {
1240         unsigned long   index;
1241         unsigned long   flags;
1242         struct kvm      *kvm;
1243         int             first_pass;
1244 };
1245
1246 #define HPTE_SIZE       (2 * sizeof(unsigned long))
1247
1248 /*
1249  * Returns 1 if this HPT entry has been modified or has pending
1250  * R/C bit changes.
1251  */
1252 static int hpte_dirty(struct revmap_entry *revp, unsigned long *hptp)
1253 {
1254         unsigned long rcbits_unset;
1255
1256         if (revp->guest_rpte & HPTE_GR_MODIFIED)
1257                 return 1;
1258
1259         /* Also need to consider changes in reference and changed bits */
1260         rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1261         if ((hptp[0] & HPTE_V_VALID) && (hptp[1] & rcbits_unset))
1262                 return 1;
1263
1264         return 0;
1265 }
1266
1267 static long record_hpte(unsigned long flags, unsigned long *hptp,
1268                         unsigned long *hpte, struct revmap_entry *revp,
1269                         int want_valid, int first_pass)
1270 {
1271         unsigned long v, r;
1272         unsigned long rcbits_unset;
1273         int ok = 1;
1274         int valid, dirty;
1275
1276         /* Unmodified entries are uninteresting except on the first pass */
1277         dirty = hpte_dirty(revp, hptp);
1278         if (!first_pass && !dirty)
1279                 return 0;
1280
1281         valid = 0;
1282         if (hptp[0] & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1283                 valid = 1;
1284                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) &&
1285                     !(hptp[0] & HPTE_V_BOLTED))
1286                         valid = 0;
1287         }
1288         if (valid != want_valid)
1289                 return 0;
1290
1291         v = r = 0;
1292         if (valid || dirty) {
1293                 /* lock the HPTE so it's stable and read it */
1294                 preempt_disable();
1295                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
1296                         cpu_relax();
1297                 v = hptp[0];
1298
1299                 /* re-evaluate valid and dirty from synchronized HPTE value */
1300                 valid = !!(v & HPTE_V_VALID);
1301                 dirty = !!(revp->guest_rpte & HPTE_GR_MODIFIED);
1302
1303                 /* Harvest R and C into guest view if necessary */
1304                 rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1305                 if (valid && (rcbits_unset & hptp[1])) {
1306                         revp->guest_rpte |= (hptp[1] & (HPTE_R_R | HPTE_R_C)) |
1307                                 HPTE_GR_MODIFIED;
1308                         dirty = 1;
1309                 }
1310
1311                 if (v & HPTE_V_ABSENT) {
1312                         v &= ~HPTE_V_ABSENT;
1313                         v |= HPTE_V_VALID;
1314                         valid = 1;
1315                 }
1316                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) && !(v & HPTE_V_BOLTED))
1317                         valid = 0;
1318
1319                 r = revp->guest_rpte;
1320                 /* only clear modified if this is the right sort of entry */
1321                 if (valid == want_valid && dirty) {
1322                         r &= ~HPTE_GR_MODIFIED;
1323                         revp->guest_rpte = r;
1324                 }
1325                 asm volatile(PPC_RELEASE_BARRIER "" : : : "memory");
1326                 hptp[0] &= ~HPTE_V_HVLOCK;
1327                 preempt_enable();
1328                 if (!(valid == want_valid && (first_pass || dirty)))
1329                         ok = 0;
1330         }
1331         hpte[0] = v;
1332         hpte[1] = r;
1333         return ok;
1334 }
1335
1336 static ssize_t kvm_htab_read(struct file *file, char __user *buf,
1337                              size_t count, loff_t *ppos)
1338 {
1339         struct kvm_htab_ctx *ctx = file->private_data;
1340         struct kvm *kvm = ctx->kvm;
1341         struct kvm_get_htab_header hdr;
1342         unsigned long *hptp;
1343         struct revmap_entry *revp;
1344         unsigned long i, nb, nw;
1345         unsigned long __user *lbuf;
1346         struct kvm_get_htab_header __user *hptr;
1347         unsigned long flags;
1348         int first_pass;
1349         unsigned long hpte[2];
1350
1351         if (!access_ok(VERIFY_WRITE, buf, count))
1352                 return -EFAULT;
1353
1354         first_pass = ctx->first_pass;
1355         flags = ctx->flags;
1356
1357         i = ctx->index;
1358         hptp = (unsigned long *)(kvm->arch.hpt_virt + (i * HPTE_SIZE));
1359         revp = kvm->arch.revmap + i;
1360         lbuf = (unsigned long __user *)buf;
1361
1362         nb = 0;
1363         while (nb + sizeof(hdr) + HPTE_SIZE < count) {
1364                 /* Initialize header */
1365                 hptr = (struct kvm_get_htab_header __user *)buf;
1366                 hdr.n_valid = 0;
1367                 hdr.n_invalid = 0;
1368                 nw = nb;
1369                 nb += sizeof(hdr);
1370                 lbuf = (unsigned long __user *)(buf + sizeof(hdr));
1371
1372                 /* Skip uninteresting entries, i.e. clean on not-first pass */
1373                 if (!first_pass) {
1374                         while (i < kvm->arch.hpt_npte &&
1375                                !hpte_dirty(revp, hptp)) {
1376                                 ++i;
1377                                 hptp += 2;
1378                                 ++revp;
1379                         }
1380                 }
1381                 hdr.index = i;
1382
1383                 /* Grab a series of valid entries */
1384                 while (i < kvm->arch.hpt_npte &&
1385                        hdr.n_valid < 0xffff &&
1386                        nb + HPTE_SIZE < count &&
1387                        record_hpte(flags, hptp, hpte, revp, 1, first_pass)) {
1388                         /* valid entry, write it out */
1389                         ++hdr.n_valid;
1390                         if (__put_user(hpte[0], lbuf) ||
1391                             __put_user(hpte[1], lbuf + 1))
1392                                 return -EFAULT;
1393                         nb += HPTE_SIZE;
1394                         lbuf += 2;
1395                         ++i;
1396                         hptp += 2;
1397                         ++revp;
1398                 }
1399                 /* Now skip invalid entries while we can */
1400                 while (i < kvm->arch.hpt_npte &&
1401                        hdr.n_invalid < 0xffff &&
1402                        record_hpte(flags, hptp, hpte, revp, 0, first_pass)) {
1403                         /* found an invalid entry */
1404                         ++hdr.n_invalid;
1405                         ++i;
1406                         hptp += 2;
1407                         ++revp;
1408                 }
1409
1410                 if (hdr.n_valid || hdr.n_invalid) {
1411                         /* write back the header */
1412                         if (__copy_to_user(hptr, &hdr, sizeof(hdr)))
1413                                 return -EFAULT;
1414                         nw = nb;
1415                         buf = (char __user *)lbuf;
1416                 } else {
1417                         nb = nw;
1418                 }
1419
1420                 /* Check if we've wrapped around the hash table */
1421                 if (i >= kvm->arch.hpt_npte) {
1422                         i = 0;
1423                         ctx->first_pass = 0;
1424                         break;
1425                 }
1426         }
1427
1428         ctx->index = i;
1429
1430         return nb;
1431 }
1432
1433 static ssize_t kvm_htab_write(struct file *file, const char __user *buf,
1434                               size_t count, loff_t *ppos)
1435 {
1436         struct kvm_htab_ctx *ctx = file->private_data;
1437         struct kvm *kvm = ctx->kvm;
1438         struct kvm_get_htab_header hdr;
1439         unsigned long i, j;
1440         unsigned long v, r;
1441         unsigned long __user *lbuf;
1442         unsigned long *hptp;
1443         unsigned long tmp[2];
1444         ssize_t nb;
1445         long int err, ret;
1446         int rma_setup;
1447
1448         if (!access_ok(VERIFY_READ, buf, count))
1449                 return -EFAULT;
1450
1451         /* lock out vcpus from running while we're doing this */
1452         mutex_lock(&kvm->lock);
1453         rma_setup = kvm->arch.rma_setup_done;
1454         if (rma_setup) {
1455                 kvm->arch.rma_setup_done = 0;   /* temporarily */
1456                 /* order rma_setup_done vs. vcpus_running */
1457                 smp_mb();
1458                 if (atomic_read(&kvm->arch.vcpus_running)) {
1459                         kvm->arch.rma_setup_done = 1;
1460                         mutex_unlock(&kvm->lock);
1461                         return -EBUSY;
1462                 }
1463         }
1464
1465         err = 0;
1466         for (nb = 0; nb + sizeof(hdr) <= count; ) {
1467                 err = -EFAULT;
1468                 if (__copy_from_user(&hdr, buf, sizeof(hdr)))
1469                         break;
1470
1471                 err = 0;
1472                 if (nb + hdr.n_valid * HPTE_SIZE > count)
1473                         break;
1474
1475                 nb += sizeof(hdr);
1476                 buf += sizeof(hdr);
1477
1478                 err = -EINVAL;
1479                 i = hdr.index;
1480                 if (i >= kvm->arch.hpt_npte ||
1481                     i + hdr.n_valid + hdr.n_invalid > kvm->arch.hpt_npte)
1482                         break;
1483
1484                 hptp = (unsigned long *)(kvm->arch.hpt_virt + (i * HPTE_SIZE));
1485                 lbuf = (unsigned long __user *)buf;
1486                 for (j = 0; j < hdr.n_valid; ++j) {
1487                         err = -EFAULT;
1488                         if (__get_user(v, lbuf) || __get_user(r, lbuf + 1))
1489                                 goto out;
1490                         err = -EINVAL;
1491                         if (!(v & HPTE_V_VALID))
1492                                 goto out;
1493                         lbuf += 2;
1494                         nb += HPTE_SIZE;
1495
1496                         if (hptp[0] & (HPTE_V_VALID | HPTE_V_ABSENT))
1497                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1498                         err = -EIO;
1499                         ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, i, v, r,
1500                                                          tmp);
1501                         if (ret != H_SUCCESS) {
1502                                 pr_err("kvm_htab_write ret %ld i=%ld v=%lx "
1503                                        "r=%lx\n", ret, i, v, r);
1504                                 goto out;
1505                         }
1506                         if (!rma_setup && is_vrma_hpte(v)) {
1507                                 unsigned long psize = hpte_page_size(v, r);
1508                                 unsigned long senc = slb_pgsize_encoding(psize);
1509                                 unsigned long lpcr;
1510
1511                                 kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
1512                                         (VRMA_VSID << SLB_VSID_SHIFT_1T);
1513                                 lpcr = kvm->arch.lpcr & ~LPCR_VRMASD;
1514                                 lpcr |= senc << (LPCR_VRMASD_SH - 4);
1515                                 kvm->arch.lpcr = lpcr;
1516                                 rma_setup = 1;
1517                         }
1518                         ++i;
1519                         hptp += 2;
1520                 }
1521
1522                 for (j = 0; j < hdr.n_invalid; ++j) {
1523                         if (hptp[0] & (HPTE_V_VALID | HPTE_V_ABSENT))
1524                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1525                         ++i;
1526                         hptp += 2;
1527                 }
1528                 err = 0;
1529         }
1530
1531  out:
1532         /* Order HPTE updates vs. rma_setup_done */
1533         smp_wmb();
1534         kvm->arch.rma_setup_done = rma_setup;
1535         mutex_unlock(&kvm->lock);
1536
1537         if (err)
1538                 return err;
1539         return nb;
1540 }
1541
1542 static int kvm_htab_release(struct inode *inode, struct file *filp)
1543 {
1544         struct kvm_htab_ctx *ctx = filp->private_data;
1545
1546         filp->private_data = NULL;
1547         if (!(ctx->flags & KVM_GET_HTAB_WRITE))
1548                 atomic_dec(&ctx->kvm->arch.hpte_mod_interest);
1549         kvm_put_kvm(ctx->kvm);
1550         kfree(ctx);
1551         return 0;
1552 }
1553
1554 static const struct file_operations kvm_htab_fops = {
1555         .read           = kvm_htab_read,
1556         .write          = kvm_htab_write,
1557         .llseek         = default_llseek,
1558         .release        = kvm_htab_release,
1559 };
1560
1561 int kvm_vm_ioctl_get_htab_fd(struct kvm *kvm, struct kvm_get_htab_fd *ghf)
1562 {
1563         int ret;
1564         struct kvm_htab_ctx *ctx;
1565         int rwflag;
1566
1567         /* reject flags we don't recognize */
1568         if (ghf->flags & ~(KVM_GET_HTAB_BOLTED_ONLY | KVM_GET_HTAB_WRITE))
1569                 return -EINVAL;
1570         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1571         if (!ctx)
1572                 return -ENOMEM;
1573         kvm_get_kvm(kvm);
1574         ctx->kvm = kvm;
1575         ctx->index = ghf->start_index;
1576         ctx->flags = ghf->flags;
1577         ctx->first_pass = 1;
1578
1579         rwflag = (ghf->flags & KVM_GET_HTAB_WRITE) ? O_WRONLY : O_RDONLY;
1580         ret = anon_inode_getfd("kvm-htab", &kvm_htab_fops, ctx, rwflag);
1581         if (ret < 0) {
1582                 kvm_put_kvm(kvm);
1583                 return ret;
1584         }
1585
1586         if (rwflag == O_RDONLY) {
1587                 mutex_lock(&kvm->slots_lock);
1588                 atomic_inc(&kvm->arch.hpte_mod_interest);
1589                 /* make sure kvmppc_do_h_enter etc. see the increment */
1590                 synchronize_srcu_expedited(&kvm->srcu);
1591                 mutex_unlock(&kvm->slots_lock);
1592         }
1593
1594         return ret;
1595 }
1596
1597 void kvmppc_mmu_book3s_hv_init(struct kvm_vcpu *vcpu)
1598 {
1599         struct kvmppc_mmu *mmu = &vcpu->arch.mmu;
1600
1601         if (cpu_has_feature(CPU_FTR_ARCH_206))
1602                 vcpu->arch.slb_nr = 32;         /* POWER7 */
1603         else
1604                 vcpu->arch.slb_nr = 64;
1605
1606         mmu->xlate = kvmppc_mmu_book3s_64_hv_xlate;
1607         mmu->reset_msr = kvmppc_mmu_book3s_64_hv_reset_msr;
1608
1609         vcpu->arch.hflags |= BOOK3S_HFLAG_SLB;
1610 }