Pileus Git - ~andy/linux/blob - 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                         unsigned int hugepage_shift;
 679                         pte_t *ptep, pte;
 680
 681                         /*
 682                          * We need to protect against page table destruction
 683                          * while looking up and updating the pte.
 684                          */
 685                         rcu_read_lock_sched();
 686                         ptep = find_linux_pte_or_hugepte(current->mm->pgd,
 687                                                          hva, &hugepage_shift);
 688                         if (ptep) {
 689                                 pte = kvmppc_read_update_linux_pte(ptep, 1,
 690                                                            hugepage_shift);
 691                                 if (pte_write(pte))
 692                                         write_ok = 1;
 693                         }
 694                         rcu_read_unlock_sched();
 695                 }
 696                 pfn = page_to_pfn(page);
 697         }
 698
 699         ret = -EFAULT;
 700         if (psize > pte_size)
 701                 goto out_put;
 702
 703         /* Check WIMG vs. the actual page we're accessing */
 704         if (!hpte_cache_flags_ok(r, is_io)) {
 705                 if (is_io)
 706                         return -EFAULT;
 707                 /*
 708                  * Allow guest to map emulated device memory as
 709                  * uncacheable, but actually make it cacheable.
 710                  */
 711                 r = (r & ~(HPTE_R_W|HPTE_R_I|HPTE_R_G)) | HPTE_R_M;
 712         }
 713
 714         /* Set the HPTE to point to pfn */
 715         r = (r & ~(HPTE_R_PP0 - pte_size)) | (pfn << PAGE_SHIFT);
 716         if (hpte_is_writable(r) && !write_ok)
 717                 r = hpte_make_readonly(r);
 718         ret = RESUME_GUEST;
 719         preempt_disable();
 720         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 721                 cpu_relax();
 722         if ((hptep[0] & ~HPTE_V_HVLOCK) != hpte[0] || hptep[1] != hpte[1] ||
 723             rev->guest_rpte != hpte[2])
 724                 /* HPTE has been changed under us; let the guest retry */
 725                 goto out_unlock;
 726         hpte[0] = (hpte[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
 727
 728         rmap = &memslot->arch.rmap[gfn - memslot->base_gfn];
 729         lock_rmap(rmap);
 730
 731         /* Check if we might have been invalidated; let the guest retry if so */
 732         ret = RESUME_GUEST;
 733         if (mmu_notifier_retry(vcpu->kvm, mmu_seq)) {
 734                 unlock_rmap(rmap);
 735                 goto out_unlock;
 736         }
 737
 738         /* Only set R/C in real HPTE if set in both *rmap and guest_rpte */
 739         rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
 740         r &= rcbits | ~(HPTE_R_R | HPTE_R_C);
 741
 742         if (hptep[0] & HPTE_V_VALID) {
 743                 /* HPTE was previously valid, so we need to invalidate it */
 744                 unlock_rmap(rmap);
 745                 hptep[0] |= HPTE_V_ABSENT;
 746                 kvmppc_invalidate_hpte(kvm, hptep, index);
 747                 /* don't lose previous R and C bits */
 748                 r |= hptep[1] & (HPTE_R_R | HPTE_R_C);
 749         } else {
 750                 kvmppc_add_revmap_chain(kvm, rev, rmap, index, 0);
 751         }
 752
 753         hptep[1] = r;
 754         eieio();
 755         hptep[0] = hpte[0];
 756         asm volatile("ptesync" : : : "memory");
 757         preempt_enable();
 758         if (page && hpte_is_writable(r))
 759                 SetPageDirty(page);
 760
 761  out_put:
 762         if (page) {
 763                 /*
 764                  * We drop pages[0] here, not page because page might
 765                  * have been set to the head page of a compound, but
 766                  * we have to drop the reference on the correct tail
 767                  * page to match the get inside gup()
 768                  */
 769                 put_page(pages[0]);
 770         }
 771         return ret;
 772
 773  out_unlock:
 774         hptep[0] &= ~HPTE_V_HVLOCK;
 775         preempt_enable();
 776         goto out_put;
 777 }
 778
 779 static void kvmppc_rmap_reset(struct kvm *kvm)
 780 {
 781         struct kvm_memslots *slots;
 782         struct kvm_memory_slot *memslot;
 783         int srcu_idx;
 784
 785         srcu_idx = srcu_read_lock(&kvm->srcu);
 786         slots = kvm->memslots;
 787         kvm_for_each_memslot(memslot, slots) {
 788                 /*
 789                  * This assumes it is acceptable to lose reference and
 790                  * change bits across a reset.
 791                  */
 792                 memset(memslot->arch.rmap, 0,
 793                        memslot->npages * sizeof(*memslot->arch.rmap));
 794         }
 795         srcu_read_unlock(&kvm->srcu, srcu_idx);
 796 }
 797
 798 static int kvm_handle_hva_range(struct kvm *kvm,
 799                                 unsigned long start,
 800                                 unsigned long end,
 801                                 int (*handler)(struct kvm *kvm,
 802                                                unsigned long *rmapp,
 803                                                unsigned long gfn))
 804 {
 805         int ret;
 806         int retval = 0;
 807         struct kvm_memslots *slots;
 808         struct kvm_memory_slot *memslot;
 809
 810         slots = kvm_memslots(kvm);
 811         kvm_for_each_memslot(memslot, slots) {
 812                 unsigned long hva_start, hva_end;
 813                 gfn_t gfn, gfn_end;
 814
 815                 hva_start = max(start, memslot->userspace_addr);
 816                 hva_end = min(end, memslot->userspace_addr +
 817                                         (memslot->npages << PAGE_SHIFT));
 818                 if (hva_start >= hva_end)
 819                         continue;
 820                 /*
 821                  * {gfn(page) | page intersects with [hva_start, hva_end)} =
 822                  * {gfn, gfn+1, ..., gfn_end-1}.
 823                  */
 824                 gfn = hva_to_gfn_memslot(hva_start, memslot);
 825                 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
 826
 827                 for (; gfn < gfn_end; ++gfn) {
 828                         gfn_t gfn_offset = gfn - memslot->base_gfn;
 829
 830                         ret = handler(kvm, &memslot->arch.rmap[gfn_offset], gfn);
 831                         retval |= ret;
 832                 }
 833         }
 834
 835         return retval;
 836 }
 837
 838 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 839                           int (*handler)(struct kvm *kvm, unsigned long *rmapp,
 840                                          unsigned long gfn))
 841 {
 842         return kvm_handle_hva_range(kvm, hva, hva + 1, handler);
 843 }
 844
 845 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
 846                            unsigned long gfn)
 847 {
 848         struct revmap_entry *rev = kvm->arch.revmap;
 849         unsigned long h, i, j;
 850         unsigned long *hptep;
 851         unsigned long ptel, psize, rcbits;
 852
 853         for (;;) {
 854                 lock_rmap(rmapp);
 855                 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 856                         unlock_rmap(rmapp);
 857                         break;
 858                 }
 859
 860                 /*
 861                  * To avoid an ABBA deadlock with the HPTE lock bit,
 862                  * we can't spin on the HPTE lock while holding the
 863                  * rmap chain lock.
 864                  */
 865                 i = *rmapp & KVMPPC_RMAP_INDEX;
 866                 hptep = (unsigned long *) (kvm->arch.hpt_virt + (i << 4));
 867                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 868                         /* unlock rmap before spinning on the HPTE lock */
 869                         unlock_rmap(rmapp);
 870                         while (hptep[0] & HPTE_V_HVLOCK)
 871                                 cpu_relax();
 872                         continue;
 873                 }
 874                 j = rev[i].forw;
 875                 if (j == i) {
 876                         /* chain is now empty */
 877                         *rmapp &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
 878                 } else {
 879                         /* remove i from chain */
 880                         h = rev[i].back;
 881                         rev[h].forw = j;
 882                         rev[j].back = h;
 883                         rev[i].forw = rev[i].back = i;
 884                         *rmapp = (*rmapp & ~KVMPPC_RMAP_INDEX) | j;
 885                 }
 886
 887                 /* Now check and modify the HPTE */
 888                 ptel = rev[i].guest_rpte;
 889                 psize = hpte_page_size(hptep[0], ptel);
 890                 if ((hptep[0] & HPTE_V_VALID) &&
 891                     hpte_rpn(ptel, psize) == gfn) {
 892                         if (kvm->arch.using_mmu_notifiers)
 893                                 hptep[0] |= HPTE_V_ABSENT;
 894                         kvmppc_invalidate_hpte(kvm, hptep, i);
 895                         /* Harvest R and C */
 896                         rcbits = hptep[1] & (HPTE_R_R | HPTE_R_C);
 897                         *rmapp |= rcbits << KVMPPC_RMAP_RC_SHIFT;
 898                         if (rcbits & ~rev[i].guest_rpte) {
 899                                 rev[i].guest_rpte = ptel | rcbits;
 900                                 note_hpte_modification(kvm, &rev[i]);
 901                         }
 902                 }
 903                 unlock_rmap(rmapp);
 904                 hptep[0] &= ~HPTE_V_HVLOCK;
 905         }
 906         return 0;
 907 }
 908
 909 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
 910 {
 911         if (kvm->arch.using_mmu_notifiers)
 912                 kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 913         return 0;
 914 }
 915
 916 int kvm_unmap_hva_range(struct kvm *kvm, unsigned long start, unsigned long end)
 917 {
 918         if (kvm->arch.using_mmu_notifiers)
 919                 kvm_handle_hva_range(kvm, start, end, kvm_unmap_rmapp);
 920         return 0;
 921 }
 922
 923 void kvmppc_core_flush_memslot(struct kvm *kvm, struct kvm_memory_slot *memslot)
 924 {
 925         unsigned long *rmapp;
 926         unsigned long gfn;
 927         unsigned long n;
 928
 929         rmapp = memslot->arch.rmap;
 930         gfn = memslot->base_gfn;
 931         for (n = memslot->npages; n; --n) {
 932                 /*
 933                  * Testing the present bit without locking is OK because
 934                  * the memslot has been marked invalid already, and hence
 935                  * no new HPTEs referencing this page can be created,
 936                  * thus the present bit can't go from 0 to 1.
 937                  */
 938                 if (*rmapp & KVMPPC_RMAP_PRESENT)
 939                         kvm_unmap_rmapp(kvm, rmapp, gfn);
 940                 ++rmapp;
 941                 ++gfn;
 942         }
 943 }
 944
 945 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
 946                          unsigned long gfn)
 947 {
 948         struct revmap_entry *rev = kvm->arch.revmap;
 949         unsigned long head, i, j;
 950         unsigned long *hptep;
 951         int ret = 0;
 952
 953  retry:
 954         lock_rmap(rmapp);
 955         if (*rmapp & KVMPPC_RMAP_REFERENCED) {
 956                 *rmapp &= ~KVMPPC_RMAP_REFERENCED;
 957                 ret = 1;
 958         }
 959         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 960                 unlock_rmap(rmapp);
 961                 return ret;
 962         }
 963
 964         i = head = *rmapp & KVMPPC_RMAP_INDEX;
 965         do {
 966                 hptep = (unsigned long *) (kvm->arch.hpt_virt + (i << 4));
 967                 j = rev[i].forw;
 968
 969                 /* If this HPTE isn't referenced, ignore it */
 970                 if (!(hptep[1] & HPTE_R_R))
 971                         continue;
 972
 973                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 974                         /* unlock rmap before spinning on the HPTE lock */
 975                         unlock_rmap(rmapp);
 976                         while (hptep[0] & HPTE_V_HVLOCK)
 977                                 cpu_relax();
 978                         goto retry;
 979                 }
 980
 981                 /* Now check and modify the HPTE */
 982                 if ((hptep[0] & HPTE_V_VALID) && (hptep[1] & HPTE_R_R)) {
 983                         kvmppc_clear_ref_hpte(kvm, hptep, i);
 984                         if (!(rev[i].guest_rpte & HPTE_R_R)) {
 985                                 rev[i].guest_rpte |= HPTE_R_R;
 986                                 note_hpte_modification(kvm, &rev[i]);
 987                         }
 988                         ret = 1;
 989                 }
 990                 hptep[0] &= ~HPTE_V_HVLOCK;
 991         } while ((i = j) != head);
 992
 993         unlock_rmap(rmapp);
 994         return ret;
 995 }
 996
 997 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
 998 {
 999         if (!kvm->arch.using_mmu_notifiers)
1000                 return 0;
1001         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
1002 }
1003
1004 static int kvm_test_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
1005                               unsigned long gfn)
1006 {
1007         struct revmap_entry *rev = kvm->arch.revmap;
1008         unsigned long head, i, j;
1009         unsigned long *hp;
1010         int ret = 1;
1011
1012         if (*rmapp & KVMPPC_RMAP_REFERENCED)
1013                 return 1;
1014
1015         lock_rmap(rmapp);
1016         if (*rmapp & KVMPPC_RMAP_REFERENCED)
1017                 goto out;
1018
1019         if (*rmapp & KVMPPC_RMAP_PRESENT) {
1020                 i = head = *rmapp & KVMPPC_RMAP_INDEX;
1021                 do {
1022                         hp = (unsigned long *)(kvm->arch.hpt_virt + (i << 4));
1023                         j = rev[i].forw;
1024                         if (hp[1] & HPTE_R_R)
1025                                 goto out;
1026                 } while ((i = j) != head);
1027         }
1028         ret = 0;
1029
1030  out:
1031         unlock_rmap(rmapp);
1032         return ret;
1033 }
1034
1035 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1036 {
1037         if (!kvm->arch.using_mmu_notifiers)
1038                 return 0;
1039         return kvm_handle_hva(kvm, hva, kvm_test_age_rmapp);
1040 }
1041
1042 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1043 {
1044         if (!kvm->arch.using_mmu_notifiers)
1045                 return;
1046         kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
1047 }
1048
1049 static int kvm_test_clear_dirty(struct kvm *kvm, unsigned long *rmapp)
1050 {
1051         struct revmap_entry *rev = kvm->arch.revmap;
1052         unsigned long head, i, j;
1053         unsigned long *hptep;
1054         int ret = 0;
1055
1056  retry:
1057         lock_rmap(rmapp);
1058         if (*rmapp & KVMPPC_RMAP_CHANGED) {
1059                 *rmapp &= ~KVMPPC_RMAP_CHANGED;
1060                 ret = 1;
1061         }
1062         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
1063                 unlock_rmap(rmapp);
1064                 return ret;
1065         }
1066
1067         i = head = *rmapp & KVMPPC_RMAP_INDEX;
1068         do {
1069                 hptep = (unsigned long *) (kvm->arch.hpt_virt + (i << 4));
1070                 j = rev[i].forw;
1071
1072                 if (!(hptep[1] & HPTE_R_C))
1073                         continue;
1074
1075                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
1076                         /* unlock rmap before spinning on the HPTE lock */
1077                         unlock_rmap(rmapp);
1078                         while (hptep[0] & HPTE_V_HVLOCK)
1079                                 cpu_relax();
1080                         goto retry;
1081                 }
1082
1083                 /* Now check and modify the HPTE */
1084                 if ((hptep[0] & HPTE_V_VALID) && (hptep[1] & HPTE_R_C)) {
1085                         /* need to make it temporarily absent to clear C */
1086                         hptep[0] |= HPTE_V_ABSENT;
1087                         kvmppc_invalidate_hpte(kvm, hptep, i);
1088                         hptep[1] &= ~HPTE_R_C;
1089                         eieio();
1090                         hptep[0] = (hptep[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
1091                         if (!(rev[i].guest_rpte & HPTE_R_C)) {
1092                                 rev[i].guest_rpte |= HPTE_R_C;
1093                                 note_hpte_modification(kvm, &rev[i]);
1094                         }
1095                         ret = 1;
1096                 }
1097                 hptep[0] &= ~HPTE_V_HVLOCK;
1098         } while ((i = j) != head);
1099
1100         unlock_rmap(rmapp);
1101         return ret;
1102 }
1103
1104 static void harvest_vpa_dirty(struct kvmppc_vpa *vpa,
1105                               struct kvm_memory_slot *memslot,
1106                               unsigned long *map)
1107 {
1108         unsigned long gfn;
1109
1110         if (!vpa->dirty || !vpa->pinned_addr)
1111                 return;
1112         gfn = vpa->gpa >> PAGE_SHIFT;
1113         if (gfn < memslot->base_gfn ||
1114             gfn >= memslot->base_gfn + memslot->npages)
1115                 return;
1116
1117         vpa->dirty = false;
1118         if (map)
1119                 __set_bit_le(gfn - memslot->base_gfn, map);
1120 }
1121
1122 long kvmppc_hv_get_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot,
1123                              unsigned long *map)
1124 {
1125         unsigned long i;
1126         unsigned long *rmapp;
1127         struct kvm_vcpu *vcpu;
1128
1129         preempt_disable();
1130         rmapp = memslot->arch.rmap;
1131         for (i = 0; i < memslot->npages; ++i) {
1132                 if (kvm_test_clear_dirty(kvm, rmapp) && map)
1133                         __set_bit_le(i, map);
1134                 ++rmapp;
1135         }
1136
1137         /* Harvest dirty bits from VPA and DTL updates */
1138         /* Note: we never modify the SLB shadow buffer areas */
1139         kvm_for_each_vcpu(i, vcpu, kvm) {
1140                 spin_lock(&vcpu->arch.vpa_update_lock);
1141                 harvest_vpa_dirty(&vcpu->arch.vpa, memslot, map);
1142                 harvest_vpa_dirty(&vcpu->arch.dtl, memslot, map);
1143                 spin_unlock(&vcpu->arch.vpa_update_lock);
1144         }
1145         preempt_enable();
1146         return 0;
1147 }
1148
1149 void *kvmppc_pin_guest_page(struct kvm *kvm, unsigned long gpa,
1150                             unsigned long *nb_ret)
1151 {
1152         struct kvm_memory_slot *memslot;
1153         unsigned long gfn = gpa >> PAGE_SHIFT;
1154         struct page *page, *pages[1];
1155         int npages;
1156         unsigned long hva, offset;
1157         unsigned long pa;
1158         unsigned long *physp;
1159         int srcu_idx;
1160
1161         srcu_idx = srcu_read_lock(&kvm->srcu);
1162         memslot = gfn_to_memslot(kvm, gfn);
1163         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
1164                 goto err;
1165         if (!kvm->arch.using_mmu_notifiers) {
1166                 physp = memslot->arch.slot_phys;
1167                 if (!physp)
1168                         goto err;
1169                 physp += gfn - memslot->base_gfn;
1170                 pa = *physp;
1171                 if (!pa) {
1172                         if (kvmppc_get_guest_page(kvm, gfn, memslot,
1173                                                   PAGE_SIZE) < 0)
1174                                 goto err;
1175                         pa = *physp;
1176                 }
1177                 page = pfn_to_page(pa >> PAGE_SHIFT);
1178                 get_page(page);
1179         } else {
1180                 hva = gfn_to_hva_memslot(memslot, gfn);
1181                 npages = get_user_pages_fast(hva, 1, 1, pages);
1182                 if (npages < 1)
1183                         goto err;
1184                 page = pages[0];
1185         }
1186         srcu_read_unlock(&kvm->srcu, srcu_idx);
1187
1188         offset = gpa & (PAGE_SIZE - 1);
1189         if (nb_ret)
1190                 *nb_ret = PAGE_SIZE - offset;
1191         return page_address(page) + offset;
1192
1193  err:
1194         srcu_read_unlock(&kvm->srcu, srcu_idx);
1195         return NULL;
1196 }
1197
1198 void kvmppc_unpin_guest_page(struct kvm *kvm, void *va, unsigned long gpa,
1199                              bool dirty)
1200 {
1201         struct page *page = virt_to_page(va);
1202         struct kvm_memory_slot *memslot;
1203         unsigned long gfn;
1204         unsigned long *rmap;
1205         int srcu_idx;
1206
1207         put_page(page);
1208
1209         if (!dirty || !kvm->arch.using_mmu_notifiers)
1210                 return;
1211
1212         /* We need to mark this page dirty in the rmap chain */
1213         gfn = gpa >> PAGE_SHIFT;
1214         srcu_idx = srcu_read_lock(&kvm->srcu);
1215         memslot = gfn_to_memslot(kvm, gfn);
1216         if (memslot) {
1217                 rmap = &memslot->arch.rmap[gfn - memslot->base_gfn];
1218                 lock_rmap(rmap);
1219                 *rmap |= KVMPPC_RMAP_CHANGED;
1220                 unlock_rmap(rmap);
1221         }
1222         srcu_read_unlock(&kvm->srcu, srcu_idx);
1223 }
1224
1225 /*
1226  * Functions for reading and writing the hash table via reads and
1227  * writes on a file descriptor.
1228  *
1229  * Reads return the guest view of the hash table, which has to be
1230  * pieced together from the real hash table and the guest_rpte
1231  * values in the revmap array.
1232  *
1233  * On writes, each HPTE written is considered in turn, and if it
1234  * is valid, it is written to the HPT as if an H_ENTER with the
1235  * exact flag set was done.  When the invalid count is non-zero
1236  * in the header written to the stream, the kernel will make
1237  * sure that that many HPTEs are invalid, and invalidate them
1238  * if not.
1239  */
1240
1241 struct kvm_htab_ctx {
1242         unsigned long   index;
1243         unsigned long   flags;
1244         struct kvm      *kvm;
1245         int             first_pass;
1246 };
1247
1248 #define HPTE_SIZE       (2 * sizeof(unsigned long))
1249
1250 /*
1251  * Returns 1 if this HPT entry has been modified or has pending
1252  * R/C bit changes.
1253  */
1254 static int hpte_dirty(struct revmap_entry *revp, unsigned long *hptp)
1255 {
1256         unsigned long rcbits_unset;
1257
1258         if (revp->guest_rpte & HPTE_GR_MODIFIED)
1259                 return 1;
1260
1261         /* Also need to consider changes in reference and changed bits */
1262         rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1263         if ((hptp[0] & HPTE_V_VALID) && (hptp[1] & rcbits_unset))
1264                 return 1;
1265
1266         return 0;
1267 }
1268
1269 static long record_hpte(unsigned long flags, unsigned long *hptp,
1270                         unsigned long *hpte, struct revmap_entry *revp,
1271                         int want_valid, int first_pass)
1272 {
1273         unsigned long v, r;
1274         unsigned long rcbits_unset;
1275         int ok = 1;
1276         int valid, dirty;
1277
1278         /* Unmodified entries are uninteresting except on the first pass */
1279         dirty = hpte_dirty(revp, hptp);
1280         if (!first_pass && !dirty)
1281                 return 0;
1282
1283         valid = 0;
1284         if (hptp[0] & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1285                 valid = 1;
1286                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) &&
1287                     !(hptp[0] & HPTE_V_BOLTED))
1288                         valid = 0;
1289         }
1290         if (valid != want_valid)
1291                 return 0;
1292
1293         v = r = 0;
1294         if (valid || dirty) {
1295                 /* lock the HPTE so it's stable and read it */
1296                 preempt_disable();
1297                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
1298                         cpu_relax();
1299                 v = hptp[0];
1300
1301                 /* re-evaluate valid and dirty from synchronized HPTE value */
1302                 valid = !!(v & HPTE_V_VALID);
1303                 dirty = !!(revp->guest_rpte & HPTE_GR_MODIFIED);
1304
1305                 /* Harvest R and C into guest view if necessary */
1306                 rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1307                 if (valid && (rcbits_unset & hptp[1])) {
1308                         revp->guest_rpte |= (hptp[1] & (HPTE_R_R | HPTE_R_C)) |
1309                                 HPTE_GR_MODIFIED;
1310                         dirty = 1;
1311                 }
1312
1313                 if (v & HPTE_V_ABSENT) {
1314                         v &= ~HPTE_V_ABSENT;
1315                         v |= HPTE_V_VALID;
1316                         valid = 1;
1317                 }
1318                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) && !(v & HPTE_V_BOLTED))
1319                         valid = 0;
1320
1321                 r = revp->guest_rpte;
1322                 /* only clear modified if this is the right sort of entry */
1323                 if (valid == want_valid && dirty) {
1324                         r &= ~HPTE_GR_MODIFIED;
1325                         revp->guest_rpte = r;
1326                 }
1327                 asm volatile(PPC_RELEASE_BARRIER "" : : : "memory");
1328                 hptp[0] &= ~HPTE_V_HVLOCK;
1329                 preempt_enable();
1330                 if (!(valid == want_valid && (first_pass || dirty)))
1331                         ok = 0;
1332         }
1333         hpte[0] = v;
1334         hpte[1] = r;
1335         return ok;
1336 }
1337
1338 static ssize_t kvm_htab_read(struct file *file, char __user *buf,
1339                              size_t count, loff_t *ppos)
1340 {
1341         struct kvm_htab_ctx *ctx = file->private_data;
1342         struct kvm *kvm = ctx->kvm;
1343         struct kvm_get_htab_header hdr;
1344         unsigned long *hptp;
1345         struct revmap_entry *revp;
1346         unsigned long i, nb, nw;
1347         unsigned long __user *lbuf;
1348         struct kvm_get_htab_header __user *hptr;
1349         unsigned long flags;
1350         int first_pass;
1351         unsigned long hpte[2];
1352
1353         if (!access_ok(VERIFY_WRITE, buf, count))
1354                 return -EFAULT;
1355
1356         first_pass = ctx->first_pass;
1357         flags = ctx->flags;
1358
1359         i = ctx->index;
1360         hptp = (unsigned long *)(kvm->arch.hpt_virt + (i * HPTE_SIZE));
1361         revp = kvm->arch.revmap + i;
1362         lbuf = (unsigned long __user *)buf;
1363
1364         nb = 0;
1365         while (nb + sizeof(hdr) + HPTE_SIZE < count) {
1366                 /* Initialize header */
1367                 hptr = (struct kvm_get_htab_header __user *)buf;
1368                 hdr.n_valid = 0;
1369                 hdr.n_invalid = 0;
1370                 nw = nb;
1371                 nb += sizeof(hdr);
1372                 lbuf = (unsigned long __user *)(buf + sizeof(hdr));
1373
1374                 /* Skip uninteresting entries, i.e. clean on not-first pass */
1375                 if (!first_pass) {
1376                         while (i < kvm->arch.hpt_npte &&
1377                                !hpte_dirty(revp, hptp)) {
1378                                 ++i;
1379                                 hptp += 2;
1380                                 ++revp;
1381                         }
1382                 }
1383                 hdr.index = i;
1384
1385                 /* Grab a series of valid entries */
1386                 while (i < kvm->arch.hpt_npte &&
1387                        hdr.n_valid < 0xffff &&
1388                        nb + HPTE_SIZE < count &&
1389                        record_hpte(flags, hptp, hpte, revp, 1, first_pass)) {
1390                         /* valid entry, write it out */
1391                         ++hdr.n_valid;
1392                         if (__put_user(hpte[0], lbuf) ||
1393                             __put_user(hpte[1], lbuf + 1))
1394                                 return -EFAULT;
1395                         nb += HPTE_SIZE;
1396                         lbuf += 2;
1397                         ++i;
1398                         hptp += 2;
1399                         ++revp;
1400                 }
1401                 /* Now skip invalid entries while we can */
1402                 while (i < kvm->arch.hpt_npte &&
1403                        hdr.n_invalid < 0xffff &&
1404                        record_hpte(flags, hptp, hpte, revp, 0, first_pass)) {
1405                         /* found an invalid entry */
1406                         ++hdr.n_invalid;
1407                         ++i;
1408                         hptp += 2;
1409                         ++revp;
1410                 }
1411
1412                 if (hdr.n_valid || hdr.n_invalid) {
1413                         /* write back the header */
1414                         if (__copy_to_user(hptr, &hdr, sizeof(hdr)))
1415                                 return -EFAULT;
1416                         nw = nb;
1417                         buf = (char __user *)lbuf;
1418                 } else {
1419                         nb = nw;
1420                 }
1421
1422                 /* Check if we've wrapped around the hash table */
1423                 if (i >= kvm->arch.hpt_npte) {
1424                         i = 0;
1425                         ctx->first_pass = 0;
1426                         break;
1427                 }
1428         }
1429
1430         ctx->index = i;
1431
1432         return nb;
1433 }
1434
1435 static ssize_t kvm_htab_write(struct file *file, const char __user *buf,
1436                               size_t count, loff_t *ppos)
1437 {
1438         struct kvm_htab_ctx *ctx = file->private_data;
1439         struct kvm *kvm = ctx->kvm;
1440         struct kvm_get_htab_header hdr;
1441         unsigned long i, j;
1442         unsigned long v, r;
1443         unsigned long __user *lbuf;
1444         unsigned long *hptp;
1445         unsigned long tmp[2];
1446         ssize_t nb;
1447         long int err, ret;
1448         int rma_setup;
1449
1450         if (!access_ok(VERIFY_READ, buf, count))
1451                 return -EFAULT;
1452
1453         /* lock out vcpus from running while we're doing this */
1454         mutex_lock(&kvm->lock);
1455         rma_setup = kvm->arch.rma_setup_done;
1456         if (rma_setup) {
1457                 kvm->arch.rma_setup_done = 0;   /* temporarily */
1458                 /* order rma_setup_done vs. vcpus_running */
1459                 smp_mb();
1460                 if (atomic_read(&kvm->arch.vcpus_running)) {
1461                         kvm->arch.rma_setup_done = 1;
1462                         mutex_unlock(&kvm->lock);
1463                         return -EBUSY;
1464                 }
1465         }
1466
1467         err = 0;
1468         for (nb = 0; nb + sizeof(hdr) <= count; ) {
1469                 err = -EFAULT;
1470                 if (__copy_from_user(&hdr, buf, sizeof(hdr)))
1471                         break;
1472
1473                 err = 0;
1474                 if (nb + hdr.n_valid * HPTE_SIZE > count)
1475                         break;
1476
1477                 nb += sizeof(hdr);
1478                 buf += sizeof(hdr);
1479
1480                 err = -EINVAL;
1481                 i = hdr.index;
1482                 if (i >= kvm->arch.hpt_npte ||
1483                     i + hdr.n_valid + hdr.n_invalid > kvm->arch.hpt_npte)
1484                         break;
1485
1486                 hptp = (unsigned long *)(kvm->arch.hpt_virt + (i * HPTE_SIZE));
1487                 lbuf = (unsigned long __user *)buf;
1488                 for (j = 0; j < hdr.n_valid; ++j) {
1489                         err = -EFAULT;
1490                         if (__get_user(v, lbuf) || __get_user(r, lbuf + 1))
1491                                 goto out;
1492                         err = -EINVAL;
1493                         if (!(v & HPTE_V_VALID))
1494                                 goto out;
1495                         lbuf += 2;
1496                         nb += HPTE_SIZE;
1497
1498                         if (hptp[0] & (HPTE_V_VALID | HPTE_V_ABSENT))
1499                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1500                         err = -EIO;
1501                         ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, i, v, r,
1502                                                          tmp);
1503                         if (ret != H_SUCCESS) {
1504                                 pr_err("kvm_htab_write ret %ld i=%ld v=%lx "
1505                                        "r=%lx\n", ret, i, v, r);
1506                                 goto out;
1507                         }
1508                         if (!rma_setup && is_vrma_hpte(v)) {
1509                                 unsigned long psize = hpte_page_size(v, r);
1510                                 unsigned long senc = slb_pgsize_encoding(psize);
1511                                 unsigned long lpcr;
1512
1513                                 kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
1514                                         (VRMA_VSID << SLB_VSID_SHIFT_1T);
1515                                 lpcr = kvm->arch.lpcr & ~LPCR_VRMASD;
1516                                 lpcr |= senc << (LPCR_VRMASD_SH - 4);
1517                                 kvm->arch.lpcr = lpcr;
1518                                 rma_setup = 1;
1519                         }
1520                         ++i;
1521                         hptp += 2;
1522                 }
1523
1524                 for (j = 0; j < hdr.n_invalid; ++j) {
1525                         if (hptp[0] & (HPTE_V_VALID | HPTE_V_ABSENT))
1526                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1527                         ++i;
1528                         hptp += 2;
1529                 }
1530                 err = 0;
1531         }
1532
1533  out:
1534         /* Order HPTE updates vs. rma_setup_done */
1535         smp_wmb();
1536         kvm->arch.rma_setup_done = rma_setup;
1537         mutex_unlock(&kvm->lock);
1538
1539         if (err)
1540                 return err;
1541         return nb;
1542 }
1543
1544 static int kvm_htab_release(struct inode *inode, struct file *filp)
1545 {
1546         struct kvm_htab_ctx *ctx = filp->private_data;
1547
1548         filp->private_data = NULL;
1549         if (!(ctx->flags & KVM_GET_HTAB_WRITE))
1550                 atomic_dec(&ctx->kvm->arch.hpte_mod_interest);
1551         kvm_put_kvm(ctx->kvm);
1552         kfree(ctx);
1553         return 0;
1554 }
1555
1556 static const struct file_operations kvm_htab_fops = {
1557         .read           = kvm_htab_read,
1558         .write          = kvm_htab_write,
1559         .llseek         = default_llseek,
1560         .release        = kvm_htab_release,
1561 };
1562
1563 int kvm_vm_ioctl_get_htab_fd(struct kvm *kvm, struct kvm_get_htab_fd *ghf)
1564 {
1565         int ret;
1566         struct kvm_htab_ctx *ctx;
1567         int rwflag;
1568
1569         /* reject flags we don't recognize */
1570         if (ghf->flags & ~(KVM_GET_HTAB_BOLTED_ONLY | KVM_GET_HTAB_WRITE))
1571                 return -EINVAL;
1572         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1573         if (!ctx)
1574                 return -ENOMEM;
1575         kvm_get_kvm(kvm);
1576         ctx->kvm = kvm;
1577         ctx->index = ghf->start_index;
1578         ctx->flags = ghf->flags;
1579         ctx->first_pass = 1;
1580
1581         rwflag = (ghf->flags & KVM_GET_HTAB_WRITE) ? O_WRONLY : O_RDONLY;
1582         ret = anon_inode_getfd("kvm-htab", &kvm_htab_fops, ctx, rwflag);
1583         if (ret < 0) {
1584                 kvm_put_kvm(kvm);
1585                 return ret;
1586         }
1587
1588         if (rwflag == O_RDONLY) {
1589                 mutex_lock(&kvm->slots_lock);
1590                 atomic_inc(&kvm->arch.hpte_mod_interest);
1591                 /* make sure kvmppc_do_h_enter etc. see the increment */
1592                 synchronize_srcu_expedited(&kvm->srcu);
1593                 mutex_unlock(&kvm->slots_lock);
1594         }
1595
1596         return ret;
1597 }
1598
1599 void kvmppc_mmu_book3s_hv_init(struct kvm_vcpu *vcpu)
1600 {
1601         struct kvmppc_mmu *mmu = &vcpu->arch.mmu;
1602
1603         if (cpu_has_feature(CPU_FTR_ARCH_206))
1604                 vcpu->arch.slb_nr = 32;         /* POWER7 */
1605         else
1606                 vcpu->arch.slb_nr = 64;
1607
1608         mmu->xlate = kvmppc_mmu_book3s_64_hv_xlate;
1609         mmu->reset_msr = kvmppc_mmu_book3s_64_hv_reset_msr;
1610
1611         vcpu->arch.hflags |= BOOK3S_HFLAG_SLB;
1612 }