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 = 0;
  54         struct revmap_entry *rev;
  55         struct page *page = NULL;
  56         long order = KVM_DEFAULT_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         kvm->arch.hpt_cma_alloc = 0;
  65         page = kvm_alloc_hpt(1ul << (order - PAGE_SHIFT));
  66         if (page) {
  67                 hpt = (unsigned long)pfn_to_kaddr(page_to_pfn(page));
  68                 memset((void *)hpt, 0, (1ul << order));
  69                 kvm->arch.hpt_cma_alloc = 1;
  70         }
  71
  72         /* Lastly try successively smaller sizes from the page allocator */
  73         while (!hpt && order > PPC_MIN_HPT_ORDER) {
  74                 hpt = __get_free_pages(GFP_KERNEL|__GFP_ZERO|__GFP_REPEAT|
  75                                        __GFP_NOWARN, order - PAGE_SHIFT);
  76                 if (!hpt)
  77                         --order;
  78         }
  79
  80         if (!hpt)
  81                 return -ENOMEM;
  82
  83         kvm->arch.hpt_virt = hpt;
  84         kvm->arch.hpt_order = order;
  85         /* HPTEs are 2**4 bytes long */
  86         kvm->arch.hpt_npte = 1ul << (order - 4);
  87         /* 128 (2**7) bytes in each HPTEG */
  88         kvm->arch.hpt_mask = (1ul << (order - 7)) - 1;
  89
  90         /* Allocate reverse map array */
  91         rev = vmalloc(sizeof(struct revmap_entry) * kvm->arch.hpt_npte);
  92         if (!rev) {
  93                 pr_err("kvmppc_alloc_hpt: Couldn't alloc reverse map array\n");
  94                 goto out_freehpt;
  95         }
  96         kvm->arch.revmap = rev;
  97         kvm->arch.sdr1 = __pa(hpt) | (order - 18);
  98
  99         pr_info("KVM guest htab at %lx (order %ld), LPID %x\n",
 100                 hpt, order, kvm->arch.lpid);
 101
 102         if (htab_orderp)
 103                 *htab_orderp = order;
 104         return 0;
 105
 106  out_freehpt:
 107         if (kvm->arch.hpt_cma_alloc)
 108                 kvm_release_hpt(page, 1 << (order - PAGE_SHIFT));
 109         else
 110                 free_pages(hpt, order - PAGE_SHIFT);
 111         return -ENOMEM;
 112 }
 113
 114 long kvmppc_alloc_reset_hpt(struct kvm *kvm, u32 *htab_orderp)
 115 {
 116         long err = -EBUSY;
 117         long order;
 118
 119         mutex_lock(&kvm->lock);
 120         if (kvm->arch.rma_setup_done) {
 121                 kvm->arch.rma_setup_done = 0;
 122                 /* order rma_setup_done vs. vcpus_running */
 123                 smp_mb();
 124                 if (atomic_read(&kvm->arch.vcpus_running)) {
 125                         kvm->arch.rma_setup_done = 1;
 126                         goto out;
 127                 }
 128         }
 129         if (kvm->arch.hpt_virt) {
 130                 order = kvm->arch.hpt_order;
 131                 /* Set the entire HPT to 0, i.e. invalid HPTEs */
 132                 memset((void *)kvm->arch.hpt_virt, 0, 1ul << order);
 133                 /*
 134                  * Reset all the reverse-mapping chains for all memslots
 135                  */
 136                 kvmppc_rmap_reset(kvm);
 137                 /* Ensure that each vcpu will flush its TLB on next entry. */
 138                 cpumask_setall(&kvm->arch.need_tlb_flush);
 139                 *htab_orderp = order;
 140                 err = 0;
 141         } else {
 142                 err = kvmppc_alloc_hpt(kvm, htab_orderp);
 143                 order = *htab_orderp;
 144         }
 145  out:
 146         mutex_unlock(&kvm->lock);
 147         return err;
 148 }
 149
 150 void kvmppc_free_hpt(struct kvm *kvm)
 151 {
 152         kvmppc_free_lpid(kvm->arch.lpid);
 153         vfree(kvm->arch.revmap);
 154         if (kvm->arch.hpt_cma_alloc)
 155                 kvm_release_hpt(virt_to_page(kvm->arch.hpt_virt),
 156                                 1 << (kvm->arch.hpt_order - PAGE_SHIFT));
 157         else
 158                 free_pages(kvm->arch.hpt_virt,
 159                            kvm->arch.hpt_order - PAGE_SHIFT);
 160 }
 161
 162 /* Bits in first HPTE dword for pagesize 4k, 64k or 16M */
 163 static inline unsigned long hpte0_pgsize_encoding(unsigned long pgsize)
 164 {
 165         return (pgsize > 0x1000) ? HPTE_V_LARGE : 0;
 166 }
 167
 168 /* Bits in second HPTE dword for pagesize 4k, 64k or 16M */
 169 static inline unsigned long hpte1_pgsize_encoding(unsigned long pgsize)
 170 {
 171         return (pgsize == 0x10000) ? 0x1000 : 0;
 172 }
 173
 174 void kvmppc_map_vrma(struct kvm_vcpu *vcpu, struct kvm_memory_slot *memslot,
 175                      unsigned long porder)
 176 {
 177         unsigned long i;
 178         unsigned long npages;
 179         unsigned long hp_v, hp_r;
 180         unsigned long addr, hash;
 181         unsigned long psize;
 182         unsigned long hp0, hp1;
 183         unsigned long idx_ret;
 184         long ret;
 185         struct kvm *kvm = vcpu->kvm;
 186
 187         psize = 1ul << porder;
 188         npages = memslot->npages >> (porder - PAGE_SHIFT);
 189
 190         /* VRMA can't be > 1TB */
 191         if (npages > 1ul << (40 - porder))
 192                 npages = 1ul << (40 - porder);
 193         /* Can't use more than 1 HPTE per HPTEG */
 194         if (npages > kvm->arch.hpt_mask + 1)
 195                 npages = kvm->arch.hpt_mask + 1;
 196
 197         hp0 = HPTE_V_1TB_SEG | (VRMA_VSID << (40 - 16)) |
 198                 HPTE_V_BOLTED | hpte0_pgsize_encoding(psize);
 199         hp1 = hpte1_pgsize_encoding(psize) |
 200                 HPTE_R_R | HPTE_R_C | HPTE_R_M | PP_RWXX;
 201
 202         for (i = 0; i < npages; ++i) {
 203                 addr = i << porder;
 204                 /* can't use hpt_hash since va > 64 bits */
 205                 hash = (i ^ (VRMA_VSID ^ (VRMA_VSID << 25))) & kvm->arch.hpt_mask;
 206                 /*
 207                  * We assume that the hash table is empty and no
 208                  * vcpus are using it at this stage.  Since we create
 209                  * at most one HPTE per HPTEG, we just assume entry 7
 210                  * is available and use it.
 211                  */
 212                 hash = (hash << 3) + 7;
 213                 hp_v = hp0 | ((addr >> 16) & ~0x7fUL);
 214                 hp_r = hp1 | addr;
 215                 ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, hash, hp_v, hp_r,
 216                                                  &idx_ret);
 217                 if (ret != H_SUCCESS) {
 218                         pr_err("KVM: map_vrma at %lx failed, ret=%ld\n",
 219                                addr, ret);
 220                         break;
 221                 }
 222         }
 223 }
 224
 225 int kvmppc_mmu_hv_init(void)
 226 {
 227         unsigned long host_lpid, rsvd_lpid;
 228
 229         if (!cpu_has_feature(CPU_FTR_HVMODE))
 230                 return -EINVAL;
 231
 232         /* POWER7 has 10-bit LPIDs, PPC970 and e500mc have 6-bit LPIDs */
 233         if (cpu_has_feature(CPU_FTR_ARCH_206)) {
 234                 host_lpid = mfspr(SPRN_LPID);   /* POWER7 */
 235                 rsvd_lpid = LPID_RSVD;
 236         } else {
 237                 host_lpid = 0;                  /* PPC970 */
 238                 rsvd_lpid = MAX_LPID_970;
 239         }
 240
 241         kvmppc_init_lpid(rsvd_lpid + 1);
 242
 243         kvmppc_claim_lpid(host_lpid);
 244         /* rsvd_lpid is reserved for use in partition switching */
 245         kvmppc_claim_lpid(rsvd_lpid);
 246
 247         return 0;
 248 }
 249
 250 static void kvmppc_mmu_book3s_64_hv_reset_msr(struct kvm_vcpu *vcpu)
 251 {
 252         unsigned long msr = vcpu->arch.intr_msr;
 253
 254         /* If transactional, change to suspend mode on IRQ delivery */
 255         if (MSR_TM_TRANSACTIONAL(vcpu->arch.shregs.msr))
 256                 msr |= MSR_TS_S;
 257         else
 258                 msr |= vcpu->arch.shregs.msr & MSR_TS_MASK;
 259         kvmppc_set_msr(vcpu, msr);
 260 }
 261
 262 /*
 263  * This is called to get a reference to a guest page if there isn't
 264  * one already in the memslot->arch.slot_phys[] array.
 265  */
 266 static long kvmppc_get_guest_page(struct kvm *kvm, unsigned long gfn,
 267                                   struct kvm_memory_slot *memslot,
 268                                   unsigned long psize)
 269 {
 270         unsigned long start;
 271         long np, err;
 272         struct page *page, *hpage, *pages[1];
 273         unsigned long s, pgsize;
 274         unsigned long *physp;
 275         unsigned int is_io, got, pgorder;
 276         struct vm_area_struct *vma;
 277         unsigned long pfn, i, npages;
 278
 279         physp = memslot->arch.slot_phys;
 280         if (!physp)
 281                 return -EINVAL;
 282         if (physp[gfn - memslot->base_gfn])
 283                 return 0;
 284
 285         is_io = 0;
 286         got = 0;
 287         page = NULL;
 288         pgsize = psize;
 289         err = -EINVAL;
 290         start = gfn_to_hva_memslot(memslot, gfn);
 291
 292         /* Instantiate and get the page we want access to */
 293         np = get_user_pages_fast(start, 1, 1, pages);
 294         if (np != 1) {
 295                 /* Look up the vma for the page */
 296                 down_read(&current->mm->mmap_sem);
 297                 vma = find_vma(current->mm, start);
 298                 if (!vma || vma->vm_start > start ||
 299                     start + psize > vma->vm_end ||
 300                     !(vma->vm_flags & VM_PFNMAP))
 301                         goto up_err;
 302                 is_io = hpte_cache_bits(pgprot_val(vma->vm_page_prot));
 303                 pfn = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 304                 /* check alignment of pfn vs. requested page size */
 305                 if (psize > PAGE_SIZE && (pfn & ((psize >> PAGE_SHIFT) - 1)))
 306                         goto up_err;
 307                 up_read(&current->mm->mmap_sem);
 308
 309         } else {
 310                 page = pages[0];
 311                 got = KVMPPC_GOT_PAGE;
 312
 313                 /* See if this is a large page */
 314                 s = PAGE_SIZE;
 315                 if (PageHuge(page)) {
 316                         hpage = compound_head(page);
 317                         s <<= compound_order(hpage);
 318                         /* Get the whole large page if slot alignment is ok */
 319                         if (s > psize && slot_is_aligned(memslot, s) &&
 320                             !(memslot->userspace_addr & (s - 1))) {
 321                                 start &= ~(s - 1);
 322                                 pgsize = s;
 323                                 get_page(hpage);
 324                                 put_page(page);
 325                                 page = hpage;
 326                         }
 327                 }
 328                 if (s < psize)
 329                         goto out;
 330                 pfn = page_to_pfn(page);
 331         }
 332
 333         npages = pgsize >> PAGE_SHIFT;
 334         pgorder = __ilog2(npages);
 335         physp += (gfn - memslot->base_gfn) & ~(npages - 1);
 336         spin_lock(&kvm->arch.slot_phys_lock);
 337         for (i = 0; i < npages; ++i) {
 338                 if (!physp[i]) {
 339                         physp[i] = ((pfn + i) << PAGE_SHIFT) +
 340                                 got + is_io + pgorder;
 341                         got = 0;
 342                 }
 343         }
 344         spin_unlock(&kvm->arch.slot_phys_lock);
 345         err = 0;
 346
 347  out:
 348         if (got)
 349                 put_page(page);
 350         return err;
 351
 352  up_err:
 353         up_read(&current->mm->mmap_sem);
 354         return err;
 355 }
 356
 357 long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
 358                                 long pte_index, unsigned long pteh,
 359                                 unsigned long ptel, unsigned long *pte_idx_ret)
 360 {
 361         unsigned long psize, gpa, gfn;
 362         struct kvm_memory_slot *memslot;
 363         long ret;
 364
 365         if (kvm->arch.using_mmu_notifiers)
 366                 goto do_insert;
 367
 368         psize = hpte_page_size(pteh, ptel);
 369         if (!psize)
 370                 return H_PARAMETER;
 371
 372         pteh &= ~(HPTE_V_HVLOCK | HPTE_V_ABSENT | HPTE_V_VALID);
 373
 374         /* Find the memslot (if any) for this address */
 375         gpa = (ptel & HPTE_R_RPN) & ~(psize - 1);
 376         gfn = gpa >> PAGE_SHIFT;
 377         memslot = gfn_to_memslot(kvm, gfn);
 378         if (memslot && !(memslot->flags & KVM_MEMSLOT_INVALID)) {
 379                 if (!slot_is_aligned(memslot, psize))
 380                         return H_PARAMETER;
 381                 if (kvmppc_get_guest_page(kvm, gfn, memslot, psize) < 0)
 382                         return H_PARAMETER;
 383         }
 384
 385  do_insert:
 386         /* Protect linux PTE lookup from page table destruction */
 387         rcu_read_lock_sched();  /* this disables preemption too */
 388         ret = kvmppc_do_h_enter(kvm, flags, pte_index, pteh, ptel,
 389                                 current->mm->pgd, false, pte_idx_ret);
 390         rcu_read_unlock_sched();
 391         if (ret == H_TOO_HARD) {
 392                 /* this can't happen */
 393                 pr_err("KVM: Oops, kvmppc_h_enter returned too hard!\n");
 394                 ret = H_RESOURCE;       /* or something */
 395         }
 396         return ret;
 397
 398 }
 399
 400 /*
 401  * We come here on a H_ENTER call from the guest when we are not
 402  * using mmu notifiers and we don't have the requested page pinned
 403  * already.
 404  */
 405 long kvmppc_virtmode_h_enter(struct kvm_vcpu *vcpu, unsigned long flags,
 406                              long pte_index, unsigned long pteh,
 407                              unsigned long ptel)
 408 {
 409         return kvmppc_virtmode_do_h_enter(vcpu->kvm, flags, pte_index,
 410                                           pteh, ptel, &vcpu->arch.gpr[4]);
 411 }
 412
 413 static struct kvmppc_slb *kvmppc_mmu_book3s_hv_find_slbe(struct kvm_vcpu *vcpu,
 414                                                          gva_t eaddr)
 415 {
 416         u64 mask;
 417         int i;
 418
 419         for (i = 0; i < vcpu->arch.slb_nr; i++) {
 420                 if (!(vcpu->arch.slb[i].orige & SLB_ESID_V))
 421                         continue;
 422
 423                 if (vcpu->arch.slb[i].origv & SLB_VSID_B_1T)
 424                         mask = ESID_MASK_1T;
 425                 else
 426                         mask = ESID_MASK;
 427
 428                 if (((vcpu->arch.slb[i].orige ^ eaddr) & mask) == 0)
 429                         return &vcpu->arch.slb[i];
 430         }
 431         return NULL;
 432 }
 433
 434 static unsigned long kvmppc_mmu_get_real_addr(unsigned long v, unsigned long r,
 435                         unsigned long ea)
 436 {
 437         unsigned long ra_mask;
 438
 439         ra_mask = hpte_page_size(v, r) - 1;
 440         return (r & HPTE_R_RPN & ~ra_mask) | (ea & ra_mask);
 441 }
 442
 443 static int kvmppc_mmu_book3s_64_hv_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
 444                         struct kvmppc_pte *gpte, bool data, bool iswrite)
 445 {
 446         struct kvm *kvm = vcpu->kvm;
 447         struct kvmppc_slb *slbe;
 448         unsigned long slb_v;
 449         unsigned long pp, key;
 450         unsigned long v, gr;
 451         __be64 *hptep;
 452         int index;
 453         int virtmode = vcpu->arch.shregs.msr & (data ? MSR_DR : MSR_IR);
 454
 455         /* Get SLB entry */
 456         if (virtmode) {
 457                 slbe = kvmppc_mmu_book3s_hv_find_slbe(vcpu, eaddr);
 458                 if (!slbe)
 459                         return -EINVAL;
 460                 slb_v = slbe->origv;
 461         } else {
 462                 /* real mode access */
 463                 slb_v = vcpu->kvm->arch.vrma_slb_v;
 464         }
 465
 466         preempt_disable();
 467         /* Find the HPTE in the hash table */
 468         index = kvmppc_hv_find_lock_hpte(kvm, eaddr, slb_v,
 469                                          HPTE_V_VALID | HPTE_V_ABSENT);
 470         if (index < 0) {
 471                 preempt_enable();
 472                 return -ENOENT;
 473         }
 474         hptep = (__be64 *)(kvm->arch.hpt_virt + (index << 4));
 475         v = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
 476         gr = kvm->arch.revmap[index].guest_rpte;
 477
 478         /* Unlock the HPTE */
 479         asm volatile("lwsync" : : : "memory");
 480         hptep[0] = cpu_to_be64(v);
 481         preempt_enable();
 482
 483         gpte->eaddr = eaddr;
 484         gpte->vpage = ((v & HPTE_V_AVPN) << 4) | ((eaddr >> 12) & 0xfff);
 485
 486         /* Get PP bits and key for permission check */
 487         pp = gr & (HPTE_R_PP0 | HPTE_R_PP);
 488         key = (vcpu->arch.shregs.msr & MSR_PR) ? SLB_VSID_KP : SLB_VSID_KS;
 489         key &= slb_v;
 490
 491         /* Calculate permissions */
 492         gpte->may_read = hpte_read_permission(pp, key);
 493         gpte->may_write = hpte_write_permission(pp, key);
 494         gpte->may_execute = gpte->may_read && !(gr & (HPTE_R_N | HPTE_R_G));
 495
 496         /* Storage key permission check for POWER7 */
 497         if (data && virtmode && cpu_has_feature(CPU_FTR_ARCH_206)) {
 498                 int amrfield = hpte_get_skey_perm(gr, vcpu->arch.amr);
 499                 if (amrfield & 1)
 500                         gpte->may_read = 0;
 501                 if (amrfield & 2)
 502                         gpte->may_write = 0;
 503         }
 504
 505         /* Get the guest physical address */
 506         gpte->raddr = kvmppc_mmu_get_real_addr(v, gr, eaddr);
 507         return 0;
 508 }
 509
 510 /*
 511  * Quick test for whether an instruction is a load or a store.
 512  * If the instruction is a load or a store, then this will indicate
 513  * which it is, at least on server processors.  (Embedded processors
 514  * have some external PID instructions that don't follow the rule
 515  * embodied here.)  If the instruction isn't a load or store, then
 516  * this doesn't return anything useful.
 517  */
 518 static int instruction_is_store(unsigned int instr)
 519 {
 520         unsigned int mask;
 521
 522         mask = 0x10000000;
 523         if ((instr & 0xfc000000) == 0x7c000000)
 524                 mask = 0x100;           /* major opcode 31 */
 525         return (instr & mask) != 0;
 526 }
 527
 528 static int kvmppc_hv_emulate_mmio(struct kvm_run *run, struct kvm_vcpu *vcpu,
 529                                   unsigned long gpa, gva_t ea, int is_store)
 530 {
 531         u32 last_inst;
 532
 533         /*
 534          * If we fail, we just return to the guest and try executing it again.
 535          */
 536         if (kvmppc_get_last_inst(vcpu, INST_GENERIC, &last_inst) !=
 537                 EMULATE_DONE)
 538                 return RESUME_GUEST;
 539
 540         /*
 541          * WARNING: We do not know for sure whether the instruction we just
 542          * read from memory is the same that caused the fault in the first
 543          * place.  If the instruction we read is neither an load or a store,
 544          * then it can't access memory, so we don't need to worry about
 545          * enforcing access permissions.  So, assuming it is a load or
 546          * store, we just check that its direction (load or store) is
 547          * consistent with the original fault, since that's what we
 548          * checked the access permissions against.  If there is a mismatch
 549          * we just return and retry the instruction.
 550          */
 551
 552         if (instruction_is_store(last_inst) != !!is_store)
 553                 return RESUME_GUEST;
 554
 555         /*
 556          * Emulated accesses are emulated by looking at the hash for
 557          * translation once, then performing the access later. The
 558          * translation could be invalidated in the meantime in which
 559          * point performing the subsequent memory access on the old
 560          * physical address could possibly be a security hole for the
 561          * guest (but not the host).
 562          *
 563          * This is less of an issue for MMIO stores since they aren't
 564          * globally visible. It could be an issue for MMIO loads to
 565          * a certain extent but we'll ignore it for now.
 566          */
 567
 568         vcpu->arch.paddr_accessed = gpa;
 569         vcpu->arch.vaddr_accessed = ea;
 570         return kvmppc_emulate_mmio(run, vcpu);
 571 }
 572
 573 int kvmppc_book3s_hv_page_fault(struct kvm_run *run, struct kvm_vcpu *vcpu,
 574                                 unsigned long ea, unsigned long dsisr)
 575 {
 576         struct kvm *kvm = vcpu->kvm;
 577         unsigned long hpte[3], r;
 578         __be64 *hptep;
 579         unsigned long mmu_seq, psize, pte_size;
 580         unsigned long gpa_base, gfn_base;
 581         unsigned long gpa, gfn, hva, pfn;
 582         struct kvm_memory_slot *memslot;
 583         unsigned long *rmap;
 584         struct revmap_entry *rev;
 585         struct page *page, *pages[1];
 586         long index, ret, npages;
 587         unsigned long is_io;
 588         unsigned int writing, write_ok;
 589         struct vm_area_struct *vma;
 590         unsigned long rcbits;
 591
 592         /*
 593          * Real-mode code has already searched the HPT and found the
 594          * entry we're interested in.  Lock the entry and check that
 595          * it hasn't changed.  If it has, just return and re-execute the
 596          * instruction.
 597          */
 598         if (ea != vcpu->arch.pgfault_addr)
 599                 return RESUME_GUEST;
 600         index = vcpu->arch.pgfault_index;
 601         hptep = (__be64 *)(kvm->arch.hpt_virt + (index << 4));
 602         rev = &kvm->arch.revmap[index];
 603         preempt_disable();
 604         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 605                 cpu_relax();
 606         hpte[0] = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
 607         hpte[1] = be64_to_cpu(hptep[1]);
 608         hpte[2] = r = rev->guest_rpte;
 609         asm volatile("lwsync" : : : "memory");
 610         hptep[0] = cpu_to_be64(hpte[0]);
 611         preempt_enable();
 612
 613         if (hpte[0] != vcpu->arch.pgfault_hpte[0] ||
 614             hpte[1] != vcpu->arch.pgfault_hpte[1])
 615                 return RESUME_GUEST;
 616
 617         /* Translate the logical address and get the page */
 618         psize = hpte_page_size(hpte[0], r);
 619         gpa_base = r & HPTE_R_RPN & ~(psize - 1);
 620         gfn_base = gpa_base >> PAGE_SHIFT;
 621         gpa = gpa_base | (ea & (psize - 1));
 622         gfn = gpa >> PAGE_SHIFT;
 623         memslot = gfn_to_memslot(kvm, gfn);
 624
 625         /* No memslot means it's an emulated MMIO region */
 626         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
 627                 return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
 628                                               dsisr & DSISR_ISSTORE);
 629
 630         if (!kvm->arch.using_mmu_notifiers)
 631                 return -EFAULT;         /* should never get here */
 632
 633         /*
 634          * This should never happen, because of the slot_is_aligned()
 635          * check in kvmppc_do_h_enter().
 636          */
 637         if (gfn_base < memslot->base_gfn)
 638                 return -EFAULT;
 639
 640         /* used to check for invalidations in progress */
 641         mmu_seq = kvm->mmu_notifier_seq;
 642         smp_rmb();
 643
 644         is_io = 0;
 645         pfn = 0;
 646         page = NULL;
 647         pte_size = PAGE_SIZE;
 648         writing = (dsisr & DSISR_ISSTORE) != 0;
 649         /* If writing != 0, then the HPTE must allow writing, if we get here */
 650         write_ok = writing;
 651         hva = gfn_to_hva_memslot(memslot, gfn);
 652         npages = get_user_pages_fast(hva, 1, writing, pages);
 653         if (npages < 1) {
 654                 /* Check if it's an I/O mapping */
 655                 down_read(&current->mm->mmap_sem);
 656                 vma = find_vma(current->mm, hva);
 657                 if (vma && vma->vm_start <= hva && hva + psize <= vma->vm_end &&
 658                     (vma->vm_flags & VM_PFNMAP)) {
 659                         pfn = vma->vm_pgoff +
 660                                 ((hva - vma->vm_start) >> PAGE_SHIFT);
 661                         pte_size = psize;
 662                         is_io = hpte_cache_bits(pgprot_val(vma->vm_page_prot));
 663                         write_ok = vma->vm_flags & VM_WRITE;
 664                 }
 665                 up_read(&current->mm->mmap_sem);
 666                 if (!pfn)
 667                         return -EFAULT;
 668         } else {
 669                 page = pages[0];
 670                 pfn = page_to_pfn(page);
 671                 if (PageHuge(page)) {
 672                         page = compound_head(page);
 673                         pte_size <<= compound_order(page);
 674                 }
 675                 /* if the guest wants write access, see if that is OK */
 676                 if (!writing && hpte_is_writable(r)) {
 677                         unsigned int hugepage_shift;
 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, &hugepage_shift);
 687                         if (ptep) {
 688                                 pte = kvmppc_read_update_linux_pte(ptep, 1,
 689                                                            hugepage_shift);
 690                                 if (pte_write(pte))
 691                                         write_ok = 1;
 692                         }
 693                         rcu_read_unlock_sched();
 694                 }
 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         /*
 713          * Set the HPTE to point to pfn.
 714          * Since the pfn is at PAGE_SIZE granularity, make sure we
 715          * don't mask out lower-order bits if psize < PAGE_SIZE.
 716          */
 717         if (psize < PAGE_SIZE)
 718                 psize = PAGE_SIZE;
 719         r = (r & ~(HPTE_R_PP0 - psize)) | ((pfn << PAGE_SHIFT) & ~(psize - 1));
 720         if (hpte_is_writable(r) && !write_ok)
 721                 r = hpte_make_readonly(r);
 722         ret = RESUME_GUEST;
 723         preempt_disable();
 724         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 725                 cpu_relax();
 726         if ((be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK) != hpte[0] ||
 727                 be64_to_cpu(hptep[1]) != hpte[1] ||
 728                 rev->guest_rpte != hpte[2])
 729                 /* HPTE has been changed under us; let the guest retry */
 730                 goto out_unlock;
 731         hpte[0] = (hpte[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
 732
 733         /* Always put the HPTE in the rmap chain for the page base address */
 734         rmap = &memslot->arch.rmap[gfn_base - memslot->base_gfn];
 735         lock_rmap(rmap);
 736
 737         /* Check if we might have been invalidated; let the guest retry if so */
 738         ret = RESUME_GUEST;
 739         if (mmu_notifier_retry(vcpu->kvm, mmu_seq)) {
 740                 unlock_rmap(rmap);
 741                 goto out_unlock;
 742         }
 743
 744         /* Only set R/C in real HPTE if set in both *rmap and guest_rpte */
 745         rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
 746         r &= rcbits | ~(HPTE_R_R | HPTE_R_C);
 747
 748         if (be64_to_cpu(hptep[0]) & HPTE_V_VALID) {
 749                 /* HPTE was previously valid, so we need to invalidate it */
 750                 unlock_rmap(rmap);
 751                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
 752                 kvmppc_invalidate_hpte(kvm, hptep, index);
 753                 /* don't lose previous R and C bits */
 754                 r |= be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
 755         } else {
 756                 kvmppc_add_revmap_chain(kvm, rev, rmap, index, 0);
 757         }
 758
 759         hptep[1] = cpu_to_be64(r);
 760         eieio();
 761         hptep[0] = cpu_to_be64(hpte[0]);
 762         asm volatile("ptesync" : : : "memory");
 763         preempt_enable();
 764         if (page && hpte_is_writable(r))
 765                 SetPageDirty(page);
 766
 767  out_put:
 768         if (page) {
 769                 /*
 770                  * We drop pages[0] here, not page because page might
 771                  * have been set to the head page of a compound, but
 772                  * we have to drop the reference on the correct tail
 773                  * page to match the get inside gup()
 774                  */
 775                 put_page(pages[0]);
 776         }
 777         return ret;
 778
 779  out_unlock:
 780         hptep[0] &= ~cpu_to_be64(HPTE_V_HVLOCK);
 781         preempt_enable();
 782         goto out_put;
 783 }
 784
 785 static void kvmppc_rmap_reset(struct kvm *kvm)
 786 {
 787         struct kvm_memslots *slots;
 788         struct kvm_memory_slot *memslot;
 789         int srcu_idx;
 790
 791         srcu_idx = srcu_read_lock(&kvm->srcu);
 792         slots = kvm->memslots;
 793         kvm_for_each_memslot(memslot, slots) {
 794                 /*
 795                  * This assumes it is acceptable to lose reference and
 796                  * change bits across a reset.
 797                  */
 798                 memset(memslot->arch.rmap, 0,
 799                        memslot->npages * sizeof(*memslot->arch.rmap));
 800         }
 801         srcu_read_unlock(&kvm->srcu, srcu_idx);
 802 }
 803
 804 static int kvm_handle_hva_range(struct kvm *kvm,
 805                                 unsigned long start,
 806                                 unsigned long end,
 807                                 int (*handler)(struct kvm *kvm,
 808                                                unsigned long *rmapp,
 809                                                unsigned long gfn))
 810 {
 811         int ret;
 812         int retval = 0;
 813         struct kvm_memslots *slots;
 814         struct kvm_memory_slot *memslot;
 815
 816         slots = kvm_memslots(kvm);
 817         kvm_for_each_memslot(memslot, slots) {
 818                 unsigned long hva_start, hva_end;
 819                 gfn_t gfn, gfn_end;
 820
 821                 hva_start = max(start, memslot->userspace_addr);
 822                 hva_end = min(end, memslot->userspace_addr +
 823                                         (memslot->npages << PAGE_SHIFT));
 824                 if (hva_start >= hva_end)
 825                         continue;
 826                 /*
 827                  * {gfn(page) | page intersects with [hva_start, hva_end)} =
 828                  * {gfn, gfn+1, ..., gfn_end-1}.
 829                  */
 830                 gfn = hva_to_gfn_memslot(hva_start, memslot);
 831                 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
 832
 833                 for (; gfn < gfn_end; ++gfn) {
 834                         gfn_t gfn_offset = gfn - memslot->base_gfn;
 835
 836                         ret = handler(kvm, &memslot->arch.rmap[gfn_offset], gfn);
 837                         retval |= ret;
 838                 }
 839         }
 840
 841         return retval;
 842 }
 843
 844 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 845                           int (*handler)(struct kvm *kvm, unsigned long *rmapp,
 846                                          unsigned long gfn))
 847 {
 848         return kvm_handle_hva_range(kvm, hva, hva + 1, handler);
 849 }
 850
 851 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
 852                            unsigned long gfn)
 853 {
 854         struct revmap_entry *rev = kvm->arch.revmap;
 855         unsigned long h, i, j;
 856         __be64 *hptep;
 857         unsigned long ptel, psize, rcbits;
 858
 859         for (;;) {
 860                 lock_rmap(rmapp);
 861                 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 862                         unlock_rmap(rmapp);
 863                         break;
 864                 }
 865
 866                 /*
 867                  * To avoid an ABBA deadlock with the HPTE lock bit,
 868                  * we can't spin on the HPTE lock while holding the
 869                  * rmap chain lock.
 870                  */
 871                 i = *rmapp & KVMPPC_RMAP_INDEX;
 872                 hptep = (__be64 *) (kvm->arch.hpt_virt + (i << 4));
 873                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 874                         /* unlock rmap before spinning on the HPTE lock */
 875                         unlock_rmap(rmapp);
 876                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
 877                                 cpu_relax();
 878                         continue;
 879                 }
 880                 j = rev[i].forw;
 881                 if (j == i) {
 882                         /* chain is now empty */
 883                         *rmapp &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
 884                 } else {
 885                         /* remove i from chain */
 886                         h = rev[i].back;
 887                         rev[h].forw = j;
 888                         rev[j].back = h;
 889                         rev[i].forw = rev[i].back = i;
 890                         *rmapp = (*rmapp & ~KVMPPC_RMAP_INDEX) | j;
 891                 }
 892
 893                 /* Now check and modify the HPTE */
 894                 ptel = rev[i].guest_rpte;
 895                 psize = hpte_page_size(be64_to_cpu(hptep[0]), ptel);
 896                 if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
 897                     hpte_rpn(ptel, psize) == gfn) {
 898                         if (kvm->arch.using_mmu_notifiers)
 899                                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
 900                         kvmppc_invalidate_hpte(kvm, hptep, i);
 901                         /* Harvest R and C */
 902                         rcbits = be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
 903                         *rmapp |= rcbits << KVMPPC_RMAP_RC_SHIFT;
 904                         if (rcbits & ~rev[i].guest_rpte) {
 905                                 rev[i].guest_rpte = ptel | rcbits;
 906                                 note_hpte_modification(kvm, &rev[i]);
 907                         }
 908                 }
 909                 unlock_rmap(rmapp);
 910                 hptep[0] &= ~cpu_to_be64(HPTE_V_HVLOCK);
 911         }
 912         return 0;
 913 }
 914
 915 int kvm_unmap_hva_hv(struct kvm *kvm, unsigned long hva)
 916 {
 917         if (kvm->arch.using_mmu_notifiers)
 918                 kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 919         return 0;
 920 }
 921
 922 int kvm_unmap_hva_range_hv(struct kvm *kvm, unsigned long start, unsigned long end)
 923 {
 924         if (kvm->arch.using_mmu_notifiers)
 925                 kvm_handle_hva_range(kvm, start, end, kvm_unmap_rmapp);
 926         return 0;
 927 }
 928
 929 void kvmppc_core_flush_memslot_hv(struct kvm *kvm,
 930                                   struct kvm_memory_slot *memslot)
 931 {
 932         unsigned long *rmapp;
 933         unsigned long gfn;
 934         unsigned long n;
 935
 936         rmapp = memslot->arch.rmap;
 937         gfn = memslot->base_gfn;
 938         for (n = memslot->npages; n; --n) {
 939                 /*
 940                  * Testing the present bit without locking is OK because
 941                  * the memslot has been marked invalid already, and hence
 942                  * no new HPTEs referencing this page can be created,
 943                  * thus the present bit can't go from 0 to 1.
 944                  */
 945                 if (*rmapp & KVMPPC_RMAP_PRESENT)
 946                         kvm_unmap_rmapp(kvm, rmapp, gfn);
 947                 ++rmapp;
 948                 ++gfn;
 949         }
 950 }
 951
 952 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
 953                          unsigned long gfn)
 954 {
 955         struct revmap_entry *rev = kvm->arch.revmap;
 956         unsigned long head, i, j;
 957         __be64 *hptep;
 958         int ret = 0;
 959
 960  retry:
 961         lock_rmap(rmapp);
 962         if (*rmapp & KVMPPC_RMAP_REFERENCED) {
 963                 *rmapp &= ~KVMPPC_RMAP_REFERENCED;
 964                 ret = 1;
 965         }
 966         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 967                 unlock_rmap(rmapp);
 968                 return ret;
 969         }
 970
 971         i = head = *rmapp & KVMPPC_RMAP_INDEX;
 972         do {
 973                 hptep = (__be64 *) (kvm->arch.hpt_virt + (i << 4));
 974                 j = rev[i].forw;
 975
 976                 /* If this HPTE isn't referenced, ignore it */
 977                 if (!(be64_to_cpu(hptep[1]) & HPTE_R_R))
 978                         continue;
 979
 980                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 981                         /* unlock rmap before spinning on the HPTE lock */
 982                         unlock_rmap(rmapp);
 983                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
 984                                 cpu_relax();
 985                         goto retry;
 986                 }
 987
 988                 /* Now check and modify the HPTE */
 989                 if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
 990                     (be64_to_cpu(hptep[1]) & HPTE_R_R)) {
 991                         kvmppc_clear_ref_hpte(kvm, hptep, i);
 992                         if (!(rev[i].guest_rpte & HPTE_R_R)) {
 993                                 rev[i].guest_rpte |= HPTE_R_R;
 994                                 note_hpte_modification(kvm, &rev[i]);
 995                         }
 996                         ret = 1;
 997                 }
 998                 hptep[0] &= ~cpu_to_be64(HPTE_V_HVLOCK);
 999         } while ((i = j) != head);
1000
1001         unlock_rmap(rmapp);
1002         return ret;
1003 }
1004
1005 int kvm_age_hva_hv(struct kvm *kvm, unsigned long start, unsigned long end)
1006 {
1007         if (!kvm->arch.using_mmu_notifiers)
1008                 return 0;
1009         return kvm_handle_hva_range(kvm, start, end, kvm_age_rmapp);
1010 }
1011
1012 static int kvm_test_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
1013                               unsigned long gfn)
1014 {
1015         struct revmap_entry *rev = kvm->arch.revmap;
1016         unsigned long head, i, j;
1017         unsigned long *hp;
1018         int ret = 1;
1019
1020         if (*rmapp & KVMPPC_RMAP_REFERENCED)
1021                 return 1;
1022
1023         lock_rmap(rmapp);
1024         if (*rmapp & KVMPPC_RMAP_REFERENCED)
1025                 goto out;
1026
1027         if (*rmapp & KVMPPC_RMAP_PRESENT) {
1028                 i = head = *rmapp & KVMPPC_RMAP_INDEX;
1029                 do {
1030                         hp = (unsigned long *)(kvm->arch.hpt_virt + (i << 4));
1031                         j = rev[i].forw;
1032                         if (be64_to_cpu(hp[1]) & HPTE_R_R)
1033                                 goto out;
1034                 } while ((i = j) != head);
1035         }
1036         ret = 0;
1037
1038  out:
1039         unlock_rmap(rmapp);
1040         return ret;
1041 }
1042
1043 int kvm_test_age_hva_hv(struct kvm *kvm, unsigned long hva)
1044 {
1045         if (!kvm->arch.using_mmu_notifiers)
1046                 return 0;
1047         return kvm_handle_hva(kvm, hva, kvm_test_age_rmapp);
1048 }
1049
1050 void kvm_set_spte_hva_hv(struct kvm *kvm, unsigned long hva, pte_t pte)
1051 {
1052         if (!kvm->arch.using_mmu_notifiers)
1053                 return;
1054         kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
1055 }
1056
1057 static int vcpus_running(struct kvm *kvm)
1058 {
1059         return atomic_read(&kvm->arch.vcpus_running) != 0;
1060 }
1061
1062 /*
1063  * Returns the number of system pages that are dirty.
1064  * This can be more than 1 if we find a huge-page HPTE.
1065  */
1066 static int kvm_test_clear_dirty_npages(struct kvm *kvm, unsigned long *rmapp)
1067 {
1068         struct revmap_entry *rev = kvm->arch.revmap;
1069         unsigned long head, i, j;
1070         unsigned long n;
1071         unsigned long v, r;
1072         __be64 *hptep;
1073         int npages_dirty = 0;
1074
1075  retry:
1076         lock_rmap(rmapp);
1077         if (*rmapp & KVMPPC_RMAP_CHANGED) {
1078                 *rmapp &= ~KVMPPC_RMAP_CHANGED;
1079                 npages_dirty = 1;
1080         }
1081         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
1082                 unlock_rmap(rmapp);
1083                 return npages_dirty;
1084         }
1085
1086         i = head = *rmapp & KVMPPC_RMAP_INDEX;
1087         do {
1088                 unsigned long hptep1;
1089                 hptep = (__be64 *) (kvm->arch.hpt_virt + (i << 4));
1090                 j = rev[i].forw;
1091
1092                 /*
1093                  * Checking the C (changed) bit here is racy since there
1094                  * is no guarantee about when the hardware writes it back.
1095                  * If the HPTE is not writable then it is stable since the
1096                  * page can't be written to, and we would have done a tlbie
1097                  * (which forces the hardware to complete any writeback)
1098                  * when making the HPTE read-only.
1099                  * If vcpus are running then this call is racy anyway
1100                  * since the page could get dirtied subsequently, so we
1101                  * expect there to be a further call which would pick up
1102                  * any delayed C bit writeback.
1103                  * Otherwise we need to do the tlbie even if C==0 in
1104                  * order to pick up any delayed writeback of C.
1105                  */
1106                 hptep1 = be64_to_cpu(hptep[1]);
1107                 if (!(hptep1 & HPTE_R_C) &&
1108                     (!hpte_is_writable(hptep1) || vcpus_running(kvm)))
1109                         continue;
1110
1111                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
1112                         /* unlock rmap before spinning on the HPTE lock */
1113                         unlock_rmap(rmapp);
1114                         while (hptep[0] & cpu_to_be64(HPTE_V_HVLOCK))
1115                                 cpu_relax();
1116                         goto retry;
1117                 }
1118
1119                 /* Now check and modify the HPTE */
1120                 if (!(hptep[0] & cpu_to_be64(HPTE_V_VALID)))
1121                         continue;
1122
1123                 /* need to make it temporarily absent so C is stable */
1124                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
1125                 kvmppc_invalidate_hpte(kvm, hptep, i);
1126                 v = be64_to_cpu(hptep[0]);
1127                 r = be64_to_cpu(hptep[1]);
1128                 if (r & HPTE_R_C) {
1129                         hptep[1] = cpu_to_be64(r & ~HPTE_R_C);
1130                         if (!(rev[i].guest_rpte & HPTE_R_C)) {
1131                                 rev[i].guest_rpte |= HPTE_R_C;
1132                                 note_hpte_modification(kvm, &rev[i]);
1133                         }
1134                         n = hpte_page_size(v, r);
1135                         n = (n + PAGE_SIZE - 1) >> PAGE_SHIFT;
1136                         if (n > npages_dirty)
1137                                 npages_dirty = n;
1138                         eieio();
1139                 }
1140                 v &= ~(HPTE_V_ABSENT | HPTE_V_HVLOCK);
1141                 v |= HPTE_V_VALID;
1142                 hptep[0] = cpu_to_be64(v);
1143         } while ((i = j) != head);
1144
1145         unlock_rmap(rmapp);
1146         return npages_dirty;
1147 }
1148
1149 static void harvest_vpa_dirty(struct kvmppc_vpa *vpa,
1150                               struct kvm_memory_slot *memslot,
1151                               unsigned long *map)
1152 {
1153         unsigned long gfn;
1154
1155         if (!vpa->dirty || !vpa->pinned_addr)
1156                 return;
1157         gfn = vpa->gpa >> PAGE_SHIFT;
1158         if (gfn < memslot->base_gfn ||
1159             gfn >= memslot->base_gfn + memslot->npages)
1160                 return;
1161
1162         vpa->dirty = false;
1163         if (map)
1164                 __set_bit_le(gfn - memslot->base_gfn, map);
1165 }
1166
1167 long kvmppc_hv_get_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot,
1168                              unsigned long *map)
1169 {
1170         unsigned long i, j;
1171         unsigned long *rmapp;
1172         struct kvm_vcpu *vcpu;
1173
1174         preempt_disable();
1175         rmapp = memslot->arch.rmap;
1176         for (i = 0; i < memslot->npages; ++i) {
1177                 int npages = kvm_test_clear_dirty_npages(kvm, rmapp);
1178                 /*
1179                  * Note that if npages > 0 then i must be a multiple of npages,
1180                  * since we always put huge-page HPTEs in the rmap chain
1181                  * corresponding to their page base address.
1182                  */
1183                 if (npages && map)
1184                         for (j = i; npages; ++j, --npages)
1185                                 __set_bit_le(j, map);
1186                 ++rmapp;
1187         }
1188
1189         /* Harvest dirty bits from VPA and DTL updates */
1190         /* Note: we never modify the SLB shadow buffer areas */
1191         kvm_for_each_vcpu(i, vcpu, kvm) {
1192                 spin_lock(&vcpu->arch.vpa_update_lock);
1193                 harvest_vpa_dirty(&vcpu->arch.vpa, memslot, map);
1194                 harvest_vpa_dirty(&vcpu->arch.dtl, memslot, map);
1195                 spin_unlock(&vcpu->arch.vpa_update_lock);
1196         }
1197         preempt_enable();
1198         return 0;
1199 }
1200
1201 void *kvmppc_pin_guest_page(struct kvm *kvm, unsigned long gpa,
1202                             unsigned long *nb_ret)
1203 {
1204         struct kvm_memory_slot *memslot;
1205         unsigned long gfn = gpa >> PAGE_SHIFT;
1206         struct page *page, *pages[1];
1207         int npages;
1208         unsigned long hva, offset;
1209         unsigned long pa;
1210         unsigned long *physp;
1211         int srcu_idx;
1212
1213         srcu_idx = srcu_read_lock(&kvm->srcu);
1214         memslot = gfn_to_memslot(kvm, gfn);
1215         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
1216                 goto err;
1217         if (!kvm->arch.using_mmu_notifiers) {
1218                 physp = memslot->arch.slot_phys;
1219                 if (!physp)
1220                         goto err;
1221                 physp += gfn - memslot->base_gfn;
1222                 pa = *physp;
1223                 if (!pa) {
1224                         if (kvmppc_get_guest_page(kvm, gfn, memslot,
1225                                                   PAGE_SIZE) < 0)
1226                                 goto err;
1227                         pa = *physp;
1228                 }
1229                 page = pfn_to_page(pa >> PAGE_SHIFT);
1230                 get_page(page);
1231         } else {
1232                 hva = gfn_to_hva_memslot(memslot, gfn);
1233                 npages = get_user_pages_fast(hva, 1, 1, pages);
1234                 if (npages < 1)
1235                         goto err;
1236                 page = pages[0];
1237         }
1238         srcu_read_unlock(&kvm->srcu, srcu_idx);
1239
1240         offset = gpa & (PAGE_SIZE - 1);
1241         if (nb_ret)
1242                 *nb_ret = PAGE_SIZE - offset;
1243         return page_address(page) + offset;
1244
1245  err:
1246         srcu_read_unlock(&kvm->srcu, srcu_idx);
1247         return NULL;
1248 }
1249
1250 void kvmppc_unpin_guest_page(struct kvm *kvm, void *va, unsigned long gpa,
1251                              bool dirty)
1252 {
1253         struct page *page = virt_to_page(va);
1254         struct kvm_memory_slot *memslot;
1255         unsigned long gfn;
1256         unsigned long *rmap;
1257         int srcu_idx;
1258
1259         put_page(page);
1260
1261         if (!dirty || !kvm->arch.using_mmu_notifiers)
1262                 return;
1263
1264         /* We need to mark this page dirty in the rmap chain */
1265         gfn = gpa >> PAGE_SHIFT;
1266         srcu_idx = srcu_read_lock(&kvm->srcu);
1267         memslot = gfn_to_memslot(kvm, gfn);
1268         if (memslot) {
1269                 rmap = &memslot->arch.rmap[gfn - memslot->base_gfn];
1270                 lock_rmap(rmap);
1271                 *rmap |= KVMPPC_RMAP_CHANGED;
1272                 unlock_rmap(rmap);
1273         }
1274         srcu_read_unlock(&kvm->srcu, srcu_idx);
1275 }
1276
1277 /*
1278  * Functions for reading and writing the hash table via reads and
1279  * writes on a file descriptor.
1280  *
1281  * Reads return the guest view of the hash table, which has to be
1282  * pieced together from the real hash table and the guest_rpte
1283  * values in the revmap array.
1284  *
1285  * On writes, each HPTE written is considered in turn, and if it
1286  * is valid, it is written to the HPT as if an H_ENTER with the
1287  * exact flag set was done.  When the invalid count is non-zero
1288  * in the header written to the stream, the kernel will make
1289  * sure that that many HPTEs are invalid, and invalidate them
1290  * if not.
1291  */
1292
1293 struct kvm_htab_ctx {
1294         unsigned long   index;
1295         unsigned long   flags;
1296         struct kvm      *kvm;
1297         int             first_pass;
1298 };
1299
1300 #define HPTE_SIZE       (2 * sizeof(unsigned long))
1301
1302 /*
1303  * Returns 1 if this HPT entry has been modified or has pending
1304  * R/C bit changes.
1305  */
1306 static int hpte_dirty(struct revmap_entry *revp, __be64 *hptp)
1307 {
1308         unsigned long rcbits_unset;
1309
1310         if (revp->guest_rpte & HPTE_GR_MODIFIED)
1311                 return 1;
1312
1313         /* Also need to consider changes in reference and changed bits */
1314         rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1315         if ((be64_to_cpu(hptp[0]) & HPTE_V_VALID) &&
1316             (be64_to_cpu(hptp[1]) & rcbits_unset))
1317                 return 1;
1318
1319         return 0;
1320 }
1321
1322 static long record_hpte(unsigned long flags, __be64 *hptp,
1323                         unsigned long *hpte, struct revmap_entry *revp,
1324                         int want_valid, int first_pass)
1325 {
1326         unsigned long v, r;
1327         unsigned long rcbits_unset;
1328         int ok = 1;
1329         int valid, dirty;
1330
1331         /* Unmodified entries are uninteresting except on the first pass */
1332         dirty = hpte_dirty(revp, hptp);
1333         if (!first_pass && !dirty)
1334                 return 0;
1335
1336         valid = 0;
1337         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1338                 valid = 1;
1339                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) &&
1340                     !(be64_to_cpu(hptp[0]) & HPTE_V_BOLTED))
1341                         valid = 0;
1342         }
1343         if (valid != want_valid)
1344                 return 0;
1345
1346         v = r = 0;
1347         if (valid || dirty) {
1348                 /* lock the HPTE so it's stable and read it */
1349                 preempt_disable();
1350                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
1351                         cpu_relax();
1352                 v = be64_to_cpu(hptp[0]);
1353
1354                 /* re-evaluate valid and dirty from synchronized HPTE value */
1355                 valid = !!(v & HPTE_V_VALID);
1356                 dirty = !!(revp->guest_rpte & HPTE_GR_MODIFIED);
1357
1358                 /* Harvest R and C into guest view if necessary */
1359                 rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1360                 if (valid && (rcbits_unset & be64_to_cpu(hptp[1]))) {
1361                         revp->guest_rpte |= (be64_to_cpu(hptp[1]) &
1362                                 (HPTE_R_R | HPTE_R_C)) | HPTE_GR_MODIFIED;
1363                         dirty = 1;
1364                 }
1365
1366                 if (v & HPTE_V_ABSENT) {
1367                         v &= ~HPTE_V_ABSENT;
1368                         v |= HPTE_V_VALID;
1369                         valid = 1;
1370                 }
1371                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) && !(v & HPTE_V_BOLTED))
1372                         valid = 0;
1373
1374                 r = revp->guest_rpte;
1375                 /* only clear modified if this is the right sort of entry */
1376                 if (valid == want_valid && dirty) {
1377                         r &= ~HPTE_GR_MODIFIED;
1378                         revp->guest_rpte = r;
1379                 }
1380                 asm volatile(PPC_RELEASE_BARRIER "" : : : "memory");
1381                 hptp[0] &= ~cpu_to_be64(HPTE_V_HVLOCK);
1382                 preempt_enable();
1383                 if (!(valid == want_valid && (first_pass || dirty)))
1384                         ok = 0;
1385         }
1386         hpte[0] = cpu_to_be64(v);
1387         hpte[1] = cpu_to_be64(r);
1388         return ok;
1389 }
1390
1391 static ssize_t kvm_htab_read(struct file *file, char __user *buf,
1392                              size_t count, loff_t *ppos)
1393 {
1394         struct kvm_htab_ctx *ctx = file->private_data;
1395         struct kvm *kvm = ctx->kvm;
1396         struct kvm_get_htab_header hdr;
1397         __be64 *hptp;
1398         struct revmap_entry *revp;
1399         unsigned long i, nb, nw;
1400         unsigned long __user *lbuf;
1401         struct kvm_get_htab_header __user *hptr;
1402         unsigned long flags;
1403         int first_pass;
1404         unsigned long hpte[2];
1405
1406         if (!access_ok(VERIFY_WRITE, buf, count))
1407                 return -EFAULT;
1408
1409         first_pass = ctx->first_pass;
1410         flags = ctx->flags;
1411
1412         i = ctx->index;
1413         hptp = (__be64 *)(kvm->arch.hpt_virt + (i * HPTE_SIZE));
1414         revp = kvm->arch.revmap + i;
1415         lbuf = (unsigned long __user *)buf;
1416
1417         nb = 0;
1418         while (nb + sizeof(hdr) + HPTE_SIZE < count) {
1419                 /* Initialize header */
1420                 hptr = (struct kvm_get_htab_header __user *)buf;
1421                 hdr.n_valid = 0;
1422                 hdr.n_invalid = 0;
1423                 nw = nb;
1424                 nb += sizeof(hdr);
1425                 lbuf = (unsigned long __user *)(buf + sizeof(hdr));
1426
1427                 /* Skip uninteresting entries, i.e. clean on not-first pass */
1428                 if (!first_pass) {
1429                         while (i < kvm->arch.hpt_npte &&
1430                                !hpte_dirty(revp, hptp)) {
1431                                 ++i;
1432                                 hptp += 2;
1433                                 ++revp;
1434                         }
1435                 }
1436                 hdr.index = i;
1437
1438                 /* Grab a series of valid entries */
1439                 while (i < kvm->arch.hpt_npte &&
1440                        hdr.n_valid < 0xffff &&
1441                        nb + HPTE_SIZE < count &&
1442                        record_hpte(flags, hptp, hpte, revp, 1, first_pass)) {
1443                         /* valid entry, write it out */
1444                         ++hdr.n_valid;
1445                         if (__put_user(hpte[0], lbuf) ||
1446                             __put_user(hpte[1], lbuf + 1))
1447                                 return -EFAULT;
1448                         nb += HPTE_SIZE;
1449                         lbuf += 2;
1450                         ++i;
1451                         hptp += 2;
1452                         ++revp;
1453                 }
1454                 /* Now skip invalid entries while we can */
1455                 while (i < kvm->arch.hpt_npte &&
1456                        hdr.n_invalid < 0xffff &&
1457                        record_hpte(flags, hptp, hpte, revp, 0, first_pass)) {
1458                         /* found an invalid entry */
1459                         ++hdr.n_invalid;
1460                         ++i;
1461                         hptp += 2;
1462                         ++revp;
1463                 }
1464
1465                 if (hdr.n_valid || hdr.n_invalid) {
1466                         /* write back the header */
1467                         if (__copy_to_user(hptr, &hdr, sizeof(hdr)))
1468                                 return -EFAULT;
1469                         nw = nb;
1470                         buf = (char __user *)lbuf;
1471                 } else {
1472                         nb = nw;
1473                 }
1474
1475                 /* Check if we've wrapped around the hash table */
1476                 if (i >= kvm->arch.hpt_npte) {
1477                         i = 0;
1478                         ctx->first_pass = 0;
1479                         break;
1480                 }
1481         }
1482
1483         ctx->index = i;
1484
1485         return nb;
1486 }
1487
1488 static ssize_t kvm_htab_write(struct file *file, const char __user *buf,
1489                               size_t count, loff_t *ppos)
1490 {
1491         struct kvm_htab_ctx *ctx = file->private_data;
1492         struct kvm *kvm = ctx->kvm;
1493         struct kvm_get_htab_header hdr;
1494         unsigned long i, j;
1495         unsigned long v, r;
1496         unsigned long __user *lbuf;
1497         __be64 *hptp;
1498         unsigned long tmp[2];
1499         ssize_t nb;
1500         long int err, ret;
1501         int rma_setup;
1502
1503         if (!access_ok(VERIFY_READ, buf, count))
1504                 return -EFAULT;
1505
1506         /* lock out vcpus from running while we're doing this */
1507         mutex_lock(&kvm->lock);
1508         rma_setup = kvm->arch.rma_setup_done;
1509         if (rma_setup) {
1510                 kvm->arch.rma_setup_done = 0;   /* temporarily */
1511                 /* order rma_setup_done vs. vcpus_running */
1512                 smp_mb();
1513                 if (atomic_read(&kvm->arch.vcpus_running)) {
1514                         kvm->arch.rma_setup_done = 1;
1515                         mutex_unlock(&kvm->lock);
1516                         return -EBUSY;
1517                 }
1518         }
1519
1520         err = 0;
1521         for (nb = 0; nb + sizeof(hdr) <= count; ) {
1522                 err = -EFAULT;
1523                 if (__copy_from_user(&hdr, buf, sizeof(hdr)))
1524                         break;
1525
1526                 err = 0;
1527                 if (nb + hdr.n_valid * HPTE_SIZE > count)
1528                         break;
1529
1530                 nb += sizeof(hdr);
1531                 buf += sizeof(hdr);
1532
1533                 err = -EINVAL;
1534                 i = hdr.index;
1535                 if (i >= kvm->arch.hpt_npte ||
1536                     i + hdr.n_valid + hdr.n_invalid > kvm->arch.hpt_npte)
1537                         break;
1538
1539                 hptp = (__be64 *)(kvm->arch.hpt_virt + (i * HPTE_SIZE));
1540                 lbuf = (unsigned long __user *)buf;
1541                 for (j = 0; j < hdr.n_valid; ++j) {
1542                         err = -EFAULT;
1543                         if (__get_user(v, lbuf) || __get_user(r, lbuf + 1))
1544                                 goto out;
1545                         err = -EINVAL;
1546                         if (!(v & HPTE_V_VALID))
1547                                 goto out;
1548                         lbuf += 2;
1549                         nb += HPTE_SIZE;
1550
1551                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1552                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1553                         err = -EIO;
1554                         ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, i, v, r,
1555                                                          tmp);
1556                         if (ret != H_SUCCESS) {
1557                                 pr_err("kvm_htab_write ret %ld i=%ld v=%lx "
1558                                        "r=%lx\n", ret, i, v, r);
1559                                 goto out;
1560                         }
1561                         if (!rma_setup && is_vrma_hpte(v)) {
1562                                 unsigned long psize = hpte_base_page_size(v, r);
1563                                 unsigned long senc = slb_pgsize_encoding(psize);
1564                                 unsigned long lpcr;
1565
1566                                 kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
1567                                         (VRMA_VSID << SLB_VSID_SHIFT_1T);
1568                                 lpcr = senc << (LPCR_VRMASD_SH - 4);
1569                                 kvmppc_update_lpcr(kvm, lpcr, LPCR_VRMASD);
1570                                 rma_setup = 1;
1571                         }
1572                         ++i;
1573                         hptp += 2;
1574                 }
1575
1576                 for (j = 0; j < hdr.n_invalid; ++j) {
1577                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1578                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1579                         ++i;
1580                         hptp += 2;
1581                 }
1582                 err = 0;
1583         }
1584
1585  out:
1586         /* Order HPTE updates vs. rma_setup_done */
1587         smp_wmb();
1588         kvm->arch.rma_setup_done = rma_setup;
1589         mutex_unlock(&kvm->lock);
1590
1591         if (err)
1592                 return err;
1593         return nb;
1594 }
1595
1596 static int kvm_htab_release(struct inode *inode, struct file *filp)
1597 {
1598         struct kvm_htab_ctx *ctx = filp->private_data;
1599
1600         filp->private_data = NULL;
1601         if (!(ctx->flags & KVM_GET_HTAB_WRITE))
1602                 atomic_dec(&ctx->kvm->arch.hpte_mod_interest);
1603         kvm_put_kvm(ctx->kvm);
1604         kfree(ctx);
1605         return 0;
1606 }
1607
1608 static const struct file_operations kvm_htab_fops = {
1609         .read           = kvm_htab_read,
1610         .write          = kvm_htab_write,
1611         .llseek         = default_llseek,
1612         .release        = kvm_htab_release,
1613 };
1614
1615 int kvm_vm_ioctl_get_htab_fd(struct kvm *kvm, struct kvm_get_htab_fd *ghf)
1616 {
1617         int ret;
1618         struct kvm_htab_ctx *ctx;
1619         int rwflag;
1620
1621         /* reject flags we don't recognize */
1622         if (ghf->flags & ~(KVM_GET_HTAB_BOLTED_ONLY | KVM_GET_HTAB_WRITE))
1623                 return -EINVAL;
1624         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1625         if (!ctx)
1626                 return -ENOMEM;
1627         kvm_get_kvm(kvm);
1628         ctx->kvm = kvm;
1629         ctx->index = ghf->start_index;
1630         ctx->flags = ghf->flags;
1631         ctx->first_pass = 1;
1632
1633         rwflag = (ghf->flags & KVM_GET_HTAB_WRITE) ? O_WRONLY : O_RDONLY;
1634         ret = anon_inode_getfd("kvm-htab", &kvm_htab_fops, ctx, rwflag | O_CLOEXEC);
1635         if (ret < 0) {
1636                 kvm_put_kvm(kvm);
1637                 return ret;
1638         }
1639
1640         if (rwflag == O_RDONLY) {
1641                 mutex_lock(&kvm->slots_lock);
1642                 atomic_inc(&kvm->arch.hpte_mod_interest);
1643                 /* make sure kvmppc_do_h_enter etc. see the increment */
1644                 synchronize_srcu_expedited(&kvm->srcu);
1645                 mutex_unlock(&kvm->slots_lock);
1646         }
1647
1648         return ret;
1649 }
1650
1651 void kvmppc_mmu_book3s_hv_init(struct kvm_vcpu *vcpu)
1652 {
1653         struct kvmppc_mmu *mmu = &vcpu->arch.mmu;
1654
1655         if (cpu_has_feature(CPU_FTR_ARCH_206))
1656                 vcpu->arch.slb_nr = 32;         /* POWER7 */
1657         else
1658                 vcpu->arch.slb_nr = 64;
1659
1660         mmu->xlate = kvmppc_mmu_book3s_64_hv_xlate;
1661         mmu->reset_msr = kvmppc_mmu_book3s_64_hv_reset_msr;
1662
1663         vcpu->arch.hflags |= BOOK3S_HFLAG_SLB;
1664 }