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 #include "book3s_hv_cma.h"
  41
  42 /* POWER7 has 10-bit LPIDs, PPC970 has 6-bit LPIDs */
  43 #define MAX_LPID_970    63
  44
  45 /* Power architecture requires HPT is at least 256kB */
  46 #define PPC_MIN_HPT_ORDER       18
  47
  48 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
  49                                 long pte_index, unsigned long pteh,
  50                                 unsigned long ptel, unsigned long *pte_idx_ret);
  51 static void kvmppc_rmap_reset(struct kvm *kvm);
  52
  53 long kvmppc_alloc_hpt(struct kvm *kvm, u32 *htab_orderp)
  54 {
  55         unsigned long hpt;
  56         struct revmap_entry *rev;
  57         struct page *page = NULL;
  58         long order = KVM_DEFAULT_HPT_ORDER;
  59
  60         if (htab_orderp) {
  61                 order = *htab_orderp;
  62                 if (order < PPC_MIN_HPT_ORDER)
  63                         order = PPC_MIN_HPT_ORDER;
  64         }
  65
  66         kvm->arch.hpt_cma_alloc = 0;
  67         /*
  68          * try first to allocate it from the kernel page allocator.
  69          * We keep the CMA reserved for failed allocation.
  70          */
  71         hpt = __get_free_pages(GFP_KERNEL | __GFP_ZERO | __GFP_REPEAT |
  72                                __GFP_NOWARN, order - PAGE_SHIFT);
  73
  74         /* Next try to allocate from the preallocated pool */
  75         if (!hpt) {
  76                 VM_BUG_ON(order < KVM_CMA_CHUNK_ORDER);
  77                 page = kvm_alloc_hpt(1 << (order - PAGE_SHIFT));
  78                 if (page) {
  79                         hpt = (unsigned long)pfn_to_kaddr(page_to_pfn(page));
  80                         kvm->arch.hpt_cma_alloc = 1;
  81                 } else
  82                         --order;
  83         }
  84
  85         /* Lastly try successively smaller sizes from the page allocator */
  86         while (!hpt && order > PPC_MIN_HPT_ORDER) {
  87                 hpt = __get_free_pages(GFP_KERNEL|__GFP_ZERO|__GFP_REPEAT|
  88                                        __GFP_NOWARN, order - PAGE_SHIFT);
  89                 if (!hpt)
  90                         --order;
  91         }
  92
  93         if (!hpt)
  94                 return -ENOMEM;
  95
  96         kvm->arch.hpt_virt = hpt;
  97         kvm->arch.hpt_order = order;
  98         /* HPTEs are 2**4 bytes long */
  99         kvm->arch.hpt_npte = 1ul << (order - 4);
 100         /* 128 (2**7) bytes in each HPTEG */
 101         kvm->arch.hpt_mask = (1ul << (order - 7)) - 1;
 102
 103         /* Allocate reverse map array */
 104         rev = vmalloc(sizeof(struct revmap_entry) * kvm->arch.hpt_npte);
 105         if (!rev) {
 106                 pr_err("kvmppc_alloc_hpt: Couldn't alloc reverse map array\n");
 107                 goto out_freehpt;
 108         }
 109         kvm->arch.revmap = rev;
 110         kvm->arch.sdr1 = __pa(hpt) | (order - 18);
 111
 112         pr_info("KVM guest htab at %lx (order %ld), LPID %x\n",
 113                 hpt, order, kvm->arch.lpid);
 114
 115         if (htab_orderp)
 116                 *htab_orderp = order;
 117         return 0;
 118
 119  out_freehpt:
 120         if (kvm->arch.hpt_cma_alloc)
 121                 kvm_release_hpt(page, 1 << (order - PAGE_SHIFT));
 122         else
 123                 free_pages(hpt, order - PAGE_SHIFT);
 124         return -ENOMEM;
 125 }
 126
 127 long kvmppc_alloc_reset_hpt(struct kvm *kvm, u32 *htab_orderp)
 128 {
 129         long err = -EBUSY;
 130         long order;
 131
 132         mutex_lock(&kvm->lock);
 133         if (kvm->arch.rma_setup_done) {
 134                 kvm->arch.rma_setup_done = 0;
 135                 /* order rma_setup_done vs. vcpus_running */
 136                 smp_mb();
 137                 if (atomic_read(&kvm->arch.vcpus_running)) {
 138                         kvm->arch.rma_setup_done = 1;
 139                         goto out;
 140                 }
 141         }
 142         if (kvm->arch.hpt_virt) {
 143                 order = kvm->arch.hpt_order;
 144                 /* Set the entire HPT to 0, i.e. invalid HPTEs */
 145                 memset((void *)kvm->arch.hpt_virt, 0, 1ul << order);
 146                 /*
 147                  * Reset all the reverse-mapping chains for all memslots
 148                  */
 149                 kvmppc_rmap_reset(kvm);
 150                 /* Ensure that each vcpu will flush its TLB on next entry. */
 151                 cpumask_setall(&kvm->arch.need_tlb_flush);
 152                 *htab_orderp = order;
 153                 err = 0;
 154         } else {
 155                 err = kvmppc_alloc_hpt(kvm, htab_orderp);
 156                 order = *htab_orderp;
 157         }
 158  out:
 159         mutex_unlock(&kvm->lock);
 160         return err;
 161 }
 162
 163 void kvmppc_free_hpt(struct kvm *kvm)
 164 {
 165         kvmppc_free_lpid(kvm->arch.lpid);
 166         vfree(kvm->arch.revmap);
 167         if (kvm->arch.hpt_cma_alloc)
 168                 kvm_release_hpt(virt_to_page(kvm->arch.hpt_virt),
 169                                 1 << (kvm->arch.hpt_order - PAGE_SHIFT));
 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 static void kvmppc_mmu_book3s_64_hv_reset_msr(struct kvm_vcpu *vcpu)
 264 {
 265         kvmppc_set_msr(vcpu, MSR_SF | MSR_ME);
 266 }
 267
 268 /*
 269  * This is called to get a reference to a guest page if there isn't
 270  * one already in the memslot->arch.slot_phys[] array.
 271  */
 272 static long kvmppc_get_guest_page(struct kvm *kvm, unsigned long gfn,
 273                                   struct kvm_memory_slot *memslot,
 274                                   unsigned long psize)
 275 {
 276         unsigned long start;
 277         long np, err;
 278         struct page *page, *hpage, *pages[1];
 279         unsigned long s, pgsize;
 280         unsigned long *physp;
 281         unsigned int is_io, got, pgorder;
 282         struct vm_area_struct *vma;
 283         unsigned long pfn, i, npages;
 284
 285         physp = memslot->arch.slot_phys;
 286         if (!physp)
 287                 return -EINVAL;
 288         if (physp[gfn - memslot->base_gfn])
 289                 return 0;
 290
 291         is_io = 0;
 292         got = 0;
 293         page = NULL;
 294         pgsize = psize;
 295         err = -EINVAL;
 296         start = gfn_to_hva_memslot(memslot, gfn);
 297
 298         /* Instantiate and get the page we want access to */
 299         np = get_user_pages_fast(start, 1, 1, pages);
 300         if (np != 1) {
 301                 /* Look up the vma for the page */
 302                 down_read(&current->mm->mmap_sem);
 303                 vma = find_vma(current->mm, start);
 304                 if (!vma || vma->vm_start > start ||
 305                     start + psize > vma->vm_end ||
 306                     !(vma->vm_flags & VM_PFNMAP))
 307                         goto up_err;
 308                 is_io = hpte_cache_bits(pgprot_val(vma->vm_page_prot));
 309                 pfn = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 310                 /* check alignment of pfn vs. requested page size */
 311                 if (psize > PAGE_SIZE && (pfn & ((psize >> PAGE_SHIFT) - 1)))
 312                         goto up_err;
 313                 up_read(&current->mm->mmap_sem);
 314
 315         } else {
 316                 page = pages[0];
 317                 got = KVMPPC_GOT_PAGE;
 318
 319                 /* See if this is a large page */
 320                 s = PAGE_SIZE;
 321                 if (PageHuge(page)) {
 322                         hpage = compound_head(page);
 323                         s <<= compound_order(hpage);
 324                         /* Get the whole large page if slot alignment is ok */
 325                         if (s > psize && slot_is_aligned(memslot, s) &&
 326                             !(memslot->userspace_addr & (s - 1))) {
 327                                 start &= ~(s - 1);
 328                                 pgsize = s;
 329                                 get_page(hpage);
 330                                 put_page(page);
 331                                 page = hpage;
 332                         }
 333                 }
 334                 if (s < psize)
 335                         goto out;
 336                 pfn = page_to_pfn(page);
 337         }
 338
 339         npages = pgsize >> PAGE_SHIFT;
 340         pgorder = __ilog2(npages);
 341         physp += (gfn - memslot->base_gfn) & ~(npages - 1);
 342         spin_lock(&kvm->arch.slot_phys_lock);
 343         for (i = 0; i < npages; ++i) {
 344                 if (!physp[i]) {
 345                         physp[i] = ((pfn + i) << PAGE_SHIFT) +
 346                                 got + is_io + pgorder;
 347                         got = 0;
 348                 }
 349         }
 350         spin_unlock(&kvm->arch.slot_phys_lock);
 351         err = 0;
 352
 353  out:
 354         if (got)
 355                 put_page(page);
 356         return err;
 357
 358  up_err:
 359         up_read(&current->mm->mmap_sem);
 360         return err;
 361 }
 362
 363 long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
 364                                 long pte_index, unsigned long pteh,
 365                                 unsigned long ptel, unsigned long *pte_idx_ret)
 366 {
 367         unsigned long psize, gpa, gfn;
 368         struct kvm_memory_slot *memslot;
 369         long ret;
 370
 371         if (kvm->arch.using_mmu_notifiers)
 372                 goto do_insert;
 373
 374         psize = hpte_page_size(pteh, ptel);
 375         if (!psize)
 376                 return H_PARAMETER;
 377
 378         pteh &= ~(HPTE_V_HVLOCK | HPTE_V_ABSENT | HPTE_V_VALID);
 379
 380         /* Find the memslot (if any) for this address */
 381         gpa = (ptel & HPTE_R_RPN) & ~(psize - 1);
 382         gfn = gpa >> PAGE_SHIFT;
 383         memslot = gfn_to_memslot(kvm, gfn);
 384         if (memslot && !(memslot->flags & KVM_MEMSLOT_INVALID)) {
 385                 if (!slot_is_aligned(memslot, psize))
 386                         return H_PARAMETER;
 387                 if (kvmppc_get_guest_page(kvm, gfn, memslot, psize) < 0)
 388                         return H_PARAMETER;
 389         }
 390
 391  do_insert:
 392         /* Protect linux PTE lookup from page table destruction */
 393         rcu_read_lock_sched();  /* this disables preemption too */
 394         ret = kvmppc_do_h_enter(kvm, flags, pte_index, pteh, ptel,
 395                                 current->mm->pgd, false, pte_idx_ret);
 396         rcu_read_unlock_sched();
 397         if (ret == H_TOO_HARD) {
 398                 /* this can't happen */
 399                 pr_err("KVM: Oops, kvmppc_h_enter returned too hard!\n");
 400                 ret = H_RESOURCE;       /* or something */
 401         }
 402         return ret;
 403
 404 }
 405
 406 /*
 407  * We come here on a H_ENTER call from the guest when we are not
 408  * using mmu notifiers and we don't have the requested page pinned
 409  * already.
 410  */
 411 long kvmppc_virtmode_h_enter(struct kvm_vcpu *vcpu, unsigned long flags,
 412                              long pte_index, unsigned long pteh,
 413                              unsigned long ptel)
 414 {
 415         return kvmppc_virtmode_do_h_enter(vcpu->kvm, flags, pte_index,
 416                                           pteh, ptel, &vcpu->arch.gpr[4]);
 417 }
 418
 419 static struct kvmppc_slb *kvmppc_mmu_book3s_hv_find_slbe(struct kvm_vcpu *vcpu,
 420                                                          gva_t eaddr)
 421 {
 422         u64 mask;
 423         int i;
 424
 425         for (i = 0; i < vcpu->arch.slb_nr; i++) {
 426                 if (!(vcpu->arch.slb[i].orige & SLB_ESID_V))
 427                         continue;
 428
 429                 if (vcpu->arch.slb[i].origv & SLB_VSID_B_1T)
 430                         mask = ESID_MASK_1T;
 431                 else
 432                         mask = ESID_MASK;
 433
 434                 if (((vcpu->arch.slb[i].orige ^ eaddr) & mask) == 0)
 435                         return &vcpu->arch.slb[i];
 436         }
 437         return NULL;
 438 }
 439
 440 static unsigned long kvmppc_mmu_get_real_addr(unsigned long v, unsigned long r,
 441                         unsigned long ea)
 442 {
 443         unsigned long ra_mask;
 444
 445         ra_mask = hpte_page_size(v, r) - 1;
 446         return (r & HPTE_R_RPN & ~ra_mask) | (ea & ra_mask);
 447 }
 448
 449 static int kvmppc_mmu_book3s_64_hv_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
 450                         struct kvmppc_pte *gpte, bool data, bool iswrite)
 451 {
 452         struct kvm *kvm = vcpu->kvm;
 453         struct kvmppc_slb *slbe;
 454         unsigned long slb_v;
 455         unsigned long pp, key;
 456         unsigned long v, gr;
 457         unsigned long *hptep;
 458         int index;
 459         int virtmode = vcpu->arch.shregs.msr & (data ? MSR_DR : MSR_IR);
 460
 461         /* Get SLB entry */
 462         if (virtmode) {
 463                 slbe = kvmppc_mmu_book3s_hv_find_slbe(vcpu, eaddr);
 464                 if (!slbe)
 465                         return -EINVAL;
 466                 slb_v = slbe->origv;
 467         } else {
 468                 /* real mode access */
 469                 slb_v = vcpu->kvm->arch.vrma_slb_v;
 470         }
 471
 472         /* Find the HPTE in the hash table */
 473         index = kvmppc_hv_find_lock_hpte(kvm, eaddr, slb_v,
 474                                          HPTE_V_VALID | HPTE_V_ABSENT);
 475         if (index < 0)
 476                 return -ENOENT;
 477         hptep = (unsigned long *)(kvm->arch.hpt_virt + (index << 4));
 478         v = hptep[0] & ~HPTE_V_HVLOCK;
 479         gr = kvm->arch.revmap[index].guest_rpte;
 480
 481         /* Unlock the HPTE */
 482         asm volatile("lwsync" : : : "memory");
 483         hptep[0] = v;
 484
 485         gpte->eaddr = eaddr;
 486         gpte->vpage = ((v & HPTE_V_AVPN) << 4) | ((eaddr >> 12) & 0xfff);
 487
 488         /* Get PP bits and key for permission check */
 489         pp = gr & (HPTE_R_PP0 | HPTE_R_PP);
 490         key = (vcpu->arch.shregs.msr & MSR_PR) ? SLB_VSID_KP : SLB_VSID_KS;
 491         key &= slb_v;
 492
 493         /* Calculate permissions */
 494         gpte->may_read = hpte_read_permission(pp, key);
 495         gpte->may_write = hpte_write_permission(pp, key);
 496         gpte->may_execute = gpte->may_read && !(gr & (HPTE_R_N | HPTE_R_G));
 497
 498         /* Storage key permission check for POWER7 */
 499         if (data && virtmode && cpu_has_feature(CPU_FTR_ARCH_206)) {
 500                 int amrfield = hpte_get_skey_perm(gr, vcpu->arch.amr);
 501                 if (amrfield & 1)
 502                         gpte->may_read = 0;
 503                 if (amrfield & 2)
 504                         gpte->may_write = 0;
 505         }
 506
 507         /* Get the guest physical address */
 508         gpte->raddr = kvmppc_mmu_get_real_addr(v, gr, eaddr);
 509         return 0;
 510 }
 511
 512 /*
 513  * Quick test for whether an instruction is a load or a store.
 514  * If the instruction is a load or a store, then this will indicate
 515  * which it is, at least on server processors.  (Embedded processors
 516  * have some external PID instructions that don't follow the rule
 517  * embodied here.)  If the instruction isn't a load or store, then
 518  * this doesn't return anything useful.
 519  */
 520 static int instruction_is_store(unsigned int instr)
 521 {
 522         unsigned int mask;
 523
 524         mask = 0x10000000;
 525         if ((instr & 0xfc000000) == 0x7c000000)
 526                 mask = 0x100;           /* major opcode 31 */
 527         return (instr & mask) != 0;
 528 }
 529
 530 static int kvmppc_hv_emulate_mmio(struct kvm_run *run, struct kvm_vcpu *vcpu,
 531                                   unsigned long gpa, gva_t ea, int is_store)
 532 {
 533         int ret;
 534         u32 last_inst;
 535         unsigned long srr0 = kvmppc_get_pc(vcpu);
 536
 537         /* We try to load the last instruction.  We don't let
 538          * emulate_instruction do it as it doesn't check what
 539          * kvmppc_ld returns.
 540          * If we fail, we just return to the guest and try executing it again.
 541          */
 542         if (vcpu->arch.last_inst == KVM_INST_FETCH_FAILED) {
 543                 ret = kvmppc_ld(vcpu, &srr0, sizeof(u32), &last_inst, false);
 544                 if (ret != EMULATE_DONE || last_inst == KVM_INST_FETCH_FAILED)
 545                         return RESUME_GUEST;
 546                 vcpu->arch.last_inst = last_inst;
 547         }
 548
 549         /*
 550          * WARNING: We do not know for sure whether the instruction we just
 551          * read from memory is the same that caused the fault in the first
 552          * place.  If the instruction we read is neither an load or a store,
 553          * then it can't access memory, so we don't need to worry about
 554          * enforcing access permissions.  So, assuming it is a load or
 555          * store, we just check that its direction (load or store) is
 556          * consistent with the original fault, since that's what we
 557          * checked the access permissions against.  If there is a mismatch
 558          * we just return and retry the instruction.
 559          */
 560
 561         if (instruction_is_store(kvmppc_get_last_inst(vcpu)) != !!is_store)
 562                 return RESUME_GUEST;
 563
 564         /*
 565          * Emulated accesses are emulated by looking at the hash for
 566          * translation once, then performing the access later. The
 567          * translation could be invalidated in the meantime in which
 568          * point performing the subsequent memory access on the old
 569          * physical address could possibly be a security hole for the
 570          * guest (but not the host).
 571          *
 572          * This is less of an issue for MMIO stores since they aren't
 573          * globally visible. It could be an issue for MMIO loads to
 574          * a certain extent but we'll ignore it for now.
 575          */
 576
 577         vcpu->arch.paddr_accessed = gpa;
 578         vcpu->arch.vaddr_accessed = ea;
 579         return kvmppc_emulate_mmio(run, vcpu);
 580 }
 581
 582 int kvmppc_book3s_hv_page_fault(struct kvm_run *run, struct kvm_vcpu *vcpu,
 583                                 unsigned long ea, unsigned long dsisr)
 584 {
 585         struct kvm *kvm = vcpu->kvm;
 586         unsigned long *hptep, hpte[3], r;
 587         unsigned long mmu_seq, psize, pte_size;
 588         unsigned long gpa, gfn, hva, pfn;
 589         struct kvm_memory_slot *memslot;
 590         unsigned long *rmap;
 591         struct revmap_entry *rev;
 592         struct page *page, *pages[1];
 593         long index, ret, npages;
 594         unsigned long is_io;
 595         unsigned int writing, write_ok;
 596         struct vm_area_struct *vma;
 597         unsigned long rcbits;
 598
 599         /*
 600          * Real-mode code has already searched the HPT and found the
 601          * entry we're interested in.  Lock the entry and check that
 602          * it hasn't changed.  If it has, just return and re-execute the
 603          * instruction.
 604          */
 605         if (ea != vcpu->arch.pgfault_addr)
 606                 return RESUME_GUEST;
 607         index = vcpu->arch.pgfault_index;
 608         hptep = (unsigned long *)(kvm->arch.hpt_virt + (index << 4));
 609         rev = &kvm->arch.revmap[index];
 610         preempt_disable();
 611         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 612                 cpu_relax();
 613         hpte[0] = hptep[0] & ~HPTE_V_HVLOCK;
 614         hpte[1] = hptep[1];
 615         hpte[2] = r = rev->guest_rpte;
 616         asm volatile("lwsync" : : : "memory");
 617         hptep[0] = hpte[0];
 618         preempt_enable();
 619
 620         if (hpte[0] != vcpu->arch.pgfault_hpte[0] ||
 621             hpte[1] != vcpu->arch.pgfault_hpte[1])
 622                 return RESUME_GUEST;
 623
 624         /* Translate the logical address and get the page */
 625         psize = hpte_page_size(hpte[0], r);
 626         gpa = (r & HPTE_R_RPN & ~(psize - 1)) | (ea & (psize - 1));
 627         gfn = gpa >> PAGE_SHIFT;
 628         memslot = gfn_to_memslot(kvm, gfn);
 629
 630         /* No memslot means it's an emulated MMIO region */
 631         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
 632                 return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
 633                                               dsisr & DSISR_ISSTORE);
 634
 635         if (!kvm->arch.using_mmu_notifiers)
 636                 return -EFAULT;         /* should never get here */
 637
 638         /* used to check for invalidations in progress */
 639         mmu_seq = kvm->mmu_notifier_seq;
 640         smp_rmb();
 641
 642         is_io = 0;
 643         pfn = 0;
 644         page = NULL;
 645         pte_size = PAGE_SIZE;
 646         writing = (dsisr & DSISR_ISSTORE) != 0;
 647         /* If writing != 0, then the HPTE must allow writing, if we get here */
 648         write_ok = writing;
 649         hva = gfn_to_hva_memslot(memslot, gfn);
 650         npages = get_user_pages_fast(hva, 1, writing, pages);
 651         if (npages < 1) {
 652                 /* Check if it's an I/O mapping */
 653                 down_read(&current->mm->mmap_sem);
 654                 vma = find_vma(current->mm, hva);
 655                 if (vma && vma->vm_start <= hva && hva + psize <= vma->vm_end &&
 656                     (vma->vm_flags & VM_PFNMAP)) {
 657                         pfn = vma->vm_pgoff +
 658                                 ((hva - vma->vm_start) >> PAGE_SHIFT);
 659                         pte_size = psize;
 660                         is_io = hpte_cache_bits(pgprot_val(vma->vm_page_prot));
 661                         write_ok = vma->vm_flags & VM_WRITE;
 662                 }
 663                 up_read(&current->mm->mmap_sem);
 664                 if (!pfn)
 665                         return -EFAULT;
 666         } else {
 667                 page = pages[0];
 668                 if (PageHuge(page)) {
 669                         page = compound_head(page);
 670                         pte_size <<= compound_order(page);
 671                 }
 672                 /* if the guest wants write access, see if that is OK */
 673                 if (!writing && hpte_is_writable(r)) {
 674                         unsigned int hugepage_shift;
 675                         pte_t *ptep, pte;
 676
 677                         /*
 678                          * We need to protect against page table destruction
 679                          * while looking up and updating the pte.
 680                          */
 681                         rcu_read_lock_sched();
 682                         ptep = find_linux_pte_or_hugepte(current->mm->pgd,
 683                                                          hva, &hugepage_shift);
 684                         if (ptep) {
 685                                 pte = kvmppc_read_update_linux_pte(ptep, 1,
 686                                                            hugepage_shift);
 687                                 if (pte_write(pte))
 688                                         write_ok = 1;
 689                         }
 690                         rcu_read_unlock_sched();
 691                 }
 692                 pfn = page_to_pfn(page);
 693         }
 694
 695         ret = -EFAULT;
 696         if (psize > pte_size)
 697                 goto out_put;
 698
 699         /* Check WIMG vs. the actual page we're accessing */
 700         if (!hpte_cache_flags_ok(r, is_io)) {
 701                 if (is_io)
 702                         return -EFAULT;
 703                 /*
 704                  * Allow guest to map emulated device memory as
 705                  * uncacheable, but actually make it cacheable.
 706                  */
 707                 r = (r & ~(HPTE_R_W|HPTE_R_I|HPTE_R_G)) | HPTE_R_M;
 708         }
 709
 710         /* Set the HPTE to point to pfn */
 711         r = (r & ~(HPTE_R_PP0 - pte_size)) | (pfn << PAGE_SHIFT);
 712         if (hpte_is_writable(r) && !write_ok)
 713                 r = hpte_make_readonly(r);
 714         ret = RESUME_GUEST;
 715         preempt_disable();
 716         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 717                 cpu_relax();
 718         if ((hptep[0] & ~HPTE_V_HVLOCK) != hpte[0] || hptep[1] != hpte[1] ||
 719             rev->guest_rpte != hpte[2])
 720                 /* HPTE has been changed under us; let the guest retry */
 721                 goto out_unlock;
 722         hpte[0] = (hpte[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
 723
 724         rmap = &memslot->arch.rmap[gfn - memslot->base_gfn];
 725         lock_rmap(rmap);
 726
 727         /* Check if we might have been invalidated; let the guest retry if so */
 728         ret = RESUME_GUEST;
 729         if (mmu_notifier_retry(vcpu->kvm, mmu_seq)) {
 730                 unlock_rmap(rmap);
 731                 goto out_unlock;
 732         }
 733
 734         /* Only set R/C in real HPTE if set in both *rmap and guest_rpte */
 735         rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
 736         r &= rcbits | ~(HPTE_R_R | HPTE_R_C);
 737
 738         if (hptep[0] & HPTE_V_VALID) {
 739                 /* HPTE was previously valid, so we need to invalidate it */
 740                 unlock_rmap(rmap);
 741                 hptep[0] |= HPTE_V_ABSENT;
 742                 kvmppc_invalidate_hpte(kvm, hptep, index);
 743                 /* don't lose previous R and C bits */
 744                 r |= hptep[1] & (HPTE_R_R | HPTE_R_C);
 745         } else {
 746                 kvmppc_add_revmap_chain(kvm, rev, rmap, index, 0);
 747         }
 748
 749         hptep[1] = r;
 750         eieio();
 751         hptep[0] = hpte[0];
 752         asm volatile("ptesync" : : : "memory");
 753         preempt_enable();
 754         if (page && hpte_is_writable(r))
 755                 SetPageDirty(page);
 756
 757  out_put:
 758         if (page) {
 759                 /*
 760                  * We drop pages[0] here, not page because page might
 761                  * have been set to the head page of a compound, but
 762                  * we have to drop the reference on the correct tail
 763                  * page to match the get inside gup()
 764                  */
 765                 put_page(pages[0]);
 766         }
 767         return ret;
 768
 769  out_unlock:
 770         hptep[0] &= ~HPTE_V_HVLOCK;
 771         preempt_enable();
 772         goto out_put;
 773 }
 774
 775 static void kvmppc_rmap_reset(struct kvm *kvm)
 776 {
 777         struct kvm_memslots *slots;
 778         struct kvm_memory_slot *memslot;
 779         int srcu_idx;
 780
 781         srcu_idx = srcu_read_lock(&kvm->srcu);
 782         slots = kvm->memslots;
 783         kvm_for_each_memslot(memslot, slots) {
 784                 /*
 785                  * This assumes it is acceptable to lose reference and
 786                  * change bits across a reset.
 787                  */
 788                 memset(memslot->arch.rmap, 0,
 789                        memslot->npages * sizeof(*memslot->arch.rmap));
 790         }
 791         srcu_read_unlock(&kvm->srcu, srcu_idx);
 792 }
 793
 794 static int kvm_handle_hva_range(struct kvm *kvm,
 795                                 unsigned long start,
 796                                 unsigned long end,
 797                                 int (*handler)(struct kvm *kvm,
 798                                                unsigned long *rmapp,
 799                                                unsigned long gfn))
 800 {
 801         int ret;
 802         int retval = 0;
 803         struct kvm_memslots *slots;
 804         struct kvm_memory_slot *memslot;
 805
 806         slots = kvm_memslots(kvm);
 807         kvm_for_each_memslot(memslot, slots) {
 808                 unsigned long hva_start, hva_end;
 809                 gfn_t gfn, gfn_end;
 810
 811                 hva_start = max(start, memslot->userspace_addr);
 812                 hva_end = min(end, memslot->userspace_addr +
 813                                         (memslot->npages << PAGE_SHIFT));
 814                 if (hva_start >= hva_end)
 815                         continue;
 816                 /*
 817                  * {gfn(page) | page intersects with [hva_start, hva_end)} =
 818                  * {gfn, gfn+1, ..., gfn_end-1}.
 819                  */
 820                 gfn = hva_to_gfn_memslot(hva_start, memslot);
 821                 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
 822
 823                 for (; gfn < gfn_end; ++gfn) {
 824                         gfn_t gfn_offset = gfn - memslot->base_gfn;
 825
 826                         ret = handler(kvm, &memslot->arch.rmap[gfn_offset], gfn);
 827                         retval |= ret;
 828                 }
 829         }
 830
 831         return retval;
 832 }
 833
 834 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 835                           int (*handler)(struct kvm *kvm, unsigned long *rmapp,
 836                                          unsigned long gfn))
 837 {
 838         return kvm_handle_hva_range(kvm, hva, hva + 1, handler);
 839 }
 840
 841 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
 842                            unsigned long gfn)
 843 {
 844         struct revmap_entry *rev = kvm->arch.revmap;
 845         unsigned long h, i, j;
 846         unsigned long *hptep;
 847         unsigned long ptel, psize, rcbits;
 848
 849         for (;;) {
 850                 lock_rmap(rmapp);
 851                 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 852                         unlock_rmap(rmapp);
 853                         break;
 854                 }
 855
 856                 /*
 857                  * To avoid an ABBA deadlock with the HPTE lock bit,
 858                  * we can't spin on the HPTE lock while holding the
 859                  * rmap chain lock.
 860                  */
 861                 i = *rmapp & KVMPPC_RMAP_INDEX;
 862                 hptep = (unsigned long *) (kvm->arch.hpt_virt + (i << 4));
 863                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 864                         /* unlock rmap before spinning on the HPTE lock */
 865                         unlock_rmap(rmapp);
 866                         while (hptep[0] & HPTE_V_HVLOCK)
 867                                 cpu_relax();
 868                         continue;
 869                 }
 870                 j = rev[i].forw;
 871                 if (j == i) {
 872                         /* chain is now empty */
 873                         *rmapp &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
 874                 } else {
 875                         /* remove i from chain */
 876                         h = rev[i].back;
 877                         rev[h].forw = j;
 878                         rev[j].back = h;
 879                         rev[i].forw = rev[i].back = i;
 880                         *rmapp = (*rmapp & ~KVMPPC_RMAP_INDEX) | j;
 881                 }
 882
 883                 /* Now check and modify the HPTE */
 884                 ptel = rev[i].guest_rpte;
 885                 psize = hpte_page_size(hptep[0], ptel);
 886                 if ((hptep[0] & HPTE_V_VALID) &&
 887                     hpte_rpn(ptel, psize) == gfn) {
 888                         if (kvm->arch.using_mmu_notifiers)
 889                                 hptep[0] |= HPTE_V_ABSENT;
 890                         kvmppc_invalidate_hpte(kvm, hptep, i);
 891                         /* Harvest R and C */
 892                         rcbits = hptep[1] & (HPTE_R_R | HPTE_R_C);
 893                         *rmapp |= rcbits << KVMPPC_RMAP_RC_SHIFT;
 894                         if (rcbits & ~rev[i].guest_rpte) {
 895                                 rev[i].guest_rpte = ptel | rcbits;
 896                                 note_hpte_modification(kvm, &rev[i]);
 897                         }
 898                 }
 899                 unlock_rmap(rmapp);
 900                 hptep[0] &= ~HPTE_V_HVLOCK;
 901         }
 902         return 0;
 903 }
 904
 905 int kvm_unmap_hva_hv(struct kvm *kvm, unsigned long hva)
 906 {
 907         if (kvm->arch.using_mmu_notifiers)
 908                 kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 909         return 0;
 910 }
 911
 912 int kvm_unmap_hva_range_hv(struct kvm *kvm, unsigned long start, unsigned long end)
 913 {
 914         if (kvm->arch.using_mmu_notifiers)
 915                 kvm_handle_hva_range(kvm, start, end, kvm_unmap_rmapp);
 916         return 0;
 917 }
 918
 919 void kvmppc_core_flush_memslot_hv(struct kvm *kvm,
 920                                   struct kvm_memory_slot *memslot)
 921 {
 922         unsigned long *rmapp;
 923         unsigned long gfn;
 924         unsigned long n;
 925
 926         rmapp = memslot->arch.rmap;
 927         gfn = memslot->base_gfn;
 928         for (n = memslot->npages; n; --n) {
 929                 /*
 930                  * Testing the present bit without locking is OK because
 931                  * the memslot has been marked invalid already, and hence
 932                  * no new HPTEs referencing this page can be created,
 933                  * thus the present bit can't go from 0 to 1.
 934                  */
 935                 if (*rmapp & KVMPPC_RMAP_PRESENT)
 936                         kvm_unmap_rmapp(kvm, rmapp, gfn);
 937                 ++rmapp;
 938                 ++gfn;
 939         }
 940 }
 941
 942 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
 943                          unsigned long gfn)
 944 {
 945         struct revmap_entry *rev = kvm->arch.revmap;
 946         unsigned long head, i, j;
 947         unsigned long *hptep;
 948         int ret = 0;
 949
 950  retry:
 951         lock_rmap(rmapp);
 952         if (*rmapp & KVMPPC_RMAP_REFERENCED) {
 953                 *rmapp &= ~KVMPPC_RMAP_REFERENCED;
 954                 ret = 1;
 955         }
 956         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 957                 unlock_rmap(rmapp);
 958                 return ret;
 959         }
 960
 961         i = head = *rmapp & KVMPPC_RMAP_INDEX;
 962         do {
 963                 hptep = (unsigned long *) (kvm->arch.hpt_virt + (i << 4));
 964                 j = rev[i].forw;
 965
 966                 /* If this HPTE isn't referenced, ignore it */
 967                 if (!(hptep[1] & HPTE_R_R))
 968                         continue;
 969
 970                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 971                         /* unlock rmap before spinning on the HPTE lock */
 972                         unlock_rmap(rmapp);
 973                         while (hptep[0] & HPTE_V_HVLOCK)
 974                                 cpu_relax();
 975                         goto retry;
 976                 }
 977
 978                 /* Now check and modify the HPTE */
 979                 if ((hptep[0] & HPTE_V_VALID) && (hptep[1] & HPTE_R_R)) {
 980                         kvmppc_clear_ref_hpte(kvm, hptep, i);
 981                         if (!(rev[i].guest_rpte & HPTE_R_R)) {
 982                                 rev[i].guest_rpte |= HPTE_R_R;
 983                                 note_hpte_modification(kvm, &rev[i]);
 984                         }
 985                         ret = 1;
 986                 }
 987                 hptep[0] &= ~HPTE_V_HVLOCK;
 988         } while ((i = j) != head);
 989
 990         unlock_rmap(rmapp);
 991         return ret;
 992 }
 993
 994 int kvm_age_hva_hv(struct kvm *kvm, unsigned long hva)
 995 {
 996         if (!kvm->arch.using_mmu_notifiers)
 997                 return 0;
 998         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
 999 }
1000
1001 static int kvm_test_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
1002                               unsigned long gfn)
1003 {
1004         struct revmap_entry *rev = kvm->arch.revmap;
1005         unsigned long head, i, j;
1006         unsigned long *hp;
1007         int ret = 1;
1008
1009         if (*rmapp & KVMPPC_RMAP_REFERENCED)
1010                 return 1;
1011
1012         lock_rmap(rmapp);
1013         if (*rmapp & KVMPPC_RMAP_REFERENCED)
1014                 goto out;
1015
1016         if (*rmapp & KVMPPC_RMAP_PRESENT) {
1017                 i = head = *rmapp & KVMPPC_RMAP_INDEX;
1018                 do {
1019                         hp = (unsigned long *)(kvm->arch.hpt_virt + (i << 4));
1020                         j = rev[i].forw;
1021                         if (hp[1] & HPTE_R_R)
1022                                 goto out;
1023                 } while ((i = j) != head);
1024         }
1025         ret = 0;
1026
1027  out:
1028         unlock_rmap(rmapp);
1029         return ret;
1030 }
1031
1032 int kvm_test_age_hva_hv(struct kvm *kvm, unsigned long hva)
1033 {
1034         if (!kvm->arch.using_mmu_notifiers)
1035                 return 0;
1036         return kvm_handle_hva(kvm, hva, kvm_test_age_rmapp);
1037 }
1038
1039 void kvm_set_spte_hva_hv(struct kvm *kvm, unsigned long hva, pte_t pte)
1040 {
1041         if (!kvm->arch.using_mmu_notifiers)
1042                 return;
1043         kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
1044 }
1045
1046 static int kvm_test_clear_dirty(struct kvm *kvm, unsigned long *rmapp)
1047 {
1048         struct revmap_entry *rev = kvm->arch.revmap;
1049         unsigned long head, i, j;
1050         unsigned long *hptep;
1051         int ret = 0;
1052
1053  retry:
1054         lock_rmap(rmapp);
1055         if (*rmapp & KVMPPC_RMAP_CHANGED) {
1056                 *rmapp &= ~KVMPPC_RMAP_CHANGED;
1057                 ret = 1;
1058         }
1059         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
1060                 unlock_rmap(rmapp);
1061                 return ret;
1062         }
1063
1064         i = head = *rmapp & KVMPPC_RMAP_INDEX;
1065         do {
1066                 hptep = (unsigned long *) (kvm->arch.hpt_virt + (i << 4));
1067                 j = rev[i].forw;
1068
1069                 if (!(hptep[1] & HPTE_R_C))
1070                         continue;
1071
1072                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
1073                         /* unlock rmap before spinning on the HPTE lock */
1074                         unlock_rmap(rmapp);
1075                         while (hptep[0] & HPTE_V_HVLOCK)
1076                                 cpu_relax();
1077                         goto retry;
1078                 }
1079
1080                 /* Now check and modify the HPTE */
1081                 if ((hptep[0] & HPTE_V_VALID) && (hptep[1] & HPTE_R_C)) {
1082                         /* need to make it temporarily absent to clear C */
1083                         hptep[0] |= HPTE_V_ABSENT;
1084                         kvmppc_invalidate_hpte(kvm, hptep, i);
1085                         hptep[1] &= ~HPTE_R_C;
1086                         eieio();
1087                         hptep[0] = (hptep[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
1088                         if (!(rev[i].guest_rpte & HPTE_R_C)) {
1089                                 rev[i].guest_rpte |= HPTE_R_C;
1090                                 note_hpte_modification(kvm, &rev[i]);
1091                         }
1092                         ret = 1;
1093                 }
1094                 hptep[0] &= ~HPTE_V_HVLOCK;
1095         } while ((i = j) != head);
1096
1097         unlock_rmap(rmapp);
1098         return ret;
1099 }
1100
1101 static void harvest_vpa_dirty(struct kvmppc_vpa *vpa,
1102                               struct kvm_memory_slot *memslot,
1103                               unsigned long *map)
1104 {
1105         unsigned long gfn;
1106
1107         if (!vpa->dirty || !vpa->pinned_addr)
1108                 return;
1109         gfn = vpa->gpa >> PAGE_SHIFT;
1110         if (gfn < memslot->base_gfn ||
1111             gfn >= memslot->base_gfn + memslot->npages)
1112                 return;
1113
1114         vpa->dirty = false;
1115         if (map)
1116                 __set_bit_le(gfn - memslot->base_gfn, map);
1117 }
1118
1119 long kvmppc_hv_get_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot,
1120                              unsigned long *map)
1121 {
1122         unsigned long i;
1123         unsigned long *rmapp;
1124         struct kvm_vcpu *vcpu;
1125
1126         preempt_disable();
1127         rmapp = memslot->arch.rmap;
1128         for (i = 0; i < memslot->npages; ++i) {
1129                 if (kvm_test_clear_dirty(kvm, rmapp) && map)
1130                         __set_bit_le(i, map);
1131                 ++rmapp;
1132         }
1133
1134         /* Harvest dirty bits from VPA and DTL updates */
1135         /* Note: we never modify the SLB shadow buffer areas */
1136         kvm_for_each_vcpu(i, vcpu, kvm) {
1137                 spin_lock(&vcpu->arch.vpa_update_lock);
1138                 harvest_vpa_dirty(&vcpu->arch.vpa, memslot, map);
1139                 harvest_vpa_dirty(&vcpu->arch.dtl, memslot, map);
1140                 spin_unlock(&vcpu->arch.vpa_update_lock);
1141         }
1142         preempt_enable();
1143         return 0;
1144 }
1145
1146 void *kvmppc_pin_guest_page(struct kvm *kvm, unsigned long gpa,
1147                             unsigned long *nb_ret)
1148 {
1149         struct kvm_memory_slot *memslot;
1150         unsigned long gfn = gpa >> PAGE_SHIFT;
1151         struct page *page, *pages[1];
1152         int npages;
1153         unsigned long hva, offset;
1154         unsigned long pa;
1155         unsigned long *physp;
1156         int srcu_idx;
1157
1158         srcu_idx = srcu_read_lock(&kvm->srcu);
1159         memslot = gfn_to_memslot(kvm, gfn);
1160         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
1161                 goto err;
1162         if (!kvm->arch.using_mmu_notifiers) {
1163                 physp = memslot->arch.slot_phys;
1164                 if (!physp)
1165                         goto err;
1166                 physp += gfn - memslot->base_gfn;
1167                 pa = *physp;
1168                 if (!pa) {
1169                         if (kvmppc_get_guest_page(kvm, gfn, memslot,
1170                                                   PAGE_SIZE) < 0)
1171                                 goto err;
1172                         pa = *physp;
1173                 }
1174                 page = pfn_to_page(pa >> PAGE_SHIFT);
1175                 get_page(page);
1176         } else {
1177                 hva = gfn_to_hva_memslot(memslot, gfn);
1178                 npages = get_user_pages_fast(hva, 1, 1, pages);
1179                 if (npages < 1)
1180                         goto err;
1181                 page = pages[0];
1182         }
1183         srcu_read_unlock(&kvm->srcu, srcu_idx);
1184
1185         offset = gpa & (PAGE_SIZE - 1);
1186         if (nb_ret)
1187                 *nb_ret = PAGE_SIZE - offset;
1188         return page_address(page) + offset;
1189
1190  err:
1191         srcu_read_unlock(&kvm->srcu, srcu_idx);
1192         return NULL;
1193 }
1194
1195 void kvmppc_unpin_guest_page(struct kvm *kvm, void *va, unsigned long gpa,
1196                              bool dirty)
1197 {
1198         struct page *page = virt_to_page(va);
1199         struct kvm_memory_slot *memslot;
1200         unsigned long gfn;
1201         unsigned long *rmap;
1202         int srcu_idx;
1203
1204         put_page(page);
1205
1206         if (!dirty || !kvm->arch.using_mmu_notifiers)
1207                 return;
1208
1209         /* We need to mark this page dirty in the rmap chain */
1210         gfn = gpa >> PAGE_SHIFT;
1211         srcu_idx = srcu_read_lock(&kvm->srcu);
1212         memslot = gfn_to_memslot(kvm, gfn);
1213         if (memslot) {
1214                 rmap = &memslot->arch.rmap[gfn - memslot->base_gfn];
1215                 lock_rmap(rmap);
1216                 *rmap |= KVMPPC_RMAP_CHANGED;
1217                 unlock_rmap(rmap);
1218         }
1219         srcu_read_unlock(&kvm->srcu, srcu_idx);
1220 }
1221
1222 /*
1223  * Functions for reading and writing the hash table via reads and
1224  * writes on a file descriptor.
1225  *
1226  * Reads return the guest view of the hash table, which has to be
1227  * pieced together from the real hash table and the guest_rpte
1228  * values in the revmap array.
1229  *
1230  * On writes, each HPTE written is considered in turn, and if it
1231  * is valid, it is written to the HPT as if an H_ENTER with the
1232  * exact flag set was done.  When the invalid count is non-zero
1233  * in the header written to the stream, the kernel will make
1234  * sure that that many HPTEs are invalid, and invalidate them
1235  * if not.
1236  */
1237
1238 struct kvm_htab_ctx {
1239         unsigned long   index;
1240         unsigned long   flags;
1241         struct kvm      *kvm;
1242         int             first_pass;
1243 };
1244
1245 #define HPTE_SIZE       (2 * sizeof(unsigned long))
1246
1247 /*
1248  * Returns 1 if this HPT entry has been modified or has pending
1249  * R/C bit changes.
1250  */
1251 static int hpte_dirty(struct revmap_entry *revp, unsigned long *hptp)
1252 {
1253         unsigned long rcbits_unset;
1254
1255         if (revp->guest_rpte & HPTE_GR_MODIFIED)
1256                 return 1;
1257
1258         /* Also need to consider changes in reference and changed bits */
1259         rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1260         if ((hptp[0] & HPTE_V_VALID) && (hptp[1] & rcbits_unset))
1261                 return 1;
1262
1263         return 0;
1264 }
1265
1266 static long record_hpte(unsigned long flags, unsigned long *hptp,
1267                         unsigned long *hpte, struct revmap_entry *revp,
1268                         int want_valid, int first_pass)
1269 {
1270         unsigned long v, r;
1271         unsigned long rcbits_unset;
1272         int ok = 1;
1273         int valid, dirty;
1274
1275         /* Unmodified entries are uninteresting except on the first pass */
1276         dirty = hpte_dirty(revp, hptp);
1277         if (!first_pass && !dirty)
1278                 return 0;
1279
1280         valid = 0;
1281         if (hptp[0] & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1282                 valid = 1;
1283                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) &&
1284                     !(hptp[0] & HPTE_V_BOLTED))
1285                         valid = 0;
1286         }
1287         if (valid != want_valid)
1288                 return 0;
1289
1290         v = r = 0;
1291         if (valid || dirty) {
1292                 /* lock the HPTE so it's stable and read it */
1293                 preempt_disable();
1294                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
1295                         cpu_relax();
1296                 v = hptp[0];
1297
1298                 /* re-evaluate valid and dirty from synchronized HPTE value */
1299                 valid = !!(v & HPTE_V_VALID);
1300                 dirty = !!(revp->guest_rpte & HPTE_GR_MODIFIED);
1301
1302                 /* Harvest R and C into guest view if necessary */
1303                 rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1304                 if (valid && (rcbits_unset & hptp[1])) {
1305                         revp->guest_rpte |= (hptp[1] & (HPTE_R_R | HPTE_R_C)) |
1306                                 HPTE_GR_MODIFIED;
1307                         dirty = 1;
1308                 }
1309
1310                 if (v & HPTE_V_ABSENT) {
1311                         v &= ~HPTE_V_ABSENT;
1312                         v |= HPTE_V_VALID;
1313                         valid = 1;
1314                 }
1315                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) && !(v & HPTE_V_BOLTED))
1316                         valid = 0;
1317
1318                 r = revp->guest_rpte;
1319                 /* only clear modified if this is the right sort of entry */
1320                 if (valid == want_valid && dirty) {
1321                         r &= ~HPTE_GR_MODIFIED;
1322                         revp->guest_rpte = r;
1323                 }
1324                 asm volatile(PPC_RELEASE_BARRIER "" : : : "memory");
1325                 hptp[0] &= ~HPTE_V_HVLOCK;
1326                 preempt_enable();
1327                 if (!(valid == want_valid && (first_pass || dirty)))
1328                         ok = 0;
1329         }
1330         hpte[0] = v;
1331         hpte[1] = r;
1332         return ok;
1333 }
1334
1335 static ssize_t kvm_htab_read(struct file *file, char __user *buf,
1336                              size_t count, loff_t *ppos)
1337 {
1338         struct kvm_htab_ctx *ctx = file->private_data;
1339         struct kvm *kvm = ctx->kvm;
1340         struct kvm_get_htab_header hdr;
1341         unsigned long *hptp;
1342         struct revmap_entry *revp;
1343         unsigned long i, nb, nw;
1344         unsigned long __user *lbuf;
1345         struct kvm_get_htab_header __user *hptr;
1346         unsigned long flags;
1347         int first_pass;
1348         unsigned long hpte[2];
1349
1350         if (!access_ok(VERIFY_WRITE, buf, count))
1351                 return -EFAULT;
1352
1353         first_pass = ctx->first_pass;
1354         flags = ctx->flags;
1355
1356         i = ctx->index;
1357         hptp = (unsigned long *)(kvm->arch.hpt_virt + (i * HPTE_SIZE));
1358         revp = kvm->arch.revmap + i;
1359         lbuf = (unsigned long __user *)buf;
1360
1361         nb = 0;
1362         while (nb + sizeof(hdr) + HPTE_SIZE < count) {
1363                 /* Initialize header */
1364                 hptr = (struct kvm_get_htab_header __user *)buf;
1365                 hdr.n_valid = 0;
1366                 hdr.n_invalid = 0;
1367                 nw = nb;
1368                 nb += sizeof(hdr);
1369                 lbuf = (unsigned long __user *)(buf + sizeof(hdr));
1370
1371                 /* Skip uninteresting entries, i.e. clean on not-first pass */
1372                 if (!first_pass) {
1373                         while (i < kvm->arch.hpt_npte &&
1374                                !hpte_dirty(revp, hptp)) {
1375                                 ++i;
1376                                 hptp += 2;
1377                                 ++revp;
1378                         }
1379                 }
1380                 hdr.index = i;
1381
1382                 /* Grab a series of valid entries */
1383                 while (i < kvm->arch.hpt_npte &&
1384                        hdr.n_valid < 0xffff &&
1385                        nb + HPTE_SIZE < count &&
1386                        record_hpte(flags, hptp, hpte, revp, 1, first_pass)) {
1387                         /* valid entry, write it out */
1388                         ++hdr.n_valid;
1389                         if (__put_user(hpte[0], lbuf) ||
1390                             __put_user(hpte[1], lbuf + 1))
1391                                 return -EFAULT;
1392                         nb += HPTE_SIZE;
1393                         lbuf += 2;
1394                         ++i;
1395                         hptp += 2;
1396                         ++revp;
1397                 }
1398                 /* Now skip invalid entries while we can */
1399                 while (i < kvm->arch.hpt_npte &&
1400                        hdr.n_invalid < 0xffff &&
1401                        record_hpte(flags, hptp, hpte, revp, 0, first_pass)) {
1402                         /* found an invalid entry */
1403                         ++hdr.n_invalid;
1404                         ++i;
1405                         hptp += 2;
1406                         ++revp;
1407                 }
1408
1409                 if (hdr.n_valid || hdr.n_invalid) {
1410                         /* write back the header */
1411                         if (__copy_to_user(hptr, &hdr, sizeof(hdr)))
1412                                 return -EFAULT;
1413                         nw = nb;
1414                         buf = (char __user *)lbuf;
1415                 } else {
1416                         nb = nw;
1417                 }
1418
1419                 /* Check if we've wrapped around the hash table */
1420                 if (i >= kvm->arch.hpt_npte) {
1421                         i = 0;
1422                         ctx->first_pass = 0;
1423                         break;
1424                 }
1425         }
1426
1427         ctx->index = i;
1428
1429         return nb;
1430 }
1431
1432 static ssize_t kvm_htab_write(struct file *file, const char __user *buf,
1433                               size_t count, loff_t *ppos)
1434 {
1435         struct kvm_htab_ctx *ctx = file->private_data;
1436         struct kvm *kvm = ctx->kvm;
1437         struct kvm_get_htab_header hdr;
1438         unsigned long i, j;
1439         unsigned long v, r;
1440         unsigned long __user *lbuf;
1441         unsigned long *hptp;
1442         unsigned long tmp[2];
1443         ssize_t nb;
1444         long int err, ret;
1445         int rma_setup;
1446
1447         if (!access_ok(VERIFY_READ, buf, count))
1448                 return -EFAULT;
1449
1450         /* lock out vcpus from running while we're doing this */
1451         mutex_lock(&kvm->lock);
1452         rma_setup = kvm->arch.rma_setup_done;
1453         if (rma_setup) {
1454                 kvm->arch.rma_setup_done = 0;   /* temporarily */
1455                 /* order rma_setup_done vs. vcpus_running */
1456                 smp_mb();
1457                 if (atomic_read(&kvm->arch.vcpus_running)) {
1458                         kvm->arch.rma_setup_done = 1;
1459                         mutex_unlock(&kvm->lock);
1460                         return -EBUSY;
1461                 }
1462         }
1463
1464         err = 0;
1465         for (nb = 0; nb + sizeof(hdr) <= count; ) {
1466                 err = -EFAULT;
1467                 if (__copy_from_user(&hdr, buf, sizeof(hdr)))
1468                         break;
1469
1470                 err = 0;
1471                 if (nb + hdr.n_valid * HPTE_SIZE > count)
1472                         break;
1473
1474                 nb += sizeof(hdr);
1475                 buf += sizeof(hdr);
1476
1477                 err = -EINVAL;
1478                 i = hdr.index;
1479                 if (i >= kvm->arch.hpt_npte ||
1480                     i + hdr.n_valid + hdr.n_invalid > kvm->arch.hpt_npte)
1481                         break;
1482
1483                 hptp = (unsigned long *)(kvm->arch.hpt_virt + (i * HPTE_SIZE));
1484                 lbuf = (unsigned long __user *)buf;
1485                 for (j = 0; j < hdr.n_valid; ++j) {
1486                         err = -EFAULT;
1487                         if (__get_user(v, lbuf) || __get_user(r, lbuf + 1))
1488                                 goto out;
1489                         err = -EINVAL;
1490                         if (!(v & HPTE_V_VALID))
1491                                 goto out;
1492                         lbuf += 2;
1493                         nb += HPTE_SIZE;
1494
1495                         if (hptp[0] & (HPTE_V_VALID | HPTE_V_ABSENT))
1496                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1497                         err = -EIO;
1498                         ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, i, v, r,
1499                                                          tmp);
1500                         if (ret != H_SUCCESS) {
1501                                 pr_err("kvm_htab_write ret %ld i=%ld v=%lx "
1502                                        "r=%lx\n", ret, i, v, r);
1503                                 goto out;
1504                         }
1505                         if (!rma_setup && is_vrma_hpte(v)) {
1506                                 unsigned long psize = hpte_page_size(v, r);
1507                                 unsigned long senc = slb_pgsize_encoding(psize);
1508                                 unsigned long lpcr;
1509
1510                                 kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
1511                                         (VRMA_VSID << SLB_VSID_SHIFT_1T);
1512                                 lpcr = senc << (LPCR_VRMASD_SH - 4);
1513                                 kvmppc_update_lpcr(kvm, lpcr, LPCR_VRMASD);
1514                                 rma_setup = 1;
1515                         }
1516                         ++i;
1517                         hptp += 2;
1518                 }
1519
1520                 for (j = 0; j < hdr.n_invalid; ++j) {
1521                         if (hptp[0] & (HPTE_V_VALID | HPTE_V_ABSENT))
1522                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1523                         ++i;
1524                         hptp += 2;
1525                 }
1526                 err = 0;
1527         }
1528
1529  out:
1530         /* Order HPTE updates vs. rma_setup_done */
1531         smp_wmb();
1532         kvm->arch.rma_setup_done = rma_setup;
1533         mutex_unlock(&kvm->lock);
1534
1535         if (err)
1536                 return err;
1537         return nb;
1538 }
1539
1540 static int kvm_htab_release(struct inode *inode, struct file *filp)
1541 {
1542         struct kvm_htab_ctx *ctx = filp->private_data;
1543
1544         filp->private_data = NULL;
1545         if (!(ctx->flags & KVM_GET_HTAB_WRITE))
1546                 atomic_dec(&ctx->kvm->arch.hpte_mod_interest);
1547         kvm_put_kvm(ctx->kvm);
1548         kfree(ctx);
1549         return 0;
1550 }
1551
1552 static const struct file_operations kvm_htab_fops = {
1553         .read           = kvm_htab_read,
1554         .write          = kvm_htab_write,
1555         .llseek         = default_llseek,
1556         .release        = kvm_htab_release,
1557 };
1558
1559 int kvm_vm_ioctl_get_htab_fd(struct kvm *kvm, struct kvm_get_htab_fd *ghf)
1560 {
1561         int ret;
1562         struct kvm_htab_ctx *ctx;
1563         int rwflag;
1564
1565         /* reject flags we don't recognize */
1566         if (ghf->flags & ~(KVM_GET_HTAB_BOLTED_ONLY | KVM_GET_HTAB_WRITE))
1567                 return -EINVAL;
1568         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1569         if (!ctx)
1570                 return -ENOMEM;
1571         kvm_get_kvm(kvm);
1572         ctx->kvm = kvm;
1573         ctx->index = ghf->start_index;
1574         ctx->flags = ghf->flags;
1575         ctx->first_pass = 1;
1576
1577         rwflag = (ghf->flags & KVM_GET_HTAB_WRITE) ? O_WRONLY : O_RDONLY;
1578         ret = anon_inode_getfd("kvm-htab", &kvm_htab_fops, ctx, rwflag | O_CLOEXEC);
1579         if (ret < 0) {
1580                 kvm_put_kvm(kvm);
1581                 return ret;
1582         }
1583
1584         if (rwflag == O_RDONLY) {
1585                 mutex_lock(&kvm->slots_lock);
1586                 atomic_inc(&kvm->arch.hpte_mod_interest);
1587                 /* make sure kvmppc_do_h_enter etc. see the increment */
1588                 synchronize_srcu_expedited(&kvm->srcu);
1589                 mutex_unlock(&kvm->slots_lock);
1590         }
1591
1592         return ret;
1593 }
1594
1595 void kvmppc_mmu_book3s_hv_init(struct kvm_vcpu *vcpu)
1596 {
1597         struct kvmppc_mmu *mmu = &vcpu->arch.mmu;
1598
1599         if (cpu_has_feature(CPU_FTR_ARCH_206))
1600                 vcpu->arch.slb_nr = 32;         /* POWER7 */
1601         else
1602                 vcpu->arch.slb_nr = 64;
1603
1604         mmu->xlate = kvmppc_mmu_book3s_64_hv_xlate;
1605         mmu->reset_msr = kvmppc_mmu_book3s_64_hv_reset_msr;
1606
1607         vcpu->arch.hflags |= BOOK3S_HFLAG_SLB;
1608 }