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 #include <linux/debugfs.h>
  31
  32 #include <asm/tlbflush.h>
  33 #include <asm/kvm_ppc.h>
  34 #include <asm/kvm_book3s.h>
  35 #include <asm/book3s/64/mmu-hash.h>
  36 #include <asm/hvcall.h>
  37 #include <asm/synch.h>
  38 #include <asm/ppc-opcode.h>
  39 #include <asm/cputable.h>
  40 #include <asm/pte-walk.h>
  41
  42 #include "trace_hv.h"
  43
  44 //#define DEBUG_RESIZE_HPT      1
  45
  46 #ifdef DEBUG_RESIZE_HPT
  47 #define resize_hpt_debug(resize, ...)                           \
  48         do {                                                    \
  49                 printk(KERN_DEBUG "RESIZE HPT %p: ", resize);   \
  50                 printk(__VA_ARGS__);                            \
  51         } while (0)
  52 #else
  53 #define resize_hpt_debug(resize, ...)                           \
  54         do { } while (0)
  55 #endif
  56
  57 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
  58                                 long pte_index, unsigned long pteh,
  59                                 unsigned long ptel, unsigned long *pte_idx_ret);
  60
  61 struct kvm_resize_hpt {
  62         /* These fields read-only after init */
  63         struct kvm *kvm;
  64         struct work_struct work;
  65         u32 order;
  66
  67         /* These fields protected by kvm->lock */
  68
  69         /* Possible values and their usage:
  70          *  <0     an error occurred during allocation,
  71          *  -EBUSY allocation is in the progress,
  72          *  0      allocation made successfuly.
  73          */
  74         int error;
  75
  76         /* Private to the work thread, until error != -EBUSY,
  77          * then protected by kvm->lock.
  78          */
  79         struct kvm_hpt_info hpt;
  80 };
  81
  82 static void kvmppc_rmap_reset(struct kvm *kvm);
  83
  84 int kvmppc_allocate_hpt(struct kvm_hpt_info *info, u32 order)
  85 {
  86         unsigned long hpt = 0;
  87         int cma = 0;
  88         struct page *page = NULL;
  89         struct revmap_entry *rev;
  90         unsigned long npte;
  91
  92         if ((order < PPC_MIN_HPT_ORDER) || (order > PPC_MAX_HPT_ORDER))
  93                 return -EINVAL;
  94
  95         page = kvm_alloc_hpt_cma(1ul << (order - PAGE_SHIFT));
  96         if (page) {
  97                 hpt = (unsigned long)pfn_to_kaddr(page_to_pfn(page));
  98                 memset((void *)hpt, 0, (1ul << order));
  99                 cma = 1;
 100         }
 101
 102         if (!hpt)
 103                 hpt = __get_free_pages(GFP_KERNEL|__GFP_ZERO|__GFP_RETRY_MAYFAIL
 104                                        |__GFP_NOWARN, order - PAGE_SHIFT);
 105
 106         if (!hpt)
 107                 return -ENOMEM;
 108
 109         /* HPTEs are 2**4 bytes long */
 110         npte = 1ul << (order - 4);
 111
 112         /* Allocate reverse map array */
 113         rev = vmalloc(sizeof(struct revmap_entry) * npte);
 114         if (!rev) {
 115                 pr_err("kvmppc_allocate_hpt: Couldn't alloc reverse map array\n");
 116                 if (cma)
 117                         kvm_free_hpt_cma(page, 1 << (order - PAGE_SHIFT));
 118                 else
 119                         free_pages(hpt, order - PAGE_SHIFT);
 120                 return -ENOMEM;
 121         }
 122
 123         info->order = order;
 124         info->virt = hpt;
 125         info->cma = cma;
 126         info->rev = rev;
 127
 128         return 0;
 129 }
 130
 131 void kvmppc_set_hpt(struct kvm *kvm, struct kvm_hpt_info *info)
 132 {
 133         atomic64_set(&kvm->arch.mmio_update, 0);
 134         kvm->arch.hpt = *info;
 135         kvm->arch.sdr1 = __pa(info->virt) | (info->order - 18);
 136
 137         pr_debug("KVM guest htab at %lx (order %ld), LPID %x\n",
 138                  info->virt, (long)info->order, kvm->arch.lpid);
 139 }
 140
 141 long kvmppc_alloc_reset_hpt(struct kvm *kvm, int order)
 142 {
 143         long err = -EBUSY;
 144         struct kvm_hpt_info info;
 145
 146         if (kvm_is_radix(kvm))
 147                 return -EINVAL;
 148
 149         mutex_lock(&kvm->lock);
 150         if (kvm->arch.hpte_setup_done) {
 151                 kvm->arch.hpte_setup_done = 0;
 152                 /* order hpte_setup_done vs. vcpus_running */
 153                 smp_mb();
 154                 if (atomic_read(&kvm->arch.vcpus_running)) {
 155                         kvm->arch.hpte_setup_done = 1;
 156                         goto out;
 157                 }
 158         }
 159         if (kvm->arch.hpt.order == order) {
 160                 /* We already have a suitable HPT */
 161
 162                 /* Set the entire HPT to 0, i.e. invalid HPTEs */
 163                 memset((void *)kvm->arch.hpt.virt, 0, 1ul << order);
 164                 /*
 165                  * Reset all the reverse-mapping chains for all memslots
 166                  */
 167                 kvmppc_rmap_reset(kvm);
 168                 err = 0;
 169                 goto out;
 170         }
 171
 172         if (kvm->arch.hpt.virt) {
 173                 kvmppc_free_hpt(&kvm->arch.hpt);
 174                 kvmppc_rmap_reset(kvm);
 175         }
 176
 177         err = kvmppc_allocate_hpt(&info, order);
 178         if (err < 0)
 179                 goto out;
 180         kvmppc_set_hpt(kvm, &info);
 181
 182 out:
 183         if (err == 0)
 184                 /* Ensure that each vcpu will flush its TLB on next entry. */
 185                 cpumask_setall(&kvm->arch.need_tlb_flush);
 186
 187         mutex_unlock(&kvm->lock);
 188         return err;
 189 }
 190
 191 void kvmppc_free_hpt(struct kvm_hpt_info *info)
 192 {
 193         vfree(info->rev);
 194         if (info->cma)
 195                 kvm_free_hpt_cma(virt_to_page(info->virt),
 196                                  1 << (info->order - PAGE_SHIFT));
 197         else if (info->virt)
 198                 free_pages(info->virt, info->order - PAGE_SHIFT);
 199         info->virt = 0;
 200         info->order = 0;
 201 }
 202
 203 /* Bits in first HPTE dword for pagesize 4k, 64k or 16M */
 204 static inline unsigned long hpte0_pgsize_encoding(unsigned long pgsize)
 205 {
 206         return (pgsize > 0x1000) ? HPTE_V_LARGE : 0;
 207 }
 208
 209 /* Bits in second HPTE dword for pagesize 4k, 64k or 16M */
 210 static inline unsigned long hpte1_pgsize_encoding(unsigned long pgsize)
 211 {
 212         return (pgsize == 0x10000) ? 0x1000 : 0;
 213 }
 214
 215 void kvmppc_map_vrma(struct kvm_vcpu *vcpu, struct kvm_memory_slot *memslot,
 216                      unsigned long porder)
 217 {
 218         unsigned long i;
 219         unsigned long npages;
 220         unsigned long hp_v, hp_r;
 221         unsigned long addr, hash;
 222         unsigned long psize;
 223         unsigned long hp0, hp1;
 224         unsigned long idx_ret;
 225         long ret;
 226         struct kvm *kvm = vcpu->kvm;
 227
 228         psize = 1ul << porder;
 229         npages = memslot->npages >> (porder - PAGE_SHIFT);
 230
 231         /* VRMA can't be > 1TB */
 232         if (npages > 1ul << (40 - porder))
 233                 npages = 1ul << (40 - porder);
 234         /* Can't use more than 1 HPTE per HPTEG */
 235         if (npages > kvmppc_hpt_mask(&kvm->arch.hpt) + 1)
 236                 npages = kvmppc_hpt_mask(&kvm->arch.hpt) + 1;
 237
 238         hp0 = HPTE_V_1TB_SEG | (VRMA_VSID << (40 - 16)) |
 239                 HPTE_V_BOLTED | hpte0_pgsize_encoding(psize);
 240         hp1 = hpte1_pgsize_encoding(psize) |
 241                 HPTE_R_R | HPTE_R_C | HPTE_R_M | PP_RWXX;
 242
 243         for (i = 0; i < npages; ++i) {
 244                 addr = i << porder;
 245                 /* can't use hpt_hash since va > 64 bits */
 246                 hash = (i ^ (VRMA_VSID ^ (VRMA_VSID << 25)))
 247                         & kvmppc_hpt_mask(&kvm->arch.hpt);
 248                 /*
 249                  * We assume that the hash table is empty and no
 250                  * vcpus are using it at this stage.  Since we create
 251                  * at most one HPTE per HPTEG, we just assume entry 7
 252                  * is available and use it.
 253                  */
 254                 hash = (hash << 3) + 7;
 255                 hp_v = hp0 | ((addr >> 16) & ~0x7fUL);
 256                 hp_r = hp1 | addr;
 257                 ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, hash, hp_v, hp_r,
 258                                                  &idx_ret);
 259                 if (ret != H_SUCCESS) {
 260                         pr_err("KVM: map_vrma at %lx failed, ret=%ld\n",
 261                                addr, ret);
 262                         break;
 263                 }
 264         }
 265 }
 266
 267 int kvmppc_mmu_hv_init(void)
 268 {
 269         unsigned long host_lpid, rsvd_lpid;
 270
 271         if (!cpu_has_feature(CPU_FTR_HVMODE))
 272                 return -EINVAL;
 273
 274         /* POWER7 has 10-bit LPIDs (12-bit in POWER8) */
 275         host_lpid = mfspr(SPRN_LPID);
 276         rsvd_lpid = LPID_RSVD;
 277
 278         kvmppc_init_lpid(rsvd_lpid + 1);
 279
 280         kvmppc_claim_lpid(host_lpid);
 281         /* rsvd_lpid is reserved for use in partition switching */
 282         kvmppc_claim_lpid(rsvd_lpid);
 283
 284         return 0;
 285 }
 286
 287 static void kvmppc_mmu_book3s_64_hv_reset_msr(struct kvm_vcpu *vcpu)
 288 {
 289         unsigned long msr = vcpu->arch.intr_msr;
 290
 291         /* If transactional, change to suspend mode on IRQ delivery */
 292         if (MSR_TM_TRANSACTIONAL(vcpu->arch.shregs.msr))
 293                 msr |= MSR_TS_S;
 294         else
 295                 msr |= vcpu->arch.shregs.msr & MSR_TS_MASK;
 296         kvmppc_set_msr(vcpu, msr);
 297 }
 298
 299 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
 300                                 long pte_index, unsigned long pteh,
 301                                 unsigned long ptel, unsigned long *pte_idx_ret)
 302 {
 303         long ret;
 304
 305         /* Protect linux PTE lookup from page table destruction */
 306         rcu_read_lock_sched();  /* this disables preemption too */
 307         ret = kvmppc_do_h_enter(kvm, flags, pte_index, pteh, ptel,
 308                                 current->mm->pgd, false, pte_idx_ret);
 309         rcu_read_unlock_sched();
 310         if (ret == H_TOO_HARD) {
 311                 /* this can't happen */
 312                 pr_err("KVM: Oops, kvmppc_h_enter returned too hard!\n");
 313                 ret = H_RESOURCE;       /* or something */
 314         }
 315         return ret;
 316
 317 }
 318
 319 static struct kvmppc_slb *kvmppc_mmu_book3s_hv_find_slbe(struct kvm_vcpu *vcpu,
 320                                                          gva_t eaddr)
 321 {
 322         u64 mask;
 323         int i;
 324
 325         for (i = 0; i < vcpu->arch.slb_nr; i++) {
 326                 if (!(vcpu->arch.slb[i].orige & SLB_ESID_V))
 327                         continue;
 328
 329                 if (vcpu->arch.slb[i].origv & SLB_VSID_B_1T)
 330                         mask = ESID_MASK_1T;
 331                 else
 332                         mask = ESID_MASK;
 333
 334                 if (((vcpu->arch.slb[i].orige ^ eaddr) & mask) == 0)
 335                         return &vcpu->arch.slb[i];
 336         }
 337         return NULL;
 338 }
 339
 340 static unsigned long kvmppc_mmu_get_real_addr(unsigned long v, unsigned long r,
 341                         unsigned long ea)
 342 {
 343         unsigned long ra_mask;
 344
 345         ra_mask = hpte_page_size(v, r) - 1;
 346         return (r & HPTE_R_RPN & ~ra_mask) | (ea & ra_mask);
 347 }
 348
 349 static int kvmppc_mmu_book3s_64_hv_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
 350                         struct kvmppc_pte *gpte, bool data, bool iswrite)
 351 {
 352         struct kvm *kvm = vcpu->kvm;
 353         struct kvmppc_slb *slbe;
 354         unsigned long slb_v;
 355         unsigned long pp, key;
 356         unsigned long v, orig_v, gr;
 357         __be64 *hptep;
 358         long int index;
 359         int virtmode = vcpu->arch.shregs.msr & (data ? MSR_DR : MSR_IR);
 360
 361         /* Get SLB entry */
 362         if (virtmode) {
 363                 slbe = kvmppc_mmu_book3s_hv_find_slbe(vcpu, eaddr);
 364                 if (!slbe)
 365                         return -EINVAL;
 366                 slb_v = slbe->origv;
 367         } else {
 368                 /* real mode access */
 369                 slb_v = vcpu->kvm->arch.vrma_slb_v;
 370         }
 371
 372         preempt_disable();
 373         /* Find the HPTE in the hash table */
 374         index = kvmppc_hv_find_lock_hpte(kvm, eaddr, slb_v,
 375                                          HPTE_V_VALID | HPTE_V_ABSENT);
 376         if (index < 0) {
 377                 preempt_enable();
 378                 return -ENOENT;
 379         }
 380         hptep = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
 381         v = orig_v = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
 382         if (cpu_has_feature(CPU_FTR_ARCH_300))
 383                 v = hpte_new_to_old_v(v, be64_to_cpu(hptep[1]));
 384         gr = kvm->arch.hpt.rev[index].guest_rpte;
 385
 386         unlock_hpte(hptep, orig_v);
 387         preempt_enable();
 388
 389         gpte->eaddr = eaddr;
 390         gpte->vpage = ((v & HPTE_V_AVPN) << 4) | ((eaddr >> 12) & 0xfff);
 391
 392         /* Get PP bits and key for permission check */
 393         pp = gr & (HPTE_R_PP0 | HPTE_R_PP);
 394         key = (vcpu->arch.shregs.msr & MSR_PR) ? SLB_VSID_KP : SLB_VSID_KS;
 395         key &= slb_v;
 396
 397         /* Calculate permissions */
 398         gpte->may_read = hpte_read_permission(pp, key);
 399         gpte->may_write = hpte_write_permission(pp, key);
 400         gpte->may_execute = gpte->may_read && !(gr & (HPTE_R_N | HPTE_R_G));
 401
 402         /* Storage key permission check for POWER7 */
 403         if (data && virtmode) {
 404                 int amrfield = hpte_get_skey_perm(gr, vcpu->arch.amr);
 405                 if (amrfield & 1)
 406                         gpte->may_read = 0;
 407                 if (amrfield & 2)
 408                         gpte->may_write = 0;
 409         }
 410
 411         /* Get the guest physical address */
 412         gpte->raddr = kvmppc_mmu_get_real_addr(v, gr, eaddr);
 413         return 0;
 414 }
 415
 416 /*
 417  * Quick test for whether an instruction is a load or a store.
 418  * If the instruction is a load or a store, then this will indicate
 419  * which it is, at least on server processors.  (Embedded processors
 420  * have some external PID instructions that don't follow the rule
 421  * embodied here.)  If the instruction isn't a load or store, then
 422  * this doesn't return anything useful.
 423  */
 424 static int instruction_is_store(unsigned int instr)
 425 {
 426         unsigned int mask;
 427
 428         mask = 0x10000000;
 429         if ((instr & 0xfc000000) == 0x7c000000)
 430                 mask = 0x100;           /* major opcode 31 */
 431         return (instr & mask) != 0;
 432 }
 433
 434 int kvmppc_hv_emulate_mmio(struct kvm_run *run, struct kvm_vcpu *vcpu,
 435                            unsigned long gpa, gva_t ea, int is_store)
 436 {
 437         u32 last_inst;
 438
 439         /*
 440          * If we fail, we just return to the guest and try executing it again.
 441          */
 442         if (kvmppc_get_last_inst(vcpu, INST_GENERIC, &last_inst) !=
 443                 EMULATE_DONE)
 444                 return RESUME_GUEST;
 445
 446         /*
 447          * WARNING: We do not know for sure whether the instruction we just
 448          * read from memory is the same that caused the fault in the first
 449          * place.  If the instruction we read is neither an load or a store,
 450          * then it can't access memory, so we don't need to worry about
 451          * enforcing access permissions.  So, assuming it is a load or
 452          * store, we just check that its direction (load or store) is
 453          * consistent with the original fault, since that's what we
 454          * checked the access permissions against.  If there is a mismatch
 455          * we just return and retry the instruction.
 456          */
 457
 458         if (instruction_is_store(last_inst) != !!is_store)
 459                 return RESUME_GUEST;
 460
 461         /*
 462          * Emulated accesses are emulated by looking at the hash for
 463          * translation once, then performing the access later. The
 464          * translation could be invalidated in the meantime in which
 465          * point performing the subsequent memory access on the old
 466          * physical address could possibly be a security hole for the
 467          * guest (but not the host).
 468          *
 469          * This is less of an issue for MMIO stores since they aren't
 470          * globally visible. It could be an issue for MMIO loads to
 471          * a certain extent but we'll ignore it for now.
 472          */
 473
 474         vcpu->arch.paddr_accessed = gpa;
 475         vcpu->arch.vaddr_accessed = ea;
 476         return kvmppc_emulate_mmio(run, vcpu);
 477 }
 478
 479 int kvmppc_book3s_hv_page_fault(struct kvm_run *run, struct kvm_vcpu *vcpu,
 480                                 unsigned long ea, unsigned long dsisr)
 481 {
 482         struct kvm *kvm = vcpu->kvm;
 483         unsigned long hpte[3], r;
 484         unsigned long hnow_v, hnow_r;
 485         __be64 *hptep;
 486         unsigned long mmu_seq, psize, pte_size;
 487         unsigned long gpa_base, gfn_base;
 488         unsigned long gpa, gfn, hva, pfn;
 489         struct kvm_memory_slot *memslot;
 490         unsigned long *rmap;
 491         struct revmap_entry *rev;
 492         struct page *page, *pages[1];
 493         long index, ret, npages;
 494         bool is_ci;
 495         unsigned int writing, write_ok;
 496         struct vm_area_struct *vma;
 497         unsigned long rcbits;
 498         long mmio_update;
 499
 500         if (kvm_is_radix(kvm))
 501                 return kvmppc_book3s_radix_page_fault(run, vcpu, ea, dsisr);
 502
 503         /*
 504          * Real-mode code has already searched the HPT and found the
 505          * entry we're interested in.  Lock the entry and check that
 506          * it hasn't changed.  If it has, just return and re-execute the
 507          * instruction.
 508          */
 509         if (ea != vcpu->arch.pgfault_addr)
 510                 return RESUME_GUEST;
 511
 512         if (vcpu->arch.pgfault_cache) {
 513                 mmio_update = atomic64_read(&kvm->arch.mmio_update);
 514                 if (mmio_update == vcpu->arch.pgfault_cache->mmio_update) {
 515                         r = vcpu->arch.pgfault_cache->rpte;
 516                         psize = hpte_page_size(vcpu->arch.pgfault_hpte[0], r);
 517                         gpa_base = r & HPTE_R_RPN & ~(psize - 1);
 518                         gfn_base = gpa_base >> PAGE_SHIFT;
 519                         gpa = gpa_base | (ea & (psize - 1));
 520                         return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
 521                                                 dsisr & DSISR_ISSTORE);
 522                 }
 523         }
 524         index = vcpu->arch.pgfault_index;
 525         hptep = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
 526         rev = &kvm->arch.hpt.rev[index];
 527         preempt_disable();
 528         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 529                 cpu_relax();
 530         hpte[0] = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
 531         hpte[1] = be64_to_cpu(hptep[1]);
 532         hpte[2] = r = rev->guest_rpte;
 533         unlock_hpte(hptep, hpte[0]);
 534         preempt_enable();
 535
 536         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
 537                 hpte[0] = hpte_new_to_old_v(hpte[0], hpte[1]);
 538                 hpte[1] = hpte_new_to_old_r(hpte[1]);
 539         }
 540         if (hpte[0] != vcpu->arch.pgfault_hpte[0] ||
 541             hpte[1] != vcpu->arch.pgfault_hpte[1])
 542                 return RESUME_GUEST;
 543
 544         /* Translate the logical address and get the page */
 545         psize = hpte_page_size(hpte[0], r);
 546         gpa_base = r & HPTE_R_RPN & ~(psize - 1);
 547         gfn_base = gpa_base >> PAGE_SHIFT;
 548         gpa = gpa_base | (ea & (psize - 1));
 549         gfn = gpa >> PAGE_SHIFT;
 550         memslot = gfn_to_memslot(kvm, gfn);
 551
 552         trace_kvm_page_fault_enter(vcpu, hpte, memslot, ea, dsisr);
 553
 554         /* No memslot means it's an emulated MMIO region */
 555         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
 556                 return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
 557                                               dsisr & DSISR_ISSTORE);
 558
 559         /*
 560          * This should never happen, because of the slot_is_aligned()
 561          * check in kvmppc_do_h_enter().
 562          */
 563         if (gfn_base < memslot->base_gfn)
 564                 return -EFAULT;
 565
 566         /* used to check for invalidations in progress */
 567         mmu_seq = kvm->mmu_notifier_seq;
 568         smp_rmb();
 569
 570         ret = -EFAULT;
 571         is_ci = false;
 572         pfn = 0;
 573         page = NULL;
 574         pte_size = PAGE_SIZE;
 575         writing = (dsisr & DSISR_ISSTORE) != 0;
 576         /* If writing != 0, then the HPTE must allow writing, if we get here */
 577         write_ok = writing;
 578         hva = gfn_to_hva_memslot(memslot, gfn);
 579         npages = get_user_pages_fast(hva, 1, writing, pages);
 580         if (npages < 1) {
 581                 /* Check if it's an I/O mapping */
 582                 down_read(&current->mm->mmap_sem);
 583                 vma = find_vma(current->mm, hva);
 584                 if (vma && vma->vm_start <= hva && hva + psize <= vma->vm_end &&
 585                     (vma->vm_flags & VM_PFNMAP)) {
 586                         pfn = vma->vm_pgoff +
 587                                 ((hva - vma->vm_start) >> PAGE_SHIFT);
 588                         pte_size = psize;
 589                         is_ci = pte_ci(__pte((pgprot_val(vma->vm_page_prot))));
 590                         write_ok = vma->vm_flags & VM_WRITE;
 591                 }
 592                 up_read(&current->mm->mmap_sem);
 593                 if (!pfn)
 594                         goto out_put;
 595         } else {
 596                 page = pages[0];
 597                 pfn = page_to_pfn(page);
 598                 if (PageHuge(page)) {
 599                         page = compound_head(page);
 600                         pte_size <<= compound_order(page);
 601                 }
 602                 /* if the guest wants write access, see if that is OK */
 603                 if (!writing && hpte_is_writable(r)) {
 604                         pte_t *ptep, pte;
 605                         unsigned long flags;
 606                         /*
 607                          * We need to protect against page table destruction
 608                          * hugepage split and collapse.
 609                          */
 610                         local_irq_save(flags);
 611                         ptep = find_current_mm_pte(current->mm->pgd,
 612                                                    hva, NULL, NULL);
 613                         if (ptep) {
 614                                 pte = kvmppc_read_update_linux_pte(ptep, 1);
 615                                 if (__pte_write(pte))
 616                                         write_ok = 1;
 617                         }
 618                         local_irq_restore(flags);
 619                 }
 620         }
 621
 622         if (psize > pte_size)
 623                 goto out_put;
 624
 625         /* Check WIMG vs. the actual page we're accessing */
 626         if (!hpte_cache_flags_ok(r, is_ci)) {
 627                 if (is_ci)
 628                         goto out_put;
 629                 /*
 630                  * Allow guest to map emulated device memory as
 631                  * uncacheable, but actually make it cacheable.
 632                  */
 633                 r = (r & ~(HPTE_R_W|HPTE_R_I|HPTE_R_G)) | HPTE_R_M;
 634         }
 635
 636         /*
 637          * Set the HPTE to point to pfn.
 638          * Since the pfn is at PAGE_SIZE granularity, make sure we
 639          * don't mask out lower-order bits if psize < PAGE_SIZE.
 640          */
 641         if (psize < PAGE_SIZE)
 642                 psize = PAGE_SIZE;
 643         r = (r & HPTE_R_KEY_HI) | (r & ~(HPTE_R_PP0 - psize)) |
 644                                         ((pfn << PAGE_SHIFT) & ~(psize - 1));
 645         if (hpte_is_writable(r) && !write_ok)
 646                 r = hpte_make_readonly(r);
 647         ret = RESUME_GUEST;
 648         preempt_disable();
 649         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 650                 cpu_relax();
 651         hnow_v = be64_to_cpu(hptep[0]);
 652         hnow_r = be64_to_cpu(hptep[1]);
 653         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
 654                 hnow_v = hpte_new_to_old_v(hnow_v, hnow_r);
 655                 hnow_r = hpte_new_to_old_r(hnow_r);
 656         }
 657
 658         /*
 659          * If the HPT is being resized, don't update the HPTE,
 660          * instead let the guest retry after the resize operation is complete.
 661          * The synchronization for hpte_setup_done test vs. set is provided
 662          * by the HPTE lock.
 663          */
 664         if (!kvm->arch.hpte_setup_done)
 665                 goto out_unlock;
 666
 667         if ((hnow_v & ~HPTE_V_HVLOCK) != hpte[0] || hnow_r != hpte[1] ||
 668             rev->guest_rpte != hpte[2])
 669                 /* HPTE has been changed under us; let the guest retry */
 670                 goto out_unlock;
 671         hpte[0] = (hpte[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
 672
 673         /* Always put the HPTE in the rmap chain for the page base address */
 674         rmap = &memslot->arch.rmap[gfn_base - memslot->base_gfn];
 675         lock_rmap(rmap);
 676
 677         /* Check if we might have been invalidated; let the guest retry if so */
 678         ret = RESUME_GUEST;
 679         if (mmu_notifier_retry(vcpu->kvm, mmu_seq)) {
 680                 unlock_rmap(rmap);
 681                 goto out_unlock;
 682         }
 683
 684         /* Only set R/C in real HPTE if set in both *rmap and guest_rpte */
 685         rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
 686         r &= rcbits | ~(HPTE_R_R | HPTE_R_C);
 687
 688         if (be64_to_cpu(hptep[0]) & HPTE_V_VALID) {
 689                 /* HPTE was previously valid, so we need to invalidate it */
 690                 unlock_rmap(rmap);
 691                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
 692                 kvmppc_invalidate_hpte(kvm, hptep, index);
 693                 /* don't lose previous R and C bits */
 694                 r |= be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
 695         } else {
 696                 kvmppc_add_revmap_chain(kvm, rev, rmap, index, 0);
 697         }
 698
 699         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
 700                 r = hpte_old_to_new_r(hpte[0], r);
 701                 hpte[0] = hpte_old_to_new_v(hpte[0]);
 702         }
 703         hptep[1] = cpu_to_be64(r);
 704         eieio();
 705         __unlock_hpte(hptep, hpte[0]);
 706         asm volatile("ptesync" : : : "memory");
 707         preempt_enable();
 708         if (page && hpte_is_writable(r))
 709                 SetPageDirty(page);
 710
 711  out_put:
 712         trace_kvm_page_fault_exit(vcpu, hpte, ret);
 713
 714         if (page) {
 715                 /*
 716                  * We drop pages[0] here, not page because page might
 717                  * have been set to the head page of a compound, but
 718                  * we have to drop the reference on the correct tail
 719                  * page to match the get inside gup()
 720                  */
 721                 put_page(pages[0]);
 722         }
 723         return ret;
 724
 725  out_unlock:
 726         __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
 727         preempt_enable();
 728         goto out_put;
 729 }
 730
 731 static void kvmppc_rmap_reset(struct kvm *kvm)
 732 {
 733         struct kvm_memslots *slots;
 734         struct kvm_memory_slot *memslot;
 735         int srcu_idx;
 736
 737         srcu_idx = srcu_read_lock(&kvm->srcu);
 738         slots = kvm_memslots(kvm);
 739         kvm_for_each_memslot(memslot, slots) {
 740                 /*
 741                  * This assumes it is acceptable to lose reference and
 742                  * change bits across a reset.
 743                  */
 744                 memset(memslot->arch.rmap, 0,
 745                        memslot->npages * sizeof(*memslot->arch.rmap));
 746         }
 747         srcu_read_unlock(&kvm->srcu, srcu_idx);
 748 }
 749
 750 typedef int (*hva_handler_fn)(struct kvm *kvm, struct kvm_memory_slot *memslot,
 751                               unsigned long gfn);
 752
 753 static int kvm_handle_hva_range(struct kvm *kvm,
 754                                 unsigned long start,
 755                                 unsigned long end,
 756                                 hva_handler_fn handler)
 757 {
 758         int ret;
 759         int retval = 0;
 760         struct kvm_memslots *slots;
 761         struct kvm_memory_slot *memslot;
 762
 763         slots = kvm_memslots(kvm);
 764         kvm_for_each_memslot(memslot, slots) {
 765                 unsigned long hva_start, hva_end;
 766                 gfn_t gfn, gfn_end;
 767
 768                 hva_start = max(start, memslot->userspace_addr);
 769                 hva_end = min(end, memslot->userspace_addr +
 770                                         (memslot->npages << PAGE_SHIFT));
 771                 if (hva_start >= hva_end)
 772                         continue;
 773                 /*
 774                  * {gfn(page) | page intersects with [hva_start, hva_end)} =
 775                  * {gfn, gfn+1, ..., gfn_end-1}.
 776                  */
 777                 gfn = hva_to_gfn_memslot(hva_start, memslot);
 778                 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
 779
 780                 for (; gfn < gfn_end; ++gfn) {
 781                         ret = handler(kvm, memslot, gfn);
 782                         retval |= ret;
 783                 }
 784         }
 785
 786         return retval;
 787 }
 788
 789 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 790                           hva_handler_fn handler)
 791 {
 792         return kvm_handle_hva_range(kvm, hva, hva + 1, handler);
 793 }
 794
 795 /* Must be called with both HPTE and rmap locked */
 796 static void kvmppc_unmap_hpte(struct kvm *kvm, unsigned long i,
 797                               unsigned long *rmapp, unsigned long gfn)
 798 {
 799         __be64 *hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
 800         struct revmap_entry *rev = kvm->arch.hpt.rev;
 801         unsigned long j, h;
 802         unsigned long ptel, psize, rcbits;
 803
 804         j = rev[i].forw;
 805         if (j == i) {
 806                 /* chain is now empty */
 807                 *rmapp &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
 808         } else {
 809                 /* remove i from chain */
 810                 h = rev[i].back;
 811                 rev[h].forw = j;
 812                 rev[j].back = h;
 813                 rev[i].forw = rev[i].back = i;
 814                 *rmapp = (*rmapp & ~KVMPPC_RMAP_INDEX) | j;
 815         }
 816
 817         /* Now check and modify the HPTE */
 818         ptel = rev[i].guest_rpte;
 819         psize = hpte_page_size(be64_to_cpu(hptep[0]), ptel);
 820         if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
 821             hpte_rpn(ptel, psize) == gfn) {
 822                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
 823                 kvmppc_invalidate_hpte(kvm, hptep, i);
 824                 hptep[1] &= ~cpu_to_be64(HPTE_R_KEY_HI | HPTE_R_KEY_LO);
 825                 /* Harvest R and C */
 826                 rcbits = be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
 827                 *rmapp |= rcbits << KVMPPC_RMAP_RC_SHIFT;
 828                 if (rcbits & HPTE_R_C)
 829                         kvmppc_update_rmap_change(rmapp, psize);
 830                 if (rcbits & ~rev[i].guest_rpte) {
 831                         rev[i].guest_rpte = ptel | rcbits;
 832                         note_hpte_modification(kvm, &rev[i]);
 833                 }
 834         }
 835 }
 836
 837 static int kvm_unmap_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
 838                            unsigned long gfn)
 839 {
 840         unsigned long i;
 841         __be64 *hptep;
 842         unsigned long *rmapp;
 843
 844         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
 845         for (;;) {
 846                 lock_rmap(rmapp);
 847                 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 848                         unlock_rmap(rmapp);
 849                         break;
 850                 }
 851
 852                 /*
 853                  * To avoid an ABBA deadlock with the HPTE lock bit,
 854                  * we can't spin on the HPTE lock while holding the
 855                  * rmap chain lock.
 856                  */
 857                 i = *rmapp & KVMPPC_RMAP_INDEX;
 858                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
 859                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 860                         /* unlock rmap before spinning on the HPTE lock */
 861                         unlock_rmap(rmapp);
 862                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
 863                                 cpu_relax();
 864                         continue;
 865                 }
 866
 867                 kvmppc_unmap_hpte(kvm, i, rmapp, gfn);
 868                 unlock_rmap(rmapp);
 869                 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
 870         }
 871         return 0;
 872 }
 873
 874 int kvm_unmap_hva_hv(struct kvm *kvm, unsigned long hva)
 875 {
 876         hva_handler_fn handler;
 877
 878         handler = kvm_is_radix(kvm) ? kvm_unmap_radix : kvm_unmap_rmapp;
 879         kvm_handle_hva(kvm, hva, handler);
 880         return 0;
 881 }
 882
 883 int kvm_unmap_hva_range_hv(struct kvm *kvm, unsigned long start, unsigned long end)
 884 {
 885         hva_handler_fn handler;
 886
 887         handler = kvm_is_radix(kvm) ? kvm_unmap_radix : kvm_unmap_rmapp;
 888         kvm_handle_hva_range(kvm, start, end, handler);
 889         return 0;
 890 }
 891
 892 void kvmppc_core_flush_memslot_hv(struct kvm *kvm,
 893                                   struct kvm_memory_slot *memslot)
 894 {
 895         unsigned long gfn;
 896         unsigned long n;
 897         unsigned long *rmapp;
 898
 899         gfn = memslot->base_gfn;
 900         rmapp = memslot->arch.rmap;
 901         for (n = memslot->npages; n; --n, ++gfn) {
 902                 if (kvm_is_radix(kvm)) {
 903                         kvm_unmap_radix(kvm, memslot, gfn);
 904                         continue;
 905                 }
 906                 /*
 907                  * Testing the present bit without locking is OK because
 908                  * the memslot has been marked invalid already, and hence
 909                  * no new HPTEs referencing this page can be created,
 910                  * thus the present bit can't go from 0 to 1.
 911                  */
 912                 if (*rmapp & KVMPPC_RMAP_PRESENT)
 913                         kvm_unmap_rmapp(kvm, memslot, gfn);
 914                 ++rmapp;
 915         }
 916 }
 917
 918 static int kvm_age_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
 919                          unsigned long gfn)
 920 {
 921         struct revmap_entry *rev = kvm->arch.hpt.rev;
 922         unsigned long head, i, j;
 923         __be64 *hptep;
 924         int ret = 0;
 925         unsigned long *rmapp;
 926
 927         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
 928  retry:
 929         lock_rmap(rmapp);
 930         if (*rmapp & KVMPPC_RMAP_REFERENCED) {
 931                 *rmapp &= ~KVMPPC_RMAP_REFERENCED;
 932                 ret = 1;
 933         }
 934         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 935                 unlock_rmap(rmapp);
 936                 return ret;
 937         }
 938
 939         i = head = *rmapp & KVMPPC_RMAP_INDEX;
 940         do {
 941                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
 942                 j = rev[i].forw;
 943
 944                 /* If this HPTE isn't referenced, ignore it */
 945                 if (!(be64_to_cpu(hptep[1]) & HPTE_R_R))
 946                         continue;
 947
 948                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 949                         /* unlock rmap before spinning on the HPTE lock */
 950                         unlock_rmap(rmapp);
 951                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
 952                                 cpu_relax();
 953                         goto retry;
 954                 }
 955
 956                 /* Now check and modify the HPTE */
 957                 if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
 958                     (be64_to_cpu(hptep[1]) & HPTE_R_R)) {
 959                         kvmppc_clear_ref_hpte(kvm, hptep, i);
 960                         if (!(rev[i].guest_rpte & HPTE_R_R)) {
 961                                 rev[i].guest_rpte |= HPTE_R_R;
 962                                 note_hpte_modification(kvm, &rev[i]);
 963                         }
 964                         ret = 1;
 965                 }
 966                 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
 967         } while ((i = j) != head);
 968
 969         unlock_rmap(rmapp);
 970         return ret;
 971 }
 972
 973 int kvm_age_hva_hv(struct kvm *kvm, unsigned long start, unsigned long end)
 974 {
 975         hva_handler_fn handler;
 976
 977         handler = kvm_is_radix(kvm) ? kvm_age_radix : kvm_age_rmapp;
 978         return kvm_handle_hva_range(kvm, start, end, handler);
 979 }
 980
 981 static int kvm_test_age_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
 982                               unsigned long gfn)
 983 {
 984         struct revmap_entry *rev = kvm->arch.hpt.rev;
 985         unsigned long head, i, j;
 986         unsigned long *hp;
 987         int ret = 1;
 988         unsigned long *rmapp;
 989
 990         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
 991         if (*rmapp & KVMPPC_RMAP_REFERENCED)
 992                 return 1;
 993
 994         lock_rmap(rmapp);
 995         if (*rmapp & KVMPPC_RMAP_REFERENCED)
 996                 goto out;
 997
 998         if (*rmapp & KVMPPC_RMAP_PRESENT) {
 999                 i = head = *rmapp & KVMPPC_RMAP_INDEX;
1000                 do {
1001                         hp = (unsigned long *)(kvm->arch.hpt.virt + (i << 4));
1002                         j = rev[i].forw;
1003                         if (be64_to_cpu(hp[1]) & HPTE_R_R)
1004                                 goto out;
1005                 } while ((i = j) != head);
1006         }
1007         ret = 0;
1008
1009  out:
1010         unlock_rmap(rmapp);
1011         return ret;
1012 }
1013
1014 int kvm_test_age_hva_hv(struct kvm *kvm, unsigned long hva)
1015 {
1016         hva_handler_fn handler;
1017
1018         handler = kvm_is_radix(kvm) ? kvm_test_age_radix : kvm_test_age_rmapp;
1019         return kvm_handle_hva(kvm, hva, handler);
1020 }
1021
1022 void kvm_set_spte_hva_hv(struct kvm *kvm, unsigned long hva, pte_t pte)
1023 {
1024         hva_handler_fn handler;
1025
1026         handler = kvm_is_radix(kvm) ? kvm_unmap_radix : kvm_unmap_rmapp;
1027         kvm_handle_hva(kvm, hva, handler);
1028 }
1029
1030 static int vcpus_running(struct kvm *kvm)
1031 {
1032         return atomic_read(&kvm->arch.vcpus_running) != 0;
1033 }
1034
1035 /*
1036  * Returns the number of system pages that are dirty.
1037  * This can be more than 1 if we find a huge-page HPTE.
1038  */
1039 static int kvm_test_clear_dirty_npages(struct kvm *kvm, unsigned long *rmapp)
1040 {
1041         struct revmap_entry *rev = kvm->arch.hpt.rev;
1042         unsigned long head, i, j;
1043         unsigned long n;
1044         unsigned long v, r;
1045         __be64 *hptep;
1046         int npages_dirty = 0;
1047
1048  retry:
1049         lock_rmap(rmapp);
1050         if (*rmapp & KVMPPC_RMAP_CHANGED) {
1051                 long change_order = (*rmapp & KVMPPC_RMAP_CHG_ORDER)
1052                         >> KVMPPC_RMAP_CHG_SHIFT;
1053                 *rmapp &= ~(KVMPPC_RMAP_CHANGED | KVMPPC_RMAP_CHG_ORDER);
1054                 npages_dirty = 1;
1055                 if (change_order > PAGE_SHIFT)
1056                         npages_dirty = 1ul << (change_order - PAGE_SHIFT);
1057         }
1058         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
1059                 unlock_rmap(rmapp);
1060                 return npages_dirty;
1061         }
1062
1063         i = head = *rmapp & KVMPPC_RMAP_INDEX;
1064         do {
1065                 unsigned long hptep1;
1066                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
1067                 j = rev[i].forw;
1068
1069                 /*
1070                  * Checking the C (changed) bit here is racy since there
1071                  * is no guarantee about when the hardware writes it back.
1072                  * If the HPTE is not writable then it is stable since the
1073                  * page can't be written to, and we would have done a tlbie
1074                  * (which forces the hardware to complete any writeback)
1075                  * when making the HPTE read-only.
1076                  * If vcpus are running then this call is racy anyway
1077                  * since the page could get dirtied subsequently, so we
1078                  * expect there to be a further call which would pick up
1079                  * any delayed C bit writeback.
1080                  * Otherwise we need to do the tlbie even if C==0 in
1081                  * order to pick up any delayed writeback of C.
1082                  */
1083                 hptep1 = be64_to_cpu(hptep[1]);
1084                 if (!(hptep1 & HPTE_R_C) &&
1085                     (!hpte_is_writable(hptep1) || vcpus_running(kvm)))
1086                         continue;
1087
1088                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
1089                         /* unlock rmap before spinning on the HPTE lock */
1090                         unlock_rmap(rmapp);
1091                         while (hptep[0] & cpu_to_be64(HPTE_V_HVLOCK))
1092                                 cpu_relax();
1093                         goto retry;
1094                 }
1095
1096                 /* Now check and modify the HPTE */
1097                 if (!(hptep[0] & cpu_to_be64(HPTE_V_VALID))) {
1098                         __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
1099                         continue;
1100                 }
1101
1102                 /* need to make it temporarily absent so C is stable */
1103                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
1104                 kvmppc_invalidate_hpte(kvm, hptep, i);
1105                 v = be64_to_cpu(hptep[0]);
1106                 r = be64_to_cpu(hptep[1]);
1107                 if (r & HPTE_R_C) {
1108                         hptep[1] = cpu_to_be64(r & ~HPTE_R_C);
1109                         if (!(rev[i].guest_rpte & HPTE_R_C)) {
1110                                 rev[i].guest_rpte |= HPTE_R_C;
1111                                 note_hpte_modification(kvm, &rev[i]);
1112                         }
1113                         n = hpte_page_size(v, r);
1114                         n = (n + PAGE_SIZE - 1) >> PAGE_SHIFT;
1115                         if (n > npages_dirty)
1116                                 npages_dirty = n;
1117                         eieio();
1118                 }
1119                 v &= ~HPTE_V_ABSENT;
1120                 v |= HPTE_V_VALID;
1121                 __unlock_hpte(hptep, v);
1122         } while ((i = j) != head);
1123
1124         unlock_rmap(rmapp);
1125         return npages_dirty;
1126 }
1127
1128 void kvmppc_harvest_vpa_dirty(struct kvmppc_vpa *vpa,
1129                               struct kvm_memory_slot *memslot,
1130                               unsigned long *map)
1131 {
1132         unsigned long gfn;
1133
1134         if (!vpa->dirty || !vpa->pinned_addr)
1135                 return;
1136         gfn = vpa->gpa >> PAGE_SHIFT;
1137         if (gfn < memslot->base_gfn ||
1138             gfn >= memslot->base_gfn + memslot->npages)
1139                 return;
1140
1141         vpa->dirty = false;
1142         if (map)
1143                 __set_bit_le(gfn - memslot->base_gfn, map);
1144 }
1145
1146 long kvmppc_hv_get_dirty_log_hpt(struct kvm *kvm,
1147                         struct kvm_memory_slot *memslot, unsigned long *map)
1148 {
1149         unsigned long i, j;
1150         unsigned long *rmapp;
1151
1152         preempt_disable();
1153         rmapp = memslot->arch.rmap;
1154         for (i = 0; i < memslot->npages; ++i) {
1155                 int npages = kvm_test_clear_dirty_npages(kvm, rmapp);
1156                 /*
1157                  * Note that if npages > 0 then i must be a multiple of npages,
1158                  * since we always put huge-page HPTEs in the rmap chain
1159                  * corresponding to their page base address.
1160                  */
1161                 if (npages && map)
1162                         for (j = i; npages; ++j, --npages)
1163                                 __set_bit_le(j, map);
1164                 ++rmapp;
1165         }
1166         preempt_enable();
1167         return 0;
1168 }
1169
1170 void *kvmppc_pin_guest_page(struct kvm *kvm, unsigned long gpa,
1171                             unsigned long *nb_ret)
1172 {
1173         struct kvm_memory_slot *memslot;
1174         unsigned long gfn = gpa >> PAGE_SHIFT;
1175         struct page *page, *pages[1];
1176         int npages;
1177         unsigned long hva, offset;
1178         int srcu_idx;
1179
1180         srcu_idx = srcu_read_lock(&kvm->srcu);
1181         memslot = gfn_to_memslot(kvm, gfn);
1182         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
1183                 goto err;
1184         hva = gfn_to_hva_memslot(memslot, gfn);
1185         npages = get_user_pages_fast(hva, 1, 1, pages);
1186         if (npages < 1)
1187                 goto err;
1188         page = pages[0];
1189         srcu_read_unlock(&kvm->srcu, srcu_idx);
1190
1191         offset = gpa & (PAGE_SIZE - 1);
1192         if (nb_ret)
1193                 *nb_ret = PAGE_SIZE - offset;
1194         return page_address(page) + offset;
1195
1196  err:
1197         srcu_read_unlock(&kvm->srcu, srcu_idx);
1198         return NULL;
1199 }
1200
1201 void kvmppc_unpin_guest_page(struct kvm *kvm, void *va, unsigned long gpa,
1202                              bool dirty)
1203 {
1204         struct page *page = virt_to_page(va);
1205         struct kvm_memory_slot *memslot;
1206         unsigned long gfn;
1207         unsigned long *rmap;
1208         int srcu_idx;
1209
1210         put_page(page);
1211
1212         if (!dirty)
1213                 return;
1214
1215         /* We need to mark this page dirty in the rmap chain */
1216         gfn = gpa >> PAGE_SHIFT;
1217         srcu_idx = srcu_read_lock(&kvm->srcu);
1218         memslot = gfn_to_memslot(kvm, gfn);
1219         if (memslot) {
1220                 if (!kvm_is_radix(kvm)) {
1221                         rmap = &memslot->arch.rmap[gfn - memslot->base_gfn];
1222                         lock_rmap(rmap);
1223                         *rmap |= KVMPPC_RMAP_CHANGED;
1224                         unlock_rmap(rmap);
1225                 } else if (memslot->dirty_bitmap) {
1226                         mark_page_dirty(kvm, gfn);
1227                 }
1228         }
1229         srcu_read_unlock(&kvm->srcu, srcu_idx);
1230 }
1231
1232 /*
1233  * HPT resizing
1234  */
1235 static int resize_hpt_allocate(struct kvm_resize_hpt *resize)
1236 {
1237         int rc;
1238
1239         rc = kvmppc_allocate_hpt(&resize->hpt, resize->order);
1240         if (rc < 0)
1241                 return rc;
1242
1243         resize_hpt_debug(resize, "resize_hpt_allocate(): HPT @ 0x%lx\n",
1244                          resize->hpt.virt);
1245
1246         return 0;
1247 }
1248
1249 static unsigned long resize_hpt_rehash_hpte(struct kvm_resize_hpt *resize,
1250                                             unsigned long idx)
1251 {
1252         struct kvm *kvm = resize->kvm;
1253         struct kvm_hpt_info *old = &kvm->arch.hpt;
1254         struct kvm_hpt_info *new = &resize->hpt;
1255         unsigned long old_hash_mask = (1ULL << (old->order - 7)) - 1;
1256         unsigned long new_hash_mask = (1ULL << (new->order - 7)) - 1;
1257         __be64 *hptep, *new_hptep;
1258         unsigned long vpte, rpte, guest_rpte;
1259         int ret;
1260         struct revmap_entry *rev;
1261         unsigned long apsize, psize, avpn, pteg, hash;
1262         unsigned long new_idx, new_pteg, replace_vpte;
1263
1264         hptep = (__be64 *)(old->virt + (idx << 4));
1265
1266         /* Guest is stopped, so new HPTEs can't be added or faulted
1267          * in, only unmapped or altered by host actions.  So, it's
1268          * safe to check this before we take the HPTE lock */
1269         vpte = be64_to_cpu(hptep[0]);
1270         if (!(vpte & HPTE_V_VALID) && !(vpte & HPTE_V_ABSENT))
1271                 return 0; /* nothing to do */
1272
1273         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
1274                 cpu_relax();
1275
1276         vpte = be64_to_cpu(hptep[0]);
1277
1278         ret = 0;
1279         if (!(vpte & HPTE_V_VALID) && !(vpte & HPTE_V_ABSENT))
1280                 /* Nothing to do */
1281                 goto out;
1282
1283         /* Unmap */
1284         rev = &old->rev[idx];
1285         guest_rpte = rev->guest_rpte;
1286
1287         ret = -EIO;
1288         apsize = hpte_page_size(vpte, guest_rpte);
1289         if (!apsize)
1290                 goto out;
1291
1292         if (vpte & HPTE_V_VALID) {
1293                 unsigned long gfn = hpte_rpn(guest_rpte, apsize);
1294                 int srcu_idx = srcu_read_lock(&kvm->srcu);
1295                 struct kvm_memory_slot *memslot =
1296                         __gfn_to_memslot(kvm_memslots(kvm), gfn);
1297
1298                 if (memslot) {
1299                         unsigned long *rmapp;
1300                         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1301
1302                         lock_rmap(rmapp);
1303                         kvmppc_unmap_hpte(kvm, idx, rmapp, gfn);
1304                         unlock_rmap(rmapp);
1305                 }
1306
1307                 srcu_read_unlock(&kvm->srcu, srcu_idx);
1308         }
1309
1310         /* Reload PTE after unmap */
1311         vpte = be64_to_cpu(hptep[0]);
1312
1313         BUG_ON(vpte & HPTE_V_VALID);
1314         BUG_ON(!(vpte & HPTE_V_ABSENT));
1315
1316         ret = 0;
1317         if (!(vpte & HPTE_V_BOLTED))
1318                 goto out;
1319
1320         rpte = be64_to_cpu(hptep[1]);
1321         psize = hpte_base_page_size(vpte, rpte);
1322         avpn = HPTE_V_AVPN_VAL(vpte) & ~((psize - 1) >> 23);
1323         pteg = idx / HPTES_PER_GROUP;
1324         if (vpte & HPTE_V_SECONDARY)
1325                 pteg = ~pteg;
1326
1327         if (!(vpte & HPTE_V_1TB_SEG)) {
1328                 unsigned long offset, vsid;
1329
1330                 /* We only have 28 - 23 bits of offset in avpn */
1331                 offset = (avpn & 0x1f) << 23;
1332                 vsid = avpn >> 5;
1333                 /* We can find more bits from the pteg value */
1334                 if (psize < (1ULL << 23))
1335                         offset |= ((vsid ^ pteg) & old_hash_mask) * psize;
1336
1337                 hash = vsid ^ (offset / psize);
1338         } else {
1339                 unsigned long offset, vsid;
1340
1341                 /* We only have 40 - 23 bits of seg_off in avpn */
1342                 offset = (avpn & 0x1ffff) << 23;
1343                 vsid = avpn >> 17;
1344                 if (psize < (1ULL << 23))
1345                         offset |= ((vsid ^ (vsid << 25) ^ pteg) & old_hash_mask) * psize;
1346
1347                 hash = vsid ^ (vsid << 25) ^ (offset / psize);
1348         }
1349
1350         new_pteg = hash & new_hash_mask;
1351         if (vpte & HPTE_V_SECONDARY)
1352                 new_pteg = ~hash & new_hash_mask;
1353
1354         new_idx = new_pteg * HPTES_PER_GROUP + (idx % HPTES_PER_GROUP);
1355         new_hptep = (__be64 *)(new->virt + (new_idx << 4));
1356
1357         replace_vpte = be64_to_cpu(new_hptep[0]);
1358
1359         if (replace_vpte & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1360                 BUG_ON(new->order >= old->order);
1361
1362                 if (replace_vpte & HPTE_V_BOLTED) {
1363                         if (vpte & HPTE_V_BOLTED)
1364                                 /* Bolted collision, nothing we can do */
1365                                 ret = -ENOSPC;
1366                         /* Discard the new HPTE */
1367                         goto out;
1368                 }
1369
1370                 /* Discard the previous HPTE */
1371         }
1372
1373         new_hptep[1] = cpu_to_be64(rpte);
1374         new->rev[new_idx].guest_rpte = guest_rpte;
1375         /* No need for a barrier, since new HPT isn't active */
1376         new_hptep[0] = cpu_to_be64(vpte);
1377         unlock_hpte(new_hptep, vpte);
1378
1379 out:
1380         unlock_hpte(hptep, vpte);
1381         return ret;
1382 }
1383
1384 static int resize_hpt_rehash(struct kvm_resize_hpt *resize)
1385 {
1386         struct kvm *kvm = resize->kvm;
1387         unsigned  long i;
1388         int rc;
1389
1390         /*
1391          * resize_hpt_rehash_hpte() doesn't handle the new-format HPTEs
1392          * that POWER9 uses, and could well hit a BUG_ON on POWER9.
1393          */
1394         if (cpu_has_feature(CPU_FTR_ARCH_300))
1395                 return -EIO;
1396         for (i = 0; i < kvmppc_hpt_npte(&kvm->arch.hpt); i++) {
1397                 rc = resize_hpt_rehash_hpte(resize, i);
1398                 if (rc != 0)
1399                         return rc;
1400         }
1401
1402         return 0;
1403 }
1404
1405 static void resize_hpt_pivot(struct kvm_resize_hpt *resize)
1406 {
1407         struct kvm *kvm = resize->kvm;
1408         struct kvm_hpt_info hpt_tmp;
1409
1410         /* Exchange the pending tables in the resize structure with
1411          * the active tables */
1412
1413         resize_hpt_debug(resize, "resize_hpt_pivot()\n");
1414
1415         spin_lock(&kvm->mmu_lock);
1416         asm volatile("ptesync" : : : "memory");
1417
1418         hpt_tmp = kvm->arch.hpt;
1419         kvmppc_set_hpt(kvm, &resize->hpt);
1420         resize->hpt = hpt_tmp;
1421
1422         spin_unlock(&kvm->mmu_lock);
1423
1424         synchronize_srcu_expedited(&kvm->srcu);
1425
1426         resize_hpt_debug(resize, "resize_hpt_pivot() done\n");
1427 }
1428
1429 static void resize_hpt_release(struct kvm *kvm, struct kvm_resize_hpt *resize)
1430 {
1431         if (WARN_ON(!mutex_is_locked(&kvm->lock)))
1432                 return;
1433
1434         if (!resize)
1435                 return;
1436
1437         if (resize->error != -EBUSY) {
1438                 if (resize->hpt.virt)
1439                         kvmppc_free_hpt(&resize->hpt);
1440                 kfree(resize);
1441         }
1442
1443         if (kvm->arch.resize_hpt == resize)
1444                 kvm->arch.resize_hpt = NULL;
1445 }
1446
1447 static void resize_hpt_prepare_work(struct work_struct *work)
1448 {
1449         struct kvm_resize_hpt *resize = container_of(work,
1450                                                      struct kvm_resize_hpt,
1451                                                      work);
1452         struct kvm *kvm = resize->kvm;
1453         int err = 0;
1454
1455         if (WARN_ON(resize->error != -EBUSY))
1456                 return;
1457
1458         mutex_lock(&kvm->lock);
1459
1460         /* Request is still current? */
1461         if (kvm->arch.resize_hpt == resize) {
1462                 /* We may request large allocations here:
1463                  * do not sleep with kvm->lock held for a while.
1464                  */
1465                 mutex_unlock(&kvm->lock);
1466
1467                 resize_hpt_debug(resize, "resize_hpt_prepare_work(): order = %d\n",
1468                                  resize->order);
1469
1470                 err = resize_hpt_allocate(resize);
1471
1472                 /* We have strict assumption about -EBUSY
1473                  * when preparing for HPT resize.
1474                  */
1475                 if (WARN_ON(err == -EBUSY))
1476                         err = -EINPROGRESS;
1477
1478                 mutex_lock(&kvm->lock);
1479                 /* It is possible that kvm->arch.resize_hpt != resize
1480                  * after we grab kvm->lock again.
1481                  */
1482         }
1483
1484         resize->error = err;
1485
1486         if (kvm->arch.resize_hpt != resize)
1487                 resize_hpt_release(kvm, resize);
1488
1489         mutex_unlock(&kvm->lock);
1490 }
1491
1492 long kvm_vm_ioctl_resize_hpt_prepare(struct kvm *kvm,
1493                                      struct kvm_ppc_resize_hpt *rhpt)
1494 {
1495         unsigned long flags = rhpt->flags;
1496         unsigned long shift = rhpt->shift;
1497         struct kvm_resize_hpt *resize;
1498         int ret;
1499
1500         if (flags != 0)
1501                 return -EINVAL;
1502
1503         if (shift && ((shift < 18) || (shift > 46)))
1504                 return -EINVAL;
1505
1506         mutex_lock(&kvm->lock);
1507
1508         resize = kvm->arch.resize_hpt;
1509
1510         if (resize) {
1511                 if (resize->order == shift) {
1512                         /* Suitable resize in progress? */
1513                         ret = resize->error;
1514                         if (ret == -EBUSY)
1515                                 ret = 100; /* estimated time in ms */
1516                         else if (ret)
1517                                 resize_hpt_release(kvm, resize);
1518
1519                         goto out;
1520                 }
1521
1522                 /* not suitable, cancel it */
1523                 resize_hpt_release(kvm, resize);
1524         }
1525
1526         ret = 0;
1527         if (!shift)
1528                 goto out; /* nothing to do */
1529
1530         /* start new resize */
1531
1532         resize = kzalloc(sizeof(*resize), GFP_KERNEL);
1533         if (!resize) {
1534                 ret = -ENOMEM;
1535                 goto out;
1536         }
1537
1538         resize->error = -EBUSY;
1539         resize->order = shift;
1540         resize->kvm = kvm;
1541         INIT_WORK(&resize->work, resize_hpt_prepare_work);
1542         kvm->arch.resize_hpt = resize;
1543
1544         schedule_work(&resize->work);
1545
1546         ret = 100; /* estimated time in ms */
1547
1548 out:
1549         mutex_unlock(&kvm->lock);
1550         return ret;
1551 }
1552
1553 static void resize_hpt_boot_vcpu(void *opaque)
1554 {
1555         /* Nothing to do, just force a KVM exit */
1556 }
1557
1558 long kvm_vm_ioctl_resize_hpt_commit(struct kvm *kvm,
1559                                     struct kvm_ppc_resize_hpt *rhpt)
1560 {
1561         unsigned long flags = rhpt->flags;
1562         unsigned long shift = rhpt->shift;
1563         struct kvm_resize_hpt *resize;
1564         long ret;
1565
1566         if (flags != 0)
1567                 return -EINVAL;
1568
1569         if (shift && ((shift < 18) || (shift > 46)))
1570                 return -EINVAL;
1571
1572         mutex_lock(&kvm->lock);
1573
1574         resize = kvm->arch.resize_hpt;
1575
1576         /* This shouldn't be possible */
1577         ret = -EIO;
1578         if (WARN_ON(!kvm->arch.hpte_setup_done))
1579                 goto out_no_hpt;
1580
1581         /* Stop VCPUs from running while we mess with the HPT */
1582         kvm->arch.hpte_setup_done = 0;
1583         smp_mb();
1584
1585         /* Boot all CPUs out of the guest so they re-read
1586          * hpte_setup_done */
1587         on_each_cpu(resize_hpt_boot_vcpu, NULL, 1);
1588
1589         ret = -ENXIO;
1590         if (!resize || (resize->order != shift))
1591                 goto out;
1592
1593         ret = resize->error;
1594         if (ret)
1595                 goto out;
1596
1597         ret = resize_hpt_rehash(resize);
1598         if (ret)
1599                 goto out;
1600
1601         resize_hpt_pivot(resize);
1602
1603 out:
1604         /* Let VCPUs run again */
1605         kvm->arch.hpte_setup_done = 1;
1606         smp_mb();
1607 out_no_hpt:
1608         resize_hpt_release(kvm, resize);
1609         mutex_unlock(&kvm->lock);
1610         return ret;
1611 }
1612
1613 /*
1614  * Functions for reading and writing the hash table via reads and
1615  * writes on a file descriptor.
1616  *
1617  * Reads return the guest view of the hash table, which has to be
1618  * pieced together from the real hash table and the guest_rpte
1619  * values in the revmap array.
1620  *
1621  * On writes, each HPTE written is considered in turn, and if it
1622  * is valid, it is written to the HPT as if an H_ENTER with the
1623  * exact flag set was done.  When the invalid count is non-zero
1624  * in the header written to the stream, the kernel will make
1625  * sure that that many HPTEs are invalid, and invalidate them
1626  * if not.
1627  */
1628
1629 struct kvm_htab_ctx {
1630         unsigned long   index;
1631         unsigned long   flags;
1632         struct kvm      *kvm;
1633         int             first_pass;
1634 };
1635
1636 #define HPTE_SIZE       (2 * sizeof(unsigned long))
1637
1638 /*
1639  * Returns 1 if this HPT entry has been modified or has pending
1640  * R/C bit changes.
1641  */
1642 static int hpte_dirty(struct revmap_entry *revp, __be64 *hptp)
1643 {
1644         unsigned long rcbits_unset;
1645
1646         if (revp->guest_rpte & HPTE_GR_MODIFIED)
1647                 return 1;
1648
1649         /* Also need to consider changes in reference and changed bits */
1650         rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1651         if ((be64_to_cpu(hptp[0]) & HPTE_V_VALID) &&
1652             (be64_to_cpu(hptp[1]) & rcbits_unset))
1653                 return 1;
1654
1655         return 0;
1656 }
1657
1658 static long record_hpte(unsigned long flags, __be64 *hptp,
1659                         unsigned long *hpte, struct revmap_entry *revp,
1660                         int want_valid, int first_pass)
1661 {
1662         unsigned long v, r, hr;
1663         unsigned long rcbits_unset;
1664         int ok = 1;
1665         int valid, dirty;
1666
1667         /* Unmodified entries are uninteresting except on the first pass */
1668         dirty = hpte_dirty(revp, hptp);
1669         if (!first_pass && !dirty)
1670                 return 0;
1671
1672         valid = 0;
1673         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1674                 valid = 1;
1675                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) &&
1676                     !(be64_to_cpu(hptp[0]) & HPTE_V_BOLTED))
1677                         valid = 0;
1678         }
1679         if (valid != want_valid)
1680                 return 0;
1681
1682         v = r = 0;
1683         if (valid || dirty) {
1684                 /* lock the HPTE so it's stable and read it */
1685                 preempt_disable();
1686                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
1687                         cpu_relax();
1688                 v = be64_to_cpu(hptp[0]);
1689                 hr = be64_to_cpu(hptp[1]);
1690                 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1691                         v = hpte_new_to_old_v(v, hr);
1692                         hr = hpte_new_to_old_r(hr);
1693                 }
1694
1695                 /* re-evaluate valid and dirty from synchronized HPTE value */
1696                 valid = !!(v & HPTE_V_VALID);
1697                 dirty = !!(revp->guest_rpte & HPTE_GR_MODIFIED);
1698
1699                 /* Harvest R and C into guest view if necessary */
1700                 rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1701                 if (valid && (rcbits_unset & hr)) {
1702                         revp->guest_rpte |= (hr &
1703                                 (HPTE_R_R | HPTE_R_C)) | HPTE_GR_MODIFIED;
1704                         dirty = 1;
1705                 }
1706
1707                 if (v & HPTE_V_ABSENT) {
1708                         v &= ~HPTE_V_ABSENT;
1709                         v |= HPTE_V_VALID;
1710                         valid = 1;
1711                 }
1712                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) && !(v & HPTE_V_BOLTED))
1713                         valid = 0;
1714
1715                 r = revp->guest_rpte;
1716                 /* only clear modified if this is the right sort of entry */
1717                 if (valid == want_valid && dirty) {
1718                         r &= ~HPTE_GR_MODIFIED;
1719                         revp->guest_rpte = r;
1720                 }
1721                 unlock_hpte(hptp, be64_to_cpu(hptp[0]));
1722                 preempt_enable();
1723                 if (!(valid == want_valid && (first_pass || dirty)))
1724                         ok = 0;
1725         }
1726         hpte[0] = cpu_to_be64(v);
1727         hpte[1] = cpu_to_be64(r);
1728         return ok;
1729 }
1730
1731 static ssize_t kvm_htab_read(struct file *file, char __user *buf,
1732                              size_t count, loff_t *ppos)
1733 {
1734         struct kvm_htab_ctx *ctx = file->private_data;
1735         struct kvm *kvm = ctx->kvm;
1736         struct kvm_get_htab_header hdr;
1737         __be64 *hptp;
1738         struct revmap_entry *revp;
1739         unsigned long i, nb, nw;
1740         unsigned long __user *lbuf;
1741         struct kvm_get_htab_header __user *hptr;
1742         unsigned long flags;
1743         int first_pass;
1744         unsigned long hpte[2];
1745
1746         if (!access_ok(VERIFY_WRITE, buf, count))
1747                 return -EFAULT;
1748
1749         first_pass = ctx->first_pass;
1750         flags = ctx->flags;
1751
1752         i = ctx->index;
1753         hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
1754         revp = kvm->arch.hpt.rev + i;
1755         lbuf = (unsigned long __user *)buf;
1756
1757         nb = 0;
1758         while (nb + sizeof(hdr) + HPTE_SIZE < count) {
1759                 /* Initialize header */
1760                 hptr = (struct kvm_get_htab_header __user *)buf;
1761                 hdr.n_valid = 0;
1762                 hdr.n_invalid = 0;
1763                 nw = nb;
1764                 nb += sizeof(hdr);
1765                 lbuf = (unsigned long __user *)(buf + sizeof(hdr));
1766
1767                 /* Skip uninteresting entries, i.e. clean on not-first pass */
1768                 if (!first_pass) {
1769                         while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1770                                !hpte_dirty(revp, hptp)) {
1771                                 ++i;
1772                                 hptp += 2;
1773                                 ++revp;
1774                         }
1775                 }
1776                 hdr.index = i;
1777
1778                 /* Grab a series of valid entries */
1779                 while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1780                        hdr.n_valid < 0xffff &&
1781                        nb + HPTE_SIZE < count &&
1782                        record_hpte(flags, hptp, hpte, revp, 1, first_pass)) {
1783                         /* valid entry, write it out */
1784                         ++hdr.n_valid;
1785                         if (__put_user(hpte[0], lbuf) ||
1786                             __put_user(hpte[1], lbuf + 1))
1787                                 return -EFAULT;
1788                         nb += HPTE_SIZE;
1789                         lbuf += 2;
1790                         ++i;
1791                         hptp += 2;
1792                         ++revp;
1793                 }
1794                 /* Now skip invalid entries while we can */
1795                 while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1796                        hdr.n_invalid < 0xffff &&
1797                        record_hpte(flags, hptp, hpte, revp, 0, first_pass)) {
1798                         /* found an invalid entry */
1799                         ++hdr.n_invalid;
1800                         ++i;
1801                         hptp += 2;
1802                         ++revp;
1803                 }
1804
1805                 if (hdr.n_valid || hdr.n_invalid) {
1806                         /* write back the header */
1807                         if (__copy_to_user(hptr, &hdr, sizeof(hdr)))
1808                                 return -EFAULT;
1809                         nw = nb;
1810                         buf = (char __user *)lbuf;
1811                 } else {
1812                         nb = nw;
1813                 }
1814
1815                 /* Check if we've wrapped around the hash table */
1816                 if (i >= kvmppc_hpt_npte(&kvm->arch.hpt)) {
1817                         i = 0;
1818                         ctx->first_pass = 0;
1819                         break;
1820                 }
1821         }
1822
1823         ctx->index = i;
1824
1825         return nb;
1826 }
1827
1828 static ssize_t kvm_htab_write(struct file *file, const char __user *buf,
1829                               size_t count, loff_t *ppos)
1830 {
1831         struct kvm_htab_ctx *ctx = file->private_data;
1832         struct kvm *kvm = ctx->kvm;
1833         struct kvm_get_htab_header hdr;
1834         unsigned long i, j;
1835         unsigned long v, r;
1836         unsigned long __user *lbuf;
1837         __be64 *hptp;
1838         unsigned long tmp[2];
1839         ssize_t nb;
1840         long int err, ret;
1841         int hpte_setup;
1842
1843         if (!access_ok(VERIFY_READ, buf, count))
1844                 return -EFAULT;
1845
1846         /* lock out vcpus from running while we're doing this */
1847         mutex_lock(&kvm->lock);
1848         hpte_setup = kvm->arch.hpte_setup_done;
1849         if (hpte_setup) {
1850                 kvm->arch.hpte_setup_done = 0;  /* temporarily */
1851                 /* order hpte_setup_done vs. vcpus_running */
1852                 smp_mb();
1853                 if (atomic_read(&kvm->arch.vcpus_running)) {
1854                         kvm->arch.hpte_setup_done = 1;
1855                         mutex_unlock(&kvm->lock);
1856                         return -EBUSY;
1857                 }
1858         }
1859
1860         err = 0;
1861         for (nb = 0; nb + sizeof(hdr) <= count; ) {
1862                 err = -EFAULT;
1863                 if (__copy_from_user(&hdr, buf, sizeof(hdr)))
1864                         break;
1865
1866                 err = 0;
1867                 if (nb + hdr.n_valid * HPTE_SIZE > count)
1868                         break;
1869
1870                 nb += sizeof(hdr);
1871                 buf += sizeof(hdr);
1872
1873                 err = -EINVAL;
1874                 i = hdr.index;
1875                 if (i >= kvmppc_hpt_npte(&kvm->arch.hpt) ||
1876                     i + hdr.n_valid + hdr.n_invalid > kvmppc_hpt_npte(&kvm->arch.hpt))
1877                         break;
1878
1879                 hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
1880                 lbuf = (unsigned long __user *)buf;
1881                 for (j = 0; j < hdr.n_valid; ++j) {
1882                         __be64 hpte_v;
1883                         __be64 hpte_r;
1884
1885                         err = -EFAULT;
1886                         if (__get_user(hpte_v, lbuf) ||
1887                             __get_user(hpte_r, lbuf + 1))
1888                                 goto out;
1889                         v = be64_to_cpu(hpte_v);
1890                         r = be64_to_cpu(hpte_r);
1891                         err = -EINVAL;
1892                         if (!(v & HPTE_V_VALID))
1893                                 goto out;
1894                         lbuf += 2;
1895                         nb += HPTE_SIZE;
1896
1897                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1898                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1899                         err = -EIO;
1900                         ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, i, v, r,
1901                                                          tmp);
1902                         if (ret != H_SUCCESS) {
1903                                 pr_err("kvm_htab_write ret %ld i=%ld v=%lx "
1904                                        "r=%lx\n", ret, i, v, r);
1905                                 goto out;
1906                         }
1907                         if (!hpte_setup && is_vrma_hpte(v)) {
1908                                 unsigned long psize = hpte_base_page_size(v, r);
1909                                 unsigned long senc = slb_pgsize_encoding(psize);
1910                                 unsigned long lpcr;
1911
1912                                 kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
1913                                         (VRMA_VSID << SLB_VSID_SHIFT_1T);
1914                                 lpcr = senc << (LPCR_VRMASD_SH - 4);
1915                                 kvmppc_update_lpcr(kvm, lpcr, LPCR_VRMASD);
1916                                 hpte_setup = 1;
1917                         }
1918                         ++i;
1919                         hptp += 2;
1920                 }
1921
1922                 for (j = 0; j < hdr.n_invalid; ++j) {
1923                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1924                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1925                         ++i;
1926                         hptp += 2;
1927                 }
1928                 err = 0;
1929         }
1930
1931  out:
1932         /* Order HPTE updates vs. hpte_setup_done */
1933         smp_wmb();
1934         kvm->arch.hpte_setup_done = hpte_setup;
1935         mutex_unlock(&kvm->lock);
1936
1937         if (err)
1938                 return err;
1939         return nb;
1940 }
1941
1942 static int kvm_htab_release(struct inode *inode, struct file *filp)
1943 {
1944         struct kvm_htab_ctx *ctx = filp->private_data;
1945
1946         filp->private_data = NULL;
1947         if (!(ctx->flags & KVM_GET_HTAB_WRITE))
1948                 atomic_dec(&ctx->kvm->arch.hpte_mod_interest);
1949         kvm_put_kvm(ctx->kvm);
1950         kfree(ctx);
1951         return 0;
1952 }
1953
1954 static const struct file_operations kvm_htab_fops = {
1955         .read           = kvm_htab_read,
1956         .write          = kvm_htab_write,
1957         .llseek         = default_llseek,
1958         .release        = kvm_htab_release,
1959 };
1960
1961 int kvm_vm_ioctl_get_htab_fd(struct kvm *kvm, struct kvm_get_htab_fd *ghf)
1962 {
1963         int ret;
1964         struct kvm_htab_ctx *ctx;
1965         int rwflag;
1966
1967         /* reject flags we don't recognize */
1968         if (ghf->flags & ~(KVM_GET_HTAB_BOLTED_ONLY | KVM_GET_HTAB_WRITE))
1969                 return -EINVAL;
1970         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1971         if (!ctx)
1972                 return -ENOMEM;
1973         kvm_get_kvm(kvm);
1974         ctx->kvm = kvm;
1975         ctx->index = ghf->start_index;
1976         ctx->flags = ghf->flags;
1977         ctx->first_pass = 1;
1978
1979         rwflag = (ghf->flags & KVM_GET_HTAB_WRITE) ? O_WRONLY : O_RDONLY;
1980         ret = anon_inode_getfd("kvm-htab", &kvm_htab_fops, ctx, rwflag | O_CLOEXEC);
1981         if (ret < 0) {
1982                 kfree(ctx);
1983                 kvm_put_kvm(kvm);
1984                 return ret;
1985         }
1986
1987         if (rwflag == O_RDONLY) {
1988                 mutex_lock(&kvm->slots_lock);
1989                 atomic_inc(&kvm->arch.hpte_mod_interest);
1990                 /* make sure kvmppc_do_h_enter etc. see the increment */
1991                 synchronize_srcu_expedited(&kvm->srcu);
1992                 mutex_unlock(&kvm->slots_lock);
1993         }
1994
1995         return ret;
1996 }
1997
1998 struct debugfs_htab_state {
1999         struct kvm      *kvm;
2000         struct mutex    mutex;
2001         unsigned long   hpt_index;
2002         int             chars_left;
2003         int             buf_index;
2004         char            buf[64];
2005 };
2006
2007 static int debugfs_htab_open(struct inode *inode, struct file *file)
2008 {
2009         struct kvm *kvm = inode->i_private;
2010         struct debugfs_htab_state *p;
2011
2012         p = kzalloc(sizeof(*p), GFP_KERNEL);
2013         if (!p)
2014                 return -ENOMEM;
2015
2016         kvm_get_kvm(kvm);
2017         p->kvm = kvm;
2018         mutex_init(&p->mutex);
2019         file->private_data = p;
2020
2021         return nonseekable_open(inode, file);
2022 }
2023
2024 static int debugfs_htab_release(struct inode *inode, struct file *file)
2025 {
2026         struct debugfs_htab_state *p = file->private_data;
2027
2028         kvm_put_kvm(p->kvm);
2029         kfree(p);
2030         return 0;
2031 }
2032
2033 static ssize_t debugfs_htab_read(struct file *file, char __user *buf,
2034                                  size_t len, loff_t *ppos)
2035 {
2036         struct debugfs_htab_state *p = file->private_data;
2037         ssize_t ret, r;
2038         unsigned long i, n;
2039         unsigned long v, hr, gr;
2040         struct kvm *kvm;
2041         __be64 *hptp;
2042
2043         ret = mutex_lock_interruptible(&p->mutex);
2044         if (ret)
2045                 return ret;
2046
2047         if (p->chars_left) {
2048                 n = p->chars_left;
2049                 if (n > len)
2050                         n = len;
2051                 r = copy_to_user(buf, p->buf + p->buf_index, n);
2052                 n -= r;
2053                 p->chars_left -= n;
2054                 p->buf_index += n;
2055                 buf += n;
2056                 len -= n;
2057                 ret = n;
2058                 if (r) {
2059                         if (!n)
2060                                 ret = -EFAULT;
2061                         goto out;
2062                 }
2063         }
2064
2065         kvm = p->kvm;
2066         i = p->hpt_index;
2067         hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
2068         for (; len != 0 && i < kvmppc_hpt_npte(&kvm->arch.hpt);
2069              ++i, hptp += 2) {
2070                 if (!(be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)))
2071                         continue;
2072
2073                 /* lock the HPTE so it's stable and read it */
2074                 preempt_disable();
2075                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
2076                         cpu_relax();
2077                 v = be64_to_cpu(hptp[0]) & ~HPTE_V_HVLOCK;
2078                 hr = be64_to_cpu(hptp[1]);
2079                 gr = kvm->arch.hpt.rev[i].guest_rpte;
2080                 unlock_hpte(hptp, v);
2081                 preempt_enable();
2082
2083                 if (!(v & (HPTE_V_VALID | HPTE_V_ABSENT)))
2084                         continue;
2085
2086                 n = scnprintf(p->buf, sizeof(p->buf),
2087                               "%6lx %.16lx %.16lx %.16lx\n",
2088                               i, v, hr, gr);
2089                 p->chars_left = n;
2090                 if (n > len)
2091                         n = len;
2092                 r = copy_to_user(buf, p->buf, n);
2093                 n -= r;
2094                 p->chars_left -= n;
2095                 p->buf_index = n;
2096                 buf += n;
2097                 len -= n;
2098                 ret += n;
2099                 if (r) {
2100                         if (!ret)
2101                                 ret = -EFAULT;
2102                         goto out;
2103                 }
2104         }
2105         p->hpt_index = i;
2106
2107  out:
2108         mutex_unlock(&p->mutex);
2109         return ret;
2110 }
2111
2112 static ssize_t debugfs_htab_write(struct file *file, const char __user *buf,
2113                            size_t len, loff_t *ppos)
2114 {
2115         return -EACCES;
2116 }
2117
2118 static const struct file_operations debugfs_htab_fops = {
2119         .owner   = THIS_MODULE,
2120         .open    = debugfs_htab_open,
2121         .release = debugfs_htab_release,
2122         .read    = debugfs_htab_read,
2123         .write   = debugfs_htab_write,
2124         .llseek  = generic_file_llseek,
2125 };
2126
2127 void kvmppc_mmu_debugfs_init(struct kvm *kvm)
2128 {
2129         kvm->arch.htab_dentry = debugfs_create_file("htab", 0400,
2130                                                     kvm->arch.debugfs_dir, kvm,
2131                                                     &debugfs_htab_fops);
2132 }
2133
2134 void kvmppc_mmu_book3s_hv_init(struct kvm_vcpu *vcpu)
2135 {
2136         struct kvmppc_mmu *mmu = &vcpu->arch.mmu;
2137
2138         vcpu->arch.slb_nr = 32;         /* POWER7/POWER8 */
2139
2140         if (kvm_is_radix(vcpu->kvm))
2141                 mmu->xlate = kvmppc_mmu_radix_xlate;
2142         else
2143                 mmu->xlate = kvmppc_mmu_book3s_64_hv_xlate;
2144         mmu->reset_msr = kvmppc_mmu_book3s_64_hv_reset_msr;
2145
2146         vcpu->arch.hflags |= BOOK3S_HFLAG_SLB;
2147 }