arch/powerpc/kvm/book3s_64_mmu_hv.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *
   4  * Copyright 2010 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
   5  */
   6
   7 #include <linux/types.h>
   8 #include <linux/string.h>
   9 #include <linux/kvm.h>
  10 #include <linux/kvm_host.h>
  11 #include <linux/highmem.h>
  12 #include <linux/gfp.h>
  13 #include <linux/slab.h>
  14 #include <linux/hugetlb.h>
  15 #include <linux/vmalloc.h>
  16 #include <linux/srcu.h>
  17 #include <linux/anon_inodes.h>
  18 #include <linux/file.h>
  19 #include <linux/debugfs.h>
  20
  21 #include <asm/kvm_ppc.h>
  22 #include <asm/kvm_book3s.h>
  23 #include <asm/book3s/64/mmu-hash.h>
  24 #include <asm/hvcall.h>
  25 #include <asm/synch.h>
  26 #include <asm/ppc-opcode.h>
  27 #include <asm/cputable.h>
  28 #include <asm/pte-walk.h>
  29
  30 #include "trace_hv.h"
  31
  32 //#define DEBUG_RESIZE_HPT      1
  33
  34 #ifdef DEBUG_RESIZE_HPT
  35 #define resize_hpt_debug(resize, ...)                           \
  36         do {                                                    \
  37                 printk(KERN_DEBUG "RESIZE HPT %p: ", resize);   \
  38                 printk(__VA_ARGS__);                            \
  39         } while (0)
  40 #else
  41 #define resize_hpt_debug(resize, ...)                           \
  42         do { } while (0)
  43 #endif
  44
  45 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
  46                                 long pte_index, unsigned long pteh,
  47                                 unsigned long ptel, unsigned long *pte_idx_ret);
  48
  49 struct kvm_resize_hpt {
  50         /* These fields read-only after init */
  51         struct kvm *kvm;
  52         struct work_struct work;
  53         u32 order;
  54
  55         /* These fields protected by kvm->arch.mmu_setup_lock */
  56
  57         /* Possible values and their usage:
  58          *  <0     an error occurred during allocation,
  59          *  -EBUSY allocation is in the progress,
  60          *  0      allocation made successfuly.
  61          */
  62         int error;
  63
  64         /* Private to the work thread, until error != -EBUSY,
  65          * then protected by kvm->arch.mmu_setup_lock.
  66          */
  67         struct kvm_hpt_info hpt;
  68 };
  69
  70 int kvmppc_allocate_hpt(struct kvm_hpt_info *info, u32 order)
  71 {
  72         unsigned long hpt = 0;
  73         int cma = 0;
  74         struct page *page = NULL;
  75         struct revmap_entry *rev;
  76         unsigned long npte;
  77
  78         if ((order < PPC_MIN_HPT_ORDER) || (order > PPC_MAX_HPT_ORDER))
  79                 return -EINVAL;
  80
  81         page = kvm_alloc_hpt_cma(1ul << (order - PAGE_SHIFT));
  82         if (page) {
  83                 hpt = (unsigned long)pfn_to_kaddr(page_to_pfn(page));
  84                 memset((void *)hpt, 0, (1ul << order));
  85                 cma = 1;
  86         }
  87
  88         if (!hpt)
  89                 hpt = __get_free_pages(GFP_KERNEL|__GFP_ZERO|__GFP_RETRY_MAYFAIL
  90                                        |__GFP_NOWARN, order - PAGE_SHIFT);
  91
  92         if (!hpt)
  93                 return -ENOMEM;
  94
  95         /* HPTEs are 2**4 bytes long */
  96         npte = 1ul << (order - 4);
  97
  98         /* Allocate reverse map array */
  99         rev = vmalloc(array_size(npte, sizeof(struct revmap_entry)));
 100         if (!rev) {
 101                 if (cma)
 102                         kvm_free_hpt_cma(page, 1 << (order - PAGE_SHIFT));
 103                 else
 104                         free_pages(hpt, order - PAGE_SHIFT);
 105                 return -ENOMEM;
 106         }
 107
 108         info->order = order;
 109         info->virt = hpt;
 110         info->cma = cma;
 111         info->rev = rev;
 112
 113         return 0;
 114 }
 115
 116 void kvmppc_set_hpt(struct kvm *kvm, struct kvm_hpt_info *info)
 117 {
 118         atomic64_set(&kvm->arch.mmio_update, 0);
 119         kvm->arch.hpt = *info;
 120         kvm->arch.sdr1 = __pa(info->virt) | (info->order - 18);
 121
 122         pr_debug("KVM guest htab at %lx (order %ld), LPID %x\n",
 123                  info->virt, (long)info->order, kvm->arch.lpid);
 124 }
 125
 126 long kvmppc_alloc_reset_hpt(struct kvm *kvm, int order)
 127 {
 128         long err = -EBUSY;
 129         struct kvm_hpt_info info;
 130
 131         mutex_lock(&kvm->arch.mmu_setup_lock);
 132         if (kvm->arch.mmu_ready) {
 133                 kvm->arch.mmu_ready = 0;
 134                 /* order mmu_ready vs. vcpus_running */
 135                 smp_mb();
 136                 if (atomic_read(&kvm->arch.vcpus_running)) {
 137                         kvm->arch.mmu_ready = 1;
 138                         goto out;
 139                 }
 140         }
 141         if (kvm_is_radix(kvm)) {
 142                 err = kvmppc_switch_mmu_to_hpt(kvm);
 143                 if (err)
 144                         goto out;
 145         }
 146
 147         if (kvm->arch.hpt.order == order) {
 148                 /* We already have a suitable HPT */
 149
 150                 /* Set the entire HPT to 0, i.e. invalid HPTEs */
 151                 memset((void *)kvm->arch.hpt.virt, 0, 1ul << order);
 152                 /*
 153                  * Reset all the reverse-mapping chains for all memslots
 154                  */
 155                 kvmppc_rmap_reset(kvm);
 156                 err = 0;
 157                 goto out;
 158         }
 159
 160         if (kvm->arch.hpt.virt) {
 161                 kvmppc_free_hpt(&kvm->arch.hpt);
 162                 kvmppc_rmap_reset(kvm);
 163         }
 164
 165         err = kvmppc_allocate_hpt(&info, order);
 166         if (err < 0)
 167                 goto out;
 168         kvmppc_set_hpt(kvm, &info);
 169
 170 out:
 171         if (err == 0)
 172                 /* Ensure that each vcpu will flush its TLB on next entry. */
 173                 cpumask_setall(&kvm->arch.need_tlb_flush);
 174
 175         mutex_unlock(&kvm->arch.mmu_setup_lock);
 176         return err;
 177 }
 178
 179 void kvmppc_free_hpt(struct kvm_hpt_info *info)
 180 {
 181         vfree(info->rev);
 182         info->rev = NULL;
 183         if (info->cma)
 184                 kvm_free_hpt_cma(virt_to_page(info->virt),
 185                                  1 << (info->order - PAGE_SHIFT));
 186         else if (info->virt)
 187                 free_pages(info->virt, info->order - PAGE_SHIFT);
 188         info->virt = 0;
 189         info->order = 0;
 190 }
 191
 192 /* Bits in first HPTE dword for pagesize 4k, 64k or 16M */
 193 static inline unsigned long hpte0_pgsize_encoding(unsigned long pgsize)
 194 {
 195         return (pgsize > 0x1000) ? HPTE_V_LARGE : 0;
 196 }
 197
 198 /* Bits in second HPTE dword for pagesize 4k, 64k or 16M */
 199 static inline unsigned long hpte1_pgsize_encoding(unsigned long pgsize)
 200 {
 201         return (pgsize == 0x10000) ? 0x1000 : 0;
 202 }
 203
 204 void kvmppc_map_vrma(struct kvm_vcpu *vcpu, struct kvm_memory_slot *memslot,
 205                      unsigned long porder)
 206 {
 207         unsigned long i;
 208         unsigned long npages;
 209         unsigned long hp_v, hp_r;
 210         unsigned long addr, hash;
 211         unsigned long psize;
 212         unsigned long hp0, hp1;
 213         unsigned long idx_ret;
 214         long ret;
 215         struct kvm *kvm = vcpu->kvm;
 216
 217         psize = 1ul << porder;
 218         npages = memslot->npages >> (porder - PAGE_SHIFT);
 219
 220         /* VRMA can't be > 1TB */
 221         if (npages > 1ul << (40 - porder))
 222                 npages = 1ul << (40 - porder);
 223         /* Can't use more than 1 HPTE per HPTEG */
 224         if (npages > kvmppc_hpt_mask(&kvm->arch.hpt) + 1)
 225                 npages = kvmppc_hpt_mask(&kvm->arch.hpt) + 1;
 226
 227         hp0 = HPTE_V_1TB_SEG | (VRMA_VSID << (40 - 16)) |
 228                 HPTE_V_BOLTED | hpte0_pgsize_encoding(psize);
 229         hp1 = hpte1_pgsize_encoding(psize) |
 230                 HPTE_R_R | HPTE_R_C | HPTE_R_M | PP_RWXX;
 231
 232         for (i = 0; i < npages; ++i) {
 233                 addr = i << porder;
 234                 /* can't use hpt_hash since va > 64 bits */
 235                 hash = (i ^ (VRMA_VSID ^ (VRMA_VSID << 25)))
 236                         & kvmppc_hpt_mask(&kvm->arch.hpt);
 237                 /*
 238                  * We assume that the hash table is empty and no
 239                  * vcpus are using it at this stage.  Since we create
 240                  * at most one HPTE per HPTEG, we just assume entry 7
 241                  * is available and use it.
 242                  */
 243                 hash = (hash << 3) + 7;
 244                 hp_v = hp0 | ((addr >> 16) & ~0x7fUL);
 245                 hp_r = hp1 | addr;
 246                 ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, hash, hp_v, hp_r,
 247                                                  &idx_ret);
 248                 if (ret != H_SUCCESS) {
 249                         pr_err("KVM: map_vrma at %lx failed, ret=%ld\n",
 250                                addr, ret);
 251                         break;
 252                 }
 253         }
 254 }
 255
 256 int kvmppc_mmu_hv_init(void)
 257 {
 258         unsigned long host_lpid, rsvd_lpid;
 259
 260         if (!mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE))
 261                 return -EINVAL;
 262
 263         /* POWER7 has 10-bit LPIDs (12-bit in POWER8) */
 264         host_lpid = 0;
 265         if (cpu_has_feature(CPU_FTR_HVMODE))
 266                 host_lpid = mfspr(SPRN_LPID);
 267         rsvd_lpid = LPID_RSVD;
 268
 269         kvmppc_init_lpid(rsvd_lpid + 1);
 270
 271         kvmppc_claim_lpid(host_lpid);
 272         /* rsvd_lpid is reserved for use in partition switching */
 273         kvmppc_claim_lpid(rsvd_lpid);
 274
 275         return 0;
 276 }
 277
 278 static void kvmppc_mmu_book3s_64_hv_reset_msr(struct kvm_vcpu *vcpu)
 279 {
 280         unsigned long msr = vcpu->arch.intr_msr;
 281
 282         /* If transactional, change to suspend mode on IRQ delivery */
 283         if (MSR_TM_TRANSACTIONAL(vcpu->arch.shregs.msr))
 284                 msr |= MSR_TS_S;
 285         else
 286                 msr |= vcpu->arch.shregs.msr & MSR_TS_MASK;
 287         kvmppc_set_msr(vcpu, msr);
 288 }
 289
 290 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
 291                                 long pte_index, unsigned long pteh,
 292                                 unsigned long ptel, unsigned long *pte_idx_ret)
 293 {
 294         long ret;
 295
 296         /* Protect linux PTE lookup from page table destruction */
 297         rcu_read_lock_sched();  /* this disables preemption too */
 298         ret = kvmppc_do_h_enter(kvm, flags, pte_index, pteh, ptel,
 299                                 current->mm->pgd, false, pte_idx_ret);
 300         rcu_read_unlock_sched();
 301         if (ret == H_TOO_HARD) {
 302                 /* this can't happen */
 303                 pr_err("KVM: Oops, kvmppc_h_enter returned too hard!\n");
 304                 ret = H_RESOURCE;       /* or something */
 305         }
 306         return ret;
 307
 308 }
 309
 310 static struct kvmppc_slb *kvmppc_mmu_book3s_hv_find_slbe(struct kvm_vcpu *vcpu,
 311                                                          gva_t eaddr)
 312 {
 313         u64 mask;
 314         int i;
 315
 316         for (i = 0; i < vcpu->arch.slb_nr; i++) {
 317                 if (!(vcpu->arch.slb[i].orige & SLB_ESID_V))
 318                         continue;
 319
 320                 if (vcpu->arch.slb[i].origv & SLB_VSID_B_1T)
 321                         mask = ESID_MASK_1T;
 322                 else
 323                         mask = ESID_MASK;
 324
 325                 if (((vcpu->arch.slb[i].orige ^ eaddr) & mask) == 0)
 326                         return &vcpu->arch.slb[i];
 327         }
 328         return NULL;
 329 }
 330
 331 static unsigned long kvmppc_mmu_get_real_addr(unsigned long v, unsigned long r,
 332                         unsigned long ea)
 333 {
 334         unsigned long ra_mask;
 335
 336         ra_mask = kvmppc_actual_pgsz(v, r) - 1;
 337         return (r & HPTE_R_RPN & ~ra_mask) | (ea & ra_mask);
 338 }
 339
 340 static int kvmppc_mmu_book3s_64_hv_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
 341                         struct kvmppc_pte *gpte, bool data, bool iswrite)
 342 {
 343         struct kvm *kvm = vcpu->kvm;
 344         struct kvmppc_slb *slbe;
 345         unsigned long slb_v;
 346         unsigned long pp, key;
 347         unsigned long v, orig_v, gr;
 348         __be64 *hptep;
 349         long int index;
 350         int virtmode = vcpu->arch.shregs.msr & (data ? MSR_DR : MSR_IR);
 351
 352         if (kvm_is_radix(vcpu->kvm))
 353                 return kvmppc_mmu_radix_xlate(vcpu, eaddr, gpte, data, iswrite);
 354
 355         /* Get SLB entry */
 356         if (virtmode) {
 357                 slbe = kvmppc_mmu_book3s_hv_find_slbe(vcpu, eaddr);
 358                 if (!slbe)
 359                         return -EINVAL;
 360                 slb_v = slbe->origv;
 361         } else {
 362                 /* real mode access */
 363                 slb_v = vcpu->kvm->arch.vrma_slb_v;
 364         }
 365
 366         preempt_disable();
 367         /* Find the HPTE in the hash table */
 368         index = kvmppc_hv_find_lock_hpte(kvm, eaddr, slb_v,
 369                                          HPTE_V_VALID | HPTE_V_ABSENT);
 370         if (index < 0) {
 371                 preempt_enable();
 372                 return -ENOENT;
 373         }
 374         hptep = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
 375         v = orig_v = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
 376         if (cpu_has_feature(CPU_FTR_ARCH_300))
 377                 v = hpte_new_to_old_v(v, be64_to_cpu(hptep[1]));
 378         gr = kvm->arch.hpt.rev[index].guest_rpte;
 379
 380         unlock_hpte(hptep, orig_v);
 381         preempt_enable();
 382
 383         gpte->eaddr = eaddr;
 384         gpte->vpage = ((v & HPTE_V_AVPN) << 4) | ((eaddr >> 12) & 0xfff);
 385
 386         /* Get PP bits and key for permission check */
 387         pp = gr & (HPTE_R_PP0 | HPTE_R_PP);
 388         key = (vcpu->arch.shregs.msr & MSR_PR) ? SLB_VSID_KP : SLB_VSID_KS;
 389         key &= slb_v;
 390
 391         /* Calculate permissions */
 392         gpte->may_read = hpte_read_permission(pp, key);
 393         gpte->may_write = hpte_write_permission(pp, key);
 394         gpte->may_execute = gpte->may_read && !(gr & (HPTE_R_N | HPTE_R_G));
 395
 396         /* Storage key permission check for POWER7 */
 397         if (data && virtmode) {
 398                 int amrfield = hpte_get_skey_perm(gr, vcpu->arch.amr);
 399                 if (amrfield & 1)
 400                         gpte->may_read = 0;
 401                 if (amrfield & 2)
 402                         gpte->may_write = 0;
 403         }
 404
 405         /* Get the guest physical address */
 406         gpte->raddr = kvmppc_mmu_get_real_addr(v, gr, eaddr);
 407         return 0;
 408 }
 409
 410 /*
 411  * Quick test for whether an instruction is a load or a store.
 412  * If the instruction is a load or a store, then this will indicate
 413  * which it is, at least on server processors.  (Embedded processors
 414  * have some external PID instructions that don't follow the rule
 415  * embodied here.)  If the instruction isn't a load or store, then
 416  * this doesn't return anything useful.
 417  */
 418 static int instruction_is_store(unsigned int instr)
 419 {
 420         unsigned int mask;
 421
 422         mask = 0x10000000;
 423         if ((instr & 0xfc000000) == 0x7c000000)
 424                 mask = 0x100;           /* major opcode 31 */
 425         return (instr & mask) != 0;
 426 }
 427
 428 int kvmppc_hv_emulate_mmio(struct kvm_run *run, struct kvm_vcpu *vcpu,
 429                            unsigned long gpa, gva_t ea, int is_store)
 430 {
 431         u32 last_inst;
 432
 433         /*
 434          * Fast path - check if the guest physical address corresponds to a
 435          * device on the FAST_MMIO_BUS, if so we can avoid loading the
 436          * instruction all together, then we can just handle it and return.
 437          */
 438         if (is_store) {
 439                 int idx, ret;
 440
 441                 idx = srcu_read_lock(&vcpu->kvm->srcu);
 442                 ret = kvm_io_bus_write(vcpu, KVM_FAST_MMIO_BUS, (gpa_t) gpa, 0,
 443                                        NULL);
 444                 srcu_read_unlock(&vcpu->kvm->srcu, idx);
 445                 if (!ret) {
 446                         kvmppc_set_pc(vcpu, kvmppc_get_pc(vcpu) + 4);
 447                         return RESUME_GUEST;
 448                 }
 449         }
 450
 451         /*
 452          * If we fail, we just return to the guest and try executing it again.
 453          */
 454         if (kvmppc_get_last_inst(vcpu, INST_GENERIC, &last_inst) !=
 455                 EMULATE_DONE)
 456                 return RESUME_GUEST;
 457
 458         /*
 459          * WARNING: We do not know for sure whether the instruction we just
 460          * read from memory is the same that caused the fault in the first
 461          * place.  If the instruction we read is neither an load or a store,
 462          * then it can't access memory, so we don't need to worry about
 463          * enforcing access permissions.  So, assuming it is a load or
 464          * store, we just check that its direction (load or store) is
 465          * consistent with the original fault, since that's what we
 466          * checked the access permissions against.  If there is a mismatch
 467          * we just return and retry the instruction.
 468          */
 469
 470         if (instruction_is_store(last_inst) != !!is_store)
 471                 return RESUME_GUEST;
 472
 473         /*
 474          * Emulated accesses are emulated by looking at the hash for
 475          * translation once, then performing the access later. The
 476          * translation could be invalidated in the meantime in which
 477          * point performing the subsequent memory access on the old
 478          * physical address could possibly be a security hole for the
 479          * guest (but not the host).
 480          *
 481          * This is less of an issue for MMIO stores since they aren't
 482          * globally visible. It could be an issue for MMIO loads to
 483          * a certain extent but we'll ignore it for now.
 484          */
 485
 486         vcpu->arch.paddr_accessed = gpa;
 487         vcpu->arch.vaddr_accessed = ea;
 488         return kvmppc_emulate_mmio(run, vcpu);
 489 }
 490
 491 int kvmppc_book3s_hv_page_fault(struct kvm_run *run, struct kvm_vcpu *vcpu,
 492                                 unsigned long ea, unsigned long dsisr)
 493 {
 494         struct kvm *kvm = vcpu->kvm;
 495         unsigned long hpte[3], r;
 496         unsigned long hnow_v, hnow_r;
 497         __be64 *hptep;
 498         unsigned long mmu_seq, psize, pte_size;
 499         unsigned long gpa_base, gfn_base;
 500         unsigned long gpa, gfn, hva, pfn;
 501         struct kvm_memory_slot *memslot;
 502         unsigned long *rmap;
 503         struct revmap_entry *rev;
 504         struct page *page, *pages[1];
 505         long index, ret, npages;
 506         bool is_ci;
 507         unsigned int writing, write_ok;
 508         struct vm_area_struct *vma;
 509         unsigned long rcbits;
 510         long mmio_update;
 511
 512         if (kvm_is_radix(kvm))
 513                 return kvmppc_book3s_radix_page_fault(run, vcpu, ea, dsisr);
 514
 515         /*
 516          * Real-mode code has already searched the HPT and found the
 517          * entry we're interested in.  Lock the entry and check that
 518          * it hasn't changed.  If it has, just return and re-execute the
 519          * instruction.
 520          */
 521         if (ea != vcpu->arch.pgfault_addr)
 522                 return RESUME_GUEST;
 523
 524         if (vcpu->arch.pgfault_cache) {
 525                 mmio_update = atomic64_read(&kvm->arch.mmio_update);
 526                 if (mmio_update == vcpu->arch.pgfault_cache->mmio_update) {
 527                         r = vcpu->arch.pgfault_cache->rpte;
 528                         psize = kvmppc_actual_pgsz(vcpu->arch.pgfault_hpte[0],
 529                                                    r);
 530                         gpa_base = r & HPTE_R_RPN & ~(psize - 1);
 531                         gfn_base = gpa_base >> PAGE_SHIFT;
 532                         gpa = gpa_base | (ea & (psize - 1));
 533                         return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
 534                                                 dsisr & DSISR_ISSTORE);
 535                 }
 536         }
 537         index = vcpu->arch.pgfault_index;
 538         hptep = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
 539         rev = &kvm->arch.hpt.rev[index];
 540         preempt_disable();
 541         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 542                 cpu_relax();
 543         hpte[0] = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
 544         hpte[1] = be64_to_cpu(hptep[1]);
 545         hpte[2] = r = rev->guest_rpte;
 546         unlock_hpte(hptep, hpte[0]);
 547         preempt_enable();
 548
 549         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
 550                 hpte[0] = hpte_new_to_old_v(hpte[0], hpte[1]);
 551                 hpte[1] = hpte_new_to_old_r(hpte[1]);
 552         }
 553         if (hpte[0] != vcpu->arch.pgfault_hpte[0] ||
 554             hpte[1] != vcpu->arch.pgfault_hpte[1])
 555                 return RESUME_GUEST;
 556
 557         /* Translate the logical address and get the page */
 558         psize = kvmppc_actual_pgsz(hpte[0], r);
 559         gpa_base = r & HPTE_R_RPN & ~(psize - 1);
 560         gfn_base = gpa_base >> PAGE_SHIFT;
 561         gpa = gpa_base | (ea & (psize - 1));
 562         gfn = gpa >> PAGE_SHIFT;
 563         memslot = gfn_to_memslot(kvm, gfn);
 564
 565         trace_kvm_page_fault_enter(vcpu, hpte, memslot, ea, dsisr);
 566
 567         /* No memslot means it's an emulated MMIO region */
 568         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
 569                 return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
 570                                               dsisr & DSISR_ISSTORE);
 571
 572         /*
 573          * This should never happen, because of the slot_is_aligned()
 574          * check in kvmppc_do_h_enter().
 575          */
 576         if (gfn_base < memslot->base_gfn)
 577                 return -EFAULT;
 578
 579         /* used to check for invalidations in progress */
 580         mmu_seq = kvm->mmu_notifier_seq;
 581         smp_rmb();
 582
 583         ret = -EFAULT;
 584         is_ci = false;
 585         pfn = 0;
 586         page = NULL;
 587         pte_size = PAGE_SIZE;
 588         writing = (dsisr & DSISR_ISSTORE) != 0;
 589         /* If writing != 0, then the HPTE must allow writing, if we get here */
 590         write_ok = writing;
 591         hva = gfn_to_hva_memslot(memslot, gfn);
 592         npages = get_user_pages_fast(hva, 1, writing ? FOLL_WRITE : 0, pages);
 593         if (npages < 1) {
 594                 /* Check if it's an I/O mapping */
 595                 down_read(&current->mm->mmap_sem);
 596                 vma = find_vma(current->mm, hva);
 597                 if (vma && vma->vm_start <= hva && hva + psize <= vma->vm_end &&
 598                     (vma->vm_flags & VM_PFNMAP)) {
 599                         pfn = vma->vm_pgoff +
 600                                 ((hva - vma->vm_start) >> PAGE_SHIFT);
 601                         pte_size = psize;
 602                         is_ci = pte_ci(__pte((pgprot_val(vma->vm_page_prot))));
 603                         write_ok = vma->vm_flags & VM_WRITE;
 604                 }
 605                 up_read(&current->mm->mmap_sem);
 606                 if (!pfn)
 607                         goto out_put;
 608         } else {
 609                 page = pages[0];
 610                 pfn = page_to_pfn(page);
 611                 if (PageHuge(page)) {
 612                         page = compound_head(page);
 613                         pte_size <<= compound_order(page);
 614                 }
 615                 /* if the guest wants write access, see if that is OK */
 616                 if (!writing && hpte_is_writable(r)) {
 617                         pte_t *ptep, pte;
 618                         unsigned long flags;
 619                         /*
 620                          * We need to protect against page table destruction
 621                          * hugepage split and collapse.
 622                          */
 623                         local_irq_save(flags);
 624                         ptep = find_current_mm_pte(current->mm->pgd,
 625                                                    hva, NULL, NULL);
 626                         if (ptep) {
 627                                 pte = kvmppc_read_update_linux_pte(ptep, 1);
 628                                 if (__pte_write(pte))
 629                                         write_ok = 1;
 630                         }
 631                         local_irq_restore(flags);
 632                 }
 633         }
 634
 635         if (psize > pte_size)
 636                 goto out_put;
 637
 638         /* Check WIMG vs. the actual page we're accessing */
 639         if (!hpte_cache_flags_ok(r, is_ci)) {
 640                 if (is_ci)
 641                         goto out_put;
 642                 /*
 643                  * Allow guest to map emulated device memory as
 644                  * uncacheable, but actually make it cacheable.
 645                  */
 646                 r = (r & ~(HPTE_R_W|HPTE_R_I|HPTE_R_G)) | HPTE_R_M;
 647         }
 648
 649         /*
 650          * Set the HPTE to point to pfn.
 651          * Since the pfn is at PAGE_SIZE granularity, make sure we
 652          * don't mask out lower-order bits if psize < PAGE_SIZE.
 653          */
 654         if (psize < PAGE_SIZE)
 655                 psize = PAGE_SIZE;
 656         r = (r & HPTE_R_KEY_HI) | (r & ~(HPTE_R_PP0 - psize)) |
 657                                         ((pfn << PAGE_SHIFT) & ~(psize - 1));
 658         if (hpte_is_writable(r) && !write_ok)
 659                 r = hpte_make_readonly(r);
 660         ret = RESUME_GUEST;
 661         preempt_disable();
 662         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 663                 cpu_relax();
 664         hnow_v = be64_to_cpu(hptep[0]);
 665         hnow_r = be64_to_cpu(hptep[1]);
 666         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
 667                 hnow_v = hpte_new_to_old_v(hnow_v, hnow_r);
 668                 hnow_r = hpte_new_to_old_r(hnow_r);
 669         }
 670
 671         /*
 672          * If the HPT is being resized, don't update the HPTE,
 673          * instead let the guest retry after the resize operation is complete.
 674          * The synchronization for mmu_ready test vs. set is provided
 675          * by the HPTE lock.
 676          */
 677         if (!kvm->arch.mmu_ready)
 678                 goto out_unlock;
 679
 680         if ((hnow_v & ~HPTE_V_HVLOCK) != hpte[0] || hnow_r != hpte[1] ||
 681             rev->guest_rpte != hpte[2])
 682                 /* HPTE has been changed under us; let the guest retry */
 683                 goto out_unlock;
 684         hpte[0] = (hpte[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
 685
 686         /* Always put the HPTE in the rmap chain for the page base address */
 687         rmap = &memslot->arch.rmap[gfn_base - memslot->base_gfn];
 688         lock_rmap(rmap);
 689
 690         /* Check if we might have been invalidated; let the guest retry if so */
 691         ret = RESUME_GUEST;
 692         if (mmu_notifier_retry(vcpu->kvm, mmu_seq)) {
 693                 unlock_rmap(rmap);
 694                 goto out_unlock;
 695         }
 696
 697         /* Only set R/C in real HPTE if set in both *rmap and guest_rpte */
 698         rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
 699         r &= rcbits | ~(HPTE_R_R | HPTE_R_C);
 700
 701         if (be64_to_cpu(hptep[0]) & HPTE_V_VALID) {
 702                 /* HPTE was previously valid, so we need to invalidate it */
 703                 unlock_rmap(rmap);
 704                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
 705                 kvmppc_invalidate_hpte(kvm, hptep, index);
 706                 /* don't lose previous R and C bits */
 707                 r |= be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
 708         } else {
 709                 kvmppc_add_revmap_chain(kvm, rev, rmap, index, 0);
 710         }
 711
 712         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
 713                 r = hpte_old_to_new_r(hpte[0], r);
 714                 hpte[0] = hpte_old_to_new_v(hpte[0]);
 715         }
 716         hptep[1] = cpu_to_be64(r);
 717         eieio();
 718         __unlock_hpte(hptep, hpte[0]);
 719         asm volatile("ptesync" : : : "memory");
 720         preempt_enable();
 721         if (page && hpte_is_writable(r))
 722                 SetPageDirty(page);
 723
 724  out_put:
 725         trace_kvm_page_fault_exit(vcpu, hpte, ret);
 726
 727         if (page) {
 728                 /*
 729                  * We drop pages[0] here, not page because page might
 730                  * have been set to the head page of a compound, but
 731                  * we have to drop the reference on the correct tail
 732                  * page to match the get inside gup()
 733                  */
 734                 put_page(pages[0]);
 735         }
 736         return ret;
 737
 738  out_unlock:
 739         __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
 740         preempt_enable();
 741         goto out_put;
 742 }
 743
 744 void kvmppc_rmap_reset(struct kvm *kvm)
 745 {
 746         struct kvm_memslots *slots;
 747         struct kvm_memory_slot *memslot;
 748         int srcu_idx;
 749
 750         srcu_idx = srcu_read_lock(&kvm->srcu);
 751         slots = kvm_memslots(kvm);
 752         kvm_for_each_memslot(memslot, slots) {
 753                 /* Mutual exclusion with kvm_unmap_hva_range etc. */
 754                 spin_lock(&kvm->mmu_lock);
 755                 /*
 756                  * This assumes it is acceptable to lose reference and
 757                  * change bits across a reset.
 758                  */
 759                 memset(memslot->arch.rmap, 0,
 760                        memslot->npages * sizeof(*memslot->arch.rmap));
 761                 spin_unlock(&kvm->mmu_lock);
 762         }
 763         srcu_read_unlock(&kvm->srcu, srcu_idx);
 764 }
 765
 766 typedef int (*hva_handler_fn)(struct kvm *kvm, struct kvm_memory_slot *memslot,
 767                               unsigned long gfn);
 768
 769 static int kvm_handle_hva_range(struct kvm *kvm,
 770                                 unsigned long start,
 771                                 unsigned long end,
 772                                 hva_handler_fn handler)
 773 {
 774         int ret;
 775         int retval = 0;
 776         struct kvm_memslots *slots;
 777         struct kvm_memory_slot *memslot;
 778
 779         slots = kvm_memslots(kvm);
 780         kvm_for_each_memslot(memslot, slots) {
 781                 unsigned long hva_start, hva_end;
 782                 gfn_t gfn, gfn_end;
 783
 784                 hva_start = max(start, memslot->userspace_addr);
 785                 hva_end = min(end, memslot->userspace_addr +
 786                                         (memslot->npages << PAGE_SHIFT));
 787                 if (hva_start >= hva_end)
 788                         continue;
 789                 /*
 790                  * {gfn(page) | page intersects with [hva_start, hva_end)} =
 791                  * {gfn, gfn+1, ..., gfn_end-1}.
 792                  */
 793                 gfn = hva_to_gfn_memslot(hva_start, memslot);
 794                 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
 795
 796                 for (; gfn < gfn_end; ++gfn) {
 797                         ret = handler(kvm, memslot, gfn);
 798                         retval |= ret;
 799                 }
 800         }
 801
 802         return retval;
 803 }
 804
 805 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 806                           hva_handler_fn handler)
 807 {
 808         return kvm_handle_hva_range(kvm, hva, hva + 1, handler);
 809 }
 810
 811 /* Must be called with both HPTE and rmap locked */
 812 static void kvmppc_unmap_hpte(struct kvm *kvm, unsigned long i,
 813                               struct kvm_memory_slot *memslot,
 814                               unsigned long *rmapp, unsigned long gfn)
 815 {
 816         __be64 *hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
 817         struct revmap_entry *rev = kvm->arch.hpt.rev;
 818         unsigned long j, h;
 819         unsigned long ptel, psize, rcbits;
 820
 821         j = rev[i].forw;
 822         if (j == i) {
 823                 /* chain is now empty */
 824                 *rmapp &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
 825         } else {
 826                 /* remove i from chain */
 827                 h = rev[i].back;
 828                 rev[h].forw = j;
 829                 rev[j].back = h;
 830                 rev[i].forw = rev[i].back = i;
 831                 *rmapp = (*rmapp & ~KVMPPC_RMAP_INDEX) | j;
 832         }
 833
 834         /* Now check and modify the HPTE */
 835         ptel = rev[i].guest_rpte;
 836         psize = kvmppc_actual_pgsz(be64_to_cpu(hptep[0]), ptel);
 837         if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
 838             hpte_rpn(ptel, psize) == gfn) {
 839                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
 840                 kvmppc_invalidate_hpte(kvm, hptep, i);
 841                 hptep[1] &= ~cpu_to_be64(HPTE_R_KEY_HI | HPTE_R_KEY_LO);
 842                 /* Harvest R and C */
 843                 rcbits = be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
 844                 *rmapp |= rcbits << KVMPPC_RMAP_RC_SHIFT;
 845                 if ((rcbits & HPTE_R_C) && memslot->dirty_bitmap)
 846                         kvmppc_update_dirty_map(memslot, gfn, psize);
 847                 if (rcbits & ~rev[i].guest_rpte) {
 848                         rev[i].guest_rpte = ptel | rcbits;
 849                         note_hpte_modification(kvm, &rev[i]);
 850                 }
 851         }
 852 }
 853
 854 static int kvm_unmap_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
 855                            unsigned long gfn)
 856 {
 857         unsigned long i;
 858         __be64 *hptep;
 859         unsigned long *rmapp;
 860
 861         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
 862         for (;;) {
 863                 lock_rmap(rmapp);
 864                 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 865                         unlock_rmap(rmapp);
 866                         break;
 867                 }
 868
 869                 /*
 870                  * To avoid an ABBA deadlock with the HPTE lock bit,
 871                  * we can't spin on the HPTE lock while holding the
 872                  * rmap chain lock.
 873                  */
 874                 i = *rmapp & KVMPPC_RMAP_INDEX;
 875                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
 876                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 877                         /* unlock rmap before spinning on the HPTE lock */
 878                         unlock_rmap(rmapp);
 879                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
 880                                 cpu_relax();
 881                         continue;
 882                 }
 883
 884                 kvmppc_unmap_hpte(kvm, i, memslot, rmapp, gfn);
 885                 unlock_rmap(rmapp);
 886                 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
 887         }
 888         return 0;
 889 }
 890
 891 int kvm_unmap_hva_range_hv(struct kvm *kvm, unsigned long start, unsigned long end)
 892 {
 893         hva_handler_fn handler;
 894
 895         handler = kvm_is_radix(kvm) ? kvm_unmap_radix : kvm_unmap_rmapp;
 896         kvm_handle_hva_range(kvm, start, end, handler);
 897         return 0;
 898 }
 899
 900 void kvmppc_core_flush_memslot_hv(struct kvm *kvm,
 901                                   struct kvm_memory_slot *memslot)
 902 {
 903         unsigned long gfn;
 904         unsigned long n;
 905         unsigned long *rmapp;
 906
 907         gfn = memslot->base_gfn;
 908         rmapp = memslot->arch.rmap;
 909         if (kvm_is_radix(kvm)) {
 910                 kvmppc_radix_flush_memslot(kvm, memslot);
 911                 return;
 912         }
 913
 914         for (n = memslot->npages; n; --n, ++gfn) {
 915                 /*
 916                  * Testing the present bit without locking is OK because
 917                  * the memslot has been marked invalid already, and hence
 918                  * no new HPTEs referencing this page can be created,
 919                  * thus the present bit can't go from 0 to 1.
 920                  */
 921                 if (*rmapp & KVMPPC_RMAP_PRESENT)
 922                         kvm_unmap_rmapp(kvm, memslot, gfn);
 923                 ++rmapp;
 924         }
 925 }
 926
 927 static int kvm_age_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
 928                          unsigned long gfn)
 929 {
 930         struct revmap_entry *rev = kvm->arch.hpt.rev;
 931         unsigned long head, i, j;
 932         __be64 *hptep;
 933         int ret = 0;
 934         unsigned long *rmapp;
 935
 936         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
 937  retry:
 938         lock_rmap(rmapp);
 939         if (*rmapp & KVMPPC_RMAP_REFERENCED) {
 940                 *rmapp &= ~KVMPPC_RMAP_REFERENCED;
 941                 ret = 1;
 942         }
 943         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 944                 unlock_rmap(rmapp);
 945                 return ret;
 946         }
 947
 948         i = head = *rmapp & KVMPPC_RMAP_INDEX;
 949         do {
 950                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
 951                 j = rev[i].forw;
 952
 953                 /* If this HPTE isn't referenced, ignore it */
 954                 if (!(be64_to_cpu(hptep[1]) & HPTE_R_R))
 955                         continue;
 956
 957                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 958                         /* unlock rmap before spinning on the HPTE lock */
 959                         unlock_rmap(rmapp);
 960                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
 961                                 cpu_relax();
 962                         goto retry;
 963                 }
 964
 965                 /* Now check and modify the HPTE */
 966                 if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
 967                     (be64_to_cpu(hptep[1]) & HPTE_R_R)) {
 968                         kvmppc_clear_ref_hpte(kvm, hptep, i);
 969                         if (!(rev[i].guest_rpte & HPTE_R_R)) {
 970                                 rev[i].guest_rpte |= HPTE_R_R;
 971                                 note_hpte_modification(kvm, &rev[i]);
 972                         }
 973                         ret = 1;
 974                 }
 975                 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
 976         } while ((i = j) != head);
 977
 978         unlock_rmap(rmapp);
 979         return ret;
 980 }
 981
 982 int kvm_age_hva_hv(struct kvm *kvm, unsigned long start, unsigned long end)
 983 {
 984         hva_handler_fn handler;
 985
 986         handler = kvm_is_radix(kvm) ? kvm_age_radix : kvm_age_rmapp;
 987         return kvm_handle_hva_range(kvm, start, end, handler);
 988 }
 989
 990 static int kvm_test_age_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
 991                               unsigned long gfn)
 992 {
 993         struct revmap_entry *rev = kvm->arch.hpt.rev;
 994         unsigned long head, i, j;
 995         unsigned long *hp;
 996         int ret = 1;
 997         unsigned long *rmapp;
 998
 999         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1000         if (*rmapp & KVMPPC_RMAP_REFERENCED)
1001                 return 1;
1002
1003         lock_rmap(rmapp);
1004         if (*rmapp & KVMPPC_RMAP_REFERENCED)
1005                 goto out;
1006
1007         if (*rmapp & KVMPPC_RMAP_PRESENT) {
1008                 i = head = *rmapp & KVMPPC_RMAP_INDEX;
1009                 do {
1010                         hp = (unsigned long *)(kvm->arch.hpt.virt + (i << 4));
1011                         j = rev[i].forw;
1012                         if (be64_to_cpu(hp[1]) & HPTE_R_R)
1013                                 goto out;
1014                 } while ((i = j) != head);
1015         }
1016         ret = 0;
1017
1018  out:
1019         unlock_rmap(rmapp);
1020         return ret;
1021 }
1022
1023 int kvm_test_age_hva_hv(struct kvm *kvm, unsigned long hva)
1024 {
1025         hva_handler_fn handler;
1026
1027         handler = kvm_is_radix(kvm) ? kvm_test_age_radix : kvm_test_age_rmapp;
1028         return kvm_handle_hva(kvm, hva, handler);
1029 }
1030
1031 void kvm_set_spte_hva_hv(struct kvm *kvm, unsigned long hva, pte_t pte)
1032 {
1033         hva_handler_fn handler;
1034
1035         handler = kvm_is_radix(kvm) ? kvm_unmap_radix : kvm_unmap_rmapp;
1036         kvm_handle_hva(kvm, hva, handler);
1037 }
1038
1039 static int vcpus_running(struct kvm *kvm)
1040 {
1041         return atomic_read(&kvm->arch.vcpus_running) != 0;
1042 }
1043
1044 /*
1045  * Returns the number of system pages that are dirty.
1046  * This can be more than 1 if we find a huge-page HPTE.
1047  */
1048 static int kvm_test_clear_dirty_npages(struct kvm *kvm, unsigned long *rmapp)
1049 {
1050         struct revmap_entry *rev = kvm->arch.hpt.rev;
1051         unsigned long head, i, j;
1052         unsigned long n;
1053         unsigned long v, r;
1054         __be64 *hptep;
1055         int npages_dirty = 0;
1056
1057  retry:
1058         lock_rmap(rmapp);
1059         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
1060                 unlock_rmap(rmapp);
1061                 return npages_dirty;
1062         }
1063
1064         i = head = *rmapp & KVMPPC_RMAP_INDEX;
1065         do {
1066                 unsigned long hptep1;
1067                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
1068                 j = rev[i].forw;
1069
1070                 /*
1071                  * Checking the C (changed) bit here is racy since there
1072                  * is no guarantee about when the hardware writes it back.
1073                  * If the HPTE is not writable then it is stable since the
1074                  * page can't be written to, and we would have done a tlbie
1075                  * (which forces the hardware to complete any writeback)
1076                  * when making the HPTE read-only.
1077                  * If vcpus are running then this call is racy anyway
1078                  * since the page could get dirtied subsequently, so we
1079                  * expect there to be a further call which would pick up
1080                  * any delayed C bit writeback.
1081                  * Otherwise we need to do the tlbie even if C==0 in
1082                  * order to pick up any delayed writeback of C.
1083                  */
1084                 hptep1 = be64_to_cpu(hptep[1]);
1085                 if (!(hptep1 & HPTE_R_C) &&
1086                     (!hpte_is_writable(hptep1) || vcpus_running(kvm)))
1087                         continue;
1088
1089                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
1090                         /* unlock rmap before spinning on the HPTE lock */
1091                         unlock_rmap(rmapp);
1092                         while (hptep[0] & cpu_to_be64(HPTE_V_HVLOCK))
1093                                 cpu_relax();
1094                         goto retry;
1095                 }
1096
1097                 /* Now check and modify the HPTE */
1098                 if (!(hptep[0] & cpu_to_be64(HPTE_V_VALID))) {
1099                         __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
1100                         continue;
1101                 }
1102
1103                 /* need to make it temporarily absent so C is stable */
1104                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
1105                 kvmppc_invalidate_hpte(kvm, hptep, i);
1106                 v = be64_to_cpu(hptep[0]);
1107                 r = be64_to_cpu(hptep[1]);
1108                 if (r & HPTE_R_C) {
1109                         hptep[1] = cpu_to_be64(r & ~HPTE_R_C);
1110                         if (!(rev[i].guest_rpte & HPTE_R_C)) {
1111                                 rev[i].guest_rpte |= HPTE_R_C;
1112                                 note_hpte_modification(kvm, &rev[i]);
1113                         }
1114                         n = kvmppc_actual_pgsz(v, r);
1115                         n = (n + PAGE_SIZE - 1) >> PAGE_SHIFT;
1116                         if (n > npages_dirty)
1117                                 npages_dirty = n;
1118                         eieio();
1119                 }
1120                 v &= ~HPTE_V_ABSENT;
1121                 v |= HPTE_V_VALID;
1122                 __unlock_hpte(hptep, v);
1123         } while ((i = j) != head);
1124
1125         unlock_rmap(rmapp);
1126         return npages_dirty;
1127 }
1128
1129 void kvmppc_harvest_vpa_dirty(struct kvmppc_vpa *vpa,
1130                               struct kvm_memory_slot *memslot,
1131                               unsigned long *map)
1132 {
1133         unsigned long gfn;
1134
1135         if (!vpa->dirty || !vpa->pinned_addr)
1136                 return;
1137         gfn = vpa->gpa >> PAGE_SHIFT;
1138         if (gfn < memslot->base_gfn ||
1139             gfn >= memslot->base_gfn + memslot->npages)
1140                 return;
1141
1142         vpa->dirty = false;
1143         if (map)
1144                 __set_bit_le(gfn - memslot->base_gfn, map);
1145 }
1146
1147 long kvmppc_hv_get_dirty_log_hpt(struct kvm *kvm,
1148                         struct kvm_memory_slot *memslot, unsigned long *map)
1149 {
1150         unsigned long i;
1151         unsigned long *rmapp;
1152
1153         preempt_disable();
1154         rmapp = memslot->arch.rmap;
1155         for (i = 0; i < memslot->npages; ++i) {
1156                 int npages = kvm_test_clear_dirty_npages(kvm, rmapp);
1157                 /*
1158                  * Note that if npages > 0 then i must be a multiple of npages,
1159                  * since we always put huge-page HPTEs in the rmap chain
1160                  * corresponding to their page base address.
1161                  */
1162                 if (npages)
1163                         set_dirty_bits(map, i, npages);
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, FOLL_WRITE, 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         int srcu_idx;
1208
1209         put_page(page);
1210
1211         if (!dirty)
1212                 return;
1213
1214         /* We need to mark this page dirty in the memslot dirty_bitmap, if any */
1215         gfn = gpa >> PAGE_SHIFT;
1216         srcu_idx = srcu_read_lock(&kvm->srcu);
1217         memslot = gfn_to_memslot(kvm, gfn);
1218         if (memslot && memslot->dirty_bitmap)
1219                 set_bit_le(gfn - memslot->base_gfn, memslot->dirty_bitmap);
1220         srcu_read_unlock(&kvm->srcu, srcu_idx);
1221 }
1222
1223 /*
1224  * HPT resizing
1225  */
1226 static int resize_hpt_allocate(struct kvm_resize_hpt *resize)
1227 {
1228         int rc;
1229
1230         rc = kvmppc_allocate_hpt(&resize->hpt, resize->order);
1231         if (rc < 0)
1232                 return rc;
1233
1234         resize_hpt_debug(resize, "resize_hpt_allocate(): HPT @ 0x%lx\n",
1235                          resize->hpt.virt);
1236
1237         return 0;
1238 }
1239
1240 static unsigned long resize_hpt_rehash_hpte(struct kvm_resize_hpt *resize,
1241                                             unsigned long idx)
1242 {
1243         struct kvm *kvm = resize->kvm;
1244         struct kvm_hpt_info *old = &kvm->arch.hpt;
1245         struct kvm_hpt_info *new = &resize->hpt;
1246         unsigned long old_hash_mask = (1ULL << (old->order - 7)) - 1;
1247         unsigned long new_hash_mask = (1ULL << (new->order - 7)) - 1;
1248         __be64 *hptep, *new_hptep;
1249         unsigned long vpte, rpte, guest_rpte;
1250         int ret;
1251         struct revmap_entry *rev;
1252         unsigned long apsize, avpn, pteg, hash;
1253         unsigned long new_idx, new_pteg, replace_vpte;
1254         int pshift;
1255
1256         hptep = (__be64 *)(old->virt + (idx << 4));
1257
1258         /* Guest is stopped, so new HPTEs can't be added or faulted
1259          * in, only unmapped or altered by host actions.  So, it's
1260          * safe to check this before we take the HPTE lock */
1261         vpte = be64_to_cpu(hptep[0]);
1262         if (!(vpte & HPTE_V_VALID) && !(vpte & HPTE_V_ABSENT))
1263                 return 0; /* nothing to do */
1264
1265         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
1266                 cpu_relax();
1267
1268         vpte = be64_to_cpu(hptep[0]);
1269
1270         ret = 0;
1271         if (!(vpte & HPTE_V_VALID) && !(vpte & HPTE_V_ABSENT))
1272                 /* Nothing to do */
1273                 goto out;
1274
1275         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1276                 rpte = be64_to_cpu(hptep[1]);
1277                 vpte = hpte_new_to_old_v(vpte, rpte);
1278         }
1279
1280         /* Unmap */
1281         rev = &old->rev[idx];
1282         guest_rpte = rev->guest_rpte;
1283
1284         ret = -EIO;
1285         apsize = kvmppc_actual_pgsz(vpte, guest_rpte);
1286         if (!apsize)
1287                 goto out;
1288
1289         if (vpte & HPTE_V_VALID) {
1290                 unsigned long gfn = hpte_rpn(guest_rpte, apsize);
1291                 int srcu_idx = srcu_read_lock(&kvm->srcu);
1292                 struct kvm_memory_slot *memslot =
1293                         __gfn_to_memslot(kvm_memslots(kvm), gfn);
1294
1295                 if (memslot) {
1296                         unsigned long *rmapp;
1297                         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1298
1299                         lock_rmap(rmapp);
1300                         kvmppc_unmap_hpte(kvm, idx, memslot, rmapp, gfn);
1301                         unlock_rmap(rmapp);
1302                 }
1303
1304                 srcu_read_unlock(&kvm->srcu, srcu_idx);
1305         }
1306
1307         /* Reload PTE after unmap */
1308         vpte = be64_to_cpu(hptep[0]);
1309         BUG_ON(vpte & HPTE_V_VALID);
1310         BUG_ON(!(vpte & HPTE_V_ABSENT));
1311
1312         ret = 0;
1313         if (!(vpte & HPTE_V_BOLTED))
1314                 goto out;
1315
1316         rpte = be64_to_cpu(hptep[1]);
1317
1318         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1319                 vpte = hpte_new_to_old_v(vpte, rpte);
1320                 rpte = hpte_new_to_old_r(rpte);
1321         }
1322
1323         pshift = kvmppc_hpte_base_page_shift(vpte, rpte);
1324         avpn = HPTE_V_AVPN_VAL(vpte) & ~(((1ul << pshift) - 1) >> 23);
1325         pteg = idx / HPTES_PER_GROUP;
1326         if (vpte & HPTE_V_SECONDARY)
1327                 pteg = ~pteg;
1328
1329         if (!(vpte & HPTE_V_1TB_SEG)) {
1330                 unsigned long offset, vsid;
1331
1332                 /* We only have 28 - 23 bits of offset in avpn */
1333                 offset = (avpn & 0x1f) << 23;
1334                 vsid = avpn >> 5;
1335                 /* We can find more bits from the pteg value */
1336                 if (pshift < 23)
1337                         offset |= ((vsid ^ pteg) & old_hash_mask) << pshift;
1338
1339                 hash = vsid ^ (offset >> pshift);
1340         } else {
1341                 unsigned long offset, vsid;
1342
1343                 /* We only have 40 - 23 bits of seg_off in avpn */
1344                 offset = (avpn & 0x1ffff) << 23;
1345                 vsid = avpn >> 17;
1346                 if (pshift < 23)
1347                         offset |= ((vsid ^ (vsid << 25) ^ pteg) & old_hash_mask) << pshift;
1348
1349                 hash = vsid ^ (vsid << 25) ^ (offset >> pshift);
1350         }
1351
1352         new_pteg = hash & new_hash_mask;
1353         if (vpte & HPTE_V_SECONDARY)
1354                 new_pteg = ~hash & new_hash_mask;
1355
1356         new_idx = new_pteg * HPTES_PER_GROUP + (idx % HPTES_PER_GROUP);
1357         new_hptep = (__be64 *)(new->virt + (new_idx << 4));
1358
1359         replace_vpte = be64_to_cpu(new_hptep[0]);
1360         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1361                 unsigned long replace_rpte = be64_to_cpu(new_hptep[1]);
1362                 replace_vpte = hpte_new_to_old_v(replace_vpte, replace_rpte);
1363         }
1364
1365         if (replace_vpte & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1366                 BUG_ON(new->order >= old->order);
1367
1368                 if (replace_vpte & HPTE_V_BOLTED) {
1369                         if (vpte & HPTE_V_BOLTED)
1370                                 /* Bolted collision, nothing we can do */
1371                                 ret = -ENOSPC;
1372                         /* Discard the new HPTE */
1373                         goto out;
1374                 }
1375
1376                 /* Discard the previous HPTE */
1377         }
1378
1379         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1380                 rpte = hpte_old_to_new_r(vpte, rpte);
1381                 vpte = hpte_old_to_new_v(vpte);
1382         }
1383
1384         new_hptep[1] = cpu_to_be64(rpte);
1385         new->rev[new_idx].guest_rpte = guest_rpte;
1386         /* No need for a barrier, since new HPT isn't active */
1387         new_hptep[0] = cpu_to_be64(vpte);
1388         unlock_hpte(new_hptep, vpte);
1389
1390 out:
1391         unlock_hpte(hptep, vpte);
1392         return ret;
1393 }
1394
1395 static int resize_hpt_rehash(struct kvm_resize_hpt *resize)
1396 {
1397         struct kvm *kvm = resize->kvm;
1398         unsigned  long i;
1399         int rc;
1400
1401         for (i = 0; i < kvmppc_hpt_npte(&kvm->arch.hpt); i++) {
1402                 rc = resize_hpt_rehash_hpte(resize, i);
1403                 if (rc != 0)
1404                         return rc;
1405         }
1406
1407         return 0;
1408 }
1409
1410 static void resize_hpt_pivot(struct kvm_resize_hpt *resize)
1411 {
1412         struct kvm *kvm = resize->kvm;
1413         struct kvm_hpt_info hpt_tmp;
1414
1415         /* Exchange the pending tables in the resize structure with
1416          * the active tables */
1417
1418         resize_hpt_debug(resize, "resize_hpt_pivot()\n");
1419
1420         spin_lock(&kvm->mmu_lock);
1421         asm volatile("ptesync" : : : "memory");
1422
1423         hpt_tmp = kvm->arch.hpt;
1424         kvmppc_set_hpt(kvm, &resize->hpt);
1425         resize->hpt = hpt_tmp;
1426
1427         spin_unlock(&kvm->mmu_lock);
1428
1429         synchronize_srcu_expedited(&kvm->srcu);
1430
1431         if (cpu_has_feature(CPU_FTR_ARCH_300))
1432                 kvmppc_setup_partition_table(kvm);
1433
1434         resize_hpt_debug(resize, "resize_hpt_pivot() done\n");
1435 }
1436
1437 static void resize_hpt_release(struct kvm *kvm, struct kvm_resize_hpt *resize)
1438 {
1439         if (WARN_ON(!mutex_is_locked(&kvm->arch.mmu_setup_lock)))
1440                 return;
1441
1442         if (!resize)
1443                 return;
1444
1445         if (resize->error != -EBUSY) {
1446                 if (resize->hpt.virt)
1447                         kvmppc_free_hpt(&resize->hpt);
1448                 kfree(resize);
1449         }
1450
1451         if (kvm->arch.resize_hpt == resize)
1452                 kvm->arch.resize_hpt = NULL;
1453 }
1454
1455 static void resize_hpt_prepare_work(struct work_struct *work)
1456 {
1457         struct kvm_resize_hpt *resize = container_of(work,
1458                                                      struct kvm_resize_hpt,
1459                                                      work);
1460         struct kvm *kvm = resize->kvm;
1461         int err = 0;
1462
1463         if (WARN_ON(resize->error != -EBUSY))
1464                 return;
1465
1466         mutex_lock(&kvm->arch.mmu_setup_lock);
1467
1468         /* Request is still current? */
1469         if (kvm->arch.resize_hpt == resize) {
1470                 /* We may request large allocations here:
1471                  * do not sleep with kvm->arch.mmu_setup_lock held for a while.
1472                  */
1473                 mutex_unlock(&kvm->arch.mmu_setup_lock);
1474
1475                 resize_hpt_debug(resize, "resize_hpt_prepare_work(): order = %d\n",
1476                                  resize->order);
1477
1478                 err = resize_hpt_allocate(resize);
1479
1480                 /* We have strict assumption about -EBUSY
1481                  * when preparing for HPT resize.
1482                  */
1483                 if (WARN_ON(err == -EBUSY))
1484                         err = -EINPROGRESS;
1485
1486                 mutex_lock(&kvm->arch.mmu_setup_lock);
1487                 /* It is possible that kvm->arch.resize_hpt != resize
1488                  * after we grab kvm->arch.mmu_setup_lock again.
1489                  */
1490         }
1491
1492         resize->error = err;
1493
1494         if (kvm->arch.resize_hpt != resize)
1495                 resize_hpt_release(kvm, resize);
1496
1497         mutex_unlock(&kvm->arch.mmu_setup_lock);
1498 }
1499
1500 long kvm_vm_ioctl_resize_hpt_prepare(struct kvm *kvm,
1501                                      struct kvm_ppc_resize_hpt *rhpt)
1502 {
1503         unsigned long flags = rhpt->flags;
1504         unsigned long shift = rhpt->shift;
1505         struct kvm_resize_hpt *resize;
1506         int ret;
1507
1508         if (flags != 0 || kvm_is_radix(kvm))
1509                 return -EINVAL;
1510
1511         if (shift && ((shift < 18) || (shift > 46)))
1512                 return -EINVAL;
1513
1514         mutex_lock(&kvm->arch.mmu_setup_lock);
1515
1516         resize = kvm->arch.resize_hpt;
1517
1518         if (resize) {
1519                 if (resize->order == shift) {
1520                         /* Suitable resize in progress? */
1521                         ret = resize->error;
1522                         if (ret == -EBUSY)
1523                                 ret = 100; /* estimated time in ms */
1524                         else if (ret)
1525                                 resize_hpt_release(kvm, resize);
1526
1527                         goto out;
1528                 }
1529
1530                 /* not suitable, cancel it */
1531                 resize_hpt_release(kvm, resize);
1532         }
1533
1534         ret = 0;
1535         if (!shift)
1536                 goto out; /* nothing to do */
1537
1538         /* start new resize */
1539
1540         resize = kzalloc(sizeof(*resize), GFP_KERNEL);
1541         if (!resize) {
1542                 ret = -ENOMEM;
1543                 goto out;
1544         }
1545
1546         resize->error = -EBUSY;
1547         resize->order = shift;
1548         resize->kvm = kvm;
1549         INIT_WORK(&resize->work, resize_hpt_prepare_work);
1550         kvm->arch.resize_hpt = resize;
1551
1552         schedule_work(&resize->work);
1553
1554         ret = 100; /* estimated time in ms */
1555
1556 out:
1557         mutex_unlock(&kvm->arch.mmu_setup_lock);
1558         return ret;
1559 }
1560
1561 static void resize_hpt_boot_vcpu(void *opaque)
1562 {
1563         /* Nothing to do, just force a KVM exit */
1564 }
1565
1566 long kvm_vm_ioctl_resize_hpt_commit(struct kvm *kvm,
1567                                     struct kvm_ppc_resize_hpt *rhpt)
1568 {
1569         unsigned long flags = rhpt->flags;
1570         unsigned long shift = rhpt->shift;
1571         struct kvm_resize_hpt *resize;
1572         long ret;
1573
1574         if (flags != 0 || kvm_is_radix(kvm))
1575                 return -EINVAL;
1576
1577         if (shift && ((shift < 18) || (shift > 46)))
1578                 return -EINVAL;
1579
1580         mutex_lock(&kvm->arch.mmu_setup_lock);
1581
1582         resize = kvm->arch.resize_hpt;
1583
1584         /* This shouldn't be possible */
1585         ret = -EIO;
1586         if (WARN_ON(!kvm->arch.mmu_ready))
1587                 goto out_no_hpt;
1588
1589         /* Stop VCPUs from running while we mess with the HPT */
1590         kvm->arch.mmu_ready = 0;
1591         smp_mb();
1592
1593         /* Boot all CPUs out of the guest so they re-read
1594          * mmu_ready */
1595         on_each_cpu(resize_hpt_boot_vcpu, NULL, 1);
1596
1597         ret = -ENXIO;
1598         if (!resize || (resize->order != shift))
1599                 goto out;
1600
1601         ret = resize->error;
1602         if (ret)
1603                 goto out;
1604
1605         ret = resize_hpt_rehash(resize);
1606         if (ret)
1607                 goto out;
1608
1609         resize_hpt_pivot(resize);
1610
1611 out:
1612         /* Let VCPUs run again */
1613         kvm->arch.mmu_ready = 1;
1614         smp_mb();
1615 out_no_hpt:
1616         resize_hpt_release(kvm, resize);
1617         mutex_unlock(&kvm->arch.mmu_setup_lock);
1618         return ret;
1619 }
1620
1621 /*
1622  * Functions for reading and writing the hash table via reads and
1623  * writes on a file descriptor.
1624  *
1625  * Reads return the guest view of the hash table, which has to be
1626  * pieced together from the real hash table and the guest_rpte
1627  * values in the revmap array.
1628  *
1629  * On writes, each HPTE written is considered in turn, and if it
1630  * is valid, it is written to the HPT as if an H_ENTER with the
1631  * exact flag set was done.  When the invalid count is non-zero
1632  * in the header written to the stream, the kernel will make
1633  * sure that that many HPTEs are invalid, and invalidate them
1634  * if not.
1635  */
1636
1637 struct kvm_htab_ctx {
1638         unsigned long   index;
1639         unsigned long   flags;
1640         struct kvm      *kvm;
1641         int             first_pass;
1642 };
1643
1644 #define HPTE_SIZE       (2 * sizeof(unsigned long))
1645
1646 /*
1647  * Returns 1 if this HPT entry has been modified or has pending
1648  * R/C bit changes.
1649  */
1650 static int hpte_dirty(struct revmap_entry *revp, __be64 *hptp)
1651 {
1652         unsigned long rcbits_unset;
1653
1654         if (revp->guest_rpte & HPTE_GR_MODIFIED)
1655                 return 1;
1656
1657         /* Also need to consider changes in reference and changed bits */
1658         rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1659         if ((be64_to_cpu(hptp[0]) & HPTE_V_VALID) &&
1660             (be64_to_cpu(hptp[1]) & rcbits_unset))
1661                 return 1;
1662
1663         return 0;
1664 }
1665
1666 static long record_hpte(unsigned long flags, __be64 *hptp,
1667                         unsigned long *hpte, struct revmap_entry *revp,
1668                         int want_valid, int first_pass)
1669 {
1670         unsigned long v, r, hr;
1671         unsigned long rcbits_unset;
1672         int ok = 1;
1673         int valid, dirty;
1674
1675         /* Unmodified entries are uninteresting except on the first pass */
1676         dirty = hpte_dirty(revp, hptp);
1677         if (!first_pass && !dirty)
1678                 return 0;
1679
1680         valid = 0;
1681         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1682                 valid = 1;
1683                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) &&
1684                     !(be64_to_cpu(hptp[0]) & HPTE_V_BOLTED))
1685                         valid = 0;
1686         }
1687         if (valid != want_valid)
1688                 return 0;
1689
1690         v = r = 0;
1691         if (valid || dirty) {
1692                 /* lock the HPTE so it's stable and read it */
1693                 preempt_disable();
1694                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
1695                         cpu_relax();
1696                 v = be64_to_cpu(hptp[0]);
1697                 hr = be64_to_cpu(hptp[1]);
1698                 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1699                         v = hpte_new_to_old_v(v, hr);
1700                         hr = hpte_new_to_old_r(hr);
1701                 }
1702
1703                 /* re-evaluate valid and dirty from synchronized HPTE value */
1704                 valid = !!(v & HPTE_V_VALID);
1705                 dirty = !!(revp->guest_rpte & HPTE_GR_MODIFIED);
1706
1707                 /* Harvest R and C into guest view if necessary */
1708                 rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1709                 if (valid && (rcbits_unset & hr)) {
1710                         revp->guest_rpte |= (hr &
1711                                 (HPTE_R_R | HPTE_R_C)) | HPTE_GR_MODIFIED;
1712                         dirty = 1;
1713                 }
1714
1715                 if (v & HPTE_V_ABSENT) {
1716                         v &= ~HPTE_V_ABSENT;
1717                         v |= HPTE_V_VALID;
1718                         valid = 1;
1719                 }
1720                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) && !(v & HPTE_V_BOLTED))
1721                         valid = 0;
1722
1723                 r = revp->guest_rpte;
1724                 /* only clear modified if this is the right sort of entry */
1725                 if (valid == want_valid && dirty) {
1726                         r &= ~HPTE_GR_MODIFIED;
1727                         revp->guest_rpte = r;
1728                 }
1729                 unlock_hpte(hptp, be64_to_cpu(hptp[0]));
1730                 preempt_enable();
1731                 if (!(valid == want_valid && (first_pass || dirty)))
1732                         ok = 0;
1733         }
1734         hpte[0] = cpu_to_be64(v);
1735         hpte[1] = cpu_to_be64(r);
1736         return ok;
1737 }
1738
1739 static ssize_t kvm_htab_read(struct file *file, char __user *buf,
1740                              size_t count, loff_t *ppos)
1741 {
1742         struct kvm_htab_ctx *ctx = file->private_data;
1743         struct kvm *kvm = ctx->kvm;
1744         struct kvm_get_htab_header hdr;
1745         __be64 *hptp;
1746         struct revmap_entry *revp;
1747         unsigned long i, nb, nw;
1748         unsigned long __user *lbuf;
1749         struct kvm_get_htab_header __user *hptr;
1750         unsigned long flags;
1751         int first_pass;
1752         unsigned long hpte[2];
1753
1754         if (!access_ok(buf, count))
1755                 return -EFAULT;
1756         if (kvm_is_radix(kvm))
1757                 return 0;
1758
1759         first_pass = ctx->first_pass;
1760         flags = ctx->flags;
1761
1762         i = ctx->index;
1763         hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
1764         revp = kvm->arch.hpt.rev + i;
1765         lbuf = (unsigned long __user *)buf;
1766
1767         nb = 0;
1768         while (nb + sizeof(hdr) + HPTE_SIZE < count) {
1769                 /* Initialize header */
1770                 hptr = (struct kvm_get_htab_header __user *)buf;
1771                 hdr.n_valid = 0;
1772                 hdr.n_invalid = 0;
1773                 nw = nb;
1774                 nb += sizeof(hdr);
1775                 lbuf = (unsigned long __user *)(buf + sizeof(hdr));
1776
1777                 /* Skip uninteresting entries, i.e. clean on not-first pass */
1778                 if (!first_pass) {
1779                         while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1780                                !hpte_dirty(revp, hptp)) {
1781                                 ++i;
1782                                 hptp += 2;
1783                                 ++revp;
1784                         }
1785                 }
1786                 hdr.index = i;
1787
1788                 /* Grab a series of valid entries */
1789                 while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1790                        hdr.n_valid < 0xffff &&
1791                        nb + HPTE_SIZE < count &&
1792                        record_hpte(flags, hptp, hpte, revp, 1, first_pass)) {
1793                         /* valid entry, write it out */
1794                         ++hdr.n_valid;
1795                         if (__put_user(hpte[0], lbuf) ||
1796                             __put_user(hpte[1], lbuf + 1))
1797                                 return -EFAULT;
1798                         nb += HPTE_SIZE;
1799                         lbuf += 2;
1800                         ++i;
1801                         hptp += 2;
1802                         ++revp;
1803                 }
1804                 /* Now skip invalid entries while we can */
1805                 while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1806                        hdr.n_invalid < 0xffff &&
1807                        record_hpte(flags, hptp, hpte, revp, 0, first_pass)) {
1808                         /* found an invalid entry */
1809                         ++hdr.n_invalid;
1810                         ++i;
1811                         hptp += 2;
1812                         ++revp;
1813                 }
1814
1815                 if (hdr.n_valid || hdr.n_invalid) {
1816                         /* write back the header */
1817                         if (__copy_to_user(hptr, &hdr, sizeof(hdr)))
1818                                 return -EFAULT;
1819                         nw = nb;
1820                         buf = (char __user *)lbuf;
1821                 } else {
1822                         nb = nw;
1823                 }
1824
1825                 /* Check if we've wrapped around the hash table */
1826                 if (i >= kvmppc_hpt_npte(&kvm->arch.hpt)) {
1827                         i = 0;
1828                         ctx->first_pass = 0;
1829                         break;
1830                 }
1831         }
1832
1833         ctx->index = i;
1834
1835         return nb;
1836 }
1837
1838 static ssize_t kvm_htab_write(struct file *file, const char __user *buf,
1839                               size_t count, loff_t *ppos)
1840 {
1841         struct kvm_htab_ctx *ctx = file->private_data;
1842         struct kvm *kvm = ctx->kvm;
1843         struct kvm_get_htab_header hdr;
1844         unsigned long i, j;
1845         unsigned long v, r;
1846         unsigned long __user *lbuf;
1847         __be64 *hptp;
1848         unsigned long tmp[2];
1849         ssize_t nb;
1850         long int err, ret;
1851         int mmu_ready;
1852         int pshift;
1853
1854         if (!access_ok(buf, count))
1855                 return -EFAULT;
1856         if (kvm_is_radix(kvm))
1857                 return -EINVAL;
1858
1859         /* lock out vcpus from running while we're doing this */
1860         mutex_lock(&kvm->arch.mmu_setup_lock);
1861         mmu_ready = kvm->arch.mmu_ready;
1862         if (mmu_ready) {
1863                 kvm->arch.mmu_ready = 0;        /* temporarily */
1864                 /* order mmu_ready vs. vcpus_running */
1865                 smp_mb();
1866                 if (atomic_read(&kvm->arch.vcpus_running)) {
1867                         kvm->arch.mmu_ready = 1;
1868                         mutex_unlock(&kvm->arch.mmu_setup_lock);
1869                         return -EBUSY;
1870                 }
1871         }
1872
1873         err = 0;
1874         for (nb = 0; nb + sizeof(hdr) <= count; ) {
1875                 err = -EFAULT;
1876                 if (__copy_from_user(&hdr, buf, sizeof(hdr)))
1877                         break;
1878
1879                 err = 0;
1880                 if (nb + hdr.n_valid * HPTE_SIZE > count)
1881                         break;
1882
1883                 nb += sizeof(hdr);
1884                 buf += sizeof(hdr);
1885
1886                 err = -EINVAL;
1887                 i = hdr.index;
1888                 if (i >= kvmppc_hpt_npte(&kvm->arch.hpt) ||
1889                     i + hdr.n_valid + hdr.n_invalid > kvmppc_hpt_npte(&kvm->arch.hpt))
1890                         break;
1891
1892                 hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
1893                 lbuf = (unsigned long __user *)buf;
1894                 for (j = 0; j < hdr.n_valid; ++j) {
1895                         __be64 hpte_v;
1896                         __be64 hpte_r;
1897
1898                         err = -EFAULT;
1899                         if (__get_user(hpte_v, lbuf) ||
1900                             __get_user(hpte_r, lbuf + 1))
1901                                 goto out;
1902                         v = be64_to_cpu(hpte_v);
1903                         r = be64_to_cpu(hpte_r);
1904                         err = -EINVAL;
1905                         if (!(v & HPTE_V_VALID))
1906                                 goto out;
1907                         pshift = kvmppc_hpte_base_page_shift(v, r);
1908                         if (pshift <= 0)
1909                                 goto out;
1910                         lbuf += 2;
1911                         nb += HPTE_SIZE;
1912
1913                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1914                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1915                         err = -EIO;
1916                         ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, i, v, r,
1917                                                          tmp);
1918                         if (ret != H_SUCCESS) {
1919                                 pr_err("kvm_htab_write ret %ld i=%ld v=%lx "
1920                                        "r=%lx\n", ret, i, v, r);
1921                                 goto out;
1922                         }
1923                         if (!mmu_ready && is_vrma_hpte(v)) {
1924                                 unsigned long senc, lpcr;
1925
1926                                 senc = slb_pgsize_encoding(1ul << pshift);
1927                                 kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
1928                                         (VRMA_VSID << SLB_VSID_SHIFT_1T);
1929                                 if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
1930                                         lpcr = senc << (LPCR_VRMASD_SH - 4);
1931                                         kvmppc_update_lpcr(kvm, lpcr,
1932                                                            LPCR_VRMASD);
1933                                 } else {
1934                                         kvmppc_setup_partition_table(kvm);
1935                                 }
1936                                 mmu_ready = 1;
1937                         }
1938                         ++i;
1939                         hptp += 2;
1940                 }
1941
1942                 for (j = 0; j < hdr.n_invalid; ++j) {
1943                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1944                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1945                         ++i;
1946                         hptp += 2;
1947                 }
1948                 err = 0;
1949         }
1950
1951  out:
1952         /* Order HPTE updates vs. mmu_ready */
1953         smp_wmb();
1954         kvm->arch.mmu_ready = mmu_ready;
1955         mutex_unlock(&kvm->arch.mmu_setup_lock);
1956
1957         if (err)
1958                 return err;
1959         return nb;
1960 }
1961
1962 static int kvm_htab_release(struct inode *inode, struct file *filp)
1963 {
1964         struct kvm_htab_ctx *ctx = filp->private_data;
1965
1966         filp->private_data = NULL;
1967         if (!(ctx->flags & KVM_GET_HTAB_WRITE))
1968                 atomic_dec(&ctx->kvm->arch.hpte_mod_interest);
1969         kvm_put_kvm(ctx->kvm);
1970         kfree(ctx);
1971         return 0;
1972 }
1973
1974 static const struct file_operations kvm_htab_fops = {
1975         .read           = kvm_htab_read,
1976         .write          = kvm_htab_write,
1977         .llseek         = default_llseek,
1978         .release        = kvm_htab_release,
1979 };
1980
1981 int kvm_vm_ioctl_get_htab_fd(struct kvm *kvm, struct kvm_get_htab_fd *ghf)
1982 {
1983         int ret;
1984         struct kvm_htab_ctx *ctx;
1985         int rwflag;
1986
1987         /* reject flags we don't recognize */
1988         if (ghf->flags & ~(KVM_GET_HTAB_BOLTED_ONLY | KVM_GET_HTAB_WRITE))
1989                 return -EINVAL;
1990         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1991         if (!ctx)
1992                 return -ENOMEM;
1993         kvm_get_kvm(kvm);
1994         ctx->kvm = kvm;
1995         ctx->index = ghf->start_index;
1996         ctx->flags = ghf->flags;
1997         ctx->first_pass = 1;
1998
1999         rwflag = (ghf->flags & KVM_GET_HTAB_WRITE) ? O_WRONLY : O_RDONLY;
2000         ret = anon_inode_getfd("kvm-htab", &kvm_htab_fops, ctx, rwflag | O_CLOEXEC);
2001         if (ret < 0) {
2002                 kfree(ctx);
2003                 kvm_put_kvm(kvm);
2004                 return ret;
2005         }
2006
2007         if (rwflag == O_RDONLY) {
2008                 mutex_lock(&kvm->slots_lock);
2009                 atomic_inc(&kvm->arch.hpte_mod_interest);
2010                 /* make sure kvmppc_do_h_enter etc. see the increment */
2011                 synchronize_srcu_expedited(&kvm->srcu);
2012                 mutex_unlock(&kvm->slots_lock);
2013         }
2014
2015         return ret;
2016 }
2017
2018 struct debugfs_htab_state {
2019         struct kvm      *kvm;
2020         struct mutex    mutex;
2021         unsigned long   hpt_index;
2022         int             chars_left;
2023         int             buf_index;
2024         char            buf[64];
2025 };
2026
2027 static int debugfs_htab_open(struct inode *inode, struct file *file)
2028 {
2029         struct kvm *kvm = inode->i_private;
2030         struct debugfs_htab_state *p;
2031
2032         p = kzalloc(sizeof(*p), GFP_KERNEL);
2033         if (!p)
2034                 return -ENOMEM;
2035
2036         kvm_get_kvm(kvm);
2037         p->kvm = kvm;
2038         mutex_init(&p->mutex);
2039         file->private_data = p;
2040
2041         return nonseekable_open(inode, file);
2042 }
2043
2044 static int debugfs_htab_release(struct inode *inode, struct file *file)
2045 {
2046         struct debugfs_htab_state *p = file->private_data;
2047
2048         kvm_put_kvm(p->kvm);
2049         kfree(p);
2050         return 0;
2051 }
2052
2053 static ssize_t debugfs_htab_read(struct file *file, char __user *buf,
2054                                  size_t len, loff_t *ppos)
2055 {
2056         struct debugfs_htab_state *p = file->private_data;
2057         ssize_t ret, r;
2058         unsigned long i, n;
2059         unsigned long v, hr, gr;
2060         struct kvm *kvm;
2061         __be64 *hptp;
2062
2063         kvm = p->kvm;
2064         if (kvm_is_radix(kvm))
2065                 return 0;
2066
2067         ret = mutex_lock_interruptible(&p->mutex);
2068         if (ret)
2069                 return ret;
2070
2071         if (p->chars_left) {
2072                 n = p->chars_left;
2073                 if (n > len)
2074                         n = len;
2075                 r = copy_to_user(buf, p->buf + p->buf_index, n);
2076                 n -= r;
2077                 p->chars_left -= n;
2078                 p->buf_index += n;
2079                 buf += n;
2080                 len -= n;
2081                 ret = n;
2082                 if (r) {
2083                         if (!n)
2084                                 ret = -EFAULT;
2085                         goto out;
2086                 }
2087         }
2088
2089         i = p->hpt_index;
2090         hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
2091         for (; len != 0 && i < kvmppc_hpt_npte(&kvm->arch.hpt);
2092              ++i, hptp += 2) {
2093                 if (!(be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)))
2094                         continue;
2095
2096                 /* lock the HPTE so it's stable and read it */
2097                 preempt_disable();
2098                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
2099                         cpu_relax();
2100                 v = be64_to_cpu(hptp[0]) & ~HPTE_V_HVLOCK;
2101                 hr = be64_to_cpu(hptp[1]);
2102                 gr = kvm->arch.hpt.rev[i].guest_rpte;
2103                 unlock_hpte(hptp, v);
2104                 preempt_enable();
2105
2106                 if (!(v & (HPTE_V_VALID | HPTE_V_ABSENT)))
2107                         continue;
2108
2109                 n = scnprintf(p->buf, sizeof(p->buf),
2110                               "%6lx %.16lx %.16lx %.16lx\n",
2111                               i, v, hr, gr);
2112                 p->chars_left = n;
2113                 if (n > len)
2114                         n = len;
2115                 r = copy_to_user(buf, p->buf, n);
2116                 n -= r;
2117                 p->chars_left -= n;
2118                 p->buf_index = n;
2119                 buf += n;
2120                 len -= n;
2121                 ret += n;
2122                 if (r) {
2123                         if (!ret)
2124                                 ret = -EFAULT;
2125                         goto out;
2126                 }
2127         }
2128         p->hpt_index = i;
2129
2130  out:
2131         mutex_unlock(&p->mutex);
2132         return ret;
2133 }
2134
2135 static ssize_t debugfs_htab_write(struct file *file, const char __user *buf,
2136                            size_t len, loff_t *ppos)
2137 {
2138         return -EACCES;
2139 }
2140
2141 static const struct file_operations debugfs_htab_fops = {
2142         .owner   = THIS_MODULE,
2143         .open    = debugfs_htab_open,
2144         .release = debugfs_htab_release,
2145         .read    = debugfs_htab_read,
2146         .write   = debugfs_htab_write,
2147         .llseek  = generic_file_llseek,
2148 };
2149
2150 void kvmppc_mmu_debugfs_init(struct kvm *kvm)
2151 {
2152         kvm->arch.htab_dentry = debugfs_create_file("htab", 0400,
2153                                                     kvm->arch.debugfs_dir, kvm,
2154                                                     &debugfs_htab_fops);
2155 }
2156
2157 void kvmppc_mmu_book3s_hv_init(struct kvm_vcpu *vcpu)
2158 {
2159         struct kvmppc_mmu *mmu = &vcpu->arch.mmu;
2160
2161         vcpu->arch.slb_nr = 32;         /* POWER7/POWER8 */
2162
2163         mmu->xlate = kvmppc_mmu_book3s_64_hv_xlate;
2164         mmu->reset_msr = kvmppc_mmu_book3s_64_hv_reset_msr;
2165
2166         vcpu->arch.hflags |= BOOK3S_HFLAG_SLB;
2167 }