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 long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
 279                                 long pte_index, unsigned long pteh,
 280                                 unsigned long ptel, unsigned long *pte_idx_ret)
 281 {
 282         long ret;
 283
 284         preempt_disable();
 285         ret = kvmppc_do_h_enter(kvm, flags, pte_index, pteh, ptel,
 286                                 kvm->mm->pgd, false, pte_idx_ret);
 287         preempt_enable();
 288         if (ret == H_TOO_HARD) {
 289                 /* this can't happen */
 290                 pr_err("KVM: Oops, kvmppc_h_enter returned too hard!\n");
 291                 ret = H_RESOURCE;       /* or something */
 292         }
 293         return ret;
 294
 295 }
 296
 297 static struct kvmppc_slb *kvmppc_mmu_book3s_hv_find_slbe(struct kvm_vcpu *vcpu,
 298                                                          gva_t eaddr)
 299 {
 300         u64 mask;
 301         int i;
 302
 303         for (i = 0; i < vcpu->arch.slb_nr; i++) {
 304                 if (!(vcpu->arch.slb[i].orige & SLB_ESID_V))
 305                         continue;
 306
 307                 if (vcpu->arch.slb[i].origv & SLB_VSID_B_1T)
 308                         mask = ESID_MASK_1T;
 309                 else
 310                         mask = ESID_MASK;
 311
 312                 if (((vcpu->arch.slb[i].orige ^ eaddr) & mask) == 0)
 313                         return &vcpu->arch.slb[i];
 314         }
 315         return NULL;
 316 }
 317
 318 static unsigned long kvmppc_mmu_get_real_addr(unsigned long v, unsigned long r,
 319                         unsigned long ea)
 320 {
 321         unsigned long ra_mask;
 322
 323         ra_mask = kvmppc_actual_pgsz(v, r) - 1;
 324         return (r & HPTE_R_RPN & ~ra_mask) | (ea & ra_mask);
 325 }
 326
 327 static int kvmppc_mmu_book3s_64_hv_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
 328                         struct kvmppc_pte *gpte, bool data, bool iswrite)
 329 {
 330         struct kvm *kvm = vcpu->kvm;
 331         struct kvmppc_slb *slbe;
 332         unsigned long slb_v;
 333         unsigned long pp, key;
 334         unsigned long v, orig_v, gr;
 335         __be64 *hptep;
 336         long int index;
 337         int virtmode = vcpu->arch.shregs.msr & (data ? MSR_DR : MSR_IR);
 338
 339         if (kvm_is_radix(vcpu->kvm))
 340                 return kvmppc_mmu_radix_xlate(vcpu, eaddr, gpte, data, iswrite);
 341
 342         /* Get SLB entry */
 343         if (virtmode) {
 344                 slbe = kvmppc_mmu_book3s_hv_find_slbe(vcpu, eaddr);
 345                 if (!slbe)
 346                         return -EINVAL;
 347                 slb_v = slbe->origv;
 348         } else {
 349                 /* real mode access */
 350                 slb_v = vcpu->kvm->arch.vrma_slb_v;
 351         }
 352
 353         preempt_disable();
 354         /* Find the HPTE in the hash table */
 355         index = kvmppc_hv_find_lock_hpte(kvm, eaddr, slb_v,
 356                                          HPTE_V_VALID | HPTE_V_ABSENT);
 357         if (index < 0) {
 358                 preempt_enable();
 359                 return -ENOENT;
 360         }
 361         hptep = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
 362         v = orig_v = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
 363         if (cpu_has_feature(CPU_FTR_ARCH_300))
 364                 v = hpte_new_to_old_v(v, be64_to_cpu(hptep[1]));
 365         gr = kvm->arch.hpt.rev[index].guest_rpte;
 366
 367         unlock_hpte(hptep, orig_v);
 368         preempt_enable();
 369
 370         gpte->eaddr = eaddr;
 371         gpte->vpage = ((v & HPTE_V_AVPN) << 4) | ((eaddr >> 12) & 0xfff);
 372
 373         /* Get PP bits and key for permission check */
 374         pp = gr & (HPTE_R_PP0 | HPTE_R_PP);
 375         key = (vcpu->arch.shregs.msr & MSR_PR) ? SLB_VSID_KP : SLB_VSID_KS;
 376         key &= slb_v;
 377
 378         /* Calculate permissions */
 379         gpte->may_read = hpte_read_permission(pp, key);
 380         gpte->may_write = hpte_write_permission(pp, key);
 381         gpte->may_execute = gpte->may_read && !(gr & (HPTE_R_N | HPTE_R_G));
 382
 383         /* Storage key permission check for POWER7 */
 384         if (data && virtmode) {
 385                 int amrfield = hpte_get_skey_perm(gr, vcpu->arch.amr);
 386                 if (amrfield & 1)
 387                         gpte->may_read = 0;
 388                 if (amrfield & 2)
 389                         gpte->may_write = 0;
 390         }
 391
 392         /* Get the guest physical address */
 393         gpte->raddr = kvmppc_mmu_get_real_addr(v, gr, eaddr);
 394         return 0;
 395 }
 396
 397 /*
 398  * Quick test for whether an instruction is a load or a store.
 399  * If the instruction is a load or a store, then this will indicate
 400  * which it is, at least on server processors.  (Embedded processors
 401  * have some external PID instructions that don't follow the rule
 402  * embodied here.)  If the instruction isn't a load or store, then
 403  * this doesn't return anything useful.
 404  */
 405 static int instruction_is_store(unsigned int instr)
 406 {
 407         unsigned int mask;
 408
 409         mask = 0x10000000;
 410         if ((instr & 0xfc000000) == 0x7c000000)
 411                 mask = 0x100;           /* major opcode 31 */
 412         return (instr & mask) != 0;
 413 }
 414
 415 int kvmppc_hv_emulate_mmio(struct kvm_run *run, struct kvm_vcpu *vcpu,
 416                            unsigned long gpa, gva_t ea, int is_store)
 417 {
 418         u32 last_inst;
 419
 420         /*
 421          * Fast path - check if the guest physical address corresponds to a
 422          * device on the FAST_MMIO_BUS, if so we can avoid loading the
 423          * instruction all together, then we can just handle it and return.
 424          */
 425         if (is_store) {
 426                 int idx, ret;
 427
 428                 idx = srcu_read_lock(&vcpu->kvm->srcu);
 429                 ret = kvm_io_bus_write(vcpu, KVM_FAST_MMIO_BUS, (gpa_t) gpa, 0,
 430                                        NULL);
 431                 srcu_read_unlock(&vcpu->kvm->srcu, idx);
 432                 if (!ret) {
 433                         kvmppc_set_pc(vcpu, kvmppc_get_pc(vcpu) + 4);
 434                         return RESUME_GUEST;
 435                 }
 436         }
 437
 438         /*
 439          * If we fail, we just return to the guest and try executing it again.
 440          */
 441         if (kvmppc_get_last_inst(vcpu, INST_GENERIC, &last_inst) !=
 442                 EMULATE_DONE)
 443                 return RESUME_GUEST;
 444
 445         /*
 446          * WARNING: We do not know for sure whether the instruction we just
 447          * read from memory is the same that caused the fault in the first
 448          * place.  If the instruction we read is neither an load or a store,
 449          * then it can't access memory, so we don't need to worry about
 450          * enforcing access permissions.  So, assuming it is a load or
 451          * store, we just check that its direction (load or store) is
 452          * consistent with the original fault, since that's what we
 453          * checked the access permissions against.  If there is a mismatch
 454          * we just return and retry the instruction.
 455          */
 456
 457         if (instruction_is_store(last_inst) != !!is_store)
 458                 return RESUME_GUEST;
 459
 460         /*
 461          * Emulated accesses are emulated by looking at the hash for
 462          * translation once, then performing the access later. The
 463          * translation could be invalidated in the meantime in which
 464          * point performing the subsequent memory access on the old
 465          * physical address could possibly be a security hole for the
 466          * guest (but not the host).
 467          *
 468          * This is less of an issue for MMIO stores since they aren't
 469          * globally visible. It could be an issue for MMIO loads to
 470          * a certain extent but we'll ignore it for now.
 471          */
 472
 473         vcpu->arch.paddr_accessed = gpa;
 474         vcpu->arch.vaddr_accessed = ea;
 475         return kvmppc_emulate_mmio(run, vcpu);
 476 }
 477
 478 int kvmppc_book3s_hv_page_fault(struct kvm_run *run, struct kvm_vcpu *vcpu,
 479                                 unsigned long ea, unsigned long dsisr)
 480 {
 481         struct kvm *kvm = vcpu->kvm;
 482         unsigned long hpte[3], r;
 483         unsigned long hnow_v, hnow_r;
 484         __be64 *hptep;
 485         unsigned long mmu_seq, psize, pte_size;
 486         unsigned long gpa_base, gfn_base;
 487         unsigned long gpa, gfn, hva, pfn, hpa;
 488         struct kvm_memory_slot *memslot;
 489         unsigned long *rmap;
 490         struct revmap_entry *rev;
 491         struct page *page;
 492         long index, ret;
 493         bool is_ci;
 494         bool writing, write_ok;
 495         unsigned int shift;
 496         unsigned long rcbits;
 497         long mmio_update;
 498         pte_t pte, *ptep;
 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 = kvmppc_actual_pgsz(vcpu->arch.pgfault_hpte[0],
 517                                                    r);
 518                         gpa_base = r & HPTE_R_RPN & ~(psize - 1);
 519                         gfn_base = gpa_base >> PAGE_SHIFT;
 520                         gpa = gpa_base | (ea & (psize - 1));
 521                         return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
 522                                                 dsisr & DSISR_ISSTORE);
 523                 }
 524         }
 525         index = vcpu->arch.pgfault_index;
 526         hptep = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
 527         rev = &kvm->arch.hpt.rev[index];
 528         preempt_disable();
 529         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 530                 cpu_relax();
 531         hpte[0] = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
 532         hpte[1] = be64_to_cpu(hptep[1]);
 533         hpte[2] = r = rev->guest_rpte;
 534         unlock_hpte(hptep, hpte[0]);
 535         preempt_enable();
 536
 537         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
 538                 hpte[0] = hpte_new_to_old_v(hpte[0], hpte[1]);
 539                 hpte[1] = hpte_new_to_old_r(hpte[1]);
 540         }
 541         if (hpte[0] != vcpu->arch.pgfault_hpte[0] ||
 542             hpte[1] != vcpu->arch.pgfault_hpte[1])
 543                 return RESUME_GUEST;
 544
 545         /* Translate the logical address and get the page */
 546         psize = kvmppc_actual_pgsz(hpte[0], r);
 547         gpa_base = r & HPTE_R_RPN & ~(psize - 1);
 548         gfn_base = gpa_base >> PAGE_SHIFT;
 549         gpa = gpa_base | (ea & (psize - 1));
 550         gfn = gpa >> PAGE_SHIFT;
 551         memslot = gfn_to_memslot(kvm, gfn);
 552
 553         trace_kvm_page_fault_enter(vcpu, hpte, memslot, ea, dsisr);
 554
 555         /* No memslot means it's an emulated MMIO region */
 556         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
 557                 return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
 558                                               dsisr & DSISR_ISSTORE);
 559
 560         /*
 561          * This should never happen, because of the slot_is_aligned()
 562          * check in kvmppc_do_h_enter().
 563          */
 564         if (gfn_base < memslot->base_gfn)
 565                 return -EFAULT;
 566
 567         /* used to check for invalidations in progress */
 568         mmu_seq = kvm->mmu_notifier_seq;
 569         smp_rmb();
 570
 571         ret = -EFAULT;
 572         page = NULL;
 573         writing = (dsisr & DSISR_ISSTORE) != 0;
 574         /* If writing != 0, then the HPTE must allow writing, if we get here */
 575         write_ok = writing;
 576         hva = gfn_to_hva_memslot(memslot, gfn);
 577
 578         /*
 579          * Do a fast check first, since __gfn_to_pfn_memslot doesn't
 580          * do it with !atomic && !async, which is how we call it.
 581          * We always ask for write permission since the common case
 582          * is that the page is writable.
 583          */
 584         if (get_user_page_fast_only(hva, FOLL_WRITE, &page)) {
 585                 write_ok = true;
 586         } else {
 587                 /* Call KVM generic code to do the slow-path check */
 588                 pfn = __gfn_to_pfn_memslot(memslot, gfn, false, NULL,
 589                                            writing, &write_ok);
 590                 if (is_error_noslot_pfn(pfn))
 591                         return -EFAULT;
 592                 page = NULL;
 593                 if (pfn_valid(pfn)) {
 594                         page = pfn_to_page(pfn);
 595                         if (PageReserved(page))
 596                                 page = NULL;
 597                 }
 598         }
 599
 600         /*
 601          * Read the PTE from the process' radix tree and use that
 602          * so we get the shift and attribute bits.
 603          */
 604         spin_lock(&kvm->mmu_lock);
 605         ptep = find_kvm_host_pte(kvm, mmu_seq, hva, &shift);
 606         pte = __pte(0);
 607         if (ptep)
 608                 pte = READ_ONCE(*ptep);
 609         spin_unlock(&kvm->mmu_lock);
 610         /*
 611          * If the PTE disappeared temporarily due to a THP
 612          * collapse, just return and let the guest try again.
 613          */
 614         if (!pte_present(pte)) {
 615                 if (page)
 616                         put_page(page);
 617                 return RESUME_GUEST;
 618         }
 619         hpa = pte_pfn(pte) << PAGE_SHIFT;
 620         pte_size = PAGE_SIZE;
 621         if (shift)
 622                 pte_size = 1ul << shift;
 623         is_ci = pte_ci(pte);
 624
 625         if (psize > pte_size)
 626                 goto out_put;
 627         if (pte_size > psize)
 628                 hpa |= hva & (pte_size - psize);
 629
 630         /* Check WIMG vs. the actual page we're accessing */
 631         if (!hpte_cache_flags_ok(r, is_ci)) {
 632                 if (is_ci)
 633                         goto out_put;
 634                 /*
 635                  * Allow guest to map emulated device memory as
 636                  * uncacheable, but actually make it cacheable.
 637                  */
 638                 r = (r & ~(HPTE_R_W|HPTE_R_I|HPTE_R_G)) | HPTE_R_M;
 639         }
 640
 641         /*
 642          * Set the HPTE to point to hpa.
 643          * Since the hpa is at PAGE_SIZE granularity, make sure we
 644          * don't mask out lower-order bits if psize < PAGE_SIZE.
 645          */
 646         if (psize < PAGE_SIZE)
 647                 psize = PAGE_SIZE;
 648         r = (r & HPTE_R_KEY_HI) | (r & ~(HPTE_R_PP0 - psize)) | hpa;
 649         if (hpte_is_writable(r) && !write_ok)
 650                 r = hpte_make_readonly(r);
 651         ret = RESUME_GUEST;
 652         preempt_disable();
 653         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 654                 cpu_relax();
 655         hnow_v = be64_to_cpu(hptep[0]);
 656         hnow_r = be64_to_cpu(hptep[1]);
 657         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
 658                 hnow_v = hpte_new_to_old_v(hnow_v, hnow_r);
 659                 hnow_r = hpte_new_to_old_r(hnow_r);
 660         }
 661
 662         /*
 663          * If the HPT is being resized, don't update the HPTE,
 664          * instead let the guest retry after the resize operation is complete.
 665          * The synchronization for mmu_ready test vs. set is provided
 666          * by the HPTE lock.
 667          */
 668         if (!kvm->arch.mmu_ready)
 669                 goto out_unlock;
 670
 671         if ((hnow_v & ~HPTE_V_HVLOCK) != hpte[0] || hnow_r != hpte[1] ||
 672             rev->guest_rpte != hpte[2])
 673                 /* HPTE has been changed under us; let the guest retry */
 674                 goto out_unlock;
 675         hpte[0] = (hpte[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
 676
 677         /* Always put the HPTE in the rmap chain for the page base address */
 678         rmap = &memslot->arch.rmap[gfn_base - memslot->base_gfn];
 679         lock_rmap(rmap);
 680
 681         /* Check if we might have been invalidated; let the guest retry if so */
 682         ret = RESUME_GUEST;
 683         if (mmu_notifier_retry(vcpu->kvm, mmu_seq)) {
 684                 unlock_rmap(rmap);
 685                 goto out_unlock;
 686         }
 687
 688         /* Only set R/C in real HPTE if set in both *rmap and guest_rpte */
 689         rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
 690         r &= rcbits | ~(HPTE_R_R | HPTE_R_C);
 691
 692         if (be64_to_cpu(hptep[0]) & HPTE_V_VALID) {
 693                 /* HPTE was previously valid, so we need to invalidate it */
 694                 unlock_rmap(rmap);
 695                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
 696                 kvmppc_invalidate_hpte(kvm, hptep, index);
 697                 /* don't lose previous R and C bits */
 698                 r |= be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
 699         } else {
 700                 kvmppc_add_revmap_chain(kvm, rev, rmap, index, 0);
 701         }
 702
 703         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
 704                 r = hpte_old_to_new_r(hpte[0], r);
 705                 hpte[0] = hpte_old_to_new_v(hpte[0]);
 706         }
 707         hptep[1] = cpu_to_be64(r);
 708         eieio();
 709         __unlock_hpte(hptep, hpte[0]);
 710         asm volatile("ptesync" : : : "memory");
 711         preempt_enable();
 712         if (page && hpte_is_writable(r))
 713                 set_page_dirty_lock(page);
 714
 715  out_put:
 716         trace_kvm_page_fault_exit(vcpu, hpte, ret);
 717
 718         if (page)
 719                 put_page(page);
 720         return ret;
 721
 722  out_unlock:
 723         __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
 724         preempt_enable();
 725         goto out_put;
 726 }
 727
 728 void kvmppc_rmap_reset(struct kvm *kvm)
 729 {
 730         struct kvm_memslots *slots;
 731         struct kvm_memory_slot *memslot;
 732         int srcu_idx;
 733
 734         srcu_idx = srcu_read_lock(&kvm->srcu);
 735         slots = kvm_memslots(kvm);
 736         kvm_for_each_memslot(memslot, slots) {
 737                 /* Mutual exclusion with kvm_unmap_hva_range etc. */
 738                 spin_lock(&kvm->mmu_lock);
 739                 /*
 740                  * This assumes it is acceptable to lose reference and
 741                  * change bits across a reset.
 742                  */
 743                 memset(memslot->arch.rmap, 0,
 744                        memslot->npages * sizeof(*memslot->arch.rmap));
 745                 spin_unlock(&kvm->mmu_lock);
 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                               struct kvm_memory_slot *memslot,
 798                               unsigned long *rmapp, unsigned long gfn)
 799 {
 800         __be64 *hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
 801         struct revmap_entry *rev = kvm->arch.hpt.rev;
 802         unsigned long j, h;
 803         unsigned long ptel, psize, rcbits;
 804
 805         j = rev[i].forw;
 806         if (j == i) {
 807                 /* chain is now empty */
 808                 *rmapp &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
 809         } else {
 810                 /* remove i from chain */
 811                 h = rev[i].back;
 812                 rev[h].forw = j;
 813                 rev[j].back = h;
 814                 rev[i].forw = rev[i].back = i;
 815                 *rmapp = (*rmapp & ~KVMPPC_RMAP_INDEX) | j;
 816         }
 817
 818         /* Now check and modify the HPTE */
 819         ptel = rev[i].guest_rpte;
 820         psize = kvmppc_actual_pgsz(be64_to_cpu(hptep[0]), ptel);
 821         if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
 822             hpte_rpn(ptel, psize) == gfn) {
 823                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
 824                 kvmppc_invalidate_hpte(kvm, hptep, i);
 825                 hptep[1] &= ~cpu_to_be64(HPTE_R_KEY_HI | HPTE_R_KEY_LO);
 826                 /* Harvest R and C */
 827                 rcbits = be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
 828                 *rmapp |= rcbits << KVMPPC_RMAP_RC_SHIFT;
 829                 if ((rcbits & HPTE_R_C) && memslot->dirty_bitmap)
 830                         kvmppc_update_dirty_map(memslot, gfn, psize);
 831                 if (rcbits & ~rev[i].guest_rpte) {
 832                         rev[i].guest_rpte = ptel | rcbits;
 833                         note_hpte_modification(kvm, &rev[i]);
 834                 }
 835         }
 836 }
 837
 838 static int kvm_unmap_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
 839                            unsigned long gfn)
 840 {
 841         unsigned long i;
 842         __be64 *hptep;
 843         unsigned long *rmapp;
 844
 845         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
 846         for (;;) {
 847                 lock_rmap(rmapp);
 848                 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 849                         unlock_rmap(rmapp);
 850                         break;
 851                 }
 852
 853                 /*
 854                  * To avoid an ABBA deadlock with the HPTE lock bit,
 855                  * we can't spin on the HPTE lock while holding the
 856                  * rmap chain lock.
 857                  */
 858                 i = *rmapp & KVMPPC_RMAP_INDEX;
 859                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
 860                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 861                         /* unlock rmap before spinning on the HPTE lock */
 862                         unlock_rmap(rmapp);
 863                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
 864                                 cpu_relax();
 865                         continue;
 866                 }
 867
 868                 kvmppc_unmap_hpte(kvm, i, memslot, rmapp, gfn);
 869                 unlock_rmap(rmapp);
 870                 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
 871         }
 872         return 0;
 873 }
 874
 875 int kvm_unmap_hva_range_hv(struct kvm *kvm, unsigned long start, unsigned long end)
 876 {
 877         hva_handler_fn handler;
 878
 879         handler = kvm_is_radix(kvm) ? kvm_unmap_radix : kvm_unmap_rmapp;
 880         kvm_handle_hva_range(kvm, start, end, handler);
 881         return 0;
 882 }
 883
 884 void kvmppc_core_flush_memslot_hv(struct kvm *kvm,
 885                                   struct kvm_memory_slot *memslot)
 886 {
 887         unsigned long gfn;
 888         unsigned long n;
 889         unsigned long *rmapp;
 890
 891         gfn = memslot->base_gfn;
 892         rmapp = memslot->arch.rmap;
 893         if (kvm_is_radix(kvm)) {
 894                 kvmppc_radix_flush_memslot(kvm, memslot);
 895                 return;
 896         }
 897
 898         for (n = memslot->npages; n; --n, ++gfn) {
 899                 /*
 900                  * Testing the present bit without locking is OK because
 901                  * the memslot has been marked invalid already, and hence
 902                  * no new HPTEs referencing this page can be created,
 903                  * thus the present bit can't go from 0 to 1.
 904                  */
 905                 if (*rmapp & KVMPPC_RMAP_PRESENT)
 906                         kvm_unmap_rmapp(kvm, memslot, gfn);
 907                 ++rmapp;
 908         }
 909 }
 910
 911 static int kvm_age_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
 912                          unsigned long gfn)
 913 {
 914         struct revmap_entry *rev = kvm->arch.hpt.rev;
 915         unsigned long head, i, j;
 916         __be64 *hptep;
 917         int ret = 0;
 918         unsigned long *rmapp;
 919
 920         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
 921  retry:
 922         lock_rmap(rmapp);
 923         if (*rmapp & KVMPPC_RMAP_REFERENCED) {
 924                 *rmapp &= ~KVMPPC_RMAP_REFERENCED;
 925                 ret = 1;
 926         }
 927         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 928                 unlock_rmap(rmapp);
 929                 return ret;
 930         }
 931
 932         i = head = *rmapp & KVMPPC_RMAP_INDEX;
 933         do {
 934                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
 935                 j = rev[i].forw;
 936
 937                 /* If this HPTE isn't referenced, ignore it */
 938                 if (!(be64_to_cpu(hptep[1]) & HPTE_R_R))
 939                         continue;
 940
 941                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 942                         /* unlock rmap before spinning on the HPTE lock */
 943                         unlock_rmap(rmapp);
 944                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
 945                                 cpu_relax();
 946                         goto retry;
 947                 }
 948
 949                 /* Now check and modify the HPTE */
 950                 if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
 951                     (be64_to_cpu(hptep[1]) & HPTE_R_R)) {
 952                         kvmppc_clear_ref_hpte(kvm, hptep, i);
 953                         if (!(rev[i].guest_rpte & HPTE_R_R)) {
 954                                 rev[i].guest_rpte |= HPTE_R_R;
 955                                 note_hpte_modification(kvm, &rev[i]);
 956                         }
 957                         ret = 1;
 958                 }
 959                 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
 960         } while ((i = j) != head);
 961
 962         unlock_rmap(rmapp);
 963         return ret;
 964 }
 965
 966 int kvm_age_hva_hv(struct kvm *kvm, unsigned long start, unsigned long end)
 967 {
 968         hva_handler_fn handler;
 969
 970         handler = kvm_is_radix(kvm) ? kvm_age_radix : kvm_age_rmapp;
 971         return kvm_handle_hva_range(kvm, start, end, handler);
 972 }
 973
 974 static int kvm_test_age_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
 975                               unsigned long gfn)
 976 {
 977         struct revmap_entry *rev = kvm->arch.hpt.rev;
 978         unsigned long head, i, j;
 979         unsigned long *hp;
 980         int ret = 1;
 981         unsigned long *rmapp;
 982
 983         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
 984         if (*rmapp & KVMPPC_RMAP_REFERENCED)
 985                 return 1;
 986
 987         lock_rmap(rmapp);
 988         if (*rmapp & KVMPPC_RMAP_REFERENCED)
 989                 goto out;
 990
 991         if (*rmapp & KVMPPC_RMAP_PRESENT) {
 992                 i = head = *rmapp & KVMPPC_RMAP_INDEX;
 993                 do {
 994                         hp = (unsigned long *)(kvm->arch.hpt.virt + (i << 4));
 995                         j = rev[i].forw;
 996                         if (be64_to_cpu(hp[1]) & HPTE_R_R)
 997                                 goto out;
 998                 } while ((i = j) != head);
 999         }
1000         ret = 0;
1001
1002  out:
1003         unlock_rmap(rmapp);
1004         return ret;
1005 }
1006
1007 int kvm_test_age_hva_hv(struct kvm *kvm, unsigned long hva)
1008 {
1009         hva_handler_fn handler;
1010
1011         handler = kvm_is_radix(kvm) ? kvm_test_age_radix : kvm_test_age_rmapp;
1012         return kvm_handle_hva(kvm, hva, handler);
1013 }
1014
1015 void kvm_set_spte_hva_hv(struct kvm *kvm, unsigned long hva, pte_t pte)
1016 {
1017         hva_handler_fn handler;
1018
1019         handler = kvm_is_radix(kvm) ? kvm_unmap_radix : kvm_unmap_rmapp;
1020         kvm_handle_hva(kvm, hva, handler);
1021 }
1022
1023 static int vcpus_running(struct kvm *kvm)
1024 {
1025         return atomic_read(&kvm->arch.vcpus_running) != 0;
1026 }
1027
1028 /*
1029  * Returns the number of system pages that are dirty.
1030  * This can be more than 1 if we find a huge-page HPTE.
1031  */
1032 static int kvm_test_clear_dirty_npages(struct kvm *kvm, unsigned long *rmapp)
1033 {
1034         struct revmap_entry *rev = kvm->arch.hpt.rev;
1035         unsigned long head, i, j;
1036         unsigned long n;
1037         unsigned long v, r;
1038         __be64 *hptep;
1039         int npages_dirty = 0;
1040
1041  retry:
1042         lock_rmap(rmapp);
1043         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
1044                 unlock_rmap(rmapp);
1045                 return npages_dirty;
1046         }
1047
1048         i = head = *rmapp & KVMPPC_RMAP_INDEX;
1049         do {
1050                 unsigned long hptep1;
1051                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
1052                 j = rev[i].forw;
1053
1054                 /*
1055                  * Checking the C (changed) bit here is racy since there
1056                  * is no guarantee about when the hardware writes it back.
1057                  * If the HPTE is not writable then it is stable since the
1058                  * page can't be written to, and we would have done a tlbie
1059                  * (which forces the hardware to complete any writeback)
1060                  * when making the HPTE read-only.
1061                  * If vcpus are running then this call is racy anyway
1062                  * since the page could get dirtied subsequently, so we
1063                  * expect there to be a further call which would pick up
1064                  * any delayed C bit writeback.
1065                  * Otherwise we need to do the tlbie even if C==0 in
1066                  * order to pick up any delayed writeback of C.
1067                  */
1068                 hptep1 = be64_to_cpu(hptep[1]);
1069                 if (!(hptep1 & HPTE_R_C) &&
1070                     (!hpte_is_writable(hptep1) || vcpus_running(kvm)))
1071                         continue;
1072
1073                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
1074                         /* unlock rmap before spinning on the HPTE lock */
1075                         unlock_rmap(rmapp);
1076                         while (hptep[0] & cpu_to_be64(HPTE_V_HVLOCK))
1077                                 cpu_relax();
1078                         goto retry;
1079                 }
1080
1081                 /* Now check and modify the HPTE */
1082                 if (!(hptep[0] & cpu_to_be64(HPTE_V_VALID))) {
1083                         __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
1084                         continue;
1085                 }
1086
1087                 /* need to make it temporarily absent so C is stable */
1088                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
1089                 kvmppc_invalidate_hpte(kvm, hptep, i);
1090                 v = be64_to_cpu(hptep[0]);
1091                 r = be64_to_cpu(hptep[1]);
1092                 if (r & HPTE_R_C) {
1093                         hptep[1] = cpu_to_be64(r & ~HPTE_R_C);
1094                         if (!(rev[i].guest_rpte & HPTE_R_C)) {
1095                                 rev[i].guest_rpte |= HPTE_R_C;
1096                                 note_hpte_modification(kvm, &rev[i]);
1097                         }
1098                         n = kvmppc_actual_pgsz(v, r);
1099                         n = (n + PAGE_SIZE - 1) >> PAGE_SHIFT;
1100                         if (n > npages_dirty)
1101                                 npages_dirty = n;
1102                         eieio();
1103                 }
1104                 v &= ~HPTE_V_ABSENT;
1105                 v |= HPTE_V_VALID;
1106                 __unlock_hpte(hptep, v);
1107         } while ((i = j) != head);
1108
1109         unlock_rmap(rmapp);
1110         return npages_dirty;
1111 }
1112
1113 void kvmppc_harvest_vpa_dirty(struct kvmppc_vpa *vpa,
1114                               struct kvm_memory_slot *memslot,
1115                               unsigned long *map)
1116 {
1117         unsigned long gfn;
1118
1119         if (!vpa->dirty || !vpa->pinned_addr)
1120                 return;
1121         gfn = vpa->gpa >> PAGE_SHIFT;
1122         if (gfn < memslot->base_gfn ||
1123             gfn >= memslot->base_gfn + memslot->npages)
1124                 return;
1125
1126         vpa->dirty = false;
1127         if (map)
1128                 __set_bit_le(gfn - memslot->base_gfn, map);
1129 }
1130
1131 long kvmppc_hv_get_dirty_log_hpt(struct kvm *kvm,
1132                         struct kvm_memory_slot *memslot, unsigned long *map)
1133 {
1134         unsigned long i;
1135         unsigned long *rmapp;
1136
1137         preempt_disable();
1138         rmapp = memslot->arch.rmap;
1139         for (i = 0; i < memslot->npages; ++i) {
1140                 int npages = kvm_test_clear_dirty_npages(kvm, rmapp);
1141                 /*
1142                  * Note that if npages > 0 then i must be a multiple of npages,
1143                  * since we always put huge-page HPTEs in the rmap chain
1144                  * corresponding to their page base address.
1145                  */
1146                 if (npages)
1147                         set_dirty_bits(map, i, npages);
1148                 ++rmapp;
1149         }
1150         preempt_enable();
1151         return 0;
1152 }
1153
1154 void *kvmppc_pin_guest_page(struct kvm *kvm, unsigned long gpa,
1155                             unsigned long *nb_ret)
1156 {
1157         struct kvm_memory_slot *memslot;
1158         unsigned long gfn = gpa >> PAGE_SHIFT;
1159         struct page *page, *pages[1];
1160         int npages;
1161         unsigned long hva, offset;
1162         int srcu_idx;
1163
1164         srcu_idx = srcu_read_lock(&kvm->srcu);
1165         memslot = gfn_to_memslot(kvm, gfn);
1166         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
1167                 goto err;
1168         hva = gfn_to_hva_memslot(memslot, gfn);
1169         npages = get_user_pages_fast(hva, 1, FOLL_WRITE, pages);
1170         if (npages < 1)
1171                 goto err;
1172         page = pages[0];
1173         srcu_read_unlock(&kvm->srcu, srcu_idx);
1174
1175         offset = gpa & (PAGE_SIZE - 1);
1176         if (nb_ret)
1177                 *nb_ret = PAGE_SIZE - offset;
1178         return page_address(page) + offset;
1179
1180  err:
1181         srcu_read_unlock(&kvm->srcu, srcu_idx);
1182         return NULL;
1183 }
1184
1185 void kvmppc_unpin_guest_page(struct kvm *kvm, void *va, unsigned long gpa,
1186                              bool dirty)
1187 {
1188         struct page *page = virt_to_page(va);
1189         struct kvm_memory_slot *memslot;
1190         unsigned long gfn;
1191         int srcu_idx;
1192
1193         put_page(page);
1194
1195         if (!dirty)
1196                 return;
1197
1198         /* We need to mark this page dirty in the memslot dirty_bitmap, if any */
1199         gfn = gpa >> PAGE_SHIFT;
1200         srcu_idx = srcu_read_lock(&kvm->srcu);
1201         memslot = gfn_to_memslot(kvm, gfn);
1202         if (memslot && memslot->dirty_bitmap)
1203                 set_bit_le(gfn - memslot->base_gfn, memslot->dirty_bitmap);
1204         srcu_read_unlock(&kvm->srcu, srcu_idx);
1205 }
1206
1207 /*
1208  * HPT resizing
1209  */
1210 static int resize_hpt_allocate(struct kvm_resize_hpt *resize)
1211 {
1212         int rc;
1213
1214         rc = kvmppc_allocate_hpt(&resize->hpt, resize->order);
1215         if (rc < 0)
1216                 return rc;
1217
1218         resize_hpt_debug(resize, "resize_hpt_allocate(): HPT @ 0x%lx\n",
1219                          resize->hpt.virt);
1220
1221         return 0;
1222 }
1223
1224 static unsigned long resize_hpt_rehash_hpte(struct kvm_resize_hpt *resize,
1225                                             unsigned long idx)
1226 {
1227         struct kvm *kvm = resize->kvm;
1228         struct kvm_hpt_info *old = &kvm->arch.hpt;
1229         struct kvm_hpt_info *new = &resize->hpt;
1230         unsigned long old_hash_mask = (1ULL << (old->order - 7)) - 1;
1231         unsigned long new_hash_mask = (1ULL << (new->order - 7)) - 1;
1232         __be64 *hptep, *new_hptep;
1233         unsigned long vpte, rpte, guest_rpte;
1234         int ret;
1235         struct revmap_entry *rev;
1236         unsigned long apsize, avpn, pteg, hash;
1237         unsigned long new_idx, new_pteg, replace_vpte;
1238         int pshift;
1239
1240         hptep = (__be64 *)(old->virt + (idx << 4));
1241
1242         /* Guest is stopped, so new HPTEs can't be added or faulted
1243          * in, only unmapped or altered by host actions.  So, it's
1244          * safe to check this before we take the HPTE lock */
1245         vpte = be64_to_cpu(hptep[0]);
1246         if (!(vpte & HPTE_V_VALID) && !(vpte & HPTE_V_ABSENT))
1247                 return 0; /* nothing to do */
1248
1249         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
1250                 cpu_relax();
1251
1252         vpte = be64_to_cpu(hptep[0]);
1253
1254         ret = 0;
1255         if (!(vpte & HPTE_V_VALID) && !(vpte & HPTE_V_ABSENT))
1256                 /* Nothing to do */
1257                 goto out;
1258
1259         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1260                 rpte = be64_to_cpu(hptep[1]);
1261                 vpte = hpte_new_to_old_v(vpte, rpte);
1262         }
1263
1264         /* Unmap */
1265         rev = &old->rev[idx];
1266         guest_rpte = rev->guest_rpte;
1267
1268         ret = -EIO;
1269         apsize = kvmppc_actual_pgsz(vpte, guest_rpte);
1270         if (!apsize)
1271                 goto out;
1272
1273         if (vpte & HPTE_V_VALID) {
1274                 unsigned long gfn = hpte_rpn(guest_rpte, apsize);
1275                 int srcu_idx = srcu_read_lock(&kvm->srcu);
1276                 struct kvm_memory_slot *memslot =
1277                         __gfn_to_memslot(kvm_memslots(kvm), gfn);
1278
1279                 if (memslot) {
1280                         unsigned long *rmapp;
1281                         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1282
1283                         lock_rmap(rmapp);
1284                         kvmppc_unmap_hpte(kvm, idx, memslot, rmapp, gfn);
1285                         unlock_rmap(rmapp);
1286                 }
1287
1288                 srcu_read_unlock(&kvm->srcu, srcu_idx);
1289         }
1290
1291         /* Reload PTE after unmap */
1292         vpte = be64_to_cpu(hptep[0]);
1293         BUG_ON(vpte & HPTE_V_VALID);
1294         BUG_ON(!(vpte & HPTE_V_ABSENT));
1295
1296         ret = 0;
1297         if (!(vpte & HPTE_V_BOLTED))
1298                 goto out;
1299
1300         rpte = be64_to_cpu(hptep[1]);
1301
1302         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1303                 vpte = hpte_new_to_old_v(vpte, rpte);
1304                 rpte = hpte_new_to_old_r(rpte);
1305         }
1306
1307         pshift = kvmppc_hpte_base_page_shift(vpte, rpte);
1308         avpn = HPTE_V_AVPN_VAL(vpte) & ~(((1ul << pshift) - 1) >> 23);
1309         pteg = idx / HPTES_PER_GROUP;
1310         if (vpte & HPTE_V_SECONDARY)
1311                 pteg = ~pteg;
1312
1313         if (!(vpte & HPTE_V_1TB_SEG)) {
1314                 unsigned long offset, vsid;
1315
1316                 /* We only have 28 - 23 bits of offset in avpn */
1317                 offset = (avpn & 0x1f) << 23;
1318                 vsid = avpn >> 5;
1319                 /* We can find more bits from the pteg value */
1320                 if (pshift < 23)
1321                         offset |= ((vsid ^ pteg) & old_hash_mask) << pshift;
1322
1323                 hash = vsid ^ (offset >> pshift);
1324         } else {
1325                 unsigned long offset, vsid;
1326
1327                 /* We only have 40 - 23 bits of seg_off in avpn */
1328                 offset = (avpn & 0x1ffff) << 23;
1329                 vsid = avpn >> 17;
1330                 if (pshift < 23)
1331                         offset |= ((vsid ^ (vsid << 25) ^ pteg) & old_hash_mask) << pshift;
1332
1333                 hash = vsid ^ (vsid << 25) ^ (offset >> pshift);
1334         }
1335
1336         new_pteg = hash & new_hash_mask;
1337         if (vpte & HPTE_V_SECONDARY)
1338                 new_pteg = ~hash & new_hash_mask;
1339
1340         new_idx = new_pteg * HPTES_PER_GROUP + (idx % HPTES_PER_GROUP);
1341         new_hptep = (__be64 *)(new->virt + (new_idx << 4));
1342
1343         replace_vpte = be64_to_cpu(new_hptep[0]);
1344         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1345                 unsigned long replace_rpte = be64_to_cpu(new_hptep[1]);
1346                 replace_vpte = hpte_new_to_old_v(replace_vpte, replace_rpte);
1347         }
1348
1349         if (replace_vpte & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1350                 BUG_ON(new->order >= old->order);
1351
1352                 if (replace_vpte & HPTE_V_BOLTED) {
1353                         if (vpte & HPTE_V_BOLTED)
1354                                 /* Bolted collision, nothing we can do */
1355                                 ret = -ENOSPC;
1356                         /* Discard the new HPTE */
1357                         goto out;
1358                 }
1359
1360                 /* Discard the previous HPTE */
1361         }
1362
1363         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1364                 rpte = hpte_old_to_new_r(vpte, rpte);
1365                 vpte = hpte_old_to_new_v(vpte);
1366         }
1367
1368         new_hptep[1] = cpu_to_be64(rpte);
1369         new->rev[new_idx].guest_rpte = guest_rpte;
1370         /* No need for a barrier, since new HPT isn't active */
1371         new_hptep[0] = cpu_to_be64(vpte);
1372         unlock_hpte(new_hptep, vpte);
1373
1374 out:
1375         unlock_hpte(hptep, vpte);
1376         return ret;
1377 }
1378
1379 static int resize_hpt_rehash(struct kvm_resize_hpt *resize)
1380 {
1381         struct kvm *kvm = resize->kvm;
1382         unsigned  long i;
1383         int rc;
1384
1385         for (i = 0; i < kvmppc_hpt_npte(&kvm->arch.hpt); i++) {
1386                 rc = resize_hpt_rehash_hpte(resize, i);
1387                 if (rc != 0)
1388                         return rc;
1389         }
1390
1391         return 0;
1392 }
1393
1394 static void resize_hpt_pivot(struct kvm_resize_hpt *resize)
1395 {
1396         struct kvm *kvm = resize->kvm;
1397         struct kvm_hpt_info hpt_tmp;
1398
1399         /* Exchange the pending tables in the resize structure with
1400          * the active tables */
1401
1402         resize_hpt_debug(resize, "resize_hpt_pivot()\n");
1403
1404         spin_lock(&kvm->mmu_lock);
1405         asm volatile("ptesync" : : : "memory");
1406
1407         hpt_tmp = kvm->arch.hpt;
1408         kvmppc_set_hpt(kvm, &resize->hpt);
1409         resize->hpt = hpt_tmp;
1410
1411         spin_unlock(&kvm->mmu_lock);
1412
1413         synchronize_srcu_expedited(&kvm->srcu);
1414
1415         if (cpu_has_feature(CPU_FTR_ARCH_300))
1416                 kvmppc_setup_partition_table(kvm);
1417
1418         resize_hpt_debug(resize, "resize_hpt_pivot() done\n");
1419 }
1420
1421 static void resize_hpt_release(struct kvm *kvm, struct kvm_resize_hpt *resize)
1422 {
1423         if (WARN_ON(!mutex_is_locked(&kvm->arch.mmu_setup_lock)))
1424                 return;
1425
1426         if (!resize)
1427                 return;
1428
1429         if (resize->error != -EBUSY) {
1430                 if (resize->hpt.virt)
1431                         kvmppc_free_hpt(&resize->hpt);
1432                 kfree(resize);
1433         }
1434
1435         if (kvm->arch.resize_hpt == resize)
1436                 kvm->arch.resize_hpt = NULL;
1437 }
1438
1439 static void resize_hpt_prepare_work(struct work_struct *work)
1440 {
1441         struct kvm_resize_hpt *resize = container_of(work,
1442                                                      struct kvm_resize_hpt,
1443                                                      work);
1444         struct kvm *kvm = resize->kvm;
1445         int err = 0;
1446
1447         if (WARN_ON(resize->error != -EBUSY))
1448                 return;
1449
1450         mutex_lock(&kvm->arch.mmu_setup_lock);
1451
1452         /* Request is still current? */
1453         if (kvm->arch.resize_hpt == resize) {
1454                 /* We may request large allocations here:
1455                  * do not sleep with kvm->arch.mmu_setup_lock held for a while.
1456                  */
1457                 mutex_unlock(&kvm->arch.mmu_setup_lock);
1458
1459                 resize_hpt_debug(resize, "resize_hpt_prepare_work(): order = %d\n",
1460                                  resize->order);
1461
1462                 err = resize_hpt_allocate(resize);
1463
1464                 /* We have strict assumption about -EBUSY
1465                  * when preparing for HPT resize.
1466                  */
1467                 if (WARN_ON(err == -EBUSY))
1468                         err = -EINPROGRESS;
1469
1470                 mutex_lock(&kvm->arch.mmu_setup_lock);
1471                 /* It is possible that kvm->arch.resize_hpt != resize
1472                  * after we grab kvm->arch.mmu_setup_lock again.
1473                  */
1474         }
1475
1476         resize->error = err;
1477
1478         if (kvm->arch.resize_hpt != resize)
1479                 resize_hpt_release(kvm, resize);
1480
1481         mutex_unlock(&kvm->arch.mmu_setup_lock);
1482 }
1483
1484 long kvm_vm_ioctl_resize_hpt_prepare(struct kvm *kvm,
1485                                      struct kvm_ppc_resize_hpt *rhpt)
1486 {
1487         unsigned long flags = rhpt->flags;
1488         unsigned long shift = rhpt->shift;
1489         struct kvm_resize_hpt *resize;
1490         int ret;
1491
1492         if (flags != 0 || kvm_is_radix(kvm))
1493                 return -EINVAL;
1494
1495         if (shift && ((shift < 18) || (shift > 46)))
1496                 return -EINVAL;
1497
1498         mutex_lock(&kvm->arch.mmu_setup_lock);
1499
1500         resize = kvm->arch.resize_hpt;
1501
1502         if (resize) {
1503                 if (resize->order == shift) {
1504                         /* Suitable resize in progress? */
1505                         ret = resize->error;
1506                         if (ret == -EBUSY)
1507                                 ret = 100; /* estimated time in ms */
1508                         else if (ret)
1509                                 resize_hpt_release(kvm, resize);
1510
1511                         goto out;
1512                 }
1513
1514                 /* not suitable, cancel it */
1515                 resize_hpt_release(kvm, resize);
1516         }
1517
1518         ret = 0;
1519         if (!shift)
1520                 goto out; /* nothing to do */
1521
1522         /* start new resize */
1523
1524         resize = kzalloc(sizeof(*resize), GFP_KERNEL);
1525         if (!resize) {
1526                 ret = -ENOMEM;
1527                 goto out;
1528         }
1529
1530         resize->error = -EBUSY;
1531         resize->order = shift;
1532         resize->kvm = kvm;
1533         INIT_WORK(&resize->work, resize_hpt_prepare_work);
1534         kvm->arch.resize_hpt = resize;
1535
1536         schedule_work(&resize->work);
1537
1538         ret = 100; /* estimated time in ms */
1539
1540 out:
1541         mutex_unlock(&kvm->arch.mmu_setup_lock);
1542         return ret;
1543 }
1544
1545 static void resize_hpt_boot_vcpu(void *opaque)
1546 {
1547         /* Nothing to do, just force a KVM exit */
1548 }
1549
1550 long kvm_vm_ioctl_resize_hpt_commit(struct kvm *kvm,
1551                                     struct kvm_ppc_resize_hpt *rhpt)
1552 {
1553         unsigned long flags = rhpt->flags;
1554         unsigned long shift = rhpt->shift;
1555         struct kvm_resize_hpt *resize;
1556         long ret;
1557
1558         if (flags != 0 || kvm_is_radix(kvm))
1559                 return -EINVAL;
1560
1561         if (shift && ((shift < 18) || (shift > 46)))
1562                 return -EINVAL;
1563
1564         mutex_lock(&kvm->arch.mmu_setup_lock);
1565
1566         resize = kvm->arch.resize_hpt;
1567
1568         /* This shouldn't be possible */
1569         ret = -EIO;
1570         if (WARN_ON(!kvm->arch.mmu_ready))
1571                 goto out_no_hpt;
1572
1573         /* Stop VCPUs from running while we mess with the HPT */
1574         kvm->arch.mmu_ready = 0;
1575         smp_mb();
1576
1577         /* Boot all CPUs out of the guest so they re-read
1578          * mmu_ready */
1579         on_each_cpu(resize_hpt_boot_vcpu, NULL, 1);
1580
1581         ret = -ENXIO;
1582         if (!resize || (resize->order != shift))
1583                 goto out;
1584
1585         ret = resize->error;
1586         if (ret)
1587                 goto out;
1588
1589         ret = resize_hpt_rehash(resize);
1590         if (ret)
1591                 goto out;
1592
1593         resize_hpt_pivot(resize);
1594
1595 out:
1596         /* Let VCPUs run again */
1597         kvm->arch.mmu_ready = 1;
1598         smp_mb();
1599 out_no_hpt:
1600         resize_hpt_release(kvm, resize);
1601         mutex_unlock(&kvm->arch.mmu_setup_lock);
1602         return ret;
1603 }
1604
1605 /*
1606  * Functions for reading and writing the hash table via reads and
1607  * writes on a file descriptor.
1608  *
1609  * Reads return the guest view of the hash table, which has to be
1610  * pieced together from the real hash table and the guest_rpte
1611  * values in the revmap array.
1612  *
1613  * On writes, each HPTE written is considered in turn, and if it
1614  * is valid, it is written to the HPT as if an H_ENTER with the
1615  * exact flag set was done.  When the invalid count is non-zero
1616  * in the header written to the stream, the kernel will make
1617  * sure that that many HPTEs are invalid, and invalidate them
1618  * if not.
1619  */
1620
1621 struct kvm_htab_ctx {
1622         unsigned long   index;
1623         unsigned long   flags;
1624         struct kvm      *kvm;
1625         int             first_pass;
1626 };
1627
1628 #define HPTE_SIZE       (2 * sizeof(unsigned long))
1629
1630 /*
1631  * Returns 1 if this HPT entry has been modified or has pending
1632  * R/C bit changes.
1633  */
1634 static int hpte_dirty(struct revmap_entry *revp, __be64 *hptp)
1635 {
1636         unsigned long rcbits_unset;
1637
1638         if (revp->guest_rpte & HPTE_GR_MODIFIED)
1639                 return 1;
1640
1641         /* Also need to consider changes in reference and changed bits */
1642         rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1643         if ((be64_to_cpu(hptp[0]) & HPTE_V_VALID) &&
1644             (be64_to_cpu(hptp[1]) & rcbits_unset))
1645                 return 1;
1646
1647         return 0;
1648 }
1649
1650 static long record_hpte(unsigned long flags, __be64 *hptp,
1651                         unsigned long *hpte, struct revmap_entry *revp,
1652                         int want_valid, int first_pass)
1653 {
1654         unsigned long v, r, hr;
1655         unsigned long rcbits_unset;
1656         int ok = 1;
1657         int valid, dirty;
1658
1659         /* Unmodified entries are uninteresting except on the first pass */
1660         dirty = hpte_dirty(revp, hptp);
1661         if (!first_pass && !dirty)
1662                 return 0;
1663
1664         valid = 0;
1665         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1666                 valid = 1;
1667                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) &&
1668                     !(be64_to_cpu(hptp[0]) & HPTE_V_BOLTED))
1669                         valid = 0;
1670         }
1671         if (valid != want_valid)
1672                 return 0;
1673
1674         v = r = 0;
1675         if (valid || dirty) {
1676                 /* lock the HPTE so it's stable and read it */
1677                 preempt_disable();
1678                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
1679                         cpu_relax();
1680                 v = be64_to_cpu(hptp[0]);
1681                 hr = be64_to_cpu(hptp[1]);
1682                 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1683                         v = hpte_new_to_old_v(v, hr);
1684                         hr = hpte_new_to_old_r(hr);
1685                 }
1686
1687                 /* re-evaluate valid and dirty from synchronized HPTE value */
1688                 valid = !!(v & HPTE_V_VALID);
1689                 dirty = !!(revp->guest_rpte & HPTE_GR_MODIFIED);
1690
1691                 /* Harvest R and C into guest view if necessary */
1692                 rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1693                 if (valid && (rcbits_unset & hr)) {
1694                         revp->guest_rpte |= (hr &
1695                                 (HPTE_R_R | HPTE_R_C)) | HPTE_GR_MODIFIED;
1696                         dirty = 1;
1697                 }
1698
1699                 if (v & HPTE_V_ABSENT) {
1700                         v &= ~HPTE_V_ABSENT;
1701                         v |= HPTE_V_VALID;
1702                         valid = 1;
1703                 }
1704                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) && !(v & HPTE_V_BOLTED))
1705                         valid = 0;
1706
1707                 r = revp->guest_rpte;
1708                 /* only clear modified if this is the right sort of entry */
1709                 if (valid == want_valid && dirty) {
1710                         r &= ~HPTE_GR_MODIFIED;
1711                         revp->guest_rpte = r;
1712                 }
1713                 unlock_hpte(hptp, be64_to_cpu(hptp[0]));
1714                 preempt_enable();
1715                 if (!(valid == want_valid && (first_pass || dirty)))
1716                         ok = 0;
1717         }
1718         hpte[0] = cpu_to_be64(v);
1719         hpte[1] = cpu_to_be64(r);
1720         return ok;
1721 }
1722
1723 static ssize_t kvm_htab_read(struct file *file, char __user *buf,
1724                              size_t count, loff_t *ppos)
1725 {
1726         struct kvm_htab_ctx *ctx = file->private_data;
1727         struct kvm *kvm = ctx->kvm;
1728         struct kvm_get_htab_header hdr;
1729         __be64 *hptp;
1730         struct revmap_entry *revp;
1731         unsigned long i, nb, nw;
1732         unsigned long __user *lbuf;
1733         struct kvm_get_htab_header __user *hptr;
1734         unsigned long flags;
1735         int first_pass;
1736         unsigned long hpte[2];
1737
1738         if (!access_ok(buf, count))
1739                 return -EFAULT;
1740         if (kvm_is_radix(kvm))
1741                 return 0;
1742
1743         first_pass = ctx->first_pass;
1744         flags = ctx->flags;
1745
1746         i = ctx->index;
1747         hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
1748         revp = kvm->arch.hpt.rev + i;
1749         lbuf = (unsigned long __user *)buf;
1750
1751         nb = 0;
1752         while (nb + sizeof(hdr) + HPTE_SIZE < count) {
1753                 /* Initialize header */
1754                 hptr = (struct kvm_get_htab_header __user *)buf;
1755                 hdr.n_valid = 0;
1756                 hdr.n_invalid = 0;
1757                 nw = nb;
1758                 nb += sizeof(hdr);
1759                 lbuf = (unsigned long __user *)(buf + sizeof(hdr));
1760
1761                 /* Skip uninteresting entries, i.e. clean on not-first pass */
1762                 if (!first_pass) {
1763                         while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1764                                !hpte_dirty(revp, hptp)) {
1765                                 ++i;
1766                                 hptp += 2;
1767                                 ++revp;
1768                         }
1769                 }
1770                 hdr.index = i;
1771
1772                 /* Grab a series of valid entries */
1773                 while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1774                        hdr.n_valid < 0xffff &&
1775                        nb + HPTE_SIZE < count &&
1776                        record_hpte(flags, hptp, hpte, revp, 1, first_pass)) {
1777                         /* valid entry, write it out */
1778                         ++hdr.n_valid;
1779                         if (__put_user(hpte[0], lbuf) ||
1780                             __put_user(hpte[1], lbuf + 1))
1781                                 return -EFAULT;
1782                         nb += HPTE_SIZE;
1783                         lbuf += 2;
1784                         ++i;
1785                         hptp += 2;
1786                         ++revp;
1787                 }
1788                 /* Now skip invalid entries while we can */
1789                 while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1790                        hdr.n_invalid < 0xffff &&
1791                        record_hpte(flags, hptp, hpte, revp, 0, first_pass)) {
1792                         /* found an invalid entry */
1793                         ++hdr.n_invalid;
1794                         ++i;
1795                         hptp += 2;
1796                         ++revp;
1797                 }
1798
1799                 if (hdr.n_valid || hdr.n_invalid) {
1800                         /* write back the header */
1801                         if (__copy_to_user(hptr, &hdr, sizeof(hdr)))
1802                                 return -EFAULT;
1803                         nw = nb;
1804                         buf = (char __user *)lbuf;
1805                 } else {
1806                         nb = nw;
1807                 }
1808
1809                 /* Check if we've wrapped around the hash table */
1810                 if (i >= kvmppc_hpt_npte(&kvm->arch.hpt)) {
1811                         i = 0;
1812                         ctx->first_pass = 0;
1813                         break;
1814                 }
1815         }
1816
1817         ctx->index = i;
1818
1819         return nb;
1820 }
1821
1822 static ssize_t kvm_htab_write(struct file *file, const char __user *buf,
1823                               size_t count, loff_t *ppos)
1824 {
1825         struct kvm_htab_ctx *ctx = file->private_data;
1826         struct kvm *kvm = ctx->kvm;
1827         struct kvm_get_htab_header hdr;
1828         unsigned long i, j;
1829         unsigned long v, r;
1830         unsigned long __user *lbuf;
1831         __be64 *hptp;
1832         unsigned long tmp[2];
1833         ssize_t nb;
1834         long int err, ret;
1835         int mmu_ready;
1836         int pshift;
1837
1838         if (!access_ok(buf, count))
1839                 return -EFAULT;
1840         if (kvm_is_radix(kvm))
1841                 return -EINVAL;
1842
1843         /* lock out vcpus from running while we're doing this */
1844         mutex_lock(&kvm->arch.mmu_setup_lock);
1845         mmu_ready = kvm->arch.mmu_ready;
1846         if (mmu_ready) {
1847                 kvm->arch.mmu_ready = 0;        /* temporarily */
1848                 /* order mmu_ready vs. vcpus_running */
1849                 smp_mb();
1850                 if (atomic_read(&kvm->arch.vcpus_running)) {
1851                         kvm->arch.mmu_ready = 1;
1852                         mutex_unlock(&kvm->arch.mmu_setup_lock);
1853                         return -EBUSY;
1854                 }
1855         }
1856
1857         err = 0;
1858         for (nb = 0; nb + sizeof(hdr) <= count; ) {
1859                 err = -EFAULT;
1860                 if (__copy_from_user(&hdr, buf, sizeof(hdr)))
1861                         break;
1862
1863                 err = 0;
1864                 if (nb + hdr.n_valid * HPTE_SIZE > count)
1865                         break;
1866
1867                 nb += sizeof(hdr);
1868                 buf += sizeof(hdr);
1869
1870                 err = -EINVAL;
1871                 i = hdr.index;
1872                 if (i >= kvmppc_hpt_npte(&kvm->arch.hpt) ||
1873                     i + hdr.n_valid + hdr.n_invalid > kvmppc_hpt_npte(&kvm->arch.hpt))
1874                         break;
1875
1876                 hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
1877                 lbuf = (unsigned long __user *)buf;
1878                 for (j = 0; j < hdr.n_valid; ++j) {
1879                         __be64 hpte_v;
1880                         __be64 hpte_r;
1881
1882                         err = -EFAULT;
1883                         if (__get_user(hpte_v, lbuf) ||
1884                             __get_user(hpte_r, lbuf + 1))
1885                                 goto out;
1886                         v = be64_to_cpu(hpte_v);
1887                         r = be64_to_cpu(hpte_r);
1888                         err = -EINVAL;
1889                         if (!(v & HPTE_V_VALID))
1890                                 goto out;
1891                         pshift = kvmppc_hpte_base_page_shift(v, r);
1892                         if (pshift <= 0)
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 (!mmu_ready && is_vrma_hpte(v)) {
1908                                 unsigned long senc, lpcr;
1909
1910                                 senc = slb_pgsize_encoding(1ul << pshift);
1911                                 kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
1912                                         (VRMA_VSID << SLB_VSID_SHIFT_1T);
1913                                 if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
1914                                         lpcr = senc << (LPCR_VRMASD_SH - 4);
1915                                         kvmppc_update_lpcr(kvm, lpcr,
1916                                                            LPCR_VRMASD);
1917                                 } else {
1918                                         kvmppc_setup_partition_table(kvm);
1919                                 }
1920                                 mmu_ready = 1;
1921                         }
1922                         ++i;
1923                         hptp += 2;
1924                 }
1925
1926                 for (j = 0; j < hdr.n_invalid; ++j) {
1927                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1928                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1929                         ++i;
1930                         hptp += 2;
1931                 }
1932                 err = 0;
1933         }
1934
1935  out:
1936         /* Order HPTE updates vs. mmu_ready */
1937         smp_wmb();
1938         kvm->arch.mmu_ready = mmu_ready;
1939         mutex_unlock(&kvm->arch.mmu_setup_lock);
1940
1941         if (err)
1942                 return err;
1943         return nb;
1944 }
1945
1946 static int kvm_htab_release(struct inode *inode, struct file *filp)
1947 {
1948         struct kvm_htab_ctx *ctx = filp->private_data;
1949
1950         filp->private_data = NULL;
1951         if (!(ctx->flags & KVM_GET_HTAB_WRITE))
1952                 atomic_dec(&ctx->kvm->arch.hpte_mod_interest);
1953         kvm_put_kvm(ctx->kvm);
1954         kfree(ctx);
1955         return 0;
1956 }
1957
1958 static const struct file_operations kvm_htab_fops = {
1959         .read           = kvm_htab_read,
1960         .write          = kvm_htab_write,
1961         .llseek         = default_llseek,
1962         .release        = kvm_htab_release,
1963 };
1964
1965 int kvm_vm_ioctl_get_htab_fd(struct kvm *kvm, struct kvm_get_htab_fd *ghf)
1966 {
1967         int ret;
1968         struct kvm_htab_ctx *ctx;
1969         int rwflag;
1970
1971         /* reject flags we don't recognize */
1972         if (ghf->flags & ~(KVM_GET_HTAB_BOLTED_ONLY | KVM_GET_HTAB_WRITE))
1973                 return -EINVAL;
1974         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1975         if (!ctx)
1976                 return -ENOMEM;
1977         kvm_get_kvm(kvm);
1978         ctx->kvm = kvm;
1979         ctx->index = ghf->start_index;
1980         ctx->flags = ghf->flags;
1981         ctx->first_pass = 1;
1982
1983         rwflag = (ghf->flags & KVM_GET_HTAB_WRITE) ? O_WRONLY : O_RDONLY;
1984         ret = anon_inode_getfd("kvm-htab", &kvm_htab_fops, ctx, rwflag | O_CLOEXEC);
1985         if (ret < 0) {
1986                 kfree(ctx);
1987                 kvm_put_kvm_no_destroy(kvm);
1988                 return ret;
1989         }
1990
1991         if (rwflag == O_RDONLY) {
1992                 mutex_lock(&kvm->slots_lock);
1993                 atomic_inc(&kvm->arch.hpte_mod_interest);
1994                 /* make sure kvmppc_do_h_enter etc. see the increment */
1995                 synchronize_srcu_expedited(&kvm->srcu);
1996                 mutex_unlock(&kvm->slots_lock);
1997         }
1998
1999         return ret;
2000 }
2001
2002 struct debugfs_htab_state {
2003         struct kvm      *kvm;
2004         struct mutex    mutex;
2005         unsigned long   hpt_index;
2006         int             chars_left;
2007         int             buf_index;
2008         char            buf[64];
2009 };
2010
2011 static int debugfs_htab_open(struct inode *inode, struct file *file)
2012 {
2013         struct kvm *kvm = inode->i_private;
2014         struct debugfs_htab_state *p;
2015
2016         p = kzalloc(sizeof(*p), GFP_KERNEL);
2017         if (!p)
2018                 return -ENOMEM;
2019
2020         kvm_get_kvm(kvm);
2021         p->kvm = kvm;
2022         mutex_init(&p->mutex);
2023         file->private_data = p;
2024
2025         return nonseekable_open(inode, file);
2026 }
2027
2028 static int debugfs_htab_release(struct inode *inode, struct file *file)
2029 {
2030         struct debugfs_htab_state *p = file->private_data;
2031
2032         kvm_put_kvm(p->kvm);
2033         kfree(p);
2034         return 0;
2035 }
2036
2037 static ssize_t debugfs_htab_read(struct file *file, char __user *buf,
2038                                  size_t len, loff_t *ppos)
2039 {
2040         struct debugfs_htab_state *p = file->private_data;
2041         ssize_t ret, r;
2042         unsigned long i, n;
2043         unsigned long v, hr, gr;
2044         struct kvm *kvm;
2045         __be64 *hptp;
2046
2047         kvm = p->kvm;
2048         if (kvm_is_radix(kvm))
2049                 return 0;
2050
2051         ret = mutex_lock_interruptible(&p->mutex);
2052         if (ret)
2053                 return ret;
2054
2055         if (p->chars_left) {
2056                 n = p->chars_left;
2057                 if (n > len)
2058                         n = len;
2059                 r = copy_to_user(buf, p->buf + p->buf_index, n);
2060                 n -= r;
2061                 p->chars_left -= n;
2062                 p->buf_index += n;
2063                 buf += n;
2064                 len -= n;
2065                 ret = n;
2066                 if (r) {
2067                         if (!n)
2068                                 ret = -EFAULT;
2069                         goto out;
2070                 }
2071         }
2072
2073         i = p->hpt_index;
2074         hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
2075         for (; len != 0 && i < kvmppc_hpt_npte(&kvm->arch.hpt);
2076              ++i, hptp += 2) {
2077                 if (!(be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)))
2078                         continue;
2079
2080                 /* lock the HPTE so it's stable and read it */
2081                 preempt_disable();
2082                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
2083                         cpu_relax();
2084                 v = be64_to_cpu(hptp[0]) & ~HPTE_V_HVLOCK;
2085                 hr = be64_to_cpu(hptp[1]);
2086                 gr = kvm->arch.hpt.rev[i].guest_rpte;
2087                 unlock_hpte(hptp, v);
2088                 preempt_enable();
2089
2090                 if (!(v & (HPTE_V_VALID | HPTE_V_ABSENT)))
2091                         continue;
2092
2093                 n = scnprintf(p->buf, sizeof(p->buf),
2094                               "%6lx %.16lx %.16lx %.16lx\n",
2095                               i, v, hr, gr);
2096                 p->chars_left = n;
2097                 if (n > len)
2098                         n = len;
2099                 r = copy_to_user(buf, p->buf, n);
2100                 n -= r;
2101                 p->chars_left -= n;
2102                 p->buf_index = n;
2103                 buf += n;
2104                 len -= n;
2105                 ret += n;
2106                 if (r) {
2107                         if (!ret)
2108                                 ret = -EFAULT;
2109                         goto out;
2110                 }
2111         }
2112         p->hpt_index = i;
2113
2114  out:
2115         mutex_unlock(&p->mutex);
2116         return ret;
2117 }
2118
2119 static ssize_t debugfs_htab_write(struct file *file, const char __user *buf,
2120                            size_t len, loff_t *ppos)
2121 {
2122         return -EACCES;
2123 }
2124
2125 static const struct file_operations debugfs_htab_fops = {
2126         .owner   = THIS_MODULE,
2127         .open    = debugfs_htab_open,
2128         .release = debugfs_htab_release,
2129         .read    = debugfs_htab_read,
2130         .write   = debugfs_htab_write,
2131         .llseek  = generic_file_llseek,
2132 };
2133
2134 void kvmppc_mmu_debugfs_init(struct kvm *kvm)
2135 {
2136         debugfs_create_file("htab", 0400, kvm->arch.debugfs_dir, kvm,
2137                             &debugfs_htab_fops);
2138 }
2139
2140 void kvmppc_mmu_book3s_hv_init(struct kvm_vcpu *vcpu)
2141 {
2142         struct kvmppc_mmu *mmu = &vcpu->arch.mmu;
2143
2144         vcpu->arch.slb_nr = 32;         /* POWER7/POWER8 */
2145
2146         mmu->xlate = kvmppc_mmu_book3s_64_hv_xlate;
2147
2148         vcpu->arch.hflags |= BOOK3S_HFLAG_SLB;
2149 }