arch/powerpc/kvm/book3s_64_mmu_hv.c

   1 /*
   2  * This program is free software; you can redistribute it and/or modify
   3  * it under the terms of the GNU General Public License, version 2, as
   4  * published by the Free Software Foundation.
   5  *
   6  * This program is distributed in the hope that it will be useful,
   7  * but WITHOUT ANY WARRANTY; without even the implied warranty of
   8  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
   9  * GNU General Public License for more details.
  10  *
  11  * You should have received a copy of the GNU General Public License
  12  * along with this program; if not, write to the Free Software
  13  * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
  14  *
  15  * Copyright 2010 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
  16  */
  17
  18 #include <linux/types.h>
  19 #include <linux/string.h>
  20 #include <linux/kvm.h>
  21 #include <linux/kvm_host.h>
  22 #include <linux/highmem.h>
  23 #include <linux/gfp.h>
  24 #include <linux/slab.h>
  25 #include <linux/hugetlb.h>
  26 #include <linux/vmalloc.h>
  27 #include <linux/srcu.h>
  28 #include <linux/anon_inodes.h>
  29 #include <linux/file.h>
  30 #include <linux/debugfs.h>
  31
  32 #include <asm/kvm_ppc.h>
  33 #include <asm/kvm_book3s.h>
  34 #include <asm/book3s/64/mmu-hash.h>
  35 #include <asm/hvcall.h>
  36 #include <asm/synch.h>
  37 #include <asm/ppc-opcode.h>
  38 #include <asm/cputable.h>
  39 #include <asm/pte-walk.h>
  40
  41 #include "trace_hv.h"
  42
  43 //#define DEBUG_RESIZE_HPT      1
  44
  45 #ifdef DEBUG_RESIZE_HPT
  46 #define resize_hpt_debug(resize, ...)                           \
  47         do {                                                    \
  48                 printk(KERN_DEBUG "RESIZE HPT %p: ", resize);   \
  49                 printk(__VA_ARGS__);                            \
  50         } while (0)
  51 #else
  52 #define resize_hpt_debug(resize, ...)                           \
  53         do { } while (0)
  54 #endif
  55
  56 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
  57                                 long pte_index, unsigned long pteh,
  58                                 unsigned long ptel, unsigned long *pte_idx_ret);
  59
  60 struct kvm_resize_hpt {
  61         /* These fields read-only after init */
  62         struct kvm *kvm;
  63         struct work_struct work;
  64         u32 order;
  65
  66         /* These fields protected by kvm->lock */
  67
  68         /* Possible values and their usage:
  69          *  <0     an error occurred during allocation,
  70          *  -EBUSY allocation is in the progress,
  71          *  0      allocation made successfuly.
  72          */
  73         int error;
  74
  75         /* Private to the work thread, until error != -EBUSY,
  76          * then protected by kvm->lock.
  77          */
  78         struct kvm_hpt_info hpt;
  79 };
  80
  81 int kvmppc_allocate_hpt(struct kvm_hpt_info *info, u32 order)
  82 {
  83         unsigned long hpt = 0;
  84         int cma = 0;
  85         struct page *page = NULL;
  86         struct revmap_entry *rev;
  87         unsigned long npte;
  88
  89         if ((order < PPC_MIN_HPT_ORDER) || (order > PPC_MAX_HPT_ORDER))
  90                 return -EINVAL;
  91
  92         page = kvm_alloc_hpt_cma(1ul << (order - PAGE_SHIFT));
  93         if (page) {
  94                 hpt = (unsigned long)pfn_to_kaddr(page_to_pfn(page));
  95                 memset((void *)hpt, 0, (1ul << order));
  96                 cma = 1;
  97         }
  98
  99         if (!hpt)
 100                 hpt = __get_free_pages(GFP_KERNEL|__GFP_ZERO|__GFP_RETRY_MAYFAIL
 101                                        |__GFP_NOWARN, order - PAGE_SHIFT);
 102
 103         if (!hpt)
 104                 return -ENOMEM;
 105
 106         /* HPTEs are 2**4 bytes long */
 107         npte = 1ul << (order - 4);
 108
 109         /* Allocate reverse map array */
 110         rev = vmalloc(array_size(npte, sizeof(struct revmap_entry)));
 111         if (!rev) {
 112                 if (cma)
 113                         kvm_free_hpt_cma(page, 1 << (order - PAGE_SHIFT));
 114                 else
 115                         free_pages(hpt, order - PAGE_SHIFT);
 116                 return -ENOMEM;
 117         }
 118
 119         info->order = order;
 120         info->virt = hpt;
 121         info->cma = cma;
 122         info->rev = rev;
 123
 124         return 0;
 125 }
 126
 127 void kvmppc_set_hpt(struct kvm *kvm, struct kvm_hpt_info *info)
 128 {
 129         atomic64_set(&kvm->arch.mmio_update, 0);
 130         kvm->arch.hpt = *info;
 131         kvm->arch.sdr1 = __pa(info->virt) | (info->order - 18);
 132
 133         pr_debug("KVM guest htab at %lx (order %ld), LPID %x\n",
 134                  info->virt, (long)info->order, kvm->arch.lpid);
 135 }
 136
 137 long kvmppc_alloc_reset_hpt(struct kvm *kvm, int order)
 138 {
 139         long err = -EBUSY;
 140         struct kvm_hpt_info info;
 141
 142         mutex_lock(&kvm->lock);
 143         if (kvm->arch.mmu_ready) {
 144                 kvm->arch.mmu_ready = 0;
 145                 /* order mmu_ready vs. vcpus_running */
 146                 smp_mb();
 147                 if (atomic_read(&kvm->arch.vcpus_running)) {
 148                         kvm->arch.mmu_ready = 1;
 149                         goto out;
 150                 }
 151         }
 152         if (kvm_is_radix(kvm)) {
 153                 err = kvmppc_switch_mmu_to_hpt(kvm);
 154                 if (err)
 155                         goto out;
 156         }
 157
 158         if (kvm->arch.hpt.order == order) {
 159                 /* We already have a suitable HPT */
 160
 161                 /* Set the entire HPT to 0, i.e. invalid HPTEs */
 162                 memset((void *)kvm->arch.hpt.virt, 0, 1ul << order);
 163                 /*
 164                  * Reset all the reverse-mapping chains for all memslots
 165                  */
 166                 kvmppc_rmap_reset(kvm);
 167                 err = 0;
 168                 goto out;
 169         }
 170
 171         if (kvm->arch.hpt.virt) {
 172                 kvmppc_free_hpt(&kvm->arch.hpt);
 173                 kvmppc_rmap_reset(kvm);
 174         }
 175
 176         err = kvmppc_allocate_hpt(&info, order);
 177         if (err < 0)
 178                 goto out;
 179         kvmppc_set_hpt(kvm, &info);
 180
 181 out:
 182         if (err == 0)
 183                 /* Ensure that each vcpu will flush its TLB on next entry. */
 184                 cpumask_setall(&kvm->arch.need_tlb_flush);
 185
 186         mutex_unlock(&kvm->lock);
 187         return err;
 188 }
 189
 190 void kvmppc_free_hpt(struct kvm_hpt_info *info)
 191 {
 192         vfree(info->rev);
 193         info->rev = NULL;
 194         if (info->cma)
 195                 kvm_free_hpt_cma(virt_to_page(info->virt),
 196                                  1 << (info->order - PAGE_SHIFT));
 197         else if (info->virt)
 198                 free_pages(info->virt, info->order - PAGE_SHIFT);
 199         info->virt = 0;
 200         info->order = 0;
 201 }
 202
 203 /* Bits in first HPTE dword for pagesize 4k, 64k or 16M */
 204 static inline unsigned long hpte0_pgsize_encoding(unsigned long pgsize)
 205 {
 206         return (pgsize > 0x1000) ? HPTE_V_LARGE : 0;
 207 }
 208
 209 /* Bits in second HPTE dword for pagesize 4k, 64k or 16M */
 210 static inline unsigned long hpte1_pgsize_encoding(unsigned long pgsize)
 211 {
 212         return (pgsize == 0x10000) ? 0x1000 : 0;
 213 }
 214
 215 void kvmppc_map_vrma(struct kvm_vcpu *vcpu, struct kvm_memory_slot *memslot,
 216                      unsigned long porder)
 217 {
 218         unsigned long i;
 219         unsigned long npages;
 220         unsigned long hp_v, hp_r;
 221         unsigned long addr, hash;
 222         unsigned long psize;
 223         unsigned long hp0, hp1;
 224         unsigned long idx_ret;
 225         long ret;
 226         struct kvm *kvm = vcpu->kvm;
 227
 228         psize = 1ul << porder;
 229         npages = memslot->npages >> (porder - PAGE_SHIFT);
 230
 231         /* VRMA can't be > 1TB */
 232         if (npages > 1ul << (40 - porder))
 233                 npages = 1ul << (40 - porder);
 234         /* Can't use more than 1 HPTE per HPTEG */
 235         if (npages > kvmppc_hpt_mask(&kvm->arch.hpt) + 1)
 236                 npages = kvmppc_hpt_mask(&kvm->arch.hpt) + 1;
 237
 238         hp0 = HPTE_V_1TB_SEG | (VRMA_VSID << (40 - 16)) |
 239                 HPTE_V_BOLTED | hpte0_pgsize_encoding(psize);
 240         hp1 = hpte1_pgsize_encoding(psize) |
 241                 HPTE_R_R | HPTE_R_C | HPTE_R_M | PP_RWXX;
 242
 243         for (i = 0; i < npages; ++i) {
 244                 addr = i << porder;
 245                 /* can't use hpt_hash since va > 64 bits */
 246                 hash = (i ^ (VRMA_VSID ^ (VRMA_VSID << 25)))
 247                         & kvmppc_hpt_mask(&kvm->arch.hpt);
 248                 /*
 249                  * We assume that the hash table is empty and no
 250                  * vcpus are using it at this stage.  Since we create
 251                  * at most one HPTE per HPTEG, we just assume entry 7
 252                  * is available and use it.
 253                  */
 254                 hash = (hash << 3) + 7;
 255                 hp_v = hp0 | ((addr >> 16) & ~0x7fUL);
 256                 hp_r = hp1 | addr;
 257                 ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, hash, hp_v, hp_r,
 258                                                  &idx_ret);
 259                 if (ret != H_SUCCESS) {
 260                         pr_err("KVM: map_vrma at %lx failed, ret=%ld\n",
 261                                addr, ret);
 262                         break;
 263                 }
 264         }
 265 }
 266
 267 int kvmppc_mmu_hv_init(void)
 268 {
 269         unsigned long host_lpid, rsvd_lpid;
 270
 271         if (!mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE))
 272                 return -EINVAL;
 273
 274         /* POWER7 has 10-bit LPIDs (12-bit in POWER8) */
 275         host_lpid = 0;
 276         if (cpu_has_feature(CPU_FTR_HVMODE))
 277                 host_lpid = mfspr(SPRN_LPID);
 278         rsvd_lpid = LPID_RSVD;
 279
 280         kvmppc_init_lpid(rsvd_lpid + 1);
 281
 282         kvmppc_claim_lpid(host_lpid);
 283         /* rsvd_lpid is reserved for use in partition switching */
 284         kvmppc_claim_lpid(rsvd_lpid);
 285
 286         return 0;
 287 }
 288
 289 static void kvmppc_mmu_book3s_64_hv_reset_msr(struct kvm_vcpu *vcpu)
 290 {
 291         unsigned long msr = vcpu->arch.intr_msr;
 292
 293         /* If transactional, change to suspend mode on IRQ delivery */
 294         if (MSR_TM_TRANSACTIONAL(vcpu->arch.shregs.msr))
 295                 msr |= MSR_TS_S;
 296         else
 297                 msr |= vcpu->arch.shregs.msr & MSR_TS_MASK;
 298         kvmppc_set_msr(vcpu, msr);
 299 }
 300
 301 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
 302                                 long pte_index, unsigned long pteh,
 303                                 unsigned long ptel, unsigned long *pte_idx_ret)
 304 {
 305         long ret;
 306
 307         /* Protect linux PTE lookup from page table destruction */
 308         rcu_read_lock_sched();  /* this disables preemption too */
 309         ret = kvmppc_do_h_enter(kvm, flags, pte_index, pteh, ptel,
 310                                 current->mm->pgd, false, pte_idx_ret);
 311         rcu_read_unlock_sched();
 312         if (ret == H_TOO_HARD) {
 313                 /* this can't happen */
 314                 pr_err("KVM: Oops, kvmppc_h_enter returned too hard!\n");
 315                 ret = H_RESOURCE;       /* or something */
 316         }
 317         return ret;
 318
 319 }
 320
 321 static struct kvmppc_slb *kvmppc_mmu_book3s_hv_find_slbe(struct kvm_vcpu *vcpu,
 322                                                          gva_t eaddr)
 323 {
 324         u64 mask;
 325         int i;
 326
 327         for (i = 0; i < vcpu->arch.slb_nr; i++) {
 328                 if (!(vcpu->arch.slb[i].orige & SLB_ESID_V))
 329                         continue;
 330
 331                 if (vcpu->arch.slb[i].origv & SLB_VSID_B_1T)
 332                         mask = ESID_MASK_1T;
 333                 else
 334                         mask = ESID_MASK;
 335
 336                 if (((vcpu->arch.slb[i].orige ^ eaddr) & mask) == 0)
 337                         return &vcpu->arch.slb[i];
 338         }
 339         return NULL;
 340 }
 341
 342 static unsigned long kvmppc_mmu_get_real_addr(unsigned long v, unsigned long r,
 343                         unsigned long ea)
 344 {
 345         unsigned long ra_mask;
 346
 347         ra_mask = kvmppc_actual_pgsz(v, r) - 1;
 348         return (r & HPTE_R_RPN & ~ra_mask) | (ea & ra_mask);
 349 }
 350
 351 static int kvmppc_mmu_book3s_64_hv_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
 352                         struct kvmppc_pte *gpte, bool data, bool iswrite)
 353 {
 354         struct kvm *kvm = vcpu->kvm;
 355         struct kvmppc_slb *slbe;
 356         unsigned long slb_v;
 357         unsigned long pp, key;
 358         unsigned long v, orig_v, gr;
 359         __be64 *hptep;
 360         long int index;
 361         int virtmode = vcpu->arch.shregs.msr & (data ? MSR_DR : MSR_IR);
 362
 363         if (kvm_is_radix(vcpu->kvm))
 364                 return kvmppc_mmu_radix_xlate(vcpu, eaddr, gpte, data, iswrite);
 365
 366         /* Get SLB entry */
 367         if (virtmode) {
 368                 slbe = kvmppc_mmu_book3s_hv_find_slbe(vcpu, eaddr);
 369                 if (!slbe)
 370                         return -EINVAL;
 371                 slb_v = slbe->origv;
 372         } else {
 373                 /* real mode access */
 374                 slb_v = vcpu->kvm->arch.vrma_slb_v;
 375         }
 376
 377         preempt_disable();
 378         /* Find the HPTE in the hash table */
 379         index = kvmppc_hv_find_lock_hpte(kvm, eaddr, slb_v,
 380                                          HPTE_V_VALID | HPTE_V_ABSENT);
 381         if (index < 0) {
 382                 preempt_enable();
 383                 return -ENOENT;
 384         }
 385         hptep = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
 386         v = orig_v = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
 387         if (cpu_has_feature(CPU_FTR_ARCH_300))
 388                 v = hpte_new_to_old_v(v, be64_to_cpu(hptep[1]));
 389         gr = kvm->arch.hpt.rev[index].guest_rpte;
 390
 391         unlock_hpte(hptep, orig_v);
 392         preempt_enable();
 393
 394         gpte->eaddr = eaddr;
 395         gpte->vpage = ((v & HPTE_V_AVPN) << 4) | ((eaddr >> 12) & 0xfff);
 396
 397         /* Get PP bits and key for permission check */
 398         pp = gr & (HPTE_R_PP0 | HPTE_R_PP);
 399         key = (vcpu->arch.shregs.msr & MSR_PR) ? SLB_VSID_KP : SLB_VSID_KS;
 400         key &= slb_v;
 401
 402         /* Calculate permissions */
 403         gpte->may_read = hpte_read_permission(pp, key);
 404         gpte->may_write = hpte_write_permission(pp, key);
 405         gpte->may_execute = gpte->may_read && !(gr & (HPTE_R_N | HPTE_R_G));
 406
 407         /* Storage key permission check for POWER7 */
 408         if (data && virtmode) {
 409                 int amrfield = hpte_get_skey_perm(gr, vcpu->arch.amr);
 410                 if (amrfield & 1)
 411                         gpte->may_read = 0;
 412                 if (amrfield & 2)
 413                         gpte->may_write = 0;
 414         }
 415
 416         /* Get the guest physical address */
 417         gpte->raddr = kvmppc_mmu_get_real_addr(v, gr, eaddr);
 418         return 0;
 419 }
 420
 421 /*
 422  * Quick test for whether an instruction is a load or a store.
 423  * If the instruction is a load or a store, then this will indicate
 424  * which it is, at least on server processors.  (Embedded processors
 425  * have some external PID instructions that don't follow the rule
 426  * embodied here.)  If the instruction isn't a load or store, then
 427  * this doesn't return anything useful.
 428  */
 429 static int instruction_is_store(unsigned int instr)
 430 {
 431         unsigned int mask;
 432
 433         mask = 0x10000000;
 434         if ((instr & 0xfc000000) == 0x7c000000)
 435                 mask = 0x100;           /* major opcode 31 */
 436         return (instr & mask) != 0;
 437 }
 438
 439 int kvmppc_hv_emulate_mmio(struct kvm_run *run, struct kvm_vcpu *vcpu,
 440                            unsigned long gpa, gva_t ea, int is_store)
 441 {
 442         u32 last_inst;
 443
 444         /*
 445          * Fast path - check if the guest physical address corresponds to a
 446          * device on the FAST_MMIO_BUS, if so we can avoid loading the
 447          * instruction all together, then we can just handle it and return.
 448          */
 449         if (is_store) {
 450                 int idx, ret;
 451
 452                 idx = srcu_read_lock(&vcpu->kvm->srcu);
 453                 ret = kvm_io_bus_write(vcpu, KVM_FAST_MMIO_BUS, (gpa_t) gpa, 0,
 454                                        NULL);
 455                 srcu_read_unlock(&vcpu->kvm->srcu, idx);
 456                 if (!ret) {
 457                         kvmppc_set_pc(vcpu, kvmppc_get_pc(vcpu) + 4);
 458                         return RESUME_GUEST;
 459                 }
 460         }
 461
 462         /*
 463          * If we fail, we just return to the guest and try executing it again.
 464          */
 465         if (kvmppc_get_last_inst(vcpu, INST_GENERIC, &last_inst) !=
 466                 EMULATE_DONE)
 467                 return RESUME_GUEST;
 468
 469         /*
 470          * WARNING: We do not know for sure whether the instruction we just
 471          * read from memory is the same that caused the fault in the first
 472          * place.  If the instruction we read is neither an load or a store,
 473          * then it can't access memory, so we don't need to worry about
 474          * enforcing access permissions.  So, assuming it is a load or
 475          * store, we just check that its direction (load or store) is
 476          * consistent with the original fault, since that's what we
 477          * checked the access permissions against.  If there is a mismatch
 478          * we just return and retry the instruction.
 479          */
 480
 481         if (instruction_is_store(last_inst) != !!is_store)
 482                 return RESUME_GUEST;
 483
 484         /*
 485          * Emulated accesses are emulated by looking at the hash for
 486          * translation once, then performing the access later. The
 487          * translation could be invalidated in the meantime in which
 488          * point performing the subsequent memory access on the old
 489          * physical address could possibly be a security hole for the
 490          * guest (but not the host).
 491          *
 492          * This is less of an issue for MMIO stores since they aren't
 493          * globally visible. It could be an issue for MMIO loads to
 494          * a certain extent but we'll ignore it for now.
 495          */
 496
 497         vcpu->arch.paddr_accessed = gpa;
 498         vcpu->arch.vaddr_accessed = ea;
 499         return kvmppc_emulate_mmio(run, vcpu);
 500 }
 501
 502 int kvmppc_book3s_hv_page_fault(struct kvm_run *run, struct kvm_vcpu *vcpu,
 503                                 unsigned long ea, unsigned long dsisr)
 504 {
 505         struct kvm *kvm = vcpu->kvm;
 506         unsigned long hpte[3], r;
 507         unsigned long hnow_v, hnow_r;
 508         __be64 *hptep;
 509         unsigned long mmu_seq, psize, pte_size;
 510         unsigned long gpa_base, gfn_base;
 511         unsigned long gpa, gfn, hva, pfn;
 512         struct kvm_memory_slot *memslot;
 513         unsigned long *rmap;
 514         struct revmap_entry *rev;
 515         struct page *page, *pages[1];
 516         long index, ret, npages;
 517         bool is_ci;
 518         unsigned int writing, write_ok;
 519         struct vm_area_struct *vma;
 520         unsigned long rcbits;
 521         long mmio_update;
 522
 523         if (kvm_is_radix(kvm))
 524                 return kvmppc_book3s_radix_page_fault(run, vcpu, ea, dsisr);
 525
 526         /*
 527          * Real-mode code has already searched the HPT and found the
 528          * entry we're interested in.  Lock the entry and check that
 529          * it hasn't changed.  If it has, just return and re-execute the
 530          * instruction.
 531          */
 532         if (ea != vcpu->arch.pgfault_addr)
 533                 return RESUME_GUEST;
 534
 535         if (vcpu->arch.pgfault_cache) {
 536                 mmio_update = atomic64_read(&kvm->arch.mmio_update);
 537                 if (mmio_update == vcpu->arch.pgfault_cache->mmio_update) {
 538                         r = vcpu->arch.pgfault_cache->rpte;
 539                         psize = kvmppc_actual_pgsz(vcpu->arch.pgfault_hpte[0],
 540                                                    r);
 541                         gpa_base = r & HPTE_R_RPN & ~(psize - 1);
 542                         gfn_base = gpa_base >> PAGE_SHIFT;
 543                         gpa = gpa_base | (ea & (psize - 1));
 544                         return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
 545                                                 dsisr & DSISR_ISSTORE);
 546                 }
 547         }
 548         index = vcpu->arch.pgfault_index;
 549         hptep = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
 550         rev = &kvm->arch.hpt.rev[index];
 551         preempt_disable();
 552         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 553                 cpu_relax();
 554         hpte[0] = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
 555         hpte[1] = be64_to_cpu(hptep[1]);
 556         hpte[2] = r = rev->guest_rpte;
 557         unlock_hpte(hptep, hpte[0]);
 558         preempt_enable();
 559
 560         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
 561                 hpte[0] = hpte_new_to_old_v(hpte[0], hpte[1]);
 562                 hpte[1] = hpte_new_to_old_r(hpte[1]);
 563         }
 564         if (hpte[0] != vcpu->arch.pgfault_hpte[0] ||
 565             hpte[1] != vcpu->arch.pgfault_hpte[1])
 566                 return RESUME_GUEST;
 567
 568         /* Translate the logical address and get the page */
 569         psize = kvmppc_actual_pgsz(hpte[0], r);
 570         gpa_base = r & HPTE_R_RPN & ~(psize - 1);
 571         gfn_base = gpa_base >> PAGE_SHIFT;
 572         gpa = gpa_base | (ea & (psize - 1));
 573         gfn = gpa >> PAGE_SHIFT;
 574         memslot = gfn_to_memslot(kvm, gfn);
 575
 576         trace_kvm_page_fault_enter(vcpu, hpte, memslot, ea, dsisr);
 577
 578         /* No memslot means it's an emulated MMIO region */
 579         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
 580                 return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
 581                                               dsisr & DSISR_ISSTORE);
 582
 583         /*
 584          * This should never happen, because of the slot_is_aligned()
 585          * check in kvmppc_do_h_enter().
 586          */
 587         if (gfn_base < memslot->base_gfn)
 588                 return -EFAULT;
 589
 590         /* used to check for invalidations in progress */
 591         mmu_seq = kvm->mmu_notifier_seq;
 592         smp_rmb();
 593
 594         ret = -EFAULT;
 595         is_ci = false;
 596         pfn = 0;
 597         page = NULL;
 598         pte_size = PAGE_SIZE;
 599         writing = (dsisr & DSISR_ISSTORE) != 0;
 600         /* If writing != 0, then the HPTE must allow writing, if we get here */
 601         write_ok = writing;
 602         hva = gfn_to_hva_memslot(memslot, gfn);
 603         npages = get_user_pages_fast(hva, 1, writing, pages);
 604         if (npages < 1) {
 605                 /* Check if it's an I/O mapping */
 606                 down_read(&current->mm->mmap_sem);
 607                 vma = find_vma(current->mm, hva);
 608                 if (vma && vma->vm_start <= hva && hva + psize <= vma->vm_end &&
 609                     (vma->vm_flags & VM_PFNMAP)) {
 610                         pfn = vma->vm_pgoff +
 611                                 ((hva - vma->vm_start) >> PAGE_SHIFT);
 612                         pte_size = psize;
 613                         is_ci = pte_ci(__pte((pgprot_val(vma->vm_page_prot))));
 614                         write_ok = vma->vm_flags & VM_WRITE;
 615                 }
 616                 up_read(&current->mm->mmap_sem);
 617                 if (!pfn)
 618                         goto out_put;
 619         } else {
 620                 page = pages[0];
 621                 pfn = page_to_pfn(page);
 622                 if (PageHuge(page)) {
 623                         page = compound_head(page);
 624                         pte_size <<= compound_order(page);
 625                 }
 626                 /* if the guest wants write access, see if that is OK */
 627                 if (!writing && hpte_is_writable(r)) {
 628                         pte_t *ptep, pte;
 629                         unsigned long flags;
 630                         /*
 631                          * We need to protect against page table destruction
 632                          * hugepage split and collapse.
 633                          */
 634                         local_irq_save(flags);
 635                         ptep = find_current_mm_pte(current->mm->pgd,
 636                                                    hva, NULL, NULL);
 637                         if (ptep) {
 638                                 pte = kvmppc_read_update_linux_pte(ptep, 1);
 639                                 if (__pte_write(pte))
 640                                         write_ok = 1;
 641                         }
 642                         local_irq_restore(flags);
 643                 }
 644         }
 645
 646         if (psize > pte_size)
 647                 goto out_put;
 648
 649         /* Check WIMG vs. the actual page we're accessing */
 650         if (!hpte_cache_flags_ok(r, is_ci)) {
 651                 if (is_ci)
 652                         goto out_put;
 653                 /*
 654                  * Allow guest to map emulated device memory as
 655                  * uncacheable, but actually make it cacheable.
 656                  */
 657                 r = (r & ~(HPTE_R_W|HPTE_R_I|HPTE_R_G)) | HPTE_R_M;
 658         }
 659
 660         /*
 661          * Set the HPTE to point to pfn.
 662          * Since the pfn is at PAGE_SIZE granularity, make sure we
 663          * don't mask out lower-order bits if psize < PAGE_SIZE.
 664          */
 665         if (psize < PAGE_SIZE)
 666                 psize = PAGE_SIZE;
 667         r = (r & HPTE_R_KEY_HI) | (r & ~(HPTE_R_PP0 - psize)) |
 668                                         ((pfn << PAGE_SHIFT) & ~(psize - 1));
 669         if (hpte_is_writable(r) && !write_ok)
 670                 r = hpte_make_readonly(r);
 671         ret = RESUME_GUEST;
 672         preempt_disable();
 673         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 674                 cpu_relax();
 675         hnow_v = be64_to_cpu(hptep[0]);
 676         hnow_r = be64_to_cpu(hptep[1]);
 677         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
 678                 hnow_v = hpte_new_to_old_v(hnow_v, hnow_r);
 679                 hnow_r = hpte_new_to_old_r(hnow_r);
 680         }
 681
 682         /*
 683          * If the HPT is being resized, don't update the HPTE,
 684          * instead let the guest retry after the resize operation is complete.
 685          * The synchronization for mmu_ready test vs. set is provided
 686          * by the HPTE lock.
 687          */
 688         if (!kvm->arch.mmu_ready)
 689                 goto out_unlock;
 690
 691         if ((hnow_v & ~HPTE_V_HVLOCK) != hpte[0] || hnow_r != hpte[1] ||
 692             rev->guest_rpte != hpte[2])
 693                 /* HPTE has been changed under us; let the guest retry */
 694                 goto out_unlock;
 695         hpte[0] = (hpte[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
 696
 697         /* Always put the HPTE in the rmap chain for the page base address */
 698         rmap = &memslot->arch.rmap[gfn_base - memslot->base_gfn];
 699         lock_rmap(rmap);
 700
 701         /* Check if we might have been invalidated; let the guest retry if so */
 702         ret = RESUME_GUEST;
 703         if (mmu_notifier_retry(vcpu->kvm, mmu_seq)) {
 704                 unlock_rmap(rmap);
 705                 goto out_unlock;
 706         }
 707
 708         /* Only set R/C in real HPTE if set in both *rmap and guest_rpte */
 709         rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
 710         r &= rcbits | ~(HPTE_R_R | HPTE_R_C);
 711
 712         if (be64_to_cpu(hptep[0]) & HPTE_V_VALID) {
 713                 /* HPTE was previously valid, so we need to invalidate it */
 714                 unlock_rmap(rmap);
 715                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
 716                 kvmppc_invalidate_hpte(kvm, hptep, index);
 717                 /* don't lose previous R and C bits */
 718                 r |= be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
 719         } else {
 720                 kvmppc_add_revmap_chain(kvm, rev, rmap, index, 0);
 721         }
 722
 723         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
 724                 r = hpte_old_to_new_r(hpte[0], r);
 725                 hpte[0] = hpte_old_to_new_v(hpte[0]);
 726         }
 727         hptep[1] = cpu_to_be64(r);
 728         eieio();
 729         __unlock_hpte(hptep, hpte[0]);
 730         asm volatile("ptesync" : : : "memory");
 731         preempt_enable();
 732         if (page && hpte_is_writable(r))
 733                 SetPageDirty(page);
 734
 735  out_put:
 736         trace_kvm_page_fault_exit(vcpu, hpte, ret);
 737
 738         if (page) {
 739                 /*
 740                  * We drop pages[0] here, not page because page might
 741                  * have been set to the head page of a compound, but
 742                  * we have to drop the reference on the correct tail
 743                  * page to match the get inside gup()
 744                  */
 745                 put_page(pages[0]);
 746         }
 747         return ret;
 748
 749  out_unlock:
 750         __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
 751         preempt_enable();
 752         goto out_put;
 753 }
 754
 755 void kvmppc_rmap_reset(struct kvm *kvm)
 756 {
 757         struct kvm_memslots *slots;
 758         struct kvm_memory_slot *memslot;
 759         int srcu_idx;
 760
 761         srcu_idx = srcu_read_lock(&kvm->srcu);
 762         slots = kvm_memslots(kvm);
 763         kvm_for_each_memslot(memslot, slots) {
 764                 /* Mutual exclusion with kvm_unmap_hva_range etc. */
 765                 spin_lock(&kvm->mmu_lock);
 766                 /*
 767                  * This assumes it is acceptable to lose reference and
 768                  * change bits across a reset.
 769                  */
 770                 memset(memslot->arch.rmap, 0,
 771                        memslot->npages * sizeof(*memslot->arch.rmap));
 772                 spin_unlock(&kvm->mmu_lock);
 773         }
 774         srcu_read_unlock(&kvm->srcu, srcu_idx);
 775 }
 776
 777 typedef int (*hva_handler_fn)(struct kvm *kvm, struct kvm_memory_slot *memslot,
 778                               unsigned long gfn);
 779
 780 static int kvm_handle_hva_range(struct kvm *kvm,
 781                                 unsigned long start,
 782                                 unsigned long end,
 783                                 hva_handler_fn handler)
 784 {
 785         int ret;
 786         int retval = 0;
 787         struct kvm_memslots *slots;
 788         struct kvm_memory_slot *memslot;
 789
 790         slots = kvm_memslots(kvm);
 791         kvm_for_each_memslot(memslot, slots) {
 792                 unsigned long hva_start, hva_end;
 793                 gfn_t gfn, gfn_end;
 794
 795                 hva_start = max(start, memslot->userspace_addr);
 796                 hva_end = min(end, memslot->userspace_addr +
 797                                         (memslot->npages << PAGE_SHIFT));
 798                 if (hva_start >= hva_end)
 799                         continue;
 800                 /*
 801                  * {gfn(page) | page intersects with [hva_start, hva_end)} =
 802                  * {gfn, gfn+1, ..., gfn_end-1}.
 803                  */
 804                 gfn = hva_to_gfn_memslot(hva_start, memslot);
 805                 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
 806
 807                 for (; gfn < gfn_end; ++gfn) {
 808                         ret = handler(kvm, memslot, gfn);
 809                         retval |= ret;
 810                 }
 811         }
 812
 813         return retval;
 814 }
 815
 816 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 817                           hva_handler_fn handler)
 818 {
 819         return kvm_handle_hva_range(kvm, hva, hva + 1, handler);
 820 }
 821
 822 /* Must be called with both HPTE and rmap locked */
 823 static void kvmppc_unmap_hpte(struct kvm *kvm, unsigned long i,
 824                               struct kvm_memory_slot *memslot,
 825                               unsigned long *rmapp, unsigned long gfn)
 826 {
 827         __be64 *hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
 828         struct revmap_entry *rev = kvm->arch.hpt.rev;
 829         unsigned long j, h;
 830         unsigned long ptel, psize, rcbits;
 831
 832         j = rev[i].forw;
 833         if (j == i) {
 834                 /* chain is now empty */
 835                 *rmapp &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
 836         } else {
 837                 /* remove i from chain */
 838                 h = rev[i].back;
 839                 rev[h].forw = j;
 840                 rev[j].back = h;
 841                 rev[i].forw = rev[i].back = i;
 842                 *rmapp = (*rmapp & ~KVMPPC_RMAP_INDEX) | j;
 843         }
 844
 845         /* Now check and modify the HPTE */
 846         ptel = rev[i].guest_rpte;
 847         psize = kvmppc_actual_pgsz(be64_to_cpu(hptep[0]), ptel);
 848         if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
 849             hpte_rpn(ptel, psize) == gfn) {
 850                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
 851                 kvmppc_invalidate_hpte(kvm, hptep, i);
 852                 hptep[1] &= ~cpu_to_be64(HPTE_R_KEY_HI | HPTE_R_KEY_LO);
 853                 /* Harvest R and C */
 854                 rcbits = be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
 855                 *rmapp |= rcbits << KVMPPC_RMAP_RC_SHIFT;
 856                 if ((rcbits & HPTE_R_C) && memslot->dirty_bitmap)
 857                         kvmppc_update_dirty_map(memslot, gfn, psize);
 858                 if (rcbits & ~rev[i].guest_rpte) {
 859                         rev[i].guest_rpte = ptel | rcbits;
 860                         note_hpte_modification(kvm, &rev[i]);
 861                 }
 862         }
 863 }
 864
 865 static int kvm_unmap_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
 866                            unsigned long gfn)
 867 {
 868         unsigned long i;
 869         __be64 *hptep;
 870         unsigned long *rmapp;
 871
 872         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
 873         for (;;) {
 874                 lock_rmap(rmapp);
 875                 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 876                         unlock_rmap(rmapp);
 877                         break;
 878                 }
 879
 880                 /*
 881                  * To avoid an ABBA deadlock with the HPTE lock bit,
 882                  * we can't spin on the HPTE lock while holding the
 883                  * rmap chain lock.
 884                  */
 885                 i = *rmapp & KVMPPC_RMAP_INDEX;
 886                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
 887                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 888                         /* unlock rmap before spinning on the HPTE lock */
 889                         unlock_rmap(rmapp);
 890                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
 891                                 cpu_relax();
 892                         continue;
 893                 }
 894
 895                 kvmppc_unmap_hpte(kvm, i, memslot, rmapp, gfn);
 896                 unlock_rmap(rmapp);
 897                 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
 898         }
 899         return 0;
 900 }
 901
 902 int kvm_unmap_hva_range_hv(struct kvm *kvm, unsigned long start, unsigned long end)
 903 {
 904         hva_handler_fn handler;
 905
 906         handler = kvm_is_radix(kvm) ? kvm_unmap_radix : kvm_unmap_rmapp;
 907         kvm_handle_hva_range(kvm, start, end, handler);
 908         return 0;
 909 }
 910
 911 void kvmppc_core_flush_memslot_hv(struct kvm *kvm,
 912                                   struct kvm_memory_slot *memslot)
 913 {
 914         unsigned long gfn;
 915         unsigned long n;
 916         unsigned long *rmapp;
 917
 918         gfn = memslot->base_gfn;
 919         rmapp = memslot->arch.rmap;
 920         if (kvm_is_radix(kvm)) {
 921                 kvmppc_radix_flush_memslot(kvm, memslot);
 922                 return;
 923         }
 924
 925         for (n = memslot->npages; n; --n, ++gfn) {
 926                 /*
 927                  * Testing the present bit without locking is OK because
 928                  * the memslot has been marked invalid already, and hence
 929                  * no new HPTEs referencing this page can be created,
 930                  * thus the present bit can't go from 0 to 1.
 931                  */
 932                 if (*rmapp & KVMPPC_RMAP_PRESENT)
 933                         kvm_unmap_rmapp(kvm, memslot, gfn);
 934                 ++rmapp;
 935         }
 936 }
 937
 938 static int kvm_age_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
 939                          unsigned long gfn)
 940 {
 941         struct revmap_entry *rev = kvm->arch.hpt.rev;
 942         unsigned long head, i, j;
 943         __be64 *hptep;
 944         int ret = 0;
 945         unsigned long *rmapp;
 946
 947         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
 948  retry:
 949         lock_rmap(rmapp);
 950         if (*rmapp & KVMPPC_RMAP_REFERENCED) {
 951                 *rmapp &= ~KVMPPC_RMAP_REFERENCED;
 952                 ret = 1;
 953         }
 954         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 955                 unlock_rmap(rmapp);
 956                 return ret;
 957         }
 958
 959         i = head = *rmapp & KVMPPC_RMAP_INDEX;
 960         do {
 961                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
 962                 j = rev[i].forw;
 963
 964                 /* If this HPTE isn't referenced, ignore it */
 965                 if (!(be64_to_cpu(hptep[1]) & HPTE_R_R))
 966                         continue;
 967
 968                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 969                         /* unlock rmap before spinning on the HPTE lock */
 970                         unlock_rmap(rmapp);
 971                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
 972                                 cpu_relax();
 973                         goto retry;
 974                 }
 975
 976                 /* Now check and modify the HPTE */
 977                 if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
 978                     (be64_to_cpu(hptep[1]) & HPTE_R_R)) {
 979                         kvmppc_clear_ref_hpte(kvm, hptep, i);
 980                         if (!(rev[i].guest_rpte & HPTE_R_R)) {
 981                                 rev[i].guest_rpte |= HPTE_R_R;
 982                                 note_hpte_modification(kvm, &rev[i]);
 983                         }
 984                         ret = 1;
 985                 }
 986                 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
 987         } while ((i = j) != head);
 988
 989         unlock_rmap(rmapp);
 990         return ret;
 991 }
 992
 993 int kvm_age_hva_hv(struct kvm *kvm, unsigned long start, unsigned long end)
 994 {
 995         hva_handler_fn handler;
 996
 997         handler = kvm_is_radix(kvm) ? kvm_age_radix : kvm_age_rmapp;
 998         return kvm_handle_hva_range(kvm, start, end, handler);
 999 }
1000
1001 static int kvm_test_age_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
1002                               unsigned long gfn)
1003 {
1004         struct revmap_entry *rev = kvm->arch.hpt.rev;
1005         unsigned long head, i, j;
1006         unsigned long *hp;
1007         int ret = 1;
1008         unsigned long *rmapp;
1009
1010         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1011         if (*rmapp & KVMPPC_RMAP_REFERENCED)
1012                 return 1;
1013
1014         lock_rmap(rmapp);
1015         if (*rmapp & KVMPPC_RMAP_REFERENCED)
1016                 goto out;
1017
1018         if (*rmapp & KVMPPC_RMAP_PRESENT) {
1019                 i = head = *rmapp & KVMPPC_RMAP_INDEX;
1020                 do {
1021                         hp = (unsigned long *)(kvm->arch.hpt.virt + (i << 4));
1022                         j = rev[i].forw;
1023                         if (be64_to_cpu(hp[1]) & HPTE_R_R)
1024                                 goto out;
1025                 } while ((i = j) != head);
1026         }
1027         ret = 0;
1028
1029  out:
1030         unlock_rmap(rmapp);
1031         return ret;
1032 }
1033
1034 int kvm_test_age_hva_hv(struct kvm *kvm, unsigned long hva)
1035 {
1036         hva_handler_fn handler;
1037
1038         handler = kvm_is_radix(kvm) ? kvm_test_age_radix : kvm_test_age_rmapp;
1039         return kvm_handle_hva(kvm, hva, handler);
1040 }
1041
1042 void kvm_set_spte_hva_hv(struct kvm *kvm, unsigned long hva, pte_t pte)
1043 {
1044         hva_handler_fn handler;
1045
1046         handler = kvm_is_radix(kvm) ? kvm_unmap_radix : kvm_unmap_rmapp;
1047         kvm_handle_hva(kvm, hva, handler);
1048 }
1049
1050 static int vcpus_running(struct kvm *kvm)
1051 {
1052         return atomic_read(&kvm->arch.vcpus_running) != 0;
1053 }
1054
1055 /*
1056  * Returns the number of system pages that are dirty.
1057  * This can be more than 1 if we find a huge-page HPTE.
1058  */
1059 static int kvm_test_clear_dirty_npages(struct kvm *kvm, unsigned long *rmapp)
1060 {
1061         struct revmap_entry *rev = kvm->arch.hpt.rev;
1062         unsigned long head, i, j;
1063         unsigned long n;
1064         unsigned long v, r;
1065         __be64 *hptep;
1066         int npages_dirty = 0;
1067
1068  retry:
1069         lock_rmap(rmapp);
1070         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
1071                 unlock_rmap(rmapp);
1072                 return npages_dirty;
1073         }
1074
1075         i = head = *rmapp & KVMPPC_RMAP_INDEX;
1076         do {
1077                 unsigned long hptep1;
1078                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
1079                 j = rev[i].forw;
1080
1081                 /*
1082                  * Checking the C (changed) bit here is racy since there
1083                  * is no guarantee about when the hardware writes it back.
1084                  * If the HPTE is not writable then it is stable since the
1085                  * page can't be written to, and we would have done a tlbie
1086                  * (which forces the hardware to complete any writeback)
1087                  * when making the HPTE read-only.
1088                  * If vcpus are running then this call is racy anyway
1089                  * since the page could get dirtied subsequently, so we
1090                  * expect there to be a further call which would pick up
1091                  * any delayed C bit writeback.
1092                  * Otherwise we need to do the tlbie even if C==0 in
1093                  * order to pick up any delayed writeback of C.
1094                  */
1095                 hptep1 = be64_to_cpu(hptep[1]);
1096                 if (!(hptep1 & HPTE_R_C) &&
1097                     (!hpte_is_writable(hptep1) || vcpus_running(kvm)))
1098                         continue;
1099
1100                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
1101                         /* unlock rmap before spinning on the HPTE lock */
1102                         unlock_rmap(rmapp);
1103                         while (hptep[0] & cpu_to_be64(HPTE_V_HVLOCK))
1104                                 cpu_relax();
1105                         goto retry;
1106                 }
1107
1108                 /* Now check and modify the HPTE */
1109                 if (!(hptep[0] & cpu_to_be64(HPTE_V_VALID))) {
1110                         __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
1111                         continue;
1112                 }
1113
1114                 /* need to make it temporarily absent so C is stable */
1115                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
1116                 kvmppc_invalidate_hpte(kvm, hptep, i);
1117                 v = be64_to_cpu(hptep[0]);
1118                 r = be64_to_cpu(hptep[1]);
1119                 if (r & HPTE_R_C) {
1120                         hptep[1] = cpu_to_be64(r & ~HPTE_R_C);
1121                         if (!(rev[i].guest_rpte & HPTE_R_C)) {
1122                                 rev[i].guest_rpte |= HPTE_R_C;
1123                                 note_hpte_modification(kvm, &rev[i]);
1124                         }
1125                         n = kvmppc_actual_pgsz(v, r);
1126                         n = (n + PAGE_SIZE - 1) >> PAGE_SHIFT;
1127                         if (n > npages_dirty)
1128                                 npages_dirty = n;
1129                         eieio();
1130                 }
1131                 v &= ~HPTE_V_ABSENT;
1132                 v |= HPTE_V_VALID;
1133                 __unlock_hpte(hptep, v);
1134         } while ((i = j) != head);
1135
1136         unlock_rmap(rmapp);
1137         return npages_dirty;
1138 }
1139
1140 void kvmppc_harvest_vpa_dirty(struct kvmppc_vpa *vpa,
1141                               struct kvm_memory_slot *memslot,
1142                               unsigned long *map)
1143 {
1144         unsigned long gfn;
1145
1146         if (!vpa->dirty || !vpa->pinned_addr)
1147                 return;
1148         gfn = vpa->gpa >> PAGE_SHIFT;
1149         if (gfn < memslot->base_gfn ||
1150             gfn >= memslot->base_gfn + memslot->npages)
1151                 return;
1152
1153         vpa->dirty = false;
1154         if (map)
1155                 __set_bit_le(gfn - memslot->base_gfn, map);
1156 }
1157
1158 long kvmppc_hv_get_dirty_log_hpt(struct kvm *kvm,
1159                         struct kvm_memory_slot *memslot, unsigned long *map)
1160 {
1161         unsigned long i;
1162         unsigned long *rmapp;
1163
1164         preempt_disable();
1165         rmapp = memslot->arch.rmap;
1166         for (i = 0; i < memslot->npages; ++i) {
1167                 int npages = kvm_test_clear_dirty_npages(kvm, rmapp);
1168                 /*
1169                  * Note that if npages > 0 then i must be a multiple of npages,
1170                  * since we always put huge-page HPTEs in the rmap chain
1171                  * corresponding to their page base address.
1172                  */
1173                 if (npages)
1174                         set_dirty_bits(map, i, npages);
1175                 ++rmapp;
1176         }
1177         preempt_enable();
1178         return 0;
1179 }
1180
1181 void *kvmppc_pin_guest_page(struct kvm *kvm, unsigned long gpa,
1182                             unsigned long *nb_ret)
1183 {
1184         struct kvm_memory_slot *memslot;
1185         unsigned long gfn = gpa >> PAGE_SHIFT;
1186         struct page *page, *pages[1];
1187         int npages;
1188         unsigned long hva, offset;
1189         int srcu_idx;
1190
1191         srcu_idx = srcu_read_lock(&kvm->srcu);
1192         memslot = gfn_to_memslot(kvm, gfn);
1193         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
1194                 goto err;
1195         hva = gfn_to_hva_memslot(memslot, gfn);
1196         npages = get_user_pages_fast(hva, 1, 1, pages);
1197         if (npages < 1)
1198                 goto err;
1199         page = pages[0];
1200         srcu_read_unlock(&kvm->srcu, srcu_idx);
1201
1202         offset = gpa & (PAGE_SIZE - 1);
1203         if (nb_ret)
1204                 *nb_ret = PAGE_SIZE - offset;
1205         return page_address(page) + offset;
1206
1207  err:
1208         srcu_read_unlock(&kvm->srcu, srcu_idx);
1209         return NULL;
1210 }
1211
1212 void kvmppc_unpin_guest_page(struct kvm *kvm, void *va, unsigned long gpa,
1213                              bool dirty)
1214 {
1215         struct page *page = virt_to_page(va);
1216         struct kvm_memory_slot *memslot;
1217         unsigned long gfn;
1218         int srcu_idx;
1219
1220         put_page(page);
1221
1222         if (!dirty)
1223                 return;
1224
1225         /* We need to mark this page dirty in the memslot dirty_bitmap, if any */
1226         gfn = gpa >> PAGE_SHIFT;
1227         srcu_idx = srcu_read_lock(&kvm->srcu);
1228         memslot = gfn_to_memslot(kvm, gfn);
1229         if (memslot && memslot->dirty_bitmap)
1230                 set_bit_le(gfn - memslot->base_gfn, memslot->dirty_bitmap);
1231         srcu_read_unlock(&kvm->srcu, srcu_idx);
1232 }
1233
1234 /*
1235  * HPT resizing
1236  */
1237 static int resize_hpt_allocate(struct kvm_resize_hpt *resize)
1238 {
1239         int rc;
1240
1241         rc = kvmppc_allocate_hpt(&resize->hpt, resize->order);
1242         if (rc < 0)
1243                 return rc;
1244
1245         resize_hpt_debug(resize, "resize_hpt_allocate(): HPT @ 0x%lx\n",
1246                          resize->hpt.virt);
1247
1248         return 0;
1249 }
1250
1251 static unsigned long resize_hpt_rehash_hpte(struct kvm_resize_hpt *resize,
1252                                             unsigned long idx)
1253 {
1254         struct kvm *kvm = resize->kvm;
1255         struct kvm_hpt_info *old = &kvm->arch.hpt;
1256         struct kvm_hpt_info *new = &resize->hpt;
1257         unsigned long old_hash_mask = (1ULL << (old->order - 7)) - 1;
1258         unsigned long new_hash_mask = (1ULL << (new->order - 7)) - 1;
1259         __be64 *hptep, *new_hptep;
1260         unsigned long vpte, rpte, guest_rpte;
1261         int ret;
1262         struct revmap_entry *rev;
1263         unsigned long apsize, avpn, pteg, hash;
1264         unsigned long new_idx, new_pteg, replace_vpte;
1265         int pshift;
1266
1267         hptep = (__be64 *)(old->virt + (idx << 4));
1268
1269         /* Guest is stopped, so new HPTEs can't be added or faulted
1270          * in, only unmapped or altered by host actions.  So, it's
1271          * safe to check this before we take the HPTE lock */
1272         vpte = be64_to_cpu(hptep[0]);
1273         if (!(vpte & HPTE_V_VALID) && !(vpte & HPTE_V_ABSENT))
1274                 return 0; /* nothing to do */
1275
1276         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
1277                 cpu_relax();
1278
1279         vpte = be64_to_cpu(hptep[0]);
1280
1281         ret = 0;
1282         if (!(vpte & HPTE_V_VALID) && !(vpte & HPTE_V_ABSENT))
1283                 /* Nothing to do */
1284                 goto out;
1285
1286         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1287                 rpte = be64_to_cpu(hptep[1]);
1288                 vpte = hpte_new_to_old_v(vpte, rpte);
1289         }
1290
1291         /* Unmap */
1292         rev = &old->rev[idx];
1293         guest_rpte = rev->guest_rpte;
1294
1295         ret = -EIO;
1296         apsize = kvmppc_actual_pgsz(vpte, guest_rpte);
1297         if (!apsize)
1298                 goto out;
1299
1300         if (vpte & HPTE_V_VALID) {
1301                 unsigned long gfn = hpte_rpn(guest_rpte, apsize);
1302                 int srcu_idx = srcu_read_lock(&kvm->srcu);
1303                 struct kvm_memory_slot *memslot =
1304                         __gfn_to_memslot(kvm_memslots(kvm), gfn);
1305
1306                 if (memslot) {
1307                         unsigned long *rmapp;
1308                         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1309
1310                         lock_rmap(rmapp);
1311                         kvmppc_unmap_hpte(kvm, idx, memslot, rmapp, gfn);
1312                         unlock_rmap(rmapp);
1313                 }
1314
1315                 srcu_read_unlock(&kvm->srcu, srcu_idx);
1316         }
1317
1318         /* Reload PTE after unmap */
1319         vpte = be64_to_cpu(hptep[0]);
1320         BUG_ON(vpte & HPTE_V_VALID);
1321         BUG_ON(!(vpte & HPTE_V_ABSENT));
1322
1323         ret = 0;
1324         if (!(vpte & HPTE_V_BOLTED))
1325                 goto out;
1326
1327         rpte = be64_to_cpu(hptep[1]);
1328
1329         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1330                 vpte = hpte_new_to_old_v(vpte, rpte);
1331                 rpte = hpte_new_to_old_r(rpte);
1332         }
1333
1334         pshift = kvmppc_hpte_base_page_shift(vpte, rpte);
1335         avpn = HPTE_V_AVPN_VAL(vpte) & ~(((1ul << pshift) - 1) >> 23);
1336         pteg = idx / HPTES_PER_GROUP;
1337         if (vpte & HPTE_V_SECONDARY)
1338                 pteg = ~pteg;
1339
1340         if (!(vpte & HPTE_V_1TB_SEG)) {
1341                 unsigned long offset, vsid;
1342
1343                 /* We only have 28 - 23 bits of offset in avpn */
1344                 offset = (avpn & 0x1f) << 23;
1345                 vsid = avpn >> 5;
1346                 /* We can find more bits from the pteg value */
1347                 if (pshift < 23)
1348                         offset |= ((vsid ^ pteg) & old_hash_mask) << pshift;
1349
1350                 hash = vsid ^ (offset >> pshift);
1351         } else {
1352                 unsigned long offset, vsid;
1353
1354                 /* We only have 40 - 23 bits of seg_off in avpn */
1355                 offset = (avpn & 0x1ffff) << 23;
1356                 vsid = avpn >> 17;
1357                 if (pshift < 23)
1358                         offset |= ((vsid ^ (vsid << 25) ^ pteg) & old_hash_mask) << pshift;
1359
1360                 hash = vsid ^ (vsid << 25) ^ (offset >> pshift);
1361         }
1362
1363         new_pteg = hash & new_hash_mask;
1364         if (vpte & HPTE_V_SECONDARY)
1365                 new_pteg = ~hash & new_hash_mask;
1366
1367         new_idx = new_pteg * HPTES_PER_GROUP + (idx % HPTES_PER_GROUP);
1368         new_hptep = (__be64 *)(new->virt + (new_idx << 4));
1369
1370         replace_vpte = be64_to_cpu(new_hptep[0]);
1371         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1372                 unsigned long replace_rpte = be64_to_cpu(new_hptep[1]);
1373                 replace_vpte = hpte_new_to_old_v(replace_vpte, replace_rpte);
1374         }
1375
1376         if (replace_vpte & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1377                 BUG_ON(new->order >= old->order);
1378
1379                 if (replace_vpte & HPTE_V_BOLTED) {
1380                         if (vpte & HPTE_V_BOLTED)
1381                                 /* Bolted collision, nothing we can do */
1382                                 ret = -ENOSPC;
1383                         /* Discard the new HPTE */
1384                         goto out;
1385                 }
1386
1387                 /* Discard the previous HPTE */
1388         }
1389
1390         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1391                 rpte = hpte_old_to_new_r(vpte, rpte);
1392                 vpte = hpte_old_to_new_v(vpte);
1393         }
1394
1395         new_hptep[1] = cpu_to_be64(rpte);
1396         new->rev[new_idx].guest_rpte = guest_rpte;
1397         /* No need for a barrier, since new HPT isn't active */
1398         new_hptep[0] = cpu_to_be64(vpte);
1399         unlock_hpte(new_hptep, vpte);
1400
1401 out:
1402         unlock_hpte(hptep, vpte);
1403         return ret;
1404 }
1405
1406 static int resize_hpt_rehash(struct kvm_resize_hpt *resize)
1407 {
1408         struct kvm *kvm = resize->kvm;
1409         unsigned  long i;
1410         int rc;
1411
1412         for (i = 0; i < kvmppc_hpt_npte(&kvm->arch.hpt); i++) {
1413                 rc = resize_hpt_rehash_hpte(resize, i);
1414                 if (rc != 0)
1415                         return rc;
1416         }
1417
1418         return 0;
1419 }
1420
1421 static void resize_hpt_pivot(struct kvm_resize_hpt *resize)
1422 {
1423         struct kvm *kvm = resize->kvm;
1424         struct kvm_hpt_info hpt_tmp;
1425
1426         /* Exchange the pending tables in the resize structure with
1427          * the active tables */
1428
1429         resize_hpt_debug(resize, "resize_hpt_pivot()\n");
1430
1431         spin_lock(&kvm->mmu_lock);
1432         asm volatile("ptesync" : : : "memory");
1433
1434         hpt_tmp = kvm->arch.hpt;
1435         kvmppc_set_hpt(kvm, &resize->hpt);
1436         resize->hpt = hpt_tmp;
1437
1438         spin_unlock(&kvm->mmu_lock);
1439
1440         synchronize_srcu_expedited(&kvm->srcu);
1441
1442         if (cpu_has_feature(CPU_FTR_ARCH_300))
1443                 kvmppc_setup_partition_table(kvm);
1444
1445         resize_hpt_debug(resize, "resize_hpt_pivot() done\n");
1446 }
1447
1448 static void resize_hpt_release(struct kvm *kvm, struct kvm_resize_hpt *resize)
1449 {
1450         if (WARN_ON(!mutex_is_locked(&kvm->lock)))
1451                 return;
1452
1453         if (!resize)
1454                 return;
1455
1456         if (resize->error != -EBUSY) {
1457                 if (resize->hpt.virt)
1458                         kvmppc_free_hpt(&resize->hpt);
1459                 kfree(resize);
1460         }
1461
1462         if (kvm->arch.resize_hpt == resize)
1463                 kvm->arch.resize_hpt = NULL;
1464 }
1465
1466 static void resize_hpt_prepare_work(struct work_struct *work)
1467 {
1468         struct kvm_resize_hpt *resize = container_of(work,
1469                                                      struct kvm_resize_hpt,
1470                                                      work);
1471         struct kvm *kvm = resize->kvm;
1472         int err = 0;
1473
1474         if (WARN_ON(resize->error != -EBUSY))
1475                 return;
1476
1477         mutex_lock(&kvm->lock);
1478
1479         /* Request is still current? */
1480         if (kvm->arch.resize_hpt == resize) {
1481                 /* We may request large allocations here:
1482                  * do not sleep with kvm->lock held for a while.
1483                  */
1484                 mutex_unlock(&kvm->lock);
1485
1486                 resize_hpt_debug(resize, "resize_hpt_prepare_work(): order = %d\n",
1487                                  resize->order);
1488
1489                 err = resize_hpt_allocate(resize);
1490
1491                 /* We have strict assumption about -EBUSY
1492                  * when preparing for HPT resize.
1493                  */
1494                 if (WARN_ON(err == -EBUSY))
1495                         err = -EINPROGRESS;
1496
1497                 mutex_lock(&kvm->lock);
1498                 /* It is possible that kvm->arch.resize_hpt != resize
1499                  * after we grab kvm->lock again.
1500                  */
1501         }
1502
1503         resize->error = err;
1504
1505         if (kvm->arch.resize_hpt != resize)
1506                 resize_hpt_release(kvm, resize);
1507
1508         mutex_unlock(&kvm->lock);
1509 }
1510
1511 long kvm_vm_ioctl_resize_hpt_prepare(struct kvm *kvm,
1512                                      struct kvm_ppc_resize_hpt *rhpt)
1513 {
1514         unsigned long flags = rhpt->flags;
1515         unsigned long shift = rhpt->shift;
1516         struct kvm_resize_hpt *resize;
1517         int ret;
1518
1519         if (flags != 0 || kvm_is_radix(kvm))
1520                 return -EINVAL;
1521
1522         if (shift && ((shift < 18) || (shift > 46)))
1523                 return -EINVAL;
1524
1525         mutex_lock(&kvm->lock);
1526
1527         resize = kvm->arch.resize_hpt;
1528
1529         if (resize) {
1530                 if (resize->order == shift) {
1531                         /* Suitable resize in progress? */
1532                         ret = resize->error;
1533                         if (ret == -EBUSY)
1534                                 ret = 100; /* estimated time in ms */
1535                         else if (ret)
1536                                 resize_hpt_release(kvm, resize);
1537
1538                         goto out;
1539                 }
1540
1541                 /* not suitable, cancel it */
1542                 resize_hpt_release(kvm, resize);
1543         }
1544
1545         ret = 0;
1546         if (!shift)
1547                 goto out; /* nothing to do */
1548
1549         /* start new resize */
1550
1551         resize = kzalloc(sizeof(*resize), GFP_KERNEL);
1552         if (!resize) {
1553                 ret = -ENOMEM;
1554                 goto out;
1555         }
1556
1557         resize->error = -EBUSY;
1558         resize->order = shift;
1559         resize->kvm = kvm;
1560         INIT_WORK(&resize->work, resize_hpt_prepare_work);
1561         kvm->arch.resize_hpt = resize;
1562
1563         schedule_work(&resize->work);
1564
1565         ret = 100; /* estimated time in ms */
1566
1567 out:
1568         mutex_unlock(&kvm->lock);
1569         return ret;
1570 }
1571
1572 static void resize_hpt_boot_vcpu(void *opaque)
1573 {
1574         /* Nothing to do, just force a KVM exit */
1575 }
1576
1577 long kvm_vm_ioctl_resize_hpt_commit(struct kvm *kvm,
1578                                     struct kvm_ppc_resize_hpt *rhpt)
1579 {
1580         unsigned long flags = rhpt->flags;
1581         unsigned long shift = rhpt->shift;
1582         struct kvm_resize_hpt *resize;
1583         long ret;
1584
1585         if (flags != 0 || kvm_is_radix(kvm))
1586                 return -EINVAL;
1587
1588         if (shift && ((shift < 18) || (shift > 46)))
1589                 return -EINVAL;
1590
1591         mutex_lock(&kvm->lock);
1592
1593         resize = kvm->arch.resize_hpt;
1594
1595         /* This shouldn't be possible */
1596         ret = -EIO;
1597         if (WARN_ON(!kvm->arch.mmu_ready))
1598                 goto out_no_hpt;
1599
1600         /* Stop VCPUs from running while we mess with the HPT */
1601         kvm->arch.mmu_ready = 0;
1602         smp_mb();
1603
1604         /* Boot all CPUs out of the guest so they re-read
1605          * mmu_ready */
1606         on_each_cpu(resize_hpt_boot_vcpu, NULL, 1);
1607
1608         ret = -ENXIO;
1609         if (!resize || (resize->order != shift))
1610                 goto out;
1611
1612         ret = resize->error;
1613         if (ret)
1614                 goto out;
1615
1616         ret = resize_hpt_rehash(resize);
1617         if (ret)
1618                 goto out;
1619
1620         resize_hpt_pivot(resize);
1621
1622 out:
1623         /* Let VCPUs run again */
1624         kvm->arch.mmu_ready = 1;
1625         smp_mb();
1626 out_no_hpt:
1627         resize_hpt_release(kvm, resize);
1628         mutex_unlock(&kvm->lock);
1629         return ret;
1630 }
1631
1632 /*
1633  * Functions for reading and writing the hash table via reads and
1634  * writes on a file descriptor.
1635  *
1636  * Reads return the guest view of the hash table, which has to be
1637  * pieced together from the real hash table and the guest_rpte
1638  * values in the revmap array.
1639  *
1640  * On writes, each HPTE written is considered in turn, and if it
1641  * is valid, it is written to the HPT as if an H_ENTER with the
1642  * exact flag set was done.  When the invalid count is non-zero
1643  * in the header written to the stream, the kernel will make
1644  * sure that that many HPTEs are invalid, and invalidate them
1645  * if not.
1646  */
1647
1648 struct kvm_htab_ctx {
1649         unsigned long   index;
1650         unsigned long   flags;
1651         struct kvm      *kvm;
1652         int             first_pass;
1653 };
1654
1655 #define HPTE_SIZE       (2 * sizeof(unsigned long))
1656
1657 /*
1658  * Returns 1 if this HPT entry has been modified or has pending
1659  * R/C bit changes.
1660  */
1661 static int hpte_dirty(struct revmap_entry *revp, __be64 *hptp)
1662 {
1663         unsigned long rcbits_unset;
1664
1665         if (revp->guest_rpte & HPTE_GR_MODIFIED)
1666                 return 1;
1667
1668         /* Also need to consider changes in reference and changed bits */
1669         rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1670         if ((be64_to_cpu(hptp[0]) & HPTE_V_VALID) &&
1671             (be64_to_cpu(hptp[1]) & rcbits_unset))
1672                 return 1;
1673
1674         return 0;
1675 }
1676
1677 static long record_hpte(unsigned long flags, __be64 *hptp,
1678                         unsigned long *hpte, struct revmap_entry *revp,
1679                         int want_valid, int first_pass)
1680 {
1681         unsigned long v, r, hr;
1682         unsigned long rcbits_unset;
1683         int ok = 1;
1684         int valid, dirty;
1685
1686         /* Unmodified entries are uninteresting except on the first pass */
1687         dirty = hpte_dirty(revp, hptp);
1688         if (!first_pass && !dirty)
1689                 return 0;
1690
1691         valid = 0;
1692         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1693                 valid = 1;
1694                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) &&
1695                     !(be64_to_cpu(hptp[0]) & HPTE_V_BOLTED))
1696                         valid = 0;
1697         }
1698         if (valid != want_valid)
1699                 return 0;
1700
1701         v = r = 0;
1702         if (valid || dirty) {
1703                 /* lock the HPTE so it's stable and read it */
1704                 preempt_disable();
1705                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
1706                         cpu_relax();
1707                 v = be64_to_cpu(hptp[0]);
1708                 hr = be64_to_cpu(hptp[1]);
1709                 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1710                         v = hpte_new_to_old_v(v, hr);
1711                         hr = hpte_new_to_old_r(hr);
1712                 }
1713
1714                 /* re-evaluate valid and dirty from synchronized HPTE value */
1715                 valid = !!(v & HPTE_V_VALID);
1716                 dirty = !!(revp->guest_rpte & HPTE_GR_MODIFIED);
1717
1718                 /* Harvest R and C into guest view if necessary */
1719                 rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1720                 if (valid && (rcbits_unset & hr)) {
1721                         revp->guest_rpte |= (hr &
1722                                 (HPTE_R_R | HPTE_R_C)) | HPTE_GR_MODIFIED;
1723                         dirty = 1;
1724                 }
1725
1726                 if (v & HPTE_V_ABSENT) {
1727                         v &= ~HPTE_V_ABSENT;
1728                         v |= HPTE_V_VALID;
1729                         valid = 1;
1730                 }
1731                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) && !(v & HPTE_V_BOLTED))
1732                         valid = 0;
1733
1734                 r = revp->guest_rpte;
1735                 /* only clear modified if this is the right sort of entry */
1736                 if (valid == want_valid && dirty) {
1737                         r &= ~HPTE_GR_MODIFIED;
1738                         revp->guest_rpte = r;
1739                 }
1740                 unlock_hpte(hptp, be64_to_cpu(hptp[0]));
1741                 preempt_enable();
1742                 if (!(valid == want_valid && (first_pass || dirty)))
1743                         ok = 0;
1744         }
1745         hpte[0] = cpu_to_be64(v);
1746         hpte[1] = cpu_to_be64(r);
1747         return ok;
1748 }
1749
1750 static ssize_t kvm_htab_read(struct file *file, char __user *buf,
1751                              size_t count, loff_t *ppos)
1752 {
1753         struct kvm_htab_ctx *ctx = file->private_data;
1754         struct kvm *kvm = ctx->kvm;
1755         struct kvm_get_htab_header hdr;
1756         __be64 *hptp;
1757         struct revmap_entry *revp;
1758         unsigned long i, nb, nw;
1759         unsigned long __user *lbuf;
1760         struct kvm_get_htab_header __user *hptr;
1761         unsigned long flags;
1762         int first_pass;
1763         unsigned long hpte[2];
1764
1765         if (!access_ok(buf, count))
1766                 return -EFAULT;
1767         if (kvm_is_radix(kvm))
1768                 return 0;
1769
1770         first_pass = ctx->first_pass;
1771         flags = ctx->flags;
1772
1773         i = ctx->index;
1774         hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
1775         revp = kvm->arch.hpt.rev + i;
1776         lbuf = (unsigned long __user *)buf;
1777
1778         nb = 0;
1779         while (nb + sizeof(hdr) + HPTE_SIZE < count) {
1780                 /* Initialize header */
1781                 hptr = (struct kvm_get_htab_header __user *)buf;
1782                 hdr.n_valid = 0;
1783                 hdr.n_invalid = 0;
1784                 nw = nb;
1785                 nb += sizeof(hdr);
1786                 lbuf = (unsigned long __user *)(buf + sizeof(hdr));
1787
1788                 /* Skip uninteresting entries, i.e. clean on not-first pass */
1789                 if (!first_pass) {
1790                         while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1791                                !hpte_dirty(revp, hptp)) {
1792                                 ++i;
1793                                 hptp += 2;
1794                                 ++revp;
1795                         }
1796                 }
1797                 hdr.index = i;
1798
1799                 /* Grab a series of valid entries */
1800                 while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1801                        hdr.n_valid < 0xffff &&
1802                        nb + HPTE_SIZE < count &&
1803                        record_hpte(flags, hptp, hpte, revp, 1, first_pass)) {
1804                         /* valid entry, write it out */
1805                         ++hdr.n_valid;
1806                         if (__put_user(hpte[0], lbuf) ||
1807                             __put_user(hpte[1], lbuf + 1))
1808                                 return -EFAULT;
1809                         nb += HPTE_SIZE;
1810                         lbuf += 2;
1811                         ++i;
1812                         hptp += 2;
1813                         ++revp;
1814                 }
1815                 /* Now skip invalid entries while we can */
1816                 while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1817                        hdr.n_invalid < 0xffff &&
1818                        record_hpte(flags, hptp, hpte, revp, 0, first_pass)) {
1819                         /* found an invalid entry */
1820                         ++hdr.n_invalid;
1821                         ++i;
1822                         hptp += 2;
1823                         ++revp;
1824                 }
1825
1826                 if (hdr.n_valid || hdr.n_invalid) {
1827                         /* write back the header */
1828                         if (__copy_to_user(hptr, &hdr, sizeof(hdr)))
1829                                 return -EFAULT;
1830                         nw = nb;
1831                         buf = (char __user *)lbuf;
1832                 } else {
1833                         nb = nw;
1834                 }
1835
1836                 /* Check if we've wrapped around the hash table */
1837                 if (i >= kvmppc_hpt_npte(&kvm->arch.hpt)) {
1838                         i = 0;
1839                         ctx->first_pass = 0;
1840                         break;
1841                 }
1842         }
1843
1844         ctx->index = i;
1845
1846         return nb;
1847 }
1848
1849 static ssize_t kvm_htab_write(struct file *file, const char __user *buf,
1850                               size_t count, loff_t *ppos)
1851 {
1852         struct kvm_htab_ctx *ctx = file->private_data;
1853         struct kvm *kvm = ctx->kvm;
1854         struct kvm_get_htab_header hdr;
1855         unsigned long i, j;
1856         unsigned long v, r;
1857         unsigned long __user *lbuf;
1858         __be64 *hptp;
1859         unsigned long tmp[2];
1860         ssize_t nb;
1861         long int err, ret;
1862         int mmu_ready;
1863         int pshift;
1864
1865         if (!access_ok(buf, count))
1866                 return -EFAULT;
1867         if (kvm_is_radix(kvm))
1868                 return -EINVAL;
1869
1870         /* lock out vcpus from running while we're doing this */
1871         mutex_lock(&kvm->lock);
1872         mmu_ready = kvm->arch.mmu_ready;
1873         if (mmu_ready) {
1874                 kvm->arch.mmu_ready = 0;        /* temporarily */
1875                 /* order mmu_ready vs. vcpus_running */
1876                 smp_mb();
1877                 if (atomic_read(&kvm->arch.vcpus_running)) {
1878                         kvm->arch.mmu_ready = 1;
1879                         mutex_unlock(&kvm->lock);
1880                         return -EBUSY;
1881                 }
1882         }
1883
1884         err = 0;
1885         for (nb = 0; nb + sizeof(hdr) <= count; ) {
1886                 err = -EFAULT;
1887                 if (__copy_from_user(&hdr, buf, sizeof(hdr)))
1888                         break;
1889
1890                 err = 0;
1891                 if (nb + hdr.n_valid * HPTE_SIZE > count)
1892                         break;
1893
1894                 nb += sizeof(hdr);
1895                 buf += sizeof(hdr);
1896
1897                 err = -EINVAL;
1898                 i = hdr.index;
1899                 if (i >= kvmppc_hpt_npte(&kvm->arch.hpt) ||
1900                     i + hdr.n_valid + hdr.n_invalid > kvmppc_hpt_npte(&kvm->arch.hpt))
1901                         break;
1902
1903                 hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
1904                 lbuf = (unsigned long __user *)buf;
1905                 for (j = 0; j < hdr.n_valid; ++j) {
1906                         __be64 hpte_v;
1907                         __be64 hpte_r;
1908
1909                         err = -EFAULT;
1910                         if (__get_user(hpte_v, lbuf) ||
1911                             __get_user(hpte_r, lbuf + 1))
1912                                 goto out;
1913                         v = be64_to_cpu(hpte_v);
1914                         r = be64_to_cpu(hpte_r);
1915                         err = -EINVAL;
1916                         if (!(v & HPTE_V_VALID))
1917                                 goto out;
1918                         pshift = kvmppc_hpte_base_page_shift(v, r);
1919                         if (pshift <= 0)
1920                                 goto out;
1921                         lbuf += 2;
1922                         nb += HPTE_SIZE;
1923
1924                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1925                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1926                         err = -EIO;
1927                         ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, i, v, r,
1928                                                          tmp);
1929                         if (ret != H_SUCCESS) {
1930                                 pr_err("kvm_htab_write ret %ld i=%ld v=%lx "
1931                                        "r=%lx\n", ret, i, v, r);
1932                                 goto out;
1933                         }
1934                         if (!mmu_ready && is_vrma_hpte(v)) {
1935                                 unsigned long senc, lpcr;
1936
1937                                 senc = slb_pgsize_encoding(1ul << pshift);
1938                                 kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
1939                                         (VRMA_VSID << SLB_VSID_SHIFT_1T);
1940                                 if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
1941                                         lpcr = senc << (LPCR_VRMASD_SH - 4);
1942                                         kvmppc_update_lpcr(kvm, lpcr,
1943                                                            LPCR_VRMASD);
1944                                 } else {
1945                                         kvmppc_setup_partition_table(kvm);
1946                                 }
1947                                 mmu_ready = 1;
1948                         }
1949                         ++i;
1950                         hptp += 2;
1951                 }
1952
1953                 for (j = 0; j < hdr.n_invalid; ++j) {
1954                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1955                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1956                         ++i;
1957                         hptp += 2;
1958                 }
1959                 err = 0;
1960         }
1961
1962  out:
1963         /* Order HPTE updates vs. mmu_ready */
1964         smp_wmb();
1965         kvm->arch.mmu_ready = mmu_ready;
1966         mutex_unlock(&kvm->lock);
1967
1968         if (err)
1969                 return err;
1970         return nb;
1971 }
1972
1973 static int kvm_htab_release(struct inode *inode, struct file *filp)
1974 {
1975         struct kvm_htab_ctx *ctx = filp->private_data;
1976
1977         filp->private_data = NULL;
1978         if (!(ctx->flags & KVM_GET_HTAB_WRITE))
1979                 atomic_dec(&ctx->kvm->arch.hpte_mod_interest);
1980         kvm_put_kvm(ctx->kvm);
1981         kfree(ctx);
1982         return 0;
1983 }
1984
1985 static const struct file_operations kvm_htab_fops = {
1986         .read           = kvm_htab_read,
1987         .write          = kvm_htab_write,
1988         .llseek         = default_llseek,
1989         .release        = kvm_htab_release,
1990 };
1991
1992 int kvm_vm_ioctl_get_htab_fd(struct kvm *kvm, struct kvm_get_htab_fd *ghf)
1993 {
1994         int ret;
1995         struct kvm_htab_ctx *ctx;
1996         int rwflag;
1997
1998         /* reject flags we don't recognize */
1999         if (ghf->flags & ~(KVM_GET_HTAB_BOLTED_ONLY | KVM_GET_HTAB_WRITE))
2000                 return -EINVAL;
2001         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
2002         if (!ctx)
2003                 return -ENOMEM;
2004         kvm_get_kvm(kvm);
2005         ctx->kvm = kvm;
2006         ctx->index = ghf->start_index;
2007         ctx->flags = ghf->flags;
2008         ctx->first_pass = 1;
2009
2010         rwflag = (ghf->flags & KVM_GET_HTAB_WRITE) ? O_WRONLY : O_RDONLY;
2011         ret = anon_inode_getfd("kvm-htab", &kvm_htab_fops, ctx, rwflag | O_CLOEXEC);
2012         if (ret < 0) {
2013                 kfree(ctx);
2014                 kvm_put_kvm(kvm);
2015                 return ret;
2016         }
2017
2018         if (rwflag == O_RDONLY) {
2019                 mutex_lock(&kvm->slots_lock);
2020                 atomic_inc(&kvm->arch.hpte_mod_interest);
2021                 /* make sure kvmppc_do_h_enter etc. see the increment */
2022                 synchronize_srcu_expedited(&kvm->srcu);
2023                 mutex_unlock(&kvm->slots_lock);
2024         }
2025
2026         return ret;
2027 }
2028
2029 struct debugfs_htab_state {
2030         struct kvm      *kvm;
2031         struct mutex    mutex;
2032         unsigned long   hpt_index;
2033         int             chars_left;
2034         int             buf_index;
2035         char            buf[64];
2036 };
2037
2038 static int debugfs_htab_open(struct inode *inode, struct file *file)
2039 {
2040         struct kvm *kvm = inode->i_private;
2041         struct debugfs_htab_state *p;
2042
2043         p = kzalloc(sizeof(*p), GFP_KERNEL);
2044         if (!p)
2045                 return -ENOMEM;
2046
2047         kvm_get_kvm(kvm);
2048         p->kvm = kvm;
2049         mutex_init(&p->mutex);
2050         file->private_data = p;
2051
2052         return nonseekable_open(inode, file);
2053 }
2054
2055 static int debugfs_htab_release(struct inode *inode, struct file *file)
2056 {
2057         struct debugfs_htab_state *p = file->private_data;
2058
2059         kvm_put_kvm(p->kvm);
2060         kfree(p);
2061         return 0;
2062 }
2063
2064 static ssize_t debugfs_htab_read(struct file *file, char __user *buf,
2065                                  size_t len, loff_t *ppos)
2066 {
2067         struct debugfs_htab_state *p = file->private_data;
2068         ssize_t ret, r;
2069         unsigned long i, n;
2070         unsigned long v, hr, gr;
2071         struct kvm *kvm;
2072         __be64 *hptp;
2073
2074         kvm = p->kvm;
2075         if (kvm_is_radix(kvm))
2076                 return 0;
2077
2078         ret = mutex_lock_interruptible(&p->mutex);
2079         if (ret)
2080                 return ret;
2081
2082         if (p->chars_left) {
2083                 n = p->chars_left;
2084                 if (n > len)
2085                         n = len;
2086                 r = copy_to_user(buf, p->buf + p->buf_index, n);
2087                 n -= r;
2088                 p->chars_left -= n;
2089                 p->buf_index += n;
2090                 buf += n;
2091                 len -= n;
2092                 ret = n;
2093                 if (r) {
2094                         if (!n)
2095                                 ret = -EFAULT;
2096                         goto out;
2097                 }
2098         }
2099
2100         i = p->hpt_index;
2101         hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
2102         for (; len != 0 && i < kvmppc_hpt_npte(&kvm->arch.hpt);
2103              ++i, hptp += 2) {
2104                 if (!(be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)))
2105                         continue;
2106
2107                 /* lock the HPTE so it's stable and read it */
2108                 preempt_disable();
2109                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
2110                         cpu_relax();
2111                 v = be64_to_cpu(hptp[0]) & ~HPTE_V_HVLOCK;
2112                 hr = be64_to_cpu(hptp[1]);
2113                 gr = kvm->arch.hpt.rev[i].guest_rpte;
2114                 unlock_hpte(hptp, v);
2115                 preempt_enable();
2116
2117                 if (!(v & (HPTE_V_VALID | HPTE_V_ABSENT)))
2118                         continue;
2119
2120                 n = scnprintf(p->buf, sizeof(p->buf),
2121                               "%6lx %.16lx %.16lx %.16lx\n",
2122                               i, v, hr, gr);
2123                 p->chars_left = n;
2124                 if (n > len)
2125                         n = len;
2126                 r = copy_to_user(buf, p->buf, n);
2127                 n -= r;
2128                 p->chars_left -= n;
2129                 p->buf_index = n;
2130                 buf += n;
2131                 len -= n;
2132                 ret += n;
2133                 if (r) {
2134                         if (!ret)
2135                                 ret = -EFAULT;
2136                         goto out;
2137                 }
2138         }
2139         p->hpt_index = i;
2140
2141  out:
2142         mutex_unlock(&p->mutex);
2143         return ret;
2144 }
2145
2146 static ssize_t debugfs_htab_write(struct file *file, const char __user *buf,
2147                            size_t len, loff_t *ppos)
2148 {
2149         return -EACCES;
2150 }
2151
2152 static const struct file_operations debugfs_htab_fops = {
2153         .owner   = THIS_MODULE,
2154         .open    = debugfs_htab_open,
2155         .release = debugfs_htab_release,
2156         .read    = debugfs_htab_read,
2157         .write   = debugfs_htab_write,
2158         .llseek  = generic_file_llseek,
2159 };
2160
2161 void kvmppc_mmu_debugfs_init(struct kvm *kvm)
2162 {
2163         kvm->arch.htab_dentry = debugfs_create_file("htab", 0400,
2164                                                     kvm->arch.debugfs_dir, kvm,
2165                                                     &debugfs_htab_fops);
2166 }
2167
2168 void kvmppc_mmu_book3s_hv_init(struct kvm_vcpu *vcpu)
2169 {
2170         struct kvmppc_mmu *mmu = &vcpu->arch.mmu;
2171
2172         vcpu->arch.slb_nr = 32;         /* POWER7/POWER8 */
2173
2174         mmu->xlate = kvmppc_mmu_book3s_64_hv_xlate;
2175         mmu->reset_msr = kvmppc_mmu_book3s_64_hv_reset_msr;
2176
2177         vcpu->arch.hflags |= BOOK3S_HFLAG_SLB;
2178 }