arch/powerpc/kvm/book3s_64_mmu_hv.c

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