arch/powerpc/kvm/book3s_hv_rm_mmu.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  * Copyright 2010-2011 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
   7  */
   8
   9 #include <linux/types.h>
  10 #include <linux/string.h>
  11 #include <linux/kvm.h>
  12 #include <linux/kvm_host.h>
  13 #include <linux/hugetlb.h>
  14 #include <linux/module.h>
  15 #include <linux/log2.h>
  16
  17 #include <asm/tlbflush.h>
  18 #include <asm/kvm_ppc.h>
  19 #include <asm/kvm_book3s.h>
  20 #include <asm/mmu-hash64.h>
  21 #include <asm/hvcall.h>
  22 #include <asm/synch.h>
  23 #include <asm/ppc-opcode.h>
  24
  25 /* Translate address of a vmalloc'd thing to a linear map address */
  26 static void *real_vmalloc_addr(void *x)
  27 {
  28         unsigned long addr = (unsigned long) x;
  29         pte_t *p;
  30         /*
  31          * assume we don't have huge pages in vmalloc space...
  32          * So don't worry about THP collapse/split. Called
  33          * Only in realmode, hence won't need irq_save/restore.
  34          */
  35         p = __find_linux_pte_or_hugepte(swapper_pg_dir, addr, NULL, NULL);
  36         if (!p || !pte_present(*p))
  37                 return NULL;
  38         addr = (pte_pfn(*p) << PAGE_SHIFT) | (addr & ~PAGE_MASK);
  39         return __va(addr);
  40 }
  41
  42 /* Return 1 if we need to do a global tlbie, 0 if we can use tlbiel */
  43 static int global_invalidates(struct kvm *kvm, unsigned long flags)
  44 {
  45         int global;
  46
  47         /*
  48          * If there is only one vcore, and it's currently running,
  49          * as indicated by local_paca->kvm_hstate.kvm_vcpu being set,
  50          * we can use tlbiel as long as we mark all other physical
  51          * cores as potentially having stale TLB entries for this lpid.
  52          * Otherwise, don't use tlbiel.
  53          */
  54         if (kvm->arch.online_vcores == 1 && local_paca->kvm_hstate.kvm_vcpu)
  55                 global = 0;
  56         else
  57                 global = 1;
  58
  59         if (!global) {
  60                 /* any other core might now have stale TLB entries... */
  61                 smp_wmb();
  62                 cpumask_setall(&kvm->arch.need_tlb_flush);
  63                 cpumask_clear_cpu(local_paca->kvm_hstate.kvm_vcore->pcpu,
  64                                   &kvm->arch.need_tlb_flush);
  65         }
  66
  67         return global;
  68 }
  69
  70 /*
  71  * Add this HPTE into the chain for the real page.
  72  * Must be called with the chain locked; it unlocks the chain.
  73  */
  74 void kvmppc_add_revmap_chain(struct kvm *kvm, struct revmap_entry *rev,
  75                              unsigned long *rmap, long pte_index, int realmode)
  76 {
  77         struct revmap_entry *head, *tail;
  78         unsigned long i;
  79
  80         if (*rmap & KVMPPC_RMAP_PRESENT) {
  81                 i = *rmap & KVMPPC_RMAP_INDEX;
  82                 head = &kvm->arch.revmap[i];
  83                 if (realmode)
  84                         head = real_vmalloc_addr(head);
  85                 tail = &kvm->arch.revmap[head->back];
  86                 if (realmode)
  87                         tail = real_vmalloc_addr(tail);
  88                 rev->forw = i;
  89                 rev->back = head->back;
  90                 tail->forw = pte_index;
  91                 head->back = pte_index;
  92         } else {
  93                 rev->forw = rev->back = pte_index;
  94                 *rmap = (*rmap & ~KVMPPC_RMAP_INDEX) |
  95                         pte_index | KVMPPC_RMAP_PRESENT;
  96         }
  97         unlock_rmap(rmap);
  98 }
  99 EXPORT_SYMBOL_GPL(kvmppc_add_revmap_chain);
 100
 101 /* Update the changed page order field of an rmap entry */
 102 void kvmppc_update_rmap_change(unsigned long *rmap, unsigned long psize)
 103 {
 104         unsigned long order;
 105
 106         if (!psize)
 107                 return;
 108         order = ilog2(psize);
 109         order <<= KVMPPC_RMAP_CHG_SHIFT;
 110         if (order > (*rmap & KVMPPC_RMAP_CHG_ORDER))
 111                 *rmap = (*rmap & ~KVMPPC_RMAP_CHG_ORDER) | order;
 112 }
 113 EXPORT_SYMBOL_GPL(kvmppc_update_rmap_change);
 114
 115 /* Returns a pointer to the revmap entry for the page mapped by a HPTE */
 116 static unsigned long *revmap_for_hpte(struct kvm *kvm, unsigned long hpte_v,
 117                                       unsigned long hpte_gr)
 118 {
 119         struct kvm_memory_slot *memslot;
 120         unsigned long *rmap;
 121         unsigned long gfn;
 122
 123         gfn = hpte_rpn(hpte_gr, hpte_page_size(hpte_v, hpte_gr));
 124         memslot = __gfn_to_memslot(kvm_memslots_raw(kvm), gfn);
 125         if (!memslot)
 126                 return NULL;
 127
 128         rmap = real_vmalloc_addr(&memslot->arch.rmap[gfn - memslot->base_gfn]);
 129         return rmap;
 130 }
 131
 132 /* Remove this HPTE from the chain for a real page */
 133 static void remove_revmap_chain(struct kvm *kvm, long pte_index,
 134                                 struct revmap_entry *rev,
 135                                 unsigned long hpte_v, unsigned long hpte_r)
 136 {
 137         struct revmap_entry *next, *prev;
 138         unsigned long ptel, head;
 139         unsigned long *rmap;
 140         unsigned long rcbits;
 141
 142         rcbits = hpte_r & (HPTE_R_R | HPTE_R_C);
 143         ptel = rev->guest_rpte |= rcbits;
 144         rmap = revmap_for_hpte(kvm, hpte_v, ptel);
 145         if (!rmap)
 146                 return;
 147         lock_rmap(rmap);
 148
 149         head = *rmap & KVMPPC_RMAP_INDEX;
 150         next = real_vmalloc_addr(&kvm->arch.revmap[rev->forw]);
 151         prev = real_vmalloc_addr(&kvm->arch.revmap[rev->back]);
 152         next->back = rev->back;
 153         prev->forw = rev->forw;
 154         if (head == pte_index) {
 155                 head = rev->forw;
 156                 if (head == pte_index)
 157                         *rmap &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
 158                 else
 159                         *rmap = (*rmap & ~KVMPPC_RMAP_INDEX) | head;
 160         }
 161         *rmap |= rcbits << KVMPPC_RMAP_RC_SHIFT;
 162         if (rcbits & HPTE_R_C)
 163                 kvmppc_update_rmap_change(rmap, hpte_page_size(hpte_v, hpte_r));
 164         unlock_rmap(rmap);
 165 }
 166
 167 long kvmppc_do_h_enter(struct kvm *kvm, unsigned long flags,
 168                        long pte_index, unsigned long pteh, unsigned long ptel,
 169                        pgd_t *pgdir, bool realmode, unsigned long *pte_idx_ret)
 170 {
 171         unsigned long i, pa, gpa, gfn, psize;
 172         unsigned long slot_fn, hva;
 173         __be64 *hpte;
 174         struct revmap_entry *rev;
 175         unsigned long g_ptel;
 176         struct kvm_memory_slot *memslot;
 177         unsigned hpage_shift;
 178         unsigned long is_io;
 179         unsigned long *rmap;
 180         pte_t *ptep;
 181         unsigned int writing;
 182         unsigned long mmu_seq;
 183         unsigned long rcbits, irq_flags = 0;
 184
 185         psize = hpte_page_size(pteh, ptel);
 186         if (!psize)
 187                 return H_PARAMETER;
 188         writing = hpte_is_writable(ptel);
 189         pteh &= ~(HPTE_V_HVLOCK | HPTE_V_ABSENT | HPTE_V_VALID);
 190         ptel &= ~HPTE_GR_RESERVED;
 191         g_ptel = ptel;
 192
 193         /* used later to detect if we might have been invalidated */
 194         mmu_seq = kvm->mmu_notifier_seq;
 195         smp_rmb();
 196
 197         /* Find the memslot (if any) for this address */
 198         gpa = (ptel & HPTE_R_RPN) & ~(psize - 1);
 199         gfn = gpa >> PAGE_SHIFT;
 200         memslot = __gfn_to_memslot(kvm_memslots_raw(kvm), gfn);
 201         pa = 0;
 202         is_io = ~0ul;
 203         rmap = NULL;
 204         if (!(memslot && !(memslot->flags & KVM_MEMSLOT_INVALID))) {
 205                 /* Emulated MMIO - mark this with key=31 */
 206                 pteh |= HPTE_V_ABSENT;
 207                 ptel |= HPTE_R_KEY_HI | HPTE_R_KEY_LO;
 208                 goto do_insert;
 209         }
 210
 211         /* Check if the requested page fits entirely in the memslot. */
 212         if (!slot_is_aligned(memslot, psize))
 213                 return H_PARAMETER;
 214         slot_fn = gfn - memslot->base_gfn;
 215         rmap = &memslot->arch.rmap[slot_fn];
 216
 217         /* Translate to host virtual address */
 218         hva = __gfn_to_hva_memslot(memslot, gfn);
 219         /*
 220          * If we had a page table table change after lookup, we would
 221          * retry via mmu_notifier_retry.
 222          */
 223         if (realmode)
 224                 ptep = __find_linux_pte_or_hugepte(pgdir, hva, NULL,
 225                                                    &hpage_shift);
 226         else {
 227                 local_irq_save(irq_flags);
 228                 ptep = find_linux_pte_or_hugepte(pgdir, hva, NULL,
 229                                                  &hpage_shift);
 230         }
 231         if (ptep) {
 232                 pte_t pte;
 233                 unsigned int host_pte_size;
 234
 235                 if (hpage_shift)
 236                         host_pte_size = 1ul << hpage_shift;
 237                 else
 238                         host_pte_size = PAGE_SIZE;
 239                 /*
 240                  * We should always find the guest page size
 241                  * to <= host page size, if host is using hugepage
 242                  */
 243                 if (host_pte_size < psize) {
 244                         if (!realmode)
 245                                 local_irq_restore(flags);
 246                         return H_PARAMETER;
 247                 }
 248                 pte = kvmppc_read_update_linux_pte(ptep, writing);
 249                 if (pte_present(pte) && !pte_protnone(pte)) {
 250                         if (writing && !pte_write(pte))
 251                                 /* make the actual HPTE be read-only */
 252                                 ptel = hpte_make_readonly(ptel);
 253                         is_io = hpte_cache_bits(pte_val(pte));
 254                         pa = pte_pfn(pte) << PAGE_SHIFT;
 255                         pa |= hva & (host_pte_size - 1);
 256                         pa |= gpa & ~PAGE_MASK;
 257                 }
 258         }
 259         if (!realmode)
 260                 local_irq_restore(irq_flags);
 261
 262         ptel &= ~(HPTE_R_PP0 - psize);
 263         ptel |= pa;
 264
 265         if (pa)
 266                 pteh |= HPTE_V_VALID;
 267         else
 268                 pteh |= HPTE_V_ABSENT;
 269
 270         /* Check WIMG */
 271         if (is_io != ~0ul && !hpte_cache_flags_ok(ptel, is_io)) {
 272                 if (is_io)
 273                         return H_PARAMETER;
 274                 /*
 275                  * Allow guest to map emulated device memory as
 276                  * uncacheable, but actually make it cacheable.
 277                  */
 278                 ptel &= ~(HPTE_R_W|HPTE_R_I|HPTE_R_G);
 279                 ptel |= HPTE_R_M;
 280         }
 281
 282         /* Find and lock the HPTEG slot to use */
 283  do_insert:
 284         if (pte_index >= kvm->arch.hpt_npte)
 285                 return H_PARAMETER;
 286         if (likely((flags & H_EXACT) == 0)) {
 287                 pte_index &= ~7UL;
 288                 hpte = (__be64 *)(kvm->arch.hpt_virt + (pte_index << 4));
 289                 for (i = 0; i < 8; ++i) {
 290                         if ((be64_to_cpu(*hpte) & HPTE_V_VALID) == 0 &&
 291                             try_lock_hpte(hpte, HPTE_V_HVLOCK | HPTE_V_VALID |
 292                                           HPTE_V_ABSENT))
 293                                 break;
 294                         hpte += 2;
 295                 }
 296                 if (i == 8) {
 297                         /*
 298                          * Since try_lock_hpte doesn't retry (not even stdcx.
 299                          * failures), it could be that there is a free slot
 300                          * but we transiently failed to lock it.  Try again,
 301                          * actually locking each slot and checking it.
 302                          */
 303                         hpte -= 16;
 304                         for (i = 0; i < 8; ++i) {
 305                                 u64 pte;
 306                                 while (!try_lock_hpte(hpte, HPTE_V_HVLOCK))
 307                                         cpu_relax();
 308                                 pte = be64_to_cpu(hpte[0]);
 309                                 if (!(pte & (HPTE_V_VALID | HPTE_V_ABSENT)))
 310                                         break;
 311                                 __unlock_hpte(hpte, pte);
 312                                 hpte += 2;
 313                         }
 314                         if (i == 8)
 315                                 return H_PTEG_FULL;
 316                 }
 317                 pte_index += i;
 318         } else {
 319                 hpte = (__be64 *)(kvm->arch.hpt_virt + (pte_index << 4));
 320                 if (!try_lock_hpte(hpte, HPTE_V_HVLOCK | HPTE_V_VALID |
 321                                    HPTE_V_ABSENT)) {
 322                         /* Lock the slot and check again */
 323                         u64 pte;
 324
 325                         while (!try_lock_hpte(hpte, HPTE_V_HVLOCK))
 326                                 cpu_relax();
 327                         pte = be64_to_cpu(hpte[0]);
 328                         if (pte & (HPTE_V_VALID | HPTE_V_ABSENT)) {
 329                                 __unlock_hpte(hpte, pte);
 330                                 return H_PTEG_FULL;
 331                         }
 332                 }
 333         }
 334
 335         /* Save away the guest's idea of the second HPTE dword */
 336         rev = &kvm->arch.revmap[pte_index];
 337         if (realmode)
 338                 rev = real_vmalloc_addr(rev);
 339         if (rev) {
 340                 rev->guest_rpte = g_ptel;
 341                 note_hpte_modification(kvm, rev);
 342         }
 343
 344         /* Link HPTE into reverse-map chain */
 345         if (pteh & HPTE_V_VALID) {
 346                 if (realmode)
 347                         rmap = real_vmalloc_addr(rmap);
 348                 lock_rmap(rmap);
 349                 /* Check for pending invalidations under the rmap chain lock */
 350                 if (mmu_notifier_retry(kvm, mmu_seq)) {
 351                         /* inval in progress, write a non-present HPTE */
 352                         pteh |= HPTE_V_ABSENT;
 353                         pteh &= ~HPTE_V_VALID;
 354                         unlock_rmap(rmap);
 355                 } else {
 356                         kvmppc_add_revmap_chain(kvm, rev, rmap, pte_index,
 357                                                 realmode);
 358                         /* Only set R/C in real HPTE if already set in *rmap */
 359                         rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
 360                         ptel &= rcbits | ~(HPTE_R_R | HPTE_R_C);
 361                 }
 362         }
 363
 364         hpte[1] = cpu_to_be64(ptel);
 365
 366         /* Write the first HPTE dword, unlocking the HPTE and making it valid */
 367         eieio();
 368         __unlock_hpte(hpte, pteh);
 369         asm volatile("ptesync" : : : "memory");
 370
 371         *pte_idx_ret = pte_index;
 372         return H_SUCCESS;
 373 }
 374 EXPORT_SYMBOL_GPL(kvmppc_do_h_enter);
 375
 376 long kvmppc_h_enter(struct kvm_vcpu *vcpu, unsigned long flags,
 377                     long pte_index, unsigned long pteh, unsigned long ptel)
 378 {
 379         return kvmppc_do_h_enter(vcpu->kvm, flags, pte_index, pteh, ptel,
 380                                  vcpu->arch.pgdir, true, &vcpu->arch.gpr[4]);
 381 }
 382
 383 #ifdef __BIG_ENDIAN__
 384 #define LOCK_TOKEN      (*(u32 *)(&get_paca()->lock_token))
 385 #else
 386 #define LOCK_TOKEN      (*(u32 *)(&get_paca()->paca_index))
 387 #endif
 388
 389 static inline int try_lock_tlbie(unsigned int *lock)
 390 {
 391         unsigned int tmp, old;
 392         unsigned int token = LOCK_TOKEN;
 393
 394         asm volatile("1:lwarx   %1,0,%2\n"
 395                      "  cmpwi   cr0,%1,0\n"
 396                      "  bne     2f\n"
 397                      "  stwcx.  %3,0,%2\n"
 398                      "  bne-    1b\n"
 399                      "  isync\n"
 400                      "2:"
 401                      : "=&r" (tmp), "=&r" (old)
 402                      : "r" (lock), "r" (token)
 403                      : "cc", "memory");
 404         return old == 0;
 405 }
 406
 407 static void do_tlbies(struct kvm *kvm, unsigned long *rbvalues,
 408                       long npages, int global, bool need_sync)
 409 {
 410         long i;
 411
 412         if (global) {
 413                 while (!try_lock_tlbie(&kvm->arch.tlbie_lock))
 414                         cpu_relax();
 415                 if (need_sync)
 416                         asm volatile("ptesync" : : : "memory");
 417                 for (i = 0; i < npages; ++i)
 418                         asm volatile(PPC_TLBIE(%1,%0) : :
 419                                      "r" (rbvalues[i]), "r" (kvm->arch.lpid));
 420                 asm volatile("eieio; tlbsync; ptesync" : : : "memory");
 421                 kvm->arch.tlbie_lock = 0;
 422         } else {
 423                 if (need_sync)
 424                         asm volatile("ptesync" : : : "memory");
 425                 for (i = 0; i < npages; ++i)
 426                         asm volatile("tlbiel %0" : : "r" (rbvalues[i]));
 427                 asm volatile("ptesync" : : : "memory");
 428         }
 429 }
 430
 431 long kvmppc_do_h_remove(struct kvm *kvm, unsigned long flags,
 432                         unsigned long pte_index, unsigned long avpn,
 433                         unsigned long *hpret)
 434 {
 435         __be64 *hpte;
 436         unsigned long v, r, rb;
 437         struct revmap_entry *rev;
 438         u64 pte;
 439
 440         if (pte_index >= kvm->arch.hpt_npte)
 441                 return H_PARAMETER;
 442         hpte = (__be64 *)(kvm->arch.hpt_virt + (pte_index << 4));
 443         while (!try_lock_hpte(hpte, HPTE_V_HVLOCK))
 444                 cpu_relax();
 445         pte = be64_to_cpu(hpte[0]);
 446         if ((pte & (HPTE_V_ABSENT | HPTE_V_VALID)) == 0 ||
 447             ((flags & H_AVPN) && (pte & ~0x7fUL) != avpn) ||
 448             ((flags & H_ANDCOND) && (pte & avpn) != 0)) {
 449                 __unlock_hpte(hpte, pte);
 450                 return H_NOT_FOUND;
 451         }
 452
 453         rev = real_vmalloc_addr(&kvm->arch.revmap[pte_index]);
 454         v = pte & ~HPTE_V_HVLOCK;
 455         if (v & HPTE_V_VALID) {
 456                 hpte[0] &= ~cpu_to_be64(HPTE_V_VALID);
 457                 rb = compute_tlbie_rb(v, be64_to_cpu(hpte[1]), pte_index);
 458                 do_tlbies(kvm, &rb, 1, global_invalidates(kvm, flags), true);
 459                 /*
 460                  * The reference (R) and change (C) bits in a HPT
 461                  * entry can be set by hardware at any time up until
 462                  * the HPTE is invalidated and the TLB invalidation
 463                  * sequence has completed.  This means that when
 464                  * removing a HPTE, we need to re-read the HPTE after
 465                  * the invalidation sequence has completed in order to
 466                  * obtain reliable values of R and C.
 467                  */
 468                 remove_revmap_chain(kvm, pte_index, rev, v,
 469                                     be64_to_cpu(hpte[1]));
 470         }
 471         r = rev->guest_rpte & ~HPTE_GR_RESERVED;
 472         note_hpte_modification(kvm, rev);
 473         unlock_hpte(hpte, 0);
 474
 475         if (v & HPTE_V_ABSENT)
 476                 v = (v & ~HPTE_V_ABSENT) | HPTE_V_VALID;
 477         hpret[0] = v;
 478         hpret[1] = r;
 479         return H_SUCCESS;
 480 }
 481 EXPORT_SYMBOL_GPL(kvmppc_do_h_remove);
 482
 483 long kvmppc_h_remove(struct kvm_vcpu *vcpu, unsigned long flags,
 484                      unsigned long pte_index, unsigned long avpn)
 485 {
 486         return kvmppc_do_h_remove(vcpu->kvm, flags, pte_index, avpn,
 487                                   &vcpu->arch.gpr[4]);
 488 }
 489
 490 long kvmppc_h_bulk_remove(struct kvm_vcpu *vcpu)
 491 {
 492         struct kvm *kvm = vcpu->kvm;
 493         unsigned long *args = &vcpu->arch.gpr[4];
 494         __be64 *hp, *hptes[4];
 495         unsigned long tlbrb[4];
 496         long int i, j, k, n, found, indexes[4];
 497         unsigned long flags, req, pte_index, rcbits;
 498         int global;
 499         long int ret = H_SUCCESS;
 500         struct revmap_entry *rev, *revs[4];
 501         u64 hp0;
 502
 503         global = global_invalidates(kvm, 0);
 504         for (i = 0; i < 4 && ret == H_SUCCESS; ) {
 505                 n = 0;
 506                 for (; i < 4; ++i) {
 507                         j = i * 2;
 508                         pte_index = args[j];
 509                         flags = pte_index >> 56;
 510                         pte_index &= ((1ul << 56) - 1);
 511                         req = flags >> 6;
 512                         flags &= 3;
 513                         if (req == 3) {         /* no more requests */
 514                                 i = 4;
 515                                 break;
 516                         }
 517                         if (req != 1 || flags == 3 ||
 518                             pte_index >= kvm->arch.hpt_npte) {
 519                                 /* parameter error */
 520                                 args[j] = ((0xa0 | flags) << 56) + pte_index;
 521                                 ret = H_PARAMETER;
 522                                 break;
 523                         }
 524                         hp = (__be64 *) (kvm->arch.hpt_virt + (pte_index << 4));
 525                         /* to avoid deadlock, don't spin except for first */
 526                         if (!try_lock_hpte(hp, HPTE_V_HVLOCK)) {
 527                                 if (n)
 528                                         break;
 529                                 while (!try_lock_hpte(hp, HPTE_V_HVLOCK))
 530                                         cpu_relax();
 531                         }
 532                         found = 0;
 533                         hp0 = be64_to_cpu(hp[0]);
 534                         if (hp0 & (HPTE_V_ABSENT | HPTE_V_VALID)) {
 535                                 switch (flags & 3) {
 536                                 case 0:         /* absolute */
 537                                         found = 1;
 538                                         break;
 539                                 case 1:         /* andcond */
 540                                         if (!(hp0 & args[j + 1]))
 541                                                 found = 1;
 542                                         break;
 543                                 case 2:         /* AVPN */
 544                                         if ((hp0 & ~0x7fUL) == args[j + 1])
 545                                                 found = 1;
 546                                         break;
 547                                 }
 548                         }
 549                         if (!found) {
 550                                 hp[0] &= ~cpu_to_be64(HPTE_V_HVLOCK);
 551                                 args[j] = ((0x90 | flags) << 56) + pte_index;
 552                                 continue;
 553                         }
 554
 555                         args[j] = ((0x80 | flags) << 56) + pte_index;
 556                         rev = real_vmalloc_addr(&kvm->arch.revmap[pte_index]);
 557                         note_hpte_modification(kvm, rev);
 558
 559                         if (!(hp0 & HPTE_V_VALID)) {
 560                                 /* insert R and C bits from PTE */
 561                                 rcbits = rev->guest_rpte & (HPTE_R_R|HPTE_R_C);
 562                                 args[j] |= rcbits << (56 - 5);
 563                                 hp[0] = 0;
 564                                 continue;
 565                         }
 566
 567                         /* leave it locked */
 568                         hp[0] &= ~cpu_to_be64(HPTE_V_VALID);
 569                         tlbrb[n] = compute_tlbie_rb(be64_to_cpu(hp[0]),
 570                                 be64_to_cpu(hp[1]), pte_index);
 571                         indexes[n] = j;
 572                         hptes[n] = hp;
 573                         revs[n] = rev;
 574                         ++n;
 575                 }
 576
 577                 if (!n)
 578                         break;
 579
 580                 /* Now that we've collected a batch, do the tlbies */
 581                 do_tlbies(kvm, tlbrb, n, global, true);
 582
 583                 /* Read PTE low words after tlbie to get final R/C values */
 584                 for (k = 0; k < n; ++k) {
 585                         j = indexes[k];
 586                         pte_index = args[j] & ((1ul << 56) - 1);
 587                         hp = hptes[k];
 588                         rev = revs[k];
 589                         remove_revmap_chain(kvm, pte_index, rev,
 590                                 be64_to_cpu(hp[0]), be64_to_cpu(hp[1]));
 591                         rcbits = rev->guest_rpte & (HPTE_R_R|HPTE_R_C);
 592                         args[j] |= rcbits << (56 - 5);
 593                         __unlock_hpte(hp, 0);
 594                 }
 595         }
 596
 597         return ret;
 598 }
 599
 600 long kvmppc_h_protect(struct kvm_vcpu *vcpu, unsigned long flags,
 601                       unsigned long pte_index, unsigned long avpn,
 602                       unsigned long va)
 603 {
 604         struct kvm *kvm = vcpu->kvm;
 605         __be64 *hpte;
 606         struct revmap_entry *rev;
 607         unsigned long v, r, rb, mask, bits;
 608         u64 pte;
 609
 610         if (pte_index >= kvm->arch.hpt_npte)
 611                 return H_PARAMETER;
 612
 613         hpte = (__be64 *)(kvm->arch.hpt_virt + (pte_index << 4));
 614         while (!try_lock_hpte(hpte, HPTE_V_HVLOCK))
 615                 cpu_relax();
 616         pte = be64_to_cpu(hpte[0]);
 617         if ((pte & (HPTE_V_ABSENT | HPTE_V_VALID)) == 0 ||
 618             ((flags & H_AVPN) && (pte & ~0x7fUL) != avpn)) {
 619                 __unlock_hpte(hpte, pte);
 620                 return H_NOT_FOUND;
 621         }
 622
 623         v = pte;
 624         bits = (flags << 55) & HPTE_R_PP0;
 625         bits |= (flags << 48) & HPTE_R_KEY_HI;
 626         bits |= flags & (HPTE_R_PP | HPTE_R_N | HPTE_R_KEY_LO);
 627
 628         /* Update guest view of 2nd HPTE dword */
 629         mask = HPTE_R_PP0 | HPTE_R_PP | HPTE_R_N |
 630                 HPTE_R_KEY_HI | HPTE_R_KEY_LO;
 631         rev = real_vmalloc_addr(&kvm->arch.revmap[pte_index]);
 632         if (rev) {
 633                 r = (rev->guest_rpte & ~mask) | bits;
 634                 rev->guest_rpte = r;
 635                 note_hpte_modification(kvm, rev);
 636         }
 637
 638         /* Update HPTE */
 639         if (v & HPTE_V_VALID) {
 640                 /*
 641                  * If the page is valid, don't let it transition from
 642                  * readonly to writable.  If it should be writable, we'll
 643                  * take a trap and let the page fault code sort it out.
 644                  */
 645                 pte = be64_to_cpu(hpte[1]);
 646                 r = (pte & ~mask) | bits;
 647                 if (hpte_is_writable(r) && !hpte_is_writable(pte))
 648                         r = hpte_make_readonly(r);
 649                 /* If the PTE is changing, invalidate it first */
 650                 if (r != pte) {
 651                         rb = compute_tlbie_rb(v, r, pte_index);
 652                         hpte[0] = cpu_to_be64((v & ~HPTE_V_VALID) |
 653                                               HPTE_V_ABSENT);
 654                         do_tlbies(kvm, &rb, 1, global_invalidates(kvm, flags),
 655                                   true);
 656                         hpte[1] = cpu_to_be64(r);
 657                 }
 658         }
 659         unlock_hpte(hpte, v & ~HPTE_V_HVLOCK);
 660         asm volatile("ptesync" : : : "memory");
 661         return H_SUCCESS;
 662 }
 663
 664 long kvmppc_h_read(struct kvm_vcpu *vcpu, unsigned long flags,
 665                    unsigned long pte_index)
 666 {
 667         struct kvm *kvm = vcpu->kvm;
 668         __be64 *hpte;
 669         unsigned long v, r;
 670         int i, n = 1;
 671         struct revmap_entry *rev = NULL;
 672
 673         if (pte_index >= kvm->arch.hpt_npte)
 674                 return H_PARAMETER;
 675         if (flags & H_READ_4) {
 676                 pte_index &= ~3;
 677                 n = 4;
 678         }
 679         rev = real_vmalloc_addr(&kvm->arch.revmap[pte_index]);
 680         for (i = 0; i < n; ++i, ++pte_index) {
 681                 hpte = (__be64 *)(kvm->arch.hpt_virt + (pte_index << 4));
 682                 v = be64_to_cpu(hpte[0]) & ~HPTE_V_HVLOCK;
 683                 r = be64_to_cpu(hpte[1]);
 684                 if (v & HPTE_V_ABSENT) {
 685                         v &= ~HPTE_V_ABSENT;
 686                         v |= HPTE_V_VALID;
 687                 }
 688                 if (v & HPTE_V_VALID) {
 689                         r = rev[i].guest_rpte | (r & (HPTE_R_R | HPTE_R_C));
 690                         r &= ~HPTE_GR_RESERVED;
 691                 }
 692                 vcpu->arch.gpr[4 + i * 2] = v;
 693                 vcpu->arch.gpr[5 + i * 2] = r;
 694         }
 695         return H_SUCCESS;
 696 }
 697
 698 long kvmppc_h_clear_ref(struct kvm_vcpu *vcpu, unsigned long flags,
 699                         unsigned long pte_index)
 700 {
 701         struct kvm *kvm = vcpu->kvm;
 702         __be64 *hpte;
 703         unsigned long v, r, gr;
 704         struct revmap_entry *rev;
 705         unsigned long *rmap;
 706         long ret = H_NOT_FOUND;
 707
 708         if (pte_index >= kvm->arch.hpt_npte)
 709                 return H_PARAMETER;
 710
 711         rev = real_vmalloc_addr(&kvm->arch.revmap[pte_index]);
 712         hpte = (__be64 *)(kvm->arch.hpt_virt + (pte_index << 4));
 713         while (!try_lock_hpte(hpte, HPTE_V_HVLOCK))
 714                 cpu_relax();
 715         v = be64_to_cpu(hpte[0]);
 716         r = be64_to_cpu(hpte[1]);
 717         if (!(v & (HPTE_V_VALID | HPTE_V_ABSENT)))
 718                 goto out;
 719
 720         gr = rev->guest_rpte;
 721         if (rev->guest_rpte & HPTE_R_R) {
 722                 rev->guest_rpte &= ~HPTE_R_R;
 723                 note_hpte_modification(kvm, rev);
 724         }
 725         if (v & HPTE_V_VALID) {
 726                 gr |= r & (HPTE_R_R | HPTE_R_C);
 727                 if (r & HPTE_R_R) {
 728                         kvmppc_clear_ref_hpte(kvm, hpte, pte_index);
 729                         rmap = revmap_for_hpte(kvm, v, gr);
 730                         if (rmap) {
 731                                 lock_rmap(rmap);
 732                                 *rmap |= KVMPPC_RMAP_REFERENCED;
 733                                 unlock_rmap(rmap);
 734                         }
 735                 }
 736         }
 737         vcpu->arch.gpr[4] = gr;
 738         ret = H_SUCCESS;
 739  out:
 740         unlock_hpte(hpte, v & ~HPTE_V_HVLOCK);
 741         return ret;
 742 }
 743
 744 long kvmppc_h_clear_mod(struct kvm_vcpu *vcpu, unsigned long flags,
 745                         unsigned long pte_index)
 746 {
 747         struct kvm *kvm = vcpu->kvm;
 748         __be64 *hpte;
 749         unsigned long v, r, gr;
 750         struct revmap_entry *rev;
 751         unsigned long *rmap;
 752         long ret = H_NOT_FOUND;
 753
 754         if (pte_index >= kvm->arch.hpt_npte)
 755                 return H_PARAMETER;
 756
 757         rev = real_vmalloc_addr(&kvm->arch.revmap[pte_index]);
 758         hpte = (__be64 *)(kvm->arch.hpt_virt + (pte_index << 4));
 759         while (!try_lock_hpte(hpte, HPTE_V_HVLOCK))
 760                 cpu_relax();
 761         v = be64_to_cpu(hpte[0]);
 762         r = be64_to_cpu(hpte[1]);
 763         if (!(v & (HPTE_V_VALID | HPTE_V_ABSENT)))
 764                 goto out;
 765
 766         gr = rev->guest_rpte;
 767         if (gr & HPTE_R_C) {
 768                 rev->guest_rpte &= ~HPTE_R_C;
 769                 note_hpte_modification(kvm, rev);
 770         }
 771         if (v & HPTE_V_VALID) {
 772                 /* need to make it temporarily absent so C is stable */
 773                 hpte[0] |= cpu_to_be64(HPTE_V_ABSENT);
 774                 kvmppc_invalidate_hpte(kvm, hpte, pte_index);
 775                 r = be64_to_cpu(hpte[1]);
 776                 gr |= r & (HPTE_R_R | HPTE_R_C);
 777                 if (r & HPTE_R_C) {
 778                         unsigned long psize = hpte_page_size(v, r);
 779                         hpte[1] = cpu_to_be64(r & ~HPTE_R_C);
 780                         eieio();
 781                         rmap = revmap_for_hpte(kvm, v, gr);
 782                         if (rmap) {
 783                                 lock_rmap(rmap);
 784                                 *rmap |= KVMPPC_RMAP_CHANGED;
 785                                 kvmppc_update_rmap_change(rmap, psize);
 786                                 unlock_rmap(rmap);
 787                         }
 788                 }
 789         }
 790         vcpu->arch.gpr[4] = gr;
 791         ret = H_SUCCESS;
 792  out:
 793         unlock_hpte(hpte, v & ~HPTE_V_HVLOCK);
 794         return ret;
 795 }
 796
 797 void kvmppc_invalidate_hpte(struct kvm *kvm, __be64 *hptep,
 798                         unsigned long pte_index)
 799 {
 800         unsigned long rb;
 801
 802         hptep[0] &= ~cpu_to_be64(HPTE_V_VALID);
 803         rb = compute_tlbie_rb(be64_to_cpu(hptep[0]), be64_to_cpu(hptep[1]),
 804                               pte_index);
 805         do_tlbies(kvm, &rb, 1, 1, true);
 806 }
 807 EXPORT_SYMBOL_GPL(kvmppc_invalidate_hpte);
 808
 809 void kvmppc_clear_ref_hpte(struct kvm *kvm, __be64 *hptep,
 810                            unsigned long pte_index)
 811 {
 812         unsigned long rb;
 813         unsigned char rbyte;
 814
 815         rb = compute_tlbie_rb(be64_to_cpu(hptep[0]), be64_to_cpu(hptep[1]),
 816                               pte_index);
 817         rbyte = (be64_to_cpu(hptep[1]) & ~HPTE_R_R) >> 8;
 818         /* modify only the second-last byte, which contains the ref bit */
 819         *((char *)hptep + 14) = rbyte;
 820         do_tlbies(kvm, &rb, 1, 1, false);
 821 }
 822 EXPORT_SYMBOL_GPL(kvmppc_clear_ref_hpte);
 823
 824 static int slb_base_page_shift[4] = {
 825         24,     /* 16M */
 826         16,     /* 64k */
 827         34,     /* 16G */
 828         20,     /* 1M, unsupported */
 829 };
 830
 831 /* When called from virtmode, this func should be protected by
 832  * preempt_disable(), otherwise, the holding of HPTE_V_HVLOCK
 833  * can trigger deadlock issue.
 834  */
 835 long kvmppc_hv_find_lock_hpte(struct kvm *kvm, gva_t eaddr, unsigned long slb_v,
 836                               unsigned long valid)
 837 {
 838         unsigned int i;
 839         unsigned int pshift;
 840         unsigned long somask;
 841         unsigned long vsid, hash;
 842         unsigned long avpn;
 843         __be64 *hpte;
 844         unsigned long mask, val;
 845         unsigned long v, r;
 846
 847         /* Get page shift, work out hash and AVPN etc. */
 848         mask = SLB_VSID_B | HPTE_V_AVPN | HPTE_V_SECONDARY;
 849         val = 0;
 850         pshift = 12;
 851         if (slb_v & SLB_VSID_L) {
 852                 mask |= HPTE_V_LARGE;
 853                 val |= HPTE_V_LARGE;
 854                 pshift = slb_base_page_shift[(slb_v & SLB_VSID_LP) >> 4];
 855         }
 856         if (slb_v & SLB_VSID_B_1T) {
 857                 somask = (1UL << 40) - 1;
 858                 vsid = (slb_v & ~SLB_VSID_B) >> SLB_VSID_SHIFT_1T;
 859                 vsid ^= vsid << 25;
 860         } else {
 861                 somask = (1UL << 28) - 1;
 862                 vsid = (slb_v & ~SLB_VSID_B) >> SLB_VSID_SHIFT;
 863         }
 864         hash = (vsid ^ ((eaddr & somask) >> pshift)) & kvm->arch.hpt_mask;
 865         avpn = slb_v & ~(somask >> 16); /* also includes B */
 866         avpn |= (eaddr & somask) >> 16;
 867
 868         if (pshift >= 24)
 869                 avpn &= ~((1UL << (pshift - 16)) - 1);
 870         else
 871                 avpn &= ~0x7fUL;
 872         val |= avpn;
 873
 874         for (;;) {
 875                 hpte = (__be64 *)(kvm->arch.hpt_virt + (hash << 7));
 876
 877                 for (i = 0; i < 16; i += 2) {
 878                         /* Read the PTE racily */
 879                         v = be64_to_cpu(hpte[i]) & ~HPTE_V_HVLOCK;
 880
 881                         /* Check valid/absent, hash, segment size and AVPN */
 882                         if (!(v & valid) || (v & mask) != val)
 883                                 continue;
 884
 885                         /* Lock the PTE and read it under the lock */
 886                         while (!try_lock_hpte(&hpte[i], HPTE_V_HVLOCK))
 887                                 cpu_relax();
 888                         v = be64_to_cpu(hpte[i]) & ~HPTE_V_HVLOCK;
 889                         r = be64_to_cpu(hpte[i+1]);
 890
 891                         /*
 892                          * Check the HPTE again, including base page size
 893                          */
 894                         if ((v & valid) && (v & mask) == val &&
 895                             hpte_base_page_size(v, r) == (1ul << pshift))
 896                                 /* Return with the HPTE still locked */
 897                                 return (hash << 3) + (i >> 1);
 898
 899                         __unlock_hpte(&hpte[i], v);
 900                 }
 901
 902                 if (val & HPTE_V_SECONDARY)
 903                         break;
 904                 val |= HPTE_V_SECONDARY;
 905                 hash = hash ^ kvm->arch.hpt_mask;
 906         }
 907         return -1;
 908 }
 909 EXPORT_SYMBOL(kvmppc_hv_find_lock_hpte);
 910
 911 /*
 912  * Called in real mode to check whether an HPTE not found fault
 913  * is due to accessing a paged-out page or an emulated MMIO page,
 914  * or if a protection fault is due to accessing a page that the
 915  * guest wanted read/write access to but which we made read-only.
 916  * Returns a possibly modified status (DSISR) value if not
 917  * (i.e. pass the interrupt to the guest),
 918  * -1 to pass the fault up to host kernel mode code, -2 to do that
 919  * and also load the instruction word (for MMIO emulation),
 920  * or 0 if we should make the guest retry the access.
 921  */
 922 long kvmppc_hpte_hv_fault(struct kvm_vcpu *vcpu, unsigned long addr,
 923                           unsigned long slb_v, unsigned int status, bool data)
 924 {
 925         struct kvm *kvm = vcpu->kvm;
 926         long int index;
 927         unsigned long v, r, gr;
 928         __be64 *hpte;
 929         unsigned long valid;
 930         struct revmap_entry *rev;
 931         unsigned long pp, key;
 932
 933         /* For protection fault, expect to find a valid HPTE */
 934         valid = HPTE_V_VALID;
 935         if (status & DSISR_NOHPTE)
 936                 valid |= HPTE_V_ABSENT;
 937
 938         index = kvmppc_hv_find_lock_hpte(kvm, addr, slb_v, valid);
 939         if (index < 0) {
 940                 if (status & DSISR_NOHPTE)
 941                         return status;  /* there really was no HPTE */
 942                 return 0;               /* for prot fault, HPTE disappeared */
 943         }
 944         hpte = (__be64 *)(kvm->arch.hpt_virt + (index << 4));
 945         v = be64_to_cpu(hpte[0]) & ~HPTE_V_HVLOCK;
 946         r = be64_to_cpu(hpte[1]);
 947         rev = real_vmalloc_addr(&kvm->arch.revmap[index]);
 948         gr = rev->guest_rpte;
 949
 950         unlock_hpte(hpte, v);
 951
 952         /* For not found, if the HPTE is valid by now, retry the instruction */
 953         if ((status & DSISR_NOHPTE) && (v & HPTE_V_VALID))
 954                 return 0;
 955
 956         /* Check access permissions to the page */
 957         pp = gr & (HPTE_R_PP0 | HPTE_R_PP);
 958         key = (vcpu->arch.shregs.msr & MSR_PR) ? SLB_VSID_KP : SLB_VSID_KS;
 959         status &= ~DSISR_NOHPTE;        /* DSISR_NOHPTE == SRR1_ISI_NOPT */
 960         if (!data) {
 961                 if (gr & (HPTE_R_N | HPTE_R_G))
 962                         return status | SRR1_ISI_N_OR_G;
 963                 if (!hpte_read_permission(pp, slb_v & key))
 964                         return status | SRR1_ISI_PROT;
 965         } else if (status & DSISR_ISSTORE) {
 966                 /* check write permission */
 967                 if (!hpte_write_permission(pp, slb_v & key))
 968                         return status | DSISR_PROTFAULT;
 969         } else {
 970                 if (!hpte_read_permission(pp, slb_v & key))
 971                         return status | DSISR_PROTFAULT;
 972         }
 973
 974         /* Check storage key, if applicable */
 975         if (data && (vcpu->arch.shregs.msr & MSR_DR)) {
 976                 unsigned int perm = hpte_get_skey_perm(gr, vcpu->arch.amr);
 977                 if (status & DSISR_ISSTORE)
 978                         perm >>= 1;
 979                 if (perm & 1)
 980                         return status | DSISR_KEYFAULT;
 981         }
 982
 983         /* Save HPTE info for virtual-mode handler */
 984         vcpu->arch.pgfault_addr = addr;
 985         vcpu->arch.pgfault_index = index;
 986         vcpu->arch.pgfault_hpte[0] = v;
 987         vcpu->arch.pgfault_hpte[1] = r;
 988
 989         /* Check the storage key to see if it is possibly emulated MMIO */
 990         if (data && (vcpu->arch.shregs.msr & MSR_IR) &&
 991             (r & (HPTE_R_KEY_HI | HPTE_R_KEY_LO)) ==
 992             (HPTE_R_KEY_HI | HPTE_R_KEY_LO))
 993                 return -2;      /* MMIO emulation - load instr word */
 994
 995         return -1;              /* send fault up to host kernel mode */
 996 }