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
  31 #include <asm/tlbflush.h>
  32 #include <asm/kvm_ppc.h>
  33 #include <asm/kvm_book3s.h>
  34 #include <asm/mmu-hash64.h>
  35 #include <asm/hvcall.h>
  36 #include <asm/synch.h>
  37 #include <asm/ppc-opcode.h>
  38 #include <asm/cputable.h>
  39
  40 #include "book3s_hv_cma.h"
  41
  42 /* POWER7 has 10-bit LPIDs, PPC970 has 6-bit LPIDs */
  43 #define MAX_LPID_970    63
  44
  45 /* Power architecture requires HPT is at least 256kB */
  46 #define PPC_MIN_HPT_ORDER       18
  47
  48 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
  49                                 long pte_index, unsigned long pteh,
  50                                 unsigned long ptel, unsigned long *pte_idx_ret);
  51 static void kvmppc_rmap_reset(struct kvm *kvm);
  52
  53 long kvmppc_alloc_hpt(struct kvm *kvm, u32 *htab_orderp)
  54 {
  55         unsigned long hpt = 0;
  56         struct revmap_entry *rev;
  57         struct page *page = NULL;
  58         long order = KVM_DEFAULT_HPT_ORDER;
  59
  60         if (htab_orderp) {
  61                 order = *htab_orderp;
  62                 if (order < PPC_MIN_HPT_ORDER)
  63                         order = PPC_MIN_HPT_ORDER;
  64         }
  65
  66         kvm->arch.hpt_cma_alloc = 0;
  67         VM_BUG_ON(order < KVM_CMA_CHUNK_ORDER);
  68         page = kvm_alloc_hpt(1 << (order - PAGE_SHIFT));
  69         if (page) {
  70                 hpt = (unsigned long)pfn_to_kaddr(page_to_pfn(page));
  71                 kvm->arch.hpt_cma_alloc = 1;
  72         }
  73
  74         /* Lastly try successively smaller sizes from the page allocator */
  75         while (!hpt && order > PPC_MIN_HPT_ORDER) {
  76                 hpt = __get_free_pages(GFP_KERNEL|__GFP_ZERO|__GFP_REPEAT|
  77                                        __GFP_NOWARN, order - PAGE_SHIFT);
  78                 if (!hpt)
  79                         --order;
  80         }
  81
  82         if (!hpt)
  83                 return -ENOMEM;
  84
  85         kvm->arch.hpt_virt = hpt;
  86         kvm->arch.hpt_order = order;
  87         /* HPTEs are 2**4 bytes long */
  88         kvm->arch.hpt_npte = 1ul << (order - 4);
  89         /* 128 (2**7) bytes in each HPTEG */
  90         kvm->arch.hpt_mask = (1ul << (order - 7)) - 1;
  91
  92         /* Allocate reverse map array */
  93         rev = vmalloc(sizeof(struct revmap_entry) * kvm->arch.hpt_npte);
  94         if (!rev) {
  95                 pr_err("kvmppc_alloc_hpt: Couldn't alloc reverse map array\n");
  96                 goto out_freehpt;
  97         }
  98         kvm->arch.revmap = rev;
  99         kvm->arch.sdr1 = __pa(hpt) | (order - 18);
 100
 101         pr_info("KVM guest htab at %lx (order %ld), LPID %x\n",
 102                 hpt, order, kvm->arch.lpid);
 103
 104         if (htab_orderp)
 105                 *htab_orderp = order;
 106         return 0;
 107
 108  out_freehpt:
 109         if (kvm->arch.hpt_cma_alloc)
 110                 kvm_release_hpt(page, 1 << (order - PAGE_SHIFT));
 111         else
 112                 free_pages(hpt, order - PAGE_SHIFT);
 113         return -ENOMEM;
 114 }
 115
 116 long kvmppc_alloc_reset_hpt(struct kvm *kvm, u32 *htab_orderp)
 117 {
 118         long err = -EBUSY;
 119         long order;
 120
 121         mutex_lock(&kvm->lock);
 122         if (kvm->arch.rma_setup_done) {
 123                 kvm->arch.rma_setup_done = 0;
 124                 /* order rma_setup_done vs. vcpus_running */
 125                 smp_mb();
 126                 if (atomic_read(&kvm->arch.vcpus_running)) {
 127                         kvm->arch.rma_setup_done = 1;
 128                         goto out;
 129                 }
 130         }
 131         if (kvm->arch.hpt_virt) {
 132                 order = kvm->arch.hpt_order;
 133                 /* Set the entire HPT to 0, i.e. invalid HPTEs */
 134                 memset((void *)kvm->arch.hpt_virt, 0, 1ul << order);
 135                 /*
 136                  * Reset all the reverse-mapping chains for all memslots
 137                  */
 138                 kvmppc_rmap_reset(kvm);
 139                 /* Ensure that each vcpu will flush its TLB on next entry. */
 140                 cpumask_setall(&kvm->arch.need_tlb_flush);
 141                 *htab_orderp = order;
 142                 err = 0;
 143         } else {
 144                 err = kvmppc_alloc_hpt(kvm, htab_orderp);
 145                 order = *htab_orderp;
 146         }
 147  out:
 148         mutex_unlock(&kvm->lock);
 149         return err;
 150 }
 151
 152 void kvmppc_free_hpt(struct kvm *kvm)
 153 {
 154         kvmppc_free_lpid(kvm->arch.lpid);
 155         vfree(kvm->arch.revmap);
 156         if (kvm->arch.hpt_cma_alloc)
 157                 kvm_release_hpt(virt_to_page(kvm->arch.hpt_virt),
 158                                 1 << (kvm->arch.hpt_order - PAGE_SHIFT));
 159         else
 160                 free_pages(kvm->arch.hpt_virt,
 161                            kvm->arch.hpt_order - PAGE_SHIFT);
 162 }
 163
 164 /* Bits in first HPTE dword for pagesize 4k, 64k or 16M */
 165 static inline unsigned long hpte0_pgsize_encoding(unsigned long pgsize)
 166 {
 167         return (pgsize > 0x1000) ? HPTE_V_LARGE : 0;
 168 }
 169
 170 /* Bits in second HPTE dword for pagesize 4k, 64k or 16M */
 171 static inline unsigned long hpte1_pgsize_encoding(unsigned long pgsize)
 172 {
 173         return (pgsize == 0x10000) ? 0x1000 : 0;
 174 }
 175
 176 void kvmppc_map_vrma(struct kvm_vcpu *vcpu, struct kvm_memory_slot *memslot,
 177                      unsigned long porder)
 178 {
 179         unsigned long i;
 180         unsigned long npages;
 181         unsigned long hp_v, hp_r;
 182         unsigned long addr, hash;
 183         unsigned long psize;
 184         unsigned long hp0, hp1;
 185         unsigned long idx_ret;
 186         long ret;
 187         struct kvm *kvm = vcpu->kvm;
 188
 189         psize = 1ul << porder;
 190         npages = memslot->npages >> (porder - PAGE_SHIFT);
 191
 192         /* VRMA can't be > 1TB */
 193         if (npages > 1ul << (40 - porder))
 194                 npages = 1ul << (40 - porder);
 195         /* Can't use more than 1 HPTE per HPTEG */
 196         if (npages > kvm->arch.hpt_mask + 1)
 197                 npages = kvm->arch.hpt_mask + 1;
 198
 199         hp0 = HPTE_V_1TB_SEG | (VRMA_VSID << (40 - 16)) |
 200                 HPTE_V_BOLTED | hpte0_pgsize_encoding(psize);
 201         hp1 = hpte1_pgsize_encoding(psize) |
 202                 HPTE_R_R | HPTE_R_C | HPTE_R_M | PP_RWXX;
 203
 204         for (i = 0; i < npages; ++i) {
 205                 addr = i << porder;
 206                 /* can't use hpt_hash since va > 64 bits */
 207                 hash = (i ^ (VRMA_VSID ^ (VRMA_VSID << 25))) & kvm->arch.hpt_mask;
 208                 /*
 209                  * We assume that the hash table is empty and no
 210                  * vcpus are using it at this stage.  Since we create
 211                  * at most one HPTE per HPTEG, we just assume entry 7
 212                  * is available and use it.
 213                  */
 214                 hash = (hash << 3) + 7;
 215                 hp_v = hp0 | ((addr >> 16) & ~0x7fUL);
 216                 hp_r = hp1 | addr;
 217                 ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, hash, hp_v, hp_r,
 218                                                  &idx_ret);
 219                 if (ret != H_SUCCESS) {
 220                         pr_err("KVM: map_vrma at %lx failed, ret=%ld\n",
 221                                addr, ret);
 222                         break;
 223                 }
 224         }
 225 }
 226
 227 int kvmppc_mmu_hv_init(void)
 228 {
 229         unsigned long host_lpid, rsvd_lpid;
 230
 231         if (!cpu_has_feature(CPU_FTR_HVMODE))
 232                 return -EINVAL;
 233
 234         /* POWER7 has 10-bit LPIDs, PPC970 and e500mc have 6-bit LPIDs */
 235         if (cpu_has_feature(CPU_FTR_ARCH_206)) {
 236                 host_lpid = mfspr(SPRN_LPID);   /* POWER7 */
 237                 rsvd_lpid = LPID_RSVD;
 238         } else {
 239                 host_lpid = 0;                  /* PPC970 */
 240                 rsvd_lpid = MAX_LPID_970;
 241         }
 242
 243         kvmppc_init_lpid(rsvd_lpid + 1);
 244
 245         kvmppc_claim_lpid(host_lpid);
 246         /* rsvd_lpid is reserved for use in partition switching */
 247         kvmppc_claim_lpid(rsvd_lpid);
 248
 249         return 0;
 250 }
 251
 252 static void kvmppc_mmu_book3s_64_hv_reset_msr(struct kvm_vcpu *vcpu)
 253 {
 254         unsigned long msr = vcpu->arch.intr_msr;
 255
 256         /* If transactional, change to suspend mode on IRQ delivery */
 257         if (MSR_TM_TRANSACTIONAL(vcpu->arch.shregs.msr))
 258                 msr |= MSR_TS_S;
 259         else
 260                 msr |= vcpu->arch.shregs.msr & MSR_TS_MASK;
 261         kvmppc_set_msr(vcpu, msr);
 262 }
 263
 264 /*
 265  * This is called to get a reference to a guest page if there isn't
 266  * one already in the memslot->arch.slot_phys[] array.
 267  */
 268 static long kvmppc_get_guest_page(struct kvm *kvm, unsigned long gfn,
 269                                   struct kvm_memory_slot *memslot,
 270                                   unsigned long psize)
 271 {
 272         unsigned long start;
 273         long np, err;
 274         struct page *page, *hpage, *pages[1];
 275         unsigned long s, pgsize;
 276         unsigned long *physp;
 277         unsigned int is_io, got, pgorder;
 278         struct vm_area_struct *vma;
 279         unsigned long pfn, i, npages;
 280
 281         physp = memslot->arch.slot_phys;
 282         if (!physp)
 283                 return -EINVAL;
 284         if (physp[gfn - memslot->base_gfn])
 285                 return 0;
 286
 287         is_io = 0;
 288         got = 0;
 289         page = NULL;
 290         pgsize = psize;
 291         err = -EINVAL;
 292         start = gfn_to_hva_memslot(memslot, gfn);
 293
 294         /* Instantiate and get the page we want access to */
 295         np = get_user_pages_fast(start, 1, 1, pages);
 296         if (np != 1) {
 297                 /* Look up the vma for the page */
 298                 down_read(&current->mm->mmap_sem);
 299                 vma = find_vma(current->mm, start);
 300                 if (!vma || vma->vm_start > start ||
 301                     start + psize > vma->vm_end ||
 302                     !(vma->vm_flags & VM_PFNMAP))
 303                         goto up_err;
 304                 is_io = hpte_cache_bits(pgprot_val(vma->vm_page_prot));
 305                 pfn = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
 306                 /* check alignment of pfn vs. requested page size */
 307                 if (psize > PAGE_SIZE && (pfn & ((psize >> PAGE_SHIFT) - 1)))
 308                         goto up_err;
 309                 up_read(&current->mm->mmap_sem);
 310
 311         } else {
 312                 page = pages[0];
 313                 got = KVMPPC_GOT_PAGE;
 314
 315                 /* See if this is a large page */
 316                 s = PAGE_SIZE;
 317                 if (PageHuge(page)) {
 318                         hpage = compound_head(page);
 319                         s <<= compound_order(hpage);
 320                         /* Get the whole large page if slot alignment is ok */
 321                         if (s > psize && slot_is_aligned(memslot, s) &&
 322                             !(memslot->userspace_addr & (s - 1))) {
 323                                 start &= ~(s - 1);
 324                                 pgsize = s;
 325                                 get_page(hpage);
 326                                 put_page(page);
 327                                 page = hpage;
 328                         }
 329                 }
 330                 if (s < psize)
 331                         goto out;
 332                 pfn = page_to_pfn(page);
 333         }
 334
 335         npages = pgsize >> PAGE_SHIFT;
 336         pgorder = __ilog2(npages);
 337         physp += (gfn - memslot->base_gfn) & ~(npages - 1);
 338         spin_lock(&kvm->arch.slot_phys_lock);
 339         for (i = 0; i < npages; ++i) {
 340                 if (!physp[i]) {
 341                         physp[i] = ((pfn + i) << PAGE_SHIFT) +
 342                                 got + is_io + pgorder;
 343                         got = 0;
 344                 }
 345         }
 346         spin_unlock(&kvm->arch.slot_phys_lock);
 347         err = 0;
 348
 349  out:
 350         if (got)
 351                 put_page(page);
 352         return err;
 353
 354  up_err:
 355         up_read(&current->mm->mmap_sem);
 356         return err;
 357 }
 358
 359 long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
 360                                 long pte_index, unsigned long pteh,
 361                                 unsigned long ptel, unsigned long *pte_idx_ret)
 362 {
 363         unsigned long psize, gpa, gfn;
 364         struct kvm_memory_slot *memslot;
 365         long ret;
 366
 367         if (kvm->arch.using_mmu_notifiers)
 368                 goto do_insert;
 369
 370         psize = hpte_page_size(pteh, ptel);
 371         if (!psize)
 372                 return H_PARAMETER;
 373
 374         pteh &= ~(HPTE_V_HVLOCK | HPTE_V_ABSENT | HPTE_V_VALID);
 375
 376         /* Find the memslot (if any) for this address */
 377         gpa = (ptel & HPTE_R_RPN) & ~(psize - 1);
 378         gfn = gpa >> PAGE_SHIFT;
 379         memslot = gfn_to_memslot(kvm, gfn);
 380         if (memslot && !(memslot->flags & KVM_MEMSLOT_INVALID)) {
 381                 if (!slot_is_aligned(memslot, psize))
 382                         return H_PARAMETER;
 383                 if (kvmppc_get_guest_page(kvm, gfn, memslot, psize) < 0)
 384                         return H_PARAMETER;
 385         }
 386
 387  do_insert:
 388         /* Protect linux PTE lookup from page table destruction */
 389         rcu_read_lock_sched();  /* this disables preemption too */
 390         ret = kvmppc_do_h_enter(kvm, flags, pte_index, pteh, ptel,
 391                                 current->mm->pgd, false, pte_idx_ret);
 392         rcu_read_unlock_sched();
 393         if (ret == H_TOO_HARD) {
 394                 /* this can't happen */
 395                 pr_err("KVM: Oops, kvmppc_h_enter returned too hard!\n");
 396                 ret = H_RESOURCE;       /* or something */
 397         }
 398         return ret;
 399
 400 }
 401
 402 /*
 403  * We come here on a H_ENTER call from the guest when we are not
 404  * using mmu notifiers and we don't have the requested page pinned
 405  * already.
 406  */
 407 long kvmppc_virtmode_h_enter(struct kvm_vcpu *vcpu, unsigned long flags,
 408                              long pte_index, unsigned long pteh,
 409                              unsigned long ptel)
 410 {
 411         return kvmppc_virtmode_do_h_enter(vcpu->kvm, flags, pte_index,
 412                                           pteh, ptel, &vcpu->arch.gpr[4]);
 413 }
 414
 415 static struct kvmppc_slb *kvmppc_mmu_book3s_hv_find_slbe(struct kvm_vcpu *vcpu,
 416                                                          gva_t eaddr)
 417 {
 418         u64 mask;
 419         int i;
 420
 421         for (i = 0; i < vcpu->arch.slb_nr; i++) {
 422                 if (!(vcpu->arch.slb[i].orige & SLB_ESID_V))
 423                         continue;
 424
 425                 if (vcpu->arch.slb[i].origv & SLB_VSID_B_1T)
 426                         mask = ESID_MASK_1T;
 427                 else
 428                         mask = ESID_MASK;
 429
 430                 if (((vcpu->arch.slb[i].orige ^ eaddr) & mask) == 0)
 431                         return &vcpu->arch.slb[i];
 432         }
 433         return NULL;
 434 }
 435
 436 static unsigned long kvmppc_mmu_get_real_addr(unsigned long v, unsigned long r,
 437                         unsigned long ea)
 438 {
 439         unsigned long ra_mask;
 440
 441         ra_mask = hpte_page_size(v, r) - 1;
 442         return (r & HPTE_R_RPN & ~ra_mask) | (ea & ra_mask);
 443 }
 444
 445 static int kvmppc_mmu_book3s_64_hv_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
 446                         struct kvmppc_pte *gpte, bool data, bool iswrite)
 447 {
 448         struct kvm *kvm = vcpu->kvm;
 449         struct kvmppc_slb *slbe;
 450         unsigned long slb_v;
 451         unsigned long pp, key;
 452         unsigned long v, gr;
 453         __be64 *hptep;
 454         int index;
 455         int virtmode = vcpu->arch.shregs.msr & (data ? MSR_DR : MSR_IR);
 456
 457         /* Get SLB entry */
 458         if (virtmode) {
 459                 slbe = kvmppc_mmu_book3s_hv_find_slbe(vcpu, eaddr);
 460                 if (!slbe)
 461                         return -EINVAL;
 462                 slb_v = slbe->origv;
 463         } else {
 464                 /* real mode access */
 465                 slb_v = vcpu->kvm->arch.vrma_slb_v;
 466         }
 467
 468         preempt_disable();
 469         /* Find the HPTE in the hash table */
 470         index = kvmppc_hv_find_lock_hpte(kvm, eaddr, slb_v,
 471                                          HPTE_V_VALID | HPTE_V_ABSENT);
 472         if (index < 0) {
 473                 preempt_enable();
 474                 return -ENOENT;
 475         }
 476         hptep = (__be64 *)(kvm->arch.hpt_virt + (index << 4));
 477         v = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
 478         gr = kvm->arch.revmap[index].guest_rpte;
 479
 480         /* Unlock the HPTE */
 481         asm volatile("lwsync" : : : "memory");
 482         hptep[0] = cpu_to_be64(v);
 483         preempt_enable();
 484
 485         gpte->eaddr = eaddr;
 486         gpte->vpage = ((v & HPTE_V_AVPN) << 4) | ((eaddr >> 12) & 0xfff);
 487
 488         /* Get PP bits and key for permission check */
 489         pp = gr & (HPTE_R_PP0 | HPTE_R_PP);
 490         key = (vcpu->arch.shregs.msr & MSR_PR) ? SLB_VSID_KP : SLB_VSID_KS;
 491         key &= slb_v;
 492
 493         /* Calculate permissions */
 494         gpte->may_read = hpte_read_permission(pp, key);
 495         gpte->may_write = hpte_write_permission(pp, key);
 496         gpte->may_execute = gpte->may_read && !(gr & (HPTE_R_N | HPTE_R_G));
 497
 498         /* Storage key permission check for POWER7 */
 499         if (data && virtmode && cpu_has_feature(CPU_FTR_ARCH_206)) {
 500                 int amrfield = hpte_get_skey_perm(gr, vcpu->arch.amr);
 501                 if (amrfield & 1)
 502                         gpte->may_read = 0;
 503                 if (amrfield & 2)
 504                         gpte->may_write = 0;
 505         }
 506
 507         /* Get the guest physical address */
 508         gpte->raddr = kvmppc_mmu_get_real_addr(v, gr, eaddr);
 509         return 0;
 510 }
 511
 512 /*
 513  * Quick test for whether an instruction is a load or a store.
 514  * If the instruction is a load or a store, then this will indicate
 515  * which it is, at least on server processors.  (Embedded processors
 516  * have some external PID instructions that don't follow the rule
 517  * embodied here.)  If the instruction isn't a load or store, then
 518  * this doesn't return anything useful.
 519  */
 520 static int instruction_is_store(unsigned int instr)
 521 {
 522         unsigned int mask;
 523
 524         mask = 0x10000000;
 525         if ((instr & 0xfc000000) == 0x7c000000)
 526                 mask = 0x100;           /* major opcode 31 */
 527         return (instr & mask) != 0;
 528 }
 529
 530 static int kvmppc_hv_emulate_mmio(struct kvm_run *run, struct kvm_vcpu *vcpu,
 531                                   unsigned long gpa, gva_t ea, int is_store)
 532 {
 533         int ret;
 534         u32 last_inst;
 535         unsigned long srr0 = kvmppc_get_pc(vcpu);
 536
 537         /* We try to load the last instruction.  We don't let
 538          * emulate_instruction do it as it doesn't check what
 539          * kvmppc_ld returns.
 540          * If we fail, we just return to the guest and try executing it again.
 541          */
 542         if (vcpu->arch.last_inst == KVM_INST_FETCH_FAILED) {
 543                 ret = kvmppc_ld(vcpu, &srr0, sizeof(u32), &last_inst, false);
 544                 if (ret != EMULATE_DONE || last_inst == KVM_INST_FETCH_FAILED)
 545                         return RESUME_GUEST;
 546                 vcpu->arch.last_inst = last_inst;
 547         }
 548
 549         /*
 550          * WARNING: We do not know for sure whether the instruction we just
 551          * read from memory is the same that caused the fault in the first
 552          * place.  If the instruction we read is neither an load or a store,
 553          * then it can't access memory, so we don't need to worry about
 554          * enforcing access permissions.  So, assuming it is a load or
 555          * store, we just check that its direction (load or store) is
 556          * consistent with the original fault, since that's what we
 557          * checked the access permissions against.  If there is a mismatch
 558          * we just return and retry the instruction.
 559          */
 560
 561         if (instruction_is_store(kvmppc_get_last_inst(vcpu)) != !!is_store)
 562                 return RESUME_GUEST;
 563
 564         /*
 565          * Emulated accesses are emulated by looking at the hash for
 566          * translation once, then performing the access later. The
 567          * translation could be invalidated in the meantime in which
 568          * point performing the subsequent memory access on the old
 569          * physical address could possibly be a security hole for the
 570          * guest (but not the host).
 571          *
 572          * This is less of an issue for MMIO stores since they aren't
 573          * globally visible. It could be an issue for MMIO loads to
 574          * a certain extent but we'll ignore it for now.
 575          */
 576
 577         vcpu->arch.paddr_accessed = gpa;
 578         vcpu->arch.vaddr_accessed = ea;
 579         return kvmppc_emulate_mmio(run, vcpu);
 580 }
 581
 582 int kvmppc_book3s_hv_page_fault(struct kvm_run *run, struct kvm_vcpu *vcpu,
 583                                 unsigned long ea, unsigned long dsisr)
 584 {
 585         struct kvm *kvm = vcpu->kvm;
 586         unsigned long hpte[3], r;
 587         __be64 *hptep;
 588         unsigned long mmu_seq, psize, pte_size;
 589         unsigned long gpa_base, gfn_base;
 590         unsigned long gpa, gfn, hva, pfn;
 591         struct kvm_memory_slot *memslot;
 592         unsigned long *rmap;
 593         struct revmap_entry *rev;
 594         struct page *page, *pages[1];
 595         long index, ret, npages;
 596         unsigned long is_io;
 597         unsigned int writing, write_ok;
 598         struct vm_area_struct *vma;
 599         unsigned long rcbits;
 600
 601         /*
 602          * Real-mode code has already searched the HPT and found the
 603          * entry we're interested in.  Lock the entry and check that
 604          * it hasn't changed.  If it has, just return and re-execute the
 605          * instruction.
 606          */
 607         if (ea != vcpu->arch.pgfault_addr)
 608                 return RESUME_GUEST;
 609         index = vcpu->arch.pgfault_index;
 610         hptep = (__be64 *)(kvm->arch.hpt_virt + (index << 4));
 611         rev = &kvm->arch.revmap[index];
 612         preempt_disable();
 613         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 614                 cpu_relax();
 615         hpte[0] = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
 616         hpte[1] = be64_to_cpu(hptep[1]);
 617         hpte[2] = r = rev->guest_rpte;
 618         asm volatile("lwsync" : : : "memory");
 619         hptep[0] = cpu_to_be64(hpte[0]);
 620         preempt_enable();
 621
 622         if (hpte[0] != vcpu->arch.pgfault_hpte[0] ||
 623             hpte[1] != vcpu->arch.pgfault_hpte[1])
 624                 return RESUME_GUEST;
 625
 626         /* Translate the logical address and get the page */
 627         psize = hpte_page_size(hpte[0], r);
 628         gpa_base = r & HPTE_R_RPN & ~(psize - 1);
 629         gfn_base = gpa_base >> PAGE_SHIFT;
 630         gpa = gpa_base | (ea & (psize - 1));
 631         gfn = gpa >> PAGE_SHIFT;
 632         memslot = gfn_to_memslot(kvm, gfn);
 633
 634         /* No memslot means it's an emulated MMIO region */
 635         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
 636                 return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
 637                                               dsisr & DSISR_ISSTORE);
 638
 639         if (!kvm->arch.using_mmu_notifiers)
 640                 return -EFAULT;         /* should never get here */
 641
 642         /*
 643          * This should never happen, because of the slot_is_aligned()
 644          * check in kvmppc_do_h_enter().
 645          */
 646         if (gfn_base < memslot->base_gfn)
 647                 return -EFAULT;
 648
 649         /* used to check for invalidations in progress */
 650         mmu_seq = kvm->mmu_notifier_seq;
 651         smp_rmb();
 652
 653         is_io = 0;
 654         pfn = 0;
 655         page = NULL;
 656         pte_size = PAGE_SIZE;
 657         writing = (dsisr & DSISR_ISSTORE) != 0;
 658         /* If writing != 0, then the HPTE must allow writing, if we get here */
 659         write_ok = writing;
 660         hva = gfn_to_hva_memslot(memslot, gfn);
 661         npages = get_user_pages_fast(hva, 1, writing, pages);
 662         if (npages < 1) {
 663                 /* Check if it's an I/O mapping */
 664                 down_read(&current->mm->mmap_sem);
 665                 vma = find_vma(current->mm, hva);
 666                 if (vma && vma->vm_start <= hva && hva + psize <= vma->vm_end &&
 667                     (vma->vm_flags & VM_PFNMAP)) {
 668                         pfn = vma->vm_pgoff +
 669                                 ((hva - vma->vm_start) >> PAGE_SHIFT);
 670                         pte_size = psize;
 671                         is_io = hpte_cache_bits(pgprot_val(vma->vm_page_prot));
 672                         write_ok = vma->vm_flags & VM_WRITE;
 673                 }
 674                 up_read(&current->mm->mmap_sem);
 675                 if (!pfn)
 676                         return -EFAULT;
 677         } else {
 678                 page = pages[0];
 679                 pfn = page_to_pfn(page);
 680                 if (PageHuge(page)) {
 681                         page = compound_head(page);
 682                         pte_size <<= compound_order(page);
 683                 }
 684                 /* if the guest wants write access, see if that is OK */
 685                 if (!writing && hpte_is_writable(r)) {
 686                         unsigned int hugepage_shift;
 687                         pte_t *ptep, pte;
 688
 689                         /*
 690                          * We need to protect against page table destruction
 691                          * while looking up and updating the pte.
 692                          */
 693                         rcu_read_lock_sched();
 694                         ptep = find_linux_pte_or_hugepte(current->mm->pgd,
 695                                                          hva, &hugepage_shift);
 696                         if (ptep) {
 697                                 pte = kvmppc_read_update_linux_pte(ptep, 1,
 698                                                            hugepage_shift);
 699                                 if (pte_write(pte))
 700                                         write_ok = 1;
 701                         }
 702                         rcu_read_unlock_sched();
 703                 }
 704         }
 705
 706         ret = -EFAULT;
 707         if (psize > pte_size)
 708                 goto out_put;
 709
 710         /* Check WIMG vs. the actual page we're accessing */
 711         if (!hpte_cache_flags_ok(r, is_io)) {
 712                 if (is_io)
 713                         return -EFAULT;
 714                 /*
 715                  * Allow guest to map emulated device memory as
 716                  * uncacheable, but actually make it cacheable.
 717                  */
 718                 r = (r & ~(HPTE_R_W|HPTE_R_I|HPTE_R_G)) | HPTE_R_M;
 719         }
 720
 721         /*
 722          * Set the HPTE to point to pfn.
 723          * Since the pfn is at PAGE_SIZE granularity, make sure we
 724          * don't mask out lower-order bits if psize < PAGE_SIZE.
 725          */
 726         if (psize < PAGE_SIZE)
 727                 psize = PAGE_SIZE;
 728         r = (r & ~(HPTE_R_PP0 - psize)) | ((pfn << PAGE_SHIFT) & ~(psize - 1));
 729         if (hpte_is_writable(r) && !write_ok)
 730                 r = hpte_make_readonly(r);
 731         ret = RESUME_GUEST;
 732         preempt_disable();
 733         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 734                 cpu_relax();
 735         if ((be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK) != hpte[0] ||
 736                 be64_to_cpu(hptep[1]) != hpte[1] ||
 737                 rev->guest_rpte != hpte[2])
 738                 /* HPTE has been changed under us; let the guest retry */
 739                 goto out_unlock;
 740         hpte[0] = (hpte[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
 741
 742         /* Always put the HPTE in the rmap chain for the page base address */
 743         rmap = &memslot->arch.rmap[gfn_base - memslot->base_gfn];
 744         lock_rmap(rmap);
 745
 746         /* Check if we might have been invalidated; let the guest retry if so */
 747         ret = RESUME_GUEST;
 748         if (mmu_notifier_retry(vcpu->kvm, mmu_seq)) {
 749                 unlock_rmap(rmap);
 750                 goto out_unlock;
 751         }
 752
 753         /* Only set R/C in real HPTE if set in both *rmap and guest_rpte */
 754         rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
 755         r &= rcbits | ~(HPTE_R_R | HPTE_R_C);
 756
 757         if (be64_to_cpu(hptep[0]) & HPTE_V_VALID) {
 758                 /* HPTE was previously valid, so we need to invalidate it */
 759                 unlock_rmap(rmap);
 760                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
 761                 kvmppc_invalidate_hpte(kvm, hptep, index);
 762                 /* don't lose previous R and C bits */
 763                 r |= be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
 764         } else {
 765                 kvmppc_add_revmap_chain(kvm, rev, rmap, index, 0);
 766         }
 767
 768         hptep[1] = cpu_to_be64(r);
 769         eieio();
 770         hptep[0] = cpu_to_be64(hpte[0]);
 771         asm volatile("ptesync" : : : "memory");
 772         preempt_enable();
 773         if (page && hpte_is_writable(r))
 774                 SetPageDirty(page);
 775
 776  out_put:
 777         if (page) {
 778                 /*
 779                  * We drop pages[0] here, not page because page might
 780                  * have been set to the head page of a compound, but
 781                  * we have to drop the reference on the correct tail
 782                  * page to match the get inside gup()
 783                  */
 784                 put_page(pages[0]);
 785         }
 786         return ret;
 787
 788  out_unlock:
 789         hptep[0] &= ~cpu_to_be64(HPTE_V_HVLOCK);
 790         preempt_enable();
 791         goto out_put;
 792 }
 793
 794 static void kvmppc_rmap_reset(struct kvm *kvm)
 795 {
 796         struct kvm_memslots *slots;
 797         struct kvm_memory_slot *memslot;
 798         int srcu_idx;
 799
 800         srcu_idx = srcu_read_lock(&kvm->srcu);
 801         slots = kvm->memslots;
 802         kvm_for_each_memslot(memslot, slots) {
 803                 /*
 804                  * This assumes it is acceptable to lose reference and
 805                  * change bits across a reset.
 806                  */
 807                 memset(memslot->arch.rmap, 0,
 808                        memslot->npages * sizeof(*memslot->arch.rmap));
 809         }
 810         srcu_read_unlock(&kvm->srcu, srcu_idx);
 811 }
 812
 813 static int kvm_handle_hva_range(struct kvm *kvm,
 814                                 unsigned long start,
 815                                 unsigned long end,
 816                                 int (*handler)(struct kvm *kvm,
 817                                                unsigned long *rmapp,
 818                                                unsigned long gfn))
 819 {
 820         int ret;
 821         int retval = 0;
 822         struct kvm_memslots *slots;
 823         struct kvm_memory_slot *memslot;
 824
 825         slots = kvm_memslots(kvm);
 826         kvm_for_each_memslot(memslot, slots) {
 827                 unsigned long hva_start, hva_end;
 828                 gfn_t gfn, gfn_end;
 829
 830                 hva_start = max(start, memslot->userspace_addr);
 831                 hva_end = min(end, memslot->userspace_addr +
 832                                         (memslot->npages << PAGE_SHIFT));
 833                 if (hva_start >= hva_end)
 834                         continue;
 835                 /*
 836                  * {gfn(page) | page intersects with [hva_start, hva_end)} =
 837                  * {gfn, gfn+1, ..., gfn_end-1}.
 838                  */
 839                 gfn = hva_to_gfn_memslot(hva_start, memslot);
 840                 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
 841
 842                 for (; gfn < gfn_end; ++gfn) {
 843                         gfn_t gfn_offset = gfn - memslot->base_gfn;
 844
 845                         ret = handler(kvm, &memslot->arch.rmap[gfn_offset], gfn);
 846                         retval |= ret;
 847                 }
 848         }
 849
 850         return retval;
 851 }
 852
 853 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 854                           int (*handler)(struct kvm *kvm, unsigned long *rmapp,
 855                                          unsigned long gfn))
 856 {
 857         return kvm_handle_hva_range(kvm, hva, hva + 1, handler);
 858 }
 859
 860 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
 861                            unsigned long gfn)
 862 {
 863         struct revmap_entry *rev = kvm->arch.revmap;
 864         unsigned long h, i, j;
 865         __be64 *hptep;
 866         unsigned long ptel, psize, rcbits;
 867
 868         for (;;) {
 869                 lock_rmap(rmapp);
 870                 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 871                         unlock_rmap(rmapp);
 872                         break;
 873                 }
 874
 875                 /*
 876                  * To avoid an ABBA deadlock with the HPTE lock bit,
 877                  * we can't spin on the HPTE lock while holding the
 878                  * rmap chain lock.
 879                  */
 880                 i = *rmapp & KVMPPC_RMAP_INDEX;
 881                 hptep = (__be64 *) (kvm->arch.hpt_virt + (i << 4));
 882                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 883                         /* unlock rmap before spinning on the HPTE lock */
 884                         unlock_rmap(rmapp);
 885                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
 886                                 cpu_relax();
 887                         continue;
 888                 }
 889                 j = rev[i].forw;
 890                 if (j == i) {
 891                         /* chain is now empty */
 892                         *rmapp &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
 893                 } else {
 894                         /* remove i from chain */
 895                         h = rev[i].back;
 896                         rev[h].forw = j;
 897                         rev[j].back = h;
 898                         rev[i].forw = rev[i].back = i;
 899                         *rmapp = (*rmapp & ~KVMPPC_RMAP_INDEX) | j;
 900                 }
 901
 902                 /* Now check and modify the HPTE */
 903                 ptel = rev[i].guest_rpte;
 904                 psize = hpte_page_size(be64_to_cpu(hptep[0]), ptel);
 905                 if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
 906                     hpte_rpn(ptel, psize) == gfn) {
 907                         if (kvm->arch.using_mmu_notifiers)
 908                                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
 909                         kvmppc_invalidate_hpte(kvm, hptep, i);
 910                         /* Harvest R and C */
 911                         rcbits = be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
 912                         *rmapp |= rcbits << KVMPPC_RMAP_RC_SHIFT;
 913                         if (rcbits & ~rev[i].guest_rpte) {
 914                                 rev[i].guest_rpte = ptel | rcbits;
 915                                 note_hpte_modification(kvm, &rev[i]);
 916                         }
 917                 }
 918                 unlock_rmap(rmapp);
 919                 hptep[0] &= ~cpu_to_be64(HPTE_V_HVLOCK);
 920         }
 921         return 0;
 922 }
 923
 924 int kvm_unmap_hva_hv(struct kvm *kvm, unsigned long hva)
 925 {
 926         if (kvm->arch.using_mmu_notifiers)
 927                 kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 928         return 0;
 929 }
 930
 931 int kvm_unmap_hva_range_hv(struct kvm *kvm, unsigned long start, unsigned long end)
 932 {
 933         if (kvm->arch.using_mmu_notifiers)
 934                 kvm_handle_hva_range(kvm, start, end, kvm_unmap_rmapp);
 935         return 0;
 936 }
 937
 938 void kvmppc_core_flush_memslot_hv(struct kvm *kvm,
 939                                   struct kvm_memory_slot *memslot)
 940 {
 941         unsigned long *rmapp;
 942         unsigned long gfn;
 943         unsigned long n;
 944
 945         rmapp = memslot->arch.rmap;
 946         gfn = memslot->base_gfn;
 947         for (n = memslot->npages; n; --n) {
 948                 /*
 949                  * Testing the present bit without locking is OK because
 950                  * the memslot has been marked invalid already, and hence
 951                  * no new HPTEs referencing this page can be created,
 952                  * thus the present bit can't go from 0 to 1.
 953                  */
 954                 if (*rmapp & KVMPPC_RMAP_PRESENT)
 955                         kvm_unmap_rmapp(kvm, rmapp, gfn);
 956                 ++rmapp;
 957                 ++gfn;
 958         }
 959 }
 960
 961 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
 962                          unsigned long gfn)
 963 {
 964         struct revmap_entry *rev = kvm->arch.revmap;
 965         unsigned long head, i, j;
 966         __be64 *hptep;
 967         int ret = 0;
 968
 969  retry:
 970         lock_rmap(rmapp);
 971         if (*rmapp & KVMPPC_RMAP_REFERENCED) {
 972                 *rmapp &= ~KVMPPC_RMAP_REFERENCED;
 973                 ret = 1;
 974         }
 975         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 976                 unlock_rmap(rmapp);
 977                 return ret;
 978         }
 979
 980         i = head = *rmapp & KVMPPC_RMAP_INDEX;
 981         do {
 982                 hptep = (__be64 *) (kvm->arch.hpt_virt + (i << 4));
 983                 j = rev[i].forw;
 984
 985                 /* If this HPTE isn't referenced, ignore it */
 986                 if (!(be64_to_cpu(hptep[1]) & HPTE_R_R))
 987                         continue;
 988
 989                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 990                         /* unlock rmap before spinning on the HPTE lock */
 991                         unlock_rmap(rmapp);
 992                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
 993                                 cpu_relax();
 994                         goto retry;
 995                 }
 996
 997                 /* Now check and modify the HPTE */
 998                 if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
 999                     (be64_to_cpu(hptep[1]) & HPTE_R_R)) {
1000                         kvmppc_clear_ref_hpte(kvm, hptep, i);
1001                         if (!(rev[i].guest_rpte & HPTE_R_R)) {
1002                                 rev[i].guest_rpte |= HPTE_R_R;
1003                                 note_hpte_modification(kvm, &rev[i]);
1004                         }
1005                         ret = 1;
1006                 }
1007                 hptep[0] &= ~cpu_to_be64(HPTE_V_HVLOCK);
1008         } while ((i = j) != head);
1009
1010         unlock_rmap(rmapp);
1011         return ret;
1012 }
1013
1014 int kvm_age_hva_hv(struct kvm *kvm, unsigned long hva)
1015 {
1016         if (!kvm->arch.using_mmu_notifiers)
1017                 return 0;
1018         return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
1019 }
1020
1021 static int kvm_test_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
1022                               unsigned long gfn)
1023 {
1024         struct revmap_entry *rev = kvm->arch.revmap;
1025         unsigned long head, i, j;
1026         unsigned long *hp;
1027         int ret = 1;
1028
1029         if (*rmapp & KVMPPC_RMAP_REFERENCED)
1030                 return 1;
1031
1032         lock_rmap(rmapp);
1033         if (*rmapp & KVMPPC_RMAP_REFERENCED)
1034                 goto out;
1035
1036         if (*rmapp & KVMPPC_RMAP_PRESENT) {
1037                 i = head = *rmapp & KVMPPC_RMAP_INDEX;
1038                 do {
1039                         hp = (unsigned long *)(kvm->arch.hpt_virt + (i << 4));
1040                         j = rev[i].forw;
1041                         if (be64_to_cpu(hp[1]) & HPTE_R_R)
1042                                 goto out;
1043                 } while ((i = j) != head);
1044         }
1045         ret = 0;
1046
1047  out:
1048         unlock_rmap(rmapp);
1049         return ret;
1050 }
1051
1052 int kvm_test_age_hva_hv(struct kvm *kvm, unsigned long hva)
1053 {
1054         if (!kvm->arch.using_mmu_notifiers)
1055                 return 0;
1056         return kvm_handle_hva(kvm, hva, kvm_test_age_rmapp);
1057 }
1058
1059 void kvm_set_spte_hva_hv(struct kvm *kvm, unsigned long hva, pte_t pte)
1060 {
1061         if (!kvm->arch.using_mmu_notifiers)
1062                 return;
1063         kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
1064 }
1065
1066 static int vcpus_running(struct kvm *kvm)
1067 {
1068         return atomic_read(&kvm->arch.vcpus_running) != 0;
1069 }
1070
1071 /*
1072  * Returns the number of system pages that are dirty.
1073  * This can be more than 1 if we find a huge-page HPTE.
1074  */
1075 static int kvm_test_clear_dirty_npages(struct kvm *kvm, unsigned long *rmapp)
1076 {
1077         struct revmap_entry *rev = kvm->arch.revmap;
1078         unsigned long head, i, j;
1079         unsigned long n;
1080         unsigned long v, r;
1081         __be64 *hptep;
1082         int npages_dirty = 0;
1083
1084  retry:
1085         lock_rmap(rmapp);
1086         if (*rmapp & KVMPPC_RMAP_CHANGED) {
1087                 *rmapp &= ~KVMPPC_RMAP_CHANGED;
1088                 npages_dirty = 1;
1089         }
1090         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
1091                 unlock_rmap(rmapp);
1092                 return npages_dirty;
1093         }
1094
1095         i = head = *rmapp & KVMPPC_RMAP_INDEX;
1096         do {
1097                 unsigned long hptep1;
1098                 hptep = (__be64 *) (kvm->arch.hpt_virt + (i << 4));
1099                 j = rev[i].forw;
1100
1101                 /*
1102                  * Checking the C (changed) bit here is racy since there
1103                  * is no guarantee about when the hardware writes it back.
1104                  * If the HPTE is not writable then it is stable since the
1105                  * page can't be written to, and we would have done a tlbie
1106                  * (which forces the hardware to complete any writeback)
1107                  * when making the HPTE read-only.
1108                  * If vcpus are running then this call is racy anyway
1109                  * since the page could get dirtied subsequently, so we
1110                  * expect there to be a further call which would pick up
1111                  * any delayed C bit writeback.
1112                  * Otherwise we need to do the tlbie even if C==0 in
1113                  * order to pick up any delayed writeback of C.
1114                  */
1115                 hptep1 = be64_to_cpu(hptep[1]);
1116                 if (!(hptep1 & HPTE_R_C) &&
1117                     (!hpte_is_writable(hptep1) || vcpus_running(kvm)))
1118                         continue;
1119
1120                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
1121                         /* unlock rmap before spinning on the HPTE lock */
1122                         unlock_rmap(rmapp);
1123                         while (hptep[0] & cpu_to_be64(HPTE_V_HVLOCK))
1124                                 cpu_relax();
1125                         goto retry;
1126                 }
1127
1128                 /* Now check and modify the HPTE */
1129                 if (!(hptep[0] & cpu_to_be64(HPTE_V_VALID)))
1130                         continue;
1131
1132                 /* need to make it temporarily absent so C is stable */
1133                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
1134                 kvmppc_invalidate_hpte(kvm, hptep, i);
1135                 v = be64_to_cpu(hptep[0]);
1136                 r = be64_to_cpu(hptep[1]);
1137                 if (r & HPTE_R_C) {
1138                         hptep[1] = cpu_to_be64(r & ~HPTE_R_C);
1139                         if (!(rev[i].guest_rpte & HPTE_R_C)) {
1140                                 rev[i].guest_rpte |= HPTE_R_C;
1141                                 note_hpte_modification(kvm, &rev[i]);
1142                         }
1143                         n = hpte_page_size(v, r);
1144                         n = (n + PAGE_SIZE - 1) >> PAGE_SHIFT;
1145                         if (n > npages_dirty)
1146                                 npages_dirty = n;
1147                         eieio();
1148                 }
1149                 v &= ~(HPTE_V_ABSENT | HPTE_V_HVLOCK);
1150                 v |= HPTE_V_VALID;
1151                 hptep[0] = cpu_to_be64(v);
1152         } while ((i = j) != head);
1153
1154         unlock_rmap(rmapp);
1155         return npages_dirty;
1156 }
1157
1158 static void harvest_vpa_dirty(struct kvmppc_vpa *vpa,
1159                               struct kvm_memory_slot *memslot,
1160                               unsigned long *map)
1161 {
1162         unsigned long gfn;
1163
1164         if (!vpa->dirty || !vpa->pinned_addr)
1165                 return;
1166         gfn = vpa->gpa >> PAGE_SHIFT;
1167         if (gfn < memslot->base_gfn ||
1168             gfn >= memslot->base_gfn + memslot->npages)
1169                 return;
1170
1171         vpa->dirty = false;
1172         if (map)
1173                 __set_bit_le(gfn - memslot->base_gfn, map);
1174 }
1175
1176 long kvmppc_hv_get_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot,
1177                              unsigned long *map)
1178 {
1179         unsigned long i, j;
1180         unsigned long *rmapp;
1181         struct kvm_vcpu *vcpu;
1182
1183         preempt_disable();
1184         rmapp = memslot->arch.rmap;
1185         for (i = 0; i < memslot->npages; ++i) {
1186                 int npages = kvm_test_clear_dirty_npages(kvm, rmapp);
1187                 /*
1188                  * Note that if npages > 0 then i must be a multiple of npages,
1189                  * since we always put huge-page HPTEs in the rmap chain
1190                  * corresponding to their page base address.
1191                  */
1192                 if (npages && map)
1193                         for (j = i; npages; ++j, --npages)
1194                                 __set_bit_le(j, map);
1195                 ++rmapp;
1196         }
1197
1198         /* Harvest dirty bits from VPA and DTL updates */
1199         /* Note: we never modify the SLB shadow buffer areas */
1200         kvm_for_each_vcpu(i, vcpu, kvm) {
1201                 spin_lock(&vcpu->arch.vpa_update_lock);
1202                 harvest_vpa_dirty(&vcpu->arch.vpa, memslot, map);
1203                 harvest_vpa_dirty(&vcpu->arch.dtl, memslot, map);
1204                 spin_unlock(&vcpu->arch.vpa_update_lock);
1205         }
1206         preempt_enable();
1207         return 0;
1208 }
1209
1210 void *kvmppc_pin_guest_page(struct kvm *kvm, unsigned long gpa,
1211                             unsigned long *nb_ret)
1212 {
1213         struct kvm_memory_slot *memslot;
1214         unsigned long gfn = gpa >> PAGE_SHIFT;
1215         struct page *page, *pages[1];
1216         int npages;
1217         unsigned long hva, offset;
1218         unsigned long pa;
1219         unsigned long *physp;
1220         int srcu_idx;
1221
1222         srcu_idx = srcu_read_lock(&kvm->srcu);
1223         memslot = gfn_to_memslot(kvm, gfn);
1224         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
1225                 goto err;
1226         if (!kvm->arch.using_mmu_notifiers) {
1227                 physp = memslot->arch.slot_phys;
1228                 if (!physp)
1229                         goto err;
1230                 physp += gfn - memslot->base_gfn;
1231                 pa = *physp;
1232                 if (!pa) {
1233                         if (kvmppc_get_guest_page(kvm, gfn, memslot,
1234                                                   PAGE_SIZE) < 0)
1235                                 goto err;
1236                         pa = *physp;
1237                 }
1238                 page = pfn_to_page(pa >> PAGE_SHIFT);
1239                 get_page(page);
1240         } else {
1241                 hva = gfn_to_hva_memslot(memslot, gfn);
1242                 npages = get_user_pages_fast(hva, 1, 1, pages);
1243                 if (npages < 1)
1244                         goto err;
1245                 page = pages[0];
1246         }
1247         srcu_read_unlock(&kvm->srcu, srcu_idx);
1248
1249         offset = gpa & (PAGE_SIZE - 1);
1250         if (nb_ret)
1251                 *nb_ret = PAGE_SIZE - offset;
1252         return page_address(page) + offset;
1253
1254  err:
1255         srcu_read_unlock(&kvm->srcu, srcu_idx);
1256         return NULL;
1257 }
1258
1259 void kvmppc_unpin_guest_page(struct kvm *kvm, void *va, unsigned long gpa,
1260                              bool dirty)
1261 {
1262         struct page *page = virt_to_page(va);
1263         struct kvm_memory_slot *memslot;
1264         unsigned long gfn;
1265         unsigned long *rmap;
1266         int srcu_idx;
1267
1268         put_page(page);
1269
1270         if (!dirty || !kvm->arch.using_mmu_notifiers)
1271                 return;
1272
1273         /* We need to mark this page dirty in the rmap chain */
1274         gfn = gpa >> PAGE_SHIFT;
1275         srcu_idx = srcu_read_lock(&kvm->srcu);
1276         memslot = gfn_to_memslot(kvm, gfn);
1277         if (memslot) {
1278                 rmap = &memslot->arch.rmap[gfn - memslot->base_gfn];
1279                 lock_rmap(rmap);
1280                 *rmap |= KVMPPC_RMAP_CHANGED;
1281                 unlock_rmap(rmap);
1282         }
1283         srcu_read_unlock(&kvm->srcu, srcu_idx);
1284 }
1285
1286 /*
1287  * Functions for reading and writing the hash table via reads and
1288  * writes on a file descriptor.
1289  *
1290  * Reads return the guest view of the hash table, which has to be
1291  * pieced together from the real hash table and the guest_rpte
1292  * values in the revmap array.
1293  *
1294  * On writes, each HPTE written is considered in turn, and if it
1295  * is valid, it is written to the HPT as if an H_ENTER with the
1296  * exact flag set was done.  When the invalid count is non-zero
1297  * in the header written to the stream, the kernel will make
1298  * sure that that many HPTEs are invalid, and invalidate them
1299  * if not.
1300  */
1301
1302 struct kvm_htab_ctx {
1303         unsigned long   index;
1304         unsigned long   flags;
1305         struct kvm      *kvm;
1306         int             first_pass;
1307 };
1308
1309 #define HPTE_SIZE       (2 * sizeof(unsigned long))
1310
1311 /*
1312  * Returns 1 if this HPT entry has been modified or has pending
1313  * R/C bit changes.
1314  */
1315 static int hpte_dirty(struct revmap_entry *revp, __be64 *hptp)
1316 {
1317         unsigned long rcbits_unset;
1318
1319         if (revp->guest_rpte & HPTE_GR_MODIFIED)
1320                 return 1;
1321
1322         /* Also need to consider changes in reference and changed bits */
1323         rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1324         if ((be64_to_cpu(hptp[0]) & HPTE_V_VALID) &&
1325             (be64_to_cpu(hptp[1]) & rcbits_unset))
1326                 return 1;
1327
1328         return 0;
1329 }
1330
1331 static long record_hpte(unsigned long flags, __be64 *hptp,
1332                         unsigned long *hpte, struct revmap_entry *revp,
1333                         int want_valid, int first_pass)
1334 {
1335         unsigned long v, r;
1336         unsigned long rcbits_unset;
1337         int ok = 1;
1338         int valid, dirty;
1339
1340         /* Unmodified entries are uninteresting except on the first pass */
1341         dirty = hpte_dirty(revp, hptp);
1342         if (!first_pass && !dirty)
1343                 return 0;
1344
1345         valid = 0;
1346         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1347                 valid = 1;
1348                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) &&
1349                     !(be64_to_cpu(hptp[0]) & HPTE_V_BOLTED))
1350                         valid = 0;
1351         }
1352         if (valid != want_valid)
1353                 return 0;
1354
1355         v = r = 0;
1356         if (valid || dirty) {
1357                 /* lock the HPTE so it's stable and read it */
1358                 preempt_disable();
1359                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
1360                         cpu_relax();
1361                 v = be64_to_cpu(hptp[0]);
1362
1363                 /* re-evaluate valid and dirty from synchronized HPTE value */
1364                 valid = !!(v & HPTE_V_VALID);
1365                 dirty = !!(revp->guest_rpte & HPTE_GR_MODIFIED);
1366
1367                 /* Harvest R and C into guest view if necessary */
1368                 rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1369                 if (valid && (rcbits_unset & be64_to_cpu(hptp[1]))) {
1370                         revp->guest_rpte |= (be64_to_cpu(hptp[1]) &
1371                                 (HPTE_R_R | HPTE_R_C)) | HPTE_GR_MODIFIED;
1372                         dirty = 1;
1373                 }
1374
1375                 if (v & HPTE_V_ABSENT) {
1376                         v &= ~HPTE_V_ABSENT;
1377                         v |= HPTE_V_VALID;
1378                         valid = 1;
1379                 }
1380                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) && !(v & HPTE_V_BOLTED))
1381                         valid = 0;
1382
1383                 r = revp->guest_rpte;
1384                 /* only clear modified if this is the right sort of entry */
1385                 if (valid == want_valid && dirty) {
1386                         r &= ~HPTE_GR_MODIFIED;
1387                         revp->guest_rpte = r;
1388                 }
1389                 asm volatile(PPC_RELEASE_BARRIER "" : : : "memory");
1390                 hptp[0] &= ~cpu_to_be64(HPTE_V_HVLOCK);
1391                 preempt_enable();
1392                 if (!(valid == want_valid && (first_pass || dirty)))
1393                         ok = 0;
1394         }
1395         hpte[0] = cpu_to_be64(v);
1396         hpte[1] = cpu_to_be64(r);
1397         return ok;
1398 }
1399
1400 static ssize_t kvm_htab_read(struct file *file, char __user *buf,
1401                              size_t count, loff_t *ppos)
1402 {
1403         struct kvm_htab_ctx *ctx = file->private_data;
1404         struct kvm *kvm = ctx->kvm;
1405         struct kvm_get_htab_header hdr;
1406         __be64 *hptp;
1407         struct revmap_entry *revp;
1408         unsigned long i, nb, nw;
1409         unsigned long __user *lbuf;
1410         struct kvm_get_htab_header __user *hptr;
1411         unsigned long flags;
1412         int first_pass;
1413         unsigned long hpte[2];
1414
1415         if (!access_ok(VERIFY_WRITE, buf, count))
1416                 return -EFAULT;
1417
1418         first_pass = ctx->first_pass;
1419         flags = ctx->flags;
1420
1421         i = ctx->index;
1422         hptp = (__be64 *)(kvm->arch.hpt_virt + (i * HPTE_SIZE));
1423         revp = kvm->arch.revmap + i;
1424         lbuf = (unsigned long __user *)buf;
1425
1426         nb = 0;
1427         while (nb + sizeof(hdr) + HPTE_SIZE < count) {
1428                 /* Initialize header */
1429                 hptr = (struct kvm_get_htab_header __user *)buf;
1430                 hdr.n_valid = 0;
1431                 hdr.n_invalid = 0;
1432                 nw = nb;
1433                 nb += sizeof(hdr);
1434                 lbuf = (unsigned long __user *)(buf + sizeof(hdr));
1435
1436                 /* Skip uninteresting entries, i.e. clean on not-first pass */
1437                 if (!first_pass) {
1438                         while (i < kvm->arch.hpt_npte &&
1439                                !hpte_dirty(revp, hptp)) {
1440                                 ++i;
1441                                 hptp += 2;
1442                                 ++revp;
1443                         }
1444                 }
1445                 hdr.index = i;
1446
1447                 /* Grab a series of valid entries */
1448                 while (i < kvm->arch.hpt_npte &&
1449                        hdr.n_valid < 0xffff &&
1450                        nb + HPTE_SIZE < count &&
1451                        record_hpte(flags, hptp, hpte, revp, 1, first_pass)) {
1452                         /* valid entry, write it out */
1453                         ++hdr.n_valid;
1454                         if (__put_user(hpte[0], lbuf) ||
1455                             __put_user(hpte[1], lbuf + 1))
1456                                 return -EFAULT;
1457                         nb += HPTE_SIZE;
1458                         lbuf += 2;
1459                         ++i;
1460                         hptp += 2;
1461                         ++revp;
1462                 }
1463                 /* Now skip invalid entries while we can */
1464                 while (i < kvm->arch.hpt_npte &&
1465                        hdr.n_invalid < 0xffff &&
1466                        record_hpte(flags, hptp, hpte, revp, 0, first_pass)) {
1467                         /* found an invalid entry */
1468                         ++hdr.n_invalid;
1469                         ++i;
1470                         hptp += 2;
1471                         ++revp;
1472                 }
1473
1474                 if (hdr.n_valid || hdr.n_invalid) {
1475                         /* write back the header */
1476                         if (__copy_to_user(hptr, &hdr, sizeof(hdr)))
1477                                 return -EFAULT;
1478                         nw = nb;
1479                         buf = (char __user *)lbuf;
1480                 } else {
1481                         nb = nw;
1482                 }
1483
1484                 /* Check if we've wrapped around the hash table */
1485                 if (i >= kvm->arch.hpt_npte) {
1486                         i = 0;
1487                         ctx->first_pass = 0;
1488                         break;
1489                 }
1490         }
1491
1492         ctx->index = i;
1493
1494         return nb;
1495 }
1496
1497 static ssize_t kvm_htab_write(struct file *file, const char __user *buf,
1498                               size_t count, loff_t *ppos)
1499 {
1500         struct kvm_htab_ctx *ctx = file->private_data;
1501         struct kvm *kvm = ctx->kvm;
1502         struct kvm_get_htab_header hdr;
1503         unsigned long i, j;
1504         unsigned long v, r;
1505         unsigned long __user *lbuf;
1506         __be64 *hptp;
1507         unsigned long tmp[2];
1508         ssize_t nb;
1509         long int err, ret;
1510         int rma_setup;
1511
1512         if (!access_ok(VERIFY_READ, buf, count))
1513                 return -EFAULT;
1514
1515         /* lock out vcpus from running while we're doing this */
1516         mutex_lock(&kvm->lock);
1517         rma_setup = kvm->arch.rma_setup_done;
1518         if (rma_setup) {
1519                 kvm->arch.rma_setup_done = 0;   /* temporarily */
1520                 /* order rma_setup_done vs. vcpus_running */
1521                 smp_mb();
1522                 if (atomic_read(&kvm->arch.vcpus_running)) {
1523                         kvm->arch.rma_setup_done = 1;
1524                         mutex_unlock(&kvm->lock);
1525                         return -EBUSY;
1526                 }
1527         }
1528
1529         err = 0;
1530         for (nb = 0; nb + sizeof(hdr) <= count; ) {
1531                 err = -EFAULT;
1532                 if (__copy_from_user(&hdr, buf, sizeof(hdr)))
1533                         break;
1534
1535                 err = 0;
1536                 if (nb + hdr.n_valid * HPTE_SIZE > count)
1537                         break;
1538
1539                 nb += sizeof(hdr);
1540                 buf += sizeof(hdr);
1541
1542                 err = -EINVAL;
1543                 i = hdr.index;
1544                 if (i >= kvm->arch.hpt_npte ||
1545                     i + hdr.n_valid + hdr.n_invalid > kvm->arch.hpt_npte)
1546                         break;
1547
1548                 hptp = (__be64 *)(kvm->arch.hpt_virt + (i * HPTE_SIZE));
1549                 lbuf = (unsigned long __user *)buf;
1550                 for (j = 0; j < hdr.n_valid; ++j) {
1551                         err = -EFAULT;
1552                         if (__get_user(v, lbuf) || __get_user(r, lbuf + 1))
1553                                 goto out;
1554                         err = -EINVAL;
1555                         if (!(v & HPTE_V_VALID))
1556                                 goto out;
1557                         lbuf += 2;
1558                         nb += HPTE_SIZE;
1559
1560                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1561                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1562                         err = -EIO;
1563                         ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, i, v, r,
1564                                                          tmp);
1565                         if (ret != H_SUCCESS) {
1566                                 pr_err("kvm_htab_write ret %ld i=%ld v=%lx "
1567                                        "r=%lx\n", ret, i, v, r);
1568                                 goto out;
1569                         }
1570                         if (!rma_setup && is_vrma_hpte(v)) {
1571                                 unsigned long psize = hpte_page_size(v, r);
1572                                 unsigned long senc = slb_pgsize_encoding(psize);
1573                                 unsigned long lpcr;
1574
1575                                 kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
1576                                         (VRMA_VSID << SLB_VSID_SHIFT_1T);
1577                                 lpcr = senc << (LPCR_VRMASD_SH - 4);
1578                                 kvmppc_update_lpcr(kvm, lpcr, LPCR_VRMASD);
1579                                 rma_setup = 1;
1580                         }
1581                         ++i;
1582                         hptp += 2;
1583                 }
1584
1585                 for (j = 0; j < hdr.n_invalid; ++j) {
1586                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1587                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1588                         ++i;
1589                         hptp += 2;
1590                 }
1591                 err = 0;
1592         }
1593
1594  out:
1595         /* Order HPTE updates vs. rma_setup_done */
1596         smp_wmb();
1597         kvm->arch.rma_setup_done = rma_setup;
1598         mutex_unlock(&kvm->lock);
1599
1600         if (err)
1601                 return err;
1602         return nb;
1603 }
1604
1605 static int kvm_htab_release(struct inode *inode, struct file *filp)
1606 {
1607         struct kvm_htab_ctx *ctx = filp->private_data;
1608
1609         filp->private_data = NULL;
1610         if (!(ctx->flags & KVM_GET_HTAB_WRITE))
1611                 atomic_dec(&ctx->kvm->arch.hpte_mod_interest);
1612         kvm_put_kvm(ctx->kvm);
1613         kfree(ctx);
1614         return 0;
1615 }
1616
1617 static const struct file_operations kvm_htab_fops = {
1618         .read           = kvm_htab_read,
1619         .write          = kvm_htab_write,
1620         .llseek         = default_llseek,
1621         .release        = kvm_htab_release,
1622 };
1623
1624 int kvm_vm_ioctl_get_htab_fd(struct kvm *kvm, struct kvm_get_htab_fd *ghf)
1625 {
1626         int ret;
1627         struct kvm_htab_ctx *ctx;
1628         int rwflag;
1629
1630         /* reject flags we don't recognize */
1631         if (ghf->flags & ~(KVM_GET_HTAB_BOLTED_ONLY | KVM_GET_HTAB_WRITE))
1632                 return -EINVAL;
1633         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1634         if (!ctx)
1635                 return -ENOMEM;
1636         kvm_get_kvm(kvm);
1637         ctx->kvm = kvm;
1638         ctx->index = ghf->start_index;
1639         ctx->flags = ghf->flags;
1640         ctx->first_pass = 1;
1641
1642         rwflag = (ghf->flags & KVM_GET_HTAB_WRITE) ? O_WRONLY : O_RDONLY;
1643         ret = anon_inode_getfd("kvm-htab", &kvm_htab_fops, ctx, rwflag | O_CLOEXEC);
1644         if (ret < 0) {
1645                 kvm_put_kvm(kvm);
1646                 return ret;
1647         }
1648
1649         if (rwflag == O_RDONLY) {
1650                 mutex_lock(&kvm->slots_lock);
1651                 atomic_inc(&kvm->arch.hpte_mod_interest);
1652                 /* make sure kvmppc_do_h_enter etc. see the increment */
1653                 synchronize_srcu_expedited(&kvm->srcu);
1654                 mutex_unlock(&kvm->slots_lock);
1655         }
1656
1657         return ret;
1658 }
1659
1660 void kvmppc_mmu_book3s_hv_init(struct kvm_vcpu *vcpu)
1661 {
1662         struct kvmppc_mmu *mmu = &vcpu->arch.mmu;
1663
1664         if (cpu_has_feature(CPU_FTR_ARCH_206))
1665                 vcpu->arch.slb_nr = 32;         /* POWER7 */
1666         else
1667                 vcpu->arch.slb_nr = 64;
1668
1669         mmu->xlate = kvmppc_mmu_book3s_64_hv_xlate;
1670         mmu->reset_msr = kvmppc_mmu_book3s_64_hv_reset_msr;
1671
1672         vcpu->arch.hflags |= BOOK3S_HFLAG_SLB;
1673 }