kernel/arch/powerpc/kvm/book3s_64_mmu_hv.c

   1 /*
   2  * This program is free software; you can redistribute it and/or modify
   3  * it under the terms of the GNU General Public License, version 2, as
   4  * published by the Free Software Foundation.
   5  *
   6  * This program is distributed in the hope that it will be useful,
   7  * but WITHOUT ANY WARRANTY; without even the implied warranty of
   8  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
   9  * GNU General Public License for more details.
  10  *
  11  * You should have received a copy of the GNU General Public License
  12  * along with this program; if not, write to the Free Software
  13  * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
  14  *
  15  * Copyright 2010 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
  16  */
  17
  18 #include <linux/types.h>
  19 #include <linux/string.h>
  20 #include <linux/kvm.h>
  21 #include <linux/kvm_host.h>
  22 #include <linux/highmem.h>
  23 #include <linux/gfp.h>
  24 #include <linux/slab.h>
  25 #include <linux/hugetlb.h>
  26 #include <linux/vmalloc.h>
  27 #include <linux/srcu.h>
  28 #include <linux/anon_inodes.h>
  29 #include <linux/file.h>
  30 #include <linux/debugfs.h>
  31
  32 #include <asm/tlbflush.h>
  33 #include <asm/kvm_ppc.h>
  34 #include <asm/kvm_book3s.h>
  35 #include <asm/mmu-hash64.h>
  36 #include <asm/hvcall.h>
  37 #include <asm/synch.h>
  38 #include <asm/ppc-opcode.h>
  39 #include <asm/cputable.h>
  40
  41 #include "trace_hv.h"
  42
  43 /* Power architecture requires HPT is at least 256kB */
  44 #define PPC_MIN_HPT_ORDER       18
  45
  46 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
  47                                 long pte_index, unsigned long pteh,
  48                                 unsigned long ptel, unsigned long *pte_idx_ret);
  49 static void kvmppc_rmap_reset(struct kvm *kvm);
  50
  51 long kvmppc_alloc_hpt(struct kvm *kvm, u32 *htab_orderp)
  52 {
  53         unsigned long hpt = 0;
  54         struct revmap_entry *rev;
  55         struct page *page = NULL;
  56         long order = KVM_DEFAULT_HPT_ORDER;
  57
  58         if (htab_orderp) {
  59                 order = *htab_orderp;
  60                 if (order < PPC_MIN_HPT_ORDER)
  61                         order = PPC_MIN_HPT_ORDER;
  62         }
  63
  64         kvm->arch.hpt_cma_alloc = 0;
  65         page = kvm_alloc_hpt(1ul << (order - PAGE_SHIFT));
  66         if (page) {
  67                 hpt = (unsigned long)pfn_to_kaddr(page_to_pfn(page));
  68                 memset((void *)hpt, 0, (1ul << order));
  69                 kvm->arch.hpt_cma_alloc = 1;
  70         }
  71
  72         /* Lastly try successively smaller sizes from the page allocator */
  73         while (!hpt && order > PPC_MIN_HPT_ORDER) {
  74                 hpt = __get_free_pages(GFP_KERNEL|__GFP_ZERO|__GFP_REPEAT|
  75                                        __GFP_NOWARN, order - PAGE_SHIFT);
  76                 if (!hpt)
  77                         --order;
  78         }
  79
  80         if (!hpt)
  81                 return -ENOMEM;
  82
  83         kvm->arch.hpt_virt = hpt;
  84         kvm->arch.hpt_order = order;
  85         /* HPTEs are 2**4 bytes long */
  86         kvm->arch.hpt_npte = 1ul << (order - 4);
  87         /* 128 (2**7) bytes in each HPTEG */
  88         kvm->arch.hpt_mask = (1ul << (order - 7)) - 1;
  89
  90         /* Allocate reverse map array */
  91         rev = vmalloc(sizeof(struct revmap_entry) * kvm->arch.hpt_npte);
  92         if (!rev) {
  93                 pr_err("kvmppc_alloc_hpt: Couldn't alloc reverse map array\n");
  94                 goto out_freehpt;
  95         }
  96         kvm->arch.revmap = rev;
  97         kvm->arch.sdr1 = __pa(hpt) | (order - 18);
  98
  99         pr_info("KVM guest htab at %lx (order %ld), LPID %x\n",
 100                 hpt, order, kvm->arch.lpid);
 101
 102         if (htab_orderp)
 103                 *htab_orderp = order;
 104         return 0;
 105
 106  out_freehpt:
 107         if (kvm->arch.hpt_cma_alloc)
 108                 kvm_release_hpt(page, 1 << (order - PAGE_SHIFT));
 109         else
 110                 free_pages(hpt, order - PAGE_SHIFT);
 111         return -ENOMEM;
 112 }
 113
 114 long kvmppc_alloc_reset_hpt(struct kvm *kvm, u32 *htab_orderp)
 115 {
 116         long err = -EBUSY;
 117         long order;
 118
 119         mutex_lock(&kvm->lock);
 120         if (kvm->arch.hpte_setup_done) {
 121                 kvm->arch.hpte_setup_done = 0;
 122                 /* order hpte_setup_done vs. vcpus_running */
 123                 smp_mb();
 124                 if (atomic_read(&kvm->arch.vcpus_running)) {
 125                         kvm->arch.hpte_setup_done = 1;
 126                         goto out;
 127                 }
 128         }
 129         if (kvm->arch.hpt_virt) {
 130                 order = kvm->arch.hpt_order;
 131                 /* Set the entire HPT to 0, i.e. invalid HPTEs */
 132                 memset((void *)kvm->arch.hpt_virt, 0, 1ul << order);
 133                 /*
 134                  * Reset all the reverse-mapping chains for all memslots
 135                  */
 136                 kvmppc_rmap_reset(kvm);
 137                 /* Ensure that each vcpu will flush its TLB on next entry. */
 138                 cpumask_setall(&kvm->arch.need_tlb_flush);
 139                 *htab_orderp = order;
 140                 err = 0;
 141         } else {
 142                 err = kvmppc_alloc_hpt(kvm, htab_orderp);
 143                 order = *htab_orderp;
 144         }
 145  out:
 146         mutex_unlock(&kvm->lock);
 147         return err;
 148 }
 149
 150 void kvmppc_free_hpt(struct kvm *kvm)
 151 {
 152         kvmppc_free_lpid(kvm->arch.lpid);
 153         vfree(kvm->arch.revmap);
 154         if (kvm->arch.hpt_cma_alloc)
 155                 kvm_release_hpt(virt_to_page(kvm->arch.hpt_virt),
 156                                 1 << (kvm->arch.hpt_order - PAGE_SHIFT));
 157         else
 158                 free_pages(kvm->arch.hpt_virt,
 159                            kvm->arch.hpt_order - PAGE_SHIFT);
 160 }
 161
 162 /* Bits in first HPTE dword for pagesize 4k, 64k or 16M */
 163 static inline unsigned long hpte0_pgsize_encoding(unsigned long pgsize)
 164 {
 165         return (pgsize > 0x1000) ? HPTE_V_LARGE : 0;
 166 }
 167
 168 /* Bits in second HPTE dword for pagesize 4k, 64k or 16M */
 169 static inline unsigned long hpte1_pgsize_encoding(unsigned long pgsize)
 170 {
 171         return (pgsize == 0x10000) ? 0x1000 : 0;
 172 }
 173
 174 void kvmppc_map_vrma(struct kvm_vcpu *vcpu, struct kvm_memory_slot *memslot,
 175                      unsigned long porder)
 176 {
 177         unsigned long i;
 178         unsigned long npages;
 179         unsigned long hp_v, hp_r;
 180         unsigned long addr, hash;
 181         unsigned long psize;
 182         unsigned long hp0, hp1;
 183         unsigned long idx_ret;
 184         long ret;
 185         struct kvm *kvm = vcpu->kvm;
 186
 187         psize = 1ul << porder;
 188         npages = memslot->npages >> (porder - PAGE_SHIFT);
 189
 190         /* VRMA can't be > 1TB */
 191         if (npages > 1ul << (40 - porder))
 192                 npages = 1ul << (40 - porder);
 193         /* Can't use more than 1 HPTE per HPTEG */
 194         if (npages > kvm->arch.hpt_mask + 1)
 195                 npages = kvm->arch.hpt_mask + 1;
 196
 197         hp0 = HPTE_V_1TB_SEG | (VRMA_VSID << (40 - 16)) |
 198                 HPTE_V_BOLTED | hpte0_pgsize_encoding(psize);
 199         hp1 = hpte1_pgsize_encoding(psize) |
 200                 HPTE_R_R | HPTE_R_C | HPTE_R_M | PP_RWXX;
 201
 202         for (i = 0; i < npages; ++i) {
 203                 addr = i << porder;
 204                 /* can't use hpt_hash since va > 64 bits */
 205                 hash = (i ^ (VRMA_VSID ^ (VRMA_VSID << 25))) & kvm->arch.hpt_mask;
 206                 /*
 207                  * We assume that the hash table is empty and no
 208                  * vcpus are using it at this stage.  Since we create
 209                  * at most one HPTE per HPTEG, we just assume entry 7
 210                  * is available and use it.
 211                  */
 212                 hash = (hash << 3) + 7;
 213                 hp_v = hp0 | ((addr >> 16) & ~0x7fUL);
 214                 hp_r = hp1 | addr;
 215                 ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, hash, hp_v, hp_r,
 216                                                  &idx_ret);
 217                 if (ret != H_SUCCESS) {
 218                         pr_err("KVM: map_vrma at %lx failed, ret=%ld\n",
 219                                addr, ret);
 220                         break;
 221                 }
 222         }
 223 }
 224
 225 int kvmppc_mmu_hv_init(void)
 226 {
 227         unsigned long host_lpid, rsvd_lpid;
 228
 229         if (!cpu_has_feature(CPU_FTR_HVMODE))
 230                 return -EINVAL;
 231
 232         /* POWER7 has 10-bit LPIDs (12-bit in POWER8) */
 233         host_lpid = mfspr(SPRN_LPID);
 234         rsvd_lpid = LPID_RSVD;
 235
 236         kvmppc_init_lpid(rsvd_lpid + 1);
 237
 238         kvmppc_claim_lpid(host_lpid);
 239         /* rsvd_lpid is reserved for use in partition switching */
 240         kvmppc_claim_lpid(rsvd_lpid);
 241
 242         return 0;
 243 }
 244
 245 static void kvmppc_mmu_book3s_64_hv_reset_msr(struct kvm_vcpu *vcpu)
 246 {
 247         unsigned long msr = vcpu->arch.intr_msr;
 248
 249         /* If transactional, change to suspend mode on IRQ delivery */
 250         if (MSR_TM_TRANSACTIONAL(vcpu->arch.shregs.msr))
 251                 msr |= MSR_TS_S;
 252         else
 253                 msr |= vcpu->arch.shregs.msr & MSR_TS_MASK;
 254         kvmppc_set_msr(vcpu, msr);
 255 }
 256
 257 long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
 258                                 long pte_index, unsigned long pteh,
 259                                 unsigned long ptel, unsigned long *pte_idx_ret)
 260 {
 261         long ret;
 262
 263         /* Protect linux PTE lookup from page table destruction */
 264         rcu_read_lock_sched();  /* this disables preemption too */
 265         ret = kvmppc_do_h_enter(kvm, flags, pte_index, pteh, ptel,
 266                                 current->mm->pgd, false, pte_idx_ret);
 267         rcu_read_unlock_sched();
 268         if (ret == H_TOO_HARD) {
 269                 /* this can't happen */
 270                 pr_err("KVM: Oops, kvmppc_h_enter returned too hard!\n");
 271                 ret = H_RESOURCE;       /* or something */
 272         }
 273         return ret;
 274
 275 }
 276
 277 static struct kvmppc_slb *kvmppc_mmu_book3s_hv_find_slbe(struct kvm_vcpu *vcpu,
 278                                                          gva_t eaddr)
 279 {
 280         u64 mask;
 281         int i;
 282
 283         for (i = 0; i < vcpu->arch.slb_nr; i++) {
 284                 if (!(vcpu->arch.slb[i].orige & SLB_ESID_V))
 285                         continue;
 286
 287                 if (vcpu->arch.slb[i].origv & SLB_VSID_B_1T)
 288                         mask = ESID_MASK_1T;
 289                 else
 290                         mask = ESID_MASK;
 291
 292                 if (((vcpu->arch.slb[i].orige ^ eaddr) & mask) == 0)
 293                         return &vcpu->arch.slb[i];
 294         }
 295         return NULL;
 296 }
 297
 298 static unsigned long kvmppc_mmu_get_real_addr(unsigned long v, unsigned long r,
 299                         unsigned long ea)
 300 {
 301         unsigned long ra_mask;
 302
 303         ra_mask = hpte_page_size(v, r) - 1;
 304         return (r & HPTE_R_RPN & ~ra_mask) | (ea & ra_mask);
 305 }
 306
 307 static int kvmppc_mmu_book3s_64_hv_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
 308                         struct kvmppc_pte *gpte, bool data, bool iswrite)
 309 {
 310         struct kvm *kvm = vcpu->kvm;
 311         struct kvmppc_slb *slbe;
 312         unsigned long slb_v;
 313         unsigned long pp, key;
 314         unsigned long v, gr;
 315         __be64 *hptep;
 316         int index;
 317         int virtmode = vcpu->arch.shregs.msr & (data ? MSR_DR : MSR_IR);
 318
 319         /* Get SLB entry */
 320         if (virtmode) {
 321                 slbe = kvmppc_mmu_book3s_hv_find_slbe(vcpu, eaddr);
 322                 if (!slbe)
 323                         return -EINVAL;
 324                 slb_v = slbe->origv;
 325         } else {
 326                 /* real mode access */
 327                 slb_v = vcpu->kvm->arch.vrma_slb_v;
 328         }
 329
 330         preempt_disable();
 331         /* Find the HPTE in the hash table */
 332         index = kvmppc_hv_find_lock_hpte(kvm, eaddr, slb_v,
 333                                          HPTE_V_VALID | HPTE_V_ABSENT);
 334         if (index < 0) {
 335                 preempt_enable();
 336                 return -ENOENT;
 337         }
 338         hptep = (__be64 *)(kvm->arch.hpt_virt + (index << 4));
 339         v = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
 340         gr = kvm->arch.revmap[index].guest_rpte;
 341
 342         unlock_hpte(hptep, v);
 343         preempt_enable();
 344
 345         gpte->eaddr = eaddr;
 346         gpte->vpage = ((v & HPTE_V_AVPN) << 4) | ((eaddr >> 12) & 0xfff);
 347
 348         /* Get PP bits and key for permission check */
 349         pp = gr & (HPTE_R_PP0 | HPTE_R_PP);
 350         key = (vcpu->arch.shregs.msr & MSR_PR) ? SLB_VSID_KP : SLB_VSID_KS;
 351         key &= slb_v;
 352
 353         /* Calculate permissions */
 354         gpte->may_read = hpte_read_permission(pp, key);
 355         gpte->may_write = hpte_write_permission(pp, key);
 356         gpte->may_execute = gpte->may_read && !(gr & (HPTE_R_N | HPTE_R_G));
 357
 358         /* Storage key permission check for POWER7 */
 359         if (data && virtmode) {
 360                 int amrfield = hpte_get_skey_perm(gr, vcpu->arch.amr);
 361                 if (amrfield & 1)
 362                         gpte->may_read = 0;
 363                 if (amrfield & 2)
 364                         gpte->may_write = 0;
 365         }
 366
 367         /* Get the guest physical address */
 368         gpte->raddr = kvmppc_mmu_get_real_addr(v, gr, eaddr);
 369         return 0;
 370 }
 371
 372 /*
 373  * Quick test for whether an instruction is a load or a store.
 374  * If the instruction is a load or a store, then this will indicate
 375  * which it is, at least on server processors.  (Embedded processors
 376  * have some external PID instructions that don't follow the rule
 377  * embodied here.)  If the instruction isn't a load or store, then
 378  * this doesn't return anything useful.
 379  */
 380 static int instruction_is_store(unsigned int instr)
 381 {
 382         unsigned int mask;
 383
 384         mask = 0x10000000;
 385         if ((instr & 0xfc000000) == 0x7c000000)
 386                 mask = 0x100;           /* major opcode 31 */
 387         return (instr & mask) != 0;
 388 }
 389
 390 static int kvmppc_hv_emulate_mmio(struct kvm_run *run, struct kvm_vcpu *vcpu,
 391                                   unsigned long gpa, gva_t ea, int is_store)
 392 {
 393         u32 last_inst;
 394
 395         /*
 396          * If we fail, we just return to the guest and try executing it again.
 397          */
 398         if (kvmppc_get_last_inst(vcpu, INST_GENERIC, &last_inst) !=
 399                 EMULATE_DONE)
 400                 return RESUME_GUEST;
 401
 402         /*
 403          * WARNING: We do not know for sure whether the instruction we just
 404          * read from memory is the same that caused the fault in the first
 405          * place.  If the instruction we read is neither an load or a store,
 406          * then it can't access memory, so we don't need to worry about
 407          * enforcing access permissions.  So, assuming it is a load or
 408          * store, we just check that its direction (load or store) is
 409          * consistent with the original fault, since that's what we
 410          * checked the access permissions against.  If there is a mismatch
 411          * we just return and retry the instruction.
 412          */
 413
 414         if (instruction_is_store(last_inst) != !!is_store)
 415                 return RESUME_GUEST;
 416
 417         /*
 418          * Emulated accesses are emulated by looking at the hash for
 419          * translation once, then performing the access later. The
 420          * translation could be invalidated in the meantime in which
 421          * point performing the subsequent memory access on the old
 422          * physical address could possibly be a security hole for the
 423          * guest (but not the host).
 424          *
 425          * This is less of an issue for MMIO stores since they aren't
 426          * globally visible. It could be an issue for MMIO loads to
 427          * a certain extent but we'll ignore it for now.
 428          */
 429
 430         vcpu->arch.paddr_accessed = gpa;
 431         vcpu->arch.vaddr_accessed = ea;
 432         return kvmppc_emulate_mmio(run, vcpu);
 433 }
 434
 435 int kvmppc_book3s_hv_page_fault(struct kvm_run *run, struct kvm_vcpu *vcpu,
 436                                 unsigned long ea, unsigned long dsisr)
 437 {
 438         struct kvm *kvm = vcpu->kvm;
 439         unsigned long hpte[3], r;
 440         __be64 *hptep;
 441         unsigned long mmu_seq, psize, pte_size;
 442         unsigned long gpa_base, gfn_base;
 443         unsigned long gpa, gfn, hva, pfn;
 444         struct kvm_memory_slot *memslot;
 445         unsigned long *rmap;
 446         struct revmap_entry *rev;
 447         struct page *page, *pages[1];
 448         long index, ret, npages;
 449         unsigned long is_io;
 450         unsigned int writing, write_ok;
 451         struct vm_area_struct *vma;
 452         unsigned long rcbits;
 453
 454         /*
 455          * Real-mode code has already searched the HPT and found the
 456          * entry we're interested in.  Lock the entry and check that
 457          * it hasn't changed.  If it has, just return and re-execute the
 458          * instruction.
 459          */
 460         if (ea != vcpu->arch.pgfault_addr)
 461                 return RESUME_GUEST;
 462         index = vcpu->arch.pgfault_index;
 463         hptep = (__be64 *)(kvm->arch.hpt_virt + (index << 4));
 464         rev = &kvm->arch.revmap[index];
 465         preempt_disable();
 466         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 467                 cpu_relax();
 468         hpte[0] = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
 469         hpte[1] = be64_to_cpu(hptep[1]);
 470         hpte[2] = r = rev->guest_rpte;
 471         unlock_hpte(hptep, hpte[0]);
 472         preempt_enable();
 473
 474         if (hpte[0] != vcpu->arch.pgfault_hpte[0] ||
 475             hpte[1] != vcpu->arch.pgfault_hpte[1])
 476                 return RESUME_GUEST;
 477
 478         /* Translate the logical address and get the page */
 479         psize = hpte_page_size(hpte[0], r);
 480         gpa_base = r & HPTE_R_RPN & ~(psize - 1);
 481         gfn_base = gpa_base >> PAGE_SHIFT;
 482         gpa = gpa_base | (ea & (psize - 1));
 483         gfn = gpa >> PAGE_SHIFT;
 484         memslot = gfn_to_memslot(kvm, gfn);
 485
 486         trace_kvm_page_fault_enter(vcpu, hpte, memslot, ea, dsisr);
 487
 488         /* No memslot means it's an emulated MMIO region */
 489         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
 490                 return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
 491                                               dsisr & DSISR_ISSTORE);
 492
 493         /*
 494          * This should never happen, because of the slot_is_aligned()
 495          * check in kvmppc_do_h_enter().
 496          */
 497         if (gfn_base < memslot->base_gfn)
 498                 return -EFAULT;
 499
 500         /* used to check for invalidations in progress */
 501         mmu_seq = kvm->mmu_notifier_seq;
 502         smp_rmb();
 503
 504         ret = -EFAULT;
 505         is_io = 0;
 506         pfn = 0;
 507         page = NULL;
 508         pte_size = PAGE_SIZE;
 509         writing = (dsisr & DSISR_ISSTORE) != 0;
 510         /* If writing != 0, then the HPTE must allow writing, if we get here */
 511         write_ok = writing;
 512         hva = gfn_to_hva_memslot(memslot, gfn);
 513         npages = get_user_pages_fast(hva, 1, writing, pages);
 514         if (npages < 1) {
 515                 /* Check if it's an I/O mapping */
 516                 down_read(&current->mm->mmap_sem);
 517                 vma = find_vma(current->mm, hva);
 518                 if (vma && vma->vm_start <= hva && hva + psize <= vma->vm_end &&
 519                     (vma->vm_flags & VM_PFNMAP)) {
 520                         pfn = vma->vm_pgoff +
 521                                 ((hva - vma->vm_start) >> PAGE_SHIFT);
 522                         pte_size = psize;
 523                         is_io = hpte_cache_bits(pgprot_val(vma->vm_page_prot));
 524                         write_ok = vma->vm_flags & VM_WRITE;
 525                 }
 526                 up_read(&current->mm->mmap_sem);
 527                 if (!pfn)
 528                         goto out_put;
 529         } else {
 530                 page = pages[0];
 531                 pfn = page_to_pfn(page);
 532                 if (PageHuge(page)) {
 533                         page = compound_head(page);
 534                         pte_size <<= compound_order(page);
 535                 }
 536                 /* if the guest wants write access, see if that is OK */
 537                 if (!writing && hpte_is_writable(r)) {
 538                         pte_t *ptep, pte;
 539                         unsigned long flags;
 540                         /*
 541                          * We need to protect against page table destruction
 542                          * hugepage split and collapse.
 543                          */
 544                         local_irq_save(flags);
 545                         ptep = find_linux_pte_or_hugepte(current->mm->pgd,
 546                                                          hva, NULL);
 547                         if (ptep) {
 548                                 pte = kvmppc_read_update_linux_pte(ptep, 1);
 549                                 if (pte_write(pte))
 550                                         write_ok = 1;
 551                         }
 552                         local_irq_restore(flags);
 553                 }
 554         }
 555
 556         if (psize > pte_size)
 557                 goto out_put;
 558
 559         /* Check WIMG vs. the actual page we're accessing */
 560         if (!hpte_cache_flags_ok(r, is_io)) {
 561                 if (is_io)
 562                         goto out_put;
 563
 564                 /*
 565                  * Allow guest to map emulated device memory as
 566                  * uncacheable, but actually make it cacheable.
 567                  */
 568                 r = (r & ~(HPTE_R_W|HPTE_R_I|HPTE_R_G)) | HPTE_R_M;
 569         }
 570
 571         /*
 572          * Set the HPTE to point to pfn.
 573          * Since the pfn is at PAGE_SIZE granularity, make sure we
 574          * don't mask out lower-order bits if psize < PAGE_SIZE.
 575          */
 576         if (psize < PAGE_SIZE)
 577                 psize = PAGE_SIZE;
 578         r = (r & ~(HPTE_R_PP0 - psize)) | ((pfn << PAGE_SHIFT) & ~(psize - 1));
 579         if (hpte_is_writable(r) && !write_ok)
 580                 r = hpte_make_readonly(r);
 581         ret = RESUME_GUEST;
 582         preempt_disable();
 583         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
 584                 cpu_relax();
 585         if ((be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK) != hpte[0] ||
 586                 be64_to_cpu(hptep[1]) != hpte[1] ||
 587                 rev->guest_rpte != hpte[2])
 588                 /* HPTE has been changed under us; let the guest retry */
 589                 goto out_unlock;
 590         hpte[0] = (hpte[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
 591
 592         /* Always put the HPTE in the rmap chain for the page base address */
 593         rmap = &memslot->arch.rmap[gfn_base - memslot->base_gfn];
 594         lock_rmap(rmap);
 595
 596         /* Check if we might have been invalidated; let the guest retry if so */
 597         ret = RESUME_GUEST;
 598         if (mmu_notifier_retry(vcpu->kvm, mmu_seq)) {
 599                 unlock_rmap(rmap);
 600                 goto out_unlock;
 601         }
 602
 603         /* Only set R/C in real HPTE if set in both *rmap and guest_rpte */
 604         rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
 605         r &= rcbits | ~(HPTE_R_R | HPTE_R_C);
 606
 607         if (be64_to_cpu(hptep[0]) & HPTE_V_VALID) {
 608                 /* HPTE was previously valid, so we need to invalidate it */
 609                 unlock_rmap(rmap);
 610                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
 611                 kvmppc_invalidate_hpte(kvm, hptep, index);
 612                 /* don't lose previous R and C bits */
 613                 r |= be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
 614         } else {
 615                 kvmppc_add_revmap_chain(kvm, rev, rmap, index, 0);
 616         }
 617
 618         hptep[1] = cpu_to_be64(r);
 619         eieio();
 620         __unlock_hpte(hptep, hpte[0]);
 621         asm volatile("ptesync" : : : "memory");
 622         preempt_enable();
 623         if (page && hpte_is_writable(r))
 624                 SetPageDirty(page);
 625
 626  out_put:
 627         trace_kvm_page_fault_exit(vcpu, hpte, ret);
 628
 629         if (page) {
 630                 /*
 631                  * We drop pages[0] here, not page because page might
 632                  * have been set to the head page of a compound, but
 633                  * we have to drop the reference on the correct tail
 634                  * page to match the get inside gup()
 635                  */
 636                 put_page(pages[0]);
 637         }
 638         return ret;
 639
 640  out_unlock:
 641         __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
 642         preempt_enable();
 643         goto out_put;
 644 }
 645
 646 static void kvmppc_rmap_reset(struct kvm *kvm)
 647 {
 648         struct kvm_memslots *slots;
 649         struct kvm_memory_slot *memslot;
 650         int srcu_idx;
 651
 652         srcu_idx = srcu_read_lock(&kvm->srcu);
 653         slots = kvm->memslots;
 654         kvm_for_each_memslot(memslot, slots) {
 655                 /*
 656                  * This assumes it is acceptable to lose reference and
 657                  * change bits across a reset.
 658                  */
 659                 memset(memslot->arch.rmap, 0,
 660                        memslot->npages * sizeof(*memslot->arch.rmap));
 661         }
 662         srcu_read_unlock(&kvm->srcu, srcu_idx);
 663 }
 664
 665 static int kvm_handle_hva_range(struct kvm *kvm,
 666                                 unsigned long start,
 667                                 unsigned long end,
 668                                 int (*handler)(struct kvm *kvm,
 669                                                unsigned long *rmapp,
 670                                                unsigned long gfn))
 671 {
 672         int ret;
 673         int retval = 0;
 674         struct kvm_memslots *slots;
 675         struct kvm_memory_slot *memslot;
 676
 677         slots = kvm_memslots(kvm);
 678         kvm_for_each_memslot(memslot, slots) {
 679                 unsigned long hva_start, hva_end;
 680                 gfn_t gfn, gfn_end;
 681
 682                 hva_start = max(start, memslot->userspace_addr);
 683                 hva_end = min(end, memslot->userspace_addr +
 684                                         (memslot->npages << PAGE_SHIFT));
 685                 if (hva_start >= hva_end)
 686                         continue;
 687                 /*
 688                  * {gfn(page) | page intersects with [hva_start, hva_end)} =
 689                  * {gfn, gfn+1, ..., gfn_end-1}.
 690                  */
 691                 gfn = hva_to_gfn_memslot(hva_start, memslot);
 692                 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
 693
 694                 for (; gfn < gfn_end; ++gfn) {
 695                         gfn_t gfn_offset = gfn - memslot->base_gfn;
 696
 697                         ret = handler(kvm, &memslot->arch.rmap[gfn_offset], gfn);
 698                         retval |= ret;
 699                 }
 700         }
 701
 702         return retval;
 703 }
 704
 705 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
 706                           int (*handler)(struct kvm *kvm, unsigned long *rmapp,
 707                                          unsigned long gfn))
 708 {
 709         return kvm_handle_hva_range(kvm, hva, hva + 1, handler);
 710 }
 711
 712 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
 713                            unsigned long gfn)
 714 {
 715         struct revmap_entry *rev = kvm->arch.revmap;
 716         unsigned long h, i, j;
 717         __be64 *hptep;
 718         unsigned long ptel, psize, rcbits;
 719
 720         for (;;) {
 721                 lock_rmap(rmapp);
 722                 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 723                         unlock_rmap(rmapp);
 724                         break;
 725                 }
 726
 727                 /*
 728                  * To avoid an ABBA deadlock with the HPTE lock bit,
 729                  * we can't spin on the HPTE lock while holding the
 730                  * rmap chain lock.
 731                  */
 732                 i = *rmapp & KVMPPC_RMAP_INDEX;
 733                 hptep = (__be64 *) (kvm->arch.hpt_virt + (i << 4));
 734                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 735                         /* unlock rmap before spinning on the HPTE lock */
 736                         unlock_rmap(rmapp);
 737                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
 738                                 cpu_relax();
 739                         continue;
 740                 }
 741                 j = rev[i].forw;
 742                 if (j == i) {
 743                         /* chain is now empty */
 744                         *rmapp &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
 745                 } else {
 746                         /* remove i from chain */
 747                         h = rev[i].back;
 748                         rev[h].forw = j;
 749                         rev[j].back = h;
 750                         rev[i].forw = rev[i].back = i;
 751                         *rmapp = (*rmapp & ~KVMPPC_RMAP_INDEX) | j;
 752                 }
 753
 754                 /* Now check and modify the HPTE */
 755                 ptel = rev[i].guest_rpte;
 756                 psize = hpte_page_size(be64_to_cpu(hptep[0]), ptel);
 757                 if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
 758                     hpte_rpn(ptel, psize) == gfn) {
 759                         hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
 760                         kvmppc_invalidate_hpte(kvm, hptep, i);
 761                         /* Harvest R and C */
 762                         rcbits = be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
 763                         *rmapp |= rcbits << KVMPPC_RMAP_RC_SHIFT;
 764                         if (rcbits & ~rev[i].guest_rpte) {
 765                                 rev[i].guest_rpte = ptel | rcbits;
 766                                 note_hpte_modification(kvm, &rev[i]);
 767                         }
 768                 }
 769                 unlock_rmap(rmapp);
 770                 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
 771         }
 772         return 0;
 773 }
 774
 775 int kvm_unmap_hva_hv(struct kvm *kvm, unsigned long hva)
 776 {
 777         kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 778         return 0;
 779 }
 780
 781 int kvm_unmap_hva_range_hv(struct kvm *kvm, unsigned long start, unsigned long end)
 782 {
 783         kvm_handle_hva_range(kvm, start, end, kvm_unmap_rmapp);
 784         return 0;
 785 }
 786
 787 void kvmppc_core_flush_memslot_hv(struct kvm *kvm,
 788                                   struct kvm_memory_slot *memslot)
 789 {
 790         unsigned long *rmapp;
 791         unsigned long gfn;
 792         unsigned long n;
 793
 794         rmapp = memslot->arch.rmap;
 795         gfn = memslot->base_gfn;
 796         for (n = memslot->npages; n; --n) {
 797                 /*
 798                  * Testing the present bit without locking is OK because
 799                  * the memslot has been marked invalid already, and hence
 800                  * no new HPTEs referencing this page can be created,
 801                  * thus the present bit can't go from 0 to 1.
 802                  */
 803                 if (*rmapp & KVMPPC_RMAP_PRESENT)
 804                         kvm_unmap_rmapp(kvm, rmapp, gfn);
 805                 ++rmapp;
 806                 ++gfn;
 807         }
 808 }
 809
 810 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
 811                          unsigned long gfn)
 812 {
 813         struct revmap_entry *rev = kvm->arch.revmap;
 814         unsigned long head, i, j;
 815         __be64 *hptep;
 816         int ret = 0;
 817
 818  retry:
 819         lock_rmap(rmapp);
 820         if (*rmapp & KVMPPC_RMAP_REFERENCED) {
 821                 *rmapp &= ~KVMPPC_RMAP_REFERENCED;
 822                 ret = 1;
 823         }
 824         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 825                 unlock_rmap(rmapp);
 826                 return ret;
 827         }
 828
 829         i = head = *rmapp & KVMPPC_RMAP_INDEX;
 830         do {
 831                 hptep = (__be64 *) (kvm->arch.hpt_virt + (i << 4));
 832                 j = rev[i].forw;
 833
 834                 /* If this HPTE isn't referenced, ignore it */
 835                 if (!(be64_to_cpu(hptep[1]) & HPTE_R_R))
 836                         continue;
 837
 838                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 839                         /* unlock rmap before spinning on the HPTE lock */
 840                         unlock_rmap(rmapp);
 841                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
 842                                 cpu_relax();
 843                         goto retry;
 844                 }
 845
 846                 /* Now check and modify the HPTE */
 847                 if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
 848                     (be64_to_cpu(hptep[1]) & HPTE_R_R)) {
 849                         kvmppc_clear_ref_hpte(kvm, hptep, i);
 850                         if (!(rev[i].guest_rpte & HPTE_R_R)) {
 851                                 rev[i].guest_rpte |= HPTE_R_R;
 852                                 note_hpte_modification(kvm, &rev[i]);
 853                         }
 854                         ret = 1;
 855                 }
 856                 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
 857         } while ((i = j) != head);
 858
 859         unlock_rmap(rmapp);
 860         return ret;
 861 }
 862
 863 int kvm_age_hva_hv(struct kvm *kvm, unsigned long start, unsigned long end)
 864 {
 865         return kvm_handle_hva_range(kvm, start, end, kvm_age_rmapp);
 866 }
 867
 868 static int kvm_test_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
 869                               unsigned long gfn)
 870 {
 871         struct revmap_entry *rev = kvm->arch.revmap;
 872         unsigned long head, i, j;
 873         unsigned long *hp;
 874         int ret = 1;
 875
 876         if (*rmapp & KVMPPC_RMAP_REFERENCED)
 877                 return 1;
 878
 879         lock_rmap(rmapp);
 880         if (*rmapp & KVMPPC_RMAP_REFERENCED)
 881                 goto out;
 882
 883         if (*rmapp & KVMPPC_RMAP_PRESENT) {
 884                 i = head = *rmapp & KVMPPC_RMAP_INDEX;
 885                 do {
 886                         hp = (unsigned long *)(kvm->arch.hpt_virt + (i << 4));
 887                         j = rev[i].forw;
 888                         if (be64_to_cpu(hp[1]) & HPTE_R_R)
 889                                 goto out;
 890                 } while ((i = j) != head);
 891         }
 892         ret = 0;
 893
 894  out:
 895         unlock_rmap(rmapp);
 896         return ret;
 897 }
 898
 899 int kvm_test_age_hva_hv(struct kvm *kvm, unsigned long hva)
 900 {
 901         return kvm_handle_hva(kvm, hva, kvm_test_age_rmapp);
 902 }
 903
 904 void kvm_set_spte_hva_hv(struct kvm *kvm, unsigned long hva, pte_t pte)
 905 {
 906         kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
 907 }
 908
 909 static int vcpus_running(struct kvm *kvm)
 910 {
 911         return atomic_read(&kvm->arch.vcpus_running) != 0;
 912 }
 913
 914 /*
 915  * Returns the number of system pages that are dirty.
 916  * This can be more than 1 if we find a huge-page HPTE.
 917  */
 918 static int kvm_test_clear_dirty_npages(struct kvm *kvm, unsigned long *rmapp)
 919 {
 920         struct revmap_entry *rev = kvm->arch.revmap;
 921         unsigned long head, i, j;
 922         unsigned long n;
 923         unsigned long v, r;
 924         __be64 *hptep;
 925         int npages_dirty = 0;
 926
 927  retry:
 928         lock_rmap(rmapp);
 929         if (*rmapp & KVMPPC_RMAP_CHANGED) {
 930                 *rmapp &= ~KVMPPC_RMAP_CHANGED;
 931                 npages_dirty = 1;
 932         }
 933         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
 934                 unlock_rmap(rmapp);
 935                 return npages_dirty;
 936         }
 937
 938         i = head = *rmapp & KVMPPC_RMAP_INDEX;
 939         do {
 940                 unsigned long hptep1;
 941                 hptep = (__be64 *) (kvm->arch.hpt_virt + (i << 4));
 942                 j = rev[i].forw;
 943
 944                 /*
 945                  * Checking the C (changed) bit here is racy since there
 946                  * is no guarantee about when the hardware writes it back.
 947                  * If the HPTE is not writable then it is stable since the
 948                  * page can't be written to, and we would have done a tlbie
 949                  * (which forces the hardware to complete any writeback)
 950                  * when making the HPTE read-only.
 951                  * If vcpus are running then this call is racy anyway
 952                  * since the page could get dirtied subsequently, so we
 953                  * expect there to be a further call which would pick up
 954                  * any delayed C bit writeback.
 955                  * Otherwise we need to do the tlbie even if C==0 in
 956                  * order to pick up any delayed writeback of C.
 957                  */
 958                 hptep1 = be64_to_cpu(hptep[1]);
 959                 if (!(hptep1 & HPTE_R_C) &&
 960                     (!hpte_is_writable(hptep1) || vcpus_running(kvm)))
 961                         continue;
 962
 963                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
 964                         /* unlock rmap before spinning on the HPTE lock */
 965                         unlock_rmap(rmapp);
 966                         while (hptep[0] & cpu_to_be64(HPTE_V_HVLOCK))
 967                                 cpu_relax();
 968                         goto retry;
 969                 }
 970
 971                 /* Now check and modify the HPTE */
 972                 if (!(hptep[0] & cpu_to_be64(HPTE_V_VALID))) {
 973                         __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
 974                         continue;
 975                 }
 976
 977                 /* need to make it temporarily absent so C is stable */
 978                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
 979                 kvmppc_invalidate_hpte(kvm, hptep, i);
 980                 v = be64_to_cpu(hptep[0]);
 981                 r = be64_to_cpu(hptep[1]);
 982                 if (r & HPTE_R_C) {
 983                         hptep[1] = cpu_to_be64(r & ~HPTE_R_C);
 984                         if (!(rev[i].guest_rpte & HPTE_R_C)) {
 985                                 rev[i].guest_rpte |= HPTE_R_C;
 986                                 note_hpte_modification(kvm, &rev[i]);
 987                         }
 988                         n = hpte_page_size(v, r);
 989                         n = (n + PAGE_SIZE - 1) >> PAGE_SHIFT;
 990                         if (n > npages_dirty)
 991                                 npages_dirty = n;
 992                         eieio();
 993                 }
 994                 v &= ~HPTE_V_ABSENT;
 995                 v |= HPTE_V_VALID;
 996                 __unlock_hpte(hptep, v);
 997         } while ((i = j) != head);
 998
 999         unlock_rmap(rmapp);
1000         return npages_dirty;
1001 }
1002
1003 static void harvest_vpa_dirty(struct kvmppc_vpa *vpa,
1004                               struct kvm_memory_slot *memslot,
1005                               unsigned long *map)
1006 {
1007         unsigned long gfn;
1008
1009         if (!vpa->dirty || !vpa->pinned_addr)
1010                 return;
1011         gfn = vpa->gpa >> PAGE_SHIFT;
1012         if (gfn < memslot->base_gfn ||
1013             gfn >= memslot->base_gfn + memslot->npages)
1014                 return;
1015
1016         vpa->dirty = false;
1017         if (map)
1018                 __set_bit_le(gfn - memslot->base_gfn, map);
1019 }
1020
1021 long kvmppc_hv_get_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot,
1022                              unsigned long *map)
1023 {
1024         unsigned long i, j;
1025         unsigned long *rmapp;
1026         struct kvm_vcpu *vcpu;
1027
1028         preempt_disable();
1029         rmapp = memslot->arch.rmap;
1030         for (i = 0; i < memslot->npages; ++i) {
1031                 int npages = kvm_test_clear_dirty_npages(kvm, rmapp);
1032                 /*
1033                  * Note that if npages > 0 then i must be a multiple of npages,
1034                  * since we always put huge-page HPTEs in the rmap chain
1035                  * corresponding to their page base address.
1036                  */
1037                 if (npages && map)
1038                         for (j = i; npages; ++j, --npages)
1039                                 __set_bit_le(j, map);
1040                 ++rmapp;
1041         }
1042
1043         /* Harvest dirty bits from VPA and DTL updates */
1044         /* Note: we never modify the SLB shadow buffer areas */
1045         kvm_for_each_vcpu(i, vcpu, kvm) {
1046                 spin_lock(&vcpu->arch.vpa_update_lock);
1047                 harvest_vpa_dirty(&vcpu->arch.vpa, memslot, map);
1048                 harvest_vpa_dirty(&vcpu->arch.dtl, memslot, map);
1049                 spin_unlock(&vcpu->arch.vpa_update_lock);
1050         }
1051         preempt_enable();
1052         return 0;
1053 }
1054
1055 void *kvmppc_pin_guest_page(struct kvm *kvm, unsigned long gpa,
1056                             unsigned long *nb_ret)
1057 {
1058         struct kvm_memory_slot *memslot;
1059         unsigned long gfn = gpa >> PAGE_SHIFT;
1060         struct page *page, *pages[1];
1061         int npages;
1062         unsigned long hva, offset;
1063         int srcu_idx;
1064
1065         srcu_idx = srcu_read_lock(&kvm->srcu);
1066         memslot = gfn_to_memslot(kvm, gfn);
1067         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
1068                 goto err;
1069         hva = gfn_to_hva_memslot(memslot, gfn);
1070         npages = get_user_pages_fast(hva, 1, 1, pages);
1071         if (npages < 1)
1072                 goto err;
1073         page = pages[0];
1074         srcu_read_unlock(&kvm->srcu, srcu_idx);
1075
1076         offset = gpa & (PAGE_SIZE - 1);
1077         if (nb_ret)
1078                 *nb_ret = PAGE_SIZE - offset;
1079         return page_address(page) + offset;
1080
1081  err:
1082         srcu_read_unlock(&kvm->srcu, srcu_idx);
1083         return NULL;
1084 }
1085
1086 void kvmppc_unpin_guest_page(struct kvm *kvm, void *va, unsigned long gpa,
1087                              bool dirty)
1088 {
1089         struct page *page = virt_to_page(va);
1090         struct kvm_memory_slot *memslot;
1091         unsigned long gfn;
1092         unsigned long *rmap;
1093         int srcu_idx;
1094
1095         put_page(page);
1096
1097         if (!dirty)
1098                 return;
1099
1100         /* We need to mark this page dirty in the rmap chain */
1101         gfn = gpa >> PAGE_SHIFT;
1102         srcu_idx = srcu_read_lock(&kvm->srcu);
1103         memslot = gfn_to_memslot(kvm, gfn);
1104         if (memslot) {
1105                 rmap = &memslot->arch.rmap[gfn - memslot->base_gfn];
1106                 lock_rmap(rmap);
1107                 *rmap |= KVMPPC_RMAP_CHANGED;
1108                 unlock_rmap(rmap);
1109         }
1110         srcu_read_unlock(&kvm->srcu, srcu_idx);
1111 }
1112
1113 /*
1114  * Functions for reading and writing the hash table via reads and
1115  * writes on a file descriptor.
1116  *
1117  * Reads return the guest view of the hash table, which has to be
1118  * pieced together from the real hash table and the guest_rpte
1119  * values in the revmap array.
1120  *
1121  * On writes, each HPTE written is considered in turn, and if it
1122  * is valid, it is written to the HPT as if an H_ENTER with the
1123  * exact flag set was done.  When the invalid count is non-zero
1124  * in the header written to the stream, the kernel will make
1125  * sure that that many HPTEs are invalid, and invalidate them
1126  * if not.
1127  */
1128
1129 struct kvm_htab_ctx {
1130         unsigned long   index;
1131         unsigned long   flags;
1132         struct kvm      *kvm;
1133         int             first_pass;
1134 };
1135
1136 #define HPTE_SIZE       (2 * sizeof(unsigned long))
1137
1138 /*
1139  * Returns 1 if this HPT entry has been modified or has pending
1140  * R/C bit changes.
1141  */
1142 static int hpte_dirty(struct revmap_entry *revp, __be64 *hptp)
1143 {
1144         unsigned long rcbits_unset;
1145
1146         if (revp->guest_rpte & HPTE_GR_MODIFIED)
1147                 return 1;
1148
1149         /* Also need to consider changes in reference and changed bits */
1150         rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1151         if ((be64_to_cpu(hptp[0]) & HPTE_V_VALID) &&
1152             (be64_to_cpu(hptp[1]) & rcbits_unset))
1153                 return 1;
1154
1155         return 0;
1156 }
1157
1158 static long record_hpte(unsigned long flags, __be64 *hptp,
1159                         unsigned long *hpte, struct revmap_entry *revp,
1160                         int want_valid, int first_pass)
1161 {
1162         unsigned long v, r;
1163         unsigned long rcbits_unset;
1164         int ok = 1;
1165         int valid, dirty;
1166
1167         /* Unmodified entries are uninteresting except on the first pass */
1168         dirty = hpte_dirty(revp, hptp);
1169         if (!first_pass && !dirty)
1170                 return 0;
1171
1172         valid = 0;
1173         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1174                 valid = 1;
1175                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) &&
1176                     !(be64_to_cpu(hptp[0]) & HPTE_V_BOLTED))
1177                         valid = 0;
1178         }
1179         if (valid != want_valid)
1180                 return 0;
1181
1182         v = r = 0;
1183         if (valid || dirty) {
1184                 /* lock the HPTE so it's stable and read it */
1185                 preempt_disable();
1186                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
1187                         cpu_relax();
1188                 v = be64_to_cpu(hptp[0]);
1189
1190                 /* re-evaluate valid and dirty from synchronized HPTE value */
1191                 valid = !!(v & HPTE_V_VALID);
1192                 dirty = !!(revp->guest_rpte & HPTE_GR_MODIFIED);
1193
1194                 /* Harvest R and C into guest view if necessary */
1195                 rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1196                 if (valid && (rcbits_unset & be64_to_cpu(hptp[1]))) {
1197                         revp->guest_rpte |= (be64_to_cpu(hptp[1]) &
1198                                 (HPTE_R_R | HPTE_R_C)) | HPTE_GR_MODIFIED;
1199                         dirty = 1;
1200                 }
1201
1202                 if (v & HPTE_V_ABSENT) {
1203                         v &= ~HPTE_V_ABSENT;
1204                         v |= HPTE_V_VALID;
1205                         valid = 1;
1206                 }
1207                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) && !(v & HPTE_V_BOLTED))
1208                         valid = 0;
1209
1210                 r = revp->guest_rpte;
1211                 /* only clear modified if this is the right sort of entry */
1212                 if (valid == want_valid && dirty) {
1213                         r &= ~HPTE_GR_MODIFIED;
1214                         revp->guest_rpte = r;
1215                 }
1216                 unlock_hpte(hptp, be64_to_cpu(hptp[0]));
1217                 preempt_enable();
1218                 if (!(valid == want_valid && (first_pass || dirty)))
1219                         ok = 0;
1220         }
1221         hpte[0] = cpu_to_be64(v);
1222         hpte[1] = cpu_to_be64(r);
1223         return ok;
1224 }
1225
1226 static ssize_t kvm_htab_read(struct file *file, char __user *buf,
1227                              size_t count, loff_t *ppos)
1228 {
1229         struct kvm_htab_ctx *ctx = file->private_data;
1230         struct kvm *kvm = ctx->kvm;
1231         struct kvm_get_htab_header hdr;
1232         __be64 *hptp;
1233         struct revmap_entry *revp;
1234         unsigned long i, nb, nw;
1235         unsigned long __user *lbuf;
1236         struct kvm_get_htab_header __user *hptr;
1237         unsigned long flags;
1238         int first_pass;
1239         unsigned long hpte[2];
1240
1241         if (!access_ok(VERIFY_WRITE, buf, count))
1242                 return -EFAULT;
1243
1244         first_pass = ctx->first_pass;
1245         flags = ctx->flags;
1246
1247         i = ctx->index;
1248         hptp = (__be64 *)(kvm->arch.hpt_virt + (i * HPTE_SIZE));
1249         revp = kvm->arch.revmap + i;
1250         lbuf = (unsigned long __user *)buf;
1251
1252         nb = 0;
1253         while (nb + sizeof(hdr) + HPTE_SIZE < count) {
1254                 /* Initialize header */
1255                 hptr = (struct kvm_get_htab_header __user *)buf;
1256                 hdr.n_valid = 0;
1257                 hdr.n_invalid = 0;
1258                 nw = nb;
1259                 nb += sizeof(hdr);
1260                 lbuf = (unsigned long __user *)(buf + sizeof(hdr));
1261
1262                 /* Skip uninteresting entries, i.e. clean on not-first pass */
1263                 if (!first_pass) {
1264                         while (i < kvm->arch.hpt_npte &&
1265                                !hpte_dirty(revp, hptp)) {
1266                                 ++i;
1267                                 hptp += 2;
1268                                 ++revp;
1269                         }
1270                 }
1271                 hdr.index = i;
1272
1273                 /* Grab a series of valid entries */
1274                 while (i < kvm->arch.hpt_npte &&
1275                        hdr.n_valid < 0xffff &&
1276                        nb + HPTE_SIZE < count &&
1277                        record_hpte(flags, hptp, hpte, revp, 1, first_pass)) {
1278                         /* valid entry, write it out */
1279                         ++hdr.n_valid;
1280                         if (__put_user(hpte[0], lbuf) ||
1281                             __put_user(hpte[1], lbuf + 1))
1282                                 return -EFAULT;
1283                         nb += HPTE_SIZE;
1284                         lbuf += 2;
1285                         ++i;
1286                         hptp += 2;
1287                         ++revp;
1288                 }
1289                 /* Now skip invalid entries while we can */
1290                 while (i < kvm->arch.hpt_npte &&
1291                        hdr.n_invalid < 0xffff &&
1292                        record_hpte(flags, hptp, hpte, revp, 0, first_pass)) {
1293                         /* found an invalid entry */
1294                         ++hdr.n_invalid;
1295                         ++i;
1296                         hptp += 2;
1297                         ++revp;
1298                 }
1299
1300                 if (hdr.n_valid || hdr.n_invalid) {
1301                         /* write back the header */
1302                         if (__copy_to_user(hptr, &hdr, sizeof(hdr)))
1303                                 return -EFAULT;
1304                         nw = nb;
1305                         buf = (char __user *)lbuf;
1306                 } else {
1307                         nb = nw;
1308                 }
1309
1310                 /* Check if we've wrapped around the hash table */
1311                 if (i >= kvm->arch.hpt_npte) {
1312                         i = 0;
1313                         ctx->first_pass = 0;
1314                         break;
1315                 }
1316         }
1317
1318         ctx->index = i;
1319
1320         return nb;
1321 }
1322
1323 static ssize_t kvm_htab_write(struct file *file, const char __user *buf,
1324                               size_t count, loff_t *ppos)
1325 {
1326         struct kvm_htab_ctx *ctx = file->private_data;
1327         struct kvm *kvm = ctx->kvm;
1328         struct kvm_get_htab_header hdr;
1329         unsigned long i, j;
1330         unsigned long v, r;
1331         unsigned long __user *lbuf;
1332         __be64 *hptp;
1333         unsigned long tmp[2];
1334         ssize_t nb;
1335         long int err, ret;
1336         int hpte_setup;
1337
1338         if (!access_ok(VERIFY_READ, buf, count))
1339                 return -EFAULT;
1340
1341         /* lock out vcpus from running while we're doing this */
1342         mutex_lock(&kvm->lock);
1343         hpte_setup = kvm->arch.hpte_setup_done;
1344         if (hpte_setup) {
1345                 kvm->arch.hpte_setup_done = 0;  /* temporarily */
1346                 /* order hpte_setup_done vs. vcpus_running */
1347                 smp_mb();
1348                 if (atomic_read(&kvm->arch.vcpus_running)) {
1349                         kvm->arch.hpte_setup_done = 1;
1350                         mutex_unlock(&kvm->lock);
1351                         return -EBUSY;
1352                 }
1353         }
1354
1355         err = 0;
1356         for (nb = 0; nb + sizeof(hdr) <= count; ) {
1357                 err = -EFAULT;
1358                 if (__copy_from_user(&hdr, buf, sizeof(hdr)))
1359                         break;
1360
1361                 err = 0;
1362                 if (nb + hdr.n_valid * HPTE_SIZE > count)
1363                         break;
1364
1365                 nb += sizeof(hdr);
1366                 buf += sizeof(hdr);
1367
1368                 err = -EINVAL;
1369                 i = hdr.index;
1370                 if (i >= kvm->arch.hpt_npte ||
1371                     i + hdr.n_valid + hdr.n_invalid > kvm->arch.hpt_npte)
1372                         break;
1373
1374                 hptp = (__be64 *)(kvm->arch.hpt_virt + (i * HPTE_SIZE));
1375                 lbuf = (unsigned long __user *)buf;
1376                 for (j = 0; j < hdr.n_valid; ++j) {
1377                         __be64 hpte_v;
1378                         __be64 hpte_r;
1379
1380                         err = -EFAULT;
1381                         if (__get_user(hpte_v, lbuf) ||
1382                             __get_user(hpte_r, lbuf + 1))
1383                                 goto out;
1384                         v = be64_to_cpu(hpte_v);
1385                         r = be64_to_cpu(hpte_r);
1386                         err = -EINVAL;
1387                         if (!(v & HPTE_V_VALID))
1388                                 goto out;
1389                         lbuf += 2;
1390                         nb += HPTE_SIZE;
1391
1392                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1393                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1394                         err = -EIO;
1395                         ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, i, v, r,
1396                                                          tmp);
1397                         if (ret != H_SUCCESS) {
1398                                 pr_err("kvm_htab_write ret %ld i=%ld v=%lx "
1399                                        "r=%lx\n", ret, i, v, r);
1400                                 goto out;
1401                         }
1402                         if (!hpte_setup && is_vrma_hpte(v)) {
1403                                 unsigned long psize = hpte_base_page_size(v, r);
1404                                 unsigned long senc = slb_pgsize_encoding(psize);
1405                                 unsigned long lpcr;
1406
1407                                 kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
1408                                         (VRMA_VSID << SLB_VSID_SHIFT_1T);
1409                                 lpcr = senc << (LPCR_VRMASD_SH - 4);
1410                                 kvmppc_update_lpcr(kvm, lpcr, LPCR_VRMASD);
1411                                 hpte_setup = 1;
1412                         }
1413                         ++i;
1414                         hptp += 2;
1415                 }
1416
1417                 for (j = 0; j < hdr.n_invalid; ++j) {
1418                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1419                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1420                         ++i;
1421                         hptp += 2;
1422                 }
1423                 err = 0;
1424         }
1425
1426  out:
1427         /* Order HPTE updates vs. hpte_setup_done */
1428         smp_wmb();
1429         kvm->arch.hpte_setup_done = hpte_setup;
1430         mutex_unlock(&kvm->lock);
1431
1432         if (err)
1433                 return err;
1434         return nb;
1435 }
1436
1437 static int kvm_htab_release(struct inode *inode, struct file *filp)
1438 {
1439         struct kvm_htab_ctx *ctx = filp->private_data;
1440
1441         filp->private_data = NULL;
1442         if (!(ctx->flags & KVM_GET_HTAB_WRITE))
1443                 atomic_dec(&ctx->kvm->arch.hpte_mod_interest);
1444         kvm_put_kvm(ctx->kvm);
1445         kfree(ctx);
1446         return 0;
1447 }
1448
1449 static const struct file_operations kvm_htab_fops = {
1450         .read           = kvm_htab_read,
1451         .write          = kvm_htab_write,
1452         .llseek         = default_llseek,
1453         .release        = kvm_htab_release,
1454 };
1455
1456 int kvm_vm_ioctl_get_htab_fd(struct kvm *kvm, struct kvm_get_htab_fd *ghf)
1457 {
1458         int ret;
1459         struct kvm_htab_ctx *ctx;
1460         int rwflag;
1461
1462         /* reject flags we don't recognize */
1463         if (ghf->flags & ~(KVM_GET_HTAB_BOLTED_ONLY | KVM_GET_HTAB_WRITE))
1464                 return -EINVAL;
1465         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1466         if (!ctx)
1467                 return -ENOMEM;
1468         kvm_get_kvm(kvm);
1469         ctx->kvm = kvm;
1470         ctx->index = ghf->start_index;
1471         ctx->flags = ghf->flags;
1472         ctx->first_pass = 1;
1473
1474         rwflag = (ghf->flags & KVM_GET_HTAB_WRITE) ? O_WRONLY : O_RDONLY;
1475         ret = anon_inode_getfd("kvm-htab", &kvm_htab_fops, ctx, rwflag | O_CLOEXEC);
1476         if (ret < 0) {
1477                 kvm_put_kvm(kvm);
1478                 return ret;
1479         }
1480
1481         if (rwflag == O_RDONLY) {
1482                 mutex_lock(&kvm->slots_lock);
1483                 atomic_inc(&kvm->arch.hpte_mod_interest);
1484                 /* make sure kvmppc_do_h_enter etc. see the increment */
1485                 synchronize_srcu_expedited(&kvm->srcu);
1486                 mutex_unlock(&kvm->slots_lock);
1487         }
1488
1489         return ret;
1490 }
1491
1492 struct debugfs_htab_state {
1493         struct kvm      *kvm;
1494         struct mutex    mutex;
1495         unsigned long   hpt_index;
1496         int             chars_left;
1497         int             buf_index;
1498         char            buf[64];
1499 };
1500
1501 static int debugfs_htab_open(struct inode *inode, struct file *file)
1502 {
1503         struct kvm *kvm = inode->i_private;
1504         struct debugfs_htab_state *p;
1505
1506         p = kzalloc(sizeof(*p), GFP_KERNEL);
1507         if (!p)
1508                 return -ENOMEM;
1509
1510         kvm_get_kvm(kvm);
1511         p->kvm = kvm;
1512         mutex_init(&p->mutex);
1513         file->private_data = p;
1514
1515         return nonseekable_open(inode, file);
1516 }
1517
1518 static int debugfs_htab_release(struct inode *inode, struct file *file)
1519 {
1520         struct debugfs_htab_state *p = file->private_data;
1521
1522         kvm_put_kvm(p->kvm);
1523         kfree(p);
1524         return 0;
1525 }
1526
1527 static ssize_t debugfs_htab_read(struct file *file, char __user *buf,
1528                                  size_t len, loff_t *ppos)
1529 {
1530         struct debugfs_htab_state *p = file->private_data;
1531         ssize_t ret, r;
1532         unsigned long i, n;
1533         unsigned long v, hr, gr;
1534         struct kvm *kvm;
1535         __be64 *hptp;
1536
1537         ret = mutex_lock_interruptible(&p->mutex);
1538         if (ret)
1539                 return ret;
1540
1541         if (p->chars_left) {
1542                 n = p->chars_left;
1543                 if (n > len)
1544                         n = len;
1545                 r = copy_to_user(buf, p->buf + p->buf_index, n);
1546                 n -= r;
1547                 p->chars_left -= n;
1548                 p->buf_index += n;
1549                 buf += n;
1550                 len -= n;
1551                 ret = n;
1552                 if (r) {
1553                         if (!n)
1554                                 ret = -EFAULT;
1555                         goto out;
1556                 }
1557         }
1558
1559         kvm = p->kvm;
1560         i = p->hpt_index;
1561         hptp = (__be64 *)(kvm->arch.hpt_virt + (i * HPTE_SIZE));
1562         for (; len != 0 && i < kvm->arch.hpt_npte; ++i, hptp += 2) {
1563                 if (!(be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)))
1564                         continue;
1565
1566                 /* lock the HPTE so it's stable and read it */
1567                 preempt_disable();
1568                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
1569                         cpu_relax();
1570                 v = be64_to_cpu(hptp[0]) & ~HPTE_V_HVLOCK;
1571                 hr = be64_to_cpu(hptp[1]);
1572                 gr = kvm->arch.revmap[i].guest_rpte;
1573                 unlock_hpte(hptp, v);
1574                 preempt_enable();
1575
1576                 if (!(v & (HPTE_V_VALID | HPTE_V_ABSENT)))
1577                         continue;
1578
1579                 n = scnprintf(p->buf, sizeof(p->buf),
1580                               "%6lx %.16lx %.16lx %.16lx\n",
1581                               i, v, hr, gr);
1582                 p->chars_left = n;
1583                 if (n > len)
1584                         n = len;
1585                 r = copy_to_user(buf, p->buf, n);
1586                 n -= r;
1587                 p->chars_left -= n;
1588                 p->buf_index = n;
1589                 buf += n;
1590                 len -= n;
1591                 ret += n;
1592                 if (r) {
1593                         if (!ret)
1594                                 ret = -EFAULT;
1595                         goto out;
1596                 }
1597         }
1598         p->hpt_index = i;
1599
1600  out:
1601         mutex_unlock(&p->mutex);
1602         return ret;
1603 }
1604
1605 ssize_t debugfs_htab_write(struct file *file, const char __user *buf,
1606                            size_t len, loff_t *ppos)
1607 {
1608         return -EACCES;
1609 }
1610
1611 static const struct file_operations debugfs_htab_fops = {
1612         .owner   = THIS_MODULE,
1613         .open    = debugfs_htab_open,
1614         .release = debugfs_htab_release,
1615         .read    = debugfs_htab_read,
1616         .write   = debugfs_htab_write,
1617         .llseek  = generic_file_llseek,
1618 };
1619
1620 void kvmppc_mmu_debugfs_init(struct kvm *kvm)
1621 {
1622         kvm->arch.htab_dentry = debugfs_create_file("htab", 0400,
1623                                                     kvm->arch.debugfs_dir, kvm,
1624                                                     &debugfs_htab_fops);
1625 }
1626
1627 void kvmppc_mmu_book3s_hv_init(struct kvm_vcpu *vcpu)
1628 {
1629         struct kvmppc_mmu *mmu = &vcpu->arch.mmu;
1630
1631         vcpu->arch.slb_nr = 32;         /* POWER7/POWER8 */
1632
1633         mmu->xlate = kvmppc_mmu_book3s_64_hv_xlate;
1634         mmu->reset_msr = kvmppc_mmu_book3s_64_hv_reset_msr;
1635
1636         vcpu->arch.hflags |= BOOK3S_HFLAG_SLB;
1637 }