Merge tag 'kvm-ppc-next-4.18-2' of git://git.kernel.org/pub/scm/linux/kernel/git...
[sfrench/cifs-2.6.git] / arch / powerpc / kvm / book3s_64_mmu_hv.c
1 /*
2  * This program is free software; you can redistribute it and/or modify
3  * it under the terms of the GNU General Public License, version 2, as
4  * published by the Free Software Foundation.
5  *
6  * This program is distributed in the hope that it will be useful,
7  * but WITHOUT ANY WARRANTY; without even the implied warranty of
8  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
9  * GNU General Public License for more details.
10  *
11  * You should have received a copy of the GNU General Public License
12  * along with this program; if not, write to the Free Software
13  * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
14  *
15  * Copyright 2010 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
16  */
17
18 #include <linux/types.h>
19 #include <linux/string.h>
20 #include <linux/kvm.h>
21 #include <linux/kvm_host.h>
22 #include <linux/highmem.h>
23 #include <linux/gfp.h>
24 #include <linux/slab.h>
25 #include <linux/hugetlb.h>
26 #include <linux/vmalloc.h>
27 #include <linux/srcu.h>
28 #include <linux/anon_inodes.h>
29 #include <linux/file.h>
30 #include <linux/debugfs.h>
31
32 #include <asm/tlbflush.h>
33 #include <asm/kvm_ppc.h>
34 #include <asm/kvm_book3s.h>
35 #include <asm/book3s/64/mmu-hash.h>
36 #include <asm/hvcall.h>
37 #include <asm/synch.h>
38 #include <asm/ppc-opcode.h>
39 #include <asm/cputable.h>
40 #include <asm/pte-walk.h>
41
42 #include "trace_hv.h"
43
44 //#define DEBUG_RESIZE_HPT      1
45
46 #ifdef DEBUG_RESIZE_HPT
47 #define resize_hpt_debug(resize, ...)                           \
48         do {                                                    \
49                 printk(KERN_DEBUG "RESIZE HPT %p: ", resize);   \
50                 printk(__VA_ARGS__);                            \
51         } while (0)
52 #else
53 #define resize_hpt_debug(resize, ...)                           \
54         do { } while (0)
55 #endif
56
57 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
58                                 long pte_index, unsigned long pteh,
59                                 unsigned long ptel, unsigned long *pte_idx_ret);
60
61 struct kvm_resize_hpt {
62         /* These fields read-only after init */
63         struct kvm *kvm;
64         struct work_struct work;
65         u32 order;
66
67         /* These fields protected by kvm->lock */
68
69         /* Possible values and their usage:
70          *  <0     an error occurred during allocation,
71          *  -EBUSY allocation is in the progress,
72          *  0      allocation made successfuly.
73          */
74         int error;
75
76         /* Private to the work thread, until error != -EBUSY,
77          * then protected by kvm->lock.
78          */
79         struct kvm_hpt_info hpt;
80 };
81
82 int kvmppc_allocate_hpt(struct kvm_hpt_info *info, u32 order)
83 {
84         unsigned long hpt = 0;
85         int cma = 0;
86         struct page *page = NULL;
87         struct revmap_entry *rev;
88         unsigned long npte;
89
90         if ((order < PPC_MIN_HPT_ORDER) || (order > PPC_MAX_HPT_ORDER))
91                 return -EINVAL;
92
93         page = kvm_alloc_hpt_cma(1ul << (order - PAGE_SHIFT));
94         if (page) {
95                 hpt = (unsigned long)pfn_to_kaddr(page_to_pfn(page));
96                 memset((void *)hpt, 0, (1ul << order));
97                 cma = 1;
98         }
99
100         if (!hpt)
101                 hpt = __get_free_pages(GFP_KERNEL|__GFP_ZERO|__GFP_RETRY_MAYFAIL
102                                        |__GFP_NOWARN, order - PAGE_SHIFT);
103
104         if (!hpt)
105                 return -ENOMEM;
106
107         /* HPTEs are 2**4 bytes long */
108         npte = 1ul << (order - 4);
109
110         /* Allocate reverse map array */
111         rev = vmalloc(array_size(npte, sizeof(struct revmap_entry)));
112         if (!rev) {
113                 if (cma)
114                         kvm_free_hpt_cma(page, 1 << (order - PAGE_SHIFT));
115                 else
116                         free_pages(hpt, order - PAGE_SHIFT);
117                 return -ENOMEM;
118         }
119
120         info->order = order;
121         info->virt = hpt;
122         info->cma = cma;
123         info->rev = rev;
124
125         return 0;
126 }
127
128 void kvmppc_set_hpt(struct kvm *kvm, struct kvm_hpt_info *info)
129 {
130         atomic64_set(&kvm->arch.mmio_update, 0);
131         kvm->arch.hpt = *info;
132         kvm->arch.sdr1 = __pa(info->virt) | (info->order - 18);
133
134         pr_debug("KVM guest htab at %lx (order %ld), LPID %x\n",
135                  info->virt, (long)info->order, kvm->arch.lpid);
136 }
137
138 long kvmppc_alloc_reset_hpt(struct kvm *kvm, int order)
139 {
140         long err = -EBUSY;
141         struct kvm_hpt_info info;
142
143         mutex_lock(&kvm->lock);
144         if (kvm->arch.mmu_ready) {
145                 kvm->arch.mmu_ready = 0;
146                 /* order mmu_ready vs. vcpus_running */
147                 smp_mb();
148                 if (atomic_read(&kvm->arch.vcpus_running)) {
149                         kvm->arch.mmu_ready = 1;
150                         goto out;
151                 }
152         }
153         if (kvm_is_radix(kvm)) {
154                 err = kvmppc_switch_mmu_to_hpt(kvm);
155                 if (err)
156                         goto out;
157         }
158
159         if (kvm->arch.hpt.order == order) {
160                 /* We already have a suitable HPT */
161
162                 /* Set the entire HPT to 0, i.e. invalid HPTEs */
163                 memset((void *)kvm->arch.hpt.virt, 0, 1ul << order);
164                 /*
165                  * Reset all the reverse-mapping chains for all memslots
166                  */
167                 kvmppc_rmap_reset(kvm);
168                 err = 0;
169                 goto out;
170         }
171
172         if (kvm->arch.hpt.virt) {
173                 kvmppc_free_hpt(&kvm->arch.hpt);
174                 kvmppc_rmap_reset(kvm);
175         }
176
177         err = kvmppc_allocate_hpt(&info, order);
178         if (err < 0)
179                 goto out;
180         kvmppc_set_hpt(kvm, &info);
181
182 out:
183         if (err == 0)
184                 /* Ensure that each vcpu will flush its TLB on next entry. */
185                 cpumask_setall(&kvm->arch.need_tlb_flush);
186
187         mutex_unlock(&kvm->lock);
188         return err;
189 }
190
191 void kvmppc_free_hpt(struct kvm_hpt_info *info)
192 {
193         vfree(info->rev);
194         info->rev = NULL;
195         if (info->cma)
196                 kvm_free_hpt_cma(virt_to_page(info->virt),
197                                  1 << (info->order - PAGE_SHIFT));
198         else if (info->virt)
199                 free_pages(info->virt, info->order - PAGE_SHIFT);
200         info->virt = 0;
201         info->order = 0;
202 }
203
204 /* Bits in first HPTE dword for pagesize 4k, 64k or 16M */
205 static inline unsigned long hpte0_pgsize_encoding(unsigned long pgsize)
206 {
207         return (pgsize > 0x1000) ? HPTE_V_LARGE : 0;
208 }
209
210 /* Bits in second HPTE dword for pagesize 4k, 64k or 16M */
211 static inline unsigned long hpte1_pgsize_encoding(unsigned long pgsize)
212 {
213         return (pgsize == 0x10000) ? 0x1000 : 0;
214 }
215
216 void kvmppc_map_vrma(struct kvm_vcpu *vcpu, struct kvm_memory_slot *memslot,
217                      unsigned long porder)
218 {
219         unsigned long i;
220         unsigned long npages;
221         unsigned long hp_v, hp_r;
222         unsigned long addr, hash;
223         unsigned long psize;
224         unsigned long hp0, hp1;
225         unsigned long idx_ret;
226         long ret;
227         struct kvm *kvm = vcpu->kvm;
228
229         psize = 1ul << porder;
230         npages = memslot->npages >> (porder - PAGE_SHIFT);
231
232         /* VRMA can't be > 1TB */
233         if (npages > 1ul << (40 - porder))
234                 npages = 1ul << (40 - porder);
235         /* Can't use more than 1 HPTE per HPTEG */
236         if (npages > kvmppc_hpt_mask(&kvm->arch.hpt) + 1)
237                 npages = kvmppc_hpt_mask(&kvm->arch.hpt) + 1;
238
239         hp0 = HPTE_V_1TB_SEG | (VRMA_VSID << (40 - 16)) |
240                 HPTE_V_BOLTED | hpte0_pgsize_encoding(psize);
241         hp1 = hpte1_pgsize_encoding(psize) |
242                 HPTE_R_R | HPTE_R_C | HPTE_R_M | PP_RWXX;
243
244         for (i = 0; i < npages; ++i) {
245                 addr = i << porder;
246                 /* can't use hpt_hash since va > 64 bits */
247                 hash = (i ^ (VRMA_VSID ^ (VRMA_VSID << 25)))
248                         & kvmppc_hpt_mask(&kvm->arch.hpt);
249                 /*
250                  * We assume that the hash table is empty and no
251                  * vcpus are using it at this stage.  Since we create
252                  * at most one HPTE per HPTEG, we just assume entry 7
253                  * is available and use it.
254                  */
255                 hash = (hash << 3) + 7;
256                 hp_v = hp0 | ((addr >> 16) & ~0x7fUL);
257                 hp_r = hp1 | addr;
258                 ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, hash, hp_v, hp_r,
259                                                  &idx_ret);
260                 if (ret != H_SUCCESS) {
261                         pr_err("KVM: map_vrma at %lx failed, ret=%ld\n",
262                                addr, ret);
263                         break;
264                 }
265         }
266 }
267
268 int kvmppc_mmu_hv_init(void)
269 {
270         unsigned long host_lpid, rsvd_lpid;
271
272         if (!cpu_has_feature(CPU_FTR_HVMODE))
273                 return -EINVAL;
274
275         if (!mmu_has_feature(MMU_FTR_LOCKLESS_TLBIE))
276                 return -EINVAL;
277
278         /* POWER7 has 10-bit LPIDs (12-bit in POWER8) */
279         host_lpid = mfspr(SPRN_LPID);
280         rsvd_lpid = LPID_RSVD;
281
282         kvmppc_init_lpid(rsvd_lpid + 1);
283
284         kvmppc_claim_lpid(host_lpid);
285         /* rsvd_lpid is reserved for use in partition switching */
286         kvmppc_claim_lpid(rsvd_lpid);
287
288         return 0;
289 }
290
291 static void kvmppc_mmu_book3s_64_hv_reset_msr(struct kvm_vcpu *vcpu)
292 {
293         unsigned long msr = vcpu->arch.intr_msr;
294
295         /* If transactional, change to suspend mode on IRQ delivery */
296         if (MSR_TM_TRANSACTIONAL(vcpu->arch.shregs.msr))
297                 msr |= MSR_TS_S;
298         else
299                 msr |= vcpu->arch.shregs.msr & MSR_TS_MASK;
300         kvmppc_set_msr(vcpu, msr);
301 }
302
303 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
304                                 long pte_index, unsigned long pteh,
305                                 unsigned long ptel, unsigned long *pte_idx_ret)
306 {
307         long ret;
308
309         /* Protect linux PTE lookup from page table destruction */
310         rcu_read_lock_sched();  /* this disables preemption too */
311         ret = kvmppc_do_h_enter(kvm, flags, pte_index, pteh, ptel,
312                                 current->mm->pgd, false, pte_idx_ret);
313         rcu_read_unlock_sched();
314         if (ret == H_TOO_HARD) {
315                 /* this can't happen */
316                 pr_err("KVM: Oops, kvmppc_h_enter returned too hard!\n");
317                 ret = H_RESOURCE;       /* or something */
318         }
319         return ret;
320
321 }
322
323 static struct kvmppc_slb *kvmppc_mmu_book3s_hv_find_slbe(struct kvm_vcpu *vcpu,
324                                                          gva_t eaddr)
325 {
326         u64 mask;
327         int i;
328
329         for (i = 0; i < vcpu->arch.slb_nr; i++) {
330                 if (!(vcpu->arch.slb[i].orige & SLB_ESID_V))
331                         continue;
332
333                 if (vcpu->arch.slb[i].origv & SLB_VSID_B_1T)
334                         mask = ESID_MASK_1T;
335                 else
336                         mask = ESID_MASK;
337
338                 if (((vcpu->arch.slb[i].orige ^ eaddr) & mask) == 0)
339                         return &vcpu->arch.slb[i];
340         }
341         return NULL;
342 }
343
344 static unsigned long kvmppc_mmu_get_real_addr(unsigned long v, unsigned long r,
345                         unsigned long ea)
346 {
347         unsigned long ra_mask;
348
349         ra_mask = kvmppc_actual_pgsz(v, r) - 1;
350         return (r & HPTE_R_RPN & ~ra_mask) | (ea & ra_mask);
351 }
352
353 static int kvmppc_mmu_book3s_64_hv_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
354                         struct kvmppc_pte *gpte, bool data, bool iswrite)
355 {
356         struct kvm *kvm = vcpu->kvm;
357         struct kvmppc_slb *slbe;
358         unsigned long slb_v;
359         unsigned long pp, key;
360         unsigned long v, orig_v, gr;
361         __be64 *hptep;
362         int index;
363         int virtmode = vcpu->arch.shregs.msr & (data ? MSR_DR : MSR_IR);
364
365         if (kvm_is_radix(vcpu->kvm))
366                 return kvmppc_mmu_radix_xlate(vcpu, eaddr, gpte, data, iswrite);
367
368         /* Get SLB entry */
369         if (virtmode) {
370                 slbe = kvmppc_mmu_book3s_hv_find_slbe(vcpu, eaddr);
371                 if (!slbe)
372                         return -EINVAL;
373                 slb_v = slbe->origv;
374         } else {
375                 /* real mode access */
376                 slb_v = vcpu->kvm->arch.vrma_slb_v;
377         }
378
379         preempt_disable();
380         /* Find the HPTE in the hash table */
381         index = kvmppc_hv_find_lock_hpte(kvm, eaddr, slb_v,
382                                          HPTE_V_VALID | HPTE_V_ABSENT);
383         if (index < 0) {
384                 preempt_enable();
385                 return -ENOENT;
386         }
387         hptep = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
388         v = orig_v = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
389         if (cpu_has_feature(CPU_FTR_ARCH_300))
390                 v = hpte_new_to_old_v(v, be64_to_cpu(hptep[1]));
391         gr = kvm->arch.hpt.rev[index].guest_rpte;
392
393         unlock_hpte(hptep, orig_v);
394         preempt_enable();
395
396         gpte->eaddr = eaddr;
397         gpte->vpage = ((v & HPTE_V_AVPN) << 4) | ((eaddr >> 12) & 0xfff);
398
399         /* Get PP bits and key for permission check */
400         pp = gr & (HPTE_R_PP0 | HPTE_R_PP);
401         key = (vcpu->arch.shregs.msr & MSR_PR) ? SLB_VSID_KP : SLB_VSID_KS;
402         key &= slb_v;
403
404         /* Calculate permissions */
405         gpte->may_read = hpte_read_permission(pp, key);
406         gpte->may_write = hpte_write_permission(pp, key);
407         gpte->may_execute = gpte->may_read && !(gr & (HPTE_R_N | HPTE_R_G));
408
409         /* Storage key permission check for POWER7 */
410         if (data && virtmode) {
411                 int amrfield = hpte_get_skey_perm(gr, vcpu->arch.amr);
412                 if (amrfield & 1)
413                         gpte->may_read = 0;
414                 if (amrfield & 2)
415                         gpte->may_write = 0;
416         }
417
418         /* Get the guest physical address */
419         gpte->raddr = kvmppc_mmu_get_real_addr(v, gr, eaddr);
420         return 0;
421 }
422
423 /*
424  * Quick test for whether an instruction is a load or a store.
425  * If the instruction is a load or a store, then this will indicate
426  * which it is, at least on server processors.  (Embedded processors
427  * have some external PID instructions that don't follow the rule
428  * embodied here.)  If the instruction isn't a load or store, then
429  * this doesn't return anything useful.
430  */
431 static int instruction_is_store(unsigned int instr)
432 {
433         unsigned int mask;
434
435         mask = 0x10000000;
436         if ((instr & 0xfc000000) == 0x7c000000)
437                 mask = 0x100;           /* major opcode 31 */
438         return (instr & mask) != 0;
439 }
440
441 int kvmppc_hv_emulate_mmio(struct kvm_run *run, struct kvm_vcpu *vcpu,
442                            unsigned long gpa, gva_t ea, int is_store)
443 {
444         u32 last_inst;
445
446         /*
447          * If we fail, we just return to the guest and try executing it again.
448          */
449         if (kvmppc_get_last_inst(vcpu, INST_GENERIC, &last_inst) !=
450                 EMULATE_DONE)
451                 return RESUME_GUEST;
452
453         /*
454          * WARNING: We do not know for sure whether the instruction we just
455          * read from memory is the same that caused the fault in the first
456          * place.  If the instruction we read is neither an load or a store,
457          * then it can't access memory, so we don't need to worry about
458          * enforcing access permissions.  So, assuming it is a load or
459          * store, we just check that its direction (load or store) is
460          * consistent with the original fault, since that's what we
461          * checked the access permissions against.  If there is a mismatch
462          * we just return and retry the instruction.
463          */
464
465         if (instruction_is_store(last_inst) != !!is_store)
466                 return RESUME_GUEST;
467
468         /*
469          * Emulated accesses are emulated by looking at the hash for
470          * translation once, then performing the access later. The
471          * translation could be invalidated in the meantime in which
472          * point performing the subsequent memory access on the old
473          * physical address could possibly be a security hole for the
474          * guest (but not the host).
475          *
476          * This is less of an issue for MMIO stores since they aren't
477          * globally visible. It could be an issue for MMIO loads to
478          * a certain extent but we'll ignore it for now.
479          */
480
481         vcpu->arch.paddr_accessed = gpa;
482         vcpu->arch.vaddr_accessed = ea;
483         return kvmppc_emulate_mmio(run, vcpu);
484 }
485
486 int kvmppc_book3s_hv_page_fault(struct kvm_run *run, struct kvm_vcpu *vcpu,
487                                 unsigned long ea, unsigned long dsisr)
488 {
489         struct kvm *kvm = vcpu->kvm;
490         unsigned long hpte[3], r;
491         unsigned long hnow_v, hnow_r;
492         __be64 *hptep;
493         unsigned long mmu_seq, psize, pte_size;
494         unsigned long gpa_base, gfn_base;
495         unsigned long gpa, gfn, hva, pfn;
496         struct kvm_memory_slot *memslot;
497         unsigned long *rmap;
498         struct revmap_entry *rev;
499         struct page *page, *pages[1];
500         long index, ret, npages;
501         bool is_ci;
502         unsigned int writing, write_ok;
503         struct vm_area_struct *vma;
504         unsigned long rcbits;
505         long mmio_update;
506
507         if (kvm_is_radix(kvm))
508                 return kvmppc_book3s_radix_page_fault(run, vcpu, ea, dsisr);
509
510         /*
511          * Real-mode code has already searched the HPT and found the
512          * entry we're interested in.  Lock the entry and check that
513          * it hasn't changed.  If it has, just return and re-execute the
514          * instruction.
515          */
516         if (ea != vcpu->arch.pgfault_addr)
517                 return RESUME_GUEST;
518
519         if (vcpu->arch.pgfault_cache) {
520                 mmio_update = atomic64_read(&kvm->arch.mmio_update);
521                 if (mmio_update == vcpu->arch.pgfault_cache->mmio_update) {
522                         r = vcpu->arch.pgfault_cache->rpte;
523                         psize = kvmppc_actual_pgsz(vcpu->arch.pgfault_hpte[0],
524                                                    r);
525                         gpa_base = r & HPTE_R_RPN & ~(psize - 1);
526                         gfn_base = gpa_base >> PAGE_SHIFT;
527                         gpa = gpa_base | (ea & (psize - 1));
528                         return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
529                                                 dsisr & DSISR_ISSTORE);
530                 }
531         }
532         index = vcpu->arch.pgfault_index;
533         hptep = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
534         rev = &kvm->arch.hpt.rev[index];
535         preempt_disable();
536         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
537                 cpu_relax();
538         hpte[0] = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
539         hpte[1] = be64_to_cpu(hptep[1]);
540         hpte[2] = r = rev->guest_rpte;
541         unlock_hpte(hptep, hpte[0]);
542         preempt_enable();
543
544         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
545                 hpte[0] = hpte_new_to_old_v(hpte[0], hpte[1]);
546                 hpte[1] = hpte_new_to_old_r(hpte[1]);
547         }
548         if (hpte[0] != vcpu->arch.pgfault_hpte[0] ||
549             hpte[1] != vcpu->arch.pgfault_hpte[1])
550                 return RESUME_GUEST;
551
552         /* Translate the logical address and get the page */
553         psize = kvmppc_actual_pgsz(hpte[0], r);
554         gpa_base = r & HPTE_R_RPN & ~(psize - 1);
555         gfn_base = gpa_base >> PAGE_SHIFT;
556         gpa = gpa_base | (ea & (psize - 1));
557         gfn = gpa >> PAGE_SHIFT;
558         memslot = gfn_to_memslot(kvm, gfn);
559
560         trace_kvm_page_fault_enter(vcpu, hpte, memslot, ea, dsisr);
561
562         /* No memslot means it's an emulated MMIO region */
563         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
564                 return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
565                                               dsisr & DSISR_ISSTORE);
566
567         /*
568          * This should never happen, because of the slot_is_aligned()
569          * check in kvmppc_do_h_enter().
570          */
571         if (gfn_base < memslot->base_gfn)
572                 return -EFAULT;
573
574         /* used to check for invalidations in progress */
575         mmu_seq = kvm->mmu_notifier_seq;
576         smp_rmb();
577
578         ret = -EFAULT;
579         is_ci = false;
580         pfn = 0;
581         page = NULL;
582         pte_size = PAGE_SIZE;
583         writing = (dsisr & DSISR_ISSTORE) != 0;
584         /* If writing != 0, then the HPTE must allow writing, if we get here */
585         write_ok = writing;
586         hva = gfn_to_hva_memslot(memslot, gfn);
587         npages = get_user_pages_fast(hva, 1, writing, pages);
588         if (npages < 1) {
589                 /* Check if it's an I/O mapping */
590                 down_read(&current->mm->mmap_sem);
591                 vma = find_vma(current->mm, hva);
592                 if (vma && vma->vm_start <= hva && hva + psize <= vma->vm_end &&
593                     (vma->vm_flags & VM_PFNMAP)) {
594                         pfn = vma->vm_pgoff +
595                                 ((hva - vma->vm_start) >> PAGE_SHIFT);
596                         pte_size = psize;
597                         is_ci = pte_ci(__pte((pgprot_val(vma->vm_page_prot))));
598                         write_ok = vma->vm_flags & VM_WRITE;
599                 }
600                 up_read(&current->mm->mmap_sem);
601                 if (!pfn)
602                         goto out_put;
603         } else {
604                 page = pages[0];
605                 pfn = page_to_pfn(page);
606                 if (PageHuge(page)) {
607                         page = compound_head(page);
608                         pte_size <<= compound_order(page);
609                 }
610                 /* if the guest wants write access, see if that is OK */
611                 if (!writing && hpte_is_writable(r)) {
612                         pte_t *ptep, pte;
613                         unsigned long flags;
614                         /*
615                          * We need to protect against page table destruction
616                          * hugepage split and collapse.
617                          */
618                         local_irq_save(flags);
619                         ptep = find_current_mm_pte(current->mm->pgd,
620                                                    hva, NULL, NULL);
621                         if (ptep) {
622                                 pte = kvmppc_read_update_linux_pte(ptep, 1);
623                                 if (__pte_write(pte))
624                                         write_ok = 1;
625                         }
626                         local_irq_restore(flags);
627                 }
628         }
629
630         if (psize > pte_size)
631                 goto out_put;
632
633         /* Check WIMG vs. the actual page we're accessing */
634         if (!hpte_cache_flags_ok(r, is_ci)) {
635                 if (is_ci)
636                         goto out_put;
637                 /*
638                  * Allow guest to map emulated device memory as
639                  * uncacheable, but actually make it cacheable.
640                  */
641                 r = (r & ~(HPTE_R_W|HPTE_R_I|HPTE_R_G)) | HPTE_R_M;
642         }
643
644         /*
645          * Set the HPTE to point to pfn.
646          * Since the pfn is at PAGE_SIZE granularity, make sure we
647          * don't mask out lower-order bits if psize < PAGE_SIZE.
648          */
649         if (psize < PAGE_SIZE)
650                 psize = PAGE_SIZE;
651         r = (r & HPTE_R_KEY_HI) | (r & ~(HPTE_R_PP0 - psize)) |
652                                         ((pfn << PAGE_SHIFT) & ~(psize - 1));
653         if (hpte_is_writable(r) && !write_ok)
654                 r = hpte_make_readonly(r);
655         ret = RESUME_GUEST;
656         preempt_disable();
657         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
658                 cpu_relax();
659         hnow_v = be64_to_cpu(hptep[0]);
660         hnow_r = be64_to_cpu(hptep[1]);
661         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
662                 hnow_v = hpte_new_to_old_v(hnow_v, hnow_r);
663                 hnow_r = hpte_new_to_old_r(hnow_r);
664         }
665
666         /*
667          * If the HPT is being resized, don't update the HPTE,
668          * instead let the guest retry after the resize operation is complete.
669          * The synchronization for mmu_ready test vs. set is provided
670          * by the HPTE lock.
671          */
672         if (!kvm->arch.mmu_ready)
673                 goto out_unlock;
674
675         if ((hnow_v & ~HPTE_V_HVLOCK) != hpte[0] || hnow_r != hpte[1] ||
676             rev->guest_rpte != hpte[2])
677                 /* HPTE has been changed under us; let the guest retry */
678                 goto out_unlock;
679         hpte[0] = (hpte[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
680
681         /* Always put the HPTE in the rmap chain for the page base address */
682         rmap = &memslot->arch.rmap[gfn_base - memslot->base_gfn];
683         lock_rmap(rmap);
684
685         /* Check if we might have been invalidated; let the guest retry if so */
686         ret = RESUME_GUEST;
687         if (mmu_notifier_retry(vcpu->kvm, mmu_seq)) {
688                 unlock_rmap(rmap);
689                 goto out_unlock;
690         }
691
692         /* Only set R/C in real HPTE if set in both *rmap and guest_rpte */
693         rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
694         r &= rcbits | ~(HPTE_R_R | HPTE_R_C);
695
696         if (be64_to_cpu(hptep[0]) & HPTE_V_VALID) {
697                 /* HPTE was previously valid, so we need to invalidate it */
698                 unlock_rmap(rmap);
699                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
700                 kvmppc_invalidate_hpte(kvm, hptep, index);
701                 /* don't lose previous R and C bits */
702                 r |= be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
703         } else {
704                 kvmppc_add_revmap_chain(kvm, rev, rmap, index, 0);
705         }
706
707         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
708                 r = hpte_old_to_new_r(hpte[0], r);
709                 hpte[0] = hpte_old_to_new_v(hpte[0]);
710         }
711         hptep[1] = cpu_to_be64(r);
712         eieio();
713         __unlock_hpte(hptep, hpte[0]);
714         asm volatile("ptesync" : : : "memory");
715         preempt_enable();
716         if (page && hpte_is_writable(r))
717                 SetPageDirty(page);
718
719  out_put:
720         trace_kvm_page_fault_exit(vcpu, hpte, ret);
721
722         if (page) {
723                 /*
724                  * We drop pages[0] here, not page because page might
725                  * have been set to the head page of a compound, but
726                  * we have to drop the reference on the correct tail
727                  * page to match the get inside gup()
728                  */
729                 put_page(pages[0]);
730         }
731         return ret;
732
733  out_unlock:
734         __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
735         preempt_enable();
736         goto out_put;
737 }
738
739 void kvmppc_rmap_reset(struct kvm *kvm)
740 {
741         struct kvm_memslots *slots;
742         struct kvm_memory_slot *memslot;
743         int srcu_idx;
744
745         srcu_idx = srcu_read_lock(&kvm->srcu);
746         slots = kvm_memslots(kvm);
747         kvm_for_each_memslot(memslot, slots) {
748                 /*
749                  * This assumes it is acceptable to lose reference and
750                  * change bits across a reset.
751                  */
752                 memset(memslot->arch.rmap, 0,
753                        memslot->npages * sizeof(*memslot->arch.rmap));
754         }
755         srcu_read_unlock(&kvm->srcu, srcu_idx);
756 }
757
758 typedef int (*hva_handler_fn)(struct kvm *kvm, struct kvm_memory_slot *memslot,
759                               unsigned long gfn);
760
761 static int kvm_handle_hva_range(struct kvm *kvm,
762                                 unsigned long start,
763                                 unsigned long end,
764                                 hva_handler_fn handler)
765 {
766         int ret;
767         int retval = 0;
768         struct kvm_memslots *slots;
769         struct kvm_memory_slot *memslot;
770
771         slots = kvm_memslots(kvm);
772         kvm_for_each_memslot(memslot, slots) {
773                 unsigned long hva_start, hva_end;
774                 gfn_t gfn, gfn_end;
775
776                 hva_start = max(start, memslot->userspace_addr);
777                 hva_end = min(end, memslot->userspace_addr +
778                                         (memslot->npages << PAGE_SHIFT));
779                 if (hva_start >= hva_end)
780                         continue;
781                 /*
782                  * {gfn(page) | page intersects with [hva_start, hva_end)} =
783                  * {gfn, gfn+1, ..., gfn_end-1}.
784                  */
785                 gfn = hva_to_gfn_memslot(hva_start, memslot);
786                 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
787
788                 for (; gfn < gfn_end; ++gfn) {
789                         ret = handler(kvm, memslot, gfn);
790                         retval |= ret;
791                 }
792         }
793
794         return retval;
795 }
796
797 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
798                           hva_handler_fn handler)
799 {
800         return kvm_handle_hva_range(kvm, hva, hva + 1, handler);
801 }
802
803 /* Must be called with both HPTE and rmap locked */
804 static void kvmppc_unmap_hpte(struct kvm *kvm, unsigned long i,
805                               struct kvm_memory_slot *memslot,
806                               unsigned long *rmapp, unsigned long gfn)
807 {
808         __be64 *hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
809         struct revmap_entry *rev = kvm->arch.hpt.rev;
810         unsigned long j, h;
811         unsigned long ptel, psize, rcbits;
812
813         j = rev[i].forw;
814         if (j == i) {
815                 /* chain is now empty */
816                 *rmapp &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
817         } else {
818                 /* remove i from chain */
819                 h = rev[i].back;
820                 rev[h].forw = j;
821                 rev[j].back = h;
822                 rev[i].forw = rev[i].back = i;
823                 *rmapp = (*rmapp & ~KVMPPC_RMAP_INDEX) | j;
824         }
825
826         /* Now check and modify the HPTE */
827         ptel = rev[i].guest_rpte;
828         psize = kvmppc_actual_pgsz(be64_to_cpu(hptep[0]), ptel);
829         if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
830             hpte_rpn(ptel, psize) == gfn) {
831                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
832                 kvmppc_invalidate_hpte(kvm, hptep, i);
833                 hptep[1] &= ~cpu_to_be64(HPTE_R_KEY_HI | HPTE_R_KEY_LO);
834                 /* Harvest R and C */
835                 rcbits = be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
836                 *rmapp |= rcbits << KVMPPC_RMAP_RC_SHIFT;
837                 if ((rcbits & HPTE_R_C) && memslot->dirty_bitmap)
838                         kvmppc_update_dirty_map(memslot, gfn, psize);
839                 if (rcbits & ~rev[i].guest_rpte) {
840                         rev[i].guest_rpte = ptel | rcbits;
841                         note_hpte_modification(kvm, &rev[i]);
842                 }
843         }
844 }
845
846 static int kvm_unmap_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
847                            unsigned long gfn)
848 {
849         unsigned long i;
850         __be64 *hptep;
851         unsigned long *rmapp;
852
853         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
854         for (;;) {
855                 lock_rmap(rmapp);
856                 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
857                         unlock_rmap(rmapp);
858                         break;
859                 }
860
861                 /*
862                  * To avoid an ABBA deadlock with the HPTE lock bit,
863                  * we can't spin on the HPTE lock while holding the
864                  * rmap chain lock.
865                  */
866                 i = *rmapp & KVMPPC_RMAP_INDEX;
867                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
868                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
869                         /* unlock rmap before spinning on the HPTE lock */
870                         unlock_rmap(rmapp);
871                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
872                                 cpu_relax();
873                         continue;
874                 }
875
876                 kvmppc_unmap_hpte(kvm, i, memslot, rmapp, gfn);
877                 unlock_rmap(rmapp);
878                 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
879         }
880         return 0;
881 }
882
883 int kvm_unmap_hva_range_hv(struct kvm *kvm, unsigned long start, unsigned long end)
884 {
885         hva_handler_fn handler;
886
887         handler = kvm_is_radix(kvm) ? kvm_unmap_radix : kvm_unmap_rmapp;
888         kvm_handle_hva_range(kvm, start, end, handler);
889         return 0;
890 }
891
892 void kvmppc_core_flush_memslot_hv(struct kvm *kvm,
893                                   struct kvm_memory_slot *memslot)
894 {
895         unsigned long gfn;
896         unsigned long n;
897         unsigned long *rmapp;
898
899         gfn = memslot->base_gfn;
900         rmapp = memslot->arch.rmap;
901         for (n = memslot->npages; n; --n, ++gfn) {
902                 if (kvm_is_radix(kvm)) {
903                         kvm_unmap_radix(kvm, memslot, gfn);
904                         continue;
905                 }
906                 /*
907                  * Testing the present bit without locking is OK because
908                  * the memslot has been marked invalid already, and hence
909                  * no new HPTEs referencing this page can be created,
910                  * thus the present bit can't go from 0 to 1.
911                  */
912                 if (*rmapp & KVMPPC_RMAP_PRESENT)
913                         kvm_unmap_rmapp(kvm, memslot, gfn);
914                 ++rmapp;
915         }
916 }
917
918 static int kvm_age_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
919                          unsigned long gfn)
920 {
921         struct revmap_entry *rev = kvm->arch.hpt.rev;
922         unsigned long head, i, j;
923         __be64 *hptep;
924         int ret = 0;
925         unsigned long *rmapp;
926
927         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
928  retry:
929         lock_rmap(rmapp);
930         if (*rmapp & KVMPPC_RMAP_REFERENCED) {
931                 *rmapp &= ~KVMPPC_RMAP_REFERENCED;
932                 ret = 1;
933         }
934         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
935                 unlock_rmap(rmapp);
936                 return ret;
937         }
938
939         i = head = *rmapp & KVMPPC_RMAP_INDEX;
940         do {
941                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
942                 j = rev[i].forw;
943
944                 /* If this HPTE isn't referenced, ignore it */
945                 if (!(be64_to_cpu(hptep[1]) & HPTE_R_R))
946                         continue;
947
948                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
949                         /* unlock rmap before spinning on the HPTE lock */
950                         unlock_rmap(rmapp);
951                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
952                                 cpu_relax();
953                         goto retry;
954                 }
955
956                 /* Now check and modify the HPTE */
957                 if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
958                     (be64_to_cpu(hptep[1]) & HPTE_R_R)) {
959                         kvmppc_clear_ref_hpte(kvm, hptep, i);
960                         if (!(rev[i].guest_rpte & HPTE_R_R)) {
961                                 rev[i].guest_rpte |= HPTE_R_R;
962                                 note_hpte_modification(kvm, &rev[i]);
963                         }
964                         ret = 1;
965                 }
966                 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
967         } while ((i = j) != head);
968
969         unlock_rmap(rmapp);
970         return ret;
971 }
972
973 int kvm_age_hva_hv(struct kvm *kvm, unsigned long start, unsigned long end)
974 {
975         hva_handler_fn handler;
976
977         handler = kvm_is_radix(kvm) ? kvm_age_radix : kvm_age_rmapp;
978         return kvm_handle_hva_range(kvm, start, end, handler);
979 }
980
981 static int kvm_test_age_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
982                               unsigned long gfn)
983 {
984         struct revmap_entry *rev = kvm->arch.hpt.rev;
985         unsigned long head, i, j;
986         unsigned long *hp;
987         int ret = 1;
988         unsigned long *rmapp;
989
990         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
991         if (*rmapp & KVMPPC_RMAP_REFERENCED)
992                 return 1;
993
994         lock_rmap(rmapp);
995         if (*rmapp & KVMPPC_RMAP_REFERENCED)
996                 goto out;
997
998         if (*rmapp & KVMPPC_RMAP_PRESENT) {
999                 i = head = *rmapp & KVMPPC_RMAP_INDEX;
1000                 do {
1001                         hp = (unsigned long *)(kvm->arch.hpt.virt + (i << 4));
1002                         j = rev[i].forw;
1003                         if (be64_to_cpu(hp[1]) & HPTE_R_R)
1004                                 goto out;
1005                 } while ((i = j) != head);
1006         }
1007         ret = 0;
1008
1009  out:
1010         unlock_rmap(rmapp);
1011         return ret;
1012 }
1013
1014 int kvm_test_age_hva_hv(struct kvm *kvm, unsigned long hva)
1015 {
1016         hva_handler_fn handler;
1017
1018         handler = kvm_is_radix(kvm) ? kvm_test_age_radix : kvm_test_age_rmapp;
1019         return kvm_handle_hva(kvm, hva, handler);
1020 }
1021
1022 void kvm_set_spte_hva_hv(struct kvm *kvm, unsigned long hva, pte_t pte)
1023 {
1024         hva_handler_fn handler;
1025
1026         handler = kvm_is_radix(kvm) ? kvm_unmap_radix : kvm_unmap_rmapp;
1027         kvm_handle_hva(kvm, hva, handler);
1028 }
1029
1030 static int vcpus_running(struct kvm *kvm)
1031 {
1032         return atomic_read(&kvm->arch.vcpus_running) != 0;
1033 }
1034
1035 /*
1036  * Returns the number of system pages that are dirty.
1037  * This can be more than 1 if we find a huge-page HPTE.
1038  */
1039 static int kvm_test_clear_dirty_npages(struct kvm *kvm, unsigned long *rmapp)
1040 {
1041         struct revmap_entry *rev = kvm->arch.hpt.rev;
1042         unsigned long head, i, j;
1043         unsigned long n;
1044         unsigned long v, r;
1045         __be64 *hptep;
1046         int npages_dirty = 0;
1047
1048  retry:
1049         lock_rmap(rmapp);
1050         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
1051                 unlock_rmap(rmapp);
1052                 return npages_dirty;
1053         }
1054
1055         i = head = *rmapp & KVMPPC_RMAP_INDEX;
1056         do {
1057                 unsigned long hptep1;
1058                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
1059                 j = rev[i].forw;
1060
1061                 /*
1062                  * Checking the C (changed) bit here is racy since there
1063                  * is no guarantee about when the hardware writes it back.
1064                  * If the HPTE is not writable then it is stable since the
1065                  * page can't be written to, and we would have done a tlbie
1066                  * (which forces the hardware to complete any writeback)
1067                  * when making the HPTE read-only.
1068                  * If vcpus are running then this call is racy anyway
1069                  * since the page could get dirtied subsequently, so we
1070                  * expect there to be a further call which would pick up
1071                  * any delayed C bit writeback.
1072                  * Otherwise we need to do the tlbie even if C==0 in
1073                  * order to pick up any delayed writeback of C.
1074                  */
1075                 hptep1 = be64_to_cpu(hptep[1]);
1076                 if (!(hptep1 & HPTE_R_C) &&
1077                     (!hpte_is_writable(hptep1) || vcpus_running(kvm)))
1078                         continue;
1079
1080                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
1081                         /* unlock rmap before spinning on the HPTE lock */
1082                         unlock_rmap(rmapp);
1083                         while (hptep[0] & cpu_to_be64(HPTE_V_HVLOCK))
1084                                 cpu_relax();
1085                         goto retry;
1086                 }
1087
1088                 /* Now check and modify the HPTE */
1089                 if (!(hptep[0] & cpu_to_be64(HPTE_V_VALID))) {
1090                         __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
1091                         continue;
1092                 }
1093
1094                 /* need to make it temporarily absent so C is stable */
1095                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
1096                 kvmppc_invalidate_hpte(kvm, hptep, i);
1097                 v = be64_to_cpu(hptep[0]);
1098                 r = be64_to_cpu(hptep[1]);
1099                 if (r & HPTE_R_C) {
1100                         hptep[1] = cpu_to_be64(r & ~HPTE_R_C);
1101                         if (!(rev[i].guest_rpte & HPTE_R_C)) {
1102                                 rev[i].guest_rpte |= HPTE_R_C;
1103                                 note_hpte_modification(kvm, &rev[i]);
1104                         }
1105                         n = kvmppc_actual_pgsz(v, r);
1106                         n = (n + PAGE_SIZE - 1) >> PAGE_SHIFT;
1107                         if (n > npages_dirty)
1108                                 npages_dirty = n;
1109                         eieio();
1110                 }
1111                 v &= ~HPTE_V_ABSENT;
1112                 v |= HPTE_V_VALID;
1113                 __unlock_hpte(hptep, v);
1114         } while ((i = j) != head);
1115
1116         unlock_rmap(rmapp);
1117         return npages_dirty;
1118 }
1119
1120 void kvmppc_harvest_vpa_dirty(struct kvmppc_vpa *vpa,
1121                               struct kvm_memory_slot *memslot,
1122                               unsigned long *map)
1123 {
1124         unsigned long gfn;
1125
1126         if (!vpa->dirty || !vpa->pinned_addr)
1127                 return;
1128         gfn = vpa->gpa >> PAGE_SHIFT;
1129         if (gfn < memslot->base_gfn ||
1130             gfn >= memslot->base_gfn + memslot->npages)
1131                 return;
1132
1133         vpa->dirty = false;
1134         if (map)
1135                 __set_bit_le(gfn - memslot->base_gfn, map);
1136 }
1137
1138 long kvmppc_hv_get_dirty_log_hpt(struct kvm *kvm,
1139                         struct kvm_memory_slot *memslot, unsigned long *map)
1140 {
1141         unsigned long i;
1142         unsigned long *rmapp;
1143
1144         preempt_disable();
1145         rmapp = memslot->arch.rmap;
1146         for (i = 0; i < memslot->npages; ++i) {
1147                 int npages = kvm_test_clear_dirty_npages(kvm, rmapp);
1148                 /*
1149                  * Note that if npages > 0 then i must be a multiple of npages,
1150                  * since we always put huge-page HPTEs in the rmap chain
1151                  * corresponding to their page base address.
1152                  */
1153                 if (npages)
1154                         set_dirty_bits(map, i, npages);
1155                 ++rmapp;
1156         }
1157         preempt_enable();
1158         return 0;
1159 }
1160
1161 void *kvmppc_pin_guest_page(struct kvm *kvm, unsigned long gpa,
1162                             unsigned long *nb_ret)
1163 {
1164         struct kvm_memory_slot *memslot;
1165         unsigned long gfn = gpa >> PAGE_SHIFT;
1166         struct page *page, *pages[1];
1167         int npages;
1168         unsigned long hva, offset;
1169         int srcu_idx;
1170
1171         srcu_idx = srcu_read_lock(&kvm->srcu);
1172         memslot = gfn_to_memslot(kvm, gfn);
1173         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
1174                 goto err;
1175         hva = gfn_to_hva_memslot(memslot, gfn);
1176         npages = get_user_pages_fast(hva, 1, 1, pages);
1177         if (npages < 1)
1178                 goto err;
1179         page = pages[0];
1180         srcu_read_unlock(&kvm->srcu, srcu_idx);
1181
1182         offset = gpa & (PAGE_SIZE - 1);
1183         if (nb_ret)
1184                 *nb_ret = PAGE_SIZE - offset;
1185         return page_address(page) + offset;
1186
1187  err:
1188         srcu_read_unlock(&kvm->srcu, srcu_idx);
1189         return NULL;
1190 }
1191
1192 void kvmppc_unpin_guest_page(struct kvm *kvm, void *va, unsigned long gpa,
1193                              bool dirty)
1194 {
1195         struct page *page = virt_to_page(va);
1196         struct kvm_memory_slot *memslot;
1197         unsigned long gfn;
1198         int srcu_idx;
1199
1200         put_page(page);
1201
1202         if (!dirty)
1203                 return;
1204
1205         /* We need to mark this page dirty in the memslot dirty_bitmap, if any */
1206         gfn = gpa >> PAGE_SHIFT;
1207         srcu_idx = srcu_read_lock(&kvm->srcu);
1208         memslot = gfn_to_memslot(kvm, gfn);
1209         if (memslot && memslot->dirty_bitmap)
1210                 set_bit_le(gfn - memslot->base_gfn, memslot->dirty_bitmap);
1211         srcu_read_unlock(&kvm->srcu, srcu_idx);
1212 }
1213
1214 /*
1215  * HPT resizing
1216  */
1217 static int resize_hpt_allocate(struct kvm_resize_hpt *resize)
1218 {
1219         int rc;
1220
1221         rc = kvmppc_allocate_hpt(&resize->hpt, resize->order);
1222         if (rc < 0)
1223                 return rc;
1224
1225         resize_hpt_debug(resize, "resize_hpt_allocate(): HPT @ 0x%lx\n",
1226                          resize->hpt.virt);
1227
1228         return 0;
1229 }
1230
1231 static unsigned long resize_hpt_rehash_hpte(struct kvm_resize_hpt *resize,
1232                                             unsigned long idx)
1233 {
1234         struct kvm *kvm = resize->kvm;
1235         struct kvm_hpt_info *old = &kvm->arch.hpt;
1236         struct kvm_hpt_info *new = &resize->hpt;
1237         unsigned long old_hash_mask = (1ULL << (old->order - 7)) - 1;
1238         unsigned long new_hash_mask = (1ULL << (new->order - 7)) - 1;
1239         __be64 *hptep, *new_hptep;
1240         unsigned long vpte, rpte, guest_rpte;
1241         int ret;
1242         struct revmap_entry *rev;
1243         unsigned long apsize, avpn, pteg, hash;
1244         unsigned long new_idx, new_pteg, replace_vpte;
1245         int pshift;
1246
1247         hptep = (__be64 *)(old->virt + (idx << 4));
1248
1249         /* Guest is stopped, so new HPTEs can't be added or faulted
1250          * in, only unmapped or altered by host actions.  So, it's
1251          * safe to check this before we take the HPTE lock */
1252         vpte = be64_to_cpu(hptep[0]);
1253         if (!(vpte & HPTE_V_VALID) && !(vpte & HPTE_V_ABSENT))
1254                 return 0; /* nothing to do */
1255
1256         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
1257                 cpu_relax();
1258
1259         vpte = be64_to_cpu(hptep[0]);
1260
1261         ret = 0;
1262         if (!(vpte & HPTE_V_VALID) && !(vpte & HPTE_V_ABSENT))
1263                 /* Nothing to do */
1264                 goto out;
1265
1266         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1267                 rpte = be64_to_cpu(hptep[1]);
1268                 vpte = hpte_new_to_old_v(vpte, rpte);
1269         }
1270
1271         /* Unmap */
1272         rev = &old->rev[idx];
1273         guest_rpte = rev->guest_rpte;
1274
1275         ret = -EIO;
1276         apsize = kvmppc_actual_pgsz(vpte, guest_rpte);
1277         if (!apsize)
1278                 goto out;
1279
1280         if (vpte & HPTE_V_VALID) {
1281                 unsigned long gfn = hpte_rpn(guest_rpte, apsize);
1282                 int srcu_idx = srcu_read_lock(&kvm->srcu);
1283                 struct kvm_memory_slot *memslot =
1284                         __gfn_to_memslot(kvm_memslots(kvm), gfn);
1285
1286                 if (memslot) {
1287                         unsigned long *rmapp;
1288                         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1289
1290                         lock_rmap(rmapp);
1291                         kvmppc_unmap_hpte(kvm, idx, memslot, rmapp, gfn);
1292                         unlock_rmap(rmapp);
1293                 }
1294
1295                 srcu_read_unlock(&kvm->srcu, srcu_idx);
1296         }
1297
1298         /* Reload PTE after unmap */
1299         vpte = be64_to_cpu(hptep[0]);
1300         BUG_ON(vpte & HPTE_V_VALID);
1301         BUG_ON(!(vpte & HPTE_V_ABSENT));
1302
1303         ret = 0;
1304         if (!(vpte & HPTE_V_BOLTED))
1305                 goto out;
1306
1307         rpte = be64_to_cpu(hptep[1]);
1308
1309         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1310                 vpte = hpte_new_to_old_v(vpte, rpte);
1311                 rpte = hpte_new_to_old_r(rpte);
1312         }
1313
1314         pshift = kvmppc_hpte_base_page_shift(vpte, rpte);
1315         avpn = HPTE_V_AVPN_VAL(vpte) & ~(((1ul << pshift) - 1) >> 23);
1316         pteg = idx / HPTES_PER_GROUP;
1317         if (vpte & HPTE_V_SECONDARY)
1318                 pteg = ~pteg;
1319
1320         if (!(vpte & HPTE_V_1TB_SEG)) {
1321                 unsigned long offset, vsid;
1322
1323                 /* We only have 28 - 23 bits of offset in avpn */
1324                 offset = (avpn & 0x1f) << 23;
1325                 vsid = avpn >> 5;
1326                 /* We can find more bits from the pteg value */
1327                 if (pshift < 23)
1328                         offset |= ((vsid ^ pteg) & old_hash_mask) << pshift;
1329
1330                 hash = vsid ^ (offset >> pshift);
1331         } else {
1332                 unsigned long offset, vsid;
1333
1334                 /* We only have 40 - 23 bits of seg_off in avpn */
1335                 offset = (avpn & 0x1ffff) << 23;
1336                 vsid = avpn >> 17;
1337                 if (pshift < 23)
1338                         offset |= ((vsid ^ (vsid << 25) ^ pteg) & old_hash_mask) << pshift;
1339
1340                 hash = vsid ^ (vsid << 25) ^ (offset >> pshift);
1341         }
1342
1343         new_pteg = hash & new_hash_mask;
1344         if (vpte & HPTE_V_SECONDARY)
1345                 new_pteg = ~hash & new_hash_mask;
1346
1347         new_idx = new_pteg * HPTES_PER_GROUP + (idx % HPTES_PER_GROUP);
1348         new_hptep = (__be64 *)(new->virt + (new_idx << 4));
1349
1350         replace_vpte = be64_to_cpu(new_hptep[0]);
1351         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1352                 unsigned long replace_rpte = be64_to_cpu(new_hptep[1]);
1353                 replace_vpte = hpte_new_to_old_v(replace_vpte, replace_rpte);
1354         }
1355
1356         if (replace_vpte & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1357                 BUG_ON(new->order >= old->order);
1358
1359                 if (replace_vpte & HPTE_V_BOLTED) {
1360                         if (vpte & HPTE_V_BOLTED)
1361                                 /* Bolted collision, nothing we can do */
1362                                 ret = -ENOSPC;
1363                         /* Discard the new HPTE */
1364                         goto out;
1365                 }
1366
1367                 /* Discard the previous HPTE */
1368         }
1369
1370         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1371                 rpte = hpte_old_to_new_r(vpte, rpte);
1372                 vpte = hpte_old_to_new_v(vpte);
1373         }
1374
1375         new_hptep[1] = cpu_to_be64(rpte);
1376         new->rev[new_idx].guest_rpte = guest_rpte;
1377         /* No need for a barrier, since new HPT isn't active */
1378         new_hptep[0] = cpu_to_be64(vpte);
1379         unlock_hpte(new_hptep, vpte);
1380
1381 out:
1382         unlock_hpte(hptep, vpte);
1383         return ret;
1384 }
1385
1386 static int resize_hpt_rehash(struct kvm_resize_hpt *resize)
1387 {
1388         struct kvm *kvm = resize->kvm;
1389         unsigned  long i;
1390         int rc;
1391
1392         for (i = 0; i < kvmppc_hpt_npte(&kvm->arch.hpt); i++) {
1393                 rc = resize_hpt_rehash_hpte(resize, i);
1394                 if (rc != 0)
1395                         return rc;
1396         }
1397
1398         return 0;
1399 }
1400
1401 static void resize_hpt_pivot(struct kvm_resize_hpt *resize)
1402 {
1403         struct kvm *kvm = resize->kvm;
1404         struct kvm_hpt_info hpt_tmp;
1405
1406         /* Exchange the pending tables in the resize structure with
1407          * the active tables */
1408
1409         resize_hpt_debug(resize, "resize_hpt_pivot()\n");
1410
1411         spin_lock(&kvm->mmu_lock);
1412         asm volatile("ptesync" : : : "memory");
1413
1414         hpt_tmp = kvm->arch.hpt;
1415         kvmppc_set_hpt(kvm, &resize->hpt);
1416         resize->hpt = hpt_tmp;
1417
1418         spin_unlock(&kvm->mmu_lock);
1419
1420         synchronize_srcu_expedited(&kvm->srcu);
1421
1422         if (cpu_has_feature(CPU_FTR_ARCH_300))
1423                 kvmppc_setup_partition_table(kvm);
1424
1425         resize_hpt_debug(resize, "resize_hpt_pivot() done\n");
1426 }
1427
1428 static void resize_hpt_release(struct kvm *kvm, struct kvm_resize_hpt *resize)
1429 {
1430         if (WARN_ON(!mutex_is_locked(&kvm->lock)))
1431                 return;
1432
1433         if (!resize)
1434                 return;
1435
1436         if (resize->error != -EBUSY) {
1437                 if (resize->hpt.virt)
1438                         kvmppc_free_hpt(&resize->hpt);
1439                 kfree(resize);
1440         }
1441
1442         if (kvm->arch.resize_hpt == resize)
1443                 kvm->arch.resize_hpt = NULL;
1444 }
1445
1446 static void resize_hpt_prepare_work(struct work_struct *work)
1447 {
1448         struct kvm_resize_hpt *resize = container_of(work,
1449                                                      struct kvm_resize_hpt,
1450                                                      work);
1451         struct kvm *kvm = resize->kvm;
1452         int err = 0;
1453
1454         if (WARN_ON(resize->error != -EBUSY))
1455                 return;
1456
1457         mutex_lock(&kvm->lock);
1458
1459         /* Request is still current? */
1460         if (kvm->arch.resize_hpt == resize) {
1461                 /* We may request large allocations here:
1462                  * do not sleep with kvm->lock held for a while.
1463                  */
1464                 mutex_unlock(&kvm->lock);
1465
1466                 resize_hpt_debug(resize, "resize_hpt_prepare_work(): order = %d\n",
1467                                  resize->order);
1468
1469                 err = resize_hpt_allocate(resize);
1470
1471                 /* We have strict assumption about -EBUSY
1472                  * when preparing for HPT resize.
1473                  */
1474                 if (WARN_ON(err == -EBUSY))
1475                         err = -EINPROGRESS;
1476
1477                 mutex_lock(&kvm->lock);
1478                 /* It is possible that kvm->arch.resize_hpt != resize
1479                  * after we grab kvm->lock again.
1480                  */
1481         }
1482
1483         resize->error = err;
1484
1485         if (kvm->arch.resize_hpt != resize)
1486                 resize_hpt_release(kvm, resize);
1487
1488         mutex_unlock(&kvm->lock);
1489 }
1490
1491 long kvm_vm_ioctl_resize_hpt_prepare(struct kvm *kvm,
1492                                      struct kvm_ppc_resize_hpt *rhpt)
1493 {
1494         unsigned long flags = rhpt->flags;
1495         unsigned long shift = rhpt->shift;
1496         struct kvm_resize_hpt *resize;
1497         int ret;
1498
1499         if (flags != 0 || kvm_is_radix(kvm))
1500                 return -EINVAL;
1501
1502         if (shift && ((shift < 18) || (shift > 46)))
1503                 return -EINVAL;
1504
1505         mutex_lock(&kvm->lock);
1506
1507         resize = kvm->arch.resize_hpt;
1508
1509         if (resize) {
1510                 if (resize->order == shift) {
1511                         /* Suitable resize in progress? */
1512                         ret = resize->error;
1513                         if (ret == -EBUSY)
1514                                 ret = 100; /* estimated time in ms */
1515                         else if (ret)
1516                                 resize_hpt_release(kvm, resize);
1517
1518                         goto out;
1519                 }
1520
1521                 /* not suitable, cancel it */
1522                 resize_hpt_release(kvm, resize);
1523         }
1524
1525         ret = 0;
1526         if (!shift)
1527                 goto out; /* nothing to do */
1528
1529         /* start new resize */
1530
1531         resize = kzalloc(sizeof(*resize), GFP_KERNEL);
1532         if (!resize) {
1533                 ret = -ENOMEM;
1534                 goto out;
1535         }
1536
1537         resize->error = -EBUSY;
1538         resize->order = shift;
1539         resize->kvm = kvm;
1540         INIT_WORK(&resize->work, resize_hpt_prepare_work);
1541         kvm->arch.resize_hpt = resize;
1542
1543         schedule_work(&resize->work);
1544
1545         ret = 100; /* estimated time in ms */
1546
1547 out:
1548         mutex_unlock(&kvm->lock);
1549         return ret;
1550 }
1551
1552 static void resize_hpt_boot_vcpu(void *opaque)
1553 {
1554         /* Nothing to do, just force a KVM exit */
1555 }
1556
1557 long kvm_vm_ioctl_resize_hpt_commit(struct kvm *kvm,
1558                                     struct kvm_ppc_resize_hpt *rhpt)
1559 {
1560         unsigned long flags = rhpt->flags;
1561         unsigned long shift = rhpt->shift;
1562         struct kvm_resize_hpt *resize;
1563         long ret;
1564
1565         if (flags != 0 || kvm_is_radix(kvm))
1566                 return -EINVAL;
1567
1568         if (shift && ((shift < 18) || (shift > 46)))
1569                 return -EINVAL;
1570
1571         mutex_lock(&kvm->lock);
1572
1573         resize = kvm->arch.resize_hpt;
1574
1575         /* This shouldn't be possible */
1576         ret = -EIO;
1577         if (WARN_ON(!kvm->arch.mmu_ready))
1578                 goto out_no_hpt;
1579
1580         /* Stop VCPUs from running while we mess with the HPT */
1581         kvm->arch.mmu_ready = 0;
1582         smp_mb();
1583
1584         /* Boot all CPUs out of the guest so they re-read
1585          * mmu_ready */
1586         on_each_cpu(resize_hpt_boot_vcpu, NULL, 1);
1587
1588         ret = -ENXIO;
1589         if (!resize || (resize->order != shift))
1590                 goto out;
1591
1592         ret = resize->error;
1593         if (ret)
1594                 goto out;
1595
1596         ret = resize_hpt_rehash(resize);
1597         if (ret)
1598                 goto out;
1599
1600         resize_hpt_pivot(resize);
1601
1602 out:
1603         /* Let VCPUs run again */
1604         kvm->arch.mmu_ready = 1;
1605         smp_mb();
1606 out_no_hpt:
1607         resize_hpt_release(kvm, resize);
1608         mutex_unlock(&kvm->lock);
1609         return ret;
1610 }
1611
1612 /*
1613  * Functions for reading and writing the hash table via reads and
1614  * writes on a file descriptor.
1615  *
1616  * Reads return the guest view of the hash table, which has to be
1617  * pieced together from the real hash table and the guest_rpte
1618  * values in the revmap array.
1619  *
1620  * On writes, each HPTE written is considered in turn, and if it
1621  * is valid, it is written to the HPT as if an H_ENTER with the
1622  * exact flag set was done.  When the invalid count is non-zero
1623  * in the header written to the stream, the kernel will make
1624  * sure that that many HPTEs are invalid, and invalidate them
1625  * if not.
1626  */
1627
1628 struct kvm_htab_ctx {
1629         unsigned long   index;
1630         unsigned long   flags;
1631         struct kvm      *kvm;
1632         int             first_pass;
1633 };
1634
1635 #define HPTE_SIZE       (2 * sizeof(unsigned long))
1636
1637 /*
1638  * Returns 1 if this HPT entry has been modified or has pending
1639  * R/C bit changes.
1640  */
1641 static int hpte_dirty(struct revmap_entry *revp, __be64 *hptp)
1642 {
1643         unsigned long rcbits_unset;
1644
1645         if (revp->guest_rpte & HPTE_GR_MODIFIED)
1646                 return 1;
1647
1648         /* Also need to consider changes in reference and changed bits */
1649         rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1650         if ((be64_to_cpu(hptp[0]) & HPTE_V_VALID) &&
1651             (be64_to_cpu(hptp[1]) & rcbits_unset))
1652                 return 1;
1653
1654         return 0;
1655 }
1656
1657 static long record_hpte(unsigned long flags, __be64 *hptp,
1658                         unsigned long *hpte, struct revmap_entry *revp,
1659                         int want_valid, int first_pass)
1660 {
1661         unsigned long v, r, hr;
1662         unsigned long rcbits_unset;
1663         int ok = 1;
1664         int valid, dirty;
1665
1666         /* Unmodified entries are uninteresting except on the first pass */
1667         dirty = hpte_dirty(revp, hptp);
1668         if (!first_pass && !dirty)
1669                 return 0;
1670
1671         valid = 0;
1672         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1673                 valid = 1;
1674                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) &&
1675                     !(be64_to_cpu(hptp[0]) & HPTE_V_BOLTED))
1676                         valid = 0;
1677         }
1678         if (valid != want_valid)
1679                 return 0;
1680
1681         v = r = 0;
1682         if (valid || dirty) {
1683                 /* lock the HPTE so it's stable and read it */
1684                 preempt_disable();
1685                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
1686                         cpu_relax();
1687                 v = be64_to_cpu(hptp[0]);
1688                 hr = be64_to_cpu(hptp[1]);
1689                 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1690                         v = hpte_new_to_old_v(v, hr);
1691                         hr = hpte_new_to_old_r(hr);
1692                 }
1693
1694                 /* re-evaluate valid and dirty from synchronized HPTE value */
1695                 valid = !!(v & HPTE_V_VALID);
1696                 dirty = !!(revp->guest_rpte & HPTE_GR_MODIFIED);
1697
1698                 /* Harvest R and C into guest view if necessary */
1699                 rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1700                 if (valid && (rcbits_unset & hr)) {
1701                         revp->guest_rpte |= (hr &
1702                                 (HPTE_R_R | HPTE_R_C)) | HPTE_GR_MODIFIED;
1703                         dirty = 1;
1704                 }
1705
1706                 if (v & HPTE_V_ABSENT) {
1707                         v &= ~HPTE_V_ABSENT;
1708                         v |= HPTE_V_VALID;
1709                         valid = 1;
1710                 }
1711                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) && !(v & HPTE_V_BOLTED))
1712                         valid = 0;
1713
1714                 r = revp->guest_rpte;
1715                 /* only clear modified if this is the right sort of entry */
1716                 if (valid == want_valid && dirty) {
1717                         r &= ~HPTE_GR_MODIFIED;
1718                         revp->guest_rpte = r;
1719                 }
1720                 unlock_hpte(hptp, be64_to_cpu(hptp[0]));
1721                 preempt_enable();
1722                 if (!(valid == want_valid && (first_pass || dirty)))
1723                         ok = 0;
1724         }
1725         hpte[0] = cpu_to_be64(v);
1726         hpte[1] = cpu_to_be64(r);
1727         return ok;
1728 }
1729
1730 static ssize_t kvm_htab_read(struct file *file, char __user *buf,
1731                              size_t count, loff_t *ppos)
1732 {
1733         struct kvm_htab_ctx *ctx = file->private_data;
1734         struct kvm *kvm = ctx->kvm;
1735         struct kvm_get_htab_header hdr;
1736         __be64 *hptp;
1737         struct revmap_entry *revp;
1738         unsigned long i, nb, nw;
1739         unsigned long __user *lbuf;
1740         struct kvm_get_htab_header __user *hptr;
1741         unsigned long flags;
1742         int first_pass;
1743         unsigned long hpte[2];
1744
1745         if (!access_ok(VERIFY_WRITE, buf, count))
1746                 return -EFAULT;
1747         if (kvm_is_radix(kvm))
1748                 return 0;
1749
1750         first_pass = ctx->first_pass;
1751         flags = ctx->flags;
1752
1753         i = ctx->index;
1754         hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
1755         revp = kvm->arch.hpt.rev + i;
1756         lbuf = (unsigned long __user *)buf;
1757
1758         nb = 0;
1759         while (nb + sizeof(hdr) + HPTE_SIZE < count) {
1760                 /* Initialize header */
1761                 hptr = (struct kvm_get_htab_header __user *)buf;
1762                 hdr.n_valid = 0;
1763                 hdr.n_invalid = 0;
1764                 nw = nb;
1765                 nb += sizeof(hdr);
1766                 lbuf = (unsigned long __user *)(buf + sizeof(hdr));
1767
1768                 /* Skip uninteresting entries, i.e. clean on not-first pass */
1769                 if (!first_pass) {
1770                         while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1771                                !hpte_dirty(revp, hptp)) {
1772                                 ++i;
1773                                 hptp += 2;
1774                                 ++revp;
1775                         }
1776                 }
1777                 hdr.index = i;
1778
1779                 /* Grab a series of valid entries */
1780                 while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1781                        hdr.n_valid < 0xffff &&
1782                        nb + HPTE_SIZE < count &&
1783                        record_hpte(flags, hptp, hpte, revp, 1, first_pass)) {
1784                         /* valid entry, write it out */
1785                         ++hdr.n_valid;
1786                         if (__put_user(hpte[0], lbuf) ||
1787                             __put_user(hpte[1], lbuf + 1))
1788                                 return -EFAULT;
1789                         nb += HPTE_SIZE;
1790                         lbuf += 2;
1791                         ++i;
1792                         hptp += 2;
1793                         ++revp;
1794                 }
1795                 /* Now skip invalid entries while we can */
1796                 while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1797                        hdr.n_invalid < 0xffff &&
1798                        record_hpte(flags, hptp, hpte, revp, 0, first_pass)) {
1799                         /* found an invalid entry */
1800                         ++hdr.n_invalid;
1801                         ++i;
1802                         hptp += 2;
1803                         ++revp;
1804                 }
1805
1806                 if (hdr.n_valid || hdr.n_invalid) {
1807                         /* write back the header */
1808                         if (__copy_to_user(hptr, &hdr, sizeof(hdr)))
1809                                 return -EFAULT;
1810                         nw = nb;
1811                         buf = (char __user *)lbuf;
1812                 } else {
1813                         nb = nw;
1814                 }
1815
1816                 /* Check if we've wrapped around the hash table */
1817                 if (i >= kvmppc_hpt_npte(&kvm->arch.hpt)) {
1818                         i = 0;
1819                         ctx->first_pass = 0;
1820                         break;
1821                 }
1822         }
1823
1824         ctx->index = i;
1825
1826         return nb;
1827 }
1828
1829 static ssize_t kvm_htab_write(struct file *file, const char __user *buf,
1830                               size_t count, loff_t *ppos)
1831 {
1832         struct kvm_htab_ctx *ctx = file->private_data;
1833         struct kvm *kvm = ctx->kvm;
1834         struct kvm_get_htab_header hdr;
1835         unsigned long i, j;
1836         unsigned long v, r;
1837         unsigned long __user *lbuf;
1838         __be64 *hptp;
1839         unsigned long tmp[2];
1840         ssize_t nb;
1841         long int err, ret;
1842         int mmu_ready;
1843         int pshift;
1844
1845         if (!access_ok(VERIFY_READ, buf, count))
1846                 return -EFAULT;
1847         if (kvm_is_radix(kvm))
1848                 return -EINVAL;
1849
1850         /* lock out vcpus from running while we're doing this */
1851         mutex_lock(&kvm->lock);
1852         mmu_ready = kvm->arch.mmu_ready;
1853         if (mmu_ready) {
1854                 kvm->arch.mmu_ready = 0;        /* temporarily */
1855                 /* order mmu_ready vs. vcpus_running */
1856                 smp_mb();
1857                 if (atomic_read(&kvm->arch.vcpus_running)) {
1858                         kvm->arch.mmu_ready = 1;
1859                         mutex_unlock(&kvm->lock);
1860                         return -EBUSY;
1861                 }
1862         }
1863
1864         err = 0;
1865         for (nb = 0; nb + sizeof(hdr) <= count; ) {
1866                 err = -EFAULT;
1867                 if (__copy_from_user(&hdr, buf, sizeof(hdr)))
1868                         break;
1869
1870                 err = 0;
1871                 if (nb + hdr.n_valid * HPTE_SIZE > count)
1872                         break;
1873
1874                 nb += sizeof(hdr);
1875                 buf += sizeof(hdr);
1876
1877                 err = -EINVAL;
1878                 i = hdr.index;
1879                 if (i >= kvmppc_hpt_npte(&kvm->arch.hpt) ||
1880                     i + hdr.n_valid + hdr.n_invalid > kvmppc_hpt_npte(&kvm->arch.hpt))
1881                         break;
1882
1883                 hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
1884                 lbuf = (unsigned long __user *)buf;
1885                 for (j = 0; j < hdr.n_valid; ++j) {
1886                         __be64 hpte_v;
1887                         __be64 hpte_r;
1888
1889                         err = -EFAULT;
1890                         if (__get_user(hpte_v, lbuf) ||
1891                             __get_user(hpte_r, lbuf + 1))
1892                                 goto out;
1893                         v = be64_to_cpu(hpte_v);
1894                         r = be64_to_cpu(hpte_r);
1895                         err = -EINVAL;
1896                         if (!(v & HPTE_V_VALID))
1897                                 goto out;
1898                         pshift = kvmppc_hpte_base_page_shift(v, r);
1899                         if (pshift <= 0)
1900                                 goto out;
1901                         lbuf += 2;
1902                         nb += HPTE_SIZE;
1903
1904                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1905                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1906                         err = -EIO;
1907                         ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, i, v, r,
1908                                                          tmp);
1909                         if (ret != H_SUCCESS) {
1910                                 pr_err("kvm_htab_write ret %ld i=%ld v=%lx "
1911                                        "r=%lx\n", ret, i, v, r);
1912                                 goto out;
1913                         }
1914                         if (!mmu_ready && is_vrma_hpte(v)) {
1915                                 unsigned long senc, lpcr;
1916
1917                                 senc = slb_pgsize_encoding(1ul << pshift);
1918                                 kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
1919                                         (VRMA_VSID << SLB_VSID_SHIFT_1T);
1920                                 if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
1921                                         lpcr = senc << (LPCR_VRMASD_SH - 4);
1922                                         kvmppc_update_lpcr(kvm, lpcr,
1923                                                            LPCR_VRMASD);
1924                                 } else {
1925                                         kvmppc_setup_partition_table(kvm);
1926                                 }
1927                                 mmu_ready = 1;
1928                         }
1929                         ++i;
1930                         hptp += 2;
1931                 }
1932
1933                 for (j = 0; j < hdr.n_invalid; ++j) {
1934                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1935                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1936                         ++i;
1937                         hptp += 2;
1938                 }
1939                 err = 0;
1940         }
1941
1942  out:
1943         /* Order HPTE updates vs. mmu_ready */
1944         smp_wmb();
1945         kvm->arch.mmu_ready = mmu_ready;
1946         mutex_unlock(&kvm->lock);
1947
1948         if (err)
1949                 return err;
1950         return nb;
1951 }
1952
1953 static int kvm_htab_release(struct inode *inode, struct file *filp)
1954 {
1955         struct kvm_htab_ctx *ctx = filp->private_data;
1956
1957         filp->private_data = NULL;
1958         if (!(ctx->flags & KVM_GET_HTAB_WRITE))
1959                 atomic_dec(&ctx->kvm->arch.hpte_mod_interest);
1960         kvm_put_kvm(ctx->kvm);
1961         kfree(ctx);
1962         return 0;
1963 }
1964
1965 static const struct file_operations kvm_htab_fops = {
1966         .read           = kvm_htab_read,
1967         .write          = kvm_htab_write,
1968         .llseek         = default_llseek,
1969         .release        = kvm_htab_release,
1970 };
1971
1972 int kvm_vm_ioctl_get_htab_fd(struct kvm *kvm, struct kvm_get_htab_fd *ghf)
1973 {
1974         int ret;
1975         struct kvm_htab_ctx *ctx;
1976         int rwflag;
1977
1978         /* reject flags we don't recognize */
1979         if (ghf->flags & ~(KVM_GET_HTAB_BOLTED_ONLY | KVM_GET_HTAB_WRITE))
1980                 return -EINVAL;
1981         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1982         if (!ctx)
1983                 return -ENOMEM;
1984         kvm_get_kvm(kvm);
1985         ctx->kvm = kvm;
1986         ctx->index = ghf->start_index;
1987         ctx->flags = ghf->flags;
1988         ctx->first_pass = 1;
1989
1990         rwflag = (ghf->flags & KVM_GET_HTAB_WRITE) ? O_WRONLY : O_RDONLY;
1991         ret = anon_inode_getfd("kvm-htab", &kvm_htab_fops, ctx, rwflag | O_CLOEXEC);
1992         if (ret < 0) {
1993                 kfree(ctx);
1994                 kvm_put_kvm(kvm);
1995                 return ret;
1996         }
1997
1998         if (rwflag == O_RDONLY) {
1999                 mutex_lock(&kvm->slots_lock);
2000                 atomic_inc(&kvm->arch.hpte_mod_interest);
2001                 /* make sure kvmppc_do_h_enter etc. see the increment */
2002                 synchronize_srcu_expedited(&kvm->srcu);
2003                 mutex_unlock(&kvm->slots_lock);
2004         }
2005
2006         return ret;
2007 }
2008
2009 struct debugfs_htab_state {
2010         struct kvm      *kvm;
2011         struct mutex    mutex;
2012         unsigned long   hpt_index;
2013         int             chars_left;
2014         int             buf_index;
2015         char            buf[64];
2016 };
2017
2018 static int debugfs_htab_open(struct inode *inode, struct file *file)
2019 {
2020         struct kvm *kvm = inode->i_private;
2021         struct debugfs_htab_state *p;
2022
2023         p = kzalloc(sizeof(*p), GFP_KERNEL);
2024         if (!p)
2025                 return -ENOMEM;
2026
2027         kvm_get_kvm(kvm);
2028         p->kvm = kvm;
2029         mutex_init(&p->mutex);
2030         file->private_data = p;
2031
2032         return nonseekable_open(inode, file);
2033 }
2034
2035 static int debugfs_htab_release(struct inode *inode, struct file *file)
2036 {
2037         struct debugfs_htab_state *p = file->private_data;
2038
2039         kvm_put_kvm(p->kvm);
2040         kfree(p);
2041         return 0;
2042 }
2043
2044 static ssize_t debugfs_htab_read(struct file *file, char __user *buf,
2045                                  size_t len, loff_t *ppos)
2046 {
2047         struct debugfs_htab_state *p = file->private_data;
2048         ssize_t ret, r;
2049         unsigned long i, n;
2050         unsigned long v, hr, gr;
2051         struct kvm *kvm;
2052         __be64 *hptp;
2053
2054         kvm = p->kvm;
2055         if (kvm_is_radix(kvm))
2056                 return 0;
2057
2058         ret = mutex_lock_interruptible(&p->mutex);
2059         if (ret)
2060                 return ret;
2061
2062         if (p->chars_left) {
2063                 n = p->chars_left;
2064                 if (n > len)
2065                         n = len;
2066                 r = copy_to_user(buf, p->buf + p->buf_index, n);
2067                 n -= r;
2068                 p->chars_left -= n;
2069                 p->buf_index += n;
2070                 buf += n;
2071                 len -= n;
2072                 ret = n;
2073                 if (r) {
2074                         if (!n)
2075                                 ret = -EFAULT;
2076                         goto out;
2077                 }
2078         }
2079
2080         i = p->hpt_index;
2081         hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
2082         for (; len != 0 && i < kvmppc_hpt_npte(&kvm->arch.hpt);
2083              ++i, hptp += 2) {
2084                 if (!(be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)))
2085                         continue;
2086
2087                 /* lock the HPTE so it's stable and read it */
2088                 preempt_disable();
2089                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
2090                         cpu_relax();
2091                 v = be64_to_cpu(hptp[0]) & ~HPTE_V_HVLOCK;
2092                 hr = be64_to_cpu(hptp[1]);
2093                 gr = kvm->arch.hpt.rev[i].guest_rpte;
2094                 unlock_hpte(hptp, v);
2095                 preempt_enable();
2096
2097                 if (!(v & (HPTE_V_VALID | HPTE_V_ABSENT)))
2098                         continue;
2099
2100                 n = scnprintf(p->buf, sizeof(p->buf),
2101                               "%6lx %.16lx %.16lx %.16lx\n",
2102                               i, v, hr, gr);
2103                 p->chars_left = n;
2104                 if (n > len)
2105                         n = len;
2106                 r = copy_to_user(buf, p->buf, n);
2107                 n -= r;
2108                 p->chars_left -= n;
2109                 p->buf_index = n;
2110                 buf += n;
2111                 len -= n;
2112                 ret += n;
2113                 if (r) {
2114                         if (!ret)
2115                                 ret = -EFAULT;
2116                         goto out;
2117                 }
2118         }
2119         p->hpt_index = i;
2120
2121  out:
2122         mutex_unlock(&p->mutex);
2123         return ret;
2124 }
2125
2126 static ssize_t debugfs_htab_write(struct file *file, const char __user *buf,
2127                            size_t len, loff_t *ppos)
2128 {
2129         return -EACCES;
2130 }
2131
2132 static const struct file_operations debugfs_htab_fops = {
2133         .owner   = THIS_MODULE,
2134         .open    = debugfs_htab_open,
2135         .release = debugfs_htab_release,
2136         .read    = debugfs_htab_read,
2137         .write   = debugfs_htab_write,
2138         .llseek  = generic_file_llseek,
2139 };
2140
2141 void kvmppc_mmu_debugfs_init(struct kvm *kvm)
2142 {
2143         kvm->arch.htab_dentry = debugfs_create_file("htab", 0400,
2144                                                     kvm->arch.debugfs_dir, kvm,
2145                                                     &debugfs_htab_fops);
2146 }
2147
2148 void kvmppc_mmu_book3s_hv_init(struct kvm_vcpu *vcpu)
2149 {
2150         struct kvmppc_mmu *mmu = &vcpu->arch.mmu;
2151
2152         vcpu->arch.slb_nr = 32;         /* POWER7/POWER8 */
2153
2154         mmu->xlate = kvmppc_mmu_book3s_64_hv_xlate;
2155         mmu->reset_msr = kvmppc_mmu_book3s_64_hv_reset_msr;
2156
2157         vcpu->arch.hflags |= BOOK3S_HFLAG_SLB;
2158 }