KVM: PPC: Book3S HV: Make HPT resizing work on POWER9
[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         int error;
69         bool prepare_done;
70
71         /* Private to the work thread, until prepare_done is true,
72          * then protected by kvm->resize_hpt_sem */
73         struct kvm_hpt_info hpt;
74 };
75
76 int kvmppc_allocate_hpt(struct kvm_hpt_info *info, u32 order)
77 {
78         unsigned long hpt = 0;
79         int cma = 0;
80         struct page *page = NULL;
81         struct revmap_entry *rev;
82         unsigned long npte;
83
84         if ((order < PPC_MIN_HPT_ORDER) || (order > PPC_MAX_HPT_ORDER))
85                 return -EINVAL;
86
87         page = kvm_alloc_hpt_cma(1ul << (order - PAGE_SHIFT));
88         if (page) {
89                 hpt = (unsigned long)pfn_to_kaddr(page_to_pfn(page));
90                 memset((void *)hpt, 0, (1ul << order));
91                 cma = 1;
92         }
93
94         if (!hpt)
95                 hpt = __get_free_pages(GFP_KERNEL|__GFP_ZERO|__GFP_RETRY_MAYFAIL
96                                        |__GFP_NOWARN, order - PAGE_SHIFT);
97
98         if (!hpt)
99                 return -ENOMEM;
100
101         /* HPTEs are 2**4 bytes long */
102         npte = 1ul << (order - 4);
103
104         /* Allocate reverse map array */
105         rev = vmalloc(sizeof(struct revmap_entry) * npte);
106         if (!rev) {
107                 if (cma)
108                         kvm_free_hpt_cma(page, 1 << (order - PAGE_SHIFT));
109                 else
110                         free_pages(hpt, order - PAGE_SHIFT);
111                 return -ENOMEM;
112         }
113
114         info->order = order;
115         info->virt = hpt;
116         info->cma = cma;
117         info->rev = rev;
118
119         return 0;
120 }
121
122 void kvmppc_set_hpt(struct kvm *kvm, struct kvm_hpt_info *info)
123 {
124         atomic64_set(&kvm->arch.mmio_update, 0);
125         kvm->arch.hpt = *info;
126         kvm->arch.sdr1 = __pa(info->virt) | (info->order - 18);
127
128         pr_debug("KVM guest htab at %lx (order %ld), LPID %x\n",
129                  info->virt, (long)info->order, kvm->arch.lpid);
130 }
131
132 long kvmppc_alloc_reset_hpt(struct kvm *kvm, int order)
133 {
134         long err = -EBUSY;
135         struct kvm_hpt_info info;
136
137         mutex_lock(&kvm->lock);
138         if (kvm->arch.mmu_ready) {
139                 kvm->arch.mmu_ready = 0;
140                 /* order mmu_ready vs. vcpus_running */
141                 smp_mb();
142                 if (atomic_read(&kvm->arch.vcpus_running)) {
143                         kvm->arch.mmu_ready = 1;
144                         goto out;
145                 }
146         }
147         if (kvm_is_radix(kvm)) {
148                 err = kvmppc_switch_mmu_to_hpt(kvm);
149                 if (err)
150                         goto out;
151         }
152
153         if (kvm->arch.hpt.order == order) {
154                 /* We already have a suitable HPT */
155
156                 /* Set the entire HPT to 0, i.e. invalid HPTEs */
157                 memset((void *)kvm->arch.hpt.virt, 0, 1ul << order);
158                 /*
159                  * Reset all the reverse-mapping chains for all memslots
160                  */
161                 kvmppc_rmap_reset(kvm);
162                 /* Ensure that each vcpu will flush its TLB on next entry. */
163                 cpumask_setall(&kvm->arch.need_tlb_flush);
164                 err = 0;
165                 goto out;
166         }
167
168         if (kvm->arch.hpt.virt) {
169                 kvmppc_free_hpt(&kvm->arch.hpt);
170                 kvmppc_rmap_reset(kvm);
171         }
172
173         err = kvmppc_allocate_hpt(&info, order);
174         if (err < 0)
175                 goto out;
176         kvmppc_set_hpt(kvm, &info);
177
178 out:
179         mutex_unlock(&kvm->lock);
180         return err;
181 }
182
183 void kvmppc_free_hpt(struct kvm_hpt_info *info)
184 {
185         vfree(info->rev);
186         info->rev = NULL;
187         if (info->cma)
188                 kvm_free_hpt_cma(virt_to_page(info->virt),
189                                  1 << (info->order - PAGE_SHIFT));
190         else if (info->virt)
191                 free_pages(info->virt, info->order - PAGE_SHIFT);
192         info->virt = 0;
193         info->order = 0;
194 }
195
196 /* Bits in first HPTE dword for pagesize 4k, 64k or 16M */
197 static inline unsigned long hpte0_pgsize_encoding(unsigned long pgsize)
198 {
199         return (pgsize > 0x1000) ? HPTE_V_LARGE : 0;
200 }
201
202 /* Bits in second HPTE dword for pagesize 4k, 64k or 16M */
203 static inline unsigned long hpte1_pgsize_encoding(unsigned long pgsize)
204 {
205         return (pgsize == 0x10000) ? 0x1000 : 0;
206 }
207
208 void kvmppc_map_vrma(struct kvm_vcpu *vcpu, struct kvm_memory_slot *memslot,
209                      unsigned long porder)
210 {
211         unsigned long i;
212         unsigned long npages;
213         unsigned long hp_v, hp_r;
214         unsigned long addr, hash;
215         unsigned long psize;
216         unsigned long hp0, hp1;
217         unsigned long idx_ret;
218         long ret;
219         struct kvm *kvm = vcpu->kvm;
220
221         psize = 1ul << porder;
222         npages = memslot->npages >> (porder - PAGE_SHIFT);
223
224         /* VRMA can't be > 1TB */
225         if (npages > 1ul << (40 - porder))
226                 npages = 1ul << (40 - porder);
227         /* Can't use more than 1 HPTE per HPTEG */
228         if (npages > kvmppc_hpt_mask(&kvm->arch.hpt) + 1)
229                 npages = kvmppc_hpt_mask(&kvm->arch.hpt) + 1;
230
231         hp0 = HPTE_V_1TB_SEG | (VRMA_VSID << (40 - 16)) |
232                 HPTE_V_BOLTED | hpte0_pgsize_encoding(psize);
233         hp1 = hpte1_pgsize_encoding(psize) |
234                 HPTE_R_R | HPTE_R_C | HPTE_R_M | PP_RWXX;
235
236         for (i = 0; i < npages; ++i) {
237                 addr = i << porder;
238                 /* can't use hpt_hash since va > 64 bits */
239                 hash = (i ^ (VRMA_VSID ^ (VRMA_VSID << 25)))
240                         & kvmppc_hpt_mask(&kvm->arch.hpt);
241                 /*
242                  * We assume that the hash table is empty and no
243                  * vcpus are using it at this stage.  Since we create
244                  * at most one HPTE per HPTEG, we just assume entry 7
245                  * is available and use it.
246                  */
247                 hash = (hash << 3) + 7;
248                 hp_v = hp0 | ((addr >> 16) & ~0x7fUL);
249                 hp_r = hp1 | addr;
250                 ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, hash, hp_v, hp_r,
251                                                  &idx_ret);
252                 if (ret != H_SUCCESS) {
253                         pr_err("KVM: map_vrma at %lx failed, ret=%ld\n",
254                                addr, ret);
255                         break;
256                 }
257         }
258 }
259
260 int kvmppc_mmu_hv_init(void)
261 {
262         unsigned long host_lpid, rsvd_lpid;
263
264         if (!cpu_has_feature(CPU_FTR_HVMODE))
265                 return -EINVAL;
266
267         /* POWER7 has 10-bit LPIDs (12-bit in POWER8) */
268         host_lpid = mfspr(SPRN_LPID);
269         rsvd_lpid = LPID_RSVD;
270
271         kvmppc_init_lpid(rsvd_lpid + 1);
272
273         kvmppc_claim_lpid(host_lpid);
274         /* rsvd_lpid is reserved for use in partition switching */
275         kvmppc_claim_lpid(rsvd_lpid);
276
277         return 0;
278 }
279
280 static void kvmppc_mmu_book3s_64_hv_reset_msr(struct kvm_vcpu *vcpu)
281 {
282         unsigned long msr = vcpu->arch.intr_msr;
283
284         /* If transactional, change to suspend mode on IRQ delivery */
285         if (MSR_TM_TRANSACTIONAL(vcpu->arch.shregs.msr))
286                 msr |= MSR_TS_S;
287         else
288                 msr |= vcpu->arch.shregs.msr & MSR_TS_MASK;
289         kvmppc_set_msr(vcpu, msr);
290 }
291
292 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
293                                 long pte_index, unsigned long pteh,
294                                 unsigned long ptel, unsigned long *pte_idx_ret)
295 {
296         long ret;
297
298         /* Protect linux PTE lookup from page table destruction */
299         rcu_read_lock_sched();  /* this disables preemption too */
300         ret = kvmppc_do_h_enter(kvm, flags, pte_index, pteh, ptel,
301                                 current->mm->pgd, false, pte_idx_ret);
302         rcu_read_unlock_sched();
303         if (ret == H_TOO_HARD) {
304                 /* this can't happen */
305                 pr_err("KVM: Oops, kvmppc_h_enter returned too hard!\n");
306                 ret = H_RESOURCE;       /* or something */
307         }
308         return ret;
309
310 }
311
312 static struct kvmppc_slb *kvmppc_mmu_book3s_hv_find_slbe(struct kvm_vcpu *vcpu,
313                                                          gva_t eaddr)
314 {
315         u64 mask;
316         int i;
317
318         for (i = 0; i < vcpu->arch.slb_nr; i++) {
319                 if (!(vcpu->arch.slb[i].orige & SLB_ESID_V))
320                         continue;
321
322                 if (vcpu->arch.slb[i].origv & SLB_VSID_B_1T)
323                         mask = ESID_MASK_1T;
324                 else
325                         mask = ESID_MASK;
326
327                 if (((vcpu->arch.slb[i].orige ^ eaddr) & mask) == 0)
328                         return &vcpu->arch.slb[i];
329         }
330         return NULL;
331 }
332
333 static unsigned long kvmppc_mmu_get_real_addr(unsigned long v, unsigned long r,
334                         unsigned long ea)
335 {
336         unsigned long ra_mask;
337
338         ra_mask = kvmppc_actual_pgsz(v, r) - 1;
339         return (r & HPTE_R_RPN & ~ra_mask) | (ea & ra_mask);
340 }
341
342 static int kvmppc_mmu_book3s_64_hv_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
343                         struct kvmppc_pte *gpte, bool data, bool iswrite)
344 {
345         struct kvm *kvm = vcpu->kvm;
346         struct kvmppc_slb *slbe;
347         unsigned long slb_v;
348         unsigned long pp, key;
349         unsigned long v, orig_v, gr;
350         __be64 *hptep;
351         int index;
352         int virtmode = vcpu->arch.shregs.msr & (data ? MSR_DR : MSR_IR);
353
354         if (kvm_is_radix(vcpu->kvm))
355                 return kvmppc_mmu_radix_xlate(vcpu, eaddr, gpte, data, iswrite);
356
357         /* Get SLB entry */
358         if (virtmode) {
359                 slbe = kvmppc_mmu_book3s_hv_find_slbe(vcpu, eaddr);
360                 if (!slbe)
361                         return -EINVAL;
362                 slb_v = slbe->origv;
363         } else {
364                 /* real mode access */
365                 slb_v = vcpu->kvm->arch.vrma_slb_v;
366         }
367
368         preempt_disable();
369         /* Find the HPTE in the hash table */
370         index = kvmppc_hv_find_lock_hpte(kvm, eaddr, slb_v,
371                                          HPTE_V_VALID | HPTE_V_ABSENT);
372         if (index < 0) {
373                 preempt_enable();
374                 return -ENOENT;
375         }
376         hptep = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
377         v = orig_v = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
378         if (cpu_has_feature(CPU_FTR_ARCH_300))
379                 v = hpte_new_to_old_v(v, be64_to_cpu(hptep[1]));
380         gr = kvm->arch.hpt.rev[index].guest_rpte;
381
382         unlock_hpte(hptep, orig_v);
383         preempt_enable();
384
385         gpte->eaddr = eaddr;
386         gpte->vpage = ((v & HPTE_V_AVPN) << 4) | ((eaddr >> 12) & 0xfff);
387
388         /* Get PP bits and key for permission check */
389         pp = gr & (HPTE_R_PP0 | HPTE_R_PP);
390         key = (vcpu->arch.shregs.msr & MSR_PR) ? SLB_VSID_KP : SLB_VSID_KS;
391         key &= slb_v;
392
393         /* Calculate permissions */
394         gpte->may_read = hpte_read_permission(pp, key);
395         gpte->may_write = hpte_write_permission(pp, key);
396         gpte->may_execute = gpte->may_read && !(gr & (HPTE_R_N | HPTE_R_G));
397
398         /* Storage key permission check for POWER7 */
399         if (data && virtmode) {
400                 int amrfield = hpte_get_skey_perm(gr, vcpu->arch.amr);
401                 if (amrfield & 1)
402                         gpte->may_read = 0;
403                 if (amrfield & 2)
404                         gpte->may_write = 0;
405         }
406
407         /* Get the guest physical address */
408         gpte->raddr = kvmppc_mmu_get_real_addr(v, gr, eaddr);
409         return 0;
410 }
411
412 /*
413  * Quick test for whether an instruction is a load or a store.
414  * If the instruction is a load or a store, then this will indicate
415  * which it is, at least on server processors.  (Embedded processors
416  * have some external PID instructions that don't follow the rule
417  * embodied here.)  If the instruction isn't a load or store, then
418  * this doesn't return anything useful.
419  */
420 static int instruction_is_store(unsigned int instr)
421 {
422         unsigned int mask;
423
424         mask = 0x10000000;
425         if ((instr & 0xfc000000) == 0x7c000000)
426                 mask = 0x100;           /* major opcode 31 */
427         return (instr & mask) != 0;
428 }
429
430 int kvmppc_hv_emulate_mmio(struct kvm_run *run, struct kvm_vcpu *vcpu,
431                            unsigned long gpa, gva_t ea, int is_store)
432 {
433         u32 last_inst;
434
435         /*
436          * If we fail, we just return to the guest and try executing it again.
437          */
438         if (kvmppc_get_last_inst(vcpu, INST_GENERIC, &last_inst) !=
439                 EMULATE_DONE)
440                 return RESUME_GUEST;
441
442         /*
443          * WARNING: We do not know for sure whether the instruction we just
444          * read from memory is the same that caused the fault in the first
445          * place.  If the instruction we read is neither an load or a store,
446          * then it can't access memory, so we don't need to worry about
447          * enforcing access permissions.  So, assuming it is a load or
448          * store, we just check that its direction (load or store) is
449          * consistent with the original fault, since that's what we
450          * checked the access permissions against.  If there is a mismatch
451          * we just return and retry the instruction.
452          */
453
454         if (instruction_is_store(last_inst) != !!is_store)
455                 return RESUME_GUEST;
456
457         /*
458          * Emulated accesses are emulated by looking at the hash for
459          * translation once, then performing the access later. The
460          * translation could be invalidated in the meantime in which
461          * point performing the subsequent memory access on the old
462          * physical address could possibly be a security hole for the
463          * guest (but not the host).
464          *
465          * This is less of an issue for MMIO stores since they aren't
466          * globally visible. It could be an issue for MMIO loads to
467          * a certain extent but we'll ignore it for now.
468          */
469
470         vcpu->arch.paddr_accessed = gpa;
471         vcpu->arch.vaddr_accessed = ea;
472         return kvmppc_emulate_mmio(run, vcpu);
473 }
474
475 int kvmppc_book3s_hv_page_fault(struct kvm_run *run, struct kvm_vcpu *vcpu,
476                                 unsigned long ea, unsigned long dsisr)
477 {
478         struct kvm *kvm = vcpu->kvm;
479         unsigned long hpte[3], r;
480         unsigned long hnow_v, hnow_r;
481         __be64 *hptep;
482         unsigned long mmu_seq, psize, pte_size;
483         unsigned long gpa_base, gfn_base;
484         unsigned long gpa, gfn, hva, pfn;
485         struct kvm_memory_slot *memslot;
486         unsigned long *rmap;
487         struct revmap_entry *rev;
488         struct page *page, *pages[1];
489         long index, ret, npages;
490         bool is_ci;
491         unsigned int writing, write_ok;
492         struct vm_area_struct *vma;
493         unsigned long rcbits;
494         long mmio_update;
495
496         if (kvm_is_radix(kvm))
497                 return kvmppc_book3s_radix_page_fault(run, vcpu, ea, dsisr);
498
499         /*
500          * Real-mode code has already searched the HPT and found the
501          * entry we're interested in.  Lock the entry and check that
502          * it hasn't changed.  If it has, just return and re-execute the
503          * instruction.
504          */
505         if (ea != vcpu->arch.pgfault_addr)
506                 return RESUME_GUEST;
507
508         if (vcpu->arch.pgfault_cache) {
509                 mmio_update = atomic64_read(&kvm->arch.mmio_update);
510                 if (mmio_update == vcpu->arch.pgfault_cache->mmio_update) {
511                         r = vcpu->arch.pgfault_cache->rpte;
512                         psize = kvmppc_actual_pgsz(vcpu->arch.pgfault_hpte[0],
513                                                    r);
514                         gpa_base = r & HPTE_R_RPN & ~(psize - 1);
515                         gfn_base = gpa_base >> PAGE_SHIFT;
516                         gpa = gpa_base | (ea & (psize - 1));
517                         return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
518                                                 dsisr & DSISR_ISSTORE);
519                 }
520         }
521         index = vcpu->arch.pgfault_index;
522         hptep = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
523         rev = &kvm->arch.hpt.rev[index];
524         preempt_disable();
525         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
526                 cpu_relax();
527         hpte[0] = be64_to_cpu(hptep[0]) & ~HPTE_V_HVLOCK;
528         hpte[1] = be64_to_cpu(hptep[1]);
529         hpte[2] = r = rev->guest_rpte;
530         unlock_hpte(hptep, hpte[0]);
531         preempt_enable();
532
533         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
534                 hpte[0] = hpte_new_to_old_v(hpte[0], hpte[1]);
535                 hpte[1] = hpte_new_to_old_r(hpte[1]);
536         }
537         if (hpte[0] != vcpu->arch.pgfault_hpte[0] ||
538             hpte[1] != vcpu->arch.pgfault_hpte[1])
539                 return RESUME_GUEST;
540
541         /* Translate the logical address and get the page */
542         psize = kvmppc_actual_pgsz(hpte[0], r);
543         gpa_base = r & HPTE_R_RPN & ~(psize - 1);
544         gfn_base = gpa_base >> PAGE_SHIFT;
545         gpa = gpa_base | (ea & (psize - 1));
546         gfn = gpa >> PAGE_SHIFT;
547         memslot = gfn_to_memslot(kvm, gfn);
548
549         trace_kvm_page_fault_enter(vcpu, hpte, memslot, ea, dsisr);
550
551         /* No memslot means it's an emulated MMIO region */
552         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
553                 return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
554                                               dsisr & DSISR_ISSTORE);
555
556         /*
557          * This should never happen, because of the slot_is_aligned()
558          * check in kvmppc_do_h_enter().
559          */
560         if (gfn_base < memslot->base_gfn)
561                 return -EFAULT;
562
563         /* used to check for invalidations in progress */
564         mmu_seq = kvm->mmu_notifier_seq;
565         smp_rmb();
566
567         ret = -EFAULT;
568         is_ci = false;
569         pfn = 0;
570         page = NULL;
571         pte_size = PAGE_SIZE;
572         writing = (dsisr & DSISR_ISSTORE) != 0;
573         /* If writing != 0, then the HPTE must allow writing, if we get here */
574         write_ok = writing;
575         hva = gfn_to_hva_memslot(memslot, gfn);
576         npages = get_user_pages_fast(hva, 1, writing, pages);
577         if (npages < 1) {
578                 /* Check if it's an I/O mapping */
579                 down_read(&current->mm->mmap_sem);
580                 vma = find_vma(current->mm, hva);
581                 if (vma && vma->vm_start <= hva && hva + psize <= vma->vm_end &&
582                     (vma->vm_flags & VM_PFNMAP)) {
583                         pfn = vma->vm_pgoff +
584                                 ((hva - vma->vm_start) >> PAGE_SHIFT);
585                         pte_size = psize;
586                         is_ci = pte_ci(__pte((pgprot_val(vma->vm_page_prot))));
587                         write_ok = vma->vm_flags & VM_WRITE;
588                 }
589                 up_read(&current->mm->mmap_sem);
590                 if (!pfn)
591                         goto out_put;
592         } else {
593                 page = pages[0];
594                 pfn = page_to_pfn(page);
595                 if (PageHuge(page)) {
596                         page = compound_head(page);
597                         pte_size <<= compound_order(page);
598                 }
599                 /* if the guest wants write access, see if that is OK */
600                 if (!writing && hpte_is_writable(r)) {
601                         pte_t *ptep, pte;
602                         unsigned long flags;
603                         /*
604                          * We need to protect against page table destruction
605                          * hugepage split and collapse.
606                          */
607                         local_irq_save(flags);
608                         ptep = find_current_mm_pte(current->mm->pgd,
609                                                    hva, NULL, NULL);
610                         if (ptep) {
611                                 pte = kvmppc_read_update_linux_pte(ptep, 1);
612                                 if (__pte_write(pte))
613                                         write_ok = 1;
614                         }
615                         local_irq_restore(flags);
616                 }
617         }
618
619         if (psize > pte_size)
620                 goto out_put;
621
622         /* Check WIMG vs. the actual page we're accessing */
623         if (!hpte_cache_flags_ok(r, is_ci)) {
624                 if (is_ci)
625                         goto out_put;
626                 /*
627                  * Allow guest to map emulated device memory as
628                  * uncacheable, but actually make it cacheable.
629                  */
630                 r = (r & ~(HPTE_R_W|HPTE_R_I|HPTE_R_G)) | HPTE_R_M;
631         }
632
633         /*
634          * Set the HPTE to point to pfn.
635          * Since the pfn is at PAGE_SIZE granularity, make sure we
636          * don't mask out lower-order bits if psize < PAGE_SIZE.
637          */
638         if (psize < PAGE_SIZE)
639                 psize = PAGE_SIZE;
640         r = (r & HPTE_R_KEY_HI) | (r & ~(HPTE_R_PP0 - psize)) |
641                                         ((pfn << PAGE_SHIFT) & ~(psize - 1));
642         if (hpte_is_writable(r) && !write_ok)
643                 r = hpte_make_readonly(r);
644         ret = RESUME_GUEST;
645         preempt_disable();
646         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
647                 cpu_relax();
648         hnow_v = be64_to_cpu(hptep[0]);
649         hnow_r = be64_to_cpu(hptep[1]);
650         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
651                 hnow_v = hpte_new_to_old_v(hnow_v, hnow_r);
652                 hnow_r = hpte_new_to_old_r(hnow_r);
653         }
654
655         /*
656          * If the HPT is being resized, don't update the HPTE,
657          * instead let the guest retry after the resize operation is complete.
658          * The synchronization for mmu_ready test vs. set is provided
659          * by the HPTE lock.
660          */
661         if (!kvm->arch.mmu_ready)
662                 goto out_unlock;
663
664         if ((hnow_v & ~HPTE_V_HVLOCK) != hpte[0] || hnow_r != hpte[1] ||
665             rev->guest_rpte != hpte[2])
666                 /* HPTE has been changed under us; let the guest retry */
667                 goto out_unlock;
668         hpte[0] = (hpte[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
669
670         /* Always put the HPTE in the rmap chain for the page base address */
671         rmap = &memslot->arch.rmap[gfn_base - memslot->base_gfn];
672         lock_rmap(rmap);
673
674         /* Check if we might have been invalidated; let the guest retry if so */
675         ret = RESUME_GUEST;
676         if (mmu_notifier_retry(vcpu->kvm, mmu_seq)) {
677                 unlock_rmap(rmap);
678                 goto out_unlock;
679         }
680
681         /* Only set R/C in real HPTE if set in both *rmap and guest_rpte */
682         rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
683         r &= rcbits | ~(HPTE_R_R | HPTE_R_C);
684
685         if (be64_to_cpu(hptep[0]) & HPTE_V_VALID) {
686                 /* HPTE was previously valid, so we need to invalidate it */
687                 unlock_rmap(rmap);
688                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
689                 kvmppc_invalidate_hpte(kvm, hptep, index);
690                 /* don't lose previous R and C bits */
691                 r |= be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
692         } else {
693                 kvmppc_add_revmap_chain(kvm, rev, rmap, index, 0);
694         }
695
696         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
697                 r = hpte_old_to_new_r(hpte[0], r);
698                 hpte[0] = hpte_old_to_new_v(hpte[0]);
699         }
700         hptep[1] = cpu_to_be64(r);
701         eieio();
702         __unlock_hpte(hptep, hpte[0]);
703         asm volatile("ptesync" : : : "memory");
704         preempt_enable();
705         if (page && hpte_is_writable(r))
706                 SetPageDirty(page);
707
708  out_put:
709         trace_kvm_page_fault_exit(vcpu, hpte, ret);
710
711         if (page) {
712                 /*
713                  * We drop pages[0] here, not page because page might
714                  * have been set to the head page of a compound, but
715                  * we have to drop the reference on the correct tail
716                  * page to match the get inside gup()
717                  */
718                 put_page(pages[0]);
719         }
720         return ret;
721
722  out_unlock:
723         __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
724         preempt_enable();
725         goto out_put;
726 }
727
728 void kvmppc_rmap_reset(struct kvm *kvm)
729 {
730         struct kvm_memslots *slots;
731         struct kvm_memory_slot *memslot;
732         int srcu_idx;
733
734         srcu_idx = srcu_read_lock(&kvm->srcu);
735         slots = kvm_memslots(kvm);
736         kvm_for_each_memslot(memslot, slots) {
737                 /*
738                  * This assumes it is acceptable to lose reference and
739                  * change bits across a reset.
740                  */
741                 memset(memslot->arch.rmap, 0,
742                        memslot->npages * sizeof(*memslot->arch.rmap));
743         }
744         srcu_read_unlock(&kvm->srcu, srcu_idx);
745 }
746
747 typedef int (*hva_handler_fn)(struct kvm *kvm, struct kvm_memory_slot *memslot,
748                               unsigned long gfn);
749
750 static int kvm_handle_hva_range(struct kvm *kvm,
751                                 unsigned long start,
752                                 unsigned long end,
753                                 hva_handler_fn handler)
754 {
755         int ret;
756         int retval = 0;
757         struct kvm_memslots *slots;
758         struct kvm_memory_slot *memslot;
759
760         slots = kvm_memslots(kvm);
761         kvm_for_each_memslot(memslot, slots) {
762                 unsigned long hva_start, hva_end;
763                 gfn_t gfn, gfn_end;
764
765                 hva_start = max(start, memslot->userspace_addr);
766                 hva_end = min(end, memslot->userspace_addr +
767                                         (memslot->npages << PAGE_SHIFT));
768                 if (hva_start >= hva_end)
769                         continue;
770                 /*
771                  * {gfn(page) | page intersects with [hva_start, hva_end)} =
772                  * {gfn, gfn+1, ..., gfn_end-1}.
773                  */
774                 gfn = hva_to_gfn_memslot(hva_start, memslot);
775                 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
776
777                 for (; gfn < gfn_end; ++gfn) {
778                         ret = handler(kvm, memslot, gfn);
779                         retval |= ret;
780                 }
781         }
782
783         return retval;
784 }
785
786 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
787                           hva_handler_fn handler)
788 {
789         return kvm_handle_hva_range(kvm, hva, hva + 1, handler);
790 }
791
792 /* Must be called with both HPTE and rmap locked */
793 static void kvmppc_unmap_hpte(struct kvm *kvm, unsigned long i,
794                               struct kvm_memory_slot *memslot,
795                               unsigned long *rmapp, unsigned long gfn)
796 {
797         __be64 *hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
798         struct revmap_entry *rev = kvm->arch.hpt.rev;
799         unsigned long j, h;
800         unsigned long ptel, psize, rcbits;
801
802         j = rev[i].forw;
803         if (j == i) {
804                 /* chain is now empty */
805                 *rmapp &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
806         } else {
807                 /* remove i from chain */
808                 h = rev[i].back;
809                 rev[h].forw = j;
810                 rev[j].back = h;
811                 rev[i].forw = rev[i].back = i;
812                 *rmapp = (*rmapp & ~KVMPPC_RMAP_INDEX) | j;
813         }
814
815         /* Now check and modify the HPTE */
816         ptel = rev[i].guest_rpte;
817         psize = kvmppc_actual_pgsz(be64_to_cpu(hptep[0]), ptel);
818         if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
819             hpte_rpn(ptel, psize) == gfn) {
820                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
821                 kvmppc_invalidate_hpte(kvm, hptep, i);
822                 hptep[1] &= ~cpu_to_be64(HPTE_R_KEY_HI | HPTE_R_KEY_LO);
823                 /* Harvest R and C */
824                 rcbits = be64_to_cpu(hptep[1]) & (HPTE_R_R | HPTE_R_C);
825                 *rmapp |= rcbits << KVMPPC_RMAP_RC_SHIFT;
826                 if ((rcbits & HPTE_R_C) && memslot->dirty_bitmap)
827                         kvmppc_update_dirty_map(memslot, gfn, psize);
828                 if (rcbits & ~rev[i].guest_rpte) {
829                         rev[i].guest_rpte = ptel | rcbits;
830                         note_hpte_modification(kvm, &rev[i]);
831                 }
832         }
833 }
834
835 static int kvm_unmap_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
836                            unsigned long gfn)
837 {
838         unsigned long i;
839         __be64 *hptep;
840         unsigned long *rmapp;
841
842         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
843         for (;;) {
844                 lock_rmap(rmapp);
845                 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
846                         unlock_rmap(rmapp);
847                         break;
848                 }
849
850                 /*
851                  * To avoid an ABBA deadlock with the HPTE lock bit,
852                  * we can't spin on the HPTE lock while holding the
853                  * rmap chain lock.
854                  */
855                 i = *rmapp & KVMPPC_RMAP_INDEX;
856                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
857                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
858                         /* unlock rmap before spinning on the HPTE lock */
859                         unlock_rmap(rmapp);
860                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
861                                 cpu_relax();
862                         continue;
863                 }
864
865                 kvmppc_unmap_hpte(kvm, i, memslot, rmapp, gfn);
866                 unlock_rmap(rmapp);
867                 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
868         }
869         return 0;
870 }
871
872 int kvm_unmap_hva_hv(struct kvm *kvm, unsigned long hva)
873 {
874         hva_handler_fn handler;
875
876         handler = kvm_is_radix(kvm) ? kvm_unmap_radix : kvm_unmap_rmapp;
877         kvm_handle_hva(kvm, hva, handler);
878         return 0;
879 }
880
881 int kvm_unmap_hva_range_hv(struct kvm *kvm, unsigned long start, unsigned long end)
882 {
883         hva_handler_fn handler;
884
885         handler = kvm_is_radix(kvm) ? kvm_unmap_radix : kvm_unmap_rmapp;
886         kvm_handle_hva_range(kvm, start, end, handler);
887         return 0;
888 }
889
890 void kvmppc_core_flush_memslot_hv(struct kvm *kvm,
891                                   struct kvm_memory_slot *memslot)
892 {
893         unsigned long gfn;
894         unsigned long n;
895         unsigned long *rmapp;
896
897         gfn = memslot->base_gfn;
898         rmapp = memslot->arch.rmap;
899         for (n = memslot->npages; n; --n, ++gfn) {
900                 if (kvm_is_radix(kvm)) {
901                         kvm_unmap_radix(kvm, memslot, gfn);
902                         continue;
903                 }
904                 /*
905                  * Testing the present bit without locking is OK because
906                  * the memslot has been marked invalid already, and hence
907                  * no new HPTEs referencing this page can be created,
908                  * thus the present bit can't go from 0 to 1.
909                  */
910                 if (*rmapp & KVMPPC_RMAP_PRESENT)
911                         kvm_unmap_rmapp(kvm, memslot, gfn);
912                 ++rmapp;
913         }
914 }
915
916 static int kvm_age_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
917                          unsigned long gfn)
918 {
919         struct revmap_entry *rev = kvm->arch.hpt.rev;
920         unsigned long head, i, j;
921         __be64 *hptep;
922         int ret = 0;
923         unsigned long *rmapp;
924
925         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
926  retry:
927         lock_rmap(rmapp);
928         if (*rmapp & KVMPPC_RMAP_REFERENCED) {
929                 *rmapp &= ~KVMPPC_RMAP_REFERENCED;
930                 ret = 1;
931         }
932         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
933                 unlock_rmap(rmapp);
934                 return ret;
935         }
936
937         i = head = *rmapp & KVMPPC_RMAP_INDEX;
938         do {
939                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
940                 j = rev[i].forw;
941
942                 /* If this HPTE isn't referenced, ignore it */
943                 if (!(be64_to_cpu(hptep[1]) & HPTE_R_R))
944                         continue;
945
946                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
947                         /* unlock rmap before spinning on the HPTE lock */
948                         unlock_rmap(rmapp);
949                         while (be64_to_cpu(hptep[0]) & HPTE_V_HVLOCK)
950                                 cpu_relax();
951                         goto retry;
952                 }
953
954                 /* Now check and modify the HPTE */
955                 if ((be64_to_cpu(hptep[0]) & HPTE_V_VALID) &&
956                     (be64_to_cpu(hptep[1]) & HPTE_R_R)) {
957                         kvmppc_clear_ref_hpte(kvm, hptep, i);
958                         if (!(rev[i].guest_rpte & HPTE_R_R)) {
959                                 rev[i].guest_rpte |= HPTE_R_R;
960                                 note_hpte_modification(kvm, &rev[i]);
961                         }
962                         ret = 1;
963                 }
964                 __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
965         } while ((i = j) != head);
966
967         unlock_rmap(rmapp);
968         return ret;
969 }
970
971 int kvm_age_hva_hv(struct kvm *kvm, unsigned long start, unsigned long end)
972 {
973         hva_handler_fn handler;
974
975         handler = kvm_is_radix(kvm) ? kvm_age_radix : kvm_age_rmapp;
976         return kvm_handle_hva_range(kvm, start, end, handler);
977 }
978
979 static int kvm_test_age_rmapp(struct kvm *kvm, struct kvm_memory_slot *memslot,
980                               unsigned long gfn)
981 {
982         struct revmap_entry *rev = kvm->arch.hpt.rev;
983         unsigned long head, i, j;
984         unsigned long *hp;
985         int ret = 1;
986         unsigned long *rmapp;
987
988         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
989         if (*rmapp & KVMPPC_RMAP_REFERENCED)
990                 return 1;
991
992         lock_rmap(rmapp);
993         if (*rmapp & KVMPPC_RMAP_REFERENCED)
994                 goto out;
995
996         if (*rmapp & KVMPPC_RMAP_PRESENT) {
997                 i = head = *rmapp & KVMPPC_RMAP_INDEX;
998                 do {
999                         hp = (unsigned long *)(kvm->arch.hpt.virt + (i << 4));
1000                         j = rev[i].forw;
1001                         if (be64_to_cpu(hp[1]) & HPTE_R_R)
1002                                 goto out;
1003                 } while ((i = j) != head);
1004         }
1005         ret = 0;
1006
1007  out:
1008         unlock_rmap(rmapp);
1009         return ret;
1010 }
1011
1012 int kvm_test_age_hva_hv(struct kvm *kvm, unsigned long hva)
1013 {
1014         hva_handler_fn handler;
1015
1016         handler = kvm_is_radix(kvm) ? kvm_test_age_radix : kvm_test_age_rmapp;
1017         return kvm_handle_hva(kvm, hva, handler);
1018 }
1019
1020 void kvm_set_spte_hva_hv(struct kvm *kvm, unsigned long hva, pte_t pte)
1021 {
1022         hva_handler_fn handler;
1023
1024         handler = kvm_is_radix(kvm) ? kvm_unmap_radix : kvm_unmap_rmapp;
1025         kvm_handle_hva(kvm, hva, handler);
1026 }
1027
1028 static int vcpus_running(struct kvm *kvm)
1029 {
1030         return atomic_read(&kvm->arch.vcpus_running) != 0;
1031 }
1032
1033 /*
1034  * Returns the number of system pages that are dirty.
1035  * This can be more than 1 if we find a huge-page HPTE.
1036  */
1037 static int kvm_test_clear_dirty_npages(struct kvm *kvm, unsigned long *rmapp)
1038 {
1039         struct revmap_entry *rev = kvm->arch.hpt.rev;
1040         unsigned long head, i, j;
1041         unsigned long n;
1042         unsigned long v, r;
1043         __be64 *hptep;
1044         int npages_dirty = 0;
1045
1046  retry:
1047         lock_rmap(rmapp);
1048         if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
1049                 unlock_rmap(rmapp);
1050                 return npages_dirty;
1051         }
1052
1053         i = head = *rmapp & KVMPPC_RMAP_INDEX;
1054         do {
1055                 unsigned long hptep1;
1056                 hptep = (__be64 *) (kvm->arch.hpt.virt + (i << 4));
1057                 j = rev[i].forw;
1058
1059                 /*
1060                  * Checking the C (changed) bit here is racy since there
1061                  * is no guarantee about when the hardware writes it back.
1062                  * If the HPTE is not writable then it is stable since the
1063                  * page can't be written to, and we would have done a tlbie
1064                  * (which forces the hardware to complete any writeback)
1065                  * when making the HPTE read-only.
1066                  * If vcpus are running then this call is racy anyway
1067                  * since the page could get dirtied subsequently, so we
1068                  * expect there to be a further call which would pick up
1069                  * any delayed C bit writeback.
1070                  * Otherwise we need to do the tlbie even if C==0 in
1071                  * order to pick up any delayed writeback of C.
1072                  */
1073                 hptep1 = be64_to_cpu(hptep[1]);
1074                 if (!(hptep1 & HPTE_R_C) &&
1075                     (!hpte_is_writable(hptep1) || vcpus_running(kvm)))
1076                         continue;
1077
1078                 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
1079                         /* unlock rmap before spinning on the HPTE lock */
1080                         unlock_rmap(rmapp);
1081                         while (hptep[0] & cpu_to_be64(HPTE_V_HVLOCK))
1082                                 cpu_relax();
1083                         goto retry;
1084                 }
1085
1086                 /* Now check and modify the HPTE */
1087                 if (!(hptep[0] & cpu_to_be64(HPTE_V_VALID))) {
1088                         __unlock_hpte(hptep, be64_to_cpu(hptep[0]));
1089                         continue;
1090                 }
1091
1092                 /* need to make it temporarily absent so C is stable */
1093                 hptep[0] |= cpu_to_be64(HPTE_V_ABSENT);
1094                 kvmppc_invalidate_hpte(kvm, hptep, i);
1095                 v = be64_to_cpu(hptep[0]);
1096                 r = be64_to_cpu(hptep[1]);
1097                 if (r & HPTE_R_C) {
1098                         hptep[1] = cpu_to_be64(r & ~HPTE_R_C);
1099                         if (!(rev[i].guest_rpte & HPTE_R_C)) {
1100                                 rev[i].guest_rpte |= HPTE_R_C;
1101                                 note_hpte_modification(kvm, &rev[i]);
1102                         }
1103                         n = kvmppc_actual_pgsz(v, r);
1104                         n = (n + PAGE_SIZE - 1) >> PAGE_SHIFT;
1105                         if (n > npages_dirty)
1106                                 npages_dirty = n;
1107                         eieio();
1108                 }
1109                 v &= ~HPTE_V_ABSENT;
1110                 v |= HPTE_V_VALID;
1111                 __unlock_hpte(hptep, v);
1112         } while ((i = j) != head);
1113
1114         unlock_rmap(rmapp);
1115         return npages_dirty;
1116 }
1117
1118 void kvmppc_harvest_vpa_dirty(struct kvmppc_vpa *vpa,
1119                               struct kvm_memory_slot *memslot,
1120                               unsigned long *map)
1121 {
1122         unsigned long gfn;
1123
1124         if (!vpa->dirty || !vpa->pinned_addr)
1125                 return;
1126         gfn = vpa->gpa >> PAGE_SHIFT;
1127         if (gfn < memslot->base_gfn ||
1128             gfn >= memslot->base_gfn + memslot->npages)
1129                 return;
1130
1131         vpa->dirty = false;
1132         if (map)
1133                 __set_bit_le(gfn - memslot->base_gfn, map);
1134 }
1135
1136 long kvmppc_hv_get_dirty_log_hpt(struct kvm *kvm,
1137                         struct kvm_memory_slot *memslot, unsigned long *map)
1138 {
1139         unsigned long i;
1140         unsigned long *rmapp;
1141
1142         preempt_disable();
1143         rmapp = memslot->arch.rmap;
1144         for (i = 0; i < memslot->npages; ++i) {
1145                 int npages = kvm_test_clear_dirty_npages(kvm, rmapp);
1146                 /*
1147                  * Note that if npages > 0 then i must be a multiple of npages,
1148                  * since we always put huge-page HPTEs in the rmap chain
1149                  * corresponding to their page base address.
1150                  */
1151                 if (npages)
1152                         set_dirty_bits(map, i, npages);
1153                 ++rmapp;
1154         }
1155         preempt_enable();
1156         return 0;
1157 }
1158
1159 void *kvmppc_pin_guest_page(struct kvm *kvm, unsigned long gpa,
1160                             unsigned long *nb_ret)
1161 {
1162         struct kvm_memory_slot *memslot;
1163         unsigned long gfn = gpa >> PAGE_SHIFT;
1164         struct page *page, *pages[1];
1165         int npages;
1166         unsigned long hva, offset;
1167         int srcu_idx;
1168
1169         srcu_idx = srcu_read_lock(&kvm->srcu);
1170         memslot = gfn_to_memslot(kvm, gfn);
1171         if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
1172                 goto err;
1173         hva = gfn_to_hva_memslot(memslot, gfn);
1174         npages = get_user_pages_fast(hva, 1, 1, pages);
1175         if (npages < 1)
1176                 goto err;
1177         page = pages[0];
1178         srcu_read_unlock(&kvm->srcu, srcu_idx);
1179
1180         offset = gpa & (PAGE_SIZE - 1);
1181         if (nb_ret)
1182                 *nb_ret = PAGE_SIZE - offset;
1183         return page_address(page) + offset;
1184
1185  err:
1186         srcu_read_unlock(&kvm->srcu, srcu_idx);
1187         return NULL;
1188 }
1189
1190 void kvmppc_unpin_guest_page(struct kvm *kvm, void *va, unsigned long gpa,
1191                              bool dirty)
1192 {
1193         struct page *page = virt_to_page(va);
1194         struct kvm_memory_slot *memslot;
1195         unsigned long gfn;
1196         int srcu_idx;
1197
1198         put_page(page);
1199
1200         if (!dirty)
1201                 return;
1202
1203         /* We need to mark this page dirty in the memslot dirty_bitmap, if any */
1204         gfn = gpa >> PAGE_SHIFT;
1205         srcu_idx = srcu_read_lock(&kvm->srcu);
1206         memslot = gfn_to_memslot(kvm, gfn);
1207         if (memslot && memslot->dirty_bitmap)
1208                 set_bit_le(gfn - memslot->base_gfn, memslot->dirty_bitmap);
1209         srcu_read_unlock(&kvm->srcu, srcu_idx);
1210 }
1211
1212 /*
1213  * HPT resizing
1214  */
1215 static int resize_hpt_allocate(struct kvm_resize_hpt *resize)
1216 {
1217         int rc;
1218
1219         rc = kvmppc_allocate_hpt(&resize->hpt, resize->order);
1220         if (rc < 0)
1221                 return rc;
1222
1223         resize_hpt_debug(resize, "resize_hpt_allocate(): HPT @ 0x%lx\n",
1224                          resize->hpt.virt);
1225
1226         return 0;
1227 }
1228
1229 static unsigned long resize_hpt_rehash_hpte(struct kvm_resize_hpt *resize,
1230                                             unsigned long idx)
1231 {
1232         struct kvm *kvm = resize->kvm;
1233         struct kvm_hpt_info *old = &kvm->arch.hpt;
1234         struct kvm_hpt_info *new = &resize->hpt;
1235         unsigned long old_hash_mask = (1ULL << (old->order - 7)) - 1;
1236         unsigned long new_hash_mask = (1ULL << (new->order - 7)) - 1;
1237         __be64 *hptep, *new_hptep;
1238         unsigned long vpte, rpte, guest_rpte;
1239         int ret;
1240         struct revmap_entry *rev;
1241         unsigned long apsize, avpn, pteg, hash;
1242         unsigned long new_idx, new_pteg, replace_vpte;
1243         int pshift;
1244
1245         hptep = (__be64 *)(old->virt + (idx << 4));
1246
1247         /* Guest is stopped, so new HPTEs can't be added or faulted
1248          * in, only unmapped or altered by host actions.  So, it's
1249          * safe to check this before we take the HPTE lock */
1250         vpte = be64_to_cpu(hptep[0]);
1251         if (!(vpte & HPTE_V_VALID) && !(vpte & HPTE_V_ABSENT))
1252                 return 0; /* nothing to do */
1253
1254         while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
1255                 cpu_relax();
1256
1257         vpte = be64_to_cpu(hptep[0]);
1258
1259         ret = 0;
1260         if (!(vpte & HPTE_V_VALID) && !(vpte & HPTE_V_ABSENT))
1261                 /* Nothing to do */
1262                 goto out;
1263
1264         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1265                 rpte = be64_to_cpu(hptep[1]);
1266                 vpte = hpte_new_to_old_v(vpte, rpte);
1267         }
1268
1269         /* Unmap */
1270         rev = &old->rev[idx];
1271         guest_rpte = rev->guest_rpte;
1272
1273         ret = -EIO;
1274         apsize = kvmppc_actual_pgsz(vpte, guest_rpte);
1275         if (!apsize)
1276                 goto out;
1277
1278         if (vpte & HPTE_V_VALID) {
1279                 unsigned long gfn = hpte_rpn(guest_rpte, apsize);
1280                 int srcu_idx = srcu_read_lock(&kvm->srcu);
1281                 struct kvm_memory_slot *memslot =
1282                         __gfn_to_memslot(kvm_memslots(kvm), gfn);
1283
1284                 if (memslot) {
1285                         unsigned long *rmapp;
1286                         rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1287
1288                         lock_rmap(rmapp);
1289                         kvmppc_unmap_hpte(kvm, idx, memslot, rmapp, gfn);
1290                         unlock_rmap(rmapp);
1291                 }
1292
1293                 srcu_read_unlock(&kvm->srcu, srcu_idx);
1294         }
1295
1296         /* Reload PTE after unmap */
1297         vpte = be64_to_cpu(hptep[0]);
1298         BUG_ON(vpte & HPTE_V_VALID);
1299         BUG_ON(!(vpte & HPTE_V_ABSENT));
1300
1301         ret = 0;
1302         if (!(vpte & HPTE_V_BOLTED))
1303                 goto out;
1304
1305         rpte = be64_to_cpu(hptep[1]);
1306
1307         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1308                 vpte = hpte_new_to_old_v(vpte, rpte);
1309                 rpte = hpte_new_to_old_r(rpte);
1310         }
1311
1312         pshift = kvmppc_hpte_base_page_shift(vpte, rpte);
1313         avpn = HPTE_V_AVPN_VAL(vpte) & ~(((1ul << pshift) - 1) >> 23);
1314         pteg = idx / HPTES_PER_GROUP;
1315         if (vpte & HPTE_V_SECONDARY)
1316                 pteg = ~pteg;
1317
1318         if (!(vpte & HPTE_V_1TB_SEG)) {
1319                 unsigned long offset, vsid;
1320
1321                 /* We only have 28 - 23 bits of offset in avpn */
1322                 offset = (avpn & 0x1f) << 23;
1323                 vsid = avpn >> 5;
1324                 /* We can find more bits from the pteg value */
1325                 if (pshift < 23)
1326                         offset |= ((vsid ^ pteg) & old_hash_mask) << pshift;
1327
1328                 hash = vsid ^ (offset >> pshift);
1329         } else {
1330                 unsigned long offset, vsid;
1331
1332                 /* We only have 40 - 23 bits of seg_off in avpn */
1333                 offset = (avpn & 0x1ffff) << 23;
1334                 vsid = avpn >> 17;
1335                 if (pshift < 23)
1336                         offset |= ((vsid ^ (vsid << 25) ^ pteg) & old_hash_mask) << pshift;
1337
1338                 hash = vsid ^ (vsid << 25) ^ (offset >> pshift);
1339         }
1340
1341         new_pteg = hash & new_hash_mask;
1342         if (vpte & HPTE_V_SECONDARY)
1343                 new_pteg = ~hash & new_hash_mask;
1344
1345         new_idx = new_pteg * HPTES_PER_GROUP + (idx % HPTES_PER_GROUP);
1346         new_hptep = (__be64 *)(new->virt + (new_idx << 4));
1347
1348         replace_vpte = be64_to_cpu(new_hptep[0]);
1349         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1350                 unsigned long replace_rpte = be64_to_cpu(new_hptep[1]);
1351                 replace_vpte = hpte_new_to_old_v(replace_vpte, replace_rpte);
1352         }
1353
1354         if (replace_vpte & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1355                 BUG_ON(new->order >= old->order);
1356
1357                 if (replace_vpte & HPTE_V_BOLTED) {
1358                         if (vpte & HPTE_V_BOLTED)
1359                                 /* Bolted collision, nothing we can do */
1360                                 ret = -ENOSPC;
1361                         /* Discard the new HPTE */
1362                         goto out;
1363                 }
1364
1365                 /* Discard the previous HPTE */
1366         }
1367
1368         if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1369                 rpte = hpte_old_to_new_r(vpte, rpte);
1370                 vpte = hpte_old_to_new_v(vpte);
1371         }
1372
1373         new_hptep[1] = cpu_to_be64(rpte);
1374         new->rev[new_idx].guest_rpte = guest_rpte;
1375         /* No need for a barrier, since new HPT isn't active */
1376         new_hptep[0] = cpu_to_be64(vpte);
1377         unlock_hpte(new_hptep, vpte);
1378
1379 out:
1380         unlock_hpte(hptep, vpte);
1381         return ret;
1382 }
1383
1384 static int resize_hpt_rehash(struct kvm_resize_hpt *resize)
1385 {
1386         struct kvm *kvm = resize->kvm;
1387         unsigned  long i;
1388         int rc;
1389
1390         for (i = 0; i < kvmppc_hpt_npte(&kvm->arch.hpt); i++) {
1391                 rc = resize_hpt_rehash_hpte(resize, i);
1392                 if (rc != 0)
1393                         return rc;
1394         }
1395
1396         return 0;
1397 }
1398
1399 static void resize_hpt_pivot(struct kvm_resize_hpt *resize)
1400 {
1401         struct kvm *kvm = resize->kvm;
1402         struct kvm_hpt_info hpt_tmp;
1403
1404         /* Exchange the pending tables in the resize structure with
1405          * the active tables */
1406
1407         resize_hpt_debug(resize, "resize_hpt_pivot()\n");
1408
1409         spin_lock(&kvm->mmu_lock);
1410         asm volatile("ptesync" : : : "memory");
1411
1412         hpt_tmp = kvm->arch.hpt;
1413         kvmppc_set_hpt(kvm, &resize->hpt);
1414         resize->hpt = hpt_tmp;
1415
1416         spin_unlock(&kvm->mmu_lock);
1417
1418         synchronize_srcu_expedited(&kvm->srcu);
1419
1420         if (cpu_has_feature(CPU_FTR_ARCH_300))
1421                 kvmppc_setup_partition_table(kvm);
1422
1423         resize_hpt_debug(resize, "resize_hpt_pivot() done\n");
1424 }
1425
1426 static void resize_hpt_release(struct kvm *kvm, struct kvm_resize_hpt *resize)
1427 {
1428         BUG_ON(kvm->arch.resize_hpt != resize);
1429
1430         if (!resize)
1431                 return;
1432
1433         if (resize->hpt.virt)
1434                 kvmppc_free_hpt(&resize->hpt);
1435
1436         kvm->arch.resize_hpt = NULL;
1437         kfree(resize);
1438 }
1439
1440 static void resize_hpt_prepare_work(struct work_struct *work)
1441 {
1442         struct kvm_resize_hpt *resize = container_of(work,
1443                                                      struct kvm_resize_hpt,
1444                                                      work);
1445         struct kvm *kvm = resize->kvm;
1446         int err;
1447
1448         resize_hpt_debug(resize, "resize_hpt_prepare_work(): order = %d\n",
1449                          resize->order);
1450
1451         err = resize_hpt_allocate(resize);
1452
1453         mutex_lock(&kvm->lock);
1454
1455         resize->error = err;
1456         resize->prepare_done = true;
1457
1458         mutex_unlock(&kvm->lock);
1459 }
1460
1461 long kvm_vm_ioctl_resize_hpt_prepare(struct kvm *kvm,
1462                                      struct kvm_ppc_resize_hpt *rhpt)
1463 {
1464         unsigned long flags = rhpt->flags;
1465         unsigned long shift = rhpt->shift;
1466         struct kvm_resize_hpt *resize;
1467         int ret;
1468
1469         if (flags != 0 || kvm_is_radix(kvm))
1470                 return -EINVAL;
1471
1472         if (shift && ((shift < 18) || (shift > 46)))
1473                 return -EINVAL;
1474
1475         mutex_lock(&kvm->lock);
1476
1477         resize = kvm->arch.resize_hpt;
1478
1479         if (resize) {
1480                 if (resize->order == shift) {
1481                         /* Suitable resize in progress */
1482                         if (resize->prepare_done) {
1483                                 ret = resize->error;
1484                                 if (ret != 0)
1485                                         resize_hpt_release(kvm, resize);
1486                         } else {
1487                                 ret = 100; /* estimated time in ms */
1488                         }
1489
1490                         goto out;
1491                 }
1492
1493                 /* not suitable, cancel it */
1494                 resize_hpt_release(kvm, resize);
1495         }
1496
1497         ret = 0;
1498         if (!shift)
1499                 goto out; /* nothing to do */
1500
1501         /* start new resize */
1502
1503         resize = kzalloc(sizeof(*resize), GFP_KERNEL);
1504         if (!resize) {
1505                 ret = -ENOMEM;
1506                 goto out;
1507         }
1508         resize->order = shift;
1509         resize->kvm = kvm;
1510         INIT_WORK(&resize->work, resize_hpt_prepare_work);
1511         kvm->arch.resize_hpt = resize;
1512
1513         schedule_work(&resize->work);
1514
1515         ret = 100; /* estimated time in ms */
1516
1517 out:
1518         mutex_unlock(&kvm->lock);
1519         return ret;
1520 }
1521
1522 static void resize_hpt_boot_vcpu(void *opaque)
1523 {
1524         /* Nothing to do, just force a KVM exit */
1525 }
1526
1527 long kvm_vm_ioctl_resize_hpt_commit(struct kvm *kvm,
1528                                     struct kvm_ppc_resize_hpt *rhpt)
1529 {
1530         unsigned long flags = rhpt->flags;
1531         unsigned long shift = rhpt->shift;
1532         struct kvm_resize_hpt *resize;
1533         long ret;
1534
1535         if (flags != 0 || kvm_is_radix(kvm))
1536                 return -EINVAL;
1537
1538         if (shift && ((shift < 18) || (shift > 46)))
1539                 return -EINVAL;
1540
1541         mutex_lock(&kvm->lock);
1542
1543         resize = kvm->arch.resize_hpt;
1544
1545         /* This shouldn't be possible */
1546         ret = -EIO;
1547         if (WARN_ON(!kvm->arch.mmu_ready))
1548                 goto out_no_hpt;
1549
1550         /* Stop VCPUs from running while we mess with the HPT */
1551         kvm->arch.mmu_ready = 0;
1552         smp_mb();
1553
1554         /* Boot all CPUs out of the guest so they re-read
1555          * mmu_ready */
1556         on_each_cpu(resize_hpt_boot_vcpu, NULL, 1);
1557
1558         ret = -ENXIO;
1559         if (!resize || (resize->order != shift))
1560                 goto out;
1561
1562         ret = -EBUSY;
1563         if (!resize->prepare_done)
1564                 goto out;
1565
1566         ret = resize->error;
1567         if (ret != 0)
1568                 goto out;
1569
1570         ret = resize_hpt_rehash(resize);
1571         if (ret != 0)
1572                 goto out;
1573
1574         resize_hpt_pivot(resize);
1575
1576 out:
1577         /* Let VCPUs run again */
1578         kvm->arch.mmu_ready = 1;
1579         smp_mb();
1580 out_no_hpt:
1581         resize_hpt_release(kvm, resize);
1582         mutex_unlock(&kvm->lock);
1583         return ret;
1584 }
1585
1586 /*
1587  * Functions for reading and writing the hash table via reads and
1588  * writes on a file descriptor.
1589  *
1590  * Reads return the guest view of the hash table, which has to be
1591  * pieced together from the real hash table and the guest_rpte
1592  * values in the revmap array.
1593  *
1594  * On writes, each HPTE written is considered in turn, and if it
1595  * is valid, it is written to the HPT as if an H_ENTER with the
1596  * exact flag set was done.  When the invalid count is non-zero
1597  * in the header written to the stream, the kernel will make
1598  * sure that that many HPTEs are invalid, and invalidate them
1599  * if not.
1600  */
1601
1602 struct kvm_htab_ctx {
1603         unsigned long   index;
1604         unsigned long   flags;
1605         struct kvm      *kvm;
1606         int             first_pass;
1607 };
1608
1609 #define HPTE_SIZE       (2 * sizeof(unsigned long))
1610
1611 /*
1612  * Returns 1 if this HPT entry has been modified or has pending
1613  * R/C bit changes.
1614  */
1615 static int hpte_dirty(struct revmap_entry *revp, __be64 *hptp)
1616 {
1617         unsigned long rcbits_unset;
1618
1619         if (revp->guest_rpte & HPTE_GR_MODIFIED)
1620                 return 1;
1621
1622         /* Also need to consider changes in reference and changed bits */
1623         rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1624         if ((be64_to_cpu(hptp[0]) & HPTE_V_VALID) &&
1625             (be64_to_cpu(hptp[1]) & rcbits_unset))
1626                 return 1;
1627
1628         return 0;
1629 }
1630
1631 static long record_hpte(unsigned long flags, __be64 *hptp,
1632                         unsigned long *hpte, struct revmap_entry *revp,
1633                         int want_valid, int first_pass)
1634 {
1635         unsigned long v, r, hr;
1636         unsigned long rcbits_unset;
1637         int ok = 1;
1638         int valid, dirty;
1639
1640         /* Unmodified entries are uninteresting except on the first pass */
1641         dirty = hpte_dirty(revp, hptp);
1642         if (!first_pass && !dirty)
1643                 return 0;
1644
1645         valid = 0;
1646         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1647                 valid = 1;
1648                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) &&
1649                     !(be64_to_cpu(hptp[0]) & HPTE_V_BOLTED))
1650                         valid = 0;
1651         }
1652         if (valid != want_valid)
1653                 return 0;
1654
1655         v = r = 0;
1656         if (valid || dirty) {
1657                 /* lock the HPTE so it's stable and read it */
1658                 preempt_disable();
1659                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
1660                         cpu_relax();
1661                 v = be64_to_cpu(hptp[0]);
1662                 hr = be64_to_cpu(hptp[1]);
1663                 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1664                         v = hpte_new_to_old_v(v, hr);
1665                         hr = hpte_new_to_old_r(hr);
1666                 }
1667
1668                 /* re-evaluate valid and dirty from synchronized HPTE value */
1669                 valid = !!(v & HPTE_V_VALID);
1670                 dirty = !!(revp->guest_rpte & HPTE_GR_MODIFIED);
1671
1672                 /* Harvest R and C into guest view if necessary */
1673                 rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1674                 if (valid && (rcbits_unset & hr)) {
1675                         revp->guest_rpte |= (hr &
1676                                 (HPTE_R_R | HPTE_R_C)) | HPTE_GR_MODIFIED;
1677                         dirty = 1;
1678                 }
1679
1680                 if (v & HPTE_V_ABSENT) {
1681                         v &= ~HPTE_V_ABSENT;
1682                         v |= HPTE_V_VALID;
1683                         valid = 1;
1684                 }
1685                 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) && !(v & HPTE_V_BOLTED))
1686                         valid = 0;
1687
1688                 r = revp->guest_rpte;
1689                 /* only clear modified if this is the right sort of entry */
1690                 if (valid == want_valid && dirty) {
1691                         r &= ~HPTE_GR_MODIFIED;
1692                         revp->guest_rpte = r;
1693                 }
1694                 unlock_hpte(hptp, be64_to_cpu(hptp[0]));
1695                 preempt_enable();
1696                 if (!(valid == want_valid && (first_pass || dirty)))
1697                         ok = 0;
1698         }
1699         hpte[0] = cpu_to_be64(v);
1700         hpte[1] = cpu_to_be64(r);
1701         return ok;
1702 }
1703
1704 static ssize_t kvm_htab_read(struct file *file, char __user *buf,
1705                              size_t count, loff_t *ppos)
1706 {
1707         struct kvm_htab_ctx *ctx = file->private_data;
1708         struct kvm *kvm = ctx->kvm;
1709         struct kvm_get_htab_header hdr;
1710         __be64 *hptp;
1711         struct revmap_entry *revp;
1712         unsigned long i, nb, nw;
1713         unsigned long __user *lbuf;
1714         struct kvm_get_htab_header __user *hptr;
1715         unsigned long flags;
1716         int first_pass;
1717         unsigned long hpte[2];
1718
1719         if (!access_ok(VERIFY_WRITE, buf, count))
1720                 return -EFAULT;
1721         if (kvm_is_radix(kvm))
1722                 return 0;
1723
1724         first_pass = ctx->first_pass;
1725         flags = ctx->flags;
1726
1727         i = ctx->index;
1728         hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
1729         revp = kvm->arch.hpt.rev + i;
1730         lbuf = (unsigned long __user *)buf;
1731
1732         nb = 0;
1733         while (nb + sizeof(hdr) + HPTE_SIZE < count) {
1734                 /* Initialize header */
1735                 hptr = (struct kvm_get_htab_header __user *)buf;
1736                 hdr.n_valid = 0;
1737                 hdr.n_invalid = 0;
1738                 nw = nb;
1739                 nb += sizeof(hdr);
1740                 lbuf = (unsigned long __user *)(buf + sizeof(hdr));
1741
1742                 /* Skip uninteresting entries, i.e. clean on not-first pass */
1743                 if (!first_pass) {
1744                         while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1745                                !hpte_dirty(revp, hptp)) {
1746                                 ++i;
1747                                 hptp += 2;
1748                                 ++revp;
1749                         }
1750                 }
1751                 hdr.index = i;
1752
1753                 /* Grab a series of valid entries */
1754                 while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1755                        hdr.n_valid < 0xffff &&
1756                        nb + HPTE_SIZE < count &&
1757                        record_hpte(flags, hptp, hpte, revp, 1, first_pass)) {
1758                         /* valid entry, write it out */
1759                         ++hdr.n_valid;
1760                         if (__put_user(hpte[0], lbuf) ||
1761                             __put_user(hpte[1], lbuf + 1))
1762                                 return -EFAULT;
1763                         nb += HPTE_SIZE;
1764                         lbuf += 2;
1765                         ++i;
1766                         hptp += 2;
1767                         ++revp;
1768                 }
1769                 /* Now skip invalid entries while we can */
1770                 while (i < kvmppc_hpt_npte(&kvm->arch.hpt) &&
1771                        hdr.n_invalid < 0xffff &&
1772                        record_hpte(flags, hptp, hpte, revp, 0, first_pass)) {
1773                         /* found an invalid entry */
1774                         ++hdr.n_invalid;
1775                         ++i;
1776                         hptp += 2;
1777                         ++revp;
1778                 }
1779
1780                 if (hdr.n_valid || hdr.n_invalid) {
1781                         /* write back the header */
1782                         if (__copy_to_user(hptr, &hdr, sizeof(hdr)))
1783                                 return -EFAULT;
1784                         nw = nb;
1785                         buf = (char __user *)lbuf;
1786                 } else {
1787                         nb = nw;
1788                 }
1789
1790                 /* Check if we've wrapped around the hash table */
1791                 if (i >= kvmppc_hpt_npte(&kvm->arch.hpt)) {
1792                         i = 0;
1793                         ctx->first_pass = 0;
1794                         break;
1795                 }
1796         }
1797
1798         ctx->index = i;
1799
1800         return nb;
1801 }
1802
1803 static ssize_t kvm_htab_write(struct file *file, const char __user *buf,
1804                               size_t count, loff_t *ppos)
1805 {
1806         struct kvm_htab_ctx *ctx = file->private_data;
1807         struct kvm *kvm = ctx->kvm;
1808         struct kvm_get_htab_header hdr;
1809         unsigned long i, j;
1810         unsigned long v, r;
1811         unsigned long __user *lbuf;
1812         __be64 *hptp;
1813         unsigned long tmp[2];
1814         ssize_t nb;
1815         long int err, ret;
1816         int mmu_ready;
1817         int pshift;
1818
1819         if (!access_ok(VERIFY_READ, buf, count))
1820                 return -EFAULT;
1821         if (kvm_is_radix(kvm))
1822                 return -EINVAL;
1823
1824         /* lock out vcpus from running while we're doing this */
1825         mutex_lock(&kvm->lock);
1826         mmu_ready = kvm->arch.mmu_ready;
1827         if (mmu_ready) {
1828                 kvm->arch.mmu_ready = 0;        /* temporarily */
1829                 /* order mmu_ready vs. vcpus_running */
1830                 smp_mb();
1831                 if (atomic_read(&kvm->arch.vcpus_running)) {
1832                         kvm->arch.mmu_ready = 1;
1833                         mutex_unlock(&kvm->lock);
1834                         return -EBUSY;
1835                 }
1836         }
1837
1838         err = 0;
1839         for (nb = 0; nb + sizeof(hdr) <= count; ) {
1840                 err = -EFAULT;
1841                 if (__copy_from_user(&hdr, buf, sizeof(hdr)))
1842                         break;
1843
1844                 err = 0;
1845                 if (nb + hdr.n_valid * HPTE_SIZE > count)
1846                         break;
1847
1848                 nb += sizeof(hdr);
1849                 buf += sizeof(hdr);
1850
1851                 err = -EINVAL;
1852                 i = hdr.index;
1853                 if (i >= kvmppc_hpt_npte(&kvm->arch.hpt) ||
1854                     i + hdr.n_valid + hdr.n_invalid > kvmppc_hpt_npte(&kvm->arch.hpt))
1855                         break;
1856
1857                 hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
1858                 lbuf = (unsigned long __user *)buf;
1859                 for (j = 0; j < hdr.n_valid; ++j) {
1860                         __be64 hpte_v;
1861                         __be64 hpte_r;
1862
1863                         err = -EFAULT;
1864                         if (__get_user(hpte_v, lbuf) ||
1865                             __get_user(hpte_r, lbuf + 1))
1866                                 goto out;
1867                         v = be64_to_cpu(hpte_v);
1868                         r = be64_to_cpu(hpte_r);
1869                         err = -EINVAL;
1870                         if (!(v & HPTE_V_VALID))
1871                                 goto out;
1872                         pshift = kvmppc_hpte_base_page_shift(v, r);
1873                         if (pshift <= 0)
1874                                 goto out;
1875                         lbuf += 2;
1876                         nb += HPTE_SIZE;
1877
1878                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1879                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1880                         err = -EIO;
1881                         ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, i, v, r,
1882                                                          tmp);
1883                         if (ret != H_SUCCESS) {
1884                                 pr_err("kvm_htab_write ret %ld i=%ld v=%lx "
1885                                        "r=%lx\n", ret, i, v, r);
1886                                 goto out;
1887                         }
1888                         if (!mmu_ready && is_vrma_hpte(v)) {
1889                                 unsigned long senc, lpcr;
1890
1891                                 senc = slb_pgsize_encoding(1ul << pshift);
1892                                 kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
1893                                         (VRMA_VSID << SLB_VSID_SHIFT_1T);
1894                                 if (!cpu_has_feature(CPU_FTR_ARCH_300)) {
1895                                         lpcr = senc << (LPCR_VRMASD_SH - 4);
1896                                         kvmppc_update_lpcr(kvm, lpcr,
1897                                                            LPCR_VRMASD);
1898                                 } else {
1899                                         kvmppc_setup_partition_table(kvm);
1900                                 }
1901                                 mmu_ready = 1;
1902                         }
1903                         ++i;
1904                         hptp += 2;
1905                 }
1906
1907                 for (j = 0; j < hdr.n_invalid; ++j) {
1908                         if (be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT))
1909                                 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1910                         ++i;
1911                         hptp += 2;
1912                 }
1913                 err = 0;
1914         }
1915
1916  out:
1917         /* Order HPTE updates vs. mmu_ready */
1918         smp_wmb();
1919         kvm->arch.mmu_ready = mmu_ready;
1920         mutex_unlock(&kvm->lock);
1921
1922         if (err)
1923                 return err;
1924         return nb;
1925 }
1926
1927 static int kvm_htab_release(struct inode *inode, struct file *filp)
1928 {
1929         struct kvm_htab_ctx *ctx = filp->private_data;
1930
1931         filp->private_data = NULL;
1932         if (!(ctx->flags & KVM_GET_HTAB_WRITE))
1933                 atomic_dec(&ctx->kvm->arch.hpte_mod_interest);
1934         kvm_put_kvm(ctx->kvm);
1935         kfree(ctx);
1936         return 0;
1937 }
1938
1939 static const struct file_operations kvm_htab_fops = {
1940         .read           = kvm_htab_read,
1941         .write          = kvm_htab_write,
1942         .llseek         = default_llseek,
1943         .release        = kvm_htab_release,
1944 };
1945
1946 int kvm_vm_ioctl_get_htab_fd(struct kvm *kvm, struct kvm_get_htab_fd *ghf)
1947 {
1948         int ret;
1949         struct kvm_htab_ctx *ctx;
1950         int rwflag;
1951
1952         /* reject flags we don't recognize */
1953         if (ghf->flags & ~(KVM_GET_HTAB_BOLTED_ONLY | KVM_GET_HTAB_WRITE))
1954                 return -EINVAL;
1955         ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1956         if (!ctx)
1957                 return -ENOMEM;
1958         kvm_get_kvm(kvm);
1959         ctx->kvm = kvm;
1960         ctx->index = ghf->start_index;
1961         ctx->flags = ghf->flags;
1962         ctx->first_pass = 1;
1963
1964         rwflag = (ghf->flags & KVM_GET_HTAB_WRITE) ? O_WRONLY : O_RDONLY;
1965         ret = anon_inode_getfd("kvm-htab", &kvm_htab_fops, ctx, rwflag | O_CLOEXEC);
1966         if (ret < 0) {
1967                 kfree(ctx);
1968                 kvm_put_kvm(kvm);
1969                 return ret;
1970         }
1971
1972         if (rwflag == O_RDONLY) {
1973                 mutex_lock(&kvm->slots_lock);
1974                 atomic_inc(&kvm->arch.hpte_mod_interest);
1975                 /* make sure kvmppc_do_h_enter etc. see the increment */
1976                 synchronize_srcu_expedited(&kvm->srcu);
1977                 mutex_unlock(&kvm->slots_lock);
1978         }
1979
1980         return ret;
1981 }
1982
1983 struct debugfs_htab_state {
1984         struct kvm      *kvm;
1985         struct mutex    mutex;
1986         unsigned long   hpt_index;
1987         int             chars_left;
1988         int             buf_index;
1989         char            buf[64];
1990 };
1991
1992 static int debugfs_htab_open(struct inode *inode, struct file *file)
1993 {
1994         struct kvm *kvm = inode->i_private;
1995         struct debugfs_htab_state *p;
1996
1997         p = kzalloc(sizeof(*p), GFP_KERNEL);
1998         if (!p)
1999                 return -ENOMEM;
2000
2001         kvm_get_kvm(kvm);
2002         p->kvm = kvm;
2003         mutex_init(&p->mutex);
2004         file->private_data = p;
2005
2006         return nonseekable_open(inode, file);
2007 }
2008
2009 static int debugfs_htab_release(struct inode *inode, struct file *file)
2010 {
2011         struct debugfs_htab_state *p = file->private_data;
2012
2013         kvm_put_kvm(p->kvm);
2014         kfree(p);
2015         return 0;
2016 }
2017
2018 static ssize_t debugfs_htab_read(struct file *file, char __user *buf,
2019                                  size_t len, loff_t *ppos)
2020 {
2021         struct debugfs_htab_state *p = file->private_data;
2022         ssize_t ret, r;
2023         unsigned long i, n;
2024         unsigned long v, hr, gr;
2025         struct kvm *kvm;
2026         __be64 *hptp;
2027
2028         kvm = p->kvm;
2029         if (kvm_is_radix(kvm))
2030                 return 0;
2031
2032         ret = mutex_lock_interruptible(&p->mutex);
2033         if (ret)
2034                 return ret;
2035
2036         if (p->chars_left) {
2037                 n = p->chars_left;
2038                 if (n > len)
2039                         n = len;
2040                 r = copy_to_user(buf, p->buf + p->buf_index, n);
2041                 n -= r;
2042                 p->chars_left -= n;
2043                 p->buf_index += n;
2044                 buf += n;
2045                 len -= n;
2046                 ret = n;
2047                 if (r) {
2048                         if (!n)
2049                                 ret = -EFAULT;
2050                         goto out;
2051                 }
2052         }
2053
2054         i = p->hpt_index;
2055         hptp = (__be64 *)(kvm->arch.hpt.virt + (i * HPTE_SIZE));
2056         for (; len != 0 && i < kvmppc_hpt_npte(&kvm->arch.hpt);
2057              ++i, hptp += 2) {
2058                 if (!(be64_to_cpu(hptp[0]) & (HPTE_V_VALID | HPTE_V_ABSENT)))
2059                         continue;
2060
2061                 /* lock the HPTE so it's stable and read it */
2062                 preempt_disable();
2063                 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
2064                         cpu_relax();
2065                 v = be64_to_cpu(hptp[0]) & ~HPTE_V_HVLOCK;
2066                 hr = be64_to_cpu(hptp[1]);
2067                 gr = kvm->arch.hpt.rev[i].guest_rpte;
2068                 unlock_hpte(hptp, v);
2069                 preempt_enable();
2070
2071                 if (!(v & (HPTE_V_VALID | HPTE_V_ABSENT)))
2072                         continue;
2073
2074                 n = scnprintf(p->buf, sizeof(p->buf),
2075                               "%6lx %.16lx %.16lx %.16lx\n",
2076                               i, v, hr, gr);
2077                 p->chars_left = n;
2078                 if (n > len)
2079                         n = len;
2080                 r = copy_to_user(buf, p->buf, n);
2081                 n -= r;
2082                 p->chars_left -= n;
2083                 p->buf_index = n;
2084                 buf += n;
2085                 len -= n;
2086                 ret += n;
2087                 if (r) {
2088                         if (!ret)
2089                                 ret = -EFAULT;
2090                         goto out;
2091                 }
2092         }
2093         p->hpt_index = i;
2094
2095  out:
2096         mutex_unlock(&p->mutex);
2097         return ret;
2098 }
2099
2100 static ssize_t debugfs_htab_write(struct file *file, const char __user *buf,
2101                            size_t len, loff_t *ppos)
2102 {
2103         return -EACCES;
2104 }
2105
2106 static const struct file_operations debugfs_htab_fops = {
2107         .owner   = THIS_MODULE,
2108         .open    = debugfs_htab_open,
2109         .release = debugfs_htab_release,
2110         .read    = debugfs_htab_read,
2111         .write   = debugfs_htab_write,
2112         .llseek  = generic_file_llseek,
2113 };
2114
2115 void kvmppc_mmu_debugfs_init(struct kvm *kvm)
2116 {
2117         kvm->arch.htab_dentry = debugfs_create_file("htab", 0400,
2118                                                     kvm->arch.debugfs_dir, kvm,
2119                                                     &debugfs_htab_fops);
2120 }
2121
2122 void kvmppc_mmu_book3s_hv_init(struct kvm_vcpu *vcpu)
2123 {
2124         struct kvmppc_mmu *mmu = &vcpu->arch.mmu;
2125
2126         vcpu->arch.slb_nr = 32;         /* POWER7/POWER8 */
2127
2128         mmu->xlate = kvmppc_mmu_book3s_64_hv_xlate;
2129         mmu->reset_msr = kvmppc_mmu_book3s_64_hv_reset_msr;
2130
2131         vcpu->arch.hflags |= BOOK3S_HFLAG_SLB;
2132 }