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.
6 * Copyright 2010-2011 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 #include <linux/types.h>
10 #include <linux/string.h>
11 #include <linux/kvm.h>
12 #include <linux/kvm_host.h>
13 #include <linux/hugetlb.h>
14 #include <linux/module.h>
15 #include <linux/log2.h>
17 #include <asm/tlbflush.h>
18 #include <asm/kvm_ppc.h>
19 #include <asm/kvm_book3s.h>
20 #include <asm/book3s/64/mmu-hash.h>
21 #include <asm/hvcall.h>
22 #include <asm/synch.h>
23 #include <asm/ppc-opcode.h>
25 /* Translate address of a vmalloc'd thing to a linear map address */
26 static void *real_vmalloc_addr(void *x)
28 unsigned long addr = (unsigned long) x;
31 * assume we don't have huge pages in vmalloc space...
32 * So don't worry about THP collapse/split. Called
33 * Only in realmode, hence won't need irq_save/restore.
35 p = __find_linux_pte_or_hugepte(swapper_pg_dir, addr, NULL, NULL);
36 if (!p || !pte_present(*p))
38 addr = (pte_pfn(*p) << PAGE_SHIFT) | (addr & ~PAGE_MASK);
42 /* Return 1 if we need to do a global tlbie, 0 if we can use tlbiel */
43 static int global_invalidates(struct kvm *kvm, unsigned long flags)
49 * If there is only one vcore, and it's currently running,
50 * as indicated by local_paca->kvm_hstate.kvm_vcpu being set,
51 * we can use tlbiel as long as we mark all other physical
52 * cores as potentially having stale TLB entries for this lpid.
53 * Otherwise, don't use tlbiel.
55 if (kvm->arch.online_vcores == 1 && local_paca->kvm_hstate.kvm_vcpu)
61 /* any other core might now have stale TLB entries... */
63 cpumask_setall(&kvm->arch.need_tlb_flush);
64 cpu = local_paca->kvm_hstate.kvm_vcore->pcpu;
66 * On POWER9, threads are independent but the TLB is shared,
67 * so use the bit for the first thread to represent the core.
69 if (cpu_has_feature(CPU_FTR_ARCH_300))
70 cpu = cpu_first_thread_sibling(cpu);
71 cpumask_clear_cpu(cpu, &kvm->arch.need_tlb_flush);
78 * Add this HPTE into the chain for the real page.
79 * Must be called with the chain locked; it unlocks the chain.
81 void kvmppc_add_revmap_chain(struct kvm *kvm, struct revmap_entry *rev,
82 unsigned long *rmap, long pte_index, int realmode)
84 struct revmap_entry *head, *tail;
87 if (*rmap & KVMPPC_RMAP_PRESENT) {
88 i = *rmap & KVMPPC_RMAP_INDEX;
89 head = &kvm->arch.hpt.rev[i];
91 head = real_vmalloc_addr(head);
92 tail = &kvm->arch.hpt.rev[head->back];
94 tail = real_vmalloc_addr(tail);
96 rev->back = head->back;
97 tail->forw = pte_index;
98 head->back = pte_index;
100 rev->forw = rev->back = pte_index;
101 *rmap = (*rmap & ~KVMPPC_RMAP_INDEX) |
102 pte_index | KVMPPC_RMAP_PRESENT;
106 EXPORT_SYMBOL_GPL(kvmppc_add_revmap_chain);
108 /* Update the changed page order field of an rmap entry */
109 void kvmppc_update_rmap_change(unsigned long *rmap, unsigned long psize)
115 order = ilog2(psize);
116 order <<= KVMPPC_RMAP_CHG_SHIFT;
117 if (order > (*rmap & KVMPPC_RMAP_CHG_ORDER))
118 *rmap = (*rmap & ~KVMPPC_RMAP_CHG_ORDER) | order;
120 EXPORT_SYMBOL_GPL(kvmppc_update_rmap_change);
122 /* Returns a pointer to the revmap entry for the page mapped by a HPTE */
123 static unsigned long *revmap_for_hpte(struct kvm *kvm, unsigned long hpte_v,
124 unsigned long hpte_gr)
126 struct kvm_memory_slot *memslot;
130 gfn = hpte_rpn(hpte_gr, hpte_page_size(hpte_v, hpte_gr));
131 memslot = __gfn_to_memslot(kvm_memslots_raw(kvm), gfn);
135 rmap = real_vmalloc_addr(&memslot->arch.rmap[gfn - memslot->base_gfn]);
139 /* Remove this HPTE from the chain for a real page */
140 static void remove_revmap_chain(struct kvm *kvm, long pte_index,
141 struct revmap_entry *rev,
142 unsigned long hpte_v, unsigned long hpte_r)
144 struct revmap_entry *next, *prev;
145 unsigned long ptel, head;
147 unsigned long rcbits;
149 rcbits = hpte_r & (HPTE_R_R | HPTE_R_C);
150 ptel = rev->guest_rpte |= rcbits;
151 rmap = revmap_for_hpte(kvm, hpte_v, ptel);
156 head = *rmap & KVMPPC_RMAP_INDEX;
157 next = real_vmalloc_addr(&kvm->arch.hpt.rev[rev->forw]);
158 prev = real_vmalloc_addr(&kvm->arch.hpt.rev[rev->back]);
159 next->back = rev->back;
160 prev->forw = rev->forw;
161 if (head == pte_index) {
163 if (head == pte_index)
164 *rmap &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
166 *rmap = (*rmap & ~KVMPPC_RMAP_INDEX) | head;
168 *rmap |= rcbits << KVMPPC_RMAP_RC_SHIFT;
169 if (rcbits & HPTE_R_C)
170 kvmppc_update_rmap_change(rmap, hpte_page_size(hpte_v, hpte_r));
174 long kvmppc_do_h_enter(struct kvm *kvm, unsigned long flags,
175 long pte_index, unsigned long pteh, unsigned long ptel,
176 pgd_t *pgdir, bool realmode, unsigned long *pte_idx_ret)
178 unsigned long i, pa, gpa, gfn, psize;
179 unsigned long slot_fn, hva;
181 struct revmap_entry *rev;
182 unsigned long g_ptel;
183 struct kvm_memory_slot *memslot;
184 unsigned hpage_shift;
188 unsigned int writing;
189 unsigned long mmu_seq;
190 unsigned long rcbits, irq_flags = 0;
192 if (kvm_is_radix(kvm))
194 psize = hpte_page_size(pteh, ptel);
197 writing = hpte_is_writable(ptel);
198 pteh &= ~(HPTE_V_HVLOCK | HPTE_V_ABSENT | HPTE_V_VALID);
199 ptel &= ~HPTE_GR_RESERVED;
202 /* used later to detect if we might have been invalidated */
203 mmu_seq = kvm->mmu_notifier_seq;
206 /* Find the memslot (if any) for this address */
207 gpa = (ptel & HPTE_R_RPN) & ~(psize - 1);
208 gfn = gpa >> PAGE_SHIFT;
209 memslot = __gfn_to_memslot(kvm_memslots_raw(kvm), gfn);
213 if (!(memslot && !(memslot->flags & KVM_MEMSLOT_INVALID))) {
214 /* Emulated MMIO - mark this with key=31 */
215 pteh |= HPTE_V_ABSENT;
216 ptel |= HPTE_R_KEY_HI | HPTE_R_KEY_LO;
220 /* Check if the requested page fits entirely in the memslot. */
221 if (!slot_is_aligned(memslot, psize))
223 slot_fn = gfn - memslot->base_gfn;
224 rmap = &memslot->arch.rmap[slot_fn];
226 /* Translate to host virtual address */
227 hva = __gfn_to_hva_memslot(memslot, gfn);
229 * If we had a page table table change after lookup, we would
230 * retry via mmu_notifier_retry.
233 ptep = __find_linux_pte_or_hugepte(pgdir, hva, NULL,
236 local_irq_save(irq_flags);
237 ptep = find_linux_pte_or_hugepte(pgdir, hva, NULL,
242 unsigned int host_pte_size;
245 host_pte_size = 1ul << hpage_shift;
247 host_pte_size = PAGE_SIZE;
249 * We should always find the guest page size
250 * to <= host page size, if host is using hugepage
252 if (host_pte_size < psize) {
254 local_irq_restore(flags);
257 pte = kvmppc_read_update_linux_pte(ptep, writing);
258 if (pte_present(pte) && !pte_protnone(pte)) {
259 if (writing && !pte_write(pte))
260 /* make the actual HPTE be read-only */
261 ptel = hpte_make_readonly(ptel);
263 pa = pte_pfn(pte) << PAGE_SHIFT;
264 pa |= hva & (host_pte_size - 1);
265 pa |= gpa & ~PAGE_MASK;
269 local_irq_restore(irq_flags);
271 ptel &= ~(HPTE_R_PP0 - psize);
275 pteh |= HPTE_V_VALID;
277 pteh |= HPTE_V_ABSENT;
278 ptel &= ~(HPTE_R_KEY_HI | HPTE_R_KEY_LO);
281 /*If we had host pte mapping then Check WIMG */
282 if (ptep && !hpte_cache_flags_ok(ptel, is_ci)) {
286 * Allow guest to map emulated device memory as
287 * uncacheable, but actually make it cacheable.
289 ptel &= ~(HPTE_R_W|HPTE_R_I|HPTE_R_G);
293 /* Find and lock the HPTEG slot to use */
295 if (pte_index >= kvmppc_hpt_npte(&kvm->arch.hpt))
297 if (likely((flags & H_EXACT) == 0)) {
299 hpte = (__be64 *)(kvm->arch.hpt.virt + (pte_index << 4));
300 for (i = 0; i < 8; ++i) {
301 if ((be64_to_cpu(*hpte) & HPTE_V_VALID) == 0 &&
302 try_lock_hpte(hpte, HPTE_V_HVLOCK | HPTE_V_VALID |
309 * Since try_lock_hpte doesn't retry (not even stdcx.
310 * failures), it could be that there is a free slot
311 * but we transiently failed to lock it. Try again,
312 * actually locking each slot and checking it.
315 for (i = 0; i < 8; ++i) {
317 while (!try_lock_hpte(hpte, HPTE_V_HVLOCK))
319 pte = be64_to_cpu(hpte[0]);
320 if (!(pte & (HPTE_V_VALID | HPTE_V_ABSENT)))
322 __unlock_hpte(hpte, pte);
330 hpte = (__be64 *)(kvm->arch.hpt.virt + (pte_index << 4));
331 if (!try_lock_hpte(hpte, HPTE_V_HVLOCK | HPTE_V_VALID |
333 /* Lock the slot and check again */
336 while (!try_lock_hpte(hpte, HPTE_V_HVLOCK))
338 pte = be64_to_cpu(hpte[0]);
339 if (pte & (HPTE_V_VALID | HPTE_V_ABSENT)) {
340 __unlock_hpte(hpte, pte);
346 /* Save away the guest's idea of the second HPTE dword */
347 rev = &kvm->arch.hpt.rev[pte_index];
349 rev = real_vmalloc_addr(rev);
351 rev->guest_rpte = g_ptel;
352 note_hpte_modification(kvm, rev);
355 /* Link HPTE into reverse-map chain */
356 if (pteh & HPTE_V_VALID) {
358 rmap = real_vmalloc_addr(rmap);
360 /* Check for pending invalidations under the rmap chain lock */
361 if (mmu_notifier_retry(kvm, mmu_seq)) {
362 /* inval in progress, write a non-present HPTE */
363 pteh |= HPTE_V_ABSENT;
364 pteh &= ~HPTE_V_VALID;
365 ptel &= ~(HPTE_R_KEY_HI | HPTE_R_KEY_LO);
368 kvmppc_add_revmap_chain(kvm, rev, rmap, pte_index,
370 /* Only set R/C in real HPTE if already set in *rmap */
371 rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
372 ptel &= rcbits | ~(HPTE_R_R | HPTE_R_C);
376 /* Convert to new format on P9 */
377 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
378 ptel = hpte_old_to_new_r(pteh, ptel);
379 pteh = hpte_old_to_new_v(pteh);
381 hpte[1] = cpu_to_be64(ptel);
383 /* Write the first HPTE dword, unlocking the HPTE and making it valid */
385 __unlock_hpte(hpte, pteh);
386 asm volatile("ptesync" : : : "memory");
388 *pte_idx_ret = pte_index;
391 EXPORT_SYMBOL_GPL(kvmppc_do_h_enter);
393 long kvmppc_h_enter(struct kvm_vcpu *vcpu, unsigned long flags,
394 long pte_index, unsigned long pteh, unsigned long ptel)
396 return kvmppc_do_h_enter(vcpu->kvm, flags, pte_index, pteh, ptel,
397 vcpu->arch.pgdir, true, &vcpu->arch.gpr[4]);
400 #ifdef __BIG_ENDIAN__
401 #define LOCK_TOKEN (*(u32 *)(&get_paca()->lock_token))
403 #define LOCK_TOKEN (*(u32 *)(&get_paca()->paca_index))
406 static inline int is_mmio_hpte(unsigned long v, unsigned long r)
408 return ((v & HPTE_V_ABSENT) &&
409 (r & (HPTE_R_KEY_HI | HPTE_R_KEY_LO)) ==
410 (HPTE_R_KEY_HI | HPTE_R_KEY_LO));
413 static inline int try_lock_tlbie(unsigned int *lock)
415 unsigned int tmp, old;
416 unsigned int token = LOCK_TOKEN;
418 asm volatile("1:lwarx %1,0,%2\n"
425 : "=&r" (tmp), "=&r" (old)
426 : "r" (lock), "r" (token)
431 static void do_tlbies(struct kvm *kvm, unsigned long *rbvalues,
432 long npages, int global, bool need_sync)
437 * We use the POWER9 5-operand versions of tlbie and tlbiel here.
438 * Since we are using RIC=0 PRS=0 R=0, and P7/P8 tlbiel ignores
439 * the RS field, this is backwards-compatible with P7 and P8.
442 while (!try_lock_tlbie(&kvm->arch.tlbie_lock))
445 asm volatile("ptesync" : : : "memory");
446 for (i = 0; i < npages; ++i)
447 asm volatile(PPC_TLBIE_5(%0,%1,0,0,0) : :
448 "r" (rbvalues[i]), "r" (kvm->arch.lpid));
449 asm volatile("eieio; tlbsync; ptesync" : : : "memory");
450 kvm->arch.tlbie_lock = 0;
453 asm volatile("ptesync" : : : "memory");
454 for (i = 0; i < npages; ++i)
455 asm volatile(PPC_TLBIEL(%0,%1,0,0,0) : :
456 "r" (rbvalues[i]), "r" (0));
457 asm volatile("ptesync" : : : "memory");
461 long kvmppc_do_h_remove(struct kvm *kvm, unsigned long flags,
462 unsigned long pte_index, unsigned long avpn,
463 unsigned long *hpret)
466 unsigned long v, r, rb;
467 struct revmap_entry *rev;
468 u64 pte, orig_pte, pte_r;
470 if (kvm_is_radix(kvm))
472 if (pte_index >= kvmppc_hpt_npte(&kvm->arch.hpt))
474 hpte = (__be64 *)(kvm->arch.hpt.virt + (pte_index << 4));
475 while (!try_lock_hpte(hpte, HPTE_V_HVLOCK))
477 pte = orig_pte = be64_to_cpu(hpte[0]);
478 pte_r = be64_to_cpu(hpte[1]);
479 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
480 pte = hpte_new_to_old_v(pte, pte_r);
481 pte_r = hpte_new_to_old_r(pte_r);
483 if ((pte & (HPTE_V_ABSENT | HPTE_V_VALID)) == 0 ||
484 ((flags & H_AVPN) && (pte & ~0x7fUL) != avpn) ||
485 ((flags & H_ANDCOND) && (pte & avpn) != 0)) {
486 __unlock_hpte(hpte, orig_pte);
490 rev = real_vmalloc_addr(&kvm->arch.hpt.rev[pte_index]);
491 v = pte & ~HPTE_V_HVLOCK;
492 if (v & HPTE_V_VALID) {
493 hpte[0] &= ~cpu_to_be64(HPTE_V_VALID);
494 rb = compute_tlbie_rb(v, pte_r, pte_index);
495 do_tlbies(kvm, &rb, 1, global_invalidates(kvm, flags), true);
497 * The reference (R) and change (C) bits in a HPT
498 * entry can be set by hardware at any time up until
499 * the HPTE is invalidated and the TLB invalidation
500 * sequence has completed. This means that when
501 * removing a HPTE, we need to re-read the HPTE after
502 * the invalidation sequence has completed in order to
503 * obtain reliable values of R and C.
505 remove_revmap_chain(kvm, pte_index, rev, v,
506 be64_to_cpu(hpte[1]));
508 r = rev->guest_rpte & ~HPTE_GR_RESERVED;
509 note_hpte_modification(kvm, rev);
510 unlock_hpte(hpte, 0);
512 if (is_mmio_hpte(v, pte_r))
513 atomic64_inc(&kvm->arch.mmio_update);
515 if (v & HPTE_V_ABSENT)
516 v = (v & ~HPTE_V_ABSENT) | HPTE_V_VALID;
521 EXPORT_SYMBOL_GPL(kvmppc_do_h_remove);
523 long kvmppc_h_remove(struct kvm_vcpu *vcpu, unsigned long flags,
524 unsigned long pte_index, unsigned long avpn)
526 return kvmppc_do_h_remove(vcpu->kvm, flags, pte_index, avpn,
530 long kvmppc_h_bulk_remove(struct kvm_vcpu *vcpu)
532 struct kvm *kvm = vcpu->kvm;
533 unsigned long *args = &vcpu->arch.gpr[4];
534 __be64 *hp, *hptes[4];
535 unsigned long tlbrb[4];
536 long int i, j, k, n, found, indexes[4];
537 unsigned long flags, req, pte_index, rcbits;
539 long int ret = H_SUCCESS;
540 struct revmap_entry *rev, *revs[4];
543 if (kvm_is_radix(kvm))
545 global = global_invalidates(kvm, 0);
546 for (i = 0; i < 4 && ret == H_SUCCESS; ) {
551 flags = pte_index >> 56;
552 pte_index &= ((1ul << 56) - 1);
555 if (req == 3) { /* no more requests */
559 if (req != 1 || flags == 3 ||
560 pte_index >= kvmppc_hpt_npte(&kvm->arch.hpt)) {
561 /* parameter error */
562 args[j] = ((0xa0 | flags) << 56) + pte_index;
566 hp = (__be64 *) (kvm->arch.hpt.virt + (pte_index << 4));
567 /* to avoid deadlock, don't spin except for first */
568 if (!try_lock_hpte(hp, HPTE_V_HVLOCK)) {
571 while (!try_lock_hpte(hp, HPTE_V_HVLOCK))
575 hp0 = be64_to_cpu(hp[0]);
576 hp1 = be64_to_cpu(hp[1]);
577 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
578 hp0 = hpte_new_to_old_v(hp0, hp1);
579 hp1 = hpte_new_to_old_r(hp1);
581 if (hp0 & (HPTE_V_ABSENT | HPTE_V_VALID)) {
583 case 0: /* absolute */
586 case 1: /* andcond */
587 if (!(hp0 & args[j + 1]))
591 if ((hp0 & ~0x7fUL) == args[j + 1])
597 hp[0] &= ~cpu_to_be64(HPTE_V_HVLOCK);
598 args[j] = ((0x90 | flags) << 56) + pte_index;
602 args[j] = ((0x80 | flags) << 56) + pte_index;
603 rev = real_vmalloc_addr(&kvm->arch.hpt.rev[pte_index]);
604 note_hpte_modification(kvm, rev);
606 if (!(hp0 & HPTE_V_VALID)) {
607 /* insert R and C bits from PTE */
608 rcbits = rev->guest_rpte & (HPTE_R_R|HPTE_R_C);
609 args[j] |= rcbits << (56 - 5);
611 if (is_mmio_hpte(hp0, hp1))
612 atomic64_inc(&kvm->arch.mmio_update);
616 /* leave it locked */
617 hp[0] &= ~cpu_to_be64(HPTE_V_VALID);
618 tlbrb[n] = compute_tlbie_rb(hp0, hp1, pte_index);
628 /* Now that we've collected a batch, do the tlbies */
629 do_tlbies(kvm, tlbrb, n, global, true);
631 /* Read PTE low words after tlbie to get final R/C values */
632 for (k = 0; k < n; ++k) {
634 pte_index = args[j] & ((1ul << 56) - 1);
637 remove_revmap_chain(kvm, pte_index, rev,
638 be64_to_cpu(hp[0]), be64_to_cpu(hp[1]));
639 rcbits = rev->guest_rpte & (HPTE_R_R|HPTE_R_C);
640 args[j] |= rcbits << (56 - 5);
641 __unlock_hpte(hp, 0);
648 long kvmppc_h_protect(struct kvm_vcpu *vcpu, unsigned long flags,
649 unsigned long pte_index, unsigned long avpn,
652 struct kvm *kvm = vcpu->kvm;
654 struct revmap_entry *rev;
655 unsigned long v, r, rb, mask, bits;
658 if (kvm_is_radix(kvm))
660 if (pte_index >= kvmppc_hpt_npte(&kvm->arch.hpt))
663 hpte = (__be64 *)(kvm->arch.hpt.virt + (pte_index << 4));
664 while (!try_lock_hpte(hpte, HPTE_V_HVLOCK))
666 v = pte_v = be64_to_cpu(hpte[0]);
667 if (cpu_has_feature(CPU_FTR_ARCH_300))
668 v = hpte_new_to_old_v(v, be64_to_cpu(hpte[1]));
669 if ((v & (HPTE_V_ABSENT | HPTE_V_VALID)) == 0 ||
670 ((flags & H_AVPN) && (v & ~0x7fUL) != avpn)) {
671 __unlock_hpte(hpte, pte_v);
675 pte_r = be64_to_cpu(hpte[1]);
676 bits = (flags << 55) & HPTE_R_PP0;
677 bits |= (flags << 48) & HPTE_R_KEY_HI;
678 bits |= flags & (HPTE_R_PP | HPTE_R_N | HPTE_R_KEY_LO);
680 /* Update guest view of 2nd HPTE dword */
681 mask = HPTE_R_PP0 | HPTE_R_PP | HPTE_R_N |
682 HPTE_R_KEY_HI | HPTE_R_KEY_LO;
683 rev = real_vmalloc_addr(&kvm->arch.hpt.rev[pte_index]);
685 r = (rev->guest_rpte & ~mask) | bits;
687 note_hpte_modification(kvm, rev);
691 if (v & HPTE_V_VALID) {
693 * If the page is valid, don't let it transition from
694 * readonly to writable. If it should be writable, we'll
695 * take a trap and let the page fault code sort it out.
697 r = (pte_r & ~mask) | bits;
698 if (hpte_is_writable(r) && !hpte_is_writable(pte_r))
699 r = hpte_make_readonly(r);
700 /* If the PTE is changing, invalidate it first */
702 rb = compute_tlbie_rb(v, r, pte_index);
703 hpte[0] = cpu_to_be64((pte_v & ~HPTE_V_VALID) |
705 do_tlbies(kvm, &rb, 1, global_invalidates(kvm, flags),
707 /* Don't lose R/C bit updates done by hardware */
708 r |= be64_to_cpu(hpte[1]) & (HPTE_R_R | HPTE_R_C);
709 hpte[1] = cpu_to_be64(r);
712 unlock_hpte(hpte, pte_v & ~HPTE_V_HVLOCK);
713 asm volatile("ptesync" : : : "memory");
714 if (is_mmio_hpte(v, pte_r))
715 atomic64_inc(&kvm->arch.mmio_update);
720 long kvmppc_h_read(struct kvm_vcpu *vcpu, unsigned long flags,
721 unsigned long pte_index)
723 struct kvm *kvm = vcpu->kvm;
727 struct revmap_entry *rev = NULL;
729 if (kvm_is_radix(kvm))
731 if (pte_index >= kvmppc_hpt_npte(&kvm->arch.hpt))
733 if (flags & H_READ_4) {
737 rev = real_vmalloc_addr(&kvm->arch.hpt.rev[pte_index]);
738 for (i = 0; i < n; ++i, ++pte_index) {
739 hpte = (__be64 *)(kvm->arch.hpt.virt + (pte_index << 4));
740 v = be64_to_cpu(hpte[0]) & ~HPTE_V_HVLOCK;
741 r = be64_to_cpu(hpte[1]);
742 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
743 v = hpte_new_to_old_v(v, r);
744 r = hpte_new_to_old_r(r);
746 if (v & HPTE_V_ABSENT) {
750 if (v & HPTE_V_VALID) {
751 r = rev[i].guest_rpte | (r & (HPTE_R_R | HPTE_R_C));
752 r &= ~HPTE_GR_RESERVED;
754 vcpu->arch.gpr[4 + i * 2] = v;
755 vcpu->arch.gpr[5 + i * 2] = r;
760 long kvmppc_h_clear_ref(struct kvm_vcpu *vcpu, unsigned long flags,
761 unsigned long pte_index)
763 struct kvm *kvm = vcpu->kvm;
765 unsigned long v, r, gr;
766 struct revmap_entry *rev;
768 long ret = H_NOT_FOUND;
770 if (kvm_is_radix(kvm))
772 if (pte_index >= kvmppc_hpt_npte(&kvm->arch.hpt))
775 rev = real_vmalloc_addr(&kvm->arch.hpt.rev[pte_index]);
776 hpte = (__be64 *)(kvm->arch.hpt.virt + (pte_index << 4));
777 while (!try_lock_hpte(hpte, HPTE_V_HVLOCK))
779 v = be64_to_cpu(hpte[0]);
780 r = be64_to_cpu(hpte[1]);
781 if (!(v & (HPTE_V_VALID | HPTE_V_ABSENT)))
784 gr = rev->guest_rpte;
785 if (rev->guest_rpte & HPTE_R_R) {
786 rev->guest_rpte &= ~HPTE_R_R;
787 note_hpte_modification(kvm, rev);
789 if (v & HPTE_V_VALID) {
790 gr |= r & (HPTE_R_R | HPTE_R_C);
792 kvmppc_clear_ref_hpte(kvm, hpte, pte_index);
793 rmap = revmap_for_hpte(kvm, v, gr);
796 *rmap |= KVMPPC_RMAP_REFERENCED;
801 vcpu->arch.gpr[4] = gr;
804 unlock_hpte(hpte, v & ~HPTE_V_HVLOCK);
808 long kvmppc_h_clear_mod(struct kvm_vcpu *vcpu, unsigned long flags,
809 unsigned long pte_index)
811 struct kvm *kvm = vcpu->kvm;
813 unsigned long v, r, gr;
814 struct revmap_entry *rev;
816 long ret = H_NOT_FOUND;
818 if (kvm_is_radix(kvm))
820 if (pte_index >= kvmppc_hpt_npte(&kvm->arch.hpt))
823 rev = real_vmalloc_addr(&kvm->arch.hpt.rev[pte_index]);
824 hpte = (__be64 *)(kvm->arch.hpt.virt + (pte_index << 4));
825 while (!try_lock_hpte(hpte, HPTE_V_HVLOCK))
827 v = be64_to_cpu(hpte[0]);
828 r = be64_to_cpu(hpte[1]);
829 if (!(v & (HPTE_V_VALID | HPTE_V_ABSENT)))
832 gr = rev->guest_rpte;
834 rev->guest_rpte &= ~HPTE_R_C;
835 note_hpte_modification(kvm, rev);
837 if (v & HPTE_V_VALID) {
838 /* need to make it temporarily absent so C is stable */
839 hpte[0] |= cpu_to_be64(HPTE_V_ABSENT);
840 kvmppc_invalidate_hpte(kvm, hpte, pte_index);
841 r = be64_to_cpu(hpte[1]);
842 gr |= r & (HPTE_R_R | HPTE_R_C);
844 unsigned long psize = hpte_page_size(v, r);
845 hpte[1] = cpu_to_be64(r & ~HPTE_R_C);
847 rmap = revmap_for_hpte(kvm, v, gr);
850 *rmap |= KVMPPC_RMAP_CHANGED;
851 kvmppc_update_rmap_change(rmap, psize);
856 vcpu->arch.gpr[4] = gr;
859 unlock_hpte(hpte, v & ~HPTE_V_HVLOCK);
863 void kvmppc_invalidate_hpte(struct kvm *kvm, __be64 *hptep,
864 unsigned long pte_index)
869 hptep[0] &= ~cpu_to_be64(HPTE_V_VALID);
870 hp0 = be64_to_cpu(hptep[0]);
871 hp1 = be64_to_cpu(hptep[1]);
872 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
873 hp0 = hpte_new_to_old_v(hp0, hp1);
874 hp1 = hpte_new_to_old_r(hp1);
876 rb = compute_tlbie_rb(hp0, hp1, pte_index);
877 do_tlbies(kvm, &rb, 1, 1, true);
879 EXPORT_SYMBOL_GPL(kvmppc_invalidate_hpte);
881 void kvmppc_clear_ref_hpte(struct kvm *kvm, __be64 *hptep,
882 unsigned long pte_index)
888 hp0 = be64_to_cpu(hptep[0]);
889 hp1 = be64_to_cpu(hptep[1]);
890 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
891 hp0 = hpte_new_to_old_v(hp0, hp1);
892 hp1 = hpte_new_to_old_r(hp1);
894 rb = compute_tlbie_rb(hp0, hp1, pte_index);
895 rbyte = (be64_to_cpu(hptep[1]) & ~HPTE_R_R) >> 8;
896 /* modify only the second-last byte, which contains the ref bit */
897 *((char *)hptep + 14) = rbyte;
898 do_tlbies(kvm, &rb, 1, 1, false);
900 EXPORT_SYMBOL_GPL(kvmppc_clear_ref_hpte);
902 static int slb_base_page_shift[4] = {
906 20, /* 1M, unsupported */
909 static struct mmio_hpte_cache_entry *mmio_cache_search(struct kvm_vcpu *vcpu,
910 unsigned long eaddr, unsigned long slb_v, long mmio_update)
912 struct mmio_hpte_cache_entry *entry = NULL;
916 for (i = 0; i < MMIO_HPTE_CACHE_SIZE; i++) {
917 entry = &vcpu->arch.mmio_cache.entry[i];
918 if (entry->mmio_update == mmio_update) {
919 pshift = entry->slb_base_pshift;
920 if ((entry->eaddr >> pshift) == (eaddr >> pshift) &&
921 entry->slb_v == slb_v)
928 static struct mmio_hpte_cache_entry *
929 next_mmio_cache_entry(struct kvm_vcpu *vcpu)
931 unsigned int index = vcpu->arch.mmio_cache.index;
933 vcpu->arch.mmio_cache.index++;
934 if (vcpu->arch.mmio_cache.index == MMIO_HPTE_CACHE_SIZE)
935 vcpu->arch.mmio_cache.index = 0;
937 return &vcpu->arch.mmio_cache.entry[index];
940 /* When called from virtmode, this func should be protected by
941 * preempt_disable(), otherwise, the holding of HPTE_V_HVLOCK
942 * can trigger deadlock issue.
944 long kvmppc_hv_find_lock_hpte(struct kvm *kvm, gva_t eaddr, unsigned long slb_v,
949 unsigned long somask;
950 unsigned long vsid, hash;
953 unsigned long mask, val;
954 unsigned long v, r, orig_v;
956 /* Get page shift, work out hash and AVPN etc. */
957 mask = SLB_VSID_B | HPTE_V_AVPN | HPTE_V_SECONDARY;
960 if (slb_v & SLB_VSID_L) {
961 mask |= HPTE_V_LARGE;
963 pshift = slb_base_page_shift[(slb_v & SLB_VSID_LP) >> 4];
965 if (slb_v & SLB_VSID_B_1T) {
966 somask = (1UL << 40) - 1;
967 vsid = (slb_v & ~SLB_VSID_B) >> SLB_VSID_SHIFT_1T;
970 somask = (1UL << 28) - 1;
971 vsid = (slb_v & ~SLB_VSID_B) >> SLB_VSID_SHIFT;
973 hash = (vsid ^ ((eaddr & somask) >> pshift)) & kvmppc_hpt_mask(&kvm->arch.hpt);
974 avpn = slb_v & ~(somask >> 16); /* also includes B */
975 avpn |= (eaddr & somask) >> 16;
978 avpn &= ~((1UL << (pshift - 16)) - 1);
984 hpte = (__be64 *)(kvm->arch.hpt.virt + (hash << 7));
986 for (i = 0; i < 16; i += 2) {
987 /* Read the PTE racily */
988 v = be64_to_cpu(hpte[i]) & ~HPTE_V_HVLOCK;
989 if (cpu_has_feature(CPU_FTR_ARCH_300))
990 v = hpte_new_to_old_v(v, be64_to_cpu(hpte[i+1]));
992 /* Check valid/absent, hash, segment size and AVPN */
993 if (!(v & valid) || (v & mask) != val)
996 /* Lock the PTE and read it under the lock */
997 while (!try_lock_hpte(&hpte[i], HPTE_V_HVLOCK))
999 v = orig_v = be64_to_cpu(hpte[i]) & ~HPTE_V_HVLOCK;
1000 r = be64_to_cpu(hpte[i+1]);
1001 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1002 v = hpte_new_to_old_v(v, r);
1003 r = hpte_new_to_old_r(r);
1007 * Check the HPTE again, including base page size
1009 if ((v & valid) && (v & mask) == val &&
1010 hpte_base_page_size(v, r) == (1ul << pshift))
1011 /* Return with the HPTE still locked */
1012 return (hash << 3) + (i >> 1);
1014 __unlock_hpte(&hpte[i], orig_v);
1017 if (val & HPTE_V_SECONDARY)
1019 val |= HPTE_V_SECONDARY;
1020 hash = hash ^ kvmppc_hpt_mask(&kvm->arch.hpt);
1024 EXPORT_SYMBOL(kvmppc_hv_find_lock_hpte);
1027 * Called in real mode to check whether an HPTE not found fault
1028 * is due to accessing a paged-out page or an emulated MMIO page,
1029 * or if a protection fault is due to accessing a page that the
1030 * guest wanted read/write access to but which we made read-only.
1031 * Returns a possibly modified status (DSISR) value if not
1032 * (i.e. pass the interrupt to the guest),
1033 * -1 to pass the fault up to host kernel mode code, -2 to do that
1034 * and also load the instruction word (for MMIO emulation),
1035 * or 0 if we should make the guest retry the access.
1037 long kvmppc_hpte_hv_fault(struct kvm_vcpu *vcpu, unsigned long addr,
1038 unsigned long slb_v, unsigned int status, bool data)
1040 struct kvm *kvm = vcpu->kvm;
1042 unsigned long v, r, gr, orig_v;
1044 unsigned long valid;
1045 struct revmap_entry *rev;
1046 unsigned long pp, key;
1047 struct mmio_hpte_cache_entry *cache_entry = NULL;
1048 long mmio_update = 0;
1050 /* For protection fault, expect to find a valid HPTE */
1051 valid = HPTE_V_VALID;
1052 if (status & DSISR_NOHPTE) {
1053 valid |= HPTE_V_ABSENT;
1054 mmio_update = atomic64_read(&kvm->arch.mmio_update);
1055 cache_entry = mmio_cache_search(vcpu, addr, slb_v, mmio_update);
1058 index = cache_entry->pte_index;
1059 v = cache_entry->hpte_v;
1060 r = cache_entry->hpte_r;
1061 gr = cache_entry->rpte;
1063 index = kvmppc_hv_find_lock_hpte(kvm, addr, slb_v, valid);
1065 if (status & DSISR_NOHPTE)
1066 return status; /* there really was no HPTE */
1067 return 0; /* for prot fault, HPTE disappeared */
1069 hpte = (__be64 *)(kvm->arch.hpt.virt + (index << 4));
1070 v = orig_v = be64_to_cpu(hpte[0]) & ~HPTE_V_HVLOCK;
1071 r = be64_to_cpu(hpte[1]);
1072 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
1073 v = hpte_new_to_old_v(v, r);
1074 r = hpte_new_to_old_r(r);
1076 rev = real_vmalloc_addr(&kvm->arch.hpt.rev[index]);
1077 gr = rev->guest_rpte;
1079 unlock_hpte(hpte, orig_v);
1082 /* For not found, if the HPTE is valid by now, retry the instruction */
1083 if ((status & DSISR_NOHPTE) && (v & HPTE_V_VALID))
1086 /* Check access permissions to the page */
1087 pp = gr & (HPTE_R_PP0 | HPTE_R_PP);
1088 key = (vcpu->arch.shregs.msr & MSR_PR) ? SLB_VSID_KP : SLB_VSID_KS;
1089 status &= ~DSISR_NOHPTE; /* DSISR_NOHPTE == SRR1_ISI_NOPT */
1091 if (gr & (HPTE_R_N | HPTE_R_G))
1092 return status | SRR1_ISI_N_OR_G;
1093 if (!hpte_read_permission(pp, slb_v & key))
1094 return status | SRR1_ISI_PROT;
1095 } else if (status & DSISR_ISSTORE) {
1096 /* check write permission */
1097 if (!hpte_write_permission(pp, slb_v & key))
1098 return status | DSISR_PROTFAULT;
1100 if (!hpte_read_permission(pp, slb_v & key))
1101 return status | DSISR_PROTFAULT;
1104 /* Check storage key, if applicable */
1105 if (data && (vcpu->arch.shregs.msr & MSR_DR)) {
1106 unsigned int perm = hpte_get_skey_perm(gr, vcpu->arch.amr);
1107 if (status & DSISR_ISSTORE)
1110 return status | DSISR_KEYFAULT;
1113 /* Save HPTE info for virtual-mode handler */
1114 vcpu->arch.pgfault_addr = addr;
1115 vcpu->arch.pgfault_index = index;
1116 vcpu->arch.pgfault_hpte[0] = v;
1117 vcpu->arch.pgfault_hpte[1] = r;
1118 vcpu->arch.pgfault_cache = cache_entry;
1120 /* Check the storage key to see if it is possibly emulated MMIO */
1121 if ((r & (HPTE_R_KEY_HI | HPTE_R_KEY_LO)) ==
1122 (HPTE_R_KEY_HI | HPTE_R_KEY_LO)) {
1124 unsigned int pshift = 12;
1125 unsigned int pshift_index;
1127 if (slb_v & SLB_VSID_L) {
1128 pshift_index = ((slb_v & SLB_VSID_LP) >> 4);
1129 pshift = slb_base_page_shift[pshift_index];
1131 cache_entry = next_mmio_cache_entry(vcpu);
1132 cache_entry->eaddr = addr;
1133 cache_entry->slb_base_pshift = pshift;
1134 cache_entry->pte_index = index;
1135 cache_entry->hpte_v = v;
1136 cache_entry->hpte_r = r;
1137 cache_entry->rpte = gr;
1138 cache_entry->slb_v = slb_v;
1139 cache_entry->mmio_update = mmio_update;
1141 if (data && (vcpu->arch.shregs.msr & MSR_IR))
1142 return -2; /* MMIO emulation - load instr word */
1145 return -1; /* send fault up to host kernel mode */