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 * 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.
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.
15 * Copyright 2010 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
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>
31 #include <asm/tlbflush.h>
32 #include <asm/kvm_ppc.h>
33 #include <asm/kvm_book3s.h>
34 #include <asm/mmu-hash64.h>
35 #include <asm/hvcall.h>
36 #include <asm/synch.h>
37 #include <asm/ppc-opcode.h>
38 #include <asm/cputable.h>
40 /* POWER7 has 10-bit LPIDs, PPC970 has 6-bit LPIDs */
41 #define MAX_LPID_970 63
43 /* Power architecture requires HPT is at least 256kB */
44 #define PPC_MIN_HPT_ORDER 18
46 static long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
47 long pte_index, unsigned long pteh,
48 unsigned long ptel, unsigned long *pte_idx_ret);
49 static void kvmppc_rmap_reset(struct kvm *kvm);
51 long kvmppc_alloc_hpt(struct kvm *kvm, u32 *htab_orderp)
54 struct revmap_entry *rev;
55 struct kvmppc_linear_info *li;
56 long order = kvm_hpt_order;
60 if (order < PPC_MIN_HPT_ORDER)
61 order = PPC_MIN_HPT_ORDER;
65 * If the user wants a different size from default,
66 * try first to allocate it from the kernel page allocator.
69 if (order != kvm_hpt_order) {
70 hpt = __get_free_pages(GFP_KERNEL|__GFP_ZERO|__GFP_REPEAT|
71 __GFP_NOWARN, order - PAGE_SHIFT);
76 /* Next try to allocate from the preallocated pool */
80 hpt = (ulong)li->base_virt;
81 kvm->arch.hpt_li = li;
82 order = kvm_hpt_order;
86 /* Lastly try successively smaller sizes from the page allocator */
87 while (!hpt && order > PPC_MIN_HPT_ORDER) {
88 hpt = __get_free_pages(GFP_KERNEL|__GFP_ZERO|__GFP_REPEAT|
89 __GFP_NOWARN, order - PAGE_SHIFT);
97 kvm->arch.hpt_virt = hpt;
98 kvm->arch.hpt_order = order;
99 /* HPTEs are 2**4 bytes long */
100 kvm->arch.hpt_npte = 1ul << (order - 4);
101 /* 128 (2**7) bytes in each HPTEG */
102 kvm->arch.hpt_mask = (1ul << (order - 7)) - 1;
104 /* Allocate reverse map array */
105 rev = vmalloc(sizeof(struct revmap_entry) * kvm->arch.hpt_npte);
107 pr_err("kvmppc_alloc_hpt: Couldn't alloc reverse map array\n");
110 kvm->arch.revmap = rev;
111 kvm->arch.sdr1 = __pa(hpt) | (order - 18);
113 pr_info("KVM guest htab at %lx (order %ld), LPID %x\n",
114 hpt, order, kvm->arch.lpid);
117 *htab_orderp = order;
121 if (kvm->arch.hpt_li)
122 kvm_release_hpt(kvm->arch.hpt_li);
124 free_pages(hpt, order - PAGE_SHIFT);
128 long kvmppc_alloc_reset_hpt(struct kvm *kvm, u32 *htab_orderp)
133 mutex_lock(&kvm->lock);
134 if (kvm->arch.rma_setup_done) {
135 kvm->arch.rma_setup_done = 0;
136 /* order rma_setup_done vs. vcpus_running */
138 if (atomic_read(&kvm->arch.vcpus_running)) {
139 kvm->arch.rma_setup_done = 1;
143 if (kvm->arch.hpt_virt) {
144 order = kvm->arch.hpt_order;
145 /* Set the entire HPT to 0, i.e. invalid HPTEs */
146 memset((void *)kvm->arch.hpt_virt, 0, 1ul << order);
148 * Reset all the reverse-mapping chains for all memslots
150 kvmppc_rmap_reset(kvm);
151 /* Ensure that each vcpu will flush its TLB on next entry. */
152 cpumask_setall(&kvm->arch.need_tlb_flush);
153 *htab_orderp = order;
156 err = kvmppc_alloc_hpt(kvm, htab_orderp);
157 order = *htab_orderp;
160 mutex_unlock(&kvm->lock);
164 void kvmppc_free_hpt(struct kvm *kvm)
166 kvmppc_free_lpid(kvm->arch.lpid);
167 vfree(kvm->arch.revmap);
168 if (kvm->arch.hpt_li)
169 kvm_release_hpt(kvm->arch.hpt_li);
171 free_pages(kvm->arch.hpt_virt,
172 kvm->arch.hpt_order - PAGE_SHIFT);
175 /* Bits in first HPTE dword for pagesize 4k, 64k or 16M */
176 static inline unsigned long hpte0_pgsize_encoding(unsigned long pgsize)
178 return (pgsize > 0x1000) ? HPTE_V_LARGE : 0;
181 /* Bits in second HPTE dword for pagesize 4k, 64k or 16M */
182 static inline unsigned long hpte1_pgsize_encoding(unsigned long pgsize)
184 return (pgsize == 0x10000) ? 0x1000 : 0;
187 void kvmppc_map_vrma(struct kvm_vcpu *vcpu, struct kvm_memory_slot *memslot,
188 unsigned long porder)
191 unsigned long npages;
192 unsigned long hp_v, hp_r;
193 unsigned long addr, hash;
195 unsigned long hp0, hp1;
196 unsigned long idx_ret;
198 struct kvm *kvm = vcpu->kvm;
200 psize = 1ul << porder;
201 npages = memslot->npages >> (porder - PAGE_SHIFT);
203 /* VRMA can't be > 1TB */
204 if (npages > 1ul << (40 - porder))
205 npages = 1ul << (40 - porder);
206 /* Can't use more than 1 HPTE per HPTEG */
207 if (npages > kvm->arch.hpt_mask + 1)
208 npages = kvm->arch.hpt_mask + 1;
210 hp0 = HPTE_V_1TB_SEG | (VRMA_VSID << (40 - 16)) |
211 HPTE_V_BOLTED | hpte0_pgsize_encoding(psize);
212 hp1 = hpte1_pgsize_encoding(psize) |
213 HPTE_R_R | HPTE_R_C | HPTE_R_M | PP_RWXX;
215 for (i = 0; i < npages; ++i) {
217 /* can't use hpt_hash since va > 64 bits */
218 hash = (i ^ (VRMA_VSID ^ (VRMA_VSID << 25))) & kvm->arch.hpt_mask;
220 * We assume that the hash table is empty and no
221 * vcpus are using it at this stage. Since we create
222 * at most one HPTE per HPTEG, we just assume entry 7
223 * is available and use it.
225 hash = (hash << 3) + 7;
226 hp_v = hp0 | ((addr >> 16) & ~0x7fUL);
228 ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, hash, hp_v, hp_r,
230 if (ret != H_SUCCESS) {
231 pr_err("KVM: map_vrma at %lx failed, ret=%ld\n",
238 int kvmppc_mmu_hv_init(void)
240 unsigned long host_lpid, rsvd_lpid;
242 if (!cpu_has_feature(CPU_FTR_HVMODE))
245 /* POWER7 has 10-bit LPIDs, PPC970 and e500mc have 6-bit LPIDs */
246 if (cpu_has_feature(CPU_FTR_ARCH_206)) {
247 host_lpid = mfspr(SPRN_LPID); /* POWER7 */
248 rsvd_lpid = LPID_RSVD;
250 host_lpid = 0; /* PPC970 */
251 rsvd_lpid = MAX_LPID_970;
254 kvmppc_init_lpid(rsvd_lpid + 1);
256 kvmppc_claim_lpid(host_lpid);
257 /* rsvd_lpid is reserved for use in partition switching */
258 kvmppc_claim_lpid(rsvd_lpid);
263 void kvmppc_mmu_destroy(struct kvm_vcpu *vcpu)
267 static void kvmppc_mmu_book3s_64_hv_reset_msr(struct kvm_vcpu *vcpu)
269 kvmppc_set_msr(vcpu, MSR_SF | MSR_ME);
273 * This is called to get a reference to a guest page if there isn't
274 * one already in the memslot->arch.slot_phys[] array.
276 static long kvmppc_get_guest_page(struct kvm *kvm, unsigned long gfn,
277 struct kvm_memory_slot *memslot,
282 struct page *page, *hpage, *pages[1];
283 unsigned long s, pgsize;
284 unsigned long *physp;
285 unsigned int is_io, got, pgorder;
286 struct vm_area_struct *vma;
287 unsigned long pfn, i, npages;
289 physp = memslot->arch.slot_phys;
292 if (physp[gfn - memslot->base_gfn])
300 start = gfn_to_hva_memslot(memslot, gfn);
302 /* Instantiate and get the page we want access to */
303 np = get_user_pages_fast(start, 1, 1, pages);
305 /* Look up the vma for the page */
306 down_read(¤t->mm->mmap_sem);
307 vma = find_vma(current->mm, start);
308 if (!vma || vma->vm_start > start ||
309 start + psize > vma->vm_end ||
310 !(vma->vm_flags & VM_PFNMAP))
312 is_io = hpte_cache_bits(pgprot_val(vma->vm_page_prot));
313 pfn = vma->vm_pgoff + ((start - vma->vm_start) >> PAGE_SHIFT);
314 /* check alignment of pfn vs. requested page size */
315 if (psize > PAGE_SIZE && (pfn & ((psize >> PAGE_SHIFT) - 1)))
317 up_read(¤t->mm->mmap_sem);
321 got = KVMPPC_GOT_PAGE;
323 /* See if this is a large page */
325 if (PageHuge(page)) {
326 hpage = compound_head(page);
327 s <<= compound_order(hpage);
328 /* Get the whole large page if slot alignment is ok */
329 if (s > psize && slot_is_aligned(memslot, s) &&
330 !(memslot->userspace_addr & (s - 1))) {
340 pfn = page_to_pfn(page);
343 npages = pgsize >> PAGE_SHIFT;
344 pgorder = __ilog2(npages);
345 physp += (gfn - memslot->base_gfn) & ~(npages - 1);
346 spin_lock(&kvm->arch.slot_phys_lock);
347 for (i = 0; i < npages; ++i) {
349 physp[i] = ((pfn + i) << PAGE_SHIFT) +
350 got + is_io + pgorder;
354 spin_unlock(&kvm->arch.slot_phys_lock);
363 up_read(¤t->mm->mmap_sem);
367 long kvmppc_virtmode_do_h_enter(struct kvm *kvm, unsigned long flags,
368 long pte_index, unsigned long pteh,
369 unsigned long ptel, unsigned long *pte_idx_ret)
371 unsigned long psize, gpa, gfn;
372 struct kvm_memory_slot *memslot;
375 if (kvm->arch.using_mmu_notifiers)
378 psize = hpte_page_size(pteh, ptel);
382 pteh &= ~(HPTE_V_HVLOCK | HPTE_V_ABSENT | HPTE_V_VALID);
384 /* Find the memslot (if any) for this address */
385 gpa = (ptel & HPTE_R_RPN) & ~(psize - 1);
386 gfn = gpa >> PAGE_SHIFT;
387 memslot = gfn_to_memslot(kvm, gfn);
388 if (memslot && !(memslot->flags & KVM_MEMSLOT_INVALID)) {
389 if (!slot_is_aligned(memslot, psize))
391 if (kvmppc_get_guest_page(kvm, gfn, memslot, psize) < 0)
396 /* Protect linux PTE lookup from page table destruction */
397 rcu_read_lock_sched(); /* this disables preemption too */
398 ret = kvmppc_do_h_enter(kvm, flags, pte_index, pteh, ptel,
399 current->mm->pgd, false, pte_idx_ret);
400 rcu_read_unlock_sched();
401 if (ret == H_TOO_HARD) {
402 /* this can't happen */
403 pr_err("KVM: Oops, kvmppc_h_enter returned too hard!\n");
404 ret = H_RESOURCE; /* or something */
411 * We come here on a H_ENTER call from the guest when we are not
412 * using mmu notifiers and we don't have the requested page pinned
415 long kvmppc_virtmode_h_enter(struct kvm_vcpu *vcpu, unsigned long flags,
416 long pte_index, unsigned long pteh,
419 return kvmppc_virtmode_do_h_enter(vcpu->kvm, flags, pte_index,
420 pteh, ptel, &vcpu->arch.gpr[4]);
423 static struct kvmppc_slb *kvmppc_mmu_book3s_hv_find_slbe(struct kvm_vcpu *vcpu,
429 for (i = 0; i < vcpu->arch.slb_nr; i++) {
430 if (!(vcpu->arch.slb[i].orige & SLB_ESID_V))
433 if (vcpu->arch.slb[i].origv & SLB_VSID_B_1T)
438 if (((vcpu->arch.slb[i].orige ^ eaddr) & mask) == 0)
439 return &vcpu->arch.slb[i];
444 static unsigned long kvmppc_mmu_get_real_addr(unsigned long v, unsigned long r,
447 unsigned long ra_mask;
449 ra_mask = hpte_page_size(v, r) - 1;
450 return (r & HPTE_R_RPN & ~ra_mask) | (ea & ra_mask);
453 static int kvmppc_mmu_book3s_64_hv_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
454 struct kvmppc_pte *gpte, bool data)
456 struct kvm *kvm = vcpu->kvm;
457 struct kvmppc_slb *slbe;
459 unsigned long pp, key;
461 unsigned long *hptep;
463 int virtmode = vcpu->arch.shregs.msr & (data ? MSR_DR : MSR_IR);
467 slbe = kvmppc_mmu_book3s_hv_find_slbe(vcpu, eaddr);
472 /* real mode access */
473 slb_v = vcpu->kvm->arch.vrma_slb_v;
477 /* Find the HPTE in the hash table */
478 index = kvmppc_hv_find_lock_hpte(kvm, eaddr, slb_v,
479 HPTE_V_VALID | HPTE_V_ABSENT);
484 hptep = (unsigned long *)(kvm->arch.hpt_virt + (index << 4));
485 v = hptep[0] & ~HPTE_V_HVLOCK;
486 gr = kvm->arch.revmap[index].guest_rpte;
488 /* Unlock the HPTE */
489 asm volatile("lwsync" : : : "memory");
494 gpte->vpage = ((v & HPTE_V_AVPN) << 4) | ((eaddr >> 12) & 0xfff);
496 /* Get PP bits and key for permission check */
497 pp = gr & (HPTE_R_PP0 | HPTE_R_PP);
498 key = (vcpu->arch.shregs.msr & MSR_PR) ? SLB_VSID_KP : SLB_VSID_KS;
501 /* Calculate permissions */
502 gpte->may_read = hpte_read_permission(pp, key);
503 gpte->may_write = hpte_write_permission(pp, key);
504 gpte->may_execute = gpte->may_read && !(gr & (HPTE_R_N | HPTE_R_G));
506 /* Storage key permission check for POWER7 */
507 if (data && virtmode && cpu_has_feature(CPU_FTR_ARCH_206)) {
508 int amrfield = hpte_get_skey_perm(gr, vcpu->arch.amr);
515 /* Get the guest physical address */
516 gpte->raddr = kvmppc_mmu_get_real_addr(v, gr, eaddr);
521 * Quick test for whether an instruction is a load or a store.
522 * If the instruction is a load or a store, then this will indicate
523 * which it is, at least on server processors. (Embedded processors
524 * have some external PID instructions that don't follow the rule
525 * embodied here.) If the instruction isn't a load or store, then
526 * this doesn't return anything useful.
528 static int instruction_is_store(unsigned int instr)
533 if ((instr & 0xfc000000) == 0x7c000000)
534 mask = 0x100; /* major opcode 31 */
535 return (instr & mask) != 0;
538 static int kvmppc_hv_emulate_mmio(struct kvm_run *run, struct kvm_vcpu *vcpu,
539 unsigned long gpa, gva_t ea, int is_store)
543 unsigned long srr0 = kvmppc_get_pc(vcpu);
545 /* We try to load the last instruction. We don't let
546 * emulate_instruction do it as it doesn't check what
548 * If we fail, we just return to the guest and try executing it again.
550 if (vcpu->arch.last_inst == KVM_INST_FETCH_FAILED) {
551 ret = kvmppc_ld(vcpu, &srr0, sizeof(u32), &last_inst, false);
552 if (ret != EMULATE_DONE || last_inst == KVM_INST_FETCH_FAILED)
554 vcpu->arch.last_inst = last_inst;
558 * WARNING: We do not know for sure whether the instruction we just
559 * read from memory is the same that caused the fault in the first
560 * place. If the instruction we read is neither an load or a store,
561 * then it can't access memory, so we don't need to worry about
562 * enforcing access permissions. So, assuming it is a load or
563 * store, we just check that its direction (load or store) is
564 * consistent with the original fault, since that's what we
565 * checked the access permissions against. If there is a mismatch
566 * we just return and retry the instruction.
569 if (instruction_is_store(vcpu->arch.last_inst) != !!is_store)
573 * Emulated accesses are emulated by looking at the hash for
574 * translation once, then performing the access later. The
575 * translation could be invalidated in the meantime in which
576 * point performing the subsequent memory access on the old
577 * physical address could possibly be a security hole for the
578 * guest (but not the host).
580 * This is less of an issue for MMIO stores since they aren't
581 * globally visible. It could be an issue for MMIO loads to
582 * a certain extent but we'll ignore it for now.
585 vcpu->arch.paddr_accessed = gpa;
586 vcpu->arch.vaddr_accessed = ea;
587 return kvmppc_emulate_mmio(run, vcpu);
590 int kvmppc_book3s_hv_page_fault(struct kvm_run *run, struct kvm_vcpu *vcpu,
591 unsigned long ea, unsigned long dsisr)
593 struct kvm *kvm = vcpu->kvm;
594 unsigned long *hptep, hpte[3], r;
595 unsigned long mmu_seq, psize, pte_size;
596 unsigned long gpa, gfn, hva, pfn;
597 struct kvm_memory_slot *memslot;
599 struct revmap_entry *rev;
600 struct page *page, *pages[1];
601 long index, ret, npages;
603 unsigned int writing, write_ok;
604 struct vm_area_struct *vma;
605 unsigned long rcbits;
608 * Real-mode code has already searched the HPT and found the
609 * entry we're interested in. Lock the entry and check that
610 * it hasn't changed. If it has, just return and re-execute the
613 if (ea != vcpu->arch.pgfault_addr)
615 index = vcpu->arch.pgfault_index;
616 hptep = (unsigned long *)(kvm->arch.hpt_virt + (index << 4));
617 rev = &kvm->arch.revmap[index];
619 while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
621 hpte[0] = hptep[0] & ~HPTE_V_HVLOCK;
623 hpte[2] = r = rev->guest_rpte;
624 asm volatile("lwsync" : : : "memory");
628 if (hpte[0] != vcpu->arch.pgfault_hpte[0] ||
629 hpte[1] != vcpu->arch.pgfault_hpte[1])
632 /* Translate the logical address and get the page */
633 psize = hpte_page_size(hpte[0], r);
634 gpa = (r & HPTE_R_RPN & ~(psize - 1)) | (ea & (psize - 1));
635 gfn = gpa >> PAGE_SHIFT;
636 memslot = gfn_to_memslot(kvm, gfn);
638 /* No memslot means it's an emulated MMIO region */
639 if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
640 return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
641 dsisr & DSISR_ISSTORE);
643 if (!kvm->arch.using_mmu_notifiers)
644 return -EFAULT; /* should never get here */
646 /* used to check for invalidations in progress */
647 mmu_seq = kvm->mmu_notifier_seq;
653 pte_size = PAGE_SIZE;
654 writing = (dsisr & DSISR_ISSTORE) != 0;
655 /* If writing != 0, then the HPTE must allow writing, if we get here */
657 hva = gfn_to_hva_memslot(memslot, gfn);
658 npages = get_user_pages_fast(hva, 1, writing, pages);
660 /* Check if it's an I/O mapping */
661 down_read(¤t->mm->mmap_sem);
662 vma = find_vma(current->mm, hva);
663 if (vma && vma->vm_start <= hva && hva + psize <= vma->vm_end &&
664 (vma->vm_flags & VM_PFNMAP)) {
665 pfn = vma->vm_pgoff +
666 ((hva - vma->vm_start) >> PAGE_SHIFT);
668 is_io = hpte_cache_bits(pgprot_val(vma->vm_page_prot));
669 write_ok = vma->vm_flags & VM_WRITE;
671 up_read(¤t->mm->mmap_sem);
676 if (PageHuge(page)) {
677 page = compound_head(page);
678 pte_size <<= compound_order(page);
680 /* if the guest wants write access, see if that is OK */
681 if (!writing && hpte_is_writable(r)) {
685 * We need to protect against page table destruction
686 * while looking up and updating the pte.
688 rcu_read_lock_sched();
689 ptep = find_linux_pte_or_hugepte(current->mm->pgd,
691 if (ptep && pte_present(*ptep)) {
692 pte = kvmppc_read_update_linux_pte(ptep, 1);
696 rcu_read_unlock_sched();
698 pfn = page_to_pfn(page);
702 if (psize > pte_size)
705 /* Check WIMG vs. the actual page we're accessing */
706 if (!hpte_cache_flags_ok(r, is_io)) {
710 * Allow guest to map emulated device memory as
711 * uncacheable, but actually make it cacheable.
713 r = (r & ~(HPTE_R_W|HPTE_R_I|HPTE_R_G)) | HPTE_R_M;
716 /* Set the HPTE to point to pfn */
717 r = (r & ~(HPTE_R_PP0 - pte_size)) | (pfn << PAGE_SHIFT);
718 if (hpte_is_writable(r) && !write_ok)
719 r = hpte_make_readonly(r);
722 while (!try_lock_hpte(hptep, HPTE_V_HVLOCK))
724 if ((hptep[0] & ~HPTE_V_HVLOCK) != hpte[0] || hptep[1] != hpte[1] ||
725 rev->guest_rpte != hpte[2])
726 /* HPTE has been changed under us; let the guest retry */
728 hpte[0] = (hpte[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
730 rmap = &memslot->arch.rmap[gfn - memslot->base_gfn];
733 /* Check if we might have been invalidated; let the guest retry if so */
735 if (mmu_notifier_retry(vcpu->kvm, mmu_seq)) {
740 /* Only set R/C in real HPTE if set in both *rmap and guest_rpte */
741 rcbits = *rmap >> KVMPPC_RMAP_RC_SHIFT;
742 r &= rcbits | ~(HPTE_R_R | HPTE_R_C);
744 if (hptep[0] & HPTE_V_VALID) {
745 /* HPTE was previously valid, so we need to invalidate it */
747 hptep[0] |= HPTE_V_ABSENT;
748 kvmppc_invalidate_hpte(kvm, hptep, index);
749 /* don't lose previous R and C bits */
750 r |= hptep[1] & (HPTE_R_R | HPTE_R_C);
752 kvmppc_add_revmap_chain(kvm, rev, rmap, index, 0);
758 asm volatile("ptesync" : : : "memory");
760 if (page && hpte_is_writable(r))
766 * We drop pages[0] here, not page because page might
767 * have been set to the head page of a compound, but
768 * we have to drop the reference on the correct tail
769 * page to match the get inside gup()
776 hptep[0] &= ~HPTE_V_HVLOCK;
781 static void kvmppc_rmap_reset(struct kvm *kvm)
783 struct kvm_memslots *slots;
784 struct kvm_memory_slot *memslot;
787 srcu_idx = srcu_read_lock(&kvm->srcu);
788 slots = kvm->memslots;
789 kvm_for_each_memslot(memslot, slots) {
791 * This assumes it is acceptable to lose reference and
792 * change bits across a reset.
794 memset(memslot->arch.rmap, 0,
795 memslot->npages * sizeof(*memslot->arch.rmap));
797 srcu_read_unlock(&kvm->srcu, srcu_idx);
800 static int kvm_handle_hva_range(struct kvm *kvm,
803 int (*handler)(struct kvm *kvm,
804 unsigned long *rmapp,
809 struct kvm_memslots *slots;
810 struct kvm_memory_slot *memslot;
812 slots = kvm_memslots(kvm);
813 kvm_for_each_memslot(memslot, slots) {
814 unsigned long hva_start, hva_end;
817 hva_start = max(start, memslot->userspace_addr);
818 hva_end = min(end, memslot->userspace_addr +
819 (memslot->npages << PAGE_SHIFT));
820 if (hva_start >= hva_end)
823 * {gfn(page) | page intersects with [hva_start, hva_end)} =
824 * {gfn, gfn+1, ..., gfn_end-1}.
826 gfn = hva_to_gfn_memslot(hva_start, memslot);
827 gfn_end = hva_to_gfn_memslot(hva_end + PAGE_SIZE - 1, memslot);
829 for (; gfn < gfn_end; ++gfn) {
830 gfn_t gfn_offset = gfn - memslot->base_gfn;
832 ret = handler(kvm, &memslot->arch.rmap[gfn_offset], gfn);
840 static int kvm_handle_hva(struct kvm *kvm, unsigned long hva,
841 int (*handler)(struct kvm *kvm, unsigned long *rmapp,
844 return kvm_handle_hva_range(kvm, hva, hva + 1, handler);
847 static int kvm_unmap_rmapp(struct kvm *kvm, unsigned long *rmapp,
850 struct revmap_entry *rev = kvm->arch.revmap;
851 unsigned long h, i, j;
852 unsigned long *hptep;
853 unsigned long ptel, psize, rcbits;
857 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
863 * To avoid an ABBA deadlock with the HPTE lock bit,
864 * we can't spin on the HPTE lock while holding the
867 i = *rmapp & KVMPPC_RMAP_INDEX;
868 hptep = (unsigned long *) (kvm->arch.hpt_virt + (i << 4));
869 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
870 /* unlock rmap before spinning on the HPTE lock */
872 while (hptep[0] & HPTE_V_HVLOCK)
878 /* chain is now empty */
879 *rmapp &= ~(KVMPPC_RMAP_PRESENT | KVMPPC_RMAP_INDEX);
881 /* remove i from chain */
885 rev[i].forw = rev[i].back = i;
886 *rmapp = (*rmapp & ~KVMPPC_RMAP_INDEX) | j;
889 /* Now check and modify the HPTE */
890 ptel = rev[i].guest_rpte;
891 psize = hpte_page_size(hptep[0], ptel);
892 if ((hptep[0] & HPTE_V_VALID) &&
893 hpte_rpn(ptel, psize) == gfn) {
894 if (kvm->arch.using_mmu_notifiers)
895 hptep[0] |= HPTE_V_ABSENT;
896 kvmppc_invalidate_hpte(kvm, hptep, i);
897 /* Harvest R and C */
898 rcbits = hptep[1] & (HPTE_R_R | HPTE_R_C);
899 *rmapp |= rcbits << KVMPPC_RMAP_RC_SHIFT;
900 if (rcbits & ~rev[i].guest_rpte) {
901 rev[i].guest_rpte = ptel | rcbits;
902 note_hpte_modification(kvm, &rev[i]);
906 hptep[0] &= ~HPTE_V_HVLOCK;
911 int kvm_unmap_hva(struct kvm *kvm, unsigned long hva)
913 if (kvm->arch.using_mmu_notifiers)
914 kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
918 int kvm_unmap_hva_range(struct kvm *kvm, unsigned long start, unsigned long end)
920 if (kvm->arch.using_mmu_notifiers)
921 kvm_handle_hva_range(kvm, start, end, kvm_unmap_rmapp);
925 void kvmppc_core_flush_memslot(struct kvm *kvm, struct kvm_memory_slot *memslot)
927 unsigned long *rmapp;
931 rmapp = memslot->arch.rmap;
932 gfn = memslot->base_gfn;
933 for (n = memslot->npages; n; --n) {
935 * Testing the present bit without locking is OK because
936 * the memslot has been marked invalid already, and hence
937 * no new HPTEs referencing this page can be created,
938 * thus the present bit can't go from 0 to 1.
940 if (*rmapp & KVMPPC_RMAP_PRESENT)
941 kvm_unmap_rmapp(kvm, rmapp, gfn);
947 static int kvm_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
950 struct revmap_entry *rev = kvm->arch.revmap;
951 unsigned long head, i, j;
952 unsigned long *hptep;
957 if (*rmapp & KVMPPC_RMAP_REFERENCED) {
958 *rmapp &= ~KVMPPC_RMAP_REFERENCED;
961 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
966 i = head = *rmapp & KVMPPC_RMAP_INDEX;
968 hptep = (unsigned long *) (kvm->arch.hpt_virt + (i << 4));
971 /* If this HPTE isn't referenced, ignore it */
972 if (!(hptep[1] & HPTE_R_R))
975 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
976 /* unlock rmap before spinning on the HPTE lock */
978 while (hptep[0] & HPTE_V_HVLOCK)
983 /* Now check and modify the HPTE */
984 if ((hptep[0] & HPTE_V_VALID) && (hptep[1] & HPTE_R_R)) {
985 kvmppc_clear_ref_hpte(kvm, hptep, i);
986 if (!(rev[i].guest_rpte & HPTE_R_R)) {
987 rev[i].guest_rpte |= HPTE_R_R;
988 note_hpte_modification(kvm, &rev[i]);
992 hptep[0] &= ~HPTE_V_HVLOCK;
993 } while ((i = j) != head);
999 int kvm_age_hva(struct kvm *kvm, unsigned long hva)
1001 if (!kvm->arch.using_mmu_notifiers)
1003 return kvm_handle_hva(kvm, hva, kvm_age_rmapp);
1006 static int kvm_test_age_rmapp(struct kvm *kvm, unsigned long *rmapp,
1009 struct revmap_entry *rev = kvm->arch.revmap;
1010 unsigned long head, i, j;
1014 if (*rmapp & KVMPPC_RMAP_REFERENCED)
1018 if (*rmapp & KVMPPC_RMAP_REFERENCED)
1021 if (*rmapp & KVMPPC_RMAP_PRESENT) {
1022 i = head = *rmapp & KVMPPC_RMAP_INDEX;
1024 hp = (unsigned long *)(kvm->arch.hpt_virt + (i << 4));
1026 if (hp[1] & HPTE_R_R)
1028 } while ((i = j) != head);
1037 int kvm_test_age_hva(struct kvm *kvm, unsigned long hva)
1039 if (!kvm->arch.using_mmu_notifiers)
1041 return kvm_handle_hva(kvm, hva, kvm_test_age_rmapp);
1044 void kvm_set_spte_hva(struct kvm *kvm, unsigned long hva, pte_t pte)
1046 if (!kvm->arch.using_mmu_notifiers)
1048 kvm_handle_hva(kvm, hva, kvm_unmap_rmapp);
1051 static int kvm_test_clear_dirty(struct kvm *kvm, unsigned long *rmapp)
1053 struct revmap_entry *rev = kvm->arch.revmap;
1054 unsigned long head, i, j;
1055 unsigned long *hptep;
1060 if (*rmapp & KVMPPC_RMAP_CHANGED) {
1061 *rmapp &= ~KVMPPC_RMAP_CHANGED;
1064 if (!(*rmapp & KVMPPC_RMAP_PRESENT)) {
1069 i = head = *rmapp & KVMPPC_RMAP_INDEX;
1071 hptep = (unsigned long *) (kvm->arch.hpt_virt + (i << 4));
1074 if (!(hptep[1] & HPTE_R_C))
1077 if (!try_lock_hpte(hptep, HPTE_V_HVLOCK)) {
1078 /* unlock rmap before spinning on the HPTE lock */
1080 while (hptep[0] & HPTE_V_HVLOCK)
1085 /* Now check and modify the HPTE */
1086 if ((hptep[0] & HPTE_V_VALID) && (hptep[1] & HPTE_R_C)) {
1087 /* need to make it temporarily absent to clear C */
1088 hptep[0] |= HPTE_V_ABSENT;
1089 kvmppc_invalidate_hpte(kvm, hptep, i);
1090 hptep[1] &= ~HPTE_R_C;
1092 hptep[0] = (hptep[0] & ~HPTE_V_ABSENT) | HPTE_V_VALID;
1093 if (!(rev[i].guest_rpte & HPTE_R_C)) {
1094 rev[i].guest_rpte |= HPTE_R_C;
1095 note_hpte_modification(kvm, &rev[i]);
1099 hptep[0] &= ~HPTE_V_HVLOCK;
1100 } while ((i = j) != head);
1106 static void harvest_vpa_dirty(struct kvmppc_vpa *vpa,
1107 struct kvm_memory_slot *memslot,
1112 if (!vpa->dirty || !vpa->pinned_addr)
1114 gfn = vpa->gpa >> PAGE_SHIFT;
1115 if (gfn < memslot->base_gfn ||
1116 gfn >= memslot->base_gfn + memslot->npages)
1121 __set_bit_le(gfn - memslot->base_gfn, map);
1124 long kvmppc_hv_get_dirty_log(struct kvm *kvm, struct kvm_memory_slot *memslot,
1128 unsigned long *rmapp;
1129 struct kvm_vcpu *vcpu;
1132 rmapp = memslot->arch.rmap;
1133 for (i = 0; i < memslot->npages; ++i) {
1134 if (kvm_test_clear_dirty(kvm, rmapp) && map)
1135 __set_bit_le(i, map);
1139 /* Harvest dirty bits from VPA and DTL updates */
1140 /* Note: we never modify the SLB shadow buffer areas */
1141 kvm_for_each_vcpu(i, vcpu, kvm) {
1142 spin_lock(&vcpu->arch.vpa_update_lock);
1143 harvest_vpa_dirty(&vcpu->arch.vpa, memslot, map);
1144 harvest_vpa_dirty(&vcpu->arch.dtl, memslot, map);
1145 spin_unlock(&vcpu->arch.vpa_update_lock);
1151 void *kvmppc_pin_guest_page(struct kvm *kvm, unsigned long gpa,
1152 unsigned long *nb_ret)
1154 struct kvm_memory_slot *memslot;
1155 unsigned long gfn = gpa >> PAGE_SHIFT;
1156 struct page *page, *pages[1];
1158 unsigned long hva, offset;
1160 unsigned long *physp;
1163 srcu_idx = srcu_read_lock(&kvm->srcu);
1164 memslot = gfn_to_memslot(kvm, gfn);
1165 if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID))
1167 if (!kvm->arch.using_mmu_notifiers) {
1168 physp = memslot->arch.slot_phys;
1171 physp += gfn - memslot->base_gfn;
1174 if (kvmppc_get_guest_page(kvm, gfn, memslot,
1179 page = pfn_to_page(pa >> PAGE_SHIFT);
1182 hva = gfn_to_hva_memslot(memslot, gfn);
1183 npages = get_user_pages_fast(hva, 1, 1, pages);
1188 srcu_read_unlock(&kvm->srcu, srcu_idx);
1190 offset = gpa & (PAGE_SIZE - 1);
1192 *nb_ret = PAGE_SIZE - offset;
1193 return page_address(page) + offset;
1196 srcu_read_unlock(&kvm->srcu, srcu_idx);
1200 void kvmppc_unpin_guest_page(struct kvm *kvm, void *va, unsigned long gpa,
1203 struct page *page = virt_to_page(va);
1204 struct kvm_memory_slot *memslot;
1206 unsigned long *rmap;
1211 if (!dirty || !kvm->arch.using_mmu_notifiers)
1214 /* We need to mark this page dirty in the rmap chain */
1215 gfn = gpa >> PAGE_SHIFT;
1216 srcu_idx = srcu_read_lock(&kvm->srcu);
1217 memslot = gfn_to_memslot(kvm, gfn);
1219 rmap = &memslot->arch.rmap[gfn - memslot->base_gfn];
1221 *rmap |= KVMPPC_RMAP_CHANGED;
1224 srcu_read_unlock(&kvm->srcu, srcu_idx);
1228 * Functions for reading and writing the hash table via reads and
1229 * writes on a file descriptor.
1231 * Reads return the guest view of the hash table, which has to be
1232 * pieced together from the real hash table and the guest_rpte
1233 * values in the revmap array.
1235 * On writes, each HPTE written is considered in turn, and if it
1236 * is valid, it is written to the HPT as if an H_ENTER with the
1237 * exact flag set was done. When the invalid count is non-zero
1238 * in the header written to the stream, the kernel will make
1239 * sure that that many HPTEs are invalid, and invalidate them
1243 struct kvm_htab_ctx {
1244 unsigned long index;
1245 unsigned long flags;
1250 #define HPTE_SIZE (2 * sizeof(unsigned long))
1253 * Returns 1 if this HPT entry has been modified or has pending
1256 static int hpte_dirty(struct revmap_entry *revp, unsigned long *hptp)
1258 unsigned long rcbits_unset;
1260 if (revp->guest_rpte & HPTE_GR_MODIFIED)
1263 /* Also need to consider changes in reference and changed bits */
1264 rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1265 if ((hptp[0] & HPTE_V_VALID) && (hptp[1] & rcbits_unset))
1271 static long record_hpte(unsigned long flags, unsigned long *hptp,
1272 unsigned long *hpte, struct revmap_entry *revp,
1273 int want_valid, int first_pass)
1276 unsigned long rcbits_unset;
1280 /* Unmodified entries are uninteresting except on the first pass */
1281 dirty = hpte_dirty(revp, hptp);
1282 if (!first_pass && !dirty)
1286 if (hptp[0] & (HPTE_V_VALID | HPTE_V_ABSENT)) {
1288 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) &&
1289 !(hptp[0] & HPTE_V_BOLTED))
1292 if (valid != want_valid)
1296 if (valid || dirty) {
1297 /* lock the HPTE so it's stable and read it */
1299 while (!try_lock_hpte(hptp, HPTE_V_HVLOCK))
1303 /* re-evaluate valid and dirty from synchronized HPTE value */
1304 valid = !!(v & HPTE_V_VALID);
1305 dirty = !!(revp->guest_rpte & HPTE_GR_MODIFIED);
1307 /* Harvest R and C into guest view if necessary */
1308 rcbits_unset = ~revp->guest_rpte & (HPTE_R_R | HPTE_R_C);
1309 if (valid && (rcbits_unset & hptp[1])) {
1310 revp->guest_rpte |= (hptp[1] & (HPTE_R_R | HPTE_R_C)) |
1315 if (v & HPTE_V_ABSENT) {
1316 v &= ~HPTE_V_ABSENT;
1320 if ((flags & KVM_GET_HTAB_BOLTED_ONLY) && !(v & HPTE_V_BOLTED))
1323 r = revp->guest_rpte;
1324 /* only clear modified if this is the right sort of entry */
1325 if (valid == want_valid && dirty) {
1326 r &= ~HPTE_GR_MODIFIED;
1327 revp->guest_rpte = r;
1329 asm volatile(PPC_RELEASE_BARRIER "" : : : "memory");
1330 hptp[0] &= ~HPTE_V_HVLOCK;
1332 if (!(valid == want_valid && (first_pass || dirty)))
1340 static ssize_t kvm_htab_read(struct file *file, char __user *buf,
1341 size_t count, loff_t *ppos)
1343 struct kvm_htab_ctx *ctx = file->private_data;
1344 struct kvm *kvm = ctx->kvm;
1345 struct kvm_get_htab_header hdr;
1346 unsigned long *hptp;
1347 struct revmap_entry *revp;
1348 unsigned long i, nb, nw;
1349 unsigned long __user *lbuf;
1350 struct kvm_get_htab_header __user *hptr;
1351 unsigned long flags;
1353 unsigned long hpte[2];
1355 if (!access_ok(VERIFY_WRITE, buf, count))
1358 first_pass = ctx->first_pass;
1362 hptp = (unsigned long *)(kvm->arch.hpt_virt + (i * HPTE_SIZE));
1363 revp = kvm->arch.revmap + i;
1364 lbuf = (unsigned long __user *)buf;
1367 while (nb + sizeof(hdr) + HPTE_SIZE < count) {
1368 /* Initialize header */
1369 hptr = (struct kvm_get_htab_header __user *)buf;
1374 lbuf = (unsigned long __user *)(buf + sizeof(hdr));
1376 /* Skip uninteresting entries, i.e. clean on not-first pass */
1378 while (i < kvm->arch.hpt_npte &&
1379 !hpte_dirty(revp, hptp)) {
1387 /* Grab a series of valid entries */
1388 while (i < kvm->arch.hpt_npte &&
1389 hdr.n_valid < 0xffff &&
1390 nb + HPTE_SIZE < count &&
1391 record_hpte(flags, hptp, hpte, revp, 1, first_pass)) {
1392 /* valid entry, write it out */
1394 if (__put_user(hpte[0], lbuf) ||
1395 __put_user(hpte[1], lbuf + 1))
1403 /* Now skip invalid entries while we can */
1404 while (i < kvm->arch.hpt_npte &&
1405 hdr.n_invalid < 0xffff &&
1406 record_hpte(flags, hptp, hpte, revp, 0, first_pass)) {
1407 /* found an invalid entry */
1414 if (hdr.n_valid || hdr.n_invalid) {
1415 /* write back the header */
1416 if (__copy_to_user(hptr, &hdr, sizeof(hdr)))
1419 buf = (char __user *)lbuf;
1424 /* Check if we've wrapped around the hash table */
1425 if (i >= kvm->arch.hpt_npte) {
1427 ctx->first_pass = 0;
1437 static ssize_t kvm_htab_write(struct file *file, const char __user *buf,
1438 size_t count, loff_t *ppos)
1440 struct kvm_htab_ctx *ctx = file->private_data;
1441 struct kvm *kvm = ctx->kvm;
1442 struct kvm_get_htab_header hdr;
1445 unsigned long __user *lbuf;
1446 unsigned long *hptp;
1447 unsigned long tmp[2];
1452 if (!access_ok(VERIFY_READ, buf, count))
1455 /* lock out vcpus from running while we're doing this */
1456 mutex_lock(&kvm->lock);
1457 rma_setup = kvm->arch.rma_setup_done;
1459 kvm->arch.rma_setup_done = 0; /* temporarily */
1460 /* order rma_setup_done vs. vcpus_running */
1462 if (atomic_read(&kvm->arch.vcpus_running)) {
1463 kvm->arch.rma_setup_done = 1;
1464 mutex_unlock(&kvm->lock);
1470 for (nb = 0; nb + sizeof(hdr) <= count; ) {
1472 if (__copy_from_user(&hdr, buf, sizeof(hdr)))
1476 if (nb + hdr.n_valid * HPTE_SIZE > count)
1484 if (i >= kvm->arch.hpt_npte ||
1485 i + hdr.n_valid + hdr.n_invalid > kvm->arch.hpt_npte)
1488 hptp = (unsigned long *)(kvm->arch.hpt_virt + (i * HPTE_SIZE));
1489 lbuf = (unsigned long __user *)buf;
1490 for (j = 0; j < hdr.n_valid; ++j) {
1492 if (__get_user(v, lbuf) || __get_user(r, lbuf + 1))
1495 if (!(v & HPTE_V_VALID))
1500 if (hptp[0] & (HPTE_V_VALID | HPTE_V_ABSENT))
1501 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1503 ret = kvmppc_virtmode_do_h_enter(kvm, H_EXACT, i, v, r,
1505 if (ret != H_SUCCESS) {
1506 pr_err("kvm_htab_write ret %ld i=%ld v=%lx "
1507 "r=%lx\n", ret, i, v, r);
1510 if (!rma_setup && is_vrma_hpte(v)) {
1511 unsigned long psize = hpte_page_size(v, r);
1512 unsigned long senc = slb_pgsize_encoding(psize);
1515 kvm->arch.vrma_slb_v = senc | SLB_VSID_B_1T |
1516 (VRMA_VSID << SLB_VSID_SHIFT_1T);
1517 lpcr = kvm->arch.lpcr & ~LPCR_VRMASD;
1518 lpcr |= senc << (LPCR_VRMASD_SH - 4);
1519 kvm->arch.lpcr = lpcr;
1526 for (j = 0; j < hdr.n_invalid; ++j) {
1527 if (hptp[0] & (HPTE_V_VALID | HPTE_V_ABSENT))
1528 kvmppc_do_h_remove(kvm, 0, i, 0, tmp);
1536 /* Order HPTE updates vs. rma_setup_done */
1538 kvm->arch.rma_setup_done = rma_setup;
1539 mutex_unlock(&kvm->lock);
1546 static int kvm_htab_release(struct inode *inode, struct file *filp)
1548 struct kvm_htab_ctx *ctx = filp->private_data;
1550 filp->private_data = NULL;
1551 if (!(ctx->flags & KVM_GET_HTAB_WRITE))
1552 atomic_dec(&ctx->kvm->arch.hpte_mod_interest);
1553 kvm_put_kvm(ctx->kvm);
1558 static const struct file_operations kvm_htab_fops = {
1559 .read = kvm_htab_read,
1560 .write = kvm_htab_write,
1561 .llseek = default_llseek,
1562 .release = kvm_htab_release,
1565 int kvm_vm_ioctl_get_htab_fd(struct kvm *kvm, struct kvm_get_htab_fd *ghf)
1568 struct kvm_htab_ctx *ctx;
1571 /* reject flags we don't recognize */
1572 if (ghf->flags & ~(KVM_GET_HTAB_BOLTED_ONLY | KVM_GET_HTAB_WRITE))
1574 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
1579 ctx->index = ghf->start_index;
1580 ctx->flags = ghf->flags;
1581 ctx->first_pass = 1;
1583 rwflag = (ghf->flags & KVM_GET_HTAB_WRITE) ? O_WRONLY : O_RDONLY;
1584 ret = anon_inode_getfd("kvm-htab", &kvm_htab_fops, ctx, rwflag);
1590 if (rwflag == O_RDONLY) {
1591 mutex_lock(&kvm->slots_lock);
1592 atomic_inc(&kvm->arch.hpte_mod_interest);
1593 /* make sure kvmppc_do_h_enter etc. see the increment */
1594 synchronize_srcu_expedited(&kvm->srcu);
1595 mutex_unlock(&kvm->slots_lock);
1601 void kvmppc_mmu_book3s_hv_init(struct kvm_vcpu *vcpu)
1603 struct kvmppc_mmu *mmu = &vcpu->arch.mmu;
1605 if (cpu_has_feature(CPU_FTR_ARCH_206))
1606 vcpu->arch.slb_nr = 32; /* POWER7 */
1608 vcpu->arch.slb_nr = 64;
1610 mmu->xlate = kvmppc_mmu_book3s_64_hv_xlate;
1611 mmu->reset_msr = kvmppc_mmu_book3s_64_hv_reset_msr;
1613 vcpu->arch.hflags |= BOOK3S_HFLAG_SLB;