1 // SPDX-License-Identifier: GPL-2.0-only
4 * Copyright 2016 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
7 #include <linux/types.h>
8 #include <linux/string.h>
10 #include <linux/kvm_host.h>
11 #include <linux/anon_inodes.h>
12 #include <linux/file.h>
13 #include <linux/debugfs.h>
14 #include <linux/pgtable.h>
16 #include <asm/kvm_ppc.h>
17 #include <asm/kvm_book3s.h>
20 #include <asm/pgalloc.h>
21 #include <asm/pte-walk.h>
22 #include <asm/ultravisor.h>
23 #include <asm/kvm_book3s_uvmem.h>
26 * Supported radix tree geometry.
27 * Like p9, we support either 5 or 9 bits at the first (lowest) level,
28 * for a page size of 64k or 4k.
30 static int p9_supported_radix_bits[4] = { 5, 9, 9, 13 };
32 unsigned long __kvmhv_copy_tofrom_guest_radix(int lpid, int pid,
33 gva_t eaddr, void *to, void *from,
36 int uninitialized_var(old_pid), old_lpid;
37 unsigned long quadrant, ret = n;
40 /* Can't access quadrants 1 or 2 in non-HV mode, call the HV to do it */
41 if (kvmhv_on_pseries())
42 return plpar_hcall_norets(H_COPY_TOFROM_GUEST, lpid, pid, eaddr,
43 __pa(to), __pa(from), n);
49 from = (void *) (eaddr | (quadrant << 62));
51 to = (void *) (eaddr | (quadrant << 62));
55 /* switch the lpid first to avoid running host with unallocated pid */
56 old_lpid = mfspr(SPRN_LPID);
58 mtspr(SPRN_LPID, lpid);
60 old_pid = mfspr(SPRN_PID);
67 ret = probe_user_read(to, (const void __user *)from, n);
69 ret = probe_user_write((void __user *)to, from, n);
71 /* switch the pid first to avoid running host with unallocated pid */
72 if (quadrant == 1 && pid != old_pid)
73 mtspr(SPRN_PID, old_pid);
75 mtspr(SPRN_LPID, old_lpid);
82 EXPORT_SYMBOL_GPL(__kvmhv_copy_tofrom_guest_radix);
84 static long kvmhv_copy_tofrom_guest_radix(struct kvm_vcpu *vcpu, gva_t eaddr,
85 void *to, void *from, unsigned long n)
87 int lpid = vcpu->kvm->arch.lpid;
88 int pid = vcpu->arch.pid;
90 /* This would cause a data segment intr so don't allow the access */
91 if (eaddr & (0x3FFUL << 52))
94 /* Should we be using the nested lpid */
95 if (vcpu->arch.nested)
96 lpid = vcpu->arch.nested->shadow_lpid;
98 /* If accessing quadrant 3 then pid is expected to be 0 */
99 if (((eaddr >> 62) & 0x3) == 0x3)
102 eaddr &= ~(0xFFFUL << 52);
104 return __kvmhv_copy_tofrom_guest_radix(lpid, pid, eaddr, to, from, n);
107 long kvmhv_copy_from_guest_radix(struct kvm_vcpu *vcpu, gva_t eaddr, void *to,
112 ret = kvmhv_copy_tofrom_guest_radix(vcpu, eaddr, to, NULL, n);
114 memset(to + (n - ret), 0, ret);
118 EXPORT_SYMBOL_GPL(kvmhv_copy_from_guest_radix);
120 long kvmhv_copy_to_guest_radix(struct kvm_vcpu *vcpu, gva_t eaddr, void *from,
123 return kvmhv_copy_tofrom_guest_radix(vcpu, eaddr, NULL, from, n);
125 EXPORT_SYMBOL_GPL(kvmhv_copy_to_guest_radix);
127 int kvmppc_mmu_walk_radix_tree(struct kvm_vcpu *vcpu, gva_t eaddr,
128 struct kvmppc_pte *gpte, u64 root,
131 struct kvm *kvm = vcpu->kvm;
133 unsigned long rts, bits, offset, index;
137 rts = ((root & RTS1_MASK) >> (RTS1_SHIFT - 3)) |
138 ((root & RTS2_MASK) >> RTS2_SHIFT);
139 bits = root & RPDS_MASK;
140 base = root & RPDB_MASK;
144 /* Current implementations only support 52-bit space */
148 /* Walk each level of the radix tree */
149 for (level = 3; level >= 0; --level) {
151 /* Check a valid size */
152 if (level && bits != p9_supported_radix_bits[level])
154 if (level == 0 && !(bits == 5 || bits == 9))
157 index = (eaddr >> offset) & ((1UL << bits) - 1);
158 /* Check that low bits of page table base are zero */
159 if (base & ((1UL << (bits + 3)) - 1))
161 /* Read the entry from guest memory */
162 addr = base + (index * sizeof(rpte));
163 ret = kvm_read_guest(kvm, addr, &rpte, sizeof(rpte));
169 pte = __be64_to_cpu(rpte);
170 if (!(pte & _PAGE_PRESENT))
172 /* Check if a leaf entry */
175 /* Get ready to walk the next level */
176 base = pte & RPDB_MASK;
177 bits = pte & RPDS_MASK;
180 /* Need a leaf at lowest level; 512GB pages not supported */
181 if (level < 0 || level == 3)
184 /* We found a valid leaf PTE */
185 /* Offset is now log base 2 of the page size */
186 gpa = pte & 0x01fffffffffff000ul;
187 if (gpa & ((1ul << offset) - 1))
189 gpa |= eaddr & ((1ul << offset) - 1);
190 for (ps = MMU_PAGE_4K; ps < MMU_PAGE_COUNT; ++ps)
191 if (offset == mmu_psize_defs[ps].shift)
193 gpte->page_size = ps;
194 gpte->page_shift = offset;
199 /* Work out permissions */
200 gpte->may_read = !!(pte & _PAGE_READ);
201 gpte->may_write = !!(pte & _PAGE_WRITE);
202 gpte->may_execute = !!(pte & _PAGE_EXEC);
204 gpte->rc = pte & (_PAGE_ACCESSED | _PAGE_DIRTY);
213 * Used to walk a partition or process table radix tree in guest memory
214 * Note: We exploit the fact that a partition table and a process
215 * table have the same layout, a partition-scoped page table and a
216 * process-scoped page table have the same layout, and the 2nd
217 * doubleword of a partition table entry has the same layout as
220 int kvmppc_mmu_radix_translate_table(struct kvm_vcpu *vcpu, gva_t eaddr,
221 struct kvmppc_pte *gpte, u64 table,
222 int table_index, u64 *pte_ret_p)
224 struct kvm *kvm = vcpu->kvm;
226 unsigned long size, ptbl, root;
227 struct prtb_entry entry;
229 if ((table & PRTS_MASK) > 24)
231 size = 1ul << ((table & PRTS_MASK) + 12);
233 /* Is the table big enough to contain this entry? */
234 if ((table_index * sizeof(entry)) >= size)
237 /* Read the table to find the root of the radix tree */
238 ptbl = (table & PRTB_MASK) + (table_index * sizeof(entry));
239 ret = kvm_read_guest(kvm, ptbl, &entry, sizeof(entry));
243 /* Root is stored in the first double word */
244 root = be64_to_cpu(entry.prtb0);
246 return kvmppc_mmu_walk_radix_tree(vcpu, eaddr, gpte, root, pte_ret_p);
249 int kvmppc_mmu_radix_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
250 struct kvmppc_pte *gpte, bool data, bool iswrite)
256 /* Work out effective PID */
257 switch (eaddr >> 62) {
259 pid = vcpu->arch.pid;
268 ret = kvmppc_mmu_radix_translate_table(vcpu, eaddr, gpte,
269 vcpu->kvm->arch.process_table, pid, &pte);
273 /* Check privilege (applies only to process scoped translations) */
274 if (kvmppc_get_msr(vcpu) & MSR_PR) {
275 if (pte & _PAGE_PRIVILEGED) {
278 gpte->may_execute = 0;
281 if (!(pte & _PAGE_PRIVILEGED)) {
282 /* Check AMR/IAMR to see if strict mode is in force */
283 if (vcpu->arch.amr & (1ul << 62))
285 if (vcpu->arch.amr & (1ul << 63))
287 if (vcpu->arch.iamr & (1ul << 62))
288 gpte->may_execute = 0;
295 void kvmppc_radix_tlbie_page(struct kvm *kvm, unsigned long addr,
296 unsigned int pshift, unsigned int lpid)
298 unsigned long psize = PAGE_SIZE;
304 psize = 1UL << pshift;
308 addr &= ~(psize - 1);
310 if (!kvmhv_on_pseries()) {
311 radix__flush_tlb_lpid_page(lpid, addr, psize);
315 psi = shift_to_mmu_psize(pshift);
316 rb = addr | (mmu_get_ap(psi) << PPC_BITLSHIFT(58));
317 rc = plpar_hcall_norets(H_TLB_INVALIDATE, H_TLBIE_P1_ENC(0, 0, 1),
320 pr_err("KVM: TLB page invalidation hcall failed, rc=%ld\n", rc);
323 static void kvmppc_radix_flush_pwc(struct kvm *kvm, unsigned int lpid)
327 if (!kvmhv_on_pseries()) {
328 radix__flush_pwc_lpid(lpid);
332 rc = plpar_hcall_norets(H_TLB_INVALIDATE, H_TLBIE_P1_ENC(1, 0, 1),
333 lpid, TLBIEL_INVAL_SET_LPID);
335 pr_err("KVM: TLB PWC invalidation hcall failed, rc=%ld\n", rc);
338 static unsigned long kvmppc_radix_update_pte(struct kvm *kvm, pte_t *ptep,
339 unsigned long clr, unsigned long set,
340 unsigned long addr, unsigned int shift)
342 return __radix_pte_update(ptep, clr, set);
345 void kvmppc_radix_set_pte_at(struct kvm *kvm, unsigned long addr,
346 pte_t *ptep, pte_t pte)
348 radix__set_pte_at(kvm->mm, addr, ptep, pte, 0);
351 static struct kmem_cache *kvm_pte_cache;
352 static struct kmem_cache *kvm_pmd_cache;
354 static pte_t *kvmppc_pte_alloc(void)
358 pte = kmem_cache_alloc(kvm_pte_cache, GFP_KERNEL);
359 /* pmd_populate() will only reference _pa(pte). */
360 kmemleak_ignore(pte);
365 static void kvmppc_pte_free(pte_t *ptep)
367 kmem_cache_free(kvm_pte_cache, ptep);
370 static pmd_t *kvmppc_pmd_alloc(void)
374 pmd = kmem_cache_alloc(kvm_pmd_cache, GFP_KERNEL);
375 /* pud_populate() will only reference _pa(pmd). */
376 kmemleak_ignore(pmd);
381 static void kvmppc_pmd_free(pmd_t *pmdp)
383 kmem_cache_free(kvm_pmd_cache, pmdp);
386 /* Called with kvm->mmu_lock held */
387 void kvmppc_unmap_pte(struct kvm *kvm, pte_t *pte, unsigned long gpa,
389 const struct kvm_memory_slot *memslot,
394 unsigned long gfn = gpa >> PAGE_SHIFT;
395 unsigned long page_size = PAGE_SIZE;
398 old = kvmppc_radix_update_pte(kvm, pte, ~0UL, 0, gpa, shift);
399 kvmppc_radix_tlbie_page(kvm, gpa, shift, lpid);
401 /* The following only applies to L1 entries */
402 if (lpid != kvm->arch.lpid)
406 memslot = gfn_to_memslot(kvm, gfn);
410 if (shift) { /* 1GB or 2MB page */
411 page_size = 1ul << shift;
412 if (shift == PMD_SHIFT)
413 kvm->stat.num_2M_pages--;
414 else if (shift == PUD_SHIFT)
415 kvm->stat.num_1G_pages--;
418 gpa &= ~(page_size - 1);
419 hpa = old & PTE_RPN_MASK;
420 kvmhv_remove_nest_rmap_range(kvm, memslot, gpa, hpa, page_size);
422 if ((old & _PAGE_DIRTY) && memslot->dirty_bitmap)
423 kvmppc_update_dirty_map(memslot, gfn, page_size);
427 * kvmppc_free_p?d are used to free existing page tables, and recursively
428 * descend and clear and free children.
429 * Callers are responsible for flushing the PWC.
431 * When page tables are being unmapped/freed as part of page fault path
432 * (full == false), valid ptes are generally not expected; however, there
433 * is one situation where they arise, which is when dirty page logging is
434 * turned off for a memslot while the VM is running. The new memslot
435 * becomes visible to page faults before the memslot commit function
436 * gets to flush the memslot, which can lead to a 2MB page mapping being
437 * installed for a guest physical address where there are already 64kB
438 * (or 4kB) mappings (of sub-pages of the same 2MB page).
440 static void kvmppc_unmap_free_pte(struct kvm *kvm, pte_t *pte, bool full,
444 memset(pte, 0, sizeof(long) << RADIX_PTE_INDEX_SIZE);
449 for (it = 0; it < PTRS_PER_PTE; ++it, ++p) {
450 if (pte_val(*p) == 0)
452 kvmppc_unmap_pte(kvm, p,
453 pte_pfn(*p) << PAGE_SHIFT,
454 PAGE_SHIFT, NULL, lpid);
458 kvmppc_pte_free(pte);
461 static void kvmppc_unmap_free_pmd(struct kvm *kvm, pmd_t *pmd, bool full,
467 for (im = 0; im < PTRS_PER_PMD; ++im, ++p) {
468 if (!pmd_present(*p))
470 if (pmd_is_leaf(*p)) {
475 kvmppc_unmap_pte(kvm, (pte_t *)p,
476 pte_pfn(*(pte_t *)p) << PAGE_SHIFT,
477 PMD_SHIFT, NULL, lpid);
482 pte = pte_offset_map(p, 0);
483 kvmppc_unmap_free_pte(kvm, pte, full, lpid);
487 kvmppc_pmd_free(pmd);
490 static void kvmppc_unmap_free_pud(struct kvm *kvm, pud_t *pud,
496 for (iu = 0; iu < PTRS_PER_PUD; ++iu, ++p) {
497 if (!pud_present(*p))
499 if (pud_is_leaf(*p)) {
504 pmd = pmd_offset(p, 0);
505 kvmppc_unmap_free_pmd(kvm, pmd, true, lpid);
509 pud_free(kvm->mm, pud);
512 void kvmppc_free_pgtable_radix(struct kvm *kvm, pgd_t *pgd, unsigned int lpid)
516 for (ig = 0; ig < PTRS_PER_PGD; ++ig, ++pgd) {
517 p4d_t *p4d = p4d_offset(pgd, 0);
520 if (!p4d_present(*p4d))
522 pud = pud_offset(p4d, 0);
523 kvmppc_unmap_free_pud(kvm, pud, lpid);
528 void kvmppc_free_radix(struct kvm *kvm)
530 if (kvm->arch.pgtable) {
531 kvmppc_free_pgtable_radix(kvm, kvm->arch.pgtable,
533 pgd_free(kvm->mm, kvm->arch.pgtable);
534 kvm->arch.pgtable = NULL;
538 static void kvmppc_unmap_free_pmd_entry_table(struct kvm *kvm, pmd_t *pmd,
539 unsigned long gpa, unsigned int lpid)
541 pte_t *pte = pte_offset_kernel(pmd, 0);
544 * Clearing the pmd entry then flushing the PWC ensures that the pte
545 * page no longer be cached by the MMU, so can be freed without
546 * flushing the PWC again.
549 kvmppc_radix_flush_pwc(kvm, lpid);
551 kvmppc_unmap_free_pte(kvm, pte, false, lpid);
554 static void kvmppc_unmap_free_pud_entry_table(struct kvm *kvm, pud_t *pud,
555 unsigned long gpa, unsigned int lpid)
557 pmd_t *pmd = pmd_offset(pud, 0);
560 * Clearing the pud entry then flushing the PWC ensures that the pmd
561 * page and any children pte pages will no longer be cached by the MMU,
562 * so can be freed without flushing the PWC again.
565 kvmppc_radix_flush_pwc(kvm, lpid);
567 kvmppc_unmap_free_pmd(kvm, pmd, false, lpid);
571 * There are a number of bits which may differ between different faults to
572 * the same partition scope entry. RC bits, in the course of cleaning and
573 * aging. And the write bit can change, either the access could have been
574 * upgraded, or a read fault could happen concurrently with a write fault
575 * that sets those bits first.
577 #define PTE_BITS_MUST_MATCH (~(_PAGE_WRITE | _PAGE_DIRTY | _PAGE_ACCESSED))
579 int kvmppc_create_pte(struct kvm *kvm, pgd_t *pgtable, pte_t pte,
580 unsigned long gpa, unsigned int level,
581 unsigned long mmu_seq, unsigned int lpid,
582 unsigned long *rmapp, struct rmap_nested **n_rmap)
586 pud_t *pud, *new_pud = NULL;
587 pmd_t *pmd, *new_pmd = NULL;
588 pte_t *ptep, *new_ptep = NULL;
591 /* Traverse the guest's 2nd-level tree, allocate new levels needed */
592 pgd = pgtable + pgd_index(gpa);
593 p4d = p4d_offset(pgd, gpa);
596 if (p4d_present(*p4d))
597 pud = pud_offset(p4d, gpa);
599 new_pud = pud_alloc_one(kvm->mm, gpa);
602 if (pud && pud_present(*pud) && !pud_is_leaf(*pud))
603 pmd = pmd_offset(pud, gpa);
605 new_pmd = kvmppc_pmd_alloc();
607 if (level == 0 && !(pmd && pmd_present(*pmd) && !pmd_is_leaf(*pmd)))
608 new_ptep = kvmppc_pte_alloc();
610 /* Check if we might have been invalidated; let the guest retry if so */
611 spin_lock(&kvm->mmu_lock);
613 if (mmu_notifier_retry(kvm, mmu_seq))
616 /* Now traverse again under the lock and change the tree */
618 if (p4d_none(*p4d)) {
621 p4d_populate(kvm->mm, p4d, new_pud);
624 pud = pud_offset(p4d, gpa);
625 if (pud_is_leaf(*pud)) {
626 unsigned long hgpa = gpa & PUD_MASK;
628 /* Check if we raced and someone else has set the same thing */
630 if (pud_raw(*pud) == pte_raw(pte)) {
634 /* Valid 1GB page here already, add our extra bits */
635 WARN_ON_ONCE((pud_val(*pud) ^ pte_val(pte)) &
636 PTE_BITS_MUST_MATCH);
637 kvmppc_radix_update_pte(kvm, (pte_t *)pud,
638 0, pte_val(pte), hgpa, PUD_SHIFT);
643 * If we raced with another CPU which has just put
644 * a 1GB pte in after we saw a pmd page, try again.
650 /* Valid 1GB page here already, remove it */
651 kvmppc_unmap_pte(kvm, (pte_t *)pud, hgpa, PUD_SHIFT, NULL,
655 if (!pud_none(*pud)) {
657 * There's a page table page here, but we wanted to
658 * install a large page, so remove and free the page
661 kvmppc_unmap_free_pud_entry_table(kvm, pud, gpa, lpid);
663 kvmppc_radix_set_pte_at(kvm, gpa, (pte_t *)pud, pte);
665 kvmhv_insert_nest_rmap(kvm, rmapp, n_rmap);
669 if (pud_none(*pud)) {
672 pud_populate(kvm->mm, pud, new_pmd);
675 pmd = pmd_offset(pud, gpa);
676 if (pmd_is_leaf(*pmd)) {
677 unsigned long lgpa = gpa & PMD_MASK;
679 /* Check if we raced and someone else has set the same thing */
681 if (pmd_raw(*pmd) == pte_raw(pte)) {
685 /* Valid 2MB page here already, add our extra bits */
686 WARN_ON_ONCE((pmd_val(*pmd) ^ pte_val(pte)) &
687 PTE_BITS_MUST_MATCH);
688 kvmppc_radix_update_pte(kvm, pmdp_ptep(pmd),
689 0, pte_val(pte), lgpa, PMD_SHIFT);
695 * If we raced with another CPU which has just put
696 * a 2MB pte in after we saw a pte page, try again.
702 /* Valid 2MB page here already, remove it */
703 kvmppc_unmap_pte(kvm, pmdp_ptep(pmd), lgpa, PMD_SHIFT, NULL,
707 if (!pmd_none(*pmd)) {
709 * There's a page table page here, but we wanted to
710 * install a large page, so remove and free the page
713 kvmppc_unmap_free_pmd_entry_table(kvm, pmd, gpa, lpid);
715 kvmppc_radix_set_pte_at(kvm, gpa, pmdp_ptep(pmd), pte);
717 kvmhv_insert_nest_rmap(kvm, rmapp, n_rmap);
721 if (pmd_none(*pmd)) {
724 pmd_populate(kvm->mm, pmd, new_ptep);
727 ptep = pte_offset_kernel(pmd, gpa);
728 if (pte_present(*ptep)) {
729 /* Check if someone else set the same thing */
730 if (pte_raw(*ptep) == pte_raw(pte)) {
734 /* Valid page here already, add our extra bits */
735 WARN_ON_ONCE((pte_val(*ptep) ^ pte_val(pte)) &
736 PTE_BITS_MUST_MATCH);
737 kvmppc_radix_update_pte(kvm, ptep, 0, pte_val(pte), gpa, 0);
741 kvmppc_radix_set_pte_at(kvm, gpa, ptep, pte);
743 kvmhv_insert_nest_rmap(kvm, rmapp, n_rmap);
747 spin_unlock(&kvm->mmu_lock);
749 pud_free(kvm->mm, new_pud);
751 kvmppc_pmd_free(new_pmd);
753 kvmppc_pte_free(new_ptep);
757 bool kvmppc_hv_handle_set_rc(struct kvm *kvm, bool nested, bool writing,
758 unsigned long gpa, unsigned int lpid)
760 unsigned long pgflags;
765 * Need to set an R or C bit in the 2nd-level tables;
766 * since we are just helping out the hardware here,
767 * it is sufficient to do what the hardware does.
769 pgflags = _PAGE_ACCESSED;
771 pgflags |= _PAGE_DIRTY;
774 ptep = find_kvm_nested_guest_pte(kvm, lpid, gpa, &shift);
776 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
778 if (ptep && pte_present(*ptep) && (!writing || pte_write(*ptep))) {
779 kvmppc_radix_update_pte(kvm, ptep, 0, pgflags, gpa, shift);
785 int kvmppc_book3s_instantiate_page(struct kvm_vcpu *vcpu,
787 struct kvm_memory_slot *memslot,
788 bool writing, bool kvm_ro,
789 pte_t *inserted_pte, unsigned int *levelp)
791 struct kvm *kvm = vcpu->kvm;
792 struct page *page = NULL;
793 unsigned long mmu_seq;
794 unsigned long hva, gfn = gpa >> PAGE_SHIFT;
795 bool upgrade_write = false;
796 bool *upgrade_p = &upgrade_write;
798 unsigned int shift, level;
802 /* used to check for invalidations in progress */
803 mmu_seq = kvm->mmu_notifier_seq;
807 * Do a fast check first, since __gfn_to_pfn_memslot doesn't
808 * do it with !atomic && !async, which is how we call it.
809 * We always ask for write permission since the common case
810 * is that the page is writable.
812 hva = gfn_to_hva_memslot(memslot, gfn);
813 if (!kvm_ro && get_user_page_fast_only(hva, FOLL_WRITE, &page)) {
814 upgrade_write = true;
818 /* Call KVM generic code to do the slow-path check */
819 pfn = __gfn_to_pfn_memslot(memslot, gfn, false, NULL,
821 if (is_error_noslot_pfn(pfn))
824 if (pfn_valid(pfn)) {
825 page = pfn_to_page(pfn);
826 if (PageReserved(page))
832 * Read the PTE from the process' radix tree and use that
833 * so we get the shift and attribute bits.
835 spin_lock(&kvm->mmu_lock);
836 ptep = find_kvm_host_pte(kvm, mmu_seq, hva, &shift);
839 pte = READ_ONCE(*ptep);
840 spin_unlock(&kvm->mmu_lock);
842 * If the PTE disappeared temporarily due to a THP
843 * collapse, just return and let the guest try again.
845 if (!pte_present(pte)) {
851 /* If we're logging dirty pages, always map single pages */
852 large_enable = !(memslot->flags & KVM_MEM_LOG_DIRTY_PAGES);
854 /* Get pte level from shift/size */
855 if (large_enable && shift == PUD_SHIFT &&
856 (gpa & (PUD_SIZE - PAGE_SIZE)) ==
857 (hva & (PUD_SIZE - PAGE_SIZE))) {
859 } else if (large_enable && shift == PMD_SHIFT &&
860 (gpa & (PMD_SIZE - PAGE_SIZE)) ==
861 (hva & (PMD_SIZE - PAGE_SIZE))) {
865 if (shift > PAGE_SHIFT) {
867 * If the pte maps more than one page, bring over
868 * bits from the virtual address to get the real
869 * address of the specific single page we want.
871 unsigned long rpnmask = (1ul << shift) - PAGE_SIZE;
872 pte = __pte(pte_val(pte) | (hva & rpnmask));
876 pte = __pte(pte_val(pte) | _PAGE_EXEC | _PAGE_ACCESSED);
877 if (writing || upgrade_write) {
878 if (pte_val(pte) & _PAGE_WRITE)
879 pte = __pte(pte_val(pte) | _PAGE_DIRTY);
881 pte = __pte(pte_val(pte) & ~(_PAGE_WRITE | _PAGE_DIRTY));
884 /* Allocate space in the tree and write the PTE */
885 ret = kvmppc_create_pte(kvm, kvm->arch.pgtable, pte, gpa, level,
886 mmu_seq, kvm->arch.lpid, NULL, NULL);
893 if (!ret && (pte_val(pte) & _PAGE_WRITE))
894 set_page_dirty_lock(page);
898 /* Increment number of large pages if we (successfully) inserted one */
901 kvm->stat.num_2M_pages++;
903 kvm->stat.num_1G_pages++;
909 int kvmppc_book3s_radix_page_fault(struct kvm_vcpu *vcpu,
910 unsigned long ea, unsigned long dsisr)
912 struct kvm *kvm = vcpu->kvm;
913 unsigned long gpa, gfn;
914 struct kvm_memory_slot *memslot;
916 bool writing = !!(dsisr & DSISR_ISSTORE);
919 /* Check for unusual errors */
920 if (dsisr & DSISR_UNSUPP_MMU) {
921 pr_err("KVM: Got unsupported MMU fault\n");
924 if (dsisr & DSISR_BADACCESS) {
925 /* Reflect to the guest as DSI */
926 pr_err("KVM: Got radix HV page fault with DSISR=%lx\n", dsisr);
927 kvmppc_core_queue_data_storage(vcpu, ea, dsisr);
931 /* Translate the logical address */
932 gpa = vcpu->arch.fault_gpa & ~0xfffUL;
933 gpa &= ~0xF000000000000000ul;
934 gfn = gpa >> PAGE_SHIFT;
935 if (!(dsisr & DSISR_PRTABLE_FAULT))
938 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
939 return kvmppc_send_page_to_uv(kvm, gfn);
941 /* Get the corresponding memslot */
942 memslot = gfn_to_memslot(kvm, gfn);
944 /* No memslot means it's an emulated MMIO region */
945 if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID)) {
946 if (dsisr & (DSISR_PRTABLE_FAULT | DSISR_BADACCESS |
949 * Bad address in guest page table tree, or other
950 * unusual error - reflect it to the guest as DSI.
952 kvmppc_core_queue_data_storage(vcpu, ea, dsisr);
955 return kvmppc_hv_emulate_mmio(vcpu, gpa, ea, writing);
958 if (memslot->flags & KVM_MEM_READONLY) {
960 /* give the guest a DSI */
961 kvmppc_core_queue_data_storage(vcpu, ea, DSISR_ISSTORE |
968 /* Failed to set the reference/change bits */
969 if (dsisr & DSISR_SET_RC) {
970 spin_lock(&kvm->mmu_lock);
971 if (kvmppc_hv_handle_set_rc(kvm, false, writing,
972 gpa, kvm->arch.lpid))
973 dsisr &= ~DSISR_SET_RC;
974 spin_unlock(&kvm->mmu_lock);
976 if (!(dsisr & (DSISR_BAD_FAULT_64S | DSISR_NOHPTE |
977 DSISR_PROTFAULT | DSISR_SET_RC)))
981 /* Try to insert a pte */
982 ret = kvmppc_book3s_instantiate_page(vcpu, gpa, memslot, writing,
985 if (ret == 0 || ret == -EAGAIN)
990 /* Called with kvm->mmu_lock held */
991 int kvm_unmap_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
995 unsigned long gpa = gfn << PAGE_SHIFT;
998 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE) {
999 uv_page_inval(kvm->arch.lpid, gpa, PAGE_SHIFT);
1003 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1004 if (ptep && pte_present(*ptep))
1005 kvmppc_unmap_pte(kvm, ptep, gpa, shift, memslot,
1010 /* Called with kvm->mmu_lock held */
1011 int kvm_age_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
1015 unsigned long gpa = gfn << PAGE_SHIFT;
1018 unsigned long old, *rmapp;
1020 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1023 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1024 if (ptep && pte_present(*ptep) && pte_young(*ptep)) {
1025 old = kvmppc_radix_update_pte(kvm, ptep, _PAGE_ACCESSED, 0,
1027 /* XXX need to flush tlb here? */
1028 /* Also clear bit in ptes in shadow pgtable for nested guests */
1029 rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1030 kvmhv_update_nest_rmap_rc_list(kvm, rmapp, _PAGE_ACCESSED, 0,
1038 /* Called with kvm->mmu_lock held */
1039 int kvm_test_age_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
1043 unsigned long gpa = gfn << PAGE_SHIFT;
1047 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1050 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1051 if (ptep && pte_present(*ptep) && pte_young(*ptep))
1056 /* Returns the number of PAGE_SIZE pages that are dirty */
1057 static int kvm_radix_test_clear_dirty(struct kvm *kvm,
1058 struct kvm_memory_slot *memslot, int pagenum)
1060 unsigned long gfn = memslot->base_gfn + pagenum;
1061 unsigned long gpa = gfn << PAGE_SHIFT;
1065 unsigned long old, *rmapp;
1067 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1071 * For performance reasons we don't hold kvm->mmu_lock while walking the
1072 * partition scoped table.
1074 ptep = find_kvm_secondary_pte_unlocked(kvm, gpa, &shift);
1078 pte = READ_ONCE(*ptep);
1079 if (pte_present(pte) && pte_dirty(pte)) {
1080 spin_lock(&kvm->mmu_lock);
1082 * Recheck the pte again
1084 if (pte_val(pte) != pte_val(*ptep)) {
1086 * We have KVM_MEM_LOG_DIRTY_PAGES enabled. Hence we can
1087 * only find PAGE_SIZE pte entries here. We can continue
1088 * to use the pte addr returned by above page table
1091 if (!pte_present(*ptep) || !pte_dirty(*ptep)) {
1092 spin_unlock(&kvm->mmu_lock);
1099 old = kvmppc_radix_update_pte(kvm, ptep, _PAGE_DIRTY, 0,
1101 kvmppc_radix_tlbie_page(kvm, gpa, shift, kvm->arch.lpid);
1102 /* Also clear bit in ptes in shadow pgtable for nested guests */
1103 rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1104 kvmhv_update_nest_rmap_rc_list(kvm, rmapp, _PAGE_DIRTY, 0,
1107 spin_unlock(&kvm->mmu_lock);
1112 long kvmppc_hv_get_dirty_log_radix(struct kvm *kvm,
1113 struct kvm_memory_slot *memslot, unsigned long *map)
1118 for (i = 0; i < memslot->npages; i = j) {
1119 npages = kvm_radix_test_clear_dirty(kvm, memslot, i);
1122 * Note that if npages > 0 then i must be a multiple of npages,
1123 * since huge pages are only used to back the guest at guest
1124 * real addresses that are a multiple of their size.
1125 * Since we have at most one PTE covering any given guest
1126 * real address, if npages > 1 we can skip to i + npages.
1130 set_dirty_bits(map, i, npages);
1137 void kvmppc_radix_flush_memslot(struct kvm *kvm,
1138 const struct kvm_memory_slot *memslot)
1145 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_START)
1146 kvmppc_uvmem_drop_pages(memslot, kvm, true);
1148 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1151 gpa = memslot->base_gfn << PAGE_SHIFT;
1152 spin_lock(&kvm->mmu_lock);
1153 for (n = memslot->npages; n; --n) {
1154 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1155 if (ptep && pte_present(*ptep))
1156 kvmppc_unmap_pte(kvm, ptep, gpa, shift, memslot,
1161 * Increase the mmu notifier sequence number to prevent any page
1162 * fault that read the memslot earlier from writing a PTE.
1164 kvm->mmu_notifier_seq++;
1165 spin_unlock(&kvm->mmu_lock);
1168 static void add_rmmu_ap_encoding(struct kvm_ppc_rmmu_info *info,
1169 int psize, int *indexp)
1171 if (!mmu_psize_defs[psize].shift)
1173 info->ap_encodings[*indexp] = mmu_psize_defs[psize].shift |
1174 (mmu_psize_defs[psize].ap << 29);
1178 int kvmhv_get_rmmu_info(struct kvm *kvm, struct kvm_ppc_rmmu_info *info)
1182 if (!radix_enabled())
1184 memset(info, 0, sizeof(*info));
1187 info->geometries[0].page_shift = 12;
1188 info->geometries[0].level_bits[0] = 9;
1189 for (i = 1; i < 4; ++i)
1190 info->geometries[0].level_bits[i] = p9_supported_radix_bits[i];
1192 info->geometries[1].page_shift = 16;
1193 for (i = 0; i < 4; ++i)
1194 info->geometries[1].level_bits[i] = p9_supported_radix_bits[i];
1197 add_rmmu_ap_encoding(info, MMU_PAGE_4K, &i);
1198 add_rmmu_ap_encoding(info, MMU_PAGE_64K, &i);
1199 add_rmmu_ap_encoding(info, MMU_PAGE_2M, &i);
1200 add_rmmu_ap_encoding(info, MMU_PAGE_1G, &i);
1205 int kvmppc_init_vm_radix(struct kvm *kvm)
1207 kvm->arch.pgtable = pgd_alloc(kvm->mm);
1208 if (!kvm->arch.pgtable)
1213 static void pte_ctor(void *addr)
1215 memset(addr, 0, RADIX_PTE_TABLE_SIZE);
1218 static void pmd_ctor(void *addr)
1220 memset(addr, 0, RADIX_PMD_TABLE_SIZE);
1223 struct debugfs_radix_state {
1234 static int debugfs_radix_open(struct inode *inode, struct file *file)
1236 struct kvm *kvm = inode->i_private;
1237 struct debugfs_radix_state *p;
1239 p = kzalloc(sizeof(*p), GFP_KERNEL);
1245 mutex_init(&p->mutex);
1246 file->private_data = p;
1248 return nonseekable_open(inode, file);
1251 static int debugfs_radix_release(struct inode *inode, struct file *file)
1253 struct debugfs_radix_state *p = file->private_data;
1255 kvm_put_kvm(p->kvm);
1260 static ssize_t debugfs_radix_read(struct file *file, char __user *buf,
1261 size_t len, loff_t *ppos)
1263 struct debugfs_radix_state *p = file->private_data;
1269 struct kvm_nested_guest *nested;
1279 if (!kvm_is_radix(kvm))
1282 ret = mutex_lock_interruptible(&p->mutex);
1286 if (p->chars_left) {
1290 r = copy_to_user(buf, p->buf + p->buf_index, n);
1307 while (len != 0 && p->lpid >= 0) {
1308 if (gpa >= RADIX_PGTABLE_RANGE) {
1312 kvmhv_put_nested(nested);
1315 p->lpid = kvmhv_nested_next_lpid(kvm, p->lpid);
1322 pgt = kvm->arch.pgtable;
1324 nested = kvmhv_get_nested(kvm, p->lpid, false);
1326 gpa = RADIX_PGTABLE_RANGE;
1329 pgt = nested->shadow_pgtable;
1335 n = scnprintf(p->buf, sizeof(p->buf),
1336 "\nNested LPID %d: ", p->lpid);
1337 n += scnprintf(p->buf + n, sizeof(p->buf) - n,
1338 "pgdir: %lx\n", (unsigned long)pgt);
1343 pgdp = pgt + pgd_index(gpa);
1344 p4dp = p4d_offset(pgdp, gpa);
1345 p4d = READ_ONCE(*p4dp);
1346 if (!(p4d_val(p4d) & _PAGE_PRESENT)) {
1347 gpa = (gpa & P4D_MASK) + P4D_SIZE;
1351 pudp = pud_offset(&p4d, gpa);
1352 pud = READ_ONCE(*pudp);
1353 if (!(pud_val(pud) & _PAGE_PRESENT)) {
1354 gpa = (gpa & PUD_MASK) + PUD_SIZE;
1357 if (pud_val(pud) & _PAGE_PTE) {
1363 pmdp = pmd_offset(&pud, gpa);
1364 pmd = READ_ONCE(*pmdp);
1365 if (!(pmd_val(pmd) & _PAGE_PRESENT)) {
1366 gpa = (gpa & PMD_MASK) + PMD_SIZE;
1369 if (pmd_val(pmd) & _PAGE_PTE) {
1375 ptep = pte_offset_kernel(&pmd, gpa);
1376 pte = pte_val(READ_ONCE(*ptep));
1377 if (!(pte & _PAGE_PRESENT)) {
1383 n = scnprintf(p->buf, sizeof(p->buf),
1384 " %lx: %lx %d\n", gpa, pte, shift);
1385 gpa += 1ul << shift;
1390 r = copy_to_user(buf, p->buf, n);
1405 kvmhv_put_nested(nested);
1408 mutex_unlock(&p->mutex);
1412 static ssize_t debugfs_radix_write(struct file *file, const char __user *buf,
1413 size_t len, loff_t *ppos)
1418 static const struct file_operations debugfs_radix_fops = {
1419 .owner = THIS_MODULE,
1420 .open = debugfs_radix_open,
1421 .release = debugfs_radix_release,
1422 .read = debugfs_radix_read,
1423 .write = debugfs_radix_write,
1424 .llseek = generic_file_llseek,
1427 void kvmhv_radix_debugfs_init(struct kvm *kvm)
1429 debugfs_create_file("radix", 0400, kvm->arch.debugfs_dir, kvm,
1430 &debugfs_radix_fops);
1433 int kvmppc_radix_init(void)
1435 unsigned long size = sizeof(void *) << RADIX_PTE_INDEX_SIZE;
1437 kvm_pte_cache = kmem_cache_create("kvm-pte", size, size, 0, pte_ctor);
1441 size = sizeof(void *) << RADIX_PMD_INDEX_SIZE;
1443 kvm_pmd_cache = kmem_cache_create("kvm-pmd", size, size, 0, pmd_ctor);
1444 if (!kvm_pmd_cache) {
1445 kmem_cache_destroy(kvm_pte_cache);
1452 void kvmppc_radix_exit(void)
1454 kmem_cache_destroy(kvm_pte_cache);
1455 kmem_cache_destroy(kvm_pmd_cache);