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>
24 #include <asm/plpar_wrappers.h>
27 * Supported radix tree geometry.
28 * Like p9, we support either 5 or 9 bits at the first (lowest) level,
29 * for a page size of 64k or 4k.
31 static int p9_supported_radix_bits[4] = { 5, 9, 9, 13 };
33 unsigned long __kvmhv_copy_tofrom_guest_radix(int lpid, int pid,
34 gva_t eaddr, void *to, void *from,
37 int old_pid, old_lpid;
38 unsigned long quadrant, ret = n;
41 /* Can't access quadrants 1 or 2 in non-HV mode, call the HV to do it */
42 if (kvmhv_on_pseries())
43 return plpar_hcall_norets(H_COPY_TOFROM_GUEST, lpid, pid, eaddr,
44 (to != NULL) ? __pa(to): 0,
45 (from != NULL) ? __pa(from): 0, n);
47 if (eaddr & (0xFFFUL << 52))
54 from = (void *) (eaddr | (quadrant << 62));
56 to = (void *) (eaddr | (quadrant << 62));
60 /* switch the lpid first to avoid running host with unallocated pid */
61 old_lpid = mfspr(SPRN_LPID);
63 mtspr(SPRN_LPID, lpid);
65 old_pid = mfspr(SPRN_PID);
73 ret = __copy_from_user_inatomic(to, (const void __user *)from, n);
75 ret = __copy_to_user_inatomic((void __user *)to, from, n);
78 /* switch the pid first to avoid running host with unallocated pid */
79 if (quadrant == 1 && pid != old_pid)
80 mtspr(SPRN_PID, old_pid);
82 mtspr(SPRN_LPID, old_lpid);
90 static long kvmhv_copy_tofrom_guest_radix(struct kvm_vcpu *vcpu, gva_t eaddr,
91 void *to, void *from, unsigned long n)
93 int lpid = vcpu->kvm->arch.lpid;
94 int pid = vcpu->arch.pid;
96 /* This would cause a data segment intr so don't allow the access */
97 if (eaddr & (0x3FFUL << 52))
100 /* Should we be using the nested lpid */
101 if (vcpu->arch.nested)
102 lpid = vcpu->arch.nested->shadow_lpid;
104 /* If accessing quadrant 3 then pid is expected to be 0 */
105 if (((eaddr >> 62) & 0x3) == 0x3)
108 eaddr &= ~(0xFFFUL << 52);
110 return __kvmhv_copy_tofrom_guest_radix(lpid, pid, eaddr, to, from, n);
113 long kvmhv_copy_from_guest_radix(struct kvm_vcpu *vcpu, gva_t eaddr, void *to,
118 ret = kvmhv_copy_tofrom_guest_radix(vcpu, eaddr, to, NULL, n);
120 memset(to + (n - ret), 0, ret);
125 long kvmhv_copy_to_guest_radix(struct kvm_vcpu *vcpu, gva_t eaddr, void *from,
128 return kvmhv_copy_tofrom_guest_radix(vcpu, eaddr, NULL, from, n);
131 int kvmppc_mmu_walk_radix_tree(struct kvm_vcpu *vcpu, gva_t eaddr,
132 struct kvmppc_pte *gpte, u64 root,
135 struct kvm *kvm = vcpu->kvm;
137 unsigned long rts, bits, offset, index;
141 rts = ((root & RTS1_MASK) >> (RTS1_SHIFT - 3)) |
142 ((root & RTS2_MASK) >> RTS2_SHIFT);
143 bits = root & RPDS_MASK;
144 base = root & RPDB_MASK;
148 /* Current implementations only support 52-bit space */
152 /* Walk each level of the radix tree */
153 for (level = 3; level >= 0; --level) {
155 /* Check a valid size */
156 if (level && bits != p9_supported_radix_bits[level])
158 if (level == 0 && !(bits == 5 || bits == 9))
161 index = (eaddr >> offset) & ((1UL << bits) - 1);
162 /* Check that low bits of page table base are zero */
163 if (base & ((1UL << (bits + 3)) - 1))
165 /* Read the entry from guest memory */
166 addr = base + (index * sizeof(rpte));
167 vcpu->srcu_idx = srcu_read_lock(&kvm->srcu);
168 ret = kvm_read_guest(kvm, addr, &rpte, sizeof(rpte));
169 srcu_read_unlock(&kvm->srcu, vcpu->srcu_idx);
175 pte = __be64_to_cpu(rpte);
176 if (!(pte & _PAGE_PRESENT))
178 /* Check if a leaf entry */
181 /* Get ready to walk the next level */
182 base = pte & RPDB_MASK;
183 bits = pte & RPDS_MASK;
186 /* Need a leaf at lowest level; 512GB pages not supported */
187 if (level < 0 || level == 3)
190 /* We found a valid leaf PTE */
191 /* Offset is now log base 2 of the page size */
192 gpa = pte & 0x01fffffffffff000ul;
193 if (gpa & ((1ul << offset) - 1))
195 gpa |= eaddr & ((1ul << offset) - 1);
196 for (ps = MMU_PAGE_4K; ps < MMU_PAGE_COUNT; ++ps)
197 if (offset == mmu_psize_defs[ps].shift)
199 gpte->page_size = ps;
200 gpte->page_shift = offset;
205 /* Work out permissions */
206 gpte->may_read = !!(pte & _PAGE_READ);
207 gpte->may_write = !!(pte & _PAGE_WRITE);
208 gpte->may_execute = !!(pte & _PAGE_EXEC);
210 gpte->rc = pte & (_PAGE_ACCESSED | _PAGE_DIRTY);
219 * Used to walk a partition or process table radix tree in guest memory
220 * Note: We exploit the fact that a partition table and a process
221 * table have the same layout, a partition-scoped page table and a
222 * process-scoped page table have the same layout, and the 2nd
223 * doubleword of a partition table entry has the same layout as
226 int kvmppc_mmu_radix_translate_table(struct kvm_vcpu *vcpu, gva_t eaddr,
227 struct kvmppc_pte *gpte, u64 table,
228 int table_index, u64 *pte_ret_p)
230 struct kvm *kvm = vcpu->kvm;
232 unsigned long size, ptbl, root;
233 struct prtb_entry entry;
235 if ((table & PRTS_MASK) > 24)
237 size = 1ul << ((table & PRTS_MASK) + 12);
239 /* Is the table big enough to contain this entry? */
240 if ((table_index * sizeof(entry)) >= size)
243 /* Read the table to find the root of the radix tree */
244 ptbl = (table & PRTB_MASK) + (table_index * sizeof(entry));
245 vcpu->srcu_idx = srcu_read_lock(&kvm->srcu);
246 ret = kvm_read_guest(kvm, ptbl, &entry, sizeof(entry));
247 srcu_read_unlock(&kvm->srcu, vcpu->srcu_idx);
251 /* Root is stored in the first double word */
252 root = be64_to_cpu(entry.prtb0);
254 return kvmppc_mmu_walk_radix_tree(vcpu, eaddr, gpte, root, pte_ret_p);
257 int kvmppc_mmu_radix_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
258 struct kvmppc_pte *gpte, bool data, bool iswrite)
264 /* Work out effective PID */
265 switch (eaddr >> 62) {
267 pid = vcpu->arch.pid;
276 ret = kvmppc_mmu_radix_translate_table(vcpu, eaddr, gpte,
277 vcpu->kvm->arch.process_table, pid, &pte);
281 /* Check privilege (applies only to process scoped translations) */
282 if (kvmppc_get_msr(vcpu) & MSR_PR) {
283 if (pte & _PAGE_PRIVILEGED) {
286 gpte->may_execute = 0;
289 if (!(pte & _PAGE_PRIVILEGED)) {
290 /* Check AMR/IAMR to see if strict mode is in force */
291 if (vcpu->arch.amr & (1ul << 62))
293 if (vcpu->arch.amr & (1ul << 63))
295 if (vcpu->arch.iamr & (1ul << 62))
296 gpte->may_execute = 0;
303 void kvmppc_radix_tlbie_page(struct kvm *kvm, unsigned long addr,
304 unsigned int pshift, unsigned int lpid)
306 unsigned long psize = PAGE_SIZE;
312 psize = 1UL << pshift;
316 addr &= ~(psize - 1);
318 if (!kvmhv_on_pseries()) {
319 radix__flush_tlb_lpid_page(lpid, addr, psize);
323 psi = shift_to_mmu_psize(pshift);
325 if (!firmware_has_feature(FW_FEATURE_RPT_INVALIDATE)) {
326 rb = addr | (mmu_get_ap(psi) << PPC_BITLSHIFT(58));
327 rc = plpar_hcall_norets(H_TLB_INVALIDATE, H_TLBIE_P1_ENC(0, 0, 1),
330 rc = pseries_rpt_invalidate(lpid, H_RPTI_TARGET_CMMU,
333 psize_to_rpti_pgsize(psi),
338 pr_err("KVM: TLB page invalidation hcall failed, rc=%ld\n", rc);
341 static void kvmppc_radix_flush_pwc(struct kvm *kvm, unsigned int lpid)
345 if (!kvmhv_on_pseries()) {
346 radix__flush_pwc_lpid(lpid);
350 if (!firmware_has_feature(FW_FEATURE_RPT_INVALIDATE))
351 rc = plpar_hcall_norets(H_TLB_INVALIDATE, H_TLBIE_P1_ENC(1, 0, 1),
352 lpid, TLBIEL_INVAL_SET_LPID);
354 rc = pseries_rpt_invalidate(lpid, H_RPTI_TARGET_CMMU,
356 H_RPTI_TYPE_PWC, H_RPTI_PAGE_ALL,
359 pr_err("KVM: TLB PWC invalidation hcall failed, rc=%ld\n", rc);
362 static unsigned long kvmppc_radix_update_pte(struct kvm *kvm, pte_t *ptep,
363 unsigned long clr, unsigned long set,
364 unsigned long addr, unsigned int shift)
366 return __radix_pte_update(ptep, clr, set);
369 static void kvmppc_radix_set_pte_at(struct kvm *kvm, unsigned long addr,
370 pte_t *ptep, pte_t pte)
372 radix__set_pte_at(kvm->mm, addr, ptep, pte, 0);
375 static struct kmem_cache *kvm_pte_cache;
376 static struct kmem_cache *kvm_pmd_cache;
378 static pte_t *kvmppc_pte_alloc(void)
382 pte = kmem_cache_alloc(kvm_pte_cache, GFP_KERNEL);
383 /* pmd_populate() will only reference _pa(pte). */
384 kmemleak_ignore(pte);
389 static void kvmppc_pte_free(pte_t *ptep)
391 kmem_cache_free(kvm_pte_cache, ptep);
394 static pmd_t *kvmppc_pmd_alloc(void)
398 pmd = kmem_cache_alloc(kvm_pmd_cache, GFP_KERNEL);
399 /* pud_populate() will only reference _pa(pmd). */
400 kmemleak_ignore(pmd);
405 static void kvmppc_pmd_free(pmd_t *pmdp)
407 kmem_cache_free(kvm_pmd_cache, pmdp);
410 /* Called with kvm->mmu_lock held */
411 void kvmppc_unmap_pte(struct kvm *kvm, pte_t *pte, unsigned long gpa,
413 const struct kvm_memory_slot *memslot,
418 unsigned long gfn = gpa >> PAGE_SHIFT;
419 unsigned long page_size = PAGE_SIZE;
422 old = kvmppc_radix_update_pte(kvm, pte, ~0UL, 0, gpa, shift);
423 kvmppc_radix_tlbie_page(kvm, gpa, shift, lpid);
425 /* The following only applies to L1 entries */
426 if (lpid != kvm->arch.lpid)
430 memslot = gfn_to_memslot(kvm, gfn);
434 if (shift) { /* 1GB or 2MB page */
435 page_size = 1ul << shift;
436 if (shift == PMD_SHIFT)
437 kvm->stat.num_2M_pages--;
438 else if (shift == PUD_SHIFT)
439 kvm->stat.num_1G_pages--;
442 gpa &= ~(page_size - 1);
443 hpa = old & PTE_RPN_MASK;
444 kvmhv_remove_nest_rmap_range(kvm, memslot, gpa, hpa, page_size);
446 if ((old & _PAGE_DIRTY) && memslot->dirty_bitmap)
447 kvmppc_update_dirty_map(memslot, gfn, page_size);
451 * kvmppc_free_p?d are used to free existing page tables, and recursively
452 * descend and clear and free children.
453 * Callers are responsible for flushing the PWC.
455 * When page tables are being unmapped/freed as part of page fault path
456 * (full == false), valid ptes are generally not expected; however, there
457 * is one situation where they arise, which is when dirty page logging is
458 * turned off for a memslot while the VM is running. The new memslot
459 * becomes visible to page faults before the memslot commit function
460 * gets to flush the memslot, which can lead to a 2MB page mapping being
461 * installed for a guest physical address where there are already 64kB
462 * (or 4kB) mappings (of sub-pages of the same 2MB page).
464 static void kvmppc_unmap_free_pte(struct kvm *kvm, pte_t *pte, bool full,
468 memset(pte, 0, sizeof(long) << RADIX_PTE_INDEX_SIZE);
473 for (it = 0; it < PTRS_PER_PTE; ++it, ++p) {
474 if (pte_val(*p) == 0)
476 kvmppc_unmap_pte(kvm, p,
477 pte_pfn(*p) << PAGE_SHIFT,
478 PAGE_SHIFT, NULL, lpid);
482 kvmppc_pte_free(pte);
485 static void kvmppc_unmap_free_pmd(struct kvm *kvm, pmd_t *pmd, bool full,
491 for (im = 0; im < PTRS_PER_PMD; ++im, ++p) {
492 if (!pmd_present(*p))
494 if (pmd_is_leaf(*p)) {
499 kvmppc_unmap_pte(kvm, (pte_t *)p,
500 pte_pfn(*(pte_t *)p) << PAGE_SHIFT,
501 PMD_SHIFT, NULL, lpid);
506 pte = pte_offset_map(p, 0);
507 kvmppc_unmap_free_pte(kvm, pte, full, lpid);
511 kvmppc_pmd_free(pmd);
514 static void kvmppc_unmap_free_pud(struct kvm *kvm, pud_t *pud,
520 for (iu = 0; iu < PTRS_PER_PUD; ++iu, ++p) {
521 if (!pud_present(*p))
523 if (pud_is_leaf(*p)) {
528 pmd = pmd_offset(p, 0);
529 kvmppc_unmap_free_pmd(kvm, pmd, true, lpid);
533 pud_free(kvm->mm, pud);
536 void kvmppc_free_pgtable_radix(struct kvm *kvm, pgd_t *pgd, unsigned int lpid)
540 for (ig = 0; ig < PTRS_PER_PGD; ++ig, ++pgd) {
541 p4d_t *p4d = p4d_offset(pgd, 0);
544 if (!p4d_present(*p4d))
546 pud = pud_offset(p4d, 0);
547 kvmppc_unmap_free_pud(kvm, pud, lpid);
552 void kvmppc_free_radix(struct kvm *kvm)
554 if (kvm->arch.pgtable) {
555 kvmppc_free_pgtable_radix(kvm, kvm->arch.pgtable,
557 pgd_free(kvm->mm, kvm->arch.pgtable);
558 kvm->arch.pgtable = NULL;
562 static void kvmppc_unmap_free_pmd_entry_table(struct kvm *kvm, pmd_t *pmd,
563 unsigned long gpa, unsigned int lpid)
565 pte_t *pte = pte_offset_kernel(pmd, 0);
568 * Clearing the pmd entry then flushing the PWC ensures that the pte
569 * page no longer be cached by the MMU, so can be freed without
570 * flushing the PWC again.
573 kvmppc_radix_flush_pwc(kvm, lpid);
575 kvmppc_unmap_free_pte(kvm, pte, false, lpid);
578 static void kvmppc_unmap_free_pud_entry_table(struct kvm *kvm, pud_t *pud,
579 unsigned long gpa, unsigned int lpid)
581 pmd_t *pmd = pmd_offset(pud, 0);
584 * Clearing the pud entry then flushing the PWC ensures that the pmd
585 * page and any children pte pages will no longer be cached by the MMU,
586 * so can be freed without flushing the PWC again.
589 kvmppc_radix_flush_pwc(kvm, lpid);
591 kvmppc_unmap_free_pmd(kvm, pmd, false, lpid);
595 * There are a number of bits which may differ between different faults to
596 * the same partition scope entry. RC bits, in the course of cleaning and
597 * aging. And the write bit can change, either the access could have been
598 * upgraded, or a read fault could happen concurrently with a write fault
599 * that sets those bits first.
601 #define PTE_BITS_MUST_MATCH (~(_PAGE_WRITE | _PAGE_DIRTY | _PAGE_ACCESSED))
603 int kvmppc_create_pte(struct kvm *kvm, pgd_t *pgtable, pte_t pte,
604 unsigned long gpa, unsigned int level,
605 unsigned long mmu_seq, unsigned int lpid,
606 unsigned long *rmapp, struct rmap_nested **n_rmap)
610 pud_t *pud, *new_pud = NULL;
611 pmd_t *pmd, *new_pmd = NULL;
612 pte_t *ptep, *new_ptep = NULL;
615 /* Traverse the guest's 2nd-level tree, allocate new levels needed */
616 pgd = pgtable + pgd_index(gpa);
617 p4d = p4d_offset(pgd, gpa);
620 if (p4d_present(*p4d))
621 pud = pud_offset(p4d, gpa);
623 new_pud = pud_alloc_one(kvm->mm, gpa);
626 if (pud && pud_present(*pud) && !pud_is_leaf(*pud))
627 pmd = pmd_offset(pud, gpa);
629 new_pmd = kvmppc_pmd_alloc();
631 if (level == 0 && !(pmd && pmd_present(*pmd) && !pmd_is_leaf(*pmd)))
632 new_ptep = kvmppc_pte_alloc();
634 /* Check if we might have been invalidated; let the guest retry if so */
635 spin_lock(&kvm->mmu_lock);
637 if (mmu_notifier_retry(kvm, mmu_seq))
640 /* Now traverse again under the lock and change the tree */
642 if (p4d_none(*p4d)) {
645 p4d_populate(kvm->mm, p4d, new_pud);
648 pud = pud_offset(p4d, gpa);
649 if (pud_is_leaf(*pud)) {
650 unsigned long hgpa = gpa & PUD_MASK;
652 /* Check if we raced and someone else has set the same thing */
654 if (pud_raw(*pud) == pte_raw(pte)) {
658 /* Valid 1GB page here already, add our extra bits */
659 WARN_ON_ONCE((pud_val(*pud) ^ pte_val(pte)) &
660 PTE_BITS_MUST_MATCH);
661 kvmppc_radix_update_pte(kvm, (pte_t *)pud,
662 0, pte_val(pte), hgpa, PUD_SHIFT);
667 * If we raced with another CPU which has just put
668 * a 1GB pte in after we saw a pmd page, try again.
674 /* Valid 1GB page here already, remove it */
675 kvmppc_unmap_pte(kvm, (pte_t *)pud, hgpa, PUD_SHIFT, NULL,
679 if (!pud_none(*pud)) {
681 * There's a page table page here, but we wanted to
682 * install a large page, so remove and free the page
685 kvmppc_unmap_free_pud_entry_table(kvm, pud, gpa, lpid);
687 kvmppc_radix_set_pte_at(kvm, gpa, (pte_t *)pud, pte);
689 kvmhv_insert_nest_rmap(kvm, rmapp, n_rmap);
693 if (pud_none(*pud)) {
696 pud_populate(kvm->mm, pud, new_pmd);
699 pmd = pmd_offset(pud, gpa);
700 if (pmd_is_leaf(*pmd)) {
701 unsigned long lgpa = gpa & PMD_MASK;
703 /* Check if we raced and someone else has set the same thing */
705 if (pmd_raw(*pmd) == pte_raw(pte)) {
709 /* Valid 2MB page here already, add our extra bits */
710 WARN_ON_ONCE((pmd_val(*pmd) ^ pte_val(pte)) &
711 PTE_BITS_MUST_MATCH);
712 kvmppc_radix_update_pte(kvm, pmdp_ptep(pmd),
713 0, pte_val(pte), lgpa, PMD_SHIFT);
719 * If we raced with another CPU which has just put
720 * a 2MB pte in after we saw a pte page, try again.
726 /* Valid 2MB page here already, remove it */
727 kvmppc_unmap_pte(kvm, pmdp_ptep(pmd), lgpa, PMD_SHIFT, NULL,
731 if (!pmd_none(*pmd)) {
733 * There's a page table page here, but we wanted to
734 * install a large page, so remove and free the page
737 kvmppc_unmap_free_pmd_entry_table(kvm, pmd, gpa, lpid);
739 kvmppc_radix_set_pte_at(kvm, gpa, pmdp_ptep(pmd), pte);
741 kvmhv_insert_nest_rmap(kvm, rmapp, n_rmap);
745 if (pmd_none(*pmd)) {
748 pmd_populate(kvm->mm, pmd, new_ptep);
751 ptep = pte_offset_kernel(pmd, gpa);
752 if (pte_present(*ptep)) {
753 /* Check if someone else set the same thing */
754 if (pte_raw(*ptep) == pte_raw(pte)) {
758 /* Valid page here already, add our extra bits */
759 WARN_ON_ONCE((pte_val(*ptep) ^ pte_val(pte)) &
760 PTE_BITS_MUST_MATCH);
761 kvmppc_radix_update_pte(kvm, ptep, 0, pte_val(pte), gpa, 0);
765 kvmppc_radix_set_pte_at(kvm, gpa, ptep, pte);
767 kvmhv_insert_nest_rmap(kvm, rmapp, n_rmap);
771 spin_unlock(&kvm->mmu_lock);
773 pud_free(kvm->mm, new_pud);
775 kvmppc_pmd_free(new_pmd);
777 kvmppc_pte_free(new_ptep);
781 bool kvmppc_hv_handle_set_rc(struct kvm *kvm, bool nested, bool writing,
782 unsigned long gpa, unsigned int lpid)
784 unsigned long pgflags;
789 * Need to set an R or C bit in the 2nd-level tables;
790 * since we are just helping out the hardware here,
791 * it is sufficient to do what the hardware does.
793 pgflags = _PAGE_ACCESSED;
795 pgflags |= _PAGE_DIRTY;
798 ptep = find_kvm_nested_guest_pte(kvm, lpid, gpa, &shift);
800 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
802 if (ptep && pte_present(*ptep) && (!writing || pte_write(*ptep))) {
803 kvmppc_radix_update_pte(kvm, ptep, 0, pgflags, gpa, shift);
809 int kvmppc_book3s_instantiate_page(struct kvm_vcpu *vcpu,
811 struct kvm_memory_slot *memslot,
812 bool writing, bool kvm_ro,
813 pte_t *inserted_pte, unsigned int *levelp)
815 struct kvm *kvm = vcpu->kvm;
816 struct page *page = NULL;
817 unsigned long mmu_seq;
818 unsigned long hva, gfn = gpa >> PAGE_SHIFT;
819 bool upgrade_write = false;
820 bool *upgrade_p = &upgrade_write;
822 unsigned int shift, level;
826 /* used to check for invalidations in progress */
827 mmu_seq = kvm->mmu_notifier_seq;
831 * Do a fast check first, since __gfn_to_pfn_memslot doesn't
832 * do it with !atomic && !async, which is how we call it.
833 * We always ask for write permission since the common case
834 * is that the page is writable.
836 hva = gfn_to_hva_memslot(memslot, gfn);
837 if (!kvm_ro && get_user_page_fast_only(hva, FOLL_WRITE, &page)) {
838 upgrade_write = true;
842 /* Call KVM generic code to do the slow-path check */
843 pfn = __gfn_to_pfn_memslot(memslot, gfn, false, NULL,
844 writing, upgrade_p, NULL);
845 if (is_error_noslot_pfn(pfn))
848 if (pfn_valid(pfn)) {
849 page = pfn_to_page(pfn);
850 if (PageReserved(page))
856 * Read the PTE from the process' radix tree and use that
857 * so we get the shift and attribute bits.
859 spin_lock(&kvm->mmu_lock);
860 ptep = find_kvm_host_pte(kvm, mmu_seq, hva, &shift);
863 pte = READ_ONCE(*ptep);
864 spin_unlock(&kvm->mmu_lock);
866 * If the PTE disappeared temporarily due to a THP
867 * collapse, just return and let the guest try again.
869 if (!pte_present(pte)) {
875 /* If we're logging dirty pages, always map single pages */
876 large_enable = !(memslot->flags & KVM_MEM_LOG_DIRTY_PAGES);
878 /* Get pte level from shift/size */
879 if (large_enable && shift == PUD_SHIFT &&
880 (gpa & (PUD_SIZE - PAGE_SIZE)) ==
881 (hva & (PUD_SIZE - PAGE_SIZE))) {
883 } else if (large_enable && shift == PMD_SHIFT &&
884 (gpa & (PMD_SIZE - PAGE_SIZE)) ==
885 (hva & (PMD_SIZE - PAGE_SIZE))) {
889 if (shift > PAGE_SHIFT) {
891 * If the pte maps more than one page, bring over
892 * bits from the virtual address to get the real
893 * address of the specific single page we want.
895 unsigned long rpnmask = (1ul << shift) - PAGE_SIZE;
896 pte = __pte(pte_val(pte) | (hva & rpnmask));
900 pte = __pte(pte_val(pte) | _PAGE_EXEC | _PAGE_ACCESSED);
901 if (writing || upgrade_write) {
902 if (pte_val(pte) & _PAGE_WRITE)
903 pte = __pte(pte_val(pte) | _PAGE_DIRTY);
905 pte = __pte(pte_val(pte) & ~(_PAGE_WRITE | _PAGE_DIRTY));
908 /* Allocate space in the tree and write the PTE */
909 ret = kvmppc_create_pte(kvm, kvm->arch.pgtable, pte, gpa, level,
910 mmu_seq, kvm->arch.lpid, NULL, NULL);
917 if (!ret && (pte_val(pte) & _PAGE_WRITE))
918 set_page_dirty_lock(page);
922 /* Increment number of large pages if we (successfully) inserted one */
925 kvm->stat.num_2M_pages++;
927 kvm->stat.num_1G_pages++;
933 int kvmppc_book3s_radix_page_fault(struct kvm_vcpu *vcpu,
934 unsigned long ea, unsigned long dsisr)
936 struct kvm *kvm = vcpu->kvm;
937 unsigned long gpa, gfn;
938 struct kvm_memory_slot *memslot;
940 bool writing = !!(dsisr & DSISR_ISSTORE);
943 /* Check for unusual errors */
944 if (dsisr & DSISR_UNSUPP_MMU) {
945 pr_err("KVM: Got unsupported MMU fault\n");
948 if (dsisr & DSISR_BADACCESS) {
949 /* Reflect to the guest as DSI */
950 pr_err("KVM: Got radix HV page fault with DSISR=%lx\n", dsisr);
951 kvmppc_core_queue_data_storage(vcpu, ea, dsisr);
955 /* Translate the logical address */
956 gpa = vcpu->arch.fault_gpa & ~0xfffUL;
957 gpa &= ~0xF000000000000000ul;
958 gfn = gpa >> PAGE_SHIFT;
959 if (!(dsisr & DSISR_PRTABLE_FAULT))
962 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
963 return kvmppc_send_page_to_uv(kvm, gfn);
965 /* Get the corresponding memslot */
966 memslot = gfn_to_memslot(kvm, gfn);
968 /* No memslot means it's an emulated MMIO region */
969 if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID)) {
970 if (dsisr & (DSISR_PRTABLE_FAULT | DSISR_BADACCESS |
973 * Bad address in guest page table tree, or other
974 * unusual error - reflect it to the guest as DSI.
976 kvmppc_core_queue_data_storage(vcpu, ea, dsisr);
979 return kvmppc_hv_emulate_mmio(vcpu, gpa, ea, writing);
982 if (memslot->flags & KVM_MEM_READONLY) {
984 /* give the guest a DSI */
985 kvmppc_core_queue_data_storage(vcpu, ea, DSISR_ISSTORE |
992 /* Failed to set the reference/change bits */
993 if (dsisr & DSISR_SET_RC) {
994 spin_lock(&kvm->mmu_lock);
995 if (kvmppc_hv_handle_set_rc(kvm, false, writing,
996 gpa, kvm->arch.lpid))
997 dsisr &= ~DSISR_SET_RC;
998 spin_unlock(&kvm->mmu_lock);
1000 if (!(dsisr & (DSISR_BAD_FAULT_64S | DSISR_NOHPTE |
1001 DSISR_PROTFAULT | DSISR_SET_RC)))
1002 return RESUME_GUEST;
1005 /* Try to insert a pte */
1006 ret = kvmppc_book3s_instantiate_page(vcpu, gpa, memslot, writing,
1007 kvm_ro, NULL, NULL);
1009 if (ret == 0 || ret == -EAGAIN)
1014 /* Called with kvm->mmu_lock held */
1015 void kvm_unmap_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
1019 unsigned long gpa = gfn << PAGE_SHIFT;
1022 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE) {
1023 uv_page_inval(kvm->arch.lpid, gpa, PAGE_SHIFT);
1027 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1028 if (ptep && pte_present(*ptep))
1029 kvmppc_unmap_pte(kvm, ptep, gpa, shift, memslot,
1033 /* Called with kvm->mmu_lock held */
1034 bool kvm_age_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
1038 unsigned long gpa = gfn << PAGE_SHIFT;
1041 unsigned long old, *rmapp;
1043 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1046 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1047 if (ptep && pte_present(*ptep) && pte_young(*ptep)) {
1048 old = kvmppc_radix_update_pte(kvm, ptep, _PAGE_ACCESSED, 0,
1050 /* XXX need to flush tlb here? */
1051 /* Also clear bit in ptes in shadow pgtable for nested guests */
1052 rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1053 kvmhv_update_nest_rmap_rc_list(kvm, rmapp, _PAGE_ACCESSED, 0,
1061 /* Called with kvm->mmu_lock held */
1062 bool kvm_test_age_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
1067 unsigned long gpa = gfn << PAGE_SHIFT;
1071 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1074 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1075 if (ptep && pte_present(*ptep) && pte_young(*ptep))
1080 /* Returns the number of PAGE_SIZE pages that are dirty */
1081 static int kvm_radix_test_clear_dirty(struct kvm *kvm,
1082 struct kvm_memory_slot *memslot, int pagenum)
1084 unsigned long gfn = memslot->base_gfn + pagenum;
1085 unsigned long gpa = gfn << PAGE_SHIFT;
1089 unsigned long old, *rmapp;
1091 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1095 * For performance reasons we don't hold kvm->mmu_lock while walking the
1096 * partition scoped table.
1098 ptep = find_kvm_secondary_pte_unlocked(kvm, gpa, &shift);
1102 pte = READ_ONCE(*ptep);
1103 if (pte_present(pte) && pte_dirty(pte)) {
1104 spin_lock(&kvm->mmu_lock);
1106 * Recheck the pte again
1108 if (pte_val(pte) != pte_val(*ptep)) {
1110 * We have KVM_MEM_LOG_DIRTY_PAGES enabled. Hence we can
1111 * only find PAGE_SIZE pte entries here. We can continue
1112 * to use the pte addr returned by above page table
1115 if (!pte_present(*ptep) || !pte_dirty(*ptep)) {
1116 spin_unlock(&kvm->mmu_lock);
1123 old = kvmppc_radix_update_pte(kvm, ptep, _PAGE_DIRTY, 0,
1125 kvmppc_radix_tlbie_page(kvm, gpa, shift, kvm->arch.lpid);
1126 /* Also clear bit in ptes in shadow pgtable for nested guests */
1127 rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1128 kvmhv_update_nest_rmap_rc_list(kvm, rmapp, _PAGE_DIRTY, 0,
1131 spin_unlock(&kvm->mmu_lock);
1136 long kvmppc_hv_get_dirty_log_radix(struct kvm *kvm,
1137 struct kvm_memory_slot *memslot, unsigned long *map)
1142 for (i = 0; i < memslot->npages; i = j) {
1143 npages = kvm_radix_test_clear_dirty(kvm, memslot, i);
1146 * Note that if npages > 0 then i must be a multiple of npages,
1147 * since huge pages are only used to back the guest at guest
1148 * real addresses that are a multiple of their size.
1149 * Since we have at most one PTE covering any given guest
1150 * real address, if npages > 1 we can skip to i + npages.
1154 set_dirty_bits(map, i, npages);
1161 void kvmppc_radix_flush_memslot(struct kvm *kvm,
1162 const struct kvm_memory_slot *memslot)
1169 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_START)
1170 kvmppc_uvmem_drop_pages(memslot, kvm, true);
1172 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1175 gpa = memslot->base_gfn << PAGE_SHIFT;
1176 spin_lock(&kvm->mmu_lock);
1177 for (n = memslot->npages; n; --n) {
1178 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1179 if (ptep && pte_present(*ptep))
1180 kvmppc_unmap_pte(kvm, ptep, gpa, shift, memslot,
1185 * Increase the mmu notifier sequence number to prevent any page
1186 * fault that read the memslot earlier from writing a PTE.
1188 kvm->mmu_notifier_seq++;
1189 spin_unlock(&kvm->mmu_lock);
1192 static void add_rmmu_ap_encoding(struct kvm_ppc_rmmu_info *info,
1193 int psize, int *indexp)
1195 if (!mmu_psize_defs[psize].shift)
1197 info->ap_encodings[*indexp] = mmu_psize_defs[psize].shift |
1198 (mmu_psize_defs[psize].ap << 29);
1202 int kvmhv_get_rmmu_info(struct kvm *kvm, struct kvm_ppc_rmmu_info *info)
1206 if (!radix_enabled())
1208 memset(info, 0, sizeof(*info));
1211 info->geometries[0].page_shift = 12;
1212 info->geometries[0].level_bits[0] = 9;
1213 for (i = 1; i < 4; ++i)
1214 info->geometries[0].level_bits[i] = p9_supported_radix_bits[i];
1216 info->geometries[1].page_shift = 16;
1217 for (i = 0; i < 4; ++i)
1218 info->geometries[1].level_bits[i] = p9_supported_radix_bits[i];
1221 add_rmmu_ap_encoding(info, MMU_PAGE_4K, &i);
1222 add_rmmu_ap_encoding(info, MMU_PAGE_64K, &i);
1223 add_rmmu_ap_encoding(info, MMU_PAGE_2M, &i);
1224 add_rmmu_ap_encoding(info, MMU_PAGE_1G, &i);
1229 int kvmppc_init_vm_radix(struct kvm *kvm)
1231 kvm->arch.pgtable = pgd_alloc(kvm->mm);
1232 if (!kvm->arch.pgtable)
1237 static void pte_ctor(void *addr)
1239 memset(addr, 0, RADIX_PTE_TABLE_SIZE);
1242 static void pmd_ctor(void *addr)
1244 memset(addr, 0, RADIX_PMD_TABLE_SIZE);
1247 struct debugfs_radix_state {
1258 static int debugfs_radix_open(struct inode *inode, struct file *file)
1260 struct kvm *kvm = inode->i_private;
1261 struct debugfs_radix_state *p;
1263 p = kzalloc(sizeof(*p), GFP_KERNEL);
1269 mutex_init(&p->mutex);
1270 file->private_data = p;
1272 return nonseekable_open(inode, file);
1275 static int debugfs_radix_release(struct inode *inode, struct file *file)
1277 struct debugfs_radix_state *p = file->private_data;
1279 kvm_put_kvm(p->kvm);
1284 static ssize_t debugfs_radix_read(struct file *file, char __user *buf,
1285 size_t len, loff_t *ppos)
1287 struct debugfs_radix_state *p = file->private_data;
1293 struct kvm_nested_guest *nested;
1303 if (!kvm_is_radix(kvm))
1306 ret = mutex_lock_interruptible(&p->mutex);
1310 if (p->chars_left) {
1314 r = copy_to_user(buf, p->buf + p->buf_index, n);
1331 while (len != 0 && p->lpid >= 0) {
1332 if (gpa >= RADIX_PGTABLE_RANGE) {
1336 kvmhv_put_nested(nested);
1339 p->lpid = kvmhv_nested_next_lpid(kvm, p->lpid);
1346 pgt = kvm->arch.pgtable;
1348 nested = kvmhv_get_nested(kvm, p->lpid, false);
1350 gpa = RADIX_PGTABLE_RANGE;
1353 pgt = nested->shadow_pgtable;
1359 n = scnprintf(p->buf, sizeof(p->buf),
1360 "\nNested LPID %d: ", p->lpid);
1361 n += scnprintf(p->buf + n, sizeof(p->buf) - n,
1362 "pgdir: %lx\n", (unsigned long)pgt);
1367 pgdp = pgt + pgd_index(gpa);
1368 p4dp = p4d_offset(pgdp, gpa);
1369 p4d = READ_ONCE(*p4dp);
1370 if (!(p4d_val(p4d) & _PAGE_PRESENT)) {
1371 gpa = (gpa & P4D_MASK) + P4D_SIZE;
1375 pudp = pud_offset(&p4d, gpa);
1376 pud = READ_ONCE(*pudp);
1377 if (!(pud_val(pud) & _PAGE_PRESENT)) {
1378 gpa = (gpa & PUD_MASK) + PUD_SIZE;
1381 if (pud_val(pud) & _PAGE_PTE) {
1387 pmdp = pmd_offset(&pud, gpa);
1388 pmd = READ_ONCE(*pmdp);
1389 if (!(pmd_val(pmd) & _PAGE_PRESENT)) {
1390 gpa = (gpa & PMD_MASK) + PMD_SIZE;
1393 if (pmd_val(pmd) & _PAGE_PTE) {
1399 ptep = pte_offset_kernel(&pmd, gpa);
1400 pte = pte_val(READ_ONCE(*ptep));
1401 if (!(pte & _PAGE_PRESENT)) {
1407 n = scnprintf(p->buf, sizeof(p->buf),
1408 " %lx: %lx %d\n", gpa, pte, shift);
1409 gpa += 1ul << shift;
1414 r = copy_to_user(buf, p->buf, n);
1429 kvmhv_put_nested(nested);
1432 mutex_unlock(&p->mutex);
1436 static ssize_t debugfs_radix_write(struct file *file, const char __user *buf,
1437 size_t len, loff_t *ppos)
1442 static const struct file_operations debugfs_radix_fops = {
1443 .owner = THIS_MODULE,
1444 .open = debugfs_radix_open,
1445 .release = debugfs_radix_release,
1446 .read = debugfs_radix_read,
1447 .write = debugfs_radix_write,
1448 .llseek = generic_file_llseek,
1451 void kvmhv_radix_debugfs_init(struct kvm *kvm)
1453 debugfs_create_file("radix", 0400, kvm->arch.debugfs_dir, kvm,
1454 &debugfs_radix_fops);
1457 int kvmppc_radix_init(void)
1459 unsigned long size = sizeof(void *) << RADIX_PTE_INDEX_SIZE;
1461 kvm_pte_cache = kmem_cache_create("kvm-pte", size, size, 0, pte_ctor);
1465 size = sizeof(void *) << RADIX_PMD_INDEX_SIZE;
1467 kvm_pmd_cache = kmem_cache_create("kvm-pmd", size, size, 0, pmd_ctor);
1468 if (!kvm_pmd_cache) {
1469 kmem_cache_destroy(kvm_pte_cache);
1476 void kvmppc_radix_exit(void)
1478 kmem_cache_destroy(kvm_pte_cache);
1479 kmem_cache_destroy(kvm_pmd_cache);