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 asm volatile("hwsync" ::: "memory");
62 /* switch the lpid first to avoid running host with unallocated pid */
63 old_lpid = mfspr(SPRN_LPID);
65 mtspr(SPRN_LPID, lpid);
67 old_pid = mfspr(SPRN_PID);
75 ret = __copy_from_user_inatomic(to, (const void __user *)from, n);
77 ret = __copy_to_user_inatomic((void __user *)to, from, n);
80 asm volatile("hwsync" ::: "memory");
82 /* switch the pid first to avoid running host with unallocated pid */
83 if (quadrant == 1 && pid != old_pid)
84 mtspr(SPRN_PID, old_pid);
86 mtspr(SPRN_LPID, old_lpid);
94 static long kvmhv_copy_tofrom_guest_radix(struct kvm_vcpu *vcpu, gva_t eaddr,
95 void *to, void *from, unsigned long n)
97 int lpid = vcpu->kvm->arch.lpid;
98 int pid = vcpu->arch.pid;
100 /* This would cause a data segment intr so don't allow the access */
101 if (eaddr & (0x3FFUL << 52))
104 /* Should we be using the nested lpid */
105 if (vcpu->arch.nested)
106 lpid = vcpu->arch.nested->shadow_lpid;
108 /* If accessing quadrant 3 then pid is expected to be 0 */
109 if (((eaddr >> 62) & 0x3) == 0x3)
112 eaddr &= ~(0xFFFUL << 52);
114 return __kvmhv_copy_tofrom_guest_radix(lpid, pid, eaddr, to, from, n);
117 long kvmhv_copy_from_guest_radix(struct kvm_vcpu *vcpu, gva_t eaddr, void *to,
122 ret = kvmhv_copy_tofrom_guest_radix(vcpu, eaddr, to, NULL, n);
124 memset(to + (n - ret), 0, ret);
129 long kvmhv_copy_to_guest_radix(struct kvm_vcpu *vcpu, gva_t eaddr, void *from,
132 return kvmhv_copy_tofrom_guest_radix(vcpu, eaddr, NULL, from, n);
135 int kvmppc_mmu_walk_radix_tree(struct kvm_vcpu *vcpu, gva_t eaddr,
136 struct kvmppc_pte *gpte, u64 root,
139 struct kvm *kvm = vcpu->kvm;
141 unsigned long rts, bits, offset, index;
145 rts = ((root & RTS1_MASK) >> (RTS1_SHIFT - 3)) |
146 ((root & RTS2_MASK) >> RTS2_SHIFT);
147 bits = root & RPDS_MASK;
148 base = root & RPDB_MASK;
152 /* Current implementations only support 52-bit space */
156 /* Walk each level of the radix tree */
157 for (level = 3; level >= 0; --level) {
159 /* Check a valid size */
160 if (level && bits != p9_supported_radix_bits[level])
162 if (level == 0 && !(bits == 5 || bits == 9))
165 index = (eaddr >> offset) & ((1UL << bits) - 1);
166 /* Check that low bits of page table base are zero */
167 if (base & ((1UL << (bits + 3)) - 1))
169 /* Read the entry from guest memory */
170 addr = base + (index * sizeof(rpte));
172 kvm_vcpu_srcu_read_lock(vcpu);
173 ret = kvm_read_guest(kvm, addr, &rpte, sizeof(rpte));
174 kvm_vcpu_srcu_read_unlock(vcpu);
180 pte = __be64_to_cpu(rpte);
181 if (!(pte & _PAGE_PRESENT))
183 /* Check if a leaf entry */
186 /* Get ready to walk the next level */
187 base = pte & RPDB_MASK;
188 bits = pte & RPDS_MASK;
191 /* Need a leaf at lowest level; 512GB pages not supported */
192 if (level < 0 || level == 3)
195 /* We found a valid leaf PTE */
196 /* Offset is now log base 2 of the page size */
197 gpa = pte & 0x01fffffffffff000ul;
198 if (gpa & ((1ul << offset) - 1))
200 gpa |= eaddr & ((1ul << offset) - 1);
201 for (ps = MMU_PAGE_4K; ps < MMU_PAGE_COUNT; ++ps)
202 if (offset == mmu_psize_defs[ps].shift)
204 gpte->page_size = ps;
205 gpte->page_shift = offset;
210 /* Work out permissions */
211 gpte->may_read = !!(pte & _PAGE_READ);
212 gpte->may_write = !!(pte & _PAGE_WRITE);
213 gpte->may_execute = !!(pte & _PAGE_EXEC);
215 gpte->rc = pte & (_PAGE_ACCESSED | _PAGE_DIRTY);
224 * Used to walk a partition or process table radix tree in guest memory
225 * Note: We exploit the fact that a partition table and a process
226 * table have the same layout, a partition-scoped page table and a
227 * process-scoped page table have the same layout, and the 2nd
228 * doubleword of a partition table entry has the same layout as
231 int kvmppc_mmu_radix_translate_table(struct kvm_vcpu *vcpu, gva_t eaddr,
232 struct kvmppc_pte *gpte, u64 table,
233 int table_index, u64 *pte_ret_p)
235 struct kvm *kvm = vcpu->kvm;
237 unsigned long size, ptbl, root;
238 struct prtb_entry entry;
240 if ((table & PRTS_MASK) > 24)
242 size = 1ul << ((table & PRTS_MASK) + 12);
244 /* Is the table big enough to contain this entry? */
245 if ((table_index * sizeof(entry)) >= size)
248 /* Read the table to find the root of the radix tree */
249 ptbl = (table & PRTB_MASK) + (table_index * sizeof(entry));
250 kvm_vcpu_srcu_read_lock(vcpu);
251 ret = kvm_read_guest(kvm, ptbl, &entry, sizeof(entry));
252 kvm_vcpu_srcu_read_unlock(vcpu);
256 /* Root is stored in the first double word */
257 root = be64_to_cpu(entry.prtb0);
259 return kvmppc_mmu_walk_radix_tree(vcpu, eaddr, gpte, root, pte_ret_p);
262 int kvmppc_mmu_radix_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
263 struct kvmppc_pte *gpte, bool data, bool iswrite)
269 /* Work out effective PID */
270 switch (eaddr >> 62) {
272 pid = vcpu->arch.pid;
281 ret = kvmppc_mmu_radix_translate_table(vcpu, eaddr, gpte,
282 vcpu->kvm->arch.process_table, pid, &pte);
286 /* Check privilege (applies only to process scoped translations) */
287 if (kvmppc_get_msr(vcpu) & MSR_PR) {
288 if (pte & _PAGE_PRIVILEGED) {
291 gpte->may_execute = 0;
294 if (!(pte & _PAGE_PRIVILEGED)) {
295 /* Check AMR/IAMR to see if strict mode is in force */
296 if (vcpu->arch.amr & (1ul << 62))
298 if (vcpu->arch.amr & (1ul << 63))
300 if (vcpu->arch.iamr & (1ul << 62))
301 gpte->may_execute = 0;
308 void kvmppc_radix_tlbie_page(struct kvm *kvm, unsigned long addr,
309 unsigned int pshift, unsigned int lpid)
311 unsigned long psize = PAGE_SIZE;
317 psize = 1UL << pshift;
321 addr &= ~(psize - 1);
323 if (!kvmhv_on_pseries()) {
324 radix__flush_tlb_lpid_page(lpid, addr, psize);
328 psi = shift_to_mmu_psize(pshift);
330 if (!firmware_has_feature(FW_FEATURE_RPT_INVALIDATE)) {
331 rb = addr | (mmu_get_ap(psi) << PPC_BITLSHIFT(58));
332 rc = plpar_hcall_norets(H_TLB_INVALIDATE, H_TLBIE_P1_ENC(0, 0, 1),
335 rc = pseries_rpt_invalidate(lpid, H_RPTI_TARGET_CMMU,
338 psize_to_rpti_pgsize(psi),
343 pr_err("KVM: TLB page invalidation hcall failed, rc=%ld\n", rc);
346 static void kvmppc_radix_flush_pwc(struct kvm *kvm, unsigned int lpid)
350 if (!kvmhv_on_pseries()) {
351 radix__flush_pwc_lpid(lpid);
355 if (!firmware_has_feature(FW_FEATURE_RPT_INVALIDATE))
356 rc = plpar_hcall_norets(H_TLB_INVALIDATE, H_TLBIE_P1_ENC(1, 0, 1),
357 lpid, TLBIEL_INVAL_SET_LPID);
359 rc = pseries_rpt_invalidate(lpid, H_RPTI_TARGET_CMMU,
361 H_RPTI_TYPE_PWC, H_RPTI_PAGE_ALL,
364 pr_err("KVM: TLB PWC invalidation hcall failed, rc=%ld\n", rc);
367 static unsigned long kvmppc_radix_update_pte(struct kvm *kvm, pte_t *ptep,
368 unsigned long clr, unsigned long set,
369 unsigned long addr, unsigned int shift)
371 return __radix_pte_update(ptep, clr, set);
374 static void kvmppc_radix_set_pte_at(struct kvm *kvm, unsigned long addr,
375 pte_t *ptep, pte_t pte)
377 radix__set_pte_at(kvm->mm, addr, ptep, pte, 0);
380 static struct kmem_cache *kvm_pte_cache;
381 static struct kmem_cache *kvm_pmd_cache;
383 static pte_t *kvmppc_pte_alloc(void)
387 pte = kmem_cache_alloc(kvm_pte_cache, GFP_KERNEL);
388 /* pmd_populate() will only reference _pa(pte). */
389 kmemleak_ignore(pte);
394 static void kvmppc_pte_free(pte_t *ptep)
396 kmem_cache_free(kvm_pte_cache, ptep);
399 static pmd_t *kvmppc_pmd_alloc(void)
403 pmd = kmem_cache_alloc(kvm_pmd_cache, GFP_KERNEL);
404 /* pud_populate() will only reference _pa(pmd). */
405 kmemleak_ignore(pmd);
410 static void kvmppc_pmd_free(pmd_t *pmdp)
412 kmem_cache_free(kvm_pmd_cache, pmdp);
415 /* Called with kvm->mmu_lock held */
416 void kvmppc_unmap_pte(struct kvm *kvm, pte_t *pte, unsigned long gpa,
418 const struct kvm_memory_slot *memslot,
423 unsigned long gfn = gpa >> PAGE_SHIFT;
424 unsigned long page_size = PAGE_SIZE;
427 old = kvmppc_radix_update_pte(kvm, pte, ~0UL, 0, gpa, shift);
428 kvmppc_radix_tlbie_page(kvm, gpa, shift, lpid);
430 /* The following only applies to L1 entries */
431 if (lpid != kvm->arch.lpid)
435 memslot = gfn_to_memslot(kvm, gfn);
439 if (shift) { /* 1GB or 2MB page */
440 page_size = 1ul << shift;
441 if (shift == PMD_SHIFT)
442 kvm->stat.num_2M_pages--;
443 else if (shift == PUD_SHIFT)
444 kvm->stat.num_1G_pages--;
447 gpa &= ~(page_size - 1);
448 hpa = old & PTE_RPN_MASK;
449 kvmhv_remove_nest_rmap_range(kvm, memslot, gpa, hpa, page_size);
451 if ((old & _PAGE_DIRTY) && memslot->dirty_bitmap)
452 kvmppc_update_dirty_map(memslot, gfn, page_size);
456 * kvmppc_free_p?d are used to free existing page tables, and recursively
457 * descend and clear and free children.
458 * Callers are responsible for flushing the PWC.
460 * When page tables are being unmapped/freed as part of page fault path
461 * (full == false), valid ptes are generally not expected; however, there
462 * is one situation where they arise, which is when dirty page logging is
463 * turned off for a memslot while the VM is running. The new memslot
464 * becomes visible to page faults before the memslot commit function
465 * gets to flush the memslot, which can lead to a 2MB page mapping being
466 * installed for a guest physical address where there are already 64kB
467 * (or 4kB) mappings (of sub-pages of the same 2MB page).
469 static void kvmppc_unmap_free_pte(struct kvm *kvm, pte_t *pte, bool full,
473 memset(pte, 0, sizeof(long) << RADIX_PTE_INDEX_SIZE);
478 for (it = 0; it < PTRS_PER_PTE; ++it, ++p) {
479 if (pte_val(*p) == 0)
481 kvmppc_unmap_pte(kvm, p,
482 pte_pfn(*p) << PAGE_SHIFT,
483 PAGE_SHIFT, NULL, lpid);
487 kvmppc_pte_free(pte);
490 static void kvmppc_unmap_free_pmd(struct kvm *kvm, pmd_t *pmd, bool full,
496 for (im = 0; im < PTRS_PER_PMD; ++im, ++p) {
497 if (!pmd_present(*p))
499 if (pmd_is_leaf(*p)) {
504 kvmppc_unmap_pte(kvm, (pte_t *)p,
505 pte_pfn(*(pte_t *)p) << PAGE_SHIFT,
506 PMD_SHIFT, NULL, lpid);
511 pte = pte_offset_map(p, 0);
512 kvmppc_unmap_free_pte(kvm, pte, full, lpid);
516 kvmppc_pmd_free(pmd);
519 static void kvmppc_unmap_free_pud(struct kvm *kvm, pud_t *pud,
525 for (iu = 0; iu < PTRS_PER_PUD; ++iu, ++p) {
526 if (!pud_present(*p))
528 if (pud_is_leaf(*p)) {
533 pmd = pmd_offset(p, 0);
534 kvmppc_unmap_free_pmd(kvm, pmd, true, lpid);
538 pud_free(kvm->mm, pud);
541 void kvmppc_free_pgtable_radix(struct kvm *kvm, pgd_t *pgd, unsigned int lpid)
545 for (ig = 0; ig < PTRS_PER_PGD; ++ig, ++pgd) {
546 p4d_t *p4d = p4d_offset(pgd, 0);
549 if (!p4d_present(*p4d))
551 pud = pud_offset(p4d, 0);
552 kvmppc_unmap_free_pud(kvm, pud, lpid);
557 void kvmppc_free_radix(struct kvm *kvm)
559 if (kvm->arch.pgtable) {
560 kvmppc_free_pgtable_radix(kvm, kvm->arch.pgtable,
562 pgd_free(kvm->mm, kvm->arch.pgtable);
563 kvm->arch.pgtable = NULL;
567 static void kvmppc_unmap_free_pmd_entry_table(struct kvm *kvm, pmd_t *pmd,
568 unsigned long gpa, unsigned int lpid)
570 pte_t *pte = pte_offset_kernel(pmd, 0);
573 * Clearing the pmd entry then flushing the PWC ensures that the pte
574 * page no longer be cached by the MMU, so can be freed without
575 * flushing the PWC again.
578 kvmppc_radix_flush_pwc(kvm, lpid);
580 kvmppc_unmap_free_pte(kvm, pte, false, lpid);
583 static void kvmppc_unmap_free_pud_entry_table(struct kvm *kvm, pud_t *pud,
584 unsigned long gpa, unsigned int lpid)
586 pmd_t *pmd = pmd_offset(pud, 0);
589 * Clearing the pud entry then flushing the PWC ensures that the pmd
590 * page and any children pte pages will no longer be cached by the MMU,
591 * so can be freed without flushing the PWC again.
594 kvmppc_radix_flush_pwc(kvm, lpid);
596 kvmppc_unmap_free_pmd(kvm, pmd, false, lpid);
600 * There are a number of bits which may differ between different faults to
601 * the same partition scope entry. RC bits, in the course of cleaning and
602 * aging. And the write bit can change, either the access could have been
603 * upgraded, or a read fault could happen concurrently with a write fault
604 * that sets those bits first.
606 #define PTE_BITS_MUST_MATCH (~(_PAGE_WRITE | _PAGE_DIRTY | _PAGE_ACCESSED))
608 int kvmppc_create_pte(struct kvm *kvm, pgd_t *pgtable, pte_t pte,
609 unsigned long gpa, unsigned int level,
610 unsigned long mmu_seq, unsigned int lpid,
611 unsigned long *rmapp, struct rmap_nested **n_rmap)
615 pud_t *pud, *new_pud = NULL;
616 pmd_t *pmd, *new_pmd = NULL;
617 pte_t *ptep, *new_ptep = NULL;
620 /* Traverse the guest's 2nd-level tree, allocate new levels needed */
621 pgd = pgtable + pgd_index(gpa);
622 p4d = p4d_offset(pgd, gpa);
625 if (p4d_present(*p4d))
626 pud = pud_offset(p4d, gpa);
628 new_pud = pud_alloc_one(kvm->mm, gpa);
631 if (pud && pud_present(*pud) && !pud_is_leaf(*pud))
632 pmd = pmd_offset(pud, gpa);
634 new_pmd = kvmppc_pmd_alloc();
636 if (level == 0 && !(pmd && pmd_present(*pmd) && !pmd_is_leaf(*pmd)))
637 new_ptep = kvmppc_pte_alloc();
639 /* Check if we might have been invalidated; let the guest retry if so */
640 spin_lock(&kvm->mmu_lock);
642 if (mmu_notifier_retry(kvm, mmu_seq))
645 /* Now traverse again under the lock and change the tree */
647 if (p4d_none(*p4d)) {
650 p4d_populate(kvm->mm, p4d, new_pud);
653 pud = pud_offset(p4d, gpa);
654 if (pud_is_leaf(*pud)) {
655 unsigned long hgpa = gpa & PUD_MASK;
657 /* Check if we raced and someone else has set the same thing */
659 if (pud_raw(*pud) == pte_raw(pte)) {
663 /* Valid 1GB page here already, add our extra bits */
664 WARN_ON_ONCE((pud_val(*pud) ^ pte_val(pte)) &
665 PTE_BITS_MUST_MATCH);
666 kvmppc_radix_update_pte(kvm, (pte_t *)pud,
667 0, pte_val(pte), hgpa, PUD_SHIFT);
672 * If we raced with another CPU which has just put
673 * a 1GB pte in after we saw a pmd page, try again.
679 /* Valid 1GB page here already, remove it */
680 kvmppc_unmap_pte(kvm, (pte_t *)pud, hgpa, PUD_SHIFT, NULL,
684 if (!pud_none(*pud)) {
686 * There's a page table page here, but we wanted to
687 * install a large page, so remove and free the page
690 kvmppc_unmap_free_pud_entry_table(kvm, pud, gpa, lpid);
692 kvmppc_radix_set_pte_at(kvm, gpa, (pte_t *)pud, pte);
694 kvmhv_insert_nest_rmap(kvm, rmapp, n_rmap);
698 if (pud_none(*pud)) {
701 pud_populate(kvm->mm, pud, new_pmd);
704 pmd = pmd_offset(pud, gpa);
705 if (pmd_is_leaf(*pmd)) {
706 unsigned long lgpa = gpa & PMD_MASK;
708 /* Check if we raced and someone else has set the same thing */
710 if (pmd_raw(*pmd) == pte_raw(pte)) {
714 /* Valid 2MB page here already, add our extra bits */
715 WARN_ON_ONCE((pmd_val(*pmd) ^ pte_val(pte)) &
716 PTE_BITS_MUST_MATCH);
717 kvmppc_radix_update_pte(kvm, pmdp_ptep(pmd),
718 0, pte_val(pte), lgpa, PMD_SHIFT);
724 * If we raced with another CPU which has just put
725 * a 2MB pte in after we saw a pte page, try again.
731 /* Valid 2MB page here already, remove it */
732 kvmppc_unmap_pte(kvm, pmdp_ptep(pmd), lgpa, PMD_SHIFT, NULL,
736 if (!pmd_none(*pmd)) {
738 * There's a page table page here, but we wanted to
739 * install a large page, so remove and free the page
742 kvmppc_unmap_free_pmd_entry_table(kvm, pmd, gpa, lpid);
744 kvmppc_radix_set_pte_at(kvm, gpa, pmdp_ptep(pmd), pte);
746 kvmhv_insert_nest_rmap(kvm, rmapp, n_rmap);
750 if (pmd_none(*pmd)) {
753 pmd_populate(kvm->mm, pmd, new_ptep);
756 ptep = pte_offset_kernel(pmd, gpa);
757 if (pte_present(*ptep)) {
758 /* Check if someone else set the same thing */
759 if (pte_raw(*ptep) == pte_raw(pte)) {
763 /* Valid page here already, add our extra bits */
764 WARN_ON_ONCE((pte_val(*ptep) ^ pte_val(pte)) &
765 PTE_BITS_MUST_MATCH);
766 kvmppc_radix_update_pte(kvm, ptep, 0, pte_val(pte), gpa, 0);
770 kvmppc_radix_set_pte_at(kvm, gpa, ptep, pte);
772 kvmhv_insert_nest_rmap(kvm, rmapp, n_rmap);
776 spin_unlock(&kvm->mmu_lock);
778 pud_free(kvm->mm, new_pud);
780 kvmppc_pmd_free(new_pmd);
782 kvmppc_pte_free(new_ptep);
786 bool kvmppc_hv_handle_set_rc(struct kvm *kvm, bool nested, bool writing,
787 unsigned long gpa, unsigned int lpid)
789 unsigned long pgflags;
794 * Need to set an R or C bit in the 2nd-level tables;
795 * since we are just helping out the hardware here,
796 * it is sufficient to do what the hardware does.
798 pgflags = _PAGE_ACCESSED;
800 pgflags |= _PAGE_DIRTY;
803 ptep = find_kvm_nested_guest_pte(kvm, lpid, gpa, &shift);
805 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
807 if (ptep && pte_present(*ptep) && (!writing || pte_write(*ptep))) {
808 kvmppc_radix_update_pte(kvm, ptep, 0, pgflags, gpa, shift);
814 int kvmppc_book3s_instantiate_page(struct kvm_vcpu *vcpu,
816 struct kvm_memory_slot *memslot,
817 bool writing, bool kvm_ro,
818 pte_t *inserted_pte, unsigned int *levelp)
820 struct kvm *kvm = vcpu->kvm;
821 struct page *page = NULL;
822 unsigned long mmu_seq;
823 unsigned long hva, gfn = gpa >> PAGE_SHIFT;
824 bool upgrade_write = false;
825 bool *upgrade_p = &upgrade_write;
827 unsigned int shift, level;
831 /* used to check for invalidations in progress */
832 mmu_seq = kvm->mmu_notifier_seq;
836 * Do a fast check first, since __gfn_to_pfn_memslot doesn't
837 * do it with !atomic && !async, which is how we call it.
838 * We always ask for write permission since the common case
839 * is that the page is writable.
841 hva = gfn_to_hva_memslot(memslot, gfn);
842 if (!kvm_ro && get_user_page_fast_only(hva, FOLL_WRITE, &page)) {
843 upgrade_write = true;
847 /* Call KVM generic code to do the slow-path check */
848 pfn = __gfn_to_pfn_memslot(memslot, gfn, false, NULL,
849 writing, upgrade_p, NULL);
850 if (is_error_noslot_pfn(pfn))
853 if (pfn_valid(pfn)) {
854 page = pfn_to_page(pfn);
855 if (PageReserved(page))
861 * Read the PTE from the process' radix tree and use that
862 * so we get the shift and attribute bits.
864 spin_lock(&kvm->mmu_lock);
865 ptep = find_kvm_host_pte(kvm, mmu_seq, hva, &shift);
868 pte = READ_ONCE(*ptep);
869 spin_unlock(&kvm->mmu_lock);
871 * If the PTE disappeared temporarily due to a THP
872 * collapse, just return and let the guest try again.
874 if (!pte_present(pte)) {
880 /* If we're logging dirty pages, always map single pages */
881 large_enable = !(memslot->flags & KVM_MEM_LOG_DIRTY_PAGES);
883 /* Get pte level from shift/size */
884 if (large_enable && shift == PUD_SHIFT &&
885 (gpa & (PUD_SIZE - PAGE_SIZE)) ==
886 (hva & (PUD_SIZE - PAGE_SIZE))) {
888 } else if (large_enable && shift == PMD_SHIFT &&
889 (gpa & (PMD_SIZE - PAGE_SIZE)) ==
890 (hva & (PMD_SIZE - PAGE_SIZE))) {
894 if (shift > PAGE_SHIFT) {
896 * If the pte maps more than one page, bring over
897 * bits from the virtual address to get the real
898 * address of the specific single page we want.
900 unsigned long rpnmask = (1ul << shift) - PAGE_SIZE;
901 pte = __pte(pte_val(pte) | (hva & rpnmask));
905 pte = __pte(pte_val(pte) | _PAGE_EXEC | _PAGE_ACCESSED);
906 if (writing || upgrade_write) {
907 if (pte_val(pte) & _PAGE_WRITE)
908 pte = __pte(pte_val(pte) | _PAGE_DIRTY);
910 pte = __pte(pte_val(pte) & ~(_PAGE_WRITE | _PAGE_DIRTY));
913 /* Allocate space in the tree and write the PTE */
914 ret = kvmppc_create_pte(kvm, kvm->arch.pgtable, pte, gpa, level,
915 mmu_seq, kvm->arch.lpid, NULL, NULL);
922 if (!ret && (pte_val(pte) & _PAGE_WRITE))
923 set_page_dirty_lock(page);
927 /* Increment number of large pages if we (successfully) inserted one */
930 kvm->stat.num_2M_pages++;
932 kvm->stat.num_1G_pages++;
938 int kvmppc_book3s_radix_page_fault(struct kvm_vcpu *vcpu,
939 unsigned long ea, unsigned long dsisr)
941 struct kvm *kvm = vcpu->kvm;
942 unsigned long gpa, gfn;
943 struct kvm_memory_slot *memslot;
945 bool writing = !!(dsisr & DSISR_ISSTORE);
948 /* Check for unusual errors */
949 if (dsisr & DSISR_UNSUPP_MMU) {
950 pr_err("KVM: Got unsupported MMU fault\n");
953 if (dsisr & DSISR_BADACCESS) {
954 /* Reflect to the guest as DSI */
955 pr_err("KVM: Got radix HV page fault with DSISR=%lx\n", dsisr);
956 kvmppc_core_queue_data_storage(vcpu, ea, dsisr);
960 /* Translate the logical address */
961 gpa = vcpu->arch.fault_gpa & ~0xfffUL;
962 gpa &= ~0xF000000000000000ul;
963 gfn = gpa >> PAGE_SHIFT;
964 if (!(dsisr & DSISR_PRTABLE_FAULT))
967 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
968 return kvmppc_send_page_to_uv(kvm, gfn);
970 /* Get the corresponding memslot */
971 memslot = gfn_to_memslot(kvm, gfn);
973 /* No memslot means it's an emulated MMIO region */
974 if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID)) {
975 if (dsisr & (DSISR_PRTABLE_FAULT | DSISR_BADACCESS |
978 * Bad address in guest page table tree, or other
979 * unusual error - reflect it to the guest as DSI.
981 kvmppc_core_queue_data_storage(vcpu, ea, dsisr);
984 return kvmppc_hv_emulate_mmio(vcpu, gpa, ea, writing);
987 if (memslot->flags & KVM_MEM_READONLY) {
989 /* give the guest a DSI */
990 kvmppc_core_queue_data_storage(vcpu, ea, DSISR_ISSTORE |
997 /* Failed to set the reference/change bits */
998 if (dsisr & DSISR_SET_RC) {
999 spin_lock(&kvm->mmu_lock);
1000 if (kvmppc_hv_handle_set_rc(kvm, false, writing,
1001 gpa, kvm->arch.lpid))
1002 dsisr &= ~DSISR_SET_RC;
1003 spin_unlock(&kvm->mmu_lock);
1005 if (!(dsisr & (DSISR_BAD_FAULT_64S | DSISR_NOHPTE |
1006 DSISR_PROTFAULT | DSISR_SET_RC)))
1007 return RESUME_GUEST;
1010 /* Try to insert a pte */
1011 ret = kvmppc_book3s_instantiate_page(vcpu, gpa, memslot, writing,
1012 kvm_ro, NULL, NULL);
1014 if (ret == 0 || ret == -EAGAIN)
1019 /* Called with kvm->mmu_lock held */
1020 void kvm_unmap_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
1024 unsigned long gpa = gfn << PAGE_SHIFT;
1027 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE) {
1028 uv_page_inval(kvm->arch.lpid, gpa, PAGE_SHIFT);
1032 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1033 if (ptep && pte_present(*ptep))
1034 kvmppc_unmap_pte(kvm, ptep, gpa, shift, memslot,
1038 /* Called with kvm->mmu_lock held */
1039 bool kvm_age_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
1043 unsigned long gpa = gfn << PAGE_SHIFT;
1046 unsigned long old, *rmapp;
1048 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1051 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1052 if (ptep && pte_present(*ptep) && pte_young(*ptep)) {
1053 old = kvmppc_radix_update_pte(kvm, ptep, _PAGE_ACCESSED, 0,
1055 /* XXX need to flush tlb here? */
1056 /* Also clear bit in ptes in shadow pgtable for nested guests */
1057 rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1058 kvmhv_update_nest_rmap_rc_list(kvm, rmapp, _PAGE_ACCESSED, 0,
1066 /* Called with kvm->mmu_lock held */
1067 bool kvm_test_age_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
1072 unsigned long gpa = gfn << PAGE_SHIFT;
1076 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1079 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1080 if (ptep && pte_present(*ptep) && pte_young(*ptep))
1085 /* Returns the number of PAGE_SIZE pages that are dirty */
1086 static int kvm_radix_test_clear_dirty(struct kvm *kvm,
1087 struct kvm_memory_slot *memslot, int pagenum)
1089 unsigned long gfn = memslot->base_gfn + pagenum;
1090 unsigned long gpa = gfn << PAGE_SHIFT;
1094 unsigned long old, *rmapp;
1096 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1100 * For performance reasons we don't hold kvm->mmu_lock while walking the
1101 * partition scoped table.
1103 ptep = find_kvm_secondary_pte_unlocked(kvm, gpa, &shift);
1107 pte = READ_ONCE(*ptep);
1108 if (pte_present(pte) && pte_dirty(pte)) {
1109 spin_lock(&kvm->mmu_lock);
1111 * Recheck the pte again
1113 if (pte_val(pte) != pte_val(*ptep)) {
1115 * We have KVM_MEM_LOG_DIRTY_PAGES enabled. Hence we can
1116 * only find PAGE_SIZE pte entries here. We can continue
1117 * to use the pte addr returned by above page table
1120 if (!pte_present(*ptep) || !pte_dirty(*ptep)) {
1121 spin_unlock(&kvm->mmu_lock);
1128 old = kvmppc_radix_update_pte(kvm, ptep, _PAGE_DIRTY, 0,
1130 kvmppc_radix_tlbie_page(kvm, gpa, shift, kvm->arch.lpid);
1131 /* Also clear bit in ptes in shadow pgtable for nested guests */
1132 rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1133 kvmhv_update_nest_rmap_rc_list(kvm, rmapp, _PAGE_DIRTY, 0,
1136 spin_unlock(&kvm->mmu_lock);
1141 long kvmppc_hv_get_dirty_log_radix(struct kvm *kvm,
1142 struct kvm_memory_slot *memslot, unsigned long *map)
1147 for (i = 0; i < memslot->npages; i = j) {
1148 npages = kvm_radix_test_clear_dirty(kvm, memslot, i);
1151 * Note that if npages > 0 then i must be a multiple of npages,
1152 * since huge pages are only used to back the guest at guest
1153 * real addresses that are a multiple of their size.
1154 * Since we have at most one PTE covering any given guest
1155 * real address, if npages > 1 we can skip to i + npages.
1159 set_dirty_bits(map, i, npages);
1166 void kvmppc_radix_flush_memslot(struct kvm *kvm,
1167 const struct kvm_memory_slot *memslot)
1174 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_START)
1175 kvmppc_uvmem_drop_pages(memslot, kvm, true);
1177 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1180 gpa = memslot->base_gfn << PAGE_SHIFT;
1181 spin_lock(&kvm->mmu_lock);
1182 for (n = memslot->npages; n; --n) {
1183 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1184 if (ptep && pte_present(*ptep))
1185 kvmppc_unmap_pte(kvm, ptep, gpa, shift, memslot,
1190 * Increase the mmu notifier sequence number to prevent any page
1191 * fault that read the memslot earlier from writing a PTE.
1193 kvm->mmu_notifier_seq++;
1194 spin_unlock(&kvm->mmu_lock);
1197 static void add_rmmu_ap_encoding(struct kvm_ppc_rmmu_info *info,
1198 int psize, int *indexp)
1200 if (!mmu_psize_defs[psize].shift)
1202 info->ap_encodings[*indexp] = mmu_psize_defs[psize].shift |
1203 (mmu_psize_defs[psize].ap << 29);
1207 int kvmhv_get_rmmu_info(struct kvm *kvm, struct kvm_ppc_rmmu_info *info)
1211 if (!radix_enabled())
1213 memset(info, 0, sizeof(*info));
1216 info->geometries[0].page_shift = 12;
1217 info->geometries[0].level_bits[0] = 9;
1218 for (i = 1; i < 4; ++i)
1219 info->geometries[0].level_bits[i] = p9_supported_radix_bits[i];
1221 info->geometries[1].page_shift = 16;
1222 for (i = 0; i < 4; ++i)
1223 info->geometries[1].level_bits[i] = p9_supported_radix_bits[i];
1226 add_rmmu_ap_encoding(info, MMU_PAGE_4K, &i);
1227 add_rmmu_ap_encoding(info, MMU_PAGE_64K, &i);
1228 add_rmmu_ap_encoding(info, MMU_PAGE_2M, &i);
1229 add_rmmu_ap_encoding(info, MMU_PAGE_1G, &i);
1234 int kvmppc_init_vm_radix(struct kvm *kvm)
1236 kvm->arch.pgtable = pgd_alloc(kvm->mm);
1237 if (!kvm->arch.pgtable)
1242 static void pte_ctor(void *addr)
1244 memset(addr, 0, RADIX_PTE_TABLE_SIZE);
1247 static void pmd_ctor(void *addr)
1249 memset(addr, 0, RADIX_PMD_TABLE_SIZE);
1252 struct debugfs_radix_state {
1263 static int debugfs_radix_open(struct inode *inode, struct file *file)
1265 struct kvm *kvm = inode->i_private;
1266 struct debugfs_radix_state *p;
1268 p = kzalloc(sizeof(*p), GFP_KERNEL);
1274 mutex_init(&p->mutex);
1275 file->private_data = p;
1277 return nonseekable_open(inode, file);
1280 static int debugfs_radix_release(struct inode *inode, struct file *file)
1282 struct debugfs_radix_state *p = file->private_data;
1284 kvm_put_kvm(p->kvm);
1289 static ssize_t debugfs_radix_read(struct file *file, char __user *buf,
1290 size_t len, loff_t *ppos)
1292 struct debugfs_radix_state *p = file->private_data;
1298 struct kvm_nested_guest *nested;
1308 if (!kvm_is_radix(kvm))
1311 ret = mutex_lock_interruptible(&p->mutex);
1315 if (p->chars_left) {
1319 r = copy_to_user(buf, p->buf + p->buf_index, n);
1336 while (len != 0 && p->lpid >= 0) {
1337 if (gpa >= RADIX_PGTABLE_RANGE) {
1341 kvmhv_put_nested(nested);
1344 p->lpid = kvmhv_nested_next_lpid(kvm, p->lpid);
1351 pgt = kvm->arch.pgtable;
1353 nested = kvmhv_get_nested(kvm, p->lpid, false);
1355 gpa = RADIX_PGTABLE_RANGE;
1358 pgt = nested->shadow_pgtable;
1364 n = scnprintf(p->buf, sizeof(p->buf),
1365 "\nNested LPID %d: ", p->lpid);
1366 n += scnprintf(p->buf + n, sizeof(p->buf) - n,
1367 "pgdir: %lx\n", (unsigned long)pgt);
1372 pgdp = pgt + pgd_index(gpa);
1373 p4dp = p4d_offset(pgdp, gpa);
1374 p4d = READ_ONCE(*p4dp);
1375 if (!(p4d_val(p4d) & _PAGE_PRESENT)) {
1376 gpa = (gpa & P4D_MASK) + P4D_SIZE;
1380 pudp = pud_offset(&p4d, gpa);
1381 pud = READ_ONCE(*pudp);
1382 if (!(pud_val(pud) & _PAGE_PRESENT)) {
1383 gpa = (gpa & PUD_MASK) + PUD_SIZE;
1386 if (pud_val(pud) & _PAGE_PTE) {
1392 pmdp = pmd_offset(&pud, gpa);
1393 pmd = READ_ONCE(*pmdp);
1394 if (!(pmd_val(pmd) & _PAGE_PRESENT)) {
1395 gpa = (gpa & PMD_MASK) + PMD_SIZE;
1398 if (pmd_val(pmd) & _PAGE_PTE) {
1404 ptep = pte_offset_kernel(&pmd, gpa);
1405 pte = pte_val(READ_ONCE(*ptep));
1406 if (!(pte & _PAGE_PRESENT)) {
1412 n = scnprintf(p->buf, sizeof(p->buf),
1413 " %lx: %lx %d\n", gpa, pte, shift);
1414 gpa += 1ul << shift;
1419 r = copy_to_user(buf, p->buf, n);
1434 kvmhv_put_nested(nested);
1437 mutex_unlock(&p->mutex);
1441 static ssize_t debugfs_radix_write(struct file *file, const char __user *buf,
1442 size_t len, loff_t *ppos)
1447 static const struct file_operations debugfs_radix_fops = {
1448 .owner = THIS_MODULE,
1449 .open = debugfs_radix_open,
1450 .release = debugfs_radix_release,
1451 .read = debugfs_radix_read,
1452 .write = debugfs_radix_write,
1453 .llseek = generic_file_llseek,
1456 void kvmhv_radix_debugfs_init(struct kvm *kvm)
1458 debugfs_create_file("radix", 0400, kvm->debugfs_dentry, kvm,
1459 &debugfs_radix_fops);
1462 int kvmppc_radix_init(void)
1464 unsigned long size = sizeof(void *) << RADIX_PTE_INDEX_SIZE;
1466 kvm_pte_cache = kmem_cache_create("kvm-pte", size, size, 0, pte_ctor);
1470 size = sizeof(void *) << RADIX_PMD_INDEX_SIZE;
1472 kvm_pmd_cache = kmem_cache_create("kvm-pmd", size, size, 0, pmd_ctor);
1473 if (!kvm_pmd_cache) {
1474 kmem_cache_destroy(kvm_pte_cache);
1481 void kvmppc_radix_exit(void)
1483 kmem_cache_destroy(kvm_pte_cache);
1484 kmem_cache_destroy(kvm_pmd_cache);