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));
171 vcpu->srcu_idx = srcu_read_lock(&kvm->srcu);
172 ret = kvm_read_guest(kvm, addr, &rpte, sizeof(rpte));
173 srcu_read_unlock(&kvm->srcu, vcpu->srcu_idx);
179 pte = __be64_to_cpu(rpte);
180 if (!(pte & _PAGE_PRESENT))
182 /* Check if a leaf entry */
185 /* Get ready to walk the next level */
186 base = pte & RPDB_MASK;
187 bits = pte & RPDS_MASK;
190 /* Need a leaf at lowest level; 512GB pages not supported */
191 if (level < 0 || level == 3)
194 /* We found a valid leaf PTE */
195 /* Offset is now log base 2 of the page size */
196 gpa = pte & 0x01fffffffffff000ul;
197 if (gpa & ((1ul << offset) - 1))
199 gpa |= eaddr & ((1ul << offset) - 1);
200 for (ps = MMU_PAGE_4K; ps < MMU_PAGE_COUNT; ++ps)
201 if (offset == mmu_psize_defs[ps].shift)
203 gpte->page_size = ps;
204 gpte->page_shift = offset;
209 /* Work out permissions */
210 gpte->may_read = !!(pte & _PAGE_READ);
211 gpte->may_write = !!(pte & _PAGE_WRITE);
212 gpte->may_execute = !!(pte & _PAGE_EXEC);
214 gpte->rc = pte & (_PAGE_ACCESSED | _PAGE_DIRTY);
223 * Used to walk a partition or process table radix tree in guest memory
224 * Note: We exploit the fact that a partition table and a process
225 * table have the same layout, a partition-scoped page table and a
226 * process-scoped page table have the same layout, and the 2nd
227 * doubleword of a partition table entry has the same layout as
230 int kvmppc_mmu_radix_translate_table(struct kvm_vcpu *vcpu, gva_t eaddr,
231 struct kvmppc_pte *gpte, u64 table,
232 int table_index, u64 *pte_ret_p)
234 struct kvm *kvm = vcpu->kvm;
236 unsigned long size, ptbl, root;
237 struct prtb_entry entry;
239 if ((table & PRTS_MASK) > 24)
241 size = 1ul << ((table & PRTS_MASK) + 12);
243 /* Is the table big enough to contain this entry? */
244 if ((table_index * sizeof(entry)) >= size)
247 /* Read the table to find the root of the radix tree */
248 ptbl = (table & PRTB_MASK) + (table_index * sizeof(entry));
249 vcpu->srcu_idx = srcu_read_lock(&kvm->srcu);
250 ret = kvm_read_guest(kvm, ptbl, &entry, sizeof(entry));
251 srcu_read_unlock(&kvm->srcu, vcpu->srcu_idx);
255 /* Root is stored in the first double word */
256 root = be64_to_cpu(entry.prtb0);
258 return kvmppc_mmu_walk_radix_tree(vcpu, eaddr, gpte, root, pte_ret_p);
261 int kvmppc_mmu_radix_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
262 struct kvmppc_pte *gpte, bool data, bool iswrite)
268 /* Work out effective PID */
269 switch (eaddr >> 62) {
271 pid = vcpu->arch.pid;
280 ret = kvmppc_mmu_radix_translate_table(vcpu, eaddr, gpte,
281 vcpu->kvm->arch.process_table, pid, &pte);
285 /* Check privilege (applies only to process scoped translations) */
286 if (kvmppc_get_msr(vcpu) & MSR_PR) {
287 if (pte & _PAGE_PRIVILEGED) {
290 gpte->may_execute = 0;
293 if (!(pte & _PAGE_PRIVILEGED)) {
294 /* Check AMR/IAMR to see if strict mode is in force */
295 if (vcpu->arch.amr & (1ul << 62))
297 if (vcpu->arch.amr & (1ul << 63))
299 if (vcpu->arch.iamr & (1ul << 62))
300 gpte->may_execute = 0;
307 void kvmppc_radix_tlbie_page(struct kvm *kvm, unsigned long addr,
308 unsigned int pshift, unsigned int lpid)
310 unsigned long psize = PAGE_SIZE;
316 psize = 1UL << pshift;
320 addr &= ~(psize - 1);
322 if (!kvmhv_on_pseries()) {
323 radix__flush_tlb_lpid_page(lpid, addr, psize);
327 psi = shift_to_mmu_psize(pshift);
329 if (!firmware_has_feature(FW_FEATURE_RPT_INVALIDATE)) {
330 rb = addr | (mmu_get_ap(psi) << PPC_BITLSHIFT(58));
331 rc = plpar_hcall_norets(H_TLB_INVALIDATE, H_TLBIE_P1_ENC(0, 0, 1),
334 rc = pseries_rpt_invalidate(lpid, H_RPTI_TARGET_CMMU,
337 psize_to_rpti_pgsize(psi),
342 pr_err("KVM: TLB page invalidation hcall failed, rc=%ld\n", rc);
345 static void kvmppc_radix_flush_pwc(struct kvm *kvm, unsigned int lpid)
349 if (!kvmhv_on_pseries()) {
350 radix__flush_pwc_lpid(lpid);
354 if (!firmware_has_feature(FW_FEATURE_RPT_INVALIDATE))
355 rc = plpar_hcall_norets(H_TLB_INVALIDATE, H_TLBIE_P1_ENC(1, 0, 1),
356 lpid, TLBIEL_INVAL_SET_LPID);
358 rc = pseries_rpt_invalidate(lpid, H_RPTI_TARGET_CMMU,
360 H_RPTI_TYPE_PWC, H_RPTI_PAGE_ALL,
363 pr_err("KVM: TLB PWC invalidation hcall failed, rc=%ld\n", rc);
366 static unsigned long kvmppc_radix_update_pte(struct kvm *kvm, pte_t *ptep,
367 unsigned long clr, unsigned long set,
368 unsigned long addr, unsigned int shift)
370 return __radix_pte_update(ptep, clr, set);
373 static void kvmppc_radix_set_pte_at(struct kvm *kvm, unsigned long addr,
374 pte_t *ptep, pte_t pte)
376 radix__set_pte_at(kvm->mm, addr, ptep, pte, 0);
379 static struct kmem_cache *kvm_pte_cache;
380 static struct kmem_cache *kvm_pmd_cache;
382 static pte_t *kvmppc_pte_alloc(void)
386 pte = kmem_cache_alloc(kvm_pte_cache, GFP_KERNEL);
387 /* pmd_populate() will only reference _pa(pte). */
388 kmemleak_ignore(pte);
393 static void kvmppc_pte_free(pte_t *ptep)
395 kmem_cache_free(kvm_pte_cache, ptep);
398 static pmd_t *kvmppc_pmd_alloc(void)
402 pmd = kmem_cache_alloc(kvm_pmd_cache, GFP_KERNEL);
403 /* pud_populate() will only reference _pa(pmd). */
404 kmemleak_ignore(pmd);
409 static void kvmppc_pmd_free(pmd_t *pmdp)
411 kmem_cache_free(kvm_pmd_cache, pmdp);
414 /* Called with kvm->mmu_lock held */
415 void kvmppc_unmap_pte(struct kvm *kvm, pte_t *pte, unsigned long gpa,
417 const struct kvm_memory_slot *memslot,
422 unsigned long gfn = gpa >> PAGE_SHIFT;
423 unsigned long page_size = PAGE_SIZE;
426 old = kvmppc_radix_update_pte(kvm, pte, ~0UL, 0, gpa, shift);
427 kvmppc_radix_tlbie_page(kvm, gpa, shift, lpid);
429 /* The following only applies to L1 entries */
430 if (lpid != kvm->arch.lpid)
434 memslot = gfn_to_memslot(kvm, gfn);
438 if (shift) { /* 1GB or 2MB page */
439 page_size = 1ul << shift;
440 if (shift == PMD_SHIFT)
441 kvm->stat.num_2M_pages--;
442 else if (shift == PUD_SHIFT)
443 kvm->stat.num_1G_pages--;
446 gpa &= ~(page_size - 1);
447 hpa = old & PTE_RPN_MASK;
448 kvmhv_remove_nest_rmap_range(kvm, memslot, gpa, hpa, page_size);
450 if ((old & _PAGE_DIRTY) && memslot->dirty_bitmap)
451 kvmppc_update_dirty_map(memslot, gfn, page_size);
455 * kvmppc_free_p?d are used to free existing page tables, and recursively
456 * descend and clear and free children.
457 * Callers are responsible for flushing the PWC.
459 * When page tables are being unmapped/freed as part of page fault path
460 * (full == false), valid ptes are generally not expected; however, there
461 * is one situation where they arise, which is when dirty page logging is
462 * turned off for a memslot while the VM is running. The new memslot
463 * becomes visible to page faults before the memslot commit function
464 * gets to flush the memslot, which can lead to a 2MB page mapping being
465 * installed for a guest physical address where there are already 64kB
466 * (or 4kB) mappings (of sub-pages of the same 2MB page).
468 static void kvmppc_unmap_free_pte(struct kvm *kvm, pte_t *pte, bool full,
472 memset(pte, 0, sizeof(long) << RADIX_PTE_INDEX_SIZE);
477 for (it = 0; it < PTRS_PER_PTE; ++it, ++p) {
478 if (pte_val(*p) == 0)
480 kvmppc_unmap_pte(kvm, p,
481 pte_pfn(*p) << PAGE_SHIFT,
482 PAGE_SHIFT, NULL, lpid);
486 kvmppc_pte_free(pte);
489 static void kvmppc_unmap_free_pmd(struct kvm *kvm, pmd_t *pmd, bool full,
495 for (im = 0; im < PTRS_PER_PMD; ++im, ++p) {
496 if (!pmd_present(*p))
498 if (pmd_is_leaf(*p)) {
503 kvmppc_unmap_pte(kvm, (pte_t *)p,
504 pte_pfn(*(pte_t *)p) << PAGE_SHIFT,
505 PMD_SHIFT, NULL, lpid);
510 pte = pte_offset_map(p, 0);
511 kvmppc_unmap_free_pte(kvm, pte, full, lpid);
515 kvmppc_pmd_free(pmd);
518 static void kvmppc_unmap_free_pud(struct kvm *kvm, pud_t *pud,
524 for (iu = 0; iu < PTRS_PER_PUD; ++iu, ++p) {
525 if (!pud_present(*p))
527 if (pud_is_leaf(*p)) {
532 pmd = pmd_offset(p, 0);
533 kvmppc_unmap_free_pmd(kvm, pmd, true, lpid);
537 pud_free(kvm->mm, pud);
540 void kvmppc_free_pgtable_radix(struct kvm *kvm, pgd_t *pgd, unsigned int lpid)
544 for (ig = 0; ig < PTRS_PER_PGD; ++ig, ++pgd) {
545 p4d_t *p4d = p4d_offset(pgd, 0);
548 if (!p4d_present(*p4d))
550 pud = pud_offset(p4d, 0);
551 kvmppc_unmap_free_pud(kvm, pud, lpid);
556 void kvmppc_free_radix(struct kvm *kvm)
558 if (kvm->arch.pgtable) {
559 kvmppc_free_pgtable_radix(kvm, kvm->arch.pgtable,
561 pgd_free(kvm->mm, kvm->arch.pgtable);
562 kvm->arch.pgtable = NULL;
566 static void kvmppc_unmap_free_pmd_entry_table(struct kvm *kvm, pmd_t *pmd,
567 unsigned long gpa, unsigned int lpid)
569 pte_t *pte = pte_offset_kernel(pmd, 0);
572 * Clearing the pmd entry then flushing the PWC ensures that the pte
573 * page no longer be cached by the MMU, so can be freed without
574 * flushing the PWC again.
577 kvmppc_radix_flush_pwc(kvm, lpid);
579 kvmppc_unmap_free_pte(kvm, pte, false, lpid);
582 static void kvmppc_unmap_free_pud_entry_table(struct kvm *kvm, pud_t *pud,
583 unsigned long gpa, unsigned int lpid)
585 pmd_t *pmd = pmd_offset(pud, 0);
588 * Clearing the pud entry then flushing the PWC ensures that the pmd
589 * page and any children pte pages will no longer be cached by the MMU,
590 * so can be freed without flushing the PWC again.
593 kvmppc_radix_flush_pwc(kvm, lpid);
595 kvmppc_unmap_free_pmd(kvm, pmd, false, lpid);
599 * There are a number of bits which may differ between different faults to
600 * the same partition scope entry. RC bits, in the course of cleaning and
601 * aging. And the write bit can change, either the access could have been
602 * upgraded, or a read fault could happen concurrently with a write fault
603 * that sets those bits first.
605 #define PTE_BITS_MUST_MATCH (~(_PAGE_WRITE | _PAGE_DIRTY | _PAGE_ACCESSED))
607 int kvmppc_create_pte(struct kvm *kvm, pgd_t *pgtable, pte_t pte,
608 unsigned long gpa, unsigned int level,
609 unsigned long mmu_seq, unsigned int lpid,
610 unsigned long *rmapp, struct rmap_nested **n_rmap)
614 pud_t *pud, *new_pud = NULL;
615 pmd_t *pmd, *new_pmd = NULL;
616 pte_t *ptep, *new_ptep = NULL;
619 /* Traverse the guest's 2nd-level tree, allocate new levels needed */
620 pgd = pgtable + pgd_index(gpa);
621 p4d = p4d_offset(pgd, gpa);
624 if (p4d_present(*p4d))
625 pud = pud_offset(p4d, gpa);
627 new_pud = pud_alloc_one(kvm->mm, gpa);
630 if (pud && pud_present(*pud) && !pud_is_leaf(*pud))
631 pmd = pmd_offset(pud, gpa);
633 new_pmd = kvmppc_pmd_alloc();
635 if (level == 0 && !(pmd && pmd_present(*pmd) && !pmd_is_leaf(*pmd)))
636 new_ptep = kvmppc_pte_alloc();
638 /* Check if we might have been invalidated; let the guest retry if so */
639 spin_lock(&kvm->mmu_lock);
641 if (mmu_notifier_retry(kvm, mmu_seq))
644 /* Now traverse again under the lock and change the tree */
646 if (p4d_none(*p4d)) {
649 p4d_populate(kvm->mm, p4d, new_pud);
652 pud = pud_offset(p4d, gpa);
653 if (pud_is_leaf(*pud)) {
654 unsigned long hgpa = gpa & PUD_MASK;
656 /* Check if we raced and someone else has set the same thing */
658 if (pud_raw(*pud) == pte_raw(pte)) {
662 /* Valid 1GB page here already, add our extra bits */
663 WARN_ON_ONCE((pud_val(*pud) ^ pte_val(pte)) &
664 PTE_BITS_MUST_MATCH);
665 kvmppc_radix_update_pte(kvm, (pte_t *)pud,
666 0, pte_val(pte), hgpa, PUD_SHIFT);
671 * If we raced with another CPU which has just put
672 * a 1GB pte in after we saw a pmd page, try again.
678 /* Valid 1GB page here already, remove it */
679 kvmppc_unmap_pte(kvm, (pte_t *)pud, hgpa, PUD_SHIFT, NULL,
683 if (!pud_none(*pud)) {
685 * There's a page table page here, but we wanted to
686 * install a large page, so remove and free the page
689 kvmppc_unmap_free_pud_entry_table(kvm, pud, gpa, lpid);
691 kvmppc_radix_set_pte_at(kvm, gpa, (pte_t *)pud, pte);
693 kvmhv_insert_nest_rmap(kvm, rmapp, n_rmap);
697 if (pud_none(*pud)) {
700 pud_populate(kvm->mm, pud, new_pmd);
703 pmd = pmd_offset(pud, gpa);
704 if (pmd_is_leaf(*pmd)) {
705 unsigned long lgpa = gpa & PMD_MASK;
707 /* Check if we raced and someone else has set the same thing */
709 if (pmd_raw(*pmd) == pte_raw(pte)) {
713 /* Valid 2MB page here already, add our extra bits */
714 WARN_ON_ONCE((pmd_val(*pmd) ^ pte_val(pte)) &
715 PTE_BITS_MUST_MATCH);
716 kvmppc_radix_update_pte(kvm, pmdp_ptep(pmd),
717 0, pte_val(pte), lgpa, PMD_SHIFT);
723 * If we raced with another CPU which has just put
724 * a 2MB pte in after we saw a pte page, try again.
730 /* Valid 2MB page here already, remove it */
731 kvmppc_unmap_pte(kvm, pmdp_ptep(pmd), lgpa, PMD_SHIFT, NULL,
735 if (!pmd_none(*pmd)) {
737 * There's a page table page here, but we wanted to
738 * install a large page, so remove and free the page
741 kvmppc_unmap_free_pmd_entry_table(kvm, pmd, gpa, lpid);
743 kvmppc_radix_set_pte_at(kvm, gpa, pmdp_ptep(pmd), pte);
745 kvmhv_insert_nest_rmap(kvm, rmapp, n_rmap);
749 if (pmd_none(*pmd)) {
752 pmd_populate(kvm->mm, pmd, new_ptep);
755 ptep = pte_offset_kernel(pmd, gpa);
756 if (pte_present(*ptep)) {
757 /* Check if someone else set the same thing */
758 if (pte_raw(*ptep) == pte_raw(pte)) {
762 /* Valid page here already, add our extra bits */
763 WARN_ON_ONCE((pte_val(*ptep) ^ pte_val(pte)) &
764 PTE_BITS_MUST_MATCH);
765 kvmppc_radix_update_pte(kvm, ptep, 0, pte_val(pte), gpa, 0);
769 kvmppc_radix_set_pte_at(kvm, gpa, ptep, pte);
771 kvmhv_insert_nest_rmap(kvm, rmapp, n_rmap);
775 spin_unlock(&kvm->mmu_lock);
777 pud_free(kvm->mm, new_pud);
779 kvmppc_pmd_free(new_pmd);
781 kvmppc_pte_free(new_ptep);
785 bool kvmppc_hv_handle_set_rc(struct kvm *kvm, bool nested, bool writing,
786 unsigned long gpa, unsigned int lpid)
788 unsigned long pgflags;
793 * Need to set an R or C bit in the 2nd-level tables;
794 * since we are just helping out the hardware here,
795 * it is sufficient to do what the hardware does.
797 pgflags = _PAGE_ACCESSED;
799 pgflags |= _PAGE_DIRTY;
802 ptep = find_kvm_nested_guest_pte(kvm, lpid, gpa, &shift);
804 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
806 if (ptep && pte_present(*ptep) && (!writing || pte_write(*ptep))) {
807 kvmppc_radix_update_pte(kvm, ptep, 0, pgflags, gpa, shift);
813 int kvmppc_book3s_instantiate_page(struct kvm_vcpu *vcpu,
815 struct kvm_memory_slot *memslot,
816 bool writing, bool kvm_ro,
817 pte_t *inserted_pte, unsigned int *levelp)
819 struct kvm *kvm = vcpu->kvm;
820 struct page *page = NULL;
821 unsigned long mmu_seq;
822 unsigned long hva, gfn = gpa >> PAGE_SHIFT;
823 bool upgrade_write = false;
824 bool *upgrade_p = &upgrade_write;
826 unsigned int shift, level;
830 /* used to check for invalidations in progress */
831 mmu_seq = kvm->mmu_notifier_seq;
835 * Do a fast check first, since __gfn_to_pfn_memslot doesn't
836 * do it with !atomic && !async, which is how we call it.
837 * We always ask for write permission since the common case
838 * is that the page is writable.
840 hva = gfn_to_hva_memslot(memslot, gfn);
841 if (!kvm_ro && get_user_page_fast_only(hva, FOLL_WRITE, &page)) {
842 upgrade_write = true;
846 /* Call KVM generic code to do the slow-path check */
847 pfn = __gfn_to_pfn_memslot(memslot, gfn, false, NULL,
848 writing, upgrade_p, NULL);
849 if (is_error_noslot_pfn(pfn))
852 if (pfn_valid(pfn)) {
853 page = pfn_to_page(pfn);
854 if (PageReserved(page))
860 * Read the PTE from the process' radix tree and use that
861 * so we get the shift and attribute bits.
863 spin_lock(&kvm->mmu_lock);
864 ptep = find_kvm_host_pte(kvm, mmu_seq, hva, &shift);
867 pte = READ_ONCE(*ptep);
868 spin_unlock(&kvm->mmu_lock);
870 * If the PTE disappeared temporarily due to a THP
871 * collapse, just return and let the guest try again.
873 if (!pte_present(pte)) {
879 /* If we're logging dirty pages, always map single pages */
880 large_enable = !(memslot->flags & KVM_MEM_LOG_DIRTY_PAGES);
882 /* Get pte level from shift/size */
883 if (large_enable && shift == PUD_SHIFT &&
884 (gpa & (PUD_SIZE - PAGE_SIZE)) ==
885 (hva & (PUD_SIZE - PAGE_SIZE))) {
887 } else if (large_enable && shift == PMD_SHIFT &&
888 (gpa & (PMD_SIZE - PAGE_SIZE)) ==
889 (hva & (PMD_SIZE - PAGE_SIZE))) {
893 if (shift > PAGE_SHIFT) {
895 * If the pte maps more than one page, bring over
896 * bits from the virtual address to get the real
897 * address of the specific single page we want.
899 unsigned long rpnmask = (1ul << shift) - PAGE_SIZE;
900 pte = __pte(pte_val(pte) | (hva & rpnmask));
904 pte = __pte(pte_val(pte) | _PAGE_EXEC | _PAGE_ACCESSED);
905 if (writing || upgrade_write) {
906 if (pte_val(pte) & _PAGE_WRITE)
907 pte = __pte(pte_val(pte) | _PAGE_DIRTY);
909 pte = __pte(pte_val(pte) & ~(_PAGE_WRITE | _PAGE_DIRTY));
912 /* Allocate space in the tree and write the PTE */
913 ret = kvmppc_create_pte(kvm, kvm->arch.pgtable, pte, gpa, level,
914 mmu_seq, kvm->arch.lpid, NULL, NULL);
921 if (!ret && (pte_val(pte) & _PAGE_WRITE))
922 set_page_dirty_lock(page);
926 /* Increment number of large pages if we (successfully) inserted one */
929 kvm->stat.num_2M_pages++;
931 kvm->stat.num_1G_pages++;
937 int kvmppc_book3s_radix_page_fault(struct kvm_vcpu *vcpu,
938 unsigned long ea, unsigned long dsisr)
940 struct kvm *kvm = vcpu->kvm;
941 unsigned long gpa, gfn;
942 struct kvm_memory_slot *memslot;
944 bool writing = !!(dsisr & DSISR_ISSTORE);
947 /* Check for unusual errors */
948 if (dsisr & DSISR_UNSUPP_MMU) {
949 pr_err("KVM: Got unsupported MMU fault\n");
952 if (dsisr & DSISR_BADACCESS) {
953 /* Reflect to the guest as DSI */
954 pr_err("KVM: Got radix HV page fault with DSISR=%lx\n", dsisr);
955 kvmppc_core_queue_data_storage(vcpu, ea, dsisr);
959 /* Translate the logical address */
960 gpa = vcpu->arch.fault_gpa & ~0xfffUL;
961 gpa &= ~0xF000000000000000ul;
962 gfn = gpa >> PAGE_SHIFT;
963 if (!(dsisr & DSISR_PRTABLE_FAULT))
966 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
967 return kvmppc_send_page_to_uv(kvm, gfn);
969 /* Get the corresponding memslot */
970 memslot = gfn_to_memslot(kvm, gfn);
972 /* No memslot means it's an emulated MMIO region */
973 if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID)) {
974 if (dsisr & (DSISR_PRTABLE_FAULT | DSISR_BADACCESS |
977 * Bad address in guest page table tree, or other
978 * unusual error - reflect it to the guest as DSI.
980 kvmppc_core_queue_data_storage(vcpu, ea, dsisr);
983 return kvmppc_hv_emulate_mmio(vcpu, gpa, ea, writing);
986 if (memslot->flags & KVM_MEM_READONLY) {
988 /* give the guest a DSI */
989 kvmppc_core_queue_data_storage(vcpu, ea, DSISR_ISSTORE |
996 /* Failed to set the reference/change bits */
997 if (dsisr & DSISR_SET_RC) {
998 spin_lock(&kvm->mmu_lock);
999 if (kvmppc_hv_handle_set_rc(kvm, false, writing,
1000 gpa, kvm->arch.lpid))
1001 dsisr &= ~DSISR_SET_RC;
1002 spin_unlock(&kvm->mmu_lock);
1004 if (!(dsisr & (DSISR_BAD_FAULT_64S | DSISR_NOHPTE |
1005 DSISR_PROTFAULT | DSISR_SET_RC)))
1006 return RESUME_GUEST;
1009 /* Try to insert a pte */
1010 ret = kvmppc_book3s_instantiate_page(vcpu, gpa, memslot, writing,
1011 kvm_ro, NULL, NULL);
1013 if (ret == 0 || ret == -EAGAIN)
1018 /* Called with kvm->mmu_lock held */
1019 void kvm_unmap_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
1023 unsigned long gpa = gfn << PAGE_SHIFT;
1026 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE) {
1027 uv_page_inval(kvm->arch.lpid, gpa, PAGE_SHIFT);
1031 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1032 if (ptep && pte_present(*ptep))
1033 kvmppc_unmap_pte(kvm, ptep, gpa, shift, memslot,
1037 /* Called with kvm->mmu_lock held */
1038 bool kvm_age_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
1042 unsigned long gpa = gfn << PAGE_SHIFT;
1045 unsigned long old, *rmapp;
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)) {
1052 old = kvmppc_radix_update_pte(kvm, ptep, _PAGE_ACCESSED, 0,
1054 /* XXX need to flush tlb here? */
1055 /* Also clear bit in ptes in shadow pgtable for nested guests */
1056 rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1057 kvmhv_update_nest_rmap_rc_list(kvm, rmapp, _PAGE_ACCESSED, 0,
1065 /* Called with kvm->mmu_lock held */
1066 bool kvm_test_age_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
1071 unsigned long gpa = gfn << PAGE_SHIFT;
1075 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1078 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1079 if (ptep && pte_present(*ptep) && pte_young(*ptep))
1084 /* Returns the number of PAGE_SIZE pages that are dirty */
1085 static int kvm_radix_test_clear_dirty(struct kvm *kvm,
1086 struct kvm_memory_slot *memslot, int pagenum)
1088 unsigned long gfn = memslot->base_gfn + pagenum;
1089 unsigned long gpa = gfn << PAGE_SHIFT;
1093 unsigned long old, *rmapp;
1095 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1099 * For performance reasons we don't hold kvm->mmu_lock while walking the
1100 * partition scoped table.
1102 ptep = find_kvm_secondary_pte_unlocked(kvm, gpa, &shift);
1106 pte = READ_ONCE(*ptep);
1107 if (pte_present(pte) && pte_dirty(pte)) {
1108 spin_lock(&kvm->mmu_lock);
1110 * Recheck the pte again
1112 if (pte_val(pte) != pte_val(*ptep)) {
1114 * We have KVM_MEM_LOG_DIRTY_PAGES enabled. Hence we can
1115 * only find PAGE_SIZE pte entries here. We can continue
1116 * to use the pte addr returned by above page table
1119 if (!pte_present(*ptep) || !pte_dirty(*ptep)) {
1120 spin_unlock(&kvm->mmu_lock);
1127 old = kvmppc_radix_update_pte(kvm, ptep, _PAGE_DIRTY, 0,
1129 kvmppc_radix_tlbie_page(kvm, gpa, shift, kvm->arch.lpid);
1130 /* Also clear bit in ptes in shadow pgtable for nested guests */
1131 rmapp = &memslot->arch.rmap[gfn - memslot->base_gfn];
1132 kvmhv_update_nest_rmap_rc_list(kvm, rmapp, _PAGE_DIRTY, 0,
1135 spin_unlock(&kvm->mmu_lock);
1140 long kvmppc_hv_get_dirty_log_radix(struct kvm *kvm,
1141 struct kvm_memory_slot *memslot, unsigned long *map)
1146 for (i = 0; i < memslot->npages; i = j) {
1147 npages = kvm_radix_test_clear_dirty(kvm, memslot, i);
1150 * Note that if npages > 0 then i must be a multiple of npages,
1151 * since huge pages are only used to back the guest at guest
1152 * real addresses that are a multiple of their size.
1153 * Since we have at most one PTE covering any given guest
1154 * real address, if npages > 1 we can skip to i + npages.
1158 set_dirty_bits(map, i, npages);
1165 void kvmppc_radix_flush_memslot(struct kvm *kvm,
1166 const struct kvm_memory_slot *memslot)
1173 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_START)
1174 kvmppc_uvmem_drop_pages(memslot, kvm, true);
1176 if (kvm->arch.secure_guest & KVMPPC_SECURE_INIT_DONE)
1179 gpa = memslot->base_gfn << PAGE_SHIFT;
1180 spin_lock(&kvm->mmu_lock);
1181 for (n = memslot->npages; n; --n) {
1182 ptep = find_kvm_secondary_pte(kvm, gpa, &shift);
1183 if (ptep && pte_present(*ptep))
1184 kvmppc_unmap_pte(kvm, ptep, gpa, shift, memslot,
1189 * Increase the mmu notifier sequence number to prevent any page
1190 * fault that read the memslot earlier from writing a PTE.
1192 kvm->mmu_notifier_seq++;
1193 spin_unlock(&kvm->mmu_lock);
1196 static void add_rmmu_ap_encoding(struct kvm_ppc_rmmu_info *info,
1197 int psize, int *indexp)
1199 if (!mmu_psize_defs[psize].shift)
1201 info->ap_encodings[*indexp] = mmu_psize_defs[psize].shift |
1202 (mmu_psize_defs[psize].ap << 29);
1206 int kvmhv_get_rmmu_info(struct kvm *kvm, struct kvm_ppc_rmmu_info *info)
1210 if (!radix_enabled())
1212 memset(info, 0, sizeof(*info));
1215 info->geometries[0].page_shift = 12;
1216 info->geometries[0].level_bits[0] = 9;
1217 for (i = 1; i < 4; ++i)
1218 info->geometries[0].level_bits[i] = p9_supported_radix_bits[i];
1220 info->geometries[1].page_shift = 16;
1221 for (i = 0; i < 4; ++i)
1222 info->geometries[1].level_bits[i] = p9_supported_radix_bits[i];
1225 add_rmmu_ap_encoding(info, MMU_PAGE_4K, &i);
1226 add_rmmu_ap_encoding(info, MMU_PAGE_64K, &i);
1227 add_rmmu_ap_encoding(info, MMU_PAGE_2M, &i);
1228 add_rmmu_ap_encoding(info, MMU_PAGE_1G, &i);
1233 int kvmppc_init_vm_radix(struct kvm *kvm)
1235 kvm->arch.pgtable = pgd_alloc(kvm->mm);
1236 if (!kvm->arch.pgtable)
1241 static void pte_ctor(void *addr)
1243 memset(addr, 0, RADIX_PTE_TABLE_SIZE);
1246 static void pmd_ctor(void *addr)
1248 memset(addr, 0, RADIX_PMD_TABLE_SIZE);
1251 struct debugfs_radix_state {
1262 static int debugfs_radix_open(struct inode *inode, struct file *file)
1264 struct kvm *kvm = inode->i_private;
1265 struct debugfs_radix_state *p;
1267 p = kzalloc(sizeof(*p), GFP_KERNEL);
1273 mutex_init(&p->mutex);
1274 file->private_data = p;
1276 return nonseekable_open(inode, file);
1279 static int debugfs_radix_release(struct inode *inode, struct file *file)
1281 struct debugfs_radix_state *p = file->private_data;
1283 kvm_put_kvm(p->kvm);
1288 static ssize_t debugfs_radix_read(struct file *file, char __user *buf,
1289 size_t len, loff_t *ppos)
1291 struct debugfs_radix_state *p = file->private_data;
1297 struct kvm_nested_guest *nested;
1307 if (!kvm_is_radix(kvm))
1310 ret = mutex_lock_interruptible(&p->mutex);
1314 if (p->chars_left) {
1318 r = copy_to_user(buf, p->buf + p->buf_index, n);
1335 while (len != 0 && p->lpid >= 0) {
1336 if (gpa >= RADIX_PGTABLE_RANGE) {
1340 kvmhv_put_nested(nested);
1343 p->lpid = kvmhv_nested_next_lpid(kvm, p->lpid);
1350 pgt = kvm->arch.pgtable;
1352 nested = kvmhv_get_nested(kvm, p->lpid, false);
1354 gpa = RADIX_PGTABLE_RANGE;
1357 pgt = nested->shadow_pgtable;
1363 n = scnprintf(p->buf, sizeof(p->buf),
1364 "\nNested LPID %d: ", p->lpid);
1365 n += scnprintf(p->buf + n, sizeof(p->buf) - n,
1366 "pgdir: %lx\n", (unsigned long)pgt);
1371 pgdp = pgt + pgd_index(gpa);
1372 p4dp = p4d_offset(pgdp, gpa);
1373 p4d = READ_ONCE(*p4dp);
1374 if (!(p4d_val(p4d) & _PAGE_PRESENT)) {
1375 gpa = (gpa & P4D_MASK) + P4D_SIZE;
1379 pudp = pud_offset(&p4d, gpa);
1380 pud = READ_ONCE(*pudp);
1381 if (!(pud_val(pud) & _PAGE_PRESENT)) {
1382 gpa = (gpa & PUD_MASK) + PUD_SIZE;
1385 if (pud_val(pud) & _PAGE_PTE) {
1391 pmdp = pmd_offset(&pud, gpa);
1392 pmd = READ_ONCE(*pmdp);
1393 if (!(pmd_val(pmd) & _PAGE_PRESENT)) {
1394 gpa = (gpa & PMD_MASK) + PMD_SIZE;
1397 if (pmd_val(pmd) & _PAGE_PTE) {
1403 ptep = pte_offset_kernel(&pmd, gpa);
1404 pte = pte_val(READ_ONCE(*ptep));
1405 if (!(pte & _PAGE_PRESENT)) {
1411 n = scnprintf(p->buf, sizeof(p->buf),
1412 " %lx: %lx %d\n", gpa, pte, shift);
1413 gpa += 1ul << shift;
1418 r = copy_to_user(buf, p->buf, n);
1433 kvmhv_put_nested(nested);
1436 mutex_unlock(&p->mutex);
1440 static ssize_t debugfs_radix_write(struct file *file, const char __user *buf,
1441 size_t len, loff_t *ppos)
1446 static const struct file_operations debugfs_radix_fops = {
1447 .owner = THIS_MODULE,
1448 .open = debugfs_radix_open,
1449 .release = debugfs_radix_release,
1450 .read = debugfs_radix_read,
1451 .write = debugfs_radix_write,
1452 .llseek = generic_file_llseek,
1455 void kvmhv_radix_debugfs_init(struct kvm *kvm)
1457 debugfs_create_file("radix", 0400, kvm->debugfs_dentry, kvm,
1458 &debugfs_radix_fops);
1461 int kvmppc_radix_init(void)
1463 unsigned long size = sizeof(void *) << RADIX_PTE_INDEX_SIZE;
1465 kvm_pte_cache = kmem_cache_create("kvm-pte", size, size, 0, pte_ctor);
1469 size = sizeof(void *) << RADIX_PMD_INDEX_SIZE;
1471 kvm_pmd_cache = kmem_cache_create("kvm-pmd", size, size, 0, pmd_ctor);
1472 if (!kvm_pmd_cache) {
1473 kmem_cache_destroy(kvm_pte_cache);
1480 void kvmppc_radix_exit(void)
1482 kmem_cache_destroy(kvm_pte_cache);
1483 kmem_cache_destroy(kvm_pmd_cache);