2 * This program is free software; you can redistribute it and/or modify
3 * it under the terms of the GNU General Public License, version 2, as
4 * published by the Free Software Foundation.
6 * Copyright 2016 Paul Mackerras, IBM Corp. <paulus@au1.ibm.com>
9 #include <linux/types.h>
10 #include <linux/string.h>
11 #include <linux/kvm.h>
12 #include <linux/kvm_host.h>
14 #include <asm/kvm_ppc.h>
15 #include <asm/kvm_book3s.h>
18 #include <asm/pgtable.h>
19 #include <asm/pgalloc.h>
22 * Supported radix tree geometry.
23 * Like p9, we support either 5 or 9 bits at the first (lowest) level,
24 * for a page size of 64k or 4k.
26 static int p9_supported_radix_bits[4] = { 5, 9, 9, 13 };
28 int kvmppc_mmu_radix_xlate(struct kvm_vcpu *vcpu, gva_t eaddr,
29 struct kvmppc_pte *gpte, bool data, bool iswrite)
31 struct kvm *kvm = vcpu->kvm;
36 unsigned long root, pte, index;
37 unsigned long rts, bits, offset;
39 unsigned long proc_tbl_size;
41 /* Work out effective PID */
42 switch (eaddr >> 62) {
52 proc_tbl_size = 1 << ((kvm->arch.process_table & PRTS_MASK) + 12);
53 if (pid * 16 >= proc_tbl_size)
56 /* Read partition table to find root of tree for effective PID */
57 ptbl = (kvm->arch.process_table & PRTB_MASK) + (pid * 16);
58 ret = kvm_read_guest(kvm, ptbl, &prte, sizeof(prte));
62 root = be64_to_cpu(prte);
63 rts = ((root & RTS1_MASK) >> (RTS1_SHIFT - 3)) |
64 ((root & RTS2_MASK) >> RTS2_SHIFT);
65 bits = root & RPDS_MASK;
66 root = root & RPDB_MASK;
68 /* P9 DD1 interprets RTS (radix tree size) differently */
70 if (cpu_has_feature(CPU_FTR_POWER9_DD1))
73 /* current implementations only support 52-bit space */
77 for (level = 3; level >= 0; --level) {
78 if (level && bits != p9_supported_radix_bits[level])
80 if (level == 0 && !(bits == 5 || bits == 9))
83 index = (eaddr >> offset) & ((1UL << bits) - 1);
84 /* check that low bits of page table base are zero */
85 if (root & ((1UL << (bits + 3)) - 1))
87 ret = kvm_read_guest(kvm, root + index * 8,
91 pte = __be64_to_cpu(rpte);
92 if (!(pte & _PAGE_PRESENT))
97 root = pte & 0x0fffffffffffff00ul;
99 /* need a leaf at lowest level; 512GB pages not supported */
100 if (level < 0 || level == 3)
103 /* offset is now log base 2 of the page size */
104 gpa = pte & 0x01fffffffffff000ul;
105 if (gpa & ((1ul << offset) - 1))
107 gpa += eaddr & ((1ul << offset) - 1);
108 for (ps = MMU_PAGE_4K; ps < MMU_PAGE_COUNT; ++ps)
109 if (offset == mmu_psize_defs[ps].shift)
111 gpte->page_size = ps;
116 /* Work out permissions */
117 gpte->may_read = !!(pte & _PAGE_READ);
118 gpte->may_write = !!(pte & _PAGE_WRITE);
119 gpte->may_execute = !!(pte & _PAGE_EXEC);
120 if (kvmppc_get_msr(vcpu) & MSR_PR) {
121 if (pte & _PAGE_PRIVILEGED) {
124 gpte->may_execute = 0;
127 if (!(pte & _PAGE_PRIVILEGED)) {
128 /* Check AMR/IAMR to see if strict mode is in force */
129 if (vcpu->arch.amr & (1ul << 62))
131 if (vcpu->arch.amr & (1ul << 63))
133 if (vcpu->arch.iamr & (1ul << 62))
134 gpte->may_execute = 0;
141 #ifdef CONFIG_PPC_64K_PAGES
142 #define MMU_BASE_PSIZE MMU_PAGE_64K
144 #define MMU_BASE_PSIZE MMU_PAGE_4K
147 static void kvmppc_radix_tlbie_page(struct kvm *kvm, unsigned long addr,
150 int psize = MMU_BASE_PSIZE;
152 if (pshift >= PMD_SHIFT)
155 addr |= mmu_psize_defs[psize].ap << 5;
156 asm volatile("ptesync": : :"memory");
157 asm volatile(PPC_TLBIE_5(%0, %1, 0, 0, 1)
158 : : "r" (addr), "r" (kvm->arch.lpid) : "memory");
159 asm volatile("ptesync": : :"memory");
162 unsigned long kvmppc_radix_update_pte(struct kvm *kvm, pte_t *ptep,
163 unsigned long clr, unsigned long set,
164 unsigned long addr, unsigned int shift)
166 unsigned long old = 0;
168 if (!(clr & _PAGE_PRESENT) && cpu_has_feature(CPU_FTR_POWER9_DD1) &&
169 pte_present(*ptep)) {
170 /* have to invalidate it first */
171 old = __radix_pte_update(ptep, _PAGE_PRESENT, 0);
172 kvmppc_radix_tlbie_page(kvm, addr, shift);
173 set |= _PAGE_PRESENT;
174 old &= _PAGE_PRESENT;
176 return __radix_pte_update(ptep, clr, set) | old;
179 void kvmppc_radix_set_pte_at(struct kvm *kvm, unsigned long addr,
180 pte_t *ptep, pte_t pte)
182 radix__set_pte_at(kvm->mm, addr, ptep, pte, 0);
185 static struct kmem_cache *kvm_pte_cache;
187 static pte_t *kvmppc_pte_alloc(void)
189 return kmem_cache_alloc(kvm_pte_cache, GFP_KERNEL);
192 static void kvmppc_pte_free(pte_t *ptep)
194 kmem_cache_free(kvm_pte_cache, ptep);
197 static int kvmppc_create_pte(struct kvm *kvm, pte_t pte, unsigned long gpa,
198 unsigned int level, unsigned long mmu_seq)
201 pud_t *pud, *new_pud = NULL;
202 pmd_t *pmd, *new_pmd = NULL;
203 pte_t *ptep, *new_ptep = NULL;
207 /* Traverse the guest's 2nd-level tree, allocate new levels needed */
208 pgd = kvm->arch.pgtable + pgd_index(gpa);
210 if (pgd_present(*pgd))
211 pud = pud_offset(pgd, gpa);
213 new_pud = pud_alloc_one(kvm->mm, gpa);
216 if (pud && pud_present(*pud))
217 pmd = pmd_offset(pud, gpa);
219 new_pmd = pmd_alloc_one(kvm->mm, gpa);
221 if (level == 0 && !(pmd && pmd_present(*pmd)))
222 new_ptep = kvmppc_pte_alloc();
224 /* Check if we might have been invalidated; let the guest retry if so */
225 spin_lock(&kvm->mmu_lock);
227 if (mmu_notifier_retry(kvm, mmu_seq))
230 /* Now traverse again under the lock and change the tree */
232 if (pgd_none(*pgd)) {
235 pgd_populate(kvm->mm, pgd, new_pud);
238 pud = pud_offset(pgd, gpa);
239 if (pud_none(*pud)) {
242 pud_populate(kvm->mm, pud, new_pmd);
245 pmd = pmd_offset(pud, gpa);
246 if (pmd_large(*pmd)) {
247 /* Someone else has instantiated a large page here; retry */
251 if (level == 1 && !pmd_none(*pmd)) {
253 * There's a page table page here, but we wanted
254 * to install a large page. Tell the caller and let
255 * it try installing a normal page if it wants.
261 if (pmd_none(*pmd)) {
264 pmd_populate(kvm->mm, pmd, new_ptep);
267 ptep = pte_offset_kernel(pmd, gpa);
268 if (pte_present(*ptep)) {
269 /* PTE was previously valid, so invalidate it */
270 old = kvmppc_radix_update_pte(kvm, ptep, _PAGE_PRESENT,
272 kvmppc_radix_tlbie_page(kvm, gpa, 0);
273 if (old & _PAGE_DIRTY)
274 mark_page_dirty(kvm, gpa >> PAGE_SHIFT);
276 kvmppc_radix_set_pte_at(kvm, gpa, ptep, pte);
278 kvmppc_radix_set_pte_at(kvm, gpa, pmdp_ptep(pmd), pte);
283 spin_unlock(&kvm->mmu_lock);
285 pud_free(kvm->mm, new_pud);
287 pmd_free(kvm->mm, new_pmd);
289 kvmppc_pte_free(new_ptep);
293 int kvmppc_book3s_radix_page_fault(struct kvm_run *run, struct kvm_vcpu *vcpu,
294 unsigned long ea, unsigned long dsisr)
296 struct kvm *kvm = vcpu->kvm;
297 unsigned long mmu_seq, pte_size;
298 unsigned long gpa, gfn, hva, pfn;
299 struct kvm_memory_slot *memslot;
300 struct page *page = NULL, *pages[1];
301 long ret, npages, ok;
302 unsigned int writing;
303 struct vm_area_struct *vma;
306 unsigned long pgflags;
307 unsigned int shift, level;
309 /* Check for unusual errors */
310 if (dsisr & DSISR_UNSUPP_MMU) {
311 pr_err("KVM: Got unsupported MMU fault\n");
314 if (dsisr & DSISR_BADACCESS) {
315 /* Reflect to the guest as DSI */
316 pr_err("KVM: Got radix HV page fault with DSISR=%lx\n", dsisr);
317 kvmppc_core_queue_data_storage(vcpu, ea, dsisr);
321 /* Translate the logical address and get the page */
322 gpa = vcpu->arch.fault_gpa & ~0xfffUL;
323 gpa &= ~0xF000000000000000ul;
324 gfn = gpa >> PAGE_SHIFT;
325 if (!(dsisr & DSISR_PGDIRFAULT))
327 memslot = gfn_to_memslot(kvm, gfn);
329 /* No memslot means it's an emulated MMIO region */
330 if (!memslot || (memslot->flags & KVM_MEMSLOT_INVALID)) {
331 if (dsisr & (DSISR_PGDIRFAULT | DSISR_BADACCESS |
334 * Bad address in guest page table tree, or other
335 * unusual error - reflect it to the guest as DSI.
337 kvmppc_core_queue_data_storage(vcpu, ea, dsisr);
340 return kvmppc_hv_emulate_mmio(run, vcpu, gpa, ea,
341 dsisr & DSISR_ISSTORE);
344 /* used to check for invalidations in progress */
345 mmu_seq = kvm->mmu_notifier_seq;
348 writing = (dsisr & DSISR_ISSTORE) != 0;
349 hva = gfn_to_hva_memslot(memslot, gfn);
350 if (dsisr & DSISR_SET_RC) {
352 * Need to set an R or C bit in the 2nd-level tables;
353 * if the relevant bits aren't already set in the linux
354 * page tables, fall through to do the gup_fast to
355 * set them in the linux page tables too.
358 pgflags = _PAGE_ACCESSED;
360 pgflags |= _PAGE_DIRTY;
361 local_irq_save(flags);
362 ptep = __find_linux_pte_or_hugepte(current->mm->pgd, hva,
365 pte = READ_ONCE(*ptep);
366 if (pte_present(pte) &&
367 (pte_val(pte) & pgflags) == pgflags)
370 local_irq_restore(flags);
372 spin_lock(&kvm->mmu_lock);
373 if (mmu_notifier_retry(vcpu->kvm, mmu_seq)) {
374 spin_unlock(&kvm->mmu_lock);
377 ptep = __find_linux_pte_or_hugepte(kvm->arch.pgtable,
379 if (ptep && pte_present(*ptep)) {
380 kvmppc_radix_update_pte(kvm, ptep, 0, pgflags,
382 spin_unlock(&kvm->mmu_lock);
385 spin_unlock(&kvm->mmu_lock);
391 pte_size = PAGE_SIZE;
392 pgflags = _PAGE_READ | _PAGE_EXEC;
394 npages = get_user_pages_fast(hva, 1, writing, pages);
396 /* Check if it's an I/O mapping */
397 down_read(¤t->mm->mmap_sem);
398 vma = find_vma(current->mm, hva);
399 if (vma && vma->vm_start <= hva && hva < vma->vm_end &&
400 (vma->vm_flags & VM_PFNMAP)) {
401 pfn = vma->vm_pgoff +
402 ((hva - vma->vm_start) >> PAGE_SHIFT);
403 pgflags = pgprot_val(vma->vm_page_prot);
405 up_read(¤t->mm->mmap_sem);
410 pfn = page_to_pfn(page);
411 if (PageHuge(page)) {
412 page = compound_head(page);
413 pte_size <<= compound_order(page);
414 /* See if we can insert a 2MB large-page PTE here */
415 if (pte_size >= PMD_SIZE &&
416 (gpa & PMD_MASK & PAGE_MASK) ==
417 (hva & PMD_MASK & PAGE_MASK)) {
419 pfn &= ~((PMD_SIZE >> PAGE_SHIFT) - 1);
422 /* See if we can provide write access */
425 * We assume gup_fast has set dirty on the host PTE.
427 pgflags |= _PAGE_WRITE;
429 local_irq_save(flags);
430 ptep = __find_linux_pte_or_hugepte(current->mm->pgd,
432 if (ptep && pte_write(*ptep) && pte_dirty(*ptep))
433 pgflags |= _PAGE_WRITE;
434 local_irq_restore(flags);
439 * Compute the PTE value that we need to insert.
441 pgflags |= _PAGE_PRESENT | _PAGE_PTE | _PAGE_ACCESSED;
442 if (pgflags & _PAGE_WRITE)
443 pgflags |= _PAGE_DIRTY;
444 pte = pfn_pte(pfn, __pgprot(pgflags));
446 /* Allocate space in the tree and write the PTE */
447 ret = kvmppc_create_pte(kvm, pte, gpa, level, mmu_seq);
450 * There's already a PMD where wanted to install a large page;
451 * for now, fall back to installing a small page.
454 pfn |= gfn & ((PMD_SIZE >> PAGE_SHIFT) - 1);
455 pte = pfn_pte(pfn, __pgprot(pgflags));
456 ret = kvmppc_create_pte(kvm, pte, gpa, level, mmu_seq);
458 if (ret == 0 || ret == -EAGAIN)
463 * We drop pages[0] here, not page because page might
464 * have been set to the head page of a compound, but
465 * we have to drop the reference on the correct tail
466 * page to match the get inside gup()
473 static void mark_pages_dirty(struct kvm *kvm, struct kvm_memory_slot *memslot,
474 unsigned long gfn, unsigned int order)
476 unsigned long i, limit;
479 if (!memslot->dirty_bitmap)
481 limit = 1ul << order;
482 if (limit < BITS_PER_LONG) {
483 for (i = 0; i < limit; ++i)
484 mark_page_dirty(kvm, gfn + i);
487 dp = memslot->dirty_bitmap + (gfn - memslot->base_gfn);
488 limit /= BITS_PER_LONG;
489 for (i = 0; i < limit; ++i)
493 /* Called with kvm->lock held */
494 int kvm_unmap_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
498 unsigned long gpa = gfn << PAGE_SHIFT;
502 ptep = __find_linux_pte_or_hugepte(kvm->arch.pgtable, gpa,
504 if (ptep && pte_present(*ptep)) {
505 old = kvmppc_radix_update_pte(kvm, ptep, _PAGE_PRESENT, 0,
507 kvmppc_radix_tlbie_page(kvm, gpa, shift);
508 if (old & _PAGE_DIRTY) {
510 mark_page_dirty(kvm, gfn);
512 mark_pages_dirty(kvm, memslot,
513 gfn, shift - PAGE_SHIFT);
519 /* Called with kvm->lock held */
520 int kvm_age_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
524 unsigned long gpa = gfn << PAGE_SHIFT;
528 ptep = __find_linux_pte_or_hugepte(kvm->arch.pgtable, gpa,
530 if (ptep && pte_present(*ptep) && pte_young(*ptep)) {
531 kvmppc_radix_update_pte(kvm, ptep, _PAGE_ACCESSED, 0,
533 /* XXX need to flush tlb here? */
539 /* Called with kvm->lock held */
540 int kvm_test_age_radix(struct kvm *kvm, struct kvm_memory_slot *memslot,
544 unsigned long gpa = gfn << PAGE_SHIFT;
548 ptep = __find_linux_pte_or_hugepte(kvm->arch.pgtable, gpa,
550 if (ptep && pte_present(*ptep) && pte_young(*ptep))
555 /* Returns the number of PAGE_SIZE pages that are dirty */
556 static int kvm_radix_test_clear_dirty(struct kvm *kvm,
557 struct kvm_memory_slot *memslot, int pagenum)
559 unsigned long gfn = memslot->base_gfn + pagenum;
560 unsigned long gpa = gfn << PAGE_SHIFT;
565 ptep = __find_linux_pte_or_hugepte(kvm->arch.pgtable, gpa,
567 if (ptep && pte_present(*ptep) && pte_dirty(*ptep)) {
570 ret = 1 << (shift - PAGE_SHIFT);
571 kvmppc_radix_update_pte(kvm, ptep, _PAGE_DIRTY, 0,
573 kvmppc_radix_tlbie_page(kvm, gpa, shift);
578 long kvmppc_hv_get_dirty_log_radix(struct kvm *kvm,
579 struct kvm_memory_slot *memslot, unsigned long *map)
586 * Radix accumulates dirty bits in the first half of the
587 * memslot's dirty_bitmap area, for when pages are paged
588 * out or modified by the host directly. Pick up these
589 * bits and add them to the map.
591 n = kvm_dirty_bitmap_bytes(memslot) / sizeof(long);
592 p = memslot->dirty_bitmap;
593 for (i = 0; i < n; ++i)
594 map[i] |= xchg(&p[i], 0);
596 for (i = 0; i < memslot->npages; i = j) {
597 npages = kvm_radix_test_clear_dirty(kvm, memslot, i);
600 * Note that if npages > 0 then i must be a multiple of npages,
601 * since huge pages are only used to back the guest at guest
602 * real addresses that are a multiple of their size.
603 * Since we have at most one PTE covering any given guest
604 * real address, if npages > 1 we can skip to i + npages.
608 for (j = i; npages; ++j, --npages)
609 __set_bit_le(j, map);
614 static void add_rmmu_ap_encoding(struct kvm_ppc_rmmu_info *info,
615 int psize, int *indexp)
617 if (!mmu_psize_defs[psize].shift)
619 info->ap_encodings[*indexp] = mmu_psize_defs[psize].shift |
620 (mmu_psize_defs[psize].ap << 29);
624 int kvmhv_get_rmmu_info(struct kvm *kvm, struct kvm_ppc_rmmu_info *info)
628 if (!radix_enabled())
630 memset(info, 0, sizeof(*info));
633 info->geometries[0].page_shift = 12;
634 info->geometries[0].level_bits[0] = 9;
635 for (i = 1; i < 4; ++i)
636 info->geometries[0].level_bits[i] = p9_supported_radix_bits[i];
638 info->geometries[1].page_shift = 16;
639 for (i = 0; i < 4; ++i)
640 info->geometries[1].level_bits[i] = p9_supported_radix_bits[i];
643 add_rmmu_ap_encoding(info, MMU_PAGE_4K, &i);
644 add_rmmu_ap_encoding(info, MMU_PAGE_64K, &i);
645 add_rmmu_ap_encoding(info, MMU_PAGE_2M, &i);
646 add_rmmu_ap_encoding(info, MMU_PAGE_1G, &i);
651 int kvmppc_init_vm_radix(struct kvm *kvm)
653 kvm->arch.pgtable = pgd_alloc(kvm->mm);
654 if (!kvm->arch.pgtable)
659 void kvmppc_free_radix(struct kvm *kvm)
661 unsigned long ig, iu, im;
667 if (!kvm->arch.pgtable)
669 pgd = kvm->arch.pgtable;
670 for (ig = 0; ig < PTRS_PER_PGD; ++ig, ++pgd) {
671 if (!pgd_present(*pgd))
673 pud = pud_offset(pgd, 0);
674 for (iu = 0; iu < PTRS_PER_PUD; ++iu, ++pud) {
675 if (!pud_present(*pud))
677 pmd = pmd_offset(pud, 0);
678 for (im = 0; im < PTRS_PER_PMD; ++im, ++pmd) {
679 if (pmd_huge(*pmd)) {
683 if (!pmd_present(*pmd))
685 pte = pte_offset_map(pmd, 0);
686 memset(pte, 0, sizeof(long) << PTE_INDEX_SIZE);
687 kvmppc_pte_free(pte);
690 pmd_free(kvm->mm, pmd_offset(pud, 0));
693 pud_free(kvm->mm, pud_offset(pgd, 0));
696 pgd_free(kvm->mm, kvm->arch.pgtable);
699 static void pte_ctor(void *addr)
701 memset(addr, 0, PTE_TABLE_SIZE);
704 int kvmppc_radix_init(void)
706 unsigned long size = sizeof(void *) << PTE_INDEX_SIZE;
708 kvm_pte_cache = kmem_cache_create("kvm-pte", size, size, 0, pte_ctor);
714 void kvmppc_radix_exit(void)
716 kmem_cache_destroy(kvm_pte_cache);