4 * This file contains the various mmu fetch and update operations.
5 * The most important job they must perform is the mapping between the
6 * domain's pfn and the overall machine mfns.
8 * Xen allows guests to directly update the pagetable, in a controlled
9 * fashion. In other words, the guest modifies the same pagetable
10 * that the CPU actually uses, which eliminates the overhead of having
11 * a separate shadow pagetable.
13 * In order to allow this, it falls on the guest domain to map its
14 * notion of a "physical" pfn - which is just a domain-local linear
15 * address - into a real "machine address" which the CPU's MMU can
18 * A pgd_t/pmd_t/pte_t will typically contain an mfn, and so can be
19 * inserted directly into the pagetable. When creating a new
20 * pte/pmd/pgd, it converts the passed pfn into an mfn. Conversely,
21 * when reading the content back with __(pgd|pmd|pte)_val, it converts
22 * the mfn back into a pfn.
24 * The other constraint is that all pages which make up a pagetable
25 * must be mapped read-only in the guest. This prevents uncontrolled
26 * guest updates to the pagetable. Xen strictly enforces this, and
27 * will disallow any pagetable update which will end up mapping a
28 * pagetable page RW, and will disallow using any writable page as a
31 * Naively, when loading %cr3 with the base of a new pagetable, Xen
32 * would need to validate the whole pagetable before going on.
33 * Naturally, this is quite slow. The solution is to "pin" a
34 * pagetable, which enforces all the constraints on the pagetable even
35 * when it is not actively in use. This menas that Xen can be assured
36 * that it is still valid when you do load it into %cr3, and doesn't
37 * need to revalidate it.
39 * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
41 #include <linux/sched.h>
42 #include <linux/highmem.h>
43 #include <linux/bug.h>
45 #include <asm/pgtable.h>
46 #include <asm/tlbflush.h>
47 #include <asm/mmu_context.h>
48 #include <asm/paravirt.h>
49 #include <asm/linkage.h>
51 #include <asm/xen/hypercall.h>
52 #include <asm/xen/hypervisor.h>
55 #include <xen/interface/xen.h>
57 #include "multicalls.h"
60 #define P2M_ENTRIES_PER_PAGE (PAGE_SIZE / sizeof(unsigned long))
61 #define TOP_ENTRIES (MAX_DOMAIN_PAGES / P2M_ENTRIES_PER_PAGE)
63 /* Placeholder for holes in the address space */
64 static unsigned long p2m_missing[P2M_ENTRIES_PER_PAGE] __page_aligned_data =
65 { [ 0 ... P2M_ENTRIES_PER_PAGE-1 ] = ~0UL };
67 /* Array of pointers to pages containing p2m entries */
68 static unsigned long *p2m_top[TOP_ENTRIES] __page_aligned_data =
69 { [ 0 ... TOP_ENTRIES - 1] = &p2m_missing[0] };
71 /* Arrays of p2m arrays expressed in mfns used for save/restore */
72 static unsigned long p2m_top_mfn[TOP_ENTRIES] __page_aligned_bss;
74 static unsigned long p2m_top_mfn_list[TOP_ENTRIES / P2M_ENTRIES_PER_PAGE]
77 static inline unsigned p2m_top_index(unsigned long pfn)
79 BUG_ON(pfn >= MAX_DOMAIN_PAGES);
80 return pfn / P2M_ENTRIES_PER_PAGE;
83 static inline unsigned p2m_index(unsigned long pfn)
85 return pfn % P2M_ENTRIES_PER_PAGE;
88 /* Build the parallel p2m_top_mfn structures */
89 void xen_setup_mfn_list_list(void)
93 for(pfn = 0; pfn < MAX_DOMAIN_PAGES; pfn += P2M_ENTRIES_PER_PAGE) {
94 unsigned topidx = p2m_top_index(pfn);
96 p2m_top_mfn[topidx] = virt_to_mfn(p2m_top[topidx]);
99 for(idx = 0; idx < ARRAY_SIZE(p2m_top_mfn_list); idx++) {
100 unsigned topidx = idx * P2M_ENTRIES_PER_PAGE;
101 p2m_top_mfn_list[idx] = virt_to_mfn(&p2m_top_mfn[topidx]);
104 BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
106 HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list =
107 virt_to_mfn(p2m_top_mfn_list);
108 HYPERVISOR_shared_info->arch.max_pfn = xen_start_info->nr_pages;
111 /* Set up p2m_top to point to the domain-builder provided p2m pages */
112 void __init xen_build_dynamic_phys_to_machine(void)
114 unsigned long *mfn_list = (unsigned long *)xen_start_info->mfn_list;
115 unsigned long max_pfn = min(MAX_DOMAIN_PAGES, xen_start_info->nr_pages);
118 for(pfn = 0; pfn < max_pfn; pfn += P2M_ENTRIES_PER_PAGE) {
119 unsigned topidx = p2m_top_index(pfn);
121 p2m_top[topidx] = &mfn_list[pfn];
125 unsigned long get_phys_to_machine(unsigned long pfn)
127 unsigned topidx, idx;
129 if (unlikely(pfn >= MAX_DOMAIN_PAGES))
130 return INVALID_P2M_ENTRY;
132 topidx = p2m_top_index(pfn);
133 idx = p2m_index(pfn);
134 return p2m_top[topidx][idx];
136 EXPORT_SYMBOL_GPL(get_phys_to_machine);
138 static void alloc_p2m(unsigned long **pp, unsigned long *mfnp)
143 p = (void *)__get_free_page(GFP_KERNEL | __GFP_NOFAIL);
146 for(i = 0; i < P2M_ENTRIES_PER_PAGE; i++)
147 p[i] = INVALID_P2M_ENTRY;
149 if (cmpxchg(pp, p2m_missing, p) != p2m_missing)
150 free_page((unsigned long)p);
152 *mfnp = virt_to_mfn(p);
155 void set_phys_to_machine(unsigned long pfn, unsigned long mfn)
157 unsigned topidx, idx;
159 if (unlikely(xen_feature(XENFEAT_auto_translated_physmap))) {
160 BUG_ON(pfn != mfn && mfn != INVALID_P2M_ENTRY);
164 if (unlikely(pfn >= MAX_DOMAIN_PAGES)) {
165 BUG_ON(mfn != INVALID_P2M_ENTRY);
169 topidx = p2m_top_index(pfn);
170 if (p2m_top[topidx] == p2m_missing) {
171 /* no need to allocate a page to store an invalid entry */
172 if (mfn == INVALID_P2M_ENTRY)
174 alloc_p2m(&p2m_top[topidx], &p2m_top_mfn[topidx]);
177 idx = p2m_index(pfn);
178 p2m_top[topidx][idx] = mfn;
181 xmaddr_t arbitrary_virt_to_machine(unsigned long address)
184 pte_t *pte = lookup_address(address, &level);
185 unsigned offset = address & ~PAGE_MASK;
189 return XMADDR((pte_mfn(*pte) << PAGE_SHIFT) + offset);
192 void make_lowmem_page_readonly(void *vaddr)
195 unsigned long address = (unsigned long)vaddr;
198 pte = lookup_address(address, &level);
201 ptev = pte_wrprotect(*pte);
203 if (HYPERVISOR_update_va_mapping(address, ptev, 0))
207 void make_lowmem_page_readwrite(void *vaddr)
210 unsigned long address = (unsigned long)vaddr;
213 pte = lookup_address(address, &level);
216 ptev = pte_mkwrite(*pte);
218 if (HYPERVISOR_update_va_mapping(address, ptev, 0))
223 static bool page_pinned(void *ptr)
225 struct page *page = virt_to_page(ptr);
227 return PagePinned(page);
230 static void extend_mmu_update(const struct mmu_update *update)
232 struct multicall_space mcs;
233 struct mmu_update *u;
235 mcs = xen_mc_extend_args(__HYPERVISOR_mmu_update, sizeof(*u));
240 mcs = __xen_mc_entry(sizeof(*u));
241 MULTI_mmu_update(mcs.mc, mcs.args, 1, NULL, DOMID_SELF);
248 void xen_set_pmd_hyper(pmd_t *ptr, pmd_t val)
256 u.ptr = virt_to_machine(ptr).maddr;
257 u.val = pmd_val_ma(val);
258 extend_mmu_update(&u);
260 xen_mc_issue(PARAVIRT_LAZY_MMU);
265 void xen_set_pmd(pmd_t *ptr, pmd_t val)
267 /* If page is not pinned, we can just update the entry
269 if (!page_pinned(ptr)) {
274 xen_set_pmd_hyper(ptr, val);
278 * Associate a virtual page frame with a given physical page frame
279 * and protection flags for that frame.
281 void set_pte_mfn(unsigned long vaddr, unsigned long mfn, pgprot_t flags)
288 pgd = swapper_pg_dir + pgd_index(vaddr);
289 if (pgd_none(*pgd)) {
293 pud = pud_offset(pgd, vaddr);
294 if (pud_none(*pud)) {
298 pmd = pmd_offset(pud, vaddr);
299 if (pmd_none(*pmd)) {
303 pte = pte_offset_kernel(pmd, vaddr);
304 /* <mfn,flags> stored as-is, to permit clearing entries */
305 xen_set_pte(pte, mfn_pte(mfn, flags));
308 * It's enough to flush this one mapping.
309 * (PGE mappings get flushed as well)
311 __flush_tlb_one(vaddr);
314 void xen_set_pte_at(struct mm_struct *mm, unsigned long addr,
315 pte_t *ptep, pte_t pteval)
317 /* updates to init_mm may be done without lock */
321 if (mm == current->mm || mm == &init_mm) {
322 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU) {
323 struct multicall_space mcs;
324 mcs = xen_mc_entry(0);
326 MULTI_update_va_mapping(mcs.mc, addr, pteval, 0);
327 xen_mc_issue(PARAVIRT_LAZY_MMU);
330 if (HYPERVISOR_update_va_mapping(addr, pteval, 0) == 0)
333 xen_set_pte(ptep, pteval);
340 pte_t xen_ptep_modify_prot_start(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
342 /* Just return the pte as-is. We preserve the bits on commit */
346 void xen_ptep_modify_prot_commit(struct mm_struct *mm, unsigned long addr,
347 pte_t *ptep, pte_t pte)
353 u.ptr = virt_to_machine(ptep).maddr | MMU_PT_UPDATE_PRESERVE_AD;
354 u.val = pte_val_ma(pte);
355 extend_mmu_update(&u);
357 xen_mc_issue(PARAVIRT_LAZY_MMU);
360 /* Assume pteval_t is equivalent to all the other *val_t types. */
361 static pteval_t pte_mfn_to_pfn(pteval_t val)
363 if (val & _PAGE_PRESENT) {
364 unsigned long mfn = (val & PTE_MASK) >> PAGE_SHIFT;
365 pteval_t flags = val & ~PTE_MASK;
366 val = ((pteval_t)mfn_to_pfn(mfn) << PAGE_SHIFT) | flags;
372 static pteval_t pte_pfn_to_mfn(pteval_t val)
374 if (val & _PAGE_PRESENT) {
375 unsigned long pfn = (val & PTE_MASK) >> PAGE_SHIFT;
376 pteval_t flags = val & ~PTE_MASK;
377 val = ((pteval_t)pfn_to_mfn(pfn) << PAGE_SHIFT) | flags;
383 pteval_t xen_pte_val(pte_t pte)
385 return pte_mfn_to_pfn(pte.pte);
388 pgdval_t xen_pgd_val(pgd_t pgd)
390 return pte_mfn_to_pfn(pgd.pgd);
393 pte_t xen_make_pte(pteval_t pte)
395 pte = pte_pfn_to_mfn(pte);
396 return native_make_pte(pte);
399 pgd_t xen_make_pgd(pgdval_t pgd)
401 pgd = pte_pfn_to_mfn(pgd);
402 return native_make_pgd(pgd);
405 pmdval_t xen_pmd_val(pmd_t pmd)
407 return pte_mfn_to_pfn(pmd.pmd);
410 void xen_set_pud_hyper(pud_t *ptr, pud_t val)
418 u.ptr = virt_to_machine(ptr).maddr;
419 u.val = pud_val_ma(val);
420 extend_mmu_update(&u);
422 xen_mc_issue(PARAVIRT_LAZY_MMU);
427 void xen_set_pud(pud_t *ptr, pud_t val)
429 /* If page is not pinned, we can just update the entry
431 if (!page_pinned(ptr)) {
436 xen_set_pud_hyper(ptr, val);
439 void xen_set_pte(pte_t *ptep, pte_t pte)
441 #ifdef CONFIG_X86_PAE
442 ptep->pte_high = pte.pte_high;
444 ptep->pte_low = pte.pte_low;
450 #ifdef CONFIG_X86_PAE
451 void xen_set_pte_atomic(pte_t *ptep, pte_t pte)
453 set_64bit((u64 *)ptep, native_pte_val(pte));
456 void xen_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
459 smp_wmb(); /* make sure low gets written first */
463 void xen_pmd_clear(pmd_t *pmdp)
465 set_pmd(pmdp, __pmd(0));
467 #endif /* CONFIG_X86_PAE */
469 pmd_t xen_make_pmd(pmdval_t pmd)
471 pmd = pte_pfn_to_mfn(pmd);
472 return native_make_pmd(pmd);
475 #if PAGETABLE_LEVELS == 4
476 pudval_t xen_pud_val(pud_t pud)
478 return pte_mfn_to_pfn(pud.pud);
481 pud_t xen_make_pud(pudval_t pud)
483 pud = pte_pfn_to_mfn(pud);
485 return native_make_pud(pud);
488 void xen_set_pgd_hyper(pgd_t *ptr, pgd_t val)
496 u.ptr = virt_to_machine(ptr).maddr;
497 u.val = pgd_val_ma(val);
498 extend_mmu_update(&u);
500 xen_mc_issue(PARAVIRT_LAZY_MMU);
505 void xen_set_pgd(pgd_t *ptr, pgd_t val)
507 /* If page is not pinned, we can just update the entry
509 if (!page_pinned(ptr)) {
514 xen_set_pgd_hyper(ptr, val);
516 #endif /* PAGETABLE_LEVELS == 4 */
519 (Yet another) pagetable walker. This one is intended for pinning a
520 pagetable. This means that it walks a pagetable and calls the
521 callback function on each page it finds making up the page table,
522 at every level. It walks the entire pagetable, but it only bothers
523 pinning pte pages which are below pte_limit. In the normal case
524 this will be TASK_SIZE, but at boot we need to pin up to
525 FIXADDR_TOP. But the important bit is that we don't pin beyond
526 there, because then we start getting into Xen's ptes.
528 static int pgd_walk(pgd_t *pgd_base, int (*func)(struct page *, enum pt_level),
531 pgd_t *pgd = pgd_base;
533 unsigned long addr = 0;
534 unsigned long pgd_next;
536 BUG_ON(limit > FIXADDR_TOP);
538 if (xen_feature(XENFEAT_auto_translated_physmap))
541 for (; addr != FIXADDR_TOP; pgd++, addr = pgd_next) {
543 unsigned long pud_limit, pud_next;
545 pgd_next = pud_limit = pgd_addr_end(addr, FIXADDR_TOP);
550 pud = pud_offset(pgd, 0);
552 if (PTRS_PER_PUD > 1) /* not folded */
553 flush |= (*func)(virt_to_page(pud), PT_PUD);
555 for (; addr != pud_limit; pud++, addr = pud_next) {
557 unsigned long pmd_limit;
559 pud_next = pud_addr_end(addr, pud_limit);
561 if (pud_next < limit)
562 pmd_limit = pud_next;
569 pmd = pmd_offset(pud, 0);
571 if (PTRS_PER_PMD > 1) /* not folded */
572 flush |= (*func)(virt_to_page(pmd), PT_PMD);
574 for (; addr != pmd_limit; pmd++) {
575 addr += (PAGE_SIZE * PTRS_PER_PTE);
576 if ((pmd_limit-1) < (addr-1)) {
584 flush |= (*func)(pmd_page(*pmd), PT_PTE);
589 flush |= (*func)(virt_to_page(pgd_base), PT_PGD);
594 static spinlock_t *lock_pte(struct page *page)
596 spinlock_t *ptl = NULL;
598 #if NR_CPUS >= CONFIG_SPLIT_PTLOCK_CPUS
599 ptl = __pte_lockptr(page);
606 static void do_unlock(void *v)
612 static void xen_do_pin(unsigned level, unsigned long pfn)
614 struct mmuext_op *op;
615 struct multicall_space mcs;
617 mcs = __xen_mc_entry(sizeof(*op));
620 op->arg1.mfn = pfn_to_mfn(pfn);
621 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
624 static int pin_page(struct page *page, enum pt_level level)
626 unsigned pgfl = TestSetPagePinned(page);
630 flush = 0; /* already pinned */
631 else if (PageHighMem(page))
632 /* kmaps need flushing if we found an unpinned
636 void *pt = lowmem_page_address(page);
637 unsigned long pfn = page_to_pfn(page);
638 struct multicall_space mcs = __xen_mc_entry(0);
645 ptl = lock_pte(page);
647 MULTI_update_va_mapping(mcs.mc, (unsigned long)pt,
648 pfn_pte(pfn, PAGE_KERNEL_RO),
649 level == PT_PGD ? UVMF_TLB_FLUSH : 0);
652 xen_do_pin(MMUEXT_PIN_L1_TABLE, pfn);
655 /* Queue a deferred unlock for when this batch
657 xen_mc_callback(do_unlock, ptl);
664 /* This is called just after a mm has been created, but it has not
665 been used yet. We need to make sure that its pagetable is all
666 read-only, and can be pinned. */
667 void xen_pgd_pin(pgd_t *pgd)
671 if (pgd_walk(pgd, pin_page, TASK_SIZE)) {
672 /* re-enable interrupts for kmap_flush_unused */
678 xen_do_pin(MMUEXT_PIN_L3_TABLE, PFN_DOWN(__pa(pgd)));
683 * On save, we need to pin all pagetables to make sure they get their
684 * mfns turned into pfns. Search the list for any unpinned pgds and pin
685 * them (unpinned pgds are not currently in use, probably because the
686 * process is under construction or destruction).
688 void xen_mm_pin_all(void)
693 spin_lock_irqsave(&pgd_lock, flags);
695 list_for_each_entry(page, &pgd_list, lru) {
696 if (!PagePinned(page)) {
697 xen_pgd_pin((pgd_t *)page_address(page));
698 SetPageSavePinned(page);
702 spin_unlock_irqrestore(&pgd_lock, flags);
706 * The init_mm pagetable is really pinned as soon as its created, but
707 * that's before we have page structures to store the bits. So do all
708 * the book-keeping now.
710 static __init int mark_pinned(struct page *page, enum pt_level level)
716 void __init xen_mark_init_mm_pinned(void)
718 pgd_walk(init_mm.pgd, mark_pinned, FIXADDR_TOP);
721 static int unpin_page(struct page *page, enum pt_level level)
723 unsigned pgfl = TestClearPagePinned(page);
725 if (pgfl && !PageHighMem(page)) {
726 void *pt = lowmem_page_address(page);
727 unsigned long pfn = page_to_pfn(page);
728 spinlock_t *ptl = NULL;
729 struct multicall_space mcs;
731 if (level == PT_PTE) {
732 ptl = lock_pte(page);
734 xen_do_pin(MMUEXT_UNPIN_TABLE, pfn);
737 mcs = __xen_mc_entry(0);
739 MULTI_update_va_mapping(mcs.mc, (unsigned long)pt,
740 pfn_pte(pfn, PAGE_KERNEL),
741 level == PT_PGD ? UVMF_TLB_FLUSH : 0);
744 /* unlock when batch completed */
745 xen_mc_callback(do_unlock, ptl);
749 return 0; /* never need to flush on unpin */
752 /* Release a pagetables pages back as normal RW */
753 static void xen_pgd_unpin(pgd_t *pgd)
757 xen_do_pin(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
759 pgd_walk(pgd, unpin_page, TASK_SIZE);
765 * On resume, undo any pinning done at save, so that the rest of the
766 * kernel doesn't see any unexpected pinned pagetables.
768 void xen_mm_unpin_all(void)
773 spin_lock_irqsave(&pgd_lock, flags);
775 list_for_each_entry(page, &pgd_list, lru) {
776 if (PageSavePinned(page)) {
777 BUG_ON(!PagePinned(page));
778 printk("unpinning pinned %p\n", page_address(page));
779 xen_pgd_unpin((pgd_t *)page_address(page));
780 ClearPageSavePinned(page);
784 spin_unlock_irqrestore(&pgd_lock, flags);
787 void xen_activate_mm(struct mm_struct *prev, struct mm_struct *next)
789 spin_lock(&next->page_table_lock);
790 xen_pgd_pin(next->pgd);
791 spin_unlock(&next->page_table_lock);
794 void xen_dup_mmap(struct mm_struct *oldmm, struct mm_struct *mm)
796 spin_lock(&mm->page_table_lock);
797 xen_pgd_pin(mm->pgd);
798 spin_unlock(&mm->page_table_lock);
803 /* Another cpu may still have their %cr3 pointing at the pagetable, so
804 we need to repoint it somewhere else before we can unpin it. */
805 static void drop_other_mm_ref(void *info)
807 struct mm_struct *mm = info;
808 struct mm_struct *active_mm;
811 active_mm = read_pda(active_mm);
813 active_mm = __get_cpu_var(cpu_tlbstate).active_mm;
817 leave_mm(smp_processor_id());
819 /* If this cpu still has a stale cr3 reference, then make sure
820 it has been flushed. */
821 if (x86_read_percpu(xen_current_cr3) == __pa(mm->pgd)) {
822 load_cr3(swapper_pg_dir);
823 arch_flush_lazy_cpu_mode();
827 static void drop_mm_ref(struct mm_struct *mm)
832 if (current->active_mm == mm) {
833 if (current->mm == mm)
834 load_cr3(swapper_pg_dir);
836 leave_mm(smp_processor_id());
837 arch_flush_lazy_cpu_mode();
840 /* Get the "official" set of cpus referring to our pagetable. */
841 mask = mm->cpu_vm_mask;
843 /* It's possible that a vcpu may have a stale reference to our
844 cr3, because its in lazy mode, and it hasn't yet flushed
845 its set of pending hypercalls yet. In this case, we can
846 look at its actual current cr3 value, and force it to flush
848 for_each_online_cpu(cpu) {
849 if (per_cpu(xen_current_cr3, cpu) == __pa(mm->pgd))
853 if (!cpus_empty(mask))
854 smp_call_function_mask(mask, drop_other_mm_ref, mm, 1);
857 static void drop_mm_ref(struct mm_struct *mm)
859 if (current->active_mm == mm)
860 load_cr3(swapper_pg_dir);
865 * While a process runs, Xen pins its pagetables, which means that the
866 * hypervisor forces it to be read-only, and it controls all updates
867 * to it. This means that all pagetable updates have to go via the
868 * hypervisor, which is moderately expensive.
870 * Since we're pulling the pagetable down, we switch to use init_mm,
871 * unpin old process pagetable and mark it all read-write, which
872 * allows further operations on it to be simple memory accesses.
874 * The only subtle point is that another CPU may be still using the
875 * pagetable because of lazy tlb flushing. This means we need need to
876 * switch all CPUs off this pagetable before we can unpin it.
878 void xen_exit_mmap(struct mm_struct *mm)
880 get_cpu(); /* make sure we don't move around */
884 spin_lock(&mm->page_table_lock);
886 /* pgd may not be pinned in the error exit path of execve */
887 if (page_pinned(mm->pgd))
888 xen_pgd_unpin(mm->pgd);
890 spin_unlock(&mm->page_table_lock);