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>
50 #include <asm/xen/hypercall.h>
51 #include <asm/xen/hypervisor.h>
54 #include <xen/interface/xen.h>
56 #include "multicalls.h"
59 #define P2M_ENTRIES_PER_PAGE (PAGE_SIZE / sizeof(unsigned long))
60 #define TOP_ENTRIES (MAX_DOMAIN_PAGES / P2M_ENTRIES_PER_PAGE)
62 /* Placeholder for holes in the address space */
63 static unsigned long p2m_missing[P2M_ENTRIES_PER_PAGE]
64 __attribute__((section(".data.page_aligned"))) =
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]
69 __attribute__((section(".data.page_aligned"))) =
70 { [ 0 ... TOP_ENTRIES - 1] = &p2m_missing[0] };
72 /* Arrays of p2m arrays expressed in mfns used for save/restore */
73 static unsigned long p2m_top_mfn[TOP_ENTRIES]
74 __attribute__((section(".bss.page_aligned")));
76 static unsigned long p2m_top_mfn_list[
77 PAGE_ALIGN(TOP_ENTRIES / P2M_ENTRIES_PER_PAGE)]
78 __attribute__((section(".bss.page_aligned")));
80 static inline unsigned p2m_top_index(unsigned long pfn)
82 BUG_ON(pfn >= MAX_DOMAIN_PAGES);
83 return pfn / P2M_ENTRIES_PER_PAGE;
86 static inline unsigned p2m_index(unsigned long pfn)
88 return pfn % P2M_ENTRIES_PER_PAGE;
91 /* Build the parallel p2m_top_mfn structures */
92 void xen_setup_mfn_list_list(void)
96 for(pfn = 0; pfn < MAX_DOMAIN_PAGES; pfn += P2M_ENTRIES_PER_PAGE) {
97 unsigned topidx = p2m_top_index(pfn);
99 p2m_top_mfn[topidx] = virt_to_mfn(p2m_top[topidx]);
102 for(idx = 0; idx < ARRAY_SIZE(p2m_top_mfn_list); idx++) {
103 unsigned topidx = idx * P2M_ENTRIES_PER_PAGE;
104 p2m_top_mfn_list[idx] = virt_to_mfn(&p2m_top_mfn[topidx]);
107 BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
109 HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list =
110 virt_to_mfn(p2m_top_mfn_list);
111 HYPERVISOR_shared_info->arch.max_pfn = xen_start_info->nr_pages;
114 /* Set up p2m_top to point to the domain-builder provided p2m pages */
115 void __init xen_build_dynamic_phys_to_machine(void)
117 unsigned long *mfn_list = (unsigned long *)xen_start_info->mfn_list;
118 unsigned long max_pfn = min(MAX_DOMAIN_PAGES, xen_start_info->nr_pages);
121 for(pfn = 0; pfn < max_pfn; pfn += P2M_ENTRIES_PER_PAGE) {
122 unsigned topidx = p2m_top_index(pfn);
124 p2m_top[topidx] = &mfn_list[pfn];
128 unsigned long get_phys_to_machine(unsigned long pfn)
130 unsigned topidx, idx;
132 if (unlikely(pfn >= MAX_DOMAIN_PAGES))
133 return INVALID_P2M_ENTRY;
135 topidx = p2m_top_index(pfn);
136 idx = p2m_index(pfn);
137 return p2m_top[topidx][idx];
139 EXPORT_SYMBOL_GPL(get_phys_to_machine);
141 static void alloc_p2m(unsigned long **pp, unsigned long *mfnp)
146 p = (void *)__get_free_page(GFP_KERNEL | __GFP_NOFAIL);
149 for(i = 0; i < P2M_ENTRIES_PER_PAGE; i++)
150 p[i] = INVALID_P2M_ENTRY;
152 if (cmpxchg(pp, p2m_missing, p) != p2m_missing)
153 free_page((unsigned long)p);
155 *mfnp = virt_to_mfn(p);
158 void set_phys_to_machine(unsigned long pfn, unsigned long mfn)
160 unsigned topidx, idx;
162 if (unlikely(xen_feature(XENFEAT_auto_translated_physmap))) {
163 BUG_ON(pfn != mfn && mfn != INVALID_P2M_ENTRY);
167 if (unlikely(pfn >= MAX_DOMAIN_PAGES)) {
168 BUG_ON(mfn != INVALID_P2M_ENTRY);
172 topidx = p2m_top_index(pfn);
173 if (p2m_top[topidx] == p2m_missing) {
174 /* no need to allocate a page to store an invalid entry */
175 if (mfn == INVALID_P2M_ENTRY)
177 alloc_p2m(&p2m_top[topidx], &p2m_top_mfn[topidx]);
180 idx = p2m_index(pfn);
181 p2m_top[topidx][idx] = mfn;
184 xmaddr_t arbitrary_virt_to_machine(unsigned long address)
187 pte_t *pte = lookup_address(address, &level);
188 unsigned offset = address & ~PAGE_MASK;
192 return XMADDR((pte_mfn(*pte) << PAGE_SHIFT) + offset);
195 void make_lowmem_page_readonly(void *vaddr)
198 unsigned long address = (unsigned long)vaddr;
201 pte = lookup_address(address, &level);
204 ptev = pte_wrprotect(*pte);
206 if (HYPERVISOR_update_va_mapping(address, ptev, 0))
210 void make_lowmem_page_readwrite(void *vaddr)
213 unsigned long address = (unsigned long)vaddr;
216 pte = lookup_address(address, &level);
219 ptev = pte_mkwrite(*pte);
221 if (HYPERVISOR_update_va_mapping(address, ptev, 0))
226 static bool page_pinned(void *ptr)
228 struct page *page = virt_to_page(ptr);
230 return PagePinned(page);
233 void xen_set_pmd_hyper(pmd_t *ptr, pmd_t val)
235 struct multicall_space mcs;
236 struct mmu_update *u;
240 mcs = xen_mc_entry(sizeof(*u));
242 u->ptr = virt_to_machine(ptr).maddr;
243 u->val = pmd_val_ma(val);
244 MULTI_mmu_update(mcs.mc, u, 1, NULL, DOMID_SELF);
246 xen_mc_issue(PARAVIRT_LAZY_MMU);
251 void xen_set_pmd(pmd_t *ptr, pmd_t val)
253 /* If page is not pinned, we can just update the entry
255 if (!page_pinned(ptr)) {
260 xen_set_pmd_hyper(ptr, val);
264 * Associate a virtual page frame with a given physical page frame
265 * and protection flags for that frame.
267 void set_pte_mfn(unsigned long vaddr, unsigned long mfn, pgprot_t flags)
274 pgd = swapper_pg_dir + pgd_index(vaddr);
275 if (pgd_none(*pgd)) {
279 pud = pud_offset(pgd, vaddr);
280 if (pud_none(*pud)) {
284 pmd = pmd_offset(pud, vaddr);
285 if (pmd_none(*pmd)) {
289 pte = pte_offset_kernel(pmd, vaddr);
290 /* <mfn,flags> stored as-is, to permit clearing entries */
291 xen_set_pte(pte, mfn_pte(mfn, flags));
294 * It's enough to flush this one mapping.
295 * (PGE mappings get flushed as well)
297 __flush_tlb_one(vaddr);
300 void xen_set_pte_at(struct mm_struct *mm, unsigned long addr,
301 pte_t *ptep, pte_t pteval)
303 /* updates to init_mm may be done without lock */
307 if (mm == current->mm || mm == &init_mm) {
308 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU) {
309 struct multicall_space mcs;
310 mcs = xen_mc_entry(0);
312 MULTI_update_va_mapping(mcs.mc, addr, pteval, 0);
313 xen_mc_issue(PARAVIRT_LAZY_MMU);
316 if (HYPERVISOR_update_va_mapping(addr, pteval, 0) == 0)
319 xen_set_pte(ptep, pteval);
326 pte_t xen_ptep_modify_prot_start(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
328 /* Just return the pte as-is. We preserve the bits on commit */
332 void xen_ptep_modify_prot_commit(struct mm_struct *mm, unsigned long addr,
333 pte_t *ptep, pte_t pte)
335 struct multicall_space mcs;
336 struct mmu_update *u;
338 mcs = xen_mc_entry(sizeof(*u));
340 u->ptr = virt_to_machine(ptep).maddr | MMU_PT_UPDATE_PRESERVE_AD;
341 u->val = pte_val_ma(pte);
342 MULTI_mmu_update(mcs.mc, u, 1, NULL, DOMID_SELF);
344 xen_mc_issue(PARAVIRT_LAZY_MMU);
347 /* Assume pteval_t is equivalent to all the other *val_t types. */
348 static pteval_t pte_mfn_to_pfn(pteval_t val)
350 if (val & _PAGE_PRESENT) {
351 unsigned long mfn = (val & PTE_MASK) >> PAGE_SHIFT;
352 pteval_t flags = val & ~PTE_MASK;
353 val = (mfn_to_pfn(mfn) << PAGE_SHIFT) | flags;
359 static pteval_t pte_pfn_to_mfn(pteval_t val)
361 if (val & _PAGE_PRESENT) {
362 unsigned long pfn = (val & PTE_MASK) >> PAGE_SHIFT;
363 pteval_t flags = val & ~PTE_MASK;
364 val = (pfn_to_mfn(pfn) << PAGE_SHIFT) | flags;
370 pteval_t xen_pte_val(pte_t pte)
372 return pte_mfn_to_pfn(pte.pte);
375 pgdval_t xen_pgd_val(pgd_t pgd)
377 return pte_mfn_to_pfn(pgd.pgd);
380 pte_t xen_make_pte(pteval_t pte)
382 pte = pte_pfn_to_mfn(pte);
383 return native_make_pte(pte);
386 pgd_t xen_make_pgd(pgdval_t pgd)
388 pgd = pte_pfn_to_mfn(pgd);
389 return native_make_pgd(pgd);
392 pmdval_t xen_pmd_val(pmd_t pmd)
394 return pte_mfn_to_pfn(pmd.pmd);
397 void xen_set_pud_hyper(pud_t *ptr, pud_t val)
399 struct multicall_space mcs;
400 struct mmu_update *u;
404 mcs = xen_mc_entry(sizeof(*u));
406 u->ptr = virt_to_machine(ptr).maddr;
407 u->val = pud_val_ma(val);
408 MULTI_mmu_update(mcs.mc, u, 1, NULL, DOMID_SELF);
410 xen_mc_issue(PARAVIRT_LAZY_MMU);
415 void xen_set_pud(pud_t *ptr, pud_t val)
417 /* If page is not pinned, we can just update the entry
419 if (!page_pinned(ptr)) {
424 xen_set_pud_hyper(ptr, val);
427 void xen_set_pte(pte_t *ptep, pte_t pte)
429 ptep->pte_high = pte.pte_high;
431 ptep->pte_low = pte.pte_low;
434 void xen_set_pte_atomic(pte_t *ptep, pte_t pte)
436 set_64bit((u64 *)ptep, pte_val_ma(pte));
439 void xen_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
442 smp_wmb(); /* make sure low gets written first */
446 void xen_pmd_clear(pmd_t *pmdp)
448 set_pmd(pmdp, __pmd(0));
451 pmd_t xen_make_pmd(pmdval_t pmd)
453 pmd = pte_pfn_to_mfn(pmd);
454 return native_make_pmd(pmd);
458 (Yet another) pagetable walker. This one is intended for pinning a
459 pagetable. This means that it walks a pagetable and calls the
460 callback function on each page it finds making up the page table,
461 at every level. It walks the entire pagetable, but it only bothers
462 pinning pte pages which are below pte_limit. In the normal case
463 this will be TASK_SIZE, but at boot we need to pin up to
464 FIXADDR_TOP. But the important bit is that we don't pin beyond
465 there, because then we start getting into Xen's ptes.
467 static int pgd_walk(pgd_t *pgd_base, int (*func)(struct page *, enum pt_level),
470 pgd_t *pgd = pgd_base;
472 unsigned long addr = 0;
473 unsigned long pgd_next;
475 BUG_ON(limit > FIXADDR_TOP);
477 if (xen_feature(XENFEAT_auto_translated_physmap))
480 for (; addr != FIXADDR_TOP; pgd++, addr = pgd_next) {
482 unsigned long pud_limit, pud_next;
484 pgd_next = pud_limit = pgd_addr_end(addr, FIXADDR_TOP);
489 pud = pud_offset(pgd, 0);
491 if (PTRS_PER_PUD > 1) /* not folded */
492 flush |= (*func)(virt_to_page(pud), PT_PUD);
494 for (; addr != pud_limit; pud++, addr = pud_next) {
496 unsigned long pmd_limit;
498 pud_next = pud_addr_end(addr, pud_limit);
500 if (pud_next < limit)
501 pmd_limit = pud_next;
508 pmd = pmd_offset(pud, 0);
510 if (PTRS_PER_PMD > 1) /* not folded */
511 flush |= (*func)(virt_to_page(pmd), PT_PMD);
513 for (; addr != pmd_limit; pmd++) {
514 addr += (PAGE_SIZE * PTRS_PER_PTE);
515 if ((pmd_limit-1) < (addr-1)) {
523 flush |= (*func)(pmd_page(*pmd), PT_PTE);
528 flush |= (*func)(virt_to_page(pgd_base), PT_PGD);
533 static spinlock_t *lock_pte(struct page *page)
535 spinlock_t *ptl = NULL;
537 #if NR_CPUS >= CONFIG_SPLIT_PTLOCK_CPUS
538 ptl = __pte_lockptr(page);
545 static void do_unlock(void *v)
551 static void xen_do_pin(unsigned level, unsigned long pfn)
553 struct mmuext_op *op;
554 struct multicall_space mcs;
556 mcs = __xen_mc_entry(sizeof(*op));
559 op->arg1.mfn = pfn_to_mfn(pfn);
560 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
563 static int pin_page(struct page *page, enum pt_level level)
565 unsigned pgfl = TestSetPagePinned(page);
569 flush = 0; /* already pinned */
570 else if (PageHighMem(page))
571 /* kmaps need flushing if we found an unpinned
575 void *pt = lowmem_page_address(page);
576 unsigned long pfn = page_to_pfn(page);
577 struct multicall_space mcs = __xen_mc_entry(0);
584 ptl = lock_pte(page);
586 MULTI_update_va_mapping(mcs.mc, (unsigned long)pt,
587 pfn_pte(pfn, PAGE_KERNEL_RO),
588 level == PT_PGD ? UVMF_TLB_FLUSH : 0);
591 xen_do_pin(MMUEXT_PIN_L1_TABLE, pfn);
594 /* Queue a deferred unlock for when this batch
596 xen_mc_callback(do_unlock, ptl);
603 /* This is called just after a mm has been created, but it has not
604 been used yet. We need to make sure that its pagetable is all
605 read-only, and can be pinned. */
606 void xen_pgd_pin(pgd_t *pgd)
610 if (pgd_walk(pgd, pin_page, TASK_SIZE)) {
611 /* re-enable interrupts for kmap_flush_unused */
617 xen_do_pin(MMUEXT_PIN_L3_TABLE, PFN_DOWN(__pa(pgd)));
622 * On save, we need to pin all pagetables to make sure they get their
623 * mfns turned into pfns. Search the list for any unpinned pgds and pin
624 * them (unpinned pgds are not currently in use, probably because the
625 * process is under construction or destruction).
627 void xen_mm_pin_all(void)
632 spin_lock_irqsave(&pgd_lock, flags);
634 list_for_each_entry(page, &pgd_list, lru) {
635 if (!PagePinned(page)) {
636 xen_pgd_pin((pgd_t *)page_address(page));
637 SetPageSavePinned(page);
641 spin_unlock_irqrestore(&pgd_lock, flags);
644 /* The init_mm pagetable is really pinned as soon as its created, but
645 that's before we have page structures to store the bits. So do all
646 the book-keeping now. */
647 static __init int mark_pinned(struct page *page, enum pt_level level)
653 void __init xen_mark_init_mm_pinned(void)
655 pgd_walk(init_mm.pgd, mark_pinned, FIXADDR_TOP);
658 static int unpin_page(struct page *page, enum pt_level level)
660 unsigned pgfl = TestClearPagePinned(page);
662 if (pgfl && !PageHighMem(page)) {
663 void *pt = lowmem_page_address(page);
664 unsigned long pfn = page_to_pfn(page);
665 spinlock_t *ptl = NULL;
666 struct multicall_space mcs;
668 if (level == PT_PTE) {
669 ptl = lock_pte(page);
671 xen_do_pin(MMUEXT_UNPIN_TABLE, pfn);
674 mcs = __xen_mc_entry(0);
676 MULTI_update_va_mapping(mcs.mc, (unsigned long)pt,
677 pfn_pte(pfn, PAGE_KERNEL),
678 level == PT_PGD ? UVMF_TLB_FLUSH : 0);
681 /* unlock when batch completed */
682 xen_mc_callback(do_unlock, ptl);
686 return 0; /* never need to flush on unpin */
689 /* Release a pagetables pages back as normal RW */
690 static void xen_pgd_unpin(pgd_t *pgd)
694 xen_do_pin(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
696 pgd_walk(pgd, unpin_page, TASK_SIZE);
702 * On resume, undo any pinning done at save, so that the rest of the
703 * kernel doesn't see any unexpected pinned pagetables.
705 void xen_mm_unpin_all(void)
710 spin_lock_irqsave(&pgd_lock, flags);
712 list_for_each_entry(page, &pgd_list, lru) {
713 if (PageSavePinned(page)) {
714 BUG_ON(!PagePinned(page));
715 printk("unpinning pinned %p\n", page_address(page));
716 xen_pgd_unpin((pgd_t *)page_address(page));
717 ClearPageSavePinned(page);
721 spin_unlock_irqrestore(&pgd_lock, flags);
724 void xen_activate_mm(struct mm_struct *prev, struct mm_struct *next)
726 spin_lock(&next->page_table_lock);
727 xen_pgd_pin(next->pgd);
728 spin_unlock(&next->page_table_lock);
731 void xen_dup_mmap(struct mm_struct *oldmm, struct mm_struct *mm)
733 spin_lock(&mm->page_table_lock);
734 xen_pgd_pin(mm->pgd);
735 spin_unlock(&mm->page_table_lock);
740 /* Another cpu may still have their %cr3 pointing at the pagetable, so
741 we need to repoint it somewhere else before we can unpin it. */
742 static void drop_other_mm_ref(void *info)
744 struct mm_struct *mm = info;
746 if (__get_cpu_var(cpu_tlbstate).active_mm == mm)
747 leave_mm(smp_processor_id());
749 /* If this cpu still has a stale cr3 reference, then make sure
750 it has been flushed. */
751 if (x86_read_percpu(xen_current_cr3) == __pa(mm->pgd)) {
752 load_cr3(swapper_pg_dir);
753 arch_flush_lazy_cpu_mode();
757 static void drop_mm_ref(struct mm_struct *mm)
762 if (current->active_mm == mm) {
763 if (current->mm == mm)
764 load_cr3(swapper_pg_dir);
766 leave_mm(smp_processor_id());
767 arch_flush_lazy_cpu_mode();
770 /* Get the "official" set of cpus referring to our pagetable. */
771 mask = mm->cpu_vm_mask;
773 /* It's possible that a vcpu may have a stale reference to our
774 cr3, because its in lazy mode, and it hasn't yet flushed
775 its set of pending hypercalls yet. In this case, we can
776 look at its actual current cr3 value, and force it to flush
778 for_each_online_cpu(cpu) {
779 if (per_cpu(xen_current_cr3, cpu) == __pa(mm->pgd))
783 if (!cpus_empty(mask))
784 xen_smp_call_function_mask(mask, drop_other_mm_ref, mm, 1);
787 static void drop_mm_ref(struct mm_struct *mm)
789 if (current->active_mm == mm)
790 load_cr3(swapper_pg_dir);
795 * While a process runs, Xen pins its pagetables, which means that the
796 * hypervisor forces it to be read-only, and it controls all updates
797 * to it. This means that all pagetable updates have to go via the
798 * hypervisor, which is moderately expensive.
800 * Since we're pulling the pagetable down, we switch to use init_mm,
801 * unpin old process pagetable and mark it all read-write, which
802 * allows further operations on it to be simple memory accesses.
804 * The only subtle point is that another CPU may be still using the
805 * pagetable because of lazy tlb flushing. This means we need need to
806 * switch all CPUs off this pagetable before we can unpin it.
808 void xen_exit_mmap(struct mm_struct *mm)
810 get_cpu(); /* make sure we don't move around */
814 spin_lock(&mm->page_table_lock);
816 /* pgd may not be pinned in the error exit path of execve */
817 if (page_pinned(mm->pgd))
818 xen_pgd_unpin(mm->pgd);
820 spin_unlock(&mm->page_table_lock);