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 pteval_t xen_pte_val(pte_t pte)
328 pteval_t ret = pte.pte;
330 if (ret & _PAGE_PRESENT)
331 ret = machine_to_phys(XMADDR(ret)).paddr | _PAGE_PRESENT;
336 pgdval_t xen_pgd_val(pgd_t pgd)
338 pgdval_t ret = pgd.pgd;
339 if (ret & _PAGE_PRESENT)
340 ret = machine_to_phys(XMADDR(ret)).paddr | _PAGE_PRESENT;
344 pte_t xen_make_pte(pteval_t pte)
346 if (pte & _PAGE_PRESENT) {
347 pte = phys_to_machine(XPADDR(pte)).maddr;
348 pte &= ~(_PAGE_PCD | _PAGE_PWT);
351 return (pte_t){ .pte = pte };
354 pgd_t xen_make_pgd(pgdval_t pgd)
356 if (pgd & _PAGE_PRESENT)
357 pgd = phys_to_machine(XPADDR(pgd)).maddr;
359 return (pgd_t){ pgd };
362 pmdval_t xen_pmd_val(pmd_t pmd)
364 pmdval_t ret = native_pmd_val(pmd);
365 if (ret & _PAGE_PRESENT)
366 ret = machine_to_phys(XMADDR(ret)).paddr | _PAGE_PRESENT;
370 void xen_set_pud_hyper(pud_t *ptr, pud_t val)
372 struct multicall_space mcs;
373 struct mmu_update *u;
377 mcs = xen_mc_entry(sizeof(*u));
379 u->ptr = virt_to_machine(ptr).maddr;
380 u->val = pud_val_ma(val);
381 MULTI_mmu_update(mcs.mc, u, 1, NULL, DOMID_SELF);
383 xen_mc_issue(PARAVIRT_LAZY_MMU);
388 void xen_set_pud(pud_t *ptr, pud_t val)
390 /* If page is not pinned, we can just update the entry
392 if (!page_pinned(ptr)) {
397 xen_set_pud_hyper(ptr, val);
400 void xen_set_pte(pte_t *ptep, pte_t pte)
402 ptep->pte_high = pte.pte_high;
404 ptep->pte_low = pte.pte_low;
407 void xen_set_pte_atomic(pte_t *ptep, pte_t pte)
409 set_64bit((u64 *)ptep, pte_val_ma(pte));
412 void xen_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
415 smp_wmb(); /* make sure low gets written first */
419 void xen_pmd_clear(pmd_t *pmdp)
421 set_pmd(pmdp, __pmd(0));
424 pmd_t xen_make_pmd(pmdval_t pmd)
426 if (pmd & _PAGE_PRESENT)
427 pmd = phys_to_machine(XPADDR(pmd)).maddr;
429 return native_make_pmd(pmd);
433 (Yet another) pagetable walker. This one is intended for pinning a
434 pagetable. This means that it walks a pagetable and calls the
435 callback function on each page it finds making up the page table,
436 at every level. It walks the entire pagetable, but it only bothers
437 pinning pte pages which are below pte_limit. In the normal case
438 this will be TASK_SIZE, but at boot we need to pin up to
439 FIXADDR_TOP. But the important bit is that we don't pin beyond
440 there, because then we start getting into Xen's ptes.
442 static int pgd_walk(pgd_t *pgd_base, int (*func)(struct page *, enum pt_level),
445 pgd_t *pgd = pgd_base;
447 unsigned long addr = 0;
448 unsigned long pgd_next;
450 BUG_ON(limit > FIXADDR_TOP);
452 if (xen_feature(XENFEAT_auto_translated_physmap))
455 for (; addr != FIXADDR_TOP; pgd++, addr = pgd_next) {
457 unsigned long pud_limit, pud_next;
459 pgd_next = pud_limit = pgd_addr_end(addr, FIXADDR_TOP);
464 pud = pud_offset(pgd, 0);
466 if (PTRS_PER_PUD > 1) /* not folded */
467 flush |= (*func)(virt_to_page(pud), PT_PUD);
469 for (; addr != pud_limit; pud++, addr = pud_next) {
471 unsigned long pmd_limit;
473 pud_next = pud_addr_end(addr, pud_limit);
475 if (pud_next < limit)
476 pmd_limit = pud_next;
483 pmd = pmd_offset(pud, 0);
485 if (PTRS_PER_PMD > 1) /* not folded */
486 flush |= (*func)(virt_to_page(pmd), PT_PMD);
488 for (; addr != pmd_limit; pmd++) {
489 addr += (PAGE_SIZE * PTRS_PER_PTE);
490 if ((pmd_limit-1) < (addr-1)) {
498 flush |= (*func)(pmd_page(*pmd), PT_PTE);
503 flush |= (*func)(virt_to_page(pgd_base), PT_PGD);
508 static spinlock_t *lock_pte(struct page *page)
510 spinlock_t *ptl = NULL;
512 #if NR_CPUS >= CONFIG_SPLIT_PTLOCK_CPUS
513 ptl = __pte_lockptr(page);
520 static void do_unlock(void *v)
526 static void xen_do_pin(unsigned level, unsigned long pfn)
528 struct mmuext_op *op;
529 struct multicall_space mcs;
531 mcs = __xen_mc_entry(sizeof(*op));
534 op->arg1.mfn = pfn_to_mfn(pfn);
535 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
538 static int pin_page(struct page *page, enum pt_level level)
540 unsigned pgfl = TestSetPagePinned(page);
544 flush = 0; /* already pinned */
545 else if (PageHighMem(page))
546 /* kmaps need flushing if we found an unpinned
550 void *pt = lowmem_page_address(page);
551 unsigned long pfn = page_to_pfn(page);
552 struct multicall_space mcs = __xen_mc_entry(0);
559 ptl = lock_pte(page);
561 MULTI_update_va_mapping(mcs.mc, (unsigned long)pt,
562 pfn_pte(pfn, PAGE_KERNEL_RO),
563 level == PT_PGD ? UVMF_TLB_FLUSH : 0);
566 xen_do_pin(MMUEXT_PIN_L1_TABLE, pfn);
569 /* Queue a deferred unlock for when this batch
571 xen_mc_callback(do_unlock, ptl);
578 /* This is called just after a mm has been created, but it has not
579 been used yet. We need to make sure that its pagetable is all
580 read-only, and can be pinned. */
581 void xen_pgd_pin(pgd_t *pgd)
585 if (pgd_walk(pgd, pin_page, TASK_SIZE)) {
586 /* re-enable interrupts for kmap_flush_unused */
592 xen_do_pin(MMUEXT_PIN_L3_TABLE, PFN_DOWN(__pa(pgd)));
597 * On save, we need to pin all pagetables to make sure they get their
598 * mfns turned into pfns. Search the list for any unpinned pgds and pin
599 * them (unpinned pgds are not currently in use, probably because the
600 * process is under construction or destruction).
602 void xen_mm_pin_all(void)
607 spin_lock_irqsave(&pgd_lock, flags);
609 list_for_each_entry(page, &pgd_list, lru) {
610 if (!PagePinned(page)) {
611 xen_pgd_pin((pgd_t *)page_address(page));
612 SetPageSavePinned(page);
616 spin_unlock_irqrestore(&pgd_lock, flags);
619 /* The init_mm pagetable is really pinned as soon as its created, but
620 that's before we have page structures to store the bits. So do all
621 the book-keeping now. */
622 static __init int mark_pinned(struct page *page, enum pt_level level)
628 void __init xen_mark_init_mm_pinned(void)
630 pgd_walk(init_mm.pgd, mark_pinned, FIXADDR_TOP);
633 static int unpin_page(struct page *page, enum pt_level level)
635 unsigned pgfl = TestClearPagePinned(page);
637 if (pgfl && !PageHighMem(page)) {
638 void *pt = lowmem_page_address(page);
639 unsigned long pfn = page_to_pfn(page);
640 spinlock_t *ptl = NULL;
641 struct multicall_space mcs;
643 if (level == PT_PTE) {
644 ptl = lock_pte(page);
646 xen_do_pin(MMUEXT_UNPIN_TABLE, pfn);
649 mcs = __xen_mc_entry(0);
651 MULTI_update_va_mapping(mcs.mc, (unsigned long)pt,
652 pfn_pte(pfn, PAGE_KERNEL),
653 level == PT_PGD ? UVMF_TLB_FLUSH : 0);
656 /* unlock when batch completed */
657 xen_mc_callback(do_unlock, ptl);
661 return 0; /* never need to flush on unpin */
664 /* Release a pagetables pages back as normal RW */
665 static void xen_pgd_unpin(pgd_t *pgd)
669 xen_do_pin(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
671 pgd_walk(pgd, unpin_page, TASK_SIZE);
677 * On resume, undo any pinning done at save, so that the rest of the
678 * kernel doesn't see any unexpected pinned pagetables.
680 void xen_mm_unpin_all(void)
685 spin_lock_irqsave(&pgd_lock, flags);
687 list_for_each_entry(page, &pgd_list, lru) {
688 if (PageSavePinned(page)) {
689 BUG_ON(!PagePinned(page));
690 printk("unpinning pinned %p\n", page_address(page));
691 xen_pgd_unpin((pgd_t *)page_address(page));
692 ClearPageSavePinned(page);
696 spin_unlock_irqrestore(&pgd_lock, flags);
699 void xen_activate_mm(struct mm_struct *prev, struct mm_struct *next)
701 spin_lock(&next->page_table_lock);
702 xen_pgd_pin(next->pgd);
703 spin_unlock(&next->page_table_lock);
706 void xen_dup_mmap(struct mm_struct *oldmm, struct mm_struct *mm)
708 spin_lock(&mm->page_table_lock);
709 xen_pgd_pin(mm->pgd);
710 spin_unlock(&mm->page_table_lock);
715 /* Another cpu may still have their %cr3 pointing at the pagetable, so
716 we need to repoint it somewhere else before we can unpin it. */
717 static void drop_other_mm_ref(void *info)
719 struct mm_struct *mm = info;
721 if (__get_cpu_var(cpu_tlbstate).active_mm == mm)
722 leave_mm(smp_processor_id());
724 /* If this cpu still has a stale cr3 reference, then make sure
725 it has been flushed. */
726 if (x86_read_percpu(xen_current_cr3) == __pa(mm->pgd)) {
727 load_cr3(swapper_pg_dir);
728 arch_flush_lazy_cpu_mode();
732 static void drop_mm_ref(struct mm_struct *mm)
737 if (current->active_mm == mm) {
738 if (current->mm == mm)
739 load_cr3(swapper_pg_dir);
741 leave_mm(smp_processor_id());
742 arch_flush_lazy_cpu_mode();
745 /* Get the "official" set of cpus referring to our pagetable. */
746 mask = mm->cpu_vm_mask;
748 /* It's possible that a vcpu may have a stale reference to our
749 cr3, because its in lazy mode, and it hasn't yet flushed
750 its set of pending hypercalls yet. In this case, we can
751 look at its actual current cr3 value, and force it to flush
753 for_each_online_cpu(cpu) {
754 if (per_cpu(xen_current_cr3, cpu) == __pa(mm->pgd))
758 if (!cpus_empty(mask))
759 xen_smp_call_function_mask(mask, drop_other_mm_ref, mm, 1);
762 static void drop_mm_ref(struct mm_struct *mm)
764 if (current->active_mm == mm)
765 load_cr3(swapper_pg_dir);
770 * While a process runs, Xen pins its pagetables, which means that the
771 * hypervisor forces it to be read-only, and it controls all updates
772 * to it. This means that all pagetable updates have to go via the
773 * hypervisor, which is moderately expensive.
775 * Since we're pulling the pagetable down, we switch to use init_mm,
776 * unpin old process pagetable and mark it all read-write, which
777 * allows further operations on it to be simple memory accesses.
779 * The only subtle point is that another CPU may be still using the
780 * pagetable because of lazy tlb flushing. This means we need need to
781 * switch all CPUs off this pagetable before we can unpin it.
783 void xen_exit_mmap(struct mm_struct *mm)
785 get_cpu(); /* make sure we don't move around */
789 spin_lock(&mm->page_table_lock);
791 /* pgd may not be pinned in the error exit path of execve */
792 if (page_pinned(mm->pgd))
793 xen_pgd_unpin(mm->pgd);
795 spin_unlock(&mm->page_table_lock);