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/debugfs.h>
44 #include <linux/bug.h>
45 #include <linux/vmalloc.h>
46 #include <linux/module.h>
47 #include <linux/gfp.h>
48 #include <linux/memblock.h>
50 #include <asm/pgtable.h>
51 #include <asm/tlbflush.h>
52 #include <asm/fixmap.h>
53 #include <asm/mmu_context.h>
54 #include <asm/setup.h>
55 #include <asm/paravirt.h>
57 #include <asm/linkage.h>
60 #include <asm/xen/hypercall.h>
61 #include <asm/xen/hypervisor.h>
65 #include <xen/interface/xen.h>
66 #include <xen/interface/hvm/hvm_op.h>
67 #include <xen/interface/version.h>
68 #include <xen/interface/memory.h>
69 #include <xen/hvc-console.h>
71 #include "multicalls.h"
75 #define MMU_UPDATE_HISTO 30
78 * Protects atomic reservation decrease/increase against concurrent increases.
79 * Also protects non-atomic updates of current_pages and driver_pages, and
82 DEFINE_SPINLOCK(xen_reservation_lock);
84 #ifdef CONFIG_XEN_DEBUG_FS
88 u32 pgd_update_pinned;
89 u32 pgd_update_batched;
92 u32 pud_update_pinned;
93 u32 pud_update_batched;
96 u32 pmd_update_pinned;
97 u32 pmd_update_batched;
100 u32 pte_update_pinned;
101 u32 pte_update_batched;
104 u32 mmu_update_extended;
105 u32 mmu_update_histo[MMU_UPDATE_HISTO];
108 u32 prot_commit_batched;
111 u32 set_pte_at_batched;
112 u32 set_pte_at_pinned;
113 u32 set_pte_at_current;
114 u32 set_pte_at_kernel;
117 static u8 zero_stats;
119 static inline void check_zero(void)
121 if (unlikely(zero_stats)) {
122 memset(&mmu_stats, 0, sizeof(mmu_stats));
127 #define ADD_STATS(elem, val) \
128 do { check_zero(); mmu_stats.elem += (val); } while(0)
130 #else /* !CONFIG_XEN_DEBUG_FS */
132 #define ADD_STATS(elem, val) do { (void)(val); } while(0)
134 #endif /* CONFIG_XEN_DEBUG_FS */
138 * Identity map, in addition to plain kernel map. This needs to be
139 * large enough to allocate page table pages to allocate the rest.
140 * Each page can map 2MB.
142 static pte_t level1_ident_pgt[PTRS_PER_PTE * 4] __page_aligned_bss;
145 /* l3 pud for userspace vsyscall mapping */
146 static pud_t level3_user_vsyscall[PTRS_PER_PUD] __page_aligned_bss;
147 #endif /* CONFIG_X86_64 */
150 * Note about cr3 (pagetable base) values:
152 * xen_cr3 contains the current logical cr3 value; it contains the
153 * last set cr3. This may not be the current effective cr3, because
154 * its update may be being lazily deferred. However, a vcpu looking
155 * at its own cr3 can use this value knowing that it everything will
156 * be self-consistent.
158 * xen_current_cr3 contains the actual vcpu cr3; it is set once the
159 * hypercall to set the vcpu cr3 is complete (so it may be a little
160 * out of date, but it will never be set early). If one vcpu is
161 * looking at another vcpu's cr3 value, it should use this variable.
163 DEFINE_PER_CPU(unsigned long, xen_cr3); /* cr3 stored as physaddr */
164 DEFINE_PER_CPU(unsigned long, xen_current_cr3); /* actual vcpu cr3 */
168 * Just beyond the highest usermode address. STACK_TOP_MAX has a
169 * redzone above it, so round it up to a PGD boundary.
171 #define USER_LIMIT ((STACK_TOP_MAX + PGDIR_SIZE - 1) & PGDIR_MASK)
174 #define P2M_ENTRIES_PER_PAGE (PAGE_SIZE / sizeof(unsigned long))
175 #define TOP_ENTRIES (MAX_DOMAIN_PAGES / P2M_ENTRIES_PER_PAGE)
177 /* Placeholder for holes in the address space */
178 static unsigned long p2m_missing[P2M_ENTRIES_PER_PAGE] __page_aligned_data =
179 { [ 0 ... P2M_ENTRIES_PER_PAGE-1 ] = ~0UL };
181 /* Array of pointers to pages containing p2m entries */
182 static unsigned long *p2m_top[TOP_ENTRIES] __page_aligned_data =
183 { [ 0 ... TOP_ENTRIES - 1] = &p2m_missing[0] };
185 /* Arrays of p2m arrays expressed in mfns used for save/restore */
186 static unsigned long p2m_top_mfn[TOP_ENTRIES] __page_aligned_bss;
188 static unsigned long p2m_top_mfn_list[TOP_ENTRIES / P2M_ENTRIES_PER_PAGE]
191 static inline unsigned p2m_top_index(unsigned long pfn)
193 BUG_ON(pfn >= MAX_DOMAIN_PAGES);
194 return pfn / P2M_ENTRIES_PER_PAGE;
197 static inline unsigned p2m_index(unsigned long pfn)
199 return pfn % P2M_ENTRIES_PER_PAGE;
202 /* Build the parallel p2m_top_mfn structures */
203 void xen_build_mfn_list_list(void)
207 for (pfn = 0; pfn < MAX_DOMAIN_PAGES; pfn += P2M_ENTRIES_PER_PAGE) {
208 unsigned topidx = p2m_top_index(pfn);
210 p2m_top_mfn[topidx] = virt_to_mfn(p2m_top[topidx]);
213 for (idx = 0; idx < ARRAY_SIZE(p2m_top_mfn_list); idx++) {
214 unsigned topidx = idx * P2M_ENTRIES_PER_PAGE;
215 p2m_top_mfn_list[idx] = virt_to_mfn(&p2m_top_mfn[topidx]);
219 void xen_setup_mfn_list_list(void)
221 BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
223 HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list =
224 virt_to_mfn(p2m_top_mfn_list);
225 HYPERVISOR_shared_info->arch.max_pfn = xen_start_info->nr_pages;
228 /* Set up p2m_top to point to the domain-builder provided p2m pages */
229 void __init xen_build_dynamic_phys_to_machine(void)
231 unsigned long *mfn_list = (unsigned long *)xen_start_info->mfn_list;
232 unsigned long max_pfn = min(MAX_DOMAIN_PAGES, xen_start_info->nr_pages);
235 for (pfn = 0; pfn < max_pfn; pfn += P2M_ENTRIES_PER_PAGE) {
236 unsigned topidx = p2m_top_index(pfn);
238 p2m_top[topidx] = &mfn_list[pfn];
241 xen_build_mfn_list_list();
244 unsigned long get_phys_to_machine(unsigned long pfn)
246 unsigned topidx, idx;
248 if (unlikely(pfn >= MAX_DOMAIN_PAGES))
249 return INVALID_P2M_ENTRY;
251 topidx = p2m_top_index(pfn);
252 idx = p2m_index(pfn);
253 return p2m_top[topidx][idx];
255 EXPORT_SYMBOL_GPL(get_phys_to_machine);
257 /* install a new p2m_top page */
258 bool install_p2mtop_page(unsigned long pfn, unsigned long *p)
260 unsigned topidx = p2m_top_index(pfn);
261 unsigned long **pfnp, *mfnp;
264 pfnp = &p2m_top[topidx];
265 mfnp = &p2m_top_mfn[topidx];
267 for (i = 0; i < P2M_ENTRIES_PER_PAGE; i++)
268 p[i] = INVALID_P2M_ENTRY;
270 if (cmpxchg(pfnp, p2m_missing, p) == p2m_missing) {
271 *mfnp = virt_to_mfn(p);
278 static void alloc_p2m(unsigned long pfn)
282 p = (void *)__get_free_page(GFP_KERNEL | __GFP_NOFAIL);
285 if (!install_p2mtop_page(pfn, p))
286 free_page((unsigned long)p);
289 /* Try to install p2m mapping; fail if intermediate bits missing */
290 bool __set_phys_to_machine(unsigned long pfn, unsigned long mfn)
292 unsigned topidx, idx;
294 if (unlikely(pfn >= MAX_DOMAIN_PAGES)) {
295 BUG_ON(mfn != INVALID_P2M_ENTRY);
299 topidx = p2m_top_index(pfn);
300 if (p2m_top[topidx] == p2m_missing) {
301 if (mfn == INVALID_P2M_ENTRY)
306 idx = p2m_index(pfn);
307 p2m_top[topidx][idx] = mfn;
312 void set_phys_to_machine(unsigned long pfn, unsigned long mfn)
314 if (unlikely(xen_feature(XENFEAT_auto_translated_physmap))) {
315 BUG_ON(pfn != mfn && mfn != INVALID_P2M_ENTRY);
319 if (unlikely(!__set_phys_to_machine(pfn, mfn))) {
322 if (!__set_phys_to_machine(pfn, mfn))
327 unsigned long arbitrary_virt_to_mfn(void *vaddr)
329 xmaddr_t maddr = arbitrary_virt_to_machine(vaddr);
331 return PFN_DOWN(maddr.maddr);
334 xmaddr_t arbitrary_virt_to_machine(void *vaddr)
336 unsigned long address = (unsigned long)vaddr;
342 * if the PFN is in the linear mapped vaddr range, we can just use
343 * the (quick) virt_to_machine() p2m lookup
345 if (virt_addr_valid(vaddr))
346 return virt_to_machine(vaddr);
348 /* otherwise we have to do a (slower) full page-table walk */
350 pte = lookup_address(address, &level);
352 offset = address & ~PAGE_MASK;
353 return XMADDR(((phys_addr_t)pte_mfn(*pte) << PAGE_SHIFT) + offset);
356 void make_lowmem_page_readonly(void *vaddr)
359 unsigned long address = (unsigned long)vaddr;
362 pte = lookup_address(address, &level);
365 ptev = pte_wrprotect(*pte);
367 if (HYPERVISOR_update_va_mapping(address, ptev, 0))
371 void make_lowmem_page_readwrite(void *vaddr)
374 unsigned long address = (unsigned long)vaddr;
377 pte = lookup_address(address, &level);
380 ptev = pte_mkwrite(*pte);
382 if (HYPERVISOR_update_va_mapping(address, ptev, 0))
387 static bool xen_page_pinned(void *ptr)
389 struct page *page = virt_to_page(ptr);
391 return PagePinned(page);
394 static bool xen_iomap_pte(pte_t pte)
396 return pte_flags(pte) & _PAGE_IOMAP;
399 static void xen_set_iomap_pte(pte_t *ptep, pte_t pteval)
401 struct multicall_space mcs;
402 struct mmu_update *u;
404 mcs = xen_mc_entry(sizeof(*u));
407 /* ptep might be kmapped when using 32-bit HIGHPTE */
408 u->ptr = arbitrary_virt_to_machine(ptep).maddr;
409 u->val = pte_val_ma(pteval);
411 MULTI_mmu_update(mcs.mc, mcs.args, 1, NULL, DOMID_IO);
413 xen_mc_issue(PARAVIRT_LAZY_MMU);
416 static void xen_extend_mmu_update(const struct mmu_update *update)
418 struct multicall_space mcs;
419 struct mmu_update *u;
421 mcs = xen_mc_extend_args(__HYPERVISOR_mmu_update, sizeof(*u));
423 if (mcs.mc != NULL) {
424 ADD_STATS(mmu_update_extended, 1);
425 ADD_STATS(mmu_update_histo[mcs.mc->args[1]], -1);
429 if (mcs.mc->args[1] < MMU_UPDATE_HISTO)
430 ADD_STATS(mmu_update_histo[mcs.mc->args[1]], 1);
432 ADD_STATS(mmu_update_histo[0], 1);
434 ADD_STATS(mmu_update, 1);
435 mcs = __xen_mc_entry(sizeof(*u));
436 MULTI_mmu_update(mcs.mc, mcs.args, 1, NULL, DOMID_SELF);
437 ADD_STATS(mmu_update_histo[1], 1);
444 void xen_set_pmd_hyper(pmd_t *ptr, pmd_t val)
452 /* ptr may be ioremapped for 64-bit pagetable setup */
453 u.ptr = arbitrary_virt_to_machine(ptr).maddr;
454 u.val = pmd_val_ma(val);
455 xen_extend_mmu_update(&u);
457 ADD_STATS(pmd_update_batched, paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU);
459 xen_mc_issue(PARAVIRT_LAZY_MMU);
464 void xen_set_pmd(pmd_t *ptr, pmd_t val)
466 ADD_STATS(pmd_update, 1);
468 /* If page is not pinned, we can just update the entry
470 if (!xen_page_pinned(ptr)) {
475 ADD_STATS(pmd_update_pinned, 1);
477 xen_set_pmd_hyper(ptr, val);
481 * Associate a virtual page frame with a given physical page frame
482 * and protection flags for that frame.
484 void set_pte_mfn(unsigned long vaddr, unsigned long mfn, pgprot_t flags)
486 set_pte_vaddr(vaddr, mfn_pte(mfn, flags));
489 void xen_set_pte_at(struct mm_struct *mm, unsigned long addr,
490 pte_t *ptep, pte_t pteval)
492 if (xen_iomap_pte(pteval)) {
493 xen_set_iomap_pte(ptep, pteval);
497 ADD_STATS(set_pte_at, 1);
498 // ADD_STATS(set_pte_at_pinned, xen_page_pinned(ptep));
499 ADD_STATS(set_pte_at_current, mm == current->mm);
500 ADD_STATS(set_pte_at_kernel, mm == &init_mm);
502 if (mm == current->mm || mm == &init_mm) {
503 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU) {
504 struct multicall_space mcs;
505 mcs = xen_mc_entry(0);
507 MULTI_update_va_mapping(mcs.mc, addr, pteval, 0);
508 ADD_STATS(set_pte_at_batched, 1);
509 xen_mc_issue(PARAVIRT_LAZY_MMU);
512 if (HYPERVISOR_update_va_mapping(addr, pteval, 0) == 0)
515 xen_set_pte(ptep, pteval);
520 pte_t xen_ptep_modify_prot_start(struct mm_struct *mm,
521 unsigned long addr, pte_t *ptep)
523 /* Just return the pte as-is. We preserve the bits on commit */
527 void xen_ptep_modify_prot_commit(struct mm_struct *mm, unsigned long addr,
528 pte_t *ptep, pte_t pte)
534 u.ptr = arbitrary_virt_to_machine(ptep).maddr | MMU_PT_UPDATE_PRESERVE_AD;
535 u.val = pte_val_ma(pte);
536 xen_extend_mmu_update(&u);
538 ADD_STATS(prot_commit, 1);
539 ADD_STATS(prot_commit_batched, paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU);
541 xen_mc_issue(PARAVIRT_LAZY_MMU);
544 /* Assume pteval_t is equivalent to all the other *val_t types. */
545 static pteval_t pte_mfn_to_pfn(pteval_t val)
547 if (val & _PAGE_PRESENT) {
548 unsigned long mfn = (val & PTE_PFN_MASK) >> PAGE_SHIFT;
549 pteval_t flags = val & PTE_FLAGS_MASK;
550 val = ((pteval_t)mfn_to_pfn(mfn) << PAGE_SHIFT) | flags;
556 static pteval_t pte_pfn_to_mfn(pteval_t val)
558 if (val & _PAGE_PRESENT) {
559 unsigned long pfn = (val & PTE_PFN_MASK) >> PAGE_SHIFT;
560 pteval_t flags = val & PTE_FLAGS_MASK;
561 val = ((pteval_t)pfn_to_mfn(pfn) << PAGE_SHIFT) | flags;
567 static pteval_t iomap_pte(pteval_t val)
569 if (val & _PAGE_PRESENT) {
570 unsigned long pfn = (val & PTE_PFN_MASK) >> PAGE_SHIFT;
571 pteval_t flags = val & PTE_FLAGS_MASK;
573 /* We assume the pte frame number is a MFN, so
574 just use it as-is. */
575 val = ((pteval_t)pfn << PAGE_SHIFT) | flags;
581 pteval_t xen_pte_val(pte_t pte)
583 if (xen_initial_domain() && (pte.pte & _PAGE_IOMAP))
586 return pte_mfn_to_pfn(pte.pte);
588 PV_CALLEE_SAVE_REGS_THUNK(xen_pte_val);
590 pgdval_t xen_pgd_val(pgd_t pgd)
592 return pte_mfn_to_pfn(pgd.pgd);
594 PV_CALLEE_SAVE_REGS_THUNK(xen_pgd_val);
596 pte_t xen_make_pte(pteval_t pte)
598 phys_addr_t addr = (pte & PTE_PFN_MASK);
601 * Unprivileged domains are allowed to do IOMAPpings for
602 * PCI passthrough, but not map ISA space. The ISA
603 * mappings are just dummy local mappings to keep other
604 * parts of the kernel happy.
606 if (unlikely(pte & _PAGE_IOMAP) &&
607 (xen_initial_domain() || addr >= ISA_END_ADDRESS)) {
608 pte = iomap_pte(pte);
611 pte = pte_pfn_to_mfn(pte);
614 return native_make_pte(pte);
616 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pte);
618 pgd_t xen_make_pgd(pgdval_t pgd)
620 pgd = pte_pfn_to_mfn(pgd);
621 return native_make_pgd(pgd);
623 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pgd);
625 pmdval_t xen_pmd_val(pmd_t pmd)
627 return pte_mfn_to_pfn(pmd.pmd);
629 PV_CALLEE_SAVE_REGS_THUNK(xen_pmd_val);
631 void xen_set_pud_hyper(pud_t *ptr, pud_t val)
639 /* ptr may be ioremapped for 64-bit pagetable setup */
640 u.ptr = arbitrary_virt_to_machine(ptr).maddr;
641 u.val = pud_val_ma(val);
642 xen_extend_mmu_update(&u);
644 ADD_STATS(pud_update_batched, paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU);
646 xen_mc_issue(PARAVIRT_LAZY_MMU);
651 void xen_set_pud(pud_t *ptr, pud_t val)
653 ADD_STATS(pud_update, 1);
655 /* If page is not pinned, we can just update the entry
657 if (!xen_page_pinned(ptr)) {
662 ADD_STATS(pud_update_pinned, 1);
664 xen_set_pud_hyper(ptr, val);
667 void xen_set_pte(pte_t *ptep, pte_t pte)
669 if (xen_iomap_pte(pte)) {
670 xen_set_iomap_pte(ptep, pte);
674 ADD_STATS(pte_update, 1);
675 // ADD_STATS(pte_update_pinned, xen_page_pinned(ptep));
676 ADD_STATS(pte_update_batched, paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU);
678 #ifdef CONFIG_X86_PAE
679 ptep->pte_high = pte.pte_high;
681 ptep->pte_low = pte.pte_low;
687 #ifdef CONFIG_X86_PAE
688 void xen_set_pte_atomic(pte_t *ptep, pte_t pte)
690 if (xen_iomap_pte(pte)) {
691 xen_set_iomap_pte(ptep, pte);
695 set_64bit((u64 *)ptep, native_pte_val(pte));
698 void xen_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
701 smp_wmb(); /* make sure low gets written first */
705 void xen_pmd_clear(pmd_t *pmdp)
707 set_pmd(pmdp, __pmd(0));
709 #endif /* CONFIG_X86_PAE */
711 pmd_t xen_make_pmd(pmdval_t pmd)
713 pmd = pte_pfn_to_mfn(pmd);
714 return native_make_pmd(pmd);
716 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pmd);
718 #if PAGETABLE_LEVELS == 4
719 pudval_t xen_pud_val(pud_t pud)
721 return pte_mfn_to_pfn(pud.pud);
723 PV_CALLEE_SAVE_REGS_THUNK(xen_pud_val);
725 pud_t xen_make_pud(pudval_t pud)
727 pud = pte_pfn_to_mfn(pud);
729 return native_make_pud(pud);
731 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pud);
733 pgd_t *xen_get_user_pgd(pgd_t *pgd)
735 pgd_t *pgd_page = (pgd_t *)(((unsigned long)pgd) & PAGE_MASK);
736 unsigned offset = pgd - pgd_page;
737 pgd_t *user_ptr = NULL;
739 if (offset < pgd_index(USER_LIMIT)) {
740 struct page *page = virt_to_page(pgd_page);
741 user_ptr = (pgd_t *)page->private;
749 static void __xen_set_pgd_hyper(pgd_t *ptr, pgd_t val)
753 u.ptr = virt_to_machine(ptr).maddr;
754 u.val = pgd_val_ma(val);
755 xen_extend_mmu_update(&u);
759 * Raw hypercall-based set_pgd, intended for in early boot before
760 * there's a page structure. This implies:
761 * 1. The only existing pagetable is the kernel's
762 * 2. It is always pinned
763 * 3. It has no user pagetable attached to it
765 void __init xen_set_pgd_hyper(pgd_t *ptr, pgd_t val)
771 __xen_set_pgd_hyper(ptr, val);
773 xen_mc_issue(PARAVIRT_LAZY_MMU);
778 void xen_set_pgd(pgd_t *ptr, pgd_t val)
780 pgd_t *user_ptr = xen_get_user_pgd(ptr);
782 ADD_STATS(pgd_update, 1);
784 /* If page is not pinned, we can just update the entry
786 if (!xen_page_pinned(ptr)) {
789 WARN_ON(xen_page_pinned(user_ptr));
795 ADD_STATS(pgd_update_pinned, 1);
796 ADD_STATS(pgd_update_batched, paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU);
798 /* If it's pinned, then we can at least batch the kernel and
799 user updates together. */
802 __xen_set_pgd_hyper(ptr, val);
804 __xen_set_pgd_hyper(user_ptr, val);
806 xen_mc_issue(PARAVIRT_LAZY_MMU);
808 #endif /* PAGETABLE_LEVELS == 4 */
811 * (Yet another) pagetable walker. This one is intended for pinning a
812 * pagetable. This means that it walks a pagetable and calls the
813 * callback function on each page it finds making up the page table,
814 * at every level. It walks the entire pagetable, but it only bothers
815 * pinning pte pages which are below limit. In the normal case this
816 * will be STACK_TOP_MAX, but at boot we need to pin up to
819 * For 32-bit the important bit is that we don't pin beyond there,
820 * because then we start getting into Xen's ptes.
822 * For 64-bit, we must skip the Xen hole in the middle of the address
823 * space, just after the big x86-64 virtual hole.
825 static int __xen_pgd_walk(struct mm_struct *mm, pgd_t *pgd,
826 int (*func)(struct mm_struct *mm, struct page *,
831 unsigned hole_low, hole_high;
832 unsigned pgdidx_limit, pudidx_limit, pmdidx_limit;
833 unsigned pgdidx, pudidx, pmdidx;
835 /* The limit is the last byte to be touched */
837 BUG_ON(limit >= FIXADDR_TOP);
839 if (xen_feature(XENFEAT_auto_translated_physmap))
843 * 64-bit has a great big hole in the middle of the address
844 * space, which contains the Xen mappings. On 32-bit these
845 * will end up making a zero-sized hole and so is a no-op.
847 hole_low = pgd_index(USER_LIMIT);
848 hole_high = pgd_index(PAGE_OFFSET);
850 pgdidx_limit = pgd_index(limit);
852 pudidx_limit = pud_index(limit);
857 pmdidx_limit = pmd_index(limit);
862 for (pgdidx = 0; pgdidx <= pgdidx_limit; pgdidx++) {
865 if (pgdidx >= hole_low && pgdidx < hole_high)
868 if (!pgd_val(pgd[pgdidx]))
871 pud = pud_offset(&pgd[pgdidx], 0);
873 if (PTRS_PER_PUD > 1) /* not folded */
874 flush |= (*func)(mm, virt_to_page(pud), PT_PUD);
876 for (pudidx = 0; pudidx < PTRS_PER_PUD; pudidx++) {
879 if (pgdidx == pgdidx_limit &&
880 pudidx > pudidx_limit)
883 if (pud_none(pud[pudidx]))
886 pmd = pmd_offset(&pud[pudidx], 0);
888 if (PTRS_PER_PMD > 1) /* not folded */
889 flush |= (*func)(mm, virt_to_page(pmd), PT_PMD);
891 for (pmdidx = 0; pmdidx < PTRS_PER_PMD; pmdidx++) {
894 if (pgdidx == pgdidx_limit &&
895 pudidx == pudidx_limit &&
896 pmdidx > pmdidx_limit)
899 if (pmd_none(pmd[pmdidx]))
902 pte = pmd_page(pmd[pmdidx]);
903 flush |= (*func)(mm, pte, PT_PTE);
909 /* Do the top level last, so that the callbacks can use it as
910 a cue to do final things like tlb flushes. */
911 flush |= (*func)(mm, virt_to_page(pgd), PT_PGD);
916 static int xen_pgd_walk(struct mm_struct *mm,
917 int (*func)(struct mm_struct *mm, struct page *,
921 return __xen_pgd_walk(mm, mm->pgd, func, limit);
924 /* If we're using split pte locks, then take the page's lock and
925 return a pointer to it. Otherwise return NULL. */
926 static spinlock_t *xen_pte_lock(struct page *page, struct mm_struct *mm)
928 spinlock_t *ptl = NULL;
930 #if USE_SPLIT_PTLOCKS
931 ptl = __pte_lockptr(page);
932 spin_lock_nest_lock(ptl, &mm->page_table_lock);
938 static void xen_pte_unlock(void *v)
944 static void xen_do_pin(unsigned level, unsigned long pfn)
946 struct mmuext_op *op;
947 struct multicall_space mcs;
949 mcs = __xen_mc_entry(sizeof(*op));
952 op->arg1.mfn = pfn_to_mfn(pfn);
953 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
956 static int xen_pin_page(struct mm_struct *mm, struct page *page,
959 unsigned pgfl = TestSetPagePinned(page);
963 flush = 0; /* already pinned */
964 else if (PageHighMem(page))
965 /* kmaps need flushing if we found an unpinned
969 void *pt = lowmem_page_address(page);
970 unsigned long pfn = page_to_pfn(page);
971 struct multicall_space mcs = __xen_mc_entry(0);
977 * We need to hold the pagetable lock between the time
978 * we make the pagetable RO and when we actually pin
979 * it. If we don't, then other users may come in and
980 * attempt to update the pagetable by writing it,
981 * which will fail because the memory is RO but not
982 * pinned, so Xen won't do the trap'n'emulate.
984 * If we're using split pte locks, we can't hold the
985 * entire pagetable's worth of locks during the
986 * traverse, because we may wrap the preempt count (8
987 * bits). The solution is to mark RO and pin each PTE
988 * page while holding the lock. This means the number
989 * of locks we end up holding is never more than a
990 * batch size (~32 entries, at present).
992 * If we're not using split pte locks, we needn't pin
993 * the PTE pages independently, because we're
994 * protected by the overall pagetable lock.
998 ptl = xen_pte_lock(page, mm);
1000 MULTI_update_va_mapping(mcs.mc, (unsigned long)pt,
1001 pfn_pte(pfn, PAGE_KERNEL_RO),
1002 level == PT_PGD ? UVMF_TLB_FLUSH : 0);
1005 xen_do_pin(MMUEXT_PIN_L1_TABLE, pfn);
1007 /* Queue a deferred unlock for when this batch
1009 xen_mc_callback(xen_pte_unlock, ptl);
1016 /* This is called just after a mm has been created, but it has not
1017 been used yet. We need to make sure that its pagetable is all
1018 read-only, and can be pinned. */
1019 static void __xen_pgd_pin(struct mm_struct *mm, pgd_t *pgd)
1023 if (__xen_pgd_walk(mm, pgd, xen_pin_page, USER_LIMIT)) {
1024 /* re-enable interrupts for flushing */
1027 kmap_flush_unused();
1032 #ifdef CONFIG_X86_64
1034 pgd_t *user_pgd = xen_get_user_pgd(pgd);
1036 xen_do_pin(MMUEXT_PIN_L4_TABLE, PFN_DOWN(__pa(pgd)));
1039 xen_pin_page(mm, virt_to_page(user_pgd), PT_PGD);
1040 xen_do_pin(MMUEXT_PIN_L4_TABLE,
1041 PFN_DOWN(__pa(user_pgd)));
1044 #else /* CONFIG_X86_32 */
1045 #ifdef CONFIG_X86_PAE
1046 /* Need to make sure unshared kernel PMD is pinnable */
1047 xen_pin_page(mm, pgd_page(pgd[pgd_index(TASK_SIZE)]),
1050 xen_do_pin(MMUEXT_PIN_L3_TABLE, PFN_DOWN(__pa(pgd)));
1051 #endif /* CONFIG_X86_64 */
1055 static void xen_pgd_pin(struct mm_struct *mm)
1057 __xen_pgd_pin(mm, mm->pgd);
1061 * On save, we need to pin all pagetables to make sure they get their
1062 * mfns turned into pfns. Search the list for any unpinned pgds and pin
1063 * them (unpinned pgds are not currently in use, probably because the
1064 * process is under construction or destruction).
1066 * Expected to be called in stop_machine() ("equivalent to taking
1067 * every spinlock in the system"), so the locking doesn't really
1068 * matter all that much.
1070 void xen_mm_pin_all(void)
1072 unsigned long flags;
1075 spin_lock_irqsave(&pgd_lock, flags);
1077 list_for_each_entry(page, &pgd_list, lru) {
1078 if (!PagePinned(page)) {
1079 __xen_pgd_pin(&init_mm, (pgd_t *)page_address(page));
1080 SetPageSavePinned(page);
1084 spin_unlock_irqrestore(&pgd_lock, flags);
1088 * The init_mm pagetable is really pinned as soon as its created, but
1089 * that's before we have page structures to store the bits. So do all
1090 * the book-keeping now.
1092 static __init int xen_mark_pinned(struct mm_struct *mm, struct page *page,
1093 enum pt_level level)
1095 SetPagePinned(page);
1099 static void __init xen_mark_init_mm_pinned(void)
1101 xen_pgd_walk(&init_mm, xen_mark_pinned, FIXADDR_TOP);
1104 static int xen_unpin_page(struct mm_struct *mm, struct page *page,
1105 enum pt_level level)
1107 unsigned pgfl = TestClearPagePinned(page);
1109 if (pgfl && !PageHighMem(page)) {
1110 void *pt = lowmem_page_address(page);
1111 unsigned long pfn = page_to_pfn(page);
1112 spinlock_t *ptl = NULL;
1113 struct multicall_space mcs;
1116 * Do the converse to pin_page. If we're using split
1117 * pte locks, we must be holding the lock for while
1118 * the pte page is unpinned but still RO to prevent
1119 * concurrent updates from seeing it in this
1120 * partially-pinned state.
1122 if (level == PT_PTE) {
1123 ptl = xen_pte_lock(page, mm);
1126 xen_do_pin(MMUEXT_UNPIN_TABLE, pfn);
1129 mcs = __xen_mc_entry(0);
1131 MULTI_update_va_mapping(mcs.mc, (unsigned long)pt,
1132 pfn_pte(pfn, PAGE_KERNEL),
1133 level == PT_PGD ? UVMF_TLB_FLUSH : 0);
1136 /* unlock when batch completed */
1137 xen_mc_callback(xen_pte_unlock, ptl);
1141 return 0; /* never need to flush on unpin */
1144 /* Release a pagetables pages back as normal RW */
1145 static void __xen_pgd_unpin(struct mm_struct *mm, pgd_t *pgd)
1149 xen_do_pin(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1151 #ifdef CONFIG_X86_64
1153 pgd_t *user_pgd = xen_get_user_pgd(pgd);
1156 xen_do_pin(MMUEXT_UNPIN_TABLE,
1157 PFN_DOWN(__pa(user_pgd)));
1158 xen_unpin_page(mm, virt_to_page(user_pgd), PT_PGD);
1163 #ifdef CONFIG_X86_PAE
1164 /* Need to make sure unshared kernel PMD is unpinned */
1165 xen_unpin_page(mm, pgd_page(pgd[pgd_index(TASK_SIZE)]),
1169 __xen_pgd_walk(mm, pgd, xen_unpin_page, USER_LIMIT);
1174 static void xen_pgd_unpin(struct mm_struct *mm)
1176 __xen_pgd_unpin(mm, mm->pgd);
1180 * On resume, undo any pinning done at save, so that the rest of the
1181 * kernel doesn't see any unexpected pinned pagetables.
1183 void xen_mm_unpin_all(void)
1185 unsigned long flags;
1188 spin_lock_irqsave(&pgd_lock, flags);
1190 list_for_each_entry(page, &pgd_list, lru) {
1191 if (PageSavePinned(page)) {
1192 BUG_ON(!PagePinned(page));
1193 __xen_pgd_unpin(&init_mm, (pgd_t *)page_address(page));
1194 ClearPageSavePinned(page);
1198 spin_unlock_irqrestore(&pgd_lock, flags);
1201 void xen_activate_mm(struct mm_struct *prev, struct mm_struct *next)
1203 spin_lock(&next->page_table_lock);
1205 spin_unlock(&next->page_table_lock);
1208 void xen_dup_mmap(struct mm_struct *oldmm, struct mm_struct *mm)
1210 spin_lock(&mm->page_table_lock);
1212 spin_unlock(&mm->page_table_lock);
1217 /* Another cpu may still have their %cr3 pointing at the pagetable, so
1218 we need to repoint it somewhere else before we can unpin it. */
1219 static void drop_other_mm_ref(void *info)
1221 struct mm_struct *mm = info;
1222 struct mm_struct *active_mm;
1224 active_mm = percpu_read(cpu_tlbstate.active_mm);
1226 if (active_mm == mm)
1227 leave_mm(smp_processor_id());
1229 /* If this cpu still has a stale cr3 reference, then make sure
1230 it has been flushed. */
1231 if (percpu_read(xen_current_cr3) == __pa(mm->pgd))
1232 load_cr3(swapper_pg_dir);
1235 static void xen_drop_mm_ref(struct mm_struct *mm)
1240 if (current->active_mm == mm) {
1241 if (current->mm == mm)
1242 load_cr3(swapper_pg_dir);
1244 leave_mm(smp_processor_id());
1247 /* Get the "official" set of cpus referring to our pagetable. */
1248 if (!alloc_cpumask_var(&mask, GFP_ATOMIC)) {
1249 for_each_online_cpu(cpu) {
1250 if (!cpumask_test_cpu(cpu, mm_cpumask(mm))
1251 && per_cpu(xen_current_cr3, cpu) != __pa(mm->pgd))
1253 smp_call_function_single(cpu, drop_other_mm_ref, mm, 1);
1257 cpumask_copy(mask, mm_cpumask(mm));
1259 /* It's possible that a vcpu may have a stale reference to our
1260 cr3, because its in lazy mode, and it hasn't yet flushed
1261 its set of pending hypercalls yet. In this case, we can
1262 look at its actual current cr3 value, and force it to flush
1264 for_each_online_cpu(cpu) {
1265 if (per_cpu(xen_current_cr3, cpu) == __pa(mm->pgd))
1266 cpumask_set_cpu(cpu, mask);
1269 if (!cpumask_empty(mask))
1270 smp_call_function_many(mask, drop_other_mm_ref, mm, 1);
1271 free_cpumask_var(mask);
1274 static void xen_drop_mm_ref(struct mm_struct *mm)
1276 if (current->active_mm == mm)
1277 load_cr3(swapper_pg_dir);
1282 * While a process runs, Xen pins its pagetables, which means that the
1283 * hypervisor forces it to be read-only, and it controls all updates
1284 * to it. This means that all pagetable updates have to go via the
1285 * hypervisor, which is moderately expensive.
1287 * Since we're pulling the pagetable down, we switch to use init_mm,
1288 * unpin old process pagetable and mark it all read-write, which
1289 * allows further operations on it to be simple memory accesses.
1291 * The only subtle point is that another CPU may be still using the
1292 * pagetable because of lazy tlb flushing. This means we need need to
1293 * switch all CPUs off this pagetable before we can unpin it.
1295 void xen_exit_mmap(struct mm_struct *mm)
1297 get_cpu(); /* make sure we don't move around */
1298 xen_drop_mm_ref(mm);
1301 spin_lock(&mm->page_table_lock);
1303 /* pgd may not be pinned in the error exit path of execve */
1304 if (xen_page_pinned(mm->pgd))
1307 spin_unlock(&mm->page_table_lock);
1310 static __init void xen_pagetable_setup_start(pgd_t *base)
1314 static void xen_post_allocator_init(void);
1316 static __init void xen_pagetable_setup_done(pgd_t *base)
1318 xen_setup_shared_info();
1319 xen_post_allocator_init();
1322 static void xen_write_cr2(unsigned long cr2)
1324 percpu_read(xen_vcpu)->arch.cr2 = cr2;
1327 static unsigned long xen_read_cr2(void)
1329 return percpu_read(xen_vcpu)->arch.cr2;
1332 unsigned long xen_read_cr2_direct(void)
1334 return percpu_read(xen_vcpu_info.arch.cr2);
1337 static void xen_flush_tlb(void)
1339 struct mmuext_op *op;
1340 struct multicall_space mcs;
1344 mcs = xen_mc_entry(sizeof(*op));
1347 op->cmd = MMUEXT_TLB_FLUSH_LOCAL;
1348 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
1350 xen_mc_issue(PARAVIRT_LAZY_MMU);
1355 static void xen_flush_tlb_single(unsigned long addr)
1357 struct mmuext_op *op;
1358 struct multicall_space mcs;
1362 mcs = xen_mc_entry(sizeof(*op));
1364 op->cmd = MMUEXT_INVLPG_LOCAL;
1365 op->arg1.linear_addr = addr & PAGE_MASK;
1366 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
1368 xen_mc_issue(PARAVIRT_LAZY_MMU);
1373 static void xen_flush_tlb_others(const struct cpumask *cpus,
1374 struct mm_struct *mm, unsigned long va)
1377 struct mmuext_op op;
1378 DECLARE_BITMAP(mask, NR_CPUS);
1380 struct multicall_space mcs;
1382 if (cpumask_empty(cpus))
1383 return; /* nothing to do */
1385 mcs = xen_mc_entry(sizeof(*args));
1387 args->op.arg2.vcpumask = to_cpumask(args->mask);
1389 /* Remove us, and any offline CPUS. */
1390 cpumask_and(to_cpumask(args->mask), cpus, cpu_online_mask);
1391 cpumask_clear_cpu(smp_processor_id(), to_cpumask(args->mask));
1393 if (va == TLB_FLUSH_ALL) {
1394 args->op.cmd = MMUEXT_TLB_FLUSH_MULTI;
1396 args->op.cmd = MMUEXT_INVLPG_MULTI;
1397 args->op.arg1.linear_addr = va;
1400 MULTI_mmuext_op(mcs.mc, &args->op, 1, NULL, DOMID_SELF);
1402 xen_mc_issue(PARAVIRT_LAZY_MMU);
1405 static unsigned long xen_read_cr3(void)
1407 return percpu_read(xen_cr3);
1410 static void set_current_cr3(void *v)
1412 percpu_write(xen_current_cr3, (unsigned long)v);
1415 static void __xen_write_cr3(bool kernel, unsigned long cr3)
1417 struct mmuext_op *op;
1418 struct multicall_space mcs;
1422 mfn = pfn_to_mfn(PFN_DOWN(cr3));
1426 WARN_ON(mfn == 0 && kernel);
1428 mcs = __xen_mc_entry(sizeof(*op));
1431 op->cmd = kernel ? MMUEXT_NEW_BASEPTR : MMUEXT_NEW_USER_BASEPTR;
1434 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
1437 percpu_write(xen_cr3, cr3);
1439 /* Update xen_current_cr3 once the batch has actually
1441 xen_mc_callback(set_current_cr3, (void *)cr3);
1445 static void xen_write_cr3(unsigned long cr3)
1447 BUG_ON(preemptible());
1449 xen_mc_batch(); /* disables interrupts */
1451 /* Update while interrupts are disabled, so its atomic with
1453 percpu_write(xen_cr3, cr3);
1455 __xen_write_cr3(true, cr3);
1457 #ifdef CONFIG_X86_64
1459 pgd_t *user_pgd = xen_get_user_pgd(__va(cr3));
1461 __xen_write_cr3(false, __pa(user_pgd));
1463 __xen_write_cr3(false, 0);
1467 xen_mc_issue(PARAVIRT_LAZY_CPU); /* interrupts restored */
1470 static int xen_pgd_alloc(struct mm_struct *mm)
1472 pgd_t *pgd = mm->pgd;
1475 BUG_ON(PagePinned(virt_to_page(pgd)));
1477 #ifdef CONFIG_X86_64
1479 struct page *page = virt_to_page(pgd);
1482 BUG_ON(page->private != 0);
1486 user_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
1487 page->private = (unsigned long)user_pgd;
1489 if (user_pgd != NULL) {
1490 user_pgd[pgd_index(VSYSCALL_START)] =
1491 __pgd(__pa(level3_user_vsyscall) | _PAGE_TABLE);
1495 BUG_ON(PagePinned(virt_to_page(xen_get_user_pgd(pgd))));
1502 static void xen_pgd_free(struct mm_struct *mm, pgd_t *pgd)
1504 #ifdef CONFIG_X86_64
1505 pgd_t *user_pgd = xen_get_user_pgd(pgd);
1508 free_page((unsigned long)user_pgd);
1512 #ifdef CONFIG_X86_32
1513 static __init pte_t mask_rw_pte(pte_t *ptep, pte_t pte)
1515 /* If there's an existing pte, then don't allow _PAGE_RW to be set */
1516 if (pte_val_ma(*ptep) & _PAGE_PRESENT)
1517 pte = __pte_ma(((pte_val_ma(*ptep) & _PAGE_RW) | ~_PAGE_RW) &
1523 /* Init-time set_pte while constructing initial pagetables, which
1524 doesn't allow RO pagetable pages to be remapped RW */
1525 static __init void xen_set_pte_init(pte_t *ptep, pte_t pte)
1527 pte = mask_rw_pte(ptep, pte);
1529 xen_set_pte(ptep, pte);
1533 static void pin_pagetable_pfn(unsigned cmd, unsigned long pfn)
1535 struct mmuext_op op;
1537 op.arg1.mfn = pfn_to_mfn(pfn);
1538 if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
1542 /* Early in boot, while setting up the initial pagetable, assume
1543 everything is pinned. */
1544 static __init void xen_alloc_pte_init(struct mm_struct *mm, unsigned long pfn)
1546 #ifdef CONFIG_FLATMEM
1547 BUG_ON(mem_map); /* should only be used early */
1549 make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
1550 pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
1553 /* Used for pmd and pud */
1554 static __init void xen_alloc_pmd_init(struct mm_struct *mm, unsigned long pfn)
1556 #ifdef CONFIG_FLATMEM
1557 BUG_ON(mem_map); /* should only be used early */
1559 make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
1562 /* Early release_pte assumes that all pts are pinned, since there's
1563 only init_mm and anything attached to that is pinned. */
1564 static __init void xen_release_pte_init(unsigned long pfn)
1566 pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
1567 make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
1570 static __init void xen_release_pmd_init(unsigned long pfn)
1572 make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
1575 /* This needs to make sure the new pte page is pinned iff its being
1576 attached to a pinned pagetable. */
1577 static void xen_alloc_ptpage(struct mm_struct *mm, unsigned long pfn, unsigned level)
1579 struct page *page = pfn_to_page(pfn);
1581 if (PagePinned(virt_to_page(mm->pgd))) {
1582 SetPagePinned(page);
1584 if (!PageHighMem(page)) {
1585 make_lowmem_page_readonly(__va(PFN_PHYS((unsigned long)pfn)));
1586 if (level == PT_PTE && USE_SPLIT_PTLOCKS)
1587 pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
1589 /* make sure there are no stray mappings of
1591 kmap_flush_unused();
1596 static void xen_alloc_pte(struct mm_struct *mm, unsigned long pfn)
1598 xen_alloc_ptpage(mm, pfn, PT_PTE);
1601 static void xen_alloc_pmd(struct mm_struct *mm, unsigned long pfn)
1603 xen_alloc_ptpage(mm, pfn, PT_PMD);
1606 /* This should never happen until we're OK to use struct page */
1607 static void xen_release_ptpage(unsigned long pfn, unsigned level)
1609 struct page *page = pfn_to_page(pfn);
1611 if (PagePinned(page)) {
1612 if (!PageHighMem(page)) {
1613 if (level == PT_PTE && USE_SPLIT_PTLOCKS)
1614 pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
1615 make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
1617 ClearPagePinned(page);
1621 static void xen_release_pte(unsigned long pfn)
1623 xen_release_ptpage(pfn, PT_PTE);
1626 static void xen_release_pmd(unsigned long pfn)
1628 xen_release_ptpage(pfn, PT_PMD);
1631 #if PAGETABLE_LEVELS == 4
1632 static void xen_alloc_pud(struct mm_struct *mm, unsigned long pfn)
1634 xen_alloc_ptpage(mm, pfn, PT_PUD);
1637 static void xen_release_pud(unsigned long pfn)
1639 xen_release_ptpage(pfn, PT_PUD);
1643 void __init xen_reserve_top(void)
1645 #ifdef CONFIG_X86_32
1646 unsigned long top = HYPERVISOR_VIRT_START;
1647 struct xen_platform_parameters pp;
1649 if (HYPERVISOR_xen_version(XENVER_platform_parameters, &pp) == 0)
1650 top = pp.virt_start;
1652 reserve_top_address(-top);
1653 #endif /* CONFIG_X86_32 */
1657 * Like __va(), but returns address in the kernel mapping (which is
1658 * all we have until the physical memory mapping has been set up.
1660 static void *__ka(phys_addr_t paddr)
1662 #ifdef CONFIG_X86_64
1663 return (void *)(paddr + __START_KERNEL_map);
1669 /* Convert a machine address to physical address */
1670 static unsigned long m2p(phys_addr_t maddr)
1674 maddr &= PTE_PFN_MASK;
1675 paddr = mfn_to_pfn(maddr >> PAGE_SHIFT) << PAGE_SHIFT;
1680 /* Convert a machine address to kernel virtual */
1681 static void *m2v(phys_addr_t maddr)
1683 return __ka(m2p(maddr));
1686 static void set_page_prot(void *addr, pgprot_t prot)
1688 unsigned long pfn = __pa(addr) >> PAGE_SHIFT;
1689 pte_t pte = pfn_pte(pfn, prot);
1691 if (HYPERVISOR_update_va_mapping((unsigned long)addr, pte, 0))
1695 static __init void xen_map_identity_early(pmd_t *pmd, unsigned long max_pfn)
1697 unsigned pmdidx, pteidx;
1703 for (pmdidx = 0; pmdidx < PTRS_PER_PMD && pfn < max_pfn; pmdidx++) {
1706 /* Reuse or allocate a page of ptes */
1707 if (pmd_present(pmd[pmdidx]))
1708 pte_page = m2v(pmd[pmdidx].pmd);
1710 /* Check for free pte pages */
1711 if (ident_pte == ARRAY_SIZE(level1_ident_pgt))
1714 pte_page = &level1_ident_pgt[ident_pte];
1715 ident_pte += PTRS_PER_PTE;
1717 pmd[pmdidx] = __pmd(__pa(pte_page) | _PAGE_TABLE);
1720 /* Install mappings */
1721 for (pteidx = 0; pteidx < PTRS_PER_PTE; pteidx++, pfn++) {
1724 if (pfn > max_pfn_mapped)
1725 max_pfn_mapped = pfn;
1727 if (!pte_none(pte_page[pteidx]))
1730 pte = pfn_pte(pfn, PAGE_KERNEL_EXEC);
1731 pte_page[pteidx] = pte;
1735 for (pteidx = 0; pteidx < ident_pte; pteidx += PTRS_PER_PTE)
1736 set_page_prot(&level1_ident_pgt[pteidx], PAGE_KERNEL_RO);
1738 set_page_prot(pmd, PAGE_KERNEL_RO);
1741 #ifdef CONFIG_X86_64
1742 static void convert_pfn_mfn(void *v)
1747 /* All levels are converted the same way, so just treat them
1749 for (i = 0; i < PTRS_PER_PTE; i++)
1750 pte[i] = xen_make_pte(pte[i].pte);
1754 * Set up the inital kernel pagetable.
1756 * We can construct this by grafting the Xen provided pagetable into
1757 * head_64.S's preconstructed pagetables. We copy the Xen L2's into
1758 * level2_ident_pgt, level2_kernel_pgt and level2_fixmap_pgt. This
1759 * means that only the kernel has a physical mapping to start with -
1760 * but that's enough to get __va working. We need to fill in the rest
1761 * of the physical mapping once some sort of allocator has been set
1764 __init pgd_t *xen_setup_kernel_pagetable(pgd_t *pgd,
1765 unsigned long max_pfn)
1770 /* Zap identity mapping */
1771 init_level4_pgt[0] = __pgd(0);
1773 /* Pre-constructed entries are in pfn, so convert to mfn */
1774 convert_pfn_mfn(init_level4_pgt);
1775 convert_pfn_mfn(level3_ident_pgt);
1776 convert_pfn_mfn(level3_kernel_pgt);
1778 l3 = m2v(pgd[pgd_index(__START_KERNEL_map)].pgd);
1779 l2 = m2v(l3[pud_index(__START_KERNEL_map)].pud);
1781 memcpy(level2_ident_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1782 memcpy(level2_kernel_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1784 l3 = m2v(pgd[pgd_index(__START_KERNEL_map + PMD_SIZE)].pgd);
1785 l2 = m2v(l3[pud_index(__START_KERNEL_map + PMD_SIZE)].pud);
1786 memcpy(level2_fixmap_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1788 /* Set up identity map */
1789 xen_map_identity_early(level2_ident_pgt, max_pfn);
1791 /* Make pagetable pieces RO */
1792 set_page_prot(init_level4_pgt, PAGE_KERNEL_RO);
1793 set_page_prot(level3_ident_pgt, PAGE_KERNEL_RO);
1794 set_page_prot(level3_kernel_pgt, PAGE_KERNEL_RO);
1795 set_page_prot(level3_user_vsyscall, PAGE_KERNEL_RO);
1796 set_page_prot(level2_kernel_pgt, PAGE_KERNEL_RO);
1797 set_page_prot(level2_fixmap_pgt, PAGE_KERNEL_RO);
1799 /* Pin down new L4 */
1800 pin_pagetable_pfn(MMUEXT_PIN_L4_TABLE,
1801 PFN_DOWN(__pa_symbol(init_level4_pgt)));
1803 /* Unpin Xen-provided one */
1804 pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1807 pgd = init_level4_pgt;
1810 * At this stage there can be no user pgd, and no page
1811 * structure to attach it to, so make sure we just set kernel
1815 __xen_write_cr3(true, __pa(pgd));
1816 xen_mc_issue(PARAVIRT_LAZY_CPU);
1818 memblock_x86_reserve_range(__pa(xen_start_info->pt_base),
1819 __pa(xen_start_info->pt_base +
1820 xen_start_info->nr_pt_frames * PAGE_SIZE),
1825 #else /* !CONFIG_X86_64 */
1826 static pmd_t level2_kernel_pgt[PTRS_PER_PMD] __page_aligned_bss;
1828 __init pgd_t *xen_setup_kernel_pagetable(pgd_t *pgd,
1829 unsigned long max_pfn)
1833 max_pfn_mapped = PFN_DOWN(__pa(xen_start_info->pt_base) +
1834 xen_start_info->nr_pt_frames * PAGE_SIZE +
1837 kernel_pmd = m2v(pgd[KERNEL_PGD_BOUNDARY].pgd);
1838 memcpy(level2_kernel_pgt, kernel_pmd, sizeof(pmd_t) * PTRS_PER_PMD);
1840 xen_map_identity_early(level2_kernel_pgt, max_pfn);
1842 memcpy(swapper_pg_dir, pgd, sizeof(pgd_t) * PTRS_PER_PGD);
1843 set_pgd(&swapper_pg_dir[KERNEL_PGD_BOUNDARY],
1844 __pgd(__pa(level2_kernel_pgt) | _PAGE_PRESENT));
1846 set_page_prot(level2_kernel_pgt, PAGE_KERNEL_RO);
1847 set_page_prot(swapper_pg_dir, PAGE_KERNEL_RO);
1848 set_page_prot(empty_zero_page, PAGE_KERNEL_RO);
1850 pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1852 xen_write_cr3(__pa(swapper_pg_dir));
1854 pin_pagetable_pfn(MMUEXT_PIN_L3_TABLE, PFN_DOWN(__pa(swapper_pg_dir)));
1856 memblock_x86_reserve_range(__pa(xen_start_info->pt_base),
1857 __pa(xen_start_info->pt_base +
1858 xen_start_info->nr_pt_frames * PAGE_SIZE),
1861 return swapper_pg_dir;
1863 #endif /* CONFIG_X86_64 */
1865 static void xen_set_fixmap(unsigned idx, phys_addr_t phys, pgprot_t prot)
1869 phys >>= PAGE_SHIFT;
1872 case FIX_BTMAP_END ... FIX_BTMAP_BEGIN:
1873 #ifdef CONFIG_X86_F00F_BUG
1876 #ifdef CONFIG_X86_32
1879 # ifdef CONFIG_HIGHMEM
1880 case FIX_KMAP_BEGIN ... FIX_KMAP_END:
1883 case VSYSCALL_LAST_PAGE ... VSYSCALL_FIRST_PAGE:
1885 #ifdef CONFIG_X86_LOCAL_APIC
1886 case FIX_APIC_BASE: /* maps dummy local APIC */
1888 case FIX_TEXT_POKE0:
1889 case FIX_TEXT_POKE1:
1890 /* All local page mappings */
1891 pte = pfn_pte(phys, prot);
1894 case FIX_PARAVIRT_BOOTMAP:
1895 /* This is an MFN, but it isn't an IO mapping from the
1897 pte = mfn_pte(phys, prot);
1901 /* By default, set_fixmap is used for hardware mappings */
1902 pte = mfn_pte(phys, __pgprot(pgprot_val(prot) | _PAGE_IOMAP));
1906 __native_set_fixmap(idx, pte);
1908 #ifdef CONFIG_X86_64
1909 /* Replicate changes to map the vsyscall page into the user
1910 pagetable vsyscall mapping. */
1911 if (idx >= VSYSCALL_LAST_PAGE && idx <= VSYSCALL_FIRST_PAGE) {
1912 unsigned long vaddr = __fix_to_virt(idx);
1913 set_pte_vaddr_pud(level3_user_vsyscall, vaddr, pte);
1918 static __init void xen_post_allocator_init(void)
1920 pv_mmu_ops.set_pte = xen_set_pte;
1921 pv_mmu_ops.set_pmd = xen_set_pmd;
1922 pv_mmu_ops.set_pud = xen_set_pud;
1923 #if PAGETABLE_LEVELS == 4
1924 pv_mmu_ops.set_pgd = xen_set_pgd;
1927 /* This will work as long as patching hasn't happened yet
1928 (which it hasn't) */
1929 pv_mmu_ops.alloc_pte = xen_alloc_pte;
1930 pv_mmu_ops.alloc_pmd = xen_alloc_pmd;
1931 pv_mmu_ops.release_pte = xen_release_pte;
1932 pv_mmu_ops.release_pmd = xen_release_pmd;
1933 #if PAGETABLE_LEVELS == 4
1934 pv_mmu_ops.alloc_pud = xen_alloc_pud;
1935 pv_mmu_ops.release_pud = xen_release_pud;
1938 #ifdef CONFIG_X86_64
1939 SetPagePinned(virt_to_page(level3_user_vsyscall));
1941 xen_mark_init_mm_pinned();
1944 static void xen_leave_lazy_mmu(void)
1948 paravirt_leave_lazy_mmu();
1952 static const struct pv_mmu_ops xen_mmu_ops __initdata = {
1953 .read_cr2 = xen_read_cr2,
1954 .write_cr2 = xen_write_cr2,
1956 .read_cr3 = xen_read_cr3,
1957 .write_cr3 = xen_write_cr3,
1959 .flush_tlb_user = xen_flush_tlb,
1960 .flush_tlb_kernel = xen_flush_tlb,
1961 .flush_tlb_single = xen_flush_tlb_single,
1962 .flush_tlb_others = xen_flush_tlb_others,
1964 .pte_update = paravirt_nop,
1965 .pte_update_defer = paravirt_nop,
1967 .pgd_alloc = xen_pgd_alloc,
1968 .pgd_free = xen_pgd_free,
1970 .alloc_pte = xen_alloc_pte_init,
1971 .release_pte = xen_release_pte_init,
1972 .alloc_pmd = xen_alloc_pmd_init,
1973 .alloc_pmd_clone = paravirt_nop,
1974 .release_pmd = xen_release_pmd_init,
1976 #ifdef CONFIG_X86_64
1977 .set_pte = xen_set_pte,
1979 .set_pte = xen_set_pte_init,
1981 .set_pte_at = xen_set_pte_at,
1982 .set_pmd = xen_set_pmd_hyper,
1984 .ptep_modify_prot_start = __ptep_modify_prot_start,
1985 .ptep_modify_prot_commit = __ptep_modify_prot_commit,
1987 .pte_val = PV_CALLEE_SAVE(xen_pte_val),
1988 .pgd_val = PV_CALLEE_SAVE(xen_pgd_val),
1990 .make_pte = PV_CALLEE_SAVE(xen_make_pte),
1991 .make_pgd = PV_CALLEE_SAVE(xen_make_pgd),
1993 #ifdef CONFIG_X86_PAE
1994 .set_pte_atomic = xen_set_pte_atomic,
1995 .pte_clear = xen_pte_clear,
1996 .pmd_clear = xen_pmd_clear,
1997 #endif /* CONFIG_X86_PAE */
1998 .set_pud = xen_set_pud_hyper,
2000 .make_pmd = PV_CALLEE_SAVE(xen_make_pmd),
2001 .pmd_val = PV_CALLEE_SAVE(xen_pmd_val),
2003 #if PAGETABLE_LEVELS == 4
2004 .pud_val = PV_CALLEE_SAVE(xen_pud_val),
2005 .make_pud = PV_CALLEE_SAVE(xen_make_pud),
2006 .set_pgd = xen_set_pgd_hyper,
2008 .alloc_pud = xen_alloc_pmd_init,
2009 .release_pud = xen_release_pmd_init,
2010 #endif /* PAGETABLE_LEVELS == 4 */
2012 .activate_mm = xen_activate_mm,
2013 .dup_mmap = xen_dup_mmap,
2014 .exit_mmap = xen_exit_mmap,
2017 .enter = paravirt_enter_lazy_mmu,
2018 .leave = xen_leave_lazy_mmu,
2021 .set_fixmap = xen_set_fixmap,
2024 void __init xen_init_mmu_ops(void)
2026 x86_init.paging.pagetable_setup_start = xen_pagetable_setup_start;
2027 x86_init.paging.pagetable_setup_done = xen_pagetable_setup_done;
2028 pv_mmu_ops = xen_mmu_ops;
2030 vmap_lazy_unmap = false;
2033 /* Protected by xen_reservation_lock. */
2034 #define MAX_CONTIG_ORDER 9 /* 2MB */
2035 static unsigned long discontig_frames[1<<MAX_CONTIG_ORDER];
2037 #define VOID_PTE (mfn_pte(0, __pgprot(0)))
2038 static void xen_zap_pfn_range(unsigned long vaddr, unsigned int order,
2039 unsigned long *in_frames,
2040 unsigned long *out_frames)
2043 struct multicall_space mcs;
2046 for (i = 0; i < (1UL<<order); i++, vaddr += PAGE_SIZE) {
2047 mcs = __xen_mc_entry(0);
2050 in_frames[i] = virt_to_mfn(vaddr);
2052 MULTI_update_va_mapping(mcs.mc, vaddr, VOID_PTE, 0);
2053 set_phys_to_machine(virt_to_pfn(vaddr), INVALID_P2M_ENTRY);
2056 out_frames[i] = virt_to_pfn(vaddr);
2062 * Update the pfn-to-mfn mappings for a virtual address range, either to
2063 * point to an array of mfns, or contiguously from a single starting
2066 static void xen_remap_exchanged_ptes(unsigned long vaddr, int order,
2067 unsigned long *mfns,
2068 unsigned long first_mfn)
2075 limit = 1u << order;
2076 for (i = 0; i < limit; i++, vaddr += PAGE_SIZE) {
2077 struct multicall_space mcs;
2080 mcs = __xen_mc_entry(0);
2084 mfn = first_mfn + i;
2086 if (i < (limit - 1))
2090 flags = UVMF_INVLPG | UVMF_ALL;
2092 flags = UVMF_TLB_FLUSH | UVMF_ALL;
2095 MULTI_update_va_mapping(mcs.mc, vaddr,
2096 mfn_pte(mfn, PAGE_KERNEL), flags);
2098 set_phys_to_machine(virt_to_pfn(vaddr), mfn);
2105 * Perform the hypercall to exchange a region of our pfns to point to
2106 * memory with the required contiguous alignment. Takes the pfns as
2107 * input, and populates mfns as output.
2109 * Returns a success code indicating whether the hypervisor was able to
2110 * satisfy the request or not.
2112 static int xen_exchange_memory(unsigned long extents_in, unsigned int order_in,
2113 unsigned long *pfns_in,
2114 unsigned long extents_out,
2115 unsigned int order_out,
2116 unsigned long *mfns_out,
2117 unsigned int address_bits)
2122 struct xen_memory_exchange exchange = {
2124 .nr_extents = extents_in,
2125 .extent_order = order_in,
2126 .extent_start = pfns_in,
2130 .nr_extents = extents_out,
2131 .extent_order = order_out,
2132 .extent_start = mfns_out,
2133 .address_bits = address_bits,
2138 BUG_ON(extents_in << order_in != extents_out << order_out);
2140 rc = HYPERVISOR_memory_op(XENMEM_exchange, &exchange);
2141 success = (exchange.nr_exchanged == extents_in);
2143 BUG_ON(!success && ((exchange.nr_exchanged != 0) || (rc == 0)));
2144 BUG_ON(success && (rc != 0));
2149 int xen_create_contiguous_region(unsigned long vstart, unsigned int order,
2150 unsigned int address_bits)
2152 unsigned long *in_frames = discontig_frames, out_frame;
2153 unsigned long flags;
2157 * Currently an auto-translated guest will not perform I/O, nor will
2158 * it require PAE page directories below 4GB. Therefore any calls to
2159 * this function are redundant and can be ignored.
2162 if (xen_feature(XENFEAT_auto_translated_physmap))
2165 if (unlikely(order > MAX_CONTIG_ORDER))
2168 memset((void *) vstart, 0, PAGE_SIZE << order);
2170 spin_lock_irqsave(&xen_reservation_lock, flags);
2172 /* 1. Zap current PTEs, remembering MFNs. */
2173 xen_zap_pfn_range(vstart, order, in_frames, NULL);
2175 /* 2. Get a new contiguous memory extent. */
2176 out_frame = virt_to_pfn(vstart);
2177 success = xen_exchange_memory(1UL << order, 0, in_frames,
2178 1, order, &out_frame,
2181 /* 3. Map the new extent in place of old pages. */
2183 xen_remap_exchanged_ptes(vstart, order, NULL, out_frame);
2185 xen_remap_exchanged_ptes(vstart, order, in_frames, 0);
2187 spin_unlock_irqrestore(&xen_reservation_lock, flags);
2189 return success ? 0 : -ENOMEM;
2191 EXPORT_SYMBOL_GPL(xen_create_contiguous_region);
2193 void xen_destroy_contiguous_region(unsigned long vstart, unsigned int order)
2195 unsigned long *out_frames = discontig_frames, in_frame;
2196 unsigned long flags;
2199 if (xen_feature(XENFEAT_auto_translated_physmap))
2202 if (unlikely(order > MAX_CONTIG_ORDER))
2205 memset((void *) vstart, 0, PAGE_SIZE << order);
2207 spin_lock_irqsave(&xen_reservation_lock, flags);
2209 /* 1. Find start MFN of contiguous extent. */
2210 in_frame = virt_to_mfn(vstart);
2212 /* 2. Zap current PTEs. */
2213 xen_zap_pfn_range(vstart, order, NULL, out_frames);
2215 /* 3. Do the exchange for non-contiguous MFNs. */
2216 success = xen_exchange_memory(1, order, &in_frame, 1UL << order,
2219 /* 4. Map new pages in place of old pages. */
2221 xen_remap_exchanged_ptes(vstart, order, out_frames, 0);
2223 xen_remap_exchanged_ptes(vstart, order, NULL, in_frame);
2225 spin_unlock_irqrestore(&xen_reservation_lock, flags);
2227 EXPORT_SYMBOL_GPL(xen_destroy_contiguous_region);
2229 #ifdef CONFIG_XEN_PVHVM
2230 static void xen_hvm_exit_mmap(struct mm_struct *mm)
2232 struct xen_hvm_pagetable_dying a;
2235 a.domid = DOMID_SELF;
2236 a.gpa = __pa(mm->pgd);
2237 rc = HYPERVISOR_hvm_op(HVMOP_pagetable_dying, &a);
2238 WARN_ON_ONCE(rc < 0);
2241 static int is_pagetable_dying_supported(void)
2243 struct xen_hvm_pagetable_dying a;
2246 a.domid = DOMID_SELF;
2248 rc = HYPERVISOR_hvm_op(HVMOP_pagetable_dying, &a);
2250 printk(KERN_DEBUG "HVMOP_pagetable_dying not supported\n");
2256 void __init xen_hvm_init_mmu_ops(void)
2258 if (is_pagetable_dying_supported())
2259 pv_mmu_ops.exit_mmap = xen_hvm_exit_mmap;
2263 #ifdef CONFIG_XEN_DEBUG_FS
2265 static struct dentry *d_mmu_debug;
2267 static int __init xen_mmu_debugfs(void)
2269 struct dentry *d_xen = xen_init_debugfs();
2274 d_mmu_debug = debugfs_create_dir("mmu", d_xen);
2276 debugfs_create_u8("zero_stats", 0644, d_mmu_debug, &zero_stats);
2278 debugfs_create_u32("pgd_update", 0444, d_mmu_debug, &mmu_stats.pgd_update);
2279 debugfs_create_u32("pgd_update_pinned", 0444, d_mmu_debug,
2280 &mmu_stats.pgd_update_pinned);
2281 debugfs_create_u32("pgd_update_batched", 0444, d_mmu_debug,
2282 &mmu_stats.pgd_update_pinned);
2284 debugfs_create_u32("pud_update", 0444, d_mmu_debug, &mmu_stats.pud_update);
2285 debugfs_create_u32("pud_update_pinned", 0444, d_mmu_debug,
2286 &mmu_stats.pud_update_pinned);
2287 debugfs_create_u32("pud_update_batched", 0444, d_mmu_debug,
2288 &mmu_stats.pud_update_pinned);
2290 debugfs_create_u32("pmd_update", 0444, d_mmu_debug, &mmu_stats.pmd_update);
2291 debugfs_create_u32("pmd_update_pinned", 0444, d_mmu_debug,
2292 &mmu_stats.pmd_update_pinned);
2293 debugfs_create_u32("pmd_update_batched", 0444, d_mmu_debug,
2294 &mmu_stats.pmd_update_pinned);
2296 debugfs_create_u32("pte_update", 0444, d_mmu_debug, &mmu_stats.pte_update);
2297 // debugfs_create_u32("pte_update_pinned", 0444, d_mmu_debug,
2298 // &mmu_stats.pte_update_pinned);
2299 debugfs_create_u32("pte_update_batched", 0444, d_mmu_debug,
2300 &mmu_stats.pte_update_pinned);
2302 debugfs_create_u32("mmu_update", 0444, d_mmu_debug, &mmu_stats.mmu_update);
2303 debugfs_create_u32("mmu_update_extended", 0444, d_mmu_debug,
2304 &mmu_stats.mmu_update_extended);
2305 xen_debugfs_create_u32_array("mmu_update_histo", 0444, d_mmu_debug,
2306 mmu_stats.mmu_update_histo, 20);
2308 debugfs_create_u32("set_pte_at", 0444, d_mmu_debug, &mmu_stats.set_pte_at);
2309 debugfs_create_u32("set_pte_at_batched", 0444, d_mmu_debug,
2310 &mmu_stats.set_pte_at_batched);
2311 debugfs_create_u32("set_pte_at_current", 0444, d_mmu_debug,
2312 &mmu_stats.set_pte_at_current);
2313 debugfs_create_u32("set_pte_at_kernel", 0444, d_mmu_debug,
2314 &mmu_stats.set_pte_at_kernel);
2316 debugfs_create_u32("prot_commit", 0444, d_mmu_debug, &mmu_stats.prot_commit);
2317 debugfs_create_u32("prot_commit_batched", 0444, d_mmu_debug,
2318 &mmu_stats.prot_commit_batched);
2322 fs_initcall(xen_mmu_debugfs);
2324 #endif /* CONFIG_XEN_DEBUG_FS */