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>
49 #include <asm/pgtable.h>
50 #include <asm/tlbflush.h>
51 #include <asm/fixmap.h>
52 #include <asm/mmu_context.h>
53 #include <asm/setup.h>
54 #include <asm/paravirt.h>
56 #include <asm/linkage.h>
59 #include <asm/xen/hypercall.h>
60 #include <asm/xen/hypervisor.h>
64 #include <xen/interface/xen.h>
65 #include <xen/interface/hvm/hvm_op.h>
66 #include <xen/interface/version.h>
67 #include <xen/interface/memory.h>
68 #include <xen/hvc-console.h>
70 #include "multicalls.h"
74 #define MMU_UPDATE_HISTO 30
77 * Protects atomic reservation decrease/increase against concurrent increases.
78 * Also protects non-atomic updates of current_pages and driver_pages, and
81 DEFINE_SPINLOCK(xen_reservation_lock);
83 #ifdef CONFIG_XEN_DEBUG_FS
87 u32 pgd_update_pinned;
88 u32 pgd_update_batched;
91 u32 pud_update_pinned;
92 u32 pud_update_batched;
95 u32 pmd_update_pinned;
96 u32 pmd_update_batched;
99 u32 pte_update_pinned;
100 u32 pte_update_batched;
103 u32 mmu_update_extended;
104 u32 mmu_update_histo[MMU_UPDATE_HISTO];
107 u32 prot_commit_batched;
110 u32 set_pte_at_batched;
111 u32 set_pte_at_pinned;
112 u32 set_pte_at_current;
113 u32 set_pte_at_kernel;
116 static u8 zero_stats;
118 static inline void check_zero(void)
120 if (unlikely(zero_stats)) {
121 memset(&mmu_stats, 0, sizeof(mmu_stats));
126 #define ADD_STATS(elem, val) \
127 do { check_zero(); mmu_stats.elem += (val); } while(0)
129 #else /* !CONFIG_XEN_DEBUG_FS */
131 #define ADD_STATS(elem, val) do { (void)(val); } while(0)
133 #endif /* CONFIG_XEN_DEBUG_FS */
137 * Identity map, in addition to plain kernel map. This needs to be
138 * large enough to allocate page table pages to allocate the rest.
139 * Each page can map 2MB.
141 static pte_t level1_ident_pgt[PTRS_PER_PTE * 4] __page_aligned_bss;
144 /* l3 pud for userspace vsyscall mapping */
145 static pud_t level3_user_vsyscall[PTRS_PER_PUD] __page_aligned_bss;
146 #endif /* CONFIG_X86_64 */
149 * Note about cr3 (pagetable base) values:
151 * xen_cr3 contains the current logical cr3 value; it contains the
152 * last set cr3. This may not be the current effective cr3, because
153 * its update may be being lazily deferred. However, a vcpu looking
154 * at its own cr3 can use this value knowing that it everything will
155 * be self-consistent.
157 * xen_current_cr3 contains the actual vcpu cr3; it is set once the
158 * hypercall to set the vcpu cr3 is complete (so it may be a little
159 * out of date, but it will never be set early). If one vcpu is
160 * looking at another vcpu's cr3 value, it should use this variable.
162 DEFINE_PER_CPU(unsigned long, xen_cr3); /* cr3 stored as physaddr */
163 DEFINE_PER_CPU(unsigned long, xen_current_cr3); /* actual vcpu cr3 */
167 * Just beyond the highest usermode address. STACK_TOP_MAX has a
168 * redzone above it, so round it up to a PGD boundary.
170 #define USER_LIMIT ((STACK_TOP_MAX + PGDIR_SIZE - 1) & PGDIR_MASK)
172 static unsigned long max_p2m_pfn __read_mostly = MAX_DOMAIN_PAGES;
174 #define P2M_ENTRIES_PER_PAGE (PAGE_SIZE / sizeof(unsigned long))
175 #define TOP_ENTRIES(pages) ((pages) / P2M_ENTRIES_PER_PAGE)
176 #define MAX_TOP_ENTRIES TOP_ENTRIES(MAX_DOMAIN_PAGES)
178 /* Placeholder for holes in the address space */
179 static RESERVE_BRK_ARRAY(unsigned long, p2m_missing, P2M_ENTRIES_PER_PAGE);
181 /* Array of pointers to pages containing p2m entries */
182 static RESERVE_BRK_ARRAY(unsigned long *, p2m_top, MAX_TOP_ENTRIES);
184 /* Arrays of p2m arrays expressed in mfns used for save/restore */
185 static RESERVE_BRK_ARRAY(unsigned long, p2m_top_mfn, MAX_TOP_ENTRIES);
187 static RESERVE_BRK_ARRAY(unsigned long, p2m_top_mfn_list,
188 (MAX_TOP_ENTRIES / P2M_ENTRIES_PER_PAGE));
190 static inline unsigned p2m_top_index(unsigned long pfn)
192 BUG_ON(pfn >= max_p2m_pfn);
193 return pfn / P2M_ENTRIES_PER_PAGE;
196 static inline unsigned p2m_index(unsigned long pfn)
198 return pfn % P2M_ENTRIES_PER_PAGE;
201 /* Build the parallel p2m_top_mfn structures */
202 void xen_build_mfn_list_list(void)
206 for (pfn = 0; pfn < max_p2m_pfn; pfn += P2M_ENTRIES_PER_PAGE) {
207 unsigned topidx = p2m_top_index(pfn);
209 p2m_top_mfn[topidx] = virt_to_mfn(p2m_top[topidx]);
213 idx < TOP_ENTRIES(max_p2m_pfn)/P2M_ENTRIES_PER_PAGE;
215 unsigned topidx = idx * P2M_ENTRIES_PER_PAGE;
216 p2m_top_mfn_list[idx] = virt_to_mfn(&p2m_top_mfn[topidx]);
220 void xen_setup_mfn_list_list(void)
222 BUG_ON(HYPERVISOR_shared_info == &xen_dummy_shared_info);
224 HYPERVISOR_shared_info->arch.pfn_to_mfn_frame_list_list =
225 virt_to_mfn(p2m_top_mfn_list);
226 HYPERVISOR_shared_info->arch.max_pfn = xen_start_info->nr_pages;
229 /* Set up p2m_top to point to the domain-builder provided p2m pages */
230 void __init xen_build_dynamic_phys_to_machine(void)
232 unsigned long *mfn_list = (unsigned long *)xen_start_info->mfn_list;
233 unsigned long max_pfn = min(MAX_DOMAIN_PAGES, xen_start_info->nr_pages);
237 max_p2m_pfn = max_pfn;
239 p2m_missing = extend_brk(sizeof(*p2m_missing) * P2M_ENTRIES_PER_PAGE,
241 for (i = 0; i < P2M_ENTRIES_PER_PAGE; i++)
242 p2m_missing[i] = ~0UL;
244 p2m_top = extend_brk(sizeof(*p2m_top) * TOP_ENTRIES(max_pfn),
246 for (i = 0; i < TOP_ENTRIES(max_pfn); i++)
247 p2m_top[i] = p2m_missing;
249 p2m_top_mfn = extend_brk(sizeof(*p2m_top_mfn) * TOP_ENTRIES(max_pfn),
251 p2m_top_mfn_list = extend_brk(sizeof(*p2m_top_mfn_list) *
252 (TOP_ENTRIES(max_pfn) / P2M_ENTRIES_PER_PAGE),
255 for (pfn = 0; pfn < max_pfn; pfn += P2M_ENTRIES_PER_PAGE) {
256 unsigned topidx = p2m_top_index(pfn);
258 p2m_top[topidx] = &mfn_list[pfn];
261 xen_build_mfn_list_list();
264 unsigned long get_phys_to_machine(unsigned long pfn)
266 unsigned topidx, idx;
268 if (unlikely(pfn >= max_p2m_pfn))
269 return INVALID_P2M_ENTRY;
271 topidx = p2m_top_index(pfn);
272 idx = p2m_index(pfn);
273 return p2m_top[topidx][idx];
275 EXPORT_SYMBOL_GPL(get_phys_to_machine);
277 /* install a new p2m_top page */
278 bool install_p2mtop_page(unsigned long pfn, unsigned long *p)
280 unsigned topidx = p2m_top_index(pfn);
281 unsigned long **pfnp, *mfnp;
284 pfnp = &p2m_top[topidx];
285 mfnp = &p2m_top_mfn[topidx];
287 for (i = 0; i < P2M_ENTRIES_PER_PAGE; i++)
288 p[i] = INVALID_P2M_ENTRY;
290 if (cmpxchg(pfnp, p2m_missing, p) == p2m_missing) {
291 *mfnp = virt_to_mfn(p);
298 static void alloc_p2m(unsigned long pfn)
302 p = (void *)__get_free_page(GFP_KERNEL | __GFP_NOFAIL);
305 if (!install_p2mtop_page(pfn, p))
306 free_page((unsigned long)p);
309 /* Try to install p2m mapping; fail if intermediate bits missing */
310 bool __set_phys_to_machine(unsigned long pfn, unsigned long mfn)
312 unsigned topidx, idx;
314 if (unlikely(pfn >= max_p2m_pfn)) {
315 BUG_ON(mfn != INVALID_P2M_ENTRY);
319 topidx = p2m_top_index(pfn);
320 if (p2m_top[topidx] == p2m_missing) {
321 if (mfn == INVALID_P2M_ENTRY)
326 idx = p2m_index(pfn);
327 p2m_top[topidx][idx] = mfn;
332 void set_phys_to_machine(unsigned long pfn, unsigned long mfn)
334 if (unlikely(xen_feature(XENFEAT_auto_translated_physmap))) {
335 BUG_ON(pfn != mfn && mfn != INVALID_P2M_ENTRY);
339 if (unlikely(!__set_phys_to_machine(pfn, mfn))) {
342 if (!__set_phys_to_machine(pfn, mfn))
347 unsigned long arbitrary_virt_to_mfn(void *vaddr)
349 xmaddr_t maddr = arbitrary_virt_to_machine(vaddr);
351 return PFN_DOWN(maddr.maddr);
354 xmaddr_t arbitrary_virt_to_machine(void *vaddr)
356 unsigned long address = (unsigned long)vaddr;
362 * if the PFN is in the linear mapped vaddr range, we can just use
363 * the (quick) virt_to_machine() p2m lookup
365 if (virt_addr_valid(vaddr))
366 return virt_to_machine(vaddr);
368 /* otherwise we have to do a (slower) full page-table walk */
370 pte = lookup_address(address, &level);
372 offset = address & ~PAGE_MASK;
373 return XMADDR(((phys_addr_t)pte_mfn(*pte) << PAGE_SHIFT) + offset);
376 void make_lowmem_page_readonly(void *vaddr)
379 unsigned long address = (unsigned long)vaddr;
382 pte = lookup_address(address, &level);
385 ptev = pte_wrprotect(*pte);
387 if (HYPERVISOR_update_va_mapping(address, ptev, 0))
391 void make_lowmem_page_readwrite(void *vaddr)
394 unsigned long address = (unsigned long)vaddr;
397 pte = lookup_address(address, &level);
400 ptev = pte_mkwrite(*pte);
402 if (HYPERVISOR_update_va_mapping(address, ptev, 0))
407 static bool xen_page_pinned(void *ptr)
409 struct page *page = virt_to_page(ptr);
411 return PagePinned(page);
414 static bool xen_iomap_pte(pte_t pte)
416 return pte_flags(pte) & _PAGE_IOMAP;
419 static void xen_set_iomap_pte(pte_t *ptep, pte_t pteval)
421 struct multicall_space mcs;
422 struct mmu_update *u;
424 mcs = xen_mc_entry(sizeof(*u));
427 /* ptep might be kmapped when using 32-bit HIGHPTE */
428 u->ptr = arbitrary_virt_to_machine(ptep).maddr;
429 u->val = pte_val_ma(pteval);
431 MULTI_mmu_update(mcs.mc, mcs.args, 1, NULL, DOMID_IO);
433 xen_mc_issue(PARAVIRT_LAZY_MMU);
436 static void xen_extend_mmu_update(const struct mmu_update *update)
438 struct multicall_space mcs;
439 struct mmu_update *u;
441 mcs = xen_mc_extend_args(__HYPERVISOR_mmu_update, sizeof(*u));
443 if (mcs.mc != NULL) {
444 ADD_STATS(mmu_update_extended, 1);
445 ADD_STATS(mmu_update_histo[mcs.mc->args[1]], -1);
449 if (mcs.mc->args[1] < MMU_UPDATE_HISTO)
450 ADD_STATS(mmu_update_histo[mcs.mc->args[1]], 1);
452 ADD_STATS(mmu_update_histo[0], 1);
454 ADD_STATS(mmu_update, 1);
455 mcs = __xen_mc_entry(sizeof(*u));
456 MULTI_mmu_update(mcs.mc, mcs.args, 1, NULL, DOMID_SELF);
457 ADD_STATS(mmu_update_histo[1], 1);
464 void xen_set_pmd_hyper(pmd_t *ptr, pmd_t val)
472 /* ptr may be ioremapped for 64-bit pagetable setup */
473 u.ptr = arbitrary_virt_to_machine(ptr).maddr;
474 u.val = pmd_val_ma(val);
475 xen_extend_mmu_update(&u);
477 ADD_STATS(pmd_update_batched, paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU);
479 xen_mc_issue(PARAVIRT_LAZY_MMU);
484 void xen_set_pmd(pmd_t *ptr, pmd_t val)
486 ADD_STATS(pmd_update, 1);
488 /* If page is not pinned, we can just update the entry
490 if (!xen_page_pinned(ptr)) {
495 ADD_STATS(pmd_update_pinned, 1);
497 xen_set_pmd_hyper(ptr, val);
501 * Associate a virtual page frame with a given physical page frame
502 * and protection flags for that frame.
504 void set_pte_mfn(unsigned long vaddr, unsigned long mfn, pgprot_t flags)
506 set_pte_vaddr(vaddr, mfn_pte(mfn, flags));
509 void xen_set_pte_at(struct mm_struct *mm, unsigned long addr,
510 pte_t *ptep, pte_t pteval)
512 if (xen_iomap_pte(pteval)) {
513 xen_set_iomap_pte(ptep, pteval);
517 ADD_STATS(set_pte_at, 1);
518 // ADD_STATS(set_pte_at_pinned, xen_page_pinned(ptep));
519 ADD_STATS(set_pte_at_current, mm == current->mm);
520 ADD_STATS(set_pte_at_kernel, mm == &init_mm);
522 if (mm == current->mm || mm == &init_mm) {
523 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU) {
524 struct multicall_space mcs;
525 mcs = xen_mc_entry(0);
527 MULTI_update_va_mapping(mcs.mc, addr, pteval, 0);
528 ADD_STATS(set_pte_at_batched, 1);
529 xen_mc_issue(PARAVIRT_LAZY_MMU);
532 if (HYPERVISOR_update_va_mapping(addr, pteval, 0) == 0)
535 xen_set_pte(ptep, pteval);
540 pte_t xen_ptep_modify_prot_start(struct mm_struct *mm,
541 unsigned long addr, pte_t *ptep)
543 /* Just return the pte as-is. We preserve the bits on commit */
547 void xen_ptep_modify_prot_commit(struct mm_struct *mm, unsigned long addr,
548 pte_t *ptep, pte_t pte)
554 u.ptr = arbitrary_virt_to_machine(ptep).maddr | MMU_PT_UPDATE_PRESERVE_AD;
555 u.val = pte_val_ma(pte);
556 xen_extend_mmu_update(&u);
558 ADD_STATS(prot_commit, 1);
559 ADD_STATS(prot_commit_batched, paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU);
561 xen_mc_issue(PARAVIRT_LAZY_MMU);
564 /* Assume pteval_t is equivalent to all the other *val_t types. */
565 static pteval_t pte_mfn_to_pfn(pteval_t val)
567 if (val & _PAGE_PRESENT) {
568 unsigned long mfn = (val & PTE_PFN_MASK) >> PAGE_SHIFT;
569 pteval_t flags = val & PTE_FLAGS_MASK;
570 val = ((pteval_t)mfn_to_pfn(mfn) << PAGE_SHIFT) | flags;
576 static pteval_t pte_pfn_to_mfn(pteval_t val)
578 if (val & _PAGE_PRESENT) {
579 unsigned long pfn = (val & PTE_PFN_MASK) >> PAGE_SHIFT;
580 pteval_t flags = val & PTE_FLAGS_MASK;
581 val = ((pteval_t)pfn_to_mfn(pfn) << PAGE_SHIFT) | flags;
587 static pteval_t iomap_pte(pteval_t val)
589 if (val & _PAGE_PRESENT) {
590 unsigned long pfn = (val & PTE_PFN_MASK) >> PAGE_SHIFT;
591 pteval_t flags = val & PTE_FLAGS_MASK;
593 /* We assume the pte frame number is a MFN, so
594 just use it as-is. */
595 val = ((pteval_t)pfn << PAGE_SHIFT) | flags;
601 pteval_t xen_pte_val(pte_t pte)
603 if (xen_initial_domain() && (pte.pte & _PAGE_IOMAP))
606 return pte_mfn_to_pfn(pte.pte);
608 PV_CALLEE_SAVE_REGS_THUNK(xen_pte_val);
610 pgdval_t xen_pgd_val(pgd_t pgd)
612 return pte_mfn_to_pfn(pgd.pgd);
614 PV_CALLEE_SAVE_REGS_THUNK(xen_pgd_val);
616 pte_t xen_make_pte(pteval_t pte)
618 phys_addr_t addr = (pte & PTE_PFN_MASK);
621 * Unprivileged domains are allowed to do IOMAPpings for
622 * PCI passthrough, but not map ISA space. The ISA
623 * mappings are just dummy local mappings to keep other
624 * parts of the kernel happy.
626 if (unlikely(pte & _PAGE_IOMAP) &&
627 (xen_initial_domain() || addr >= ISA_END_ADDRESS)) {
628 pte = iomap_pte(pte);
631 pte = pte_pfn_to_mfn(pte);
634 return native_make_pte(pte);
636 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pte);
638 pgd_t xen_make_pgd(pgdval_t pgd)
640 pgd = pte_pfn_to_mfn(pgd);
641 return native_make_pgd(pgd);
643 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pgd);
645 pmdval_t xen_pmd_val(pmd_t pmd)
647 return pte_mfn_to_pfn(pmd.pmd);
649 PV_CALLEE_SAVE_REGS_THUNK(xen_pmd_val);
651 void xen_set_pud_hyper(pud_t *ptr, pud_t val)
659 /* ptr may be ioremapped for 64-bit pagetable setup */
660 u.ptr = arbitrary_virt_to_machine(ptr).maddr;
661 u.val = pud_val_ma(val);
662 xen_extend_mmu_update(&u);
664 ADD_STATS(pud_update_batched, paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU);
666 xen_mc_issue(PARAVIRT_LAZY_MMU);
671 void xen_set_pud(pud_t *ptr, pud_t val)
673 ADD_STATS(pud_update, 1);
675 /* If page is not pinned, we can just update the entry
677 if (!xen_page_pinned(ptr)) {
682 ADD_STATS(pud_update_pinned, 1);
684 xen_set_pud_hyper(ptr, val);
687 void xen_set_pte(pte_t *ptep, pte_t pte)
689 if (xen_iomap_pte(pte)) {
690 xen_set_iomap_pte(ptep, pte);
694 ADD_STATS(pte_update, 1);
695 // ADD_STATS(pte_update_pinned, xen_page_pinned(ptep));
696 ADD_STATS(pte_update_batched, paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU);
698 #ifdef CONFIG_X86_PAE
699 ptep->pte_high = pte.pte_high;
701 ptep->pte_low = pte.pte_low;
707 #ifdef CONFIG_X86_PAE
708 void xen_set_pte_atomic(pte_t *ptep, pte_t pte)
710 if (xen_iomap_pte(pte)) {
711 xen_set_iomap_pte(ptep, pte);
715 set_64bit((u64 *)ptep, native_pte_val(pte));
718 void xen_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
721 smp_wmb(); /* make sure low gets written first */
725 void xen_pmd_clear(pmd_t *pmdp)
727 set_pmd(pmdp, __pmd(0));
729 #endif /* CONFIG_X86_PAE */
731 pmd_t xen_make_pmd(pmdval_t pmd)
733 pmd = pte_pfn_to_mfn(pmd);
734 return native_make_pmd(pmd);
736 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pmd);
738 #if PAGETABLE_LEVELS == 4
739 pudval_t xen_pud_val(pud_t pud)
741 return pte_mfn_to_pfn(pud.pud);
743 PV_CALLEE_SAVE_REGS_THUNK(xen_pud_val);
745 pud_t xen_make_pud(pudval_t pud)
747 pud = pte_pfn_to_mfn(pud);
749 return native_make_pud(pud);
751 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pud);
753 pgd_t *xen_get_user_pgd(pgd_t *pgd)
755 pgd_t *pgd_page = (pgd_t *)(((unsigned long)pgd) & PAGE_MASK);
756 unsigned offset = pgd - pgd_page;
757 pgd_t *user_ptr = NULL;
759 if (offset < pgd_index(USER_LIMIT)) {
760 struct page *page = virt_to_page(pgd_page);
761 user_ptr = (pgd_t *)page->private;
769 static void __xen_set_pgd_hyper(pgd_t *ptr, pgd_t val)
773 u.ptr = virt_to_machine(ptr).maddr;
774 u.val = pgd_val_ma(val);
775 xen_extend_mmu_update(&u);
779 * Raw hypercall-based set_pgd, intended for in early boot before
780 * there's a page structure. This implies:
781 * 1. The only existing pagetable is the kernel's
782 * 2. It is always pinned
783 * 3. It has no user pagetable attached to it
785 void __init xen_set_pgd_hyper(pgd_t *ptr, pgd_t val)
791 __xen_set_pgd_hyper(ptr, val);
793 xen_mc_issue(PARAVIRT_LAZY_MMU);
798 void xen_set_pgd(pgd_t *ptr, pgd_t val)
800 pgd_t *user_ptr = xen_get_user_pgd(ptr);
802 ADD_STATS(pgd_update, 1);
804 /* If page is not pinned, we can just update the entry
806 if (!xen_page_pinned(ptr)) {
809 WARN_ON(xen_page_pinned(user_ptr));
815 ADD_STATS(pgd_update_pinned, 1);
816 ADD_STATS(pgd_update_batched, paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU);
818 /* If it's pinned, then we can at least batch the kernel and
819 user updates together. */
822 __xen_set_pgd_hyper(ptr, val);
824 __xen_set_pgd_hyper(user_ptr, val);
826 xen_mc_issue(PARAVIRT_LAZY_MMU);
828 #endif /* PAGETABLE_LEVELS == 4 */
831 * (Yet another) pagetable walker. This one is intended for pinning a
832 * pagetable. This means that it walks a pagetable and calls the
833 * callback function on each page it finds making up the page table,
834 * at every level. It walks the entire pagetable, but it only bothers
835 * pinning pte pages which are below limit. In the normal case this
836 * will be STACK_TOP_MAX, but at boot we need to pin up to
839 * For 32-bit the important bit is that we don't pin beyond there,
840 * because then we start getting into Xen's ptes.
842 * For 64-bit, we must skip the Xen hole in the middle of the address
843 * space, just after the big x86-64 virtual hole.
845 static int __xen_pgd_walk(struct mm_struct *mm, pgd_t *pgd,
846 int (*func)(struct mm_struct *mm, struct page *,
851 unsigned hole_low, hole_high;
852 unsigned pgdidx_limit, pudidx_limit, pmdidx_limit;
853 unsigned pgdidx, pudidx, pmdidx;
855 /* The limit is the last byte to be touched */
857 BUG_ON(limit >= FIXADDR_TOP);
859 if (xen_feature(XENFEAT_auto_translated_physmap))
863 * 64-bit has a great big hole in the middle of the address
864 * space, which contains the Xen mappings. On 32-bit these
865 * will end up making a zero-sized hole and so is a no-op.
867 hole_low = pgd_index(USER_LIMIT);
868 hole_high = pgd_index(PAGE_OFFSET);
870 pgdidx_limit = pgd_index(limit);
872 pudidx_limit = pud_index(limit);
877 pmdidx_limit = pmd_index(limit);
882 for (pgdidx = 0; pgdidx <= pgdidx_limit; pgdidx++) {
885 if (pgdidx >= hole_low && pgdidx < hole_high)
888 if (!pgd_val(pgd[pgdidx]))
891 pud = pud_offset(&pgd[pgdidx], 0);
893 if (PTRS_PER_PUD > 1) /* not folded */
894 flush |= (*func)(mm, virt_to_page(pud), PT_PUD);
896 for (pudidx = 0; pudidx < PTRS_PER_PUD; pudidx++) {
899 if (pgdidx == pgdidx_limit &&
900 pudidx > pudidx_limit)
903 if (pud_none(pud[pudidx]))
906 pmd = pmd_offset(&pud[pudidx], 0);
908 if (PTRS_PER_PMD > 1) /* not folded */
909 flush |= (*func)(mm, virt_to_page(pmd), PT_PMD);
911 for (pmdidx = 0; pmdidx < PTRS_PER_PMD; pmdidx++) {
914 if (pgdidx == pgdidx_limit &&
915 pudidx == pudidx_limit &&
916 pmdidx > pmdidx_limit)
919 if (pmd_none(pmd[pmdidx]))
922 pte = pmd_page(pmd[pmdidx]);
923 flush |= (*func)(mm, pte, PT_PTE);
929 /* Do the top level last, so that the callbacks can use it as
930 a cue to do final things like tlb flushes. */
931 flush |= (*func)(mm, virt_to_page(pgd), PT_PGD);
936 static int xen_pgd_walk(struct mm_struct *mm,
937 int (*func)(struct mm_struct *mm, struct page *,
941 return __xen_pgd_walk(mm, mm->pgd, func, limit);
944 /* If we're using split pte locks, then take the page's lock and
945 return a pointer to it. Otherwise return NULL. */
946 static spinlock_t *xen_pte_lock(struct page *page, struct mm_struct *mm)
948 spinlock_t *ptl = NULL;
950 #if USE_SPLIT_PTLOCKS
951 ptl = __pte_lockptr(page);
952 spin_lock_nest_lock(ptl, &mm->page_table_lock);
958 static void xen_pte_unlock(void *v)
964 static void xen_do_pin(unsigned level, unsigned long pfn)
966 struct mmuext_op *op;
967 struct multicall_space mcs;
969 mcs = __xen_mc_entry(sizeof(*op));
972 op->arg1.mfn = pfn_to_mfn(pfn);
973 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
976 static int xen_pin_page(struct mm_struct *mm, struct page *page,
979 unsigned pgfl = TestSetPagePinned(page);
983 flush = 0; /* already pinned */
984 else if (PageHighMem(page))
985 /* kmaps need flushing if we found an unpinned
989 void *pt = lowmem_page_address(page);
990 unsigned long pfn = page_to_pfn(page);
991 struct multicall_space mcs = __xen_mc_entry(0);
997 * We need to hold the pagetable lock between the time
998 * we make the pagetable RO and when we actually pin
999 * it. If we don't, then other users may come in and
1000 * attempt to update the pagetable by writing it,
1001 * which will fail because the memory is RO but not
1002 * pinned, so Xen won't do the trap'n'emulate.
1004 * If we're using split pte locks, we can't hold the
1005 * entire pagetable's worth of locks during the
1006 * traverse, because we may wrap the preempt count (8
1007 * bits). The solution is to mark RO and pin each PTE
1008 * page while holding the lock. This means the number
1009 * of locks we end up holding is never more than a
1010 * batch size (~32 entries, at present).
1012 * If we're not using split pte locks, we needn't pin
1013 * the PTE pages independently, because we're
1014 * protected by the overall pagetable lock.
1017 if (level == PT_PTE)
1018 ptl = xen_pte_lock(page, mm);
1020 MULTI_update_va_mapping(mcs.mc, (unsigned long)pt,
1021 pfn_pte(pfn, PAGE_KERNEL_RO),
1022 level == PT_PGD ? UVMF_TLB_FLUSH : 0);
1025 xen_do_pin(MMUEXT_PIN_L1_TABLE, pfn);
1027 /* Queue a deferred unlock for when this batch
1029 xen_mc_callback(xen_pte_unlock, ptl);
1036 /* This is called just after a mm has been created, but it has not
1037 been used yet. We need to make sure that its pagetable is all
1038 read-only, and can be pinned. */
1039 static void __xen_pgd_pin(struct mm_struct *mm, pgd_t *pgd)
1043 if (__xen_pgd_walk(mm, pgd, xen_pin_page, USER_LIMIT)) {
1044 /* re-enable interrupts for flushing */
1047 kmap_flush_unused();
1052 #ifdef CONFIG_X86_64
1054 pgd_t *user_pgd = xen_get_user_pgd(pgd);
1056 xen_do_pin(MMUEXT_PIN_L4_TABLE, PFN_DOWN(__pa(pgd)));
1059 xen_pin_page(mm, virt_to_page(user_pgd), PT_PGD);
1060 xen_do_pin(MMUEXT_PIN_L4_TABLE,
1061 PFN_DOWN(__pa(user_pgd)));
1064 #else /* CONFIG_X86_32 */
1065 #ifdef CONFIG_X86_PAE
1066 /* Need to make sure unshared kernel PMD is pinnable */
1067 xen_pin_page(mm, pgd_page(pgd[pgd_index(TASK_SIZE)]),
1070 xen_do_pin(MMUEXT_PIN_L3_TABLE, PFN_DOWN(__pa(pgd)));
1071 #endif /* CONFIG_X86_64 */
1075 static void xen_pgd_pin(struct mm_struct *mm)
1077 __xen_pgd_pin(mm, mm->pgd);
1081 * On save, we need to pin all pagetables to make sure they get their
1082 * mfns turned into pfns. Search the list for any unpinned pgds and pin
1083 * them (unpinned pgds are not currently in use, probably because the
1084 * process is under construction or destruction).
1086 * Expected to be called in stop_machine() ("equivalent to taking
1087 * every spinlock in the system"), so the locking doesn't really
1088 * matter all that much.
1090 void xen_mm_pin_all(void)
1092 unsigned long flags;
1095 spin_lock_irqsave(&pgd_lock, flags);
1097 list_for_each_entry(page, &pgd_list, lru) {
1098 if (!PagePinned(page)) {
1099 __xen_pgd_pin(&init_mm, (pgd_t *)page_address(page));
1100 SetPageSavePinned(page);
1104 spin_unlock_irqrestore(&pgd_lock, flags);
1108 * The init_mm pagetable is really pinned as soon as its created, but
1109 * that's before we have page structures to store the bits. So do all
1110 * the book-keeping now.
1112 static __init int xen_mark_pinned(struct mm_struct *mm, struct page *page,
1113 enum pt_level level)
1115 SetPagePinned(page);
1119 static void __init xen_mark_init_mm_pinned(void)
1121 xen_pgd_walk(&init_mm, xen_mark_pinned, FIXADDR_TOP);
1124 static int xen_unpin_page(struct mm_struct *mm, struct page *page,
1125 enum pt_level level)
1127 unsigned pgfl = TestClearPagePinned(page);
1129 if (pgfl && !PageHighMem(page)) {
1130 void *pt = lowmem_page_address(page);
1131 unsigned long pfn = page_to_pfn(page);
1132 spinlock_t *ptl = NULL;
1133 struct multicall_space mcs;
1136 * Do the converse to pin_page. If we're using split
1137 * pte locks, we must be holding the lock for while
1138 * the pte page is unpinned but still RO to prevent
1139 * concurrent updates from seeing it in this
1140 * partially-pinned state.
1142 if (level == PT_PTE) {
1143 ptl = xen_pte_lock(page, mm);
1146 xen_do_pin(MMUEXT_UNPIN_TABLE, pfn);
1149 mcs = __xen_mc_entry(0);
1151 MULTI_update_va_mapping(mcs.mc, (unsigned long)pt,
1152 pfn_pte(pfn, PAGE_KERNEL),
1153 level == PT_PGD ? UVMF_TLB_FLUSH : 0);
1156 /* unlock when batch completed */
1157 xen_mc_callback(xen_pte_unlock, ptl);
1161 return 0; /* never need to flush on unpin */
1164 /* Release a pagetables pages back as normal RW */
1165 static void __xen_pgd_unpin(struct mm_struct *mm, pgd_t *pgd)
1169 xen_do_pin(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1171 #ifdef CONFIG_X86_64
1173 pgd_t *user_pgd = xen_get_user_pgd(pgd);
1176 xen_do_pin(MMUEXT_UNPIN_TABLE,
1177 PFN_DOWN(__pa(user_pgd)));
1178 xen_unpin_page(mm, virt_to_page(user_pgd), PT_PGD);
1183 #ifdef CONFIG_X86_PAE
1184 /* Need to make sure unshared kernel PMD is unpinned */
1185 xen_unpin_page(mm, pgd_page(pgd[pgd_index(TASK_SIZE)]),
1189 __xen_pgd_walk(mm, pgd, xen_unpin_page, USER_LIMIT);
1194 static void xen_pgd_unpin(struct mm_struct *mm)
1196 __xen_pgd_unpin(mm, mm->pgd);
1200 * On resume, undo any pinning done at save, so that the rest of the
1201 * kernel doesn't see any unexpected pinned pagetables.
1203 void xen_mm_unpin_all(void)
1205 unsigned long flags;
1208 spin_lock_irqsave(&pgd_lock, flags);
1210 list_for_each_entry(page, &pgd_list, lru) {
1211 if (PageSavePinned(page)) {
1212 BUG_ON(!PagePinned(page));
1213 __xen_pgd_unpin(&init_mm, (pgd_t *)page_address(page));
1214 ClearPageSavePinned(page);
1218 spin_unlock_irqrestore(&pgd_lock, flags);
1221 void xen_activate_mm(struct mm_struct *prev, struct mm_struct *next)
1223 spin_lock(&next->page_table_lock);
1225 spin_unlock(&next->page_table_lock);
1228 void xen_dup_mmap(struct mm_struct *oldmm, struct mm_struct *mm)
1230 spin_lock(&mm->page_table_lock);
1232 spin_unlock(&mm->page_table_lock);
1237 /* Another cpu may still have their %cr3 pointing at the pagetable, so
1238 we need to repoint it somewhere else before we can unpin it. */
1239 static void drop_other_mm_ref(void *info)
1241 struct mm_struct *mm = info;
1242 struct mm_struct *active_mm;
1244 active_mm = percpu_read(cpu_tlbstate.active_mm);
1246 if (active_mm == mm)
1247 leave_mm(smp_processor_id());
1249 /* If this cpu still has a stale cr3 reference, then make sure
1250 it has been flushed. */
1251 if (percpu_read(xen_current_cr3) == __pa(mm->pgd))
1252 load_cr3(swapper_pg_dir);
1255 static void xen_drop_mm_ref(struct mm_struct *mm)
1260 if (current->active_mm == mm) {
1261 if (current->mm == mm)
1262 load_cr3(swapper_pg_dir);
1264 leave_mm(smp_processor_id());
1267 /* Get the "official" set of cpus referring to our pagetable. */
1268 if (!alloc_cpumask_var(&mask, GFP_ATOMIC)) {
1269 for_each_online_cpu(cpu) {
1270 if (!cpumask_test_cpu(cpu, mm_cpumask(mm))
1271 && per_cpu(xen_current_cr3, cpu) != __pa(mm->pgd))
1273 smp_call_function_single(cpu, drop_other_mm_ref, mm, 1);
1277 cpumask_copy(mask, mm_cpumask(mm));
1279 /* It's possible that a vcpu may have a stale reference to our
1280 cr3, because its in lazy mode, and it hasn't yet flushed
1281 its set of pending hypercalls yet. In this case, we can
1282 look at its actual current cr3 value, and force it to flush
1284 for_each_online_cpu(cpu) {
1285 if (per_cpu(xen_current_cr3, cpu) == __pa(mm->pgd))
1286 cpumask_set_cpu(cpu, mask);
1289 if (!cpumask_empty(mask))
1290 smp_call_function_many(mask, drop_other_mm_ref, mm, 1);
1291 free_cpumask_var(mask);
1294 static void xen_drop_mm_ref(struct mm_struct *mm)
1296 if (current->active_mm == mm)
1297 load_cr3(swapper_pg_dir);
1302 * While a process runs, Xen pins its pagetables, which means that the
1303 * hypervisor forces it to be read-only, and it controls all updates
1304 * to it. This means that all pagetable updates have to go via the
1305 * hypervisor, which is moderately expensive.
1307 * Since we're pulling the pagetable down, we switch to use init_mm,
1308 * unpin old process pagetable and mark it all read-write, which
1309 * allows further operations on it to be simple memory accesses.
1311 * The only subtle point is that another CPU may be still using the
1312 * pagetable because of lazy tlb flushing. This means we need need to
1313 * switch all CPUs off this pagetable before we can unpin it.
1315 void xen_exit_mmap(struct mm_struct *mm)
1317 get_cpu(); /* make sure we don't move around */
1318 xen_drop_mm_ref(mm);
1321 spin_lock(&mm->page_table_lock);
1323 /* pgd may not be pinned in the error exit path of execve */
1324 if (xen_page_pinned(mm->pgd))
1327 spin_unlock(&mm->page_table_lock);
1330 static __init void xen_pagetable_setup_start(pgd_t *base)
1334 static void xen_post_allocator_init(void);
1336 static __init void xen_pagetable_setup_done(pgd_t *base)
1338 xen_setup_shared_info();
1339 xen_post_allocator_init();
1342 static void xen_write_cr2(unsigned long cr2)
1344 percpu_read(xen_vcpu)->arch.cr2 = cr2;
1347 static unsigned long xen_read_cr2(void)
1349 return percpu_read(xen_vcpu)->arch.cr2;
1352 unsigned long xen_read_cr2_direct(void)
1354 return percpu_read(xen_vcpu_info.arch.cr2);
1357 static void xen_flush_tlb(void)
1359 struct mmuext_op *op;
1360 struct multicall_space mcs;
1364 mcs = xen_mc_entry(sizeof(*op));
1367 op->cmd = MMUEXT_TLB_FLUSH_LOCAL;
1368 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
1370 xen_mc_issue(PARAVIRT_LAZY_MMU);
1375 static void xen_flush_tlb_single(unsigned long addr)
1377 struct mmuext_op *op;
1378 struct multicall_space mcs;
1382 mcs = xen_mc_entry(sizeof(*op));
1384 op->cmd = MMUEXT_INVLPG_LOCAL;
1385 op->arg1.linear_addr = addr & PAGE_MASK;
1386 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
1388 xen_mc_issue(PARAVIRT_LAZY_MMU);
1393 static void xen_flush_tlb_others(const struct cpumask *cpus,
1394 struct mm_struct *mm, unsigned long va)
1397 struct mmuext_op op;
1398 DECLARE_BITMAP(mask, NR_CPUS);
1400 struct multicall_space mcs;
1402 if (cpumask_empty(cpus))
1403 return; /* nothing to do */
1405 mcs = xen_mc_entry(sizeof(*args));
1407 args->op.arg2.vcpumask = to_cpumask(args->mask);
1409 /* Remove us, and any offline CPUS. */
1410 cpumask_and(to_cpumask(args->mask), cpus, cpu_online_mask);
1411 cpumask_clear_cpu(smp_processor_id(), to_cpumask(args->mask));
1413 if (va == TLB_FLUSH_ALL) {
1414 args->op.cmd = MMUEXT_TLB_FLUSH_MULTI;
1416 args->op.cmd = MMUEXT_INVLPG_MULTI;
1417 args->op.arg1.linear_addr = va;
1420 MULTI_mmuext_op(mcs.mc, &args->op, 1, NULL, DOMID_SELF);
1422 xen_mc_issue(PARAVIRT_LAZY_MMU);
1425 static unsigned long xen_read_cr3(void)
1427 return percpu_read(xen_cr3);
1430 static void set_current_cr3(void *v)
1432 percpu_write(xen_current_cr3, (unsigned long)v);
1435 static void __xen_write_cr3(bool kernel, unsigned long cr3)
1437 struct mmuext_op *op;
1438 struct multicall_space mcs;
1442 mfn = pfn_to_mfn(PFN_DOWN(cr3));
1446 WARN_ON(mfn == 0 && kernel);
1448 mcs = __xen_mc_entry(sizeof(*op));
1451 op->cmd = kernel ? MMUEXT_NEW_BASEPTR : MMUEXT_NEW_USER_BASEPTR;
1454 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
1457 percpu_write(xen_cr3, cr3);
1459 /* Update xen_current_cr3 once the batch has actually
1461 xen_mc_callback(set_current_cr3, (void *)cr3);
1465 static void xen_write_cr3(unsigned long cr3)
1467 BUG_ON(preemptible());
1469 xen_mc_batch(); /* disables interrupts */
1471 /* Update while interrupts are disabled, so its atomic with
1473 percpu_write(xen_cr3, cr3);
1475 __xen_write_cr3(true, cr3);
1477 #ifdef CONFIG_X86_64
1479 pgd_t *user_pgd = xen_get_user_pgd(__va(cr3));
1481 __xen_write_cr3(false, __pa(user_pgd));
1483 __xen_write_cr3(false, 0);
1487 xen_mc_issue(PARAVIRT_LAZY_CPU); /* interrupts restored */
1490 static int xen_pgd_alloc(struct mm_struct *mm)
1492 pgd_t *pgd = mm->pgd;
1495 BUG_ON(PagePinned(virt_to_page(pgd)));
1497 #ifdef CONFIG_X86_64
1499 struct page *page = virt_to_page(pgd);
1502 BUG_ON(page->private != 0);
1506 user_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
1507 page->private = (unsigned long)user_pgd;
1509 if (user_pgd != NULL) {
1510 user_pgd[pgd_index(VSYSCALL_START)] =
1511 __pgd(__pa(level3_user_vsyscall) | _PAGE_TABLE);
1515 BUG_ON(PagePinned(virt_to_page(xen_get_user_pgd(pgd))));
1522 static void xen_pgd_free(struct mm_struct *mm, pgd_t *pgd)
1524 #ifdef CONFIG_X86_64
1525 pgd_t *user_pgd = xen_get_user_pgd(pgd);
1528 free_page((unsigned long)user_pgd);
1532 #ifdef CONFIG_X86_32
1533 static __init pte_t mask_rw_pte(pte_t *ptep, pte_t pte)
1535 /* If there's an existing pte, then don't allow _PAGE_RW to be set */
1536 if (pte_val_ma(*ptep) & _PAGE_PRESENT)
1537 pte = __pte_ma(((pte_val_ma(*ptep) & _PAGE_RW) | ~_PAGE_RW) &
1543 /* Init-time set_pte while constructing initial pagetables, which
1544 doesn't allow RO pagetable pages to be remapped RW */
1545 static __init void xen_set_pte_init(pte_t *ptep, pte_t pte)
1547 pte = mask_rw_pte(ptep, pte);
1549 xen_set_pte(ptep, pte);
1553 static void pin_pagetable_pfn(unsigned cmd, unsigned long pfn)
1555 struct mmuext_op op;
1557 op.arg1.mfn = pfn_to_mfn(pfn);
1558 if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
1562 /* Early in boot, while setting up the initial pagetable, assume
1563 everything is pinned. */
1564 static __init void xen_alloc_pte_init(struct mm_struct *mm, unsigned long pfn)
1566 #ifdef CONFIG_FLATMEM
1567 BUG_ON(mem_map); /* should only be used early */
1569 make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
1570 pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
1573 /* Used for pmd and pud */
1574 static __init void xen_alloc_pmd_init(struct mm_struct *mm, unsigned long pfn)
1576 #ifdef CONFIG_FLATMEM
1577 BUG_ON(mem_map); /* should only be used early */
1579 make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
1582 /* Early release_pte assumes that all pts are pinned, since there's
1583 only init_mm and anything attached to that is pinned. */
1584 static __init void xen_release_pte_init(unsigned long pfn)
1586 pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
1587 make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
1590 static __init void xen_release_pmd_init(unsigned long pfn)
1592 make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
1595 /* This needs to make sure the new pte page is pinned iff its being
1596 attached to a pinned pagetable. */
1597 static void xen_alloc_ptpage(struct mm_struct *mm, unsigned long pfn, unsigned level)
1599 struct page *page = pfn_to_page(pfn);
1601 if (PagePinned(virt_to_page(mm->pgd))) {
1602 SetPagePinned(page);
1604 if (!PageHighMem(page)) {
1605 make_lowmem_page_readonly(__va(PFN_PHYS((unsigned long)pfn)));
1606 if (level == PT_PTE && USE_SPLIT_PTLOCKS)
1607 pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
1609 /* make sure there are no stray mappings of
1611 kmap_flush_unused();
1616 static void xen_alloc_pte(struct mm_struct *mm, unsigned long pfn)
1618 xen_alloc_ptpage(mm, pfn, PT_PTE);
1621 static void xen_alloc_pmd(struct mm_struct *mm, unsigned long pfn)
1623 xen_alloc_ptpage(mm, pfn, PT_PMD);
1626 /* This should never happen until we're OK to use struct page */
1627 static void xen_release_ptpage(unsigned long pfn, unsigned level)
1629 struct page *page = pfn_to_page(pfn);
1631 if (PagePinned(page)) {
1632 if (!PageHighMem(page)) {
1633 if (level == PT_PTE && USE_SPLIT_PTLOCKS)
1634 pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
1635 make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
1637 ClearPagePinned(page);
1641 static void xen_release_pte(unsigned long pfn)
1643 xen_release_ptpage(pfn, PT_PTE);
1646 static void xen_release_pmd(unsigned long pfn)
1648 xen_release_ptpage(pfn, PT_PMD);
1651 #if PAGETABLE_LEVELS == 4
1652 static void xen_alloc_pud(struct mm_struct *mm, unsigned long pfn)
1654 xen_alloc_ptpage(mm, pfn, PT_PUD);
1657 static void xen_release_pud(unsigned long pfn)
1659 xen_release_ptpage(pfn, PT_PUD);
1663 void __init xen_reserve_top(void)
1665 #ifdef CONFIG_X86_32
1666 unsigned long top = HYPERVISOR_VIRT_START;
1667 struct xen_platform_parameters pp;
1669 if (HYPERVISOR_xen_version(XENVER_platform_parameters, &pp) == 0)
1670 top = pp.virt_start;
1672 reserve_top_address(-top);
1673 #endif /* CONFIG_X86_32 */
1677 * Like __va(), but returns address in the kernel mapping (which is
1678 * all we have until the physical memory mapping has been set up.
1680 static void *__ka(phys_addr_t paddr)
1682 #ifdef CONFIG_X86_64
1683 return (void *)(paddr + __START_KERNEL_map);
1689 /* Convert a machine address to physical address */
1690 static unsigned long m2p(phys_addr_t maddr)
1694 maddr &= PTE_PFN_MASK;
1695 paddr = mfn_to_pfn(maddr >> PAGE_SHIFT) << PAGE_SHIFT;
1700 /* Convert a machine address to kernel virtual */
1701 static void *m2v(phys_addr_t maddr)
1703 return __ka(m2p(maddr));
1706 static void set_page_prot(void *addr, pgprot_t prot)
1708 unsigned long pfn = __pa(addr) >> PAGE_SHIFT;
1709 pte_t pte = pfn_pte(pfn, prot);
1711 if (HYPERVISOR_update_va_mapping((unsigned long)addr, pte, 0))
1715 static __init void xen_map_identity_early(pmd_t *pmd, unsigned long max_pfn)
1717 unsigned pmdidx, pteidx;
1723 for (pmdidx = 0; pmdidx < PTRS_PER_PMD && pfn < max_pfn; pmdidx++) {
1726 /* Reuse or allocate a page of ptes */
1727 if (pmd_present(pmd[pmdidx]))
1728 pte_page = m2v(pmd[pmdidx].pmd);
1730 /* Check for free pte pages */
1731 if (ident_pte == ARRAY_SIZE(level1_ident_pgt))
1734 pte_page = &level1_ident_pgt[ident_pte];
1735 ident_pte += PTRS_PER_PTE;
1737 pmd[pmdidx] = __pmd(__pa(pte_page) | _PAGE_TABLE);
1740 /* Install mappings */
1741 for (pteidx = 0; pteidx < PTRS_PER_PTE; pteidx++, pfn++) {
1744 if (pfn > max_pfn_mapped)
1745 max_pfn_mapped = pfn;
1747 if (!pte_none(pte_page[pteidx]))
1750 pte = pfn_pte(pfn, PAGE_KERNEL_EXEC);
1751 pte_page[pteidx] = pte;
1755 for (pteidx = 0; pteidx < ident_pte; pteidx += PTRS_PER_PTE)
1756 set_page_prot(&level1_ident_pgt[pteidx], PAGE_KERNEL_RO);
1758 set_page_prot(pmd, PAGE_KERNEL_RO);
1761 #ifdef CONFIG_X86_64
1762 static void convert_pfn_mfn(void *v)
1767 /* All levels are converted the same way, so just treat them
1769 for (i = 0; i < PTRS_PER_PTE; i++)
1770 pte[i] = xen_make_pte(pte[i].pte);
1774 * Set up the inital kernel pagetable.
1776 * We can construct this by grafting the Xen provided pagetable into
1777 * head_64.S's preconstructed pagetables. We copy the Xen L2's into
1778 * level2_ident_pgt, level2_kernel_pgt and level2_fixmap_pgt. This
1779 * means that only the kernel has a physical mapping to start with -
1780 * but that's enough to get __va working. We need to fill in the rest
1781 * of the physical mapping once some sort of allocator has been set
1784 __init pgd_t *xen_setup_kernel_pagetable(pgd_t *pgd,
1785 unsigned long max_pfn)
1790 /* Zap identity mapping */
1791 init_level4_pgt[0] = __pgd(0);
1793 /* Pre-constructed entries are in pfn, so convert to mfn */
1794 convert_pfn_mfn(init_level4_pgt);
1795 convert_pfn_mfn(level3_ident_pgt);
1796 convert_pfn_mfn(level3_kernel_pgt);
1798 l3 = m2v(pgd[pgd_index(__START_KERNEL_map)].pgd);
1799 l2 = m2v(l3[pud_index(__START_KERNEL_map)].pud);
1801 memcpy(level2_ident_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1802 memcpy(level2_kernel_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1804 l3 = m2v(pgd[pgd_index(__START_KERNEL_map + PMD_SIZE)].pgd);
1805 l2 = m2v(l3[pud_index(__START_KERNEL_map + PMD_SIZE)].pud);
1806 memcpy(level2_fixmap_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1808 /* Set up identity map */
1809 xen_map_identity_early(level2_ident_pgt, max_pfn);
1811 /* Make pagetable pieces RO */
1812 set_page_prot(init_level4_pgt, PAGE_KERNEL_RO);
1813 set_page_prot(level3_ident_pgt, PAGE_KERNEL_RO);
1814 set_page_prot(level3_kernel_pgt, PAGE_KERNEL_RO);
1815 set_page_prot(level3_user_vsyscall, PAGE_KERNEL_RO);
1816 set_page_prot(level2_kernel_pgt, PAGE_KERNEL_RO);
1817 set_page_prot(level2_fixmap_pgt, PAGE_KERNEL_RO);
1819 /* Pin down new L4 */
1820 pin_pagetable_pfn(MMUEXT_PIN_L4_TABLE,
1821 PFN_DOWN(__pa_symbol(init_level4_pgt)));
1823 /* Unpin Xen-provided one */
1824 pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1827 pgd = init_level4_pgt;
1830 * At this stage there can be no user pgd, and no page
1831 * structure to attach it to, so make sure we just set kernel
1835 __xen_write_cr3(true, __pa(pgd));
1836 xen_mc_issue(PARAVIRT_LAZY_CPU);
1838 reserve_early(__pa(xen_start_info->pt_base),
1839 __pa(xen_start_info->pt_base +
1840 xen_start_info->nr_pt_frames * PAGE_SIZE),
1845 #else /* !CONFIG_X86_64 */
1846 static RESERVE_BRK_ARRAY(pmd_t, level2_kernel_pgt, PTRS_PER_PMD);
1848 __init pgd_t *xen_setup_kernel_pagetable(pgd_t *pgd,
1849 unsigned long max_pfn)
1853 level2_kernel_pgt = extend_brk(sizeof(pmd_t *) * PTRS_PER_PMD, PAGE_SIZE);
1855 max_pfn_mapped = PFN_DOWN(__pa(xen_start_info->pt_base) +
1856 xen_start_info->nr_pt_frames * PAGE_SIZE +
1859 kernel_pmd = m2v(pgd[KERNEL_PGD_BOUNDARY].pgd);
1860 memcpy(level2_kernel_pgt, kernel_pmd, sizeof(pmd_t) * PTRS_PER_PMD);
1862 xen_map_identity_early(level2_kernel_pgt, max_pfn);
1864 memcpy(swapper_pg_dir, pgd, sizeof(pgd_t) * PTRS_PER_PGD);
1865 set_pgd(&swapper_pg_dir[KERNEL_PGD_BOUNDARY],
1866 __pgd(__pa(level2_kernel_pgt) | _PAGE_PRESENT));
1868 set_page_prot(level2_kernel_pgt, PAGE_KERNEL_RO);
1869 set_page_prot(swapper_pg_dir, PAGE_KERNEL_RO);
1870 set_page_prot(empty_zero_page, PAGE_KERNEL_RO);
1872 pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1874 xen_write_cr3(__pa(swapper_pg_dir));
1876 pin_pagetable_pfn(MMUEXT_PIN_L3_TABLE, PFN_DOWN(__pa(swapper_pg_dir)));
1878 reserve_early(__pa(xen_start_info->pt_base),
1879 __pa(xen_start_info->pt_base +
1880 xen_start_info->nr_pt_frames * PAGE_SIZE),
1883 return swapper_pg_dir;
1885 #endif /* CONFIG_X86_64 */
1887 static void xen_set_fixmap(unsigned idx, phys_addr_t phys, pgprot_t prot)
1891 phys >>= PAGE_SHIFT;
1894 case FIX_BTMAP_END ... FIX_BTMAP_BEGIN:
1895 #ifdef CONFIG_X86_F00F_BUG
1898 #ifdef CONFIG_X86_32
1901 # ifdef CONFIG_HIGHMEM
1902 case FIX_KMAP_BEGIN ... FIX_KMAP_END:
1905 case VSYSCALL_LAST_PAGE ... VSYSCALL_FIRST_PAGE:
1907 #ifdef CONFIG_X86_LOCAL_APIC
1908 case FIX_APIC_BASE: /* maps dummy local APIC */
1910 case FIX_TEXT_POKE0:
1911 case FIX_TEXT_POKE1:
1912 /* All local page mappings */
1913 pte = pfn_pte(phys, prot);
1916 case FIX_PARAVIRT_BOOTMAP:
1917 /* This is an MFN, but it isn't an IO mapping from the
1919 pte = mfn_pte(phys, prot);
1923 /* By default, set_fixmap is used for hardware mappings */
1924 pte = mfn_pte(phys, __pgprot(pgprot_val(prot) | _PAGE_IOMAP));
1928 __native_set_fixmap(idx, pte);
1930 #ifdef CONFIG_X86_64
1931 /* Replicate changes to map the vsyscall page into the user
1932 pagetable vsyscall mapping. */
1933 if (idx >= VSYSCALL_LAST_PAGE && idx <= VSYSCALL_FIRST_PAGE) {
1934 unsigned long vaddr = __fix_to_virt(idx);
1935 set_pte_vaddr_pud(level3_user_vsyscall, vaddr, pte);
1940 static __init void xen_post_allocator_init(void)
1942 pv_mmu_ops.set_pte = xen_set_pte;
1943 pv_mmu_ops.set_pmd = xen_set_pmd;
1944 pv_mmu_ops.set_pud = xen_set_pud;
1945 #if PAGETABLE_LEVELS == 4
1946 pv_mmu_ops.set_pgd = xen_set_pgd;
1949 /* This will work as long as patching hasn't happened yet
1950 (which it hasn't) */
1951 pv_mmu_ops.alloc_pte = xen_alloc_pte;
1952 pv_mmu_ops.alloc_pmd = xen_alloc_pmd;
1953 pv_mmu_ops.release_pte = xen_release_pte;
1954 pv_mmu_ops.release_pmd = xen_release_pmd;
1955 #if PAGETABLE_LEVELS == 4
1956 pv_mmu_ops.alloc_pud = xen_alloc_pud;
1957 pv_mmu_ops.release_pud = xen_release_pud;
1960 #ifdef CONFIG_X86_64
1961 SetPagePinned(virt_to_page(level3_user_vsyscall));
1963 xen_mark_init_mm_pinned();
1966 static void xen_leave_lazy_mmu(void)
1970 paravirt_leave_lazy_mmu();
1974 static const struct pv_mmu_ops xen_mmu_ops __initdata = {
1975 .read_cr2 = xen_read_cr2,
1976 .write_cr2 = xen_write_cr2,
1978 .read_cr3 = xen_read_cr3,
1979 .write_cr3 = xen_write_cr3,
1981 .flush_tlb_user = xen_flush_tlb,
1982 .flush_tlb_kernel = xen_flush_tlb,
1983 .flush_tlb_single = xen_flush_tlb_single,
1984 .flush_tlb_others = xen_flush_tlb_others,
1986 .pte_update = paravirt_nop,
1987 .pte_update_defer = paravirt_nop,
1989 .pgd_alloc = xen_pgd_alloc,
1990 .pgd_free = xen_pgd_free,
1992 .alloc_pte = xen_alloc_pte_init,
1993 .release_pte = xen_release_pte_init,
1994 .alloc_pmd = xen_alloc_pmd_init,
1995 .alloc_pmd_clone = paravirt_nop,
1996 .release_pmd = xen_release_pmd_init,
1998 #ifdef CONFIG_X86_64
1999 .set_pte = xen_set_pte,
2001 .set_pte = xen_set_pte_init,
2003 .set_pte_at = xen_set_pte_at,
2004 .set_pmd = xen_set_pmd_hyper,
2006 .ptep_modify_prot_start = __ptep_modify_prot_start,
2007 .ptep_modify_prot_commit = __ptep_modify_prot_commit,
2009 .pte_val = PV_CALLEE_SAVE(xen_pte_val),
2010 .pgd_val = PV_CALLEE_SAVE(xen_pgd_val),
2012 .make_pte = PV_CALLEE_SAVE(xen_make_pte),
2013 .make_pgd = PV_CALLEE_SAVE(xen_make_pgd),
2015 #ifdef CONFIG_X86_PAE
2016 .set_pte_atomic = xen_set_pte_atomic,
2017 .pte_clear = xen_pte_clear,
2018 .pmd_clear = xen_pmd_clear,
2019 #endif /* CONFIG_X86_PAE */
2020 .set_pud = xen_set_pud_hyper,
2022 .make_pmd = PV_CALLEE_SAVE(xen_make_pmd),
2023 .pmd_val = PV_CALLEE_SAVE(xen_pmd_val),
2025 #if PAGETABLE_LEVELS == 4
2026 .pud_val = PV_CALLEE_SAVE(xen_pud_val),
2027 .make_pud = PV_CALLEE_SAVE(xen_make_pud),
2028 .set_pgd = xen_set_pgd_hyper,
2030 .alloc_pud = xen_alloc_pmd_init,
2031 .release_pud = xen_release_pmd_init,
2032 #endif /* PAGETABLE_LEVELS == 4 */
2034 .activate_mm = xen_activate_mm,
2035 .dup_mmap = xen_dup_mmap,
2036 .exit_mmap = xen_exit_mmap,
2039 .enter = paravirt_enter_lazy_mmu,
2040 .leave = xen_leave_lazy_mmu,
2043 .set_fixmap = xen_set_fixmap,
2046 void __init xen_init_mmu_ops(void)
2048 x86_init.paging.pagetable_setup_start = xen_pagetable_setup_start;
2049 x86_init.paging.pagetable_setup_done = xen_pagetable_setup_done;
2050 pv_mmu_ops = xen_mmu_ops;
2052 vmap_lazy_unmap = false;
2055 /* Protected by xen_reservation_lock. */
2056 #define MAX_CONTIG_ORDER 9 /* 2MB */
2057 static unsigned long discontig_frames[1<<MAX_CONTIG_ORDER];
2059 #define VOID_PTE (mfn_pte(0, __pgprot(0)))
2060 static void xen_zap_pfn_range(unsigned long vaddr, unsigned int order,
2061 unsigned long *in_frames,
2062 unsigned long *out_frames)
2065 struct multicall_space mcs;
2068 for (i = 0; i < (1UL<<order); i++, vaddr += PAGE_SIZE) {
2069 mcs = __xen_mc_entry(0);
2072 in_frames[i] = virt_to_mfn(vaddr);
2074 MULTI_update_va_mapping(mcs.mc, vaddr, VOID_PTE, 0);
2075 set_phys_to_machine(virt_to_pfn(vaddr), INVALID_P2M_ENTRY);
2078 out_frames[i] = virt_to_pfn(vaddr);
2084 * Update the pfn-to-mfn mappings for a virtual address range, either to
2085 * point to an array of mfns, or contiguously from a single starting
2088 static void xen_remap_exchanged_ptes(unsigned long vaddr, int order,
2089 unsigned long *mfns,
2090 unsigned long first_mfn)
2097 limit = 1u << order;
2098 for (i = 0; i < limit; i++, vaddr += PAGE_SIZE) {
2099 struct multicall_space mcs;
2102 mcs = __xen_mc_entry(0);
2106 mfn = first_mfn + i;
2108 if (i < (limit - 1))
2112 flags = UVMF_INVLPG | UVMF_ALL;
2114 flags = UVMF_TLB_FLUSH | UVMF_ALL;
2117 MULTI_update_va_mapping(mcs.mc, vaddr,
2118 mfn_pte(mfn, PAGE_KERNEL), flags);
2120 set_phys_to_machine(virt_to_pfn(vaddr), mfn);
2127 * Perform the hypercall to exchange a region of our pfns to point to
2128 * memory with the required contiguous alignment. Takes the pfns as
2129 * input, and populates mfns as output.
2131 * Returns a success code indicating whether the hypervisor was able to
2132 * satisfy the request or not.
2134 static int xen_exchange_memory(unsigned long extents_in, unsigned int order_in,
2135 unsigned long *pfns_in,
2136 unsigned long extents_out,
2137 unsigned int order_out,
2138 unsigned long *mfns_out,
2139 unsigned int address_bits)
2144 struct xen_memory_exchange exchange = {
2146 .nr_extents = extents_in,
2147 .extent_order = order_in,
2148 .extent_start = pfns_in,
2152 .nr_extents = extents_out,
2153 .extent_order = order_out,
2154 .extent_start = mfns_out,
2155 .address_bits = address_bits,
2160 BUG_ON(extents_in << order_in != extents_out << order_out);
2162 rc = HYPERVISOR_memory_op(XENMEM_exchange, &exchange);
2163 success = (exchange.nr_exchanged == extents_in);
2165 BUG_ON(!success && ((exchange.nr_exchanged != 0) || (rc == 0)));
2166 BUG_ON(success && (rc != 0));
2171 int xen_create_contiguous_region(unsigned long vstart, unsigned int order,
2172 unsigned int address_bits)
2174 unsigned long *in_frames = discontig_frames, out_frame;
2175 unsigned long flags;
2179 * Currently an auto-translated guest will not perform I/O, nor will
2180 * it require PAE page directories below 4GB. Therefore any calls to
2181 * this function are redundant and can be ignored.
2184 if (xen_feature(XENFEAT_auto_translated_physmap))
2187 if (unlikely(order > MAX_CONTIG_ORDER))
2190 memset((void *) vstart, 0, PAGE_SIZE << order);
2192 spin_lock_irqsave(&xen_reservation_lock, flags);
2194 /* 1. Zap current PTEs, remembering MFNs. */
2195 xen_zap_pfn_range(vstart, order, in_frames, NULL);
2197 /* 2. Get a new contiguous memory extent. */
2198 out_frame = virt_to_pfn(vstart);
2199 success = xen_exchange_memory(1UL << order, 0, in_frames,
2200 1, order, &out_frame,
2203 /* 3. Map the new extent in place of old pages. */
2205 xen_remap_exchanged_ptes(vstart, order, NULL, out_frame);
2207 xen_remap_exchanged_ptes(vstart, order, in_frames, 0);
2209 spin_unlock_irqrestore(&xen_reservation_lock, flags);
2211 return success ? 0 : -ENOMEM;
2213 EXPORT_SYMBOL_GPL(xen_create_contiguous_region);
2215 void xen_destroy_contiguous_region(unsigned long vstart, unsigned int order)
2217 unsigned long *out_frames = discontig_frames, in_frame;
2218 unsigned long flags;
2221 if (xen_feature(XENFEAT_auto_translated_physmap))
2224 if (unlikely(order > MAX_CONTIG_ORDER))
2227 memset((void *) vstart, 0, PAGE_SIZE << order);
2229 spin_lock_irqsave(&xen_reservation_lock, flags);
2231 /* 1. Find start MFN of contiguous extent. */
2232 in_frame = virt_to_mfn(vstart);
2234 /* 2. Zap current PTEs. */
2235 xen_zap_pfn_range(vstart, order, NULL, out_frames);
2237 /* 3. Do the exchange for non-contiguous MFNs. */
2238 success = xen_exchange_memory(1, order, &in_frame, 1UL << order,
2241 /* 4. Map new pages in place of old pages. */
2243 xen_remap_exchanged_ptes(vstart, order, out_frames, 0);
2245 xen_remap_exchanged_ptes(vstart, order, NULL, in_frame);
2247 spin_unlock_irqrestore(&xen_reservation_lock, flags);
2249 EXPORT_SYMBOL_GPL(xen_destroy_contiguous_region);
2251 #ifdef CONFIG_XEN_PVHVM
2252 static void xen_hvm_exit_mmap(struct mm_struct *mm)
2254 struct xen_hvm_pagetable_dying a;
2257 a.domid = DOMID_SELF;
2258 a.gpa = __pa(mm->pgd);
2259 rc = HYPERVISOR_hvm_op(HVMOP_pagetable_dying, &a);
2260 WARN_ON_ONCE(rc < 0);
2263 static int is_pagetable_dying_supported(void)
2265 struct xen_hvm_pagetable_dying a;
2268 a.domid = DOMID_SELF;
2270 rc = HYPERVISOR_hvm_op(HVMOP_pagetable_dying, &a);
2272 printk(KERN_DEBUG "HVMOP_pagetable_dying not supported\n");
2278 void __init xen_hvm_init_mmu_ops(void)
2280 if (is_pagetable_dying_supported())
2281 pv_mmu_ops.exit_mmap = xen_hvm_exit_mmap;
2285 #ifdef CONFIG_XEN_DEBUG_FS
2287 static struct dentry *d_mmu_debug;
2289 static int __init xen_mmu_debugfs(void)
2291 struct dentry *d_xen = xen_init_debugfs();
2296 d_mmu_debug = debugfs_create_dir("mmu", d_xen);
2298 debugfs_create_u8("zero_stats", 0644, d_mmu_debug, &zero_stats);
2300 debugfs_create_u32("pgd_update", 0444, d_mmu_debug, &mmu_stats.pgd_update);
2301 debugfs_create_u32("pgd_update_pinned", 0444, d_mmu_debug,
2302 &mmu_stats.pgd_update_pinned);
2303 debugfs_create_u32("pgd_update_batched", 0444, d_mmu_debug,
2304 &mmu_stats.pgd_update_pinned);
2306 debugfs_create_u32("pud_update", 0444, d_mmu_debug, &mmu_stats.pud_update);
2307 debugfs_create_u32("pud_update_pinned", 0444, d_mmu_debug,
2308 &mmu_stats.pud_update_pinned);
2309 debugfs_create_u32("pud_update_batched", 0444, d_mmu_debug,
2310 &mmu_stats.pud_update_pinned);
2312 debugfs_create_u32("pmd_update", 0444, d_mmu_debug, &mmu_stats.pmd_update);
2313 debugfs_create_u32("pmd_update_pinned", 0444, d_mmu_debug,
2314 &mmu_stats.pmd_update_pinned);
2315 debugfs_create_u32("pmd_update_batched", 0444, d_mmu_debug,
2316 &mmu_stats.pmd_update_pinned);
2318 debugfs_create_u32("pte_update", 0444, d_mmu_debug, &mmu_stats.pte_update);
2319 // debugfs_create_u32("pte_update_pinned", 0444, d_mmu_debug,
2320 // &mmu_stats.pte_update_pinned);
2321 debugfs_create_u32("pte_update_batched", 0444, d_mmu_debug,
2322 &mmu_stats.pte_update_pinned);
2324 debugfs_create_u32("mmu_update", 0444, d_mmu_debug, &mmu_stats.mmu_update);
2325 debugfs_create_u32("mmu_update_extended", 0444, d_mmu_debug,
2326 &mmu_stats.mmu_update_extended);
2327 xen_debugfs_create_u32_array("mmu_update_histo", 0444, d_mmu_debug,
2328 mmu_stats.mmu_update_histo, 20);
2330 debugfs_create_u32("set_pte_at", 0444, d_mmu_debug, &mmu_stats.set_pte_at);
2331 debugfs_create_u32("set_pte_at_batched", 0444, d_mmu_debug,
2332 &mmu_stats.set_pte_at_batched);
2333 debugfs_create_u32("set_pte_at_current", 0444, d_mmu_debug,
2334 &mmu_stats.set_pte_at_current);
2335 debugfs_create_u32("set_pte_at_kernel", 0444, d_mmu_debug,
2336 &mmu_stats.set_pte_at_kernel);
2338 debugfs_create_u32("prot_commit", 0444, d_mmu_debug, &mmu_stats.prot_commit);
2339 debugfs_create_u32("prot_commit_batched", 0444, d_mmu_debug,
2340 &mmu_stats.prot_commit_batched);
2344 fs_initcall(xen_mmu_debugfs);
2346 #endif /* CONFIG_XEN_DEBUG_FS */