4 * This file contains the various mmu fetch and update operations.
5 * The most important job they must perform is the mapping between the
6 * domain's pfn and the overall machine mfns.
8 * Xen allows guests to directly update the pagetable, in a controlled
9 * fashion. In other words, the guest modifies the same pagetable
10 * that the CPU actually uses, which eliminates the overhead of having
11 * a separate shadow pagetable.
13 * In order to allow this, it falls on the guest domain to map its
14 * notion of a "physical" pfn - which is just a domain-local linear
15 * address - into a real "machine address" which the CPU's MMU can
18 * A pgd_t/pmd_t/pte_t will typically contain an mfn, and so can be
19 * inserted directly into the pagetable. When creating a new
20 * pte/pmd/pgd, it converts the passed pfn into an mfn. Conversely,
21 * when reading the content back with __(pgd|pmd|pte)_val, it converts
22 * the mfn back into a pfn.
24 * The other constraint is that all pages which make up a pagetable
25 * must be mapped read-only in the guest. This prevents uncontrolled
26 * guest updates to the pagetable. Xen strictly enforces this, and
27 * will disallow any pagetable update which will end up mapping a
28 * pagetable page RW, and will disallow using any writable page as a
31 * Naively, when loading %cr3 with the base of a new pagetable, Xen
32 * would need to validate the whole pagetable before going on.
33 * Naturally, this is quite slow. The solution is to "pin" a
34 * pagetable, which enforces all the constraints on the pagetable even
35 * when it is not actively in use. This menas that Xen can be assured
36 * that it is still valid when you do load it into %cr3, and doesn't
37 * need to revalidate it.
39 * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
41 #include <linux/sched.h>
42 #include <linux/highmem.h>
43 #include <linux/bug.h>
45 #include <asm/pgtable.h>
46 #include <asm/tlbflush.h>
47 #include <asm/mmu_context.h>
48 #include <asm/paravirt.h>
50 #include <asm/xen/hypercall.h>
51 #include <asm/xen/hypervisor.h>
54 #include <xen/interface/xen.h>
56 #include "multicalls.h"
59 xmaddr_t arbitrary_virt_to_machine(unsigned long address)
62 pte_t *pte = lookup_address(address, &level);
63 unsigned offset = address & PAGE_MASK;
67 return XMADDR((pte_mfn(*pte) << PAGE_SHIFT) + offset);
70 void make_lowmem_page_readonly(void *vaddr)
73 unsigned long address = (unsigned long)vaddr;
76 pte = lookup_address(address, &level);
79 ptev = pte_wrprotect(*pte);
81 if (HYPERVISOR_update_va_mapping(address, ptev, 0))
85 void make_lowmem_page_readwrite(void *vaddr)
88 unsigned long address = (unsigned long)vaddr;
91 pte = lookup_address(address, &level);
94 ptev = pte_mkwrite(*pte);
96 if (HYPERVISOR_update_va_mapping(address, ptev, 0))
101 void xen_set_pmd(pmd_t *ptr, pmd_t val)
103 struct multicall_space mcs;
104 struct mmu_update *u;
108 mcs = xen_mc_entry(sizeof(*u));
110 u->ptr = virt_to_machine(ptr).maddr;
111 u->val = pmd_val_ma(val);
112 MULTI_mmu_update(mcs.mc, u, 1, NULL, DOMID_SELF);
114 xen_mc_issue(PARAVIRT_LAZY_MMU);
120 * Associate a virtual page frame with a given physical page frame
121 * and protection flags for that frame.
123 void set_pte_mfn(unsigned long vaddr, unsigned long mfn, pgprot_t flags)
130 pgd = swapper_pg_dir + pgd_index(vaddr);
131 if (pgd_none(*pgd)) {
135 pud = pud_offset(pgd, vaddr);
136 if (pud_none(*pud)) {
140 pmd = pmd_offset(pud, vaddr);
141 if (pmd_none(*pmd)) {
145 pte = pte_offset_kernel(pmd, vaddr);
146 /* <mfn,flags> stored as-is, to permit clearing entries */
147 xen_set_pte(pte, mfn_pte(mfn, flags));
150 * It's enough to flush this one mapping.
151 * (PGE mappings get flushed as well)
153 __flush_tlb_one(vaddr);
156 void xen_set_pte_at(struct mm_struct *mm, unsigned long addr,
157 pte_t *ptep, pte_t pteval)
159 /* updates to init_mm may be done without lock */
163 if (mm == current->mm || mm == &init_mm) {
164 if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU) {
165 struct multicall_space mcs;
166 mcs = xen_mc_entry(0);
168 MULTI_update_va_mapping(mcs.mc, addr, pteval, 0);
169 xen_mc_issue(PARAVIRT_LAZY_MMU);
172 if (HYPERVISOR_update_va_mapping(addr, pteval, 0) == 0)
175 xen_set_pte(ptep, pteval);
182 pteval_t xen_pte_val(pte_t pte)
184 pteval_t ret = pte.pte;
186 if (ret & _PAGE_PRESENT)
187 ret = machine_to_phys(XMADDR(ret)).paddr | _PAGE_PRESENT;
192 pgdval_t xen_pgd_val(pgd_t pgd)
194 pgdval_t ret = pgd.pgd;
195 if (ret & _PAGE_PRESENT)
196 ret = machine_to_phys(XMADDR(ret)).paddr | _PAGE_PRESENT;
200 pte_t xen_make_pte(pteval_t pte)
202 if (pte & _PAGE_PRESENT) {
203 pte = phys_to_machine(XPADDR(pte)).maddr;
204 pte &= ~(_PAGE_PCD | _PAGE_PWT);
207 return (pte_t){ .pte = pte };
210 pgd_t xen_make_pgd(pgdval_t pgd)
212 if (pgd & _PAGE_PRESENT)
213 pgd = phys_to_machine(XPADDR(pgd)).maddr;
215 return (pgd_t){ pgd };
218 pmdval_t xen_pmd_val(pmd_t pmd)
220 pmdval_t ret = native_pmd_val(pmd);
221 if (ret & _PAGE_PRESENT)
222 ret = machine_to_phys(XMADDR(ret)).paddr | _PAGE_PRESENT;
225 #ifdef CONFIG_X86_PAE
226 void xen_set_pud(pud_t *ptr, pud_t val)
228 struct multicall_space mcs;
229 struct mmu_update *u;
233 mcs = xen_mc_entry(sizeof(*u));
235 u->ptr = virt_to_machine(ptr).maddr;
236 u->val = pud_val_ma(val);
237 MULTI_mmu_update(mcs.mc, u, 1, NULL, DOMID_SELF);
239 xen_mc_issue(PARAVIRT_LAZY_MMU);
244 void xen_set_pte(pte_t *ptep, pte_t pte)
246 ptep->pte_high = pte.pte_high;
248 ptep->pte_low = pte.pte_low;
251 void xen_set_pte_atomic(pte_t *ptep, pte_t pte)
253 set_64bit((u64 *)ptep, pte_val_ma(pte));
256 void xen_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
259 smp_wmb(); /* make sure low gets written first */
263 void xen_pmd_clear(pmd_t *pmdp)
265 xen_set_pmd(pmdp, __pmd(0));
268 pmd_t xen_make_pmd(pmdval_t pmd)
270 if (pmd & _PAGE_PRESENT)
271 pmd = phys_to_machine(XPADDR(pmd)).maddr;
273 return native_make_pmd(pmd);
276 void xen_set_pte(pte_t *ptep, pte_t pte)
280 #endif /* CONFIG_X86_PAE */
283 (Yet another) pagetable walker. This one is intended for pinning a
284 pagetable. This means that it walks a pagetable and calls the
285 callback function on each page it finds making up the page table,
286 at every level. It walks the entire pagetable, but it only bothers
287 pinning pte pages which are below pte_limit. In the normal case
288 this will be TASK_SIZE, but at boot we need to pin up to
289 FIXADDR_TOP. But the important bit is that we don't pin beyond
290 there, because then we start getting into Xen's ptes.
292 static int pgd_walk(pgd_t *pgd_base, int (*func)(struct page *, enum pt_level),
295 pgd_t *pgd = pgd_base;
297 unsigned long addr = 0;
298 unsigned long pgd_next;
300 BUG_ON(limit > FIXADDR_TOP);
302 if (xen_feature(XENFEAT_auto_translated_physmap))
305 for (; addr != FIXADDR_TOP; pgd++, addr = pgd_next) {
307 unsigned long pud_limit, pud_next;
309 pgd_next = pud_limit = pgd_addr_end(addr, FIXADDR_TOP);
314 pud = pud_offset(pgd, 0);
316 if (PTRS_PER_PUD > 1) /* not folded */
317 flush |= (*func)(virt_to_page(pud), PT_PUD);
319 for (; addr != pud_limit; pud++, addr = pud_next) {
321 unsigned long pmd_limit;
323 pud_next = pud_addr_end(addr, pud_limit);
325 if (pud_next < limit)
326 pmd_limit = pud_next;
333 pmd = pmd_offset(pud, 0);
335 if (PTRS_PER_PMD > 1) /* not folded */
336 flush |= (*func)(virt_to_page(pmd), PT_PMD);
338 for (; addr != pmd_limit; pmd++) {
339 addr += (PAGE_SIZE * PTRS_PER_PTE);
340 if ((pmd_limit-1) < (addr-1)) {
348 flush |= (*func)(pmd_page(*pmd), PT_PTE);
353 flush |= (*func)(virt_to_page(pgd_base), PT_PGD);
358 static spinlock_t *lock_pte(struct page *page)
360 spinlock_t *ptl = NULL;
362 #if NR_CPUS >= CONFIG_SPLIT_PTLOCK_CPUS
363 ptl = __pte_lockptr(page);
370 static void do_unlock(void *v)
376 static void xen_do_pin(unsigned level, unsigned long pfn)
378 struct mmuext_op *op;
379 struct multicall_space mcs;
381 mcs = __xen_mc_entry(sizeof(*op));
384 op->arg1.mfn = pfn_to_mfn(pfn);
385 MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
388 static int pin_page(struct page *page, enum pt_level level)
390 unsigned pgfl = TestSetPagePinned(page);
394 flush = 0; /* already pinned */
395 else if (PageHighMem(page))
396 /* kmaps need flushing if we found an unpinned
400 void *pt = lowmem_page_address(page);
401 unsigned long pfn = page_to_pfn(page);
402 struct multicall_space mcs = __xen_mc_entry(0);
409 ptl = lock_pte(page);
411 MULTI_update_va_mapping(mcs.mc, (unsigned long)pt,
412 pfn_pte(pfn, PAGE_KERNEL_RO),
413 level == PT_PGD ? UVMF_TLB_FLUSH : 0);
416 xen_do_pin(MMUEXT_PIN_L1_TABLE, pfn);
419 /* Queue a deferred unlock for when this batch
421 xen_mc_callback(do_unlock, ptl);
428 /* This is called just after a mm has been created, but it has not
429 been used yet. We need to make sure that its pagetable is all
430 read-only, and can be pinned. */
431 void xen_pgd_pin(pgd_t *pgd)
437 if (pgd_walk(pgd, pin_page, TASK_SIZE)) {
438 /* re-enable interrupts for kmap_flush_unused */
444 #ifdef CONFIG_X86_PAE
445 level = MMUEXT_PIN_L3_TABLE;
447 level = MMUEXT_PIN_L2_TABLE;
450 xen_do_pin(level, PFN_DOWN(__pa(pgd)));
455 /* The init_mm pagetable is really pinned as soon as its created, but
456 that's before we have page structures to store the bits. So do all
457 the book-keeping now. */
458 static __init int mark_pinned(struct page *page, enum pt_level level)
464 void __init xen_mark_init_mm_pinned(void)
466 pgd_walk(init_mm.pgd, mark_pinned, FIXADDR_TOP);
469 static int unpin_page(struct page *page, enum pt_level level)
471 unsigned pgfl = TestClearPagePinned(page);
473 if (pgfl && !PageHighMem(page)) {
474 void *pt = lowmem_page_address(page);
475 unsigned long pfn = page_to_pfn(page);
476 spinlock_t *ptl = NULL;
477 struct multicall_space mcs;
479 if (level == PT_PTE) {
480 ptl = lock_pte(page);
482 xen_do_pin(MMUEXT_UNPIN_TABLE, pfn);
485 mcs = __xen_mc_entry(0);
487 MULTI_update_va_mapping(mcs.mc, (unsigned long)pt,
488 pfn_pte(pfn, PAGE_KERNEL),
489 level == PT_PGD ? UVMF_TLB_FLUSH : 0);
492 /* unlock when batch completed */
493 xen_mc_callback(do_unlock, ptl);
497 return 0; /* never need to flush on unpin */
500 /* Release a pagetables pages back as normal RW */
501 static void xen_pgd_unpin(pgd_t *pgd)
505 xen_do_pin(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
507 pgd_walk(pgd, unpin_page, TASK_SIZE);
512 void xen_activate_mm(struct mm_struct *prev, struct mm_struct *next)
514 spin_lock(&next->page_table_lock);
515 xen_pgd_pin(next->pgd);
516 spin_unlock(&next->page_table_lock);
519 void xen_dup_mmap(struct mm_struct *oldmm, struct mm_struct *mm)
521 spin_lock(&mm->page_table_lock);
522 xen_pgd_pin(mm->pgd);
523 spin_unlock(&mm->page_table_lock);
528 /* Another cpu may still have their %cr3 pointing at the pagetable, so
529 we need to repoint it somewhere else before we can unpin it. */
530 static void drop_other_mm_ref(void *info)
532 struct mm_struct *mm = info;
534 if (__get_cpu_var(cpu_tlbstate).active_mm == mm)
535 leave_mm(smp_processor_id());
537 /* If this cpu still has a stale cr3 reference, then make sure
538 it has been flushed. */
539 if (x86_read_percpu(xen_current_cr3) == __pa(mm->pgd)) {
540 load_cr3(swapper_pg_dir);
541 arch_flush_lazy_cpu_mode();
545 static void drop_mm_ref(struct mm_struct *mm)
550 if (current->active_mm == mm) {
551 if (current->mm == mm)
552 load_cr3(swapper_pg_dir);
554 leave_mm(smp_processor_id());
555 arch_flush_lazy_cpu_mode();
558 /* Get the "official" set of cpus referring to our pagetable. */
559 mask = mm->cpu_vm_mask;
561 /* It's possible that a vcpu may have a stale reference to our
562 cr3, because its in lazy mode, and it hasn't yet flushed
563 its set of pending hypercalls yet. In this case, we can
564 look at its actual current cr3 value, and force it to flush
566 for_each_online_cpu(cpu) {
567 if (per_cpu(xen_current_cr3, cpu) == __pa(mm->pgd))
571 if (!cpus_empty(mask))
572 xen_smp_call_function_mask(mask, drop_other_mm_ref, mm, 1);
575 static void drop_mm_ref(struct mm_struct *mm)
577 if (current->active_mm == mm)
578 load_cr3(swapper_pg_dir);
583 * While a process runs, Xen pins its pagetables, which means that the
584 * hypervisor forces it to be read-only, and it controls all updates
585 * to it. This means that all pagetable updates have to go via the
586 * hypervisor, which is moderately expensive.
588 * Since we're pulling the pagetable down, we switch to use init_mm,
589 * unpin old process pagetable and mark it all read-write, which
590 * allows further operations on it to be simple memory accesses.
592 * The only subtle point is that another CPU may be still using the
593 * pagetable because of lazy tlb flushing. This means we need need to
594 * switch all CPUs off this pagetable before we can unpin it.
596 void xen_exit_mmap(struct mm_struct *mm)
598 get_cpu(); /* make sure we don't move around */
602 spin_lock(&mm->page_table_lock);
604 /* pgd may not be pinned in the error exit path of execve */
605 if (PagePinned(virt_to_page(mm->pgd)))
606 xen_pgd_unpin(mm->pgd);
608 spin_unlock(&mm->page_table_lock);