xref: /openbmc/linux/arch/x86/xen/mmu.c (revision 4800cd83)
1 /*
2  * Xen mmu operations
3  *
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.
7  *
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.
12  *
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
16  * use.
17  *
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.
23  *
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
29  * pagetable.
30  *
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.
38  *
39  * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007
40  */
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>
49 
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>
56 #include <asm/e820.h>
57 #include <asm/linkage.h>
58 #include <asm/page.h>
59 #include <asm/init.h>
60 #include <asm/pat.h>
61 
62 #include <asm/xen/hypercall.h>
63 #include <asm/xen/hypervisor.h>
64 
65 #include <xen/xen.h>
66 #include <xen/page.h>
67 #include <xen/interface/xen.h>
68 #include <xen/interface/hvm/hvm_op.h>
69 #include <xen/interface/version.h>
70 #include <xen/interface/memory.h>
71 #include <xen/hvc-console.h>
72 
73 #include "multicalls.h"
74 #include "mmu.h"
75 #include "debugfs.h"
76 
77 #define MMU_UPDATE_HISTO	30
78 
79 /*
80  * Protects atomic reservation decrease/increase against concurrent increases.
81  * Also protects non-atomic updates of current_pages and driver_pages, and
82  * balloon lists.
83  */
84 DEFINE_SPINLOCK(xen_reservation_lock);
85 
86 #ifdef CONFIG_XEN_DEBUG_FS
87 
88 static struct {
89 	u32 pgd_update;
90 	u32 pgd_update_pinned;
91 	u32 pgd_update_batched;
92 
93 	u32 pud_update;
94 	u32 pud_update_pinned;
95 	u32 pud_update_batched;
96 
97 	u32 pmd_update;
98 	u32 pmd_update_pinned;
99 	u32 pmd_update_batched;
100 
101 	u32 pte_update;
102 	u32 pte_update_pinned;
103 	u32 pte_update_batched;
104 
105 	u32 mmu_update;
106 	u32 mmu_update_extended;
107 	u32 mmu_update_histo[MMU_UPDATE_HISTO];
108 
109 	u32 prot_commit;
110 	u32 prot_commit_batched;
111 
112 	u32 set_pte_at;
113 	u32 set_pte_at_batched;
114 	u32 set_pte_at_pinned;
115 	u32 set_pte_at_current;
116 	u32 set_pte_at_kernel;
117 } mmu_stats;
118 
119 static u8 zero_stats;
120 
121 static inline void check_zero(void)
122 {
123 	if (unlikely(zero_stats)) {
124 		memset(&mmu_stats, 0, sizeof(mmu_stats));
125 		zero_stats = 0;
126 	}
127 }
128 
129 #define ADD_STATS(elem, val)			\
130 	do { check_zero(); mmu_stats.elem += (val); } while(0)
131 
132 #else  /* !CONFIG_XEN_DEBUG_FS */
133 
134 #define ADD_STATS(elem, val)	do { (void)(val); } while(0)
135 
136 #endif /* CONFIG_XEN_DEBUG_FS */
137 
138 
139 /*
140  * Identity map, in addition to plain kernel map.  This needs to be
141  * large enough to allocate page table pages to allocate the rest.
142  * Each page can map 2MB.
143  */
144 #define LEVEL1_IDENT_ENTRIES	(PTRS_PER_PTE * 4)
145 static RESERVE_BRK_ARRAY(pte_t, level1_ident_pgt, LEVEL1_IDENT_ENTRIES);
146 
147 #ifdef CONFIG_X86_64
148 /* l3 pud for userspace vsyscall mapping */
149 static pud_t level3_user_vsyscall[PTRS_PER_PUD] __page_aligned_bss;
150 #endif /* CONFIG_X86_64 */
151 
152 /*
153  * Note about cr3 (pagetable base) values:
154  *
155  * xen_cr3 contains the current logical cr3 value; it contains the
156  * last set cr3.  This may not be the current effective cr3, because
157  * its update may be being lazily deferred.  However, a vcpu looking
158  * at its own cr3 can use this value knowing that it everything will
159  * be self-consistent.
160  *
161  * xen_current_cr3 contains the actual vcpu cr3; it is set once the
162  * hypercall to set the vcpu cr3 is complete (so it may be a little
163  * out of date, but it will never be set early).  If one vcpu is
164  * looking at another vcpu's cr3 value, it should use this variable.
165  */
166 DEFINE_PER_CPU(unsigned long, xen_cr3);	 /* cr3 stored as physaddr */
167 DEFINE_PER_CPU(unsigned long, xen_current_cr3);	 /* actual vcpu cr3 */
168 
169 
170 /*
171  * Just beyond the highest usermode address.  STACK_TOP_MAX has a
172  * redzone above it, so round it up to a PGD boundary.
173  */
174 #define USER_LIMIT	((STACK_TOP_MAX + PGDIR_SIZE - 1) & PGDIR_MASK)
175 
176 unsigned long arbitrary_virt_to_mfn(void *vaddr)
177 {
178 	xmaddr_t maddr = arbitrary_virt_to_machine(vaddr);
179 
180 	return PFN_DOWN(maddr.maddr);
181 }
182 
183 xmaddr_t arbitrary_virt_to_machine(void *vaddr)
184 {
185 	unsigned long address = (unsigned long)vaddr;
186 	unsigned int level;
187 	pte_t *pte;
188 	unsigned offset;
189 
190 	/*
191 	 * if the PFN is in the linear mapped vaddr range, we can just use
192 	 * the (quick) virt_to_machine() p2m lookup
193 	 */
194 	if (virt_addr_valid(vaddr))
195 		return virt_to_machine(vaddr);
196 
197 	/* otherwise we have to do a (slower) full page-table walk */
198 
199 	pte = lookup_address(address, &level);
200 	BUG_ON(pte == NULL);
201 	offset = address & ~PAGE_MASK;
202 	return XMADDR(((phys_addr_t)pte_mfn(*pte) << PAGE_SHIFT) + offset);
203 }
204 EXPORT_SYMBOL_GPL(arbitrary_virt_to_machine);
205 
206 void make_lowmem_page_readonly(void *vaddr)
207 {
208 	pte_t *pte, ptev;
209 	unsigned long address = (unsigned long)vaddr;
210 	unsigned int level;
211 
212 	pte = lookup_address(address, &level);
213 	if (pte == NULL)
214 		return;		/* vaddr missing */
215 
216 	ptev = pte_wrprotect(*pte);
217 
218 	if (HYPERVISOR_update_va_mapping(address, ptev, 0))
219 		BUG();
220 }
221 
222 void make_lowmem_page_readwrite(void *vaddr)
223 {
224 	pte_t *pte, ptev;
225 	unsigned long address = (unsigned long)vaddr;
226 	unsigned int level;
227 
228 	pte = lookup_address(address, &level);
229 	if (pte == NULL)
230 		return;		/* vaddr missing */
231 
232 	ptev = pte_mkwrite(*pte);
233 
234 	if (HYPERVISOR_update_va_mapping(address, ptev, 0))
235 		BUG();
236 }
237 
238 
239 static bool xen_page_pinned(void *ptr)
240 {
241 	struct page *page = virt_to_page(ptr);
242 
243 	return PagePinned(page);
244 }
245 
246 static bool xen_iomap_pte(pte_t pte)
247 {
248 	return pte_flags(pte) & _PAGE_IOMAP;
249 }
250 
251 void xen_set_domain_pte(pte_t *ptep, pte_t pteval, unsigned domid)
252 {
253 	struct multicall_space mcs;
254 	struct mmu_update *u;
255 
256 	mcs = xen_mc_entry(sizeof(*u));
257 	u = mcs.args;
258 
259 	/* ptep might be kmapped when using 32-bit HIGHPTE */
260 	u->ptr = arbitrary_virt_to_machine(ptep).maddr;
261 	u->val = pte_val_ma(pteval);
262 
263 	MULTI_mmu_update(mcs.mc, mcs.args, 1, NULL, domid);
264 
265 	xen_mc_issue(PARAVIRT_LAZY_MMU);
266 }
267 EXPORT_SYMBOL_GPL(xen_set_domain_pte);
268 
269 static void xen_set_iomap_pte(pte_t *ptep, pte_t pteval)
270 {
271 	xen_set_domain_pte(ptep, pteval, DOMID_IO);
272 }
273 
274 static void xen_extend_mmu_update(const struct mmu_update *update)
275 {
276 	struct multicall_space mcs;
277 	struct mmu_update *u;
278 
279 	mcs = xen_mc_extend_args(__HYPERVISOR_mmu_update, sizeof(*u));
280 
281 	if (mcs.mc != NULL) {
282 		ADD_STATS(mmu_update_extended, 1);
283 		ADD_STATS(mmu_update_histo[mcs.mc->args[1]], -1);
284 
285 		mcs.mc->args[1]++;
286 
287 		if (mcs.mc->args[1] < MMU_UPDATE_HISTO)
288 			ADD_STATS(mmu_update_histo[mcs.mc->args[1]], 1);
289 		else
290 			ADD_STATS(mmu_update_histo[0], 1);
291 	} else {
292 		ADD_STATS(mmu_update, 1);
293 		mcs = __xen_mc_entry(sizeof(*u));
294 		MULTI_mmu_update(mcs.mc, mcs.args, 1, NULL, DOMID_SELF);
295 		ADD_STATS(mmu_update_histo[1], 1);
296 	}
297 
298 	u = mcs.args;
299 	*u = *update;
300 }
301 
302 void xen_set_pmd_hyper(pmd_t *ptr, pmd_t val)
303 {
304 	struct mmu_update u;
305 
306 	preempt_disable();
307 
308 	xen_mc_batch();
309 
310 	/* ptr may be ioremapped for 64-bit pagetable setup */
311 	u.ptr = arbitrary_virt_to_machine(ptr).maddr;
312 	u.val = pmd_val_ma(val);
313 	xen_extend_mmu_update(&u);
314 
315 	ADD_STATS(pmd_update_batched, paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU);
316 
317 	xen_mc_issue(PARAVIRT_LAZY_MMU);
318 
319 	preempt_enable();
320 }
321 
322 void xen_set_pmd(pmd_t *ptr, pmd_t val)
323 {
324 	ADD_STATS(pmd_update, 1);
325 
326 	/* If page is not pinned, we can just update the entry
327 	   directly */
328 	if (!xen_page_pinned(ptr)) {
329 		*ptr = val;
330 		return;
331 	}
332 
333 	ADD_STATS(pmd_update_pinned, 1);
334 
335 	xen_set_pmd_hyper(ptr, val);
336 }
337 
338 /*
339  * Associate a virtual page frame with a given physical page frame
340  * and protection flags for that frame.
341  */
342 void set_pte_mfn(unsigned long vaddr, unsigned long mfn, pgprot_t flags)
343 {
344 	set_pte_vaddr(vaddr, mfn_pte(mfn, flags));
345 }
346 
347 void xen_set_pte_at(struct mm_struct *mm, unsigned long addr,
348 		    pte_t *ptep, pte_t pteval)
349 {
350 	if (xen_iomap_pte(pteval)) {
351 		xen_set_iomap_pte(ptep, pteval);
352 		goto out;
353 	}
354 
355 	ADD_STATS(set_pte_at, 1);
356 //	ADD_STATS(set_pte_at_pinned, xen_page_pinned(ptep));
357 	ADD_STATS(set_pte_at_current, mm == current->mm);
358 	ADD_STATS(set_pte_at_kernel, mm == &init_mm);
359 
360 	if (mm == current->mm || mm == &init_mm) {
361 		if (paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU) {
362 			struct multicall_space mcs;
363 			mcs = xen_mc_entry(0);
364 
365 			MULTI_update_va_mapping(mcs.mc, addr, pteval, 0);
366 			ADD_STATS(set_pte_at_batched, 1);
367 			xen_mc_issue(PARAVIRT_LAZY_MMU);
368 			goto out;
369 		} else
370 			if (HYPERVISOR_update_va_mapping(addr, pteval, 0) == 0)
371 				goto out;
372 	}
373 	xen_set_pte(ptep, pteval);
374 
375 out:	return;
376 }
377 
378 pte_t xen_ptep_modify_prot_start(struct mm_struct *mm,
379 				 unsigned long addr, pte_t *ptep)
380 {
381 	/* Just return the pte as-is.  We preserve the bits on commit */
382 	return *ptep;
383 }
384 
385 void xen_ptep_modify_prot_commit(struct mm_struct *mm, unsigned long addr,
386 				 pte_t *ptep, pte_t pte)
387 {
388 	struct mmu_update u;
389 
390 	xen_mc_batch();
391 
392 	u.ptr = arbitrary_virt_to_machine(ptep).maddr | MMU_PT_UPDATE_PRESERVE_AD;
393 	u.val = pte_val_ma(pte);
394 	xen_extend_mmu_update(&u);
395 
396 	ADD_STATS(prot_commit, 1);
397 	ADD_STATS(prot_commit_batched, paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU);
398 
399 	xen_mc_issue(PARAVIRT_LAZY_MMU);
400 }
401 
402 /* Assume pteval_t is equivalent to all the other *val_t types. */
403 static pteval_t pte_mfn_to_pfn(pteval_t val)
404 {
405 	if (val & _PAGE_PRESENT) {
406 		unsigned long mfn = (val & PTE_PFN_MASK) >> PAGE_SHIFT;
407 		pteval_t flags = val & PTE_FLAGS_MASK;
408 		val = ((pteval_t)mfn_to_pfn(mfn) << PAGE_SHIFT) | flags;
409 	}
410 
411 	return val;
412 }
413 
414 static pteval_t pte_pfn_to_mfn(pteval_t val)
415 {
416 	if (val & _PAGE_PRESENT) {
417 		unsigned long pfn = (val & PTE_PFN_MASK) >> PAGE_SHIFT;
418 		pteval_t flags = val & PTE_FLAGS_MASK;
419 		unsigned long mfn = pfn_to_mfn(pfn);
420 
421 		/*
422 		 * If there's no mfn for the pfn, then just create an
423 		 * empty non-present pte.  Unfortunately this loses
424 		 * information about the original pfn, so
425 		 * pte_mfn_to_pfn is asymmetric.
426 		 */
427 		if (unlikely(mfn == INVALID_P2M_ENTRY)) {
428 			mfn = 0;
429 			flags = 0;
430 		}
431 
432 		val = ((pteval_t)mfn << PAGE_SHIFT) | flags;
433 	}
434 
435 	return val;
436 }
437 
438 static pteval_t iomap_pte(pteval_t val)
439 {
440 	if (val & _PAGE_PRESENT) {
441 		unsigned long pfn = (val & PTE_PFN_MASK) >> PAGE_SHIFT;
442 		pteval_t flags = val & PTE_FLAGS_MASK;
443 
444 		/* We assume the pte frame number is a MFN, so
445 		   just use it as-is. */
446 		val = ((pteval_t)pfn << PAGE_SHIFT) | flags;
447 	}
448 
449 	return val;
450 }
451 
452 pteval_t xen_pte_val(pte_t pte)
453 {
454 	pteval_t pteval = pte.pte;
455 
456 	/* If this is a WC pte, convert back from Xen WC to Linux WC */
457 	if ((pteval & (_PAGE_PAT | _PAGE_PCD | _PAGE_PWT)) == _PAGE_PAT) {
458 		WARN_ON(!pat_enabled);
459 		pteval = (pteval & ~_PAGE_PAT) | _PAGE_PWT;
460 	}
461 
462 	if (xen_initial_domain() && (pteval & _PAGE_IOMAP))
463 		return pteval;
464 
465 	return pte_mfn_to_pfn(pteval);
466 }
467 PV_CALLEE_SAVE_REGS_THUNK(xen_pte_val);
468 
469 pgdval_t xen_pgd_val(pgd_t pgd)
470 {
471 	return pte_mfn_to_pfn(pgd.pgd);
472 }
473 PV_CALLEE_SAVE_REGS_THUNK(xen_pgd_val);
474 
475 /*
476  * Xen's PAT setup is part of its ABI, though I assume entries 6 & 7
477  * are reserved for now, to correspond to the Intel-reserved PAT
478  * types.
479  *
480  * We expect Linux's PAT set as follows:
481  *
482  * Idx  PTE flags        Linux    Xen    Default
483  * 0                     WB       WB     WB
484  * 1            PWT      WC       WT     WT
485  * 2        PCD          UC-      UC-    UC-
486  * 3        PCD PWT      UC       UC     UC
487  * 4    PAT              WB       WC     WB
488  * 5    PAT     PWT      WC       WP     WT
489  * 6    PAT PCD          UC-      UC     UC-
490  * 7    PAT PCD PWT      UC       UC     UC
491  */
492 
493 void xen_set_pat(u64 pat)
494 {
495 	/* We expect Linux to use a PAT setting of
496 	 * UC UC- WC WB (ignoring the PAT flag) */
497 	WARN_ON(pat != 0x0007010600070106ull);
498 }
499 
500 pte_t xen_make_pte(pteval_t pte)
501 {
502 	phys_addr_t addr = (pte & PTE_PFN_MASK);
503 
504 	/* If Linux is trying to set a WC pte, then map to the Xen WC.
505 	 * If _PAGE_PAT is set, then it probably means it is really
506 	 * _PAGE_PSE, so avoid fiddling with the PAT mapping and hope
507 	 * things work out OK...
508 	 *
509 	 * (We should never see kernel mappings with _PAGE_PSE set,
510 	 * but we could see hugetlbfs mappings, I think.).
511 	 */
512 	if (pat_enabled && !WARN_ON(pte & _PAGE_PAT)) {
513 		if ((pte & (_PAGE_PCD | _PAGE_PWT)) == _PAGE_PWT)
514 			pte = (pte & ~(_PAGE_PCD | _PAGE_PWT)) | _PAGE_PAT;
515 	}
516 
517 	/*
518 	 * Unprivileged domains are allowed to do IOMAPpings for
519 	 * PCI passthrough, but not map ISA space.  The ISA
520 	 * mappings are just dummy local mappings to keep other
521 	 * parts of the kernel happy.
522 	 */
523 	if (unlikely(pte & _PAGE_IOMAP) &&
524 	    (xen_initial_domain() || addr >= ISA_END_ADDRESS)) {
525 		pte = iomap_pte(pte);
526 	} else {
527 		pte &= ~_PAGE_IOMAP;
528 		pte = pte_pfn_to_mfn(pte);
529 	}
530 
531 	return native_make_pte(pte);
532 }
533 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pte);
534 
535 pgd_t xen_make_pgd(pgdval_t pgd)
536 {
537 	pgd = pte_pfn_to_mfn(pgd);
538 	return native_make_pgd(pgd);
539 }
540 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pgd);
541 
542 pmdval_t xen_pmd_val(pmd_t pmd)
543 {
544 	return pte_mfn_to_pfn(pmd.pmd);
545 }
546 PV_CALLEE_SAVE_REGS_THUNK(xen_pmd_val);
547 
548 void xen_set_pud_hyper(pud_t *ptr, pud_t val)
549 {
550 	struct mmu_update u;
551 
552 	preempt_disable();
553 
554 	xen_mc_batch();
555 
556 	/* ptr may be ioremapped for 64-bit pagetable setup */
557 	u.ptr = arbitrary_virt_to_machine(ptr).maddr;
558 	u.val = pud_val_ma(val);
559 	xen_extend_mmu_update(&u);
560 
561 	ADD_STATS(pud_update_batched, paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU);
562 
563 	xen_mc_issue(PARAVIRT_LAZY_MMU);
564 
565 	preempt_enable();
566 }
567 
568 void xen_set_pud(pud_t *ptr, pud_t val)
569 {
570 	ADD_STATS(pud_update, 1);
571 
572 	/* If page is not pinned, we can just update the entry
573 	   directly */
574 	if (!xen_page_pinned(ptr)) {
575 		*ptr = val;
576 		return;
577 	}
578 
579 	ADD_STATS(pud_update_pinned, 1);
580 
581 	xen_set_pud_hyper(ptr, val);
582 }
583 
584 void xen_set_pte(pte_t *ptep, pte_t pte)
585 {
586 	if (xen_iomap_pte(pte)) {
587 		xen_set_iomap_pte(ptep, pte);
588 		return;
589 	}
590 
591 	ADD_STATS(pte_update, 1);
592 //	ADD_STATS(pte_update_pinned, xen_page_pinned(ptep));
593 	ADD_STATS(pte_update_batched, paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU);
594 
595 #ifdef CONFIG_X86_PAE
596 	ptep->pte_high = pte.pte_high;
597 	smp_wmb();
598 	ptep->pte_low = pte.pte_low;
599 #else
600 	*ptep = pte;
601 #endif
602 }
603 
604 #ifdef CONFIG_X86_PAE
605 void xen_set_pte_atomic(pte_t *ptep, pte_t pte)
606 {
607 	if (xen_iomap_pte(pte)) {
608 		xen_set_iomap_pte(ptep, pte);
609 		return;
610 	}
611 
612 	set_64bit((u64 *)ptep, native_pte_val(pte));
613 }
614 
615 void xen_pte_clear(struct mm_struct *mm, unsigned long addr, pte_t *ptep)
616 {
617 	ptep->pte_low = 0;
618 	smp_wmb();		/* make sure low gets written first */
619 	ptep->pte_high = 0;
620 }
621 
622 void xen_pmd_clear(pmd_t *pmdp)
623 {
624 	set_pmd(pmdp, __pmd(0));
625 }
626 #endif	/* CONFIG_X86_PAE */
627 
628 pmd_t xen_make_pmd(pmdval_t pmd)
629 {
630 	pmd = pte_pfn_to_mfn(pmd);
631 	return native_make_pmd(pmd);
632 }
633 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pmd);
634 
635 #if PAGETABLE_LEVELS == 4
636 pudval_t xen_pud_val(pud_t pud)
637 {
638 	return pte_mfn_to_pfn(pud.pud);
639 }
640 PV_CALLEE_SAVE_REGS_THUNK(xen_pud_val);
641 
642 pud_t xen_make_pud(pudval_t pud)
643 {
644 	pud = pte_pfn_to_mfn(pud);
645 
646 	return native_make_pud(pud);
647 }
648 PV_CALLEE_SAVE_REGS_THUNK(xen_make_pud);
649 
650 pgd_t *xen_get_user_pgd(pgd_t *pgd)
651 {
652 	pgd_t *pgd_page = (pgd_t *)(((unsigned long)pgd) & PAGE_MASK);
653 	unsigned offset = pgd - pgd_page;
654 	pgd_t *user_ptr = NULL;
655 
656 	if (offset < pgd_index(USER_LIMIT)) {
657 		struct page *page = virt_to_page(pgd_page);
658 		user_ptr = (pgd_t *)page->private;
659 		if (user_ptr)
660 			user_ptr += offset;
661 	}
662 
663 	return user_ptr;
664 }
665 
666 static void __xen_set_pgd_hyper(pgd_t *ptr, pgd_t val)
667 {
668 	struct mmu_update u;
669 
670 	u.ptr = virt_to_machine(ptr).maddr;
671 	u.val = pgd_val_ma(val);
672 	xen_extend_mmu_update(&u);
673 }
674 
675 /*
676  * Raw hypercall-based set_pgd, intended for in early boot before
677  * there's a page structure.  This implies:
678  *  1. The only existing pagetable is the kernel's
679  *  2. It is always pinned
680  *  3. It has no user pagetable attached to it
681  */
682 void __init xen_set_pgd_hyper(pgd_t *ptr, pgd_t val)
683 {
684 	preempt_disable();
685 
686 	xen_mc_batch();
687 
688 	__xen_set_pgd_hyper(ptr, val);
689 
690 	xen_mc_issue(PARAVIRT_LAZY_MMU);
691 
692 	preempt_enable();
693 }
694 
695 void xen_set_pgd(pgd_t *ptr, pgd_t val)
696 {
697 	pgd_t *user_ptr = xen_get_user_pgd(ptr);
698 
699 	ADD_STATS(pgd_update, 1);
700 
701 	/* If page is not pinned, we can just update the entry
702 	   directly */
703 	if (!xen_page_pinned(ptr)) {
704 		*ptr = val;
705 		if (user_ptr) {
706 			WARN_ON(xen_page_pinned(user_ptr));
707 			*user_ptr = val;
708 		}
709 		return;
710 	}
711 
712 	ADD_STATS(pgd_update_pinned, 1);
713 	ADD_STATS(pgd_update_batched, paravirt_get_lazy_mode() == PARAVIRT_LAZY_MMU);
714 
715 	/* If it's pinned, then we can at least batch the kernel and
716 	   user updates together. */
717 	xen_mc_batch();
718 
719 	__xen_set_pgd_hyper(ptr, val);
720 	if (user_ptr)
721 		__xen_set_pgd_hyper(user_ptr, val);
722 
723 	xen_mc_issue(PARAVIRT_LAZY_MMU);
724 }
725 #endif	/* PAGETABLE_LEVELS == 4 */
726 
727 /*
728  * (Yet another) pagetable walker.  This one is intended for pinning a
729  * pagetable.  This means that it walks a pagetable and calls the
730  * callback function on each page it finds making up the page table,
731  * at every level.  It walks the entire pagetable, but it only bothers
732  * pinning pte pages which are below limit.  In the normal case this
733  * will be STACK_TOP_MAX, but at boot we need to pin up to
734  * FIXADDR_TOP.
735  *
736  * For 32-bit the important bit is that we don't pin beyond there,
737  * because then we start getting into Xen's ptes.
738  *
739  * For 64-bit, we must skip the Xen hole in the middle of the address
740  * space, just after the big x86-64 virtual hole.
741  */
742 static int __xen_pgd_walk(struct mm_struct *mm, pgd_t *pgd,
743 			  int (*func)(struct mm_struct *mm, struct page *,
744 				      enum pt_level),
745 			  unsigned long limit)
746 {
747 	int flush = 0;
748 	unsigned hole_low, hole_high;
749 	unsigned pgdidx_limit, pudidx_limit, pmdidx_limit;
750 	unsigned pgdidx, pudidx, pmdidx;
751 
752 	/* The limit is the last byte to be touched */
753 	limit--;
754 	BUG_ON(limit >= FIXADDR_TOP);
755 
756 	if (xen_feature(XENFEAT_auto_translated_physmap))
757 		return 0;
758 
759 	/*
760 	 * 64-bit has a great big hole in the middle of the address
761 	 * space, which contains the Xen mappings.  On 32-bit these
762 	 * will end up making a zero-sized hole and so is a no-op.
763 	 */
764 	hole_low = pgd_index(USER_LIMIT);
765 	hole_high = pgd_index(PAGE_OFFSET);
766 
767 	pgdidx_limit = pgd_index(limit);
768 #if PTRS_PER_PUD > 1
769 	pudidx_limit = pud_index(limit);
770 #else
771 	pudidx_limit = 0;
772 #endif
773 #if PTRS_PER_PMD > 1
774 	pmdidx_limit = pmd_index(limit);
775 #else
776 	pmdidx_limit = 0;
777 #endif
778 
779 	for (pgdidx = 0; pgdidx <= pgdidx_limit; pgdidx++) {
780 		pud_t *pud;
781 
782 		if (pgdidx >= hole_low && pgdidx < hole_high)
783 			continue;
784 
785 		if (!pgd_val(pgd[pgdidx]))
786 			continue;
787 
788 		pud = pud_offset(&pgd[pgdidx], 0);
789 
790 		if (PTRS_PER_PUD > 1) /* not folded */
791 			flush |= (*func)(mm, virt_to_page(pud), PT_PUD);
792 
793 		for (pudidx = 0; pudidx < PTRS_PER_PUD; pudidx++) {
794 			pmd_t *pmd;
795 
796 			if (pgdidx == pgdidx_limit &&
797 			    pudidx > pudidx_limit)
798 				goto out;
799 
800 			if (pud_none(pud[pudidx]))
801 				continue;
802 
803 			pmd = pmd_offset(&pud[pudidx], 0);
804 
805 			if (PTRS_PER_PMD > 1) /* not folded */
806 				flush |= (*func)(mm, virt_to_page(pmd), PT_PMD);
807 
808 			for (pmdidx = 0; pmdidx < PTRS_PER_PMD; pmdidx++) {
809 				struct page *pte;
810 
811 				if (pgdidx == pgdidx_limit &&
812 				    pudidx == pudidx_limit &&
813 				    pmdidx > pmdidx_limit)
814 					goto out;
815 
816 				if (pmd_none(pmd[pmdidx]))
817 					continue;
818 
819 				pte = pmd_page(pmd[pmdidx]);
820 				flush |= (*func)(mm, pte, PT_PTE);
821 			}
822 		}
823 	}
824 
825 out:
826 	/* Do the top level last, so that the callbacks can use it as
827 	   a cue to do final things like tlb flushes. */
828 	flush |= (*func)(mm, virt_to_page(pgd), PT_PGD);
829 
830 	return flush;
831 }
832 
833 static int xen_pgd_walk(struct mm_struct *mm,
834 			int (*func)(struct mm_struct *mm, struct page *,
835 				    enum pt_level),
836 			unsigned long limit)
837 {
838 	return __xen_pgd_walk(mm, mm->pgd, func, limit);
839 }
840 
841 /* If we're using split pte locks, then take the page's lock and
842    return a pointer to it.  Otherwise return NULL. */
843 static spinlock_t *xen_pte_lock(struct page *page, struct mm_struct *mm)
844 {
845 	spinlock_t *ptl = NULL;
846 
847 #if USE_SPLIT_PTLOCKS
848 	ptl = __pte_lockptr(page);
849 	spin_lock_nest_lock(ptl, &mm->page_table_lock);
850 #endif
851 
852 	return ptl;
853 }
854 
855 static void xen_pte_unlock(void *v)
856 {
857 	spinlock_t *ptl = v;
858 	spin_unlock(ptl);
859 }
860 
861 static void xen_do_pin(unsigned level, unsigned long pfn)
862 {
863 	struct mmuext_op *op;
864 	struct multicall_space mcs;
865 
866 	mcs = __xen_mc_entry(sizeof(*op));
867 	op = mcs.args;
868 	op->cmd = level;
869 	op->arg1.mfn = pfn_to_mfn(pfn);
870 	MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
871 }
872 
873 static int xen_pin_page(struct mm_struct *mm, struct page *page,
874 			enum pt_level level)
875 {
876 	unsigned pgfl = TestSetPagePinned(page);
877 	int flush;
878 
879 	if (pgfl)
880 		flush = 0;		/* already pinned */
881 	else if (PageHighMem(page))
882 		/* kmaps need flushing if we found an unpinned
883 		   highpage */
884 		flush = 1;
885 	else {
886 		void *pt = lowmem_page_address(page);
887 		unsigned long pfn = page_to_pfn(page);
888 		struct multicall_space mcs = __xen_mc_entry(0);
889 		spinlock_t *ptl;
890 
891 		flush = 0;
892 
893 		/*
894 		 * We need to hold the pagetable lock between the time
895 		 * we make the pagetable RO and when we actually pin
896 		 * it.  If we don't, then other users may come in and
897 		 * attempt to update the pagetable by writing it,
898 		 * which will fail because the memory is RO but not
899 		 * pinned, so Xen won't do the trap'n'emulate.
900 		 *
901 		 * If we're using split pte locks, we can't hold the
902 		 * entire pagetable's worth of locks during the
903 		 * traverse, because we may wrap the preempt count (8
904 		 * bits).  The solution is to mark RO and pin each PTE
905 		 * page while holding the lock.  This means the number
906 		 * of locks we end up holding is never more than a
907 		 * batch size (~32 entries, at present).
908 		 *
909 		 * If we're not using split pte locks, we needn't pin
910 		 * the PTE pages independently, because we're
911 		 * protected by the overall pagetable lock.
912 		 */
913 		ptl = NULL;
914 		if (level == PT_PTE)
915 			ptl = xen_pte_lock(page, mm);
916 
917 		MULTI_update_va_mapping(mcs.mc, (unsigned long)pt,
918 					pfn_pte(pfn, PAGE_KERNEL_RO),
919 					level == PT_PGD ? UVMF_TLB_FLUSH : 0);
920 
921 		if (ptl) {
922 			xen_do_pin(MMUEXT_PIN_L1_TABLE, pfn);
923 
924 			/* Queue a deferred unlock for when this batch
925 			   is completed. */
926 			xen_mc_callback(xen_pte_unlock, ptl);
927 		}
928 	}
929 
930 	return flush;
931 }
932 
933 /* This is called just after a mm has been created, but it has not
934    been used yet.  We need to make sure that its pagetable is all
935    read-only, and can be pinned. */
936 static void __xen_pgd_pin(struct mm_struct *mm, pgd_t *pgd)
937 {
938 	xen_mc_batch();
939 
940 	if (__xen_pgd_walk(mm, pgd, xen_pin_page, USER_LIMIT)) {
941 		/* re-enable interrupts for flushing */
942 		xen_mc_issue(0);
943 
944 		kmap_flush_unused();
945 
946 		xen_mc_batch();
947 	}
948 
949 #ifdef CONFIG_X86_64
950 	{
951 		pgd_t *user_pgd = xen_get_user_pgd(pgd);
952 
953 		xen_do_pin(MMUEXT_PIN_L4_TABLE, PFN_DOWN(__pa(pgd)));
954 
955 		if (user_pgd) {
956 			xen_pin_page(mm, virt_to_page(user_pgd), PT_PGD);
957 			xen_do_pin(MMUEXT_PIN_L4_TABLE,
958 				   PFN_DOWN(__pa(user_pgd)));
959 		}
960 	}
961 #else /* CONFIG_X86_32 */
962 #ifdef CONFIG_X86_PAE
963 	/* Need to make sure unshared kernel PMD is pinnable */
964 	xen_pin_page(mm, pgd_page(pgd[pgd_index(TASK_SIZE)]),
965 		     PT_PMD);
966 #endif
967 	xen_do_pin(MMUEXT_PIN_L3_TABLE, PFN_DOWN(__pa(pgd)));
968 #endif /* CONFIG_X86_64 */
969 	xen_mc_issue(0);
970 }
971 
972 static void xen_pgd_pin(struct mm_struct *mm)
973 {
974 	__xen_pgd_pin(mm, mm->pgd);
975 }
976 
977 /*
978  * On save, we need to pin all pagetables to make sure they get their
979  * mfns turned into pfns.  Search the list for any unpinned pgds and pin
980  * them (unpinned pgds are not currently in use, probably because the
981  * process is under construction or destruction).
982  *
983  * Expected to be called in stop_machine() ("equivalent to taking
984  * every spinlock in the system"), so the locking doesn't really
985  * matter all that much.
986  */
987 void xen_mm_pin_all(void)
988 {
989 	struct page *page;
990 
991 	spin_lock(&pgd_lock);
992 
993 	list_for_each_entry(page, &pgd_list, lru) {
994 		if (!PagePinned(page)) {
995 			__xen_pgd_pin(&init_mm, (pgd_t *)page_address(page));
996 			SetPageSavePinned(page);
997 		}
998 	}
999 
1000 	spin_unlock(&pgd_lock);
1001 }
1002 
1003 /*
1004  * The init_mm pagetable is really pinned as soon as its created, but
1005  * that's before we have page structures to store the bits.  So do all
1006  * the book-keeping now.
1007  */
1008 static __init int xen_mark_pinned(struct mm_struct *mm, struct page *page,
1009 				  enum pt_level level)
1010 {
1011 	SetPagePinned(page);
1012 	return 0;
1013 }
1014 
1015 static void __init xen_mark_init_mm_pinned(void)
1016 {
1017 	xen_pgd_walk(&init_mm, xen_mark_pinned, FIXADDR_TOP);
1018 }
1019 
1020 static int xen_unpin_page(struct mm_struct *mm, struct page *page,
1021 			  enum pt_level level)
1022 {
1023 	unsigned pgfl = TestClearPagePinned(page);
1024 
1025 	if (pgfl && !PageHighMem(page)) {
1026 		void *pt = lowmem_page_address(page);
1027 		unsigned long pfn = page_to_pfn(page);
1028 		spinlock_t *ptl = NULL;
1029 		struct multicall_space mcs;
1030 
1031 		/*
1032 		 * Do the converse to pin_page.  If we're using split
1033 		 * pte locks, we must be holding the lock for while
1034 		 * the pte page is unpinned but still RO to prevent
1035 		 * concurrent updates from seeing it in this
1036 		 * partially-pinned state.
1037 		 */
1038 		if (level == PT_PTE) {
1039 			ptl = xen_pte_lock(page, mm);
1040 
1041 			if (ptl)
1042 				xen_do_pin(MMUEXT_UNPIN_TABLE, pfn);
1043 		}
1044 
1045 		mcs = __xen_mc_entry(0);
1046 
1047 		MULTI_update_va_mapping(mcs.mc, (unsigned long)pt,
1048 					pfn_pte(pfn, PAGE_KERNEL),
1049 					level == PT_PGD ? UVMF_TLB_FLUSH : 0);
1050 
1051 		if (ptl) {
1052 			/* unlock when batch completed */
1053 			xen_mc_callback(xen_pte_unlock, ptl);
1054 		}
1055 	}
1056 
1057 	return 0;		/* never need to flush on unpin */
1058 }
1059 
1060 /* Release a pagetables pages back as normal RW */
1061 static void __xen_pgd_unpin(struct mm_struct *mm, pgd_t *pgd)
1062 {
1063 	xen_mc_batch();
1064 
1065 	xen_do_pin(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1066 
1067 #ifdef CONFIG_X86_64
1068 	{
1069 		pgd_t *user_pgd = xen_get_user_pgd(pgd);
1070 
1071 		if (user_pgd) {
1072 			xen_do_pin(MMUEXT_UNPIN_TABLE,
1073 				   PFN_DOWN(__pa(user_pgd)));
1074 			xen_unpin_page(mm, virt_to_page(user_pgd), PT_PGD);
1075 		}
1076 	}
1077 #endif
1078 
1079 #ifdef CONFIG_X86_PAE
1080 	/* Need to make sure unshared kernel PMD is unpinned */
1081 	xen_unpin_page(mm, pgd_page(pgd[pgd_index(TASK_SIZE)]),
1082 		       PT_PMD);
1083 #endif
1084 
1085 	__xen_pgd_walk(mm, pgd, xen_unpin_page, USER_LIMIT);
1086 
1087 	xen_mc_issue(0);
1088 }
1089 
1090 static void xen_pgd_unpin(struct mm_struct *mm)
1091 {
1092 	__xen_pgd_unpin(mm, mm->pgd);
1093 }
1094 
1095 /*
1096  * On resume, undo any pinning done at save, so that the rest of the
1097  * kernel doesn't see any unexpected pinned pagetables.
1098  */
1099 void xen_mm_unpin_all(void)
1100 {
1101 	struct page *page;
1102 
1103 	spin_lock(&pgd_lock);
1104 
1105 	list_for_each_entry(page, &pgd_list, lru) {
1106 		if (PageSavePinned(page)) {
1107 			BUG_ON(!PagePinned(page));
1108 			__xen_pgd_unpin(&init_mm, (pgd_t *)page_address(page));
1109 			ClearPageSavePinned(page);
1110 		}
1111 	}
1112 
1113 	spin_unlock(&pgd_lock);
1114 }
1115 
1116 void xen_activate_mm(struct mm_struct *prev, struct mm_struct *next)
1117 {
1118 	spin_lock(&next->page_table_lock);
1119 	xen_pgd_pin(next);
1120 	spin_unlock(&next->page_table_lock);
1121 }
1122 
1123 void xen_dup_mmap(struct mm_struct *oldmm, struct mm_struct *mm)
1124 {
1125 	spin_lock(&mm->page_table_lock);
1126 	xen_pgd_pin(mm);
1127 	spin_unlock(&mm->page_table_lock);
1128 }
1129 
1130 
1131 #ifdef CONFIG_SMP
1132 /* Another cpu may still have their %cr3 pointing at the pagetable, so
1133    we need to repoint it somewhere else before we can unpin it. */
1134 static void drop_other_mm_ref(void *info)
1135 {
1136 	struct mm_struct *mm = info;
1137 	struct mm_struct *active_mm;
1138 
1139 	active_mm = percpu_read(cpu_tlbstate.active_mm);
1140 
1141 	if (active_mm == mm)
1142 		leave_mm(smp_processor_id());
1143 
1144 	/* If this cpu still has a stale cr3 reference, then make sure
1145 	   it has been flushed. */
1146 	if (percpu_read(xen_current_cr3) == __pa(mm->pgd))
1147 		load_cr3(swapper_pg_dir);
1148 }
1149 
1150 static void xen_drop_mm_ref(struct mm_struct *mm)
1151 {
1152 	cpumask_var_t mask;
1153 	unsigned cpu;
1154 
1155 	if (current->active_mm == mm) {
1156 		if (current->mm == mm)
1157 			load_cr3(swapper_pg_dir);
1158 		else
1159 			leave_mm(smp_processor_id());
1160 	}
1161 
1162 	/* Get the "official" set of cpus referring to our pagetable. */
1163 	if (!alloc_cpumask_var(&mask, GFP_ATOMIC)) {
1164 		for_each_online_cpu(cpu) {
1165 			if (!cpumask_test_cpu(cpu, mm_cpumask(mm))
1166 			    && per_cpu(xen_current_cr3, cpu) != __pa(mm->pgd))
1167 				continue;
1168 			smp_call_function_single(cpu, drop_other_mm_ref, mm, 1);
1169 		}
1170 		return;
1171 	}
1172 	cpumask_copy(mask, mm_cpumask(mm));
1173 
1174 	/* It's possible that a vcpu may have a stale reference to our
1175 	   cr3, because its in lazy mode, and it hasn't yet flushed
1176 	   its set of pending hypercalls yet.  In this case, we can
1177 	   look at its actual current cr3 value, and force it to flush
1178 	   if needed. */
1179 	for_each_online_cpu(cpu) {
1180 		if (per_cpu(xen_current_cr3, cpu) == __pa(mm->pgd))
1181 			cpumask_set_cpu(cpu, mask);
1182 	}
1183 
1184 	if (!cpumask_empty(mask))
1185 		smp_call_function_many(mask, drop_other_mm_ref, mm, 1);
1186 	free_cpumask_var(mask);
1187 }
1188 #else
1189 static void xen_drop_mm_ref(struct mm_struct *mm)
1190 {
1191 	if (current->active_mm == mm)
1192 		load_cr3(swapper_pg_dir);
1193 }
1194 #endif
1195 
1196 /*
1197  * While a process runs, Xen pins its pagetables, which means that the
1198  * hypervisor forces it to be read-only, and it controls all updates
1199  * to it.  This means that all pagetable updates have to go via the
1200  * hypervisor, which is moderately expensive.
1201  *
1202  * Since we're pulling the pagetable down, we switch to use init_mm,
1203  * unpin old process pagetable and mark it all read-write, which
1204  * allows further operations on it to be simple memory accesses.
1205  *
1206  * The only subtle point is that another CPU may be still using the
1207  * pagetable because of lazy tlb flushing.  This means we need need to
1208  * switch all CPUs off this pagetable before we can unpin it.
1209  */
1210 void xen_exit_mmap(struct mm_struct *mm)
1211 {
1212 	get_cpu();		/* make sure we don't move around */
1213 	xen_drop_mm_ref(mm);
1214 	put_cpu();
1215 
1216 	spin_lock(&mm->page_table_lock);
1217 
1218 	/* pgd may not be pinned in the error exit path of execve */
1219 	if (xen_page_pinned(mm->pgd))
1220 		xen_pgd_unpin(mm);
1221 
1222 	spin_unlock(&mm->page_table_lock);
1223 }
1224 
1225 static __init void xen_pagetable_setup_start(pgd_t *base)
1226 {
1227 }
1228 
1229 static void xen_post_allocator_init(void);
1230 
1231 static __init void xen_pagetable_setup_done(pgd_t *base)
1232 {
1233 	xen_setup_shared_info();
1234 	xen_post_allocator_init();
1235 }
1236 
1237 static void xen_write_cr2(unsigned long cr2)
1238 {
1239 	percpu_read(xen_vcpu)->arch.cr2 = cr2;
1240 }
1241 
1242 static unsigned long xen_read_cr2(void)
1243 {
1244 	return percpu_read(xen_vcpu)->arch.cr2;
1245 }
1246 
1247 unsigned long xen_read_cr2_direct(void)
1248 {
1249 	return percpu_read(xen_vcpu_info.arch.cr2);
1250 }
1251 
1252 static void xen_flush_tlb(void)
1253 {
1254 	struct mmuext_op *op;
1255 	struct multicall_space mcs;
1256 
1257 	preempt_disable();
1258 
1259 	mcs = xen_mc_entry(sizeof(*op));
1260 
1261 	op = mcs.args;
1262 	op->cmd = MMUEXT_TLB_FLUSH_LOCAL;
1263 	MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
1264 
1265 	xen_mc_issue(PARAVIRT_LAZY_MMU);
1266 
1267 	preempt_enable();
1268 }
1269 
1270 static void xen_flush_tlb_single(unsigned long addr)
1271 {
1272 	struct mmuext_op *op;
1273 	struct multicall_space mcs;
1274 
1275 	preempt_disable();
1276 
1277 	mcs = xen_mc_entry(sizeof(*op));
1278 	op = mcs.args;
1279 	op->cmd = MMUEXT_INVLPG_LOCAL;
1280 	op->arg1.linear_addr = addr & PAGE_MASK;
1281 	MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
1282 
1283 	xen_mc_issue(PARAVIRT_LAZY_MMU);
1284 
1285 	preempt_enable();
1286 }
1287 
1288 static void xen_flush_tlb_others(const struct cpumask *cpus,
1289 				 struct mm_struct *mm, unsigned long va)
1290 {
1291 	struct {
1292 		struct mmuext_op op;
1293 		DECLARE_BITMAP(mask, NR_CPUS);
1294 	} *args;
1295 	struct multicall_space mcs;
1296 
1297 	if (cpumask_empty(cpus))
1298 		return;		/* nothing to do */
1299 
1300 	mcs = xen_mc_entry(sizeof(*args));
1301 	args = mcs.args;
1302 	args->op.arg2.vcpumask = to_cpumask(args->mask);
1303 
1304 	/* Remove us, and any offline CPUS. */
1305 	cpumask_and(to_cpumask(args->mask), cpus, cpu_online_mask);
1306 	cpumask_clear_cpu(smp_processor_id(), to_cpumask(args->mask));
1307 
1308 	if (va == TLB_FLUSH_ALL) {
1309 		args->op.cmd = MMUEXT_TLB_FLUSH_MULTI;
1310 	} else {
1311 		args->op.cmd = MMUEXT_INVLPG_MULTI;
1312 		args->op.arg1.linear_addr = va;
1313 	}
1314 
1315 	MULTI_mmuext_op(mcs.mc, &args->op, 1, NULL, DOMID_SELF);
1316 
1317 	xen_mc_issue(PARAVIRT_LAZY_MMU);
1318 }
1319 
1320 static unsigned long xen_read_cr3(void)
1321 {
1322 	return percpu_read(xen_cr3);
1323 }
1324 
1325 static void set_current_cr3(void *v)
1326 {
1327 	percpu_write(xen_current_cr3, (unsigned long)v);
1328 }
1329 
1330 static void __xen_write_cr3(bool kernel, unsigned long cr3)
1331 {
1332 	struct mmuext_op *op;
1333 	struct multicall_space mcs;
1334 	unsigned long mfn;
1335 
1336 	if (cr3)
1337 		mfn = pfn_to_mfn(PFN_DOWN(cr3));
1338 	else
1339 		mfn = 0;
1340 
1341 	WARN_ON(mfn == 0 && kernel);
1342 
1343 	mcs = __xen_mc_entry(sizeof(*op));
1344 
1345 	op = mcs.args;
1346 	op->cmd = kernel ? MMUEXT_NEW_BASEPTR : MMUEXT_NEW_USER_BASEPTR;
1347 	op->arg1.mfn = mfn;
1348 
1349 	MULTI_mmuext_op(mcs.mc, op, 1, NULL, DOMID_SELF);
1350 
1351 	if (kernel) {
1352 		percpu_write(xen_cr3, cr3);
1353 
1354 		/* Update xen_current_cr3 once the batch has actually
1355 		   been submitted. */
1356 		xen_mc_callback(set_current_cr3, (void *)cr3);
1357 	}
1358 }
1359 
1360 static void xen_write_cr3(unsigned long cr3)
1361 {
1362 	BUG_ON(preemptible());
1363 
1364 	xen_mc_batch();  /* disables interrupts */
1365 
1366 	/* Update while interrupts are disabled, so its atomic with
1367 	   respect to ipis */
1368 	percpu_write(xen_cr3, cr3);
1369 
1370 	__xen_write_cr3(true, cr3);
1371 
1372 #ifdef CONFIG_X86_64
1373 	{
1374 		pgd_t *user_pgd = xen_get_user_pgd(__va(cr3));
1375 		if (user_pgd)
1376 			__xen_write_cr3(false, __pa(user_pgd));
1377 		else
1378 			__xen_write_cr3(false, 0);
1379 	}
1380 #endif
1381 
1382 	xen_mc_issue(PARAVIRT_LAZY_CPU);  /* interrupts restored */
1383 }
1384 
1385 static int xen_pgd_alloc(struct mm_struct *mm)
1386 {
1387 	pgd_t *pgd = mm->pgd;
1388 	int ret = 0;
1389 
1390 	BUG_ON(PagePinned(virt_to_page(pgd)));
1391 
1392 #ifdef CONFIG_X86_64
1393 	{
1394 		struct page *page = virt_to_page(pgd);
1395 		pgd_t *user_pgd;
1396 
1397 		BUG_ON(page->private != 0);
1398 
1399 		ret = -ENOMEM;
1400 
1401 		user_pgd = (pgd_t *)__get_free_page(GFP_KERNEL | __GFP_ZERO);
1402 		page->private = (unsigned long)user_pgd;
1403 
1404 		if (user_pgd != NULL) {
1405 			user_pgd[pgd_index(VSYSCALL_START)] =
1406 				__pgd(__pa(level3_user_vsyscall) | _PAGE_TABLE);
1407 			ret = 0;
1408 		}
1409 
1410 		BUG_ON(PagePinned(virt_to_page(xen_get_user_pgd(pgd))));
1411 	}
1412 #endif
1413 
1414 	return ret;
1415 }
1416 
1417 static void xen_pgd_free(struct mm_struct *mm, pgd_t *pgd)
1418 {
1419 #ifdef CONFIG_X86_64
1420 	pgd_t *user_pgd = xen_get_user_pgd(pgd);
1421 
1422 	if (user_pgd)
1423 		free_page((unsigned long)user_pgd);
1424 #endif
1425 }
1426 
1427 static __init pte_t mask_rw_pte(pte_t *ptep, pte_t pte)
1428 {
1429 	unsigned long pfn = pte_pfn(pte);
1430 
1431 #ifdef CONFIG_X86_32
1432 	/* If there's an existing pte, then don't allow _PAGE_RW to be set */
1433 	if (pte_val_ma(*ptep) & _PAGE_PRESENT)
1434 		pte = __pte_ma(((pte_val_ma(*ptep) & _PAGE_RW) | ~_PAGE_RW) &
1435 			       pte_val_ma(pte));
1436 #endif
1437 
1438 	/*
1439 	 * If the new pfn is within the range of the newly allocated
1440 	 * kernel pagetable, and it isn't being mapped into an
1441 	 * early_ioremap fixmap slot, make sure it is RO.
1442 	 */
1443 	if (!is_early_ioremap_ptep(ptep) &&
1444 	    pfn >= e820_table_start && pfn < e820_table_end)
1445 		pte = pte_wrprotect(pte);
1446 
1447 	return pte;
1448 }
1449 
1450 /* Init-time set_pte while constructing initial pagetables, which
1451    doesn't allow RO pagetable pages to be remapped RW */
1452 static __init void xen_set_pte_init(pte_t *ptep, pte_t pte)
1453 {
1454 	pte = mask_rw_pte(ptep, pte);
1455 
1456 	xen_set_pte(ptep, pte);
1457 }
1458 
1459 static void pin_pagetable_pfn(unsigned cmd, unsigned long pfn)
1460 {
1461 	struct mmuext_op op;
1462 	op.cmd = cmd;
1463 	op.arg1.mfn = pfn_to_mfn(pfn);
1464 	if (HYPERVISOR_mmuext_op(&op, 1, NULL, DOMID_SELF))
1465 		BUG();
1466 }
1467 
1468 /* Early in boot, while setting up the initial pagetable, assume
1469    everything is pinned. */
1470 static __init void xen_alloc_pte_init(struct mm_struct *mm, unsigned long pfn)
1471 {
1472 #ifdef CONFIG_FLATMEM
1473 	BUG_ON(mem_map);	/* should only be used early */
1474 #endif
1475 	make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
1476 	pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
1477 }
1478 
1479 /* Used for pmd and pud */
1480 static __init void xen_alloc_pmd_init(struct mm_struct *mm, unsigned long pfn)
1481 {
1482 #ifdef CONFIG_FLATMEM
1483 	BUG_ON(mem_map);	/* should only be used early */
1484 #endif
1485 	make_lowmem_page_readonly(__va(PFN_PHYS(pfn)));
1486 }
1487 
1488 /* Early release_pte assumes that all pts are pinned, since there's
1489    only init_mm and anything attached to that is pinned. */
1490 static __init void xen_release_pte_init(unsigned long pfn)
1491 {
1492 	pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
1493 	make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
1494 }
1495 
1496 static __init void xen_release_pmd_init(unsigned long pfn)
1497 {
1498 	make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
1499 }
1500 
1501 /* This needs to make sure the new pte page is pinned iff its being
1502    attached to a pinned pagetable. */
1503 static void xen_alloc_ptpage(struct mm_struct *mm, unsigned long pfn, unsigned level)
1504 {
1505 	struct page *page = pfn_to_page(pfn);
1506 
1507 	if (PagePinned(virt_to_page(mm->pgd))) {
1508 		SetPagePinned(page);
1509 
1510 		if (!PageHighMem(page)) {
1511 			make_lowmem_page_readonly(__va(PFN_PHYS((unsigned long)pfn)));
1512 			if (level == PT_PTE && USE_SPLIT_PTLOCKS)
1513 				pin_pagetable_pfn(MMUEXT_PIN_L1_TABLE, pfn);
1514 		} else {
1515 			/* make sure there are no stray mappings of
1516 			   this page */
1517 			kmap_flush_unused();
1518 		}
1519 	}
1520 }
1521 
1522 static void xen_alloc_pte(struct mm_struct *mm, unsigned long pfn)
1523 {
1524 	xen_alloc_ptpage(mm, pfn, PT_PTE);
1525 }
1526 
1527 static void xen_alloc_pmd(struct mm_struct *mm, unsigned long pfn)
1528 {
1529 	xen_alloc_ptpage(mm, pfn, PT_PMD);
1530 }
1531 
1532 /* This should never happen until we're OK to use struct page */
1533 static void xen_release_ptpage(unsigned long pfn, unsigned level)
1534 {
1535 	struct page *page = pfn_to_page(pfn);
1536 
1537 	if (PagePinned(page)) {
1538 		if (!PageHighMem(page)) {
1539 			if (level == PT_PTE && USE_SPLIT_PTLOCKS)
1540 				pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, pfn);
1541 			make_lowmem_page_readwrite(__va(PFN_PHYS(pfn)));
1542 		}
1543 		ClearPagePinned(page);
1544 	}
1545 }
1546 
1547 static void xen_release_pte(unsigned long pfn)
1548 {
1549 	xen_release_ptpage(pfn, PT_PTE);
1550 }
1551 
1552 static void xen_release_pmd(unsigned long pfn)
1553 {
1554 	xen_release_ptpage(pfn, PT_PMD);
1555 }
1556 
1557 #if PAGETABLE_LEVELS == 4
1558 static void xen_alloc_pud(struct mm_struct *mm, unsigned long pfn)
1559 {
1560 	xen_alloc_ptpage(mm, pfn, PT_PUD);
1561 }
1562 
1563 static void xen_release_pud(unsigned long pfn)
1564 {
1565 	xen_release_ptpage(pfn, PT_PUD);
1566 }
1567 #endif
1568 
1569 void __init xen_reserve_top(void)
1570 {
1571 #ifdef CONFIG_X86_32
1572 	unsigned long top = HYPERVISOR_VIRT_START;
1573 	struct xen_platform_parameters pp;
1574 
1575 	if (HYPERVISOR_xen_version(XENVER_platform_parameters, &pp) == 0)
1576 		top = pp.virt_start;
1577 
1578 	reserve_top_address(-top);
1579 #endif	/* CONFIG_X86_32 */
1580 }
1581 
1582 /*
1583  * Like __va(), but returns address in the kernel mapping (which is
1584  * all we have until the physical memory mapping has been set up.
1585  */
1586 static void *__ka(phys_addr_t paddr)
1587 {
1588 #ifdef CONFIG_X86_64
1589 	return (void *)(paddr + __START_KERNEL_map);
1590 #else
1591 	return __va(paddr);
1592 #endif
1593 }
1594 
1595 /* Convert a machine address to physical address */
1596 static unsigned long m2p(phys_addr_t maddr)
1597 {
1598 	phys_addr_t paddr;
1599 
1600 	maddr &= PTE_PFN_MASK;
1601 	paddr = mfn_to_pfn(maddr >> PAGE_SHIFT) << PAGE_SHIFT;
1602 
1603 	return paddr;
1604 }
1605 
1606 /* Convert a machine address to kernel virtual */
1607 static void *m2v(phys_addr_t maddr)
1608 {
1609 	return __ka(m2p(maddr));
1610 }
1611 
1612 /* Set the page permissions on an identity-mapped pages */
1613 static void set_page_prot(void *addr, pgprot_t prot)
1614 {
1615 	unsigned long pfn = __pa(addr) >> PAGE_SHIFT;
1616 	pte_t pte = pfn_pte(pfn, prot);
1617 
1618 	if (HYPERVISOR_update_va_mapping((unsigned long)addr, pte, 0))
1619 		BUG();
1620 }
1621 
1622 static __init void xen_map_identity_early(pmd_t *pmd, unsigned long max_pfn)
1623 {
1624 	unsigned pmdidx, pteidx;
1625 	unsigned ident_pte;
1626 	unsigned long pfn;
1627 
1628 	level1_ident_pgt = extend_brk(sizeof(pte_t) * LEVEL1_IDENT_ENTRIES,
1629 				      PAGE_SIZE);
1630 
1631 	ident_pte = 0;
1632 	pfn = 0;
1633 	for (pmdidx = 0; pmdidx < PTRS_PER_PMD && pfn < max_pfn; pmdidx++) {
1634 		pte_t *pte_page;
1635 
1636 		/* Reuse or allocate a page of ptes */
1637 		if (pmd_present(pmd[pmdidx]))
1638 			pte_page = m2v(pmd[pmdidx].pmd);
1639 		else {
1640 			/* Check for free pte pages */
1641 			if (ident_pte == LEVEL1_IDENT_ENTRIES)
1642 				break;
1643 
1644 			pte_page = &level1_ident_pgt[ident_pte];
1645 			ident_pte += PTRS_PER_PTE;
1646 
1647 			pmd[pmdidx] = __pmd(__pa(pte_page) | _PAGE_TABLE);
1648 		}
1649 
1650 		/* Install mappings */
1651 		for (pteidx = 0; pteidx < PTRS_PER_PTE; pteidx++, pfn++) {
1652 			pte_t pte;
1653 
1654 			if (pfn > max_pfn_mapped)
1655 				max_pfn_mapped = pfn;
1656 
1657 			if (!pte_none(pte_page[pteidx]))
1658 				continue;
1659 
1660 			pte = pfn_pte(pfn, PAGE_KERNEL_EXEC);
1661 			pte_page[pteidx] = pte;
1662 		}
1663 	}
1664 
1665 	for (pteidx = 0; pteidx < ident_pte; pteidx += PTRS_PER_PTE)
1666 		set_page_prot(&level1_ident_pgt[pteidx], PAGE_KERNEL_RO);
1667 
1668 	set_page_prot(pmd, PAGE_KERNEL_RO);
1669 }
1670 
1671 void __init xen_setup_machphys_mapping(void)
1672 {
1673 	struct xen_machphys_mapping mapping;
1674 	unsigned long machine_to_phys_nr_ents;
1675 
1676 	if (HYPERVISOR_memory_op(XENMEM_machphys_mapping, &mapping) == 0) {
1677 		machine_to_phys_mapping = (unsigned long *)mapping.v_start;
1678 		machine_to_phys_nr_ents = mapping.max_mfn + 1;
1679 	} else {
1680 		machine_to_phys_nr_ents = MACH2PHYS_NR_ENTRIES;
1681 	}
1682 	machine_to_phys_order = fls(machine_to_phys_nr_ents - 1);
1683 }
1684 
1685 #ifdef CONFIG_X86_64
1686 static void convert_pfn_mfn(void *v)
1687 {
1688 	pte_t *pte = v;
1689 	int i;
1690 
1691 	/* All levels are converted the same way, so just treat them
1692 	   as ptes. */
1693 	for (i = 0; i < PTRS_PER_PTE; i++)
1694 		pte[i] = xen_make_pte(pte[i].pte);
1695 }
1696 
1697 /*
1698  * Set up the inital kernel pagetable.
1699  *
1700  * We can construct this by grafting the Xen provided pagetable into
1701  * head_64.S's preconstructed pagetables.  We copy the Xen L2's into
1702  * level2_ident_pgt, level2_kernel_pgt and level2_fixmap_pgt.  This
1703  * means that only the kernel has a physical mapping to start with -
1704  * but that's enough to get __va working.  We need to fill in the rest
1705  * of the physical mapping once some sort of allocator has been set
1706  * up.
1707  */
1708 __init pgd_t *xen_setup_kernel_pagetable(pgd_t *pgd,
1709 					 unsigned long max_pfn)
1710 {
1711 	pud_t *l3;
1712 	pmd_t *l2;
1713 
1714 	/* Zap identity mapping */
1715 	init_level4_pgt[0] = __pgd(0);
1716 
1717 	/* Pre-constructed entries are in pfn, so convert to mfn */
1718 	convert_pfn_mfn(init_level4_pgt);
1719 	convert_pfn_mfn(level3_ident_pgt);
1720 	convert_pfn_mfn(level3_kernel_pgt);
1721 
1722 	l3 = m2v(pgd[pgd_index(__START_KERNEL_map)].pgd);
1723 	l2 = m2v(l3[pud_index(__START_KERNEL_map)].pud);
1724 
1725 	memcpy(level2_ident_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1726 	memcpy(level2_kernel_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1727 
1728 	l3 = m2v(pgd[pgd_index(__START_KERNEL_map + PMD_SIZE)].pgd);
1729 	l2 = m2v(l3[pud_index(__START_KERNEL_map + PMD_SIZE)].pud);
1730 	memcpy(level2_fixmap_pgt, l2, sizeof(pmd_t) * PTRS_PER_PMD);
1731 
1732 	/* Set up identity map */
1733 	xen_map_identity_early(level2_ident_pgt, max_pfn);
1734 
1735 	/* Make pagetable pieces RO */
1736 	set_page_prot(init_level4_pgt, PAGE_KERNEL_RO);
1737 	set_page_prot(level3_ident_pgt, PAGE_KERNEL_RO);
1738 	set_page_prot(level3_kernel_pgt, PAGE_KERNEL_RO);
1739 	set_page_prot(level3_user_vsyscall, PAGE_KERNEL_RO);
1740 	set_page_prot(level2_kernel_pgt, PAGE_KERNEL_RO);
1741 	set_page_prot(level2_fixmap_pgt, PAGE_KERNEL_RO);
1742 
1743 	/* Pin down new L4 */
1744 	pin_pagetable_pfn(MMUEXT_PIN_L4_TABLE,
1745 			  PFN_DOWN(__pa_symbol(init_level4_pgt)));
1746 
1747 	/* Unpin Xen-provided one */
1748 	pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1749 
1750 	/* Switch over */
1751 	pgd = init_level4_pgt;
1752 
1753 	/*
1754 	 * At this stage there can be no user pgd, and no page
1755 	 * structure to attach it to, so make sure we just set kernel
1756 	 * pgd.
1757 	 */
1758 	xen_mc_batch();
1759 	__xen_write_cr3(true, __pa(pgd));
1760 	xen_mc_issue(PARAVIRT_LAZY_CPU);
1761 
1762 	memblock_x86_reserve_range(__pa(xen_start_info->pt_base),
1763 		      __pa(xen_start_info->pt_base +
1764 			   xen_start_info->nr_pt_frames * PAGE_SIZE),
1765 		      "XEN PAGETABLES");
1766 
1767 	return pgd;
1768 }
1769 #else	/* !CONFIG_X86_64 */
1770 static RESERVE_BRK_ARRAY(pmd_t, initial_kernel_pmd, PTRS_PER_PMD);
1771 static RESERVE_BRK_ARRAY(pmd_t, swapper_kernel_pmd, PTRS_PER_PMD);
1772 
1773 static __init void xen_write_cr3_init(unsigned long cr3)
1774 {
1775 	unsigned long pfn = PFN_DOWN(__pa(swapper_pg_dir));
1776 
1777 	BUG_ON(read_cr3() != __pa(initial_page_table));
1778 	BUG_ON(cr3 != __pa(swapper_pg_dir));
1779 
1780 	/*
1781 	 * We are switching to swapper_pg_dir for the first time (from
1782 	 * initial_page_table) and therefore need to mark that page
1783 	 * read-only and then pin it.
1784 	 *
1785 	 * Xen disallows sharing of kernel PMDs for PAE
1786 	 * guests. Therefore we must copy the kernel PMD from
1787 	 * initial_page_table into a new kernel PMD to be used in
1788 	 * swapper_pg_dir.
1789 	 */
1790 	swapper_kernel_pmd =
1791 		extend_brk(sizeof(pmd_t) * PTRS_PER_PMD, PAGE_SIZE);
1792 	memcpy(swapper_kernel_pmd, initial_kernel_pmd,
1793 	       sizeof(pmd_t) * PTRS_PER_PMD);
1794 	swapper_pg_dir[KERNEL_PGD_BOUNDARY] =
1795 		__pgd(__pa(swapper_kernel_pmd) | _PAGE_PRESENT);
1796 	set_page_prot(swapper_kernel_pmd, PAGE_KERNEL_RO);
1797 
1798 	set_page_prot(swapper_pg_dir, PAGE_KERNEL_RO);
1799 	xen_write_cr3(cr3);
1800 	pin_pagetable_pfn(MMUEXT_PIN_L3_TABLE, pfn);
1801 
1802 	pin_pagetable_pfn(MMUEXT_UNPIN_TABLE,
1803 			  PFN_DOWN(__pa(initial_page_table)));
1804 	set_page_prot(initial_page_table, PAGE_KERNEL);
1805 	set_page_prot(initial_kernel_pmd, PAGE_KERNEL);
1806 
1807 	pv_mmu_ops.write_cr3 = &xen_write_cr3;
1808 }
1809 
1810 __init pgd_t *xen_setup_kernel_pagetable(pgd_t *pgd,
1811 					 unsigned long max_pfn)
1812 {
1813 	pmd_t *kernel_pmd;
1814 
1815 	initial_kernel_pmd =
1816 		extend_brk(sizeof(pmd_t) * PTRS_PER_PMD, PAGE_SIZE);
1817 
1818 	max_pfn_mapped = PFN_DOWN(__pa(xen_start_info->pt_base) +
1819 				  xen_start_info->nr_pt_frames * PAGE_SIZE +
1820 				  512*1024);
1821 
1822 	kernel_pmd = m2v(pgd[KERNEL_PGD_BOUNDARY].pgd);
1823 	memcpy(initial_kernel_pmd, kernel_pmd, sizeof(pmd_t) * PTRS_PER_PMD);
1824 
1825 	xen_map_identity_early(initial_kernel_pmd, max_pfn);
1826 
1827 	memcpy(initial_page_table, pgd, sizeof(pgd_t) * PTRS_PER_PGD);
1828 	initial_page_table[KERNEL_PGD_BOUNDARY] =
1829 		__pgd(__pa(initial_kernel_pmd) | _PAGE_PRESENT);
1830 
1831 	set_page_prot(initial_kernel_pmd, PAGE_KERNEL_RO);
1832 	set_page_prot(initial_page_table, PAGE_KERNEL_RO);
1833 	set_page_prot(empty_zero_page, PAGE_KERNEL_RO);
1834 
1835 	pin_pagetable_pfn(MMUEXT_UNPIN_TABLE, PFN_DOWN(__pa(pgd)));
1836 
1837 	pin_pagetable_pfn(MMUEXT_PIN_L3_TABLE,
1838 			  PFN_DOWN(__pa(initial_page_table)));
1839 	xen_write_cr3(__pa(initial_page_table));
1840 
1841 	memblock_x86_reserve_range(__pa(xen_start_info->pt_base),
1842 		      __pa(xen_start_info->pt_base +
1843 			   xen_start_info->nr_pt_frames * PAGE_SIZE),
1844 		      "XEN PAGETABLES");
1845 
1846 	return initial_page_table;
1847 }
1848 #endif	/* CONFIG_X86_64 */
1849 
1850 static unsigned char dummy_mapping[PAGE_SIZE] __page_aligned_bss;
1851 
1852 static void xen_set_fixmap(unsigned idx, phys_addr_t phys, pgprot_t prot)
1853 {
1854 	pte_t pte;
1855 
1856 	phys >>= PAGE_SHIFT;
1857 
1858 	switch (idx) {
1859 	case FIX_BTMAP_END ... FIX_BTMAP_BEGIN:
1860 #ifdef CONFIG_X86_F00F_BUG
1861 	case FIX_F00F_IDT:
1862 #endif
1863 #ifdef CONFIG_X86_32
1864 	case FIX_WP_TEST:
1865 	case FIX_VDSO:
1866 # ifdef CONFIG_HIGHMEM
1867 	case FIX_KMAP_BEGIN ... FIX_KMAP_END:
1868 # endif
1869 #else
1870 	case VSYSCALL_LAST_PAGE ... VSYSCALL_FIRST_PAGE:
1871 #endif
1872 	case FIX_TEXT_POKE0:
1873 	case FIX_TEXT_POKE1:
1874 		/* All local page mappings */
1875 		pte = pfn_pte(phys, prot);
1876 		break;
1877 
1878 #ifdef CONFIG_X86_LOCAL_APIC
1879 	case FIX_APIC_BASE:	/* maps dummy local APIC */
1880 		pte = pfn_pte(PFN_DOWN(__pa(dummy_mapping)), PAGE_KERNEL);
1881 		break;
1882 #endif
1883 
1884 #ifdef CONFIG_X86_IO_APIC
1885 	case FIX_IO_APIC_BASE_0 ... FIX_IO_APIC_BASE_END:
1886 		/*
1887 		 * We just don't map the IO APIC - all access is via
1888 		 * hypercalls.  Keep the address in the pte for reference.
1889 		 */
1890 		pte = pfn_pte(PFN_DOWN(__pa(dummy_mapping)), PAGE_KERNEL);
1891 		break;
1892 #endif
1893 
1894 	case FIX_PARAVIRT_BOOTMAP:
1895 		/* This is an MFN, but it isn't an IO mapping from the
1896 		   IO domain */
1897 		pte = mfn_pte(phys, prot);
1898 		break;
1899 
1900 	default:
1901 		/* By default, set_fixmap is used for hardware mappings */
1902 		pte = mfn_pte(phys, __pgprot(pgprot_val(prot) | _PAGE_IOMAP));
1903 		break;
1904 	}
1905 
1906 	__native_set_fixmap(idx, pte);
1907 
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);
1914 	}
1915 #endif
1916 }
1917 
1918 __init void xen_ident_map_ISA(void)
1919 {
1920 	unsigned long pa;
1921 
1922 	/*
1923 	 * If we're dom0, then linear map the ISA machine addresses into
1924 	 * the kernel's address space.
1925 	 */
1926 	if (!xen_initial_domain())
1927 		return;
1928 
1929 	xen_raw_printk("Xen: setup ISA identity maps\n");
1930 
1931 	for (pa = ISA_START_ADDRESS; pa < ISA_END_ADDRESS; pa += PAGE_SIZE) {
1932 		pte_t pte = mfn_pte(PFN_DOWN(pa), PAGE_KERNEL_IO);
1933 
1934 		if (HYPERVISOR_update_va_mapping(PAGE_OFFSET + pa, pte, 0))
1935 			BUG();
1936 	}
1937 
1938 	xen_flush_tlb();
1939 }
1940 
1941 static __init void xen_post_allocator_init(void)
1942 {
1943 	pv_mmu_ops.set_pte = xen_set_pte;
1944 	pv_mmu_ops.set_pmd = xen_set_pmd;
1945 	pv_mmu_ops.set_pud = xen_set_pud;
1946 #if PAGETABLE_LEVELS == 4
1947 	pv_mmu_ops.set_pgd = xen_set_pgd;
1948 #endif
1949 
1950 	/* This will work as long as patching hasn't happened yet
1951 	   (which it hasn't) */
1952 	pv_mmu_ops.alloc_pte = xen_alloc_pte;
1953 	pv_mmu_ops.alloc_pmd = xen_alloc_pmd;
1954 	pv_mmu_ops.release_pte = xen_release_pte;
1955 	pv_mmu_ops.release_pmd = xen_release_pmd;
1956 #if PAGETABLE_LEVELS == 4
1957 	pv_mmu_ops.alloc_pud = xen_alloc_pud;
1958 	pv_mmu_ops.release_pud = xen_release_pud;
1959 #endif
1960 
1961 #ifdef CONFIG_X86_64
1962 	SetPagePinned(virt_to_page(level3_user_vsyscall));
1963 #endif
1964 	xen_mark_init_mm_pinned();
1965 }
1966 
1967 static void xen_leave_lazy_mmu(void)
1968 {
1969 	preempt_disable();
1970 	xen_mc_flush();
1971 	paravirt_leave_lazy_mmu();
1972 	preempt_enable();
1973 }
1974 
1975 static const struct pv_mmu_ops xen_mmu_ops __initdata = {
1976 	.read_cr2 = xen_read_cr2,
1977 	.write_cr2 = xen_write_cr2,
1978 
1979 	.read_cr3 = xen_read_cr3,
1980 #ifdef CONFIG_X86_32
1981 	.write_cr3 = xen_write_cr3_init,
1982 #else
1983 	.write_cr3 = xen_write_cr3,
1984 #endif
1985 
1986 	.flush_tlb_user = xen_flush_tlb,
1987 	.flush_tlb_kernel = xen_flush_tlb,
1988 	.flush_tlb_single = xen_flush_tlb_single,
1989 	.flush_tlb_others = xen_flush_tlb_others,
1990 
1991 	.pte_update = paravirt_nop,
1992 	.pte_update_defer = paravirt_nop,
1993 
1994 	.pgd_alloc = xen_pgd_alloc,
1995 	.pgd_free = xen_pgd_free,
1996 
1997 	.alloc_pte = xen_alloc_pte_init,
1998 	.release_pte = xen_release_pte_init,
1999 	.alloc_pmd = xen_alloc_pmd_init,
2000 	.release_pmd = xen_release_pmd_init,
2001 
2002 	.set_pte = xen_set_pte_init,
2003 	.set_pte_at = xen_set_pte_at,
2004 	.set_pmd = xen_set_pmd_hyper,
2005 
2006 	.ptep_modify_prot_start = __ptep_modify_prot_start,
2007 	.ptep_modify_prot_commit = __ptep_modify_prot_commit,
2008 
2009 	.pte_val = PV_CALLEE_SAVE(xen_pte_val),
2010 	.pgd_val = PV_CALLEE_SAVE(xen_pgd_val),
2011 
2012 	.make_pte = PV_CALLEE_SAVE(xen_make_pte),
2013 	.make_pgd = PV_CALLEE_SAVE(xen_make_pgd),
2014 
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,
2021 
2022 	.make_pmd = PV_CALLEE_SAVE(xen_make_pmd),
2023 	.pmd_val = PV_CALLEE_SAVE(xen_pmd_val),
2024 
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,
2029 
2030 	.alloc_pud = xen_alloc_pmd_init,
2031 	.release_pud = xen_release_pmd_init,
2032 #endif	/* PAGETABLE_LEVELS == 4 */
2033 
2034 	.activate_mm = xen_activate_mm,
2035 	.dup_mmap = xen_dup_mmap,
2036 	.exit_mmap = xen_exit_mmap,
2037 
2038 	.lazy_mode = {
2039 		.enter = paravirt_enter_lazy_mmu,
2040 		.leave = xen_leave_lazy_mmu,
2041 	},
2042 
2043 	.set_fixmap = xen_set_fixmap,
2044 };
2045 
2046 void __init xen_init_mmu_ops(void)
2047 {
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;
2051 
2052 	memset(dummy_mapping, 0xff, PAGE_SIZE);
2053 }
2054 
2055 /* Protected by xen_reservation_lock. */
2056 #define MAX_CONTIG_ORDER 9 /* 2MB */
2057 static unsigned long discontig_frames[1<<MAX_CONTIG_ORDER];
2058 
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)
2063 {
2064 	int i;
2065 	struct multicall_space mcs;
2066 
2067 	xen_mc_batch();
2068 	for (i = 0; i < (1UL<<order); i++, vaddr += PAGE_SIZE) {
2069 		mcs = __xen_mc_entry(0);
2070 
2071 		if (in_frames)
2072 			in_frames[i] = virt_to_mfn(vaddr);
2073 
2074 		MULTI_update_va_mapping(mcs.mc, vaddr, VOID_PTE, 0);
2075 		set_phys_to_machine(virt_to_pfn(vaddr), INVALID_P2M_ENTRY);
2076 
2077 		if (out_frames)
2078 			out_frames[i] = virt_to_pfn(vaddr);
2079 	}
2080 	xen_mc_issue(0);
2081 }
2082 
2083 /*
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
2086  * mfn.
2087  */
2088 static void xen_remap_exchanged_ptes(unsigned long vaddr, int order,
2089 				     unsigned long *mfns,
2090 				     unsigned long first_mfn)
2091 {
2092 	unsigned i, limit;
2093 	unsigned long mfn;
2094 
2095 	xen_mc_batch();
2096 
2097 	limit = 1u << order;
2098 	for (i = 0; i < limit; i++, vaddr += PAGE_SIZE) {
2099 		struct multicall_space mcs;
2100 		unsigned flags;
2101 
2102 		mcs = __xen_mc_entry(0);
2103 		if (mfns)
2104 			mfn = mfns[i];
2105 		else
2106 			mfn = first_mfn + i;
2107 
2108 		if (i < (limit - 1))
2109 			flags = 0;
2110 		else {
2111 			if (order == 0)
2112 				flags = UVMF_INVLPG | UVMF_ALL;
2113 			else
2114 				flags = UVMF_TLB_FLUSH | UVMF_ALL;
2115 		}
2116 
2117 		MULTI_update_va_mapping(mcs.mc, vaddr,
2118 				mfn_pte(mfn, PAGE_KERNEL), flags);
2119 
2120 		set_phys_to_machine(virt_to_pfn(vaddr), mfn);
2121 	}
2122 
2123 	xen_mc_issue(0);
2124 }
2125 
2126 /*
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.
2130  *
2131  * Returns a success code indicating whether the hypervisor was able to
2132  * satisfy the request or not.
2133  */
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)
2140 {
2141 	long rc;
2142 	int success;
2143 
2144 	struct xen_memory_exchange exchange = {
2145 		.in = {
2146 			.nr_extents   = extents_in,
2147 			.extent_order = order_in,
2148 			.extent_start = pfns_in,
2149 			.domid        = DOMID_SELF
2150 		},
2151 		.out = {
2152 			.nr_extents   = extents_out,
2153 			.extent_order = order_out,
2154 			.extent_start = mfns_out,
2155 			.address_bits = address_bits,
2156 			.domid        = DOMID_SELF
2157 		}
2158 	};
2159 
2160 	BUG_ON(extents_in << order_in != extents_out << order_out);
2161 
2162 	rc = HYPERVISOR_memory_op(XENMEM_exchange, &exchange);
2163 	success = (exchange.nr_exchanged == extents_in);
2164 
2165 	BUG_ON(!success && ((exchange.nr_exchanged != 0) || (rc == 0)));
2166 	BUG_ON(success && (rc != 0));
2167 
2168 	return success;
2169 }
2170 
2171 int xen_create_contiguous_region(unsigned long vstart, unsigned int order,
2172 				 unsigned int address_bits)
2173 {
2174 	unsigned long *in_frames = discontig_frames, out_frame;
2175 	unsigned long  flags;
2176 	int            success;
2177 
2178 	/*
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.
2182 	 */
2183 
2184 	if (xen_feature(XENFEAT_auto_translated_physmap))
2185 		return 0;
2186 
2187 	if (unlikely(order > MAX_CONTIG_ORDER))
2188 		return -ENOMEM;
2189 
2190 	memset((void *) vstart, 0, PAGE_SIZE << order);
2191 
2192 	spin_lock_irqsave(&xen_reservation_lock, flags);
2193 
2194 	/* 1. Zap current PTEs, remembering MFNs. */
2195 	xen_zap_pfn_range(vstart, order, in_frames, NULL);
2196 
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,
2201 				      address_bits);
2202 
2203 	/* 3. Map the new extent in place of old pages. */
2204 	if (success)
2205 		xen_remap_exchanged_ptes(vstart, order, NULL, out_frame);
2206 	else
2207 		xen_remap_exchanged_ptes(vstart, order, in_frames, 0);
2208 
2209 	spin_unlock_irqrestore(&xen_reservation_lock, flags);
2210 
2211 	return success ? 0 : -ENOMEM;
2212 }
2213 EXPORT_SYMBOL_GPL(xen_create_contiguous_region);
2214 
2215 void xen_destroy_contiguous_region(unsigned long vstart, unsigned int order)
2216 {
2217 	unsigned long *out_frames = discontig_frames, in_frame;
2218 	unsigned long  flags;
2219 	int success;
2220 
2221 	if (xen_feature(XENFEAT_auto_translated_physmap))
2222 		return;
2223 
2224 	if (unlikely(order > MAX_CONTIG_ORDER))
2225 		return;
2226 
2227 	memset((void *) vstart, 0, PAGE_SIZE << order);
2228 
2229 	spin_lock_irqsave(&xen_reservation_lock, flags);
2230 
2231 	/* 1. Find start MFN of contiguous extent. */
2232 	in_frame = virt_to_mfn(vstart);
2233 
2234 	/* 2. Zap current PTEs. */
2235 	xen_zap_pfn_range(vstart, order, NULL, out_frames);
2236 
2237 	/* 3. Do the exchange for non-contiguous MFNs. */
2238 	success = xen_exchange_memory(1, order, &in_frame, 1UL << order,
2239 					0, out_frames, 0);
2240 
2241 	/* 4. Map new pages in place of old pages. */
2242 	if (success)
2243 		xen_remap_exchanged_ptes(vstart, order, out_frames, 0);
2244 	else
2245 		xen_remap_exchanged_ptes(vstart, order, NULL, in_frame);
2246 
2247 	spin_unlock_irqrestore(&xen_reservation_lock, flags);
2248 }
2249 EXPORT_SYMBOL_GPL(xen_destroy_contiguous_region);
2250 
2251 #ifdef CONFIG_XEN_PVHVM
2252 static void xen_hvm_exit_mmap(struct mm_struct *mm)
2253 {
2254 	struct xen_hvm_pagetable_dying a;
2255 	int rc;
2256 
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);
2261 }
2262 
2263 static int is_pagetable_dying_supported(void)
2264 {
2265 	struct xen_hvm_pagetable_dying a;
2266 	int rc = 0;
2267 
2268 	a.domid = DOMID_SELF;
2269 	a.gpa = 0x00;
2270 	rc = HYPERVISOR_hvm_op(HVMOP_pagetable_dying, &a);
2271 	if (rc < 0) {
2272 		printk(KERN_DEBUG "HVMOP_pagetable_dying not supported\n");
2273 		return 0;
2274 	}
2275 	return 1;
2276 }
2277 
2278 void __init xen_hvm_init_mmu_ops(void)
2279 {
2280 	if (is_pagetable_dying_supported())
2281 		pv_mmu_ops.exit_mmap = xen_hvm_exit_mmap;
2282 }
2283 #endif
2284 
2285 #define REMAP_BATCH_SIZE 16
2286 
2287 struct remap_data {
2288 	unsigned long mfn;
2289 	pgprot_t prot;
2290 	struct mmu_update *mmu_update;
2291 };
2292 
2293 static int remap_area_mfn_pte_fn(pte_t *ptep, pgtable_t token,
2294 				 unsigned long addr, void *data)
2295 {
2296 	struct remap_data *rmd = data;
2297 	pte_t pte = pte_mkspecial(pfn_pte(rmd->mfn++, rmd->prot));
2298 
2299 	rmd->mmu_update->ptr = arbitrary_virt_to_machine(ptep).maddr;
2300 	rmd->mmu_update->val = pte_val_ma(pte);
2301 	rmd->mmu_update++;
2302 
2303 	return 0;
2304 }
2305 
2306 int xen_remap_domain_mfn_range(struct vm_area_struct *vma,
2307 			       unsigned long addr,
2308 			       unsigned long mfn, int nr,
2309 			       pgprot_t prot, unsigned domid)
2310 {
2311 	struct remap_data rmd;
2312 	struct mmu_update mmu_update[REMAP_BATCH_SIZE];
2313 	int batch;
2314 	unsigned long range;
2315 	int err = 0;
2316 
2317 	prot = __pgprot(pgprot_val(prot) | _PAGE_IOMAP);
2318 
2319 	BUG_ON(!((vma->vm_flags & (VM_PFNMAP | VM_RESERVED | VM_IO)) ==
2320 				(VM_PFNMAP | VM_RESERVED | VM_IO)));
2321 
2322 	rmd.mfn = mfn;
2323 	rmd.prot = prot;
2324 
2325 	while (nr) {
2326 		batch = min(REMAP_BATCH_SIZE, nr);
2327 		range = (unsigned long)batch << PAGE_SHIFT;
2328 
2329 		rmd.mmu_update = mmu_update;
2330 		err = apply_to_page_range(vma->vm_mm, addr, range,
2331 					  remap_area_mfn_pte_fn, &rmd);
2332 		if (err)
2333 			goto out;
2334 
2335 		err = -EFAULT;
2336 		if (HYPERVISOR_mmu_update(mmu_update, batch, NULL, domid) < 0)
2337 			goto out;
2338 
2339 		nr -= batch;
2340 		addr += range;
2341 	}
2342 
2343 	err = 0;
2344 out:
2345 
2346 	flush_tlb_all();
2347 
2348 	return err;
2349 }
2350 EXPORT_SYMBOL_GPL(xen_remap_domain_mfn_range);
2351 
2352 #ifdef CONFIG_XEN_DEBUG_FS
2353 
2354 static struct dentry *d_mmu_debug;
2355 
2356 static int __init xen_mmu_debugfs(void)
2357 {
2358 	struct dentry *d_xen = xen_init_debugfs();
2359 
2360 	if (d_xen == NULL)
2361 		return -ENOMEM;
2362 
2363 	d_mmu_debug = debugfs_create_dir("mmu", d_xen);
2364 
2365 	debugfs_create_u8("zero_stats", 0644, d_mmu_debug, &zero_stats);
2366 
2367 	debugfs_create_u32("pgd_update", 0444, d_mmu_debug, &mmu_stats.pgd_update);
2368 	debugfs_create_u32("pgd_update_pinned", 0444, d_mmu_debug,
2369 			   &mmu_stats.pgd_update_pinned);
2370 	debugfs_create_u32("pgd_update_batched", 0444, d_mmu_debug,
2371 			   &mmu_stats.pgd_update_pinned);
2372 
2373 	debugfs_create_u32("pud_update", 0444, d_mmu_debug, &mmu_stats.pud_update);
2374 	debugfs_create_u32("pud_update_pinned", 0444, d_mmu_debug,
2375 			   &mmu_stats.pud_update_pinned);
2376 	debugfs_create_u32("pud_update_batched", 0444, d_mmu_debug,
2377 			   &mmu_stats.pud_update_pinned);
2378 
2379 	debugfs_create_u32("pmd_update", 0444, d_mmu_debug, &mmu_stats.pmd_update);
2380 	debugfs_create_u32("pmd_update_pinned", 0444, d_mmu_debug,
2381 			   &mmu_stats.pmd_update_pinned);
2382 	debugfs_create_u32("pmd_update_batched", 0444, d_mmu_debug,
2383 			   &mmu_stats.pmd_update_pinned);
2384 
2385 	debugfs_create_u32("pte_update", 0444, d_mmu_debug, &mmu_stats.pte_update);
2386 //	debugfs_create_u32("pte_update_pinned", 0444, d_mmu_debug,
2387 //			   &mmu_stats.pte_update_pinned);
2388 	debugfs_create_u32("pte_update_batched", 0444, d_mmu_debug,
2389 			   &mmu_stats.pte_update_pinned);
2390 
2391 	debugfs_create_u32("mmu_update", 0444, d_mmu_debug, &mmu_stats.mmu_update);
2392 	debugfs_create_u32("mmu_update_extended", 0444, d_mmu_debug,
2393 			   &mmu_stats.mmu_update_extended);
2394 	xen_debugfs_create_u32_array("mmu_update_histo", 0444, d_mmu_debug,
2395 				     mmu_stats.mmu_update_histo, 20);
2396 
2397 	debugfs_create_u32("set_pte_at", 0444, d_mmu_debug, &mmu_stats.set_pte_at);
2398 	debugfs_create_u32("set_pte_at_batched", 0444, d_mmu_debug,
2399 			   &mmu_stats.set_pte_at_batched);
2400 	debugfs_create_u32("set_pte_at_current", 0444, d_mmu_debug,
2401 			   &mmu_stats.set_pte_at_current);
2402 	debugfs_create_u32("set_pte_at_kernel", 0444, d_mmu_debug,
2403 			   &mmu_stats.set_pte_at_kernel);
2404 
2405 	debugfs_create_u32("prot_commit", 0444, d_mmu_debug, &mmu_stats.prot_commit);
2406 	debugfs_create_u32("prot_commit_batched", 0444, d_mmu_debug,
2407 			   &mmu_stats.prot_commit_batched);
2408 
2409 	return 0;
2410 }
2411 fs_initcall(xen_mmu_debugfs);
2412 
2413 #endif	/* CONFIG_XEN_DEBUG_FS */
2414