2 * Copyright 2002 Andi Kleen, SuSE Labs.
3 * Thanks to Ben LaHaise for precious feedback.
5 #include <linux/highmem.h>
6 #include <linux/bootmem.h>
7 #include <linux/module.h>
8 #include <linux/sched.h>
10 #include <linux/interrupt.h>
11 #include <linux/seq_file.h>
12 #include <linux/debugfs.h>
13 #include <linux/pfn.h>
14 #include <linux/percpu.h>
15 #include <linux/gfp.h>
16 #include <linux/pci.h>
19 #include <asm/processor.h>
20 #include <asm/tlbflush.h>
21 #include <asm/sections.h>
22 #include <asm/setup.h>
23 #include <asm/uaccess.h>
24 #include <asm/pgalloc.h>
25 #include <asm/proto.h>
29 * The current flushing context - we pass it instead of 5 arguments:
38 unsigned force_split : 1;
44 * Serialize cpa() (for !DEBUG_PAGEALLOC which uses large identity mappings)
45 * using cpa_lock. So that we don't allow any other cpu, with stale large tlb
46 * entries change the page attribute in parallel to some other cpu
47 * splitting a large page entry along with changing the attribute.
49 static DEFINE_SPINLOCK(cpa_lock);
51 #define CPA_FLUSHTLB 1
53 #define CPA_PAGES_ARRAY 4
56 static unsigned long direct_pages_count[PG_LEVEL_NUM];
58 void update_page_count(int level, unsigned long pages)
60 /* Protect against CPA */
62 direct_pages_count[level] += pages;
63 spin_unlock(&pgd_lock);
66 static void split_page_count(int level)
68 direct_pages_count[level]--;
69 direct_pages_count[level - 1] += PTRS_PER_PTE;
72 void arch_report_meminfo(struct seq_file *m)
74 seq_printf(m, "DirectMap4k: %8lu kB\n",
75 direct_pages_count[PG_LEVEL_4K] << 2);
76 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
77 seq_printf(m, "DirectMap2M: %8lu kB\n",
78 direct_pages_count[PG_LEVEL_2M] << 11);
80 seq_printf(m, "DirectMap4M: %8lu kB\n",
81 direct_pages_count[PG_LEVEL_2M] << 12);
85 seq_printf(m, "DirectMap1G: %8lu kB\n",
86 direct_pages_count[PG_LEVEL_1G] << 20);
90 static inline void split_page_count(int level) { }
95 static inline unsigned long highmap_start_pfn(void)
97 return __pa(_text) >> PAGE_SHIFT;
100 static inline unsigned long highmap_end_pfn(void)
102 return __pa(roundup(_brk_end, PMD_SIZE)) >> PAGE_SHIFT;
107 #ifdef CONFIG_DEBUG_PAGEALLOC
108 # define debug_pagealloc 1
110 # define debug_pagealloc 0
114 within(unsigned long addr, unsigned long start, unsigned long end)
116 return addr >= start && addr < end;
124 * clflush_cache_range - flush a cache range with clflush
125 * @addr: virtual start address
126 * @size: number of bytes to flush
128 * clflush is an unordered instruction which needs fencing with mfence
129 * to avoid ordering issues.
131 void clflush_cache_range(void *vaddr, unsigned int size)
133 void *vend = vaddr + size - 1;
137 for (; vaddr < vend; vaddr += boot_cpu_data.x86_clflush_size)
140 * Flush any possible final partial cacheline:
146 EXPORT_SYMBOL_GPL(clflush_cache_range);
148 static void __cpa_flush_all(void *arg)
150 unsigned long cache = (unsigned long)arg;
153 * Flush all to work around Errata in early athlons regarding
154 * large page flushing.
158 if (cache && boot_cpu_data.x86 >= 4)
162 static void cpa_flush_all(unsigned long cache)
164 BUG_ON(irqs_disabled());
166 on_each_cpu(__cpa_flush_all, (void *) cache, 1);
169 static void __cpa_flush_range(void *arg)
172 * We could optimize that further and do individual per page
173 * tlb invalidates for a low number of pages. Caveat: we must
174 * flush the high aliases on 64bit as well.
179 static void cpa_flush_range(unsigned long start, int numpages, int cache)
181 unsigned int i, level;
184 BUG_ON(irqs_disabled());
185 WARN_ON(PAGE_ALIGN(start) != start);
187 on_each_cpu(__cpa_flush_range, NULL, 1);
193 * We only need to flush on one CPU,
194 * clflush is a MESI-coherent instruction that
195 * will cause all other CPUs to flush the same
198 for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
199 pte_t *pte = lookup_address(addr, &level);
202 * Only flush present addresses:
204 if (pte && (__pte_val(*pte) & _PAGE_PRESENT))
205 clflush_cache_range((void *) addr, PAGE_SIZE);
209 static void cpa_flush_array(unsigned long *start, int numpages, int cache,
210 int in_flags, struct page **pages)
212 unsigned int i, level;
213 unsigned long do_wbinvd = cache && numpages >= 1024; /* 4M threshold */
215 BUG_ON(irqs_disabled());
217 on_each_cpu(__cpa_flush_all, (void *) do_wbinvd, 1);
219 if (!cache || do_wbinvd)
223 * We only need to flush on one CPU,
224 * clflush is a MESI-coherent instruction that
225 * will cause all other CPUs to flush the same
228 for (i = 0; i < numpages; i++) {
232 if (in_flags & CPA_PAGES_ARRAY)
233 addr = (unsigned long)page_address(pages[i]);
237 pte = lookup_address(addr, &level);
240 * Only flush present addresses:
242 if (pte && (__pte_val(*pte) & _PAGE_PRESENT))
243 clflush_cache_range((void *)addr, PAGE_SIZE);
248 * Certain areas of memory on x86 require very specific protection flags,
249 * for example the BIOS area or kernel text. Callers don't always get this
250 * right (again, ioremap() on BIOS memory is not uncommon) so this function
251 * checks and fixes these known static required protection bits.
253 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
256 pgprot_t forbidden = __pgprot(0);
259 * The BIOS area between 640k and 1Mb needs to be executable for
260 * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
262 #ifdef CONFIG_PCI_BIOS
263 if (pcibios_enabled && within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
264 pgprot_val(forbidden) |= _PAGE_NX;
268 * The kernel text needs to be executable for obvious reasons
269 * Does not cover __inittext since that is gone later on. On
270 * 64bit we do not enforce !NX on the low mapping
272 if (within(address, (unsigned long)_text, (unsigned long)_etext))
273 pgprot_val(forbidden) |= _PAGE_NX;
276 * The .rodata section needs to be read-only. Using the pfn
277 * catches all aliases.
279 if (within(pfn, __pa((unsigned long)__start_rodata) >> PAGE_SHIFT,
280 __pa((unsigned long)__end_rodata) >> PAGE_SHIFT))
281 pgprot_val(forbidden) |= _PAGE_RW;
283 #if defined(CONFIG_X86_64) && defined(CONFIG_DEBUG_RODATA) && !defined(CONFIG_XEN)
285 * Once the kernel maps the text as RO (kernel_set_to_readonly is set),
286 * kernel text mappings for the large page aligned text, rodata sections
287 * will be always read-only. For the kernel identity mappings covering
288 * the holes caused by this alignment can be anything that user asks.
290 * This will preserve the large page mappings for kernel text/data
293 if (kernel_set_to_readonly &&
294 within(address, (unsigned long)_text,
295 (unsigned long)__end_rodata_hpage_align)) {
299 * Don't enforce the !RW mapping for the kernel text mapping,
300 * if the current mapping is already using small page mapping.
301 * No need to work hard to preserve large page mappings in this
304 * This also fixes the Linux Xen paravirt guest boot failure
305 * (because of unexpected read-only mappings for kernel identity
306 * mappings). In this paravirt guest case, the kernel text
307 * mapping and the kernel identity mapping share the same
308 * page-table pages. Thus we can't really use different
309 * protections for the kernel text and identity mappings. Also,
310 * these shared mappings are made of small page mappings.
311 * Thus this don't enforce !RW mapping for small page kernel
312 * text mapping logic will help Linux Xen parvirt guest boot
315 if (lookup_address(address, &level) && (level != PG_LEVEL_4K))
316 pgprot_val(forbidden) |= _PAGE_RW;
320 prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
326 * Lookup the page table entry for a virtual address. Return a pointer
327 * to the entry and the level of the mapping.
329 * Note: We return pud and pmd either when the entry is marked large
330 * or when the present bit is not set. Otherwise we would return a
331 * pointer to a nonexisting mapping.
333 pte_t *lookup_address(unsigned long address, unsigned int *level)
335 pgd_t *pgd = pgd_offset_k(address);
339 *level = PG_LEVEL_NONE;
344 pud = pud_offset(pgd, address);
348 *level = PG_LEVEL_1G;
349 if (pud_large(*pud) || !pud_present(*pud))
352 pmd = pmd_offset(pud, address);
356 *level = PG_LEVEL_2M;
357 if (pmd_large(*pmd) || !pmd_present(*pmd))
360 *level = PG_LEVEL_4K;
362 return pte_offset_kernel(pmd, address);
364 EXPORT_SYMBOL_GPL(lookup_address);
367 * Set the new pmd in all the pgds we know about:
369 static void __set_pmd_pte(pte_t *kpte, unsigned long address,
370 unsigned int level, pte_t pte)
375 xen_l2_entry_update((pmd_t *)kpte, __pmd_ma(__pte_val(pte)));
379 xen_l3_entry_update((pud_t *)kpte, __pud_ma(__pte_val(pte)));
386 if (!SHARED_KERNEL_PMD) {
389 list_for_each_entry(page, &pgd_list, lru) {
394 pgd = (pgd_t *)page_address(page) + pgd_index(address);
395 pud = pud_offset(pgd, address);
396 pmd = pmd_offset(pud, address);
397 xen_l2_entry_update(pmd, __pmd_ma(__pte_val(pte)));
404 try_preserve_large_page(pte_t *kpte, unsigned long address,
405 struct cpa_data *cpa)
407 unsigned long nextpage_addr, numpages, pmask, psize, addr, pfn;
408 pte_t new_pte, old_pte, *tmp;
409 pgprot_t old_prot, new_prot, req_prot;
413 if (cpa->force_split)
416 spin_lock(&pgd_lock);
418 * Check for races, another CPU might have split this page
421 tmp = lookup_address(address, &level);
427 psize = PMD_PAGE_SIZE;
428 pmask = PMD_PAGE_MASK;
432 psize = PUD_PAGE_SIZE;
433 pmask = PUD_PAGE_MASK;
442 * Calculate the number of pages, which fit into this large
443 * page starting at address:
445 nextpage_addr = (address + psize) & pmask;
446 numpages = (nextpage_addr - address) >> PAGE_SHIFT;
447 if (numpages < cpa->numpages)
448 cpa->numpages = numpages;
451 * We are safe now. Check whether the new pgprot is the same:
454 old_prot = new_prot = req_prot = pte_pgprot(old_pte);
456 pgprot_val(req_prot) &= ~pgprot_val(cpa->mask_clr);
457 pgprot_val(req_prot) |= pgprot_val(cpa->mask_set);
460 * old_pte points to the large page base address. So we need
461 * to add the offset of the virtual address:
463 pfn = pte_pfn(old_pte) + ((address & (psize - 1)) >> PAGE_SHIFT);
466 new_prot = static_protections(req_prot, address, pfn);
469 * We need to check the full range, whether
470 * static_protection() requires a different pgprot for one of
471 * the pages in the range we try to preserve:
473 addr = address & pmask;
474 pfn = pte_pfn(old_pte);
475 for (i = 0; i < (psize >> PAGE_SHIFT) && pfn < max_mapnr;
476 i++, addr += PAGE_SIZE, pfn++) {
477 pgprot_t chk_prot = static_protections(req_prot, addr, pfn);
479 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
484 * If there are no changes, return. maxpages has been updated
487 if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
493 * We need to change the attributes. Check, whether we can
494 * change the large page in one go. We request a split, when
495 * the address is not aligned and the number of pages is
496 * smaller than the number of pages in the large page. Note
497 * that we limited the number of possible pages already to
498 * the number of pages in the large page.
500 if (address == (address & pmask) && cpa->numpages == (psize >> PAGE_SHIFT)) {
502 * The address is aligned and the number of pages
503 * covers the full page.
505 new_pte = pfn_pte_ma(__pte_mfn(old_pte), canon_pgprot(new_prot));
506 __set_pmd_pte(kpte, address, level, new_pte);
507 cpa->flags |= CPA_FLUSHTLB;
512 spin_unlock(&pgd_lock);
517 static int split_large_page(pte_t *kpte, unsigned long address)
519 unsigned long mfn, mfninc = 1;
520 unsigned int i, level;
525 if (!debug_pagealloc)
526 spin_unlock(&cpa_lock);
527 base = alloc_pages(GFP_KERNEL | __GFP_NOTRACK, 0);
528 if (!debug_pagealloc)
529 spin_lock(&cpa_lock);
533 spin_lock(&pgd_lock);
535 * Check for races, another CPU might have split this page
538 tmp = lookup_address(address, &level);
542 pbase = (pte_t *)page_address(base);
543 paravirt_alloc_pte(&init_mm, page_to_pfn(base));
544 ref_prot = pte_pgprot(pte_clrhuge(*kpte));
546 * If we ever want to utilize the PAT bit, we need to
547 * update this function to make sure it's converted from
548 * bit 12 to bit 7 when we cross from the 2MB level to
551 WARN_ON_ONCE(pgprot_val(ref_prot) & _PAGE_PAT_LARGE);
554 if (level == PG_LEVEL_1G) {
555 mfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
556 pgprot_val(ref_prot) |= _PAGE_PSE;
560 if (address >= (unsigned long)__va(0) &&
561 address < (unsigned long)__va(max_low_pfn_mapped << PAGE_SHIFT))
562 split_page_count(level);
565 if (address >= (unsigned long)__va(1UL<<32) &&
566 address < (unsigned long)__va(max_pfn_mapped << PAGE_SHIFT))
567 split_page_count(level);
571 * Get the target mfn from the original entry:
573 mfn = __pte_mfn(*kpte);
574 for (i = 0; i < PTRS_PER_PTE; i++, mfn += mfninc)
575 set_pte(&pbase[i], pfn_pte_ma(mfn, ref_prot));
578 * Install the new, split up pagetable.
580 * We use the standard kernel pagetable protections for the new
581 * pagetable protections, the actual ptes set above control the
582 * primary protection behavior:
584 if (!xen_feature(XENFEAT_writable_page_tables) &&
585 HYPERVISOR_update_va_mapping((unsigned long)pbase,
586 mk_pte(base, PAGE_KERNEL_RO), 0))
588 __set_pmd_pte(kpte, address, level, mk_pte(base, __pgprot(_KERNPG_TABLE)));
591 * Intel Atom errata AAH41 workaround.
593 * The real fix should be in hw or in a microcode update, but
594 * we also probabilistically try to reduce the window of having
595 * a large TLB mixed with 4K TLBs while instruction fetches are
604 * If we dropped out via the lookup_address check under
605 * pgd_lock then stick the page back into the pool:
609 spin_unlock(&pgd_lock);
614 static int __cpa_process_fault(struct cpa_data *cpa, unsigned long vaddr,
618 * Ignore all non primary paths.
624 * Ignore the NULL PTE for kernel identity mapping, as it is expected
626 * Also set numpages to '1' indicating that we processed cpa req for
627 * one virtual address page and its pfn. TBD: numpages can be set based
628 * on the initial value and the level returned by lookup_address().
630 if (within(vaddr, PAGE_OFFSET,
631 PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
633 cpa->pfn = __pa(vaddr) >> PAGE_SHIFT;
636 WARN(1, KERN_WARNING "CPA: called for zero pte. "
637 "vaddr = %lx cpa->vaddr = %lx\n", vaddr,
644 static int __change_page_attr(struct cpa_data *cpa, int primary)
646 unsigned long address;
649 pte_t *kpte, old_pte;
651 if (cpa->flags & CPA_PAGES_ARRAY) {
652 struct page *page = cpa->pages[cpa->curpage];
653 if (unlikely(PageHighMem(page)))
655 address = (unsigned long)page_address(page);
656 } else if (cpa->flags & CPA_ARRAY)
657 address = cpa->vaddr[cpa->curpage];
659 address = *cpa->vaddr;
661 kpte = lookup_address(address, &level);
663 return __cpa_process_fault(cpa, address, primary);
666 if (!__pte_val(old_pte))
667 return __cpa_process_fault(cpa, address, primary);
669 if (level == PG_LEVEL_4K) {
671 pgprot_t new_prot = pte_pgprot(old_pte);
672 unsigned long mfn = __pte_mfn(old_pte);
674 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
675 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
677 new_prot = static_protections(new_prot, address,
678 mfn_to_local_pfn(mfn));
681 * We need to keep the mfn from the existing PTE,
682 * after all we're only going to change it's attributes
683 * not the memory it points to
685 new_pte = pfn_pte_ma(mfn, canon_pgprot(new_prot));
686 cpa->pfn = mfn_to_local_pfn(mfn);
688 * Do we really change anything ?
690 if (__pte_val(old_pte) != __pte_val(new_pte)) {
693 u.ptr = virt_to_machine(kpte);
694 u.val = __pte_val(new_pte);
695 WARN_ON_ONCE(arch_use_lazy_mmu_mode());
697 err = HYPERVISOR_mmu_update(&u, 1, NULL,
704 BUG_ON(!((pgprot_val(cpa->mask_set) |
705 pgprot_val(cpa->mask_clr)) &
707 if (hypervisor_oom())
714 cpa->flags |= CPA_FLUSHTLB;
721 * Check, whether we can keep the large page intact
722 * and just change the pte:
724 do_split = try_preserve_large_page(kpte, address, cpa);
726 * When the range fits into the existing large page,
727 * return. cp->numpages and cpa->tlbflush have been updated in
734 * We have to split the large page:
736 err = split_large_page(kpte, address);
739 * Do a global flush tlb after splitting the large page
740 * and before we do the actual change page attribute in the PTE.
742 * With out this, we violate the TLB application note, that says
743 * "The TLBs may contain both ordinary and large-page
744 * translations for a 4-KByte range of linear addresses. This
745 * may occur if software modifies the paging structures so that
746 * the page size used for the address range changes. If the two
747 * translations differ with respect to page frame or attributes
748 * (e.g., permissions), processor behavior is undefined and may
749 * be implementation-specific."
751 * We do this global tlb flush inside the cpa_lock, so that we
752 * don't allow any other cpu, with stale tlb entries change the
753 * page attribute in parallel, that also falls into the
754 * just split large page entry.
763 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
765 static int cpa_process_alias(struct cpa_data *cpa)
767 struct cpa_data alias_cpa;
768 unsigned long laddr = (unsigned long)__va(cpa->pfn << PAGE_SHIFT);
772 if (cpa->pfn >= max_pfn_mapped)
776 if (cpa->pfn >= max_low_pfn_mapped && cpa->pfn < (1UL<<(32-PAGE_SHIFT)))
780 * No need to redo, when the primary call touched the direct
783 if (cpa->flags & CPA_PAGES_ARRAY) {
784 struct page *page = cpa->pages[cpa->curpage];
785 if (unlikely(PageHighMem(page)))
787 vaddr = (unsigned long)page_address(page);
788 } else if (cpa->flags & CPA_ARRAY)
789 vaddr = cpa->vaddr[cpa->curpage];
793 if (!(within(vaddr, PAGE_OFFSET,
794 PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT)))) {
797 alias_cpa.vaddr = &laddr;
798 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
800 ret = __change_page_attr_set_clr(&alias_cpa, 0);
807 * If the primary call didn't touch the high mapping already
808 * and the physical address is inside the kernel map, we need
809 * to touch the high mapped kernel as well:
811 if (!within(vaddr, (unsigned long)_text, _brk_end) &&
812 within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn())) {
813 unsigned long temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) +
816 alias_cpa.vaddr = &temp_cpa_vaddr;
817 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
820 * The high mapping range is imprecise, so ignore the
823 __change_page_attr_set_clr(&alias_cpa, 0);
830 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
832 int ret, numpages = cpa->numpages;
836 * Store the remaining nr of pages for the large page
837 * preservation check.
839 cpa->numpages = numpages;
840 /* for array changes, we can't use large page */
841 if (cpa->flags & (CPA_ARRAY | CPA_PAGES_ARRAY))
844 if (!debug_pagealloc)
845 spin_lock(&cpa_lock);
846 ret = __change_page_attr(cpa, checkalias);
847 if (!debug_pagealloc)
848 spin_unlock(&cpa_lock);
853 ret = cpa_process_alias(cpa);
859 * Adjust the number of pages with the result of the
860 * CPA operation. Either a large page has been
861 * preserved or a single page update happened.
863 BUG_ON(cpa->numpages > numpages);
864 numpages -= cpa->numpages;
865 if (cpa->flags & (CPA_PAGES_ARRAY | CPA_ARRAY))
868 *cpa->vaddr += cpa->numpages * PAGE_SIZE;
874 static inline int cache_attr(pgprot_t attr)
876 return pgprot_val(attr) &
877 (_PAGE_PAT | _PAGE_PAT_LARGE | _PAGE_PWT | _PAGE_PCD);
880 static int change_page_attr_set_clr(unsigned long *addr, int numpages,
881 pgprot_t mask_set, pgprot_t mask_clr,
882 int force_split, int in_flag,
886 int ret, cache, checkalias;
887 unsigned long baddr = 0;
890 * Check, if we are requested to change a not supported
893 mask_set = canon_pgprot(mask_set);
894 mask_clr = canon_pgprot(mask_clr);
895 if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
898 /* Ensure we are PAGE_SIZE aligned */
899 if (in_flag & CPA_ARRAY) {
901 for (i = 0; i < numpages; i++) {
902 if (addr[i] & ~PAGE_MASK) {
903 addr[i] &= PAGE_MASK;
907 } else if (!(in_flag & CPA_PAGES_ARRAY)) {
909 * in_flag of CPA_PAGES_ARRAY implies it is aligned.
910 * No need to cehck in that case
912 if (*addr & ~PAGE_MASK) {
915 * People should not be passing in unaligned addresses:
920 * Save address for cache flush. *addr is modified in the call
921 * to __change_page_attr_set_clr() below.
926 /* Must avoid aliasing mappings in the highmem code */
933 cpa.numpages = numpages;
934 cpa.mask_set = mask_set;
935 cpa.mask_clr = mask_clr;
938 cpa.force_split = force_split;
940 if (in_flag & (CPA_ARRAY | CPA_PAGES_ARRAY))
941 cpa.flags |= in_flag;
943 /* No alias checking for _NX bit modifications */
944 checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
946 ret = __change_page_attr_set_clr(&cpa, checkalias);
949 * Check whether we really changed something:
951 if (!(cpa.flags & CPA_FLUSHTLB))
955 * No need to flush, when we did not set any of the caching
958 cache = cache_attr(mask_set);
961 * On success we use clflush, when the CPU supports it to
962 * avoid the wbindv. If the CPU does not support it and in the
963 * error case we fall back to cpa_flush_all (which uses
966 if (!ret && cpu_has_clflush) {
967 if (cpa.flags & (CPA_PAGES_ARRAY | CPA_ARRAY)) {
968 cpa_flush_array(addr, numpages, cache,
971 cpa_flush_range(baddr, numpages, cache);
973 cpa_flush_all(cache);
979 static inline int change_page_attr_set(unsigned long *addr, int numpages,
980 pgprot_t mask, int array)
982 return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
983 (array ? CPA_ARRAY : 0), NULL);
986 static inline int change_page_attr_clear(unsigned long *addr, int numpages,
987 pgprot_t mask, int array)
989 return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
990 (array ? CPA_ARRAY : 0), NULL);
993 static inline int cpa_set_pages_array(struct page **pages, int numpages,
996 return change_page_attr_set_clr(NULL, numpages, mask, __pgprot(0), 0,
997 CPA_PAGES_ARRAY, pages);
1000 static inline int cpa_clear_pages_array(struct page **pages, int numpages,
1003 return change_page_attr_set_clr(NULL, numpages, __pgprot(0), mask, 0,
1004 CPA_PAGES_ARRAY, pages);
1008 static void _free_memtype(u64 pstart, u64 pend)
1010 u64 pa = pstart &= __PHYSICAL_MASK;
1011 u64 ma = phys_to_machine(pa);
1013 while ((pa += PAGE_SIZE) < pend) {
1014 if (phys_to_machine(pa) != ma + (pa - pstart)) {
1015 free_memtype(ma, ma + (pa - pstart));
1017 ma = phys_to_machine(pa);
1020 free_memtype(ma, ma + (pend - pstart));
1022 #define free_memtype _free_memtype
1024 static int _reserve_memtype(u64 pstart, u64 pend, unsigned long req_type)
1026 u64 pcur = pstart &= __PHYSICAL_MASK, pa = pcur;
1027 u64 ma = phys_to_machine(pa);
1030 while ((pa += PAGE_SIZE) < pend) {
1031 if (phys_to_machine(pa) != ma + (pa - pcur)) {
1032 rc = reserve_memtype(ma, ma + (pa - pcur),
1037 ma = phys_to_machine(pa);
1041 rc = reserve_memtype(ma, ma + (pend - pcur), req_type, NULL);
1043 if (unlikely(!rc) && pstart < pcur)
1044 _free_memtype(pstart, pcur);
1048 #define reserve_memtype(s, e, r, n) \
1049 _reserve_memtype(s, e, BUILD_BUG_ON_ZERO(n) ?: (r))
1052 int _set_memory_uc(unsigned long addr, int numpages)
1055 * for now UC MINUS. see comments in ioremap_nocache()
1057 return change_page_attr_set(&addr, numpages,
1058 __pgprot(_PAGE_CACHE_UC_MINUS), 0);
1061 int set_memory_uc(unsigned long addr, int numpages)
1066 * for now UC MINUS. see comments in ioremap_nocache()
1068 ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1069 _PAGE_CACHE_UC_MINUS, NULL);
1073 ret = _set_memory_uc(addr, numpages);
1080 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1084 EXPORT_SYMBOL(set_memory_uc);
1086 static int _set_memory_array(unsigned long *addr, int addrinarray,
1087 unsigned long new_type)
1093 * for now UC MINUS. see comments in ioremap_nocache()
1095 for (i = 0; i < addrinarray; i++) {
1096 ret = reserve_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE,
1102 ret = change_page_attr_set(addr, addrinarray,
1103 __pgprot(_PAGE_CACHE_UC_MINUS), 1);
1105 if (!ret && new_type == _PAGE_CACHE_WC)
1106 ret = change_page_attr_set_clr(addr, addrinarray,
1107 __pgprot(_PAGE_CACHE_WC),
1108 __pgprot(_PAGE_CACHE_MASK),
1109 0, CPA_ARRAY, NULL);
1116 for (j = 0; j < i; j++)
1117 free_memtype(__pa(addr[j]), __pa(addr[j]) + PAGE_SIZE);
1122 int set_memory_array_uc(unsigned long *addr, int addrinarray)
1124 return _set_memory_array(addr, addrinarray, _PAGE_CACHE_UC_MINUS);
1126 EXPORT_SYMBOL(set_memory_array_uc);
1128 int set_memory_array_wc(unsigned long *addr, int addrinarray)
1130 return _set_memory_array(addr, addrinarray, _PAGE_CACHE_WC);
1132 EXPORT_SYMBOL(set_memory_array_wc);
1134 int _set_memory_wc(unsigned long addr, int numpages)
1137 unsigned long addr_copy = addr;
1139 ret = change_page_attr_set(&addr, numpages,
1140 __pgprot(_PAGE_CACHE_UC_MINUS), 0);
1142 ret = change_page_attr_set_clr(&addr_copy, numpages,
1143 __pgprot(_PAGE_CACHE_WC),
1144 __pgprot(_PAGE_CACHE_MASK),
1150 int set_memory_wc(unsigned long addr, int numpages)
1155 return set_memory_uc(addr, numpages);
1157 ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1158 _PAGE_CACHE_WC, NULL);
1162 ret = _set_memory_wc(addr, numpages);
1169 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1173 EXPORT_SYMBOL(set_memory_wc);
1175 int _set_memory_wb(unsigned long addr, int numpages)
1177 return change_page_attr_clear(&addr, numpages,
1178 __pgprot(_PAGE_CACHE_MASK), 0);
1181 int set_memory_wb(unsigned long addr, int numpages)
1185 ret = _set_memory_wb(addr, numpages);
1189 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1192 EXPORT_SYMBOL(set_memory_wb);
1194 int set_memory_array_wb(unsigned long *addr, int addrinarray)
1199 ret = change_page_attr_clear(addr, addrinarray,
1200 __pgprot(_PAGE_CACHE_MASK), 1);
1204 for (i = 0; i < addrinarray; i++)
1205 free_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE);
1209 EXPORT_SYMBOL(set_memory_array_wb);
1211 int set_memory_x(unsigned long addr, int numpages)
1213 if (!(__supported_pte_mask & _PAGE_NX))
1216 return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1218 EXPORT_SYMBOL(set_memory_x);
1220 int set_memory_nx(unsigned long addr, int numpages)
1222 if (!(__supported_pte_mask & _PAGE_NX))
1225 return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1227 EXPORT_SYMBOL(set_memory_nx);
1229 int set_memory_ro(unsigned long addr, int numpages)
1231 return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1233 EXPORT_SYMBOL_GPL(set_memory_ro);
1235 int set_memory_rw(unsigned long addr, int numpages)
1237 return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
1239 EXPORT_SYMBOL_GPL(set_memory_rw);
1241 int set_memory_np(unsigned long addr, int numpages)
1243 return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
1246 int set_memory_4k(unsigned long addr, int numpages)
1248 return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1249 __pgprot(0), 1, 0, NULL);
1252 int set_pages_uc(struct page *page, int numpages)
1254 unsigned long addr = (unsigned long)page_address(page);
1256 return set_memory_uc(addr, numpages);
1258 EXPORT_SYMBOL(set_pages_uc);
1260 static int _set_pages_array(struct page **pages, int addrinarray,
1261 unsigned long new_type)
1263 unsigned long start;
1269 for (i = 0; i < addrinarray; i++) {
1270 if (PageHighMem(pages[i]))
1272 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1273 end = start + PAGE_SIZE;
1274 if (reserve_memtype(start, end, new_type, NULL))
1278 ret = cpa_set_pages_array(pages, addrinarray,
1279 __pgprot(_PAGE_CACHE_UC_MINUS));
1280 if (!ret && new_type == _PAGE_CACHE_WC)
1281 ret = change_page_attr_set_clr(NULL, addrinarray,
1282 __pgprot(_PAGE_CACHE_WC),
1283 __pgprot(_PAGE_CACHE_MASK),
1284 0, CPA_PAGES_ARRAY, pages);
1287 return 0; /* Success */
1290 for (i = 0; i < free_idx; i++) {
1291 if (PageHighMem(pages[i]))
1293 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1294 end = start + PAGE_SIZE;
1295 free_memtype(start, end);
1300 int set_pages_array_uc(struct page **pages, int addrinarray)
1302 return _set_pages_array(pages, addrinarray, _PAGE_CACHE_UC_MINUS);
1304 EXPORT_SYMBOL(set_pages_array_uc);
1306 int set_pages_array_wc(struct page **pages, int addrinarray)
1308 return _set_pages_array(pages, addrinarray, _PAGE_CACHE_WC);
1310 EXPORT_SYMBOL(set_pages_array_wc);
1312 int set_pages_wb(struct page *page, int numpages)
1314 unsigned long addr = (unsigned long)page_address(page);
1316 return set_memory_wb(addr, numpages);
1318 EXPORT_SYMBOL(set_pages_wb);
1320 int set_pages_array_wb(struct page **pages, int addrinarray)
1323 unsigned long start;
1327 retval = cpa_clear_pages_array(pages, addrinarray,
1328 __pgprot(_PAGE_CACHE_MASK));
1332 for (i = 0; i < addrinarray; i++) {
1333 if (PageHighMem(pages[i]))
1335 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
1336 end = start + PAGE_SIZE;
1337 free_memtype(start, end);
1342 EXPORT_SYMBOL(set_pages_array_wb);
1344 int set_pages_x(struct page *page, int numpages)
1346 unsigned long addr = (unsigned long)page_address(page);
1348 return set_memory_x(addr, numpages);
1350 EXPORT_SYMBOL(set_pages_x);
1352 int set_pages_nx(struct page *page, int numpages)
1354 unsigned long addr = (unsigned long)page_address(page);
1356 return set_memory_nx(addr, numpages);
1358 EXPORT_SYMBOL(set_pages_nx);
1360 int set_pages_ro(struct page *page, int numpages)
1362 unsigned long addr = (unsigned long)page_address(page);
1364 return set_memory_ro(addr, numpages);
1367 int set_pages_rw(struct page *page, int numpages)
1369 unsigned long addr = (unsigned long)page_address(page);
1371 return set_memory_rw(addr, numpages);
1374 #ifdef CONFIG_DEBUG_PAGEALLOC
1376 static int __set_pages_p(struct page *page, int numpages)
1378 unsigned long tempaddr = (unsigned long) page_address(page);
1379 struct cpa_data cpa = { .vaddr = &tempaddr,
1380 .numpages = numpages,
1381 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1382 .mask_clr = __pgprot(0),
1386 * No alias checking needed for setting present flag. otherwise,
1387 * we may need to break large pages for 64-bit kernel text
1388 * mappings (this adds to complexity if we want to do this from
1389 * atomic context especially). Let's keep it simple!
1391 return __change_page_attr_set_clr(&cpa, 0);
1394 static int __set_pages_np(struct page *page, int numpages)
1396 unsigned long tempaddr = (unsigned long) page_address(page);
1397 struct cpa_data cpa = { .vaddr = &tempaddr,
1398 .numpages = numpages,
1399 .mask_set = __pgprot(0),
1400 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
1404 * No alias checking needed for setting not present flag. otherwise,
1405 * we may need to break large pages for 64-bit kernel text
1406 * mappings (this adds to complexity if we want to do this from
1407 * atomic context especially). Let's keep it simple!
1409 return __change_page_attr_set_clr(&cpa, 0);
1412 void kernel_map_pages(struct page *page, int numpages, int enable)
1414 if (PageHighMem(page))
1417 debug_check_no_locks_freed(page_address(page),
1418 numpages * PAGE_SIZE);
1422 * The return value is ignored as the calls cannot fail.
1423 * Large pages for identity mappings are not used at boot time
1424 * and hence no memory allocations during large page split.
1427 __set_pages_p(page, numpages);
1429 __set_pages_np(page, numpages);
1432 * We should perform an IPI and flush all tlbs,
1433 * but that can deadlock->flush only current cpu:
1438 #ifdef CONFIG_HIBERNATION
1440 bool kernel_page_present(struct page *page)
1445 if (PageHighMem(page))
1448 pte = lookup_address((unsigned long)page_address(page), &level);
1449 return (__pte_val(*pte) & _PAGE_PRESENT);
1452 #endif /* CONFIG_HIBERNATION */
1454 #endif /* CONFIG_DEBUG_PAGEALLOC */
1456 static inline int in_secondary_range(unsigned long va)
1458 #ifdef CONFIG_X86_64
1459 return va >= VMALLOC_START && va < VMALLOC_END;
1461 return va >= (unsigned long)high_memory;
1465 static void __make_page_readonly(unsigned long va)
1470 pte = lookup_address(va, &level);
1471 BUG_ON(!pte || level != PG_LEVEL_4K);
1472 if (HYPERVISOR_update_va_mapping(va, pte_wrprotect(*pte), 0))
1474 if (in_secondary_range(va)) {
1475 unsigned long pfn = pte_pfn(*pte);
1477 #ifdef CONFIG_HIGHMEM
1478 if (pfn >= highstart_pfn)
1479 kmap_flush_unused(); /* flush stale writable kmaps */
1482 __make_page_readonly((unsigned long)__va(pfn << PAGE_SHIFT));
1486 static void __make_page_writable(unsigned long va)
1491 pte = lookup_address(va, &level);
1492 BUG_ON(!pte || level != PG_LEVEL_4K);
1493 if (HYPERVISOR_update_va_mapping(va, pte_mkwrite(*pte), UVMF_INVLPG))
1495 if (in_secondary_range(va)) {
1496 unsigned long pfn = pte_pfn(*pte);
1498 #ifdef CONFIG_HIGHMEM
1499 if (pfn < highstart_pfn)
1501 __make_page_writable((unsigned long)__va(pfn << PAGE_SHIFT));
1505 void make_page_readonly(void *va, unsigned int feature)
1507 if (!xen_feature(feature))
1508 __make_page_readonly((unsigned long)va);
1511 void make_page_writable(void *va, unsigned int feature)
1513 if (!xen_feature(feature))
1514 __make_page_writable((unsigned long)va);
1517 void make_pages_readonly(void *va, unsigned int nr, unsigned int feature)
1521 if (xen_feature(feature))
1524 for (addr = (unsigned long)va; nr--; addr += PAGE_SIZE)
1525 __make_page_readonly(addr);
1528 void make_pages_writable(void *va, unsigned int nr, unsigned int feature)
1532 if (xen_feature(feature))
1535 for (addr = (unsigned long)va; nr--; addr += PAGE_SIZE)
1536 __make_page_writable(addr);
1540 * The testcases use internal knowledge of the implementation that shouldn't
1541 * be exposed to the rest of the kernel. Include these directly here.
1543 #ifdef CONFIG_CPA_DEBUG
1544 #include "pageattr-test.c"