1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright 2002 Andi Kleen, SuSE Labs.
4 * Thanks to Ben LaHaise for precious feedback.
6 #include <linux/highmem.h>
7 #include <linux/memblock.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>
17 #include <linux/vmalloc.h>
18 #include <linux/libnvdimm.h>
20 #include <asm/e820/api.h>
21 #include <asm/processor.h>
22 #include <asm/tlbflush.h>
23 #include <asm/sections.h>
24 #include <asm/setup.h>
25 #include <linux/uaccess.h>
26 #include <asm/pgalloc.h>
27 #include <asm/proto.h>
28 #include <asm/memtype.h>
29 #include <asm/set_memory.h>
31 #include "../mm_internal.h"
34 * The current flushing context - we pass it instead of 5 arguments:
41 unsigned long numpages;
42 unsigned long curpage;
45 unsigned int force_split : 1,
46 force_static_prot : 1,
57 static const int cpa_warn_level = CPA_PROTECT;
60 * Serialize cpa() (for !DEBUG_PAGEALLOC which uses large identity mappings)
61 * using cpa_lock. So that we don't allow any other cpu, with stale large tlb
62 * entries change the page attribute in parallel to some other cpu
63 * splitting a large page entry along with changing the attribute.
65 static DEFINE_SPINLOCK(cpa_lock);
67 #define CPA_FLUSHTLB 1
69 #define CPA_PAGES_ARRAY 4
70 #define CPA_NO_CHECK_ALIAS 8 /* Do not search for aliases */
73 static unsigned long direct_pages_count[PG_LEVEL_NUM];
75 void update_page_count(int level, unsigned long pages)
77 /* Protect against CPA */
79 direct_pages_count[level] += pages;
80 spin_unlock(&pgd_lock);
83 static void split_page_count(int level)
85 if (direct_pages_count[level] == 0)
88 direct_pages_count[level]--;
89 direct_pages_count[level - 1] += PTRS_PER_PTE;
92 void arch_report_meminfo(struct seq_file *m)
94 seq_printf(m, "DirectMap4k: %8lu kB\n",
95 direct_pages_count[PG_LEVEL_4K] << 2);
96 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
97 seq_printf(m, "DirectMap2M: %8lu kB\n",
98 direct_pages_count[PG_LEVEL_2M] << 11);
100 seq_printf(m, "DirectMap4M: %8lu kB\n",
101 direct_pages_count[PG_LEVEL_2M] << 12);
104 seq_printf(m, "DirectMap1G: %8lu kB\n",
105 direct_pages_count[PG_LEVEL_1G] << 20);
108 static inline void split_page_count(int level) { }
111 #ifdef CONFIG_X86_CPA_STATISTICS
113 static unsigned long cpa_1g_checked;
114 static unsigned long cpa_1g_sameprot;
115 static unsigned long cpa_1g_preserved;
116 static unsigned long cpa_2m_checked;
117 static unsigned long cpa_2m_sameprot;
118 static unsigned long cpa_2m_preserved;
119 static unsigned long cpa_4k_install;
121 static inline void cpa_inc_1g_checked(void)
126 static inline void cpa_inc_2m_checked(void)
131 static inline void cpa_inc_4k_install(void)
136 static inline void cpa_inc_lp_sameprot(int level)
138 if (level == PG_LEVEL_1G)
144 static inline void cpa_inc_lp_preserved(int level)
146 if (level == PG_LEVEL_1G)
152 static int cpastats_show(struct seq_file *m, void *p)
154 seq_printf(m, "1G pages checked: %16lu\n", cpa_1g_checked);
155 seq_printf(m, "1G pages sameprot: %16lu\n", cpa_1g_sameprot);
156 seq_printf(m, "1G pages preserved: %16lu\n", cpa_1g_preserved);
157 seq_printf(m, "2M pages checked: %16lu\n", cpa_2m_checked);
158 seq_printf(m, "2M pages sameprot: %16lu\n", cpa_2m_sameprot);
159 seq_printf(m, "2M pages preserved: %16lu\n", cpa_2m_preserved);
160 seq_printf(m, "4K pages set-checked: %16lu\n", cpa_4k_install);
164 static int cpastats_open(struct inode *inode, struct file *file)
166 return single_open(file, cpastats_show, NULL);
169 static const struct file_operations cpastats_fops = {
170 .open = cpastats_open,
173 .release = single_release,
176 static int __init cpa_stats_init(void)
178 debugfs_create_file("cpa_stats", S_IRUSR, arch_debugfs_dir, NULL,
182 late_initcall(cpa_stats_init);
184 static inline void cpa_inc_1g_checked(void) { }
185 static inline void cpa_inc_2m_checked(void) { }
186 static inline void cpa_inc_4k_install(void) { }
187 static inline void cpa_inc_lp_sameprot(int level) { }
188 static inline void cpa_inc_lp_preserved(int level) { }
193 within(unsigned long addr, unsigned long start, unsigned long end)
195 return addr >= start && addr < end;
199 within_inclusive(unsigned long addr, unsigned long start, unsigned long end)
201 return addr >= start && addr <= end;
206 static inline unsigned long highmap_start_pfn(void)
208 return __pa_symbol(_text) >> PAGE_SHIFT;
211 static inline unsigned long highmap_end_pfn(void)
213 /* Do not reference physical address outside the kernel. */
214 return __pa_symbol(roundup(_brk_end, PMD_SIZE) - 1) >> PAGE_SHIFT;
217 static bool __cpa_pfn_in_highmap(unsigned long pfn)
220 * Kernel text has an alias mapping at a high address, known
223 return within_inclusive(pfn, highmap_start_pfn(), highmap_end_pfn());
228 static bool __cpa_pfn_in_highmap(unsigned long pfn)
230 /* There is no highmap on 32-bit */
237 * See set_mce_nospec().
239 * Machine check recovery code needs to change cache mode of poisoned pages to
240 * UC to avoid speculative access logging another error. But passing the
241 * address of the 1:1 mapping to set_memory_uc() is a fine way to encourage a
242 * speculative access. So we cheat and flip the top bit of the address. This
243 * works fine for the code that updates the page tables. But at the end of the
244 * process we need to flush the TLB and cache and the non-canonical address
245 * causes a #GP fault when used by the INVLPG and CLFLUSH instructions.
247 * But in the common case we already have a canonical address. This code
248 * will fix the top bit if needed and is a no-op otherwise.
250 static inline unsigned long fix_addr(unsigned long addr)
253 return (long)(addr << 1) >> 1;
259 static unsigned long __cpa_addr(struct cpa_data *cpa, unsigned long idx)
261 if (cpa->flags & CPA_PAGES_ARRAY) {
262 struct page *page = cpa->pages[idx];
264 if (unlikely(PageHighMem(page)))
267 return (unsigned long)page_address(page);
270 if (cpa->flags & CPA_ARRAY)
271 return cpa->vaddr[idx];
273 return *cpa->vaddr + idx * PAGE_SIZE;
280 static void clflush_cache_range_opt(void *vaddr, unsigned int size)
282 const unsigned long clflush_size = boot_cpu_data.x86_clflush_size;
283 void *p = (void *)((unsigned long)vaddr & ~(clflush_size - 1));
284 void *vend = vaddr + size;
289 for (; p < vend; p += clflush_size)
294 * clflush_cache_range - flush a cache range with clflush
295 * @vaddr: virtual start address
296 * @size: number of bytes to flush
298 * CLFLUSHOPT is an unordered instruction which needs fencing with MFENCE or
299 * SFENCE to avoid ordering issues.
301 void clflush_cache_range(void *vaddr, unsigned int size)
304 clflush_cache_range_opt(vaddr, size);
307 EXPORT_SYMBOL_GPL(clflush_cache_range);
309 #ifdef CONFIG_ARCH_HAS_PMEM_API
310 void arch_invalidate_pmem(void *addr, size_t size)
312 clflush_cache_range(addr, size);
314 EXPORT_SYMBOL_GPL(arch_invalidate_pmem);
317 static void __cpa_flush_all(void *arg)
319 unsigned long cache = (unsigned long)arg;
322 * Flush all to work around Errata in early athlons regarding
323 * large page flushing.
327 if (cache && boot_cpu_data.x86 >= 4)
331 static void cpa_flush_all(unsigned long cache)
333 BUG_ON(irqs_disabled() && !early_boot_irqs_disabled);
335 on_each_cpu(__cpa_flush_all, (void *) cache, 1);
338 static void __cpa_flush_tlb(void *data)
340 struct cpa_data *cpa = data;
343 for (i = 0; i < cpa->numpages; i++)
344 __flush_tlb_one_kernel(fix_addr(__cpa_addr(cpa, i)));
347 static void cpa_flush(struct cpa_data *data, int cache)
349 struct cpa_data *cpa = data;
352 BUG_ON(irqs_disabled() && !early_boot_irqs_disabled);
354 if (cache && !static_cpu_has(X86_FEATURE_CLFLUSH)) {
355 cpa_flush_all(cache);
359 if (cpa->force_flush_all || cpa->numpages > tlb_single_page_flush_ceiling)
362 on_each_cpu(__cpa_flush_tlb, cpa, 1);
368 for (i = 0; i < cpa->numpages; i++) {
369 unsigned long addr = __cpa_addr(cpa, i);
372 pte_t *pte = lookup_address(addr, &level);
375 * Only flush present addresses:
377 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
378 clflush_cache_range_opt((void *)fix_addr(addr), PAGE_SIZE);
383 static bool overlaps(unsigned long r1_start, unsigned long r1_end,
384 unsigned long r2_start, unsigned long r2_end)
386 return (r1_start <= r2_end && r1_end >= r2_start) ||
387 (r2_start <= r1_end && r2_end >= r1_start);
390 #ifdef CONFIG_PCI_BIOS
392 * The BIOS area between 640k and 1Mb needs to be executable for PCI BIOS
393 * based config access (CONFIG_PCI_GOBIOS) support.
395 #define BIOS_PFN PFN_DOWN(BIOS_BEGIN)
396 #define BIOS_PFN_END PFN_DOWN(BIOS_END - 1)
398 static pgprotval_t protect_pci_bios(unsigned long spfn, unsigned long epfn)
400 if (pcibios_enabled && overlaps(spfn, epfn, BIOS_PFN, BIOS_PFN_END))
405 static pgprotval_t protect_pci_bios(unsigned long spfn, unsigned long epfn)
412 * The .rodata section needs to be read-only. Using the pfn catches all
413 * aliases. This also includes __ro_after_init, so do not enforce until
414 * kernel_set_to_readonly is true.
416 static pgprotval_t protect_rodata(unsigned long spfn, unsigned long epfn)
418 unsigned long epfn_ro, spfn_ro = PFN_DOWN(__pa_symbol(__start_rodata));
421 * Note: __end_rodata is at page aligned and not inclusive, so
422 * subtract 1 to get the last enforced PFN in the rodata area.
424 epfn_ro = PFN_DOWN(__pa_symbol(__end_rodata)) - 1;
426 if (kernel_set_to_readonly && overlaps(spfn, epfn, spfn_ro, epfn_ro))
432 * Protect kernel text against becoming non executable by forbidding
433 * _PAGE_NX. This protects only the high kernel mapping (_text -> _etext)
434 * out of which the kernel actually executes. Do not protect the low
437 * This does not cover __inittext since that is gone after boot.
439 static pgprotval_t protect_kernel_text(unsigned long start, unsigned long end)
441 unsigned long t_end = (unsigned long)_etext - 1;
442 unsigned long t_start = (unsigned long)_text;
444 if (overlaps(start, end, t_start, t_end))
449 #if defined(CONFIG_X86_64)
451 * Once the kernel maps the text as RO (kernel_set_to_readonly is set),
452 * kernel text mappings for the large page aligned text, rodata sections
453 * will be always read-only. For the kernel identity mappings covering the
454 * holes caused by this alignment can be anything that user asks.
456 * This will preserve the large page mappings for kernel text/data at no
459 static pgprotval_t protect_kernel_text_ro(unsigned long start,
462 unsigned long t_end = (unsigned long)__end_rodata_hpage_align - 1;
463 unsigned long t_start = (unsigned long)_text;
466 if (!kernel_set_to_readonly || !overlaps(start, end, t_start, t_end))
469 * Don't enforce the !RW mapping for the kernel text mapping, if
470 * the current mapping is already using small page mapping. No
471 * need to work hard to preserve large page mappings in this case.
473 * This also fixes the Linux Xen paravirt guest boot failure caused
474 * by unexpected read-only mappings for kernel identity
475 * mappings. In this paravirt guest case, the kernel text mapping
476 * and the kernel identity mapping share the same page-table pages,
477 * so the protections for kernel text and identity mappings have to
480 if (lookup_address(start, &level) && (level != PG_LEVEL_4K))
485 static pgprotval_t protect_kernel_text_ro(unsigned long start,
492 static inline bool conflicts(pgprot_t prot, pgprotval_t val)
494 return (pgprot_val(prot) & ~val) != pgprot_val(prot);
497 static inline void check_conflict(int warnlvl, pgprot_t prot, pgprotval_t val,
498 unsigned long start, unsigned long end,
499 unsigned long pfn, const char *txt)
501 static const char *lvltxt[] = {
502 [CPA_CONFLICT] = "conflict",
503 [CPA_PROTECT] = "protect",
504 [CPA_DETECT] = "detect",
507 if (warnlvl > cpa_warn_level || !conflicts(prot, val))
510 pr_warn("CPA %8s %10s: 0x%016lx - 0x%016lx PFN %lx req %016llx prevent %016llx\n",
511 lvltxt[warnlvl], txt, start, end, pfn, (unsigned long long)pgprot_val(prot),
512 (unsigned long long)val);
516 * Certain areas of memory on x86 require very specific protection flags,
517 * for example the BIOS area or kernel text. Callers don't always get this
518 * right (again, ioremap() on BIOS memory is not uncommon) so this function
519 * checks and fixes these known static required protection bits.
521 static inline pgprot_t static_protections(pgprot_t prot, unsigned long start,
522 unsigned long pfn, unsigned long npg,
523 unsigned long lpsize, int warnlvl)
525 pgprotval_t forbidden, res;
529 * There is no point in checking RW/NX conflicts when the requested
530 * mapping is setting the page !PRESENT.
532 if (!(pgprot_val(prot) & _PAGE_PRESENT))
535 /* Operate on the virtual address */
536 end = start + npg * PAGE_SIZE - 1;
538 res = protect_kernel_text(start, end);
539 check_conflict(warnlvl, prot, res, start, end, pfn, "Text NX");
543 * Special case to preserve a large page. If the change spawns the
544 * full large page mapping then there is no point to split it
545 * up. Happens with ftrace and is going to be removed once ftrace
546 * switched to text_poke().
548 if (lpsize != (npg * PAGE_SIZE) || (start & (lpsize - 1))) {
549 res = protect_kernel_text_ro(start, end);
550 check_conflict(warnlvl, prot, res, start, end, pfn, "Text RO");
554 /* Check the PFN directly */
555 res = protect_pci_bios(pfn, pfn + npg - 1);
556 check_conflict(warnlvl, prot, res, start, end, pfn, "PCIBIOS NX");
559 res = protect_rodata(pfn, pfn + npg - 1);
560 check_conflict(warnlvl, prot, res, start, end, pfn, "Rodata RO");
563 return __pgprot(pgprot_val(prot) & ~forbidden);
567 * Lookup the page table entry for a virtual address in a specific pgd.
568 * Return a pointer to the entry and the level of the mapping.
570 pte_t *lookup_address_in_pgd(pgd_t *pgd, unsigned long address,
577 *level = PG_LEVEL_NONE;
582 p4d = p4d_offset(pgd, address);
586 *level = PG_LEVEL_512G;
587 if (p4d_large(*p4d) || !p4d_present(*p4d))
590 pud = pud_offset(p4d, address);
594 *level = PG_LEVEL_1G;
595 if (pud_large(*pud) || !pud_present(*pud))
598 pmd = pmd_offset(pud, address);
602 *level = PG_LEVEL_2M;
603 if (pmd_large(*pmd) || !pmd_present(*pmd))
606 *level = PG_LEVEL_4K;
608 return pte_offset_kernel(pmd, address);
612 * Lookup the page table entry for a virtual address. Return a pointer
613 * to the entry and the level of the mapping.
615 * Note: We return pud and pmd either when the entry is marked large
616 * or when the present bit is not set. Otherwise we would return a
617 * pointer to a nonexisting mapping.
619 pte_t *lookup_address(unsigned long address, unsigned int *level)
621 return lookup_address_in_pgd(pgd_offset_k(address), address, level);
623 EXPORT_SYMBOL_GPL(lookup_address);
626 * Lookup the page table entry for a virtual address in a given mm. Return a
627 * pointer to the entry and the level of the mapping.
629 pte_t *lookup_address_in_mm(struct mm_struct *mm, unsigned long address,
632 return lookup_address_in_pgd(pgd_offset(mm, address), address, level);
634 EXPORT_SYMBOL_GPL(lookup_address_in_mm);
636 static pte_t *_lookup_address_cpa(struct cpa_data *cpa, unsigned long address,
640 return lookup_address_in_pgd(cpa->pgd + pgd_index(address),
643 return lookup_address(address, level);
647 * Lookup the PMD entry for a virtual address. Return a pointer to the entry
648 * or NULL if not present.
650 pmd_t *lookup_pmd_address(unsigned long address)
656 pgd = pgd_offset_k(address);
660 p4d = p4d_offset(pgd, address);
661 if (p4d_none(*p4d) || p4d_large(*p4d) || !p4d_present(*p4d))
664 pud = pud_offset(p4d, address);
665 if (pud_none(*pud) || pud_large(*pud) || !pud_present(*pud))
668 return pmd_offset(pud, address);
672 * This is necessary because __pa() does not work on some
673 * kinds of memory, like vmalloc() or the alloc_remap()
674 * areas on 32-bit NUMA systems. The percpu areas can
675 * end up in this kind of memory, for instance.
677 * This could be optimized, but it is only intended to be
678 * used at inititalization time, and keeping it
679 * unoptimized should increase the testing coverage for
680 * the more obscure platforms.
682 phys_addr_t slow_virt_to_phys(void *__virt_addr)
684 unsigned long virt_addr = (unsigned long)__virt_addr;
685 phys_addr_t phys_addr;
686 unsigned long offset;
690 pte = lookup_address(virt_addr, &level);
694 * pXX_pfn() returns unsigned long, which must be cast to phys_addr_t
695 * before being left-shifted PAGE_SHIFT bits -- this trick is to
696 * make 32-PAE kernel work correctly.
700 phys_addr = (phys_addr_t)pud_pfn(*(pud_t *)pte) << PAGE_SHIFT;
701 offset = virt_addr & ~PUD_PAGE_MASK;
704 phys_addr = (phys_addr_t)pmd_pfn(*(pmd_t *)pte) << PAGE_SHIFT;
705 offset = virt_addr & ~PMD_PAGE_MASK;
708 phys_addr = (phys_addr_t)pte_pfn(*pte) << PAGE_SHIFT;
709 offset = virt_addr & ~PAGE_MASK;
712 return (phys_addr_t)(phys_addr | offset);
714 EXPORT_SYMBOL_GPL(slow_virt_to_phys);
717 * Set the new pmd in all the pgds we know about:
719 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
722 set_pte_atomic(kpte, pte);
724 if (!SHARED_KERNEL_PMD) {
727 list_for_each_entry(page, &pgd_list, lru) {
733 pgd = (pgd_t *)page_address(page) + pgd_index(address);
734 p4d = p4d_offset(pgd, address);
735 pud = pud_offset(p4d, address);
736 pmd = pmd_offset(pud, address);
737 set_pte_atomic((pte_t *)pmd, pte);
743 static pgprot_t pgprot_clear_protnone_bits(pgprot_t prot)
746 * _PAGE_GLOBAL means "global page" for present PTEs.
747 * But, it is also used to indicate _PAGE_PROTNONE
748 * for non-present PTEs.
750 * This ensures that a _PAGE_GLOBAL PTE going from
751 * present to non-present is not confused as
754 if (!(pgprot_val(prot) & _PAGE_PRESENT))
755 pgprot_val(prot) &= ~_PAGE_GLOBAL;
760 static int __should_split_large_page(pte_t *kpte, unsigned long address,
761 struct cpa_data *cpa)
763 unsigned long numpages, pmask, psize, lpaddr, pfn, old_pfn;
764 pgprot_t old_prot, new_prot, req_prot, chk_prot;
769 * Check for races, another CPU might have split this page
772 tmp = _lookup_address_cpa(cpa, address, &level);
778 old_prot = pmd_pgprot(*(pmd_t *)kpte);
779 old_pfn = pmd_pfn(*(pmd_t *)kpte);
780 cpa_inc_2m_checked();
783 old_prot = pud_pgprot(*(pud_t *)kpte);
784 old_pfn = pud_pfn(*(pud_t *)kpte);
785 cpa_inc_1g_checked();
791 psize = page_level_size(level);
792 pmask = page_level_mask(level);
795 * Calculate the number of pages, which fit into this large
796 * page starting at address:
798 lpaddr = (address + psize) & pmask;
799 numpages = (lpaddr - address) >> PAGE_SHIFT;
800 if (numpages < cpa->numpages)
801 cpa->numpages = numpages;
804 * We are safe now. Check whether the new pgprot is the same:
805 * Convert protection attributes to 4k-format, as cpa->mask* are set
809 /* Clear PSE (aka _PAGE_PAT) and move PAT bit to correct position */
810 req_prot = pgprot_large_2_4k(old_prot);
812 pgprot_val(req_prot) &= ~pgprot_val(cpa->mask_clr);
813 pgprot_val(req_prot) |= pgprot_val(cpa->mask_set);
816 * req_prot is in format of 4k pages. It must be converted to large
817 * page format: the caching mode includes the PAT bit located at
818 * different bit positions in the two formats.
820 req_prot = pgprot_4k_2_large(req_prot);
821 req_prot = pgprot_clear_protnone_bits(req_prot);
822 if (pgprot_val(req_prot) & _PAGE_PRESENT)
823 pgprot_val(req_prot) |= _PAGE_PSE;
826 * old_pfn points to the large page base pfn. So we need to add the
827 * offset of the virtual address:
829 pfn = old_pfn + ((address & (psize - 1)) >> PAGE_SHIFT);
833 * Calculate the large page base address and the number of 4K pages
836 lpaddr = address & pmask;
837 numpages = psize >> PAGE_SHIFT;
840 * Sanity check that the existing mapping is correct versus the static
841 * protections. static_protections() guards against !PRESENT, so no
842 * extra conditional required here.
844 chk_prot = static_protections(old_prot, lpaddr, old_pfn, numpages,
845 psize, CPA_CONFLICT);
847 if (WARN_ON_ONCE(pgprot_val(chk_prot) != pgprot_val(old_prot))) {
849 * Split the large page and tell the split code to
850 * enforce static protections.
852 cpa->force_static_prot = 1;
857 * Optimization: If the requested pgprot is the same as the current
858 * pgprot, then the large page can be preserved and no updates are
859 * required independent of alignment and length of the requested
860 * range. The above already established that the current pgprot is
861 * correct, which in consequence makes the requested pgprot correct
862 * as well if it is the same. The static protection scan below will
863 * not come to a different conclusion.
865 if (pgprot_val(req_prot) == pgprot_val(old_prot)) {
866 cpa_inc_lp_sameprot(level);
871 * If the requested range does not cover the full page, split it up
873 if (address != lpaddr || cpa->numpages != numpages)
877 * Check whether the requested pgprot is conflicting with a static
878 * protection requirement in the large page.
880 new_prot = static_protections(req_prot, lpaddr, old_pfn, numpages,
884 * If there is a conflict, split the large page.
886 * There used to be a 4k wise evaluation trying really hard to
887 * preserve the large pages, but experimentation has shown, that this
888 * does not help at all. There might be corner cases which would
889 * preserve one large page occasionally, but it's really not worth the
890 * extra code and cycles for the common case.
892 if (pgprot_val(req_prot) != pgprot_val(new_prot))
895 /* All checks passed. Update the large page mapping. */
896 new_pte = pfn_pte(old_pfn, new_prot);
897 __set_pmd_pte(kpte, address, new_pte);
898 cpa->flags |= CPA_FLUSHTLB;
899 cpa_inc_lp_preserved(level);
903 static int should_split_large_page(pte_t *kpte, unsigned long address,
904 struct cpa_data *cpa)
908 if (cpa->force_split)
911 spin_lock(&pgd_lock);
912 do_split = __should_split_large_page(kpte, address, cpa);
913 spin_unlock(&pgd_lock);
918 static void split_set_pte(struct cpa_data *cpa, pte_t *pte, unsigned long pfn,
919 pgprot_t ref_prot, unsigned long address,
922 unsigned int npg = PFN_DOWN(size);
926 * If should_split_large_page() discovered an inconsistent mapping,
927 * remove the invalid protection in the split mapping.
929 if (!cpa->force_static_prot)
932 /* Hand in lpsize = 0 to enforce the protection mechanism */
933 prot = static_protections(ref_prot, address, pfn, npg, 0, CPA_PROTECT);
935 if (pgprot_val(prot) == pgprot_val(ref_prot))
939 * If this is splitting a PMD, fix it up. PUD splits cannot be
940 * fixed trivially as that would require to rescan the newly
941 * installed PMD mappings after returning from split_large_page()
942 * so an eventual further split can allocate the necessary PTE
943 * pages. Warn for now and revisit it in case this actually
946 if (size == PAGE_SIZE)
949 pr_warn_once("CPA: Cannot fixup static protections for PUD split\n");
951 set_pte(pte, pfn_pte(pfn, ref_prot));
955 __split_large_page(struct cpa_data *cpa, pte_t *kpte, unsigned long address,
958 unsigned long lpaddr, lpinc, ref_pfn, pfn, pfninc = 1;
959 pte_t *pbase = (pte_t *)page_address(base);
960 unsigned int i, level;
964 spin_lock(&pgd_lock);
966 * Check for races, another CPU might have split this page
969 tmp = _lookup_address_cpa(cpa, address, &level);
971 spin_unlock(&pgd_lock);
975 paravirt_alloc_pte(&init_mm, page_to_pfn(base));
979 ref_prot = pmd_pgprot(*(pmd_t *)kpte);
981 * Clear PSE (aka _PAGE_PAT) and move
982 * PAT bit to correct position.
984 ref_prot = pgprot_large_2_4k(ref_prot);
985 ref_pfn = pmd_pfn(*(pmd_t *)kpte);
986 lpaddr = address & PMD_MASK;
991 ref_prot = pud_pgprot(*(pud_t *)kpte);
992 ref_pfn = pud_pfn(*(pud_t *)kpte);
993 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
994 lpaddr = address & PUD_MASK;
997 * Clear the PSE flags if the PRESENT flag is not set
998 * otherwise pmd_present/pmd_huge will return true
999 * even on a non present pmd.
1001 if (!(pgprot_val(ref_prot) & _PAGE_PRESENT))
1002 pgprot_val(ref_prot) &= ~_PAGE_PSE;
1006 spin_unlock(&pgd_lock);
1010 ref_prot = pgprot_clear_protnone_bits(ref_prot);
1013 * Get the target pfn from the original entry:
1016 for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc, lpaddr += lpinc)
1017 split_set_pte(cpa, pbase + i, pfn, ref_prot, lpaddr, lpinc);
1019 if (virt_addr_valid(address)) {
1020 unsigned long pfn = PFN_DOWN(__pa(address));
1022 if (pfn_range_is_mapped(pfn, pfn + 1))
1023 split_page_count(level);
1027 * Install the new, split up pagetable.
1029 * We use the standard kernel pagetable protections for the new
1030 * pagetable protections, the actual ptes set above control the
1031 * primary protection behavior:
1033 __set_pmd_pte(kpte, address, mk_pte(base, __pgprot(_KERNPG_TABLE)));
1036 * Do a global flush tlb after splitting the large page
1037 * and before we do the actual change page attribute in the PTE.
1039 * Without this, we violate the TLB application note, that says:
1040 * "The TLBs may contain both ordinary and large-page
1041 * translations for a 4-KByte range of linear addresses. This
1042 * may occur if software modifies the paging structures so that
1043 * the page size used for the address range changes. If the two
1044 * translations differ with respect to page frame or attributes
1045 * (e.g., permissions), processor behavior is undefined and may
1046 * be implementation-specific."
1048 * We do this global tlb flush inside the cpa_lock, so that we
1049 * don't allow any other cpu, with stale tlb entries change the
1050 * page attribute in parallel, that also falls into the
1051 * just split large page entry.
1054 spin_unlock(&pgd_lock);
1059 static int split_large_page(struct cpa_data *cpa, pte_t *kpte,
1060 unsigned long address)
1064 if (!debug_pagealloc_enabled())
1065 spin_unlock(&cpa_lock);
1066 base = alloc_pages(GFP_KERNEL, 0);
1067 if (!debug_pagealloc_enabled())
1068 spin_lock(&cpa_lock);
1072 if (__split_large_page(cpa, kpte, address, base))
1078 static bool try_to_free_pte_page(pte_t *pte)
1082 for (i = 0; i < PTRS_PER_PTE; i++)
1083 if (!pte_none(pte[i]))
1086 free_page((unsigned long)pte);
1090 static bool try_to_free_pmd_page(pmd_t *pmd)
1094 for (i = 0; i < PTRS_PER_PMD; i++)
1095 if (!pmd_none(pmd[i]))
1098 free_page((unsigned long)pmd);
1102 static bool unmap_pte_range(pmd_t *pmd, unsigned long start, unsigned long end)
1104 pte_t *pte = pte_offset_kernel(pmd, start);
1106 while (start < end) {
1107 set_pte(pte, __pte(0));
1113 if (try_to_free_pte_page((pte_t *)pmd_page_vaddr(*pmd))) {
1120 static void __unmap_pmd_range(pud_t *pud, pmd_t *pmd,
1121 unsigned long start, unsigned long end)
1123 if (unmap_pte_range(pmd, start, end))
1124 if (try_to_free_pmd_page((pmd_t *)pud_page_vaddr(*pud)))
1128 static void unmap_pmd_range(pud_t *pud, unsigned long start, unsigned long end)
1130 pmd_t *pmd = pmd_offset(pud, start);
1133 * Not on a 2MB page boundary?
1135 if (start & (PMD_SIZE - 1)) {
1136 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
1137 unsigned long pre_end = min_t(unsigned long, end, next_page);
1139 __unmap_pmd_range(pud, pmd, start, pre_end);
1146 * Try to unmap in 2M chunks.
1148 while (end - start >= PMD_SIZE) {
1149 if (pmd_large(*pmd))
1152 __unmap_pmd_range(pud, pmd, start, start + PMD_SIZE);
1162 return __unmap_pmd_range(pud, pmd, start, end);
1165 * Try again to free the PMD page if haven't succeeded above.
1167 if (!pud_none(*pud))
1168 if (try_to_free_pmd_page((pmd_t *)pud_page_vaddr(*pud)))
1172 static void unmap_pud_range(p4d_t *p4d, unsigned long start, unsigned long end)
1174 pud_t *pud = pud_offset(p4d, start);
1177 * Not on a GB page boundary?
1179 if (start & (PUD_SIZE - 1)) {
1180 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
1181 unsigned long pre_end = min_t(unsigned long, end, next_page);
1183 unmap_pmd_range(pud, start, pre_end);
1190 * Try to unmap in 1G chunks?
1192 while (end - start >= PUD_SIZE) {
1194 if (pud_large(*pud))
1197 unmap_pmd_range(pud, start, start + PUD_SIZE);
1207 unmap_pmd_range(pud, start, end);
1210 * No need to try to free the PUD page because we'll free it in
1211 * populate_pgd's error path
1215 static int alloc_pte_page(pmd_t *pmd)
1217 pte_t *pte = (pte_t *)get_zeroed_page(GFP_KERNEL);
1221 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
1225 static int alloc_pmd_page(pud_t *pud)
1227 pmd_t *pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL);
1231 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
1235 static void populate_pte(struct cpa_data *cpa,
1236 unsigned long start, unsigned long end,
1237 unsigned num_pages, pmd_t *pmd, pgprot_t pgprot)
1241 pte = pte_offset_kernel(pmd, start);
1243 pgprot = pgprot_clear_protnone_bits(pgprot);
1245 while (num_pages-- && start < end) {
1246 set_pte(pte, pfn_pte(cpa->pfn, pgprot));
1254 static long populate_pmd(struct cpa_data *cpa,
1255 unsigned long start, unsigned long end,
1256 unsigned num_pages, pud_t *pud, pgprot_t pgprot)
1260 pgprot_t pmd_pgprot;
1263 * Not on a 2M boundary?
1265 if (start & (PMD_SIZE - 1)) {
1266 unsigned long pre_end = start + (num_pages << PAGE_SHIFT);
1267 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
1269 pre_end = min_t(unsigned long, pre_end, next_page);
1270 cur_pages = (pre_end - start) >> PAGE_SHIFT;
1271 cur_pages = min_t(unsigned int, num_pages, cur_pages);
1276 pmd = pmd_offset(pud, start);
1278 if (alloc_pte_page(pmd))
1281 populate_pte(cpa, start, pre_end, cur_pages, pmd, pgprot);
1287 * We mapped them all?
1289 if (num_pages == cur_pages)
1292 pmd_pgprot = pgprot_4k_2_large(pgprot);
1294 while (end - start >= PMD_SIZE) {
1297 * We cannot use a 1G page so allocate a PMD page if needed.
1300 if (alloc_pmd_page(pud))
1303 pmd = pmd_offset(pud, start);
1305 set_pmd(pmd, pmd_mkhuge(pfn_pmd(cpa->pfn,
1306 canon_pgprot(pmd_pgprot))));
1309 cpa->pfn += PMD_SIZE >> PAGE_SHIFT;
1310 cur_pages += PMD_SIZE >> PAGE_SHIFT;
1314 * Map trailing 4K pages.
1317 pmd = pmd_offset(pud, start);
1319 if (alloc_pte_page(pmd))
1322 populate_pte(cpa, start, end, num_pages - cur_pages,
1328 static int populate_pud(struct cpa_data *cpa, unsigned long start, p4d_t *p4d,
1334 pgprot_t pud_pgprot;
1336 end = start + (cpa->numpages << PAGE_SHIFT);
1339 * Not on a Gb page boundary? => map everything up to it with
1342 if (start & (PUD_SIZE - 1)) {
1343 unsigned long pre_end;
1344 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
1346 pre_end = min_t(unsigned long, end, next_page);
1347 cur_pages = (pre_end - start) >> PAGE_SHIFT;
1348 cur_pages = min_t(int, (int)cpa->numpages, cur_pages);
1350 pud = pud_offset(p4d, start);
1356 if (alloc_pmd_page(pud))
1359 cur_pages = populate_pmd(cpa, start, pre_end, cur_pages,
1367 /* We mapped them all? */
1368 if (cpa->numpages == cur_pages)
1371 pud = pud_offset(p4d, start);
1372 pud_pgprot = pgprot_4k_2_large(pgprot);
1375 * Map everything starting from the Gb boundary, possibly with 1G pages
1377 while (boot_cpu_has(X86_FEATURE_GBPAGES) && end - start >= PUD_SIZE) {
1378 set_pud(pud, pud_mkhuge(pfn_pud(cpa->pfn,
1379 canon_pgprot(pud_pgprot))));
1382 cpa->pfn += PUD_SIZE >> PAGE_SHIFT;
1383 cur_pages += PUD_SIZE >> PAGE_SHIFT;
1387 /* Map trailing leftover */
1391 pud = pud_offset(p4d, start);
1393 if (alloc_pmd_page(pud))
1396 tmp = populate_pmd(cpa, start, end, cpa->numpages - cur_pages,
1407 * Restrictions for kernel page table do not necessarily apply when mapping in
1410 static int populate_pgd(struct cpa_data *cpa, unsigned long addr)
1412 pgprot_t pgprot = __pgprot(_KERNPG_TABLE);
1413 pud_t *pud = NULL; /* shut up gcc */
1418 pgd_entry = cpa->pgd + pgd_index(addr);
1420 if (pgd_none(*pgd_entry)) {
1421 p4d = (p4d_t *)get_zeroed_page(GFP_KERNEL);
1425 set_pgd(pgd_entry, __pgd(__pa(p4d) | _KERNPG_TABLE));
1429 * Allocate a PUD page and hand it down for mapping.
1431 p4d = p4d_offset(pgd_entry, addr);
1432 if (p4d_none(*p4d)) {
1433 pud = (pud_t *)get_zeroed_page(GFP_KERNEL);
1437 set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE));
1440 pgprot_val(pgprot) &= ~pgprot_val(cpa->mask_clr);
1441 pgprot_val(pgprot) |= pgprot_val(cpa->mask_set);
1443 ret = populate_pud(cpa, addr, p4d, pgprot);
1446 * Leave the PUD page in place in case some other CPU or thread
1447 * already found it, but remove any useless entries we just
1450 unmap_pud_range(p4d, addr,
1451 addr + (cpa->numpages << PAGE_SHIFT));
1455 cpa->numpages = ret;
1459 static int __cpa_process_fault(struct cpa_data *cpa, unsigned long vaddr,
1464 * Right now, we only execute this code path when mapping
1465 * the EFI virtual memory map regions, no other users
1466 * provide a ->pgd value. This may change in the future.
1468 return populate_pgd(cpa, vaddr);
1472 * Ignore all non primary paths.
1480 * Ignore the NULL PTE for kernel identity mapping, as it is expected
1482 * Also set numpages to '1' indicating that we processed cpa req for
1483 * one virtual address page and its pfn. TBD: numpages can be set based
1484 * on the initial value and the level returned by lookup_address().
1486 if (within(vaddr, PAGE_OFFSET,
1487 PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
1489 cpa->pfn = __pa(vaddr) >> PAGE_SHIFT;
1492 } else if (__cpa_pfn_in_highmap(cpa->pfn)) {
1493 /* Faults in the highmap are OK, so do not warn: */
1496 WARN(1, KERN_WARNING "CPA: called for zero pte. "
1497 "vaddr = %lx cpa->vaddr = %lx\n", vaddr,
1504 static int __change_page_attr(struct cpa_data *cpa, int primary)
1506 unsigned long address;
1509 pte_t *kpte, old_pte;
1511 address = __cpa_addr(cpa, cpa->curpage);
1513 kpte = _lookup_address_cpa(cpa, address, &level);
1515 return __cpa_process_fault(cpa, address, primary);
1518 if (pte_none(old_pte))
1519 return __cpa_process_fault(cpa, address, primary);
1521 if (level == PG_LEVEL_4K) {
1523 pgprot_t new_prot = pte_pgprot(old_pte);
1524 unsigned long pfn = pte_pfn(old_pte);
1526 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
1527 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
1529 cpa_inc_4k_install();
1530 /* Hand in lpsize = 0 to enforce the protection mechanism */
1531 new_prot = static_protections(new_prot, address, pfn, 1, 0,
1534 new_prot = pgprot_clear_protnone_bits(new_prot);
1537 * We need to keep the pfn from the existing PTE,
1538 * after all we're only going to change it's attributes
1539 * not the memory it points to
1541 new_pte = pfn_pte(pfn, new_prot);
1544 * Do we really change anything ?
1546 if (pte_val(old_pte) != pte_val(new_pte)) {
1547 set_pte_atomic(kpte, new_pte);
1548 cpa->flags |= CPA_FLUSHTLB;
1555 * Check, whether we can keep the large page intact
1556 * and just change the pte:
1558 do_split = should_split_large_page(kpte, address, cpa);
1560 * When the range fits into the existing large page,
1561 * return. cp->numpages and cpa->tlbflush have been updated in
1568 * We have to split the large page:
1570 err = split_large_page(cpa, kpte, address);
1577 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
1579 static int cpa_process_alias(struct cpa_data *cpa)
1581 struct cpa_data alias_cpa;
1582 unsigned long laddr = (unsigned long)__va(cpa->pfn << PAGE_SHIFT);
1583 unsigned long vaddr;
1586 if (!pfn_range_is_mapped(cpa->pfn, cpa->pfn + 1))
1590 * No need to redo, when the primary call touched the direct
1593 vaddr = __cpa_addr(cpa, cpa->curpage);
1594 if (!(within(vaddr, PAGE_OFFSET,
1595 PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT)))) {
1598 alias_cpa.vaddr = &laddr;
1599 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1600 alias_cpa.curpage = 0;
1602 cpa->force_flush_all = 1;
1604 ret = __change_page_attr_set_clr(&alias_cpa, 0);
1609 #ifdef CONFIG_X86_64
1611 * If the primary call didn't touch the high mapping already
1612 * and the physical address is inside the kernel map, we need
1613 * to touch the high mapped kernel as well:
1615 if (!within(vaddr, (unsigned long)_text, _brk_end) &&
1616 __cpa_pfn_in_highmap(cpa->pfn)) {
1617 unsigned long temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) +
1618 __START_KERNEL_map - phys_base;
1620 alias_cpa.vaddr = &temp_cpa_vaddr;
1621 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1622 alias_cpa.curpage = 0;
1624 cpa->force_flush_all = 1;
1626 * The high mapping range is imprecise, so ignore the
1629 __change_page_attr_set_clr(&alias_cpa, 0);
1636 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
1638 unsigned long numpages = cpa->numpages;
1639 unsigned long rempages = numpages;
1644 * Store the remaining nr of pages for the large page
1645 * preservation check.
1647 cpa->numpages = rempages;
1648 /* for array changes, we can't use large page */
1649 if (cpa->flags & (CPA_ARRAY | CPA_PAGES_ARRAY))
1652 if (!debug_pagealloc_enabled())
1653 spin_lock(&cpa_lock);
1654 ret = __change_page_attr(cpa, checkalias);
1655 if (!debug_pagealloc_enabled())
1656 spin_unlock(&cpa_lock);
1661 ret = cpa_process_alias(cpa);
1667 * Adjust the number of pages with the result of the
1668 * CPA operation. Either a large page has been
1669 * preserved or a single page update happened.
1671 BUG_ON(cpa->numpages > rempages || !cpa->numpages);
1672 rempages -= cpa->numpages;
1673 cpa->curpage += cpa->numpages;
1677 /* Restore the original numpages */
1678 cpa->numpages = numpages;
1682 static int change_page_attr_set_clr(unsigned long *addr, int numpages,
1683 pgprot_t mask_set, pgprot_t mask_clr,
1684 int force_split, int in_flag,
1685 struct page **pages)
1687 struct cpa_data cpa;
1688 int ret, cache, checkalias;
1690 memset(&cpa, 0, sizeof(cpa));
1693 * Check, if we are requested to set a not supported
1694 * feature. Clearing non-supported features is OK.
1696 mask_set = canon_pgprot(mask_set);
1698 if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
1701 /* Ensure we are PAGE_SIZE aligned */
1702 if (in_flag & CPA_ARRAY) {
1704 for (i = 0; i < numpages; i++) {
1705 if (addr[i] & ~PAGE_MASK) {
1706 addr[i] &= PAGE_MASK;
1710 } else if (!(in_flag & CPA_PAGES_ARRAY)) {
1712 * in_flag of CPA_PAGES_ARRAY implies it is aligned.
1713 * No need to check in that case
1715 if (*addr & ~PAGE_MASK) {
1718 * People should not be passing in unaligned addresses:
1724 /* Must avoid aliasing mappings in the highmem code */
1725 kmap_flush_unused();
1731 cpa.numpages = numpages;
1732 cpa.mask_set = mask_set;
1733 cpa.mask_clr = mask_clr;
1736 cpa.force_split = force_split;
1738 if (in_flag & (CPA_ARRAY | CPA_PAGES_ARRAY))
1739 cpa.flags |= in_flag;
1741 /* No alias checking for _NX bit modifications */
1742 checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
1743 /* Has caller explicitly disabled alias checking? */
1744 if (in_flag & CPA_NO_CHECK_ALIAS)
1747 ret = __change_page_attr_set_clr(&cpa, checkalias);
1750 * Check whether we really changed something:
1752 if (!(cpa.flags & CPA_FLUSHTLB))
1756 * No need to flush, when we did not set any of the caching
1759 cache = !!pgprot2cachemode(mask_set);
1762 * On error; flush everything to be sure.
1765 cpa_flush_all(cache);
1769 cpa_flush(&cpa, cache);
1774 static inline int change_page_attr_set(unsigned long *addr, int numpages,
1775 pgprot_t mask, int array)
1777 return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
1778 (array ? CPA_ARRAY : 0), NULL);
1781 static inline int change_page_attr_clear(unsigned long *addr, int numpages,
1782 pgprot_t mask, int array)
1784 return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
1785 (array ? CPA_ARRAY : 0), NULL);
1788 static inline int cpa_set_pages_array(struct page **pages, int numpages,
1791 return change_page_attr_set_clr(NULL, numpages, mask, __pgprot(0), 0,
1792 CPA_PAGES_ARRAY, pages);
1795 static inline int cpa_clear_pages_array(struct page **pages, int numpages,
1798 return change_page_attr_set_clr(NULL, numpages, __pgprot(0), mask, 0,
1799 CPA_PAGES_ARRAY, pages);
1803 * _set_memory_prot is an internal helper for callers that have been passed
1804 * a pgprot_t value from upper layers and a reservation has already been taken.
1805 * If you want to set the pgprot to a specific page protocol, use the
1806 * set_memory_xx() functions.
1808 int __set_memory_prot(unsigned long addr, int numpages, pgprot_t prot)
1810 return change_page_attr_set_clr(&addr, numpages, prot,
1811 __pgprot(~pgprot_val(prot)), 0, 0,
1815 int _set_memory_uc(unsigned long addr, int numpages)
1818 * for now UC MINUS. see comments in ioremap()
1819 * If you really need strong UC use ioremap_uc(), but note
1820 * that you cannot override IO areas with set_memory_*() as
1821 * these helpers cannot work with IO memory.
1823 return change_page_attr_set(&addr, numpages,
1824 cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1828 int set_memory_uc(unsigned long addr, int numpages)
1833 * for now UC MINUS. see comments in ioremap()
1835 ret = memtype_reserve(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1836 _PAGE_CACHE_MODE_UC_MINUS, NULL);
1840 ret = _set_memory_uc(addr, numpages);
1847 memtype_free(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1851 EXPORT_SYMBOL(set_memory_uc);
1853 int _set_memory_wc(unsigned long addr, int numpages)
1857 ret = change_page_attr_set(&addr, numpages,
1858 cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1861 ret = change_page_attr_set_clr(&addr, numpages,
1862 cachemode2pgprot(_PAGE_CACHE_MODE_WC),
1863 __pgprot(_PAGE_CACHE_MASK),
1869 int set_memory_wc(unsigned long addr, int numpages)
1873 ret = memtype_reserve(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1874 _PAGE_CACHE_MODE_WC, NULL);
1878 ret = _set_memory_wc(addr, numpages);
1880 memtype_free(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1884 EXPORT_SYMBOL(set_memory_wc);
1886 int _set_memory_wt(unsigned long addr, int numpages)
1888 return change_page_attr_set(&addr, numpages,
1889 cachemode2pgprot(_PAGE_CACHE_MODE_WT), 0);
1892 int _set_memory_wb(unsigned long addr, int numpages)
1894 /* WB cache mode is hard wired to all cache attribute bits being 0 */
1895 return change_page_attr_clear(&addr, numpages,
1896 __pgprot(_PAGE_CACHE_MASK), 0);
1899 int set_memory_wb(unsigned long addr, int numpages)
1903 ret = _set_memory_wb(addr, numpages);
1907 memtype_free(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1910 EXPORT_SYMBOL(set_memory_wb);
1912 int set_memory_x(unsigned long addr, int numpages)
1914 if (!(__supported_pte_mask & _PAGE_NX))
1917 return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1920 int set_memory_nx(unsigned long addr, int numpages)
1922 if (!(__supported_pte_mask & _PAGE_NX))
1925 return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1928 int set_memory_ro(unsigned long addr, int numpages)
1930 return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1933 int set_memory_rw(unsigned long addr, int numpages)
1935 return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
1938 int set_memory_np(unsigned long addr, int numpages)
1940 return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
1943 int set_memory_np_noalias(unsigned long addr, int numpages)
1945 int cpa_flags = CPA_NO_CHECK_ALIAS;
1947 return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1948 __pgprot(_PAGE_PRESENT), 0,
1952 int set_memory_4k(unsigned long addr, int numpages)
1954 return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1955 __pgprot(0), 1, 0, NULL);
1958 int set_memory_nonglobal(unsigned long addr, int numpages)
1960 return change_page_attr_clear(&addr, numpages,
1961 __pgprot(_PAGE_GLOBAL), 0);
1964 int set_memory_global(unsigned long addr, int numpages)
1966 return change_page_attr_set(&addr, numpages,
1967 __pgprot(_PAGE_GLOBAL), 0);
1970 static int __set_memory_enc_dec(unsigned long addr, int numpages, bool enc)
1972 struct cpa_data cpa;
1975 /* Nothing to do if memory encryption is not active */
1976 if (!mem_encrypt_active())
1979 /* Should not be working on unaligned addresses */
1980 if (WARN_ONCE(addr & ~PAGE_MASK, "misaligned address: %#lx\n", addr))
1983 memset(&cpa, 0, sizeof(cpa));
1985 cpa.numpages = numpages;
1986 cpa.mask_set = enc ? __pgprot(_PAGE_ENC) : __pgprot(0);
1987 cpa.mask_clr = enc ? __pgprot(0) : __pgprot(_PAGE_ENC);
1988 cpa.pgd = init_mm.pgd;
1990 /* Must avoid aliasing mappings in the highmem code */
1991 kmap_flush_unused();
1995 * Before changing the encryption attribute, we need to flush caches.
1999 ret = __change_page_attr_set_clr(&cpa, 1);
2002 * After changing the encryption attribute, we need to flush TLBs again
2003 * in case any speculative TLB caching occurred (but no need to flush
2004 * caches again). We could just use cpa_flush_all(), but in case TLB
2005 * flushing gets optimized in the cpa_flush() path use the same logic
2013 int set_memory_encrypted(unsigned long addr, int numpages)
2015 return __set_memory_enc_dec(addr, numpages, true);
2017 EXPORT_SYMBOL_GPL(set_memory_encrypted);
2019 int set_memory_decrypted(unsigned long addr, int numpages)
2021 return __set_memory_enc_dec(addr, numpages, false);
2023 EXPORT_SYMBOL_GPL(set_memory_decrypted);
2025 int set_pages_uc(struct page *page, int numpages)
2027 unsigned long addr = (unsigned long)page_address(page);
2029 return set_memory_uc(addr, numpages);
2031 EXPORT_SYMBOL(set_pages_uc);
2033 static int _set_pages_array(struct page **pages, int numpages,
2034 enum page_cache_mode new_type)
2036 unsigned long start;
2038 enum page_cache_mode set_type;
2043 for (i = 0; i < numpages; i++) {
2044 if (PageHighMem(pages[i]))
2046 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
2047 end = start + PAGE_SIZE;
2048 if (memtype_reserve(start, end, new_type, NULL))
2052 /* If WC, set to UC- first and then WC */
2053 set_type = (new_type == _PAGE_CACHE_MODE_WC) ?
2054 _PAGE_CACHE_MODE_UC_MINUS : new_type;
2056 ret = cpa_set_pages_array(pages, numpages,
2057 cachemode2pgprot(set_type));
2058 if (!ret && new_type == _PAGE_CACHE_MODE_WC)
2059 ret = change_page_attr_set_clr(NULL, numpages,
2061 _PAGE_CACHE_MODE_WC),
2062 __pgprot(_PAGE_CACHE_MASK),
2063 0, CPA_PAGES_ARRAY, pages);
2066 return 0; /* Success */
2069 for (i = 0; i < free_idx; i++) {
2070 if (PageHighMem(pages[i]))
2072 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
2073 end = start + PAGE_SIZE;
2074 memtype_free(start, end);
2079 int set_pages_array_uc(struct page **pages, int numpages)
2081 return _set_pages_array(pages, numpages, _PAGE_CACHE_MODE_UC_MINUS);
2083 EXPORT_SYMBOL(set_pages_array_uc);
2085 int set_pages_array_wc(struct page **pages, int numpages)
2087 return _set_pages_array(pages, numpages, _PAGE_CACHE_MODE_WC);
2089 EXPORT_SYMBOL(set_pages_array_wc);
2091 int set_pages_array_wt(struct page **pages, int numpages)
2093 return _set_pages_array(pages, numpages, _PAGE_CACHE_MODE_WT);
2095 EXPORT_SYMBOL_GPL(set_pages_array_wt);
2097 int set_pages_wb(struct page *page, int numpages)
2099 unsigned long addr = (unsigned long)page_address(page);
2101 return set_memory_wb(addr, numpages);
2103 EXPORT_SYMBOL(set_pages_wb);
2105 int set_pages_array_wb(struct page **pages, int numpages)
2108 unsigned long start;
2112 /* WB cache mode is hard wired to all cache attribute bits being 0 */
2113 retval = cpa_clear_pages_array(pages, numpages,
2114 __pgprot(_PAGE_CACHE_MASK));
2118 for (i = 0; i < numpages; i++) {
2119 if (PageHighMem(pages[i]))
2121 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
2122 end = start + PAGE_SIZE;
2123 memtype_free(start, end);
2128 EXPORT_SYMBOL(set_pages_array_wb);
2130 int set_pages_ro(struct page *page, int numpages)
2132 unsigned long addr = (unsigned long)page_address(page);
2134 return set_memory_ro(addr, numpages);
2137 int set_pages_rw(struct page *page, int numpages)
2139 unsigned long addr = (unsigned long)page_address(page);
2141 return set_memory_rw(addr, numpages);
2144 static int __set_pages_p(struct page *page, int numpages)
2146 unsigned long tempaddr = (unsigned long) page_address(page);
2147 struct cpa_data cpa = { .vaddr = &tempaddr,
2149 .numpages = numpages,
2150 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
2151 .mask_clr = __pgprot(0),
2155 * No alias checking needed for setting present flag. otherwise,
2156 * we may need to break large pages for 64-bit kernel text
2157 * mappings (this adds to complexity if we want to do this from
2158 * atomic context especially). Let's keep it simple!
2160 return __change_page_attr_set_clr(&cpa, 0);
2163 static int __set_pages_np(struct page *page, int numpages)
2165 unsigned long tempaddr = (unsigned long) page_address(page);
2166 struct cpa_data cpa = { .vaddr = &tempaddr,
2168 .numpages = numpages,
2169 .mask_set = __pgprot(0),
2170 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
2174 * No alias checking needed for setting not present flag. otherwise,
2175 * we may need to break large pages for 64-bit kernel text
2176 * mappings (this adds to complexity if we want to do this from
2177 * atomic context especially). Let's keep it simple!
2179 return __change_page_attr_set_clr(&cpa, 0);
2182 int set_direct_map_invalid_noflush(struct page *page)
2184 return __set_pages_np(page, 1);
2187 int set_direct_map_default_noflush(struct page *page)
2189 return __set_pages_p(page, 1);
2192 void __kernel_map_pages(struct page *page, int numpages, int enable)
2194 if (PageHighMem(page))
2197 debug_check_no_locks_freed(page_address(page),
2198 numpages * PAGE_SIZE);
2202 * The return value is ignored as the calls cannot fail.
2203 * Large pages for identity mappings are not used at boot time
2204 * and hence no memory allocations during large page split.
2207 __set_pages_p(page, numpages);
2209 __set_pages_np(page, numpages);
2212 * We should perform an IPI and flush all tlbs,
2213 * but that can deadlock->flush only current cpu.
2214 * Preemption needs to be disabled around __flush_tlb_all() due to
2215 * CR3 reload in __native_flush_tlb().
2221 arch_flush_lazy_mmu_mode();
2224 #ifdef CONFIG_HIBERNATION
2225 bool kernel_page_present(struct page *page)
2230 if (PageHighMem(page))
2233 pte = lookup_address((unsigned long)page_address(page), &level);
2234 return (pte_val(*pte) & _PAGE_PRESENT);
2236 #endif /* CONFIG_HIBERNATION */
2238 int __init kernel_map_pages_in_pgd(pgd_t *pgd, u64 pfn, unsigned long address,
2239 unsigned numpages, unsigned long page_flags)
2241 int retval = -EINVAL;
2243 struct cpa_data cpa = {
2247 .numpages = numpages,
2248 .mask_set = __pgprot(0),
2249 .mask_clr = __pgprot(~page_flags & (_PAGE_NX|_PAGE_RW)),
2253 WARN_ONCE(num_online_cpus() > 1, "Don't call after initializing SMP");
2255 if (!(__supported_pte_mask & _PAGE_NX))
2258 if (!(page_flags & _PAGE_ENC))
2259 cpa.mask_clr = pgprot_encrypted(cpa.mask_clr);
2261 cpa.mask_set = __pgprot(_PAGE_PRESENT | page_flags);
2263 retval = __change_page_attr_set_clr(&cpa, 0);
2271 * __flush_tlb_all() flushes mappings only on current CPU and hence this
2272 * function shouldn't be used in an SMP environment. Presently, it's used only
2273 * during boot (way before smp_init()) by EFI subsystem and hence is ok.
2275 int __init kernel_unmap_pages_in_pgd(pgd_t *pgd, unsigned long address,
2276 unsigned long numpages)
2281 * The typical sequence for unmapping is to find a pte through
2282 * lookup_address_in_pgd() (ideally, it should never return NULL because
2283 * the address is already mapped) and change it's protections. As pfn is
2284 * the *target* of a mapping, it's not useful while unmapping.
2286 struct cpa_data cpa = {
2290 .numpages = numpages,
2291 .mask_set = __pgprot(0),
2292 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
2296 WARN_ONCE(num_online_cpus() > 1, "Don't call after initializing SMP");
2298 retval = __change_page_attr_set_clr(&cpa, 0);
2305 * The testcases use internal knowledge of the implementation that shouldn't
2306 * be exposed to the rest of the kernel. Include these directly here.
2308 #ifdef CONFIG_CPA_DEBUG
2309 #include "cpa-test.c"