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>
19 #include <asm/e820/api.h>
20 #include <asm/processor.h>
21 #include <asm/tlbflush.h>
22 #include <asm/sections.h>
23 #include <asm/setup.h>
24 #include <linux/uaccess.h>
25 #include <asm/pgalloc.h>
26 #include <asm/proto.h>
28 #include <asm/set_memory.h>
30 #include "mm_internal.h"
33 * The current flushing context - we pass it instead of 5 arguments:
40 unsigned long numpages;
41 unsigned long curpage;
44 unsigned int force_split : 1,
45 force_static_prot : 1;
55 static const int cpa_warn_level = CPA_PROTECT;
58 * Serialize cpa() (for !DEBUG_PAGEALLOC which uses large identity mappings)
59 * using cpa_lock. So that we don't allow any other cpu, with stale large tlb
60 * entries change the page attribute in parallel to some other cpu
61 * splitting a large page entry along with changing the attribute.
63 static DEFINE_SPINLOCK(cpa_lock);
65 #define CPA_FLUSHTLB 1
67 #define CPA_PAGES_ARRAY 4
68 #define CPA_NO_CHECK_ALIAS 8 /* Do not search for aliases */
71 static unsigned long direct_pages_count[PG_LEVEL_NUM];
73 void update_page_count(int level, unsigned long pages)
75 /* Protect against CPA */
77 direct_pages_count[level] += pages;
78 spin_unlock(&pgd_lock);
81 static void split_page_count(int level)
83 if (direct_pages_count[level] == 0)
86 direct_pages_count[level]--;
87 direct_pages_count[level - 1] += PTRS_PER_PTE;
90 void arch_report_meminfo(struct seq_file *m)
92 seq_printf(m, "DirectMap4k: %8lu kB\n",
93 direct_pages_count[PG_LEVEL_4K] << 2);
94 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
95 seq_printf(m, "DirectMap2M: %8lu kB\n",
96 direct_pages_count[PG_LEVEL_2M] << 11);
98 seq_printf(m, "DirectMap4M: %8lu kB\n",
99 direct_pages_count[PG_LEVEL_2M] << 12);
102 seq_printf(m, "DirectMap1G: %8lu kB\n",
103 direct_pages_count[PG_LEVEL_1G] << 20);
106 static inline void split_page_count(int level) { }
109 #ifdef CONFIG_X86_CPA_STATISTICS
111 static unsigned long cpa_1g_checked;
112 static unsigned long cpa_1g_sameprot;
113 static unsigned long cpa_1g_preserved;
114 static unsigned long cpa_2m_checked;
115 static unsigned long cpa_2m_sameprot;
116 static unsigned long cpa_2m_preserved;
117 static unsigned long cpa_4k_install;
119 static inline void cpa_inc_1g_checked(void)
124 static inline void cpa_inc_2m_checked(void)
129 static inline void cpa_inc_4k_install(void)
134 static inline void cpa_inc_lp_sameprot(int level)
136 if (level == PG_LEVEL_1G)
142 static inline void cpa_inc_lp_preserved(int level)
144 if (level == PG_LEVEL_1G)
150 static int cpastats_show(struct seq_file *m, void *p)
152 seq_printf(m, "1G pages checked: %16lu\n", cpa_1g_checked);
153 seq_printf(m, "1G pages sameprot: %16lu\n", cpa_1g_sameprot);
154 seq_printf(m, "1G pages preserved: %16lu\n", cpa_1g_preserved);
155 seq_printf(m, "2M pages checked: %16lu\n", cpa_2m_checked);
156 seq_printf(m, "2M pages sameprot: %16lu\n", cpa_2m_sameprot);
157 seq_printf(m, "2M pages preserved: %16lu\n", cpa_2m_preserved);
158 seq_printf(m, "4K pages set-checked: %16lu\n", cpa_4k_install);
162 static int cpastats_open(struct inode *inode, struct file *file)
164 return single_open(file, cpastats_show, NULL);
167 static const struct file_operations cpastats_fops = {
168 .open = cpastats_open,
171 .release = single_release,
174 static int __init cpa_stats_init(void)
176 debugfs_create_file("cpa_stats", S_IRUSR, arch_debugfs_dir, NULL,
180 late_initcall(cpa_stats_init);
182 static inline void cpa_inc_1g_checked(void) { }
183 static inline void cpa_inc_2m_checked(void) { }
184 static inline void cpa_inc_4k_install(void) { }
185 static inline void cpa_inc_lp_sameprot(int level) { }
186 static inline void cpa_inc_lp_preserved(int level) { }
191 within(unsigned long addr, unsigned long start, unsigned long end)
193 return addr >= start && addr < end;
197 within_inclusive(unsigned long addr, unsigned long start, unsigned long end)
199 return addr >= start && addr <= end;
204 static inline unsigned long highmap_start_pfn(void)
206 return __pa_symbol(_text) >> PAGE_SHIFT;
209 static inline unsigned long highmap_end_pfn(void)
211 /* Do not reference physical address outside the kernel. */
212 return __pa_symbol(roundup(_brk_end, PMD_SIZE) - 1) >> PAGE_SHIFT;
215 static bool __cpa_pfn_in_highmap(unsigned long pfn)
218 * Kernel text has an alias mapping at a high address, known
221 return within_inclusive(pfn, highmap_start_pfn(), highmap_end_pfn());
226 static bool __cpa_pfn_in_highmap(unsigned long pfn)
228 /* There is no highmap on 32-bit */
235 * See set_mce_nospec().
237 * Machine check recovery code needs to change cache mode of poisoned pages to
238 * UC to avoid speculative access logging another error. But passing the
239 * address of the 1:1 mapping to set_memory_uc() is a fine way to encourage a
240 * speculative access. So we cheat and flip the top bit of the address. This
241 * works fine for the code that updates the page tables. But at the end of the
242 * process we need to flush the TLB and cache and the non-canonical address
243 * causes a #GP fault when used by the INVLPG and CLFLUSH instructions.
245 * But in the common case we already have a canonical address. This code
246 * will fix the top bit if needed and is a no-op otherwise.
248 static inline unsigned long fix_addr(unsigned long addr)
251 return (long)(addr << 1) >> 1;
257 static unsigned long __cpa_addr(struct cpa_data *cpa, unsigned long idx)
259 if (cpa->flags & CPA_PAGES_ARRAY) {
260 struct page *page = cpa->pages[idx];
262 if (unlikely(PageHighMem(page)))
265 return (unsigned long)page_address(page);
268 if (cpa->flags & CPA_ARRAY)
269 return cpa->vaddr[idx];
271 return *cpa->vaddr + idx * PAGE_SIZE;
278 static void clflush_cache_range_opt(void *vaddr, unsigned int size)
280 const unsigned long clflush_size = boot_cpu_data.x86_clflush_size;
281 void *p = (void *)((unsigned long)vaddr & ~(clflush_size - 1));
282 void *vend = vaddr + size;
287 for (; p < vend; p += clflush_size)
292 * clflush_cache_range - flush a cache range with clflush
293 * @vaddr: virtual start address
294 * @size: number of bytes to flush
296 * CLFLUSHOPT is an unordered instruction which needs fencing with MFENCE or
297 * SFENCE to avoid ordering issues.
299 void clflush_cache_range(void *vaddr, unsigned int size)
302 clflush_cache_range_opt(vaddr, size);
305 EXPORT_SYMBOL_GPL(clflush_cache_range);
307 void arch_invalidate_pmem(void *addr, size_t size)
309 clflush_cache_range(addr, size);
311 EXPORT_SYMBOL_GPL(arch_invalidate_pmem);
313 static void __cpa_flush_all(void *arg)
315 unsigned long cache = (unsigned long)arg;
318 * Flush all to work around Errata in early athlons regarding
319 * large page flushing.
323 if (cache && boot_cpu_data.x86 >= 4)
327 static void cpa_flush_all(unsigned long cache)
329 BUG_ON(irqs_disabled() && !early_boot_irqs_disabled);
331 on_each_cpu(__cpa_flush_all, (void *) cache, 1);
334 void __cpa_flush_tlb(void *data)
336 struct cpa_data *cpa = data;
339 for (i = 0; i < cpa->numpages; i++)
340 __flush_tlb_one_kernel(fix_addr(__cpa_addr(cpa, i)));
343 static void cpa_flush(struct cpa_data *data, int cache)
345 struct cpa_data *cpa = data;
348 BUG_ON(irqs_disabled() && !early_boot_irqs_disabled);
350 if (cache && !static_cpu_has(X86_FEATURE_CLFLUSH)) {
351 cpa_flush_all(cache);
355 if (cpa->numpages <= tlb_single_page_flush_ceiling)
356 on_each_cpu(__cpa_flush_tlb, cpa, 1);
364 for (i = 0; i < cpa->numpages; i++) {
365 unsigned long addr = __cpa_addr(cpa, i);
368 pte_t *pte = lookup_address(addr, &level);
371 * Only flush present addresses:
373 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
374 clflush_cache_range_opt((void *)fix_addr(addr), PAGE_SIZE);
379 static bool overlaps(unsigned long r1_start, unsigned long r1_end,
380 unsigned long r2_start, unsigned long r2_end)
382 return (r1_start <= r2_end && r1_end >= r2_start) ||
383 (r2_start <= r1_end && r2_end >= r1_start);
386 #ifdef CONFIG_PCI_BIOS
388 * The BIOS area between 640k and 1Mb needs to be executable for PCI BIOS
389 * based config access (CONFIG_PCI_GOBIOS) support.
391 #define BIOS_PFN PFN_DOWN(BIOS_BEGIN)
392 #define BIOS_PFN_END PFN_DOWN(BIOS_END - 1)
394 static pgprotval_t protect_pci_bios(unsigned long spfn, unsigned long epfn)
396 if (pcibios_enabled && overlaps(spfn, epfn, BIOS_PFN, BIOS_PFN_END))
401 static pgprotval_t protect_pci_bios(unsigned long spfn, unsigned long epfn)
408 * The .rodata section needs to be read-only. Using the pfn catches all
409 * aliases. This also includes __ro_after_init, so do not enforce until
410 * kernel_set_to_readonly is true.
412 static pgprotval_t protect_rodata(unsigned long spfn, unsigned long epfn)
414 unsigned long epfn_ro, spfn_ro = PFN_DOWN(__pa_symbol(__start_rodata));
417 * Note: __end_rodata is at page aligned and not inclusive, so
418 * subtract 1 to get the last enforced PFN in the rodata area.
420 epfn_ro = PFN_DOWN(__pa_symbol(__end_rodata)) - 1;
422 if (kernel_set_to_readonly && overlaps(spfn, epfn, spfn_ro, epfn_ro))
428 * Protect kernel text against becoming non executable by forbidding
429 * _PAGE_NX. This protects only the high kernel mapping (_text -> _etext)
430 * out of which the kernel actually executes. Do not protect the low
433 * This does not cover __inittext since that is gone after boot.
435 static pgprotval_t protect_kernel_text(unsigned long start, unsigned long end)
437 unsigned long t_end = (unsigned long)_etext - 1;
438 unsigned long t_start = (unsigned long)_text;
440 if (overlaps(start, end, t_start, t_end))
445 #if defined(CONFIG_X86_64)
447 * Once the kernel maps the text as RO (kernel_set_to_readonly is set),
448 * kernel text mappings for the large page aligned text, rodata sections
449 * will be always read-only. For the kernel identity mappings covering the
450 * holes caused by this alignment can be anything that user asks.
452 * This will preserve the large page mappings for kernel text/data at no
455 static pgprotval_t protect_kernel_text_ro(unsigned long start,
458 unsigned long t_end = (unsigned long)__end_rodata_hpage_align - 1;
459 unsigned long t_start = (unsigned long)_text;
462 if (!kernel_set_to_readonly || !overlaps(start, end, t_start, t_end))
465 * Don't enforce the !RW mapping for the kernel text mapping, if
466 * the current mapping is already using small page mapping. No
467 * need to work hard to preserve large page mappings in this case.
469 * This also fixes the Linux Xen paravirt guest boot failure caused
470 * by unexpected read-only mappings for kernel identity
471 * mappings. In this paravirt guest case, the kernel text mapping
472 * and the kernel identity mapping share the same page-table pages,
473 * so the protections for kernel text and identity mappings have to
476 if (lookup_address(start, &level) && (level != PG_LEVEL_4K))
481 static pgprotval_t protect_kernel_text_ro(unsigned long start,
488 static inline bool conflicts(pgprot_t prot, pgprotval_t val)
490 return (pgprot_val(prot) & ~val) != pgprot_val(prot);
493 static inline void check_conflict(int warnlvl, pgprot_t prot, pgprotval_t val,
494 unsigned long start, unsigned long end,
495 unsigned long pfn, const char *txt)
497 static const char *lvltxt[] = {
498 [CPA_CONFLICT] = "conflict",
499 [CPA_PROTECT] = "protect",
500 [CPA_DETECT] = "detect",
503 if (warnlvl > cpa_warn_level || !conflicts(prot, val))
506 pr_warn("CPA %8s %10s: 0x%016lx - 0x%016lx PFN %lx req %016llx prevent %016llx\n",
507 lvltxt[warnlvl], txt, start, end, pfn, (unsigned long long)pgprot_val(prot),
508 (unsigned long long)val);
512 * Certain areas of memory on x86 require very specific protection flags,
513 * for example the BIOS area or kernel text. Callers don't always get this
514 * right (again, ioremap() on BIOS memory is not uncommon) so this function
515 * checks and fixes these known static required protection bits.
517 static inline pgprot_t static_protections(pgprot_t prot, unsigned long start,
518 unsigned long pfn, unsigned long npg,
519 unsigned long lpsize, int warnlvl)
521 pgprotval_t forbidden, res;
525 * There is no point in checking RW/NX conflicts when the requested
526 * mapping is setting the page !PRESENT.
528 if (!(pgprot_val(prot) & _PAGE_PRESENT))
531 /* Operate on the virtual address */
532 end = start + npg * PAGE_SIZE - 1;
534 res = protect_kernel_text(start, end);
535 check_conflict(warnlvl, prot, res, start, end, pfn, "Text NX");
539 * Special case to preserve a large page. If the change spawns the
540 * full large page mapping then there is no point to split it
541 * up. Happens with ftrace and is going to be removed once ftrace
542 * switched to text_poke().
544 if (lpsize != (npg * PAGE_SIZE) || (start & (lpsize - 1))) {
545 res = protect_kernel_text_ro(start, end);
546 check_conflict(warnlvl, prot, res, start, end, pfn, "Text RO");
550 /* Check the PFN directly */
551 res = protect_pci_bios(pfn, pfn + npg - 1);
552 check_conflict(warnlvl, prot, res, start, end, pfn, "PCIBIOS NX");
555 res = protect_rodata(pfn, pfn + npg - 1);
556 check_conflict(warnlvl, prot, res, start, end, pfn, "Rodata RO");
559 return __pgprot(pgprot_val(prot) & ~forbidden);
563 * Lookup the page table entry for a virtual address in a specific pgd.
564 * Return a pointer to the entry and the level of the mapping.
566 pte_t *lookup_address_in_pgd(pgd_t *pgd, unsigned long address,
573 *level = PG_LEVEL_NONE;
578 p4d = p4d_offset(pgd, address);
582 *level = PG_LEVEL_512G;
583 if (p4d_large(*p4d) || !p4d_present(*p4d))
586 pud = pud_offset(p4d, address);
590 *level = PG_LEVEL_1G;
591 if (pud_large(*pud) || !pud_present(*pud))
594 pmd = pmd_offset(pud, address);
598 *level = PG_LEVEL_2M;
599 if (pmd_large(*pmd) || !pmd_present(*pmd))
602 *level = PG_LEVEL_4K;
604 return pte_offset_kernel(pmd, address);
608 * Lookup the page table entry for a virtual address. Return a pointer
609 * to the entry and the level of the mapping.
611 * Note: We return pud and pmd either when the entry is marked large
612 * or when the present bit is not set. Otherwise we would return a
613 * pointer to a nonexisting mapping.
615 pte_t *lookup_address(unsigned long address, unsigned int *level)
617 return lookup_address_in_pgd(pgd_offset_k(address), address, level);
619 EXPORT_SYMBOL_GPL(lookup_address);
621 static pte_t *_lookup_address_cpa(struct cpa_data *cpa, unsigned long address,
625 return lookup_address_in_pgd(cpa->pgd + pgd_index(address),
628 return lookup_address(address, level);
632 * Lookup the PMD entry for a virtual address. Return a pointer to the entry
633 * or NULL if not present.
635 pmd_t *lookup_pmd_address(unsigned long address)
641 pgd = pgd_offset_k(address);
645 p4d = p4d_offset(pgd, address);
646 if (p4d_none(*p4d) || p4d_large(*p4d) || !p4d_present(*p4d))
649 pud = pud_offset(p4d, address);
650 if (pud_none(*pud) || pud_large(*pud) || !pud_present(*pud))
653 return pmd_offset(pud, address);
657 * This is necessary because __pa() does not work on some
658 * kinds of memory, like vmalloc() or the alloc_remap()
659 * areas on 32-bit NUMA systems. The percpu areas can
660 * end up in this kind of memory, for instance.
662 * This could be optimized, but it is only intended to be
663 * used at inititalization time, and keeping it
664 * unoptimized should increase the testing coverage for
665 * the more obscure platforms.
667 phys_addr_t slow_virt_to_phys(void *__virt_addr)
669 unsigned long virt_addr = (unsigned long)__virt_addr;
670 phys_addr_t phys_addr;
671 unsigned long offset;
675 pte = lookup_address(virt_addr, &level);
679 * pXX_pfn() returns unsigned long, which must be cast to phys_addr_t
680 * before being left-shifted PAGE_SHIFT bits -- this trick is to
681 * make 32-PAE kernel work correctly.
685 phys_addr = (phys_addr_t)pud_pfn(*(pud_t *)pte) << PAGE_SHIFT;
686 offset = virt_addr & ~PUD_PAGE_MASK;
689 phys_addr = (phys_addr_t)pmd_pfn(*(pmd_t *)pte) << PAGE_SHIFT;
690 offset = virt_addr & ~PMD_PAGE_MASK;
693 phys_addr = (phys_addr_t)pte_pfn(*pte) << PAGE_SHIFT;
694 offset = virt_addr & ~PAGE_MASK;
697 return (phys_addr_t)(phys_addr | offset);
699 EXPORT_SYMBOL_GPL(slow_virt_to_phys);
702 * Set the new pmd in all the pgds we know about:
704 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
707 set_pte_atomic(kpte, pte);
709 if (!SHARED_KERNEL_PMD) {
712 list_for_each_entry(page, &pgd_list, lru) {
718 pgd = (pgd_t *)page_address(page) + pgd_index(address);
719 p4d = p4d_offset(pgd, address);
720 pud = pud_offset(p4d, address);
721 pmd = pmd_offset(pud, address);
722 set_pte_atomic((pte_t *)pmd, pte);
728 static pgprot_t pgprot_clear_protnone_bits(pgprot_t prot)
731 * _PAGE_GLOBAL means "global page" for present PTEs.
732 * But, it is also used to indicate _PAGE_PROTNONE
733 * for non-present PTEs.
735 * This ensures that a _PAGE_GLOBAL PTE going from
736 * present to non-present is not confused as
739 if (!(pgprot_val(prot) & _PAGE_PRESENT))
740 pgprot_val(prot) &= ~_PAGE_GLOBAL;
745 static int __should_split_large_page(pte_t *kpte, unsigned long address,
746 struct cpa_data *cpa)
748 unsigned long numpages, pmask, psize, lpaddr, pfn, old_pfn;
749 pgprot_t old_prot, new_prot, req_prot, chk_prot;
754 * Check for races, another CPU might have split this page
757 tmp = _lookup_address_cpa(cpa, address, &level);
763 old_prot = pmd_pgprot(*(pmd_t *)kpte);
764 old_pfn = pmd_pfn(*(pmd_t *)kpte);
765 cpa_inc_2m_checked();
768 old_prot = pud_pgprot(*(pud_t *)kpte);
769 old_pfn = pud_pfn(*(pud_t *)kpte);
770 cpa_inc_1g_checked();
776 psize = page_level_size(level);
777 pmask = page_level_mask(level);
780 * Calculate the number of pages, which fit into this large
781 * page starting at address:
783 lpaddr = (address + psize) & pmask;
784 numpages = (lpaddr - address) >> PAGE_SHIFT;
785 if (numpages < cpa->numpages)
786 cpa->numpages = numpages;
789 * We are safe now. Check whether the new pgprot is the same:
790 * Convert protection attributes to 4k-format, as cpa->mask* are set
794 /* Clear PSE (aka _PAGE_PAT) and move PAT bit to correct position */
795 req_prot = pgprot_large_2_4k(old_prot);
797 pgprot_val(req_prot) &= ~pgprot_val(cpa->mask_clr);
798 pgprot_val(req_prot) |= pgprot_val(cpa->mask_set);
801 * req_prot is in format of 4k pages. It must be converted to large
802 * page format: the caching mode includes the PAT bit located at
803 * different bit positions in the two formats.
805 req_prot = pgprot_4k_2_large(req_prot);
806 req_prot = pgprot_clear_protnone_bits(req_prot);
807 if (pgprot_val(req_prot) & _PAGE_PRESENT)
808 pgprot_val(req_prot) |= _PAGE_PSE;
811 * old_pfn points to the large page base pfn. So we need to add the
812 * offset of the virtual address:
814 pfn = old_pfn + ((address & (psize - 1)) >> PAGE_SHIFT);
818 * Calculate the large page base address and the number of 4K pages
821 lpaddr = address & pmask;
822 numpages = psize >> PAGE_SHIFT;
825 * Sanity check that the existing mapping is correct versus the static
826 * protections. static_protections() guards against !PRESENT, so no
827 * extra conditional required here.
829 chk_prot = static_protections(old_prot, lpaddr, old_pfn, numpages,
830 psize, CPA_CONFLICT);
832 if (WARN_ON_ONCE(pgprot_val(chk_prot) != pgprot_val(old_prot))) {
834 * Split the large page and tell the split code to
835 * enforce static protections.
837 cpa->force_static_prot = 1;
842 * Optimization: If the requested pgprot is the same as the current
843 * pgprot, then the large page can be preserved and no updates are
844 * required independent of alignment and length of the requested
845 * range. The above already established that the current pgprot is
846 * correct, which in consequence makes the requested pgprot correct
847 * as well if it is the same. The static protection scan below will
848 * not come to a different conclusion.
850 if (pgprot_val(req_prot) == pgprot_val(old_prot)) {
851 cpa_inc_lp_sameprot(level);
856 * If the requested range does not cover the full page, split it up
858 if (address != lpaddr || cpa->numpages != numpages)
862 * Check whether the requested pgprot is conflicting with a static
863 * protection requirement in the large page.
865 new_prot = static_protections(req_prot, lpaddr, old_pfn, numpages,
869 * If there is a conflict, split the large page.
871 * There used to be a 4k wise evaluation trying really hard to
872 * preserve the large pages, but experimentation has shown, that this
873 * does not help at all. There might be corner cases which would
874 * preserve one large page occasionally, but it's really not worth the
875 * extra code and cycles for the common case.
877 if (pgprot_val(req_prot) != pgprot_val(new_prot))
880 /* All checks passed. Update the large page mapping. */
881 new_pte = pfn_pte(old_pfn, new_prot);
882 __set_pmd_pte(kpte, address, new_pte);
883 cpa->flags |= CPA_FLUSHTLB;
884 cpa_inc_lp_preserved(level);
888 static int should_split_large_page(pte_t *kpte, unsigned long address,
889 struct cpa_data *cpa)
893 if (cpa->force_split)
896 spin_lock(&pgd_lock);
897 do_split = __should_split_large_page(kpte, address, cpa);
898 spin_unlock(&pgd_lock);
903 static void split_set_pte(struct cpa_data *cpa, pte_t *pte, unsigned long pfn,
904 pgprot_t ref_prot, unsigned long address,
907 unsigned int npg = PFN_DOWN(size);
911 * If should_split_large_page() discovered an inconsistent mapping,
912 * remove the invalid protection in the split mapping.
914 if (!cpa->force_static_prot)
917 /* Hand in lpsize = 0 to enforce the protection mechanism */
918 prot = static_protections(ref_prot, address, pfn, npg, 0, CPA_PROTECT);
920 if (pgprot_val(prot) == pgprot_val(ref_prot))
924 * If this is splitting a PMD, fix it up. PUD splits cannot be
925 * fixed trivially as that would require to rescan the newly
926 * installed PMD mappings after returning from split_large_page()
927 * so an eventual further split can allocate the necessary PTE
928 * pages. Warn for now and revisit it in case this actually
931 if (size == PAGE_SIZE)
934 pr_warn_once("CPA: Cannot fixup static protections for PUD split\n");
936 set_pte(pte, pfn_pte(pfn, ref_prot));
940 __split_large_page(struct cpa_data *cpa, pte_t *kpte, unsigned long address,
943 unsigned long lpaddr, lpinc, ref_pfn, pfn, pfninc = 1;
944 pte_t *pbase = (pte_t *)page_address(base);
945 unsigned int i, level;
949 spin_lock(&pgd_lock);
951 * Check for races, another CPU might have split this page
954 tmp = _lookup_address_cpa(cpa, address, &level);
956 spin_unlock(&pgd_lock);
960 paravirt_alloc_pte(&init_mm, page_to_pfn(base));
964 ref_prot = pmd_pgprot(*(pmd_t *)kpte);
966 * Clear PSE (aka _PAGE_PAT) and move
967 * PAT bit to correct position.
969 ref_prot = pgprot_large_2_4k(ref_prot);
970 ref_pfn = pmd_pfn(*(pmd_t *)kpte);
971 lpaddr = address & PMD_MASK;
976 ref_prot = pud_pgprot(*(pud_t *)kpte);
977 ref_pfn = pud_pfn(*(pud_t *)kpte);
978 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
979 lpaddr = address & PUD_MASK;
982 * Clear the PSE flags if the PRESENT flag is not set
983 * otherwise pmd_present/pmd_huge will return true
984 * even on a non present pmd.
986 if (!(pgprot_val(ref_prot) & _PAGE_PRESENT))
987 pgprot_val(ref_prot) &= ~_PAGE_PSE;
991 spin_unlock(&pgd_lock);
995 ref_prot = pgprot_clear_protnone_bits(ref_prot);
998 * Get the target pfn from the original entry:
1001 for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc, lpaddr += lpinc)
1002 split_set_pte(cpa, pbase + i, pfn, ref_prot, lpaddr, lpinc);
1004 if (virt_addr_valid(address)) {
1005 unsigned long pfn = PFN_DOWN(__pa(address));
1007 if (pfn_range_is_mapped(pfn, pfn + 1))
1008 split_page_count(level);
1012 * Install the new, split up pagetable.
1014 * We use the standard kernel pagetable protections for the new
1015 * pagetable protections, the actual ptes set above control the
1016 * primary protection behavior:
1018 __set_pmd_pte(kpte, address, mk_pte(base, __pgprot(_KERNPG_TABLE)));
1021 * Do a global flush tlb after splitting the large page
1022 * and before we do the actual change page attribute in the PTE.
1024 * Without this, we violate the TLB application note, that says:
1025 * "The TLBs may contain both ordinary and large-page
1026 * translations for a 4-KByte range of linear addresses. This
1027 * may occur if software modifies the paging structures so that
1028 * the page size used for the address range changes. If the two
1029 * translations differ with respect to page frame or attributes
1030 * (e.g., permissions), processor behavior is undefined and may
1031 * be implementation-specific."
1033 * We do this global tlb flush inside the cpa_lock, so that we
1034 * don't allow any other cpu, with stale tlb entries change the
1035 * page attribute in parallel, that also falls into the
1036 * just split large page entry.
1039 spin_unlock(&pgd_lock);
1044 static int split_large_page(struct cpa_data *cpa, pte_t *kpte,
1045 unsigned long address)
1049 if (!debug_pagealloc_enabled())
1050 spin_unlock(&cpa_lock);
1051 base = alloc_pages(GFP_KERNEL, 0);
1052 if (!debug_pagealloc_enabled())
1053 spin_lock(&cpa_lock);
1057 if (__split_large_page(cpa, kpte, address, base))
1063 static bool try_to_free_pte_page(pte_t *pte)
1067 for (i = 0; i < PTRS_PER_PTE; i++)
1068 if (!pte_none(pte[i]))
1071 free_page((unsigned long)pte);
1075 static bool try_to_free_pmd_page(pmd_t *pmd)
1079 for (i = 0; i < PTRS_PER_PMD; i++)
1080 if (!pmd_none(pmd[i]))
1083 free_page((unsigned long)pmd);
1087 static bool unmap_pte_range(pmd_t *pmd, unsigned long start, unsigned long end)
1089 pte_t *pte = pte_offset_kernel(pmd, start);
1091 while (start < end) {
1092 set_pte(pte, __pte(0));
1098 if (try_to_free_pte_page((pte_t *)pmd_page_vaddr(*pmd))) {
1105 static void __unmap_pmd_range(pud_t *pud, pmd_t *pmd,
1106 unsigned long start, unsigned long end)
1108 if (unmap_pte_range(pmd, start, end))
1109 if (try_to_free_pmd_page((pmd_t *)pud_page_vaddr(*pud)))
1113 static void unmap_pmd_range(pud_t *pud, unsigned long start, unsigned long end)
1115 pmd_t *pmd = pmd_offset(pud, start);
1118 * Not on a 2MB page boundary?
1120 if (start & (PMD_SIZE - 1)) {
1121 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
1122 unsigned long pre_end = min_t(unsigned long, end, next_page);
1124 __unmap_pmd_range(pud, pmd, start, pre_end);
1131 * Try to unmap in 2M chunks.
1133 while (end - start >= PMD_SIZE) {
1134 if (pmd_large(*pmd))
1137 __unmap_pmd_range(pud, pmd, start, start + PMD_SIZE);
1147 return __unmap_pmd_range(pud, pmd, start, end);
1150 * Try again to free the PMD page if haven't succeeded above.
1152 if (!pud_none(*pud))
1153 if (try_to_free_pmd_page((pmd_t *)pud_page_vaddr(*pud)))
1157 static void unmap_pud_range(p4d_t *p4d, unsigned long start, unsigned long end)
1159 pud_t *pud = pud_offset(p4d, start);
1162 * Not on a GB page boundary?
1164 if (start & (PUD_SIZE - 1)) {
1165 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
1166 unsigned long pre_end = min_t(unsigned long, end, next_page);
1168 unmap_pmd_range(pud, start, pre_end);
1175 * Try to unmap in 1G chunks?
1177 while (end - start >= PUD_SIZE) {
1179 if (pud_large(*pud))
1182 unmap_pmd_range(pud, start, start + PUD_SIZE);
1192 unmap_pmd_range(pud, start, end);
1195 * No need to try to free the PUD page because we'll free it in
1196 * populate_pgd's error path
1200 static int alloc_pte_page(pmd_t *pmd)
1202 pte_t *pte = (pte_t *)get_zeroed_page(GFP_KERNEL);
1206 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
1210 static int alloc_pmd_page(pud_t *pud)
1212 pmd_t *pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL);
1216 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
1220 static void populate_pte(struct cpa_data *cpa,
1221 unsigned long start, unsigned long end,
1222 unsigned num_pages, pmd_t *pmd, pgprot_t pgprot)
1226 pte = pte_offset_kernel(pmd, start);
1228 pgprot = pgprot_clear_protnone_bits(pgprot);
1230 while (num_pages-- && start < end) {
1231 set_pte(pte, pfn_pte(cpa->pfn, pgprot));
1239 static long populate_pmd(struct cpa_data *cpa,
1240 unsigned long start, unsigned long end,
1241 unsigned num_pages, pud_t *pud, pgprot_t pgprot)
1245 pgprot_t pmd_pgprot;
1248 * Not on a 2M boundary?
1250 if (start & (PMD_SIZE - 1)) {
1251 unsigned long pre_end = start + (num_pages << PAGE_SHIFT);
1252 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
1254 pre_end = min_t(unsigned long, pre_end, next_page);
1255 cur_pages = (pre_end - start) >> PAGE_SHIFT;
1256 cur_pages = min_t(unsigned int, num_pages, cur_pages);
1261 pmd = pmd_offset(pud, start);
1263 if (alloc_pte_page(pmd))
1266 populate_pte(cpa, start, pre_end, cur_pages, pmd, pgprot);
1272 * We mapped them all?
1274 if (num_pages == cur_pages)
1277 pmd_pgprot = pgprot_4k_2_large(pgprot);
1279 while (end - start >= PMD_SIZE) {
1282 * We cannot use a 1G page so allocate a PMD page if needed.
1285 if (alloc_pmd_page(pud))
1288 pmd = pmd_offset(pud, start);
1290 set_pmd(pmd, pmd_mkhuge(pfn_pmd(cpa->pfn,
1291 canon_pgprot(pmd_pgprot))));
1294 cpa->pfn += PMD_SIZE >> PAGE_SHIFT;
1295 cur_pages += PMD_SIZE >> PAGE_SHIFT;
1299 * Map trailing 4K pages.
1302 pmd = pmd_offset(pud, start);
1304 if (alloc_pte_page(pmd))
1307 populate_pte(cpa, start, end, num_pages - cur_pages,
1313 static int populate_pud(struct cpa_data *cpa, unsigned long start, p4d_t *p4d,
1319 pgprot_t pud_pgprot;
1321 end = start + (cpa->numpages << PAGE_SHIFT);
1324 * Not on a Gb page boundary? => map everything up to it with
1327 if (start & (PUD_SIZE - 1)) {
1328 unsigned long pre_end;
1329 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
1331 pre_end = min_t(unsigned long, end, next_page);
1332 cur_pages = (pre_end - start) >> PAGE_SHIFT;
1333 cur_pages = min_t(int, (int)cpa->numpages, cur_pages);
1335 pud = pud_offset(p4d, start);
1341 if (alloc_pmd_page(pud))
1344 cur_pages = populate_pmd(cpa, start, pre_end, cur_pages,
1352 /* We mapped them all? */
1353 if (cpa->numpages == cur_pages)
1356 pud = pud_offset(p4d, start);
1357 pud_pgprot = pgprot_4k_2_large(pgprot);
1360 * Map everything starting from the Gb boundary, possibly with 1G pages
1362 while (boot_cpu_has(X86_FEATURE_GBPAGES) && end - start >= PUD_SIZE) {
1363 set_pud(pud, pud_mkhuge(pfn_pud(cpa->pfn,
1364 canon_pgprot(pud_pgprot))));
1367 cpa->pfn += PUD_SIZE >> PAGE_SHIFT;
1368 cur_pages += PUD_SIZE >> PAGE_SHIFT;
1372 /* Map trailing leftover */
1376 pud = pud_offset(p4d, start);
1378 if (alloc_pmd_page(pud))
1381 tmp = populate_pmd(cpa, start, end, cpa->numpages - cur_pages,
1392 * Restrictions for kernel page table do not necessarily apply when mapping in
1395 static int populate_pgd(struct cpa_data *cpa, unsigned long addr)
1397 pgprot_t pgprot = __pgprot(_KERNPG_TABLE);
1398 pud_t *pud = NULL; /* shut up gcc */
1403 pgd_entry = cpa->pgd + pgd_index(addr);
1405 if (pgd_none(*pgd_entry)) {
1406 p4d = (p4d_t *)get_zeroed_page(GFP_KERNEL);
1410 set_pgd(pgd_entry, __pgd(__pa(p4d) | _KERNPG_TABLE));
1414 * Allocate a PUD page and hand it down for mapping.
1416 p4d = p4d_offset(pgd_entry, addr);
1417 if (p4d_none(*p4d)) {
1418 pud = (pud_t *)get_zeroed_page(GFP_KERNEL);
1422 set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE));
1425 pgprot_val(pgprot) &= ~pgprot_val(cpa->mask_clr);
1426 pgprot_val(pgprot) |= pgprot_val(cpa->mask_set);
1428 ret = populate_pud(cpa, addr, p4d, pgprot);
1431 * Leave the PUD page in place in case some other CPU or thread
1432 * already found it, but remove any useless entries we just
1435 unmap_pud_range(p4d, addr,
1436 addr + (cpa->numpages << PAGE_SHIFT));
1440 cpa->numpages = ret;
1444 static int __cpa_process_fault(struct cpa_data *cpa, unsigned long vaddr,
1449 * Right now, we only execute this code path when mapping
1450 * the EFI virtual memory map regions, no other users
1451 * provide a ->pgd value. This may change in the future.
1453 return populate_pgd(cpa, vaddr);
1457 * Ignore all non primary paths.
1465 * Ignore the NULL PTE for kernel identity mapping, as it is expected
1467 * Also set numpages to '1' indicating that we processed cpa req for
1468 * one virtual address page and its pfn. TBD: numpages can be set based
1469 * on the initial value and the level returned by lookup_address().
1471 if (within(vaddr, PAGE_OFFSET,
1472 PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
1474 cpa->pfn = __pa(vaddr) >> PAGE_SHIFT;
1477 } else if (__cpa_pfn_in_highmap(cpa->pfn)) {
1478 /* Faults in the highmap are OK, so do not warn: */
1481 WARN(1, KERN_WARNING "CPA: called for zero pte. "
1482 "vaddr = %lx cpa->vaddr = %lx\n", vaddr,
1489 static int __change_page_attr(struct cpa_data *cpa, int primary)
1491 unsigned long address;
1494 pte_t *kpte, old_pte;
1496 address = __cpa_addr(cpa, cpa->curpage);
1498 kpte = _lookup_address_cpa(cpa, address, &level);
1500 return __cpa_process_fault(cpa, address, primary);
1503 if (pte_none(old_pte))
1504 return __cpa_process_fault(cpa, address, primary);
1506 if (level == PG_LEVEL_4K) {
1508 pgprot_t new_prot = pte_pgprot(old_pte);
1509 unsigned long pfn = pte_pfn(old_pte);
1511 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
1512 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
1514 cpa_inc_4k_install();
1515 /* Hand in lpsize = 0 to enforce the protection mechanism */
1516 new_prot = static_protections(new_prot, address, pfn, 1, 0,
1519 new_prot = pgprot_clear_protnone_bits(new_prot);
1522 * We need to keep the pfn from the existing PTE,
1523 * after all we're only going to change it's attributes
1524 * not the memory it points to
1526 new_pte = pfn_pte(pfn, new_prot);
1529 * Do we really change anything ?
1531 if (pte_val(old_pte) != pte_val(new_pte)) {
1532 set_pte_atomic(kpte, new_pte);
1533 cpa->flags |= CPA_FLUSHTLB;
1540 * Check, whether we can keep the large page intact
1541 * and just change the pte:
1543 do_split = should_split_large_page(kpte, address, cpa);
1545 * When the range fits into the existing large page,
1546 * return. cp->numpages and cpa->tlbflush have been updated in
1553 * We have to split the large page:
1555 err = split_large_page(cpa, kpte, address);
1562 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
1564 static int cpa_process_alias(struct cpa_data *cpa)
1566 struct cpa_data alias_cpa;
1567 unsigned long laddr = (unsigned long)__va(cpa->pfn << PAGE_SHIFT);
1568 unsigned long vaddr;
1571 if (!pfn_range_is_mapped(cpa->pfn, cpa->pfn + 1))
1575 * No need to redo, when the primary call touched the direct
1578 vaddr = __cpa_addr(cpa, cpa->curpage);
1579 if (!(within(vaddr, PAGE_OFFSET,
1580 PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT)))) {
1583 alias_cpa.vaddr = &laddr;
1584 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1585 alias_cpa.curpage = 0;
1587 ret = __change_page_attr_set_clr(&alias_cpa, 0);
1592 #ifdef CONFIG_X86_64
1594 * If the primary call didn't touch the high mapping already
1595 * and the physical address is inside the kernel map, we need
1596 * to touch the high mapped kernel as well:
1598 if (!within(vaddr, (unsigned long)_text, _brk_end) &&
1599 __cpa_pfn_in_highmap(cpa->pfn)) {
1600 unsigned long temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) +
1601 __START_KERNEL_map - phys_base;
1603 alias_cpa.vaddr = &temp_cpa_vaddr;
1604 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1605 alias_cpa.curpage = 0;
1608 * The high mapping range is imprecise, so ignore the
1611 __change_page_attr_set_clr(&alias_cpa, 0);
1618 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
1620 unsigned long numpages = cpa->numpages;
1621 unsigned long rempages = numpages;
1626 * Store the remaining nr of pages for the large page
1627 * preservation check.
1629 cpa->numpages = rempages;
1630 /* for array changes, we can't use large page */
1631 if (cpa->flags & (CPA_ARRAY | CPA_PAGES_ARRAY))
1634 if (!debug_pagealloc_enabled())
1635 spin_lock(&cpa_lock);
1636 ret = __change_page_attr(cpa, checkalias);
1637 if (!debug_pagealloc_enabled())
1638 spin_unlock(&cpa_lock);
1643 ret = cpa_process_alias(cpa);
1649 * Adjust the number of pages with the result of the
1650 * CPA operation. Either a large page has been
1651 * preserved or a single page update happened.
1653 BUG_ON(cpa->numpages > rempages || !cpa->numpages);
1654 rempages -= cpa->numpages;
1655 cpa->curpage += cpa->numpages;
1659 /* Restore the original numpages */
1660 cpa->numpages = numpages;
1664 static int change_page_attr_set_clr(unsigned long *addr, int numpages,
1665 pgprot_t mask_set, pgprot_t mask_clr,
1666 int force_split, int in_flag,
1667 struct page **pages)
1669 struct cpa_data cpa;
1670 int ret, cache, checkalias;
1672 memset(&cpa, 0, sizeof(cpa));
1675 * Check, if we are requested to set a not supported
1676 * feature. Clearing non-supported features is OK.
1678 mask_set = canon_pgprot(mask_set);
1680 if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
1683 /* Ensure we are PAGE_SIZE aligned */
1684 if (in_flag & CPA_ARRAY) {
1686 for (i = 0; i < numpages; i++) {
1687 if (addr[i] & ~PAGE_MASK) {
1688 addr[i] &= PAGE_MASK;
1692 } else if (!(in_flag & CPA_PAGES_ARRAY)) {
1694 * in_flag of CPA_PAGES_ARRAY implies it is aligned.
1695 * No need to check in that case
1697 if (*addr & ~PAGE_MASK) {
1700 * People should not be passing in unaligned addresses:
1706 /* Must avoid aliasing mappings in the highmem code */
1707 kmap_flush_unused();
1713 cpa.numpages = numpages;
1714 cpa.mask_set = mask_set;
1715 cpa.mask_clr = mask_clr;
1718 cpa.force_split = force_split;
1720 if (in_flag & (CPA_ARRAY | CPA_PAGES_ARRAY))
1721 cpa.flags |= in_flag;
1723 /* No alias checking for _NX bit modifications */
1724 checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
1725 /* Has caller explicitly disabled alias checking? */
1726 if (in_flag & CPA_NO_CHECK_ALIAS)
1729 ret = __change_page_attr_set_clr(&cpa, checkalias);
1732 * Check whether we really changed something:
1734 if (!(cpa.flags & CPA_FLUSHTLB))
1738 * No need to flush, when we did not set any of the caching
1741 cache = !!pgprot2cachemode(mask_set);
1744 * On error; flush everything to be sure.
1747 cpa_flush_all(cache);
1751 cpa_flush(&cpa, cache);
1756 static inline int change_page_attr_set(unsigned long *addr, int numpages,
1757 pgprot_t mask, int array)
1759 return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
1760 (array ? CPA_ARRAY : 0), NULL);
1763 static inline int change_page_attr_clear(unsigned long *addr, int numpages,
1764 pgprot_t mask, int array)
1766 return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
1767 (array ? CPA_ARRAY : 0), NULL);
1770 static inline int cpa_set_pages_array(struct page **pages, int numpages,
1773 return change_page_attr_set_clr(NULL, numpages, mask, __pgprot(0), 0,
1774 CPA_PAGES_ARRAY, pages);
1777 static inline int cpa_clear_pages_array(struct page **pages, int numpages,
1780 return change_page_attr_set_clr(NULL, numpages, __pgprot(0), mask, 0,
1781 CPA_PAGES_ARRAY, pages);
1784 int _set_memory_uc(unsigned long addr, int numpages)
1787 * for now UC MINUS. see comments in ioremap_nocache()
1788 * If you really need strong UC use ioremap_uc(), but note
1789 * that you cannot override IO areas with set_memory_*() as
1790 * these helpers cannot work with IO memory.
1792 return change_page_attr_set(&addr, numpages,
1793 cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1797 int set_memory_uc(unsigned long addr, int numpages)
1802 * for now UC MINUS. see comments in ioremap_nocache()
1804 ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1805 _PAGE_CACHE_MODE_UC_MINUS, NULL);
1809 ret = _set_memory_uc(addr, numpages);
1816 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1820 EXPORT_SYMBOL(set_memory_uc);
1822 static int _set_memory_array(unsigned long *addr, int numpages,
1823 enum page_cache_mode new_type)
1825 enum page_cache_mode set_type;
1829 for (i = 0; i < numpages; i++) {
1830 ret = reserve_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE,
1836 /* If WC, set to UC- first and then WC */
1837 set_type = (new_type == _PAGE_CACHE_MODE_WC) ?
1838 _PAGE_CACHE_MODE_UC_MINUS : new_type;
1840 ret = change_page_attr_set(addr, numpages,
1841 cachemode2pgprot(set_type), 1);
1843 if (!ret && new_type == _PAGE_CACHE_MODE_WC)
1844 ret = change_page_attr_set_clr(addr, numpages,
1846 _PAGE_CACHE_MODE_WC),
1847 __pgprot(_PAGE_CACHE_MASK),
1848 0, CPA_ARRAY, NULL);
1855 for (j = 0; j < i; j++)
1856 free_memtype(__pa(addr[j]), __pa(addr[j]) + PAGE_SIZE);
1861 int set_memory_array_uc(unsigned long *addr, int numpages)
1863 return _set_memory_array(addr, numpages, _PAGE_CACHE_MODE_UC_MINUS);
1865 EXPORT_SYMBOL(set_memory_array_uc);
1867 int set_memory_array_wc(unsigned long *addr, int numpages)
1869 return _set_memory_array(addr, numpages, _PAGE_CACHE_MODE_WC);
1871 EXPORT_SYMBOL(set_memory_array_wc);
1873 int set_memory_array_wt(unsigned long *addr, int numpages)
1875 return _set_memory_array(addr, numpages, _PAGE_CACHE_MODE_WT);
1877 EXPORT_SYMBOL_GPL(set_memory_array_wt);
1879 int _set_memory_wc(unsigned long addr, int numpages)
1883 ret = change_page_attr_set(&addr, numpages,
1884 cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1887 ret = change_page_attr_set_clr(&addr, numpages,
1888 cachemode2pgprot(_PAGE_CACHE_MODE_WC),
1889 __pgprot(_PAGE_CACHE_MASK),
1895 int set_memory_wc(unsigned long addr, int numpages)
1899 ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1900 _PAGE_CACHE_MODE_WC, NULL);
1904 ret = _set_memory_wc(addr, numpages);
1906 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1910 EXPORT_SYMBOL(set_memory_wc);
1912 int _set_memory_wt(unsigned long addr, int numpages)
1914 return change_page_attr_set(&addr, numpages,
1915 cachemode2pgprot(_PAGE_CACHE_MODE_WT), 0);
1918 int set_memory_wt(unsigned long addr, int numpages)
1922 ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1923 _PAGE_CACHE_MODE_WT, NULL);
1927 ret = _set_memory_wt(addr, numpages);
1929 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1933 EXPORT_SYMBOL_GPL(set_memory_wt);
1935 int _set_memory_wb(unsigned long addr, int numpages)
1937 /* WB cache mode is hard wired to all cache attribute bits being 0 */
1938 return change_page_attr_clear(&addr, numpages,
1939 __pgprot(_PAGE_CACHE_MASK), 0);
1942 int set_memory_wb(unsigned long addr, int numpages)
1946 ret = _set_memory_wb(addr, numpages);
1950 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1953 EXPORT_SYMBOL(set_memory_wb);
1955 int set_memory_array_wb(unsigned long *addr, int numpages)
1960 /* WB cache mode is hard wired to all cache attribute bits being 0 */
1961 ret = change_page_attr_clear(addr, numpages,
1962 __pgprot(_PAGE_CACHE_MASK), 1);
1966 for (i = 0; i < numpages; i++)
1967 free_memtype(__pa(addr[i]), __pa(addr[i]) + PAGE_SIZE);
1971 EXPORT_SYMBOL(set_memory_array_wb);
1973 int set_memory_x(unsigned long addr, int numpages)
1975 if (!(__supported_pte_mask & _PAGE_NX))
1978 return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1980 EXPORT_SYMBOL(set_memory_x);
1982 int set_memory_nx(unsigned long addr, int numpages)
1984 if (!(__supported_pte_mask & _PAGE_NX))
1987 return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1989 EXPORT_SYMBOL(set_memory_nx);
1991 int set_memory_ro(unsigned long addr, int numpages)
1993 return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1996 int set_memory_rw(unsigned long addr, int numpages)
1998 return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
2001 int set_memory_np(unsigned long addr, int numpages)
2003 return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
2006 int set_memory_np_noalias(unsigned long addr, int numpages)
2008 int cpa_flags = CPA_NO_CHECK_ALIAS;
2010 return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
2011 __pgprot(_PAGE_PRESENT), 0,
2015 int set_memory_4k(unsigned long addr, int numpages)
2017 return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
2018 __pgprot(0), 1, 0, NULL);
2021 int set_memory_nonglobal(unsigned long addr, int numpages)
2023 return change_page_attr_clear(&addr, numpages,
2024 __pgprot(_PAGE_GLOBAL), 0);
2027 int set_memory_global(unsigned long addr, int numpages)
2029 return change_page_attr_set(&addr, numpages,
2030 __pgprot(_PAGE_GLOBAL), 0);
2033 static int __set_memory_enc_dec(unsigned long addr, int numpages, bool enc)
2035 struct cpa_data cpa;
2038 /* Nothing to do if memory encryption is not active */
2039 if (!mem_encrypt_active())
2042 /* Should not be working on unaligned addresses */
2043 if (WARN_ONCE(addr & ~PAGE_MASK, "misaligned address: %#lx\n", addr))
2046 memset(&cpa, 0, sizeof(cpa));
2048 cpa.numpages = numpages;
2049 cpa.mask_set = enc ? __pgprot(_PAGE_ENC) : __pgprot(0);
2050 cpa.mask_clr = enc ? __pgprot(0) : __pgprot(_PAGE_ENC);
2051 cpa.pgd = init_mm.pgd;
2053 /* Must avoid aliasing mappings in the highmem code */
2054 kmap_flush_unused();
2058 * Before changing the encryption attribute, we need to flush caches.
2062 ret = __change_page_attr_set_clr(&cpa, 1);
2065 * After changing the encryption attribute, we need to flush TLBs again
2066 * in case any speculative TLB caching occurred (but no need to flush
2067 * caches again). We could just use cpa_flush_all(), but in case TLB
2068 * flushing gets optimized in the cpa_flush() path use the same logic
2076 int set_memory_encrypted(unsigned long addr, int numpages)
2078 return __set_memory_enc_dec(addr, numpages, true);
2080 EXPORT_SYMBOL_GPL(set_memory_encrypted);
2082 int set_memory_decrypted(unsigned long addr, int numpages)
2084 return __set_memory_enc_dec(addr, numpages, false);
2086 EXPORT_SYMBOL_GPL(set_memory_decrypted);
2088 int set_pages_uc(struct page *page, int numpages)
2090 unsigned long addr = (unsigned long)page_address(page);
2092 return set_memory_uc(addr, numpages);
2094 EXPORT_SYMBOL(set_pages_uc);
2096 static int _set_pages_array(struct page **pages, int numpages,
2097 enum page_cache_mode new_type)
2099 unsigned long start;
2101 enum page_cache_mode set_type;
2106 for (i = 0; i < numpages; i++) {
2107 if (PageHighMem(pages[i]))
2109 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
2110 end = start + PAGE_SIZE;
2111 if (reserve_memtype(start, end, new_type, NULL))
2115 /* If WC, set to UC- first and then WC */
2116 set_type = (new_type == _PAGE_CACHE_MODE_WC) ?
2117 _PAGE_CACHE_MODE_UC_MINUS : new_type;
2119 ret = cpa_set_pages_array(pages, numpages,
2120 cachemode2pgprot(set_type));
2121 if (!ret && new_type == _PAGE_CACHE_MODE_WC)
2122 ret = change_page_attr_set_clr(NULL, numpages,
2124 _PAGE_CACHE_MODE_WC),
2125 __pgprot(_PAGE_CACHE_MASK),
2126 0, CPA_PAGES_ARRAY, pages);
2129 return 0; /* Success */
2132 for (i = 0; i < free_idx; i++) {
2133 if (PageHighMem(pages[i]))
2135 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
2136 end = start + PAGE_SIZE;
2137 free_memtype(start, end);
2142 int set_pages_array_uc(struct page **pages, int numpages)
2144 return _set_pages_array(pages, numpages, _PAGE_CACHE_MODE_UC_MINUS);
2146 EXPORT_SYMBOL(set_pages_array_uc);
2148 int set_pages_array_wc(struct page **pages, int numpages)
2150 return _set_pages_array(pages, numpages, _PAGE_CACHE_MODE_WC);
2152 EXPORT_SYMBOL(set_pages_array_wc);
2154 int set_pages_array_wt(struct page **pages, int numpages)
2156 return _set_pages_array(pages, numpages, _PAGE_CACHE_MODE_WT);
2158 EXPORT_SYMBOL_GPL(set_pages_array_wt);
2160 int set_pages_wb(struct page *page, int numpages)
2162 unsigned long addr = (unsigned long)page_address(page);
2164 return set_memory_wb(addr, numpages);
2166 EXPORT_SYMBOL(set_pages_wb);
2168 int set_pages_array_wb(struct page **pages, int numpages)
2171 unsigned long start;
2175 /* WB cache mode is hard wired to all cache attribute bits being 0 */
2176 retval = cpa_clear_pages_array(pages, numpages,
2177 __pgprot(_PAGE_CACHE_MASK));
2181 for (i = 0; i < numpages; i++) {
2182 if (PageHighMem(pages[i]))
2184 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
2185 end = start + PAGE_SIZE;
2186 free_memtype(start, end);
2191 EXPORT_SYMBOL(set_pages_array_wb);
2193 int set_pages_x(struct page *page, int numpages)
2195 unsigned long addr = (unsigned long)page_address(page);
2197 return set_memory_x(addr, numpages);
2199 EXPORT_SYMBOL(set_pages_x);
2201 int set_pages_nx(struct page *page, int numpages)
2203 unsigned long addr = (unsigned long)page_address(page);
2205 return set_memory_nx(addr, numpages);
2207 EXPORT_SYMBOL(set_pages_nx);
2209 int set_pages_ro(struct page *page, int numpages)
2211 unsigned long addr = (unsigned long)page_address(page);
2213 return set_memory_ro(addr, numpages);
2216 int set_pages_rw(struct page *page, int numpages)
2218 unsigned long addr = (unsigned long)page_address(page);
2220 return set_memory_rw(addr, numpages);
2223 static int __set_pages_p(struct page *page, int numpages)
2225 unsigned long tempaddr = (unsigned long) page_address(page);
2226 struct cpa_data cpa = { .vaddr = &tempaddr,
2228 .numpages = numpages,
2229 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
2230 .mask_clr = __pgprot(0),
2234 * No alias checking needed for setting present flag. otherwise,
2235 * we may need to break large pages for 64-bit kernel text
2236 * mappings (this adds to complexity if we want to do this from
2237 * atomic context especially). Let's keep it simple!
2239 return __change_page_attr_set_clr(&cpa, 0);
2242 static int __set_pages_np(struct page *page, int numpages)
2244 unsigned long tempaddr = (unsigned long) page_address(page);
2245 struct cpa_data cpa = { .vaddr = &tempaddr,
2247 .numpages = numpages,
2248 .mask_set = __pgprot(0),
2249 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
2253 * No alias checking needed for setting not present flag. otherwise,
2254 * we may need to break large pages for 64-bit kernel text
2255 * mappings (this adds to complexity if we want to do this from
2256 * atomic context especially). Let's keep it simple!
2258 return __change_page_attr_set_clr(&cpa, 0);
2261 int set_direct_map_invalid_noflush(struct page *page)
2263 return __set_pages_np(page, 1);
2266 int set_direct_map_default_noflush(struct page *page)
2268 return __set_pages_p(page, 1);
2271 void __kernel_map_pages(struct page *page, int numpages, int enable)
2273 if (PageHighMem(page))
2276 debug_check_no_locks_freed(page_address(page),
2277 numpages * PAGE_SIZE);
2281 * The return value is ignored as the calls cannot fail.
2282 * Large pages for identity mappings are not used at boot time
2283 * and hence no memory allocations during large page split.
2286 __set_pages_p(page, numpages);
2288 __set_pages_np(page, numpages);
2291 * We should perform an IPI and flush all tlbs,
2292 * but that can deadlock->flush only current cpu.
2293 * Preemption needs to be disabled around __flush_tlb_all() due to
2294 * CR3 reload in __native_flush_tlb().
2300 arch_flush_lazy_mmu_mode();
2303 #ifdef CONFIG_HIBERNATION
2304 bool kernel_page_present(struct page *page)
2309 if (PageHighMem(page))
2312 pte = lookup_address((unsigned long)page_address(page), &level);
2313 return (pte_val(*pte) & _PAGE_PRESENT);
2315 #endif /* CONFIG_HIBERNATION */
2317 int __init kernel_map_pages_in_pgd(pgd_t *pgd, u64 pfn, unsigned long address,
2318 unsigned numpages, unsigned long page_flags)
2320 int retval = -EINVAL;
2322 struct cpa_data cpa = {
2326 .numpages = numpages,
2327 .mask_set = __pgprot(0),
2328 .mask_clr = __pgprot(0),
2332 WARN_ONCE(num_online_cpus() > 1, "Don't call after initializing SMP");
2334 if (!(__supported_pte_mask & _PAGE_NX))
2337 if (!(page_flags & _PAGE_NX))
2338 cpa.mask_clr = __pgprot(_PAGE_NX);
2340 if (!(page_flags & _PAGE_RW))
2341 cpa.mask_clr = __pgprot(_PAGE_RW);
2343 if (!(page_flags & _PAGE_ENC))
2344 cpa.mask_clr = pgprot_encrypted(cpa.mask_clr);
2346 cpa.mask_set = __pgprot(_PAGE_PRESENT | page_flags);
2348 retval = __change_page_attr_set_clr(&cpa, 0);
2356 * __flush_tlb_all() flushes mappings only on current CPU and hence this
2357 * function shouldn't be used in an SMP environment. Presently, it's used only
2358 * during boot (way before smp_init()) by EFI subsystem and hence is ok.
2360 int __init kernel_unmap_pages_in_pgd(pgd_t *pgd, unsigned long address,
2361 unsigned long numpages)
2366 * The typical sequence for unmapping is to find a pte through
2367 * lookup_address_in_pgd() (ideally, it should never return NULL because
2368 * the address is already mapped) and change it's protections. As pfn is
2369 * the *target* of a mapping, it's not useful while unmapping.
2371 struct cpa_data cpa = {
2375 .numpages = numpages,
2376 .mask_set = __pgprot(0),
2377 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
2381 WARN_ONCE(num_online_cpus() > 1, "Don't call after initializing SMP");
2383 retval = __change_page_attr_set_clr(&cpa, 0);
2390 * The testcases use internal knowledge of the implementation that shouldn't
2391 * be exposed to the rest of the kernel. Include these directly here.
2393 #ifdef CONFIG_CPA_DEBUG
2394 #include "pageattr-test.c"