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,
56 static const int cpa_warn_level = CPA_PROTECT;
59 * Serialize cpa() (for !DEBUG_PAGEALLOC which uses large identity mappings)
60 * using cpa_lock. So that we don't allow any other cpu, with stale large tlb
61 * entries change the page attribute in parallel to some other cpu
62 * splitting a large page entry along with changing the attribute.
64 static DEFINE_SPINLOCK(cpa_lock);
66 #define CPA_FLUSHTLB 1
68 #define CPA_PAGES_ARRAY 4
69 #define CPA_NO_CHECK_ALIAS 8 /* Do not search for aliases */
72 static unsigned long direct_pages_count[PG_LEVEL_NUM];
74 void update_page_count(int level, unsigned long pages)
76 /* Protect against CPA */
78 direct_pages_count[level] += pages;
79 spin_unlock(&pgd_lock);
82 static void split_page_count(int level)
84 if (direct_pages_count[level] == 0)
87 direct_pages_count[level]--;
88 direct_pages_count[level - 1] += PTRS_PER_PTE;
91 void arch_report_meminfo(struct seq_file *m)
93 seq_printf(m, "DirectMap4k: %8lu kB\n",
94 direct_pages_count[PG_LEVEL_4K] << 2);
95 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
96 seq_printf(m, "DirectMap2M: %8lu kB\n",
97 direct_pages_count[PG_LEVEL_2M] << 11);
99 seq_printf(m, "DirectMap4M: %8lu kB\n",
100 direct_pages_count[PG_LEVEL_2M] << 12);
103 seq_printf(m, "DirectMap1G: %8lu kB\n",
104 direct_pages_count[PG_LEVEL_1G] << 20);
107 static inline void split_page_count(int level) { }
110 #ifdef CONFIG_X86_CPA_STATISTICS
112 static unsigned long cpa_1g_checked;
113 static unsigned long cpa_1g_sameprot;
114 static unsigned long cpa_1g_preserved;
115 static unsigned long cpa_2m_checked;
116 static unsigned long cpa_2m_sameprot;
117 static unsigned long cpa_2m_preserved;
118 static unsigned long cpa_4k_install;
120 static inline void cpa_inc_1g_checked(void)
125 static inline void cpa_inc_2m_checked(void)
130 static inline void cpa_inc_4k_install(void)
135 static inline void cpa_inc_lp_sameprot(int level)
137 if (level == PG_LEVEL_1G)
143 static inline void cpa_inc_lp_preserved(int level)
145 if (level == PG_LEVEL_1G)
151 static int cpastats_show(struct seq_file *m, void *p)
153 seq_printf(m, "1G pages checked: %16lu\n", cpa_1g_checked);
154 seq_printf(m, "1G pages sameprot: %16lu\n", cpa_1g_sameprot);
155 seq_printf(m, "1G pages preserved: %16lu\n", cpa_1g_preserved);
156 seq_printf(m, "2M pages checked: %16lu\n", cpa_2m_checked);
157 seq_printf(m, "2M pages sameprot: %16lu\n", cpa_2m_sameprot);
158 seq_printf(m, "2M pages preserved: %16lu\n", cpa_2m_preserved);
159 seq_printf(m, "4K pages set-checked: %16lu\n", cpa_4k_install);
163 static int cpastats_open(struct inode *inode, struct file *file)
165 return single_open(file, cpastats_show, NULL);
168 static const struct file_operations cpastats_fops = {
169 .open = cpastats_open,
172 .release = single_release,
175 static int __init cpa_stats_init(void)
177 debugfs_create_file("cpa_stats", S_IRUSR, arch_debugfs_dir, NULL,
181 late_initcall(cpa_stats_init);
183 static inline void cpa_inc_1g_checked(void) { }
184 static inline void cpa_inc_2m_checked(void) { }
185 static inline void cpa_inc_4k_install(void) { }
186 static inline void cpa_inc_lp_sameprot(int level) { }
187 static inline void cpa_inc_lp_preserved(int level) { }
192 within(unsigned long addr, unsigned long start, unsigned long end)
194 return addr >= start && addr < end;
198 within_inclusive(unsigned long addr, unsigned long start, unsigned long end)
200 return addr >= start && addr <= end;
205 static inline unsigned long highmap_start_pfn(void)
207 return __pa_symbol(_text) >> PAGE_SHIFT;
210 static inline unsigned long highmap_end_pfn(void)
212 /* Do not reference physical address outside the kernel. */
213 return __pa_symbol(roundup(_brk_end, PMD_SIZE) - 1) >> PAGE_SHIFT;
216 static bool __cpa_pfn_in_highmap(unsigned long pfn)
219 * Kernel text has an alias mapping at a high address, known
222 return within_inclusive(pfn, highmap_start_pfn(), highmap_end_pfn());
227 static bool __cpa_pfn_in_highmap(unsigned long pfn)
229 /* There is no highmap on 32-bit */
236 * See set_mce_nospec().
238 * Machine check recovery code needs to change cache mode of poisoned pages to
239 * UC to avoid speculative access logging another error. But passing the
240 * address of the 1:1 mapping to set_memory_uc() is a fine way to encourage a
241 * speculative access. So we cheat and flip the top bit of the address. This
242 * works fine for the code that updates the page tables. But at the end of the
243 * process we need to flush the TLB and cache and the non-canonical address
244 * causes a #GP fault when used by the INVLPG and CLFLUSH instructions.
246 * But in the common case we already have a canonical address. This code
247 * will fix the top bit if needed and is a no-op otherwise.
249 static inline unsigned long fix_addr(unsigned long addr)
252 return (long)(addr << 1) >> 1;
258 static unsigned long __cpa_addr(struct cpa_data *cpa, unsigned long idx)
260 if (cpa->flags & CPA_PAGES_ARRAY) {
261 struct page *page = cpa->pages[idx];
263 if (unlikely(PageHighMem(page)))
266 return (unsigned long)page_address(page);
269 if (cpa->flags & CPA_ARRAY)
270 return cpa->vaddr[idx];
272 return *cpa->vaddr + idx * PAGE_SIZE;
279 static void clflush_cache_range_opt(void *vaddr, unsigned int size)
281 const unsigned long clflush_size = boot_cpu_data.x86_clflush_size;
282 void *p = (void *)((unsigned long)vaddr & ~(clflush_size - 1));
283 void *vend = vaddr + size;
288 for (; p < vend; p += clflush_size)
293 * clflush_cache_range - flush a cache range with clflush
294 * @vaddr: virtual start address
295 * @size: number of bytes to flush
297 * CLFLUSHOPT is an unordered instruction which needs fencing with MFENCE or
298 * SFENCE to avoid ordering issues.
300 void clflush_cache_range(void *vaddr, unsigned int size)
303 clflush_cache_range_opt(vaddr, size);
306 EXPORT_SYMBOL_GPL(clflush_cache_range);
308 void arch_invalidate_pmem(void *addr, size_t size)
310 clflush_cache_range(addr, size);
312 EXPORT_SYMBOL_GPL(arch_invalidate_pmem);
314 static void __cpa_flush_all(void *arg)
316 unsigned long cache = (unsigned long)arg;
319 * Flush all to work around Errata in early athlons regarding
320 * large page flushing.
324 if (cache && boot_cpu_data.x86 >= 4)
328 static void cpa_flush_all(unsigned long cache)
330 BUG_ON(irqs_disabled() && !early_boot_irqs_disabled);
332 on_each_cpu(__cpa_flush_all, (void *) cache, 1);
335 void __cpa_flush_tlb(void *data)
337 struct cpa_data *cpa = data;
340 for (i = 0; i < cpa->numpages; i++)
341 __flush_tlb_one_kernel(fix_addr(__cpa_addr(cpa, i)));
344 static void cpa_flush(struct cpa_data *data, int cache)
346 struct cpa_data *cpa = data;
349 BUG_ON(irqs_disabled() && !early_boot_irqs_disabled);
351 if (cache && !static_cpu_has(X86_FEATURE_CLFLUSH)) {
352 cpa_flush_all(cache);
356 if (cpa->force_flush_all || cpa->numpages > tlb_single_page_flush_ceiling)
359 on_each_cpu(__cpa_flush_tlb, cpa, 1);
365 for (i = 0; i < cpa->numpages; i++) {
366 unsigned long addr = __cpa_addr(cpa, i);
369 pte_t *pte = lookup_address(addr, &level);
372 * Only flush present addresses:
374 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
375 clflush_cache_range_opt((void *)fix_addr(addr), PAGE_SIZE);
380 static bool overlaps(unsigned long r1_start, unsigned long r1_end,
381 unsigned long r2_start, unsigned long r2_end)
383 return (r1_start <= r2_end && r1_end >= r2_start) ||
384 (r2_start <= r1_end && r2_end >= r1_start);
387 #ifdef CONFIG_PCI_BIOS
389 * The BIOS area between 640k and 1Mb needs to be executable for PCI BIOS
390 * based config access (CONFIG_PCI_GOBIOS) support.
392 #define BIOS_PFN PFN_DOWN(BIOS_BEGIN)
393 #define BIOS_PFN_END PFN_DOWN(BIOS_END - 1)
395 static pgprotval_t protect_pci_bios(unsigned long spfn, unsigned long epfn)
397 if (pcibios_enabled && overlaps(spfn, epfn, BIOS_PFN, BIOS_PFN_END))
402 static pgprotval_t protect_pci_bios(unsigned long spfn, unsigned long epfn)
409 * The .rodata section needs to be read-only. Using the pfn catches all
410 * aliases. This also includes __ro_after_init, so do not enforce until
411 * kernel_set_to_readonly is true.
413 static pgprotval_t protect_rodata(unsigned long spfn, unsigned long epfn)
415 unsigned long epfn_ro, spfn_ro = PFN_DOWN(__pa_symbol(__start_rodata));
418 * Note: __end_rodata is at page aligned and not inclusive, so
419 * subtract 1 to get the last enforced PFN in the rodata area.
421 epfn_ro = PFN_DOWN(__pa_symbol(__end_rodata)) - 1;
423 if (kernel_set_to_readonly && overlaps(spfn, epfn, spfn_ro, epfn_ro))
429 * Protect kernel text against becoming non executable by forbidding
430 * _PAGE_NX. This protects only the high kernel mapping (_text -> _etext)
431 * out of which the kernel actually executes. Do not protect the low
434 * This does not cover __inittext since that is gone after boot.
436 static pgprotval_t protect_kernel_text(unsigned long start, unsigned long end)
438 unsigned long t_end = (unsigned long)_etext - 1;
439 unsigned long t_start = (unsigned long)_text;
441 if (overlaps(start, end, t_start, t_end))
446 #if defined(CONFIG_X86_64)
448 * Once the kernel maps the text as RO (kernel_set_to_readonly is set),
449 * kernel text mappings for the large page aligned text, rodata sections
450 * will be always read-only. For the kernel identity mappings covering the
451 * holes caused by this alignment can be anything that user asks.
453 * This will preserve the large page mappings for kernel text/data at no
456 static pgprotval_t protect_kernel_text_ro(unsigned long start,
459 unsigned long t_end = (unsigned long)__end_rodata_hpage_align - 1;
460 unsigned long t_start = (unsigned long)_text;
463 if (!kernel_set_to_readonly || !overlaps(start, end, t_start, t_end))
466 * Don't enforce the !RW mapping for the kernel text mapping, if
467 * the current mapping is already using small page mapping. No
468 * need to work hard to preserve large page mappings in this case.
470 * This also fixes the Linux Xen paravirt guest boot failure caused
471 * by unexpected read-only mappings for kernel identity
472 * mappings. In this paravirt guest case, the kernel text mapping
473 * and the kernel identity mapping share the same page-table pages,
474 * so the protections for kernel text and identity mappings have to
477 if (lookup_address(start, &level) && (level != PG_LEVEL_4K))
482 static pgprotval_t protect_kernel_text_ro(unsigned long start,
489 static inline bool conflicts(pgprot_t prot, pgprotval_t val)
491 return (pgprot_val(prot) & ~val) != pgprot_val(prot);
494 static inline void check_conflict(int warnlvl, pgprot_t prot, pgprotval_t val,
495 unsigned long start, unsigned long end,
496 unsigned long pfn, const char *txt)
498 static const char *lvltxt[] = {
499 [CPA_CONFLICT] = "conflict",
500 [CPA_PROTECT] = "protect",
501 [CPA_DETECT] = "detect",
504 if (warnlvl > cpa_warn_level || !conflicts(prot, val))
507 pr_warn("CPA %8s %10s: 0x%016lx - 0x%016lx PFN %lx req %016llx prevent %016llx\n",
508 lvltxt[warnlvl], txt, start, end, pfn, (unsigned long long)pgprot_val(prot),
509 (unsigned long long)val);
513 * Certain areas of memory on x86 require very specific protection flags,
514 * for example the BIOS area or kernel text. Callers don't always get this
515 * right (again, ioremap() on BIOS memory is not uncommon) so this function
516 * checks and fixes these known static required protection bits.
518 static inline pgprot_t static_protections(pgprot_t prot, unsigned long start,
519 unsigned long pfn, unsigned long npg,
520 unsigned long lpsize, int warnlvl)
522 pgprotval_t forbidden, res;
526 * There is no point in checking RW/NX conflicts when the requested
527 * mapping is setting the page !PRESENT.
529 if (!(pgprot_val(prot) & _PAGE_PRESENT))
532 /* Operate on the virtual address */
533 end = start + npg * PAGE_SIZE - 1;
535 res = protect_kernel_text(start, end);
536 check_conflict(warnlvl, prot, res, start, end, pfn, "Text NX");
540 * Special case to preserve a large page. If the change spawns the
541 * full large page mapping then there is no point to split it
542 * up. Happens with ftrace and is going to be removed once ftrace
543 * switched to text_poke().
545 if (lpsize != (npg * PAGE_SIZE) || (start & (lpsize - 1))) {
546 res = protect_kernel_text_ro(start, end);
547 check_conflict(warnlvl, prot, res, start, end, pfn, "Text RO");
551 /* Check the PFN directly */
552 res = protect_pci_bios(pfn, pfn + npg - 1);
553 check_conflict(warnlvl, prot, res, start, end, pfn, "PCIBIOS NX");
556 res = protect_rodata(pfn, pfn + npg - 1);
557 check_conflict(warnlvl, prot, res, start, end, pfn, "Rodata RO");
560 return __pgprot(pgprot_val(prot) & ~forbidden);
564 * Lookup the page table entry for a virtual address in a specific pgd.
565 * Return a pointer to the entry and the level of the mapping.
567 pte_t *lookup_address_in_pgd(pgd_t *pgd, unsigned long address,
574 *level = PG_LEVEL_NONE;
579 p4d = p4d_offset(pgd, address);
583 *level = PG_LEVEL_512G;
584 if (p4d_large(*p4d) || !p4d_present(*p4d))
587 pud = pud_offset(p4d, address);
591 *level = PG_LEVEL_1G;
592 if (pud_large(*pud) || !pud_present(*pud))
595 pmd = pmd_offset(pud, address);
599 *level = PG_LEVEL_2M;
600 if (pmd_large(*pmd) || !pmd_present(*pmd))
603 *level = PG_LEVEL_4K;
605 return pte_offset_kernel(pmd, address);
609 * Lookup the page table entry for a virtual address. Return a pointer
610 * to the entry and the level of the mapping.
612 * Note: We return pud and pmd either when the entry is marked large
613 * or when the present bit is not set. Otherwise we would return a
614 * pointer to a nonexisting mapping.
616 pte_t *lookup_address(unsigned long address, unsigned int *level)
618 return lookup_address_in_pgd(pgd_offset_k(address), address, level);
620 EXPORT_SYMBOL_GPL(lookup_address);
622 static pte_t *_lookup_address_cpa(struct cpa_data *cpa, unsigned long address,
626 return lookup_address_in_pgd(cpa->pgd + pgd_index(address),
629 return lookup_address(address, level);
633 * Lookup the PMD entry for a virtual address. Return a pointer to the entry
634 * or NULL if not present.
636 pmd_t *lookup_pmd_address(unsigned long address)
642 pgd = pgd_offset_k(address);
646 p4d = p4d_offset(pgd, address);
647 if (p4d_none(*p4d) || p4d_large(*p4d) || !p4d_present(*p4d))
650 pud = pud_offset(p4d, address);
651 if (pud_none(*pud) || pud_large(*pud) || !pud_present(*pud))
654 return pmd_offset(pud, address);
658 * This is necessary because __pa() does not work on some
659 * kinds of memory, like vmalloc() or the alloc_remap()
660 * areas on 32-bit NUMA systems. The percpu areas can
661 * end up in this kind of memory, for instance.
663 * This could be optimized, but it is only intended to be
664 * used at inititalization time, and keeping it
665 * unoptimized should increase the testing coverage for
666 * the more obscure platforms.
668 phys_addr_t slow_virt_to_phys(void *__virt_addr)
670 unsigned long virt_addr = (unsigned long)__virt_addr;
671 phys_addr_t phys_addr;
672 unsigned long offset;
676 pte = lookup_address(virt_addr, &level);
680 * pXX_pfn() returns unsigned long, which must be cast to phys_addr_t
681 * before being left-shifted PAGE_SHIFT bits -- this trick is to
682 * make 32-PAE kernel work correctly.
686 phys_addr = (phys_addr_t)pud_pfn(*(pud_t *)pte) << PAGE_SHIFT;
687 offset = virt_addr & ~PUD_PAGE_MASK;
690 phys_addr = (phys_addr_t)pmd_pfn(*(pmd_t *)pte) << PAGE_SHIFT;
691 offset = virt_addr & ~PMD_PAGE_MASK;
694 phys_addr = (phys_addr_t)pte_pfn(*pte) << PAGE_SHIFT;
695 offset = virt_addr & ~PAGE_MASK;
698 return (phys_addr_t)(phys_addr | offset);
700 EXPORT_SYMBOL_GPL(slow_virt_to_phys);
703 * Set the new pmd in all the pgds we know about:
705 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
708 set_pte_atomic(kpte, pte);
710 if (!SHARED_KERNEL_PMD) {
713 list_for_each_entry(page, &pgd_list, lru) {
719 pgd = (pgd_t *)page_address(page) + pgd_index(address);
720 p4d = p4d_offset(pgd, address);
721 pud = pud_offset(p4d, address);
722 pmd = pmd_offset(pud, address);
723 set_pte_atomic((pte_t *)pmd, pte);
729 static pgprot_t pgprot_clear_protnone_bits(pgprot_t prot)
732 * _PAGE_GLOBAL means "global page" for present PTEs.
733 * But, it is also used to indicate _PAGE_PROTNONE
734 * for non-present PTEs.
736 * This ensures that a _PAGE_GLOBAL PTE going from
737 * present to non-present is not confused as
740 if (!(pgprot_val(prot) & _PAGE_PRESENT))
741 pgprot_val(prot) &= ~_PAGE_GLOBAL;
746 static int __should_split_large_page(pte_t *kpte, unsigned long address,
747 struct cpa_data *cpa)
749 unsigned long numpages, pmask, psize, lpaddr, pfn, old_pfn;
750 pgprot_t old_prot, new_prot, req_prot, chk_prot;
755 * Check for races, another CPU might have split this page
758 tmp = _lookup_address_cpa(cpa, address, &level);
764 old_prot = pmd_pgprot(*(pmd_t *)kpte);
765 old_pfn = pmd_pfn(*(pmd_t *)kpte);
766 cpa_inc_2m_checked();
769 old_prot = pud_pgprot(*(pud_t *)kpte);
770 old_pfn = pud_pfn(*(pud_t *)kpte);
771 cpa_inc_1g_checked();
777 psize = page_level_size(level);
778 pmask = page_level_mask(level);
781 * Calculate the number of pages, which fit into this large
782 * page starting at address:
784 lpaddr = (address + psize) & pmask;
785 numpages = (lpaddr - address) >> PAGE_SHIFT;
786 if (numpages < cpa->numpages)
787 cpa->numpages = numpages;
790 * We are safe now. Check whether the new pgprot is the same:
791 * Convert protection attributes to 4k-format, as cpa->mask* are set
795 /* Clear PSE (aka _PAGE_PAT) and move PAT bit to correct position */
796 req_prot = pgprot_large_2_4k(old_prot);
798 pgprot_val(req_prot) &= ~pgprot_val(cpa->mask_clr);
799 pgprot_val(req_prot) |= pgprot_val(cpa->mask_set);
802 * req_prot is in format of 4k pages. It must be converted to large
803 * page format: the caching mode includes the PAT bit located at
804 * different bit positions in the two formats.
806 req_prot = pgprot_4k_2_large(req_prot);
807 req_prot = pgprot_clear_protnone_bits(req_prot);
808 if (pgprot_val(req_prot) & _PAGE_PRESENT)
809 pgprot_val(req_prot) |= _PAGE_PSE;
812 * old_pfn points to the large page base pfn. So we need to add the
813 * offset of the virtual address:
815 pfn = old_pfn + ((address & (psize - 1)) >> PAGE_SHIFT);
819 * Calculate the large page base address and the number of 4K pages
822 lpaddr = address & pmask;
823 numpages = psize >> PAGE_SHIFT;
826 * Sanity check that the existing mapping is correct versus the static
827 * protections. static_protections() guards against !PRESENT, so no
828 * extra conditional required here.
830 chk_prot = static_protections(old_prot, lpaddr, old_pfn, numpages,
831 psize, CPA_CONFLICT);
833 if (WARN_ON_ONCE(pgprot_val(chk_prot) != pgprot_val(old_prot))) {
835 * Split the large page and tell the split code to
836 * enforce static protections.
838 cpa->force_static_prot = 1;
843 * Optimization: If the requested pgprot is the same as the current
844 * pgprot, then the large page can be preserved and no updates are
845 * required independent of alignment and length of the requested
846 * range. The above already established that the current pgprot is
847 * correct, which in consequence makes the requested pgprot correct
848 * as well if it is the same. The static protection scan below will
849 * not come to a different conclusion.
851 if (pgprot_val(req_prot) == pgprot_val(old_prot)) {
852 cpa_inc_lp_sameprot(level);
857 * If the requested range does not cover the full page, split it up
859 if (address != lpaddr || cpa->numpages != numpages)
863 * Check whether the requested pgprot is conflicting with a static
864 * protection requirement in the large page.
866 new_prot = static_protections(req_prot, lpaddr, old_pfn, numpages,
870 * If there is a conflict, split the large page.
872 * There used to be a 4k wise evaluation trying really hard to
873 * preserve the large pages, but experimentation has shown, that this
874 * does not help at all. There might be corner cases which would
875 * preserve one large page occasionally, but it's really not worth the
876 * extra code and cycles for the common case.
878 if (pgprot_val(req_prot) != pgprot_val(new_prot))
881 /* All checks passed. Update the large page mapping. */
882 new_pte = pfn_pte(old_pfn, new_prot);
883 __set_pmd_pte(kpte, address, new_pte);
884 cpa->flags |= CPA_FLUSHTLB;
885 cpa_inc_lp_preserved(level);
889 static int should_split_large_page(pte_t *kpte, unsigned long address,
890 struct cpa_data *cpa)
894 if (cpa->force_split)
897 spin_lock(&pgd_lock);
898 do_split = __should_split_large_page(kpte, address, cpa);
899 spin_unlock(&pgd_lock);
904 static void split_set_pte(struct cpa_data *cpa, pte_t *pte, unsigned long pfn,
905 pgprot_t ref_prot, unsigned long address,
908 unsigned int npg = PFN_DOWN(size);
912 * If should_split_large_page() discovered an inconsistent mapping,
913 * remove the invalid protection in the split mapping.
915 if (!cpa->force_static_prot)
918 /* Hand in lpsize = 0 to enforce the protection mechanism */
919 prot = static_protections(ref_prot, address, pfn, npg, 0, CPA_PROTECT);
921 if (pgprot_val(prot) == pgprot_val(ref_prot))
925 * If this is splitting a PMD, fix it up. PUD splits cannot be
926 * fixed trivially as that would require to rescan the newly
927 * installed PMD mappings after returning from split_large_page()
928 * so an eventual further split can allocate the necessary PTE
929 * pages. Warn for now and revisit it in case this actually
932 if (size == PAGE_SIZE)
935 pr_warn_once("CPA: Cannot fixup static protections for PUD split\n");
937 set_pte(pte, pfn_pte(pfn, ref_prot));
941 __split_large_page(struct cpa_data *cpa, pte_t *kpte, unsigned long address,
944 unsigned long lpaddr, lpinc, ref_pfn, pfn, pfninc = 1;
945 pte_t *pbase = (pte_t *)page_address(base);
946 unsigned int i, level;
950 spin_lock(&pgd_lock);
952 * Check for races, another CPU might have split this page
955 tmp = _lookup_address_cpa(cpa, address, &level);
957 spin_unlock(&pgd_lock);
961 paravirt_alloc_pte(&init_mm, page_to_pfn(base));
965 ref_prot = pmd_pgprot(*(pmd_t *)kpte);
967 * Clear PSE (aka _PAGE_PAT) and move
968 * PAT bit to correct position.
970 ref_prot = pgprot_large_2_4k(ref_prot);
971 ref_pfn = pmd_pfn(*(pmd_t *)kpte);
972 lpaddr = address & PMD_MASK;
977 ref_prot = pud_pgprot(*(pud_t *)kpte);
978 ref_pfn = pud_pfn(*(pud_t *)kpte);
979 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
980 lpaddr = address & PUD_MASK;
983 * Clear the PSE flags if the PRESENT flag is not set
984 * otherwise pmd_present/pmd_huge will return true
985 * even on a non present pmd.
987 if (!(pgprot_val(ref_prot) & _PAGE_PRESENT))
988 pgprot_val(ref_prot) &= ~_PAGE_PSE;
992 spin_unlock(&pgd_lock);
996 ref_prot = pgprot_clear_protnone_bits(ref_prot);
999 * Get the target pfn from the original entry:
1002 for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc, lpaddr += lpinc)
1003 split_set_pte(cpa, pbase + i, pfn, ref_prot, lpaddr, lpinc);
1005 if (virt_addr_valid(address)) {
1006 unsigned long pfn = PFN_DOWN(__pa(address));
1008 if (pfn_range_is_mapped(pfn, pfn + 1))
1009 split_page_count(level);
1013 * Install the new, split up pagetable.
1015 * We use the standard kernel pagetable protections for the new
1016 * pagetable protections, the actual ptes set above control the
1017 * primary protection behavior:
1019 __set_pmd_pte(kpte, address, mk_pte(base, __pgprot(_KERNPG_TABLE)));
1022 * Do a global flush tlb after splitting the large page
1023 * and before we do the actual change page attribute in the PTE.
1025 * Without this, we violate the TLB application note, that says:
1026 * "The TLBs may contain both ordinary and large-page
1027 * translations for a 4-KByte range of linear addresses. This
1028 * may occur if software modifies the paging structures so that
1029 * the page size used for the address range changes. If the two
1030 * translations differ with respect to page frame or attributes
1031 * (e.g., permissions), processor behavior is undefined and may
1032 * be implementation-specific."
1034 * We do this global tlb flush inside the cpa_lock, so that we
1035 * don't allow any other cpu, with stale tlb entries change the
1036 * page attribute in parallel, that also falls into the
1037 * just split large page entry.
1040 spin_unlock(&pgd_lock);
1045 static int split_large_page(struct cpa_data *cpa, pte_t *kpte,
1046 unsigned long address)
1050 if (!debug_pagealloc_enabled())
1051 spin_unlock(&cpa_lock);
1052 base = alloc_pages(GFP_KERNEL, 0);
1053 if (!debug_pagealloc_enabled())
1054 spin_lock(&cpa_lock);
1058 if (__split_large_page(cpa, kpte, address, base))
1064 static bool try_to_free_pte_page(pte_t *pte)
1068 for (i = 0; i < PTRS_PER_PTE; i++)
1069 if (!pte_none(pte[i]))
1072 free_page((unsigned long)pte);
1076 static bool try_to_free_pmd_page(pmd_t *pmd)
1080 for (i = 0; i < PTRS_PER_PMD; i++)
1081 if (!pmd_none(pmd[i]))
1084 free_page((unsigned long)pmd);
1088 static bool unmap_pte_range(pmd_t *pmd, unsigned long start, unsigned long end)
1090 pte_t *pte = pte_offset_kernel(pmd, start);
1092 while (start < end) {
1093 set_pte(pte, __pte(0));
1099 if (try_to_free_pte_page((pte_t *)pmd_page_vaddr(*pmd))) {
1106 static void __unmap_pmd_range(pud_t *pud, pmd_t *pmd,
1107 unsigned long start, unsigned long end)
1109 if (unmap_pte_range(pmd, start, end))
1110 if (try_to_free_pmd_page((pmd_t *)pud_page_vaddr(*pud)))
1114 static void unmap_pmd_range(pud_t *pud, unsigned long start, unsigned long end)
1116 pmd_t *pmd = pmd_offset(pud, start);
1119 * Not on a 2MB page boundary?
1121 if (start & (PMD_SIZE - 1)) {
1122 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
1123 unsigned long pre_end = min_t(unsigned long, end, next_page);
1125 __unmap_pmd_range(pud, pmd, start, pre_end);
1132 * Try to unmap in 2M chunks.
1134 while (end - start >= PMD_SIZE) {
1135 if (pmd_large(*pmd))
1138 __unmap_pmd_range(pud, pmd, start, start + PMD_SIZE);
1148 return __unmap_pmd_range(pud, pmd, start, end);
1151 * Try again to free the PMD page if haven't succeeded above.
1153 if (!pud_none(*pud))
1154 if (try_to_free_pmd_page((pmd_t *)pud_page_vaddr(*pud)))
1158 static void unmap_pud_range(p4d_t *p4d, unsigned long start, unsigned long end)
1160 pud_t *pud = pud_offset(p4d, start);
1163 * Not on a GB page boundary?
1165 if (start & (PUD_SIZE - 1)) {
1166 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
1167 unsigned long pre_end = min_t(unsigned long, end, next_page);
1169 unmap_pmd_range(pud, start, pre_end);
1176 * Try to unmap in 1G chunks?
1178 while (end - start >= PUD_SIZE) {
1180 if (pud_large(*pud))
1183 unmap_pmd_range(pud, start, start + PUD_SIZE);
1193 unmap_pmd_range(pud, start, end);
1196 * No need to try to free the PUD page because we'll free it in
1197 * populate_pgd's error path
1201 static int alloc_pte_page(pmd_t *pmd)
1203 pte_t *pte = (pte_t *)get_zeroed_page(GFP_KERNEL);
1207 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
1211 static int alloc_pmd_page(pud_t *pud)
1213 pmd_t *pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL);
1217 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
1221 static void populate_pte(struct cpa_data *cpa,
1222 unsigned long start, unsigned long end,
1223 unsigned num_pages, pmd_t *pmd, pgprot_t pgprot)
1227 pte = pte_offset_kernel(pmd, start);
1229 pgprot = pgprot_clear_protnone_bits(pgprot);
1231 while (num_pages-- && start < end) {
1232 set_pte(pte, pfn_pte(cpa->pfn, pgprot));
1240 static long populate_pmd(struct cpa_data *cpa,
1241 unsigned long start, unsigned long end,
1242 unsigned num_pages, pud_t *pud, pgprot_t pgprot)
1246 pgprot_t pmd_pgprot;
1249 * Not on a 2M boundary?
1251 if (start & (PMD_SIZE - 1)) {
1252 unsigned long pre_end = start + (num_pages << PAGE_SHIFT);
1253 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
1255 pre_end = min_t(unsigned long, pre_end, next_page);
1256 cur_pages = (pre_end - start) >> PAGE_SHIFT;
1257 cur_pages = min_t(unsigned int, num_pages, cur_pages);
1262 pmd = pmd_offset(pud, start);
1264 if (alloc_pte_page(pmd))
1267 populate_pte(cpa, start, pre_end, cur_pages, pmd, pgprot);
1273 * We mapped them all?
1275 if (num_pages == cur_pages)
1278 pmd_pgprot = pgprot_4k_2_large(pgprot);
1280 while (end - start >= PMD_SIZE) {
1283 * We cannot use a 1G page so allocate a PMD page if needed.
1286 if (alloc_pmd_page(pud))
1289 pmd = pmd_offset(pud, start);
1291 set_pmd(pmd, pmd_mkhuge(pfn_pmd(cpa->pfn,
1292 canon_pgprot(pmd_pgprot))));
1295 cpa->pfn += PMD_SIZE >> PAGE_SHIFT;
1296 cur_pages += PMD_SIZE >> PAGE_SHIFT;
1300 * Map trailing 4K pages.
1303 pmd = pmd_offset(pud, start);
1305 if (alloc_pte_page(pmd))
1308 populate_pte(cpa, start, end, num_pages - cur_pages,
1314 static int populate_pud(struct cpa_data *cpa, unsigned long start, p4d_t *p4d,
1320 pgprot_t pud_pgprot;
1322 end = start + (cpa->numpages << PAGE_SHIFT);
1325 * Not on a Gb page boundary? => map everything up to it with
1328 if (start & (PUD_SIZE - 1)) {
1329 unsigned long pre_end;
1330 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
1332 pre_end = min_t(unsigned long, end, next_page);
1333 cur_pages = (pre_end - start) >> PAGE_SHIFT;
1334 cur_pages = min_t(int, (int)cpa->numpages, cur_pages);
1336 pud = pud_offset(p4d, start);
1342 if (alloc_pmd_page(pud))
1345 cur_pages = populate_pmd(cpa, start, pre_end, cur_pages,
1353 /* We mapped them all? */
1354 if (cpa->numpages == cur_pages)
1357 pud = pud_offset(p4d, start);
1358 pud_pgprot = pgprot_4k_2_large(pgprot);
1361 * Map everything starting from the Gb boundary, possibly with 1G pages
1363 while (boot_cpu_has(X86_FEATURE_GBPAGES) && end - start >= PUD_SIZE) {
1364 set_pud(pud, pud_mkhuge(pfn_pud(cpa->pfn,
1365 canon_pgprot(pud_pgprot))));
1368 cpa->pfn += PUD_SIZE >> PAGE_SHIFT;
1369 cur_pages += PUD_SIZE >> PAGE_SHIFT;
1373 /* Map trailing leftover */
1377 pud = pud_offset(p4d, start);
1379 if (alloc_pmd_page(pud))
1382 tmp = populate_pmd(cpa, start, end, cpa->numpages - cur_pages,
1393 * Restrictions for kernel page table do not necessarily apply when mapping in
1396 static int populate_pgd(struct cpa_data *cpa, unsigned long addr)
1398 pgprot_t pgprot = __pgprot(_KERNPG_TABLE);
1399 pud_t *pud = NULL; /* shut up gcc */
1404 pgd_entry = cpa->pgd + pgd_index(addr);
1406 if (pgd_none(*pgd_entry)) {
1407 p4d = (p4d_t *)get_zeroed_page(GFP_KERNEL);
1411 set_pgd(pgd_entry, __pgd(__pa(p4d) | _KERNPG_TABLE));
1415 * Allocate a PUD page and hand it down for mapping.
1417 p4d = p4d_offset(pgd_entry, addr);
1418 if (p4d_none(*p4d)) {
1419 pud = (pud_t *)get_zeroed_page(GFP_KERNEL);
1423 set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE));
1426 pgprot_val(pgprot) &= ~pgprot_val(cpa->mask_clr);
1427 pgprot_val(pgprot) |= pgprot_val(cpa->mask_set);
1429 ret = populate_pud(cpa, addr, p4d, pgprot);
1432 * Leave the PUD page in place in case some other CPU or thread
1433 * already found it, but remove any useless entries we just
1436 unmap_pud_range(p4d, addr,
1437 addr + (cpa->numpages << PAGE_SHIFT));
1441 cpa->numpages = ret;
1445 static int __cpa_process_fault(struct cpa_data *cpa, unsigned long vaddr,
1450 * Right now, we only execute this code path when mapping
1451 * the EFI virtual memory map regions, no other users
1452 * provide a ->pgd value. This may change in the future.
1454 return populate_pgd(cpa, vaddr);
1458 * Ignore all non primary paths.
1466 * Ignore the NULL PTE for kernel identity mapping, as it is expected
1468 * Also set numpages to '1' indicating that we processed cpa req for
1469 * one virtual address page and its pfn. TBD: numpages can be set based
1470 * on the initial value and the level returned by lookup_address().
1472 if (within(vaddr, PAGE_OFFSET,
1473 PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
1475 cpa->pfn = __pa(vaddr) >> PAGE_SHIFT;
1478 } else if (__cpa_pfn_in_highmap(cpa->pfn)) {
1479 /* Faults in the highmap are OK, so do not warn: */
1482 WARN(1, KERN_WARNING "CPA: called for zero pte. "
1483 "vaddr = %lx cpa->vaddr = %lx\n", vaddr,
1490 static int __change_page_attr(struct cpa_data *cpa, int primary)
1492 unsigned long address;
1495 pte_t *kpte, old_pte;
1497 address = __cpa_addr(cpa, cpa->curpage);
1499 kpte = _lookup_address_cpa(cpa, address, &level);
1501 return __cpa_process_fault(cpa, address, primary);
1504 if (pte_none(old_pte))
1505 return __cpa_process_fault(cpa, address, primary);
1507 if (level == PG_LEVEL_4K) {
1509 pgprot_t new_prot = pte_pgprot(old_pte);
1510 unsigned long pfn = pte_pfn(old_pte);
1512 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
1513 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
1515 cpa_inc_4k_install();
1516 /* Hand in lpsize = 0 to enforce the protection mechanism */
1517 new_prot = static_protections(new_prot, address, pfn, 1, 0,
1520 new_prot = pgprot_clear_protnone_bits(new_prot);
1523 * We need to keep the pfn from the existing PTE,
1524 * after all we're only going to change it's attributes
1525 * not the memory it points to
1527 new_pte = pfn_pte(pfn, new_prot);
1530 * Do we really change anything ?
1532 if (pte_val(old_pte) != pte_val(new_pte)) {
1533 set_pte_atomic(kpte, new_pte);
1534 cpa->flags |= CPA_FLUSHTLB;
1541 * Check, whether we can keep the large page intact
1542 * and just change the pte:
1544 do_split = should_split_large_page(kpte, address, cpa);
1546 * When the range fits into the existing large page,
1547 * return. cp->numpages and cpa->tlbflush have been updated in
1554 * We have to split the large page:
1556 err = split_large_page(cpa, kpte, address);
1563 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
1565 static int cpa_process_alias(struct cpa_data *cpa)
1567 struct cpa_data alias_cpa;
1568 unsigned long laddr = (unsigned long)__va(cpa->pfn << PAGE_SHIFT);
1569 unsigned long vaddr;
1572 if (!pfn_range_is_mapped(cpa->pfn, cpa->pfn + 1))
1576 * No need to redo, when the primary call touched the direct
1579 vaddr = __cpa_addr(cpa, cpa->curpage);
1580 if (!(within(vaddr, PAGE_OFFSET,
1581 PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT)))) {
1584 alias_cpa.vaddr = &laddr;
1585 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1586 alias_cpa.curpage = 0;
1588 cpa->force_flush_all = 1;
1590 ret = __change_page_attr_set_clr(&alias_cpa, 0);
1595 #ifdef CONFIG_X86_64
1597 * If the primary call didn't touch the high mapping already
1598 * and the physical address is inside the kernel map, we need
1599 * to touch the high mapped kernel as well:
1601 if (!within(vaddr, (unsigned long)_text, _brk_end) &&
1602 __cpa_pfn_in_highmap(cpa->pfn)) {
1603 unsigned long temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) +
1604 __START_KERNEL_map - phys_base;
1606 alias_cpa.vaddr = &temp_cpa_vaddr;
1607 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1608 alias_cpa.curpage = 0;
1610 cpa->force_flush_all = 1;
1612 * The high mapping range is imprecise, so ignore the
1615 __change_page_attr_set_clr(&alias_cpa, 0);
1622 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
1624 unsigned long numpages = cpa->numpages;
1625 unsigned long rempages = numpages;
1630 * Store the remaining nr of pages for the large page
1631 * preservation check.
1633 cpa->numpages = rempages;
1634 /* for array changes, we can't use large page */
1635 if (cpa->flags & (CPA_ARRAY | CPA_PAGES_ARRAY))
1638 if (!debug_pagealloc_enabled())
1639 spin_lock(&cpa_lock);
1640 ret = __change_page_attr(cpa, checkalias);
1641 if (!debug_pagealloc_enabled())
1642 spin_unlock(&cpa_lock);
1647 ret = cpa_process_alias(cpa);
1653 * Adjust the number of pages with the result of the
1654 * CPA operation. Either a large page has been
1655 * preserved or a single page update happened.
1657 BUG_ON(cpa->numpages > rempages || !cpa->numpages);
1658 rempages -= cpa->numpages;
1659 cpa->curpage += cpa->numpages;
1663 /* Restore the original numpages */
1664 cpa->numpages = numpages;
1668 static int change_page_attr_set_clr(unsigned long *addr, int numpages,
1669 pgprot_t mask_set, pgprot_t mask_clr,
1670 int force_split, int in_flag,
1671 struct page **pages)
1673 struct cpa_data cpa;
1674 int ret, cache, checkalias;
1676 memset(&cpa, 0, sizeof(cpa));
1679 * Check, if we are requested to set a not supported
1680 * feature. Clearing non-supported features is OK.
1682 mask_set = canon_pgprot(mask_set);
1684 if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
1687 /* Ensure we are PAGE_SIZE aligned */
1688 if (in_flag & CPA_ARRAY) {
1690 for (i = 0; i < numpages; i++) {
1691 if (addr[i] & ~PAGE_MASK) {
1692 addr[i] &= PAGE_MASK;
1696 } else if (!(in_flag & CPA_PAGES_ARRAY)) {
1698 * in_flag of CPA_PAGES_ARRAY implies it is aligned.
1699 * No need to check in that case
1701 if (*addr & ~PAGE_MASK) {
1704 * People should not be passing in unaligned addresses:
1710 /* Must avoid aliasing mappings in the highmem code */
1711 kmap_flush_unused();
1717 cpa.numpages = numpages;
1718 cpa.mask_set = mask_set;
1719 cpa.mask_clr = mask_clr;
1722 cpa.force_split = force_split;
1724 if (in_flag & (CPA_ARRAY | CPA_PAGES_ARRAY))
1725 cpa.flags |= in_flag;
1727 /* No alias checking for _NX bit modifications */
1728 checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
1729 /* Has caller explicitly disabled alias checking? */
1730 if (in_flag & CPA_NO_CHECK_ALIAS)
1733 ret = __change_page_attr_set_clr(&cpa, checkalias);
1736 * Check whether we really changed something:
1738 if (!(cpa.flags & CPA_FLUSHTLB))
1742 * No need to flush, when we did not set any of the caching
1745 cache = !!pgprot2cachemode(mask_set);
1748 * On error; flush everything to be sure.
1751 cpa_flush_all(cache);
1755 cpa_flush(&cpa, cache);
1760 static inline int change_page_attr_set(unsigned long *addr, int numpages,
1761 pgprot_t mask, int array)
1763 return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
1764 (array ? CPA_ARRAY : 0), NULL);
1767 static inline int change_page_attr_clear(unsigned long *addr, int numpages,
1768 pgprot_t mask, int array)
1770 return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
1771 (array ? CPA_ARRAY : 0), NULL);
1774 static inline int cpa_set_pages_array(struct page **pages, int numpages,
1777 return change_page_attr_set_clr(NULL, numpages, mask, __pgprot(0), 0,
1778 CPA_PAGES_ARRAY, pages);
1781 static inline int cpa_clear_pages_array(struct page **pages, int numpages,
1784 return change_page_attr_set_clr(NULL, numpages, __pgprot(0), mask, 0,
1785 CPA_PAGES_ARRAY, pages);
1788 int _set_memory_uc(unsigned long addr, int numpages)
1791 * for now UC MINUS. see comments in ioremap_nocache()
1792 * If you really need strong UC use ioremap_uc(), but note
1793 * that you cannot override IO areas with set_memory_*() as
1794 * these helpers cannot work with IO memory.
1796 return change_page_attr_set(&addr, numpages,
1797 cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1801 int set_memory_uc(unsigned long addr, int numpages)
1806 * for now UC MINUS. see comments in ioremap_nocache()
1808 ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1809 _PAGE_CACHE_MODE_UC_MINUS, NULL);
1813 ret = _set_memory_uc(addr, numpages);
1820 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1824 EXPORT_SYMBOL(set_memory_uc);
1826 int _set_memory_wc(unsigned long addr, int numpages)
1830 ret = change_page_attr_set(&addr, numpages,
1831 cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1834 ret = change_page_attr_set_clr(&addr, numpages,
1835 cachemode2pgprot(_PAGE_CACHE_MODE_WC),
1836 __pgprot(_PAGE_CACHE_MASK),
1842 int set_memory_wc(unsigned long addr, int numpages)
1846 ret = reserve_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1847 _PAGE_CACHE_MODE_WC, NULL);
1851 ret = _set_memory_wc(addr, numpages);
1853 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1857 EXPORT_SYMBOL(set_memory_wc);
1859 int _set_memory_wt(unsigned long addr, int numpages)
1861 return change_page_attr_set(&addr, numpages,
1862 cachemode2pgprot(_PAGE_CACHE_MODE_WT), 0);
1865 int _set_memory_wb(unsigned long addr, int numpages)
1867 /* WB cache mode is hard wired to all cache attribute bits being 0 */
1868 return change_page_attr_clear(&addr, numpages,
1869 __pgprot(_PAGE_CACHE_MASK), 0);
1872 int set_memory_wb(unsigned long addr, int numpages)
1876 ret = _set_memory_wb(addr, numpages);
1880 free_memtype(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1883 EXPORT_SYMBOL(set_memory_wb);
1885 int set_memory_x(unsigned long addr, int numpages)
1887 if (!(__supported_pte_mask & _PAGE_NX))
1890 return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1893 int set_memory_nx(unsigned long addr, int numpages)
1895 if (!(__supported_pte_mask & _PAGE_NX))
1898 return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1901 int set_memory_ro(unsigned long addr, int numpages)
1903 return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1906 int set_memory_rw(unsigned long addr, int numpages)
1908 return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
1911 int set_memory_np(unsigned long addr, int numpages)
1913 return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
1916 int set_memory_np_noalias(unsigned long addr, int numpages)
1918 int cpa_flags = CPA_NO_CHECK_ALIAS;
1920 return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1921 __pgprot(_PAGE_PRESENT), 0,
1925 int set_memory_4k(unsigned long addr, int numpages)
1927 return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1928 __pgprot(0), 1, 0, NULL);
1931 int set_memory_nonglobal(unsigned long addr, int numpages)
1933 return change_page_attr_clear(&addr, numpages,
1934 __pgprot(_PAGE_GLOBAL), 0);
1937 int set_memory_global(unsigned long addr, int numpages)
1939 return change_page_attr_set(&addr, numpages,
1940 __pgprot(_PAGE_GLOBAL), 0);
1943 static int __set_memory_enc_dec(unsigned long addr, int numpages, bool enc)
1945 struct cpa_data cpa;
1948 /* Nothing to do if memory encryption is not active */
1949 if (!mem_encrypt_active())
1952 /* Should not be working on unaligned addresses */
1953 if (WARN_ONCE(addr & ~PAGE_MASK, "misaligned address: %#lx\n", addr))
1956 memset(&cpa, 0, sizeof(cpa));
1958 cpa.numpages = numpages;
1959 cpa.mask_set = enc ? __pgprot(_PAGE_ENC) : __pgprot(0);
1960 cpa.mask_clr = enc ? __pgprot(0) : __pgprot(_PAGE_ENC);
1961 cpa.pgd = init_mm.pgd;
1963 /* Must avoid aliasing mappings in the highmem code */
1964 kmap_flush_unused();
1968 * Before changing the encryption attribute, we need to flush caches.
1970 cpa_flush(&cpa, !this_cpu_has(X86_FEATURE_SME_COHERENT));
1972 ret = __change_page_attr_set_clr(&cpa, 1);
1975 * After changing the encryption attribute, we need to flush TLBs again
1976 * in case any speculative TLB caching occurred (but no need to flush
1977 * caches again). We could just use cpa_flush_all(), but in case TLB
1978 * flushing gets optimized in the cpa_flush() path use the same logic
1986 int set_memory_encrypted(unsigned long addr, int numpages)
1988 return __set_memory_enc_dec(addr, numpages, true);
1990 EXPORT_SYMBOL_GPL(set_memory_encrypted);
1992 int set_memory_decrypted(unsigned long addr, int numpages)
1994 return __set_memory_enc_dec(addr, numpages, false);
1996 EXPORT_SYMBOL_GPL(set_memory_decrypted);
1998 int set_pages_uc(struct page *page, int numpages)
2000 unsigned long addr = (unsigned long)page_address(page);
2002 return set_memory_uc(addr, numpages);
2004 EXPORT_SYMBOL(set_pages_uc);
2006 static int _set_pages_array(struct page **pages, int numpages,
2007 enum page_cache_mode new_type)
2009 unsigned long start;
2011 enum page_cache_mode set_type;
2016 for (i = 0; i < numpages; i++) {
2017 if (PageHighMem(pages[i]))
2019 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
2020 end = start + PAGE_SIZE;
2021 if (reserve_memtype(start, end, new_type, NULL))
2025 /* If WC, set to UC- first and then WC */
2026 set_type = (new_type == _PAGE_CACHE_MODE_WC) ?
2027 _PAGE_CACHE_MODE_UC_MINUS : new_type;
2029 ret = cpa_set_pages_array(pages, numpages,
2030 cachemode2pgprot(set_type));
2031 if (!ret && new_type == _PAGE_CACHE_MODE_WC)
2032 ret = change_page_attr_set_clr(NULL, numpages,
2034 _PAGE_CACHE_MODE_WC),
2035 __pgprot(_PAGE_CACHE_MASK),
2036 0, CPA_PAGES_ARRAY, pages);
2039 return 0; /* Success */
2042 for (i = 0; i < free_idx; i++) {
2043 if (PageHighMem(pages[i]))
2045 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
2046 end = start + PAGE_SIZE;
2047 free_memtype(start, end);
2052 int set_pages_array_uc(struct page **pages, int numpages)
2054 return _set_pages_array(pages, numpages, _PAGE_CACHE_MODE_UC_MINUS);
2056 EXPORT_SYMBOL(set_pages_array_uc);
2058 int set_pages_array_wc(struct page **pages, int numpages)
2060 return _set_pages_array(pages, numpages, _PAGE_CACHE_MODE_WC);
2062 EXPORT_SYMBOL(set_pages_array_wc);
2064 int set_pages_array_wt(struct page **pages, int numpages)
2066 return _set_pages_array(pages, numpages, _PAGE_CACHE_MODE_WT);
2068 EXPORT_SYMBOL_GPL(set_pages_array_wt);
2070 int set_pages_wb(struct page *page, int numpages)
2072 unsigned long addr = (unsigned long)page_address(page);
2074 return set_memory_wb(addr, numpages);
2076 EXPORT_SYMBOL(set_pages_wb);
2078 int set_pages_array_wb(struct page **pages, int numpages)
2081 unsigned long start;
2085 /* WB cache mode is hard wired to all cache attribute bits being 0 */
2086 retval = cpa_clear_pages_array(pages, numpages,
2087 __pgprot(_PAGE_CACHE_MASK));
2091 for (i = 0; i < numpages; i++) {
2092 if (PageHighMem(pages[i]))
2094 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
2095 end = start + PAGE_SIZE;
2096 free_memtype(start, end);
2101 EXPORT_SYMBOL(set_pages_array_wb);
2103 int set_pages_ro(struct page *page, int numpages)
2105 unsigned long addr = (unsigned long)page_address(page);
2107 return set_memory_ro(addr, numpages);
2110 int set_pages_rw(struct page *page, int numpages)
2112 unsigned long addr = (unsigned long)page_address(page);
2114 return set_memory_rw(addr, numpages);
2117 static int __set_pages_p(struct page *page, int numpages)
2119 unsigned long tempaddr = (unsigned long) page_address(page);
2120 struct cpa_data cpa = { .vaddr = &tempaddr,
2122 .numpages = numpages,
2123 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
2124 .mask_clr = __pgprot(0),
2128 * No alias checking needed for setting present flag. otherwise,
2129 * we may need to break large pages for 64-bit kernel text
2130 * mappings (this adds to complexity if we want to do this from
2131 * atomic context especially). Let's keep it simple!
2133 return __change_page_attr_set_clr(&cpa, 0);
2136 static int __set_pages_np(struct page *page, int numpages)
2138 unsigned long tempaddr = (unsigned long) page_address(page);
2139 struct cpa_data cpa = { .vaddr = &tempaddr,
2141 .numpages = numpages,
2142 .mask_set = __pgprot(0),
2143 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
2147 * No alias checking needed for setting not present flag. otherwise,
2148 * we may need to break large pages for 64-bit kernel text
2149 * mappings (this adds to complexity if we want to do this from
2150 * atomic context especially). Let's keep it simple!
2152 return __change_page_attr_set_clr(&cpa, 0);
2155 int set_direct_map_invalid_noflush(struct page *page)
2157 return __set_pages_np(page, 1);
2160 int set_direct_map_default_noflush(struct page *page)
2162 return __set_pages_p(page, 1);
2165 void __kernel_map_pages(struct page *page, int numpages, int enable)
2167 if (PageHighMem(page))
2170 debug_check_no_locks_freed(page_address(page),
2171 numpages * PAGE_SIZE);
2175 * The return value is ignored as the calls cannot fail.
2176 * Large pages for identity mappings are not used at boot time
2177 * and hence no memory allocations during large page split.
2180 __set_pages_p(page, numpages);
2182 __set_pages_np(page, numpages);
2185 * We should perform an IPI and flush all tlbs,
2186 * but that can deadlock->flush only current cpu.
2187 * Preemption needs to be disabled around __flush_tlb_all() due to
2188 * CR3 reload in __native_flush_tlb().
2194 arch_flush_lazy_mmu_mode();
2197 #ifdef CONFIG_HIBERNATION
2198 bool kernel_page_present(struct page *page)
2203 if (PageHighMem(page))
2206 pte = lookup_address((unsigned long)page_address(page), &level);
2207 return (pte_val(*pte) & _PAGE_PRESENT);
2209 #endif /* CONFIG_HIBERNATION */
2211 int __init kernel_map_pages_in_pgd(pgd_t *pgd, u64 pfn, unsigned long address,
2212 unsigned numpages, unsigned long page_flags)
2214 int retval = -EINVAL;
2216 struct cpa_data cpa = {
2220 .numpages = numpages,
2221 .mask_set = __pgprot(0),
2222 .mask_clr = __pgprot(~page_flags & (_PAGE_NX|_PAGE_RW)),
2226 WARN_ONCE(num_online_cpus() > 1, "Don't call after initializing SMP");
2228 if (!(__supported_pte_mask & _PAGE_NX))
2231 if (!(page_flags & _PAGE_ENC))
2232 cpa.mask_clr = pgprot_encrypted(cpa.mask_clr);
2234 cpa.mask_set = __pgprot(_PAGE_PRESENT | page_flags);
2236 retval = __change_page_attr_set_clr(&cpa, 0);
2244 * __flush_tlb_all() flushes mappings only on current CPU and hence this
2245 * function shouldn't be used in an SMP environment. Presently, it's used only
2246 * during boot (way before smp_init()) by EFI subsystem and hence is ok.
2248 int __init kernel_unmap_pages_in_pgd(pgd_t *pgd, unsigned long address,
2249 unsigned long numpages)
2254 * The typical sequence for unmapping is to find a pte through
2255 * lookup_address_in_pgd() (ideally, it should never return NULL because
2256 * the address is already mapped) and change it's protections. As pfn is
2257 * the *target* of a mapping, it's not useful while unmapping.
2259 struct cpa_data cpa = {
2263 .numpages = numpages,
2264 .mask_set = __pgprot(0),
2265 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
2269 WARN_ONCE(num_online_cpus() > 1, "Don't call after initializing SMP");
2271 retval = __change_page_attr_set_clr(&cpa, 0);
2278 * The testcases use internal knowledge of the implementation that shouldn't
2279 * be exposed to the rest of the kernel. Include these directly here.
2281 #ifdef CONFIG_CPA_DEBUG
2282 #include "pageattr-test.c"