1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright 2002 Andi Kleen, SuSE Labs.
4 * Thanks to Ben LaHaise for precious feedback.
6 #include <linux/highmem.h>
7 #include <linux/memblock.h>
8 #include <linux/sched.h>
10 #include <linux/interrupt.h>
11 #include <linux/seq_file.h>
12 #include <linux/debugfs.h>
13 #include <linux/pfn.h>
14 #include <linux/percpu.h>
15 #include <linux/gfp.h>
16 #include <linux/pci.h>
17 #include <linux/vmalloc.h>
18 #include <linux/libnvdimm.h>
19 #include <linux/vmstat.h>
20 #include <linux/kernel.h>
21 #include <linux/cc_platform.h>
23 #include <asm/e820/api.h>
24 #include <asm/processor.h>
25 #include <asm/tlbflush.h>
26 #include <asm/sections.h>
27 #include <asm/setup.h>
28 #include <linux/uaccess.h>
29 #include <asm/pgalloc.h>
30 #include <asm/proto.h>
31 #include <asm/memtype.h>
32 #include <asm/set_memory.h>
33 #include <asm/hyperv-tlfs.h>
34 #include <asm/mshyperv.h>
36 #include "../mm_internal.h"
39 * The current flushing context - we pass it instead of 5 arguments:
46 unsigned long numpages;
47 unsigned long curpage;
50 unsigned int force_split : 1,
51 force_static_prot : 1,
62 static const int cpa_warn_level = CPA_PROTECT;
65 * Serialize cpa() (for !DEBUG_PAGEALLOC which uses large identity mappings)
66 * using cpa_lock. So that we don't allow any other cpu, with stale large tlb
67 * entries change the page attribute in parallel to some other cpu
68 * splitting a large page entry along with changing the attribute.
70 static DEFINE_SPINLOCK(cpa_lock);
72 #define CPA_FLUSHTLB 1
74 #define CPA_PAGES_ARRAY 4
75 #define CPA_NO_CHECK_ALIAS 8 /* Do not search for aliases */
77 static inline pgprot_t cachemode2pgprot(enum page_cache_mode pcm)
79 return __pgprot(cachemode2protval(pcm));
83 static unsigned long direct_pages_count[PG_LEVEL_NUM];
85 void update_page_count(int level, unsigned long pages)
87 /* Protect against CPA */
89 direct_pages_count[level] += pages;
90 spin_unlock(&pgd_lock);
93 static void split_page_count(int level)
95 if (direct_pages_count[level] == 0)
98 direct_pages_count[level]--;
99 if (system_state == SYSTEM_RUNNING) {
100 if (level == PG_LEVEL_2M)
101 count_vm_event(DIRECT_MAP_LEVEL2_SPLIT);
102 else if (level == PG_LEVEL_1G)
103 count_vm_event(DIRECT_MAP_LEVEL3_SPLIT);
105 direct_pages_count[level - 1] += PTRS_PER_PTE;
108 void arch_report_meminfo(struct seq_file *m)
110 seq_printf(m, "DirectMap4k: %8lu kB\n",
111 direct_pages_count[PG_LEVEL_4K] << 2);
112 #if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
113 seq_printf(m, "DirectMap2M: %8lu kB\n",
114 direct_pages_count[PG_LEVEL_2M] << 11);
116 seq_printf(m, "DirectMap4M: %8lu kB\n",
117 direct_pages_count[PG_LEVEL_2M] << 12);
120 seq_printf(m, "DirectMap1G: %8lu kB\n",
121 direct_pages_count[PG_LEVEL_1G] << 20);
124 static inline void split_page_count(int level) { }
127 #ifdef CONFIG_X86_CPA_STATISTICS
129 static unsigned long cpa_1g_checked;
130 static unsigned long cpa_1g_sameprot;
131 static unsigned long cpa_1g_preserved;
132 static unsigned long cpa_2m_checked;
133 static unsigned long cpa_2m_sameprot;
134 static unsigned long cpa_2m_preserved;
135 static unsigned long cpa_4k_install;
137 static inline void cpa_inc_1g_checked(void)
142 static inline void cpa_inc_2m_checked(void)
147 static inline void cpa_inc_4k_install(void)
149 data_race(cpa_4k_install++);
152 static inline void cpa_inc_lp_sameprot(int level)
154 if (level == PG_LEVEL_1G)
160 static inline void cpa_inc_lp_preserved(int level)
162 if (level == PG_LEVEL_1G)
168 static int cpastats_show(struct seq_file *m, void *p)
170 seq_printf(m, "1G pages checked: %16lu\n", cpa_1g_checked);
171 seq_printf(m, "1G pages sameprot: %16lu\n", cpa_1g_sameprot);
172 seq_printf(m, "1G pages preserved: %16lu\n", cpa_1g_preserved);
173 seq_printf(m, "2M pages checked: %16lu\n", cpa_2m_checked);
174 seq_printf(m, "2M pages sameprot: %16lu\n", cpa_2m_sameprot);
175 seq_printf(m, "2M pages preserved: %16lu\n", cpa_2m_preserved);
176 seq_printf(m, "4K pages set-checked: %16lu\n", cpa_4k_install);
180 static int cpastats_open(struct inode *inode, struct file *file)
182 return single_open(file, cpastats_show, NULL);
185 static const struct file_operations cpastats_fops = {
186 .open = cpastats_open,
189 .release = single_release,
192 static int __init cpa_stats_init(void)
194 debugfs_create_file("cpa_stats", S_IRUSR, arch_debugfs_dir, NULL,
198 late_initcall(cpa_stats_init);
200 static inline void cpa_inc_1g_checked(void) { }
201 static inline void cpa_inc_2m_checked(void) { }
202 static inline void cpa_inc_4k_install(void) { }
203 static inline void cpa_inc_lp_sameprot(int level) { }
204 static inline void cpa_inc_lp_preserved(int level) { }
209 within(unsigned long addr, unsigned long start, unsigned long end)
211 return addr >= start && addr < end;
215 within_inclusive(unsigned long addr, unsigned long start, unsigned long end)
217 return addr >= start && addr <= end;
222 static inline unsigned long highmap_start_pfn(void)
224 return __pa_symbol(_text) >> PAGE_SHIFT;
227 static inline unsigned long highmap_end_pfn(void)
229 /* Do not reference physical address outside the kernel. */
230 return __pa_symbol(roundup(_brk_end, PMD_SIZE) - 1) >> PAGE_SHIFT;
233 static bool __cpa_pfn_in_highmap(unsigned long pfn)
236 * Kernel text has an alias mapping at a high address, known
239 return within_inclusive(pfn, highmap_start_pfn(), highmap_end_pfn());
244 static bool __cpa_pfn_in_highmap(unsigned long pfn)
246 /* There is no highmap on 32-bit */
253 * See set_mce_nospec().
255 * Machine check recovery code needs to change cache mode of poisoned pages to
256 * UC to avoid speculative access logging another error. But passing the
257 * address of the 1:1 mapping to set_memory_uc() is a fine way to encourage a
258 * speculative access. So we cheat and flip the top bit of the address. This
259 * works fine for the code that updates the page tables. But at the end of the
260 * process we need to flush the TLB and cache and the non-canonical address
261 * causes a #GP fault when used by the INVLPG and CLFLUSH instructions.
263 * But in the common case we already have a canonical address. This code
264 * will fix the top bit if needed and is a no-op otherwise.
266 static inline unsigned long fix_addr(unsigned long addr)
269 return (long)(addr << 1) >> 1;
275 static unsigned long __cpa_addr(struct cpa_data *cpa, unsigned long idx)
277 if (cpa->flags & CPA_PAGES_ARRAY) {
278 struct page *page = cpa->pages[idx];
280 if (unlikely(PageHighMem(page)))
283 return (unsigned long)page_address(page);
286 if (cpa->flags & CPA_ARRAY)
287 return cpa->vaddr[idx];
289 return *cpa->vaddr + idx * PAGE_SIZE;
296 static void clflush_cache_range_opt(void *vaddr, unsigned int size)
298 const unsigned long clflush_size = boot_cpu_data.x86_clflush_size;
299 void *p = (void *)((unsigned long)vaddr & ~(clflush_size - 1));
300 void *vend = vaddr + size;
305 for (; p < vend; p += clflush_size)
310 * clflush_cache_range - flush a cache range with clflush
311 * @vaddr: virtual start address
312 * @size: number of bytes to flush
314 * CLFLUSHOPT is an unordered instruction which needs fencing with MFENCE or
315 * SFENCE to avoid ordering issues.
317 void clflush_cache_range(void *vaddr, unsigned int size)
320 clflush_cache_range_opt(vaddr, size);
323 EXPORT_SYMBOL_GPL(clflush_cache_range);
325 #ifdef CONFIG_ARCH_HAS_PMEM_API
326 void arch_invalidate_pmem(void *addr, size_t size)
328 clflush_cache_range(addr, size);
330 EXPORT_SYMBOL_GPL(arch_invalidate_pmem);
333 static void __cpa_flush_all(void *arg)
335 unsigned long cache = (unsigned long)arg;
338 * Flush all to work around Errata in early athlons regarding
339 * large page flushing.
343 if (cache && boot_cpu_data.x86 >= 4)
347 static void cpa_flush_all(unsigned long cache)
349 BUG_ON(irqs_disabled() && !early_boot_irqs_disabled);
351 on_each_cpu(__cpa_flush_all, (void *) cache, 1);
354 static void __cpa_flush_tlb(void *data)
356 struct cpa_data *cpa = data;
359 for (i = 0; i < cpa->numpages; i++)
360 flush_tlb_one_kernel(fix_addr(__cpa_addr(cpa, i)));
363 static void cpa_flush(struct cpa_data *data, int cache)
365 struct cpa_data *cpa = data;
368 BUG_ON(irqs_disabled() && !early_boot_irqs_disabled);
370 if (cache && !static_cpu_has(X86_FEATURE_CLFLUSH)) {
371 cpa_flush_all(cache);
375 if (cpa->force_flush_all || cpa->numpages > tlb_single_page_flush_ceiling)
378 on_each_cpu(__cpa_flush_tlb, cpa, 1);
384 for (i = 0; i < cpa->numpages; i++) {
385 unsigned long addr = __cpa_addr(cpa, i);
388 pte_t *pte = lookup_address(addr, &level);
391 * Only flush present addresses:
393 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
394 clflush_cache_range_opt((void *)fix_addr(addr), PAGE_SIZE);
399 static bool overlaps(unsigned long r1_start, unsigned long r1_end,
400 unsigned long r2_start, unsigned long r2_end)
402 return (r1_start <= r2_end && r1_end >= r2_start) ||
403 (r2_start <= r1_end && r2_end >= r1_start);
406 #ifdef CONFIG_PCI_BIOS
408 * The BIOS area between 640k and 1Mb needs to be executable for PCI BIOS
409 * based config access (CONFIG_PCI_GOBIOS) support.
411 #define BIOS_PFN PFN_DOWN(BIOS_BEGIN)
412 #define BIOS_PFN_END PFN_DOWN(BIOS_END - 1)
414 static pgprotval_t protect_pci_bios(unsigned long spfn, unsigned long epfn)
416 if (pcibios_enabled && overlaps(spfn, epfn, BIOS_PFN, BIOS_PFN_END))
421 static pgprotval_t protect_pci_bios(unsigned long spfn, unsigned long epfn)
428 * The .rodata section needs to be read-only. Using the pfn catches all
429 * aliases. This also includes __ro_after_init, so do not enforce until
430 * kernel_set_to_readonly is true.
432 static pgprotval_t protect_rodata(unsigned long spfn, unsigned long epfn)
434 unsigned long epfn_ro, spfn_ro = PFN_DOWN(__pa_symbol(__start_rodata));
437 * Note: __end_rodata is at page aligned and not inclusive, so
438 * subtract 1 to get the last enforced PFN in the rodata area.
440 epfn_ro = PFN_DOWN(__pa_symbol(__end_rodata)) - 1;
442 if (kernel_set_to_readonly && overlaps(spfn, epfn, spfn_ro, epfn_ro))
448 * Protect kernel text against becoming non executable by forbidding
449 * _PAGE_NX. This protects only the high kernel mapping (_text -> _etext)
450 * out of which the kernel actually executes. Do not protect the low
453 * This does not cover __inittext since that is gone after boot.
455 static pgprotval_t protect_kernel_text(unsigned long start, unsigned long end)
457 unsigned long t_end = (unsigned long)_etext - 1;
458 unsigned long t_start = (unsigned long)_text;
460 if (overlaps(start, end, t_start, t_end))
465 #if defined(CONFIG_X86_64)
467 * Once the kernel maps the text as RO (kernel_set_to_readonly is set),
468 * kernel text mappings for the large page aligned text, rodata sections
469 * will be always read-only. For the kernel identity mappings covering the
470 * holes caused by this alignment can be anything that user asks.
472 * This will preserve the large page mappings for kernel text/data at no
475 static pgprotval_t protect_kernel_text_ro(unsigned long start,
478 unsigned long t_end = (unsigned long)__end_rodata_hpage_align - 1;
479 unsigned long t_start = (unsigned long)_text;
482 if (!kernel_set_to_readonly || !overlaps(start, end, t_start, t_end))
485 * Don't enforce the !RW mapping for the kernel text mapping, if
486 * the current mapping is already using small page mapping. No
487 * need to work hard to preserve large page mappings in this case.
489 * This also fixes the Linux Xen paravirt guest boot failure caused
490 * by unexpected read-only mappings for kernel identity
491 * mappings. In this paravirt guest case, the kernel text mapping
492 * and the kernel identity mapping share the same page-table pages,
493 * so the protections for kernel text and identity mappings have to
496 if (lookup_address(start, &level) && (level != PG_LEVEL_4K))
501 static pgprotval_t protect_kernel_text_ro(unsigned long start,
508 static inline bool conflicts(pgprot_t prot, pgprotval_t val)
510 return (pgprot_val(prot) & ~val) != pgprot_val(prot);
513 static inline void check_conflict(int warnlvl, pgprot_t prot, pgprotval_t val,
514 unsigned long start, unsigned long end,
515 unsigned long pfn, const char *txt)
517 static const char *lvltxt[] = {
518 [CPA_CONFLICT] = "conflict",
519 [CPA_PROTECT] = "protect",
520 [CPA_DETECT] = "detect",
523 if (warnlvl > cpa_warn_level || !conflicts(prot, val))
526 pr_warn("CPA %8s %10s: 0x%016lx - 0x%016lx PFN %lx req %016llx prevent %016llx\n",
527 lvltxt[warnlvl], txt, start, end, pfn, (unsigned long long)pgprot_val(prot),
528 (unsigned long long)val);
532 * Certain areas of memory on x86 require very specific protection flags,
533 * for example the BIOS area or kernel text. Callers don't always get this
534 * right (again, ioremap() on BIOS memory is not uncommon) so this function
535 * checks and fixes these known static required protection bits.
537 static inline pgprot_t static_protections(pgprot_t prot, unsigned long start,
538 unsigned long pfn, unsigned long npg,
539 unsigned long lpsize, int warnlvl)
541 pgprotval_t forbidden, res;
545 * There is no point in checking RW/NX conflicts when the requested
546 * mapping is setting the page !PRESENT.
548 if (!(pgprot_val(prot) & _PAGE_PRESENT))
551 /* Operate on the virtual address */
552 end = start + npg * PAGE_SIZE - 1;
554 res = protect_kernel_text(start, end);
555 check_conflict(warnlvl, prot, res, start, end, pfn, "Text NX");
559 * Special case to preserve a large page. If the change spawns the
560 * full large page mapping then there is no point to split it
561 * up. Happens with ftrace and is going to be removed once ftrace
562 * switched to text_poke().
564 if (lpsize != (npg * PAGE_SIZE) || (start & (lpsize - 1))) {
565 res = protect_kernel_text_ro(start, end);
566 check_conflict(warnlvl, prot, res, start, end, pfn, "Text RO");
570 /* Check the PFN directly */
571 res = protect_pci_bios(pfn, pfn + npg - 1);
572 check_conflict(warnlvl, prot, res, start, end, pfn, "PCIBIOS NX");
575 res = protect_rodata(pfn, pfn + npg - 1);
576 check_conflict(warnlvl, prot, res, start, end, pfn, "Rodata RO");
579 return __pgprot(pgprot_val(prot) & ~forbidden);
583 * Lookup the page table entry for a virtual address in a specific pgd.
584 * Return a pointer to the entry and the level of the mapping.
586 pte_t *lookup_address_in_pgd(pgd_t *pgd, unsigned long address,
593 *level = PG_LEVEL_NONE;
598 p4d = p4d_offset(pgd, address);
602 *level = PG_LEVEL_512G;
603 if (p4d_large(*p4d) || !p4d_present(*p4d))
606 pud = pud_offset(p4d, address);
610 *level = PG_LEVEL_1G;
611 if (pud_large(*pud) || !pud_present(*pud))
614 pmd = pmd_offset(pud, address);
618 *level = PG_LEVEL_2M;
619 if (pmd_large(*pmd) || !pmd_present(*pmd))
622 *level = PG_LEVEL_4K;
624 return pte_offset_kernel(pmd, address);
628 * Lookup the page table entry for a virtual address. Return a pointer
629 * to the entry and the level of the mapping.
631 * Note: We return pud and pmd either when the entry is marked large
632 * or when the present bit is not set. Otherwise we would return a
633 * pointer to a nonexisting mapping.
635 pte_t *lookup_address(unsigned long address, unsigned int *level)
637 return lookup_address_in_pgd(pgd_offset_k(address), address, level);
639 EXPORT_SYMBOL_GPL(lookup_address);
642 * Lookup the page table entry for a virtual address in a given mm. Return a
643 * pointer to the entry and the level of the mapping.
645 pte_t *lookup_address_in_mm(struct mm_struct *mm, unsigned long address,
648 return lookup_address_in_pgd(pgd_offset(mm, address), address, level);
650 EXPORT_SYMBOL_GPL(lookup_address_in_mm);
652 static pte_t *_lookup_address_cpa(struct cpa_data *cpa, unsigned long address,
656 return lookup_address_in_pgd(cpa->pgd + pgd_index(address),
659 return lookup_address(address, level);
663 * Lookup the PMD entry for a virtual address. Return a pointer to the entry
664 * or NULL if not present.
666 pmd_t *lookup_pmd_address(unsigned long address)
672 pgd = pgd_offset_k(address);
676 p4d = p4d_offset(pgd, address);
677 if (p4d_none(*p4d) || p4d_large(*p4d) || !p4d_present(*p4d))
680 pud = pud_offset(p4d, address);
681 if (pud_none(*pud) || pud_large(*pud) || !pud_present(*pud))
684 return pmd_offset(pud, address);
688 * This is necessary because __pa() does not work on some
689 * kinds of memory, like vmalloc() or the alloc_remap()
690 * areas on 32-bit NUMA systems. The percpu areas can
691 * end up in this kind of memory, for instance.
693 * This could be optimized, but it is only intended to be
694 * used at initialization time, and keeping it
695 * unoptimized should increase the testing coverage for
696 * the more obscure platforms.
698 phys_addr_t slow_virt_to_phys(void *__virt_addr)
700 unsigned long virt_addr = (unsigned long)__virt_addr;
701 phys_addr_t phys_addr;
702 unsigned long offset;
706 pte = lookup_address(virt_addr, &level);
710 * pXX_pfn() returns unsigned long, which must be cast to phys_addr_t
711 * before being left-shifted PAGE_SHIFT bits -- this trick is to
712 * make 32-PAE kernel work correctly.
716 phys_addr = (phys_addr_t)pud_pfn(*(pud_t *)pte) << PAGE_SHIFT;
717 offset = virt_addr & ~PUD_PAGE_MASK;
720 phys_addr = (phys_addr_t)pmd_pfn(*(pmd_t *)pte) << PAGE_SHIFT;
721 offset = virt_addr & ~PMD_PAGE_MASK;
724 phys_addr = (phys_addr_t)pte_pfn(*pte) << PAGE_SHIFT;
725 offset = virt_addr & ~PAGE_MASK;
728 return (phys_addr_t)(phys_addr | offset);
730 EXPORT_SYMBOL_GPL(slow_virt_to_phys);
733 * Set the new pmd in all the pgds we know about:
735 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
738 set_pte_atomic(kpte, pte);
740 if (!SHARED_KERNEL_PMD) {
743 list_for_each_entry(page, &pgd_list, lru) {
749 pgd = (pgd_t *)page_address(page) + pgd_index(address);
750 p4d = p4d_offset(pgd, address);
751 pud = pud_offset(p4d, address);
752 pmd = pmd_offset(pud, address);
753 set_pte_atomic((pte_t *)pmd, pte);
759 static pgprot_t pgprot_clear_protnone_bits(pgprot_t prot)
762 * _PAGE_GLOBAL means "global page" for present PTEs.
763 * But, it is also used to indicate _PAGE_PROTNONE
764 * for non-present PTEs.
766 * This ensures that a _PAGE_GLOBAL PTE going from
767 * present to non-present is not confused as
770 if (!(pgprot_val(prot) & _PAGE_PRESENT))
771 pgprot_val(prot) &= ~_PAGE_GLOBAL;
776 static int __should_split_large_page(pte_t *kpte, unsigned long address,
777 struct cpa_data *cpa)
779 unsigned long numpages, pmask, psize, lpaddr, pfn, old_pfn;
780 pgprot_t old_prot, new_prot, req_prot, chk_prot;
785 * Check for races, another CPU might have split this page
788 tmp = _lookup_address_cpa(cpa, address, &level);
794 old_prot = pmd_pgprot(*(pmd_t *)kpte);
795 old_pfn = pmd_pfn(*(pmd_t *)kpte);
796 cpa_inc_2m_checked();
799 old_prot = pud_pgprot(*(pud_t *)kpte);
800 old_pfn = pud_pfn(*(pud_t *)kpte);
801 cpa_inc_1g_checked();
807 psize = page_level_size(level);
808 pmask = page_level_mask(level);
811 * Calculate the number of pages, which fit into this large
812 * page starting at address:
814 lpaddr = (address + psize) & pmask;
815 numpages = (lpaddr - address) >> PAGE_SHIFT;
816 if (numpages < cpa->numpages)
817 cpa->numpages = numpages;
820 * We are safe now. Check whether the new pgprot is the same:
821 * Convert protection attributes to 4k-format, as cpa->mask* are set
825 /* Clear PSE (aka _PAGE_PAT) and move PAT bit to correct position */
826 req_prot = pgprot_large_2_4k(old_prot);
828 pgprot_val(req_prot) &= ~pgprot_val(cpa->mask_clr);
829 pgprot_val(req_prot) |= pgprot_val(cpa->mask_set);
832 * req_prot is in format of 4k pages. It must be converted to large
833 * page format: the caching mode includes the PAT bit located at
834 * different bit positions in the two formats.
836 req_prot = pgprot_4k_2_large(req_prot);
837 req_prot = pgprot_clear_protnone_bits(req_prot);
838 if (pgprot_val(req_prot) & _PAGE_PRESENT)
839 pgprot_val(req_prot) |= _PAGE_PSE;
842 * old_pfn points to the large page base pfn. So we need to add the
843 * offset of the virtual address:
845 pfn = old_pfn + ((address & (psize - 1)) >> PAGE_SHIFT);
849 * Calculate the large page base address and the number of 4K pages
852 lpaddr = address & pmask;
853 numpages = psize >> PAGE_SHIFT;
856 * Sanity check that the existing mapping is correct versus the static
857 * protections. static_protections() guards against !PRESENT, so no
858 * extra conditional required here.
860 chk_prot = static_protections(old_prot, lpaddr, old_pfn, numpages,
861 psize, CPA_CONFLICT);
863 if (WARN_ON_ONCE(pgprot_val(chk_prot) != pgprot_val(old_prot))) {
865 * Split the large page and tell the split code to
866 * enforce static protections.
868 cpa->force_static_prot = 1;
873 * Optimization: If the requested pgprot is the same as the current
874 * pgprot, then the large page can be preserved and no updates are
875 * required independent of alignment and length of the requested
876 * range. The above already established that the current pgprot is
877 * correct, which in consequence makes the requested pgprot correct
878 * as well if it is the same. The static protection scan below will
879 * not come to a different conclusion.
881 if (pgprot_val(req_prot) == pgprot_val(old_prot)) {
882 cpa_inc_lp_sameprot(level);
887 * If the requested range does not cover the full page, split it up
889 if (address != lpaddr || cpa->numpages != numpages)
893 * Check whether the requested pgprot is conflicting with a static
894 * protection requirement in the large page.
896 new_prot = static_protections(req_prot, lpaddr, old_pfn, numpages,
900 * If there is a conflict, split the large page.
902 * There used to be a 4k wise evaluation trying really hard to
903 * preserve the large pages, but experimentation has shown, that this
904 * does not help at all. There might be corner cases which would
905 * preserve one large page occasionally, but it's really not worth the
906 * extra code and cycles for the common case.
908 if (pgprot_val(req_prot) != pgprot_val(new_prot))
911 /* All checks passed. Update the large page mapping. */
912 new_pte = pfn_pte(old_pfn, new_prot);
913 __set_pmd_pte(kpte, address, new_pte);
914 cpa->flags |= CPA_FLUSHTLB;
915 cpa_inc_lp_preserved(level);
919 static int should_split_large_page(pte_t *kpte, unsigned long address,
920 struct cpa_data *cpa)
924 if (cpa->force_split)
927 spin_lock(&pgd_lock);
928 do_split = __should_split_large_page(kpte, address, cpa);
929 spin_unlock(&pgd_lock);
934 static void split_set_pte(struct cpa_data *cpa, pte_t *pte, unsigned long pfn,
935 pgprot_t ref_prot, unsigned long address,
938 unsigned int npg = PFN_DOWN(size);
942 * If should_split_large_page() discovered an inconsistent mapping,
943 * remove the invalid protection in the split mapping.
945 if (!cpa->force_static_prot)
948 /* Hand in lpsize = 0 to enforce the protection mechanism */
949 prot = static_protections(ref_prot, address, pfn, npg, 0, CPA_PROTECT);
951 if (pgprot_val(prot) == pgprot_val(ref_prot))
955 * If this is splitting a PMD, fix it up. PUD splits cannot be
956 * fixed trivially as that would require to rescan the newly
957 * installed PMD mappings after returning from split_large_page()
958 * so an eventual further split can allocate the necessary PTE
959 * pages. Warn for now and revisit it in case this actually
962 if (size == PAGE_SIZE)
965 pr_warn_once("CPA: Cannot fixup static protections for PUD split\n");
967 set_pte(pte, pfn_pte(pfn, ref_prot));
971 __split_large_page(struct cpa_data *cpa, pte_t *kpte, unsigned long address,
974 unsigned long lpaddr, lpinc, ref_pfn, pfn, pfninc = 1;
975 pte_t *pbase = (pte_t *)page_address(base);
976 unsigned int i, level;
980 spin_lock(&pgd_lock);
982 * Check for races, another CPU might have split this page
985 tmp = _lookup_address_cpa(cpa, address, &level);
987 spin_unlock(&pgd_lock);
991 paravirt_alloc_pte(&init_mm, page_to_pfn(base));
995 ref_prot = pmd_pgprot(*(pmd_t *)kpte);
997 * Clear PSE (aka _PAGE_PAT) and move
998 * PAT bit to correct position.
1000 ref_prot = pgprot_large_2_4k(ref_prot);
1001 ref_pfn = pmd_pfn(*(pmd_t *)kpte);
1002 lpaddr = address & PMD_MASK;
1007 ref_prot = pud_pgprot(*(pud_t *)kpte);
1008 ref_pfn = pud_pfn(*(pud_t *)kpte);
1009 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
1010 lpaddr = address & PUD_MASK;
1013 * Clear the PSE flags if the PRESENT flag is not set
1014 * otherwise pmd_present/pmd_huge will return true
1015 * even on a non present pmd.
1017 if (!(pgprot_val(ref_prot) & _PAGE_PRESENT))
1018 pgprot_val(ref_prot) &= ~_PAGE_PSE;
1022 spin_unlock(&pgd_lock);
1026 ref_prot = pgprot_clear_protnone_bits(ref_prot);
1029 * Get the target pfn from the original entry:
1032 for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc, lpaddr += lpinc)
1033 split_set_pte(cpa, pbase + i, pfn, ref_prot, lpaddr, lpinc);
1035 if (virt_addr_valid(address)) {
1036 unsigned long pfn = PFN_DOWN(__pa(address));
1038 if (pfn_range_is_mapped(pfn, pfn + 1))
1039 split_page_count(level);
1043 * Install the new, split up pagetable.
1045 * We use the standard kernel pagetable protections for the new
1046 * pagetable protections, the actual ptes set above control the
1047 * primary protection behavior:
1049 __set_pmd_pte(kpte, address, mk_pte(base, __pgprot(_KERNPG_TABLE)));
1052 * Do a global flush tlb after splitting the large page
1053 * and before we do the actual change page attribute in the PTE.
1055 * Without this, we violate the TLB application note, that says:
1056 * "The TLBs may contain both ordinary and large-page
1057 * translations for a 4-KByte range of linear addresses. This
1058 * may occur if software modifies the paging structures so that
1059 * the page size used for the address range changes. If the two
1060 * translations differ with respect to page frame or attributes
1061 * (e.g., permissions), processor behavior is undefined and may
1062 * be implementation-specific."
1064 * We do this global tlb flush inside the cpa_lock, so that we
1065 * don't allow any other cpu, with stale tlb entries change the
1066 * page attribute in parallel, that also falls into the
1067 * just split large page entry.
1070 spin_unlock(&pgd_lock);
1075 static int split_large_page(struct cpa_data *cpa, pte_t *kpte,
1076 unsigned long address)
1080 if (!debug_pagealloc_enabled())
1081 spin_unlock(&cpa_lock);
1082 base = alloc_pages(GFP_KERNEL, 0);
1083 if (!debug_pagealloc_enabled())
1084 spin_lock(&cpa_lock);
1088 if (__split_large_page(cpa, kpte, address, base))
1094 static bool try_to_free_pte_page(pte_t *pte)
1098 for (i = 0; i < PTRS_PER_PTE; i++)
1099 if (!pte_none(pte[i]))
1102 free_page((unsigned long)pte);
1106 static bool try_to_free_pmd_page(pmd_t *pmd)
1110 for (i = 0; i < PTRS_PER_PMD; i++)
1111 if (!pmd_none(pmd[i]))
1114 free_page((unsigned long)pmd);
1118 static bool unmap_pte_range(pmd_t *pmd, unsigned long start, unsigned long end)
1120 pte_t *pte = pte_offset_kernel(pmd, start);
1122 while (start < end) {
1123 set_pte(pte, __pte(0));
1129 if (try_to_free_pte_page((pte_t *)pmd_page_vaddr(*pmd))) {
1136 static void __unmap_pmd_range(pud_t *pud, pmd_t *pmd,
1137 unsigned long start, unsigned long end)
1139 if (unmap_pte_range(pmd, start, end))
1140 if (try_to_free_pmd_page(pud_pgtable(*pud)))
1144 static void unmap_pmd_range(pud_t *pud, unsigned long start, unsigned long end)
1146 pmd_t *pmd = pmd_offset(pud, start);
1149 * Not on a 2MB page boundary?
1151 if (start & (PMD_SIZE - 1)) {
1152 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
1153 unsigned long pre_end = min_t(unsigned long, end, next_page);
1155 __unmap_pmd_range(pud, pmd, start, pre_end);
1162 * Try to unmap in 2M chunks.
1164 while (end - start >= PMD_SIZE) {
1165 if (pmd_large(*pmd))
1168 __unmap_pmd_range(pud, pmd, start, start + PMD_SIZE);
1178 return __unmap_pmd_range(pud, pmd, start, end);
1181 * Try again to free the PMD page if haven't succeeded above.
1183 if (!pud_none(*pud))
1184 if (try_to_free_pmd_page(pud_pgtable(*pud)))
1188 static void unmap_pud_range(p4d_t *p4d, unsigned long start, unsigned long end)
1190 pud_t *pud = pud_offset(p4d, start);
1193 * Not on a GB page boundary?
1195 if (start & (PUD_SIZE - 1)) {
1196 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
1197 unsigned long pre_end = min_t(unsigned long, end, next_page);
1199 unmap_pmd_range(pud, start, pre_end);
1206 * Try to unmap in 1G chunks?
1208 while (end - start >= PUD_SIZE) {
1210 if (pud_large(*pud))
1213 unmap_pmd_range(pud, start, start + PUD_SIZE);
1223 unmap_pmd_range(pud, start, end);
1226 * No need to try to free the PUD page because we'll free it in
1227 * populate_pgd's error path
1231 static int alloc_pte_page(pmd_t *pmd)
1233 pte_t *pte = (pte_t *)get_zeroed_page(GFP_KERNEL);
1237 set_pmd(pmd, __pmd(__pa(pte) | _KERNPG_TABLE));
1241 static int alloc_pmd_page(pud_t *pud)
1243 pmd_t *pmd = (pmd_t *)get_zeroed_page(GFP_KERNEL);
1247 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
1251 static void populate_pte(struct cpa_data *cpa,
1252 unsigned long start, unsigned long end,
1253 unsigned num_pages, pmd_t *pmd, pgprot_t pgprot)
1257 pte = pte_offset_kernel(pmd, start);
1259 pgprot = pgprot_clear_protnone_bits(pgprot);
1261 while (num_pages-- && start < end) {
1262 set_pte(pte, pfn_pte(cpa->pfn, pgprot));
1270 static long populate_pmd(struct cpa_data *cpa,
1271 unsigned long start, unsigned long end,
1272 unsigned num_pages, pud_t *pud, pgprot_t pgprot)
1276 pgprot_t pmd_pgprot;
1279 * Not on a 2M boundary?
1281 if (start & (PMD_SIZE - 1)) {
1282 unsigned long pre_end = start + (num_pages << PAGE_SHIFT);
1283 unsigned long next_page = (start + PMD_SIZE) & PMD_MASK;
1285 pre_end = min_t(unsigned long, pre_end, next_page);
1286 cur_pages = (pre_end - start) >> PAGE_SHIFT;
1287 cur_pages = min_t(unsigned int, num_pages, cur_pages);
1292 pmd = pmd_offset(pud, start);
1294 if (alloc_pte_page(pmd))
1297 populate_pte(cpa, start, pre_end, cur_pages, pmd, pgprot);
1303 * We mapped them all?
1305 if (num_pages == cur_pages)
1308 pmd_pgprot = pgprot_4k_2_large(pgprot);
1310 while (end - start >= PMD_SIZE) {
1313 * We cannot use a 1G page so allocate a PMD page if needed.
1316 if (alloc_pmd_page(pud))
1319 pmd = pmd_offset(pud, start);
1321 set_pmd(pmd, pmd_mkhuge(pfn_pmd(cpa->pfn,
1322 canon_pgprot(pmd_pgprot))));
1325 cpa->pfn += PMD_SIZE >> PAGE_SHIFT;
1326 cur_pages += PMD_SIZE >> PAGE_SHIFT;
1330 * Map trailing 4K pages.
1333 pmd = pmd_offset(pud, start);
1335 if (alloc_pte_page(pmd))
1338 populate_pte(cpa, start, end, num_pages - cur_pages,
1344 static int populate_pud(struct cpa_data *cpa, unsigned long start, p4d_t *p4d,
1350 pgprot_t pud_pgprot;
1352 end = start + (cpa->numpages << PAGE_SHIFT);
1355 * Not on a Gb page boundary? => map everything up to it with
1358 if (start & (PUD_SIZE - 1)) {
1359 unsigned long pre_end;
1360 unsigned long next_page = (start + PUD_SIZE) & PUD_MASK;
1362 pre_end = min_t(unsigned long, end, next_page);
1363 cur_pages = (pre_end - start) >> PAGE_SHIFT;
1364 cur_pages = min_t(int, (int)cpa->numpages, cur_pages);
1366 pud = pud_offset(p4d, start);
1372 if (alloc_pmd_page(pud))
1375 cur_pages = populate_pmd(cpa, start, pre_end, cur_pages,
1383 /* We mapped them all? */
1384 if (cpa->numpages == cur_pages)
1387 pud = pud_offset(p4d, start);
1388 pud_pgprot = pgprot_4k_2_large(pgprot);
1391 * Map everything starting from the Gb boundary, possibly with 1G pages
1393 while (boot_cpu_has(X86_FEATURE_GBPAGES) && end - start >= PUD_SIZE) {
1394 set_pud(pud, pud_mkhuge(pfn_pud(cpa->pfn,
1395 canon_pgprot(pud_pgprot))));
1398 cpa->pfn += PUD_SIZE >> PAGE_SHIFT;
1399 cur_pages += PUD_SIZE >> PAGE_SHIFT;
1403 /* Map trailing leftover */
1407 pud = pud_offset(p4d, start);
1409 if (alloc_pmd_page(pud))
1412 tmp = populate_pmd(cpa, start, end, cpa->numpages - cur_pages,
1423 * Restrictions for kernel page table do not necessarily apply when mapping in
1426 static int populate_pgd(struct cpa_data *cpa, unsigned long addr)
1428 pgprot_t pgprot = __pgprot(_KERNPG_TABLE);
1429 pud_t *pud = NULL; /* shut up gcc */
1434 pgd_entry = cpa->pgd + pgd_index(addr);
1436 if (pgd_none(*pgd_entry)) {
1437 p4d = (p4d_t *)get_zeroed_page(GFP_KERNEL);
1441 set_pgd(pgd_entry, __pgd(__pa(p4d) | _KERNPG_TABLE));
1445 * Allocate a PUD page and hand it down for mapping.
1447 p4d = p4d_offset(pgd_entry, addr);
1448 if (p4d_none(*p4d)) {
1449 pud = (pud_t *)get_zeroed_page(GFP_KERNEL);
1453 set_p4d(p4d, __p4d(__pa(pud) | _KERNPG_TABLE));
1456 pgprot_val(pgprot) &= ~pgprot_val(cpa->mask_clr);
1457 pgprot_val(pgprot) |= pgprot_val(cpa->mask_set);
1459 ret = populate_pud(cpa, addr, p4d, pgprot);
1462 * Leave the PUD page in place in case some other CPU or thread
1463 * already found it, but remove any useless entries we just
1466 unmap_pud_range(p4d, addr,
1467 addr + (cpa->numpages << PAGE_SHIFT));
1471 cpa->numpages = ret;
1475 static int __cpa_process_fault(struct cpa_data *cpa, unsigned long vaddr,
1480 * Right now, we only execute this code path when mapping
1481 * the EFI virtual memory map regions, no other users
1482 * provide a ->pgd value. This may change in the future.
1484 return populate_pgd(cpa, vaddr);
1488 * Ignore all non primary paths.
1496 * Ignore the NULL PTE for kernel identity mapping, as it is expected
1498 * Also set numpages to '1' indicating that we processed cpa req for
1499 * one virtual address page and its pfn. TBD: numpages can be set based
1500 * on the initial value and the level returned by lookup_address().
1502 if (within(vaddr, PAGE_OFFSET,
1503 PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
1505 cpa->pfn = __pa(vaddr) >> PAGE_SHIFT;
1508 } else if (__cpa_pfn_in_highmap(cpa->pfn)) {
1509 /* Faults in the highmap are OK, so do not warn: */
1512 WARN(1, KERN_WARNING "CPA: called for zero pte. "
1513 "vaddr = %lx cpa->vaddr = %lx\n", vaddr,
1520 static int __change_page_attr(struct cpa_data *cpa, int primary)
1522 unsigned long address;
1525 pte_t *kpte, old_pte;
1527 address = __cpa_addr(cpa, cpa->curpage);
1529 kpte = _lookup_address_cpa(cpa, address, &level);
1531 return __cpa_process_fault(cpa, address, primary);
1534 if (pte_none(old_pte))
1535 return __cpa_process_fault(cpa, address, primary);
1537 if (level == PG_LEVEL_4K) {
1539 pgprot_t new_prot = pte_pgprot(old_pte);
1540 unsigned long pfn = pte_pfn(old_pte);
1542 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
1543 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
1545 cpa_inc_4k_install();
1546 /* Hand in lpsize = 0 to enforce the protection mechanism */
1547 new_prot = static_protections(new_prot, address, pfn, 1, 0,
1550 new_prot = pgprot_clear_protnone_bits(new_prot);
1553 * We need to keep the pfn from the existing PTE,
1554 * after all we're only going to change it's attributes
1555 * not the memory it points to
1557 new_pte = pfn_pte(pfn, new_prot);
1560 * Do we really change anything ?
1562 if (pte_val(old_pte) != pte_val(new_pte)) {
1563 set_pte_atomic(kpte, new_pte);
1564 cpa->flags |= CPA_FLUSHTLB;
1571 * Check, whether we can keep the large page intact
1572 * and just change the pte:
1574 do_split = should_split_large_page(kpte, address, cpa);
1576 * When the range fits into the existing large page,
1577 * return. cp->numpages and cpa->tlbflush have been updated in
1584 * We have to split the large page:
1586 err = split_large_page(cpa, kpte, address);
1593 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
1595 static int cpa_process_alias(struct cpa_data *cpa)
1597 struct cpa_data alias_cpa;
1598 unsigned long laddr = (unsigned long)__va(cpa->pfn << PAGE_SHIFT);
1599 unsigned long vaddr;
1602 if (!pfn_range_is_mapped(cpa->pfn, cpa->pfn + 1))
1606 * No need to redo, when the primary call touched the direct
1609 vaddr = __cpa_addr(cpa, cpa->curpage);
1610 if (!(within(vaddr, PAGE_OFFSET,
1611 PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT)))) {
1614 alias_cpa.vaddr = &laddr;
1615 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1616 alias_cpa.curpage = 0;
1618 cpa->force_flush_all = 1;
1620 ret = __change_page_attr_set_clr(&alias_cpa, 0);
1625 #ifdef CONFIG_X86_64
1627 * If the primary call didn't touch the high mapping already
1628 * and the physical address is inside the kernel map, we need
1629 * to touch the high mapped kernel as well:
1631 if (!within(vaddr, (unsigned long)_text, _brk_end) &&
1632 __cpa_pfn_in_highmap(cpa->pfn)) {
1633 unsigned long temp_cpa_vaddr = (cpa->pfn << PAGE_SHIFT) +
1634 __START_KERNEL_map - phys_base;
1636 alias_cpa.vaddr = &temp_cpa_vaddr;
1637 alias_cpa.flags &= ~(CPA_PAGES_ARRAY | CPA_ARRAY);
1638 alias_cpa.curpage = 0;
1640 cpa->force_flush_all = 1;
1642 * The high mapping range is imprecise, so ignore the
1645 __change_page_attr_set_clr(&alias_cpa, 0);
1652 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
1654 unsigned long numpages = cpa->numpages;
1655 unsigned long rempages = numpages;
1660 * Store the remaining nr of pages for the large page
1661 * preservation check.
1663 cpa->numpages = rempages;
1664 /* for array changes, we can't use large page */
1665 if (cpa->flags & (CPA_ARRAY | CPA_PAGES_ARRAY))
1668 if (!debug_pagealloc_enabled())
1669 spin_lock(&cpa_lock);
1670 ret = __change_page_attr(cpa, checkalias);
1671 if (!debug_pagealloc_enabled())
1672 spin_unlock(&cpa_lock);
1677 ret = cpa_process_alias(cpa);
1683 * Adjust the number of pages with the result of the
1684 * CPA operation. Either a large page has been
1685 * preserved or a single page update happened.
1687 BUG_ON(cpa->numpages > rempages || !cpa->numpages);
1688 rempages -= cpa->numpages;
1689 cpa->curpage += cpa->numpages;
1693 /* Restore the original numpages */
1694 cpa->numpages = numpages;
1698 static int change_page_attr_set_clr(unsigned long *addr, int numpages,
1699 pgprot_t mask_set, pgprot_t mask_clr,
1700 int force_split, int in_flag,
1701 struct page **pages)
1703 struct cpa_data cpa;
1704 int ret, cache, checkalias;
1706 memset(&cpa, 0, sizeof(cpa));
1709 * Check, if we are requested to set a not supported
1710 * feature. Clearing non-supported features is OK.
1712 mask_set = canon_pgprot(mask_set);
1714 if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
1717 /* Ensure we are PAGE_SIZE aligned */
1718 if (in_flag & CPA_ARRAY) {
1720 for (i = 0; i < numpages; i++) {
1721 if (addr[i] & ~PAGE_MASK) {
1722 addr[i] &= PAGE_MASK;
1726 } else if (!(in_flag & CPA_PAGES_ARRAY)) {
1728 * in_flag of CPA_PAGES_ARRAY implies it is aligned.
1729 * No need to check in that case
1731 if (*addr & ~PAGE_MASK) {
1734 * People should not be passing in unaligned addresses:
1740 /* Must avoid aliasing mappings in the highmem code */
1741 kmap_flush_unused();
1747 cpa.numpages = numpages;
1748 cpa.mask_set = mask_set;
1749 cpa.mask_clr = mask_clr;
1752 cpa.force_split = force_split;
1754 if (in_flag & (CPA_ARRAY | CPA_PAGES_ARRAY))
1755 cpa.flags |= in_flag;
1757 /* No alias checking for _NX bit modifications */
1758 checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
1759 /* Has caller explicitly disabled alias checking? */
1760 if (in_flag & CPA_NO_CHECK_ALIAS)
1763 ret = __change_page_attr_set_clr(&cpa, checkalias);
1766 * Check whether we really changed something:
1768 if (!(cpa.flags & CPA_FLUSHTLB))
1772 * No need to flush, when we did not set any of the caching
1775 cache = !!pgprot2cachemode(mask_set);
1778 * On error; flush everything to be sure.
1781 cpa_flush_all(cache);
1785 cpa_flush(&cpa, cache);
1790 static inline int change_page_attr_set(unsigned long *addr, int numpages,
1791 pgprot_t mask, int array)
1793 return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0,
1794 (array ? CPA_ARRAY : 0), NULL);
1797 static inline int change_page_attr_clear(unsigned long *addr, int numpages,
1798 pgprot_t mask, int array)
1800 return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0,
1801 (array ? CPA_ARRAY : 0), NULL);
1804 static inline int cpa_set_pages_array(struct page **pages, int numpages,
1807 return change_page_attr_set_clr(NULL, numpages, mask, __pgprot(0), 0,
1808 CPA_PAGES_ARRAY, pages);
1811 static inline int cpa_clear_pages_array(struct page **pages, int numpages,
1814 return change_page_attr_set_clr(NULL, numpages, __pgprot(0), mask, 0,
1815 CPA_PAGES_ARRAY, pages);
1819 * _set_memory_prot is an internal helper for callers that have been passed
1820 * a pgprot_t value from upper layers and a reservation has already been taken.
1821 * If you want to set the pgprot to a specific page protocol, use the
1822 * set_memory_xx() functions.
1824 int __set_memory_prot(unsigned long addr, int numpages, pgprot_t prot)
1826 return change_page_attr_set_clr(&addr, numpages, prot,
1827 __pgprot(~pgprot_val(prot)), 0, 0,
1831 int _set_memory_uc(unsigned long addr, int numpages)
1834 * for now UC MINUS. see comments in ioremap()
1835 * If you really need strong UC use ioremap_uc(), but note
1836 * that you cannot override IO areas with set_memory_*() as
1837 * these helpers cannot work with IO memory.
1839 return change_page_attr_set(&addr, numpages,
1840 cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1844 int set_memory_uc(unsigned long addr, int numpages)
1849 * for now UC MINUS. see comments in ioremap()
1851 ret = memtype_reserve(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1852 _PAGE_CACHE_MODE_UC_MINUS, NULL);
1856 ret = _set_memory_uc(addr, numpages);
1863 memtype_free(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1867 EXPORT_SYMBOL(set_memory_uc);
1869 int _set_memory_wc(unsigned long addr, int numpages)
1873 ret = change_page_attr_set(&addr, numpages,
1874 cachemode2pgprot(_PAGE_CACHE_MODE_UC_MINUS),
1877 ret = change_page_attr_set_clr(&addr, numpages,
1878 cachemode2pgprot(_PAGE_CACHE_MODE_WC),
1879 __pgprot(_PAGE_CACHE_MASK),
1885 int set_memory_wc(unsigned long addr, int numpages)
1889 ret = memtype_reserve(__pa(addr), __pa(addr) + numpages * PAGE_SIZE,
1890 _PAGE_CACHE_MODE_WC, NULL);
1894 ret = _set_memory_wc(addr, numpages);
1896 memtype_free(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1900 EXPORT_SYMBOL(set_memory_wc);
1902 int _set_memory_wt(unsigned long addr, int numpages)
1904 return change_page_attr_set(&addr, numpages,
1905 cachemode2pgprot(_PAGE_CACHE_MODE_WT), 0);
1908 int _set_memory_wb(unsigned long addr, int numpages)
1910 /* WB cache mode is hard wired to all cache attribute bits being 0 */
1911 return change_page_attr_clear(&addr, numpages,
1912 __pgprot(_PAGE_CACHE_MASK), 0);
1915 int set_memory_wb(unsigned long addr, int numpages)
1919 ret = _set_memory_wb(addr, numpages);
1923 memtype_free(__pa(addr), __pa(addr) + numpages * PAGE_SIZE);
1926 EXPORT_SYMBOL(set_memory_wb);
1928 int set_memory_x(unsigned long addr, int numpages)
1930 if (!(__supported_pte_mask & _PAGE_NX))
1933 return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_NX), 0);
1936 int set_memory_nx(unsigned long addr, int numpages)
1938 if (!(__supported_pte_mask & _PAGE_NX))
1941 return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_NX), 0);
1944 int set_memory_ro(unsigned long addr, int numpages)
1946 return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_RW), 0);
1949 int set_memory_rw(unsigned long addr, int numpages)
1951 return change_page_attr_set(&addr, numpages, __pgprot(_PAGE_RW), 0);
1954 int set_memory_np(unsigned long addr, int numpages)
1956 return change_page_attr_clear(&addr, numpages, __pgprot(_PAGE_PRESENT), 0);
1959 int set_memory_np_noalias(unsigned long addr, int numpages)
1961 int cpa_flags = CPA_NO_CHECK_ALIAS;
1963 return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1964 __pgprot(_PAGE_PRESENT), 0,
1968 int set_memory_4k(unsigned long addr, int numpages)
1970 return change_page_attr_set_clr(&addr, numpages, __pgprot(0),
1971 __pgprot(0), 1, 0, NULL);
1974 int set_memory_nonglobal(unsigned long addr, int numpages)
1976 return change_page_attr_clear(&addr, numpages,
1977 __pgprot(_PAGE_GLOBAL), 0);
1980 int set_memory_global(unsigned long addr, int numpages)
1982 return change_page_attr_set(&addr, numpages,
1983 __pgprot(_PAGE_GLOBAL), 0);
1987 * __set_memory_enc_pgtable() is used for the hypervisors that get
1988 * informed about "encryption" status via page tables.
1990 static int __set_memory_enc_pgtable(unsigned long addr, int numpages, bool enc)
1992 struct cpa_data cpa;
1995 /* Should not be working on unaligned addresses */
1996 if (WARN_ONCE(addr & ~PAGE_MASK, "misaligned address: %#lx\n", addr))
1999 memset(&cpa, 0, sizeof(cpa));
2001 cpa.numpages = numpages;
2002 cpa.mask_set = enc ? __pgprot(_PAGE_ENC) : __pgprot(0);
2003 cpa.mask_clr = enc ? __pgprot(0) : __pgprot(_PAGE_ENC);
2004 cpa.pgd = init_mm.pgd;
2006 /* Must avoid aliasing mappings in the highmem code */
2007 kmap_flush_unused();
2011 * Before changing the encryption attribute, we need to flush caches.
2013 cpa_flush(&cpa, !this_cpu_has(X86_FEATURE_SME_COHERENT));
2015 ret = __change_page_attr_set_clr(&cpa, 1);
2018 * After changing the encryption attribute, we need to flush TLBs again
2019 * in case any speculative TLB caching occurred (but no need to flush
2020 * caches again). We could just use cpa_flush_all(), but in case TLB
2021 * flushing gets optimized in the cpa_flush() path use the same logic
2027 * Notify hypervisor that a given memory range is mapped encrypted
2030 notify_range_enc_status_changed(addr, numpages, enc);
2035 static int __set_memory_enc_dec(unsigned long addr, int numpages, bool enc)
2037 if (hv_is_isolation_supported())
2038 return hv_set_mem_host_visibility(addr, numpages, !enc);
2040 if (cc_platform_has(CC_ATTR_MEM_ENCRYPT))
2041 return __set_memory_enc_pgtable(addr, numpages, enc);
2046 int set_memory_encrypted(unsigned long addr, int numpages)
2048 return __set_memory_enc_dec(addr, numpages, true);
2050 EXPORT_SYMBOL_GPL(set_memory_encrypted);
2052 int set_memory_decrypted(unsigned long addr, int numpages)
2054 return __set_memory_enc_dec(addr, numpages, false);
2056 EXPORT_SYMBOL_GPL(set_memory_decrypted);
2058 int set_pages_uc(struct page *page, int numpages)
2060 unsigned long addr = (unsigned long)page_address(page);
2062 return set_memory_uc(addr, numpages);
2064 EXPORT_SYMBOL(set_pages_uc);
2066 static int _set_pages_array(struct page **pages, int numpages,
2067 enum page_cache_mode new_type)
2069 unsigned long start;
2071 enum page_cache_mode set_type;
2076 for (i = 0; i < numpages; i++) {
2077 if (PageHighMem(pages[i]))
2079 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
2080 end = start + PAGE_SIZE;
2081 if (memtype_reserve(start, end, new_type, NULL))
2085 /* If WC, set to UC- first and then WC */
2086 set_type = (new_type == _PAGE_CACHE_MODE_WC) ?
2087 _PAGE_CACHE_MODE_UC_MINUS : new_type;
2089 ret = cpa_set_pages_array(pages, numpages,
2090 cachemode2pgprot(set_type));
2091 if (!ret && new_type == _PAGE_CACHE_MODE_WC)
2092 ret = change_page_attr_set_clr(NULL, numpages,
2094 _PAGE_CACHE_MODE_WC),
2095 __pgprot(_PAGE_CACHE_MASK),
2096 0, CPA_PAGES_ARRAY, pages);
2099 return 0; /* Success */
2102 for (i = 0; i < free_idx; i++) {
2103 if (PageHighMem(pages[i]))
2105 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
2106 end = start + PAGE_SIZE;
2107 memtype_free(start, end);
2112 int set_pages_array_uc(struct page **pages, int numpages)
2114 return _set_pages_array(pages, numpages, _PAGE_CACHE_MODE_UC_MINUS);
2116 EXPORT_SYMBOL(set_pages_array_uc);
2118 int set_pages_array_wc(struct page **pages, int numpages)
2120 return _set_pages_array(pages, numpages, _PAGE_CACHE_MODE_WC);
2122 EXPORT_SYMBOL(set_pages_array_wc);
2124 int set_pages_array_wt(struct page **pages, int numpages)
2126 return _set_pages_array(pages, numpages, _PAGE_CACHE_MODE_WT);
2128 EXPORT_SYMBOL_GPL(set_pages_array_wt);
2130 int set_pages_wb(struct page *page, int numpages)
2132 unsigned long addr = (unsigned long)page_address(page);
2134 return set_memory_wb(addr, numpages);
2136 EXPORT_SYMBOL(set_pages_wb);
2138 int set_pages_array_wb(struct page **pages, int numpages)
2141 unsigned long start;
2145 /* WB cache mode is hard wired to all cache attribute bits being 0 */
2146 retval = cpa_clear_pages_array(pages, numpages,
2147 __pgprot(_PAGE_CACHE_MASK));
2151 for (i = 0; i < numpages; i++) {
2152 if (PageHighMem(pages[i]))
2154 start = page_to_pfn(pages[i]) << PAGE_SHIFT;
2155 end = start + PAGE_SIZE;
2156 memtype_free(start, end);
2161 EXPORT_SYMBOL(set_pages_array_wb);
2163 int set_pages_ro(struct page *page, int numpages)
2165 unsigned long addr = (unsigned long)page_address(page);
2167 return set_memory_ro(addr, numpages);
2170 int set_pages_rw(struct page *page, int numpages)
2172 unsigned long addr = (unsigned long)page_address(page);
2174 return set_memory_rw(addr, numpages);
2177 static int __set_pages_p(struct page *page, int numpages)
2179 unsigned long tempaddr = (unsigned long) page_address(page);
2180 struct cpa_data cpa = { .vaddr = &tempaddr,
2182 .numpages = numpages,
2183 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
2184 .mask_clr = __pgprot(0),
2188 * No alias checking needed for setting present flag. otherwise,
2189 * we may need to break large pages for 64-bit kernel text
2190 * mappings (this adds to complexity if we want to do this from
2191 * atomic context especially). Let's keep it simple!
2193 return __change_page_attr_set_clr(&cpa, 0);
2196 static int __set_pages_np(struct page *page, int numpages)
2198 unsigned long tempaddr = (unsigned long) page_address(page);
2199 struct cpa_data cpa = { .vaddr = &tempaddr,
2201 .numpages = numpages,
2202 .mask_set = __pgprot(0),
2203 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
2207 * No alias checking needed for setting not present flag. otherwise,
2208 * we may need to break large pages for 64-bit kernel text
2209 * mappings (this adds to complexity if we want to do this from
2210 * atomic context especially). Let's keep it simple!
2212 return __change_page_attr_set_clr(&cpa, 0);
2215 int set_direct_map_invalid_noflush(struct page *page)
2217 return __set_pages_np(page, 1);
2220 int set_direct_map_default_noflush(struct page *page)
2222 return __set_pages_p(page, 1);
2225 #ifdef CONFIG_DEBUG_PAGEALLOC
2226 void __kernel_map_pages(struct page *page, int numpages, int enable)
2228 if (PageHighMem(page))
2231 debug_check_no_locks_freed(page_address(page),
2232 numpages * PAGE_SIZE);
2236 * The return value is ignored as the calls cannot fail.
2237 * Large pages for identity mappings are not used at boot time
2238 * and hence no memory allocations during large page split.
2241 __set_pages_p(page, numpages);
2243 __set_pages_np(page, numpages);
2246 * We should perform an IPI and flush all tlbs,
2247 * but that can deadlock->flush only current cpu.
2248 * Preemption needs to be disabled around __flush_tlb_all() due to
2249 * CR3 reload in __native_flush_tlb().
2255 arch_flush_lazy_mmu_mode();
2257 #endif /* CONFIG_DEBUG_PAGEALLOC */
2259 bool kernel_page_present(struct page *page)
2264 if (PageHighMem(page))
2267 pte = lookup_address((unsigned long)page_address(page), &level);
2268 return (pte_val(*pte) & _PAGE_PRESENT);
2271 int __init kernel_map_pages_in_pgd(pgd_t *pgd, u64 pfn, unsigned long address,
2272 unsigned numpages, unsigned long page_flags)
2274 int retval = -EINVAL;
2276 struct cpa_data cpa = {
2280 .numpages = numpages,
2281 .mask_set = __pgprot(0),
2282 .mask_clr = __pgprot(~page_flags & (_PAGE_NX|_PAGE_RW)),
2286 WARN_ONCE(num_online_cpus() > 1, "Don't call after initializing SMP");
2288 if (!(__supported_pte_mask & _PAGE_NX))
2291 if (!(page_flags & _PAGE_ENC))
2292 cpa.mask_clr = pgprot_encrypted(cpa.mask_clr);
2294 cpa.mask_set = __pgprot(_PAGE_PRESENT | page_flags);
2296 retval = __change_page_attr_set_clr(&cpa, 0);
2304 * __flush_tlb_all() flushes mappings only on current CPU and hence this
2305 * function shouldn't be used in an SMP environment. Presently, it's used only
2306 * during boot (way before smp_init()) by EFI subsystem and hence is ok.
2308 int __init kernel_unmap_pages_in_pgd(pgd_t *pgd, unsigned long address,
2309 unsigned long numpages)
2314 * The typical sequence for unmapping is to find a pte through
2315 * lookup_address_in_pgd() (ideally, it should never return NULL because
2316 * the address is already mapped) and change it's protections. As pfn is
2317 * the *target* of a mapping, it's not useful while unmapping.
2319 struct cpa_data cpa = {
2323 .numpages = numpages,
2324 .mask_set = __pgprot(0),
2325 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW),
2329 WARN_ONCE(num_online_cpus() > 1, "Don't call after initializing SMP");
2331 retval = __change_page_attr_set_clr(&cpa, 0);
2338 * The testcases use internal knowledge of the implementation that shouldn't
2339 * be exposed to the rest of the kernel. Include these directly here.
2341 #ifdef CONFIG_CPA_DEBUG
2342 #include "cpa-test.c"