2 * Copyright 2002 Andi Kleen, SuSE Labs.
3 * Thanks to Ben LaHaise for precious feedback.
5 #include <linux/highmem.h>
6 #include <linux/bootmem.h>
7 #include <linux/module.h>
8 #include <linux/sched.h>
9 #include <linux/slab.h>
11 #include <linux/interrupt.h>
12 #include <linux/seq_file.h>
13 #include <linux/debugfs.h>
16 #include <asm/processor.h>
17 #include <asm/tlbflush.h>
18 #include <asm/sections.h>
19 #include <asm/uaccess.h>
20 #include <asm/pgalloc.h>
21 #include <asm/proto.h>
25 * The current flushing context - we pass it instead of 5 arguments:
34 unsigned force_split : 1;
39 static inline unsigned long highmap_start_pfn(void)
41 return __pa(_text) >> PAGE_SHIFT;
44 static inline unsigned long highmap_end_pfn(void)
46 return __pa(round_up((unsigned long)_end, PMD_SIZE)) >> PAGE_SHIFT;
51 #ifdef CONFIG_DEBUG_PAGEALLOC
52 # define debug_pagealloc 1
54 # define debug_pagealloc 0
58 within(unsigned long addr, unsigned long start, unsigned long end)
60 return addr >= start && addr < end;
68 * clflush_cache_range - flush a cache range with clflush
69 * @addr: virtual start address
70 * @size: number of bytes to flush
72 * clflush is an unordered instruction which needs fencing with mfence
73 * to avoid ordering issues.
75 void clflush_cache_range(void *vaddr, unsigned int size)
77 void *vend = vaddr + size - 1;
81 for (; vaddr < vend; vaddr += boot_cpu_data.x86_clflush_size)
84 * Flush any possible final partial cacheline:
91 static void __cpa_flush_all(void *arg)
93 unsigned long cache = (unsigned long)arg;
96 * Flush all to work around Errata in early athlons regarding
97 * large page flushing.
101 if (cache && boot_cpu_data.x86_model >= 4)
105 static void cpa_flush_all(unsigned long cache)
107 BUG_ON(irqs_disabled());
109 on_each_cpu(__cpa_flush_all, (void *) cache, 1, 1);
112 static void __cpa_flush_range(void *arg)
115 * We could optimize that further and do individual per page
116 * tlb invalidates for a low number of pages. Caveat: we must
117 * flush the high aliases on 64bit as well.
122 static void cpa_flush_range(unsigned long start, int numpages, int cache)
124 unsigned int i, level;
127 BUG_ON(irqs_disabled());
128 WARN_ON(PAGE_ALIGN(start) != start);
130 on_each_cpu(__cpa_flush_range, NULL, 1, 1);
136 * We only need to flush on one CPU,
137 * clflush is a MESI-coherent instruction that
138 * will cause all other CPUs to flush the same
141 for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
142 pte_t *pte = lookup_address(addr, &level);
145 * Only flush present addresses:
147 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
148 clflush_cache_range((void *) addr, PAGE_SIZE);
153 * Certain areas of memory on x86 require very specific protection flags,
154 * for example the BIOS area or kernel text. Callers don't always get this
155 * right (again, ioremap() on BIOS memory is not uncommon) so this function
156 * checks and fixes these known static required protection bits.
158 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
161 pgprot_t forbidden = __pgprot(0);
164 * The BIOS area between 640k and 1Mb needs to be executable for
165 * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
167 if (within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
168 pgprot_val(forbidden) |= _PAGE_NX;
171 * The kernel text needs to be executable for obvious reasons
172 * Does not cover __inittext since that is gone later on. On
173 * 64bit we do not enforce !NX on the low mapping
175 if (within(address, (unsigned long)_text, (unsigned long)_etext))
176 pgprot_val(forbidden) |= _PAGE_NX;
179 * The .rodata section needs to be read-only. Using the pfn
180 * catches all aliases.
182 if (within(pfn, __pa((unsigned long)__start_rodata) >> PAGE_SHIFT,
183 __pa((unsigned long)__end_rodata) >> PAGE_SHIFT))
184 pgprot_val(forbidden) |= _PAGE_RW;
186 prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
192 * Lookup the page table entry for a virtual address. Return a pointer
193 * to the entry and the level of the mapping.
195 * Note: We return pud and pmd either when the entry is marked large
196 * or when the present bit is not set. Otherwise we would return a
197 * pointer to a nonexisting mapping.
199 pte_t *lookup_address(unsigned long address, unsigned int *level)
201 pgd_t *pgd = pgd_offset_k(address);
205 *level = PG_LEVEL_NONE;
210 pud = pud_offset(pgd, address);
214 *level = PG_LEVEL_1G;
215 if (pud_large(*pud) || !pud_present(*pud))
218 pmd = pmd_offset(pud, address);
222 *level = PG_LEVEL_2M;
223 if (pmd_large(*pmd) || !pmd_present(*pmd))
226 *level = PG_LEVEL_4K;
228 return pte_offset_kernel(pmd, address);
230 EXPORT_SYMBOL_GPL(lookup_address);
233 * Set the new pmd in all the pgds we know about:
235 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
238 set_pte_atomic(kpte, pte);
240 if (!SHARED_KERNEL_PMD) {
243 list_for_each_entry(page, &pgd_list, lru) {
248 pgd = (pgd_t *)page_address(page) + pgd_index(address);
249 pud = pud_offset(pgd, address);
250 pmd = pmd_offset(pud, address);
251 set_pte_atomic((pte_t *)pmd, pte);
258 try_preserve_large_page(pte_t *kpte, unsigned long address,
259 struct cpa_data *cpa)
261 unsigned long nextpage_addr, numpages, pmask, psize, flags, addr, pfn;
262 pte_t new_pte, old_pte, *tmp;
263 pgprot_t old_prot, new_prot;
267 if (cpa->force_split)
270 spin_lock_irqsave(&pgd_lock, flags);
272 * Check for races, another CPU might have split this page
275 tmp = lookup_address(address, &level);
281 psize = PMD_PAGE_SIZE;
282 pmask = PMD_PAGE_MASK;
286 psize = PUD_PAGE_SIZE;
287 pmask = PUD_PAGE_MASK;
296 * Calculate the number of pages, which fit into this large
297 * page starting at address:
299 nextpage_addr = (address + psize) & pmask;
300 numpages = (nextpage_addr - address) >> PAGE_SHIFT;
301 if (numpages < cpa->numpages)
302 cpa->numpages = numpages;
305 * We are safe now. Check whether the new pgprot is the same:
308 old_prot = new_prot = pte_pgprot(old_pte);
310 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
311 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
314 * old_pte points to the large page base address. So we need
315 * to add the offset of the virtual address:
317 pfn = pte_pfn(old_pte) + ((address & (psize - 1)) >> PAGE_SHIFT);
320 new_prot = static_protections(new_prot, address, pfn);
323 * We need to check the full range, whether
324 * static_protection() requires a different pgprot for one of
325 * the pages in the range we try to preserve:
327 addr = address + PAGE_SIZE;
329 for (i = 1; i < cpa->numpages; i++, addr += PAGE_SIZE, pfn++) {
330 pgprot_t chk_prot = static_protections(new_prot, addr, pfn);
332 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
337 * If there are no changes, return. maxpages has been updated
340 if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
346 * We need to change the attributes. Check, whether we can
347 * change the large page in one go. We request a split, when
348 * the address is not aligned and the number of pages is
349 * smaller than the number of pages in the large page. Note
350 * that we limited the number of possible pages already to
351 * the number of pages in the large page.
353 if (address == (nextpage_addr - psize) && cpa->numpages == numpages) {
355 * The address is aligned and the number of pages
356 * covers the full page.
358 new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
359 __set_pmd_pte(kpte, address, new_pte);
365 spin_unlock_irqrestore(&pgd_lock, flags);
370 static LIST_HEAD(page_pool);
371 static unsigned long pool_size, pool_pages, pool_low;
372 static unsigned long pool_used, pool_failed;
374 static void cpa_fill_pool(struct page **ret)
376 gfp_t gfp = GFP_KERNEL;
381 * Avoid recursion (on debug-pagealloc) and also signal
382 * our priority to get to these pagetables:
384 if (current->flags & PF_MEMALLOC)
386 current->flags |= PF_MEMALLOC;
389 * Allocate atomically from atomic contexts:
391 if (in_atomic() || irqs_disabled() || debug_pagealloc)
392 gfp = GFP_ATOMIC | __GFP_NORETRY | __GFP_NOWARN;
394 while (pool_pages < pool_size || (ret && !*ret)) {
395 p = alloc_pages(gfp, 0);
401 * If the call site needs a page right now, provide it:
407 spin_lock_irqsave(&pgd_lock, flags);
408 list_add(&p->lru, &page_pool);
410 spin_unlock_irqrestore(&pgd_lock, flags);
413 current->flags &= ~PF_MEMALLOC;
416 #define SHIFT_MB (20 - PAGE_SHIFT)
417 #define ROUND_MB_GB ((1 << 10) - 1)
418 #define SHIFT_MB_GB 10
419 #define POOL_PAGES_PER_GB 16
421 void __init cpa_init(void)
428 * Calculate the number of pool pages:
430 * Convert totalram (nr of pages) to MiB and round to the next
431 * GiB. Shift MiB to Gib and multiply the result by
434 if (debug_pagealloc) {
435 gb = ((si.totalram >> SHIFT_MB) + ROUND_MB_GB) >> SHIFT_MB_GB;
436 pool_size = POOL_PAGES_PER_GB * gb;
440 pool_low = pool_size;
444 "CPA: page pool initialized %lu of %lu pages preallocated\n",
445 pool_pages, pool_size);
448 static int split_large_page(pte_t *kpte, unsigned long address)
450 unsigned long flags, pfn, pfninc = 1;
451 unsigned int i, level;
457 * Get a page from the pool. The pool list is protected by the
458 * pgd_lock, which we have to take anyway for the split
461 spin_lock_irqsave(&pgd_lock, flags);
462 if (list_empty(&page_pool)) {
463 spin_unlock_irqrestore(&pgd_lock, flags);
465 cpa_fill_pool(&base);
468 spin_lock_irqsave(&pgd_lock, flags);
470 base = list_first_entry(&page_pool, struct page, lru);
471 list_del(&base->lru);
474 if (pool_pages < pool_low)
475 pool_low = pool_pages;
479 * Check for races, another CPU might have split this page
482 tmp = lookup_address(address, &level);
486 pbase = (pte_t *)page_address(base);
487 paravirt_alloc_pte(&init_mm, page_to_pfn(base));
488 ref_prot = pte_pgprot(pte_clrhuge(*kpte));
491 if (level == PG_LEVEL_1G) {
492 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
493 pgprot_val(ref_prot) |= _PAGE_PSE;
498 * Get the target pfn from the original entry:
500 pfn = pte_pfn(*kpte);
501 for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
502 set_pte(&pbase[i], pfn_pte(pfn, ref_prot));
505 * Install the new, split up pagetable. Important details here:
507 * On Intel the NX bit of all levels must be cleared to make a
508 * page executable. See section 4.13.2 of Intel 64 and IA-32
509 * Architectures Software Developer's Manual).
511 * Mark the entry present. The current mapping might be
512 * set to not present, which we preserved above.
514 ref_prot = pte_pgprot(pte_mkexec(pte_clrhuge(*kpte)));
515 pgprot_val(ref_prot) |= _PAGE_PRESENT;
516 __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
521 * If we dropped out via the lookup_address check under
522 * pgd_lock then stick the page back into the pool:
525 list_add(&base->lru, &page_pool);
529 spin_unlock_irqrestore(&pgd_lock, flags);
534 static int __change_page_attr(struct cpa_data *cpa, int primary)
536 unsigned long address = cpa->vaddr;
539 pte_t *kpte, old_pte;
542 kpte = lookup_address(address, &level);
547 if (!pte_val(old_pte)) {
550 printk(KERN_WARNING "CPA: called for zero pte. "
551 "vaddr = %lx cpa->vaddr = %lx\n", address,
557 if (level == PG_LEVEL_4K) {
559 pgprot_t new_prot = pte_pgprot(old_pte);
560 unsigned long pfn = pte_pfn(old_pte);
562 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
563 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
565 new_prot = static_protections(new_prot, address, pfn);
568 * We need to keep the pfn from the existing PTE,
569 * after all we're only going to change it's attributes
570 * not the memory it points to
572 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
575 * Do we really change anything ?
577 if (pte_val(old_pte) != pte_val(new_pte)) {
578 set_pte_atomic(kpte, new_pte);
586 * Check, whether we can keep the large page intact
587 * and just change the pte:
589 do_split = try_preserve_large_page(kpte, address, cpa);
591 * When the range fits into the existing large page,
592 * return. cp->numpages and cpa->tlbflush have been updated in
599 * We have to split the large page:
601 err = split_large_page(kpte, address);
610 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
612 static int cpa_process_alias(struct cpa_data *cpa)
614 struct cpa_data alias_cpa;
617 if (cpa->pfn > max_pfn_mapped)
621 * No need to redo, when the primary call touched the direct
624 if (!within(cpa->vaddr, PAGE_OFFSET,
625 PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
628 alias_cpa.vaddr = (unsigned long) __va(cpa->pfn << PAGE_SHIFT);
630 ret = __change_page_attr_set_clr(&alias_cpa, 0);
637 * No need to redo, when the primary call touched the high
640 if (within(cpa->vaddr, (unsigned long) _text, (unsigned long) _end))
644 * If the physical address is inside the kernel map, we need
645 * to touch the high mapped kernel as well:
647 if (!within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn()))
652 (cpa->pfn << PAGE_SHIFT) + __START_KERNEL_map - phys_base;
655 * The high mapping range is imprecise, so ignore the return value.
657 __change_page_attr_set_clr(&alias_cpa, 0);
662 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
664 int ret, numpages = cpa->numpages;
668 * Store the remaining nr of pages for the large page
669 * preservation check.
671 cpa->numpages = numpages;
673 ret = __change_page_attr(cpa, checkalias);
678 ret = cpa_process_alias(cpa);
684 * Adjust the number of pages with the result of the
685 * CPA operation. Either a large page has been
686 * preserved or a single page update happened.
688 BUG_ON(cpa->numpages > numpages);
689 numpages -= cpa->numpages;
690 cpa->vaddr += cpa->numpages * PAGE_SIZE;
695 static inline int cache_attr(pgprot_t attr)
697 return pgprot_val(attr) &
698 (_PAGE_PAT | _PAGE_PAT_LARGE | _PAGE_PWT | _PAGE_PCD);
701 static int change_page_attr_set_clr(unsigned long addr, int numpages,
702 pgprot_t mask_set, pgprot_t mask_clr,
706 int ret, cache, checkalias;
709 * Check, if we are requested to change a not supported
712 mask_set = canon_pgprot(mask_set);
713 mask_clr = canon_pgprot(mask_clr);
714 if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
717 /* Ensure we are PAGE_SIZE aligned */
718 if (addr & ~PAGE_MASK) {
721 * People should not be passing in unaligned addresses:
727 cpa.numpages = numpages;
728 cpa.mask_set = mask_set;
729 cpa.mask_clr = mask_clr;
731 cpa.force_split = force_split;
733 /* No alias checking for _NX bit modifications */
734 checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
736 ret = __change_page_attr_set_clr(&cpa, checkalias);
739 * Check whether we really changed something:
745 * No need to flush, when we did not set any of the caching
748 cache = cache_attr(mask_set);
751 * On success we use clflush, when the CPU supports it to
752 * avoid the wbindv. If the CPU does not support it and in the
753 * error case we fall back to cpa_flush_all (which uses
756 if (!ret && cpu_has_clflush)
757 cpa_flush_range(addr, numpages, cache);
759 cpa_flush_all(cache);
767 static inline int change_page_attr_set(unsigned long addr, int numpages,
770 return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0);
773 static inline int change_page_attr_clear(unsigned long addr, int numpages,
776 return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0);
779 int _set_memory_uc(unsigned long addr, int numpages)
782 * for now UC MINUS. see comments in ioremap_nocache()
784 return change_page_attr_set(addr, numpages,
785 __pgprot(_PAGE_CACHE_UC_MINUS));
788 int set_memory_uc(unsigned long addr, int numpages)
791 * for now UC MINUS. see comments in ioremap_nocache()
793 if (reserve_memtype(addr, addr + numpages * PAGE_SIZE,
794 _PAGE_CACHE_UC_MINUS, NULL))
797 return _set_memory_uc(addr, numpages);
799 EXPORT_SYMBOL(set_memory_uc);
801 int _set_memory_wc(unsigned long addr, int numpages)
803 return change_page_attr_set(addr, numpages,
804 __pgprot(_PAGE_CACHE_WC));
807 int set_memory_wc(unsigned long addr, int numpages)
810 return set_memory_uc(addr, numpages);
812 if (reserve_memtype(addr, addr + numpages * PAGE_SIZE,
813 _PAGE_CACHE_WC, NULL))
816 return _set_memory_wc(addr, numpages);
818 EXPORT_SYMBOL(set_memory_wc);
820 int _set_memory_wb(unsigned long addr, int numpages)
822 return change_page_attr_clear(addr, numpages,
823 __pgprot(_PAGE_CACHE_MASK));
826 int set_memory_wb(unsigned long addr, int numpages)
828 free_memtype(addr, addr + numpages * PAGE_SIZE);
830 return _set_memory_wb(addr, numpages);
832 EXPORT_SYMBOL(set_memory_wb);
834 int set_memory_x(unsigned long addr, int numpages)
836 return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_NX));
838 EXPORT_SYMBOL(set_memory_x);
840 int set_memory_nx(unsigned long addr, int numpages)
842 return change_page_attr_set(addr, numpages, __pgprot(_PAGE_NX));
844 EXPORT_SYMBOL(set_memory_nx);
846 int set_memory_ro(unsigned long addr, int numpages)
848 return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_RW));
851 int set_memory_rw(unsigned long addr, int numpages)
853 return change_page_attr_set(addr, numpages, __pgprot(_PAGE_RW));
856 int set_memory_np(unsigned long addr, int numpages)
858 return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_PRESENT));
861 int set_memory_4k(unsigned long addr, int numpages)
863 return change_page_attr_set_clr(addr, numpages, __pgprot(0),
867 int set_pages_uc(struct page *page, int numpages)
869 unsigned long addr = (unsigned long)page_address(page);
871 return set_memory_uc(addr, numpages);
873 EXPORT_SYMBOL(set_pages_uc);
875 int set_pages_wb(struct page *page, int numpages)
877 unsigned long addr = (unsigned long)page_address(page);
879 return set_memory_wb(addr, numpages);
881 EXPORT_SYMBOL(set_pages_wb);
883 int set_pages_x(struct page *page, int numpages)
885 unsigned long addr = (unsigned long)page_address(page);
887 return set_memory_x(addr, numpages);
889 EXPORT_SYMBOL(set_pages_x);
891 int set_pages_nx(struct page *page, int numpages)
893 unsigned long addr = (unsigned long)page_address(page);
895 return set_memory_nx(addr, numpages);
897 EXPORT_SYMBOL(set_pages_nx);
899 int set_pages_ro(struct page *page, int numpages)
901 unsigned long addr = (unsigned long)page_address(page);
903 return set_memory_ro(addr, numpages);
906 int set_pages_rw(struct page *page, int numpages)
908 unsigned long addr = (unsigned long)page_address(page);
910 return set_memory_rw(addr, numpages);
913 #ifdef CONFIG_DEBUG_PAGEALLOC
915 static int __set_pages_p(struct page *page, int numpages)
917 struct cpa_data cpa = { .vaddr = (unsigned long) page_address(page),
918 .numpages = numpages,
919 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
920 .mask_clr = __pgprot(0)};
922 return __change_page_attr_set_clr(&cpa, 1);
925 static int __set_pages_np(struct page *page, int numpages)
927 struct cpa_data cpa = { .vaddr = (unsigned long) page_address(page),
928 .numpages = numpages,
929 .mask_set = __pgprot(0),
930 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW)};
932 return __change_page_attr_set_clr(&cpa, 1);
935 void kernel_map_pages(struct page *page, int numpages, int enable)
937 if (PageHighMem(page))
940 debug_check_no_locks_freed(page_address(page),
941 numpages * PAGE_SIZE);
945 * If page allocator is not up yet then do not call c_p_a():
947 if (!debug_pagealloc_enabled)
951 * The return value is ignored as the calls cannot fail.
952 * Large pages are kept enabled at boot time, and are
953 * split up quickly with DEBUG_PAGEALLOC. If a splitup
954 * fails here (due to temporary memory shortage) no damage
955 * is done because we just keep the largepage intact up
956 * to the next attempt when it will likely be split up:
959 __set_pages_p(page, numpages);
961 __set_pages_np(page, numpages);
964 * We should perform an IPI and flush all tlbs,
965 * but that can deadlock->flush only current cpu:
970 * Try to refill the page pool here. We can do this only after
976 #ifdef CONFIG_DEBUG_FS
977 static int dpa_show(struct seq_file *m, void *v)
979 seq_puts(m, "DEBUG_PAGEALLOC\n");
980 seq_printf(m, "pool_size : %lu\n", pool_size);
981 seq_printf(m, "pool_pages : %lu\n", pool_pages);
982 seq_printf(m, "pool_low : %lu\n", pool_low);
983 seq_printf(m, "pool_used : %lu\n", pool_used);
984 seq_printf(m, "pool_failed : %lu\n", pool_failed);
989 static int dpa_open(struct inode *inode, struct file *filp)
991 return single_open(filp, dpa_show, NULL);
994 static const struct file_operations dpa_fops = {
998 .release = single_release,
1001 static int __init debug_pagealloc_proc_init(void)
1005 de = debugfs_create_file("debug_pagealloc", 0600, NULL, NULL,
1012 __initcall(debug_pagealloc_proc_init);
1015 #ifdef CONFIG_HIBERNATION
1017 bool kernel_page_present(struct page *page)
1022 if (PageHighMem(page))
1025 pte = lookup_address((unsigned long)page_address(page), &level);
1026 return (pte_val(*pte) & _PAGE_PRESENT);
1029 #endif /* CONFIG_HIBERNATION */
1031 #endif /* CONFIG_DEBUG_PAGEALLOC */
1034 * The testcases use internal knowledge of the implementation that shouldn't
1035 * be exposed to the rest of the kernel. Include these directly here.
1037 #ifdef CONFIG_CPA_DEBUG
1038 #include "pageattr-test.c"