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:
38 static inline unsigned long highmap_start_pfn(void)
40 return __pa(_text) >> PAGE_SHIFT;
43 static inline unsigned long highmap_end_pfn(void)
45 return __pa(round_up((unsigned long)_end, PMD_SIZE)) >> PAGE_SHIFT;
50 #ifdef CONFIG_DEBUG_PAGEALLOC
51 # define debug_pagealloc 1
53 # define debug_pagealloc 0
57 within(unsigned long addr, unsigned long start, unsigned long end)
59 return addr >= start && addr < end;
67 * clflush_cache_range - flush a cache range with clflush
68 * @addr: virtual start address
69 * @size: number of bytes to flush
71 * clflush is an unordered instruction which needs fencing with mfence
72 * to avoid ordering issues.
74 void clflush_cache_range(void *vaddr, unsigned int size)
76 void *vend = vaddr + size - 1;
80 for (; vaddr < vend; vaddr += boot_cpu_data.x86_clflush_size)
83 * Flush any possible final partial cacheline:
90 static void __cpa_flush_all(void *arg)
92 unsigned long cache = (unsigned long)arg;
95 * Flush all to work around Errata in early athlons regarding
96 * large page flushing.
100 if (cache && boot_cpu_data.x86_model >= 4)
104 static void cpa_flush_all(unsigned long cache)
106 BUG_ON(irqs_disabled());
108 on_each_cpu(__cpa_flush_all, (void *) cache, 1, 1);
111 static void __cpa_flush_range(void *arg)
114 * We could optimize that further and do individual per page
115 * tlb invalidates for a low number of pages. Caveat: we must
116 * flush the high aliases on 64bit as well.
121 static void cpa_flush_range(unsigned long start, int numpages, int cache)
123 unsigned int i, level;
126 BUG_ON(irqs_disabled());
127 WARN_ON(PAGE_ALIGN(start) != start);
129 on_each_cpu(__cpa_flush_range, NULL, 1, 1);
135 * We only need to flush on one CPU,
136 * clflush is a MESI-coherent instruction that
137 * will cause all other CPUs to flush the same
140 for (i = 0, addr = start; i < numpages; i++, addr += PAGE_SIZE) {
141 pte_t *pte = lookup_address(addr, &level);
144 * Only flush present addresses:
146 if (pte && (pte_val(*pte) & _PAGE_PRESENT))
147 clflush_cache_range((void *) addr, PAGE_SIZE);
152 * Certain areas of memory on x86 require very specific protection flags,
153 * for example the BIOS area or kernel text. Callers don't always get this
154 * right (again, ioremap() on BIOS memory is not uncommon) so this function
155 * checks and fixes these known static required protection bits.
157 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address,
160 pgprot_t forbidden = __pgprot(0);
163 * The BIOS area between 640k and 1Mb needs to be executable for
164 * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
166 if (within(pfn, BIOS_BEGIN >> PAGE_SHIFT, BIOS_END >> PAGE_SHIFT))
167 pgprot_val(forbidden) |= _PAGE_NX;
170 * The kernel text needs to be executable for obvious reasons
171 * Does not cover __inittext since that is gone later on. On
172 * 64bit we do not enforce !NX on the low mapping
174 if (within(address, (unsigned long)_text, (unsigned long)_etext))
175 pgprot_val(forbidden) |= _PAGE_NX;
178 * The .rodata section needs to be read-only. Using the pfn
179 * catches all aliases.
181 if (within(pfn, __pa((unsigned long)__start_rodata) >> PAGE_SHIFT,
182 __pa((unsigned long)__end_rodata) >> PAGE_SHIFT))
183 pgprot_val(forbidden) |= _PAGE_RW;
185 prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
191 * Lookup the page table entry for a virtual address. Return a pointer
192 * to the entry and the level of the mapping.
194 * Note: We return pud and pmd either when the entry is marked large
195 * or when the present bit is not set. Otherwise we would return a
196 * pointer to a nonexisting mapping.
198 pte_t *lookup_address(unsigned long address, unsigned int *level)
200 pgd_t *pgd = pgd_offset_k(address);
204 *level = PG_LEVEL_NONE;
209 pud = pud_offset(pgd, address);
213 *level = PG_LEVEL_1G;
214 if (pud_large(*pud) || !pud_present(*pud))
217 pmd = pmd_offset(pud, address);
221 *level = PG_LEVEL_2M;
222 if (pmd_large(*pmd) || !pmd_present(*pmd))
225 *level = PG_LEVEL_4K;
227 return pte_offset_kernel(pmd, address);
231 * Set the new pmd in all the pgds we know about:
233 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
236 set_pte_atomic(kpte, pte);
238 if (!SHARED_KERNEL_PMD) {
241 list_for_each_entry(page, &pgd_list, lru) {
246 pgd = (pgd_t *)page_address(page) + pgd_index(address);
247 pud = pud_offset(pgd, address);
248 pmd = pmd_offset(pud, address);
249 set_pte_atomic((pte_t *)pmd, pte);
256 try_preserve_large_page(pte_t *kpte, unsigned long address,
257 struct cpa_data *cpa)
259 unsigned long nextpage_addr, numpages, pmask, psize, flags, addr, pfn;
260 pte_t new_pte, old_pte, *tmp;
261 pgprot_t old_prot, new_prot;
265 spin_lock_irqsave(&pgd_lock, flags);
267 * Check for races, another CPU might have split this page
270 tmp = lookup_address(address, &level);
276 psize = PMD_PAGE_SIZE;
277 pmask = PMD_PAGE_MASK;
281 psize = PUD_PAGE_SIZE;
282 pmask = PUD_PAGE_MASK;
291 * Calculate the number of pages, which fit into this large
292 * page starting at address:
294 nextpage_addr = (address + psize) & pmask;
295 numpages = (nextpage_addr - address) >> PAGE_SHIFT;
296 if (numpages < cpa->numpages)
297 cpa->numpages = numpages;
300 * We are safe now. Check whether the new pgprot is the same:
303 old_prot = new_prot = pte_pgprot(old_pte);
305 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
306 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
309 * old_pte points to the large page base address. So we need
310 * to add the offset of the virtual address:
312 pfn = pte_pfn(old_pte) + ((address & (psize - 1)) >> PAGE_SHIFT);
315 new_prot = static_protections(new_prot, address, pfn);
318 * We need to check the full range, whether
319 * static_protection() requires a different pgprot for one of
320 * the pages in the range we try to preserve:
322 addr = address + PAGE_SIZE;
324 for (i = 1; i < cpa->numpages; i++, addr += PAGE_SIZE, pfn++) {
325 pgprot_t chk_prot = static_protections(new_prot, addr, pfn);
327 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
332 * If there are no changes, return. maxpages has been updated
335 if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
341 * We need to change the attributes. Check, whether we can
342 * change the large page in one go. We request a split, when
343 * the address is not aligned and the number of pages is
344 * smaller than the number of pages in the large page. Note
345 * that we limited the number of possible pages already to
346 * the number of pages in the large page.
348 if (address == (nextpage_addr - psize) && cpa->numpages == numpages) {
350 * The address is aligned and the number of pages
351 * covers the full page.
353 new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
354 __set_pmd_pte(kpte, address, new_pte);
360 spin_unlock_irqrestore(&pgd_lock, flags);
365 static LIST_HEAD(page_pool);
366 static unsigned long pool_size, pool_pages, pool_low;
367 static unsigned long pool_used, pool_failed;
369 static void cpa_fill_pool(struct page **ret)
371 gfp_t gfp = GFP_KERNEL;
376 * Avoid recursion (on debug-pagealloc) and also signal
377 * our priority to get to these pagetables:
379 if (current->flags & PF_MEMALLOC)
381 current->flags |= PF_MEMALLOC;
384 * Allocate atomically from atomic contexts:
386 if (in_atomic() || irqs_disabled() || debug_pagealloc)
387 gfp = GFP_ATOMIC | __GFP_NORETRY | __GFP_NOWARN;
389 while (pool_pages < pool_size || (ret && !*ret)) {
390 p = alloc_pages(gfp, 0);
396 * If the call site needs a page right now, provide it:
402 spin_lock_irqsave(&pgd_lock, flags);
403 list_add(&p->lru, &page_pool);
405 spin_unlock_irqrestore(&pgd_lock, flags);
408 current->flags &= ~PF_MEMALLOC;
411 #define SHIFT_MB (20 - PAGE_SHIFT)
412 #define ROUND_MB_GB ((1 << 10) - 1)
413 #define SHIFT_MB_GB 10
414 #define POOL_PAGES_PER_GB 16
416 void __init cpa_init(void)
423 * Calculate the number of pool pages:
425 * Convert totalram (nr of pages) to MiB and round to the next
426 * GiB. Shift MiB to Gib and multiply the result by
429 if (debug_pagealloc) {
430 gb = ((si.totalram >> SHIFT_MB) + ROUND_MB_GB) >> SHIFT_MB_GB;
431 pool_size = POOL_PAGES_PER_GB * gb;
435 pool_low = pool_size;
439 "CPA: page pool initialized %lu of %lu pages preallocated\n",
440 pool_pages, pool_size);
443 static int split_large_page(pte_t *kpte, unsigned long address)
445 unsigned long flags, pfn, pfninc = 1;
446 unsigned int i, level;
452 * Get a page from the pool. The pool list is protected by the
453 * pgd_lock, which we have to take anyway for the split
456 spin_lock_irqsave(&pgd_lock, flags);
457 if (list_empty(&page_pool)) {
458 spin_unlock_irqrestore(&pgd_lock, flags);
460 cpa_fill_pool(&base);
463 spin_lock_irqsave(&pgd_lock, flags);
465 base = list_first_entry(&page_pool, struct page, lru);
466 list_del(&base->lru);
469 if (pool_pages < pool_low)
470 pool_low = pool_pages;
474 * Check for races, another CPU might have split this page
477 tmp = lookup_address(address, &level);
481 pbase = (pte_t *)page_address(base);
483 paravirt_alloc_pt(&init_mm, page_to_pfn(base));
485 ref_prot = pte_pgprot(pte_clrhuge(*kpte));
488 if (level == PG_LEVEL_1G) {
489 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
490 pgprot_val(ref_prot) |= _PAGE_PSE;
495 * Get the target pfn from the original entry:
497 pfn = pte_pfn(*kpte);
498 for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
499 set_pte(&pbase[i], pfn_pte(pfn, ref_prot));
502 * Install the new, split up pagetable. Important details here:
504 * On Intel the NX bit of all levels must be cleared to make a
505 * page executable. See section 4.13.2 of Intel 64 and IA-32
506 * Architectures Software Developer's Manual).
508 * Mark the entry present. The current mapping might be
509 * set to not present, which we preserved above.
511 ref_prot = pte_pgprot(pte_mkexec(pte_clrhuge(*kpte)));
512 pgprot_val(ref_prot) |= _PAGE_PRESENT;
513 __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
518 * If we dropped out via the lookup_address check under
519 * pgd_lock then stick the page back into the pool:
522 list_add(&base->lru, &page_pool);
526 spin_unlock_irqrestore(&pgd_lock, flags);
531 static int __change_page_attr(struct cpa_data *cpa, int primary)
533 unsigned long address = cpa->vaddr;
536 pte_t *kpte, old_pte;
539 kpte = lookup_address(address, &level);
541 return primary ? -EINVAL : 0;
544 if (!pte_val(old_pte)) {
547 printk(KERN_WARNING "CPA: called for zero pte. "
548 "vaddr = %lx cpa->vaddr = %lx\n", address,
554 if (level == PG_LEVEL_4K) {
556 pgprot_t new_prot = pte_pgprot(old_pte);
557 unsigned long pfn = pte_pfn(old_pte);
559 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
560 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
562 new_prot = static_protections(new_prot, address, pfn);
565 * We need to keep the pfn from the existing PTE,
566 * after all we're only going to change it's attributes
567 * not the memory it points to
569 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
572 * Do we really change anything ?
574 if (pte_val(old_pte) != pte_val(new_pte)) {
575 set_pte_atomic(kpte, new_pte);
583 * Check, whether we can keep the large page intact
584 * and just change the pte:
586 do_split = try_preserve_large_page(kpte, address, cpa);
588 * When the range fits into the existing large page,
589 * return. cp->numpages and cpa->tlbflush have been updated in
596 * We have to split the large page:
598 err = split_large_page(kpte, address);
607 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
609 static int cpa_process_alias(struct cpa_data *cpa)
611 struct cpa_data alias_cpa;
614 if (cpa->pfn > max_pfn_mapped)
618 * No need to redo, when the primary call touched the direct
621 if (!within(cpa->vaddr, PAGE_OFFSET,
622 PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
625 alias_cpa.vaddr = (unsigned long) __va(cpa->pfn << PAGE_SHIFT);
627 ret = __change_page_attr_set_clr(&alias_cpa, 0);
634 * No need to redo, when the primary call touched the high
637 if (within(cpa->vaddr, (unsigned long) _text, (unsigned long) _end))
641 * If the physical address is inside the kernel map, we need
642 * to touch the high mapped kernel as well:
644 if (!within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn()))
649 (cpa->pfn << PAGE_SHIFT) + __START_KERNEL_map - phys_base;
652 * The high mapping range is imprecise, so ignore the return value.
654 __change_page_attr_set_clr(&alias_cpa, 0);
659 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
661 int ret, numpages = cpa->numpages;
665 * Store the remaining nr of pages for the large page
666 * preservation check.
668 cpa->numpages = numpages;
670 ret = __change_page_attr(cpa, checkalias);
675 ret = cpa_process_alias(cpa);
681 * Adjust the number of pages with the result of the
682 * CPA operation. Either a large page has been
683 * preserved or a single page update happened.
685 BUG_ON(cpa->numpages > numpages);
686 numpages -= cpa->numpages;
687 cpa->vaddr += cpa->numpages * PAGE_SIZE;
692 static inline int cache_attr(pgprot_t attr)
694 return pgprot_val(attr) &
695 (_PAGE_PAT | _PAGE_PAT_LARGE | _PAGE_PWT | _PAGE_PCD);
698 static int change_page_attr_set_clr(unsigned long addr, int numpages,
699 pgprot_t mask_set, pgprot_t mask_clr)
702 int ret, cache, checkalias;
705 * Check, if we are requested to change a not supported
708 mask_set = canon_pgprot(mask_set);
709 mask_clr = canon_pgprot(mask_clr);
710 if (!pgprot_val(mask_set) && !pgprot_val(mask_clr))
713 /* Ensure we are PAGE_SIZE aligned */
714 if (addr & ~PAGE_MASK) {
717 * People should not be passing in unaligned addresses:
723 cpa.numpages = numpages;
724 cpa.mask_set = mask_set;
725 cpa.mask_clr = mask_clr;
728 /* No alias checking for _NX bit modifications */
729 checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
731 ret = __change_page_attr_set_clr(&cpa, checkalias);
734 * Check whether we really changed something:
740 * No need to flush, when we did not set any of the caching
743 cache = cache_attr(mask_set);
746 * On success we use clflush, when the CPU supports it to
747 * avoid the wbindv. If the CPU does not support it and in the
748 * error case we fall back to cpa_flush_all (which uses
751 if (!ret && cpu_has_clflush)
752 cpa_flush_range(addr, numpages, cache);
754 cpa_flush_all(cache);
762 static inline int change_page_attr_set(unsigned long addr, int numpages,
765 return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0));
768 static inline int change_page_attr_clear(unsigned long addr, int numpages,
771 return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask);
774 int _set_memory_uc(unsigned long addr, int numpages)
776 return change_page_attr_set(addr, numpages,
777 __pgprot(_PAGE_CACHE_UC));
780 int set_memory_uc(unsigned long addr, int numpages)
782 if (reserve_memtype(addr, addr + numpages * PAGE_SIZE,
783 _PAGE_CACHE_UC, NULL))
786 return _set_memory_uc(addr, numpages);
788 EXPORT_SYMBOL(set_memory_uc);
790 int _set_memory_wc(unsigned long addr, int numpages)
792 return change_page_attr_set(addr, numpages,
793 __pgprot(_PAGE_CACHE_WC));
796 int set_memory_wc(unsigned long addr, int numpages)
799 return set_memory_uc(addr, numpages);
801 if (reserve_memtype(addr, addr + numpages * PAGE_SIZE,
802 _PAGE_CACHE_WC, NULL))
805 return _set_memory_wc(addr, numpages);
807 EXPORT_SYMBOL(set_memory_wc);
809 int _set_memory_wb(unsigned long addr, int numpages)
811 return change_page_attr_clear(addr, numpages,
812 __pgprot(_PAGE_CACHE_MASK));
815 int set_memory_wb(unsigned long addr, int numpages)
817 free_memtype(addr, addr + numpages * PAGE_SIZE);
819 return _set_memory_wb(addr, numpages);
821 EXPORT_SYMBOL(set_memory_wb);
823 int set_memory_x(unsigned long addr, int numpages)
825 return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_NX));
827 EXPORT_SYMBOL(set_memory_x);
829 int set_memory_nx(unsigned long addr, int numpages)
831 return change_page_attr_set(addr, numpages, __pgprot(_PAGE_NX));
833 EXPORT_SYMBOL(set_memory_nx);
835 int set_memory_ro(unsigned long addr, int numpages)
837 return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_RW));
840 int set_memory_rw(unsigned long addr, int numpages)
842 return change_page_attr_set(addr, numpages, __pgprot(_PAGE_RW));
845 int set_memory_np(unsigned long addr, int numpages)
847 return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_PRESENT));
850 int set_pages_uc(struct page *page, int numpages)
852 unsigned long addr = (unsigned long)page_address(page);
854 return set_memory_uc(addr, numpages);
856 EXPORT_SYMBOL(set_pages_uc);
858 int set_pages_wb(struct page *page, int numpages)
860 unsigned long addr = (unsigned long)page_address(page);
862 return set_memory_wb(addr, numpages);
864 EXPORT_SYMBOL(set_pages_wb);
866 int set_pages_x(struct page *page, int numpages)
868 unsigned long addr = (unsigned long)page_address(page);
870 return set_memory_x(addr, numpages);
872 EXPORT_SYMBOL(set_pages_x);
874 int set_pages_nx(struct page *page, int numpages)
876 unsigned long addr = (unsigned long)page_address(page);
878 return set_memory_nx(addr, numpages);
880 EXPORT_SYMBOL(set_pages_nx);
882 int set_pages_ro(struct page *page, int numpages)
884 unsigned long addr = (unsigned long)page_address(page);
886 return set_memory_ro(addr, numpages);
889 int set_pages_rw(struct page *page, int numpages)
891 unsigned long addr = (unsigned long)page_address(page);
893 return set_memory_rw(addr, numpages);
896 #ifdef CONFIG_DEBUG_PAGEALLOC
898 static int __set_pages_p(struct page *page, int numpages)
900 struct cpa_data cpa = { .vaddr = (unsigned long) page_address(page),
901 .numpages = numpages,
902 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
903 .mask_clr = __pgprot(0)};
905 return __change_page_attr_set_clr(&cpa, 1);
908 static int __set_pages_np(struct page *page, int numpages)
910 struct cpa_data cpa = { .vaddr = (unsigned long) page_address(page),
911 .numpages = numpages,
912 .mask_set = __pgprot(0),
913 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW)};
915 return __change_page_attr_set_clr(&cpa, 1);
918 void kernel_map_pages(struct page *page, int numpages, int enable)
920 if (PageHighMem(page))
923 debug_check_no_locks_freed(page_address(page),
924 numpages * PAGE_SIZE);
928 * If page allocator is not up yet then do not call c_p_a():
930 if (!debug_pagealloc_enabled)
934 * The return value is ignored as the calls cannot fail.
935 * Large pages are kept enabled at boot time, and are
936 * split up quickly with DEBUG_PAGEALLOC. If a splitup
937 * fails here (due to temporary memory shortage) no damage
938 * is done because we just keep the largepage intact up
939 * to the next attempt when it will likely be split up:
942 __set_pages_p(page, numpages);
944 __set_pages_np(page, numpages);
947 * We should perform an IPI and flush all tlbs,
948 * but that can deadlock->flush only current cpu:
953 * Try to refill the page pool here. We can do this only after
959 #ifdef CONFIG_DEBUG_FS
960 static int dpa_show(struct seq_file *m, void *v)
962 seq_puts(m, "DEBUG_PAGEALLOC\n");
963 seq_printf(m, "pool_size : %lu\n", pool_size);
964 seq_printf(m, "pool_pages : %lu\n", pool_pages);
965 seq_printf(m, "pool_low : %lu\n", pool_low);
966 seq_printf(m, "pool_used : %lu\n", pool_used);
967 seq_printf(m, "pool_failed : %lu\n", pool_failed);
972 static int dpa_open(struct inode *inode, struct file *filp)
974 return single_open(filp, dpa_show, NULL);
977 static const struct file_operations dpa_fops = {
981 .release = single_release,
984 int __init debug_pagealloc_proc_init(void)
988 de = debugfs_create_file("debug_pagealloc", 0600, NULL, NULL,
995 __initcall(debug_pagealloc_proc_init);
998 #ifdef CONFIG_HIBERNATION
1000 bool kernel_page_present(struct page *page)
1005 if (PageHighMem(page))
1008 pte = lookup_address((unsigned long)page_address(page), &level);
1009 return (pte_val(*pte) & _PAGE_PRESENT);
1012 #endif /* CONFIG_HIBERNATION */
1014 #endif /* CONFIG_DEBUG_PAGEALLOC */
1017 * The testcases use internal knowledge of the implementation that shouldn't
1018 * be exposed to the rest of the kernel. Include these directly here.
1020 #ifdef CONFIG_CPA_DEBUG
1021 #include "pageattr-test.c"