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);
232 * Set the new pmd in all the pgds we know about:
234 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
237 set_pte_atomic(kpte, pte);
239 if (!SHARED_KERNEL_PMD) {
242 list_for_each_entry(page, &pgd_list, lru) {
247 pgd = (pgd_t *)page_address(page) + pgd_index(address);
248 pud = pud_offset(pgd, address);
249 pmd = pmd_offset(pud, address);
250 set_pte_atomic((pte_t *)pmd, pte);
257 try_preserve_large_page(pte_t *kpte, unsigned long address,
258 struct cpa_data *cpa)
260 unsigned long nextpage_addr, numpages, pmask, psize, flags, addr, pfn;
261 pte_t new_pte, old_pte, *tmp;
262 pgprot_t old_prot, new_prot;
266 if (cpa->force_split)
269 spin_lock_irqsave(&pgd_lock, flags);
271 * Check for races, another CPU might have split this page
274 tmp = lookup_address(address, &level);
280 psize = PMD_PAGE_SIZE;
281 pmask = PMD_PAGE_MASK;
285 psize = PUD_PAGE_SIZE;
286 pmask = PUD_PAGE_MASK;
295 * Calculate the number of pages, which fit into this large
296 * page starting at address:
298 nextpage_addr = (address + psize) & pmask;
299 numpages = (nextpage_addr - address) >> PAGE_SHIFT;
300 if (numpages < cpa->numpages)
301 cpa->numpages = numpages;
304 * We are safe now. Check whether the new pgprot is the same:
307 old_prot = new_prot = pte_pgprot(old_pte);
309 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
310 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
313 * old_pte points to the large page base address. So we need
314 * to add the offset of the virtual address:
316 pfn = pte_pfn(old_pte) + ((address & (psize - 1)) >> PAGE_SHIFT);
319 new_prot = static_protections(new_prot, address, pfn);
322 * We need to check the full range, whether
323 * static_protection() requires a different pgprot for one of
324 * the pages in the range we try to preserve:
326 addr = address + PAGE_SIZE;
328 for (i = 1; i < cpa->numpages; i++, addr += PAGE_SIZE, pfn++) {
329 pgprot_t chk_prot = static_protections(new_prot, addr, pfn);
331 if (pgprot_val(chk_prot) != pgprot_val(new_prot))
336 * If there are no changes, return. maxpages has been updated
339 if (pgprot_val(new_prot) == pgprot_val(old_prot)) {
345 * We need to change the attributes. Check, whether we can
346 * change the large page in one go. We request a split, when
347 * the address is not aligned and the number of pages is
348 * smaller than the number of pages in the large page. Note
349 * that we limited the number of possible pages already to
350 * the number of pages in the large page.
352 if (address == (nextpage_addr - psize) && cpa->numpages == numpages) {
354 * The address is aligned and the number of pages
355 * covers the full page.
357 new_pte = pfn_pte(pte_pfn(old_pte), canon_pgprot(new_prot));
358 __set_pmd_pte(kpte, address, new_pte);
364 spin_unlock_irqrestore(&pgd_lock, flags);
369 static LIST_HEAD(page_pool);
370 static unsigned long pool_size, pool_pages, pool_low;
371 static unsigned long pool_used, pool_failed;
373 static void cpa_fill_pool(struct page **ret)
375 gfp_t gfp = GFP_KERNEL;
380 * Avoid recursion (on debug-pagealloc) and also signal
381 * our priority to get to these pagetables:
383 if (current->flags & PF_MEMALLOC)
385 current->flags |= PF_MEMALLOC;
388 * Allocate atomically from atomic contexts:
390 if (in_atomic() || irqs_disabled() || debug_pagealloc)
391 gfp = GFP_ATOMIC | __GFP_NORETRY | __GFP_NOWARN;
393 while (pool_pages < pool_size || (ret && !*ret)) {
394 p = alloc_pages(gfp, 0);
400 * If the call site needs a page right now, provide it:
406 spin_lock_irqsave(&pgd_lock, flags);
407 list_add(&p->lru, &page_pool);
409 spin_unlock_irqrestore(&pgd_lock, flags);
412 current->flags &= ~PF_MEMALLOC;
415 #define SHIFT_MB (20 - PAGE_SHIFT)
416 #define ROUND_MB_GB ((1 << 10) - 1)
417 #define SHIFT_MB_GB 10
418 #define POOL_PAGES_PER_GB 16
420 void __init cpa_init(void)
427 * Calculate the number of pool pages:
429 * Convert totalram (nr of pages) to MiB and round to the next
430 * GiB. Shift MiB to Gib and multiply the result by
433 if (debug_pagealloc) {
434 gb = ((si.totalram >> SHIFT_MB) + ROUND_MB_GB) >> SHIFT_MB_GB;
435 pool_size = POOL_PAGES_PER_GB * gb;
439 pool_low = pool_size;
443 "CPA: page pool initialized %lu of %lu pages preallocated\n",
444 pool_pages, pool_size);
447 static int split_large_page(pte_t *kpte, unsigned long address)
449 unsigned long flags, pfn, pfninc = 1;
450 unsigned int i, level;
456 * Get a page from the pool. The pool list is protected by the
457 * pgd_lock, which we have to take anyway for the split
460 spin_lock_irqsave(&pgd_lock, flags);
461 if (list_empty(&page_pool)) {
462 spin_unlock_irqrestore(&pgd_lock, flags);
464 cpa_fill_pool(&base);
467 spin_lock_irqsave(&pgd_lock, flags);
469 base = list_first_entry(&page_pool, struct page, lru);
470 list_del(&base->lru);
473 if (pool_pages < pool_low)
474 pool_low = pool_pages;
478 * Check for races, another CPU might have split this page
481 tmp = lookup_address(address, &level);
485 pbase = (pte_t *)page_address(base);
487 paravirt_alloc_pt(&init_mm, page_to_pfn(base));
489 ref_prot = pte_pgprot(pte_clrhuge(*kpte));
492 if (level == PG_LEVEL_1G) {
493 pfninc = PMD_PAGE_SIZE >> PAGE_SHIFT;
494 pgprot_val(ref_prot) |= _PAGE_PSE;
499 * Get the target pfn from the original entry:
501 pfn = pte_pfn(*kpte);
502 for (i = 0; i < PTRS_PER_PTE; i++, pfn += pfninc)
503 set_pte(&pbase[i], pfn_pte(pfn, ref_prot));
506 * Install the new, split up pagetable. Important details here:
508 * On Intel the NX bit of all levels must be cleared to make a
509 * page executable. See section 4.13.2 of Intel 64 and IA-32
510 * Architectures Software Developer's Manual).
512 * Mark the entry present. The current mapping might be
513 * set to not present, which we preserved above.
515 ref_prot = pte_pgprot(pte_mkexec(pte_clrhuge(*kpte)));
516 pgprot_val(ref_prot) |= _PAGE_PRESENT;
517 __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
522 * If we dropped out via the lookup_address check under
523 * pgd_lock then stick the page back into the pool:
526 list_add(&base->lru, &page_pool);
530 spin_unlock_irqrestore(&pgd_lock, flags);
535 static int __change_page_attr(struct cpa_data *cpa, int primary)
537 unsigned long address = cpa->vaddr;
540 pte_t *kpte, old_pte;
543 kpte = lookup_address(address, &level);
545 return primary ? -EINVAL : 0;
548 if (!pte_val(old_pte)) {
551 printk(KERN_WARNING "CPA: called for zero pte. "
552 "vaddr = %lx cpa->vaddr = %lx\n", address,
558 if (level == PG_LEVEL_4K) {
560 pgprot_t new_prot = pte_pgprot(old_pte);
561 unsigned long pfn = pte_pfn(old_pte);
563 pgprot_val(new_prot) &= ~pgprot_val(cpa->mask_clr);
564 pgprot_val(new_prot) |= pgprot_val(cpa->mask_set);
566 new_prot = static_protections(new_prot, address, pfn);
569 * We need to keep the pfn from the existing PTE,
570 * after all we're only going to change it's attributes
571 * not the memory it points to
573 new_pte = pfn_pte(pfn, canon_pgprot(new_prot));
576 * Do we really change anything ?
578 if (pte_val(old_pte) != pte_val(new_pte)) {
579 set_pte_atomic(kpte, new_pte);
587 * Check, whether we can keep the large page intact
588 * and just change the pte:
590 do_split = try_preserve_large_page(kpte, address, cpa);
592 * When the range fits into the existing large page,
593 * return. cp->numpages and cpa->tlbflush have been updated in
600 * We have to split the large page:
602 err = split_large_page(kpte, address);
611 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias);
613 static int cpa_process_alias(struct cpa_data *cpa)
615 struct cpa_data alias_cpa;
618 if (cpa->pfn > max_pfn_mapped)
622 * No need to redo, when the primary call touched the direct
625 if (!within(cpa->vaddr, PAGE_OFFSET,
626 PAGE_OFFSET + (max_pfn_mapped << PAGE_SHIFT))) {
629 alias_cpa.vaddr = (unsigned long) __va(cpa->pfn << PAGE_SHIFT);
631 ret = __change_page_attr_set_clr(&alias_cpa, 0);
638 * No need to redo, when the primary call touched the high
641 if (within(cpa->vaddr, (unsigned long) _text, (unsigned long) _end))
645 * If the physical address is inside the kernel map, we need
646 * to touch the high mapped kernel as well:
648 if (!within(cpa->pfn, highmap_start_pfn(), highmap_end_pfn()))
653 (cpa->pfn << PAGE_SHIFT) + __START_KERNEL_map - phys_base;
656 * The high mapping range is imprecise, so ignore the return value.
658 __change_page_attr_set_clr(&alias_cpa, 0);
663 static int __change_page_attr_set_clr(struct cpa_data *cpa, int checkalias)
665 int ret, numpages = cpa->numpages;
669 * Store the remaining nr of pages for the large page
670 * preservation check.
672 cpa->numpages = numpages;
674 ret = __change_page_attr(cpa, checkalias);
679 ret = cpa_process_alias(cpa);
685 * Adjust the number of pages with the result of the
686 * CPA operation. Either a large page has been
687 * preserved or a single page update happened.
689 BUG_ON(cpa->numpages > numpages);
690 numpages -= cpa->numpages;
691 cpa->vaddr += cpa->numpages * PAGE_SIZE;
696 static inline int cache_attr(pgprot_t attr)
698 return pgprot_val(attr) &
699 (_PAGE_PAT | _PAGE_PAT_LARGE | _PAGE_PWT | _PAGE_PCD);
702 static int change_page_attr_set_clr(unsigned long addr, int numpages,
703 pgprot_t mask_set, pgprot_t mask_clr,
707 int ret, cache, checkalias;
710 * Check, if we are requested to change a not supported
713 mask_set = canon_pgprot(mask_set);
714 mask_clr = canon_pgprot(mask_clr);
715 if (!pgprot_val(mask_set) && !pgprot_val(mask_clr) && !force_split)
718 /* Ensure we are PAGE_SIZE aligned */
719 if (addr & ~PAGE_MASK) {
722 * People should not be passing in unaligned addresses:
728 cpa.numpages = numpages;
729 cpa.mask_set = mask_set;
730 cpa.mask_clr = mask_clr;
732 cpa.force_split = force_split;
734 /* No alias checking for _NX bit modifications */
735 checkalias = (pgprot_val(mask_set) | pgprot_val(mask_clr)) != _PAGE_NX;
737 ret = __change_page_attr_set_clr(&cpa, checkalias);
740 * Check whether we really changed something:
746 * No need to flush, when we did not set any of the caching
749 cache = cache_attr(mask_set);
752 * On success we use clflush, when the CPU supports it to
753 * avoid the wbindv. If the CPU does not support it and in the
754 * error case we fall back to cpa_flush_all (which uses
757 if (!ret && cpu_has_clflush)
758 cpa_flush_range(addr, numpages, cache);
760 cpa_flush_all(cache);
768 static inline int change_page_attr_set(unsigned long addr, int numpages,
771 return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0), 0);
774 static inline int change_page_attr_clear(unsigned long addr, int numpages,
777 return change_page_attr_set_clr(addr, numpages, __pgprot(0), mask, 0);
780 int _set_memory_uc(unsigned long addr, int numpages)
782 return change_page_attr_set(addr, numpages,
783 __pgprot(_PAGE_CACHE_UC));
786 int set_memory_uc(unsigned long addr, int numpages)
788 if (reserve_memtype(addr, addr + numpages * PAGE_SIZE,
789 _PAGE_CACHE_UC, NULL))
792 return _set_memory_uc(addr, numpages);
794 EXPORT_SYMBOL(set_memory_uc);
796 int _set_memory_wc(unsigned long addr, int numpages)
798 return change_page_attr_set(addr, numpages,
799 __pgprot(_PAGE_CACHE_WC));
802 int set_memory_wc(unsigned long addr, int numpages)
805 return set_memory_uc(addr, numpages);
807 if (reserve_memtype(addr, addr + numpages * PAGE_SIZE,
808 _PAGE_CACHE_WC, NULL))
811 return _set_memory_wc(addr, numpages);
813 EXPORT_SYMBOL(set_memory_wc);
815 int _set_memory_wb(unsigned long addr, int numpages)
817 return change_page_attr_clear(addr, numpages,
818 __pgprot(_PAGE_CACHE_MASK));
821 int set_memory_wb(unsigned long addr, int numpages)
823 free_memtype(addr, addr + numpages * PAGE_SIZE);
825 return _set_memory_wb(addr, numpages);
827 EXPORT_SYMBOL(set_memory_wb);
829 int set_memory_x(unsigned long addr, int numpages)
831 return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_NX));
833 EXPORT_SYMBOL(set_memory_x);
835 int set_memory_nx(unsigned long addr, int numpages)
837 return change_page_attr_set(addr, numpages, __pgprot(_PAGE_NX));
839 EXPORT_SYMBOL(set_memory_nx);
841 int set_memory_ro(unsigned long addr, int numpages)
843 return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_RW));
846 int set_memory_rw(unsigned long addr, int numpages)
848 return change_page_attr_set(addr, numpages, __pgprot(_PAGE_RW));
851 int set_memory_np(unsigned long addr, int numpages)
853 return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_PRESENT));
856 int set_memory_4k(unsigned long addr, int numpages)
858 return change_page_attr_set_clr(addr, numpages, __pgprot(0),
862 int set_pages_uc(struct page *page, int numpages)
864 unsigned long addr = (unsigned long)page_address(page);
866 return set_memory_uc(addr, numpages);
868 EXPORT_SYMBOL(set_pages_uc);
870 int set_pages_wb(struct page *page, int numpages)
872 unsigned long addr = (unsigned long)page_address(page);
874 return set_memory_wb(addr, numpages);
876 EXPORT_SYMBOL(set_pages_wb);
878 int set_pages_x(struct page *page, int numpages)
880 unsigned long addr = (unsigned long)page_address(page);
882 return set_memory_x(addr, numpages);
884 EXPORT_SYMBOL(set_pages_x);
886 int set_pages_nx(struct page *page, int numpages)
888 unsigned long addr = (unsigned long)page_address(page);
890 return set_memory_nx(addr, numpages);
892 EXPORT_SYMBOL(set_pages_nx);
894 int set_pages_ro(struct page *page, int numpages)
896 unsigned long addr = (unsigned long)page_address(page);
898 return set_memory_ro(addr, numpages);
901 int set_pages_rw(struct page *page, int numpages)
903 unsigned long addr = (unsigned long)page_address(page);
905 return set_memory_rw(addr, numpages);
908 #ifdef CONFIG_DEBUG_PAGEALLOC
910 static int __set_pages_p(struct page *page, int numpages)
912 struct cpa_data cpa = { .vaddr = (unsigned long) page_address(page),
913 .numpages = numpages,
914 .mask_set = __pgprot(_PAGE_PRESENT | _PAGE_RW),
915 .mask_clr = __pgprot(0)};
917 return __change_page_attr_set_clr(&cpa, 1);
920 static int __set_pages_np(struct page *page, int numpages)
922 struct cpa_data cpa = { .vaddr = (unsigned long) page_address(page),
923 .numpages = numpages,
924 .mask_set = __pgprot(0),
925 .mask_clr = __pgprot(_PAGE_PRESENT | _PAGE_RW)};
927 return __change_page_attr_set_clr(&cpa, 1);
930 void kernel_map_pages(struct page *page, int numpages, int enable)
932 if (PageHighMem(page))
935 debug_check_no_locks_freed(page_address(page),
936 numpages * PAGE_SIZE);
940 * If page allocator is not up yet then do not call c_p_a():
942 if (!debug_pagealloc_enabled)
946 * The return value is ignored as the calls cannot fail.
947 * Large pages are kept enabled at boot time, and are
948 * split up quickly with DEBUG_PAGEALLOC. If a splitup
949 * fails here (due to temporary memory shortage) no damage
950 * is done because we just keep the largepage intact up
951 * to the next attempt when it will likely be split up:
954 __set_pages_p(page, numpages);
956 __set_pages_np(page, numpages);
959 * We should perform an IPI and flush all tlbs,
960 * but that can deadlock->flush only current cpu:
965 * Try to refill the page pool here. We can do this only after
971 #ifdef CONFIG_DEBUG_FS
972 static int dpa_show(struct seq_file *m, void *v)
974 seq_puts(m, "DEBUG_PAGEALLOC\n");
975 seq_printf(m, "pool_size : %lu\n", pool_size);
976 seq_printf(m, "pool_pages : %lu\n", pool_pages);
977 seq_printf(m, "pool_low : %lu\n", pool_low);
978 seq_printf(m, "pool_used : %lu\n", pool_used);
979 seq_printf(m, "pool_failed : %lu\n", pool_failed);
984 static int dpa_open(struct inode *inode, struct file *filp)
986 return single_open(filp, dpa_show, NULL);
989 static const struct file_operations dpa_fops = {
993 .release = single_release,
996 int __init debug_pagealloc_proc_init(void)
1000 de = debugfs_create_file("debug_pagealloc", 0600, NULL, NULL,
1007 __initcall(debug_pagealloc_proc_init);
1010 #ifdef CONFIG_HIBERNATION
1012 bool kernel_page_present(struct page *page)
1017 if (PageHighMem(page))
1020 pte = lookup_address((unsigned long)page_address(page), &level);
1021 return (pte_val(*pte) & _PAGE_PRESENT);
1024 #endif /* CONFIG_HIBERNATION */
1026 #endif /* CONFIG_DEBUG_PAGEALLOC */
1029 * The testcases use internal knowledge of the implementation that shouldn't
1030 * be exposed to the rest of the kernel. Include these directly here.
1032 #ifdef CONFIG_CPA_DEBUG
1033 #include "pageattr-test.c"