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>
13 #include <asm/processor.h>
14 #include <asm/tlbflush.h>
15 #include <asm/sections.h>
16 #include <asm/uaccess.h>
17 #include <asm/pgalloc.h>
20 within(unsigned long addr, unsigned long start, unsigned long end)
22 return addr >= start && addr < end;
28 void clflush_cache_range(void *addr, int size)
32 for (i = 0; i < size; i += boot_cpu_data.x86_clflush_size)
36 static void flush_kernel_map(void *arg)
39 * Flush all to work around Errata in early athlons regarding
40 * large page flushing.
44 if (boot_cpu_data.x86_model >= 4)
48 static void global_flush_tlb(void)
50 BUG_ON(irqs_disabled());
52 on_each_cpu(flush_kernel_map, NULL, 1, 1);
56 * Certain areas of memory on x86 require very specific protection flags,
57 * for example the BIOS area or kernel text. Callers don't always get this
58 * right (again, ioremap() on BIOS memory is not uncommon) so this function
59 * checks and fixes these known static required protection bits.
61 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address)
63 pgprot_t forbidden = __pgprot(0);
66 * The BIOS area between 640k and 1Mb needs to be executable for
67 * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
69 if (within(__pa(address), BIOS_BEGIN, BIOS_END))
70 pgprot_val(forbidden) |= _PAGE_NX;
73 * The kernel text needs to be executable for obvious reasons
74 * Does not cover __inittext since that is gone later on
76 if (within(address, (unsigned long)_text, (unsigned long)_etext))
77 pgprot_val(forbidden) |= _PAGE_NX;
79 #ifdef CONFIG_DEBUG_RODATA
80 /* The .rodata section needs to be read-only */
81 if (within(address, (unsigned long)__start_rodata,
82 (unsigned long)__end_rodata))
83 pgprot_val(forbidden) |= _PAGE_RW;
86 prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
91 pte_t *lookup_address(unsigned long address, int *level)
93 pgd_t *pgd = pgd_offset_k(address);
97 *level = PG_LEVEL_NONE;
101 pud = pud_offset(pgd, address);
104 pmd = pmd_offset(pud, address);
108 *level = PG_LEVEL_2M;
112 *level = PG_LEVEL_4K;
113 return pte_offset_kernel(pmd, address);
116 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
119 set_pte_atomic(kpte, pte);
121 if (!SHARED_KERNEL_PMD) {
124 for (page = pgd_list; page; page = (struct page *)page->index) {
129 pgd = (pgd_t *)page_address(page) + pgd_index(address);
130 pud = pud_offset(pgd, address);
131 pmd = pmd_offset(pud, address);
132 set_pte_atomic((pte_t *)pmd, pte);
138 static int split_large_page(pte_t *kpte, unsigned long address)
140 pgprot_t ref_prot = pte_pgprot(pte_clrhuge(*kpte));
141 gfp_t gfp_flags = GFP_KERNEL;
148 #ifdef CONFIG_DEBUG_PAGEALLOC
149 gfp_flags = GFP_ATOMIC;
151 base = alloc_pages(gfp_flags, 0);
155 spin_lock_irqsave(&pgd_lock, flags);
157 * Check for races, another CPU might have split this page
160 tmp = lookup_address(address, &level);
166 address = __pa(address);
167 addr = address & LARGE_PAGE_MASK;
168 pbase = (pte_t *)page_address(base);
170 paravirt_alloc_pt(&init_mm, page_to_pfn(base));
173 for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE)
174 set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT, ref_prot));
177 * Install the new, split up pagetable. Important detail here:
179 * On Intel the NX bit of all levels must be cleared to make a
180 * page executable. See section 4.13.2 of Intel 64 and IA-32
181 * Architectures Software Developer's Manual).
183 ref_prot = pte_pgprot(pte_mkexec(pte_clrhuge(*kpte)));
184 __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
188 spin_unlock_irqrestore(&pgd_lock, flags);
191 __free_pages(base, 0);
197 __change_page_attr(unsigned long address, unsigned long pfn, pgprot_t prot)
199 struct page *kpte_page;
204 BUG_ON(pfn > max_low_pfn);
208 kpte = lookup_address(address, &level);
212 kpte_page = virt_to_page(kpte);
213 BUG_ON(PageLRU(kpte_page));
214 BUG_ON(PageCompound(kpte_page));
216 prot = static_protections(prot, address);
218 if (level == PG_LEVEL_4K) {
219 WARN_ON_ONCE(pgprot_val(prot) & _PAGE_PSE);
220 set_pte_atomic(kpte, pfn_pte(pfn, canon_pgprot(prot)));
222 /* Clear the PSE bit for the 4k level pages ! */
223 pgprot_val(prot) = pgprot_val(prot) & ~_PAGE_PSE;
225 err = split_large_page(kpte, address);
233 * change_page_attr_addr - Change page table attributes in linear mapping
234 * @address: Virtual address in linear mapping.
235 * @prot: New page table attribute (PAGE_*)
237 * Change page attributes of a page in the direct mapping. This is a variant
238 * of change_page_attr() that also works on memory holes that do not have
239 * mem_map entry (pfn_valid() is false).
241 * See change_page_attr() documentation for more details.
243 * Modules and drivers should use the set_memory_* APIs instead.
246 static int change_page_attr_addr(unsigned long address, pgprot_t prot)
248 int err = 0, kernel_map = 0;
249 unsigned long pfn = __pa(address) >> PAGE_SHIFT;
252 if (address >= __START_KERNEL_map &&
253 address < __START_KERNEL_map + KERNEL_TEXT_SIZE) {
255 address = (unsigned long)__va(__pa(address));
260 if (!kernel_map || pte_present(pfn_pte(0, prot))) {
261 err = __change_page_attr(address, pfn, prot);
268 * Handle kernel mapping too which aliases part of
271 if (__pa(address) < KERNEL_TEXT_SIZE) {
275 addr2 = __START_KERNEL_map + __pa(address);
276 /* Make sure the kernel mappings stay executable */
277 prot2 = pte_pgprot(pte_mkexec(pfn_pte(0, prot)));
278 err = __change_page_attr(addr2, pfn, prot2);
286 * change_page_attr_set - Change page table attributes in the linear mapping.
287 * @addr: Virtual address in linear mapping.
288 * @numpages: Number of pages to change
289 * @prot: Protection/caching type bits to set (PAGE_*)
291 * Returns 0 on success, otherwise a negated errno.
293 * This should be used when a page is mapped with a different caching policy
294 * than write-back somewhere - some CPUs do not like it when mappings with
295 * different caching policies exist. This changes the page attributes of the
296 * in kernel linear mapping too.
298 * The caller needs to ensure that there are no conflicting mappings elsewhere
299 * (e.g. in user space) * This function only deals with the kernel linear map.
301 * This function is different from change_page_attr() in that only selected bits
302 * are impacted, all other bits remain as is.
304 static int __change_page_attr_set(unsigned long addr, int numpages,
307 pgprot_t current_prot, new_prot;
312 for (i = 0; i < numpages ; i++) {
314 pte = lookup_address(addr, &level);
318 current_prot = pte_pgprot(*pte);
320 pgprot_val(new_prot) =
321 pgprot_val(current_prot) | pgprot_val(prot);
323 ret = change_page_attr_addr(addr, new_prot);
332 static int change_page_attr_set(unsigned long addr, int numpages, pgprot_t prot)
334 int ret = __change_page_attr_set(addr, numpages, prot);
342 * change_page_attr_clear - Change page table attributes in the linear mapping.
343 * @addr: Virtual address in linear mapping.
344 * @numpages: Number of pages to change
345 * @prot: Protection/caching type bits to clear (PAGE_*)
347 * Returns 0 on success, otherwise a negated errno.
349 * This should be used when a page is mapped with a different caching policy
350 * than write-back somewhere - some CPUs do not like it when mappings with
351 * different caching policies exist. This changes the page attributes of the
352 * in kernel linear mapping too.
354 * The caller needs to ensure that there are no conflicting mappings elsewhere
355 * (e.g. in user space) * This function only deals with the kernel linear map.
357 * This function is different from change_page_attr() in that only selected bits
358 * are impacted, all other bits remain as is.
360 static int __change_page_attr_clear(unsigned long addr, int numpages,
363 pgprot_t current_prot, new_prot;
368 for (i = 0; i < numpages; i++) {
370 pte = lookup_address(addr, &level);
374 current_prot = pte_pgprot(*pte);
376 pgprot_val(new_prot) =
377 pgprot_val(current_prot) & ~pgprot_val(prot);
379 ret = change_page_attr_addr(addr, new_prot);
388 static int change_page_attr_clear(unsigned long addr, int numpages,
391 int ret = __change_page_attr_clear(addr, numpages, prot);
398 int set_memory_uc(unsigned long addr, int numpages)
400 return change_page_attr_set(addr, numpages,
401 __pgprot(_PAGE_PCD | _PAGE_PWT));
403 EXPORT_SYMBOL(set_memory_uc);
405 int set_memory_wb(unsigned long addr, int numpages)
407 return change_page_attr_clear(addr, numpages,
408 __pgprot(_PAGE_PCD | _PAGE_PWT));
410 EXPORT_SYMBOL(set_memory_wb);
412 int set_memory_x(unsigned long addr, int numpages)
414 return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_NX));
416 EXPORT_SYMBOL(set_memory_x);
418 int set_memory_nx(unsigned long addr, int numpages)
420 return change_page_attr_set(addr, numpages, __pgprot(_PAGE_NX));
422 EXPORT_SYMBOL(set_memory_nx);
424 int set_memory_ro(unsigned long addr, int numpages)
426 return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_RW));
429 int set_memory_rw(unsigned long addr, int numpages)
431 return change_page_attr_set(addr, numpages, __pgprot(_PAGE_RW));
434 int set_memory_np(unsigned long addr, int numpages)
436 return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_PRESENT));
439 int set_pages_uc(struct page *page, int numpages)
441 unsigned long addr = (unsigned long)page_address(page);
443 return set_memory_uc(addr, numpages);
445 EXPORT_SYMBOL(set_pages_uc);
447 int set_pages_wb(struct page *page, int numpages)
449 unsigned long addr = (unsigned long)page_address(page);
451 return set_memory_wb(addr, numpages);
453 EXPORT_SYMBOL(set_pages_wb);
455 int set_pages_x(struct page *page, int numpages)
457 unsigned long addr = (unsigned long)page_address(page);
459 return set_memory_x(addr, numpages);
461 EXPORT_SYMBOL(set_pages_x);
463 int set_pages_nx(struct page *page, int numpages)
465 unsigned long addr = (unsigned long)page_address(page);
467 return set_memory_nx(addr, numpages);
469 EXPORT_SYMBOL(set_pages_nx);
471 int set_pages_ro(struct page *page, int numpages)
473 unsigned long addr = (unsigned long)page_address(page);
475 return set_memory_ro(addr, numpages);
478 int set_pages_rw(struct page *page, int numpages)
480 unsigned long addr = (unsigned long)page_address(page);
482 return set_memory_rw(addr, numpages);
486 #ifdef CONFIG_DEBUG_PAGEALLOC
488 static int __set_pages_p(struct page *page, int numpages)
490 unsigned long addr = (unsigned long)page_address(page);
492 return __change_page_attr_set(addr, numpages,
493 __pgprot(_PAGE_PRESENT | _PAGE_RW));
496 static int __set_pages_np(struct page *page, int numpages)
498 unsigned long addr = (unsigned long)page_address(page);
500 return __change_page_attr_clear(addr, numpages,
501 __pgprot(_PAGE_PRESENT));
504 void kernel_map_pages(struct page *page, int numpages, int enable)
506 if (PageHighMem(page))
509 debug_check_no_locks_freed(page_address(page),
510 numpages * PAGE_SIZE);
514 * If page allocator is not up yet then do not call c_p_a():
516 if (!debug_pagealloc_enabled)
520 * The return value is ignored - the calls cannot fail,
521 * large pages are disabled at boot time:
524 __set_pages_p(page, numpages);
526 __set_pages_np(page, numpages);
529 * We should perform an IPI and flush all tlbs,
530 * but that can deadlock->flush only current cpu:
537 * The testcases use internal knowledge of the implementation that shouldn't
538 * be exposed to the rest of the kernel. Include these directly here.
540 #ifdef CONFIG_CPA_DEBUG
541 #include "pageattr-test.c"