x86: cpa use the new set_clr function
[linux-2.6] / arch / x86 / mm / pageattr.c
1 /*
2  * Copyright 2002 Andi Kleen, SuSE Labs.
3  * Thanks to Ben LaHaise for precious feedback.
4  */
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>
10 #include <linux/mm.h>
11
12 #include <asm/e820.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>
18
19 static inline int
20 within(unsigned long addr, unsigned long start, unsigned long end)
21 {
22         return addr >= start && addr < end;
23 }
24
25 /*
26  * Flushing functions
27  */
28 void clflush_cache_range(void *addr, int size)
29 {
30         int i;
31
32         for (i = 0; i < size; i += boot_cpu_data.x86_clflush_size)
33                 clflush(addr+i);
34 }
35
36 static void flush_kernel_map(void *arg)
37 {
38         /*
39          * Flush all to work around Errata in early athlons regarding
40          * large page flushing.
41          */
42         __flush_tlb_all();
43
44         if (boot_cpu_data.x86_model >= 4)
45                 wbinvd();
46 }
47
48 static void global_flush_tlb(void)
49 {
50         BUG_ON(irqs_disabled());
51
52         on_each_cpu(flush_kernel_map, NULL, 1, 1);
53 }
54
55 /*
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.
60  */
61 static inline pgprot_t static_protections(pgprot_t prot, unsigned long address)
62 {
63         pgprot_t forbidden = __pgprot(0);
64
65         /*
66          * The BIOS area between 640k and 1Mb needs to be executable for
67          * PCI BIOS based config access (CONFIG_PCI_GOBIOS) support.
68          */
69         if (within(__pa(address), BIOS_BEGIN, BIOS_END))
70                 pgprot_val(forbidden) |= _PAGE_NX;
71
72         /*
73          * The kernel text needs to be executable for obvious reasons
74          * Does not cover __inittext since that is gone later on
75          */
76         if (within(address, (unsigned long)_text, (unsigned long)_etext))
77                 pgprot_val(forbidden) |= _PAGE_NX;
78
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;
84 #endif
85
86         prot = __pgprot(pgprot_val(prot) & ~pgprot_val(forbidden));
87
88         return prot;
89 }
90
91 pte_t *lookup_address(unsigned long address, int *level)
92 {
93         pgd_t *pgd = pgd_offset_k(address);
94         pud_t *pud;
95         pmd_t *pmd;
96
97         *level = PG_LEVEL_NONE;
98
99         if (pgd_none(*pgd))
100                 return NULL;
101         pud = pud_offset(pgd, address);
102         if (pud_none(*pud))
103                 return NULL;
104         pmd = pmd_offset(pud, address);
105         if (pmd_none(*pmd))
106                 return NULL;
107
108         *level = PG_LEVEL_2M;
109         if (pmd_large(*pmd))
110                 return (pte_t *)pmd;
111
112         *level = PG_LEVEL_4K;
113         return pte_offset_kernel(pmd, address);
114 }
115
116 static void __set_pmd_pte(pte_t *kpte, unsigned long address, pte_t pte)
117 {
118         /* change init_mm */
119         set_pte_atomic(kpte, pte);
120 #ifdef CONFIG_X86_32
121         if (!SHARED_KERNEL_PMD) {
122                 struct page *page;
123
124                 for (page = pgd_list; page; page = (struct page *)page->index) {
125                         pgd_t *pgd;
126                         pud_t *pud;
127                         pmd_t *pmd;
128
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);
133                 }
134         }
135 #endif
136 }
137
138 static int split_large_page(pte_t *kpte, unsigned long address)
139 {
140         pgprot_t ref_prot = pte_pgprot(pte_clrhuge(*kpte));
141         gfp_t gfp_flags = GFP_KERNEL;
142         unsigned long flags;
143         unsigned long addr;
144         pte_t *pbase, *tmp;
145         struct page *base;
146         int i, level;
147
148 #ifdef CONFIG_DEBUG_PAGEALLOC
149         gfp_flags = GFP_ATOMIC;
150 #endif
151         base = alloc_pages(gfp_flags, 0);
152         if (!base)
153                 return -ENOMEM;
154
155         spin_lock_irqsave(&pgd_lock, flags);
156         /*
157          * Check for races, another CPU might have split this page
158          * up for us already:
159          */
160         tmp = lookup_address(address, &level);
161         if (tmp != kpte) {
162                 WARN_ON_ONCE(1);
163                 goto out_unlock;
164         }
165
166         address = __pa(address);
167         addr = address & LARGE_PAGE_MASK;
168         pbase = (pte_t *)page_address(base);
169 #ifdef CONFIG_X86_32
170         paravirt_alloc_pt(&init_mm, page_to_pfn(base));
171 #endif
172
173         for (i = 0; i < PTRS_PER_PTE; i++, addr += PAGE_SIZE)
174                 set_pte(&pbase[i], pfn_pte(addr >> PAGE_SHIFT, ref_prot));
175
176         /*
177          * Install the new, split up pagetable. Important detail here:
178          *
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).
182          */
183         ref_prot = pte_pgprot(pte_mkexec(pte_clrhuge(*kpte)));
184         __set_pmd_pte(kpte, address, mk_pte(base, ref_prot));
185         base = NULL;
186
187 out_unlock:
188         spin_unlock_irqrestore(&pgd_lock, flags);
189
190         if (base)
191                 __free_pages(base, 0);
192
193         return 0;
194 }
195
196 static int
197 __change_page_attr(unsigned long address, unsigned long pfn, pgprot_t prot)
198 {
199         struct page *kpte_page;
200         int level, err = 0;
201         pte_t *kpte;
202
203 #ifdef CONFIG_X86_32
204         BUG_ON(pfn > max_low_pfn);
205 #endif
206
207 repeat:
208         kpte = lookup_address(address, &level);
209         if (!kpte)
210                 return -EINVAL;
211
212         kpte_page = virt_to_page(kpte);
213         BUG_ON(PageLRU(kpte_page));
214         BUG_ON(PageCompound(kpte_page));
215
216         prot = static_protections(prot, address);
217
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)));
221         } else {
222                 /* Clear the PSE bit for the 4k level pages ! */
223                 pgprot_val(prot) = pgprot_val(prot) & ~_PAGE_PSE;
224
225                 err = split_large_page(kpte, address);
226                 if (!err)
227                         goto repeat;
228         }
229         return err;
230 }
231
232 /**
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_*)
236  *
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).
240  *
241  * See change_page_attr() documentation for more details.
242  *
243  * Modules and drivers should use the set_memory_* APIs instead.
244  */
245
246 static int change_page_attr_addr(unsigned long address, pgprot_t prot)
247 {
248         int err = 0, kernel_map = 0;
249         unsigned long pfn = __pa(address) >> PAGE_SHIFT;
250
251 #ifdef CONFIG_X86_64
252         if (address >= __START_KERNEL_map &&
253                         address < __START_KERNEL_map + KERNEL_TEXT_SIZE) {
254
255                 address = (unsigned long)__va(__pa(address));
256                 kernel_map = 1;
257         }
258 #endif
259
260         if (!kernel_map || pte_present(pfn_pte(0, prot))) {
261                 err = __change_page_attr(address, pfn, prot);
262                 if (err)
263                         return err;
264         }
265
266 #ifdef CONFIG_X86_64
267         /*
268          * Handle kernel mapping too which aliases part of
269          * lowmem:
270          */
271         if (__pa(address) < KERNEL_TEXT_SIZE) {
272                 unsigned long addr2;
273                 pgprot_t prot2;
274
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);
279         }
280 #endif
281
282         return err;
283 }
284
285 static int __change_page_attr_set_clr(unsigned long addr, int numpages,
286                                       pgprot_t mask_set, pgprot_t mask_clr)
287 {
288         pgprot_t new_prot;
289         int level;
290         pte_t *pte;
291         int i, ret;
292
293         for (i = 0; i < numpages ; i++) {
294
295                 pte = lookup_address(addr, &level);
296                 if (!pte)
297                         return -EINVAL;
298
299                 new_prot = pte_pgprot(*pte);
300
301                 pgprot_val(new_prot) &= ~pgprot_val(mask_clr);
302                 pgprot_val(new_prot) |= pgprot_val(mask_set);
303
304                 ret = change_page_attr_addr(addr, new_prot);
305                 if (ret)
306                         return ret;
307                 addr += PAGE_SIZE;
308         }
309
310         return 0;
311 }
312
313 static int change_page_attr_set_clr(unsigned long addr, int numpages,
314                                     pgprot_t mask_set, pgprot_t mask_clr)
315 {
316         int ret = __change_page_attr_set_clr(addr, numpages, mask_set,
317                                              mask_clr);
318
319         global_flush_tlb();
320
321         return ret;
322 }
323
324 static inline int change_page_attr_set(unsigned long addr, int numpages,
325                                        pgprot_t mask)
326 {
327         return change_page_attr_set_clr(addr, numpages, mask, __pgprot(0));
328 }
329
330 static inline int change_page_attr_clear(unsigned long addr, int numpages,
331                                          pgprot_t mask)
332 {
333         return __change_page_attr_set_clr(addr, numpages, __pgprot(0), mask);
334
335 }
336
337 int set_memory_uc(unsigned long addr, int numpages)
338 {
339         return change_page_attr_set(addr, numpages,
340                                     __pgprot(_PAGE_PCD | _PAGE_PWT));
341 }
342 EXPORT_SYMBOL(set_memory_uc);
343
344 int set_memory_wb(unsigned long addr, int numpages)
345 {
346         return change_page_attr_clear(addr, numpages,
347                                       __pgprot(_PAGE_PCD | _PAGE_PWT));
348 }
349 EXPORT_SYMBOL(set_memory_wb);
350
351 int set_memory_x(unsigned long addr, int numpages)
352 {
353         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_NX));
354 }
355 EXPORT_SYMBOL(set_memory_x);
356
357 int set_memory_nx(unsigned long addr, int numpages)
358 {
359         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_NX));
360 }
361 EXPORT_SYMBOL(set_memory_nx);
362
363 int set_memory_ro(unsigned long addr, int numpages)
364 {
365         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_RW));
366 }
367
368 int set_memory_rw(unsigned long addr, int numpages)
369 {
370         return change_page_attr_set(addr, numpages, __pgprot(_PAGE_RW));
371 }
372
373 int set_memory_np(unsigned long addr, int numpages)
374 {
375         return change_page_attr_clear(addr, numpages, __pgprot(_PAGE_PRESENT));
376 }
377
378 int set_pages_uc(struct page *page, int numpages)
379 {
380         unsigned long addr = (unsigned long)page_address(page);
381
382         return set_memory_uc(addr, numpages);
383 }
384 EXPORT_SYMBOL(set_pages_uc);
385
386 int set_pages_wb(struct page *page, int numpages)
387 {
388         unsigned long addr = (unsigned long)page_address(page);
389
390         return set_memory_wb(addr, numpages);
391 }
392 EXPORT_SYMBOL(set_pages_wb);
393
394 int set_pages_x(struct page *page, int numpages)
395 {
396         unsigned long addr = (unsigned long)page_address(page);
397
398         return set_memory_x(addr, numpages);
399 }
400 EXPORT_SYMBOL(set_pages_x);
401
402 int set_pages_nx(struct page *page, int numpages)
403 {
404         unsigned long addr = (unsigned long)page_address(page);
405
406         return set_memory_nx(addr, numpages);
407 }
408 EXPORT_SYMBOL(set_pages_nx);
409
410 int set_pages_ro(struct page *page, int numpages)
411 {
412         unsigned long addr = (unsigned long)page_address(page);
413
414         return set_memory_ro(addr, numpages);
415 }
416
417 int set_pages_rw(struct page *page, int numpages)
418 {
419         unsigned long addr = (unsigned long)page_address(page);
420
421         return set_memory_rw(addr, numpages);
422 }
423
424
425 #if defined(CONFIG_DEBUG_PAGEALLOC) || defined(CONFIG_CPA_DEBUG)
426 static inline int __change_page_attr_set(unsigned long addr, int numpages,
427                                          pgprot_t mask)
428 {
429         return __change_page_attr_set_clr(addr, numpages, mask, __pgprot(0));
430 }
431
432 static inline int __change_page_attr_clear(unsigned long addr, int numpages,
433                                            pgprot_t mask)
434 {
435         return __change_page_attr_set_clr(addr, numpages, __pgprot(0), mask);
436 }
437 #endif
438
439 #ifdef CONFIG_DEBUG_PAGEALLOC
440
441 static int __set_pages_p(struct page *page, int numpages)
442 {
443         unsigned long addr = (unsigned long)page_address(page);
444
445         return __change_page_attr_set(addr, numpages,
446                                       __pgprot(_PAGE_PRESENT | _PAGE_RW));
447 }
448
449 static int __set_pages_np(struct page *page, int numpages)
450 {
451         unsigned long addr = (unsigned long)page_address(page);
452
453         return __change_page_attr_clear(addr, numpages,
454                                         __pgprot(_PAGE_PRESENT));
455 }
456
457 void kernel_map_pages(struct page *page, int numpages, int enable)
458 {
459         if (PageHighMem(page))
460                 return;
461         if (!enable) {
462                 debug_check_no_locks_freed(page_address(page),
463                                            numpages * PAGE_SIZE);
464         }
465
466         /*
467          * If page allocator is not up yet then do not call c_p_a():
468          */
469         if (!debug_pagealloc_enabled)
470                 return;
471
472         /*
473          * The return value is ignored - the calls cannot fail,
474          * large pages are disabled at boot time:
475          */
476         if (enable)
477                 __set_pages_p(page, numpages);
478         else
479                 __set_pages_np(page, numpages);
480
481         /*
482          * We should perform an IPI and flush all tlbs,
483          * but that can deadlock->flush only current cpu:
484          */
485         __flush_tlb_all();
486 }
487 #endif
488
489 /*
490  * The testcases use internal knowledge of the implementation that shouldn't
491  * be exposed to the rest of the kernel. Include these directly here.
492  */
493 #ifdef CONFIG_CPA_DEBUG
494 #include "pageattr-test.c"
495 #endif