Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/arch/cris/mm/fault.c | |
3 | * | |
4 | * Copyright (C) 2000, 2001 Axis Communications AB | |
5 | * | |
6 | * Authors: Bjorn Wesen | |
7 | * | |
8 | * $Log: fault.c,v $ | |
4f18cfbf MS |
9 | * Revision 1.20 2005/03/04 08:16:18 starvik |
10 | * Merge of Linux 2.6.11. | |
11 | * | |
12 | * Revision 1.19 2005/01/14 10:07:59 starvik | |
13 | * Fixed warning. | |
14 | * | |
15 | * Revision 1.18 2005/01/12 08:10:14 starvik | |
49b4ff33 | 16 | * Re-added the change of frametype when handling kernel page fault fixup |
4f18cfbf MS |
17 | * for v10. This is necessary to avoid that the CPU remakes the faulting |
18 | * access. | |
19 | * | |
20 | * Revision 1.17 2005/01/11 13:53:05 starvik | |
21 | * Use raw_printk. | |
22 | * | |
23 | * Revision 1.16 2004/12/17 11:39:41 starvik | |
24 | * SMP support. | |
25 | * | |
26 | * Revision 1.15 2004/11/23 18:36:18 starvik | |
27 | * Stack is now non-executable. | |
28 | * Signal handler trampolines are placed in a reserved page mapped into all | |
29 | * processes. | |
30 | * | |
31 | * Revision 1.14 2004/11/23 07:10:21 starvik | |
32 | * Moved find_fixup_code to generic code. | |
33 | * | |
34 | * Revision 1.13 2004/11/23 07:00:54 starvik | |
35 | * Actually use the execute permission bit in the MMU. This makes it possible | |
36 | * to prevent e.g. attacks where executable code is put on the stack. | |
37 | * | |
38 | * Revision 1.12 2004/09/29 06:16:04 starvik | |
39 | * Use instruction_pointer | |
40 | * | |
1da177e4 LT |
41 | * Revision 1.11 2004/05/14 07:58:05 starvik |
42 | * Merge of changes from 2.4 | |
43 | * | |
44 | * Revision 1.10 2003/10/27 14:51:24 starvik | |
45 | * Removed debugcode | |
46 | * | |
47 | * Revision 1.9 2003/10/27 14:50:42 starvik | |
48 | * Changed do_page_fault signature | |
49 | * | |
50 | * Revision 1.8 2003/07/04 13:02:48 tobiasa | |
51 | * Moved code snippet from arch/cris/mm/fault.c that searches for fixup code | |
49b4ff33 | 52 | * to separate function in arch-specific files. |
1da177e4 LT |
53 | * |
54 | * Revision 1.7 2003/01/22 06:48:38 starvik | |
55 | * Fixed warnings issued by GCC 3.2.1 | |
56 | * | |
57 | * Revision 1.6 2003/01/09 14:42:52 starvik | |
58 | * Merge of Linux 2.5.55 | |
59 | * | |
60 | * Revision 1.5 2002/12/11 14:44:48 starvik | |
61 | * Extracted v10 (ETRAX 100LX) specific stuff to arch/cris/arch-v10/mm | |
62 | * | |
63 | * Revision 1.4 2002/11/13 15:10:28 starvik | |
64 | * pte_offset has been renamed to pte_offset_kernel | |
65 | * | |
66 | * Revision 1.3 2002/11/05 06:45:13 starvik | |
67 | * Merge of Linux 2.5.45 | |
68 | * | |
69 | * Revision 1.2 2001/12/18 13:35:22 bjornw | |
70 | * Applied the 2.4.13->2.4.16 CRIS patch to 2.5.1 (is a copy of 2.4.15). | |
71 | * | |
72 | * Revision 1.20 2001/11/22 13:34:06 bjornw | |
73 | * * Bug workaround (LX TR89): force a rerun of the whole of an interrupted | |
74 | * unaligned write, because the second half of the write will be corrupted | |
75 | * otherwise. Affected unaligned writes spanning not-yet mapped pages. | |
76 | * * Optimization: use the wr_rd bit in R_MMU_CAUSE to know whether a miss | |
77 | * was due to a read or a write (before we didn't know this until the next | |
78 | * restart of the interrupted instruction, thus wasting one fault-irq) | |
79 | * | |
80 | * Revision 1.19 2001/11/12 19:02:10 pkj | |
81 | * Fixed compiler warnings. | |
82 | * | |
83 | * Revision 1.18 2001/07/18 22:14:32 bjornw | |
84 | * Enable interrupts in the bulk of do_page_fault | |
85 | * | |
86 | * Revision 1.17 2001/07/18 13:07:23 bjornw | |
87 | * * Detect non-existant PTE's in vmalloc pmd synchronization | |
88 | * * Remove comment about fast-paths for VMALLOC_START etc, because all that | |
89 | * was totally bogus anyway it turned out :) | |
90 | * * Fix detection of vmalloc-area synchronization | |
91 | * * Add some comments | |
92 | * | |
93 | * Revision 1.16 2001/06/13 00:06:08 bjornw | |
94 | * current_pgd should be volatile | |
95 | * | |
96 | * Revision 1.15 2001/06/13 00:02:23 bjornw | |
97 | * Use a separate variable to store the current pgd to avoid races in schedule | |
98 | * | |
99 | * Revision 1.14 2001/05/16 17:41:07 hp | |
100 | * Last comment tweak further tweaked. | |
101 | * | |
102 | * Revision 1.13 2001/05/15 00:58:44 hp | |
103 | * Expand a bit on the comment why we compare address >= TASK_SIZE rather | |
104 | * than >= VMALLOC_START. | |
105 | * | |
106 | * Revision 1.12 2001/04/04 10:51:14 bjornw | |
107 | * mmap_sem is grabbed for reading | |
108 | * | |
109 | * Revision 1.11 2001/03/23 07:36:07 starvik | |
110 | * Corrected according to review remarks | |
111 | * | |
112 | * Revision 1.10 2001/03/21 16:10:11 bjornw | |
113 | * CRIS_FRAME_FIXUP not needed anymore, use FRAME_NORMAL | |
114 | * | |
115 | * Revision 1.9 2001/03/05 13:22:20 bjornw | |
116 | * Spell-fix and fix in vmalloc_fault handling | |
117 | * | |
118 | * Revision 1.8 2000/11/22 14:45:31 bjornw | |
119 | * * 2.4.0-test10 removed the set_pgdir instantaneous kernel global mapping | |
120 | * into all processes. Instead we fill in the missing PTE entries on demand. | |
121 | * | |
122 | * Revision 1.7 2000/11/21 16:39:09 bjornw | |
123 | * fixup switches frametype | |
124 | * | |
125 | * Revision 1.6 2000/11/17 16:54:08 bjornw | |
126 | * More detailed siginfo reporting | |
127 | * | |
128 | * | |
129 | */ | |
130 | ||
131 | #include <linux/mm.h> | |
132 | #include <linux/interrupt.h> | |
133 | #include <linux/module.h> | |
134 | #include <asm/uaccess.h> | |
135 | ||
136 | extern int find_fixup_code(struct pt_regs *); | |
137 | extern void die_if_kernel(const char *, struct pt_regs *, long); | |
4f18cfbf | 138 | extern int raw_printk(const char *fmt, ...); |
1da177e4 LT |
139 | |
140 | /* debug of low-level TLB reload */ | |
141 | #undef DEBUG | |
142 | ||
143 | #ifdef DEBUG | |
144 | #define D(x) x | |
145 | #else | |
146 | #define D(x) | |
147 | #endif | |
148 | ||
149 | /* debug of higher-level faults */ | |
150 | #define DPG(x) | |
151 | ||
152 | /* current active page directory */ | |
153 | ||
4f18cfbf MS |
154 | volatile DEFINE_PER_CPU(pgd_t *,current_pgd); |
155 | unsigned long cris_signal_return_page; | |
1da177e4 LT |
156 | |
157 | /* | |
158 | * This routine handles page faults. It determines the address, | |
159 | * and the problem, and then passes it off to one of the appropriate | |
160 | * routines. | |
161 | * | |
162 | * Notice that the address we're given is aligned to the page the fault | |
163 | * occurred in, since we only get the PFN in R_MMU_CAUSE not the complete | |
164 | * address. | |
165 | * | |
166 | * error_code: | |
167 | * bit 0 == 0 means no page found, 1 means protection fault | |
168 | * bit 1 == 0 means read, 1 means write | |
169 | * | |
170 | * If this routine detects a bad access, it returns 1, otherwise it | |
171 | * returns 0. | |
172 | */ | |
173 | ||
174 | asmlinkage void | |
175 | do_page_fault(unsigned long address, struct pt_regs *regs, | |
176 | int protection, int writeaccess) | |
177 | { | |
178 | struct task_struct *tsk; | |
179 | struct mm_struct *mm; | |
180 | struct vm_area_struct * vma; | |
181 | siginfo_t info; | |
83c54070 | 182 | int fault; |
1da177e4 | 183 | |
4f18cfbf MS |
184 | D(printk("Page fault for %lX on %X at %lX, prot %d write %d\n", |
185 | address, smp_processor_id(), instruction_pointer(regs), | |
186 | protection, writeaccess)); | |
1da177e4 LT |
187 | |
188 | tsk = current; | |
189 | ||
190 | /* | |
191 | * We fault-in kernel-space virtual memory on-demand. The | |
192 | * 'reference' page table is init_mm.pgd. | |
193 | * | |
194 | * NOTE! We MUST NOT take any locks for this case. We may | |
195 | * be in an interrupt or a critical region, and should | |
196 | * only copy the information from the master page table, | |
197 | * nothing more. | |
198 | * | |
199 | * NOTE2: This is done so that, when updating the vmalloc | |
200 | * mappings we don't have to walk all processes pgdirs and | |
201 | * add the high mappings all at once. Instead we do it as they | |
202 | * are used. However vmalloc'ed page entries have the PAGE_GLOBAL | |
203 | * bit set so sometimes the TLB can use a lingering entry. | |
204 | * | |
205 | * This verifies that the fault happens in kernel space | |
206 | * and that the fault was not a protection error (error_code & 1). | |
207 | */ | |
208 | ||
209 | if (address >= VMALLOC_START && | |
210 | !protection && | |
211 | !user_mode(regs)) | |
212 | goto vmalloc_fault; | |
213 | ||
4f18cfbf MS |
214 | /* When stack execution is not allowed we store the signal |
215 | * trampolines in the reserved cris_signal_return_page. | |
216 | * Handle this in the exact same way as vmalloc (we know | |
217 | * that the mapping is there and is valid so no need to | |
218 | * call handle_mm_fault). | |
219 | */ | |
220 | if (cris_signal_return_page && | |
221 | address == cris_signal_return_page && | |
222 | !protection && user_mode(regs)) | |
223 | goto vmalloc_fault; | |
224 | ||
1da177e4 | 225 | /* we can and should enable interrupts at this point */ |
4f18cfbf | 226 | local_irq_enable(); |
1da177e4 LT |
227 | |
228 | mm = tsk->mm; | |
229 | info.si_code = SEGV_MAPERR; | |
230 | ||
231 | /* | |
232 | * If we're in an interrupt or have no user | |
233 | * context, we must not take the fault.. | |
234 | */ | |
235 | ||
6edaf68a | 236 | if (in_atomic() || !mm) |
1da177e4 LT |
237 | goto no_context; |
238 | ||
239 | down_read(&mm->mmap_sem); | |
240 | vma = find_vma(mm, address); | |
241 | if (!vma) | |
242 | goto bad_area; | |
243 | if (vma->vm_start <= address) | |
244 | goto good_area; | |
245 | if (!(vma->vm_flags & VM_GROWSDOWN)) | |
246 | goto bad_area; | |
247 | if (user_mode(regs)) { | |
248 | /* | |
249 | * accessing the stack below usp is always a bug. | |
250 | * we get page-aligned addresses so we can only check | |
251 | * if we're within a page from usp, but that might be | |
252 | * enough to catch brutal errors at least. | |
253 | */ | |
254 | if (address + PAGE_SIZE < rdusp()) | |
255 | goto bad_area; | |
256 | } | |
257 | if (expand_stack(vma, address)) | |
258 | goto bad_area; | |
259 | ||
260 | /* | |
261 | * Ok, we have a good vm_area for this memory access, so | |
262 | * we can handle it.. | |
263 | */ | |
264 | ||
265 | good_area: | |
266 | info.si_code = SEGV_ACCERR; | |
267 | ||
268 | /* first do some preliminary protection checks */ | |
269 | ||
4f18cfbf MS |
270 | if (writeaccess == 2){ |
271 | if (!(vma->vm_flags & VM_EXEC)) | |
272 | goto bad_area; | |
273 | } else if (writeaccess == 1) { | |
1da177e4 LT |
274 | if (!(vma->vm_flags & VM_WRITE)) |
275 | goto bad_area; | |
276 | } else { | |
277 | if (!(vma->vm_flags & (VM_READ | VM_EXEC))) | |
278 | goto bad_area; | |
279 | } | |
280 | ||
281 | /* | |
282 | * If for any reason at all we couldn't handle the fault, | |
283 | * make sure we exit gracefully rather than endlessly redo | |
284 | * the fault. | |
285 | */ | |
286 | ||
83c54070 NP |
287 | fault = handle_mm_fault(mm, vma, address, writeaccess & 1); |
288 | if (unlikely(fault & VM_FAULT_ERROR)) { | |
289 | if (fault & VM_FAULT_OOM) | |
290 | goto out_of_memory; | |
291 | else if (fault & VM_FAULT_SIGBUS) | |
292 | goto do_sigbus; | |
293 | BUG(); | |
1da177e4 | 294 | } |
83c54070 NP |
295 | if (fault & VM_FAULT_MAJOR) |
296 | tsk->maj_flt++; | |
297 | else | |
298 | tsk->min_flt++; | |
1da177e4 LT |
299 | |
300 | up_read(&mm->mmap_sem); | |
301 | return; | |
302 | ||
303 | /* | |
304 | * Something tried to access memory that isn't in our memory map.. | |
305 | * Fix it, but check if it's kernel or user first.. | |
306 | */ | |
307 | ||
308 | bad_area: | |
309 | up_read(&mm->mmap_sem); | |
310 | ||
311 | bad_area_nosemaphore: | |
312 | DPG(show_registers(regs)); | |
313 | ||
314 | /* User mode accesses just cause a SIGSEGV */ | |
315 | ||
316 | if (user_mode(regs)) { | |
317 | info.si_signo = SIGSEGV; | |
318 | info.si_errno = 0; | |
319 | /* info.si_code has been set above */ | |
320 | info.si_addr = (void *)address; | |
321 | force_sig_info(SIGSEGV, &info, tsk); | |
322 | return; | |
323 | } | |
324 | ||
325 | no_context: | |
326 | ||
327 | /* Are we prepared to handle this kernel fault? | |
328 | * | |
329 | * (The kernel has valid exception-points in the source | |
330 | * when it acesses user-memory. When it fails in one | |
331 | * of those points, we find it in a table and do a jump | |
332 | * to some fixup code that loads an appropriate error | |
333 | * code) | |
334 | */ | |
335 | ||
336 | if (find_fixup_code(regs)) | |
337 | return; | |
338 | ||
339 | /* | |
340 | * Oops. The kernel tried to access some bad page. We'll have to | |
341 | * terminate things with extreme prejudice. | |
342 | */ | |
343 | ||
344 | if ((unsigned long) (address) < PAGE_SIZE) | |
4f18cfbf | 345 | raw_printk(KERN_ALERT "Unable to handle kernel NULL pointer dereference"); |
1da177e4 | 346 | else |
4f18cfbf MS |
347 | raw_printk(KERN_ALERT "Unable to handle kernel access"); |
348 | raw_printk(" at virtual address %08lx\n",address); | |
1da177e4 LT |
349 | |
350 | die_if_kernel("Oops", regs, (writeaccess << 1) | protection); | |
351 | ||
352 | do_exit(SIGKILL); | |
353 | ||
354 | /* | |
355 | * We ran out of memory, or some other thing happened to us that made | |
356 | * us unable to handle the page fault gracefully. | |
357 | */ | |
358 | ||
359 | out_of_memory: | |
360 | up_read(&mm->mmap_sem); | |
361 | printk("VM: killing process %s\n", tsk->comm); | |
362 | if (user_mode(regs)) | |
dcca2bde | 363 | do_group_exit(SIGKILL); |
1da177e4 LT |
364 | goto no_context; |
365 | ||
366 | do_sigbus: | |
367 | up_read(&mm->mmap_sem); | |
368 | ||
369 | /* | |
370 | * Send a sigbus, regardless of whether we were in kernel | |
371 | * or user mode. | |
372 | */ | |
373 | info.si_signo = SIGBUS; | |
374 | info.si_errno = 0; | |
375 | info.si_code = BUS_ADRERR; | |
376 | info.si_addr = (void *)address; | |
377 | force_sig_info(SIGBUS, &info, tsk); | |
378 | ||
379 | /* Kernel mode? Handle exceptions or die */ | |
380 | if (!user_mode(regs)) | |
381 | goto no_context; | |
382 | return; | |
383 | ||
384 | vmalloc_fault: | |
385 | { | |
386 | /* | |
387 | * Synchronize this task's top level page-table | |
388 | * with the 'reference' page table. | |
389 | * | |
390 | * Use current_pgd instead of tsk->active_mm->pgd | |
391 | * since the latter might be unavailable if this | |
392 | * code is executed in a misfortunately run irq | |
393 | * (like inside schedule() between switch_mm and | |
394 | * switch_to...). | |
395 | */ | |
396 | ||
397 | int offset = pgd_index(address); | |
398 | pgd_t *pgd, *pgd_k; | |
4f18cfbf | 399 | pud_t *pud, *pud_k; |
1da177e4 LT |
400 | pmd_t *pmd, *pmd_k; |
401 | pte_t *pte_k; | |
402 | ||
4f18cfbf | 403 | pgd = (pgd_t *)per_cpu(current_pgd, smp_processor_id()) + offset; |
1da177e4 LT |
404 | pgd_k = init_mm.pgd + offset; |
405 | ||
406 | /* Since we're two-level, we don't need to do both | |
407 | * set_pgd and set_pmd (they do the same thing). If | |
408 | * we go three-level at some point, do the right thing | |
409 | * with pgd_present and set_pgd here. | |
410 | * | |
411 | * Also, since the vmalloc area is global, we don't | |
412 | * need to copy individual PTE's, it is enough to | |
413 | * copy the pgd pointer into the pte page of the | |
414 | * root task. If that is there, we'll find our pte if | |
415 | * it exists. | |
416 | */ | |
417 | ||
4f18cfbf MS |
418 | pud = pud_offset(pgd, address); |
419 | pud_k = pud_offset(pgd_k, address); | |
420 | if (!pud_present(*pud_k)) | |
421 | goto no_context; | |
422 | ||
423 | pmd = pmd_offset(pud, address); | |
424 | pmd_k = pmd_offset(pud_k, address); | |
1da177e4 LT |
425 | |
426 | if (!pmd_present(*pmd_k)) | |
427 | goto bad_area_nosemaphore; | |
428 | ||
429 | set_pmd(pmd, *pmd_k); | |
430 | ||
431 | /* Make sure the actual PTE exists as well to | |
432 | * catch kernel vmalloc-area accesses to non-mapped | |
433 | * addresses. If we don't do this, this will just | |
434 | * silently loop forever. | |
435 | */ | |
436 | ||
437 | pte_k = pte_offset_kernel(pmd_k, address); | |
438 | if (!pte_present(*pte_k)) | |
439 | goto no_context; | |
440 | ||
441 | return; | |
442 | } | |
443 | } | |
4f18cfbf MS |
444 | |
445 | /* Find fixup code. */ | |
446 | int | |
447 | find_fixup_code(struct pt_regs *regs) | |
448 | { | |
449 | const struct exception_table_entry *fixup; | |
450 | ||
451 | if ((fixup = search_exception_tables(instruction_pointer(regs))) != 0) { | |
452 | /* Adjust the instruction pointer in the stackframe. */ | |
453 | instruction_pointer(regs) = fixup->fixup; | |
454 | arch_fixup(regs); | |
455 | return 1; | |
456 | } | |
457 | ||
458 | return 0; | |
459 | } |