Merge commit 'v2.6.27-rc8' into x86/setup
[linux-2.6] / arch / sh / mm / fault_32.c
1 /*
2  * Page fault handler for SH with an MMU.
3  *
4  *  Copyright (C) 1999  Niibe Yutaka
5  *  Copyright (C) 2003 - 2007  Paul Mundt
6  *
7  *  Based on linux/arch/i386/mm/fault.c:
8  *   Copyright (C) 1995  Linus Torvalds
9  *
10  * This file is subject to the terms and conditions of the GNU General Public
11  * License.  See the file "COPYING" in the main directory of this archive
12  * for more details.
13  */
14 #include <linux/kernel.h>
15 #include <linux/mm.h>
16 #include <linux/hardirq.h>
17 #include <linux/kprobes.h>
18 #include <asm/io_trapped.h>
19 #include <asm/system.h>
20 #include <asm/mmu_context.h>
21 #include <asm/tlbflush.h>
22 #include <asm/kgdb.h>
23
24 /*
25  * This routine handles page faults.  It determines the address,
26  * and the problem, and then passes it off to one of the appropriate
27  * routines.
28  */
29 asmlinkage void __kprobes do_page_fault(struct pt_regs *regs,
30                                         unsigned long writeaccess,
31                                         unsigned long address)
32 {
33         struct task_struct *tsk;
34         struct mm_struct *mm;
35         struct vm_area_struct * vma;
36         int si_code;
37         int fault;
38         siginfo_t info;
39
40 #ifdef CONFIG_SH_KGDB
41         if (kgdb_nofault && kgdb_bus_err_hook)
42                 kgdb_bus_err_hook();
43 #endif
44
45         tsk = current;
46         si_code = SEGV_MAPERR;
47
48         if (unlikely(address >= TASK_SIZE)) {
49                 /*
50                  * Synchronize this task's top level page-table
51                  * with the 'reference' page table.
52                  *
53                  * Do _not_ use "tsk" here. We might be inside
54                  * an interrupt in the middle of a task switch..
55                  */
56                 int offset = pgd_index(address);
57                 pgd_t *pgd, *pgd_k;
58                 pud_t *pud, *pud_k;
59                 pmd_t *pmd, *pmd_k;
60
61                 pgd = get_TTB() + offset;
62                 pgd_k = swapper_pg_dir + offset;
63
64                 /* This will never happen with the folded page table. */
65                 if (!pgd_present(*pgd)) {
66                         if (!pgd_present(*pgd_k))
67                                 goto bad_area_nosemaphore;
68                         set_pgd(pgd, *pgd_k);
69                         return;
70                 }
71
72                 pud = pud_offset(pgd, address);
73                 pud_k = pud_offset(pgd_k, address);
74                 if (pud_present(*pud) || !pud_present(*pud_k))
75                         goto bad_area_nosemaphore;
76                 set_pud(pud, *pud_k);
77
78                 pmd = pmd_offset(pud, address);
79                 pmd_k = pmd_offset(pud_k, address);
80                 if (pmd_present(*pmd) || !pmd_present(*pmd_k))
81                         goto bad_area_nosemaphore;
82                 set_pmd(pmd, *pmd_k);
83
84                 return;
85         }
86
87         /* Only enable interrupts if they were on before the fault */
88         if ((regs->sr & SR_IMASK) != SR_IMASK) {
89                 trace_hardirqs_on();
90                 local_irq_enable();
91         }
92
93         mm = tsk->mm;
94
95         /*
96          * If we're in an interrupt or have no user
97          * context, we must not take the fault..
98          */
99         if (in_atomic() || !mm)
100                 goto no_context;
101
102         down_read(&mm->mmap_sem);
103
104         vma = find_vma(mm, address);
105         if (!vma)
106                 goto bad_area;
107         if (vma->vm_start <= address)
108                 goto good_area;
109         if (!(vma->vm_flags & VM_GROWSDOWN))
110                 goto bad_area;
111         if (expand_stack(vma, address))
112                 goto bad_area;
113 /*
114  * Ok, we have a good vm_area for this memory access, so
115  * we can handle it..
116  */
117 good_area:
118         si_code = SEGV_ACCERR;
119         if (writeaccess) {
120                 if (!(vma->vm_flags & VM_WRITE))
121                         goto bad_area;
122         } else {
123                 if (!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE)))
124                         goto bad_area;
125         }
126
127         /*
128          * If for any reason at all we couldn't handle the fault,
129          * make sure we exit gracefully rather than endlessly redo
130          * the fault.
131          */
132 survive:
133         fault = handle_mm_fault(mm, vma, address, writeaccess);
134         if (unlikely(fault & VM_FAULT_ERROR)) {
135                 if (fault & VM_FAULT_OOM)
136                         goto out_of_memory;
137                 else if (fault & VM_FAULT_SIGBUS)
138                         goto do_sigbus;
139                 BUG();
140         }
141         if (fault & VM_FAULT_MAJOR)
142                 tsk->maj_flt++;
143         else
144                 tsk->min_flt++;
145
146         up_read(&mm->mmap_sem);
147         return;
148
149 /*
150  * Something tried to access memory that isn't in our memory map..
151  * Fix it, but check if it's kernel or user first..
152  */
153 bad_area:
154         up_read(&mm->mmap_sem);
155
156 bad_area_nosemaphore:
157         if (user_mode(regs)) {
158                 info.si_signo = SIGSEGV;
159                 info.si_errno = 0;
160                 info.si_code = si_code;
161                 info.si_addr = (void *) address;
162                 force_sig_info(SIGSEGV, &info, tsk);
163                 return;
164         }
165
166 no_context:
167         /* Are we prepared to handle this kernel fault?  */
168         if (fixup_exception(regs))
169                 return;
170
171         if (handle_trapped_io(regs, address))
172                 return;
173 /*
174  * Oops. The kernel tried to access some bad page. We'll have to
175  * terminate things with extreme prejudice.
176  *
177  */
178
179         bust_spinlocks(1);
180
181         if (oops_may_print()) {
182                 unsigned long page;
183
184                 if (address < PAGE_SIZE)
185                         printk(KERN_ALERT "Unable to handle kernel NULL "
186                                           "pointer dereference");
187                 else
188                         printk(KERN_ALERT "Unable to handle kernel paging "
189                                           "request");
190                 printk(" at virtual address %08lx\n", address);
191                 printk(KERN_ALERT "pc = %08lx\n", regs->pc);
192                 page = (unsigned long)get_TTB();
193                 if (page) {
194                         page = ((__typeof__(page) *)page)[address >> PGDIR_SHIFT];
195                         printk(KERN_ALERT "*pde = %08lx\n", page);
196                         if (page & _PAGE_PRESENT) {
197                                 page &= PAGE_MASK;
198                                 address &= 0x003ff000;
199                                 page = ((__typeof__(page) *)
200                                                 __va(page))[address >>
201                                                             PAGE_SHIFT];
202                                 printk(KERN_ALERT "*pte = %08lx\n", page);
203                         }
204                 }
205         }
206
207         die("Oops", regs, writeaccess);
208         bust_spinlocks(0);
209         do_exit(SIGKILL);
210
211 /*
212  * We ran out of memory, or some other thing happened to us that made
213  * us unable to handle the page fault gracefully.
214  */
215 out_of_memory:
216         up_read(&mm->mmap_sem);
217         if (is_global_init(current)) {
218                 yield();
219                 down_read(&mm->mmap_sem);
220                 goto survive;
221         }
222         printk("VM: killing process %s\n", tsk->comm);
223         if (user_mode(regs))
224                 do_group_exit(SIGKILL);
225         goto no_context;
226
227 do_sigbus:
228         up_read(&mm->mmap_sem);
229
230         /*
231          * Send a sigbus, regardless of whether we were in kernel
232          * or user mode.
233          */
234         info.si_signo = SIGBUS;
235         info.si_errno = 0;
236         info.si_code = BUS_ADRERR;
237         info.si_addr = (void *)address;
238         force_sig_info(SIGBUS, &info, tsk);
239
240         /* Kernel mode? Handle exceptions or die */
241         if (!user_mode(regs))
242                 goto no_context;
243 }
244
245 #ifdef CONFIG_SH_STORE_QUEUES
246 /*
247  * This is a special case for the SH-4 store queues, as pages for this
248  * space still need to be faulted in before it's possible to flush the
249  * store queue cache for writeout to the remapped region.
250  */
251 #define P3_ADDR_MAX             (P4SEG_STORE_QUE + 0x04000000)
252 #else
253 #define P3_ADDR_MAX             P4SEG
254 #endif
255
256 /*
257  * Called with interrupts disabled.
258  */
259 asmlinkage int __kprobes __do_page_fault(struct pt_regs *regs,
260                                          unsigned long writeaccess,
261                                          unsigned long address)
262 {
263         pgd_t *pgd;
264         pud_t *pud;
265         pmd_t *pmd;
266         pte_t *pte;
267         pte_t entry;
268
269 #ifdef CONFIG_SH_KGDB
270         if (kgdb_nofault && kgdb_bus_err_hook)
271                 kgdb_bus_err_hook();
272 #endif
273
274         /*
275          * We don't take page faults for P1, P2, and parts of P4, these
276          * are always mapped, whether it be due to legacy behaviour in
277          * 29-bit mode, or due to PMB configuration in 32-bit mode.
278          */
279         if (address >= P3SEG && address < P3_ADDR_MAX) {
280                 pgd = pgd_offset_k(address);
281         } else {
282                 if (unlikely(address >= TASK_SIZE || !current->mm))
283                         return 1;
284
285                 pgd = pgd_offset(current->mm, address);
286         }
287
288         pud = pud_offset(pgd, address);
289         if (pud_none_or_clear_bad(pud))
290                 return 1;
291         pmd = pmd_offset(pud, address);
292         if (pmd_none_or_clear_bad(pmd))
293                 return 1;
294
295         pte = pte_offset_kernel(pmd, address);
296         entry = *pte;
297         if (unlikely(pte_none(entry) || pte_not_present(entry)))
298                 return 1;
299         if (unlikely(writeaccess && !pte_write(entry)))
300                 return 1;
301
302         if (writeaccess)
303                 entry = pte_mkdirty(entry);
304         entry = pte_mkyoung(entry);
305
306 #if defined(CONFIG_CPU_SH4) && !defined(CONFIG_SMP)
307         /*
308          * ITLB is not affected by "ldtlb" instruction.
309          * So, we need to flush the entry by ourselves.
310          */
311         local_flush_tlb_one(get_asid(), address & PAGE_MASK);
312 #endif
313
314         set_pte(pte, entry);
315         update_mmu_cache(NULL, address, entry);
316
317         return 0;
318 }