Commit | Line | Data |
---|---|---|
867e359b CM |
1 | /* |
2 | * Copyright 2010 Tilera Corporation. All Rights Reserved. | |
3 | * | |
4 | * This program is free software; you can redistribute it and/or | |
5 | * modify it under the terms of the GNU General Public License | |
6 | * as published by the Free Software Foundation, version 2. | |
7 | * | |
8 | * This program is distributed in the hope that it will be useful, but | |
9 | * WITHOUT ANY WARRANTY; without even the implied warranty of | |
10 | * MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or | |
11 | * NON INFRINGEMENT. See the GNU General Public License for | |
12 | * more details. | |
13 | * | |
14 | * From i386 code copyright (C) 1995 Linus Torvalds | |
15 | */ | |
16 | ||
17 | #include <linux/signal.h> | |
18 | #include <linux/sched.h> | |
19 | #include <linux/kernel.h> | |
20 | #include <linux/errno.h> | |
21 | #include <linux/string.h> | |
22 | #include <linux/types.h> | |
23 | #include <linux/ptrace.h> | |
24 | #include <linux/mman.h> | |
25 | #include <linux/mm.h> | |
26 | #include <linux/smp.h> | |
867e359b CM |
27 | #include <linux/interrupt.h> |
28 | #include <linux/init.h> | |
29 | #include <linux/tty.h> | |
30 | #include <linux/vt_kern.h> /* For unblank_screen() */ | |
31 | #include <linux/highmem.h> | |
32 | #include <linux/module.h> | |
33 | #include <linux/kprobes.h> | |
34 | #include <linux/hugetlb.h> | |
35 | #include <linux/syscalls.h> | |
36 | #include <linux/uaccess.h> | |
3fa17c39 | 37 | #include <linux/kdebug.h> |
867e359b | 38 | |
867e359b CM |
39 | #include <asm/pgalloc.h> |
40 | #include <asm/sections.h> | |
0707ad30 CM |
41 | #include <asm/traps.h> |
42 | #include <asm/syscalls.h> | |
867e359b CM |
43 | |
44 | #include <arch/interrupts.h> | |
45 | ||
571d76ac CM |
46 | static noinline void force_sig_info_fault(const char *type, int si_signo, |
47 | int si_code, unsigned long address, | |
48 | int fault_num, | |
49 | struct task_struct *tsk, | |
50 | struct pt_regs *regs) | |
867e359b CM |
51 | { |
52 | siginfo_t info; | |
53 | ||
54 | if (unlikely(tsk->pid < 2)) { | |
55 | panic("Signal %d (code %d) at %#lx sent to %s!", | |
56 | si_signo, si_code & 0xffff, address, | |
a95f8817 | 57 | is_idle_task(tsk) ? "the idle task" : "init"); |
867e359b CM |
58 | } |
59 | ||
60 | info.si_signo = si_signo; | |
61 | info.si_errno = 0; | |
62 | info.si_code = si_code; | |
63 | info.si_addr = (void __user *)address; | |
64 | info.si_trapno = fault_num; | |
571d76ac | 65 | trace_unhandled_signal(type, regs, address, si_signo); |
867e359b CM |
66 | force_sig_info(si_signo, &info, tsk); |
67 | } | |
68 | ||
69 | #ifndef __tilegx__ | |
70 | /* | |
71 | * Synthesize the fault a PL0 process would get by doing a word-load of | |
d929b6ae | 72 | * an unaligned address or a high kernel address. |
867e359b | 73 | */ |
6b14e419 | 74 | SYSCALL_DEFINE1(cmpxchg_badaddr, unsigned long, address) |
867e359b | 75 | { |
6b14e419 CM |
76 | struct pt_regs *regs = current_pt_regs(); |
77 | ||
867e359b | 78 | if (address >= PAGE_OFFSET) |
571d76ac CM |
79 | force_sig_info_fault("atomic segfault", SIGSEGV, SEGV_MAPERR, |
80 | address, INT_DTLB_MISS, current, regs); | |
867e359b | 81 | else |
571d76ac CM |
82 | force_sig_info_fault("atomic alignment fault", SIGBUS, |
83 | BUS_ADRALN, address, | |
84 | INT_UNALIGN_DATA, current, regs); | |
867e359b CM |
85 | |
86 | /* | |
87 | * Adjust pc to point at the actual instruction, which is unusual | |
88 | * for syscalls normally, but is appropriate when we are claiming | |
89 | * that a syscall swint1 caused a page fault or bus error. | |
90 | */ | |
91 | regs->pc -= 8; | |
92 | ||
93 | /* | |
94 | * Mark this as a caller-save interrupt, like a normal page fault, | |
95 | * so that when we go through the signal handler path we will | |
96 | * properly restore r0, r1, and r2 for the signal handler arguments. | |
97 | */ | |
98 | regs->flags |= PT_FLAGS_CALLER_SAVES; | |
99 | ||
100 | return 0; | |
101 | } | |
102 | #endif | |
103 | ||
104 | static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address) | |
105 | { | |
106 | unsigned index = pgd_index(address); | |
107 | pgd_t *pgd_k; | |
108 | pud_t *pud, *pud_k; | |
109 | pmd_t *pmd, *pmd_k; | |
110 | ||
111 | pgd += index; | |
112 | pgd_k = init_mm.pgd + index; | |
113 | ||
114 | if (!pgd_present(*pgd_k)) | |
115 | return NULL; | |
116 | ||
117 | pud = pud_offset(pgd, address); | |
118 | pud_k = pud_offset(pgd_k, address); | |
119 | if (!pud_present(*pud_k)) | |
120 | return NULL; | |
121 | ||
122 | pmd = pmd_offset(pud, address); | |
123 | pmd_k = pmd_offset(pud_k, address); | |
124 | if (!pmd_present(*pmd_k)) | |
125 | return NULL; | |
1182b69c | 126 | if (!pmd_present(*pmd)) |
867e359b | 127 | set_pmd(pmd, *pmd_k); |
1182b69c | 128 | else |
867e359b CM |
129 | BUG_ON(pmd_ptfn(*pmd) != pmd_ptfn(*pmd_k)); |
130 | return pmd_k; | |
131 | } | |
132 | ||
133 | /* | |
51bcdf88 | 134 | * Handle a fault on the vmalloc area. |
867e359b CM |
135 | */ |
136 | static inline int vmalloc_fault(pgd_t *pgd, unsigned long address) | |
137 | { | |
138 | pmd_t *pmd_k; | |
139 | pte_t *pte_k; | |
140 | ||
141 | /* Make sure we are in vmalloc area */ | |
142 | if (!(address >= VMALLOC_START && address < VMALLOC_END)) | |
143 | return -1; | |
144 | ||
145 | /* | |
146 | * Synchronize this task's top level page-table | |
147 | * with the 'reference' page table. | |
148 | */ | |
149 | pmd_k = vmalloc_sync_one(pgd, address); | |
150 | if (!pmd_k) | |
151 | return -1; | |
867e359b CM |
152 | pte_k = pte_offset_kernel(pmd_k, address); |
153 | if (!pte_present(*pte_k)) | |
154 | return -1; | |
155 | return 0; | |
156 | } | |
157 | ||
158 | /* Wait until this PTE has completed migration. */ | |
159 | static void wait_for_migration(pte_t *pte) | |
160 | { | |
161 | if (pte_migrating(*pte)) { | |
162 | /* | |
163 | * Wait until the migrater fixes up this pte. | |
164 | * We scale the loop count by the clock rate so we'll wait for | |
165 | * a few seconds here. | |
166 | */ | |
167 | int retries = 0; | |
168 | int bound = get_clock_rate(); | |
169 | while (pte_migrating(*pte)) { | |
170 | barrier(); | |
171 | if (++retries > bound) | |
f4743673 | 172 | panic("Hit migrating PTE (%#llx) and page PFN %#lx still migrating", |
867e359b CM |
173 | pte->val, pte_pfn(*pte)); |
174 | } | |
175 | } | |
176 | } | |
177 | ||
178 | /* | |
179 | * It's not generally safe to use "current" to get the page table pointer, | |
180 | * since we might be running an oprofile interrupt in the middle of a | |
181 | * task switch. | |
182 | */ | |
183 | static pgd_t *get_current_pgd(void) | |
184 | { | |
185 | HV_Context ctx = hv_inquire_context(); | |
186 | unsigned long pgd_pfn = ctx.page_table >> PAGE_SHIFT; | |
187 | struct page *pgd_page = pfn_to_page(pgd_pfn); | |
621b1955 | 188 | BUG_ON(PageHighMem(pgd_page)); |
867e359b CM |
189 | return (pgd_t *) __va(ctx.page_table); |
190 | } | |
191 | ||
192 | /* | |
193 | * We can receive a page fault from a migrating PTE at any time. | |
194 | * Handle it by just waiting until the fault resolves. | |
195 | * | |
196 | * It's also possible to get a migrating kernel PTE that resolves | |
197 | * itself during the downcall from hypervisor to Linux. We just check | |
198 | * here to see if the PTE seems valid, and if so we retry it. | |
199 | * | |
200 | * NOTE! We MUST NOT take any locks for this case. We may be in an | |
201 | * interrupt or a critical region, and must do as little as possible. | |
202 | * Similarly, we can't use atomic ops here, since we may be handling a | |
203 | * fault caused by an atomic op access. | |
48292738 CM |
204 | * |
205 | * If we find a migrating PTE while we're in an NMI context, and we're | |
206 | * at a PC that has a registered exception handler, we don't wait, | |
207 | * since this thread may (e.g.) have been interrupted while migrating | |
208 | * its own stack, which would then cause us to self-deadlock. | |
867e359b CM |
209 | */ |
210 | static int handle_migrating_pte(pgd_t *pgd, int fault_num, | |
48292738 | 211 | unsigned long address, unsigned long pc, |
867e359b CM |
212 | int is_kernel_mode, int write) |
213 | { | |
214 | pud_t *pud; | |
215 | pmd_t *pmd; | |
216 | pte_t *pte; | |
217 | pte_t pteval; | |
218 | ||
219 | if (pgd_addr_invalid(address)) | |
220 | return 0; | |
221 | ||
222 | pgd += pgd_index(address); | |
223 | pud = pud_offset(pgd, address); | |
224 | if (!pud || !pud_present(*pud)) | |
225 | return 0; | |
226 | pmd = pmd_offset(pud, address); | |
227 | if (!pmd || !pmd_present(*pmd)) | |
228 | return 0; | |
229 | pte = pmd_huge_page(*pmd) ? ((pte_t *)pmd) : | |
230 | pte_offset_kernel(pmd, address); | |
231 | pteval = *pte; | |
232 | if (pte_migrating(pteval)) { | |
48292738 CM |
233 | if (in_nmi() && search_exception_tables(pc)) |
234 | return 0; | |
867e359b CM |
235 | wait_for_migration(pte); |
236 | return 1; | |
237 | } | |
238 | ||
239 | if (!is_kernel_mode || !pte_present(pteval)) | |
240 | return 0; | |
241 | if (fault_num == INT_ITLB_MISS) { | |
242 | if (pte_exec(pteval)) | |
243 | return 1; | |
244 | } else if (write) { | |
245 | if (pte_write(pteval)) | |
246 | return 1; | |
247 | } else { | |
248 | if (pte_read(pteval)) | |
249 | return 1; | |
250 | } | |
251 | ||
252 | return 0; | |
253 | } | |
254 | ||
255 | /* | |
256 | * This routine is responsible for faulting in user pages. | |
257 | * It passes the work off to one of the appropriate routines. | |
258 | * It returns true if the fault was successfully handled. | |
259 | */ | |
260 | static int handle_page_fault(struct pt_regs *regs, | |
261 | int fault_num, | |
262 | int is_page_fault, | |
263 | unsigned long address, | |
264 | int write) | |
265 | { | |
266 | struct task_struct *tsk; | |
267 | struct mm_struct *mm; | |
268 | struct vm_area_struct *vma; | |
269 | unsigned long stack_offset; | |
270 | int fault; | |
271 | int si_code; | |
272 | int is_kernel_mode; | |
273 | pgd_t *pgd; | |
4ce6bea2 | 274 | unsigned int flags; |
867e359b CM |
275 | |
276 | /* on TILE, protection faults are always writes */ | |
277 | if (!is_page_fault) | |
278 | write = 1; | |
279 | ||
759496ba | 280 | flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE; |
4ce6bea2 | 281 | |
051168df | 282 | is_kernel_mode = !user_mode(regs); |
867e359b CM |
283 | |
284 | tsk = validate_current(); | |
285 | ||
286 | /* | |
287 | * Check to see if we might be overwriting the stack, and bail | |
288 | * out if so. The page fault code is a relatively likely | |
289 | * place to get trapped in an infinite regress, and once we | |
290 | * overwrite the whole stack, it becomes very hard to recover. | |
291 | */ | |
292 | stack_offset = stack_pointer & (THREAD_SIZE-1); | |
293 | if (stack_offset < THREAD_SIZE / 8) { | |
f4743673 | 294 | pr_alert("Potential stack overrun: sp %#lx\n", stack_pointer); |
867e359b | 295 | show_regs(regs); |
0707ad30 | 296 | pr_alert("Killing current process %d/%s\n", |
f4743673 | 297 | tsk->pid, tsk->comm); |
867e359b CM |
298 | do_group_exit(SIGKILL); |
299 | } | |
300 | ||
301 | /* | |
302 | * Early on, we need to check for migrating PTE entries; | |
303 | * see homecache.c. If we find a migrating PTE, we wait until | |
25985edc | 304 | * the backing page claims to be done migrating, then we proceed. |
867e359b CM |
305 | * For kernel PTEs, we rewrite the PTE and return and retry. |
306 | * Otherwise, we treat the fault like a normal "no PTE" fault, | |
307 | * rather than trying to patch up the existing PTE. | |
308 | */ | |
309 | pgd = get_current_pgd(); | |
48292738 | 310 | if (handle_migrating_pte(pgd, fault_num, address, regs->pc, |
867e359b CM |
311 | is_kernel_mode, write)) |
312 | return 1; | |
313 | ||
314 | si_code = SEGV_MAPERR; | |
315 | ||
316 | /* | |
317 | * We fault-in kernel-space virtual memory on-demand. The | |
318 | * 'reference' page table is init_mm.pgd. | |
319 | * | |
320 | * NOTE! We MUST NOT take any locks for this case. We may | |
321 | * be in an interrupt or a critical region, and should | |
322 | * only copy the information from the master page table, | |
323 | * nothing more. | |
324 | * | |
325 | * This verifies that the fault happens in kernel space | |
326 | * and that the fault was not a protection fault. | |
327 | */ | |
328 | if (unlikely(address >= TASK_SIZE && | |
329 | !is_arch_mappable_range(address, 0))) { | |
330 | if (is_kernel_mode && is_page_fault && | |
331 | vmalloc_fault(pgd, address) >= 0) | |
332 | return 1; | |
333 | /* | |
334 | * Don't take the mm semaphore here. If we fixup a prefetch | |
335 | * fault we could otherwise deadlock. | |
336 | */ | |
337 | mm = NULL; /* happy compiler */ | |
338 | vma = NULL; | |
339 | goto bad_area_nosemaphore; | |
340 | } | |
341 | ||
342 | /* | |
343 | * If we're trying to touch user-space addresses, we must | |
344 | * be either at PL0, or else with interrupts enabled in the | |
b230ff2d CM |
345 | * kernel, so either way we can re-enable interrupts here |
346 | * unless we are doing atomic access to user space with | |
347 | * interrupts disabled. | |
867e359b | 348 | */ |
b230ff2d CM |
349 | if (!(regs->flags & PT_FLAGS_DISABLE_IRQ)) |
350 | local_irq_enable(); | |
867e359b CM |
351 | |
352 | mm = tsk->mm; | |
353 | ||
354 | /* | |
355 | * If we're in an interrupt, have no user context or are running in an | |
70ffdb93 | 356 | * region with pagefaults disabled then we must not take the fault. |
867e359b | 357 | */ |
70ffdb93 | 358 | if (pagefault_disabled() || !mm) { |
867e359b CM |
359 | vma = NULL; /* happy compiler */ |
360 | goto bad_area_nosemaphore; | |
361 | } | |
362 | ||
759496ba JW |
363 | if (!is_kernel_mode) |
364 | flags |= FAULT_FLAG_USER; | |
365 | ||
867e359b CM |
366 | /* |
367 | * When running in the kernel we expect faults to occur only to | |
368 | * addresses in user space. All other faults represent errors in the | |
369 | * kernel and should generate an OOPS. Unfortunately, in the case of an | |
370 | * erroneous fault occurring in a code path which already holds mmap_sem | |
371 | * we will deadlock attempting to validate the fault against the | |
372 | * address space. Luckily the kernel only validly references user | |
373 | * space from well defined areas of code, which are listed in the | |
374 | * exceptions table. | |
375 | * | |
376 | * As the vast majority of faults will be valid we will only perform | |
377 | * the source reference check when there is a possibility of a deadlock. | |
378 | * Attempt to lock the address space, if we cannot we then validate the | |
379 | * source. If this is invalid we can skip the address space check, | |
380 | * thus avoiding the deadlock. | |
381 | */ | |
382 | if (!down_read_trylock(&mm->mmap_sem)) { | |
383 | if (is_kernel_mode && | |
384 | !search_exception_tables(regs->pc)) { | |
385 | vma = NULL; /* happy compiler */ | |
386 | goto bad_area_nosemaphore; | |
387 | } | |
4ce6bea2 KC |
388 | |
389 | retry: | |
867e359b CM |
390 | down_read(&mm->mmap_sem); |
391 | } | |
392 | ||
393 | vma = find_vma(mm, address); | |
394 | if (!vma) | |
395 | goto bad_area; | |
396 | if (vma->vm_start <= address) | |
397 | goto good_area; | |
398 | if (!(vma->vm_flags & VM_GROWSDOWN)) | |
399 | goto bad_area; | |
400 | if (regs->sp < PAGE_OFFSET) { | |
401 | /* | |
402 | * accessing the stack below sp is always a bug. | |
403 | */ | |
404 | if (address < regs->sp) | |
405 | goto bad_area; | |
406 | } | |
407 | if (expand_stack(vma, address)) | |
408 | goto bad_area; | |
409 | ||
410 | /* | |
411 | * Ok, we have a good vm_area for this memory access, so | |
412 | * we can handle it.. | |
413 | */ | |
414 | good_area: | |
415 | si_code = SEGV_ACCERR; | |
416 | if (fault_num == INT_ITLB_MISS) { | |
417 | if (!(vma->vm_flags & VM_EXEC)) | |
418 | goto bad_area; | |
419 | } else if (write) { | |
420 | #ifdef TEST_VERIFY_AREA | |
421 | if (!is_page_fault && regs->cs == KERNEL_CS) | |
f4743673 | 422 | pr_err("WP fault at " REGFMT "\n", regs->eip); |
867e359b CM |
423 | #endif |
424 | if (!(vma->vm_flags & VM_WRITE)) | |
425 | goto bad_area; | |
759496ba | 426 | flags |= FAULT_FLAG_WRITE; |
867e359b CM |
427 | } else { |
428 | if (!is_page_fault || !(vma->vm_flags & VM_READ)) | |
429 | goto bad_area; | |
430 | } | |
431 | ||
867e359b CM |
432 | /* |
433 | * If for any reason at all we couldn't handle the fault, | |
434 | * make sure we exit gracefully rather than endlessly redo | |
435 | * the fault. | |
436 | */ | |
dcddffd4 | 437 | fault = handle_mm_fault(vma, address, flags); |
4ce6bea2 KC |
438 | |
439 | if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)) | |
440 | return 0; | |
441 | ||
867e359b CM |
442 | if (unlikely(fault & VM_FAULT_ERROR)) { |
443 | if (fault & VM_FAULT_OOM) | |
444 | goto out_of_memory; | |
33692f27 LT |
445 | else if (fault & VM_FAULT_SIGSEGV) |
446 | goto bad_area; | |
867e359b CM |
447 | else if (fault & VM_FAULT_SIGBUS) |
448 | goto do_sigbus; | |
449 | BUG(); | |
450 | } | |
4ce6bea2 KC |
451 | if (flags & FAULT_FLAG_ALLOW_RETRY) { |
452 | if (fault & VM_FAULT_MAJOR) | |
453 | tsk->maj_flt++; | |
454 | else | |
455 | tsk->min_flt++; | |
456 | if (fault & VM_FAULT_RETRY) { | |
457 | flags &= ~FAULT_FLAG_ALLOW_RETRY; | |
45cac65b | 458 | flags |= FAULT_FLAG_TRIED; |
4ce6bea2 KC |
459 | |
460 | /* | |
461 | * No need to up_read(&mm->mmap_sem) as we would | |
462 | * have already released it in __lock_page_or_retry | |
463 | * in mm/filemap.c. | |
464 | */ | |
465 | goto retry; | |
466 | } | |
467 | } | |
867e359b | 468 | |
867e359b | 469 | #if CHIP_HAS_TILE_DMA() |
d7c96611 CM |
470 | /* If this was a DMA TLB fault, restart the DMA engine. */ |
471 | switch (fault_num) { | |
867e359b CM |
472 | case INT_DMATLB_MISS: |
473 | case INT_DMATLB_MISS_DWNCL: | |
474 | case INT_DMATLB_ACCESS: | |
475 | case INT_DMATLB_ACCESS_DWNCL: | |
476 | __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__REQUEST_MASK); | |
477 | break; | |
867e359b | 478 | } |
0707ad30 | 479 | #endif |
867e359b CM |
480 | |
481 | up_read(&mm->mmap_sem); | |
482 | return 1; | |
483 | ||
484 | /* | |
485 | * Something tried to access memory that isn't in our memory map.. | |
486 | * Fix it, but check if it's kernel or user first.. | |
487 | */ | |
488 | bad_area: | |
489 | up_read(&mm->mmap_sem); | |
490 | ||
491 | bad_area_nosemaphore: | |
492 | /* User mode accesses just cause a SIGSEGV */ | |
493 | if (!is_kernel_mode) { | |
494 | /* | |
495 | * It's possible to have interrupts off here. | |
496 | */ | |
497 | local_irq_enable(); | |
498 | ||
571d76ac CM |
499 | force_sig_info_fault("segfault", SIGSEGV, si_code, address, |
500 | fault_num, tsk, regs); | |
867e359b CM |
501 | return 0; |
502 | } | |
503 | ||
504 | no_context: | |
505 | /* Are we prepared to handle this kernel fault? */ | |
506 | if (fixup_exception(regs)) | |
507 | return 0; | |
508 | ||
509 | /* | |
510 | * Oops. The kernel tried to access some bad page. We'll have to | |
511 | * terminate things with extreme prejudice. | |
512 | */ | |
513 | ||
514 | bust_spinlocks(1); | |
515 | ||
516 | /* FIXME: no lookup_address() yet */ | |
517 | #ifdef SUPPORT_LOOKUP_ADDRESS | |
518 | if (fault_num == INT_ITLB_MISS) { | |
519 | pte_t *pte = lookup_address(address); | |
520 | ||
521 | if (pte && pte_present(*pte) && !pte_exec_kernel(*pte)) | |
f4743673 JP |
522 | pr_crit("kernel tried to execute non-executable page - exploit attempt? (uid: %d)\n", |
523 | current->uid); | |
867e359b CM |
524 | } |
525 | #endif | |
526 | if (address < PAGE_SIZE) | |
0707ad30 | 527 | pr_alert("Unable to handle kernel NULL pointer dereference\n"); |
867e359b | 528 | else |
0707ad30 | 529 | pr_alert("Unable to handle kernel paging request\n"); |
f4743673 | 530 | pr_alert(" at virtual address " REGFMT ", pc " REGFMT "\n", |
0707ad30 | 531 | address, regs->pc); |
867e359b CM |
532 | |
533 | show_regs(regs); | |
534 | ||
535 | if (unlikely(tsk->pid < 2)) { | |
536 | panic("Kernel page fault running %s!", | |
a95f8817 | 537 | is_idle_task(tsk) ? "the idle task" : "init"); |
867e359b CM |
538 | } |
539 | ||
540 | /* | |
541 | * More FIXME: we should probably copy the i386 here and | |
542 | * implement a generic die() routine. Not today. | |
543 | */ | |
544 | #ifdef SUPPORT_DIE | |
545 | die("Oops", regs); | |
546 | #endif | |
547 | bust_spinlocks(1); | |
548 | ||
549 | do_group_exit(SIGKILL); | |
550 | ||
551 | /* | |
552 | * We ran out of memory, or some other thing happened to us that made | |
553 | * us unable to handle the page fault gracefully. | |
554 | */ | |
555 | out_of_memory: | |
556 | up_read(&mm->mmap_sem); | |
609838cf JW |
557 | if (is_kernel_mode) |
558 | goto no_context; | |
559 | pagefault_out_of_memory(); | |
560 | return 0; | |
867e359b CM |
561 | |
562 | do_sigbus: | |
563 | up_read(&mm->mmap_sem); | |
564 | ||
565 | /* Kernel mode? Handle exceptions or die */ | |
566 | if (is_kernel_mode) | |
567 | goto no_context; | |
568 | ||
571d76ac CM |
569 | force_sig_info_fault("bus error", SIGBUS, BUS_ADRERR, address, |
570 | fault_num, tsk, regs); | |
867e359b CM |
571 | return 0; |
572 | } | |
573 | ||
574 | #ifndef __tilegx__ | |
575 | ||
867e359b | 576 | /* We must release ICS before panicking or we won't get anywhere. */ |
f4743673 JP |
577 | #define ics_panic(fmt, ...) \ |
578 | do { \ | |
579 | __insn_mtspr(SPR_INTERRUPT_CRITICAL_SECTION, 0); \ | |
580 | panic(fmt, ##__VA_ARGS__); \ | |
867e359b CM |
581 | } while (0) |
582 | ||
867e359b CM |
583 | /* |
584 | * When we take an ITLB or DTLB fault or access violation in the | |
585 | * supervisor while the critical section bit is set, the hypervisor is | |
a78c942d | 586 | * reluctant to write new values into the EX_CONTEXT_K_x registers, |
867e359b CM |
587 | * since that might indicate we have not yet squirreled the SPR |
588 | * contents away and can thus safely take a recursive interrupt. | |
a78c942d | 589 | * Accordingly, the hypervisor passes us the PC via SYSTEM_SAVE_K_2. |
c745a8a1 CM |
590 | * |
591 | * Note that this routine is called before homecache_tlb_defer_enter(), | |
592 | * which means that we can properly unlock any atomics that might | |
593 | * be used there (good), but also means we must be very sensitive | |
594 | * to not touch any data structures that might be located in memory | |
595 | * that could migrate, as we could be entering the kernel on a dataplane | |
596 | * cpu that has been deferring kernel TLB updates. This means, for | |
597 | * example, that we can't migrate init_mm or its pgd. | |
867e359b CM |
598 | */ |
599 | struct intvec_state do_page_fault_ics(struct pt_regs *regs, int fault_num, | |
600 | unsigned long address, | |
601 | unsigned long info) | |
602 | { | |
603 | unsigned long pc = info & ~1; | |
604 | int write = info & 1; | |
605 | pgd_t *pgd = get_current_pgd(); | |
606 | ||
607 | /* Retval is 1 at first since we will handle the fault fully. */ | |
608 | struct intvec_state state = { | |
609 | do_page_fault, fault_num, address, write, 1 | |
610 | }; | |
611 | ||
612 | /* Validate that we are plausibly in the right routine. */ | |
613 | if ((pc & 0x7) != 0 || pc < PAGE_OFFSET || | |
614 | (fault_num != INT_DTLB_MISS && | |
615 | fault_num != INT_DTLB_ACCESS)) { | |
616 | unsigned long old_pc = regs->pc; | |
617 | regs->pc = pc; | |
f4743673 | 618 | ics_panic("Bad ICS page fault args: old PC %#lx, fault %d/%d at %#lx", |
867e359b CM |
619 | old_pc, fault_num, write, address); |
620 | } | |
621 | ||
622 | /* We might be faulting on a vmalloc page, so check that first. */ | |
623 | if (fault_num != INT_DTLB_ACCESS && vmalloc_fault(pgd, address) >= 0) | |
624 | return state; | |
625 | ||
626 | /* | |
627 | * If we faulted with ICS set in sys_cmpxchg, we are providing | |
628 | * a user syscall service that should generate a signal on | |
629 | * fault. We didn't set up a kernel stack on initial entry to | |
630 | * sys_cmpxchg, but instead had one set up by the fault, which | |
631 | * (because sys_cmpxchg never releases ICS) came to us via the | |
a78c942d | 632 | * SYSTEM_SAVE_K_2 mechanism, and thus EX_CONTEXT_K_[01] are |
867e359b CM |
633 | * still referencing the original user code. We release the |
634 | * atomic lock and rewrite pt_regs so that it appears that we | |
635 | * came from user-space directly, and after we finish the | |
636 | * fault we'll go back to user space and re-issue the swint. | |
637 | * This way the backtrace information is correct if we need to | |
638 | * emit a stack dump at any point while handling this. | |
639 | * | |
640 | * Must match register use in sys_cmpxchg(). | |
641 | */ | |
642 | if (pc >= (unsigned long) sys_cmpxchg && | |
643 | pc < (unsigned long) __sys_cmpxchg_end) { | |
644 | #ifdef CONFIG_SMP | |
645 | /* Don't unlock before we could have locked. */ | |
646 | if (pc >= (unsigned long)__sys_cmpxchg_grab_lock) { | |
647 | int *lock_ptr = (int *)(regs->regs[ATOMIC_LOCK_REG]); | |
648 | __atomic_fault_unlock(lock_ptr); | |
649 | } | |
650 | #endif | |
651 | regs->sp = regs->regs[27]; | |
652 | } | |
653 | ||
654 | /* | |
655 | * We can also fault in the atomic assembly, in which | |
656 | * case we use the exception table to do the first-level fixup. | |
657 | * We may re-fixup again in the real fault handler if it | |
658 | * turns out the faulting address is just bad, and not, | |
659 | * for example, migrating. | |
660 | */ | |
661 | else if (pc >= (unsigned long) __start_atomic_asm_code && | |
662 | pc < (unsigned long) __end_atomic_asm_code) { | |
663 | const struct exception_table_entry *fixup; | |
664 | #ifdef CONFIG_SMP | |
665 | /* Unlock the atomic lock. */ | |
666 | int *lock_ptr = (int *)(regs->regs[ATOMIC_LOCK_REG]); | |
667 | __atomic_fault_unlock(lock_ptr); | |
668 | #endif | |
669 | fixup = search_exception_tables(pc); | |
670 | if (!fixup) | |
f4743673 JP |
671 | ics_panic("ICS atomic fault not in table: PC %#lx, fault %d", |
672 | pc, fault_num); | |
867e359b CM |
673 | regs->pc = fixup->fixup; |
674 | regs->ex1 = PL_ICS_EX1(KERNEL_PL, 0); | |
675 | } | |
676 | ||
867e359b CM |
677 | /* |
678 | * Now that we have released the atomic lock (if necessary), | |
679 | * it's safe to spin if the PTE that caused the fault was migrating. | |
680 | */ | |
681 | if (fault_num == INT_DTLB_ACCESS) | |
682 | write = 1; | |
48292738 | 683 | if (handle_migrating_pte(pgd, fault_num, address, pc, 1, write)) |
867e359b CM |
684 | return state; |
685 | ||
686 | /* Return zero so that we continue on with normal fault handling. */ | |
687 | state.retval = 0; | |
688 | return state; | |
689 | } | |
690 | ||
691 | #endif /* !__tilegx__ */ | |
692 | ||
693 | /* | |
694 | * This routine handles page faults. It determines the address, and the | |
695 | * problem, and then passes it handle_page_fault() for normal DTLB and | |
696 | * ITLB issues, and for DMA or SN processor faults when we are in user | |
697 | * space. For the latter, if we're in kernel mode, we just save the | |
698 | * interrupt away appropriately and return immediately. We can't do | |
699 | * page faults for user code while in kernel mode. | |
700 | */ | |
5316a64c CM |
701 | static inline void __do_page_fault(struct pt_regs *regs, int fault_num, |
702 | unsigned long address, unsigned long write) | |
867e359b CM |
703 | { |
704 | int is_page_fault; | |
705 | ||
3fa17c39 TL |
706 | #ifdef CONFIG_KPROBES |
707 | /* | |
708 | * This is to notify the fault handler of the kprobes. The | |
709 | * exception code is redundant as it is also carried in REGS, | |
710 | * but we pass it anyhow. | |
711 | */ | |
712 | if (notify_die(DIE_PAGE_FAULT, "page fault", regs, -1, | |
713 | regs->faultnum, SIGSEGV) == NOTIFY_STOP) | |
5316a64c | 714 | return; |
3fa17c39 TL |
715 | #endif |
716 | ||
2f9ac29e CM |
717 | #ifdef __tilegx__ |
718 | /* | |
719 | * We don't need early do_page_fault_ics() support, since unlike | |
720 | * Pro we don't need to worry about unlocking the atomic locks. | |
721 | * There is only one current case in GX where we touch any memory | |
722 | * under ICS other than our own kernel stack, and we handle that | |
723 | * here. (If we crash due to trying to touch our own stack, | |
724 | * we're in too much trouble for C code to help out anyway.) | |
725 | */ | |
726 | if (write & ~1) { | |
727 | unsigned long pc = write & ~1; | |
728 | if (pc >= (unsigned long) __start_unalign_asm_code && | |
729 | pc < (unsigned long) __end_unalign_asm_code) { | |
730 | struct thread_info *ti = current_thread_info(); | |
731 | /* | |
732 | * Our EX_CONTEXT is still what it was from the | |
733 | * initial unalign exception, but now we've faulted | |
734 | * on the JIT page. We would like to complete the | |
735 | * page fault however is appropriate, and then retry | |
736 | * the instruction that caused the unalign exception. | |
737 | * Our state has been "corrupted" by setting the low | |
738 | * bit in "sp", and stashing r0..r3 in the | |
739 | * thread_info area, so we revert all of that, then | |
740 | * continue as if this were a normal page fault. | |
741 | */ | |
742 | regs->sp &= ~1UL; | |
743 | regs->regs[0] = ti->unalign_jit_tmp[0]; | |
744 | regs->regs[1] = ti->unalign_jit_tmp[1]; | |
745 | regs->regs[2] = ti->unalign_jit_tmp[2]; | |
746 | regs->regs[3] = ti->unalign_jit_tmp[3]; | |
747 | write &= 1; | |
748 | } else { | |
749 | pr_alert("%s/%d: ICS set at page fault at %#lx: %#lx\n", | |
750 | current->comm, current->pid, pc, address); | |
751 | show_regs(regs); | |
752 | do_group_exit(SIGKILL); | |
2f9ac29e CM |
753 | } |
754 | } | |
755 | #else | |
867e359b CM |
756 | /* This case should have been handled by do_page_fault_ics(). */ |
757 | BUG_ON(write & ~1); | |
2f9ac29e | 758 | #endif |
867e359b CM |
759 | |
760 | #if CHIP_HAS_TILE_DMA() | |
761 | /* | |
762 | * If it's a DMA fault, suspend the transfer while we're | |
763 | * handling the miss; we'll restart after it's handled. If we | |
764 | * don't suspend, it's possible that this process could swap | |
765 | * out and back in, and restart the engine since the DMA is | |
766 | * still 'running'. | |
767 | */ | |
768 | if (fault_num == INT_DMATLB_MISS || | |
769 | fault_num == INT_DMATLB_ACCESS || | |
770 | fault_num == INT_DMATLB_MISS_DWNCL || | |
771 | fault_num == INT_DMATLB_ACCESS_DWNCL) { | |
772 | __insn_mtspr(SPR_DMA_CTR, SPR_DMA_CTR__SUSPEND_MASK); | |
773 | while (__insn_mfspr(SPR_DMA_USER_STATUS) & | |
774 | SPR_DMA_STATUS__BUSY_MASK) | |
775 | ; | |
776 | } | |
777 | #endif | |
778 | ||
779 | /* Validate fault num and decide if this is a first-time page fault. */ | |
780 | switch (fault_num) { | |
781 | case INT_ITLB_MISS: | |
782 | case INT_DTLB_MISS: | |
783 | #if CHIP_HAS_TILE_DMA() | |
784 | case INT_DMATLB_MISS: | |
785 | case INT_DMATLB_MISS_DWNCL: | |
867e359b CM |
786 | #endif |
787 | is_page_fault = 1; | |
788 | break; | |
789 | ||
790 | case INT_DTLB_ACCESS: | |
791 | #if CHIP_HAS_TILE_DMA() | |
792 | case INT_DMATLB_ACCESS: | |
793 | case INT_DMATLB_ACCESS_DWNCL: | |
794 | #endif | |
795 | is_page_fault = 0; | |
796 | break; | |
797 | ||
798 | default: | |
799 | panic("Bad fault number %d in do_page_fault", fault_num); | |
800 | } | |
801 | ||
d7c96611 | 802 | #if CHIP_HAS_TILE_DMA() |
051168df | 803 | if (!user_mode(regs)) { |
867e359b CM |
804 | struct async_tlb *async; |
805 | switch (fault_num) { | |
806 | #if CHIP_HAS_TILE_DMA() | |
807 | case INT_DMATLB_MISS: | |
808 | case INT_DMATLB_ACCESS: | |
809 | case INT_DMATLB_MISS_DWNCL: | |
810 | case INT_DMATLB_ACCESS_DWNCL: | |
811 | async = ¤t->thread.dma_async_tlb; | |
812 | break; | |
867e359b CM |
813 | #endif |
814 | default: | |
815 | async = NULL; | |
816 | } | |
817 | if (async) { | |
818 | ||
819 | /* | |
820 | * No vmalloc check required, so we can allow | |
821 | * interrupts immediately at this point. | |
822 | */ | |
823 | local_irq_enable(); | |
824 | ||
825 | set_thread_flag(TIF_ASYNC_TLB); | |
826 | if (async->fault_num != 0) { | |
f4743673 | 827 | panic("Second async fault %d; old fault was %d (%#lx/%ld)", |
867e359b CM |
828 | fault_num, async->fault_num, |
829 | address, write); | |
830 | } | |
831 | BUG_ON(fault_num == 0); | |
832 | async->fault_num = fault_num; | |
833 | async->is_fault = is_page_fault; | |
834 | async->is_write = write; | |
835 | async->address = address; | |
5316a64c | 836 | return; |
867e359b CM |
837 | } |
838 | } | |
313ce674 | 839 | #endif |
867e359b CM |
840 | |
841 | handle_page_fault(regs, fault_num, is_page_fault, address, write); | |
5316a64c | 842 | } |
49e4e156 | 843 | |
5316a64c CM |
844 | void do_page_fault(struct pt_regs *regs, int fault_num, |
845 | unsigned long address, unsigned long write) | |
846 | { | |
5316a64c | 847 | __do_page_fault(regs, fault_num, address, write); |
867e359b CM |
848 | } |
849 | ||
d7c96611 | 850 | #if CHIP_HAS_TILE_DMA() |
867e359b | 851 | /* |
d7c96611 CM |
852 | * This routine effectively re-issues asynchronous page faults |
853 | * when we are returning to user space. | |
867e359b | 854 | */ |
d7c96611 | 855 | void do_async_page_fault(struct pt_regs *regs) |
867e359b | 856 | { |
d7c96611 CM |
857 | struct async_tlb *async = ¤t->thread.dma_async_tlb; |
858 | ||
859 | /* | |
860 | * Clear thread flag early. If we re-interrupt while processing | |
861 | * code here, we will reset it and recall this routine before | |
862 | * returning to user space. | |
863 | */ | |
864 | clear_thread_flag(TIF_ASYNC_TLB); | |
865 | ||
867e359b CM |
866 | if (async->fault_num) { |
867 | /* | |
868 | * Clear async->fault_num before calling the page-fault | |
869 | * handler so that if we re-interrupt before returning | |
870 | * from the function we have somewhere to put the | |
871 | * information from the new interrupt. | |
872 | */ | |
873 | int fault_num = async->fault_num; | |
874 | async->fault_num = 0; | |
875 | handle_page_fault(regs, fault_num, async->is_fault, | |
876 | async->address, async->is_write); | |
877 | } | |
878 | } | |
d7c96611 | 879 | #endif /* CHIP_HAS_TILE_DMA() */ |
313ce674 | 880 | |
867e359b CM |
881 | |
882 | void vmalloc_sync_all(void) | |
883 | { | |
884 | #ifdef __tilegx__ | |
885 | /* Currently all L1 kernel pmd's are static and shared. */ | |
e5f7bd43 CM |
886 | BUILD_BUG_ON(pgd_index(VMALLOC_END - PAGE_SIZE) != |
887 | pgd_index(VMALLOC_START)); | |
867e359b CM |
888 | #else |
889 | /* | |
890 | * Note that races in the updates of insync and start aren't | |
891 | * problematic: insync can only get set bits added, and updates to | |
892 | * start are only improving performance (without affecting correctness | |
893 | * if undone). | |
894 | */ | |
895 | static DECLARE_BITMAP(insync, PTRS_PER_PGD); | |
896 | static unsigned long start = PAGE_OFFSET; | |
897 | unsigned long address; | |
898 | ||
899 | BUILD_BUG_ON(PAGE_OFFSET & ~PGDIR_MASK); | |
900 | for (address = start; address >= PAGE_OFFSET; address += PGDIR_SIZE) { | |
901 | if (!test_bit(pgd_index(address), insync)) { | |
902 | unsigned long flags; | |
903 | struct list_head *pos; | |
904 | ||
905 | spin_lock_irqsave(&pgd_lock, flags); | |
906 | list_for_each(pos, &pgd_list) | |
907 | if (!vmalloc_sync_one(list_to_pgd(pos), | |
908 | address)) { | |
909 | /* Must be at first entry in list. */ | |
910 | BUG_ON(pos != pgd_list.next); | |
911 | break; | |
912 | } | |
913 | spin_unlock_irqrestore(&pgd_lock, flags); | |
914 | if (pos != pgd_list.next) | |
915 | set_bit(pgd_index(address), insync); | |
916 | } | |
917 | if (address == start && test_bit(pgd_index(address), insync)) | |
918 | start = address + PGDIR_SIZE; | |
919 | } | |
920 | #endif | |
921 | } |