arch/sh/mm/fault.c

   1 /*
   2  * Page fault handler for SH with an MMU.
   3  *
   4  *  Copyright (C) 1999  Niibe Yutaka
   5  *  Copyright (C) 2003 - 2012  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 <linux/perf_event.h>
  19 #include <linux/kdebug.h>
  20 #include <asm/io_trapped.h>
  21 #include <asm/mmu_context.h>
  22 #include <asm/tlbflush.h>
  23 #include <asm/traps.h>
  24
  25 static inline int notify_page_fault(struct pt_regs *regs, int trap)
  26 {
  27         int ret = 0;
  28
  29         if (kprobes_built_in() && !user_mode(regs)) {
  30                 preempt_disable();
  31                 if (kprobe_running() && kprobe_fault_handler(regs, trap))
  32                         ret = 1;
  33                 preempt_enable();
  34         }
  35
  36         return ret;
  37 }
  38
  39 static void
  40 force_sig_info_fault(int si_signo, int si_code, unsigned long address,
  41                      struct task_struct *tsk)
  42 {
  43         siginfo_t info;
  44
  45         info.si_signo   = si_signo;
  46         info.si_errno   = 0;
  47         info.si_code    = si_code;
  48         info.si_addr    = (void __user *)address;
  49
  50         force_sig_info(si_signo, &info, tsk);
  51 }
  52
  53 /*
  54  * This is useful to dump out the page tables associated with
  55  * 'addr' in mm 'mm'.
  56  */
  57 static void show_pte(struct mm_struct *mm, unsigned long addr)
  58 {
  59         pgd_t *pgd;
  60
  61         if (mm)
  62                 pgd = mm->pgd;
  63         else
  64                 pgd = get_TTB();
  65
  66         printk(KERN_ALERT "pgd = %p\n", pgd);
  67         pgd += pgd_index(addr);
  68         printk(KERN_ALERT "[%08lx] *pgd=%0*Lx", addr,
  69                (u32)(sizeof(*pgd) * 2), (u64)pgd_val(*pgd));
  70
  71         do {
  72                 pud_t *pud;
  73                 pmd_t *pmd;
  74                 pte_t *pte;
  75
  76                 if (pgd_none(*pgd))
  77                         break;
  78
  79                 if (pgd_bad(*pgd)) {
  80                         printk("(bad)");
  81                         break;
  82                 }
  83
  84                 pud = pud_offset(pgd, addr);
  85                 if (PTRS_PER_PUD != 1)
  86                         printk(", *pud=%0*Lx", (u32)(sizeof(*pud) * 2),
  87                                (u64)pud_val(*pud));
  88
  89                 if (pud_none(*pud))
  90                         break;
  91
  92                 if (pud_bad(*pud)) {
  93                         printk("(bad)");
  94                         break;
  95                 }
  96
  97                 pmd = pmd_offset(pud, addr);
  98                 if (PTRS_PER_PMD != 1)
  99                         printk(", *pmd=%0*Lx", (u32)(sizeof(*pmd) * 2),
 100                                (u64)pmd_val(*pmd));
 101
 102                 if (pmd_none(*pmd))
 103                         break;
 104
 105                 if (pmd_bad(*pmd)) {
 106                         printk("(bad)");
 107                         break;
 108                 }
 109
 110                 /* We must not map this if we have highmem enabled */
 111                 if (PageHighMem(pfn_to_page(pmd_val(*pmd) >> PAGE_SHIFT)))
 112                         break;
 113
 114                 pte = pte_offset_kernel(pmd, addr);
 115                 printk(", *pte=%0*Lx", (u32)(sizeof(*pte) * 2),
 116                        (u64)pte_val(*pte));
 117         } while (0);
 118
 119         printk("\n");
 120 }
 121
 122 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 123 {
 124         unsigned index = pgd_index(address);
 125         pgd_t *pgd_k;
 126         pud_t *pud, *pud_k;
 127         pmd_t *pmd, *pmd_k;
 128
 129         pgd += index;
 130         pgd_k = init_mm.pgd + index;
 131
 132         if (!pgd_present(*pgd_k))
 133                 return NULL;
 134
 135         pud = pud_offset(pgd, address);
 136         pud_k = pud_offset(pgd_k, address);
 137         if (!pud_present(*pud_k))
 138                 return NULL;
 139
 140         if (!pud_present(*pud))
 141             set_pud(pud, *pud_k);
 142
 143         pmd = pmd_offset(pud, address);
 144         pmd_k = pmd_offset(pud_k, address);
 145         if (!pmd_present(*pmd_k))
 146                 return NULL;
 147
 148         if (!pmd_present(*pmd))
 149                 set_pmd(pmd, *pmd_k);
 150         else {
 151                 /*
 152                  * The page tables are fully synchronised so there must
 153                  * be another reason for the fault. Return NULL here to
 154                  * signal that we have not taken care of the fault.
 155                  */
 156                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 157                 return NULL;
 158         }
 159
 160         return pmd_k;
 161 }
 162
 163 #ifdef CONFIG_SH_STORE_QUEUES
 164 #define __FAULT_ADDR_LIMIT      P3_ADDR_MAX
 165 #else
 166 #define __FAULT_ADDR_LIMIT      VMALLOC_END
 167 #endif
 168
 169 /*
 170  * Handle a fault on the vmalloc or module mapping area
 171  */
 172 static noinline int vmalloc_fault(unsigned long address)
 173 {
 174         pgd_t *pgd_k;
 175         pmd_t *pmd_k;
 176         pte_t *pte_k;
 177
 178         /* Make sure we are in vmalloc/module/P3 area: */
 179         if (!(address >= VMALLOC_START && address < __FAULT_ADDR_LIMIT))
 180                 return -1;
 181
 182         /*
 183          * Synchronize this task's top level page-table
 184          * with the 'reference' page table.
 185          *
 186          * Do _not_ use "current" here. We might be inside
 187          * an interrupt in the middle of a task switch..
 188          */
 189         pgd_k = get_TTB();
 190         pmd_k = vmalloc_sync_one(pgd_k, address);
 191         if (!pmd_k)
 192                 return -1;
 193
 194         pte_k = pte_offset_kernel(pmd_k, address);
 195         if (!pte_present(*pte_k))
 196                 return -1;
 197
 198         return 0;
 199 }
 200
 201 static void
 202 show_fault_oops(struct pt_regs *regs, unsigned long address)
 203 {
 204         if (!oops_may_print())
 205                 return;
 206
 207         printk(KERN_ALERT "BUG: unable to handle kernel ");
 208         if (address < PAGE_SIZE)
 209                 printk(KERN_CONT "NULL pointer dereference");
 210         else
 211                 printk(KERN_CONT "paging request");
 212
 213         printk(KERN_CONT " at %08lx\n", address);
 214         printk(KERN_ALERT "PC:");
 215         printk_address(regs->pc, 1);
 216
 217         show_pte(NULL, address);
 218 }
 219
 220 static noinline void
 221 no_context(struct pt_regs *regs, unsigned long error_code,
 222            unsigned long address)
 223 {
 224         /* Are we prepared to handle this kernel fault?  */
 225         if (fixup_exception(regs))
 226                 return;
 227
 228         if (handle_trapped_io(regs, address))
 229                 return;
 230
 231         /*
 232          * Oops. The kernel tried to access some bad page. We'll have to
 233          * terminate things with extreme prejudice.
 234          */
 235         bust_spinlocks(1);
 236
 237         show_fault_oops(regs, address);
 238
 239         die("Oops", regs, error_code);
 240         bust_spinlocks(0);
 241         do_exit(SIGKILL);
 242 }
 243
 244 static void
 245 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 246                        unsigned long address, int si_code)
 247 {
 248         struct task_struct *tsk = current;
 249
 250         /* User mode accesses just cause a SIGSEGV */
 251         if (user_mode(regs)) {
 252                 /*
 253                  * It's possible to have interrupts off here:
 254                  */
 255                 local_irq_enable();
 256
 257                 force_sig_info_fault(SIGSEGV, si_code, address, tsk);
 258
 259                 return;
 260         }
 261
 262         no_context(regs, error_code, address);
 263 }
 264
 265 static noinline void
 266 bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 267                      unsigned long address)
 268 {
 269         __bad_area_nosemaphore(regs, error_code, address, SEGV_MAPERR);
 270 }
 271
 272 static void
 273 __bad_area(struct pt_regs *regs, unsigned long error_code,
 274            unsigned long address, int si_code)
 275 {
 276         struct mm_struct *mm = current->mm;
 277
 278         /*
 279          * Something tried to access memory that isn't in our memory map..
 280          * Fix it, but check if it's kernel or user first..
 281          */
 282         up_read(&mm->mmap_sem);
 283
 284         __bad_area_nosemaphore(regs, error_code, address, si_code);
 285 }
 286
 287 static noinline void
 288 bad_area(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 289 {
 290         __bad_area(regs, error_code, address, SEGV_MAPERR);
 291 }
 292
 293 static noinline void
 294 bad_area_access_error(struct pt_regs *regs, unsigned long error_code,
 295                       unsigned long address)
 296 {
 297         __bad_area(regs, error_code, address, SEGV_ACCERR);
 298 }
 299
 300 static void out_of_memory(void)
 301 {
 302         /*
 303          * We ran out of memory, call the OOM killer, and return the userspace
 304          * (which will retry the fault, or kill us if we got oom-killed):
 305          */
 306         up_read(&current->mm->mmap_sem);
 307
 308         pagefault_out_of_memory();
 309 }
 310
 311 static void
 312 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 313 {
 314         struct task_struct *tsk = current;
 315         struct mm_struct *mm = tsk->mm;
 316
 317         up_read(&mm->mmap_sem);
 318
 319         /* Kernel mode? Handle exceptions or die: */
 320         if (!user_mode(regs))
 321                 no_context(regs, error_code, address);
 322
 323         force_sig_info_fault(SIGBUS, BUS_ADRERR, address, tsk);
 324 }
 325
 326 static noinline int
 327 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
 328                unsigned long address, unsigned int fault)
 329 {
 330         /*
 331          * Pagefault was interrupted by SIGKILL. We have no reason to
 332          * continue pagefault.
 333          */
 334         if (fatal_signal_pending(current)) {
 335                 if (!(fault & VM_FAULT_RETRY))
 336                         up_read(&current->mm->mmap_sem);
 337                 if (!user_mode(regs))
 338                         no_context(regs, error_code, address);
 339                 return 1;
 340         }
 341
 342         if (!(fault & VM_FAULT_ERROR))
 343                 return 0;
 344
 345         if (fault & VM_FAULT_OOM) {
 346                 /* Kernel mode? Handle exceptions or die: */
 347                 if (!user_mode(regs)) {
 348                         up_read(&current->mm->mmap_sem);
 349                         no_context(regs, error_code, address);
 350                         return 1;
 351                 }
 352
 353                 out_of_memory();
 354         } else {
 355                 if (fault & VM_FAULT_SIGBUS)
 356                         do_sigbus(regs, error_code, address);
 357                 else
 358                         BUG();
 359         }
 360
 361         return 1;
 362 }
 363
 364 static inline int access_error(int error_code, struct vm_area_struct *vma)
 365 {
 366         if (error_code & FAULT_CODE_WRITE) {
 367                 /* write, present and write, not present: */
 368                 if (unlikely(!(vma->vm_flags & VM_WRITE)))
 369                         return 1;
 370                 return 0;
 371         }
 372
 373         /* ITLB miss on NX page */
 374         if (unlikely((error_code & FAULT_CODE_ITLB) &&
 375                      !(vma->vm_flags & VM_EXEC)))
 376                 return 1;
 377
 378         /* read, not present: */
 379         if (unlikely(!(vma->vm_flags & (VM_READ | VM_EXEC | VM_WRITE))))
 380                 return 1;
 381
 382         return 0;
 383 }
 384
 385 static int fault_in_kernel_space(unsigned long address)
 386 {
 387         return address >= TASK_SIZE;
 388 }
 389
 390 /*
 391  * This routine handles page faults.  It determines the address,
 392  * and the problem, and then passes it off to one of the appropriate
 393  * routines.
 394  */
 395 asmlinkage void __kprobes do_page_fault(struct pt_regs *regs,
 396                                         unsigned long error_code,
 397                                         unsigned long address)
 398 {
 399         unsigned long vec;
 400         struct task_struct *tsk;
 401         struct mm_struct *mm;
 402         struct vm_area_struct * vma;
 403         int fault;
 404         int write = error_code & FAULT_CODE_WRITE;
 405         unsigned int flags = (FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE |
 406                               (write ? FAULT_FLAG_WRITE : 0));
 407
 408         tsk = current;
 409         mm = tsk->mm;
 410         vec = lookup_exception_vector();
 411
 412         /*
 413          * We fault-in kernel-space virtual memory on-demand. The
 414          * 'reference' page table is init_mm.pgd.
 415          *
 416          * NOTE! We MUST NOT take any locks for this case. We may
 417          * be in an interrupt or a critical region, and should
 418          * only copy the information from the master page table,
 419          * nothing more.
 420          */
 421         if (unlikely(fault_in_kernel_space(address))) {
 422                 if (vmalloc_fault(address) >= 0)
 423                         return;
 424                 if (notify_page_fault(regs, vec))
 425                         return;
 426
 427                 bad_area_nosemaphore(regs, error_code, address);
 428                 return;
 429         }
 430
 431         if (unlikely(notify_page_fault(regs, vec)))
 432                 return;
 433
 434         /* Only enable interrupts if they were on before the fault */
 435         if ((regs->sr & SR_IMASK) != SR_IMASK)
 436                 local_irq_enable();
 437
 438         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
 439
 440         /*
 441          * If we're in an interrupt, have no user context or are running
 442          * in an atomic region then we must not take the fault:
 443          */
 444         if (unlikely(in_atomic() || !mm)) {
 445                 bad_area_nosemaphore(regs, error_code, address);
 446                 return;
 447         }
 448
 449 retry:
 450         down_read(&mm->mmap_sem);
 451
 452         vma = find_vma(mm, address);
 453         if (unlikely(!vma)) {
 454                 bad_area(regs, error_code, address);
 455                 return;
 456         }
 457         if (likely(vma->vm_start <= address))
 458                 goto good_area;
 459         if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
 460                 bad_area(regs, error_code, address);
 461                 return;
 462         }
 463         if (unlikely(expand_stack(vma, address))) {
 464                 bad_area(regs, error_code, address);
 465                 return;
 466         }
 467
 468         /*
 469          * Ok, we have a good vm_area for this memory access, so
 470          * we can handle it..
 471          */
 472 good_area:
 473         if (unlikely(access_error(error_code, vma))) {
 474                 bad_area_access_error(regs, error_code, address);
 475                 return;
 476         }
 477
 478         set_thread_fault_code(error_code);
 479
 480         /*
 481          * If for any reason at all we couldn't handle the fault,
 482          * make sure we exit gracefully rather than endlessly redo
 483          * the fault.
 484          */
 485         fault = handle_mm_fault(mm, vma, address, flags);
 486
 487         if (unlikely(fault & (VM_FAULT_RETRY | VM_FAULT_ERROR)))
 488                 if (mm_fault_error(regs, error_code, address, fault))
 489                         return;
 490
 491         if (flags & FAULT_FLAG_ALLOW_RETRY) {
 492                 if (fault & VM_FAULT_MAJOR) {
 493                         tsk->maj_flt++;
 494                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
 495                                       regs, address);
 496                 } else {
 497                         tsk->min_flt++;
 498                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
 499                                       regs, address);
 500                 }
 501                 if (fault & VM_FAULT_RETRY) {
 502                         flags &= ~FAULT_FLAG_ALLOW_RETRY;
 503
 504                         /*
 505                          * No need to up_read(&mm->mmap_sem) as we would
 506                          * have already released it in __lock_page_or_retry
 507                          * in mm/filemap.c.
 508                          */
 509                         goto retry;
 510                 }
 511         }
 512
 513         up_read(&mm->mmap_sem);
 514 }