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/sched/signal.h>
  17 #include <linux/hardirq.h>
  18 #include <linux/kprobes.h>
  19 #include <linux/perf_event.h>
  20 #include <linux/kdebug.h>
  21 #include <linux/uaccess.h>
  22 #include <asm/io_trapped.h>
  23 #include <asm/mmu_context.h>
  24 #include <asm/tlbflush.h>
  25 #include <asm/traps.h>
  26
  27 static void
  28 force_sig_info_fault(int si_signo, int si_code, unsigned long address)
  29 {
  30         force_sig_fault(si_signo, si_code, (void __user *)address);
  31 }
  32
  33 /*
  34  * This is useful to dump out the page tables associated with
  35  * 'addr' in mm 'mm'.
  36  */
  37 static void show_pte(struct mm_struct *mm, unsigned long addr)
  38 {
  39         pgd_t *pgd;
  40
  41         if (mm) {
  42                 pgd = mm->pgd;
  43         } else {
  44                 pgd = get_TTB();
  45
  46                 if (unlikely(!pgd))
  47                         pgd = swapper_pg_dir;
  48         }
  49
  50         pr_alert("pgd = %p\n", pgd);
  51         pgd += pgd_index(addr);
  52         pr_alert("[%08lx] *pgd=%0*llx", addr, (u32)(sizeof(*pgd) * 2),
  53                  (u64)pgd_val(*pgd));
  54
  55         do {
  56                 p4d_t *p4d;
  57                 pud_t *pud;
  58                 pmd_t *pmd;
  59                 pte_t *pte;
  60
  61                 if (pgd_none(*pgd))
  62                         break;
  63
  64                 if (pgd_bad(*pgd)) {
  65                         pr_cont("(bad)");
  66                         break;
  67                 }
  68
  69                 p4d = p4d_offset(pgd, addr);
  70                 if (PTRS_PER_P4D != 1)
  71                         pr_cont(", *p4d=%0*Lx", (u32)(sizeof(*p4d) * 2),
  72                                 (u64)p4d_val(*p4d));
  73
  74                 if (p4d_none(*p4d))
  75                         break;
  76
  77                 if (p4d_bad(*p4d)) {
  78                         pr_cont("(bad)");
  79                         break;
  80                 }
  81
  82                 pud = pud_offset(p4d, addr);
  83                 if (PTRS_PER_PUD != 1)
  84                         pr_cont(", *pud=%0*llx", (u32)(sizeof(*pud) * 2),
  85                                 (u64)pud_val(*pud));
  86
  87                 if (pud_none(*pud))
  88                         break;
  89
  90                 if (pud_bad(*pud)) {
  91                         pr_cont("(bad)");
  92                         break;
  93                 }
  94
  95                 pmd = pmd_offset(pud, addr);
  96                 if (PTRS_PER_PMD != 1)
  97                         pr_cont(", *pmd=%0*llx", (u32)(sizeof(*pmd) * 2),
  98                                 (u64)pmd_val(*pmd));
  99
 100                 if (pmd_none(*pmd))
 101                         break;
 102
 103                 if (pmd_bad(*pmd)) {
 104                         pr_cont("(bad)");
 105                         break;
 106                 }
 107
 108                 /* We must not map this if we have highmem enabled */
 109                 if (PageHighMem(pfn_to_page(pmd_val(*pmd) >> PAGE_SHIFT)))
 110                         break;
 111
 112                 pte = pte_offset_kernel(pmd, addr);
 113                 pr_cont(", *pte=%0*llx", (u32)(sizeof(*pte) * 2),
 114                         (u64)pte_val(*pte));
 115         } while (0);
 116
 117         pr_cont("\n");
 118 }
 119
 120 static inline pmd_t *vmalloc_sync_one(pgd_t *pgd, unsigned long address)
 121 {
 122         unsigned index = pgd_index(address);
 123         pgd_t *pgd_k;
 124         p4d_t *p4d, *p4d_k;
 125         pud_t *pud, *pud_k;
 126         pmd_t *pmd, *pmd_k;
 127
 128         pgd += index;
 129         pgd_k = init_mm.pgd + index;
 130
 131         if (!pgd_present(*pgd_k))
 132                 return NULL;
 133
 134         p4d = p4d_offset(pgd, address);
 135         p4d_k = p4d_offset(pgd_k, address);
 136         if (!p4d_present(*p4d_k))
 137                 return NULL;
 138
 139         pud = pud_offset(p4d, address);
 140         pud_k = pud_offset(p4d_k, address);
 141         if (!pud_present(*pud_k))
 142                 return NULL;
 143
 144         if (!pud_present(*pud))
 145             set_pud(pud, *pud_k);
 146
 147         pmd = pmd_offset(pud, address);
 148         pmd_k = pmd_offset(pud_k, address);
 149         if (!pmd_present(*pmd_k))
 150                 return NULL;
 151
 152         if (!pmd_present(*pmd))
 153                 set_pmd(pmd, *pmd_k);
 154         else {
 155                 /*
 156                  * The page tables are fully synchronised so there must
 157                  * be another reason for the fault. Return NULL here to
 158                  * signal that we have not taken care of the fault.
 159                  */
 160                 BUG_ON(pmd_page(*pmd) != pmd_page(*pmd_k));
 161                 return NULL;
 162         }
 163
 164         return pmd_k;
 165 }
 166
 167 #ifdef CONFIG_SH_STORE_QUEUES
 168 #define __FAULT_ADDR_LIMIT      P3_ADDR_MAX
 169 #else
 170 #define __FAULT_ADDR_LIMIT      VMALLOC_END
 171 #endif
 172
 173 /*
 174  * Handle a fault on the vmalloc or module mapping area
 175  */
 176 static noinline int vmalloc_fault(unsigned long address)
 177 {
 178         pgd_t *pgd_k;
 179         pmd_t *pmd_k;
 180         pte_t *pte_k;
 181
 182         /* Make sure we are in vmalloc/module/P3 area: */
 183         if (!(address >= VMALLOC_START && address < __FAULT_ADDR_LIMIT))
 184                 return -1;
 185
 186         /*
 187          * Synchronize this task's top level page-table
 188          * with the 'reference' page table.
 189          *
 190          * Do _not_ use "current" here. We might be inside
 191          * an interrupt in the middle of a task switch..
 192          */
 193         pgd_k = get_TTB();
 194         pmd_k = vmalloc_sync_one(pgd_k, address);
 195         if (!pmd_k)
 196                 return -1;
 197
 198         pte_k = pte_offset_kernel(pmd_k, address);
 199         if (!pte_present(*pte_k))
 200                 return -1;
 201
 202         return 0;
 203 }
 204
 205 static void
 206 show_fault_oops(struct pt_regs *regs, unsigned long address)
 207 {
 208         if (!oops_may_print())
 209                 return;
 210
 211         printk(KERN_ALERT "PC:");
 212         pr_alert("BUG: unable to handle kernel %s at %08lx\n",
 213                  address < PAGE_SIZE ? "NULL pointer dereference"
 214                                      : "paging request",
 215                  address);
 216         pr_alert("PC:");
 217         printk_address(regs->pc, 1, KERN_ALERT);
 218
 219         show_pte(NULL, address);
 220 }
 221
 222 static noinline void
 223 no_context(struct pt_regs *regs, unsigned long error_code,
 224            unsigned long address)
 225 {
 226         /* Are we prepared to handle this kernel fault?  */
 227         if (fixup_exception(regs))
 228                 return;
 229
 230         if (handle_trapped_io(regs, address))
 231                 return;
 232
 233         /*
 234          * Oops. The kernel tried to access some bad page. We'll have to
 235          * terminate things with extreme prejudice.
 236          */
 237         bust_spinlocks(1);
 238
 239         show_fault_oops(regs, address);
 240
 241         die("Oops", regs, error_code);
 242         bust_spinlocks(0);
 243         do_exit(SIGKILL);
 244 }
 245
 246 static void
 247 __bad_area_nosemaphore(struct pt_regs *regs, unsigned long error_code,
 248                        unsigned long address, int si_code)
 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);
 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         mmap_read_unlock(mm);
 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
 301 do_sigbus(struct pt_regs *regs, unsigned long error_code, unsigned long address)
 302 {
 303         struct task_struct *tsk = current;
 304         struct mm_struct *mm = tsk->mm;
 305
 306         mmap_read_unlock(mm);
 307
 308         /* Kernel mode? Handle exceptions or die: */
 309         if (!user_mode(regs))
 310                 no_context(regs, error_code, address);
 311
 312         force_sig_info_fault(SIGBUS, BUS_ADRERR, address);
 313 }
 314
 315 static noinline int
 316 mm_fault_error(struct pt_regs *regs, unsigned long error_code,
 317                unsigned long address, vm_fault_t fault)
 318 {
 319         /*
 320          * Pagefault was interrupted by SIGKILL. We have no reason to
 321          * continue pagefault.
 322          */
 323         if (fault_signal_pending(fault, regs)) {
 324                 if (!user_mode(regs))
 325                         no_context(regs, error_code, address);
 326                 return 1;
 327         }
 328
 329         /* Release mmap_lock first if necessary */
 330         if (!(fault & VM_FAULT_RETRY))
 331                 mmap_read_unlock(current->mm);
 332
 333         if (!(fault & VM_FAULT_ERROR))
 334                 return 0;
 335
 336         if (fault & VM_FAULT_OOM) {
 337                 /* Kernel mode? Handle exceptions or die: */
 338                 if (!user_mode(regs)) {
 339                         no_context(regs, error_code, address);
 340                         return 1;
 341                 }
 342
 343                 /*
 344                  * We ran out of memory, call the OOM killer, and return the
 345                  * userspace (which will retry the fault, or kill us if we got
 346                  * oom-killed):
 347                  */
 348                 pagefault_out_of_memory();
 349         } else {
 350                 if (fault & VM_FAULT_SIGBUS)
 351                         do_sigbus(regs, error_code, address);
 352                 else if (fault & VM_FAULT_SIGSEGV)
 353                         bad_area(regs, error_code, address);
 354                 else
 355                         BUG();
 356         }
 357
 358         return 1;
 359 }
 360
 361 static inline int access_error(int error_code, struct vm_area_struct *vma)
 362 {
 363         if (error_code & FAULT_CODE_WRITE) {
 364                 /* write, present and write, not present: */
 365                 if (unlikely(!(vma->vm_flags & VM_WRITE)))
 366                         return 1;
 367                 return 0;
 368         }
 369
 370         /* ITLB miss on NX page */
 371         if (unlikely((error_code & FAULT_CODE_ITLB) &&
 372                      !(vma->vm_flags & VM_EXEC)))
 373                 return 1;
 374
 375         /* read, not present: */
 376         if (unlikely(!vma_is_accessible(vma)))
 377                 return 1;
 378
 379         return 0;
 380 }
 381
 382 static int fault_in_kernel_space(unsigned long address)
 383 {
 384         return address >= TASK_SIZE;
 385 }
 386
 387 /*
 388  * This routine handles page faults.  It determines the address,
 389  * and the problem, and then passes it off to one of the appropriate
 390  * routines.
 391  */
 392 asmlinkage void __kprobes do_page_fault(struct pt_regs *regs,
 393                                         unsigned long error_code,
 394                                         unsigned long address)
 395 {
 396         unsigned long vec;
 397         struct task_struct *tsk;
 398         struct mm_struct *mm;
 399         struct vm_area_struct * vma;
 400         vm_fault_t fault;
 401         unsigned int flags = FAULT_FLAG_DEFAULT;
 402
 403         tsk = current;
 404         mm = tsk->mm;
 405         vec = lookup_exception_vector();
 406
 407         /*
 408          * We fault-in kernel-space virtual memory on-demand. The
 409          * 'reference' page table is init_mm.pgd.
 410          *
 411          * NOTE! We MUST NOT take any locks for this case. We may
 412          * be in an interrupt or a critical region, and should
 413          * only copy the information from the master page table,
 414          * nothing more.
 415          */
 416         if (unlikely(fault_in_kernel_space(address))) {
 417                 if (vmalloc_fault(address) >= 0)
 418                         return;
 419                 if (kprobe_page_fault(regs, vec))
 420                         return;
 421
 422                 bad_area_nosemaphore(regs, error_code, address);
 423                 return;
 424         }
 425
 426         if (unlikely(kprobe_page_fault(regs, vec)))
 427                 return;
 428
 429         /* Only enable interrupts if they were on before the fault */
 430         if ((regs->sr & SR_IMASK) != SR_IMASK)
 431                 local_irq_enable();
 432
 433         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
 434
 435         /*
 436          * If we're in an interrupt, have no user context or are running
 437          * with pagefaults disabled then we must not take the fault:
 438          */
 439         if (unlikely(faulthandler_disabled() || !mm)) {
 440                 bad_area_nosemaphore(regs, error_code, address);
 441                 return;
 442         }
 443
 444 retry:
 445         mmap_read_lock(mm);
 446
 447         vma = find_vma(mm, address);
 448         if (unlikely(!vma)) {
 449                 bad_area(regs, error_code, address);
 450                 return;
 451         }
 452         if (likely(vma->vm_start <= address))
 453                 goto good_area;
 454         if (unlikely(!(vma->vm_flags & VM_GROWSDOWN))) {
 455                 bad_area(regs, error_code, address);
 456                 return;
 457         }
 458         if (unlikely(expand_stack(vma, address))) {
 459                 bad_area(regs, error_code, address);
 460                 return;
 461         }
 462
 463         /*
 464          * Ok, we have a good vm_area for this memory access, so
 465          * we can handle it..
 466          */
 467 good_area:
 468         if (unlikely(access_error(error_code, vma))) {
 469                 bad_area_access_error(regs, error_code, address);
 470                 return;
 471         }
 472
 473         set_thread_fault_code(error_code);
 474
 475         if (user_mode(regs))
 476                 flags |= FAULT_FLAG_USER;
 477         if (error_code & FAULT_CODE_WRITE)
 478                 flags |= FAULT_FLAG_WRITE;
 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(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_TRIED;
 503
 504                         /*
 505                          * No need to mmap_read_unlock(mm) 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         mmap_read_unlock(mm);
 514 }