arch/sparc/mm/fault.c

   1 /* $Id: fault.c,v 1.122 2001/11/17 07:19:26 davem Exp $
   2  * fault.c:  Page fault handlers for the Sparc.
   3  *
   4  * Copyright (C) 1995 David S. Miller (davem@caip.rutgers.edu)
   5  * Copyright (C) 1996 Eddie C. Dost (ecd@skynet.be)
   6  * Copyright (C) 1997 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
   7  */
   8
   9 #include <asm/head.h>
  10
  11 #include <linux/string.h>
  12 #include <linux/types.h>
  13 #include <linux/sched.h>
  14 #include <linux/ptrace.h>
  15 #include <linux/mman.h>
  16 #include <linux/threads.h>
  17 #include <linux/kernel.h>
  18 #include <linux/signal.h>
  19 #include <linux/mm.h>
  20 #include <linux/smp.h>
  21 #include <linux/smp_lock.h>
  22 #include <linux/interrupt.h>
  23 #include <linux/module.h>
  24
  25 #include <asm/system.h>
  26 #include <asm/segment.h>
  27 #include <asm/page.h>
  28 #include <asm/pgtable.h>
  29 #include <asm/memreg.h>
  30 #include <asm/openprom.h>
  31 #include <asm/oplib.h>
  32 #include <asm/smp.h>
  33 #include <asm/traps.h>
  34 #include <asm/kdebug.h>
  35 #include <asm/uaccess.h>
  36
  37 #define ELEMENTS(arr) (sizeof (arr)/sizeof (arr[0]))
  38
  39 extern int prom_node_root;
  40
  41 /* At boot time we determine these two values necessary for setting
  42  * up the segment maps and page table entries (pte's).
  43  */
  44
  45 int num_segmaps, num_contexts;
  46 int invalid_segment;
  47
  48 /* various Virtual Address Cache parameters we find at boot time... */
  49
  50 int vac_size, vac_linesize, vac_do_hw_vac_flushes;
  51 int vac_entries_per_context, vac_entries_per_segment;
  52 int vac_entries_per_page;
  53
  54 /* Nice, simple, prom library does all the sweating for us. ;) */
  55 int prom_probe_memory (void)
  56 {
  57         register struct linux_mlist_v0 *mlist;
  58         register unsigned long bytes, base_paddr, tally;
  59         register int i;
  60
  61         i = 0;
  62         mlist= *prom_meminfo()->v0_available;
  63         bytes = tally = mlist->num_bytes;
  64         base_paddr = (unsigned long) mlist->start_adr;
  65
  66         sp_banks[0].base_addr = base_paddr;
  67         sp_banks[0].num_bytes = bytes;
  68
  69         while (mlist->theres_more != (void *) 0){
  70                 i++;
  71                 mlist = mlist->theres_more;
  72                 bytes = mlist->num_bytes;
  73                 tally += bytes;
  74                 if (i > SPARC_PHYS_BANKS-1) {
  75                         printk ("The machine has more banks than "
  76                                 "this kernel can support\n"
  77                                 "Increase the SPARC_PHYS_BANKS "
  78                                 "setting (currently %d)\n",
  79                                 SPARC_PHYS_BANKS);
  80                         i = SPARC_PHYS_BANKS-1;
  81                         break;
  82                 }
  83
  84                 sp_banks[i].base_addr = (unsigned long) mlist->start_adr;
  85                 sp_banks[i].num_bytes = mlist->num_bytes;
  86         }
  87
  88         i++;
  89         sp_banks[i].base_addr = 0xdeadbeef;
  90         sp_banks[i].num_bytes = 0;
  91
  92         /* Now mask all bank sizes on a page boundary, it is all we can
  93          * use anyways.
  94          */
  95         for(i=0; sp_banks[i].num_bytes != 0; i++)
  96                 sp_banks[i].num_bytes &= PAGE_MASK;
  97
  98         return tally;
  99 }
 100
 101 /* Traverse the memory lists in the prom to see how much physical we
 102  * have.
 103  */
 104 unsigned long
 105 probe_memory(void)
 106 {
 107         int total;
 108
 109         total = prom_probe_memory();
 110
 111         /* Oh man, much nicer, keep the dirt in promlib. */
 112         return total;
 113 }
 114
 115 extern void sun4c_complete_all_stores(void);
 116
 117 /* Whee, a level 15 NMI interrupt memory error.  Let's have fun... */
 118 asmlinkage void sparc_lvl15_nmi(struct pt_regs *regs, unsigned long serr,
 119                                 unsigned long svaddr, unsigned long aerr,
 120                                 unsigned long avaddr)
 121 {
 122         sun4c_complete_all_stores();
 123         printk("FAULT: NMI received\n");
 124         printk("SREGS: Synchronous Error %08lx\n", serr);
 125         printk("       Synchronous Vaddr %08lx\n", svaddr);
 126         printk("      Asynchronous Error %08lx\n", aerr);
 127         printk("      Asynchronous Vaddr %08lx\n", avaddr);
 128         if (sun4c_memerr_reg)
 129                 printk("     Memory Parity Error %08lx\n", *sun4c_memerr_reg);
 130         printk("REGISTER DUMP:\n");
 131         show_regs(regs);
 132         prom_halt();
 133 }
 134
 135 static void unhandled_fault(unsigned long, struct task_struct *,
 136                 struct pt_regs *) __attribute__ ((noreturn));
 137
 138 static void unhandled_fault(unsigned long address, struct task_struct *tsk,
 139                      struct pt_regs *regs)
 140 {
 141         if((unsigned long) address < PAGE_SIZE) {
 142                 printk(KERN_ALERT
 143                     "Unable to handle kernel NULL pointer dereference\n");
 144         } else {
 145                 printk(KERN_ALERT "Unable to handle kernel paging request "
 146                        "at virtual address %08lx\n", address);
 147         }
 148         printk(KERN_ALERT "tsk->{mm,active_mm}->context = %08lx\n",
 149                 (tsk->mm ? tsk->mm->context : tsk->active_mm->context));
 150         printk(KERN_ALERT "tsk->{mm,active_mm}->pgd = %08lx\n",
 151                 (tsk->mm ? (unsigned long) tsk->mm->pgd :
 152                         (unsigned long) tsk->active_mm->pgd));
 153         die_if_kernel("Oops", regs);
 154 }
 155
 156 asmlinkage int lookup_fault(unsigned long pc, unsigned long ret_pc,
 157                             unsigned long address)
 158 {
 159         struct pt_regs regs;
 160         unsigned long g2;
 161         unsigned int insn;
 162         int i;
 163
 164         i = search_extables_range(ret_pc, &g2);
 165         switch (i) {
 166         case 3:
 167                 /* load & store will be handled by fixup */
 168                 return 3;
 169
 170         case 1:
 171                 /* store will be handled by fixup, load will bump out */
 172                 /* for _to_ macros */
 173                 insn = *((unsigned int *) pc);
 174                 if ((insn >> 21) & 1)
 175                         return 1;
 176                 break;
 177
 178         case 2:
 179                 /* load will be handled by fixup, store will bump out */
 180                 /* for _from_ macros */
 181                 insn = *((unsigned int *) pc);
 182                 if (!((insn >> 21) & 1) || ((insn>>19)&0x3f) == 15)
 183                         return 2;
 184                 break;
 185
 186         default:
 187                 break;
 188         };
 189
 190         memset(&regs, 0, sizeof (regs));
 191         regs.pc = pc;
 192         regs.npc = pc + 4;
 193         __asm__ __volatile__(
 194                 "rd %%psr, %0\n\t"
 195                 "nop\n\t"
 196                 "nop\n\t"
 197                 "nop\n" : "=r" (regs.psr));
 198         unhandled_fault(address, current, &regs);
 199
 200         /* Not reached */
 201         return 0;
 202 }
 203
 204 asmlinkage void do_sparc_fault(struct pt_regs *regs, int text_fault, int write,
 205                                unsigned long address)
 206 {
 207         struct vm_area_struct *vma;
 208         struct task_struct *tsk = current;
 209         struct mm_struct *mm = tsk->mm;
 210         unsigned int fixup;
 211         unsigned long g2;
 212         siginfo_t info;
 213         int from_user = !(regs->psr & PSR_PS);
 214
 215         if(text_fault)
 216                 address = regs->pc;
 217
 218         /*
 219          * We fault-in kernel-space virtual memory on-demand. The
 220          * 'reference' page table is init_mm.pgd.
 221          *
 222          * NOTE! We MUST NOT take any locks for this case. We may
 223          * be in an interrupt or a critical region, and should
 224          * only copy the information from the master page table,
 225          * nothing more.
 226          */
 227         if (!ARCH_SUN4C_SUN4 && address >= TASK_SIZE)
 228                 goto vmalloc_fault;
 229
 230         info.si_code = SEGV_MAPERR;
 231
 232         /*
 233          * If we're in an interrupt or have no user
 234          * context, we must not take the fault..
 235          */
 236         if (in_atomic() || !mm)
 237                 goto no_context;
 238
 239         down_read(&mm->mmap_sem);
 240
 241         /*
 242          * The kernel referencing a bad kernel pointer can lock up
 243          * a sun4c machine completely, so we must attempt recovery.
 244          */
 245         if(!from_user && address >= PAGE_OFFSET)
 246                 goto bad_area;
 247
 248         vma = find_vma(mm, address);
 249         if(!vma)
 250                 goto bad_area;
 251         if(vma->vm_start <= address)
 252                 goto good_area;
 253         if(!(vma->vm_flags & VM_GROWSDOWN))
 254                 goto bad_area;
 255         if(expand_stack(vma, address))
 256                 goto bad_area;
 257         /*
 258          * Ok, we have a good vm_area for this memory access, so
 259          * we can handle it..
 260          */
 261 good_area:
 262         info.si_code = SEGV_ACCERR;
 263         if(write) {
 264                 if(!(vma->vm_flags & VM_WRITE))
 265                         goto bad_area;
 266         } else {
 267                 /* Allow reads even for write-only mappings */
 268                 if(!(vma->vm_flags & (VM_READ | VM_EXEC)))
 269                         goto bad_area;
 270         }
 271
 272         /*
 273          * If for any reason at all we couldn't handle the fault,
 274          * make sure we exit gracefully rather than endlessly redo
 275          * the fault.
 276          */
 277         switch (handle_mm_fault(mm, vma, address, write)) {
 278         case 1:
 279                 current->min_flt++;
 280                 break;
 281         case 2:
 282                 current->maj_flt++;
 283                 break;
 284         case 0:
 285                 goto do_sigbus;
 286         default:
 287                 goto out_of_memory;
 288         }
 289         up_read(&mm->mmap_sem);
 290         return;
 291
 292         /*
 293          * Something tried to access memory that isn't in our memory map..
 294          * Fix it, but check if it's kernel or user first..
 295          */
 296 bad_area:
 297         up_read(&mm->mmap_sem);
 298
 299 bad_area_nosemaphore:
 300         /* User mode accesses just cause a SIGSEGV */
 301         if(from_user) {
 302 #if 0
 303                 printk("Fault whee %s [%d]: segfaults at %08lx pc=%08lx\n",
 304                        tsk->comm, tsk->pid, address, regs->pc);
 305 #endif
 306                 info.si_signo = SIGSEGV;
 307                 info.si_errno = 0;
 308                 /* info.si_code set above to make clear whether
 309                    this was a SEGV_MAPERR or SEGV_ACCERR fault.  */
 310                 info.si_addr = (void *)address;
 311                 info.si_trapno = 0;
 312                 force_sig_info (SIGSEGV, &info, tsk);
 313                 return;
 314         }
 315
 316         /* Is this in ex_table? */
 317 no_context:
 318         g2 = regs->u_regs[UREG_G2];
 319         if (!from_user && (fixup = search_extables_range(regs->pc, &g2))) {
 320                 if (fixup > 10) { /* Values below are reserved for other things */
 321                         extern const unsigned __memset_start[];
 322                         extern const unsigned __memset_end[];
 323                         extern const unsigned __csum_partial_copy_start[];
 324                         extern const unsigned __csum_partial_copy_end[];
 325
 326 #ifdef DEBUG_EXCEPTIONS
 327                         printk("Exception: PC<%08lx> faddr<%08lx>\n", regs->pc, address);
 328                         printk("EX_TABLE: insn<%08lx> fixup<%08x> g2<%08lx>\n",
 329                                 regs->pc, fixup, g2);
 330 #endif
 331                         if ((regs->pc >= (unsigned long)__memset_start &&
 332                              regs->pc < (unsigned long)__memset_end) ||
 333                             (regs->pc >= (unsigned long)__csum_partial_copy_start &&
 334                              regs->pc < (unsigned long)__csum_partial_copy_end)) {
 335                                 regs->u_regs[UREG_I4] = address;
 336                                 regs->u_regs[UREG_I5] = regs->pc;
 337                         }
 338                         regs->u_regs[UREG_G2] = g2;
 339                         regs->pc = fixup;
 340                         regs->npc = regs->pc + 4;
 341                         return;
 342                 }
 343         }
 344
 345         unhandled_fault (address, tsk, regs);
 346         do_exit(SIGKILL);
 347
 348 /*
 349  * We ran out of memory, or some other thing happened to us that made
 350  * us unable to handle the page fault gracefully.
 351  */
 352 out_of_memory:
 353         up_read(&mm->mmap_sem);
 354         printk("VM: killing process %s\n", tsk->comm);
 355         if (from_user)
 356                 do_exit(SIGKILL);
 357         goto no_context;
 358
 359 do_sigbus:
 360         up_read(&mm->mmap_sem);
 361         info.si_signo = SIGBUS;
 362         info.si_errno = 0;
 363         info.si_code = BUS_ADRERR;
 364         info.si_addr = (void *)address;
 365         info.si_trapno = 0;
 366         force_sig_info (SIGBUS, &info, tsk);
 367         if (!from_user)
 368                 goto no_context;
 369
 370 vmalloc_fault:
 371         {
 372                 /*
 373                  * Synchronize this task's top level page-table
 374                  * with the 'reference' page table.
 375                  */
 376                 int offset = pgd_index(address);
 377                 pgd_t *pgd, *pgd_k;
 378                 pmd_t *pmd, *pmd_k;
 379
 380                 pgd = tsk->active_mm->pgd + offset;
 381                 pgd_k = init_mm.pgd + offset;
 382
 383                 if (!pgd_present(*pgd)) {
 384                         if (!pgd_present(*pgd_k))
 385                                 goto bad_area_nosemaphore;
 386                         pgd_val(*pgd) = pgd_val(*pgd_k);
 387                         return;
 388                 }
 389
 390                 pmd = pmd_offset(pgd, address);
 391                 pmd_k = pmd_offset(pgd_k, address);
 392
 393                 if (pmd_present(*pmd) || !pmd_present(*pmd_k))
 394                         goto bad_area_nosemaphore;
 395                 *pmd = *pmd_k;
 396                 return;
 397         }
 398 }
 399
 400 asmlinkage void do_sun4c_fault(struct pt_regs *regs, int text_fault, int write,
 401                                unsigned long address)
 402 {
 403         extern void sun4c_update_mmu_cache(struct vm_area_struct *,
 404                                            unsigned long,pte_t);
 405         extern pte_t *sun4c_pte_offset_kernel(pmd_t *,unsigned long);
 406         struct task_struct *tsk = current;
 407         struct mm_struct *mm = tsk->mm;
 408         pgd_t *pgdp;
 409         pte_t *ptep;
 410
 411         if (text_fault) {
 412                 address = regs->pc;
 413         } else if (!write &&
 414                    !(regs->psr & PSR_PS)) {
 415                 unsigned int insn, *ip;
 416
 417                 ip = (unsigned int *)regs->pc;
 418                 if (! get_user(insn, ip)) {
 419                         if ((insn & 0xc1680000) == 0xc0680000)
 420                                 write = 1;
 421                 }
 422         }
 423
 424         if (!mm) {
 425                 /* We are oopsing. */
 426                 do_sparc_fault(regs, text_fault, write, address);
 427                 BUG();  /* P3 Oops already, you bitch */
 428         }
 429
 430         pgdp = pgd_offset(mm, address);
 431         ptep = sun4c_pte_offset_kernel((pmd_t *) pgdp, address);
 432
 433         if (pgd_val(*pgdp)) {
 434             if (write) {
 435                 if ((pte_val(*ptep) & (_SUN4C_PAGE_WRITE|_SUN4C_PAGE_PRESENT))
 436                                    == (_SUN4C_PAGE_WRITE|_SUN4C_PAGE_PRESENT)) {
 437                         unsigned long flags;
 438
 439                         *ptep = __pte(pte_val(*ptep) | _SUN4C_PAGE_ACCESSED |
 440                                       _SUN4C_PAGE_MODIFIED |
 441                                       _SUN4C_PAGE_VALID |
 442                                       _SUN4C_PAGE_DIRTY);
 443
 444                         local_irq_save(flags);
 445                         if (sun4c_get_segmap(address) != invalid_segment) {
 446                                 sun4c_put_pte(address, pte_val(*ptep));
 447                                 local_irq_restore(flags);
 448                                 return;
 449                         }
 450                         local_irq_restore(flags);
 451                 }
 452             } else {
 453                 if ((pte_val(*ptep) & (_SUN4C_PAGE_READ|_SUN4C_PAGE_PRESENT))
 454                                    == (_SUN4C_PAGE_READ|_SUN4C_PAGE_PRESENT)) {
 455                         unsigned long flags;
 456
 457                         *ptep = __pte(pte_val(*ptep) | _SUN4C_PAGE_ACCESSED |
 458                                       _SUN4C_PAGE_VALID);
 459
 460                         local_irq_save(flags);
 461                         if (sun4c_get_segmap(address) != invalid_segment) {
 462                                 sun4c_put_pte(address, pte_val(*ptep));
 463                                 local_irq_restore(flags);
 464                                 return;
 465                         }
 466                         local_irq_restore(flags);
 467                 }
 468             }
 469         }
 470
 471         /* This conditional is 'interesting'. */
 472         if (pgd_val(*pgdp) && !(write && !(pte_val(*ptep) & _SUN4C_PAGE_WRITE))
 473             && (pte_val(*ptep) & _SUN4C_PAGE_VALID))
 474                 /* Note: It is safe to not grab the MMAP semaphore here because
 475                  *       we know that update_mmu_cache() will not sleep for
 476                  *       any reason (at least not in the current implementation)
 477                  *       and therefore there is no danger of another thread getting
 478                  *       on the CPU and doing a shrink_mmap() on this vma.
 479                  */
 480                 sun4c_update_mmu_cache (find_vma(current->mm, address), address,
 481                                         *ptep);
 482         else
 483                 do_sparc_fault(regs, text_fault, write, address);
 484 }
 485
 486 /* This always deals with user addresses. */
 487 inline void force_user_fault(unsigned long address, int write)
 488 {
 489         struct vm_area_struct *vma;
 490         struct task_struct *tsk = current;
 491         struct mm_struct *mm = tsk->mm;
 492         siginfo_t info;
 493
 494         info.si_code = SEGV_MAPERR;
 495
 496 #if 0
 497         printk("wf<pid=%d,wr=%d,addr=%08lx>\n",
 498                tsk->pid, write, address);
 499 #endif
 500         down_read(&mm->mmap_sem);
 501         vma = find_vma(mm, address);
 502         if(!vma)
 503                 goto bad_area;
 504         if(vma->vm_start <= address)
 505                 goto good_area;
 506         if(!(vma->vm_flags & VM_GROWSDOWN))
 507                 goto bad_area;
 508         if(expand_stack(vma, address))
 509                 goto bad_area;
 510 good_area:
 511         info.si_code = SEGV_ACCERR;
 512         if(write) {
 513                 if(!(vma->vm_flags & VM_WRITE))
 514                         goto bad_area;
 515         } else {
 516                 if(!(vma->vm_flags & (VM_READ | VM_EXEC)))
 517                         goto bad_area;
 518         }
 519         if (!handle_mm_fault(mm, vma, address, write))
 520                 goto do_sigbus;
 521         up_read(&mm->mmap_sem);
 522         return;
 523 bad_area:
 524         up_read(&mm->mmap_sem);
 525 #if 0
 526         printk("Window whee %s [%d]: segfaults at %08lx\n",
 527                tsk->comm, tsk->pid, address);
 528 #endif
 529         info.si_signo = SIGSEGV;
 530         info.si_errno = 0;
 531         /* info.si_code set above to make clear whether
 532            this was a SEGV_MAPERR or SEGV_ACCERR fault.  */
 533         info.si_addr = (void *)address;
 534         info.si_trapno = 0;
 535         force_sig_info (SIGSEGV, &info, tsk);
 536         return;
 537
 538 do_sigbus:
 539         up_read(&mm->mmap_sem);
 540         info.si_signo = SIGBUS;
 541         info.si_errno = 0;
 542         info.si_code = BUS_ADRERR;
 543         info.si_addr = (void *)address;
 544         info.si_trapno = 0;
 545         force_sig_info (SIGBUS, &info, tsk);
 546 }
 547
 548 void window_overflow_fault(void)
 549 {
 550         unsigned long sp;
 551
 552         sp = current_thread_info()->rwbuf_stkptrs[0];
 553         if(((sp + 0x38) & PAGE_MASK) != (sp & PAGE_MASK))
 554                 force_user_fault(sp + 0x38, 1);
 555         force_user_fault(sp, 1);
 556 }
 557
 558 void window_underflow_fault(unsigned long sp)
 559 {
 560         if(((sp + 0x38) & PAGE_MASK) != (sp & PAGE_MASK))
 561                 force_user_fault(sp + 0x38, 0);
 562         force_user_fault(sp, 0);
 563 }
 564
 565 void window_ret_fault(struct pt_regs *regs)
 566 {
 567         unsigned long sp;
 568
 569         sp = regs->u_regs[UREG_FP];
 570         if(((sp + 0x38) & PAGE_MASK) != (sp & PAGE_MASK))
 571                 force_user_fault(sp + 0x38, 0);
 572         force_user_fault(sp, 0);
 573 }