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 extern unsigned long safe_compute_effective_address(struct pt_regs *,
 205                                                     unsigned int);
 206
 207 static unsigned long compute_si_addr(struct pt_regs *regs, int text_fault)
 208 {
 209         unsigned int insn;
 210
 211         if (text_fault)
 212                 return regs->pc;
 213
 214         if (regs->psr & PSR_PS) {
 215                 insn = *(unsigned int *) regs->pc;
 216         } else {
 217                 __get_user(insn, (unsigned int *) regs->pc);
 218         }
 219
 220         return safe_compute_effective_address(regs, insn);
 221 }
 222
 223 asmlinkage void do_sparc_fault(struct pt_regs *regs, int text_fault, int write,
 224                                unsigned long address)
 225 {
 226         struct vm_area_struct *vma;
 227         struct task_struct *tsk = current;
 228         struct mm_struct *mm = tsk->mm;
 229         unsigned int fixup;
 230         unsigned long g2;
 231         siginfo_t info;
 232         int from_user = !(regs->psr & PSR_PS);
 233
 234         if(text_fault)
 235                 address = regs->pc;
 236
 237         /*
 238          * We fault-in kernel-space virtual memory on-demand. The
 239          * 'reference' page table is init_mm.pgd.
 240          *
 241          * NOTE! We MUST NOT take any locks for this case. We may
 242          * be in an interrupt or a critical region, and should
 243          * only copy the information from the master page table,
 244          * nothing more.
 245          */
 246         if (!ARCH_SUN4C_SUN4 && address >= TASK_SIZE)
 247                 goto vmalloc_fault;
 248
 249         info.si_code = SEGV_MAPERR;
 250
 251         /*
 252          * If we're in an interrupt or have no user
 253          * context, we must not take the fault..
 254          */
 255         if (in_atomic() || !mm)
 256                 goto no_context;
 257
 258         down_read(&mm->mmap_sem);
 259
 260         /*
 261          * The kernel referencing a bad kernel pointer can lock up
 262          * a sun4c machine completely, so we must attempt recovery.
 263          */
 264         if(!from_user && address >= PAGE_OFFSET)
 265                 goto bad_area;
 266
 267         vma = find_vma(mm, address);
 268         if(!vma)
 269                 goto bad_area;
 270         if(vma->vm_start <= address)
 271                 goto good_area;
 272         if(!(vma->vm_flags & VM_GROWSDOWN))
 273                 goto bad_area;
 274         if(expand_stack(vma, address))
 275                 goto bad_area;
 276         /*
 277          * Ok, we have a good vm_area for this memory access, so
 278          * we can handle it..
 279          */
 280 good_area:
 281         info.si_code = SEGV_ACCERR;
 282         if(write) {
 283                 if(!(vma->vm_flags & VM_WRITE))
 284                         goto bad_area;
 285         } else {
 286                 /* Allow reads even for write-only mappings */
 287                 if(!(vma->vm_flags & (VM_READ | VM_EXEC)))
 288                         goto bad_area;
 289         }
 290
 291         /*
 292          * If for any reason at all we couldn't handle the fault,
 293          * make sure we exit gracefully rather than endlessly redo
 294          * the fault.
 295          */
 296         switch (handle_mm_fault(mm, vma, address, write)) {
 297         case VM_FAULT_SIGBUS:
 298                 goto do_sigbus;
 299         case VM_FAULT_OOM:
 300                 goto out_of_memory;
 301         case VM_FAULT_MAJOR:
 302                 current->maj_flt++;
 303                 break;
 304         case VM_FAULT_MINOR:
 305         default:
 306                 current->min_flt++;
 307                 break;
 308         }
 309         up_read(&mm->mmap_sem);
 310         return;
 311
 312         /*
 313          * Something tried to access memory that isn't in our memory map..
 314          * Fix it, but check if it's kernel or user first..
 315          */
 316 bad_area:
 317         up_read(&mm->mmap_sem);
 318
 319 bad_area_nosemaphore:
 320         /* User mode accesses just cause a SIGSEGV */
 321         if(from_user) {
 322 #if 0
 323                 printk("Fault whee %s [%d]: segfaults at %08lx pc=%08lx\n",
 324                        tsk->comm, tsk->pid, address, regs->pc);
 325 #endif
 326                 info.si_signo = SIGSEGV;
 327                 info.si_errno = 0;
 328                 /* info.si_code set above to make clear whether
 329                    this was a SEGV_MAPERR or SEGV_ACCERR fault.  */
 330                 info.si_addr = (void __user *)compute_si_addr(regs, text_fault);
 331                 info.si_trapno = 0;
 332                 force_sig_info (SIGSEGV, &info, tsk);
 333                 return;
 334         }
 335
 336         /* Is this in ex_table? */
 337 no_context:
 338         g2 = regs->u_regs[UREG_G2];
 339         if (!from_user && (fixup = search_extables_range(regs->pc, &g2))) {
 340                 if (fixup > 10) { /* Values below are reserved for other things */
 341                         extern const unsigned __memset_start[];
 342                         extern const unsigned __memset_end[];
 343                         extern const unsigned __csum_partial_copy_start[];
 344                         extern const unsigned __csum_partial_copy_end[];
 345
 346 #ifdef DEBUG_EXCEPTIONS
 347                         printk("Exception: PC<%08lx> faddr<%08lx>\n", regs->pc, address);
 348                         printk("EX_TABLE: insn<%08lx> fixup<%08x> g2<%08lx>\n",
 349                                 regs->pc, fixup, g2);
 350 #endif
 351                         if ((regs->pc >= (unsigned long)__memset_start &&
 352                              regs->pc < (unsigned long)__memset_end) ||
 353                             (regs->pc >= (unsigned long)__csum_partial_copy_start &&
 354                              regs->pc < (unsigned long)__csum_partial_copy_end)) {
 355                                 regs->u_regs[UREG_I4] = address;
 356                                 regs->u_regs[UREG_I5] = regs->pc;
 357                         }
 358                         regs->u_regs[UREG_G2] = g2;
 359                         regs->pc = fixup;
 360                         regs->npc = regs->pc + 4;
 361                         return;
 362                 }
 363         }
 364
 365         unhandled_fault (address, tsk, regs);
 366         do_exit(SIGKILL);
 367
 368 /*
 369  * We ran out of memory, or some other thing happened to us that made
 370  * us unable to handle the page fault gracefully.
 371  */
 372 out_of_memory:
 373         up_read(&mm->mmap_sem);
 374         printk("VM: killing process %s\n", tsk->comm);
 375         if (from_user)
 376                 do_exit(SIGKILL);
 377         goto no_context;
 378
 379 do_sigbus:
 380         up_read(&mm->mmap_sem);
 381         info.si_signo = SIGBUS;
 382         info.si_errno = 0;
 383         info.si_code = BUS_ADRERR;
 384         info.si_addr = (void __user *) compute_si_addr(regs, text_fault);
 385         info.si_trapno = 0;
 386         force_sig_info (SIGBUS, &info, tsk);
 387         if (!from_user)
 388                 goto no_context;
 389
 390 vmalloc_fault:
 391         {
 392                 /*
 393                  * Synchronize this task's top level page-table
 394                  * with the 'reference' page table.
 395                  */
 396                 int offset = pgd_index(address);
 397                 pgd_t *pgd, *pgd_k;
 398                 pmd_t *pmd, *pmd_k;
 399
 400                 pgd = tsk->active_mm->pgd + offset;
 401                 pgd_k = init_mm.pgd + offset;
 402
 403                 if (!pgd_present(*pgd)) {
 404                         if (!pgd_present(*pgd_k))
 405                                 goto bad_area_nosemaphore;
 406                         pgd_val(*pgd) = pgd_val(*pgd_k);
 407                         return;
 408                 }
 409
 410                 pmd = pmd_offset(pgd, address);
 411                 pmd_k = pmd_offset(pgd_k, address);
 412
 413                 if (pmd_present(*pmd) || !pmd_present(*pmd_k))
 414                         goto bad_area_nosemaphore;
 415                 *pmd = *pmd_k;
 416                 return;
 417         }
 418 }
 419
 420 asmlinkage void do_sun4c_fault(struct pt_regs *regs, int text_fault, int write,
 421                                unsigned long address)
 422 {
 423         extern void sun4c_update_mmu_cache(struct vm_area_struct *,
 424                                            unsigned long,pte_t);
 425         extern pte_t *sun4c_pte_offset_kernel(pmd_t *,unsigned long);
 426         struct task_struct *tsk = current;
 427         struct mm_struct *mm = tsk->mm;
 428         pgd_t *pgdp;
 429         pte_t *ptep;
 430
 431         if (text_fault) {
 432                 address = regs->pc;
 433         } else if (!write &&
 434                    !(regs->psr & PSR_PS)) {
 435                 unsigned int insn, __user *ip;
 436
 437                 ip = (unsigned int __user *)regs->pc;
 438                 if (!get_user(insn, ip)) {
 439                         if ((insn & 0xc1680000) == 0xc0680000)
 440                                 write = 1;
 441                 }
 442         }
 443
 444         if (!mm) {
 445                 /* We are oopsing. */
 446                 do_sparc_fault(regs, text_fault, write, address);
 447                 BUG();  /* P3 Oops already, you bitch */
 448         }
 449
 450         pgdp = pgd_offset(mm, address);
 451         ptep = sun4c_pte_offset_kernel((pmd_t *) pgdp, address);
 452
 453         if (pgd_val(*pgdp)) {
 454             if (write) {
 455                 if ((pte_val(*ptep) & (_SUN4C_PAGE_WRITE|_SUN4C_PAGE_PRESENT))
 456                                    == (_SUN4C_PAGE_WRITE|_SUN4C_PAGE_PRESENT)) {
 457                         unsigned long flags;
 458
 459                         *ptep = __pte(pte_val(*ptep) | _SUN4C_PAGE_ACCESSED |
 460                                       _SUN4C_PAGE_MODIFIED |
 461                                       _SUN4C_PAGE_VALID |
 462                                       _SUN4C_PAGE_DIRTY);
 463
 464                         local_irq_save(flags);
 465                         if (sun4c_get_segmap(address) != invalid_segment) {
 466                                 sun4c_put_pte(address, pte_val(*ptep));
 467                                 local_irq_restore(flags);
 468                                 return;
 469                         }
 470                         local_irq_restore(flags);
 471                 }
 472             } else {
 473                 if ((pte_val(*ptep) & (_SUN4C_PAGE_READ|_SUN4C_PAGE_PRESENT))
 474                                    == (_SUN4C_PAGE_READ|_SUN4C_PAGE_PRESENT)) {
 475                         unsigned long flags;
 476
 477                         *ptep = __pte(pte_val(*ptep) | _SUN4C_PAGE_ACCESSED |
 478                                       _SUN4C_PAGE_VALID);
 479
 480                         local_irq_save(flags);
 481                         if (sun4c_get_segmap(address) != invalid_segment) {
 482                                 sun4c_put_pte(address, pte_val(*ptep));
 483                                 local_irq_restore(flags);
 484                                 return;
 485                         }
 486                         local_irq_restore(flags);
 487                 }
 488             }
 489         }
 490
 491         /* This conditional is 'interesting'. */
 492         if (pgd_val(*pgdp) && !(write && !(pte_val(*ptep) & _SUN4C_PAGE_WRITE))
 493             && (pte_val(*ptep) & _SUN4C_PAGE_VALID))
 494                 /* Note: It is safe to not grab the MMAP semaphore here because
 495                  *       we know that update_mmu_cache() will not sleep for
 496                  *       any reason (at least not in the current implementation)
 497                  *       and therefore there is no danger of another thread getting
 498                  *       on the CPU and doing a shrink_mmap() on this vma.
 499                  */
 500                 sun4c_update_mmu_cache (find_vma(current->mm, address), address,
 501                                         *ptep);
 502         else
 503                 do_sparc_fault(regs, text_fault, write, address);
 504 }
 505
 506 /* This always deals with user addresses. */
 507 inline void force_user_fault(unsigned long address, int write)
 508 {
 509         struct vm_area_struct *vma;
 510         struct task_struct *tsk = current;
 511         struct mm_struct *mm = tsk->mm;
 512         siginfo_t info;
 513
 514         info.si_code = SEGV_MAPERR;
 515
 516 #if 0
 517         printk("wf<pid=%d,wr=%d,addr=%08lx>\n",
 518                tsk->pid, write, address);
 519 #endif
 520         down_read(&mm->mmap_sem);
 521         vma = find_vma(mm, address);
 522         if(!vma)
 523                 goto bad_area;
 524         if(vma->vm_start <= address)
 525                 goto good_area;
 526         if(!(vma->vm_flags & VM_GROWSDOWN))
 527                 goto bad_area;
 528         if(expand_stack(vma, address))
 529                 goto bad_area;
 530 good_area:
 531         info.si_code = SEGV_ACCERR;
 532         if(write) {
 533                 if(!(vma->vm_flags & VM_WRITE))
 534                         goto bad_area;
 535         } else {
 536                 if(!(vma->vm_flags & (VM_READ | VM_EXEC)))
 537                         goto bad_area;
 538         }
 539         switch (handle_mm_fault(mm, vma, address, write)) {
 540         case VM_FAULT_SIGBUS:
 541         case VM_FAULT_OOM:
 542                 goto do_sigbus;
 543         }
 544         up_read(&mm->mmap_sem);
 545         return;
 546 bad_area:
 547         up_read(&mm->mmap_sem);
 548 #if 0
 549         printk("Window whee %s [%d]: segfaults at %08lx\n",
 550                tsk->comm, tsk->pid, address);
 551 #endif
 552         info.si_signo = SIGSEGV;
 553         info.si_errno = 0;
 554         /* info.si_code set above to make clear whether
 555            this was a SEGV_MAPERR or SEGV_ACCERR fault.  */
 556         info.si_addr = (void __user *) address;
 557         info.si_trapno = 0;
 558         force_sig_info (SIGSEGV, &info, tsk);
 559         return;
 560
 561 do_sigbus:
 562         up_read(&mm->mmap_sem);
 563         info.si_signo = SIGBUS;
 564         info.si_errno = 0;
 565         info.si_code = BUS_ADRERR;
 566         info.si_addr = (void __user *) address;
 567         info.si_trapno = 0;
 568         force_sig_info (SIGBUS, &info, tsk);
 569 }
 570
 571 void window_overflow_fault(void)
 572 {
 573         unsigned long sp;
 574
 575         sp = current_thread_info()->rwbuf_stkptrs[0];
 576         if(((sp + 0x38) & PAGE_MASK) != (sp & PAGE_MASK))
 577                 force_user_fault(sp + 0x38, 1);
 578         force_user_fault(sp, 1);
 579 }
 580
 581 void window_underflow_fault(unsigned long sp)
 582 {
 583         if(((sp + 0x38) & PAGE_MASK) != (sp & PAGE_MASK))
 584                 force_user_fault(sp + 0x38, 0);
 585         force_user_fault(sp, 0);
 586 }
 587
 588 void window_ret_fault(struct pt_regs *regs)
 589 {
 590         unsigned long sp;
 591
 592         sp = regs->u_regs[UREG_FP];
 593         if(((sp + 0x38) & PAGE_MASK) != (sp & PAGE_MASK))
 594                 force_user_fault(sp + 0x38, 0);
 595         force_user_fault(sp, 0);
 596 }