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/page.h>
  27 #include <asm/pgtable.h>
  28 #include <asm/memreg.h>
  29 #include <asm/openprom.h>
  30 #include <asm/oplib.h>
  31 #include <asm/smp.h>
  32 #include <asm/traps.h>
  33 #include <asm/kdebug.h>
  34 #include <asm/uaccess.h>
  35
  36 #define ELEMENTS(arr) (sizeof (arr)/sizeof (arr[0]))
  37
  38 extern int prom_node_root;
  39
  40 /* At boot time we determine these two values necessary for setting
  41  * up the segment maps and page table entries (pte's).
  42  */
  43
  44 int num_segmaps, num_contexts;
  45 int invalid_segment;
  46
  47 /* various Virtual Address Cache parameters we find at boot time... */
  48
  49 int vac_size, vac_linesize, vac_do_hw_vac_flushes;
  50 int vac_entries_per_context, vac_entries_per_segment;
  51 int vac_entries_per_page;
  52
  53 /* Nice, simple, prom library does all the sweating for us. ;) */
  54 int prom_probe_memory (void)
  55 {
  56         register struct linux_mlist_v0 *mlist;
  57         register unsigned long bytes, base_paddr, tally;
  58         register int i;
  59
  60         i = 0;
  61         mlist= *prom_meminfo()->v0_available;
  62         bytes = tally = mlist->num_bytes;
  63         base_paddr = (unsigned long) mlist->start_adr;
  64
  65         sp_banks[0].base_addr = base_paddr;
  66         sp_banks[0].num_bytes = bytes;
  67
  68         while (mlist->theres_more != (void *) 0){
  69                 i++;
  70                 mlist = mlist->theres_more;
  71                 bytes = mlist->num_bytes;
  72                 tally += bytes;
  73                 if (i > SPARC_PHYS_BANKS-1) {
  74                         printk ("The machine has more banks than "
  75                                 "this kernel can support\n"
  76                                 "Increase the SPARC_PHYS_BANKS "
  77                                 "setting (currently %d)\n",
  78                                 SPARC_PHYS_BANKS);
  79                         i = SPARC_PHYS_BANKS-1;
  80                         break;
  81                 }
  82
  83                 sp_banks[i].base_addr = (unsigned long) mlist->start_adr;
  84                 sp_banks[i].num_bytes = mlist->num_bytes;
  85         }
  86
  87         i++;
  88         sp_banks[i].base_addr = 0xdeadbeef;
  89         sp_banks[i].num_bytes = 0;
  90
  91         /* Now mask all bank sizes on a page boundary, it is all we can
  92          * use anyways.
  93          */
  94         for(i=0; sp_banks[i].num_bytes != 0; i++)
  95                 sp_banks[i].num_bytes &= PAGE_MASK;
  96
  97         return tally;
  98 }
  99
 100 /* Traverse the memory lists in the prom to see how much physical we
 101  * have.
 102  */
 103 unsigned long
 104 probe_memory(void)
 105 {
 106         int total;
 107
 108         total = prom_probe_memory();
 109
 110         /* Oh man, much nicer, keep the dirt in promlib. */
 111         return total;
 112 }
 113
 114 extern void sun4c_complete_all_stores(void);
 115
 116 /* Whee, a level 15 NMI interrupt memory error.  Let's have fun... */
 117 asmlinkage void sparc_lvl15_nmi(struct pt_regs *regs, unsigned long serr,
 118                                 unsigned long svaddr, unsigned long aerr,
 119                                 unsigned long avaddr)
 120 {
 121         sun4c_complete_all_stores();
 122         printk("FAULT: NMI received\n");
 123         printk("SREGS: Synchronous Error %08lx\n", serr);
 124         printk("       Synchronous Vaddr %08lx\n", svaddr);
 125         printk("      Asynchronous Error %08lx\n", aerr);
 126         printk("      Asynchronous Vaddr %08lx\n", avaddr);
 127         if (sun4c_memerr_reg)
 128                 printk("     Memory Parity Error %08lx\n", *sun4c_memerr_reg);
 129         printk("REGISTER DUMP:\n");
 130         show_regs(regs);
 131         prom_halt();
 132 }
 133
 134 static void unhandled_fault(unsigned long, struct task_struct *,
 135                 struct pt_regs *) __attribute__ ((noreturn));
 136
 137 static void unhandled_fault(unsigned long address, struct task_struct *tsk,
 138                      struct pt_regs *regs)
 139 {
 140         if((unsigned long) address < PAGE_SIZE) {
 141                 printk(KERN_ALERT
 142                     "Unable to handle kernel NULL pointer dereference\n");
 143         } else {
 144                 printk(KERN_ALERT "Unable to handle kernel paging request "
 145                        "at virtual address %08lx\n", address);
 146         }
 147         printk(KERN_ALERT "tsk->{mm,active_mm}->context = %08lx\n",
 148                 (tsk->mm ? tsk->mm->context : tsk->active_mm->context));
 149         printk(KERN_ALERT "tsk->{mm,active_mm}->pgd = %08lx\n",
 150                 (tsk->mm ? (unsigned long) tsk->mm->pgd :
 151                         (unsigned long) tsk->active_mm->pgd));
 152         die_if_kernel("Oops", regs);
 153 }
 154
 155 asmlinkage int lookup_fault(unsigned long pc, unsigned long ret_pc,
 156                             unsigned long address)
 157 {
 158         struct pt_regs regs;
 159         unsigned long g2;
 160         unsigned int insn;
 161         int i;
 162
 163         i = search_extables_range(ret_pc, &g2);
 164         switch (i) {
 165         case 3:
 166                 /* load & store will be handled by fixup */
 167                 return 3;
 168
 169         case 1:
 170                 /* store will be handled by fixup, load will bump out */
 171                 /* for _to_ macros */
 172                 insn = *((unsigned int *) pc);
 173                 if ((insn >> 21) & 1)
 174                         return 1;
 175                 break;
 176
 177         case 2:
 178                 /* load will be handled by fixup, store will bump out */
 179                 /* for _from_ macros */
 180                 insn = *((unsigned int *) pc);
 181                 if (!((insn >> 21) & 1) || ((insn>>19)&0x3f) == 15)
 182                         return 2;
 183                 break;
 184
 185         default:
 186                 break;
 187         };
 188
 189         memset(&regs, 0, sizeof (regs));
 190         regs.pc = pc;
 191         regs.npc = pc + 4;
 192         __asm__ __volatile__(
 193                 "rd %%psr, %0\n\t"
 194                 "nop\n\t"
 195                 "nop\n\t"
 196                 "nop\n" : "=r" (regs.psr));
 197         unhandled_fault(address, current, &regs);
 198
 199         /* Not reached */
 200         return 0;
 201 }
 202
 203 extern unsigned long safe_compute_effective_address(struct pt_regs *,
 204                                                     unsigned int);
 205
 206 static unsigned long compute_si_addr(struct pt_regs *regs, int text_fault)
 207 {
 208         unsigned int insn;
 209
 210         if (text_fault)
 211                 return regs->pc;
 212
 213         if (regs->psr & PSR_PS) {
 214                 insn = *(unsigned int *) regs->pc;
 215         } else {
 216                 __get_user(insn, (unsigned int *) regs->pc);
 217         }
 218
 219         return safe_compute_effective_address(regs, insn);
 220 }
 221
 222 asmlinkage void do_sparc_fault(struct pt_regs *regs, int text_fault, int write,
 223                                unsigned long address)
 224 {
 225         struct vm_area_struct *vma;
 226         struct task_struct *tsk = current;
 227         struct mm_struct *mm = tsk->mm;
 228         unsigned int fixup;
 229         unsigned long g2;
 230         siginfo_t info;
 231         int from_user = !(regs->psr & PSR_PS);
 232
 233         if(text_fault)
 234                 address = regs->pc;
 235
 236         /*
 237          * We fault-in kernel-space virtual memory on-demand. The
 238          * 'reference' page table is init_mm.pgd.
 239          *
 240          * NOTE! We MUST NOT take any locks for this case. We may
 241          * be in an interrupt or a critical region, and should
 242          * only copy the information from the master page table,
 243          * nothing more.
 244          */
 245         if (!ARCH_SUN4C_SUN4 && address >= TASK_SIZE)
 246                 goto vmalloc_fault;
 247
 248         info.si_code = SEGV_MAPERR;
 249
 250         /*
 251          * If we're in an interrupt or have no user
 252          * context, we must not take the fault..
 253          */
 254         if (in_atomic() || !mm)
 255                 goto no_context;
 256
 257         down_read(&mm->mmap_sem);
 258
 259         /*
 260          * The kernel referencing a bad kernel pointer can lock up
 261          * a sun4c machine completely, so we must attempt recovery.
 262          */
 263         if(!from_user && address >= PAGE_OFFSET)
 264                 goto bad_area;
 265
 266         vma = find_vma(mm, address);
 267         if(!vma)
 268                 goto bad_area;
 269         if(vma->vm_start <= address)
 270                 goto good_area;
 271         if(!(vma->vm_flags & VM_GROWSDOWN))
 272                 goto bad_area;
 273         if(expand_stack(vma, address))
 274                 goto bad_area;
 275         /*
 276          * Ok, we have a good vm_area for this memory access, so
 277          * we can handle it..
 278          */
 279 good_area:
 280         info.si_code = SEGV_ACCERR;
 281         if(write) {
 282                 if(!(vma->vm_flags & VM_WRITE))
 283                         goto bad_area;
 284         } else {
 285                 /* Allow reads even for write-only mappings */
 286                 if(!(vma->vm_flags & (VM_READ | VM_EXEC)))
 287                         goto bad_area;
 288         }
 289
 290         /*
 291          * If for any reason at all we couldn't handle the fault,
 292          * make sure we exit gracefully rather than endlessly redo
 293          * the fault.
 294          */
 295         switch (handle_mm_fault(mm, vma, address, write)) {
 296         case VM_FAULT_SIGBUS:
 297                 goto do_sigbus;
 298         case VM_FAULT_OOM:
 299                 goto out_of_memory;
 300         case VM_FAULT_MAJOR:
 301                 current->maj_flt++;
 302                 break;
 303         case VM_FAULT_MINOR:
 304         default:
 305                 current->min_flt++;
 306                 break;
 307         }
 308         up_read(&mm->mmap_sem);
 309         return;
 310
 311         /*
 312          * Something tried to access memory that isn't in our memory map..
 313          * Fix it, but check if it's kernel or user first..
 314          */
 315 bad_area:
 316         up_read(&mm->mmap_sem);
 317
 318 bad_area_nosemaphore:
 319         /* User mode accesses just cause a SIGSEGV */
 320         if(from_user) {
 321 #if 0
 322                 printk("Fault whee %s [%d]: segfaults at %08lx pc=%08lx\n",
 323                        tsk->comm, tsk->pid, address, regs->pc);
 324 #endif
 325                 info.si_signo = SIGSEGV;
 326                 info.si_errno = 0;
 327                 /* info.si_code set above to make clear whether
 328                    this was a SEGV_MAPERR or SEGV_ACCERR fault.  */
 329                 info.si_addr = (void __user *)compute_si_addr(regs, text_fault);
 330                 info.si_trapno = 0;
 331                 force_sig_info (SIGSEGV, &info, tsk);
 332                 return;
 333         }
 334
 335         /* Is this in ex_table? */
 336 no_context:
 337         g2 = regs->u_regs[UREG_G2];
 338         if (!from_user && (fixup = search_extables_range(regs->pc, &g2))) {
 339                 if (fixup > 10) { /* Values below are reserved for other things */
 340                         extern const unsigned __memset_start[];
 341                         extern const unsigned __memset_end[];
 342                         extern const unsigned __csum_partial_copy_start[];
 343                         extern const unsigned __csum_partial_copy_end[];
 344
 345 #ifdef DEBUG_EXCEPTIONS
 346                         printk("Exception: PC<%08lx> faddr<%08lx>\n", regs->pc, address);
 347                         printk("EX_TABLE: insn<%08lx> fixup<%08x> g2<%08lx>\n",
 348                                 regs->pc, fixup, g2);
 349 #endif
 350                         if ((regs->pc >= (unsigned long)__memset_start &&
 351                              regs->pc < (unsigned long)__memset_end) ||
 352                             (regs->pc >= (unsigned long)__csum_partial_copy_start &&
 353                              regs->pc < (unsigned long)__csum_partial_copy_end)) {
 354                                 regs->u_regs[UREG_I4] = address;
 355                                 regs->u_regs[UREG_I5] = regs->pc;
 356                         }
 357                         regs->u_regs[UREG_G2] = g2;
 358                         regs->pc = fixup;
 359                         regs->npc = regs->pc + 4;
 360                         return;
 361                 }
 362         }
 363
 364         unhandled_fault (address, tsk, regs);
 365         do_exit(SIGKILL);
 366
 367 /*
 368  * We ran out of memory, or some other thing happened to us that made
 369  * us unable to handle the page fault gracefully.
 370  */
 371 out_of_memory:
 372         up_read(&mm->mmap_sem);
 373         printk("VM: killing process %s\n", tsk->comm);
 374         if (from_user)
 375                 do_exit(SIGKILL);
 376         goto no_context;
 377
 378 do_sigbus:
 379         up_read(&mm->mmap_sem);
 380         info.si_signo = SIGBUS;
 381         info.si_errno = 0;
 382         info.si_code = BUS_ADRERR;
 383         info.si_addr = (void __user *) compute_si_addr(regs, text_fault);
 384         info.si_trapno = 0;
 385         force_sig_info (SIGBUS, &info, tsk);
 386         if (!from_user)
 387                 goto no_context;
 388
 389 vmalloc_fault:
 390         {
 391                 /*
 392                  * Synchronize this task's top level page-table
 393                  * with the 'reference' page table.
 394                  */
 395                 int offset = pgd_index(address);
 396                 pgd_t *pgd, *pgd_k;
 397                 pmd_t *pmd, *pmd_k;
 398
 399                 pgd = tsk->active_mm->pgd + offset;
 400                 pgd_k = init_mm.pgd + offset;
 401
 402                 if (!pgd_present(*pgd)) {
 403                         if (!pgd_present(*pgd_k))
 404                                 goto bad_area_nosemaphore;
 405                         pgd_val(*pgd) = pgd_val(*pgd_k);
 406                         return;
 407                 }
 408
 409                 pmd = pmd_offset(pgd, address);
 410                 pmd_k = pmd_offset(pgd_k, address);
 411
 412                 if (pmd_present(*pmd) || !pmd_present(*pmd_k))
 413                         goto bad_area_nosemaphore;
 414                 *pmd = *pmd_k;
 415                 return;
 416         }
 417 }
 418
 419 asmlinkage void do_sun4c_fault(struct pt_regs *regs, int text_fault, int write,
 420                                unsigned long address)
 421 {
 422         extern void sun4c_update_mmu_cache(struct vm_area_struct *,
 423                                            unsigned long,pte_t);
 424         extern pte_t *sun4c_pte_offset_kernel(pmd_t *,unsigned long);
 425         struct task_struct *tsk = current;
 426         struct mm_struct *mm = tsk->mm;
 427         pgd_t *pgdp;
 428         pte_t *ptep;
 429
 430         if (text_fault) {
 431                 address = regs->pc;
 432         } else if (!write &&
 433                    !(regs->psr & PSR_PS)) {
 434                 unsigned int insn, __user *ip;
 435
 436                 ip = (unsigned int __user *)regs->pc;
 437                 if (!get_user(insn, ip)) {
 438                         if ((insn & 0xc1680000) == 0xc0680000)
 439                                 write = 1;
 440                 }
 441         }
 442
 443         if (!mm) {
 444                 /* We are oopsing. */
 445                 do_sparc_fault(regs, text_fault, write, address);
 446                 BUG();  /* P3 Oops already, you bitch */
 447         }
 448
 449         pgdp = pgd_offset(mm, address);
 450         ptep = sun4c_pte_offset_kernel((pmd_t *) pgdp, address);
 451
 452         if (pgd_val(*pgdp)) {
 453             if (write) {
 454                 if ((pte_val(*ptep) & (_SUN4C_PAGE_WRITE|_SUN4C_PAGE_PRESENT))
 455                                    == (_SUN4C_PAGE_WRITE|_SUN4C_PAGE_PRESENT)) {
 456                         unsigned long flags;
 457
 458                         *ptep = __pte(pte_val(*ptep) | _SUN4C_PAGE_ACCESSED |
 459                                       _SUN4C_PAGE_MODIFIED |
 460                                       _SUN4C_PAGE_VALID |
 461                                       _SUN4C_PAGE_DIRTY);
 462
 463                         local_irq_save(flags);
 464                         if (sun4c_get_segmap(address) != invalid_segment) {
 465                                 sun4c_put_pte(address, pte_val(*ptep));
 466                                 local_irq_restore(flags);
 467                                 return;
 468                         }
 469                         local_irq_restore(flags);
 470                 }
 471             } else {
 472                 if ((pte_val(*ptep) & (_SUN4C_PAGE_READ|_SUN4C_PAGE_PRESENT))
 473                                    == (_SUN4C_PAGE_READ|_SUN4C_PAGE_PRESENT)) {
 474                         unsigned long flags;
 475
 476                         *ptep = __pte(pte_val(*ptep) | _SUN4C_PAGE_ACCESSED |
 477                                       _SUN4C_PAGE_VALID);
 478
 479                         local_irq_save(flags);
 480                         if (sun4c_get_segmap(address) != invalid_segment) {
 481                                 sun4c_put_pte(address, pte_val(*ptep));
 482                                 local_irq_restore(flags);
 483                                 return;
 484                         }
 485                         local_irq_restore(flags);
 486                 }
 487             }
 488         }
 489
 490         /* This conditional is 'interesting'. */
 491         if (pgd_val(*pgdp) && !(write && !(pte_val(*ptep) & _SUN4C_PAGE_WRITE))
 492             && (pte_val(*ptep) & _SUN4C_PAGE_VALID))
 493                 /* Note: It is safe to not grab the MMAP semaphore here because
 494                  *       we know that update_mmu_cache() will not sleep for
 495                  *       any reason (at least not in the current implementation)
 496                  *       and therefore there is no danger of another thread getting
 497                  *       on the CPU and doing a shrink_mmap() on this vma.
 498                  */
 499                 sun4c_update_mmu_cache (find_vma(current->mm, address), address,
 500                                         *ptep);
 501         else
 502                 do_sparc_fault(regs, text_fault, write, address);
 503 }
 504
 505 /* This always deals with user addresses. */
 506 inline void force_user_fault(unsigned long address, int write)
 507 {
 508         struct vm_area_struct *vma;
 509         struct task_struct *tsk = current;
 510         struct mm_struct *mm = tsk->mm;
 511         siginfo_t info;
 512
 513         info.si_code = SEGV_MAPERR;
 514
 515 #if 0
 516         printk("wf<pid=%d,wr=%d,addr=%08lx>\n",
 517                tsk->pid, write, address);
 518 #endif
 519         down_read(&mm->mmap_sem);
 520         vma = find_vma(mm, address);
 521         if(!vma)
 522                 goto bad_area;
 523         if(vma->vm_start <= address)
 524                 goto good_area;
 525         if(!(vma->vm_flags & VM_GROWSDOWN))
 526                 goto bad_area;
 527         if(expand_stack(vma, address))
 528                 goto bad_area;
 529 good_area:
 530         info.si_code = SEGV_ACCERR;
 531         if(write) {
 532                 if(!(vma->vm_flags & VM_WRITE))
 533                         goto bad_area;
 534         } else {
 535                 if(!(vma->vm_flags & (VM_READ | VM_EXEC)))
 536                         goto bad_area;
 537         }
 538         switch (handle_mm_fault(mm, vma, address, write)) {
 539         case VM_FAULT_SIGBUS:
 540         case VM_FAULT_OOM:
 541                 goto do_sigbus;
 542         }
 543         up_read(&mm->mmap_sem);
 544         return;
 545 bad_area:
 546         up_read(&mm->mmap_sem);
 547 #if 0
 548         printk("Window whee %s [%d]: segfaults at %08lx\n",
 549                tsk->comm, tsk->pid, address);
 550 #endif
 551         info.si_signo = SIGSEGV;
 552         info.si_errno = 0;
 553         /* info.si_code set above to make clear whether
 554            this was a SEGV_MAPERR or SEGV_ACCERR fault.  */
 555         info.si_addr = (void __user *) address;
 556         info.si_trapno = 0;
 557         force_sig_info (SIGSEGV, &info, tsk);
 558         return;
 559
 560 do_sigbus:
 561         up_read(&mm->mmap_sem);
 562         info.si_signo = SIGBUS;
 563         info.si_errno = 0;
 564         info.si_code = BUS_ADRERR;
 565         info.si_addr = (void __user *) address;
 566         info.si_trapno = 0;
 567         force_sig_info (SIGBUS, &info, tsk);
 568 }
 569
 570 void window_overflow_fault(void)
 571 {
 572         unsigned long sp;
 573
 574         sp = current_thread_info()->rwbuf_stkptrs[0];
 575         if(((sp + 0x38) & PAGE_MASK) != (sp & PAGE_MASK))
 576                 force_user_fault(sp + 0x38, 1);
 577         force_user_fault(sp, 1);
 578 }
 579
 580 void window_underflow_fault(unsigned long sp)
 581 {
 582         if(((sp + 0x38) & PAGE_MASK) != (sp & PAGE_MASK))
 583                 force_user_fault(sp + 0x38, 0);
 584         force_user_fault(sp, 0);
 585 }
 586
 587 void window_ret_fault(struct pt_regs *regs)
 588 {
 589         unsigned long sp;
 590
 591         sp = regs->u_regs[UREG_FP];
 592         if(((sp + 0x38) & PAGE_MASK) != (sp & PAGE_MASK))
 593                 force_user_fault(sp + 0x38, 0);
 594         force_user_fault(sp, 0);
 595 }