arch/sparc64/mm/fault.c

   1 /* $Id: fault.c,v 1.59 2002/02/09 19:49:31 davem Exp $
   2  * arch/sparc64/mm/fault.c: Page fault handlers for the 64-bit Sparc.
   3  *
   4  * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
   5  * Copyright (C) 1997, 1999 Jakub Jelinek (jj@ultra.linux.cz)
   6  */
   7
   8 #include <asm/head.h>
   9
  10 #include <linux/string.h>
  11 #include <linux/types.h>
  12 #include <linux/sched.h>
  13 #include <linux/ptrace.h>
  14 #include <linux/mman.h>
  15 #include <linux/signal.h>
  16 #include <linux/mm.h>
  17 #include <linux/module.h>
  18 #include <linux/smp_lock.h>
  19 #include <linux/init.h>
  20 #include <linux/interrupt.h>
  21
  22 #include <asm/page.h>
  23 #include <asm/pgtable.h>
  24 #include <asm/openprom.h>
  25 #include <asm/oplib.h>
  26 #include <asm/uaccess.h>
  27 #include <asm/asi.h>
  28 #include <asm/lsu.h>
  29 #include <asm/sections.h>
  30 #include <asm/kdebug.h>
  31
  32 #define ELEMENTS(arr) (sizeof (arr)/sizeof (arr[0]))
  33
  34 extern struct sparc_phys_banks sp_banks[SPARC_PHYS_BANKS];
  35
  36 /*
  37  * To debug kernel during syscall entry.
  38  */
  39 void syscall_trace_entry(struct pt_regs *regs)
  40 {
  41         printk("scall entry: %s[%d]/cpu%d: %d\n", current->comm, current->pid, smp_processor_id(), (int) regs->u_regs[UREG_G1]);
  42 }
  43
  44 /*
  45  * To debug kernel during syscall exit.
  46  */
  47 void syscall_trace_exit(struct pt_regs *regs)
  48 {
  49         printk("scall exit: %s[%d]/cpu%d: %d\n", current->comm, current->pid, smp_processor_id(), (int) regs->u_regs[UREG_G1]);
  50 }
  51
  52 /*
  53  * To debug kernel to catch accesses to certain virtual/physical addresses.
  54  * Mode = 0 selects physical watchpoints, mode = 1 selects virtual watchpoints.
  55  * flags = VM_READ watches memread accesses, flags = VM_WRITE watches memwrite accesses.
  56  * Caller passes in a 64bit aligned addr, with mask set to the bytes that need to be
  57  * watched. This is only useful on a single cpu machine for now. After the watchpoint
  58  * is detected, the process causing it will be killed, thus preventing an infinite loop.
  59  */
  60 void set_brkpt(unsigned long addr, unsigned char mask, int flags, int mode)
  61 {
  62         unsigned long lsubits;
  63
  64         __asm__ __volatile__("ldxa [%%g0] %1, %0"
  65                              : "=r" (lsubits)
  66                              : "i" (ASI_LSU_CONTROL));
  67         lsubits &= ~(LSU_CONTROL_PM | LSU_CONTROL_VM |
  68                      LSU_CONTROL_PR | LSU_CONTROL_VR |
  69                      LSU_CONTROL_PW | LSU_CONTROL_VW);
  70
  71         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
  72                              "membar    #Sync"
  73                              : /* no outputs */
  74                              : "r" (addr), "r" (mode ? VIRT_WATCHPOINT : PHYS_WATCHPOINT),
  75                                "i" (ASI_DMMU));
  76
  77         lsubits |= ((unsigned long)mask << (mode ? 25 : 33));
  78         if (flags & VM_READ)
  79                 lsubits |= (mode ? LSU_CONTROL_VR : LSU_CONTROL_PR);
  80         if (flags & VM_WRITE)
  81                 lsubits |= (mode ? LSU_CONTROL_VW : LSU_CONTROL_PW);
  82         __asm__ __volatile__("stxa %0, [%%g0] %1\n\t"
  83                              "membar #Sync"
  84                              : /* no outputs */
  85                              : "r" (lsubits), "i" (ASI_LSU_CONTROL)
  86                              : "memory");
  87 }
  88
  89 /* Nice, simple, prom library does all the sweating for us. ;) */
  90 unsigned long __init prom_probe_memory (void)
  91 {
  92         register struct linux_mlist_p1275 *mlist;
  93         register unsigned long bytes, base_paddr, tally;
  94         register int i;
  95
  96         i = 0;
  97         mlist = *prom_meminfo()->p1275_available;
  98         bytes = tally = mlist->num_bytes;
  99         base_paddr = mlist->start_adr;
 100
 101         sp_banks[0].base_addr = base_paddr;
 102         sp_banks[0].num_bytes = bytes;
 103
 104         while (mlist->theres_more != (void *) 0) {
 105                 i++;
 106                 mlist = mlist->theres_more;
 107                 bytes = mlist->num_bytes;
 108                 tally += bytes;
 109                 if (i >= SPARC_PHYS_BANKS-1) {
 110                         printk ("The machine has more banks than "
 111                                 "this kernel can support\n"
 112                                 "Increase the SPARC_PHYS_BANKS "
 113                                 "setting (currently %d)\n",
 114                                 SPARC_PHYS_BANKS);
 115                         i = SPARC_PHYS_BANKS-1;
 116                         break;
 117                 }
 118
 119                 sp_banks[i].base_addr = mlist->start_adr;
 120                 sp_banks[i].num_bytes = mlist->num_bytes;
 121         }
 122
 123         i++;
 124         sp_banks[i].base_addr = 0xdeadbeefbeefdeadUL;
 125         sp_banks[i].num_bytes = 0;
 126
 127         /* Now mask all bank sizes on a page boundary, it is all we can
 128          * use anyways.
 129          */
 130         for (i = 0; sp_banks[i].num_bytes != 0; i++)
 131                 sp_banks[i].num_bytes &= PAGE_MASK;
 132
 133         return tally;
 134 }
 135
 136 static void unhandled_fault(unsigned long address, struct task_struct *tsk,
 137                             struct pt_regs *regs)
 138 {
 139         if ((unsigned long) address < PAGE_SIZE) {
 140                 printk(KERN_ALERT "Unable to handle kernel NULL "
 141                        "pointer dereference\n");
 142         } else {
 143                 printk(KERN_ALERT "Unable to handle kernel paging request "
 144                        "at virtual address %016lx\n", (unsigned long)address);
 145         }
 146         printk(KERN_ALERT "tsk->{mm,active_mm}->context = %016lx\n",
 147                (tsk->mm ?
 148                 CTX_HWBITS(tsk->mm->context) :
 149                 CTX_HWBITS(tsk->active_mm->context)));
 150         printk(KERN_ALERT "tsk->{mm,active_mm}->pgd = %016lx\n",
 151                (tsk->mm ? (unsigned long) tsk->mm->pgd :
 152                           (unsigned long) tsk->active_mm->pgd));
 153         if (notify_die(DIE_GPF, "general protection fault", regs,
 154                        0, 0, SIGSEGV) == NOTIFY_STOP)
 155                 return;
 156         die_if_kernel("Oops", regs);
 157 }
 158
 159 static void bad_kernel_pc(struct pt_regs *regs)
 160 {
 161         unsigned long *ksp;
 162
 163         printk(KERN_CRIT "OOPS: Bogus kernel PC [%016lx] in fault handler\n",
 164                regs->tpc);
 165         __asm__("mov %%sp, %0" : "=r" (ksp));
 166         show_stack(current, ksp);
 167         unhandled_fault(regs->tpc, current, regs);
 168 }
 169
 170 /*
 171  * We now make sure that mmap_sem is held in all paths that call
 172  * this. Additionally, to prevent kswapd from ripping ptes from
 173  * under us, raise interrupts around the time that we look at the
 174  * pte, kswapd will have to wait to get his smp ipi response from
 175  * us. This saves us having to get page_table_lock.
 176  */
 177 static unsigned int get_user_insn(unsigned long tpc)
 178 {
 179         pgd_t *pgdp = pgd_offset(current->mm, tpc);
 180         pud_t *pudp;
 181         pmd_t *pmdp;
 182         pte_t *ptep, pte;
 183         unsigned long pa;
 184         u32 insn = 0;
 185         unsigned long pstate;
 186
 187         if (pgd_none(*pgdp))
 188                 goto outret;
 189         pudp = pud_offset(pgdp, tpc);
 190         if (pud_none(*pudp))
 191                 goto outret;
 192         pmdp = pmd_offset(pudp, tpc);
 193         if (pmd_none(*pmdp))
 194                 goto outret;
 195
 196         /* This disables preemption for us as well. */
 197         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 198         __asm__ __volatile__("wrpr %0, %1, %%pstate"
 199                                 : : "r" (pstate), "i" (PSTATE_IE));
 200         ptep = pte_offset_map(pmdp, tpc);
 201         pte = *ptep;
 202         if (!pte_present(pte))
 203                 goto out;
 204
 205         pa  = (pte_val(pte) & _PAGE_PADDR);
 206         pa += (tpc & ~PAGE_MASK);
 207
 208         /* Use phys bypass so we don't pollute dtlb/dcache. */
 209         __asm__ __volatile__("lduwa [%1] %2, %0"
 210                              : "=r" (insn)
 211                              : "r" (pa), "i" (ASI_PHYS_USE_EC));
 212
 213 out:
 214         pte_unmap(ptep);
 215         __asm__ __volatile__("wrpr %0, 0x0, %%pstate" : : "r" (pstate));
 216 outret:
 217         return insn;
 218 }
 219
 220 extern unsigned long compute_effective_address(struct pt_regs *, unsigned int, unsigned int);
 221
 222 static void do_fault_siginfo(int code, int sig, struct pt_regs *regs,
 223                              unsigned int insn, int fault_code)
 224 {
 225         siginfo_t info;
 226
 227         info.si_code = code;
 228         info.si_signo = sig;
 229         info.si_errno = 0;
 230         if (fault_code & FAULT_CODE_ITLB)
 231                 info.si_addr = (void __user *) regs->tpc;
 232         else
 233                 info.si_addr = (void __user *)
 234                         compute_effective_address(regs, insn, 0);
 235         info.si_trapno = 0;
 236         force_sig_info(sig, &info, current);
 237 }
 238
 239 extern int handle_ldf_stq(u32, struct pt_regs *);
 240 extern int handle_ld_nf(u32, struct pt_regs *);
 241
 242 static unsigned int get_fault_insn(struct pt_regs *regs, unsigned int insn)
 243 {
 244         if (!insn) {
 245                 if (!regs->tpc || (regs->tpc & 0x3))
 246                         return 0;
 247                 if (regs->tstate & TSTATE_PRIV) {
 248                         insn = *(unsigned int *) regs->tpc;
 249                 } else {
 250                         insn = get_user_insn(regs->tpc);
 251                 }
 252         }
 253         return insn;
 254 }
 255
 256 static void do_kernel_fault(struct pt_regs *regs, int si_code, int fault_code,
 257                             unsigned int insn, unsigned long address)
 258 {
 259         unsigned long g2;
 260         unsigned char asi = ASI_P;
 261
 262         if ((!insn) && (regs->tstate & TSTATE_PRIV))
 263                 goto cannot_handle;
 264
 265         /* If user insn could be read (thus insn is zero), that
 266          * is fine.  We will just gun down the process with a signal
 267          * in that case.
 268          */
 269
 270         if (!(fault_code & (FAULT_CODE_WRITE|FAULT_CODE_ITLB)) &&
 271             (insn & 0xc0800000) == 0xc0800000) {
 272                 if (insn & 0x2000)
 273                         asi = (regs->tstate >> 24);
 274                 else
 275                         asi = (insn >> 5);
 276                 if ((asi & 0xf2) == 0x82) {
 277                         if (insn & 0x1000000) {
 278                                 handle_ldf_stq(insn, regs);
 279                         } else {
 280                                 /* This was a non-faulting load. Just clear the
 281                                  * destination register(s) and continue with the next
 282                                  * instruction. -jj
 283                                  */
 284                                 handle_ld_nf(insn, regs);
 285                         }
 286                         return;
 287                 }
 288         }
 289
 290         g2 = regs->u_regs[UREG_G2];
 291
 292         /* Is this in ex_table? */
 293         if (regs->tstate & TSTATE_PRIV) {
 294                 unsigned long fixup;
 295
 296                 if (asi == ASI_P && (insn & 0xc0800000) == 0xc0800000) {
 297                         if (insn & 0x2000)
 298                                 asi = (regs->tstate >> 24);
 299                         else
 300                                 asi = (insn >> 5);
 301                 }
 302
 303                 /* Look in asi.h: All _S asis have LS bit set */
 304                 if ((asi & 0x1) &&
 305                     (fixup = search_extables_range(regs->tpc, &g2))) {
 306                         regs->tpc = fixup;
 307                         regs->tnpc = regs->tpc + 4;
 308                         regs->u_regs[UREG_G2] = g2;
 309                         return;
 310                 }
 311         } else {
 312                 /* The si_code was set to make clear whether
 313                  * this was a SEGV_MAPERR or SEGV_ACCERR fault.
 314                  */
 315                 do_fault_siginfo(si_code, SIGSEGV, regs, insn, fault_code);
 316                 return;
 317         }
 318
 319 cannot_handle:
 320         unhandled_fault (address, current, regs);
 321 }
 322
 323 asmlinkage void do_sparc64_fault(struct pt_regs *regs)
 324 {
 325         struct mm_struct *mm = current->mm;
 326         struct vm_area_struct *vma;
 327         unsigned int insn = 0;
 328         int si_code, fault_code;
 329         unsigned long address;
 330
 331         fault_code = get_thread_fault_code();
 332
 333         if (notify_die(DIE_PAGE_FAULT, "page_fault", regs,
 334                        fault_code, 0, SIGSEGV) == NOTIFY_STOP)
 335                 return;
 336
 337         si_code = SEGV_MAPERR;
 338         address = current_thread_info()->fault_address;
 339
 340         if ((fault_code & FAULT_CODE_ITLB) &&
 341             (fault_code & FAULT_CODE_DTLB))
 342                 BUG();
 343
 344         if (regs->tstate & TSTATE_PRIV) {
 345                 unsigned long tpc = regs->tpc;
 346
 347                 /* Sanity check the PC. */
 348                 if ((tpc >= KERNBASE && tpc < (unsigned long) _etext) ||
 349                     (tpc >= MODULES_VADDR && tpc < MODULES_END)) {
 350                         /* Valid, no problems... */
 351                 } else {
 352                         bad_kernel_pc(regs);
 353                         return;
 354                 }
 355         }
 356
 357         /*
 358          * If we're in an interrupt or have no user
 359          * context, we must not take the fault..
 360          */
 361         if (in_atomic() || !mm)
 362                 goto intr_or_no_mm;
 363
 364         if (test_thread_flag(TIF_32BIT)) {
 365                 if (!(regs->tstate & TSTATE_PRIV))
 366                         regs->tpc &= 0xffffffff;
 367                 address &= 0xffffffff;
 368         }
 369
 370         if (!down_read_trylock(&mm->mmap_sem)) {
 371                 if ((regs->tstate & TSTATE_PRIV) &&
 372                     !search_exception_tables(regs->tpc)) {
 373                         insn = get_fault_insn(regs, insn);
 374                         goto handle_kernel_fault;
 375                 }
 376                 down_read(&mm->mmap_sem);
 377         }
 378
 379         vma = find_vma(mm, address);
 380         if (!vma)
 381                 goto bad_area;
 382
 383         /* Pure DTLB misses do not tell us whether the fault causing
 384          * load/store/atomic was a write or not, it only says that there
 385          * was no match.  So in such a case we (carefully) read the
 386          * instruction to try and figure this out.  It's an optimization
 387          * so it's ok if we can't do this.
 388          *
 389          * Special hack, window spill/fill knows the exact fault type.
 390          */
 391         if (((fault_code &
 392               (FAULT_CODE_DTLB | FAULT_CODE_WRITE | FAULT_CODE_WINFIXUP)) == FAULT_CODE_DTLB) &&
 393             (vma->vm_flags & VM_WRITE) != 0) {
 394                 insn = get_fault_insn(regs, 0);
 395                 if (!insn)
 396                         goto continue_fault;
 397                 if ((insn & 0xc0200000) == 0xc0200000 &&
 398                     (insn & 0x1780000) != 0x1680000) {
 399                         /* Don't bother updating thread struct value,
 400                          * because update_mmu_cache only cares which tlb
 401                          * the access came from.
 402                          */
 403                         fault_code |= FAULT_CODE_WRITE;
 404                 }
 405         }
 406 continue_fault:
 407
 408         if (vma->vm_start <= address)
 409                 goto good_area;
 410         if (!(vma->vm_flags & VM_GROWSDOWN))
 411                 goto bad_area;
 412         if (!(fault_code & FAULT_CODE_WRITE)) {
 413                 /* Non-faulting loads shouldn't expand stack. */
 414                 insn = get_fault_insn(regs, insn);
 415                 if ((insn & 0xc0800000) == 0xc0800000) {
 416                         unsigned char asi;
 417
 418                         if (insn & 0x2000)
 419                                 asi = (regs->tstate >> 24);
 420                         else
 421                                 asi = (insn >> 5);
 422                         if ((asi & 0xf2) == 0x82)
 423                                 goto bad_area;
 424                 }
 425         }
 426         if (expand_stack(vma, address))
 427                 goto bad_area;
 428         /*
 429          * Ok, we have a good vm_area for this memory access, so
 430          * we can handle it..
 431          */
 432 good_area:
 433         si_code = SEGV_ACCERR;
 434
 435         /* If we took a ITLB miss on a non-executable page, catch
 436          * that here.
 437          */
 438         if ((fault_code & FAULT_CODE_ITLB) && !(vma->vm_flags & VM_EXEC)) {
 439                 BUG_ON(address != regs->tpc);
 440                 BUG_ON(regs->tstate & TSTATE_PRIV);
 441                 goto bad_area;
 442         }
 443
 444         if (fault_code & FAULT_CODE_WRITE) {
 445                 if (!(vma->vm_flags & VM_WRITE))
 446                         goto bad_area;
 447
 448                 /* Spitfire has an icache which does not snoop
 449                  * processor stores.  Later processors do...
 450                  */
 451                 if (tlb_type == spitfire &&
 452                     (vma->vm_flags & VM_EXEC) != 0 &&
 453                     vma->vm_file != NULL)
 454                         set_thread_fault_code(fault_code |
 455                                               FAULT_CODE_BLKCOMMIT);
 456         } else {
 457                 /* Allow reads even for write-only mappings */
 458                 if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
 459                         goto bad_area;
 460         }
 461
 462         switch (handle_mm_fault(mm, vma, address, (fault_code & FAULT_CODE_WRITE))) {
 463         case VM_FAULT_MINOR:
 464                 current->min_flt++;
 465                 break;
 466         case VM_FAULT_MAJOR:
 467                 current->maj_flt++;
 468                 break;
 469         case VM_FAULT_SIGBUS:
 470                 goto do_sigbus;
 471         case VM_FAULT_OOM:
 472                 goto out_of_memory;
 473         default:
 474                 BUG();
 475         }
 476
 477         up_read(&mm->mmap_sem);
 478         goto fault_done;
 479
 480         /*
 481          * Something tried to access memory that isn't in our memory map..
 482          * Fix it, but check if it's kernel or user first..
 483          */
 484 bad_area:
 485         insn = get_fault_insn(regs, insn);
 486         up_read(&mm->mmap_sem);
 487
 488 handle_kernel_fault:
 489         do_kernel_fault(regs, si_code, fault_code, insn, address);
 490
 491         goto fault_done;
 492
 493 /*
 494  * We ran out of memory, or some other thing happened to us that made
 495  * us unable to handle the page fault gracefully.
 496  */
 497 out_of_memory:
 498         insn = get_fault_insn(regs, insn);
 499         up_read(&mm->mmap_sem);
 500         printk("VM: killing process %s\n", current->comm);
 501         if (!(regs->tstate & TSTATE_PRIV))
 502                 do_exit(SIGKILL);
 503         goto handle_kernel_fault;
 504
 505 intr_or_no_mm:
 506         insn = get_fault_insn(regs, 0);
 507         goto handle_kernel_fault;
 508
 509 do_sigbus:
 510         insn = get_fault_insn(regs, insn);
 511         up_read(&mm->mmap_sem);
 512
 513         /*
 514          * Send a sigbus, regardless of whether we were in kernel
 515          * or user mode.
 516          */
 517         do_fault_siginfo(BUS_ADRERR, SIGBUS, regs, insn, fault_code);
 518
 519         /* Kernel mode? Handle exceptions or die */
 520         if (regs->tstate & TSTATE_PRIV)
 521                 goto handle_kernel_fault;
 522
 523 fault_done:
 524         /* These values are no longer needed, clear them. */
 525         set_thread_fault_code(0);
 526         current_thread_info()->fault_address = 0;
 527 }