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/init.h>
  19 #include <linux/interrupt.h>
  20 #include <linux/kprobes.h>
  21 #include <linux/kallsyms.h>
  22 #include <linux/kdebug.h>
  23
  24 #include <asm/page.h>
  25 #include <asm/pgtable.h>
  26 #include <asm/openprom.h>
  27 #include <asm/oplib.h>
  28 #include <asm/uaccess.h>
  29 #include <asm/asi.h>
  30 #include <asm/lsu.h>
  31 #include <asm/sections.h>
  32 #include <asm/mmu_context.h>
  33
  34 #ifdef CONFIG_KPROBES
  35 static inline int notify_page_fault(struct pt_regs *regs)
  36 {
  37         int ret = 0;
  38
  39         /* kprobe_running() needs smp_processor_id() */
  40         if (!user_mode(regs)) {
  41                 preempt_disable();
  42                 if (kprobe_running() && kprobe_fault_handler(regs, 0))
  43                         ret = 1;
  44                 preempt_enable();
  45         }
  46         return ret;
  47 }
  48 #else
  49 static inline int notify_page_fault(struct pt_regs *regs)
  50 {
  51         return 0;
  52 }
  53 #endif
  54
  55 /*
  56  * To debug kernel to catch accesses to certain virtual/physical addresses.
  57  * Mode = 0 selects physical watchpoints, mode = 1 selects virtual watchpoints.
  58  * flags = VM_READ watches memread accesses, flags = VM_WRITE watches memwrite accesses.
  59  * Caller passes in a 64bit aligned addr, with mask set to the bytes that need to be
  60  * watched. This is only useful on a single cpu machine for now. After the watchpoint
  61  * is detected, the process causing it will be killed, thus preventing an infinite loop.
  62  */
  63 void set_brkpt(unsigned long addr, unsigned char mask, int flags, int mode)
  64 {
  65         unsigned long lsubits;
  66
  67         __asm__ __volatile__("ldxa [%%g0] %1, %0"
  68                              : "=r" (lsubits)
  69                              : "i" (ASI_LSU_CONTROL));
  70         lsubits &= ~(LSU_CONTROL_PM | LSU_CONTROL_VM |
  71                      LSU_CONTROL_PR | LSU_CONTROL_VR |
  72                      LSU_CONTROL_PW | LSU_CONTROL_VW);
  73
  74         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
  75                              "membar    #Sync"
  76                              : /* no outputs */
  77                              : "r" (addr), "r" (mode ? VIRT_WATCHPOINT : PHYS_WATCHPOINT),
  78                                "i" (ASI_DMMU));
  79
  80         lsubits |= ((unsigned long)mask << (mode ? 25 : 33));
  81         if (flags & VM_READ)
  82                 lsubits |= (mode ? LSU_CONTROL_VR : LSU_CONTROL_PR);
  83         if (flags & VM_WRITE)
  84                 lsubits |= (mode ? LSU_CONTROL_VW : LSU_CONTROL_PW);
  85         __asm__ __volatile__("stxa %0, [%%g0] %1\n\t"
  86                              "membar #Sync"
  87                              : /* no outputs */
  88                              : "r" (lsubits), "i" (ASI_LSU_CONTROL)
  89                              : "memory");
  90 }
  91
  92 static void __kprobes unhandled_fault(unsigned long address,
  93                                       struct task_struct *tsk,
  94                                       struct pt_regs *regs)
  95 {
  96         if ((unsigned long) address < PAGE_SIZE) {
  97                 printk(KERN_ALERT "Unable to handle kernel NULL "
  98                        "pointer dereference\n");
  99         } else {
 100                 printk(KERN_ALERT "Unable to handle kernel paging request "
 101                        "at virtual address %016lx\n", (unsigned long)address);
 102         }
 103         printk(KERN_ALERT "tsk->{mm,active_mm}->context = %016lx\n",
 104                (tsk->mm ?
 105                 CTX_HWBITS(tsk->mm->context) :
 106                 CTX_HWBITS(tsk->active_mm->context)));
 107         printk(KERN_ALERT "tsk->{mm,active_mm}->pgd = %016lx\n",
 108                (tsk->mm ? (unsigned long) tsk->mm->pgd :
 109                           (unsigned long) tsk->active_mm->pgd));
 110         die_if_kernel("Oops", regs);
 111 }
 112
 113 static void bad_kernel_pc(struct pt_regs *regs, unsigned long vaddr)
 114 {
 115         unsigned long *ksp;
 116
 117         printk(KERN_CRIT "OOPS: Bogus kernel PC [%016lx] in fault handler\n",
 118                regs->tpc);
 119         printk(KERN_CRIT "OOPS: RPC [%016lx]\n", regs->u_regs[15]);
 120         print_symbol("RPC: <%s>\n", regs->u_regs[15]);
 121         printk(KERN_CRIT "OOPS: Fault was to vaddr[%lx]\n", vaddr);
 122         __asm__("mov %%sp, %0" : "=r" (ksp));
 123         show_stack(current, ksp);
 124         unhandled_fault(regs->tpc, current, regs);
 125 }
 126
 127 /*
 128  * We now make sure that mmap_sem is held in all paths that call
 129  * this. Additionally, to prevent kswapd from ripping ptes from
 130  * under us, raise interrupts around the time that we look at the
 131  * pte, kswapd will have to wait to get his smp ipi response from
 132  * us. vmtruncate likewise. This saves us having to get pte lock.
 133  */
 134 static unsigned int get_user_insn(unsigned long tpc)
 135 {
 136         pgd_t *pgdp = pgd_offset(current->mm, tpc);
 137         pud_t *pudp;
 138         pmd_t *pmdp;
 139         pte_t *ptep, pte;
 140         unsigned long pa;
 141         u32 insn = 0;
 142         unsigned long pstate;
 143
 144         if (pgd_none(*pgdp))
 145                 goto outret;
 146         pudp = pud_offset(pgdp, tpc);
 147         if (pud_none(*pudp))
 148                 goto outret;
 149         pmdp = pmd_offset(pudp, tpc);
 150         if (pmd_none(*pmdp))
 151                 goto outret;
 152
 153         /* This disables preemption for us as well. */
 154         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 155         __asm__ __volatile__("wrpr %0, %1, %%pstate"
 156                                 : : "r" (pstate), "i" (PSTATE_IE));
 157         ptep = pte_offset_map(pmdp, tpc);
 158         pte = *ptep;
 159         if (!pte_present(pte))
 160                 goto out;
 161
 162         pa  = (pte_pfn(pte) << PAGE_SHIFT);
 163         pa += (tpc & ~PAGE_MASK);
 164
 165         /* Use phys bypass so we don't pollute dtlb/dcache. */
 166         __asm__ __volatile__("lduwa [%1] %2, %0"
 167                              : "=r" (insn)
 168                              : "r" (pa), "i" (ASI_PHYS_USE_EC));
 169
 170 out:
 171         pte_unmap(ptep);
 172         __asm__ __volatile__("wrpr %0, 0x0, %%pstate" : : "r" (pstate));
 173 outret:
 174         return insn;
 175 }
 176
 177 extern unsigned long compute_effective_address(struct pt_regs *, unsigned int, unsigned int);
 178
 179 static void do_fault_siginfo(int code, int sig, struct pt_regs *regs,
 180                              unsigned int insn, int fault_code)
 181 {
 182         siginfo_t info;
 183
 184         info.si_code = code;
 185         info.si_signo = sig;
 186         info.si_errno = 0;
 187         if (fault_code & FAULT_CODE_ITLB)
 188                 info.si_addr = (void __user *) regs->tpc;
 189         else
 190                 info.si_addr = (void __user *)
 191                         compute_effective_address(regs, insn, 0);
 192         info.si_trapno = 0;
 193         force_sig_info(sig, &info, current);
 194 }
 195
 196 extern int handle_ldf_stq(u32, struct pt_regs *);
 197 extern int handle_ld_nf(u32, struct pt_regs *);
 198
 199 static unsigned int get_fault_insn(struct pt_regs *regs, unsigned int insn)
 200 {
 201         if (!insn) {
 202                 if (!regs->tpc || (regs->tpc & 0x3))
 203                         return 0;
 204                 if (regs->tstate & TSTATE_PRIV) {
 205                         insn = *(unsigned int *) regs->tpc;
 206                 } else {
 207                         insn = get_user_insn(regs->tpc);
 208                 }
 209         }
 210         return insn;
 211 }
 212
 213 static void do_kernel_fault(struct pt_regs *regs, int si_code, int fault_code,
 214                             unsigned int insn, unsigned long address)
 215 {
 216         unsigned char asi = ASI_P;
 217
 218         if ((!insn) && (regs->tstate & TSTATE_PRIV))
 219                 goto cannot_handle;
 220
 221         /* If user insn could be read (thus insn is zero), that
 222          * is fine.  We will just gun down the process with a signal
 223          * in that case.
 224          */
 225
 226         if (!(fault_code & (FAULT_CODE_WRITE|FAULT_CODE_ITLB)) &&
 227             (insn & 0xc0800000) == 0xc0800000) {
 228                 if (insn & 0x2000)
 229                         asi = (regs->tstate >> 24);
 230                 else
 231                         asi = (insn >> 5);
 232                 if ((asi & 0xf2) == 0x82) {
 233                         if (insn & 0x1000000) {
 234                                 handle_ldf_stq(insn, regs);
 235                         } else {
 236                                 /* This was a non-faulting load. Just clear the
 237                                  * destination register(s) and continue with the next
 238                                  * instruction. -jj
 239                                  */
 240                                 handle_ld_nf(insn, regs);
 241                         }
 242                         return;
 243                 }
 244         }
 245
 246         /* Is this in ex_table? */
 247         if (regs->tstate & TSTATE_PRIV) {
 248                 const struct exception_table_entry *entry;
 249
 250                 if (asi == ASI_P && (insn & 0xc0800000) == 0xc0800000) {
 251                         if (insn & 0x2000)
 252                                 asi = (regs->tstate >> 24);
 253                         else
 254                                 asi = (insn >> 5);
 255                 }
 256
 257                 /* Look in asi.h: All _S asis have LS bit set */
 258                 if ((asi & 0x1) &&
 259                     (entry = search_exception_tables(regs->tpc))) {
 260                         regs->tpc = entry->fixup;
 261                         regs->tnpc = regs->tpc + 4;
 262                         return;
 263                 }
 264         } else {
 265                 /* The si_code was set to make clear whether
 266                  * this was a SEGV_MAPERR or SEGV_ACCERR fault.
 267                  */
 268                 do_fault_siginfo(si_code, SIGSEGV, regs, insn, fault_code);
 269                 return;
 270         }
 271
 272 cannot_handle:
 273         unhandled_fault (address, current, regs);
 274 }
 275
 276 asmlinkage void __kprobes do_sparc64_fault(struct pt_regs *regs)
 277 {
 278         struct mm_struct *mm = current->mm;
 279         struct vm_area_struct *vma;
 280         unsigned int insn = 0;
 281         int si_code, fault_code;
 282         unsigned long address, mm_rss;
 283
 284         fault_code = get_thread_fault_code();
 285
 286         if (notify_page_fault(regs))
 287                 return;
 288
 289         si_code = SEGV_MAPERR;
 290         address = current_thread_info()->fault_address;
 291
 292         if ((fault_code & FAULT_CODE_ITLB) &&
 293             (fault_code & FAULT_CODE_DTLB))
 294                 BUG();
 295
 296         if (regs->tstate & TSTATE_PRIV) {
 297                 unsigned long tpc = regs->tpc;
 298
 299                 /* Sanity check the PC. */
 300                 if ((tpc >= KERNBASE && tpc < (unsigned long) _etext) ||
 301                     (tpc >= MODULES_VADDR && tpc < MODULES_END)) {
 302                         /* Valid, no problems... */
 303                 } else {
 304                         bad_kernel_pc(regs, address);
 305                         return;
 306                 }
 307         }
 308
 309         /*
 310          * If we're in an interrupt or have no user
 311          * context, we must not take the fault..
 312          */
 313         if (in_atomic() || !mm)
 314                 goto intr_or_no_mm;
 315
 316         if (test_thread_flag(TIF_32BIT)) {
 317                 if (!(regs->tstate & TSTATE_PRIV))
 318                         regs->tpc &= 0xffffffff;
 319                 address &= 0xffffffff;
 320         }
 321
 322         if (!down_read_trylock(&mm->mmap_sem)) {
 323                 if ((regs->tstate & TSTATE_PRIV) &&
 324                     !search_exception_tables(regs->tpc)) {
 325                         insn = get_fault_insn(regs, insn);
 326                         goto handle_kernel_fault;
 327                 }
 328                 down_read(&mm->mmap_sem);
 329         }
 330
 331         vma = find_vma(mm, address);
 332         if (!vma)
 333                 goto bad_area;
 334
 335         /* Pure DTLB misses do not tell us whether the fault causing
 336          * load/store/atomic was a write or not, it only says that there
 337          * was no match.  So in such a case we (carefully) read the
 338          * instruction to try and figure this out.  It's an optimization
 339          * so it's ok if we can't do this.
 340          *
 341          * Special hack, window spill/fill knows the exact fault type.
 342          */
 343         if (((fault_code &
 344               (FAULT_CODE_DTLB | FAULT_CODE_WRITE | FAULT_CODE_WINFIXUP)) == FAULT_CODE_DTLB) &&
 345             (vma->vm_flags & VM_WRITE) != 0) {
 346                 insn = get_fault_insn(regs, 0);
 347                 if (!insn)
 348                         goto continue_fault;
 349                 /* All loads, stores and atomics have bits 30 and 31 both set
 350                  * in the instruction.  Bit 21 is set in all stores, but we
 351                  * have to avoid prefetches which also have bit 21 set.
 352                  */
 353                 if ((insn & 0xc0200000) == 0xc0200000 &&
 354                     (insn & 0x01780000) != 0x01680000) {
 355                         /* Don't bother updating thread struct value,
 356                          * because update_mmu_cache only cares which tlb
 357                          * the access came from.
 358                          */
 359                         fault_code |= FAULT_CODE_WRITE;
 360                 }
 361         }
 362 continue_fault:
 363
 364         if (vma->vm_start <= address)
 365                 goto good_area;
 366         if (!(vma->vm_flags & VM_GROWSDOWN))
 367                 goto bad_area;
 368         if (!(fault_code & FAULT_CODE_WRITE)) {
 369                 /* Non-faulting loads shouldn't expand stack. */
 370                 insn = get_fault_insn(regs, insn);
 371                 if ((insn & 0xc0800000) == 0xc0800000) {
 372                         unsigned char asi;
 373
 374                         if (insn & 0x2000)
 375                                 asi = (regs->tstate >> 24);
 376                         else
 377                                 asi = (insn >> 5);
 378                         if ((asi & 0xf2) == 0x82)
 379                                 goto bad_area;
 380                 }
 381         }
 382         if (expand_stack(vma, address))
 383                 goto bad_area;
 384         /*
 385          * Ok, we have a good vm_area for this memory access, so
 386          * we can handle it..
 387          */
 388 good_area:
 389         si_code = SEGV_ACCERR;
 390
 391         /* If we took a ITLB miss on a non-executable page, catch
 392          * that here.
 393          */
 394         if ((fault_code & FAULT_CODE_ITLB) && !(vma->vm_flags & VM_EXEC)) {
 395                 BUG_ON(address != regs->tpc);
 396                 BUG_ON(regs->tstate & TSTATE_PRIV);
 397                 goto bad_area;
 398         }
 399
 400         if (fault_code & FAULT_CODE_WRITE) {
 401                 if (!(vma->vm_flags & VM_WRITE))
 402                         goto bad_area;
 403
 404                 /* Spitfire has an icache which does not snoop
 405                  * processor stores.  Later processors do...
 406                  */
 407                 if (tlb_type == spitfire &&
 408                     (vma->vm_flags & VM_EXEC) != 0 &&
 409                     vma->vm_file != NULL)
 410                         set_thread_fault_code(fault_code |
 411                                               FAULT_CODE_BLKCOMMIT);
 412         } else {
 413                 /* Allow reads even for write-only mappings */
 414                 if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
 415                         goto bad_area;
 416         }
 417
 418         switch (handle_mm_fault(mm, vma, address, (fault_code & FAULT_CODE_WRITE))) {
 419         case VM_FAULT_MINOR:
 420                 current->min_flt++;
 421                 break;
 422         case VM_FAULT_MAJOR:
 423                 current->maj_flt++;
 424                 break;
 425         case VM_FAULT_SIGBUS:
 426                 goto do_sigbus;
 427         case VM_FAULT_OOM:
 428                 goto out_of_memory;
 429         default:
 430                 BUG();
 431         }
 432
 433         up_read(&mm->mmap_sem);
 434
 435         mm_rss = get_mm_rss(mm);
 436 #ifdef CONFIG_HUGETLB_PAGE
 437         mm_rss -= (mm->context.huge_pte_count * (HPAGE_SIZE / PAGE_SIZE));
 438 #endif
 439         if (unlikely(mm_rss >
 440                      mm->context.tsb_block[MM_TSB_BASE].tsb_rss_limit))
 441                 tsb_grow(mm, MM_TSB_BASE, mm_rss);
 442 #ifdef CONFIG_HUGETLB_PAGE
 443         mm_rss = mm->context.huge_pte_count;
 444         if (unlikely(mm_rss >
 445                      mm->context.tsb_block[MM_TSB_HUGE].tsb_rss_limit))
 446                 tsb_grow(mm, MM_TSB_HUGE, mm_rss);
 447 #endif
 448         return;
 449
 450         /*
 451          * Something tried to access memory that isn't in our memory map..
 452          * Fix it, but check if it's kernel or user first..
 453          */
 454 bad_area:
 455         insn = get_fault_insn(regs, insn);
 456         up_read(&mm->mmap_sem);
 457
 458 handle_kernel_fault:
 459         do_kernel_fault(regs, si_code, fault_code, insn, address);
 460         return;
 461
 462 /*
 463  * We ran out of memory, or some other thing happened to us that made
 464  * us unable to handle the page fault gracefully.
 465  */
 466 out_of_memory:
 467         insn = get_fault_insn(regs, insn);
 468         up_read(&mm->mmap_sem);
 469         printk("VM: killing process %s\n", current->comm);
 470         if (!(regs->tstate & TSTATE_PRIV))
 471                 do_exit(SIGKILL);
 472         goto handle_kernel_fault;
 473
 474 intr_or_no_mm:
 475         insn = get_fault_insn(regs, 0);
 476         goto handle_kernel_fault;
 477
 478 do_sigbus:
 479         insn = get_fault_insn(regs, insn);
 480         up_read(&mm->mmap_sem);
 481
 482         /*
 483          * Send a sigbus, regardless of whether we were in kernel
 484          * or user mode.
 485          */
 486         do_fault_siginfo(BUS_ADRERR, SIGBUS, regs, insn, fault_code);
 487
 488         /* Kernel mode? Handle exceptions or die */
 489         if (regs->tstate & TSTATE_PRIV)
 490                 goto handle_kernel_fault;
 491 }