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         printk(KERN_CRIT "OOPS: Bogus kernel PC [%016lx] in fault handler\n",
 116                regs->tpc);
 117         printk(KERN_CRIT "OOPS: RPC [%016lx]\n", regs->u_regs[15]);
 118         print_symbol("RPC: <%s>\n", regs->u_regs[15]);
 119         printk(KERN_CRIT "OOPS: Fault was to vaddr[%lx]\n", vaddr);
 120         dump_stack();
 121         unhandled_fault(regs->tpc, current, regs);
 122 }
 123
 124 /*
 125  * We now make sure that mmap_sem is held in all paths that call
 126  * this. Additionally, to prevent kswapd from ripping ptes from
 127  * under us, raise interrupts around the time that we look at the
 128  * pte, kswapd will have to wait to get his smp ipi response from
 129  * us. vmtruncate likewise. This saves us having to get pte lock.
 130  */
 131 static unsigned int get_user_insn(unsigned long tpc)
 132 {
 133         pgd_t *pgdp = pgd_offset(current->mm, tpc);
 134         pud_t *pudp;
 135         pmd_t *pmdp;
 136         pte_t *ptep, pte;
 137         unsigned long pa;
 138         u32 insn = 0;
 139         unsigned long pstate;
 140
 141         if (pgd_none(*pgdp))
 142                 goto outret;
 143         pudp = pud_offset(pgdp, tpc);
 144         if (pud_none(*pudp))
 145                 goto outret;
 146         pmdp = pmd_offset(pudp, tpc);
 147         if (pmd_none(*pmdp))
 148                 goto outret;
 149
 150         /* This disables preemption for us as well. */
 151         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 152         __asm__ __volatile__("wrpr %0, %1, %%pstate"
 153                                 : : "r" (pstate), "i" (PSTATE_IE));
 154         ptep = pte_offset_map(pmdp, tpc);
 155         pte = *ptep;
 156         if (!pte_present(pte))
 157                 goto out;
 158
 159         pa  = (pte_pfn(pte) << PAGE_SHIFT);
 160         pa += (tpc & ~PAGE_MASK);
 161
 162         /* Use phys bypass so we don't pollute dtlb/dcache. */
 163         __asm__ __volatile__("lduwa [%1] %2, %0"
 164                              : "=r" (insn)
 165                              : "r" (pa), "i" (ASI_PHYS_USE_EC));
 166
 167 out:
 168         pte_unmap(ptep);
 169         __asm__ __volatile__("wrpr %0, 0x0, %%pstate" : : "r" (pstate));
 170 outret:
 171         return insn;
 172 }
 173
 174 extern unsigned long compute_effective_address(struct pt_regs *, unsigned int, unsigned int);
 175
 176 static void do_fault_siginfo(int code, int sig, struct pt_regs *regs,
 177                              unsigned int insn, int fault_code)
 178 {
 179         siginfo_t info;
 180
 181         info.si_code = code;
 182         info.si_signo = sig;
 183         info.si_errno = 0;
 184         if (fault_code & FAULT_CODE_ITLB)
 185                 info.si_addr = (void __user *) regs->tpc;
 186         else
 187                 info.si_addr = (void __user *)
 188                         compute_effective_address(regs, insn, 0);
 189         info.si_trapno = 0;
 190         force_sig_info(sig, &info, current);
 191 }
 192
 193 extern int handle_ldf_stq(u32, struct pt_regs *);
 194 extern int handle_ld_nf(u32, struct pt_regs *);
 195
 196 static unsigned int get_fault_insn(struct pt_regs *regs, unsigned int insn)
 197 {
 198         if (!insn) {
 199                 if (!regs->tpc || (regs->tpc & 0x3))
 200                         return 0;
 201                 if (regs->tstate & TSTATE_PRIV) {
 202                         insn = *(unsigned int *) regs->tpc;
 203                 } else {
 204                         insn = get_user_insn(regs->tpc);
 205                 }
 206         }
 207         return insn;
 208 }
 209
 210 static void do_kernel_fault(struct pt_regs *regs, int si_code, int fault_code,
 211                             unsigned int insn, unsigned long address)
 212 {
 213         unsigned char asi = ASI_P;
 214
 215         if ((!insn) && (regs->tstate & TSTATE_PRIV))
 216                 goto cannot_handle;
 217
 218         /* If user insn could be read (thus insn is zero), that
 219          * is fine.  We will just gun down the process with a signal
 220          * in that case.
 221          */
 222
 223         if (!(fault_code & (FAULT_CODE_WRITE|FAULT_CODE_ITLB)) &&
 224             (insn & 0xc0800000) == 0xc0800000) {
 225                 if (insn & 0x2000)
 226                         asi = (regs->tstate >> 24);
 227                 else
 228                         asi = (insn >> 5);
 229                 if ((asi & 0xf2) == 0x82) {
 230                         if (insn & 0x1000000) {
 231                                 handle_ldf_stq(insn, regs);
 232                         } else {
 233                                 /* This was a non-faulting load. Just clear the
 234                                  * destination register(s) and continue with the next
 235                                  * instruction. -jj
 236                                  */
 237                                 handle_ld_nf(insn, regs);
 238                         }
 239                         return;
 240                 }
 241         }
 242
 243         /* Is this in ex_table? */
 244         if (regs->tstate & TSTATE_PRIV) {
 245                 const struct exception_table_entry *entry;
 246
 247                 entry = search_exception_tables(regs->tpc);
 248                 if (entry) {
 249                         regs->tpc = entry->fixup;
 250                         regs->tnpc = regs->tpc + 4;
 251                         return;
 252                 }
 253         } else {
 254                 /* The si_code was set to make clear whether
 255                  * this was a SEGV_MAPERR or SEGV_ACCERR fault.
 256                  */
 257                 do_fault_siginfo(si_code, SIGSEGV, regs, insn, fault_code);
 258                 return;
 259         }
 260
 261 cannot_handle:
 262         unhandled_fault (address, current, regs);
 263 }
 264
 265 asmlinkage void __kprobes do_sparc64_fault(struct pt_regs *regs)
 266 {
 267         struct mm_struct *mm = current->mm;
 268         struct vm_area_struct *vma;
 269         unsigned int insn = 0;
 270         int si_code, fault_code, fault;
 271         unsigned long address, mm_rss;
 272
 273         fault_code = get_thread_fault_code();
 274
 275         if (notify_page_fault(regs))
 276                 return;
 277
 278         si_code = SEGV_MAPERR;
 279         address = current_thread_info()->fault_address;
 280
 281         if ((fault_code & FAULT_CODE_ITLB) &&
 282             (fault_code & FAULT_CODE_DTLB))
 283                 BUG();
 284
 285         if (regs->tstate & TSTATE_PRIV) {
 286                 unsigned long tpc = regs->tpc;
 287
 288                 /* Sanity check the PC. */
 289                 if ((tpc >= KERNBASE && tpc < (unsigned long) __init_end) ||
 290                     (tpc >= MODULES_VADDR && tpc < MODULES_END)) {
 291                         /* Valid, no problems... */
 292                 } else {
 293                         bad_kernel_pc(regs, address);
 294                         return;
 295                 }
 296         }
 297
 298         /*
 299          * If we're in an interrupt or have no user
 300          * context, we must not take the fault..
 301          */
 302         if (in_atomic() || !mm)
 303                 goto intr_or_no_mm;
 304
 305         if (test_thread_flag(TIF_32BIT)) {
 306                 if (!(regs->tstate & TSTATE_PRIV))
 307                         regs->tpc &= 0xffffffff;
 308                 address &= 0xffffffff;
 309         }
 310
 311         if (!down_read_trylock(&mm->mmap_sem)) {
 312                 if ((regs->tstate & TSTATE_PRIV) &&
 313                     !search_exception_tables(regs->tpc)) {
 314                         insn = get_fault_insn(regs, insn);
 315                         goto handle_kernel_fault;
 316                 }
 317                 down_read(&mm->mmap_sem);
 318         }
 319
 320         vma = find_vma(mm, address);
 321         if (!vma)
 322                 goto bad_area;
 323
 324         /* Pure DTLB misses do not tell us whether the fault causing
 325          * load/store/atomic was a write or not, it only says that there
 326          * was no match.  So in such a case we (carefully) read the
 327          * instruction to try and figure this out.  It's an optimization
 328          * so it's ok if we can't do this.
 329          *
 330          * Special hack, window spill/fill knows the exact fault type.
 331          */
 332         if (((fault_code &
 333               (FAULT_CODE_DTLB | FAULT_CODE_WRITE | FAULT_CODE_WINFIXUP)) == FAULT_CODE_DTLB) &&
 334             (vma->vm_flags & VM_WRITE) != 0) {
 335                 insn = get_fault_insn(regs, 0);
 336                 if (!insn)
 337                         goto continue_fault;
 338                 /* All loads, stores and atomics have bits 30 and 31 both set
 339                  * in the instruction.  Bit 21 is set in all stores, but we
 340                  * have to avoid prefetches which also have bit 21 set.
 341                  */
 342                 if ((insn & 0xc0200000) == 0xc0200000 &&
 343                     (insn & 0x01780000) != 0x01680000) {
 344                         /* Don't bother updating thread struct value,
 345                          * because update_mmu_cache only cares which tlb
 346                          * the access came from.
 347                          */
 348                         fault_code |= FAULT_CODE_WRITE;
 349                 }
 350         }
 351 continue_fault:
 352
 353         if (vma->vm_start <= address)
 354                 goto good_area;
 355         if (!(vma->vm_flags & VM_GROWSDOWN))
 356                 goto bad_area;
 357         if (!(fault_code & FAULT_CODE_WRITE)) {
 358                 /* Non-faulting loads shouldn't expand stack. */
 359                 insn = get_fault_insn(regs, insn);
 360                 if ((insn & 0xc0800000) == 0xc0800000) {
 361                         unsigned char asi;
 362
 363                         if (insn & 0x2000)
 364                                 asi = (regs->tstate >> 24);
 365                         else
 366                                 asi = (insn >> 5);
 367                         if ((asi & 0xf2) == 0x82)
 368                                 goto bad_area;
 369                 }
 370         }
 371         if (expand_stack(vma, address))
 372                 goto bad_area;
 373         /*
 374          * Ok, we have a good vm_area for this memory access, so
 375          * we can handle it..
 376          */
 377 good_area:
 378         si_code = SEGV_ACCERR;
 379
 380         /* If we took a ITLB miss on a non-executable page, catch
 381          * that here.
 382          */
 383         if ((fault_code & FAULT_CODE_ITLB) && !(vma->vm_flags & VM_EXEC)) {
 384                 BUG_ON(address != regs->tpc);
 385                 BUG_ON(regs->tstate & TSTATE_PRIV);
 386                 goto bad_area;
 387         }
 388
 389         if (fault_code & FAULT_CODE_WRITE) {
 390                 if (!(vma->vm_flags & VM_WRITE))
 391                         goto bad_area;
 392
 393                 /* Spitfire has an icache which does not snoop
 394                  * processor stores.  Later processors do...
 395                  */
 396                 if (tlb_type == spitfire &&
 397                     (vma->vm_flags & VM_EXEC) != 0 &&
 398                     vma->vm_file != NULL)
 399                         set_thread_fault_code(fault_code |
 400                                               FAULT_CODE_BLKCOMMIT);
 401         } else {
 402                 /* Allow reads even for write-only mappings */
 403                 if (!(vma->vm_flags & (VM_READ | VM_EXEC)))
 404                         goto bad_area;
 405         }
 406
 407         fault = handle_mm_fault(mm, vma, address, (fault_code & FAULT_CODE_WRITE));
 408         if (unlikely(fault & VM_FAULT_ERROR)) {
 409                 if (fault & VM_FAULT_OOM)
 410                         goto out_of_memory;
 411                 else if (fault & VM_FAULT_SIGBUS)
 412                         goto do_sigbus;
 413                 BUG();
 414         }
 415         if (fault & VM_FAULT_MAJOR)
 416                 current->maj_flt++;
 417         else
 418                 current->min_flt++;
 419
 420         up_read(&mm->mmap_sem);
 421
 422         mm_rss = get_mm_rss(mm);
 423 #ifdef CONFIG_HUGETLB_PAGE
 424         mm_rss -= (mm->context.huge_pte_count * (HPAGE_SIZE / PAGE_SIZE));
 425 #endif
 426         if (unlikely(mm_rss >
 427                      mm->context.tsb_block[MM_TSB_BASE].tsb_rss_limit))
 428                 tsb_grow(mm, MM_TSB_BASE, mm_rss);
 429 #ifdef CONFIG_HUGETLB_PAGE
 430         mm_rss = mm->context.huge_pte_count;
 431         if (unlikely(mm_rss >
 432                      mm->context.tsb_block[MM_TSB_HUGE].tsb_rss_limit))
 433                 tsb_grow(mm, MM_TSB_HUGE, mm_rss);
 434 #endif
 435         return;
 436
 437         /*
 438          * Something tried to access memory that isn't in our memory map..
 439          * Fix it, but check if it's kernel or user first..
 440          */
 441 bad_area:
 442         insn = get_fault_insn(regs, insn);
 443         up_read(&mm->mmap_sem);
 444
 445 handle_kernel_fault:
 446         do_kernel_fault(regs, si_code, fault_code, insn, address);
 447         return;
 448
 449 /*
 450  * We ran out of memory, or some other thing happened to us that made
 451  * us unable to handle the page fault gracefully.
 452  */
 453 out_of_memory:
 454         insn = get_fault_insn(regs, insn);
 455         up_read(&mm->mmap_sem);
 456         printk("VM: killing process %s\n", current->comm);
 457         if (!(regs->tstate & TSTATE_PRIV))
 458                 do_group_exit(SIGKILL);
 459         goto handle_kernel_fault;
 460
 461 intr_or_no_mm:
 462         insn = get_fault_insn(regs, 0);
 463         goto handle_kernel_fault;
 464
 465 do_sigbus:
 466         insn = get_fault_insn(regs, insn);
 467         up_read(&mm->mmap_sem);
 468
 469         /*
 470          * Send a sigbus, regardless of whether we were in kernel
 471          * or user mode.
 472          */
 473         do_fault_siginfo(BUS_ADRERR, SIGBUS, regs, insn, fault_code);
 474
 475         /* Kernel mode? Handle exceptions or die */
 476         if (regs->tstate & TSTATE_PRIV)
 477                 goto handle_kernel_fault;
 478 }