arch/s390/mm/fault.c

   1 /*
   2  *  S390 version
   3  *    Copyright IBM Corp. 1999
   4  *    Author(s): Hartmut Penner (hp@de.ibm.com)
   5  *               Ulrich Weigand (uweigand@de.ibm.com)
   6  *
   7  *  Derived from "arch/i386/mm/fault.c"
   8  *    Copyright (C) 1995  Linus Torvalds
   9  */
  10
  11 #include <linux/kernel_stat.h>
  12 #include <linux/perf_event.h>
  13 #include <linux/signal.h>
  14 #include <linux/sched.h>
  15 #include <linux/kernel.h>
  16 #include <linux/errno.h>
  17 #include <linux/string.h>
  18 #include <linux/types.h>
  19 #include <linux/ptrace.h>
  20 #include <linux/mman.h>
  21 #include <linux/mm.h>
  22 #include <linux/compat.h>
  23 #include <linux/smp.h>
  24 #include <linux/kdebug.h>
  25 #include <linux/init.h>
  26 #include <linux/console.h>
  27 #include <linux/module.h>
  28 #include <linux/hardirq.h>
  29 #include <linux/kprobes.h>
  30 #include <linux/uaccess.h>
  31 #include <linux/hugetlb.h>
  32 #include <asm/asm-offsets.h>
  33 #include <asm/diag.h>
  34 #include <asm/pgtable.h>
  35 #include <asm/irq.h>
  36 #include <asm/mmu_context.h>
  37 #include <asm/facility.h>
  38 #include "../kernel/entry.h"
  39
  40 #define __FAIL_ADDR_MASK -4096L
  41 #define __SUBCODE_MASK 0x0600
  42 #define __PF_RES_FIELD 0x8000000000000000ULL
  43
  44 #define VM_FAULT_BADCONTEXT     0x010000
  45 #define VM_FAULT_BADMAP         0x020000
  46 #define VM_FAULT_BADACCESS      0x040000
  47 #define VM_FAULT_SIGNAL         0x080000
  48 #define VM_FAULT_PFAULT         0x100000
  49
  50 static unsigned long store_indication __read_mostly;
  51
  52 static int __init fault_init(void)
  53 {
  54         if (test_facility(75))
  55                 store_indication = 0xc00;
  56         return 0;
  57 }
  58 early_initcall(fault_init);
  59
  60 static inline int notify_page_fault(struct pt_regs *regs)
  61 {
  62         int ret = 0;
  63
  64         /* kprobe_running() needs smp_processor_id() */
  65         if (kprobes_built_in() && !user_mode(regs)) {
  66                 preempt_disable();
  67                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  68                         ret = 1;
  69                 preempt_enable();
  70         }
  71         return ret;
  72 }
  73
  74
  75 /*
  76  * Unlock any spinlocks which will prevent us from getting the
  77  * message out.
  78  */
  79 void bust_spinlocks(int yes)
  80 {
  81         if (yes) {
  82                 oops_in_progress = 1;
  83         } else {
  84                 int loglevel_save = console_loglevel;
  85                 console_unblank();
  86                 oops_in_progress = 0;
  87                 /*
  88                  * OK, the message is on the console.  Now we call printk()
  89                  * without oops_in_progress set so that printk will give klogd
  90                  * a poke.  Hold onto your hats...
  91                  */
  92                 console_loglevel = 15;
  93                 printk(" ");
  94                 console_loglevel = loglevel_save;
  95         }
  96 }
  97
  98 /*
  99  * Returns the address space associated with the fault.
 100  * Returns 0 for kernel space and 1 for user space.
 101  */
 102 static inline int user_space_fault(struct pt_regs *regs)
 103 {
 104         unsigned long trans_exc_code;
 105
 106         /*
 107          * The lowest two bits of the translation exception
 108          * identification indicate which paging table was used.
 109          */
 110         trans_exc_code = regs->int_parm_long & 3;
 111         if (trans_exc_code == 3) /* home space -> kernel */
 112                 return 0;
 113         if (user_mode(regs))
 114                 return 1;
 115         if (trans_exc_code == 2) /* secondary space -> set_fs */
 116                 return current->thread.mm_segment.ar4;
 117         if (current->flags & PF_VCPU)
 118                 return 1;
 119         return 0;
 120 }
 121
 122 static int bad_address(void *p)
 123 {
 124         unsigned long dummy;
 125
 126         return probe_kernel_address((unsigned long *)p, dummy);
 127 }
 128
 129 static void dump_pagetable(unsigned long asce, unsigned long address)
 130 {
 131         unsigned long *table = __va(asce & PAGE_MASK);
 132
 133         pr_alert("AS:%016lx ", asce);
 134         switch (asce & _ASCE_TYPE_MASK) {
 135         case _ASCE_TYPE_REGION1:
 136                 table = table + ((address >> 53) & 0x7ff);
 137                 if (bad_address(table))
 138                         goto bad;
 139                 pr_cont("R1:%016lx ", *table);
 140                 if (*table & _REGION_ENTRY_INVALID)
 141                         goto out;
 142                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 143                 /* fallthrough */
 144         case _ASCE_TYPE_REGION2:
 145                 table = table + ((address >> 42) & 0x7ff);
 146                 if (bad_address(table))
 147                         goto bad;
 148                 pr_cont("R2:%016lx ", *table);
 149                 if (*table & _REGION_ENTRY_INVALID)
 150                         goto out;
 151                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 152                 /* fallthrough */
 153         case _ASCE_TYPE_REGION3:
 154                 table = table + ((address >> 31) & 0x7ff);
 155                 if (bad_address(table))
 156                         goto bad;
 157                 pr_cont("R3:%016lx ", *table);
 158                 if (*table & (_REGION_ENTRY_INVALID | _REGION3_ENTRY_LARGE))
 159                         goto out;
 160                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 161                 /* fallthrough */
 162         case _ASCE_TYPE_SEGMENT:
 163                 table = table + ((address >> 20) & 0x7ff);
 164                 if (bad_address(table))
 165                         goto bad;
 166                 pr_cont("S:%016lx ", *table);
 167                 if (*table & (_SEGMENT_ENTRY_INVALID | _SEGMENT_ENTRY_LARGE))
 168                         goto out;
 169                 table = (unsigned long *)(*table & _SEGMENT_ENTRY_ORIGIN);
 170         }
 171         table = table + ((address >> 12) & 0xff);
 172         if (bad_address(table))
 173                 goto bad;
 174         pr_cont("P:%016lx ", *table);
 175 out:
 176         pr_cont("\n");
 177         return;
 178 bad:
 179         pr_cont("BAD\n");
 180 }
 181
 182 static void dump_fault_info(struct pt_regs *regs)
 183 {
 184         unsigned long asce;
 185
 186         pr_alert("Fault in ");
 187         switch (regs->int_parm_long & 3) {
 188         case 3:
 189                 pr_cont("home space ");
 190                 break;
 191         case 2:
 192                 pr_cont("secondary space ");
 193                 break;
 194         case 1:
 195                 pr_cont("access register ");
 196                 break;
 197         case 0:
 198                 pr_cont("primary space ");
 199                 break;
 200         }
 201         pr_cont("mode while using ");
 202         if (!user_space_fault(regs)) {
 203                 asce = S390_lowcore.kernel_asce;
 204                 pr_cont("kernel ");
 205         }
 206 #ifdef CONFIG_PGSTE
 207         else if ((current->flags & PF_VCPU) && S390_lowcore.gmap) {
 208                 struct gmap *gmap = (struct gmap *)S390_lowcore.gmap;
 209                 asce = gmap->asce;
 210                 pr_cont("gmap ");
 211         }
 212 #endif
 213         else {
 214                 asce = S390_lowcore.user_asce;
 215                 pr_cont("user ");
 216         }
 217         pr_cont("ASCE.\n");
 218         dump_pagetable(asce, regs->int_parm_long & __FAIL_ADDR_MASK);
 219 }
 220
 221 static inline void report_user_fault(struct pt_regs *regs, long signr)
 222 {
 223         if ((task_pid_nr(current) > 1) && !show_unhandled_signals)
 224                 return;
 225         if (!unhandled_signal(current, signr))
 226                 return;
 227         if (!printk_ratelimit())
 228                 return;
 229         printk(KERN_ALERT "User process fault: interruption code %04x ilc:%d ",
 230                regs->int_code & 0xffff, regs->int_code >> 17);
 231         print_vma_addr(KERN_CONT "in ", regs->psw.addr);
 232         printk(KERN_CONT "\n");
 233         printk(KERN_ALERT "failing address: %016lx TEID: %016lx\n",
 234                regs->int_parm_long & __FAIL_ADDR_MASK, regs->int_parm_long);
 235         dump_fault_info(regs);
 236         show_regs(regs);
 237 }
 238
 239 /*
 240  * Send SIGSEGV to task.  This is an external routine
 241  * to keep the stack usage of do_page_fault small.
 242  */
 243 static noinline void do_sigsegv(struct pt_regs *regs, int si_code)
 244 {
 245         struct siginfo si;
 246
 247         report_user_fault(regs, SIGSEGV);
 248         si.si_signo = SIGSEGV;
 249         si.si_code = si_code;
 250         si.si_addr = (void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK);
 251         force_sig_info(SIGSEGV, &si, current);
 252 }
 253
 254 static noinline void do_no_context(struct pt_regs *regs)
 255 {
 256         const struct exception_table_entry *fixup;
 257
 258         /* Are we prepared to handle this kernel fault?  */
 259         fixup = search_exception_tables(regs->psw.addr);
 260         if (fixup) {
 261                 regs->psw.addr = extable_fixup(fixup);
 262                 return;
 263         }
 264
 265         /*
 266          * Oops. The kernel tried to access some bad page. We'll have to
 267          * terminate things with extreme prejudice.
 268          */
 269         if (!user_space_fault(regs))
 270                 printk(KERN_ALERT "Unable to handle kernel pointer dereference"
 271                        " in virtual kernel address space\n");
 272         else
 273                 printk(KERN_ALERT "Unable to handle kernel paging request"
 274                        " in virtual user address space\n");
 275         printk(KERN_ALERT "failing address: %016lx TEID: %016lx\n",
 276                regs->int_parm_long & __FAIL_ADDR_MASK, regs->int_parm_long);
 277         dump_fault_info(regs);
 278         die(regs, "Oops");
 279         do_exit(SIGKILL);
 280 }
 281
 282 static noinline void do_low_address(struct pt_regs *regs)
 283 {
 284         /* Low-address protection hit in kernel mode means
 285            NULL pointer write access in kernel mode.  */
 286         if (regs->psw.mask & PSW_MASK_PSTATE) {
 287                 /* Low-address protection hit in user mode 'cannot happen'. */
 288                 die (regs, "Low-address protection");
 289                 do_exit(SIGKILL);
 290         }
 291
 292         do_no_context(regs);
 293 }
 294
 295 static noinline void do_sigbus(struct pt_regs *regs)
 296 {
 297         struct task_struct *tsk = current;
 298         struct siginfo si;
 299
 300         /*
 301          * Send a sigbus, regardless of whether we were in kernel
 302          * or user mode.
 303          */
 304         si.si_signo = SIGBUS;
 305         si.si_errno = 0;
 306         si.si_code = BUS_ADRERR;
 307         si.si_addr = (void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK);
 308         force_sig_info(SIGBUS, &si, tsk);
 309 }
 310
 311 static noinline void do_fault_error(struct pt_regs *regs, int fault)
 312 {
 313         int si_code;
 314
 315         switch (fault) {
 316         case VM_FAULT_BADACCESS:
 317         case VM_FAULT_BADMAP:
 318                 /* Bad memory access. Check if it is kernel or user space. */
 319                 if (user_mode(regs)) {
 320                         /* User mode accesses just cause a SIGSEGV */
 321                         si_code = (fault == VM_FAULT_BADMAP) ?
 322                                 SEGV_MAPERR : SEGV_ACCERR;
 323                         do_sigsegv(regs, si_code);
 324                         return;
 325                 }
 326         case VM_FAULT_BADCONTEXT:
 327         case VM_FAULT_PFAULT:
 328                 do_no_context(regs);
 329                 break;
 330         case VM_FAULT_SIGNAL:
 331                 if (!user_mode(regs))
 332                         do_no_context(regs);
 333                 break;
 334         default: /* fault & VM_FAULT_ERROR */
 335                 if (fault & VM_FAULT_OOM) {
 336                         if (!user_mode(regs))
 337                                 do_no_context(regs);
 338                         else
 339                                 pagefault_out_of_memory();
 340                 } else if (fault & VM_FAULT_SIGSEGV) {
 341                         /* Kernel mode? Handle exceptions or die */
 342                         if (!user_mode(regs))
 343                                 do_no_context(regs);
 344                         else
 345                                 do_sigsegv(regs, SEGV_MAPERR);
 346                 } else if (fault & VM_FAULT_SIGBUS) {
 347                         /* Kernel mode? Handle exceptions or die */
 348                         if (!user_mode(regs))
 349                                 do_no_context(regs);
 350                         else
 351                                 do_sigbus(regs);
 352                 } else
 353                         BUG();
 354                 break;
 355         }
 356 }
 357
 358 /*
 359  * This routine handles page faults.  It determines the address,
 360  * and the problem, and then passes it off to one of the appropriate
 361  * routines.
 362  *
 363  * interruption code (int_code):
 364  *   04       Protection           ->  Write-Protection  (suprression)
 365  *   10       Segment translation  ->  Not present       (nullification)
 366  *   11       Page translation     ->  Not present       (nullification)
 367  *   3b       Region third trans.  ->  Not present       (nullification)
 368  */
 369 static inline int do_exception(struct pt_regs *regs, int access)
 370 {
 371 #ifdef CONFIG_PGSTE
 372         struct gmap *gmap;
 373 #endif
 374         struct task_struct *tsk;
 375         struct mm_struct *mm;
 376         struct vm_area_struct *vma;
 377         unsigned long trans_exc_code;
 378         unsigned long address;
 379         unsigned int flags;
 380         int fault;
 381
 382         tsk = current;
 383         /*
 384          * The instruction that caused the program check has
 385          * been nullified. Don't signal single step via SIGTRAP.
 386          */
 387         clear_pt_regs_flag(regs, PIF_PER_TRAP);
 388
 389         if (notify_page_fault(regs))
 390                 return 0;
 391
 392         mm = tsk->mm;
 393         trans_exc_code = regs->int_parm_long;
 394
 395         /*
 396          * Verify that the fault happened in user space, that
 397          * we are not in an interrupt and that there is a
 398          * user context.
 399          */
 400         fault = VM_FAULT_BADCONTEXT;
 401         if (unlikely(!user_space_fault(regs) || faulthandler_disabled() || !mm))
 402                 goto out;
 403
 404         address = trans_exc_code & __FAIL_ADDR_MASK;
 405         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
 406         flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
 407         if (user_mode(regs))
 408                 flags |= FAULT_FLAG_USER;
 409         if (access == VM_WRITE || (trans_exc_code & store_indication) == 0x400)
 410                 flags |= FAULT_FLAG_WRITE;
 411         down_read(&mm->mmap_sem);
 412
 413 #ifdef CONFIG_PGSTE
 414         gmap = (current->flags & PF_VCPU) ?
 415                 (struct gmap *) S390_lowcore.gmap : NULL;
 416         if (gmap) {
 417                 current->thread.gmap_addr = address;
 418                 address = __gmap_translate(gmap, address);
 419                 if (address == -EFAULT) {
 420                         fault = VM_FAULT_BADMAP;
 421                         goto out_up;
 422                 }
 423                 if (gmap->pfault_enabled)
 424                         flags |= FAULT_FLAG_RETRY_NOWAIT;
 425         }
 426 #endif
 427
 428 retry:
 429         fault = VM_FAULT_BADMAP;
 430         vma = find_vma(mm, address);
 431         if (!vma)
 432                 goto out_up;
 433
 434         if (unlikely(vma->vm_start > address)) {
 435                 if (!(vma->vm_flags & VM_GROWSDOWN))
 436                         goto out_up;
 437                 if (expand_stack(vma, address))
 438                         goto out_up;
 439         }
 440
 441         /*
 442          * Ok, we have a good vm_area for this memory access, so
 443          * we can handle it..
 444          */
 445         fault = VM_FAULT_BADACCESS;
 446         if (unlikely(!(vma->vm_flags & access)))
 447                 goto out_up;
 448
 449         if (is_vm_hugetlb_page(vma))
 450                 address &= HPAGE_MASK;
 451         /*
 452          * If for any reason at all we couldn't handle the fault,
 453          * make sure we exit gracefully rather than endlessly redo
 454          * the fault.
 455          */
 456         fault = handle_mm_fault(mm, vma, address, flags);
 457         /* No reason to continue if interrupted by SIGKILL. */
 458         if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)) {
 459                 fault = VM_FAULT_SIGNAL;
 460                 goto out;
 461         }
 462         if (unlikely(fault & VM_FAULT_ERROR))
 463                 goto out_up;
 464
 465         /*
 466          * Major/minor page fault accounting is only done on the
 467          * initial attempt. If we go through a retry, it is extremely
 468          * likely that the page will be found in page cache at that point.
 469          */
 470         if (flags & FAULT_FLAG_ALLOW_RETRY) {
 471                 if (fault & VM_FAULT_MAJOR) {
 472                         tsk->maj_flt++;
 473                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
 474                                       regs, address);
 475                 } else {
 476                         tsk->min_flt++;
 477                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
 478                                       regs, address);
 479                 }
 480                 if (fault & VM_FAULT_RETRY) {
 481 #ifdef CONFIG_PGSTE
 482                         if (gmap && (flags & FAULT_FLAG_RETRY_NOWAIT)) {
 483                                 /* FAULT_FLAG_RETRY_NOWAIT has been set,
 484                                  * mmap_sem has not been released */
 485                                 current->thread.gmap_pfault = 1;
 486                                 fault = VM_FAULT_PFAULT;
 487                                 goto out_up;
 488                         }
 489 #endif
 490                         /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
 491                          * of starvation. */
 492                         flags &= ~(FAULT_FLAG_ALLOW_RETRY |
 493                                    FAULT_FLAG_RETRY_NOWAIT);
 494                         flags |= FAULT_FLAG_TRIED;
 495                         down_read(&mm->mmap_sem);
 496                         goto retry;
 497                 }
 498         }
 499 #ifdef CONFIG_PGSTE
 500         if (gmap) {
 501                 address =  __gmap_link(gmap, current->thread.gmap_addr,
 502                                        address);
 503                 if (address == -EFAULT) {
 504                         fault = VM_FAULT_BADMAP;
 505                         goto out_up;
 506                 }
 507                 if (address == -ENOMEM) {
 508                         fault = VM_FAULT_OOM;
 509                         goto out_up;
 510                 }
 511         }
 512 #endif
 513         fault = 0;
 514 out_up:
 515         up_read(&mm->mmap_sem);
 516 out:
 517         return fault;
 518 }
 519
 520 void do_protection_exception(struct pt_regs *regs)
 521 {
 522         unsigned long trans_exc_code;
 523         int fault;
 524
 525         trans_exc_code = regs->int_parm_long;
 526         /*
 527          * Protection exceptions are suppressing, decrement psw address.
 528          * The exception to this rule are aborted transactions, for these
 529          * the PSW already points to the correct location.
 530          */
 531         if (!(regs->int_code & 0x200))
 532                 regs->psw.addr = __rewind_psw(regs->psw, regs->int_code >> 16);
 533         /*
 534          * Check for low-address protection.  This needs to be treated
 535          * as a special case because the translation exception code
 536          * field is not guaranteed to contain valid data in this case.
 537          */
 538         if (unlikely(!(trans_exc_code & 4))) {
 539                 do_low_address(regs);
 540                 return;
 541         }
 542         fault = do_exception(regs, VM_WRITE);
 543         if (unlikely(fault))
 544                 do_fault_error(regs, fault);
 545 }
 546 NOKPROBE_SYMBOL(do_protection_exception);
 547
 548 void do_dat_exception(struct pt_regs *regs)
 549 {
 550         int access, fault;
 551
 552         access = VM_READ | VM_EXEC | VM_WRITE;
 553         fault = do_exception(regs, access);
 554         if (unlikely(fault))
 555                 do_fault_error(regs, fault);
 556 }
 557 NOKPROBE_SYMBOL(do_dat_exception);
 558
 559 #ifdef CONFIG_PFAULT
 560 /*
 561  * 'pfault' pseudo page faults routines.
 562  */
 563 static int pfault_disable;
 564
 565 static int __init nopfault(char *str)
 566 {
 567         pfault_disable = 1;
 568         return 1;
 569 }
 570
 571 __setup("nopfault", nopfault);
 572
 573 struct pfault_refbk {
 574         u16 refdiagc;
 575         u16 reffcode;
 576         u16 refdwlen;
 577         u16 refversn;
 578         u64 refgaddr;
 579         u64 refselmk;
 580         u64 refcmpmk;
 581         u64 reserved;
 582 } __attribute__ ((packed, aligned(8)));
 583
 584 int pfault_init(void)
 585 {
 586         struct pfault_refbk refbk = {
 587                 .refdiagc = 0x258,
 588                 .reffcode = 0,
 589                 .refdwlen = 5,
 590                 .refversn = 2,
 591                 .refgaddr = __LC_LPP,
 592                 .refselmk = 1ULL << 48,
 593                 .refcmpmk = 1ULL << 48,
 594                 .reserved = __PF_RES_FIELD };
 595         int rc;
 596
 597         if (pfault_disable)
 598                 return -1;
 599         diag_stat_inc(DIAG_STAT_X258);
 600         asm volatile(
 601                 "       diag    %1,%0,0x258\n"
 602                 "0:     j       2f\n"
 603                 "1:     la      %0,8\n"
 604                 "2:\n"
 605                 EX_TABLE(0b,1b)
 606                 : "=d" (rc) : "a" (&refbk), "m" (refbk) : "cc");
 607         return rc;
 608 }
 609
 610 void pfault_fini(void)
 611 {
 612         struct pfault_refbk refbk = {
 613                 .refdiagc = 0x258,
 614                 .reffcode = 1,
 615                 .refdwlen = 5,
 616                 .refversn = 2,
 617         };
 618
 619         if (pfault_disable)
 620                 return;
 621         diag_stat_inc(DIAG_STAT_X258);
 622         asm volatile(
 623                 "       diag    %0,0,0x258\n"
 624                 "0:\n"
 625                 EX_TABLE(0b,0b)
 626                 : : "a" (&refbk), "m" (refbk) : "cc");
 627 }
 628
 629 static DEFINE_SPINLOCK(pfault_lock);
 630 static LIST_HEAD(pfault_list);
 631
 632 static void pfault_interrupt(struct ext_code ext_code,
 633                              unsigned int param32, unsigned long param64)
 634 {
 635         struct task_struct *tsk;
 636         __u16 subcode;
 637         pid_t pid;
 638
 639         /*
 640          * Get the external interruption subcode & pfault
 641          * initial/completion signal bit. VM stores this
 642          * in the 'cpu address' field associated with the
 643          * external interrupt.
 644          */
 645         subcode = ext_code.subcode;
 646         if ((subcode & 0xff00) != __SUBCODE_MASK)
 647                 return;
 648         inc_irq_stat(IRQEXT_PFL);
 649         /* Get the token (= pid of the affected task). */
 650         pid = param64 & LPP_PFAULT_PID_MASK;
 651         rcu_read_lock();
 652         tsk = find_task_by_pid_ns(pid, &init_pid_ns);
 653         if (tsk)
 654                 get_task_struct(tsk);
 655         rcu_read_unlock();
 656         if (!tsk)
 657                 return;
 658         spin_lock(&pfault_lock);
 659         if (subcode & 0x0080) {
 660                 /* signal bit is set -> a page has been swapped in by VM */
 661                 if (tsk->thread.pfault_wait == 1) {
 662                         /* Initial interrupt was faster than the completion
 663                          * interrupt. pfault_wait is valid. Set pfault_wait
 664                          * back to zero and wake up the process. This can
 665                          * safely be done because the task is still sleeping
 666                          * and can't produce new pfaults. */
 667                         tsk->thread.pfault_wait = 0;
 668                         list_del(&tsk->thread.list);
 669                         wake_up_process(tsk);
 670                         put_task_struct(tsk);
 671                 } else {
 672                         /* Completion interrupt was faster than initial
 673                          * interrupt. Set pfault_wait to -1 so the initial
 674                          * interrupt doesn't put the task to sleep.
 675                          * If the task is not running, ignore the completion
 676                          * interrupt since it must be a leftover of a PFAULT
 677                          * CANCEL operation which didn't remove all pending
 678                          * completion interrupts. */
 679                         if (tsk->state == TASK_RUNNING)
 680                                 tsk->thread.pfault_wait = -1;
 681                 }
 682         } else {
 683                 /* signal bit not set -> a real page is missing. */
 684                 if (WARN_ON_ONCE(tsk != current))
 685                         goto out;
 686                 if (tsk->thread.pfault_wait == 1) {
 687                         /* Already on the list with a reference: put to sleep */
 688                         __set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 689                         set_tsk_need_resched(tsk);
 690                 } else if (tsk->thread.pfault_wait == -1) {
 691                         /* Completion interrupt was faster than the initial
 692                          * interrupt (pfault_wait == -1). Set pfault_wait
 693                          * back to zero and exit. */
 694                         tsk->thread.pfault_wait = 0;
 695                 } else {
 696                         /* Initial interrupt arrived before completion
 697                          * interrupt. Let the task sleep.
 698                          * An extra task reference is needed since a different
 699                          * cpu may set the task state to TASK_RUNNING again
 700                          * before the scheduler is reached. */
 701                         get_task_struct(tsk);
 702                         tsk->thread.pfault_wait = 1;
 703                         list_add(&tsk->thread.list, &pfault_list);
 704                         __set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 705                         set_tsk_need_resched(tsk);
 706                 }
 707         }
 708 out:
 709         spin_unlock(&pfault_lock);
 710         put_task_struct(tsk);
 711 }
 712
 713 static int pfault_cpu_notify(struct notifier_block *self, unsigned long action,
 714                              void *hcpu)
 715 {
 716         struct thread_struct *thread, *next;
 717         struct task_struct *tsk;
 718
 719         switch (action & ~CPU_TASKS_FROZEN) {
 720         case CPU_DEAD:
 721                 spin_lock_irq(&pfault_lock);
 722                 list_for_each_entry_safe(thread, next, &pfault_list, list) {
 723                         thread->pfault_wait = 0;
 724                         list_del(&thread->list);
 725                         tsk = container_of(thread, struct task_struct, thread);
 726                         wake_up_process(tsk);
 727                         put_task_struct(tsk);
 728                 }
 729                 spin_unlock_irq(&pfault_lock);
 730                 break;
 731         default:
 732                 break;
 733         }
 734         return NOTIFY_OK;
 735 }
 736
 737 static int __init pfault_irq_init(void)
 738 {
 739         int rc;
 740
 741         rc = register_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
 742         if (rc)
 743                 goto out_extint;
 744         rc = pfault_init() == 0 ? 0 : -EOPNOTSUPP;
 745         if (rc)
 746                 goto out_pfault;
 747         irq_subclass_register(IRQ_SUBCLASS_SERVICE_SIGNAL);
 748         hotcpu_notifier(pfault_cpu_notify, 0);
 749         return 0;
 750
 751 out_pfault:
 752         unregister_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
 753 out_extint:
 754         pfault_disable = 1;
 755         return rc;
 756 }
 757 early_initcall(pfault_irq_init);
 758
 759 #endif /* CONFIG_PFAULT */