kernel/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 & PSW_ADDR_INSN);
 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         unsigned long address;
 258
 259         /* Are we prepared to handle this kernel fault?  */
 260         fixup = search_exception_tables(regs->psw.addr & PSW_ADDR_INSN);
 261         if (fixup) {
 262                 regs->psw.addr = extable_fixup(fixup) | PSW_ADDR_AMODE;
 263                 return;
 264         }
 265
 266         /*
 267          * Oops. The kernel tried to access some bad page. We'll have to
 268          * terminate things with extreme prejudice.
 269          */
 270         address = regs->int_parm_long & __FAIL_ADDR_MASK;
 271         if (!user_space_fault(regs))
 272                 printk(KERN_ALERT "Unable to handle kernel pointer dereference"
 273                        " in virtual kernel address space\n");
 274         else
 275                 printk(KERN_ALERT "Unable to handle kernel paging request"
 276                        " in virtual user address space\n");
 277         printk(KERN_ALERT "failing address: %016lx TEID: %016lx\n",
 278                regs->int_parm_long & __FAIL_ADDR_MASK, regs->int_parm_long);
 279         dump_fault_info(regs);
 280         die(regs, "Oops");
 281         do_exit(SIGKILL);
 282 }
 283
 284 static noinline void do_low_address(struct pt_regs *regs)
 285 {
 286         /* Low-address protection hit in kernel mode means
 287            NULL pointer write access in kernel mode.  */
 288         if (regs->psw.mask & PSW_MASK_PSTATE) {
 289                 /* Low-address protection hit in user mode 'cannot happen'. */
 290                 die (regs, "Low-address protection");
 291                 do_exit(SIGKILL);
 292         }
 293
 294         do_no_context(regs);
 295 }
 296
 297 static noinline void do_sigbus(struct pt_regs *regs)
 298 {
 299         struct task_struct *tsk = current;
 300         struct siginfo si;
 301
 302         /*
 303          * Send a sigbus, regardless of whether we were in kernel
 304          * or user mode.
 305          */
 306         si.si_signo = SIGBUS;
 307         si.si_errno = 0;
 308         si.si_code = BUS_ADRERR;
 309         si.si_addr = (void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK);
 310         force_sig_info(SIGBUS, &si, tsk);
 311 }
 312
 313 static noinline void do_fault_error(struct pt_regs *regs, int fault)
 314 {
 315         int si_code;
 316
 317         switch (fault) {
 318         case VM_FAULT_BADACCESS:
 319         case VM_FAULT_BADMAP:
 320                 /* Bad memory access. Check if it is kernel or user space. */
 321                 if (user_mode(regs)) {
 322                         /* User mode accesses just cause a SIGSEGV */
 323                         si_code = (fault == VM_FAULT_BADMAP) ?
 324                                 SEGV_MAPERR : SEGV_ACCERR;
 325                         do_sigsegv(regs, si_code);
 326                         return;
 327                 }
 328         case VM_FAULT_BADCONTEXT:
 329         case VM_FAULT_PFAULT:
 330                 do_no_context(regs);
 331                 break;
 332         case VM_FAULT_SIGNAL:
 333                 if (!user_mode(regs))
 334                         do_no_context(regs);
 335                 break;
 336         default: /* fault & VM_FAULT_ERROR */
 337                 if (fault & VM_FAULT_OOM) {
 338                         if (!user_mode(regs))
 339                                 do_no_context(regs);
 340                         else
 341                                 pagefault_out_of_memory();
 342                 } else if (fault & VM_FAULT_SIGSEGV) {
 343                         /* Kernel mode? Handle exceptions or die */
 344                         if (!user_mode(regs))
 345                                 do_no_context(regs);
 346                         else
 347                                 do_sigsegv(regs, SEGV_MAPERR);
 348                 } else if (fault & VM_FAULT_SIGBUS) {
 349                         /* Kernel mode? Handle exceptions or die */
 350                         if (!user_mode(regs))
 351                                 do_no_context(regs);
 352                         else
 353                                 do_sigbus(regs);
 354                 } else
 355                         BUG();
 356                 break;
 357         }
 358 }
 359
 360 /*
 361  * This routine handles page faults.  It determines the address,
 362  * and the problem, and then passes it off to one of the appropriate
 363  * routines.
 364  *
 365  * interruption code (int_code):
 366  *   04       Protection           ->  Write-Protection  (suprression)
 367  *   10       Segment translation  ->  Not present       (nullification)
 368  *   11       Page translation     ->  Not present       (nullification)
 369  *   3b       Region third trans.  ->  Not present       (nullification)
 370  */
 371 static inline int do_exception(struct pt_regs *regs, int access)
 372 {
 373 #ifdef CONFIG_PGSTE
 374         struct gmap *gmap;
 375 #endif
 376         struct task_struct *tsk;
 377         struct mm_struct *mm;
 378         struct vm_area_struct *vma;
 379         unsigned long trans_exc_code;
 380         unsigned long address;
 381         unsigned int flags;
 382         int fault;
 383
 384         tsk = current;
 385         /*
 386          * The instruction that caused the program check has
 387          * been nullified. Don't signal single step via SIGTRAP.
 388          */
 389         clear_pt_regs_flag(regs, PIF_PER_TRAP);
 390
 391         if (notify_page_fault(regs))
 392                 return 0;
 393
 394         mm = tsk->mm;
 395         trans_exc_code = regs->int_parm_long;
 396
 397         /*
 398          * Verify that the fault happened in user space, that
 399          * we are not in an interrupt and that there is a
 400          * user context.
 401          */
 402         fault = VM_FAULT_BADCONTEXT;
 403         if (unlikely(!user_space_fault(regs) || faulthandler_disabled() || !mm))
 404                 goto out;
 405
 406         address = trans_exc_code & __FAIL_ADDR_MASK;
 407         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
 408         flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
 409         if (user_mode(regs))
 410                 flags |= FAULT_FLAG_USER;
 411         if (access == VM_WRITE || (trans_exc_code & store_indication) == 0x400)
 412                 flags |= FAULT_FLAG_WRITE;
 413         down_read(&mm->mmap_sem);
 414
 415 #ifdef CONFIG_PGSTE
 416         gmap = (current->flags & PF_VCPU) ?
 417                 (struct gmap *) S390_lowcore.gmap : NULL;
 418         if (gmap) {
 419                 current->thread.gmap_addr = address;
 420                 address = __gmap_translate(gmap, address);
 421                 if (address == -EFAULT) {
 422                         fault = VM_FAULT_BADMAP;
 423                         goto out_up;
 424                 }
 425                 if (gmap->pfault_enabled)
 426                         flags |= FAULT_FLAG_RETRY_NOWAIT;
 427         }
 428 #endif
 429
 430 retry:
 431         fault = VM_FAULT_BADMAP;
 432         vma = find_vma(mm, address);
 433         if (!vma)
 434                 goto out_up;
 435
 436         if (unlikely(vma->vm_start > address)) {
 437                 if (!(vma->vm_flags & VM_GROWSDOWN))
 438                         goto out_up;
 439                 if (expand_stack(vma, address))
 440                         goto out_up;
 441         }
 442
 443         /*
 444          * Ok, we have a good vm_area for this memory access, so
 445          * we can handle it..
 446          */
 447         fault = VM_FAULT_BADACCESS;
 448         if (unlikely(!(vma->vm_flags & access)))
 449                 goto out_up;
 450
 451         if (is_vm_hugetlb_page(vma))
 452                 address &= HPAGE_MASK;
 453         /*
 454          * If for any reason at all we couldn't handle the fault,
 455          * make sure we exit gracefully rather than endlessly redo
 456          * the fault.
 457          */
 458         fault = handle_mm_fault(mm, vma, address, flags);
 459         /* No reason to continue if interrupted by SIGKILL. */
 460         if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)) {
 461                 fault = VM_FAULT_SIGNAL;
 462                 goto out;
 463         }
 464         if (unlikely(fault & VM_FAULT_ERROR))
 465                 goto out_up;
 466
 467         /*
 468          * Major/minor page fault accounting is only done on the
 469          * initial attempt. If we go through a retry, it is extremely
 470          * likely that the page will be found in page cache at that point.
 471          */
 472         if (flags & FAULT_FLAG_ALLOW_RETRY) {
 473                 if (fault & VM_FAULT_MAJOR) {
 474                         tsk->maj_flt++;
 475                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
 476                                       regs, address);
 477                 } else {
 478                         tsk->min_flt++;
 479                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
 480                                       regs, address);
 481                 }
 482                 if (fault & VM_FAULT_RETRY) {
 483 #ifdef CONFIG_PGSTE
 484                         if (gmap && (flags & FAULT_FLAG_RETRY_NOWAIT)) {
 485                                 /* FAULT_FLAG_RETRY_NOWAIT has been set,
 486                                  * mmap_sem has not been released */
 487                                 current->thread.gmap_pfault = 1;
 488                                 fault = VM_FAULT_PFAULT;
 489                                 goto out_up;
 490                         }
 491 #endif
 492                         /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
 493                          * of starvation. */
 494                         flags &= ~(FAULT_FLAG_ALLOW_RETRY |
 495                                    FAULT_FLAG_RETRY_NOWAIT);
 496                         flags |= FAULT_FLAG_TRIED;
 497                         down_read(&mm->mmap_sem);
 498                         goto retry;
 499                 }
 500         }
 501 #ifdef CONFIG_PGSTE
 502         if (gmap) {
 503                 address =  __gmap_link(gmap, current->thread.gmap_addr,
 504                                        address);
 505                 if (address == -EFAULT) {
 506                         fault = VM_FAULT_BADMAP;
 507                         goto out_up;
 508                 }
 509                 if (address == -ENOMEM) {
 510                         fault = VM_FAULT_OOM;
 511                         goto out_up;
 512                 }
 513         }
 514 #endif
 515         fault = 0;
 516 out_up:
 517         up_read(&mm->mmap_sem);
 518 out:
 519         return fault;
 520 }
 521
 522 void do_protection_exception(struct pt_regs *regs)
 523 {
 524         unsigned long trans_exc_code;
 525         int fault;
 526
 527         trans_exc_code = regs->int_parm_long;
 528         /*
 529          * Protection exceptions are suppressing, decrement psw address.
 530          * The exception to this rule are aborted transactions, for these
 531          * the PSW already points to the correct location.
 532          */
 533         if (!(regs->int_code & 0x200))
 534                 regs->psw.addr = __rewind_psw(regs->psw, regs->int_code >> 16);
 535         /*
 536          * Check for low-address protection.  This needs to be treated
 537          * as a special case because the translation exception code
 538          * field is not guaranteed to contain valid data in this case.
 539          */
 540         if (unlikely(!(trans_exc_code & 4))) {
 541                 do_low_address(regs);
 542                 return;
 543         }
 544         fault = do_exception(regs, VM_WRITE);
 545         if (unlikely(fault))
 546                 do_fault_error(regs, fault);
 547 }
 548 NOKPROBE_SYMBOL(do_protection_exception);
 549
 550 void do_dat_exception(struct pt_regs *regs)
 551 {
 552         int access, fault;
 553
 554         access = VM_READ | VM_EXEC | VM_WRITE;
 555         fault = do_exception(regs, access);
 556         if (unlikely(fault))
 557                 do_fault_error(regs, fault);
 558 }
 559 NOKPROBE_SYMBOL(do_dat_exception);
 560
 561 #ifdef CONFIG_PFAULT
 562 /*
 563  * 'pfault' pseudo page faults routines.
 564  */
 565 static int pfault_disable;
 566
 567 static int __init nopfault(char *str)
 568 {
 569         pfault_disable = 1;
 570         return 1;
 571 }
 572
 573 __setup("nopfault", nopfault);
 574
 575 struct pfault_refbk {
 576         u16 refdiagc;
 577         u16 reffcode;
 578         u16 refdwlen;
 579         u16 refversn;
 580         u64 refgaddr;
 581         u64 refselmk;
 582         u64 refcmpmk;
 583         u64 reserved;
 584 } __attribute__ ((packed, aligned(8)));
 585
 586 int pfault_init(void)
 587 {
 588         struct pfault_refbk refbk = {
 589                 .refdiagc = 0x258,
 590                 .reffcode = 0,
 591                 .refdwlen = 5,
 592                 .refversn = 2,
 593                 .refgaddr = __LC_LPP,
 594                 .refselmk = 1ULL << 48,
 595                 .refcmpmk = 1ULL << 48,
 596                 .reserved = __PF_RES_FIELD };
 597         int rc;
 598
 599         if (pfault_disable)
 600                 return -1;
 601         diag_stat_inc(DIAG_STAT_X258);
 602         asm volatile(
 603                 "       diag    %1,%0,0x258\n"
 604                 "0:     j       2f\n"
 605                 "1:     la      %0,8\n"
 606                 "2:\n"
 607                 EX_TABLE(0b,1b)
 608                 : "=d" (rc) : "a" (&refbk), "m" (refbk) : "cc");
 609         return rc;
 610 }
 611
 612 void pfault_fini(void)
 613 {
 614         struct pfault_refbk refbk = {
 615                 .refdiagc = 0x258,
 616                 .reffcode = 1,
 617                 .refdwlen = 5,
 618                 .refversn = 2,
 619         };
 620
 621         if (pfault_disable)
 622                 return;
 623         diag_stat_inc(DIAG_STAT_X258);
 624         asm volatile(
 625                 "       diag    %0,0,0x258\n"
 626                 "0:\n"
 627                 EX_TABLE(0b,0b)
 628                 : : "a" (&refbk), "m" (refbk) : "cc");
 629 }
 630
 631 static DEFINE_SPINLOCK(pfault_lock);
 632 static LIST_HEAD(pfault_list);
 633
 634 static void pfault_interrupt(struct ext_code ext_code,
 635                              unsigned int param32, unsigned long param64)
 636 {
 637         struct task_struct *tsk;
 638         __u16 subcode;
 639         pid_t pid;
 640
 641         /*
 642          * Get the external interruption subcode & pfault
 643          * initial/completion signal bit. VM stores this
 644          * in the 'cpu address' field associated with the
 645          * external interrupt.
 646          */
 647         subcode = ext_code.subcode;
 648         if ((subcode & 0xff00) != __SUBCODE_MASK)
 649                 return;
 650         inc_irq_stat(IRQEXT_PFL);
 651         /* Get the token (= pid of the affected task). */
 652         pid = param64 & LPP_PFAULT_PID_MASK;
 653         rcu_read_lock();
 654         tsk = find_task_by_pid_ns(pid, &init_pid_ns);
 655         if (tsk)
 656                 get_task_struct(tsk);
 657         rcu_read_unlock();
 658         if (!tsk)
 659                 return;
 660         spin_lock(&pfault_lock);
 661         if (subcode & 0x0080) {
 662                 /* signal bit is set -> a page has been swapped in by VM */
 663                 if (tsk->thread.pfault_wait == 1) {
 664                         /* Initial interrupt was faster than the completion
 665                          * interrupt. pfault_wait is valid. Set pfault_wait
 666                          * back to zero and wake up the process. This can
 667                          * safely be done because the task is still sleeping
 668                          * and can't produce new pfaults. */
 669                         tsk->thread.pfault_wait = 0;
 670                         list_del(&tsk->thread.list);
 671                         wake_up_process(tsk);
 672                         put_task_struct(tsk);
 673                 } else {
 674                         /* Completion interrupt was faster than initial
 675                          * interrupt. Set pfault_wait to -1 so the initial
 676                          * interrupt doesn't put the task to sleep.
 677                          * If the task is not running, ignore the completion
 678                          * interrupt since it must be a leftover of a PFAULT
 679                          * CANCEL operation which didn't remove all pending
 680                          * completion interrupts. */
 681                         if (tsk->state == TASK_RUNNING)
 682                                 tsk->thread.pfault_wait = -1;
 683                 }
 684         } else {
 685                 /* signal bit not set -> a real page is missing. */
 686                 if (WARN_ON_ONCE(tsk != current))
 687                         goto out;
 688                 if (tsk->thread.pfault_wait == 1) {
 689                         /* Already on the list with a reference: put to sleep */
 690                         __set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 691                         set_tsk_need_resched(tsk);
 692                 } else if (tsk->thread.pfault_wait == -1) {
 693                         /* Completion interrupt was faster than the initial
 694                          * interrupt (pfault_wait == -1). Set pfault_wait
 695                          * back to zero and exit. */
 696                         tsk->thread.pfault_wait = 0;
 697                 } else {
 698                         /* Initial interrupt arrived before completion
 699                          * interrupt. Let the task sleep.
 700                          * An extra task reference is needed since a different
 701                          * cpu may set the task state to TASK_RUNNING again
 702                          * before the scheduler is reached. */
 703                         get_task_struct(tsk);
 704                         tsk->thread.pfault_wait = 1;
 705                         list_add(&tsk->thread.list, &pfault_list);
 706                         __set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 707                         set_tsk_need_resched(tsk);
 708                 }
 709         }
 710 out:
 711         spin_unlock(&pfault_lock);
 712         put_task_struct(tsk);
 713 }
 714
 715 static int pfault_cpu_notify(struct notifier_block *self, unsigned long action,
 716                              void *hcpu)
 717 {
 718         struct thread_struct *thread, *next;
 719         struct task_struct *tsk;
 720
 721         switch (action & ~CPU_TASKS_FROZEN) {
 722         case CPU_DEAD:
 723                 spin_lock_irq(&pfault_lock);
 724                 list_for_each_entry_safe(thread, next, &pfault_list, list) {
 725                         thread->pfault_wait = 0;
 726                         list_del(&thread->list);
 727                         tsk = container_of(thread, struct task_struct, thread);
 728                         wake_up_process(tsk);
 729                         put_task_struct(tsk);
 730                 }
 731                 spin_unlock_irq(&pfault_lock);
 732                 break;
 733         default:
 734                 break;
 735         }
 736         return NOTIFY_OK;
 737 }
 738
 739 static int __init pfault_irq_init(void)
 740 {
 741         int rc;
 742
 743         rc = register_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
 744         if (rc)
 745                 goto out_extint;
 746         rc = pfault_init() == 0 ? 0 : -EOPNOTSUPP;
 747         if (rc)
 748                 goto out_pfault;
 749         irq_subclass_register(IRQ_SUBCLASS_SERVICE_SIGNAL);
 750         hotcpu_notifier(pfault_cpu_notify, 0);
 751         return 0;
 752
 753 out_pfault:
 754         unregister_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
 755 out_extint:
 756         pfault_disable = 1;
 757         return rc;
 758 }
 759 early_initcall(pfault_irq_init);
 760
 761 #endif /* CONFIG_PFAULT */