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