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 #ifndef CONFIG_64BIT
  40 #define __FAIL_ADDR_MASK 0x7ffff000
  41 #define __SUBCODE_MASK 0x0200
  42 #define __PF_RES_FIELD 0ULL
  43 #else /* CONFIG_64BIT */
  44 #define __FAIL_ADDR_MASK -4096L
  45 #define __SUBCODE_MASK 0x0600
  46 #define __PF_RES_FIELD 0x8000000000000000ULL
  47 #endif /* CONFIG_64BIT */
  48
  49 #define VM_FAULT_BADCONTEXT     0x010000
  50 #define VM_FAULT_BADMAP         0x020000
  51 #define VM_FAULT_BADACCESS      0x040000
  52 #define VM_FAULT_SIGNAL         0x080000
  53 #define VM_FAULT_PFAULT         0x100000
  54
  55 static unsigned long store_indication __read_mostly;
  56
  57 #ifdef CONFIG_64BIT
  58 static int __init fault_init(void)
  59 {
  60         if (test_facility(75))
  61                 store_indication = 0xc00;
  62         return 0;
  63 }
  64 early_initcall(fault_init);
  65 #endif
  66
  67 static inline int notify_page_fault(struct pt_regs *regs)
  68 {
  69         int ret = 0;
  70
  71         /* kprobe_running() needs smp_processor_id() */
  72         if (kprobes_built_in() && !user_mode(regs)) {
  73                 preempt_disable();
  74                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  75                         ret = 1;
  76                 preempt_enable();
  77         }
  78         return ret;
  79 }
  80
  81
  82 /*
  83  * Unlock any spinlocks which will prevent us from getting the
  84  * message out.
  85  */
  86 void bust_spinlocks(int yes)
  87 {
  88         if (yes) {
  89                 oops_in_progress = 1;
  90         } else {
  91                 int loglevel_save = console_loglevel;
  92                 console_unblank();
  93                 oops_in_progress = 0;
  94                 /*
  95                  * OK, the message is on the console.  Now we call printk()
  96                  * without oops_in_progress set so that printk will give klogd
  97                  * a poke.  Hold onto your hats...
  98                  */
  99                 console_loglevel = 15;
 100                 printk(" ");
 101                 console_loglevel = loglevel_save;
 102         }
 103 }
 104
 105 /*
 106  * Returns the address space associated with the fault.
 107  * Returns 0 for kernel space and 1 for user space.
 108  */
 109 static inline int user_space_fault(struct pt_regs *regs)
 110 {
 111         unsigned long trans_exc_code;
 112
 113         /*
 114          * The lowest two bits of the translation exception
 115          * identification indicate which paging table was used.
 116          */
 117         trans_exc_code = regs->int_parm_long & 3;
 118         if (trans_exc_code == 3) /* home space -> kernel */
 119                 return 0;
 120         if (user_mode(regs))
 121                 return 1;
 122         if (trans_exc_code == 2) /* secondary space -> set_fs */
 123                 return current->thread.mm_segment.ar4;
 124         if (current->flags & PF_VCPU)
 125                 return 1;
 126         return 0;
 127 }
 128
 129 static int bad_address(void *p)
 130 {
 131         unsigned long dummy;
 132
 133         return probe_kernel_address((unsigned long *)p, dummy);
 134 }
 135
 136 #ifdef CONFIG_64BIT
 137 static void dump_pagetable(unsigned long asce, unsigned long address)
 138 {
 139         unsigned long *table = __va(asce & PAGE_MASK);
 140
 141         pr_alert("AS:%016lx ", asce);
 142         switch (asce & _ASCE_TYPE_MASK) {
 143         case _ASCE_TYPE_REGION1:
 144                 table = table + ((address >> 53) & 0x7ff);
 145                 if (bad_address(table))
 146                         goto bad;
 147                 pr_cont("R1:%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_REGION2:
 153                 table = table + ((address >> 42) & 0x7ff);
 154                 if (bad_address(table))
 155                         goto bad;
 156                 pr_cont("R2:%016lx ", *table);
 157                 if (*table & _REGION_ENTRY_INVALID)
 158                         goto out;
 159                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 160                 /* fallthrough */
 161         case _ASCE_TYPE_REGION3:
 162                 table = table + ((address >> 31) & 0x7ff);
 163                 if (bad_address(table))
 164                         goto bad;
 165                 pr_cont("R3:%016lx ", *table);
 166                 if (*table & (_REGION_ENTRY_INVALID | _REGION3_ENTRY_LARGE))
 167                         goto out;
 168                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 169                 /* fallthrough */
 170         case _ASCE_TYPE_SEGMENT:
 171                 table = table + ((address >> 20) & 0x7ff);
 172                 if (bad_address(table))
 173                         goto bad;
 174                 pr_cont(KERN_CONT "S:%016lx ", *table);
 175                 if (*table & (_SEGMENT_ENTRY_INVALID | _SEGMENT_ENTRY_LARGE))
 176                         goto out;
 177                 table = (unsigned long *)(*table & _SEGMENT_ENTRY_ORIGIN);
 178         }
 179         table = table + ((address >> 12) & 0xff);
 180         if (bad_address(table))
 181                 goto bad;
 182         pr_cont("P:%016lx ", *table);
 183 out:
 184         pr_cont("\n");
 185         return;
 186 bad:
 187         pr_cont("BAD\n");
 188 }
 189
 190 #else /* CONFIG_64BIT */
 191
 192 static void dump_pagetable(unsigned long asce, unsigned long address)
 193 {
 194         unsigned long *table = __va(asce & PAGE_MASK);
 195
 196         pr_alert("AS:%08lx ", asce);
 197         table = table + ((address >> 20) & 0x7ff);
 198         if (bad_address(table))
 199                 goto bad;
 200         pr_cont("S:%08lx ", *table);
 201         if (*table & _SEGMENT_ENTRY_INVALID)
 202                 goto out;
 203         table = (unsigned long *)(*table & _SEGMENT_ENTRY_ORIGIN);
 204         table = table + ((address >> 12) & 0xff);
 205         if (bad_address(table))
 206                 goto bad;
 207         pr_cont("P:%08lx ", *table);
 208 out:
 209         pr_cont("\n");
 210         return;
 211 bad:
 212         pr_cont("BAD\n");
 213 }
 214
 215 #endif /* CONFIG_64BIT */
 216
 217 static void dump_fault_info(struct pt_regs *regs)
 218 {
 219         unsigned long asce;
 220
 221         pr_alert("Fault in ");
 222         switch (regs->int_parm_long & 3) {
 223         case 3:
 224                 pr_cont("home space ");
 225                 break;
 226         case 2:
 227                 pr_cont("secondary space ");
 228                 break;
 229         case 1:
 230                 pr_cont("access register ");
 231                 break;
 232         case 0:
 233                 pr_cont("primary space ");
 234                 break;
 235         }
 236         pr_cont("mode while using ");
 237         if (!user_space_fault(regs)) {
 238                 asce = S390_lowcore.kernel_asce;
 239                 pr_cont("kernel ");
 240         }
 241 #ifdef CONFIG_PGSTE
 242         else if ((current->flags & PF_VCPU) && S390_lowcore.gmap) {
 243                 struct gmap *gmap = (struct gmap *)S390_lowcore.gmap;
 244                 asce = gmap->asce;
 245                 pr_cont("gmap ");
 246         }
 247 #endif
 248         else {
 249                 asce = S390_lowcore.user_asce;
 250                 pr_cont("user ");
 251         }
 252         pr_cont("ASCE.\n");
 253         dump_pagetable(asce, regs->int_parm_long & __FAIL_ADDR_MASK);
 254 }
 255
 256 static inline void report_user_fault(struct pt_regs *regs, long signr)
 257 {
 258         if ((task_pid_nr(current) > 1) && !show_unhandled_signals)
 259                 return;
 260         if (!unhandled_signal(current, signr))
 261                 return;
 262         if (!printk_ratelimit())
 263                 return;
 264         printk(KERN_ALERT "User process fault: interruption code %04x ilc:%d",
 265                regs->int_code & 0xffff, regs->int_code >> 17);
 266         print_vma_addr(KERN_CONT "in ", regs->psw.addr & PSW_ADDR_INSN);
 267         printk(KERN_CONT "\n");
 268         printk(KERN_ALERT "failing address: %016lx TEID: %016lx\n",
 269                regs->int_parm_long & __FAIL_ADDR_MASK, regs->int_parm_long);
 270         dump_fault_info(regs);
 271         show_regs(regs);
 272 }
 273
 274 /*
 275  * Send SIGSEGV to task.  This is an external routine
 276  * to keep the stack usage of do_page_fault small.
 277  */
 278 static noinline void do_sigsegv(struct pt_regs *regs, int si_code)
 279 {
 280         struct siginfo si;
 281
 282         report_user_fault(regs, SIGSEGV);
 283         si.si_signo = SIGSEGV;
 284         si.si_code = si_code;
 285         si.si_addr = (void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK);
 286         force_sig_info(SIGSEGV, &si, current);
 287 }
 288
 289 static noinline void do_no_context(struct pt_regs *regs)
 290 {
 291         const struct exception_table_entry *fixup;
 292         unsigned long address;
 293
 294         /* Are we prepared to handle this kernel fault?  */
 295         fixup = search_exception_tables(regs->psw.addr & PSW_ADDR_INSN);
 296         if (fixup) {
 297                 regs->psw.addr = extable_fixup(fixup) | PSW_ADDR_AMODE;
 298                 return;
 299         }
 300
 301         /*
 302          * Oops. The kernel tried to access some bad page. We'll have to
 303          * terminate things with extreme prejudice.
 304          */
 305         address = regs->int_parm_long & __FAIL_ADDR_MASK;
 306         if (!user_space_fault(regs))
 307                 printk(KERN_ALERT "Unable to handle kernel pointer dereference"
 308                        " in virtual kernel address space\n");
 309         else
 310                 printk(KERN_ALERT "Unable to handle kernel paging request"
 311                        " in virtual user address space\n");
 312         printk(KERN_ALERT "failing address: %016lx TEID: %016lx\n",
 313                regs->int_parm_long & __FAIL_ADDR_MASK, regs->int_parm_long);
 314         dump_fault_info(regs);
 315         die(regs, "Oops");
 316         do_exit(SIGKILL);
 317 }
 318
 319 static noinline void do_low_address(struct pt_regs *regs)
 320 {
 321         /* Low-address protection hit in kernel mode means
 322            NULL pointer write access in kernel mode.  */
 323         if (regs->psw.mask & PSW_MASK_PSTATE) {
 324                 /* Low-address protection hit in user mode 'cannot happen'. */
 325                 die (regs, "Low-address protection");
 326                 do_exit(SIGKILL);
 327         }
 328
 329         do_no_context(regs);
 330 }
 331
 332 static noinline void do_sigbus(struct pt_regs *regs)
 333 {
 334         struct task_struct *tsk = current;
 335         struct siginfo si;
 336
 337         /*
 338          * Send a sigbus, regardless of whether we were in kernel
 339          * or user mode.
 340          */
 341         si.si_signo = SIGBUS;
 342         si.si_errno = 0;
 343         si.si_code = BUS_ADRERR;
 344         si.si_addr = (void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK);
 345         force_sig_info(SIGBUS, &si, tsk);
 346 }
 347
 348 static noinline void do_fault_error(struct pt_regs *regs, int fault)
 349 {
 350         int si_code;
 351
 352         switch (fault) {
 353         case VM_FAULT_BADACCESS:
 354         case VM_FAULT_BADMAP:
 355                 /* Bad memory access. Check if it is kernel or user space. */
 356                 if (user_mode(regs)) {
 357                         /* User mode accesses just cause a SIGSEGV */
 358                         si_code = (fault == VM_FAULT_BADMAP) ?
 359                                 SEGV_MAPERR : SEGV_ACCERR;
 360                         do_sigsegv(regs, si_code);
 361                         return;
 362                 }
 363         case VM_FAULT_BADCONTEXT:
 364         case VM_FAULT_PFAULT:
 365                 do_no_context(regs);
 366                 break;
 367         case VM_FAULT_SIGNAL:
 368                 if (!user_mode(regs))
 369                         do_no_context(regs);
 370                 break;
 371         default: /* fault & VM_FAULT_ERROR */
 372                 if (fault & VM_FAULT_OOM) {
 373                         if (!user_mode(regs))
 374                                 do_no_context(regs);
 375                         else
 376                                 pagefault_out_of_memory();
 377                 } else if (fault & VM_FAULT_SIGBUS) {
 378                         /* Kernel mode? Handle exceptions or die */
 379                         if (!user_mode(regs))
 380                                 do_no_context(regs);
 381                         else
 382                                 do_sigbus(regs);
 383                 } else
 384                         BUG();
 385                 break;
 386         }
 387 }
 388
 389 /*
 390  * This routine handles page faults.  It determines the address,
 391  * and the problem, and then passes it off to one of the appropriate
 392  * routines.
 393  *
 394  * interruption code (int_code):
 395  *   04       Protection           ->  Write-Protection  (suprression)
 396  *   10       Segment translation  ->  Not present       (nullification)
 397  *   11       Page translation     ->  Not present       (nullification)
 398  *   3b       Region third trans.  ->  Not present       (nullification)
 399  */
 400 static inline int do_exception(struct pt_regs *regs, int access)
 401 {
 402 #ifdef CONFIG_PGSTE
 403         struct gmap *gmap;
 404 #endif
 405         struct task_struct *tsk;
 406         struct mm_struct *mm;
 407         struct vm_area_struct *vma;
 408         unsigned long trans_exc_code;
 409         unsigned long address;
 410         unsigned int flags;
 411         int fault;
 412
 413         tsk = current;
 414         /*
 415          * The instruction that caused the program check has
 416          * been nullified. Don't signal single step via SIGTRAP.
 417          */
 418         clear_pt_regs_flag(regs, PIF_PER_TRAP);
 419
 420         if (notify_page_fault(regs))
 421                 return 0;
 422
 423         mm = tsk->mm;
 424         trans_exc_code = regs->int_parm_long;
 425
 426         /*
 427          * Verify that the fault happened in user space, that
 428          * we are not in an interrupt and that there is a
 429          * user context.
 430          */
 431         fault = VM_FAULT_BADCONTEXT;
 432         if (unlikely(!user_space_fault(regs) || in_atomic() || !mm))
 433                 goto out;
 434
 435         address = trans_exc_code & __FAIL_ADDR_MASK;
 436         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
 437         flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
 438         if (user_mode(regs))
 439                 flags |= FAULT_FLAG_USER;
 440         if (access == VM_WRITE || (trans_exc_code & store_indication) == 0x400)
 441                 flags |= FAULT_FLAG_WRITE;
 442         down_read(&mm->mmap_sem);
 443
 444 #ifdef CONFIG_PGSTE
 445         gmap = (current->flags & PF_VCPU) ?
 446                 (struct gmap *) S390_lowcore.gmap : NULL;
 447         if (gmap) {
 448                 current->thread.gmap_addr = address;
 449                 address = __gmap_translate(gmap, address);
 450                 if (address == -EFAULT) {
 451                         fault = VM_FAULT_BADMAP;
 452                         goto out_up;
 453                 }
 454                 if (gmap->pfault_enabled)
 455                         flags |= FAULT_FLAG_RETRY_NOWAIT;
 456         }
 457 #endif
 458
 459 retry:
 460         fault = VM_FAULT_BADMAP;
 461         vma = find_vma(mm, address);
 462         if (!vma)
 463                 goto out_up;
 464
 465         if (unlikely(vma->vm_start > address)) {
 466                 if (!(vma->vm_flags & VM_GROWSDOWN))
 467                         goto out_up;
 468                 if (expand_stack(vma, address))
 469                         goto out_up;
 470         }
 471
 472         /*
 473          * Ok, we have a good vm_area for this memory access, so
 474          * we can handle it..
 475          */
 476         fault = VM_FAULT_BADACCESS;
 477         if (unlikely(!(vma->vm_flags & access)))
 478                 goto out_up;
 479
 480         if (is_vm_hugetlb_page(vma))
 481                 address &= HPAGE_MASK;
 482         /*
 483          * If for any reason at all we couldn't handle the fault,
 484          * make sure we exit gracefully rather than endlessly redo
 485          * the fault.
 486          */
 487         fault = handle_mm_fault(mm, vma, address, flags);
 488         /* No reason to continue if interrupted by SIGKILL. */
 489         if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)) {
 490                 fault = VM_FAULT_SIGNAL;
 491                 goto out;
 492         }
 493         if (unlikely(fault & VM_FAULT_ERROR))
 494                 goto out_up;
 495
 496         /*
 497          * Major/minor page fault accounting is only done on the
 498          * initial attempt. If we go through a retry, it is extremely
 499          * likely that the page will be found in page cache at that point.
 500          */
 501         if (flags & FAULT_FLAG_ALLOW_RETRY) {
 502                 if (fault & VM_FAULT_MAJOR) {
 503                         tsk->maj_flt++;
 504                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
 505                                       regs, address);
 506                 } else {
 507                         tsk->min_flt++;
 508                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
 509                                       regs, address);
 510                 }
 511                 if (fault & VM_FAULT_RETRY) {
 512 #ifdef CONFIG_PGSTE
 513                         if (gmap && (flags & FAULT_FLAG_RETRY_NOWAIT)) {
 514                                 /* FAULT_FLAG_RETRY_NOWAIT has been set,
 515                                  * mmap_sem has not been released */
 516                                 current->thread.gmap_pfault = 1;
 517                                 fault = VM_FAULT_PFAULT;
 518                                 goto out_up;
 519                         }
 520 #endif
 521                         /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
 522                          * of starvation. */
 523                         flags &= ~(FAULT_FLAG_ALLOW_RETRY |
 524                                    FAULT_FLAG_RETRY_NOWAIT);
 525                         flags |= FAULT_FLAG_TRIED;
 526                         down_read(&mm->mmap_sem);
 527                         goto retry;
 528                 }
 529         }
 530 #ifdef CONFIG_PGSTE
 531         if (gmap) {
 532                 address =  __gmap_link(gmap, current->thread.gmap_addr,
 533                                        address);
 534                 if (address == -EFAULT) {
 535                         fault = VM_FAULT_BADMAP;
 536                         goto out_up;
 537                 }
 538                 if (address == -ENOMEM) {
 539                         fault = VM_FAULT_OOM;
 540                         goto out_up;
 541                 }
 542         }
 543 #endif
 544         fault = 0;
 545 out_up:
 546         up_read(&mm->mmap_sem);
 547 out:
 548         return fault;
 549 }
 550
 551 void do_protection_exception(struct pt_regs *regs)
 552 {
 553         unsigned long trans_exc_code;
 554         int fault;
 555
 556         trans_exc_code = regs->int_parm_long;
 557         /*
 558          * Protection exceptions are suppressing, decrement psw address.
 559          * The exception to this rule are aborted transactions, for these
 560          * the PSW already points to the correct location.
 561          */
 562         if (!(regs->int_code & 0x200))
 563                 regs->psw.addr = __rewind_psw(regs->psw, regs->int_code >> 16);
 564         /*
 565          * Check for low-address protection.  This needs to be treated
 566          * as a special case because the translation exception code
 567          * field is not guaranteed to contain valid data in this case.
 568          */
 569         if (unlikely(!(trans_exc_code & 4))) {
 570                 do_low_address(regs);
 571                 return;
 572         }
 573         fault = do_exception(regs, VM_WRITE);
 574         if (unlikely(fault))
 575                 do_fault_error(regs, fault);
 576 }
 577 NOKPROBE_SYMBOL(do_protection_exception);
 578
 579 void do_dat_exception(struct pt_regs *regs)
 580 {
 581         int access, fault;
 582
 583         access = VM_READ | VM_EXEC | VM_WRITE;
 584         fault = do_exception(regs, access);
 585         if (unlikely(fault))
 586                 do_fault_error(regs, fault);
 587 }
 588 NOKPROBE_SYMBOL(do_dat_exception);
 589
 590 #ifdef CONFIG_PFAULT
 591 /*
 592  * 'pfault' pseudo page faults routines.
 593  */
 594 static int pfault_disable;
 595
 596 static int __init nopfault(char *str)
 597 {
 598         pfault_disable = 1;
 599         return 1;
 600 }
 601
 602 __setup("nopfault", nopfault);
 603
 604 struct pfault_refbk {
 605         u16 refdiagc;
 606         u16 reffcode;
 607         u16 refdwlen;
 608         u16 refversn;
 609         u64 refgaddr;
 610         u64 refselmk;
 611         u64 refcmpmk;
 612         u64 reserved;
 613 } __attribute__ ((packed, aligned(8)));
 614
 615 int pfault_init(void)
 616 {
 617         struct pfault_refbk refbk = {
 618                 .refdiagc = 0x258,
 619                 .reffcode = 0,
 620                 .refdwlen = 5,
 621                 .refversn = 2,
 622                 .refgaddr = __LC_CURRENT_PID,
 623                 .refselmk = 1ULL << 48,
 624                 .refcmpmk = 1ULL << 48,
 625                 .reserved = __PF_RES_FIELD };
 626         int rc;
 627
 628         if (pfault_disable)
 629                 return -1;
 630         asm volatile(
 631                 "       diag    %1,%0,0x258\n"
 632                 "0:     j       2f\n"
 633                 "1:     la      %0,8\n"
 634                 "2:\n"
 635                 EX_TABLE(0b,1b)
 636                 : "=d" (rc) : "a" (&refbk), "m" (refbk) : "cc");
 637         return rc;
 638 }
 639
 640 void pfault_fini(void)
 641 {
 642         struct pfault_refbk refbk = {
 643                 .refdiagc = 0x258,
 644                 .reffcode = 1,
 645                 .refdwlen = 5,
 646                 .refversn = 2,
 647         };
 648
 649         if (pfault_disable)
 650                 return;
 651         asm volatile(
 652                 "       diag    %0,0,0x258\n"
 653                 "0:\n"
 654                 EX_TABLE(0b,0b)
 655                 : : "a" (&refbk), "m" (refbk) : "cc");
 656 }
 657
 658 static DEFINE_SPINLOCK(pfault_lock);
 659 static LIST_HEAD(pfault_list);
 660
 661 static void pfault_interrupt(struct ext_code ext_code,
 662                              unsigned int param32, unsigned long param64)
 663 {
 664         struct task_struct *tsk;
 665         __u16 subcode;
 666         pid_t pid;
 667
 668         /*
 669          * Get the external interruption subcode & pfault
 670          * initial/completion signal bit. VM stores this
 671          * in the 'cpu address' field associated with the
 672          * external interrupt.
 673          */
 674         subcode = ext_code.subcode;
 675         if ((subcode & 0xff00) != __SUBCODE_MASK)
 676                 return;
 677         inc_irq_stat(IRQEXT_PFL);
 678         /* Get the token (= pid of the affected task). */
 679         pid = sizeof(void *) == 4 ? param32 : param64;
 680         rcu_read_lock();
 681         tsk = find_task_by_pid_ns(pid, &init_pid_ns);
 682         if (tsk)
 683                 get_task_struct(tsk);
 684         rcu_read_unlock();
 685         if (!tsk)
 686                 return;
 687         spin_lock(&pfault_lock);
 688         if (subcode & 0x0080) {
 689                 /* signal bit is set -> a page has been swapped in by VM */
 690                 if (tsk->thread.pfault_wait == 1) {
 691                         /* Initial interrupt was faster than the completion
 692                          * interrupt. pfault_wait is valid. Set pfault_wait
 693                          * back to zero and wake up the process. This can
 694                          * safely be done because the task is still sleeping
 695                          * and can't produce new pfaults. */
 696                         tsk->thread.pfault_wait = 0;
 697                         list_del(&tsk->thread.list);
 698                         wake_up_process(tsk);
 699                         put_task_struct(tsk);
 700                 } else {
 701                         /* Completion interrupt was faster than initial
 702                          * interrupt. Set pfault_wait to -1 so the initial
 703                          * interrupt doesn't put the task to sleep.
 704                          * If the task is not running, ignore the completion
 705                          * interrupt since it must be a leftover of a PFAULT
 706                          * CANCEL operation which didn't remove all pending
 707                          * completion interrupts. */
 708                         if (tsk->state == TASK_RUNNING)
 709                                 tsk->thread.pfault_wait = -1;
 710                 }
 711         } else {
 712                 /* signal bit not set -> a real page is missing. */
 713                 if (WARN_ON_ONCE(tsk != current))
 714                         goto out;
 715                 if (tsk->thread.pfault_wait == 1) {
 716                         /* Already on the list with a reference: put to sleep */
 717                         __set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 718                         set_tsk_need_resched(tsk);
 719                 } else if (tsk->thread.pfault_wait == -1) {
 720                         /* Completion interrupt was faster than the initial
 721                          * interrupt (pfault_wait == -1). Set pfault_wait
 722                          * back to zero and exit. */
 723                         tsk->thread.pfault_wait = 0;
 724                 } else {
 725                         /* Initial interrupt arrived before completion
 726                          * interrupt. Let the task sleep.
 727                          * An extra task reference is needed since a different
 728                          * cpu may set the task state to TASK_RUNNING again
 729                          * before the scheduler is reached. */
 730                         get_task_struct(tsk);
 731                         tsk->thread.pfault_wait = 1;
 732                         list_add(&tsk->thread.list, &pfault_list);
 733                         __set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 734                         set_tsk_need_resched(tsk);
 735                 }
 736         }
 737 out:
 738         spin_unlock(&pfault_lock);
 739         put_task_struct(tsk);
 740 }
 741
 742 static int pfault_cpu_notify(struct notifier_block *self, unsigned long action,
 743                              void *hcpu)
 744 {
 745         struct thread_struct *thread, *next;
 746         struct task_struct *tsk;
 747
 748         switch (action & ~CPU_TASKS_FROZEN) {
 749         case CPU_DEAD:
 750                 spin_lock_irq(&pfault_lock);
 751                 list_for_each_entry_safe(thread, next, &pfault_list, list) {
 752                         thread->pfault_wait = 0;
 753                         list_del(&thread->list);
 754                         tsk = container_of(thread, struct task_struct, thread);
 755                         wake_up_process(tsk);
 756                         put_task_struct(tsk);
 757                 }
 758                 spin_unlock_irq(&pfault_lock);
 759                 break;
 760         default:
 761                 break;
 762         }
 763         return NOTIFY_OK;
 764 }
 765
 766 static int __init pfault_irq_init(void)
 767 {
 768         int rc;
 769
 770         rc = register_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
 771         if (rc)
 772                 goto out_extint;
 773         rc = pfault_init() == 0 ? 0 : -EOPNOTSUPP;
 774         if (rc)
 775                 goto out_pfault;
 776         irq_subclass_register(IRQ_SUBCLASS_SERVICE_SIGNAL);
 777         hotcpu_notifier(pfault_cpu_notify, 0);
 778         return 0;
 779
 780 out_pfault:
 781         unregister_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
 782 out_extint:
 783         pfault_disable = 1;
 784         return rc;
 785 }
 786 early_initcall(pfault_irq_init);
 787
 788 #endif /* CONFIG_PFAULT */