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("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_SIGSEGV) {
 378                         /* Kernel mode? Handle exceptions or die */
 379                         if (!user_mode(regs))
 380                                 do_no_context(regs);
 381                         else
 382                                 do_sigsegv(regs, SEGV_MAPERR);
 383                 } else if (fault & VM_FAULT_SIGBUS) {
 384                         /* Kernel mode? Handle exceptions or die */
 385                         if (!user_mode(regs))
 386                                 do_no_context(regs);
 387                         else
 388                                 do_sigbus(regs);
 389                 } else
 390                         BUG();
 391                 break;
 392         }
 393 }
 394
 395 /*
 396  * This routine handles page faults.  It determines the address,
 397  * and the problem, and then passes it off to one of the appropriate
 398  * routines.
 399  *
 400  * interruption code (int_code):
 401  *   04       Protection           ->  Write-Protection  (suprression)
 402  *   10       Segment translation  ->  Not present       (nullification)
 403  *   11       Page translation     ->  Not present       (nullification)
 404  *   3b       Region third trans.  ->  Not present       (nullification)
 405  */
 406 static inline int do_exception(struct pt_regs *regs, int access)
 407 {
 408 #ifdef CONFIG_PGSTE
 409         struct gmap *gmap;
 410 #endif
 411         struct task_struct *tsk;
 412         struct mm_struct *mm;
 413         struct vm_area_struct *vma;
 414         unsigned long trans_exc_code;
 415         unsigned long address;
 416         unsigned int flags;
 417         int fault;
 418
 419         tsk = current;
 420         /*
 421          * The instruction that caused the program check has
 422          * been nullified. Don't signal single step via SIGTRAP.
 423          */
 424         clear_pt_regs_flag(regs, PIF_PER_TRAP);
 425
 426         if (notify_page_fault(regs))
 427                 return 0;
 428
 429         mm = tsk->mm;
 430         trans_exc_code = regs->int_parm_long;
 431
 432         /*
 433          * Verify that the fault happened in user space, that
 434          * we are not in an interrupt and that there is a
 435          * user context.
 436          */
 437         fault = VM_FAULT_BADCONTEXT;
 438         if (unlikely(!user_space_fault(regs) || in_atomic() || !mm))
 439                 goto out;
 440
 441         address = trans_exc_code & __FAIL_ADDR_MASK;
 442         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
 443         flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
 444         if (user_mode(regs))
 445                 flags |= FAULT_FLAG_USER;
 446         if (access == VM_WRITE || (trans_exc_code & store_indication) == 0x400)
 447                 flags |= FAULT_FLAG_WRITE;
 448         down_read(&mm->mmap_sem);
 449
 450 #ifdef CONFIG_PGSTE
 451         gmap = (current->flags & PF_VCPU) ?
 452                 (struct gmap *) S390_lowcore.gmap : NULL;
 453         if (gmap) {
 454                 current->thread.gmap_addr = address;
 455                 address = __gmap_translate(gmap, address);
 456                 if (address == -EFAULT) {
 457                         fault = VM_FAULT_BADMAP;
 458                         goto out_up;
 459                 }
 460                 if (gmap->pfault_enabled)
 461                         flags |= FAULT_FLAG_RETRY_NOWAIT;
 462         }
 463 #endif
 464
 465 retry:
 466         fault = VM_FAULT_BADMAP;
 467         vma = find_vma(mm, address);
 468         if (!vma)
 469                 goto out_up;
 470
 471         if (unlikely(vma->vm_start > address)) {
 472                 if (!(vma->vm_flags & VM_GROWSDOWN))
 473                         goto out_up;
 474                 if (expand_stack(vma, address))
 475                         goto out_up;
 476         }
 477
 478         /*
 479          * Ok, we have a good vm_area for this memory access, so
 480          * we can handle it..
 481          */
 482         fault = VM_FAULT_BADACCESS;
 483         if (unlikely(!(vma->vm_flags & access)))
 484                 goto out_up;
 485
 486         if (is_vm_hugetlb_page(vma))
 487                 address &= HPAGE_MASK;
 488         /*
 489          * If for any reason at all we couldn't handle the fault,
 490          * make sure we exit gracefully rather than endlessly redo
 491          * the fault.
 492          */
 493         fault = handle_mm_fault(mm, vma, address, flags);
 494         /* No reason to continue if interrupted by SIGKILL. */
 495         if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)) {
 496                 fault = VM_FAULT_SIGNAL;
 497                 goto out;
 498         }
 499         if (unlikely(fault & VM_FAULT_ERROR))
 500                 goto out_up;
 501
 502         /*
 503          * Major/minor page fault accounting is only done on the
 504          * initial attempt. If we go through a retry, it is extremely
 505          * likely that the page will be found in page cache at that point.
 506          */
 507         if (flags & FAULT_FLAG_ALLOW_RETRY) {
 508                 if (fault & VM_FAULT_MAJOR) {
 509                         tsk->maj_flt++;
 510                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
 511                                       regs, address);
 512                 } else {
 513                         tsk->min_flt++;
 514                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
 515                                       regs, address);
 516                 }
 517                 if (fault & VM_FAULT_RETRY) {
 518 #ifdef CONFIG_PGSTE
 519                         if (gmap && (flags & FAULT_FLAG_RETRY_NOWAIT)) {
 520                                 /* FAULT_FLAG_RETRY_NOWAIT has been set,
 521                                  * mmap_sem has not been released */
 522                                 current->thread.gmap_pfault = 1;
 523                                 fault = VM_FAULT_PFAULT;
 524                                 goto out_up;
 525                         }
 526 #endif
 527                         /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
 528                          * of starvation. */
 529                         flags &= ~(FAULT_FLAG_ALLOW_RETRY |
 530                                    FAULT_FLAG_RETRY_NOWAIT);
 531                         flags |= FAULT_FLAG_TRIED;
 532                         down_read(&mm->mmap_sem);
 533                         goto retry;
 534                 }
 535         }
 536 #ifdef CONFIG_PGSTE
 537         if (gmap) {
 538                 address =  __gmap_link(gmap, current->thread.gmap_addr,
 539                                        address);
 540                 if (address == -EFAULT) {
 541                         fault = VM_FAULT_BADMAP;
 542                         goto out_up;
 543                 }
 544                 if (address == -ENOMEM) {
 545                         fault = VM_FAULT_OOM;
 546                         goto out_up;
 547                 }
 548         }
 549 #endif
 550         fault = 0;
 551 out_up:
 552         up_read(&mm->mmap_sem);
 553 out:
 554         return fault;
 555 }
 556
 557 void do_protection_exception(struct pt_regs *regs)
 558 {
 559         unsigned long trans_exc_code;
 560         int fault;
 561
 562         trans_exc_code = regs->int_parm_long;
 563         /*
 564          * Protection exceptions are suppressing, decrement psw address.
 565          * The exception to this rule are aborted transactions, for these
 566          * the PSW already points to the correct location.
 567          */
 568         if (!(regs->int_code & 0x200))
 569                 regs->psw.addr = __rewind_psw(regs->psw, regs->int_code >> 16);
 570         /*
 571          * Check for low-address protection.  This needs to be treated
 572          * as a special case because the translation exception code
 573          * field is not guaranteed to contain valid data in this case.
 574          */
 575         if (unlikely(!(trans_exc_code & 4))) {
 576                 do_low_address(regs);
 577                 return;
 578         }
 579         fault = do_exception(regs, VM_WRITE);
 580         if (unlikely(fault))
 581                 do_fault_error(regs, fault);
 582 }
 583 NOKPROBE_SYMBOL(do_protection_exception);
 584
 585 void do_dat_exception(struct pt_regs *regs)
 586 {
 587         int access, fault;
 588
 589         access = VM_READ | VM_EXEC | VM_WRITE;
 590         fault = do_exception(regs, access);
 591         if (unlikely(fault))
 592                 do_fault_error(regs, fault);
 593 }
 594 NOKPROBE_SYMBOL(do_dat_exception);
 595
 596 #ifdef CONFIG_PFAULT
 597 /*
 598  * 'pfault' pseudo page faults routines.
 599  */
 600 static int pfault_disable;
 601
 602 static int __init nopfault(char *str)
 603 {
 604         pfault_disable = 1;
 605         return 1;
 606 }
 607
 608 __setup("nopfault", nopfault);
 609
 610 struct pfault_refbk {
 611         u16 refdiagc;
 612         u16 reffcode;
 613         u16 refdwlen;
 614         u16 refversn;
 615         u64 refgaddr;
 616         u64 refselmk;
 617         u64 refcmpmk;
 618         u64 reserved;
 619 } __attribute__ ((packed, aligned(8)));
 620
 621 int pfault_init(void)
 622 {
 623         struct pfault_refbk refbk = {
 624                 .refdiagc = 0x258,
 625                 .reffcode = 0,
 626                 .refdwlen = 5,
 627                 .refversn = 2,
 628                 .refgaddr = __LC_CURRENT_PID,
 629                 .refselmk = 1ULL << 48,
 630                 .refcmpmk = 1ULL << 48,
 631                 .reserved = __PF_RES_FIELD };
 632         int rc;
 633
 634         if (pfault_disable)
 635                 return -1;
 636         asm volatile(
 637                 "       diag    %1,%0,0x258\n"
 638                 "0:     j       2f\n"
 639                 "1:     la      %0,8\n"
 640                 "2:\n"
 641                 EX_TABLE(0b,1b)
 642                 : "=d" (rc) : "a" (&refbk), "m" (refbk) : "cc");
 643         return rc;
 644 }
 645
 646 void pfault_fini(void)
 647 {
 648         struct pfault_refbk refbk = {
 649                 .refdiagc = 0x258,
 650                 .reffcode = 1,
 651                 .refdwlen = 5,
 652                 .refversn = 2,
 653         };
 654
 655         if (pfault_disable)
 656                 return;
 657         asm volatile(
 658                 "       diag    %0,0,0x258\n"
 659                 "0:\n"
 660                 EX_TABLE(0b,0b)
 661                 : : "a" (&refbk), "m" (refbk) : "cc");
 662 }
 663
 664 static DEFINE_SPINLOCK(pfault_lock);
 665 static LIST_HEAD(pfault_list);
 666
 667 static void pfault_interrupt(struct ext_code ext_code,
 668                              unsigned int param32, unsigned long param64)
 669 {
 670         struct task_struct *tsk;
 671         __u16 subcode;
 672         pid_t pid;
 673
 674         /*
 675          * Get the external interruption subcode & pfault
 676          * initial/completion signal bit. VM stores this
 677          * in the 'cpu address' field associated with the
 678          * external interrupt.
 679          */
 680         subcode = ext_code.subcode;
 681         if ((subcode & 0xff00) != __SUBCODE_MASK)
 682                 return;
 683         inc_irq_stat(IRQEXT_PFL);
 684         /* Get the token (= pid of the affected task). */
 685         pid = sizeof(void *) == 4 ? param32 : param64;
 686         rcu_read_lock();
 687         tsk = find_task_by_pid_ns(pid, &init_pid_ns);
 688         if (tsk)
 689                 get_task_struct(tsk);
 690         rcu_read_unlock();
 691         if (!tsk)
 692                 return;
 693         spin_lock(&pfault_lock);
 694         if (subcode & 0x0080) {
 695                 /* signal bit is set -> a page has been swapped in by VM */
 696                 if (tsk->thread.pfault_wait == 1) {
 697                         /* Initial interrupt was faster than the completion
 698                          * interrupt. pfault_wait is valid. Set pfault_wait
 699                          * back to zero and wake up the process. This can
 700                          * safely be done because the task is still sleeping
 701                          * and can't produce new pfaults. */
 702                         tsk->thread.pfault_wait = 0;
 703                         list_del(&tsk->thread.list);
 704                         wake_up_process(tsk);
 705                         put_task_struct(tsk);
 706                 } else {
 707                         /* Completion interrupt was faster than initial
 708                          * interrupt. Set pfault_wait to -1 so the initial
 709                          * interrupt doesn't put the task to sleep.
 710                          * If the task is not running, ignore the completion
 711                          * interrupt since it must be a leftover of a PFAULT
 712                          * CANCEL operation which didn't remove all pending
 713                          * completion interrupts. */
 714                         if (tsk->state == TASK_RUNNING)
 715                                 tsk->thread.pfault_wait = -1;
 716                 }
 717         } else {
 718                 /* signal bit not set -> a real page is missing. */
 719                 if (WARN_ON_ONCE(tsk != current))
 720                         goto out;
 721                 if (tsk->thread.pfault_wait == 1) {
 722                         /* Already on the list with a reference: put to sleep */
 723                         __set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 724                         set_tsk_need_resched(tsk);
 725                 } else if (tsk->thread.pfault_wait == -1) {
 726                         /* Completion interrupt was faster than the initial
 727                          * interrupt (pfault_wait == -1). Set pfault_wait
 728                          * back to zero and exit. */
 729                         tsk->thread.pfault_wait = 0;
 730                 } else {
 731                         /* Initial interrupt arrived before completion
 732                          * interrupt. Let the task sleep.
 733                          * An extra task reference is needed since a different
 734                          * cpu may set the task state to TASK_RUNNING again
 735                          * before the scheduler is reached. */
 736                         get_task_struct(tsk);
 737                         tsk->thread.pfault_wait = 1;
 738                         list_add(&tsk->thread.list, &pfault_list);
 739                         __set_task_state(tsk, TASK_UNINTERRUPTIBLE);
 740                         set_tsk_need_resched(tsk);
 741                 }
 742         }
 743 out:
 744         spin_unlock(&pfault_lock);
 745         put_task_struct(tsk);
 746 }
 747
 748 static int pfault_cpu_notify(struct notifier_block *self, unsigned long action,
 749                              void *hcpu)
 750 {
 751         struct thread_struct *thread, *next;
 752         struct task_struct *tsk;
 753
 754         switch (action & ~CPU_TASKS_FROZEN) {
 755         case CPU_DEAD:
 756                 spin_lock_irq(&pfault_lock);
 757                 list_for_each_entry_safe(thread, next, &pfault_list, list) {
 758                         thread->pfault_wait = 0;
 759                         list_del(&thread->list);
 760                         tsk = container_of(thread, struct task_struct, thread);
 761                         wake_up_process(tsk);
 762                         put_task_struct(tsk);
 763                 }
 764                 spin_unlock_irq(&pfault_lock);
 765                 break;
 766         default:
 767                 break;
 768         }
 769         return NOTIFY_OK;
 770 }
 771
 772 static int __init pfault_irq_init(void)
 773 {
 774         int rc;
 775
 776         rc = register_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
 777         if (rc)
 778                 goto out_extint;
 779         rc = pfault_init() == 0 ? 0 : -EOPNOTSUPP;
 780         if (rc)
 781                 goto out_pfault;
 782         irq_subclass_register(IRQ_SUBCLASS_SERVICE_SIGNAL);
 783         hotcpu_notifier(pfault_cpu_notify, 0);
 784         return 0;
 785
 786 out_pfault:
 787         unregister_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
 788 out_extint:
 789         pfault_disable = 1;
 790         return rc;
 791 }
 792 early_initcall(pfault_irq_init);
 793
 794 #endif /* CONFIG_PFAULT */