arch/s390/mm/fault.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  *  S390 version
   4  *    Copyright IBM Corp. 1999
   5  *    Author(s): Hartmut Penner (hp@de.ibm.com)
   6  *               Ulrich Weigand (uweigand@de.ibm.com)
   7  *
   8  *  Derived from "arch/i386/mm/fault.c"
   9  *    Copyright (C) 1995  Linus Torvalds
  10  */
  11
  12 #include <linux/kernel_stat.h>
  13 #include <linux/perf_event.h>
  14 #include <linux/signal.h>
  15 #include <linux/sched.h>
  16 #include <linux/sched/debug.h>
  17 #include <linux/kernel.h>
  18 #include <linux/errno.h>
  19 #include <linux/string.h>
  20 #include <linux/types.h>
  21 #include <linux/ptrace.h>
  22 #include <linux/mman.h>
  23 #include <linux/mm.h>
  24 #include <linux/compat.h>
  25 #include <linux/smp.h>
  26 #include <linux/kdebug.h>
  27 #include <linux/init.h>
  28 #include <linux/console.h>
  29 #include <linux/extable.h>
  30 #include <linux/hardirq.h>
  31 #include <linux/kprobes.h>
  32 #include <linux/uaccess.h>
  33 #include <linux/hugetlb.h>
  34 #include <asm/asm-offsets.h>
  35 #include <asm/diag.h>
  36 #include <asm/pgtable.h>
  37 #include <asm/gmap.h>
  38 #include <asm/irq.h>
  39 #include <asm/mmu_context.h>
  40 #include <asm/facility.h>
  41 #include "../kernel/entry.h"
  42
  43 #define __FAIL_ADDR_MASK -4096L
  44 #define __SUBCODE_MASK 0x0600
  45 #define __PF_RES_FIELD 0x8000000000000000ULL
  46
  47 #define VM_FAULT_BADCONTEXT     0x010000
  48 #define VM_FAULT_BADMAP         0x020000
  49 #define VM_FAULT_BADACCESS      0x040000
  50 #define VM_FAULT_SIGNAL         0x080000
  51 #define VM_FAULT_PFAULT         0x100000
  52
  53 enum fault_type {
  54         KERNEL_FAULT,
  55         USER_FAULT,
  56         VDSO_FAULT,
  57         GMAP_FAULT,
  58 };
  59
  60 static unsigned long store_indication __read_mostly;
  61
  62 static int __init fault_init(void)
  63 {
  64         if (test_facility(75))
  65                 store_indication = 0xc00;
  66         return 0;
  67 }
  68 early_initcall(fault_init);
  69
  70 static inline int notify_page_fault(struct pt_regs *regs)
  71 {
  72         int ret = 0;
  73
  74         /* kprobe_running() needs smp_processor_id() */
  75         if (kprobes_built_in() && !user_mode(regs)) {
  76                 preempt_disable();
  77                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  78                         ret = 1;
  79                 preempt_enable();
  80         }
  81         return ret;
  82 }
  83
  84
  85 /*
  86  * Unlock any spinlocks which will prevent us from getting the
  87  * message out.
  88  */
  89 void bust_spinlocks(int yes)
  90 {
  91         if (yes) {
  92                 oops_in_progress = 1;
  93         } else {
  94                 int loglevel_save = console_loglevel;
  95                 console_unblank();
  96                 oops_in_progress = 0;
  97                 /*
  98                  * OK, the message is on the console.  Now we call printk()
  99                  * without oops_in_progress set so that printk will give klogd
 100                  * a poke.  Hold onto your hats...
 101                  */
 102                 console_loglevel = 15;
 103                 printk(" ");
 104                 console_loglevel = loglevel_save;
 105         }
 106 }
 107
 108 /*
 109  * Find out which address space caused the exception.
 110  * Access register mode is impossible, ignore space == 3.
 111  */
 112 static inline enum fault_type get_fault_type(struct pt_regs *regs)
 113 {
 114         unsigned long trans_exc_code;
 115
 116         trans_exc_code = regs->int_parm_long & 3;
 117         if (likely(trans_exc_code == 0)) {
 118                 /* primary space exception */
 119                 if (IS_ENABLED(CONFIG_PGSTE) &&
 120                     test_pt_regs_flag(regs, PIF_GUEST_FAULT))
 121                         return GMAP_FAULT;
 122                 if (current->thread.mm_segment == USER_DS)
 123                         return USER_FAULT;
 124                 return KERNEL_FAULT;
 125         }
 126         if (trans_exc_code == 2) {
 127                 /* secondary space exception */
 128                 if (current->thread.mm_segment & 1) {
 129                         if (current->thread.mm_segment == USER_DS_SACF)
 130                                 return USER_FAULT;
 131                         return KERNEL_FAULT;
 132                 }
 133                 return VDSO_FAULT;
 134         }
 135         /* home space exception -> access via kernel ASCE */
 136         return KERNEL_FAULT;
 137 }
 138
 139 static int bad_address(void *p)
 140 {
 141         unsigned long dummy;
 142
 143         return probe_kernel_address((unsigned long *)p, dummy);
 144 }
 145
 146 static void dump_pagetable(unsigned long asce, unsigned long address)
 147 {
 148         unsigned long *table = __va(asce & _ASCE_ORIGIN);
 149
 150         pr_alert("AS:%016lx ", asce);
 151         switch (asce & _ASCE_TYPE_MASK) {
 152         case _ASCE_TYPE_REGION1:
 153                 table += (address & _REGION1_INDEX) >> _REGION1_SHIFT;
 154                 if (bad_address(table))
 155                         goto bad;
 156                 pr_cont("R1:%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_REGION2:
 162                 table += (address & _REGION2_INDEX) >> _REGION2_SHIFT;
 163                 if (bad_address(table))
 164                         goto bad;
 165                 pr_cont("R2:%016lx ", *table);
 166                 if (*table & _REGION_ENTRY_INVALID)
 167                         goto out;
 168                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 169                 /* fallthrough */
 170         case _ASCE_TYPE_REGION3:
 171                 table += (address & _REGION3_INDEX) >> _REGION3_SHIFT;
 172                 if (bad_address(table))
 173                         goto bad;
 174                 pr_cont("R3:%016lx ", *table);
 175                 if (*table & (_REGION_ENTRY_INVALID | _REGION3_ENTRY_LARGE))
 176                         goto out;
 177                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 178                 /* fallthrough */
 179         case _ASCE_TYPE_SEGMENT:
 180                 table += (address & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
 181                 if (bad_address(table))
 182                         goto bad;
 183                 pr_cont("S:%016lx ", *table);
 184                 if (*table & (_SEGMENT_ENTRY_INVALID | _SEGMENT_ENTRY_LARGE))
 185                         goto out;
 186                 table = (unsigned long *)(*table & _SEGMENT_ENTRY_ORIGIN);
 187         }
 188         table += (address & _PAGE_INDEX) >> _PAGE_SHIFT;
 189         if (bad_address(table))
 190                 goto bad;
 191         pr_cont("P:%016lx ", *table);
 192 out:
 193         pr_cont("\n");
 194         return;
 195 bad:
 196         pr_cont("BAD\n");
 197 }
 198
 199 static void dump_fault_info(struct pt_regs *regs)
 200 {
 201         unsigned long asce;
 202
 203         pr_alert("Failing address: %016lx TEID: %016lx\n",
 204                  regs->int_parm_long & __FAIL_ADDR_MASK, regs->int_parm_long);
 205         pr_alert("Fault in ");
 206         switch (regs->int_parm_long & 3) {
 207         case 3:
 208                 pr_cont("home space ");
 209                 break;
 210         case 2:
 211                 pr_cont("secondary space ");
 212                 break;
 213         case 1:
 214                 pr_cont("access register ");
 215                 break;
 216         case 0:
 217                 pr_cont("primary space ");
 218                 break;
 219         }
 220         pr_cont("mode while using ");
 221         switch (get_fault_type(regs)) {
 222         case USER_FAULT:
 223                 asce = S390_lowcore.user_asce;
 224                 pr_cont("user ");
 225                 break;
 226         case VDSO_FAULT:
 227                 asce = S390_lowcore.vdso_asce;
 228                 pr_cont("vdso ");
 229                 break;
 230         case GMAP_FAULT:
 231                 asce = ((struct gmap *) S390_lowcore.gmap)->asce;
 232                 pr_cont("gmap ");
 233                 break;
 234         case KERNEL_FAULT:
 235                 asce = S390_lowcore.kernel_asce;
 236                 pr_cont("kernel ");
 237                 break;
 238         }
 239         pr_cont("ASCE.\n");
 240         dump_pagetable(asce, regs->int_parm_long & __FAIL_ADDR_MASK);
 241 }
 242
 243 int show_unhandled_signals = 1;
 244
 245 void report_user_fault(struct pt_regs *regs, long signr, int is_mm_fault)
 246 {
 247         if ((task_pid_nr(current) > 1) && !show_unhandled_signals)
 248                 return;
 249         if (!unhandled_signal(current, signr))
 250                 return;
 251         if (!printk_ratelimit())
 252                 return;
 253         printk(KERN_ALERT "User process fault: interruption code %04x ilc:%d ",
 254                regs->int_code & 0xffff, regs->int_code >> 17);
 255         print_vma_addr(KERN_CONT "in ", regs->psw.addr);
 256         printk(KERN_CONT "\n");
 257         if (is_mm_fault)
 258                 dump_fault_info(regs);
 259         show_regs(regs);
 260 }
 261
 262 /*
 263  * Send SIGSEGV to task.  This is an external routine
 264  * to keep the stack usage of do_page_fault small.
 265  */
 266 static noinline void do_sigsegv(struct pt_regs *regs, int si_code)
 267 {
 268         report_user_fault(regs, SIGSEGV, 1);
 269         force_sig_fault(SIGSEGV, si_code,
 270                         (void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK),
 271                         current);
 272 }
 273
 274 static noinline void do_no_context(struct pt_regs *regs)
 275 {
 276         const struct exception_table_entry *fixup;
 277
 278         /* Are we prepared to handle this kernel fault?  */
 279         fixup = search_exception_tables(regs->psw.addr);
 280         if (fixup) {
 281                 regs->psw.addr = extable_fixup(fixup);
 282                 return;
 283         }
 284
 285         /*
 286          * Oops. The kernel tried to access some bad page. We'll have to
 287          * terminate things with extreme prejudice.
 288          */
 289         if (get_fault_type(regs) == KERNEL_FAULT)
 290                 printk(KERN_ALERT "Unable to handle kernel pointer dereference"
 291                        " in virtual kernel address space\n");
 292         else
 293                 printk(KERN_ALERT "Unable to handle kernel paging request"
 294                        " in virtual user address space\n");
 295         dump_fault_info(regs);
 296         die(regs, "Oops");
 297         do_exit(SIGKILL);
 298 }
 299
 300 static noinline void do_low_address(struct pt_regs *regs)
 301 {
 302         /* Low-address protection hit in kernel mode means
 303            NULL pointer write access in kernel mode.  */
 304         if (regs->psw.mask & PSW_MASK_PSTATE) {
 305                 /* Low-address protection hit in user mode 'cannot happen'. */
 306                 die (regs, "Low-address protection");
 307                 do_exit(SIGKILL);
 308         }
 309
 310         do_no_context(regs);
 311 }
 312
 313 static noinline void do_sigbus(struct pt_regs *regs)
 314 {
 315         /*
 316          * Send a sigbus, regardless of whether we were in kernel
 317          * or user mode.
 318          */
 319         force_sig_fault(SIGBUS, BUS_ADRERR,
 320                         (void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK),
 321                         current);
 322 }
 323
 324 static noinline int signal_return(struct pt_regs *regs)
 325 {
 326         u16 instruction;
 327         int rc;
 328
 329         rc = __get_user(instruction, (u16 __user *) regs->psw.addr);
 330         if (rc)
 331                 return rc;
 332         if (instruction == 0x0a77) {
 333                 set_pt_regs_flag(regs, PIF_SYSCALL);
 334                 regs->int_code = 0x00040077;
 335                 return 0;
 336         } else if (instruction == 0x0aad) {
 337                 set_pt_regs_flag(regs, PIF_SYSCALL);
 338                 regs->int_code = 0x000400ad;
 339                 return 0;
 340         }
 341         return -EACCES;
 342 }
 343
 344 static noinline void do_fault_error(struct pt_regs *regs, int access,
 345                                         vm_fault_t fault)
 346 {
 347         int si_code;
 348
 349         switch (fault) {
 350         case VM_FAULT_BADACCESS:
 351                 if (access == VM_EXEC && signal_return(regs) == 0)
 352                         break;
 353         case VM_FAULT_BADMAP:
 354                 /* Bad memory access. Check if it is kernel or user space. */
 355                 if (user_mode(regs)) {
 356                         /* User mode accesses just cause a SIGSEGV */
 357                         si_code = (fault == VM_FAULT_BADMAP) ?
 358                                 SEGV_MAPERR : SEGV_ACCERR;
 359                         do_sigsegv(regs, si_code);
 360                         break;
 361                 }
 362         case VM_FAULT_BADCONTEXT:
 363         case VM_FAULT_PFAULT:
 364                 do_no_context(regs);
 365                 break;
 366         case VM_FAULT_SIGNAL:
 367                 if (!user_mode(regs))
 368                         do_no_context(regs);
 369                 break;
 370         default: /* fault & VM_FAULT_ERROR */
 371                 if (fault & VM_FAULT_OOM) {
 372                         if (!user_mode(regs))
 373                                 do_no_context(regs);
 374                         else
 375                                 pagefault_out_of_memory();
 376                 } else if (fault & VM_FAULT_SIGSEGV) {
 377                         /* Kernel mode? Handle exceptions or die */
 378                         if (!user_mode(regs))
 379                                 do_no_context(regs);
 380                         else
 381                                 do_sigsegv(regs, SEGV_MAPERR);
 382                 } else if (fault & VM_FAULT_SIGBUS) {
 383                         /* Kernel mode? Handle exceptions or die */
 384                         if (!user_mode(regs))
 385                                 do_no_context(regs);
 386                         else
 387                                 do_sigbus(regs);
 388                 } else
 389                         BUG();
 390                 break;
 391         }
 392 }
 393
 394 /*
 395  * This routine handles page faults.  It determines the address,
 396  * and the problem, and then passes it off to one of the appropriate
 397  * routines.
 398  *
 399  * interruption code (int_code):
 400  *   04       Protection           ->  Write-Protection  (suprression)
 401  *   10       Segment translation  ->  Not present       (nullification)
 402  *   11       Page translation     ->  Not present       (nullification)
 403  *   3b       Region third trans.  ->  Not present       (nullification)
 404  */
 405 static inline vm_fault_t do_exception(struct pt_regs *regs, int access)
 406 {
 407         struct gmap *gmap;
 408         struct task_struct *tsk;
 409         struct mm_struct *mm;
 410         struct vm_area_struct *vma;
 411         enum fault_type type;
 412         unsigned long trans_exc_code;
 413         unsigned long address;
 414         unsigned int flags;
 415         vm_fault_t fault;
 416
 417         tsk = current;
 418         /*
 419          * The instruction that caused the program check has
 420          * been nullified. Don't signal single step via SIGTRAP.
 421          */
 422         clear_pt_regs_flag(regs, PIF_PER_TRAP);
 423
 424         if (notify_page_fault(regs))
 425                 return 0;
 426
 427         mm = tsk->mm;
 428         trans_exc_code = regs->int_parm_long;
 429
 430         /*
 431          * Verify that the fault happened in user space, that
 432          * we are not in an interrupt and that there is a
 433          * user context.
 434          */
 435         fault = VM_FAULT_BADCONTEXT;
 436         type = get_fault_type(regs);
 437         switch (type) {
 438         case KERNEL_FAULT:
 439                 goto out;
 440         case VDSO_FAULT:
 441                 fault = VM_FAULT_BADMAP;
 442                 goto out;
 443         case USER_FAULT:
 444         case GMAP_FAULT:
 445                 if (faulthandler_disabled() || !mm)
 446                         goto out;
 447                 break;
 448         }
 449
 450         address = trans_exc_code & __FAIL_ADDR_MASK;
 451         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
 452         flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
 453         if (user_mode(regs))
 454                 flags |= FAULT_FLAG_USER;
 455         if (access == VM_WRITE || (trans_exc_code & store_indication) == 0x400)
 456                 flags |= FAULT_FLAG_WRITE;
 457         down_read(&mm->mmap_sem);
 458
 459         gmap = NULL;
 460         if (IS_ENABLED(CONFIG_PGSTE) && type == GMAP_FAULT) {
 461                 gmap = (struct gmap *) S390_lowcore.gmap;
 462                 current->thread.gmap_addr = address;
 463                 current->thread.gmap_write_flag = !!(flags & FAULT_FLAG_WRITE);
 464                 current->thread.gmap_int_code = regs->int_code & 0xffff;
 465                 address = __gmap_translate(gmap, address);
 466                 if (address == -EFAULT) {
 467                         fault = VM_FAULT_BADMAP;
 468                         goto out_up;
 469                 }
 470                 if (gmap->pfault_enabled)
 471                         flags |= FAULT_FLAG_RETRY_NOWAIT;
 472         }
 473
 474 retry:
 475         fault = VM_FAULT_BADMAP;
 476         vma = find_vma(mm, address);
 477         if (!vma)
 478                 goto out_up;
 479
 480         if (unlikely(vma->vm_start > address)) {
 481                 if (!(vma->vm_flags & VM_GROWSDOWN))
 482                         goto out_up;
 483                 if (expand_stack(vma, address))
 484                         goto out_up;
 485         }
 486
 487         /*
 488          * Ok, we have a good vm_area for this memory access, so
 489          * we can handle it..
 490          */
 491         fault = VM_FAULT_BADACCESS;
 492         if (unlikely(!(vma->vm_flags & access)))
 493                 goto out_up;
 494
 495         if (is_vm_hugetlb_page(vma))
 496                 address &= HPAGE_MASK;
 497         /*
 498          * If for any reason at all we couldn't handle the fault,
 499          * make sure we exit gracefully rather than endlessly redo
 500          * the fault.
 501          */
 502         fault = handle_mm_fault(vma, address, flags);
 503         /* No reason to continue if interrupted by SIGKILL. */
 504         if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)) {
 505                 fault = VM_FAULT_SIGNAL;
 506                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
 507                         goto out_up;
 508                 goto out;
 509         }
 510         if (unlikely(fault & VM_FAULT_ERROR))
 511                 goto out_up;
 512
 513         /*
 514          * Major/minor page fault accounting is only done on the
 515          * initial attempt. If we go through a retry, it is extremely
 516          * likely that the page will be found in page cache at that point.
 517          */
 518         if (flags & FAULT_FLAG_ALLOW_RETRY) {
 519                 if (fault & VM_FAULT_MAJOR) {
 520                         tsk->maj_flt++;
 521                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
 522                                       regs, address);
 523                 } else {
 524                         tsk->min_flt++;
 525                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
 526                                       regs, address);
 527                 }
 528                 if (fault & VM_FAULT_RETRY) {
 529                         if (IS_ENABLED(CONFIG_PGSTE) && gmap &&
 530                             (flags & FAULT_FLAG_RETRY_NOWAIT)) {
 531                                 /* FAULT_FLAG_RETRY_NOWAIT has been set,
 532                                  * mmap_sem has not been released */
 533                                 current->thread.gmap_pfault = 1;
 534                                 fault = VM_FAULT_PFAULT;
 535                                 goto out_up;
 536                         }
 537                         /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
 538                          * of starvation. */
 539                         flags &= ~(FAULT_FLAG_ALLOW_RETRY |
 540                                    FAULT_FLAG_RETRY_NOWAIT);
 541                         flags |= FAULT_FLAG_TRIED;
 542                         down_read(&mm->mmap_sem);
 543                         goto retry;
 544                 }
 545         }
 546         if (IS_ENABLED(CONFIG_PGSTE) && gmap) {
 547                 address =  __gmap_link(gmap, current->thread.gmap_addr,
 548                                        address);
 549                 if (address == -EFAULT) {
 550                         fault = VM_FAULT_BADMAP;
 551                         goto out_up;
 552                 }
 553                 if (address == -ENOMEM) {
 554                         fault = VM_FAULT_OOM;
 555                         goto out_up;
 556                 }
 557         }
 558         fault = 0;
 559 out_up:
 560         up_read(&mm->mmap_sem);
 561 out:
 562         return fault;
 563 }
 564
 565 void do_protection_exception(struct pt_regs *regs)
 566 {
 567         unsigned long trans_exc_code;
 568         int access;
 569         vm_fault_t fault;
 570
 571         trans_exc_code = regs->int_parm_long;
 572         /*
 573          * Protection exceptions are suppressing, decrement psw address.
 574          * The exception to this rule are aborted transactions, for these
 575          * the PSW already points to the correct location.
 576          */
 577         if (!(regs->int_code & 0x200))
 578                 regs->psw.addr = __rewind_psw(regs->psw, regs->int_code >> 16);
 579         /*
 580          * Check for low-address protection.  This needs to be treated
 581          * as a special case because the translation exception code
 582          * field is not guaranteed to contain valid data in this case.
 583          */
 584         if (unlikely(!(trans_exc_code & 4))) {
 585                 do_low_address(regs);
 586                 return;
 587         }
 588         if (unlikely(MACHINE_HAS_NX && (trans_exc_code & 0x80))) {
 589                 regs->int_parm_long = (trans_exc_code & ~PAGE_MASK) |
 590                                         (regs->psw.addr & PAGE_MASK);
 591                 access = VM_EXEC;
 592                 fault = VM_FAULT_BADACCESS;
 593         } else {
 594                 access = VM_WRITE;
 595                 fault = do_exception(regs, access);
 596         }
 597         if (unlikely(fault))
 598                 do_fault_error(regs, access, fault);
 599 }
 600 NOKPROBE_SYMBOL(do_protection_exception);
 601
 602 void do_dat_exception(struct pt_regs *regs)
 603 {
 604         int access;
 605         vm_fault_t fault;
 606
 607         access = VM_READ | VM_EXEC | VM_WRITE;
 608         fault = do_exception(regs, access);
 609         if (unlikely(fault))
 610                 do_fault_error(regs, access, fault);
 611 }
 612 NOKPROBE_SYMBOL(do_dat_exception);
 613
 614 #ifdef CONFIG_PFAULT
 615 /*
 616  * 'pfault' pseudo page faults routines.
 617  */
 618 static int pfault_disable;
 619
 620 static int __init nopfault(char *str)
 621 {
 622         pfault_disable = 1;
 623         return 1;
 624 }
 625
 626 __setup("nopfault", nopfault);
 627
 628 struct pfault_refbk {
 629         u16 refdiagc;
 630         u16 reffcode;
 631         u16 refdwlen;
 632         u16 refversn;
 633         u64 refgaddr;
 634         u64 refselmk;
 635         u64 refcmpmk;
 636         u64 reserved;
 637 } __attribute__ ((packed, aligned(8)));
 638
 639 static struct pfault_refbk pfault_init_refbk = {
 640         .refdiagc = 0x258,
 641         .reffcode = 0,
 642         .refdwlen = 5,
 643         .refversn = 2,
 644         .refgaddr = __LC_LPP,
 645         .refselmk = 1ULL << 48,
 646         .refcmpmk = 1ULL << 48,
 647         .reserved = __PF_RES_FIELD
 648 };
 649
 650 int pfault_init(void)
 651 {
 652         int rc;
 653
 654         if (pfault_disable)
 655                 return -1;
 656         diag_stat_inc(DIAG_STAT_X258);
 657         asm volatile(
 658                 "       diag    %1,%0,0x258\n"
 659                 "0:     j       2f\n"
 660                 "1:     la      %0,8\n"
 661                 "2:\n"
 662                 EX_TABLE(0b,1b)
 663                 : "=d" (rc)
 664                 : "a" (&pfault_init_refbk), "m" (pfault_init_refbk) : "cc");
 665         return rc;
 666 }
 667
 668 static struct pfault_refbk pfault_fini_refbk = {
 669         .refdiagc = 0x258,
 670         .reffcode = 1,
 671         .refdwlen = 5,
 672         .refversn = 2,
 673 };
 674
 675 void pfault_fini(void)
 676 {
 677
 678         if (pfault_disable)
 679                 return;
 680         diag_stat_inc(DIAG_STAT_X258);
 681         asm volatile(
 682                 "       diag    %0,0,0x258\n"
 683                 "0:     nopr    %%r7\n"
 684                 EX_TABLE(0b,0b)
 685                 : : "a" (&pfault_fini_refbk), "m" (pfault_fini_refbk) : "cc");
 686 }
 687
 688 static DEFINE_SPINLOCK(pfault_lock);
 689 static LIST_HEAD(pfault_list);
 690
 691 #define PF_COMPLETE     0x0080
 692
 693 /*
 694  * The mechanism of our pfault code: if Linux is running as guest, runs a user
 695  * space process and the user space process accesses a page that the host has
 696  * paged out we get a pfault interrupt.
 697  *
 698  * This allows us, within the guest, to schedule a different process. Without
 699  * this mechanism the host would have to suspend the whole virtual cpu until
 700  * the page has been paged in.
 701  *
 702  * So when we get such an interrupt then we set the state of the current task
 703  * to uninterruptible and also set the need_resched flag. Both happens within
 704  * interrupt context(!). If we later on want to return to user space we
 705  * recognize the need_resched flag and then call schedule().  It's not very
 706  * obvious how this works...
 707  *
 708  * Of course we have a lot of additional fun with the completion interrupt (->
 709  * host signals that a page of a process has been paged in and the process can
 710  * continue to run). This interrupt can arrive on any cpu and, since we have
 711  * virtual cpus, actually appear before the interrupt that signals that a page
 712  * is missing.
 713  */
 714 static void pfault_interrupt(struct ext_code ext_code,
 715                              unsigned int param32, unsigned long param64)
 716 {
 717         struct task_struct *tsk;
 718         __u16 subcode;
 719         pid_t pid;
 720
 721         /*
 722          * Get the external interruption subcode & pfault initial/completion
 723          * signal bit. VM stores this in the 'cpu address' field associated
 724          * with the external interrupt.
 725          */
 726         subcode = ext_code.subcode;
 727         if ((subcode & 0xff00) != __SUBCODE_MASK)
 728                 return;
 729         inc_irq_stat(IRQEXT_PFL);
 730         /* Get the token (= pid of the affected task). */
 731         pid = param64 & LPP_PID_MASK;
 732         rcu_read_lock();
 733         tsk = find_task_by_pid_ns(pid, &init_pid_ns);
 734         if (tsk)
 735                 get_task_struct(tsk);
 736         rcu_read_unlock();
 737         if (!tsk)
 738                 return;
 739         spin_lock(&pfault_lock);
 740         if (subcode & PF_COMPLETE) {
 741                 /* signal bit is set -> a page has been swapped in by VM */
 742                 if (tsk->thread.pfault_wait == 1) {
 743                         /* Initial interrupt was faster than the completion
 744                          * interrupt. pfault_wait is valid. Set pfault_wait
 745                          * back to zero and wake up the process. This can
 746                          * safely be done because the task is still sleeping
 747                          * and can't produce new pfaults. */
 748                         tsk->thread.pfault_wait = 0;
 749                         list_del(&tsk->thread.list);
 750                         wake_up_process(tsk);
 751                         put_task_struct(tsk);
 752                 } else {
 753                         /* Completion interrupt was faster than initial
 754                          * interrupt. Set pfault_wait to -1 so the initial
 755                          * interrupt doesn't put the task to sleep.
 756                          * If the task is not running, ignore the completion
 757                          * interrupt since it must be a leftover of a PFAULT
 758                          * CANCEL operation which didn't remove all pending
 759                          * completion interrupts. */
 760                         if (tsk->state == TASK_RUNNING)
 761                                 tsk->thread.pfault_wait = -1;
 762                 }
 763         } else {
 764                 /* signal bit not set -> a real page is missing. */
 765                 if (WARN_ON_ONCE(tsk != current))
 766                         goto out;
 767                 if (tsk->thread.pfault_wait == 1) {
 768                         /* Already on the list with a reference: put to sleep */
 769                         goto block;
 770                 } else if (tsk->thread.pfault_wait == -1) {
 771                         /* Completion interrupt was faster than the initial
 772                          * interrupt (pfault_wait == -1). Set pfault_wait
 773                          * back to zero and exit. */
 774                         tsk->thread.pfault_wait = 0;
 775                 } else {
 776                         /* Initial interrupt arrived before completion
 777                          * interrupt. Let the task sleep.
 778                          * An extra task reference is needed since a different
 779                          * cpu may set the task state to TASK_RUNNING again
 780                          * before the scheduler is reached. */
 781                         get_task_struct(tsk);
 782                         tsk->thread.pfault_wait = 1;
 783                         list_add(&tsk->thread.list, &pfault_list);
 784 block:
 785                         /* Since this must be a userspace fault, there
 786                          * is no kernel task state to trample. Rely on the
 787                          * return to userspace schedule() to block. */
 788                         __set_current_state(TASK_UNINTERRUPTIBLE);
 789                         set_tsk_need_resched(tsk);
 790                         set_preempt_need_resched();
 791                 }
 792         }
 793 out:
 794         spin_unlock(&pfault_lock);
 795         put_task_struct(tsk);
 796 }
 797
 798 static int pfault_cpu_dead(unsigned int cpu)
 799 {
 800         struct thread_struct *thread, *next;
 801         struct task_struct *tsk;
 802
 803         spin_lock_irq(&pfault_lock);
 804         list_for_each_entry_safe(thread, next, &pfault_list, list) {
 805                 thread->pfault_wait = 0;
 806                 list_del(&thread->list);
 807                 tsk = container_of(thread, struct task_struct, thread);
 808                 wake_up_process(tsk);
 809                 put_task_struct(tsk);
 810         }
 811         spin_unlock_irq(&pfault_lock);
 812         return 0;
 813 }
 814
 815 static int __init pfault_irq_init(void)
 816 {
 817         int rc;
 818
 819         rc = register_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
 820         if (rc)
 821                 goto out_extint;
 822         rc = pfault_init() == 0 ? 0 : -EOPNOTSUPP;
 823         if (rc)
 824                 goto out_pfault;
 825         irq_subclass_register(IRQ_SUBCLASS_SERVICE_SIGNAL);
 826         cpuhp_setup_state_nocalls(CPUHP_S390_PFAULT_DEAD, "s390/pfault:dead",
 827                                   NULL, pfault_cpu_dead);
 828         return 0;
 829
 830 out_pfault:
 831         unregister_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
 832 out_extint:
 833         pfault_disable = 1;
 834         return rc;
 835 }
 836 early_initcall(pfault_irq_init);
 837
 838 #endif /* CONFIG_PFAULT */