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  * Find out which address space caused the exception.
  86  * Access register mode is impossible, ignore space == 3.
  87  */
  88 static inline enum fault_type get_fault_type(struct pt_regs *regs)
  89 {
  90         unsigned long trans_exc_code;
  91
  92         trans_exc_code = regs->int_parm_long & 3;
  93         if (likely(trans_exc_code == 0)) {
  94                 /* primary space exception */
  95                 if (IS_ENABLED(CONFIG_PGSTE) &&
  96                     test_pt_regs_flag(regs, PIF_GUEST_FAULT))
  97                         return GMAP_FAULT;
  98                 if (current->thread.mm_segment == USER_DS)
  99                         return USER_FAULT;
 100                 return KERNEL_FAULT;
 101         }
 102         if (trans_exc_code == 2) {
 103                 /* secondary space exception */
 104                 if (current->thread.mm_segment & 1) {
 105                         if (current->thread.mm_segment == USER_DS_SACF)
 106                                 return USER_FAULT;
 107                         return KERNEL_FAULT;
 108                 }
 109                 return VDSO_FAULT;
 110         }
 111         /* home space exception -> access via kernel ASCE */
 112         return KERNEL_FAULT;
 113 }
 114
 115 static int bad_address(void *p)
 116 {
 117         unsigned long dummy;
 118
 119         return probe_kernel_address((unsigned long *)p, dummy);
 120 }
 121
 122 static void dump_pagetable(unsigned long asce, unsigned long address)
 123 {
 124         unsigned long *table = __va(asce & _ASCE_ORIGIN);
 125
 126         pr_alert("AS:%016lx ", asce);
 127         switch (asce & _ASCE_TYPE_MASK) {
 128         case _ASCE_TYPE_REGION1:
 129                 table += (address & _REGION1_INDEX) >> _REGION1_SHIFT;
 130                 if (bad_address(table))
 131                         goto bad;
 132                 pr_cont("R1:%016lx ", *table);
 133                 if (*table & _REGION_ENTRY_INVALID)
 134                         goto out;
 135                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 136                 /* fallthrough */
 137         case _ASCE_TYPE_REGION2:
 138                 table += (address & _REGION2_INDEX) >> _REGION2_SHIFT;
 139                 if (bad_address(table))
 140                         goto bad;
 141                 pr_cont("R2:%016lx ", *table);
 142                 if (*table & _REGION_ENTRY_INVALID)
 143                         goto out;
 144                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 145                 /* fallthrough */
 146         case _ASCE_TYPE_REGION3:
 147                 table += (address & _REGION3_INDEX) >> _REGION3_SHIFT;
 148                 if (bad_address(table))
 149                         goto bad;
 150                 pr_cont("R3:%016lx ", *table);
 151                 if (*table & (_REGION_ENTRY_INVALID | _REGION3_ENTRY_LARGE))
 152                         goto out;
 153                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 154                 /* fallthrough */
 155         case _ASCE_TYPE_SEGMENT:
 156                 table += (address & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
 157                 if (bad_address(table))
 158                         goto bad;
 159                 pr_cont("S:%016lx ", *table);
 160                 if (*table & (_SEGMENT_ENTRY_INVALID | _SEGMENT_ENTRY_LARGE))
 161                         goto out;
 162                 table = (unsigned long *)(*table & _SEGMENT_ENTRY_ORIGIN);
 163         }
 164         table += (address & _PAGE_INDEX) >> _PAGE_SHIFT;
 165         if (bad_address(table))
 166                 goto bad;
 167         pr_cont("P:%016lx ", *table);
 168 out:
 169         pr_cont("\n");
 170         return;
 171 bad:
 172         pr_cont("BAD\n");
 173 }
 174
 175 static void dump_fault_info(struct pt_regs *regs)
 176 {
 177         unsigned long asce;
 178
 179         pr_alert("Failing address: %016lx TEID: %016lx\n",
 180                  regs->int_parm_long & __FAIL_ADDR_MASK, regs->int_parm_long);
 181         pr_alert("Fault in ");
 182         switch (regs->int_parm_long & 3) {
 183         case 3:
 184                 pr_cont("home space ");
 185                 break;
 186         case 2:
 187                 pr_cont("secondary space ");
 188                 break;
 189         case 1:
 190                 pr_cont("access register ");
 191                 break;
 192         case 0:
 193                 pr_cont("primary space ");
 194                 break;
 195         }
 196         pr_cont("mode while using ");
 197         switch (get_fault_type(regs)) {
 198         case USER_FAULT:
 199                 asce = S390_lowcore.user_asce;
 200                 pr_cont("user ");
 201                 break;
 202         case VDSO_FAULT:
 203                 asce = S390_lowcore.vdso_asce;
 204                 pr_cont("vdso ");
 205                 break;
 206         case GMAP_FAULT:
 207                 asce = ((struct gmap *) S390_lowcore.gmap)->asce;
 208                 pr_cont("gmap ");
 209                 break;
 210         case KERNEL_FAULT:
 211                 asce = S390_lowcore.kernel_asce;
 212                 pr_cont("kernel ");
 213                 break;
 214         }
 215         pr_cont("ASCE.\n");
 216         dump_pagetable(asce, regs->int_parm_long & __FAIL_ADDR_MASK);
 217 }
 218
 219 int show_unhandled_signals = 1;
 220
 221 void report_user_fault(struct pt_regs *regs, long signr, int is_mm_fault)
 222 {
 223         if ((task_pid_nr(current) > 1) && !show_unhandled_signals)
 224                 return;
 225         if (!unhandled_signal(current, signr))
 226                 return;
 227         if (!printk_ratelimit())
 228                 return;
 229         printk(KERN_ALERT "User process fault: interruption code %04x ilc:%d ",
 230                regs->int_code & 0xffff, regs->int_code >> 17);
 231         print_vma_addr(KERN_CONT "in ", regs->psw.addr);
 232         printk(KERN_CONT "\n");
 233         if (is_mm_fault)
 234                 dump_fault_info(regs);
 235         show_regs(regs);
 236 }
 237
 238 /*
 239  * Send SIGSEGV to task.  This is an external routine
 240  * to keep the stack usage of do_page_fault small.
 241  */
 242 static noinline void do_sigsegv(struct pt_regs *regs, int si_code)
 243 {
 244         report_user_fault(regs, SIGSEGV, 1);
 245         force_sig_fault(SIGSEGV, si_code,
 246                         (void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK),
 247                         current);
 248 }
 249
 250 const struct exception_table_entry *s390_search_extables(unsigned long addr)
 251 {
 252         const struct exception_table_entry *fixup;
 253
 254         fixup = search_extable(__start_dma_ex_table,
 255                                __stop_dma_ex_table - __start_dma_ex_table,
 256                                addr);
 257         if (!fixup)
 258                 fixup = search_exception_tables(addr);
 259         return fixup;
 260 }
 261
 262 static noinline void do_no_context(struct pt_regs *regs)
 263 {
 264         const struct exception_table_entry *fixup;
 265
 266         /* Are we prepared to handle this kernel fault?  */
 267         fixup = s390_search_extables(regs->psw.addr);
 268         if (fixup) {
 269                 regs->psw.addr = extable_fixup(fixup);
 270                 return;
 271         }
 272
 273         /*
 274          * Oops. The kernel tried to access some bad page. We'll have to
 275          * terminate things with extreme prejudice.
 276          */
 277         if (get_fault_type(regs) == KERNEL_FAULT)
 278                 printk(KERN_ALERT "Unable to handle kernel pointer dereference"
 279                        " in virtual kernel address space\n");
 280         else
 281                 printk(KERN_ALERT "Unable to handle kernel paging request"
 282                        " in virtual user address space\n");
 283         dump_fault_info(regs);
 284         die(regs, "Oops");
 285         do_exit(SIGKILL);
 286 }
 287
 288 static noinline void do_low_address(struct pt_regs *regs)
 289 {
 290         /* Low-address protection hit in kernel mode means
 291            NULL pointer write access in kernel mode.  */
 292         if (regs->psw.mask & PSW_MASK_PSTATE) {
 293                 /* Low-address protection hit in user mode 'cannot happen'. */
 294                 die (regs, "Low-address protection");
 295                 do_exit(SIGKILL);
 296         }
 297
 298         do_no_context(regs);
 299 }
 300
 301 static noinline void do_sigbus(struct pt_regs *regs)
 302 {
 303         /*
 304          * Send a sigbus, regardless of whether we were in kernel
 305          * or user mode.
 306          */
 307         force_sig_fault(SIGBUS, BUS_ADRERR,
 308                         (void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK),
 309                         current);
 310 }
 311
 312 static noinline int signal_return(struct pt_regs *regs)
 313 {
 314         u16 instruction;
 315         int rc;
 316
 317         rc = __get_user(instruction, (u16 __user *) regs->psw.addr);
 318         if (rc)
 319                 return rc;
 320         if (instruction == 0x0a77) {
 321                 set_pt_regs_flag(regs, PIF_SYSCALL);
 322                 regs->int_code = 0x00040077;
 323                 return 0;
 324         } else if (instruction == 0x0aad) {
 325                 set_pt_regs_flag(regs, PIF_SYSCALL);
 326                 regs->int_code = 0x000400ad;
 327                 return 0;
 328         }
 329         return -EACCES;
 330 }
 331
 332 static noinline void do_fault_error(struct pt_regs *regs, int access,
 333                                         vm_fault_t fault)
 334 {
 335         int si_code;
 336
 337         switch (fault) {
 338         case VM_FAULT_BADACCESS:
 339                 if (access == VM_EXEC && signal_return(regs) == 0)
 340                         break;
 341         case VM_FAULT_BADMAP:
 342                 /* Bad memory access. Check if it is kernel or user space. */
 343                 if (user_mode(regs)) {
 344                         /* User mode accesses just cause a SIGSEGV */
 345                         si_code = (fault == VM_FAULT_BADMAP) ?
 346                                 SEGV_MAPERR : SEGV_ACCERR;
 347                         do_sigsegv(regs, si_code);
 348                         break;
 349                 }
 350         case VM_FAULT_BADCONTEXT:
 351         case VM_FAULT_PFAULT:
 352                 do_no_context(regs);
 353                 break;
 354         case VM_FAULT_SIGNAL:
 355                 if (!user_mode(regs))
 356                         do_no_context(regs);
 357                 break;
 358         default: /* fault & VM_FAULT_ERROR */
 359                 if (fault & VM_FAULT_OOM) {
 360                         if (!user_mode(regs))
 361                                 do_no_context(regs);
 362                         else
 363                                 pagefault_out_of_memory();
 364                 } else if (fault & VM_FAULT_SIGSEGV) {
 365                         /* Kernel mode? Handle exceptions or die */
 366                         if (!user_mode(regs))
 367                                 do_no_context(regs);
 368                         else
 369                                 do_sigsegv(regs, SEGV_MAPERR);
 370                 } else if (fault & VM_FAULT_SIGBUS) {
 371                         /* Kernel mode? Handle exceptions or die */
 372                         if (!user_mode(regs))
 373                                 do_no_context(regs);
 374                         else
 375                                 do_sigbus(regs);
 376                 } else
 377                         BUG();
 378                 break;
 379         }
 380 }
 381
 382 /*
 383  * This routine handles page faults.  It determines the address,
 384  * and the problem, and then passes it off to one of the appropriate
 385  * routines.
 386  *
 387  * interruption code (int_code):
 388  *   04       Protection           ->  Write-Protection  (suprression)
 389  *   10       Segment translation  ->  Not present       (nullification)
 390  *   11       Page translation     ->  Not present       (nullification)
 391  *   3b       Region third trans.  ->  Not present       (nullification)
 392  */
 393 static inline vm_fault_t do_exception(struct pt_regs *regs, int access)
 394 {
 395         struct gmap *gmap;
 396         struct task_struct *tsk;
 397         struct mm_struct *mm;
 398         struct vm_area_struct *vma;
 399         enum fault_type type;
 400         unsigned long trans_exc_code;
 401         unsigned long address;
 402         unsigned int flags;
 403         vm_fault_t fault;
 404
 405         tsk = current;
 406         /*
 407          * The instruction that caused the program check has
 408          * been nullified. Don't signal single step via SIGTRAP.
 409          */
 410         clear_pt_regs_flag(regs, PIF_PER_TRAP);
 411
 412         if (notify_page_fault(regs))
 413                 return 0;
 414
 415         mm = tsk->mm;
 416         trans_exc_code = regs->int_parm_long;
 417
 418         /*
 419          * Verify that the fault happened in user space, that
 420          * we are not in an interrupt and that there is a
 421          * user context.
 422          */
 423         fault = VM_FAULT_BADCONTEXT;
 424         type = get_fault_type(regs);
 425         switch (type) {
 426         case KERNEL_FAULT:
 427                 goto out;
 428         case VDSO_FAULT:
 429                 fault = VM_FAULT_BADMAP;
 430                 goto out;
 431         case USER_FAULT:
 432         case GMAP_FAULT:
 433                 if (faulthandler_disabled() || !mm)
 434                         goto out;
 435                 break;
 436         }
 437
 438         address = trans_exc_code & __FAIL_ADDR_MASK;
 439         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
 440         flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
 441         if (user_mode(regs))
 442                 flags |= FAULT_FLAG_USER;
 443         if (access == VM_WRITE || (trans_exc_code & store_indication) == 0x400)
 444                 flags |= FAULT_FLAG_WRITE;
 445         down_read(&mm->mmap_sem);
 446
 447         gmap = NULL;
 448         if (IS_ENABLED(CONFIG_PGSTE) && type == GMAP_FAULT) {
 449                 gmap = (struct gmap *) S390_lowcore.gmap;
 450                 current->thread.gmap_addr = address;
 451                 current->thread.gmap_write_flag = !!(flags & FAULT_FLAG_WRITE);
 452                 current->thread.gmap_int_code = regs->int_code & 0xffff;
 453                 address = __gmap_translate(gmap, address);
 454                 if (address == -EFAULT) {
 455                         fault = VM_FAULT_BADMAP;
 456                         goto out_up;
 457                 }
 458                 if (gmap->pfault_enabled)
 459                         flags |= FAULT_FLAG_RETRY_NOWAIT;
 460         }
 461
 462 retry:
 463         fault = VM_FAULT_BADMAP;
 464         vma = find_vma(mm, address);
 465         if (!vma)
 466                 goto out_up;
 467
 468         if (unlikely(vma->vm_start > address)) {
 469                 if (!(vma->vm_flags & VM_GROWSDOWN))
 470                         goto out_up;
 471                 if (expand_stack(vma, address))
 472                         goto out_up;
 473         }
 474
 475         /*
 476          * Ok, we have a good vm_area for this memory access, so
 477          * we can handle it..
 478          */
 479         fault = VM_FAULT_BADACCESS;
 480         if (unlikely(!(vma->vm_flags & access)))
 481                 goto out_up;
 482
 483         if (is_vm_hugetlb_page(vma))
 484                 address &= HPAGE_MASK;
 485         /*
 486          * If for any reason at all we couldn't handle the fault,
 487          * make sure we exit gracefully rather than endlessly redo
 488          * the fault.
 489          */
 490         fault = handle_mm_fault(vma, address, flags);
 491         /* No reason to continue if interrupted by SIGKILL. */
 492         if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)) {
 493                 fault = VM_FAULT_SIGNAL;
 494                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
 495                         goto out_up;
 496                 goto out;
 497         }
 498         if (unlikely(fault & VM_FAULT_ERROR))
 499                 goto out_up;
 500
 501         /*
 502          * Major/minor page fault accounting is only done on the
 503          * initial attempt. If we go through a retry, it is extremely
 504          * likely that the page will be found in page cache at that point.
 505          */
 506         if (flags & FAULT_FLAG_ALLOW_RETRY) {
 507                 if (fault & VM_FAULT_MAJOR) {
 508                         tsk->maj_flt++;
 509                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
 510                                       regs, address);
 511                 } else {
 512                         tsk->min_flt++;
 513                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
 514                                       regs, address);
 515                 }
 516                 if (fault & VM_FAULT_RETRY) {
 517                         if (IS_ENABLED(CONFIG_PGSTE) && gmap &&
 518                             (flags & FAULT_FLAG_RETRY_NOWAIT)) {
 519                                 /* FAULT_FLAG_RETRY_NOWAIT has been set,
 520                                  * mmap_sem has not been released */
 521                                 current->thread.gmap_pfault = 1;
 522                                 fault = VM_FAULT_PFAULT;
 523                                 goto out_up;
 524                         }
 525                         /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
 526                          * of starvation. */
 527                         flags &= ~(FAULT_FLAG_ALLOW_RETRY |
 528                                    FAULT_FLAG_RETRY_NOWAIT);
 529                         flags |= FAULT_FLAG_TRIED;
 530                         down_read(&mm->mmap_sem);
 531                         goto retry;
 532                 }
 533         }
 534         if (IS_ENABLED(CONFIG_PGSTE) && gmap) {
 535                 address =  __gmap_link(gmap, current->thread.gmap_addr,
 536                                        address);
 537                 if (address == -EFAULT) {
 538                         fault = VM_FAULT_BADMAP;
 539                         goto out_up;
 540                 }
 541                 if (address == -ENOMEM) {
 542                         fault = VM_FAULT_OOM;
 543                         goto out_up;
 544                 }
 545         }
 546         fault = 0;
 547 out_up:
 548         up_read(&mm->mmap_sem);
 549 out:
 550         return fault;
 551 }
 552
 553 void do_protection_exception(struct pt_regs *regs)
 554 {
 555         unsigned long trans_exc_code;
 556         int access;
 557         vm_fault_t fault;
 558
 559         trans_exc_code = regs->int_parm_long;
 560         /*
 561          * Protection exceptions are suppressing, decrement psw address.
 562          * The exception to this rule are aborted transactions, for these
 563          * the PSW already points to the correct location.
 564          */
 565         if (!(regs->int_code & 0x200))
 566                 regs->psw.addr = __rewind_psw(regs->psw, regs->int_code >> 16);
 567         /*
 568          * Check for low-address protection.  This needs to be treated
 569          * as a special case because the translation exception code
 570          * field is not guaranteed to contain valid data in this case.
 571          */
 572         if (unlikely(!(trans_exc_code & 4))) {
 573                 do_low_address(regs);
 574                 return;
 575         }
 576         if (unlikely(MACHINE_HAS_NX && (trans_exc_code & 0x80))) {
 577                 regs->int_parm_long = (trans_exc_code & ~PAGE_MASK) |
 578                                         (regs->psw.addr & PAGE_MASK);
 579                 access = VM_EXEC;
 580                 fault = VM_FAULT_BADACCESS;
 581         } else {
 582                 access = VM_WRITE;
 583                 fault = do_exception(regs, access);
 584         }
 585         if (unlikely(fault))
 586                 do_fault_error(regs, access, fault);
 587 }
 588 NOKPROBE_SYMBOL(do_protection_exception);
 589
 590 void do_dat_exception(struct pt_regs *regs)
 591 {
 592         int access;
 593         vm_fault_t fault;
 594
 595         access = VM_READ | VM_EXEC | VM_WRITE;
 596         fault = do_exception(regs, access);
 597         if (unlikely(fault))
 598                 do_fault_error(regs, access, fault);
 599 }
 600 NOKPROBE_SYMBOL(do_dat_exception);
 601
 602 #ifdef CONFIG_PFAULT
 603 /*
 604  * 'pfault' pseudo page faults routines.
 605  */
 606 static int pfault_disable;
 607
 608 static int __init nopfault(char *str)
 609 {
 610         pfault_disable = 1;
 611         return 1;
 612 }
 613
 614 __setup("nopfault", nopfault);
 615
 616 struct pfault_refbk {
 617         u16 refdiagc;
 618         u16 reffcode;
 619         u16 refdwlen;
 620         u16 refversn;
 621         u64 refgaddr;
 622         u64 refselmk;
 623         u64 refcmpmk;
 624         u64 reserved;
 625 } __attribute__ ((packed, aligned(8)));
 626
 627 static struct pfault_refbk pfault_init_refbk = {
 628         .refdiagc = 0x258,
 629         .reffcode = 0,
 630         .refdwlen = 5,
 631         .refversn = 2,
 632         .refgaddr = __LC_LPP,
 633         .refselmk = 1ULL << 48,
 634         .refcmpmk = 1ULL << 48,
 635         .reserved = __PF_RES_FIELD
 636 };
 637
 638 int pfault_init(void)
 639 {
 640         int rc;
 641
 642         if (pfault_disable)
 643                 return -1;
 644         diag_stat_inc(DIAG_STAT_X258);
 645         asm volatile(
 646                 "       diag    %1,%0,0x258\n"
 647                 "0:     j       2f\n"
 648                 "1:     la      %0,8\n"
 649                 "2:\n"
 650                 EX_TABLE(0b,1b)
 651                 : "=d" (rc)
 652                 : "a" (&pfault_init_refbk), "m" (pfault_init_refbk) : "cc");
 653         return rc;
 654 }
 655
 656 static struct pfault_refbk pfault_fini_refbk = {
 657         .refdiagc = 0x258,
 658         .reffcode = 1,
 659         .refdwlen = 5,
 660         .refversn = 2,
 661 };
 662
 663 void pfault_fini(void)
 664 {
 665
 666         if (pfault_disable)
 667                 return;
 668         diag_stat_inc(DIAG_STAT_X258);
 669         asm volatile(
 670                 "       diag    %0,0,0x258\n"
 671                 "0:     nopr    %%r7\n"
 672                 EX_TABLE(0b,0b)
 673                 : : "a" (&pfault_fini_refbk), "m" (pfault_fini_refbk) : "cc");
 674 }
 675
 676 static DEFINE_SPINLOCK(pfault_lock);
 677 static LIST_HEAD(pfault_list);
 678
 679 #define PF_COMPLETE     0x0080
 680
 681 /*
 682  * The mechanism of our pfault code: if Linux is running as guest, runs a user
 683  * space process and the user space process accesses a page that the host has
 684  * paged out we get a pfault interrupt.
 685  *
 686  * This allows us, within the guest, to schedule a different process. Without
 687  * this mechanism the host would have to suspend the whole virtual cpu until
 688  * the page has been paged in.
 689  *
 690  * So when we get such an interrupt then we set the state of the current task
 691  * to uninterruptible and also set the need_resched flag. Both happens within
 692  * interrupt context(!). If we later on want to return to user space we
 693  * recognize the need_resched flag and then call schedule().  It's not very
 694  * obvious how this works...
 695  *
 696  * Of course we have a lot of additional fun with the completion interrupt (->
 697  * host signals that a page of a process has been paged in and the process can
 698  * continue to run). This interrupt can arrive on any cpu and, since we have
 699  * virtual cpus, actually appear before the interrupt that signals that a page
 700  * is missing.
 701  */
 702 static void pfault_interrupt(struct ext_code ext_code,
 703                              unsigned int param32, unsigned long param64)
 704 {
 705         struct task_struct *tsk;
 706         __u16 subcode;
 707         pid_t pid;
 708
 709         /*
 710          * Get the external interruption subcode & pfault initial/completion
 711          * signal bit. VM stores this in the 'cpu address' field associated
 712          * with the external interrupt.
 713          */
 714         subcode = ext_code.subcode;
 715         if ((subcode & 0xff00) != __SUBCODE_MASK)
 716                 return;
 717         inc_irq_stat(IRQEXT_PFL);
 718         /* Get the token (= pid of the affected task). */
 719         pid = param64 & LPP_PID_MASK;
 720         rcu_read_lock();
 721         tsk = find_task_by_pid_ns(pid, &init_pid_ns);
 722         if (tsk)
 723                 get_task_struct(tsk);
 724         rcu_read_unlock();
 725         if (!tsk)
 726                 return;
 727         spin_lock(&pfault_lock);
 728         if (subcode & PF_COMPLETE) {
 729                 /* signal bit is set -> a page has been swapped in by VM */
 730                 if (tsk->thread.pfault_wait == 1) {
 731                         /* Initial interrupt was faster than the completion
 732                          * interrupt. pfault_wait is valid. Set pfault_wait
 733                          * back to zero and wake up the process. This can
 734                          * safely be done because the task is still sleeping
 735                          * and can't produce new pfaults. */
 736                         tsk->thread.pfault_wait = 0;
 737                         list_del(&tsk->thread.list);
 738                         wake_up_process(tsk);
 739                         put_task_struct(tsk);
 740                 } else {
 741                         /* Completion interrupt was faster than initial
 742                          * interrupt. Set pfault_wait to -1 so the initial
 743                          * interrupt doesn't put the task to sleep.
 744                          * If the task is not running, ignore the completion
 745                          * interrupt since it must be a leftover of a PFAULT
 746                          * CANCEL operation which didn't remove all pending
 747                          * completion interrupts. */
 748                         if (tsk->state == TASK_RUNNING)
 749                                 tsk->thread.pfault_wait = -1;
 750                 }
 751         } else {
 752                 /* signal bit not set -> a real page is missing. */
 753                 if (WARN_ON_ONCE(tsk != current))
 754                         goto out;
 755                 if (tsk->thread.pfault_wait == 1) {
 756                         /* Already on the list with a reference: put to sleep */
 757                         goto block;
 758                 } else if (tsk->thread.pfault_wait == -1) {
 759                         /* Completion interrupt was faster than the initial
 760                          * interrupt (pfault_wait == -1). Set pfault_wait
 761                          * back to zero and exit. */
 762                         tsk->thread.pfault_wait = 0;
 763                 } else {
 764                         /* Initial interrupt arrived before completion
 765                          * interrupt. Let the task sleep.
 766                          * An extra task reference is needed since a different
 767                          * cpu may set the task state to TASK_RUNNING again
 768                          * before the scheduler is reached. */
 769                         get_task_struct(tsk);
 770                         tsk->thread.pfault_wait = 1;
 771                         list_add(&tsk->thread.list, &pfault_list);
 772 block:
 773                         /* Since this must be a userspace fault, there
 774                          * is no kernel task state to trample. Rely on the
 775                          * return to userspace schedule() to block. */
 776                         __set_current_state(TASK_UNINTERRUPTIBLE);
 777                         set_tsk_need_resched(tsk);
 778                         set_preempt_need_resched();
 779                 }
 780         }
 781 out:
 782         spin_unlock(&pfault_lock);
 783         put_task_struct(tsk);
 784 }
 785
 786 static int pfault_cpu_dead(unsigned int cpu)
 787 {
 788         struct thread_struct *thread, *next;
 789         struct task_struct *tsk;
 790
 791         spin_lock_irq(&pfault_lock);
 792         list_for_each_entry_safe(thread, next, &pfault_list, list) {
 793                 thread->pfault_wait = 0;
 794                 list_del(&thread->list);
 795                 tsk = container_of(thread, struct task_struct, thread);
 796                 wake_up_process(tsk);
 797                 put_task_struct(tsk);
 798         }
 799         spin_unlock_irq(&pfault_lock);
 800         return 0;
 801 }
 802
 803 static int __init pfault_irq_init(void)
 804 {
 805         int rc;
 806
 807         rc = register_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
 808         if (rc)
 809                 goto out_extint;
 810         rc = pfault_init() == 0 ? 0 : -EOPNOTSUPP;
 811         if (rc)
 812                 goto out_pfault;
 813         irq_subclass_register(IRQ_SUBCLASS_SERVICE_SIGNAL);
 814         cpuhp_setup_state_nocalls(CPUHP_S390_PFAULT_DEAD, "s390/pfault:dead",
 815                                   NULL, pfault_cpu_dead);
 816         return 0;
 817
 818 out_pfault:
 819         unregister_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
 820 out_extint:
 821         pfault_disable = 1;
 822         return rc;
 823 }
 824 early_initcall(pfault_irq_init);
 825
 826 #endif /* CONFIG_PFAULT */