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 static unsigned long store_indication __read_mostly;
  54
  55 static int __init fault_init(void)
  56 {
  57         if (test_facility(75))
  58                 store_indication = 0xc00;
  59         return 0;
  60 }
  61 early_initcall(fault_init);
  62
  63 static inline int notify_page_fault(struct pt_regs *regs)
  64 {
  65         int ret = 0;
  66
  67         /* kprobe_running() needs smp_processor_id() */
  68         if (kprobes_built_in() && !user_mode(regs)) {
  69                 preempt_disable();
  70                 if (kprobe_running() && kprobe_fault_handler(regs, 14))
  71                         ret = 1;
  72                 preempt_enable();
  73         }
  74         return ret;
  75 }
  76
  77
  78 /*
  79  * Unlock any spinlocks which will prevent us from getting the
  80  * message out.
  81  */
  82 void bust_spinlocks(int yes)
  83 {
  84         if (yes) {
  85                 oops_in_progress = 1;
  86         } else {
  87                 int loglevel_save = console_loglevel;
  88                 console_unblank();
  89                 oops_in_progress = 0;
  90                 /*
  91                  * OK, the message is on the console.  Now we call printk()
  92                  * without oops_in_progress set so that printk will give klogd
  93                  * a poke.  Hold onto your hats...
  94                  */
  95                 console_loglevel = 15;
  96                 printk(" ");
  97                 console_loglevel = loglevel_save;
  98         }
  99 }
 100
 101 /*
 102  * Returns the address space associated with the fault.
 103  * Returns 0 for kernel space and 1 for user space.
 104  */
 105 static inline int user_space_fault(struct pt_regs *regs)
 106 {
 107         unsigned long trans_exc_code;
 108
 109         /*
 110          * The lowest two bits of the translation exception
 111          * identification indicate which paging table was used.
 112          */
 113         trans_exc_code = regs->int_parm_long & 3;
 114         if (trans_exc_code == 3) /* home space -> kernel */
 115                 return 0;
 116         if (user_mode(regs))
 117                 return 1;
 118         if (trans_exc_code == 2) /* secondary space -> set_fs */
 119                 return current->thread.mm_segment.ar4;
 120         if (current->flags & PF_VCPU)
 121                 return 1;
 122         return 0;
 123 }
 124
 125 static int bad_address(void *p)
 126 {
 127         unsigned long dummy;
 128
 129         return probe_kernel_address((unsigned long *)p, dummy);
 130 }
 131
 132 static void dump_pagetable(unsigned long asce, unsigned long address)
 133 {
 134         unsigned long *table = __va(asce & _ASCE_ORIGIN);
 135
 136         pr_alert("AS:%016lx ", asce);
 137         switch (asce & _ASCE_TYPE_MASK) {
 138         case _ASCE_TYPE_REGION1:
 139                 table += (address & _REGION1_INDEX) >> _REGION1_SHIFT;
 140                 if (bad_address(table))
 141                         goto bad;
 142                 pr_cont("R1:%016lx ", *table);
 143                 if (*table & _REGION_ENTRY_INVALID)
 144                         goto out;
 145                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 146                 /* fallthrough */
 147         case _ASCE_TYPE_REGION2:
 148                 table += (address & _REGION2_INDEX) >> _REGION2_SHIFT;
 149                 if (bad_address(table))
 150                         goto bad;
 151                 pr_cont("R2:%016lx ", *table);
 152                 if (*table & _REGION_ENTRY_INVALID)
 153                         goto out;
 154                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 155                 /* fallthrough */
 156         case _ASCE_TYPE_REGION3:
 157                 table += (address & _REGION3_INDEX) >> _REGION3_SHIFT;
 158                 if (bad_address(table))
 159                         goto bad;
 160                 pr_cont("R3:%016lx ", *table);
 161                 if (*table & (_REGION_ENTRY_INVALID | _REGION3_ENTRY_LARGE))
 162                         goto out;
 163                 table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
 164                 /* fallthrough */
 165         case _ASCE_TYPE_SEGMENT:
 166                 table += (address & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
 167                 if (bad_address(table))
 168                         goto bad;
 169                 pr_cont("S:%016lx ", *table);
 170                 if (*table & (_SEGMENT_ENTRY_INVALID | _SEGMENT_ENTRY_LARGE))
 171                         goto out;
 172                 table = (unsigned long *)(*table & _SEGMENT_ENTRY_ORIGIN);
 173         }
 174         table += (address & _PAGE_INDEX) >> _PAGE_SHIFT;
 175         if (bad_address(table))
 176                 goto bad;
 177         pr_cont("P:%016lx ", *table);
 178 out:
 179         pr_cont("\n");
 180         return;
 181 bad:
 182         pr_cont("BAD\n");
 183 }
 184
 185 static void dump_fault_info(struct pt_regs *regs)
 186 {
 187         unsigned long asce;
 188
 189         pr_alert("Failing address: %016lx TEID: %016lx\n",
 190                  regs->int_parm_long & __FAIL_ADDR_MASK, regs->int_parm_long);
 191         pr_alert("Fault in ");
 192         switch (regs->int_parm_long & 3) {
 193         case 3:
 194                 pr_cont("home space ");
 195                 break;
 196         case 2:
 197                 pr_cont("secondary space ");
 198                 break;
 199         case 1:
 200                 pr_cont("access register ");
 201                 break;
 202         case 0:
 203                 pr_cont("primary space ");
 204                 break;
 205         }
 206         pr_cont("mode while using ");
 207         if (!user_space_fault(regs)) {
 208                 asce = S390_lowcore.kernel_asce;
 209                 pr_cont("kernel ");
 210         }
 211 #ifdef CONFIG_PGSTE
 212         else if ((current->flags & PF_VCPU) && S390_lowcore.gmap) {
 213                 struct gmap *gmap = (struct gmap *)S390_lowcore.gmap;
 214                 asce = gmap->asce;
 215                 pr_cont("gmap ");
 216         }
 217 #endif
 218         else {
 219                 asce = S390_lowcore.user_asce;
 220                 pr_cont("user ");
 221         }
 222         pr_cont("ASCE.\n");
 223         dump_pagetable(asce, regs->int_parm_long & __FAIL_ADDR_MASK);
 224 }
 225
 226 int show_unhandled_signals = 1;
 227
 228 void report_user_fault(struct pt_regs *regs, long signr, int is_mm_fault)
 229 {
 230         if ((task_pid_nr(current) > 1) && !show_unhandled_signals)
 231                 return;
 232         if (!unhandled_signal(current, signr))
 233                 return;
 234         if (!printk_ratelimit())
 235                 return;
 236         printk(KERN_ALERT "User process fault: interruption code %04x ilc:%d ",
 237                regs->int_code & 0xffff, regs->int_code >> 17);
 238         print_vma_addr(KERN_CONT "in ", regs->psw.addr);
 239         printk(KERN_CONT "\n");
 240         if (is_mm_fault)
 241                 dump_fault_info(regs);
 242         show_regs(regs);
 243 }
 244
 245 /*
 246  * Send SIGSEGV to task.  This is an external routine
 247  * to keep the stack usage of do_page_fault small.
 248  */
 249 static noinline void do_sigsegv(struct pt_regs *regs, int si_code)
 250 {
 251         struct siginfo si;
 252
 253         report_user_fault(regs, SIGSEGV, 1);
 254         si.si_signo = SIGSEGV;
 255         si.si_errno = 0;
 256         si.si_code = si_code;
 257         si.si_addr = (void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK);
 258         force_sig_info(SIGSEGV, &si, current);
 259 }
 260
 261 static noinline void do_no_context(struct pt_regs *regs)
 262 {
 263         const struct exception_table_entry *fixup;
 264
 265         /* Are we prepared to handle this kernel fault?  */
 266         fixup = search_exception_tables(regs->psw.addr);
 267         if (fixup) {
 268                 regs->psw.addr = extable_fixup(fixup);
 269                 return;
 270         }
 271
 272         /*
 273          * Oops. The kernel tried to access some bad page. We'll have to
 274          * terminate things with extreme prejudice.
 275          */
 276         if (!user_space_fault(regs))
 277                 printk(KERN_ALERT "Unable to handle kernel pointer dereference"
 278                        " in virtual kernel address space\n");
 279         else
 280                 printk(KERN_ALERT "Unable to handle kernel paging request"
 281                        " in virtual user address space\n");
 282         dump_fault_info(regs);
 283         die(regs, "Oops");
 284         do_exit(SIGKILL);
 285 }
 286
 287 static noinline void do_low_address(struct pt_regs *regs)
 288 {
 289         /* Low-address protection hit in kernel mode means
 290            NULL pointer write access in kernel mode.  */
 291         if (regs->psw.mask & PSW_MASK_PSTATE) {
 292                 /* Low-address protection hit in user mode 'cannot happen'. */
 293                 die (regs, "Low-address protection");
 294                 do_exit(SIGKILL);
 295         }
 296
 297         do_no_context(regs);
 298 }
 299
 300 static noinline void do_sigbus(struct pt_regs *regs)
 301 {
 302         struct task_struct *tsk = current;
 303         struct siginfo si;
 304
 305         /*
 306          * Send a sigbus, regardless of whether we were in kernel
 307          * or user mode.
 308          */
 309         si.si_signo = SIGBUS;
 310         si.si_errno = 0;
 311         si.si_code = BUS_ADRERR;
 312         si.si_addr = (void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK);
 313         force_sig_info(SIGBUS, &si, tsk);
 314 }
 315
 316 static noinline int signal_return(struct pt_regs *regs)
 317 {
 318         u16 instruction;
 319         int rc;
 320
 321         rc = __get_user(instruction, (u16 __user *) regs->psw.addr);
 322         if (rc)
 323                 return rc;
 324         if (instruction == 0x0a77) {
 325                 set_pt_regs_flag(regs, PIF_SYSCALL);
 326                 regs->int_code = 0x00040077;
 327                 return 0;
 328         } else if (instruction == 0x0aad) {
 329                 set_pt_regs_flag(regs, PIF_SYSCALL);
 330                 regs->int_code = 0x000400ad;
 331                 return 0;
 332         }
 333         return -EACCES;
 334 }
 335
 336 static noinline void do_fault_error(struct pt_regs *regs, int access, int fault)
 337 {
 338         int si_code;
 339
 340         switch (fault) {
 341         case VM_FAULT_BADACCESS:
 342                 if (access == VM_EXEC && signal_return(regs) == 0)
 343                         break;
 344         case VM_FAULT_BADMAP:
 345                 /* Bad memory access. Check if it is kernel or user space. */
 346                 if (user_mode(regs)) {
 347                         /* User mode accesses just cause a SIGSEGV */
 348                         si_code = (fault == VM_FAULT_BADMAP) ?
 349                                 SEGV_MAPERR : SEGV_ACCERR;
 350                         do_sigsegv(regs, si_code);
 351                         break;
 352                 }
 353         case VM_FAULT_BADCONTEXT:
 354         case VM_FAULT_PFAULT:
 355                 do_no_context(regs);
 356                 break;
 357         case VM_FAULT_SIGNAL:
 358                 if (!user_mode(regs))
 359                         do_no_context(regs);
 360                 break;
 361         default: /* fault & VM_FAULT_ERROR */
 362                 if (fault & VM_FAULT_OOM) {
 363                         if (!user_mode(regs))
 364                                 do_no_context(regs);
 365                         else
 366                                 pagefault_out_of_memory();
 367                 } else if (fault & VM_FAULT_SIGSEGV) {
 368                         /* Kernel mode? Handle exceptions or die */
 369                         if (!user_mode(regs))
 370                                 do_no_context(regs);
 371                         else
 372                                 do_sigsegv(regs, SEGV_MAPERR);
 373                 } else if (fault & VM_FAULT_SIGBUS) {
 374                         /* Kernel mode? Handle exceptions or die */
 375                         if (!user_mode(regs))
 376                                 do_no_context(regs);
 377                         else
 378                                 do_sigbus(regs);
 379                 } else
 380                         BUG();
 381                 break;
 382         }
 383 }
 384
 385 /*
 386  * This routine handles page faults.  It determines the address,
 387  * and the problem, and then passes it off to one of the appropriate
 388  * routines.
 389  *
 390  * interruption code (int_code):
 391  *   04       Protection           ->  Write-Protection  (suprression)
 392  *   10       Segment translation  ->  Not present       (nullification)
 393  *   11       Page translation     ->  Not present       (nullification)
 394  *   3b       Region third trans.  ->  Not present       (nullification)
 395  */
 396 static inline int do_exception(struct pt_regs *regs, int access)
 397 {
 398 #ifdef CONFIG_PGSTE
 399         struct gmap *gmap;
 400 #endif
 401         struct task_struct *tsk;
 402         struct mm_struct *mm;
 403         struct vm_area_struct *vma;
 404         unsigned long trans_exc_code;
 405         unsigned long address;
 406         unsigned int flags;
 407         int fault;
 408
 409         tsk = current;
 410         /*
 411          * The instruction that caused the program check has
 412          * been nullified. Don't signal single step via SIGTRAP.
 413          */
 414         clear_pt_regs_flag(regs, PIF_PER_TRAP);
 415
 416         if (notify_page_fault(regs))
 417                 return 0;
 418
 419         mm = tsk->mm;
 420         trans_exc_code = regs->int_parm_long;
 421
 422         /*
 423          * Verify that the fault happened in user space, that
 424          * we are not in an interrupt and that there is a
 425          * user context.
 426          */
 427         fault = VM_FAULT_BADCONTEXT;
 428         if (unlikely(!user_space_fault(regs) || faulthandler_disabled() || !mm))
 429                 goto out;
 430
 431         address = trans_exc_code & __FAIL_ADDR_MASK;
 432         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
 433         flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
 434         if (user_mode(regs))
 435                 flags |= FAULT_FLAG_USER;
 436         if (access == VM_WRITE || (trans_exc_code & store_indication) == 0x400)
 437                 flags |= FAULT_FLAG_WRITE;
 438         down_read(&mm->mmap_sem);
 439
 440 #ifdef CONFIG_PGSTE
 441         gmap = (current->flags & PF_VCPU) ?
 442                 (struct gmap *) S390_lowcore.gmap : NULL;
 443         if (gmap) {
 444                 current->thread.gmap_addr = address;
 445                 current->thread.gmap_write_flag = !!(flags & FAULT_FLAG_WRITE);
 446                 current->thread.gmap_int_code = regs->int_code & 0xffff;
 447                 address = __gmap_translate(gmap, address);
 448                 if (address == -EFAULT) {
 449                         fault = VM_FAULT_BADMAP;
 450                         goto out_up;
 451                 }
 452                 if (gmap->pfault_enabled)
 453                         flags |= FAULT_FLAG_RETRY_NOWAIT;
 454         }
 455 #endif
 456
 457 retry:
 458         fault = VM_FAULT_BADMAP;
 459         vma = find_vma(mm, address);
 460         if (!vma)
 461                 goto out_up;
 462
 463         if (unlikely(vma->vm_start > address)) {
 464                 if (!(vma->vm_flags & VM_GROWSDOWN))
 465                         goto out_up;
 466                 if (expand_stack(vma, address))
 467                         goto out_up;
 468         }
 469
 470         /*
 471          * Ok, we have a good vm_area for this memory access, so
 472          * we can handle it..
 473          */
 474         fault = VM_FAULT_BADACCESS;
 475         if (unlikely(!(vma->vm_flags & access)))
 476                 goto out_up;
 477
 478         if (is_vm_hugetlb_page(vma))
 479                 address &= HPAGE_MASK;
 480         /*
 481          * If for any reason at all we couldn't handle the fault,
 482          * make sure we exit gracefully rather than endlessly redo
 483          * the fault.
 484          */
 485         fault = handle_mm_fault(vma, address, flags);
 486         /* No reason to continue if interrupted by SIGKILL. */
 487         if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)) {
 488                 fault = VM_FAULT_SIGNAL;
 489                 if (flags & FAULT_FLAG_RETRY_NOWAIT)
 490                         goto out_up;
 491                 goto out;
 492         }
 493         if (unlikely(fault & VM_FAULT_ERROR))
 494                 goto out_up;
 495
 496         /*
 497          * Major/minor page fault accounting is only done on the
 498          * initial attempt. If we go through a retry, it is extremely
 499          * likely that the page will be found in page cache at that point.
 500          */
 501         if (flags & FAULT_FLAG_ALLOW_RETRY) {
 502                 if (fault & VM_FAULT_MAJOR) {
 503                         tsk->maj_flt++;
 504                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
 505                                       regs, address);
 506                 } else {
 507                         tsk->min_flt++;
 508                         perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
 509                                       regs, address);
 510                 }
 511                 if (fault & VM_FAULT_RETRY) {
 512 #ifdef CONFIG_PGSTE
 513                         if (gmap && (flags & FAULT_FLAG_RETRY_NOWAIT)) {
 514                                 /* FAULT_FLAG_RETRY_NOWAIT has been set,
 515                                  * mmap_sem has not been released */
 516                                 current->thread.gmap_pfault = 1;
 517                                 fault = VM_FAULT_PFAULT;
 518                                 goto out_up;
 519                         }
 520 #endif
 521                         /* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
 522                          * of starvation. */
 523                         flags &= ~(FAULT_FLAG_ALLOW_RETRY |
 524                                    FAULT_FLAG_RETRY_NOWAIT);
 525                         flags |= FAULT_FLAG_TRIED;
 526                         down_read(&mm->mmap_sem);
 527                         goto retry;
 528                 }
 529         }
 530 #ifdef CONFIG_PGSTE
 531         if (gmap) {
 532                 address =  __gmap_link(gmap, current->thread.gmap_addr,
 533                                        address);
 534                 if (address == -EFAULT) {
 535                         fault = VM_FAULT_BADMAP;
 536                         goto out_up;
 537                 }
 538                 if (address == -ENOMEM) {
 539                         fault = VM_FAULT_OOM;
 540                         goto out_up;
 541                 }
 542         }
 543 #endif
 544         fault = 0;
 545 out_up:
 546         up_read(&mm->mmap_sem);
 547 out:
 548         return fault;
 549 }
 550
 551 void do_protection_exception(struct pt_regs *regs)
 552 {
 553         unsigned long trans_exc_code;
 554         int access, fault;
 555
 556         trans_exc_code = regs->int_parm_long;
 557         /*
 558          * Protection exceptions are suppressing, decrement psw address.
 559          * The exception to this rule are aborted transactions, for these
 560          * the PSW already points to the correct location.
 561          */
 562         if (!(regs->int_code & 0x200))
 563                 regs->psw.addr = __rewind_psw(regs->psw, regs->int_code >> 16);
 564         /*
 565          * Check for low-address protection.  This needs to be treated
 566          * as a special case because the translation exception code
 567          * field is not guaranteed to contain valid data in this case.
 568          */
 569         if (unlikely(!(trans_exc_code & 4))) {
 570                 do_low_address(regs);
 571                 return;
 572         }
 573         if (unlikely(MACHINE_HAS_NX && (trans_exc_code & 0x80))) {
 574                 regs->int_parm_long = (trans_exc_code & ~PAGE_MASK) |
 575                                         (regs->psw.addr & PAGE_MASK);
 576                 access = VM_EXEC;
 577                 fault = VM_FAULT_BADACCESS;
 578         } else {
 579                 access = VM_WRITE;
 580                 fault = do_exception(regs, access);
 581         }
 582         if (unlikely(fault))
 583                 do_fault_error(regs, access, fault);
 584 }
 585 NOKPROBE_SYMBOL(do_protection_exception);
 586
 587 void do_dat_exception(struct pt_regs *regs)
 588 {
 589         int access, fault;
 590
 591         access = VM_READ | VM_EXEC | VM_WRITE;
 592         fault = do_exception(regs, access);
 593         if (unlikely(fault))
 594                 do_fault_error(regs, access, fault);
 595 }
 596 NOKPROBE_SYMBOL(do_dat_exception);
 597
 598 #ifdef CONFIG_PFAULT
 599 /*
 600  * 'pfault' pseudo page faults routines.
 601  */
 602 static int pfault_disable;
 603
 604 static int __init nopfault(char *str)
 605 {
 606         pfault_disable = 1;
 607         return 1;
 608 }
 609
 610 __setup("nopfault", nopfault);
 611
 612 struct pfault_refbk {
 613         u16 refdiagc;
 614         u16 reffcode;
 615         u16 refdwlen;
 616         u16 refversn;
 617         u64 refgaddr;
 618         u64 refselmk;
 619         u64 refcmpmk;
 620         u64 reserved;
 621 } __attribute__ ((packed, aligned(8)));
 622
 623 int pfault_init(void)
 624 {
 625         struct pfault_refbk refbk = {
 626                 .refdiagc = 0x258,
 627                 .reffcode = 0,
 628                 .refdwlen = 5,
 629                 .refversn = 2,
 630                 .refgaddr = __LC_LPP,
 631                 .refselmk = 1ULL << 48,
 632                 .refcmpmk = 1ULL << 48,
 633                 .reserved = __PF_RES_FIELD };
 634         int rc;
 635
 636         if (pfault_disable)
 637                 return -1;
 638         diag_stat_inc(DIAG_STAT_X258);
 639         asm volatile(
 640                 "       diag    %1,%0,0x258\n"
 641                 "0:     j       2f\n"
 642                 "1:     la      %0,8\n"
 643                 "2:\n"
 644                 EX_TABLE(0b,1b)
 645                 : "=d" (rc) : "a" (&refbk), "m" (refbk) : "cc");
 646         return rc;
 647 }
 648
 649 void pfault_fini(void)
 650 {
 651         struct pfault_refbk refbk = {
 652                 .refdiagc = 0x258,
 653                 .reffcode = 1,
 654                 .refdwlen = 5,
 655                 .refversn = 2,
 656         };
 657
 658         if (pfault_disable)
 659                 return;
 660         diag_stat_inc(DIAG_STAT_X258);
 661         asm volatile(
 662                 "       diag    %0,0,0x258\n"
 663                 "0:     nopr    %%r7\n"
 664                 EX_TABLE(0b,0b)
 665                 : : "a" (&refbk), "m" (refbk) : "cc");
 666 }
 667
 668 static DEFINE_SPINLOCK(pfault_lock);
 669 static LIST_HEAD(pfault_list);
 670
 671 #define PF_COMPLETE     0x0080
 672
 673 /*
 674  * The mechanism of our pfault code: if Linux is running as guest, runs a user
 675  * space process and the user space process accesses a page that the host has
 676  * paged out we get a pfault interrupt.
 677  *
 678  * This allows us, within the guest, to schedule a different process. Without
 679  * this mechanism the host would have to suspend the whole virtual cpu until
 680  * the page has been paged in.
 681  *
 682  * So when we get such an interrupt then we set the state of the current task
 683  * to uninterruptible and also set the need_resched flag. Both happens within
 684  * interrupt context(!). If we later on want to return to user space we
 685  * recognize the need_resched flag and then call schedule().  It's not very
 686  * obvious how this works...
 687  *
 688  * Of course we have a lot of additional fun with the completion interrupt (->
 689  * host signals that a page of a process has been paged in and the process can
 690  * continue to run). This interrupt can arrive on any cpu and, since we have
 691  * virtual cpus, actually appear before the interrupt that signals that a page
 692  * is missing.
 693  */
 694 static void pfault_interrupt(struct ext_code ext_code,
 695                              unsigned int param32, unsigned long param64)
 696 {
 697         struct task_struct *tsk;
 698         __u16 subcode;
 699         pid_t pid;
 700
 701         /*
 702          * Get the external interruption subcode & pfault initial/completion
 703          * signal bit. VM stores this in the 'cpu address' field associated
 704          * with the external interrupt.
 705          */
 706         subcode = ext_code.subcode;
 707         if ((subcode & 0xff00) != __SUBCODE_MASK)
 708                 return;
 709         inc_irq_stat(IRQEXT_PFL);
 710         /* Get the token (= pid of the affected task). */
 711         pid = param64 & LPP_PFAULT_PID_MASK;
 712         rcu_read_lock();
 713         tsk = find_task_by_pid_ns(pid, &init_pid_ns);
 714         if (tsk)
 715                 get_task_struct(tsk);
 716         rcu_read_unlock();
 717         if (!tsk)
 718                 return;
 719         spin_lock(&pfault_lock);
 720         if (subcode & PF_COMPLETE) {
 721                 /* signal bit is set -> a page has been swapped in by VM */
 722                 if (tsk->thread.pfault_wait == 1) {
 723                         /* Initial interrupt was faster than the completion
 724                          * interrupt. pfault_wait is valid. Set pfault_wait
 725                          * back to zero and wake up the process. This can
 726                          * safely be done because the task is still sleeping
 727                          * and can't produce new pfaults. */
 728                         tsk->thread.pfault_wait = 0;
 729                         list_del(&tsk->thread.list);
 730                         wake_up_process(tsk);
 731                         put_task_struct(tsk);
 732                 } else {
 733                         /* Completion interrupt was faster than initial
 734                          * interrupt. Set pfault_wait to -1 so the initial
 735                          * interrupt doesn't put the task to sleep.
 736                          * If the task is not running, ignore the completion
 737                          * interrupt since it must be a leftover of a PFAULT
 738                          * CANCEL operation which didn't remove all pending
 739                          * completion interrupts. */
 740                         if (tsk->state == TASK_RUNNING)
 741                                 tsk->thread.pfault_wait = -1;
 742                 }
 743         } else {
 744                 /* signal bit not set -> a real page is missing. */
 745                 if (WARN_ON_ONCE(tsk != current))
 746                         goto out;
 747                 if (tsk->thread.pfault_wait == 1) {
 748                         /* Already on the list with a reference: put to sleep */
 749                         goto block;
 750                 } else if (tsk->thread.pfault_wait == -1) {
 751                         /* Completion interrupt was faster than the initial
 752                          * interrupt (pfault_wait == -1). Set pfault_wait
 753                          * back to zero and exit. */
 754                         tsk->thread.pfault_wait = 0;
 755                 } else {
 756                         /* Initial interrupt arrived before completion
 757                          * interrupt. Let the task sleep.
 758                          * An extra task reference is needed since a different
 759                          * cpu may set the task state to TASK_RUNNING again
 760                          * before the scheduler is reached. */
 761                         get_task_struct(tsk);
 762                         tsk->thread.pfault_wait = 1;
 763                         list_add(&tsk->thread.list, &pfault_list);
 764 block:
 765                         /* Since this must be a userspace fault, there
 766                          * is no kernel task state to trample. Rely on the
 767                          * return to userspace schedule() to block. */
 768                         __set_current_state(TASK_UNINTERRUPTIBLE);
 769                         set_tsk_need_resched(tsk);
 770                         set_preempt_need_resched();
 771                 }
 772         }
 773 out:
 774         spin_unlock(&pfault_lock);
 775         put_task_struct(tsk);
 776 }
 777
 778 static int pfault_cpu_dead(unsigned int cpu)
 779 {
 780         struct thread_struct *thread, *next;
 781         struct task_struct *tsk;
 782
 783         spin_lock_irq(&pfault_lock);
 784         list_for_each_entry_safe(thread, next, &pfault_list, list) {
 785                 thread->pfault_wait = 0;
 786                 list_del(&thread->list);
 787                 tsk = container_of(thread, struct task_struct, thread);
 788                 wake_up_process(tsk);
 789                 put_task_struct(tsk);
 790         }
 791         spin_unlock_irq(&pfault_lock);
 792         return 0;
 793 }
 794
 795 static int __init pfault_irq_init(void)
 796 {
 797         int rc;
 798
 799         rc = register_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
 800         if (rc)
 801                 goto out_extint;
 802         rc = pfault_init() == 0 ? 0 : -EOPNOTSUPP;
 803         if (rc)
 804                 goto out_pfault;
 805         irq_subclass_register(IRQ_SUBCLASS_SERVICE_SIGNAL);
 806         cpuhp_setup_state_nocalls(CPUHP_S390_PFAULT_DEAD, "s390/pfault:dead",
 807                                   NULL, pfault_cpu_dead);
 808         return 0;
 809
 810 out_pfault:
 811         unregister_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
 812 out_extint:
 813         pfault_disable = 1;
 814         return rc;
 815 }
 816 early_initcall(pfault_irq_init);
 817
 818 #endif /* CONFIG_PFAULT */