xref: /linux/arch/s390/mm/fault.c (revision b6ebbac51bedf9e98e837688bc838f400196da5e)
1 /*
2  *  S390 version
3  *    Copyright IBM Corp. 1999
4  *    Author(s): Hartmut Penner (hp@de.ibm.com)
5  *               Ulrich Weigand (uweigand@de.ibm.com)
6  *
7  *  Derived from "arch/i386/mm/fault.c"
8  *    Copyright (C) 1995  Linus Torvalds
9  */
10 
11 #include <linux/kernel_stat.h>
12 #include <linux/perf_event.h>
13 #include <linux/signal.h>
14 #include <linux/sched.h>
15 #include <linux/kernel.h>
16 #include <linux/errno.h>
17 #include <linux/string.h>
18 #include <linux/types.h>
19 #include <linux/ptrace.h>
20 #include <linux/mman.h>
21 #include <linux/mm.h>
22 #include <linux/compat.h>
23 #include <linux/smp.h>
24 #include <linux/kdebug.h>
25 #include <linux/init.h>
26 #include <linux/console.h>
27 #include <linux/module.h>
28 #include <linux/hardirq.h>
29 #include <linux/kprobes.h>
30 #include <linux/uaccess.h>
31 #include <linux/hugetlb.h>
32 #include <asm/asm-offsets.h>
33 #include <asm/diag.h>
34 #include <asm/pgtable.h>
35 #include <asm/gmap.h>
36 #include <asm/irq.h>
37 #include <asm/mmu_context.h>
38 #include <asm/facility.h>
39 #include "../kernel/entry.h"
40 
41 #define __FAIL_ADDR_MASK -4096L
42 #define __SUBCODE_MASK 0x0600
43 #define __PF_RES_FIELD 0x8000000000000000ULL
44 
45 #define VM_FAULT_BADCONTEXT	0x010000
46 #define VM_FAULT_BADMAP		0x020000
47 #define VM_FAULT_BADACCESS	0x040000
48 #define VM_FAULT_SIGNAL		0x080000
49 #define VM_FAULT_PFAULT		0x100000
50 
51 static unsigned long store_indication __read_mostly;
52 
53 static int __init fault_init(void)
54 {
55 	if (test_facility(75))
56 		store_indication = 0xc00;
57 	return 0;
58 }
59 early_initcall(fault_init);
60 
61 static inline int notify_page_fault(struct pt_regs *regs)
62 {
63 	int ret = 0;
64 
65 	/* kprobe_running() needs smp_processor_id() */
66 	if (kprobes_built_in() && !user_mode(regs)) {
67 		preempt_disable();
68 		if (kprobe_running() && kprobe_fault_handler(regs, 14))
69 			ret = 1;
70 		preempt_enable();
71 	}
72 	return ret;
73 }
74 
75 
76 /*
77  * Unlock any spinlocks which will prevent us from getting the
78  * message out.
79  */
80 void bust_spinlocks(int yes)
81 {
82 	if (yes) {
83 		oops_in_progress = 1;
84 	} else {
85 		int loglevel_save = console_loglevel;
86 		console_unblank();
87 		oops_in_progress = 0;
88 		/*
89 		 * OK, the message is on the console.  Now we call printk()
90 		 * without oops_in_progress set so that printk will give klogd
91 		 * a poke.  Hold onto your hats...
92 		 */
93 		console_loglevel = 15;
94 		printk(" ");
95 		console_loglevel = loglevel_save;
96 	}
97 }
98 
99 /*
100  * Returns the address space associated with the fault.
101  * Returns 0 for kernel space and 1 for user space.
102  */
103 static inline int user_space_fault(struct pt_regs *regs)
104 {
105 	unsigned long trans_exc_code;
106 
107 	/*
108 	 * The lowest two bits of the translation exception
109 	 * identification indicate which paging table was used.
110 	 */
111 	trans_exc_code = regs->int_parm_long & 3;
112 	if (trans_exc_code == 3) /* home space -> kernel */
113 		return 0;
114 	if (user_mode(regs))
115 		return 1;
116 	if (trans_exc_code == 2) /* secondary space -> set_fs */
117 		return current->thread.mm_segment.ar4;
118 	if (current->flags & PF_VCPU)
119 		return 1;
120 	return 0;
121 }
122 
123 static int bad_address(void *p)
124 {
125 	unsigned long dummy;
126 
127 	return probe_kernel_address((unsigned long *)p, dummy);
128 }
129 
130 static void dump_pagetable(unsigned long asce, unsigned long address)
131 {
132 	unsigned long *table = __va(asce & PAGE_MASK);
133 
134 	pr_alert("AS:%016lx ", asce);
135 	switch (asce & _ASCE_TYPE_MASK) {
136 	case _ASCE_TYPE_REGION1:
137 		table = table + ((address >> 53) & 0x7ff);
138 		if (bad_address(table))
139 			goto bad;
140 		pr_cont("R1:%016lx ", *table);
141 		if (*table & _REGION_ENTRY_INVALID)
142 			goto out;
143 		table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
144 		/* fallthrough */
145 	case _ASCE_TYPE_REGION2:
146 		table = table + ((address >> 42) & 0x7ff);
147 		if (bad_address(table))
148 			goto bad;
149 		pr_cont("R2:%016lx ", *table);
150 		if (*table & _REGION_ENTRY_INVALID)
151 			goto out;
152 		table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
153 		/* fallthrough */
154 	case _ASCE_TYPE_REGION3:
155 		table = table + ((address >> 31) & 0x7ff);
156 		if (bad_address(table))
157 			goto bad;
158 		pr_cont("R3:%016lx ", *table);
159 		if (*table & (_REGION_ENTRY_INVALID | _REGION3_ENTRY_LARGE))
160 			goto out;
161 		table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
162 		/* fallthrough */
163 	case _ASCE_TYPE_SEGMENT:
164 		table = table + ((address >> 20) & 0x7ff);
165 		if (bad_address(table))
166 			goto bad;
167 		pr_cont("S:%016lx ", *table);
168 		if (*table & (_SEGMENT_ENTRY_INVALID | _SEGMENT_ENTRY_LARGE))
169 			goto out;
170 		table = (unsigned long *)(*table & _SEGMENT_ENTRY_ORIGIN);
171 	}
172 	table = table + ((address >> 12) & 0xff);
173 	if (bad_address(table))
174 		goto bad;
175 	pr_cont("P:%016lx ", *table);
176 out:
177 	pr_cont("\n");
178 	return;
179 bad:
180 	pr_cont("BAD\n");
181 }
182 
183 static void dump_fault_info(struct pt_regs *regs)
184 {
185 	unsigned long asce;
186 
187 	pr_alert("Failing address: %016lx TEID: %016lx\n",
188 		 regs->int_parm_long & __FAIL_ADDR_MASK, regs->int_parm_long);
189 	pr_alert("Fault in ");
190 	switch (regs->int_parm_long & 3) {
191 	case 3:
192 		pr_cont("home space ");
193 		break;
194 	case 2:
195 		pr_cont("secondary space ");
196 		break;
197 	case 1:
198 		pr_cont("access register ");
199 		break;
200 	case 0:
201 		pr_cont("primary space ");
202 		break;
203 	}
204 	pr_cont("mode while using ");
205 	if (!user_space_fault(regs)) {
206 		asce = S390_lowcore.kernel_asce;
207 		pr_cont("kernel ");
208 	}
209 #ifdef CONFIG_PGSTE
210 	else if ((current->flags & PF_VCPU) && S390_lowcore.gmap) {
211 		struct gmap *gmap = (struct gmap *)S390_lowcore.gmap;
212 		asce = gmap->asce;
213 		pr_cont("gmap ");
214 	}
215 #endif
216 	else {
217 		asce = S390_lowcore.user_asce;
218 		pr_cont("user ");
219 	}
220 	pr_cont("ASCE.\n");
221 	dump_pagetable(asce, regs->int_parm_long & __FAIL_ADDR_MASK);
222 }
223 
224 int show_unhandled_signals = 1;
225 
226 void report_user_fault(struct pt_regs *regs, long signr, int is_mm_fault)
227 {
228 	if ((task_pid_nr(current) > 1) && !show_unhandled_signals)
229 		return;
230 	if (!unhandled_signal(current, signr))
231 		return;
232 	if (!printk_ratelimit())
233 		return;
234 	printk(KERN_ALERT "User process fault: interruption code %04x ilc:%d ",
235 	       regs->int_code & 0xffff, regs->int_code >> 17);
236 	print_vma_addr(KERN_CONT "in ", regs->psw.addr);
237 	printk(KERN_CONT "\n");
238 	if (is_mm_fault)
239 		dump_fault_info(regs);
240 	show_regs(regs);
241 }
242 
243 /*
244  * Send SIGSEGV to task.  This is an external routine
245  * to keep the stack usage of do_page_fault small.
246  */
247 static noinline void do_sigsegv(struct pt_regs *regs, int si_code)
248 {
249 	struct siginfo si;
250 
251 	report_user_fault(regs, SIGSEGV, 1);
252 	si.si_signo = SIGSEGV;
253 	si.si_errno = 0;
254 	si.si_code = si_code;
255 	si.si_addr = (void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK);
256 	force_sig_info(SIGSEGV, &si, current);
257 }
258 
259 static noinline void do_no_context(struct pt_regs *regs)
260 {
261 	const struct exception_table_entry *fixup;
262 
263 	/* Are we prepared to handle this kernel fault?  */
264 	fixup = search_exception_tables(regs->psw.addr);
265 	if (fixup) {
266 		regs->psw.addr = extable_fixup(fixup);
267 		return;
268 	}
269 
270 	/*
271 	 * Oops. The kernel tried to access some bad page. We'll have to
272 	 * terminate things with extreme prejudice.
273 	 */
274 	if (!user_space_fault(regs))
275 		printk(KERN_ALERT "Unable to handle kernel pointer dereference"
276 		       " in virtual kernel address space\n");
277 	else
278 		printk(KERN_ALERT "Unable to handle kernel paging request"
279 		       " in virtual user address space\n");
280 	dump_fault_info(regs);
281 	die(regs, "Oops");
282 	do_exit(SIGKILL);
283 }
284 
285 static noinline void do_low_address(struct pt_regs *regs)
286 {
287 	/* Low-address protection hit in kernel mode means
288 	   NULL pointer write access in kernel mode.  */
289 	if (regs->psw.mask & PSW_MASK_PSTATE) {
290 		/* Low-address protection hit in user mode 'cannot happen'. */
291 		die (regs, "Low-address protection");
292 		do_exit(SIGKILL);
293 	}
294 
295 	do_no_context(regs);
296 }
297 
298 static noinline void do_sigbus(struct pt_regs *regs)
299 {
300 	struct task_struct *tsk = current;
301 	struct siginfo si;
302 
303 	/*
304 	 * Send a sigbus, regardless of whether we were in kernel
305 	 * or user mode.
306 	 */
307 	si.si_signo = SIGBUS;
308 	si.si_errno = 0;
309 	si.si_code = BUS_ADRERR;
310 	si.si_addr = (void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK);
311 	force_sig_info(SIGBUS, &si, tsk);
312 }
313 
314 static noinline void do_fault_error(struct pt_regs *regs, int fault)
315 {
316 	int si_code;
317 
318 	switch (fault) {
319 	case VM_FAULT_BADACCESS:
320 	case VM_FAULT_BADMAP:
321 		/* Bad memory access. Check if it is kernel or user space. */
322 		if (user_mode(regs)) {
323 			/* User mode accesses just cause a SIGSEGV */
324 			si_code = (fault == VM_FAULT_BADMAP) ?
325 				SEGV_MAPERR : SEGV_ACCERR;
326 			do_sigsegv(regs, si_code);
327 			return;
328 		}
329 	case VM_FAULT_BADCONTEXT:
330 	case VM_FAULT_PFAULT:
331 		do_no_context(regs);
332 		break;
333 	case VM_FAULT_SIGNAL:
334 		if (!user_mode(regs))
335 			do_no_context(regs);
336 		break;
337 	default: /* fault & VM_FAULT_ERROR */
338 		if (fault & VM_FAULT_OOM) {
339 			if (!user_mode(regs))
340 				do_no_context(regs);
341 			else
342 				pagefault_out_of_memory();
343 		} else if (fault & VM_FAULT_SIGSEGV) {
344 			/* Kernel mode? Handle exceptions or die */
345 			if (!user_mode(regs))
346 				do_no_context(regs);
347 			else
348 				do_sigsegv(regs, SEGV_MAPERR);
349 		} else if (fault & VM_FAULT_SIGBUS) {
350 			/* Kernel mode? Handle exceptions or die */
351 			if (!user_mode(regs))
352 				do_no_context(regs);
353 			else
354 				do_sigbus(regs);
355 		} else
356 			BUG();
357 		break;
358 	}
359 }
360 
361 /*
362  * This routine handles page faults.  It determines the address,
363  * and the problem, and then passes it off to one of the appropriate
364  * routines.
365  *
366  * interruption code (int_code):
367  *   04       Protection           ->  Write-Protection  (suprression)
368  *   10       Segment translation  ->  Not present       (nullification)
369  *   11       Page translation     ->  Not present       (nullification)
370  *   3b       Region third trans.  ->  Not present       (nullification)
371  */
372 static inline int do_exception(struct pt_regs *regs, int access)
373 {
374 #ifdef CONFIG_PGSTE
375 	struct gmap *gmap;
376 #endif
377 	struct task_struct *tsk;
378 	struct mm_struct *mm;
379 	struct vm_area_struct *vma;
380 	unsigned long trans_exc_code;
381 	unsigned long address;
382 	unsigned int flags;
383 	int fault;
384 
385 	tsk = current;
386 	/*
387 	 * The instruction that caused the program check has
388 	 * been nullified. Don't signal single step via SIGTRAP.
389 	 */
390 	clear_pt_regs_flag(regs, PIF_PER_TRAP);
391 
392 	if (notify_page_fault(regs))
393 		return 0;
394 
395 	mm = tsk->mm;
396 	trans_exc_code = regs->int_parm_long;
397 
398 	/*
399 	 * Verify that the fault happened in user space, that
400 	 * we are not in an interrupt and that there is a
401 	 * user context.
402 	 */
403 	fault = VM_FAULT_BADCONTEXT;
404 	if (unlikely(!user_space_fault(regs) || faulthandler_disabled() || !mm))
405 		goto out;
406 
407 	address = trans_exc_code & __FAIL_ADDR_MASK;
408 	perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
409 	flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
410 	if (user_mode(regs))
411 		flags |= FAULT_FLAG_USER;
412 	if (access == VM_WRITE || (trans_exc_code & store_indication) == 0x400)
413 		flags |= FAULT_FLAG_WRITE;
414 	down_read(&mm->mmap_sem);
415 
416 #ifdef CONFIG_PGSTE
417 	gmap = (current->flags & PF_VCPU) ?
418 		(struct gmap *) S390_lowcore.gmap : NULL;
419 	if (gmap) {
420 		current->thread.gmap_addr = address;
421 		current->thread.gmap_write_flag = !!(flags & FAULT_FLAG_WRITE);
422 		current->thread.gmap_int_code = regs->int_code & 0xffff;
423 		address = __gmap_translate(gmap, address);
424 		if (address == -EFAULT) {
425 			fault = VM_FAULT_BADMAP;
426 			goto out_up;
427 		}
428 		if (gmap->pfault_enabled)
429 			flags |= FAULT_FLAG_RETRY_NOWAIT;
430 	}
431 #endif
432 
433 retry:
434 	fault = VM_FAULT_BADMAP;
435 	vma = find_vma(mm, address);
436 	if (!vma)
437 		goto out_up;
438 
439 	if (unlikely(vma->vm_start > address)) {
440 		if (!(vma->vm_flags & VM_GROWSDOWN))
441 			goto out_up;
442 		if (expand_stack(vma, address))
443 			goto out_up;
444 	}
445 
446 	/*
447 	 * Ok, we have a good vm_area for this memory access, so
448 	 * we can handle it..
449 	 */
450 	fault = VM_FAULT_BADACCESS;
451 	if (unlikely(!(vma->vm_flags & access)))
452 		goto out_up;
453 
454 	if (is_vm_hugetlb_page(vma))
455 		address &= HPAGE_MASK;
456 	/*
457 	 * If for any reason at all we couldn't handle the fault,
458 	 * make sure we exit gracefully rather than endlessly redo
459 	 * the fault.
460 	 */
461 	fault = handle_mm_fault(vma, address, flags);
462 	/* No reason to continue if interrupted by SIGKILL. */
463 	if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)) {
464 		fault = VM_FAULT_SIGNAL;
465 		goto out;
466 	}
467 	if (unlikely(fault & VM_FAULT_ERROR))
468 		goto out_up;
469 
470 	/*
471 	 * Major/minor page fault accounting is only done on the
472 	 * initial attempt. If we go through a retry, it is extremely
473 	 * likely that the page will be found in page cache at that point.
474 	 */
475 	if (flags & FAULT_FLAG_ALLOW_RETRY) {
476 		if (fault & VM_FAULT_MAJOR) {
477 			tsk->maj_flt++;
478 			perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
479 				      regs, address);
480 		} else {
481 			tsk->min_flt++;
482 			perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
483 				      regs, address);
484 		}
485 		if (fault & VM_FAULT_RETRY) {
486 #ifdef CONFIG_PGSTE
487 			if (gmap && (flags & FAULT_FLAG_RETRY_NOWAIT)) {
488 				/* FAULT_FLAG_RETRY_NOWAIT has been set,
489 				 * mmap_sem has not been released */
490 				current->thread.gmap_pfault = 1;
491 				fault = VM_FAULT_PFAULT;
492 				goto out_up;
493 			}
494 #endif
495 			/* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
496 			 * of starvation. */
497 			flags &= ~(FAULT_FLAG_ALLOW_RETRY |
498 				   FAULT_FLAG_RETRY_NOWAIT);
499 			flags |= FAULT_FLAG_TRIED;
500 			down_read(&mm->mmap_sem);
501 			goto retry;
502 		}
503 	}
504 #ifdef CONFIG_PGSTE
505 	if (gmap) {
506 		address =  __gmap_link(gmap, current->thread.gmap_addr,
507 				       address);
508 		if (address == -EFAULT) {
509 			fault = VM_FAULT_BADMAP;
510 			goto out_up;
511 		}
512 		if (address == -ENOMEM) {
513 			fault = VM_FAULT_OOM;
514 			goto out_up;
515 		}
516 	}
517 #endif
518 	fault = 0;
519 out_up:
520 	up_read(&mm->mmap_sem);
521 out:
522 	return fault;
523 }
524 
525 void do_protection_exception(struct pt_regs *regs)
526 {
527 	unsigned long trans_exc_code;
528 	int fault;
529 
530 	trans_exc_code = regs->int_parm_long;
531 	/*
532 	 * Protection exceptions are suppressing, decrement psw address.
533 	 * The exception to this rule are aborted transactions, for these
534 	 * the PSW already points to the correct location.
535 	 */
536 	if (!(regs->int_code & 0x200))
537 		regs->psw.addr = __rewind_psw(regs->psw, regs->int_code >> 16);
538 	/*
539 	 * Check for low-address protection.  This needs to be treated
540 	 * as a special case because the translation exception code
541 	 * field is not guaranteed to contain valid data in this case.
542 	 */
543 	if (unlikely(!(trans_exc_code & 4))) {
544 		do_low_address(regs);
545 		return;
546 	}
547 	fault = do_exception(regs, VM_WRITE);
548 	if (unlikely(fault))
549 		do_fault_error(regs, fault);
550 }
551 NOKPROBE_SYMBOL(do_protection_exception);
552 
553 void do_dat_exception(struct pt_regs *regs)
554 {
555 	int access, fault;
556 
557 	access = VM_READ | VM_EXEC | VM_WRITE;
558 	fault = do_exception(regs, access);
559 	if (unlikely(fault))
560 		do_fault_error(regs, fault);
561 }
562 NOKPROBE_SYMBOL(do_dat_exception);
563 
564 #ifdef CONFIG_PFAULT
565 /*
566  * 'pfault' pseudo page faults routines.
567  */
568 static int pfault_disable;
569 
570 static int __init nopfault(char *str)
571 {
572 	pfault_disable = 1;
573 	return 1;
574 }
575 
576 __setup("nopfault", nopfault);
577 
578 struct pfault_refbk {
579 	u16 refdiagc;
580 	u16 reffcode;
581 	u16 refdwlen;
582 	u16 refversn;
583 	u64 refgaddr;
584 	u64 refselmk;
585 	u64 refcmpmk;
586 	u64 reserved;
587 } __attribute__ ((packed, aligned(8)));
588 
589 int pfault_init(void)
590 {
591 	struct pfault_refbk refbk = {
592 		.refdiagc = 0x258,
593 		.reffcode = 0,
594 		.refdwlen = 5,
595 		.refversn = 2,
596 		.refgaddr = __LC_LPP,
597 		.refselmk = 1ULL << 48,
598 		.refcmpmk = 1ULL << 48,
599 		.reserved = __PF_RES_FIELD };
600         int rc;
601 
602 	if (pfault_disable)
603 		return -1;
604 	diag_stat_inc(DIAG_STAT_X258);
605 	asm volatile(
606 		"	diag	%1,%0,0x258\n"
607 		"0:	j	2f\n"
608 		"1:	la	%0,8\n"
609 		"2:\n"
610 		EX_TABLE(0b,1b)
611 		: "=d" (rc) : "a" (&refbk), "m" (refbk) : "cc");
612         return rc;
613 }
614 
615 void pfault_fini(void)
616 {
617 	struct pfault_refbk refbk = {
618 		.refdiagc = 0x258,
619 		.reffcode = 1,
620 		.refdwlen = 5,
621 		.refversn = 2,
622 	};
623 
624 	if (pfault_disable)
625 		return;
626 	diag_stat_inc(DIAG_STAT_X258);
627 	asm volatile(
628 		"	diag	%0,0,0x258\n"
629 		"0:	nopr	%%r7\n"
630 		EX_TABLE(0b,0b)
631 		: : "a" (&refbk), "m" (refbk) : "cc");
632 }
633 
634 static DEFINE_SPINLOCK(pfault_lock);
635 static LIST_HEAD(pfault_list);
636 
637 #define PF_COMPLETE	0x0080
638 
639 /*
640  * The mechanism of our pfault code: if Linux is running as guest, runs a user
641  * space process and the user space process accesses a page that the host has
642  * paged out we get a pfault interrupt.
643  *
644  * This allows us, within the guest, to schedule a different process. Without
645  * this mechanism the host would have to suspend the whole virtual cpu until
646  * the page has been paged in.
647  *
648  * So when we get such an interrupt then we set the state of the current task
649  * to uninterruptible and also set the need_resched flag. Both happens within
650  * interrupt context(!). If we later on want to return to user space we
651  * recognize the need_resched flag and then call schedule().  It's not very
652  * obvious how this works...
653  *
654  * Of course we have a lot of additional fun with the completion interrupt (->
655  * host signals that a page of a process has been paged in and the process can
656  * continue to run). This interrupt can arrive on any cpu and, since we have
657  * virtual cpus, actually appear before the interrupt that signals that a page
658  * is missing.
659  */
660 static void pfault_interrupt(struct ext_code ext_code,
661 			     unsigned int param32, unsigned long param64)
662 {
663 	struct task_struct *tsk;
664 	__u16 subcode;
665 	pid_t pid;
666 
667 	/*
668 	 * Get the external interruption subcode & pfault initial/completion
669 	 * signal bit. VM stores this in the 'cpu address' field associated
670 	 * with the external interrupt.
671 	 */
672 	subcode = ext_code.subcode;
673 	if ((subcode & 0xff00) != __SUBCODE_MASK)
674 		return;
675 	inc_irq_stat(IRQEXT_PFL);
676 	/* Get the token (= pid of the affected task). */
677 	pid = param64 & LPP_PFAULT_PID_MASK;
678 	rcu_read_lock();
679 	tsk = find_task_by_pid_ns(pid, &init_pid_ns);
680 	if (tsk)
681 		get_task_struct(tsk);
682 	rcu_read_unlock();
683 	if (!tsk)
684 		return;
685 	spin_lock(&pfault_lock);
686 	if (subcode & PF_COMPLETE) {
687 		/* signal bit is set -> a page has been swapped in by VM */
688 		if (tsk->thread.pfault_wait == 1) {
689 			/* Initial interrupt was faster than the completion
690 			 * interrupt. pfault_wait is valid. Set pfault_wait
691 			 * back to zero and wake up the process. This can
692 			 * safely be done because the task is still sleeping
693 			 * and can't produce new pfaults. */
694 			tsk->thread.pfault_wait = 0;
695 			list_del(&tsk->thread.list);
696 			wake_up_process(tsk);
697 			put_task_struct(tsk);
698 		} else {
699 			/* Completion interrupt was faster than initial
700 			 * interrupt. Set pfault_wait to -1 so the initial
701 			 * interrupt doesn't put the task to sleep.
702 			 * If the task is not running, ignore the completion
703 			 * interrupt since it must be a leftover of a PFAULT
704 			 * CANCEL operation which didn't remove all pending
705 			 * completion interrupts. */
706 			if (tsk->state == TASK_RUNNING)
707 				tsk->thread.pfault_wait = -1;
708 		}
709 	} else {
710 		/* signal bit not set -> a real page is missing. */
711 		if (WARN_ON_ONCE(tsk != current))
712 			goto out;
713 		if (tsk->thread.pfault_wait == 1) {
714 			/* Already on the list with a reference: put to sleep */
715 			goto block;
716 		} else if (tsk->thread.pfault_wait == -1) {
717 			/* Completion interrupt was faster than the initial
718 			 * interrupt (pfault_wait == -1). Set pfault_wait
719 			 * back to zero and exit. */
720 			tsk->thread.pfault_wait = 0;
721 		} else {
722 			/* Initial interrupt arrived before completion
723 			 * interrupt. Let the task sleep.
724 			 * An extra task reference is needed since a different
725 			 * cpu may set the task state to TASK_RUNNING again
726 			 * before the scheduler is reached. */
727 			get_task_struct(tsk);
728 			tsk->thread.pfault_wait = 1;
729 			list_add(&tsk->thread.list, &pfault_list);
730 block:
731 			/* Since this must be a userspace fault, there
732 			 * is no kernel task state to trample. Rely on the
733 			 * return to userspace schedule() to block. */
734 			__set_current_state(TASK_UNINTERRUPTIBLE);
735 			set_tsk_need_resched(tsk);
736 		}
737 	}
738 out:
739 	spin_unlock(&pfault_lock);
740 	put_task_struct(tsk);
741 }
742 
743 static int pfault_cpu_notify(struct notifier_block *self, unsigned long action,
744 			     void *hcpu)
745 {
746 	struct thread_struct *thread, *next;
747 	struct task_struct *tsk;
748 
749 	switch (action & ~CPU_TASKS_FROZEN) {
750 	case CPU_DEAD:
751 		spin_lock_irq(&pfault_lock);
752 		list_for_each_entry_safe(thread, next, &pfault_list, list) {
753 			thread->pfault_wait = 0;
754 			list_del(&thread->list);
755 			tsk = container_of(thread, struct task_struct, thread);
756 			wake_up_process(tsk);
757 			put_task_struct(tsk);
758 		}
759 		spin_unlock_irq(&pfault_lock);
760 		break;
761 	default:
762 		break;
763 	}
764 	return NOTIFY_OK;
765 }
766 
767 static int __init pfault_irq_init(void)
768 {
769 	int rc;
770 
771 	rc = register_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
772 	if (rc)
773 		goto out_extint;
774 	rc = pfault_init() == 0 ? 0 : -EOPNOTSUPP;
775 	if (rc)
776 		goto out_pfault;
777 	irq_subclass_register(IRQ_SUBCLASS_SERVICE_SIGNAL);
778 	hotcpu_notifier(pfault_cpu_notify, 0);
779 	return 0;
780 
781 out_pfault:
782 	unregister_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
783 out_extint:
784 	pfault_disable = 1;
785 	return rc;
786 }
787 early_initcall(pfault_irq_init);
788 
789 #endif /* CONFIG_PFAULT */
790