xref: /openbmc/linux/arch/s390/mm/fault.c (revision 8e8e69d6)
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 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 	default:
215 		unreachable();
216 	}
217 	pr_cont("ASCE.\n");
218 	dump_pagetable(asce, regs->int_parm_long & __FAIL_ADDR_MASK);
219 }
220 
221 int show_unhandled_signals = 1;
222 
223 void report_user_fault(struct pt_regs *regs, long signr, int is_mm_fault)
224 {
225 	if ((task_pid_nr(current) > 1) && !show_unhandled_signals)
226 		return;
227 	if (!unhandled_signal(current, signr))
228 		return;
229 	if (!printk_ratelimit())
230 		return;
231 	printk(KERN_ALERT "User process fault: interruption code %04x ilc:%d ",
232 	       regs->int_code & 0xffff, regs->int_code >> 17);
233 	print_vma_addr(KERN_CONT "in ", regs->psw.addr);
234 	printk(KERN_CONT "\n");
235 	if (is_mm_fault)
236 		dump_fault_info(regs);
237 	show_regs(regs);
238 }
239 
240 /*
241  * Send SIGSEGV to task.  This is an external routine
242  * to keep the stack usage of do_page_fault small.
243  */
244 static noinline void do_sigsegv(struct pt_regs *regs, int si_code)
245 {
246 	report_user_fault(regs, SIGSEGV, 1);
247 	force_sig_fault(SIGSEGV, si_code,
248 			(void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK),
249 			current);
250 }
251 
252 const struct exception_table_entry *s390_search_extables(unsigned long addr)
253 {
254 	const struct exception_table_entry *fixup;
255 
256 	fixup = search_extable(__start_dma_ex_table,
257 			       __stop_dma_ex_table - __start_dma_ex_table,
258 			       addr);
259 	if (!fixup)
260 		fixup = search_exception_tables(addr);
261 	return fixup;
262 }
263 
264 static noinline void do_no_context(struct pt_regs *regs)
265 {
266 	const struct exception_table_entry *fixup;
267 
268 	/* Are we prepared to handle this kernel fault?  */
269 	fixup = s390_search_extables(regs->psw.addr);
270 	if (fixup) {
271 		regs->psw.addr = extable_fixup(fixup);
272 		return;
273 	}
274 
275 	/*
276 	 * Oops. The kernel tried to access some bad page. We'll have to
277 	 * terminate things with extreme prejudice.
278 	 */
279 	if (get_fault_type(regs) == KERNEL_FAULT)
280 		printk(KERN_ALERT "Unable to handle kernel pointer dereference"
281 		       " in virtual kernel address space\n");
282 	else
283 		printk(KERN_ALERT "Unable to handle kernel paging request"
284 		       " in virtual user address space\n");
285 	dump_fault_info(regs);
286 	die(regs, "Oops");
287 	do_exit(SIGKILL);
288 }
289 
290 static noinline void do_low_address(struct pt_regs *regs)
291 {
292 	/* Low-address protection hit in kernel mode means
293 	   NULL pointer write access in kernel mode.  */
294 	if (regs->psw.mask & PSW_MASK_PSTATE) {
295 		/* Low-address protection hit in user mode 'cannot happen'. */
296 		die (regs, "Low-address protection");
297 		do_exit(SIGKILL);
298 	}
299 
300 	do_no_context(regs);
301 }
302 
303 static noinline void do_sigbus(struct pt_regs *regs)
304 {
305 	/*
306 	 * Send a sigbus, regardless of whether we were in kernel
307 	 * or user mode.
308 	 */
309 	force_sig_fault(SIGBUS, BUS_ADRERR,
310 			(void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK),
311 			current);
312 }
313 
314 static noinline int signal_return(struct pt_regs *regs)
315 {
316 	u16 instruction;
317 	int rc;
318 
319 	rc = __get_user(instruction, (u16 __user *) regs->psw.addr);
320 	if (rc)
321 		return rc;
322 	if (instruction == 0x0a77) {
323 		set_pt_regs_flag(regs, PIF_SYSCALL);
324 		regs->int_code = 0x00040077;
325 		return 0;
326 	} else if (instruction == 0x0aad) {
327 		set_pt_regs_flag(regs, PIF_SYSCALL);
328 		regs->int_code = 0x000400ad;
329 		return 0;
330 	}
331 	return -EACCES;
332 }
333 
334 static noinline void do_fault_error(struct pt_regs *regs, int access,
335 					vm_fault_t fault)
336 {
337 	int si_code;
338 
339 	switch (fault) {
340 	case VM_FAULT_BADACCESS:
341 		if (access == VM_EXEC && signal_return(regs) == 0)
342 			break;
343 	case VM_FAULT_BADMAP:
344 		/* Bad memory access. Check if it is kernel or user space. */
345 		if (user_mode(regs)) {
346 			/* User mode accesses just cause a SIGSEGV */
347 			si_code = (fault == VM_FAULT_BADMAP) ?
348 				SEGV_MAPERR : SEGV_ACCERR;
349 			do_sigsegv(regs, si_code);
350 			break;
351 		}
352 	case VM_FAULT_BADCONTEXT:
353 	case VM_FAULT_PFAULT:
354 		do_no_context(regs);
355 		break;
356 	case VM_FAULT_SIGNAL:
357 		if (!user_mode(regs))
358 			do_no_context(regs);
359 		break;
360 	default: /* fault & VM_FAULT_ERROR */
361 		if (fault & VM_FAULT_OOM) {
362 			if (!user_mode(regs))
363 				do_no_context(regs);
364 			else
365 				pagefault_out_of_memory();
366 		} else if (fault & VM_FAULT_SIGSEGV) {
367 			/* Kernel mode? Handle exceptions or die */
368 			if (!user_mode(regs))
369 				do_no_context(regs);
370 			else
371 				do_sigsegv(regs, SEGV_MAPERR);
372 		} else if (fault & VM_FAULT_SIGBUS) {
373 			/* Kernel mode? Handle exceptions or die */
374 			if (!user_mode(regs))
375 				do_no_context(regs);
376 			else
377 				do_sigbus(regs);
378 		} else
379 			BUG();
380 		break;
381 	}
382 }
383 
384 /*
385  * This routine handles page faults.  It determines the address,
386  * and the problem, and then passes it off to one of the appropriate
387  * routines.
388  *
389  * interruption code (int_code):
390  *   04       Protection           ->  Write-Protection  (suprression)
391  *   10       Segment translation  ->  Not present       (nullification)
392  *   11       Page translation     ->  Not present       (nullification)
393  *   3b       Region third trans.  ->  Not present       (nullification)
394  */
395 static inline vm_fault_t do_exception(struct pt_regs *regs, int access)
396 {
397 	struct gmap *gmap;
398 	struct task_struct *tsk;
399 	struct mm_struct *mm;
400 	struct vm_area_struct *vma;
401 	enum fault_type type;
402 	unsigned long trans_exc_code;
403 	unsigned long address;
404 	unsigned int flags;
405 	vm_fault_t fault;
406 
407 	tsk = current;
408 	/*
409 	 * The instruction that caused the program check has
410 	 * been nullified. Don't signal single step via SIGTRAP.
411 	 */
412 	clear_pt_regs_flag(regs, PIF_PER_TRAP);
413 
414 	if (notify_page_fault(regs))
415 		return 0;
416 
417 	mm = tsk->mm;
418 	trans_exc_code = regs->int_parm_long;
419 
420 	/*
421 	 * Verify that the fault happened in user space, that
422 	 * we are not in an interrupt and that there is a
423 	 * user context.
424 	 */
425 	fault = VM_FAULT_BADCONTEXT;
426 	type = get_fault_type(regs);
427 	switch (type) {
428 	case KERNEL_FAULT:
429 		goto out;
430 	case VDSO_FAULT:
431 		fault = VM_FAULT_BADMAP;
432 		goto out;
433 	case USER_FAULT:
434 	case GMAP_FAULT:
435 		if (faulthandler_disabled() || !mm)
436 			goto out;
437 		break;
438 	}
439 
440 	address = trans_exc_code & __FAIL_ADDR_MASK;
441 	perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
442 	flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
443 	if (user_mode(regs))
444 		flags |= FAULT_FLAG_USER;
445 	if (access == VM_WRITE || (trans_exc_code & store_indication) == 0x400)
446 		flags |= FAULT_FLAG_WRITE;
447 	down_read(&mm->mmap_sem);
448 
449 	gmap = NULL;
450 	if (IS_ENABLED(CONFIG_PGSTE) && type == GMAP_FAULT) {
451 		gmap = (struct gmap *) S390_lowcore.gmap;
452 		current->thread.gmap_addr = address;
453 		current->thread.gmap_write_flag = !!(flags & FAULT_FLAG_WRITE);
454 		current->thread.gmap_int_code = regs->int_code & 0xffff;
455 		address = __gmap_translate(gmap, address);
456 		if (address == -EFAULT) {
457 			fault = VM_FAULT_BADMAP;
458 			goto out_up;
459 		}
460 		if (gmap->pfault_enabled)
461 			flags |= FAULT_FLAG_RETRY_NOWAIT;
462 	}
463 
464 retry:
465 	fault = VM_FAULT_BADMAP;
466 	vma = find_vma(mm, address);
467 	if (!vma)
468 		goto out_up;
469 
470 	if (unlikely(vma->vm_start > address)) {
471 		if (!(vma->vm_flags & VM_GROWSDOWN))
472 			goto out_up;
473 		if (expand_stack(vma, address))
474 			goto out_up;
475 	}
476 
477 	/*
478 	 * Ok, we have a good vm_area for this memory access, so
479 	 * we can handle it..
480 	 */
481 	fault = VM_FAULT_BADACCESS;
482 	if (unlikely(!(vma->vm_flags & access)))
483 		goto out_up;
484 
485 	if (is_vm_hugetlb_page(vma))
486 		address &= HPAGE_MASK;
487 	/*
488 	 * If for any reason at all we couldn't handle the fault,
489 	 * make sure we exit gracefully rather than endlessly redo
490 	 * the fault.
491 	 */
492 	fault = handle_mm_fault(vma, address, flags);
493 	/* No reason to continue if interrupted by SIGKILL. */
494 	if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)) {
495 		fault = VM_FAULT_SIGNAL;
496 		if (flags & FAULT_FLAG_RETRY_NOWAIT)
497 			goto out_up;
498 		goto out;
499 	}
500 	if (unlikely(fault & VM_FAULT_ERROR))
501 		goto out_up;
502 
503 	/*
504 	 * Major/minor page fault accounting is only done on the
505 	 * initial attempt. If we go through a retry, it is extremely
506 	 * likely that the page will be found in page cache at that point.
507 	 */
508 	if (flags & FAULT_FLAG_ALLOW_RETRY) {
509 		if (fault & VM_FAULT_MAJOR) {
510 			tsk->maj_flt++;
511 			perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
512 				      regs, address);
513 		} else {
514 			tsk->min_flt++;
515 			perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
516 				      regs, address);
517 		}
518 		if (fault & VM_FAULT_RETRY) {
519 			if (IS_ENABLED(CONFIG_PGSTE) && gmap &&
520 			    (flags & FAULT_FLAG_RETRY_NOWAIT)) {
521 				/* FAULT_FLAG_RETRY_NOWAIT has been set,
522 				 * mmap_sem has not been released */
523 				current->thread.gmap_pfault = 1;
524 				fault = VM_FAULT_PFAULT;
525 				goto out_up;
526 			}
527 			/* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
528 			 * of starvation. */
529 			flags &= ~(FAULT_FLAG_ALLOW_RETRY |
530 				   FAULT_FLAG_RETRY_NOWAIT);
531 			flags |= FAULT_FLAG_TRIED;
532 			down_read(&mm->mmap_sem);
533 			goto retry;
534 		}
535 	}
536 	if (IS_ENABLED(CONFIG_PGSTE) && gmap) {
537 		address =  __gmap_link(gmap, current->thread.gmap_addr,
538 				       address);
539 		if (address == -EFAULT) {
540 			fault = VM_FAULT_BADMAP;
541 			goto out_up;
542 		}
543 		if (address == -ENOMEM) {
544 			fault = VM_FAULT_OOM;
545 			goto out_up;
546 		}
547 	}
548 	fault = 0;
549 out_up:
550 	up_read(&mm->mmap_sem);
551 out:
552 	return fault;
553 }
554 
555 void do_protection_exception(struct pt_regs *regs)
556 {
557 	unsigned long trans_exc_code;
558 	int access;
559 	vm_fault_t fault;
560 
561 	trans_exc_code = regs->int_parm_long;
562 	/*
563 	 * Protection exceptions are suppressing, decrement psw address.
564 	 * The exception to this rule are aborted transactions, for these
565 	 * the PSW already points to the correct location.
566 	 */
567 	if (!(regs->int_code & 0x200))
568 		regs->psw.addr = __rewind_psw(regs->psw, regs->int_code >> 16);
569 	/*
570 	 * Check for low-address protection.  This needs to be treated
571 	 * as a special case because the translation exception code
572 	 * field is not guaranteed to contain valid data in this case.
573 	 */
574 	if (unlikely(!(trans_exc_code & 4))) {
575 		do_low_address(regs);
576 		return;
577 	}
578 	if (unlikely(MACHINE_HAS_NX && (trans_exc_code & 0x80))) {
579 		regs->int_parm_long = (trans_exc_code & ~PAGE_MASK) |
580 					(regs->psw.addr & PAGE_MASK);
581 		access = VM_EXEC;
582 		fault = VM_FAULT_BADACCESS;
583 	} else {
584 		access = VM_WRITE;
585 		fault = do_exception(regs, access);
586 	}
587 	if (unlikely(fault))
588 		do_fault_error(regs, access, fault);
589 }
590 NOKPROBE_SYMBOL(do_protection_exception);
591 
592 void do_dat_exception(struct pt_regs *regs)
593 {
594 	int access;
595 	vm_fault_t fault;
596 
597 	access = VM_READ | VM_EXEC | VM_WRITE;
598 	fault = do_exception(regs, access);
599 	if (unlikely(fault))
600 		do_fault_error(regs, access, fault);
601 }
602 NOKPROBE_SYMBOL(do_dat_exception);
603 
604 #ifdef CONFIG_PFAULT
605 /*
606  * 'pfault' pseudo page faults routines.
607  */
608 static int pfault_disable;
609 
610 static int __init nopfault(char *str)
611 {
612 	pfault_disable = 1;
613 	return 1;
614 }
615 
616 __setup("nopfault", nopfault);
617 
618 struct pfault_refbk {
619 	u16 refdiagc;
620 	u16 reffcode;
621 	u16 refdwlen;
622 	u16 refversn;
623 	u64 refgaddr;
624 	u64 refselmk;
625 	u64 refcmpmk;
626 	u64 reserved;
627 } __attribute__ ((packed, aligned(8)));
628 
629 static struct pfault_refbk pfault_init_refbk = {
630 	.refdiagc = 0x258,
631 	.reffcode = 0,
632 	.refdwlen = 5,
633 	.refversn = 2,
634 	.refgaddr = __LC_LPP,
635 	.refselmk = 1ULL << 48,
636 	.refcmpmk = 1ULL << 48,
637 	.reserved = __PF_RES_FIELD
638 };
639 
640 int pfault_init(void)
641 {
642         int rc;
643 
644 	if (pfault_disable)
645 		return -1;
646 	diag_stat_inc(DIAG_STAT_X258);
647 	asm volatile(
648 		"	diag	%1,%0,0x258\n"
649 		"0:	j	2f\n"
650 		"1:	la	%0,8\n"
651 		"2:\n"
652 		EX_TABLE(0b,1b)
653 		: "=d" (rc)
654 		: "a" (&pfault_init_refbk), "m" (pfault_init_refbk) : "cc");
655         return rc;
656 }
657 
658 static struct pfault_refbk pfault_fini_refbk = {
659 	.refdiagc = 0x258,
660 	.reffcode = 1,
661 	.refdwlen = 5,
662 	.refversn = 2,
663 };
664 
665 void pfault_fini(void)
666 {
667 
668 	if (pfault_disable)
669 		return;
670 	diag_stat_inc(DIAG_STAT_X258);
671 	asm volatile(
672 		"	diag	%0,0,0x258\n"
673 		"0:	nopr	%%r7\n"
674 		EX_TABLE(0b,0b)
675 		: : "a" (&pfault_fini_refbk), "m" (pfault_fini_refbk) : "cc");
676 }
677 
678 static DEFINE_SPINLOCK(pfault_lock);
679 static LIST_HEAD(pfault_list);
680 
681 #define PF_COMPLETE	0x0080
682 
683 /*
684  * The mechanism of our pfault code: if Linux is running as guest, runs a user
685  * space process and the user space process accesses a page that the host has
686  * paged out we get a pfault interrupt.
687  *
688  * This allows us, within the guest, to schedule a different process. Without
689  * this mechanism the host would have to suspend the whole virtual cpu until
690  * the page has been paged in.
691  *
692  * So when we get such an interrupt then we set the state of the current task
693  * to uninterruptible and also set the need_resched flag. Both happens within
694  * interrupt context(!). If we later on want to return to user space we
695  * recognize the need_resched flag and then call schedule().  It's not very
696  * obvious how this works...
697  *
698  * Of course we have a lot of additional fun with the completion interrupt (->
699  * host signals that a page of a process has been paged in and the process can
700  * continue to run). This interrupt can arrive on any cpu and, since we have
701  * virtual cpus, actually appear before the interrupt that signals that a page
702  * is missing.
703  */
704 static void pfault_interrupt(struct ext_code ext_code,
705 			     unsigned int param32, unsigned long param64)
706 {
707 	struct task_struct *tsk;
708 	__u16 subcode;
709 	pid_t pid;
710 
711 	/*
712 	 * Get the external interruption subcode & pfault initial/completion
713 	 * signal bit. VM stores this in the 'cpu address' field associated
714 	 * with the external interrupt.
715 	 */
716 	subcode = ext_code.subcode;
717 	if ((subcode & 0xff00) != __SUBCODE_MASK)
718 		return;
719 	inc_irq_stat(IRQEXT_PFL);
720 	/* Get the token (= pid of the affected task). */
721 	pid = param64 & LPP_PID_MASK;
722 	rcu_read_lock();
723 	tsk = find_task_by_pid_ns(pid, &init_pid_ns);
724 	if (tsk)
725 		get_task_struct(tsk);
726 	rcu_read_unlock();
727 	if (!tsk)
728 		return;
729 	spin_lock(&pfault_lock);
730 	if (subcode & PF_COMPLETE) {
731 		/* signal bit is set -> a page has been swapped in by VM */
732 		if (tsk->thread.pfault_wait == 1) {
733 			/* Initial interrupt was faster than the completion
734 			 * interrupt. pfault_wait is valid. Set pfault_wait
735 			 * back to zero and wake up the process. This can
736 			 * safely be done because the task is still sleeping
737 			 * and can't produce new pfaults. */
738 			tsk->thread.pfault_wait = 0;
739 			list_del(&tsk->thread.list);
740 			wake_up_process(tsk);
741 			put_task_struct(tsk);
742 		} else {
743 			/* Completion interrupt was faster than initial
744 			 * interrupt. Set pfault_wait to -1 so the initial
745 			 * interrupt doesn't put the task to sleep.
746 			 * If the task is not running, ignore the completion
747 			 * interrupt since it must be a leftover of a PFAULT
748 			 * CANCEL operation which didn't remove all pending
749 			 * completion interrupts. */
750 			if (tsk->state == TASK_RUNNING)
751 				tsk->thread.pfault_wait = -1;
752 		}
753 	} else {
754 		/* signal bit not set -> a real page is missing. */
755 		if (WARN_ON_ONCE(tsk != current))
756 			goto out;
757 		if (tsk->thread.pfault_wait == 1) {
758 			/* Already on the list with a reference: put to sleep */
759 			goto block;
760 		} else if (tsk->thread.pfault_wait == -1) {
761 			/* Completion interrupt was faster than the initial
762 			 * interrupt (pfault_wait == -1). Set pfault_wait
763 			 * back to zero and exit. */
764 			tsk->thread.pfault_wait = 0;
765 		} else {
766 			/* Initial interrupt arrived before completion
767 			 * interrupt. Let the task sleep.
768 			 * An extra task reference is needed since a different
769 			 * cpu may set the task state to TASK_RUNNING again
770 			 * before the scheduler is reached. */
771 			get_task_struct(tsk);
772 			tsk->thread.pfault_wait = 1;
773 			list_add(&tsk->thread.list, &pfault_list);
774 block:
775 			/* Since this must be a userspace fault, there
776 			 * is no kernel task state to trample. Rely on the
777 			 * return to userspace schedule() to block. */
778 			__set_current_state(TASK_UNINTERRUPTIBLE);
779 			set_tsk_need_resched(tsk);
780 			set_preempt_need_resched();
781 		}
782 	}
783 out:
784 	spin_unlock(&pfault_lock);
785 	put_task_struct(tsk);
786 }
787 
788 static int pfault_cpu_dead(unsigned int cpu)
789 {
790 	struct thread_struct *thread, *next;
791 	struct task_struct *tsk;
792 
793 	spin_lock_irq(&pfault_lock);
794 	list_for_each_entry_safe(thread, next, &pfault_list, list) {
795 		thread->pfault_wait = 0;
796 		list_del(&thread->list);
797 		tsk = container_of(thread, struct task_struct, thread);
798 		wake_up_process(tsk);
799 		put_task_struct(tsk);
800 	}
801 	spin_unlock_irq(&pfault_lock);
802 	return 0;
803 }
804 
805 static int __init pfault_irq_init(void)
806 {
807 	int rc;
808 
809 	rc = register_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
810 	if (rc)
811 		goto out_extint;
812 	rc = pfault_init() == 0 ? 0 : -EOPNOTSUPP;
813 	if (rc)
814 		goto out_pfault;
815 	irq_subclass_register(IRQ_SUBCLASS_SERVICE_SIGNAL);
816 	cpuhp_setup_state_nocalls(CPUHP_S390_PFAULT_DEAD, "s390/pfault:dead",
817 				  NULL, pfault_cpu_dead);
818 	return 0;
819 
820 out_pfault:
821 	unregister_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
822 out_extint:
823 	pfault_disable = 1;
824 	return rc;
825 }
826 early_initcall(pfault_irq_init);
827 
828 #endif /* CONFIG_PFAULT */
829