xref: /openbmc/linux/arch/s390/mm/fault.c (revision faa16bc4)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  *  S390 version
4  *    Copyright IBM Corp. 1999
5  *    Author(s): Hartmut Penner (hp@de.ibm.com)
6  *               Ulrich Weigand (uweigand@de.ibm.com)
7  *
8  *  Derived from "arch/i386/mm/fault.c"
9  *    Copyright (C) 1995  Linus Torvalds
10  */
11 
12 #include <linux/kernel_stat.h>
13 #include <linux/perf_event.h>
14 #include <linux/signal.h>
15 #include <linux/sched.h>
16 #include <linux/sched/debug.h>
17 #include <linux/kernel.h>
18 #include <linux/errno.h>
19 #include <linux/string.h>
20 #include <linux/types.h>
21 #include <linux/ptrace.h>
22 #include <linux/mman.h>
23 #include <linux/mm.h>
24 #include <linux/compat.h>
25 #include <linux/smp.h>
26 #include <linux/kdebug.h>
27 #include <linux/init.h>
28 #include <linux/console.h>
29 #include <linux/extable.h>
30 #include <linux/hardirq.h>
31 #include <linux/kprobes.h>
32 #include <linux/uaccess.h>
33 #include <linux/hugetlb.h>
34 #include <asm/asm-offsets.h>
35 #include <asm/diag.h>
36 #include <asm/pgtable.h>
37 #include <asm/gmap.h>
38 #include <asm/irq.h>
39 #include <asm/mmu_context.h>
40 #include <asm/facility.h>
41 #include "../kernel/entry.h"
42 
43 #define __FAIL_ADDR_MASK -4096L
44 #define __SUBCODE_MASK 0x0600
45 #define __PF_RES_FIELD 0x8000000000000000ULL
46 
47 #define VM_FAULT_BADCONTEXT	0x010000
48 #define VM_FAULT_BADMAP		0x020000
49 #define VM_FAULT_BADACCESS	0x040000
50 #define VM_FAULT_SIGNAL		0x080000
51 #define VM_FAULT_PFAULT		0x100000
52 
53 enum fault_type {
54 	KERNEL_FAULT,
55 	USER_FAULT,
56 	VDSO_FAULT,
57 	GMAP_FAULT,
58 };
59 
60 static unsigned long store_indication __read_mostly;
61 
62 static int __init fault_init(void)
63 {
64 	if (test_facility(75))
65 		store_indication = 0xc00;
66 	return 0;
67 }
68 early_initcall(fault_init);
69 
70 static inline int notify_page_fault(struct pt_regs *regs)
71 {
72 	int ret = 0;
73 
74 	/* kprobe_running() needs smp_processor_id() */
75 	if (kprobes_built_in() && !user_mode(regs)) {
76 		preempt_disable();
77 		if (kprobe_running() && kprobe_fault_handler(regs, 14))
78 			ret = 1;
79 		preempt_enable();
80 	}
81 	return ret;
82 }
83 
84 
85 /*
86  * Unlock any spinlocks which will prevent us from getting the
87  * message out.
88  */
89 void bust_spinlocks(int yes)
90 {
91 	if (yes) {
92 		oops_in_progress = 1;
93 	} else {
94 		int loglevel_save = console_loglevel;
95 		console_unblank();
96 		oops_in_progress = 0;
97 		/*
98 		 * OK, the message is on the console.  Now we call printk()
99 		 * without oops_in_progress set so that printk will give klogd
100 		 * a poke.  Hold onto your hats...
101 		 */
102 		console_loglevel = 15;
103 		printk(" ");
104 		console_loglevel = loglevel_save;
105 	}
106 }
107 
108 /*
109  * Find out which address space caused the exception.
110  * Access register mode is impossible, ignore space == 3.
111  */
112 static inline enum fault_type get_fault_type(struct pt_regs *regs)
113 {
114 	unsigned long trans_exc_code;
115 
116 	trans_exc_code = regs->int_parm_long & 3;
117 	if (likely(trans_exc_code == 0)) {
118 		/* primary space exception */
119 		if (IS_ENABLED(CONFIG_PGSTE) &&
120 		    test_pt_regs_flag(regs, PIF_GUEST_FAULT))
121 			return GMAP_FAULT;
122 		if (current->thread.mm_segment == USER_DS)
123 			return USER_FAULT;
124 		return KERNEL_FAULT;
125 	}
126 	if (trans_exc_code == 2) {
127 		/* secondary space exception */
128 		if (current->thread.mm_segment & 1) {
129 			if (current->thread.mm_segment == USER_DS_SACF)
130 				return USER_FAULT;
131 			return KERNEL_FAULT;
132 		}
133 		return VDSO_FAULT;
134 	}
135 	/* home space exception -> access via kernel ASCE */
136 	return KERNEL_FAULT;
137 }
138 
139 static int bad_address(void *p)
140 {
141 	unsigned long dummy;
142 
143 	return probe_kernel_address((unsigned long *)p, dummy);
144 }
145 
146 static void dump_pagetable(unsigned long asce, unsigned long address)
147 {
148 	unsigned long *table = __va(asce & _ASCE_ORIGIN);
149 
150 	pr_alert("AS:%016lx ", asce);
151 	switch (asce & _ASCE_TYPE_MASK) {
152 	case _ASCE_TYPE_REGION1:
153 		table += (address & _REGION1_INDEX) >> _REGION1_SHIFT;
154 		if (bad_address(table))
155 			goto bad;
156 		pr_cont("R1:%016lx ", *table);
157 		if (*table & _REGION_ENTRY_INVALID)
158 			goto out;
159 		table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
160 		/* fallthrough */
161 	case _ASCE_TYPE_REGION2:
162 		table += (address & _REGION2_INDEX) >> _REGION2_SHIFT;
163 		if (bad_address(table))
164 			goto bad;
165 		pr_cont("R2:%016lx ", *table);
166 		if (*table & _REGION_ENTRY_INVALID)
167 			goto out;
168 		table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
169 		/* fallthrough */
170 	case _ASCE_TYPE_REGION3:
171 		table += (address & _REGION3_INDEX) >> _REGION3_SHIFT;
172 		if (bad_address(table))
173 			goto bad;
174 		pr_cont("R3:%016lx ", *table);
175 		if (*table & (_REGION_ENTRY_INVALID | _REGION3_ENTRY_LARGE))
176 			goto out;
177 		table = (unsigned long *)(*table & _REGION_ENTRY_ORIGIN);
178 		/* fallthrough */
179 	case _ASCE_TYPE_SEGMENT:
180 		table += (address & _SEGMENT_INDEX) >> _SEGMENT_SHIFT;
181 		if (bad_address(table))
182 			goto bad;
183 		pr_cont("S:%016lx ", *table);
184 		if (*table & (_SEGMENT_ENTRY_INVALID | _SEGMENT_ENTRY_LARGE))
185 			goto out;
186 		table = (unsigned long *)(*table & _SEGMENT_ENTRY_ORIGIN);
187 	}
188 	table += (address & _PAGE_INDEX) >> _PAGE_SHIFT;
189 	if (bad_address(table))
190 		goto bad;
191 	pr_cont("P:%016lx ", *table);
192 out:
193 	pr_cont("\n");
194 	return;
195 bad:
196 	pr_cont("BAD\n");
197 }
198 
199 static void dump_fault_info(struct pt_regs *regs)
200 {
201 	unsigned long asce;
202 
203 	pr_alert("Failing address: %016lx TEID: %016lx\n",
204 		 regs->int_parm_long & __FAIL_ADDR_MASK, regs->int_parm_long);
205 	pr_alert("Fault in ");
206 	switch (regs->int_parm_long & 3) {
207 	case 3:
208 		pr_cont("home space ");
209 		break;
210 	case 2:
211 		pr_cont("secondary space ");
212 		break;
213 	case 1:
214 		pr_cont("access register ");
215 		break;
216 	case 0:
217 		pr_cont("primary space ");
218 		break;
219 	}
220 	pr_cont("mode while using ");
221 	switch (get_fault_type(regs)) {
222 	case USER_FAULT:
223 		asce = S390_lowcore.user_asce;
224 		pr_cont("user ");
225 		break;
226 	case VDSO_FAULT:
227 		asce = S390_lowcore.vdso_asce;
228 		pr_cont("vdso ");
229 		break;
230 	case GMAP_FAULT:
231 		asce = ((struct gmap *) S390_lowcore.gmap)->asce;
232 		pr_cont("gmap ");
233 		break;
234 	case KERNEL_FAULT:
235 		asce = S390_lowcore.kernel_asce;
236 		pr_cont("kernel ");
237 		break;
238 	}
239 	pr_cont("ASCE.\n");
240 	dump_pagetable(asce, regs->int_parm_long & __FAIL_ADDR_MASK);
241 }
242 
243 int show_unhandled_signals = 1;
244 
245 void report_user_fault(struct pt_regs *regs, long signr, int is_mm_fault)
246 {
247 	if ((task_pid_nr(current) > 1) && !show_unhandled_signals)
248 		return;
249 	if (!unhandled_signal(current, signr))
250 		return;
251 	if (!printk_ratelimit())
252 		return;
253 	printk(KERN_ALERT "User process fault: interruption code %04x ilc:%d ",
254 	       regs->int_code & 0xffff, regs->int_code >> 17);
255 	print_vma_addr(KERN_CONT "in ", regs->psw.addr);
256 	printk(KERN_CONT "\n");
257 	if (is_mm_fault)
258 		dump_fault_info(regs);
259 	show_regs(regs);
260 }
261 
262 /*
263  * Send SIGSEGV to task.  This is an external routine
264  * to keep the stack usage of do_page_fault small.
265  */
266 static noinline void do_sigsegv(struct pt_regs *regs, int si_code)
267 {
268 	report_user_fault(regs, SIGSEGV, 1);
269 	force_sig_fault(SIGSEGV, si_code,
270 			(void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK),
271 			current);
272 }
273 
274 static noinline void do_no_context(struct pt_regs *regs)
275 {
276 	const struct exception_table_entry *fixup;
277 
278 	/* Are we prepared to handle this kernel fault?  */
279 	fixup = search_exception_tables(regs->psw.addr);
280 	if (fixup) {
281 		regs->psw.addr = extable_fixup(fixup);
282 		return;
283 	}
284 
285 	/*
286 	 * Oops. The kernel tried to access some bad page. We'll have to
287 	 * terminate things with extreme prejudice.
288 	 */
289 	if (get_fault_type(regs) == KERNEL_FAULT)
290 		printk(KERN_ALERT "Unable to handle kernel pointer dereference"
291 		       " in virtual kernel address space\n");
292 	else
293 		printk(KERN_ALERT "Unable to handle kernel paging request"
294 		       " in virtual user address space\n");
295 	dump_fault_info(regs);
296 	die(regs, "Oops");
297 	do_exit(SIGKILL);
298 }
299 
300 static noinline void do_low_address(struct pt_regs *regs)
301 {
302 	/* Low-address protection hit in kernel mode means
303 	   NULL pointer write access in kernel mode.  */
304 	if (regs->psw.mask & PSW_MASK_PSTATE) {
305 		/* Low-address protection hit in user mode 'cannot happen'. */
306 		die (regs, "Low-address protection");
307 		do_exit(SIGKILL);
308 	}
309 
310 	do_no_context(regs);
311 }
312 
313 static noinline void do_sigbus(struct pt_regs *regs)
314 {
315 	/*
316 	 * Send a sigbus, regardless of whether we were in kernel
317 	 * or user mode.
318 	 */
319 	force_sig_fault(SIGBUS, BUS_ADRERR,
320 			(void __user *)(regs->int_parm_long & __FAIL_ADDR_MASK),
321 			current);
322 }
323 
324 static noinline int signal_return(struct pt_regs *regs)
325 {
326 	u16 instruction;
327 	int rc;
328 
329 	rc = __get_user(instruction, (u16 __user *) regs->psw.addr);
330 	if (rc)
331 		return rc;
332 	if (instruction == 0x0a77) {
333 		set_pt_regs_flag(regs, PIF_SYSCALL);
334 		regs->int_code = 0x00040077;
335 		return 0;
336 	} else if (instruction == 0x0aad) {
337 		set_pt_regs_flag(regs, PIF_SYSCALL);
338 		regs->int_code = 0x000400ad;
339 		return 0;
340 	}
341 	return -EACCES;
342 }
343 
344 static noinline void do_fault_error(struct pt_regs *regs, int access, int fault)
345 {
346 	int si_code;
347 
348 	switch (fault) {
349 	case VM_FAULT_BADACCESS:
350 		if (access == VM_EXEC && signal_return(regs) == 0)
351 			break;
352 	case VM_FAULT_BADMAP:
353 		/* Bad memory access. Check if it is kernel or user space. */
354 		if (user_mode(regs)) {
355 			/* User mode accesses just cause a SIGSEGV */
356 			si_code = (fault == VM_FAULT_BADMAP) ?
357 				SEGV_MAPERR : SEGV_ACCERR;
358 			do_sigsegv(regs, si_code);
359 			break;
360 		}
361 	case VM_FAULT_BADCONTEXT:
362 	case VM_FAULT_PFAULT:
363 		do_no_context(regs);
364 		break;
365 	case VM_FAULT_SIGNAL:
366 		if (!user_mode(regs))
367 			do_no_context(regs);
368 		break;
369 	default: /* fault & VM_FAULT_ERROR */
370 		if (fault & VM_FAULT_OOM) {
371 			if (!user_mode(regs))
372 				do_no_context(regs);
373 			else
374 				pagefault_out_of_memory();
375 		} else if (fault & VM_FAULT_SIGSEGV) {
376 			/* Kernel mode? Handle exceptions or die */
377 			if (!user_mode(regs))
378 				do_no_context(regs);
379 			else
380 				do_sigsegv(regs, SEGV_MAPERR);
381 		} else if (fault & VM_FAULT_SIGBUS) {
382 			/* Kernel mode? Handle exceptions or die */
383 			if (!user_mode(regs))
384 				do_no_context(regs);
385 			else
386 				do_sigbus(regs);
387 		} else
388 			BUG();
389 		break;
390 	}
391 }
392 
393 /*
394  * This routine handles page faults.  It determines the address,
395  * and the problem, and then passes it off to one of the appropriate
396  * routines.
397  *
398  * interruption code (int_code):
399  *   04       Protection           ->  Write-Protection  (suprression)
400  *   10       Segment translation  ->  Not present       (nullification)
401  *   11       Page translation     ->  Not present       (nullification)
402  *   3b       Region third trans.  ->  Not present       (nullification)
403  */
404 static inline int do_exception(struct pt_regs *regs, int access)
405 {
406 	struct gmap *gmap;
407 	struct task_struct *tsk;
408 	struct mm_struct *mm;
409 	struct vm_area_struct *vma;
410 	enum fault_type type;
411 	unsigned long trans_exc_code;
412 	unsigned long address;
413 	unsigned int flags;
414 	int fault;
415 
416 	tsk = current;
417 	/*
418 	 * The instruction that caused the program check has
419 	 * been nullified. Don't signal single step via SIGTRAP.
420 	 */
421 	clear_pt_regs_flag(regs, PIF_PER_TRAP);
422 
423 	if (notify_page_fault(regs))
424 		return 0;
425 
426 	mm = tsk->mm;
427 	trans_exc_code = regs->int_parm_long;
428 
429 	/*
430 	 * Verify that the fault happened in user space, that
431 	 * we are not in an interrupt and that there is a
432 	 * user context.
433 	 */
434 	fault = VM_FAULT_BADCONTEXT;
435 	type = get_fault_type(regs);
436 	switch (type) {
437 	case KERNEL_FAULT:
438 		goto out;
439 	case VDSO_FAULT:
440 		fault = VM_FAULT_BADMAP;
441 		goto out;
442 	case USER_FAULT:
443 	case GMAP_FAULT:
444 		if (faulthandler_disabled() || !mm)
445 			goto out;
446 		break;
447 	}
448 
449 	address = trans_exc_code & __FAIL_ADDR_MASK;
450 	perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
451 	flags = FAULT_FLAG_ALLOW_RETRY | FAULT_FLAG_KILLABLE;
452 	if (user_mode(regs))
453 		flags |= FAULT_FLAG_USER;
454 	if (access == VM_WRITE || (trans_exc_code & store_indication) == 0x400)
455 		flags |= FAULT_FLAG_WRITE;
456 	down_read(&mm->mmap_sem);
457 
458 	gmap = NULL;
459 	if (IS_ENABLED(CONFIG_PGSTE) && type == GMAP_FAULT) {
460 		gmap = (struct gmap *) S390_lowcore.gmap;
461 		current->thread.gmap_addr = address;
462 		current->thread.gmap_write_flag = !!(flags & FAULT_FLAG_WRITE);
463 		current->thread.gmap_int_code = regs->int_code & 0xffff;
464 		address = __gmap_translate(gmap, address);
465 		if (address == -EFAULT) {
466 			fault = VM_FAULT_BADMAP;
467 			goto out_up;
468 		}
469 		if (gmap->pfault_enabled)
470 			flags |= FAULT_FLAG_RETRY_NOWAIT;
471 	}
472 
473 retry:
474 	fault = VM_FAULT_BADMAP;
475 	vma = find_vma(mm, address);
476 	if (!vma)
477 		goto out_up;
478 
479 	if (unlikely(vma->vm_start > address)) {
480 		if (!(vma->vm_flags & VM_GROWSDOWN))
481 			goto out_up;
482 		if (expand_stack(vma, address))
483 			goto out_up;
484 	}
485 
486 	/*
487 	 * Ok, we have a good vm_area for this memory access, so
488 	 * we can handle it..
489 	 */
490 	fault = VM_FAULT_BADACCESS;
491 	if (unlikely(!(vma->vm_flags & access)))
492 		goto out_up;
493 
494 	if (is_vm_hugetlb_page(vma))
495 		address &= HPAGE_MASK;
496 	/*
497 	 * If for any reason at all we couldn't handle the fault,
498 	 * make sure we exit gracefully rather than endlessly redo
499 	 * the fault.
500 	 */
501 	fault = handle_mm_fault(vma, address, flags);
502 	/* No reason to continue if interrupted by SIGKILL. */
503 	if ((fault & VM_FAULT_RETRY) && fatal_signal_pending(current)) {
504 		fault = VM_FAULT_SIGNAL;
505 		goto out;
506 	}
507 	if (unlikely(fault & VM_FAULT_ERROR))
508 		goto out_up;
509 
510 	/*
511 	 * Major/minor page fault accounting is only done on the
512 	 * initial attempt. If we go through a retry, it is extremely
513 	 * likely that the page will be found in page cache at that point.
514 	 */
515 	if (flags & FAULT_FLAG_ALLOW_RETRY) {
516 		if (fault & VM_FAULT_MAJOR) {
517 			tsk->maj_flt++;
518 			perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MAJ, 1,
519 				      regs, address);
520 		} else {
521 			tsk->min_flt++;
522 			perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS_MIN, 1,
523 				      regs, address);
524 		}
525 		if (fault & VM_FAULT_RETRY) {
526 			if (IS_ENABLED(CONFIG_PGSTE) && gmap &&
527 			    (flags & FAULT_FLAG_RETRY_NOWAIT)) {
528 				/* FAULT_FLAG_RETRY_NOWAIT has been set,
529 				 * mmap_sem has not been released */
530 				current->thread.gmap_pfault = 1;
531 				fault = VM_FAULT_PFAULT;
532 				goto out_up;
533 			}
534 			/* Clear FAULT_FLAG_ALLOW_RETRY to avoid any risk
535 			 * of starvation. */
536 			flags &= ~(FAULT_FLAG_ALLOW_RETRY |
537 				   FAULT_FLAG_RETRY_NOWAIT);
538 			flags |= FAULT_FLAG_TRIED;
539 			down_read(&mm->mmap_sem);
540 			goto retry;
541 		}
542 	}
543 	if (IS_ENABLED(CONFIG_PGSTE) && gmap) {
544 		address =  __gmap_link(gmap, current->thread.gmap_addr,
545 				       address);
546 		if (address == -EFAULT) {
547 			fault = VM_FAULT_BADMAP;
548 			goto out_up;
549 		}
550 		if (address == -ENOMEM) {
551 			fault = VM_FAULT_OOM;
552 			goto out_up;
553 		}
554 	}
555 	fault = 0;
556 out_up:
557 	up_read(&mm->mmap_sem);
558 out:
559 	return fault;
560 }
561 
562 void do_protection_exception(struct pt_regs *regs)
563 {
564 	unsigned long trans_exc_code;
565 	int access, fault;
566 
567 	trans_exc_code = regs->int_parm_long;
568 	/*
569 	 * Protection exceptions are suppressing, decrement psw address.
570 	 * The exception to this rule are aborted transactions, for these
571 	 * the PSW already points to the correct location.
572 	 */
573 	if (!(regs->int_code & 0x200))
574 		regs->psw.addr = __rewind_psw(regs->psw, regs->int_code >> 16);
575 	/*
576 	 * Check for low-address protection.  This needs to be treated
577 	 * as a special case because the translation exception code
578 	 * field is not guaranteed to contain valid data in this case.
579 	 */
580 	if (unlikely(!(trans_exc_code & 4))) {
581 		do_low_address(regs);
582 		return;
583 	}
584 	if (unlikely(MACHINE_HAS_NX && (trans_exc_code & 0x80))) {
585 		regs->int_parm_long = (trans_exc_code & ~PAGE_MASK) |
586 					(regs->psw.addr & PAGE_MASK);
587 		access = VM_EXEC;
588 		fault = VM_FAULT_BADACCESS;
589 	} else {
590 		access = VM_WRITE;
591 		fault = do_exception(regs, access);
592 	}
593 	if (unlikely(fault))
594 		do_fault_error(regs, access, fault);
595 }
596 NOKPROBE_SYMBOL(do_protection_exception);
597 
598 void do_dat_exception(struct pt_regs *regs)
599 {
600 	int access, fault;
601 
602 	access = VM_READ | VM_EXEC | VM_WRITE;
603 	fault = do_exception(regs, access);
604 	if (unlikely(fault))
605 		do_fault_error(regs, access, fault);
606 }
607 NOKPROBE_SYMBOL(do_dat_exception);
608 
609 #ifdef CONFIG_PFAULT
610 /*
611  * 'pfault' pseudo page faults routines.
612  */
613 static int pfault_disable;
614 
615 static int __init nopfault(char *str)
616 {
617 	pfault_disable = 1;
618 	return 1;
619 }
620 
621 __setup("nopfault", nopfault);
622 
623 struct pfault_refbk {
624 	u16 refdiagc;
625 	u16 reffcode;
626 	u16 refdwlen;
627 	u16 refversn;
628 	u64 refgaddr;
629 	u64 refselmk;
630 	u64 refcmpmk;
631 	u64 reserved;
632 } __attribute__ ((packed, aligned(8)));
633 
634 int pfault_init(void)
635 {
636 	struct pfault_refbk refbk = {
637 		.refdiagc = 0x258,
638 		.reffcode = 0,
639 		.refdwlen = 5,
640 		.refversn = 2,
641 		.refgaddr = __LC_LPP,
642 		.refselmk = 1ULL << 48,
643 		.refcmpmk = 1ULL << 48,
644 		.reserved = __PF_RES_FIELD };
645         int rc;
646 
647 	if (pfault_disable)
648 		return -1;
649 	diag_stat_inc(DIAG_STAT_X258);
650 	asm volatile(
651 		"	diag	%1,%0,0x258\n"
652 		"0:	j	2f\n"
653 		"1:	la	%0,8\n"
654 		"2:\n"
655 		EX_TABLE(0b,1b)
656 		: "=d" (rc) : "a" (&refbk), "m" (refbk) : "cc");
657         return rc;
658 }
659 
660 void pfault_fini(void)
661 {
662 	struct pfault_refbk refbk = {
663 		.refdiagc = 0x258,
664 		.reffcode = 1,
665 		.refdwlen = 5,
666 		.refversn = 2,
667 	};
668 
669 	if (pfault_disable)
670 		return;
671 	diag_stat_inc(DIAG_STAT_X258);
672 	asm volatile(
673 		"	diag	%0,0,0x258\n"
674 		"0:	nopr	%%r7\n"
675 		EX_TABLE(0b,0b)
676 		: : "a" (&refbk), "m" (refbk) : "cc");
677 }
678 
679 static DEFINE_SPINLOCK(pfault_lock);
680 static LIST_HEAD(pfault_list);
681 
682 #define PF_COMPLETE	0x0080
683 
684 /*
685  * The mechanism of our pfault code: if Linux is running as guest, runs a user
686  * space process and the user space process accesses a page that the host has
687  * paged out we get a pfault interrupt.
688  *
689  * This allows us, within the guest, to schedule a different process. Without
690  * this mechanism the host would have to suspend the whole virtual cpu until
691  * the page has been paged in.
692  *
693  * So when we get such an interrupt then we set the state of the current task
694  * to uninterruptible and also set the need_resched flag. Both happens within
695  * interrupt context(!). If we later on want to return to user space we
696  * recognize the need_resched flag and then call schedule().  It's not very
697  * obvious how this works...
698  *
699  * Of course we have a lot of additional fun with the completion interrupt (->
700  * host signals that a page of a process has been paged in and the process can
701  * continue to run). This interrupt can arrive on any cpu and, since we have
702  * virtual cpus, actually appear before the interrupt that signals that a page
703  * is missing.
704  */
705 static void pfault_interrupt(struct ext_code ext_code,
706 			     unsigned int param32, unsigned long param64)
707 {
708 	struct task_struct *tsk;
709 	__u16 subcode;
710 	pid_t pid;
711 
712 	/*
713 	 * Get the external interruption subcode & pfault initial/completion
714 	 * signal bit. VM stores this in the 'cpu address' field associated
715 	 * with the external interrupt.
716 	 */
717 	subcode = ext_code.subcode;
718 	if ((subcode & 0xff00) != __SUBCODE_MASK)
719 		return;
720 	inc_irq_stat(IRQEXT_PFL);
721 	/* Get the token (= pid of the affected task). */
722 	pid = param64 & LPP_PID_MASK;
723 	rcu_read_lock();
724 	tsk = find_task_by_pid_ns(pid, &init_pid_ns);
725 	if (tsk)
726 		get_task_struct(tsk);
727 	rcu_read_unlock();
728 	if (!tsk)
729 		return;
730 	spin_lock(&pfault_lock);
731 	if (subcode & PF_COMPLETE) {
732 		/* signal bit is set -> a page has been swapped in by VM */
733 		if (tsk->thread.pfault_wait == 1) {
734 			/* Initial interrupt was faster than the completion
735 			 * interrupt. pfault_wait is valid. Set pfault_wait
736 			 * back to zero and wake up the process. This can
737 			 * safely be done because the task is still sleeping
738 			 * and can't produce new pfaults. */
739 			tsk->thread.pfault_wait = 0;
740 			list_del(&tsk->thread.list);
741 			wake_up_process(tsk);
742 			put_task_struct(tsk);
743 		} else {
744 			/* Completion interrupt was faster than initial
745 			 * interrupt. Set pfault_wait to -1 so the initial
746 			 * interrupt doesn't put the task to sleep.
747 			 * If the task is not running, ignore the completion
748 			 * interrupt since it must be a leftover of a PFAULT
749 			 * CANCEL operation which didn't remove all pending
750 			 * completion interrupts. */
751 			if (tsk->state == TASK_RUNNING)
752 				tsk->thread.pfault_wait = -1;
753 		}
754 	} else {
755 		/* signal bit not set -> a real page is missing. */
756 		if (WARN_ON_ONCE(tsk != current))
757 			goto out;
758 		if (tsk->thread.pfault_wait == 1) {
759 			/* Already on the list with a reference: put to sleep */
760 			goto block;
761 		} else if (tsk->thread.pfault_wait == -1) {
762 			/* Completion interrupt was faster than the initial
763 			 * interrupt (pfault_wait == -1). Set pfault_wait
764 			 * back to zero and exit. */
765 			tsk->thread.pfault_wait = 0;
766 		} else {
767 			/* Initial interrupt arrived before completion
768 			 * interrupt. Let the task sleep.
769 			 * An extra task reference is needed since a different
770 			 * cpu may set the task state to TASK_RUNNING again
771 			 * before the scheduler is reached. */
772 			get_task_struct(tsk);
773 			tsk->thread.pfault_wait = 1;
774 			list_add(&tsk->thread.list, &pfault_list);
775 block:
776 			/* Since this must be a userspace fault, there
777 			 * is no kernel task state to trample. Rely on the
778 			 * return to userspace schedule() to block. */
779 			__set_current_state(TASK_UNINTERRUPTIBLE);
780 			set_tsk_need_resched(tsk);
781 			set_preempt_need_resched();
782 		}
783 	}
784 out:
785 	spin_unlock(&pfault_lock);
786 	put_task_struct(tsk);
787 }
788 
789 static int pfault_cpu_dead(unsigned int cpu)
790 {
791 	struct thread_struct *thread, *next;
792 	struct task_struct *tsk;
793 
794 	spin_lock_irq(&pfault_lock);
795 	list_for_each_entry_safe(thread, next, &pfault_list, list) {
796 		thread->pfault_wait = 0;
797 		list_del(&thread->list);
798 		tsk = container_of(thread, struct task_struct, thread);
799 		wake_up_process(tsk);
800 		put_task_struct(tsk);
801 	}
802 	spin_unlock_irq(&pfault_lock);
803 	return 0;
804 }
805 
806 static int __init pfault_irq_init(void)
807 {
808 	int rc;
809 
810 	rc = register_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
811 	if (rc)
812 		goto out_extint;
813 	rc = pfault_init() == 0 ? 0 : -EOPNOTSUPP;
814 	if (rc)
815 		goto out_pfault;
816 	irq_subclass_register(IRQ_SUBCLASS_SERVICE_SIGNAL);
817 	cpuhp_setup_state_nocalls(CPUHP_S390_PFAULT_DEAD, "s390/pfault:dead",
818 				  NULL, pfault_cpu_dead);
819 	return 0;
820 
821 out_pfault:
822 	unregister_external_irq(EXT_IRQ_CP_SERVICE, pfault_interrupt);
823 out_extint:
824 	pfault_disable = 1;
825 	return rc;
826 }
827 early_initcall(pfault_irq_init);
828 
829 #endif /* CONFIG_PFAULT */
830