xref: /openbmc/linux/arch/parisc/mm/fault.c (revision b868a02e)
1 /*
2  * This file is subject to the terms and conditions of the GNU General Public
3  * License.  See the file "COPYING" in the main directory of this archive
4  * for more details.
5  *
6  *
7  * Copyright (C) 1995, 1996, 1997, 1998 by Ralf Baechle
8  * Copyright 1999 SuSE GmbH (Philipp Rumpf, prumpf@tux.org)
9  * Copyright 1999 Hewlett Packard Co.
10  *
11  */
12 
13 #include <linux/mm.h>
14 #include <linux/ptrace.h>
15 #include <linux/sched.h>
16 #include <linux/sched/debug.h>
17 #include <linux/interrupt.h>
18 #include <linux/extable.h>
19 #include <linux/uaccess.h>
20 #include <linux/hugetlb.h>
21 #include <linux/perf_event.h>
22 
23 #include <asm/traps.h>
24 
25 #define DEBUG_NATLB 0
26 
27 /* Various important other fields */
28 #define bit22set(x)		(x & 0x00000200)
29 #define bits23_25set(x)		(x & 0x000001c0)
30 #define isGraphicsFlushRead(x)	((x & 0xfc003fdf) == 0x04001a80)
31 				/* extended opcode is 0x6a */
32 
33 #define BITSSET		0x1c0	/* for identifying LDCW */
34 
35 
36 int show_unhandled_signals = 1;
37 
38 /*
39  * parisc_acctyp(unsigned int inst) --
40  *    Given a PA-RISC memory access instruction, determine if the
41  *    instruction would perform a memory read or memory write
42  *    operation.
43  *
44  *    This function assumes that the given instruction is a memory access
45  *    instruction (i.e. you should really only call it if you know that
46  *    the instruction has generated some sort of a memory access fault).
47  *
48  * Returns:
49  *   VM_READ  if read operation
50  *   VM_WRITE if write operation
51  *   VM_EXEC  if execute operation
52  */
53 unsigned long
54 parisc_acctyp(unsigned long code, unsigned int inst)
55 {
56 	if (code == 6 || code == 16)
57 	    return VM_EXEC;
58 
59 	switch (inst & 0xf0000000) {
60 	case 0x40000000: /* load */
61 	case 0x50000000: /* new load */
62 		return VM_READ;
63 
64 	case 0x60000000: /* store */
65 	case 0x70000000: /* new store */
66 		return VM_WRITE;
67 
68 	case 0x20000000: /* coproc */
69 	case 0x30000000: /* coproc2 */
70 		if (bit22set(inst))
71 			return VM_WRITE;
72 		fallthrough;
73 
74 	case 0x0: /* indexed/memory management */
75 		if (bit22set(inst)) {
76 			/*
77 			 * Check for the 'Graphics Flush Read' instruction.
78 			 * It resembles an FDC instruction, except for bits
79 			 * 20 and 21. Any combination other than zero will
80 			 * utilize the block mover functionality on some
81 			 * older PA-RISC platforms.  The case where a block
82 			 * move is performed from VM to graphics IO space
83 			 * should be treated as a READ.
84 			 *
85 			 * The significance of bits 20,21 in the FDC
86 			 * instruction is:
87 			 *
88 			 *   00  Flush data cache (normal instruction behavior)
89 			 *   01  Graphics flush write  (IO space -> VM)
90 			 *   10  Graphics flush read   (VM -> IO space)
91 			 *   11  Graphics flush read/write (VM <-> IO space)
92 			 */
93 			if (isGraphicsFlushRead(inst))
94 				return VM_READ;
95 			return VM_WRITE;
96 		} else {
97 			/*
98 			 * Check for LDCWX and LDCWS (semaphore instructions).
99 			 * If bits 23 through 25 are all 1's it is one of
100 			 * the above two instructions and is a write.
101 			 *
102 			 * Note: With the limited bits we are looking at,
103 			 * this will also catch PROBEW and PROBEWI. However,
104 			 * these should never get in here because they don't
105 			 * generate exceptions of the type:
106 			 *   Data TLB miss fault/data page fault
107 			 *   Data memory protection trap
108 			 */
109 			if (bits23_25set(inst) == BITSSET)
110 				return VM_WRITE;
111 		}
112 		return VM_READ; /* Default */
113 	}
114 	return VM_READ; /* Default */
115 }
116 
117 #undef bit22set
118 #undef bits23_25set
119 #undef isGraphicsFlushRead
120 #undef BITSSET
121 
122 
123 #if 0
124 /* This is the treewalk to find a vma which is the highest that has
125  * a start < addr.  We're using find_vma_prev instead right now, but
126  * we might want to use this at some point in the future.  Probably
127  * not, but I want it committed to CVS so I don't lose it :-)
128  */
129 			while (tree != vm_avl_empty) {
130 				if (tree->vm_start > addr) {
131 					tree = tree->vm_avl_left;
132 				} else {
133 					prev = tree;
134 					if (prev->vm_next == NULL)
135 						break;
136 					if (prev->vm_next->vm_start > addr)
137 						break;
138 					tree = tree->vm_avl_right;
139 				}
140 			}
141 #endif
142 
143 int fixup_exception(struct pt_regs *regs)
144 {
145 	const struct exception_table_entry *fix;
146 
147 	fix = search_exception_tables(regs->iaoq[0]);
148 	if (fix) {
149 		/*
150 		 * Fix up get_user() and put_user().
151 		 * ASM_EXCEPTIONTABLE_ENTRY_EFAULT() sets the least-significant
152 		 * bit in the relative address of the fixup routine to indicate
153 		 * that gr[ASM_EXCEPTIONTABLE_REG] should be loaded with
154 		 * -EFAULT to report a userspace access error.
155 		 */
156 		if (fix->fixup & 1) {
157 			regs->gr[ASM_EXCEPTIONTABLE_REG] = -EFAULT;
158 
159 			/* zero target register for get_user() */
160 			if (parisc_acctyp(0, regs->iir) == VM_READ) {
161 				int treg = regs->iir & 0x1f;
162 				BUG_ON(treg == 0);
163 				regs->gr[treg] = 0;
164 			}
165 		}
166 
167 		regs->iaoq[0] = (unsigned long)&fix->fixup + fix->fixup;
168 		regs->iaoq[0] &= ~3;
169 		/*
170 		 * NOTE: In some cases the faulting instruction
171 		 * may be in the delay slot of a branch. We
172 		 * don't want to take the branch, so we don't
173 		 * increment iaoq[1], instead we set it to be
174 		 * iaoq[0]+4, and clear the B bit in the PSW
175 		 */
176 		regs->iaoq[1] = regs->iaoq[0] + 4;
177 		regs->gr[0] &= ~PSW_B; /* IPSW in gr[0] */
178 
179 		return 1;
180 	}
181 
182 	return 0;
183 }
184 
185 /*
186  * parisc hardware trap list
187  *
188  * Documented in section 3 "Addressing and Access Control" of the
189  * "PA-RISC 1.1 Architecture and Instruction Set Reference Manual"
190  * https://parisc.wiki.kernel.org/index.php/File:Pa11_acd.pdf
191  *
192  * For implementation see handle_interruption() in traps.c
193  */
194 static const char * const trap_description[] = {
195 	[1] "High-priority machine check (HPMC)",
196 	[2] "Power failure interrupt",
197 	[3] "Recovery counter trap",
198 	[5] "Low-priority machine check",
199 	[6] "Instruction TLB miss fault",
200 	[7] "Instruction access rights / protection trap",
201 	[8] "Illegal instruction trap",
202 	[9] "Break instruction trap",
203 	[10] "Privileged operation trap",
204 	[11] "Privileged register trap",
205 	[12] "Overflow trap",
206 	[13] "Conditional trap",
207 	[14] "FP Assist Exception trap",
208 	[15] "Data TLB miss fault",
209 	[16] "Non-access ITLB miss fault",
210 	[17] "Non-access DTLB miss fault",
211 	[18] "Data memory protection/unaligned access trap",
212 	[19] "Data memory break trap",
213 	[20] "TLB dirty bit trap",
214 	[21] "Page reference trap",
215 	[22] "Assist emulation trap",
216 	[25] "Taken branch trap",
217 	[26] "Data memory access rights trap",
218 	[27] "Data memory protection ID trap",
219 	[28] "Unaligned data reference trap",
220 };
221 
222 const char *trap_name(unsigned long code)
223 {
224 	const char *t = NULL;
225 
226 	if (code < ARRAY_SIZE(trap_description))
227 		t = trap_description[code];
228 
229 	return t ? t : "Unknown trap";
230 }
231 
232 /*
233  * Print out info about fatal segfaults, if the show_unhandled_signals
234  * sysctl is set:
235  */
236 static inline void
237 show_signal_msg(struct pt_regs *regs, unsigned long code,
238 		unsigned long address, struct task_struct *tsk,
239 		struct vm_area_struct *vma)
240 {
241 	if (!unhandled_signal(tsk, SIGSEGV))
242 		return;
243 
244 	if (!printk_ratelimit())
245 		return;
246 
247 	pr_warn("\n");
248 	pr_warn("do_page_fault() command='%s' type=%lu address=0x%08lx",
249 	    tsk->comm, code, address);
250 	print_vma_addr(KERN_CONT " in ", regs->iaoq[0]);
251 
252 	pr_cont("\ntrap #%lu: %s%c", code, trap_name(code),
253 		vma ? ',':'\n');
254 
255 	if (vma)
256 		pr_cont(" vm_start = 0x%08lx, vm_end = 0x%08lx\n",
257 			vma->vm_start, vma->vm_end);
258 
259 	show_regs(regs);
260 }
261 
262 void do_page_fault(struct pt_regs *regs, unsigned long code,
263 			      unsigned long address)
264 {
265 	struct vm_area_struct *vma, *prev_vma;
266 	struct task_struct *tsk;
267 	struct mm_struct *mm;
268 	unsigned long acc_type;
269 	vm_fault_t fault = 0;
270 	unsigned int flags;
271 	char *msg;
272 
273 	tsk = current;
274 	mm = tsk->mm;
275 	if (!mm) {
276 		msg = "Page fault: no context";
277 		goto no_context;
278 	}
279 
280 	flags = FAULT_FLAG_DEFAULT;
281 	if (user_mode(regs))
282 		flags |= FAULT_FLAG_USER;
283 
284 	acc_type = parisc_acctyp(code, regs->iir);
285 	if (acc_type & VM_WRITE)
286 		flags |= FAULT_FLAG_WRITE;
287 	perf_sw_event(PERF_COUNT_SW_PAGE_FAULTS, 1, regs, address);
288 retry:
289 	mmap_read_lock(mm);
290 	vma = find_vma_prev(mm, address, &prev_vma);
291 	if (!vma || address < vma->vm_start)
292 		goto check_expansion;
293 /*
294  * Ok, we have a good vm_area for this memory access. We still need to
295  * check the access permissions.
296  */
297 
298 good_area:
299 
300 	if ((vma->vm_flags & acc_type) != acc_type)
301 		goto bad_area;
302 
303 	/*
304 	 * If for any reason at all we couldn't handle the fault, make
305 	 * sure we exit gracefully rather than endlessly redo the
306 	 * fault.
307 	 */
308 
309 	fault = handle_mm_fault(vma, address, flags, regs);
310 
311 	if (fault_signal_pending(fault, regs))
312 		return;
313 
314 	/* The fault is fully completed (including releasing mmap lock) */
315 	if (fault & VM_FAULT_COMPLETED)
316 		return;
317 
318 	if (unlikely(fault & VM_FAULT_ERROR)) {
319 		/*
320 		 * We hit a shared mapping outside of the file, or some
321 		 * other thing happened to us that made us unable to
322 		 * handle the page fault gracefully.
323 		 */
324 		if (fault & VM_FAULT_OOM)
325 			goto out_of_memory;
326 		else if (fault & VM_FAULT_SIGSEGV)
327 			goto bad_area;
328 		else if (fault & (VM_FAULT_SIGBUS|VM_FAULT_HWPOISON|
329 				  VM_FAULT_HWPOISON_LARGE))
330 			goto bad_area;
331 		BUG();
332 	}
333 	if (fault & VM_FAULT_RETRY) {
334 		/*
335 		 * No need to mmap_read_unlock(mm) as we would
336 		 * have already released it in __lock_page_or_retry
337 		 * in mm/filemap.c.
338 		 */
339 		flags |= FAULT_FLAG_TRIED;
340 		goto retry;
341 	}
342 	mmap_read_unlock(mm);
343 	return;
344 
345 check_expansion:
346 	vma = prev_vma;
347 	if (vma && (expand_stack(vma, address) == 0))
348 		goto good_area;
349 
350 /*
351  * Something tried to access memory that isn't in our memory map..
352  */
353 bad_area:
354 	mmap_read_unlock(mm);
355 
356 	if (user_mode(regs)) {
357 		int signo, si_code;
358 
359 		switch (code) {
360 		case 15:	/* Data TLB miss fault/Data page fault */
361 			/* send SIGSEGV when outside of vma */
362 			if (!vma ||
363 			    address < vma->vm_start || address >= vma->vm_end) {
364 				signo = SIGSEGV;
365 				si_code = SEGV_MAPERR;
366 				break;
367 			}
368 
369 			/* send SIGSEGV for wrong permissions */
370 			if ((vma->vm_flags & acc_type) != acc_type) {
371 				signo = SIGSEGV;
372 				si_code = SEGV_ACCERR;
373 				break;
374 			}
375 
376 			/* probably address is outside of mapped file */
377 			fallthrough;
378 		case 17:	/* NA data TLB miss / page fault */
379 		case 18:	/* Unaligned access - PCXS only */
380 			signo = SIGBUS;
381 			si_code = (code == 18) ? BUS_ADRALN : BUS_ADRERR;
382 			break;
383 		case 16:	/* Non-access instruction TLB miss fault */
384 		case 26:	/* PCXL: Data memory access rights trap */
385 		default:
386 			signo = SIGSEGV;
387 			si_code = (code == 26) ? SEGV_ACCERR : SEGV_MAPERR;
388 			break;
389 		}
390 #ifdef CONFIG_MEMORY_FAILURE
391 		if (fault & (VM_FAULT_HWPOISON|VM_FAULT_HWPOISON_LARGE)) {
392 			unsigned int lsb = 0;
393 			printk(KERN_ERR
394 	"MCE: Killing %s:%d due to hardware memory corruption fault at %08lx\n",
395 			tsk->comm, tsk->pid, address);
396 			/*
397 			 * Either small page or large page may be poisoned.
398 			 * In other words, VM_FAULT_HWPOISON_LARGE and
399 			 * VM_FAULT_HWPOISON are mutually exclusive.
400 			 */
401 			if (fault & VM_FAULT_HWPOISON_LARGE)
402 				lsb = hstate_index_to_shift(VM_FAULT_GET_HINDEX(fault));
403 			else if (fault & VM_FAULT_HWPOISON)
404 				lsb = PAGE_SHIFT;
405 
406 			force_sig_mceerr(BUS_MCEERR_AR, (void __user *) address,
407 					 lsb);
408 			return;
409 		}
410 #endif
411 		show_signal_msg(regs, code, address, tsk, vma);
412 
413 		force_sig_fault(signo, si_code, (void __user *) address);
414 		return;
415 	}
416 	msg = "Page fault: bad address";
417 
418 no_context:
419 
420 	if (!user_mode(regs) && fixup_exception(regs)) {
421 		return;
422 	}
423 
424 	parisc_terminate(msg, regs, code, address);
425 
426 out_of_memory:
427 	mmap_read_unlock(mm);
428 	if (!user_mode(regs)) {
429 		msg = "Page fault: out of memory";
430 		goto no_context;
431 	}
432 	pagefault_out_of_memory();
433 }
434 
435 /* Handle non-access data TLB miss faults.
436  *
437  * For probe instructions, accesses to userspace are considered allowed
438  * if they lie in a valid VMA and the access type matches. We are not
439  * allowed to handle MM faults here so there may be situations where an
440  * actual access would fail even though a probe was successful.
441  */
442 int
443 handle_nadtlb_fault(struct pt_regs *regs)
444 {
445 	unsigned long insn = regs->iir;
446 	int breg, treg, xreg, val = 0;
447 	struct vm_area_struct *vma, *prev_vma;
448 	struct task_struct *tsk;
449 	struct mm_struct *mm;
450 	unsigned long address;
451 	unsigned long acc_type;
452 
453 	switch (insn & 0x380) {
454 	case 0x280:
455 		/* FDC instruction */
456 		fallthrough;
457 	case 0x380:
458 		/* PDC and FIC instructions */
459 		if (DEBUG_NATLB && printk_ratelimit()) {
460 			pr_warn("WARNING: nullifying cache flush/purge instruction\n");
461 			show_regs(regs);
462 		}
463 		if (insn & 0x20) {
464 			/* Base modification */
465 			breg = (insn >> 21) & 0x1f;
466 			xreg = (insn >> 16) & 0x1f;
467 			if (breg && xreg)
468 				regs->gr[breg] += regs->gr[xreg];
469 		}
470 		regs->gr[0] |= PSW_N;
471 		return 1;
472 
473 	case 0x180:
474 		/* PROBE instruction */
475 		treg = insn & 0x1f;
476 		if (regs->isr) {
477 			tsk = current;
478 			mm = tsk->mm;
479 			if (mm) {
480 				/* Search for VMA */
481 				address = regs->ior;
482 				mmap_read_lock(mm);
483 				vma = find_vma_prev(mm, address, &prev_vma);
484 				mmap_read_unlock(mm);
485 
486 				/*
487 				 * Check if access to the VMA is okay.
488 				 * We don't allow for stack expansion.
489 				 */
490 				acc_type = (insn & 0x40) ? VM_WRITE : VM_READ;
491 				if (vma
492 				    && address >= vma->vm_start
493 				    && (vma->vm_flags & acc_type) == acc_type)
494 					val = 1;
495 			}
496 		}
497 		if (treg)
498 			regs->gr[treg] = val;
499 		regs->gr[0] |= PSW_N;
500 		return 1;
501 
502 	case 0x300:
503 		/* LPA instruction */
504 		if (insn & 0x20) {
505 			/* Base modification */
506 			breg = (insn >> 21) & 0x1f;
507 			xreg = (insn >> 16) & 0x1f;
508 			if (breg && xreg)
509 				regs->gr[breg] += regs->gr[xreg];
510 		}
511 		treg = insn & 0x1f;
512 		if (treg)
513 			regs->gr[treg] = 0;
514 		regs->gr[0] |= PSW_N;
515 		return 1;
516 
517 	default:
518 		break;
519 	}
520 
521 	return 0;
522 }
523