platform.c

  1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Persistent Storage - platform driver interface parts.
  4 *
  5 * Copyright (C) 2007-2008 Google, Inc.
  6 * Copyright (C) 2010 Intel Corporation <tony.luck@intel.com>
  7 */
  8
  9#define pr_fmt(fmt) "pstore: " fmt
 10
 11#include <linux/atomic.h>
 12#include <linux/types.h>
 13#include <linux/errno.h>
 14#include <linux/init.h>
 15#include <linux/kmsg_dump.h>
 16#include <linux/console.h>
 17#include <linux/mm.h>
 18#include <linux/module.h>
 19#include <linux/pstore.h>
 20#include <linux/string.h>
 21#include <linux/timer.h>
 22#include <linux/slab.h>
 23#include <linux/uaccess.h>
 24#include <linux/jiffies.h>
 25#include <linux/vmalloc.h>
 26#include <linux/workqueue.h>
 27#include <linux/zlib.h>
 28
 29#include "internal.h"
 30
 31/*
 32 * We defer making "oops" entries appear in pstore - see
 33 * whether the system is actually still running well enough
 34 * to let someone see the entry
 35 */
 36static int pstore_update_ms = -1;
 37module_param_named(update_ms, pstore_update_ms, int, 0600);
 38MODULE_PARM_DESC(update_ms, "milliseconds before pstore updates its content "
 39		 "(default is -1, which means runtime updates are disabled; "
 40		 "enabling this option may not be safe; it may lead to further "
 41		 "corruption on Oopses)");
 42
 43/* Names should be in the same order as the enum pstore_type_id */
 44static const char * const pstore_type_names[] = {
 45	"dmesg",
 46	"mce",
 47	"console",
 48	"ftrace",
 49	"rtas",
 50	"powerpc-ofw",
 51	"powerpc-common",
 52	"pmsg",
 53	"powerpc-opal",
 54};
 55
 56static int pstore_new_entry;
 57
 58static void pstore_timefunc(struct timer_list *);
 59static DEFINE_TIMER(pstore_timer, pstore_timefunc);
 60
 61static void pstore_dowork(struct work_struct *);
 62static DECLARE_WORK(pstore_work, pstore_dowork);
 63
 64/*
 65 * psinfo_lock protects "psinfo" during calls to
 66 * pstore_register(), pstore_unregister(), and
 67 * the filesystem mount/unmount routines.
 68 */
 69static DEFINE_MUTEX(psinfo_lock);
 70struct pstore_info *psinfo;
 71
 72static char *backend;
 73module_param(backend, charp, 0444);
 74MODULE_PARM_DESC(backend, "specific backend to use");
 75
 76/*
 77 * pstore no longer implements compression via the crypto API, and only
 78 * supports zlib deflate compression implemented using the zlib library
 79 * interface. This removes additional complexity which is hard to justify for a
 80 * diagnostic facility that has to operate in conditions where the system may
 81 * have become unstable. Zlib deflate is comparatively small in terms of code
 82 * size, and compresses ASCII text comparatively well. In terms of compression
 83 * speed, deflate is not the best performer but for recording the log output on
 84 * a kernel panic, this is not considered critical.
 85 *
 86 * The only remaining arguments supported by the compress= module parameter are
 87 * 'deflate' and 'none'. To retain compatibility with existing installations,
 88 * all other values are logged and replaced with 'deflate'.
 89 */
 90static char *compress = "deflate";
 91module_param(compress, charp, 0444);
 92MODULE_PARM_DESC(compress, "compression to use");
 93
 94/* How much of the kernel log to snapshot */
 95unsigned int kmsg_bytes = CONFIG_PSTORE_DEFAULT_KMSG_BYTES;
 96module_param(kmsg_bytes, uint, 0444);
 97MODULE_PARM_DESC(kmsg_bytes, "amount of kernel log to snapshot (in bytes)");
 98
 99static void *compress_workspace;
100
101/*
102 * Compression is only used for dmesg output, which consists of low-entropy
103 * ASCII text, and so we can assume worst-case 60%.
104 */
105#define DMESG_COMP_PERCENT	60
106
107static char *big_oops_buf;
108static size_t max_compressed_size;
109
110void pstore_set_kmsg_bytes(unsigned int bytes)
111{
112	WRITE_ONCE(kmsg_bytes, bytes);
113}
114
115/* Tag each group of saved records with a sequence number */
116static int	oopscount;
117
118const char *pstore_type_to_name(enum pstore_type_id type)
119{
120	BUILD_BUG_ON(ARRAY_SIZE(pstore_type_names) != PSTORE_TYPE_MAX);
121
122	if (WARN_ON_ONCE(type >= PSTORE_TYPE_MAX))
123		return "unknown";
124
125	return pstore_type_names[type];
126}
127EXPORT_SYMBOL_GPL(pstore_type_to_name);
128
129enum pstore_type_id pstore_name_to_type(const char *name)
130{
131	int i;
132
133	for (i = 0; i < PSTORE_TYPE_MAX; i++) {
134		if (!strcmp(pstore_type_names[i], name))
135			return i;
136	}
137
138	return PSTORE_TYPE_MAX;
139}
140EXPORT_SYMBOL_GPL(pstore_name_to_type);
141
142static void pstore_timer_kick(void)
143{
144	if (pstore_update_ms < 0)
145		return;
146
147	mod_timer(&pstore_timer, jiffies + msecs_to_jiffies(pstore_update_ms));
148}
149
150static bool pstore_cannot_block_path(enum kmsg_dump_reason reason)
151{
152	/*
153	 * In case of NMI path, pstore shouldn't be blocked
154	 * regardless of reason.
155	 */
156	if (in_nmi())
157		return true;
158
159	switch (reason) {
160	/* In panic case, other cpus are stopped by smp_send_stop(). */
161	case KMSG_DUMP_PANIC:
162	/*
163	 * Emergency restart shouldn't be blocked by spinning on
164	 * pstore_info::buf_lock.
165	 */
166	case KMSG_DUMP_EMERG:
167		return true;
168	default:
169		return false;
170	}
171}
172
173static int pstore_compress(const void *in, void *out,
174			   unsigned int inlen, unsigned int outlen)
175{
176	struct z_stream_s zstream = {
177		.next_in	= in,
178		.avail_in	= inlen,
179		.next_out	= out,
180		.avail_out	= outlen,
181		.workspace	= compress_workspace,
182	};
183	int ret;
184
185	if (!IS_ENABLED(CONFIG_PSTORE_COMPRESS))
186		return -EINVAL;
187
188	ret = zlib_deflateInit2(&zstream, Z_DEFAULT_COMPRESSION, Z_DEFLATED,
189				-MAX_WBITS, DEF_MEM_LEVEL, Z_DEFAULT_STRATEGY);
190	if (ret != Z_OK)
191		return -EINVAL;
192
193	ret = zlib_deflate(&zstream, Z_FINISH);
194	if (ret != Z_STREAM_END)
195		return -EINVAL;
196
197	ret = zlib_deflateEnd(&zstream);
198	if (ret != Z_OK)
199		pr_warn_once("zlib_deflateEnd() failed: %d\n", ret);
200
201	return zstream.total_out;
202}
203
204static void allocate_buf_for_compression(void)
205{
206	size_t compressed_size;
207	char *buf;
208
209	/* Skip if not built-in or compression disabled. */
210	if (!IS_ENABLED(CONFIG_PSTORE_COMPRESS) || !compress ||
211	    !strcmp(compress, "none")) {
212		compress = NULL;
213		return;
214	}
215
216	if (strcmp(compress, "deflate")) {
217		pr_err("Unsupported compression '%s', falling back to deflate\n",
218		       compress);
219		compress = "deflate";
220	}
221
222	/*
223	 * The compression buffer only needs to be as large as the maximum
224	 * uncompressed record size, since any record that would be expanded by
225	 * compression is just stored uncompressed.
226	 */
227	compressed_size = (psinfo->bufsize * 100) / DMESG_COMP_PERCENT;
228	buf = kvzalloc(compressed_size, GFP_KERNEL);
229	if (!buf) {
230		pr_err("Failed %zu byte compression buffer allocation for: %s\n",
231		       psinfo->bufsize, compress);
232		return;
233	}
234
235	compress_workspace =
236		vmalloc(zlib_deflate_workspacesize(MAX_WBITS, DEF_MEM_LEVEL));
237	if (!compress_workspace) {
238		pr_err("Failed to allocate zlib deflate workspace\n");
239		kvfree(buf);
240		return;
241	}
242
243	/* A non-NULL big_oops_buf indicates compression is available. */
244	big_oops_buf = buf;
245	max_compressed_size = compressed_size;
246
247	pr_info("Using crash dump compression: %s\n", compress);
248}
249
250static void free_buf_for_compression(void)
251{
252	if (IS_ENABLED(CONFIG_PSTORE_COMPRESS) && compress_workspace) {
253		vfree(compress_workspace);
254		compress_workspace = NULL;
255	}
256
257	kvfree(big_oops_buf);
258	big_oops_buf = NULL;
259	max_compressed_size = 0;
260}
261
262void pstore_record_init(struct pstore_record *record,
263			struct pstore_info *psinfo)
264{
265	memset(record, 0, sizeof(*record));
266
267	record->psi = psinfo;
268
269	/* Report zeroed timestamp if called before timekeeping has resumed. */
270	record->time = ns_to_timespec64(ktime_get_real_fast_ns());
271}
272
273/*
274 * callback from kmsg_dump. Save as much as we can (up to kmsg_bytes) from the
275 * end of the buffer.
276 */
277static void pstore_dump(struct kmsg_dumper *dumper,
278			struct kmsg_dump_detail *detail)
279{
280	struct kmsg_dump_iter iter;
281	unsigned int	remaining = READ_ONCE(kmsg_bytes);
282	unsigned long	total = 0;
283	const char	*why;
284	unsigned int	part = 1;
285	unsigned long	flags = 0;
286	int		saved_ret = 0;
287	int		ret;
288
289	why = kmsg_dump_reason_str(detail->reason);
290
291	if (pstore_cannot_block_path(detail->reason)) {
292		if (!raw_spin_trylock_irqsave(&psinfo->buf_lock, flags)) {
293			pr_err("dump skipped in %s path because of concurrent dump\n",
294					in_nmi() ? "NMI" : why);
295			return;
296		}
297	} else {
298		raw_spin_lock_irqsave(&psinfo->buf_lock, flags);
299	}
300
301	kmsg_dump_rewind(&iter);
302
303	oopscount++;
304	while (total < remaining) {
305		char *dst;
306		size_t dst_size;
307		int header_size;
308		int zipped_len = -1;
309		size_t dump_size;
310		struct pstore_record record;
311
312		pstore_record_init(&record, psinfo);
313		record.type = PSTORE_TYPE_DMESG;
314		record.count = oopscount;
315		record.reason = detail->reason;
316		record.part = part;
317		record.buf = psinfo->buf;
318
319		dst = big_oops_buf ?: psinfo->buf;
320		dst_size = max_compressed_size ?: psinfo->bufsize;
321
322		/* Write dump header. */
323		header_size = snprintf(dst, dst_size, "%s#%d Part%u\n", why,
324				 oopscount, part);
325		dst_size -= header_size;
326
327		/* Write dump contents. */
328		if (!kmsg_dump_get_buffer(&iter, true, dst + header_size,
329					  dst_size, &dump_size))
330			break;
331
332		if (big_oops_buf) {
333			zipped_len = pstore_compress(dst, psinfo->buf,
334						header_size + dump_size,
335						psinfo->bufsize);
336
337			if (zipped_len > 0) {
338				record.compressed = true;
339				record.size = zipped_len;
340			} else {
341				/*
342				 * Compression failed, so the buffer is most
343				 * likely filled with binary data that does not
344				 * compress as well as ASCII text. Copy as much
345				 * of the uncompressed data as possible into
346				 * the pstore record, and discard the rest.
347				 */
348				record.size = psinfo->bufsize;
349				memcpy(psinfo->buf, dst, psinfo->bufsize);
350			}
351		} else {
352			record.size = header_size + dump_size;
353		}
354
355		ret = psinfo->write(&record);
356		if (ret == 0 && detail->reason == KMSG_DUMP_OOPS) {
357			pstore_new_entry = 1;
358			pstore_timer_kick();
359		} else {
360			/* Preserve only the first non-zero returned value. */
361			if (!saved_ret)
362				saved_ret = ret;
363		}
364
365		total += record.size;
366		part++;
367	}
368	raw_spin_unlock_irqrestore(&psinfo->buf_lock, flags);
369
370	if (saved_ret) {
371		pr_err_once("backend (%s) writing error (%d)\n", psinfo->name,
372			    saved_ret);
373	}
374}
375
376static struct kmsg_dumper pstore_dumper = {
377	.dump = pstore_dump,
378};
379
380/*
381 * Register with kmsg_dump to save last part of console log on panic.
382 */
383static void pstore_register_kmsg(void)
384{
385	kmsg_dump_register(&pstore_dumper);
386}
387
388static void pstore_unregister_kmsg(void)
389{
390	kmsg_dump_unregister(&pstore_dumper);
391}
392
393#ifdef CONFIG_PSTORE_CONSOLE
394static void pstore_console_write(struct console *con, const char *s, unsigned c)
395{
396	struct pstore_record record;
397
398	if (!c)
399		return;
400
401	pstore_record_init(&record, psinfo);
402	record.type = PSTORE_TYPE_CONSOLE;
403
404	record.buf = (char *)s;
405	record.size = c;
406	psinfo->write(&record);
407}
408
409static struct console pstore_console = {
410	.write	= pstore_console_write,
411	.index	= -1,
412};
413
414static void pstore_register_console(void)
415{
416	/* Show which backend is going to get console writes. */
417	strscpy(pstore_console.name, psinfo->name,
418		sizeof(pstore_console.name));
419	/*
420	 * Always initialize flags here since prior unregister_console()
421	 * calls may have changed settings (specifically CON_ENABLED).
422	 */
423	pstore_console.flags = CON_PRINTBUFFER | CON_ENABLED | CON_ANYTIME;
424	register_console(&pstore_console);
425}
426
427static void pstore_unregister_console(void)
428{
429	unregister_console(&pstore_console);
430}
431#else
432static void pstore_register_console(void) {}
433static void pstore_unregister_console(void) {}
434#endif
435
436static int pstore_write_user_compat(struct pstore_record *record,
437				    const char __user *buf)
438{
439	int ret = 0;
440
441	if (record->buf)
442		return -EINVAL;
443
444	record->buf = vmemdup_user(buf, record->size);
445	if (IS_ERR(record->buf)) {
446		ret = PTR_ERR(record->buf);
447		goto out;
448	}
449
450	ret = record->psi->write(record);
451
452	kvfree(record->buf);
453out:
454	record->buf = NULL;
455
456	return unlikely(ret < 0) ? ret : record->size;
457}
458
459/*
460 * platform specific persistent storage driver registers with
461 * us here. If pstore is already mounted, call the platform
462 * read function right away to populate the file system. If not
463 * then the pstore mount code will call us later to fill out
464 * the file system.
465 */
466int pstore_register(struct pstore_info *psi)
467{
468	char *new_backend;
469
470	if (backend && strcmp(backend, psi->name)) {
471		pr_warn("backend '%s' already in use: ignoring '%s'\n",
472			backend, psi->name);
473		return -EBUSY;
474	}
475
476	/* Sanity check flags. */
477	if (!psi->flags) {
478		pr_warn("backend '%s' must support at least one frontend\n",
479			psi->name);
480		return -EINVAL;
481	}
482
483	/* Check for required functions. */
484	if (!psi->read || !psi->write) {
485		pr_warn("backend '%s' must implement read() and write()\n",
486			psi->name);
487		return -EINVAL;
488	}
489
490	new_backend = kstrdup(psi->name, GFP_KERNEL);
491	if (!new_backend)
492		return -ENOMEM;
493
494	mutex_lock(&psinfo_lock);
495	if (psinfo) {
496		pr_warn("backend '%s' already loaded: ignoring '%s'\n",
497			psinfo->name, psi->name);
498		mutex_unlock(&psinfo_lock);
499		kfree(new_backend);
500		return -EBUSY;
501	}
502
503	if (!psi->write_user)
504		psi->write_user = pstore_write_user_compat;
505	psinfo = psi;
506	mutex_init(&psinfo->read_mutex);
507	raw_spin_lock_init(&psinfo->buf_lock);
508
509	if (psi->flags & PSTORE_FLAGS_DMESG)
510		allocate_buf_for_compression();
511
512	pstore_get_records(0);
513
514	if (psi->flags & PSTORE_FLAGS_DMESG) {
515		pstore_dumper.max_reason = psinfo->max_reason;
516		pstore_register_kmsg();
517	}
518	if (psi->flags & PSTORE_FLAGS_CONSOLE)
519		pstore_register_console();
520	if (psi->flags & PSTORE_FLAGS_FTRACE)
521		pstore_register_ftrace();
522	if (psi->flags & PSTORE_FLAGS_PMSG)
523		pstore_register_pmsg();
524
525	/* Start watching for new records, if desired. */
526	pstore_timer_kick();
527
528	/*
529	 * Update the module parameter backend, so it is visible
530	 * through /sys/module/pstore/parameters/backend
531	 */
532	backend = new_backend;
533
534	pr_info("Registered %s as persistent store backend\n", psi->name);
535
536	mutex_unlock(&psinfo_lock);
537	return 0;
538}
539EXPORT_SYMBOL_GPL(pstore_register);
540
541void pstore_unregister(struct pstore_info *psi)
542{
543	/* It's okay to unregister nothing. */
544	if (!psi)
545		return;
546
547	mutex_lock(&psinfo_lock);
548
549	/* Only one backend can be registered at a time. */
550	if (WARN_ON(psi != psinfo)) {
551		mutex_unlock(&psinfo_lock);
552		return;
553	}
554
555	/* Unregister all callbacks. */
556	if (psi->flags & PSTORE_FLAGS_PMSG)
557		pstore_unregister_pmsg();
558	if (psi->flags & PSTORE_FLAGS_FTRACE)
559		pstore_unregister_ftrace();
560	if (psi->flags & PSTORE_FLAGS_CONSOLE)
561		pstore_unregister_console();
562	if (psi->flags & PSTORE_FLAGS_DMESG)
563		pstore_unregister_kmsg();
564
565	/* Stop timer and make sure all work has finished. */
566	timer_delete_sync(&pstore_timer);
567	flush_work(&pstore_work);
568
569	/* Remove all backend records from filesystem tree. */
570	pstore_put_backend_records(psi);
571
572	free_buf_for_compression();
573
574	psinfo = NULL;
575	kfree(backend);
576	backend = NULL;
577
578	pr_info("Unregistered %s as persistent store backend\n", psi->name);
579	mutex_unlock(&psinfo_lock);
580}
581EXPORT_SYMBOL_GPL(pstore_unregister);
582
583static void decompress_record(struct pstore_record *record,
584			      struct z_stream_s *zstream)
585{
586	int ret;
587	int unzipped_len;
588	char *unzipped, *workspace;
589	size_t max_uncompressed_size;
590
591	if (!IS_ENABLED(CONFIG_PSTORE_COMPRESS) || !record->compressed)
592		return;
593
594	/* Only PSTORE_TYPE_DMESG support compression. */
595	if (record->type != PSTORE_TYPE_DMESG) {
596		pr_warn("ignored compressed record type %d\n", record->type);
597		return;
598	}
599
600	/* Missing compression buffer means compression was not initialized. */
601	if (!zstream->workspace) {
602		pr_warn("no decompression method initialized!\n");
603		return;
604	}
605
606	ret = zlib_inflateReset(zstream);
607	if (ret != Z_OK) {
608		pr_err("zlib_inflateReset() failed, ret = %d!\n", ret);
609		return;
610	}
611
612	/* Allocate enough space to hold max decompression and ECC. */
613	max_uncompressed_size = 3 * psinfo->bufsize;
614	workspace = kvzalloc(max_uncompressed_size + record->ecc_notice_size,
615			     GFP_KERNEL);
616	if (!workspace)
617		return;
618
619	zstream->next_in	= record->buf;
620	zstream->avail_in	= record->size;
621	zstream->next_out	= workspace;
622	zstream->avail_out	= max_uncompressed_size;
623
624	ret = zlib_inflate(zstream, Z_FINISH);
625	if (ret != Z_STREAM_END) {
626		pr_err_ratelimited("zlib_inflate() failed, ret = %d!\n", ret);
627		kvfree(workspace);
628		return;
629	}
630
631	unzipped_len = zstream->total_out;
632
633	/* Append ECC notice to decompressed buffer. */
634	memcpy(workspace + unzipped_len, record->buf + record->size,
635	       record->ecc_notice_size);
636
637	/* Copy decompressed contents into an minimum-sized allocation. */
638	unzipped = kvmemdup(workspace, unzipped_len + record->ecc_notice_size,
639			    GFP_KERNEL);
640	kvfree(workspace);
641	if (!unzipped)
642		return;
643
644	/* Swap out compressed contents with decompressed contents. */
645	kvfree(record->buf);
646	record->buf = unzipped;
647	record->size = unzipped_len;
648	record->compressed = false;
649}
650
651/*
652 * Read all the records from one persistent store backend. Create
653 * files in our filesystem.  Don't warn about -EEXIST errors
654 * when we are re-scanning the backing store looking to add new
655 * error records.
656 */
657void pstore_get_backend_records(struct pstore_info *psi,
658				struct dentry *root, int quiet)
659{
660	int failed = 0;
661	unsigned int stop_loop = 65536;
662	struct z_stream_s zstream = {};
663
664	if (!psi || !root)
665		return;
666
667	if (IS_ENABLED(CONFIG_PSTORE_COMPRESS) && compress) {
668		zstream.workspace = kvmalloc(zlib_inflate_workspacesize(),
669					     GFP_KERNEL);
670		zlib_inflateInit2(&zstream, -DEF_WBITS);
671	}
672
673	mutex_lock(&psi->read_mutex);
674	if (psi->open && psi->open(psi))
675		goto out;
676
677	/*
678	 * Backend callback read() allocates record.buf. decompress_record()
679	 * may reallocate record.buf. On success, pstore_mkfile() will keep
680	 * the record.buf, so free it only on failure.
681	 */
682	for (; stop_loop; stop_loop--) {
683		struct pstore_record *record;
684		int rc;
685
686		record = kzalloc(sizeof(*record), GFP_KERNEL);
687		if (!record) {
688			pr_err("out of memory creating record\n");
689			break;
690		}
691		pstore_record_init(record, psi);
692
693		record->size = psi->read(record);
694
695		/* No more records left in backend? */
696		if (record->size <= 0) {
697			kfree(record);
698			break;
699		}
700
701		decompress_record(record, &zstream);
702		rc = pstore_mkfile(root, record);
703		if (rc) {
704			/* pstore_mkfile() did not take record, so free it. */
705			kvfree(record->buf);
706			kfree(record->priv);
707			kfree(record);
708			if (rc != -EEXIST || !quiet)
709				failed++;
710		}
711	}
712	if (psi->close)
713		psi->close(psi);
714out:
715	mutex_unlock(&psi->read_mutex);
716
717	if (IS_ENABLED(CONFIG_PSTORE_COMPRESS) && compress) {
718		if (zlib_inflateEnd(&zstream) != Z_OK)
719			pr_warn("zlib_inflateEnd() failed\n");
720		kvfree(zstream.workspace);
721	}
722
723	if (failed)
724		pr_warn("failed to create %d record(s) from '%s'\n",
725			failed, psi->name);
726	if (!stop_loop)
727		pr_err("looping? Too many records seen from '%s'\n",
728			psi->name);
729}
730
731static void pstore_dowork(struct work_struct *work)
732{
733	pstore_get_records(1);
734}
735
736static void pstore_timefunc(struct timer_list *unused)
737{
738	if (pstore_new_entry) {
739		pstore_new_entry = 0;
740		schedule_work(&pstore_work);
741	}
742
743	pstore_timer_kick();
744}
745
746static int __init pstore_init(void)
747{
748	int ret;
749
750	ret = pstore_init_fs();
751	if (ret)
752		free_buf_for_compression();
753
754	return ret;
755}
756late_initcall(pstore_init);
757
758static void __exit pstore_exit(void)
759{
760	pstore_exit_fs();
761}
762module_exit(pstore_exit)
763
764MODULE_AUTHOR("Tony Luck <tony.luck@intel.com>");
765MODULE_DESCRIPTION("Persistent Storage - platform driver interface");
766MODULE_LICENSE("GPL");