pstore/zone: Provide way to skip "broken" zone for MTD devices

One requirement to support MTD devices in pstore/zone is having a
way to declare certain regions as broken. Add this support to
pstore/zone.

The MTD driver should return -ENOMSG when encountering a bad region,
which tells pstore/zone to skip and try the next one.
Signed-off-by: WeiXiong Liao <liaoweixiong@allwinnertech.com>
Link: https://lore.kernel.org/lkml/20200511233229.27745-8-keescook@chromium.org/Co-developed-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Link: //lore.kernel.org/lkml/20200512173801.222666-1-colin.king@canonical.com
Signed-off-by: Kees Cook <keescook@chromium.org>

fs/pstore/blk.c

@@ -109,9 +109,12 @@ struct bdev_info {
  *		means error.
  * @write:	The same as @read, but the following error number:
  *		-EBUSY means try to write again later.
+ *		-ENOMSG means to try next zone.
  * @panic_write:The write operation only used for panic case. It's optional
- *		if you do not care panic log. The parameters and return value
- *		are the same as @read.
+ *		if you do not care panic log. The parameters are relative
+ *		value to storage.
+ *		On success, the number of bytes should be returned, others
+ *		excluding -ENOMSG mean error. -ENOMSG means to try next zone.
  */
 struct pstore_device_info {
 	unsigned long total_size;
@@ -337,6 +340,9 @@ static ssize_t psblk_blk_panic_write(const char *buf, size_t size,
 	/* size and off must align to SECTOR_SIZE for block device */
 	ret = blkdev_panic_write(buf, off >> SECTOR_SHIFT,
 			size >> SECTOR_SHIFT);
+	/* try next zone */
+	if (ret == -ENOMSG)
+		return ret;
 	return ret ? -EIO : size;
 }
 

fs/pstore/zone.c

@@ -249,6 +249,9 @@ static int psz_zone_write(struct pstore_zone *zone,
 
 	return 0;
 dirty:
+	/* no need to mark dirty if going to try next zone */
+	if (wcnt == -ENOMSG)
+		return -ENOMSG;
 	atomic_set(&zone->dirty, true);
 	/* flush dirty zones nicely */
 	if (wcnt == -EBUSY && !is_on_panic())
@@ -391,7 +394,11 @@ static int psz_kmsg_recover_meta(struct psz_context *cxt)
 			return -EINVAL;
 
 		rcnt = info->read((char *)buf, len, zone->off);
-		if (rcnt != len) {
+		if (rcnt == -ENOMSG) {
+			pr_debug("%s with id %lu may be broken, skip\n",
+					zone->name, i);
+			continue;
+		} else if (rcnt != len) {
 			pr_err("read %s with id %lu failed\n", zone->name, i);
 			return (int)rcnt < 0 ? (int)rcnt : -EIO;
 		}
@@ -725,24 +732,58 @@ static void psz_write_kmsg_hdr(struct pstore_zone *zone,
 		hdr->counter = 0;
 }
 
+/*
+ * In case zone is broken, which may occur to MTD device, we try each zones,
+ * start at cxt->kmsg_write_cnt.
+ */
 static inline int notrace psz_kmsg_write_record(struct psz_context *cxt,
 		struct pstore_record *record)
 {
 	size_t size, hlen;
 	struct pstore_zone *zone;
-	unsigned int zonenum;
+	unsigned int i;
 
-	zonenum = cxt->kmsg_write_cnt;
-	zone = cxt->kpszs[zonenum];
-	if (unlikely(!zone))
-		return -ENOSPC;
-	cxt->kmsg_write_cnt = (zonenum + 1) % cxt->kmsg_max_cnt;
+	for (i = 0; i < cxt->kmsg_max_cnt; i++) {
+		unsigned int zonenum, len;
+		int ret;
 
-	pr_debug("write %s to zone id %d\n", zone->name, zonenum);
-	psz_write_kmsg_hdr(zone, record);
-	hlen = sizeof(struct psz_kmsg_header);
-	size = min_t(size_t, record->size, zone->buffer_size - hlen);
-	return psz_zone_write(zone, FLUSH_ALL, record->buf, size, hlen);
+		zonenum = (cxt->kmsg_write_cnt + i) % cxt->kmsg_max_cnt;
+		zone = cxt->kpszs[zonenum];
+		if (unlikely(!zone))
+			return -ENOSPC;
+
+		/* avoid destroying old data, allocate a new one */
+		len = zone->buffer_size + sizeof(*zone->buffer);
+		zone->oldbuf = zone->buffer;
+		zone->buffer = kzalloc(len, GFP_KERNEL);
+		if (!zone->buffer) {
+			zone->buffer = zone->oldbuf;
+			return -ENOMEM;
+		}
+		zone->buffer->sig = zone->oldbuf->sig;
+
+		pr_debug("write %s to zone id %d\n", zone->name, zonenum);
+		psz_write_kmsg_hdr(zone, record);
+		hlen = sizeof(struct psz_kmsg_header);
+		size = min_t(size_t, record->size, zone->buffer_size - hlen);
+		ret = psz_zone_write(zone, FLUSH_ALL, record->buf, size, hlen);
+		if (likely(!ret || ret != -ENOMSG)) {
+			cxt->kmsg_write_cnt = zonenum + 1;
+			cxt->kmsg_write_cnt %= cxt->kmsg_max_cnt;
+			/* no need to try next zone, free last zone buffer */
+			kfree(zone->oldbuf);
+			zone->oldbuf = NULL;
+			return ret;
+		}
+
+		pr_debug("zone %u may be broken, try next dmesg zone\n",
+				zonenum);
+		kfree(zone->buffer);
+		zone->buffer = zone->oldbuf;
+		zone->oldbuf = NULL;
+	}
+
+	return -EBUSY;
 }
 
 static int notrace psz_kmsg_write(struct psz_context *cxt,

include/linux/pstore_blk.h

@@ -14,7 +14,8 @@
  * @start_sect: start sector to block device
  * @sects: sectors count on buf
  *
- * Return: On success, zero should be returned. Others mean error.
+ * Return: On success, zero should be returned. Others excluding -ENOMSG
+ * mean error. -ENOMSG means to try next zone.
  *
  * Panic write to block device must be aligned to SECTOR_SIZE.
  */

include/linux/pstore_zone.h

@@ -23,11 +23,15 @@ typedef ssize_t (*pstore_zone_write_op)(const char *, size_t, loff_t);
  * @read:	The general read operation. Both of the function parameters
  *		@size and @offset are relative value to storage.
  *		On success, the number of bytes should be returned, others
- *		means error.
- * @write:	The same as @read, but -EBUSY means try to write again later.
+ *		mean error.
+ * @write:	The same as @read, but the following error number:
+ *		-EBUSY means try to write again later.
+ *		-ENOMSG means to try next zone.
  * @panic_write:The write operation only used for panic case. It's optional
- *		if you do not care panic log. The parameters and return value
- *		are the same as @read.
+ *		if you do not care panic log. The parameters are relative
+ *		value to storage.
+ *		On success, the number of bytes should be returned, others
+ *		excluding -ENOMSG mean error. -ENOMSG means to try next zone.
  */
 struct pstore_zone_info {
 	struct module *owner;