From ffb6b7e4fdef715061859651fe46cd27afc6acec Mon Sep 17 00:00:00 2001 From: Artem Bityutskiy Date: Tue, 12 May 2009 15:43:44 +0300 Subject: UBI: fix races in I/O debugging checks When paranoid checs are enabled, the 'io_paral' test from the 'mtd-utils' package fails. The symptoms are: UBI error: paranoid_check_all_ff: flash region at PEB 3973:512, length 15872 does not contain all 0xFF bytes UBI error: paranoid_check_all_ff: paranoid check failed for PEB 3973 UBI: hex dump of the 512-16384 region It turned out to be a bug in the checking function. Suppose there are 2 tasks - A and B. Task A is the wear-levelling working ('wear_leveling_worker()'). It is reading the VID header to find which LEB this PEB belongs to. Say, task A is reading header of PEB X. Suppose PEB X is unmapped, and has no VID header. Task B is trying to write to PEB X. Task A: in 'ubi_io_read_vid_hdr()': reads the VID header from PEB X. The read data contain all 0xFF bytes. Task B: writes VID header and some data to PEB X Task A: assumes PEB X is empty, calls 'paranoid_check_all_ff()', which fails. The solution for this problem is to make 'paranoid_check_all_ff()' re-read the VID header, re-check it, and only if it is not there, check the rest. This now implemented by the 'paranoid_check_empty()' function. Signed-off-by: Artem Bityutskiy diff --git a/drivers/mtd/ubi/io.c b/drivers/mtd/ubi/io.c index fe81039..ac6604a 100644 --- a/drivers/mtd/ubi/io.c +++ b/drivers/mtd/ubi/io.c @@ -100,6 +100,7 @@ static int paranoid_check_vid_hdr(const struct ubi_device *ubi, int pnum, const struct ubi_vid_hdr *vid_hdr); static int paranoid_check_all_ff(struct ubi_device *ubi, int pnum, int offset, int len); +static int paranoid_check_empty(struct ubi_device *ubi, int pnum); #else #define paranoid_check_not_bad(ubi, pnum) 0 #define paranoid_check_peb_ec_hdr(ubi, pnum) 0 @@ -107,6 +108,7 @@ static int paranoid_check_all_ff(struct ubi_device *ubi, int pnum, int offset, #define paranoid_check_peb_vid_hdr(ubi, pnum) 0 #define paranoid_check_vid_hdr(ubi, pnum, vid_hdr) 0 #define paranoid_check_all_ff(ubi, pnum, offset, len) 0 +#define paranoid_check_empty(ubi, pnum) 0 #endif /** @@ -670,11 +672,6 @@ int ubi_io_read_ec_hdr(struct ubi_device *ubi, int pnum, if (read_err != -EBADMSG && check_pattern(ec_hdr, 0xFF, UBI_EC_HDR_SIZE)) { /* The physical eraseblock is supposedly empty */ - - /* - * The below is just a paranoid check, it has to be - * compiled out if paranoid checks are disabled. - */ err = paranoid_check_all_ff(ubi, pnum, 0, ubi->peb_size); if (err) @@ -955,8 +952,7 @@ int ubi_io_read_vid_hdr(struct ubi_device *ubi, int pnum, * The below is just a paranoid check, it has to be * compiled out if paranoid checks are disabled. */ - err = paranoid_check_all_ff(ubi, pnum, ubi->leb_start, - ubi->leb_size); + err = paranoid_check_empty(ubi, pnum); if (err) return err > 0 ? UBI_IO_BAD_VID_HDR : err; @@ -1280,4 +1276,74 @@ error: return err; } +/** + * paranoid_check_empty - whether a PEB is empty. + * @ubi: UBI device description object + * @pnum: the physical eraseblock number to check + * + * This function makes sure PEB @pnum is empty, which means it contains only + * %0xFF data bytes. Returns zero if the PEB is empty, %1 if not, and a + * negative error code in case of failure. + * + * Empty PEBs have the EC header, and do not have the VID header. The caller of + * this function should have already made sure the PEB does not have the VID + * header. However, this function re-checks that, because it is possible that + * the header and data has already been written to the PEB. + * + * Let's consider a possible scenario. Suppose there are 2 tasks - A and B. + * Task A is in 'wear_leveling_worker()'. It is reading VID header of PEB X to + * find which LEB it corresponds to. PEB X is currently unmapped, and has no + * VID header. Task B is trying to write to PEB X. + * + * Task A: in 'ubi_io_read_vid_hdr()': reads the VID header from PEB X. The + * read data contain all 0xFF bytes; + * Task B: writes VID header and some data to PEB X; + * Task A: assumes PEB X is empty, calls 'paranoid_check_empty()'. And if we + * do not re-read the VID header, and do not cancel the checking if it + * is there, we fail. + */ +static int paranoid_check_empty(struct ubi_device *ubi, int pnum) +{ + int err, offs = ubi->vid_hdr_aloffset, len = ubi->vid_hdr_alsize; + size_t read; + uint32_t magic; + const struct ubi_vid_hdr *vid_hdr; + + mutex_lock(&ubi->dbg_buf_mutex); + err = ubi->mtd->read(ubi->mtd, offs, len, &read, ubi->dbg_peb_buf); + if (err && err != -EUCLEAN) { + ubi_err("error %d while reading %d bytes from PEB %d:%d, " + "read %zd bytes", err, len, pnum, offs, read); + goto error; + } + + vid_hdr = ubi->dbg_peb_buf; + magic = be32_to_cpu(vid_hdr->magic); + if (magic == UBI_VID_HDR_MAGIC) + /* The PEB contains VID header, so it is not empty */ + goto out; + + err = check_pattern(ubi->dbg_peb_buf, 0xFF, len); + if (err == 0) { + ubi_err("flash region at PEB %d:%d, length %d does not " + "contain all 0xFF bytes", pnum, offs, len); + goto fail; + } + +out: + mutex_unlock(&ubi->dbg_buf_mutex); + return 0; + +fail: + ubi_err("paranoid check failed for PEB %d", pnum); + ubi_msg("hex dump of the %d-%d region", offs, offs + len); + print_hex_dump(KERN_DEBUG, "", DUMP_PREFIX_OFFSET, 32, 1, + ubi->dbg_peb_buf, len, 1); + err = 1; +error: + ubi_dbg_dump_stack(); + mutex_unlock(&ubi->dbg_buf_mutex); + return err; +} + #endif /* CONFIG_MTD_UBI_DEBUG_PARANOID */ -- cgit v0.10.2