With more than one device, dmz_zone_is_cache() only matches the zones of
the regular (cache) device. A chunk mapped to a conventional zone of the
zoned device therefore reaches the write pointer test in
dmz_check_mapped_zone_bitmap(). The kernel reports the write pointer of a
conventional zone as -1, so wp_block becomes garbage and the bitmap loop
reads far past its two-block buffer: dmzadm --check segfaults
("Checking zone bitmaps..." is the last line), and --repair could clear
bits based on that memory before crashing.

Treat conventional and emulated (regular-device) zones as having no write
pointer: every block may be valid.

Found on a cache-device + 27 TB host-managed drive set after reclaim had
moved chunks into conventional zones (stock 2.2.2: exit 139; with the
patch: "No error detected", also on an emulated ZBC drive with the same
layout).

Reproducer (scsi_debug zbc=managed, no special hardware):
https://github.com/illofspeed/synology-host-managed-smr/blob/main/repro/1-dmzadm-check-two-devices.sh

Assisted-by: LLM
Signed-off-by: Kyrylo Biryuk <[email protected]>
---

Notes:
    v3: no code changes. From and Signed-off-by now carry my full name
    instead of the GitHub handle "illofspeed" used for v2 (same author).
    
    The addresses in README.md are out of date (wdc.com, [email protected]),
    hence this goes to [email protected] and [email protected].

 src/dmz_check.c | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)

diff --git a/src/dmz_check.c b/src/dmz_check.c
index b13d025..943308b 100644
--- a/src/dmz_check.c
+++ b/src/dmz_check.c
@@ -509,9 +509,18 @@ static int dmz_check_mapped_zone_bitmap(struct dmz_dev 
*dev,
        if (ret != 0)
                return -1;
 
-       /* No valid block should be present after the write pointer */
+       /*
+        * No valid block should be present after the write pointer.
+        * Conventional zones have no write pointer (the kernel reports -1):
+        * every block may be valid. With more than one device,
+        * dmz_zone_is_cache() only matches the cache device's zones, so the
+        * zoned device's conventional zones reach this point.
+        */
        bad_bits = 0;
-       wp_block = dmz_sect2blk(zone->wp - zone->start);
+       if (dmz_zone_conv(zone) || dmz_zone_unknown(zone))
+               wp_block = dev->zone_nr_blocks;
+       else
+               wp_block = dmz_sect2blk(zone->wp - zone->start);
        for (b = 0; b < wp_block; b++) {
                if (dmz_test_bit(dbuf, b))
                        dzone_weight++;
-- 
2.43.0


Reply via email to