Description of problem: The status of the region having write failure is marked as "in-sync" if the mirror device is suspended before the dmeventd takes action. This causes data corruption because the data is written successfully except for the failed device and recovery isn't kicked even if the failed device is restored. Version-Release number of selected component: kernel-2.6.9-34.EL How reproducible: Always Steps to Reproduce: 1. Modify /etc/lvm/lvm.conf not to launch the dmeventd like below. dmeventd { mirror_library = "none" } 2. Prepare some PVs (more than 2) and create VG from them. Example) - /dev/sda, /dev/sdb, /dev/sdc as PVs - vg0 contains these 3 PVs 3. Create a mirror LV and activate it. # lvcreate -L 12M -n lv0 -m 1 vg0 4. Wait until recovery is done. 5. Disconnect one of PVs used for the mirror LV. # echo offline > /sys/block/sdb/device/state 6. Write distinguishable data to the mirror LV. # echo -n "deadbeef" | dd of=/dev/mapper/vg0-lv0 7. Suspend and resume the mirror LV. # dmsetup suspend vg0-lv0 # dmsetup resume vg0-lv0 8. Re-connect the PV. # echo running > /sys/block/sdb/device/state 9. Deactivate and reactivate the mirror LV. # lvchange -an vg0/lv0 # lvchange -ay vg0/lv0 10. Check if mirror images are consistent. # cmp -l /dev/mapper/vg0-lv0_mimage_0 /dev/mapper/vg0-lv0_mimage_1 Actual results: cmp detects these two mirror images are inconsistent, while 'dmsetup status' shows that recovery of the mirror LV is completed. This happens because the kmir_mon calls bio_endio() with no error for all failed write in write_failure_handler(). Therefore, the write is considered as success and the corresponding region is marked as "in-sync". Expected results: cmp should not detect any error. In the kernel side, the region having had write errors should not been marked as "in-sync" while the failed device is still on the mirror map. And recovery should be correctly kicked if the failed device is restored. Additional info:
At least, we have to avoid the data corruption. It's possible by not marking the failed region as clean until the failed device is removed. Possible solution is to mark the failed region (which should be RH_DIRTY at that point) as RH_NOSYNC. With a proposed fix in BZ#177067 Comment#10, the RH_NOSYNC region will stay in region hash and be kept unclean until suspend happens. Upon suspend, dirty state of the region is flushed to log. Then after resume, the RH_NOSYNC region will be eventually recovered. If the recoverying process correctly handles write errors (BZ#185785), we can make sure the region state not becoming clean until failed devices are removed.
committed in stream U4 build 34.26. A test kernel with this patch is available from http://people.redhat.com/~jbaron/rhel4/
Needs supplimental patch to properly take account of cluster mirroring. Sent to kernel list on 05/15/2006.
Created attachment 129057 [details] supplimental patch This patch applies ontop of what is already there.
I've created a new bugzilla to track the supplimental patch.
An advisory has been issued which should help the problem described in this bug report. This report is therefore being closed with a resolution of ERRATA. For more information on the solution and/or where to find the updated files, please follow the link below. You may reopen this bug report if the solution does not work for you. http://rhn.redhat.com/errata/RHSA-2006-0575.html