A mirror logical volume can itself have a mirrored log device. When a device in an image of the mirror and its log failed at the same time, it was possible for I/O errors to appear on the mirror LV when they should have been handled. That is, the kernel would not absorb the I/O errors from the failed device by relying on the remaining device. This would then cause file systems built on the device to respond to the I/O errors (turn read-only in the case of the ext3/4 file systems).
The cause was found to be that the mirror was not suspended for repair using the 'noflush' flag. This flag allows the kernel to requeue I/O requests that need to be retried. Because the kernel was not allowed to requeue the requests, it had no choice but to return the I/O as errored. This issue has been corrected and the mirror is now properly suspended with the 'noflush' flag.