Christophe, This patch to dm_simplecmd of multipath-tools/libmultipath/devmapper.c will disable the use of lock_fs during I/O suspension while changing dm-multipath device mapper maps via the DM_DEVICE_RESUME ioctl of libdevmapper. Disabling lock_fs will result in not flushing I/O to the dm-multipath block device before changing the maps for the mapped device. Flushing I/O at this time appears not only unnecessary for dm-multipath devices but causes some troublesome delays due to SCSI HBA driver target-side timeouts when these maps are changed as a result of paths coming and going while upgrading software (a non-disruptive ucode upgrade) on a storage system. As is the case for any change which is dependent on similar changes in libdevmapper and the kernel, this patch will introduce a dependency on libdevmapper and kernel. Not sure which versions first had this support but at least libdevmapper 1.02.02 and linux kernel version 2.6.18-rc2 have the required support. diff --git a/libmultipath/devmapper.c b/libmultipath/devmapper.c index c0765ae..67c7740 100644 --- a/libmultipath/devmapper.c +++ b/libmultipath/devmapper.c @@ -105,6 +105,7 @@ dm_simplecmd (int task, const char *name goto out; dm_task_no_open_count(dmt); + dm_task_skip_lockfs(dmt); /* for DM_DEVICE_RESUME */ r = dm_task_run (dmt); @@ -183,6 +184,41 @@ out: } Thanks, Ed -- dm-devel@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/dm-devel
