Multipath and device-mapper - redundant paths to SAN LUNs · Linux

Multipath is the standard way to provide redundant paths from a host to a SAN LUN. With multipathd running, the kernel automatically fails over between paths when one becomes unavailable, and the I/O continues uninterrupted.

How multipath works

Read-only / Safemultipath topology— Four paths (sda, sdb, sdc, sdd) to one LUN, arranged in two path groups. The group with prio=50 is status=active and carries the I/O; the prio=10 group is status=enabled, meaning healthy but not in use. Two groups with different priorities is ALUA reporting which controller owns the LUN: the higher-priority group is the optimised path set. If every path in the active group fails, multipathd promotes the enabled group and I/O continues over the non-optimised controller — working, but slower.

$ multipath -ll /dev/mapper/mpath0

mpath0 (3600508...) dm-2 NETAPP,LUN
size=500G features='1 queue_if_no_path' wp=rw
|-+- policy='service-time 0' prio=50 status=active
| |- 1:0:0:0 sda 8:0   active ready running
| `- 1:0:1:0 sdb 8:16  active ready running
`-+- policy='service-time 0' prio=10 status=enabled
|- 1:0:2:0 sdc 8:32  active ready running
`- 1:0:3:0 sdd 8:48  active ready running

Configuring multipath

Read-only / Safemultipath config— This is the config that produces the topology above. user_friendly_names gives names like mpath0 instead of a raw WWN. find_multipaths yes creates multipath devices only where multiple paths actually exist. group_by_prio with prio alua asks the array which controller owns the LUN and puts the optimised paths in their own higher-priority group - that is what produces the prio=50 and prio=10 groups. path_checker tur sends a TEST UNIT READY to each path on every polling interval. failback immediate returns to the optimised group as soon as it recovers. no_path_retry 18 is the setting worth arguing about; see below.

$ cat /etc/multipath.conf | head -30

defaults {
user_friendly_names   yes
find_multipaths       yes
path_grouping_policy  group_by_prio
prio                  alua
path_selector         "service-time 0"
path_checker          tur
failback              immediate
no_path_retry         18
}

no_path_retry decides what happens when every path is gone

This one keyword chooses the failure mode of every filesystem on the LUN, and the two obvious values are both traps.

no_path_retry fail returns I/O errors the instant the last path drops. A SAN controller failover, a fabric reconvergence or a switch reload — events that resolve themselves in ten to thirty seconds — surface as write errors, and ext4 remounts read-only underneath a running database.

no_path_retry queue never gives up. I/O blocks forever, the processes touching it go into uninterruptible D state where no signal reaches them, and the host may refuse to shut down cleanly because the shutdown itself needs to write. man 5 multipath.conf carries this under KNOWN ISSUES for exactly that reason.

A finite count is the usual production answer. Each retry is one polling_interval (5 seconds by default), so no_path_retry 18 queues for roughly 90 seconds — long enough to ride out a controller failover, short enough that a genuine outage still surfaces as an error rather than a hang.

The exception is a clustered node that will be fenced. There you want fail, because a node blocked on queued I/O cannot complete its side of the fence.

Note that the deprecated queue_if_no_path feature and no_path_retry are the same control expressed two ways; if both are set, no_path_retry wins.

Data-loss riskstart multipathd— Enable and start multipathd. The -r flag rescans existing devices and creates multipath devices for them. New LUNs that appear after multipathd starts are automatically detected and added to multipath devices.

$ systemctl enable --now multipathd; multipath -r; ls /dev/mapper/mpath*

Created symlink /etc/systemd/system/multipath-tools.service...
Created symlink /etc/systemd/system/systemd...service...
mpath0
mpath1

How multipath works

Configuring multipath

Knowledge check