Multipath I/O
Multipath I/O manages multiple paths to the same storage device for failover and, where supported, load distribution.
Redundant paths are alternative routes to the same logical storage, not additional replicas. Linux v6.12 looks up shared namespace paths within the same NVM subsystem and checks identifier consistency; an identifier match alone is not a global merge rule. [7] Its native policies prefer ANA-optimized paths and include numa, round-robin and queue-depth. [2]
Path selection and access exclusion solve different problems, because a surviving route does not establish which application node may write. For shared storage, check the deployed reservation or fencing mechanism across all relevant paths. Block-layer persistent reservations can limit writes, or all access, to initiators holding a registered key. [3]
All-paths-down queueing policy is a separate availability decision; inspect it alongside recovery timers. In the pinned device-mapper multipath manual, polling_interval defaults to 5 seconds and no_path_retry to fail; hardware-specific settings may differ. This is not a fixed application timeout. [5]
For SCSI path grouping, the pinned multipath manual uses a selected WWID and defaults uid_attribute to ID_SERIAL. That property is not necessarily a literal unit serial: systemd v257 can derive it from page 0x83. Preserve the identifier's source and scope when checking whether paths refer to one logical unit. [5] [8] [9]
A matching uid_attrs rule overrides other configured WWID methods. Otherwise uid_attribute selects a udev property; the SCSI default is ID_SERIAL. The sysfs method reads VPD page 0x83. Record the effective configuration and identifier source with the value instead of assuming every displayed WWID is a raw device designator. [5]
Relations
- related
- Dual-port drive, Fibre Channel, NVMe, ALUA, SCSI initiator and target, Asymmetric Namespace Access, I/O fencing, SCSI persistent reservations, I/O timeout, no_path_retry, LVM, ANA group, fast_io_fail_tmo (FC path failure), dev_loss_tmo (transport device loss), SCSI Device Identification, SCSI Unit Serial Number, World Wide Name (Fibre Channel)
Advantages
- Native NVMe path policies prefer ANA-optimized access when available; that preference does not measure the path's latency. [2]
Limitations
- A positive no_path_retry counts retries rather than seconds, so the time before I/O fails also depends on the polling interval. [5]
Citations
- Linux kernel documentation: Linux NVMe multipath, introduction and policies.
- Linux kernel documentation: Block layer support for Persistent Reservations.
- multipath-tools: multipath.conf(5), no_path_retry and FC timeout options; pinned commit d53932bdbee02f440518ea384e63d7c4d3f66685.
- Linux v6.12: nvme_find_ns_head and nvme_init_ns_head, subsystem-scoped shared-namespace and identifier checks.
- systemd v257: scsi_serial.c, check_fill_0x83_id, do_scsi_page80_inquiry and scsi_get_serial
- systemd v257: scsi_id.c, sanitized ID_SERIAL and ID_SERIAL_SHORT exports
Further reading
- Multipath I/O
- Storage fencing: why a lost heartbeat does not stop writes
- Storage timeouts: an expired timer is not a cancelled write
Sources & caveats
- linux-ana is the current Linux NVMe multipath page, read 2026-10-23 for the namespace merging and the numa, round-robin and queue-depth policies. linux-pr was read for the reservation types.
- multipath-man is multipath.conf(5) at the pinned commit, read for polling_interval, no_path_retry, fast_io_fail_tmo and dev_loss_tmo.
- Companion analyses of fencing and timeouts are listed but not used for any fact here.
- The nvme-multipath-io-policy entry links here because a policy is the rule that picks one path among those a multipath device manages.
- The 23 October revision qualifies namespace merging against Linux v6.12 source and removes a workload-performance generalization; no benchmark was performed.