systemd/extras/multipath-tools

Dependancies :
==============

o libdevmapper : comes with device-mapper-XXXX.tar.gz
  See www.sistina.com
  This lib has been dropped in the multipath tree
o libsysfs : comes with sysutils or udev
  See ftp.kernel.org/pub/linux/utils/kernel/hotplug/
o Linux kernel 2.6.0 with udm5 patchset
  http://people.sistina.com/~thornber/dm/
o udev
  See ftp.kernel.org/pub/linux/utils/kernel/hotplug/

How it works :
==============

Fill the all_paths array. Each path store this info :

struct path {
        char dev[FILE_NAME_SIZE];
        char sg_dev[FILE_NAME_SIZE];
        struct scsi_idlun scsi_id;
        struct sg_id sg_id;
        int state;
        char wwid[WWID_SIZE];
};

scsi_id, sg_dev and sg_id are only really useful for 2.4 
kernels, for which SG cmnds must go through sg devs.
In 2.5+ we have the nice opportunity to send SG cmnds
through SCSI bdevs.

For 2.4 compat, we pivot on idlun tupple to map sg devs
to SCSI bdevs.

2.4 does not do device enumeration, so we must scan a
defined number of sg devs and scsi bdevs. Good enough.
In 2.5+, we rely on libsysfs (sysutils) to access to
sysfs device enums.

the wwid is retrieved by a switch fonction. Only White
Listed HW can filled this field. For now every FC array 
HW listed in kernel's devinfo.c is White Listed, assuming
the WWID is stored is the SCSI-3 standard 0x83 EVPD page.

When all_paths is filled, we coalesce the paths and store
the result in mp array. Each mp is a struct like this :

struct multipath {
        char wwid[WWID_SIZE];
        int npaths;
        int pindex[MAX_MP_PATHS];
};

When mp is filled, the device maps are fed to the kernel
through libdevmapper.

The naming of the corresponding block device is handeld 
by udev with the help of the devmap_name proggy. It is 
called by the following rule in /etc/udev/udev.rules :
KERNEL="dm-[0-9]*", PROGRAM="/sbin/devmap_name %M %m", \
NAME="%k", SYMLINK="%c"

Notes :
=======

o On 2.4, make sure you have enough /dev/sg* nodes
  (/dev/MAKEDEV if necesary). Same goes for /dev/sd*

o path coalescing relies on a path unique id being found.
  This unique id, lacking a standard method, is vendor
  specific. A switch function (get_unique_id) is present
  and an example function is provided for storageworks
  arrays (get_evpd_wwid). Feel free to enrich
  with hardware you have at hand :)

o The kernel does NOT manage properly ghosts paths
  with StorageWorks HW. Seems nobody cares after a load
  of posts to linux-scsi.

o 2.4.21 version of DM does not like even segment size.
  if you enconter pbs with this, upgrade DM.

Credits :
=========

o Heavy cut'n paste from sg_utils. Thanks goes to D. 
  Gilbert.
o Light cut'n paste from dmsetup. Thanks Joe Thornber.
o Greg KH for the nice sysfs API.