[dm-devel] Multipath v. 0.4.7 trips over USB flash drive

[Mike Anderson]

usvyatsky_ilya at emc.com <usvyatsky_ilya at emc.com> wrote:
> Hi there,
> 
> I am using DM MPIO for quite a while now and in general it works.
> Lately I encountered a problem with one of my boxes that is equipped
> with USB flash drive and is connected to a disk array.
> Namely, the flash drive has an empty bcdDevice entry, which translates
> into empty /sys/block/sdXX/device/rev file.
> As a result, the 'multipath -ll' command trips over this device and
> bails out without listing any array mpathXX devices.
> The devices are created in /dev/disk/by-name/ directory, so it's just
> the userland tool issue.
> 
> I've built debug version of multipath and found the following:
> 
> 1. The function sysfs_get_rev() (which is generated from macro
> declare_sysfs_get_str()) assumes that the rev has to have at least 2
> characters. Since in the case of my device this assumption is false, the
> function returns 1.
> 2. The caller function is scsi_sysfs_pathinfo() and it returns 1 as
> well.
> 3. The caller function is sysfs_pathinfo(), which returns 1 as well.
> 4. The caller function is pathinfo(), which also returns 1.
> 5. The caller function is store_pathinfo(). Here the pathinfo is not
> stored and NULL is returned to the caller.
> 6. The caller is path_discover() that returns 1.
> 7. The caller is path_discovery() that returns non-zero result.
> 8. The caller is configure() that bails out on this result.
> 
> I have tried to work around this issue by blacklisting the device, but
> unfortunately the only blacklist check that is done prior to call to
> store_pathinfo() in the path_discover() is based on the devnode. 
> This means that effectively I cannot use this workaround since with USB
> device I cannot guarantee that it will always be mapped to the same
> devnode.
> 
> I'd like to propose a couple of options for a fix for this issue:
> 
> 1. Re-implement sysfs_get_rev() in such a way that rev filed minimal
> length is no longer required. 
>    Please see attached a patch that is supposed to take care of this.
> 2. Re-implement pathinfo() to check whether vendor and/or product are
> blacklisted. 
>    This way non-compliant devices may be blacklisted based on vendor
> and/or product name.
> 
> I am not sure, which option is preferable, but I clearly need at least
> one of them in the product.
> 
> Your comments are greatly appreciated. 
> 
> Please note that my patch is against v. 0.4.7 of the multipath tool. 
> I realize that upstream may differ significantly, so if option (1) makes
> sense I'll revalidate with the latest version and submit the patch if
> necessary.

The upstream version already removes the minimum size.

You should try the multipath-tools git tree version to see if it addresses
your issue.

http://git.kernel.org/?p=linux/storage/multipath-tools/.git;a=summary

-andmike
--
Michael Anderson
andmike at linux.vnet.ibm.com