[dm-devel] failed multipaths are never re-validated

[christophe.varoqui at free.fr]

> 
> I was hitting the same bug. Here is a patch that correctly clears out the
> currpath structure, and fixes things.
> 
your patch seems truncated,
but anyway I fixed that today

> But there is a tiny bug remaining when using the qla2xxx driver it will
> set the host byte but when you do sg io to a /dev/sdX (vs /dev/sgX) entry
> the block/scsi_ioctl.c sgio does not pass these errors up. So you end up
> with a device like this:
> 
> dmsetup deps
> 
> 00000000000000000000000000000000: 4 dependencies        : (8, 144) (8,
> 128) (8, 112) (8, 96)
> 

Yes, I'm bitten hard by this sgio problem : TUR seem reliable when done on
/dev/sg* and not reliable when done on /dev/sd*.

As I don't want to add to complexity of the /dev/sd -> /dev/sg translation, I
would like to know if it is possible to correct this in the kernel ...

Another, related, problem is that readsector0 is unsuable too : read() seems
very upset with the "reservation" of the individual paths by the device mapper.

Insights ?