Re: DM's filesystem lookup in dm_get_dev_t() [was: Re: linux-next: manual merge of the device-mapper tree with Linus' tree]

[Christoph Hellwig]

On Tue, Dec 22, 2020 at 06:24:09PM +0100, Hannes Reinecke wrote:
> Ok. The problem from my perspective is that device-mapper needs to
> a) ensure that the arbitrary string passed in with the table definition 
> refers to a valid block device
> and
> b) the block device can be opened with O_EXCL, so that device-mapper can 
> then use it.
>
> Originally (ie prior to commit 644bda6f3460) dm_get_device() just converted 
> the string to a 'dev_t' representation, and then the block device itself 
> was checked and opened in dm_get_table_device().
> 'lookup_bdev' was just being used to convert the path if the string was not 
> in the canonical major:minor format, as then it was assumed that it 
> referred to a block device node, and then lookup_bdev kinda makes sense.

Yes, 644bda6f3460 is the cause of all evil, as it uses an API in a place
where it should not be used.  It and the prep patch
(e6e20a7a5f3f49bfee518d5c6849107398d83912) which did the grave mistake
of making name_to_dev_t available outside of the early init code really
need to be reverted.

> However, lookup_bdev() now always recurses into the filesystem, causing 
> multipath to stall in an all-paths-down scenario.

lookup_bdev always did a file system lookup, and always only accepted
a valid name in the file system.

> Alternatively, if Mike says that only major:minor is the valid format for a 
> table definition we can kill that code completely. But clearly _I_ cannot 
> make the call here.

Before 644bda6f3460 the table definitions only accepted a valid name in
the file system.  Which is the proper interface.