Re: [PATCH 0/4 v3] fs: Remove i_devices from struct inode

[Andy Lutomirski <luto@amacapital.net>]

On Tue, Nov 4, 2014 at 11:55 AM, Jan Kara <jack@suse.cz> wrote:
> On Tue 04-11-14 11:47:21, Andy Lutomirski wrote:
>> On 11/04/2014 07:39 AM, Al Viro wrote:
>> > On Tue, Nov 04, 2014 at 11:27:27AM +0100, Jan Kara wrote:
>> >>   Hello,
>> >>
>> >>   this patch set removes use of i_devices from block and character device
>> >> code and thus we can remove the list head from struct inode thus saving two
>> >> pointers in it. As Christoph has reviewed the series, can you please merge
>> >> it Al? Thanks!
>> >>
>> >> Since v2 I've added reviewed-by tags from Christoph and changed one variable
>> >> name in cdev_forget().
>> >>
>> >> Since v1 I have split the patches and properly handled character devices (I
>> >> broke them last time as Christoph pointed out).
>> >
>> > My problem with that is in buggered module refcounts (which was the reason
>> > for doing those non-counting references back then).  Suppose you open
>> > /dev/some_char_device and close it; having the module pinned down until
>> > the inode of that sucker gets evicted by dcache/icache memory pressure
>> > would be wrong - it _isn't_ in use, and there's no way short of forcing
>> > the full eviction of VFS caches to get it possible to unload...
>> >
>>
>> At the risk of asking what may be a rather dumb question...
>>
>> Why do device node inodes need to be cached at all?  In other words,
>> when you try open a device node, can't the kernel materialize the inode
>> from just information that's in the dentry without touching the
>> filesystem at all?  If that's true, couldn't all device inodes be
>> dropped from icache as soon as they're unreferenced?
>>
>> (Yes, there's mtime, but I never understood why tracking mtime on device
>> nodes made any sense in the first place.)
>   I can see a few reasons:
> 1) positive dentry without inode - no-no for dcache.
> 2) how would you get the information which device the dentry references?
> 3) what would you gain to outweight the complications and special code
>    paths?
>

Yeah, this idea clearly doesn't work.  But I wonder whether the
revised variant (which may be just moving cd_forget and bd_forget from
evict to iput_final) would work.

--Andy

>                                                                 Honza
> --
> Jan Kara <jack@suse.cz>
> SUSE Labs, CR



-- 
Andy Lutomirski
AMA Capital Management, LLC