Re: [PATCH] prune_icache_sb

["Nate Diller" <nate.diller@gmail.com>]

On 11/30/06, Wendy Cheng <wcheng@redhat.com> wrote:
> How about a simple and plain change with this uploaded patch ....
>
> The idea is, instead of unconditionally dropping every buffer associated
> with the particular mount point (that defeats the purpose of page
> caching), base kernel exports the "drop_pagecache_sb()" call that allows
> page cache to be trimmed. More importantly, it is changed to offer the
> choice of not randomly purging any buffer but the ones that seem to be
> unused (i_state is NULL and i_count is zero). This will encourage
> filesystem(s) to pro actively response to vm memory shortage if they
> choose so.
>
>  From our end (cluster locks are expensive - that's why we cache them),
> one of our kernel daemons will invoke this newly exported call based on
> a set of pre-defined tunables. It is then followed by a lock reclaim
> logic to trim the locks by checking the page cache associated with the
> inode (that this cluster lock is created for). If nothing is attached to
> the inode (based on i_mapping->nrpages count), we know it is a good
> candidate for trimming and will subsequently drop this lock (instead of
> waiting until the end of vfs inode life cycle).

I have a patch that is a more comprehensive version of this idea, but
it is not fully debugged, and has suffered some bitrot in the past
couple months.  This turns out to be a good performance improvement in
the general case too, but is more complex than your idea because there
are real locking changes needed to avoid deadlocks.  I can send you a
copy of the patch if you are interested.

> Note that I could do invalidate_inode_pages() within our kernel modules
> to accomplish what drop_pagecache_sb() does (without coming here to bug
> people) but I don't have access to inode_lock as an external kernel
> module. So either EXPORT_SYMBOL(inode_lock) or this patch ?

like i said above, you have to be careful when touching inode_lock,
dcache_lock, and the mapping's tree_lock, because of potential
deadlocks.  the mapping's lock can be taken from softirq context, but
the inode and dcache locks cannot.

NATE