Re: [RFC Patch] fs: implement per-file drop caches

["John Stoffel" <john@stoffel.org>]

>>>>> "Cong" == Cong Wang <amwang@redhat.com> writes:

Cong> On Thu, 2012-05-31 at 15:09 -0400, KOSAKI Motohiro wrote:
>> (5/31/12 8:11 AM), Cong Wang wrote:
>> > On Thu, 2012-05-31 at 02:30 -0400, KOSAKI Motohiro wrote:
>> >> (5/31/12 2:20 AM), Cong Wang wrote:
>> >>> On Wed, 2012-05-30 at 16:14 +0100, Pádraig Brady wrote:
>> >>>> On 05/30/2012 02:38 PM, Cong Wang wrote:
>> >>>>> This is a draft patch of implementing per-file drop caches.
>> >>>>>
>> >>>>> It introduces a new fcntl command  F_DROP_CACHES to drop
>> >>>>> file caches of a specific file. The reason is that currently
>> >>>>> we only have a system-wide drop caches interface, it could
>> >>>>> cause system-wide performance down if we drop all page caches
>> >>>>> when we actually want to drop the caches of some huge file.
>> >>>>
>> >>>> This is useful functionality.
>> >>>> Though isn't it already provided with POSIX_FADV_DONTNEED?
>> >>>
>> >>> Thanks for teaching this!
>> >>>
>> >>> However, from the source code of madvise_dontneed() it looks like it is
>> >>> using a totally different way to drop page caches, that is to invalidate
>> >>> the page mapping, and trigger a re-mapping of the file pages after a
>> >>> page fault. So, yeah, this could probably drop the page caches too (I am
>> >>> not so sure, haven't checked the code in details), but with my patch, it
>> >>> flushes the page caches directly, what's more, it can also prune
>> >>> dcache/icache of the file.
>> >>
>> >> madvise should work. I don't think we need duplicate interface. Moreomover
>> >> madvise(2) is cleaner than fcntl(2).
>> >>
>> >
>> > I think madvise(DONTNEED) attacks the problem in a different approach,
>> > it munmaps the file mapping and by the way drops the page caches, my
>> > approach is to drop the page caches directly similar to what sysctl
>> > drop_caches.
>> >
>> > What about private file mapping? Could madvise(DONTNEED) drop the page
>> > caches too even when the other process is doing the same private file
>> > mapping? At least my patch could do this.
>> 
>> Right. But a process can makes another mappings if a process have enough
>> permission. and if it doesn't, a process shouldn't be able to drop a shared
>> cache.
>> 

Cong> Ok, then this patch is not a dup of madvise(DONTNEED).

>> 
>> > I am not sure if fcntl() is a good interface either, this is why the
>> > patch is marked as RFC. :-D
>> 
>> But, if you can find certain usecase, I'm not against anymore.
>> 

Cong> Yeah, at least John Stoffel expressed his interests on this, as
Cong> a sysadmin. So I believe there are some people need it.

I expressed an interest if there was a way to usefully *find* the
processes that are hogging cache.  Without a reporting mechanism of
cache usage on per-file or per-process manner, then I don't see a
great use for this.  It's just simpler to drop all the caches when you
hit a wall.  

Cong> Now the problem is that I don't find a proper existing utility
Cong> to patch, maybe Pádraig has any hints on this? Could this
Cong> feature be merged into some core utility? Or I have to write a
Cong> new utility for this?

I'd write a new tutorial utility, maybe you could call it 'cache_top'
and have it both show the biggest users of cache, as well as exposing
your new ability to drop the cache on a per-fd basis.

It's really not much use unless we can measure it.

John

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>