>>>>> "Cong" == Cong Wang <amwang@xxxxxxxxxx> 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 from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
