Messages in this thread | | | Date | Fri, 03 Dec 2010 13:24:59 -0500 | From | Ric Wheeler <> | Subject | Re: [PATCH 1/2] fs: Do not dispatch FITRIM through separate super_operation |
| |
On 11/18/2010 12:55 PM, Chris Mason wrote: > Excerpts from James Bottomley's message of 2010-11-18 12:19:10 -0500: >> On Thu, 2010-11-18 at 09:29 -0500, Christoph Hellwig wrote: >>> On Thu, Nov 18, 2010 at 07:19:58AM -0700, Matthew Wilcox wrote: >>>> I guess I was assuming that, on receiving a FALLOC_FL_PUNCH_HOLE, a >>>> filesystem that was TRIM-aware would pass that information down to the >>>> block device that it's mounted on. I strongly feel that we shouldn't >>>> have two interfaces to do essentially the same thing. >>>> >>>> I guess I'm saying that you're going to have to learn about TRIM :-) >>> Did you actually look Lukas FITRIM code (not the slight reordering here, >>> but the original one). It's the ext4 version of the batched discard >>> model, that is a userspace ioctl to discard free space in the >>> filesystem. >>> >>> hole punching will free the blocks into the free space pool. If you do >>> online discard it will also get discarded, but a filesystem that has >>> online discard enabled doesn't need FITRIM. >> Not stepping into the debate: I'm happy to see punch go to the mapping >> data and FITRIM pick it up later. >> >> However, I think it's time to question whether we actually still want to >> allow online discard at all. Most of the benchmarks show it to be a net >> lose to almost everything (either SSD or Thinly Provisioned arrays), so >> it's become an "enable this to degrade performance" option with no >> upside. > I think we want to keep it. In general we've (except for hch) spent > almost zero time actually tuning online discard, and the benchmarking > needs to be redone with the shiny new barrier code. > > -chris >
Very belated response - I agree that we should keep the online discard support in (but off by default).
Some of the devices we have tested perform well with it and I expect that hardware vendors will get better now that we have the support for them to test with.
Ric
| |