"Stephen C. Tweedie" wrote: > > Hi, > > On Thu, Feb 28, 2002 at 01:30:13PM -0800, Andrew Morton wrote: > > > There are a few things you can do: > > > > You'll want > > http://www.zip.com.au/~akpm/linux/patches/2.4/2.4.18-rc1/make_request.patch > > to fix the elevator starvation problem. This is officially in 2.4.19-pre1. > > >From which: > > + * a) A READA requester fails if free_requests < batch_requests > + * > + * We don't want READA requests to prevent sleepers from ever > + * waking. Note that READA is used extremely rarely - a few > + * filesystems use it for directory readahead. > > I'd actually like to see READA used more, but isn't it dangerous to > drop such requests? Under high IO load, readahead is actually likely > to reduce the number of seeks we incur, so surely it gets more, not > less, important to honour readahead as the IO queues fill up? > I've in fact proposed that READA simply be scrapped. I can't think of a situation in which it's the right thing to do. Or, at least, not enough for it to matter. And yes, dropping READA requests on the floor when we're under high load is the worst time to do it. But given that read requests come from a different queue from writes, I expect that you'd need a rather large number of reading processes before any requests were lost. But 2.4.19 isn't the place to be experimenting with READA, I guess. The above patch is in fact leaving the READA handling unaltered. -
