[dm-devel] Re: Why does __do_page_cache_readahead submit READ, not READA?

[Jens Axboe]

(not sure why your mailer stripped me from the CC?)


On Mon, Aug 03 2009, James Bottomley wrote:
> On Mon, 2009-08-03 at 09:59 +0200, Jens Axboe wrote:
> > On Mon, Aug 03 2009, Wu Fengguang wrote:
> > > On Thu, Jul 30, 2009 at 08:06:49AM +0200, Jens Axboe wrote:
> > > > On Wed, Jul 29 2009, Chris Mason wrote:
> > > > > On Wed, Jul 29, 2009 at 11:18:45PM +0200, Jens Axboe wrote:
> > > > > > On Wed, Jul 29 2009, Lars Ellenberg wrote:
> > > > > > > I naively assumed, from the "readahead" in the name, that readahead
> > > > > > > would be submitting READA bios. It does not.
> > > > > > > 
> > > > > > > I recently did some statistics on how many READ and READA requests
> > > > > > > we actually see on the block device level.
> > > > > > > I was suprised that READA is basically only used for file system
> > > > > > > internal meta data (and not even for all file systems),
> > > > > > > but _never_ for file data.
> > > > > > > 
> > > > > > > A simple
> > > > > > > 	dd if=bigfile of=/dev/null bs=4k count=1
> > > > > > > will absolutely cause readahead of the configured amount, no problem.
> > > > > > > But on the block device level, these are READ requests, where I'd
> > > > > > > expected them to be READA requests, based on the name.
> > > > > > > 
> > > > > > > This is because __do_page_cache_readahead() calls read_pages(),
> > > > > > > which in turn is mapping->a_ops->readpages(), or, as fallback,
> > > > > > > mapping->a_ops->readpage().
> > > > > > > 
> > > > > > > On that level, all variants end up submitting as READ.
> > > > > > > 
> > > > > > > This may even be intentional.
> > > > > > > But if so, I'd like to understand that.
> > > > > > 
> > > > > > I don't think it's intentional, and if memory serves, we used to use
> > > > > > READA when submitting read-ahead. Not sure how best to improve the
> > > > > > situation, since (as you describe), we lose the read-ahead vs normal
> > > > > > read at that level. I did some experimentation some time ago for
> > > > > > flagging this, see:
> > > > > > 
> > > > > > http://git.kernel.dk/?p=linux-2.6-block.git;a=commitdiff;h=16cfe64e3568cda412b3cf6b7b891331946b595e
> > > > > > 
> > > > > > which should pass down READA properly.
> > > > > 
> > > > > One of the problems in the past was that reada would fail if there
> > > > > wasn't a free request when we actually wanted it to go ahead and wait.
> > > > > Or something.  We've switched it around a few times I think.
> > > > 
> > > > Yes, we did used to do that, whether it was 2.2 or 2.4 I
> > > > don't recall :-)
> > > > 
> > > > It should be safe to enable know, whether there's a prettier way
> > > > than the above, I don't know. It works by detecting the read-ahead
> > > > marker, but it's a bit of a fragile design.
> > > 
> > > Another consideration is io-priority reversion and the overheads
> > > required to avoid it:
> > > 
> > >         readahead(pages A-Z)    => READA IO for pages A-Z
> > >         <short time later>
> > >         read(page A) => blocked => find the request that contains page A
> > >                                    and requeue/kick it as READ IO
> > > 
> > > The page-to-request lookups are not always required but nevertheless
> > > the complexity and overheads won't be trivial.
> > > 
> > > The page-to-request lookup feature would be also useful for "advanced"
> > > features like io-canceling (if implemented, hwpoison could be its
> > > first user ;)
> > 
> > I added that 3-4 years ago or so, to experiment with in-kernel
> > cancellation for things like truncate(). Tracking pages is not cheap,
> > and since the write cancelling wasn't really very sucessful, I didn't go
> > ahead with it.
> 
> Cancellation also came up several years ago with aio, which also has
> requirements for it.

Are they any real-world use cases for cancellation with aio?

> > So I'm not sure it's a viable alternative, even if we restricted it to
> > just tracking READA's, for instance.
> > 
> > But I don't think we have any priority inversion to worry about, at
> > least not from the CFQ perspective.
> 
> The basic problem with cancellation when implemented at the storage
> layer is that its an unusual operation.  The storage primitives which
> implement it aren't often invoked, so there's a lot of wariness to
> implementing them in practice.  For example, although SCSI has the abort
> function, it's not implemented by a lot of controllers, so we'd have to
> drop all pending I/O on the floor with a reset and then try and pick up
> the pieces we wanted.  Also, cancellation is racy since you never quite
> know if the I/O hit the storage or not.
> 
> On the back of this, we thought in 2003 or so that the best way to
> implement cancellation was simply to do nothing and have something wait
> around for the completion and throw it away.  This has exactly the same
> properties as storage implemented cancellation, but the benefit is that
> we don't have to perturb the storage stack to do it.

The approach I took (and exactly for the reasons you outline I think
it's the only feasible one) is to allow cancellations of unstarted IO
only. I can just imagine the pandora box we would open if we where to do
this at the hardware/device level as well, so just chose not to go that
far.

-- 
Jens Axboe