[lkml]   [2000]   [Dec]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: test12-pre5

On Tue, 5 Dec 2000, Daniel Phillips wrote:
> OK, I see - this isn't easy at all. You start the io if necessary, and
> some time later it completes.

Right. You don't know when. Once completed, it will unlock the page and
wake up waiters. It will also set PG_Uptodate if the read was successful
(and obviously it might not have been).

If you only care about the contents of the page, you can just test the
Uptodate flag - if the page is up-to-date you may not care whether IO is
outstanding on the page, or whether somebody is modifying page state
(removing page buffers etc). So it's entirely legal to look up a page in
the page cache and only look at uptodate.

(Of course, if it _isn't_ uptodate, you will still have to get the page
lock and re-test and possibly start the IO if it still isn't up-to-date
after you got the page lock. This is what all the generic_file_read()
stuff does for you).

> The locking state is therefore
> indeterminate after ->readpage or ->writepage; all we know is that
> after some finite amount of time the lock bit will go down.

Yes. It might have completed immediately (ramdisks or what not), or be
fast enough that by the time you get back it's already done. But the
likely case is that the page will be locked, and you'd be better off going
away and doing something else in the meantime (this is certainly important
for the VM layer - it needs to continue doing swap-outs so that it
doesn't just dribble the pages out one by one).

The main reason for the page locking changes for the writepage() case were

- the swapping code serializes accesses by using the page lock, and
doesn't have any other underlying serializing primitives. So we needed
to let the brw_page() code leave the lock set until the write has
physically completed.

- the VM code really wants to have a notion of "I have this many pages in
flight", so that it can sanely make a decision on when to start waiting
on page writeout completion, instead of just writing as much as it can.
The block device layer does some of this, of course, but the VM layer
can do more by just using the "nr_async_pages" thing. However, before
the unlock changes, this could not work for shared file mappings.

I hope that explains the change,


To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
Please read the FAQ at

 \ /
  Last update: 2005-03-22 12:51    [W:0.034 / U:0.012 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site