On Mon, Sep 22, 2008 at 1:04 PM, Aaron Straus <aaron-bYFJunmd+ZV8UrSeD/[email protected]> wrote:
> On Sep 22 12:35 PM, Trond Myklebust wrote:
>> Revert _what_ exactly?
> Yep. I narrowed the problem down to an offending hunk in a particular
> patch. Removing that hunk did eliminate the problem. However,
> reverting that hunk is likely wrong and the code has changed _a lot_
> since that commit.
>> My understanding was that this is a consequence of unordered writes
>> causing the file to be extended while some other task is reading.
> Yes. I added some debugging statements to look at the writeout path.
> I think the following happens:
> - page 0 gets partially written to by app
> - VM writes out partial dirty page 0
> - page 0 gets fully written by app
> - page 1 gets partially written by app
> - VM writes out dirty page 1
> At this point there is a hole in the file. The tail end of page 0 is
> still not written to server.
> - VM writes out dirty page 0
>> AFAICS, this sort of behaviour has _always_ been possible. I can't see
>> how reverting anything will fix it.
> Here is the crux. It was possible previously but unlikely e.g. our app
> never saw this behavior. The new writeout semantics causes visible
> holes in files often.
> Anyway, I agree the new writeout semantics are allowed and possibly
> saner than the previous writeout path. The problem is that it is
> __annoying__ for this use case (log files).
> I'm not sure if there is an easy solution. We want the VM to writeout
> the address space in order. Maybe we can start the scan for dirty
> pages at the last page we wrote out i.e. page 0 in the example above?
Why can't you use O_SYNC | O_APPEND ?