Re: Questions regarding logging upon fsync in btrfs

[Josef Bacik <jbacik@fusionio.com>]

On Wed, Oct 02, 2013 at 10:12:20PM +0200, Aastha Mehta wrote:
> On 2 October 2013 13:52, Josef Bacik <jbacik@fusionio.com> wrote:
> > On Tue, Oct 01, 2013 at 10:13:25PM +0200, Aastha Mehta wrote:
> >> On 1 October 2013 21:42, Aastha Mehta <aasthakm@gmail.com> wrote:
> >> > On 1 October 2013 21:40, Aastha Mehta <aasthakm@gmail.com> wrote:
> >> >> On 1 October 2013 19:34, Josef Bacik <jbacik@fusionio.com> wrote:
> >> >>> On Mon, Sep 30, 2013 at 11:07:20PM +0200, Aastha Mehta wrote:
> >> >>>> On 30 September 2013 22:47, Josef Bacik <jbacik@fusionio.com> wrote:
> >> >>>> > On Mon, Sep 30, 2013 at 10:30:59PM +0200, Aastha Mehta wrote:
> >> >>>> >> On 30 September 2013 22:11, Josef Bacik <jbacik@fusionio.com> wrote:
> >> >>>> >> > On Mon, Sep 30, 2013 at 09:32:54PM +0200, Aastha Mehta wrote:
> >> >>>> >> >> On 29 September 2013 15:12, Josef Bacik <jbacik@fusionio.com> wrote:
> >> >>>> >> >> > On Sun, Sep 29, 2013 at 11:22:36AM +0200, Aastha Mehta wrote:
> >> >>>> >> >> >> Thank you very much for the reply. That clarifies a lot of things.
> >> >>>> >> >> >>
> >> >>>> >> >> >> I was trying a small test case that opens a file, writes a block of
> >> >>>> >> >> >> data, calls fsync and then closes the file. If I understand correctly,
> >> >>>> >> >> >> fsync would return only after all in-memory buffers have been
> >> >>>> >> >> >> committed to disk. I have added few print statements in the
> >> >>>> >> >> >> __extent_writepage function, and I notice that the function gets
> >> >>>> >> >> >> called a bit later after fsync returns. It seems that I am not
> >> >>>> >> >> >> guaranteed to see the data going to disk by the time fsync returns.
> >> >>>> >> >> >>
> >> >>>> >> >> >> Am I doing something wrong, or am I looking at the wrong place for
> >> >>>> >> >> >> disk write? This happens both with tree logging enabled as well as
> >> >>>> >> >> >> with notreelog.
> >> >>>> >> >> >>
> >> >>>> >> >> >
> >> >>>> >> >> > So 3.1 was a long time ago and to be sure it had issues I don't think it was
> >> >>>> >> >> > _that_ broken.  You are probably better off instrumenting a recent kernel, 3.11
> >> >>>> >> >> > or just build btrfs-next from git.  But if I were to make a guess I'd say that
> >> >>>> >> >> > __extent_writepage was how both data and metadata was written out at the time (I
> >> >>>> >> >> > don't think I changed it until 3.2 or something later) so what you are likely
> >> >>>> >> >> > seeing is the normal transaction commit after the fsync.  In the case of
> >> >>>> >> >> > notreelog we are likely starting another transaction and you are seeing that
> >> >>>> >> >> > commit (at the time the transaction kthread would start a transaction even if
> >> >>>> >> >> > none had been started yet.)  Thanks,
> >> >>>> >> >> >
> >> >>>> >> >> > Josef
> >> >>>> >> >>
> >> >>>> >> >> Is there any special handling for very small file write, less than 4K? As
> >> >>>> >> >> I understand there is an optimization to inline the first extent in a file if
> >> >>>> >> >> it is smaller than 4K, does it affect the writeback on fsync as well? I did
> >> >>>> >> >> set the max_inline mount option to 0, but even then it seems there is
> >> >>>> >> >> some difference in fsync behaviour for writing first extent of less than 4K
> >> >>>> >> >> size and writing 4K or more.
> >> >>>> >> >>
> >> >>>> >> >
> >> >>>> >> > Yeah if the file is an inline extent then it will be copied into the log
> >> >>>> >> > directly and the log will be written out, no going through the data write path
> >> >>>> >> > at all.  Max inline == 0 should make it so we don't inline, so if it isn't
> >> >>>> >> > honoring that then that may be a bug.  Thanks,
> >> >>>> >> >
> >> >>>> >> > Josef
> >> >>>> >>
> >> >>>> >> I tried it on 3.12-rc2 release, and it seems there is a bug then.
> >> >>>> >> Please find attached logs to confirm.
> >> >>>> >> Also, probably on the older release.
> >> >>>> >>
> >> >>>> >
> >> >>>> > Oooh ok I understand, you have your printk's in the wrong place ;).
> >> >>>> > do_writepages doesn't necessarily mean you are writing something.  If you want
> >> >>>> > to see if stuff got written to the disk I'd put a printk at run_delalloc_range
> >> >>>> > and have it spit out the range it is writing out since thats what we think is
> >> >>>> > actually dirty.  Thanks,
> >> >>>> >
> >> >>>> > Josef
> >> >>>>
> >> >>>> No, but I also placed dump_stack() in the beginning of
> >> >>>> __extent_writepage. run_delalloc_range is being called only from
> >> >>>> __extent_writepage, if it were to be called, the dump_stack() at the
> >> >>>> top of __extent_writepage would have printed as well, no?
> >> >>>>
> >> >>>
> >> >>> Ok I've done the same thing and I'm not seeing what you are seeing.  Are you
> >> >>> using any mount options other than notreelog and max_inline=0?  Could you adjust
> >> >>> your printk to print out the root objectid for the inode as well?  It could be
> >> >>> possible that this is the writeout for the space cache or inode cache.  Thanks,
> >> >>>
> >> >>> Josef
> >> >>
> >> >> I actually printed the stack only when the root objectid is 5. I have
> >> >> attached another log for writing the first 500 bytes in a file. I also
> >> >> print the root objectid for the inode in run_delalloc and
> >> >> __extent_writepage.
> >> >>
> >> >> Thanks
> >> >>
> >> >
> >> > Just to clarify, in the latest logs, I allowed printing of debug
> >> > printk's and stack dump for all root objectid's.
> >>
> >> Actually, it is the same behaviour when I write anything less than 4K
> >> long, no matter what offset, except if I straddle the page boundary.
> >> To summarise:
> >> 1. write 4K -> write in the fsync path
> >> 2. write less than 4K, within a single page -> bdi_writeback by flush worker
> >> 3. small write that straddles a page boundary or write 4K+delta -> the
> >> first page gets written in the fsync path, the remaining length that
> >> straddles the page boundary is written in the bdi_writeback path
> >>
> >> Please let me know, if I am trying out incorrect cases.
> >>
> >> Sorry for too many mails.
> >>
> >
> > This has been bugging me so much I was dreaming about it and now here I am
> > writing an email at 4:45 in the morning ;).  So I couldn't reproduce earlier
> > with any of these scenarios and then I realized something, I'm doing something
> > like this
> >
> > xfs_io -f -c "pwrite 0 54" -c "fsync" /mnt/btrfs-test/foo
> >
> > and it is working perfectly.  But I bet what you are doing is something like
> > this
> >
> > file = fopen("/mnt/btrfs-test/foo");
> > fwrite(buf, 54, 1, file);
> > fsync(fileno(file));
> > fclose(file);
> >
> > right?  Please say yes :).  If this is the case then it is likely that these
> > small writes are getting buffered in the userspace buffering that comes with
> > fwrite, and so when you fsync it is only flushing the data that is actually in
> > the kernel, not what is buffered in userspace.  Then when you fclose it flushes
> > what is in the userspace buffers out to the kernel and then later on the
> > background writer comes in and writes out the dirty data.  To fix this you want
> > to do fflush() and then fsync().  Hopefully that is what you are doing and I can
> > go back to sleep, thanks,
> >
> > Josef
> 
> Indeed!! :)
> 
> I did mention I am using f* version of the POSIX API. Sorry for the
> confusion. Calling fflush before fsync seems to write everything
> perfectly. It works even without notreelog option, as it should have.
> I was under the misconception that fflush and fsync do the same thing.
> 
> Thanks a lot for your quick help.
> 

Yeah sorry all I read was "there is a bug in fsync" and just assumed we broke
something ;).  Glad I could help,

Josef