Re: [PATCH] blk queue io tracing support

[Tom Zanussi <zanussi@us.ibm.com>]

Nathan Scott writes:
 > Hi Tom,
 > 
 > On Tue, Aug 30, 2005 at 11:19:04PM -0500, Tom Zanussi wrote:
 > > You're right, it should be using simple_rmdir rather than
 > > simple_unlink for removing directories.  Thanks for sending the patch,
 > 
 > No problem.
 > 
 > > which I've modified a bit to avoid splitting the rmdir/unlink cases
 > > into separate functions, since they're almost the same except for what
 > > they end up calling.  relayfs_remove_dir now doesn't do anything but
 > > call relayfs_remove (it didn't do much more than that before anyway),
 > > but it makes sense to me to keep it, as the counterpart to
 > > relayfs_create_dir.  Let me know if you see any problems with it.
 > 
 > Looks OK, I'll give it a spin.
 > 
 > On an unrelated note, are there any known issues with using epoll
 > on relayfs file descriptors?  I'm having a few troubles, and just
 > wondering if its me doing something silly, or if its known to not
 > work...?  Symptoms of the problem are epoll continually reaching
 > its timeout with no modified fds found (when I know the inode has
 > modified trace buffers attached) ... and the epoll code is a bit
 > too hairy for me to go find a quick fix - seems like it should be
 > able to work though since relayfs has a ->poll implementation.

Well, the relayfs poll implementation is based on completed
sub-buffers, so you can be writing events into a buffer, but until a
buffer switch happens, you won't be notified that anything's changed.
The reason for the sub-buffer granularity is that relayfs was
originally meant for use only with mmap(), but now that there's a
read(), I'll probably have to make some changes to the poll
implementation as well.

Tom