Re: [take33 10/10] kevent: Kevent based AIO (aio_sendfile()/aio_sendfile_path()).

[Evgeniy Polyakov <johnpol@2ka.mipt.ru>]

On Fri, Jan 19, 2007 at 11:57:00AM +0530, Suparna Bhattacharya (suparna@in.ibm.com) wrote:
> > > Since you are implementing new APIs here, have you considered doing an
> > > aio_sendfilev to be able to send a header with the data ?
> > 
> > It is doable, but why people do not like corking?
> > With Linux less than microsecond syscall overhead it is better and more
> > flexible solution, doesn't it?
> 
> That is what I used to think as well. However ...
> 
> The problem as I understand it now is not about bunching data together, but
> of ensuring some sort of atomicity between the header and the data, when
> there can be multiple outstanding aio requests on the same socket - i.e
> ensuring strict ordering without other data coming in between, when data
> to be sent is not already in cache, and in the meantime another sendfile
> or aio write requests comes in for the same socket. Without having to lock
> the socket when reading data from disk.

No, socket locking is not solution at all here.
But the same applies to header - it will be copied into socket queue,
then socket will be unlocked and populated VFS data will be put into
that queue too, but there is a window between socket unlock after header
copy and file data copy. If we will hold socket lock after header is
copied, it is possible to lock it for too long - bad sectors on disk,
and reading might take forever.

> There are alternate ways to address this, aio_sendfilev is one of the options
> I have heard people requesting.

I bet those people worked with different Unix systems, which have much
slower syscalls, so they combine several operations into one call.

Only from this perspective I see any benefit from having header in the
syscall related to file transfer. Since I already "optimized" open()
syscall into file sending, things can not became worse if I will put there
header pointer too. I will schedule new kevent release with this change
somewhere after current work on M-on-N threading model.

> Regards
> Suparna

-- 
	Evgeniy Polyakov
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/