Re: [patch rfc] towards supporting O_NONBLOCK on regular files

["Stephen C. Tweedie" <sct@redhat.com>]

Hi,

On Fri, 2004-10-15 at 16:44, Jeff Moyer wrote:

> I got the partial read case wrong in the last patch.  In fact, it looks
> like this code path would perform infinite retries before.  This should
> address that by returning upon the first partial read.  Attached is a new
> version of the patch.

...

> --- linux-2.6.9-rc4-mm1/mm/filemap.c.orig	2004-10-15 10:33:24.986209880 -0400
> +++ linux-2.6.9-rc4-mm1/mm/filemap.c	2004-10-15 11:38:50.869384920 -0400
> @@ -976,10 +987,10 @@ __generic_file_aio_read(struct kiocb *io
>  			desc.error = 0;
>  			do_generic_file_read(filp,ppos,&desc,file_read_actor);
>  			retval += desc.written;
> -			if (!retval) {
> +			if (!retval)
>  				retval = desc.error;
> +			if (desc.written != iov[seg].iov_len)
>  				break;
> -			}
>  		}

Yep; this chunk used to break out only on !retval, so at worst it's a
data corrupter for multi-segment iovecs.  If one do_generic_file_read()
returned a short read we'd update retval but then move on to the next
segment of the iovec; if the next one succeeded we'd be reading into the
next segment but ppos wouldn't be correctly advanced.

The fix looks correct --- we still only return an error if no data at
all has been returned, but we break out of the IO completely at the
first short read.

--Stephen