On wed, 17 Nov 2010 11:55:28 -0500, Chris Mason wrote:
Excerpts from Josef Bacik's message of 2010-11-17 07:50:11 -0500:
On Wed, Nov 17, 2010 at 06:11:03PM +0800, Miao Xie wrote:
Hi, Josef
On wed, 17 Nov 2010 04:37:21 -0500, Josef Bacik wrote:
Heh so I was going to fix this after the hole punching stuff. The fact is btrfs
maps everything that is ok to do in one IO via get_blocks(). So all we need to
do is add another DIO_ flag to tell us to treat each get_blocks() call as
discrete. I wanted to use buffer_boundary for this, but I think it's too
drastic of a change for people who already use buffer_boundary();
What happens today is that say we map 4k, we do submit_page_section, but if this
is our first bit of IO we just set dio->cur_page and such and then loop again.
Say there is 4k-hole-4k, we do the next mapping and set buffer_boundary again,
and come into submit_page_section and because cur_page is set, we do
dio_send_cur_page. Because there is no dio->bio we setup a new bio, but when we
do that we clear dio->boundary, and leave the bio all setup. So the next time
we loop around the tail 4k gets added to our previously setup bio and boom we
hit this problem with btrfs.
If we can add a DIO_GET_BLOCKS_DISCRETE or some other such non-sense then we can
easily kill all the logical offset code I had and just make some simple changes
to make the DIO stuff work for us. All we do is in get_more_blocks we do
if ((dio->flags& DIO_GET_BLOCKS_DISCRETE)&& dio->bio)
dio_submit_bio(dio);
Right after I went to bed I realized this should be
if (dio->flags& DIO_GET_BLOCKS_DISCRETE) {
if (dio->cur_page) {
dio_send_cur_page(dio);
page_cache_release(dio->cur_page);
dio->cur_page = NULL;
}
if (dio->bio)
dio_submit_bio(dio);
}
As far as I know, get_block() can not make sure the IO doesn't span the chunks or
stripes. Maybe we can do this check in get_blocks(). In this way, we needn't change
vfs.
Right thats the idea, if we can't span chunks/stripes we should be doing that
limiting in our get_blocks call and that way we don't have to screw with the
generic direct io stuff too much. Thanks,
In this case we're adding complexity to the O_DIRECT mapping code, when
we really should be adding it to the btrfs submit bio hook. It can
easily break up the bio into smaller units, which will leave us with a
smaller number of get_blocks calls overall.
I'm working that out now.
Do you mean you are fixing this bug now?
Thanks
Miao
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
