On Wed, Jun 11, 2008 at 12:07:49PM -0700, Andrew Morton wrote:
> On Wed, 11 Jun 2008 20:38:45 +0530
> "Aneesh Kumar K.V" <aneesh.kumar@xxxxxxxxxxxxxxxxxx> wrote:
>
> > On Thu, Jun 05, 2008 at 12:30:45PM -0700, Andrew Morton wrote:
> > > On Thu, 5 Jun 2008 22:35:12 +0530
> > > "Aneesh Kumar K.V" <aneesh.kumar@xxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > > We would like to get notified when we are doing a write on mmap
> > > > section. The changes are needed to handle ENOSPC when writing to an
> > > > mmap section of files with holes.
> > > >
> > >
> > > Whoa. You didn't copy anything like enough mailing lists for a change
> > > of this magnitude. I added some.
> > >
> > > This is a large change in behaviour!
> > >
> > > a) applications will now get a synchronous SIGBUS when modifying a
> > > page over an ENOSPC filesystem. Whereas previously they could have
> > > proceeded to completion and then detected the error via an fsync().
> >
> > Or not detect the error at all if we don't call fsync() right ? Isn't a
> > synchronous SIGBUS the right behaviour ?
> >
>
> Not according to POSIX. Or at least posix-several-years-ago, when this
> last was discussed. The spec doesn't have much useful to say about any
> of this.
>
> It's a significant change in the userspace interface.
>
> >
> > >
> > > It's going to take more than one skimpy little paragraph to
> > > justify this, and to demonstrate that it is preferable, and to
> > > convince us that nothing will break from this user-visible behaviour
> > > change.
> > >
> > > b) we're now doing fs operations (and some I/O) in the pagefault
> > > code. This has several implications:
> > >
> > > - performance changes
> > >
> > > - potential for deadlocks when a process takes the fault from
> > > within a copy_to_user() in, say, mm/filemap.c
> > >
> > > - performing additional memory allocations within that
> > > copy_to_user(). Possibility that these will reenter the
> > > filesystem.
> > >
> > > And that's just ext2.
> > >
> > > For ext3 things are even more complex, because we have the
> > > journal_start/journal_end pair which is effectively another "lock" for
> > > ranking/deadlock purposes. And now we're taking i_alloc_sem and
> > > lock_page and we're doing ->writepage() and its potential
> > > journal_start(), all potentially within the context of a
> > > copy_to_user().
> >
> > One of the reason why we would need this in ext3/ext4 is that we cannot
> > do block allocation in the writepage with the recent locking changes.
>
> Perhaps those recent locking changes were wrong.
>
> > The locking changes involve changing the locking order of journal_start
> > and page_lock. With writepage we are already called with page_lock and
> > we can't start new transaction needed for block allocation.
>
> ext3_write_begin() has journal_start() nesting inside the lock_page().
>
All those are changed as a part of lock inversion changes.
> > But if we agree that we should not do block allocation in page_mkwrite
> > we need to add writepages and allocate blocks in writepages.
>
> I'm not sure what writepages has to do with pagefaults?
>
The idea is to have ext3/4_writepages. In writepages start a transaction
and iterate over the pages take the lock and do block allocation. With
that change we should be able to not do block allocation in the
page_mkwrite path. We may still want to do block reservation there.
Something like.
ext4_writepages()
{
journal_start()
for_each_page()
lock_page
if (bh_unmapped()...)
block_alloc()
unlock_page
journal_stop()
}
ext4_writepage()
{
for_each_buffer_head()
if (bh_unmapped()) {
redirty_page
unlock_page
return;
}
}
--
To unsubscribe from this list: send the line "unsubscribe linux-ext4" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
