> On Jul 27, 2016, at 07:55, Benjamin Coddington <bcodding@xxxxxxxxxx> wrote:
>
> After adding more debugging, I see that all of that is working correctly,
> but the first LAYOUTCOMMIT is taking the size back down to 4096 from the
> last nfs_writeback_done(), and the second LAYOUTCOMMIT never brings it back
> up again.
>
Excellent! Thanks for debugging that.
> Now I see that we should be marking the block extents as written atomically with
> setting LAYOUTCOMMIT and nfsi->layout->plh_lwb, otherwise a LAYOUTCOMMIT can
> collect extents just added from the next bl_write_cleanup(). Then, the next
> LAYOUTCOMMIT fails, and all we're left with is the size from the first
> LAYOUTCOMMIT. Not sure if that particular problem is the whole fix, but
> that's something to work on.
>
> I see ways to fix that:
>
> - make a new pnfs_set_layoutcommit_locked() that can be used to call
> ext_tree_mark_written() inside the i_lock
>
> - make another pnfs_layoutdriver_type operation to be used within
> pnfs_set_layoutcommit (mark_layoutcommit? set_layoutcommit?), and call
> ext_tree_mark_written() within that..
>
> - have .prepare_layoutcommit return a new positive plh_lwb that would
> extend the current LAYOUTCOMMIT
>
> - make ext_tree_prepare_commit only encode up to plh_lwb
I see no reason why ext_tree_prepare_commit() shouldn’t be allowed to extend the args->lastbytewritten. This is a metadata operation that is owned by the pNFS layout driver.
The only thing I’d note is you should then rewrite the failure case in pnfs_layoutcommit_inode() so that it doesn’t rely on the saved “end_pos”, but uses args->lastbytewritten instead (with a comment to the effect why)…
Cheers
Trond��.n��������+%������w��{.n�����{��w��� jg��� �����ݢj����G���� ���j:+v���w�m������w�������h�����٥
