Re: [Qemu-devel] [PATCH v2 17/21] block: Move cache options into options QDict

[Kevin Wolf]

Am 27.11.2015 um 20:57 hat Max Reitz geschrieben:
> On 23.11.2015 16:59, Kevin Wolf wrote:
> > This adds the cache mode options to the QDict, so that they can be
> > specified for child nodes (e.g. backing.cache.direct=off).
> > 
> > The cache modes are not removed from the flags at this point; instead,
> > options and flags are kept in sync. If the user specifies both flags and
> > options, the options take precedence.
> > 
> > Child node inherit cache modes as options now, they don't use flags any
> > more.
> > 
> > Note that this forbids specifying the cache mode for empty drives. It
> > didn't make sense anyway to specify it there, because it didn't have any
> > effect. blockdev_init() considers the cache options now bdrv_open()
> > options and therefore doesn't create an empty drive any more but calls
> > into bdrv_open(). This in turn will fail with no driver and filename
> > specified.
> > 
> > Signed-off-by: Kevin Wolf <address@hidden>
> > ---
> >  block.c    | 100 
> > ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++---
> >  blockdev.c |  52 ++++++++++----------------------
> >  2 files changed, 111 insertions(+), 41 deletions(-)
> > 
> > diff --git a/block.c b/block.c
> > index eff0c19..397014a 100644
> > --- a/block.c
> > +++ b/block.c
> 
> [...]
> 
> > @@ -1747,12 +1816,22 @@ static BlockReopenQueue 
> > *bdrv_reopen_queue_child(BlockReopenQueue *bs_queue,
> >      /*
> >       * Precedence of options:
> >       * 1. Explicitly passed in options (highest)
> > -     * 2. TODO Set in flags (only for top level)
> > +     * 2. Set in flags (only for top level)
> >       * 3. Retained from explicitly set options of bs
> >       * 4. Inherited from parent node
> >       * 5. Retained from effective options of bs
> >       */
> >  
> > +    if (!parent_options) {
> > +        /*
> > +         * Any setting represented by flags is always updated. If the
> > +         * corresponding QDict option is set, it takes precedence. 
> > Otherwise
> > +         * the flag is translated into a QDict option. The old setting of 
> > bs is
> > +         * not considered.
> > +         */
> > +        update_options_from_flags(options, flags);
> > +    }
> > +
> >      /* Old explicitly set values (don't overwrite by inherited value) */
> >      old_options = qdict_clone_shallow(bs->explicit_options);
> >      bdrv_join_options(bs, options, old_options);
> > @@ -1923,6 +2002,19 @@ int bdrv_reopen_prepare(BDRVReopenState 
> > *reopen_state, BlockReopenQueue *queue,
> >          goto error;
> >      }
> >  
> > +    update_flags_from_options(&reopen_state->flags, opts);
> > +
> > +    /* If a guest device is attached, it owns WCE */
> > +    if (reopen_state->bs->blk && 
> > blk_get_attached_dev(reopen_state->bs->blk)) {
> > +        bool old_wce = bdrv_enable_write_cache(reopen_state->bs);
> > +        bool new_wce = (reopen_state->flags & BDRV_O_CACHE_WB);
> > +        if (old_wce != new_wce) {
> > +            error_setg(errp, "Cannot change cache.writeback: Device 
> > attached");
> > +            ret = -EINVAL;
> > +            goto error;
> > +        }
> > +    }
> > +
> 
> Time to get back to my question regarding bdrv_set_enable_write_cache():

Alles kaputt. :-)

> 1. bdrv_set_enable_write_cache() sets/unsets BDRV_O_CACHE_WB in
>    bs->open_flags so that it is preserved through a reopen.
> 2. On reopen, bdrv_reopen_queue_child() calls
>    update_options_from_flags() unless @parent_options is set. If that
>    function is called, it will set the appropriate caching options in
>    @options as dictated by bdrv_set_enable_write_cache().
> 3. If @parent_options was NULL, the update_flags_from_options() call
>    here in bdrv_reopen_prepare() will set BDRV_O_CACHE_WB just as
>    dictated by bdrv_set_enable_write_cache() (unless overwritten).
>    That is what we want.
> 4. If @parent_options was not NULL, the caching options in
>    bs->open_flags are completely overwritten, discarding everything that
>    had been set by bdrv_set_enable_write_cache(). That's not so good.
> 
> @parent_options is tested so that update_options_from_flags() is called
> only for root BDSs. It is possible to call bdrv_set_enable_write_cache()
> on non-root BDSs (e.g. commit_active_start() does it on the commit base
> via mirror_start_job()), so I do think overriding the setting made by
> bdrv_set_enable_write_cache() on reopen for any non-root BDSs is not
> correct.

I think you didn't interpret @parent_options correctly. It is NULL for
the node that bdrv_reopen() is called for. It is non-NULL for
(recursive) child nodes of that node that inherit options from it. Not
sure if it matters for this case, though.

In my opinion, what these block jobs are doing is insane. They should
never touch the cache mode of a node as long as this cache mode can be
shared with other users. It doesn't matter much with block jobs that can
only work on a BDS that they just created, but with blockdev-backup I'd
consider this a real bug.

> Maybe bdrv_set_enable_write_cache() should set cache.writeback in
> bs->explicit_options to on or off instead of using bs->open_flags. But
> in that case, we can no longer use bdrv_set_enable_write_cache() in
> bdrv_open_common(), because the cache setting there may be implicit.

The _real_ thing would be moving cache.writeback to the BB. That's out
of scope for this series, though.

I don't like to enbable the guest to change bs->explicit_options,
because that doesn't feel very explicit, but I might consider it if we
find a use case where it's necessary in order to get the desired
behaviour. In the light of the above, however, I'm inclined to think
that it's an (accidental) bug mitigation feature rather than a bug.

Kevin

pgp3xTH58tWYq.pgp
Description: PGP signature