Re: [PATCH] ALSA: hda: Use correct start/count for sysfs init

[Takashi Iwai]

On Wed, 26 Jun 2019 22:34:28 +0200,
Evan Green wrote:
> 
> On Wed, Jun 26, 2019 at 1:27 AM Takashi Iwai <tiwai@xxxxxxx> wrote:
> >
> > On Tue, 25 Jun 2019 23:54:18 +0200,
> > Evan Green wrote:
> > >
> > > The normal flow through the widget sysfs codepath is that
> > > snd_hdac_refresh_widgets() is called once without the sysfs bool set
> > > to set up codec->num_nodes and friends, then another time with the
> > > bool set to actually allocate all the sysfs widgets. However, during
> > > the first time allocation, hda_widget_sysfs_reinit() ignores the new
> > > num_nodes passed in via parameter and just calls hda_widget_sysfs_init(),
> > > using whatever was in codec->num_nodes before the update. This is not
> > > correct in cases where num_nodes changes. Here's an example:
> > >
> > > Sometime earlier:
> > > snd_hdac_refresh_widgets(hdac, false)
> > >   sets codec->num_nodes to 2, widgets is still not allocated
> > >
> > > Now:
> > > snd_hdac_refresh_widgets(hdac, true)
> > >   hda_widget_sysfs_reinit(num_nodes=7)
> > >     hda_widget_sysfs_init()
> > >       widget_tree_create()
> > >         alloc(codec->num_nodes) // this is still 2
> > >   codec->num_nodes = 7
> > >
> > > Pass num_nodes and start_nid down into widget_tree_create() so that
> > > the right number of nodes are allocated in all cases.
> > >
> > > Signed-off-by: Evan Green <evgreen@xxxxxxxxxxxx>
> >
> > Thanks for the patch.  That's indeed a problem, but I guess a simpler
> > approach is just to return if sysfs didn't exist.  If the sysfs
> > entries aren't present at the second call with sysfs=true, it implies
> > that the codec object will be exposed anyway later, and the sysfs will
> > be created there.  So, something like below would work instead?
> 
> Hi Takashi,
> Thanks for taking a look. I'm not sure you'd want to do that because
> then you end up returning from sysfs_reinit without having allocated
> any of the sysfs widgets. You'd be relying on the implicit behavior
> that another call to init is coming later (despite having updated
> num_nodes and start node), which is difficult to follow and easy to
> break. In my opinion the slight bit of extra diffs is well worth the
> clarity of having widget_tree_create always allocate the correct
> start/count.

Well, skipping is the right behavior, actually.  The whole need of the
refresh function is just to refresh the widget list, and the current
behavior to create a sysfs is rather superfluous.  This action has
never been used, so better to get removed for avoiding misuse.

> Actually, in looking at the widget lock patch, I don't think it's
> sufficient either. It adds a lock around sysfs_reinit, but the setting
> of codec->num_nodes and codec->start_nid is unprotected by the lock.
> So you could have the two threads politely serialize through
> sysfs_reinit, but then get reordered before setting codec->num_nodes,
> landing you with an array whose length doesn't match num_nodes.

The usage of snd_hdac_refresh_widgets() is supposed to be done only at
the codec probe phase, hence there is no lock done in the core code;
IOW, any concurrent access has to be protected in the caller side in
general.

Have you actually seen such concurrent accesses?  If yes, that's a
problem in the usage.


thanks,

Takashi

> 
> Let me craft up an additional patch to fix the locking.
> -Evan
> 
> >
> >
> > thanks,
> >
> > Takashi
> >
> > --- a/sound/hda/hdac_sysfs.c
> > +++ b/sound/hda/hdac_sysfs.c
> > @@ -428,7 +428,7 @@ int hda_widget_sysfs_reinit(struct hdac_device *codec,
> >         int i;
> >
> >         if (!codec->widgets)
> > -               return hda_widget_sysfs_init(codec);
> > +               return 0;
> >
> >         tree = kmemdup(codec->widgets, sizeof(*tree), GFP_KERNEL);
> >         if (!tree)
>