[PATCH 1/2] cpukit/jffs2: Set extracted nodes off-chain

[Kinsey Moore]

On Thu, Sep 21, 2023 at 10:23 AM Sebastian Huber <
sebastian.huber at embedded-brains.de> wrote:

> On 21.09.23 17:06, Kinsey Moore wrote:
> > On Thu, Sep 21, 2023 at 10:01 AM Sebastian Huber
> > <sebastian.huber at embedded-brains.de
> > <mailto:sebastian.huber at embedded-brains.de>> wrote:
> >
> >     On 21.09.23 16:52, Kinsey Moore wrote:
> >      > On Thu, Sep 21, 2023 at 9:47 AM Sebastian Huber
> >      > <sebastian.huber at embedded-brains.de
> >     <mailto:sebastian.huber at embedded-brains.de>
> >      > <mailto:sebastian.huber at embedded-brains.de
> >     <mailto:sebastian.huber at embedded-brains.de>>> wrote:
> >      >
> >      >     On 20.09.23 20:35, Kinsey Moore wrote:
> >      >     [...]
> >      >      > @@ -1306,8 +1307,22 @@ static void
> process_delayed_work(void)
> >      >      >       while (!rtems_chain_is_tail(&process_work_chain,
> >     node)) {
> >      >      >               work = (struct delayed_work*) node;
> >      >      >               rtems_chain_node* next_node =
> >     rtems_chain_next(node);
> >      >      > +
> >      >      > +             /*
> >      >      > +              * Don't leave extracted node exposed to
> other
> >      >     operations
> >      >      > +              * under RTEMS_DEBUG
> >      >      > +              */
> >      >      > +#ifdef RTEMS_DEBUG
> >      >      > +             mutex_lock(&delayed_work_mutex);
> >      >      > +#endif
> >      >      >               rtems_chain_extract(node);
> >      >      > +#ifdef RTEMS_DEBUG
> >      >      > +             node->next = node;
> >      >      > +             mutex_unlock(&delayed_work_mutex);
> >      >      > +#endif
> >      >      > +
> >      >      >               work->callback(&work->work);
> >      >      > +             rtems_chain_set_off_chain(node);
> >      >      >               node = next_node;
> >      >      >       }
> >      >      >   }
> >      >
> >      >     This change makes no sense to me. The code should work
> >     regardless if
> >      >     RTEMS_DEBUG is defined or not.
> >      >
> >      >
> >      > RTEMS_DEBUG introduces a behavioral change in
> >     rtems_chain_extract() such
> >      > that extracted nodes are automatically set off-chain. When
> >     RTEMS_DEBUG
> >      > is not set, node->next is left untouched. This has to be managed
> >     because
> >      > this code needs the node to remain on-chain so that it is not
> >     re-queued
> >      > during the operation.
> >
> >     Yes, but while a node is on a chain you must not call
> >     rtems_chain_set_off_chain(). If you want to use the off-chain state,
> >     then you have to use this:
> >
> >     rtems_chain_extract(node);
> >     rtems_chain_set_off_chain(node);
> >
> >     or
> >
> >     rtems_chain_extract_unprotected(node);
> >     rtems_chain_set_off_chain(node);
> >
> >     The automatic set off-chain in RTEMS_DEBUG is just to ensure that
> basic
> >     chain operations are used in the right state.
> >
> >
> > Yes, there is no behavior here where rtems_chain_set_off_chain() is
> > being called on a node that is still in a chain. This section of code is
> > entirely managing the side-effect of RTEMS_DEBUG that sets the node in
> > the off-chain state post-extraction. In this case, that side-effect is
> > undesirable and so a lock is in place to prevent that temporary
> > side-effect from leaking to other parts of the system since all other
> > off-chain checks are behind the same lock.
>
> Ok, so you want to delay the set-off state visibility in case
> RTEMS_DEBUG is defined. This is the first place in the code base which
> needs this behaviour.
>

Would it be more acceptable to always lock it and ensure that it's
recognized as on-chain instead of checking for RTEMS_DEBUG?

>
> We basically don't use internal locking in JFFS2 in RTEMS. Why can't you
> use the global file system lock of the JFFS2 instance to work carry out
> this delayed work?
>

The JFFS2 instance to which the delayed work belongs is a detail internal
to the delayed work itself. Reaching that information is possible, but
would be fragile. The other option I had in mind was to create one delayed
work thread per JFFS2 FS instance so that the delayed work processor can
know reliably which JFFS2 instance it is working on, but that seemed
wasteful.

Kinsey
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.rtems.org/pipermail/devel/attachments/20230921/a691c0a9/attachment-0001.htm>