RE: [PATCH 1/2] mtd: use refcount to prevent corruption

[Usyskin, Alexander]

Hi Miquel,
> 
> Hi Alexander,
> 
> alexander.usyskin@xxxxxxxxx wrote on Mon, 24 Jul 2023 11:43:59 +0000:
> 
> > > > > > With this patch applied, when I load up the module, I get the same 3
> > > > > > devices:
> > > > > > /dev/mtd0
> > > > > > /dev/mtd0ro
> > > > > > /dev/mtdblock0
> > > > > >
> > > > > > Upon removal, the below 2 devices still hang around:
> > > > > > /dev/mtd0
> > > > > > /dev/mtd0ro
> > > > >
> > > > Our use-case do not produce mtdblock, maybe there are some
> imbalances
> > > of get/put?
> > > > I have somewhere version with pr_debug after every kref_get/put. That
> may
> > > help to catch where
> > > > it missed, I hope.
> > >
> > > I believe mtdblock is the good citizen here. Just disable
> > > CONFIG_MTD_BLOCK from your configuration and you will likely observe
> > > the same issue, just a bit narrowed, perhaps. Indeed, if you manage to
> > > follow all the get/put calls it can help to find an imbalance.
> > >
> > > Thanks,
> > > Miquèl
> >
> > Miquel, do you have CONFIG_MTD_PARTITIONED_MASTER set in your
> config?
> 
> Not sure I get your question. You can enable or disable it, it should
> work in both cases (yet, the handling is of course a bit different as
> the top level device will be retained/not retained).
> 
> Thanks,
> Miquèl

I'm trying to understand why I can't reproduce the problem in my scenario.
I found an important difference in upstreamed patch and internal version:
The IS_ENABLED(CONFIG_MTD_PARTITIONED_MASTER) check around kref_get/put does not exists in the internal tree.
The code before my patch do not have such check, so I tend to assume that this check should be removed.
If you reproduce happens with CONFIG_MTD_PARTITIONED_MASTER disabled that may explain problems that you see.

-- 
Thanks,
Sasha