Re: [PATCH v5 3/3] mtd: core: introduce of support for dynamic partitions

[han.xu]

On 22/10/17 11:52PM, Rafał Miłecki wrote:
> On 17.10.2022 21:59, han.xu wrote:
> > On 22/06/06 05:14PM, Ansuel Smith wrote:
> > > We have many parser that register mtd partitions at runtime. One example
> > > is the cmdlinepart or the smem-part parser where the compatible is defined
> > > in the dts and the partitions gets detected and registered by the
> > > parser. This is problematic for the NVMEM subsystem that requires an OF node
> > > to detect NVMEM cells.
> > > 
> > > To fix this problem, introduce an additional logic that will try to
> > > assign an OF node to the MTD if declared.
> > > 
> > > On MTD addition, it will be checked if the MTD has an OF node and if
> > > not declared will check if a partition with the same label is
> > > declared in DTS. If an exact match is found, the partition dynamically
> > > allocated by the parser will have a connected OF node.
> > > 
> > > The NVMEM subsystem will detect the OF node and register any NVMEM cells
> > > declared statically in the DTS.
> > > 
> > > Signed-off-by: Ansuel Smith <ansuelsmth@xxxxxxxxx>
> > > ---
> > >   drivers/mtd/mtdcore.c | 49 +++++++++++++++++++++++++++++++++++++++++++
> > >   1 file changed, 49 insertions(+)
> > > 
> > > diff --git a/drivers/mtd/mtdcore.c b/drivers/mtd/mtdcore.c
> > > index 7731796024e0..807194efb580 100644
> > > --- a/drivers/mtd/mtdcore.c
> > > +++ b/drivers/mtd/mtdcore.c
> > > @@ -546,6 +546,54 @@ static int mtd_nvmem_add(struct mtd_info *mtd)
> > >   	return 0;
> > >   }
> > > +static void mtd_check_of_node(struct mtd_info *mtd)
> > > +{
> > > +	struct device_node *partitions, *parent_dn, *mtd_dn = NULL;
> > > +	struct mtd_info *parent;
> > > +	const char *mtd_name;
> > > +	bool found = false;
> > > +	int plen;
> > > +
> > > +	/* Check if MTD already has a device node */
> > > +	if (dev_of_node(&mtd->dev))
> > > +		return;
> > > +
> > > +	/* Check if a partitions node exist */
> > > +	parent = mtd->parent;
> > > +	parent_dn = dev_of_node(&parent->dev);
> > > +	if (!parent_dn)
> > > +		return;
> > > +
> > > +	partitions = of_get_child_by_name(parent_dn, "partitions");
> > > +	if (!partitions)
> > > +		goto exit_parent;
> > > +
> > > +	/* Search if a partition is defined with the same name */
> > > +	for_each_child_of_node(partitions, mtd_dn) {
> > > +		/* Skip partition with no label */
> > > +		mtd_name = of_get_property(mtd_dn, "label", &plen);
> > > +		if (!mtd_name)
> > > +			continue;
> > > +
> > > +		if (!strncmp(mtd->name, mtd_name, plen)) {
> > > +			found = true;
> > > +			break;
> > > +		}
> > > +	}
> > > +
> > > +	if (!found)
> > > +		goto exit_partitions;
> > > +
> > > +	/* Set of_node only for nvmem */
> > > +	if (of_device_is_compatible(mtd_dn, "nvmem-cells"))
> > > +		mtd_set_of_node(mtd, mtd_dn);
> > > +
> > > +exit_partitions:
> > > +	of_node_put(partitions);
> > > +exit_parent:
> > > +	of_node_put(parent_dn);
> > > +}
> > > +
> > >   /**
> > >    *	add_mtd_device - register an MTD device
> > >    *	@mtd: pointer to new MTD device info structure
> > > @@ -651,6 +699,7 @@ int add_mtd_device(struct mtd_info *mtd)
> > >   	mtd->dev.devt = MTD_DEVT(i);
> > >   	dev_set_name(&mtd->dev, "mtd%d", i);
> > >   	dev_set_drvdata(&mtd->dev, mtd);
> > > +	mtd_check_of_node(mtd);
> > >   	of_node_get(mtd_get_of_node(mtd));
> > >   	error = device_register(&mtd->dev);
> > >   	if (error)
> > 
> > NXP GPMI NAND controller with 6 cmdline partitions meets refcount underflow
> > with this patch. The of_node_put(parent_dn) doesn't work with cmdline parser and
> > causes the issue.
> 
> Can you try:
> 
> diff --git a/drivers/mtd/mtdcore.c b/drivers/mtd/mtdcore.c
> index 18aa54460d36..0b4ca0aa4132 100644
> --- a/drivers/mtd/mtdcore.c
> +++ b/drivers/mtd/mtdcore.c
> @@ -562,7 +562,7 @@ static void mtd_check_of_node(struct mtd_info *mtd)
>  	if (!mtd_is_partition(mtd))
>  		return;
>  	parent = mtd->parent;
> -	parent_dn = dev_of_node(&parent->dev);
> +	parent_dn = of_node_get(dev_of_node(&parent->dev));
>  	if (!parent_dn)
>  		return;

Yes, with of_node_get() the refcount issue gone.

> 
>