[PATCH v4] mtd: parsers: ofpart: add workaround for #size-cells 0

[Francesco Dolcini]

Il 3 febbraio 2023 16:12:02 CET, Miquel Raynal <miquel.raynal at bootlin.com> ha scritto:
>Hi Francesco,
>
>francesco at dolcini.it wrote on Thu, 2 Feb 2023 12:33:34 +0100:
>
>> On Thu, Jan 26, 2023 at 10:12:04AM +0100, Miquel Raynal wrote:
>> > gregkh at linuxfoundation.org wrote on Thu, 26 Jan 2023 10:01:02 +0100:
>> >   
>> > > On Wed, Jan 25, 2023 at 10:06:57PM +0100, Francesco Dolcini wrote:  
>> > > > Hello Miquel, Greg and all
>> > > > 
>> > > > On Tue, Jan 24, 2023 at 04:38:59PM +0100, Greg Kroah-Hartman wrote:    
>> > > > > On Tue, Jan 24, 2023 at 11:44:44AM +0100, Francesco Dolcini wrote:    
>> > > > > > From: Francesco Dolcini <francesco.dolcini at toradex.com>
>> > > > > > 
>> > > > > > Add a mechanism to handle the case in which partitions are present as
>> > > > > > direct child of the nand controller node and #size-cells is set to <0>.
>> > > > > > 
>> > > > > > This could happen if the nand-controller node in the DTS is supposed to
>> > > > > > have #size-cells set to 0, but for some historical reason/bug it was set
>> > > > > > to 1 in the past, and the firmware (e.g. U-Boot) is adding the partition
>> > > > > > as direct children of the nand-controller defaulting to #size-cells
>> > > > > > being to 1.
>> > > > > > 
>> > > > > > This prevents a real boot failure on colibri-imx7 that happened during v6.1
>> > > > > > development cycles.
>> > > > > > 
>> > > > > > Link: https://lore.kernel.org/all/Y4dgBTGNWpM6SQXI@francesco-nb.int.toradex.com/
>> > > > > > Link: https://lore.kernel.org/all/20221202071900.1143950-1-francesco@dolcini.it/
>> > > > > > Signed-off-by: Francesco Dolcini <francesco.dolcini at toradex.com>
>> > > > > > Reviewed-by: Greg Kroah-Hartman <gregkh at linuxfoundation.org>
>> > > > > > ---
>> > > > > > I do not expect this patch to be backported to stable, however I would expect
>> > > > > > that we do not backport nand-controller dts cleanups neither.
>> > > > > > 
>> > > > > > v4:
>> > > > > >  fixed wrong English spelling in the comment
>> > > > > > 
>> > > > > > v3:
>> > > > > >  minor formatting change, removed not needed new-line and space. 
>> > > > > > 
>> > > > > > v2:
>> > > > > >  fixup size-cells only when partitions are direct children of the nand-controller
>> > > > > >  completely revised commit message, comments and warning print
>> > > > > >  use pr_warn instead of pr_warn_once
>> > > > > >  added Reviewed-by Greg
>> > > > > >  removed cc:stable@ and fixes tag, since the problematic commit was reverted
>> > > > > > ---
>> > > > > >  drivers/mtd/parsers/ofpart_core.c | 19 +++++++++++++++++++
>> > > > > >  1 file changed, 19 insertions(+)
>> > > > > > 
>> > > > > > diff --git a/drivers/mtd/parsers/ofpart_core.c b/drivers/mtd/parsers/ofpart_core.c
>> > > > > > index 192190c42fc8..e7b8e9d0a910 100644
>> > > > > > --- a/drivers/mtd/parsers/ofpart_core.c
>> > > > > > +++ b/drivers/mtd/parsers/ofpart_core.c
>> > > > > > @@ -122,6 +122,25 @@ static int parse_fixed_partitions(struct mtd_info *master,
>> > > > > >  
>> > > > > >  		a_cells = of_n_addr_cells(pp);
>> > > > > >  		s_cells = of_n_size_cells(pp);
>> > > > > > +		if (!dedicated && s_cells == 0) {
>> > > > > > +			/*
>> > > > > > +			 * This is a ugly workaround to not create
>> > > > > > +			 * regression on devices that are still creating
>> > > > > > +			 * partitions as direct children of the nand controller.
>> > > > > > +			 * This can happen in case the nand controller node has
>> > > > > > +			 * #size-cells equal to 0 and the firmware (e.g.
>> > > > > > +			 * U-Boot) just add the partitions there assuming
>> > > > > > +			 * 32-bit addressing.
>> > > > > > +			 *
>> > > > > > +			 * If you get this warning your firmware and/or DTS
>> > > > > > +			 * should be really fixed.
>> > > > > > +			 *
>> > > > > > +			 * This is working only for devices smaller than 4GiB.
>> > > > > > +			 */
>> > > > > > +			pr_warn("%s: ofpart partition %pOF (%pOF) #size-cells is wrongly set to <0>, assuming <1> for parsing partitions.\n",
>> > > > > > +				master->name, pp, mtd_node);    
>> > > > > 
>> > > > > This is a driver, always use dev_*() calls, not pr_*() calls so that we
>> > > > > know what is being referred to exactly.    
>> > > > 
>> > > > Is this reasonable here? Where can I get the struct device?    
>> > > 
>> > > Walk back up the call chain, there has to be a device somewhere
>> > > controlling this, right?
>> > >   
>> > > > In general this file uses only pr_* debug API and messages are about OF
>> > > > nodes/properties, not about a device.    
>> > > 
>> > > OF nodes and properties are part of a device's properties :)  
>> > 
>> > Yes but the warning comes from a wrong DT description, hence it felt
>> > better suited to warn against the node name which is easily identifiable
>> > in a text file and must be fixed rather than the device which is a pure
>> > software component.
>> > 
>> > Anyway, Francesco, please show us the resultant line and if it feels
>> > meaningful enough we'll take the dev_warn approach.  
>> 
>> So, I tried, but I guess I failed.
>> 
>> Both
>> 
>>   dev_warn(&mtd_get_master(master)->dev, ...);
>> 
>> and
>> 
>>   dev_warn(&master->dev, ...);
>> 
>> are NULL.
>
>mtd->dev (in raw NAND) is populated by the controller drivers, so the
>master mtd device is pointing to the bus "struct device" in its
>dev.parent field. This happens at the end of the probe of the
>controller, after setting the dev entry, so we expect the name of the
>controller to appear.
>
>> (null): gpmi-nand: ofpart partition /soc/nand-controller at 33002000/partition at 0 (/soc/nand-controller at 33002000) #size-cells is wrongly set to <0>, assuming <1> for parsing partitions.
>
>Second field looks right, first field does not (bus or class id?) I
>have no idea why it has not been populated at this point (end of the
>controller probe). But it's not a big deal, at least we have the device
>name, so it's ok for me.

If I understand correctly you are fine with the current patch v4 that just print (on colibri-imx7)


gpmi-nand: ofpart partition /soc/nand-controller at 33002000/partition at 0 (/soc/nand-controller at 33002000) #size-cells is wrongly set to <0>, assuming <1> for parsing partitions.

Correct? Do you want a v5 with the reviewed-by greg removed or you remove it yourself?

Francesco