Re: [PATCH v2] mtd: Fix mtd not the same name not registered if nvmem

From: Boris Brezillon
Date: Thu Apr 30 2020 - 08:52:13 EST


On Thu, 30 Apr 2020 14:26:51 +0200
Ricardo Ribalda Delgado <ribalda@xxxxxxxxxx> wrote:

> Hi
>
> On Mon, Apr 27, 2020 at 9:10 PM Boris Brezillon
> <boris.brezillon@xxxxxxxxxxxxx> wrote:
> >
> > On Mon, 27 Apr 2020 16:49:22 +0200
> > Miquel Raynal <miquel.raynal@xxxxxxxxxxx> wrote:
> >
> > > Hi Boris,
> > >
> > > Boris Brezillon <boris.brezillon@xxxxxxxxxxxxx> wrote on Mon, 27 Apr
> > > 2020 16:37:11 +0200:
> > >
> > > > On Mon, 27 Apr 2020 16:22:22 +0200
> > > > Miquel Raynal <miquel.raynal@xxxxxxxxxxx> wrote:
> > > >
> > > > > Hi Ricardo,
> > > > >
> > > > > Ricardo Ribalda Delgado <ribalda@xxxxxxxxxx> wrote on Tue, 14 Apr 2020
> > > > > 15:47:23 +0200:
> > > > >
> > > > > > Ping?
> > > > > >
> > > > > > On Thu, Apr 2, 2020 at 8:59 AM Ricardo Ribalda Delgado
> > > > > > <ribalda@xxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > When the nvmem framework is enabled, a nvmem device is created per mtd
> > > > > > > device/partition.
> > > > > > >
> > > > > > > It is not uncommon that a device can have multiple mtd devices with
> > > > > > > partitions that have the same name. Eg, when there DT overlay is allowed
> > > > > > > and the same device with mtd is attached twice.
> > > > > > >
> > > > > > > Under that circumstances, the mtd fails to register due to a name
> > > > > > > duplication on the nvmem framework.
> > > > > > >
> > > > > > > With this patch we add a _1, _2, _X to the subsequent names if there is
> > > > > > > a collition, and throw a warning, instead of not starting the mtd
> > > > > > > device.
> > > > > > >
> > > > > > > [ 8.948991] sysfs: cannot create duplicate filename '/bus/nvmem/devices/Production Data'
> > > > > > > [ 8.948992] CPU: 7 PID: 246 Comm: systemd-udevd Not tainted 5.5.0-qtec-standard #13
> > > > > > > [ 8.948993] Hardware name: AMD Dibbler/Dibbler, BIOS 05.22.04.0019 10/26/2019
> > > > > > > [ 8.948994] Call Trace:
> > > > > > > [ 8.948996] dump_stack+0x50/0x70
> > > > > > > [ 8.948998] sysfs_warn_dup.cold+0x17/0x2d
> > > > > > > [ 8.949000] sysfs_do_create_link_sd.isra.0+0xc2/0xd0
> > > > > > > [ 8.949002] bus_add_device+0x74/0x140
> > > > > > > [ 8.949004] device_add+0x34b/0x850
> > > > > > > [ 8.949006] nvmem_register.part.0+0x1bf/0x640
> > > > > > > ...
> > > > > > > [ 8.948926] mtd mtd8: Failed to register NVMEM device
> > > > > > >
> > > > > > > Signed-off-by: Ricardo Ribalda Delgado <ribalda@xxxxxxxxxx>
> > > > >
> > > > > Thanks for proposing this change. Indeed we are aware of the problem
> > > > > and the best solution that we could come up with was to create an
> > > > > additional "unique_name" field to the mtd_info structure. This new
> > > > > field would have the form:
> > > > >
> > > > > [<parent-unique-name><separator>]<mtd-name>
> > > > >
> > > > > The separator might be '~' (but I am completely open on that), and that
> > > > > would give for instance:
> > > > >
> > > > > my-controller~my-device~my-part~mysub-part
> > > >
> > > > I'd prefer something slightly more standard for the separator, like '/',
> > > > which is what we usually use when we want to represent a path in a tree.
> > > > I do agree on the general approach though.
> > >
> > > I am not sure / is a valid separator here we would use this
> > > name to create a sysfs entry. Would it work?
> >
> > Hm, you're probably right, I didn't check how the name was used by
> > nvmem. I also see that partition names can contain spaces, which is
> > not that great. So, if we want to use mtd->unique_name we should
> > probably sanitize it before passing it to nvmem. All this makes me
> > reconsider your initial proposal: use 'mtdX' as the nvmem name. It's
> > unique and it's simple enough to not require this extra sanitization
> > pass, on the other hand, guessing the nvmem partition based on such an
> > opaque name is not simple, not to mention that the mtd device name can
> > change depending on the probe order.
> >
> > I also wonder if creating nvmem devs unconditionally for all mtd
> > devices is a good idea. Sounds like we should only do that if there's an
> > explicit request to have one partition exposed as an nvmem.
> >
> > Note that, no matter what we decide about nvmem, I think having unique
> > names at the MTD level is a good thing.
>
> I have no preference one way or another.
>
> The issue is that our current master leads to mtds not working fine
> and making the system unusable. Whatever we decide it must be fast and
> the patch backported.
>
> My original patch follows the same logic as led does where there is a
> duplicated name. I can send a separate patch that uses mtdX and then
> you decide what to pick. Or we can go with a third way, but it needs
> to be soon ;).

Please send a patch using dev_name(&mtd->dev), and let's hope it
doesn't break someone else use case.