Re: [PATCH v3 0/2] mtd: spi-nor: fix use-after-free on unbind with open handles

From: NG, TZE YEE

Date: Thu Nov 19 2026 - 21:21:55 EST


Hi Michael,

Thanks for taking a look.

On 18/11/2026 5:36 pm, Michael Walle wrote:
Hi,

On Wed Nov 18, 2026 at 9:07 AM CET, Tze Yee Ng wrote:
Gentle ping on this patch. Any chance for a review when you have time?

I had a quick glance, and honestly, I really don't like all the
reference and resource handling splattered all over the place.

I agree it is spread across spi-nor. It comes from one thing, which the MTD core still uses the driver object after unbind, so spi-nor has to keep that object and its data alive and fence its own operations.

Also do we really need the low level kref_ handling?

A core-owned lifetime would avoid it. v2 tried that, with a release hook
and a removed flag in the core. Review found problems in that design:
the master can be freed inside __put_mtd_device(), the driver object
leaked with CONFIG_MTD_PARTITIONED_MASTER off, and a lockless removed
check raced with remove(). The kref in spi-nor is what works without a
larger change in the core; patch 1 is the only core change.

And what about spi-nand and all the other MTD subsystems? They have the same
problem, no?

-michael

Yes. spinand allocates its object with devm_kzalloc() and frees it after
mtd_device_unregister(), and many other drivers that embed mtd_info do
the same. A generic fix belongs in the core, but it is a much larger
change that touches every MTD driver's lifetime rules.

I would prefer to keep this series as a spi-nor-only fix, since it solves the problem for spi-nor today. James Hilliard's raw NAND series [1] adds a generic opt-in lifetime to the core (mtd_set_release()). I would be happy to look at moving spi-nor onto it if it lands, but maybe in another series. Are you okay with keeping this series as is?

Thanks,
Tze Yee

[1] https://lore.kernel.org/all/20261115-submit-rawnand-retained-handle-lifetime-v1-v5-0-cbc49c5c5e74@xxxxxxxxx/