Re: [PATCH v2] ubi: gluebi: Fix NULL pointer dereference caused by ftl notifier

From: Richard Weinberger
Date: Thu Oct 19 2023 - 16:27:37 EST


----- Ursprüngliche Mail -----
> Von: "ZhaoLong Wang" <wangzhaolong1@xxxxxxxxxx>
> An: "richard" <richard@xxxxxx>, "Miquel Raynal" <miquel.raynal@xxxxxxxxxxx>, "Vignesh Raghavendra" <vigneshr@xxxxxx>,
> dpervushin@xxxxxxxxxxxxxxxxx, "Artem Bityutskiy" <Artem.Bityutskiy@xxxxxxxxx>
> CC: "linux-mtd" <linux-mtd@xxxxxxxxxxxxxxxxxxx>, "linux-kernel" <linux-kernel@xxxxxxxxxxxxxxx>, "chengzhihao1"
> <chengzhihao1@xxxxxxxxxx>, "ZhaoLong Wang" <wangzhaolong1@xxxxxxxxxx>, "yi zhang" <yi.zhang@xxxxxxxxxx>, "yangerkun"
> <yangerkun@xxxxxxxxxx>
> Gesendet: Mittwoch, 18. Oktober 2023 14:16:18
> Betreff: [PATCH v2] ubi: gluebi: Fix NULL pointer dereference caused by ftl notifier

> If both flt.ko and gluebi.ko are loaded, the notiier of ftl
> triggers NULL pointer dereference when trying to access
> ‘gluebi->desc’ in gluebi_read().
>
> ubi_gluebi_init
> ubi_register_volume_notifier
> ubi_enumerate_volumes
> ubi_notify_all
> gluebi_notify nb->notifier_call()
> gluebi_create
> mtd_device_register
> mtd_device_parse_register
> add_mtd_device
> blktrans_notify_add not->add()
> ftl_add_mtd tr->add_mtd()
> scan_header
> mtd_read
> mtd_read
> mtd_read_oob
> gluebi_read mtd->read()
> gluebi->desc - NULL
>
> Detailed reproduction information available at the link[1],
>
> In the normal case, obtain gluebi->desc in the gluebi_get_device(),
> and accesses gluebi->desc in the gluebi_read(). However,
> gluebi_get_device() is not executed in advance in the
> ftl_add_mtd() process, which leads to NULL pointer dereference.
>
> The value of gluebi->desc may also be a negative error code, which
> triggers the page fault error.
>
> This patch has the following modifications:
>
> 1. Do not assign gluebi->desc to the error code. Use the NULL instead.
>
> 2. Always check the validity of gluebi->desc in gluebi_read() If the
> gluebi->desc is NULL, try to get MTD device.
>
> Such a modification currently works because the mutex "mtd_table_mutex"
> is held on all necessary paths, including the ftl_add_mtd() call path,
> open and close paths. Therefore, many race condition can be avoided.

I see the problem, but I'm not really satisfied by the solution.
Adding this hack to gluebi_read() is not nice at all.

Is there a strong reason why have to defer ubi_open_volume() to
gluebi_get_device()?

Miquel, what do you think?

Thanks,
//richard