From: Boris Brezillon <[email protected]>
The BBT scan logic use the MTD helpers before the MTD layer had a
chance to initialize the device, and that leads to issues when
accessing the uninitialized suspend lock. Let's temporarily set the
suspend/resume hooks to NULL to skip the lock acquire/release step.
Fixes: 013e6292aaf5 ("mtd: rawnand: Simplify the locking")
Tested-by: Sean Nyekjaer <[email protected]>
Signed-off-by: Boris Brezillon <[email protected]>
Signed-off-by: Sean Nyekjaer <[email protected]>
---
drivers/mtd/nand/raw/nand_bbt.c | 28 +++++++++++++++++++++++++++-
1 file changed, 27 insertions(+), 1 deletion(-)
diff --git a/drivers/mtd/nand/raw/nand_bbt.c b/drivers/mtd/nand/raw/nand_bbt.c
index b7ad030225f8..93d385703469 100644
--- a/drivers/mtd/nand/raw/nand_bbt.c
+++ b/drivers/mtd/nand/raw/nand_bbt.c
@@ -1397,8 +1397,28 @@ static int nand_create_badblock_pattern(struct nand_chip *this)
*/
int nand_create_bbt(struct nand_chip *this)
{
+ struct mtd_info *mtd = nand_to_mtd(this);
+ int (*suspend) (struct mtd_info *) = mtd->_suspend;
+ void (*resume) (struct mtd_info *) = mtd->_resume;
int ret;
+ /*
+ * The BBT scan logic use the MTD helpers before the MTD layer had a
+ * chance to initialize the device, and that leads to issues when
+ * accessing the uninitialized suspend lock. Let's temporarily set the
+ * suspend/resume hooks to NULL to skip the lock acquire/release step.
+ *
+ * FIXME: This is an ugly hack, so please don't copy this pattern to
+ * other MTD implementations. The proper fix would be to implement a
+ * generic BBT scan logic at the NAND level that's not using any of the
+ * MTD helpers to access pages. We also might consider doing a two
+ * step initialization at the MTD level (mtd_device_init() +
+ * mtd_device_register()) so some of the fields are initialized
+ * early.
+ */
+ mtd->_suspend = NULL;
+ mtd->_resume = NULL;
+
/* Is a flash based bad block table requested? */
if (this->bbt_options & NAND_BBT_USE_FLASH) {
/* Use the default pattern descriptors */
@@ -1422,7 +1442,13 @@ int nand_create_bbt(struct nand_chip *this)
return ret;
}
- return nand_scan_bbt(this, this->badblock_pattern);
+ ret = nand_scan_bbt(this, this->badblock_pattern);
+
+ /* Restore the suspend/resume hooks. */
+ mtd->_suspend = suspend;
+ mtd->_resume = resume;
+
+ return ret;
}
EXPORT_SYMBOL(nand_create_bbt);
--
2.33.0
On Tue, 26 Oct 2021 07:55:48 +0200
Sean Nyekjaer <[email protected]> wrote:
> From: Boris Brezillon <[email protected]>
>
> The BBT scan logic use the MTD helpers before the MTD layer had a
> chance to initialize the device, and that leads to issues when
> accessing the uninitialized suspend lock. Let's temporarily set the
> suspend/resume hooks to NULL to skip the lock acquire/release step.
>
> Fixes: 013e6292aaf5 ("mtd: rawnand: Simplify the locking")
I think I already mentioned this Fixes tag should not be there.
> Tested-by: Sean Nyekjaer <[email protected]>
> Signed-off-by: Boris Brezillon <[email protected]>
> Signed-off-by: Sean Nyekjaer <[email protected]>
> ---
> drivers/mtd/nand/raw/nand_bbt.c | 28 +++++++++++++++++++++++++++-
> 1 file changed, 27 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/mtd/nand/raw/nand_bbt.c b/drivers/mtd/nand/raw/nand_bbt.c
> index b7ad030225f8..93d385703469 100644
> --- a/drivers/mtd/nand/raw/nand_bbt.c
> +++ b/drivers/mtd/nand/raw/nand_bbt.c
> @@ -1397,8 +1397,28 @@ static int nand_create_badblock_pattern(struct nand_chip *this)
> */
> int nand_create_bbt(struct nand_chip *this)
> {
> + struct mtd_info *mtd = nand_to_mtd(this);
> + int (*suspend) (struct mtd_info *) = mtd->_suspend;
> + void (*resume) (struct mtd_info *) = mtd->_resume;
> int ret;
>
> + /*
> + * The BBT scan logic use the MTD helpers before the MTD layer had a
> + * chance to initialize the device, and that leads to issues when
> + * accessing the uninitialized suspend lock. Let's temporarily set the
> + * suspend/resume hooks to NULL to skip the lock acquire/release step.
> + *
> + * FIXME: This is an ugly hack, so please don't copy this pattern to
> + * other MTD implementations. The proper fix would be to implement a
> + * generic BBT scan logic at the NAND level that's not using any of the
> + * MTD helpers to access pages. We also might consider doing a two
> + * step initialization at the MTD level (mtd_device_init() +
> + * mtd_device_register()) so some of the fields are initialized
> + * early.
> + */
> + mtd->_suspend = NULL;
> + mtd->_resume = NULL;
> +
> /* Is a flash based bad block table requested? */
> if (this->bbt_options & NAND_BBT_USE_FLASH) {
> /* Use the default pattern descriptors */
> @@ -1422,7 +1442,13 @@ int nand_create_bbt(struct nand_chip *this)
> return ret;
> }
>
> - return nand_scan_bbt(this, this->badblock_pattern);
> + ret = nand_scan_bbt(this, this->badblock_pattern);
> +
> + /* Restore the suspend/resume hooks. */
> + mtd->_suspend = suspend;
> + mtd->_resume = resume;
> +
> + return ret;
> }
> EXPORT_SYMBOL(nand_create_bbt);
>
On Mon, Nov 01, 2021 at 08:38:24AM +0100, Boris Brezillon wrote:
> On Tue, 26 Oct 2021 07:55:48 +0200
> Sean Nyekjaer <[email protected]> wrote:
>
> > From: Boris Brezillon <[email protected]>
> >
> > The BBT scan logic use the MTD helpers before the MTD layer had a
> > chance to initialize the device, and that leads to issues when
> > accessing the uninitialized suspend lock. Let's temporarily set the
> > suspend/resume hooks to NULL to skip the lock acquire/release step.
> >
> > Fixes: 013e6292aaf5 ("mtd: rawnand: Simplify the locking")
>
> I think I already mentioned this Fixes tag should not be there.
>
No, I didn't recall that, but nevermind :)
Hmm when ('mtd: core: protect access to MTD devices while in suspend')
is backported we want to live with the use before init of the rwsem?
/Sean
On Mon, 1 Nov 2021 09:46:23 +0100
Sean Nyekjaer <[email protected]> wrote:
> On Mon, Nov 01, 2021 at 08:38:24AM +0100, Boris Brezillon wrote:
> > On Tue, 26 Oct 2021 07:55:48 +0200
> > Sean Nyekjaer <[email protected]> wrote:
> >
> > > From: Boris Brezillon <[email protected]>
> > >
> > > The BBT scan logic use the MTD helpers before the MTD layer had a
> > > chance to initialize the device, and that leads to issues when
> > > accessing the uninitialized suspend lock. Let's temporarily set the
> > > suspend/resume hooks to NULL to skip the lock acquire/release step.
> > >
> > > Fixes: 013e6292aaf5 ("mtd: rawnand: Simplify the locking")
> >
> > I think I already mentioned this Fixes tag should not be there.
> >
>
> No, I didn't recall that, but nevermind :)
>
> Hmm when ('mtd: core: protect access to MTD devices while in suspend')
> is backported we want to live with the use before init of the rwsem?
It's not meant to be backported. You'll need something simpler...
>
> /Sean