[PATCH] mwifiex: fix kernel crash after shutdown command timeout

[PATCH] mwifiex: fix kernel crash after shutdown command timeout

[Amitkumar Karwar]

We observed a SHUTDOWN command timeout during reboot stress test due
to a corner case firmware bug. It leads to use-after-free on adapter
structure pointer and crash.

We already have a cancel_work_sync() call in teardown thread. This
issue is fixed by having this call just before mwifiex_remove_card().
At this point no further work will be scheduled.

Signed-off-by: Amitkumar Karwar <akarwar@marvell.com>
Signed-off-by: Cathy Luo <cluo@marvell.com>
---
 drivers/net/wireless/marvell/mwifiex/pcie.c | 3 +--
 drivers/net/wireless/marvell/mwifiex/sdio.c | 3 +--
 2 files changed, 2 insertions(+), 4 deletions(-)

diff --git a/drivers/net/wireless/marvell/mwifiex/pcie.c b/drivers/net/wireless/marvell/mwifiex/pcie.c
index a0d9180..f31c5ea 100644
--- a/drivers/net/wireless/marvell/mwifiex/pcie.c
+++ b/drivers/net/wireless/marvell/mwifiex/pcie.c
@@ -294,8 +294,6 @@ static void mwifiex_pcie_remove(struct pci_dev *pdev)
 	if (!adapter || !adapter->priv_num)
 		return;
 
-	cancel_work_sync(&card->work);
-
 	reg = card->pcie.reg;
 	if (reg)
 		ret = mwifiex_read_reg(adapter, reg->fw_status, &fw_status);
@@ -312,6 +310,7 @@ static void mwifiex_pcie_remove(struct pci_dev *pdev)
 		mwifiex_init_shutdown_fw(priv, MWIFIEX_FUNC_SHUTDOWN);
 	}
 
+	cancel_work_sync(&card->work);
 	mwifiex_remove_card(adapter);
 }
 
diff --git a/drivers/net/wireless/marvell/mwifiex/sdio.c b/drivers/net/wireless/marvell/mwifiex/sdio.c
index a4b356d..9534b47 100644
--- a/drivers/net/wireless/marvell/mwifiex/sdio.c
+++ b/drivers/net/wireless/marvell/mwifiex/sdio.c
@@ -387,8 +387,6 @@ static int mwifiex_check_winner_status(struct mwifiex_adapter *adapter)
 	if (!adapter || !adapter->priv_num)
 		return;
 
-	cancel_work_sync(&card->work);
-
 	mwifiex_dbg(adapter, INFO, "info: SDIO func num=%d\n", func->num);
 
 	ret = mwifiex_sdio_read_fw_status(adapter, &firmware_stat);
@@ -400,6 +398,7 @@ static int mwifiex_check_winner_status(struct mwifiex_adapter *adapter)
 		mwifiex_init_shutdown_fw(priv, MWIFIEX_FUNC_SHUTDOWN);
 	}
 
+	cancel_work_sync(&card->work);
 	mwifiex_remove_card(adapter);
 }
 
-- 
1.9.1

Re: [PATCH] mwifiex: fix kernel crash after shutdown command timeout

[Brian Norris]

Hi Amit,

You managed to CC several other Google folks, but not me, the one who
actually reviewed most of this! You might consider including me in the
future :)

On Fri, Feb 24, 2017 at 06:59:24PM +0530, Amitkumar Karwar wrote:
> We observed a SHUTDOWN command timeout during reboot stress test due
> to a corner case firmware bug. It leads to use-after-free on adapter
> structure pointer and crash.
> 
> We already have a cancel_work_sync() call in teardown thread. This
> issue is fixed by having this call just before mwifiex_remove_card().
> At this point no further work will be scheduled.
> 
> Signed-off-by: Amitkumar Karwar <akarwar@marvell.com>
> Signed-off-by: Cathy Luo <cluo@marvell.com>

I'm testing this artificially by testing things like this concurrently:

rmmod mwifiex_pcie &
cat /sys/kernel/debug/mwifiex/mlan0/device_dump

I'm using a 4.4-based kernel (plus quite a few backports) at the moment
and I'm having problems (I can retest on upstream if really needed), and
pretty sure this patch is buggy.

> ---
>  drivers/net/wireless/marvell/mwifiex/pcie.c | 3 +--
>  drivers/net/wireless/marvell/mwifiex/sdio.c | 3 +--
>  2 files changed, 2 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/net/wireless/marvell/mwifiex/pcie.c b/drivers/net/wireless/marvell/mwifiex/pcie.c
> index a0d9180..f31c5ea 100644
> --- a/drivers/net/wireless/marvell/mwifiex/pcie.c
> +++ b/drivers/net/wireless/marvell/mwifiex/pcie.c
> @@ -294,8 +294,6 @@ static void mwifiex_pcie_remove(struct pci_dev *pdev)
>  	if (!adapter || !adapter->priv_num)
>  		return;
>  
> -	cancel_work_sync(&card->work);
> -
>  	reg = card->pcie.reg;
>  	if (reg)
>  		ret = mwifiex_read_reg(adapter, reg->fw_status, &fw_status);
> @@ -312,6 +310,7 @@ static void mwifiex_pcie_remove(struct pci_dev *pdev)
>  		mwifiex_init_shutdown_fw(priv, MWIFIEX_FUNC_SHUTDOWN);
>  	}
>  
> +	cancel_work_sync(&card->work);

I don't think we want to move the cancellation to be this far; see the
mwifiex_init_shutdown_fw() above! If I add a msleep(3000) below this,
then run:

  rmmod mwifiex_pcie & sleep 0.5; cat /sys/kernel/debug/mwifiex/mlan0/device_dump

I can trigger an abort in mwifiex_pcie_rdwr_firmware(). The problem is
that you still allow a command timeout + firmware dump worker to still
race with the shutdown -- in this case, I think it's
mwifiex_init_shutdown_fw() that's disabling the device.

I think the real solution is to, somewhere before we shutdown the
firmware, *really* prevent any further work to be scheduled to
&card->work. Maybe that means adding another flag so that the worker
will just abort quickly in that case? So it's something like:

	card->worker_flags |= DONT_RUN_ANY_MORE;
	cancel_work_sync(&card->work);

	... (this can be done either above the FIRMWARE_READY_PCIE
	check, or else you need to write a different version for
	FIRMWARE_READY_PCIE vs. !FIRMWARE_READY_PCIE) ... but definitely
	before mwifiex_init_shutdown_fw() ) ...

And in mwifiex_pcie_work():

	if (card->worker_flags & DONT_RUN_ANY_MORE)
		return;

IOW, NAK to this patch.

Brian

>  	mwifiex_remove_card(adapter);
>  }
>  
> diff --git a/drivers/net/wireless/marvell/mwifiex/sdio.c b/drivers/net/wireless/marvell/mwifiex/sdio.c
> index a4b356d..9534b47 100644
> --- a/drivers/net/wireless/marvell/mwifiex/sdio.c
> +++ b/drivers/net/wireless/marvell/mwifiex/sdio.c
> @@ -387,8 +387,6 @@ static int mwifiex_check_winner_status(struct mwifiex_adapter *adapter)
>  	if (!adapter || !adapter->priv_num)
>  		return;
>  
> -	cancel_work_sync(&card->work);
> -
>  	mwifiex_dbg(adapter, INFO, "info: SDIO func num=%d\n", func->num);
>  
>  	ret = mwifiex_sdio_read_fw_status(adapter, &firmware_stat);
> @@ -400,6 +398,7 @@ static int mwifiex_check_winner_status(struct mwifiex_adapter *adapter)
>  		mwifiex_init_shutdown_fw(priv, MWIFIEX_FUNC_SHUTDOWN);
>  	}
>  
> +	cancel_work_sync(&card->work);
>  	mwifiex_remove_card(adapter);
>  }
>  
> -- 
> 1.9.1
> 

Re: [PATCH] mwifiex: fix kernel crash after shutdown command timeout

[Brian Norris]

On Tue, Mar 14, 2017 at 11:33 AM, Brian Norris <briannorris@chromium.org> wrote:
>
> Hi Amit,
>
> You managed to CC several other Google folks, but not me, the one who
> actually reviewed most of this! You might consider including me in the
> future :)


Oops, in fact you did CC me... my bad. But I usually use my
@chromium.org for this stuff.

Re: [PATCH] mwifiex: fix kernel crash after shutdown command timeout

[Amitkumar Karwar]

> From: Brian Norris [mailto:briannorris@chromium.org]
> Sent: Wednesday, March 15, 2017 12:03 AM
> To: Amitkumar Karwar
> Cc: linux-wireless@vger.kernel.org; Cathy Luo; Nishant Sarmukadam;
> rajatja@google.com; dmitry.torokhov@gmail.com
> Subject: [EXT] Re: [PATCH] mwifiex: fix kernel crash after shutdown
> command timeout
>  
> On Fri, Feb 24, 2017 at 06:59:24PM +0530, Amitkumar Karwar wrote:
> > We observed a SHUTDOWN command timeout during reboot stress test due
> > to a corner case firmware bug. It leads to use-after-free on adapter
> > structure pointer and crash.
> >
> > We already have a cancel_work_sync() call in teardown thread. This
> > issue is fixed by having this call just before mwifiex_remove_card().
> > At this point no further work will be scheduled.
> >
> > Signed-off-by: Amitkumar Karwar <akarwar@marvell.com>
> > Signed-off-by: Cathy Luo <cluo@marvell.com>
> 
> I'm testing this artificially by testing things like this concurrently:
> 
> rmmod mwifiex_pcie &
> cat /sys/kernel/debug/mwifiex/mlan0/device_dump
> 
> I'm using a 4.4-based kernel (plus quite a few backports) at the moment
> and I'm having problems (I can retest on upstream if really needed),
> and pretty sure this patch is buggy.
> 
> > ---
> >  drivers/net/wireless/marvell/mwifiex/pcie.c | 3 +--
> > drivers/net/wireless/marvell/mwifiex/sdio.c | 3 +--
> >  2 files changed, 2 insertions(+), 4 deletions(-)
> >
> > diff --git a/drivers/net/wireless/marvell/mwifiex/pcie.c
> > b/drivers/net/wireless/marvell/mwifiex/pcie.c
> > index a0d9180..f31c5ea 100644
> > --- a/drivers/net/wireless/marvell/mwifiex/pcie.c
> > +++ b/drivers/net/wireless/marvell/mwifiex/pcie.c
> > @@ -294,8 +294,6 @@ static void mwifiex_pcie_remove(struct pci_dev
> *pdev)
> >  	if (!adapter || !adapter->priv_num)
> >  		return;
> >
> > -	cancel_work_sync(&card->work);
> > -
> >  	reg = card->pcie.reg;
> >  	if (reg)
> >  		ret = mwifiex_read_reg(adapter, reg->fw_status,
> &fw_status); @@
> > -312,6 +310,7 @@ static void mwifiex_pcie_remove(struct pci_dev
> *pdev)
> >  		mwifiex_init_shutdown_fw(priv, MWIFIEX_FUNC_SHUTDOWN);
> >  	}
> >
> > +	cancel_work_sync(&card->work);
> 
> I don't think we want to move the cancellation to be this far; see the
> mwifiex_init_shutdown_fw() above! If I add a msleep(3000) below this,
> then run:
> 
>   rmmod mwifiex_pcie & sleep 0.5; cat
> /sys/kernel/debug/mwifiex/mlan0/device_dump
> 
> I can trigger an abort in mwifiex_pcie_rdwr_firmware(). The problem is
> that you still allow a command timeout + firmware dump worker to still
> race with the shutdown -- in this case, I think it's
> mwifiex_init_shutdown_fw() that's disabling the device.
> 
> I think the real solution is to, somewhere before we shutdown the
> firmware, *really* prevent any further work to be scheduled to &card-
> >work. Maybe that means adding another flag so that the worker will
> just abort quickly in that case? So it's something like:
> 
> 	card->worker_flags |= DONT_RUN_ANY_MORE;
> 	cancel_work_sync(&card->work);
> 
> 	... (this can be done either above the FIRMWARE_READY_PCIE
> 	check, or else you need to write a different version for
> 	FIRMWARE_READY_PCIE vs. !FIRMWARE_READY_PCIE) ... but definitely
> 	before mwifiex_init_shutdown_fw() ) ...
> 
> And in mwifiex_pcie_work():
> 
> 	if (card->worker_flags & DONT_RUN_ANY_MORE)
> 		return;
> 

Thanks for the review.
You are right. This can be cleanly fixed with a extra worker flag(DONT_RUN_ANY_MORE)
I will submit updated version with this approach.

Regards,
Amitkumar