Re: ping // [PATCH] mtd:cfi_cmdset_0002: fix atomic sleep bug when CONFIG_MTD_XIP=y

From: Vignesh Raghavendra <vigneshr@ti.com>
To: Miquel Raynal <miquel.raynal@bootlin.com>,
	Xiaoming Ni <nixiaoming@huawei.com>
Cc: wangle6@huawei.com, tudor.ambarus@microchip.com, richard@nod.at,
	linux-kernel@vger.kernel.org, stable@vger.kernel.org,
	linux-mtd@lists.infradead.org, gregkh@linuxfoundation.org,
	tglx@linutronix.de
Subject: Re: ping // [PATCH] mtd:cfi_cmdset_0002: fix atomic sleep bug when CONFIG_MTD_XIP=y
Date: Tue, 8 Dec 2020 00:29:21 +0530	[thread overview]
Message-ID: <73b539eb-616e-64d8-07d8-4606da2ea2ea@ti.com> (raw)
In-Reply-To: <20201207115228.0a6de398@xps13>

Hi Xiaoming,

On 12/7/20 4:23 PM, Miquel Raynal wrote:
> Hi Xiaoming,
> 
> Xiaoming Ni <nixiaoming@huawei.com> wrote on Mon, 7 Dec 2020 18:48:33
> +0800:
> 
>> ping
>>
>> On 2020/11/27 21:07, Xiaoming Ni wrote:
>>> When CONFIG_MTD_XIP=y, local_irq_disable() is called in xip_disable().
>>> To avoid sleep in interrupt context, we need to call local_irq_enable()
>>> before schedule().
>>>
>>> The problem call stack is as follows:
>>> bug1:
>>> 	do_write_oneword_retry()
>>> 		xip_disable()
>>> 			local_irq_disable()
>>> 		do_write_oneword_once()
>>> 			schedule()
>>> bug2:
>>> 	do_write_buffer()
>>> 		xip_disable()
>>> 			local_irq_disable()
>>> 		do_write_buffer_wait()
>>> 			schedule()
>>> bug3:
>>> 	do_erase_chip()
>>> 		xip_disable()
>>> 			local_irq_disable()
>>> 		schedule()
>>> bug4:
>>> 	do_erase_oneblock()
>>> 		xip_disable()
>>> 			local_irq_disable()
>>> 		schedule()
>>>
>>> Fixes: 02b15e343aee ("[MTD] XIP for AMD CFI flash.")
>>> Cc: stable@vger.kernel.org # v2.6.13
>>> Signed-off-by: Xiaoming Ni <nixiaoming@huawei.com>
>>> ---
>>>   drivers/mtd/chips/cfi_cmdset_0002.c | 16 ++++++++++++++++
>>>   1 file changed, 16 insertions(+)
>>>
>>> diff --git a/drivers/mtd/chips/cfi_cmdset_0002.c b/drivers/mtd/chips/cfi_cmdset_0002.c
>>> index a1f3e1031c3d..12c3776f093a 100644
>>> --- a/drivers/mtd/chips/cfi_cmdset_0002.c
>>> +++ b/drivers/mtd/chips/cfi_cmdset_0002.c
>>> @@ -1682,7 +1682,11 @@ static int __xipram do_write_oneword_once(struct map_info *map,
>>>   			set_current_state(TASK_UNINTERRUPTIBLE);
>>>   			add_wait_queue(&chip->wq, &wait);
>>>   			mutex_unlock(&chip->mutex);
>>> +			if (IS_ENABLED(CONFIG_MTD_XIP))
>>> +				local_irq_enable();
>>>   			schedule();
>>> +			if (IS_ENABLED(CONFIG_MTD_XIP))
>>> +				local_irq_disable();
> 
> The fix really seems strange to me. I will let Vignesh decide but I
> think we should consider updating/fixing xip_disable instead.

Agree with Miquel. Have you done any testing 
or is this purely based on code inspection?

What about comment before xip_disable() function:

/*
 * No interrupt what so ever can be serviced while the flash isn't in array
 * mode.  This is ensured by the xip_disable() and xip_enable() functions
 * enclosing any code path where the flash is known not to be in array mode.
 * And within a XIP disabled code path, only functions marked with __xipram
 * may be called and nothing else (it's a good thing to inspect generated
 * assembly to make sure inline functions were actually inlined and that gcc
 * didn't emit calls to its own support functions). Also configuring MTD CFI
 * support to a single buswidth and a single interleave is also recommended.
 */

So, I don't think the fix is as simple as this patch.

Regards
Vignesh

______________________________________________________
Linux MTD discussion mailing list
http://lists.infradead.org/mailman/listinfo/linux-mtd/