Patchwork [v2] mtd: cmdlinepart: fix the wrong check condition

Hi Huang, Artem,

On Sat, 25 Aug 2012 16:06:50 -0400 Huang Shijie <shijie8@gmail.com> wrote:
> diff --git a/drivers/mtd/cmdlinepart.c b/drivers/mtd/cmdlinepart.c
> index fc960a3..216d751 100644
> --- a/drivers/mtd/cmdlinepart.c
> +++ b/drivers/mtd/cmdlinepart.c
> @@ -322,13 +322,16 @@ static int parse_cmdline_partitions(struct mtd_info *master,
>  	struct cmdline_mtd_partition *part;
>  	const char *mtd_id = master->name;
>  
> +	if (!mtd_id)
> +		return 0;
> +
>  	/* parse command line */
>  	if (!cmdline_parsed)
>  		mtdpart_setup_real(cmdline);
>  
>  	for(part = partitions; part; part = part->next)
>  	{
> -		if ((!mtd_id) || (!strcmp(part->mtd_id, mtd_id)))
> +		if (!strcmp(part->mtd_id, mtd_id))
>  		{
>  			for(i = 0, offset = 0; i < part->num_parts; i++)
>  			{

This changes the behavior of cmdling parsing, which might affect users
expecting the old behavior.

According to the remark above 'parse_cmdline_partitions':

 * It returns partitions for the requested mtd device, or
 * the first one in the chain if a NULL mtd_id is passed in.

I think the purpose of a NULL 'mtd_id' was to support simple systems
where there's a single driver and a single chip.
The driver could be dumb, not specifying its 'mtd_info->name'
(thus, a NULL mtd_id is passed).

In this case, since the system is simply configured (one driver, one
chip), 'parse_cmdline_partitions' simply disregards the "mtd-id" name
specified in the cmdline string, allowing the user to present some
arbitrary string there.

I quite remember seeing this pattern somewhere in the past, I don't know
if it's still used, though.

Obviously if you have many drivers (and many chips) in a system, that
won't work; the drivers must initialize 'mtd_info->name' and the user
should present a cmdline that has explicit 'mtd-id's.

So question is, would we like to prohibit NULL mtd-id?

If so, we must make sure all drivers are properly assigning their
'mtd_info->name', and all users correctly specifying 'mtd-id' in their
"mtdparts" cmdline strings.

Regards,
Shmulik

The `mtd_id` is set by the name of a mtd device driver.

As a nand controller driver, even we do not set the @name of
the mtd_info{}, the nand_get_flash_type() will set it with
the nand type's name. So the `mtd_id` can never be NULL in this
case.

But as a nor controller driver which may does not call the
nand_get_flash_type(), there is a risk that the `mtd_id` becames NULL.

If the `mtd_id` is NULL, the check condition will be true.
If we accidentally set some partitions in the kernel command line,
just like:
          #gpmi-nand:20m(boot),20m(kernel),1g(rootfs),-(user)

The cmdlinepart may parses out several mtd partitions right now.
This is obviously wrong. We even do not enable the gpmi-nand in
this case.

The patch comes from Artem's suggestion code which is better then mine.

Signed-off-by: Huang Shijie <shijie8@gmail.com>
---
 drivers/mtd/cmdlinepart.c |    5 ++++-
 1 files changed, 4 insertions(+), 1 deletions(-)

On Sat, Aug 25, 2012 at 5:31 AM, Shmulik Ladkani
<shmulik.ladkani@gmail.com> wrote:
> Hi Huang, Artem,
>
> On Sat, 25 Aug 2012 16:06:50 -0400 Huang Shijie <shijie8@gmail.com> wrote:
>> diff --git a/drivers/mtd/cmdlinepart.c b/drivers/mtd/cmdlinepart.c
>> index fc960a3..216d751 100644
>> --- a/drivers/mtd/cmdlinepart.c
>> +++ b/drivers/mtd/cmdlinepart.c
>> @@ -322,13 +322,16 @@ static int parse_cmdline_partitions(struct mtd_info *master,
>>       struct cmdline_mtd_partition *part;
>>       const char *mtd_id = master->name;
>>
>> +     if (!mtd_id)
>> +             return 0;
>> +
>>       /* parse command line */
>>       if (!cmdline_parsed)
>>               mtdpart_setup_real(cmdline);
>>
>>       for(part = partitions; part; part = part->next)
>>       {
>> -             if ((!mtd_id) || (!strcmp(part->mtd_id, mtd_id)))
>> +             if (!strcmp(part->mtd_id, mtd_id))
>>               {
>>                       for(i = 0, offset = 0; i < part->num_parts; i++)
>>                       {
>
> This changes the behavior of cmdling parsing, which might affect users
> expecting the old behavior.
>
> According to the remark above 'parse_cmdline_partitions':
>
>  * It returns partitions for the requested mtd device, or
>  * the first one in the chain if a NULL mtd_id is passed in.
>
> I think the purpose of a NULL 'mtd_id' was to support simple systems
> where there's a single driver and a single chip.
> The driver could be dumb, not specifying its 'mtd_info->name'
> (thus, a NULL mtd_id is passed).
>
> In this case, since the system is simply configured (one driver, one
> chip), 'parse_cmdline_partitions' simply disregards the "mtd-id" name
> specified in the cmdline string, allowing the user to present some
> arbitrary string there.
>
> I quite remember seeing this pattern somewhere in the past, I don't know
> if it's still used, though.
>
thanks for the detail explanation.


> Obviously if you have many drivers (and many chips) in a system, that
> won't work; the drivers must initialize 'mtd_info->name' and the user
> should present a cmdline that has explicit 'mtd-id's.
>
> So question is, would we like to prohibit NULL mtd-id?

I prefer to prohibit the NULL mtd-id. The code looks strange enough.


>
> If so, we must make sure all drivers are properly assigning their
> 'mtd_info->name', and all users correctly specifying 'mtd-id' in their
> "mtdparts" cmdline strings.
yes. I agree that all the driver should set the mtd_info->name properly.

Best Regards
Huang Shijie

>
> Regards,
> Shmulik

On Sat, 2012-08-25 at 12:31 +0300, Shmulik Ladkani wrote:
> Hi Huang, Artem,
> 
> On Sat, 25 Aug 2012 16:06:50 -0400 Huang Shijie <shijie8@gmail.com> wrote:
> > diff --git a/drivers/mtd/cmdlinepart.c b/drivers/mtd/cmdlinepart.c
> > index fc960a3..216d751 100644
> > --- a/drivers/mtd/cmdlinepart.c
> > +++ b/drivers/mtd/cmdlinepart.c
> > @@ -322,13 +322,16 @@ static int parse_cmdline_partitions(struct mtd_info *master,
> >  	struct cmdline_mtd_partition *part;
> >  	const char *mtd_id = master->name;
> >  
> > +	if (!mtd_id)
> > +		return 0;
> > +
> >  	/* parse command line */
> >  	if (!cmdline_parsed)
> >  		mtdpart_setup_real(cmdline);
> >  
> >  	for(part = partitions; part; part = part->next)
> >  	{
> > -		if ((!mtd_id) || (!strcmp(part->mtd_id, mtd_id)))
> > +		if (!strcmp(part->mtd_id, mtd_id))
> >  		{
> >  			for(i = 0, offset = 0; i < part->num_parts; i++)
> >  			{
> 
> This changes the behavior of cmdling parsing, which might affect users
> expecting the old behavior.

Yes, you are right, we should not change the mtd_id hack unless we have
checked all the users.