diff mbox series

[RESEND,v2] mfd: syscon: Use a unique name with regmap_config

Message ID 20200727211008.24225-1-s-anna@ti.com (mailing list archive)
State Mainlined
Commit e15d7f2b81d2e7d93115d46fa931b366c1cdebc2
Headers show
Series [RESEND,v2] mfd: syscon: Use a unique name with regmap_config | expand

Commit Message

Suman Anna July 27, 2020, 9:10 p.m. UTC
The DT node full name is currently being used in regmap_config
which in turn is used to create the regmap debugfs directories.
This name however is not guaranteed to be unique and the regmap
debugfs registration can fail in the cases where the syscon nodes
have the same unit-address but are present in different DT node
hierarchies. Replace this logic using the syscon reg resource
address instead (inspired from logic used while creating platform
devices) to ensure a unique name is given for each syscon.

Signed-off-by: Suman Anna <s-anna@ti.com>
---
Hi Arnd,
Lee is looking for your review on this patch. Can you please
review and provide your comments.

This is a resend of the patch that was posted previously, rebased
now onto latest kernel.

v2: https://patchwork.kernel.org/patch/11353355/
 - Fix build warning reported by kbuild test bot
v1: https://patchwork.kernel.org/patch/11346363/

 drivers/mfd/syscon.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

Comments

Arnd Bergmann July 28, 2020, 7:44 a.m. UTC | #1
On Mon, Jul 27, 2020 at 11:10 PM Suman Anna <s-anna@ti.com> wrote:
>
> The DT node full name is currently being used in regmap_config
> which in turn is used to create the regmap debugfs directories.
> This name however is not guaranteed to be unique and the regmap
> debugfs registration can fail in the cases where the syscon nodes
> have the same unit-address but are present in different DT node
> hierarchies. Replace this logic using the syscon reg resource
> address instead (inspired from logic used while creating platform
> devices) to ensure a unique name is given for each syscon.
>
> Signed-off-by: Suman Anna <s-anna@ti.com>
> ---
> Hi Arnd,
> Lee is looking for your review on this patch. Can you please
> review and provide your comments.

Sorry for missing this earlier. I think this makes sense, and I don't
expect the name change to cause problems, so

Reviewed-by: Arnd Bergmann <arnd@arndb.de>

> --- a/drivers/mfd/syscon.c
> +++ b/drivers/mfd/syscon.c
> @@ -101,12 +101,14 @@ static struct syscon *of_syscon_register(struct device_node *np, bool check_clk)
>                 }
>         }
>
> -       syscon_config.name = of_node_full_name(np);
> +       syscon_config.name = kasprintf(GFP_KERNEL, "%pOFn@%llx", np,
> +                                      (u64)res.start);

Note that you could avoid the cast by using "%pOFn@%pa", and
passing res.start by reference. Not important though, the result should
be similar, and you might not like the '0x' that this adds.

      Arnd
Lee Jones July 28, 2020, 8:01 a.m. UTC | #2
On Mon, 27 Jul 2020, Suman Anna wrote:

> The DT node full name is currently being used in regmap_config
> which in turn is used to create the regmap debugfs directories.
> This name however is not guaranteed to be unique and the regmap
> debugfs registration can fail in the cases where the syscon nodes
> have the same unit-address but are present in different DT node
> hierarchies. Replace this logic using the syscon reg resource
> address instead (inspired from logic used while creating platform
> devices) to ensure a unique name is given for each syscon.
> 
> Signed-off-by: Suman Anna <s-anna@ti.com>
> ---
> Hi Arnd,
> Lee is looking for your review on this patch. Can you please
> review and provide your comments.
> 
> This is a resend of the patch that was posted previously, rebased
> now onto latest kernel.
> 
> v2: https://patchwork.kernel.org/patch/11353355/
>  - Fix build warning reported by kbuild test bot
> v1: https://patchwork.kernel.org/patch/11346363/
> 
>  drivers/mfd/syscon.c | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)

Applied, thanks.
Suman Anna July 28, 2020, 3:34 p.m. UTC | #3
On 7/28/20 2:44 AM, Arnd Bergmann wrote:
> On Mon, Jul 27, 2020 at 11:10 PM Suman Anna <s-anna@ti.com> wrote:
>>
>> The DT node full name is currently being used in regmap_config
>> which in turn is used to create the regmap debugfs directories.
>> This name however is not guaranteed to be unique and the regmap
>> debugfs registration can fail in the cases where the syscon nodes
>> have the same unit-address but are present in different DT node
>> hierarchies. Replace this logic using the syscon reg resource
>> address instead (inspired from logic used while creating platform
>> devices) to ensure a unique name is given for each syscon.
>>
>> Signed-off-by: Suman Anna <s-anna@ti.com>
>> ---
>> Hi Arnd,
>> Lee is looking for your review on this patch. Can you please
>> review and provide your comments.
> 
> Sorry for missing this earlier. I think this makes sense, and I don't
> expect the name change to cause problems, so
> 
> Reviewed-by: Arnd Bergmann <arnd@arndb.de>

Thanks Arnd.

> 
>> --- a/drivers/mfd/syscon.c
>> +++ b/drivers/mfd/syscon.c
>> @@ -101,12 +101,14 @@ static struct syscon *of_syscon_register(struct device_node *np, bool check_clk)
>>                  }
>>          }
>>
>> -       syscon_config.name = of_node_full_name(np);
>> +       syscon_config.name = kasprintf(GFP_KERNEL, "%pOFn@%llx", np,
>> +                                      (u64)res.start);
> 
> Note that you could avoid the cast by using "%pOFn@%pa", and
> passing res.start by reference. Not important though, the result should
> be similar, and you might not like the '0x' that this adds.

Yeah, preference was not to add the leading 0x or any leading 0s. We did 
discuss about this on the original v2 submission [1].

regards
Suman

[1] https://patchwork.kernel.org/comment/23129393/
Marc Zyngier Aug. 27, 2020, 2:46 p.m. UTC | #4
Hi all,

On 2020-07-27 22:10, Suman Anna wrote:
> The DT node full name is currently being used in regmap_config
> which in turn is used to create the regmap debugfs directories.
> This name however is not guaranteed to be unique and the regmap
> debugfs registration can fail in the cases where the syscon nodes
> have the same unit-address but are present in different DT node
> hierarchies. Replace this logic using the syscon reg resource
> address instead (inspired from logic used while creating platform
> devices) to ensure a unique name is given for each syscon.
> 
> Signed-off-by: Suman Anna <s-anna@ti.com>
> ---
> Hi Arnd,
> Lee is looking for your review on this patch. Can you please
> review and provide your comments.
> 
> This is a resend of the patch that was posted previously, rebased
> now onto latest kernel.
> 
> v2: https://patchwork.kernel.org/patch/11353355/
>  - Fix build warning reported by kbuild test bot
> v1: https://patchwork.kernel.org/patch/11346363/
> 
>  drivers/mfd/syscon.c | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/mfd/syscon.c b/drivers/mfd/syscon.c
> index 3a97816d0cba..75859e492984 100644
> --- a/drivers/mfd/syscon.c
> +++ b/drivers/mfd/syscon.c
> @@ -101,12 +101,14 @@ static struct syscon *of_syscon_register(struct
> device_node *np, bool check_clk)
>  		}
>  	}
> 
> -	syscon_config.name = of_node_full_name(np);
> +	syscon_config.name = kasprintf(GFP_KERNEL, "%pOFn@%llx", np,
> +				       (u64)res.start);
>  	syscon_config.reg_stride = reg_io_width;
>  	syscon_config.val_bits = reg_io_width * 8;
>  	syscon_config.max_register = resource_size(&res) - reg_io_width;
> 
>  	regmap = regmap_init_mmio(NULL, base, &syscon_config);
> +	kfree(syscon_config.name);
>  	if (IS_ERR(regmap)) {
>  		pr_err("regmap init failed\n");
>  		ret = PTR_ERR(regmap);

This patch triggers some illegal memory accesses when debugfs is
enabled, as regmap does rely on config->name to be persistent
when the debugfs registration is deferred via regmap_debugfs_early_list
(__regmap_init() -> regmap_attach_dev() -> regmap_debugfs_init()...),
leading to a KASAN splat on demand.

I came up with the following patch that solves the issue for me.

Thanks,

         M.

 From fd3f5f2bf72df53be18d13914fe349a34f81f16b Mon Sep 17 00:00:00 2001
 From: Marc Zyngier <maz@kernel.org>
Date: Thu, 27 Aug 2020 14:45:34 +0100
Subject: [PATCH] mfd: syscon: Don't free allocated name for 
regmap_config

The name allocated for the regmap_config structure is freed
pretty early, right after the registration of the MMIO region.

Unfortunately, that doesn't follow the life cycle that debugfs
expects, as it can access the name field long after the free
has occured.

Move the free on the error path, and keep it forever otherwise.

Fixes: e15d7f2b81d2 ("mfd: syscon: Use a unique name with 
regmap_config")
Signed-off-by: Marc Zyngier <maz@kernel.org>
---
  drivers/mfd/syscon.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/mfd/syscon.c b/drivers/mfd/syscon.c
index 75859e492984..7a660411c562 100644
--- a/drivers/mfd/syscon.c
+++ b/drivers/mfd/syscon.c
@@ -108,7 +108,6 @@ static struct syscon *of_syscon_register(struct 
device_node *np, bool check_clk)
  	syscon_config.max_register = resource_size(&res) - reg_io_width;

  	regmap = regmap_init_mmio(NULL, base, &syscon_config);
-	kfree(syscon_config.name);
  	if (IS_ERR(regmap)) {
  		pr_err("regmap init failed\n");
  		ret = PTR_ERR(regmap);
@@ -145,6 +144,7 @@ static struct syscon *of_syscon_register(struct 
device_node *np, bool check_clk)
  	regmap_exit(regmap);
  err_regmap:
  	iounmap(base);
+	kfree(syscon_config.name);
  err_map:
  	kfree(syscon);
  	return ERR_PTR(ret);
Suman Anna Aug. 27, 2020, 6:28 p.m. UTC | #5
Hi Marc,

On 8/27/20 9:46 AM, Marc Zyngier wrote:
> Hi all,
> 
> On 2020-07-27 22:10, Suman Anna wrote:
>> The DT node full name is currently being used in regmap_config
>> which in turn is used to create the regmap debugfs directories.
>> This name however is not guaranteed to be unique and the regmap
>> debugfs registration can fail in the cases where the syscon nodes
>> have the same unit-address but are present in different DT node
>> hierarchies. Replace this logic using the syscon reg resource
>> address instead (inspired from logic used while creating platform
>> devices) to ensure a unique name is given for each syscon.
>>
>> Signed-off-by: Suman Anna <s-anna@ti.com>
>> ---
>> Hi Arnd,
>> Lee is looking for your review on this patch. Can you please
>> review and provide your comments.
>>
>> This is a resend of the patch that was posted previously, rebased
>> now onto latest kernel.
>>
>> v2: https://patchwork.kernel.org/patch/11353355/
>>  - Fix build warning reported by kbuild test bot
>> v1: https://patchwork.kernel.org/patch/11346363/
>>
>>  drivers/mfd/syscon.c | 4 +++-
>>  1 file changed, 3 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/mfd/syscon.c b/drivers/mfd/syscon.c
>> index 3a97816d0cba..75859e492984 100644
>> --- a/drivers/mfd/syscon.c
>> +++ b/drivers/mfd/syscon.c
>> @@ -101,12 +101,14 @@ static struct syscon *of_syscon_register(struct
>> device_node *np, bool check_clk)
>>          }
>>      }
>>
>> -    syscon_config.name = of_node_full_name(np);
>> +    syscon_config.name = kasprintf(GFP_KERNEL, "%pOFn@%llx", np,
>> +                       (u64)res.start);
>>      syscon_config.reg_stride = reg_io_width;
>>      syscon_config.val_bits = reg_io_width * 8;
>>      syscon_config.max_register = resource_size(&res) - reg_io_width;
>>
>>      regmap = regmap_init_mmio(NULL, base, &syscon_config);
>> +    kfree(syscon_config.name);
>>      if (IS_ERR(regmap)) {
>>          pr_err("regmap init failed\n");
>>          ret = PTR_ERR(regmap);
> 
> This patch triggers some illegal memory accesses when debugfs is
> enabled, as regmap does rely on config->name to be persistent
> when the debugfs registration is deferred via regmap_debugfs_early_list
> (__regmap_init() -> regmap_attach_dev() -> regmap_debugfs_init()...),
> leading to a KASAN splat on demand.
> 

Thanks, I missed the subtlety around the debugfs registration.

> I came up with the following patch that solves the issue for me.
> 
> Thanks,
> 
>         M.
> 
> From fd3f5f2bf72df53be18d13914fe349a34f81f16b Mon Sep 17 00:00:00 2001
> From: Marc Zyngier <maz@kernel.org>
> Date: Thu, 27 Aug 2020 14:45:34 +0100
> Subject: [PATCH] mfd: syscon: Don't free allocated name for regmap_config
> 
> The name allocated for the regmap_config structure is freed
> pretty early, right after the registration of the MMIO region.
> 
> Unfortunately, that doesn't follow the life cycle that debugfs
> expects, as it can access the name field long after the free
> has occured.
> 
> Move the free on the error path, and keep it forever otherwise.

Hmm, this is exactly what I was trying to avoid. The regmap_init does duplicate
the name into map->name if config->name is given, and the regmap debugfs makes
another copy of its own into debugfs_name when actually registered. If the rules
for regmap_init is that the config->name should be persistent, then I guess we
have no choice but to go with the below fix.

Does something like below help?

diff --git a/drivers/base/regmap/regmap.c b/drivers/base/regmap/regmap.c
index e93700af7e6e..96d8a0161c89 100644
--- a/drivers/base/regmap/regmap.c
+++ b/drivers/base/regmap/regmap.c
@@ -1137,7 +1137,7 @@ struct regmap *__regmap_init(struct device *dev,
                if (ret != 0)
                        goto err_regcache;
        } else {
-               regmap_debugfs_init(map, config->name);
+               regmap_debugfs_init(map, map->name);

But there are couple of other places in regmap code that uses config->name, but
those won't be exercised with the syscon code.

regards
Suman

> 
> Fixes: e15d7f2b81d2 ("mfd: syscon: Use a unique name with regmap_config")
> Signed-off-by: Marc Zyngier <maz@kernel.org>
> ---
>  drivers/mfd/syscon.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/drivers/mfd/syscon.c b/drivers/mfd/syscon.c
> index 75859e492984..7a660411c562 100644
> --- a/drivers/mfd/syscon.c
> +++ b/drivers/mfd/syscon.c
> @@ -108,7 +108,6 @@ static struct syscon *of_syscon_register(struct device_node
> *np, bool check_clk)
>      syscon_config.max_register = resource_size(&res) - reg_io_width;
> 
>      regmap = regmap_init_mmio(NULL, base, &syscon_config);
> -    kfree(syscon_config.name);
>      if (IS_ERR(regmap)) {
>          pr_err("regmap init failed\n");
>          ret = PTR_ERR(regmap);
> @@ -145,6 +144,7 @@ static struct syscon *of_syscon_register(struct device_node
> *np, bool check_clk)
>      regmap_exit(regmap);
>  err_regmap:
>      iounmap(base);
> +    kfree(syscon_config.name);
>  err_map:
>      kfree(syscon);
>      return ERR_PTR(ret);
Marc Zyngier Aug. 27, 2020, 8:06 p.m. UTC | #6
Hi Suman,

On 2020-08-27 19:28, Suman Anna wrote:
> Hi Marc,
> 
> On 8/27/20 9:46 AM, Marc Zyngier wrote:
>> Hi all,
>> 
>> On 2020-07-27 22:10, Suman Anna wrote:
>>> The DT node full name is currently being used in regmap_config
>>> which in turn is used to create the regmap debugfs directories.
>>> This name however is not guaranteed to be unique and the regmap
>>> debugfs registration can fail in the cases where the syscon nodes
>>> have the same unit-address but are present in different DT node
>>> hierarchies. Replace this logic using the syscon reg resource
>>> address instead (inspired from logic used while creating platform
>>> devices) to ensure a unique name is given for each syscon.
>>> 
>>> Signed-off-by: Suman Anna <s-anna@ti.com>
>>> ---
>>> Hi Arnd,
>>> Lee is looking for your review on this patch. Can you please
>>> review and provide your comments.
>>> 
>>> This is a resend of the patch that was posted previously, rebased
>>> now onto latest kernel.
>>> 
>>> v2: https://patchwork.kernel.org/patch/11353355/
>>>  - Fix build warning reported by kbuild test bot
>>> v1: https://patchwork.kernel.org/patch/11346363/
>>> 
>>>  drivers/mfd/syscon.c | 4 +++-
>>>  1 file changed, 3 insertions(+), 1 deletion(-)
>>> 
>>> diff --git a/drivers/mfd/syscon.c b/drivers/mfd/syscon.c
>>> index 3a97816d0cba..75859e492984 100644
>>> --- a/drivers/mfd/syscon.c
>>> +++ b/drivers/mfd/syscon.c
>>> @@ -101,12 +101,14 @@ static struct syscon *of_syscon_register(struct
>>> device_node *np, bool check_clk)
>>>          }
>>>      }
>>> 
>>> -    syscon_config.name = of_node_full_name(np);
>>> +    syscon_config.name = kasprintf(GFP_KERNEL, "%pOFn@%llx", np,
>>> +                       (u64)res.start);
>>>      syscon_config.reg_stride = reg_io_width;
>>>      syscon_config.val_bits = reg_io_width * 8;
>>>      syscon_config.max_register = resource_size(&res) - reg_io_width;
>>> 
>>>      regmap = regmap_init_mmio(NULL, base, &syscon_config);
>>> +    kfree(syscon_config.name);
>>>      if (IS_ERR(regmap)) {
>>>          pr_err("regmap init failed\n");
>>>          ret = PTR_ERR(regmap);
>> 
>> This patch triggers some illegal memory accesses when debugfs is
>> enabled, as regmap does rely on config->name to be persistent
>> when the debugfs registration is deferred via 
>> regmap_debugfs_early_list
>> (__regmap_init() -> regmap_attach_dev() -> regmap_debugfs_init()...),
>> leading to a KASAN splat on demand.
>> 
> 
> Thanks, I missed the subtlety around the debugfs registration.
> 
>> I came up with the following patch that solves the issue for me.
>> 
>> Thanks,
>> 
>>         M.
>> 
>> From fd3f5f2bf72df53be18d13914fe349a34f81f16b Mon Sep 17 00:00:00 2001
>> From: Marc Zyngier <maz@kernel.org>
>> Date: Thu, 27 Aug 2020 14:45:34 +0100
>> Subject: [PATCH] mfd: syscon: Don't free allocated name for 
>> regmap_config
>> 
>> The name allocated for the regmap_config structure is freed
>> pretty early, right after the registration of the MMIO region.
>> 
>> Unfortunately, that doesn't follow the life cycle that debugfs
>> expects, as it can access the name field long after the free
>> has occured.
>> 
>> Move the free on the error path, and keep it forever otherwise.
> 
> Hmm, this is exactly what I was trying to avoid. The regmap_init does 
> duplicate
> the name into map->name if config->name is given, and the regmap 
> debugfs makes
> another copy of its own into debugfs_name when actually registered. If 
> the rules
> for regmap_init is that the config->name should be persistent, then I 
> guess we
> have no choice but to go with the below fix.
> 
> Does something like below help?
> 
> diff --git a/drivers/base/regmap/regmap.c 
> b/drivers/base/regmap/regmap.c
> index e93700af7e6e..96d8a0161c89 100644
> --- a/drivers/base/regmap/regmap.c
> +++ b/drivers/base/regmap/regmap.c
> @@ -1137,7 +1137,7 @@ struct regmap *__regmap_init(struct device *dev,
>                 if (ret != 0)
>                         goto err_regcache;
>         } else {
> -               regmap_debugfs_init(map, config->name);
> +               regmap_debugfs_init(map, map->name);
> 
> But there are couple of other places in regmap code that uses 
> config->name, but
> those won't be exercised with the syscon code.

Is config->name always the same as map->name? If so, why don't you just
pass map once and for all? Is the lifetime of map->name the same as
that of config->name?

My worry with this approach is that we start changing stuff in a rush,
and this would IMHO deserve a thorough investigation of whether this
change is actually safe.

I'd rather take the safe approach of either keeping the memory around
until we clearly understand what the implications are (and probably
this should involve the regmap maintainer), or to revert this patch
until we figure out the actual life cycle of the various names.

Thanks,

         M.
Suman Anna Aug. 27, 2020, 8:32 p.m. UTC | #7
Hi Marc,

+ Mark Brown

On 8/27/20 3:06 PM, Marc Zyngier wrote:
> Hi Suman,
> 
> On 2020-08-27 19:28, Suman Anna wrote:
>> Hi Marc,
>>
>> On 8/27/20 9:46 AM, Marc Zyngier wrote:
>>> Hi all,
>>>
>>> On 2020-07-27 22:10, Suman Anna wrote:
>>>> The DT node full name is currently being used in regmap_config
>>>> which in turn is used to create the regmap debugfs directories.
>>>> This name however is not guaranteed to be unique and the regmap
>>>> debugfs registration can fail in the cases where the syscon nodes
>>>> have the same unit-address but are present in different DT node
>>>> hierarchies. Replace this logic using the syscon reg resource
>>>> address instead (inspired from logic used while creating platform
>>>> devices) to ensure a unique name is given for each syscon.
>>>>
>>>> Signed-off-by: Suman Anna <s-anna@ti.com>
>>>> ---
>>>> Hi Arnd,
>>>> Lee is looking for your review on this patch. Can you please
>>>> review and provide your comments.
>>>>
>>>> This is a resend of the patch that was posted previously, rebased
>>>> now onto latest kernel.
>>>>
>>>> v2: https://patchwork.kernel.org/patch/11353355/
>>>>  - Fix build warning reported by kbuild test bot
>>>> v1: https://patchwork.kernel.org/patch/11346363/
>>>>
>>>>  drivers/mfd/syscon.c | 4 +++-
>>>>  1 file changed, 3 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/mfd/syscon.c b/drivers/mfd/syscon.c
>>>> index 3a97816d0cba..75859e492984 100644
>>>> --- a/drivers/mfd/syscon.c
>>>> +++ b/drivers/mfd/syscon.c
>>>> @@ -101,12 +101,14 @@ static struct syscon *of_syscon_register(struct
>>>> device_node *np, bool check_clk)
>>>>          }
>>>>      }
>>>>
>>>> -    syscon_config.name = of_node_full_name(np);
>>>> +    syscon_config.name = kasprintf(GFP_KERNEL, "%pOFn@%llx", np,
>>>> +                       (u64)res.start);
>>>>      syscon_config.reg_stride = reg_io_width;
>>>>      syscon_config.val_bits = reg_io_width * 8;
>>>>      syscon_config.max_register = resource_size(&res) - reg_io_width;
>>>>
>>>>      regmap = regmap_init_mmio(NULL, base, &syscon_config);
>>>> +    kfree(syscon_config.name);
>>>>      if (IS_ERR(regmap)) {
>>>>          pr_err("regmap init failed\n");
>>>>          ret = PTR_ERR(regmap);
>>>
>>> This patch triggers some illegal memory accesses when debugfs is
>>> enabled, as regmap does rely on config->name to be persistent
>>> when the debugfs registration is deferred via regmap_debugfs_early_list
>>> (__regmap_init() -> regmap_attach_dev() -> regmap_debugfs_init()...),
>>> leading to a KASAN splat on demand.
>>>
>>
>> Thanks, I missed the subtlety around the debugfs registration.
>>
>>> I came up with the following patch that solves the issue for me.
>>>
>>> Thanks,
>>>
>>>         M.
>>>
>>> From fd3f5f2bf72df53be18d13914fe349a34f81f16b Mon Sep 17 00:00:00 2001
>>> From: Marc Zyngier <maz@kernel.org>
>>> Date: Thu, 27 Aug 2020 14:45:34 +0100
>>> Subject: [PATCH] mfd: syscon: Don't free allocated name for regmap_config
>>>
>>> The name allocated for the regmap_config structure is freed
>>> pretty early, right after the registration of the MMIO region.
>>>
>>> Unfortunately, that doesn't follow the life cycle that debugfs
>>> expects, as it can access the name field long after the free
>>> has occured.
>>>
>>> Move the free on the error path, and keep it forever otherwise.
>>
>> Hmm, this is exactly what I was trying to avoid. The regmap_init does duplicate
>> the name into map->name if config->name is given, and the regmap debugfs makes
>> another copy of its own into debugfs_name when actually registered. If the rules
>> for regmap_init is that the config->name should be persistent, then I guess we
>> have no choice but to go with the below fix.
>>
>> Does something like below help?
>>
>> diff --git a/drivers/base/regmap/regmap.c b/drivers/base/regmap/regmap.c
>> index e93700af7e6e..96d8a0161c89 100644
>> --- a/drivers/base/regmap/regmap.c
>> +++ b/drivers/base/regmap/regmap.c
>> @@ -1137,7 +1137,7 @@ struct regmap *__regmap_init(struct device *dev,
>>                 if (ret != 0)
>>                         goto err_regcache;
>>         } else {
>> -               regmap_debugfs_init(map, config->name);
>> +               regmap_debugfs_init(map, map->name);
>>
>> But there are couple of other places in regmap code that uses config->name, but
>> those won't be exercised with the syscon code.
> 
> Is config->name always the same as map->name? If so, why don't you just
> pass map once and for all? Is the lifetime of map->name the same as
> that of config->name?

map->name is created (kstrdup_const) from config->name if not NULL, so above
replacement should be exactly equivalent, map is filled in _regmap_init. But it
does make the regmap_debugfs_init callsites in the file look dissimilar.

> 
> My worry with this approach is that we start changing stuff in a rush,
> and this would IMHO deserve a thorough investigation of whether this
> change is actually safe.
> 
> I'd rather take the safe approach of either keeping the memory around
> until we clearly understand what the implications are (and probably
> this should involve the regmap maintainer), or to revert this patch
> until we figure out the actual life cycle of the various names.

Yeah, agreed. Let's see what Mark suggests.

Mark,
Can you clarify the lifecycle expectations on the config->name and do you have
any suggestions here?

regards
Suman
Mark Brown Aug. 28, 2020, 10:40 a.m. UTC | #8
On Thu, Aug 27, 2020 at 03:32:05PM -0500, Suman Anna wrote:

> Can you clarify the lifecycle expectations on the config->name and do you have
> any suggestions here?

The regmap name is expected to be managed by the caller and should be
live as long as the regmap is live, it is almost always static data.
Marc Zyngier Sept. 3, 2020, 8:26 a.m. UTC | #9
On 2020-08-27 21:32, Suman Anna wrote:
> Hi Marc,
> 
> + Mark Brown
> 
> On 8/27/20 3:06 PM, Marc Zyngier wrote:
>> Hi Suman,
>> 
>> On 2020-08-27 19:28, Suman Anna wrote:
>>> Hi Marc,
>>> 
>>> On 8/27/20 9:46 AM, Marc Zyngier wrote:

[...]

>>>> This patch triggers some illegal memory accesses when debugfs is
>>>> enabled, as regmap does rely on config->name to be persistent
>>>> when the debugfs registration is deferred via 
>>>> regmap_debugfs_early_list
>>>> (__regmap_init() -> regmap_attach_dev() -> 
>>>> regmap_debugfs_init()...),
>>>> leading to a KASAN splat on demand.
>>>> 
>>> 
>>> Thanks, I missed the subtlety around the debugfs registration.
>>> 
>>>> I came up with the following patch that solves the issue for me.
>>>> 
>>>> Thanks,
>>>> 
>>>>         M.
>>>> 
>>>> From fd3f5f2bf72df53be18d13914fe349a34f81f16b Mon Sep 17 00:00:00 
>>>> 2001
>>>> From: Marc Zyngier <maz@kernel.org>
>>>> Date: Thu, 27 Aug 2020 14:45:34 +0100
>>>> Subject: [PATCH] mfd: syscon: Don't free allocated name for 
>>>> regmap_config
>>>> 
>>>> The name allocated for the regmap_config structure is freed
>>>> pretty early, right after the registration of the MMIO region.
>>>> 
>>>> Unfortunately, that doesn't follow the life cycle that debugfs
>>>> expects, as it can access the name field long after the free
>>>> has occured.
>>>> 
>>>> Move the free on the error path, and keep it forever otherwise.
>>> 
>>> Hmm, this is exactly what I was trying to avoid. The regmap_init does 
>>> duplicate
>>> the name into map->name if config->name is given, and the regmap 
>>> debugfs makes
>>> another copy of its own into debugfs_name when actually registered. 
>>> If the rules
>>> for regmap_init is that the config->name should be persistent, then I 
>>> guess we
>>> have no choice but to go with the below fix.
>>> 
>>> Does something like below help?
>>> 
>>> diff --git a/drivers/base/regmap/regmap.c 
>>> b/drivers/base/regmap/regmap.c
>>> index e93700af7e6e..96d8a0161c89 100644
>>> --- a/drivers/base/regmap/regmap.c
>>> +++ b/drivers/base/regmap/regmap.c
>>> @@ -1137,7 +1137,7 @@ struct regmap *__regmap_init(struct device 
>>> *dev,
>>>                 if (ret != 0)
>>>                         goto err_regcache;
>>>         } else {
>>> -               regmap_debugfs_init(map, config->name);
>>> +               regmap_debugfs_init(map, map->name);
>>> 
>>> But there are couple of other places in regmap code that uses 
>>> config->name, but
>>> those won't be exercised with the syscon code.
>> 
>> Is config->name always the same as map->name? If so, why don't you 
>> just
>> pass map once and for all? Is the lifetime of map->name the same as
>> that of config->name?
> 
> map->name is created (kstrdup_const) from config->name if not NULL, so 
> above
> replacement should be exactly equivalent, map is filled in 
> _regmap_init. But it
> does make the regmap_debugfs_init callsites in the file look 
> dissimilar.
> 
>> 
>> My worry with this approach is that we start changing stuff in a rush,
>> and this would IMHO deserve a thorough investigation of whether this
>> change is actually safe.
>> 
>> I'd rather take the safe approach of either keeping the memory around
>> until we clearly understand what the implications are (and probably
>> this should involve the regmap maintainer), or to revert this patch
>> until we figure out the actual life cycle of the various names.
> 
> Yeah, agreed. Let's see what Mark suggests.
> 
> Mark,
> Can you clarify the lifecycle expectations on the config->name and do 
> you have
> any suggestions here?

Have we reached a conclusion here? Can we get a fix in mainline?

Thanks,

         M.
Suman Anna Sept. 3, 2020, 1:25 p.m. UTC | #10
On 9/3/20 3:26 AM, Marc Zyngier wrote:
> On 2020-08-27 21:32, Suman Anna wrote:
>> Hi Marc,
>>
>> + Mark Brown
>>
>> On 8/27/20 3:06 PM, Marc Zyngier wrote:
>>> Hi Suman,
>>>
>>> On 2020-08-27 19:28, Suman Anna wrote:
>>>> Hi Marc,
>>>>
>>>> On 8/27/20 9:46 AM, Marc Zyngier wrote:
> 
> [...]
> 
>>>>> This patch triggers some illegal memory accesses when debugfs is
>>>>> enabled, as regmap does rely on config->name to be persistent
>>>>> when the debugfs registration is deferred via regmap_debugfs_early_list
>>>>> (__regmap_init() -> regmap_attach_dev() -> regmap_debugfs_init()...),
>>>>> leading to a KASAN splat on demand.
>>>>>
>>>>
>>>> Thanks, I missed the subtlety around the debugfs registration.
>>>>
>>>>> I came up with the following patch that solves the issue for me.
>>>>>
>>>>> Thanks,
>>>>>
>>>>>         M.
>>>>>
>>>>> From fd3f5f2bf72df53be18d13914fe349a34f81f16b Mon Sep 17 00:00:00 2001
>>>>> From: Marc Zyngier <maz@kernel.org>
>>>>> Date: Thu, 27 Aug 2020 14:45:34 +0100
>>>>> Subject: [PATCH] mfd: syscon: Don't free allocated name for regmap_config
>>>>>
>>>>> The name allocated for the regmap_config structure is freed
>>>>> pretty early, right after the registration of the MMIO region.
>>>>>
>>>>> Unfortunately, that doesn't follow the life cycle that debugfs
>>>>> expects, as it can access the name field long after the free
>>>>> has occured.
>>>>>
>>>>> Move the free on the error path, and keep it forever otherwise.
>>>>
>>>> Hmm, this is exactly what I was trying to avoid. The regmap_init does duplicate
>>>> the name into map->name if config->name is given, and the regmap debugfs makes
>>>> another copy of its own into debugfs_name when actually registered. If the
>>>> rules
>>>> for regmap_init is that the config->name should be persistent, then I guess we
>>>> have no choice but to go with the below fix.
>>>>
>>>> Does something like below help?
>>>>
>>>> diff --git a/drivers/base/regmap/regmap.c b/drivers/base/regmap/regmap.c
>>>> index e93700af7e6e..96d8a0161c89 100644
>>>> --- a/drivers/base/regmap/regmap.c
>>>> +++ b/drivers/base/regmap/regmap.c
>>>> @@ -1137,7 +1137,7 @@ struct regmap *__regmap_init(struct device *dev,
>>>>                 if (ret != 0)
>>>>                         goto err_regcache;
>>>>         } else {
>>>> -               regmap_debugfs_init(map, config->name);
>>>> +               regmap_debugfs_init(map, map->name);
>>>>
>>>> But there are couple of other places in regmap code that uses config->name, but
>>>> those won't be exercised with the syscon code.
>>>
>>> Is config->name always the same as map->name? If so, why don't you just
>>> pass map once and for all? Is the lifetime of map->name the same as
>>> that of config->name?
>>
>> map->name is created (kstrdup_const) from config->name if not NULL, so above
>> replacement should be exactly equivalent, map is filled in _regmap_init. But it
>> does make the regmap_debugfs_init callsites in the file look dissimilar.
>>
>>>
>>> My worry with this approach is that we start changing stuff in a rush,
>>> and this would IMHO deserve a thorough investigation of whether this
>>> change is actually safe.
>>>
>>> I'd rather take the safe approach of either keeping the memory around
>>> until we clearly understand what the implications are (and probably
>>> this should involve the regmap maintainer), or to revert this patch
>>> until we figure out the actual life cycle of the various names.
>>
>> Yeah, agreed. Let's see what Mark suggests.
>>
>> Mark,
>> Can you clarify the lifecycle expectations on the config->name and do you have
>> any suggestions here?
> 
> Have we reached a conclusion here? Can we get a fix in mainline?

Marc, we can go with your patch based on Mark's response.

regards
Suman
Marc Zyngier Sept. 3, 2020, 4:04 p.m. UTC | #11
On 2020-09-03 14:25, Suman Anna wrote:
> On 9/3/20 3:26 AM, Marc Zyngier wrote:

[...]

>> Have we reached a conclusion here? Can we get a fix in mainline?
> 
> Marc, we can go with your patch based on Mark's response.

Patch resent[1].

         M.

[1] https://lore.kernel.org/r/20200903160237.932818-1-maz@kernel.org
diff mbox series

Patch

diff --git a/drivers/mfd/syscon.c b/drivers/mfd/syscon.c
index 3a97816d0cba..75859e492984 100644
--- a/drivers/mfd/syscon.c
+++ b/drivers/mfd/syscon.c
@@ -101,12 +101,14 @@  static struct syscon *of_syscon_register(struct device_node *np, bool check_clk)
 		}
 	}
 
-	syscon_config.name = of_node_full_name(np);
+	syscon_config.name = kasprintf(GFP_KERNEL, "%pOFn@%llx", np,
+				       (u64)res.start);
 	syscon_config.reg_stride = reg_io_width;
 	syscon_config.val_bits = reg_io_width * 8;
 	syscon_config.max_register = resource_size(&res) - reg_io_width;
 
 	regmap = regmap_init_mmio(NULL, base, &syscon_config);
+	kfree(syscon_config.name);
 	if (IS_ERR(regmap)) {
 		pr_err("regmap init failed\n");
 		ret = PTR_ERR(regmap);