Patchwork UBI: fix refill_wl_user_pool()

login
register
mail settings
Submitter Richard Weinberger
Date Aug. 26, 2013, 10:08 a.m.
Message ID <1377511730-14834-1-git-send-email-richard@nod.at>
Download mbox | patch
Permalink /patch/269854/
State New
Headers show

Comments

Richard Weinberger - Aug. 26, 2013, 10:08 a.m.
If no free PEBs are available refill_wl_user_pool() must not
return with -ENOSPC immediately.
It has to block till produce_free_peb() produced a free PEB.

Reported-and-Tested-by: Richard Genoud <richard.genoud@gmail.com>
Signed-off-by: Richard Weinberger <richard@nod.at>
---
 drivers/mtd/ubi/wl.c | 4 ----
 1 file changed, 4 deletions(-)
Richard Weinberger - Sept. 4, 2013, 8:20 a.m.
Am 26.08.2013 12:08, schrieb Richard Weinberger:
> If no free PEBs are available refill_wl_user_pool() must not
> return with -ENOSPC immediately.
> It has to block till produce_free_peb() produced a free PEB.
> 
> Reported-and-Tested-by: Richard Genoud <richard.genoud@gmail.com>
> Signed-off-by: Richard Weinberger <richard@nod.at>
> ---
>  drivers/mtd/ubi/wl.c | 4 ----
>  1 file changed, 4 deletions(-)
> 
> diff --git a/drivers/mtd/ubi/wl.c b/drivers/mtd/ubi/wl.c
> index c95bfb1..02317c1 100644
> --- a/drivers/mtd/ubi/wl.c
> +++ b/drivers/mtd/ubi/wl.c
> @@ -599,10 +599,6 @@ static void refill_wl_user_pool(struct ubi_device *ubi)
>  	return_unused_pool_pebs(ubi, pool);
>  
>  	for (pool->size = 0; pool->size < pool->max_size; pool->size++) {
> -		if (!ubi->free.rb_node ||
> -		   (ubi->free_count - ubi->beb_rsvd_pebs < 1))
> -			break;
> -
>  		pool->pebs[pool->size] = __wl_get_peb(ubi);
>  		if (pool->pebs[pool->size] < 0)
>  			break;
> 

Artem, please don't forget do apply this patch.

Thanks,
//richard
David Mosberger-Tang - Sept. 4, 2013, 3:50 p.m.
Richard,

Is there more background on the bug this patch fixes?
We are working on root-causing an issue where a UBIFS write fails with
ENOSPC, even though there is no plausible reason that the filesystem
should be out of space.  The ENOSPC errors persist across multiple
reboots, only to disappear magically later on.  We haven't found a
reliable way to reproduce this so far, so we're still tapping a bit in
the dark.

  --david

On Wed, Sep 4, 2013 at 2:20 AM, Richard Weinberger <richard@nod.at> wrote:
> Am 26.08.2013 12:08, schrieb Richard Weinberger:
>> If no free PEBs are available refill_wl_user_pool() must not
>> return with -ENOSPC immediately.
>> It has to block till produce_free_peb() produced a free PEB.
>>
>> Reported-and-Tested-by: Richard Genoud <richard.genoud@gmail.com>
>> Signed-off-by: Richard Weinberger <richard@nod.at>
>> ---
>>  drivers/mtd/ubi/wl.c | 4 ----
>>  1 file changed, 4 deletions(-)
>>
>> diff --git a/drivers/mtd/ubi/wl.c b/drivers/mtd/ubi/wl.c
>> index c95bfb1..02317c1 100644
>> --- a/drivers/mtd/ubi/wl.c
>> +++ b/drivers/mtd/ubi/wl.c
>> @@ -599,10 +599,6 @@ static void refill_wl_user_pool(struct ubi_device *ubi)
>>       return_unused_pool_pebs(ubi, pool);
>>
>>       for (pool->size = 0; pool->size < pool->max_size; pool->size++) {
>> -             if (!ubi->free.rb_node ||
>> -                (ubi->free_count - ubi->beb_rsvd_pebs < 1))
>> -                     break;
>> -
>>               pool->pebs[pool->size] = __wl_get_peb(ubi);
>>               if (pool->pebs[pool->size] < 0)
>>                       break;
>>
>
> Artem, please don't forget do apply this patch.
>
> Thanks,
> //richard
>
> ______________________________________________________
> Linux MTD discussion mailing list
> http://lists.infradead.org/mailman/listinfo/linux-mtd/
Richard Weinberger - Sept. 4, 2013, 4:01 p.m.
David,

Am 04.09.2013 17:50, schrieb David Mosberger-Tang:
> Richard,
> 
> Is there more background on the bug this patch fixes?
> We are working on root-causing an issue where a UBIFS write fails with
> ENOSPC, even though there is no plausible reason that the filesystem
> should be out of space.  The ENOSPC errors persist across multiple
> reboots, only to disappear magically later on.  We haven't found a
> reliable way to reproduce this so far, so we're still tapping a bit in
> the dark.

The said patch fixes an issue if your Kernel was build with experimental Fastmap support.
If you have CONFIG_MTD_UBI_FASTMAP=n you are not affected.

Fastmap uses two pools of available PEBs. Every time a pool is filled with used PEBs
it writes out a new fastmap to the flash.
The bug in refill_wl_user_pool() was that the function did not try hard enough to get a free PEB
and failed too early with -ENOSPC.

Thanks,
//richard

>   --david
> 
> On Wed, Sep 4, 2013 at 2:20 AM, Richard Weinberger <richard@nod.at> wrote:
>> Am 26.08.2013 12:08, schrieb Richard Weinberger:
>>> If no free PEBs are available refill_wl_user_pool() must not
>>> return with -ENOSPC immediately.
>>> It has to block till produce_free_peb() produced a free PEB.
>>>
>>> Reported-and-Tested-by: Richard Genoud <richard.genoud@gmail.com>
>>> Signed-off-by: Richard Weinberger <richard@nod.at>
>>> ---
>>>  drivers/mtd/ubi/wl.c | 4 ----
>>>  1 file changed, 4 deletions(-)
>>>
>>> diff --git a/drivers/mtd/ubi/wl.c b/drivers/mtd/ubi/wl.c
>>> index c95bfb1..02317c1 100644
>>> --- a/drivers/mtd/ubi/wl.c
>>> +++ b/drivers/mtd/ubi/wl.c
>>> @@ -599,10 +599,6 @@ static void refill_wl_user_pool(struct ubi_device *ubi)
>>>       return_unused_pool_pebs(ubi, pool);
>>>
>>>       for (pool->size = 0; pool->size < pool->max_size; pool->size++) {
>>> -             if (!ubi->free.rb_node ||
>>> -                (ubi->free_count - ubi->beb_rsvd_pebs < 1))
>>> -                     break;
>>> -
>>>               pool->pebs[pool->size] = __wl_get_peb(ubi);
>>>               if (pool->pebs[pool->size] < 0)
>>>                       break;
>>>
>>
>> Artem, please don't forget do apply this patch.
>>
>> Thanks,
>> //richard
>>
>> ______________________________________________________
>> Linux MTD discussion mailing list
>> http://lists.infradead.org/mailman/listinfo/linux-mtd/
David Mosberger-Tang - Sept. 4, 2013, 4:03 p.m.
Richard,

Thanks for the info.  We are not using FASTMAP, so it doesn't explain
what we're seeing.

Best regards,

  --david

On Wed, Sep 4, 2013 at 10:01 AM, Richard Weinberger <richard@nod.at> wrote:
> David,
>
> Am 04.09.2013 17:50, schrieb David Mosberger-Tang:
>> Richard,
>>
>> Is there more background on the bug this patch fixes?
>> We are working on root-causing an issue where a UBIFS write fails with
>> ENOSPC, even though there is no plausible reason that the filesystem
>> should be out of space.  The ENOSPC errors persist across multiple
>> reboots, only to disappear magically later on.  We haven't found a
>> reliable way to reproduce this so far, so we're still tapping a bit in
>> the dark.
>
> The said patch fixes an issue if your Kernel was build with experimental Fastmap support.
> If you have CONFIG_MTD_UBI_FASTMAP=n you are not affected.
>
> Fastmap uses two pools of available PEBs. Every time a pool is filled with used PEBs
> it writes out a new fastmap to the flash.
> The bug in refill_wl_user_pool() was that the function did not try hard enough to get a free PEB
> and failed too early with -ENOSPC.
>
> Thanks,
> //richard
>
>>   --david
>>
>> On Wed, Sep 4, 2013 at 2:20 AM, Richard Weinberger <richard@nod.at> wrote:
>>> Am 26.08.2013 12:08, schrieb Richard Weinberger:
>>>> If no free PEBs are available refill_wl_user_pool() must not
>>>> return with -ENOSPC immediately.
>>>> It has to block till produce_free_peb() produced a free PEB.
>>>>
>>>> Reported-and-Tested-by: Richard Genoud <richard.genoud@gmail.com>
>>>> Signed-off-by: Richard Weinberger <richard@nod.at>
>>>> ---
>>>>  drivers/mtd/ubi/wl.c | 4 ----
>>>>  1 file changed, 4 deletions(-)
>>>>
>>>> diff --git a/drivers/mtd/ubi/wl.c b/drivers/mtd/ubi/wl.c
>>>> index c95bfb1..02317c1 100644
>>>> --- a/drivers/mtd/ubi/wl.c
>>>> +++ b/drivers/mtd/ubi/wl.c
>>>> @@ -599,10 +599,6 @@ static void refill_wl_user_pool(struct ubi_device *ubi)
>>>>       return_unused_pool_pebs(ubi, pool);
>>>>
>>>>       for (pool->size = 0; pool->size < pool->max_size; pool->size++) {
>>>> -             if (!ubi->free.rb_node ||
>>>> -                (ubi->free_count - ubi->beb_rsvd_pebs < 1))
>>>> -                     break;
>>>> -
>>>>               pool->pebs[pool->size] = __wl_get_peb(ubi);
>>>>               if (pool->pebs[pool->size] < 0)
>>>>                       break;
>>>>
>>>
>>> Artem, please don't forget do apply this patch.
>>>
>>> Thanks,
>>> //richard
>>>
>>> ______________________________________________________
>>> Linux MTD discussion mailing list
>>> http://lists.infradead.org/mailman/listinfo/linux-mtd/
>
Richard Weinberger - Sept. 5, 2013, 6:09 a.m.
David,

Am 04.09.2013 18:03, schrieb David Mosberger-Tang:
> Thanks for the info.  We are not using FASTMAP, so it doesn't explain
> what we're seeing.

Maybe you are suffering from the issue fixed by:
http://git.infradead.org/linux-ubi.git/commit/5ef4414f4bc26a19cfd5cd11aee9697a863e4d51

This patch will hopefully make it into 3.12-rc1 and then back into -stable as soon as possible.

Thanks,
//richard

> Best regards,
> 
>   --david
> 
> On Wed, Sep 4, 2013 at 10:01 AM, Richard Weinberger <richard@nod.at> wrote:
>> David,
>>
>> Am 04.09.2013 17:50, schrieb David Mosberger-Tang:
>>> Richard,
>>>
>>> Is there more background on the bug this patch fixes?
>>> We are working on root-causing an issue where a UBIFS write fails with
>>> ENOSPC, even though there is no plausible reason that the filesystem
>>> should be out of space.  The ENOSPC errors persist across multiple
>>> reboots, only to disappear magically later on.  We haven't found a
>>> reliable way to reproduce this so far, so we're still tapping a bit in
>>> the dark.
>>
>> The said patch fixes an issue if your Kernel was build with experimental Fastmap support.
>> If you have CONFIG_MTD_UBI_FASTMAP=n you are not affected.
>>
>> Fastmap uses two pools of available PEBs. Every time a pool is filled with used PEBs
>> it writes out a new fastmap to the flash.
>> The bug in refill_wl_user_pool() was that the function did not try hard enough to get a free PEB
>> and failed too early with -ENOSPC.
>>
>> Thanks,
>> //richard
>>
>>>   --david
>>>
>>> On Wed, Sep 4, 2013 at 2:20 AM, Richard Weinberger <richard@nod.at> wrote:
>>>> Am 26.08.2013 12:08, schrieb Richard Weinberger:
>>>>> If no free PEBs are available refill_wl_user_pool() must not
>>>>> return with -ENOSPC immediately.
>>>>> It has to block till produce_free_peb() produced a free PEB.
>>>>>
>>>>> Reported-and-Tested-by: Richard Genoud <richard.genoud@gmail.com>
>>>>> Signed-off-by: Richard Weinberger <richard@nod.at>
>>>>> ---
>>>>>  drivers/mtd/ubi/wl.c | 4 ----
>>>>>  1 file changed, 4 deletions(-)
>>>>>
>>>>> diff --git a/drivers/mtd/ubi/wl.c b/drivers/mtd/ubi/wl.c
>>>>> index c95bfb1..02317c1 100644
>>>>> --- a/drivers/mtd/ubi/wl.c
>>>>> +++ b/drivers/mtd/ubi/wl.c
>>>>> @@ -599,10 +599,6 @@ static void refill_wl_user_pool(struct ubi_device *ubi)
>>>>>       return_unused_pool_pebs(ubi, pool);
>>>>>
>>>>>       for (pool->size = 0; pool->size < pool->max_size; pool->size++) {
>>>>> -             if (!ubi->free.rb_node ||
>>>>> -                (ubi->free_count - ubi->beb_rsvd_pebs < 1))
>>>>> -                     break;
>>>>> -
>>>>>               pool->pebs[pool->size] = __wl_get_peb(ubi);
>>>>>               if (pool->pebs[pool->size] < 0)
>>>>>                       break;
>>>>>
>>>>
>>>> Artem, please don't forget do apply this patch.
>>>>
>>>> Thanks,
>>>> //richard
>>>>
>>>> ______________________________________________________
>>>> Linux MTD discussion mailing list
>>>> http://lists.infradead.org/mailman/listinfo/linux-mtd/
>>
David Mosberger-Tang - Sept. 5, 2013, 5:16 p.m.
The platform we're seeing this on is running an old kernel: 2.6.27.10.
We don't know yet whether it happens on newer kernels (we are trying
to reproduce the issue on both older and newer kernels, but that may
take a month; so far, we have only seen the issue on systems in the
field).

  --david


On Thu, Sep 5, 2013 at 12:09 AM, Richard Weinberger <richard@nod.at> wrote:
> David,
>
> Am 04.09.2013 18:03, schrieb David Mosberger-Tang:
>> Thanks for the info.  We are not using FASTMAP, so it doesn't explain
>> what we're seeing.
>
> Maybe you are suffering from the issue fixed by:
> http://git.infradead.org/linux-ubi.git/commit/5ef4414f4bc26a19cfd5cd11aee9697a863e4d51
>
> This patch will hopefully make it into 3.12-rc1 and then back into -stable as soon as possible.
>
> Thanks,
> //richard
>
>> Best regards,
>>
>>   --david
>>
>> On Wed, Sep 4, 2013 at 10:01 AM, Richard Weinberger <richard@nod.at> wrote:
>>> David,
>>>
>>> Am 04.09.2013 17:50, schrieb David Mosberger-Tang:
>>>> Richard,
>>>>
>>>> Is there more background on the bug this patch fixes?
>>>> We are working on root-causing an issue where a UBIFS write fails with
>>>> ENOSPC, even though there is no plausible reason that the filesystem
>>>> should be out of space.  The ENOSPC errors persist across multiple
>>>> reboots, only to disappear magically later on.  We haven't found a
>>>> reliable way to reproduce this so far, so we're still tapping a bit in
>>>> the dark.
>>>
>>> The said patch fixes an issue if your Kernel was build with experimental Fastmap support.
>>> If you have CONFIG_MTD_UBI_FASTMAP=n you are not affected.
>>>
>>> Fastmap uses two pools of available PEBs. Every time a pool is filled with used PEBs
>>> it writes out a new fastmap to the flash.
>>> The bug in refill_wl_user_pool() was that the function did not try hard enough to get a free PEB
>>> and failed too early with -ENOSPC.
>>>
>>> Thanks,
>>> //richard
>>>
>>>>   --david
>>>>
>>>> On Wed, Sep 4, 2013 at 2:20 AM, Richard Weinberger <richard@nod.at> wrote:
>>>>> Am 26.08.2013 12:08, schrieb Richard Weinberger:
>>>>>> If no free PEBs are available refill_wl_user_pool() must not
>>>>>> return with -ENOSPC immediately.
>>>>>> It has to block till produce_free_peb() produced a free PEB.
>>>>>>
>>>>>> Reported-and-Tested-by: Richard Genoud <richard.genoud@gmail.com>
>>>>>> Signed-off-by: Richard Weinberger <richard@nod.at>
>>>>>> ---
>>>>>>  drivers/mtd/ubi/wl.c | 4 ----
>>>>>>  1 file changed, 4 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/mtd/ubi/wl.c b/drivers/mtd/ubi/wl.c
>>>>>> index c95bfb1..02317c1 100644
>>>>>> --- a/drivers/mtd/ubi/wl.c
>>>>>> +++ b/drivers/mtd/ubi/wl.c
>>>>>> @@ -599,10 +599,6 @@ static void refill_wl_user_pool(struct ubi_device *ubi)
>>>>>>       return_unused_pool_pebs(ubi, pool);
>>>>>>
>>>>>>       for (pool->size = 0; pool->size < pool->max_size; pool->size++) {
>>>>>> -             if (!ubi->free.rb_node ||
>>>>>> -                (ubi->free_count - ubi->beb_rsvd_pebs < 1))
>>>>>> -                     break;
>>>>>> -
>>>>>>               pool->pebs[pool->size] = __wl_get_peb(ubi);
>>>>>>               if (pool->pebs[pool->size] < 0)
>>>>>>                       break;
>>>>>>
>>>>>
>>>>> Artem, please don't forget do apply this patch.
>>>>>
>>>>> Thanks,
>>>>> //richard
>>>>>
>>>>> ______________________________________________________
>>>>> Linux MTD discussion mailing list
>>>>> http://lists.infradead.org/mailman/listinfo/linux-mtd/
>>>
>
richard -rw- weinberger - Sept. 20, 2013, 4:09 p.m.
On Wed, Sep 4, 2013 at 10:20 AM, Richard Weinberger <richard@nod.at> wrote:
> Am 26.08.2013 12:08, schrieb Richard Weinberger:
>> If no free PEBs are available refill_wl_user_pool() must not
>> return with -ENOSPC immediately.
>> It has to block till produce_free_peb() produced a free PEB.
>>
>> Reported-and-Tested-by: Richard Genoud <richard.genoud@gmail.com>
>> Signed-off-by: Richard Weinberger <richard@nod.at>
>> ---
>>  drivers/mtd/ubi/wl.c | 4 ----
>>  1 file changed, 4 deletions(-)
>>
>> diff --git a/drivers/mtd/ubi/wl.c b/drivers/mtd/ubi/wl.c
>> index c95bfb1..02317c1 100644
>> --- a/drivers/mtd/ubi/wl.c
>> +++ b/drivers/mtd/ubi/wl.c
>> @@ -599,10 +599,6 @@ static void refill_wl_user_pool(struct ubi_device *ubi)
>>       return_unused_pool_pebs(ubi, pool);
>>
>>       for (pool->size = 0; pool->size < pool->max_size; pool->size++) {
>> -             if (!ubi->free.rb_node ||
>> -                (ubi->free_count - ubi->beb_rsvd_pebs < 1))
>> -                     break;
>> -
>>               pool->pebs[pool->size] = __wl_get_peb(ubi);
>>               if (pool->pebs[pool->size] < 0)
>>                       break;
>>
>
> Artem, please don't forget do apply this patch.

*ping*

Patch

diff --git a/drivers/mtd/ubi/wl.c b/drivers/mtd/ubi/wl.c
index c95bfb1..02317c1 100644
--- a/drivers/mtd/ubi/wl.c
+++ b/drivers/mtd/ubi/wl.c
@@ -599,10 +599,6 @@  static void refill_wl_user_pool(struct ubi_device *ubi)
 	return_unused_pool_pebs(ubi, pool);
 
 	for (pool->size = 0; pool->size < pool->max_size; pool->size++) {
-		if (!ubi->free.rb_node ||
-		   (ubi->free_count - ubi->beb_rsvd_pebs < 1))
-			break;
-
 		pool->pebs[pool->size] = __wl_get_peb(ubi);
 		if (pool->pebs[pool->size] < 0)
 			break;