[4.19.y PATCH] tmpfs: fix unable to remount nr_inodes from limited to unlimited

[4.19.y PATCH] tmpfs: fix unable to remount nr_inodes from limited to unlimited

[yu kuai]

tmpfs support 'size', 'nr_blocks' and 'nr_inodes' mount options. mount or
remount them to zero means unlimited. 'size' and 'br_blocks' can remount
from limited to unlimited, while 'nr_inodes' can't.

The problem is fixed since upstream commit 0b5071dd323d ("
shmem_parse_options(): use a separate structure to keep the results"). But
in order to backport it, the amount of related patches need to backport is
huge. 

So, I made some local changes to fix the problem.

Signed-off-by: yu kuai <yukuai3@huawei.com>
---
 mm/shmem.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/mm/shmem.c b/mm/shmem.c
index 3c8742655756..966fc69ee8fb 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -3444,7 +3444,7 @@ static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
 	inodes = sbinfo->max_inodes - sbinfo->free_inodes;
 	if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
 		goto out;
-	if (config.max_inodes < inodes)
+	if (config.max_inodes && config.max_inodes < inodes)
 		goto out;
 	/*
 	 * Those tests disallow limited->unlimited while any are in use;
@@ -3460,7 +3460,10 @@ static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
 	sbinfo->huge = config.huge;
 	sbinfo->max_blocks  = config.max_blocks;
 	sbinfo->max_inodes  = config.max_inodes;
-	sbinfo->free_inodes = config.max_inodes - inodes;
+	if (!config.max_inodes)
+		sbinfo->free_inodes = 0;
+	else
+		sbinfo->free_inodes = config.max_inodes - inodes;
 
 	/*
 	 * Preserve previous mempolicy unless mpol remount option was specified.
-- 
2.17.2

Re: [4.19.y PATCH] tmpfs: fix unable to remount nr_inodes from limited to unlimited

[Greg KH]

On Wed, Dec 04, 2019 at 09:11:37PM +0800, yu kuai wrote:
> tmpfs support 'size', 'nr_blocks' and 'nr_inodes' mount options. mount or
> remount them to zero means unlimited. 'size' and 'br_blocks' can remount
> from limited to unlimited, while 'nr_inodes' can't.
> 
> The problem is fixed since upstream commit 0b5071dd323d ("
> shmem_parse_options(): use a separate structure to keep the results"). But
> in order to backport it, the amount of related patches need to backport is
> huge. 
> 
> So, I made some local changes to fix the problem.
> 
> Signed-off-by: yu kuai <yukuai3@huawei.com>
> ---
>  mm/shmem.c | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)


<formletter>

This is not the correct way to submit patches for inclusion in the
stable kernel tree.  Please read:
    https://www.kernel.org/doc/html/latest/process/stable-kernel-rules.html
for how to do this properly.

</formletter>

Re: [4.19.y PATCH] tmpfs: fix unable to remount nr_inodes from limited to unlimited

[Greg KH]

On Wed, Dec 04, 2019 at 09:11:37PM +0800, yu kuai wrote:
> tmpfs support 'size', 'nr_blocks' and 'nr_inodes' mount options. mount or
> remount them to zero means unlimited. 'size' and 'br_blocks' can remount
> from limited to unlimited, while 'nr_inodes' can't.
> 
> The problem is fixed since upstream commit 0b5071dd323d ("
> shmem_parse_options(): use a separate structure to keep the results"). But
> in order to backport it, the amount of related patches need to backport is
> huge. 
> 
> So, I made some local changes to fix the problem.
> 
> Signed-off-by: yu kuai <yukuai3@huawei.com>
> ---
>  mm/shmem.c | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
> 
> diff --git a/mm/shmem.c b/mm/shmem.c
> index 3c8742655756..966fc69ee8fb 100644
> --- a/mm/shmem.c
> +++ b/mm/shmem.c
> @@ -3444,7 +3444,7 @@ static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
>  	inodes = sbinfo->max_inodes - sbinfo->free_inodes;
>  	if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
>  		goto out;
> -	if (config.max_inodes < inodes)
> +	if (config.max_inodes && config.max_inodes < inodes)
>  		goto out;
>  	/*
>  	 * Those tests disallow limited->unlimited while any are in use;
> @@ -3460,7 +3460,10 @@ static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
>  	sbinfo->huge = config.huge;
>  	sbinfo->max_blocks  = config.max_blocks;
>  	sbinfo->max_inodes  = config.max_inodes;
> -	sbinfo->free_inodes = config.max_inodes - inodes;
> +	if (!config.max_inodes)
> +		sbinfo->free_inodes = 0;
> +	else
> +		sbinfo->free_inodes = config.max_inodes - inodes;
>  
>  	/*
>  	 * Preserve previous mempolicy unless mpol remount option was specified.
> -- 
> 2.17.2
> 

Hm, sorry about my bot, this looked like an odd one-off patch.

What about 5.3.y, should this patch also go there as well?

But is it really an issue as this is a new "feature" that 5.4 now has,
can't you just use 5.4.y if you need this type of thing?  It's never
worked in the past, right?

thanks,

greg k-h

Re: [4.19.y PATCH] tmpfs: fix unable to remount nr_inodes from limited to unlimited

[Hugh Dickins]

On Wed, 4 Dec 2019, Greg KH wrote:
> On Wed, Dec 04, 2019 at 09:11:37PM +0800, yu kuai wrote:
> > tmpfs support 'size', 'nr_blocks' and 'nr_inodes' mount options. mount or
> > remount them to zero means unlimited. 'size' and 'br_blocks' can remount
> > from limited to unlimited, while 'nr_inodes' can't.
> > 
> > The problem is fixed since upstream commit 0b5071dd323d ("
> > shmem_parse_options(): use a separate structure to keep the results"). But
> > in order to backport it, the amount of related patches need to backport is
> > huge. 
> > 
> > So, I made some local changes to fix the problem.
> > 
> > Signed-off-by: yu kuai <yukuai3@huawei.com>
> > ---
> >  mm/shmem.c | 7 +++++--
> >  1 file changed, 5 insertions(+), 2 deletions(-)
> > 
> > diff --git a/mm/shmem.c b/mm/shmem.c
> > index 3c8742655756..966fc69ee8fb 100644
> > --- a/mm/shmem.c
> > +++ b/mm/shmem.c
> > @@ -3444,7 +3444,7 @@ static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
> >  	inodes = sbinfo->max_inodes - sbinfo->free_inodes;
> >  	if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
> >  		goto out;
> > -	if (config.max_inodes < inodes)
> > +	if (config.max_inodes && config.max_inodes < inodes)
> >  		goto out;
> >  	/*
> >  	 * Those tests disallow limited->unlimited while any are in use;
> > @@ -3460,7 +3460,10 @@ static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
> >  	sbinfo->huge = config.huge;
> >  	sbinfo->max_blocks  = config.max_blocks;
> >  	sbinfo->max_inodes  = config.max_inodes;
> > -	sbinfo->free_inodes = config.max_inodes - inodes;
> > +	if (!config.max_inodes)
> > +		sbinfo->free_inodes = 0;
> > +	else
> > +		sbinfo->free_inodes = config.max_inodes - inodes;
> >  
> >  	/*
> >  	 * Preserve previous mempolicy unless mpol remount option was specified.
> > -- 
> > 2.17.2
> > 
> 
> Hm, sorry about my bot, this looked like an odd one-off patch.
> 
> What about 5.3.y, should this patch also go there as well?
> 
> But is it really an issue as this is a new "feature" that 5.4 now has,
> can't you just use 5.4.y if you need this type of thing?  It's never
> worked in the past, right?

Yes, please ignore this for stable, Greg: it appears to be a new feature
in 5.4: one that I should have noticed when testing, but failed to do so
(and it may even be something that I foisted unthinkingly on Al when
suggesting mods to his and David's originals).

Yu Kuai: many thanks for noticing and reporting this, I was unconscious
of changing behavior here.  Notice how the 5.4 shmem_reconfigure() still
has a comment above it saying "we disallow change from limited->unlimited
blocks/inodes while any are in use" - and root inode is always in use.
Notice how your 4.19 patch does nothing for max_blocks, so remounting
with nr_blocks=0 will still fail, once a non-empty file has been created.

I agree that it's not obvious why limited->unlimited needs to fail,
and perhaps a nice (worthwhile?) little enhancement to allow that;
but it was unintentional, and now (but not today) I have to go back
to remind myself why 2.6.13 implemented it with that restriction,
and whether there are any fixes needed to the new behavior in 5.4
(at the least, we ought to fix that comment in 5.5).

Hugh

Re: [4.19.y PATCH] tmpfs: fix unable to remount nr_inodes from limited to unlimited

[yukuai (C)]

On 2019/12/5 4:13, Hugh Dickins wrote:
> On Wed, 4 Dec 2019, Greg KH wrote:
>> On Wed, Dec 04, 2019 at 09:11:37PM +0800, yu kuai wrote:
>>> tmpfs support 'size', 'nr_blocks' and 'nr_inodes' mount options. mount or
>>> remount them to zero means unlimited. 'size' and 'br_blocks' can remount
>>> from limited to unlimited, while 'nr_inodes' can't.
>>>
>>> The problem is fixed since upstream commit 0b5071dd323d ("
>>> shmem_parse_options(): use a separate structure to keep the results"). But
>>> in order to backport it, the amount of related patches need to backport is
>>> huge.
>>>
>>> So, I made some local changes to fix the problem.
>>>
>>> Signed-off-by: yu kuai <yukuai3@huawei.com>
>>> ---
>>>   mm/shmem.c | 7 +++++--
>>>   1 file changed, 5 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/mm/shmem.c b/mm/shmem.c
>>> index 3c8742655756..966fc69ee8fb 100644
>>> --- a/mm/shmem.c
>>> +++ b/mm/shmem.c
>>> @@ -3444,7 +3444,7 @@ static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
>>>   	inodes = sbinfo->max_inodes - sbinfo->free_inodes;
>>>   	if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
>>>   		goto out;
>>> -	if (config.max_inodes < inodes)
>>> +	if (config.max_inodes && config.max_inodes < inodes)
>>>   		goto out;
>>>   	/*
>>>   	 * Those tests disallow limited->unlimited while any are in use;
>>> @@ -3460,7 +3460,10 @@ static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
>>>   	sbinfo->huge = config.huge;
>>>   	sbinfo->max_blocks  = config.max_blocks;
>>>   	sbinfo->max_inodes  = config.max_inodes;
>>> -	sbinfo->free_inodes = config.max_inodes - inodes;
>>> +	if (!config.max_inodes)
>>> +		sbinfo->free_inodes = 0;
>>> +	else
>>> +		sbinfo->free_inodes = config.max_inodes - inodes;
>>>   
>>>   	/*
>>>   	 * Preserve previous mempolicy unless mpol remount option was specified.
>>> -- 
>>> 2.17.2
>>>
>>
>> Hm, sorry about my bot, this looked like an odd one-off patch.
>>
>> What about 5.3.y, should this patch also go there as well?
>>
>> But is it really an issue as this is a new "feature" that 5.4 now has,
>> can't you just use 5.4.y if you need this type of thing?  It's never
>> worked in the past, right?
> 
> Yes, please ignore this for stable, Greg: it appears to be a new feature
> in 5.4: one that I should have noticed when testing, but failed to do so
> (and it may even be something that I foisted unthinkingly on Al when
> suggesting mods to his and David's originals).
> 
> Yu Kuai: many thanks for noticing and reporting this, I was unconscious
> of changing behavior here.  Notice how the 5.4 shmem_reconfigure() still
> has a comment above it saying "we disallow change from limited->unlimited
> blocks/inodes while any are in use" - and root inode is always in use.
> Notice how your 4.19 patch does nothing for max_blocks, so remounting
> with nr_blocks=0 will still fail, once a non-empty file has been created.
Thank you for your response, I missed that in my test.
> 
> I agree that it's not obvious why limited->unlimited needs to fail,
> and perhaps a nice (worthwhile?) little enhancement to allow that;
> but it was unintentional, and now (but not today) I have to go back
> to remind myself why 2.6.13 implemented it with that restriction,
> and whether there are any fixes needed to the new behavior in 5.4
> (at the least, we ought to fix that comment in 5.5).

I was confused about the comment in 5.5: "we disallow change from 
limited->unlimited blocks/inodes while any are in use", it seems change 
from limited to unlimited will always succed. Maybe this is
more appropriate: "Note that we allow change from limited to unlimited 
blocks/inodes. But we disallow change from unlimited->limited, because 
in that case we have no record of how much is already in use. And we 
disallow change blocks/inodes to less than the amount that is aready in use"

Thanks
Yu Kuai

Re: [4.19.y PATCH] tmpfs: fix unable to remount nr_inodes from limited to unlimited

[yukuai (C)]

On 2019/12/5 1:33, Greg KH wrote:
> On Wed, Dec 04, 2019 at 09:11:37PM +0800, yu kuai wrote:
>> tmpfs support 'size', 'nr_blocks' and 'nr_inodes' mount options. mount or
>> remount them to zero means unlimited. 'size' and 'br_blocks' can remount
>> from limited to unlimited, while 'nr_inodes' can't.
>>
>> The problem is fixed since upstream commit 0b5071dd323d ("
>> shmem_parse_options(): use a separate structure to keep the results"). But
>> in order to backport it, the amount of related patches need to backport is
>> huge.
>>
>> So, I made some local changes to fix the problem.
>>
>> Signed-off-by: yu kuai <yukuai3@huawei.com>
>> ---
>>   mm/shmem.c | 7 +++++--
>>   1 file changed, 5 insertions(+), 2 deletions(-)
>>
>> diff --git a/mm/shmem.c b/mm/shmem.c
>> index 3c8742655756..966fc69ee8fb 100644
>> --- a/mm/shmem.c
>> +++ b/mm/shmem.c
>> @@ -3444,7 +3444,7 @@ static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
>>   	inodes = sbinfo->max_inodes - sbinfo->free_inodes;
>>   	if (percpu_counter_compare(&sbinfo->used_blocks, config.max_blocks) > 0)
>>   		goto out;
>> -	if (config.max_inodes < inodes)
>> +	if (config.max_inodes && config.max_inodes < inodes)
>>   		goto out;
>>   	/*
>>   	 * Those tests disallow limited->unlimited while any are in use;
>> @@ -3460,7 +3460,10 @@ static int shmem_remount_fs(struct super_block *sb, int *flags, char *data)
>>   	sbinfo->huge = config.huge;
>>   	sbinfo->max_blocks  = config.max_blocks;
>>   	sbinfo->max_inodes  = config.max_inodes;
>> -	sbinfo->free_inodes = config.max_inodes - inodes;
>> +	if (!config.max_inodes)
>> +		sbinfo->free_inodes = 0;
>> +	else
>> +		sbinfo->free_inodes = config.max_inodes - inodes;
>>   
>>   	/*
>>   	 * Preserve previous mempolicy unless mpol remount option was specified.
>> -- 
>> 2.17.2
>>
> 
> Hm, sorry about my bot, this looked like an odd one-off patch.
> 
> What about 5.3.y, should this patch also go there as well?
Yes, 4.4y and 5.3y have the same problem.
> 
> But is it really an issue as this is a new "feature" that 5.4 now has,
> can't you just use 5.4.y if you need this type of thing?  It's never
> worked in the past, right?
> 
It's true that it never worked in the past. I thoult it might be a bug 
because the behavior is not like what the document said.

Thanks for your response.
Yu Kuai