Re: ext4: cache all of an extent tree's leaf block upon reading

Re: ext4: cache all of an extent tree's leaf block upon reading

[Dave Jones]

On Thu, Sep 05, 2013 at 01:38:48AM +0000, Linux Kernel wrote:
 > Gitweb:     http://git.kernel.org/linus/;a=commit;h=107a7bd31ac003e42c0f966aa8e5b26947de6024
 > Commit:     107a7bd31ac003e42c0f966aa8e5b26947de6024
 > Parent:     3be78c73179c9347bdc0a92b2898063bd2300ff7
 > Author:     Theodore Ts'o <tytso@mit.edu>
 > AuthorDate: Fri Aug 16 21:23:41 2013 -0400
 > Committer:  Theodore Ts'o <tytso@mit.edu>
 > CommitDate: Fri Aug 16 21:23:41 2013 -0400
 > 
 >     ext4: cache all of an extent tree's leaf block upon reading
 

 > + * ext4_es_cache_extent() inserts information into the extent status
 > + * tree if and only if there isn't information about the range in
 > + * question already.
 > + */
 > +void ext4_es_cache_extent(struct inode *inode, ext4_lblk_t lblk,
 > +			  ext4_lblk_t len, ext4_fsblk_t pblk,
 > +			  unsigned int status)
 > +{
 > +	struct extent_status *es;
 > +	struct extent_status newes;
 > +	ext4_lblk_t end = lblk + len - 1;
 > +
 > +	newes.es_lblk = lblk;
 > +	newes.es_len = len;
 > +	ext4_es_store_pblock(&newes, pblk);


ext4_es_store_pblock or's the pblk with the existing contents of the struct member.
(albeit masked with ES_MASK)

Should there be a 

	newes.es_pblk = 0;

up there too ?

It seems like if the stack happened to contain any of ES_WRITTEN | ES_UNWRITTEN | ES_DELAYED | ES_HOLE
then it could leak through into the new extent status.

	Dave

Re: ext4: cache all of an extent tree's leaf block upon reading

[Theodore Ts'o]

On Thu, Sep 05, 2013 at 10:37:49AM -0400, Dave Jones wrote:
>  > +void ext4_es_cache_extent(struct inode *inode, ext4_lblk_t lblk,
>  > +			  ext4_lblk_t len, ext4_fsblk_t pblk,
>  > +			  unsigned int status)
>  > +{
>  > +	struct extent_status *es;
>  > +	struct extent_status newes;
>  > +	ext4_lblk_t end = lblk + len - 1;
>  > +
>  > +	newes.es_lblk = lblk;
>  > +	newes.es_len = len;
>  > +	ext4_es_store_pblock(&newes, pblk);
> 
> 
> ext4_es_store_pblock or's the pblk with the existing contents of the struct member.
> (albeit masked with ES_MASK)
> 
> Should there be a 
> 
> 	newes.es_pblk = 0;
> 
> up there too ?

The next line after ext4_es_store_pblock() is:

        ext4_es_store_status(&newes, status);

This will set remaining ES_WRITTEN | ES_UNWRITTEN... bits.

So the only reason to add a line explicitly setting es_pblk to zero
would be to suppress a warning from some insufficiently smart static
code analysis tool.  I didn't see a warning from gcc, but it's
possible that this is something which is causing Coverity or some
other code scanner heartburn.

Cheers,

					- Ted

Re: ext4: cache all of an extent tree's leaf block upon reading

[Dave Jones]

On Thu, Sep 05, 2013 at 10:53:34AM -0400, Theodore Ts'o wrote:

 > > ext4_es_store_pblock or's the pblk with the existing contents of the struct member.
 > > (albeit masked with ES_MASK)
 > > 
 > > Should there be a 
 > > 
 > > 	newes.es_pblk = 0;
 > > 
 > > up there too ?
 > 
 > The next line after ext4_es_store_pblock() is:
 > 
 >         ext4_es_store_status(&newes, status);
 > 
 > This will set remaining ES_WRITTEN | ES_UNWRITTEN... bits.
 > 
 > So the only reason to add a line explicitly setting es_pblk to zero
 > would be to suppress a warning from some insufficiently smart static
 > code analysis tool.  I didn't see a warning from gcc, but it's
 > possible that this is something which is causing Coverity or some
 > other code scanner heartburn.

Yep, that's what picked it up.  I'll add a 'not a bug' annotation to stop
it getting flagged again.  This was the only ext* issue that Coverity
picked up from yesterdays merge btw, which I guess is good news ;)

	Dave

Re: ext4: cache all of an extent tree's leaf block upon reading

[Theodore Ts'o]

On Thu, Sep 05, 2013 at 11:14:57AM -0400, Dave Jones wrote:
>  > So the only reason to add a line explicitly setting es_pblk to zero
>  > would be to suppress a warning from some insufficiently smart static
>  > code analysis tool.  I didn't see a warning from gcc, but it's
>  > possible that this is something which is causing Coverity or some
>  > other code scanner heartburn.
> 
> Yep, that's what picked it up.  I'll add a 'not a bug' annotation to stop
> it getting flagged again.  This was the only ext* issue that Coverity
> picked up from yesterdays merge btw, which I guess is good news ;)

Indeed.

Hmm...  I could add a new inline function
"ext4_es_store_pblock_status()" which sets both parts of the es_pblk
word at once, and which doesn't depend looking at its original value
at all.  I doubt we would never measure a difference in performance,
but in theory it would be more efficient.  And if it eliminates a
potential static code analysis complaint, maybe the two justifications
is good enough to add the new function.

Thanks for checking the coverity results!!

						- Ted