Re: [PATCH] [v2] nvdimm-btt: fix several memleaks

[Ira Weiny <ira.weiny@intel.com>]

dinghao.liu@ wrote:
> > Ira Weiny wrote:
> > > Dinghao Liu wrote:
> > 
> > [snip]
> > 
> > -static int btt_freelist_init(struct arena_info *arena)
> > +static int btt_freelist_init(struct device *dev, struct arena_info *arena)
> > 
> > Both struct arena_info and struct btt contain references to struct nd_btt
> > which is the device you are passing down this call stack.
> > 
> > Just use the device in the arena/btt rather than passing a device
> > structure.  That makes the code easier to read and ensures that the device
> > associated with this arena or btt is used.
> 
> Thanks for this suggestion! I will fix this in the v3 patch.
> 
> > [snip]
> > > >  
> > > > -static int btt_maplocks_init(struct arena_info *arena)
> > > > +static int btt_maplocks_init(struct device *dev, struct arena_info *arena)
> > > >  {
> > > >  	u32 i;
> > > >  
> > > > -	arena->map_locks = kcalloc(arena->nfree, sizeof(struct aligned_lock),
> > > > +	arena->map_locks = devm_kcalloc(dev, arena->nfree, sizeof(struct aligned_lock),
> > > >  				GFP_KERNEL);
> > > >  	if (!arena->map_locks)
> > > >  		return -ENOMEM;
> > > > @@ -805,9 +805,6 @@ static void free_arenas(struct btt *btt)
> > > >  
> > > >  	list_for_each_entry_safe(arena, next, &btt->arena_list, list) {
> > > >  		list_del(&arena->list);
> > > > -		kfree(arena->rtt);
> > > > -		kfree(arena->map_locks);
> > > > -		kfree(arena->freelist);
> > > 
> > > This does not quite work.
> > > 
> > > free_arenas() is used in the error paths of create_arenas() and
> > > discover_arenas().  In those cases devm_kfree() is probably a better way
> > > to clean up this.
> 
> Here I'm a little confused about when devm_kfree() should be used.

Over all it should be used whenever memory is allocated for the lifetime
of the device.

> Code in btt_init() implies that resources allocated by devm_* could be
> auto freed in both error and success paths of probe/attach (e.g., btt 
> allocated by devm_kzalloc is never freed by devm_kfree).
> Using devm_kfree() in free_arenas() is ok for me, but I want to know
> whether not using devm_kfree() constitutes a bug.

Unfortunately I'm not familiar enough with this code to know for sure.

After my quick checks before I thought it was.  But each time I look at it
I get confused.  This is why I was thinking maybe not using devm_*() and
using no_free_ptr() may be a better solution to make sure things don't
leak without changing the success path (which is likely working fine
because no bugs have been found.)

> 
> > > 
> > > However...
> > > 
> > > >  		debugfs_remove_recursive(arena->debugfs_dir);
> > > >  		kfree(arena);
> > > 
> > > Why can't arena be allocated with devm_*?
> > > 
> > > We need to change this up a bit more to handle the error path vs regular
> > > device shutdown free (automatic devm frees).
> > 
> 
> At first, I think the use of arena is correct. Therefore, allocating arena
> with devm_* should be a code style optimization. However, I rechecked
> discover_arenas and found arena might also be leaked (e.g., if
> alloc_arena() fails in the second loop, the previously allocated
> resources in the loop is leaked). The correct code could be found in
> create_arenas(), where free_arenas is called on failure of
> alloc_arena().
 
Yea I've not reached that level of detail in my analysis.

>
> To fix this issue, I think it's better to change the 'goto out_super'
> tag to 'goto out'. We could also use devm_* for arena to simplify the
> error path in discover_arenas(). 

I think it is your call at this point as I don't have time to dig in more
than I have.  Sorry.

> 
> > We might want to look at using no_free_ptr() in this code.  See this
> > patch[1] for an example of how to inhibit the cleanup and pass the
> > pointer on when the function succeeds.
> > 
> > [1]
> > https://lore.kernel.org/all/170261791914.1714654.6447680285357545638.stgit@dwillia2-xfh.jf.intel.com/
> > 
> > Ira
> 
> Thanks for this example. But it seems that no_free_ptr() is used to
> handle the scope based resource management. Changes in this patch does
> not introduce this feature. Do I understand this correctly?

You do understand but I was thinking that perhaps using no_free_ptr()
rather than devm_*() might be an easier way to fix this bug without trying
to decode the lifetime of everything.

Ira

> 
> Regards,
> Dinghao