Re: [PATCH] scsi: be_iscsi: fix possible memory leak and refactor code

[James Bottomley <James.Bottomley@HansenPartnership.com>]

On Mon, 2013-11-18 at 14:18 -0200, Geyslan Gregório Bem wrote:
> 2013/11/18 James Bottomley <James.Bottomley@hansenpartnership.com>:
> > On Sun, 2013-11-17 at 23:12 -0200, Geyslan Gregório Bem wrote:
> >> 2013/11/17 James Bottomley <James.Bottomley@hansenpartnership.com>:
> >> > On Sun, 2013-11-17 at 19:09 -0200, Geyslan Gregório Bem wrote:
> >> >> 2013/11/17 James Bottomley <James.Bottomley@hansenpartnership.com>:
> >> >> > On Sun, 2013-11-17 at 15:51 -0300, Geyslan G. Bem wrote:
> >> >> >> This patch fix memory leakage in cases 'ISCSI_NET_PARAM_VLAN_ID' and
> >> >> >> 'ISCSI_NET_PARAM_VLAN_PRIORITY' and refactors code 'going out' when
> >> >> >> necessary.
> >> >> >
> >> >> > You pointlessly renamed a variable, which makes the diff hard to read.
> >> >> > Please don't do that.
> >> >>
> >> >> Ok, I can agree. 'len' means length? What is returned in case of non
> >> >> error?
> >> >
> >> > it returns the length of buf written to or negative error.
> >> >
> >> >> > You missed the fact that the passed in pointer is unmodified if
> >> >> > mgmt_get_if_info() returns non zero, so the kfree frees junk and would
> >> >> > oops.
> >> >> >
> >> >> > There's no need for a goto; len = -EINVAL; does everything that's
> >> >> > needed.
> >> >>
> >> >> Well, that is a coverity catch. CID 1128954. Check it.
> >> >
> >> > I didn't say coverity was wrong, I said your patch was (well not wrong,
> >> > just over complex and incomplete).  This is the way to fix both
> >> > problems.
> >> >
> >> > James
> >> >
> >> > ---
> >> >
> >> > diff --git a/drivers/scsi/be2iscsi/be_iscsi.c b/drivers/scsi/be2iscsi/be_iscsi.c
> >> > index ffadbee..9dcbdfa 100644
> >> > --- a/drivers/scsi/be2iscsi/be_iscsi.c
> >> > +++ b/drivers/scsi/be2iscsi/be_iscsi.c
> >> > @@ -541,10 +541,8 @@ static int be2iscsi_get_if_param(struct beiscsi_hba *phba,
> >> >                 ip_type = BE2_IPV6;
> >>
> >> James, this approach will not prevent the leakage.
> >
> > I don't see why not.  The -EINVAL case goes through the kfree() now too,
> > no?
> 
> I'm refering to the removal of kfree in your suggestion.

That's the second bug I pointed out via code inspection.  If the
function returns an error (any non zero return) then the pointer isn't
altered, so we return without the free.  It's a standard error pattern.

> >
> >>  We can initialize the if_info with NULL and always kfree it without
> >> to care about junk.
> >
> > Why?  Error return means no allocation.
> Setting if_info to NULL allow to kfree without concerns.
> 
> Eg.:
> 
> - struct be_cmd_get_if_info_resp *if_info;
> + struct be_cmd_get_if_info_resp *if_info = NULL;
> 
> ...
> 
> +       if (len)
> +               goto out;
> 
> ...
> 
> -               if (if_info->vlan_priority == BEISCSI_VLAN_DISABLE)
> -                       return -EINVAL;
> +               if (if_info->vlan_priority == BEISCSI_VLAN_DISABLE) {
> +                       len = -EINVAL;
> +                       goto out;
> +               }

What's the point of that?  Just removing the goto out; has the code
going to the same place because of the break below.

James