M_NOFREE removal (was Re: svn commit: r254520 - in head/sys: kern sys)

[Andre Oppermann]

On 21.08.2013 22:52, Navdeep Parhar wrote:
> On 08/21/13 13:44, Andre Oppermann wrote:
>> On 21.08.2013 21:40, Navdeep Parhar wrote:
>>> On 08/21/13 12:22, Andre Oppermann wrote:
>>>> On 21.08.2013 20:23, Navdeep Parhar wrote:
>>>>> I believe we need an extra patch to get M_NOFREE correct.  I've had it
>>>>> forever in some of my internal repos but never committed it upstream
>>>>> (just plain forgot).  Since this stuff is fresh in your mind, can you
>>>>> review this:
>>>>>
>>>>> diff -r cd78031b7885 sys/sys/mbuf.h
>>>>> --- a/sys/sys/mbuf.h    Fri Aug 16 13:35:26 2013 -0700
>>>>> +++ b/sys/sys/mbuf.h    Wed Aug 21 10:55:57 2013 -0700
>>>>> @@ -535,6 +535,8 @@ m_free(struct mbuf *m)
>>>>>     {
>>>>>         struct mbuf *n = m->m_next;
>>>>>
>>>>> +    if ((m->m_flags & (M_PKTHDR|M_NOFREE)) == (M_PKTHDR|M_NOFREE))
>>>>> +        m_tag_delete_chain(m, NULL);
>>>>>         if (m->m_flags & M_EXT)
>>>>>             mb_free_ext(m);
>>>>>         else if ((m->m_flags & M_NOFREE) == 0)
>>>>>
>>>>> It prevents leaks of tags from M_NOFREE+pkthdr mbufs.
>>>>
>>>> Good point.  Looks correct.
>>>>
>>>> But then I wonder if it is really a smart thing to allow single
>>>> mbufs (w/o M_EXT) to be M_NOFREE at the same time.  They easily
>>>> get lost.  If it doesn't have an external buffer attached there
>>>> is no way to know when and if it was freed.
>>>>
>>>> If M_NOFREE is only allowed together with M_EXT then the tag chain
>>>> delete should happen in mb_ext_free() next to 'skipmbuf' instead.
>>>>
>>>> Index: kern/uipc_mbuf.c
>>>> ===================================================================
>>>> --- kern/uipc_mbuf.c    (revision 254605)
>>>> +++ kern/uipc_mbuf.c    (working copy)
>>>> @@ -283,11 +283,6 @@
>>>>           KASSERT((m->m_flags & M_EXT) == M_EXT, ("%s: M_EXT not set",
>>>> __func__));
>>>>           KASSERT(m->m_ext.ref_cnt != NULL, ("%s: ref_cnt not set",
>>>> __func__));
>>>>
>>>> -       /*
>>>> -        * check if the header is embedded in the cluster
>>>> -        */
>>>> -       skipmbuf = (m->m_flags & M_NOFREE);
>>>> -
>>>>           /* Free attached storage if this mbuf is the only reference to
>>>> it. */
>>>>           if (*(m->m_ext.ref_cnt) == 1 ||
>>>>               atomic_fetchadd_int(m->m_ext.ref_cnt, -1) == 1) {
>>>> @@ -328,8 +323,14 @@
>>>>                                   ("%s: unknown ext_type", __func__));
>>>>                   }
>>>>           }
>>>> -       if (skipmbuf)
>>>> +
>>>> +       /*
>>>> +        * Do not free if the mbuf is embedded in the cluster.
>>>> +        */
>>>> +       if (m->m_flags & M_NOFREE) {
>>>> +               m_tag_delete_chain(m, NULL);
>>>>                   return;
>>>> +       }
>>>
>>> The problem with this is that the mbuf may already be gone if it was
>>> embedded in the cluster that was just freed by the ext free.  That's why
>>> skipmbuf is used to cache the M_NOFREE bit.
>>
>> Next try: ;)
>
> It is most flexible to let M_NOFREE work without any assumptions
> attached (must be M_EXT, etc.)  So I still prefer my patch to this.  If
> you don't have any strong preferences may I commit that one instead?

I don't mind having your patch.  Though I don't see how it possibly can't leak
mbufs if they are not M_EXT.

-- 
Andre