On Fri, Feb 19, 2016 at 08:29:35PM +0900, Kengo NAKAHARA wrote: > Hi, > > On 2016/02/19 19:38, Joerg Sonnenberger wrote: > > On Fri, Feb 19, 2016 at 06:26:14PM +0900, Kengo NAKAHARA wrote: > >> Of course, m_tag_get() is used by packet processing path, this overhead > >> would have a certain influence on packet throughput and latency. > >> So, I think m_tag should use a dedicated pool cache like mbuf to call > >> pool_cache_get() directly. > > > > Which tags are you actually seeing allocated? It would be better to > > remove the tags in first place if they are hot enough... > > I am seeing a struct altq_pktattr tag, which is used to refactor ALTQ > in my working modification. I think ALTQ code would become better than > now by using altq_pktattr tag... > # The altq_pktattr tag is also used by sys/dist/pf/net/pf.c, so I think > # the result affect NetBSD-current itself. > > I agree it should be removed in first place. However, I think there is > several cases which cannot avoid to use m_tag. So, it may be required > to rescue such cases.
We don't have that many random consumers of tags in the tree. For the hot cases, it makes a lot more sense to put a field or two into the mbuf packet header -- if more data is needed, a specific pool is needed anyway. Joerg
