Re: RFC: overheads of using mb_unmapped_to_ext?

On Sun, May 10, 2026 at 02:00:29PM +0800, Zhenlei Huang wrote:
> 
> 
> > On May 10, 2026, at 6:47 AM, Rick Macklem <[email protected]> wrote:
> > 
> > On Sat, May 9, 2026 at 3:43 PM Rick Macklem <[email protected]> wrote:
> >> 
> >> Hi,
> >> 
> >> I have a simple patch that makes the NFS server generate
> >> read replies as a list of m_extpg mbufs. (The code already
> >> knows how to do this for the KTLS, but this patch enables
> >> it for non-TLS use.)
> >> 
> >> I am wondering if the overhead of calling
> >> mb_unmapped_to_ext() on the mbuf chain when the
> >> NIC does not have IFCAP_MEXTPG makes doing
> >> this an undesirable alternative to generating a longer
> >> mbuf chain of mbuf clusters?
> >> 
> >> For my hardware, things just run at wire speed (1Gbps),
> >> so I do not see any difference between the two cases.
> >> I do see a 15% improvement for localhost mounts, but
> >> those hardly matter.
> >> 
> >> I tester did determine that performance was improved for
> >> a Mellanox NIC that has IFCAP_MEXTPG. (5-15% depending
> >> on setup)
> >> 
> >> I can enable this two ways:
> >> 1 - A sysctl that is enabled/disabled by default (which means most
> >>     cases will call mb_unmapped_to_ext() if enabled).
> >> or
> >> 2 - Try and check to see if the outbound NIC has IFCAP_MEXTPG.
> >>    (Note that this does not have to be 100% correct, since the NFS
> >>     reply works either way.  The reviewer was concerned about routing
> >>     changes, etc, affecting the result.)
> >> 
> >> I implemented #2 using the suggested approach, but the
> >> reviewer (who suggested the approach) decided that doing so
> >> was not recommended. (D55203 on phabricator.)
> >> 
> >> So, what do you think is the preferred alternative?
> > I may have partially answered my own question.
> > I did a crude measurement of cpu overheads and found that,
> > for a simple amd64 system with a 1Gbps Intel NIC, that the
> > overheads appear to be about 10% better (aka lower) when
> > M_EXTPG mbufs are used.
> 
> I think the overheads is acceptable on slow NICs.
> 
> For the 1st approach, if a user find performance regression, he / she 
> can just tune sysctl to turn off mb_unmapped_to_ext() . That is much
> simple and robust, and easy to maintain.

Problem starts when user finds a performance regression, but cannot
identify the cause and does not know that the solution for his particular
case is to tweak some knob.

From this PoV, IFCAP_MEXTPG lookup might be more friendly.
> 
> > 
> > So, unless others think there are cases where using M_EXTPG
> > mbufs for read replies is undesirable, I think enabling them by
> > default regardless of the NIC, might be ok.
> > 
> > rick
> > 
> >> 
> >> rick
> > 
> 
> Best regards,
> Zhenlei
>