On Mon, Sep 22, 2014 at 10:28:02PM +0300, Yuval Shaia wrote:
> On Tue, Sep 16, 2014 at 09:47:43AM +0300, Or Gerlitz wrote:
> >
> > On 9/15/2014 9:55 PM, Yuval Shaia wrote:
> > >On Mon, Sep 15, 2014 at 05:47:19PM +0300, Or Gerlitz wrote:
> > >>>On Sun, Sep 14, 2014 at 9:46 PM, Yuval Shaia<yuval.sh...@oracle.com>
> > >>>wrote:
> > >>>>>By default, IPoIB-CM driver uses 64k MTU. Larger MTU gives better
> > >>>>>performance.
> > >>>>>This MTU plus overhead puts the memory allocation for IP based packets
> > >>>>>at 32 4k pages (order 5),
> > >>
> > >>>So if we make sure that the advertized netdevice MTU is 64K minus that
> > >>>over head we're back to order four
> > >>>allocation and problem is solved? note that RFC 4755 makes sure that
> > >>>the MTU is negotiated in both directions,
> > >>>so it can have any value, specifically 64K - that epsilon which will
> > >>>hopefully make you happy
> > >Interesting point. But please note that in any case, when not using
> > >scatter/gather we force the allocation of large contiguous physical memory.
> >
> > On the post you wrote "[...] resolve the issue by removing the physically
> > contiguous memory requirement using Scatter/Gather feature that exists in
> > Linux".
> >
> > I assume you refer to NETIF_F_SG, right? so your claim is that Linux will
> > not effectively use the driver ability to serve SG skbs unless the driver
> > also advertizes (say) NETIF_F_IP_CSUM?!
> >
> > I thought it's the other way around -- that is supporting checksum
> > offloading is useless unless SG is supported. Can you provide pointer into
> > the network stack code/documentation that supports your claim?
> > Or.
> While porting the patch to latest kernel i have learned that this limitation
> is no longer exist.
> The limitation i was talking about was in older versions of the function
> netdev_fix_features().
> /* Fix illegal SG+CSUM combinations. */
> if ((features & NETIF_F_SG) &&
> !(features & NETIF_F_ALL_CSUM)) {
> netdev_dbg(dev,
> "Dropping NETIF_F_SG since no checksum feature.\n");
> features &= ~NETIF_F_SG;
> }
> Now, with the removal of this limitation the argument of "...In order to use
> SG....IPoIB-CM must support IP checksum offload" is no longer applicable.
> We now left with the performance benefit argument only, can we ignore the
> performance improvement achieved by eliminating the need to calculate
> checksum in SW?
Another benefit we should take into account is that that IB-CRC is **much
more** reliable than checksum!!
> >
> >
> >
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html