Re: [PATCH v2] net: mana: Use page pool fragments for RX buffers instead of full pages to improve memory efficiency.

[Chris Arges <carges@cloudflare.com>]

On 2025-07-31 18:36:04, Jesper Dangaard Brouer wrote:
> 
> 
> On 30/07/2025 09.31, Dragos Tatulea wrote:
> > On Tue, Jul 29, 2025 at 01:20:07PM -0700, Dipayaan Roy wrote:
> > > On Tue, Jul 29, 2025 at 12:15:23PM +0200, Jesper Dangaard Brouer wrote:
> > > > 
> > > > 
> > > > On 23/07/2025 21.07, Dipayaan Roy wrote:
> > > > > This patch enhances RX buffer handling in the mana driver by allocating
> > > > > pages from a page pool and slicing them into MTU-sized fragments, rather
> > > > > than dedicating a full page per packet. This approach is especially
> > > > > beneficial on systems with large page sizes like 64KB.
> > > > > 
> > > > > Key improvements:
> > > > > 
> > > > > - Proper integration of page pool for RX buffer allocations.
> > > > > - MTU-sized buffer slicing to improve memory utilization.
> > > > > - Reduce overall per Rx queue memory footprint.
> > > > > - Automatic fallback to full-page buffers when:
> > > > >     * Jumbo frames are enabled (MTU > PAGE_SIZE / 2).
> > > > >     * The XDP path is active, to avoid complexities with fragment reuse.
> > > > > - Removal of redundant pre-allocated RX buffers used in scenarios like MTU
> > > > >    changes, ensuring consistency in RX buffer allocation.
> > > > > 
> > > > > Testing on VMs with 64KB pages shows around 200% throughput improvement.
> > > > > Memory efficiency is significantly improved due to reduced wastage in page
> > > > > allocations. Example: We are now able to fit 35 rx buffers in a single 64kb
> > > > > page for MTU size of 1500, instead of 1 rx buffer per page previously.
> > > > > 
> > > > > Tested:
> > > > > 
> > > > > - iperf3, iperf2, and nttcp benchmarks.
> > > > > - Jumbo frames with MTU 9000.
> > > > > - Native XDP programs (XDP_PASS, XDP_DROP, XDP_TX, XDP_REDIRECT) for
> > > > >    testing the XDP path in driver.
> > > > > - Page leak detection (kmemleak).
> > > > > - Driver load/unload, reboot, and stress scenarios.
> > > > 
> > > > Chris (Cc) discovered a crash/bug[1] with page pool fragments used
> > > > from the mlx5 driver.
> > > > He put together a BPF program that reproduces the issue here:
> > > > - [2] https://github.com/arges/xdp-redirector
> > > > 
> > > > Can I ask you to test that your driver against this reproducer?
> > > > 
> > > > 
> > > > [1] https://lore.kernel.org/all/aIEuZy6fUj_4wtQ6@861G6M3/
> > > > 
> > > > --Jesper
> > > > 
> > > 
> > > Hi Jesper,
> > > 
> > > I was unable to reproduce this issue on mana driver.
> > > 
> > Please note that I had to make a few adjustments to get reprodduction on
> > mlx5:
> > 
> > - Make sure that the veth MACs are recognized by the device. Otherwise
> >    traffic might be dropped by the device.
> > 
> > - Enable GRO on the veth device. Otherwise packets get dropped before
> >    they reach the devmap BPF program.
> > 
> > Try starting the test program with one thread and see if you see packets
> > coming through veth1-ns1 end of the veth pair.
> > 
> 
> Hi Dipayaan,
> 
> Enabling GRO on the veth device is quite important for the test to be valid.
> 
> I've asked Chris to fix this in the reproducer. He can report back when
> he have done this, so you can re-run the test.  It is also good advice
> from Dragos that you should check packets are coming through the veth
> pair, to make sure the test is working.
> 
> The setup.sh script also need to be modified, as it is loading xdp on a
> net_device called "ext0" [0], which is specific to our systems (which
> default also have GRO enabled for veth).
> 
> [0] https://github.com/arges/xdp-redirector/blob/main/setup.sh#L28
> 
> --Jesper


I pushed some updates to the setup script to make it easier to use. If you have
issues running the script, please share the output.

--chris