Re: [REGRESSION][BISECTED] tun/tap & vhost-net: multi-threaded network performance

[Simon Schippers]

On 7/2/26 09:42, Michael S. Tsirkin wrote:
> On Thu, Jul 02, 2026 at 09:24:59AM +0200, Simon Schippers wrote:
>> On 7/1/26 22:56, Michael S. Tsirkin wrote:
>>> On Wed, Jul 01, 2026 at 09:16:48PM +0200, Brett Sheffield wrote:
>>>> TL;DR - Commit 1d6e569b7d0c0b2736636749e4be0a27f3cefcb3 causes
>>>> significant performance regressions with TAP interfaces and multithreaded
>>>> network code. Please revert.
>>>>
>>>>
>>>> Librecast is an IPv6 multicast library. One of the tests (0055) fails under
>>>> Linux 7.2-rc1. The test performs data synchronization over IPv6 multicast using a TAP
>>>> interface. This test has run successfully on every stable, LTS and mainline RC
>>>> released in the past year. Every kernel with my Tested-by has run this test.
>>>>
>>>> There have been a bunch of changes to MLDv2 so I started bisecting there, but
>>>> the culprit is actually 1d6e569b7d0c0b2736636749e4be0a27f3cefcb3 "tun/tap &
>>>> vhost-net: avoid ptr_ring tail-drop when a qdisc is present"
>>>>
>>>> Reverting this commit fixes the test.
>>>>
>>>> To eliminate my code and any multicast weirdness, I ran tests with iperf3
>>>> comparing the same host running 7.2-rc1 both with and without 1d6e569b7d0
>>>> reverted.
>>
>> Thank you very much for your bisect!
>>
>> As the author, I am sorry for that regression!
>>
>>>
>>> Thanks a lot for the bisect! Reverting is not out of question, but
>>> just before we do, it is worth analyzing the situation.
>>>
>>> Could you pls tell us
>>> - do you see packet drops?
>>
>> iperf3 shows no TCP retransmissions, so there were never packet drops
>> when the patch was enabled.
>> It is the number after the sender data rate (example: threads 1,
>> reverted has 368 retransmissions/drops).
>>
>>> - does it help to increase the tun queue size?
>>
>> I agree, this would be great to know.
>>
>> However, even then we must act. I am considering IFF_BACKPRESSURE
>> as a feature flag, defaulting to off. It would just enable/disable
>> the stopping logic in tun_net_xmit() and the waking logic
>> in __tun_wake_queue(). If disabled, it would result in the same logic
>> as before.
>>
>> I could provide such a patch as [net] material.
>>
>> Thanks again!
> 
> Or BQL? I quickly wrote a prototype of that and it seems to work well -
> could you help test maybe?

BQL won't fix it I think.
A bigger TUN queue would probably fix it but BQL can only just adjust
to a smaller queue than 500 packets.

Unless you don't take the ptr_ring size as an upper limit for BQL and
resize the ptr_ring on the fly? I don't think this is viable.

I am currently working with others to get BQL working for veth.
TLDR: The issue with software interfaces is that there is no *periodic*
completion process [2]. For hardware there is.
I applied the same veth approach for TUN and it showed to fix
bufferbloat... But again won't fix the issue here I think.

[1] Link: https://lore.kernel.org/netdev/20260612083530.1650245-1-hawk@xxxxxxxxxx/T/#u
[2] Link: https://lwn.net/Articles/469651/

> 
> 
>>>
>>> Thanks a lot!
>>>
>>>
>>>> CPU: AMD Ryzen 9 9950X

This processor has 2 CCDs. This probably makes the issue worse then it
was for me. My (older) Ryzen 5 5600X only has a single CCD.

Thanks.

>>>>
>>>> [ host ] - [ bridge ] - [ tap ] - [ guest (qemu) ]
>>>>
>>>> Running matching kernels on host and guest, I started iperf3 in server mode on
>>>> the guest and tested from the host so traffic passes through the tap interface.
>>>>
>>>> iperf3 -s -V                 # server
>>>> iperf3 -c guest -P nthreads  # client
>>>>
>>>> 7.2.0-rc1 (threads 1):
>>>>
>>>> [  5]   0.00-10.00  sec  20.2 GBytes  17.4 Gbits/sec    0            sender
>>>> [  5]   0.00-10.00  sec  2.00 GBytes  1.72 Gbits/sec                  receiver
>>>>
>>>> 7.2.0-rc1 (threads 1, reverted):
>>>>
>>>> [  5]   0.00-10.00  sec  15.3 GBytes  13.1 Gbits/sec  368            sender
>>>> [  5]   0.00-10.00  sec  2.00 GBytes  1.72 Gbits/sec                  receiver
>>>>
>>>> 7.2.0-rc1 (threads 2):
>>>>
>>>> [SUM]   0.00-10.00  sec  10.9 GBytes  9.33 Gbits/sec    0             sender
>>>> [SUM]   0.00-10.00  sec  4.00 GBytes  3.43 Gbits/sec                  receiver
>>>>
>>>> 7.2.0-rc1 (threads 2, reverted):
>>>>
>>>> [SUM]   0.00-10.00  sec  15.9 GBytes  13.7 Gbits/sec  1567             sender
>>>> [SUM]   0.00-10.00  sec  4.00 GBytes  3.43 Gbits/sec                  receiver
>>>>
>>>> 7.2.0-rc1 (threads 4):
>>>>
>>>> [SUM]   0.00-10.00  sec  10.9 GBytes  9.33 Gbits/sec    0             sender
>>>> [SUM]   0.00-10.00  sec  8.00 GBytes  6.87 Gbits/sec                  receiver
>>>>
>>>> 7.2.0-rc1 (threads 4, reverted):
>>>>
>>>> [SUM]   0.00-10.00  sec  16.5 GBytes  14.1 Gbits/sec  6701             sender
>>>> [SUM]   0.00-10.00  sec  8.00 GBytes  6.87 Gbits/sec                  receiver
>>>>
>>>> 7.2.0-rc1 (threads 8):
>>>>
>>>> [SUM]   0.00-10.00  sec  10.7 GBytes  9.15 Gbits/sec    0             sender
>>>> [SUM]   0.00-10.01  sec  10.6 GBytes  9.13 Gbits/sec                  receiver
>>>>
>>>> 7.2.0-rc1 (threads 8, reverted):
>>>>
>>>> [SUM]   0.00-10.00  sec  16.2 GBytes  14.0 Gbits/sec  19319             sender
>>>> [SUM]   0.00-10.00  sec  15.7 GBytes  13.5 Gbits/sec                  receiver
>>>>
>>>> 7.2.0-rc1 (threads 16):
>>>>
>>>> [SUM]   0.00-10.00  sec  10.9 GBytes  9.35 Gbits/sec    0             sender
>>>> [SUM]   0.00-10.01  sec  10.9 GBytes  9.32 Gbits/sec                  receiver
>>>>
>>>> 7.2.0-rc1 (threads 16, reverted):
>>>>
>>>> [SUM]   0.00-10.00  sec  14.4 GBytes  12.4 Gbits/sec  43593             sender
>>>> [SUM]   0.00-10.00  sec  14.4 GBytes  12.4 Gbits/sec                  receiver
>>>>
>>>>
>>>> As you can see, the new code works for single threaded, but for all other cases
>>>> there's a significant performance drop. I see this trade-off is mentioned in the
>>>> commit, but the performance drop off is much worse than suggested with the
>>>> current patch.
>>>>
>>>> In our multicast use case data is sent by multiple threads to multiple groups
>>>> simultaneously, this just breaks things to the extent that a <2 second test
>>>> times out after 5 minutes.
>>>>
>>>>
>>>> git bisect start
>>>> # status: waiting for both good and bad commits
>>>> # bad: [dc59e4fea9d83f03bad6bddf3fa2e52491777482] Linux 7.2-rc1
>>>> git bisect bad dc59e4fea9d83f03bad6bddf3fa2e52491777482
>>>> # status: waiting for good commit(s), bad commit known
>>>> # good: [36bdc0e815b4e8a05b9028d8ef8a25e1ead35cc1] net: usb: asix: ax88772: re-add usbnet_link_change() in phylink callbacks
>>>> git bisect good 36bdc0e815b4e8a05b9028d8ef8a25e1ead35cc1
>>>> # good: [db314398f618a3a23315f73c87f7d318eaf06c1b] Merge branch 'net-bridge-mcast-support-exponential-field-encoding'
>>>> git bisect good db314398f618a3a23315f73c87f7d318eaf06c1b
>>>> # bad: [079a028d6327e68cfa5d38b36123637b321c19a7] string: Remove strncpy() from the kernel
>>>> git bisect bad 079a028d6327e68cfa5d38b36123637b321c19a7
>>>> # bad: [f396f4005180928cd9e15e352a6512865d3bc908] Bluetooth: btmtk: fix URB leak in alloc_mtk_intr_urb error path
>>>> git bisect bad f396f4005180928cd9e15e352a6512865d3bc908
>>>> # bad: [ec1806a730a1c0b3d68a7f9afe81514fb0dd7991] netfilter: x_tables: disable 32bit compat interface in user namespaces
>>>> git bisect bad ec1806a730a1c0b3d68a7f9afe81514fb0dd7991
>>>> # good: [50c2d91c5dfa0e465826ec1f8dbad9cdc254bd85] mptcp: do not drop partial packets
>>>> git bisect good 50c2d91c5dfa0e465826ec1f8dbad9cdc254bd85
>>>> # good: [68993ced0f618e36cf33388f1e50223e5e6e78cc] Merge tag 'net-7.1-rc5' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net
>>>> git bisect good 68993ced0f618e36cf33388f1e50223e5e6e78cc
>>>> # good: [34c78dff59a25110a4ce50c208e42a91490fe615] Merge branch 'net-use-ip_outnoroutes-drop-reason'
>>>> git bisect good 34c78dff59a25110a4ce50c208e42a91490fe615
>>>> # bad: [9587ed8137fb83d93f84b858337412f4500b21e9] Merge branch 'gve-add-support-for-ptp-gettimex64'
>>>> git bisect bad 9587ed8137fb83d93f84b858337412f4500b21e9
>>>> # bad: [83ea7fd73b11dd8cbf4416507a5eac3890b49fb0] net: dsa: microchip: remove unused phylink_mac_link_up() callback
>>>> git bisect bad 83ea7fd73b11dd8cbf4416507a5eac3890b49fb0
>>>> # bad: [f0de88303d5e7e04a1224bc7a00512b5a1c4fe7a] net: make is_skb_wmem() available to modules
>>>> git bisect bad f0de88303d5e7e04a1224bc7a00512b5a1c4fe7a
>>>> # bad: [c411baa463e85a779a7e68a00ba6298770b58c4c] netconsole: move push_ipv6() from netpoll
>>>> git bisect bad c411baa463e85a779a7e68a00ba6298770b58c4c
>>>> # good: [fba362c17d9d9211fc51f272156bb84fc23bdf98] ptr_ring: move free-space check into separate helper
>>>> git bisect good fba362c17d9d9211fc51f272156bb84fc23bdf98
>>>> # bad: [d0273dbe8be1640e597552f81faf1d6c9997d3e3] ipvlan: use netif_receive_skb() in ipvlan_process_multicast()
>>>> git bisect bad d0273dbe8be1640e597552f81faf1d6c9997d3e3
>>>> # bad: [3803065cd6b0630d4161d86aa04e2d1db0f3a0b5] Merge branch 'tun-tap-vhost-net-apply-qdisc-backpressure-on-full-ptr_ring-to-reduce-tx-drops'
>>>> git bisect bad 3803065cd6b0630d4161d86aa04e2d1db0f3a0b5
>>>> # bad: [1d6e569b7d0c0b2736636749e4be0a27f3cefcb3] tun/tap & vhost-net: avoid ptr_ring tail-drop when a qdisc is present
>>>> git bisect bad 1d6e569b7d0c0b2736636749e4be0a27f3cefcb3
>>>> # first bad commit: [1d6e569b7d0c0b2736636749e4be0a27f3cefcb3] tun/tap & vhost-net: avoid ptr_ring tail-drop when a qdisc is present
>>>>
>>>> -- 
>>>> Brett Sheffield (he/him)
>>>> Librecast - Decentralising the Internet with Multicast
>>>> https://librecast.net/
>>>> https://blog.brettsheffield.com/
>>>
>