On Wed, Dec 3, 2025 at 5:40 PM Magnus Karlsson
[off-list ref] wrote:

On Wed, 3 Dec 2025 at 10:25, Paolo Abeni [off-list ref] wrote:

quoted

On 12/3/25 7:56 AM, Jason Xing wrote:

quoted

On Sat, Nov 29, 2025 at 8:55 AM Jason Xing [off-list ref] wrote:

quoted

On Fri, Nov 28, 2025 at 10:20 PM Paolo Abeni [off-list ref] wrote:

quoted

On 11/28/25 2:46 PM, Jason Xing wrote:

quoted

From: Jason Xing <kernelxing@tencent.com>

Use atomic_try_cmpxchg operations to replace spin lock. Technically
CAS (Compare And Swap) is better than a coarse way like spin-lock
especially when we only need to perform a few simple operations.
Similar idea can also be found in the recent commit 100dfa74cad9
("net: dev_queue_xmit() llist adoption") that implements the lockless
logic with the help of try_cmpxchg.

Signed-off-by: Jason Xing <kernelxing@tencent.com>
---
Paolo, sorry that I didn't try to move the lock to struct xsk_queue
because after investigation I reckon try_cmpxchg can add less overhead
when multiple xsks contend at this point. So I hope this approach
can be adopted.

I still think that moving the lock would be preferable, because it makes
sense also from a maintenance perspective.

I can see your point here. Sure, moving the lock is relatively easier
to understand. But my take is that atomic changes here are not that
hard to read :) It has the same effect as spin lock because it will
atomically check, compare and set in try_cmpxchg().

quoted

Can you report the difference
you measure atomics vs moving the spin lock?

No problem, hopefully I will give a detailed report next week because
I'm going to apply it directly in production where we have multiple
xsk sharing the same umem.

I'm done with the test in production where a few applications rely on
multiple xsks sharing the same pool to send UDP packets. Here are
significant numbers from bcc tool that recorded the latency caused by
these particular functions:

1. use spin lock
$ sudo ./funclatency xsk_cq_reserve_locked
Tracing 1 functions for "xsk_cq_reserve_locked"... Hit Ctrl-C to end.
^C
     nsecs               : count     distribution
         0 -> 1          : 0        |                                        |
         2 -> 3          : 0        |                                        |
         4 -> 7          : 0        |                                        |
         8 -> 15         : 0        |                                        |
        16 -> 31         : 0        |                                        |
        32 -> 63         : 0        |                                        |
        64 -> 127        : 0        |                                        |
       128 -> 255        : 25308114 |**                                      |
       256 -> 511        : 283924647 |**********************                  |
       512 -> 1023       : 501589652 |****************************************|
      1024 -> 2047       : 93045664 |*******                                 |
      2048 -> 4095       : 746395   |                                        |
      4096 -> 8191       : 424053   |                                        |
      8192 -> 16383      : 1041     |                                        |
     16384 -> 32767      : 0        |                                        |
     32768 -> 65535      : 0        |                                        |
     65536 -> 131071     : 0        |                                        |
    131072 -> 262143     : 0        |                                        |
    262144 -> 524287     : 0        |                                        |
    524288 -> 1048575    : 6        |                                        |
   1048576 -> 2097151    : 2        |                                        |

avg = 664 nsecs, total: 601186432273 nsecs, count: 905039574

2. use atomic
$ sudo ./funclatency xsk_cq_cached_prod_reserve
Tracing 1 functions for "xsk_cq_cached_prod_reserve"... Hit Ctrl-C to end.
^C
     nsecs               : count     distribution
         0 -> 1          : 0        |                                        |
         2 -> 3          : 0        |                                        |
         4 -> 7          : 0        |                                        |
         8 -> 15         : 0        |                                        |
        16 -> 31         : 0        |                                        |
        32 -> 63         : 0        |                                        |
        64 -> 127        : 0        |                                        |
       128 -> 255        : 109815401 |*********                               |
       256 -> 511        : 485028947 |****************************************|
       512 -> 1023       : 320121627 |**************************              |
      1024 -> 2047       : 38538584 |***                                     |
      2048 -> 4095       : 377026   |                                        |
      4096 -> 8191       : 340961   |                                        |
      8192 -> 16383      : 549      |                                        |
     16384 -> 32767      : 0        |                                        |
     32768 -> 65535      : 0        |                                        |
     65536 -> 131071     : 0        |                                        |
    131072 -> 262143     : 0        |                                        |
    262144 -> 524287     : 0        |                                        |
    524288 -> 1048575    : 10       |                                        |

avg = 496 nsecs, total: 473682265261 nsecs, count: 954223105

And those numbers were verified over and over again which means they
are quite stable.

You can see that when using atomic, the avg is smaller and the count
of [128 -> 255] is larger, which shows better performance.

I will add the above numbers in the commit log after the merge window is open.

It's not just a matter of performance. Spinlock additionally give you
fairness and lockdep guarantees, beyond being easier to graps for
however is going to touch this code in the future, while raw atomic none
of them.

Right.

quoted

From a maintainability perspective spinlocks are much more preferable.

IMHO micro-benchmarking is not a strong enough argument to counter the
spinlock adavantages: at very _least_ large performance gain should be
observed in relevant test-cases and/or real live workloads.

The problem is that I have no good benchmark to see the minor
improvement because it requires multiple xsks. Xdpsock has bugs and
doesn't allow two xsks running in parallel. Unlike one xsk scenario,
it's really difficult for me currently to measure the improvement. So
I resorted to observing latency.

With that said, I will follow your suggestion to move that lock :)

quoted

quoted

quoted

quoted

Have you tried moving cq_prod_lock, too?

Not yet, thanks for reminding me. It should not affect the sending
rate but the tx completion time, I think.

I also tried moving this lock, but sadly I noticed that in completion
time the lock was set which led to invalidation of the cache line of
another thread sending packets. It can be obviously proved by perf
cycles:ppp:
1. before
8.70% xsk_cq_cached_prod_reserve

2. after
12.31% xsk_cq_cached_prod_reserve

So I decided not to bring such a modification. Anyway, thanks for your
valuable suggestions and I learnt a lot from those interesting
experiments.

The goal of such change would be reducing the number of touched
cachelines;

Yep.

quoted

when I suggested the above, I did not dive into the producer
specifics, I assumed the relevant producer data were inside the
xsk_queue struct.

It looks like the data is actually inside 'struct xdp_ring', so the
producer lock should be moved there, specifically:

struct xdp_ring {
        u32 producer ____cacheline_aligned_in_smp;
        spinlock_t producer_lock;
        // ...

This struct is reflected to user space, so we should not put the spin
lock there.

Agree on this point.

But you could put it in struct xsk_queue, but perhaps you

My concern is that only cq uses this lock while this structure is used
by all types of queues.

would want to call it something more generic as there would be a lock
present in all queues/rings, though you would only use it for the cq.
Some of the members in xsk_queue are nearly always used when
manipulating the ring, so the cache line should be hot.

I am just thinking aloud if this would be correct. There is one pool
per cq. When a pool is shared, the cq belonging to that pool is also
always shared, so I think that would be correct moving the lock from
the pool to the cq.

Exactly, that is how Paolo suggested previously.

quoted

I'm a bit lost in the structs indirection, but I think the above would
be beneficial even for the ZC path.

Spot on, that will be part of my future plan. Thanks!

Thanks,
Jason