Skip to contentSkip to conversation

xsk: introduce atomic for cq in generic path

PATCH Superseded Jason Xing 2 participants 10 messages Nov 25, 2025 – Nov 27, 2025

Subsystems (4)

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Jason Xing

author
#0 lore raw

v2 cover xsk: introduce atomic for cq in generic path

From: Jason Xing <[email protected]>

In the hot path (that is __xsk_generic_xmit()), playing with spin lock is time consuming. So this series replaces spin lock with atomic operations to get better performance.

---
V2

1. use separate functions rather than branches within shared routines. (Maciej)
2. make each patch as simple as possible for easier review

Jason Xing (3):
xsk: add atomic cached_prod for copy mode
xsk: use atomic operations around cached_prod for copy mode
xsk: remove spin lock protection of cached_prod

 include/net/xsk_buff_pool.h |  5 -----
 net/xdp/xsk.c               | 23 +++++------------------
 net/xdp/xsk_buff_pool.c     |  1 -
 net/xdp/xsk_queue.h         | 27 +++++++++++++++++++++++----
 4 files changed, 28 insertions(+), 28 deletions(-)
Signature · 1 line
2.41.3
author
#1 lore raw

PATCH v2 1/3 xsk: add atomic cached_prod for copy mode

From: Jason Xing <[email protected]>

Add a union member for completion queue only in copy mode for now. The purpose is to replace the cq_cached_prod_lock with atomic operation to improve performance. Note that completion queue in zerocopy mode doesn't need to be converted because the whole process is lockless.

---
 net/xdp/xsk_queue.h | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)
net/xdp/xsk_queue.h +5 -1
index 1eb8d9f8b104..44cc01555c0b 100644
@@ -40,7 +40,11 @@ struct xdp_umem_ring {
 struct xsk_queue {
 	u32 ring_mask;
 	u32 nentries;
-	u32 cached_prod;
+	union {
+		u32 cached_prod;
+		/* Used for cq in copy mode only */
+		atomic_t cached_prod_atomic;
+	};
 	u32 cached_cons;
 	struct xdp_ring *ring;
 	u64 invalid_descs;
Signature · 1 line
2.41.3
author
#2 lore raw

PATCH v2 2/3 xsk: use atomic operations around cached_prod for copy mode

In reply to Jason Xing

From: Jason Xing <[email protected]>

Use some exclusive functions for cached_prod in generic path instead of extending unified functions to avoid affecting zerocopy feature.

Use atomic operations.

---
 net/xdp/xsk.c       |  4 ++--
 net/xdp/xsk_queue.h | 21 ++++++++++++++++++---
 2 files changed, 20 insertions(+), 5 deletions(-)
net/xdp/xsk.c +2 -2
index bcfd400e9cf8..b63409b1422e 100644
@@ -551,7 +551,7 @@ static int xsk_cq_reserve_locked(struct xsk_buff_pool *pool)
 	int ret;
 
 	spin_lock(&pool->cq_cached_prod_lock);
-	ret = xskq_prod_reserve(pool->cq);
+	ret = xsk_cq_cached_prod_reserve(pool->cq);
 	spin_unlock(&pool->cq_cached_prod_lock);
 
 	return ret;
@@ -588,7 +588,7 @@ static void xsk_cq_submit_addr_locked(struct xsk_buff_pool *pool,
 static void xsk_cq_cancel_locked(struct xsk_buff_pool *pool, u32 n)
 {
 	spin_lock(&pool->cq_cached_prod_lock);
-	xskq_prod_cancel_n(pool->cq, n);
+	atomic_sub(n, &pool->cq->cached_prod_atomic);
 	spin_unlock(&pool->cq_cached_prod_lock);
 }
 
net/xdp/xsk_queue.h +18 -3
index 44cc01555c0b..3a023791b273 100644
@@ -402,13 +402,28 @@ static inline void xskq_prod_cancel_n(struct xsk_queue *q, u32 cnt)
 	q->cached_prod -= cnt;
 }
 
-static inline int xskq_prod_reserve(struct xsk_queue *q)
+static inline bool xsk_cq_cached_prod_nb_free(struct xsk_queue *q)
 {
-	if (xskq_prod_is_full(q))
+	u32 cached_prod = atomic_read(&q->cached_prod_atomic);
+	u32 free_entries = q->nentries - (cached_prod - q->cached_cons);
+
+	if (free_entries)
+		return true;
+
+	/* Refresh the local tail pointer */
+	q->cached_cons = READ_ONCE(q->ring->consumer);
+	free_entries = q->nentries - (cached_prod - q->cached_cons);
+
+	return free_entries ? true : false;
+}
+
+static inline int xsk_cq_cached_prod_reserve(struct xsk_queue *q)
+{
+	if (!xsk_cq_cached_prod_nb_free(q))
 		return -ENOSPC;
 
 	/* A, matches D */
-	q->cached_prod++;
+	atomic_inc(&q->cached_prod_atomic);
 	return 0;
 }
 
Signature · 1 line
2.41.3
maintainer
#3 lore raw

Re: [PATCH net-next v2 2/3] xsk: use atomic operations around cached_prod for copy mode

Quoted 14 lines · Jason Xing
net/xdp/xsk_queue.h +3 -2
index 44cc01555c0b..3a023791b273 100644
@@ -402,13 +402,28 @@ static inline void xskq_prod_cancel_n(struct xsk_queue *q, u32 cnt)
 	q->cached_prod -= cnt;
 }
 
-static inline int xskq_prod_reserve(struct xsk_queue *q)
+static inline bool xsk_cq_cached_prod_nb_free(struct xsk_queue *q)
 {
-	if (xskq_prod_is_full(q))
+	u32 cached_prod = atomic_read(&q->cached_prod_atomic);
+	u32 free_entries = q->nentries - (cached_prod - q->cached_cons);

net/xdp/xsk_queue.h static inline void xskq_prod_cancel_n(struct xsk_queue *q, u32 cnt) · line 418

Quoted 10 lines
net/xdp/xsk_queue.h +10 -0
+
+	if (free_entries)
+		return true;
+
+	/* Refresh the local tail pointer */
+	q->cached_cons = READ_ONCE(q->ring->consumer);
+	free_entries = q->nentries - (cached_prod - q->cached_cons);
+
+	return free_entries ? true : false;
+}

_If_ different CPUs can call xsk_cq_cached_prod_reserve() simultaneously (as the spinlock existence suggests) the above change introduce a race:

xsk_cq_cached_prod_nb_free() can return true when num_free == 1 on CPU1, and xsk_cq_cached_prod_reserve increment cached_prod_atomic on CPU2 before CPU1 completed xsk_cq_cached_prod_reserve().

/P

author
#4 lore raw

Re: [PATCH net-next v2 2/3] xsk: use atomic operations around cached_prod for copy mode

Quoted 29 lines · Paolo Abeni
Quoted 14 lines · Jason Xing
net/xdp/xsk_queue.h +3 -2
index 44cc01555c0b..3a023791b273 100644
@@ -402,13 +402,28 @@ static inline void xskq_prod_cancel_n(struct xsk_queue *q, u32 cnt)
      q->cached_prod -= cnt;
 }
 
-static inline int xskq_prod_reserve(struct xsk_queue *q)
+static inline bool xsk_cq_cached_prod_nb_free(struct xsk_queue *q)
 {
-     if (xskq_prod_is_full(q))
+     u32 cached_prod = atomic_read(&q->cached_prod_atomic);
+     u32 free_entries = q->nentries - (cached_prod - q->cached_cons);

net/xdp/xsk_queue.h static inline void xskq_prod_cancel_n(struct xsk_queue *q, u32 cnt) · line 418

Quoted 10 lines
net/xdp/xsk_queue.h +10 -0
+
+     if (free_entries)
+             return true;
+
+     /* Refresh the local tail pointer */
+     q->cached_cons = READ_ONCE(q->ring->consumer);
+     free_entries = q->nentries - (cached_prod - q->cached_cons);
+
+     return free_entries ? true : false;
+}

_If_ different CPUs can call xsk_cq_cached_prod_reserve() simultaneously (as the spinlock existence suggests) the above change introduce a race:

Quoted 3 lines

xsk_cq_cached_prod_nb_free() can return true when num_free == 1 on CPU1, and xsk_cq_cached_prod_reserve increment cached_prod_atomic on CPU2 before CPU1 completed xsk_cq_cached_prod_reserve().

I think you're right... I will give it more thought tomorrow morning.

I presume using try_cmpxchg() should work as it can detect if another process changes @cached_prod simultaneously. They both work similarly. But does it make any difference compared to spin lock? I don't have any handy benchmark to stably measure two xsk sharing the same umem, probably going to implement one.

Or like what you suggested in another thread, move that lock to struct xsk_queue?

Thanks,
Jason

maintainer
#5 lore raw

Re: [PATCH net-next v2 2/3] xsk: use atomic operations around cached_prod for copy mode

Quoted 35 lines · Jason Xing
Quoted 28 lines · Paolo Abeni
Quoted 14 lines · Jason Xing
net/xdp/xsk_queue.h +3 -2
index 44cc01555c0b..3a023791b273 100644
@@ -402,13 +402,28 @@ static inline void xskq_prod_cancel_n(struct xsk_queue *q, u32 cnt)
      q->cached_prod -= cnt;
 }
 
-static inline int xskq_prod_reserve(struct xsk_queue *q)
+static inline bool xsk_cq_cached_prod_nb_free(struct xsk_queue *q)
 {
-     if (xskq_prod_is_full(q))
+     u32 cached_prod = atomic_read(&q->cached_prod_atomic);
+     u32 free_entries = q->nentries - (cached_prod - q->cached_cons);

net/xdp/xsk_queue.h static inline void xskq_prod_cancel_n(struct xsk_queue *q, u32 cnt) · line 418

Quoted 10 lines
net/xdp/xsk_queue.h +10 -0
+
+     if (free_entries)
+             return true;
+
+     /* Refresh the local tail pointer */
+     q->cached_cons = READ_ONCE(q->ring->consumer);
+     free_entries = q->nentries - (cached_prod - q->cached_cons);
+
+     return free_entries ? true : false;
+}

_If_ different CPUs can call xsk_cq_cached_prod_reserve() simultaneously (as the spinlock existence suggests) the above change introduce a race:

Quoted 3 lines

xsk_cq_cached_prod_nb_free() can return true when num_free == 1 on CPU1, and xsk_cq_cached_prod_reserve increment cached_prod_atomic on CPU2 before CPU1 completed xsk_cq_cached_prod_reserve().

I think you're right... I will give it more thought tomorrow morning.

Quoted 8 lines

I presume using try_cmpxchg() should work as it can detect if another process changes @cached_prod simultaneously. They both work similarly. But does it make any difference compared to spin lock? I don't have any handy benchmark to stably measure two xsk sharing the same umem, probably going to implement one.

Or like what you suggested in another thread, move that lock to struct xsk_queue?

I think moving the lock should be preferable: I think it makes sense from a maintenance perspective to bundle the lock in the structure it protects, and I hope it should make the whole patch simpler.

Cheers,

Paolo

author
#6 lore raw

Re: [PATCH net-next v2 2/3] xsk: use atomic operations around cached_prod for copy mode

Quoted 43 lines · Paolo Abeni
Quoted 41 lines · Jason Xing
Quoted 28 lines · Paolo Abeni
Quoted 14 lines · Jason Xing
net/xdp/xsk_queue.h +3 -2
index 44cc01555c0b..3a023791b273 100644
@@ -402,13 +402,28 @@ static inline void xskq_prod_cancel_n(struct xsk_queue *q, u32 cnt)
      q->cached_prod -= cnt;
 }
 
-static inline int xskq_prod_reserve(struct xsk_queue *q)
+static inline bool xsk_cq_cached_prod_nb_free(struct xsk_queue *q)
 {
-     if (xskq_prod_is_full(q))
+     u32 cached_prod = atomic_read(&q->cached_prod_atomic);
+     u32 free_entries = q->nentries - (cached_prod - q->cached_cons);

net/xdp/xsk_queue.h static inline void xskq_prod_cancel_n(struct xsk_queue *q, u32 cnt) · line 418

Quoted 10 lines
net/xdp/xsk_queue.h +10 -0
+
+     if (free_entries)
+             return true;
+
+     /* Refresh the local tail pointer */
+     q->cached_cons = READ_ONCE(q->ring->consumer);
+     free_entries = q->nentries - (cached_prod - q->cached_cons);
+
+     return free_entries ? true : false;
+}

_If_ different CPUs can call xsk_cq_cached_prod_reserve() simultaneously (as the spinlock existence suggests) the above change introduce a race:

Quoted 3 lines

xsk_cq_cached_prod_nb_free() can return true when num_free == 1 on CPU1, and xsk_cq_cached_prod_reserve increment cached_prod_atomic on CPU2 before CPU1 completed xsk_cq_cached_prod_reserve().

I think you're right... I will give it more thought tomorrow morning.

I presume using try_cmpxchg() should work as it can detect if another process changes @cached_prod simultaneously. They both work similarly. But does it make any difference compared to spin lock? I don't have any handy benchmark to stably measure two xsk sharing the same umem, probably going to implement one.

Quoted 6 lines
Quoted 2 lines

Or like what you suggested in another thread, move that lock to struct xsk_queue?

I think moving the lock should be preferable: I think it makes sense from a maintenance perspective to bundle the lock in the structure it protects, and I hope it should make the whole patch simpler.

Agreed. At least so far I cannot see the benefits of using try_cmpxchg() instead as the protected area is really small. Probably in the future I will try a better way after successfully spotting the contention causing the performance problem.

I'm going to add your suggested-by tag since you provide this good idea :) Thanks!

Thanks,
Jason

author
#7 lore raw

PATCH v2 3/3 xsk: remove spin lock protection of cached_prod

In reply to Jason Xing

From: Jason Xing <[email protected]>

Remove the spin lock protection along with some functions adjusted.

Now cached_prod is fully converted to atomic, which improves the performance by around 5% over different platforms.

---
 include/net/xsk_buff_pool.h |  5 -----
 net/xdp/xsk.c               | 21 ++++-----------------
 net/xdp/xsk_buff_pool.c     |  1 -
 3 files changed, 4 insertions(+), 23 deletions(-)
include/net/xsk_buff_pool.h +0 -5
index 92a2358c6ce3..0b1abdb99c9e 100644
@@ -90,11 +90,6 @@ struct xsk_buff_pool {
 	 * destructor callback.
 	 */
 	spinlock_t cq_prod_lock;
-	/* Mutual exclusion of the completion ring in the SKB mode.
-	 * Protect: when sockets share a single cq when the same netdev
-	 * and queue id is shared.
-	 */
-	spinlock_t cq_cached_prod_lock;
 	struct xdp_buff_xsk *free_heads[];
 };
 
net/xdp/xsk.c +4 -17
index b63409b1422e..ae8a92c168b8 100644
@@ -546,17 +546,6 @@ static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
 	return dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
 }
 
-static int xsk_cq_reserve_locked(struct xsk_buff_pool *pool)
-{
-	int ret;
-
-	spin_lock(&pool->cq_cached_prod_lock);
-	ret = xsk_cq_cached_prod_reserve(pool->cq);
-	spin_unlock(&pool->cq_cached_prod_lock);
-
-	return ret;
-}
-
 static void xsk_cq_submit_addr_locked(struct xsk_buff_pool *pool,
 				      struct sk_buff *skb)
 {
@@ -585,11 +574,9 @@ static void xsk_cq_submit_addr_locked(struct xsk_buff_pool *pool,
 	spin_unlock_irqrestore(&pool->cq_prod_lock, flags);
 }
 
-static void xsk_cq_cancel_locked(struct xsk_buff_pool *pool, u32 n)
+static void xsk_cq_cached_prod_cancel(struct xsk_buff_pool *pool, u32 n)
 {
-	spin_lock(&pool->cq_cached_prod_lock);
 	atomic_sub(n, &pool->cq->cached_prod_atomic);
-	spin_unlock(&pool->cq_cached_prod_lock);
 }
 
 static void xsk_inc_num_desc(struct sk_buff *skb)
@@ -643,7 +630,7 @@ static void xsk_consume_skb(struct sk_buff *skb)
 	}
 
 	skb->destructor = sock_wfree;
-	xsk_cq_cancel_locked(xs->pool, num_descs);
+	xsk_cq_cached_prod_cancel(xs->pool, num_descs);
 	/* Free skb without triggering the perf drop trace */
 	consume_skb(skb);
 	xs->skb = NULL;
@@ -860,7 +847,7 @@ static struct sk_buff *xsk_build_skb(struct xdp_sock *xs,
 		xskq_cons_release(xs->tx);
 	} else {
 		/* Let application retry */
-		xsk_cq_cancel_locked(xs->pool, 1);
+		xsk_cq_cached_prod_cancel(xs->pool, 1);
 	}
 
 	return ERR_PTR(err);
@@ -898,7 +885,7 @@ static int __xsk_generic_xmit(struct sock *sk)
 		 * if there is space in it. This avoids having to implement
 		 * any buffering in the Tx path.
 		 */
-		err = xsk_cq_reserve_locked(xs->pool);
+		err = xsk_cq_cached_prod_reserve(xs->pool->cq);
 		if (err) {
 			err = -EAGAIN;
 			goto out;
net/xdp/xsk_buff_pool.c +0 -1
index 51526034c42a..9539f121b290 100644
@@ -91,7 +91,6 @@ struct xsk_buff_pool *xp_create_and_assign_umem(struct xdp_sock *xs,
 	INIT_LIST_HEAD(&pool->xsk_tx_list);
 	spin_lock_init(&pool->xsk_tx_list_lock);
 	spin_lock_init(&pool->cq_prod_lock);
-	spin_lock_init(&pool->cq_cached_prod_lock);
 	refcount_set(&pool->users, 1);
 
 	pool->fq = xs->fq_tmp;
Signature · 1 line
2.41.3
maintainer
#8 lore raw

Re: [PATCH net-next v2 3/3] xsk: remove spin lock protection of cached_prod

Quoted 6 lines · Jason Xing

From: Jason Xing <[email protected]>

Remove the spin lock protection along with some functions adjusted.

Now cached_prod is fully converted to atomic, which improves the performance by around 5% over different platforms.

I must admit that I'm surprised of the above delta; AFAIK replacing 1to1 spinlock with atomic should not impact performances measurably, as the thread should still see the same contention, and will use the same number of atomic operation on the bus.

net/xdp/xsk.c static void xsk_cq_submit_addr_locked(struct xsk_buff_pool *pool, · line 579

Quoted 9 lines
net/xdp/xsk.c +1 -2
@@ -585,11 +574,9 @@ static void xsk_cq_submit_addr_locked(struct xsk_buff_pool *pool,
 	spin_unlock_irqrestore(&pool->cq_prod_lock, flags);
 }
 
-static void xsk_cq_cancel_locked(struct xsk_buff_pool *pool, u32 n)
+static void xsk_cq_cached_prod_cancel(struct xsk_buff_pool *pool, u32 n)
 {
-	spin_lock(&pool->cq_cached_prod_lock);
 	atomic_sub(n, &pool->cq->cached_prod_atomic);

It looks like that the spinlock and the protected data are on different structs.

I wild guess/suspect the real gain comes from avoiding touching an additional cacheline.
struct xsk_queue size is 48 bytes and such struct is allocated via kmalloc. Adding up to 16 bytes there will not change the slub used and thus the actual memory usage.

I think that moving the cq_cached* spinlock(s) in xsk_queue should give the same gain, with much less code churn. Could you please have a look at such option?

/P

author
#9 lore raw

Re: [PATCH net-next v2 3/3] xsk: remove spin lock protection of cached_prod

Quoted 6 lines · Paolo Abeni
Quoted 4 lines · Jason Xing

From: Jason Xing <[email protected]>

Remove the spin lock protection along with some functions adjusted.

Quoted 7 lines
Quoted 2 lines

Now cached_prod is fully converted to atomic, which improves the performance by around 5% over different platforms.

I must admit that I'm surprised of the above delta; AFAIK replacing 1to1 spinlock with atomic should not impact performances measurably, as the thread should still see the same contention, and will use the same number of atomic operation on the bus.

Interesting point.

Quoted 21 lines

net/xdp/xsk.c static void xsk_cq_submit_addr_locked(struct xsk_buff_pool *pool, · line 579

Quoted 11 lines
net/xdp/xsk.c +1 -2
@@ -585,11 +574,9 @@ static void xsk_cq_submit_addr_locked(struct xsk_buff_pool *pool,
      spin_unlock_irqrestore(&pool->cq_prod_lock, flags);
 }
 
-static void xsk_cq_cancel_locked(struct xsk_buff_pool *pool, u32 n)
+static void xsk_cq_cached_prod_cancel(struct xsk_buff_pool *pool, u32 n)
 {
-     spin_lock(&pool->cq_cached_prod_lock);
      atomic_sub(n, &pool->cq->cached_prod_atomic);

It looks like that the spinlock and the protected data are on different structs.

I wild guess/suspect the real gain comes from avoiding touching an additional cacheline.
struct xsk_queue size is 48 bytes and such struct is allocated via kmalloc. Adding up to 16 bytes there will not change the slub used and thus the actual memory usage.

Quoted 3 lines

I think that moving the cq_cached* spinlock(s) in xsk_queue should give the same gain, with much less code churn. Could you please have a look at such option?

I just did some tests and observed the same result as you predicted. Thanks for the lesson!

Thanks,
Jason