net/core/devmem.c | 38 +++++++++++++++++++++++++++++++++----- net/core/devmem.h | 18 ++++++++++-------- 2 files changed, 43 insertions(+), 13 deletions(-)
From: Bobby Eshleman <bobbyeshleman@meta.com>
Convert net_devmem_dmabuf_binding refcount from refcount_t to percpu_ref
to optimize common-case reference counting on the hot path.
The typical devmem workflow involves binding a dmabuf to a queue
(acquiring the initial reference on binding->ref), followed by
high-volume traffic where every skb fragment acquires a reference.
Eventually traffic stops and the unbind operation releases the initial
reference. Additionally, the high traffic hot path is often multi-core.
This access pattern is ideal for percpu_ref as the first and last
reference during bind/unbind and normally book-ends activity in the hot
path.
__net_devmem_dmabuf_binding_free becomes the percpu_ref callback invoked
when the last reference is dropped.
kperf test:
- 4MB message sizes
- 60s of workload each run
- 5 runs
- 4 flows
Throughput:
Before: 45.31 GB/s (+/- 3.17 GB/s)
After: 48.67 GB/s (+/- 0.01 GB/s)
Picking throughput-matched kperf runs (both before and after matched at
~48 GB/s) for apples-to-apples comparison:
Summary (averaged across 4 workers):
TX worker CPU idle %:
Before: 34.44%
After: 87.13%
RX worker CPU idle %:
Before: 5.38%
After: 9.73%
kperf before:
client: == Source
client: Tx 98.100 Gbps (735764807680 bytes in 60001149 usec)
client: Tx102.798 Gbps (770996961280 bytes in 60001149 usec)
client: Tx101.534 Gbps (761517834240 bytes in 60001149 usec)
client: Tx 82.794 Gbps (620966707200 bytes in 60001149 usec)
client: net CPU 56: usr: 0.01% sys: 0.12% idle:17.06% iow: 0.00% irq: 9.89% sirq:72.91%
client: app CPU 60: usr: 0.08% sys:63.30% idle:36.24% iow: 0.00% irq: 0.30% sirq: 0.06%
client: net CPU 57: usr: 0.03% sys: 0.08% idle:75.68% iow: 0.00% irq: 2.96% sirq:21.23%
client: app CPU 61: usr: 0.06% sys:67.67% idle:31.94% iow: 0.00% irq: 0.28% sirq: 0.03%
client: net CPU 58: usr: 0.01% sys: 0.06% idle:76.87% iow: 0.00% irq: 2.84% sirq:20.19%
client: app CPU 62: usr: 0.06% sys:69.78% idle:29.79% iow: 0.00% irq: 0.30% sirq: 0.05%
client: net CPU 59: usr: 0.06% sys: 0.16% idle:74.97% iow: 0.00% irq: 3.76% sirq:21.03%
client: app CPU 63: usr: 0.06% sys:59.82% idle:39.80% iow: 0.00% irq: 0.25% sirq: 0.05%
client: == Target
client: Rx 98.092 Gbps (735764807680 bytes in 60006084 usec)
client: Rx102.785 Gbps (770962161664 bytes in 60006084 usec)
client: Rx101.523 Gbps (761499566080 bytes in 60006084 usec)
client: Rx 82.783 Gbps (620933136384 bytes in 60006084 usec)
client: net CPU 2: usr: 0.00% sys: 0.01% idle:24.51% iow: 0.00% irq: 1.67% sirq:73.79%
client: app CPU 6: usr: 1.51% sys:96.43% idle: 1.13% iow: 0.00% irq: 0.36% sirq: 0.55%
client: net CPU 1: usr: 0.00% sys: 0.01% idle:25.18% iow: 0.00% irq: 1.99% sirq:72.80%
client: app CPU 5: usr: 2.21% sys:94.54% idle: 2.54% iow: 0.00% irq: 0.38% sirq: 0.30%
client: net CPU 3: usr: 0.00% sys: 0.01% idle:26.34% iow: 0.00% irq: 2.12% sirq:71.51%
client: app CPU 7: usr: 2.22% sys:94.28% idle: 2.52% iow: 0.00% irq: 0.59% sirq: 0.37%
client: net CPU 0: usr: 0.00% sys: 0.03% idle: 0.00% iow: 0.00% irq:10.44% sirq:89.51%
client: app CPU 4: usr: 2.39% sys:81.46% idle:15.33% iow: 0.00% irq: 0.50% sirq: 0.30%
kperf after:
client: == Source
client: Tx 99.257 Gbps (744447016960 bytes in 60001303 usec)
client: Tx101.013 Gbps (757617131520 bytes in 60001303 usec)
client: Tx 88.179 Gbps (661357854720 bytes in 60001303 usec)
client: Tx101.002 Gbps (757533245440 bytes in 60001303 usec)
client: net CPU 56: usr: 0.00% sys: 0.01% idle: 6.22% iow: 0.00% irq: 8.68% sirq:85.06%
client: app CPU 60: usr: 0.08% sys:12.56% idle:87.21% iow: 0.00% irq: 0.08% sirq: 0.05%
client: net CPU 57: usr: 0.00% sys: 0.05% idle:69.53% iow: 0.00% irq: 2.02% sirq:28.38%
client: app CPU 61: usr: 0.11% sys:13.40% idle:86.36% iow: 0.00% irq: 0.08% sirq: 0.03%
client: net CPU 58: usr: 0.00% sys: 0.03% idle:70.04% iow: 0.00% irq: 3.38% sirq:26.53%
client: app CPU 62: usr: 0.10% sys:11.46% idle:88.31% iow: 0.00% irq: 0.08% sirq: 0.03%
client: net CPU 59: usr: 0.01% sys: 0.06% idle:71.18% iow: 0.00% irq: 1.97% sirq:26.75%
client: app CPU 63: usr: 0.10% sys:13.10% idle:86.64% iow: 0.00% irq: 0.10% sirq: 0.05%
client: == Target
client: Rx 99.250 Gbps (744415182848 bytes in 60003297 usec)
client: Rx101.006 Gbps (757589737472 bytes in 60003297 usec)
client: Rx 88.171 Gbps (661319475200 bytes in 60003297 usec)
client: Rx100.996 Gbps (757514792960 bytes in 60003297 usec)
client: net CPU 2: usr: 0.00% sys: 0.01% idle:28.02% iow: 0.00% irq: 1.95% sirq:70.00%
client: app CPU 6: usr: 2.03% sys:87.20% idle:10.04% iow: 0.00% irq: 0.37% sirq: 0.33%
client: net CPU 3: usr: 0.00% sys: 0.00% idle:27.63% iow: 0.00% irq: 1.90% sirq:70.45%
client: app CPU 7: usr: 1.78% sys:89.70% idle: 7.79% iow: 0.00% irq: 0.37% sirq: 0.34%
client: net CPU 0: usr: 0.00% sys: 0.01% idle: 0.00% iow: 0.00% irq: 9.96% sirq:90.01%
client: app CPU 4: usr: 2.33% sys:83.51% idle:13.24% iow: 0.00% irq: 0.64% sirq: 0.26%
client: net CPU 1: usr: 0.00% sys: 0.01% idle:27.60% iow: 0.00% irq: 1.94% sirq:70.43%
client: app CPU 5: usr: 1.88% sys:89.61% idle: 7.86% iow: 0.00% irq: 0.35% sirq: 0.27%
Signed-off-by: Bobby Eshleman <bobbyeshleman@meta.com>
---
net/core/devmem.c | 38 +++++++++++++++++++++++++++++++++-----
net/core/devmem.h | 18 ++++++++++--------
2 files changed, 43 insertions(+), 13 deletions(-)
diff --git a/net/core/devmem.c b/net/core/devmem.c
index 1d04754bc756..83989cf4a987 100644
--- a/net/core/devmem.c
+++ b/net/core/devmem.c
@@ -54,10 +54,26 @@ static dma_addr_t net_devmem_get_dma_addr(const struct net_iov *niov)
((dma_addr_t)net_iov_idx(niov) << PAGE_SHIFT);
}
-void __net_devmem_dmabuf_binding_free(struct work_struct *wq)
+/*
+ * percpu_ref release callback invoked when the last reference to the binding
+ * is dropped. Schedules the actual cleanup in a workqueue because
+ * ref->release() cb is not allowed to sleep as it may be called in RCU
+ * callback context.
+ */
+static void net_devmem_dmabuf_binding_release(struct percpu_ref *ref)
{
- struct net_devmem_dmabuf_binding *binding = container_of(wq, typeof(*binding), unbind_w);
+ struct net_devmem_dmabuf_binding *binding =
+ container_of(ref, struct net_devmem_dmabuf_binding, ref);
+
+ INIT_WORK(&binding->unbind_w, __net_devmem_dmabuf_binding_free);
+ schedule_work(&binding->unbind_w);
+}
+/* Workqueue callback that performs the actual cleanup of the binding. */
+void __net_devmem_dmabuf_binding_free(struct work_struct *work)
+{
+ struct net_devmem_dmabuf_binding *binding =
+ container_of(work, struct net_devmem_dmabuf_binding, unbind_w);
size_t size, avail;
gen_pool_for_each_chunk(binding->chunk_pool,
@@ -75,6 +91,7 @@ void __net_devmem_dmabuf_binding_free(struct work_struct *wq)
dma_buf_detach(binding->dmabuf, binding->attachment);
dma_buf_put(binding->dmabuf);
xa_destroy(&binding->bound_rxqs);
+ percpu_ref_exit(&binding->ref);
kvfree(binding->tx_vec);
kfree(binding);
}
@@ -143,7 +160,11 @@ void net_devmem_unbind_dmabuf(struct net_devmem_dmabuf_binding *binding)
__net_mp_close_rxq(binding->dev, rxq_idx, &mp_params);
}
- net_devmem_dmabuf_binding_put(binding);
+ /* Switch to atomic mode and drop the initial reference we acquired at
+ * bind time. This will invoke __net_devmem_dmabuf_binding_free if this
+ * is the last reference.
+ */
+ percpu_ref_kill(&binding->ref);
}
int net_devmem_bind_dmabuf_to_queue(struct net_device *dev, u32 rxq_idx,
@@ -209,7 +230,12 @@ net_devmem_bind_dmabuf(struct net_device *dev,
binding->dev = dev;
xa_init_flags(&binding->bound_rxqs, XA_FLAGS_ALLOC);
- refcount_set(&binding->ref, 1);
+ err = percpu_ref_init(&binding->ref,
+ net_devmem_dmabuf_binding_release,
+ 0, GFP_KERNEL);
+
+ if (err < 0)
+ goto err_free_binding;
mutex_init(&binding->lock);
@@ -220,7 +246,7 @@ net_devmem_bind_dmabuf(struct net_device *dev,
if (IS_ERR(binding->attachment)) {
err = PTR_ERR(binding->attachment);
NL_SET_ERR_MSG(extack, "Failed to bind dmabuf to device");
- goto err_free_binding;
+ goto err_exit_ref;
}
binding->sgt = dma_buf_map_attachment_unlocked(binding->attachment,
@@ -322,6 +348,8 @@ net_devmem_bind_dmabuf(struct net_device *dev,
direction);
err_detach:
dma_buf_detach(dmabuf, binding->attachment);
+err_exit_ref:
+ percpu_ref_exit(&binding->ref);
err_free_binding:
kfree(binding);
err_put_dmabuf:
diff --git a/net/core/devmem.h b/net/core/devmem.h
index 0b43a648cd2e..e9f900dd4060 100644
--- a/net/core/devmem.h
+++ b/net/core/devmem.h
@@ -41,7 +41,7 @@ struct net_devmem_dmabuf_binding {
* retransmits) hold a reference to the binding until the skb holding
* them is freed.
*/
- refcount_t ref;
+ struct percpu_ref ref;
/* The list of bindings currently active. Used for netlink to notify us
* of the user dropping the bind.
@@ -82,7 +82,7 @@ struct dmabuf_genpool_chunk_owner {
dma_addr_t base_dma_addr;
};
-void __net_devmem_dmabuf_binding_free(struct work_struct *wq);
+void __net_devmem_dmabuf_binding_free(struct work_struct *work);
struct net_devmem_dmabuf_binding *
net_devmem_bind_dmabuf(struct net_device *dev,
struct device *dma_dev,
@@ -125,17 +125,19 @@ static inline unsigned long net_iov_virtual_addr(const struct net_iov *niov)
static inline bool
net_devmem_dmabuf_binding_get(struct net_devmem_dmabuf_binding *binding)
{
- return refcount_inc_not_zero(&binding->ref);
+ return percpu_ref_tryget(&binding->ref);
}
static inline void
net_devmem_dmabuf_binding_put(struct net_devmem_dmabuf_binding *binding)
{
- if (!refcount_dec_and_test(&binding->ref))
- return;
-
- INIT_WORK(&binding->unbind_w, __net_devmem_dmabuf_binding_free);
- schedule_work(&binding->unbind_w);
+ /* When the ref count reaches 0, the percpu_ref release callback will
+ * schedule __net_devmem_dmabuf_binding_free() to run on a workqueue.
+ * More often the dmabuf is still bound and so the reference count
+ * remains above 0 here. Consequently, the cleanup is typically
+ * triggered by net_devmem_unbind_dmabuf via percpu_ref_kill().
+ */
+ percpu_ref_put(&binding->ref);
}
void net_devmem_get_net_iov(struct net_iov *niov);
---
base-commit: ab084f0b8d6d2ee4b1c6a28f39a2a7430bdfa7f0
change-id: 20251126-upstream-percpu-ref-401327f62b4b
Best regards,
--
Bobby Eshleman <bobbyeshleman@meta.com>On 11/26, Bobby Eshleman wrote: > From: Bobby Eshleman <bobbyeshleman@meta.com> > > Convert net_devmem_dmabuf_binding refcount from refcount_t to percpu_ref > to optimize common-case reference counting on the hot path. > > The typical devmem workflow involves binding a dmabuf to a queue > (acquiring the initial reference on binding->ref), followed by > high-volume traffic where every skb fragment acquires a reference. > Eventually traffic stops and the unbind operation releases the initial > reference. Additionally, the high traffic hot path is often multi-core. > This access pattern is ideal for percpu_ref as the first and last > reference during bind/unbind and normally book-ends activity in the hot > path. > > __net_devmem_dmabuf_binding_free becomes the percpu_ref callback invoked > when the last reference is dropped. > > kperf test: > - 4MB message sizes > - 60s of workload each run > - 5 runs > - 4 flows > > Throughput: > Before: 45.31 GB/s (+/- 3.17 GB/s) > After: 48.67 GB/s (+/- 0.01 GB/s) > > Picking throughput-matched kperf runs (both before and after matched at > ~48 GB/s) for apples-to-apples comparison: > > Summary (averaged across 4 workers): > > TX worker CPU idle %: > Before: 34.44% > After: 87.13% > > RX worker CPU idle %: > Before: 5.38% > After: 9.73% > > kperf before: > > client: == Source > client: Tx 98.100 Gbps (735764807680 bytes in 60001149 usec) > client: Tx102.798 Gbps (770996961280 bytes in 60001149 usec) > client: Tx101.534 Gbps (761517834240 bytes in 60001149 usec) > client: Tx 82.794 Gbps (620966707200 bytes in 60001149 usec) > client: net CPU 56: usr: 0.01% sys: 0.12% idle:17.06% iow: 0.00% irq: 9.89% sirq:72.91% > client: app CPU 60: usr: 0.08% sys:63.30% idle:36.24% iow: 0.00% irq: 0.30% sirq: 0.06% > client: net CPU 57: usr: 0.03% sys: 0.08% idle:75.68% iow: 0.00% irq: 2.96% sirq:21.23% > client: app CPU 61: usr: 0.06% sys:67.67% idle:31.94% iow: 0.00% irq: 0.28% sirq: 0.03% > client: net CPU 58: usr: 0.01% sys: 0.06% idle:76.87% iow: 0.00% irq: 2.84% sirq:20.19% > client: app CPU 62: usr: 0.06% sys:69.78% idle:29.79% iow: 0.00% irq: 0.30% sirq: 0.05% > client: net CPU 59: usr: 0.06% sys: 0.16% idle:74.97% iow: 0.00% irq: 3.76% sirq:21.03% > client: app CPU 63: usr: 0.06% sys:59.82% idle:39.80% iow: 0.00% irq: 0.25% sirq: 0.05% > client: == Target > client: Rx 98.092 Gbps (735764807680 bytes in 60006084 usec) > client: Rx102.785 Gbps (770962161664 bytes in 60006084 usec) > client: Rx101.523 Gbps (761499566080 bytes in 60006084 usec) > client: Rx 82.783 Gbps (620933136384 bytes in 60006084 usec) > client: net CPU 2: usr: 0.00% sys: 0.01% idle:24.51% iow: 0.00% irq: 1.67% sirq:73.79% > client: app CPU 6: usr: 1.51% sys:96.43% idle: 1.13% iow: 0.00% irq: 0.36% sirq: 0.55% > client: net CPU 1: usr: 0.00% sys: 0.01% idle:25.18% iow: 0.00% irq: 1.99% sirq:72.80% > client: app CPU 5: usr: 2.21% sys:94.54% idle: 2.54% iow: 0.00% irq: 0.38% sirq: 0.30% > client: net CPU 3: usr: 0.00% sys: 0.01% idle:26.34% iow: 0.00% irq: 2.12% sirq:71.51% > client: app CPU 7: usr: 2.22% sys:94.28% idle: 2.52% iow: 0.00% irq: 0.59% sirq: 0.37% > client: net CPU 0: usr: 0.00% sys: 0.03% idle: 0.00% iow: 0.00% irq:10.44% sirq:89.51% > client: app CPU 4: usr: 2.39% sys:81.46% idle:15.33% iow: 0.00% irq: 0.50% sirq: 0.30% > > kperf after: > > client: == Source > client: Tx 99.257 Gbps (744447016960 bytes in 60001303 usec) > client: Tx101.013 Gbps (757617131520 bytes in 60001303 usec) > client: Tx 88.179 Gbps (661357854720 bytes in 60001303 usec) > client: Tx101.002 Gbps (757533245440 bytes in 60001303 usec) > client: net CPU 56: usr: 0.00% sys: 0.01% idle: 6.22% iow: 0.00% irq: 8.68% sirq:85.06% > client: app CPU 60: usr: 0.08% sys:12.56% idle:87.21% iow: 0.00% irq: 0.08% sirq: 0.05% > client: net CPU 57: usr: 0.00% sys: 0.05% idle:69.53% iow: 0.00% irq: 2.02% sirq:28.38% > client: app CPU 61: usr: 0.11% sys:13.40% idle:86.36% iow: 0.00% irq: 0.08% sirq: 0.03% > client: net CPU 58: usr: 0.00% sys: 0.03% idle:70.04% iow: 0.00% irq: 3.38% sirq:26.53% > client: app CPU 62: usr: 0.10% sys:11.46% idle:88.31% iow: 0.00% irq: 0.08% sirq: 0.03% > client: net CPU 59: usr: 0.01% sys: 0.06% idle:71.18% iow: 0.00% irq: 1.97% sirq:26.75% > client: app CPU 63: usr: 0.10% sys:13.10% idle:86.64% iow: 0.00% irq: 0.10% sirq: 0.05% > client: == Target > client: Rx 99.250 Gbps (744415182848 bytes in 60003297 usec) > client: Rx101.006 Gbps (757589737472 bytes in 60003297 usec) > client: Rx 88.171 Gbps (661319475200 bytes in 60003297 usec) > client: Rx100.996 Gbps (757514792960 bytes in 60003297 usec) > client: net CPU 2: usr: 0.00% sys: 0.01% idle:28.02% iow: 0.00% irq: 1.95% sirq:70.00% > client: app CPU 6: usr: 2.03% sys:87.20% idle:10.04% iow: 0.00% irq: 0.37% sirq: 0.33% > client: net CPU 3: usr: 0.00% sys: 0.00% idle:27.63% iow: 0.00% irq: 1.90% sirq:70.45% > client: app CPU 7: usr: 1.78% sys:89.70% idle: 7.79% iow: 0.00% irq: 0.37% sirq: 0.34% > client: net CPU 0: usr: 0.00% sys: 0.01% idle: 0.00% iow: 0.00% irq: 9.96% sirq:90.01% > client: app CPU 4: usr: 2.33% sys:83.51% idle:13.24% iow: 0.00% irq: 0.64% sirq: 0.26% > client: net CPU 1: usr: 0.00% sys: 0.01% idle:27.60% iow: 0.00% irq: 1.94% sirq:70.43% > client: app CPU 5: usr: 1.88% sys:89.61% idle: 7.86% iow: 0.00% irq: 0.35% sirq: 0.27% > > Signed-off-by: Bobby Eshleman <bobbyeshleman@meta.com> > --- > net/core/devmem.c | 38 +++++++++++++++++++++++++++++++++----- > net/core/devmem.h | 18 ++++++++++-------- > 2 files changed, 43 insertions(+), 13 deletions(-) > > diff --git a/net/core/devmem.c b/net/core/devmem.c > index 1d04754bc756..83989cf4a987 100644 > --- a/net/core/devmem.c > +++ b/net/core/devmem.c > @@ -54,10 +54,26 @@ static dma_addr_t net_devmem_get_dma_addr(const struct net_iov *niov) > ((dma_addr_t)net_iov_idx(niov) << PAGE_SHIFT); > } > > -void __net_devmem_dmabuf_binding_free(struct work_struct *wq) > +/* > + * percpu_ref release callback invoked when the last reference to the binding > + * is dropped. Schedules the actual cleanup in a workqueue because > + * ref->release() cb is not allowed to sleep as it may be called in RCU > + * callback context. > + */ Can we drop this and the rest of the comments? I feel like they mostly explain how percpu_ref works, nothing devmem specific. refcnt-wise, feels like the only place that deserves a comment is net_devmem_get_net_iov (why it's safe to ignore net_devmem_dmabuf_binding_get return value, but you are not touching that..) Otherwise LGTM!
© 2016 - 2025 Red Hat, Inc.