[PATCH] net: devmem: Remove dst (ENODEV) check in net_devmem_get_binding

[PATCH] net: devmem: Remove dst (ENODEV) check in net_devmem_get_binding

[Shivaji Kant]

The Devmem TX binding lookup function, performs a strict
check against the socket's destination cache (`dst`) to
ensure the bound `dmabuf_id` corresponds to the correct
network device (`dst->dev->ifindex == binding->dev->ifindex`).

However, this check incorrectly fails and returns `-ENODEV`
if the socket's route cache entry (`dst`) is merely missing
or expired (`dst == NULL`). This scenario is observed during
network events, such as when flow steering rules are deleted,
leading to a temporary route cache invalidation.

The parent caller, `tcp_sendmsg_locked()`, is already
responsible for acquiring or validating the route (`dst_entry`).
If `dst` is `NULL`, `tcp_sendmsg_locked()` will correctly
derive the route before transmission.

This patch removes the `dst` validation from
`net_devmem_get_binding()`. The function now only validates
the existence of the binding and its TX vector, relying on the
calling context for device/route correctness. This allows
temporary route cache misses to be handled gracefully by the
TCP/IP stack without ENODEV error on the Devmem TX path.

Reported-by: Eric Dumazet <edumazet@google.com>
Reported-by: Vedant Mathur <vedantmathur@google.com>
Suggested-by: Eric Dumazet <edumazet@google.com>
Fixes: bd61848900bf ("net: devmem: Implement TX path")
Signed-off-by: Shivaji Kant <shivajikant@google.com>
---
 net/core/devmem.c | 27 ++++++++++++++++++++++++---
 1 file changed, 24 insertions(+), 3 deletions(-)

diff --git a/net/core/devmem.c b/net/core/devmem.c
index d9de31a6cc7f..1d04754bc756 100644
--- a/net/core/devmem.c
+++ b/net/core/devmem.c
@@ -17,6 +17,7 @@
 #include <net/page_pool/helpers.h>
 #include <net/page_pool/memory_provider.h>
 #include <net/sock.h>
+#include <net/tcp.h>
 #include <trace/events/page_pool.h>
 
 #include "devmem.h"
@@ -357,7 +358,8 @@ struct net_devmem_dmabuf_binding *net_devmem_get_binding(struct sock *sk,
 							 unsigned int dmabuf_id)
 {
 	struct net_devmem_dmabuf_binding *binding;
-	struct dst_entry *dst = __sk_dst_get(sk);
+	struct net_device *dst_dev;
+	struct dst_entry *dst;
 	int err = 0;
 
 	binding = net_devmem_lookup_dmabuf(dmabuf_id);
@@ -366,16 +368,35 @@ struct net_devmem_dmabuf_binding *net_devmem_get_binding(struct sock *sk,
 		goto out_err;
 	}
 
+	rcu_read_lock();
+	dst = __sk_dst_get(sk);
+	/* If dst is NULL (route expired), attempt to rebuild it. */
+	if (unlikely(!dst)) {
+		if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) {
+			err = -EHOSTUNREACH;
+			goto out_unlock;
+		}
+		dst = __sk_dst_get(sk);
+		if (unlikely(!dst)) {
+			err = -ENODEV;
+			goto out_unlock;
+		}
+	}
+
 	/* The dma-addrs in this binding are only reachable to the corresponding
 	 * net_device.
 	 */
-	if (!dst || !dst->dev || dst->dev->ifindex != binding->dev->ifindex) {
+	dst_dev = dst_dev_rcu(dst);
+	if (unlikely(!dst_dev) || unlikely(dst_dev != binding->dev)) {
 		err = -ENODEV;
-		goto out_err;
+		goto out_unlock;
 	}
 
+	rcu_read_unlock();
 	return binding;
 
+out_unlock:
+	rcu_read_unlock();
 out_err:
 	if (binding)
 		net_devmem_dmabuf_binding_put(binding);
-- 
2.51.1.838.g19442a804e-goog

Re: [PATCH] net: devmem: Remove dst (ENODEV) check in net_devmem_get_binding

[Eric Dumazet]

On Mon, Oct 27, 2025 at 11:07 PM Shivaji Kant <shivajikant@google.com> wrote:
>
> The Devmem TX binding lookup function, performs a strict
> check against the socket's destination cache (`dst`) to
> ensure the bound `dmabuf_id` corresponds to the correct
> network device (`dst->dev->ifindex == binding->dev->ifindex`).
>
> However, this check incorrectly fails and returns `-ENODEV`
> if the socket's route cache entry (`dst`) is merely missing
> or expired (`dst == NULL`). This scenario is observed during
> network events, such as when flow steering rules are deleted,
> leading to a temporary route cache invalidation.
>
> The parent caller, `tcp_sendmsg_locked()`, is already
> responsible for acquiring or validating the route (`dst_entry`).
> If `dst` is `NULL`, `tcp_sendmsg_locked()` will correctly
> derive the route before transmission.
>
> This patch removes the `dst` validation from
> `net_devmem_get_binding()`. The function now only validates
> the existence of the binding and its TX vector, relying on the
> calling context for device/route correctness. This allows
> temporary route cache misses to be handled gracefully by the
> TCP/IP stack without ENODEV error on the Devmem TX path.
>
> Reported-by: Eric Dumazet <edumazet@google.com>
> Reported-by: Vedant Mathur <vedantmathur@google.com>
> Suggested-by: Eric Dumazet <edumazet@google.com>
> Fixes: bd61848900bf ("net: devmem: Implement TX path")
> Signed-off-by: Shivaji Kant <shivajikant@google.com>
> ---
>  net/core/devmem.c | 27 ++++++++++++++++++++++++---
>  1 file changed, 24 insertions(+), 3 deletions(-)
>

Patch looks good, but the title should be improved a bit.

"Remove dst (ENODEV) check in net_devmem_get_binding"

It is not about removing the check, more about refreshing the dst if necessary ?

Please wait ~24 hours before sending an updated version, to  give time
for other reviewers to chime in.

Thanks !

Re: [PATCH] net: devmem: Remove dst (ENODEV) check in net_devmem_get_binding

[Shivaji Kant]

Thanks Eric for pointing it out.
The description of the patch needs to be updated.
Will send a v2 with updated description in 24 hrs.


On Tue, Oct 28, 2025 at 12:33 PM Eric Dumazet <edumazet@google.com> wrote:
>
> On Mon, Oct 27, 2025 at 11:07 PM Shivaji Kant <shivajikant@google.com> wrote:
> >
> > The Devmem TX binding lookup function, performs a strict
> > check against the socket's destination cache (`dst`) to
> > ensure the bound `dmabuf_id` corresponds to the correct
> > network device (`dst->dev->ifindex == binding->dev->ifindex`).
> >
> > However, this check incorrectly fails and returns `-ENODEV`
> > if the socket's route cache entry (`dst`) is merely missing
> > or expired (`dst == NULL`). This scenario is observed during
> > network events, such as when flow steering rules are deleted,
> > leading to a temporary route cache invalidation.
> >
> > The parent caller, `tcp_sendmsg_locked()`, is already
> > responsible for acquiring or validating the route (`dst_entry`).
> > If `dst` is `NULL`, `tcp_sendmsg_locked()` will correctly
> > derive the route before transmission.
> >
> > This patch removes the `dst` validation from
> > `net_devmem_get_binding()`. The function now only validates
> > the existence of the binding and its TX vector, relying on the
> > calling context for device/route correctness. This allows
> > temporary route cache misses to be handled gracefully by the
> > TCP/IP stack without ENODEV error on the Devmem TX path.
> >
> > Reported-by: Eric Dumazet <edumazet@google.com>
> > Reported-by: Vedant Mathur <vedantmathur@google.com>
> > Suggested-by: Eric Dumazet <edumazet@google.com>
> > Fixes: bd61848900bf ("net: devmem: Implement TX path")
> > Signed-off-by: Shivaji Kant <shivajikant@google.com>
> > ---
> >  net/core/devmem.c | 27 ++++++++++++++++++++++++---
> >  1 file changed, 24 insertions(+), 3 deletions(-)
> >
>
> Patch looks good, but the title should be improved a bit.
>
> "Remove dst (ENODEV) check in net_devmem_get_binding"
>
> It is not about removing the check, more about refreshing the dst if necessary ?
>
> Please wait ~24 hours before sending an updated version, to  give time
> for other reviewers to chime in.
>
> Thanks !

Re: [PATCH] net: devmem: Remove dst (ENODEV) check in net_devmem_get_binding

[Bobby Eshleman]

On Tue, Oct 28, 2025 at 06:07:14AM +0000, Shivaji Kant wrote:
> The Devmem TX binding lookup function, performs a strict
> check against the socket's destination cache (`dst`) to
> ensure the bound `dmabuf_id` corresponds to the correct
> network device (`dst->dev->ifindex == binding->dev->ifindex`).
> 
> However, this check incorrectly fails and returns `-ENODEV`
> if the socket's route cache entry (`dst`) is merely missing
> or expired (`dst == NULL`). This scenario is observed during
> network events, such as when flow steering rules are deleted,
> leading to a temporary route cache invalidation.
> 
> The parent caller, `tcp_sendmsg_locked()`, is already
> responsible for acquiring or validating the route (`dst_entry`).
> If `dst` is `NULL`, `tcp_sendmsg_locked()` will correctly
> derive the route before transmission.
> 
> This patch removes the `dst` validation from
> `net_devmem_get_binding()`. The function now only validates
> the existence of the binding and its TX vector, relying on the
> calling context for device/route correctness. This allows
> temporary route cache misses to be handled gracefully by the
> TCP/IP stack without ENODEV error on the Devmem TX path.
> 
> Reported-by: Eric Dumazet <edumazet@google.com>
> Reported-by: Vedant Mathur <vedantmathur@google.com>
> Suggested-by: Eric Dumazet <edumazet@google.com>
> Fixes: bd61848900bf ("net: devmem: Implement TX path")
> Signed-off-by: Shivaji Kant <shivajikant@google.com>
> ---
>  net/core/devmem.c | 27 ++++++++++++++++++++++++---
>  1 file changed, 24 insertions(+), 3 deletions(-)
> 
> diff --git a/net/core/devmem.c b/net/core/devmem.c
> index d9de31a6cc7f..1d04754bc756 100644
> --- a/net/core/devmem.c
> +++ b/net/core/devmem.c
> @@ -17,6 +17,7 @@
>  #include <net/page_pool/helpers.h>
>  #include <net/page_pool/memory_provider.h>
>  #include <net/sock.h>
> +#include <net/tcp.h>
>  #include <trace/events/page_pool.h>
>  
>  #include "devmem.h"
> @@ -357,7 +358,8 @@ struct net_devmem_dmabuf_binding *net_devmem_get_binding(struct sock *sk,
>  							 unsigned int dmabuf_id)
>  {
>  	struct net_devmem_dmabuf_binding *binding;
> -	struct dst_entry *dst = __sk_dst_get(sk);
> +	struct net_device *dst_dev;
> +	struct dst_entry *dst;
>  	int err = 0;
>  
>  	binding = net_devmem_lookup_dmabuf(dmabuf_id);
> @@ -366,16 +368,35 @@ struct net_devmem_dmabuf_binding *net_devmem_get_binding(struct sock *sk,
>  		goto out_err;
>  	}
>  
> +	rcu_read_lock();
> +	dst = __sk_dst_get(sk);
> +	/* If dst is NULL (route expired), attempt to rebuild it. */
> +	if (unlikely(!dst)) {
> +		if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) {
> +			err = -EHOSTUNREACH;
> +			goto out_unlock;
> +		}

Echoing your discussion with Eric, I think the message might want to
call out this part. Besides that, all looks good!

Pending that nit:

Reviewed-by: Bobby Eshleman <bobbyeshleman@meta.com>

Best,
Bobby

Re: [PATCH] net: devmem: Remove dst (ENODEV) check in net_devmem_get_binding

[Shivaji Kant]

On Tue, Oct 28, 2025 at 11:41 PM Bobby Eshleman <bobbyeshleman@gmail.com> wrote:
>
> On Tue, Oct 28, 2025 at 06:07:14AM +0000, Shivaji Kant wrote:
> > The Devmem TX binding lookup function, performs a strict
> > check against the socket's destination cache (`dst`) to
> > ensure the bound `dmabuf_id` corresponds to the correct
> > network device (`dst->dev->ifindex == binding->dev->ifindex`).
> >
> > However, this check incorrectly fails and returns `-ENODEV`
> > if the socket's route cache entry (`dst`) is merely missing
> > or expired (`dst == NULL`). This scenario is observed during
> > network events, such as when flow steering rules are deleted,
> > leading to a temporary route cache invalidation.
> >
> > The parent caller, `tcp_sendmsg_locked()`, is already
> > responsible for acquiring or validating the route (`dst_entry`).
> > If `dst` is `NULL`, `tcp_sendmsg_locked()` will correctly
> > derive the route before transmission.
> >
> > This patch removes the `dst` validation from
> > `net_devmem_get_binding()`. The function now only validates
> > the existence of the binding and its TX vector, relying on the
> > calling context for device/route correctness. This allows
> > temporary route cache misses to be handled gracefully by the
> > TCP/IP stack without ENODEV error on the Devmem TX path.
> >
> > Reported-by: Eric Dumazet <edumazet@google.com>
> > Reported-by: Vedant Mathur <vedantmathur@google.com>
> > Suggested-by: Eric Dumazet <edumazet@google.com>
> > Fixes: bd61848900bf ("net: devmem: Implement TX path")
> > Signed-off-by: Shivaji Kant <shivajikant@google.com>
> > ---
> >  net/core/devmem.c | 27 ++++++++++++++++++++++++---
> >  1 file changed, 24 insertions(+), 3 deletions(-)
> >
> > diff --git a/net/core/devmem.c b/net/core/devmem.c
> > index d9de31a6cc7f..1d04754bc756 100644
> > --- a/net/core/devmem.c
> > +++ b/net/core/devmem.c
> > @@ -17,6 +17,7 @@
> >  #include <net/page_pool/helpers.h>
> >  #include <net/page_pool/memory_provider.h>
> >  #include <net/sock.h>
> > +#include <net/tcp.h>
> >  #include <trace/events/page_pool.h>
> >
> >  #include "devmem.h"
> > @@ -357,7 +358,8 @@ struct net_devmem_dmabuf_binding *net_devmem_get_binding(struct sock *sk,
> >                                                        unsigned int dmabuf_id)
> >  {
> >       struct net_devmem_dmabuf_binding *binding;
> > -     struct dst_entry *dst = __sk_dst_get(sk);
> > +     struct net_device *dst_dev;
> > +     struct dst_entry *dst;
> >       int err = 0;
> >
> >       binding = net_devmem_lookup_dmabuf(dmabuf_id);
> > @@ -366,16 +368,35 @@ struct net_devmem_dmabuf_binding *net_devmem_get_binding(struct sock *sk,
> >               goto out_err;
> >       }
> >
> > +     rcu_read_lock();
> > +     dst = __sk_dst_get(sk);
> > +     /* If dst is NULL (route expired), attempt to rebuild it. */
> > +     if (unlikely(!dst)) {
> > +             if (inet_csk(sk)->icsk_af_ops->rebuild_header(sk)) {
> > +                     err = -EHOSTUNREACH;
> > +                     goto out_unlock;
> > +             }
>
> Echoing your discussion with Eric, I think the message might want to
> call out this part. Besides that, all looks good!
>
> Pending that nit:
>
> Reviewed-by: Bobby Eshleman <bobbyeshleman@meta.com>
>
> Best,
> Bobby

Thanks Bobby for the review.
Will surely update the description in the v2 patch.