[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20250605141624.GG7435@unreal>
Date: Thu, 5 Jun 2025 17:16:24 +0300
From: Leon Romanovsky <leon@...nel.org>
To: Sabrina Dubroca <sd@...asysnail.net>
Cc: Steffen Klassert <steffen.klassert@...unet.com>,
"David S. Miller" <davem@...emloft.net>,
Eric Dumazet <edumazet@...gle.com>,
Herbert Xu <herbert@...dor.apana.org.au>,
Jakub Kicinski <kuba@...nel.org>, netdev@...r.kernel.org,
Paolo Abeni <pabeni@...hat.com>
Subject: Re: [PATCH ipsec-next v1 1/5] xfrm: delay initialization of offload
path till its actually requested
On Thu, Jun 05, 2025 at 03:09:19PM +0200, Sabrina Dubroca wrote:
> Hello,
>
> I think we need to revert this patch. It causes a severe performance
> regression for SW IPsec (around 40-50%).
>
> 2025-02-19, 15:50:57 +0200, Leon Romanovsky wrote:
> > From: Leon Romanovsky <leonro@...dia.com>
> >
> > XFRM offload path is probed even if offload isn't needed at all. Let's
> > make sure that x->type_offload pointer stays NULL for such path to
> > reduce ambiguity.
>
> x->type_offload is used for GRO with SW IPsec, not just for HW offload.
Thanks for the report, can you please try the following fix?
diff --git a/include/net/xfrm.h b/include/net/xfrm.h
index a21e276dbe447..e45a275fca26d 100644
--- a/include/net/xfrm.h
+++ b/include/net/xfrm.h
@@ -474,7 +474,7 @@ struct xfrm_type_offload {
int xfrm_register_type_offload(const struct xfrm_type_offload *type, unsigned short family);
void xfrm_unregister_type_offload(const struct xfrm_type_offload *type, unsigned short family);
-void xfrm_set_type_offload(struct xfrm_state *x);
+void xfrm_set_type_offload(struct xfrm_state *x, bool try_load);
static inline void xfrm_unset_type_offload(struct xfrm_state *x)
{
if (!x->type_offload)
diff --git a/net/xfrm/xfrm_device.c b/net/xfrm/xfrm_device.c
index 81fd486b5e566..d2819baea414f 100644
--- a/net/xfrm/xfrm_device.c
+++ b/net/xfrm/xfrm_device.c
@@ -305,7 +305,6 @@ int xfrm_dev_state_add(struct net *net, struct xfrm_state *x,
return -EINVAL;
}
- xfrm_set_type_offload(x);
if (!x->type_offload) {
NL_SET_ERR_MSG(extack, "Type doesn't support offload");
dev_put(dev);
diff --git a/net/xfrm/xfrm_state.c b/net/xfrm/xfrm_state.c
index 203b585c2ae28..2da7281f8344b 100644
--- a/net/xfrm/xfrm_state.c
+++ b/net/xfrm/xfrm_state.c
@@ -424,11 +424,10 @@ void xfrm_unregister_type_offload(const struct xfrm_type_offload *type,
}
EXPORT_SYMBOL(xfrm_unregister_type_offload);
-void xfrm_set_type_offload(struct xfrm_state *x)
+void xfrm_set_type_offload(struct xfrm_state *x, bool try_load)
{
const struct xfrm_type_offload *type = NULL;
struct xfrm_state_afinfo *afinfo;
- bool try_load = true;
retry:
afinfo = xfrm_state_get_afinfo(x->props.family);
@@ -607,6 +606,7 @@ static void ___xfrm_state_destroy(struct xfrm_state *x)
kfree(x->coaddr);
kfree(x->replay_esn);
kfree(x->preplay_esn);
+ xfrm_unset_type_offload(x);
if (x->type) {
x->type->destructor(x);
xfrm_put_type(x->type);
@@ -780,8 +780,6 @@ void xfrm_dev_state_free(struct xfrm_state *x)
struct xfrm_dev_offload *xso = &x->xso;
struct net_device *dev = READ_ONCE(xso->dev);
- xfrm_unset_type_offload(x);
-
if (dev && dev->xfrmdev_ops) {
spin_lock_bh(&xfrm_state_dev_gc_lock);
if (!hlist_unhashed(&x->dev_gclist))
diff --git a/net/xfrm/xfrm_user.c b/net/xfrm/xfrm_user.c
index 59f258daf8305..1db18f470f429 100644
--- a/net/xfrm/xfrm_user.c
+++ b/net/xfrm/xfrm_user.c
@@ -977,6 +977,7 @@ static struct xfrm_state *xfrm_state_construct(struct net *net,
/* override default values from above */
xfrm_update_ae_params(x, attrs, 0);
+ xfrm_set_type_offload(x, attrs[XFRMA_OFFLOAD_DEV]);
/* configure the hardware if offload is requested */
if (attrs[XFRMA_OFFLOAD_DEV]) {
err = xfrm_dev_state_add(net, x,
>
> --
> Sabrina
Powered by blists - more mailing lists