Skip to contentSkip to conversation

xsk: Enable shared umem support.

PATCH In review William Tu 2 participants 3 messages Nov 5, 2019 – Nov 6, 2019

Subsystems (5)
lore
Share & tools
lore

William Tu

author
#0 lore raw

PATCH 1/1 xsk: Enable shared umem support.

Currently the shared umem feature is not supported in libbpf. The patch removes the refcount check in libbpf to enable use of shared umem. Also, a umem can be shared by multiple netdevs, so remove the checking at xsk_bind.

Tested using OVS at:
https://mail.openvswitch.org/pipermail/ovs-dev/2019-November/364392.html

---
 net/xdp/xsk.c       |  5 -----
 tools/lib/bpf/xsk.c | 10 +++++-----
 2 files changed, 5 insertions(+), 10 deletions(-)
net/xdp/xsk.c +0 -5
index 6040bc2b0088..0f2b16e275e3 100644
@@ -697,11 +697,6 @@ static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
 			sockfd_put(sock);
 			goto out_unlock;
 		}
-		if (umem_xs->dev != dev || umem_xs->queue_id != qid) {
-			err = -EINVAL;
-			sockfd_put(sock);
-			goto out_unlock;
-		}
 
 		xdp_get_umem(umem_xs->umem);
 		WRITE_ONCE(xs->umem, umem_xs->umem);
tools/lib/bpf/xsk.c +5 -5
index 74d84f36a5b2..e6c4eb077dcd 100644
@@ -579,16 +579,13 @@ int xsk_socket__create(struct xsk_socket **xsk_ptr, const char *ifname,
 	struct sockaddr_xdp sxdp = {};
 	struct xdp_mmap_offsets off;
 	struct xsk_socket *xsk;
+	bool shared;
 	int err;
 
 	if (!umem || !xsk_ptr || !rx || !tx)
 		return -EFAULT;
 
-	if (umem->refcount) {
-		pr_warn("Error: shared umems not supported by libbpf.\n");
-		return -EBUSY;
-	}
-
+	shared = !!(usr_config->bind_flags & XDP_SHARED_UMEM);
 	xsk = calloc(1, sizeof(*xsk));
 	if (!xsk)
 		return -ENOMEM;
@@ -687,6 +684,9 @@ int xsk_socket__create(struct xsk_socket **xsk_ptr, const char *ifname,
 	sxdp.sxdp_queue_id = xsk->queue_id;
 	sxdp.sxdp_flags = xsk->config.bind_flags;
 
+	if (shared)
+		sxdp.sxdp_shared_umem_fd = umem->fd;
+
 	err = bind(xsk->fd, (struct sockaddr *)&sxdp, sizeof(sxdp));
 	if (err) {
 		err = -errno;
Signature · 1 line
2.7.4
#1 lore raw
Quoted 5 lines · William Tu

Currently the shared umem feature is not supported in libbpf. The patch removes the refcount check in libbpf to enable use of shared umem. Also, a umem can be shared by multiple netdevs, so remove the checking at xsk_bind.

Hi William,

I do have a five part patch set sitting on the shelf that implements this as well as a sample, added documentation and support for Rx-only and Tx-only sockets. Let me just rebase it, retest it then submit it to the list and you can see what you think of it. It is along the lines of what you are suggesting here. I am travelling at the moment, so this might not happen until tomorrow.

Structure of the patch set that I will submit:

Patch 1: Adds shared umem support to libbpf
Patch 2: Shared umem support and example XPD program added to xdpsock sample Patch 3: Adds Rx-only and Tx-only support to libbpf
Patch 4: Uses Rx-only sockets for rxdrop and Tx-only sockets for txpush in
the xdpsock sample
Patch 5: Add documentation entries for these two features

Quoted 61 lines

Tested using OVS at:
https://mail.openvswitch.org/pipermail/ovs-dev/2019-November/364392.html

---
 net/xdp/xsk.c       |  5 -----
 tools/lib/bpf/xsk.c | 10 +++++-----
 2 files changed, 5 insertions(+), 10 deletions(-)
net/xdp/xsk.c +0 -5
index 6040bc2b0088..0f2b16e275e3 100644
@@ -697,11 +697,6 @@ static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
                        sockfd_put(sock);
                        goto out_unlock;
                }
-               if (umem_xs->dev != dev || umem_xs->queue_id != qid) {
-                       err = -EINVAL;
-                       sockfd_put(sock);
-                       goto out_unlock;
-               }
 
                xdp_get_umem(umem_xs->umem);
                WRITE_ONCE(xs->umem, umem_xs->umem);
tools/lib/bpf/xsk.c +5 -6
index 74d84f36a5b2..e6c4eb077dcd 100644
@@ -579,16 +579,13 @@ int xsk_socket__create(struct xsk_socket **xsk_ptr, const char *ifname,
        struct sockaddr_xdp sxdp = {};
        struct xdp_mmap_offsets off;
        struct xsk_socket *xsk;
+       bool shared;
        int err;
 
        if (!umem || !xsk_ptr || !rx || !tx)
                return -EFAULT;
 
-       if (umem->refcount) {
-               pr_warn("Error: shared umems not supported by libbpf.\n");
-               return -EBUSY;
-       }
-
+       shared = !!(usr_config->bind_flags & XDP_SHARED_UMEM);
        xsk = calloc(1, sizeof(*xsk));
        if (!xsk)
                return -ENOMEM;
@@ -687,6 +684,9 @@ int xsk_socket__create(struct xsk_socket **xsk_ptr, const char *ifname,
        sxdp.sxdp_queue_id = xsk->queue_id;
        sxdp.sxdp_flags = xsk->config.bind_flags;
 
+       if (shared)
+               sxdp.sxdp_shared_umem_fd = umem->fd;
+
        err = bind(xsk->fd, (struct sockaddr *)&sxdp, sizeof(sxdp));
        if (err) {
                err = -errno;
--

2.7.4

author
#2 lore raw
Quoted 16 lines · Magnus Karlsson
Quoted 5 lines · William Tu

Currently the shared umem feature is not supported in libbpf. The patch removes the refcount check in libbpf to enable use of shared umem. Also, a umem can be shared by multiple netdevs, so remove the checking at xsk_bind.

Hi William,

I do have a five part patch set sitting on the shelf that implements this as well as a sample, added documentation and support for Rx-only and Tx-only sockets. Let me just rebase it, retest it then submit it to the list and you can see what you think of it. It is along the lines of what you are suggesting here. I am travelling at the moment, so this might not happen until tomorrow.

Quoted 9 lines

Structure of the patch set that I will submit:

Patch 1: Adds shared umem support to libbpf
Patch 2: Shared umem support and example XPD program added to xdpsock sample Patch 3: Adds Rx-only and Tx-only support to libbpf
Patch 4: Uses Rx-only sockets for rxdrop and Tx-only sockets for txpush in
the xdpsock sample
Patch 5: Add documentation entries for these two features

Hi Magnus,

Thank you. I will wait for your patch set.

William