Fix the problem that rxe can not work in net namespace
Subsystems (2)
v6 cover Fix the problem that rxe can not work in net namespace
From: Zhu Yanjun <[email protected]>
When run "ip link add" command to add a rxe rdma link in a net namespace, normally this rxe rdma link can not work in a net name space.
The root cause is that a sock listening on udp port 4791 is created in init_net when the rdma_rxe module is loaded into kernel. That is, the sock listening on udp port 4791 is created in init_net. Other net namespace is difficult to use this sock.
The following commits will solve this problem.
In the first commit, move the creating sock listening on udp port 4791 from module_init function to rdma link creating functions. That is, after the module rdma_rxe is loaded, the sock will not be created. When run "rdma link add ..." command, the sock will be created. So when creating a rdma link in the net namespace, the sock will be created in this net namespace.
In the second commit, the functions udp4_lib_lookup and udp6_lib_lookup will check the sock exists in the net namespace or not. If yes, rdma link will increase the reference count of this sock, then continue other jobs instead of creating a new sock to listen on udp port 4791. Since the network notifier is global, when the module rdma_rxe is loaded, this notifier will be registered.
After the rdma link is created, the command "rdma link del" is to delete rdma link at the same time the sock is checked. If the reference count of this sock is greater than the sock reference count needed by udp tunnel, the sock reference count is decreased by one. If equal, it indicates that this rdma link is the last one. As such, the udp tunnel is shut down and the sock is closed. The above work should be implemented in linkdel function. But currently no dellink function in rxe. So the 3rd commit addes dellink function pointer. And the 4th commit implements the dellink function in rxe.
To now, it is not necessary to keep a global variable to store the sock listening udp port 4791. This global variable can be replaced by the functions udp4_lib_lookup and udp6_lib_lookup totally. Because the function udp6_lib_lookup is in the fast path, a member variable l_sk6 is added to store the sock. If l_sk6 is NULL, udp6_lib_lookup is called to lookup the sock, then the sock is stored in l_sk6, in the future,it can be used directly.
All the above work has been done in init_net. And it can also work in the net namespace. So the init_net is replaced by the individual net namespace. This is what the 6th commit does. Because rxe device is dependent on the net device and the sock listening on udp port 4791, every rxe device is in exclusive mode in the individual net namespace. Other rdma netns operations will be considerred in the future.
In the 7th commit, the register_pernet_subsys/unregister_pernet_subsys functions are added. When a new net namespace is created, the init function will initialize the sk4 and sk6 socks. Then the 2 socks will be released when the net namespace is destroyed. The functions rxe_ns_pernet_sk4/rxe_ns_pernet_set_sk4 will get and set sk4 in the net namespace. The functions rxe_ns_pernet_sk6/rxe_ns_pernet_set_sk6 will handle sk6. Then sk4 and sk6 are used in the previous commits.
As the sk4 and sk6 in pernet namespace can be accessed, it is not necessary to add a new l_sk6. As such, in the 8th commit, the l_sk6 is replaced with the sk6 in pernet namespace.
Test steps:
1) Suppose that 2 NICs are in 2 different net namespaces.
# ip netns exec net0 ip link
3: eno2: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc mq state UP
link/ether 00:1e:67:a0:22:3f brd ff:ff:ff:ff:ff:ff
altname enp5s0 # ip netns exec net1 ip link
4: eno3: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc fq_codel
link/ether f8:e4:3b:3b:e4:10 brd ff:ff:ff:ff:ff:ff2) Add rdma link in the different net namespace
net0:
# ip netns exec net0 rdma link add rxe0 type rxe netdev eno2
net1:
# ip netns exec net1 rdma link add rxe1 type rxe netdev eno3
3) Run rping test.
net0
# ip netns exec net0 rping -s -a 192.168.2.1 -C 1&
[1] 1737
# ip netns exec net1 rping -c -a 192.168.2.1 -d -v -C 1
verbose
count 1
...
ping data: rdma-ping-0: ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqr
...4) Remove the rdma links from the net namespaces.
net0:
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
UNCONN 0 0 0.0.0.0:4791 0.0.0.0:*
UNCONN 0 0 [::]:4791 [::]:*
# ip netns exec net0 rdma link del rxe0
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
net1:
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
UNCONN 0 0 0.0.0.0:4791 0.0.0.0:*
UNCONN 0 0 [::]:4791 [::]:*
# ip netns exec net1 rdma link del rxe1
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
V5->V6: Fix resource leak problem when rxe_net_init fails. And fix some
style problems.
V4->V5: Rebase the commits to V6.4-rc1
V3->V4: Rebase the commits to rdma-next;
V2->V3: 1) Add "rdma link del" example in the cover letter, and use "ss -lu" to
verify rdma link is removed.
2) Add register_pernet_subsys/unregister_pernet_subsys net namespace
3) Replace l_sk6 with sk6 of pernet_name_space
V1->V2: Add the explicit initialization of sk6.
Zhu Yanjun (8):
RDMA/rxe: Creating listening sock in newlink function RDMA/rxe: Support more rdma links in init_net RDMA/nldev: Add dellink function pointer
RDMA/rxe: Implement dellink in rxe
RDMA/rxe: Replace global variable with sock lookup functions RDMA/rxe: add the support of net namespace RDMA/rxe: Add the support of net namespace notifier RDMA/rxe: Replace l_sk6 with sk6 in net namespace
drivers/infiniband/core/nldev.c | 6 ++ drivers/infiniband/sw/rxe/Makefile | 3 +- drivers/infiniband/sw/rxe/rxe.c | 35 +++++++- drivers/infiniband/sw/rxe/rxe_net.c | 119 ++++++++++++++++++------ drivers/infiniband/sw/rxe/rxe_net.h | 9 +- drivers/infiniband/sw/rxe/rxe_ns.c | 134 ++++++++++++++++++++++++++++ drivers/infiniband/sw/rxe/rxe_ns.h | 17 ++++ include/rdma/rdma_netlink.h | 2 + 8 files changed, 285 insertions(+), 40 deletions(-) create mode 100644 drivers/infiniband/sw/rxe/rxe_ns.c create mode 100644 drivers/infiniband/sw/rxe/rxe_ns.h
Signature · 1 line
2.27.0
PATCH v6 1/8 RDMA/rxe: Creating listening sock in newlink function
From: Zhu Yanjun <[email protected]>
Originally when the module rdma_rxe is loaded, the sock listening on udp port 4791 is created. Currently moving the creating listening port to newlink function.
So when running "rdma link add" command, the sock listening on udp port 4791 is created.
--- drivers/infiniband/sw/rxe/rxe.c | 10 ++++------ 1 file changed, 4 insertions(+), 6 deletions(-)
drivers/infiniband/sw/rxe/rxe.c +4 -6
@@ -194,6 +194,10 @@ static int rxe_newlink(const char *ibdev_name, struct net_device *ndev) goto err; } + err =rxe_net_init (); +if (err) +return err; + err =rxe_net_add (ibdev_name, ndev);if (err) {rxe_err ("failed to add %s\n" , ndev->name);
@@ -210,12 +214,6 @@ static struct rdma_link_ops rxe_link_ops = {static int __init rxe_module_init (void ) { -int err; - - err =rxe_net_init (); -if (err) -return err; -rdma_link_register (&rxe_link_ops);pr_info ("loaded\n" );return 0 ;
Signature · 1 line
2.27.0
PATCH v6 2/8 RDMA/rxe: Support more rdma links in init_net
In reply to Zhu YanjunFrom: Zhu Yanjun <[email protected]>
In init_net, when several rdma links are created with the command "rdma link add", newlink will check whether the udp port 4791 is listening or not.
If not, creating a sock listening on udp port 4791. If yes, increasing the reference count of the sock.
--- drivers/infiniband/sw/rxe/rxe.c | 12 ++++++- drivers/infiniband/sw/rxe/rxe_net.c | 55 +++++++++++++++++++++-------- drivers/infiniband/sw/rxe/rxe_net.h | 1 + 3 files changed, 52 insertions(+), 16 deletions(-)
drivers/infiniband/sw/rxe/rxe.c +11 -1
@@ -8,6 +8,7 @@#include <net/addrconf.h> #include "rxe.h" #include "rxe_loc.h" +#include "rxe_net.h" MODULE_AUTHOR ("Bob Pearson, Frank Zago, John Groves, Kamal Heib" );MODULE_DESCRIPTION ("Soft RDMA transport" );
@@ -207,14 +208,23 @@ static int rxe_newlink(const char *ibdev_name, struct net_device *ndev)return err; } -static struct rdma_link_ops rxe_link_ops = { +struct rdma_link_ops rxe_link_ops = { .type ="rxe" , .newlink = rxe_newlink, };static int __init rxe_module_init (void ) { +int err; +rdma_link_register (&rxe_link_ops); + + err =rxe_register_notifier (); +if (err) { +pr_err ("Failed to register netdev notifier\n" ); +return -1 ; + } +pr_info ("loaded\n" );return 0 ; }
drivers/infiniband/sw/rxe/rxe_net.c +40 -15
96 lines not on this page · raw message
drivers/infiniband/sw/rxe/rxe_net.h +1 -0
@@ -18,6 +18,7 @@ struct rxe_recv_sockets {int rxe_net_add (const char *ibdev_name,struct net_device *ndev); +int rxe_register_notifier (void );int rxe_net_init (void );void rxe_net_exit (void );
Signature · 1 line
2.27.0
PATCH v6 3/8 RDMA/nldev: Add dellink function pointer
In reply to Zhu YanjunFrom: Zhu Yanjun <[email protected]>
The newlink function pointer is added. And the sock listening on port 4791 is added in the newlink function. So the dellink function is needed to remove the sock.
--- drivers/infiniband/core/nldev.c | 6 ++++++ include/rdma/rdma_netlink.h | 2 ++ 2 files changed, 8 insertions(+)
drivers/infiniband/core/nldev.c +6 -0
@@ -1758,6 +1758,12 @@ static int nldev_dellink(struct sk_buff *skb, struct nlmsghdr *nlh,return -EINVAL ; } +if (device->link_ops) { + err = device->link_ops->dellink (device); +if (err) +return err; + } +ib_unregister_device_and_put (device);return 0 ; }
include/rdma/rdma_netlink.h +2 -0
@@ -5,6 +5,7 @@#include <linux/netlink.h> #include <uapi/rdma/rdma_netlink.h> +#include <rdma/ib_verbs.h> enum {RDMA_NLDEV_ATTR_EMPTY_STRING =1 ,
@@ -114,6 +115,7 @@ struct rdma_link_ops {struct list_head list;const char *type;int (*newlink)(const char *ibdev_name,struct net_device *ndev); +int (*dellink)(struct ib_device *dev); };void rdma_link_register (struct rdma_link_ops *ops);
Signature · 1 line
2.27.0
RE: [PATCH v6 3/8] RDMA/nldev: Add dellink function pointer
Quoted 31 lines
-----Original Message-----
From: Zhu Yanjun <[email protected]>
Sent: Friday, 23 June 2023 11:58
To: [email protected]; [email protected]; [email protected]; linux- [email protected]; [email protected]; [email protected]; [email protected]
Cc: Zhu Yanjun <[email protected]>; Rain River <[email protected]>
Subject: [EXTERNAL] [PATCH v6 3/8] RDMA/nldev: Add dellink function pointerFrom: Zhu Yanjun <[email protected]>
The newlink function pointer is added. And the sock listening on port 4791 is added in the newlink function. So the dellink function is needed to remove the sock.
Tested-by Rain River <[email protected]>Signed-off-by Zhu Yanjun <[email protected]>--- drivers/infiniband/core/nldev.c | 6 ++++++ include/rdma/rdma_netlink.h | 2 ++ 2 files changed, 8 insertions(+)drivers/infiniband/core/nldev.c +0
-0
drivers/infiniband/core/nldev.c static int nldev_dellink(struct sk_buff *skb, struct · line 1763
Quoted 6 lines
drivers/infiniband/core/nldev.c +2
-0@@ -1758,6 +1758,12 @@ static int nldev_dellink(struct sk_buff *skb, struct nlmsghdr *nlh, return -EINVAL ; } +if (device->link_ops) { + err = device->link_ops->dellink (device);
That assumes delink() being populated by all drivers having link_ops in place. It crashes for drivers like siw.
Quoted 29 lines
drivers/infiniband/core/nldev.c +4
-0+if (err) +return err; + } +ib_unregister_device_and_put (device);return 0 ;}
include/rdma/rdma_netlink.h +2
-1@@ -5,6 +5,7 @@#include <linux/netlink.h> #include <uapi/rdma/rdma_netlink.h> +#include <rdma/ib_verbs.h> enum {RDMA_NLDEV_ATTR_EMPTY_STRING =1 ,@@ -114,6 +115,7 @@ struct rdma_link_ops {struct list_head list;const char *type;int (*newlink)(const char *ibdev_name,struct net_device *ndev); +int (*dellink)(struct ib_device *dev); };void rdma_link_register (struct rdma_link_ops *ops); --2.27.0
Re: RE: [PATCH v6 3/8] RDMA/nldev: Add dellink function pointer
Quoted 35 lines · Bernard Metzler
Quoted 35 lines
-----Original Message-----
From: Zhu Yanjun <[email protected]>
Sent: Friday, 23 June 2023 11:58
To: [email protected]; [email protected]; [email protected]; linux- [email protected]; [email protected]; [email protected]; [email protected]
Cc: Zhu Yanjun <[email protected]>; Rain River <[email protected]>
Subject: [EXTERNAL] [PATCH v6 3/8] RDMA/nldev: Add dellink function pointerFrom: Zhu Yanjun <[email protected]>
The newlink function pointer is added. And the sock listening on port 4791 is added in the newlink function. So the dellink function is needed to remove the sock.
Tested-by Rain River <[email protected]>Signed-off-by Zhu Yanjun <[email protected]>--- drivers/infiniband/core/nldev.c | 6 ++++++ include/rdma/rdma_netlink.h | 2 ++ 2 files changed, 8 insertions(+)drivers/infiniband/core/nldev.c +0
-0@@ -1758,6 +1758,12 @@ static int nldev_dellink(struct sk_buff *skb, struct nlmsghdr *nlh, return -EINVAL ; }
Quoted 6 lines
drivers/infiniband/core/nldev.c static int nldev_dellink(struct sk_buff *skb, struct nlmsghdr *nlh, · line 1762
Quoted 2 lines
drivers/infiniband/core/nldev.c +2
-0+if (device->link_ops) { + err = device->link_ops->dellink (device);That assumes
delink()being populated by all drivers havinglink_opsin place. It crashes for drivers like siw.
Good catch. Thanks, I will fix it.
Zhu Yanjun
Quoted 32 lines
Quoted 31 lines
drivers/infiniband/core/nldev.c +4
-0+if (err) +return err; + } +ib_unregister_device_and_put (device);return 0 ;}include/rdma/rdma_netlink.h +2
-1@@ -5,6 +5,7 @@#include <linux/netlink.h> #include <uapi/rdma/rdma_netlink.h> +#include <rdma/ib_verbs.h> enum {RDMA_NLDEV_ATTR_EMPTY_STRING =1 ,@@ -114,6 +115,7 @@ struct rdma_link_ops {struct list_head list;const char *type;int (*newlink)(const char *ibdev_name,struct net_device *ndev); +int (*dellink)(struct ib_device *dev); };void rdma_link_register (struct rdma_link_ops *ops); --2.27.0
PATCH v6 4/8 RDMA/rxe: Implement dellink in rxe
In reply to Zhu YanjunFrom: Zhu Yanjun <[email protected]>
When running "rdma link del" command, dellink function will be called. If the sock refcnt is greater than the refcnt needed for udp tunnel, the sock refcnt will be decreased by 1.
If equal, the last rdma link is removed. The udp tunnel will be destroyed.
--- drivers/infiniband/sw/rxe/rxe.c | 12 +++++++++++- drivers/infiniband/sw/rxe/rxe_net.c | 17 +++++++++++++++-- drivers/infiniband/sw/rxe/rxe_net.h | 1 + 3 files changed, 27 insertions(+), 3 deletions(-)
drivers/infiniband/sw/rxe/rxe.c +11 -1
@@ -168,10 +168,12 @@ void rxe_set_mtu(struct rxe_dev *rxe, unsigned int ndev_mtu)/* called by ifc layer to create new rxe device. * The caller should allocate memory for rxe by calling ib_alloc_device. */ +static struct rdma_link_ops rxe_link_ops;int rxe_add (struct rxe_dev *rxe,unsigned int mtu,const char *ibdev_name) {rxe_init (rxe);rxe_set_mtu (rxe, mtu); + rxe->ib_dev.link_ops = &rxe_link_ops;return rxe_register_device (rxe, ibdev_name); }
@@ -208,9 +210,17 @@ static int rxe_newlink(const char *ibdev_name, struct net_device *ndev)return err; } -struct rdma_link_ops rxe_link_ops = { +static int rxe_dellink (struct ib_device *dev) +{ +rxe_net_del (dev); + +return 0 ; +} + +static struct rdma_link_ops rxe_link_ops = { .type ="rxe" , .newlink = rxe_newlink, + .dellink = rxe_dellink, };static int __init rxe_module_init (void )
drivers/infiniband/sw/rxe/rxe_net.c +15 -2
@@ -534,6 +534,21 @@ int rxe_net_add(const char *ibdev_name, struct net_device *ndev)return 0 ; } +#define SK_REF_FOR_TUNNEL 2 +void rxe_net_del (struct ib_device *dev) +{ +if (refcount_read (&recv_sockets.sk6->sk->sk_refcnt) >SK_REF_FOR_TUNNEL ) +__sock_put (recv_sockets.sk6->sk); +else +rxe_release_udp_tunnel (recv_sockets.sk6); + +if (refcount_read (&recv_sockets.sk4->sk->sk_refcnt) >SK_REF_FOR_TUNNEL ) +__sock_put (recv_sockets.sk4->sk); +else +rxe_release_udp_tunnel (recv_sockets.sk4); +} +#undef SK_REF_FOR_TUNNEL +static void rxe_port_event (struct rxe_dev *rxe,enum ib_event_type event) {
@@ -693,8 +708,6 @@ int rxe_register_notifier(void)void rxe_net_exit (void ) { -rxe_release_udp_tunnel (recv_sockets.sk6); -rxe_release_udp_tunnel (recv_sockets.sk4);unregister_netdevice_notifier (&rxe_net_notifier); }
drivers/infiniband/sw/rxe/rxe_net.h +1 -0
@@ -17,6 +17,7 @@ struct rxe_recv_sockets { };int rxe_net_add (const char *ibdev_name,struct net_device *ndev); +void rxe_net_del (struct ib_device *dev);int rxe_register_notifier (void );int rxe_net_init (void );
Signature · 1 line
2.27.0
PATCH v6 5/8 RDMA/rxe: Replace global variable with sock lookup functions
In reply to Zhu YanjunFrom: Zhu Yanjun <[email protected]>
Originally a global variable is to keep the sock of udp listening on port 4791. In fact, sock lookup functions can be used to get the sock.
--- drivers/infiniband/sw/rxe/rxe.c | 1 + drivers/infiniband/sw/rxe/rxe_net.c | 58 ++++++++++++++++++++------- drivers/infiniband/sw/rxe/rxe_net.h | 5 --- drivers/infiniband/sw/rxe/rxe_verbs.h | 1 + 4 files changed, 45 insertions(+), 20 deletions(-)
drivers/infiniband/sw/rxe/rxe.c +1 -0
@@ -74,6 +74,7 @@ static void rxe_init_device_param(struct rxe_dev *rxe) rxe->ndev->dev_addr); rxe->max_ucontext =RXE_MAX_UCONTEXT ; + rxe->l_sk6 =NULL ; }/* initialize port attributes */
drivers/infiniband/sw/rxe/rxe_net.c +43 -15
107 lines not on this page · raw message
drivers/infiniband/sw/rxe/rxe_net.h +0 -5
@@ -11,11 +11,6 @@#include <net/if_inet6.h> #include <linux/module.h> -struct rxe_recv_sockets { -struct socket *sk4; -struct socket *sk6; -}; -int rxe_net_add (const char *ibdev_name,struct net_device *ndev);void rxe_net_del (struct ib_device *dev);
drivers/infiniband/sw/rxe/rxe_verbs.h +1 -0
@@ -382,6 +382,7 @@ struct rxe_dev {struct rxe_port port;struct crypto_shash *tfm; +struct socket *l_sk6; };static inline void rxe_counter_inc (struct rxe_dev *rxe,enum rxe_counters index)
Signature · 1 line
2.27.0
PATCH v6 6/8 RDMA/rxe: add the support of net namespace
In reply to Zhu YanjunFrom: Zhu Yanjun <[email protected]>
Originally init_net is used to indicate the current net namespace. Currently more net namespaces are supported.
--- drivers/infiniband/sw/rxe/rxe.c | 2 +- drivers/infiniband/sw/rxe/rxe_net.c | 33 +++++++++++++++++------------ drivers/infiniband/sw/rxe/rxe_net.h | 2 +- 3 files changed, 22 insertions(+), 15 deletions(-)
drivers/infiniband/sw/rxe/rxe.c +1 -1
@@ -198,7 +198,7 @@ static int rxe_newlink(const char *ibdev_name, struct net_device *ndev) goto err; } - err =rxe_net_init (); + err =rxe_net_init (ndev);if (err)return err;
drivers/infiniband/sw/rxe/rxe_net.c +20 -13
@@ -32,7 +32,7 @@ static struct dst_entry *rxe_find_route4(struct rxe_qp *qp,memcpy (&fl.daddr, daddr,sizeof (*daddr)); fl.flowi4_proto =IPPROTO_UDP ; - rt =ip_route_output_key (&init_net, &fl); + rt =ip_route_output_key (dev_net (ndev), &fl);if (IS_ERR (rt)) {rxe_dbg_qp (qp,"no route to %pI4\n" , &daddr->s_addr);return NULL ;
@@ -56,7 +56,8 @@ static struct dst_entry *rxe_find_route6(struct rxe_qp *qp,struct sock *sk;rcu_read_lock (); - sk =udp6_lib_lookup (&init_net,NULL ,0 , &in6addr_any,htons (ROCE_V2_UDP_DPORT ),0 ); + sk =udp6_lib_lookup (dev_net (ndev),NULL ,0 , &in6addr_any, +htons (ROCE_V2_UDP_DPORT ),0 );rcu_read_unlock ();if (!sk) {rxe_dbg_qp (qp,"file: %s +%d, error\n" , __FILE__, __LINE__);
@@ -553,9 +554,13 @@ int rxe_net_add(const char *ibdev_name, struct net_device *ndev)void rxe_net_del (struct ib_device *dev) {struct sock *sk; +struct rxe_dev *rxe; + + rxe = container_of(dev, struct rxe_dev, ib_dev);rcu_read_lock (); - sk =udp4_lib_lookup (&init_net,0 ,0 ,htonl (INADDR_ANY ),htons (ROCE_V2_UDP_DPORT ),0 ); + sk =udp4_lib_lookup (dev_net (rxe->ndev),0 ,0 ,htonl (INADDR_ANY ), +htons (ROCE_V2_UDP_DPORT ),0 );rcu_read_unlock ();if (!sk)return ;
@@ -568,7 +573,8 @@ void rxe_net_del(struct ib_device *dev)rxe_release_udp_tunnel (sk->sk_socket);rcu_read_lock (); - sk =udp6_lib_lookup (&init_net,NULL ,0 , &in6addr_any,htons (ROCE_V2_UDP_DPORT ),0 ); + sk =udp6_lib_lookup (dev_net (rxe->ndev),NULL ,0 , &in6addr_any, +htons (ROCE_V2_UDP_DPORT ),0 );rcu_read_unlock ();if (!sk)return ;
@@ -640,6 +646,7 @@ static int rxe_notify(struct notifier_block *not_blk,switch (event) {case NETDEV_UNREGISTER :ib_unregister_device_queued (&rxe->ib_dev); +rxe_net_del (&rxe->ib_dev);break ;case NETDEV_UP :rxe_port_up (rxe);
@@ -673,19 +680,19 @@ static struct notifier_block rxe_net_notifier = { .notifier_call = rxe_notify, }; -static int rxe_net_ipv4_init (void ) +static int rxe_net_ipv4_init (struct net_device *ndev) {struct sock *sk;struct socket *sock;rcu_read_lock (); - sk =udp4_lib_lookup (&init_net,0 ,0 ,htonl (INADDR_ANY ), + sk =udp4_lib_lookup (dev_net (ndev),0 ,0 ,htonl (INADDR_ANY ),htons (ROCE_V2_UDP_DPORT ),0 );rcu_read_unlock ();if (sk)return 0 ; - sock =rxe_setup_udp_tunnel (&init_net,htons (ROCE_V2_UDP_DPORT ), false); + sock =rxe_setup_udp_tunnel (dev_net (ndev),htons (ROCE_V2_UDP_DPORT ), false);if (IS_ERR (sock)) {pr_err ("Failed to create IPv4 UDP tunnel\n" );return -1 ;
@@ -694,20 +701,20 @@ static int rxe_net_ipv4_init(void)return 0 ; } -static int rxe_net_ipv6_init (void ) +static int rxe_net_ipv6_init (struct net_device *ndev) {#if IS_ENABLED (CONFIG_IPV6 )struct sock *sk;struct socket *sock;rcu_read_lock (); - sk =udp6_lib_lookup (&init_net,NULL ,0 , &in6addr_any, + sk =udp6_lib_lookup (dev_net (ndev),NULL ,0 , &in6addr_any,htons (ROCE_V2_UDP_DPORT ),0 );rcu_read_unlock ();if (sk)return 0 ; - sock =rxe_setup_udp_tunnel (&init_net,htons (ROCE_V2_UDP_DPORT ), true); + sock =rxe_setup_udp_tunnel (dev_net (ndev),htons (ROCE_V2_UDP_DPORT ), true);if (PTR_ERR (sock) == -EAFNOSUPPORT ) {pr_warn ("IPv6 is not supported, can not create a UDPv6 socket\n" );return 0 ;
@@ -739,14 +746,14 @@ void rxe_net_exit(void)unregister_netdevice_notifier (&rxe_net_notifier); } -int rxe_net_init (void ) +int rxe_net_init (struct net_device *ndev) {int err; - err =rxe_net_ipv4_init (); + err =rxe_net_ipv4_init (ndev);if (err)return err; - err =rxe_net_ipv6_init (); + err =rxe_net_ipv6_init (ndev);if (err) goto err_out;return 0 ;
drivers/infiniband/sw/rxe/rxe_net.h +1 -1
@@ -15,7 +15,7 @@ int rxe_net_add(const char *ibdev_name, struct net_device *ndev);void rxe_net_del (struct ib_device *dev);int rxe_register_notifier (void ); -int rxe_net_init (void ); +int rxe_net_init (struct net_device *ndev);void rxe_net_exit (void );#endif /* RXE_NET_H */
Signature · 1 line
2.27.0
PATCH v6 7/8 RDMA/rxe: Add the support of net namespace notifier
In reply to Zhu YanjunFrom: Zhu Yanjun <[email protected]>
The functions register_pernet_subsys/unregister_pernet_subsys register a notifier of net namespace. When a new net namespace is created, the init function of rxe will be called to initialize sk4 and sk6 socks. When a net namespace is destroyed, the exit function will be called to handle sk4 and sk6 socks.
The functions rxe_ns_pernet_sk4 and rxe_ns_pernet_sk6 are used to get sk4 and sk6 socks.
The functions rxe_ns_pernet_set_sk4 and rxe_ns_pernet_set_sk6 are used to set sk4 and sk6 socks.
--- drivers/infiniband/sw/rxe/Makefile | 3 +- drivers/infiniband/sw/rxe/rxe.c | 9 ++ drivers/infiniband/sw/rxe/rxe_net.c | 56 ++++++------ drivers/infiniband/sw/rxe/rxe_ns.c | 134 ++++++++++++++++++++++++++++ drivers/infiniband/sw/rxe/rxe_ns.h | 17 ++++ 5 files changed, 193 insertions(+), 26 deletions(-) create mode 100644 drivers/infiniband/sw/rxe/rxe_ns.c create mode 100644 drivers/infiniband/sw/rxe/rxe_ns.h
drivers/infiniband/sw/rxe/Makefile +2 -1
@@ -22,4 +22,5 @@ rdma_rxe-y := \ rxe_mcast.o \ rxe_task.o \ rxe_net.o \ - rxe_hw_counters.o + rxe_hw_counters.o \ + rxe_ns.o
drivers/infiniband/sw/rxe/rxe.c +9 -0
@@ -9,6 +9,7 @@#include "rxe.h" #include "rxe_loc.h" #include "rxe_net.h" +#include "rxe_ns.h" MODULE_AUTHOR ("Bob Pearson, Frank Zago, John Groves, Kamal Heib" );MODULE_DESCRIPTION ("Soft RDMA transport" );
@@ -236,6 +237,12 @@ static int __init rxe_module_init(void)return -1 ; } + err =rxe_namespace_init (); +if (err) { +pr_err ("Failed to register net namespace notifier\n" ); +return -1 ; + } +pr_info ("loaded\n" );return 0 ; }
@@ -246,6 +253,8 @@ static void __exit rxe_module_exit(void)ib_unregister_driver (RDMA_DRIVER_RXE );rxe_net_exit (); +rxe_namespace_exit (); +pr_info ("unloaded\n" ); }
drivers/infiniband/sw/rxe/rxe_net.c +31 -25
117 lines not on this page · raw message
drivers/infiniband/sw/rxe/rxe_ns.c new file +134 -0
134 lines not on this page · raw message
drivers/infiniband/sw/rxe/rxe_ns.h new file +17 -0
@@ -0,0 +1,17 @@ +// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB +/* +* Copyright (c) 2016 Mellanox Technologies Ltd. All rights reserved. +* Copyright (c) 2015 System Fabric Works, Inc. All rights reserved. +*/ + +#ifndef RXE_NS_H +#define RXE_NS_H + +struct sock *rxe_ns_pernet_sk4 (struct net *net); +struct sock *rxe_ns_pernet_sk6 (struct net *net); +void rxe_ns_pernet_set_sk4 (struct net *net,struct sock *sk); +void rxe_ns_pernet_set_sk6 (struct net *net,struct sock *sk); +int __init rxe_namespace_init (void ); +void __exit rxe_namespace_exit(void); + +#endif /* RXE_NS_H */
Signature · 1 line
2.27.0
PATCH v6 8/8 RDMA/rxe: Replace l_sk6 with sk6 in net namespace
In reply to Zhu YanjunFrom: Zhu Yanjun <[email protected]>
The net namespace variable sk6 can be used. As such, l_sk6 can be replaced with it.
--- drivers/infiniband/sw/rxe/rxe.c | 1 - drivers/infiniband/sw/rxe/rxe_net.c | 20 +------------------- drivers/infiniband/sw/rxe/rxe_verbs.h | 1 - 3 files changed, 1 insertion(+), 21 deletions(-)
drivers/infiniband/sw/rxe/rxe.c +0 -1
@@ -75,7 +75,6 @@ static void rxe_init_device_param(struct rxe_dev *rxe) rxe->ndev->dev_addr); rxe->max_ucontext =RXE_MAX_UCONTEXT ; - rxe->l_sk6 =NULL ; }/* initialize port attributes */
drivers/infiniband/sw/rxe/rxe_net.c +1 -19
@@ -50,24 +50,6 @@ static struct dst_entry *rxe_find_route6(struct rxe_qp *qp, {struct dst_entry *ndst;struct flowi6 fl6 = { {0 } }; -struct rxe_dev *rxe; - - rxe =rxe_get_dev_from_net (ndev); -if (!rxe->l_sk6) { -struct sock *sk; - -rcu_read_lock (); - sk =udp6_lib_lookup (dev_net (ndev),NULL ,0 , &in6addr_any, -htons (ROCE_V2_UDP_DPORT ),0 ); -rcu_read_unlock (); -if (!sk) { -rxe_dbg_qp (qp,"file: %s +%d, error\n" , __FILE__, __LINE__); -return (struct dst_entry *)sk; - } -__sock_put (sk); - rxe->l_sk6 = sk->sk_socket; - } -memset (&fl6,0 ,sizeof (fl6)); fl6.flowi6_oif = ndev->ifindex;
@@ -76,7 +58,7 @@ static struct dst_entry *rxe_find_route6(struct rxe_qp *qp, fl6.flowi6_proto =IPPROTO_UDP ; ndst = ipv6_stub->ipv6_dst_lookup_flow (dev_net (ndev), - rxe->l_sk6->sk, &fl6, +rxe_ns_pernet_sk6 (dev_net (ndev)), &fl6,NULL );if (IS_ERR (ndst)) {rxe_dbg_qp (qp,"no route to %pI6\n" , daddr);
drivers/infiniband/sw/rxe/rxe_verbs.h +0 -1
@@ -382,7 +382,6 @@ struct rxe_dev {struct rxe_port port;struct crypto_shash *tfm; -struct socket *l_sk6; };static inline void rxe_counter_inc (struct rxe_dev *rxe,enum rxe_counters index)
Signature · 1 line
2.27.0
Quoted 141 lines · Zhu Yanjun
From: Zhu Yanjun <[email protected]>
When run "ip link add" command to add a rxe rdma link in a net namespace, normally this rxe rdma link can not work in a net name space.
The root cause is that a sock listening on udp port 4791 is created in
init_netwhen therdma_rxemodule is loaded into kernel. That is, the sock listening on udp port 4791 is created ininit_net. Other net namespace is difficult to use this sock.The following commits will solve this problem.
In the first commit, move the creating sock listening on udp port 4791 from
module_initfunction to rdma link creating functions. That is, after the modulerdma_rxeis loaded, the sock will not be created. When run "rdma link add ..." command, the sock will be created. So when creating a rdma link in the net namespace, the sock will be created in this net namespace.In the second commit, the functions
udp4_lib_lookupandudp6_lib_lookupwill check the sock exists in the net namespace or not. If yes, rdma link will increase the reference count of this sock, then continue other jobs instead of creating a new sock to listen on udp port 4791. Since the network notifier is global, when the modulerdma_rxeis loaded, this notifier will be registered.After the rdma link is created, the command "rdma link del" is to delete rdma link at the same time the sock is checked. If the reference count of this sock is greater than the sock reference count needed by udp tunnel, the sock reference count is decreased by one. If equal, it indicates that this rdma link is the last one. As such, the udp tunnel is shut down and the sock is closed. The above work should be implemented in linkdel function. But currently no dellink function in rxe. So the 3rd commit addes dellink function pointer. And the 4th commit implements the dellink function in rxe.
To now, it is not necessary to keep a global variable to store the sock listening udp port 4791. This global variable can be replaced by the functions
udp4_lib_lookupandudp6_lib_lookuptotally. Because the functionudp6_lib_lookupis in the fast path, a member variablel_sk6is added to store the sock. Ifl_sk6is NULL,udp6_lib_lookupis called to lookup the sock, then the sock is stored inl_sk6, in the future,it can be used directly.All the above work has been done in
init_net. And it can also work in the net namespace. So theinit_netis replaced by the individual net namespace. This is what the 6th commit does. Because rxe device is dependent on the net device and the sock listening on udp port 4791, every rxe device is in exclusive mode in the individual net namespace. Other rdma netns operations will be considerred in the future.In the 7th commit, the register_pernet_subsys/unregister_pernet_subsys functions are added. When a new net namespace is created, the init function will initialize the sk4 and sk6 socks. Then the 2 socks will be released when the net namespace is destroyed. The functions rxe_ns_pernet_sk4/rxe_ns_pernet_set_sk4 will get and set sk4 in the net namespace. The functions rxe_ns_pernet_sk6/rxe_ns_pernet_set_sk6 will handle sk6. Then sk4 and sk6 are used in the previous commits.
As the sk4 and sk6 in pernet namespace can be accessed, it is not necessary to add a new
l_sk6. As such, in the 8th commit, thel_sk6is replaced with the sk6 in pernet namespace.Test steps:
1) Suppose that 2 NICs are in 2 different net namespaces.# ip netns exec net0 ip link 3: eno2: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc mq state UP link/ether 00:1e:67:a0:22:3f brd ff:ff:ff:ff:ff:ff altname enp5s0# ip netns exec net1 ip link 4: eno3: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc fq_codel link/ether f8:e4:3b:3b:e4:10 brd ff:ff:ff:ff:ff:ff2) Add rdma link in the different net namespace
net0:
# ip netns exec net0 rdma link add rxe0 type rxe netdev eno2net1:
# ip netns exec net1 rdma link add rxe1 type rxe netdev eno33) Run rping test.
net0 # ip netns exec net0 rping -s -a 192.168.2.1 -C 1& [1] 1737 # ip netns exec net1 rping -c -a 192.168.2.1 -d -v -C 1 verbose count 1 ... ping data: rdma-ping-0: ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqr ...4) Remove the rdma links from the net namespaces.
net0:
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
UNCONN 0 0 0.0.0.0:4791 0.0.0.0:*
UNCONN 0 0 [::]:4791 [::]:*# ip netns exec net0 rdma link del rxe0
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Processnet1:
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
UNCONN 0 0 0.0.0.0:4791 0.0.0.0:*
UNCONN 0 0 [::]:4791 [::]:*# ip netns exec net1 rdma link del rxe1
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
V5->V6: Fix resource leak problem whenrxe_net_initfails. And fix some
style problems.
V4->V5: Rebase the commits to V6.4-rc1
V3->V4: Rebase the commits to rdma-next;
V2->V3: 1) Add "rdma link del" example in the cover letter, and use "ss -lu" to
verify rdma link is removed.
2) Add register_pernet_subsys/unregister_pernet_subsys net namespace
3) Replacel_sk6with sk6 ofpernet_name_space
V1->V2: Add the explicit initialization of sk6.Zhu Yanjun (8):
RDMA/rxe: Creating listening sock in newlink function RDMA/rxe: Support more rdma links ininit_netRDMA/nldev: Add dellink function pointer
RDMA/rxe: Implement dellink in rxe
RDMA/rxe: Replace global variable with sock lookup functions RDMA/rxe: add the support of net namespace RDMA/rxe: Add the support of net namespace notifier RDMA/rxe: Replacel_sk6with sk6 in net namespace
Quoted 12 lines
drivers/infiniband/core/nldev.c | 6 ++ drivers/infiniband/sw/rxe/Makefile | 3 +- drivers/infiniband/sw/rxe/rxe.c | 35 +++++++- drivers/infiniband/sw/rxe/rxe_net.c | 119 ++++++++++++++++++------ drivers/infiniband/sw/rxe/rxe_net.h | 9 +- drivers/infiniband/sw/rxe/rxe_ns.c | 134 ++++++++++++++++++++++++++++ drivers/infiniband/sw/rxe/rxe_ns.h | 17 ++++ include/rdma/rdma_netlink.h | 2 + 8 files changed, 285 insertions(+), 40 deletions(-) create mode 100644 drivers/infiniband/sw/rxe/rxe_ns.c create mode 100644 drivers/infiniband/sw/rxe/rxe_ns.h
I am OOO until Monday evening. I will try it then.
Bob
Quoted 155 lines · Bob Pearson
Quoted 150 lines · Zhu Yanjun
From: Zhu Yanjun <[email protected]>
When run "ip link add" command to add a rxe rdma link in a net namespace, normally this rxe rdma link can not work in a net name space.
The root cause is that a sock listening on udp port 4791 is created in
init_netwhen therdma_rxemodule is loaded into kernel. That is, the sock listening on udp port 4791 is created ininit_net. Other net namespace is difficult to use this sock.The following commits will solve this problem.
In the first commit, move the creating sock listening on udp port 4791 from
module_initfunction to rdma link creating functions. That is, after the modulerdma_rxeis loaded, the sock will not be created. When run "rdma link add ..." command, the sock will be created. So when creating a rdma link in the net namespace, the sock will be created in this net namespace.In the second commit, the functions
udp4_lib_lookupandudp6_lib_lookupwill check the sock exists in the net namespace or not. If yes, rdma link will increase the reference count of this sock, then continue other jobs instead of creating a new sock to listen on udp port 4791. Since the network notifier is global, when the modulerdma_rxeis loaded, this notifier will be registered.After the rdma link is created, the command "rdma link del" is to delete rdma link at the same time the sock is checked. If the reference count of this sock is greater than the sock reference count needed by udp tunnel, the sock reference count is decreased by one. If equal, it indicates that this rdma link is the last one. As such, the udp tunnel is shut down and the sock is closed. The above work should be implemented in linkdel function. But currently no dellink function in rxe. So the 3rd commit addes dellink function pointer. And the 4th commit implements the dellink function in rxe.
To now, it is not necessary to keep a global variable to store the sock listening udp port 4791. This global variable can be replaced by the functions
udp4_lib_lookupandudp6_lib_lookuptotally. Because the functionudp6_lib_lookupis in the fast path, a member variablel_sk6is added to store the sock. Ifl_sk6is NULL,udp6_lib_lookupis called to lookup the sock, then the sock is stored inl_sk6, in the future,it can be used directly.All the above work has been done in
init_net. And it can also work in the net namespace. So theinit_netis replaced by the individual net namespace. This is what the 6th commit does. Because rxe device is dependent on the net device and the sock listening on udp port 4791, every rxe device is in exclusive mode in the individual net namespace. Other rdma netns operations will be considerred in the future.In the 7th commit, the register_pernet_subsys/unregister_pernet_subsys functions are added. When a new net namespace is created, the init function will initialize the sk4 and sk6 socks. Then the 2 socks will be released when the net namespace is destroyed. The functions rxe_ns_pernet_sk4/rxe_ns_pernet_set_sk4 will get and set sk4 in the net namespace. The functions rxe_ns_pernet_sk6/rxe_ns_pernet_set_sk6 will handle sk6. Then sk4 and sk6 are used in the previous commits.
As the sk4 and sk6 in pernet namespace can be accessed, it is not necessary to add a new
l_sk6. As such, in the 8th commit, thel_sk6is replaced with the sk6 in pernet namespace.Test steps:
1) Suppose that 2 NICs are in 2 different net namespaces.# ip netns exec net0 ip link 3: eno2: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc mq state UP link/ether 00:1e:67:a0:22:3f brd ff:ff:ff:ff:ff:ff altname enp5s0# ip netns exec net1 ip link 4: eno3: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc fq_codel link/ether f8:e4:3b:3b:e4:10 brd ff:ff:ff:ff:ff:ff2) Add rdma link in the different net namespace
net0:
# ip netns exec net0 rdma link add rxe0 type rxe netdev eno2net1:
# ip netns exec net1 rdma link add rxe1 type rxe netdev eno33) Run rping test.
net0 # ip netns exec net0 rping -s -a 192.168.2.1 -C 1& [1] 1737 # ip netns exec net1 rping -c -a 192.168.2.1 -d -v -C 1 verbose count 1 ... ping data: rdma-ping-0: ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqr ...4) Remove the rdma links from the net namespaces.
net0:
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
UNCONN 0 0 0.0.0.0:4791 0.0.0.0:*
UNCONN 0 0 [::]:4791 [::]:*# ip netns exec net0 rdma link del rxe0
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Processnet1:
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
UNCONN 0 0 0.0.0.0:4791 0.0.0.0:*
UNCONN 0 0 [::]:4791 [::]:*# ip netns exec net1 rdma link del rxe1
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
V5->V6: Fix resource leak problem whenrxe_net_initfails. And fix some
style problems.
V4->V5: Rebase the commits to V6.4-rc1
V3->V4: Rebase the commits to rdma-next;
V2->V3: 1) Add "rdma link del" example in the cover letter, and use "ss -lu" to
verify rdma link is removed.
2) Add register_pernet_subsys/unregister_pernet_subsys net namespace
3) Replacel_sk6with sk6 ofpernet_name_space
V1->V2: Add the explicit initialization of sk6.Zhu Yanjun (8):
RDMA/rxe: Creating listening sock in newlink function RDMA/rxe: Support more rdma links ininit_netRDMA/nldev: Add dellink function pointer
RDMA/rxe: Implement dellink in rxe
RDMA/rxe: Replace global variable with sock lookup functions RDMA/rxe: add the support of net namespace RDMA/rxe: Add the support of net namespace notifier RDMA/rxe: Replacel_sk6with sk6 in net namespacedrivers/infiniband/core/nldev.c | 6 ++ drivers/infiniband/sw/rxe/Makefile | 3 +- drivers/infiniband/sw/rxe/rxe.c | 35 +++++++- drivers/infiniband/sw/rxe/rxe_net.c | 119 ++++++++++++++++++------ drivers/infiniband/sw/rxe/rxe_net.h | 9 +- drivers/infiniband/sw/rxe/rxe_ns.c | 134 ++++++++++++++++++++++++++++ drivers/infiniband/sw/rxe/rxe_ns.h | 17 ++++ include/rdma/rdma_netlink.h | 2 + 8 files changed, 285 insertions(+), 40 deletions(-)Quoted 3 lines
create mode 100644
drivers/infiniband/sw/rxe/rxe_ns.ccreate mode 100644drivers/infiniband/sw/rxe/rxe_ns.h
Quoted 1 line
I am OOO until Monday evening. I will try it then.
Got it. When you make tests with net namespace, please do not use lo because lo can not work with rxe. GID of rxe can not be generated correctly. Please use a physical NIC. Use the command "ip link set xxx netns test" to set the physical NIC xxx in the net namespace test.
Thanks a lot.
Zhu Yanjun
Quoted 2 lines
Bob
Re: '[v6,0/8] Fix the problem that rxe can not work in net namespace'
In reply to Zhu YanjunQuoted 141 lines
From: Zhu Yanjun <[email protected]>
When run "ip link add" command to add a rxe rdma link in a net namespace, normally this rxe rdma link can not work in a net name space.
The root cause is that a sock listening on udp port 4791 is created in
init_netwhen therdma_rxemodule is loaded into kernel. That is, the sock listening on udp port 4791 is created ininit_net. Other net namespace is difficult to use this sock.The following commits will solve this problem.
In the first commit, move the creating sock listening on udp port 4791 from
module_initfunction to rdma link creating functions. That is, after the modulerdma_rxeis loaded, the sock will not be created. When run "rdma link add ..." command, the sock will be created. So when creating a rdma link in the net namespace, the sock will be created in this net namespace.In the second commit, the functions
udp4_lib_lookupandudp6_lib_lookupwill check the sock exists in the net namespace or not. If yes, rdma link will increase the reference count of this sock, then continue other jobs instead of creating a new sock to listen on udp port 4791. Since the network notifier is global, when the modulerdma_rxeis loaded, this notifier will be registered.After the rdma link is created, the command "rdma link del" is to delete rdma link at the same time the sock is checked. If the reference count of this sock is greater than the sock reference count needed by udp tunnel, the sock reference count is decreased by one. If equal, it indicates that this rdma link is the last one. As such, the udp tunnel is shut down and the sock is closed. The above work should be implemented in linkdel function. But currently no dellink function in rxe. So the 3rd commit addes dellink function pointer. And the 4th commit implements the dellink function in rxe.
To now, it is not necessary to keep a global variable to store the sock listening udp port 4791. This global variable can be replaced by the functions
udp4_lib_lookupandudp6_lib_lookuptotally. Because the functionudp6_lib_lookupis in the fast path, a member variablel_sk6is added to store the sock. Ifl_sk6is NULL,udp6_lib_lookupis called to lookup the sock, then the sock is stored inl_sk6, in the future,it can be used directly.All the above work has been done in
init_net. And it can also work in the net namespace. So theinit_netis replaced by the individual net namespace. This is what the 6th commit does. Because rxe device is dependent on the net device and the sock listening on udp port 4791, every rxe device is in exclusive mode in the individual net namespace. Other rdma netns operations will be considerred in the future.In the 7th commit, the register_pernet_subsys/unregister_pernet_subsys functions are added. When a new net namespace is created, the init function will initialize the sk4 and sk6 socks. Then the 2 socks will be released when the net namespace is destroyed. The functions rxe_ns_pernet_sk4/rxe_ns_pernet_set_sk4 will get and set sk4 in the net namespace. The functions rxe_ns_pernet_sk6/rxe_ns_pernet_set_sk6 will handle sk6. Then sk4 and sk6 are used in the previous commits.
As the sk4 and sk6 in pernet namespace can be accessed, it is not necessary to add a new
l_sk6. As such, in the 8th commit, thel_sk6is replaced with the sk6 in pernet namespace.Test steps:
1) Suppose that 2 NICs are in 2 different net namespaces.# ip netns exec net0 ip link 3: eno2: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc mq state UP link/ether 00:1e:67:a0:22:3f brd ff:ff:ff:ff:ff:ff altname enp5s0# ip netns exec net1 ip link 4: eno3: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc fq_codel link/ether f8:e4:3b:3b:e4:10 brd ff:ff:ff:ff:ff:ff2) Add rdma link in the different net namespace
net0:
# ip netns exec net0 rdma link add rxe0 type rxe netdev eno2net1:
# ip netns exec net1 rdma link add rxe1 type rxe netdev eno33) Run rping test.
net0 # ip netns exec net0 rping -s -a 192.168.2.1 -C 1& [1] 1737 # ip netns exec net1 rping -c -a 192.168.2.1 -d -v -C 1 verbose count 1 ... ping data: rdma-ping-0: ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqr ...4) Remove the rdma links from the net namespaces.
net0:
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
UNCONN 0 0 0.0.0.0:4791 0.0.0.0:*
UNCONN 0 0 [::]:4791 [::]:*# ip netns exec net0 rdma link del rxe0
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Processnet1:
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
UNCONN 0 0 0.0.0.0:4791 0.0.0.0:*
UNCONN 0 0 [::]:4791 [::]:*# ip netns exec net1 rdma link del rxe1
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
V5->V6: Fix resource leak problem whenrxe_net_initfails. And fix some
style problems.
V4->V5: Rebase the commits to V6.4-rc1
V3->V4: Rebase the commits to rdma-next;
V2->V3: 1) Add "rdma link del" example in the cover letter, and use "ss -lu" to
verify rdma link is removed.
2) Add register_pernet_subsys/unregister_pernet_subsys net namespace
3) Replacel_sk6with sk6 ofpernet_name_space
V1->V2: Add the explicit initialization of sk6.Zhu Yanjun (8):
RDMA/rxe: Creating listening sock in newlink function RDMA/rxe: Support more rdma links ininit_netRDMA/nldev: Add dellink function pointer
RDMA/rxe: Implement dellink in rxe
RDMA/rxe: Replace global variable with sock lookup functions RDMA/rxe: add the support of net namespace RDMA/rxe: Add the support of net namespace notifier RDMA/rxe: Replacel_sk6with sk6 in net namespace
Quoted 12 lines
drivers/infiniband/core/nldev.c | 6 ++ drivers/infiniband/sw/rxe/Makefile | 3 +- drivers/infiniband/sw/rxe/rxe.c | 35 +++++++- drivers/infiniband/sw/rxe/rxe_net.c | 119 ++++++++++++++++++------ drivers/infiniband/sw/rxe/rxe_net.h | 9 +- drivers/infiniband/sw/rxe/rxe_ns.c | 134 ++++++++++++++++++++++++++++ drivers/infiniband/sw/rxe/rxe_ns.h | 17 ++++ include/rdma/rdma_netlink.h | 2 + 8 files changed, 285 insertions(+), 40 deletions(-) create mode 100644 drivers/infiniband/sw/rxe/rxe_ns.c create mode 100644 drivers/infiniband/sw/rxe/rxe_ns.h
Hi mates, thanks for your work. I'm working on cloud native web related and I really need this patch :).
Nowadays AI + Cloud Native is becoming more and more popular and there are more and more AI clusters running on Kubernetes. Most AI clusters use RDMA technology to communicate across nodes between training tasks, which speeds up training and reduces network latency. However, RDMA hardware devices are very expensive and costly for those who want to experience or test the entire AI training process (involving scheduling, using RDMA networks, and using nccl-like communication libraries, etc.). So we plan to develop a Kubernetes CNI plugin that can run RDMA locally, which will mount the rdma device on the host inside the container so that the AI tasks can also use the RDMA network.
I understand that Soft Roce is able to virtualize rdma devices without RDMA hardware, and I have verified this locally, being able to test it with tools like ib_send_bw. However, I further moved the NIC on the host or the NIC on the host using macvlan technology to the network namespace inside the container, and then manually added the rdma device inside the container using the rdma link add command, and then finally found out that I couldn't see the virtualized rdma device inside the container by using the ibv_devices command, and I switched the system mode of the rdma respectively(shared or exclusive), and neither of them worked. I've also seen other developers report similar issues, refer [nccl-tests container can't use Soft-RoCE interfaces](https://github.com/NVIDIA/deepops/issues/772).
I guess the Linux kernel upstream doesn't support soft roce containerization yet. I read the linux kernel upstream code and related patches carefully and confirmed my guess. So I contacted Yanjun, I really hope this patch can be merged into upstream, thanks again for your work.
Best Regards
Cyclinder kuo
Re: '[v6,0/8] Fix the problem that rxe can not work in net namespace'
In reply to Zhu YanjunQuoted 141 lines
From: Zhu Yanjun <[email protected]>
When run "ip link add" command to add a rxe rdma link in a net namespace, normally this rxe rdma link can not work in a net name space.
The root cause is that a sock listening on udp port 4791 is created in
init_netwhen therdma_rxemodule is loaded into kernel. That is, the sock listening on udp port 4791 is created ininit_net. Other net namespace is difficult to use this sock.The following commits will solve this problem.
In the first commit, move the creating sock listening on udp port 4791 from
module_initfunction to rdma link creating functions. That is, after the modulerdma_rxeis loaded, the sock will not be created. When run "rdma link add ..." command, the sock will be created. So when creating a rdma link in the net namespace, the sock will be created in this net namespace.In the second commit, the functions
udp4_lib_lookupandudp6_lib_lookupwill check the sock exists in the net namespace or not. If yes, rdma link will increase the reference count of this sock, then continue other jobs instead of creating a new sock to listen on udp port 4791. Since the network notifier is global, when the modulerdma_rxeis loaded, this notifier will be registered.After the rdma link is created, the command "rdma link del" is to delete rdma link at the same time the sock is checked. If the reference count of this sock is greater than the sock reference count needed by udp tunnel, the sock reference count is decreased by one. If equal, it indicates that this rdma link is the last one. As such, the udp tunnel is shut down and the sock is closed. The above work should be implemented in linkdel function. But currently no dellink function in rxe. So the 3rd commit addes dellink function pointer. And the 4th commit implements the dellink function in rxe.
To now, it is not necessary to keep a global variable to store the sock listening udp port 4791. This global variable can be replaced by the functions
udp4_lib_lookupandudp6_lib_lookuptotally. Because the functionudp6_lib_lookupis in the fast path, a member variablel_sk6is added to store the sock. Ifl_sk6is NULL,udp6_lib_lookupis called to lookup the sock, then the sock is stored inl_sk6, in the future,it can be used directly.All the above work has been done in
init_net. And it can also work in the net namespace. So theinit_netis replaced by the individual net namespace. This is what the 6th commit does. Because rxe device is dependent on the net device and the sock listening on udp port 4791, every rxe device is in exclusive mode in the individual net namespace. Other rdma netns operations will be considerred in the future.In the 7th commit, the register_pernet_subsys/unregister_pernet_subsys functions are added. When a new net namespace is created, the init function will initialize the sk4 and sk6 socks. Then the 2 socks will be released when the net namespace is destroyed. The functions rxe_ns_pernet_sk4/rxe_ns_pernet_set_sk4 will get and set sk4 in the net namespace. The functions rxe_ns_pernet_sk6/rxe_ns_pernet_set_sk6 will handle sk6. Then sk4 and sk6 are used in the previous commits.
As the sk4 and sk6 in pernet namespace can be accessed, it is not necessary to add a new
l_sk6. As such, in the 8th commit, thel_sk6is replaced with the sk6 in pernet namespace.Test steps:
1) Suppose that 2 NICs are in 2 different net namespaces.# ip netns exec net0 ip link 3: eno2: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc mq state UP link/ether 00:1e:67:a0:22:3f brd ff:ff:ff:ff:ff:ff altname enp5s0# ip netns exec net1 ip link 4: eno3: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 1500 qdisc fq_codel link/ether f8:e4:3b:3b:e4:10 brd ff:ff:ff:ff:ff:ff2) Add rdma link in the different net namespace
net0:
# ip netns exec net0 rdma link add rxe0 type rxe netdev eno2net1:
# ip netns exec net1 rdma link add rxe1 type rxe netdev eno33) Run rping test.
net0 # ip netns exec net0 rping -s -a 192.168.2.1 -C 1& [1] 1737 # ip netns exec net1 rping -c -a 192.168.2.1 -d -v -C 1 verbose count 1 ... ping data: rdma-ping-0: ABCDEFGHIJKLMNOPQRSTUVWXYZ[\]^_`abcdefghijklmnopqr ...4) Remove the rdma links from the net namespaces.
net0:
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
UNCONN 0 0 0.0.0.0:4791 0.0.0.0:*
UNCONN 0 0 [::]:4791 [::]:*# ip netns exec net0 rdma link del rxe0
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Processnet1:
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
UNCONN 0 0 0.0.0.0:4791 0.0.0.0:*
UNCONN 0 0 [::]:4791 [::]:*# ip netns exec net1 rdma link del rxe1
# ip netns exec net0 ss -lu
State Recv-Q Send-Q Local Address:Port Peer Address:Port Process
V5->V6: Fix resource leak problem whenrxe_net_initfails. And fix some
style problems.
V4->V5: Rebase the commits to V6.4-rc1
V3->V4: Rebase the commits to rdma-next;
V2->V3: 1) Add "rdma link del" example in the cover letter, and use "ss -lu" to
verify rdma link is removed.
2) Add register_pernet_subsys/unregister_pernet_subsys net namespace
3) Replacel_sk6with sk6 ofpernet_name_space
V1->V2: Add the explicit initialization of sk6.Zhu Yanjun (8):
RDMA/rxe: Creating listening sock in newlink function RDMA/rxe: Support more rdma links ininit_netRDMA/nldev: Add dellink function pointer
RDMA/rxe: Implement dellink in rxe
RDMA/rxe: Replace global variable with sock lookup functions RDMA/rxe: add the support of net namespace RDMA/rxe: Add the support of net namespace notifier RDMA/rxe: Replacel_sk6with sk6 in net namespace
Quoted 12 lines
drivers/infiniband/core/nldev.c | 6 ++ drivers/infiniband/sw/rxe/Makefile | 3 +- drivers/infiniband/sw/rxe/rxe.c | 35 +++++++- drivers/infiniband/sw/rxe/rxe_net.c | 119 ++++++++++++++++++------ drivers/infiniband/sw/rxe/rxe_net.h | 9 +- drivers/infiniband/sw/rxe/rxe_ns.c | 134 ++++++++++++++++++++++++++++ drivers/infiniband/sw/rxe/rxe_ns.h | 17 ++++ include/rdma/rdma_netlink.h | 2 + 8 files changed, 285 insertions(+), 40 deletions(-) create mode 100644 drivers/infiniband/sw/rxe/rxe_ns.c create mode 100644 drivers/infiniband/sw/rxe/rxe_ns.h
Hi mates, thanks for your work. I'm working on cloud native web related and I really need this patch :).
Nowadays AI + Cloud Native is becoming more and more popular and there are more and more AI clusters running on Kubernetes. Most AI clusters use RDMA technology to communicate across nodes between training tasks, which speeds up training and reduces network latency. However, RDMA hardware devices are very expensive and costly for those who want to experience or test the entire AI training process (involving scheduling, using RDMA networks, and using nccl-like communication libraries, etc.). So we plan to develop a Kubernetes CNI plugin that can run RDMA locally, which will mount the rdma device on the host inside the container so that the AI tasks can also use the RDMA network.
I understand that Soft Roce is able to virtualize rdma devices without RDMA hardware, and I have verified this locally, being able to test it with tools like ib_send_bw. However, I further moved the NIC on the host or the NIC on the host using macvlan technology to the network namespace inside the container, and then manually added the rdma device inside the container using the rdma link add command, and then finally found out that I couldn't see the virtualized rdma device inside the container by using the ibv_devices command, and I switched the system mode of the rdma respectively(shared or exclusive), and neither of them worked. I've also seen other developers report similar issues, refer [nccl-tests container can't use Soft-RoCE interfaces](https://github.com/NVIDIA/deepops/issues/772).
I guess the Linux kernel upstream doesn't support soft roce containerization yet. I read the linux kernel upstream code and related patches carefully and confirmed my guess. So I contacted Yanjun, I really hope this patch can be merged into upstream, thanks again for your work.
Best Regards
Cyclinder kuo