lwn.git/net/rds/ib.c, branch docs-5.3

net: rds: fix memory leak when unload rds_rdma

2019-06-06T00:08:14+00:00

When KASAN is enabled, after several rds connections are created, then "rmmod rds_rdma" is run. The following will appear. " BUG rds_ib_incoming (Not tainted): Objects remaining in rds_ib_incoming on __kmem_cache_shutdown() Call Trace: dump_stack+0x71/0xab slab_err+0xad/0xd0 __kmem_cache_shutdown+0x17d/0x370 shutdown_cache+0x17/0x130 kmem_cache_destroy+0x1df/0x210 rds_ib_recv_exit+0x11/0x20 [rds_rdma] rds_ib_exit+0x7a/0x90 [rds_rdma] __x64_sys_delete_module+0x224/0x2c0 ? __ia32_sys_delete_module+0x2c0/0x2c0 do_syscall_64+0x73/0x190 entry_SYSCALL_64_after_hwframe+0x44/0xa9 " This is rds connection memory leak. The root cause is: When "rmmod rds_rdma" is run, rds_ib_remove_one will call rds_ib_dev_shutdown to drop the rds connections. rds_ib_dev_shutdown will call rds_conn_drop to drop rds connections as below. " rds_conn_path_drop(&conn->c_path[0], false); " In the above, destroy is set to false. void rds_conn_path_drop(struct rds_conn_path *cp, bool destroy) { atomic_set(&cp->cp_state, RDS_CONN_ERROR); rcu_read_lock(); if (!destroy && rds_destroy_pending(cp->cp_conn)) { rcu_read_unlock(); return; } queue_work(rds_wq, &cp->cp_down_w); rcu_read_unlock(); } In the above function, destroy is set to false. rds_destroy_pending is called. This does not move rds connections to ib_nodev_conns. So destroy is set to true to move rds connections to ib_nodev_conns. In rds_ib_unregister_client, flush_workqueue is called to make rds_wq finsh shutdown rds connections. The function rds_ib_destroy_nodev_conns is called to shutdown rds connections finally. Then rds_ib_recv_exit is called to destroy slab. void rds_ib_recv_exit(void) { kmem_cache_destroy(rds_ib_incoming_slab); kmem_cache_destroy(rds_ib_frag_slab); } The above slab memory leak will not occur again. >From tests, 256 rds connections [root@ca-dev14 ~]# time rmmod rds_rdma real 0m16.522s user 0m0.000s sys 0m8.152s 512 rds connections [root@ca-dev14 ~]# time rmmod rds_rdma real 0m32.054s user 0m0.000s sys 0m15.568s To rmmod rds_rdma with 256 rds connections, about 16 seconds are needed. And with 512 rds connections, about 32 seconds are needed. >From ftrace, when one rds connection is destroyed, " 19) | rds_conn_destroy [rds]() { 19) 7.782 us | rds_conn_path_drop [rds](); 15) | rds_shutdown_worker [rds]() { 15) | rds_conn_shutdown [rds]() { 15) 1.651 us | rds_send_path_reset [rds](); 15) 7.195 us | } 15) + 11.434 us | } 19) 2.285 us | rds_cong_remove_conn [rds](); 19) * 24062.76 us | } " So if many rds connections will be destroyed, this function rds_ib_destroy_nodev_conns uses most of time. Suggested-by: Håkon Bugge Signed-off-by: Zhu Yanjun Signed-off-by: David S. Miller

rds: add transport specific tos_map hook

2019-02-04T22:59:13+00:00

RDMA transport maps user tos to underline virtual lanes(VL) for IB or DSCP values. RDMA CM transport abstract thats for RDS. TCP transport makes use of default priority 0 and maps all user tos values to it. Reviewed-by: Sowmini Varadhan Signed-off-by: Santosh Shilimkar [yanjun.zhu@oracle.com: Adapted original patch with ipv6 changes] Signed-off-by: Zhu Yanjun

rds: add type of service(tos) infrastructure

2019-02-04T22:59:12+00:00

RDS Service type (TOS) is user-defined and needs to be configured via RDS IOCTL interface. It must be set before initiating any traffic and once set the TOS can not be changed. All out-going traffic from the socket will be associated with its TOS. Reviewed-by: Sowmini Varadhan Signed-off-by: Santosh Shilimkar [yanjun.zhu@oracle.com: Adapted original patch with ipv6 changes] Signed-off-by: Zhu Yanjun

RDMA: Start use ib_device_ops

2018-12-12T14:40:16+00:00

Make all the required change to start use the ib_device_ops structure. Signed-off-by: Kamal Heib Signed-off-by: Jason Gunthorpe

net/rds: Use rdma_read_gids to get connection SGID/DGID in IPv6

2018-08-27T22:26:01+00:00

In IPv4, the newly introduced rdma_read_gids is used to read the SGID/DGID for the connection which returns GID correctly for RoCE transport as well. In IPv6, rdma_read_gids is also used. The following are why rdma_read_gids is introduced. rdma_addr_get_dgid() for RoCE for client side connections returns MAC address, instead of DGID. rdma_addr_get_sgid() for RoCE doesn't return correct SGID for IPv6 and when more than one IP address is assigned to the netdevice. So the transport agnostic rdma_read_gids() API is provided by rdma_cm module. Signed-off-by: Zhu Yanjun Acked-by: Santosh Shilimkar Signed-off-by: David S. Miller

Merge branch 'linus/master' into rdma.git for-next

2018-08-16T20:21:29+00:00

rdma.git merge resolution for the 4.19 merge window Conflicts: drivers/infiniband/core/rdma_core.c - Use the rdma code and revise with the new spelling for atomic_fetch_add_unless drivers/nvme/host/rdma.c - Replace max_sge with max_send_sge in new blk code drivers/nvme/target/rdma.c - Use the blk code and revise to use NULL for ib_post_recv when appropriate - Replace max_sge with max_recv_sge in new blk code net/rds/ib_send.c - Use the net code and revise to use NULL for ib_post_recv when appropriate Signed-off-by: Jason Gunthorpe

rds: Remove IPv6 dependency

2018-08-01T16:32:35+00:00

This patch removes the IPv6 dependency from RDS. Signed-off-by: Ka-Cheong Poon Acked-by: Santosh Shilimkar Signed-off-by: David S. Miller

rds: remove trailing whitespace and blank lines

2018-07-24T21:10:42+00:00

Signed-off-by: Stephen Hemminger Signed-off-by: David S. Miller

rds: Extend RDS API for IPv6 support

2018-07-24T04:17:44+00:00

There are many data structures (RDS socket options) used by RDS apps which use a 32 bit integer to store IP address. To support IPv6, struct in6_addr needs to be used. To ensure backward compatibility, a new data structure is introduced for each of those data structures which use a 32 bit integer to represent an IP address. And new socket options are introduced to use those new structures. This means that existing apps should work without a problem with the new RDS module. For apps which want to use IPv6, those new data structures and socket options can be used. IPv4 mapped address is used to represent IPv4 address in the new data structures. v4: Revert changes to SO_RDS_TRANSPORT Signed-off-by: Ka-Cheong Poon Acked-by: Santosh Shilimkar Signed-off-by: David S. Miller

rds: Enable RDS IPv6 support

2018-07-24T04:17:44+00:00

This patch enables RDS to use IPv6 addresses. For RDS/TCP, the listener is now an IPv6 endpoint which accepts both IPv4 and IPv6 connection requests. RDS/RDMA/IB uses a private data (struct rds_ib_connect_private) exchange between endpoints at RDS connection establishment time to support RDMA. This private data exchange uses a 32 bit integer to represent an IP address. This needs to be changed in order to support IPv6. A new private data struct rds6_ib_connect_private is introduced to handle this. To ensure backward compatibility, an IPv6 capable RDS stack uses another RDMA listener port (RDS_CM_PORT) to accept IPv6 connection. And it continues to use the original RDS_PORT for IPv4 RDS connections. When it needs to communicate with an IPv6 peer, it uses the RDS_CM_PORT to send the connection set up request. v5: Fixed syntax problem (David Miller). v4: Changed port history comments in rds.h (Sowmini Varadhan). v3: Added support to set up IPv4 connection using mapped address (David Miller). Added support to set up connection between link local and non-link addresses. Various review comments from Santosh Shilimkar and Sowmini Varadhan. v2: Fixed bound and peer address scope mismatched issue. Added back rds_connect() IPv6 changes. Signed-off-by: Ka-Cheong Poon Acked-by: Santosh Shilimkar Signed-off-by: David S. Miller