From c156618e15101a9cc8c815108fec0300a0ec6637 Mon Sep 17 00:00:00 2001 From: Scott Mayhew Date: Tue, 5 Dec 2017 13:55:44 -0500 Subject: [PATCH 1/4] nfs: fix a deadlock in nfs client initialization The following deadlock can occur between a process waiting for a client to initialize in while walking the client list during nfsv4 server trunking detection and another process waiting for the nfs_clid_init_mutex so it can initialize that client: Process 1 Process 2 --------- --------- spin_lock(&nn->nfs_client_lock); list_add_tail(&CLIENTA->cl_share_link, &nn->nfs_client_list); spin_unlock(&nn->nfs_client_lock); spin_lock(&nn->nfs_client_lock); list_add_tail(&CLIENTB->cl_share_link, &nn->nfs_client_list); spin_unlock(&nn->nfs_client_lock); mutex_lock(&nfs_clid_init_mutex); nfs41_walk_client_list(clp, result, cred); nfs_wait_client_init_complete(CLIENTA); (waiting for nfs_clid_init_mutex) Make sure nfs_match_client() only evaluates clients that have completed initialization in order to prevent that deadlock. This patch also fixes v4.0 trunking behavior by not marking the client NFS_CS_READY until the clientid has been confirmed. Signed-off-by: Scott Mayhew Signed-off-by: Anna Schumaker --- fs/nfs/client.c | 11 +++++++++++ fs/nfs/nfs4client.c | 17 +++++++++++++---- 2 files changed, 24 insertions(+), 4 deletions(-) diff --git a/fs/nfs/client.c b/fs/nfs/client.c index 0ac2fb1c6b63..b9129e2befea 100644 --- a/fs/nfs/client.c +++ b/fs/nfs/client.c @@ -291,12 +291,23 @@ static struct nfs_client *nfs_match_client(const struct nfs_client_initdata *dat const struct sockaddr *sap = data->addr; struct nfs_net *nn = net_generic(data->net, nfs_net_id); +again: list_for_each_entry(clp, &nn->nfs_client_list, cl_share_link) { const struct sockaddr *clap = (struct sockaddr *)&clp->cl_addr; /* Don't match clients that failed to initialise properly */ if (clp->cl_cons_state < 0) continue; + /* If a client is still initializing then we need to wait */ + if (clp->cl_cons_state > NFS_CS_READY) { + refcount_inc(&clp->cl_count); + spin_unlock(&nn->nfs_client_lock); + nfs_wait_client_init_complete(clp); + nfs_put_client(clp); + spin_lock(&nn->nfs_client_lock); + goto again; + } + /* Different NFS versions cannot share the same nfs_client */ if (clp->rpc_ops != data->nfs_mod->rpc_ops) continue; diff --git a/fs/nfs/nfs4client.c b/fs/nfs/nfs4client.c index 12bbab0becb4..65a7e5da508c 100644 --- a/fs/nfs/nfs4client.c +++ b/fs/nfs/nfs4client.c @@ -404,15 +404,19 @@ struct nfs_client *nfs4_init_client(struct nfs_client *clp, if (error < 0) goto error; - if (!nfs4_has_session(clp)) - nfs_mark_client_ready(clp, NFS_CS_READY); - error = nfs4_discover_server_trunking(clp, &old); if (error < 0) goto error; - if (clp != old) + if (clp != old) { clp->cl_preserve_clid = true; + /* + * Mark the client as having failed initialization so other + * processes walking the nfs_client_list in nfs_match_client() + * won't try to use it. + */ + nfs_mark_client_ready(clp, -EPERM); + } nfs_put_client(clp); clear_bit(NFS_CS_TSM_POSSIBLE, &clp->cl_flags); return old; @@ -539,6 +543,9 @@ int nfs40_walk_client_list(struct nfs_client *new, spin_lock(&nn->nfs_client_lock); list_for_each_entry(pos, &nn->nfs_client_list, cl_share_link) { + if (pos == new) + goto found; + status = nfs4_match_client(pos, new, &prev, nn); if (status < 0) goto out_unlock; @@ -559,6 +566,7 @@ int nfs40_walk_client_list(struct nfs_client *new, * way that a SETCLIENTID_CONFIRM to pos can succeed is * if new and pos point to the same server: */ +found: refcount_inc(&pos->cl_count); spin_unlock(&nn->nfs_client_lock); @@ -572,6 +580,7 @@ int nfs40_walk_client_list(struct nfs_client *new, case 0: nfs4_swap_callback_idents(pos, new); pos->cl_confirm = new->cl_confirm; + nfs_mark_client_ready(pos, NFS_CS_READY); prev = NULL; *result = pos; From ccede7598588ae344143f82fb763912535648d58 Mon Sep 17 00:00:00 2001 From: Chuck Lever Date: Mon, 4 Dec 2017 14:04:04 -0500 Subject: [PATCH 2/4] xprtrdma: Spread reply processing over more CPUs Commit d8f532d20ee4 ("xprtrdma: Invoke rpcrdma_reply_handler directly from RECV completion") introduced a performance regression for NFS I/O small enough to not need memory registration. In multi- threaded benchmarks that generate primarily small I/O requests, IOPS throughput is reduced by nearly a third. This patch restores the previous level of throughput. Because workqueues are typically BOUND (in particular ib_comp_wq, nfsiod_workqueue, and rpciod_workqueue), NFS/RDMA workloads tend to aggregate on the CPU that is handling Receive completions. The usual approach to addressing this problem is to create a QP and CQ for each CPU, and then schedule transactions on the QP for the CPU where you want the transaction to complete. The transaction then does not require an extra context switch during completion to end up on the same CPU where the transaction was started. This approach doesn't work for the Linux NFS/RDMA client because currently the Linux NFS client does not support multiple connections per client-server pair, and the RDMA core API does not make it straightforward for ULPs to determine which CPU is responsible for handling Receive completions for a CQ. So for the moment, record the CPU number in the rpcrdma_req before the transport sends each RPC Call. Then during Receive completion, queue the RPC completion on that same CPU. Additionally, move all RPC completion processing to the deferred handler so that even RPCs with simple small replies complete on the CPU that sent the corresponding RPC Call. Fixes: d8f532d20ee4 ("xprtrdma: Invoke rpcrdma_reply_handler ...") Signed-off-by: Chuck Lever Signed-off-by: Anna Schumaker --- net/sunrpc/xprtrdma/rpc_rdma.c | 6 +----- net/sunrpc/xprtrdma/transport.c | 2 ++ net/sunrpc/xprtrdma/verbs.c | 2 +- net/sunrpc/xprtrdma/xprt_rdma.h | 1 + 4 files changed, 5 insertions(+), 6 deletions(-) diff --git a/net/sunrpc/xprtrdma/rpc_rdma.c b/net/sunrpc/xprtrdma/rpc_rdma.c index ed34dc0f144c..a3f2ab283aeb 100644 --- a/net/sunrpc/xprtrdma/rpc_rdma.c +++ b/net/sunrpc/xprtrdma/rpc_rdma.c @@ -1408,11 +1408,7 @@ void rpcrdma_reply_handler(struct rpcrdma_rep *rep) dprintk("RPC: %s: reply %p completes request %p (xid 0x%08x)\n", __func__, rep, req, be32_to_cpu(rep->rr_xid)); - if (list_empty(&req->rl_registered) && - !test_bit(RPCRDMA_REQ_F_TX_RESOURCES, &req->rl_flags)) - rpcrdma_complete_rqst(rep); - else - queue_work(rpcrdma_receive_wq, &rep->rr_work); + queue_work_on(req->rl_cpu, rpcrdma_receive_wq, &rep->rr_work); return; out_badstatus: diff --git a/net/sunrpc/xprtrdma/transport.c b/net/sunrpc/xprtrdma/transport.c index 646c24494ea7..6ee1ad8978f3 100644 --- a/net/sunrpc/xprtrdma/transport.c +++ b/net/sunrpc/xprtrdma/transport.c @@ -52,6 +52,7 @@ #include #include #include +#include #include "xprt_rdma.h" @@ -656,6 +657,7 @@ xprt_rdma_allocate(struct rpc_task *task) task->tk_pid, __func__, rqst->rq_callsize, rqst->rq_rcvsize, req); + req->rl_cpu = smp_processor_id(); req->rl_connect_cookie = 0; /* our reserved value */ rpcrdma_set_xprtdata(rqst, req); rqst->rq_buffer = req->rl_sendbuf->rg_base; diff --git a/net/sunrpc/xprtrdma/verbs.c b/net/sunrpc/xprtrdma/verbs.c index 710b3f77db82..8607c029c0dd 100644 --- a/net/sunrpc/xprtrdma/verbs.c +++ b/net/sunrpc/xprtrdma/verbs.c @@ -83,7 +83,7 @@ rpcrdma_alloc_wq(void) struct workqueue_struct *recv_wq; recv_wq = alloc_workqueue("xprtrdma_receive", - WQ_MEM_RECLAIM | WQ_UNBOUND | WQ_HIGHPRI, + WQ_MEM_RECLAIM | WQ_HIGHPRI, 0); if (!recv_wq) return -ENOMEM; diff --git a/net/sunrpc/xprtrdma/xprt_rdma.h b/net/sunrpc/xprtrdma/xprt_rdma.h index 51686d9eac5f..1342f743f1c4 100644 --- a/net/sunrpc/xprtrdma/xprt_rdma.h +++ b/net/sunrpc/xprtrdma/xprt_rdma.h @@ -342,6 +342,7 @@ enum { struct rpcrdma_buffer; struct rpcrdma_req { struct list_head rl_list; + int rl_cpu; unsigned int rl_connect_cookie; struct rpcrdma_buffer *rl_buffer; struct rpcrdma_rep *rl_reply; From dc4fd9ab01ab379ae5af522b3efd4187a7c30a31 Mon Sep 17 00:00:00 2001 From: Scott Mayhew Date: Fri, 8 Dec 2017 16:00:12 -0500 Subject: [PATCH 3/4] nfs: don't wait on commit in nfs_commit_inode() if there were no commit requests If there were no commit requests, then nfs_commit_inode() should not wait on the commit or mark the inode dirty, otherwise the following BUG_ON can be triggered: [ 1917.130762] kernel BUG at fs/inode.c:578! [ 1917.130766] Oops: Exception in kernel mode, sig: 5 [#1] [ 1917.130768] SMP NR_CPUS=2048 NUMA pSeries [ 1917.130772] Modules linked in: iscsi_tcp libiscsi_tcp libiscsi scsi_transport_iscsi blocklayoutdriver rpcsec_gss_krb5 auth_rpcgss nfsv4 dns_resolver nfs lockd grace fscache sunrpc sg nx_crypto pseries_rng ip_tables xfs libcrc32c sd_mod crc_t10dif crct10dif_generic crct10dif_common ibmvscsi scsi_transport_srp ibmveth scsi_tgt dm_mirror dm_region_hash dm_log dm_mod [ 1917.130805] CPU: 2 PID: 14923 Comm: umount.nfs4 Tainted: G ------------ T 3.10.0-768.el7.ppc64 #1 [ 1917.130810] task: c0000005ecd88040 ti: c00000004cea0000 task.ti: c00000004cea0000 [ 1917.130813] NIP: c000000000354178 LR: c000000000354160 CTR: c00000000012db80 [ 1917.130816] REGS: c00000004cea3720 TRAP: 0700 Tainted: G ------------ T (3.10.0-768.el7.ppc64) [ 1917.130820] MSR: 8000000100029032 CR: 22002822 XER: 20000000 [ 1917.130828] CFAR: c00000000011f594 SOFTE: 1 GPR00: c000000000354160 c00000004cea39a0 c0000000014c4700 c0000000018cc750 GPR04: 000000000000c750 80c0000000000000 0600000000000000 04eeb76bea749a03 GPR08: 0000000000000034 c0000000018cc758 0000000000000001 d000000005e619e8 GPR12: c00000000012db80 c000000007b31200 0000000000000000 0000000000000000 GPR16: 0000000000000000 0000000000000000 0000000000000000 0000000000000000 GPR20: 0000000000000000 0000000000000000 0000000000000000 0000000000000000 GPR24: 0000000000000000 c000000000dfc3ec 0000000000000000 c0000005eefc02c0 GPR28: d0000000079dbd50 c0000005b94a02c0 c0000005b94a0250 c0000005b94a01c8 [ 1917.130867] NIP [c000000000354178] .evict+0x1c8/0x350 [ 1917.130871] LR [c000000000354160] .evict+0x1b0/0x350 [ 1917.130873] Call Trace: [ 1917.130876] [c00000004cea39a0] [c000000000354160] .evict+0x1b0/0x350 (unreliable) [ 1917.130880] [c00000004cea3a30] [c0000000003558cc] .evict_inodes+0x13c/0x270 [ 1917.130884] [c00000004cea3af0] [c000000000327d20] .kill_anon_super+0x70/0x1e0 [ 1917.130896] [c00000004cea3b80] [d000000005e43e30] .nfs_kill_super+0x20/0x60 [nfs] [ 1917.130900] [c00000004cea3c00] [c000000000328a20] .deactivate_locked_super+0xa0/0x1b0 [ 1917.130903] [c00000004cea3c80] [c00000000035ba54] .cleanup_mnt+0xd4/0x180 [ 1917.130907] [c00000004cea3d10] [c000000000119034] .task_work_run+0x114/0x150 [ 1917.130912] [c00000004cea3db0] [c00000000001ba6c] .do_notify_resume+0xcc/0x100 [ 1917.130916] [c00000004cea3e30] [c00000000000a7b0] .ret_from_except_lite+0x5c/0x60 [ 1917.130919] Instruction dump: [ 1917.130921] 7fc3f378 486734b5 60000000 387f00a0 38800003 4bdcb365 60000000 e95f00a0 [ 1917.130927] 694a0060 7d4a0074 794ad182 694a0001 <0b0a0000> 892d02a4 2f890000 40de0134 Signed-off-by: Scott Mayhew Cc: stable@vger.kernel.org # 4.5+ Signed-off-by: Anna Schumaker --- fs/nfs/write.c | 2 ++ 1 file changed, 2 insertions(+) diff --git a/fs/nfs/write.c b/fs/nfs/write.c index 5b5f464f6f2a..4a379d7918f2 100644 --- a/fs/nfs/write.c +++ b/fs/nfs/write.c @@ -1890,6 +1890,8 @@ int nfs_commit_inode(struct inode *inode, int how) if (res) error = nfs_generic_commit_list(inode, &head, how, &cinfo); nfs_commit_end(cinfo.mds); + if (res == 0) + return res; if (error < 0) goto out_error; if (!may_wait) From 90d91b0cd371193d9dbfa9beacab8ab9a4cb75e0 Mon Sep 17 00:00:00 2001 From: Trond Myklebust Date: Thu, 14 Dec 2017 21:24:08 -0500 Subject: [PATCH 4/4] SUNRPC: Fix a race in the receive code path We must ensure that the call to rpc_sleep_on() in xprt_transmit() cannot race with the call to xprt_complete_rqst(). Reported-by: Chuck Lever Link: https://bugzilla.linux-nfs.org/show_bug.cgi?id=317 Fixes: ce7c252a8c74 ("SUNRPC: Add a separate spinlock to protect..") Cc: stable@vger.kernel.org # 4.14+ Reviewed-by: Chuck Lever Signed-off-by: Trond Myklebust Signed-off-by: Anna Schumaker --- net/sunrpc/xprt.c | 36 +++++++++++++++++++++++------------- 1 file changed, 23 insertions(+), 13 deletions(-) diff --git a/net/sunrpc/xprt.c b/net/sunrpc/xprt.c index 02a9bacb239b..5b06f6906a27 100644 --- a/net/sunrpc/xprt.c +++ b/net/sunrpc/xprt.c @@ -1001,6 +1001,7 @@ void xprt_transmit(struct rpc_task *task) { struct rpc_rqst *req = task->tk_rqstp; struct rpc_xprt *xprt = req->rq_xprt; + unsigned int connect_cookie; int status, numreqs; dprintk("RPC: %5u xprt_transmit(%u)\n", task->tk_pid, req->rq_slen); @@ -1024,6 +1025,7 @@ void xprt_transmit(struct rpc_task *task) } else if (!req->rq_bytes_sent) return; + connect_cookie = xprt->connect_cookie; req->rq_xtime = ktime_get(); status = xprt->ops->send_request(task); trace_xprt_transmit(xprt, req->rq_xid, status); @@ -1047,20 +1049,28 @@ void xprt_transmit(struct rpc_task *task) xprt->stat.bklog_u += xprt->backlog.qlen; xprt->stat.sending_u += xprt->sending.qlen; xprt->stat.pending_u += xprt->pending.qlen; - - /* Don't race with disconnect */ - if (!xprt_connected(xprt)) - task->tk_status = -ENOTCONN; - else { - /* - * Sleep on the pending queue since - * we're expecting a reply. - */ - if (!req->rq_reply_bytes_recvd && rpc_reply_expected(task)) - rpc_sleep_on(&xprt->pending, task, xprt_timer); - req->rq_connect_cookie = xprt->connect_cookie; - } spin_unlock_bh(&xprt->transport_lock); + + req->rq_connect_cookie = connect_cookie; + if (rpc_reply_expected(task) && !READ_ONCE(req->rq_reply_bytes_recvd)) { + /* + * Sleep on the pending queue if we're expecting a reply. + * The spinlock ensures atomicity between the test of + * req->rq_reply_bytes_recvd, and the call to rpc_sleep_on(). + */ + spin_lock(&xprt->recv_lock); + if (!req->rq_reply_bytes_recvd) { + rpc_sleep_on(&xprt->pending, task, xprt_timer); + /* + * Send an extra queue wakeup call if the + * connection was dropped in case the call to + * rpc_sleep_on() raced. + */ + if (!xprt_connected(xprt)) + xprt_wake_pending_tasks(xprt, -ENOTCONN); + } + spin_unlock(&xprt->recv_lock); + } } static void xprt_add_backlog(struct rpc_xprt *xprt, struct rpc_task *task)