diff options
author | Liang Zhen <liang.zhen@intel.com> | 2016-06-20 16:55:31 -0400 |
---|---|---|
committer | Greg Kroah-Hartman <gregkh@linuxfoundation.org> | 2016-06-20 14:28:39 -0700 |
commit | 9faa2ade3fdaf00c6c45654071b325d3c88f9230 (patch) | |
tree | 0ff8d317949ddadfe19752201ef0b45bae9963d9 /drivers/staging/lustre/lustre/ptlrpc/niobuf.c | |
parent | 32c8728d87dcd59949800d838e4e7dce0c625062 (diff) |
staging/lustre/ptlrpc: missing wakeup for ptlrpc_check_set
This patch changes a few things:
- There is no guarantee that request_out_callback will happen
before reply_in_callback, if a request got reply and unlinked
reply buffer before request_out_callback is called, then the
thread waiting on ptlrpc_request_set will miss wakeup event.
This may seriously impact performance of some IO workloads or
result in RPC timeout
- To make code more easier to understand, this patch changes
action-bits "rq_req_unlink" and "rq_reply_unlink" to
status-bits "rq_req_unlinked" and "rq_reply_unlinked"
Signed-off-by: Liang Zhen <liang.zhen@intel.com>
Reviewed-on: http://review.whamcloud.com/12158
Intel-bug-id: https://jira.hpdd.intel.com/browse/LU-5696
Reviewed-by: Johann Lombardi <johann.lombardi@intel.com>
Reviewed-by: Li Wei <wei.g.li@intel.com>
Reviewed-by: Mike Pershin <mike.pershin@intel.com>
Signed-off-by: Oleg Drokin <green@linuxhacker.ru>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Diffstat (limited to 'drivers/staging/lustre/lustre/ptlrpc/niobuf.c')
-rw-r--r-- | drivers/staging/lustre/lustre/ptlrpc/niobuf.c | 14 |
1 files changed, 7 insertions, 7 deletions
diff --git a/drivers/staging/lustre/lustre/ptlrpc/niobuf.c b/drivers/staging/lustre/lustre/ptlrpc/niobuf.c index 4e7d68f223a5..8c49f5ed2fca 100644 --- a/drivers/staging/lustre/lustre/ptlrpc/niobuf.c +++ b/drivers/staging/lustre/lustre/ptlrpc/niobuf.c @@ -577,19 +577,18 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) } spin_lock(&request->rq_lock); - /* If the MD attach succeeds, there _will_ be a reply_in callback */ - request->rq_receiving_reply = !noreply; - request->rq_req_unlink = 1; /* We are responsible for unlinking the reply buffer */ - request->rq_reply_unlink = !noreply; + request->rq_reply_unlinked = noreply; + request->rq_receiving_reply = !noreply; /* Clear any flags that may be present from previous sends. */ + request->rq_req_unlinked = 0; request->rq_replied = 0; request->rq_err = 0; request->rq_timedout = 0; request->rq_net_err = 0; request->rq_resend = 0; request->rq_restart = 0; - request->rq_reply_truncate = 0; + request->rq_reply_truncated = 0; spin_unlock(&request->rq_lock); if (!noreply) { @@ -604,7 +603,7 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) reply_md.user_ptr = &request->rq_reply_cbid; reply_md.eq_handle = ptlrpc_eq_h; - /* We must see the unlink callback to unset rq_reply_unlink, + /* We must see the unlink callback to set rq_reply_unlinked, * so we can't auto-unlink */ rc = LNetMDAttach(reply_me_h, reply_md, LNET_RETAIN, @@ -651,9 +650,10 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply) connection, request->rq_request_portal, request->rq_xid, 0); - if (rc == 0) + if (likely(rc == 0)) goto out; + request->rq_req_unlinked = 1; ptlrpc_req_finished(request); if (noreply) goto out; |