Last week Trond reported a server crash in svc_rdma_sendto() that
seemed related to deferral of a request incoming on an RPC/RDMA
transport. A link to that report appears in the first patch
below.
The fundamental problem seems to be that there is no way to
guarantee test coverage of SUNRPC request deferral. "SUNRPC: Cache
deferral injection" proposes a way to address that issue.
With an on-demand mechanism for triggering a deferral, it was
straightforward to find and address the issues that led to Trond's
reported crash. The first patch in this series addresses the crash,
and the second fixes a related observability crash.
The final three patches are clean-ups that are possible now that
the deferral issues have been addressed.
Comments, opinions, and Reviewed-by's are appreciated!
---
Chuck Lever (6):
SUNRPC: Fix NFSD's request deferral on RDMA transports
SUNRPC: Fix the svc_deferred_event trace class
SUNRPC: Cache deferral injection
SUNRPC: Make cache_req::thread_wait an unsigned long
SUNRPC: Remove dead code in svc_tcp_release_rqst()
SUNRPC: Remove svc_rqst::rq_xprt_hlen
include/linux/sunrpc/cache.h | 7 +++----
include/linux/sunrpc/svc.h | 3 +--
include/trace/events/sunrpc.h | 11 +++++------
net/sunrpc/cache.c | 18 +++++++++++++++++-
net/sunrpc/debugfs.c | 3 +++
net/sunrpc/fail.h | 2 +-
net/sunrpc/svc_xprt.c | 11 ++++++-----
net/sunrpc/svcsock.c | 11 -----------
net/sunrpc/xprtrdma/svc_rdma_recvfrom.c | 1 -
9 files changed, 36 insertions(+), 31 deletions(-)
--
Chuck Lever
