v1.9.0-pre41

• Add support for static build (see https://openucx.readthedocs.io/en/master/faq.html#build-user-application-with-ucx)
• Add support for different number of RoCE LAG paths on server and client
• Add support for user-provided memory handle #247
• Add support for setting local bind address on client endpoint #250
• Add support for building debug RPM #221
• Support setting UCT level parameters from UCP API #209
• Set UCX_SOCKADDR_CM_ENABLE=y by default
• Disable DevX objects by default #185 #178
• Fixes in keepalive protocol #243 #205 #196 #184
• Fixes in rendezvous cancel protocol #233 #230 #227
• Fix hang in case of device fatal error #231
• Fix simultaneous disconnect #200
• Fix low bandwidth with multi-rail eager #194
• Fix uct_rdmacm_cm_cqs hash key #237
• Multiple fixes in io_demo test application
• Logging and memory tracker improvements

v1.9.0-pre40

Features:

• 100252a, 08e59bf Add ucp_tag_send_nbx() API with eager/rndv flags

Bugfixes:

• ea0288b Fix string buffer grow
• 13d124c Fix err code for rdma_<establish|accept> failure
• ac26299 Fix error handling in ucp_stream_am_handler
• 0aef7b3 Fix RC scatter-to-cqe configuration
• a864933, 472a471 Remove ucp_ep_flush progress callback when endpoint is closed
• de3247c, 529eb4f Track memory usage by application
• 271e82a Do not progress rendezvous operation if endpoint already failed
• 41500a7 Move keepalive to a separate progress callback
• df2dead Do not enable KA on CM lane
• 6aeb722 Do not access UCP endpoint after it's destroyed or its error callback has been invoked

v1.9.0-pre39

Features:

• Support random path for RoCE high availability, enable by UCX_IB_ROCE_RANDOM_PATH=y #127
• IODEMO: Non-blocking connect/disconnect #128 #139

Bugfixes:

• Fix error handling for AM_ZCOPY #131
• Fix premature keepalive on unconnected endpoint #133
• Fix TX_INLINE_RESP=0 behavior on DevX #136
• Fix a race condition when connecting to sockaddr #137
• IODEMO: Fix compilation on RH 7.2 #129
• IODEMO: Fix unknown conn_id #141

v1.9.0-pre38

Bugfixes:

• Fix post_recv called without buffers to post #117
• Reduce log level of simultaneous error during disconnect #119
• Fix RoCE LAG detection on MLNX_OFED 5.x #124
• Process registration cache invalidation queue during progress, to release stale regions #123
• Fix assertion checks in am_zcopy flow #125
• Add API to return registration cache information and counters #126

v1.9.0-pre37

Features:

• Add non-blocking resource cleanup by a background process, enabled by UCX_IB_CLEANUP_THREAD=y
• Make rdma_cm address/route resolve timeout configurable, instead of always using 1 second. Example: UCX_RDMACM_TIMEOUT=10s

Bugfixes:

• Fix heap corruption caused by ucm_set_event_handler() in multi-threaded application
• Fix RoCE LAG detection: take GID index from iface to find associated netdev, instead of always checking gid 0
• Disable backtrace by default to avoid deadlock with malloc/free
• Fix leak of listener requests during ucp_listener_destroy()
• Added lock on rc_iface->eps access, to fix race condition between main thread and pack_cb called from progress thread
• Use ibv_pd handle instead of device name as CQ hash key, for rdma_cm temporary QP

v1.9.0-pre36

Fixes for simultaneous disconnect (peer failure during ep_close):

• Flush operation failed with assertion due to endpoint changed number of lanes (#90, #91)
• Assertion failed that not all pending operations were removed, since flush_internal exited prematurely (#92)
• ep_close operation does not complete, since when ep_flushed returned error the lane was not accounted for (#93)

v1.9.0-pre35

• Fix potential race condition when keepalive is sent from main thread, while RC endpoint is created from progress thread, and keepalive tries to use a partially initialized endpoint:
  • UCT/RC: Protect rc_iface->ep_list with a spinlock
  • UCT/RC: Initialize ep->connected=0 before keepalive could run
• Fix race condition between installing malloc() hooks and IB async event thread, which can lead to segfault during application start

v1.9.0-pre34

• Reserve 1 CQ credit for qp-flush NOP, to avoid disabling CQ moderation
• Fix memory leak in IO demo test
• Simplify active connection counting in IO demo test

v1.9.0-pre33

• Fix reordering when destroying RC endpoint, due to releasing CQ and RDMA_READ credits
• Enhance logging in rdma_cm

v1.9-pre32

• Add limit for memory registration cache, configured by UCX_IB_RCACHE_MAX_REGIONS and UCX_IB_RCACHE_MAX_SIZE
• Fix keepalive protocol - don't send when QP is in INIT state
• Fix RPM build - missing io_demo installed file