forked from pool/libfabric
Update to v2.4.0 #4
No Assignees
Notifications
Due Date
No due date set.
Dependencies
No dependencies set.
Reference: HPC/libfabric#4
Reference in New Issue
Block a user
Delete Branch "NMorey/libfabric:main"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
FI_GETWAITOBJforfi_controlfi_tostr()datatypesep_list_lockduring counter progressSigned-off-by: Nicolas Morey nmorey@suse.com
- Core - hmem/cuda: Adding more robust libgdrapi libpaths - Update bindings/rust/README.md to reflect the recommended build process. - Update build.rs to support both cargo build & cargo publish work directories. - Update Cargo.toml in preparation for crates.io publishing. - configure: Fix sanitizer detection logic - Introduce a lightweight Rust bindings for Libfabric, using bindgen. - include/ofi_indexer: introduce new ofi_array_at_max function - man/fi_cxi: fixup info for FI_CXI_RDZV_GET_MIN - man/fi_getinfo: Update the capabilities with mode bits requirements - man/fi_cq: Document `FI_GETWAITOBJ` for `fi_control` - man/fi_fabric: Update `fi_tostr()` datatypes - CXI - Bump provider support up to libfabric 2.4 - Add domain rx match mode override - Set rendezvous eager size default to 2K - Change cuda dmabuf default to enabled - Do not abort if MR match count do not reconcile - Allow CP for triggered CQ to remap to Best Effort - Fix sl-driver path for testing - Set max domain TX CQs to 14 - Use cxil_alloc_trig_cp to distinguish trig and tx cmdqs - Add FI_EBUSY debug messages - Fix validation of service id - Fix criterion test_sw tap files - Cxip_cmdq_cp_modify fix - Fix RNR protocol send byte/error counting - Release TX credit when pending RNR retry - Update rocr test fine grained flags - Fix DEVICE in fi_info_test - Introduce non-debug tracing - Reset timer on rx of ARM packet - Fix performance issue with close_mc() - Increase vni range in auth_key tests - Support auth_key ranges - Fix use of hw_cps and memory leak - EFA - Fix cq data size in efa-rdm pkt post - fix test_efa_rdm_mr_reg_cuda_memory unit test - adjust the memory barrier positions - Optimize RTW packet sending by replacing efa_rdm_ope_post_send - Adjust logging level for txe releases - Add tracepoints for handshake - Add flags to MR logs - Grow efa_tx_pkt_pool and ope_pool during rdm ep creation - Do not use rdma write when unsolicited recv support is inconsistent - Determine whether using device rdma based on p2p - Introduce pke generation counter for protocol path - Enable data path direct for efa-rdm - Update the function signature for efa_data_path_direct_cq_initialize - Move efa_cq_open_ibv_cq to efa_cq.c - Do not track rx pkt pool for non-debug build - Temporarily disable FI_OPT_EFA_SENDRECV_IN_ORDER_ALIGNED_128_BYTES support for efa protocol - do not ignore local read completion - Add missing lttng tps in efa_post_send - Fix the remote cq data flags for zcpy recv - Optimize the WQE post in data path direct - fix typos in error messages - Only show help message for OPE warn logs - configure: replace no-brake space with regular space character - Remove unused function declarations - Acquire CQ's `ep_list_lock` during counter progress - Add asserts to detect erroneous CQE dereferences - Ignore rma completion to a removed peer - Remove the incorrect check for device max_msg_size - Fix function signature mismatch - Set FI_RX_CQ_DATA for efa direct with NULL hints - Do not fail fi_getinfo for the wrong fabric - Log warnings only for internal OPE failures or if CQ error entry not written - Add unit tests for LRU AH eviction - Evict AH with no explicit AV entries when AH limit reached - Add locking assertions and update unit tests - Remove efa_conn_release unsafe - Require FI_RX_CQ_DATA on devices without unsolicited write recv - Add LLTng tracepoints for direct data path operations - Don't warn users about non-EFA devices - Support FI_RX_CQ_DATA for efa-direct - Fix deadlocks in AV insert/remove/close and CQ read paths - Don't try to release a lock that is not taken - set RUNPATH if custom rdma-core provided - Remove rx_msg_flags from efa_rdm_msg_recv/efa_rdm_msg_recvv - Update tracepoints in the receive path - Slide recv-win on RTM/RTA error - Insert read and write packets to tx debug list - LNX - remove force setting DEVICE_ONLY flag - set core hints proto to UNSPEC - remove iov count failures - add wait object implementation - OPX - Don't fail configure when OPX unhappy - Add note to FI_OPX_SDMA_MIN_PAYLOAD_BYTES doc - Simplify uapi configuration - Unionize 9B and 16B packet SCB models in endpoint structs. - Support shared contexts in hfisvc bts - Fix replays for multi-packet eager - Don't retry forever in send rendezvous. - Don't ACK packets that were never received - Segfault in opx_hfi_rdma_context_open() on 2nd endpoint opened - Fix seg fault in finalize - Fix SDMA writev error when RDMA core functions are being used. - Add back accidentally removed opx_domain_hfisvc_poll() - Add missing function pointers for HFI service - Check uapi for hfisvc/HFI1 direct verbs - Rename hfisvc to opx-hfisvc - Move submodule to rdma core - Remove stx/srx support in OPX - Register MRs with HFI service - Ensure SDMA packet lengths are 8-byte multiples - Use HFI service by default if enabled in the driver. - fixup goto labels that need statements - Update hfisvc_client to 64-bit atomics - HFISVC: Fix replay payload - Disable HFI Service by default. - Disable use of HFI service when driver does not support it. - Update hfisvc_client to latest patch - Only open IPC cache if HMEM initialized and IPC enabled - Handle extended rx bits in common 9B code - Add IPC to 16B header path - Make sriov-alpha limitations CN5000-only - Remove cmake build for hfisvc_client library - Handle completion errors from HFI service - Fix setting of rc in deferred recv rts - Additional HFI Service support changes - HFI Service initial support - Asynchronous HMEM memcopy for IPC Signed-off-by: Nicolas Morey <nmorey@suse.com>Closing here because the associated Project PR has been closed.
Closing here because the associated Project PR has been closed.