Nicolas Morey
fd5f1a4feb
- Core
- ofi_atomic_queue: convert from MPMC to MPSC
- hmem: Add additional CUDA wrapper functions
- man: Fix reference to fi_mr_raw_attr()
- common: Fix byte order of AF_IB ofi_addr_set_port
- common: Do not assert on truncated buffers
- man: Update efa man page to reflect zero copy mode deprecation
- configure.ac: Fix --enable-xpmem=yes
- man: Clarify FI_INJECT + FI_HMEM restriction
- core: Replace indirect held, lock, and unlock function calls
- CXI
- Fix spelling of "receive" across CXI provider
- Extend timeout for sw_max_recv msg test
- Refuse cancel for receives with RDMA in flight
- Add test for TLE pool sharing from the default service
- Added FI_SUCCESS return to rnr recv cb
- Minor tweak to append_oflow test
- Handle LE append failure for RNR recv
- Fix minor size check in RNR send_common
- Check FI_CXI_CTRL_RX_EQ_MAX_SIZE for 0
- New mr test to check buffer attributes
- EFA
- use cmocka's "void ** state" signature in unit tests for C23
- honor homogeneous_peers in efa_rdm_interop_rdma_read
- Drive real completion handler in local-read unit test
- NULL local_read_pkt_entry after data-copied hand-off
- Track efa_rx_pkts_held with a per-pkt flag
- Fix pkt leak in efa_rdm_ep_post_queued_pkts on error
- Fix direct_ope leak on error in efa-direct send/recv/rma
- remove trailing whitespaces in the mr unit test.
- Gate the pps interface with feature flag
- Add rdma-core path support for WQE-level PPS hints
- Implement WQE level PPS hints in data-path-direct
- Add efa-specific operational flags for high PPS
- Guard hardware counter with FI_EFA_USE_HW_CNTR
- Add feature flag FI_EFA_USE_HW_CNTR
- Use hardware counter in fi_cntr_open
- Add fi_cntr_wait for hardware counter
- Attach hardware completion counter to QP
- Extend GDA domain op to support hardware cntr with external memory
- Add hardware counter operations
- Advertise hardware counter max value in fi_getinfo
- Set the correct cntr_cnt in domain_attr
- Baseline the read nack extra feature
- Baseline the runt extra feature
- Baseline the connid header extra request
- Validate dmabuf provided when setting FI_MR_DMABUF
- Fix pkt_entry double-free during endpoint teardown
- Add missing txe release check in dc ctsdata packet
- Destroy buffer pools when pke_vec alloc fails in ep_open
- Fix info and util_ep leak on efa_recv_wr_vec calloc failure
- Add assert guards for iov_count bounds in atomic, rma, and srx
- Propagate robuf allocation failure from peer_construct
- Check util_cq refcount before destroying ibv_cq in close
- Clear EFA_RDM_PKE_RNR_RETRANSMIT in peer_destruct
- Bounds-check fi_mr_attr.iface before indexing g_efa_hmem_info
- Initialize srx_lock early in efa_domain_open
- Stop returning early from efa_domain_close
- Use ret directly in efa_domain_open
- Do a proper cleanup in case of a fork handler failure
- Do a proper cleanup in case of an unsupported endpoint type
- Drop dead assignment to local efa_domain
- Remove trailing whitespace in efa_domain.c
- Write CQ error instead of EQ error for unsolicited write recv
- Extract efa_rdm_cq_write_error() helper for CQ error with EQ fallback
- Always build rdma-core WQE post functions in header
- Fix length field in data_path_direct tracepoint
- Fix out-of-bounds array access in tracepoint post_send
- Add unit-test for util_foreach_unspec
- Add FI_EFA_FEATURE_OPS for runtime feature discovery
- test: Verify EAGAIN when pre-handshake queue is full
- Separate CNTR into independent efa-direct and efa-rdm implementations
- Suppress duplicate error CQ/counter for ops errored synchronously
- Add test confirming multi-packet send is not susceptible to partial-post
- Fix same double-free pattern in multi-segment RDMA read path
- Fix double-free on partial multi-segment RDMA write failure
- Store EFA-internal txe flags in internal_flags, not fi_flags
- Fix -Wpointer-arith warnings on ARM
- Remove handshake requirement for DC
- Remove zero copy receive path, keep send path for compat
- Move memory alignment helper to header as static inline
- Inline and remove efa_rdm_ep_alloc_txe
- Check if the send queue is full earlier
- Add ep->send_pkt_entry_vec_size
- Check all descs for HMEM in efa_post_send inline path
- test: Fix flaky QPN collision in implicit AV unit tests
- Fix NULL deref in efa_av_reverse_av_remove on QPN collision
- Fix FI_INJECT in efa direct
- Remove incorrect assert in test_efa_data_path_direct_qp_gen_initialization
- Improve error message for ibv_create_ah failure
- Subtract prefix size in inject assertion for dgram
- Optimize shm address retrieval in RDM operations
- test: Cast the qp gen when doing the comparison
- Fix race between fi_av_lookup and fi_av_remove
- Fix the max_msg_size validation for efa-direct
- Preserve shm MR close error in efa_rdm_mr_close
- Fix ineffective error check on ofi_get_page_size()
- Fix use-after-free when TX and RX share the same CQ
- Fix unchecked strndup return in get_sysfs_path
- Improve the warn log in efa_mr/efa_rdm_mr
- Introduce a warning macro for fi errno
- Fix unused return codes
- Honor user-requested QP sizes within device limits
- Fix missing mem_desc and iface initialization in non-p2p path
- Separate core MR logic from RDM-specific MR implementation
- LNX
- Update handling of deprecated FI_AV_MAP
- Fix lnx capability settings and checking
- Move environment variables into a global struct for easy access
- Cleanup and reorganization
- OPX
- Lower GDRcopy threshold/Allow HMEM MP Eager
- Initialize deferred HFISVC receive contexts
- Make OPENED MR notify non-owning for rzv completion
- Document HMEM-dependent FI_OPX_RZV_MIN_PAYLOAD_BYTES defaults
- Reject incompatible MR registrations when HFISVC is enabled
- Handle MP eager FI_CLAIM receives
- Drop mm lock before memory operations
- Remove SDMA queue ring size workaround
- Remove global hfi_local_info
- Fix origin_rx in realibility ping, ack and nack
- Store HFI selection per-domain
- Use generation-agnostic fabric name and RDMA device domain name
- Populate fid_nic in fi_getinfo with device/bus/link attributes
- Remove fabric from linked list in close paths
- Restore fd_verbs (hfi direct) support
- Dual plane: Send only context
- Dual plane: Environment variables
- Fix no common tx context error
- Dual Plane: Stripe data across two HFISVC clients in dual/single plane
- Fix reliability origin_rx access
- sriov support for lmc/lid hairpin
- Dual plane: Reliability and reply path changes
- tracer BEGIN/END instrumentation fixes
- Unsubscribe when entry is removed from cache.
- FI_MR_DMABUF mr_regattr using invalid addr
- Catch hfi config error earlier and fail
- Add valid() calls to opx caching
- Do not fail kdreg2
- Don't use FI_DELIVERY_COMPLETE for MSG/Tagged sends
- Add full FI_DELIVERY_COMPLETE support for MSG/Tagged sends
- HSA_STATUS_ERROR_INVALID_ARGUMENT IPC RZV Send
- SLES compilation fix
- Remove #if 0 dead code
- MP Eager data validation fix
- Dual plane - addressing, plane selection and shm
- Populate PCI bus attributes in fi_info from sysfs
- OPX Tracer v2
- Dual plane: Multiple tx and ibv context
- Restore debugging #ifdef
- change all CUDA calls to use ofi wrapper functions
- Reduce warnings
- Fix use_cnt/mmu windows
- Fix rendezvous partial registration
- Unlock mm in shm signal
- Open HFISVC completion queue on RX CQ when separate from TX CQ (#1321)
- Fix extended addr allocation during av insert
- Compute sbuf_offset for non-dmabuf MRs in hfisvc RTS
- Fix CTS replay payload for RMA GET
- Support 9B HFI Service
- store dmabuf base_addr as page-aligned fd start
- Always poll domain and CQ's HFI service completion queues during CQ Read
- Fix DMABUF HFISVC addr/offset calculation
- Fix MP Eager replays for 9B headers
- Sync HMEM stream before destroy
- Prevent RDMA lib close in uninitialized state
- Only close DMABUF fd when created by OPX
- Remove dead code in rendezvous CTS path
- Fixing issue in av_insert when using av_table
- Trivial FI_WARN update
- Initialize OPX ref counters
- Fix build error in hmem path for av_type
- Use ofi_atomics for reference counting.
- Changing av_map to work like av_table internally
- Invalid RDMA ops ref counter decrement
- Segfault with NCCL/RCCL Plugin & HFISVC
- Enable container build
- Fix non-powers of 2 rcvhdrcnt
- Fix mismatched alignment attribute between packet header & payload unions.
- Use correct device with DMABUF support
- Fix use of DMABUF fd in the OPX mr
- Update SDMA/RZV threshold for AMD GPUs
- fix fi_opx_open_command_queues() segfault
- Only require DMABUF for HFISVC for HMEM
- Added hmem_dev_reg_handle so receive side can use ROCr Copy
- Remove debug log of every reliability inject
- Add DMA-BUF offset to buffer offset calculation
- DMABUF support for HFI service
- MR Registration with HFI service
- Fix reliability key in reliability debug logs
- Remove unnecessary csr reads
- Dynamically growable access_key pool for HFI service.
- PSM3
- Internal polling must use timeout to prevent infinite loop during resources acquiring
- RXD
- Remove rxd_mr_verify dead code
- RXM
- Fix ignored domain_attr->cq_data_size hint
- SHM
- DL initialize mem and monitors
- Register max_gdrcopy_size env var
- Do not return local cmd copy in smr_discard
- Remove init_fn from cmd queue release path
- Fix the error flag propagation
- Let atomic_inline use smr_flags for format
- Revert to a lock-unlock inject pool
- Fix compile warning coming from shm
- Do not take inject for op_read_req
- Do not check for total_len < inject_size on rma_fast
- Push cmd back to stack on error
- remove cmd_entry ptr
- Remove 0-byte copy SAR
- Move inject pool above command stack
- Use hdr.sm
Signed-off-by: Nicolas Morey <nmorey@suse.com>
Build Results
Current state of libfabric in openSUSE:Factory is
The current state of libfabric in the devel project build (science:HPC)