Tags: cornelisnetworks/opa-psm2
Tags
Reset rpm_release_extension to "1" for psm-v12.0 series Signed-off-by: Brendan Cunningham <bcunningham@cornelisnetworks.com>
Fix silent data error on FECN-marked TID packet for unaligned receive. When a TID-receive packet is marked with the FECN bit, PSM2 is responsible for copying the payload to the destination memory rather than the TID hardware. This TID-FECN path in PSM2 did not account for unaligned receive buffers, buffers whose start address is not a 4B multiple. As a result, the payload would be copied to a location 1-3 bytes off from the correct location. Fix in this commit accounts for unaligned buffer start and copies data to correct location. FECN-marking only occurs when OPA Congestion Control Architecture (CCA) is enabled on the fabric. Only OPA users who have explicitly enabled the CCA feature via the FM XML are exposed to this issue. Bump PSM2 version to 11.2.230 to distinguish libpsm2 with this fix from prior versions. Fixes #64. Signed-off-by: Brendan Cunningham <bcunningham@cornelisnetworks.com>
Bump PSM2 version to 11.2.229. Signed-off-by: Brendan Cunningham <bcunningham@cornelisnetworks.com>
Updates the PSM2 library to match release 11.2.228
Corresponds to OPXS release OPENIB_INSTALL_10_11_1_0_10.
Changes:
* hfi_userinit_internal: change _HFI_INFO to _HFI_ERROR or add
"Warning:".
* Always print packet opcode as 0x%x.
* Print bad opcode at INFO instead of DBG level.
* Do not use cuMemcpy*() for host<>GPU transfers when the
PSM2_MQ_FLAG_GDRCPY_ONLY flag is passed into a psm2_mq*{send,recv}*()
function.
* Moved gpu_p2p_supported() into a block where for sure it is in a CUDA
context.
* Fixes crash when GDR FD was closed but PTL had outstanding CUDA
messages that required a valid GDR FD.
* Safely share PSM2 GDR file descriptor with many EPs.
* Fix format string using 32-bit %d for 64-bit argument.
* Do not try to pin 0-length GPU pages on GDRCopy.
* Add CUDADBG tracemask (0x2000) for most CUDA calls, events.
* Validate gdr_convert_gpu_to_host_addr() arguments.
* Code cleanup, update copyright headers, update README.
Signed-off-by: Brendan Cunningham <bcunningham@cornelisnetworks.com>
ips_cmpxchg: remove compile-breaking bad cast-to-struct. uint32_t ptr argument in compare-and-swap code snippet should not be cast to pointer-to-uint32_t[100] struct. Signed-off-by: Brendan Cunningham <bcunningham@cornelisnetworks.com>
Updates the PSM2 library to match release 11.2.205 Corresponds to IFS release OPENIB_INSTALL_LINUX_opa-10_11_0_1_10_11_0_1_2 with additional post-release fix. Signed-off-by: Brendan Cunningham <bcunningham@cornelisnetworks.com>
Bump PSM2 version to 11.2.203. Signed-off-by: Brendan Cunningham <bcunningham@cornelisnetworks.com>
Updates the PSM2 library to match release 11.2.202 Corresponds to IFS release OPENIB_INSTALL_LINUX_opa-10_11_10_11_0_0_578 with additional post-release fix. Changes: * Check PSM2_GPUDIRECT_RDMA_SEND, PSM2_GPUDIRECT_RDMA_RECV before disabling GDRCopy when PSM2_GPUDIRECT=0. * Remove bashisms from Makefile. H/t Brian T. Smith. * Fix syntax errors in Makefile. H/t Chuck Cranor. * Replace CUDA events with cuStreamSynchronize. H/t Jonas Hahnfeld. * Lazily allocate buffers, events, and streams for GPUs. This allows PSM2 to be initialized before GPUs are chosen. H/t Jonas Hahnfeld. * Remove ad_context from psmi_mpool_create_for_cuda; ad_context was not used. H/t Jonas Hahnfeld. * Other CUDA code cleanup and improvements to allow CUDA deferred initialization. * Fix crash on uninitialized CUDA IPC variable. H/t Rémi Lacroix. * Other minor fixes. * Update contributors. Signed-off-by: Brendan Cunningham <bcunningham@cornelisnetworks.com>
Updates the PSM2 library to match release 11.2.201 Corresponds to IFS release OPENIB_INSTALL_LINUX_opa-10_11_10_11_0_0_578 Changes: * Check PSM2_GPUDIRECT_RDMA_SEND, PSM2_GPUDIRECT_RDMA_RECV before disabling GDRCopy when PSM2_GPUDIRECT=0. * Remove bashisms from Makefile. H/t Brian T. Smith. * Fix syntax errors in Makefile. H/t Chuck Cranor. * Replace CUDA events with cuStreamSynchronize. H/t Jonas Hahnfeld. * Lazily allocate buffers, events, and streams for GPUs. This allows PSM2 to be initialized before GPUs are chosen. H/t Jonas Hahnfeld. * Remove ad_context from psmi_mpool_create_for_cuda; ad_context was not used. H/t Jonas Hahnfeld. * Other CUDA code cleanup and improvements to allow CUDA deferred initialization. * Other minor fixes. * Update contributors. Signed-off-by: Brendan Cunningham <bcunningham@cornelisnetworks.com>
PreviousNext