[ofiwg] OFIWG 8/5/2025 Minutes

[Xiong, Jianxin]

8/5/2025

* Participants *

Alex McKinley (Intel)
Ben Lynam [Cornelis]
Jerome Soumagne (HPE)
Jianxin Xiong (Intel)
John Byrne
Kaushik Argonne
Ken Raffenetti (ANL)
Sai Sunku [AWS]
Shi Jin (AWS)
Stephen Oost (Intel)
Steve Welch [HPE]
Zach Dworkin (Intel)
Rajalaxmi (Intel)

* Summary *

There is a thread-safety issue (#10943) with the memhooks monitor for system memory.  The memhooks monitor patches the symbol table to redirect system calls to custom code segments. The patching / restore process is not thread safe and can cause other thread calling libc memory allocation related functions to segfault. There is no elegant way to resolve this. The proposed approach is to re-oder the system memory monitors so that memhooks is not set to the default if other options are available. Comments from John, Steve, and Ken indicated that userfault fd and kdereg2 are preferable in general. 

There are a few DAOS related PRs (#10718, #10783, #10997) might no longer be needed. Jerome will take a further look and take appropriate actions. The RPC API PR (#10965) is on hold, pending more evaluation from the DAOS team for it's necessity and implementation concerns.

----------------
Jianxin Xiong
Fabric Software
Intel Corporation
Jianxin.xiong at intel.com