Syscall + libfabric bug on buffer reuse

  • Using syscall intercept in their SOD2D+MultiVDB pipeline causes To Do% deadlocks

  • The error is "decompress_and_parse_entries_standard() Unexpected end of buffer while parsing name"

  • The issue doesn't occur with FUSE

  • A suggested cause is libfabric's MR cache returning reused pointers in forward_get_dirents()

  • Setting FI_MR_CACHE_MAX_COUNT=0 fixes the issue

Solved using a thread_local buffer

Assignee Loading
Time tracking Loading