On 1/29/14, 11:07 PM, Richard Yao wrote: > I caught this last week when running valgrind on ZoL's zdb. it turned > out to be a false positive because zdb killed itself to test ZFS' > resilence against kernel crashes. If I recall correctly, I concluded > that this is freed properly in the zio_root.
Your analysis sounds about right. We should always come through l2arc_write_done from the zio callback and free any buffers in there (there was a bug previously where freeing was predicated on grabbing the hash lock - that's fixed now). Regular ARC eviction should never destroy buffer headers which are in an ARC_L2_WRITING state - if it does, then that's a potential leak. -- Saso _______________________________________________ developer mailing list [email protected] http://lists.open-zfs.org/mailman/listinfo/developer
