I have modified the fix a little bit: http://cr.opensolaris.org/~pavelf/6897605-xattr-v2
It will skip VN_RELE(xattr) only if the r_hasq is same. The advantage is that more memory is freed. Originally, I was resistant to use the value of r_hasq - since it should be a private data of lower layer and opaque to the nfs_reclaim code, but nfs4_active_reclaim() is already aware of the implementation details of rnode table. Cheers, Pavel On 11/09/09 14:22, Calum Mackay wrote: > On 09/11/09 13:06, Frank Batschulat (Home) wrote: >>> 2) introduce VN_RELE_ASYNC(xattr) >>> pros: - more memory is reclaimed >>> cons: - slower code (extra work when the system is low on memory) >>> - adding a new code - need to create a new taskq (pool?) >> >> yepp, we'd have to create a new taskq as well, we probaly do not want >> to use >> the system taskq here. > > yup. > > We looked at this a year or two ago, to work-around an issue where the > network layer was calling into VFS from interrupt context, where we > run into problems if we end up doing rnode inactivation. > > It opened up a larger can of worms (e.g taskq setup from interrupt > context) than we were trying to fix, however, and since calling into > VFS from interrupt is poor practice in any case, we decided that > adding the async rele wasn't really worth the effort, at least in that > case. > > cheers, > c.
