Re: [RFC 00/22] target: se_node_acl LUN list RCU conversion

[Christoph Hellwig <hch@infradead.org>]

On Tue, Mar 31, 2015 at 11:51:56PM -0700, Nicholas A. Bellinger wrote:
> Since last week, enable/disable device_list code has been converted to
> use mempool + call_rcu() and performs the RCU pointer swap in
> se_node_acl->lun_entry_hlist[] under se_node_acl->lun_entry_mutex.

Why use a mempool there?  Please take a look at my branch, it's much
simpler and logical than the old version.

> There are a few more target_core_pr.c pieces that need to be updated to
> handle NULL se_node_acl->lun_entry_hlist[] pointers, along with proper
> rcu_dereference_protected() and rcu_access_pointer() notations for
> se_dev_entry pointer access outside of RCU read lock.

In my branch I've been over all the places to make sure they handle
the NULL case.

> > that implements this scheme.  Note that I dropped the percpu refcount
> > changes - the reference is only taken in the SCSI3-PR slow path, and the
> > percpu refcount API is extremly cumbersome.
> > 
> 
> Well, there still needs to be some manner of reference counting on
> se_dev_entry once rcu_read_lock() is released for the REGISTER w/ I_PORT
> and REGISTER_AND_MOVE cases with a non-local I_T_L se_dev_entry
> reference.

The refcount is still there in the good old atomic_t-based version,
which is perfectl fine for the PRIN and PROUT subcommands which aren't
the I/O fast path.