James Smart wrote:
Joe Eykholt wrote:
It seems like there are two slightly different sets of rules depending
on whether dd_fcrport_size is zero or not, as specified by the LLDD.
In the first model, where dd_fcrport_size is zero, the transport
never sets dd_data at all. My understanding now is that its OK
for the LLDD to set it non-NULL, but not OK to change it after that.
I guess it would be OK but unnecessary to NULL it at dev_loss timeout
just before freeing the attached context. These are the usage rules
I didn't fully understand. These rules are really established by
how the LLDDs I/O routines use dd_data.
No. In all cases where you see dd_data, the dd_data field is a transport
owned field. The LLDD can change the contents of data pointed to by
dd_data, but an LLDD is not supposed to change the dd_data field itself.
I didn't know that was the intent. The comment "exported data" seems
to indicate the LLDD can use it.
In that case, there is no other good hook the LLDD could use for its
private data and this model would never be used. Since the transport
never sets dd_data to anything (except NULL on initial allocation) in
this model, I think it would be OK for the LLDD to set it once, but it
does cause usage problems for the I/O routines. I think most/all
LLDDs have non-zero dd_fcrport_size so this model isn't used.
In the second model, it doesn't seem like the LLDD has full control over
the contents pointed to by dd_data either, since when the remote
port is re-added the area pointed to by dd_data is cleared by the
transport, so we always start fresh. This is fine, but has
implications on how the context is used during devloss. For example,
it shouldn't be used for list linkage unless it's unlinked
before fc_remote_port_add. All that's in the LLDDs control, so it's OK.
The only place the transport should be zero'ing the contents are:
a) when it's allocated
b) when it was deleted but kept around to hold the target id binding, then
reallocated. Note: by deleted, I mean it fully transitioned through
devloss_tmo_callbk() so for all intents and purposes, its as if it was
freed and realloc'd.
If there's another case, it's bug and we should be fixing it.
I didn't see any other case, and didn't mean to imply that there was.
For libfc, I'm leaning towards continuing to use a non-zero
dd_fcrport_size
and the fc_rport_libfc_priv struct. libfc could use a separately
allocated struct like fc_disc_rport for the discovery and
rport (PLOGI, PRLI, etc.) state machines.
makes sense
This is all an effort to clean up some issues caused by creating "rogue"
fc_rports in libfc so that we would always have both an
fc_rport_libfc_priv
and an fc_rport allocated together, even before fc_remote_port_add().
It causes issues when we do remote_port_add and have to transition
the state from the rogue to the "real" rport.
In the meantime, the rogue could still be accessed by incoming requests,
or new RSCNs, and those changes wouldn't get reflected to the real rport.
It's messy, and hard to analyze all the potential problems, so I'm
trying to fix that.
I really appreciate your help! Thanks a bunch!
Joe
Ok.. Good Luck.
-- james
Thanks again.
Joe
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
