Re: [HPDD-discuss] [PATCH] staging: lustre: obd_support: Add obd_cpt_alloc function
On May 3, 2015, at 1:53 AM, Julia Lawall wrote:
On Sat, 2 May 2015, Drokin, Oleg wrote:
>
> On May 2, 2015, at 5:16 PM, Julia Lawall wrote:
>
>> Summarize OBD_CPT_ALLOC_GFP, OBD_CPT_ALLOC, and OBD_CPT_ALLOC_PTR as a
>> function, obd_cpt_alloc.
>>
>> Signed-off-by: Julia Lawall <Julia.Lawall(a)lip6.fr>
>>
>> ---
>>
>> Some questions: Is the name OK? Is the NULL test needed? If not, should
>> the call to kzalloc_node with the call to cfs_cpt_spread_node just be
>> inlined into the call sites?
>
> I think we don't need this function at all, we can use kzalloc/kzalloc_node
directly with cfs_cpt_spread_node call in.
So everywhere the CPT macro is called, it is known that the value is not
NULL? I looked at some call sites, but it's not obvious to determine
that.
It's not obvious, but I believe this is true now.
Basically in lustre/ptlrpc/service.c we use service->srv_cptable
and that comes from ptlrpc_register_service:
cptable = cconf->cc_cptable;
if (cptable == NULL)
cptable = cfs_cpt_table;
….
service->srv_cptable = cptable;
service->srv_cpts = cpts;
service->srv_ncpts = ncpts;
that on the client it's only called from:
lustre/ldlm/ldlm_lockd.c::ldlm_setup() where we unconditionally assign
.psc_cpt = {
.cc_pattern = ldlm_cpts,
},
But even if there was a different caller, we always use cfs_cpt_table as fallback.
It's also directly used in ptlrpc_hr_init():
ptlrpc_hr.hr_cpt_table = cfs_cpt_table;
Two callers in lustre/ptlrpc/nrs.c use the same stuff as above.
One caller in lustre/ptlrpc/nrs_fifo.c uses nrs_pol2cptab which is defined as
nrs_pol2svc(policy)->srv_cptable which is again same as above.
There are no other callers.
Bye,
Oleg