Summarize OBD_CPT_ALLOC_GFP, OBD_CPT_ALLOC, and OBD_CPT_ALLOC_PTR as a
function, obd_cpt_alloc.
Signed-off-by: Julia Lawall <[email protected]>
---
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?
drivers/staging/lustre/lustre/include/obd_support.h | 9 +++++++++
1 file changed, 9 insertions(+)
diff --git a/drivers/staging/lustre/lustre/include/obd_support.h b/drivers/staging/lustre/lustre/include/obd_support.h
index 2991d2e..3d380f0 100644
--- a/drivers/staging/lustre/lustre/include/obd_support.h
+++ b/drivers/staging/lustre/lustre/include/obd_support.h
@@ -655,6 +655,15 @@ do { \
#define OBD_CPT_ALLOC_PTR(ptr, cptab, cpt) \
OBD_CPT_ALLOC(ptr, cptab, cpt, sizeof(*(ptr)))
+static inline void *obd_cpt_alloc(struct cfs_cpt_table *cptab, int cpt,
+ size_t size, gfp_t flags)
+{
+ return (cptab) == NULL ?
+ kzalloc(size, flags) :
+ kzalloc_node(size, flags, cfs_cpt_spread_node(cptab, cpt));
+}
+
+
# define __OBD_VMALLOC_VEROBSE(ptr, cptab, cpt, size) \
do { \
(ptr) = cptab == NULL ? \
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 <[email protected]>
>
> ---
>
> 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.
What we do need is obd_cpt_alloc_large similar to how we need obd_alloc_large (I know I still owe you a proper patch with that).
The only differences between the two would then be passing down of the cpt (and it's use) or not.
Bye,
Oleg-
On Sat, May 02, 2015 at 11:16:48PM +0200, Julia Lawall wrote:
> 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?
>
> drivers/staging/lustre/lustre/include/obd_support.h | 9 +++++++++
> 1 file changed, 9 insertions(+)
>
> diff --git a/drivers/staging/lustre/lustre/include/obd_support.h b/drivers/staging/lustre/lustre/include/obd_support.h
> index 2991d2e..3d380f0 100644
> --- a/drivers/staging/lustre/lustre/include/obd_support.h
> +++ b/drivers/staging/lustre/lustre/include/obd_support.h
> @@ -655,6 +655,15 @@ do { \
> #define OBD_CPT_ALLOC_PTR(ptr, cptab, cpt) \
> OBD_CPT_ALLOC(ptr, cptab, cpt, sizeof(*(ptr)))
>
> +static inline void *obd_cpt_alloc(struct cfs_cpt_table *cptab, int cpt,
> + size_t size, gfp_t flags)
> +{
> + return (cptab) == NULL ?
These parens aren't needed any more.
I feel like people shouldn't deliberately call this with dptab == NULL.
I looked at it a bit and wasn't sure, (was sleepy though), so it's maybe
safest to keep the test.
I wish that cfs_cpt_spread_node() accepted NULL pointers so that we
didn't have to have the check for "cptab == NULL". But your patch seems
like the way forward for now.
> + kzalloc(size, flags) :
> + kzalloc_node(size, flags, cfs_cpt_spread_node(cptab, cpt));
> +}
> +
regards,
dan carpenter
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 <[email protected]>
> >
> > ---
> >
> > 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.
> What we do need is obd_cpt_alloc_large similar to how we need
> obd_alloc_large (I know I still owe you a proper patch with that). The
> only differences between the two would then be passing down of the cpt
> (and it's use) or not.
I saw that patch. Thanks.
julia
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 <[email protected]>
>>>
>>> ---
>>>
>>> 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-