Small optimization for ib_map_mr_sg() and ib_map_mr_sg_pi()

[PATCH for-next 0/2] Small optimization for ib_map_mr_sg() and ib_map_mr_sg_pi()

Posted by Junxian Huang 2 weeks, 4 days ago

ib_map_mr_sg() and ib_map_mr_sg_pi() allow ULPs to specify NULL as
the sg_offset/data_sg_offset/meta_sg_offset arguments. Drivers who
need to derefernce these arguments have to add NULL pointer checks
to avoid crashing the kernel.

This can be optimized by adding dummy sg_offset pointer to these
two APIs. When the sg_offset arguments are NULL, pass the pointer
of dummy to drivers. Drivers can always get a valid pointer, so no
need to add NULL pointer checks.

Junxian Huang (2):
  RDMA/core: Add dummy sg_offset pointer for ib_map_mr_sg() and
    ib_map_mr_sg_pi()
  RDMA: Delete NULL pointer checks for sg_offset in .map_mr_sg ops

 drivers/infiniband/core/verbs.c         | 12 +++++++++---
 drivers/infiniband/hw/mlx5/mr.c         | 18 ++++++------------
 drivers/infiniband/sw/rdmavt/trace_mr.h |  2 +-
 3 files changed, 16 insertions(+), 16 deletions(-)

--
2.33.0

Re: [PATCH for-next 0/2] Small optimization for ib_map_mr_sg() and ib_map_mr_sg_pi()

Posted by Leon Romanovsky 2 weeks, 3 days ago

On Tue, Nov 05, 2024 at 08:08:39PM +0800, Junxian Huang wrote:
> ib_map_mr_sg() and ib_map_mr_sg_pi() allow ULPs to specify NULL as
> the sg_offset/data_sg_offset/meta_sg_offset arguments. Drivers who
> need to derefernce these arguments have to add NULL pointer checks
> to avoid crashing the kernel.
> 
> This can be optimized by adding dummy sg_offset pointer to these
> two APIs. When the sg_offset arguments are NULL, pass the pointer
> of dummy to drivers. Drivers can always get a valid pointer, so no
> need to add NULL pointer checks.
> 
> Junxian Huang (2):
>   RDMA/core: Add dummy sg_offset pointer for ib_map_mr_sg() and
>     ib_map_mr_sg_pi()
>   RDMA: Delete NULL pointer checks for sg_offset in .map_mr_sg ops
> 
>  drivers/infiniband/core/verbs.c         | 12 +++++++++---
>  drivers/infiniband/hw/mlx5/mr.c         | 18 ++++++------------
>  drivers/infiniband/sw/rdmavt/trace_mr.h |  2 +-
>  3 files changed, 16 insertions(+), 16 deletions(-)

So what does this change give us?
We have same functionality, same number of lines, same everything ...

Thanks

> 
> --
> 2.33.0
> 
>

Re: [PATCH for-next 0/2] Small optimization for ib_map_mr_sg() and ib_map_mr_sg_pi()

Posted by Junxian Huang 2 weeks, 3 days ago


On 2024/11/6 20:08, Leon Romanovsky wrote:
> On Tue, Nov 05, 2024 at 08:08:39PM +0800, Junxian Huang wrote:
>> ib_map_mr_sg() and ib_map_mr_sg_pi() allow ULPs to specify NULL as
>> the sg_offset/data_sg_offset/meta_sg_offset arguments. Drivers who
>> need to derefernce these arguments have to add NULL pointer checks
>> to avoid crashing the kernel.
>>
>> This can be optimized by adding dummy sg_offset pointer to these
>> two APIs. When the sg_offset arguments are NULL, pass the pointer
>> of dummy to drivers. Drivers can always get a valid pointer, so no
>> need to add NULL pointer checks.
>>
>> Junxian Huang (2):
>>   RDMA/core: Add dummy sg_offset pointer for ib_map_mr_sg() and
>>     ib_map_mr_sg_pi()
>>   RDMA: Delete NULL pointer checks for sg_offset in .map_mr_sg ops
>>
>>  drivers/infiniband/core/verbs.c         | 12 +++++++++---
>>  drivers/infiniband/hw/mlx5/mr.c         | 18 ++++++------------
>>  drivers/infiniband/sw/rdmavt/trace_mr.h |  2 +-
>>  3 files changed, 16 insertions(+), 16 deletions(-)
> 
> So what does this change give us?
> We have same functionality, same number of lines, same everything ...
> 

Actually this is inspired by an hns bug. When ib_map_mr_sg() passes a NULL
sg_offset pointer to hns_roce_map_mr_sg(), we dereference this pointer
without a NULL check.

Of course we can fix it by adding NULL check in hns, but I think this
patch may be a better solution since the sg_offset is guaranteed to be
a valid pointer. This could benefit future drivers who also want to
dereference sg_offset, they won't need to care about NULL checks.

Junxian

> Thanks
> 
>>
>> --
>> 2.33.0
>>
>>

Re: [PATCH for-next 0/2] Small optimization for ib_map_mr_sg() and ib_map_mr_sg_pi()

Posted by Leon Romanovsky 2 weeks, 3 days ago

On Wed, Nov 06, 2024 at 09:12:47PM +0800, Junxian Huang wrote:
> 
> 
> On 2024/11/6 20:08, Leon Romanovsky wrote:
> > On Tue, Nov 05, 2024 at 08:08:39PM +0800, Junxian Huang wrote:
> >> ib_map_mr_sg() and ib_map_mr_sg_pi() allow ULPs to specify NULL as
> >> the sg_offset/data_sg_offset/meta_sg_offset arguments. Drivers who
> >> need to derefernce these arguments have to add NULL pointer checks
> >> to avoid crashing the kernel.
> >>
> >> This can be optimized by adding dummy sg_offset pointer to these
> >> two APIs. When the sg_offset arguments are NULL, pass the pointer
> >> of dummy to drivers. Drivers can always get a valid pointer, so no
> >> need to add NULL pointer checks.
> >>
> >> Junxian Huang (2):
> >>   RDMA/core: Add dummy sg_offset pointer for ib_map_mr_sg() and
> >>     ib_map_mr_sg_pi()
> >>   RDMA: Delete NULL pointer checks for sg_offset in .map_mr_sg ops
> >>
> >>  drivers/infiniband/core/verbs.c         | 12 +++++++++---
> >>  drivers/infiniband/hw/mlx5/mr.c         | 18 ++++++------------
> >>  drivers/infiniband/sw/rdmavt/trace_mr.h |  2 +-
> >>  3 files changed, 16 insertions(+), 16 deletions(-)
> > 
> > So what does this change give us?
> > We have same functionality, same number of lines, same everything ...
> > 
> 
> Actually this is inspired by an hns bug. When ib_map_mr_sg() passes a NULL
> sg_offset pointer to hns_roce_map_mr_sg(), we dereference this pointer
> without a NULL check.
> 
> Of course we can fix it by adding NULL check in hns, but I think this
> patch may be a better solution since the sg_offset is guaranteed to be
> a valid pointer. This could benefit future drivers who also want to
> dereference sg_offset, they won't need to care about NULL checks.

Let's fix hns please. We are moving away from SG in RDMA.

> 
> Junxian
> 
> > Thanks
> > 
> >>
> >> --
> >> 2.33.0
> >>
> >>

Re: [PATCH for-next 0/2] Small optimization for ib_map_mr_sg() and ib_map_mr_sg_pi()

Posted by Junxian Huang 2 weeks, 3 days ago


On 2024/11/6 21:36, Leon Romanovsky wrote:
> On Wed, Nov 06, 2024 at 09:12:47PM +0800, Junxian Huang wrote:
>>
>>
>> On 2024/11/6 20:08, Leon Romanovsky wrote:
>>> On Tue, Nov 05, 2024 at 08:08:39PM +0800, Junxian Huang wrote:
>>>> ib_map_mr_sg() and ib_map_mr_sg_pi() allow ULPs to specify NULL as
>>>> the sg_offset/data_sg_offset/meta_sg_offset arguments. Drivers who
>>>> need to derefernce these arguments have to add NULL pointer checks
>>>> to avoid crashing the kernel.
>>>>
>>>> This can be optimized by adding dummy sg_offset pointer to these
>>>> two APIs. When the sg_offset arguments are NULL, pass the pointer
>>>> of dummy to drivers. Drivers can always get a valid pointer, so no
>>>> need to add NULL pointer checks.
>>>>
>>>> Junxian Huang (2):
>>>>   RDMA/core: Add dummy sg_offset pointer for ib_map_mr_sg() and
>>>>     ib_map_mr_sg_pi()
>>>>   RDMA: Delete NULL pointer checks for sg_offset in .map_mr_sg ops
>>>>
>>>>  drivers/infiniband/core/verbs.c         | 12 +++++++++---
>>>>  drivers/infiniband/hw/mlx5/mr.c         | 18 ++++++------------
>>>>  drivers/infiniband/sw/rdmavt/trace_mr.h |  2 +-
>>>>  3 files changed, 16 insertions(+), 16 deletions(-)
>>>
>>> So what does this change give us?
>>> We have same functionality, same number of lines, same everything ...
>>>
>>
>> Actually this is inspired by an hns bug. When ib_map_mr_sg() passes a NULL
>> sg_offset pointer to hns_roce_map_mr_sg(), we dereference this pointer
>> without a NULL check.
>>
>> Of course we can fix it by adding NULL check in hns, but I think this
>> patch may be a better solution since the sg_offset is guaranteed to be
>> a valid pointer. This could benefit future drivers who also want to
>> dereference sg_offset, they won't need to care about NULL checks.
> 
> Let's fix hns please. We are moving away from SG in RDMA.
> 

Sure, thanks

Junxian

>>
>> Junxian
>>
>>> Thanks
>>>
>>>>
>>>> --
>>>> 2.33.0
>>>>
>>>>