Re: Re: [PATCH] RDMA/siw: Relax from kmap_atomic() use in TX path

["Bernard Metzler" <BMT@zurich.ibm.com>]

-----"Steve Wise" <larrystevenwise@gmail.com> wrote: -----

>To: "Bernard Metzler" <bmt@zurich.ibm.com>
>From: "Steve Wise" <larrystevenwise@gmail.com>
>Date: 09/07/2019 12:31AM
>Cc: linux-rdma@vger.kernel.org, dledford@redhat.com
>Subject: [EXTERNAL] Re: [PATCH] RDMA/siw: Relax from kmap_atomic()
>use in TX path
>
>Hey Bernard,  is the code below called directly from the driver’s
>post_send function?  If so it must use atomic...I think.
>

Hey Steve,

Only a kernel TX thread executes the TX path. It is not
executed directly from the drivers post_send() method,
which only queues work for TX processing and wakes up
the TX thread if it currently does not run.
The single exception is TX processing
out of a user space application context, if no TX
thread is currently running, but only for a small
burst. In any case, all TX processing is done out of
a non-atomic context.

Best,
Bernard.

>Stevo
>
>> On Sep 6, 2019, at 6:18 AM, Bernard Metzler <bmt@zurich.ibm.com>
>wrote:
>> 
>> Since the transmit path is never executed in an atomic
>> context, we do not need kmap_atomic() and can always
>> use less demanding kmap().
>> 
>> Signed-off-by: Bernard Metzler <bmt@zurich.ibm.com>
>> ---
>> drivers/infiniband/sw/siw/siw_qp_tx.c | 12 +++++-------
>> 1 file changed, 5 insertions(+), 7 deletions(-)
>> 
>> diff --git a/drivers/infiniband/sw/siw/siw_qp_tx.c
>b/drivers/infiniband/sw/siw/siw_qp_tx.c
>> index 8e72f955921d..5d97bba0ce6d 100644
>> --- a/drivers/infiniband/sw/siw/siw_qp_tx.c
>> +++ b/drivers/infiniband/sw/siw/siw_qp_tx.c
>> @@ -76,16 +76,15 @@ static int siw_try_1seg(struct siw_iwarp_tx
>*c_tx, void *paddr)
>>            if (unlikely(!p))
>>                return -EFAULT;
>> 
>> -            buffer = kmap_atomic(p);
>> +            buffer = kmap(p);
>> 
>>            if (likely(PAGE_SIZE - off >= bytes)) {
>>                memcpy(paddr, buffer + off, bytes);
>> -                kunmap_atomic(buffer);
>>            } else {
>>                unsigned long part = bytes - (PAGE_SIZE - off);
>> 
>>                memcpy(paddr, buffer + off, part);
>> -                kunmap_atomic(buffer);
>> +                kunmap(p);
>> 
>>                if (!mem->is_pbl)
>>                    p = siw_get_upage(mem->umem,
>> @@ -97,11 +96,10 @@ static int siw_try_1seg(struct siw_iwarp_tx
>*c_tx, void *paddr)
>>                if (unlikely(!p))
>>                    return -EFAULT;
>> 
>> -                buffer = kmap_atomic(p);
>> -                memcpy(paddr + part, buffer,
>> -                       bytes - part);
>> -                kunmap_atomic(buffer);
>> +                buffer = kmap(p);
>> +                memcpy(paddr + part, buffer, bytes - part);
>>            }
>> +            kunmap(p);
>>        }
>>    }
>>    return (int)bytes;
>> -- 
>> 2.17.2
>> 
>
>