[PATCH 1/1] net/rds: suppress page allocation failure error in recv buffer refill

[PATCH 1/1] net/rds: suppress page allocation failure error in recv buffer refill

[Manjunath Patil]

RDS/IB tries to refill the recv buffer in softirq context using
GFP_NOWAIT flag. However alloc failure is handled by queueing a work to
refill the recv buffer with GFP_KERNEL flag. This means failure to
allocate with GFP_NOWAIT isn't fatal. Do not print the PAF warnings if
softirq context fails to refill the recv buffer, instead print a one
line warning once a day.

Signed-off-by: Manjunath Patil <manjunath.b.patil@oracle.com>
Reviewed-by: Aruna Ramakrishna <aruna.ramakrishna@oracle.com>
---
 net/rds/ib_recv.c | 16 +++++++++++++---
 1 file changed, 13 insertions(+), 3 deletions(-)

diff --git a/net/rds/ib_recv.c b/net/rds/ib_recv.c
index 694d411dc72f..38d2894f6bb2 100644
--- a/net/rds/ib_recv.c
+++ b/net/rds/ib_recv.c
@@ -310,8 +310,8 @@ static int rds_ib_recv_refill_one(struct rds_connection *conn,
 	struct rds_ib_connection *ic = conn->c_transport_data;
 	struct ib_sge *sge;
 	int ret = -ENOMEM;
-	gfp_t slab_mask = GFP_NOWAIT;
-	gfp_t page_mask = GFP_NOWAIT;
+	gfp_t slab_mask = gfp;
+	gfp_t page_mask = gfp;
 
 	if (gfp & __GFP_DIRECT_RECLAIM) {
 		slab_mask = GFP_KERNEL;
@@ -406,6 +406,16 @@ void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp)
 		recv = &ic->i_recvs[pos];
 		ret = rds_ib_recv_refill_one(conn, recv, gfp);
 		if (ret) {
+			static unsigned long warn_time;
+			/* warn max once per day. This should be enough to
+			 * warn users about low mem situation.
+			 */
+			if (printk_timed_ratelimit(&warn_time,
+						   24 * 60 * 60 * 1000))
+				pr_warn("RDS/IB: failed to refill recv buffer for <%pI6c,%pI6c,%d>, waking worker\n",
+					&conn->c_laddr, &conn->c_faddr,
+					conn->c_tos);
+
 			must_wake = true;
 			break;
 		}
@@ -1020,7 +1030,7 @@ void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic,
 		rds_ib_stats_inc(s_ib_rx_ring_empty);
 
 	if (rds_ib_ring_low(&ic->i_recv_ring)) {
-		rds_ib_recv_refill(conn, 0, GFP_NOWAIT);
+		rds_ib_recv_refill(conn, 0, GFP_NOWAIT | __GFP_NOWARN);
 		rds_ib_stats_inc(s_ib_rx_refill_from_cq);
 	}
 }
-- 
2.27.0.112.g101b320

Re: [PATCH 1/1] net/rds: suppress page allocation failure error in recv buffer refill

[santosh.shilimkar]

On 9/16/20 12:08 PM, Manjunath Patil wrote:
> RDS/IB tries to refill the recv buffer in softirq context using
> GFP_NOWAIT flag. However alloc failure is handled by queueing a work to
> refill the recv buffer with GFP_KERNEL flag. This means failure to
> allocate with GFP_NOWAIT isn't fatal. Do not print the PAF warnings if
> softirq context fails to refill the recv buffer, instead print a one
> line warning once a day.
> 
> Signed-off-by: Manjunath Patil <manjunath.b.patil@oracle.com>
> Reviewed-by: Aruna Ramakrishna <aruna.ramakrishna@oracle.com>
> ---
>   net/rds/ib_recv.c | 16 +++++++++++++---
>   1 file changed, 13 insertions(+), 3 deletions(-)
> 
> diff --git a/net/rds/ib_recv.c b/net/rds/ib_recv.c
> index 694d411dc72f..38d2894f6bb2 100644
> --- a/net/rds/ib_recv.c
> +++ b/net/rds/ib_recv.c
> @@ -310,8 +310,8 @@ static int rds_ib_recv_refill_one(struct rds_connection *conn,
>   	struct rds_ib_connection *ic = conn->c_transport_data;
>   	struct ib_sge *sge;
>   	int ret = -ENOMEM;
> -	gfp_t slab_mask = GFP_NOWAIT;
> -	gfp_t page_mask = GFP_NOWAIT;
> +	gfp_t slab_mask = gfp;
> +	gfp_t page_mask = gfp;
>   
>   	if (gfp & __GFP_DIRECT_RECLAIM) {
>   		slab_mask = GFP_KERNEL;
> @@ -406,6 +406,16 @@ void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp)
>   		recv = &ic->i_recvs[pos];
>   		ret = rds_ib_recv_refill_one(conn, recv, gfp);
>   		if (ret) {
> +			static unsigned long warn_time;
Comment should start on next line.
> +			/* warn max once per day. This should be enough to
> +			 * warn users about low mem situation.
> +			 */
> +			if (printk_timed_ratelimit(&warn_time,
> +						   24 * 60 * 60 * 1000))
> +				pr_warn("RDS/IB: failed to refill recv buffer for <%pI6c,%pI6c,%d>, waking worker\n",
> +					&conn->c_laddr, &conn->c_faddr,
> +					conn->c_tos);
Didn't notice this before.
Why not just use "pr_warn_ratelimited()" ?
> +
>   			must_wake = true;
>   			break;
>   		}
> @@ -1020,7 +1030,7 @@ void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic,
>   		rds_ib_stats_inc(s_ib_rx_ring_empty);
>   
>   	if (rds_ib_ring_low(&ic->i_recv_ring)) {
> -		rds_ib_recv_refill(conn, 0, GFP_NOWAIT);
> +		rds_ib_recv_refill(conn, 0, GFP_NOWAIT | __GFP_NOWARN);
>   		rds_ib_stats_inc(s_ib_rx_refill_from_cq);
>   	}
>   }
> 

Re: [PATCH 1/1] net/rds: suppress page allocation failure error in recv buffer refill

[Manjunath Patil]

Hi Santosh,

inline.
On 9/16/2020 12:27 PM, santosh.shilimkar@oracle.com wrote:
> On 9/16/20 12:08 PM, Manjunath Patil wrote:
>> RDS/IB tries to refill the recv buffer in softirq context using
>> GFP_NOWAIT flag. However alloc failure is handled by queueing a work to
>> refill the recv buffer with GFP_KERNEL flag. This means failure to
>> allocate with GFP_NOWAIT isn't fatal. Do not print the PAF warnings if
>> softirq context fails to refill the recv buffer, instead print a one
>> line warning once a day.
>>
>> Signed-off-by: Manjunath Patil <manjunath.b.patil@oracle.com>
>> Reviewed-by: Aruna Ramakrishna <aruna.ramakrishna@oracle.com>
>> ---
>>   net/rds/ib_recv.c | 16 +++++++++++++---
>>   1 file changed, 13 insertions(+), 3 deletions(-)
>>
>> diff --git a/net/rds/ib_recv.c b/net/rds/ib_recv.c
>> index 694d411dc72f..38d2894f6bb2 100644
>> --- a/net/rds/ib_recv.c
>> +++ b/net/rds/ib_recv.c
>> @@ -310,8 +310,8 @@ static int rds_ib_recv_refill_one(struct 
>> rds_connection *conn,
>>       struct rds_ib_connection *ic = conn->c_transport_data;
>>       struct ib_sge *sge;
>>       int ret = -ENOMEM;
>> -    gfp_t slab_mask = GFP_NOWAIT;
>> -    gfp_t page_mask = GFP_NOWAIT;
>> +    gfp_t slab_mask = gfp;
>> +    gfp_t page_mask = gfp;
>>         if (gfp & __GFP_DIRECT_RECLAIM) {
>>           slab_mask = GFP_KERNEL;
>> @@ -406,6 +406,16 @@ void rds_ib_recv_refill(struct rds_connection 
>> *conn, int prefill, gfp_t gfp)
>>           recv = &ic->i_recvs[pos];
>>           ret = rds_ib_recv_refill_one(conn, recv, gfp);
>>           if (ret) {
>> +            static unsigned long warn_time;
> Comment should start on next line.
I will add new line. checkpatch.pl didn't find it though.
>> +            /* warn max once per day. This should be enough to
>> +             * warn users about low mem situation.
>> +             */
>> +            if (printk_timed_ratelimit(&warn_time,
>> +                           24 * 60 * 60 * 1000))
>> +                pr_warn("RDS/IB: failed to refill recv buffer for 
>> <%pI6c,%pI6c,%d>, waking worker\n",
>> +                    &conn->c_laddr, &conn->c_faddr,
>> +                    conn->c_tos);
> Didn't notice this before.
> Why not just use "pr_warn_ratelimited()" ?
I think you meant, get rid of if clause and use "pr_warn_ratelimited()" 
instead.
That can still produce more than needed logs during low memory situation.

-Thanks,
Manjunath
>> +
>>               must_wake = true;
>>               break;
>>           }
>> @@ -1020,7 +1030,7 @@ void rds_ib_recv_cqe_handler(struct 
>> rds_ib_connection *ic,
>>           rds_ib_stats_inc(s_ib_rx_ring_empty);
>>         if (rds_ib_ring_low(&ic->i_recv_ring)) {
>> -        rds_ib_recv_refill(conn, 0, GFP_NOWAIT);
>> +        rds_ib_recv_refill(conn, 0, GFP_NOWAIT | __GFP_NOWARN);
>>           rds_ib_stats_inc(s_ib_rx_refill_from_cq);
>>       }
>>   }
>>

Re: [PATCH 1/1] net/rds: suppress page allocation failure error in recv buffer refill

[santosh.shilimkar]

On 9/16/20 2:15 PM, Manjunath Patil wrote:
> Hi Santosh,
> 
> inline.
> On 9/16/2020 12:27 PM, santosh.shilimkar@oracle.com wrote:
>> On 9/16/20 12:08 PM, Manjunath Patil wrote:
>>> RDS/IB tries to refill the recv buffer in softirq context using
>>> GFP_NOWAIT flag. However alloc failure is handled by queueing a work to
>>> refill the recv buffer with GFP_KERNEL flag. This means failure to
>>> allocate with GFP_NOWAIT isn't fatal. Do not print the PAF warnings if
>>> softirq context fails to refill the recv buffer, instead print a one
>>> line warning once a day.
>>>
>>> Signed-off-by: Manjunath Patil <manjunath.b.patil@oracle.com>
>>> Reviewed-by: Aruna Ramakrishna <aruna.ramakrishna@oracle.com>
>>> ---
>>>   net/rds/ib_recv.c | 16 +++++++++++++---
>>>   1 file changed, 13 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/net/rds/ib_recv.c b/net/rds/ib_recv.c
>>> index 694d411dc72f..38d2894f6bb2 100644
>>> --- a/net/rds/ib_recv.c
>>> +++ b/net/rds/ib_recv.c
>>> @@ -310,8 +310,8 @@ static int rds_ib_recv_refill_one(struct 
>>> rds_connection *conn,
>>>       struct rds_ib_connection *ic = conn->c_transport_data;
>>>       struct ib_sge *sge;
>>>       int ret = -ENOMEM;
>>> -    gfp_t slab_mask = GFP_NOWAIT;
>>> -    gfp_t page_mask = GFP_NOWAIT;
>>> +    gfp_t slab_mask = gfp;
>>> +    gfp_t page_mask = gfp;
>>>         if (gfp & __GFP_DIRECT_RECLAIM) {
>>>           slab_mask = GFP_KERNEL;
>>> @@ -406,6 +406,16 @@ void rds_ib_recv_refill(struct rds_connection 
>>> *conn, int prefill, gfp_t gfp)
>>>           recv = &ic->i_recvs[pos];
>>>           ret = rds_ib_recv_refill_one(conn, recv, gfp);
>>>           if (ret) {
>>> +            static unsigned long warn_time;
>> Comment should start on next line.
> I will add new line. checkpatch.pl didn't find it though.
>>> +            /* warn max once per day. This should be enough to
>>> +             * warn users about low mem situation.
>>> +             */
>>> +            if (printk_timed_ratelimit(&warn_time,
>>> +                           24 * 60 * 60 * 1000))
>>> +                pr_warn("RDS/IB: failed to refill recv buffer for 
>>> <%pI6c,%pI6c,%d>, waking worker\n",
>>> +                    &conn->c_laddr, &conn->c_faddr,
>>> +                    conn->c_tos);
>> Didn't notice this before.
>> Why not just use "pr_warn_ratelimited()" ?
> I think you meant, get rid of if clause and use "pr_warn_ratelimited()" 
> instead.
> That can still produce more than needed logs during low memory situation.
> 
Try it out. It will do the same job as what you are trying to do.

Re: [PATCH 1/1] net/rds: suppress page allocation failure error in recv buffer refill

[Manjunath Patil]

On 9/16/2020 2:25 PM, santosh.shilimkar@oracle.com wrote:
> On 9/16/20 2:15 PM, Manjunath Patil wrote:
>> Hi Santosh,
>>
>> inline.
>> On 9/16/2020 12:27 PM, santosh.shilimkar@oracle.com wrote:
>>> On 9/16/20 12:08 PM, Manjunath Patil wrote:
>>>> RDS/IB tries to refill the recv buffer in softirq context using
>>>> GFP_NOWAIT flag. However alloc failure is handled by queueing a 
>>>> work to
>>>> refill the recv buffer with GFP_KERNEL flag. This means failure to
>>>> allocate with GFP_NOWAIT isn't fatal. Do not print the PAF warnings if
>>>> softirq context fails to refill the recv buffer, instead print a one
>>>> line warning once a day.
>>>>
>>>> Signed-off-by: Manjunath Patil <manjunath.b.patil@oracle.com>
>>>> Reviewed-by: Aruna Ramakrishna <aruna.ramakrishna@oracle.com>
>>>> ---
>>>>   net/rds/ib_recv.c | 16 +++++++++++++---
>>>>   1 file changed, 13 insertions(+), 3 deletions(-)
>>>>
>>>> diff --git a/net/rds/ib_recv.c b/net/rds/ib_recv.c
>>>> index 694d411dc72f..38d2894f6bb2 100644
>>>> --- a/net/rds/ib_recv.c
>>>> +++ b/net/rds/ib_recv.c
>>>> @@ -310,8 +310,8 @@ static int rds_ib_recv_refill_one(struct 
>>>> rds_connection *conn,
>>>>       struct rds_ib_connection *ic = conn->c_transport_data;
>>>>       struct ib_sge *sge;
>>>>       int ret = -ENOMEM;
>>>> -    gfp_t slab_mask = GFP_NOWAIT;
>>>> -    gfp_t page_mask = GFP_NOWAIT;
>>>> +    gfp_t slab_mask = gfp;
>>>> +    gfp_t page_mask = gfp;
>>>>         if (gfp & __GFP_DIRECT_RECLAIM) {
>>>>           slab_mask = GFP_KERNEL;
>>>> @@ -406,6 +406,16 @@ void rds_ib_recv_refill(struct rds_connection 
>>>> *conn, int prefill, gfp_t gfp)
>>>>           recv = &ic->i_recvs[pos];
>>>>           ret = rds_ib_recv_refill_one(conn, recv, gfp);
>>>>           if (ret) {
>>>> +            static unsigned long warn_time;
>>> Comment should start on next line.
>> I will add new line. checkpatch.pl didn't find it though.
>>>> +            /* warn max once per day. This should be enough to
>>>> +             * warn users about low mem situation.
>>>> +             */
>>>> +            if (printk_timed_ratelimit(&warn_time,
>>>> +                           24 * 60 * 60 * 1000))
>>>> +                pr_warn("RDS/IB: failed to refill recv buffer for 
>>>> <%pI6c,%pI6c,%d>, waking worker\n",
>>>> +                    &conn->c_laddr, &conn->c_faddr,
>>>> +                    conn->c_tos);
>>> Didn't notice this before.
>>> Why not just use "pr_warn_ratelimited()" ?
>> I think you meant, get rid of if clause and use 
>> "pr_warn_ratelimited()" instead.
>> That can still produce more than needed logs during low memory 
>> situation.
>>
> Try it out. It will do the same job as what you are trying to do.
Sure. I will use it and see. I will submit next version after my testing.

-Thanks,
Manjunath

[PATCH 1/1] net/rds: suppress page allocation failure error in recv buffer refill

[Manjunath Patil]

RDS/IB tries to refill the recv buffer in softirq context using
GFP_NOWAIT flag. However alloc failure is handled by queueing a work to
refill the recv buffer with GFP_KERNEL flag. This means failure to
allocate with GFP_NOWAIT isn't fatal. Do not print the PAF warnings if
softirq context fails to refill the recv buffer, instead print rate
limited warnings.

Signed-off-by: Manjunath Patil <manjunath.b.patil@oracle.com>
Reviewed-by: Aruna Ramakrishna <aruna.ramakrishna@oracle.com>
---
 net/rds/ib_recv.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

diff --git a/net/rds/ib_recv.c b/net/rds/ib_recv.c
index 694d411dc72f..b4ed421acc7b 100644
--- a/net/rds/ib_recv.c
+++ b/net/rds/ib_recv.c
@@ -310,8 +310,8 @@ static int rds_ib_recv_refill_one(struct rds_connection *conn,
 	struct rds_ib_connection *ic = conn->c_transport_data;
 	struct ib_sge *sge;
 	int ret = -ENOMEM;
-	gfp_t slab_mask = GFP_NOWAIT;
-	gfp_t page_mask = GFP_NOWAIT;
+	gfp_t slab_mask = gfp;
+	gfp_t page_mask = gfp;
 
 	if (gfp & __GFP_DIRECT_RECLAIM) {
 		slab_mask = GFP_KERNEL;
@@ -406,6 +406,9 @@ void rds_ib_recv_refill(struct rds_connection *conn, int prefill, gfp_t gfp)
 		recv = &ic->i_recvs[pos];
 		ret = rds_ib_recv_refill_one(conn, recv, gfp);
 		if (ret) {
+			pr_warn_ratelimited("RDS/IB: failed to refill recv buffer for <%pI6c,%pI6c,%d>, waking worker\n",
+					    &conn->c_laddr, &conn->c_faddr,
+					    conn->c_tos);
 			must_wake = true;
 			break;
 		}
@@ -1020,7 +1023,7 @@ void rds_ib_recv_cqe_handler(struct rds_ib_connection *ic,
 		rds_ib_stats_inc(s_ib_rx_ring_empty);
 
 	if (rds_ib_ring_low(&ic->i_recv_ring)) {
-		rds_ib_recv_refill(conn, 0, GFP_NOWAIT);
+		rds_ib_recv_refill(conn, 0, GFP_NOWAIT | __GFP_NOWARN);
 		rds_ib_stats_inc(s_ib_rx_refill_from_cq);
 	}
 }
-- 
2.27.0.112.g101b320

Re: [PATCH 1/1] net/rds: suppress page allocation failure error in recv buffer refill

[santosh.shilimkar]

On 10/2/20 1:05 PM, Manjunath Patil wrote:
> RDS/IB tries to refill the recv buffer in softirq context using
> GFP_NOWAIT flag. However alloc failure is handled by queueing a work to
> refill the recv buffer with GFP_KERNEL flag. This means failure to
> allocate with GFP_NOWAIT isn't fatal. Do not print the PAF warnings if
> softirq context fails to refill the recv buffer, instead print rate
> limited warnings.
> 
> Signed-off-by: Manjunath Patil <manjunath.b.patil@oracle.com>
> Reviewed-by: Aruna Ramakrishna <aruna.ramakrishna@oracle.com>
> ---
Thanks for the updated version. Whenever you send updated patch,
you should add version so that it helps for archiving as well as
review.

Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>

Re: [PATCH 1/1] net/rds: suppress page allocation failure error in recv buffer refill

[Manjunath Patil]

Thanks for ack'ing.

Yeah, sorry about version. I had it in my mind to add it when I started, 
but forgot it at the last moment.

-Thanks,
Manjunath
On 10/2/2020 1:10 PM, santosh.shilimkar@oracle.com wrote:
> On 10/2/20 1:05 PM, Manjunath Patil wrote:
>> RDS/IB tries to refill the recv buffer in softirq context using
>> GFP_NOWAIT flag. However alloc failure is handled by queueing a work to
>> refill the recv buffer with GFP_KERNEL flag. This means failure to
>> allocate with GFP_NOWAIT isn't fatal. Do not print the PAF warnings if
>> softirq context fails to refill the recv buffer, instead print rate
>> limited warnings.
>>
>> Signed-off-by: Manjunath Patil <manjunath.b.patil@oracle.com>
>> Reviewed-by: Aruna Ramakrishna <aruna.ramakrishna@oracle.com>
>> ---
> Thanks for the updated version. Whenever you send updated patch,
> you should add version so that it helps for archiving as well as
> review.
>
> Acked-by: Santosh Shilimkar <santosh.shilimkar@oracle.com>

Re: [PATCH 1/1] net/rds: suppress page allocation failure error in recv buffer refill

[David Miller]

From: Manjunath Patil <manjunath.b.patil@oracle.com>
Date: Fri,  2 Oct 2020 13:05:45 -0700

> RDS/IB tries to refill the recv buffer in softirq context using
> GFP_NOWAIT flag. However alloc failure is handled by queueing a work to
> refill the recv buffer with GFP_KERNEL flag. This means failure to
> allocate with GFP_NOWAIT isn't fatal. Do not print the PAF warnings if
> softirq context fails to refill the recv buffer, instead print rate
> limited warnings.
> 
> Signed-off-by: Manjunath Patil <manjunath.b.patil@oracle.com>
> Reviewed-by: Aruna Ramakrishna <aruna.ramakrishna@oracle.com>

Honestly I don't think the subsystem should print any warning at all.

Either it's a softirq failure, and that's ok because you will push
the allocation to GFP_KERNEL via a work job.  Or it's a GFP_KERNEL
failure in non-softirq context and the kernel will print a warning
and a stack backtrace from the memory allocator.

Therefore, please remove all of the warnings in the rds code.

Thanks.

Re: [PATCH 1/1] net/rds: suppress page allocation failure error in recv buffer refill

[Manjunath Patil]

Thanks David for your feedback.

I will submit v3 of this patch removing the warning.

-Manjunath
On 10/3/2020 5:26 PM, David Miller wrote:
> From: Manjunath Patil <manjunath.b.patil@oracle.com>
> Date: Fri,  2 Oct 2020 13:05:45 -0700
>
>> RDS/IB tries to refill the recv buffer in softirq context using
>> GFP_NOWAIT flag. However alloc failure is handled by queueing a work to
>> refill the recv buffer with GFP_KERNEL flag. This means failure to
>> allocate with GFP_NOWAIT isn't fatal. Do not print the PAF warnings if
>> softirq context fails to refill the recv buffer, instead print rate
>> limited warnings.
>>
>> Signed-off-by: Manjunath Patil <manjunath.b.patil@oracle.com>
>> Reviewed-by: Aruna Ramakrishna <aruna.ramakrishna@oracle.com>
> Honestly I don't think the subsystem should print any warning at all.
>
> Either it's a softirq failure, and that's ok because you will push
> the allocation to GFP_KERNEL via a work job.  Or it's a GFP_KERNEL
> failure in non-softirq context and the kernel will print a warning
> and a stack backtrace from the memory allocator.
>
> Therefore, please remove all of the warnings in the rds code.
>
> Thanks.