octeontx2 driver calls page_pool_create() during driver probe()
and fails if queue size > 32k. Page pool infra uses these buffers
as shock absorbers for burst traffic. These pages are pinned down
over time as working sets varies, due to the recycling nature
of page pool, given page pool (currently) don't have a shrinker
mechanism, the pages remain pinned down in ptr_ring.
Instead of clamping page_pool size to 32k at
most, limit it even more to 2k to avoid wasting memory.

This have been tested on octeontx2 CN10KA hardware.
TCP and UDP tests using iperf shows no performance regressions.

Fixes: b2e3406a38f0 ("octeontx2-pf: Add support for page pool")
Suggested-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Reviewed-by: Sunil Goutham <sgoutham@marvell.com>
Signed-off-by: Ratheesh Kannoth <rkannoth@marvell.com>
---

ChangeLogs:

v1->v2: Commit message changes and typo fixes
v0->v1: Commit message changes.
---
 drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c | 2 +-
 drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h | 2 ++
 2 files changed, 3 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
index 77c8f650f7ac..3e1c70c74622 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
@@ -1432,7 +1432,7 @@ int otx2_pool_init(struct otx2_nic *pfvf, u16 pool_id,
 	}
 
 	pp_params.flags = PP_FLAG_PAGE_FRAG | PP_FLAG_DMA_MAP;
-	pp_params.pool_size = numptrs;
+	pp_params.pool_size = min(OTX2_PAGE_POOL_SZ, numptrs);
 	pp_params.nid = NUMA_NO_NODE;
 	pp_params.dev = pfvf->dev;
 	pp_params.dma_dir = DMA_FROM_DEVICE;
diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
index ba8091131ec0..f6fea43617ff 100644
--- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
+++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
@@ -30,6 +30,8 @@
 #include <rvu_trace.h>
 #include "qos.h"
 
+#define OTX2_PAGE_POOL_SZ 2048
+
 /* IPv4 flag more fragment bit */
 #define IPV4_FLAG_MORE				0x20
 
-- 
2.25.1

From: Ratheesh Kannoth <rkannoth@marvell.com>
Date: Wed, 23 Aug 2023 08:23:25 +0530

> octeontx2 driver calls page_pool_create() during driver probe()
> and fails if queue size > 32k. Page pool infra uses these buffers
> as shock absorbers for burst traffic. These pages are pinned down
> over time as working sets varies, due to the recycling nature
> of page pool, given page pool (currently) don't have a shrinker
> mechanism, the pages remain pinned down in ptr_ring.
> Instead of clamping page_pool size to 32k at
> most, limit it even more to 2k to avoid wasting memory.
> 
> This have been tested on octeontx2 CN10KA hardware.
> TCP and UDP tests using iperf shows no performance regressions.

See now? I told ya you don't need that many to provide good recycling :>
But nice to see it was confirmed 2k is enough.

> 
> Fixes: b2e3406a38f0 ("octeontx2-pf: Add support for page pool")
> Suggested-by: Alexander Lobakin <aleksander.lobakin@intel.com>
> Reviewed-by: Sunil Goutham <sgoutham@marvell.com>
> Signed-off-by: Ratheesh Kannoth <rkannoth@marvell.com>
> ---
> 
> ChangeLogs:
> 
> v1->v2: Commit message changes and typo fixes
> v0->v1: Commit message changes.
> ---
>  drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c | 2 +-
>  drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h | 2 ++
>  2 files changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
> index 77c8f650f7ac..3e1c70c74622 100644
> --- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
> +++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.c
> @@ -1432,7 +1432,7 @@ int otx2_pool_init(struct otx2_nic *pfvf, u16 pool_id,
>  	}
>  
>  	pp_params.flags = PP_FLAG_PAGE_FRAG | PP_FLAG_DMA_MAP;
> -	pp_params.pool_size = numptrs;
> +	pp_params.pool_size = min(OTX2_PAGE_POOL_SZ, numptrs);
>  	pp_params.nid = NUMA_NO_NODE;
>  	pp_params.dev = pfvf->dev;
>  	pp_params.dma_dir = DMA_FROM_DEVICE;
> diff --git a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
> index ba8091131ec0..f6fea43617ff 100644
> --- a/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h
> +++ b/drivers/net/ethernet/marvell/octeontx2/nic/otx2_common.h

Isn't otx2_txrx.h a better place for this definition? For example, I'd
place it somewhere between CGX_CHAN_BASE and OTX2_DATA_ALIGN. Or after
MAX_XDP_MTU, doesn't matter much.

> @@ -30,6 +30,8 @@
>  #include <rvu_trace.h>
>  #include "qos.h"
>  
> +#define OTX2_PAGE_POOL_SZ 2048

Don't forget about alignment. Otx2 code has very inconsistent alignment
of definitions -- just one space in some places and proper tab alignment
in others. I still believe we prefer the second for netdev code.

> +
>  /* IPv4 flag more fragment bit */
>  #define IPV4_FLAG_MORE				0x20
>  

Thanks,
Olek

> From: Alexander Lobakin <aleksander.lobakin@intel.com>
> Sent: Wednesday, August 23, 2023 3:43 PM
> Subject: [EXT] Re: [PATCH v2 net] octeontx2-pf: fix page_pool creation fail
> for rings > 32k
> See now? I told ya you don't need that many to provide good recycling :> But
> nice to see it was confirmed 2k is enough.
Thanks !
 
> Isn't otx2_txrx.h a better place for this definition? For example, I'd place it
> somewhere between CGX_CHAN_BASE and OTX2_DATA_ALIGN. Or after
> MAX_XDP_MTU, doesn't matter much.
ACK.

> Don't forget about alignment. Otx2 code has very inconsistent alignment of
> definitions -- just one space in some places and proper tab alignment in
> others. I still believe we prefer the second for netdev code.
ACK.  Is it documented somewhere. I mean the coding std for netdev ?

On 23/08/2023 04.53, Ratheesh Kannoth wrote:
> octeontx2 driver calls page_pool_create() during driver probe()
> and fails if queue size > 32k. Page pool infra uses these buffers
> as shock absorbers for burst traffic. These pages are pinned down
> over time as working sets varies, due to the recycling nature
> of page pool, given page pool (currently) don't have a shrinker
> mechanism, the pages remain pinned down in ptr_ring.
> Instead of clamping page_pool size to 32k at
> most, limit it even more to 2k to avoid wasting memory.
> 
> This have been tested on octeontx2 CN10KA hardware.
> TCP and UDP tests using iperf shows no performance regressions.
> 
> Fixes: b2e3406a38f0 ("octeontx2-pf: Add support for page pool")
> Suggested-by: Alexander Lobakin<aleksander.lobakin@intel.com>
> Reviewed-by: Sunil Goutham<sgoutham@marvell.com>
> Signed-off-by: Ratheesh Kannoth<rkannoth@marvell.com>

LGTM

Acked-by: Jesper Dangaard Brouer <hawk@kernel.org>