eth: bnxt: take page size into account for page pool recycling rings
Jakub Kicinski
maintainerPATCH 1/1 eth: bnxt: take page size into account for page pool recycling rings
The Rx rings are filled with Rx buffers. Which are supposed to fit packet headers (or MTU if HW-GRO is disabled). The aggregation buffers are filled with "device pages". Adjust the sizes of the page pool recycling ring appropriately, based on ratio of the size of the buffer on given ring vs system page size. Otherwise on a system with 64kB pages we end up with >700MB of memory sitting in every single page pool cache.
Correct the size calculation for the head_pool. Since the buffers there are always small I'm pretty sure I meant to cap the size at 1k, rather than make it the lowest possible size. With 64k pages 1k cache with a 1k ring is 64x larger than we need.
---
CC: [email protected]
CC: [email protected]
--- drivers/net/ethernet/broadcom/bnxt/bnxt.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)
drivers/net/ethernet/broadcom/bnxt/bnxt.c +5 -3
@@ -3807,12 +3807,14 @@ static int bnxt_alloc_rx_page_pool(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, int numa_node) { +const unsigned int agg_size_fac =PAGE_SIZE /BNXT_RX_PAGE_SIZE ; +const unsigned int rx_size_fac =PAGE_SIZE /SZ_4K ;struct page_pool_params pp = {0 };struct page_pool *pool; - pp.pool_size = bp->rx_agg_ring_size; + pp.pool_size = bp->rx_agg_ring_size / agg_size_fac;if (BNXT_RX_PAGE_MODE (bp)) - pp.pool_size += bp->rx_ring_size; + pp.pool_size += bp->rx_ring_size / rx_size_fac; pp.nid = numa_node; pp.napi = &rxr->bnapi->napi; pp.netdev = bp->dev;
@@ -3830,7 +3832,7 @@ static int bnxt_alloc_rx_page_pool(struct bnxt *bp, rxr->need_head_pool =page_pool_is_unreadable (pool);if (bnxt_separate_head_pool (rxr)) { - pp.pool_size =max (bp->rx_ring_size,1024 ); + pp.pool_size =min (bp->rx_ring_size / rx_size_fac,1024 ); pp.flags =PP_FLAG_DMA_MAP |PP_FLAG_DMA_SYNC_DEV ; pool =page_pool_create (&pp);if (IS_ERR (pool))
Signature · 1 line
2.50.0
Michael Chan
maintainerQuoted 27 lines · Jakub Kicinski
The Rx rings are filled with Rx buffers. Which are supposed to fit packet headers (or MTU if HW-GRO is disabled). The aggregation buffers are filled with "device pages". Adjust the sizes of the page pool recycling ring appropriately, based on ratio of the size of the buffer on given ring vs system page size. Otherwise on a system with 64kB pages we end up with >700MB of memory sitting in every single page pool cache.
Correct the size calculation for the
head_pool. Since the buffers there are always small I'm pretty sure I meant to cap the size at 1k, rather than make it the lowest possible size. With 64k pages 1k cache with a 1k ring is 64x larger than we need.Signed-off-by Jakub Kicinski <[email protected]>---
CC: [email protected]
CC: [email protected]--- drivers/net/ethernet/broadcom/bnxt/bnxt.c | 8 +++++--- 1 file changed, 5 insertions(+), 3 deletions(-)drivers/net/ethernet/broadcom/bnxt/bnxt.c +0
-0
drivers/net/ethernet/broadcom/bnxt/bnxt.c static int bnxt_alloc_rx_page_pool(struct bnxt *bp, · line 3815
Quoted 10 lines
drivers/net/ethernet/broadcom/bnxt/bnxt.c +3
-1@@ -3807,12 +3807,14 @@ static int bnxt_alloc_rx_page_pool(struct bnxt *bp, struct bnxt_rx_ring_info *rxr, int numa_node) { +const unsigned int agg_size_fac =PAGE_SIZE /BNXT_RX_PAGE_SIZE ; +const unsigned int rx_size_fac =PAGE_SIZE /SZ_4K ;struct page_pool_params pp = {0 };struct page_pool *pool; - pp.pool_size = bp->rx_agg_ring_size; + pp.pool_size = bp->rx_agg_ring_size / agg_size_fac;
The bp->rx_agg_ring_size has already taken the system PAGE_SIZE into consideration to some extent in bnxt_set_ring_params(). The jumbo_factor and agg_factor will be smaller when PAGE_SIZE is larger. Will this overcompensate?
Jakub Kicinski
maintainerQuoted 6 lines · On Thu, 26 Jun 2025 14:52:17 -0700 Michael Chan wrote:
Quoted 6 lines
{
drivers/net/ethernet/broadcom/bnxt/bnxt.c +2
-0+const unsigned int agg_size_fac =PAGE_SIZE /BNXT_RX_PAGE_SIZE ; +const unsigned int rx_size_fac =PAGE_SIZE /SZ_4K ;struct page_pool_params pp = {0 };struct page_pool *pool;
Quoted 7 lines
drivers/net/ethernet/broadcom/bnxt/bnxt.c static int bnxt_alloc_rx_page_pool(struct bnxt *bp, · line 3815
Quoted 2 lines
drivers/net/ethernet/broadcom/bnxt/bnxt.c +1
-1- pp.pool_size = bp->rx_agg_ring_size; + pp.pool_size = bp->rx_agg_ring_size / agg_size_fac;The
bp->rx_agg_ring_sizehas already taken the systemPAGE_SIZEinto consideration to some extent inbnxt_set_ring_params(). Thejumbo_factorandagg_factorwill be smaller whenPAGE_SIZEis larger. Will this overcompensate?
My understanding is basically that bnxt_set_ring_params() operates on BNXT_RX_PAGE_SIZE so it takes care of 4k .. 32k range pretty well. But for 64k pages we will use 32k buffers, so 2 agg ring entries per system page. If our heuristic is that we want the same number of pages on the device ring as in the pp cache we should divide the cache size by two. Hope that makes sense.
My initial temptation was to say that agg ring is always shown to the user in 4kB units, regardless of system page. The driver would divide and multiply the parameter in the ethtool callbacks. Otherwise even with this patch existing configs for bnxt have to be adjusted based on system page size :( But I suspect you may have existing users on systems with 64kB pages, so this would be too risky? WDYT?
Quoted 16 lines · Jakub Kicinski
Quoted 6 lines · On Thu, 26 Jun 2025 14:52:17 -0700 Michael Chan wrote:
Quoted 6 lines
{
drivers/net/ethernet/broadcom/bnxt/bnxt.c +2
-0+const unsigned int agg_size_fac =PAGE_SIZE /BNXT_RX_PAGE_SIZE ; +const unsigned int rx_size_fac =PAGE_SIZE /SZ_4K ;struct page_pool_params pp = {0 };struct page_pool *pool;Quoted 7 lines
drivers/net/ethernet/broadcom/bnxt/bnxt.c static int bnxt_alloc_rx_page_pool(struct bnxt *bp, · line 3815
Quoted 2 lines
drivers/net/ethernet/broadcom/bnxt/bnxt.c +1
-1- pp.pool_size = bp->rx_agg_ring_size; + pp.pool_size = bp->rx_agg_ring_size / agg_size_fac;The
bp->rx_agg_ring_sizehas already taken the systemPAGE_SIZEinto consideration to some extent inbnxt_set_ring_params(). Thejumbo_factorandagg_factorwill be smaller whenPAGE_SIZEis larger. Will this overcompensate?
Quoted 6 lines
My understanding is basically that
bnxt_set_ring_params()operates onBNXT_RX_PAGE_SIZEso it takes care of 4k .. 32k range pretty well. But for 64k pages we will use 32k buffers, so 2 agg ring entries per system page. If our heuristic is that we want the same number of pages on the device ring as in the pp cache we should divide the cache size by two. Hope that makes sense.
Ah, got it. agg_size_fac will be at most 2 when PAGE_SIZE is 64K.
Quoted 7 lines
My initial temptation was to say that agg ring is always shown to the user in 4kB units, regardless of system page. The driver would divide and multiply the parameter in the ethtool callbacks. Otherwise even with this patch existing configs for bnxt have to be adjusted based on system page size :( But I suspect you may have existing users on systems with 64kB pages, so this would be too risky? WDYT?
Agreed. Too risky and too confusing. Thanks.
This patch was applied to netdev/net-next.git (main) by Jakub Kicinski <[email protected]>:
Hello:
This patch was applied to netdev/net-next.git (main) by Jakub Kicinski <[email protected]>:
Quoted 9 lines · On Thu, 26 Jun 2025 09:54:41 -0700 you wrote:
The Rx rings are filled with Rx buffers. Which are supposed to fit packet headers (or MTU if HW-GRO is disabled). The aggregation buffers are filled with "device pages". Adjust the sizes of the page pool recycling ring appropriately, based on ratio of the size of the buffer on given ring vs system page size. Otherwise on a system with 64kB pages we end up with >700MB of memory sitting in every single page pool cache.
[...]
Here is the summary with links:
- [net-next] eth: bnxt: take page size into account for page pool recycling rings
https://git.kernel.org/netdev/net-next/c/f7dbedba6312
You are awesome, thank you!
Signature · 2 lines
Deet-doot-dot, I am a bot. https://korg.docs.kernel.org/patchwork/pwbot.html