[PATCH net] gro_cells: reduce number of synchronize_net() calls

[PATCH net] gro_cells: reduce number of synchronize_net() calls

[Eric Dumazet]

From: Eric Dumazet <edumazet@google.com>

After cited commit, gro_cells_destroy() became damn slow
on hosts with a lot of cores.

This is because we have one additional synchronize_net() per cpu as
stated in the changelog.

gro_cells_init() is setting NAPI_STATE_NO_BUSY_POLL, and this was enough
to not have one synchronize_net() call per netif_napi_del()

We can factorize all the synchronize_net() to a single one,
right before freeing per-cpu memory.

Fixes: 5198d545dba8 ("net: remove napi_hash_del() from driver-facing API")
Signed-off-by: Eric Dumazet <edumazet@google.com>
---
 net/core/gro_cells.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/net/core/gro_cells.c b/net/core/gro_cells.c
index e095fb871d9120787bfdf62149f4d82e0e3b0a51..6eb2e5ec2c5068e1d798557e55d084b785187a9b 100644
--- a/net/core/gro_cells.c
+++ b/net/core/gro_cells.c
@@ -99,9 +99,14 @@ void gro_cells_destroy(struct gro_cells *gcells)
 		struct gro_cell *cell = per_cpu_ptr(gcells->cells, i);
 
 		napi_disable(&cell->napi);
-		netif_napi_del(&cell->napi);
+		__netif_napi_del(&cell->napi);
 		__skb_queue_purge(&cell->napi_skbs);
 	}
+	/* This barrier is needed because netpoll could access dev->napi_list
+	 * under rcu protection.
+	 */
+	synchronize_net();
+
 	free_percpu(gcells->cells);
 	gcells->cells = NULL;
 }
-- 
2.29.2.454.gaff20da3a2-goog

Re: [PATCH net] gro_cells: reduce number of synchronize_net() calls

[Jakub Kicinski]

On Tue, 24 Nov 2020 12:38:22 -0800 Eric Dumazet wrote:
> From: Eric Dumazet <edumazet@google.com>
> 
> After cited commit, gro_cells_destroy() became damn slow
> on hosts with a lot of cores.
> 
> This is because we have one additional synchronize_net() per cpu as
> stated in the changelog.
> 
> gro_cells_init() is setting NAPI_STATE_NO_BUSY_POLL, and this was enough
> to not have one synchronize_net() call per netif_napi_del()
> 
> We can factorize all the synchronize_net() to a single one,
> right before freeing per-cpu memory.
> 
> Fixes: 5198d545dba8 ("net: remove napi_hash_del() from driver-facing API")
> Signed-off-by: Eric Dumazet <edumazet@google.com>

Applied, thank you!