diff mbox series

[net-next,13/15] vxlan: Do not treat dst cache initialization errors as fatal

Message ID 20250415121143.345227-14-idosch@nvidia.com (mailing list archive)
State New
Delegated to: Netdev Maintainers
Headers show
Series vxlan: Convert FDB table to rhashtable | expand

Checks

Context Check Description
netdev/series_format success Posting correctly formatted
netdev/tree_selection success Clearly marked for net-next
netdev/ynl success Generated files up to date; no warnings/errors; no diff in generated;
netdev/fixes_present success Fixes tag not required for -next series
netdev/header_inline success No static functions without inline keyword in header files
netdev/build_32bit success Errors and warnings before: 1 this patch: 1
netdev/build_tools success No tools touched, skip
netdev/cc_maintainers success CCed 8 of 8 maintainers
netdev/build_clang success Errors and warnings before: 2 this patch: 2
netdev/verify_signedoff success Signed-off-by tag matches author and committer
netdev/deprecated_api success None detected
netdev/check_selftest success No net selftest shell script
netdev/verify_fixes success No Fixes tag
netdev/build_allmodconfig_warn success Errors and warnings before: 3 this patch: 3
netdev/checkpatch success total: 0 errors, 0 warnings, 0 checks, 14 lines checked
netdev/build_clang_rust success No Rust files in patch. Skipping build
netdev/kdoc success Errors and warnings before: 0 this patch: 0
netdev/source_inline success Was 0 now: 0
netdev/contest warning net-next-2025-04-16--09-00 (tests: 795)

Commit Message

Ido Schimmel April 15, 2025, 12:11 p.m. UTC 
FDB entries are allocated in an atomic context as they can be added from
the data path when learning is enabled.

After converting the FDB hash table to rhashtable, the insertion rate
will be much higher (*) which will entail a much higher rate of per-CPU
allocations via dst_cache_init().

When adding a large number of entries (e.g., 256k) in a batch, a small
percentage (< 0.02%) of these per-CPU allocations will fail [1]. This
does not happen with the current code since the insertion rate is low
enough to give the per-CPU allocator a chance to asynchronously create
new chunks of per-CPU memory.

Given that:

a. Only a small percentage of these per-CPU allocations fail.

b. The scenario where this happens might not be the most realistic one.

c. The driver can work correctly without dst caches. The dst_cache_*()
APIs first check that the dst cache was properly initialized.

d. The dst caches are not always used (e.g., 'tos inherit').

It seems reasonable to not treat these allocation failures as fatal.

Therefore, do not bail when dst_cache_init() fails and suppress warnings
by specifying '__GFP_NOWARN'.

[1] percpu: allocation failed, size=40 align=8 atomic=1, atomic alloc failed, no space left

(*) 97% reduction in average latency of vxlan_fdb_update() when adding
256k entries in a batch.

Reviewed-by: Petr Machata <petrm@nvidia.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
---
 drivers/net/vxlan/vxlan_core.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)
diff mbox series

Patch

diff --git a/drivers/net/vxlan/vxlan_core.c b/drivers/net/vxlan/vxlan_core.c
index 2846c8c5234e..5c0752161529 100644
--- a/drivers/net/vxlan/vxlan_core.c
+++ b/drivers/net/vxlan/vxlan_core.c
@@ -619,10 +619,10 @@  static int vxlan_fdb_append(struct vxlan_fdb *f,
 	if (rd == NULL)
 		return -ENOMEM;
 
-	if (dst_cache_init(&rd->dst_cache, GFP_ATOMIC)) {
-		kfree(rd);
-		return -ENOMEM;
-	}
+	/* The driver can work correctly without a dst cache, so do not treat
+	 * dst cache initialization errors as fatal.
+	 */
+	dst_cache_init(&rd->dst_cache, GFP_ATOMIC | __GFP_NOWARN);
 
 	rd->remote_ip = *ip;
 	rd->remote_port = port;