Re: [PATCH v2] crypto: powerpc/p10-aes-gcm - Fix IPsec stall on bond failover
From: Danny Tsen
Date: Fri Nov 20 2026 - 15:40:59 EST
On 11/20/26 2:14 PM, Breno Leitao wrote:
On Fri, Nov 20, 2026 at 01:16:57PM -0500, Danny Tsen wrote:
The Power10 AES-GCM driver registers its gcm(aes) andReviewed-by: Breno Leitao <leitao@xxxxxxxxxx>
rfc4106(gcm(aes)) algorithms through the crypto SIMD helper. When
crypto_simd_usable() is false, for example in softirq context, the
helper defers requests to cryptd and completes them asynchronously
with -EINPROGRESS.
This breaks IPsec traffic during a bond failover. With the crypto
offload device active, traffic is switched to a passive device that
has only software encryption, and the ESP requests then go through
aes_gcm_p10_crypto. Under heavy load, some or all of the TCP streams
drop to zero throughput and do not recover. The stack is seen to drop
the ACKs as invalid, which points to the cryptd queueing delaying or
reordering the packets.
Fix this by registering the algorithms directly as aead_alg with
CRYPTO_ALG_NEED_FALLBACK, and allocating a "gcm(aes)" or
"rfc4106(gcm(aes))" fallback tfm in ->init(). When VSX cannot be used
(!crypto_simd_usable()), the key was not set up for SIMD, or
cryptlen < 64, the request is handed to the fallback and completed
synchronously in the caller's context. The driver no longer defers
work to cryptd.
Remove Kconfig option CRYPTO_SIMD and add CRYPTO_AES and CRYPTO_GCM.
Fixes: c954b252dee9 ("crypto: powerpc/p10-aes-gcm - Register modules as SIMD")
Signed-off-by: Danny Tsen <dtsen@xxxxxxxxxxxxx>
- * Copyright 2022- IBM Inc. All rights reservedDo we really want to change this in this "fix" commit?
+ * Copyright IBM Corp. 2022, 2023, 2024, 2025, 2026
This is following IBM's copyright style.
Thanks.
-Danny