It has been observed that in zero-loss benchmarks, when a slow traffic rate is being tested, the IRQ timer coalescing parameter was set too high, and the ethernet controller would start dropping packets because the job ring back half wouldn't be executed in time before the ethernet controller would fill its buffers, thereby significantly reducing the zero-loss performance figures. Empirical testing has shown that the best zero-loss performance is achieved when IRQ coalescing is set to minimum values and/or turned off, since apparently the job ring driver already implements an adequately-performing general-purpose IRQ mitigation strategy in software. Whilst we could go with minimal count (2-8) and timing settings (192-256), we prefer just turning h/w coalescing altogether off to minimize setkey latency (due to split key generation), and for consistent cross-SoC performance (the SEC vs. core clock ratio changes). Signed-off-by: Kim Phillips <kim.phillips@xxxxxxxxxxxxx> --- drivers/crypto/caam/Kconfig | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/drivers/crypto/caam/Kconfig b/drivers/crypto/caam/Kconfig index 502a7f3..65c7668 100644 --- a/drivers/crypto/caam/Kconfig +++ b/drivers/crypto/caam/Kconfig @@ -32,10 +32,13 @@ config CRYPTO_DEV_FSL_CAAM_RINGSIZE config CRYPTO_DEV_FSL_CAAM_INTC bool "Job Ring interrupt coalescing" depends on CRYPTO_DEV_FSL_CAAM - default y + default n help Enable the Job Ring's interrupt coalescing feature. + Note: the driver already provides adequate + interrupt coalescing in software. + config CRYPTO_DEV_FSL_CAAM_INTC_COUNT_THLD int "Job Ring interrupt coalescing count threshold" depends on CRYPTO_DEV_FSL_CAAM_INTC -- 1.7.11.1 -- To unsubscribe from this list: send the line "unsubscribe linux-crypto" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
