Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Feedback-ID: idfb84289:Fastmail
From:   Andrew Jeffery <andrew@aj.id.au>
To:     openipmi-developer@lists.sourceforge.net
Cc:     minyard@acm.org, joel@jms.id.au,
        linux-arm-kernel@lists.infradead.org,
        linux-aspeed@lists.ozlabs.org, linux-kernel@vger.kernel.org,
        openbmc@lists.ozlabs.org
Subject: [PATCH] ipmi: kcs: Poll OBF briefly to reduce OBE latency
Date:   Sat, 13 Aug 2022 00:17:41 +0930
Message-Id: <20220812144741.240315-1-andrew@aj.id.au>
MIME-Version: 1.0
Content-Transfer-Encoding: 8bit
Precedence: bulk

The ASPEED KCS devices don't provide a BMC-side interrupt for the host
reading the output data register (ODR). The act of the host reading ODR
clears the output buffer full (OBF) flag in the status register (STR),
informing the BMC it can transmit a subsequent byte.

On the BMC side the KCS client must enable the OBE event *and* perform a
subsequent read of STR anyway to avoid races - the polling provides a
window for the host to read ODR if data was freshly written while
minimising BMC-side latency.

Signed-off-by: Andrew Jeffery <andrew@aj.id.au>
---
 drivers/char/ipmi/kcs_bmc_aspeed.c | 24 +++++++++++++++++++++---
 1 file changed, 21 insertions(+), 3 deletions(-)

diff --git a/drivers/char/ipmi/kcs_bmc_aspeed.c b/drivers/char/ipmi/kcs_bmc_aspeed.c
index cdc88cde1e9a..417e5a3ccfae 100644
--- a/drivers/char/ipmi/kcs_bmc_aspeed.c
+++ b/drivers/char/ipmi/kcs_bmc_aspeed.c
@@ -399,13 +399,31 @@ static void aspeed_kcs_check_obe(struct timer_list *timer)
 static void aspeed_kcs_irq_mask_update(struct kcs_bmc_device *kcs_bmc, u8 mask, u8 state)
 {
 	struct aspeed_kcs_bmc *priv = to_aspeed_kcs_bmc(kcs_bmc);
+	int rc;
+	u8 str;
 
 	/* We don't have an OBE IRQ, emulate it */
 	if (mask & KCS_BMC_EVENT_TYPE_OBE) {
-		if (KCS_BMC_EVENT_TYPE_OBE & state)
-			mod_timer(&priv->obe.timer, jiffies + OBE_POLL_PERIOD);
-		else
+		if (KCS_BMC_EVENT_TYPE_OBE & state) {
+			/*
+			 * Given we don't have an OBE IRQ, delay by polling briefly to see if we can
+			 * observe such an event before returning to the caller. This is not
+			 * incorrect because OBF may have already become clear before enabling the
+			 * IRQ if we had one, under which circumstance no event will be propagated
+			 * anyway.
+			 *
+			 * The onus is on the client to perform a race-free check that it hasn't
+			 * missed the event.
+			 */
+			rc = read_poll_timeout_atomic(aspeed_kcs_inb, str,
+						      !(str & KCS_BMC_STR_OBF), 1, 100, false,
+						      &priv->kcs_bmc, priv->kcs_bmc.ioreg.str);
+			/* Time for the slow path? */
+			if (rc == -ETIMEDOUT)
+				mod_timer(&priv->obe.timer, jiffies + OBE_POLL_PERIOD);
+		} else {
 			del_timer(&priv->obe.timer);
+		}
 	}
 
 	if (mask & KCS_BMC_EVENT_TYPE_IBF) {
-- 
2.34.1