netdev - [PATCH v2] i40e: avoid spin loop in i40e_asq_send

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [day] [month] [year] [list]

Message-Id: <20211101173808.1735144-1-csander@purestorage.com>
Date:   Mon,  1 Nov 2021 11:38:08 -0600
From:   Caleb Sander <csander@...estorage.com>
To:     intel-wired-lan@...ts.osuosl.org
Cc:     Tony Nguyen <anthony.l.nguyen@...el.com>,
        Caleb Sander <csander@...estorage.com>,
        Joern Engel <joern@...estorage.com>,
        Tony Brelinski <tony.brelinski@...el.com>, davem@...emloft.net,
        Jakub Kicinski <kuba@...nel.org>, netdev@...r.kernel.org,
        sassmann@...hat.com
Subject: [PATCH v2] i40e: avoid spin loop in i40e_asq_send_command()

Previously, the kernel could spend up to 250 ms waiting for a command to
be submitted to an admin queue. This function is also called in a loop,
e.g., in i40e_get_module_eeprom() (through i40e_aq_get_phy_register()),
so the time spent in the kernel may be even higher. We observed
scheduling delays of over 2 seconds in production,
with stacktraces pointing to this code as the culprit.

Use usleep_range() instead of udelay() so the loop can yield the CPU.
Also compute the elapsed time using the jiffies counter rather than
assuming udelay() waits exactly the time interval requested.

Signed-off-by: Caleb Sander <csander@...estorage.com>
Reviewed-by: Joern Engel <joern@...estorage.com>
---
 drivers/net/ethernet/intel/i40e/i40e_adminq.c | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

Changed from v1:
Use usleep_range() instead of udelay() + cond_resched(),
to avoid using the CPU while waiting.
Use 50 us as the max for the range since hrtimers schedules the sleep
for the max (unless another timer interrupt occurs after the min).
Since checking if the command is done too frequently would waste time
context-switching, use half of the max (25 us) as the min for the range.

diff --git a/drivers/net/ethernet/intel/i40e/i40e_adminq.c b/drivers/net/ethernet/intel/i40e/i40e_adminq.c
index 593912b17..b2c27ab3b 100644
--- a/drivers/net/ethernet/intel/i40e/i40e_adminq.c
+++ b/drivers/net/ethernet/intel/i40e/i40e_adminq.c
@@ -902,7 +902,7 @@ i40e_status i40e_asq_send_command(struct i40e_hw *hw,
 	 * we need to wait for desc write back
 	 */
 	if (!details->async && !details->postpone) {
-		u32 total_delay = 0;
+		unsigned long timeout_end = jiffies + usecs_to_jiffies(hw->aq.asq_cmd_timeout);

 		do {
 			/* AQ designers suggest use of head for better
@@ -910,9 +910,8 @@ i40e_status i40e_asq_send_command(struct i40e_hw *hw,
 			 */
 			if (i40e_asq_done(hw))
 				break;
-			udelay(50);
-			total_delay += 50;
-		} while (total_delay < hw->aq.asq_cmd_timeout);
+			usleep_range(25, 50);
+		} while (time_before(jiffies, timeout_end));
 	}

 	/* if ready, copy the desc back to temp */
-- 
2.25.1