Thread (15 messages) 15 messages, 4 authors, 2024-05-02

Re: [PATCH net-next v2 5/6] bnxt_en: Optimize recovery path ULP locking in the driver

From: Simon Horman <horms@kernel.org>
Date: 2024-05-02 10:07:24

On Tue, Apr 30, 2024 at 05:30:55PM -0700, Michael Chan wrote:
From: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>

In the error recovery path (AER, firmware recovery, etc), the
driver notifies the RoCE driver via ULP_STOP before the reset
and via ULP_START after the reset, all under RTNL_LOCK.  The
RoCE driver can take a long time if there are a lot of QPs to
destroy, so it is not ideal to hold the global RTNL lock.

Rely on the new en_dev_lock mutex instead for ULP_STOP and
ULP_START.  For the most part, we move the ULP_STOP call before
we take the RTNL lock and move the ULP_START after RTNL unlock.
Note that SRIOV re-enablement must be done after ULP_START
or RoCE on the VFs will not resume properly after reset.

The one scenario in bnxt_hwrm_if_change() where the RTNL lock
is already taken in the .ndo_open() context requires the ULP
restart to be deferred to the bnxt_sp_task() workqueue.

Reviewed-by: Selvin Thyparampil Xavier <selvin.xavier@broadcom.com>
Reviewed-by: Vikas Gupta <vikas.gupta@broadcom.com>
Reviewed-by: Pavan Chebbi <pavan.chebbi@broadcom.com>
Signed-off-by: Kalesh AP <kalesh-anakkur.purayil@broadcom.com>
Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Reviewed-by: Simon Horman <horms@kernel.org>
quoted hunk ↗ jump to hunk
diff --git a/drivers/net/ethernet/broadcom/bnxt/bnxt_devlink.c b/drivers/net/ethernet/broadcom/bnxt/bnxt_devlink.c
index d9ea6fa23923..4cb0fabf977e 100644
--- a/drivers/net/ethernet/broadcom/bnxt/bnxt_devlink.c
+++ b/drivers/net/ethernet/broadcom/bnxt/bnxt_devlink.c
@@ -437,18 +437,20 @@ static int bnxt_dl_reload_down(struct devlink *dl, bool netns_change,
 
 	switch (action) {
 	case DEVLINK_RELOAD_ACTION_DRIVER_REINIT: {
+		bnxt_ulp_stop(bp);
 		rtnl_lock();
 		if (bnxt_sriov_cfg(bp)) {
 			NL_SET_ERR_MSG_MOD(extack,
 					   "reload is unsupported while VFs are allocated or being configured");
 			rtnl_unlock();
+			bnxt_ulp_start(bp, 0);
 			return -EOPNOTSUPP;
 		}
 		if (bp->dev->reg_state == NETREG_UNREGISTERED) {
 			rtnl_unlock();
+			bnxt_ulp_start(bp, 0);
 			return -ENODEV;
Hi Selvin, Michael, all,

FWIIW, I would have used a goto to unwind this and the previous error.
No need to need to respin because of this.
quoted hunk ↗ jump to hunk
 		}
-		bnxt_ulp_stop(bp);
 		if (netif_running(bp->dev))
 			bnxt_close_nic(bp, true, true);
 		bnxt_vf_reps_free(bp);
@@ -516,7 +518,6 @@ static int bnxt_dl_reload_up(struct devlink *dl, enum devlink_reload_action acti
 		bnxt_vf_reps_alloc(bp);
 		if (netif_running(bp->dev))
 			rc = bnxt_open_nic(bp, true, true);
-		bnxt_ulp_start(bp, rc);
 		if (!rc) {
 			bnxt_reenable_sriov(bp);
 			bnxt_ptp_reapply_pps(bp);
@@ -570,6 +571,8 @@ static int bnxt_dl_reload_up(struct devlink *dl, enum devlink_reload_action acti
 		dev_close(bp->dev);
 	}
 	rtnl_unlock();
+	if (action == DEVLINK_RELOAD_ACTION_DRIVER_REINIT)
+		bnxt_ulp_start(bp, rc);
 	return rc;
 }
  
Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help