frames followed by these errors in log. [sdp] Result: hostbyte=DID_OK driverbyte=DRIVER_SENSE,SUGGEST_OK [sdp] Sense Key : Aborted Command [current] [sdp] Add. Sense: Data phase error This was causing some test apps to exit due to write failure under heavy load. This was due to a race around adding and removing tx frame skb in fcoe_pending_queue, Chris Leech helped me to find that brief unlocking period when pulling skb from fcoe_pending_queue in various contexts (fcoe_watchdog and fcoe_xmit) and then adding skb back into fcoe_pending_queue up on a failed fcoe_start_io could change skb/tx frame order in fcoe_pending_queue. Thanks Chris. This patch allows only single context to pull skb from fcoe_pending_queue at any time to prevent above described ordering issue/race by use of fcoe_pending_queue_active flag. This patch simplified fcoe_watchdog with modified fcoe_check_wait_queue by use of FCOE_LOW_QUEUE_DEPTH instead previously used several conditionals to clear and set lp->qfull. I think FCOE_MAX_QUEUE_DEPTH with FCOE_LOW_QUEUE_DEPTH will work better in re/setting lp->qfull and these could be fine tuned for performance.
Showing the changes to include (1 of 3 files). Show all
Unified · Split
| @@ -46,6 +46,7 @@ struct fcoe_softc { · 45 unmodified lines | ||
| 46 | 46 | struct net_device *phys_dev; /* device with ethtool_ops */ |
| 47 | 47 | struct packet_type fcoe_packet_type; |
| 48 | 48 | struct sk_buff_head fcoe_pending_queue; |
| 49 | + u8 fcoe_pending_queue_active; | |
| 49 | 50 | |
| 50 | 51 | u8 dest_addr[ETH_ALEN]; |
| 51 | 52 | u8 ctl_src_addr[ETH_ALEN]; |