From e4b433f4a74196476ccf226e450c4582428641c1 Mon Sep 17 00:00:00 2001 From: Jacob Keller Date: Thu, 13 Apr 2017 04:45:52 -0400 Subject: i40e: reset all VFs in parallel when rebuilding PF When there are a lot of active VFs, it can take multiple seconds to finish resetting all of them during certain flows., which can cause some VFs to fail to wait long enough for the reset to occur. The user might see messages like "Never saw reset" or "Reset never finished" and the VF driver will stop functioning properly. The naive solution would be to simply increase the wait timer. We can get much more clever. Notice that i40e_reset_vf is run in a serialized fashion, and includes lots of delays. There are two prominent delays which take most of the time. First, when we begin resetting VFs, we have multiple 10ms delays which accrue because we reset each VF in a serial fashion. These delays accumulate to almost 4 seconds when handling the maximum number of VFs (128). Secondly, there is a massive 50ms delay for each time we disable queues on a VSI. This delay is necessary to allow HW to finish disabling queues before we restore functionality. However, just like with the first case, we are paying the cost for each VF, rather than disabling all VFs and waiting once. Both of these can be fixed, but required some previous refactoring to handle the special case. First, we will need the i40e_vsi_wait_queues_disabled function which was previously DCB specific. Second, we will need to implement our own i40e_vsi_stop_rings_no_wait function which will handle the stopping of rings without the delays. Finally, implement an i40e_reset_all_vfs function, which will first start the reset of all VFs, and pay the wait cost all at once, rather than serially waiting for each VF before we start processing then next one. After the VF has been reset, we'll disable all the VF queues, and then wait for them to disable. Again, we'll organize the flow such that we pay the wait cost only once. Finally, after we've disabled queues we'll go ahead and begin restoring VF functionality. The result is reducing the wait time by a large factor and ensuring that VFs do not timeout when waiting in the VF driver. Change-ID: Ia6e8cf8d98131b78aec89db78afb8d905c9b12be Signed-off-by: Jacob Keller Tested-by: Andrew Bowers Signed-off-by: Jeff Kirsher --- drivers/net/ethernet/intel/i40e/i40e.h | 2 ++ 1 file changed, 2 insertions(+) (limited to 'drivers/net/ethernet/intel/i40e/i40e.h') diff --git a/drivers/net/ethernet/intel/i40e/i40e.h b/drivers/net/ethernet/intel/i40e/i40e.h index b629f3a..dfab7b0 100644 --- a/drivers/net/ethernet/intel/i40e/i40e.h +++ b/drivers/net/ethernet/intel/i40e/i40e.h @@ -868,6 +868,8 @@ void i40e_notify_client_of_vf_msg(struct i40e_vsi *vsi, u32 vf_id, int i40e_vsi_start_rings(struct i40e_vsi *vsi); void i40e_vsi_stop_rings(struct i40e_vsi *vsi); +void i40e_vsi_stop_rings_no_wait(struct i40e_vsi *vsi); +int i40e_vsi_wait_queues_disabled(struct i40e_vsi *vsi); int i40e_reconfig_rss_queues(struct i40e_pf *pf, int queue_count); struct i40e_veb *i40e_veb_setup(struct i40e_pf *pf, u16 flags, u16 uplink_seid, u16 downlink_seid, u8 enabled_tc); -- cgit v1.1