Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[orchagent]: Added support of PFC WD for BFN platform #823

Merged
merged 5 commits into from
Jun 7, 2019
Merged
Show file tree
Hide file tree
Changes from 2 commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion orchagent/orchdaemon.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -217,7 +217,7 @@ bool OrchDaemon::init()
CFG_PFC_WD_TABLE_NAME
};

if (platform == MLNX_PLATFORM_SUBSTRING)
if ((platform == MLNX_PLATFORM_SUBSTRING) || (platform == BFN_PLATFORM_SUBSTRING))
{

static const vector<sai_port_stat_t> portStatIds =
Expand Down
117 changes: 64 additions & 53 deletions orchagent/pfc_detect_barefoot.lua
Original file line number Diff line number Diff line change
Expand Up @@ -21,70 +21,81 @@ for i = n, 1, -1 do
local is_deadlock = false
local pfc_wd_status = redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'PFC_WD_STATUS')
local pfc_wd_action = redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'PFC_WD_ACTION')
if pfc_wd_status == 'operational' or pfc_wd_action == 'alert' then
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is this file identical to Mellanox too? If yes, will the symlink work?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Since it was there before and it's related to a different platform It's better to leave it as is.

local detection_time = tonumber(redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'PFC_WD_DETECTION_TIME'))
local time_left = redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'PFC_WD_DETECTION_TIME_LEFT')
if not time_left then
time_left = detection_time
else
time_left = tonumber(time_left)
end

local queue_index = redis.call('HGET', 'COUNTERS_QUEUE_INDEX_MAP', KEYS[i])
local port_id = redis.call('HGET', 'COUNTERS_QUEUE_PORT_MAP', KEYS[i])
local pfc_rx_pkt_key = 'SAI_PORT_STAT_PFC_' .. queue_index .. '_RX_PKTS'
local pfc_on2off_key = 'SAI_PORT_STAT_PFC_' .. queue_index .. '_ON2OFF_RX_PKTS'

local big_red_switch_mode = redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'BIG_RED_SWITCH_MODE')
if not big_red_switch_mode and (pfc_wd_status == 'operational' or pfc_wd_action == 'alert') then
local detection_time = redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'PFC_WD_DETECTION_TIME')
if detection_time then
detection_time = tonumber(detection_time)
local time_left = redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'PFC_WD_DETECTION_TIME_LEFT')
if not time_left then
time_left = detection_time
else
time_left = tonumber(time_left)
end

-- Get all counters
local occupancy_bytes = tonumber(redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'SAI_QUEUE_STAT_CURR_OCCUPANCY_BYTES'))
local packets = tonumber(redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'SAI_QUEUE_STAT_PACKETS'))
local pfc_rx_packets = tonumber(redis.call('HGET', counters_table_name .. ':' .. port_id, pfc_rx_pkt_key))
local pfc_on2off = tonumber(redis.call('HGET', counters_table_name .. ':' .. port_id, pfc_on2off_key))
local queue_pause_status = redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'SAI_QUEUE_ATTR_PAUSE_STATUS')

local packets_last = redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'SAI_QUEUE_STAT_PACKETS_last')
local pfc_rx_packets_last = redis.call('HGET', counters_table_name .. ':' .. port_id, pfc_rx_pkt_key .. '_last')
local pfc_on2off_last = redis.call('HGET', counters_table_name .. ':' .. port_id, pfc_on2off_key .. '_last')
local queue_pause_status_last = redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'SAI_QUEUE_ATTR_PAUSE_STATUS_last')
local queue_index = redis.call('HGET', 'COUNTERS_QUEUE_INDEX_MAP', KEYS[i])
local port_id = redis.call('HGET', 'COUNTERS_QUEUE_PORT_MAP', KEYS[i])
local pfc_rx_pkt_key = 'SAI_PORT_STAT_PFC_' .. queue_index .. '_RX_PKTS'
local pfc_duration_key = 'SAI_PORT_STAT_PFC_' .. queue_index .. '_RX_PAUSE_DURATION'

-- DEBUG CODE START. Uncomment to enable
local debug_storm = redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'DEBUG_STORM')
-- DEBUG CODE END.
-- Get all counters
local occupancy_bytes = redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'SAI_QUEUE_STAT_CURR_OCCUPANCY_BYTES')
local packets = redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'SAI_QUEUE_STAT_PACKETS')
local pfc_rx_packets = redis.call('HGET', counters_table_name .. ':' .. port_id, pfc_rx_pkt_key)
local pfc_duration = redis.call('HGET', counters_table_name .. ':' .. port_id, pfc_duration_key)

-- If this is not a first run, then we have last values available
if packets_last and pfc_rx_packets_last and pfc_on2off_last and queue_pause_status_last then
packets_last = tonumber(packets_last)
pfc_rx_packets_last = tonumber(pfc_rx_packets_last)
pfc_on2off_last = tonumber(pfc_on2off_last)
if occupancy_bytes and packets and pfc_rx_packets and pfc_duration then
occupancy_bytes = tonumber(occupancy_bytes)
packets = tonumber(packets)
pfc_rx_packets = tonumber(pfc_rx_packets)
pfc_duration = tonumber(pfc_duration)

-- Check actual condition of queue being in PFC storm
if (occupancy_bytes > 0 and packets - packets_last == 0 and pfc_rx_packets - pfc_rx_packets_last > 0) or
local packets_last = redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'SAI_QUEUE_STAT_PACKETS_last')
local pfc_rx_packets_last = redis.call('HGET', counters_table_name .. ':' .. port_id, pfc_rx_pkt_key .. '_last')
local pfc_duration_last = redis.call('HGET', counters_table_name .. ':' .. port_id, pfc_duration_key .. '_last')
-- DEBUG CODE START. Uncomment to enable
(debug_storm == "enabled") or
local debug_storm = redis.call('HGET', counters_table_name .. ':' .. KEYS[i], 'DEBUG_STORM')
-- DEBUG CODE END.
(occupancy_bytes == 0 and pfc_rx_packets - pfc_rx_packets_last > 0 and pfc_on2off - pfc_on2off_last == 0 and queue_pause_status_last == 'true' and queue_pause_status == 'true') then
if time_left <= poll_time then
redis.call('PUBLISH', 'PFC_WD', '["' .. KEYS[i] .. '","storm"]')
is_deadlock = true
time_left = detection_time
else
time_left = time_left - poll_time

-- If this is not a first run, then we have last values available
if packets_last and pfc_rx_packets_last and pfc_duration_last then
packets_last = tonumber(packets_last)
pfc_rx_packets_last = tonumber(pfc_rx_packets_last)
pfc_duration_last = tonumber(pfc_duration_last)

-- Check actual condition of queue being in PFC storm
if (occupancy_bytes > 0 and packets - packets_last == 0 and pfc_rx_packets - pfc_rx_packets_last > 0) or
-- DEBUG CODE START. Uncomment to enable
(debug_storm == "enabled") or
-- DEBUG CODE END.
(occupancy_bytes == 0 and packets - packets_last == 0 and (pfc_duration - pfc_duration_last) > poll_time * 0.8) then
if time_left <= poll_time then
redis.call('HDEL', counters_table_name .. ':' .. port_id, pfc_rx_pkt_key .. '_last')
redis.call('HDEL', counters_table_name .. ':' .. port_id, pfc_duration_key .. '_last')
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

You do not need to learn hdel _RX_PKTS_last and _RX_PAUSE_DURATION_last on lines 74 & 75 from the Mellanox script. This is to solve the double storm signaling that was observed on Mellanox platforms. I assume barefoot platforms do not have this issue. #697

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@andriymoroz-mlnx for comments

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

this is not Mellanox issue. It is common for this approach. If Barefoot uses the same solution (approach, counters, etc) they can use the same script

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The issue should not be related to the counters set Mellanox chooses. Excluding the counter factors, this should also happen on brcm platforms. However, we do not observe the issue on brcm platforms so far.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think Broadcom uses different way to detect storm and thus do not have this issue

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Difference is only in the counter set. The idea behind is the same---no packets going out of the queue and the queue is paused.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This dual storm signaling should only happen if the drop action is not installed by the orchagent fast enough. pfcactionhandler initCounters() will flip the PFC_WD_STATUS from operational to stormed as part of its drop actions. If PFC_WD_STATUS is not operational, the detect script logic will exit early on line 26 without running the actual detect state machine. So there should be no occurrence of dual storm signaling at all. The possible cause I can think of is that the control-plane cpu is not fast enough to schedule running the orchagent drop actions within a polling interval of 200 ms https://github.com/Azure/sonic-swss/blob/f22fb80bdfa10ea38e718996235c99233e08c31a/orchagent/pfc_detect_barefoot.lua#L26

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Let's better keep it as is. In case race condition really happens it will be very hard to catch and fix.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I suggest we start from a clean implementation rather than patch it here and there. The double storming signaling can be captured by PFC watchdog test. This is how mlnx found the problem on their platforms. Function-wise, double signaling does not affect the proper detection and the proper restoration.

Last time, you said bfn still uses manual test for PFC watchdog validation. If you have proof that it does exist also on bfn, we can later add the patch back.

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@wendani the ct test passed for bfn with these changes. If we are talking about fast or not fast cpu I would say it hard to catch the race condition on all possible CPUs, as bfn sdk could run on different platform vendors and this value could differ. Could we leave this check to avoid further patching and would see in context this feature will evolve?

redis.call('PUBLISH', 'PFC_WD', '["' .. KEYS[i] .. '","storm"]')
is_deadlock = true
time_left = detection_time
else
time_left = time_left - poll_time
end
else
if pfc_wd_action == 'alert' and pfc_wd_status ~= 'operational' then
redis.call('PUBLISH', 'PFC_WD', '["' .. KEYS[i] .. '","restore"]')
wendani marked this conversation as resolved.
Show resolved Hide resolved
end
time_left = detection_time
end
end
else
if pfc_wd_action == 'alert' and pfc_wd_status ~= 'operational' then
redis.call('PUBLISH', 'PFC_WD', '["' .. KEYS[i] .. '","restore"]')

-- Save values for next run
redis.call('HSET', counters_table_name .. ':' .. KEYS[i], 'SAI_QUEUE_STAT_PACKETS_last', packets)
redis.call('HSET', counters_table_name .. ':' .. KEYS[i], 'PFC_WD_DETECTION_TIME_LEFT', time_left)
if is_deadlock == false then
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Drop the 'if is_deadlock == false' condition. #697

redis.call('HSET', counters_table_name .. ':' .. port_id, pfc_rx_pkt_key .. '_last', pfc_rx_packets)
redis.call('HSET', counters_table_name .. ':' .. port_id, pfc_duration_key .. '_last', pfc_duration)
end
time_left = detection_time
end
end

-- Save values for next run
redis.call('HSET', counters_table_name .. ':' .. KEYS[i], 'SAI_QUEUE_ATTR_PAUSE_STATUS_last', queue_pause_status)
redis.call('HSET', counters_table_name .. ':' .. KEYS[i], 'SAI_QUEUE_STAT_PACKETS_last', packets)
redis.call('HSET', counters_table_name .. ':' .. KEYS[i], 'PFC_WD_DETECTION_TIME_LEFT', time_left)
redis.call('HSET', counters_table_name .. ':' .. port_id, pfc_rx_pkt_key .. '_last', pfc_rx_packets)
redis.call('HSET', counters_table_name .. ':' .. port_id, pfc_on2off_key .. '_last', pfc_on2off)
end
end

Expand Down