diff mbox

[net-next] gianfar: Simplify MQ polling to avoid soft lockup

Message ID 20140327125303.GA22117@breakpoint.cc
State RFC, archived
Delegated to: David Miller
Headers show

Commit Message

Sebastian Andrzej Siewior March 27, 2014, 12:53 p.m. UTC 
On 2013-10-14 18:11:15 [+0300], Claudiu Manoil wrote:
> >>BUG: soft lockup - CPU#0 stuck for 23s! [iperf:2847]
> >>NIP [c0255b6c] find_next_bit+0xb8/0xc4
> >>LR [c0367ae8] gfar_poll+0xc8/0x1d8
> >It seems there is a race condition, and this patch only makes it happen
> >less often ?
> >
> >return faster means what exactly ?
> >
> 
> Hi Eric,
> Because of the outer while loop, gfar_poll may not return due
> to continuous tx work. The later implementation of gfar_poll
> allows only one iteration of the Tx queues before returning
> control to net_rx_action(), that's what I meant with "returns faster".

We talk here about 23secs of cleanup. RX is limited by NAPI and TX is
limited because it can't be refilled on your UP system.
Does your box recover from this condition without this patch? Mine does
not. But I run -RT and stumbled uppon something different.

What I observe is that the TX queue is not empty but does not make any
progress. That means tx_queue->tx_skbuff[tx_queue->skb_dirtytx] is true
and gfar_clean_tx_ring() cleans up zero packages because it is not yet
complete.

My problem is that when gfar_start_xmit() is preemted after the 
tx_queue->tx_skbuff[tx_queue->skb_curtx] is set but before the DMA is started
then the NAPI-poll never completes because it sees a packet which never
completes because the DMA engine did no start yet and won't.
On non-RT SMP systems this isn't a big problem because on the first iteration
the DMA engine might be idle but on the second the other CPU most likely
started the DMA engine and on the fifth the packet might be gone so you
stop (finally).

What happens on your slow link setup is probably the following:
You enqueue hundrets of packets which need TX cleanup. Since that link
is *that* slow, you spent a bunch of cycles calling gfar_clean_tx_ring()
with zero cleanup and you can't leave the poll routine because there is
TX skb not cleaned up. What amazes me is hat it keeps you CPU busy for
as long as 23secs.
*IF* the link goes down in the middle of a cleanup you should see
similar stall because that TX packet won't leave the device so you never
cleanup. So what happens here? Do you get an error interrupt which
purges that skb or do you wait for ndo_tx_timeout()? One of these two
will save your ass but it ain't pretty.

To fix properly with something that works on -RT and mainline I suggest
to revert this patch and add the following:
- do not set has_tx_work unless gfar_clean_tx_ring() unless atleast one
  skb has been cleaned up.
- take the TX cleanup into NAPI accounting. I am not sure if it is
  realistic that one CPU is filling the queue and the other cleans up
  continuously assuming a GBIT link and small packets. However this should
  put a limit here.

which looks in C like this:

> Thanks,
> Claudiu

Sebastian
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Comments

Claudiu Manoil March 28, 2014, 8:19 a.m. UTC | #1 
On 3/27/2014 2:53 PM, Sebastian Andrzej Siewior wrote:
> On 2013-10-14 18:11:15 [+0300], Claudiu Manoil wrote:
>>>> BUG: soft lockup - CPU#0 stuck for 23s! [iperf:2847]
>>>> NIP [c0255b6c] find_next_bit+0xb8/0xc4
>>>> LR [c0367ae8] gfar_poll+0xc8/0x1d8
>>> It seems there is a race condition, and this patch only makes it happen
>>> less often ?
>>>
>>> return faster means what exactly ?
>>>
>>
>> Hi Eric,
>> Because of the outer while loop, gfar_poll may not return due
>> to continuous tx work. The later implementation of gfar_poll
>> allows only one iteration of the Tx queues before returning
>> control to net_rx_action(), that's what I meant with "returns faster".
>
> We talk here about 23secs of cleanup. RX is limited by NAPI and TX is
> limited because it can't be refilled on your UP system.
> Does your box recover from this condition without this patch? Mine does
> not. But I run -RT and stumbled uppon something different.
>
> What I observe is that the TX queue is not empty but does not make any
> progress. That means tx_queue->tx_skbuff[tx_queue->skb_dirtytx] is true
> and gfar_clean_tx_ring() cleans up zero packages because it is not yet
> complete.
>
> My problem is that when gfar_start_xmit() is preemted after the
> tx_queue->tx_skbuff[tx_queue->skb_curtx] is set but before the DMA is started
> then the NAPI-poll never completes because it sees a packet which never
> completes because the DMA engine did no start yet and won't.

False, that code section from start_xmit() cannot be preempted, because
it has spin_lock_irqsave()/restore() around it (unless you modified
your code).  Will check though if on SMP, for some reason,
clean_tx_ring() enters with 0 skbs to clean.

[...]

> To fix properly with something that works on -RT and mainline I suggest
> to revert this patch and add the following:

This patch cannot be reverted. (why would you?)
This patch fixes the issue from description.  I'm seeing no issues with
P1010 now (on any kind of traffic), and the openwrt/tp-link guys also
confirmed (on the powerpc list) that this patch addresses the issue on
their end.
If you encounter problems with the latest driver code, please submit a
proper issue description indicating the code base you're using and so
on.  Also make sure that the problem you're seeing wasn't already fixed
by one of the latest gianfar fixes from net-next:
http://git.kernel.org/cgit/linux/kernel/git/davem/net-next.git


--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Sebastian Andrzej Siewior March 28, 2014, 8:34 a.m. UTC | #2 
On 2014-03-28 10:19:07 [+0200], Claudiu Manoil wrote:
> >My problem is that when gfar_start_xmit() is preemted after the
> >tx_queue->tx_skbuff[tx_queue->skb_curtx] is set but before the DMA is started
> >then the NAPI-poll never completes because it sees a packet which never
> >completes because the DMA engine did no start yet and won't.
> 
> False, that code section from start_xmit() cannot be preempted, because
> it has spin_lock_irqsave()/restore() around it (unless you modified
> your code).  Will check though if on SMP, for some reason,
> clean_tx_ring() enters with 0 skbs to clean.

I said on -RT. On mainline it can't be preempted as I said. If for
some reason you can't get your packet out (on a slow link as you in your
case) it will return with 0 cleanups.
This has been broken since c233cf4 ("gianfar: Fix tx napi polling")
since you drop the return value.

> [...]
> 
> >To fix properly with something that works on -RT and mainline I suggest
> >to revert this patch and add the following:
> 
> This patch cannot be reverted. (why would you?)
Because it does not fix a thing it simply duck tapes the issue that a TX
transfer does not cleanup a thing and you assume that it did something.
You have budget a reserved for RX cleanup which you do not use up if possible.
You simple do one loop and leave.

> This patch fixes the issue from description.  I'm seeing no issues with
> P1010 now (on any kind of traffic), and the openwrt/tp-link guys also
> confirmed (on the powerpc list) that this patch addresses the issue on
> their end.
Simply because the stall is gone doesn't make it good. As you had no
idea why.

> If you encounter problems with the latest driver code, please submit a
> proper issue description indicating the code base you're using and so
> on.  Also make sure that the problem you're seeing wasn't already fixed
> by one of the latest gianfar fixes from net-next:
> http://git.kernel.org/cgit/linux/kernel/git/davem/net-next.git

I pointed out _why_ you saw the stall and the fix involved not to
endless loop on TX clean up on yet transmitted packages. The removal of
outer loop was not required. The issue is present since c233cf4 which
made it in v3.10 into the kernel. 

Sebastian
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Claudiu Manoil March 28, 2014, 9:46 a.m. UTC | #3 
On 3/28/2014 10:34 AM, Sebastian Andrzej Siewior wrote:
> On 2014-03-28 10:19:07 [+0200], Claudiu Manoil wrote:
>>> My problem is that when gfar_start_xmit() is preemted after the
>>> tx_queue->tx_skbuff[tx_queue->skb_curtx] is set but before the DMA is started
>>> then the NAPI-poll never completes because it sees a packet which never
>>> completes because the DMA engine did no start yet and won't.
>>
>> False, that code section from start_xmit() cannot be preempted, because
>> it has spin_lock_irqsave()/restore() around it (unless you modified
>> your code).  Will check though if on SMP, for some reason,
>> clean_tx_ring() enters with 0 skbs to clean.
>
> I said on -RT. On mainline it can't be preempted as I said. If for
> some reason you can't get your packet out (on a slow link as you in your
> case) it will return with 0 cleanups.
> This has been broken since c233cf4 ("gianfar: Fix tx napi polling")
> since you drop the return value.
>
>> [...]
>>
>>> To fix properly with something that works on -RT and mainline I suggest
>>> to revert this patch and add the following:
>>
>> This patch cannot be reverted. (why would you?)
> Because it does not fix a thing it simply duck tapes the issue that a TX
> transfer does not cleanup a thing and you assume that it did something.
> You have budget a reserved for RX cleanup which you do not use up if possible.
> You simple do one loop and leave.

Your proposed fix doesn't fix the root cause either, it's just a 
workaround that came late.  Do you suggest consuming Rx budget for
Tx processing as a better workaround?
Note that the NAPI processing code has been changed in the meanwhile:
http://git.kernel.org/cgit/linux/kernel/git/davem/net-next.git
to address other issues (see aeb12c5ef7cb08d879af22fc0a56cab9e70689ea,
and 71ff9e3df7e1c5d3293af6b595309124e8c97412).


--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
diff mbox

Patch

diff --git a/drivers/net/ethernet/freescale/gianfar.c b/drivers/net/ethernet/freescale/gianfar.c
index 1799ff0..19192c4 100644
--- a/drivers/net/ethernet/freescale/gianfar.c
+++ b/drivers/net/ethernet/freescale/gianfar.c
@@ -132,7 +132,6 @@  static int gfar_poll(struct napi_struct *napi, int budget);
 static void gfar_netpoll(struct net_device *dev);
 #endif
 int gfar_clean_rx_ring(struct gfar_priv_rx_q *rx_queue, int rx_work_limit);
-static void gfar_clean_tx_ring(struct gfar_priv_tx_q *tx_queue);
 static void gfar_process_frame(struct net_device *dev, struct sk_buff *skb,
 			       int amount_pull, struct napi_struct *napi);
 void gfar_halt(struct net_device *dev);
@@ -2473,7 +2472,7 @@  static void gfar_align_skb(struct sk_buff *skb)
 }
 
 /* Interrupt Handler for Transmit complete */
-static void gfar_clean_tx_ring(struct gfar_priv_tx_q *tx_queue)
+static int gfar_clean_tx_ring(struct gfar_priv_tx_q *tx_queue)
 {
 	struct net_device *dev = tx_queue->dev;
 	struct netdev_queue *txq;
@@ -2854,10 +2853,14 @@  static int gfar_poll(struct napi_struct *napi, int budget)
 			tx_queue = priv->tx_queue[i];
 			/* run Tx cleanup to completion */
 			if (tx_queue->tx_skbuff[tx_queue->skb_dirtytx]) {
-				gfar_clean_tx_ring(tx_queue);
-				has_tx_work = 1;
+				int ret;
+
+				ret = gfar_clean_tx_ring(tx_queue);
+				if (ret)
+					has_tx_work++;
 			}
 		}
+		work_done += has_tx_work;
 
 		for_each_set_bit(i, &gfargrp->rx_bit_map, priv->num_rx_queues) {
 			/* skip queue if not active */