NAPI: Polling function requesting extra calls

* NAPI: Polling function requesting extra calls - difference between 3.18.11 and 4.0.0
@ 2015-04-23 18:18 Rafał Miłecki
  2015-04-23 18:28 ` Eric Dumazet
  0 siblings, 1 reply; 6+ messages in thread
From: Rafał Miłecki @ 2015-04-23 18:18 UTC (permalink / raw)
  To: Network Development; +Cc: Felix Fietkau

[-- Attachment #1: Type: text/plain, Size: 4042 bytes --]

Hi,

Recently Felix improved bgmac driver to be smarter in situation where
new packets were Tx/Rx-ed during bgmac_poll execution. It was handled
in:
eb64e29 bgmac: leave interrupts disabled as long as there is work to do
http://git.kernel.org/cgit/linux/kernel/git/torvalds/linux.git/commit/?id=eb64e2923a886441c7b322f138b36029f3fa6a36
With above change, after handling all skb-s in bgmac_poll, bgmac
checks if there are new sbk-s to be handled. If so, it doesn't call
napi_complete & it doesn't enable interrupts.

Above commit was merged for 4.1 kernel, but I work with OpenWrt based
on older releases, so I have all bgmac changes backported to 3.8.11
and 4.0.


With 3.8.11 Felix's change seems to be working fine, I did some debugging:

[   44.970000] [bgmac_interrupt] int_status & int_mask = 0x01000000

[   44.980000] [bgmac_poll] START weight:64
[   44.990000] [bgmac_poll] Read handled:0 skb-s
[   44.990000] [bgmac_poll] Read IRQ status again BGMAC_IS_TX0:0 BGMAC_IS_RX:0
[   45.000000] [bgmac_poll] Read skb-s < weight, calling napi_complete, en IRQs
[   45.010000] [bgmac_poll] END returning handled:0

[   48.040000] [bgmac_interrupt] int_status & int_mask = 0x01000000

[   48.050000] [bgmac_poll] START weight:64
[   48.060000] [bgmac_poll] Read handled:0 skb-s
[   48.060000] [bgmac_poll] Read IRQ status again BGMAC_IS_TX0:0 BGMAC_IS_RX:0
[   48.070000] [bgmac_poll] Read skb-s < weight, calling napi_complete, en IRQs
[   48.080000] [bgmac_poll] END returning handled:0

[   48.100000] [bgmac_poll] START weight:64
[   48.110000] [bgmac_poll] Read handled:1 skb-s
[   48.110000] [bgmac_poll] Read IRQ status again BGMAC_IS_TX0:1 BGMAC_IS_RX:1
[   48.120000] [bgmac_poll] END more skb-s to handle, returning handled:1

[   48.130000] [bgmac_poll] START weight:64
[   48.130000] [bgmac_poll] Read handled:1 skb-s
[   48.140000] [bgmac_poll] Read IRQ status again BGMAC_IS_TX0:0 BGMAC_IS_RX:0
[   48.150000] [bgmac_poll] Read skb-s < weight, calling napi_complete, en IRQs
[   48.160000] [bgmac_poll] END returning handled:1

Please note that at 48.120000 bgmac decided bgmac_poll should be
called again and it didn't call napi_complete & didn't enable IRQs. It
simply returned 1.
NAPI behaved just like we expected, bgmac_poll was called again a
moment later. It handled skb-s that appeared meanwhile and ended
calling napi_complete and returning 1.


Now, it fails badly for me at all when using kernel 4.0.0:

[   52.800000] [bgmac_interrupt] int_status & int_mask = 0x01000000

[   52.800000] [bgmac_poll] START weight:64
[   52.810000] [bgmac_poll] Read handled:0 skb-s
[   52.810000] [bgmac_poll] Read IRQ status again BGMAC_IS_TX0:0 BGMAC_IS_RX:0
[   52.820000] [bgmac_poll] Read skb-s < weight, calling napi_complete, en IRQs
[   52.830000] [bgmac_poll] END returning handled:0

[   54.610000] [bgmac_interrupt] int_status & int_mask = 0x01000000

[   54.620000] [bgmac_poll] START weight:64
[   54.630000] [bgmac_poll] Read handled:0 skb-s
[   54.630000] [bgmac_poll] Read IRQ status again BGMAC_IS_TX0:0 BGMAC_IS_RX:0
[   54.640000] [bgmac_poll] Read skb-s < weight, calling napi_complete, en IRQs
[   54.650000] [bgmac_poll] END returning handled:0

[   55.900000] [bgmac_interrupt] int_status & int_mask = 0x01000000

[   55.910000] [bgmac_poll] START weight:64
[   55.920000] [bgmac_poll] Read handled:1 skb-s
[   55.920000] [bgmac_poll] Read IRQ status again BGMAC_IS_TX0:1 BGMAC_IS_RX:1
[   55.930000] [bgmac_poll] END more skb-s to handle, returning handled:1

As you can see, at 55.930000 bgmac also decided bgmac_poll should be
called again. It didn't call napi_complete, didn't enable IRQs. It
simply returned 1. Just like with kernel 3.8.11.
It this case however, NAPI never called bgmac_poll again. It resulted
in bgmac hanging (IRQs are off, bgmac expects bgmac_poll to be
called).


Can you help us with this, please? Does bgmac use NAPI incorrectly?
Were there some important changes in 3.19 or 4.0?

-- 
Rafał

[-- Attachment #2: tmp.diff --]
[-- Type: text/plain, Size: 1843 bytes --]

diff --git a/drivers/net/ethernet/broadcom/bgmac.c b/drivers/net/ethernet/broadcom/bgmac.c
index 4929932..40ce34d 100644
--- a/drivers/net/ethernet/broadcom/bgmac.c
+++ b/drivers/net/ethernet/broadcom/bgmac.c
@@ -1219,6 +1219,7 @@ static irqreturn_t bgmac_interrupt(int irq, void *dev_id)
 	struct bgmac *bgmac = netdev_priv(dev_id);
 
 	u32 int_status = bgmac_read(bgmac, BGMAC_INT_STATUS);
+	pr_info("[%s] int_status:0x%08x & int_mask:0x%08x = 0x%08x\n", __FUNCTION__, int_status, bgmac->int_mask, int_status & bgmac->int_mask);
 	int_status &= bgmac->int_mask;
 
 	if (!int_status)
@@ -1240,22 +1241,31 @@ static int bgmac_poll(struct napi_struct *napi, int weight)
 {
 	struct bgmac *bgmac = container_of(napi, struct bgmac, napi);
 	int handled = 0;
+	u32 int_status;
 
+	pr_info("[%s] START weight:%d\n", __FUNCTION__, weight);
 	/* Ack */
 	bgmac_write(bgmac, BGMAC_INT_STATUS, ~0);
 
 	bgmac_dma_tx_free(bgmac, &bgmac->tx_ring[0]);
 	handled += bgmac_dma_rx_read(bgmac, &bgmac->rx_ring[0], weight);
+	pr_info("[%s] Read handled:%d skb-s\n", __FUNCTION__, handled);
 
 	/* Poll again if more events arrived in the meantime */
-	if (bgmac_read(bgmac, BGMAC_INT_STATUS) & (BGMAC_IS_TX0 | BGMAC_IS_RX))
+	int_status = bgmac_read(bgmac, BGMAC_INT_STATUS);
+	pr_info("[%s] Read IRQ status again BGMAC_IS_TX0:%d BGMAC_IS_RX:%d\n", __FUNCTION__, !!(int_status & BGMAC_IS_TX0), !!(int_status & BGMAC_IS_RX));
+	if (int_status & (BGMAC_IS_TX0 | BGMAC_IS_RX)) {
+		pr_info("[%s] END more skb-s to handle, returning handled:%d\n", __FUNCTION__, handled);
 		return handled;
+	}
 
 	if (handled < weight) {
+		pr_info("[%s] Read skb-s < weight, calling napi_complete, en IRQs\n", __FUNCTION__);
 		napi_complete(napi);
 		bgmac_chip_intrs_on(bgmac);
 	}
 
+	pr_info("[%s] END returning handled:%d\n", __FUNCTION__, handled);
 	return handled;
 }
 

^ permalink raw reply related	[flat|nested] 6+ messages in thread