Skip to contentSkip to conversation

b43: use rx desc underrun interrupt

PATCH New Thommy Jakobsson 5 participants 45 messages Apr 15, 2013 – Apr 23, 2013 2 messages not in the archive Networking drivers (wireless)

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Hi,

long user of b43 driver under openwrt. Not sure if it has been posted previously but using slow and old hardware (as routers running openwrt most of the time is), the CPU is sometimes slower than the NIC. So recently I, and several others, has had problem of wifi going down. See ticket 7552 at openwrt for more info
(https://dev.openwrt.org/ticket/7552)

I have tracked down the problem to the nic getting into a rx descriptor underrun. Since the driver don't listen to that (or take care of it in any other way), the Wifi seemingly just dies. The CPU is waiting for data from the nic and the nic is waiting for a confirmation that it can keep sending.

I created a patch for handling of that interrupt. I have tested it by reducing the number of rx slots to 16 and sending a couple of 100GB of data. Previously Wifi went down within a minute when maxing out the uplink (so rx for AP), now it has been working flawless for a 1.5week.

BR,
Thommy Jakobsson

Index: drivers/net/wireless/b43/dma.c
===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/dma.c

2013-04-09 17:33:54.325322046 +0200
+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/dma.c
2013-04-09 17:34:43.473565843 +0200
@@ -1688,6 +1688,21 @@
         b43_poison_rx_buffer(ring, skb);
         sync_descbuffer_for_device(ring, dmaaddr, ring->rx_buffersize);

}
+void b43_dma_rx_discard(struct b43_dmaring *ring) +{
+ B43_WARN_ON(ring->tx);
+
+ /* Device has filled all buffers, drop all packets in buffers

+     * and let TCP decrease speed.
+     * Set index to one desc after the last one (which is marked)
+     * so the device will see all slots as free again
+     */

+ /*
+ *TODO: How to increase rx_drop in mac80211
+ */
+ b43_dma_write(ring, B43_DMA32_RXINDEX, ring->nr_slots *
+ sizeof(struct b43_dmadesc32));
+}

void b43_dma_rx(struct b43_dmaring *ring)
{

Index: drivers/net/wireless/b43/dma.h
===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/dma.h

2013-04-09 17:34:09.561397686 +0200
+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/dma.h
2013-04-09 17:34:43.461565780 +0200
@@ -10,7 +10,8 @@
 #define B43_DMAIRQ_FATALMASK    ((1 << 10) | (1 << 11) | (1 << 12) \
                                          | (1 << 14) | (1 << 15))

#define B43_DMAIRQ_NONFATALMASK (1 << 13)
-#define B43_DMAIRQ_RX_DONE (1 << 16)
+#define B43_DMAIRQ_RX_DONE (1 << 16)
+#define B43_DMAIRQ_RDESC_UFLOW (1 << 13)

/*** 32-bit DMA Engine. ***/

@@ -295,6 +296,8 @@
void b43_dma_handle_txstatus(struct b43_wldev *dev,

                              const struct b43_txstatus *status);

+void b43_dma_rx_discard(struct b43_dmaring *ring); +
void b43_dma_rx(struct b43_dmaring *ring);

void b43_dma_direct_fifo_rx(struct b43_wldev *dev,

Index: drivers/net/wireless/b43/main.c
===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/main.c

2013-04-09 17:33:54.325322046 +0200
+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/main.c
2013-04-09 18:08:01.011471215 +0200
@@ -1894,14 +1894,6 @@
                         b43_controller_restart(dev, "DMA error");
                         return;
                 }
-                if (merged_dma_reason & B43_DMAIRQ_NONFATALMASK) {
-                        b43err(dev->wl, "DMA error: "
-                               "0x%08X, 0x%08X, 0x%08X, "
-                               "0x%08X, 0x%08X, 0x%08X\n",
-                               dma_reason[0], dma_reason[1],
-                               dma_reason[2], dma_reason[3],
-                               dma_reason[4], dma_reason[5]);

- }
}

if (unlikely(reason & B43_IRQ_UCODE_DEBUG))
@@ -1920,6 +1912,14 @@

                 handle_irq_noise(dev);

         /* Check the DMA reason registers for received data. */

+ if (dma_reason[0] & B43_DMAIRQ_RDESC_UFLOW) {
+ //only print 256 time to not flood log
+ if(!(dev->stats.rxdesc_underruns++&0xFF)){
+ b43warn(dev->wl, "Rx descriptor underrun
(high cpu load?), throwing packets\n");
+ }
+ b43_dma_rx_discard(dev->dma.rx_ring);
+
+ }

         if (dma_reason[0] & B43_DMAIRQ_RX_DONE) {
                 if (b43_using_pio_transfers(dev))
                         b43_pio_rx(dev->pio.rx_queue);
@@ -1977,7 +1977,7 @@
                 return IRQ_NONE;

         dev->dma_reason[0] = b43_read32(dev, B43_MMIO_DMA0_REASON)
-            & 0x0001DC00;
+            & 0x0001FC00;
         dev->dma_reason[1] = b43_read32(dev, B43_MMIO_DMA1_REASON)
             & 0x0000DC00;
         dev->dma_reason[2] = b43_read32(dev, B43_MMIO_DMA2_REASON)
@@ -3081,7 +3081,7 @@
                 b43_write32(dev, 0x018C, 0x02000000);
         }
         b43_write32(dev, B43_MMIO_GEN_IRQ_REASON, 0x00004000);
-        b43_write32(dev, B43_MMIO_DMA0_IRQ_MASK, 0x0001DC00);
+        b43_write32(dev, B43_MMIO_DMA0_IRQ_MASK, 0x0001FC00);
         b43_write32(dev, B43_MMIO_DMA1_IRQ_MASK, 0x0000DC00);
         b43_write32(dev, B43_MMIO_DMA2_IRQ_MASK, 0x0000DC00);
         b43_write32(dev, B43_MMIO_DMA3_IRQ_MASK, 0x0001DC00);

===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/b43.h

2013-04-09 17:04:51.552680190 +0200
+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/b43.h
2013-04-09 17:46:08.472962612 +0200

@@ -671,6 +671,7 @@

 struct b43_stats {
         u8 link_noise;

+ u32 rxdesc_underruns;
};

struct b43_key {

#1 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Hi,

Quoted 15 lines · Thommy Jakobsson

Hi,

long user of b43 driver under openwrt. Not sure if it has been posted previously but using slow and old hardware (as routers running openwrt most of the time is), the CPU is sometimes slower than the NIC. So recently I, and several others, has had problem of wifi going down. See ticket 7552 at openwrt for more info
(https://dev.openwrt.org/ticket/7552)

I have tracked down the problem to the nic getting into a rx descriptor underrun. Since the driver don't listen to that (or take care of it in any other way), the Wifi seemingly just dies. The CPU is waiting for data from the nic and the nic is waiting for a confirmation that it can keep sending.

Quoted 4 lines

I created a patch for handling of that interrupt. I have tested it by reducing the number of rx slots to 16 and sending a couple of 100GB of data. Previously Wifi went down within a minute when maxing out the uplink (so rx for AP), now it has been working flawless for a 1.5week.

Nice, very good to hear. Looks like you found the real issue the rx descriptor count increase was just papering over. Some comments regarding your patch.

Quoted 4 lines

BR,
Thommy Jakobsson

You are missing a "Signed-off-by" - without it the patch can never be accepted. Also your patch is has broken tabs (they were converted to spaces), as well as is line-broken, so it doesn't apply anymore. You can't use gmail's web interface for sending patches.

Quoted 5 lines

Index: drivers/net/wireless/b43/dma.c
===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/dma.c

2013-04-09 17:33:54.325322046 +0200

+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/dma.c

You should make your patch against the newest git tree, so usually wireless-testing or Rafa?'s b43-next.

Quoted 31 lines
2013-04-09 17:34:43.473565843 +0200
(quoted diff) +15 -0
@@ -1688,6 +1688,21 @@
         b43_poison_rx_buffer(ring, skb);
         sync_descbuffer_for_device(ring, dmaaddr, ring->rx_buffersize);
 }
+void b43_dma_rx_discard(struct b43_dmaring *ring)
+{
+   B43_WARN_ON(ring->tx);
+
+    /* Device has filled all buffers, drop all packets in buffers
+     * and let TCP decrease speed.
+     * Set index to one desc after the last one (which is marked)
+     * so the device will see all slots as free again
+     */
+    /*
+     *TODO: How to increase rx_drop in mac80211
+     */
+   b43_dma_write(ring, B43_DMA32_RXINDEX, ring->nr_slots *
+                                     sizeof(struct b43_dmadesc32));
+}
 
 void b43_dma_rx(struct b43_dmaring *ring)
 {

Index: drivers/net/wireless/b43/dma.h
===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/dma.h

2013-04-09 17:34:09.561397686 +0200
+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/dma.h
2013-04-09 17:34:43.461565780 +0200

Review line 13

Quoted 5 lines
(quoted diff) +1 -1
@@ -10,7 +10,8 @@
 #define B43_DMAIRQ_FATALMASK    ((1 << 10) | (1 << 11) | (1 << 12) \
                                          | (1 << 14) | (1 << 15))
 #define B43_DMAIRQ_NONFATALMASK (1 << 13)
-#define B43_DMAIRQ_RX_DONE              (1 << 16)
+#define B43_DMAIRQ_RX_DONE      (1 << 16)

Unnecessary whitespace change.

Review line 14

Quoted 1 line
(quoted diff) +1 -0
+#define B43_DMAIRQ_RDESC_UFLOW  (1 << 13)

Why not just rename B43_DMAIRQ_NONFATALMASK ?

Quoted 21 lines
(quoted diff) +2 -0
 /*** 32-bit DMA Engine. ***/
 
@@ -295,6 +296,8 @@
 void b43_dma_handle_txstatus(struct b43_wldev *dev,
                              const struct b43_txstatus *status);
 
+void b43_dma_rx_discard(struct b43_dmaring *ring);
+
 void b43_dma_rx(struct b43_dmaring *ring);
 
 void b43_dma_direct_fifo_rx(struct b43_wldev *dev,

Index: drivers/net/wireless/b43/main.c
===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/main.c

2013-04-09 17:33:54.325322046 +0200
+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/main.c
2013-04-09 18:08:01.011471215 +0200
(quoted diff) +0 -0
@@ -1894,14 +1894,6 @@
                         b43_controller_restart(dev, "DMA error");

Review line 1904

Quoted 10 lines
(quoted diff) +0 -8
                         return;
                 }
-                if (merged_dma_reason & B43_DMAIRQ_NONFATALMASK) {
-                        b43err(dev->wl, "DMA error: "
-                               "0x%08X, 0x%08X, 0x%08X, "
-                               "0x%08X, 0x%08X, 0x%08X\n",
-                               dma_reason[0], dma_reason[1],
-                               dma_reason[2], dma_reason[3],
-                               dma_reason[4], dma_reason[5]);
-                }

You should say why you remove this one, and you should remove the B43_DMAIRQ_NONFATALMASK from the if (unlikely()) enclosing it.

Review line 1919

Quoted 12 lines
(quoted diff) +4 -0
         }
 
         if (unlikely(reason & B43_IRQ_UCODE_DEBUG))
@@ -1920,6 +1912,14 @@
                 handle_irq_noise(dev);
 
         /* Check the DMA reason registers for received data. */
+        if (dma_reason[0] & B43_DMAIRQ_RDESC_UFLOW) {
+           //only print 256 time to not flood log
+           if(!(dev->stats.rxdesc_underruns++&0xFF)){
+                        b43warn(dev->wl, "Rx descriptor underrun
 (high cpu load?), throwing packets\n");

Apart from the style problems, maybe doing a

	if (printk_ratelimit())
		b43warn(...);

which will ensure that it won't flood the log, but still won't stop reporting them after the 256th time.

Quoted 32 lines
(quoted diff) +4 -0
+           }
+           b43_dma_rx_discard(dev->dma.rx_ring);
+
+        }
         if (dma_reason[0] & B43_DMAIRQ_RX_DONE) {
                 if (b43_using_pio_transfers(dev))
                         b43_pio_rx(dev->pio.rx_queue);
(quoted diff) +2 -2
@@ -1977,7 +1977,7 @@
                 return IRQ_NONE;
 
         dev->dma_reason[0] = b43_read32(dev, B43_MMIO_DMA0_REASON)
-            & 0x0001DC00;
+            & 0x0001FC00;
         dev->dma_reason[1] = b43_read32(dev, B43_MMIO_DMA1_REASON)
             & 0x0000DC00;
         dev->dma_reason[2] = b43_read32(dev, B43_MMIO_DMA2_REASON)
@@ -3081,7 +3081,7 @@
                 b43_write32(dev, 0x018C, 0x02000000);
         }
         b43_write32(dev, B43_MMIO_GEN_IRQ_REASON, 0x00004000);
-        b43_write32(dev, B43_MMIO_DMA0_IRQ_MASK, 0x0001DC00);
+        b43_write32(dev, B43_MMIO_DMA0_IRQ_MASK, 0x0001FC00);
         b43_write32(dev, B43_MMIO_DMA1_IRQ_MASK, 0x0000DC00);
         b43_write32(dev, B43_MMIO_DMA2_IRQ_MASK, 0x0000DC00);
         b43_write32(dev, B43_MMIO_DMA3_IRQ_MASK, 0x0001DC00);

===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/b43.h

2013-04-09 17:04:51.552680190 +0200
+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/b43.h
2013-04-09 17:46:08.472962612 +0200

Review line 677

Quoted 7 lines
(quoted diff) +1 -0
@@ -671,6 +671,7 @@
 
 struct b43_stats {
         u8 link_noise;
+        u32 rxdesc_underruns;
 };
 
 struct b43_key {

Jonas

#2 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 6 lines · On Fri, 19 Apr 2013 16:32:00 +0200 Jonas Gorski <[email protected]> wrote:

Apart from the style problems, maybe doing a

	if (printk_ratelimit())
		b43warn(...);

which will ensure that it won't flood the log, but still won't stop reporting them after the 256th time.

All b43 log helpers are ratelimited already. So no need for that.

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130419/a57f356e/attachment.sig>
#3 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 11 lines · Michael B?sch
Quoted 6 lines · On Fri, 19 Apr 2013 16:32:00 +0200 Jonas Gorski <[email protected]> wrote:

Apart from the style problems, maybe doing a

      if (printk_ratelimit())
              b43warn(...);

which will ensure that it won't flood the log, but still won't stop reporting them after the 256th time.

All b43 log helpers are ratelimited already. So no need for that.

Even better :-)

Jonas

#4 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

In reply to Jonas Gorski
Quoted 178 lines · Jonas Gorski

Hi,

Quoted 15 lines · Thommy Jakobsson

Hi,

long user of b43 driver under openwrt. Not sure if it has been posted previously but using slow and old hardware (as routers running openwrt most of the time is), the CPU is sometimes slower than the NIC. So recently I, and several others, has had problem of wifi going down. See ticket 7552 at openwrt for more info
(https://dev.openwrt.org/ticket/7552)

I have tracked down the problem to the nic getting into a rx descriptor underrun. Since the driver don't listen to that (or take care of it in any other way), the Wifi seemingly just dies. The CPU is waiting for data from the nic and the nic is waiting for a confirmation that it can keep sending.

Quoted 4 lines

I created a patch for handling of that interrupt. I have tested it by reducing the number of rx slots to 16 and sending a couple of 100GB of data. Previously Wifi went down within a minute when maxing out the uplink (so rx for AP), now it has been working flawless for a 1.5week.

Nice, very good to hear. Looks like you found the real issue the rx descriptor count increase was just papering over. Some comments regarding your patch.

Quoted 4 lines

BR,
Thommy Jakobsson

You are missing a "Signed-off-by" - without it the patch can never be accepted. Also your patch is has broken tabs (they were converted to spaces), as well as is line-broken, so it doesn't apply anymore. You can't use gmail's web interface for sending patches.

Quoted 5 lines

Index: drivers/net/wireless/b43/dma.c
===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/dma.c

2013-04-09 17:33:54.325322046 +0200

+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/dma.c

You should make your patch against the newest git tree, so usually wireless-testing or Rafa?'s b43-next.

Quoted 31 lines
2013-04-09 17:34:43.473565843 +0200
(quoted diff) +15 -0
@@ -1688,6 +1688,21 @@
          b43_poison_rx_buffer(ring, skb);
          sync_descbuffer_for_device(ring, dmaaddr, ring->rx_buffersize);
  }
+void b43_dma_rx_discard(struct b43_dmaring *ring)
+{
+   B43_WARN_ON(ring->tx);
+
+    /* Device has filled all buffers, drop all packets in buffers
+     * and let TCP decrease speed.
+     * Set index to one desc after the last one (which is marked)
+     * so the device will see all slots as free again
+     */
+    /*
+     *TODO: How to increase rx_drop in mac80211
+     */
+   b43_dma_write(ring, B43_DMA32_RXINDEX, ring->nr_slots *
+                                     sizeof(struct b43_dmadesc32));
+}
 
  void b43_dma_rx(struct b43_dmaring *ring)
  {

Index: drivers/net/wireless/b43/dma.h
===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/dma.h

2013-04-09 17:34:09.561397686 +0200
+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/dma.h
2013-04-09 17:34:43.461565780 +0200

Review line 13

Quoted 5 lines
(quoted diff) +1 -1
@@ -10,7 +10,8 @@
  #define B43_DMAIRQ_FATALMASK    ((1 << 10) | (1 << 11) | (1 << 12) \
                                           | (1 << 14) | (1 << 15))
  #define B43_DMAIRQ_NONFATALMASK (1 << 13)
-#define B43_DMAIRQ_RX_DONE              (1 << 16)
+#define B43_DMAIRQ_RX_DONE      (1 << 16)

Unnecessary whitespace change.

Review line 14

Quoted 1 line
(quoted diff) +1 -0
+#define B43_DMAIRQ_RDESC_UFLOW  (1 << 13)

Why not just rename B43_DMAIRQ_NONFATALMASK ?

Quoted 21 lines
(quoted diff) +2 -0
  /*** 32-bit DMA Engine. ***/
 
@@ -295,6 +296,8 @@
  void b43_dma_handle_txstatus(struct b43_wldev *dev,
                               const struct b43_txstatus *status);
 
+void b43_dma_rx_discard(struct b43_dmaring *ring);
+
  void b43_dma_rx(struct b43_dmaring *ring);
 
  void b43_dma_direct_fifo_rx(struct b43_wldev *dev,

Index: drivers/net/wireless/b43/main.c
===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/main.c

2013-04-09 17:33:54.325322046 +0200
+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/main.c
2013-04-09 18:08:01.011471215 +0200
(quoted diff) +0 -0
@@ -1894,14 +1894,6 @@
                          b43_controller_restart(dev, "DMA error");

Review line 1904

Quoted 10 lines
(quoted diff) +0 -8
                          return;
                  }
-                if (merged_dma_reason & B43_DMAIRQ_NONFATALMASK) {
-                        b43err(dev->wl, "DMA error: "
-                               "0x%08X, 0x%08X, 0x%08X, "
-                               "0x%08X, 0x%08X, 0x%08X\n",
-                               dma_reason[0], dma_reason[1],
-                               dma_reason[2], dma_reason[3],
-                               dma_reason[4], dma_reason[5]);
-                }

You should say why you remove this one, and you should remove the B43_DMAIRQ_NONFATALMASK from the if (unlikely()) enclosing it.

Review line 1919

Quoted 12 lines
(quoted diff) +4 -0
          }
 
          if (unlikely(reason & B43_IRQ_UCODE_DEBUG))
@@ -1920,6 +1912,14 @@
                  handle_irq_noise(dev);
 
          /* Check the DMA reason registers for received data. */
+        if (dma_reason[0] & B43_DMAIRQ_RDESC_UFLOW) {
+           //only print 256 time to not flood log
+           if(!(dev->stats.rxdesc_underruns++&0xFF)){
+                        b43warn(dev->wl, "Rx descriptor underrun
 (high cpu load?), throwing packets\n");

Apart from the style problems, maybe doing a

	if (printk_ratelimit())
		b43warn(...);

which will ensure that it won't flood the log, but still won't stop reporting them after the 256th time.

Quoted 33 lines
(quoted diff) +4 -0
+           }
+           b43_dma_rx_discard(dev->dma.rx_ring);
+
+        }
          if (dma_reason[0] & B43_DMAIRQ_RX_DONE) {
                  if (b43_using_pio_transfers(dev))
                          b43_pio_rx(dev->pio.rx_queue);
(quoted diff) +2 -2
@@ -1977,7 +1977,7 @@
                  return IRQ_NONE;
 
          dev->dma_reason[0] = b43_read32(dev, B43_MMIO_DMA0_REASON)
-            & 0x0001DC00;
+            & 0x0001FC00;
          dev->dma_reason[1] = b43_read32(dev, B43_MMIO_DMA1_REASON)
              & 0x0000DC00;
          dev->dma_reason[2] = b43_read32(dev, B43_MMIO_DMA2_REASON)
@@ -3081,7 +3081,7 @@
                  b43_write32(dev, 0x018C, 0x02000000);
          }
          b43_write32(dev, B43_MMIO_GEN_IRQ_REASON, 0x00004000);
-        b43_write32(dev, B43_MMIO_DMA0_IRQ_MASK, 0x0001DC00);
+        b43_write32(dev, B43_MMIO_DMA0_IRQ_MASK, 0x0001FC00);
          b43_write32(dev, B43_MMIO_DMA1_IRQ_MASK, 0x0000DC00);
          b43_write32(dev, B43_MMIO_DMA2_IRQ_MASK, 0x0000DC00);
          b43_write32(dev, B43_MMIO_DMA3_IRQ_MASK, 0x0001DC00);

===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/b43.h

2013-04-09 17:04:51.552680190 +0200
+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/b43.h
2013-04-09 17:46:08.472962612 +0200
(quoted diff) +0 -0
@@ -671,6 +671,7 @@
 
Quoted 6 lines

Review line 677

Quoted 6 lines
(quoted diff) +1 -0
  struct b43_stats {
          u8 link_noise;
+        u32 rxdesc_underruns;
  };
 
  struct b43_key {

Chris,

With this patch, can the number of RX descriptors be reduced back to 64 on your system? I realize it will be be hard to apply with the white space errors and line wraps, but perhaps we will get a clean version soon.

Larry

#5 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 3 lines · On Fri, 19 Apr 2013 13:17:03 -0500 Larry Finger <[email protected]>…

With this patch, can the number of RX descriptors be reduced back to 64 on your system? I realize it will be be hard to apply with the white space errors and line wraps, but perhaps we will get a clean version soon.

It would be good to lower the number of RX descriptors as low as possible. Maybe even lower than 64. We should do more tests on that. That certainly could improve performance on low-end devices.

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130419/fe2ab06e/attachment.sig>
#6 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

As I mentioned in the original post I lowered it to 16 on my system without any problem (and I'm still running like that). I think that was perhaps a bit too low though, didn't do any real measurements but it felt like it cut off every other burst coming in. But I guess that also depends on the target system.

//Thommy

Quoted 13 lines · Michael B?sch
Quoted 3 lines · On Fri, 19 Apr 2013 13:17:03 -0500 Larry Finger <[email protected]>…

With this patch, can the number of RX descriptors be reduced back to 64 on your system? I realize it will be be hard to apply with the white space errors and line wraps, but perhaps we will get a clean version soon.

It would be good to lower the number of RX descriptors as low as possible. Maybe even lower than 64. We should do more tests on that. That certainly could improve performance on low-end devices.

--
Michael

#7 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 5 lines · On Fri, 19 Apr 2013 22:37:28 +0200 Thommy Jakobsson <[email protected]> wrote:

As I mentioned in the original post I lowered it to 16 on my system without any problem (and I'm still running like that). I think that was perhaps a bit too low though, didn't do any real measurements but it felt like it cut off every other burst coming in. But I guess that also depends on the target system.

Yes 16 sounds a little low, but 32 could be a good idea to test.

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130419/18347906/attachment.sig>
#8 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 10 lines · Michael B?sch
Quoted 5 lines · On Fri, 19 Apr 2013 22:37:28 +0200 Thommy Jakobsson <[email protected]> wrote:

As I mentioned in the original post I lowered it to 16 on my system without any problem (and I'm still running like that). I think that was perhaps a bit too low though, didn't do any real measurements but it felt like it cut off every other burst coming in. But I guess that also depends on the target system.

Yes 16 sounds a little low, but 32 could be a good idea to test.

I agree.

Larry

#9 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

In reply to Jonas Gorski

Hi Jonas,

thanks for input. As you see I'm not used to send in patches, but I'll give it a new shot as soon as I succeed to download and patch the latest git tree. Otherwise see my comments inline

//Thommy

Quoted 115 lines · Jonas Gorski

Hi,

Quoted 15 lines · Thommy Jakobsson

Hi,

long user of b43 driver under openwrt. Not sure if it has been posted previously but using slow and old hardware (as routers running openwrt most of the time is), the CPU is sometimes slower than the NIC. So recently I, and several others, has had problem of wifi going down. See ticket 7552 at openwrt for more info
(https://dev.openwrt.org/ticket/7552)

I have tracked down the problem to the nic getting into a rx descriptor underrun. Since the driver don't listen to that (or take care of it in any other way), the Wifi seemingly just dies. The CPU is waiting for data from the nic and the nic is waiting for a confirmation that it can keep sending.

Quoted 4 lines

I created a patch for handling of that interrupt. I have tested it by reducing the number of rx slots to 16 and sending a couple of 100GB of data. Previously Wifi went down within a minute when maxing out the uplink (so rx for AP), now it has been working flawless for a 1.5week.

Nice, very good to hear. Looks like you found the real issue the rx descriptor count increase was just papering over. Some comments regarding your patch.

Quoted 4 lines

BR,
Thommy Jakobsson

You are missing a "Signed-off-by" - without it the patch can never be accepted. Also your patch is has broken tabs (they were converted to spaces), as well as is line-broken, so it doesn't apply anymore. You can't use gmail's web interface for sending patches.

Quoted 5 lines

Index: drivers/net/wireless/b43/dma.c
===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/dma.c

2013-04-09 17:33:54.325322046 +0200

+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/dma.c

You should make your patch against the newest git tree, so usually wireless-testing or Rafa?'s b43-next.

Quoted 31 lines
2013-04-09 17:34:43.473565843 +0200
(quoted diff) +15 -0
@@ -1688,6 +1688,21 @@
         b43_poison_rx_buffer(ring, skb);
         sync_descbuffer_for_device(ring, dmaaddr, ring->rx_buffersize);
 }
+void b43_dma_rx_discard(struct b43_dmaring *ring)
+{
+   B43_WARN_ON(ring->tx);
+
+    /* Device has filled all buffers, drop all packets in buffers
+     * and let TCP decrease speed.
+     * Set index to one desc after the last one (which is marked)
+     * so the device will see all slots as free again
+     */
+    /*
+     *TODO: How to increase rx_drop in mac80211
+     */
+   b43_dma_write(ring, B43_DMA32_RXINDEX, ring->nr_slots *
+                                     sizeof(struct b43_dmadesc32));
+}
 
 void b43_dma_rx(struct b43_dmaring *ring)
 {

Index: drivers/net/wireless/b43/dma.h
===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/dma.h

2013-04-09 17:34:09.561397686 +0200
+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/dma.h
2013-04-09 17:34:43.461565780 +0200

Review line 13

Quoted 5 lines
(quoted diff) +1 -1
@@ -10,7 +10,8 @@
 #define B43_DMAIRQ_FATALMASK    ((1 << 10) | (1 << 11) | (1 << 12) \
                                          | (1 << 14) | (1 << 15))
 #define B43_DMAIRQ_NONFATALMASK (1 << 13)
-#define B43_DMAIRQ_RX_DONE              (1 << 16)
+#define B43_DMAIRQ_RX_DONE      (1 << 16)

Unnecessary whitespace change.

Review line 14

Quoted 1 line
(quoted diff) +1 -0
+#define B43_DMAIRQ_RDESC_UFLOW  (1 << 13)

Why not just rename B43_DMAIRQ_NONFATALMASK ?

Quoted 26 lines
(quoted diff) +2 -0
 /*** 32-bit DMA Engine. ***/
 
@@ -295,6 +296,8 @@
 void b43_dma_handle_txstatus(struct b43_wldev *dev,
                              const struct b43_txstatus *status);
 
+void b43_dma_rx_discard(struct b43_dmaring *ring);
+
 void b43_dma_rx(struct b43_dmaring *ring);
 
 void b43_dma_direct_fifo_rx(struct b43_wldev *dev,

Index: drivers/net/wireless/b43/main.c
===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/main.c

2013-04-09 17:33:54.325322046 +0200
+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/main.c
2013-04-09 18:08:01.011471215 +0200
(quoted diff) +0 -3
@@ -1894,14 +1894,6 @@
                         b43_controller_restart(dev, "DMA error");
                         return;
                 }
-                if (merged_dma_reason & B43_DMAIRQ_NONFATALMASK) {
-                        b43err(dev->wl, "DMA error: "
-                               "0x%08X, 0x%08X, 0x%08X, "
Quoted 8 lines

Review line 1904

Quoted 5 lines
(quoted diff) +0 -5
-                               "0x%08X, 0x%08X, 0x%08X\n",
-                               dma_reason[0], dma_reason[1],
-                               dma_reason[2], dma_reason[3],
-                               dma_reason[4], dma_reason[5]);
-                }

You should say why you remove this one, and you should remove the B43_DMAIRQ_NONFATALMASK from the if (unlikely()) enclosing it.

[thommy]: I removed it because I didn't want to have a printout each time some non fatal happened. That would mean a printout every time a rx descriptor underrun would happen. Also the printout says "DMA error", which sounds very serious when in fact it isn't (or not in my opinion anyway).

Quoted 15 lines

Review line 1919

Quoted 13 lines
(quoted diff) +4 -0
         }
 
         if (unlikely(reason & B43_IRQ_UCODE_DEBUG))
@@ -1920,6 +1912,14 @@
                 handle_irq_noise(dev);
 
         /* Check the DMA reason registers for received data. */
+        if (dma_reason[0] & B43_DMAIRQ_RDESC_UFLOW) {
+           //only print 256 time to not flood log
+           if(!(dev->stats.rxdesc_underruns++&0xFF)){
+                        b43warn(dev->wl, "Rx descriptor underrun
 (high cpu load?), throwing packets\n");

Apart from the style problems, maybe doing a

Quoted 5 lines
        if (printk_ratelimit())
                b43warn(...);

which will ensure that it won't flood the log, but still won't stop reporting them after the 256th time.

[thommy]: My comment in the code is wrong. The meaning was to print it every 256th time it happens, not to only to print the 256 first times. As it seems from other messages there are better way of rate limiting that I was unaware of.

Quoted 44 lines
Quoted 34 lines
(quoted diff) +4 -0
+           }
+           b43_dma_rx_discard(dev->dma.rx_ring);
+
+        }
         if (dma_reason[0] & B43_DMAIRQ_RX_DONE) {
                 if (b43_using_pio_transfers(dev))
                         b43_pio_rx(dev->pio.rx_queue);
(quoted diff) +2 -2
@@ -1977,7 +1977,7 @@
                 return IRQ_NONE;
 
         dev->dma_reason[0] = b43_read32(dev, B43_MMIO_DMA0_REASON)
-            & 0x0001DC00;
+            & 0x0001FC00;
         dev->dma_reason[1] = b43_read32(dev, B43_MMIO_DMA1_REASON)
             & 0x0000DC00;
         dev->dma_reason[2] = b43_read32(dev, B43_MMIO_DMA2_REASON)
@@ -3081,7 +3081,7 @@
                 b43_write32(dev, 0x018C, 0x02000000);
         }
         b43_write32(dev, B43_MMIO_GEN_IRQ_REASON, 0x00004000);
-        b43_write32(dev, B43_MMIO_DMA0_IRQ_MASK, 0x0001DC00);
+        b43_write32(dev, B43_MMIO_DMA0_IRQ_MASK, 0x0001FC00);
         b43_write32(dev, B43_MMIO_DMA1_IRQ_MASK, 0x0000DC00);
         b43_write32(dev, B43_MMIO_DMA2_IRQ_MASK, 0x0000DC00);
         b43_write32(dev, B43_MMIO_DMA3_IRQ_MASK, 0x0001DC00);

===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/b43.h

2013-04-09 17:04:51.552680190 +0200
+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/b43.h
2013-04-09 17:46:08.472962612 +0200

Review line 677

Quoted 7 lines
(quoted diff) +1 -0
@@ -671,6 +671,7 @@
 
 struct b43_stats {
         u8 link_noise;
+        u32 rxdesc_underruns;
 };
 
 struct b43_key {

Jonas

author
#10 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

In reply to Jonas Gorski

Hi,

new try sending in the patch (with pine instead of gmail webclient this time). Patch created on latest wireless-testing

BR,
Thommy Jakobsson

---
 drivers/net/wireless/b43/b43.h  |    1 +
 drivers/net/wireless/b43/dma.c  |   16 +++++++++++++++
 drivers/net/wireless/b43/dma.h  |    4 +++-
 drivers/net/wireless/b43/main.c |   43 ++++++++++++++++-----------------------
 4 files changed, 38 insertions(+), 26 deletions(-)
drivers/net/wireless/b43/b43.h +1 -0
index 7f3d461..fdcd00f 100644
@@ -680,6 +680,7 @@ struct b43_noise_calculation {
 
 struct b43_stats {
 	u8 link_noise;
+	u32 rxdesc_underruns;
 };
 
 struct b43_key {
drivers/net/wireless/b43/dma.c +16 -0
index 1221469..2ae00dd 100644
@@ -1733,6 +1733,22 @@ drop_recycle_buffer:
 	sync_descbuffer_for_device(ring, dmaaddr, ring->rx_buffersize);
 }
 
+void b43_dma_rx_discard(struct b43_dmaring *ring)
+{
+	B43_WARN_ON(ring->tx);
+
+	/* Device has filled all buffers, drop all packets in buffers
+	* and let TCP decrease speed.
+	* Set index to one desc after the last one
+	* so the device will see all slots as free again
+	*/
+	/*
+	*TODO: How to increase rx_drop in mac80211?
+	*/
+	b43_dma_write(ring, B43_DMA32_RXINDEX, ring->nr_slots *
+			sizeof(struct b43_dmadesc32));
+}
+
 void b43_dma_rx(struct b43_dmaring *ring)
 {
 	const struct b43_dma_ops *ops = ring->ops;
drivers/net/wireless/b43/dma.h +3 -1
index 9fdd198..fed8163 100644
@@ -9,7 +9,7 @@
 /* DMA-Interrupt reasons. */
 #define B43_DMAIRQ_FATALMASK	((1 << 10) | (1 << 11) | (1 << 12) \
 					 | (1 << 14) | (1 << 15))
-#define B43_DMAIRQ_NONFATALMASK	(1 << 13)
+#define B43_DMAIRQ_RDESC_UFLOW		(1 << 13)
 #define B43_DMAIRQ_RX_DONE		(1 << 16)
 
 /*** 32-bit DMA Engine. ***/
@@ -295,6 +295,8 @@ int b43_dma_tx(struct b43_wldev *dev,
 void b43_dma_handle_txstatus(struct b43_wldev *dev,
 			     const struct b43_txstatus *status);
 
+void b43_dma_rx_discard(struct b43_dmaring *ring);
+
 void b43_dma_rx(struct b43_dmaring *ring);
 
 void b43_dma_direct_fifo_rx(struct b43_wldev *dev,
drivers/net/wireless/b43/main.c +18 -25

68 lines not on this page · raw message

Signature · 1 line
1.7.9.5
Quoted 188 lines · Jonas Gorski

Hi,

Quoted 15 lines · Thommy Jakobsson

Hi,

long user of b43 driver under openwrt. Not sure if it has been posted previously but using slow and old hardware (as routers running openwrt most of the time is), the CPU is sometimes slower than the NIC. So recently I, and several others, has had problem of wifi going down. See ticket 7552 at openwrt for more info
(https://dev.openwrt.org/ticket/7552)

I have tracked down the problem to the nic getting into a rx descriptor underrun. Since the driver don't listen to that (or take care of it in any other way), the Wifi seemingly just dies. The CPU is waiting for data from the nic and the nic is waiting for a confirmation that it can keep sending.

Quoted 4 lines

I created a patch for handling of that interrupt. I have tested it by reducing the number of rx slots to 16 and sending a couple of 100GB of data. Previously Wifi went down within a minute when maxing out the uplink (so rx for AP), now it has been working flawless for a 1.5week.

Nice, very good to hear. Looks like you found the real issue the rx descriptor count increase was just papering over. Some comments regarding your patch.

Quoted 4 lines

BR,
Thommy Jakobsson

You are missing a "Signed-off-by" - without it the patch can never be accepted. Also your patch is has broken tabs (they were converted to spaces), as well as is line-broken, so it doesn't apply anymore. You can't use gmail's web interface for sending patches.

Quoted 5 lines

Index: drivers/net/wireless/b43/dma.c
===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/dma.c

2013-04-09 17:33:54.325322046 +0200

+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/dma.c

You should make your patch against the newest git tree, so usually wireless-testing or Rafa?'s b43-next.

Quoted 31 lines
2013-04-09 17:34:43.473565843 +0200
(quoted diff) +15 -0
@@ -1688,6 +1688,21 @@
         b43_poison_rx_buffer(ring, skb);
         sync_descbuffer_for_device(ring, dmaaddr, ring->rx_buffersize);
 }
+void b43_dma_rx_discard(struct b43_dmaring *ring)
+{
+   B43_WARN_ON(ring->tx);
+
+    /* Device has filled all buffers, drop all packets in buffers
+     * and let TCP decrease speed.
+     * Set index to one desc after the last one (which is marked)
+     * so the device will see all slots as free again
+     */
+    /*
+     *TODO: How to increase rx_drop in mac80211
+     */
+   b43_dma_write(ring, B43_DMA32_RXINDEX, ring->nr_slots *
+                                     sizeof(struct b43_dmadesc32));
+}
 
 void b43_dma_rx(struct b43_dmaring *ring)
 {

Index: drivers/net/wireless/b43/dma.h
===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/dma.h

2013-04-09 17:34:09.561397686 +0200
+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/dma.h
2013-04-09 17:34:43.461565780 +0200

Review line 13

Quoted 5 lines
(quoted diff) +1 -1
@@ -10,7 +10,8 @@
 #define B43_DMAIRQ_FATALMASK    ((1 << 10) | (1 << 11) | (1 << 12) \
                                          | (1 << 14) | (1 << 15))
 #define B43_DMAIRQ_NONFATALMASK (1 << 13)
-#define B43_DMAIRQ_RX_DONE              (1 << 16)
+#define B43_DMAIRQ_RX_DONE      (1 << 16)

Unnecessary whitespace change.

Review line 14

Quoted 1 line
(quoted diff) +1 -0
+#define B43_DMAIRQ_RDESC_UFLOW  (1 << 13)

Why not just rename B43_DMAIRQ_NONFATALMASK ?

Quoted 21 lines
(quoted diff) +2 -0
 /*** 32-bit DMA Engine. ***/
 
@@ -295,6 +296,8 @@
 void b43_dma_handle_txstatus(struct b43_wldev *dev,
                              const struct b43_txstatus *status);
 
+void b43_dma_rx_discard(struct b43_dmaring *ring);
+
 void b43_dma_rx(struct b43_dmaring *ring);
 
 void b43_dma_direct_fifo_rx(struct b43_wldev *dev,

Index: drivers/net/wireless/b43/main.c
===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/main.c

2013-04-09 17:33:54.325322046 +0200
+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/main.c
2013-04-09 18:08:01.011471215 +0200
(quoted diff) +0 -0
@@ -1894,14 +1894,6 @@
                         b43_controller_restart(dev, "DMA error");

Review line 1904

Quoted 10 lines
(quoted diff) +0 -8
                         return;
                 }
-                if (merged_dma_reason & B43_DMAIRQ_NONFATALMASK) {
-                        b43err(dev->wl, "DMA error: "
-                               "0x%08X, 0x%08X, 0x%08X, "
-                               "0x%08X, 0x%08X, 0x%08X\n",
-                               dma_reason[0], dma_reason[1],
-                               dma_reason[2], dma_reason[3],
-                               dma_reason[4], dma_reason[5]);
-                }

You should say why you remove this one, and you should remove the B43_DMAIRQ_NONFATALMASK from the if (unlikely()) enclosing it.

Review line 1919

Quoted 12 lines
(quoted diff) +4 -0
         }
 
         if (unlikely(reason & B43_IRQ_UCODE_DEBUG))
@@ -1920,6 +1912,14 @@
                 handle_irq_noise(dev);
 
         /* Check the DMA reason registers for received data. */
+        if (dma_reason[0] & B43_DMAIRQ_RDESC_UFLOW) {
+           //only print 256 time to not flood log
+           if(!(dev->stats.rxdesc_underruns++&0xFF)){
+                        b43warn(dev->wl, "Rx descriptor underrun
 (high cpu load?), throwing packets\n");

Apart from the style problems, maybe doing a

	if (printk_ratelimit())
		b43warn(...);

which will ensure that it won't flood the log, but still won't stop reporting them after the 256th time.

Quoted 32 lines
(quoted diff) +4 -0
+           }
+           b43_dma_rx_discard(dev->dma.rx_ring);
+
+        }
         if (dma_reason[0] & B43_DMAIRQ_RX_DONE) {
                 if (b43_using_pio_transfers(dev))
                         b43_pio_rx(dev->pio.rx_queue);
(quoted diff) +2 -2
@@ -1977,7 +1977,7 @@
                 return IRQ_NONE;
 
         dev->dma_reason[0] = b43_read32(dev, B43_MMIO_DMA0_REASON)
-            & 0x0001DC00;
+            & 0x0001FC00;
         dev->dma_reason[1] = b43_read32(dev, B43_MMIO_DMA1_REASON)
             & 0x0000DC00;
         dev->dma_reason[2] = b43_read32(dev, B43_MMIO_DMA2_REASON)
@@ -3081,7 +3081,7 @@
                 b43_write32(dev, 0x018C, 0x02000000);
         }
         b43_write32(dev, B43_MMIO_GEN_IRQ_REASON, 0x00004000);
-        b43_write32(dev, B43_MMIO_DMA0_IRQ_MASK, 0x0001DC00);
+        b43_write32(dev, B43_MMIO_DMA0_IRQ_MASK, 0x0001FC00);
         b43_write32(dev, B43_MMIO_DMA1_IRQ_MASK, 0x0000DC00);
         b43_write32(dev, B43_MMIO_DMA2_IRQ_MASK, 0x0000DC00);
         b43_write32(dev, B43_MMIO_DMA3_IRQ_MASK, 0x0001DC00);

===================================================================
--- compat-wireless-2012-09-07.orig/drivers/net/wireless/b43/b43.h

2013-04-09 17:04:51.552680190 +0200
+++ compat-wireless-2012-09-07/drivers/net/wireless/b43/b43.h
2013-04-09 17:46:08.472962612 +0200

Review line 677

Quoted 7 lines
(quoted diff) +1 -0
@@ -671,6 +671,7 @@
 
 struct b43_stats {
         u8 link_noise;
+        u32 rxdesc_underruns;
 };
 
 struct b43_key {

Jonas

#11 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 145 lines · Thommy

Hi,

new try sending in the patch (with pine instead of gmail webclient this time). Patch created on latest wireless-testing

BR,
Thommy Jakobsson

---
  drivers/net/wireless/b43/b43.h  |    1 +
  drivers/net/wireless/b43/dma.c  |   16 +++++++++++++++
  drivers/net/wireless/b43/dma.h  |    4 +++-
  drivers/net/wireless/b43/main.c |   43 ++++++++++++++++-----------------------
  4 files changed, 38 insertions(+), 26 deletions(-)
drivers/net/wireless/b43/b43.h +1 -0
index 7f3d461..fdcd00f 100644
@@ -680,6 +680,7 @@ struct b43_noise_calculation {
 
  struct b43_stats {
  	u8 link_noise;
+	u32 rxdesc_underruns;
  };
 
  struct b43_key {
drivers/net/wireless/b43/dma.c +16 -0
index 1221469..2ae00dd 100644
@@ -1733,6 +1733,22 @@ drop_recycle_buffer:
  	sync_descbuffer_for_device(ring, dmaaddr, ring->rx_buffersize);
  }
 
+void b43_dma_rx_discard(struct b43_dmaring *ring)
+{
+	B43_WARN_ON(ring->tx);
+
+	/* Device has filled all buffers, drop all packets in buffers
+	* and let TCP decrease speed.
+	* Set index to one desc after the last one
+	* so the device will see all slots as free again
+	*/
+	/*
+	*TODO: How to increase rx_drop in mac80211?
+	*/
+	b43_dma_write(ring, B43_DMA32_RXINDEX, ring->nr_slots *
+			sizeof(struct b43_dmadesc32));
+}
+
  void b43_dma_rx(struct b43_dmaring *ring)
  {
  	const struct b43_dma_ops *ops = ring->ops;
drivers/net/wireless/b43/dma.h +3 -1
index 9fdd198..fed8163 100644
@@ -9,7 +9,7 @@
  /* DMA-Interrupt reasons. */
  #define B43_DMAIRQ_FATALMASK	((1 << 10) | (1 << 11) | (1 << 12) \
  					 | (1 << 14) | (1 << 15))
-#define B43_DMAIRQ_NONFATALMASK	(1 << 13)
+#define B43_DMAIRQ_RDESC_UFLOW		(1 << 13)
  #define B43_DMAIRQ_RX_DONE		(1 << 16)
 
  /*** 32-bit DMA Engine. ***/
@@ -295,6 +295,8 @@ int b43_dma_tx(struct b43_wldev *dev,
  void b43_dma_handle_txstatus(struct b43_wldev *dev,
  			     const struct b43_txstatus *status);
 
+void b43_dma_rx_discard(struct b43_dmaring *ring);
+
  void b43_dma_rx(struct b43_dmaring *ring);
 
  void b43_dma_direct_fifo_rx(struct b43_wldev *dev,
drivers/net/wireless/b43/main.c +17 -24
index d377f77..277fe75 100644
@@ -1902,30 +1902,19 @@ static void b43_do_interrupt_thread(struct b43_wldev *dev)
  		}
  	}
 
-	if (unlikely(merged_dma_reason & (B43_DMAIRQ_FATALMASK |
-					  B43_DMAIRQ_NONFATALMASK))) {
-		if (merged_dma_reason & B43_DMAIRQ_FATALMASK) {
-			b43err(dev->wl, "Fatal DMA error: "
-			       "0x%08X, 0x%08X, 0x%08X, "
-			       "0x%08X, 0x%08X, 0x%08X\n",
-			       dma_reason[0], dma_reason[1],
-			       dma_reason[2], dma_reason[3],
-			       dma_reason[4], dma_reason[5]);
-			b43err(dev->wl, "This device does not support DMA "
+	if (unlikely(merged_dma_reason & (B43_DMAIRQ_FATALMASK))) {
+		b43err(dev->wl, "Fatal DMA error: "
+		       "0x%08X, 0x%08X, 0x%08X, "
+		       "0x%08X, 0x%08X, 0x%08X\n",
+		       dma_reason[0], dma_reason[1],
+		       dma_reason[2], dma_reason[3],
+		       dma_reason[4], dma_reason[5]);
+		b43err(dev->wl, "This device does not support DMA "
  			       "on your system. It will now be switched to PIO.\n");
-			/* Fall back to PIO transfers if we get fatal DMA errors! */
-			dev->use_pio = true;
-			b43_controller_restart(dev, "DMA error");
-			return;
-		}
-		if (merged_dma_reason & B43_DMAIRQ_NONFATALMASK) {
-			b43err(dev->wl, "DMA error: "
-			       "0x%08X, 0x%08X, 0x%08X, "
-			       "0x%08X, 0x%08X, 0x%08X\n",
-			       dma_reason[0], dma_reason[1],
-			       dma_reason[2], dma_reason[3],
-			       dma_reason[4], dma_reason[5]);
-		}
+		/* Fall back to PIO transfers if we get fatal DMA errors! */
+		dev->use_pio = true;
+		b43_controller_restart(dev, "DMA error");
+		return;
  	}
 
  	if (unlikely(reason & B43_IRQ_UCODE_DEBUG))
@@ -1944,6 +1933,10 @@ static void b43_do_interrupt_thread(struct b43_wldev *dev)
  		handle_irq_noise(dev);
 
  	/* Check the DMA reason registers for received data. */
+	if (dma_reason[0] & B43_DMAIRQ_RDESC_UFLOW) {
+		b43warn(dev->wl, "RX descriptor underrun, dropping packets\n");
+		b43_dma_rx_discard(dev->dma.rx_ring);
+	}
  	if (dma_reason[0] & B43_DMAIRQ_RX_DONE) {
  		if (b43_using_pio_transfers(dev))
  			b43_pio_rx(dev->pio.rx_queue);
@@ -2001,7 +1994,7 @@ static irqreturn_t b43_do_interrupt(struct b43_wldev *dev)
  		return IRQ_NONE;
 
  	dev->dma_reason[0] = b43_read32(dev, B43_MMIO_DMA0_REASON)
-	    & 0x0001DC00;
+	    & 0x0001FC00;
  	dev->dma_reason[1] = b43_read32(dev, B43_MMIO_DMA1_REASON)
  	    & 0x0000DC00;
  	dev->dma_reason[2] = b43_read32(dev, B43_MMIO_DMA2_REASON)

drivers/net/wireless/b43/main.c static int b43_chip_init(struct b43_wldev *dev) · line 3129

Quoted 8 lines
drivers/net/wireless/b43/main.c +1 -1
@@ -3130,7 +3123,7 @@ static int b43_chip_init(struct b43_wldev *dev)
  		b43_write32(dev, 0x018C, 0x02000000);
  	}
  	b43_write32(dev, B43_MMIO_GEN_IRQ_REASON, 0x00004000);
-	b43_write32(dev, B43_MMIO_DMA0_IRQ_MASK, 0x0001DC00);
+	b43_write32(dev, B43_MMIO_DMA0_IRQ_MASK, 0x0001FC00);
  	b43_write32(dev, B43_MMIO_DMA1_IRQ_MASK, 0x0000DC00);
  	b43_write32(dev, B43_MMIO_DMA2_IRQ_MASK, 0x0000DC00);
  	b43_write32(dev, B43_MMIO_DMA3_IRQ_MASK, 0x0001DC00);

Thommy,

This version did apply cleanly; however, there are still a couple of things that need fixing.

With scripts/checkpatch.pl, there is a warning that a quoted string is split over two lines. Although testing for that condition is a recent addition to the script, it is probably best to honor that requirement with new code. To do so, the lines starting at 1907 in main.c should be

                 b43err(dev->wl,
                        "Fatal DMA error: 0x%08X, 0x%08X, 0x%08X, 0x%08X, 

0x%08X, 0x%08X\n",

My mailer will likely wrap the string, but ignore that. Yes, that line is longer than 80 characters, but that is OK for quoted strings. The idea is that if they are not split, it is easier to grep for them. In this case, that is not an issue, but I prefer to make all new code be free of errors, warnings, and checks.

Your subject and commit message are not appropriate. Have you read Documentation/SubmittingPatches? The info you have at the start would not be appropriate for the permanent kernel record. If you want to add such informal remarks, include a line with --- after the formal commit message, and place informal remarks below that line. Everything is available at patch submission time, but the scripts strip it off before it is included in the git repositories. The b43-dev list is OK for discussions of the patch, but the final patch needs to be sent to John Linville <[email protected]>, with Cc: to b43-dev, and the linux-wireless mailing list <[email protected]>.

When the patch is appropriate for backporting to earlier, stable kernels, and this one certainly fits that criterion, follow the "Signed-off-by" line with one that says "Cc: Stable <[email protected]>". That way, the patch will automatically be included in earlier versions as soon as it is incorporated in the mainline version (Linus's tree).

Lastly, I have a question about the patch. Is it necessary to log every underrun? Perhaps you might want to place the b43_warn statement inside an "if (B43_DEBUG)" conditional. That way the condition will be logged for people really interested in debugging the driver, but the general user of a distro kernel will never see the message.

Thanks for finding this issue.

Larry

#12 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 43 lines · Larry Finger

Thommy,

This version did apply cleanly; however, there are still a couple of things that need fixing.

With scripts/checkpatch.pl, there is a warning that a quoted string is split over two lines. Although testing for that condition is a recent addition to the script, it is probably best to honor that requirement with new code. To do so, the lines starting at 1907 in main.c should be

                b43err(dev->wl,
                       "Fatal DMA error: 0x%08X, 0x%08X, 0x%08X, 0x%08X,

0x%08X, 0x%08X\n",

My mailer will likely wrap the string, but ignore that. Yes, that line is longer than 80 characters, but that is OK for quoted strings. The idea is that if they are not split, it is easier to grep for them. In this case, that is not an issue, but I prefer to make all new code be free of errors, warnings, and checks.

Your subject and commit message are not appropriate. Have you read Documentation/SubmittingPatches? The info you have at the start would not be appropriate for the permanent kernel record. If you want to add such informal remarks, include a line with --- after the formal commit message, and place informal remarks below that line. Everything is available at patch submission time, but the scripts strip it off before it is included in the git repositories. The b43-dev list is OK for discussions of the patch, but the final patch needs to be sent to John Linville <[email protected]>, with
Cc: to b43-dev, and the linux-wireless mailing list <[email protected]>.

When the patch is appropriate for backporting to earlier, stable kernels, and this one certainly fits that criterion, follow the "Signed-off-by" line with one that says "Cc: Stable <[email protected]>". That way, the patch will automatically be included in earlier versions as soon as it is incorporated in the mainline version (Linus's tree).

Lastly, I have a question about the patch. Is it necessary to log every underrun? Perhaps you might want to place the b43_warn statement inside an "if (B43_DEBUG)" conditional. That way the condition will be logged for people really interested in debugging the driver, but the general user of a distro kernel will never see the message.

Quoted 5 lines

Thanks for finding this issue.

Larry

New try, think I have fixed everything but have probably missed something. If not Piotras or someone else find a better way I planned to send this in to John.

Subject:[PATCH] B43: Handle DMA RX descriptor underrun
From: Thommy Jakobsson <[email protected]>

Add handling of rx descriptor underflow. This fixes a fault that could happen on slow machines, where data is received faster than the CPU can handle. In such a case the device will use up all rx descriptors and refuse to send any more data before confirming that it is ok. This patch enables necessary interrupt to discover such a situation and will handle them by dropping everything in the ring buffer.

---
 drivers/net/wireless/b43/b43.h  |    1 +
 drivers/net/wireless/b43/dma.c  |   15 ++++++++++++++
 drivers/net/wireless/b43/dma.h  |    4 +++-
 drivers/net/wireless/b43/main.c |   43 ++++++++++++++++-----------------------
 4 files changed, 37 insertions(+), 26 deletions(-)
drivers/net/wireless/b43/b43.h +1 -0
index 7f3d461..fdcd00f 100644
@@ -680,6 +680,7 @@ struct b43_noise_calculation {
 
 struct b43_stats {
 	u8 link_noise;
+	u32 rxdesc_underruns;
 };
 
 struct b43_key {
drivers/net/wireless/b43/dma.c +15 -0
index 1221469..df929b3 100644
@@ -1733,6 +1733,21 @@ drop_recycle_buffer:
 	sync_descbuffer_for_device(ring, dmaaddr, ring->rx_buffersize);
 }
 
+void b43_dma_rx_discard(struct b43_dmaring *ring)
+{
+	B43_WARN_ON(ring->tx);
+
+	/* Device has filled all buffers, drop all packets in buffers
+	* and let TCP decrease speed.
+	* Set index to one desc after the last one
+	* so the device will see all slots as free again
+	*/
+	/*
+	*TODO: How to increase rx_drop in mac80211?
+	*/
+	ring->ops->set_current_rxslot(ring, ring->nr_slots);
+}
+
 void b43_dma_rx(struct b43_dmaring *ring)
 {
 	const struct b43_dma_ops *ops = ring->ops;
drivers/net/wireless/b43/dma.h +3 -1
index 9fdd198..fed8163 100644
@@ -9,7 +9,7 @@
 /* DMA-Interrupt reasons. */
 #define B43_DMAIRQ_FATALMASK	((1 << 10) | (1 << 11) | (1 << 12) \
 					 | (1 << 14) | (1 << 15))
-#define B43_DMAIRQ_NONFATALMASK	(1 << 13)
+#define B43_DMAIRQ_RDESC_UFLOW		(1 << 13)
 #define B43_DMAIRQ_RX_DONE		(1 << 16)
 
 /*** 32-bit DMA Engine. ***/
@@ -295,6 +295,8 @@ int b43_dma_tx(struct b43_wldev *dev,
 void b43_dma_handle_txstatus(struct b43_wldev *dev,
 			     const struct b43_txstatus *status);
 
+void b43_dma_rx_discard(struct b43_dmaring *ring);
+
 void b43_dma_rx(struct b43_dmaring *ring);
 
 void b43_dma_direct_fifo_rx(struct b43_wldev *dev,
drivers/net/wireless/b43/main.c +18 -25

68 lines not on this page · raw message

Signature · 1 line
1.7.9.5
#13 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

In reply to Thommy
Quoted 5 lines · On Sat, 20 Apr 2013 16:14:09 +0200 (CEST) Thommy <[email protected]> wrote:
drivers/net/wireless/b43/dma.c +5 -0
+void b43_dma_rx_discard(struct b43_dmaring *ring)
+{
+	B43_WARN_ON(ring->tx);
+
+	/* Device has filled all buffers, drop all packets in buffers

drivers/net/wireless/b43/dma.c drop_recycle_buffer: · line 1750

Quoted 10 lines
drivers/net/wireless/b43/dma.c +10 -0
+	* and let TCP decrease speed.
+	* Set index to one desc after the last one
+	* so the device will see all slots as free again
+	*/
+	/*
+	*TODO: How to increase rx_drop in mac80211?
+	*/
+	b43_dma_write(ring, B43_DMA32_RXINDEX, ring->nr_slots *
+			sizeof(struct b43_dmadesc32));
+}

You need to check whether this is a 32bit or 64bit DMA engine and write to B43_DMA32_RXINDEX or B43_DMA64_RXINDEX.
You could simply use the ring->ops->set_current_rxslot() dmaop for that.

And I think you need to set ring->current_slot to the value of ring->ops->get_current_slot(). Not 100% sure, though, since it is years since I worked on that code.

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130420/8697d8d4/attachment.sig>
#14 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 14 lines · Michael B?sch
Quoted 11 lines · On Sat, 20 Apr 2013 16:14:09 +0200 (CEST) Thommy <[email protected]> wrote:
drivers/net/wireless/b43/dma.c +11 -0
+void b43_dma_rx_discard(struct b43_dmaring *ring)
+{
+	B43_WARN_ON(ring->tx);
+
+	/* Device has filled all buffers, drop all packets in buffers
+	* and let TCP decrease speed.
+	* Set index to one desc after the last one
+	* so the device will see all slots as free again
+	*/
+	/*
+	*TODO: How to increase rx_drop in mac80211?
Quoted 9 lines

drivers/net/wireless/b43/dma.c drop_recycle_buffer: · line 1750

Quoted 4 lines
drivers/net/wireless/b43/dma.c +4 -0
+	*/
+	b43_dma_write(ring, B43_DMA32_RXINDEX, ring->nr_slots *
+			sizeof(struct b43_dmadesc32));
+}

You need to check whether this is a 32bit or 64bit DMA engine and write to B43_DMA32_RXINDEX or B43_DMA64_RXINDEX.
You could simply use the ring->ops->set_current_rxslot() dmaop for that.

Good point, I'll fix that

Quoted 3 lines

And I think you need to set ring->current_slot to the value of ring->ops->get_current_slot(). Not 100% sure, though, since it is years since I worked on that code.

Is that really needed? They should already be the same since the device has thrown the underflow interrupt, or do I miss some race condition?

//thommy

#15 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 5 lines · On Sat, 20 Apr 2013 21:47:10 +0200 (CEST) Thommy Jakobsson <[email protected]>…
Quoted 3 lines

And I think you need to set ring->current_slot to the value of ring->ops->get_current_slot(). Not 100% sure, though, since it is years since I worked on that code.

Is that really needed? They should already be the same since the device has thrown the underflow interrupt

Hm not sure. But should be easy to verify.

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130420/333eb764/attachment.sig>
#16 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 10 lines · Michael B?sch
Quoted 5 lines · On Sat, 20 Apr 2013 21:47:10 +0200 (CEST) Thommy Jakobsson <[email protected]>…
Quoted 3 lines

And I think you need to set ring->current_slot to the value of ring->ops->get_current_slot(). Not 100% sure, though, since it is years since I worked on that code.

Is that really needed? They should already be the same since the device has thrown the underflow interrupt

Hm not sure. But should be easy to verify.

I did check it in the beginning when I was testing the patch, just printed the indexes in the kernel log. But of course if there is a race condition it could be hard to catch. Have used the patch for several weeks and several 100Gb now, so it does seem to work.

The reason for getting this interrupt is that the current descriptor pointer (RXDPTR) has the same value as the descriptor index (RXINDEX). The descriptor index is set at the end of b43_dma_rx with ops->set_current_rxslot. At the next line, and with the same value, the ring->current slot is set. So in my opinion I don't need to set it. But I could have missed something.

BR,
Thommy

#17 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 10 lines · On Sat, 20 Apr 2013 22:16:26 +0200 (CEST) Thommy Jakobsson <[email protected]>…
Quoted 6 lines · Michael B?sch
Quoted 3 lines · On Sat, 20 Apr 2013 21:47:10 +0200 (CEST) Thommy Jakobsson <[email protected]>…
Quoted 3 lines

And I think you need to set ring->current_slot to the value of ring->ops->get_current_slot(). Not 100% sure, though, since it is years since I worked on that code.

Quoted 7 lines
Quoted 4 lines
Quoted 2 lines

Is that really needed? They should already be the same since the device has thrown the underflow interrupt

Hm not sure. But should be easy to verify.

I did check it in the beginning when I was testing the patch, just printed the indexes in the kernel log. But of course if there is a race condition it could be hard to catch.

Ok. I don't think there should be a race.
Thanks for checking.

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130420/cefb6f3a/attachment-0001.sig>
#18 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

In reply to Thommy Jakobsson

(resending with some corrections after original email bounced)

Hi all,

I tested original version of this patch and it happen to fix the issue on Linksys WRT54GL (based on Broadcom 5352 SoC). However I think that the bug is still present.

After closer look it seems that DMA buffers may be used incorrectly in b43_dma_rx. I don't have documentation for Broadcom WLAN chipsets, but DMA engine used by for BCM440X apparently has similar design. Looking at "RECEIVE DESCRIPTOR TABLE POINTER REGISTER"
(page 35 of http://www.broadcom.com/collateral/pg/440X-PG02-R.pdf),
B43_DMA32_RXINDEX should be left pointing after last valid descriptor as initialized in dmacontroller_setup. Notice that the last valid descriptor already has B43_DMA32_DCTL_DTABLEEND set, so DMA should operate in a round-robin fashion.

B43_DMA32_RXINDEX is reprogrammed from b43_dma_rx, by calling ops->set_current_rxslot(ring, slot). The patch from Thommy simply reprogrammes the register back with the same value as initialized in dmacontroller_setup.

I suspect that with set_current_rxslot call removed, the patch from Thommy will not be necessary anymore. I'm currently testing such fix and will report back in a few days.

Regards,

Piotr

Quoted 33 lines · Thommy Jakobsson
Quoted 10 lines · Michael B?sch
Quoted 5 lines · On Sat, 20 Apr 2013 21:47:10 +0200 (CEST) Thommy Jakobsson <[email protected]>…
Quoted 3 lines

And I think you need to set ring->current_slot to the value of ring->ops->get_current_slot(). Not 100% sure, though, since it is years since I worked on that code.

Is that really needed? They should already be the same since the device has thrown the underflow interrupt

Hm not sure. But should be easy to verify.

I did check it in the beginning when I was testing the patch, just printed the indexes in the kernel log. But of course if there is a race condition it could be hard to catch. Have used the patch for several weeks and several 100Gb now, so it does seem to work.

The reason for getting this interrupt is that the current descriptor pointer (RXDPTR) has the same value as the descriptor index (RXINDEX). The descriptor index is set at the end of b43_dma_rx with ops->set_current_rxslot. At the next line, and with the same value, the ring->current slot is set. So in my opinion I don't need to set it. But I could have missed something.

BR,
Thommy
_______________________________________________
b43-dev mailing list
b43-dev at lists.infradead.org
http://lists.infradead.org/mailman/listinfo/b43-dev

A message that is not in the archive (sent off-list or not yet received): CAFzhf4rSUO9mn_RKw4Vehw2LKggFdC_W0v-K_KKCZJez1AwynQ@mail.gmail.com

#20 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 20 lines · Piotras

Hi all,

I tested original version of this patch and it happen to fix the issue on Linksys WRT54GL (based on Broadcom 5352 SoC). However I think that the original cause is still present.

After closer look it seems that DMA buffers may be used incorrectly in b43_dma_rx. I don't have documentation for Broadcom WLAN chipsets, but DMA engine used by for BCM440X apparently has similar design. Looking at "RECEIVE DESCRIPTOR TABLE POINTER REGISTER"
(page 35 of http://www.broadcom.com/collateral/pg/440X-PG02-R.pdf),
B43_DMA32_RXINDEX should be left pointing after last valid descriptor as initialized in dmacontroller_setup (the referred descriptor already has B43_DMA32_DCTL_DTABLEEND set).

B43_DMA32_RXINDEX is reprogrammed from b43_dma_rx, by calling ops->set_current_rxslot(ring, slot). The patch from Thommy simply reprogramms the register back to the value initialized used in dmacontroller_setup.

Quoted 9 lines

I suspect that with set_current_rxslot call removed, the patch from Thommy will not be necessary anymore. I'm currently testing such fix and will report back in a few days.

Regards,

Piotr

Hi,

yes I also saw that datasheet and funny enough I also reacted on the same thing. But if you keep the rxindex pointing after the last descriptor all the time, how can you know if you had an underflow? Of course you could argue that it doesn't matter, the device will just keep on sending data since the current descriptor index will never reach the descriptor index. But that also means that you could start reading a packet at the same time as the device is overwriting it with an other, right? I didn't read it to carefully but the datasheet doesn't prohibit a change of the rxpointer, does it?

//Thommy

#21 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 45 lines · Thommy Jakobsson
Quoted 20 lines · Piotras

Hi all,

I tested original version of this patch and it happen to fix the issue on Linksys WRT54GL (based on Broadcom 5352 SoC). However I think that the original cause is still present.

After closer look it seems that DMA buffers may be used incorrectly in b43_dma_rx. I don't have documentation for Broadcom WLAN chipsets, but DMA engine used by for BCM440X apparently has similar design. Looking at "RECEIVE DESCRIPTOR TABLE POINTER REGISTER"
(page 35 of http://www.broadcom.com/collateral/pg/440X-PG02-R.pdf),
B43_DMA32_RXINDEX should be left pointing after last valid descriptor as initialized in dmacontroller_setup (the referred descriptor already has B43_DMA32_DCTL_DTABLEEND set).

B43_DMA32_RXINDEX is reprogrammed from b43_dma_rx, by calling ops->set_current_rxslot(ring, slot). The patch from Thommy simply reprogramms the register back to the value initialized used in dmacontroller_setup.

Quoted 9 lines

I suspect that with set_current_rxslot call removed, the patch from Thommy will not be necessary anymore. I'm currently testing such fix and will report back in a few days.

Regards,

Piotr

Hi,

yes I also saw that datasheet and funny enough I also reacted on the same thing. But if you keep the rxindex pointing after the last descriptor all the time, how can you know if you had an underflow? Of course you could argue that it doesn't matter, the device will just keep on sending data since the current descriptor index will never reach the descriptor index. But that also means that you could start reading a packet at the same time as the device is overwriting it with an other, right? I didn't read it to carefully but the datasheet doesn't prohibit a change of the rxpointer, does it?

Quoted 2 lines

//Thommy

Sorry I mean that the datasheet doesn't prohibit a change of the rx descriptor index, does it? Where the rx descriptor index in the reversed broadcom documentation and code would be translated to the LastDscr in the BCM440X datasheet (even has the same offset).

//Thommy

#22 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 49 lines · Thommy Jakobsson
Quoted 45 lines · Thommy Jakobsson
Quoted 20 lines · Piotras

Hi all,

I tested original version of this patch and it happen to fix the issue on Linksys WRT54GL (based on Broadcom 5352 SoC). However I think that the original cause is still present.

After closer look it seems that DMA buffers may be used incorrectly in b43_dma_rx. I don't have documentation for Broadcom WLAN chipsets, but DMA engine used by for BCM440X apparently has similar design. Looking at "RECEIVE DESCRIPTOR TABLE POINTER REGISTER"
(page 35 of http://www.broadcom.com/collateral/pg/440X-PG02-R.pdf),
B43_DMA32_RXINDEX should be left pointing after last valid descriptor as initialized in dmacontroller_setup (the referred descriptor already has B43_DMA32_DCTL_DTABLEEND set).

B43_DMA32_RXINDEX is reprogrammed from b43_dma_rx, by calling ops->set_current_rxslot(ring, slot). The patch from Thommy simply reprogramms the register back to the value initialized used in dmacontroller_setup.

Quoted 9 lines

I suspect that with set_current_rxslot call removed, the patch from Thommy will not be necessary anymore. I'm currently testing such fix and will report back in a few days.

Regards,

Piotr

Hi,

yes I also saw that datasheet and funny enough I also reacted on the same thing. But if you keep the rxindex pointing after the last descriptor all the time, how can you know if you had an underflow? Of course you could argue that it doesn't matter, the device will just keep on sending data since the current descriptor index will never reach the descriptor index. But that also means that you could start reading a packet at the same time as the device is overwriting it with an other, right? I didn't read it to carefully but the datasheet doesn't prohibit a change of the rxpointer, does it?

Quoted 9 lines
Quoted 2 lines

//Thommy

Sorry I mean that the datasheet doesn't prohibit a change of the rx descriptor index, does it? Where the rx descriptor index in the reversed broadcom documentation and code would be translated to the LastDscr in the BCM440X datasheet (even has the same offset).

//Thommy

Looking a bit more carefully in the datasheet I find

"Software makes the new descriptors available for use by updating the LastDscr field of the XmtPtr register (see ?Receive Descriptor Table Pointer Register (RcvPtr, offset 0x218)? on page 35) to point to the entry after the new last valid descriptor." on page 40

Isn't that exactly what we do in b43_dma_rx? The initial rx descriptor index is just to make the device to start. Theoretically you could get into a fault situation if the device succeds in using up all descriptors between b43_dma_rx_discard (added by my patch) and the call to b43_dma_rx. Because then you could start reading the same packet as the device is writing to.

//Thommy

#23 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 6 lines · On Sat, 20 Apr 2013 23:23:54 +0200 (CEST) Thommy Jakobsson <[email protected]>…

Isn't that exactly what we do in b43_dma_rx? The initial rx descriptor index is just to make the device to start. Theoretically you could get into a fault situation if the device succeds in using up all descriptors between b43_dma_rx_discard (added by my patch) and the call to b43_dma_rx. Because then you could start reading the same packet as the device is writing to.

Yes this is true.
And thus I'm currently unsure why we need this patch at all. _Why_ does the DMA stall as soon as the ring is filled up?

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130421/201e5823/attachment.sig>
#24 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 10 lines · Michael B?sch
Quoted 6 lines · On Sat, 20 Apr 2013 23:23:54 +0200 (CEST) Thommy Jakobsson <[email protected]>…

Isn't that exactly what we do in b43_dma_rx? The initial rx descriptor index is just to make the device to start. Theoretically you could get into a fault situation if the device succeds in using up all descriptors between b43_dma_rx_discard (added by my patch) and the call to b43_dma_rx. Because then you could start reading the same packet as the device is writing to.

Quoted 3 lines

Yes this is true.
And thus I'm currently unsure why we need this patch at all. _Why_ does the DMA stall as soon as the ring is filled up?

The device stops sending when it hits the rx index. Thats what rx index is, the address after the last slot the device is allowed to use. In the driver it is used as the last slot handled, so if the device would continue it will overwrite slots that hasn't been read. So you need to handle that in someway. One way is using my patch, another is to never let it hit rx index (as Piotras is working on).

This is how I see the problem:
1. the driver get an interrupt that there is data in the ring, the interrupt thread is scheduled
2. the interrupt thread (in b43_dma_rx) checks how many slots that the device has filled
3. it then loops through them and sends the packets in the ring to the kernel (dma_rx). It also allocates new skb_buf:s
4. after all slots has been handled the rx index is updated
5. In the meantime the device continues to pump data in to the ring
6. If the cpu is slow, and we don't have that many slots, there is a chance that the device have time to loop through all and hitting rx index. If we don't update the rx index in such a situation the device will never continue. Since it has hit the end of available descriptors, it doesn't have much of a choice, where would it write next?

Normally the device never hit the rx index, and in such case it will just use the next descriptor in the table. If it finds a descriptor with END set, it will jump back to the beginning. Forming the ring

//Thomy

#25 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 6 lines · On Sun, 21 Apr 2013 10:22:51 +0200 (CEST) Thommy Jakobsson <[email protected]>…

The device stops sending when it hits the rx index. Thats what rx index is, the address after the last slot the device is allowed to use. In the driver it is used as the last slot handled, so if the device would continue it will overwrite slots that hasn't been read. So you need to handle that in someway. One way is using my patch, another is to never let it hit rx index (as Piotras is working on).

Yeah I get that. But why do we need to handle that by putting the stop index right beyond the ring? That requires us to get another RX interrupt before the ring fills up again. Why can't we set it so that the ring is fully writable again, but not set it beyond the ring boundary? I guess that is decrementing it by one, perhaps, honoring 0-index-wrapover.

Quoted 3 lines

Normally the device never hit the rx index, and in such case it will just use the next descriptor in the table. If it finds a descriptor with END set, it will jump back to the beginning. Forming the ring

I know how that basic stuff works, because I wrote most of the code ;D

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130421/59279114/attachment.sig>
#26 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 10 lines · Michael B?sch
Quoted 6 lines · On Sun, 21 Apr 2013 10:22:51 +0200 (CEST) Thommy Jakobsson <[email protected]>…

The device stops sending when it hits the rx index. Thats what rx index is, the address after the last slot the device is allowed to use. In the driver it is used as the last slot handled, so if the device would continue it will overwrite slots that hasn't been read. So you need to handle that in someway. One way is using my patch, another is to never let it hit rx index (as Piotras is working on).

Quoted 6 lines

Yeah I get that. But why do we need to handle that by putting the stop index right beyond the ring? That requires us to get another RX interrupt before the ring fills up again. Why can't we set it so that the ring is fully writable again, but not set it beyond the ring boundary? I guess that is decrementing it by one, perhaps, honoring 0-index-wrapover.

All times that I had extra debugging on I always got a normal rx interrupt at the same time as the underflow interrupt. So the b43_dma_rx is called just after the b43_rx_discard, and the rx index is updated again to something better. Writing the rx index to index-1 was my initial plan, the only reason why I chose beyond the ring boundary was because thats how it is initialised. I never tried index-1, so I don't know if it works though. It did try to write the same value back again, but of course that didn't work.

Why is it set to to beyond the ring when it is initilised?

Quoted 5 lines
Quoted 3 lines

Normally the device never hit the rx index, and in such case it will just use the next descriptor in the table. If it finds a descriptor with END set, it will jump back to the beginning. Forming the ring

I know how that basic stuff works, because I wrote most of the code ;D

and yet you didn't fix this problem... =) Just kidding, to be honest I didn't understand what in my patch your question concerned. So I pretty much wrote down all I know about how it works (or how I think it works).

Thanks,
Thommy

#27 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 1 line · On Sun, 21 Apr 2013 11:01:11 +0200 (CEST) Thommy Jakobsson <[email protected]>…

Why is it set to to beyond the ring when it is initilised?

That's what broadcom does in their implementation. I guess that's the only reason.

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130421/a25c76e8/attachment.sig>
#28 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 7 lines · Michael B?sch
Quoted 1 line · On Sun, 21 Apr 2013 11:01:11 +0200 (CEST) Thommy Jakobsson <[email protected]>…

Why is it set to to beyond the ring when it is initilised?

That's what broadcom does in their implementation. I guess that's the only reason.

Check, I'll give index-1 a go. It should be quite obvious if it works or not.

//thommy

#29 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 7 lines · Thommy Jakobsson
Quoted 3 lines · Michael B?sch

On Sun, 21 Apr 2013 11:01:11 +0200 (CEST)
Thommy Jakobsson <[email protected]> wrote:

Quoted 6 lines
Quoted 4 lines
Quoted 1 line

Why is it set to to beyond the ring when it is initilised?

That's what broadcom does in their implementation. I guess that's the only reason.

Check, I'll give index-1 a go. It should be quite obvious if it works or not.

Thommy,

I had the same thought, but notice that you also need to update ring->current_slot in b43_dma_rx_discard.

When ring->current_slot update is missing (as in your original patch), you will introduce the same problem I hit by removing ops->set_current_rxslot(ring, slot) from b43_dma_rx (disabling RX underflow detection). Basically we break assumptions about ownership of descriptors (which descriptors are exclusively used by host and which can be written by DMA processor). So next time b43_dma_rx is called it may process descriptors that are updated by DMA processor at the same time (race condition).

I guess the following could be used to mark all descriptors for DMA use from RX underflow interrupt or whenever we decide that number of free descriptors is too low (not tested):

  slot = prev_slot(ring, ops->get_current_rxslot(ring));
  wmb();
  ops->set_current_rxslot(ring, slot);
  ring->current_slot = slot;

The above code simply puts back all the already filled descriptors back for DMA processor to be overwritten with new data.

I'm still concerned about FIFO overruns. When descriptor underflow condition occurs while CPU is busy with b43_dma_rx, we still wait for b43_dma_rx to finish processing all received packets before any descriptors become available for DMA to use. Depending on FIFO size and CPU load we may still hit fatal receive FIFO underflow error. This should cause driver to reset (possible explanation of why "discards" and "slot_usage" files disappeared when Richlv tested your patch in https://dev.openwrt.org/ticket/7552#comment:94).

I suspect that driver reset after FIFO underflow interrupt triggers is not handled correctly, but not sure why. Possibly this is not expected by hostapd, so it needs restart?

Not sure if FIFO overruns are really an issue and how to prevent them. Maybe RX underflow should be handled in top half and reprogram device to suspend receives (is this possible)? Or maybe we could clean and set B43_DMA32_RXENABLE bit in B43_DMA32_RXCTL to reset DMA channel without resetting the driver?

Ideas?

Piotr

Quoted 2 lines

//thommy

#30 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 12 lines · On Sun, 21 Apr 2013 15:46:03 +0100 Piotras <[email protected]> wrote:
Quoted 11 lines · Thommy Jakobsson
Quoted 7 lines · Michael B?sch
Quoted 1 line · On Sun, 21 Apr 2013 11:01:11 +0200 (CEST) Thommy Jakobsson <[email protected]>…

Why is it set to to beyond the ring when it is initilised?

That's what broadcom does in their implementation. I guess that's the only reason.

Quoted 7 lines
Quoted 2 lines

Check, I'll give index-1 a go. It should be quite obvious if it works or not.

Thommy,

I had the same thought, but notice that you also need to update ring->current_slot in b43_dma_rx_discard.

I don't think so.
We don't modify the descriptor pointer, but the stop index.

Quoted 7 lines

I guess the following could be used to mark all descriptors for DMA use from RX underflow interrupt or whenever we decide that number of free descriptors is too low (not tested):

  slot = prev_slot(ring, ops->get_current_rxslot(ring));
  wmb();
  ops->set_current_rxslot(ring, slot);

get_current_rxslot() reads a different register than set_current_rxslot().

get_current_rxslot() returns the descriptor pointer that the device is pointing to. set_current_rxslot() sets the _stop_ index.
Yes, at least the last one is misnamed.

Quoted 3 lines

Not sure if FIFO overruns are really an issue and how to prevent them. Maybe RX underflow should be handled in top half and reprogram device to suspend receives (is this possible)?

What is "top half"?

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130421/4c5e94bf/attachment.sig>
#31 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 18 lines · Michael B?sch
Quoted 15 lines · On Sun, 21 Apr 2013 15:46:03 +0100 Piotras <[email protected]> wrote:
Quoted 13 lines · Thommy Jakobsson
Quoted 7 lines · Michael B?sch
Quoted 1 line · On Sun, 21 Apr 2013 11:01:11 +0200 (CEST) Thommy Jakobsson <[email protected]>…

Why is it set to to beyond the ring when it is initilised?

That's what broadcom does in their implementation. I guess that's the only reason.

Check, I'll give index-1 a go. It should be quite obvious if it works or not.

Quoted 7 lines
Quoted 4 lines

Thommy,

I had the same thought, but notice that you also need to update ring->current_slot in b43_dma_rx_discard.

I don't think so.
We don't modify the descriptor pointer, but the stop index.

Yes.

My code was incorrect as I should ensure that ring->current_slot is equal ops->get_current_rxslot(ring). As we are in the RX underflow, B43_DMA32_RXSTATUS, B43_DMA32_RXINDEX and ring->current_slot should already point to the same descriptor, so only set_current_rxslot is needed.

Quoted 15 lines
Quoted 8 lines

I guess the following could be used to mark all descriptors for DMA use from RX underflow interrupt or whenever we decide that number of free descriptors is too low (not tested):

  slot = prev_slot(ring, ops->get_current_rxslot(ring));
  wmb();
  ops->set_current_rxslot(ring, slot);

get_current_rxslot() reads a different register than set_current_rxslot().

get_current_rxslot() returns the descriptor pointer that the device is pointing to. set_current_rxslot() sets the _stop_ index.
Yes, at least the last one is misnamed.

Quoted 5 lines
Quoted 3 lines

Not sure if FIFO overruns are really an issue and how to prevent them. Maybe RX underflow should be handled in top half and reprogram device to suspend receives (is this possible)?

What is "top half"?

Top half interrupt handler (b43_do_interrupt).

Quoted 3 lines

--
Michael

#32 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 7 lines · On Sun, 21 Apr 2013 16:24:18 +0100 Piotras <[email protected]> wrote:
Quoted 5 lines
Quoted 3 lines

Not sure if FIFO overruns are really an issue and how to prevent them. Maybe RX underflow should be handled in top half and reprogram device to suspend receives (is this possible)?

What is "top half"?

Top half interrupt handler (b43_do_interrupt).

No way.
Interrupts are properly blocked before leaving the top-half handler and re-enabled before leaving the bottom-half handler already. Adding stuff to the top-half handler would only introduce serious locking and latency issues and it would fix nothing.

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130421/36056889/attachment.sig>
#33 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

In reply to Piotras
Quoted 5 lines

I had the same thought, but notice that you also need to update ring->current_slot in b43_dma_rx_discard.

Quoted 8 lines

When ring->current_slot update is missing (as in your original patch), you will introduce the same problem I hit by removing ops->set_current_rxslot(ring, slot) from b43_dma_rx (disabling RX underflow detection). Basically we break assumptions about ownership of descriptors (which descriptors are exclusively used by host and which can be written by DMA processor). So next time b43_dma_rx is called it may process descriptors that are updated by DMA processor at the same time (race condition).

Why would I update ring->current_slot? That should point where we are in the ring from the drivers perspective. Since that i always updated together with the rx index (set_current_rxslot) it will always be the same as the current index in hw (get_current_rxslot) in the b43_rx_discard. That is also where it should be.

It is true that I'll introduce a possibility for ownership problem with the original patch, but that will only last until the next time b43_dma_rx is called. Which is just after the b43_dma_rx_discard has exited since you always will get a rx interrupt at the same time as the underflow interrupt. This is also the case at each startup since rx index is initilised to the same thing as I used.

I have tested the index-1 that me and michael talked about, from the initial test is working just fine. Im running with only 8 rxslots to try to find any corner cases, so I see a lot of traffic drops though. This would also fix the ownership problem.

BTW, it is confusing that set and get current_rxslot doesn't updated and read the same register.

About the FIFO errors I haven't looked at those at all. I did get some other errors when I pumped data at link speed into my router, but that was for TX.

//Thommy

#34 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 2 lines · On Sun, 21 Apr 2013 17:11:23 +0200 (CEST) Thommy Jakobsson <[email protected]>…

BTW, it is confusing that set and get current_rxslot doesn't updated and read the same register.

Yes the set_current_rxslot() op should be renamed to set_rx_stop_slot() or something like that.

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130421/ad24cd10/attachment.sig>
#35 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

In reply to Michael Büsch
Quoted 6 lines · Michael B?sch

Yeah I get that. But why do we need to handle that by putting the stop index right beyond the ring? That requires us to get another RX interrupt before the ring fills up again. Why can't we set it so that the ring is fully writable again, but not set it beyond the ring boundary? I guess that is decrementing it by one, perhaps, honoring 0-index-wrapover.

Hi again,
below you can find the patch that I'm currently trying out. The change is according to the discussion above. Instead of setting the rx index as the driver does during startup, it decrements it by one.

It has been running at ~1300KB/s in RX and 100KB/s in TX for the last couple of hours without problems. Approximately there have been one underflow per second. CPU load is constantly over 80%, pretty much just the interrupt thread that is using the CPU.

I can see that at the entry of b43_dma_discard all three indexes, ring->current_slot(ring), get_current_rxslot(ring) and b43_dma_read(ring,B43_DMA32_RXINDEX) are always the same. Of course the rx index is in bytes. So this is as expected.
B
I also has to back my statement that I always see a normal rx interrupt at the same time as the underflow interrupt. I have seen a couple of cases where that didn't happen. So this seems like a better way of doing things than setting rx index to max slots when having underflows. This way device and driver collisions are better avoided. Good idea Michael.

//Thommy

Subject:[PATCH] B43: Handle DMA RX descriptor underrun
From: Thommy Jakobsson <[email protected]>

Add handling of rx descriptor underflow. This fixes a fault that could happen on slow machines, where data is received faster than the CPU can handle. In such a case the device will use up all rx descriptors and refuse to send any more data before confirming that it is ok. This patch enables necessary interrupt to discover such a situation and will handle them by dropping everything in the ring buffer.

---
 drivers/net/wireless/b43/b43.h  |    1 +
 drivers/net/wireless/b43/dma.c  |   19 +++++++++++++++++
 drivers/net/wireless/b43/dma.h  |    4 +++-
 drivers/net/wireless/b43/main.c |   43 ++++++++++++++++-----------------------
 4 files changed, 41 insertions(+), 26 deletions(-)
drivers/net/wireless/b43/b43.h +1 -0
index 7f3d461..fdcd00f 100644
@@ -680,6 +680,7 @@ struct b43_noise_calculation {
 
 struct b43_stats {
 	u8 link_noise;
+	u32 rxdesc_underruns;
 };
 
 struct b43_key {
drivers/net/wireless/b43/dma.c +19 -0
index 1221469..139305e 100644
@@ -1733,6 +1733,25 @@ drop_recycle_buffer:
 	sync_descbuffer_for_device(ring, dmaaddr, ring->rx_buffersize);
 }
 
+void b43_dma_rx_discard(struct b43_dmaring *ring)
+{
+	int current_slot, previous_slot;
+
+	B43_WARN_ON(ring->tx);
+
+	/* Device has filled all buffers, drop all packets in buffers
+	* and let TCP decrease speed.
+	* Set index to one desc after the last one
+	* so the device will see all slots as free again
+	*/
+	/*
+	*TODO: How to increase rx_drop in mac80211?
+	*/
+	current_slot = ring->ops->get_current_rxslot(ring);
+	previous_slot = prev_slot(ring, current_slot);
+	ring->ops->set_current_rxslot(ring, previous_slot);
+}
+
 void b43_dma_rx(struct b43_dmaring *ring)
 {
 	const struct b43_dma_ops *ops = ring->ops;
drivers/net/wireless/b43/dma.h +3 -1
index 9fdd198..fed8163 100644
@@ -9,7 +9,7 @@
 /* DMA-Interrupt reasons. */
 #define B43_DMAIRQ_FATALMASK	((1 << 10) | (1 << 11) | (1 << 12) \
 					 | (1 << 14) | (1 << 15))
-#define B43_DMAIRQ_NONFATALMASK	(1 << 13)
+#define B43_DMAIRQ_RDESC_UFLOW		(1 << 13)
 #define B43_DMAIRQ_RX_DONE		(1 << 16)
 
 /*** 32-bit DMA Engine. ***/
@@ -295,6 +295,8 @@ int b43_dma_tx(struct b43_wldev *dev,
 void b43_dma_handle_txstatus(struct b43_wldev *dev,
 			     const struct b43_txstatus *status);
 
+void b43_dma_rx_discard(struct b43_dmaring *ring);
+
 void b43_dma_rx(struct b43_dmaring *ring);
 
 void b43_dma_direct_fifo_rx(struct b43_wldev *dev,
drivers/net/wireless/b43/main.c +18 -25

68 lines not on this page · raw message

Signature · 1 line
1.7.9.5
#36 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 7 lines · On Sun, 21 Apr 2013 20:07:39 +0200 (CEST) Thommy Jakobsson <[email protected]>…
drivers/net/wireless/b43/dma.c +7 -0
+void b43_dma_rx_discard(struct b43_dmaring *ring)
+{
+	int current_slot, previous_slot;
+
+	B43_WARN_ON(ring->tx);
+
+	/* Device has filled all buffers, drop all packets in buffers

drivers/net/wireless/b43/dma.c drop_recycle_buffer: · line 1752

Quoted 10 lines
drivers/net/wireless/b43/dma.c +10 -0
+	* and let TCP decrease speed.
+	* Set index to one desc after the last one
+	* so the device will see all slots as free again
+	*/
+	/*
+	*TODO: How to increase rx_drop in mac80211?
+	*/
+	current_slot = ring->ops->get_current_rxslot(ring);
+	previous_slot = prev_slot(ring, current_slot);
+	ring->ops->set_current_rxslot(ring, previous_slot);

Hmmm. While this does work (because the register and ring->current_slot contain the same value at this point), I'd prefer if you write ring->current_slot - 1
to the stop-index-register.

Also, the comment needs to be updated.

PS: I'll send a patch that renames set_current_rxslot() later, because it's really confusing.

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130421/30d76e76/attachment.sig>
#37 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

New response since my phone answered with html

Quoted 3 lines · Michael B?sch

Hmmm. While this does work (because the register and ring->current_slot contain the same value at this point), I'd prefer if you write ring->current_slot - 1
to the stop-index-register.

The reason for the interrupt is because the current index in the device is the same as the last index. That's why I read the value from the device. But I fix it, don't have a strong opinion about it.

Quoted 1 line

Also, the comment needs to be updated

Crap, missed that.

/thommy

A message that is not in the archive (sent off-list or not yet received): CAOTQ16LCWpAOndQsHBxN1-X_8SLpcU8V=mM41gvcNRocA=GhsA@mail.gmail.com

#39 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 24 lines · On Sun, 21 Apr 2013 20:39:49 +0200 Thommy Jakobsson <[email protected]> wrote:
Quoted 15 lines · Michael B?sch
Quoted 11 lines · On Sun, 21 Apr 2013 20:07:39 +0200 (CEST) Thommy Jakobsson <[email protected]>…
drivers/net/wireless/b43/dma.c +11 -0
+void b43_dma_rx_discard(struct b43_dmaring *ring)
+{
+     int current_slot, previous_slot;
+
+     B43_WARN_ON(ring->tx);
+
+     /* Device has filled all buffers, drop all packets in buffers
+     * and let TCP decrease speed.
+     * Set index to one desc after the last one
+     * so the device will see all slots as free again
+     */
Quoted 8 lines

drivers/net/wireless/b43/dma.c drop_recycle_buffer: · line 1752

Quoted 6 lines
drivers/net/wireless/b43/dma.c +6 -0
+     /*
+     *TODO: How to increase rx_drop in mac80211?
+     */
+     current_slot = ring->ops->get_current_rxslot(ring);
+     previous_slot = prev_slot(ring, current_slot);
+     ring->ops->set_current_rxslot(ring, previous_slot);

Hmmm. While this does work (because the register and ring->current_slot

Quoted 6 lines

contain the same

Quoted 2 lines

value at this point), I'd prefer if you write ring->current_slot - 1 to the stop-index-register.

The reason for the interrupt is because the current index in the device is the same as the last index. That's why I read the value from the device. But I fix it, don't have a strong opinion about it.

Hm, ok. Well, sounds plausible. So you may keep it as-is.

But you should probably rename the function to b43_dma_handle_rx_overflow(), because the discard does not work properly, if the function is called without an overflow in progress.

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130421/a18e0cde/attachment-0001.sig>
#40 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 7 lines

Hm, ok. Well, sounds plausible. So you may keep it as-is.

But you should probably rename the function to b43_dma_handle_rx_overflow(), because the discard does not work properly, if the function is called without an overflow in progress.

Check, better?

//Thommy

Subject:[PATCH] B43: Handle DMA RX descriptor underrun
From: Thommy Jakobsson <[email protected]>

Add handling of rx descriptor underflow. This fixes a fault that could happen on slow machines, where data is received faster than the CPU can handle. In such a case the device will use up all rx descriptors and refuse to send any more data before confirming that it is ok. This patch enables necessary interrupt to discover such a situation and will handle them by dropping everything in the ring buffer.

---
 drivers/net/wireless/b43/b43.h  |    1 +
 drivers/net/wireless/b43/dma.c  |   19 +++++++++++++++++
 drivers/net/wireless/b43/dma.h  |    4 +++-
 drivers/net/wireless/b43/main.c |   43 ++++++++++++++++-----------------------
 4 files changed, 41 insertions(+), 26 deletions(-)
drivers/net/wireless/b43/b43.h +1 -0
index 7f3d461..fdcd00f 100644
@@ -680,6 +680,7 @@ struct b43_noise_calculation {
 
 struct b43_stats {
 	u8 link_noise;
+	u32 rxdesc_underruns;
 };
 
 struct b43_key {
drivers/net/wireless/b43/dma.c +19 -0
index 1221469..84af9e1 100644
@@ -1733,6 +1733,25 @@ drop_recycle_buffer:
 	sync_descbuffer_for_device(ring, dmaaddr, ring->rx_buffersize);
 }
 
+void b43_dma_handle_rx_overflow(struct b43_dmaring *ring)
+{
+	int current_slot, previous_slot;
+
+	B43_WARN_ON(ring->tx);
+
+	/* Device has filled all buffers, drop all packets and let TCP
+	 * decrease speed.
+	 * Decrement RX index by one will let the device to see all slots
+	 * as free again
+	 */
+	/*
+	*TODO: How to increase rx_drop in mac80211?
+	*/
+	current_slot = ring->ops->get_current_rxslot(ring);
+	previous_slot = prev_slot(ring, current_slot);
+	ring->ops->set_current_rxslot(ring, previous_slot);
+}
+
 void b43_dma_rx(struct b43_dmaring *ring)
 {
 	const struct b43_dma_ops *ops = ring->ops;
drivers/net/wireless/b43/dma.h +3 -1
index 9fdd198..df8c8cd 100644
@@ -9,7 +9,7 @@
 /* DMA-Interrupt reasons. */
 #define B43_DMAIRQ_FATALMASK	((1 << 10) | (1 << 11) | (1 << 12) \
 					 | (1 << 14) | (1 << 15))
-#define B43_DMAIRQ_NONFATALMASK	(1 << 13)
+#define B43_DMAIRQ_RDESC_UFLOW		(1 << 13)
 #define B43_DMAIRQ_RX_DONE		(1 << 16)
 
 /*** 32-bit DMA Engine. ***/
@@ -295,6 +295,8 @@ int b43_dma_tx(struct b43_wldev *dev,
 void b43_dma_handle_txstatus(struct b43_wldev *dev,
 			     const struct b43_txstatus *status);
 
+void b43_dma_handle_rx_overflow(struct b43_dmaring *ring);
+
 void b43_dma_rx(struct b43_dmaring *ring);
 
 void b43_dma_direct_fifo_rx(struct b43_wldev *dev,
drivers/net/wireless/b43/main.c +18 -25

68 lines not on this page · raw message

Signature · 1 line
1.7.9.5
#41 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

drivers/net/wireless/b43/b43.h struct b43_noise_calculation { · line 683

Quoted 3 lines · On Sun, 21 Apr 2013 22:13:43 +0200 (CEST) Thommy Jakobsson <[email protected]>…

struct b43_stats {

drivers/net/wireless/b43/b43.h +1 -0
 	u8 link_noise;
+	u32 rxdesc_underruns;

Why do we add that? It's not incremented and used anywhere.

The rest looks good.

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130421/8eb7d860/attachment.sig>
#42 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 6 lines · Michael B?sch

Review

Quoted 3 lines

struct b43_stats {

(quoted diff) +1 -0
 	u8 link_noise;
+	u32 rxdesc_underruns;

Why do we add that? It's not incremented and used anywhere.

Sorry, seems like I dont do that of a good joob revewing my own patches. It slipped in from one of my debug builts. Should I send in this to John or do you think we have to test it further? It has been running during the entire night with high load on my device. So that seems to work, but that is only on one device of course.
//Thommy

Subject:[PATCH] B43: Handle DMA RX descriptor underrun
From: Thommy Jakobsson <[email protected]>

Add handling of rx descriptor underflow. This fixes a fault that could happen on slow machines, where data is received faster than the CPU can handle. In such a case the device will use up all rx descriptors and refuse to send any more data before confirming that it is ok. This patch enables necessary interrupt to discover such a situation and will handle them by dropping everything in the ring buffer.

---
 drivers/net/wireless/b43/dma.c  |   19 +++++++++++++++++
 drivers/net/wireless/b43/dma.h  |    4 +++-
 drivers/net/wireless/b43/main.c |   43 ++++++++++++++++-----------------------
 3 files changed, 40 insertions(+), 26 deletions(-)
drivers/net/wireless/b43/dma.c +19 -0
index 1221469..ee3d640 100644
@@ -1733,6 +1733,25 @@ drop_recycle_buffer:
 	sync_descbuffer_for_device(ring, dmaaddr, ring->rx_buffersize);
 }
 
+void b43_dma_handle_rx_overflow(struct b43_dmaring *ring)
+{
+	int current_slot, previous_slot;
+
+	B43_WARN_ON(ring->tx);
+
+	/* Device has filled all buffers, drop all packets and let TCP
+	 * decrease speed.
+	 * Decrement RX index by one will let the device to see all slots
+	 * as free again
+	 */
+	/*
+	*TODO: How to increase rx_drop in mac80211?
+	*/
+	current_slot = ring->ops->get_current_rxslot(ring);
+	previous_slot = prev_slot(ring, current_slot);
+	ring->ops->set_current_rxslot(ring, previous_slot);
+}
+
 void b43_dma_rx(struct b43_dmaring *ring)
 {
 	const struct b43_dma_ops *ops = ring->ops;
drivers/net/wireless/b43/dma.h +3 -1
index 9fdd198..df8c8cd 100644
@@ -9,7 +9,7 @@
 /* DMA-Interrupt reasons. */
 #define B43_DMAIRQ_FATALMASK	((1 << 10) | (1 << 11) | (1 << 12) \
 					 | (1 << 14) | (1 << 15))
-#define B43_DMAIRQ_NONFATALMASK	(1 << 13)
+#define B43_DMAIRQ_RDESC_UFLOW		(1 << 13)
 #define B43_DMAIRQ_RX_DONE		(1 << 16)
 
 /*** 32-bit DMA Engine. ***/
@@ -295,6 +295,8 @@ int b43_dma_tx(struct b43_wldev *dev,
 void b43_dma_handle_txstatus(struct b43_wldev *dev,
 			     const struct b43_txstatus *status);
 
+void b43_dma_handle_rx_overflow(struct b43_dmaring *ring);
+
 void b43_dma_rx(struct b43_dmaring *ring);
 
 void b43_dma_direct_fifo_rx(struct b43_wldev *dev,
drivers/net/wireless/b43/main.c +18 -25

68 lines not on this page · raw message

Signature · 1 line
1.7.9.5
#43 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 8 lines · On Mon, 22 Apr 2013 10:11:41 +0200 (CEST) Thommy Jakobsson <[email protected]>…
Quoted 6 lines · Michael B?sch

Review

Quoted 3 lines

struct b43_stats {

(quoted diff) +1 -0
 	u8 link_noise;
+	u32 rxdesc_underruns;

Why do we add that? It's not incremented and used anywhere.

Quoted 5 lines

Sorry, seems like I dont do that of a good joob revewing my own patches. It slipped in from one of my debug builts. Should I send in this to John or do you think we have to test it further? It has been running during the entire night with high load on my device. So that seems to work, but that is only on one device of course.

Please remove rxdesc_underruns and send the patch (properly signed-off) to John Linville. Cc the linux-wireless mailinglist, b43-dev mailinglist and probably everybody involved in the review discussion.

You can also add a

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130422/48db6429/attachment.sig>
#44 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 17 lines · On Mon, 22 Apr 2013 12:00:47 +0200 Michael B?sch <[email protected]> wrote:
Quoted 8 lines · On Mon, 22 Apr 2013 10:11:41 +0200 (CEST) Thommy Jakobsson <[email protected]>…
Quoted 6 lines · Michael B?sch

Review

Quoted 3 lines

struct b43_stats {

(quoted diff) +1 -0
 	u8 link_noise;
+	u32 rxdesc_underruns;

Why do we add that? It's not incremented and used anywhere.

Quoted 5 lines

Sorry, seems like I dont do that of a good joob revewing my own patches. It slipped in from one of my debug builts. Should I send in this to John or do you think we have to test it further? It has been running during the entire night with high load on my device. So that seems to work, but that is only on one device of course.

Quoted 3 lines

Please remove rxdesc_underruns and send the patch (properly signed-off) to John Linville. Cc the linux-wireless mailinglist, b43-dev mailinglist and probably everybody involved in the review discussion.

whoops I forgot. Please also add
Cc: stable at vger.kernel.org
to the signed-off list

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130423/e1c123eb/attachment.sig>
#45 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

In reply to Michael Büsch

Michael,

Based on BCM440X Programmer's Reference Guide that may or may not apply to DMA processor used in WLAN chipsets, B43_DMA32_RXSTATUS may be updated by hardware to point to the same descriptor as B43_DMA32_RXINDEX.

For DMA processor this means that no more free descriptors are available. However b43_dma_rx assumes no data needs to be processed (see how ring->current_slot is updated and used in loop condition on next interrupt). This could explain stalls.

I don't expect that changing loop condition is sufficient, as without dropping some packets we will not prevent FIFO overruns.

Regards,

Piotr

Quoted 16 lines · Michael B?sch
Quoted 6 lines · On Sat, 20 Apr 2013 23:23:54 +0200 (CEST) Thommy Jakobsson <[email protected]>…

Isn't that exactly what we do in b43_dma_rx? The initial rx descriptor index is just to make the device to start. Theoretically you could get into a fault situation if the device succeds in using up all descriptors between b43_dma_rx_discard (added by my patch) and the call to b43_dma_rx. Because then you could start reading the same packet as the device is writing to.

Yes this is true.
And thus I'm currently unsure why we need this patch at all. _Why_ does the DMA stall as soon as the ring is filled up?

--
Michael

#46 lore raw

PATCH 1/1 b43: use rx desc underrun interrupt

Quoted 9 lines · On Sun, 21 Apr 2013 15:27:49 +0100 Piotras <[email protected]> wrote:

Based on BCM440X Programmer's Reference Guide that may or may not apply to DMA processor used in WLAN chipsets, B43_DMA32_RXSTATUS may be updated by hardware to point to the same descriptor as B43_DMA32_RXINDEX.

For DMA processor this means that no more free descriptors are available. However b43_dma_rx assumes no data needs to be processed (see how ring->current_slot is updated and used in loop condition on next interrupt). This could explain stalls.

Yes this is why it stalls. But how to fix this without additional counting? I'd rather prefer wiping the whole ring in that case.

Signature · 8 lines
Michael
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 836 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/b43-dev/attachments/20130421/9341395e/attachment.sig>