On 03/27/2013 03:07 PM, Christian Klein wrote:
Hello Larry,
thanks, I'll revert the number of slots back to 64 and see if that helps.
How much memory does one slot take? May it be worthwile to decrease the
B43_TXRING_SLOTS value, too?
Each slot is only the size of an RX or TX descriptor, which is only 12 bytes.
The real use of memory comes because a buffer is allocated for every RX slot
when initializing. For TX, the buffers are attached to the slot when the data is
ready. Reducing the TX ring size will reduce the peak memory demand at
throughput expense as the queues will be stopped more.
The device has 4MB Flash and 16 MB Ram. Although this is not much, according
to the openwrt supported hardware wiki page there seem to be quite a few
routers with 16MB that run with kernel 2.6 and b43.
From what I found on the internet it seems that this router never really
worked with 2.6+ kernels, at least in AP mode. Most Info I saw blamed it on
the Broadcom 5354 SOC. One mail I found from 06/2010 suggested that the
LP-PHY of this SoC may not be fully supported.
(http://comments.gmane.org/gmane.linux.drivers.bcm54xx.devel/10671)
Until the RX ring size was increased, my LP PHY devices were not very reliable
as STAs. I never tried them as an AP.
Is there any way to find out what may cause the problems and if it really is
a memory problem? I saw the OOM-Killer only with the 666.2 firmware, for
the 410.2160 firmware it seemed to slow to a crawl and die without memory
troubles. What seemed strange to me in the dmesg output was
[ 240.560000] b43-phy0 debug: Adding Interface type 2
[ 240.568000] b43-phy0 debug: Removing Interface type 2
[ 240.568000] b43-phy0 debug: Wireless interface stopped
[...]
[ 247.184000] b43-phy0 debug: Adding Interface type 3
Why would the driver switch interface types here?
I do not know why it would change interface types.
One thing that might be useful in your code would be the patch to work around RX
buffer overflows. A copy is attached.
Larry
Index: wireless-testing-rebased/drivers/net/wireless/b43/dma.c
===================================================================
--- wireless-testing-rebased.orig/drivers/net/wireless/b43/dma.c
+++ wireless-testing-rebased/drivers/net/wireless/b43/dma.c
@@ -1692,6 +1692,50 @@ drop_recycle_buffer:
sync_descbuffer_for_device(ring, dmaaddr, ring->rx_buffersize);
}
+/* check for overflow of the RX descriptor ring. If found, reset the DMA
+ * controller and return true.
+ */
+static bool dma_rx_check_overflow(struct b43_dmaring *ring)
+{
+ if (ring->type == B43_DMA_64BIT) {
+ u64 state;
+ u64 rxctl;
+
+ state = b43_dma_read(ring, B43_DMA64_RXSTATUS) &
+ B43_DMA64_RXSTAT;
+ if (state != B43_DMA64_RXSTAT_IDLEWAIT)
+ return false;
+ rxctl = b43_dma_read(ring, B43_DMA64_RXCTL);
+ b43_dmacontroller_rx_reset(ring->dev, ring->mmio_base,
+ ring->type);
+
+ b43_dma_write(ring, B43_DMA64_RXCTL, rxctl);
+ b43_dma_write(ring, B43_DMA64_RXINDEX, ring->nr_slots *
+ sizeof(struct b43_dmadesc64));
+ } else {
+ u32 state;
+ u32 rxctl;
+
+ state = b43_dma_read(ring, B43_DMA32_RXSTATUS) &
+ B43_DMA32_RXSTAT;
+ if (state != B43_DMA32_RXSTAT_IDLEWAIT)
+ return false;
+
+ rxctl = b43_dma_read(ring, B43_DMA32_RXCTL);
+ b43_dmacontroller_rx_reset(ring->dev, ring->mmio_base,
+ ring->type);
+
+ b43_dma_write(ring, B43_DMA32_RXCTL, rxctl);
+ b43_dma_write(ring, B43_DMA32_RXINDEX, ring->nr_slots *
+ sizeof(struct b43_dmadesc32));
+ }
+ ring->current_slot = 0;
+
+ b43err(ring->dev->wl, "DMA RX reset due to overflow\n");
+
+ return true;
+}
+
void b43_dma_rx(struct b43_dmaring *ring)
{
const struct b43_dma_ops *ops = ring->ops;
@@ -1703,9 +1747,23 @@ void b43_dma_rx(struct b43_dmaring *ring
B43_WARN_ON(!(current_slot >= 0 && current_slot < ring->nr_slots));
slot = ring->current_slot;
- for (; slot != current_slot; slot = next_slot(ring, slot)) {
- dma_rx(ring, &slot);
- update_max_used_slots(ring, ++used_slots);
+
+ /* XXX: BRCM4318(?) dirty workaround:
+ * it seems sometimes the RX ring overflows due to interrupt
+ * latencies; particularly for systems with slow CPUs and tight
+ * memory constraints
+ */
+ if (slot == current_slot) {
+ /* Try to reset the RX channel, will cost us few lost frames,
+ * but will recover from an eternal stall
+ */
+ if (dma_rx_check_overflow(ring))
+ return; /* exit on overflow and reset */
+ }
+
+ for (; slot != current_slot; slot = next_slot(ring, slot)) {
+ dma_rx(ring, &slot);
+ update_max_used_slots(ring, ++used_slots);
}
wmb();
ops->set_current_rxslot(ring, slot);
_______________________________________________
b43-dev mailing list
b43-dev@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/b43-dev
