On Mon, Mar 2, 2009 at 9:27 AM, saeed bishara <saeed.bishara@xxxxxxxxx> wrote: > Hi Dan, > The log below show a kernel crash when using the mv_xor dma engine > and enabling the NET_DMA, the test that I run is iperf as server. > after doing some experiences, I found that the reason is because this > engine doesn't support requests less than 16 bytes, the driver returns > NULL when the request len is too short. > is this the right thing do when the prepare function needs to reject a request? Yes, this is the right thing to do, but this particular backtrace is something different. In this case we are hitting the valid checks in dma_cache_maint. I suspect it is failing the dma_map_single call as it is unlikely we have a bad struct page here, but your function offsets do not correlate with my local build so it is hard to tell. This leaves a skb buffer that does not pass the virt_addr_valid checks? That seems unlikely as well since, as far as I understand, skb->data is always kmalloc'd? Can you determine what the failing address is and where it came from? NET_DMA=n is the default on ARM for a reason. The large overhead of get_user_pages makes NET_DMA a losing proposition versus cpu-copy. A cache coherent architecture is a minimum requirement to see a performance or cpu utilization gain with NET_DMA. Regardless of whether it provides a benefit, it should not be crashing, so let's dig a bit deeper. Thanks, Dan -- To unsubscribe from this list: send the line "unsubscribe linux-raid" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
