Hello, 16.08.2016, 13:56, Felipe Balbi kirjoitti: > Jussi Kivilinna <jussi.kivilinna@xxxxxxxxxxx> writes: >>> Jussi Kivilinna <jussi.kivilinna@xxxxxxxxxxx> writes: >>>> Enabling SG allows enabling GSO (generic segmentation offload) feature >>>> of linux networking layer. This increases TCP throughput with NCM >>>> on Cortex-A15+USB3380 based device from 300 Mbit/s to 1.1 Gbit/s. >>>> >>>> Signed-off-by: Jussi Kivilinna <jussi.kivilinna@xxxxxxxxxxx> >>> >>> this is AWESOME!! :-) But here's the thing, any chance we can build this >>> in u_ether.c ? Also, NETIF_F_SG should be conditional on >>> gadget->sg_supported so that we don't break UDCs that don't support >>> sglists. >>> >> >> Actually, no sglists are passed to UDC. Reason why this work >> with minimal changes for NCM is that NCM does tx buffering >> in its wrap function.. 'ncm_wrap_ntb' copies input skbuffs to >> larger skbuff, so enabling SG is only matter of changing that >> skbuff data copy from 'memcpy' to 'skb_copy_bits' (and changing >> CRC calculation work with skbuff fragments). Since NCM already >> does copying, SG can be enabled for NCM without extra overhead. > > aha, understood. Now what if we skip copying altogether? If we have an > sg and a UDC that supports sg (gadget->sg_supported = 1), then we can > avoid copying, right? Skip copying might difficult for NCM because buffering, but might work for other networking gadgets. > >> To see if NETIF_F_SG with skbuff copying made difference with >> other networking gadgets, I made quick test for RNDIS gadget to >> enable NETIF_F_SG by adding skb_linearize_cow call to >> 'rndis_add_header' wrap function. TCP transfer from device to >> host: >> without SG: 265 Mbit/s >> with SG: 326 Mbit/s >> >> So, adding NETIF_F_SG with skbuff linearization in u_ether >> could improve performance little bit. > > interesting. Does USB3380 support sglists? We could check how much more > we get if we skip copying altogether. SG support net2280.c does not appear to be complete, for example PIO-mode 'write_fifo' accesses (struct usb_request).buf directly. Data book reads that USB3380 has Scatter/Gather DMA mode support. However errata list has two "DMA channel may hang under specific conditions" problems for rev.AA with workaround: use single DMA mode instead. -Jussi -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
