On Fri, 22 Aug 2014, Dan Williams wrote: > v1.0 hosts require that TD-fragments (portions of a TD that do not end > on a MPB boundary) not cross a TRB segment boundary. This constraint is > in addition to the constraint that a TRB may not specify a transfer that > crosses a 64K boundary. This enabling permits the driver to accept > scatterlists of nearly any geometry. "Nearly" because there is one > unlikely remaining degenerate case of a driver submitting a transfer > that consumes all the TRBs in a segment before hitting an MBP boundary. > That case is trapped and the transfer is rejected. That last part sounds problematic. The issue will arise at unpredictable times, depending on the lengths of the scatterlists that have been submitted in the past, so it's not easily reproducible. What is the caller supposed to do when this happens? As for the likelihood of this occurring... A ring segment is one page, so 4096 bytes. Each TRB is 16 bytes, so a segment can hold 256 TRBs. With a bulk maxpacket size of 1024 and a (typical?) maxburst size of 8, MPB boundaries occur every 8192 bytes. Therefore if a scatterlist contains more than 256 entries, with an average length < 32 bytes, it is likely trigger this condition (depending on the exact alignment with respect to the MPB boundary). I don't know exactly how unlikely such a situation is, but it's not hard to imagine a network packet composed of lots of little pieces. Instead of failing the submission, we ought to set up some sort of bounce buffers. I have no good suggestions on how to implement that, however. Alan Stern -- 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
