On Tue, 25 Sep 2012, Sarah Sharp wrote:
> Alan, I'm wondering if the xHCI ring expansion is causing issues with
> USB hard drives under xHCI. Testing with a Buffalo USB 3.0 hard drive
> with an NEC uPD720200 xHCI host, I see that the usb-storage and SCSI
> initialization produces I/O errors on random sectors in 3.4.0, 3.4.6,
> and 3.5.0. I can't get those errors to be reproduced in 3.3.1.
>
> The xHCI ring expansion was added in 3.4, and we changed the xHCI's
> sg_tablesize:
>
> int xhci_gen_setup(struct usb_hcd *hcd, xhci_get_quirks_t get_quirks)
> {
> ...
> /* Accept arbitrarily long scatter-gather lists */
> hcd->self.sg_tablesize = ~0;
>
> The usb-storage driver sets the tablesize thus:
>
> static unsigned int usb_stor_sg_tablesize(struct usb_interface *intf)
> {
> struct usb_device *usb_dev = interface_to_usbdev(intf);
>
> if (usb_dev->bus->sg_tablesize) {
> return usb_dev->bus->sg_tablesize;
> }
> return SG_ALL;
> }
>
> I notice that SG_ALL is set to SCSI_MAX_SG_SEGMENTS, which is only 128.
> Should we be passing an arbitrarily large number to the SCSI core?
Yes, there's no reason not to. The block layer will make sure that
each individual request has a sufficiently small number of segments.
> There's some wording in include/scsi/scsi.h about also limiting the
> number of chained sgs to 2048. I'm wondering if we're hitting some bugs
> in the SCSI layer because we're setting the sg_tablesize so high.
I doubt it. Anyway, this stuff is handled by the block layer now, not
the SCSI layer. If you look through drivers/scsi, you'll see that
SG_ALL is used only in various SCSI interface drivers, not in the core.
> Alternately, we could be hitting bugs in the USB 3.0 firmware when we
> attempt to issue a read or write that's too big. The read on Adrian's
> hard drive failed on a bigger read request (122880 bytes). It would be
> interesting to see if it works fine if the xHCI sg_tablesize is limited.
> I'm going to try that with my own drive on 3.5.4 and see if it helps.
There were examples in the earlier usbmon traces where 122880-byte
reads succeeded, for whatever that's worth...
I doubt very much that you are anywhere close to hitting that limit.
If a 120-KB transfer has more than 128 SG segments then on average each
segment would be under 1024 bytes, a lot smaller than a page, which
seems unlikely. I don't think I've ever seen a transfer needing more
than about 8 segments.
Alan Stern
--
To unsubscribe from this list: send the line "unsubscribe linux-scsi" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
