Slow USB3 drive (Intel amd64)

[Mark Carroll <mtbc%ixod.org@localhost>]

I am trying out NetBSD 8.0 (GENERIC) on an Intel Broadwell NUC5i3RYK
with USB3 ports on the front and back and finding them surprisingly slow
with an external USB3 HDD. dmesg reports,

xhci0 at pci0 dev 20 function 0: vendor 8086 product 9cb1 (rev. 0x03)
xhci0: interrupting at msi1 vec 0
xhci0: xHCI version 1.0
usb0 at xhci0: USB revision 3.0
usb1 at xhci0: USB revision 2.0

uhub0 at usb0: vendor 8086 (0x8086) xHCI Root Hub (0000), class 9/0, rev 1.00/1.00, addr 0
uhub1 at usb1: vendor 8086 (0x8086) xHCI Root Hub (0000), class 9/0, rev 2.00/1.00, addr 0

pcictl mentions,

000:20:0: Intel Core 5G (mobile) USB xHCI (USB serial bus, xHCI, revision 0x03)
000:20:0: 0x9cb18086 (0x0c033003)

and if I plug the Western Digital Elements drive in I get messages like,

umass0 at uhub0 port 1 configuration 1 interface 0
umass0: Western Digital (0x1058) Elements 25A2 (0x25a2), rev 3.00/10.19, addr 1
umass0: using SCSI over Bulk-Only
scsibus0 at umass0: 2 targets, 1 lun per target
sd0 at scsibus0 target 0 lun 0: <WD, Elements 25A2, 1019> disk fixed

usbdevs -d -v reports,

addr 0: super speed, self powered, config 1, xHCI Root Hub(0x0000), vendor 8086(0x8086), rev 1.00(0x0100)
  uhub0
 port 1 addr 1: super speed, power 224 mA, config 1, Elements 25A2(0x25a2), Western Digital(0x1058), rev 10.19(0x1019), serial 57584D3141353733345A415A
   umass0

I don't see anything that makes me worry that I won't get USB3 speeds
but,

# time dd if=/dev/sd0 of=/dev/null bs=1m count=1k
1024+0 records in
1024+0 records out
1073741824 bytes transferred in 393.518 secs (2728571 bytes/sec)
      399.26 real         0.00 user         1.22 sys

is rather slower than I had hoped (<3Mb/s reads?): at that speed it'll
take a long time to cover the 2TB of the disk!

I also tried a USB3 Seagate drive,
 port 1 addr 5: super speed, power 224 mA, config 1, Expansion(0x231a), Seagate(0x0bc2), rev 7.06(0x0706), serial NA829NLT
and got around the same data read rate.

From the internal SSD it gets 58684036 bytes/sec so at least the problem
seems limited to USB, ahcisata seems fine.

I am new to NetBSD: I am not sure what to expect and I am probably
missing some obvious diagnostics or perhaps even somehow missed some
helpful documentation page or past list thread. Should I expect better
USB mass storage performance from a Broadwell-based system? Is there
something I should read or try next to get it? Naturally I acknowledge
the caveat,

> xhci driver is at best experimental.

-- Mark