kernel/linux
0
Fork 0
mirror of https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/ synced 2026-10-16 22:39:48 +00:00
Code Issues Releases Packages Activity

xhci: Revert TD fragment hacks.

Browse source 
Hi Greg,
 
 Recently, we found that commit "35773dac5f "usb: xhci: Link TRB must not
 occur within a USB payload burst" causes a userspace regression.  It will
 cause larger transfers submitted through usbfs that would have succeeded
 on older kernels to be rejected.
 
 Commit 35773dac5f was designed to address an issue where an ASIX USB
 ethernet device would get wedged when it was connected to a 1.0 xHCI
 host.  Only this particular ethernet device was impacted, because only the
 ax88179_178a driver implemented scatter-gather in 3.12.  The xHCI driver
 doesn't currently support TD fragment rules, and commit 35773dac5f was a
 quick hack that partially implemented one of the rules.
 
 This is the third regression this patch has caused.  There's yet another
 quick hack to work around the issue, but I really want to support TD
 fragments properly, rather than hacking around it.  It will take us a
 kernel release or two to get it implemented, since it is a big
 architectural change.
 
 This patchset backs out commit 35773dac5f, and the two bug fix patches
 for it.  The first patch limits arbitrarily aligned scatter-gather under
 xHCI 1.0 hosts.
 
 As a result of this patchset:
 
 1. usb-storage and uas will still be able to use scatter-gather, since
    they submit max packet sized aligned transfers.
 2. usbfs will behave exactly as before, no more userspace regressions.
 3. The ax88179_178a driver works fine without scatter-gather (Mark Lord
    confirms this).
 
 Users of the ASIX chipset may still see occasional packet loss on 1.0 xHCI
 hosts, if the xHCI driver needs to split a TRB across 64-KB boundaries,
 and a link TRB happens to fall between those two TRBs.  I expect this
 corner case to be infrequent.
 
 Sarah Sharp
 -----BEGIN PGP SIGNATURE-----
 Version: GnuPG v1.4.14 (GNU/Linux)
 
 iQIcBAABAgAGBQJS+lcgAAoJEBMGWMLi1Gc5dt4P/1M4HtDKavU4ReJROonILiM3
 eYcSVJPkYm6KeZx3YtNhzuHE4CEbMwCSmu4qrvffM87Ya8hckrS5iafciCJlwQpB
 AEZCEeuD/ACHF/F/BKnwMkZJCIIdjhdklqxUqUCR0tKC2BwUB6+JWrig6tHcGglV
 y9C1FLI910hq1UGxvF0fzK6/DpEEAIs4UzysdZ4B0u++cOI0u4cQmFLMwrY/CoqY
 gokzHQOALwcnbmbT8+LcjFwumGPAFsTKkbPyw+PWroiPoGlqcFQZErfSaSwojo4B
 ZQj3zWvYeO6plHv8rdBk1Q5sr6F+h2Mf9j6YXMLOw/i1sHn+Zryonf6/350cKi2p
 TjmJfOkmtQpZ+leAC2LUaTEVTNUE8FUc4tLY3KPWhR3nDLcx73UL3Lmh3monhiIz
 nB52e4EjAS+XN5v/Ef1FbtEUBccHiaNWUK9v44oUAur/bRl5MkN7JkDc1eqih92d
 7gHtM3K9IXIzbLCOndnu8hDj/vlgw9kf7wSMC0W4+1vdRqK8PoB9GaBf4VrgWGnm
 toY26pG8DpRhZJdN0gVPA3YgoEcgucLB5XUHDxr3SRV3dM5zVHAz/1ZOC1phwO/Y
 0CiS/5hpXd0hVp8d93RWKGtp5O1RvEKtUnFN6dW5XYJtXxgLTldxR7N340Z5DOWy
 /c219n1V4XIIU2DsL1Qe
 =0yTh
 -----END PGP SIGNATURE-----

Merge tag 'for-usb-linus-2014-02-11' of git://git.kernel.org/pub/scm/linux/kernel/git/sarah/xhci into usb-linus

Sarah writes:

xhci: Revert TD fragment hacks.

Hi Greg,

Recently, we found that commit "35773dac5f "usb: xhci: Link TRB must not
occur within a USB payload burst" causes a userspace regression.  It will
cause larger transfers submitted through usbfs that would have succeeded
on older kernels to be rejected.

Commit 35773dac5f was designed to address an issue where an ASIX USB
ethernet device would get wedged when it was connected to a 1.0 xHCI
host.  Only this particular ethernet device was impacted, because only the
ax88179_178a driver implemented scatter-gather in 3.12.  The xHCI driver
doesn't currently support TD fragment rules, and commit 35773dac5f was a
quick hack that partially implemented one of the rules.

This is the third regression this patch has caused.  There's yet another
quick hack to work around the issue, but I really want to support TD
fragments properly, rather than hacking around it.  It will take us a
kernel release or two to get it implemented, since it is a big
architectural change.

This patchset backs out commit 35773dac5f, and the two bug fix patches
for it.  The first patch limits arbitrarily aligned scatter-gather under
xHCI 1.0 hosts.

As a result of this patchset:

1. usb-storage and uas will still be able to use scatter-gather, since
   they submit max packet sized aligned transfers.
2. usbfs will behave exactly as before, no more userspace regressions.
3. The ax88179_178a driver works fine without scatter-gather (Mark Lord
   confirms this).

Users of the ASIX chipset may still see occasional packet loss on 1.0 xHCI
hosts, if the xHCI driver needs to split a TRB across 64-KB boundaries,
and a link TRB happens to fall between those two TRBs.  I expect this
corner case to be infrequent.

Sarah Sharp
This commit is contained in:
Greg Kroah-Hartman 2014-02-11 15:23:53 -08:00
commit 2991942f56
4 changed files with 16 additions and 60 deletions
Download patch file Download diff file Expand all files Collapse all files