Bug#586269: More info and confirmation from: [E1000-devel] e1000e latency issues
Problem found in kernel 2.6.34 and 2.6.35
[e1000-devel, the Intel wired ethernet developers mailing list]
[E1000-devel] e1000e latency issues
https://sourceforge.net/mailarchive/forum.php?thread_name=8DD2590731AB5D4C9DBF71A877482A900164D9262F%40orsmsx509.amr.corp.intel.com&forum_name=e1000-devel
From: Allan, Bruce W <bruce.allan@in...> - 2010-08-10 02:03
>So, yes, according to the lspci output, ASPM L1 is enabled. Disabling
>ASPM config option in the kernel may not necessarily disable it in
>hardware unfortunately. We made a change to the driver to forcibly
>disable ASPM (L0s for standard frames and both L0s and L1 for jumbo
>frames) on the adapter you have - this change went into 2.6.34. We
>suspect your issue may be due to ASPM L1 which can be confirmed by
>disabling it by either 1) enabling jumbo frames on the adapter, or 2)
>forcing ASPM L1 disabled via setpci. If you don't have jumbo frames
>enabled for your network segment, I recommend trying the latter option
>as follows (assuming your adapter is still PCI bus/device/number
>02:00.0 as indicated in the lspci output you provided earlier):
>
>1) First check the hexadecimal value of the LnkCtl register -
># setpci -s 2:0.0 0xf0
>
>2) Disable ASPM (both L0s and L1) by zeroing out bits 0 and 1 in the
> value returned in the previous step. For example, if it returned
> 42 (hex 42, that is) -
>
># setpci -s 2:0.0 0xf0=0x40
>3) Confirm ASPM is disabled by checking the output from lspci again.
From: Allan, Bruce W <bruce.allan@in...> - 2010-08-10 14:55
>That's great news! Thanks for the confirmation. We'll fix the issue
>by forcibly disabling L1 in the driver for that adapter.
>
>The SourceForge driver and the latest in-kernel driver are very close
>to one another (except the SourceForge driver has a compatibility
>layer that allows it to run on older kernels). They should work about
>the same, but sometimes it might take a little longer for some fixes
>to get rolled up into the latest stable kernel(s). I need to get
>better at updating the version number in the in-kernel driver.
Reply to: