** Also affects: linux (Ubuntu Artful) Importance: Undecided Status: New
** Changed in: linux (Ubuntu) Status: Fix Released => Confirmed ** Summary changed: - PCIe: Allow drivers to use Relaxed Ordering on capable root ports + ixgbe/PCIe: Allow drivers to use Relaxed Ordering on capable root ports ** Changed in: linux (Ubuntu) Status: Confirmed => In Progress ** Changed in: linux (Ubuntu Artful) Status: New => In Progress ** Changed in: linux (Ubuntu Artful) Assignee: (unassigned) => dann frazier (dannf) ** Changed in: linux (Ubuntu) Assignee: (unassigned) => dann frazier (dannf) ** Changed in: linux (Ubuntu Zesty) Importance: Undecided => Medium ** Changed in: linux (Ubuntu Artful) Importance: Undecided => Medium ** Description changed: [Impact] A signficant performance gain can be achieved with the iperf benchmark by allowing the ixgbe driver to take advantage of PCIe Relaxed Ordering, as observed on the HiSilicon D05 system. [Test Case] = server = taskset -c 0-15 iperf -s -B 192.168.80.11 taskset -c 16-31 iperf -s -B 192.168.90.11 = client = taskset -c 0-15 iperf -c 192.168.80.11 -i 3 -t 600 -P 8 & taskset -c 16-31 iperf -c 192.168.90.11 -i 3 -t 600 -P 8 & Without these patches, we observe an aggregate performance of 8.54 Gbit/sec. With these patches, we observe an aggregate performance of 18.82 Gbit/sec. - [Regression Risk] The following risk assessment applies to both artful and zesty - but note that the bulk of these changes are *already in artful*. The proposed SRU for artful would just allow an additional driver to take advantage of it. For zesty, an SRU would need to both introduce this feature and enable it for both the cxgb4 and ixgbe drivers. The patchset is careful to only enable Relaxed Ordering if advertised up through the root port. However, it is possible that enabling RO on hardware that supports it could actually regress performance, or that there is hw out there that advertises RO, but has bugs in the implementation. The patch series includes quirks for HW where RO is advertised but known to have issues[*] - but this list maybe incomplete. This would only impact devices w/ drivers that take advantage of RO - currently cxgb4 and (not yet upstream) ixgbe. Also, there is a possibility of a crash caused by a coding error. One such issue has already been found and fixed upstream [**], and is part of this series. The risk of both of these regressions is reduced by the fact that this - code is now in an upstream release (v4.13) and has therefore gotten a - lot more testing that has not resulted in regressions reports. + code is upstream, and has therefore gotten a lot more testing that has + not resulted in regressions reports. [*] commit 077fa19c5dfa06a6ae04fb1661680940ff837612 Author: dingtianhong <dingtianh...@huawei.com> Date: Tue Aug 15 11:23:25 2017 +0800 PCI: Disable Relaxed Ordering Attributes for AMD A110 commit 87e09cdec4dae08acdb4aa49beb793c19d73e73e Author: dingtianhong <dingtianh...@huawei.com> Date: Tue Aug 15 11:23:24 2017 +0800 PCI: Disable Relaxed Ordering for some Intel processors [**] commit 0e405232871d67bf1b238d56b6b3d500e69ebbf3 Author: dingtianhong <dingtianh...@huawei.com> Date: Tue Aug 15 23:24:48 2017 +0800 PCI: fix oops when try to find Root Port for a PCI device ** Description changed: [Impact] A signficant performance gain can be achieved with the iperf benchmark by allowing the ixgbe driver to take advantage of PCIe Relaxed Ordering, as observed on the HiSilicon D05 system. [Test Case] + Setup two servers with dual ixgbe 10G ports, wired back-to-back. + = server = taskset -c 0-15 iperf -s -B 192.168.80.11 taskset -c 16-31 iperf -s -B 192.168.90.11 = client = taskset -c 0-15 iperf -c 192.168.80.11 -i 3 -t 600 -P 8 & taskset -c 16-31 iperf -c 192.168.90.11 -i 3 -t 600 -P 8 & Without these patches, we observe an aggregate performance of 8.54 Gbit/sec. With these patches, we observe an aggregate performance of 18.82 Gbit/sec. [Regression Risk] The following risk assessment applies to both artful and zesty - but note that the bulk of these changes are *already in artful*. The proposed SRU for artful would just allow an additional driver to take advantage of it. For zesty, an SRU would need to both introduce this feature and enable it for both the cxgb4 and ixgbe drivers. The patchset is careful to only enable Relaxed Ordering if advertised up through the root port. However, it is possible that enabling RO on hardware that supports it could actually regress performance, or that there is hw out there that advertises RO, but has bugs in the implementation. The patch series includes quirks for HW where RO is advertised but known to have issues[*] - but this list maybe incomplete. This would only impact devices w/ drivers that take advantage of RO - currently cxgb4 and (not yet upstream) ixgbe. Also, there is a possibility of a crash caused by a coding error. One such issue has already been found and fixed upstream [**], and is part of this series. The risk of both of these regressions is reduced by the fact that this code is upstream, and has therefore gotten a lot more testing that has not resulted in regressions reports. [*] commit 077fa19c5dfa06a6ae04fb1661680940ff837612 Author: dingtianhong <dingtianh...@huawei.com> Date: Tue Aug 15 11:23:25 2017 +0800 PCI: Disable Relaxed Ordering Attributes for AMD A110 commit 87e09cdec4dae08acdb4aa49beb793c19d73e73e Author: dingtianhong <dingtianh...@huawei.com> Date: Tue Aug 15 11:23:24 2017 +0800 PCI: Disable Relaxed Ordering for some Intel processors [**] commit 0e405232871d67bf1b238d56b6b3d500e69ebbf3 Author: dingtianhong <dingtianh...@huawei.com> Date: Tue Aug 15 23:24:48 2017 +0800 PCI: fix oops when try to find Root Port for a PCI device ** Description changed: [Impact] A signficant performance gain can be achieved with the iperf benchmark by allowing the ixgbe driver to take advantage of PCIe Relaxed Ordering, as observed on the HiSilicon D05 system. [Test Case] - Setup two servers with dual ixgbe 10G ports, wired back-to-back. + Setup two servers with dual-port PCIe ixgbe cards, connected back-to-back. = server = taskset -c 0-15 iperf -s -B 192.168.80.11 taskset -c 16-31 iperf -s -B 192.168.90.11 = client = taskset -c 0-15 iperf -c 192.168.80.11 -i 3 -t 600 -P 8 & taskset -c 16-31 iperf -c 192.168.90.11 -i 3 -t 600 -P 8 & Without these patches, we observe an aggregate performance of 8.54 Gbit/sec. With these patches, we observe an aggregate performance of 18.82 Gbit/sec. [Regression Risk] The following risk assessment applies to both artful and zesty - but note that the bulk of these changes are *already in artful*. The proposed SRU for artful would just allow an additional driver to take advantage of it. For zesty, an SRU would need to both introduce this feature and enable it for both the cxgb4 and ixgbe drivers. The patchset is careful to only enable Relaxed Ordering if advertised up through the root port. However, it is possible that enabling RO on hardware that supports it could actually regress performance, or that there is hw out there that advertises RO, but has bugs in the implementation. The patch series includes quirks for HW where RO is advertised but known to have issues[*] - but this list maybe incomplete. This would only impact devices w/ drivers that take advantage of RO - currently cxgb4 and (not yet upstream) ixgbe. Also, there is a possibility of a crash caused by a coding error. One such issue has already been found and fixed upstream [**], and is part of this series. The risk of both of these regressions is reduced by the fact that this code is upstream, and has therefore gotten a lot more testing that has not resulted in regressions reports. [*] commit 077fa19c5dfa06a6ae04fb1661680940ff837612 Author: dingtianhong <dingtianh...@huawei.com> Date: Tue Aug 15 11:23:25 2017 +0800 PCI: Disable Relaxed Ordering Attributes for AMD A110 commit 87e09cdec4dae08acdb4aa49beb793c19d73e73e Author: dingtianhong <dingtianh...@huawei.com> Date: Tue Aug 15 11:23:24 2017 +0800 PCI: Disable Relaxed Ordering for some Intel processors [**] commit 0e405232871d67bf1b238d56b6b3d500e69ebbf3 Author: dingtianhong <dingtianh...@huawei.com> Date: Tue Aug 15 23:24:48 2017 +0800 PCI: fix oops when try to find Root Port for a PCI device -- You received this bug notification because you are a member of Ubuntu Bugs, which is subscribed to Ubuntu. https://bugs.launchpad.net/bugs/1721365 Title: ixgbe/PCIe: Allow drivers to use Relaxed Ordering on capable root ports To manage notifications about this bug go to: https://bugs.launchpad.net/ubuntu/+source/linux/+bug/1721365/+subscriptions -- ubuntu-bugs mailing list ubuntu-bugs@lists.ubuntu.com https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs