| Message ID | 20230727184600.26768-1-longman@redhat.com |
|---|---|
| Headers | show |
| Series | x86/speculation: Disable IBRS when idle | expand |
On 7/27/23 14:45, Waiman Long wrote: > v6: > - Fix allyesconfig build error by moving __update_spec_ctrl() > helper from nospec-branch.h to spec-ctrl.h and include it in files > that need the helper. > > v5: > - Update comment in patch 1. > - Minor doc update and code twist in patch 4 as suggested by Peter and > Randy. > > v4: > - Add a new __update_spec_ctrl() helper in patch 1. > - Rebased to the latest linux kernel. > > v3: > - Drop patches 1 ("x86/speculation: Provide a debugfs file to dump > SPEC_CTRL MSRs") and 5 ("x86/idle: Disable IBRS entering mwait idle > and enable it on wakeup") for now. > - Drop the MSR restoration code in ("x86/idle: Disable IBRS when cpu > is offline") as native_play_dead() does not return. > - For patch ("intel_idle: Add ibrs_off module parameter to force > disable IBRS"), change the name from "no_ibrs" to "ibrs_off" and > document the new parameter in intel_idle.rst. > > For Intel processors that need to turn on IBRS to protect against > Spectre v2 and Retbleed, the IBRS bit in the SPEC_CTRL MSR affects > the performance of the whole core even if only one thread is turning > it on when running in the kernel. For user space heavy applications, > the performance impact of occasionally turning IBRS on during syscalls > shouldn't be significant. Unfortunately, that is not the case when the > sibling thread is idling in the kernel. In that case, the performance > impact can be significant. > > When DPDK is running on an isolated CPU thread processing network packets > in user space while its sibling thread is idle. The performance of the > busy DPDK thread with IBRS on and off in the sibling idle thread are: > > IBRS on IBRS off > ------- -------- > packets/second: 7.8M 10.4M > avg tsc cycles/packet: 282.26 209.86 > > This is a 25% performance degradation. The test system is a Intel Xeon > 4114 CPU @ 2.20GHz. > > Commit bf5835bcdb96 ("intel_idle: Disable IBRS during long idle") > disables IBRS when the CPU enters long idle (C6 or below). However, there > are existing users out there who have set "intel_idle.max_cstate=1" > to decrease latency. Those users won't be able to benefit from this > commit. This patch series extends this commit by providing a new > "intel_idle.ibrs_off" module parameter to force disable IBRS even when > "intel_idle.max_cstate=1" at the expense of increased IRQ response > latency. It also includes a commit to allow the disabling of IBRS when > a CPU becomes offline. > > Waiman Long (4): > x86/speculation: Add __update_spec_ctrl() helper > x86/idle: Disable IBRS when cpu is offline > intel_idle: Use __update_spec_ctrl() in intel_idle_ibrs() > intel_idle: Add ibrs_off module parameter to force disable IBRS > > Documentation/admin-guide/pm/intel_idle.rst | 17 ++++++++++++++++- > arch/x86/include/asm/spec-ctrl.h | 11 +++++++++++ > arch/x86/kernel/smpboot.c | 8 ++++++++ > drivers/idle/intel_idle.c | 18 +++++++++++++----- > 4 files changed, 48 insertions(+), 6 deletions(-) > Peter, Is this patch series good enough to be merged? Thanks, Longman
* Waiman Long <longman@redhat.com> wrote: > For Intel processors that need to turn on IBRS to protect against > Spectre v2 and Retbleed, the IBRS bit in the SPEC_CTRL MSR affects > the performance of the whole core even if only one thread is turning > it on when running in the kernel. For user space heavy applications, > the performance impact of occasionally turning IBRS on during syscalls > shouldn't be significant. Unfortunately, that is not the case when the > sibling thread is idling in the kernel. In that case, the performance > impact can be significant. > > When DPDK is running on an isolated CPU thread processing network packets > in user space while its sibling thread is idle. The performance of the > busy DPDK thread with IBRS on and off in the sibling idle thread are: > > IBRS on IBRS off > ------- -------- > packets/second: 7.8M 10.4M > avg tsc cycles/packet: 282.26 209.86 > > This is a 25% performance degradation. The test system is a Intel Xeon > 4114 CPU @ 2.20GHz. Ok, that's a solid improvement, and the feature has no obvious downsides, so I've applied your series to tip:sched/core with a few edits here and there. Thanks! Ingo
v6: - Fix allyesconfig build error by moving __update_spec_ctrl() helper from nospec-branch.h to spec-ctrl.h and include it in files that need the helper. v5: - Update comment in patch 1. - Minor doc update and code twist in patch 4 as suggested by Peter and Randy. v4: - Add a new __update_spec_ctrl() helper in patch 1. - Rebased to the latest linux kernel. v3: - Drop patches 1 ("x86/speculation: Provide a debugfs file to dump SPEC_CTRL MSRs") and 5 ("x86/idle: Disable IBRS entering mwait idle and enable it on wakeup") for now. - Drop the MSR restoration code in ("x86/idle: Disable IBRS when cpu is offline") as native_play_dead() does not return. - For patch ("intel_idle: Add ibrs_off module parameter to force disable IBRS"), change the name from "no_ibrs" to "ibrs_off" and document the new parameter in intel_idle.rst. For Intel processors that need to turn on IBRS to protect against Spectre v2 and Retbleed, the IBRS bit in the SPEC_CTRL MSR affects the performance of the whole core even if only one thread is turning it on when running in the kernel. For user space heavy applications, the performance impact of occasionally turning IBRS on during syscalls shouldn't be significant. Unfortunately, that is not the case when the sibling thread is idling in the kernel. In that case, the performance impact can be significant. When DPDK is running on an isolated CPU thread processing network packets in user space while its sibling thread is idle. The performance of the busy DPDK thread with IBRS on and off in the sibling idle thread are: IBRS on IBRS off ------- -------- packets/second: 7.8M 10.4M avg tsc cycles/packet: 282.26 209.86 This is a 25% performance degradation. The test system is a Intel Xeon 4114 CPU @ 2.20GHz. Commit bf5835bcdb96 ("intel_idle: Disable IBRS during long idle") disables IBRS when the CPU enters long idle (C6 or below). However, there are existing users out there who have set "intel_idle.max_cstate=1" to decrease latency. Those users won't be able to benefit from this commit. This patch series extends this commit by providing a new "intel_idle.ibrs_off" module parameter to force disable IBRS even when "intel_idle.max_cstate=1" at the expense of increased IRQ response latency. It also includes a commit to allow the disabling of IBRS when a CPU becomes offline. Waiman Long (4): x86/speculation: Add __update_spec_ctrl() helper x86/idle: Disable IBRS when cpu is offline intel_idle: Use __update_spec_ctrl() in intel_idle_ibrs() intel_idle: Add ibrs_off module parameter to force disable IBRS Documentation/admin-guide/pm/intel_idle.rst | 17 ++++++++++++++++- arch/x86/include/asm/spec-ctrl.h | 11 +++++++++++ arch/x86/kernel/smpboot.c | 8 ++++++++ drivers/idle/intel_idle.c | 18 +++++++++++++----- 4 files changed, 48 insertions(+), 6 deletions(-)