On Sat, Oct 07, 2023 at 09:22:55PM -0400, Joel Fernandes wrote:
On Fri, Oct 6, 2023 at 2:20 PM Paul E. McKenney paulmck@kernel.org wrote:
On Fri, Oct 06, 2023 at 01:57:14PM -0400, Liam R. Howlett wrote:
- Paul E. McKenney paulmck@kernel.org [231006 12:47]:
On Fri, Oct 06, 2023 at 12:20:38PM -0400, Liam R. Howlett wrote:
- Naresh Kamboju naresh.kamboju@linaro.org [231005 13:49]:
On Wed, 4 Oct 2023 at 23:33, Greg Kroah-Hartman gregkh@linuxfoundation.org wrote: > > This is the start of the stable review cycle for the 5.15.134 release. > There are 183 patches in this series, all will be posted as a response > to this one. If anyone has any issues with these being applied, please > let me know. > > Responses should be made by Fri, 06 Oct 2023 17:51:12 +0000. > Anything received after that time might be too late. > > The whole patch series can be found in one patch at: > https://www.kernel.org/pub/linux/kernel/v5.x/stable-review/patch-5.15.134-rc... > or in the git tree and branch at: > git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux-stable-rc.git linux-5.15.y > and the diffstat can be found below. > > thanks, > > greg k-h
Results from Linaro’s test farm. Regressions on x86.
Following kernel warning noticed on x86 while booting stable-rc 5.15.134-rc1 with selftest merge config built kernel.
Reported-by: Linux Kernel Functional Testing lkft@linaro.org
Anyone noticed this kernel warning ?
This is always reproducible while booting x86 with a given config.
From that config:
# # RCU Subsystem # CONFIG_TREE_RCU=y # CONFIG_RCU_EXPERT is not set CONFIG_SRCU=y CONFIG_TREE_SRCU=y CONFIG_TASKS_RCU_GENERIC=y CONFIG_TASKS_RUDE_RCU=y CONFIG_TASKS_TRACE_RCU=y CONFIG_RCU_STALL_COMMON=y CONFIG_RCU_NEED_SEGCBLIST=y # end of RCU Subsystem
# # RCU Debugging # CONFIG_PROVE_RCU=y # CONFIG_RCU_SCALE_TEST is not set # CONFIG_RCU_TORTURE_TEST is not set # CONFIG_RCU_REF_SCALE_TEST is not set CONFIG_RCU_CPU_STALL_TIMEOUT=21 CONFIG_RCU_TRACE=y # CONFIG_RCU_EQS_DEBUG is not set # end of RCU Debugging
x86 boot log:
[ 0.000000] Linux version 5.15.134-rc1 (tuxmake@tuxmake) (x86_64-linux-gnu-gcc (Debian 12.2.0-14) 12.2.0, GNU ld (GNU Binutils for Debian) 2.40) #1 SMP @1696443178 ... [ 1.480701] ------------[ cut here ]------------ [ 1.481296] WARNING: CPU: 0 PID: 13 at kernel/rcu/tasks.h:958 trc_inspect_reader+0x80/0xb0 [ 1.481296] Modules linked in: [ 1.481296] CPU: 0 PID: 13 Comm: rcu_tasks_trace Not tainted 5.15.134-rc1 #1 [ 1.481296] Hardware name: Supermicro SYS-5019S-ML/X11SSH-F, BIOS 2.5 11/26/2020 [ 1.481296] RIP: 0010:trc_inspect_reader+0x80/0xb0
This function has changed a lot, including the dropping of this WARN_ON_ONCE(). The warning was replaced in 897ba84dc5aa ("rcu-tasks: Handle idle tasks for recently offlined CPUs") with something that looks equivalent so I'm not sure why it would not trigger in newer revisions.
Obviously the behaviour I changed was the test for the task being idle. I am not sure how best to short-circuit that test from happening during boot as I am not familiar with the RCU code.
The usual test for RCU's notion of early boot being completed is (rcu_scheduler_active != RCU_SCHEDULER_INIT).
Except that "ofl" should always be false that early in boot, at least in mainline.
Is this still true in the final version of the patch where we set the boot task as !idle until just before the early boot is finished? I wouldn't think of this as 'early in boot' anymore as much as the entire kernel setup. Maybe we need to shorten the time we stay in !idle mode for earlier kernels?
In mainline, the ofl variable is defined as cpu_is_offline(cpu), and during boot, the boot CPU is guaranteed to be online. (As opposed to the boot CPU's idle-task state.)
How frequent is this function called? We could check something for early boot... or track down where the cpu is put online and restore idle before that happens?
Once per RCU Tasks Trace grace period per reader seen to be blocking that grace period. Its performance is as issue, but not to anywhere near the same extent as (say) rcu_read_lock_trace().
It's also worth noting that the bug this fixes wasn't exposed until the maple tree (added in v6.1) was used for the IRQ descriptors (added in v6.5).
Lots of latent bugs, to be sure, even with rcutorture. :-/
The Right Thing is to fix the bug all the way back to the introduction, but what fallout makes the backport less desirable than living with the unexposed bug?
You are quite right that it is possible for the risk of a backport to exceed the risk of the original bug.
I defer to Joel (CCed) on how best to resolve this in -stable.
Maybe I am missing something but this issue should also be happening in mainline right?
Even though mainline has 897ba84dc5aa ("rcu-tasks: Handle idle tasks for recently offlined CPUs") , the warning should still be happening due to Liam's "kernel/sched: Modify initial boot task idle setup" because the warning is just rearranged a bit but essentially the same.
IMHO, the right thing to do then is to drop Liam's patch from 5.15 and fix it in mainline (using the ideas described in this thread), then backport both that new fix and Liam's patch to 5.15.
Or is there a reason this warning does not show up on the mainline?
My impression is that dropping Liam's patch for the stable release and revisiting it later is a better approach since tiny RCU is used way less in the wild than tree/tasks RCU. Thoughts?
I think that this one is strange enough that we need to write down the situation in detail, make sure we have all the corner cases covered in both mainline and -stable, and decide what to do from there.
Yes, I know, this email thread contains much of this information, but a little organizing of it would be good.
Would you like to put that together, or should I? If me, I will get a draft out by the end of this coming Tuesday, Pacific Time.
Thanx, Paul