Re: [BISECTED - impi related]: acpi_power_meter: power*_average sysfs read hangs, mutex deadlock in hwmon_attr_show since v6.18.y

[Corey Minyard <corey@minyard.net>]

On Thu, Feb 12, 2026 at 01:27:41PM +0100, Rafael J. Wysocki wrote:
> On Thu, Feb 12, 2026 at 10:11???AM Jaroslav Pulchart
> <jaroslav.pulchart@gooddata.com> wrote:
> >
> > >
> > > On Fri, Feb 6, 2026 at 4:58???PM Corey Minyard <corey@minyard.net> wrote:
> > > >
> > > > On Fri, Feb 06, 2026 at 01:08:56PM +0100, Rafael J. Wysocki wrote:
> > > > > On Thu, Feb 5, 2026 at 11:34???PM Guenter Roeck <linux@roeck-us.net> wrote:
> > > > > >
> > > > > > On Thu, Feb 05, 2026 at 08:04:12PM +0100, Rafael J. Wysocki wrote:
> > > > > > > Cc: Corey
> > > > > > >
> > > > > > > On Thu, Feb 5, 2026 at 6:51???PM Guenter Roeck <linux@roeck-us.net> wrote:
> > > > > > > >
> > > > > > > > On Thu, Feb 05, 2026 at 08:25:57AM +0100, Igor Raits wrote:
> > > > > > > > > On Wed, Feb 4, 2026 at 11:49???PM Guenter Roeck <linux@roeck-us.net> wrote:
> > > > > > > > > >
> > > > > > > > > > On 2/4/26 11:54, Igor Raits wrote:
> > > > > > > > > > > I have written a patch with the help of AI and it fixes the problem. Attached.
> > > > > > > > > > >
> > > > > > > > > >
> > > > > > > > > > "No MIME, no links, no compression, no attachments.  Just plain text"
> > > > > > > > >
> > > > > > > > > Sorry for that, I had assumed that attaching the file would make it in-line.
> > > > > > > > >
> > > > > > > > > > ... which means I can not provide inline feedback, which is the whole
> > > > > > > > > > point of the above.
> > > > > > > > > >
> > > > > > > > > > Your patch crosses subsystems, so it will need to be split in two
> > > > > > > > > > (assuming the ACPI side is even needed). Also, references to iDRAC
> > > > > > > > > > in common code seem inappropriate.
> > > > > > > > >
> > > > > > > > > Yes, this I believe was the essential part (it was the last piece in
> > > > > > > > > my testing which fixed the hanging):
> > > > > > > > >
> > > > > > > >
> > > > > > > > Then I'll need to ask differently: What happens if you drop the IPMI code,
> > > > > > > > and just keep the wait_for_completion -> wait_for_completion_timeout
> > > > > > > > change ? Would that be sufficient to solve the problem ?
> > > > > > >
> > > > > > > I'd rather say "Would that be sufficient to make the symptoms go
> > > > > > > away?" as it most likely papers over the real problem.
> > > > > > >
> > > > > >
> > > > > > Good point. Worse, it may result in UAF or memory leaks.
> > > > > >
> > > > > > > > Either case, the need for this change suggests that the ipmi change
> > > > > > > > may not be complete, since it should send a completion with an error.
> > > > > > >
> > > > > > > I think that reverting commit bc3a9d217755 ("ipmi:si: Gracefully
> > > > > > > handle if the BMC is non-functional") should also be considered as a
> > > > > > > possible way forward because it clearly did not improve things as
> > > > > > > expected, at least in this particular case.
> > > > > > >
> > > > > >
> > > > > > I tend to agree. I ran a number of AI code reviews over the patch, and
> > > > > > each time it finds new (and different) problems. The fact that the acpi
> > > > > > patch is still needed even after applying the ipmi changes suggests that
> > > > > > something is still missing in the ipmi code.
> > > > > >
> > > > > > > It evidently did something that confuses things quite a bit.  Either
> > > > > > > it is returning IPMI_BUS_ERR instead of IPMI_ERR_UNSPECIFIED, or it is
> > > > > > > the "hosed" state and refusing to accept messages.
> > > > > > >
> > > > > >
> > > > > > More than that. My latest AI results are below, just for reference
> > > > > > (using Gemini 3 with Chris Mason's debug prompts). The prompt I used
> > > > > > for this run is:
> > > > >
> > > > > Well, I guess it's time to send a revert patch then.
> > > >
> > > > Thanks for the CC.
> > > >
> > > > Let's fix it right in the IPMI driver.
> > > >
> > > > >
> > > > > > "
> > > > > > The top commit in the linux/ directory results in hung tasks if the BMC
> > > > > > stops responding. Using @review-prompts/kernel/debugging.md analyze the
> > > > > > patch, identify the reason for the hung task problem, suggest and implement
> > > > > > a fix. Note that there may be more than one problem in the patch, so analyze
> > > > > > the complete patch and do not stop after fiding the first regression.
> > > > > > "
> > > > > >
> > > > > > I think that catches most of the problem, but not all of it.
> > > > > >
> > > > > > Guenter
> > > > > >
> > > > > > ---
> > > > > >
> > > > > > Summary of crash or warning:
> > > > > > Hung task detected in ipmi_si driver when BMC becomes non-functional.
> > > > > > Processes waiting for IPMI responses (e.g. ipmitool, monitoring agents) enter D state and never recover.
> > > > > >
> > > > > > Kernel version if available:
> > > > > > Top of tree (commit bc3a9d217755f65c137f145600f23bf1d6c31ea9)
> > > > > >
> > > > > > Machine type if available:
> > > > > > Generic Server with BMC
> > > > > >
> > > > > > Cleaned up copy of oops or stack trace:
> > > > > > [  120.123456] INFO: task ipmitool:1234 blocked for more than 120 seconds.
> > > > > > [  120.123457]       Not tainted 6.14.0-rc1 #1
> > > > > > [  120.123458] "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message.
> > > > > > [  120.123459] task:ipmitool        state:D stack:    0 pid: 1234 ppid:  100 flags:0x00000000
> > > > > > [  120.123460] Call Trace:
> > > > > > [  120.123461]  <TASK>
> > > > > > [  120.123462]  __schedule+0x123/0x456
> > > > > > [  120.123463]  schedule+0x45/0x78
> > > > > > [  120.123464]  schedule_timeout+0x9a/0xbc
> > > > > > [  120.123465]  wait_for_completion+0xde/0xf0
> > > > > > [  120.123466]  ipmi_request_settime+0x123/0x145
> > > > > > [  120.123467]  ...
> > > > > > [  120.123468]  </TASK>
> > > > > >
> > > > > > Any other kernel messages you found relevant:
> > > > > > N/A
> > > > > >
> > > > > > Explanation of the problem:
> > > > > > 1. Hung Task:
> > > > > > The patch "ipmi:si: Gracefully handle if the BMC is non-functional" introduces a new state `SI_HOSED` to handle BMC failures. When the driver detects that the BMC is not responding, it transitions to `SI_HOSED` and fails the currently processing message (`curr_msg`). However, if a new message is queued via `sender()` (populating `waiting_msg`) during a recovery probe (state `SI_GETTING_FLAGS`), and that probe subsequently fails, the state machine transitions back to `SI_HOSED`. In this transition, the driver checks and fails `curr_msg`, but it neglects to check or fail `waiting_msg`. As a result, the `waiting_msg` remains in the queue indefinitely, causing the waiting process to hang.
> > > > >
> > > > > That's quite convincing and it would explain the observed symptoms.
> > > >
> > > > Yes, and it's a fairly easy fix, I think.  The waiting message just
> > > > needs to be returned in that case.  The following patch should do it:
> > >
> > > Jaroslav, it would be good to test the patch below on top of 6.19.  I
> > > can put it on a test git branch if need be, so please let me know.
> > >
> > > > diff --git a/drivers/char/ipmi/ipmi_si_intf.c b/drivers/char/ipmi/ipmi_si_intf.c
> > > > index 5459ffdde8dc..ff159b1162b9 100644
> > > > --- a/drivers/char/ipmi/ipmi_si_intf.c
> > > > +++ b/drivers/char/ipmi/ipmi_si_intf.c
> > > > @@ -809,6 +809,12 @@ static enum si_sm_result smi_event_handler(struct smi_info *smi_info,
> > > >                          */
> > > >                         return_hosed_msg(smi_info, IPMI_BUS_ERR);
> > > >                 }
> > > > +               if (smi_info->waiting_msg != NULL) {
> > > > +                       /* Also handle if there was a message waiting. */
> > > > +                       smi_info->curr_msg = smi_info->waiting_msg;
> > > > +                       smi_info->waiting_msg = NULL;
> > > > +                       return_hosed_msg(smi_info, IPMI_BUS_ERR);
> > > > +               }
> > > >                 smi_mod_timer(smi_info, jiffies + SI_TIMEOUT_HOSED);
> > > >                 goto out;
> > > >         }
> >
> > I apply ^ patch to both 6.18.10 and 6.19 and reproduced the issue on
> > both, so it does not fix the problem.
> 
> Thanks!
> 
> With all due respect to everyone involved (including the AI), this
> means that we are not anywhere close to fixing the problem and it
> would be a shame to ship 7.0 with it.
> 
> I'm sending a revert patch shortly.

Unfortunately, that patch fixed an issue others were having.

I'm going to try to hack on qemu and see if I can reproduce this.  I
don't have any hardware where this is feasible.

-corey