lkml.org 
[lkml]   [2020]   [Jul]   [14]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: power-off delay/hang due to commit 6d25be57 (mainline)
On 2020-06-24 23:49:52 [+0200], Stephen Berman wrote:

Let me summarize the thread here:

On Stephen's system, ACPI informs the thermal zone driver to poll the
temperature every second and the driver does so.
The driver queries the temperature by invoking acpi_evaluate_integer()
which invokes (at some point) acpi_ev_queue_notify_request().
This then invokes acpi_os_execute_deferred() via
queue_work_on(, kacpi_notify_wq, )

acpi_os_execute_deferred() invokes acpi_ev_notify_dispatch() and this is
no longer synchronised with the initial acpi_evaluate_integer() request.

Before commit
6d25be5782e48 ("sched/core, workqueues: Distangle worker accounting from rq lock")

that function took on average 1.023993 seconds to complete. The
interval when the thermal driver invokes acpi_evaluate_integer() isn't
exactly 1 second but almost (it is not an absolute timer so). Still it
looks that one function slowly overtakes the other. After 5 Minutes
uptime there is:

| kworker/0:1-12 [000] 312.315565: acpi_ev_queue_notify_request: Dispatching Notify on [TZ10] (Thermal) Value 0x81 (Information Change) Node ffff9935ab20f7a8
| kworker/0:1-12 [000] 312.315567: acpi_os_execute: Adding acpi_ev_notify_dispatch+0x0/0x5a ffff9935a6c64050 <ffff9935a7eb5e00>

Enqueue worker with job ffff9935a7eb5e00

| kworker/0:1-12 [000] 312.315596: acpi_os_execute_deferred_notify: End ffff9935a7eb5c80 acpi_ev_notify_dispatch+0x0/0x5a(ffff9935a6c64c80)
| kworker/0:1-12 [000] 312.315607: acpi_os_execute_deferred_notify: Start ffff9935a7eb5d80 acpi_ev_notify_dispatch+0x0/0x5a(ffff9935a6c64dc0)

previous worker completed, another (already enqueued) started.

| kworker/0:1-12 [000] 313.339564: acpi_ev_queue_notify_request: Dispatching Notify on [TZ10] (Thermal) Value 0x81 (Information Change) Node ffff9935ab20f7a8
| kworker/0:1-12 [000] 313.339566: acpi_os_execute: Adding acpi_ev_notify_dispatch+0x0/0x5a ffff9935a6c64f00 <ffff9935a7eb5c80>

another one enqueued.

| kworker/0:1-12 [000] 313.339595: acpi_os_execute_deferred_notify: End ffff9935a7eb5d80 acpi_ev_notify_dispatch+0x0/0x5a(ffff9935a6c64dc0)
| kworker/0:1-12 [000] 313.339597: acpi_os_execute_deferred_notify: Start ffff9935a7eb5e00 acpi_ev_notify_dispatch+0x0/0x5a(ffff9935a6c64050)

finally, job ffff9935a7eb5e00 started (one second after enqueue).

| kworker/0:1-12 [000] 314.363571: acpi_ev_queue_notify_request: Dispatching Notify on [TZ10] (Thermal) Value 0x81 (Information Change) Node ffff9935ab20f7a8
| kworker/0:1-12 [000] 314.363574: acpi_os_execute: Adding acpi_ev_notify_dispatch+0x0/0x5a ffff9935a6c646e0 <ffff9935a7eb5d80>
| kworker/0:1-12 [000] 314.363608: acpi_os_execute_deferred_notify: End ffff9935a7eb5e00 acpi_ev_notify_dispatch+0x0/0x5a(ffff9935a6c64050)

and ended, two seconds after enqueue. Before it ended, the system
enqueued two more jobs. And this just within 5 Minutes of uptime. The
worker pile slowly up.

According to Stephen, after previously mentioned commit the situation
gets worse. According to tracing the execution time of
acpi_ev_notify_dispatch() varies between 1.023990 (like before) and
2.048005 seconds. While I don't have an explanation for the changed
behaviour (or see anything wrong the commit in question), the pile up of
worker increased.
Stephen noticed this because system shutdown flushes kacpi_notify_wq
and this takes quite some time (with that amount of worker pending).

I tried a few test cases to reproduce that behaviour but failed.
However, even before that commit in question the situation is far from
perfect:
- There is the lack of synchronisation between thermal driver's
requests and its tail (that acpi_ev_notify_dispatch() part).
- Do we have to respect ACPI interval of one seconds. Is the hardware
really so accurate that it can change noticeable in one second.
- The requests are already back to back which keeps the CPU busy (maybe
sched-switch will reveal that the CPU is idle most of the time).

So...

Is there a simple way to synchronize the ACPI part? The obvious thing
would be

flush_workqueue(kacpi_notify_wq);
or
acpi_os_wait_events_complete()

in thermal_get_temp().
Would it make sense to ensure that the polling interval is no less than
10 seconds?

> Steve Berman

Sebastian

\
 
 \ /
  Last update: 2020-07-14 15:44    [W:0.115 / U:2.936 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site