Re: [PATCH] x86/split_lock: fix delayed detection enabling

From: Maksim Davydov
Date: Fri Apr 19 2024 - 06:16:21 EST



On 3/31/24 20:07, Guilherme G. Piccoli wrote:
On 21/03/2024 16:55, Maksim Davydov wrote:
If the warn mode with disabled mitigation mode is used, then on each cpu
where the split lock occurred detection will be disabled in order to make
progress and delayed work will be scheduled, which then will enable
detection back. Now it turns out that all CPUs use one global delayed
work structure. This leads to the fact that if a split lock occurs on
several CPUs at the same time (within 2 jiffies), only one cpu will
schedule delayed work, but the rest will not. The return value of
schedule_delayed_work_on() would have shown this, but it is not checked
in the code
In order to fix the warn mode with disabled mitigation mode, delayed work
has to be a per-cpu.

Fixes: 727209376f49 ("x86/split_lock: Add sysctl to control the misery mode")
Thanks Maksim! I confess I (think I) understand the theory behind the
possible problem, but I'm not seeing how it happens - probably just me
being silly , but can you help me to understand it clearly?

Let's say we have 2 CPUs, CPU0 and CPU1 and we're running with
sld_mitigate = 0, meaning we don't have "the misery".

If the code running in CPU0 reaches split_lock_warn(), my understanding
is that it warns the user, schedule the sld reenable [via and
schedule_delayed_work_on()] and disables the feature with
sld_update_msr(false), correct? So, does this disabling happens only at
core level, or it disables for the whole CPU including all cores?

But back to our example, if CPU1 detects the split lock, it'll run the
same procedure as CPU0 did - so are you saying we have a race there if
CPU1 face a split lock before CPU0 disabled the MSR?

Maybe a more clear example of the issue would be even helpful in the
commit message, showing the path both CPUs would take and how the
problem happens exactly.

Thanks in advance,


Guilherme

Sorry for a late reply.

I made a diagram to explain how this bug occurs. If it makes it clearer,
then I will include the diagram in the commit description.

Some information that should be taken into account:
* sld_update_msr() enables/disables SLD on both CPUs on the same core
* schedule_delayed_work_on() internally checks WORK_STRUCT_PENDING_BIT. If
  a work has the 'pending' status, then schedule_delayed_work_on() will
  return an error code and, most importantly, the work will not be placed
  in the workqueue

Let's say we have a multicore system on which split_lock_mitigate=0 and
a multithreaded application is running that calls splitlock in multiple
threads. Due to the fact that sld_update_msr() affects the entire core (both
CPUs), we will consider 2 CPUs from different cores. Let the 2 threads of
this application schedule to CPU0 (core 0) and to CPU 2 (core 1), then:

| ||                                      |
|             CPU 0 (core 0)          ||             CPU 2 (core 1)           |
|_____________________________________||______________________________________|
| ||                                      |
| 1) SPLIT LOCK occured ||                                      |
| ||                                      |
| 2) split_lock_warn() ||                                      |
| ||                                      |
| 3) sysctl_sld_mitigate == 0 ||                                      |
|    (work = &sl_reenable) ||                                      |
| ||                                      |
| 4) schedule_delayed_work_on() ||                                      |
|    (reenable will be called after ||                                      |
|     2 jiffies on CPU 0) ||                                      |
| ||                                      |
| 5) disable SLD for core 0 ||                                      |
| ||                                      |
|    ----------------------------- ||                                      |
| ||                                      |
|                                     || 6) SPLIT LOCK occured                |
| ||                                      |
|                                     || 7) split_lock_warn()                 |
| ||                                      |
|                                     || 8) sysctl_sld_mitigate == 0          |
|                                     ||    (work = &sl_reenable,             |
|                                     ||     the same address as in 3) )      |
| ||                                      |
|             2 jiffies               || 9) schedule_delayed_work_on() fails  |
|                                     ||    because the work is in the        |
|                                     ||    pending state since 4). The work  |
|                                     ||    wasn't placed to the workqueue.   |
|                                     ||    reenable won't be called on CPU 2 |
| ||                                      |
|                                     || 10) disable SLD for core 0           |
| ||                                      |
|                                     ||     From now on, SLD will            |
|                                     ||     never be reenabled on core 1     |
| ||                                      |
|    ----------------------------- ||                                      |
| ||                                      |
|    11) enable SLD for core 0 by ||                                      |
|        __split_lock_reenable ||                                      |
| ||                                      |


If the application threads can be scheduled to all processor cores, then over
time there will be only one core left, on which SLD will be enabled and split
lock can be detected; and on all other cores SLD will be disabled all the
time.
Most likely, this bug has not been noticed for so long because
sysctl_sld_mitigate default value is 1, and in this case a semaphore is used
that does not allow 2 different cores to have SLD disabled at the same time,
that is, strictly only one work is placed in the workqueue.

--
Best regards,
Maksim Davydov