On Fri, 10 Jul 2026 08:33:57 -0400
Steven Rostedt <[email protected]> wrote:

> On Fri, 10 Jul 2026 12:22:31 +0900
> Masami Hiramatsu (Google) <[email protected]> wrote:
> 
> > Hm, in my view, this warning indicates that the circuit breaker has
> > triggered correctly, so that is not a bug. Under the heavy memory
> > pressure and low-memory situation, the page can be reclaimed soon
> > after it is copied.
> 
> So you are saying that every time the copy_from_user() is executed, the
> page is reclaimed? And this causes a schedule?
> 
> Now, I did have a version that used sched_switch and only incremented the
> counter when a non-kernel thread was scheduled in. Then the test would
> check if the counter increased by 2 or more. As an increase by 1 meant that
> only kernel threads scheduled in which would not corrupt the buffer. The 1
> increment was the current task scheduling back.
> 
> This is based on that work (I'm glad I save old versions in my git tree :-)
> 
> Funny, the comments were from the original change I did back in August of
> 2025, which mentions kernel threads scheduling in to handle the fault.
> 
> I also kept this around in case it was needed. Looks like it may be needed.

Hmm, it's not clear whether this is the result of a scheduling storm
caused by kernel threads. But maybe worth to try.

Jeongho, can you try this patch? Or/and you maybe also try using
mlock the LMKD process memory which is passed to trace_marker.

Thank you, 

> 
> -- Steve
> 
> diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
> index 18710c190c92..19354fe2fca1 100644
> --- a/kernel/trace/trace.c
> +++ b/kernel/trace/trace.c
> @@ -53,6 +53,8 @@
>  #include <linux/io.h> /* vmap_page_range() */
>  #include <linux/fs_context.h>
>  
> +#include <trace/events/sched.h>
> +
>  #include <asm/setup.h> /* COMMAND_LINE_SIZE */
>  
>  #include "trace.h"
> @@ -5984,6 +5986,32 @@ struct trace_user_buf {
>  static DEFINE_MUTEX(trace_user_buffer_mutex);
>  static struct trace_user_buf_info *trace_user_buffer;
>  
> +static DEFINE_PER_CPU(unsigned long, sched_switch_cnt);
> +
> +/*
> + * The per CPU buffer trace_user_buffer is written to optimstically.
> + * The counter sched_switch_cnt is taken, preemption is enabled,
> + * the copying of the user space memory is placed into the trace_user_buffer,
> + * Preeption is re-enabled and the count is read again. If the count is 
> greater
> + * than one from its previous reading, it means that another user space
> + * task scheduled in and the buffer is unreliable for use.
> + */
> +static void
> +probe_sched_switch(void *ignore, bool preempt,
> +                struct task_struct *prev, struct task_struct *next,
> +                unsigned int prev_state)
> +{
> +     /*
> +      * The buffer can only be corrupted by another user space task.
> +      * Ignore kernel tasks that may be scheduled in order to process
> +      * the faulting memory.
> +      */
> +     if (!is_user_task(next))
> +             return;
> +
> +     this_cpu_inc(sched_switch_cnt);
> +}
> +
>  /**
>   * trace_user_fault_destroy - free up allocated memory of a trace user buffer
>   * @tinfo: The descriptor to free up
> @@ -6003,6 +6031,8 @@ void trace_user_fault_destroy(struct 
> trace_user_buf_info *tinfo)
>               kfree(buf);
>       }
>       free_percpu(tinfo->tbuf);
> +
> +     unregister_trace_sched_switch(probe_sched_switch, NULL);
>  }
>  
>  static int user_fault_buffer_enable(struct trace_user_buf_info *tinfo, 
> size_t size)
> @@ -6053,11 +6083,17 @@ static int user_buffer_init(struct 
> trace_user_buf_info **tinfo, size_t size)
>  
>       lockdep_assert_held(&trace_user_buffer_mutex);
>  
> +     ret = register_trace_sched_switch(probe_sched_switch, NULL);
> +     if (ret < 0)
> +             return ret;
> +
>       if (!*tinfo) {
>               alloc = true;
>               *tinfo = kzalloc_obj(**tinfo);
> -             if (!*tinfo)
> +             if (!*tinfo) {
> +                     unregister_trace_sched_switch(probe_sched_switch, NULL);
>                       return -ENOMEM;
> +             }
>       }
>  
>       ret = user_fault_buffer_enable(*tinfo, size);
> @@ -6241,7 +6277,7 @@ char *trace_user_fault_read(struct trace_user_buf_info 
> *tinfo,
>                       return NULL;
>  
>               /* Read the current CPU context switch counter */
> -             cnt = nr_context_switches_cpu(cpu);
> +             cnt = this_cpu_read(sched_switch_cnt);
>  
>               /*
>                * Preemption is going to be enabled, but this task must
> @@ -6272,12 +6308,19 @@ char *trace_user_fault_read(struct 
> trace_user_buf_info *tinfo,
>                       return NULL;
>  
>               /*
> -              * Preemption is disabled again, now check the per CPU context
> -              * switch counter. If it doesn't match, then another user space
> -              * process may have schedule in and corrupted our buffer. In 
> that
> -              * case the copying must be retried.
> +              * Preemption is disabled again, now check the sched_switch_cnt.
> +              * If it increased by two or more, then another user space 
> process
> +              * may have schedule in and corrupted our buffer. In that case
> +              * the copying must be retried.
> +              *
> +              * Note, if this task was scheduled out and only kernel threads
> +              * were scheduled in (maybe to process the fault), then the
> +              * counter would increment again when this task scheduled in.
> +              * If this task scheduled out and another user task scheduled
> +              * in, this task would still need to be scheduled back in and
> +              * the counter would increment by at least two.
>                */
> -     } while (nr_context_switches_cpu(cpu) != cnt);
> +     } while (this_cpu_read(sched_switch_cnt) > cnt + 1);
>  
>       return buffer;
>  }


-- 
Masami Hiramatsu (Google) <[email protected]>

Reply via email to