public inbox for linux-trace-kernel@vger.kernel.org
 help / color / mirror / Atom feed
* [bug report] ring-buffer: Limit time with disabled interrupts in rb_check_pages()
@ 2024-10-22  9:09 Dan Carpenter
  2024-10-22  9:50 ` Petr Pavlu
  2024-10-22 10:17 ` Steven Rostedt
  0 siblings, 2 replies; 3+ messages in thread
From: Dan Carpenter @ 2024-10-22  9:09 UTC (permalink / raw)
  To: Petr Pavlu; +Cc: linux-trace-kernel

Hello Petr Pavlu,

Commit 1f1c2bc9d075 ("ring-buffer: Limit time with disabled
interrupts in rb_check_pages()") from Jul 15, 2024 (linux-next),
leads to the following Smatch static checker warning:

	kernel/trace/ring_buffer.c:1540 rb_check_pages()
	warn: ignoring unreachable code.

kernel/trace/ring_buffer.c
    1501 static void rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
    1502 {
    1503         struct list_head *head, *tmp;
    1504         unsigned long buffer_cnt;
    1505         unsigned long flags;
    1506         int nr_loops = 0;
    1507 
    1508         /*
    1509          * Walk the linked list underpinning the ring buffer and validate all
    1510          * its next and prev links.
    1511          *
    1512          * The check acquires the reader_lock to avoid concurrent processing
    1513          * with code that could be modifying the list. However, the lock cannot
    1514          * be held for the entire duration of the walk, as this would make the
    1515          * time when interrupts are disabled non-deterministic, dependent on the
    1516          * ring buffer size. Therefore, the code releases and re-acquires the
    1517          * lock after checking each page. The ring_buffer_per_cpu.cnt variable
    1518          * is then used to detect if the list was modified while the lock was
    1519          * not held, in which case the check needs to be restarted.
    1520          *
    1521          * The code attempts to perform the check at most three times before
    1522          * giving up. This is acceptable because this is only a self-validation
    1523          * to detect problems early on. In practice, the list modification
    1524          * operations are fairly spaced, and so this check typically succeeds at
    1525          * most on the second try.
    1526          */
    1527 again:
    1528         if (++nr_loops > 3)
    1529                 return;
    1530 
    1531         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
    1532         head = rb_list_head(cpu_buffer->pages);
    1533         if (!rb_check_links(cpu_buffer, head))
    1534                 goto out_locked;
    1535         buffer_cnt = cpu_buffer->cnt;
    1536         tmp = head;
    1537         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
    1538                 return;
                         ^^^^^^
You probably intended to delete this return?

    1539 
--> 1540         while (true) {
    1541                 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
    1542 
    1543                 if (buffer_cnt != cpu_buffer->cnt) {
    1544                         /* The list was updated, try again. */
    1545                         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
    1546                         goto again;
    1547                 }
    1548 
    1549                 tmp = rb_list_head(tmp->next);
    1550                 if (tmp == head)
    1551                         /* The iteration circled back, all is done. */
    1552                         goto out_locked;
    1553 
    1554                 if (!rb_check_links(cpu_buffer, tmp))
    1555                         goto out_locked;
    1556 
    1557                 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
    1558         }
    1559 
    1560 out_locked:
    1561         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
    1562 }

regards,
dan carpenter

^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: [bug report] ring-buffer: Limit time with disabled interrupts in rb_check_pages()
  2024-10-22  9:09 [bug report] ring-buffer: Limit time with disabled interrupts in rb_check_pages() Dan Carpenter
@ 2024-10-22  9:50 ` Petr Pavlu
  2024-10-22 10:17 ` Steven Rostedt
  1 sibling, 0 replies; 3+ messages in thread
From: Petr Pavlu @ 2024-10-22  9:50 UTC (permalink / raw)
  To: Dan Carpenter; +Cc: linux-trace-kernel

On 10/22/24 11:09, Dan Carpenter wrote:
> Hello Petr Pavlu,
> 
> Commit 1f1c2bc9d075 ("ring-buffer: Limit time with disabled
> interrupts in rb_check_pages()") from Jul 15, 2024 (linux-next),
> leads to the following Smatch static checker warning:
> 
> 	kernel/trace/ring_buffer.c:1540 rb_check_pages()
> 	warn: ignoring unreachable code.
> 
> kernel/trace/ring_buffer.c
>     1501 static void rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
>     1502 {
>     1503         struct list_head *head, *tmp;
>     1504         unsigned long buffer_cnt;
>     1505         unsigned long flags;
>     1506         int nr_loops = 0;
>     1507 
>     1508         /*
>     1509          * Walk the linked list underpinning the ring buffer and validate all
>     1510          * its next and prev links.
>     1511          *
>     1512          * The check acquires the reader_lock to avoid concurrent processing
>     1513          * with code that could be modifying the list. However, the lock cannot
>     1514          * be held for the entire duration of the walk, as this would make the
>     1515          * time when interrupts are disabled non-deterministic, dependent on the
>     1516          * ring buffer size. Therefore, the code releases and re-acquires the
>     1517          * lock after checking each page. The ring_buffer_per_cpu.cnt variable
>     1518          * is then used to detect if the list was modified while the lock was
>     1519          * not held, in which case the check needs to be restarted.
>     1520          *
>     1521          * The code attempts to perform the check at most three times before
>     1522          * giving up. This is acceptable because this is only a self-validation
>     1523          * to detect problems early on. In practice, the list modification
>     1524          * operations are fairly spaced, and so this check typically succeeds at
>     1525          * most on the second try.
>     1526          */
>     1527 again:
>     1528         if (++nr_loops > 3)
>     1529                 return;
>     1530 
>     1531         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
>     1532         head = rb_list_head(cpu_buffer->pages);
>     1533         if (!rb_check_links(cpu_buffer, head))
>     1534                 goto out_locked;
>     1535         buffer_cnt = cpu_buffer->cnt;
>     1536         tmp = head;
>     1537         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
>     1538                 return;
>                          ^^^^^^
> You probably intended to delete this return?

There was a problem with rebasing the patch, see
https://lore.kernel.org/oe-kbuild-all/20241019110743.72fa5f29@gandalf.local.home/

It has been fixed in ftrace/for-next.

-- 
Thanks,
Petr

^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: [bug report] ring-buffer: Limit time with disabled interrupts in rb_check_pages()
  2024-10-22  9:09 [bug report] ring-buffer: Limit time with disabled interrupts in rb_check_pages() Dan Carpenter
  2024-10-22  9:50 ` Petr Pavlu
@ 2024-10-22 10:17 ` Steven Rostedt
  1 sibling, 0 replies; 3+ messages in thread
From: Steven Rostedt @ 2024-10-22 10:17 UTC (permalink / raw)
  To: Dan Carpenter; +Cc: Petr Pavlu, Linux trace kernel

On Tue, 22 Oct 2024 12:09:47 +0300
Dan Carpenter <dan.carpenter@linaro.org> wrote:

> Hello Petr Pavlu,

It's not Petr's fault, but mine.

> 
> Commit 1f1c2bc9d075 ("ring-buffer: Limit time with disabled
> interrupts in rb_check_pages()") from Jul 15, 2024 (linux-next),
> leads to the following Smatch static checker warning:
> 

>     1531         raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
>     1532         head = rb_list_head(cpu_buffer->pages);
>     1533         if (!rb_check_links(cpu_buffer, head))
>     1534                 goto out_locked;
>     1535         buffer_cnt = cpu_buffer->cnt;
>     1536         tmp = head;
>     1537         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
>     1538                 return;
>                          ^^^^^^
> You probably intended to delete this return?

He sent two patches where the second patch was a fix to mainline. I
asked him to reverse the order so I could send the fix to Linus, in
which he did. In the mean time I switched the order myself for testing
and broke it with this mistake. When he sent the update I took his
patches but accidentally tested and pushed the broken branch to
linux-next.

When Linus took the fix, Stephen Rothwell reported a bad merge between
next and Linus's tree. That's when I realized I had pushed the wrong
branch.

This is the broken branch. I already updated my for-next branch with
the correct commits and when Stephen syncs the linux-next tree, it
should be fixed.

-- Steve


> 
>     1539 
> --> 1540         while (true) {  
>     1541                 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
>     1542 
>     1543                 if (buffer_cnt != cpu_buffer->cnt) {
>     1544                         /* The list was updated, try again. */
>     1545                         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
>     1546                         goto again;
>     1547                 }
>     1548 
>     1549                 tmp = rb_list_head(tmp->next);
>     1550                 if (tmp == head)
>     1551                         /* The iteration circled back, all is done. */
>     1552                         goto out_locked;
>     1553 
>     1554                 if (!rb_check_links(cpu_buffer, tmp))
>     1555                         goto out_locked;
>     1556 
>     1557                 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
>     1558         }
>     1559 
>     1560 out_locked:
>     1561         raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
>     1562 }
> 
> regards,
> dan carpenter


^ permalink raw reply	[flat|nested] 3+ messages in thread

end of thread, other threads:[~2024-10-22 10:17 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2024-10-22  9:09 [bug report] ring-buffer: Limit time with disabled interrupts in rb_check_pages() Dan Carpenter
2024-10-22  9:50 ` Petr Pavlu
2024-10-22 10:17 ` Steven Rostedt

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox