Re: [PATCH v2] mm: memory-failure: use rcu lock instead of tasklist_lock when collect_procs()

[Matthew Wilcox <willy@infradead.org>]

On Mon, Aug 21, 2023 at 05:13:12PM +0800, Tong Tiangen wrote:
> We found a softlock issue in our test, analyzed the logs, and found that
> the relevant CPU call trace as follows:
> 
> CPU0:
>   _do_fork
>     -> copy_process()
>       -> write_lock_irq(&tasklist_lock)  //Disable irq,waiting for
>       					 //tasklist_lock
> 
> CPU1:
>   wp_page_copy()
>     ->pte_offset_map_lock()
>       -> spin_lock(&page->ptl);        //Hold page->ptl
>     -> ptep_clear_flush()
>       -> flush_tlb_others() ...
>         -> smp_call_function_many()
>           -> arch_send_call_function_ipi_mask()
>             -> csd_lock_wait()         //Waiting for other CPUs respond
> 	                               //IPI
> 
> CPU2:
>   collect_procs_anon()
>     -> read_lock(&tasklist_lock)       //Hold tasklist_lock
>       ->for_each_process(tsk)
>         -> page_mapped_in_vma()
>           -> page_vma_mapped_walk()
> 	    -> map_pte()
>               ->spin_lock(&page->ptl)  //Waiting for page->ptl
> 
> We can see that CPU1 waiting for CPU0 respond IPI，CPU0 waiting for CPU2
> unlock tasklist_lock, CPU2 waiting for CPU1 unlock page->ptl. As a result,
> softlockup is triggered.
> 
> For collect_procs_anon(), we will not modify the tasklist, but only perform
> read traversal. Therefore, we can use rcu lock instead of spin lock
> tasklist_lock, from this, we can break the softlock chain above.

The only thing that's giving me pause is that there's no discussion
about why this is safe.  "We're not modifying it" isn't really enough
to justify going from read_lock() to rcu_read_lock().  When you take a
normal read_lock(), writers are not permitted and so you see an atomic
snapshot of the list.  With rcu_read_lock() you can see inconsistencies.
For example, if new tasks can be added to the tasklist, they may not
be seen by an iteration.  Is this OK?  Tasks may be removed from the
tasklist after they have been seen by the iteration.  Is this OK?

As I understand the list RCU code, it guarantees that all tasks which
were on the list before rcu_read_lock() and remain on the list after
rcu_read_unlock() will be seen by a list iteration, while tasks which
are added or removed during that time may or may not be seen.