Re: madvise(MADV_REMOVE) deadlocks on shmem THP

[Hugh Dickins <hughd@google.com>]

On Thu, 14 Jan 2021, Sergey Senozhatsky wrote:

> Hi,
> 
> We are running into lockups during the memory pressure tests on our
> boards, which essentially NMI panic them. In short the test case is
> 
> - THP shmem
>     echo advise > /sys/kernel/mm/transparent_hugepage/shmem_enabled
> 
> - And a user-space process doing madvise(MADV_HUGEPAGE) on new mappings,
>   and madvise(MADV_REMOVE) when it wants to remove the page range
> 
> The problem boils down to the reverse locking chain:
> 	kswapd does
> 
> 		lock_page(page) -> down_read(page->mapping->i_mmap_rwsem)
> 
> 	madvise() process does
> 
> 		down_write(page->mapping->i_mmap_rwsem) -> lock_page(page)
> 
> 
> 
> CPU0                                                       CPU1
> 
> kswapd                                                     vfs_fallocate()
>  shrink_node()                                              shmem_fallocate()
>   shrink_active_list()                                       unmap_mapping_range()
>    page_referenced() << lock page:PG_locked >>                unmap_mapping_pages()  << down_write(mapping->i_mmap_rwsem) >>
>     rmap_walk_file()                                           zap_page_range_single()
>      down_read(mapping->i_mmap_rwsem) << W-locked on CPU1>>     unmap_page_range()
>       rwsem_down_read_failed()                                   __split_huge_pmd()
>        __rwsem_down_read_failed_common()                          __lock_page()  << PG_locked on CPU0 >>
>         schedule()                                                 wait_on_page_bit_common()
>                                                                     io_schedule()

Very interesting, Sergey: many thanks for this report.

There is no doubt that kswapd is right in its lock ordering:
__split_huge_pmd() is in the wrong to be attempting lock_page().

Which used not to be done, but was added in 5.8's c444eb564fb1 ("mm:
thp: make the THP mapcount atomic against __split_huge_pmd_locked()").

Which explains why this deadlock was not seen years ago: that
surprised me at first, since the case you show to reproduce it is good,
but I'd expect more common ways in which that deadlock could show up.

And your report is remarkably timely too: I have two other reasons
for looking at that change at the moment (I'm currently catching up
with recent discussion of page_count versus mapcount when deciding
COW page reuse).

I won't say more tonight, but should have more to add tomorrow.

Hugh