Re: [RFC 0/1] memfd: Support mapping to zero page on reading

[Hugh Dickins]

On Wed, 22 Dec 2021, Peng Liang wrote:

> Hi all,
> 
> Recently we are working on implementing CRIU [1] for QEMU based on
> Steven's work [2].  It will use memfd to allocate guest memory in order
> to restore (inherit) it in the new QEMU process.  However, memfd will
> allocate a new page for reading while anonymous memory will map to zero
> page for reading.  For QEMU, memfd may cause that all memory are
> allocated during the migration because QEMU will read all pages in
> migration.  It may lead to OOM if over-committed memory is enabled,
> which is usually enabled in public cloud.
> 
> In this patch I try to add support mapping to zero pages on reading
> memfd.  On reading, memfd will map to zero page instead of allocating a
> new page.  Then COW it when a write occurs.
> 
> For now it's just a demo for discussion.  There are lots of work to do,
> e.g.:
> 1. don't support THP;
> 2. don't support shared reading and writing, only for inherit.  For
>    example:
>      task1                        | task2
>        1) read from addr          |
>                                   |   2) write to addr
>        3) read from addr again    |
>    then 3) will read 0 instead of the data task2 writed in 2).
> 
> Would something similar be welcome in the Linux?

David has made good suggestions on better avoiding the need for
such a change, for the use case you have in mind.

And I don't care for the particular RFC patch that you posted.

But I have to say that use of ZERO_PAGE for shmem/memfd/tmpfs read-fault
might (potentially) be very welcome.  Not as some MFD_ZEROPAGE special
case, but as how it would always work.  Deleting the shmem_recalc_inode()
cruft, which is there to correct accounting for the unmodified read-only
pages, after page reclaim has got around to freeing them later.

It does require more work than you gave it in 1/1: mainly, as you call
out above, there's a need to note in the mapping's XArray when ZERO_PAGE
has been used at an offset, and do an rmap walk to unmap those ptes when
a writable page is substituted - see __xip_unmap() in Linux 3.19's
mm/filemap_xip.c for such an rmap walk.

Though when this came up before (in the "no-fault mmap" MAP_NOSIGBUS
thread last year - which then got forgotten), Linus was wary of that
unmapping, and it was dropped for a simple MAP_PRIVATE implementation.

And I've never scoped out what is needed to protect the page from
writing in all circumstances: in principle, it ought to be easy by
giving shmem_vm_ops a page_mkwrite; but that's likely to come with
a performance penalty, which may not be justified for this case.

I didn't check what you did for write protection: maybe what you
did was enough, but one has to be very careful about that.

Making this change to ZERO_PAGE has never quite justified the effort
so far: temporarily allocated pages have worked well enough in most
circumstances.

Hugh

> 
> Thanks,
> Peng
> 
> [1] https://criu.org/Checkpoint/Restore
> [2] https://patchwork.kernel.org/project/qemu-devel/cover/1628286241-217457-1-git-send-email-steven.sistare@xxxxxxxxxx/
> 
> Peng Liang (1):
>   memfd: Support mapping to zero page on reading memfd
> 
>  include/linux/fs.h         |  2 ++
>  include/uapi/linux/memfd.h |  1 +
>  mm/memfd.c                 |  8 ++++++--
>  mm/memory.c                | 37 ++++++++++++++++++++++++++++++++++---
>  mm/shmem.c                 | 10 ++++++++--
>  5 files changed, 51 insertions(+), 7 deletions(-)
> 
> -- 
> 2.33.1