On Tue, Apr 27, 2021 at 05:58:16PM -0700, Hugh Dickins wrote:

On Tue, 27 Apr 2021, Axel Rasmussen wrote:

quoted

In a previous commit, we added the mcopy_atomic_install_pte() helper.
This helper does the job of setting up PTEs for an existing page, to map
it into a given VMA. It deals with both the anon and shmem cases, as
well as the shared and private cases.

In other words, shmem_mcopy_atomic_pte() duplicates a case it already
handles. So, expose it, and let shmem_mcopy_atomic_pte() use it
directly, to reduce code duplication.

This requires that we refactor shmem_mcopy_atomic_pte() a bit:

Instead of doing accounting (shmem_recalc_inode() et al) part-way
through the PTE setup, do it afterward. This frees up
mcopy_atomic_install_pte() from having to care about this accounting,
and means we don't need to e.g. shmem_uncharge() in the error path.

A side effect is this switches shmem_mcopy_atomic_pte() to use
lru_cache_add_inactive_or_unevictable() instead of just lru_cache_add().
This wrapper does some extra accounting in an exceptional case, if
appropriate, so it's actually the more correct thing to use.

Signed-off-by: Axel Rasmussen <axelrasmussen@google.com>

Not quite. Two things.

One, in this version, delete_from_page_cache(page) has vanished
from the particular error path which needs it.

Agreed.  I also spotted that the set_page_dirty() seems to have been overlooked
when reusing mcopy_atomic_install_pte(), which afaiu should be move into the
helper.

Two, and I think this predates your changes (so needs a separate
fix patch first, for backport to stable? a user with bad intentions
might be able to trigger the BUG), in pondering the new error paths
and that /* don't free the page */ one in particular, isn't it the
case that the shmem_inode_acct_block() on entry might succeed the
first time, but atomic copy fail so -ENOENT, then something else
fill up the tmpfs before the retry comes in, so that retry then
fail with -ENOMEM, and hit the BUG_ON(page) in __mcopy_atomic()?

(As I understand it, the shmem_inode_unacct_blocks() has to be
done before returning, because the caller may be unable to retry.)

What the right fix is rather depends on other uses of __mcopy_atomic():
if they obviously cannot hit that BUG_ON(page), you may prefer to leave
it in, and fix it here where shmem_inode_acct_block() fails. Or you may
prefer instead to delete that "else BUG_ON(page);" - looks as if that
would end up doing the right thing.  Peter may have a preference.

To me, the BUG_ON(page) wanted to guarantee mfill_atomic_pte() should have
consumed the page properly when possible.  Removing the BUG_ON() looks good
already, it will just stop covering the case when e.g. ret==0.

So maybe slightly better to release the page when shmem_inode_acct_block()
fails (so as to still keep some guard on the page)?

Thanks,

-- 
Peter Xu