vm: use atomic fetchadd in vm_page_sunbusy

[mjg]

Summary

This also fixes a bug where not-last unbusy failed to post a release
fence.

[markj]

sys/vm/vm_page.c

981

We should keep this assertion (and additionally verify that the page is not xbusy).

[kib]

The 'lack' of the release fence for non-last sunbusy is intentional, it is only needed for the last unlock.

[mjg]

sys/vm/vm_page.c

981

you want this asserted before or after (as in on the result of fetchadd). note there is already vm_page_assert_sbusied in there

[mjg]

In D36084#819356, @kib wrote:

The 'lack' of the release fence for non-last sunbusy is intentional, it is only needed for the last unlock.

that's the same bug as 890611286ee256314407bbcf64ad74956939eac7

[markj]

sys/vm/vm_page.c

981

The assertion should be on the state returned by fetchadd.

The existing assertion is looks sufficient, but it's racy and that can easily be fixed.

[mjg]

sys/vm/vm_page.c

981

something of this sort?

x = atomic_fetchadd_int(&m->busy_lock, -VPB_ONE_SHARER);
KASSERT(x != VPB_FREED, ("page %p is freed", m));
KASSERT(x != VPB_UNBUSIED && (x & VPB_BIT_SHARED) != 0,
    ("page %p not sbusied", m));

[markj]

sys/vm/vm_page.c

981

Looks right to me.