> On Wed, 12 Sep 2001, Marcelo Tosatti wrote:
> > On Tue, 11 Sep 2001, Hugh Dickins wrote:
> > > It may be made more likely by my swapoff changes (not bumping swap
> > > count in valid_swaphandles), but it's not been introduced by those
> > > changes. Though usually swapin_readahead/valid_swaphandles covers
> > > (includes) the particular swap entry which do_swap_page actually
> > > wants to bring in, under pressure that's not always so, and then
> > > the race you outline can occur with the "bare" read_swap_cache_async
> > > for which there was no bumping. Furthermore, you can play your
> > > scenario with valid_swaphandles through to add_to_swap_cache on CPU0
> > > interposed between the get_swap_page and add_to_swap_cache on CPU1
> > > (if interrupt on CPU1 diverts it).
> > I don't think so. A "bare" read_swap_cache_async() call only happens on
> > swap entries which already have additional references. That is, its
> > guaranteed that a "bare" read_swap_cache_async() call only happens for
> > swap map entries which already have a reference, so we're guaranteed that
> > it cannot be reused.
> Almost agreed, but there may be a long interval between when that reference
> was observed in the page table, and when read_swap_cache_async upon it is
> actually performed (waiting for BKL, waiting to allocate pages for prior
> swapin_readahead). In that interval the reference can be removed:
> certainly by swapoff, certainly by vm_swap_full removal, anything else?
Not sure about swapoff().
vm_swap_full() is only going to remove the reference _after_ we did the
swapin, so I don't see how the race can happen with it.
> That's why the pte_same tests were added into do_swap_page in 2.4.8.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to firstname.lastname@example.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/