On Thu, Apr 17, 2025 at 03:36:43PM +0100, Matthew Wilcox wrote: > Let's suppose we're allocating the PI buffer in anonymous memory. > Also, we're under memory pressure. We've already swapped out the page > containing the PI buffer once, so it's in the swap cache and marked > as clean. We do a READ from the device, and the new metadata is written > to the page. Then a new round of memory reclaim happens and this page > is chosen. If it's still clean, the new contents will not be written > to swap and the page will simply be discarded. When we go to validate > the PI data, the page will be swapped back in, but it will have old PI > information in it so the verification will fail. > > What we need to do is mark the folio dirty at pin time. I believe > O_DIRECT does this properly, and I'm not sure whether this code does it > properly or not. O_DIRECT reads dirty right after pinning and then check if the dirty bit has been cleared in the I/O completion handler and redirty from a workqueue if so. We're currently trying to figure out if we still need that redirtying with proper pinning. Either way metadata should follow the data behavior 1:1 here and preferably also share the code for that as much as possible.
