On Tue, Sep 06, 2022 at 12:58:59AM -0700, John Hubbard wrote: > > For FOLL_PIN callers, never pin bvec and kvec pages: For file systems > > not acquiring a reference is obviously safe, and the other callers will > > need an audit, but I can't think of why it woul ever be unsafe. > > In order to do that, one would need to be confident that such bvec and kvec > pages do not get passed down to bio_release_pages() (or the new > bio_unpin_pages()). Also, I'm missing a key point, because today bvec pages get > a get_page() reference from __iov_iter_get_pages_alloc(). If I just skip > that, then the get/put calls are unbalanced... Except that for the most relevant callers (bdev and iomap) we never end up in __iov_iter_get_pages_alloc. What I suggest is that, with the move to the pin helper, we codify that logic. For callers of the bio_iov_iter_pin_pages helper, the corresponding bio_unpin_pages helper will encapsulate that logic. For direct calls to iov_iter_pin_pages, we should add a iov_iter_unpin_pages helper that also encapsulates the logic. The beauty is that this means the caller itself does not have to care about any of thise, and the logic is in one (well two due to the block special case that reuses the bio_vec array for the pages space) place. > I can just hear Al Viro repeating his points about splice() and vmsplice(), > heh. :) splice does take these references before calling into the file system (see my reply to Jan).
