On Wed, Feb 24, 2021 at 05:12:36PM +0100, Christoph Hellwig wrote: > On Wed, Feb 24, 2021 at 07:19:52PM +0800, Ming Lei wrote: > > > if (bio_is_passthrough(bio_src)) > > > - bio = bio_kmalloc(GFP_NOIO, bio_segments(bio_src)); > > > + bio = bio_kmalloc(GFP_NOIO | __GFP_NOFAIL, > > > + bio_segments(bio_src)); > > > > bio_kmalloc() still may fail if bio_segments(bio_src) is > UIO_MAXIOV. > > Yes, but bio_kmalloc is what is used to allocate the passthrough > requests to start with, so we'd not even make it here. The original bio_kmalloc() may start with allowed nr_iovecs , but later more pages are retrieved from iov_iter and added to the bio, see bio_map_user_iov(). Then bounce_clone_bio() will see too big bio_segments(bio_src) to be held in UIO_MAXIOV vecs. This behavior is similar with blkdev_direct_IO(). -- Ming
