On Fri, Oct 04, 2024 at 05:56:52PM +0100, David Howells wrote: > Ming Lei <ming.lei@xxxxxxxxxx> wrote: > > > All iov_iter_bvec() users only want to extract virt-contiguous pages from > > iov_iter_extract_pages() instead physical-contiguous pages. > > What do you mean by "virt-contiguous"? Virtual according to what mapping? The term is from comment iov_iter_extract_kvec_pages(), seems it is invented by you, :-) Actually iov_iter_extract_pages() requires that there isn't gap in the extracted pages, so 'offset' only exists in the 1st page, then these pages can be mapped to one virtual(contiguous) address. > > The reason for physical contiguity is that you can pass a set of physical > contiguous pages as a single DMA descriptor. Therefore, at some point, you > might end up screwing up skb_splice_from_iter(). Currently, that's limited to > a PAGE_SIZE per fragment, but hopefully that will be fixed at some point. If any user wants to extract physical pages, new interface can be added for returning single page instead of page array, cause it is physically contiguous. Other kind of iterators(UBUF, KVEC, ...) do return non physically-contiguous pages. The point is that one bvec often point to one page except for huge page case, so iov_iter_extract_pages() just returns single page each time no matter how big maxpages & maxsize is passed in. It is actually one regression: Fixes: a7e689dd1c06 ("block: Convert bio_iov_iter_get_pages to use iov_iter_extract_pages") Thanks, Ming