On Sat, Sep 23, 2017 at 09:33:23PM +0100, Al Viro wrote:

On Sat, Sep 23, 2017 at 06:19:26PM +0100, Al Viro wrote:

quoted

On Sat, Sep 23, 2017 at 05:55:37PM +0100, Al Viro wrote:

quoted

IOW, the loop on failure exit should go through the bio, like __bio_unmap_user()
does.  We *also* need to put everything left unused in pages[], but only from the
last iteration through iov_for_each().

Frankly, I would prefer to reuse the pages[], rather than append to it on each
iteration.  Used iov_iter_get_pages_alloc(), actually.

Something like completely untested diff below, perhaps...

quoted

+				unsigned n = PAGE_SIZE - offs;
+				unsigned prev_bi_vcnt = bio->bi_vcnt;

Sorry, that should've been followed by
				if (n > bytes)
					n = bytes;

Anyway, a carved-up variant is in vfs.git#work.iov_iter.  It still needs
review and testing; the patch Vitaly has posted in this thread plus 6
followups, hopefully more readable than aggregate diff.

Comments?

BTW, there's something fishy in bio_copy_user_iov().  If the area we'd asked for
had been too large for a single bio, we are going to create a bio and have
 bio_add_pc_page() eventually fill it up to limit.  Then we return into
__blk_rq_map_user_iov(), advance iter and call bio_copy_user_iov() again.
Fine, but... now we might have non-zero iter->iov_offset.  And this
        bmd->is_our_pages = map_data ? 0 : 1;
        memcpy(bmd->iov, iter->iov, sizeof(struct iovec) * iter->nr_segs);
        iov_iter_init(&bmd->iter, iter->type, bmd->iov,
                        iter->nr_segs, iter->count);
does not even look at iter->iov_offset.  As the result, when it gets to
bio_uncopy_user(), we copy the data from each bio into the *beginning* of
the user area, overwriting that from the other bio.

At the very least, we need bmd->iter = *iter; bmd->iter.iov = bmd->iov;
instead of that iov_iter_init() in there.  I'm not sure how far back does
it go; looks like "block: support large requests in blk_rq_map_user_iov"
is the earliest possible point, but it might need more digging to make
sure.  v4.5+, if that's when the problems began...

Anyway, I'd added the obvious fix to #work.iov_iter, reordered it and
force-pushed the result.