On Wed, Nov 11, 2020 at 02:27:02PM -0800, Linus Torvalds wrote: > On Wed, Nov 11, 2020 at 2:21 PM Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote: > > > > Something like below (build-tested only): > > Apart from my usual "oh, Gods, the iter model really does confuse me" > this looks more like what I expected, yes. > > Considering the original bug, I'm clearly not the only one confused by > the iov_iter helper functions and the rules.. copy_to_iter() returns the amount it has actually copied, that's all; the cause of that bug is not the primitives used, it's the rules for ->read_iter(). The rules are actually fairly simple - "->read_iter() should not report less data than it has actually left there". For read(2) it's a matter of QoI - if we hit an unmapped page, POSIX pretty much says that all bets are off; read(fd, unmapped - 5, 8) might copy 5 bytes and return 4. It is allowed (and read(2) on those files used to do just that), but it's nicer not to do so. For generic_file_splice_read(), OTOH, it's a bug - we end up with stray data spewed into pipe. So converting to ->read_iter() needs some care. Probably something along those lines should go into D/f/porting...
