On Fri, Oct 28, 2022 at 11:35:06AM -0700, Linus Torvalds wrote: > > Umm... How are you going to e.g. copy from ITER_DISCARD? I've no problem > > with WARN_ON_ONCE(), but when the operation really can't be done, what > > can we do except returning an error? > > Fair enough. But it's the "people got the direction wrong, but the > code worked" case that I would want tyo make sure still works - just > with a warning. > > Clearly the ITER_DISCARD didn't work before either, but all the cases > in patches 1-10 were things that _worked_, just with entirely the > wrong ->data_source (aka iov_iter_rw()) value. > > So things like copy_to_iter() should warn if it's not a READ (or > ITER_DEST), but it should still copy into the destination described by > the iter, in order to keep broken code working. > > That's simply because I worry that your patches 1-10 didn't actually > catch every single case. I'm not actually sure how you found them all > - did you have some automation, or was it with "boot and find warnings > from the first version of patch 11/12"? Went through the callers, replaced each with the right ITER_... (there's not that many of them and they are fairly easy to review), then went through mismatches and split their fixups into the beginning of the series (READ -> ITER_SOURCE becoming READ -> WRITE -> ITER_SOURCE, that is). FWIW, there used to be one case where we really tried to copy the wrong way - fixed a couple of cycles ago (f615625a44c4 "9p: handling Rerror without copy_from_iter_full()"). No such catches this time...
