On Wed, Jun 24, 2020 at 11:20:26AM -0700, Linus Torvalds wrote: > On Wed, Jun 24, 2020 at 11:14 AM Christoph Hellwig <h...@lst.de> wrote: > > > > So we'd need new user copy functions for just those cases > > No. We'd open-code them. They'd look at "oh, I'm supposed to use a > kernel pointer" and just use those. > > IOW, basically IN THE CODE that cares (and the whole argument is that > this code is one or two special cases) you do > > /* This has not been converted to the new world order */ > if (get_fs() == KERNEL_DS) memcpy(..) else copy_from_user(); > > You're overdesigning things. You're making them more complex than they > need to be.
I wish it was so simple. I really don't like overdesigns, trust me. But please take a look at setsockopt and all the different instances (count 90 .setsockopt wireups, and they then branch out into various subroutines as well). I really don't want to open code that there, but we could do helper specific to setsockopt. Honestly my preference would be to say that no eBPF isn't actually a user API and just rip out the crap added to it, but I fear that is not an option. Because in that case we'd basically be done. > Basically, I do *NOT* want to pollute the VFS layer with new > interfaces that shouldn't exist in the long run. I'd much rather make > the eventual goal be to get rid of 'read/write' entirely in favour of > the 'iter' things, but what I absolutely do *NOT* want to see is to > make a _third_ interface for reading and writing. Quite the reverse. > We should strive to make it a _single_ interface, not add a new one. Completele agreement on this. I actually hate the new fops, and only added them reluctantly as I mis-interpreted what you said.
