On Tue, 7 May 2019 21:50:52 -0700 Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > > It's been a bane of mine for some time. > > Guys, I have basically a one-liner patch for your hangups. > > It's called "rename 'sp' to 'user_sp' on x86-32". > > Then we make the 'sp' field on x86-64 be a union, so that you can call > it user_sp or sp as you wish. > > Yeah, it's really more than one line, because obviously the users will > need chaning, but honestly, that would be a _real_ cleanup. Make the > register match what it actually is. But is it? Sure, it will be a reminder that it's different for x86-32, but that still doesn't take away the fact that pt_regs on x86_32 is an anomaly! Where else do we have part of a data structure that can't be read because it can possibly fault? If regs is a valid pointer, one would think that simply reading regs->sp (or regs->user_sp) would be no more dangerous than reading regs->ip. The difference between entry_32.S from entry_64.S causes it to spill into C code, making the x86_64 code more difficult to deal with. Sure, 10 to 15 years ago, all your arguments would make sense. But today, who uses x86_32? Perhaps you may use it in a VM, but I asked a few developers when was the last time they used one, they told me 5 to 7 years ago. I only boot x86_32 to test to make sure I didn't break it. Yes, your diffstat is really nice to the changes to entry_32.S, but at what cost? To make the x86_64 code more complex? That whole returning the regs in the int3 handler makes no sense on x86_64, but yet we would need to do it to handle x86_32. Why burden the architecture of today and tomorrow with the architecture of yesterday? x86_32 is becoming more obsolete by the day. It baffles me why we wouldn't want to contain its complexity in a single file then to spread it out like wildfire across the generic x86 code. The ®s->sp hack is just one more rung in the complex learning curve ladder of becoming a Linux kernel developer. -- Steve
