On Thu, Oct 10, 2019 at 05:31:13PM -0700, Linus Torvalds wrote: > So the code actually needs to properly return the error early, or > initialize the segments that didn't get loaded to 0, or something. > > And when I posted that, Luto said "just get rid of the get_user_ex() > entirely, instead of changing semantics of the existing ones to be > sane. > > Which is probably right. There aren't that many. > > I *thought* there were also cases of us doing some questionably things > inside the get_user_try sections, but those seem to have gotten fixed > already independently, so it's really just the "make try/catch really > try/catch" change that needs some editing of our current broken stuff > that depends on it not actually *catching* exceptions, but on just > continuing on to the next one. Umm... TBH, I wonder if we would be better off if restore_sigcontext() (i.e. sigreturn()/rt_sigreturn()) would flat-out copy_from_user() the entire[*] struct sigcontext into a local variable and then copied fields to pt_regs... The thing is small enough for not blowing the stack (256 bytes max. and it's on a shallow stack) and big enough to make "fancy memcpy + let the compiler think how to combine in-kernel copies" potentially better than hardwired sequence of 64bit loads/stores... [*] OK, sans ->reserved part in the very end on 64bit. 192 bytes to copy. Same for do_sys_vm86(), perhaps - we want regs/flags/cpu_type and screen_bitmap there, i.e. the beginning of struct vm86plus_struct and of struct vm86_struct... 24*32bit. IOW, 96-byte memcpy + gcc-visible field-by-field copying vs. hardwired sequence of 32bit loads (with some 16bit ones thrown in, for extra fun) and compiler told not to reorder anything. And these (32bit and 64bit restore_sigcontext() and do_sys_vm86()) are the only get_user_ex() users anywhere...
