On Mon, May 06, 2019 at 10:19:51AM +0200, Peter Zijlstra wrote: > +.Lfrom_usermode_no_fixup_\@: > +.endm > + > +.macro IRET_FRAME > + > + /* orig_eax is already POP'ed when we're here */ > + > + testl $CS_FROM_KERNEL, 1*4(%esp) > + jz .Lfinished_frame_\@ > + > + pushl %eax > + >From there.. > + lea 10*4(%esp), %eax # address of <previous context> > + cmpl %eax, 4*4(%esp) # if ->sp is unmodified > + jnz .Lmodified_sp_do_fixup_\@ > + > + /* > + * Fast path; regs->sp wasn't modified, reuse the original IRET frame. > + */ > + pop %eax > + add $6*4, %esp > + jmp .Lfinished_frame_\@; > + > +.Lmodified_sp_do_fixup_\@: ... until here, needs to go, it is buggy. While a clever idea, it looses updates to regs->ip and ->flags. > + > + /* > + * Reconstruct the 3 entry IRET frame right after the (modified) > + * regs->sp without lowering %esp in between, such that an NMI in the > + * middle doesn't scribble our stack. > + */ > + pushl %ecx > + movl 5*4(%esp), %eax # (modified) regs->sp > + > + movl 4*4(%esp), %ecx # flags > + movl %ecx, -4(%eax) > + > + movl 3*4(%esp), %ecx # cs > + andl $0x0000ffff, %ecx > + movl %ecx, -8(%eax) > + > + movl 2*4(%esp), %ecx # ip > + movl %ecx, -12(%eax) > + > + movl 1*4(%esp), %ecx # eax > + movl %ecx, -16(%eax) > + > + popl %ecx > + lea -16(%eax), %esp > + popl %eax > + > +.Lfinished_frame_\@: > +.endm