Re: [PATCHv5 perf/core 09/22] uprobes/x86: Add uprobe syscall to speed up uprobe

[Masami Hiramatsu (Google) <mhiramat@kernel.org>]

On Tue, 15 Jul 2025 14:16:33 +0200
Jiri Olsa <olsajiri@gmail.com> wrote:

> On Tue, Jul 15, 2025 at 08:54:51AM +0900, Masami Hiramatsu wrote:
> > On Mon, 14 Jul 2025 23:28:58 +0200
> > Jiri Olsa <olsajiri@gmail.com> wrote:
> > 
> > > On Mon, Jul 14, 2025 at 07:19:35PM +0900, Masami Hiramatsu wrote:
> > > > On Mon, 14 Jul 2025 11:39:03 +0200
> > > > Peter Zijlstra <peterz@infradead.org> wrote:
> > > > 
> > > > > On Mon, Jul 14, 2025 at 05:39:15PM +0900, Masami Hiramatsu wrote:
> > > > > 
> > > > > > > +	/*
> > > > > > > +	 * Some of the uprobe consumers has changed sp, we can do nothing,
> > > > > > > +	 * just return via iret.
> > > > > > > +	 */
> > > > > > 
> > > > > > Do we allow consumers to change the `sp`? It seems dangerous
> > > > > > because consumer needs to know whether it is called from
> > > > > > breakpoint or syscall. Note that it has to set up ax, r11
> > > > > > and cx on the stack correctly only if it is called from syscall,
> > > > > > that is not compatible with breakpoint mode.
> > > > > > 
> > > > > > > +	if (regs->sp != sp)
> > > > > > > +		return regs->ax;
> > > > > > 
> > > > > > Shouldn't we recover regs->ip? Or in this case does consumer has
> > > > > > to change ip (== return address from trampline) too?
> > > > > > 
> > > > > > IMHO, it should not allow to change the `sp` and `ip` directly
> > > > > > in syscall mode. In case of kprobes, kprobe jump optimization
> > > > > > must be disabled explicitly (e.g. setting dummy post_handler)
> > > > > > if the handler changes `ip`.
> > > > > > 
> > > > > > Or, even if allowing to modify `sp` and `ip`, it should be helped
> > > > > > by this function, e.g. stack up the dummy regs->ax/r11/cx on the
> > > > > > new stack at the new `regs->sp`. This will allow modifying those
> > > > > > registries transparently as same as breakpoint mode.
> > > > > > In this case, I think we just need to remove above 2 lines.
> > > > > 
> > > > > There are two syscall return paths; the 'normal' is sysret and for that
> > > > > you need to undo all things just right.
> > > > > 
> > > > > The other is IRET. At which point we can have whatever state we want,
> > > > > including modified SP.
> > > > > 
> > > > > See arch/x86/entry/syscall_64.c:do_syscall_64() and
> > > > > arch/x86/entry/entry_64.S:entry_SYSCALL_64
> > > > > 
> > > > > The IRET path should return pt_regs as is from an interrupt/exception
> > > > > very much like INT3.
> > > > 
> > > > OK, so SYSRET case, we need to follow;
> > > > 
> > > > sys_uprobe -> do_syscall_64 -> entry_SYSCALL_64 -> trampoline -> retaddr
> > > > 
> > > > But using IRET to return, we can skip returning to trampoline,
> > > > 
> > > > sys_uprobe -> do_syscall_64 -> entry_SYSCALL_64 -> regs->ip
> > > 
> > > the handler gets the original breakpoint address, it's set in:
> > > 
> > >         regs->ip  = ax_r11_cx_ip[3] - 5;
> > > 
> > > and at the point we do:
> > > 
> > >         /*
> > >          * Some of the uprobe consumers has changed sp, we can do nothing,
> > >          * just return via iret.
> > >          */
> > >         if (regs->sp != sp)
> > >                 return regs->ax;
> > > 
> > > 
> > > .. regs->ip value wasn't restored for the trampoline's return address,
> > > so iret will skip the trampoline
> > 
> > Ah, OK. So unless we restore regs->cx = regs->ip and 
> > regs->r11 = regs->flags, it automatically use IRET. Got it.
> > 
> > > 
> > > but perhaps we could do the extra check below to land on the next instruction?
> > 
> > Hmm, can you clarify the required condition of changing regs
> > in the consumers? regs->sp change need to be handled by the
> > IRET, but other changes can be handled by trampoline. Is that
> > correct?
> 
> yes,
> if handler changes regs->sp we return through iret
> if handler changes regs->ip (the only other tricky one IIUC), we return through
> the trampoline and jump to regs->ip via trampoline's 'ret' instruction


Got it. Thanks for confirming it.

Thank you,

> 
> jirka


-- 
Masami Hiramatsu (Google) <mhiramat@kernel.org>