Re: [PATCH] x86: KPROBE_ENTRY should be paired wth KPROBE_END

[Cyrill Gorcunov <gorcunov@gmail.com>]

[Ingo Molnar - Sun, Nov 23, 2008 at 03:55:54PM +0100]
| 
| * Cyrill Gorcunov <gorcunov@gmail.com> wrote:
| 
| > [Cyrill Gorcunov - Sun, Nov 23, 2008 at 04:51:34PM +0300]
| > | [Ingo Molnar - Sun, Nov 23, 2008 at 02:27:52PM +0100]
| > | | 
| > | | * Alexander van Heukelum <heukelum@mailshack.com> wrote:
| > | | 
| > | | > Impact: moves some code out of .kprobes.text
| > | | > 
| > | | > KPROBE_ENTRY switches code generation to .kprobes.text, and KPROBE_END
| > | | > uses .popsection to get back to the previous section (.text, normally).
| > | | > Also replace ENDPROC by END, for consistency.
| > | | > 
| > | | > Signed-off-by: Alexander van Heukelum <heukelum@fastmail.fm>
| > | | 
| > | | applied to tip/x86/irq, thanks Alexander!
| > | | 
| > | | > One more small change for today. The xen-related functions 
| > | | > xen_do_hypervisor_callback and xen_failsafe_callback are put
| > | | > in the .kprobes.text even in the current kernel: ignore_sysret
| > | | > is enclosed in KPROBE_ENTRY / ENDPROC, instead of KPROBE_ENTRY /
| > | | > KPROBE_END, but I guess the situation is harmless.
| > | | 
| > | | yeah. It narrows no-kprobes protection for that code, but it should 
| > | | indeed be fine (and that's the intention as well).
| > | | 
| > | | Note that this is a reoccuring bug type, and rather long-lived. Can 
| > | | you think of any way to get automated nesting protection of both the 
| > | | .cfi_startproc/endproc macros and kprobes start/end? A poor man's 
| > | | solution would be to grep the number of start and end methods and 
| > | | enforce that they are equal.
| > | | 
| > | | 	Ingo
| > | | 
| > | 
| > | I think we could play with preprocessor and check if ENTRY/END matches.
| > | Looking now.
| > | 
| > | 		- Cyrill -
| > 
| > Here is what I've done
| > 
| > 1) Add some macros like:
| > 
| > 	.macro __set_entry
| > 	.set _ENTRY_IN, 1
| > 	.endm
| > 
| > 	.macro __unset_entry
| > 	.set _ENTRY_IN, 0
| > 	.endm
| > 
| > 	.macro __check_entry
| > 	.ifeq _ENTRY_IN
| > 	.error "END should be used"
| > 	.abort
| > 	.endif
| > 	.endm
| > 
| > So the code
| > 
| > ENTRY(mcount)
| > 	__unset_entry
| > 	retq
| > 	__check_entry
| > END(mcount)
| > 
| > will fail like
| > 
| > cyrill@lenovo linux-2.6.git $ make arch/x86/kernel/entry_64.o
| >   CHK     include/linux/version.h
| >   CHK     include/linux/utsrelease.h
| >   SYMLINK include/asm -> include/asm-x86
| >   CALL    scripts/checksyscalls.sh
| >   AS      arch/x86/kernel/entry_64.o
| > arch/x86/kernel/entry_64.S: Assembler messages:
| > arch/x86/kernel/entry_64.S:84: Error: END should be used
| > arch/x86/kernel/entry_64.S:84: Fatal error: .abort detected.  Abandoning ship.
| > make[1]: *** [arch/x86/kernel/entry_64.o] Error 1
| > make: *** [arch/x86/kernel/entry_64.o] Error 2
| > cyrill@lenovo linux-2.6.git $ 
| > 
| > So if such an approach is acceptable (in general) -- I could take a 
| > more deeper look. So every ENTRY would check if other ENTRY/KPROBE 
| > is active and report that.
| 
| looks good!
| 
| Can we somehow detect a missing .cfi_endproc? That's another pattern 
| i've seen.
| 
| 	Ingo
| 

As only I finish with this pattern we could add anything we need :)
Will keep you in touch.

		- Cyrill -