Re: [PATCH v3] powerpc: implement CONFIG_DEBUG_VIRTUAL

[Christophe Leroy]

Le 12/12/2018 Ã 01:23, Michael Ellerman a ÃcritÂ:
> Christophe Leroy <christophe.leroy@xxxxxx> writes:
>
> > This patch implements CONFIG_DEBUG_VIRTUAL to warn about
> > incorrect use of virt_to_phys() and page_to_phys()
> >
> > Below is the result of test_debug_virtual:
> >
> > [    1.438746] WARNING: CPU: 0 PID: 1 at ./arch/powerpc/include/asm/io.h:808 test_debug_virtual_init+0x3c/0xd4
> > [    1.448156] CPU: 0 PID: 1 Comm: swapper Not tainted 4.20.0-rc5-00560-g6bfb52e23a00-dirty #532
> > [    1.457259] NIP:  c066c550 LR: c0650ccc CTR: c066c514
> > [    1.462257] REGS: c900bdb0 TRAP: 0700   Not tainted  (4.20.0-rc5-00560-g6bfb52e23a00-dirty)
> > [    1.471184] MSR:  00029032 <EE,ME,IR,DR,RI>  CR: 48000422  XER: 20000000
> > [    1.477811]
> > [    1.477811] GPR00: c0650ccc c900be60 c60d0000 00000000 006000c0 c9000000 00009032 c7fa0020
> > [    1.477811] GPR08: 00002400 00000001 09000000 00000000 c07b5d04 00000000 c00037d8 00000000
> > [    1.477811] GPR16: 00000000 00000000 00000000 00000000 c0760000 c0740000 00000092 c0685bb0
> > [    1.477811] GPR24: c065042c c068a734 c0685b8c 00000006 00000000 c0760000 c075c3c0 ffffffff
> > [    1.512711] NIP [c066c550] test_debug_virtual_init+0x3c/0xd4
> > [    1.518315] LR [c0650ccc] do_one_initcall+0x8c/0x1cc
> > [    1.523163] Call Trace:
> > [    1.525595] [c900be60] [c0567340] 0xc0567340 (unreliable)
> > [    1.530954] [c900be90] [c0650ccc] do_one_initcall+0x8c/0x1cc
> > [    1.536551] [c900bef0] [c0651000] kernel_init_freeable+0x1f4/0x2cc
> > [    1.542658] [c900bf30] [c00037ec] kernel_init+0x14/0x110
> > [    1.547913] [c900bf40] [c000e1d0] ret_from_kernel_thread+0x14/0x1c
> > [    1.553971] Instruction dump:
> > [    1.556909] 3ca50100 bfa10024 54a5000e 3fa0c076 7c0802a6 3d454000 813dc204 554893be
> > [    1.564566] 7d294010 7d294910 90010034 39290001 <0f090000> 7c3e0b78 955e0008 3fe0c062
> > [    1.572425] ---[ end trace 6f6984225b280ad6 ]---
> > [    1.577467] PA: 0x09000000 for VA: 0xc9000000
> > [    1.581799] PA: 0x061e8f50 for VA: 0xc61e8f50
> >
> > Signed-off-by: Christophe Leroy <christophe.leroy@xxxxxx>
> > ---
> >   v3: Added missing linux/mm.h
> >       I realised that a driver may use DMA on stack after checking with virt_addr_valid(), so the new
> >       verification might induce false positives. I remove it for now, will add it again later in a more
> >       controled way.
>
> What is this comment referring to?
>
> I can't see any difference to v2 except the linux/mm.h include.

v2 was:


@@ -804,6 +806,11 @@ extern void __iounmap_at(void *ea, unsigned long size);
  */
 static inline unsigned long virt_to_phys(volatile void * address)
 {
+	if (IS_ENABLED(CONFIG_DEBUG_VIRTUAL) &&
+	    !WARN_ON(IS_ENABLED(CONFIG_HAVE_ARCH_VMAP_STACK) && current->pid &&
+		     object_is_on_stack((const void*)address)))
+			WARN_ON(!virt_addr_valid(address));
+
 	return __pa((unsigned long)address);
 }


v3 is: (same as v1)


@@ -804,6 +806,8 @@ extern void __iounmap_at(void *ea, unsigned long size);
  */
 static inline unsigned long virt_to_phys(volatile void * address)
 {
+	WARN_ON(IS_ENABLED(CONFIG_DEBUG_VIRTUAL) && !virt_addr_valid(address));
+
 	return __pa((unsigned long)address);
 }


The idea in v2 was to detect objects on stack used for DMA before 
activating CONFIG_VMAP_STACK, but if the driver uses virt_addr_valid() 
to decide if it can DMA map it, then we'll get false positives.
So I think this should be added with a dedicated DEBUG CONFIG option, 
not implicitely.

Christophe