Re: [BUG -tip] kmemleak and stacktrace cause page faul

[Cyrill Gorcunov]

On Wed, Oct 23, 2019 at 03:21:05PM +0200, Thomas Gleixner wrote:
> On Tue, 22 Oct 2019, Cyrill Gorcunov wrote:
> > On Tue, Oct 22, 2019 at 05:23:25PM +0300, Cyrill Gorcunov wrote:
> > > 
> > > I presume the kmemleak tries to save stack trace too early when estack_pages are not
> > > yet filled.
> > 
> > Indeed, at this stage of boot the percpu_setup_exception_stacks has not been called
> > yet and estack_pages full of crap
> > 
> > [    0.157502] stk 0x1008 k 1 begin 0x0 end 0xd000 estack_pages 0xffffffff82014880 ep 0xffffffff82014888
> > [    0.159395] estack_pages[0] = 0x0
> > [    0.160046] estack_pages[1] = 0x5100000001000
> > [    0.160881] estack_pages[2] = 0x0
> > [    0.161530] estack_pages[3] = 0x6100000003000
> > [    0.162343] estack_pages[4] = 0x0
> > [    0.162962] estack_pages[5] = 0x0
> > [    0.163523] estack_pages[6] = 0x0
> > [    0.164065] estack_pages[7] = 0x8100000007000
> > [    0.164978] estack_pages[8] = 0x0
> > [    0.165624] estack_pages[9] = 0x9100000009000
> > [    0.166448] estack_pages[10] = 0x0
> > [    0.167064] estack_pages[11] = 0xa10000000b000
> > [    0.168055] estack_pages[12] = 0x0
> 
> Errm. estack_pages is statically initialized and it's an array of:.
> 
> struct estack_pages {
>         u32     offs;
>         u16     size;
>         u16     type;
> };
> 
> [0,2,4,5,6,8,10,12] are guard pages so 0 is not that crappy at all

Wait, Thomas, I might be wrong, but per-cpu is initialized to the pointer,
the memory for this estack_pages has not yet been allocated, no?

> The rest looks completely valid if you actually decode it proper.

The diff I made to fetch the values are

diff --git a/arch/x86/kernel/dumpstack_64.c b/arch/x86/kernel/dumpstack_64.c
index 753b8cfe8b8a..bf0d755b6079 100644
--- a/arch/x86/kernel/dumpstack_64.c
+++ b/arch/x86/kernel/dumpstack_64.c
@@ -101,8 +101,18 @@ static bool in_exception_stack(unsigned long *stack, struct stack_info *info)
 
 	/* Calc page offset from start of exception stacks */
 	k = (stk - begin) >> PAGE_SHIFT;
+
 	/* Lookup the page descriptor */
 	ep = &estack_pages[k];
+
+	printk("stk 0x%lx k %u begin 0x%lx end 0x%lx estack_pages 0x%lx ep 0x%lx\n",
+	       stk, k, begin, end, (long)(void *)&estack_pages[0], (long)(void *)ep);
+
+	for (k = 0; k < CEA_ESTACK_PAGES; k++) {
+		long v = *(long *)(void *)&estack_pages[k];
+		printk("estack_pages[%d] = 0x%lx\n", k, v);
+	}
+
 	/* Guard page? */
 	if (!ep->size)
 		return false;


> 
> e.g. 0x51000 00001000
> 
>      bit  0-31: 00001000		Offset 0x1000: 1 Page
>      bit 32-47: 1000			Size 0x1000:   1 Page
>      bit 48-63: 5			Type 5: STACK_TYPE_EXCEPTION + ESTACK_DF
> 
> So, no. This is NOT the problem.

I drop the left of your reply. True, I agreed with anything you said.
You know I didn't manage to dive more into this problem yesterday
but if time permits I'll continue today. It is easily triggering
under kvm (the kernel I'm building is almost without modules so
I simply upload bzImage into the guest). FWIW, the config I'm
using is https://gist.github.com/cyrillos/7cd5d2510a99af8ea872f07ac6f9095b