Re: kmemleak: Unable to handle kernel paging request

On Fri, Jun 13, 2014 at 08:12:08AM +0100, Denis Kirjanov wrote:

On 6/12/14, Catalin Marinas [off-list ref] wrote:

quoted

On Thu, Jun 12, 2014 at 01:00:57PM +0100, Denis Kirjanov wrote:

quoted

On 6/12/14, Denis Kirjanov [off-list ref] wrote:

quoted

On 6/12/14, Catalin Marinas [off-list ref] wrote:

quoted

On 11 Jun 2014, at 21:04, Denis Kirjanov [off-list ref]
wrote:

quoted

On 6/11/14, Catalin Marinas [off-list ref] wrote:

quoted

On Wed, Jun 11, 2014 at 04:13:07PM +0400, Denis Kirjanov wrote:

quoted

I got a trace while running 3.15.0-08556-gdfb9454:

[  104.534026] Unable to handle kernel paging request for data at
address 0xc00000007f000000

Were there any kmemleak messages prior to this, like "kmemleak
disabled"? There could be a race when kmemleak is disabled because
of
some fatal (for kmemleak) error while the scanning is taking place
(which needs some more thinking to fix properly).

No. I checked for the similar problem and didn't find anything
relevant.
I'll try to bisect it.

Does this happen soon after boot? I guess it’s the first scan
(scheduled at around 1min after boot). Something seems to be telling
kmemleak that there is a valid memory block at 0xc00000007f000000.

Yeah, it happens after a while with a booted system so that's the
first kmemleak scan.

I've bisected to this commit: d4c54919ed86302094c0ca7d48a8cbd4ee753e92
"mm: add !pte_present() check on existing hugetlb_entry callbacks".
Reverting the commit fixes the issue

I can't figure how this causes the problem but I have more questions. Is
0xc00000007f000000 address always the same in all crashes? If yes, you
could comment out start_scan_thread() in kmemleak_late_init() to avoid
the scanning thread starting. Once booted, you can run:

  echo dump=0xc00000007f000000 > /sys/kernel/debug/kmemleak

and check the dmesg for what kmemleak knows about that address, when it
was allocated and whether it should be mapped or not.

The address is always the same.

[  179.466239] kmemleak: Object 0xc00000007f000000 (size 16777216):
[  179.466503] kmemleak:   comm "swapper/0", pid 0, jiffies 4294892300
[  179.466508] kmemleak:   min_count = 0
[  179.466512] kmemleak:   count = 0
[  179.466517] kmemleak:   flags = 0x1
[  179.466522] kmemleak:   checksum = 0
[  179.466526] kmemleak:   backtrace:
[  179.466531]      [<c000000000afc3dc>] .memblock_alloc_range_nid+0x68/0x88
[  179.466544]      [<c000000000afc444>] .memblock_alloc_base+0x20/0x58
[  179.466553]      [<c000000000ae96cc>] .alloc_dart_table+0x5c/0xb0
[  179.466561]      [<c000000000aea300>] .pmac_probe+0x38/0xa0
[  179.466569]      [<000000000002166c>] 0x2166c
[  179.466579]      [<0000000000ae0e68>] 0xae0e68
[  179.466587]      [<0000000000009bc4>] 0x9bc4

OK, so that's the DART table allocated via alloc_dart_table(). Is
dart_tablebase removed from the kernel linear mapping after allocation?
If that's the case, we need to tell kmemleak to ignore this block (see
patch below, untested). But I still can't explain how commit
d4c54919ed863020 causes this issue.

(also cc'ing the powerpc list and maintainers)

---------------8<--------------------------

From 09a7f1c97166c7bdca7ca4e8a4ff2774f3706ea3 Mon Sep 17 00:00:00 2001

From: Catalin Marinas <catalin.marinas@arm.com>
Date: Fri, 13 Jun 2014 09:44:21 +0100
Subject: [PATCH] powerpc/kmemleak: Do not scan the DART table

The DART table allocation is registered to kmemleak via the
memblock_alloc_base() call. However, the DART table is later unmapped
and dart_tablebase VA no longer accessible. This patch tells kmemleak
not to scan this block and avoid an unhandled paging request.

Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <redacted>
---
 arch/powerpc/sysdev/dart_iommu.c | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/arch/powerpc/sysdev/dart_iommu.c b/arch/powerpc/sysdev/dart_iommu.c
index 62c47bb76517..9e5353ff6d1b 100644
--- a/arch/powerpc/sysdev/dart_iommu.c
+++ b/arch/powerpc/sysdev/dart_iommu.c

@@ -476,6 +476,11 @@ void __init alloc_dart_table(void)
 	 */
 	dart_tablebase = (unsigned long)
 		__va(memblock_alloc_base(1UL<<24, 1UL<<24, 0x80000000L));
+	/*
+	 * The DART space is later unmapped from the kernel linear mapping and
+	 * accessing dart_tablebase during kmemleak scanning will fault.
+	 */
+	kmemleak_no_scan((void *)dart_tablebase);
 
 	printk(KERN_INFO "DART table allocated at: %lx\n", dart_tablebase);
 }