Re: [pcpu] BUG: KASAN: use-after-scope in pcpu_setup_first_chunk+0x1e3b/0x29e2

From: Fengguang Wu
Date: Tue Nov 28 2017 - 07:45:44 EST



Hi Dennis,

On Mon, Nov 27, 2017 at 03:03:01PM -0600, Dennis Zhou wrote:
Hi Fangguang,

On Sun, Nov 26, 2017 at 02:31:17PM +0800, Fengguang Wu wrote:
Hello,

FYI this happens in mainline kernel 4.14.0-13151-g5a78775.
This looks like a new regression after 4.14.

I have reproduced this with the commit and scripts attached. The
offending line is the INIT_LIST_HEAD call in the for loop below. Both
pcpu_nr_slots and pcpu_slot are global variables.

pcpu_nr_slots = __pcpu_size_to_slot(pcpu_unit_size) + 2;
pcpu_slot = memblock_virt_alloc(
pcpu_nr_slots * sizeof(pcpu_slot[0]), 0);
for (i = 0; i < pcpu_nr_slots; i++)
INIT_LIST_HEAD(&pcpu_slot[i]);

The management of the percpu slots was not changed with the bitmap
allocator in v4.14. That line actually is from Tejun's rewrite in 2009,
commit fbf59bc9d.

Just to be safe, I did revert my commits and reproduced the same error.

GCC version: gcc-7 (Ubuntu 7.2.0-1ubuntu1~16.04) 7.2.0

[ 0.000000] ==================================================================
[ 0.000000] BUG: KASAN: use-after-scope in pcpu_setup_first_chunk+0x1e3b/0x29e2:
pcpu_setup_first_chunk at mm/percpu.c:2118 (discriminator 3)
[ 0.000000] Write of size 8 at addr ffffffff83c07d38 by task swapper/0
[ 0.000000]
[ 0.000000] CPU: 0 PID: 0 Comm: swapper Not tainted 4.14.0-13151-g5a78775 #2
[ 0.000000] Call Trace:
[ 0.000000] print_address_description+0x2d/0x3d0:
print_address_description at mm/kasan/report.c:253
[ 0.000000] ? pcpu_setup_first_chunk+0x1e3b/0x29e2:
pcpu_setup_first_chunk at mm/percpu.c:2118 (discriminator 3)
[ 0.000000] kasan_report+0x1f4/0x3b0:
kasan_report_error at mm/kasan/report.c:352
(inlined by) kasan_report at mm/kasan/report.c:409
[ 0.000000] pcpu_setup_first_chunk+0x1e3b/0x29e2:
pcpu_setup_first_chunk at mm/percpu.c:2118 (discriminator 3)

Interestingly, while the INIT_LIST_HEAD call is what is causing
problems, I manually unrolled the loop and the location of the bug
changed. This happens both inside and outside of the for loop as long as
it only has 1 iteration.

[ 0.000000] ==================================================================
[ 0.000000] BUG: KASAN: use-after-scope in pageset_init (/users/dennisz/linux-linus//include/linux/listh:28 /users/dennisz/linux-linus/mm/page_allocc:5509)
[ 0.000000] Write of size 8 at addr ffffffff83c07dc8 by task swapper/0
[ 0.000000]
[ 0.000000] CPU: 0 PID: 0 Comm: swapper Not tainted 4.14.0-13151-g5a78775-dirty #53
[ 0.000000] Call Trace:
[ 0.000000] print_address_description (/users/dennisz/linux-linus/mm/kasan/reportc:253)
[ 0.000000] ? pageset_init (/users/dennisz/linux-linus//include/linux/listh:28 /users/dennisz/linux-linus/mm/page_allocc:5509)
[ 0.000000] kasan_report (/users/dennisz/linux-linus/mm/kasan/reportc:352 /users/dennisz/linux-linus/mm/kasan/reportc:409)
[ 0.000000] pageset_init (/users/dennisz/linux-linus//include/linux/listh:28 /users/dennisz/linux-linus/mm/page_allocc:5509)
[ 0.000000] ? is_free_buddy_page (/users/dennisz/linux-linus/mm/page_allocc:5500)
[ 0.000000] build_all_zonelists_init (/users/dennisz/linux-linus/mm/page_allocc:5483 /users/dennisz/linux-linus/mm/page_allocc:5496 /users/dennisz/linux-linus/mm/page_allocc:5515 /users/dennisz/linux-linus/mm/page_allocc:5262)
[ 0.000000] build_all_zonelists (/users/dennisz/linux-linus/mm/page_allocc:5277)
[ 0.000000] start_kernel (/users/dennisz/linux-linus/init/mainc:546)
[ 0.000000] ? thread_stack_cache_init+0x2e/0x2e
[ 0.000000] ? memcpy_orig (/users/dennisz/linux-linus/arch/x86/lib/memcpy_64S:106)
[ 0.000000] secondary_startup_64 (/users/dennisz/linux-linus/arch/x86/kernel/head_64S:237)
[ 0.000000]
[ 0.000000]
[ 0.000000] Memory state around the buggy address:
[ 0.000000] ffffffff83c07c80: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
[ 0.000000] ffffffff83c07d00: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
[ 0.000000] >ffffffff83c07d80: 00 00 00 00 00 f1 f1 f1 f1 f8 f2 f2 f2 00 00 00
[ 0.000000] ^
[ 0.000000] ffffffff83c07e00: 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00 00
[ 0.000000] ffffffff83c07e80: 00 00 00 00 00 00 00 00 00 00 00 00 f8 00 00 00
[ 0.000000] ==================================================================

Next, I did a partial unroll where the loop went through 2 iterations
and then manually called INIT_LIST_HEAD for the remainder, the bug
returns. I've left this trace out as it is the same as reported by
Fangguang.

I'm not familiar with the details of KASAN, but it seems that there is a
weird interaction with KASAN use-after-scope and subsequent invocations
of inlined functions called from inside a for loop.

I don't think I have enough background to continue debugging this
myself, who would be a good person to loop in for help?

CC KASAN people. I managed to bisect it to commit f7dd2507893cc3
("gcc-plugins: structleak: add option to init all vars used as byref args")

Thanks,
Fengguang