Re: [syzbot] KASAN: vmalloc-out-of-bounds Read in __bpf_prog_put

On Mon, 24 Jan 2022 at 12:42, Vegard Nossum [off-list ref] wrote:

On Thu, 20 Jan 2022 at 15:17, syzbot
[off-list ref] wrote:

quoted

syzbot suspects this issue was fixed by commit:

commit 218d747a4142f281a256687bb513a135c905867b
Author: John Fastabend [off-list ref]
Date:   Tue Jan 4 21:46:45 2022 +0000

    bpf, sockmap: Fix double bpf_prog_put on error case in map_link

I can confirm the above commit fixes the issue, but it references a
slightly different report. Looks like the only difference is
__bpf_prog_put instead of bpf_prog_put:

KASAN: vmalloc-out-of-bounds Read in __bpf_prog_put
KASAN: vmalloc-out-of-bounds Read in bpf_prog_put

However, looking at the stack traces for the two bugs shows that
__bpf_prog_put() is really the location for both reports, see:

https://syzkaller.appspot.com/bug?id=797cd651dd0d9bd921e4fa51b792f5afdc3f390f
 kasan_report.cold+0x83/0xdf mm/kasan/report.c:450 mm/kasan/report.c:450
 __bpf_prog_put.constprop.0+0x1dd/0x220 kernel/bpf/syscall.c:1812
kernel/bpf/syscall.c:1812
 bpf_prog_put kernel/bpf/syscall.c:1829 [inline]
 bpf_prog_put kernel/bpf/syscall.c:1829 [inline] kernel/bpf/syscall.c:1837

vs.

https://syzkaller.appspot.com/bug?extid=bb73e71cf4b8fd376a4f
 kasan_report+0x19a/0x1f0 mm/kasan/report.c:450 mm/kasan/report.c:450
 __bpf_prog_put kernel/bpf/syscall.c:1812 [inline]
 __bpf_prog_put kernel/bpf/syscall.c:1812 [inline] kernel/bpf/syscall.c:1829
 bpf_prog_put+0x8c/0x4f0 kernel/bpf/syscall.c:1829 kernel/bpf/syscall.c:1829

Looks to me like the compiler's inlining decision caused syzbot to see
__bpf_prog_put() instead of bpf_prog_put(), but I can't tell if it's
because it got inlined or because of the .constprop.0 suffix... I
guess syzbot skips the [inline] entries when deciding which function
to report the bug in?

Not sure if you are still interested in this or not...
But, yes, it's inline frames that are a problem, ".constprop.0" should
be stripped.
syzkaller parses non-symbolized kernel output w/o inlined frames to
extract the title. This was a very early decision, not sure if it's
the right one or not. On the other hand using inline frames can cause
attribution to all the common one-liners.
Now it's somewhat hard to change b/c if we change it, new crashes will
be parsed differently and it will cause a storm of duplicates for
already reported bugs.

In any case:

#syz dup: KASAN: vmalloc-out-of-bounds Read in bpf_prog_put

Vegard