On Wed, Mar 17, 2021 at 9:32 AM Arnd Bergmann <arnd@xxxxxxxx> wrote: > > > > > <linux@xxxxxxxxxxxxxxx> wrote: > > > > > > On Tue, Mar 16, 2021 at 04:44:45PM +0100, Arnd Bergmann wrote: > > > > > > > On Tue, Mar 16, 2021 at 11:17 AM Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote: > > > > > > > > The compiler is gcc version 10.2.1 20210110 (Debian 10.2.1-6) > > > > > > > > > > > > > > Ok, building with Ubuntu 10.2.1-1ubuntu1 20201207 locally, that's > > > > > > > the closest I have installed, and I think the Debian and Ubuntu versions > > > > > > > are generally quite close in case of gcc since they are maintained by > > > > > > > the same packagers. > > > > > > > > > > > > ... which shouldn't be a problem - that's just over 1/4 of the stack > > > > > > space. Could it be the syzbot's gcc is doing something weird and > > > > > > inflating the stack frames? > > > > > > > > > > It's possible, I think that's really unlikely given that it's just Debian's > > > > > gcc, which is as close to mainline as the version I was using. > > > > > > > > > > Uwe's DEBUG_STACKOVERFLOW patch from a while ago might > > > > > help if this was the problem though: > > > > > https://lore.kernel.org/linux-arm-kernel/20200108082913.29710-1-u.kleine-koenig@xxxxxxxxxxxxxx/ > > > > > > > > > > My best guess is something going wrong in the interrupt > > > > > that triggered the preempt_schedule() which ended up calling > > > > > task_stack_end_corrupted() in schedule_debug(), as you suggested > > > > > earlier. > > > > > > > > FWIW I see slightly larger frames with the config: > > > > > > > > 073ab64 <ima_calc_field_array_hash_tfm>: > > > > 8073ab64: e1a0c00d mov ip, sp > > > > 8073ab68: e92ddff0 push {r4, r5, r6, r7, r8, r9, sl, > > > > fp, ip, lr, pc} > > > > 8073ab6c: e24cb004 sub fp, ip, #4 > > > > 8073ab70: e24ddfa7 sub sp, sp, #668 ; 0x29c > > > > > > Yes, this is the one that the compiler complained about when warning > > > for stack over 600 bytes. It's not called in this call chain though. > > > > > > > page_alloc can also do reclaim, I had the impression that reclaim can > > > > be quite heavy-weight in all respects. > > > > > > Yes, that is another possibility. What writable file systems or swap > > > do you normally have mounted that it could be writing to, and on > > > what storage device? > > > > The root fs is ext4 on virtio-blk. > > > > There are also several dozens of shrinkers that can be called during reclaim: > > https://elixir.bootlin.com/linux/latest/C/ident/unregister_shrinker > > Right, unfortunately I don't see a smoking gun there either, unless you are > also using NFS or devicemapper. > > Implementing VMAP_STACK as you suggested earlier is probably the > best way to figure out if there is an actual overrun of the stack. > Alternatively, adding support for GCC_PLUGIN_STACKLEAK might > also help find out if we ever get close to the limit. This is probably > less work, but it might not actually help in this case. VMAP_STACK is quite intrusive as far as I understand. For KASAN I considered a simpler option: have a debug config that allocates an extra page after the stack and mprotect's it. It wastes a physical page per task (fine for a debug config), but I would assume should be radically simpler to implement. In the end somebody implemented proper VMAP_STACK support for KASAN, but I still think it may be a reasonable compromise between time investment and value.
