Re: khugepaged might_sleep() warn due to CONFIG_HIGHPTE=y

From: Jason Gunthorpe
Date: Tue Oct 29 2019 - 19:15:23 EST


On Tue, Oct 29, 2019 at 10:15:13PM +0200, Ville SyrjÃlà wrote:
> Hi,
>
> I got some khugepaged spew on a 32bit x86:
>
> [ 217.490026] BUG: sleeping function called from invalid context at include/linux/mmu_notifier.h:346
> [ 217.492826] in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 25, name: khugepaged
> [ 217.495589] INFO: lockdep is turned off.
> [ 217.498371] CPU: 1 PID: 25 Comm: khugepaged Not tainted 5.4.0-rc5-elk+ #206
> [ 217.501233] Hardware name: System manufacturer P5Q-EM/P5Q-EM, BIOS 2203 07/08/2009
> [ 217.501697] Call Trace:
> [ 217.501697] dump_stack+0x66/0x8e
> [ 217.501697] ___might_sleep.cold.96+0x95/0xa6
> [ 217.501697] __might_sleep+0x2e/0x80
> [ 217.501697] collapse_huge_page.isra.51+0x5ac/0x1360
> [ 217.501697] ? __alloc_pages_nodemask+0xec/0xf80
> [ 217.501697] ? __alloc_pages_nodemask+0x191/0xf80
> [ 217.501697] ? trace_hardirqs_on+0x4a/0xf0
> [ 217.501697] khugepaged+0x9a9/0x20f0
> [ 217.501697] ? _raw_spin_unlock+0x21/0x30
> [ 217.501697] ? trace_hardirqs_on+0x4a/0xf0
> [ 217.501697] ? wait_woken+0xa0/0xa0
> [ 217.501697] kthread+0xf5/0x110
> [ 217.501697] ? collapse_pte_mapped_thp+0x3b0/0x3b0
> [ 217.501697] ? kthread_create_worker_on_cpu+0x20/0x20
> [ 217.501697] ret_from_fork+0x2e/0x38
>
> Looks like it's due to CONFIG_HIGHPTE=y pte_offset_map()->kmap_atomic() vs.
> mmu_notifier_invalidate_range_start().
>
> My naive idea would be to just reorder those things, but not sure
> if there's some magic ordering constraint here. At least the machine
> still boots when I do it :)
>
> diff --git a/mm/khugepaged.c b/mm/khugepaged.c
> index 0a1b4b484ac5..f05d27b7183d 100644
> +++ b/mm/khugepaged.c
> @@ -1028,12 +1028,13 @@ static void collapse_huge_page(struct mm_struct *mm,
>
> anon_vma_lock_write(vma->anon_vma);
>
> - pte = pte_offset_map(pmd, address);
> - pte_ptl = pte_lockptr(mm, pmd);
> -
> mmu_notifier_range_init(&range, MMU_NOTIFY_CLEAR, 0, NULL, mm,
> address, address + HPAGE_PMD_SIZE);
> mmu_notifier_invalidate_range_start(&range);
> +
> + pte = pte_offset_map(pmd, address);
> + pte_ptl = pte_lockptr(mm, pmd);
> +

Since pte and pte_ptl don't leak into invalidate_range_start this
seems reasonable to me..

Good catch with the new debugging!

Jason