When handle MADV_WILLNEED in madvise(), the soflockup may be occurred
in swapin_walk_pmd_entry() if swapin lots of memory on slow device,
add a cond_resched() into it to avoid the possible softlockup.
Fixes: 1998cc048901 ("mm: make madvise(MADV_WILLNEED) support swap file prefetch")
Signed-off-by: Kefeng Wang <[email protected]>
---
mm/madvise.c | 1 +
1 file changed, 1 insertion(+)
diff --git a/mm/madvise.c b/mm/madvise.c
index b913ba6efc10..fea589d8a2fb 100644
--- a/mm/madvise.c
+++ b/mm/madvise.c
@@ -226,6 +226,7 @@ static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
put_page(page);
}
swap_read_unplug(splug);
+ cond_resched();
return 0;
}
--
2.35.3
On Mon, 5 Dec 2022 22:03:27 +0800 Kefeng Wang <[email protected]> wrote:
> When handle MADV_WILLNEED in madvise(), the soflockup may be occurred
> in swapin_walk_pmd_entry() if swapin lots of memory on slow device,
> add a cond_resched() into it to avoid the possible softlockup.
>
> ...
>
> --- a/mm/madvise.c
> +++ b/mm/madvise.c
> @@ -226,6 +226,7 @@ static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
> put_page(page);
> }
> swap_read_unplug(splug);
> + cond_resched();
>
> return 0;
> }
I wonder if this would be better in walk_pmd_range(), to address other
very large walk attempts.
On 2022/12/6 5:03, Andrew Morton wrote:
> On Mon, 5 Dec 2022 22:03:27 +0800 Kefeng Wang <[email protected]> wrote:
>
>> When handle MADV_WILLNEED in madvise(), the soflockup may be occurred
>> in swapin_walk_pmd_entry() if swapin lots of memory on slow device,
>> add a cond_resched() into it to avoid the possible softlockup.
>>
>> ...
>>
>> --- a/mm/madvise.c
>> +++ b/mm/madvise.c
>> @@ -226,6 +226,7 @@ static int swapin_walk_pmd_entry(pmd_t *pmd, unsigned long start,
>> put_page(page);
>> }
>> swap_read_unplug(splug);
>> + cond_resched();
>>
>> return 0;
>> }
> I wonder if this would be better in walk_pmd_range(), to address other
> very large walk attempts.
mm/madvise.c:287: walk_page_range(vma->vm_mm, start, end,
&swapin_walk_ops, vma);
mm/madvise.c:514: walk_page_range(vma->vm_mm, addr, end,
&cold_walk_ops, &walk_private);
mm/madvise.c:762: walk_page_range(vma->vm_mm, range.start, range.end,
mm/madvise.c-763- &madvise_free_walk_ops, &tlb);
The cold_walk_ops and madvise_free_walk_ops are already with cond_resched()
in theirs pmd_entry walk, maybe there's no need for a precautionary increase
a cond_resched() for now
>
> .