Re: [v2 PATCH] mm: hugetlb: avoid soft lockup when mprotect to large memory area

[Dev Jain <dev.jain@arm.com>]

On 30/09/25 11:38 pm, Yang Shi wrote:
>
>
> On 9/29/25 10:26 PM, Dev Jain wrote:
>>
>> On 30/09/25 1:54 am, Yang Shi wrote:
>>> When calling mprotect() to a large hugetlb memory area in our 
>>> customer's
>>> workload (~300GB hugetlb memory), soft lockup was observed:
>>>
>>> watchdog: BUG: soft lockup - CPU#98 stuck for 23s! [t2_new_sysv:126916]
>>>
>>> CPU: 98 PID: 126916 Comm: t2_new_sysv Kdump: loaded Not tainted 
>>> 6.17-rc7
>>> Hardware name: GIGACOMPUTING R2A3-T40-AAV1/Jefferson CIO, BIOS 
>>> 5.4.4.1 07/15/2025
>>> pstate: 20400009 (nzCv daif +PAN -UAO -TCO -DIT -SSBS BTYPE=--)
>>> pc : mte_clear_page_tags+0x14/0x24
>>> lr : mte_sync_tags+0x1c0/0x240
>>> sp : ffff80003150bb80
>>> x29: ffff80003150bb80 x28: ffff00739e9705a8 x27: 0000ffd2d6a00000
>>> x26: 0000ff8e4bc00000 x25: 00e80046cde00f45 x24: 0000000000022458
>>> x23: 0000000000000000 x22: 0000000000000004 x21: 000000011b380000
>>> x20: ffff000000000000 x19: 000000011b379f40 x18: 0000000000000000
>>> x17: 0000000000000000 x16: 0000000000000000 x15: 0000000000000000
>>> x14: 0000000000000000 x13: 0000000000000000 x12: 0000000000000000
>>> x11: 0000000000000000 x10: 0000000000000000 x9 : ffffc875e0aa5e2c
>>> x8 : 0000000000000000 x7 : 0000000000000000 x6 : 0000000000000000
>>> x5 : fffffc01ce7a5c00 x4 : 00000000046cde00 x3 : fffffc0000000000
>>> x2 : 0000000000000004 x1 : 0000000000000040 x0 : ffff0046cde7c000
>>>
>>> Call trace:
>>>    mte_clear_page_tags+0x14/0x24
>>>    set_huge_pte_at+0x25c/0x280
>>>    hugetlb_change_protection+0x220/0x430
>>>    change_protection+0x5c/0x8c
>>>    mprotect_fixup+0x10c/0x294
>>>    do_mprotect_pkey.constprop.0+0x2e0/0x3d4
>>>    __arm64_sys_mprotect+0x24/0x44
>>>    invoke_syscall+0x50/0x160
>>>    el0_svc_common+0x48/0x144
>>>    do_el0_svc+0x30/0xe0
>>>    el0_svc+0x30/0xf0
>>>    el0t_64_sync_handler+0xc4/0x148
>>>    el0t_64_sync+0x1a4/0x1a8
>>>
>>> Soft lockup is not triggered with THP or base page because there is
>>> cond_resched() called for each PMD size.
>>>
>>> Although the soft lockup was triggered by MTE, it should be not MTE
>>> specific. The other processing which takes long time in the loop may
>>> trigger soft lockup too.
>>>
>>> So add cond_resched() for hugetlb to avoid soft lockup.
>>>
>>> Fixes: 8f860591ffb2 ("[PATCH] Enable mprotect on huge pages")
>>> Tested-by: Carl Worth <carl@os.amperecomputing.com>
>>> Reviewed-by: Christoph Lameter (Ampere) <cl@gentwo.org>
>>> Reviewed-by: Catalin Marinas <catalin.marinas@arm.com>
>>> Acked-by: David Hildenbrand <david@redhat.com>
>>> Acked-by: Oscar Salvador <osalvador@suse.de>
>>> Reviewed-by: Anshuman Khandual <anshuman.khandual@arm.com>
>>> Signed-off-by: Yang Shi <yang@os.amperecomputing.com>
>>> ---
>>> v2: - Made the subject and commit message less MTE specific and fixed
>>>        the fixes tag.
>>>      - Collected all R-bs and A-bs.
>>>
>>>   mm/hugetlb.c | 2 ++
>>>   1 file changed, 2 insertions(+)
>>>
>>> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
>>> index cb5c4e79e0b8..fe6606d91b31 100644
>>> --- a/mm/hugetlb.c
>>> +++ b/mm/hugetlb.c
>>> @@ -7242,6 +7242,8 @@ long hugetlb_change_protection(struct 
>>> vm_area_struct *vma,
>>>                           psize);
>>>           }
>>>           spin_unlock(ptl);
>>> +
>>> +        cond_resched();
>>>       }
>>>       /*
>>>        * Must flush TLB before releasing i_mmap_rwsem: x86's 
>>> huge_pmd_unshare
>>
>> Reviewed-by: Dev Jain <dev.jain@arm.com>
>
> Thank you.
>
>>
>> Does it make sense to also do cond_resched() in the 
>> huge_pmd_unshare() branch?
>> That also amounts to clearing a page. And I can see for example, 
>> zap_huge_pmd()
>> and change_huge_pmd() consume a cond_resched().
>
> Thanks for raising this. I did think about it. But I didn't convince 
> myself because shared pmd should be not that common IMHO (If I'm 
> wrong, please feel free to correct me). At least PMD can't be shared 
> if the memory is tagged IIRC. So I'd like to keep the patch minimal 
> for now and defer adding cond_resched() until it is hit by some real 
> life workload.

If we have large swathes of hugetlb memory like in your workload, and it 
is MAP_SHARED, then there should be high chances of sharing the PMD. 
Although, I incorrectly

observed that we are clearing a page there - we are only clearing the 
pud entry which is 8 bytes. So yes a soft lockup should be highly 
unlikely. But since cond_resched()

is cheap (I assume this is the case since it is liberally sprinkled all 
over the codebase) I think we should be consistent. Probably not an 
immediate concern and not a matter

of this patch.


>
> Yang
>
>