Re: [PATCH v4 4/5] mm/shmem: fix infinite loop when swap in shmem error at swapoff time

[Miaohe Lin <linmiaohe@huawei.com>]

On 2022/5/20 14:34, HORIGUCHI NAOYA(堀口 直也) wrote:
> On Thu, May 19, 2022 at 08:50:29PM +0800, Miaohe Lin wrote:
>> When swap in shmem error at swapoff time, there would be a infinite loop
>> in the while loop in shmem_unuse_inode(). It's because swapin error is
>> deliberately ignored now and thus info->swapped will never reach 0. So
>> we can't escape the loop in shmem_unuse().
>>
>> In order to fix the issue, swapin_error entry is stored in the mapping
>> when swapin error occurs. So the swapcache page can be freed and the
>> user won't end up with a permanently mounted swap because a sector is
>> bad. If the page is accessed later, the user process will be killed
>> so that corrupted data is never consumed. On the other hand, if the
>> page is never accessed, the user won't even notice it.
>>
>> Reported-by: Naoya Horiguchi <naoya.horiguchi@nec.com>
>> Signed-off-by: Miaohe Lin <linmiaohe@huawei.com>
> 
> Hi Miaohe,
> 
> Thank you for the update.  I might miss something, but I still see the same
> problem (I checked it on mm-everything-2022-05-19-00-03 + this patchset).

Hi Naoya,
I reproduce the issue in the linux-next-next-20220520 version. And I found even if
I *do not inject the swapin error*, the deadloop still occurs. After investigating the
code for a long while, I found the root cause:

diff --git a/mm/shmem.c b/mm/shmem.c
index d55dd972023a..6d23ed4d23cb 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -1182,7 +1182,7 @@ static int shmem_find_swap_entries(struct address_space *mapping,
                if (swp_type(entry) != type)
                        continue;

-               indices[ret] = xas.xa_index;
+               indices[ret++] = xas.xa_index;
                if (!folio_batch_add(fbatch, folio))
                        break;

The origin code does not increment the ret value when a folio is found. I will send a patch
to fix this next week. Thanks! :)

BTW: With the above change, deadloop doesn't occur when swapin error is injected. I will take
a more close look at next week.

Thanks!

> 
> This patch has the effect to change the return value of shmem_swapin_folio(),
> -EIO (without this patch) to -EEXIST (with this patch).
> But shmem_unuse_swap_entries() checks neither, so no change from caller's view point.
> Maybe breaking in errors (rather than ENOMEM) in for loop in shmem_unuse_swap_entries()
> solves the issue?  I briefly checked with the below change, then swapoff can return
> with failure.
> 
> @@ -1222,7 +1222,7 @@ static int shmem_unuse_swap_entries(struct inode *inode,
>                         folio_put(folio);
>                         ret++;
>                 }
> -               if (error == -ENOMEM)
> +               if (error < 0)
>                         break;
>                 error = 0;
>         }
> 
>> ---
>>  mm/shmem.c | 39 +++++++++++++++++++++++++++++++++++++++
>>  1 file changed, 39 insertions(+)
>>
>> diff --git a/mm/shmem.c b/mm/shmem.c
>> index d3c7970e0179..d55dd972023a 100644
>> --- a/mm/shmem.c
>> +++ b/mm/shmem.c
>> @@ -1175,6 +1175,10 @@ static int shmem_find_swap_entries(struct address_space *mapping,
>>  			continue;
>>  
>>  		entry = radix_to_swp_entry(folio);
>> +		/*
>> +		 * swapin error entries can be found in the mapping. But they're
>> +		 * deliberately ignored here as we've done everything we can do.
>> +		 */
>>  		if (swp_type(entry) != type)
>>  			continue;
>>  
>> @@ -1672,6 +1676,36 @@ static int shmem_replace_page(struct page **pagep, gfp_t gfp,
>>  	return error;
>>  }
>>  
>> +static void shmem_set_folio_swapin_error(struct inode *inode, pgoff_t index,
>> +					 struct folio *folio, swp_entry_t swap)
>> +{
>> +	struct address_space *mapping = inode->i_mapping;
>> +	struct shmem_inode_info *info = SHMEM_I(inode);
>> +	swp_entry_t swapin_error;
>> +	void *old;
>> +
>> +	swapin_error = make_swapin_error_entry(&folio->page);
>> +	old = xa_cmpxchg_irq(&mapping->i_pages, index,
>> +			     swp_to_radix_entry(swap),
>> +			     swp_to_radix_entry(swapin_error), 0);
>> +	if (old != swp_to_radix_entry(swap))
>> +		return;
>> +
>> +	folio_wait_writeback(folio);
>> +	delete_from_swap_cache(&folio->page);
>> +	spin_lock_irq(&info->lock);
>> +	/*
>> +	 * Don't treat swapin error folio as alloced. Otherwise inode->i_blocks won't
>> +	 * be 0 when inode is released and thus trigger WARN_ON(inode->i_blocks) in
>> +	 * shmem_evict_inode.
>> +	 */
>> +	info->alloced--;
>> +	info->swapped--;
>> +	shmem_recalc_inode(inode);
>> +	spin_unlock_irq(&info->lock);
>> +	swap_free(swap);
>> +}
>> +
>>  /*
>>   * Swap in the page pointed to by *pagep.
>>   * Caller has to make sure that *pagep contains a valid swapped page.
> 
> (off-topic a little) BTW, the comment on shmem_swapin_folio() still mentions
> *pagep, but maybe it can be updated to *foliop.
> 
> Thanks,
> Naoya Horiguchi
> 
>> @@ -1695,6 +1729,9 @@ static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
>>  	swap = radix_to_swp_entry(*foliop);
>>  	*foliop = NULL;
>>  
>> +	if (is_swapin_error_entry(swap))
>> +		return -EIO;
>> +
>>  	/* Look it up and read it in.. */
>>  	page = lookup_swap_cache(swap, NULL, 0);
>>  	if (!page) {
>> @@ -1762,6 +1799,8 @@ static int shmem_swapin_folio(struct inode *inode, pgoff_t index,
>>  failed:
>>  	if (!shmem_confirm_swap(mapping, index, swap))
>>  		error = -EEXIST;
>> +	if (error == -EIO)
>> +		shmem_set_folio_swapin_error(inode, index, folio, swap);