Re: [RFC PATCH] mm: filemap: fix nr_pages calculation overflow in filemap_map_pages()

[Baolin Wang <baolin.wang@linux.alibaba.com>]

On 3/16/26 10:06 PM, David Hildenbrand (Arm) wrote:
> On 3/13/26 04:45, Baolin Wang wrote:
>> When running stress-ng on my Arm64 machine with v7.0-rc3 kernel, I encountered
>> some very strange crash issues showing up as "Bad page state":
>>
>> "
>> [  734.496287] BUG: Bad page state in process stress-ng-env  pfn:415735fb
>> [  734.496427] page: refcount:0 mapcount:1 mapping:0000000000000000 index:0x4cf316 pfn:0x415735fb
>> [  734.496434] flags: 0x57fffe000000800(owner_2|node=1|zone=2|lastcpupid=0x3ffff)
>> [  734.496439] raw: 057fffe000000800 0000000000000000 dead000000000122 0000000000000000
>> [  734.496440] raw: 00000000004cf316 0000000000000000 0000000000000000 0000000000000000
>> [  734.496442] page dumped because: nonzero mapcount
>> "
>>
>> After analyzing this page’s state, it is hard to understand why the mapcount
>> is not 0 while the refcount is 0, since this page is not where the issue first
>> occurred. By enabling the CONFIG_DEBUG_VM config, I can reproduce the crash as
>> well and captured the first warning where the issue appears:
>>
>> "
>> [  734.469226] page: refcount:33 mapcount:0 mapping:00000000bef2d187 index:0x81a0 pfn:0x415735c0
>> [  734.469304] head: order:5 mapcount:0 entire_mapcount:0 nr_pages_mapped:0 pincount:0
>> [  734.469315] memcg:ffff000807a8ec00
>> [  734.469320] aops:ext4_da_aops ino:100b6f dentry name(?):"stress-ng-mmaptorture-9397-0-2736200540"
>> [  734.469335] flags: 0x57fffe400000069(locked|uptodate|lru|head|node=1|zone=2|lastcpupid=0x3ffff)
>> ......
>> [  734.469364] page dumped because: VM_WARN_ON_FOLIO((_Generic((page + nr_pages - 1),
>> const struct page *: (const struct folio *)_compound_head(page + nr_pages - 1), struct page *:
>> (struct folio *)_compound_head(page + nr_pages - 1))) != folio)
>> [  734.469390] ------------[ cut here ]------------
>> [  734.469393] WARNING: ./include/linux/rmap.h:351 at folio_add_file_rmap_ptes+0x3b8/0x468,
>> CPU#90: stress-ng-mlock/9430
>> [  734.469551]  folio_add_file_rmap_ptes+0x3b8/0x468 (P)
>> [  734.469555]  set_pte_range+0xd8/0x2f8
>> [  734.469566]  filemap_map_folio_range+0x190/0x400
>> [  734.469579]  filemap_map_pages+0x348/0x638
>> [  734.469583]  do_fault_around+0x140/0x198
>> ......
>> [  734.469640]  el0t_64_sync+0x184/0x188
>> "
>>
>> The code that triggers the warning is: "VM_WARN_ON_FOLIO(page_folio(page + nr_pages - 1) != folio, folio)",
>> which indicates that set_pte_range() tried to map beyond the large folio’s
>> size.
> 
> We had a bunch of similar reports throughout the last year from syzbot,
> but always without a reproducer.

I can reproduce it by running stress-ng for a few hours on my 128‑core 
machine, but I can’t reproduce it on smaller machines (e.g., 32‑core 
systems). It seems we need sufficient concurrency to trigger this race.

> In particular this here:
> 
> https://syzkaller.appspot.com/bug?extid=c0673e1f1f054fac28c2
> 
> via
> 
> https://lore.kernel.org/all/6758f0cc.050a0220.17f54a.0001.GAE@google.com/
> 
> Could that be the same root cause?

It appears so. Its warning stack calltrace matches what I reproduced.

>> By adding more debug information, I found that 'nr_pages' had overflowed in
>> filemap_map_pages(), causing set_pte_range() to establish mappings for a range
>> exceeding the folio size, potentially corrupting fields of pages that do not
>> belong to this folio (e.g., page->_mapcount).
> 
> Sounds quite bad.
> 
>>
>> After above analysis, I think the possible race is as follows:
>>
>> CPU 0                                                  CPU 1
>> filemap_map_pages()                                   ext4_setattr()
>>     //get and lock folio with old inode->i_size
>>     next_uptodate_folio()
>>
>>                                                            .......
>>                                                            //shrink the inode->i_size
>>                                                            i_size_write(inode, attr->ia_size);
>>
>>     //calculate the end_pgoff with the new inode->i_size
>>     file_end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE) - 1;
>>     end_pgoff = min(end_pgoff, file_end);
>>
>>     ......
>>     //nr_pages can be overflowed, cause xas.xa_index > end_pgoff
>>     end = folio_next_index(folio) - 1;
>>     nr_pages = min(end, end_pgoff) - xas.xa_index + 1;
>>
>>     ......
>>     //map large folio
>>     filemap_map_folio_range()
>>                                                            ......
>>                                                            //truncate folios
>>                                                            truncate_pagecache(inode, inode->i_size);
>>
>> To fix this issue, move the 'end_pgoff' calculation before next_uptodate_folio(),
>> so the retrieved folio stays consistent with the file end to avoid 'nr_pages'
>> calculation overflow. After this patch, the crash issue is gone.
>>
>> Fixes: 743a2753a02e ("filemap: cap PTE range to be created to allowed zero fill in folio_map_range()")
> 
> This certainly sounds like stable material :)

Yes, will cc stable.

>> Reported-by: Yuanhe Shu <xiangzao@linux.alibaba.com>
>> Tested-by: Yuanhe Shu <xiangzao@linux.alibaba.com>
>> Signed-off-by: Baolin Wang <baolin.wang@linux.alibaba.com>
>> ---
>>   mm/filemap.c | 6 +++---
>>   1 file changed, 3 insertions(+), 3 deletions(-)
>>
>> diff --git a/mm/filemap.c b/mm/filemap.c
>> index bc6775084744..923d28e59642 100644
>> --- a/mm/filemap.c
>> +++ b/mm/filemap.c
>> @@ -3879,14 +3879,14 @@ vm_fault_t filemap_map_pages(struct vm_fault *vmf,
>>   	unsigned int nr_pages = 0, folio_type;
>>   	unsigned short mmap_miss = 0, mmap_miss_saved;
>>   
>> +	file_end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE) - 1;
>> +	end_pgoff = min(end_pgoff, file_end);
>> +
>>   	rcu_read_lock();
>>   	folio = next_uptodate_folio(&xas, mapping, end_pgoff);
>>   	if (!folio)
>>   		goto out;
>>   
>> -	file_end = DIV_ROUND_UP(i_size_read(mapping->host), PAGE_SIZE) - 1;
>> -	end_pgoff = min(end_pgoff, file_end);
>> -
>>   	/*
>>   	 * Do not allow to map with PMD across i_size to preserve
>>   	 * SIGBUS semantics.
> 
> 
> LGTM. I do wonder if we want to add a comment above the file_end,
> stating that this really must happen before the next_uptodate_folio()
> to handle concurrent truncation.

Ack. How about adding the following comments?

"
Recalculate end_pgoff based on file_end before calling 
next_uptodate_folio() to avoid races with concurrent truncation.
"