Re: [syzbot] [ext4?] WARNING in mb_cache_destroy

Re: [syzbot] [ext4?] WARNING in mb_cache_destroy

[syzbot]

syzbot has bisected this issue to:

commit 67d7d8ad99beccd9fe92d585b87f1760dc9018e3
Author: Baokun Li <libaokun1@huawei.com>
Date:   Thu Jun 16 02:13:56 2022 +0000

    ext4: fix use-after-free in ext4_xattr_set_entry

bisection log:  https://syzkaller.appspot.com/x/bisect.txt?x=1011bcf8980000
start commit:   b947cc5bf6d7 Merge tag 'erofs-for-6.9-rc7-fixes' of git://..
git tree:       upstream
final oops:     https://syzkaller.appspot.com/x/report.txt?x=1211bcf8980000
console output: https://syzkaller.appspot.com/x/log.txt?x=1411bcf8980000
kernel config:  https://syzkaller.appspot.com/x/.config?x=d2f00edef461175
dashboard link: https://syzkaller.appspot.com/bug?extid=dd43bd0f7474512edc47
syz repro:      https://syzkaller.appspot.com/x/repro.syz?x=11d2957f180000
C reproducer:   https://syzkaller.appspot.com/x/repro.c?x=1620ca40980000

Reported-by: syzbot+dd43bd0f7474512edc47@syzkaller.appspotmail.com
Fixes: 67d7d8ad99be ("ext4: fix use-after-free in ext4_xattr_set_entry")

For information about bisection process see: https://goo.gl/tpsmEJ#bisection

Re: [syzbot] [ext4?] WARNING in mb_cache_destroy

[Jan Kara]

On Tue 30-04-24 08:04:03, syzbot wrote:
> syzbot has bisected this issue to:
> 
> commit 67d7d8ad99beccd9fe92d585b87f1760dc9018e3
> Author: Baokun Li <libaokun1@huawei.com>
> Date:   Thu Jun 16 02:13:56 2022 +0000
> 
>     ext4: fix use-after-free in ext4_xattr_set_entry

So I'm not sure the bisect is correct since the change is looking harmless.
But it is sufficiently related that there indeed may be some relationship.
Anyway, the kernel log has:

[   44.932900][ T1063] EXT4-fs warning (device loop0): ext4_evict_inode:297: xattr delete (err -12)
[   44.943316][ T1063] EXT4-fs (loop0): unmounting filesystem.
[   44.949531][ T1063] ------------[ cut here ]------------
[   44.955050][ T1063] WARNING: CPU: 0 PID: 1063 at fs/mbcache.c:409 mb_cache_destroy+0xda/0x110

So ext4_xattr_delete_inode() called when removing inode has failed with
ENOMEM and later mb_cache_destroy() was eventually complaining about having
mbcache entry with increased refcount. So likely some error cleanup path is
forgetting to drop mbcache entry reference somewhere but at this point I
cannot find where. We'll likely need to play with the reproducer to debug
that. Baokun, any chance for looking into this?

								Honza
-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR

Re: [syzbot] [ext4?] WARNING in mb_cache_destroy

[Baokun Li]

On 2024/5/2 18:33, Jan Kara wrote:
> On Tue 30-04-24 08:04:03, syzbot wrote:
>> syzbot has bisected this issue to:
>>
>> commit 67d7d8ad99beccd9fe92d585b87f1760dc9018e3
>> Author: Baokun Li <libaokun1@huawei.com>
>> Date:   Thu Jun 16 02:13:56 2022 +0000
>>
>>      ext4: fix use-after-free in ext4_xattr_set_entry
> So I'm not sure the bisect is correct since the change is looking harmless.
> But it is sufficiently related that there indeed may be some relationship.
> Anyway, the kernel log has:
>
> [   44.932900][ T1063] EXT4-fs warning (device loop0): ext4_evict_inode:297: xattr delete (err -12)
> [   44.943316][ T1063] EXT4-fs (loop0): unmounting filesystem.
> [   44.949531][ T1063] ------------[ cut here ]------------
> [   44.955050][ T1063] WARNING: CPU: 0 PID: 1063 at fs/mbcache.c:409 mb_cache_destroy+0xda/0x110
>
> So ext4_xattr_delete_inode() called when removing inode has failed with
> ENOMEM and later mb_cache_destroy() was eventually complaining about having
> mbcache entry with increased refcount. So likely some error cleanup path is
> forgetting to drop mbcache entry reference somewhere but at this point I
> cannot find where. We'll likely need to play with the reproducer to debug
> that. Baokun, any chance for looking into this?
>
> 								Honza

Hi Honza,

Sorry for the late reply, the first two days were May Day holidays.
Thanks for the heads up, I'll take a look at this soon.

Regards,
Baokun

Re: [syzbot] [ext4?] WARNING in mb_cache_destroy

[Baokun Li]

Hi Honza,

On 2024/5/2 18:33, Jan Kara wrote:
> On Tue 30-04-24 08:04:03, syzbot wrote:
>> syzbot has bisected this issue to:
>>
>> commit 67d7d8ad99beccd9fe92d585b87f1760dc9018e3
>> Author: Baokun Li <libaokun1@huawei.com>
>> Date:   Thu Jun 16 02:13:56 2022 +0000
>>
>>      ext4: fix use-after-free in ext4_xattr_set_entry
> So I'm not sure the bisect is correct since the change is looking harmless.
Yes, the root cause of the problem has nothing to do with this patch,
and please see the detailed analysis below.
> But it is sufficiently related that there indeed may be some relationship.
> Anyway, the kernel log has:
>
> [   44.932900][ T1063] EXT4-fs warning (device loop0): ext4_evict_inode:297: xattr delete (err -12)
> [   44.943316][ T1063] EXT4-fs (loop0): unmounting filesystem.
> [   44.949531][ T1063] ------------[ cut here ]------------
> [   44.955050][ T1063] WARNING: CPU: 0 PID: 1063 at fs/mbcache.c:409 mb_cache_destroy+0xda/0x110
>
> So ext4_xattr_delete_inode() called when removing inode has failed with
> ENOMEM and later mb_cache_destroy() was eventually complaining about having
> mbcache entry with increased refcount. So likely some error cleanup path is
> forgetting to drop mbcache entry reference somewhere but at this point I
> cannot find where. We'll likely need to play with the reproducer to debug
> that. Baokun, any chance for looking into this?
>
> 								Honza
As you guessed, when -ENOMEM is returned in ext4_sb_bread(),
the reference count of ce is not properly released, as follows.

ext4_create
  __ext4_new_inode
   security_inode_init_security
    ext4_initxattrs
     ext4_xattr_set_handle
      ext4_xattr_block_find
      ext4_xattr_block_set
       ext4_xattr_block_cache_find
         ce = mb_cache_entry_find_first
             __entry_find
             atomic_inc_not_zero(&entry->e_refcnt)
         bh = ext4_sb_bread(inode->i_sb, ce->e_value, REQ_PRIO);
         if (PTR_ERR(bh) == -ENOMEM)
             return NULL;

Before merging into commit 67d7d8ad99be("ext4: fix use-after-free
in ext4_xattr_set_entry"), it will not return early in 
ext4_xattr_ibody_find(),
so it tries to find it in iboy, fails the check in xattr_check_inode() and
returns without executing ext4_xattr_block_find(). Thus it will bisect
the patch, but actually has nothing to do with it.

ext4_xattr_ibody_get
  xattr_check_inode
   __xattr_check_inode
    check_xattrs
     if (end - (void *)header < sizeof(*header) + sizeof(u32))
       "in-inode xattr block too small"

Here's the patch in testing, I'll send it out officially after it is tested.
(PS:  I'm not sure if propagating the ext4_xattr_block_cache_find() 
errors would be better.)

Regards,
Baokun


From: Baokun Li <libaokun1@huawei.com>
Date: Fri, 3 May 2024 16:51:43 +0800
Subject: [PATCH] ext4: fix mb_cache_entry's e_refcnt leak in
  ext4_xattr_block_cache_find()

Syzbot reports a warning as follows:

============================================
WARNING: CPU: 0 PID: 5075 at fs/mbcache.c:419 mb_cache_destroy+0x224/0x290
Modules linked in:
CPU: 0 PID: 5075 Comm: syz-executor199 Not tainted 6.9.0-rc6-gb947cc5bf6d7
RIP: 0010:mb_cache_destroy+0x224/0x290 fs/mbcache.c:419
Call Trace:
  <TASK>
  ext4_put_super+0x6d4/0xcd0 fs/ext4/super.c:1375
  generic_shutdown_super+0x136/0x2d0 fs/super.c:641
  kill_block_super+0x44/0x90 fs/super.c:1675
  ext4_kill_sb+0x68/0xa0 fs/ext4/super.c:7327
[...]
============================================

This is because when finding an entry in ext4_xattr_block_cache_find(), if
ext4_sb_bread() returns -ENOMEM, the ce's e_refcnt, which has already grown
in the __entry_find(), won't be put away, and eventually trigger the above
issue in mb_cache_destroy() due to reference count leakage. So correct the
handling of the -ENOMEM error branch to avoid the above issue.

Reported-by: syzbot+dd43bd0f7474512edc47@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=dd43bd0f7474512edc47
Fixes: fb265c9cb49e ("ext4: add ext4_sb_bread() to disambiguate ENOMEM 
cases")
Cc: stable@kernel.org # v5.0-rc1
Signed-off-by: Baokun Li <libaokun1@huawei.com>
---
  fs/ext4/xattr.c | 7 +++----
  1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/fs/ext4/xattr.c b/fs/ext4/xattr.c
index b67a176bfcf9..5c9e751915fd 100644
--- a/fs/ext4/xattr.c
+++ b/fs/ext4/xattr.c
@@ -3113,11 +3113,10 @@ ext4_xattr_block_cache_find(struct inode *inode,

          bh = ext4_sb_bread(inode->i_sb, ce->e_value, REQ_PRIO);
          if (IS_ERR(bh)) {
-            if (PTR_ERR(bh) == -ENOMEM)
-                return NULL;
+            if (PTR_ERR(bh) != -ENOMEM)
+                EXT4_ERROR_INODE(inode, "block %lu read error",
+                         (unsigned long)ce->e_value);
              bh = NULL;
-            EXT4_ERROR_INODE(inode, "block %lu read error",
-                     (unsigned long)ce->e_value);
          } else if (ext4_xattr_cmp(header, BHDR(bh)) == 0) {
              *pce = ce;
              return bh;
-- 
2.39.2

Re: [syzbot] [ext4?] WARNING in mb_cache_destroy

[Jan Kara]

Hi!

On Fri 03-05-24 17:51:07, Baokun Li wrote:
> On 2024/5/2 18:33, Jan Kara wrote:
> > On Tue 30-04-24 08:04:03, syzbot wrote:
> > > syzbot has bisected this issue to:
> > > 
> > > commit 67d7d8ad99beccd9fe92d585b87f1760dc9018e3
> > > Author: Baokun Li <libaokun1@huawei.com>
> > > Date:   Thu Jun 16 02:13:56 2022 +0000
> > > 
> > >      ext4: fix use-after-free in ext4_xattr_set_entry
> > So I'm not sure the bisect is correct since the change is looking harmless.
> Yes, the root cause of the problem has nothing to do with this patch,
> and please see the detailed analysis below.
> > But it is sufficiently related that there indeed may be some relationship.
> > Anyway, the kernel log has:
> > 
> > [   44.932900][ T1063] EXT4-fs warning (device loop0): ext4_evict_inode:297: xattr delete (err -12)
> > [   44.943316][ T1063] EXT4-fs (loop0): unmounting filesystem.
> > [   44.949531][ T1063] ------------[ cut here ]------------
> > [   44.955050][ T1063] WARNING: CPU: 0 PID: 1063 at fs/mbcache.c:409 mb_cache_destroy+0xda/0x110
> > 
> > So ext4_xattr_delete_inode() called when removing inode has failed with
> > ENOMEM and later mb_cache_destroy() was eventually complaining about having
> > mbcache entry with increased refcount. So likely some error cleanup path is
> > forgetting to drop mbcache entry reference somewhere but at this point I
> > cannot find where. We'll likely need to play with the reproducer to debug
> > that. Baokun, any chance for looking into this?
> > 
> > 								Honza
> As you guessed, when -ENOMEM is returned in ext4_sb_bread(),
> the reference count of ce is not properly released, as follows.
> 
> ext4_create
>  __ext4_new_inode
>   security_inode_init_security
>    ext4_initxattrs
>     ext4_xattr_set_handle
>      ext4_xattr_block_find
>      ext4_xattr_block_set
>       ext4_xattr_block_cache_find
>         ce = mb_cache_entry_find_first
>             __entry_find
>             atomic_inc_not_zero(&entry->e_refcnt)
>         bh = ext4_sb_bread(inode->i_sb, ce->e_value, REQ_PRIO);
>         if (PTR_ERR(bh) == -ENOMEM)
>             return NULL;
> 
> Before merging into commit 67d7d8ad99be("ext4: fix use-after-free
> in ext4_xattr_set_entry"), it will not return early in
> ext4_xattr_ibody_find(),
> so it tries to find it in iboy, fails the check in xattr_check_inode() and
> returns without executing ext4_xattr_block_find(). Thus it will bisect
> the patch, but actually has nothing to do with it.
> 
> ext4_xattr_ibody_get
>  xattr_check_inode
>   __xattr_check_inode
>    check_xattrs
>     if (end - (void *)header < sizeof(*header) + sizeof(u32))
>       "in-inode xattr block too small"
> 
> Here's the patch in testing, I'll send it out officially after it is tested.
> (PS:  I'm not sure if propagating the ext4_xattr_block_cache_find() errors
> would be better.)

Great! Thanks for debugging this! Some comments to your fix below:

> diff --git a/fs/ext4/xattr.c b/fs/ext4/xattr.c
> index b67a176bfcf9..5c9e751915fd 100644
> --- a/fs/ext4/xattr.c
> +++ b/fs/ext4/xattr.c
> @@ -3113,11 +3113,10 @@ ext4_xattr_block_cache_find(struct inode *inode,
> 
>          bh = ext4_sb_bread(inode->i_sb, ce->e_value, REQ_PRIO);
>          if (IS_ERR(bh)) {
> -            if (PTR_ERR(bh) == -ENOMEM)
> -                return NULL;
> +            if (PTR_ERR(bh) != -ENOMEM)
> +                EXT4_ERROR_INODE(inode, "block %lu read error",
> +                         (unsigned long)ce->e_value);
>              bh = NULL;
> -            EXT4_ERROR_INODE(inode, "block %lu read error",
> -                     (unsigned long)ce->e_value);
>          } else if (ext4_xattr_cmp(header, BHDR(bh)) == 0) {
>              *pce = ce;
>              return bh;

So if we get the ENOMEM error, continuing the iteration seems to be
pointless as we'll likely get it for the following entries as well. I think
the original behavior of aborting the iteration in case of ENOMEM is
actually better. We just have to do mb_cache_entry_put(ea_block_cache, ce)
before returning...

								Honza

-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR

Re: [syzbot] [ext4?] WARNING in mb_cache_destroy

[Baokun Li]

On 2024/5/3 18:23, Jan Kara wrote:
> Hi!
>
> On Fri 03-05-24 17:51:07, Baokun Li wrote:
>> On 2024/5/2 18:33, Jan Kara wrote:
>>> On Tue 30-04-24 08:04:03, syzbot wrote:
>>>> syzbot has bisected this issue to:
>>>>
>>>> commit 67d7d8ad99beccd9fe92d585b87f1760dc9018e3
>>>> Author: Baokun Li <libaokun1@huawei.com>
>>>> Date:   Thu Jun 16 02:13:56 2022 +0000
>>>>
>>>>       ext4: fix use-after-free in ext4_xattr_set_entry
>>> So I'm not sure the bisect is correct since the change is looking harmless.
>> Yes, the root cause of the problem has nothing to do with this patch,
>> and please see the detailed analysis below.
>>> But it is sufficiently related that there indeed may be some relationship.
>>> Anyway, the kernel log has:
>>>
>>> [   44.932900][ T1063] EXT4-fs warning (device loop0): ext4_evict_inode:297: xattr delete (err -12)
>>> [   44.943316][ T1063] EXT4-fs (loop0): unmounting filesystem.
>>> [   44.949531][ T1063] ------------[ cut here ]------------
>>> [   44.955050][ T1063] WARNING: CPU: 0 PID: 1063 at fs/mbcache.c:409 mb_cache_destroy+0xda/0x110
>>>
>>> So ext4_xattr_delete_inode() called when removing inode has failed with
>>> ENOMEM and later mb_cache_destroy() was eventually complaining about having
>>> mbcache entry with increased refcount. So likely some error cleanup path is
>>> forgetting to drop mbcache entry reference somewhere but at this point I
>>> cannot find where. We'll likely need to play with the reproducer to debug
>>> that. Baokun, any chance for looking into this?
>>>
>>> 								Honza
>> As you guessed, when -ENOMEM is returned in ext4_sb_bread(),
>> the reference count of ce is not properly released, as follows.
>>
>> ext4_create
>>   __ext4_new_inode
>>    security_inode_init_security
>>     ext4_initxattrs
>>      ext4_xattr_set_handle
>>       ext4_xattr_block_find
>>       ext4_xattr_block_set
>>        ext4_xattr_block_cache_find
>>          ce = mb_cache_entry_find_first
>>              __entry_find
>>              atomic_inc_not_zero(&entry->e_refcnt)
>>          bh = ext4_sb_bread(inode->i_sb, ce->e_value, REQ_PRIO);
>>          if (PTR_ERR(bh) == -ENOMEM)
>>              return NULL;
>>
>> Before merging into commit 67d7d8ad99be("ext4: fix use-after-free
>> in ext4_xattr_set_entry"), it will not return early in
>> ext4_xattr_ibody_find(),
>> so it tries to find it in iboy, fails the check in xattr_check_inode() and
>> returns without executing ext4_xattr_block_find(). Thus it will bisect
>> the patch, but actually has nothing to do with it.
>>
>> ext4_xattr_ibody_get
>>   xattr_check_inode
>>    __xattr_check_inode
>>     check_xattrs
>>      if (end - (void *)header < sizeof(*header) + sizeof(u32))
>>        "in-inode xattr block too small"
>>
>> Here's the patch in testing, I'll send it out officially after it is tested.
>> (PS:  I'm not sure if propagating the ext4_xattr_block_cache_find() errors
>> would be better.)
> Great! Thanks for debugging this! Some comments to your fix below:
>
>> diff --git a/fs/ext4/xattr.c b/fs/ext4/xattr.c
>> index b67a176bfcf9..5c9e751915fd 100644
>> --- a/fs/ext4/xattr.c
>> +++ b/fs/ext4/xattr.c
>> @@ -3113,11 +3113,10 @@ ext4_xattr_block_cache_find(struct inode *inode,
>>
>>           bh = ext4_sb_bread(inode->i_sb, ce->e_value, REQ_PRIO);
>>           if (IS_ERR(bh)) {
>> -            if (PTR_ERR(bh) == -ENOMEM)
>> -                return NULL;
>> +            if (PTR_ERR(bh) != -ENOMEM)
>> +                EXT4_ERROR_INODE(inode, "block %lu read error",
>> +                         (unsigned long)ce->e_value);
>>               bh = NULL;
>> -            EXT4_ERROR_INODE(inode, "block %lu read error",
>> -                     (unsigned long)ce->e_value);
>>           } else if (ext4_xattr_cmp(header, BHDR(bh)) == 0) {
>>               *pce = ce;
>>               return bh;
> So if we get the ENOMEM error, continuing the iteration seems to be
> pointless as we'll likely get it for the following entries as well. I think
> the original behavior of aborting the iteration in case of ENOMEM is
> actually better. We just have to do mb_cache_entry_put(ea_block_cache, ce)
> before returning...
>
> 								Honza
Returning NULL here would normally attempt to allocate a new
xattr_block in ext4_xattr_block_set(), and when ext4_sb_bread() fails,
allocating the new block and inserting it would most likely fail as well,
so my initial thought was to propagate the error from ext4_sb_bread()
to also make ext4_xattr_block_set() fail when ext4_sb_bread() fails.

But I noticed that before Ted added the special handling for -ENOMEM,
EXT4_ERROR_INODE was called to set the ERROR_FS flag no matter
what error ext4_sb_bread() returned, and after we can distinguish
between -EIO and -ENOMEM, we don't have to set the ERROR_FS flag
in the case of -ENOMEM. So there's this conservative fix now.

In short, in my personal opinion, for -EIO and -ENOMEM, they should
be the same except whether or not the ERROR_FS flag is set.
Otherwise, I think adding mb_cache_entry_put() directly is the easiest
and most straightforward fix.  Honza, do you have any other thoughts?

Thanks,
Baokun

Re: [syzbot] [ext4?] WARNING in mb_cache_destroy

[Jan Kara]

On Fri 03-05-24 19:38:21, Baokun Li wrote:
> On 2024/5/3 18:23, Jan Kara wrote:
> > Hi!
> > 
> > On Fri 03-05-24 17:51:07, Baokun Li wrote:
> > > On 2024/5/2 18:33, Jan Kara wrote:
> > > > On Tue 30-04-24 08:04:03, syzbot wrote:
> > > > > syzbot has bisected this issue to:
> > > > > 
> > > > > commit 67d7d8ad99beccd9fe92d585b87f1760dc9018e3
> > > > > Author: Baokun Li <libaokun1@huawei.com>
> > > > > Date:   Thu Jun 16 02:13:56 2022 +0000
> > > > > 
> > > > >       ext4: fix use-after-free in ext4_xattr_set_entry
> > > > So I'm not sure the bisect is correct since the change is looking harmless.
> > > Yes, the root cause of the problem has nothing to do with this patch,
> > > and please see the detailed analysis below.
> > > > But it is sufficiently related that there indeed may be some relationship.
> > > > Anyway, the kernel log has:
> > > > 
> > > > [   44.932900][ T1063] EXT4-fs warning (device loop0): ext4_evict_inode:297: xattr delete (err -12)
> > > > [   44.943316][ T1063] EXT4-fs (loop0): unmounting filesystem.
> > > > [   44.949531][ T1063] ------------[ cut here ]------------
> > > > [   44.955050][ T1063] WARNING: CPU: 0 PID: 1063 at fs/mbcache.c:409 mb_cache_destroy+0xda/0x110
> > > > 
> > > > So ext4_xattr_delete_inode() called when removing inode has failed with
> > > > ENOMEM and later mb_cache_destroy() was eventually complaining about having
> > > > mbcache entry with increased refcount. So likely some error cleanup path is
> > > > forgetting to drop mbcache entry reference somewhere but at this point I
> > > > cannot find where. We'll likely need to play with the reproducer to debug
> > > > that. Baokun, any chance for looking into this?
> > > > 
> > > > 								Honza
> > > As you guessed, when -ENOMEM is returned in ext4_sb_bread(),
> > > the reference count of ce is not properly released, as follows.
> > > 
> > > ext4_create
> > >   __ext4_new_inode
> > >    security_inode_init_security
> > >     ext4_initxattrs
> > >      ext4_xattr_set_handle
> > >       ext4_xattr_block_find
> > >       ext4_xattr_block_set
> > >        ext4_xattr_block_cache_find
> > >          ce = mb_cache_entry_find_first
> > >              __entry_find
> > >              atomic_inc_not_zero(&entry->e_refcnt)
> > >          bh = ext4_sb_bread(inode->i_sb, ce->e_value, REQ_PRIO);
> > >          if (PTR_ERR(bh) == -ENOMEM)
> > >              return NULL;
> > > 
> > > Before merging into commit 67d7d8ad99be("ext4: fix use-after-free
> > > in ext4_xattr_set_entry"), it will not return early in
> > > ext4_xattr_ibody_find(),
> > > so it tries to find it in iboy, fails the check in xattr_check_inode() and
> > > returns without executing ext4_xattr_block_find(). Thus it will bisect
> > > the patch, but actually has nothing to do with it.
> > > 
> > > ext4_xattr_ibody_get
> > >   xattr_check_inode
> > >    __xattr_check_inode
> > >     check_xattrs
> > >      if (end - (void *)header < sizeof(*header) + sizeof(u32))
> > >        "in-inode xattr block too small"
> > > 
> > > Here's the patch in testing, I'll send it out officially after it is tested.
> > > (PS:  I'm not sure if propagating the ext4_xattr_block_cache_find() errors
> > > would be better.)
> > Great! Thanks for debugging this! Some comments to your fix below:
> > 
> > > diff --git a/fs/ext4/xattr.c b/fs/ext4/xattr.c
> > > index b67a176bfcf9..5c9e751915fd 100644
> > > --- a/fs/ext4/xattr.c
> > > +++ b/fs/ext4/xattr.c
> > > @@ -3113,11 +3113,10 @@ ext4_xattr_block_cache_find(struct inode *inode,
> > > 
> > >           bh = ext4_sb_bread(inode->i_sb, ce->e_value, REQ_PRIO);
> > >           if (IS_ERR(bh)) {
> > > -            if (PTR_ERR(bh) == -ENOMEM)
> > > -                return NULL;
> > > +            if (PTR_ERR(bh) != -ENOMEM)
> > > +                EXT4_ERROR_INODE(inode, "block %lu read error",
> > > +                         (unsigned long)ce->e_value);
> > >               bh = NULL;
> > > -            EXT4_ERROR_INODE(inode, "block %lu read error",
> > > -                     (unsigned long)ce->e_value);
> > >           } else if (ext4_xattr_cmp(header, BHDR(bh)) == 0) {
> > >               *pce = ce;
> > >               return bh;
> > So if we get the ENOMEM error, continuing the iteration seems to be
> > pointless as we'll likely get it for the following entries as well. I think
> > the original behavior of aborting the iteration in case of ENOMEM is
> > actually better. We just have to do mb_cache_entry_put(ea_block_cache, ce)
> > before returning...
> > 
> > 								Honza
> Returning NULL here would normally attempt to allocate a new
> xattr_block in ext4_xattr_block_set(), and when ext4_sb_bread() fails,
> allocating the new block and inserting it would most likely fail as well,
> so my initial thought was to propagate the error from ext4_sb_bread()
> to also make ext4_xattr_block_set() fail when ext4_sb_bread() fails.

Yes, this would be probably even better solution.

> But I noticed that before Ted added the special handling for -ENOMEM,
> EXT4_ERROR_INODE was called to set the ERROR_FS flag no matter
> what error ext4_sb_bread() returned, and after we can distinguish
> between -EIO and -ENOMEM, we don't have to set the ERROR_FS flag
> in the case of -ENOMEM. So there's this conservative fix now.
> 
> In short, in my personal opinion, for -EIO and -ENOMEM, they should
> be the same except whether or not the ERROR_FS flag is set.
> Otherwise, I think adding mb_cache_entry_put() directly is the easiest
> and most straightforward fix.  Honza, do you have any other thoughts?

Yeah. I'd go for adding mb_cache_entry_put() now as a quick fix and then
work on propagating the error from ext4_xattr_block_cache_find() as a
cleaner solution...

								Honza

-- 
Jan Kara <jack@suse.com>
SUSE Labs, CR

Re: [syzbot] [ext4?] WARNING in mb_cache_destroy

[Baokun Li]

On 2024/5/3 22:09, Jan Kara wrote:
> On Fri 03-05-24 19:38:21, Baokun Li wrote:
>> On 2024/5/3 18:23, Jan Kara wrote:
>>> Hi!
>>>
>>> On Fri 03-05-24 17:51:07, Baokun Li wrote:
>>>> On 2024/5/2 18:33, Jan Kara wrote:
>>>>> On Tue 30-04-24 08:04:03, syzbot wrote:
>>>>>> syzbot has bisected this issue to:
>>>>>>
>>>>>> commit 67d7d8ad99beccd9fe92d585b87f1760dc9018e3
>>>>>> Author: Baokun Li <libaokun1@huawei.com>
>>>>>> Date:   Thu Jun 16 02:13:56 2022 +0000
>>>>>>
>>>>>>        ext4: fix use-after-free in ext4_xattr_set_entry
>>>>> So I'm not sure the bisect is correct since the change is looking harmless.
>>>> Yes, the root cause of the problem has nothing to do with this patch,
>>>> and please see the detailed analysis below.
>>>>> But it is sufficiently related that there indeed may be some relationship.
>>>>> Anyway, the kernel log has:
>>>>>
>>>>> [   44.932900][ T1063] EXT4-fs warning (device loop0): ext4_evict_inode:297: xattr delete (err -12)
>>>>> [   44.943316][ T1063] EXT4-fs (loop0): unmounting filesystem.
>>>>> [   44.949531][ T1063] ------------[ cut here ]------------
>>>>> [   44.955050][ T1063] WARNING: CPU: 0 PID: 1063 at fs/mbcache.c:409 mb_cache_destroy+0xda/0x110
>>>>>
>>>>> So ext4_xattr_delete_inode() called when removing inode has failed with
>>>>> ENOMEM and later mb_cache_destroy() was eventually complaining about having
>>>>> mbcache entry with increased refcount. So likely some error cleanup path is
>>>>> forgetting to drop mbcache entry reference somewhere but at this point I
>>>>> cannot find where. We'll likely need to play with the reproducer to debug
>>>>> that. Baokun, any chance for looking into this?
>>>>>
>>>>> 								Honza
>>>> As you guessed, when -ENOMEM is returned in ext4_sb_bread(),
>>>> the reference count of ce is not properly released, as follows.
>>>>
>>>> ext4_create
>>>>    __ext4_new_inode
>>>>     security_inode_init_security
>>>>      ext4_initxattrs
>>>>       ext4_xattr_set_handle
>>>>        ext4_xattr_block_find
>>>>        ext4_xattr_block_set
>>>>         ext4_xattr_block_cache_find
>>>>           ce = mb_cache_entry_find_first
>>>>               __entry_find
>>>>               atomic_inc_not_zero(&entry->e_refcnt)
>>>>           bh = ext4_sb_bread(inode->i_sb, ce->e_value, REQ_PRIO);
>>>>           if (PTR_ERR(bh) == -ENOMEM)
>>>>               return NULL;
>>>>
>>>> Before merging into commit 67d7d8ad99be("ext4: fix use-after-free
>>>> in ext4_xattr_set_entry"), it will not return early in
>>>> ext4_xattr_ibody_find(),
>>>> so it tries to find it in iboy, fails the check in xattr_check_inode() and
>>>> returns without executing ext4_xattr_block_find(). Thus it will bisect
>>>> the patch, but actually has nothing to do with it.
>>>>
>>>> ext4_xattr_ibody_get
>>>>    xattr_check_inode
>>>>     __xattr_check_inode
>>>>      check_xattrs
>>>>       if (end - (void *)header < sizeof(*header) + sizeof(u32))
>>>>         "in-inode xattr block too small"
>>>>
>>>> Here's the patch in testing, I'll send it out officially after it is tested.
>>>> (PS:  I'm not sure if propagating the ext4_xattr_block_cache_find() errors
>>>> would be better.)
>>> Great! Thanks for debugging this! Some comments to your fix below:
>>>
>>>> diff --git a/fs/ext4/xattr.c b/fs/ext4/xattr.c
>>>> index b67a176bfcf9..5c9e751915fd 100644
>>>> --- a/fs/ext4/xattr.c
>>>> +++ b/fs/ext4/xattr.c
>>>> @@ -3113,11 +3113,10 @@ ext4_xattr_block_cache_find(struct inode *inode,
>>>>
>>>>            bh = ext4_sb_bread(inode->i_sb, ce->e_value, REQ_PRIO);
>>>>            if (IS_ERR(bh)) {
>>>> -            if (PTR_ERR(bh) == -ENOMEM)
>>>> -                return NULL;
>>>> +            if (PTR_ERR(bh) != -ENOMEM)
>>>> +                EXT4_ERROR_INODE(inode, "block %lu read error",
>>>> +                         (unsigned long)ce->e_value);
>>>>                bh = NULL;
>>>> -            EXT4_ERROR_INODE(inode, "block %lu read error",
>>>> -                     (unsigned long)ce->e_value);
>>>>            } else if (ext4_xattr_cmp(header, BHDR(bh)) == 0) {
>>>>                *pce = ce;
>>>>                return bh;
>>> So if we get the ENOMEM error, continuing the iteration seems to be
>>> pointless as we'll likely get it for the following entries as well. I think
>>> the original behavior of aborting the iteration in case of ENOMEM is
>>> actually better. We just have to do mb_cache_entry_put(ea_block_cache, ce)
>>> before returning...
>>>
>>> 								Honza
>> Returning NULL here would normally attempt to allocate a new
>> xattr_block in ext4_xattr_block_set(), and when ext4_sb_bread() fails,
>> allocating the new block and inserting it would most likely fail as well,
>> so my initial thought was to propagate the error from ext4_sb_bread()
>> to also make ext4_xattr_block_set() fail when ext4_sb_bread() fails.
> Yes, this would be probably even better solution.
Okay.
>
>> But I noticed that before Ted added the special handling for -ENOMEM,
>> EXT4_ERROR_INODE was called to set the ERROR_FS flag no matter
>> what error ext4_sb_bread() returned, and after we can distinguish
>> between -EIO and -ENOMEM, we don't have to set the ERROR_FS flag
>> in the case of -ENOMEM. So there's this conservative fix now.
>>
>> In short, in my personal opinion, for -EIO and -ENOMEM, they should
>> be the same except whether or not the ERROR_FS flag is set.
>> Otherwise, I think adding mb_cache_entry_put() directly is the easiest
>> and most straightforward fix.  Honza, do you have any other thoughts?
> Yeah. I'd go for adding mb_cache_entry_put() now as a quick fix and then
> work on propagating the error from ext4_xattr_block_cache_find() as a
> cleaner solution...
>
> 								Honza
>
Ok, thank you very much for the suggestion!
I'll send the quick fix out right away.

Cheers,
Baokun