Skip to contentSkip to conversation

/proc/<pid>/exe symlink behavior change in >=3.15.

DISCUSSION Piotr Karbowski 5 participants 8 messages Sep 6, 2014 – Sep 15, 2014 VFS

lore
Share & tools
lore
#0 lore raw

Hi,

Starting with kernel 3.15 the 'exe' symlink under /proc/<pid>/ acts diffrent than it used to in all the pre-3.15 kernels.

The usecase:

run /root/testbin (app that just sleeps)
cp /root/testbin /root/testbin.new
mv /root/testbin.new /root/testbin
ls -al /proc/pidof testbin/exe

<=3.14: /root/testbin (deleted)

Quoted 1 line

=3.15: /root/testbin.new (deleted)

Was the change intentional? It does render my system unusable and I failed to find a information about such change in the ChangeLog.

-- Piotr.

#1 lore raw
Quoted 12 lines · Piotr Karbowski

Hi,

Starting with kernel 3.15 the 'exe' symlink under /proc/<pid>/ acts diffrent than it used to in all the pre-3.15 kernels.

The usecase:

run /root/testbin (app that just sleeps)
cp /root/testbin /root/testbin.new
mv /root/testbin.new /root/testbin
ls -al /proc/pidof testbin/exe

Quoted 5 lines

<=3.14: /root/testbin (deleted)

Quoted 1 line

=3.15: /root/testbin.new (deleted)

Was the change intentional? It does render my system unusable and I failed to find a information about such change in the ChangeLog.

Can you please find the offending commit?
Using an automated bisect run this should be easily doable.

Signature · 2 lines
Thanks,
//richard
#3 lore raw
In reply to Piotr Karbowski
Quoted 12 lines · Piotr Karbowski

Hi,

Starting with kernel 3.15 the 'exe' symlink under /proc/<pid>/ acts diffrent than it used to in all the pre-3.15 kernels.

The usecase:

run /root/testbin (app that just sleeps)
cp /root/testbin /root/testbin.new
mv /root/testbin.new /root/testbin
ls -al /proc/pidof testbin/exe

Quoted 6 lines

<=3.14: /root/testbin (deleted)

Quoted 1 line

=3.15: /root/testbin.new (deleted)

Was the change intentional? It does render my system unusable and I failed to find a information about such change in the ChangeLog.

It looks like this was already broken for "long" (> DNAME_INLINE_LEN) names.

Short names share the problem since da1ce0670c14d8 "vfs: add cross-rename".

The following change to switch_names is the culprit:

- memcpy(dentry->d_iname, target->d_name.name,
- target->d_name.len + 1);

-       dentry->d_name.len = target->d_name.len;
-       return;
+       unsigned int i;
+       BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
+       for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {

+ swap(((long *) &dentry->d_iname)[i],
+ ((long *) &target->d_iname)[i]);
+ }

Dentries can have names from embedded structure or from an external buffer.

If you take a look around you will see the code just swaps pointers for "both external" case. But this results in the same behavoiur you are seeing.

Not sure how to fix it. Name in 'target' needs to be preserved, but memory allocation which may be needed for this purpose can fail and switch_names returns void, just like its callers (not to mention locks held around this).

One crap idea would be to have external buffers with a reference counter. d_inode would still be set to the buffer and freeing funcs would use container_of to get to the counter.

I can implement that later if it sounds sane enough.

Note this behaviour seems to be a requirement for cross-rename to work.

At least restoring previous behaviour while keeping cross-rename is not hard, I can write it later.

Signature · 1 line
Mateusz Guzik
author Reviewed-by Jan Kara Reviewed-by J. Bruce Fields
#2 lore raw
In reply to Richard Weinberger

Hi,

Quoted 2 lines · Richard Weinberger

Can you please find the offending commit?
Using an automated bisect run this should be easily doable.

That would be it:

da1ce0670c14d8380e423a3239e562a1dc15fa9e is the first bad commit commit da1ce0670c14d8380e423a3239e562a1dc15fa9e

Author: Miklos Szeredi <[email protected]>
Date:   Tue Apr 1 17:08:43 2014 +0200

vfs: add cross-rename

If flags contain RENAME_EXCHANGE then exchange source and
destination files.
There's no restriction on the type of the files; e.g. a directory
can be
exchanged with a symlink.

:040000 040000 86e9be1a42cb91c1068b76700c74ec7fdba5443c e0269a1fffefe60cbd6d56ccab6485ff383d728d M fs
:040000 040000 e677e31ebb0a5b355cefe199cba9453800624905 7a083642c2ea0698d85bfc9e65b99e8cfb4a7440 M include
:040000 040000 266252c531debd56ae52e000ed1579dca19ddb35 bfcb75158276de3feafd45d726f3297dfc139389 M security

-- Piotr.

#4 lore raw
In reply to Mateusz Guzik

Hi Mateusz,

Quoted 5 lines

I can implement that later if it sounds sane enough.
Note this behaviour seems to be a requirement for cross-rename to work.

At least restoring previous behaviour while keeping cross-rename is not hard, I can write it later.

Have you find a proper solution for this very issue or anything that can push forward addressing it?

Thankfully this issue has been addressed by a workaround in fs/dcache.c in grsecurity patch for 3.16.2 (grsecurity-3.0-3.16.2-201409082129.patch) but it would be great to have a fix upstream.

-- Piotr.

#5 lore raw
In reply to Mateusz Guzik
Quoted 40 lines · On Sun, 7 Sep 2014 09:56:08 +0200 Mateusz Guzik <[email protected]> wrote:
Quoted 12 lines · Piotr Karbowski

Hi,

Starting with kernel 3.15 the 'exe' symlink under /proc/<pid>/ acts diffrent than it used to in all the pre-3.15 kernels.

The usecase:

run /root/testbin (app that just sleeps)
cp /root/testbin /root/testbin.new
mv /root/testbin.new /root/testbin
ls -al /proc/pidof testbin/exe

Quoted 6 lines

<=3.14: /root/testbin (deleted)

Quoted 1 line

=3.15: /root/testbin.new (deleted)

Was the change intentional? It does render my system unusable and I failed to find a information about such change in the ChangeLog.

It looks like this was already broken for "long" (> DNAME_INLINE_LEN) names.

Short names share the problem since da1ce0670c14d8 "vfs: add cross-rename".

The following change to switch_names is the culprit:

(quoted diff) +6 -4
-       memcpy(dentry->d_iname, target->d_name.name,
-                   target->d_name.len + 1);
-       dentry->d_name.len = target->d_name.len;
-       return;
+       unsigned int i;
+       BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
+       for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
+     		  swap(((long *) &dentry->d_iname)[i],
+  			((long *) &target->d_iname)[i]);
+       }
Quoted 5 lines

Dentries can have names from embedded structure or from an external buffer.

If you take a look around you will see the code just swaps pointers for "both external" case. But this results in the same behavoiur you are seeing.

Looks like the real problem here is that __d_materialise_dentry() needs the old behavior of switch_names() . At least that's how it got fixed in grsecurity.

#6 lore raw
Quoted 43 lines · Chuck Ebbert
Quoted 40 lines · On Sun, 7 Sep 2014 09:56:08 +0200 Mateusz Guzik <[email protected]> wrote:
Quoted 12 lines · Piotr Karbowski

Hi,

Starting with kernel 3.15 the 'exe' symlink under /proc/<pid>/ acts diffrent than it used to in all the pre-3.15 kernels.

The usecase:

run /root/testbin (app that just sleeps)
cp /root/testbin /root/testbin.new
mv /root/testbin.new /root/testbin
ls -al /proc/pidof testbin/exe

Quoted 6 lines

<=3.14: /root/testbin (deleted)

Quoted 1 line

=3.15: /root/testbin.new (deleted)

Was the change intentional? It does render my system unusable and I failed to find a information about such change in the ChangeLog.

It looks like this was already broken for "long" (> DNAME_INLINE_LEN) names.

Short names share the problem since da1ce0670c14d8 "vfs: add cross-rename".

The following change to switch_names is the culprit:

(quoted diff) +6 -4
-       memcpy(dentry->d_iname, target->d_name.name,
-                   target->d_name.len + 1);
-       dentry->d_name.len = target->d_name.len;
-       return;
+       unsigned int i;
+       BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
+       for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
+     		  swap(((long *) &dentry->d_iname)[i],
+  			((long *) &target->d_iname)[i]);
+       }
Quoted 8 lines
Quoted 5 lines

Dentries can have names from embedded structure or from an external buffer.

If you take a look around you will see the code just swaps pointers for "both external" case. But this results in the same behavoiur you are seeing.

Looks like the real problem here is that __d_materialise_dentry() needs the old behavior of switch_names() . At least that's how it got fixed in grsecurity.

No.

Regression in question is an effect of swap instead of memcpy in switch_names, as called by d_move. Fix in grsecurity reverts to previous behaviour when needed and imho should be applied for the time being.

The real problem is that __d_move always switches parent dentry and calls switch_names, which actually switches names in some cases.

Without the regression you get expected results only for short names when you move stuff around within the same directory.

For instance with current code:
mv /foo/bar/baz /1/2/3

will replace the whole path.

Previous behavoiur would result in /foo/bar/3 as the new path, which is clearly still incorrect

Leaving the old dentry under the same parent would mean that the "tree" associated with now moved dentry will possibly need to be freed.

In addition to that one has to deal with the need of having renamed dentry the new name which possibly came from an external buffer. An idea I came up with (atomic_t refcount; char name[0]; with ->name assigned to dentry) may require adding an additional field to struct dentry, which would be bad.

I didn't have the time yet to look at this stuff properly.

Signature · 1 line
Mateusz Guzik
#7 lore raw
Quoted 17 lines · Mateusz Guzik
Quoted 16 lines · Chuck Ebbert
Quoted 13 lines · On Sun, 7 Sep 2014 09:56:08 +0200 Mateusz Guzik <[email protected]> wrote:
Quoted 12 lines · Piotr Karbowski

Hi,

Starting with kernel 3.15 the 'exe' symlink under /proc/<pid>/ acts diffrent than it used to in all the pre-3.15 kernels.

The usecase:

run /root/testbin (app that just sleeps)
cp /root/testbin /root/testbin.new
mv /root/testbin.new /root/testbin
ls -al /proc/pidof testbin/exe

Quoted 5 lines
Quoted 5 lines
Quoted 5 lines
Quoted 5 lines

<=3.14: /root/testbin (deleted)

Quoted 1 line

=3.15: /root/testbin.new (deleted)

Was the change intentional? It does render my system unusable and I failed to find a information about such change in the ChangeLog.

Piotr, what exactly happens? How does this break your system?

Quoted 28 lines
Quoted 28 lines
Quoted 27 lines
Quoted 1 line

It looks like this was already broken for "long" (> DNAME_INLINE_LEN) names.

Short names share the problem since da1ce0670c14d8 "vfs: add cross-rename".

The following change to switch_names is the culprit:

(quoted diff) +6 -4
-       memcpy(dentry->d_iname, target->d_name.name,
-                   target->d_name.len + 1);
-       dentry->d_name.len = target->d_name.len;
-       return;
+       unsigned int i;
+       BUILD_BUG_ON(!IS_ALIGNED(DNAME_INLINE_LEN, sizeof(long)));
+       for (i = 0; i < DNAME_INLINE_LEN / sizeof(long); i++) {
+                     swap(((long *) &dentry->d_iname)[i],
+                   ((long *) &target->d_iname)[i]);
+       }

Dentries can have names from embedded structure or from an external buffer.

If you take a look around you will see the code just swaps pointers for "both external" case. But this results in the same behavoiur you are seeing.

Quoted 8 lines
Quoted 2 lines

Looks like the real problem here is that __d_materialise_dentry() needs the old behavior of switch_names() . At least that's how it got fixed in grsecurity.

No.

Regression in question is an effect of swap instead of memcpy in switch_names, as called by d_move. Fix in grsecurity reverts to previous behaviour when needed and imho should be applied for the time being.

Ack for that. Linus will happily take this on the grounds of backward compatibility, even if the old behavior was arguably more crazy than the new one.

Quoted 10 lines

The real problem is that __d_move always switches parent dentry and calls switch_names, which actually switches names in some cases.

Without the regression you get expected results only for short names when you move stuff around within the same directory.

For instance with current code:
mv /foo/bar/baz /1/2/3

Quoted 7 lines

will replace the whole path.

Previous behavoiur would result in /foo/bar/3 as the new path, which is clearly still incorrect

Leaving the old dentry under the same parent would mean that the "tree" associated with now moved dentry will possibly need to be freed.

It's done by dput(). But callers need to hold ref to old parent anyway (because of locking) so it's not going to go away in d_move(), only after everything is done.

Quoted 6 lines

In addition to that one has to deal with the need of having renamed dentry the new name which possibly came from an external buffer. An idea I came up with (atomic_t refcount; char name[0]; with ->name assigned to dentry) may require adding an additional field to struct dentry, which would be bad.

You can do that without an extra field: e.g. use container_of() to get the refcounted struct from the name.

Consider using kref for refcounting.

Thanks,
Miklos