linux-kernel - Re: [PATCH 053/190] Revert "ecryptfs: replace BUG

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite for Android: free password hash cracker in your pocket

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <20210421233239.GA177816@sequoia>
Date:   Wed, 21 Apr 2021 18:32:39 -0500
From:   Tyler Hicks <code@...icks.com>
To:     Al Viro <viro@...iv.linux.org.uk>
Cc:     Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        linux-kernel@...r.kernel.org, Aditya Pakki <pakki001@....edu>,
        Michael Halcrow <mhalcrow@...gle.com>
Subject: Re: [PATCH 053/190] Revert "ecryptfs: replace BUG_ON with error
 handling code"

On 2021-04-21 17:03:24, Al Viro wrote:
> On Wed, Apr 21, 2021 at 11:13:29AM -0500, Tyler Hicks wrote:
> 
> > > It *is* functionally harmless, AFAICS, but only because the condition
> > > is really impossible.  However,
> > > 	* it refers to vague (s)tool they'd produced, nevermind that
> > > all they really do is "find BUG_ON(), replace with returning an error".
> > > 	* unlike BUG_ON(), the replacement does *NOT* document the
> > > fact that condition should be impossible.
> > > IMO either should be sufficient for rejecting the patch.
> > 
> > I agree that it was not a malicious change. There are other places
> > within the same function that return -EINVAL and the expectation is that
> > errors from this function should be handled safely.
> 
> Umm...  Assuming that failure exits in the callers will function properly
> if those conditions are true.  Which is not obvious at all.
> 
> > That said, I can find no real-world reports of this BUG_ON() ever being
> > a problem and I don't think that there's any actual need for this
> > change. So, I'm alright with it being reverted considering the
> > circumstances.
> 
> AFAICS, at least some parts of that BUG_ON() are provably impossible
> (e.g. NULL crypt_stat would've oopsed well upstream of the only call
> of that function).  ECRYPTFS_STRUCT_INITIALIZED is set after
> ecryptfs_alloc_inode() and never cleared, i.e. it should be present
> in ecryptfs_inode_to_private(ecryptfs_inode)->crypt_stat.flags for
> all inodes.  And crypt_stat we are passing to that thing is
> calculated as &(ecryptfs_inode_to_private(ecryptfs_inode)->crypt_stat),
> which is another reason why it can't be NULL.

I agree.

> 
> Incidentally, what's ecryptfs_setattr() doing with similar check?
> It had been introduced in e10f281bca03 "eCryptfs: initialize crypt_stat
> in setattr", which claims
>     Recent changes in eCryptfs have made it possible to get to ecryptfs_setattr()
>     with an uninitialized crypt_stat struct.  This results in a wide and colorful
>     variety of unpleasantries.  This patch properly initializes the crypt_stat
>     structure in ecryptfs_setattr() when it is necessary to do so.
> and AFAICS at that point the call of ecryptfs_init_crypt_stat() in
> ecryptfs_alloc_inode() had already been there and EXCRYPTFS_STRUCT_INITIALIZED
> had been (unconditionally) set by it.  So how could that check trigger in
> ecryptfs_setattr()?  No direct calls of that function (then as well as now),
> it's only reachable as ecryptfs_{symlink,dir,main}_iops.setattr.  The first
> two could only end up set by ecryptfs_interpose(), for inode returned by
> iget5_locked() (i.e. one that had been returned by ->alloc_inode()),
> the last is set by ecryptfs_init_inode(), called by ecryptfs_inode_set(), 
> passed as callback to iget5_locked() by the same ecryptfs_interpose().
> IOW, again, the inode must have been returned by ->alloc_inode().
> 
> I realize that it had been a long time ago, but... could somebody
> recall what that patch had been about?  Michael?

I looked through the commits that proceeded e10f281bca03 and the only
thing I can think of is the addition of "passthrough" mode where the
lower, encrypted data can be directly read from the eCryptfs mount. It
was introduced in commit e77a56ddceee ("[PATCH] eCryptfs: Encrypted
passthrough"), several months before e10f281bca03. However, I don't see
how it would have left us with an uninitialized crypt_stat in setattr.

Tyler

> Commit in question contains another (and much bigger) chunk; do
> the comments in commit message refer to it?  Because it really
> looks like
> 	if (!(crypt_stat->flags & ECRYPTFS_STRUCT_INITIALIZED))
> 		ecryptfs_init_crypt_stat(crypt_stat);
> part in ecryptfs_setattr() is a confusing no-op...