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] [day] [month] [year] [list]
Date:   Thu, 10 Feb 2022 12:08:07 -0800
From:   "Paul E. McKenney" <paulmck@...nel.org>
To:     Jan Kara <jack@...e.cz>
Cc:     Qian Cai <quic_qiancai@...cinc.com>, Theodore Ts'o <tytso@....edu>,
        Jan Kara <jack@...e.com>,
        Neeraj Upadhyay <quic_neeraju@...cinc.com>,
        Joel Fernandes <joel@...lfernandes.org>,
        Boqun Feng <boqun.feng@...il.com>, linux-ext4@...r.kernel.org,
        rcu@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [RFC PATCH] jbd2: avoid __GFP_ZERO with SLAB_TYPESAFE_BY_RCU

On Thu, Feb 10, 2022 at 12:06:34PM -0800, Paul E. McKenney wrote:
> On Thu, Feb 10, 2022 at 08:12:52PM +0100, Jan Kara wrote:
> > On Thu 10-02-22 06:57:13, Paul E. McKenney wrote:
> > > On Thu, Feb 10, 2022 at 10:16:48AM +0100, Jan Kara wrote:
> > > > On Wed 09-02-22 12:11:37, Paul E. McKenney wrote:
> > > > > On Wed, Feb 09, 2022 at 07:10:10PM +0100, Jan Kara wrote:
> > > > > > On Wed 09-02-22 11:57:42, Qian Cai wrote:
> > > > > > > Since the linux-next commit 120aa5e57479 (mm: Check for
> > > > > > > SLAB_TYPESAFE_BY_RCU and __GFP_ZERO slab allocation), we will get a
> > > > > > > boot warning. Avoid it by calling synchronize_rcu() before the zeroing.
> > > > > > > 
> > > > > > > Signed-off-by: Qian Cai <quic_qiancai@...cinc.com>
> > > > > > 
> > > > > > No, the performance impact of this would be just horrible. Can you
> > > > > > ellaborate a bit why SLAB_TYPESAFE_BY_RCU + __GFP_ZERO is a problem and why
> > > > > > synchronize_rcu() would be needed here before the memset() please? I mean
> > > > > > how is zeroing here any different from the memory just being used?
> > > > > 
> > > > > Suppose a reader picks up a pointer to a memory block, then that memory
> > > > > is freed.  No problem, given that this is a SLAB_TYPESAFE_BY_RCU slab,
> > > > > so the memory won't be freed while the reader is accessing it.  But while
> > > > > the reader is in the process of validating the block, it is zeroed.
> > > > > 
> > > > > How does the validation step handle this in all cases?
> > > > > 
> > > > > If you have a way of handling this, I will of course drop the patch.
> > > > > And learn something new, which is always a good thing.  ;-)
> > > > 
> > > > So I maybe missed something when implementing the usage of journal_heads
> > > > under RCU but let's have a look. An example of RCU user of journal heads
> > > > is fs/jbd2/transaction.c:jbd2_write_access_granted(). It does:
> > > > 
> > > >         rcu_read_lock();
> > > > 
> > > > 	// This part fetches journal_head from buffer_head - not related to
> > > > 	// our slab RCU discussion
> > > > 
> > > >         if (!buffer_jbd(bh))
> > > >                 goto out;
> > > >         /* This should be bh2jh() but that doesn't work with inline functions */
> > > >         jh = READ_ONCE(bh->b_private);
> > > >         if (!jh)
> > > >                 goto out;
> > > > 
> > > > 	// The validation comes here
> > > > 
> > > >         /* For undo access buffer must have data copied */
> > > >         if (undo && !jh->b_committed_data)
> > > >                 goto out;
> > > 
> > > OK, so if *jh was freed and re-zallocated in the meantime, this test
> > > should fail.  One concern would be if the zeroing was not at least eight
> > > bytes at a time, maybe due to overly eager use of fancy SIMD hardware.
> > > Though perhaps you also do something about ->b_committed_data on
> > > the free path, the commit-done path, or whatever?  (I do see a
> > > "jh->b_committed_data = NULL" on what might well be the commit-done path.)
> > >
> > > >         if (READ_ONCE(jh->b_transaction) != handle->h_transaction &&
> > > >             READ_ONCE(jh->b_next_transaction) != handle->h_transaction)
> > > >                 goto out;
> > > 
> > > And same with these guys.
> > 
> > Yes, on commit-done path we zero out jh->b_transaction (or set
> > jh->b_transaction = jh->b_next_transaction; jh->b_next_transaction = NULL).
> > So these fields are actually guaranteed to be zero on free.
> 
> Very good, thank you!
> 
> One more thing...
> 
> This assumes that when the slab allocator gets a fresh slab from mm,
> that slab has been zeroed, right?  Or is there some other trick that
> you are using to somehow accommodate randomly initialized memory?
> 
> Or am I just blind today and missing where the zeroing always happens
> (other than for pages destined for userspace)?

Never mind!!!

This is of course exactly why you are using zalloc().

I will go grab that brown paper bag...

							Thanx, Paul

> > > > 	// Then some more checks unrelated to the slab itself.
> > > > 
> > > >         /*
> > > >          * There are two reasons for the barrier here:
> > > >          * 1) Make sure to fetch b_bh after we did previous checks so that we
> > > >          * detect when jh went through free, realloc, attach to transaction
> > > >          * while we were checking. Paired with implicit barrier in that path.
> > > >          * 2) So that access to bh done after jbd2_write_access_granted()
> > > >          * doesn't get reordered and see inconsistent state of concurrent
> > > >          * do_get_write_access().
> > > >          */
> > > >         smp_mb();
> > > >         if (unlikely(jh->b_bh != bh))
> > > >                 goto out;
> > > > 
> > > > 	// If all passed
> > > > 
> > > > 	rcu_read_unlock();
> > > > 	return true;
> > > > 
> > > > So if we are going to return true from the function, we know that 'jh' was
> > > > attached to handle->h_transaction at some point. And when 'jh' was attached
> > > > to handle->h_transaction, the transaction was holding reference to the 'jh'
> > > > and our 'handle' holds reference to the transaction so 'jh' could not be
> > > > freed since that moment. I.e., we are sure our reference to the handle keeps
> > > > 'jh' alive and we can safely use it.
> > > > 
> > > > I don't see how any amount of scribbling over 'jh' could break this
> > > > validation. But maybe it is just a lack of my imagination :).
> > > 
> > > Regardless of whether you are suffering a lack of imagination, you
> > > have clearly demonstrated that it is possible to correctly use the
> > > SLAB_TYPESAFE_BY_RCU flag in conjunction with kmem_cache_alloc(), thus
> > > demonstrating that I was suffering from a lack of imagination.  ;-)
> > > 
> > > I have therefore reverted my commit.  Please accept my apologies for
> > > the hassle!
> > 
> > No problem. Thanks for reverting the patch. I can imagine that jbd2's use
> > of SLAB_TYPESAFE_BY_RCU is an unusual one...
> 
> I do like the very lightweight validation checks!
> 
> My mental model of this sort of validation always involved expensive
> atomic operations.  ;-)
> 
> 							Thanx, Paul

Powered by blists - more mailing lists