Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.31 as permitted sender) client-ip=23.128.96.31;
Date:   Wed, 13 Dec 2023 13:53:58 +0100
From:   David Sterba <dsterba@suse.cz>
To:     "Paul E. McKenney" <paulmck@kernel.org>
Cc:     kernel test robot <lkp@intel.com>, llvm@lists.linux.dev,
        oe-kbuild-all@lists.linux.dev, linux-kernel@vger.kernel.org,
        Josef Bacik <josef@toxicpanda.com>, julia.lawall@inria.fr,
        clm@fb.com, dsterba@suse.com, baptiste.lepers@gmail.com
Subject: Re: [paulmck-rcu:frederic.2023.12.08a 29/37]
 fs/btrfs/transaction.c:496:6: error: call to '__compiletime_assert_329'
 declared with 'error' attribute: Need native word sized stores/loads for
 atomicity.
Message-ID: <20231213125358.GB3001@suse.cz>
Reply-To: dsterba@suse.cz
References: <202312091837.cKaPw0Tf-lkp@intel.com>
 <0487d7cc-b906-4a4a-b284-9c79700b4ede@paulmck-laptop>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <0487d7cc-b906-4a4a-b284-9c79700b4ede@paulmck-laptop>
User-Agent: Mutt/1.5.23.1-rc1 (2014-03-12)
Precedence: bulk

On Sat, Dec 09, 2023 at 07:51:30AM -0800, Paul E. McKenney wrote:
> On Sat, Dec 09, 2023 at 06:20:37PM +0800, kernel test robot wrote:
> > tree:   https://git.kernel.org/pub/scm/linux/kernel/git/paulmck/linux-rcu.git frederic.2023.12.08a
> > head:   37843b5f561a08ae899fb791eeeb5abd992eabe2
> > commit: 7dd87072d40809e26503f04b79d63290288dbbac [29/37] btrfs: Adjust ->last_trans ordering in btrfs_record_root_in_trans()
> > config: riscv-rv32_defconfig (https://download.01.org/0day-ci/archive/20231209/202312091837.cKaPw0Tf-lkp@intel.com/config)
> > compiler: clang version 17.0.0 (https://github.com/llvm/llvm-project.git 4a5ac14ee968ff0ad5d2cc1ffa0299048db4c88a)
> > reproduce (this is a W=1 build): (https://download.01.org/0day-ci/archive/20231209/202312091837.cKaPw0Tf-lkp@intel.com/reproduce)
> > 
> > If you fix the issue in a separate patch/commit (i.e. not just a new version of
> > the same patch/commit), kindly add following tags
> > | Reported-by: kernel test robot <lkp@intel.com>
> > | Closes: https://lore.kernel.org/oe-kbuild-all/202312091837.cKaPw0Tf-lkp@intel.com/
> > 
> > All errors (new ones prefixed by >>):
> > 
> >    warning: unknown warning option '-Wpacked-not-aligned'; did you mean '-Wpacked-non-pod'? [-Wunknown-warning-option]
> >    warning: unknown warning option '-Wstringop-truncation'; did you mean '-Wstring-concatenation'? [-Wunknown-warning-option]
> >    warning: unknown warning option '-Wmaybe-uninitialized'; did you mean '-Wuninitialized'? [-Wunknown-warning-option]
> > >> fs/btrfs/transaction.c:496:6: error: call to '__compiletime_assert_329' declared with 'error' attribute: Need native word sized stores/loads for atomicity.
> >      496 |         if (smp_load_acquire(&root->last_trans) == trans->transid && /* ^^^ */
> >          |             ^
> 
> Ooooh!!!  :-/
> 
> From what I can see, the current code can tear this load on 32-bit
> systems, which can result in bad comparisons and then in failure to wait
> for a partially complete transaction.
> 
> So is btrfs actually supported on 32-bit systems?  If not, would the
> following patch be appropriate?

There are limitations on 32bit systems, eg. due to shorter inode numbers
(ino_t is unsigned long) and that radix-tree/xarray does support only
unsigned long keys, while we have 64bit identifiers for inodes or tree
roots.

So far we support that and dropping it completely is I think a big deal,
like with any possibly used feature. What I've seen there are NAS boxes
with low power ARM that are 32bit.

> If btrfs is to be supported on 32-bit systems, from what I can see some
> major surgery is required, even if a 32-bit counter is wrap-safe for
> this particular type of transaction.  (But SSDs?  In-memory btrfs
> filesystems?)

We won't probably do any major surgery to support 32bit systems.

> 
> 							Thanx, Paul
> 
> ------------------------------------------------------------------------
> 
> diff --git a/fs/btrfs/Kconfig b/fs/btrfs/Kconfig
> index 4fb925e8c981..4d56158c34f9 100644
> --- a/fs/btrfs/Kconfig
> +++ b/fs/btrfs/Kconfig
> @@ -19,6 +19,7 @@ config BTRFS_FS
>  	select RAID6_PQ
>  	select XOR_BLOCKS
>  	depends on PAGE_SIZE_LESS_THAN_256KB
> +	depends on 64BIT

Can we keep the current inefficient smp_* barriers instead of dropping
the whole 32bit support as an alternative. If the smp_load_acquire are
better but not strictly necessary for the correctness (from the barriers
POV) I'd suggest to leave it as-is. We can put comments in case somebody
wants to optimize it in the future again.