Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
Date:   Thu, 1 Apr 2021 18:06:15 +0100
From:   Matthew Wilcox <willy@infradead.org>
To:     Hugh Dickins <hughd@google.com>
Cc:     Andrew Morton <akpm@linux-foundation.org>, linux-mm@kvack.org,
        linux-fsdevel@vger.kernel.org, linux-nvdimm@lists.01.org,
        linux-kernel@vger.kernel.org
Subject: Re: BUG_ON(!mapping_empty(&inode->i_data))
Message-ID: <20210401170615.GH351017@casper.infradead.org>
References: <alpine.LSU.2.11.2103301654520.2648@eggly.anvils>
 <20210331024913.GS351017@casper.infradead.org>
 <alpine.LSU.2.11.2103311413560.1201@eggly.anvils>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <alpine.LSU.2.11.2103311413560.1201@eggly.anvils>
Precedence: bulk

On Wed, Mar 31, 2021 at 02:58:12PM -0700, Hugh Dickins wrote:
> I suspect there's a bug in the XArray handling in collapse_file(),
> which sometimes leaves empty nodes behind.

Urp, yes, that can easily happen.

        /* This will be less messy when we use multi-index entries */
        do {
                xas_lock_irq(&xas);
                xas_create_range(&xas);
                if (!xas_error(&xas))
                        break;
                if (!xas_nomem(&xas, GFP_KERNEL)) {
                        result = SCAN_FAIL;
                        goto out;
                }

xas_create_range() can absolutely create nodes with zero entries.
So if we create m/n nodes and then it runs out of memory (or cgroup
denies it), we can leave nodes in the tree with zero entries.

There are three options for fixing it ...
 - Switch to using multi-index entries.  We need to do this anyway, but
   I don't yet have a handle on the bugs that you found last time I
   pushed this into linux-next.  At -rc5 seems like a late stage to be
   trying this solution.
 - Add an xas_prune_range() that gets called on failure.  Should be
   straightforward to write, but will be obsolete as soon as we do the
   above and it's a pain for the callers.
 - Change how xas_create_range() works to merely preallocate the xa_nodes
   and not insert them into the tree until we're trying to insert data into
   them.  I favour this option, and this scenario is amenable to writing
   a test that will simulate failure halfway through.

I'm going to start on option 3 now.