Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
Date:   Mon, 10 Feb 2020 21:05:46 +0300
From:   "Kirill A. Shutemov" <kirill@shutemov.name>
To:     Matthew Wilcox <willy@infradead.org>
Cc:     Qian Cai <cai@lca.pw>, akpm@linux-foundation.org, elver@google.com,
        linux-mm@kvack.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH -next] mm/filemap: fix a data race in filemap_fault()
Message-ID: <20200210180546.vt7yhdjav5oinij7@box>
References: <1581354029-20154-1-git-send-email-cai@lca.pw>
 <20200210172511.GL8731@bombadil.infradead.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20200210172511.GL8731@bombadil.infradead.org>
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Mon, Feb 10, 2020 at 09:25:11AM -0800, Matthew Wilcox wrote:
> On Mon, Feb 10, 2020 at 12:00:29PM -0500, Qian Cai wrote:
> > @@ -2622,7 +2622,7 @@ void filemap_map_pages(struct vm_fault *vmf,
> >  		if (page->index >= max_idx)
> >  			goto unlock;
> >  
> > -		if (file->f_ra.mmap_miss > 0)
> > +		if (data_race(file->f_ra.mmap_miss > 0))
> >  			file->f_ra.mmap_miss--;
> 
> How is this safe?  Two threads can each see 1, and then both decrement the
> in-memory copy, causing it to end up at -1.

Right, it is bogus.

Below is my completely untested attempt on fix this. It still allows
races, but they will only lead to missed accounting, but not underflow.


diff --git a/mm/filemap.c b/mm/filemap.c
index 1784478270e1..1919d37c646a 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -2365,6 +2365,7 @@ static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
 	struct address_space *mapping = file->f_mapping;
 	struct file *fpin = NULL;
 	pgoff_t offset = vmf->pgoff;
+	unsigned mmap_miss;
 
 	/* If we don't want any read-ahead, don't bother */
 	if (vmf->vma->vm_flags & VM_RAND_READ)
@@ -2380,14 +2381,15 @@ static struct file *do_sync_mmap_readahead(struct vm_fault *vmf)
 	}
 
 	/* Avoid banging the cache line if not needed */
-	if (ra->mmap_miss < MMAP_LOTSAMISS * 10)
-		ra->mmap_miss++;
+	mmap_miss = READ_ONCE(ra->mmap_miss);
+	if (mmap_miss < MMAP_LOTSAMISS * 10)
+		WRITE_ONCE(ra->mmap_miss, ++mmap_miss);
 
 	/*
 	 * Do we miss much more than hit in this file? If so,
 	 * stop bothering with read-ahead. It will only hurt.
 	 */
-	if (ra->mmap_miss > MMAP_LOTSAMISS)
+	if (mmap_miss > MMAP_LOTSAMISS)
 		return fpin;
 
 	/*
@@ -2413,13 +2415,15 @@ static struct file *do_async_mmap_readahead(struct vm_fault *vmf,
 	struct file_ra_state *ra = &file->f_ra;
 	struct address_space *mapping = file->f_mapping;
 	struct file *fpin = NULL;
+	unsigned int mmap_miss;
 	pgoff_t offset = vmf->pgoff;
 
 	/* If we don't want any read-ahead, don't bother */
 	if (vmf->vma->vm_flags & VM_RAND_READ)
 		return fpin;
-	if (ra->mmap_miss > 0)
-		ra->mmap_miss--;
+	mmap_miss = READ_ONCE(ra->mmap_miss);
+	if (mmap_miss)
+		WRITE_ONCE(ra->mmap_miss, --mmap_miss);
 	if (PageReadahead(page)) {
 		fpin = maybe_unlock_mmap_for_io(vmf, fpin);
 		page_cache_async_readahead(mapping, ra, file,
@@ -2586,7 +2590,9 @@ void filemap_map_pages(struct vm_fault *vmf,
 	unsigned long max_idx;
 	XA_STATE(xas, &mapping->i_pages, start_pgoff);
 	struct page *page;
+	unsigned long mmap_miss;
 
+	mmap_miss = READ_ONCE(file->f_ra.mmap_miss);
 	rcu_read_lock();
 	xas_for_each(&xas, page, end_pgoff) {
 		if (xas_retry(&xas, page))
@@ -2622,8 +2628,8 @@ void filemap_map_pages(struct vm_fault *vmf,
 		if (page->index >= max_idx)
 			goto unlock;
 
-		if (file->f_ra.mmap_miss > 0)
-			file->f_ra.mmap_miss--;
+		if (mmap_miss > 0)
+			mmap_miss--;
 
 		vmf->address += (xas.xa_index - last_pgoff) << PAGE_SHIFT;
 		if (vmf->pte)
@@ -2643,6 +2649,7 @@ void filemap_map_pages(struct vm_fault *vmf,
 			break;
 	}
 	rcu_read_unlock();
+	WRITE_ONCE(file->f_ra.mmap_miss, mmap_miss);
 }
 EXPORT_SYMBOL(filemap_map_pages);
 
-- 
 Kirill A. Shutemov