file: convert to SLAB_TYPESAFE_BY_RCU

In recent discussions around some performance improvements in the file
handling area we discussed switching the file cache to rely on
SLAB_TYPESAFE_BY_RCU which allows us to get rid of call_rcu() based
freeing for files completely. This is a pretty sensitive change overall
but it might actually be worth doing.

The main downside is the subtlety. The other one is that we should
really wait for Jann's patch to land that enables KASAN to handle
SLAB_TYPESAFE_BY_RCU UAFs. Currently it doesn't but a patch for this
exists.

With SLAB_TYPESAFE_BY_RCU objects may be freed and reused multiple times
which requires a few changes. So it isn't sufficient anymore to just
acquire a reference to the file in question under rcu using
atomic_long_inc_not_zero() since the file might have already been
recycled and someone else might have bumped the reference.

In other words, callers might see reference count bumps from newer
users. For this reason it is necessary to verify that the pointer is the
same before and after the reference count increment. This pattern can be
seen in get_file_rcu() and __files_get_rcu().

In addition, it isn't possible to access or check fields in struct file
without first aqcuiring a reference on it. Not doing that was always
very dodgy and it was only usable for non-pointer data in struct file.
With SLAB_TYPESAFE_BY_RCU it is necessary that callers first acquire a
reference under rcu or they must hold the files_lock of the fdtable.
Failing to do either one of this is a bug.

Thanks to Jann for pointing out that we need to ensure memory ordering
between reallocations and pointer check by ensuring that all subsequent
loads have a dependency on the second load in get_file_rcu() and
providing a fixup that was folded into this patch.

Cc: Jann Horn <jannh@google.com>
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Christian Brauner <brauner@kernel.org>

Documentation/filesystems/files.rst

@@ -62,7 +62,7 @@ the fdtable structure -
    be held.
 
 4. To look up the file structure given an fd, a reader
-   must use either lookup_fd_rcu() or files_lookup_fd_rcu() APIs. These
+   must use either lookup_fdget_rcu() or files_lookup_fdget_rcu() APIs. These
    take care of barrier requirements due to lock-free lookup.
 
    An example::
@@ -70,43 +70,22 @@ the fdtable structure -
 	struct file *file;
 
 	rcu_read_lock();
-	file = lookup_fd_rcu(fd);
-	if (file) {
-		...
-	}
-	....
+	file = lookup_fdget_rcu(fd);
 	rcu_read_unlock();
-
-5. Handling of the file structures is special. Since the look-up
-   of the fd (fget()/fget_light()) are lock-free, it is possible
-   that look-up may race with the last put() operation on the
-   file structure. This is avoided using atomic_long_inc_not_zero()
-   on ->f_count::
-
-	rcu_read_lock();
-	file = files_lookup_fd_rcu(files, fd);
 	if (file) {
-		if (atomic_long_inc_not_zero(&file->f_count))
-			*fput_needed = 1;
-		else
-		/* Didn't get the reference, someone's freed */
-			file = NULL;
+		...
+                fput(file);
 	}
-	rcu_read_unlock();
 	....
-	return file;
-
-   atomic_long_inc_not_zero() detects if refcounts is already zero or
-   goes to zero during increment. If it does, we fail
-   fget()/fget_light().
 
-6. Since both fdtable and file structures can be looked up
+5. Since both fdtable and file structures can be looked up
    lock-free, they must be installed using rcu_assign_pointer()
    API. If they are looked up lock-free, rcu_dereference()
    must be used. However it is advisable to use files_fdtable()
-   and lookup_fd_rcu()/files_lookup_fd_rcu() which take care of these issues.
+   and lookup_fdget_rcu()/files_lookup_fdget_rcu() which take care of these
+   issues.
 
-7. While updating, the fdtable pointer must be looked up while
+6. While updating, the fdtable pointer must be looked up while
    holding files->file_lock. If ->file_lock is dropped, then
    another thread expand the files thereby creating a new
    fdtable and making the earlier fdtable pointer stale.
@@ -126,3 +105,19 @@ the fdtable structure -
    Since locate_fd() can drop ->file_lock (and reacquire ->file_lock),
    the fdtable pointer (fdt) must be loaded after locate_fd().
 
+On newer kernels rcu based file lookup has been switched to rely on
+SLAB_TYPESAFE_BY_RCU instead of call_rcu(). It isn't sufficient anymore
+to just acquire a reference to the file in question under rcu using
+atomic_long_inc_not_zero() since the file might have already been
+recycled and someone else might have bumped the reference. In other
+words, callers might see reference count bumps from newer users. For
+this is reason it is necessary to verify that the pointer is the same
+before and after the reference count increment. This pattern can be seen
+in get_file_rcu() and __files_get_rcu().
+
+In addition, it isn't possible to access or check fields in struct file
+without first aqcuiring a reference on it under rcu lookup. Not doing
+that was always very dodgy and it was only usable for non-pointer data
+in struct file. With SLAB_TYPESAFE_BY_RCU it is necessary that callers
+either first acquire a reference or they must hold the files_lock of the
+fdtable.

arch/powerpc/platforms/cell/spufs/coredump.c

@@ -66,7 +66,7 @@ static int match_context(const void *v, struct file *file, unsigned fd)
  */
 static struct spu_context *coredump_next_context(int *fd)
 {
-	struct spu_context *ctx;
+	struct spu_context *ctx = NULL;
 	struct file *file;
 	int n = iterate_fd(current->files, *fd, match_context, NULL);
 	if (!n)
@@ -74,10 +74,13 @@ static struct spu_context *coredump_next_context(int *fd)
 	*fd = n - 1;
 
 	rcu_read_lock();
-	file = lookup_fd_rcu(*fd);
-	ctx = SPUFS_I(file_inode(file))->i_ctx;
-	get_spu_context(ctx);
+	file = lookup_fdget_rcu(*fd);
 	rcu_read_unlock();
+	if (file) {
+		ctx = SPUFS_I(file_inode(file))->i_ctx;
+		get_spu_context(ctx);
+		fput(file);
+	}
 
 	return ctx;
 }

drivers/gpu/drm/i915/gem/i915_gem_mman.c

@@ -916,9 +916,7 @@ static struct file *mmap_singleton(struct drm_i915_private *i915)
 	struct file *file;
 
 	rcu_read_lock();
-	file = READ_ONCE(i915->gem.mmap_singleton);
-	if (file && !get_file_rcu(file))
-		file = NULL;
+	file = get_file_rcu(&i915->gem.mmap_singleton);
 	rcu_read_unlock();
 	if (file)
 		return file;

fs/file.c

@@ -853,8 +853,79 @@ void do_close_on_exec(struct files_struct *files)
 	spin_unlock(&files->file_lock);
 }
 
+static struct file *__get_file_rcu(struct file __rcu **f)
+{
+	struct file __rcu *file;
+	struct file __rcu *file_reloaded;
+	struct file __rcu *file_reloaded_cmp;
+
+	file = rcu_dereference_raw(*f);
+	if (!file)
+		return NULL;
+
+	if (unlikely(!atomic_long_inc_not_zero(&file->f_count)))
+		return ERR_PTR(-EAGAIN);
+
+	file_reloaded = rcu_dereference_raw(*f);
+
+	/*
+	 * Ensure that all accesses have a dependency on the load from
+	 * rcu_dereference_raw() above so we get correct ordering
+	 * between reuse/allocation and the pointer check below.
+	 */
+	file_reloaded_cmp = file_reloaded;
+	OPTIMIZER_HIDE_VAR(file_reloaded_cmp);
+
+	/*
+	 * atomic_long_inc_not_zero() above provided a full memory
+	 * barrier when we acquired a reference.
+	 *
+	 * This is paired with the write barrier from assigning to the
+	 * __rcu protected file pointer so that if that pointer still
+	 * matches the current file, we know we have successfully
+	 * acquired a reference to the right file.
+	 *
+	 * If the pointers don't match the file has been reallocated by
+	 * SLAB_TYPESAFE_BY_RCU.
+	 */
+	if (file == file_reloaded_cmp)
+		return file_reloaded;
+
+	fput(file);
+	return ERR_PTR(-EAGAIN);
+}
+
+/**
+ * get_file_rcu - try go get a reference to a file under rcu
+ * @f: the file to get a reference on
+ *
+ * This function tries to get a reference on @f carefully verifying that
+ * @f hasn't been reused.
+ *
+ * This function should rarely have to be used and only by users who
+ * understand the implications of SLAB_TYPESAFE_BY_RCU. Try to avoid it.
+ *
+ * Return: Returns @f with the reference count increased or NULL.
+ */
+struct file *get_file_rcu(struct file __rcu **f)
+{
+	for (;;) {
+		struct file __rcu *file;
+
+		file = __get_file_rcu(f);
+		if (unlikely(!file))
+			return NULL;
+
+		if (unlikely(IS_ERR(file)))
+			continue;
+
+		return file;
+	}
+}
+EXPORT_SYMBOL_GPL(get_file_rcu);
+
 static inline struct file *__fget_files_rcu(struct files_struct *files,
-	unsigned int fd, fmode_t mask)
+       unsigned int fd, fmode_t mask)
 {
 	for (;;) {
 		struct file *file;
@@ -865,12 +936,6 @@ static inline struct file *__fget_files_rcu(struct files_struct *files,
 			return NULL;
 
 		fdentry = fdt->fd + array_index_nospec(fd, fdt->max_fds);
-		file = rcu_dereference_raw(*fdentry);
-		if (unlikely(!file))
-			return NULL;
-
-		if (unlikely(file->f_mode & mask))
-			return NULL;
 
 		/*
 		 * Ok, we have a file pointer. However, because we do
@@ -879,10 +944,15 @@ static inline struct file *__fget_files_rcu(struct files_struct *files,
 		 *
 		 * Such a race can take two forms:
 		 *
-		 *  (a) the file ref already went down to zero,
-		 *      and get_file_rcu() fails. Just try again:
+		 *  (a) the file ref already went down to zero and the
+		 *      file hasn't been reused yet or the file count
+		 *      isn't zero but the file has already been reused.
 		 */
-		if (unlikely(!get_file_rcu(file)))
+		file = __get_file_rcu(fdentry);
+		if (unlikely(!file))
+			return NULL;
+
+		if (unlikely(IS_ERR(file)))
 			continue;
 
 		/*
@@ -893,12 +963,20 @@ static inline struct file *__fget_files_rcu(struct files_struct *files,
 		 *
 		 * If so, we need to put our ref and try again.
 		 */
-		if (unlikely(rcu_dereference_raw(files->fdt) != fdt) ||
-		    unlikely(rcu_dereference_raw(*fdentry) != file)) {
+		if (unlikely(rcu_dereference_raw(files->fdt) != fdt)) {
 			fput(file);
 			continue;
 		}
 
+		/*
+		 * This isn't the file we're looking for or we're not
+		 * allowed to get a reference to it.
+		 */
+		if (unlikely(file->f_mode & mask)) {
+			fput(file);
+			return NULL;
+		}
+
 		/*
 		 * Ok, we have a ref to the file, and checked that it
 		 * still exists.
@@ -948,7 +1026,14 @@ struct file *fget_task(struct task_struct *task, unsigned int fd)
 	return file;
 }
 
-struct file *task_lookup_fd_rcu(struct task_struct *task, unsigned int fd)
+struct file *lookup_fdget_rcu(unsigned int fd)
+{
+	return __fget_files_rcu(current->files, fd, 0);
+
+}
+EXPORT_SYMBOL_GPL(lookup_fdget_rcu);
+
+struct file *task_lookup_fdget_rcu(struct task_struct *task, unsigned int fd)
 {
 	/* Must be called with rcu_read_lock held */
 	struct files_struct *files;
@@ -957,13 +1042,13 @@ struct file *task_lookup_fd_rcu(struct task_struct *task, unsigned int fd)
 	task_lock(task);
 	files = task->files;
 	if (files)
-		file = files_lookup_fd_rcu(files, fd);
+		file = __fget_files_rcu(files, fd, 0);
 	task_unlock(task);
 
 	return file;
 }
 
-struct file *task_lookup_next_fd_rcu(struct task_struct *task, unsigned int *ret_fd)
+struct file *task_lookup_next_fdget_rcu(struct task_struct *task, unsigned int *ret_fd)
 {
 	/* Must be called with rcu_read_lock held */
 	struct files_struct *files;
@@ -974,7 +1059,7 @@ struct file *task_lookup_next_fd_rcu(struct task_struct *task, unsigned int *ret
 	files = task->files;
 	if (files) {
 		for (; fd < files_fdtable(files)->max_fds; fd++) {
-			file = files_lookup_fd_rcu(files, fd);
+			file = __fget_files_rcu(files, fd, 0);
 			if (file)
 				break;
 		}
@@ -983,7 +1068,7 @@ struct file *task_lookup_next_fd_rcu(struct task_struct *task, unsigned int *ret
 	*ret_fd = fd;
 	return file;
 }
-EXPORT_SYMBOL(task_lookup_next_fd_rcu);
+EXPORT_SYMBOL(task_lookup_next_fdget_rcu);
 
 /*
  * Lightweight file lookup - no refcnt increment if fd table isn't shared.
@@ -1272,12 +1357,16 @@ SYSCALL_DEFINE2(dup2, unsigned int, oldfd, unsigned int, newfd)
 {
 	if (unlikely(newfd == oldfd)) { /* corner case */
 		struct files_struct *files = current->files;
+		struct file *f;
 		int retval = oldfd;
 
 		rcu_read_lock();
-		if (!files_lookup_fd_rcu(files, oldfd))
+		f = __fget_files_rcu(files, oldfd, 0);
+		if (!f)
 			retval = -EBADF;
 		rcu_read_unlock();
+		if (f)
+			fput(f);
 		return retval;
 	}
 	return ksys_dup3(oldfd, newfd, 0);

fs/file_table.c

@@ -65,33 +65,33 @@ static void file_free_rcu(struct rcu_head *head)
 {
 	struct file *f = container_of(head, struct file, f_rcuhead);
 
-	put_cred(f->f_cred);
-	if (unlikely(f->f_mode & FMODE_BACKING))
-		kfree(backing_file(f));
-	else
-		kmem_cache_free(filp_cachep, f);
+	kfree(backing_file(f));
 }
 
 static inline void file_free(struct file *f)
 {
 	security_file_free(f);
-	if (unlikely(f->f_mode & FMODE_BACKING))
-		path_put(backing_file_real_path(f));
 	if (likely(!(f->f_mode & FMODE_NOACCOUNT)))
 		percpu_counter_dec(&nr_files);
-	call_rcu(&f->f_rcuhead, file_free_rcu);
+	put_cred(f->f_cred);
+	if (unlikely(f->f_mode & FMODE_BACKING)) {
+		path_put(backing_file_real_path(f));
+		call_rcu(&f->f_rcuhead, file_free_rcu);
+	} else {
+		kmem_cache_free(filp_cachep, f);
+	}
 }
 
 void release_empty_file(struct file *f)
 {
 	WARN_ON_ONCE(f->f_mode & (FMODE_BACKING | FMODE_OPENED));
-	/* Uhm, we better find out who grabs references to an unopened file. */
-	WARN_ON_ONCE(atomic_long_cmpxchg(&f->f_count, 1, 0) != 1);
-	security_file_free(f);
-	put_cred(f->f_cred);
-	if (likely(!(f->f_mode & FMODE_NOACCOUNT)))
-		percpu_counter_dec(&nr_files);
-	kmem_cache_free(filp_cachep, f);
+	if (atomic_long_dec_and_test(&f->f_count)) {
+		security_file_free(f);
+		put_cred(f->f_cred);
+		if (likely(!(f->f_mode & FMODE_NOACCOUNT)))
+			percpu_counter_dec(&nr_files);
+		kmem_cache_free(filp_cachep, f);
+	}
 }
 
 /*
@@ -176,7 +176,6 @@ static int init_file(struct file *f, int flags, const struct cred *cred)
 		return error;
 	}
 
-	atomic_long_set(&f->f_count, 1);
 	rwlock_init(&f->f_owner.lock);
 	spin_lock_init(&f->f_lock);
 	mutex_init(&f->f_pos_lock);
@@ -184,6 +183,12 @@ static int init_file(struct file *f, int flags, const struct cred *cred)
 	f->f_mode = OPEN_FMODE(flags);
 	/* f->f_version: 0 */
 
+	/*
+	 * We're SLAB_TYPESAFE_BY_RCU so initialize f_count last. While
+	 * fget-rcu pattern users need to be able to handle spurious
+	 * refcount bumps we should reinitialize the reused file first.
+	 */
+	atomic_long_set(&f->f_count, 1);
 	return 0;
 }
 
@@ -483,7 +488,8 @@ EXPORT_SYMBOL(__fput_sync);
 void __init files_init(void)
 {
 	filp_cachep = kmem_cache_create("filp", sizeof(struct file), 0,
-			SLAB_HWCACHE_ALIGN | SLAB_PANIC | SLAB_ACCOUNT, NULL);
+				SLAB_TYPESAFE_BY_RCU | SLAB_HWCACHE_ALIGN |
+				SLAB_PANIC | SLAB_ACCOUNT, NULL);
 	percpu_counter_init(&nr_files, 0, GFP_KERNEL);
 }
 

fs/gfs2/glock.c

@@ -2717,16 +2717,19 @@ static struct file *gfs2_glockfd_next_file(struct gfs2_glockfd_iter *i)
 	for(;; i->fd++) {
 		struct inode *inode;
 
-		i->file = task_lookup_next_fd_rcu(i->task, &i->fd);
+		i->file = task_lookup_next_fdget_rcu(i->task, &i->fd);
 		if (!i->file) {
 			i->fd = 0;
 			break;
 		}
+
 		inode = file_inode(i->file);
-		if (inode->i_sb != i->sb)
-			continue;
-		if (get_file_rcu(i->file))
+		if (inode->i_sb == i->sb)
 			break;
+
+		rcu_read_unlock();
+		fput(i->file);
+		rcu_read_lock();
 	}
 	rcu_read_unlock();
 	return i->file;

fs/notify/dnotify/dnotify.c

@@ -265,7 +265,7 @@ int fcntl_dirnotify(int fd, struct file *filp, unsigned int arg)
 	struct dnotify_struct *dn;
 	struct inode *inode;
 	fl_owner_t id = current->files;
-	struct file *f;
+	struct file *f = NULL;
 	int destroy = 0, error = 0;
 	__u32 mask;
 
@@ -345,7 +345,7 @@ int fcntl_dirnotify(int fd, struct file *filp, unsigned int arg)
 	}
 
 	rcu_read_lock();
-	f = lookup_fd_rcu(fd);
+	f = lookup_fdget_rcu(fd);
 	rcu_read_unlock();
 
 	/* if (f != filp) means that we lost a race and another task/thread
@@ -392,6 +392,8 @@ int fcntl_dirnotify(int fd, struct file *filp, unsigned int arg)
 		fsnotify_put_mark(new_fsn_mark);
 	if (dn)
 		kmem_cache_free(dnotify_struct_cache, dn);
+	if (f)
+		fput(f);
 	return error;
 }
 

fs/proc/fd.c

@@ -113,10 +113,12 @@ static bool tid_fd_mode(struct task_struct *task, unsigned fd, fmode_t *mode)
 	struct file *file;
 
 	rcu_read_lock();
-	file = task_lookup_fd_rcu(task, fd);
-	if (file)
-		*mode = file->f_mode;
+	file = task_lookup_fdget_rcu(task, fd);
 	rcu_read_unlock();
+	if (file) {
+		*mode = file->f_mode;
+		fput(file);
+	}
 	return !!file;
 }
 
@@ -259,12 +261,13 @@ static int proc_readfd_common(struct file *file, struct dir_context *ctx,
 		char name[10 + 1];
 		unsigned int len;
 
-		f = task_lookup_next_fd_rcu(p, &fd);
+		f = task_lookup_next_fdget_rcu(p, &fd);
 		ctx->pos = fd + 2LL;
 		if (!f)
 			break;
 		data.mode = f->f_mode;
 		rcu_read_unlock();
+		fput(f);
 		data.fd = fd;
 
 		len = snprintf(name, sizeof(name), "%u", fd);

include/linux/fdtable.h

@@ -98,20 +98,9 @@ static inline struct file *files_lookup_fd_locked(struct files_struct *files, un
 	return files_lookup_fd_raw(files, fd);
 }
 
-static inline struct file *files_lookup_fd_rcu(struct files_struct *files, unsigned int fd)
-{
-	RCU_LOCKDEP_WARN(!rcu_read_lock_held(),
-			   "suspicious rcu_dereference_check() usage");
-	return files_lookup_fd_raw(files, fd);
-}
-
-static inline struct file *lookup_fd_rcu(unsigned int fd)
-{
-	return files_lookup_fd_rcu(current->files, fd);
-}
-
-struct file *task_lookup_fd_rcu(struct task_struct *task, unsigned int fd);
-struct file *task_lookup_next_fd_rcu(struct task_struct *task, unsigned int *fd);
+struct file *lookup_fdget_rcu(unsigned int fd);
+struct file *task_lookup_fdget_rcu(struct task_struct *task, unsigned int fd);
+struct file *task_lookup_next_fdget_rcu(struct task_struct *task, unsigned int *fd);
 
 struct task_struct;
 

include/linux/fs.h

@@ -1042,7 +1042,9 @@ static inline struct file *get_file(struct file *f)
 	atomic_long_inc(&f->f_count);
 	return f;
 }
-#define get_file_rcu(x) atomic_long_inc_not_zero(&(x)->f_count)
+
+struct file *get_file_rcu(struct file __rcu **f);
+
 #define file_count(x)	atomic_long_read(&(x)->f_count)
 
 #define	MAX_NON_LFS	((1UL<<31) - 1)

kernel/bpf/task_iter.c

@@ -308,11 +308,9 @@ task_file_seq_get_next(struct bpf_iter_seq_task_file_info *info)
 	rcu_read_lock();
 	for (;; curr_fd++) {
 		struct file *f;
-		f = task_lookup_next_fd_rcu(curr_task, &curr_fd);
+		f = task_lookup_next_fdget_rcu(curr_task, &curr_fd);
 		if (!f)
 			break;
-		if (!get_file_rcu(f))
-			continue;
 
 		/* set info->fd */
 		info->fd = curr_fd;

kernel/fork.c

@@ -1492,9 +1492,7 @@ struct file *get_mm_exe_file(struct mm_struct *mm)
 	struct file *exe_file;
 
 	rcu_read_lock();
-	exe_file = rcu_dereference(mm->exe_file);
-	if (exe_file && !get_file_rcu(exe_file))
-		exe_file = NULL;
+	exe_file = get_file_rcu(&mm->exe_file);
 	rcu_read_unlock();
 	return exe_file;
 }

kernel/kcmp.c

@@ -64,8 +64,10 @@ get_file_raw_ptr(struct task_struct *task, unsigned int idx)
 	struct file *file;
 
 	rcu_read_lock();
-	file = task_lookup_fd_rcu(task, idx);
+	file = task_lookup_fdget_rcu(task, idx);
 	rcu_read_unlock();
+	if (file)
+		fput(file);
 
 	return file;
 }