1.. SPDX-License-Identifier: GPL-2.0
2
3===================================
4File management in the Linux kernel
5===================================
6
7This document describes how locking for files (struct file)
8and file descriptor table (struct files) works.
9
10Up until 2.6.12, the file descriptor table has been protected
11with a lock (files->file_lock) and reference count (files->count).
12->file_lock protected accesses to all the file related fields
13of the table. ->count was used for sharing the file descriptor
14table between tasks cloned with CLONE_FILES flag. Typically
15this would be the case for posix threads. As with the common
16refcounting model in the kernel, the last task doing
17a put_files_struct() frees the file descriptor (fd) table.
18The files (struct file) themselves are protected using
19reference count (->f_count).
20
21In the new lock-free model of file descriptor management,
22the reference counting is similar, but the locking is
23based on RCU. The file descriptor table contains multiple
24elements - the fd sets (open_fds and close_on_exec, the
25array of file pointers, the sizes of the sets and the array
26etc.). In order for the updates to appear atomic to
27a lock-free reader, all the elements of the file descriptor
28table are in a separate structure - struct fdtable.
29files_struct contains a pointer to struct fdtable through
30which the actual fd table is accessed. Initially the
31fdtable is embedded in files_struct itself. On a subsequent
32expansion of fdtable, a new fdtable structure is allocated
33and files->fdtab points to the new structure. The fdtable
34structure is freed with RCU and lock-free readers either
35see the old fdtable or the new fdtable making the update
36appear atomic. Here are the locking rules for
37the fdtable structure -
38
391. All references to the fdtable must be done through
40   the files_fdtable() macro::
41
42	struct fdtable *fdt;
43
44	rcu_read_lock();
45
46	fdt = files_fdtable(files);
47	....
48	if (n <= fdt->max_fds)
49		....
50	...
51	rcu_read_unlock();
52
53   files_fdtable() uses rcu_dereference() macro which takes care of
54   the memory barrier requirements for lock-free dereference.
55   The fdtable pointer must be read within the read-side
56   critical section.
57
582. Reading of the fdtable as described above must be protected
59   by rcu_read_lock()/rcu_read_unlock().
60
613. For any update to the fd table, files->file_lock must
62   be held.
63
644. To look up the file structure given an fd, a reader
65   must use either lookup_fdget_rcu() or files_lookup_fdget_rcu() APIs. These
66   take care of barrier requirements due to lock-free lookup.
67
68   An example::
69
70	struct file *file;
71
72	rcu_read_lock();
73	file = lookup_fdget_rcu(fd);
74	rcu_read_unlock();
75	if (file) {
76		...
77                fput(file);
78	}
79	....
80
815. Since both fdtable and file structures can be looked up
82   lock-free, they must be installed using rcu_assign_pointer()
83   API. If they are looked up lock-free, rcu_dereference()
84   must be used. However it is advisable to use files_fdtable()
85   and lookup_fdget_rcu()/files_lookup_fdget_rcu() which take care of these
86   issues.
87
886. While updating, the fdtable pointer must be looked up while
89   holding files->file_lock. If ->file_lock is dropped, then
90   another thread expand the files thereby creating a new
91   fdtable and making the earlier fdtable pointer stale.
92
93   For example::
94
95	spin_lock(&files->file_lock);
96	fd = locate_fd(files, file, start);
97	if (fd >= 0) {
98		/* locate_fd() may have expanded fdtable, load the ptr */
99		fdt = files_fdtable(files);
100		__set_open_fd(fd, fdt);
101		__clear_close_on_exec(fd, fdt);
102		spin_unlock(&files->file_lock);
103	.....
104
105   Since locate_fd() can drop ->file_lock (and reacquire ->file_lock),
106   the fdtable pointer (fdt) must be loaded after locate_fd().
107
108On newer kernels rcu based file lookup has been switched to rely on
109SLAB_TYPESAFE_BY_RCU instead of call_rcu(). It isn't sufficient anymore
110to just acquire a reference to the file in question under rcu using
111atomic_long_inc_not_zero() since the file might have already been
112recycled and someone else might have bumped the reference. In other
113words, callers might see reference count bumps from newer users. For
114this is reason it is necessary to verify that the pointer is the same
115before and after the reference count increment. This pattern can be seen
116in get_file_rcu() and __files_get_rcu().
117
118In addition, it isn't possible to access or check fields in struct file
119without first acquiring a reference on it under rcu lookup. Not doing
120that was always very dodgy and it was only usable for non-pointer data
121in struct file. With SLAB_TYPESAFE_BY_RCU it is necessary that callers
122either first acquire a reference or they must hold the files_lock of the
123fdtable.
124