1.. SPDX-License-Identifier: GPL-2.0
2
3====
4FUSE
5====
6
7Definitions
8===========
9
10Userspace filesystem:
11  A filesystem in which data and metadata are provided by an ordinary
12  userspace process.  The filesystem can be accessed normally through
13  the kernel interface.
14
15Filesystem daemon:
16  The process(es) providing the data and metadata of the filesystem.
17
18Non-privileged mount (or user mount):
19  A userspace filesystem mounted by a non-privileged (non-root) user.
20  The filesystem daemon is running with the privileges of the mounting
21  user.  NOTE: this is not the same as mounts allowed with the "user"
22  option in /etc/fstab, which is not discussed here.
23
24Filesystem connection:
25  A connection between the filesystem daemon and the kernel.  The
26  connection exists until either the daemon dies, or the filesystem is
27  umounted.  Note that detaching (or lazy umounting) the filesystem
28  does *not* break the connection, in this case it will exist until
29  the last reference to the filesystem is released.
30
31Mount owner:
32  The user who does the mounting.
33
34User:
35  The user who is performing filesystem operations.
36
37What is FUSE?
38=============
39
40FUSE is a userspace filesystem framework.  It consists of a kernel
41module (fuse.ko), a userspace library (libfuse.*) and a mount utility
42(fusermount).
43
44One of the most important features of FUSE is allowing secure,
45non-privileged mounts.  This opens up new possibilities for the use of
46filesystems.  A good example is sshfs: a secure network filesystem
47using the sftp protocol.
48
49The userspace library and utilities are available from the
50`FUSE homepage: <https://github.com/libfuse/>`_
51
52Filesystem type
53===============
54
55The filesystem type given to mount(2) can be one of the following:
56
57    fuse
58      This is the usual way to mount a FUSE filesystem.  The first
59      argument of the mount system call may contain an arbitrary string,
60      which is not interpreted by the kernel.
61
62    fuseblk
63      The filesystem is block device based.  The first argument of the
64      mount system call is interpreted as the name of the device.
65
66Mount options
67=============
68
69fd=N
70  The file descriptor to use for communication between the userspace
71  filesystem and the kernel.  The file descriptor must have been
72  obtained by opening the FUSE device ('/dev/fuse').
73
74rootmode=M
75  The file mode of the filesystem's root in octal representation.
76
77user_id=N
78  The numeric user id of the mount owner.
79
80group_id=N
81  The numeric group id of the mount owner.
82
83default_permissions
84  By default FUSE doesn't check file access permissions, the
85  filesystem is free to implement its access policy or leave it to
86  the underlying file access mechanism (e.g. in case of network
87  filesystems).  This option enables permission checking, restricting
88  access based on file mode.  It is usually useful together with the
89  'allow_other' mount option.
90
91allow_other
92  This option overrides the security measure restricting file access
93  to the user mounting the filesystem.  This option is by default only
94  allowed to root, but this restriction can be removed with a
95  (userspace) configuration option.
96
97max_read=N
98  With this option the maximum size of read operations can be set.
99  The default is infinite.  Note that the size of read requests is
100  limited anyway to 32 pages (which is 128kbyte on i386).
101
102blksize=N
103  Set the block size for the filesystem.  The default is 512.  This
104  option is only valid for 'fuseblk' type mounts.
105
106Control filesystem
107==================
108
109There's a control filesystem for FUSE, which can be mounted by::
110
111  mount -t fusectl none /sys/fs/fuse/connections
112
113Mounting it under the '/sys/fs/fuse/connections' directory makes it
114backwards compatible with earlier versions.
115
116Under the fuse control filesystem each connection has a directory
117named by a unique number.
118
119For each connection the following files exist within this directory:
120
121	waiting
122	  The number of requests which are waiting to be transferred to
123	  userspace or being processed by the filesystem daemon.  If there is
124	  no filesystem activity and 'waiting' is non-zero, then the
125	  filesystem is hung or deadlocked.
126
127	abort
128	  Writing anything into this file will abort the filesystem
129	  connection.  This means that all waiting requests will be aborted an
130	  error returned for all aborted and new requests.
131
132Only the owner of the mount may read or write these files.
133
134Interrupting filesystem operations
135##################################
136
137If a process issuing a FUSE filesystem request is interrupted, the
138following will happen:
139
140  -  If the request is not yet sent to userspace AND the signal is
141     fatal (SIGKILL or unhandled fatal signal), then the request is
142     dequeued and returns immediately.
143
144  -  If the request is not yet sent to userspace AND the signal is not
145     fatal, then an interrupted flag is set for the request.  When
146     the request has been successfully transferred to userspace and
147     this flag is set, an INTERRUPT request is queued.
148
149  -  If the request is already sent to userspace, then an INTERRUPT
150     request is queued.
151
152INTERRUPT requests take precedence over other requests, so the
153userspace filesystem will receive queued INTERRUPTs before any others.
154
155The userspace filesystem may ignore the INTERRUPT requests entirely,
156or may honor them by sending a reply to the *original* request, with
157the error set to EINTR.
158
159It is also possible that there's a race between processing the
160original request and its INTERRUPT request.  There are two possibilities:
161
162  1. The INTERRUPT request is processed before the original request is
163     processed
164
165  2. The INTERRUPT request is processed after the original request has
166     been answered
167
168If the filesystem cannot find the original request, it should wait for
169some timeout and/or a number of new requests to arrive, after which it
170should reply to the INTERRUPT request with an EAGAIN error.  In case
1711) the INTERRUPT request will be requeued.  In case 2) the INTERRUPT
172reply will be ignored.
173
174Aborting a filesystem connection
175================================
176
177It is possible to get into certain situations where the filesystem is
178not responding.  Reasons for this may be:
179
180  a) Broken userspace filesystem implementation
181
182  b) Network connection down
183
184  c) Accidental deadlock
185
186  d) Malicious deadlock
187
188(For more on c) and d) see later sections)
189
190In either of these cases it may be useful to abort the connection to
191the filesystem.  There are several ways to do this:
192
193  - Kill the filesystem daemon.  Works in case of a) and b)
194
195  - Kill the filesystem daemon and all users of the filesystem.  Works
196    in all cases except some malicious deadlocks
197
198  - Use forced umount (umount -f).  Works in all cases but only if
199    filesystem is still attached (it hasn't been lazy unmounted)
200
201  - Abort filesystem through the FUSE control filesystem.  Most
202    powerful method, always works.
203
204How do non-privileged mounts work?
205==================================
206
207Since the mount() system call is a privileged operation, a helper
208program (fusermount) is needed, which is installed setuid root.
209
210The implication of providing non-privileged mounts is that the mount
211owner must not be able to use this capability to compromise the
212system.  Obvious requirements arising from this are:
213
214 A) mount owner should not be able to get elevated privileges with the
215    help of the mounted filesystem
216
217 B) mount owner should not get illegitimate access to information from
218    other users' and the super user's processes
219
220 C) mount owner should not be able to induce undesired behavior in
221    other users' or the super user's processes
222
223How are requirements fulfilled?
224===============================
225
226 A) The mount owner could gain elevated privileges by either:
227
228    1. creating a filesystem containing a device file, then opening this device
229
230    2. creating a filesystem containing a suid or sgid application, then executing this application
231
232    The solution is not to allow opening device files and ignore
233    setuid and setgid bits when executing programs.  To ensure this
234    fusermount always adds "nosuid" and "nodev" to the mount options
235    for non-privileged mounts.
236
237 B) If another user is accessing files or directories in the
238    filesystem, the filesystem daemon serving requests can record the
239    exact sequence and timing of operations performed.  This
240    information is otherwise inaccessible to the mount owner, so this
241    counts as an information leak.
242
243    The solution to this problem will be presented in point 2) of C).
244
245 C) There are several ways in which the mount owner can induce
246    undesired behavior in other users' processes, such as:
247
248     1) mounting a filesystem over a file or directory which the mount
249        owner could otherwise not be able to modify (or could only
250        make limited modifications).
251
252        This is solved in fusermount, by checking the access
253        permissions on the mountpoint and only allowing the mount if
254        the mount owner can do unlimited modification (has write
255        access to the mountpoint, and mountpoint is not a "sticky"
256        directory)
257
258     2) Even if 1) is solved the mount owner can change the behavior
259        of other users' processes.
260
261         i) It can slow down or indefinitely delay the execution of a
262            filesystem operation creating a DoS against the user or the
263            whole system.  For example a suid application locking a
264            system file, and then accessing a file on the mount owner's
265            filesystem could be stopped, and thus causing the system
266            file to be locked forever.
267
268         ii) It can present files or directories of unlimited length, or
269             directory structures of unlimited depth, possibly causing a
270             system process to eat up diskspace, memory or other
271             resources, again causing *DoS*.
272
273	The solution to this as well as B) is not to allow processes
274	to access the filesystem, which could otherwise not be
275	monitored or manipulated by the mount owner.  Since if the
276	mount owner can ptrace a process, it can do all of the above
277	without using a FUSE mount, the same criteria as used in
278	ptrace can be used to check if a process is allowed to access
279	the filesystem or not.
280
281	Note that the *ptrace* check is not strictly necessary to
282	prevent C/2/i, it is enough to check if mount owner has enough
283	privilege to send signal to the process accessing the
284	filesystem, since *SIGSTOP* can be used to get a similar effect.
285
286I think these limitations are unacceptable?
287===========================================
288
289If a sysadmin trusts the users enough, or can ensure through other
290measures, that system processes will never enter non-privileged
291mounts, it can relax the last limitation in several ways:
292
293  - With the 'user_allow_other' config option. If this config option is
294    set, the mounting user can add the 'allow_other' mount option which
295    disables the check for other users' processes.
296
297    User namespaces have an unintuitive interaction with 'allow_other':
298    an unprivileged user - normally restricted from mounting with
299    'allow_other' - could do so in a user namespace where they're
300    privileged. If any process could access such an 'allow_other' mount
301    this would give the mounting user the ability to manipulate
302    processes in user namespaces where they're unprivileged. For this
303    reason 'allow_other' restricts access to users in the same userns
304    or a descendant.
305
306  - With the 'allow_sys_admin_access' module option. If this option is
307    set, super user's processes have unrestricted access to mounts
308    irrespective of allow_other setting or user namespace of the
309    mounting user.
310
311Note that both of these relaxations expose the system to potential
312information leak or *DoS* as described in points B and C/2/i-ii in the
313preceding section.
314
315Kernel - userspace interface
316============================
317
318The following diagram shows how a filesystem operation (in this
319example unlink) is performed in FUSE. ::
320
321
322 |  "rm /mnt/fuse/file"               |  FUSE filesystem daemon
323 |                                    |
324 |                                    |  >sys_read()
325 |                                    |    >fuse_dev_read()
326 |                                    |      >request_wait()
327 |                                    |        [sleep on fc->waitq]
328 |                                    |
329 |  >sys_unlink()                     |
330 |    >fuse_unlink()                  |
331 |      [get request from             |
332 |       fc->unused_list]             |
333 |      >request_send()               |
334 |        [queue req on fc->pending]  |
335 |        [wake up fc->waitq]         |        [woken up]
336 |        >request_wait_answer()      |
337 |          [sleep on req->waitq]     |
338 |                                    |      <request_wait()
339 |                                    |      [remove req from fc->pending]
340 |                                    |      [copy req to read buffer]
341 |                                    |      [add req to fc->processing]
342 |                                    |    <fuse_dev_read()
343 |                                    |  <sys_read()
344 |                                    |
345 |                                    |  [perform unlink]
346 |                                    |
347 |                                    |  >sys_write()
348 |                                    |    >fuse_dev_write()
349 |                                    |      [look up req in fc->processing]
350 |                                    |      [remove from fc->processing]
351 |                                    |      [copy write buffer to req]
352 |          [woken up]                |      [wake up req->waitq]
353 |                                    |    <fuse_dev_write()
354 |                                    |  <sys_write()
355 |        <request_wait_answer()      |
356 |      <request_send()               |
357 |      [add request to               |
358 |       fc->unused_list]             |
359 |    <fuse_unlink()                  |
360 |  <sys_unlink()                     |
361
362.. note:: Everything in the description above is greatly simplified
363
364There are a couple of ways in which to deadlock a FUSE filesystem.
365Since we are talking about unprivileged userspace programs,
366something must be done about these.
367
368**Scenario 1 -  Simple deadlock**::
369
370 |  "rm /mnt/fuse/file"               |  FUSE filesystem daemon
371 |                                    |
372 |  >sys_unlink("/mnt/fuse/file")     |
373 |    [acquire inode semaphore        |
374 |     for "file"]                    |
375 |    >fuse_unlink()                  |
376 |      [sleep on req->waitq]         |
377 |                                    |  <sys_read()
378 |                                    |  >sys_unlink("/mnt/fuse/file")
379 |                                    |    [acquire inode semaphore
380 |                                    |     for "file"]
381 |                                    |    *DEADLOCK*
382
383The solution for this is to allow the filesystem to be aborted.
384
385**Scenario 2 - Tricky deadlock**
386
387
388This one needs a carefully crafted filesystem.  It's a variation on
389the above, only the call back to the filesystem is not explicit,
390but is caused by a pagefault. ::
391
392 |  Kamikaze filesystem thread 1      |  Kamikaze filesystem thread 2
393 |                                    |
394 |  [fd = open("/mnt/fuse/file")]     |  [request served normally]
395 |  [mmap fd to 'addr']               |
396 |  [close fd]                        |  [FLUSH triggers 'magic' flag]
397 |  [read a byte from addr]           |
398 |    >do_page_fault()                |
399 |      [find or create page]         |
400 |      [lock page]                   |
401 |      >fuse_readpage()              |
402 |         [queue READ request]       |
403 |         [sleep on req->waitq]      |
404 |                                    |  [read request to buffer]
405 |                                    |  [create reply header before addr]
406 |                                    |  >sys_write(addr - headerlength)
407 |                                    |    >fuse_dev_write()
408 |                                    |      [look up req in fc->processing]
409 |                                    |      [remove from fc->processing]
410 |                                    |      [copy write buffer to req]
411 |                                    |        >do_page_fault()
412 |                                    |           [find or create page]
413 |                                    |           [lock page]
414 |                                    |           * DEADLOCK *
415
416The solution is basically the same as above.
417
418An additional problem is that while the write buffer is being copied
419to the request, the request must not be interrupted/aborted.  This is
420because the destination address of the copy may not be valid after the
421request has returned.
422
423This is solved with doing the copy atomically, and allowing abort
424while the page(s) belonging to the write buffer are faulted with
425get_user_pages().  The 'req->locked' flag indicates when the copy is
426taking place, and abort is delayed until this flag is unset.
427