0001 .. SPDX-License-Identifier: GPL-2.0
0002
0003 ====
0004 FUSE
0005 ====
0006
0007 Definitions
0008 ===========
0009
0010 Userspace filesystem:
0011 A filesystem in which data and metadata are provided by an ordinary
0012 userspace process. The filesystem can be accessed normally through
0013 the kernel interface.
0014
0015 Filesystem daemon:
0016 The process(es) providing the data and metadata of the filesystem.
0017
0018 Non-privileged mount (or user mount):
0019 A userspace filesystem mounted by a non-privileged (non-root) user.
0020 The filesystem daemon is running with the privileges of the mounting
0021 user. NOTE: this is not the same as mounts allowed with the "user"
0022 option in /etc/fstab, which is not discussed here.
0023
0024 Filesystem connection:
0025 A connection between the filesystem daemon and the kernel. The
0026 connection exists until either the daemon dies, or the filesystem is
0027 umounted. Note that detaching (or lazy umounting) the filesystem
0028 does *not* break the connection, in this case it will exist until
0029 the last reference to the filesystem is released.
0030
0031 Mount owner:
0032 The user who does the mounting.
0033
0034 User:
0035 The user who is performing filesystem operations.
0036
0037 What is FUSE?
0038 =============
0039
0040 FUSE is a userspace filesystem framework. It consists of a kernel
0041 module (fuse.ko), a userspace library (libfuse.*) and a mount utility
0042 (fusermount).
0043
0044 One of the most important features of FUSE is allowing secure,
0045 non-privileged mounts. This opens up new possibilities for the use of
0046 filesystems. A good example is sshfs: a secure network filesystem
0047 using the sftp protocol.
0048
0049 The userspace library and utilities are available from the
0050 `FUSE homepage: <https://github.com/libfuse/>`_
0051
0052 Filesystem type
0053 ===============
0054
0055 The filesystem type given to mount(2) can be one of the following:
0056
0057 fuse
0058 This is the usual way to mount a FUSE filesystem. The first
0059 argument of the mount system call may contain an arbitrary string,
0060 which is not interpreted by the kernel.
0061
0062 fuseblk
0063 The filesystem is block device based. The first argument of the
0064 mount system call is interpreted as the name of the device.
0065
0066 Mount options
0067 =============
0068
0069 fd=N
0070 The file descriptor to use for communication between the userspace
0071 filesystem and the kernel. The file descriptor must have been
0072 obtained by opening the FUSE device ('/dev/fuse').
0073
0074 rootmode=M
0075 The file mode of the filesystem's root in octal representation.
0076
0077 user_id=N
0078 The numeric user id of the mount owner.
0079
0080 group_id=N
0081 The numeric group id of the mount owner.
0082
0083 default_permissions
0084 By default FUSE doesn't check file access permissions, the
0085 filesystem is free to implement its access policy or leave it to
0086 the underlying file access mechanism (e.g. in case of network
0087 filesystems). This option enables permission checking, restricting
0088 access based on file mode. It is usually useful together with the
0089 'allow_other' mount option.
0090
0091 allow_other
0092 This option overrides the security measure restricting file access
0093 to the user mounting the filesystem. This option is by default only
0094 allowed to root, but this restriction can be removed with a
0095 (userspace) configuration option.
0096
0097 max_read=N
0098 With this option the maximum size of read operations can be set.
0099 The default is infinite. Note that the size of read requests is
0100 limited anyway to 32 pages (which is 128kbyte on i386).
0101
0102 blksize=N
0103 Set the block size for the filesystem. The default is 512. This
0104 option is only valid for 'fuseblk' type mounts.
0105
0106 Control filesystem
0107 ==================
0108
0109 There's a control filesystem for FUSE, which can be mounted by::
0110
0111 mount -t fusectl none /sys/fs/fuse/connections
0112
0113 Mounting it under the '/sys/fs/fuse/connections' directory makes it
0114 backwards compatible with earlier versions.
0115
0116 Under the fuse control filesystem each connection has a directory
0117 named by a unique number.
0118
0119 For each connection the following files exist within this directory:
0120
0121 waiting
0122 The number of requests which are waiting to be transferred to
0123 userspace or being processed by the filesystem daemon. If there is
0124 no filesystem activity and 'waiting' is non-zero, then the
0125 filesystem is hung or deadlocked.
0126
0127 abort
0128 Writing anything into this file will abort the filesystem
0129 connection. This means that all waiting requests will be aborted an
0130 error returned for all aborted and new requests.
0131
0132 Only the owner of the mount may read or write these files.
0133
0134 Interrupting filesystem operations
0135 ##################################
0136
0137 If a process issuing a FUSE filesystem request is interrupted, the
0138 following will happen:
0139
0140 - If the request is not yet sent to userspace AND the signal is
0141 fatal (SIGKILL or unhandled fatal signal), then the request is
0142 dequeued and returns immediately.
0143
0144 - If the request is not yet sent to userspace AND the signal is not
0145 fatal, then an interrupted flag is set for the request. When
0146 the request has been successfully transferred to userspace and
0147 this flag is set, an INTERRUPT request is queued.
0148
0149 - If the request is already sent to userspace, then an INTERRUPT
0150 request is queued.
0151
0152 INTERRUPT requests take precedence over other requests, so the
0153 userspace filesystem will receive queued INTERRUPTs before any others.
0154
0155 The userspace filesystem may ignore the INTERRUPT requests entirely,
0156 or may honor them by sending a reply to the *original* request, with
0157 the error set to EINTR.
0158
0159 It is also possible that there's a race between processing the
0160 original request and its INTERRUPT request. There are two possibilities:
0161
0162 1. The INTERRUPT request is processed before the original request is
0163 processed
0164
0165 2. The INTERRUPT request is processed after the original request has
0166 been answered
0167
0168 If the filesystem cannot find the original request, it should wait for
0169 some timeout and/or a number of new requests to arrive, after which it
0170 should reply to the INTERRUPT request with an EAGAIN error. In case
0171 1) the INTERRUPT request will be requeued. In case 2) the INTERRUPT
0172 reply will be ignored.
0173
0174 Aborting a filesystem connection
0175 ================================
0176
0177 It is possible to get into certain situations where the filesystem is
0178 not responding. Reasons for this may be:
0179
0180 a) Broken userspace filesystem implementation
0181
0182 b) Network connection down
0183
0184 c) Accidental deadlock
0185
0186 d) Malicious deadlock
0187
0188 (For more on c) and d) see later sections)
0189
0190 In either of these cases it may be useful to abort the connection to
0191 the filesystem. There are several ways to do this:
0192
0193 - Kill the filesystem daemon. Works in case of a) and b)
0194
0195 - Kill the filesystem daemon and all users of the filesystem. Works
0196 in all cases except some malicious deadlocks
0197
0198 - Use forced umount (umount -f). Works in all cases but only if
0199 filesystem is still attached (it hasn't been lazy unmounted)
0200
0201 - Abort filesystem through the FUSE control filesystem. Most
0202 powerful method, always works.
0203
0204 How do non-privileged mounts work?
0205 ==================================
0206
0207 Since the mount() system call is a privileged operation, a helper
0208 program (fusermount) is needed, which is installed setuid root.
0209
0210 The implication of providing non-privileged mounts is that the mount
0211 owner must not be able to use this capability to compromise the
0212 system. Obvious requirements arising from this are:
0213
0214 A) mount owner should not be able to get elevated privileges with the
0215 help of the mounted filesystem
0216
0217 B) mount owner should not get illegitimate access to information from
0218 other users' and the super user's processes
0219
0220 C) mount owner should not be able to induce undesired behavior in
0221 other users' or the super user's processes
0222
0223 How are requirements fulfilled?
0224 ===============================
0225
0226 A) The mount owner could gain elevated privileges by either:
0227
0228 1. creating a filesystem containing a device file, then opening this device
0229
0230 2. creating a filesystem containing a suid or sgid application, then executing this application
0231
0232 The solution is not to allow opening device files and ignore
0233 setuid and setgid bits when executing programs. To ensure this
0234 fusermount always adds "nosuid" and "nodev" to the mount options
0235 for non-privileged mounts.
0236
0237 B) If another user is accessing files or directories in the
0238 filesystem, the filesystem daemon serving requests can record the
0239 exact sequence and timing of operations performed. This
0240 information is otherwise inaccessible to the mount owner, so this
0241 counts as an information leak.
0242
0243 The solution to this problem will be presented in point 2) of C).
0244
0245 C) There are several ways in which the mount owner can induce
0246 undesired behavior in other users' processes, such as:
0247
0248 1) mounting a filesystem over a file or directory which the mount
0249 owner could otherwise not be able to modify (or could only
0250 make limited modifications).
0251
0252 This is solved in fusermount, by checking the access
0253 permissions on the mountpoint and only allowing the mount if
0254 the mount owner can do unlimited modification (has write
0255 access to the mountpoint, and mountpoint is not a "sticky"
0256 directory)
0257
0258 2) Even if 1) is solved the mount owner can change the behavior
0259 of other users' processes.
0260
0261 i) It can slow down or indefinitely delay the execution of a
0262 filesystem operation creating a DoS against the user or the
0263 whole system. For example a suid application locking a
0264 system file, and then accessing a file on the mount owner's
0265 filesystem could be stopped, and thus causing the system
0266 file to be locked forever.
0267
0268 ii) It can present files or directories of unlimited length, or
0269 directory structures of unlimited depth, possibly causing a
0270 system process to eat up diskspace, memory or other
0271 resources, again causing *DoS*.
0272
0273 The solution to this as well as B) is not to allow processes
0274 to access the filesystem, which could otherwise not be
0275 monitored or manipulated by the mount owner. Since if the
0276 mount owner can ptrace a process, it can do all of the above
0277 without using a FUSE mount, the same criteria as used in
0278 ptrace can be used to check if a process is allowed to access
0279 the filesystem or not.
0280
0281 Note that the *ptrace* check is not strictly necessary to
0282 prevent C/2/i, it is enough to check if mount owner has enough
0283 privilege to send signal to the process accessing the
0284 filesystem, since *SIGSTOP* can be used to get a similar effect.
0285
0286 I think these limitations are unacceptable?
0287 ===========================================
0288
0289 If a sysadmin trusts the users enough, or can ensure through other
0290 measures, that system processes will never enter non-privileged
0291 mounts, it can relax the last limitation in several ways:
0292
0293 - With the 'user_allow_other' config option. If this config option is
0294 set, the mounting user can add the 'allow_other' mount option which
0295 disables the check for other users' processes.
0296
0297 User namespaces have an unintuitive interaction with 'allow_other':
0298 an unprivileged user - normally restricted from mounting with
0299 'allow_other' - could do so in a user namespace where they're
0300 privileged. If any process could access such an 'allow_other' mount
0301 this would give the mounting user the ability to manipulate
0302 processes in user namespaces where they're unprivileged. For this
0303 reason 'allow_other' restricts access to users in the same userns
0304 or a descendant.
0305
0306 - With the 'allow_sys_admin_access' module option. If this option is
0307 set, super user's processes have unrestricted access to mounts
0308 irrespective of allow_other setting or user namespace of the
0309 mounting user.
0310
0311 Note that both of these relaxations expose the system to potential
0312 information leak or *DoS* as described in points B and C/2/i-ii in the
0313 preceding section.
0314
0315 Kernel - userspace interface
0316 ============================
0317
0318 The following diagram shows how a filesystem operation (in this
0319 example unlink) is performed in FUSE. ::
0320
0321
0322 | "rm /mnt/fuse/file" | FUSE filesystem daemon
0323 | |
0324 | | >sys_read()
0325 | | >fuse_dev_read()
0326 | | >request_wait()
0327 | | [sleep on fc->waitq]
0328 | |
0329 | >sys_unlink() |
0330 | >fuse_unlink() |
0331 | [get request from |
0332 | fc->unused_list] |
0333 | >request_send() |
0334 | [queue req on fc->pending] |
0335 | [wake up fc->waitq] | [woken up]
0336 | >request_wait_answer() |
0337 | [sleep on req->waitq] |
0338 | | <request_wait()
0339 | | [remove req from fc->pending]
0340 | | [copy req to read buffer]
0341 | | [add req to fc->processing]
0342 | | <fuse_dev_read()
0343 | | <sys_read()
0344 | |
0345 | | [perform unlink]
0346 | |
0347 | | >sys_write()
0348 | | >fuse_dev_write()
0349 | | [look up req in fc->processing]
0350 | | [remove from fc->processing]
0351 | | [copy write buffer to req]
0352 | [woken up] | [wake up req->waitq]
0353 | | <fuse_dev_write()
0354 | | <sys_write()
0355 | <request_wait_answer() |
0356 | <request_send() |
0357 | [add request to |
0358 | fc->unused_list] |
0359 | <fuse_unlink() |
0360 | <sys_unlink() |
0361
0362 .. note:: Everything in the description above is greatly simplified
0363
0364 There are a couple of ways in which to deadlock a FUSE filesystem.
0365 Since we are talking about unprivileged userspace programs,
0366 something must be done about these.
0367
0368 **Scenario 1 - Simple deadlock**::
0369
0370 | "rm /mnt/fuse/file" | FUSE filesystem daemon
0371 | |
0372 | >sys_unlink("/mnt/fuse/file") |
0373 | [acquire inode semaphore |
0374 | for "file"] |
0375 | >fuse_unlink() |
0376 | [sleep on req->waitq] |
0377 | | <sys_read()
0378 | | >sys_unlink("/mnt/fuse/file")
0379 | | [acquire inode semaphore
0380 | | for "file"]
0381 | | *DEADLOCK*
0382
0383 The solution for this is to allow the filesystem to be aborted.
0384
0385 **Scenario 2 - Tricky deadlock**
0386
0387
0388 This one needs a carefully crafted filesystem. It's a variation on
0389 the above, only the call back to the filesystem is not explicit,
0390 but is caused by a pagefault. ::
0391
0392 | Kamikaze filesystem thread 1 | Kamikaze filesystem thread 2
0393 | |
0394 | [fd = open("/mnt/fuse/file")] | [request served normally]
0395 | [mmap fd to 'addr'] |
0396 | [close fd] | [FLUSH triggers 'magic' flag]
0397 | [read a byte from addr] |
0398 | >do_page_fault() |
0399 | [find or create page] |
0400 | [lock page] |
0401 | >fuse_readpage() |
0402 | [queue READ request] |
0403 | [sleep on req->waitq] |
0404 | | [read request to buffer]
0405 | | [create reply header before addr]
0406 | | >sys_write(addr - headerlength)
0407 | | >fuse_dev_write()
0408 | | [look up req in fc->processing]
0409 | | [remove from fc->processing]
0410 | | [copy write buffer to req]
0411 | | >do_page_fault()
0412 | | [find or create page]
0413 | | [lock page]
0414 | | * DEADLOCK *
0415
0416 The solution is basically the same as above.
0417
0418 An additional problem is that while the write buffer is being copied
0419 to the request, the request must not be interrupted/aborted. This is
0420 because the destination address of the copy may not be valid after the
0421 request has returned.
0422
0423 This is solved with doing the copy atomically, and allowing abort
0424 while the page(s) belonging to the write buffer are faulted with
0425 get_user_pages(). The 'req->locked' flag indicates when the copy is
0426 taking place, and abort is delayed until this flag is unset.
