10 Userfaults allow the implementation of on-demand paging from userland
11 and more generally they allow userland to take control of various
12 memory page faults, something otherwise only the kernel code could do.
14 For example userfaults allows a proper and more optimal implementation
15 of the ``PROT_NONE+SIGSEGV`` trick.
20 Userfaults are delivered and resolved through the ``userfaultfd`` syscall.
22 The ``userfaultfd`` (aside from registering and unregistering virtual
23 memory ranges) provides two primary functionalities:
25 1) ``read/POLLIN`` protocol to notify a userland thread of the faults
28 2) various ``UFFDIO_*`` ioctls that can manage the virtual memory regions
29 registered in the ``userfaultfd`` that allows userland to efficiently
30 resolve the userfaults it receives via 1) or to manage the virtual
31 memory in the background
33 The real advantage of userfaults if compared to regular virtual memory
34 management of mremap/mprotect is that the userfaults in all their
35 operations never involve heavyweight structures like vmas (in fact the
36 ``userfaultfd`` runtime load never takes the mmap_lock for writing).
38 Vmas are not suitable for page- (or hugepage) granular fault tracking
39 when dealing with virtual address spaces that could span
40 Terabytes. Too many vmas would be needed for that.
42 The ``userfaultfd`` once opened by invoking the syscall, can also be
43 passed using unix domain sockets to a manager process, so the same
44 manager process could handle the userfaults of a multitude of
45 different processes without them being aware about what is going on
46 (well of course unless they later try to use the ``userfaultfd``
47 themselves on the same region the manager is already tracking, which
48 is a corner case that would currently return ``-EBUSY``).
53 When first opened the ``userfaultfd`` must be enabled invoking the
54 ``UFFDIO_API`` ioctl specifying a ``uffdio_api.api`` value set to ``UFFD_API`` (or
55 a later API version) which will specify the ``read/POLLIN`` protocol
56 userland intends to speak on the ``UFFD`` and the ``uffdio_api.features``
57 userland requires. The ``UFFDIO_API`` ioctl if successful (i.e. if the
58 requested ``uffdio_api.api`` is spoken also by the running kernel and the
59 requested features are going to be enabled) will return into
60 ``uffdio_api.features`` and ``uffdio_api.ioctls`` two 64bit bitmasks of
61 respectively all the available features of the read(2) protocol and
62 the generic ioctl available.
64 The ``uffdio_api.features`` bitmask returned by the ``UFFDIO_API`` ioctl
65 defines what memory types are supported by the ``userfaultfd`` and what
66 events, except page fault notifications, may be generated.
68 If the kernel supports registering ``userfaultfd`` ranges on hugetlbfs
69 virtual memory areas, ``UFFD_FEATURE_MISSING_HUGETLBFS`` will be set in
70 ``uffdio_api.features``. Similarly, ``UFFD_FEATURE_MISSING_SHMEM`` will be
71 set if the kernel supports registering ``userfaultfd`` ranges on shared
72 memory (covering all shmem APIs, i.e. tmpfs, ``IPCSHM``, ``/dev/zero``,
73 ``MAP_SHARED``, ``memfd_create``, etc).
75 The userland application that wants to use ``userfaultfd`` with hugetlbfs
76 or shared memory need to set the corresponding flag in
77 ``uffdio_api.features`` to enable those features.
79 If the userland desires to receive notifications for events other than
80 page faults, it has to verify that ``uffdio_api.features`` has appropriate
81 ``UFFD_FEATURE_EVENT_*`` bits set. These events are described in more
82 detail below in `Non-cooperative userfaultfd`_ section.
84 Once the ``userfaultfd`` has been enabled the ``UFFDIO_REGISTER`` ioctl should
85 be invoked (if present in the returned ``uffdio_api.ioctls`` bitmask) to
86 register a memory range in the ``userfaultfd`` by setting the
87 uffdio_register structure accordingly. The ``uffdio_register.mode``
88 bitmask will specify to the kernel which kind of faults to track for
89 the range (``UFFDIO_REGISTER_MODE_MISSING`` would track missing
90 pages). The ``UFFDIO_REGISTER`` ioctl will return the
91 ``uffdio_register.ioctls`` bitmask of ioctls that are suitable to resolve
92 userfaults on the range registered. Not all ioctls will necessarily be
93 supported for all memory types depending on the underlying virtual
94 memory backend (anonymous memory vs tmpfs vs real filebacked
97 Userland can use the ``uffdio_register.ioctls`` to manage the virtual
98 address space in the background (to add or potentially also remove
99 memory from the ``userfaultfd`` registered range). This means a userfault
100 could be triggering just before userland maps in the background the
103 The primary ioctl to resolve userfaults is ``UFFDIO_COPY``. That
104 atomically copies a page into the userfault registered range and wakes
105 up the blocked userfaults
106 (unless ``uffdio_copy.mode & UFFDIO_COPY_MODE_DONTWAKE`` is set).
107 Other ioctl works similarly to ``UFFDIO_COPY``. They're atomic as in
108 guaranteeing that nothing can see an half copied page since it'll
109 keep userfaulting until the copy has finished.
113 - If you requested ``UFFDIO_REGISTER_MODE_MISSING`` when registering then
114 you must provide some kind of page in your thread after reading from
115 the uffd. You must provide either ``UFFDIO_COPY`` or ``UFFDIO_ZEROPAGE``.
116 The normal behavior of the OS automatically providing a zero page on
117 an anonymous mmaping is not in place.
119 - None of the page-delivering ioctls default to the range that you
120 registered with. You must fill in all fields for the appropriate
121 ioctl struct including the range.
123 - You get the address of the access that triggered the missing page
124 event out of a struct uffd_msg that you read in the thread from the
125 uffd. You can supply as many pages as you want with ``UFFDIO_COPY`` or
126 ``UFFDIO_ZEROPAGE``. Keep in mind that unless you used DONTWAKE then
127 the first of any of those IOCTLs wakes up the faulting thread.
129 - Be sure to test for all errors including
130 (``pollfd[0].revents & POLLERR``). This can happen, e.g. when ranges
131 supplied were incorrect.
133 Write Protect Notifications
134 ---------------------------
136 This is equivalent to (but faster than) using mprotect and a SIGSEGV
139 Firstly you need to register a range with ``UFFDIO_REGISTER_MODE_WP``.
140 Instead of using mprotect(2) you use
141 ``ioctl(uffd, UFFDIO_WRITEPROTECT, struct *uffdio_writeprotect)``
142 while ``mode = UFFDIO_WRITEPROTECT_MODE_WP``
143 in the struct passed in. The range does not default to and does not
144 have to be identical to the range you registered with. You can write
145 protect as many ranges as you like (inside the registered range).
146 Then, in the thread reading from uffd the struct will have
147 ``msg.arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WP`` set. Now you send
148 ``ioctl(uffd, UFFDIO_WRITEPROTECT, struct *uffdio_writeprotect)``
149 again while ``pagefault.mode`` does not have ``UFFDIO_WRITEPROTECT_MODE_WP``
150 set. This wakes up the thread which will continue to run with writes. This
151 allows you to do the bookkeeping about the write in the uffd reading
152 thread before the ioctl.
154 If you registered with both ``UFFDIO_REGISTER_MODE_MISSING`` and
155 ``UFFDIO_REGISTER_MODE_WP`` then you need to think about the sequence in
156 which you supply a page and undo write protect. Note that there is a
157 difference between writes into a WP area and into a !WP area. The
158 former will have ``UFFD_PAGEFAULT_FLAG_WP`` set, the latter
159 ``UFFD_PAGEFAULT_FLAG_WRITE``. The latter did not fail on protection but
160 you still need to supply a page when ``UFFDIO_REGISTER_MODE_MISSING`` was
166 QEMU/KVM is using the ``userfaultfd`` syscall to implement postcopy live
167 migration. Postcopy live migration is one form of memory
168 externalization consisting of a virtual machine running with part or
169 all of its memory residing on a different node in the cloud. The
170 ``userfaultfd`` abstraction is generic enough that not a single line of
171 KVM kernel code had to be modified in order to add postcopy live
174 Guest async page faults, ``FOLL_NOWAIT`` and all other ``GUP*`` features work
175 just fine in combination with userfaults. Userfaults trigger async
176 page faults in the guest scheduler so those guest processes that
177 aren't waiting for userfaults (i.e. network bound) can keep running in
180 It is generally beneficial to run one pass of precopy live migration
181 just before starting postcopy live migration, in order to avoid
182 generating userfaults for readonly guest regions.
184 The implementation of postcopy live migration currently uses one
185 single bidirectional socket but in the future two different sockets
186 will be used (to reduce the latency of the userfaults to the minimum
187 possible without having to decrease ``/proc/sys/net/ipv4/tcp_wmem``).
189 The QEMU in the source node writes all pages that it knows are missing
190 in the destination node, into the socket, and the migration thread of
191 the QEMU running in the destination node runs ``UFFDIO_COPY|ZEROPAGE``
192 ioctls on the ``userfaultfd`` in order to map the received pages into the
193 guest (``UFFDIO_ZEROCOPY`` is used if the source page was a zero page).
195 A different postcopy thread in the destination node listens with
196 poll() to the ``userfaultfd`` in parallel. When a ``POLLIN`` event is
197 generated after a userfault triggers, the postcopy thread read() from
198 the ``userfaultfd`` and receives the fault address (or ``-EAGAIN`` in case the
199 userfault was already resolved and waken by a ``UFFDIO_COPY|ZEROPAGE`` run
200 by the parallel QEMU migration thread).
202 After the QEMU postcopy thread (running in the destination node) gets
203 the userfault address it writes the information about the missing page
204 into the socket. The QEMU source node receives the information and
205 roughly "seeks" to that page address and continues sending all
206 remaining missing pages from that new page offset. Soon after that
207 (just the time to flush the tcp_wmem queue through the network) the
208 migration thread in the QEMU running in the destination node will
209 receive the page that triggered the userfault and it'll map it as
210 usual with the ``UFFDIO_COPY|ZEROPAGE`` (without actually knowing if it
211 was spontaneously sent by the source or if it was an urgent page
212 requested through a userfault).
214 By the time the userfaults start, the QEMU in the destination node
215 doesn't need to keep any per-page state bitmap relative to the live
216 migration around and a single per-page bitmap has to be maintained in
217 the QEMU running in the source node to know which pages are still
218 missing in the destination node. The bitmap in the source node is
219 checked to find which missing pages to send in round robin and we seek
220 over it when receiving incoming userfaults. After sending each page of
221 course the bitmap is updated accordingly. It's also useful to avoid
222 sending the same page twice (in case the userfault is read by the
223 postcopy thread just before ``UFFDIO_COPY|ZEROPAGE`` runs in the migration
226 Non-cooperative userfaultfd
227 ===========================
229 When the ``userfaultfd`` is monitored by an external manager, the manager
230 must be able to track changes in the process virtual memory
231 layout. Userfaultfd can notify the manager about such changes using
232 the same read(2) protocol as for the page fault notifications. The
233 manager has to explicitly enable these events by setting appropriate
234 bits in ``uffdio_api.features`` passed to ``UFFDIO_API`` ioctl:
236 ``UFFD_FEATURE_EVENT_FORK``
237 enable ``userfaultfd`` hooks for fork(). When this feature is
238 enabled, the ``userfaultfd`` context of the parent process is
239 duplicated into the newly created process. The manager
240 receives ``UFFD_EVENT_FORK`` with file descriptor of the new
241 ``userfaultfd`` context in the ``uffd_msg.fork``.
243 ``UFFD_FEATURE_EVENT_REMAP``
244 enable notifications about mremap() calls. When the
245 non-cooperative process moves a virtual memory area to a
246 different location, the manager will receive
247 ``UFFD_EVENT_REMAP``. The ``uffd_msg.remap`` will contain the old and
248 new addresses of the area and its original length.
250 ``UFFD_FEATURE_EVENT_REMOVE``
251 enable notifications about madvise(MADV_REMOVE) and
252 madvise(MADV_DONTNEED) calls. The event ``UFFD_EVENT_REMOVE`` will
253 be generated upon these calls to madvise(). The ``uffd_msg.remove``
254 will contain start and end addresses of the removed area.
256 ``UFFD_FEATURE_EVENT_UNMAP``
257 enable notifications about memory unmapping. The manager will
258 get ``UFFD_EVENT_UNMAP`` with ``uffd_msg.remove`` containing start and
259 end addresses of the unmapped area.
261 Although the ``UFFD_FEATURE_EVENT_REMOVE`` and ``UFFD_FEATURE_EVENT_UNMAP``
262 are pretty similar, they quite differ in the action expected from the
263 ``userfaultfd`` manager. In the former case, the virtual memory is
264 removed, but the area is not, the area remains monitored by the
265 ``userfaultfd``, and if a page fault occurs in that area it will be
266 delivered to the manager. The proper resolution for such page fault is
267 to zeromap the faulting address. However, in the latter case, when an
268 area is unmapped, either explicitly (with munmap() system call), or
269 implicitly (e.g. during mremap()), the area is removed and in turn the
270 ``userfaultfd`` context for such area disappears too and the manager will
271 not get further userland page faults from the removed area. Still, the
272 notification is required in order to prevent manager from using
273 ``UFFDIO_COPY`` on the unmapped area.
275 Unlike userland page faults which have to be synchronous and require
276 explicit or implicit wakeup, all the events are delivered
277 asynchronously and the non-cooperative process resumes execution as
278 soon as manager executes read(). The ``userfaultfd`` manager should
279 carefully synchronize calls to ``UFFDIO_COPY`` with the events
280 processing. To aid the synchronization, the ``UFFDIO_COPY`` ioctl will
281 return ``-ENOSPC`` when the monitored process exits at the time of
282 ``UFFDIO_COPY``, and ``-ENOENT``, when the non-cooperative process has changed
283 its virtual memory layout simultaneously with outstanding ``UFFDIO_COPY``
286 The current asynchronous model of the event delivery is optimal for
287 single threaded non-cooperative ``userfaultfd`` manager implementations. A
288 synchronous event delivery model can be added later as a new
289 ``userfaultfd`` feature to facilitate multithreading enhancements of the
290 non cooperative manager, for example to allow ``UFFDIO_COPY`` ioctls to
291 run in parallel to the event reception. Single threaded
292 implementations should continue to use the current async event
293 delivery model instead.