1 =========== 1 ===========
2 Userfaultfd 2 Userfaultfd
3 =========== 3 ===========
4 4
5 Objective 5 Objective
6 ========= 6 =========
7 7
8 Userfaults allow the implementation of on-dema 8 Userfaults allow the implementation of on-demand paging from userland
9 and more generally they allow userland to take 9 and more generally they allow userland to take control of various
10 memory page faults, something otherwise only t 10 memory page faults, something otherwise only the kernel code could do.
11 11
12 For example userfaults allows a proper and mor 12 For example userfaults allows a proper and more optimal implementation
13 of the ``PROT_NONE+SIGSEGV`` trick. 13 of the ``PROT_NONE+SIGSEGV`` trick.
14 14
15 Design 15 Design
16 ====== 16 ======
17 17
18 Userspace creates a new userfaultfd, initializ 18 Userspace creates a new userfaultfd, initializes it, and registers one or more
19 regions of virtual memory with it. Then, any p 19 regions of virtual memory with it. Then, any page faults which occur within the
20 region(s) result in a message being delivered 20 region(s) result in a message being delivered to the userfaultfd, notifying
21 userspace of the fault. 21 userspace of the fault.
22 22
23 The ``userfaultfd`` (aside from registering an 23 The ``userfaultfd`` (aside from registering and unregistering virtual
24 memory ranges) provides two primary functional 24 memory ranges) provides two primary functionalities:
25 25
26 1) ``read/POLLIN`` protocol to notify a userla 26 1) ``read/POLLIN`` protocol to notify a userland thread of the faults
27 happening 27 happening
28 28
29 2) various ``UFFDIO_*`` ioctls that can manage 29 2) various ``UFFDIO_*`` ioctls that can manage the virtual memory regions
30 registered in the ``userfaultfd`` that allo 30 registered in the ``userfaultfd`` that allows userland to efficiently
31 resolve the userfaults it receives via 1) o 31 resolve the userfaults it receives via 1) or to manage the virtual
32 memory in the background 32 memory in the background
33 33
34 The real advantage of userfaults if compared t 34 The real advantage of userfaults if compared to regular virtual memory
35 management of mremap/mprotect is that the user 35 management of mremap/mprotect is that the userfaults in all their
36 operations never involve heavyweight structure 36 operations never involve heavyweight structures like vmas (in fact the
37 ``userfaultfd`` runtime load never takes the m 37 ``userfaultfd`` runtime load never takes the mmap_lock for writing).
38 Vmas are not suitable for page- (or hugepage) 38 Vmas are not suitable for page- (or hugepage) granular fault tracking
39 when dealing with virtual address spaces that 39 when dealing with virtual address spaces that could span
40 Terabytes. Too many vmas would be needed for t 40 Terabytes. Too many vmas would be needed for that.
41 41
42 The ``userfaultfd``, once created, can also be 42 The ``userfaultfd``, once created, can also be
43 passed using unix domain sockets to a manager 43 passed using unix domain sockets to a manager process, so the same
44 manager process could handle the userfaults of 44 manager process could handle the userfaults of a multitude of
45 different processes without them being aware a 45 different processes without them being aware about what is going on
46 (well of course unless they later try to use t 46 (well of course unless they later try to use the ``userfaultfd``
47 themselves on the same region the manager is a 47 themselves on the same region the manager is already tracking, which
48 is a corner case that would currently return ` 48 is a corner case that would currently return ``-EBUSY``).
49 49
50 API 50 API
51 === 51 ===
52 52
53 Creating a userfaultfd 53 Creating a userfaultfd
54 ---------------------- 54 ----------------------
55 55
56 There are two ways to create a new userfaultfd 56 There are two ways to create a new userfaultfd, each of which provide ways to
57 restrict access to this functionality (since h 57 restrict access to this functionality (since historically userfaultfds which
58 handle kernel page faults have been a useful t 58 handle kernel page faults have been a useful tool for exploiting the kernel).
59 59
60 The first way, supported since userfaultfd was 60 The first way, supported since userfaultfd was introduced, is the
61 userfaultfd(2) syscall. Access to this is cont 61 userfaultfd(2) syscall. Access to this is controlled in several ways:
62 62
63 - Any user can always create a userfaultfd whi 63 - Any user can always create a userfaultfd which traps userspace page faults
64 only. Such a userfaultfd can be created usin 64 only. Such a userfaultfd can be created using the userfaultfd(2) syscall
65 with the flag UFFD_USER_MODE_ONLY. 65 with the flag UFFD_USER_MODE_ONLY.
66 66
67 - In order to also trap kernel page faults for 67 - In order to also trap kernel page faults for the address space, either the
68 process needs the CAP_SYS_PTRACE capability, 68 process needs the CAP_SYS_PTRACE capability, or the system must have
69 vm.unprivileged_userfaultfd set to 1. By def 69 vm.unprivileged_userfaultfd set to 1. By default, vm.unprivileged_userfaultfd
70 is set to 0. 70 is set to 0.
71 71
72 The second way, added to the kernel more recen 72 The second way, added to the kernel more recently, is by opening
73 /dev/userfaultfd and issuing a USERFAULTFD_IOC 73 /dev/userfaultfd and issuing a USERFAULTFD_IOC_NEW ioctl to it. This method
74 yields equivalent userfaultfds to the userfaul 74 yields equivalent userfaultfds to the userfaultfd(2) syscall.
75 75
76 Unlike userfaultfd(2), access to /dev/userfaul 76 Unlike userfaultfd(2), access to /dev/userfaultfd is controlled via normal
77 filesystem permissions (user/group/mode), whic 77 filesystem permissions (user/group/mode), which gives fine grained access to
78 userfaultfd specifically, without also grantin 78 userfaultfd specifically, without also granting other unrelated privileges at
79 the same time (as e.g. granting CAP_SYS_PTRACE 79 the same time (as e.g. granting CAP_SYS_PTRACE would do). Users who have access
80 to /dev/userfaultfd can always create userfaul 80 to /dev/userfaultfd can always create userfaultfds that trap kernel page faults;
81 vm.unprivileged_userfaultfd is not considered. 81 vm.unprivileged_userfaultfd is not considered.
82 82
83 Initializing a userfaultfd 83 Initializing a userfaultfd
84 -------------------------- 84 --------------------------
85 85
86 When first opened the ``userfaultfd`` must be 86 When first opened the ``userfaultfd`` must be enabled invoking the
87 ``UFFDIO_API`` ioctl specifying a ``uffdio_api 87 ``UFFDIO_API`` ioctl specifying a ``uffdio_api.api`` value set to ``UFFD_API`` (or
88 a later API version) which will specify the `` 88 a later API version) which will specify the ``read/POLLIN`` protocol
89 userland intends to speak on the ``UFFD`` and 89 userland intends to speak on the ``UFFD`` and the ``uffdio_api.features``
90 userland requires. The ``UFFDIO_API`` ioctl if 90 userland requires. The ``UFFDIO_API`` ioctl if successful (i.e. if the
91 requested ``uffdio_api.api`` is spoken also by 91 requested ``uffdio_api.api`` is spoken also by the running kernel and the
92 requested features are going to be enabled) wi 92 requested features are going to be enabled) will return into
93 ``uffdio_api.features`` and ``uffdio_api.ioctl 93 ``uffdio_api.features`` and ``uffdio_api.ioctls`` two 64bit bitmasks of
94 respectively all the available features of the 94 respectively all the available features of the read(2) protocol and
95 the generic ioctl available. 95 the generic ioctl available.
96 96
97 The ``uffdio_api.features`` bitmask returned b 97 The ``uffdio_api.features`` bitmask returned by the ``UFFDIO_API`` ioctl
98 defines what memory types are supported by the 98 defines what memory types are supported by the ``userfaultfd`` and what
99 events, except page fault notifications, may b 99 events, except page fault notifications, may be generated:
100 100
101 - The ``UFFD_FEATURE_EVENT_*`` flags indicate 101 - The ``UFFD_FEATURE_EVENT_*`` flags indicate that various other events
102 other than page faults are supported. These 102 other than page faults are supported. These events are described in more
103 detail below in the `Non-cooperative userfau 103 detail below in the `Non-cooperative userfaultfd`_ section.
104 104
105 - ``UFFD_FEATURE_MISSING_HUGETLBFS`` and ``UFF 105 - ``UFFD_FEATURE_MISSING_HUGETLBFS`` and ``UFFD_FEATURE_MISSING_SHMEM``
106 indicate that the kernel supports ``UFFDIO_R 106 indicate that the kernel supports ``UFFDIO_REGISTER_MODE_MISSING``
107 registrations for hugetlbfs and shared memor 107 registrations for hugetlbfs and shared memory (covering all shmem APIs,
108 i.e. tmpfs, ``IPCSHM``, ``/dev/zero``, ``MAP 108 i.e. tmpfs, ``IPCSHM``, ``/dev/zero``, ``MAP_SHARED``, ``memfd_create``,
109 etc) virtual memory areas, respectively. 109 etc) virtual memory areas, respectively.
110 110
111 - ``UFFD_FEATURE_MINOR_HUGETLBFS`` indicates t 111 - ``UFFD_FEATURE_MINOR_HUGETLBFS`` indicates that the kernel supports
112 ``UFFDIO_REGISTER_MODE_MINOR`` registration 112 ``UFFDIO_REGISTER_MODE_MINOR`` registration for hugetlbfs virtual memory
113 areas. ``UFFD_FEATURE_MINOR_SHMEM`` is the a 113 areas. ``UFFD_FEATURE_MINOR_SHMEM`` is the analogous feature indicating
114 support for shmem virtual memory areas. 114 support for shmem virtual memory areas.
115 115
116 - ``UFFD_FEATURE_MOVE`` indicates that the ker <<
117 existing page contents from userspace. <<
118 <<
119 The userland application should set the featur 116 The userland application should set the feature flags it intends to use
120 when invoking the ``UFFDIO_API`` ioctl, to req 117 when invoking the ``UFFDIO_API`` ioctl, to request that those features be
121 enabled if supported. 118 enabled if supported.
122 119
123 Once the ``userfaultfd`` API has been enabled 120 Once the ``userfaultfd`` API has been enabled the ``UFFDIO_REGISTER``
124 ioctl should be invoked (if present in the ret 121 ioctl should be invoked (if present in the returned ``uffdio_api.ioctls``
125 bitmask) to register a memory range in the ``u 122 bitmask) to register a memory range in the ``userfaultfd`` by setting the
126 uffdio_register structure accordingly. The ``u 123 uffdio_register structure accordingly. The ``uffdio_register.mode``
127 bitmask will specify to the kernel which kind 124 bitmask will specify to the kernel which kind of faults to track for
128 the range. The ``UFFDIO_REGISTER`` ioctl will 125 the range. The ``UFFDIO_REGISTER`` ioctl will return the
129 ``uffdio_register.ioctls`` bitmask of ioctls t 126 ``uffdio_register.ioctls`` bitmask of ioctls that are suitable to resolve
130 userfaults on the range registered. Not all io 127 userfaults on the range registered. Not all ioctls will necessarily be
131 supported for all memory types (e.g. anonymous 128 supported for all memory types (e.g. anonymous memory vs. shmem vs.
132 hugetlbfs), or all types of intercepted faults 129 hugetlbfs), or all types of intercepted faults.
133 130
134 Userland can use the ``uffdio_register.ioctls` 131 Userland can use the ``uffdio_register.ioctls`` to manage the virtual
135 address space in the background (to add or pot 132 address space in the background (to add or potentially also remove
136 memory from the ``userfaultfd`` registered ran 133 memory from the ``userfaultfd`` registered range). This means a userfault
137 could be triggering just before userland maps 134 could be triggering just before userland maps in the background the
138 user-faulted page. 135 user-faulted page.
139 136
140 Resolving Userfaults 137 Resolving Userfaults
141 -------------------- 138 --------------------
142 139
143 There are three basic ways to resolve userfaul 140 There are three basic ways to resolve userfaults:
144 141
145 - ``UFFDIO_COPY`` atomically copies some exist 142 - ``UFFDIO_COPY`` atomically copies some existing page contents from
146 userspace. 143 userspace.
147 144
148 - ``UFFDIO_ZEROPAGE`` atomically zeros the new 145 - ``UFFDIO_ZEROPAGE`` atomically zeros the new page.
149 146
150 - ``UFFDIO_CONTINUE`` maps an existing, previo 147 - ``UFFDIO_CONTINUE`` maps an existing, previously-populated page.
151 148
152 These operations are atomic in the sense that 149 These operations are atomic in the sense that they guarantee nothing can
153 see a half-populated page, since readers will 150 see a half-populated page, since readers will keep userfaulting until the
154 operation has finished. 151 operation has finished.
155 152
156 By default, these wake up userfaults blocked o 153 By default, these wake up userfaults blocked on the range in question.
157 They support a ``UFFDIO_*_MODE_DONTWAKE`` ``mo 154 They support a ``UFFDIO_*_MODE_DONTWAKE`` ``mode`` flag, which indicates
158 that waking will be done separately at some la 155 that waking will be done separately at some later time.
159 156
160 Which ioctl to choose depends on the kind of p 157 Which ioctl to choose depends on the kind of page fault, and what we'd
161 like to do to resolve it: 158 like to do to resolve it:
162 159
163 - For ``UFFDIO_REGISTER_MODE_MISSING`` faults, 160 - For ``UFFDIO_REGISTER_MODE_MISSING`` faults, the fault needs to be
164 resolved by either providing a new page (``U 161 resolved by either providing a new page (``UFFDIO_COPY``), or mapping
165 the zero page (``UFFDIO_ZEROPAGE``). By defa 162 the zero page (``UFFDIO_ZEROPAGE``). By default, the kernel would map
166 the zero page for a missing fault. With user 163 the zero page for a missing fault. With userfaultfd, userspace can
167 decide what content to provide before the fa 164 decide what content to provide before the faulting thread continues.
168 165
169 - For ``UFFDIO_REGISTER_MODE_MINOR`` faults, t 166 - For ``UFFDIO_REGISTER_MODE_MINOR`` faults, there is an existing page (in
170 the page cache). Userspace has the option of 167 the page cache). Userspace has the option of modifying the page's
171 contents before resolving the fault. Once th 168 contents before resolving the fault. Once the contents are correct
172 (modified or not), userspace asks the kernel 169 (modified or not), userspace asks the kernel to map the page and let the
173 faulting thread continue with ``UFFDIO_CONTI 170 faulting thread continue with ``UFFDIO_CONTINUE``.
174 171
175 Notes: 172 Notes:
176 173
177 - You can tell which kind of fault occurred by 174 - You can tell which kind of fault occurred by examining
178 ``pagefault.flags`` within the ``uffd_msg``, 175 ``pagefault.flags`` within the ``uffd_msg``, checking for the
179 ``UFFD_PAGEFAULT_FLAG_*`` flags. 176 ``UFFD_PAGEFAULT_FLAG_*`` flags.
180 177
181 - None of the page-delivering ioctls default t 178 - None of the page-delivering ioctls default to the range that you
182 registered with. You must fill in all field 179 registered with. You must fill in all fields for the appropriate
183 ioctl struct including the range. 180 ioctl struct including the range.
184 181
185 - You get the address of the access that trigg 182 - You get the address of the access that triggered the missing page
186 event out of a struct uffd_msg that you read 183 event out of a struct uffd_msg that you read in the thread from the
187 uffd. You can supply as many pages as you w 184 uffd. You can supply as many pages as you want with these IOCTLs.
188 Keep in mind that unless you used DONTWAKE t 185 Keep in mind that unless you used DONTWAKE then the first of any of
189 those IOCTLs wakes up the faulting thread. 186 those IOCTLs wakes up the faulting thread.
190 187
191 - Be sure to test for all errors including 188 - Be sure to test for all errors including
192 (``pollfd[0].revents & POLLERR``). This can 189 (``pollfd[0].revents & POLLERR``). This can happen, e.g. when ranges
193 supplied were incorrect. 190 supplied were incorrect.
194 191
195 Write Protect Notifications 192 Write Protect Notifications
196 --------------------------- 193 ---------------------------
197 194
198 This is equivalent to (but faster than) using 195 This is equivalent to (but faster than) using mprotect and a SIGSEGV
199 signal handler. 196 signal handler.
200 197
201 Firstly you need to register a range with ``UF 198 Firstly you need to register a range with ``UFFDIO_REGISTER_MODE_WP``.
202 Instead of using mprotect(2) you use 199 Instead of using mprotect(2) you use
203 ``ioctl(uffd, UFFDIO_WRITEPROTECT, struct *uff 200 ``ioctl(uffd, UFFDIO_WRITEPROTECT, struct *uffdio_writeprotect)``
204 while ``mode = UFFDIO_WRITEPROTECT_MODE_WP`` 201 while ``mode = UFFDIO_WRITEPROTECT_MODE_WP``
205 in the struct passed in. The range does not d 202 in the struct passed in. The range does not default to and does not
206 have to be identical to the range you register 203 have to be identical to the range you registered with. You can write
207 protect as many ranges as you like (inside the 204 protect as many ranges as you like (inside the registered range).
208 Then, in the thread reading from uffd the stru 205 Then, in the thread reading from uffd the struct will have
209 ``msg.arg.pagefault.flags & UFFD_PAGEFAULT_FLA 206 ``msg.arg.pagefault.flags & UFFD_PAGEFAULT_FLAG_WP`` set. Now you send
210 ``ioctl(uffd, UFFDIO_WRITEPROTECT, struct *uff 207 ``ioctl(uffd, UFFDIO_WRITEPROTECT, struct *uffdio_writeprotect)``
211 again while ``pagefault.mode`` does not have ` 208 again while ``pagefault.mode`` does not have ``UFFDIO_WRITEPROTECT_MODE_WP``
212 set. This wakes up the thread which will conti 209 set. This wakes up the thread which will continue to run with writes. This
213 allows you to do the bookkeeping about the wri 210 allows you to do the bookkeeping about the write in the uffd reading
214 thread before the ioctl. 211 thread before the ioctl.
215 212
216 If you registered with both ``UFFDIO_REGISTER_ 213 If you registered with both ``UFFDIO_REGISTER_MODE_MISSING`` and
217 ``UFFDIO_REGISTER_MODE_WP`` then you need to t 214 ``UFFDIO_REGISTER_MODE_WP`` then you need to think about the sequence in
218 which you supply a page and undo write protect 215 which you supply a page and undo write protect. Note that there is a
219 difference between writes into a WP area and i 216 difference between writes into a WP area and into a !WP area. The
220 former will have ``UFFD_PAGEFAULT_FLAG_WP`` se 217 former will have ``UFFD_PAGEFAULT_FLAG_WP`` set, the latter
221 ``UFFD_PAGEFAULT_FLAG_WRITE``. The latter did 218 ``UFFD_PAGEFAULT_FLAG_WRITE``. The latter did not fail on protection but
222 you still need to supply a page when ``UFFDIO_ 219 you still need to supply a page when ``UFFDIO_REGISTER_MODE_MISSING`` was
223 used. 220 used.
224 221
225 Userfaultfd write-protect mode currently behav 222 Userfaultfd write-protect mode currently behave differently on none ptes
226 (when e.g. page is missing) over different typ 223 (when e.g. page is missing) over different types of memories.
227 224
228 For anonymous memory, ``ioctl(UFFDIO_WRITEPROT 225 For anonymous memory, ``ioctl(UFFDIO_WRITEPROTECT)`` will ignore none ptes
229 (e.g. when pages are missing and not populated 226 (e.g. when pages are missing and not populated). For file-backed memories
230 like shmem and hugetlbfs, none ptes will be wr 227 like shmem and hugetlbfs, none ptes will be write protected just like a
231 present pte. In other words, there will be a 228 present pte. In other words, there will be a userfaultfd write fault
232 message generated when writing to a missing pa 229 message generated when writing to a missing page on file typed memories,
233 as long as the page range was write-protected 230 as long as the page range was write-protected before. Such a message will
234 not be generated on anonymous memories by defa 231 not be generated on anonymous memories by default.
235 232
236 If the application wants to be able to write p 233 If the application wants to be able to write protect none ptes on anonymous
237 memory, one can pre-populate the memory with e 234 memory, one can pre-populate the memory with e.g. MADV_POPULATE_READ. On
238 newer kernels, one can also detect the feature 235 newer kernels, one can also detect the feature UFFD_FEATURE_WP_UNPOPULATED
239 and set the feature bit in advance to make sur 236 and set the feature bit in advance to make sure none ptes will also be
240 write protected even upon anonymous memory. 237 write protected even upon anonymous memory.
241 238
242 When using ``UFFDIO_REGISTER_MODE_WP`` in comb 239 When using ``UFFDIO_REGISTER_MODE_WP`` in combination with either
243 ``UFFDIO_REGISTER_MODE_MISSING`` or ``UFFDIO_R 240 ``UFFDIO_REGISTER_MODE_MISSING`` or ``UFFDIO_REGISTER_MODE_MINOR``, when
244 resolving missing / minor faults with ``UFFDIO 241 resolving missing / minor faults with ``UFFDIO_COPY`` or ``UFFDIO_CONTINUE``
245 respectively, it may be desirable for the new 242 respectively, it may be desirable for the new page / mapping to be
246 write-protected (so future writes will also re 243 write-protected (so future writes will also result in a WP fault). These ioctls
247 support a mode flag (``UFFDIO_COPY_MODE_WP`` o 244 support a mode flag (``UFFDIO_COPY_MODE_WP`` or ``UFFDIO_CONTINUE_MODE_WP``
248 respectively) to configure the mapping this wa 245 respectively) to configure the mapping this way.
249 246
250 If the userfaultfd context has ``UFFD_FEATURE_ 247 If the userfaultfd context has ``UFFD_FEATURE_WP_ASYNC`` feature bit set,
251 any vma registered with write-protection will 248 any vma registered with write-protection will work in async mode rather
252 than the default sync mode. 249 than the default sync mode.
253 250
254 In async mode, there will be no message genera 251 In async mode, there will be no message generated when a write operation
255 happens, meanwhile the write-protection will b 252 happens, meanwhile the write-protection will be resolved automatically by
256 the kernel. It can be seen as a more accurate 253 the kernel. It can be seen as a more accurate version of soft-dirty
257 tracking and it can be different in a few ways 254 tracking and it can be different in a few ways:
258 255
259 - The dirty result will not be affected by v 256 - The dirty result will not be affected by vma changes (e.g. vma
260 merging) because the dirty is only tracked 257 merging) because the dirty is only tracked by the pte.
261 258
262 - It supports range operations by default, s 259 - It supports range operations by default, so one can enable tracking on
263 any range of memory as long as page aligne 260 any range of memory as long as page aligned.
264 261
265 - Dirty information will not get lost if the 262 - Dirty information will not get lost if the pte was zapped due to
266 various reasons (e.g. during split of a sh 263 various reasons (e.g. during split of a shmem transparent huge page).
267 264
268 - Due to a reverted meaning of soft-dirty (p 265 - Due to a reverted meaning of soft-dirty (page clean when uffd-wp bit
269 set; dirty when uffd-wp bit cleared), it h 266 set; dirty when uffd-wp bit cleared), it has different semantics on
270 some of the memory operations. For exampl 267 some of the memory operations. For example: ``MADV_DONTNEED`` on
271 anonymous (or ``MADV_REMOVE`` on a file ma 268 anonymous (or ``MADV_REMOVE`` on a file mapping) will be treated as
272 dirtying of memory by dropping uffd-wp bit 269 dirtying of memory by dropping uffd-wp bit during the procedure.
273 270
274 The user app can collect the "written/dirty" s 271 The user app can collect the "written/dirty" status by looking up the
275 uffd-wp bit for the pages being interested in 272 uffd-wp bit for the pages being interested in /proc/pagemap.
276 273
277 The page will not be under track of uffd-wp as 274 The page will not be under track of uffd-wp async mode until the page is
278 explicitly write-protected by ``ioctl(UFFDIO_W 275 explicitly write-protected by ``ioctl(UFFDIO_WRITEPROTECT)`` with the mode
279 flag ``UFFDIO_WRITEPROTECT_MODE_WP`` set. Try 276 flag ``UFFDIO_WRITEPROTECT_MODE_WP`` set. Trying to resolve a page fault
280 that was tracked by async mode userfaultfd-wp 277 that was tracked by async mode userfaultfd-wp is invalid.
281 278
282 When userfaultfd-wp async mode is used alone, 279 When userfaultfd-wp async mode is used alone, it can be applied to all
283 kinds of memory. 280 kinds of memory.
284 281
285 Memory Poisioning Emulation 282 Memory Poisioning Emulation
286 --------------------------- 283 ---------------------------
287 284
288 In response to a fault (either missing or mino 285 In response to a fault (either missing or minor), an action userspace can
289 take to "resolve" it is to issue a ``UFFDIO_PO 286 take to "resolve" it is to issue a ``UFFDIO_POISON``. This will cause any
290 future faulters to either get a SIGBUS, or in 287 future faulters to either get a SIGBUS, or in KVM's case the guest will
291 receive an MCE as if there were hardware memor 288 receive an MCE as if there were hardware memory poisoning.
292 289
293 This is used to emulate hardware memory poison 290 This is used to emulate hardware memory poisoning. Imagine a VM running on a
294 machine which experiences a real hardware memo 291 machine which experiences a real hardware memory error. Later, we live migrate
295 the VM to another physical machine. Since we w 292 the VM to another physical machine. Since we want the migration to be
296 transparent to the guest, we want that same ad 293 transparent to the guest, we want that same address range to act as if it was
297 still poisoned, even though it's on a new phys 294 still poisoned, even though it's on a new physical host which ostensibly
298 doesn't have a memory error in the exact same 295 doesn't have a memory error in the exact same spot.
299 296
300 QEMU/KVM 297 QEMU/KVM
301 ======== 298 ========
302 299
303 QEMU/KVM is using the ``userfaultfd`` syscall 300 QEMU/KVM is using the ``userfaultfd`` syscall to implement postcopy live
304 migration. Postcopy live migration is one form 301 migration. Postcopy live migration is one form of memory
305 externalization consisting of a virtual machin 302 externalization consisting of a virtual machine running with part or
306 all of its memory residing on a different node 303 all of its memory residing on a different node in the cloud. The
307 ``userfaultfd`` abstraction is generic enough 304 ``userfaultfd`` abstraction is generic enough that not a single line of
308 KVM kernel code had to be modified in order to 305 KVM kernel code had to be modified in order to add postcopy live
309 migration to QEMU. 306 migration to QEMU.
310 307
311 Guest async page faults, ``FOLL_NOWAIT`` and a 308 Guest async page faults, ``FOLL_NOWAIT`` and all other ``GUP*`` features work
312 just fine in combination with userfaults. User 309 just fine in combination with userfaults. Userfaults trigger async
313 page faults in the guest scheduler so those gu 310 page faults in the guest scheduler so those guest processes that
314 aren't waiting for userfaults (i.e. network bo 311 aren't waiting for userfaults (i.e. network bound) can keep running in
315 the guest vcpus. 312 the guest vcpus.
316 313
317 It is generally beneficial to run one pass of 314 It is generally beneficial to run one pass of precopy live migration
318 just before starting postcopy live migration, 315 just before starting postcopy live migration, in order to avoid
319 generating userfaults for readonly guest regio 316 generating userfaults for readonly guest regions.
320 317
321 The implementation of postcopy live migration 318 The implementation of postcopy live migration currently uses one
322 single bidirectional socket but in the future 319 single bidirectional socket but in the future two different sockets
323 will be used (to reduce the latency of the use 320 will be used (to reduce the latency of the userfaults to the minimum
324 possible without having to decrease ``/proc/sy 321 possible without having to decrease ``/proc/sys/net/ipv4/tcp_wmem``).
325 322
326 The QEMU in the source node writes all pages t 323 The QEMU in the source node writes all pages that it knows are missing
327 in the destination node, into the socket, and 324 in the destination node, into the socket, and the migration thread of
328 the QEMU running in the destination node runs 325 the QEMU running in the destination node runs ``UFFDIO_COPY|ZEROPAGE``
329 ioctls on the ``userfaultfd`` in order to map 326 ioctls on the ``userfaultfd`` in order to map the received pages into the
330 guest (``UFFDIO_ZEROCOPY`` is used if the sour 327 guest (``UFFDIO_ZEROCOPY`` is used if the source page was a zero page).
331 328
332 A different postcopy thread in the destination 329 A different postcopy thread in the destination node listens with
333 poll() to the ``userfaultfd`` in parallel. Whe 330 poll() to the ``userfaultfd`` in parallel. When a ``POLLIN`` event is
334 generated after a userfault triggers, the post 331 generated after a userfault triggers, the postcopy thread read() from
335 the ``userfaultfd`` and receives the fault add 332 the ``userfaultfd`` and receives the fault address (or ``-EAGAIN`` in case the
336 userfault was already resolved and waken by a 333 userfault was already resolved and waken by a ``UFFDIO_COPY|ZEROPAGE`` run
337 by the parallel QEMU migration thread). 334 by the parallel QEMU migration thread).
338 335
339 After the QEMU postcopy thread (running in the 336 After the QEMU postcopy thread (running in the destination node) gets
340 the userfault address it writes the informatio 337 the userfault address it writes the information about the missing page
341 into the socket. The QEMU source node receives 338 into the socket. The QEMU source node receives the information and
342 roughly "seeks" to that page address and conti 339 roughly "seeks" to that page address and continues sending all
343 remaining missing pages from that new page off 340 remaining missing pages from that new page offset. Soon after that
344 (just the time to flush the tcp_wmem queue thr 341 (just the time to flush the tcp_wmem queue through the network) the
345 migration thread in the QEMU running in the de 342 migration thread in the QEMU running in the destination node will
346 receive the page that triggered the userfault 343 receive the page that triggered the userfault and it'll map it as
347 usual with the ``UFFDIO_COPY|ZEROPAGE`` (witho 344 usual with the ``UFFDIO_COPY|ZEROPAGE`` (without actually knowing if it
348 was spontaneously sent by the source or if it 345 was spontaneously sent by the source or if it was an urgent page
349 requested through a userfault). 346 requested through a userfault).
350 347
351 By the time the userfaults start, the QEMU in 348 By the time the userfaults start, the QEMU in the destination node
352 doesn't need to keep any per-page state bitmap 349 doesn't need to keep any per-page state bitmap relative to the live
353 migration around and a single per-page bitmap 350 migration around and a single per-page bitmap has to be maintained in
354 the QEMU running in the source node to know wh 351 the QEMU running in the source node to know which pages are still
355 missing in the destination node. The bitmap in 352 missing in the destination node. The bitmap in the source node is
356 checked to find which missing pages to send in 353 checked to find which missing pages to send in round robin and we seek
357 over it when receiving incoming userfaults. Af 354 over it when receiving incoming userfaults. After sending each page of
358 course the bitmap is updated accordingly. It's 355 course the bitmap is updated accordingly. It's also useful to avoid
359 sending the same page twice (in case the userf 356 sending the same page twice (in case the userfault is read by the
360 postcopy thread just before ``UFFDIO_COPY|ZERO 357 postcopy thread just before ``UFFDIO_COPY|ZEROPAGE`` runs in the migration
361 thread). 358 thread).
362 359
363 Non-cooperative userfaultfd 360 Non-cooperative userfaultfd
364 =========================== 361 ===========================
365 362
366 When the ``userfaultfd`` is monitored by an ex 363 When the ``userfaultfd`` is monitored by an external manager, the manager
367 must be able to track changes in the process v 364 must be able to track changes in the process virtual memory
368 layout. Userfaultfd can notify the manager abo 365 layout. Userfaultfd can notify the manager about such changes using
369 the same read(2) protocol as for the page faul 366 the same read(2) protocol as for the page fault notifications. The
370 manager has to explicitly enable these events 367 manager has to explicitly enable these events by setting appropriate
371 bits in ``uffdio_api.features`` passed to ``UF 368 bits in ``uffdio_api.features`` passed to ``UFFDIO_API`` ioctl:
372 369
373 ``UFFD_FEATURE_EVENT_FORK`` 370 ``UFFD_FEATURE_EVENT_FORK``
374 enable ``userfaultfd`` hooks for fork( 371 enable ``userfaultfd`` hooks for fork(). When this feature is
375 enabled, the ``userfaultfd`` context o 372 enabled, the ``userfaultfd`` context of the parent process is
376 duplicated into the newly created proc 373 duplicated into the newly created process. The manager
377 receives ``UFFD_EVENT_FORK`` with file 374 receives ``UFFD_EVENT_FORK`` with file descriptor of the new
378 ``userfaultfd`` context in the ``uffd_ 375 ``userfaultfd`` context in the ``uffd_msg.fork``.
379 376
380 ``UFFD_FEATURE_EVENT_REMAP`` 377 ``UFFD_FEATURE_EVENT_REMAP``
381 enable notifications about mremap() ca 378 enable notifications about mremap() calls. When the
382 non-cooperative process moves a virtua 379 non-cooperative process moves a virtual memory area to a
383 different location, the manager will r 380 different location, the manager will receive
384 ``UFFD_EVENT_REMAP``. The ``uffd_msg.r 381 ``UFFD_EVENT_REMAP``. The ``uffd_msg.remap`` will contain the old and
385 new addresses of the area and its orig 382 new addresses of the area and its original length.
386 383
387 ``UFFD_FEATURE_EVENT_REMOVE`` 384 ``UFFD_FEATURE_EVENT_REMOVE``
388 enable notifications about madvise(MAD 385 enable notifications about madvise(MADV_REMOVE) and
389 madvise(MADV_DONTNEED) calls. The even 386 madvise(MADV_DONTNEED) calls. The event ``UFFD_EVENT_REMOVE`` will
390 be generated upon these calls to madvi 387 be generated upon these calls to madvise(). The ``uffd_msg.remove``
391 will contain start and end addresses o 388 will contain start and end addresses of the removed area.
392 389
393 ``UFFD_FEATURE_EVENT_UNMAP`` 390 ``UFFD_FEATURE_EVENT_UNMAP``
394 enable notifications about memory unma 391 enable notifications about memory unmapping. The manager will
395 get ``UFFD_EVENT_UNMAP`` with ``uffd_m 392 get ``UFFD_EVENT_UNMAP`` with ``uffd_msg.remove`` containing start and
396 end addresses of the unmapped area. 393 end addresses of the unmapped area.
397 394
398 Although the ``UFFD_FEATURE_EVENT_REMOVE`` and 395 Although the ``UFFD_FEATURE_EVENT_REMOVE`` and ``UFFD_FEATURE_EVENT_UNMAP``
399 are pretty similar, they quite differ in the a 396 are pretty similar, they quite differ in the action expected from the
400 ``userfaultfd`` manager. In the former case, t 397 ``userfaultfd`` manager. In the former case, the virtual memory is
401 removed, but the area is not, the area remains 398 removed, but the area is not, the area remains monitored by the
402 ``userfaultfd``, and if a page fault occurs in 399 ``userfaultfd``, and if a page fault occurs in that area it will be
403 delivered to the manager. The proper resolutio 400 delivered to the manager. The proper resolution for such page fault is
404 to zeromap the faulting address. However, in t 401 to zeromap the faulting address. However, in the latter case, when an
405 area is unmapped, either explicitly (with munm 402 area is unmapped, either explicitly (with munmap() system call), or
406 implicitly (e.g. during mremap()), the area is 403 implicitly (e.g. during mremap()), the area is removed and in turn the
407 ``userfaultfd`` context for such area disappea 404 ``userfaultfd`` context for such area disappears too and the manager will
408 not get further userland page faults from the 405 not get further userland page faults from the removed area. Still, the
409 notification is required in order to prevent m 406 notification is required in order to prevent manager from using
410 ``UFFDIO_COPY`` on the unmapped area. 407 ``UFFDIO_COPY`` on the unmapped area.
411 408
412 Unlike userland page faults which have to be s 409 Unlike userland page faults which have to be synchronous and require
413 explicit or implicit wakeup, all the events ar 410 explicit or implicit wakeup, all the events are delivered
414 asynchronously and the non-cooperative process 411 asynchronously and the non-cooperative process resumes execution as
415 soon as manager executes read(). The ``userfau 412 soon as manager executes read(). The ``userfaultfd`` manager should
416 carefully synchronize calls to ``UFFDIO_COPY`` 413 carefully synchronize calls to ``UFFDIO_COPY`` with the events
417 processing. To aid the synchronization, the `` 414 processing. To aid the synchronization, the ``UFFDIO_COPY`` ioctl will
418 return ``-ENOSPC`` when the monitored process 415 return ``-ENOSPC`` when the monitored process exits at the time of
419 ``UFFDIO_COPY``, and ``-ENOENT``, when the non 416 ``UFFDIO_COPY``, and ``-ENOENT``, when the non-cooperative process has changed
420 its virtual memory layout simultaneously with 417 its virtual memory layout simultaneously with outstanding ``UFFDIO_COPY``
421 operation. 418 operation.
422 419
423 The current asynchronous model of the event de 420 The current asynchronous model of the event delivery is optimal for
424 single threaded non-cooperative ``userfaultfd` 421 single threaded non-cooperative ``userfaultfd`` manager implementations. A
425 synchronous event delivery model can be added 422 synchronous event delivery model can be added later as a new
426 ``userfaultfd`` feature to facilitate multithr 423 ``userfaultfd`` feature to facilitate multithreading enhancements of the
427 non cooperative manager, for example to allow 424 non cooperative manager, for example to allow ``UFFDIO_COPY`` ioctls to
428 run in parallel to the event reception. Single 425 run in parallel to the event reception. Single threaded
429 implementations should continue to use the cur 426 implementations should continue to use the current async event
430 delivery model instead. 427 delivery model instead.
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.