Documentation/vm/userfaultfd.txt

   1 = Userfaultfd =
   2
   3 == Objective ==
   4
   5 Userfaults allow the implementation of on-demand paging from userland
   6 and more generally they allow userland to take control of various
   7 memory page faults, something otherwise only the kernel code could do.
   8
   9 For example userfaults allows a proper and more optimal implementation
  10 of the PROT_NONE+SIGSEGV trick.
  11
  12 == Design ==
  13
  14 Userfaults are delivered and resolved through the userfaultfd syscall.
  15
  16 The userfaultfd (aside from registering and unregistering virtual
  17 memory ranges) provides two primary functionalities:
  18
  19 1) read/POLLIN protocol to notify a userland thread of the faults
  20    happening
  21
  22 2) various UFFDIO_* ioctls that can manage the virtual memory regions
  23    registered in the userfaultfd that allows userland to efficiently
  24    resolve the userfaults it receives via 1) or to manage the virtual
  25    memory in the background
  26
  27 The real advantage of userfaults if compared to regular virtual memory
  28 management of mremap/mprotect is that the userfaults in all their
  29 operations never involve heavyweight structures like vmas (in fact the
  30 userfaultfd runtime load never takes the mmap_sem for writing).
  31
  32 Vmas are not suitable for page- (or hugepage) granular fault tracking
  33 when dealing with virtual address spaces that could span
  34 Terabytes. Too many vmas would be needed for that.
  35
  36 The userfaultfd once opened by invoking the syscall, can also be
  37 passed using unix domain sockets to a manager process, so the same
  38 manager process could handle the userfaults of a multitude of
  39 different processes without them being aware about what is going on
  40 (well of course unless they later try to use the userfaultfd
  41 themselves on the same region the manager is already tracking, which
  42 is a corner case that would currently return -EBUSY).
  43
  44 == API ==
  45
  46 When first opened the userfaultfd must be enabled invoking the
  47 UFFDIO_API ioctl specifying a uffdio_api.api value set to UFFD_API (or
  48 a later API version) which will specify the read/POLLIN protocol
  49 userland intends to speak on the UFFD and the uffdio_api.features
  50 userland requires. The UFFDIO_API ioctl if successful (i.e. if the
  51 requested uffdio_api.api is spoken also by the running kernel and the
  52 requested features are going to be enabled) will return into
  53 uffdio_api.features and uffdio_api.ioctls two 64bit bitmasks of
  54 respectively all the available features of the read(2) protocol and
  55 the generic ioctl available.
  56
  57 The uffdio_api.features bitmask returned by the UFFDIO_API ioctl
  58 defines what memory types are supported by the userfaultfd and what
  59 events, except page fault notifications, may be generated.
  60
  61 If the kernel supports registering userfaultfd ranges on hugetlbfs
  62 virtual memory areas, UFFD_FEATURE_MISSING_HUGETLBFS will be set in
  63 uffdio_api.features. Similarly, UFFD_FEATURE_MISSING_SHMEM will be
  64 set if the kernel supports registering userfaultfd ranges on shared
  65 memory (covering all shmem APIs, i.e. tmpfs, IPCSHM, /dev/zero
  66 MAP_SHARED, memfd_create, etc).
  67
  68 The userland application that wants to use userfaultfd with hugetlbfs
  69 or shared memory need to set the corresponding flag in
  70 uffdio_api.features to enable those features.
  71
  72 If the userland desires to receive notifications for events other than
  73 page faults, it has to verify that uffdio_api.features has appropriate
  74 UFFD_FEATURE_EVENT_* bits set. These events are described in more
  75 detail below in "Non-cooperative userfaultfd" section.
  76
  77 Once the userfaultfd has been enabled the UFFDIO_REGISTER ioctl should
  78 be invoked (if present in the returned uffdio_api.ioctls bitmask) to
  79 register a memory range in the userfaultfd by setting the
  80 uffdio_register structure accordingly. The uffdio_register.mode
  81 bitmask will specify to the kernel which kind of faults to track for
  82 the range (UFFDIO_REGISTER_MODE_MISSING would track missing
  83 pages). The UFFDIO_REGISTER ioctl will return the
  84 uffdio_register.ioctls bitmask of ioctls that are suitable to resolve
  85 userfaults on the range registered. Not all ioctls will necessarily be
  86 supported for all memory types depending on the underlying virtual
  87 memory backend (anonymous memory vs tmpfs vs real filebacked
  88 mappings).
  89
  90 Userland can use the uffdio_register.ioctls to manage the virtual
  91 address space in the background (to add or potentially also remove
  92 memory from the userfaultfd registered range). This means a userfault
  93 could be triggering just before userland maps in the background the
  94 user-faulted page.
  95
  96 The primary ioctl to resolve userfaults is UFFDIO_COPY. That
  97 atomically copies a page into the userfault registered range and wakes
  98 up the blocked userfaults (unless uffdio_copy.mode &
  99 UFFDIO_COPY_MODE_DONTWAKE is set). Other ioctl works similarly to
 100 UFFDIO_COPY. They're atomic as in guaranteeing that nothing can see an
 101 half copied page since it'll keep userfaulting until the copy has
 102 finished.
 103
 104 == QEMU/KVM ==
 105
 106 QEMU/KVM is using the userfaultfd syscall to implement postcopy live
 107 migration. Postcopy live migration is one form of memory
 108 externalization consisting of a virtual machine running with part or
 109 all of its memory residing on a different node in the cloud. The
 110 userfaultfd abstraction is generic enough that not a single line of
 111 KVM kernel code had to be modified in order to add postcopy live
 112 migration to QEMU.
 113
 114 Guest async page faults, FOLL_NOWAIT and all other GUP features work
 115 just fine in combination with userfaults. Userfaults trigger async
 116 page faults in the guest scheduler so those guest processes that
 117 aren't waiting for userfaults (i.e. network bound) can keep running in
 118 the guest vcpus.
 119
 120 It is generally beneficial to run one pass of precopy live migration
 121 just before starting postcopy live migration, in order to avoid
 122 generating userfaults for readonly guest regions.
 123
 124 The implementation of postcopy live migration currently uses one
 125 single bidirectional socket but in the future two different sockets
 126 will be used (to reduce the latency of the userfaults to the minimum
 127 possible without having to decrease /proc/sys/net/ipv4/tcp_wmem).
 128
 129 The QEMU in the source node writes all pages that it knows are missing
 130 in the destination node, into the socket, and the migration thread of
 131 the QEMU running in the destination node runs UFFDIO_COPY|ZEROPAGE
 132 ioctls on the userfaultfd in order to map the received pages into the
 133 guest (UFFDIO_ZEROCOPY is used if the source page was a zero page).
 134
 135 A different postcopy thread in the destination node listens with
 136 poll() to the userfaultfd in parallel. When a POLLIN event is
 137 generated after a userfault triggers, the postcopy thread read() from
 138 the userfaultfd and receives the fault address (or -EAGAIN in case the
 139 userfault was already resolved and waken by a UFFDIO_COPY|ZEROPAGE run
 140 by the parallel QEMU migration thread).
 141
 142 After the QEMU postcopy thread (running in the destination node) gets
 143 the userfault address it writes the information about the missing page
 144 into the socket. The QEMU source node receives the information and
 145 roughly "seeks" to that page address and continues sending all
 146 remaining missing pages from that new page offset. Soon after that
 147 (just the time to flush the tcp_wmem queue through the network) the
 148 migration thread in the QEMU running in the destination node will
 149 receive the page that triggered the userfault and it'll map it as
 150 usual with the UFFDIO_COPY|ZEROPAGE (without actually knowing if it
 151 was spontaneously sent by the source or if it was an urgent page
 152 requested through a userfault).
 153
 154 By the time the userfaults start, the QEMU in the destination node
 155 doesn't need to keep any per-page state bitmap relative to the live
 156 migration around and a single per-page bitmap has to be maintained in
 157 the QEMU running in the source node to know which pages are still
 158 missing in the destination node. The bitmap in the source node is
 159 checked to find which missing pages to send in round robin and we seek
 160 over it when receiving incoming userfaults. After sending each page of
 161 course the bitmap is updated accordingly. It's also useful to avoid
 162 sending the same page twice (in case the userfault is read by the
 163 postcopy thread just before UFFDIO_COPY|ZEROPAGE runs in the migration
 164 thread).
 165
 166 == Non-cooperative userfaultfd ==
 167
 168 When the userfaultfd is monitored by an external manager, the manager
 169 must be able to track changes in the process virtual memory
 170 layout. Userfaultfd can notify the manager about such changes using
 171 the same read(2) protocol as for the page fault notifications. The
 172 manager has to explicitly enable these events by setting appropriate
 173 bits in uffdio_api.features passed to UFFDIO_API ioctl:
 174
 175 UFFD_FEATURE_EVENT_FORK - enable userfaultfd hooks for fork(). When
 176 this feature is enabled, the userfaultfd context of the parent process
 177 is duplicated into the newly created process. The manager receives
 178 UFFD_EVENT_FORK with file descriptor of the new userfaultfd context in
 179 the uffd_msg.fork.
 180
 181 UFFD_FEATURE_EVENT_REMAP - enable notifications about mremap()
 182 calls. When the non-cooperative process moves a virtual memory area to
 183 a different location, the manager will receive UFFD_EVENT_REMAP. The
 184 uffd_msg.remap will contain the old and new addresses of the area and
 185 its original length.
 186
 187 UFFD_FEATURE_EVENT_REMOVE - enable notifications about
 188 madvise(MADV_REMOVE) and madvise(MADV_DONTNEED) calls. The event
 189 UFFD_EVENT_REMOVE will be generated upon these calls to madvise. The
 190 uffd_msg.remove will contain start and end addresses of the removed
 191 area.
 192
 193 UFFD_FEATURE_EVENT_UNMAP - enable notifications about memory
 194 unmapping. The manager will get UFFD_EVENT_UNMAP with uffd_msg.remove
 195 containing start and end addresses of the unmapped area.
 196
 197 Although the UFFD_FEATURE_EVENT_REMOVE and UFFD_FEATURE_EVENT_UNMAP
 198 are pretty similar, they quite differ in the action expected from the
 199 userfaultfd manager. In the former case, the virtual memory is
 200 removed, but the area is not, the area remains monitored by the
 201 userfaultfd, and if a page fault occurs in that area it will be
 202 delivered to the manager. The proper resolution for such page fault is
 203 to zeromap the faulting address. However, in the latter case, when an
 204 area is unmapped, either explicitly (with munmap() system call), or
 205 implicitly (e.g. during mremap()), the area is removed and in turn the
 206 userfaultfd context for such area disappears too and the manager will
 207 not get further userland page faults from the removed area. Still, the
 208 notification is required in order to prevent manager from using
 209 UFFDIO_COPY on the unmapped area.
 210
 211 Unlike userland page faults which have to be synchronous and require
 212 explicit or implicit wakeup, all the events are delivered
 213 asynchronously and the non-cooperative process resumes execution as
 214 soon as manager executes read(). The userfaultfd manager should
 215 carefully synchronize calls to UFFDIO_COPY with the events
 216 processing. To aid the synchronization, the UFFDIO_COPY ioctl will
 217 return -ENOSPC when the monitored process exits at the time of
 218 UFFDIO_COPY, and -ENOENT, when the non-cooperative process has changed
 219 its virtual memory layout simultaneously with outstanding UFFDIO_COPY
 220 operation.
 221
 222 The current asynchronous model of the event delivery is optimal for
 223 single threaded non-cooperative userfaultfd manager implementations. A
 224 synchronous event delivery model can be added later as a new
 225 userfaultfd feature to facilitate multithreading enhancements of the
 226 non cooperative manager, for example to allow UFFDIO_COPY ioctls to
 227 run in parallel to the event reception. Single threaded
 228 implementations should continue to use the current async event
 229 delivery model instead.