dma-buf.rst 8.9 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250
  1. Buffer Sharing and Synchronization
  2. ==================================
  3. The dma-buf subsystem provides the framework for sharing buffers for
  4. hardware (DMA) access across multiple device drivers and subsystems, and
  5. for synchronizing asynchronous hardware access.
  6. This is used, for example, by drm "prime" multi-GPU support, but is of
  7. course not limited to GPU use cases.
  8. The three main components of this are: (1) dma-buf, representing a
  9. sg_table and exposed to userspace as a file descriptor to allow passing
  10. between devices, (2) fence, which provides a mechanism to signal when
  11. one device has finished access, and (3) reservation, which manages the
  12. shared or exclusive fence(s) associated with the buffer.
  13. Shared DMA Buffers
  14. ------------------
  15. This document serves as a guide to device-driver writers on what is the dma-buf
  16. buffer sharing API, how to use it for exporting and using shared buffers.
  17. Any device driver which wishes to be a part of DMA buffer sharing, can do so as
  18. either the 'exporter' of buffers, or the 'user' or 'importer' of buffers.
  19. Say a driver A wants to use buffers created by driver B, then we call B as the
  20. exporter, and A as buffer-user/importer.
  21. The exporter
  22. - implements and manages operations in :c:type:`struct dma_buf_ops
  23. <dma_buf_ops>` for the buffer,
  24. - allows other users to share the buffer by using dma_buf sharing APIs,
  25. - manages the details of buffer allocation, wrapped in a :c:type:`struct
  26. dma_buf <dma_buf>`,
  27. - decides about the actual backing storage where this allocation happens,
  28. - and takes care of any migration of scatterlist - for all (shared) users of
  29. this buffer.
  30. The buffer-user
  31. - is one of (many) sharing users of the buffer.
  32. - doesn't need to worry about how the buffer is allocated, or where.
  33. - and needs a mechanism to get access to the scatterlist that makes up this
  34. buffer in memory, mapped into its own address space, so it can access the
  35. same area of memory. This interface is provided by :c:type:`struct
  36. dma_buf_attachment <dma_buf_attachment>`.
  37. Any exporters or users of the dma-buf buffer sharing framework must have a
  38. 'select DMA_SHARED_BUFFER' in their respective Kconfigs.
  39. Userspace Interface Notes
  40. ~~~~~~~~~~~~~~~~~~~~~~~~~
  41. Mostly a DMA buffer file descriptor is simply an opaque object for userspace,
  42. and hence the generic interface exposed is very minimal. There's a few things to
  43. consider though:
  44. - Since kernel 3.12 the dma-buf FD supports the llseek system call, but only
  45. with offset=0 and whence=SEEK_END|SEEK_SET. SEEK_SET is supported to allow
  46. the usual size discover pattern size = SEEK_END(0); SEEK_SET(0). Every other
  47. llseek operation will report -EINVAL.
  48. If llseek on dma-buf FDs isn't support the kernel will report -ESPIPE for all
  49. cases. Userspace can use this to detect support for discovering the dma-buf
  50. size using llseek.
  51. - In order to avoid fd leaks on exec, the FD_CLOEXEC flag must be set
  52. on the file descriptor. This is not just a resource leak, but a
  53. potential security hole. It could give the newly exec'd application
  54. access to buffers, via the leaked fd, to which it should otherwise
  55. not be permitted access.
  56. The problem with doing this via a separate fcntl() call, versus doing it
  57. atomically when the fd is created, is that this is inherently racy in a
  58. multi-threaded app[3]. The issue is made worse when it is library code
  59. opening/creating the file descriptor, as the application may not even be
  60. aware of the fd's.
  61. To avoid this problem, userspace must have a way to request O_CLOEXEC
  62. flag be set when the dma-buf fd is created. So any API provided by
  63. the exporting driver to create a dmabuf fd must provide a way to let
  64. userspace control setting of O_CLOEXEC flag passed in to dma_buf_fd().
  65. - Memory mapping the contents of the DMA buffer is also supported. See the
  66. discussion below on `CPU Access to DMA Buffer Objects`_ for the full details.
  67. - The DMA buffer FD is also pollable, see `Implicit Fence Poll Support`_ below for
  68. details.
  69. Basic Operation and Device DMA Access
  70. ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  71. .. kernel-doc:: drivers/dma-buf/dma-buf.c
  72. :doc: dma buf device access
  73. CPU Access to DMA Buffer Objects
  74. ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  75. .. kernel-doc:: drivers/dma-buf/dma-buf.c
  76. :doc: cpu access
  77. Implicit Fence Poll Support
  78. ~~~~~~~~~~~~~~~~~~~~~~~~~~~
  79. .. kernel-doc:: drivers/dma-buf/dma-buf.c
  80. :doc: implicit fence polling
  81. Kernel Functions and Structures Reference
  82. ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  83. .. kernel-doc:: drivers/dma-buf/dma-buf.c
  84. :export:
  85. .. kernel-doc:: include/linux/dma-buf.h
  86. :internal:
  87. Reservation Objects
  88. -------------------
  89. .. kernel-doc:: drivers/dma-buf/dma-resv.c
  90. :doc: Reservation Object Overview
  91. .. kernel-doc:: drivers/dma-buf/dma-resv.c
  92. :export:
  93. .. kernel-doc:: include/linux/dma-resv.h
  94. :internal:
  95. DMA Fences
  96. ----------
  97. .. kernel-doc:: drivers/dma-buf/dma-fence.c
  98. :doc: DMA fences overview
  99. DMA Fence Cross-Driver Contract
  100. ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  101. .. kernel-doc:: drivers/dma-buf/dma-fence.c
  102. :doc: fence cross-driver contract
  103. DMA Fence Signalling Annotations
  104. ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  105. .. kernel-doc:: drivers/dma-buf/dma-fence.c
  106. :doc: fence signalling annotation
  107. DMA Fences Functions Reference
  108. ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  109. .. kernel-doc:: drivers/dma-buf/dma-fence.c
  110. :export:
  111. .. kernel-doc:: include/linux/dma-fence.h
  112. :internal:
  113. Seqno Hardware Fences
  114. ~~~~~~~~~~~~~~~~~~~~~
  115. .. kernel-doc:: include/linux/seqno-fence.h
  116. :internal:
  117. DMA Fence Array
  118. ~~~~~~~~~~~~~~~
  119. .. kernel-doc:: drivers/dma-buf/dma-fence-array.c
  120. :export:
  121. .. kernel-doc:: include/linux/dma-fence-array.h
  122. :internal:
  123. DMA Fence uABI/Sync File
  124. ~~~~~~~~~~~~~~~~~~~~~~~~
  125. .. kernel-doc:: drivers/dma-buf/sync_file.c
  126. :export:
  127. .. kernel-doc:: include/linux/sync_file.h
  128. :internal:
  129. Indefinite DMA Fences
  130. ~~~~~~~~~~~~~~~~~~~~~
  131. At various times &dma_fence with an indefinite time until dma_fence_wait()
  132. finishes have been proposed. Examples include:
  133. * Future fences, used in HWC1 to signal when a buffer isn't used by the display
  134. any longer, and created with the screen update that makes the buffer visible.
  135. The time this fence completes is entirely under userspace's control.
  136. * Proxy fences, proposed to handle &drm_syncobj for which the fence has not yet
  137. been set. Used to asynchronously delay command submission.
  138. * Userspace fences or gpu futexes, fine-grained locking within a command buffer
  139. that userspace uses for synchronization across engines or with the CPU, which
  140. are then imported as a DMA fence for integration into existing winsys
  141. protocols.
  142. * Long-running compute command buffers, while still using traditional end of
  143. batch DMA fences for memory management instead of context preemption DMA
  144. fences which get reattached when the compute job is rescheduled.
  145. Common to all these schemes is that userspace controls the dependencies of these
  146. fences and controls when they fire. Mixing indefinite fences with normal
  147. in-kernel DMA fences does not work, even when a fallback timeout is included to
  148. protect against malicious userspace:
  149. * Only the kernel knows about all DMA fence dependencies, userspace is not aware
  150. of dependencies injected due to memory management or scheduler decisions.
  151. * Only userspace knows about all dependencies in indefinite fences and when
  152. exactly they will complete, the kernel has no visibility.
  153. Furthermore the kernel has to be able to hold up userspace command submission
  154. for memory management needs, which means we must support indefinite fences being
  155. dependent upon DMA fences. If the kernel also support indefinite fences in the
  156. kernel like a DMA fence, like any of the above proposal would, there is the
  157. potential for deadlocks.
  158. .. kernel-render:: DOT
  159. :alt: Indefinite Fencing Dependency Cycle
  160. :caption: Indefinite Fencing Dependency Cycle
  161. digraph "Fencing Cycle" {
  162. node [shape=box bgcolor=grey style=filled]
  163. kernel [label="Kernel DMA Fences"]
  164. userspace [label="userspace controlled fences"]
  165. kernel -> userspace [label="memory management"]
  166. userspace -> kernel [label="Future fence, fence proxy, ..."]
  167. { rank=same; kernel userspace }
  168. }
  169. This means that the kernel might accidentally create deadlocks
  170. through memory management dependencies which userspace is unaware of, which
  171. randomly hangs workloads until the timeout kicks in. Workloads, which from
  172. userspace's perspective, do not contain a deadlock. In such a mixed fencing
  173. architecture there is no single entity with knowledge of all dependencies.
  174. Thefore preventing such deadlocks from within the kernel is not possible.
  175. The only solution to avoid dependencies loops is by not allowing indefinite
  176. fences in the kernel. This means:
  177. * No future fences, proxy fences or userspace fences imported as DMA fences,
  178. with or without a timeout.
  179. * No DMA fences that signal end of batchbuffer for command submission where
  180. userspace is allowed to use userspace fencing or long running compute
  181. workloads. This also means no implicit fencing for shared buffers in these
  182. cases.