botching-up-ioctls.rst 11 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225
  1. =================================
  2. (How to avoid) Botching up ioctls
  3. =================================
  4. From: https://blog.ffwll.ch/2013/11/botching-up-ioctls.html
  5. By: Daniel Vetter, Copyright © 2013 Intel Corporation
  6. One clear insight kernel graphics hackers gained in the past few years is that
  7. trying to come up with a unified interface to manage the execution units and
  8. memory on completely different GPUs is a futile effort. So nowadays every
  9. driver has its own set of ioctls to allocate memory and submit work to the GPU.
  10. Which is nice, since there's no more insanity in the form of fake-generic, but
  11. actually only used once interfaces. But the clear downside is that there's much
  12. more potential to screw things up.
  13. To avoid repeating all the same mistakes again I've written up some of the
  14. lessons learned while botching the job for the drm/i915 driver. Most of these
  15. only cover technicalities and not the big-picture issues like what the command
  16. submission ioctl exactly should look like. Learning these lessons is probably
  17. something every GPU driver has to do on its own.
  18. Prerequisites
  19. -------------
  20. First the prerequisites. Without these you have already failed, because you
  21. will need to add a 32-bit compat layer:
  22. * Only use fixed sized integers. To avoid conflicts with typedefs in userspace
  23. the kernel has special types like __u32, __s64. Use them.
  24. * Align everything to the natural size and use explicit padding. 32-bit
  25. platforms don't necessarily align 64-bit values to 64-bit boundaries, but
  26. 64-bit platforms do. So we always need padding to the natural size to get
  27. this right.
  28. * Pad the entire struct to a multiple of 64-bits if the structure contains
  29. 64-bit types - the structure size will otherwise differ on 32-bit versus
  30. 64-bit. Having a different structure size hurts when passing arrays of
  31. structures to the kernel, or if the kernel checks the structure size, which
  32. e.g. the drm core does.
  33. * Pointers are __u64, cast from/to a uintprt_t on the userspace side and
  34. from/to a void __user * in the kernel. Try really hard not to delay this
  35. conversion or worse, fiddle the raw __u64 through your code since that
  36. diminishes the checking tools like sparse can provide. The macro
  37. u64_to_user_ptr can be used in the kernel to avoid warnings about integers
  38. and pointers of different sizes.
  39. Basics
  40. ------
  41. With the joys of writing a compat layer avoided we can take a look at the basic
  42. fumbles. Neglecting these will make backward and forward compatibility a real
  43. pain. And since getting things wrong on the first attempt is guaranteed you
  44. will have a second iteration or at least an extension for any given interface.
  45. * Have a clear way for userspace to figure out whether your new ioctl or ioctl
  46. extension is supported on a given kernel. If you can't rely on old kernels
  47. rejecting the new flags/modes or ioctls (since doing that was botched in the
  48. past) then you need a driver feature flag or revision number somewhere.
  49. * Have a plan for extending ioctls with new flags or new fields at the end of
  50. the structure. The drm core checks the passed-in size for each ioctl call
  51. and zero-extends any mismatches between kernel and userspace. That helps,
  52. but isn't a complete solution since newer userspace on older kernels won't
  53. notice that the newly added fields at the end get ignored. So this still
  54. needs a new driver feature flags.
  55. * Check all unused fields and flags and all the padding for whether it's 0,
  56. and reject the ioctl if that's not the case. Otherwise your nice plan for
  57. future extensions is going right down the gutters since someone will submit
  58. an ioctl struct with random stack garbage in the yet unused parts. Which
  59. then bakes in the ABI that those fields can never be used for anything else
  60. but garbage. This is also the reason why you must explicitly pad all
  61. structures, even if you never use them in an array - the padding the compiler
  62. might insert could contain garbage.
  63. * Have simple testcases for all of the above.
  64. Fun with Error Paths
  65. --------------------
  66. Nowadays we don't have any excuse left any more for drm drivers being neat
  67. little root exploits. This means we both need full input validation and solid
  68. error handling paths - GPUs will die eventually in the oddmost corner cases
  69. anyway:
  70. * The ioctl must check for array overflows. Also it needs to check for
  71. over/underflows and clamping issues of integer values in general. The usual
  72. example is sprite positioning values fed directly into the hardware with the
  73. hardware just having 12 bits or so. Works nicely until some odd display
  74. server doesn't bother with clamping itself and the cursor wraps around the
  75. screen.
  76. * Have simple testcases for every input validation failure case in your ioctl.
  77. Check that the error code matches your expectations. And finally make sure
  78. that you only test for one single error path in each subtest by submitting
  79. otherwise perfectly valid data. Without this an earlier check might reject
  80. the ioctl already and shadow the codepath you actually want to test, hiding
  81. bugs and regressions.
  82. * Make all your ioctls restartable. First X really loves signals and second
  83. this will allow you to test 90% of all error handling paths by just
  84. interrupting your main test suite constantly with signals. Thanks to X's
  85. love for signal you'll get an excellent base coverage of all your error
  86. paths pretty much for free for graphics drivers. Also, be consistent with
  87. how you handle ioctl restarting - e.g. drm has a tiny drmIoctl helper in its
  88. userspace library. The i915 driver botched this with the set_tiling ioctl,
  89. now we're stuck forever with some arcane semantics in both the kernel and
  90. userspace.
  91. * If you can't make a given codepath restartable make a stuck task at least
  92. killable. GPUs just die and your users won't like you more if you hang their
  93. entire box (by means of an unkillable X process). If the state recovery is
  94. still too tricky have a timeout or hangcheck safety net as a last-ditch
  95. effort in case the hardware has gone bananas.
  96. * Have testcases for the really tricky corner cases in your error recovery code
  97. - it's way too easy to create a deadlock between your hangcheck code and
  98. waiters.
  99. Time, Waiting and Missing it
  100. ----------------------------
  101. GPUs do most everything asynchronously, so we have a need to time operations and
  102. wait for outstanding ones. This is really tricky business; at the moment none of
  103. the ioctls supported by the drm/i915 get this fully right, which means there's
  104. still tons more lessons to learn here.
  105. * Use CLOCK_MONOTONIC as your reference time, always. It's what alsa, drm and
  106. v4l use by default nowadays. But let userspace know which timestamps are
  107. derived from different clock domains like your main system clock (provided
  108. by the kernel) or some independent hardware counter somewhere else. Clocks
  109. will mismatch if you look close enough, but if performance measuring tools
  110. have this information they can at least compensate. If your userspace can
  111. get at the raw values of some clocks (e.g. through in-command-stream
  112. performance counter sampling instructions) consider exposing those also.
  113. * Use __s64 seconds plus __u64 nanoseconds to specify time. It's not the most
  114. convenient time specification, but it's mostly the standard.
  115. * Check that input time values are normalized and reject them if not. Note
  116. that the kernel native struct ktime has a signed integer for both seconds
  117. and nanoseconds, so beware here.
  118. * For timeouts, use absolute times. If you're a good fellow and made your
  119. ioctl restartable relative timeouts tend to be too coarse and can
  120. indefinitely extend your wait time due to rounding on each restart.
  121. Especially if your reference clock is something really slow like the display
  122. frame counter. With a spec lawyer hat on this isn't a bug since timeouts can
  123. always be extended - but users will surely hate you if their neat animations
  124. starts to stutter due to this.
  125. * Consider ditching any synchronous wait ioctls with timeouts and just deliver
  126. an asynchronous event on a pollable file descriptor. It fits much better
  127. into event driven applications' main loop.
  128. * Have testcases for corner-cases, especially whether the return values for
  129. already-completed events, successful waits and timed-out waits are all sane
  130. and suiting to your needs.
  131. Leaking Resources, Not
  132. ----------------------
  133. A full-blown drm driver essentially implements a little OS, but specialized to
  134. the given GPU platforms. This means a driver needs to expose tons of handles
  135. for different objects and other resources to userspace. Doing that right
  136. entails its own little set of pitfalls:
  137. * Always attach the lifetime of your dynamically created resources to the
  138. lifetime of a file descriptor. Consider using a 1:1 mapping if your resource
  139. needs to be shared across processes - fd-passing over unix domain sockets
  140. also simplifies lifetime management for userspace.
  141. * Always have O_CLOEXEC support.
  142. * Ensure that you have sufficient insulation between different clients. By
  143. default pick a private per-fd namespace which forces any sharing to be done
  144. explicitly. Only go with a more global per-device namespace if the objects
  145. are truly device-unique. One counterexample in the drm modeset interfaces is
  146. that the per-device modeset objects like connectors share a namespace with
  147. framebuffer objects, which mostly are not shared at all. A separate
  148. namespace, private by default, for framebuffers would have been more
  149. suitable.
  150. * Think about uniqueness requirements for userspace handles. E.g. for most drm
  151. drivers it's a userspace bug to submit the same object twice in the same
  152. command submission ioctl. But then if objects are shareable userspace needs
  153. to know whether it has seen an imported object from a different process
  154. already or not. I haven't tried this myself yet due to lack of a new class
  155. of objects, but consider using inode numbers on your shared file descriptors
  156. as unique identifiers - it's how real files are told apart, too.
  157. Unfortunately this requires a full-blown virtual filesystem in the kernel.
  158. Last, but not Least
  159. -------------------
  160. Not every problem needs a new ioctl:
  161. * Think hard whether you really want a driver-private interface. Of course
  162. it's much quicker to push a driver-private interface than engaging in
  163. lengthy discussions for a more generic solution. And occasionally doing a
  164. private interface to spearhead a new concept is what's required. But in the
  165. end, once the generic interface comes around you'll end up maintainer two
  166. interfaces. Indefinitely.
  167. * Consider other interfaces than ioctls. A sysfs attribute is much better for
  168. per-device settings, or for child objects with fairly static lifetimes (like
  169. output connectors in drm with all the detection override attributes). Or
  170. maybe only your testsuite needs this interface, and then debugfs with its
  171. disclaimer of not having a stable ABI would be better.
  172. Finally, the name of the game is to get it right on the first attempt, since if
  173. your driver proves popular and your hardware platforms long-lived then you'll
  174. be stuck with a given ioctl essentially forever. You can try to deprecate
  175. horrible ioctls on newer iterations of your hardware, but generally it takes
  176. years to accomplish this. And then again years until the last user able to
  177. complain about regressions disappears, too.