summaryrefslogtreecommitdiff
path: root/Documentation/ioctl/botching-up-ioctls.txt
blob: d02cfb48901c5c5c2b357822e51da3e2a0621fde (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
(How to avoid) Botching up ioctls
=================================

From: http://blog.ffwll.ch/2013/11/botching-up-ioctls.html

By: Daniel Vetter, Copyright © 2013 Intel Corporation

One clear insight kernel graphics hackers gained in the past few years is that
trying to come up with a unified interface to manage the execution units and
memory on completely different GPUs is a futile effort. So nowadays every
driver has its own set of ioctls to allocate memory and submit work to the GPU.
Which is nice, since there's no more insanity in the form of fake-generic, but
actually only used once interfaces. But the clear downside is that there's much
more potential to screw things up.

To avoid repeating all the same mistakes again I've written up some of the
lessons learned while botching the job for the drm/i915 driver. Most of these
only cover technicalities and not the big-picture issues like what the command
submission ioctl exactly should look like. Learning these lessons is probably
something every GPU driver has to do on its own.


Prerequisites
-------------

First the prerequisites. Without these you have already failed, because you
will need to add a 32-bit compat layer:

 * Only use fixed sized integers. To avoid conflicts with typedefs in userspace
   the kernel has special types like __u32, __s64. Use them.

 * Align everything to the natural size and use explicit padding. 32-bit
   platforms don't necessarily align 64-bit values to 64-bit boundaries, but
   64-bit platforms do. So we always need padding to the natural size to get
   this right.

 * Pad the entire struct to a multiple of 64-bits if the structure contains
   64-bit types - the structure size will otherwise differ on 32-bit versus
   64-bit. Having a different structure size hurts when passing arrays of
   structures to the kernel, or if the kernel checks the structure size, which
   e.g. the drm core does.

 * Pointers are __u64, cast from/to a uintprt_t on the userspace side and
   from/to a void __user * in the kernel. Try really hard not to delay this
   conversion or worse, fiddle the raw __u64 through your code since that
   diminishes the checking tools like sparse can provide. The macro
   u64_to_user_ptr can be used in the kernel to avoid warnings about integers
   and pointres of different sizes.


Basics
------

With the joys of writing a compat layer avoided we can take a look at the basic
fumbles. Neglecting these will make backward and forward compatibility a real
pain. And since getting things wrong on the first attempt is guaranteed you
will have a second iteration or at least an extension for any given interface.

 * Have a clear way for userspace to figure out whether your new ioctl or ioctl
   extension is supported on a given kernel. If you can't rely on old kernels
   rejecting the new flags/modes or ioctls (since doing that was botched in the
   past) then you need a driver feature flag or revision number somewhere.

 * Have a plan for extending ioctls with new flags or new fields at the end of
   the structure. The drm core checks the passed-in size for each ioctl call
   and zero-extends any mismatches between kernel and userspace. That helps,
   but isn't a complete solution since newer userspace on older kernels won't
   notice that the newly added fields at the end get ignored. So this still
   needs a new driver feature flags.

 * Check all unused fields and flags and all the padding for whether it's 0,
   and reject the ioctl if that's not the case. Otherwise your nice plan for
   future extensions is going right down the gutters since someone will submit
   an ioctl struct with random stack garbage in the yet unused parts. Which
   then bakes in the ABI that those fields can never be used for anything else
   but garbage.

 * Have simple testcases for all of the above.


Fun with Error Paths
--------------------

Nowadays we don't have any excuse left any more for drm drivers being neat
little root exploits. This means we both need full input validation and solid
error handling paths - GPUs will die eventually in the oddmost corner cases
anyway:

 * The ioctl must check for array overflows. Also it needs to check for
   over/underflows and clamping issues of integer values in general. The usual
   example is sprite positioning values fed directly into the hardware with the
   hardware just having 12 bits or so. Works nicely until some odd display
   server doesn't bother with clamping itself and the cursor wraps around the
   screen.

 * Have simple testcases for every input validation failure case in your ioctl.
   Check that the error code matches your expectations. And finally make sure
   that you only test for one single error path in each subtest by submitting
   otherwise perfectly valid data. Without this an earlier check might reject
   the ioctl already and shadow the codepath you actually want to test, hiding
   bugs and regressions.

 * Make all your ioctls restartable. First X really loves signals and second
   this will allow you to test 90% of all error handling paths by just
   interrupting your main test suite constantly with signals. Thanks to X's
   love for signal you'll get an excellent base coverage of all your error
   paths pretty much for free for graphics drivers. Also, be consistent with
   how you handle ioctl restarting - e.g. drm has a tiny drmIoctl helper in its
   userspace library. The i915 driver botched this with the set_tiling ioctl,
   now we're stuck forever with some arcane semantics in both the kernel and
   userspace.

 * If you can't make a given codepath restartable make a stuck task at least
   killable. GPUs just die and your users won't like you more if you hang their
   entire box (by means of an unkillable X process). If the state recovery is
   still too tricky have a timeout or hangcheck safety net as a last-ditch
   effort in case the hardware has gone bananas.

 * Have testcases for the really tricky corner cases in your error recovery code
   - it's way too easy to create a deadlock between your hangcheck code and
   waiters.


Time, Waiting and Missing it
----------------------------

GPUs do most everything asynchronously, so we have a need to time operations and
wait for outstanding ones. This is really tricky business; at the moment none of
the ioctls supported by the drm/i915 get this fully right, which means there's
still tons more lessons to learn here.

 * Use CLOCK_MONOTONIC as your reference time, always. It's what alsa, drm and
   v4l use by default nowadays. But let userspace know which timestamps are
   derived from different clock domains like your main system clock (provided
   by the kernel) or some independent hardware counter somewhere else. Clocks
   will mismatch if you look close enough, but if performance measuring tools
   have this information they can at least compensate. If your userspace can
   get at the raw values of some clocks (e.g. through in-command-stream
   performance counter sampling instructions) consider exposing those also.

 * Use __s64 seconds plus __u64 nanoseconds to specify time. It's not the most
   convenient time specification, but it's mostly the standard.

 * Check that input time values are normalized and reject them if not. Note
   that the kernel native struct ktime has a signed integer for both seconds
   and nanoseconds, so beware here.

 * For timeouts, use absolute times. If you're a good fellow and made your
   ioctl restartable relative timeouts tend to be too coarse and can
   indefinitely extend your wait time due to rounding on each restart.
   Especially if your reference clock is something really slow like the display
   frame counter. With a spec lawyer hat on this isn't a bug since timeouts can
   always be extended - but users will surely hate you if their neat animations
   starts to stutter due to this.

 * Consider ditching any synchronous wait ioctls with timeouts and just deliver
   an asynchronous event on a pollable file descriptor. It fits much better
   into event driven applications' main loop.

 * Have testcases for corner-cases, especially whether the return values for
   already-completed events, successful waits and timed-out waits are all sane
   and suiting to your needs.


Leaking Resources, Not
----------------------

A full-blown drm driver essentially implements a little OS, but specialized to
the given GPU platforms. This means a driver needs to expose tons of handles
for different objects and other resources to userspace. Doing that right
entails its own little set of pitfalls:

 * Always attach the lifetime of your dynamically created resources to the
   lifetime of a file descriptor. Consider using a 1:1 mapping if your resource
   needs to be shared across processes -  fd-passing over unix domain sockets
   also simplifies lifetime management for userspace.

 * Always have O_CLOEXEC support.

 * Ensure that you have sufficient insulation between different clients. By
   default pick a private per-fd namespace which forces any sharing to be done
   explicitly. Only go with a more global per-device namespace if the objects
   are truly device-unique. One counterexample in the drm modeset interfaces is
   that the per-device modeset objects like connectors share a namespace with
   framebuffer objects, which mostly are not shared at all. A separate
   namespace, private by default, for framebuffers would have been more
   suitable.

 * Think about uniqueness requirements for userspace handles. E.g. for most drm
   drivers it's a userspace bug to submit the same object twice in the same
   command submission ioctl. But then if objects are shareable userspace needs
   to know whether it has seen an imported object from a different process
   already or not. I haven't tried this myself yet due to lack of a new class
   of objects, but consider using inode numbers on your shared file descriptors
   as unique identifiers - it's how real files are told apart, too.
   Unfortunately this requires a full-blown virtual filesystem in the kernel.


Last, but not Least
-------------------

Not every problem needs a new ioctl:

 * Think hard whether you really want a driver-private interface. Of course
   it's much quicker to push a driver-private interface than engaging in
   lengthy discussions for a more generic solution. And occasionally doing a
   private interface to spearhead a new concept is what's required. But in the
   end, once the generic interface comes around you'll end up maintainer two
   interfaces. Indefinitely.

 * Consider other interfaces than ioctls. A sysfs attribute is much better for
   per-device settings, or for child objects with fairly static lifetimes (like
   output connectors in drm with all the detection override attributes). Or
   maybe only your testsuite needs this interface, and then debugfs with its
   disclaimer of not having a stable ABI would be better.

Finally, the name of the game is to get it right on the first attempt, since if
your driver proves popular and your hardware platforms long-lived then you'll
be stuck with a given ioctl essentially forever. You can try to deprecate
horrible ioctls on newer iterations of your hardware, but generally it takes
years to accomplish this. And then again years until the last user able to
complain about regressions disappears, too.