projects/linux·blob master · refs

477 lines · 18.5 KiB

1.. SPDX-License-Identifier: GPL-2.0-only
2
3============
4UAPI Checker
5============
6
7The UAPI checker (``scripts/check-uapi.sh``) is a shell script which
8checks UAPI header files for userspace backwards-compatibility across
9the git tree.
10
11Options
12=======
13
14This section will describe the options with which ``check-uapi.sh``
15can be run.
16
17Usage::
18
19 check-uapi.sh [-b BASE_REF] [-p PAST_REF] [-j N] [-l ERROR_LOG] [-i] [-q] [-v]
20
21Available options::
22
23 -b BASE_REF Base git reference to use for comparison. If unspecified or empty,
24 will use any dirty changes in tree to UAPI files. If there are no
25 dirty changes, HEAD will be used.
26 -p PAST_REF Compare BASE_REF to PAST_REF (e.g. -p v6.1). If unspecified or empty,
27 will use BASE_REF^1. Must be an ancestor of BASE_REF. Only headers
28 that exist on PAST_REF will be checked for compatibility.
29 -j JOBS Number of checks to run in parallel (default: number of CPU cores).
30 -l ERROR_LOG Write error log to file (default: no error log is generated).
31 -i Ignore ambiguous changes that may or may not break UAPI compatibility.
32 -q Quiet operation.
33 -v Verbose operation (print more information about each header being checked).
34
35Environmental args::
36
37 ABIDIFF Custom path to abidiff binary
38 CC C compiler (default is "gcc")
39 ARCH Target architecture of C compiler (default is host arch)
40
41Exit codes::
42
43 0) Success
44 1) ABI difference detected
45 2) Prerequisite not met
46
47Examples
48========
49
50Basic Usage
51-----------
52
53First, let's try making a change to a UAPI header file that obviously
54won't break userspace::
55
56 cat << 'EOF' | patch -l -p1
57 --- a/include/uapi/linux/acct.h
58 +++ b/include/uapi/linux/acct.h
59 @@ -21,7 +21,9 @@
60 #include <asm/param.h>
61 #include <asm/byteorder.h>
62
63 -/*
64 +#define FOO
65 +
66 +/*
67 * comp_t is a 16-bit "floating" point number with a 3-bit base 8
68 * exponent and a 13-bit fraction.
69 * comp2_t is 24-bit with 5-bit base 2 exponent and 20 bit fraction
70 diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
71 EOF
72
73Now, let's use the script to validate::
74
75 % ./scripts/check-uapi.sh
76 Installing user-facing UAPI headers from dirty tree... OK
77 Installing user-facing UAPI headers from HEAD... OK
78 Checking changes to UAPI headers between HEAD and dirty tree...
79 All 912 UAPI headers compatible with x86 appear to be backwards compatible
80
81Let's add another change that *might* break userspace::
82
83 cat << 'EOF' | patch -l -p1
84 --- a/include/uapi/linux/bpf.h
85 +++ b/include/uapi/linux/bpf.h
86 @@ -74,7 +74,7 @@ struct bpf_insn {
87 __u8 dst_reg:4; /* dest register */
88 __u8 src_reg:4; /* source register */
89 __s16 off; /* signed offset */
90 - __s32 imm; /* signed immediate constant */
91 + __u32 imm; /* unsigned immediate constant */
92 };
93
94 /* Key of an a BPF_MAP_TYPE_LPM_TRIE entry */
95 EOF
96
97The script will catch this::
98
99 % ./scripts/check-uapi.sh
100 Installing user-facing UAPI headers from dirty tree... OK
101 Installing user-facing UAPI headers from HEAD... OK
102 Checking changes to UAPI headers between HEAD and dirty tree...
103 ==== ABI differences detected in include/linux/bpf.h from HEAD -> dirty tree ====
104 [C] 'struct bpf_insn' changed:
105 type size hasn't changed
106 1 data member change:
107 type of '__s32 imm' changed:
108 typedef name changed from __s32 to __u32 at int-ll64.h:27:1
109 underlying type 'int' changed:
110 type name changed from 'int' to 'unsigned int'
111 type size hasn't changed
112 ==================================================================================
113
114 error - 1/912 UAPI headers compatible with x86 appear _not_ to be backwards compatible
115
116In this case, the script is reporting the type change because it could
117break a userspace program that passes in a negative number. Now, let's
118say you know that no userspace program could possibly be using a negative
119value in ``imm``, so changing to an unsigned type there shouldn't hurt
120anything. You can pass the ``-i`` flag to the script to ignore changes
121in which the userspace backwards compatibility is ambiguous::
122
123 % ./scripts/check-uapi.sh -i
124 Installing user-facing UAPI headers from dirty tree... OK
125 Installing user-facing UAPI headers from HEAD... OK
126 Checking changes to UAPI headers between HEAD and dirty tree...
127 All 912 UAPI headers compatible with x86 appear to be backwards compatible
128
129Now, let's make a similar change that *will* break userspace::
130
131 cat << 'EOF' | patch -l -p1
132 --- a/include/uapi/linux/bpf.h
133 +++ b/include/uapi/linux/bpf.h
134 @@ -71,8 +71,8 @@ enum {
135
136 struct bpf_insn {
137 __u8 code; /* opcode */
138 - __u8 dst_reg:4; /* dest register */
139 __u8 src_reg:4; /* source register */
140 + __u8 dst_reg:4; /* dest register */
141 __s16 off; /* signed offset */
142 __s32 imm; /* signed immediate constant */
143 };
144 EOF
145
146Since we're re-ordering an existing struct member, there's no ambiguity,
147and the script will report the breakage even if you pass ``-i``::
148
149 % ./scripts/check-uapi.sh -i
150 Installing user-facing UAPI headers from dirty tree... OK
151 Installing user-facing UAPI headers from HEAD... OK
152 Checking changes to UAPI headers between HEAD and dirty tree...
153 ==== ABI differences detected in include/linux/bpf.h from HEAD -> dirty tree ====
154 [C] 'struct bpf_insn' changed:
155 type size hasn't changed
156 2 data member changes:
157 '__u8 dst_reg' offset changed from 8 to 12 (in bits) (by +4 bits)
158 '__u8 src_reg' offset changed from 12 to 8 (in bits) (by -4 bits)
159 ==================================================================================
160
161 error - 1/912 UAPI headers compatible with x86 appear _not_ to be backwards compatible
162
163Let's commit the breaking change, then commit the innocuous change::
164
165 % git commit -m 'Breaking UAPI change' include/uapi/linux/bpf.h
166 [detached HEAD f758e574663a] Breaking UAPI change
167 1 file changed, 1 insertion(+), 1 deletion(-)
168 % git commit -m 'Innocuous UAPI change' include/uapi/linux/acct.h
169 [detached HEAD 2e87df769081] Innocuous UAPI change
170 1 file changed, 3 insertions(+), 1 deletion(-)
171
172Now, let's run the script again with no arguments::
173
174 % ./scripts/check-uapi.sh
175 Installing user-facing UAPI headers from HEAD... OK
176 Installing user-facing UAPI headers from HEAD^1... OK
177 Checking changes to UAPI headers between HEAD^1 and HEAD...
178 All 912 UAPI headers compatible with x86 appear to be backwards compatible
179
180It doesn't catch any breaking change because, by default, it only
181compares ``HEAD`` to ``HEAD^1``. The breaking change was committed on
182``HEAD~2``. If we wanted the search scope to go back further, we'd have to
183use the ``-p`` option to pass a different past reference. In this case,
184let's pass ``-p HEAD~2`` to the script so it checks UAPI changes between
185``HEAD~2`` and ``HEAD``::
186
187 % ./scripts/check-uapi.sh -p HEAD~2
188 Installing user-facing UAPI headers from HEAD... OK
189 Installing user-facing UAPI headers from HEAD~2... OK
190 Checking changes to UAPI headers between HEAD~2 and HEAD...
191 ==== ABI differences detected in include/linux/bpf.h from HEAD~2 -> HEAD ====
192 [C] 'struct bpf_insn' changed:
193 type size hasn't changed
194 2 data member changes:
195 '__u8 dst_reg' offset changed from 8 to 12 (in bits) (by +4 bits)
196 '__u8 src_reg' offset changed from 12 to 8 (in bits) (by -4 bits)
197 ==============================================================================
198
199 error - 1/912 UAPI headers compatible with x86 appear _not_ to be backwards compatible
200
201Alternatively, we could have also run with ``-b HEAD~``. This would set the
202base reference to ``HEAD~`` so then the script would compare it to ``HEAD~^1``.
203
204Architecture-specific Headers
205-----------------------------
206
207Consider this change::
208
209 cat << 'EOF' | patch -l -p1
210 --- a/arch/arm64/include/uapi/asm/sigcontext.h
211 +++ b/arch/arm64/include/uapi/asm/sigcontext.h
212 @@ -70,6 +70,7 @@ struct sigcontext {
213 struct _aarch64_ctx {
214 __u32 magic;
215 __u32 size;
216 + __u32 new_var;
217 };
218
219 #define FPSIMD_MAGIC 0x46508001
220 EOF
221
222This is a change to an arm64-specific UAPI header file. In this example, I'm
223running the script from an x86 machine with an x86 compiler, so, by default,
224the script only checks x86-compatible UAPI header files::
225
226 % ./scripts/check-uapi.sh
227 Installing user-facing UAPI headers from dirty tree... OK
228 Installing user-facing UAPI headers from HEAD... OK
229 No changes to UAPI headers were applied between HEAD and dirty tree
230
231With an x86 compiler, we can't check header files in ``arch/arm64``, so the
232script doesn't even try.
233
234If we want to check the header file, we'll have to use an arm64 compiler and
235set ``ARCH`` accordingly::
236
237 % CC=aarch64-linux-gnu-gcc ARCH=arm64 ./scripts/check-uapi.sh
238 Installing user-facing UAPI headers from dirty tree... OK
239 Installing user-facing UAPI headers from HEAD... OK
240 Checking changes to UAPI headers between HEAD and dirty tree...
241 ==== ABI differences detected in include/asm/sigcontext.h from HEAD -> dirty tree ====
242 [C] 'struct _aarch64_ctx' changed:
243 type size changed from 64 to 96 (in bits)
244 1 data member insertion:
245 '__u32 new_var', at offset 64 (in bits) at sigcontext.h:73:1
246 -- snip --
247 [C] 'struct zt_context' changed:
248 type size changed from 128 to 160 (in bits)
249 2 data member changes (1 filtered):
250 '__u16 nregs' offset changed from 64 to 96 (in bits) (by +32 bits)
251 '__u16 __reserved[3]' offset changed from 80 to 112 (in bits) (by +32 bits)
252 =======================================================================================
253
254 error - 1/884 UAPI headers compatible with arm64 appear _not_ to be backwards compatible
255
256We can see with ``ARCH`` and ``CC`` set properly for the file, the ABI
257change is reported properly. Also notice that the total number of UAPI
258header files checked by the script changes. This is because the number
259of headers installed for arm64 platforms is different than x86.
260
261Cross-Dependency Breakages
262--------------------------
263
264Consider this change::
265
266 cat << 'EOF' | patch -l -p1
267 --- a/include/uapi/linux/types.h
268 +++ b/include/uapi/linux/types.h
269 @@ -52,7 +52,7 @@ typedef __u32 __bitwise __wsum;
270 #define __aligned_be64 __be64 __attribute__((aligned(8)))
271 #define __aligned_le64 __le64 __attribute__((aligned(8)))
272
273 -typedef unsigned __bitwise __poll_t;
274 +typedef unsigned short __bitwise __poll_t;
275
276 #endif /* __ASSEMBLY__ */
277 #endif /* _UAPI_LINUX_TYPES_H */
278 EOF
279
280Here, we're changing a ``typedef`` in ``types.h``. This doesn't break
281a UAPI in ``types.h``, but other UAPIs in the tree may break due to
282this change::
283
284 % ./scripts/check-uapi.sh
285 Installing user-facing UAPI headers from dirty tree... OK
286 Installing user-facing UAPI headers from HEAD... OK
287 Checking changes to UAPI headers between HEAD and dirty tree...
288 ==== ABI differences detected in include/linux/eventpoll.h from HEAD -> dirty tree ====
289 [C] 'struct epoll_event' changed:
290 type size changed from 96 to 80 (in bits)
291 2 data member changes:
292 type of '__poll_t events' changed:
293 underlying type 'unsigned int' changed:
294 type name changed from 'unsigned int' to 'unsigned short int'
295 type size changed from 32 to 16 (in bits)
296 '__u64 data' offset changed from 32 to 16 (in bits) (by -16 bits)
297 ========================================================================================
298 include/linux/eventpoll.h did not change between HEAD and dirty tree...
299 It's possible a change to one of the headers it includes caused this error:
300 #include <linux/fcntl.h>
301 #include <linux/types.h>
302
303Note that the script noticed the failing header file did not change,
304so it assumes one of its includes must have caused the breakage. Indeed,
305we can see ``linux/types.h`` is used from ``eventpoll.h``.
306
307UAPI Header Removals
308--------------------
309
310Consider this change::
311
312 cat << 'EOF' | patch -l -p1
313 diff --git a/include/uapi/asm-generic/Kbuild b/include/uapi/asm-generic/Kbuild
314 index ebb180aac74e..a9c88b0a8b3b 100644
315 --- a/include/uapi/asm-generic/Kbuild
316 +++ b/include/uapi/asm-generic/Kbuild
317 @@ -31,6 +31,6 @@ mandatory-y += stat.h
318 mandatory-y += statfs.h
319 mandatory-y += swab.h
320 mandatory-y += termbits.h
321 -mandatory-y += termios.h
322 +#mandatory-y += termios.h
323 mandatory-y += types.h
324 mandatory-y += unistd.h
325 EOF
326
327This script removes a UAPI header file from the install list. Let's run
328the script::
329
330 % ./scripts/check-uapi.sh
331 Installing user-facing UAPI headers from dirty tree... OK
332 Installing user-facing UAPI headers from HEAD... OK
333 Checking changes to UAPI headers between HEAD and dirty tree...
334 ==== UAPI header include/asm/termios.h was removed between HEAD and dirty tree ====
335
336 error - 1/912 UAPI headers compatible with x86 appear _not_ to be backwards compatible
337
338Removing a UAPI header is considered a breaking change, and the script
339will flag it as such.
340
341Checking Historic UAPI Compatibility
342------------------------------------
343
344You can use the ``-b`` and ``-p`` options to examine different chunks of your
345git tree. For example, to check all changed UAPI header files between tags
346v6.0 and v6.1, you'd run::
347
348 % ./scripts/check-uapi.sh -b v6.1 -p v6.0
349 Installing user-facing UAPI headers from v6.1... OK
350 Installing user-facing UAPI headers from v6.0... OK
351 Checking changes to UAPI headers between v6.0 and v6.1...
352
353 --- snip ---
354 error - 37/907 UAPI headers compatible with x86 appear _not_ to be backwards compatible
355
356Note: Before v5.3, a header file needed by the script is not present,
357so the script is unable to check changes before then.
358
359You'll notice that the script detected many UAPI changes that are not
360backwards compatible. Knowing that kernel UAPIs are supposed to be stable
361forever, this is an alarming result. This brings us to the next section:
362caveats.
363
364Caveats
365=======
366
367The UAPI checker makes no assumptions about the author's intention, so some
368types of changes may be flagged even though they intentionally break UAPI.
369
370Removals For Refactoring or Deprecation
371---------------------------------------
372
373Sometimes drivers for very old hardware are removed, such as in this example::
374
375 % ./scripts/check-uapi.sh -b ba47652ba655
376 Installing user-facing UAPI headers from ba47652ba655... OK
377 Installing user-facing UAPI headers from ba47652ba655^1... OK
378 Checking changes to UAPI headers between ba47652ba655^1 and ba47652ba655...
379 ==== UAPI header include/linux/meye.h was removed between ba47652ba655^1 and ba47652ba655 ====
380
381 error - 1/910 UAPI headers compatible with x86 appear _not_ to be backwards compatible
382
383The script will always flag removals (even if they're intentional).
384
385Struct Expansions
386-----------------
387
388Depending on how a structure is handled in kernelspace, a change which
389expands a struct could be non-breaking.
390
391If a struct is used as the argument to an ioctl, then the kernel driver
392must be able to handle ioctl commands of any size. Beyond that, you need
393to be careful when copying data from the user. Say, for example, that
394``struct foo`` is changed like this::
395
396 struct foo {
397 __u64 a; /* added in version 1 */
398 + __u32 b; /* added in version 2 */
399 + __u32 c; /* added in version 2 */
400 }
401
402By default, the script will flag this kind of change for further review::
403
404 [C] 'struct foo' changed:
405 type size changed from 64 to 128 (in bits)
406 2 data member insertions:
407 '__u32 b', at offset 64 (in bits)
408 '__u32 c', at offset 96 (in bits)
409
410However, it is possible that this change was made safely.
411
412If a userspace program was built with version 1, it will think
413``sizeof(struct foo)`` is 8. That size will be encoded in the
414ioctl value that gets sent to the kernel. If the kernel is built
415with version 2, it will think the ``sizeof(struct foo)`` is 16.
416
417The kernel can use the ``_IOC_SIZE`` macro to get the size encoded
418in the ioctl code that the user passed in and then use
419``copy_struct_from_user()`` to safely copy the value::
420
421 int handle_ioctl(unsigned long cmd, unsigned long arg)
422 {
423 switch _IOC_NR(cmd) {
424 0x01: {
425 struct foo my_cmd; /* size 16 in the kernel */
426
427 ret = copy_struct_from_user(&my_cmd, arg, sizeof(struct foo), _IOC_SIZE(cmd));
428 ...
429
430``copy_struct_from_user`` will zero the struct in the kernel and then copy
431only the bytes passed in from the user (leaving new members zeroized).
432If the user passed in a larger struct, the extra members are ignored.
433
434If you know this situation is accounted for in the kernel code, you can
435pass ``-i`` to the script, and struct expansions like this will be ignored.
436
437Flex Array Migration
438--------------------
439
440While the script handles expansion into an existing flex array, it does
441still flag initial migration to flex arrays from 1-element fake flex
442arrays. For example::
443
444 struct foo {
445 __u32 x;
446 - __u32 flex[1]; /* fake flex */
447 + __u32 flex[]; /* real flex */
448 };
449
450This change would be flagged by the script::
451
452 [C] 'struct foo' changed:
453 type size changed from 64 to 32 (in bits)
454 1 data member change:
455 type of '__u32 flex[1]' changed:
456 type name changed from '__u32[1]' to '__u32[]'
457 array type size changed from 32 to 'unknown'
458 array type subrange 1 changed length from 1 to 'unknown'
459
460At this time, there's no way to filter these types of changes, so be
461aware of this possible false positive.
462
463Summary
464-------
465
466While many types of false positives are filtered out by the script,
467it's possible there are some cases where the script flags a change
468which does not break UAPI. It's also possible a change which *does*
469break userspace would not be flagged by this script. While the script
470has been run on much of the kernel history, there could still be corner
471cases that are not accounted for.
472
473The intention is for this script to be used as a quick check for
474maintainers or automated tooling, not as the end-all authority on
475patch compatibility. It's best to remember: use your best judgment
476(and ideally a unit test in userspace) to make sure your UAPI changes
477are backwards-compatible!