• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1# The MB (Meta-Build wrapper) design spec
2
3[TOC]
4
5## Intro
6
7MB is intended to address two major aspects of the GYP -> GN transition
8for Chromium:
9
101. "bot toggling" - make it so that we can easily flip a given bot
11   back and forth between GN and GYP.
12
132. "bot configuration" - provide a single source of truth for all of
14   the different configurations (os/arch/`gyp_define` combinations) of
15   Chromium that are supported.
16
17MB must handle at least the `gen` and `analyze` steps on the bots, i.e.,
18we need to wrap both the `gyp_chromium` invocation to generate the
19Ninja files, and the `analyze` step that takes a list of modified files
20and a list of targets to build and returns which targets are affected by
21the files.
22
23For more information on how to actually use MB, see
24[the user guide](user_guide.md).
25
26## Design
27
28MB is intended to be as simple as possible, and to defer as much work as
29possible to GN or GYP. It should live as a very simple Python wrapper
30that offers little in the way of surprises.
31
32### Command line
33
34It is structured as a single binary that supports a list of subcommands:
35
36* `mb gen -c linux_rel_bot //out/Release`
37* `mb analyze -m tryserver.chromium.linux -b linux_rel /tmp/input.json /tmp/output.json`
38
39### Configurations
40
41`mb` will first look for a bot config file in a set of different locations
42(initially just in //ios/build/bots). Bot config files are JSON files that
43contain keys for 'GYP_DEFINES' (a list of strings that will be joined together
44with spaces and passed to GYP, or a dict that will be similarly converted),
45'gn_args' (a list of strings that will be joined together), and an
46'mb_type' field that says whether to use GN or GYP. Bot config files
47require the full list of settings to be given explicitly.
48
49If no matching bot config file is found, `mb` looks in the
50`//tools/mb/mb_config.pyl` config file to determine whether to use GYP or GN
51for a particular build directory, and what set of flags (`GYP_DEFINES` or `gn
52args`) to use.
53
54A config can either be specified directly (useful for testing) or by specifying
55the master name and builder name (useful on the bots so that they do not need
56to specify a config directly and can be hidden from the details).
57
58See the [user guide](user_guide.md#mb_config.pyl) for details.
59
60### Handling the analyze step
61
62The interface to `mb analyze` is described in the
63[user\_guide](user_guide.md#mb_analyze).
64
65The way analyze works can be subtle and complicated (see below).
66
67Since the interface basically mirrors the way the "analyze" step on the bots
68invokes `gyp_chromium` today, when the config is found to be a gyp config,
69the arguments are passed straight through.
70
71It implements the equivalent functionality in GN by calling `gn refs
72[list of files] --type=executable --all --as=output` and filtering the
73output to match the list of targets.
74
75## Analyze
76
77The goal of the `analyze` step is to speed up the cycle time of the try servers
78by only building and running the tests affected by the files in a patch, rather
79than everything that might be out of date. Doing this ends up being tricky.
80
81We start with the following requirements and observations:
82
83* In an ideal (un-resource-constrained) world, we would build and test
84  everything that a patch affected on every patch. This does not
85  necessarily mean that we would build 'all' on every patch (see below).
86
87* In the real world, however, we do not have an infinite number of machines,
88  and try jobs are not infinitely fast, so we need to balance the desire
89  to get maximum test coverage against the desire to have reasonable cycle
90  times, given the number of machines we have.
91
92* Also, since we run most try jobs against tip-of-tree Chromium, by
93  the time one job completes on the bot, new patches have probably landed,
94  rendering the build out of date.
95
96* This means that the next try job may have to do a build that is out of
97  date due to a combination of files affected by a given patch, and files
98  affected for unrelated reasons. We want to rebuild and test only the
99  targets affected by the patch, so that we don't blame or punish the
100  patch author for unrelated changes.
101
102So:
103
1041. We need a way to indicate which changed files we care about and which
105   we don't (the affected files of a patch).
106
1072. We need to know which tests we might potentially want to run, and how
108   those are mapped onto build targets. For some kinds of tests (like
109   GTest-based tests), the mapping is 1:1 - if you want to run base_unittests,
110   you need to build base_unittests. For others (like the telemetry and
111   layout tests), you might need to build several executables in order to
112   run the tests, and that mapping might best be captured by a *meta*
113   target (a GN group or a GYP 'none' target like `webkit_tests`) that
114   depends on the right list of files. Because the GN and GYP files know
115   nothing about test steps, we have to have some way of mapping back
116   and forth between test steps and build targets. That mapping
117   is *not* currently available to MB (or GN or GYP), and so we have to
118   enough information to make it possible for the caller to do the mapping.
119
1203. We might also want to know when test targets are affected by data files
121   that aren't compiled (python scripts, or the layout tests themselves).
122   There's no good way to do this in GYP, but GN supports this.
123
1244. We also want to ensure that particular targets still compile even if they
125   are not actually tested; consider testing the installers themselves, or
126   targets that don't yet have good test coverage. We might want to use meta
127   targets for this purpose as well.
128
1295. However, for some meta targets, we don't necessarily want to rebuild the
130   meta target itself, perhaps just the dependencies of the meta target that
131   are affected by the patch. For example, if you have a meta target like
132   `blink_tests` that might depend on ten different test binaries. If a patch
133   only affects one of them (say `wtf_unittests`), you don't want to
134   build `blink_tests`, because that might actually also build the other nine
135   targets.  In other words, some meta targets are *prunable*.
136
1376. As noted above, in the ideal case we actually have enough resources and
138   things are fast enough that we can afford to build everything affected by a
139   patch, but listing every possible target explicitly would be painful. The
140   GYP and GN Ninja generators provide an 'all' target that captures (nearly,
141   see [crbug.com/503241](crbug.com/503241)) everything, but unfortunately
142   neither GN nor GYP actually represents 'all' as a meta target in the build
143   graph, so we will need to write code to handle that specially.
144
1457. In some cases, we will not be able to correctly analyze the build graph to
146   determine the impact of a patch, and need to bail out (e.g,. if you change a
147   build file itself, it may not be easy to tell how that affects the graph).
148   In that case we should simply build and run everything.
149
150The interaction between 2) and 5) means that we need to treat meta targets
151two different ways, and so we need to know which targets should be
152pruned in the sense of 5) and which targets should be returned unchanged
153so that we can map them back to the appropriate tests.
154
155So, we need three things as input:
156
157* `files`: the list of files in the patch
158* `test_targets`: the list of ninja targets which, if affected by a patch,
159  should be reported back so that we can map them back to the appropriate
160  tests to run. Any meta targets in this list should *not* be pruned.
161* `additional_compile_targets`: the list of ninja targets we wish to compile
162  *in addition to* the list in `test_targets`. Any meta targets
163  present in this list should be pruned (we don't need to return the
164  meta targets because they aren't mapped back to tests, and we don't want
165  to build them because we might build too much).
166
167We can then return two lists as output:
168
169* `compile_targets`, which is a list of pruned targets to be
170  passed to Ninja to build. It is acceptable to replace a list of
171  pruned targets by a meta target if it turns out that all of the
172  dependencies of the target are affected by the patch (i.e.,
173  all ten binaries that blink_tests depends on), but doing so is
174  not required.
175* `test_targets`, which is a list of unpruned targets to be mapped
176  back to determine which tests to run.
177
178There may be substantial overlap between the two lists, but there is
179no guarantee that one is a subset of the other and the two cannot be
180used interchangeably or merged together without losing information and
181causing the wrong thing to happen.
182
183The implementation is responsible for recognizing 'all' as a magic string
184and mapping it onto the list of all root nodes in the build graph.
185
186There may be files listed in the input that don't actually exist in the build
187graph: this could be either the result of an error (the file should be in the
188build graph, but isn't), or perfectly fine (the file doesn't affect the build
189graph at all). We can't tell these two apart, so we should ignore missing
190files.
191
192There may be targets listed in the input that don't exist in the build
193graph; unlike missing files, this can only indicate a configuration error,
194and so we should return which targets are missing so the caller can
195treat this as an error, if so desired.
196
197Any of the three inputs may be an empty list:
198
199* It normally doesn't make sense to call analyze at all if no files
200  were modified, but in rare cases we can hit a race where we try to
201  test a patch after it has already been committed, in which case
202  the list of modified files is empty. We should return 'no dependency'
203  in that case.
204
205* Passing an empty list for one or the other of test_targets and
206  additional_compile_targets is perfectly sensible: in the former case,
207  it can indicate that you don't want to run any tests, and in the latter,
208  it can indicate that you don't want to do build anything else in
209  addition to the test targets.
210
211* It doesn't make sense to call analyze if you don't want to compile
212  anything at all, so passing [] for both test_targets and
213  additional_compile_targets should probably return an error.
214
215In the output case, an empty list indicates that there was nothing to
216build, or that there were no affected test targets as appropriate.
217
218Note that passing no arguments to Ninja is equivalent to passing
219`all` to Ninja (at least given how GN and GYP work); however, we
220don't want to take advantage of this in most cases because we don't
221actually want to build every out of date target, only the targets
222potentially affected by the files. One could try to indicate
223to analyze that we wanted to use no arguments instead of an empty
224list, but using the existing fields for this seems fragile and/or
225confusing, and adding a new field for this seems unwarranted at this time.
226
227There is an "error" field in case something goes wrong (like the
228empty file list case, above, or an internal error in MB/GYP/GN). The
229analyze code should also return an error code to the shell if appropriate
230to indicate that the command failed.
231
232In the case where build files themselves are modified and analyze may
233not be able to determine a correct answer (point 7 above, where we return
234"Found dependency (all)"), we should also return the `test_targets` unmodified
235and return the union of `test_targets` and `additional_compile_targets` for
236`compile_targets`, to avoid confusion.
237
238### Examples
239
240Continuing the example given above, suppose we have the following build
241graph:
242
243* `blink_tests` is a meta target that depends on `webkit_unit_tests`,
244  `wtf_unittests`, and `webkit_tests` and represents all of the targets
245  needed to fully test Blink. Each of those is a separate test step.
246* `webkit_tests` is also a meta target; it depends on `content_shell`
247  and `image_diff`.
248* `base_unittests` is a separate test binary.
249* `wtf_unittests` depends on `Assertions.cpp` and `AssertionsTest.cpp`.
250* `webkit_unit_tests` depends on `WebNode.cpp` and `WebNodeTest.cpp`.
251* `content_shell` depends on `WebNode.cpp` and `Assertions.cpp`.
252* `base_unittests` depends on `logging.cc` and `logging_unittest.cc`.
253
254#### Example 1
255
256We wish to run 'wtf_unittests' and 'webkit_tests' on a bot, but not
257compile any additional targets.
258
259If a patch touches WebNode.cpp, then analyze gets as input:
260
261    {
262      "files": ["WebNode.cpp"],
263      "test_targets": ["wtf_unittests", "webkit_tests"],
264      "additional_compile_targets": []
265    }
266
267and should return as output:
268
269    {
270      "status": "Found dependency",
271      "compile_targets": ["webkit_unit_tests"],
272      "test_targets": ["webkit_tests"]
273    }
274
275Note how `webkit_tests` was pruned in compile_targets but not in test_targets.
276
277#### Example 2
278
279Using the same patch as Example 1, assume we wish to run only `wtf_unittests`,
280but additionally build everything needed to test Blink (`blink_tests`):
281
282We pass as input:
283
284    {
285      "files": ["WebNode.cpp"],
286      "test_targets": ["wtf_unittests"],
287      "additional_compile_targets": ["blink_tests"]
288    }
289
290And should get as output:
291
292    {
293      "status": "Found dependency",
294      "compile_targets": ["webkit_unit_tests"],
295      "test_targets": []
296    }
297
298Here `blink_tests` was pruned in the output compile_targets, and
299test_targets was empty, since blink_tests was not listed in the input
300test_targets.
301
302#### Example 3
303
304Build everything, but do not run any tests.
305
306Input:
307
308    {
309      "files": ["WebNode.cpp"],
310      "test_targets": [],
311      "additional_compile_targets": ["all"]
312    }
313
314Output:
315
316    {
317      "status": "Found dependency",
318      "compile_targets": ["webkit_unit_tests", "content_shell"],
319      "test_targets": []
320    }
321
322#### Example 4
323
324Same as Example 2, but a build file was modified instead of a source file.
325
326Input:
327
328    {
329      "files": ["BUILD.gn"],
330      "test_targets": ["wtf_unittests"],
331      "additional_compile_targets": ["blink_tests"]
332    }
333
334Output:
335
336    {
337      "status": "Found dependency (all)",
338      "compile_targets": ["webkit_unit_tests", "wtf_unittests"],
339      "test_targets": ["wtf_unittests"]
340    }
341
342test_targets was returned unchanged, compile_targets was pruned.
343
344## Random Requirements and Rationale
345
346This section is collection of semi-organized notes on why MB is the way
347it is ...
348
349### in-tree or out-of-tree
350
351The first issue is whether or not this should exist as a script in
352Chromium at all; an alternative would be to simply change the bot
353configurations to know whether to use GYP or GN, and which flags to
354pass.
355
356That would certainly work, but experience over the past two years
357suggests a few things:
358
359  * we should push as much logic as we can into the source repositories
360    so that they can be versioned and changed atomically with changes to
361    the product code; having to coordinate changes between src/ and
362    build/ is at best annoying and can lead to weird errors.
363  * the infra team would really like to move to providing
364    product-independent services (i.e., not have to do one thing for
365    Chromium, another for NaCl, a third for V8, etc.).
366  * we found that during the SVN->GIT migration the ability to flip bot
367    configurations between the two via changes to a file in chromium
368    was very useful.
369
370All of this suggests that the interface between bots and Chromium should
371be a simple one, hiding as much of the chromium logic as possible.
372
373### Why not have MB be smarter about de-duping flags?
374
375This just adds complexity to the MB implementation, and duplicates logic
376that GYP and GN already have to support anyway; in particular, it might
377require MB to know how to parse GYP and GN values. The belief is that
378if MB does *not* do this, it will lead to fewer surprises.
379
380It will not be hard to change this if need be.
381
382### Integration w/ gclient runhooks
383
384On the bots, we will disable `gyp_chromium` as part of runhooks (using
385`GYP_CHROMIUM_NO_ACTION=1`), so that mb shows up as a separate step.
386
387At the moment, we expect most developers to either continue to use
388`gyp_chromium` in runhooks or to disable at as above if they have no
389use for GYP at all. We may revisit how this works once we encourage more
390people to use GN full-time (i.e., we might take `gyp_chromium` out of
391runhooks altogether).
392
393### Config per flag set or config per (os/arch/flag set)?
394
395Currently, mb_config.pyl does not specify the host_os, target_os, host_cpu, or
396target_cpu values for every config that Chromium runs on, it only specifies
397them for when the values need to be explicitly set on the command line.
398
399Instead, we have one config per unique combination of flags only.
400
401In other words, rather than having `linux_rel_bot`, `win_rel_bot`, and
402`mac_rel_bot`, we just have `rel_bot`.
403
404This design allows us to determine easily all of the different sets
405of flags that we need to support, but *not* which flags are used on which
406host/target combinations.
407
408It may be that we should really track the latter. Doing so is just a
409config file change, however.
410
411### Non-goals
412
413* MB is not intended to replace direct invocation of GN or GYP for
414  complicated build scenarios (a.k.a. Chrome OS), where multiple flags need
415  to be set to user-defined paths for specific toolchains (e.g., where
416  Chrome OS needs to specify specific board types and compilers).
417
418* MB is not intended at this time to be something developers use frequently,
419  or to add a lot of features to. We hope to be able to get rid of it once
420  the GYP->GN migration is done, and so we should not add things for
421  developers that can't easily be added to GN itself.
422
423* MB is not intended to replace the
424  [CR tool](https://code.google.com/p/chromium/wiki/CRUserManual). Not
425  only is it only intended to replace the gyp\_chromium part of `'gclient
426  runhooks'`, it is not really meant as a developer-facing tool.
427