Re-derive the API 37 emulator failure, and harden the local sweep #47
Merged
JMR-dev
merged 7 commits from 2026-08-23 05:21:26 +00:00
tools/api-37-emulator into main
7
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 JMR-dev
|
f8e6bfa2a3 | Merge remote-tracking branch 'origin/main' into tools/api-37-emulator | ||
|
JMR-dev
|
7ae660ee37 | Merge remote-tracking branch 'origin/main' into tools/api-37-emulator | ||
 JMR-devandClaude Opus 5
|
775a44753b |
Say that the default sweep is red on purpose, and narrow two claims
R16 / #25 -- the branch put API 37 into the default APIS list, where it is permanently two
failures short of green, so a bare `run-e2e.sh` exits 1 by design and nothing said so.
Somebody running it from habit, a wrapper or a hook gets a red exit forever and either
stops reading exit codes or debugs a normal state.
Documented rather than suppressed. The script's own comment already argued that an
expected-red level belongs in the exit code -- reversing that is the branch owner's call,
not a correction -- and the review's alternative needs an exact-set comparison of the
failing test names before it can subtract 37's contribution, which is a new mechanism that
cannot be validated without a device. So:
- the header now states the exit code (0 all green / 1 any level red / 2 refused to
start), says a bare run is 1 by design and why, and gives `run-e2e.sh 33 34 35 36` as
the sweep that can be green;
- a red sweep prints one note after the summary saying the same thing, because the exit
code is read in the terminal and not in the docs -- but ONLY when 37.x is the only level
that went red. `overall` is set by any red level, so a note keyed on "37 was in the
list" would have called a genuine API 34 failure "by design", which is the defect this
is meant to prevent, one layer up. mark_red records which level it was, where the loop
already knows;
- docs/local-emulator.md says it where the default is documented.
R27 / #36 --
|
||
|
JMR-devandClaude Opus 5
|
961cfa72a2 |
Derive the suite size instead of writing it down in two documents
R4 / #13 and R20 / #29 are one defect: an absolute test total in an unregenerated document, written the same day it went stale. This branch was cut at |
||
|
JMR-devandClaude Opus 5
|
da6f2807e9 |
Stop an interrupted sweep leaking the emulator, the AVD and the port
Four corrections to the harness, none of which changes what a successful sweep does. R17 / #26 -- no trap. Ctrl-C during a sweep (now up to five boots long) left headless qemu on console port 5560 and an lmc_e2e_apiNN AVD behind. The next run's `emulator -port` then collides with the orphan and `emu_adb` can resolve to it -- on a workstation with the Pixel plugged in, exactly the ambiguity the ANDROID_SERIAL pinning exists to prevent. `cleanup` (stop_emulator + delete_created_avds, KEEP_AVD honoured) is now on EXIT, INT and TERM. It is idempotent and the normal path calls it explicitly before the summary, so cleanup output cannot land after the summary and the EXIT trap finds nothing to redo. The interrupt path passes a 6-second grace rather than 30: Ctrl-C has already reached the emulator through the foreground process group, so that wait is only for it to finish writing, and `kill -9` follows regardless. `exit "$overall"` stays the last line, so the exit code an EXIT trap could have swallowed is still the one that escapes. The emulator logs in $LOG_DIR are deliberately kept -- they are the only evidence a failed boot leaves. R31 / #40 -- `kill -9 "${EMU_PID:-0}"`. EMU_PID is empty, not unset, if the background launch never produced a job, so `:-0` converted "nothing to kill" into pid 0, which POSIX reads as the sender's whole process group. The `kill -0` wait loop had the same shape and would have spent its full grace period testing the group. All three sites now take a bare `$EMU_PID` behind one `[ -n ... ] || return 0` guard. boot_emulator's own `kill -0` is left alone: it runs only after the assignment and cannot reach the group form. R33 / #42 -- ensure_avd wrote CI's RAM and disk pins to a hardcoded $HOME/.android/avd/... path and checked nothing. With ANDROID_AVD_HOME (or ANDROID_USER_HOME, or ANDROID_SDK_HOME) set, the sed failed and the level ran on at default RAM and userdata, which surfaces much later as "not enough space" and reads as a device problem. `avd_config_path` now looks in every directory avdmanager honours -- no precedence is asserted, the existence check decides -- and a level that cannot be found or written fails instead of running unpinned. R34 / #43 -- disable_region_sampling's "one blocking wait on the device" did not wait: `adb shell stop` does not clear sys.boot_completed, so the property still read 1 and the loop returned at once. Deleted, and the comment now names the service-check loop below it as the actual wait -- which polls the better thing anyway, since `Can't find service: package` is the failure it exists to prevent. That loop also says so when it gives up after 150 s instead of proceeding silently. Deliberately not doing the `setprop sys.boot_completed 0` variant: the loop tested for an empty value, so a 0 would not have made it wait either, and the `!= 1` form it would need is an unbounded loop inside `adb shell` with no timeout. Checked with `bash -n` and with two stub harnesses in place of a device (shellcheck is not installed here): one drives the extracted lifecycle functions against fake binaries and asserts pid 0 really does hit the sender's process group, that an empty EMU_PID now signals nothing and returns at once, that SIGINT cleans up once and exits 130 within seconds, that KEEP_AVD survives the trap path, and that an explicit exit status survives the EXIT trap; the other runs the real script end to end on the boot-failure path, which stops short of e2e-run.sh, and checks the pins land in config.ini, the created AVD is removed, a misplaced config.ini fails the level, and `set -u` is not tripped anywhere. No emulator was booted. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
JMR-devandClaude Opus 5
|
792286a2d7 |
Stop three claims in the API 37 doc outrunning their evidence
Three corrections, all narrowing: - angle_indirect and swangle_indirect are not two independent renderers here. Both logged gles_mode_selected:swangle with the same adapter, differing only in the Vulkan backend underneath -- unlike at API 33-36, where angle_indirect resolves to ANGLE on llvmpipe. What is 7-for-7 is the host-GLES-versus-not split, not "two renderers agree". - "Disabling SystemUI stops the crashes entirely" was one 180-second measurement on a device that had been up twelve minutes. The harness path reproduces a rate collapse, not a zero: its own quiet check printed 1 abort in 45 s and 4 across the run. A 47-second Gradle run survives that; a five-minute one might not. - "Reproduced twice" conflated two routes. The 49/2/0/2 came back from a hand-driven sequence and from the harness, which corroborates the numbers, but the harness path itself has one green measurement. Also records what the doc never said: from 37.1 onward Google ships only 16 KB-page x86_64 images, so page-size alignment is a prerequisite for that path rather than a detail. All 20 libraries in the committed FFmpeg AAR are 0x4000-aligned, checked before the first ps16k boot -- which is why 37.1 reproducing the abort means the gralloc bug and not a page-size mismatch. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |
||
|
JMR-devandClaude Opus 5
|
739bffa5a0 |
Re-derive the API 37 emulator failure: it is the renderer, not the image
docs/api-37-emulator-crash.md claimed "Both swiftshader_indirect and host crash... The crash is in the gralloc mapper, below the renderer." Re-measured, seven runs, one variable each: that is wrong. The mapper is below the renderer, but whether its bad path is reached is not. -gpu host gles_mode_selected:host never boots (57-71 aborts, looping) -gpu swangle_indirect gles_mode_selected:swangle boots, 85 s (1 abort) -gpu angle_indirect gles_mode_selected:swangle boots, 112 s (2 aborts) The old claim rested on two samples of two different things, neither of them ANGLE: the local swiftshader_indirect sample was void, because on this host every SwiftShader-GLES launch segfaults the emulator before the guest matters (the execheap bug in docs/local-emulator.md, not understood when that file was written), and the CI sample was a single swiftshader_indirect run. Also re-derived, and null: android-37.1 rev 8 -- a stable REL image the doc's own "new image revision" trigger was too narrow to catch -- fails identically; -feature -GLDMA,-GLDMA2,-GLDirectMem is accepted and changes nothing; the image's advancedFeatures.ini is byte-identical to API 36's but for one camera line; and there is still no ATD image above API 36. The mechanism, end to end: SystemUI registers a nav-bar luma-sampling listener, SurfaceFlinger's RegionSamplingThread locks a GraphicBuffer, Gralloc5 routes into GoldfishMapper::readFromHost, which asserts, and init SIGKILLs zygote in response -- so the framework restarts under the test run. Disabling SystemUI removes the listener and the aborts stop dead: 0 in 180 s, against 10-11 per 150 s. So run-e2e.sh now covers API 37: renderer chosen per level (33-36 need host, 37 must not have it), dotted image labels, SystemUI disabled followed by a deliberate stop/start, and an abort count printed on every 37 row. The result is 49 tests, 2 failures, 0 errors, 2 skipped, reproduced twice. The two failures are Media3EngineTest on c2.goldfish.h264.decoder; API 35 under the identical renderer is 49/0/0/2 green, so they are the image and not the renderer. CI's matrix should still stop at 36, for reasons now written down rather than assumed. CLAUDE.md is left alone; a replacement bullet is proposed in the doc. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> |