neiam /lambent-flex-c6

public · Issues · Pulls · Labels · Forks · Compare · Actions success · Badges · Packages

Pin the nightly, and keep the toolchain off the image layer

478e173 · Gabriel Morell · 2026-09-16 18:46

6 files +68 -22

Message

{commit_body(@commit)}

Files changed

modified .forgejo/workflows/firmware.yml

+28 −13

@@ -101,13 +101,15 @@ jobs:

101	101	set -euo pipefail
102	102
103	103	# Espressif's own Rust image, pinned to a version rather than
104		− # `esp32c6_latest`. The pin is load-bearing twice over: it fixes the
105		− # ESP-IDF-capable toolchain, and it fixes the nightly the image
106		− # installed when it was built. `rust-toolchain.toml` asks for
107		− # "nightly", rustup uses whatever nightly is already installed rather
108		− # than fetching a newer one, and a current nightly does not build
109		− # `std` for this target at all (see CLAUDE.md, failure 1 —
110		− # `AT_FDCWD`).
	104	+ # `esp32c6_latest`, for the ESP-IDF-capable toolchain it carries:
	105	+ # rustup, ldproxy and espflash.
	106	+ #
	107	+ # It is NOT what pins the compiler — rust-toolchain.toml is. Run #5
	108	+ # is why that distinction is written down: the image's `nightly` was
	109	+ # missing rust-src, so rustup resynced the channel to satisfy the
	110	+ # toolchain file and pulled the current nightly, which cannot build
	111	+ # `std` for this target at all (CLAUDE.md, failure 1 — `AT_FDCWD`).
	112	+ # A dated channel has nothing to resync to.
111	113	#
112	114	# Kept next to the command that uses it. Run #4 lost the image
113	115	# argument from the `docker run` below in an edit, podman quietly

@@ -183,18 +185,25 @@ jobs:

183	185	# and the image's `esp` user need not agree, and root can write
184	186	# either way.
185	187	#
186		− # HOME points inside the workspace so ~/.espressif and ~/.cargo are
187		− # cacheable; RUSTUP_HOME stays in the image, where the pinned
188		− # nightly already lives — pointing it at the workspace would make
189		− # rustup fetch a current nightly, which is the one thing this pin
190		− # exists to prevent.
	188	+ # HOME, CARGO_HOME and RUSTUP_HOME all point inside the workspace,
	189	+ # so they are cacheable — and so that nothing writes into the image.
	190	+ # Run #5 tried the image's own /home/esp/.rustup and rustup failed
	191	+ # with "Invalid cross-device link (os error 18)": renaming a file out
	192	+ # of a lower overlayfs layer is EXDEV under rootless podman.
	193	+ #
	194	+ # The toolchain that lands there is pinned by rust-toolchain.toml to
	195	+ # a dated nightly. Run #5 also showed why that matters: the image's
	196	+ # `nightly` was missing rust-src, so rustup went to satisfy the
	197	+ # toolchain file and resynced the channel — pulling today\'s
	198	+ # nightly, the one that cannot build std for this target at all.
	199	+ # A dated channel has nothing to resync to.
191	200	docker run --rm \
192	201	--user 0:0 \
193	202	"${mount[@]}" \
194	203	-w "$PWD/lambent-flex-c6" \
195	204	-e HOME="$PWD/.home" \
196	205	-e CARGO_HOME="$PWD/.home/.cargo" \
197		− -e RUSTUP_HOME=/home/esp/.rustup \
	206	+ -e RUSTUP_HOME="$PWD/.home/.rustup" \
198	207	-e LAMBENT_FLEX_VERSION="$VERSION" \
199	208	-e HOST_UID="$(id -u)" \
200	209	-e HOST_GID="$(id -g)" \

@@ -217,6 +226,12 @@ jobs:

217	226
218	227	echo "container: ulimit-u=$(ulimit -u) pids.max=$(cat /sys/fs/cgroup/pids.max 2>/dev/null \|\| echo n/a) jobs=$CARGO_BUILD_JOBS"
219	228
	229	+ # Install it explicitly, so a toolchain problem reports itself
	230	+ # here rather than as a confusing failure inside cargo. Reads
	231	+ # the channel from rust-toolchain.toml; idempotent once cached.
	232	+ rustup show active-toolchain \|\| rustup toolchain install --profile minimal
	233	+ rustup show
	234	+
220	235	cargo build --release
221	236
222	237	mkdir -p dist

modified CLAUDE.md

+19 −6

@@ -81,7 +81,8 @@ cargo build --release

81	81	cargo run --release # runner = ./flash.sh (espflash + the OTA layout)
82	82	```
83	83
84		−Needs nightly + `rust-src` (see `rust-toolchain.toml`) for `build-std`, plus
	84	+Needs the pinned nightly + `rust-src` (see `rust-toolchain.toml`; the pin is
	85	+load-bearing — failure 1 below) for `build-std`, plus
85	86	`ldproxy` and `espflash` on PATH. `~/.espressif` holds ESP-IDF v5.1.1 and
86	87	v5.2.3 side by side; v5.2.3 is the one in use.
87	88

@@ -185,11 +186,23 @@ five of these were hit in sequence; the first is still a live local patch.

185	186	says ESP-IDF should skip `O_NOFOLLOW`, so the intended behaviour is plain
186	187	`chmod`.
187	188
188		− Worked around by patching the rustup toolchain in place, adding an
189		− `any(target_os = "espidf", target_os = "horizon")` arm that calls
190		− `set_perm`. Backup at `unix.rs.orig` beside it. **`rustup update` silently
191		− reverts this and the build breaks again** — if `AT_FDCWD` reappears, that's
192		− why. Not fixed by the dependency bump; it is purely a `std` problem.
	189	+ ~~Worked around by patching the rustup toolchain in place~~ —
	190	+ superseded on 2026-09-16 by pinning the nightly. `rust-toolchain.toml`
	191	+ now says `channel = "nightly-2026-04-27"`, whose `unix.rs` has no
	192	+ `set_perm_nofollow` in it at all (checked against the published `rust-src`
	193	+ for that date, not assumed). So there is nothing to patch, `rustup update`
	194	+ cannot break the build any more, and the same compiler builds it on a
	195	+ laptop and in CI — it is also the nightly the CI image ships.
	196	+
	197	+ The old workaround was an in-place edit of the toolchain's `unix.rs`
	198	+ adding an `any(target_os = "espidf", target_os = "horizon")` arm calling
	199	+ `set_perm`, with a backup at `unix.rs.orig` beside it. Harmless to leave;
	200	+ nothing reads that toolchain now.
	201	+
	202	+ Moving the pin means re-checking this. Take the date's `rust-src`
	203	+ tarball and grep `library/std/src/sys/fs/unix.rs` for `set_perm_nofollow`
	204	+ before trusting a newer nightly. It is purely a `std` problem — the
	205	+ dependency bump had nothing to do with it.
193	206
194	207	2. ~40 bindgen types emitted as opaque `{ _address: u8 }`, including
195	208	`spi_transaction_t`, `struct stat`, `netif`, `esp_netif_config`,

modified README.md

+3 −2

@@ -126,8 +126,9 @@ cargo build --release

126	126	cargo run --release # -> flash.sh: espflash with the OTA layout
127	127	```
128	128
129		−Requires the nightly toolchain with `rust-src` (see `rust-toolchain.toml`) for
130		−`build-std`, plus `ldproxy` and `espflash`.
	129	+Requires the pinned nightly with `rust-src` (`rust-toolchain.toml` — rustup
	130	+installs it for you), plus `ldproxy` and `espflash`. The pin is deliberate:
	131	+current nightlies do not build `std` for this target.
131	132
132	133	`cargo run` goes through `flash.sh`, which adds three things OTA needs: the
133	134	two-slot `partitions.csv`, our own bootloader (rollback is a bootloader

modified rust-toolchain.toml

+18 −1

@@ -1,3 +1,20 @@

1	1	[toolchain]
2		−channel = "nightly"
	2	+# Pinned, not "nightly".
	3	+#
	4	+# `build-std` needs a nightly, and an unpinned one means the build depends on
	5	+# whatever was published this morning. That is not hypothetical here: the
	6	+# current nightly's `set_perm_nofollow` in library/std/src/sys/fs/unix.rs calls
	7	+# `libc::fchmodat(libc::AT_FDCWD, …)` on a `_` arm that espidf falls through
	8	+# to, and the espidf libc module defines neither symbol — so `std` does not
	9	+# compile at all (CLAUDE.md, failure 1). This date predates that: its
	10	+# unix.rs has no `set_perm_nofollow` in it, verified against the published
	11	+# rust-src for 2026-04-27 rather than assumed.
	12	+#
	13	+# It is also the nightly `espressif/idf-rust:esp32c6_1.95.0.0` ships, which is
	14	+# what CI builds in — so the pin makes a laptop and CI agree instead of merely
	15	+# being reproducible on their own.
	16	+#
	17	+# Moving this means re-checking that `std` still builds for a RISC-V espidf
	18	+# target. `rustup update` no longer silently decides that for you.
	19	+channel = "nightly-2026-04-27"
3	20	components = ["rust-src"]

Parents: 5614525