fixes for --no-buckd

[martinpitt]

I am currently experimenting with --no-buckd mode. This is currently broken, and buck itself needs some fixes for that which we need to pull into our fork if we want to go ahead with this: facebook/buck2#1551

This series fixes a conceptual problem with long-running processes through buck run, see the individual commits for details. We don't really need that if we decide to keep daemon mode, but OTOH this is relatively small and conceptually nice anyway.

With this series, I can patch ci.sh to always use daemon-less mode, I had a watch pgrep -af [b]uck2d running in parallel:

--- tools/ci.sh
+++ tools/ci.sh
@@ -6,7 +6,7 @@
 # The full CI pipeline. Arguments select groups, e.g. `tools/ci.sh check build`
 set -Eeuo pipefail
 cd "$(dirname "$0")/.."
-buck=(bin/tine buck)
+buck=(bin/tine buck --no-buckd)
 only_groups=("$@")
 
 # GitHub folds each group into a log group; an interactive terminal gets a bold banner; anything else
@@ -134,7 +134,7 @@ secureboot_pkcs11() {
     "${buck[@]}" build tine//tools:swtpm-signing.box
 
     pkcs11_dir=$(mktemp -d)
-    bin/tine exec tine//tools:signing-server -- "$pkcs11_dir" > "$pkcs11_dir/log" 2>&1 &
+    bin/tine exec tine//tools:signing-server --no-buckd -- "$pkcs11_dir" > "$pkcs11_dir/log" 2>&1 &
     server_pid=$!
     trap pkcs11_cleanup EXIT
     until [ -S "$pkcs11_dir/sock/pkcs11" ]; do
@@ -204,4 +204,4 @@ group build             -- whole_cell build
 group image-tests       -- whole_cell test --include image "${vm_filter[@]}"
 group secureboot-pkcs11 -- secureboot_pkcs11
 # The shared cache with a real Buck on both ends; in its own isolation dir.
-group remote-cache      -- "${buck[@]}" run tine//tests:cache-roundtrip
+#group remote-cache      -- "${buck[@]}" run tine//tests:cache-roundtrip

This needs the patched buck from the above PR. I built that locally and told tine to use it:

❱❱❱ cat tine.local.toml 
[buck2.platforms.Linux-x86_64]
sha256 = "b532412119f41dff0d35bd618cbf6d6036f5d2345003ef48b0bc3bbc66289e5f"

(note that this breaks the remote-cache test group, as that wants to write its own tine.local.toml).

I had claude research how much it would cost to use daemon-less mode by default, and how much we can lose in both tine and our buck2 fork. My main motivation is to drop the inotify shenanigans, which are a gift that keeps on giving. The detailed plan: plan-tine-daemonless.md

TL/DR:

• -700 lines in our buck fork (all the inotify patches), but + ~ 50 to add a "null" file watcher which --no-buckd would use by default.
• biggest caveat: parallel bin/tine build invocations break. What the daemon does is to serialize these requests, with daemonless they'd step on each other, so tine needs to do the locking itself (flock on buck-out).
• tine.py would not be a net shrinkage: we can lose a few dozen lines of daemon management, but the flock stuff brings back about an equal amount.
• the main advantage is robustness with dropping inotify, the tine complexity does not really change.

Verdict/personal recommendation: this PR is cheap enough that we should land it to make --no-buckd functional for people who use it explicitly. Going the full way is not the big win that we hoped for, but might still be worth it to avoid the trouble we ran into yesterday with Rodrigo. Let's discuss it, but not do it in a whim.

[martinpitt]

facebook/buck2#1551 updated. With that, we don't need this any more.