Two related races on recipes/<r>/target/<target>/sysroot, hit by
llvm-native during a -j 20 redbear-full build:
repo: failed to build: Renaming failed from '.../sysroot.tmp' to
'.../sysroot': File exists (os error 17)
llvm-native, llvm-native.dev and llvm-native.runtime are package
variants of one recipe and share that target directory. The cook_jobs
pool runs them as concurrent threads in a single process, so:
1. Shared staging path. Every variant used the same '<sysroot>.tmp', and
create_dir_clean() at the start of the staging block would wipe a
sibling's half-populated tree while it was still extracting pkgars
into it. Staging is now per-call, named from an atomic counter (a pid
would not disambiguate -- the pool is threads, not processes).
2. TOCTOU on the publish. The '!deps_dir.is_dir()' guard is checked long
before the rename, so all variants can pass it and then race; rename(2)
returns EEXIST/ENOTEMPTY once a winner has published a non-empty
directory, and the losers failed the whole recipe. Losing is harmless
-- every variant populates from the same recipe's dep set, so the
winner's tree is equivalent -- so the losers now discard their staging
and use it. The destination is deliberately not removed first, since a
concurrent cook may already be reading it.
vasilito
713131367f