the family root's bootstrap.py and siblings.py: under GNU make 4.4 and later, a member's make loses the parallelism it may have had through the holder's fifo: jobserver, since #950 takes every jobserver out of MAKEFLAGS #987

Goal: under the holder's make -jN bootstrap (or validate, park, resume), a member's make runs its jobs together where it can share the holder's jobserver safely, and one at a time, with no warning, everywhere else.

Filed by lane openRepoShape-4 from lane 5's checklist candidate C1 in its review of #950 (#888) at 4758e645, #950 issuecomment-6253417294. Lane 4's round-1 plan said it would be filed (#950 issuecomment-6253442578). Lane 5's re-check found #950 READY at 1e94ce04 (#950 issuecomment-6254597377); this issue is read against that head.

Line

• Under GNU make 4.4 and later, a member's make runs one job at a time where it probably ran them together before the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950.
  • the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950 (the family root's Makefile: under make -jN, every member's make warns jobserver unavailable: using -j1 and runs one job at a time #888) takes the jobserver word and every -j out of the MAKEFLAGS a member's make inherits (member_make_env, in both scripts), wherever MAKEFLAGS names a jobserver. On GNU Make 4.3 (measured) that only removes the jobserver unavailable: using -j1. Add '+' to parent make rule. warning: the member's make ran serially before and still does.
  • GNU make 4.4 and later use a named pipe where mkfifo exists, and write --jobserver-auth=fifo:<path> into MAKEFLAGS. Their NEWS says a sub-make then connects to the jobserver even where it is not marked recursive. So before the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950, a member's make under a 4.4 holder probably opened the holder's fifo by path and ran its jobs together. Now it runs them one at a time. the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's body says so under "Choices for the reviewer" (1 and 2).
  • Not measured: no GNU make 4.4 is on lane 4's host.
  • Why the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950 did not keep fifo:: the member's make is $MAKE or make on PATH, not the make that ran the holder (GNU make exports no MAKE to a recipe; measured). A GNU make 4.2 or 4.3 given 4.4's fifo: word stops with internal error: invalid --jobserver-auth string ..., exit 2 (measured on 4.3).
  • Direction, one of:
    • keep the fifo: word and the -jN beside it where the member's make is known to be GNU make 4.4 or later. That could come from one <make> --version per run, read before the first member (GNU make exports neither MAKE nor, by default, MAKE_VERSION to a recipe). Take them out everywhere else, as now;
    • or document that a member's make runs one job at a time under any -j, 4.4 included, and close this.
  • Whatever is kept: never a pipe's descriptor numbers (3,4), which Popen closes and the holder's rules do not pass; never a -j with no jobserver beside it, which alone means unbounded jobs; and nothing a member's make of another version cannot read.
  • Whatever changes keeps the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's reading of MAKEFLAGS: words split as make splits them, only at a space or a tab, a backslash keeping either in a value, so a value holding U+00A0, U+202F or a newline stays one word (the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950's round 1); and the two scripts' copies of member_make_env held to one answer by tests/test_family_member_make_jobserver.py.

Verify:

• Under GNU make 4.4+ (--jobserver-style=fifo, the default where mkfifo exists), with the holder's make -j2 bootstrap and a member whose bootstrap target has two independent prerequisites that each write a trace: the two run together and no warning is printed.
• Under GNU make 4.3, and under a 4.4 holder whose make on PATH is 4.3: no warning, no invalid --jobserver-auth string, and the member runs serially, as after the family root's bootstrap.py and siblings.py: a member's make inherits no jobserver from the holder's make -jN, so make -j2 bootstrap, validate, park and resume no longer warn jobserver unavailable of a rule the person does not own (#888) #950.
• tests/test_family_member_make_jobserver.py and the trio pass. Each changed pinned file costs every holder one upstream-changed row (FAMILY_COPIED_VERBATIM).

Milestone: none set; lane 3's call. Under 4.4 this is lost parallelism with no warning and no wrong result; no verdict changes.

Lane: openreposhape-4 (openRepoShape-4)