Using anonymous pipes for jobserver support has some advantages:
for example there is nothing on disk that needs to be cleaned up.
However it has many obscure problems, related to the fact that in
order for it to work we need to ensure these resources are properly
passed through to child processes that want to use the jobserver.
At the same time we don't want to pass the pipe to process which
DON'T know about the jobserver.
Other processes can open file descriptors which we then think are
our jobserver, but aren't. And, we open the pipe file descriptors
in blocking mode which doesn't work for all users.
See issues such as SV 57178, SV 57242, and SV 62397
To avoid these issues, use named pipes (on systems where they are
available) instead of anonoymous pipes. This simplifies many things:
we never need to pass open file descriptors to our children; they
can open the jobserver named pipe. We don't need to worry about
recursive vs. non-recursive children. Users don't have to "pass
through" the resources if they are invoking sub-makes. Each child
can open its own file descriptor and set blocking as needed.
The downside is the named pipe exists on disk and so must be cleaned
up when the "top-level" make instance exits.
In order to allow make to continue to be used in build systems where
older versions of GNU make, or other tools that want to use the
jobserver, but don't understand named pipes, introduce a new option
--jobserver-style that allows the user to choose anonymous pipes.
* NEWS: Announce the change and the --jobserver-style option.
* doc/make.1: Add --jobserver-style documentation.
* doc/make.texi (Special Variables): Add missing items to .FEATURES.
(Options Summary): Add --jobserver-style.
(POSIX Jobserver): Named pipes, changes to --jobserver-auth, and the
--jobserver-style option.
(Windows Jobserver): Document --jobserver-style for Windows.
* configure.ac: Check for mkfifo.
* src/config.h-vms.template: Undefined HAVE_MKFIFO.
* src/config.h.W32.template: Ditto.
* src/main.c: Add jobserver-style as a new command line option.
(main): Add jobserver-fifo to .FEATURES if supported. Pass the style
option to jobserver_setup().
* src/os.h (jobserver_setup): Accept a style string option.
* src/posixos.c (enum js_type): Enumeration of the jobserver style.
(js_type): Which style we are currently using.
(fifo_name): The path to the named pipe (if in use).
(jobserver_setup): If no style is given, or "fifo" is given, set up a
named pipe: get a temporary file and use mkfifo() on it, then open it
for reading and writing. If something fails fall back to anonymous
pipes.
(jobserver_parse_auth): Parse jobserver-auth to determine the style.
If we are using a named pipe, open it. If we're using anonymous pipes
ensure they're valid as before.
(jobserver_get_invalid_auth): Don't invalidate the jobserver when
using named pipes.
(jobserver_clear): Clean up memory used for named pipes.
(jobserver_acquire_all): Unlink the named pipe when done.
* src/w32/w32os.c (jobserver_setup): Check the style argument.
* tests/scripts/features/jobserver: Use --jobserver-style to test
the anonymous pipe behavior, and also test named pipe/semaphore
behavior. Check invalid jobserver-style options.
* tests/scripts/functions/shell: Use --jobserver-style to test the
anonymous pipe behavior, and also test named pipe/semaphore
behavior.
Move all the logic on creating temporary files into misc.c, and add
a new function get_tmppath() that returns the pathname to a temporary
file without creating or opening it.
* src/makeint.h: Add a declaration for get_tmppath(). Remove the
template argument from get_tmpfile(): it will compute its own.
* src/main.c (main): Remove the logic for computing templates.
* src/vmsjobs.c (child_execute_job): Ditto.
* src/misc.c (get_tmptemplate): New function to return an allocated
template string for use with various mktemp-style functions.
(get_tmppath): Return an allocated path to a temporary file, but do
not create it. Generally this should be avoided due to TOCTOU issues.
(get_tmpfile): Use get_tmptemplate() to generate a template rather
than using one passed in. If we don't have mkstemp() then use
get_tmppath() to compute the path of a temp file.
Second-expand only the prerequisites of the targets being built.
Defer second-expanding the prerequisites of targets until we need
to decide if they should be built.
* NEWS: Mention the change in behavior.
* doc/make.texi (Secondary Expansion): Document the new behavior.
* src/filedef.h (struct file): Add flag snapped.
(expand_deps): Declare a function to second expand the
prerequisites of a target.
* src/file.c (rehash_file): Merge flag snapped.
(expand_deps): Remove qualifier static. Check flag snapped.
(snap_deps): Remove the loop which performed second expansion for all
targets.
* src/remake.c (update_file_1): Second expand the prerequisites of
the considered target.
* tests/scripts/features/se_explicit: Add tests.
* tests/scripts/features/se_implicit: Ditto.
* tests/scripts/features/se_statpat: Ditto.
Savannah issues such as SV 57242 and SV 62397 show how passing
references to closed file descriptors via the --jobserver-auth option
in MAKEFLAGS can lead to problematic outcomes.
When computing the child environment for a non-recursive shell, add
an extra option to MAKEFLAGS to disable the file descriptors for the
jobserver.
Unfortunately this doesn't modify the value of the make variable
MAKEFLAGS, it only modifies the value of the sub-shell environment
variable MAKEFLAGS. This can lead to confusion if the user is not
considering the distinction.
* src/makeint.h: Publish the jobserver-auth value. Add a global
definition of the name of the command line option.
* src/os.h (jobserver_get_invalid_auth): New function to return a
string invalidating the jobserver-auth option.
* src/w32/w32os.c (jobserver_get_invaid_auth): Implement it. On
Windows we use a semaphore so there's no need to invalidate.
* src/posixos.c (jobserver_parse_auth): If we parse the invalid
auth string, don't set up the jobserver.
(jobserver_get_invalid_auth): Return an invalid option.
* src/variable.h (target_environment): Specify if the target
environment is for a recursive shell or non-recursive shell.
* src/variable.c (target_environment): Move checking for MAKELEVEL
into the loop rather than doing it at the end.
Along with this, check for MAKEFLAGS and MFLAGS, and update them
based on whether we're invoking a recursive or non-recursive child,
and also on whether it's necessary to invalidate the jobserver.
* src/function.c (func_shell_base): Shell functions can never be
recursive to pass 0 to target_environment().
* src/job.c (start_job_command): Specify whether the child is
recursive when calling target_environment().
* src/main.c: Export jobserver_auth. sync_mutex doesn't need to
be exported. Use the global definition for the option name.
* tests/scripts/variables/MAKEFLAGS: Add tests for $MAKEFLAGS.
* src/main.c (main): Don't reset the jobserver if the number of
slots has not changed.
(define_makeflags): Add all normal flags even when ALL is not set.
* tests/scripts/functions/shell: Test invoking make in $(shell ...).
* tests/scripts/variables/MAKEFLAGS: Test the value of MAKEFLAGS in
$(shell ...).
* src/makeint.h (stpcpy): Add missing declaration.
* src/amiga.c (MyExecute): Avoid extra strlen using stpcpy.
* src/function.c (func_shell_base): Ditto.
(func_error): Use memcpy() not strcpy() when we know the length.
* src/job.c (construct_command_argv_internal): Use stpcpy().
* src/main.c (main): Ditto.
(define_makeflags): Ditto.
* src/variable.c (print_target_variables): Use memcpy() when we know
the length.
* src/commands.c (set_file_variables): Use mempcpy().
* src/expand.c (variable_buffer_output): Ditto.
* src/file.c (expand_deps): Ditto.
* src/function.c (abspath): Ditto.
(handle_function): Ditto.
* src/implicit.c (pattern_search): Ditto.
* src/job.c (construct_command_argv_internal): Use mempcpy() and
don't add multiple spaces when there are no shell flags.
* src/main.c (decode_env_switches): Use mempcpy() to simplify.
(define_makeflags): Ditto.
* src/variable.c (selective_vpath_search): Ditto.
We want to know if getloadavg() is declared, not if we have it.
* configure.ac: Add it as a define (why is this not the default?)
* src/job.c: Conditionalize declaration on HAVE_DECL_GETLOADAVG.
* config.ami.template: Rename HAVE_GETLOADAVG to HAVE_DECL_GETLOADAVG.
* config.h-vms.template: Ditto.
* config.h.W32.template: Ditto.
Clean up and add new elements from latest config.h.
The gl subdirectory contains our local versions of gnulib module
implementations, so move m4/acinclude.m4 and m4/dospaths.m4 there.
* gl/modules/make-macros: Create a new module to handle the macros.
* bootstrap.conf: Add the new module.
* configure.ac: Macro invocation is moved to make-macros.
* m4/.gitignore: Delete unnecessary ignore file: m4 is empty.
* .gitignore: Add m4/ as an ignored directory.
The gl subdirectory contains our local versions of gnulib module
implementations, so move fnmatch* and glob* from lib to gl/lib.
* gl/modules/make-glob: Add a proper Files: section.
* lib/.gitignore: Delete unnecessary ignore file: lib is empty.
* .gitignore: Add lib/ as an ignored directory.
This macro is obsolete: no useful system has this problem anymore.
* src/output.c (output_init): Remove reference to SETVBUF_REVERSED.
* src/config.ami.template: Remove undef of SETVBUF_REVERSED.
* src/config.h-vms.template: Ditto.
* src/config.h.W32.template: Ditto.
This was needed when getloadavg required privileged access; in this
case GNU make needed to be installed as a setgid program. But this
hasn't been supported by gnulib getloadavg() since 2011 and systems
are no longer using it, so remove it.
* src/makeint.h (user_access): Remove unnecessary function.
(make_access): Ditto.
(child_access): Ditto.
* src/misc.c: Remove implementations of the *_access() functions.
* src/main.c (main): Remove unneeded call to user_access().
* src/job.c (load_too_high): Remove calls to {make,user}_access().
(exec_command): Remove call to child_access().
* src/remote-cstms.c: Remove calls to these methods. I suppose it
might be possible this is needed and was piggy-backing on the
privileged setting but since that's been broken for a while I doubt
this is needed. If so we can bring back the implementation into
this source file.
* src/config.h.W32.template: Remove GETLOADAVG_PRIVILEGED undef.
* src/config.h-vms.template: Ditto.
* src/config.ami.template: Ditto.
We can no longer pass our mondo-warnings options to the builds,
as these will impact all the code including gnulib code, and this
won't work. Also allow the caller to disable either the build.sh
or makefile invocation, for testing.
* Makefile.am: Allow the caller to reset the path to the make
binary to be tested. Remove c90 test: gnulib doesn't support it.
This pulls in entirely too much stuff we don't need. Instead grab
just the gnulib source file, then include it in src/misc.c.
* bootstrap.conf: Add just the lib/strtol.c source file.
* configure.ac: Check for strtoll.
* src/misc.c: Include strtol.c, with QUAD set, if needed.
This pulls in a metric ton of stuff that we otherwise don't need, just
for a one-liner that we already have a replacement for in src/misc.c.
* bootstrap.conf: Remove mempcpy
* configure.ac: Add mempcpy to AC_CHECK_FUNCS
We can assume that the return type of a signal handler is void.
We can assume that, if sys/time.h exists, it can be included
with time.h.
* bootstrap: Get the latest version
* configure.ac: Require a newer version of autoconf.
Remove unnecessary AC_PROG_CC_C99 (already have AC_PROC_CC).
Remove unnecessary AC_AIX, AC_ISC_POSIX, AC_MINIX.
Remove unnecessary AC_HEADER_STDC, AC_HEADER_TIME, AC_TYPE_SIGNAL.
Use strerror to search for the cposix library.
* src/commands.c (fatal_error_signal): Assume return type is void.
* src/commands.h: Ditto.
* src/job.c: Ditto.
* src/job.h: Ditto.
* src/main.c: Ditto.
* src/makeint.h: Ditto.
Don't bother with TIME_WITH_SYS_TIME.
* src/remote-cstms.c: Check HAVE_SYS_TIME_H.
* src/config.ami.template: Remove RETSIGTYPE.
* src/config.h-vms.template: Ditto.
* src/config.h.W32.template: Ditto.
Remove TIME_WITH_SYS_TIME.
Export all variables, including exported makefile variables, when
invoking a shell for the $(shell ...) function. If we detect a
recursive variable expansion, silently ignore that variable and do
not export it. We do print a debug message.
* NEWS: Announce the potential backward-incompatibility.
* doc/make.texi (Shell Function): Document the export behavior.
* src/main.c (main): Add "shell-export" to .FEATURES.
* src/job.h: New function to free struct childbase.
* src/job.c (free_childbase): Implement it; call from free_child.
* src/function.c (func_shell_base): Use target_environment() to
obtain the proper environment for the shell function.
Use free_childbase() to free memory.
(windows32_openpipe): Don't reset the environment: the caller
already provided a proper PATH variable in envp.
* src/variable.c (target_environment): If we detect a recursive
expansion and we're called from func_shell, ignore the variable.
(sync_Path_environment): Simplify and reduce memory allocation.
* tests/scripts/functions/shell: Add tests for this.
* src/file.c (rehash_file): Fix warning message.
(rehash_file): Fix comment to match the behavior.
* tests/scripts/features/se_explicit: Fix test.
* tests/scripts/features/mult_rules: Add a new test.
Commit f2771aa614 introduced a bug where some switches were left out
of MAKEFLAGS. Instead of resetting switches, get the same results by
filtering out duplicates.
* src/makeint.h: Remove reset_switches.
* src/main.c: (reset_switches): Remove reset_switches.
* (main): Remove call to reset_switches.
* (decode_switches): Filter out duplicate flags.
* src/variable.c: (set_special_var): Remove call to reset_switches.
* tests/scripts/variables/MAKEFLAGS: Verify that duplicate flags are
properly filtered out.
If GNUMAKEFLAGS was not present in the environment when we started,
don't add it.
* src/main.c (main): Don't mess with GNUMAKEFLAGS unless it exists.
* tests/scripts/variables/GNUMAKEFLAGS: Test this behavior.
Reported, with initial patch, by Gisle Vanem <gvanem@online.no>
* src/main.c (find_and_set_default_shell) [W32]: Pass search_token
directly to w32ify: no need to make a copy first. When we need to
construct a path, use snprintf() to be sure we don't overwrite
the locally-allocated buffer.
* src/w32/pathstuff.c (w32ify) [W32]: Use the malloc version of
_fullpath(), followed by strncpy(), to avoid overwriting buffers.
Introduce non-deterministic ordering into goal and prerequisite
traversal to help tease out inconsistent failures that may happen
when running in parallel build mode.
Introduce second order into each dependency chain:
1. Existing order is syntactic order reachable via 'dep->next'
2. New order is shuffled order stored as 'dep->shuf' in each 'dep'
When updating goals and prerequisites and '--shuffle' is provided,
use the shuffled order to walk the graph. When automatic variable
are set always use the syntactic order of parameters.
* Makefile.am: Add new src/shuffle.c and src/shuffle.h file.
* build_w32.bat: Ditto.
* builddos.bat: Ditto.
* makefile.com: Ditto.
* po/POTFILES.in: Ditto.
* doc/make.texi: Add documentation for --shuffle.
* doc/make.1: Ditto.
* src/dep.h (DEP): Add the shuf pointer.
* src/filedef.h (struct file): Add was_shuffled flag.
* src/main.c: (shuffle_mode): Global flag for the shuffle mode.
(usage): Add the --shuffle option.
(switches): Ditto.
(main): Set shuffle_mode based on the command line parameter.
Reshuffle prerequisites if requested.
* src/remake.c (update_goal_chain): Walk the shuffled list if enabled.
(update_file_1): Ditto.
* src/shuffle.h: Provide an interface for shuffling prerequisites.
* src/shuffle.c: Implement option parsing and prerequisite shuffling.
* tests/scripts/options/shuffle: Test shuffle option and modes.
* src/makeint.h: Declare it.
* src/arscan.c (ar_scan): Replace atoi() calls with make_toui().
Modify some integral types to be more correct.
* src/job.c (load_too_high): Replace atoi() calls with make_toui().
* src/main.c (main): Ditto.
(decode_switches): Ditto.
memcmp() is always faster than strcmp(). In places where we already
know that both buffers have sufficient size, replace strcmp() with
memcmp().
* src/main.c (main): Replace strncmp() with memcmp()==0.
* src/read.c (word1eq): Ditto.
* src/commands.c (set_file_variables): Ditto.
* src/function.c (func_filter_filterout): Ditto.
(a_word_hash_cmp): Use STRING_N_COMPARE since we know the length.
(func_sort): Replace strcmp() with memcmp().
The standard requires that tolower() returns its argument if there's
no valid lowercase conversion: it's not necessary to check isupper()
before invoking tolower().
Comparing the performance of the old ISTRING_HASH to the new one
on all the files present in GLIBC.
Data Type Old Time New Time
Alternating case 24985.754 13883.422
random case 35211.777 13569.051
all lower 18818.974 13706.645
all upper 38859.454 13506.315
* src/hash.h (ISTRING_HASH_1): Omit isupper() check.
(ISTRING_HASH_2): Ditto.
During second expansion of pattern rules only the first pattern in
each "group" was being substituted. E.g. in this makefile:
.SECONDEXPANSION:
all: hello.x
%.x: $$(wordlist 1, 99, %.1 %.%.2) ; $(info $@ from $^)
hello.1 hello.\%.2 \%.1 \%.\%.2: ;
the output would build "hello.1" and "%.%.2" because each function
is considered a single "word" and only the first pattern is replaced.
Fix the expansion so each whitespace-separated string is considered a
word and the first pattern is replaced, giving "hello.1" and
"hello.%.2".
* src/rule.c (snap_implicit_rules): Keep enough space to replace %
with $(*F) if necessary.
* src/implicit.c (pattern_search): During second expansion break each
get_next_word result into individual words and replace the first % in
each with $* or $(*F) as needed.
* tests/scripts/features/patternrules: Add tests for variations.
The hash function we use can yield different results on big- and
little-endian systems which makes test output different. Choose
names to avoid this.
* tests/scripts/features/patternrules: Choose portable target names.
* tests/scripts/features/se_explicit: Ditto.
* tests/scripts/features/se_implicit: Ditto.
