git.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
* [PATCH 0/1] RUNTIME_PREFIX on POSIX systems.
@ 2017-11-16 17:05 Dan Jacques
  2017-11-16 17:05 ` [PATCH 1/1] exec_cmd: RUNTIME_PREFIX on some " Dan Jacques
  0 siblings, 1 reply; 3+ messages in thread
From: Dan Jacques @ 2017-11-16 17:05 UTC (permalink / raw)
  To: git; +Cc: Dan Jacques

Hello! This would be my first contribution to the Git project, so please
accept my apology in advance for any mistakes and let me know what I can
do better.

This patch expands support for the RUNTIME_PREFIX configuration flag,
currently only used on Windows builds, to include Linux, Darwin, and
FreeBSD. When Git is built with RUNTIME_PREFIX enabled, it resolves its
ancillary paths relative to the runtime location of its executable
rather than hard-coding them at compile-time, allowing a Git
installation to be deployed to a path other than the one in which it
was installed.

It was useful to create Git distribution bundles that could unpack
fully-functional Git deployments to arbitrary locations in support of the
Chromium project. Chromium has been using Git bundles built with a patch
similar to this one on its Linux and Mac continuous integration fleet (plus
some developer systems) for almost a year now.

RUNTIME_PREFIX remains an optional configuration flag, so standard Git
builds will not see any changes. However, with this patch applied,
Linux, Darwin, and FreeBSD users can now optionally use "config.mak" to
enable RUNTIME_PREFIX and build relocatable Git distributions. An
example "config.mak" that builds relocatable Git binaries for Linux/Mac
is:

# BEGIN: config.mak
RUNTIME_PREFIX = YesPlease
gitexecdir = libexec/git-core
template_dir = share/git-core/templates
sysconfdir = etc
# END: config.mak

Implementation notes:

It is unfortunately not straightforward to resolve the full absolute path
of the currently-running binary. On some operating systems, notably
Windows, this path is executively supplied as argv[0]. On other
operating systems, however, argv[0] is supplied by the invoker (shell,
script, kernel, etc.), and is not a reliable source of information about
the running Git binary.

The specific method that this patch employs for binary directory resolution
varies depending on the operating system. On Linux and FreeBSD,
Git resolves "/proc/self/exe" and "/proc/curproc/file" respectively. On
Darwin, Git uses the "_NSGetExecutablePath" function. On all operating
systems, notably Windows, Git continues to fall back to resolution
against argv[0] when it is an absolute path.

When RUNTIME_PREFIX is enabled, the resolved runtime path needs to be
passed to ancillary Git tools for their own resolution requirements:

- C-source Git programs will use the EXEC_PATH_ENVIRONMENT environment
  variable that Git already exports, ensuring that any launched tools use
  the same runtime prefix as the entry point.

- PERL tooling needs to know how to locate Git's support libraries. When
  RUNTIME_PREFIX is configured, Git now exports the GITPERLLIB environment
  variable, a mechanism that Git's PERL tooling supports that appears to be
  built for testing. PERL scripts installed using MakeMaker incorporate the
  builder system's PERL version into their installation path, making
  it inconsistent to hard-code; consequently, this patch opts to disable
  MakeMaker for RUNTIME_PREFIX builds in order to deterministically control
  the destination of Git's support libraries.

- Git also exports the GIT_TEXTDOMAINDIR environment variable when
  RUNTIME_PREFIX is set so that its locale configuration can be leveraged
  by Git tooling gettext().

Please note that this patch affects Windows Git builds, since the Windows
Git project uses RUNTIME_PREFIX to support arbitrary installation paths.
Notably, PERL scripts are now always installed without MakeMaker (if they
weren't before), and EXEC_PATH_ENVIRONMENT is preferred by tools instead of
re-resolving argv[0]. Chromium uses the stock redistributable Windows Git
package, so I haven't had an opportunity to test this patch on that
platform.

Please take a look and let me know what you think. Thanks!
-Dan

Dan Jacques (1):
  exec_cmd: RUNTIME_PREFIX on some POSIX systems

 Makefile               |  29 ++++++--
 builtin/receive-pack.c |   2 +-
 cache.h                |   2 +
 common-main.c          |   4 +-
 config.mak.uname       |   4 ++
 exec_cmd.c             | 189 +++++++++++++++++++++++++++++++++++++++++++------
 exec_cmd.h             |   6 +-
 gettext.c              |   7 +-
 git.c                  |   7 +-
 http-backend.c         |   2 +-
 shell.c                |   2 +-
 upload-pack.c          |   2 +-
 12 files changed, 217 insertions(+), 39 deletions(-)

-- 
2.15.0.448.gf294e3d99a-goog


^ permalink raw reply	[flat|nested] 3+ messages in thread

* [PATCH 1/1] exec_cmd: RUNTIME_PREFIX on some POSIX systems
  2017-11-16 17:05 [PATCH 0/1] RUNTIME_PREFIX on POSIX systems Dan Jacques
@ 2017-11-16 17:05 ` Dan Jacques
  2017-11-17  5:08   ` Junio C Hamano
  0 siblings, 1 reply; 3+ messages in thread
From: Dan Jacques @ 2017-11-16 17:05 UTC (permalink / raw)
  To: git; +Cc: Dan Jacques

Enable Git to resolve its own binary location using a variety of
OS-specific and generic methods, including:

- procfs via "/proc/self/exe" (Linux)
- _NSGetExecutablePath (Darwin)
- argv0, if absolute (all, including Windows).

This is used to enable RUNTIME_PREFIX support for non-Windows systems,
notably Linux and Darwin. When configured with RUNTIME_PREFIX, Git will
do a best-effort resolution of its executable path and automatically use
this as its "exec_path" for relative helper and data lookups, unless
explicitly overridden.

Git will also always export and consume its resolved "exec_path" using
the EXEC_PATH_ENVIRONMENT regardless of whether the user has overridden
it, simplifying future lookups and ensuring consistency in Git tooling
execution.

When building with a runtime prefix, Git's PERL libraries are now
installed to a consistently-named directory. This path is resolved and
exported to Git's delegate PERL invocations using the GITPERLLIB
environment variable. This enables Git's delegate PERL scripts to import
Git's own PERL libraries from a path relative to the executable.

Small incidental formatting cleanup of "exec_cmd.c".

Signed-off-by: Dan Jacques <dnj@google.com>
---
 Makefile               |  29 ++++++--
 builtin/receive-pack.c |   2 +-
 cache.h                |   2 +
 common-main.c          |   4 +-
 config.mak.uname       |   4 ++
 exec_cmd.c             | 189 +++++++++++++++++++++++++++++++++++++++++++------
 exec_cmd.h             |   6 +-
 gettext.c              |   7 +-
 git.c                  |   7 +-
 http-backend.c         |   2 +-
 shell.c                |   2 +-
 upload-pack.c          |   2 +-
 12 files changed, 217 insertions(+), 39 deletions(-)

diff --git a/Makefile b/Makefile
index ee9d5eb11..80db01706 100644
--- a/Makefile
+++ b/Makefile
@@ -296,7 +296,8 @@ all::
 # Define PERL_PATH to the path of your Perl binary (usually /usr/bin/perl).
 #
 # Define NO_PERL_MAKEMAKER if you cannot use Makefiles generated by perl's
-# MakeMaker (e.g. using ActiveState under Cygwin).
+# MakeMaker (e.g. using ActiveState under Cygwin, or building with a fixed
+# runtime prefix).
 #
 # Define NO_PERL if you do not want Perl scripts or libraries at all.
 #
@@ -462,6 +463,7 @@ ARFLAGS = rcs
 #   mandir
 #   infodir
 #   htmldir
+#   localedir
 # This can help installing the suite in a relocatable way.
 
 prefix = $(HOME)
@@ -485,6 +487,7 @@ pathsep = :
 mandir_relative = $(patsubst $(prefix)/%,%,$(mandir))
 infodir_relative = $(patsubst $(prefix)/%,%,$(infodir))
 htmldir_relative = $(patsubst $(prefix)/%,%,$(htmldir))
+localedir_relative = $(patsubst $(prefix)/%,%,$(localedir))
 
 export prefix bindir sharedir sysconfdir gitwebdir localedir
 
@@ -1522,9 +1525,6 @@ ifdef SHA1_MAX_BLOCK_SIZE
 	LIB_OBJS += compat/sha1-chunked.o
 	BASIC_CFLAGS += -DSHA1_MAX_BLOCK_SIZE="$(SHA1_MAX_BLOCK_SIZE)"
 endif
-ifdef NO_PERL_MAKEMAKER
-	export NO_PERL_MAKEMAKER
-endif
 ifdef NO_HSTRERROR
 	COMPAT_CFLAGS += -DNO_HSTRERROR
 	COMPAT_OBJS += compat/hstrerror.o
@@ -1547,6 +1547,14 @@ else
 endif
 ifdef RUNTIME_PREFIX
 	COMPAT_CFLAGS += -DRUNTIME_PREFIX
+
+	# Control PERL library location so it can be referenced by relocatable
+	# code.
+	NO_PERL_MAKEMAKER = YesPlease
+endif
+
+ifdef NO_PERL_MAKEMAKER
+	export NO_PERL_MAKEMAKER
 endif
 
 ifdef NO_PTHREADS
@@ -1632,6 +1640,15 @@ ifdef HAVE_GETDELIM
 	BASIC_CFLAGS += -DHAVE_GETDELIM
 endif
 
+ifneq ($(PROCFS_EXECUTABLE_PATH),)
+	procfs_executable_path_SQ = $(subst ','\'',$(PROCFS_EXECUTABLE_PATH))
+	BASIC_CFLAGS += '-DPROCFS_EXECUTABLE_PATH="$(procfs_executable_path_SQ)"'
+endif
+
+ifdef HAVE_NS_GET_EXECUTABLE_PATH
+	BASIC_CFLAGS += -DHAVE_NS_GET_EXECUTABLE_PATH
+endif
+
 ifeq ($(TCLTK_PATH),)
 NO_TCLTK = NoThanks
 endif
@@ -1714,6 +1731,7 @@ bindir_relative_SQ = $(subst ','\'',$(bindir_relative))
 mandir_relative_SQ = $(subst ','\'',$(mandir_relative))
 infodir_relative_SQ = $(subst ','\'',$(infodir_relative))
 localedir_SQ = $(subst ','\'',$(localedir))
+localedir_relative_SQ = $(subst ','\'',$(localedir_relative))
 gitexecdir_SQ = $(subst ','\'',$(gitexecdir))
 template_dir_SQ = $(subst ','\'',$(template_dir))
 htmldir_relative_SQ = $(subst ','\'',$(htmldir_relative))
@@ -2130,6 +2148,7 @@ endif
 exec_cmd.sp exec_cmd.s exec_cmd.o: GIT-PREFIX
 exec_cmd.sp exec_cmd.s exec_cmd.o: EXTRA_CPPFLAGS = \
 	'-DGIT_EXEC_PATH="$(gitexecdir_SQ)"' \
+	'-DGIT_LOCALE_PATH="$(localedir_relative_SQ)"' \
 	'-DBINDIR="$(bindir_relative_SQ)"' \
 	'-DPREFIX="$(prefix_SQ)"'
 
@@ -2147,7 +2166,7 @@ attr.sp attr.s attr.o: EXTRA_CPPFLAGS = \
 
 gettext.sp gettext.s gettext.o: GIT-PREFIX
 gettext.sp gettext.s gettext.o: EXTRA_CPPFLAGS = \
-	-DGIT_LOCALE_PATH='"$(localedir_SQ)"'
+	-DGIT_LOCALE_PATH='"$(localedir_relative_SQ)"'
 
 http-push.sp http.sp http-walker.sp remote-curl.sp imap-send.sp: SPARSE_FLAGS += \
 	-DCURL_DISABLE_TYPECHECK
diff --git a/builtin/receive-pack.c b/builtin/receive-pack.c
index 4d37a160d..51203cc3d 100644
--- a/builtin/receive-pack.c
+++ b/builtin/receive-pack.c
@@ -1947,7 +1947,7 @@ int cmd_receive_pack(int argc, const char **argv, const char *prefix)
 
 	service_dir = argv[0];
 
-	setup_path();
+	setup_path_and_env();
 
 	if (!enter_repo(service_dir, 0))
 		die("'%s' does not appear to be a git repository", service_dir);
diff --git a/cache.h b/cache.h
index cb7fb7c00..9ef59f1cc 100644
--- a/cache.h
+++ b/cache.h
@@ -445,6 +445,8 @@ static inline enum object_type object_type(unsigned int mode)
 #define GIT_ICASE_PATHSPECS_ENVIRONMENT "GIT_ICASE_PATHSPECS"
 #define GIT_QUARANTINE_ENVIRONMENT "GIT_QUARANTINE_PATH"
 #define GIT_OPTIONAL_LOCKS_ENVIRONMENT "GIT_OPTIONAL_LOCKS"
+#define GIT_PERL_LIB_ENVIRONMENT "GITPERLLIB"
+#define GIT_TEXT_DOMAIN_DIR_ENVIRONMENT "GIT_TEXTDOMAINDIR"
 
 /*
  * This environment variable is expected to contain a boolean indicating
diff --git a/common-main.c b/common-main.c
index 6a689007e..6516a1f89 100644
--- a/common-main.c
+++ b/common-main.c
@@ -32,12 +32,12 @@ int main(int argc, const char **argv)
 	 */
 	sanitize_stdfds();
 
+	git_resolve_executable_dir(argv[0]);
+
 	git_setup_gettext();
 
 	attr_start();
 
-	git_extract_argv0_path(argv[0]);
-
 	restore_sigpipe_to_default();
 
 	return cmd_main(argc, argv);
diff --git a/config.mak.uname b/config.mak.uname
index 685a80d13..bd5c326d4 100644
--- a/config.mak.uname
+++ b/config.mak.uname
@@ -37,6 +37,7 @@ ifeq ($(uname_S),Linux)
 	HAVE_GETDELIM = YesPlease
 	SANE_TEXT_GREP=-a
 	FREAD_READS_DIRECTORIES = UnfortunatelyYes
+	PROCFS_EXECUTABLE_PATH = /proc/self/exe
 endif
 ifeq ($(uname_S),GNU/kFreeBSD)
 	HAVE_ALLOCA_H = YesPlease
@@ -111,6 +112,7 @@ ifeq ($(uname_S),Darwin)
 	BASIC_CFLAGS += -DPROTECT_HFS_DEFAULT=1
 	HAVE_BSD_SYSCTL = YesPlease
 	FREAD_READS_DIRECTORIES = UnfortunatelyYes
+	HAVE_NS_GET_EXECUTABLE_PATH = YesPlease
 endif
 ifeq ($(uname_S),SunOS)
 	NEEDS_SOCKET = YesPlease
@@ -218,6 +220,7 @@ ifeq ($(uname_S),OpenBSD)
 	BASIC_LDFLAGS += -L/usr/local/lib
 	HAVE_PATHS_H = YesPlease
 	HAVE_BSD_SYSCTL = YesPlease
+	PROCFS_EXECUTABLE_PATH = /proc/curproc/file
 endif
 ifeq ($(uname_S),MirBSD)
 	NO_STRCASESTR = YesPlease
@@ -236,6 +239,7 @@ ifeq ($(uname_S),NetBSD)
 	USE_ST_TIMESPEC = YesPlease
 	HAVE_PATHS_H = YesPlease
 	HAVE_BSD_SYSCTL = YesPlease
+	PROCFS_EXECUTABLE_PATH = /proc/curproc/exe
 endif
 ifeq ($(uname_S),AIX)
 	DEFAULT_PAGER = more
diff --git a/exec_cmd.c b/exec_cmd.c
index ce192a2d6..10ca7a75a 100644
--- a/exec_cmd.c
+++ b/exec_cmd.c
@@ -2,53 +2,187 @@
 #include "exec_cmd.h"
 #include "quote.h"
 #include "argv-array.h"
+
+#if defined(RUNTIME_PREFIX) && defined(HAVE_NS_GET_EXECUTABLE_PATH)
+#include <mach-o/dyld.h>
+#endif
+
 #define MAX_ARGS	32
 
 static const char *argv_exec_path;
 
+static const char *system_prefix(void);
+
 #ifdef RUNTIME_PREFIX
-static const char *argv0_path;
+
+/**
+ * When using a runtime prefix, Git dynamically resolves paths relative to its
+ * executable.
+ *
+ * The method for determining the path of the executable is highly
+ * platform-specific.
+ */
+
+/**
+ * Path to the current Git executable. Resolved on startup by
+ * 'git_resolve_executable_dir'.
+ */
+static const char *executable_dirname;
 
 static const char *system_prefix(void)
 {
 	static const char *prefix;
 
-	assert(argv0_path);
-	assert(is_absolute_path(argv0_path));
+	assert(executable_dirname);
+	assert(is_absolute_path(executable_dirname));
 
 	if (!prefix &&
-	    !(prefix = strip_path_suffix(argv0_path, GIT_EXEC_PATH)) &&
-	    !(prefix = strip_path_suffix(argv0_path, BINDIR)) &&
-	    !(prefix = strip_path_suffix(argv0_path, "git"))) {
+	    !(prefix = strip_path_suffix(executable_dirname, GIT_EXEC_PATH)) &&
+	    !(prefix = strip_path_suffix(executable_dirname, BINDIR)) &&
+	    !(prefix = strip_path_suffix(executable_dirname, "git"))) {
 		prefix = PREFIX;
 		trace_printf("RUNTIME_PREFIX requested, "
-				"but prefix computation failed.  "
-				"Using static fallback '%s'.\n", prefix);
+			     "but prefix computation failed.  "
+			     "Using static fallback '%s'.\n",
+			     prefix);
 	}
 	return prefix;
 }
 
-void git_extract_argv0_path(const char *argv0)
+/*
+ * Resolves the executable path from argv[0], only if it is absolute.
+ *
+ * Returns 0 on success, -1 on failure.
+ */
+static int git_get_exec_path_from_argv0(struct strbuf *buf, const char *argv0)
 {
 	const char *slash;
 
 	if (!argv0 || !*argv0)
-		return;
+		return -1;
 
 	slash = find_last_dir_sep(argv0);
+	if (slash) {
+		trace_printf("Determined executable path from argv0: %s\n",
+			     argv0);
+		strbuf_add_absolute_path(buf, argv0);
+		return 0;
+	}
+	return -1;
+}
+
+#ifdef PROCFS_EXECUTABLE_PATH
+/*
+ * Resolves the executable path by examining a procfs symlink.
+ *
+ * Returns 0 on success, -1 on failure.
+ */
+static int git_get_exec_path_procfs(struct strbuf *buf)
+{
+	char *path = realpath(PROCFS_EXECUTABLE_PATH, NULL);
+	if (path) {
+		trace_printf("Determined executable path from procfs: %s\n",
+			     path);
+		strbuf_addstr(buf, path);
+		free(path);
+		return 0;
+	}
+	return -1;
+}
+#endif /* PROCFS_EXECUTABLE_PATH */
+
+#ifdef HAVE_NS_GET_EXECUTABLE_PATH
+/*
+ * Resolves the executable path by querying Darwin applicaton stack.
+ *
+ * Returns 0 on success, -1 on failure.
+ */
+static int git_get_exec_path_darwin(struct strbuf *buf)
+{
+	char path[PATH_MAX];
+	uint32_t size = sizeof(path);
+	if (!_NSGetExecutablePath(path, &size)) {
+		trace_printf(
+			"Determined executable path from Darwin stack: %s\n",
+			path);
+		strbuf_addstr(buf, path);
+		return 0;
+	}
+	return -1;
+}
+#endif /* HAVE_NS_GET_EXECUTABLE_PATH */
+
+/*
+ * Resolves the absolute path of the current executable by employing
+ * one or more platform-specific methods.
+ *
+ * Returns 0 on success, -1 on failure.
+ */
+static int git_get_exec_path(struct strbuf *buf, const char *argv0)
+{
+	if (
+#ifdef PROCFS_EXECUTABLE_PATH
+		git_get_exec_path_procfs(buf) &&
+#endif /* PROCFS_EXECUTABLE_PATH */
+#ifdef HAVE_NS_GET_EXECUTABLE_PATH
+		git_get_exec_path_darwin(buf) &&
+#endif /* HAVE_NS_GET_EXECUTABLE_PATH */
+		git_get_exec_path_from_argv0(buf, argv0)) {
+		return -1;
+	}
+
+	if (strbuf_normalize_path(buf)) {
+		trace_printf("Could not normalize path: %s\n", buf->buf);
+		return -1;
+	}
+
+	return 0;
+}
+
+void git_resolve_executable_dir(const char *argv0)
+{
+	struct strbuf buf = STRBUF_INIT;
+	char *resolved;
+	const char *slash;
+	const char *path;
+
+	path = getenv(EXEC_PATH_ENVIRONMENT);
+	if (path) {
+		trace_printf("Determined executable path from ENV: %s\n", path);
+		executable_dirname = path;
+	} else {
+		if (git_get_exec_path(&buf, argv0)) {
+			trace_printf(
+				"Could not determine executable path from: %s\n",
+				argv0);
+			strbuf_release(&buf);
+			return;
+		}
+
+		resolved = strbuf_detach(&buf, NULL);
+		slash = find_last_dir_sep(resolved);
+		if (slash)
+			resolved[slash - resolved] = '\0';
+
+		executable_dirname = resolved;
+	}
 
-	if (slash)
-		argv0_path = xstrndup(argv0, slash - argv0);
+	trace_printf("Determined executable dir: %s\n", executable_dirname);
 }
 
 #else
 
+/**
+ * When not using a runtime prefix, Git uses a hard-coded path, and there is
+ * nothing to resolve.
+ */
+
 static const char *system_prefix(void)
 {
 	return PREFIX;
 }
 
-void git_extract_argv0_path(const char *argv0)
+void git_resolve_executable_dir(const char *argv0)
 {
 }
 
@@ -65,7 +199,7 @@ char *system_path(const char *path)
 	return strbuf_detach(&d, NULL);
 }
 
-void git_set_argv_exec_path(const char *exec_path)
+void git_set_exec_path(const char *exec_path)
 {
 	argv_exec_path = exec_path;
 	/*
@@ -74,7 +208,6 @@ void git_set_argv_exec_path(const char *exec_path)
 	setenv(EXEC_PATH_ENVIRONMENT, exec_path, 1);
 }
 
-
 /* Returns the highest-priority, location to look for git programs. */
 const char *git_exec_path(void)
 {
@@ -101,12 +234,14 @@ static void add_path(struct strbuf *out, const char *path)
 	}
 }
 
-void setup_path(void)
+void setup_path_and_env(void)
 {
+	const char *exec_path = git_exec_path();
 	const char *old_path = getenv("PATH");
 	struct strbuf new_path = STRBUF_INIT;
 
-	add_path(&new_path, git_exec_path());
+	git_set_exec_path(exec_path);
+	add_path(&new_path, exec_path);
 
 	if (old_path)
 		strbuf_addstr(&new_path, old_path);
@@ -116,6 +251,20 @@ void setup_path(void)
 	setenv("PATH", new_path.buf, 1);
 
 	strbuf_release(&new_path);
+
+	/*
+	 * If RUNTIME_PREFIX is set, the relative PERL library and locale paths
+	 * must be exported for invoked programs to inherit the calculated
+	 * runtime path.
+	 *
+	 * When RUNTIME_PREFIX is defined, Git's PERL include path is always
+	 * "$(PREFIX)/lib" (by setting NO_PERL_MAKEMAKER).
+	 */
+#ifdef RUNTIME_PREFIX
+	setenv(GIT_TEXT_DOMAIN_DIR_ENVIRONMENT, system_path(GIT_LOCALE_PATH),
+	       1);
+	setenv(GIT_PERL_LIB_ENVIRONMENT, system_path("lib"), 1);
+#endif /* RUNTIME_PREFIX */
 }
 
 const char **prepare_git_cmd(struct argv_array *out, const char **argv)
@@ -125,7 +274,8 @@ const char **prepare_git_cmd(struct argv_array *out, const char **argv)
 	return out->argv;
 }
 
-int execv_git_cmd(const char **argv) {
+int execv_git_cmd(const char **argv)
+{
 	struct argv_array nargv = ARGV_ARRAY_INIT;
 
 	prepare_git_cmd(&nargv, argv);
@@ -140,8 +290,7 @@ int execv_git_cmd(const char **argv) {
 	return -1;
 }
 
-
-int execl_git_cmd(const char *cmd,...)
+int execl_git_cmd(const char *cmd, ...)
 {
 	int argc;
 	const char *argv[MAX_ARGS + 1];
diff --git a/exec_cmd.h b/exec_cmd.h
index ff0b48048..490493745 100644
--- a/exec_cmd.h
+++ b/exec_cmd.h
@@ -3,10 +3,10 @@
 
 struct argv_array;
 
-extern void git_set_argv_exec_path(const char *exec_path);
-extern void git_extract_argv0_path(const char *path);
+extern void git_set_exec_path(const char *exec_path);
+extern void git_resolve_executable_dir(const char *path);
 extern const char *git_exec_path(void);
-extern void setup_path(void);
+extern void setup_path_and_env(void);
 extern const char **prepare_git_cmd(struct argv_array *out, const char **argv);
 extern int execv_git_cmd(const char **argv); /* NULL terminated */
 LAST_ARG_MUST_BE_NULL
diff --git a/gettext.c b/gettext.c
index db727ea02..f1c162cd8 100644
--- a/gettext.c
+++ b/gettext.c
@@ -2,7 +2,8 @@
  * Copyright (c) 2010 Ævar Arnfjörð Bjarmason
  */
 
-#include "git-compat-util.h"
+#include "cache.h"
+#include "exec_cmd.h"
 #include "gettext.h"
 #include "strbuf.h"
 #include "utf8.h"
@@ -157,10 +158,10 @@ static void init_gettext_charset(const char *domain)
 
 void git_setup_gettext(void)
 {
-	const char *podir = getenv("GIT_TEXTDOMAINDIR");
+	const char *podir = getenv(GIT_TEXT_DOMAIN_DIR_ENVIRONMENT);
 
 	if (!podir)
-		podir = GIT_LOCALE_PATH;
+		podir = system_path(GIT_LOCALE_PATH);
 	bindtextdomain("git", podir);
 	setlocale(LC_MESSAGES, "");
 	setlocale(LC_TIME, "");
diff --git a/git.c b/git.c
index 9e96dd409..f6336bca0 100644
--- a/git.c
+++ b/git.c
@@ -65,7 +65,7 @@ static int handle_options(const char ***argv, int *argc, int *envchanged)
 		 */
 		if (skip_prefix(cmd, "--exec-path", &cmd)) {
 			if (*cmd == '=')
-				git_set_argv_exec_path(cmd + 1);
+				git_set_exec_path(cmd + 1);
 			else {
 				puts(git_exec_path());
 				exit(0);
@@ -676,8 +676,11 @@ int cmd_main(int argc, const char **argv)
 	 * precedence paths: the "--exec-path" option, the GIT_EXEC_PATH
 	 * environment, and the $(gitexecdir) from the Makefile at build
 	 * time.
+	 *
+	 * If RUNTIME_PREFIX is defined, we export some additional environment
+	 * variables to help subprocess code identify runtime-derived paths.
 	 */
-	setup_path();
+	setup_path_and_env();
 
 	while (1) {
 		int was_alias = run_argv(&argc, &argv);
diff --git a/http-backend.c b/http-backend.c
index f3dc218b2..200d4dc13 100644
--- a/http-backend.c
+++ b/http-backend.c
@@ -697,7 +697,7 @@ int cmd_main(int argc, const char **argv)
 	if (!cmd)
 		not_found(&hdr, "Request not supported: '%s'", dir);
 
-	setup_path();
+	setup_path_and_env();
 	if (!enter_repo(dir, 0))
 		not_found(&hdr, "Not a git repository: '%s'", dir);
 	if (!getenv("GIT_HTTP_EXPORT_ALL") &&
diff --git a/shell.c b/shell.c
index 234b2d4f1..82177f5bc 100644
--- a/shell.c
+++ b/shell.c
@@ -12,7 +12,7 @@ static int do_generic_cmd(const char *me, char *arg)
 {
 	const char *my_argv[4];
 
-	setup_path();
+	setup_path_and_env();
 	if (!arg || !(arg = sq_dequote(arg)) || *arg == '-')
 		die("bad argument");
 	if (!starts_with(me, "git-"))
diff --git a/upload-pack.c b/upload-pack.c
index 6d5f3c0d3..aa3d693ab 100644
--- a/upload-pack.c
+++ b/upload-pack.c
@@ -1058,7 +1058,7 @@ int cmd_main(int argc, const char **argv)
 	if (timeout)
 		daemon_mode = 1;
 
-	setup_path();
+	setup_path_and_env();
 
 	dir = argv[0];
 
-- 
2.15.0.448.gf294e3d99a-goog


^ permalink raw reply related	[flat|nested] 3+ messages in thread

* Re: [PATCH 1/1] exec_cmd: RUNTIME_PREFIX on some POSIX systems
  2017-11-16 17:05 ` [PATCH 1/1] exec_cmd: RUNTIME_PREFIX on some " Dan Jacques
@ 2017-11-17  5:08   ` Junio C Hamano
  0 siblings, 0 replies; 3+ messages in thread
From: Junio C Hamano @ 2017-11-17  5:08 UTC (permalink / raw)
  To: Dan Jacques; +Cc: git

Dan Jacques <dnj@google.com> writes:

> Enable Git to resolve its own binary location using a variety of
> OS-specific and generic methods, including:
>
> - procfs via "/proc/self/exe" (Linux)
> - _NSGetExecutablePath (Darwin)
> - argv0, if absolute (all, including Windows).
>
> This is used to enable RUNTIME_PREFIX support for non-Windows systems,
> notably Linux and Darwin. When configured with RUNTIME_PREFIX, Git will
> do a best-effort resolution of its executable path and automatically use
> this as its "exec_path" for relative helper and data lookups, unless
> explicitly overridden.

Yay.

> Git will also always export and consume its resolved "exec_path" using
> the EXEC_PATH_ENVIRONMENT regardless of whether the user has overridden
> it, simplifying future lookups and ensuring consistency in Git tooling
> execution.

The "regardless of whether the user has overridden it" part sounded
alarming and made me wince twice.  I think you meant

 - If the user already has GIT_EXEC_PATH in the environment pointing
   somewhere, when we end up calling git_set_exec_path() to export
   the variable, the value we have at hand to be exported is what
   originally came from the user, bypassing the auto-detection logic
   this patch adds.

 - If the user did not have GIT_EXEC_PATH, the auto-detection logic
   is exercised, and the result is exported when git_set_exec_path()
   is called.  Any program that is spawned by us as a subprocess
   will inherit the same GIT_EXEC_PATH we detected.

As I have multiple installations of various versions of Git, I find
the latter somewhat disturbing.  

Suppose that I designate one stable version of Git and install it at
$HOME/git-stable/{bin,libexec,...}/, and want that version to always
be used in my hooks and other helper scripts that are spawned by
Git, even when I am trying out a newer version of Git that is under
testing.

My hooks would be running $HOME/git-stable/bin/git subcommand" (or
more realistically, it would do "PATH=$HOME/git-stable/bin:$PATH"
upfront), but with this patch (and with runtime-prefix layout) they
would use GIT_EXEC_PATH that was discovered and exported by the
version of Git under testing that invoked the hook, leading to an
inconsistent and hard to debug behaviour, no?

On the other hand, as long as existing GIT_EXEC_PATH is passed-thru
(i.e. the first point above), I think that is a sensible thing to
do.

> When building with a runtime prefix, Git's PERL libraries are now
> installed to a consistently-named directory. This path is resolved and
> exported to Git's delegate PERL invocations using the GITPERLLIB
> environment variable. This enables Git's delegate PERL scripts to import
> Git's own PERL libraries from a path relative to the executable.

Sounds good.

> Small incidental formatting cleanup of "exec_cmd.c".

We usually frown upon these because they often gets distracting, but
I didn't find the ones in this patch too bad.

> diff --git a/Makefile b/Makefile
> index ee9d5eb11..80db01706 100644
> --- a/Makefile
> +++ b/Makefile
> @@ -296,7 +296,8 @@ all::
>  # Define PERL_PATH to the path of your Perl binary (usually /usr/bin/perl).
>  #
>  # Define NO_PERL_MAKEMAKER if you cannot use Makefiles generated by perl's
> -# MakeMaker (e.g. using ActiveState under Cygwin).
> +# MakeMaker (e.g. using ActiveState under Cygwin, or building with a fixed
> +# runtime prefix).

Windows folks may want to comment (either positively or negatively)
as they are the only ones that are currently using runtime-prefix
layout, especially on this part:

> @@ -1547,6 +1547,14 @@ else
>  endif
>  ifdef RUNTIME_PREFIX
>  	COMPAT_CFLAGS += -DRUNTIME_PREFIX
> +
> +	# Control PERL library location so it can be referenced by relocatable
> +	# code.
> +	NO_PERL_MAKEMAKER = YesPlease
> +endif


> diff --git a/cache.h b/cache.h
> index cb7fb7c00..9ef59f1cc 100644
> --- a/cache.h
> +++ b/cache.h
> @@ -445,6 +445,8 @@ static inline enum object_type object_type(unsigned int mode)
>  #define GIT_ICASE_PATHSPECS_ENVIRONMENT "GIT_ICASE_PATHSPECS"
>  #define GIT_QUARANTINE_ENVIRONMENT "GIT_QUARANTINE_PATH"
>  #define GIT_OPTIONAL_LOCKS_ENVIRONMENT "GIT_OPTIONAL_LOCKS"
> +#define GIT_PERL_LIB_ENVIRONMENT "GITPERLLIB"
> +#define GIT_TEXT_DOMAIN_DIR_ENVIRONMENT "GIT_TEXTDOMAINDIR"

This is a nice touch.  The other day I noticed that we stopped
defining these when we start using a new enviornment variable, which
should be rectified (this is not within the scope of this patch--I
am just welcoming this change that fixes the existing issue a bit
without getting distracing).

> diff --git a/common-main.c b/common-main.c
> index 6a689007e..6516a1f89 100644
> --- a/common-main.c
> +++ b/common-main.c
> @@ -32,12 +32,12 @@ int main(int argc, const char **argv)
>  	 */
>  	sanitize_stdfds();
>  
> +	git_resolve_executable_dir(argv[0]);
> +
>  	git_setup_gettext();
>  
>  	attr_start();
>  
> -	git_extract_argv0_path(argv[0]);
> -

I presume that this is because we may need to know where to find the
locale stuff before calling git_setup_gettext(); makes sense.

> diff --git a/exec_cmd.c b/exec_cmd.c
> index ce192a2d6..10ca7a75a 100644
> --- a/exec_cmd.c
> +++ b/exec_cmd.c
> @@ -2,53 +2,187 @@
>  #include "exec_cmd.h"
>  #include "quote.h"
>  #include "argv-array.h"
> +
> +#if defined(RUNTIME_PREFIX) && defined(HAVE_NS_GET_EXECUTABLE_PATH)
> +#include <mach-o/dyld.h>
> +#endif
> +
>  #define MAX_ARGS	32
>  
>  static const char *argv_exec_path;
>  
> +static const char *system_prefix(void);
> +
>  #ifdef RUNTIME_PREFIX
> -static const char *argv0_path;
> +
> +/**
> + * When using a runtime prefix, Git dynamically resolves paths relative to its
> + * executable.
> + *
> + * The method for determining the path of the executable is highly
> + * platform-specific.
> + */
> +
> +/**
> + * Path to the current Git executable. Resolved on startup by
> + * 'git_resolve_executable_dir'.
> + */
> +static const char *executable_dirname;

OK, so that is a more appropriate name for the variable that is a
logical successor of argv0_path.  I wonder if the file-scope static
variable argv_exec_path we see above would want to move to somewhere
closer to one of these "platform specific methods", though.

>  static const char *system_prefix(void)
>  {
>  	static const char *prefix;
>  
> -	assert(argv0_path);
> -	assert(is_absolute_path(argv0_path));
> +	assert(executable_dirname);
> +	assert(is_absolute_path(executable_dirname));
>  
>  	if (!prefix &&
> -	    !(prefix = strip_path_suffix(argv0_path, GIT_EXEC_PATH)) &&
> -	    !(prefix = strip_path_suffix(argv0_path, BINDIR)) &&
> -	    !(prefix = strip_path_suffix(argv0_path, "git"))) {
> +	    !(prefix = strip_path_suffix(executable_dirname, GIT_EXEC_PATH)) &&
> +	    !(prefix = strip_path_suffix(executable_dirname, BINDIR)) &&
> +	    !(prefix = strip_path_suffix(executable_dirname, "git"))) {
>  		prefix = PREFIX;
>  		trace_printf("RUNTIME_PREFIX requested, "
> -				"but prefix computation failed.  "
> -				"Using static fallback '%s'.\n", prefix);
> +			     "but prefix computation failed.  "
> +			     "Using static fallback '%s'.\n",
> +			     prefix);
>  	}
>  	return prefix;
>  }
>  
> -void git_extract_argv0_path(const char *argv0)
> +/*
> + * Resolves the executable path from argv[0], only if it is absolute.
> + *
> + * Returns 0 on success, -1 on failure.
> + */
> +static int git_get_exec_path_from_argv0(struct strbuf *buf, const char *argv0)
>  {
>  	const char *slash;
>  
>  	if (!argv0 || !*argv0)
> -		return;
> +		return -1;
>  
>  	slash = find_last_dir_sep(argv0);
> +	if (slash) {
> +		trace_printf("Determined executable path from argv0: %s\n",
> +			     argv0);
> +		strbuf_add_absolute_path(buf, argv0);
> +		return 0;
> +	}
> +	return -1;
> +}
> +
> +#ifdef PROCFS_EXECUTABLE_PATH
> +/*
> + * Resolves the executable path by examining a procfs symlink.
> + *
> + * Returns 0 on success, -1 on failure.
> + */
> +static int git_get_exec_path_procfs(struct strbuf *buf)
> +{
> +	char *path = realpath(PROCFS_EXECUTABLE_PATH, NULL);

I think this is our first use of realpath(), which is XSI.

    http://pubs.opengroup.org/onlinepubs/9699919799/functions/realpath.html

I'd assume it is available on any platform on which we would want to
use PROCFS_EXECUTABLE_PATH, so this probably is OK.

> +	if (path) {
> +		trace_printf("Determined executable path from procfs: %s\n",
> +			     path);
> +		strbuf_addstr(buf, path);
> +		free(path);
> +		return 0;
> +	}
> +	return -1;
> +}
> +#endif /* PROCFS_EXECUTABLE_PATH */

> +/*
> + * Resolves the absolute path of the current executable by employing
> + * one or more platform-specific methods.
> + *
> + * Returns 0 on success, -1 on failure.
> + */
> +static int git_get_exec_path(struct strbuf *buf, const char *argv0)
> +{
> +	if (
> +#ifdef PROCFS_EXECUTABLE_PATH
> +		git_get_exec_path_procfs(buf) &&
> +#endif /* PROCFS_EXECUTABLE_PATH */
> +#ifdef HAVE_NS_GET_EXECUTABLE_PATH
> +		git_get_exec_path_darwin(buf) &&
> +#endif /* HAVE_NS_GET_EXECUTABLE_PATH */
> +		git_get_exec_path_from_argv0(buf, argv0)) {

The contract between this caller and these platform-specific helpers
is for the callee to return 0 when they succeed, and -1 when they
punt with "I dunno", so the &&-cascade stops after the first one
that returns 0.  I wonder why argv0 (i.e. the full-path case) is not
the first one to try, though---isn't that one the simplest?

Also, I wonder if this caller gets simpler to read and understand if
each of these "platform specific" ones are done like so:

	#ifdef FOO_EXECUTABLE_PATH
	static int git_get_exec_path_foo(struct strbuf *buf)
	{
		...
	}
	#else /* FOO_EXECUTABLE_PATH */
	#define git_get_exec_path_foo(buf) (-1)
	#endif

Then the caller can lose the ifdef/endif that is noisier than the
actual calls we see above, i.e.

	if (!git_get_exec_path_from_argv0(buf, argv0) ||
	    !git_get_exec_path_procfs(buf) ||
            !git_get_exec_path_darwin(buf)) {
		/* at least one of them found what we wanted */
		... do the success case ...
	} else {
		... oops, we failed to find ...
	}

> +void git_resolve_executable_dir(const char *argv0)
> +{
> +	struct strbuf buf = STRBUF_INIT;
> +	char *resolved;
> +	const char *slash;
> +	const char *path;
> +
> +	path = getenv(EXEC_PATH_ENVIRONMENT);
> +	if (path) {
> +		trace_printf("Determined executable path from ENV: %s\n", path);
> +		executable_dirname = path;
> +	} else {
> +		if (git_get_exec_path(&buf, argv0)) {
> +			trace_printf(
> +				"Could not determine executable path from: %s\n",
> +				argv0);
> +			strbuf_release(&buf);
> +			return;
> +		}
> +
> +		resolved = strbuf_detach(&buf, NULL);
> +		slash = find_last_dir_sep(resolved);
> +		if (slash)
> +			resolved[slash - resolved] = '\0';
> +
> +		executable_dirname = resolved;
> +	}
>  
> -	if (slash)
> -		argv0_path = xstrndup(argv0, slash - argv0);
> +	trace_printf("Determined executable dir: %s\n", executable_dirname);
>  }

OK.

> -void git_set_argv_exec_path(const char *exec_path)
> +void git_set_exec_path(const char *exec_path)
>  {
>  	argv_exec_path = exec_path;
>  	/*

I wonder if we want to clean up this part of the system, in that
git_exec_path() is the only user of argv_exec_path which is now a
gross misnomer (the way you figured out exec_path may not be based
on argv at all).  Once resolve-executable-dir is called and we have
a value in executable_dirname, this argv_exec_path variable feels
more or less redundant.  But perhaps that is outside the scope of
this patch.  I dunno.

> @@ -101,12 +234,14 @@ static void add_path(struct strbuf *out, const char *path)
>  	}
>  }
>  
> -void setup_path(void)
> +void setup_path_and_env(void)
>  {
> +	const char *exec_path = git_exec_path();
>  	const char *old_path = getenv("PATH");
>  	struct strbuf new_path = STRBUF_INIT;
>  
> -	add_path(&new_path, git_exec_path());
> +	git_set_exec_path(exec_path);
> +	add_path(&new_path, exec_path);
>  
>  	if (old_path)
>  		strbuf_addstr(&new_path, old_path);
> @@ -116,6 +251,20 @@ void setup_path(void)
>  	setenv("PATH", new_path.buf, 1);
>  
>  	strbuf_release(&new_path);
> +
> +	/*
> +	 * If RUNTIME_PREFIX is set, the relative PERL library and locale paths
> +	 * must be exported for invoked programs to inherit the calculated
> +	 * runtime path.
> +	 *
> +	 * When RUNTIME_PREFIX is defined, Git's PERL include path is always
> +	 * "$(PREFIX)/lib" (by setting NO_PERL_MAKEMAKER).
> +	 */
> +#ifdef RUNTIME_PREFIX
> +	setenv(GIT_TEXT_DOMAIN_DIR_ENVIRONMENT, system_path(GIT_LOCALE_PATH),
> +	       1);
> +	setenv(GIT_PERL_LIB_ENVIRONMENT, system_path("lib"), 1);
> +#endif /* RUNTIME_PREFIX */
>  }

So..., now this function exports textdomain and perllib environment
when we are compiled with runtime-prefix, and exports GIT_EXEC_PATH
even when we are *not* compiled with runtime-prefix (by calling the
git_set_exec_path() function that does the exporting)?  That makes
the worry I expressed in an earlier part of the message even worse,
as I assumed this gratuitous exporting would be in effect only when
runtime-prefix is in use.  I hope I am mistaken...

> @@ -125,7 +274,8 @@ const char **prepare_git_cmd(struct argv_array *out, const char **argv)
>  	return out->argv;
>  }
>  
> -int execv_git_cmd(const char **argv) {
> +int execv_git_cmd(const char **argv)
> +{
>  	struct argv_array nargv = ARGV_ARRAY_INIT;
>  
>  	prepare_git_cmd(&nargv, argv);
> @@ -140,8 +290,7 @@ int execv_git_cmd(const char **argv) {
>  	return -1;
>  }
>  
> -
> -int execl_git_cmd(const char *cmd,...)
> +int execl_git_cmd(const char *cmd, ...)
>  {
>  	int argc;
>  	const char *argv[MAX_ARGS + 1];

These two are good changes (reducing the inter-function spacing
might not be in that "we'd probably prefer there wasn't an extra
blank line from the beginning, but it is not worth a patch noise to
remove it" sense).  Even though they are somewhat distracting,
seeing them after finishing reading the real changes is much less
bad than having to see distracting changes in the middle of real
changes during a review.

Thanks.

^ permalink raw reply	[flat|nested] 3+ messages in thread

end of thread, other threads:[~2017-11-17  5:09 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2017-11-16 17:05 [PATCH 0/1] RUNTIME_PREFIX on POSIX systems Dan Jacques
2017-11-16 17:05 ` [PATCH 1/1] exec_cmd: RUNTIME_PREFIX on some " Dan Jacques
2017-11-17  5:08   ` Junio C Hamano

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).