[PATCH] Makefile: add a knob to turn off hardlinks within same directory

Jonathan Nieder· Oct 12, 2011, 08:38 UTC · lore

The git builtins in $(gitexecdir) are implemented as hard links to a single git binary by default, so even the overhead of symlink resolution is not needed to run them. However, the trick can be harmful, in two cases:

 - on Windows, some tools to estimate directory size hugely
   overestimate the size of git (each hardlink counts as taking up the
   same amount of space as git.exe)

 - various filesystems have limits on the number of hardlinks that
   can be made to a particular file --- Linux's LINK_MAX is 127,
   _POSIX_LINK_MAX is 8, and btrfs has a limit of 4096 /
   ($len_filename + 8) or so links to a given inode in a single
   directory.

Normally that second case is not a problem (when ln fails, "make install" falls back to "ln -s"), but if git is tar'ed up on one filesystem and then extracted on a more limited one, it can result in "Too many links" errors.

Nowadays people are encouraged to (and typically do) run builtins using the "git" command name directly rather than those dashed forms in scripts, making the use of hardlinks for the dashed forms a somewhat pointless optimization. Introduce a new knob to allow people to turn it off with a simple "make install NO_HARDLINKS=YesPlease".

Typically someone using this setting would want to set NO_CROSS_DIRECTORY_HARDLINKS, too, but that is not enforced, so you can make $(bindir)/git and $(gitexecdir)/git into hardlinks to the same inode and still make sure your tarball avoids the btrfs limits if you want.

Requested-by: Bastian Blank <waldi@debian.org>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
---
Hi,

See <http://bugs.debian.org/642603> for context.  Sane?

 Makefile |    7 +++++++
 1 files changed, 7 insertions(+), 0 deletions(-)

Show changes to Makefile +7 −0

diff --git a/Makefile b/Makefile
index 9afdcf2a..ab64ff4c 100644
--- a/Makefile
+++ b/Makefile
@@ -226,6 +226,10 @@ all::
 # Define NO_CROSS_DIRECTORY_HARDLINKS if you plan to distribute the installed
 # programs as a tar, where bin/ and libexec/ might be on different file systems.
 #
+# Define NO_HARDLINKS if you plan to distribute the installed programs as a tar
+# that might be extracted on a filesystem like btrfs that does not cope well
+# with many links to one inode in one directory.
+#
 # Define USE_NED_ALLOCATOR if you want to replace the platforms default
 # memory allocators with the nedmalloc allocator written by Niall Douglas.
 #
@@ -2326,12 +2330,14 @@ endif
 	} && \
 	for p in $(filter $(install_bindir_programs),$(BUILT_INS)); do \
 		$(RM) "$$bindir/$$p" && \
+		test -z "$(NO_HARDLINKS)" && \
 		ln "$$bindir/git$X" "$$bindir/$$p" 2>/dev/null || \
 		ln -s "git$X" "$$bindir/$$p" 2>/dev/null || \
 		cp "$$bindir/git$X" "$$bindir/$$p" || exit; \
 	done && \
 	for p in $(BUILT_INS); do \
 		$(RM) "$$execdir/$$p" && \
+		test -z "$(NO_HARDLINKS)" && \
 		ln "$$execdir/git$X" "$$execdir/$$p" 2>/dev/null || \
 		ln -s "git$X" "$$execdir/$$p" 2>/dev/null || \
 		cp "$$execdir/git$X" "$$execdir/$$p" || exit; \
@@ -2339,6 +2345,7 @@ endif
 	remote_curl_aliases="$(REMOTE_CURL_ALIASES)" && \
 	for p in $$remote_curl_aliases; do \
 		$(RM) "$$execdir/$$p" && \
+		test -z "$(NO_HARDLINKS)" && \
 		ln "$$execdir/git-remote-http$X" "$$execdir/$$p" 2>/dev/null || \
 		ln -s "git-remote-http$X" "$$execdir/$$p" 2>/dev/null || \
 		cp "$$execdir/git-remote-http$X" "$$execdir/$$p" || exit; \

-- 
1.7.7

Junio C Hamano· Oct 12, 2011, 19:02 UTC · re: Jonathan Nieder · lore

Re: [PATCH] Makefile: add a knob to turn off hardlinks within same directory

Jonathan Nieder <jrnieder@gmail.com> writes:

Show 28 quoted lines

> Typically someone using this setting would want to set
> NO_CROSS_DIRECTORY_HARDLINKS, too, but that is not enforced, so you
> can make $(bindir)/git and $(gitexecdir)/git into hardlinks to the
> same inode and still make sure your tarball avoids the btrfs limits if
> you want.
>
> Requested-by: Bastian Blank <waldi@debian.org>
> Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
> ---
> Hi,
>
> See <http://bugs.debian.org/642603> for context.  Sane?
>
>  Makefile |    7 +++++++
>  1 files changed, 7 insertions(+), 0 deletions(-)
>
> diff --git a/Makefile b/Makefile
> index 9afdcf2a..ab64ff4c 100644
> --- a/Makefile
> +++ b/Makefile
> @@ -226,6 +226,10 @@ all::
>  # Define NO_CROSS_DIRECTORY_HARDLINKS if you plan to distribute the installed
>  # programs as a tar, where bin/ and libexec/ might be on different file systems.
>  #
> +# Define NO_HARDLINKS if you plan to distribute the installed programs as a tar
> +# that might be extracted on a filesystem like btrfs that does not cope well
> +# with many links to one inode in one directory.
> +#

Hmm.... I would understand if the change were to eventually remove these "git-foo" anywhere from the filesystem perhaps after a large version bump in Git 2.0, but that is not what you are trying to do.

Wouldn't your use case be better served with

    $ tar zcf dist.tar.gz --hard-dereference $list_of_files_to_tar_up

instead?

Jonathan Nieder· Oct 12, 2011, 19:57 UTC · re: Junio C Hamano · lore

Re: [PATCH] Makefile: add a knob to turn off hardlinks within same directory

Junio C Hamano wrote:

Show 5 quoted lines

> Wouldn't your use case be better served with
>
>     $ tar zcf dist.tar.gz --hard-dereference $list_of_files_to_tar_up
>
> instead?

No, because that duplicates the files instead of making symlinks.

patchMakefile: add a knob to turn off hardlinks within same directory