Re: [BUG] `make install' partly ignores `NO_INSTALL_HARDLINKS'

["Đoàn Trần Công Danh" <congdanhqx@gmail.com>]

On 2020-08-17 09:40:23-0700, Junio C Hamano <gitster@pobox.com> wrote:
> Đoàn Trần Công Danh  <congdanhqx@gmail.com> writes:
> 
> > On 2020-08-14 10:26:24-0700, Junio C Hamano <gitster@pobox.com> wrote:
> >> > Since both git-citool and git-gui will be installed into same
> >> > directory "$(libexecdir)", I think it would make more sense to use:
> >> >
> >> > 	LN = ln -s
> >> >
> >> > here instead?
> >> 
> >> In the top-level Makefile, INSTALL_SYMLINKS make macro does exist,
> >> but it is not exported to submakes.  If it were, something like
> >> 
> >>     ifdef INSTALL_SYMLINKS
> >> 	LN = ln -s
> >>     else
> >>     ifdef NO_INSTALL_HARDLINKS
> >> 	LN = cp
> >>     else
> >> 	LN = ln
> >>     endif
> >>     endif
> >> 
> >> might become possible, but you'd need to audit what is fed to $(LN)
> >> at the locations the macro is used and make necessary adjustment
> >> accordingly.  "cp A ../B" or "ln A ../B" will make a usable copy of
> >> file A appear inside ../B directory, but "ln -s A ../B" will not,
> >> and I didn't see if all uses of $(LN) was to give synonyms to what
> >> is already installed, or some of them were truly installing from the
> >> build location when I gave the "something along this line" example.
> >
> > Yes, the top-level Makefile seems to have a special branch for
> > BUILT_INS, in which, we will create symlink for those builtin in
> > libexecdr if NO_INSTALL_HARDLINKS is defined.
> 
> Did you mean pieces like this?
> 
> 	for p in $(filter $(install_bindir_programs),$(BUILT_INS)); do \
> 		$(RM) "$$bindir/$$p" && \
> 		test -n "$(INSTALL_SYMLINKS)" && \
> 		ln -s "git$X" "$$bindir/$$p" || \
> 		{ test -z "$(NO_INSTALL_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 && \

Yes, I meant this piece.

> The symlinks happen ONLY when INSTALL_SYMLINKS is asked for.

Not what I understand from that code.
When `INSTALL_SYMLINKS` is not asked, the shell will jump to this block:

 		{ test -z "$(NO_INSTALL_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; } \

When NO_INSTALL_HARDLINKS is asked, the shell will jump to last
2 clauses:

 		  ln -s "git$X" "$$bindir/$$p" 2>/dev/null || \
 		  cp "$$bindir/git$X" "$$bindir/$$p" || exit; } \

shell try to "ln -s" first, when it failed to symlink, cp will be used.

In fact, this's what we used for packaging in distro, we often ask for:

	NO_INSTALL_HARDLINKS

> Not all filesystems support symbolic links,

This is correct, but most sane package manager will do the right thing.

> hardlinks never suffer from
> dangling link problem, and often they are cheaper.

Part of the reason we don't ask for INSTALL_SYMLINKS,
because with INSTALL_SYMLINKS, symlink will be make from
$(libexecdir)/git to $(bindir)/git.

But, NO_INSTALL_HARDLINKS will only symlink inside $(libexecdir), e.g:

	$ ls -al $(/usr/bin/git --exec-path )/git
	-rwxr-xr-x 1 root root 3193704 2020-08-12 21:37 /usr/libexec/git-core/git
	$ ls -al $(/usr/bin/git --exec-path )/git-ls-files
	lrwxrwxrwx 1 root root 3 2020-08-12 21:37 /usr/libexec/git-core/git-ls-files -> git

And those symlinks are packaged and managed by package manager,
so, no, no dangling problem.


-- 
Danh