Make ld-elf.so.1 directly executable

[kib]

Summary

The map_object.c change fixes the AT_EXECFD functionality, which is the prerequisite for the main part of the patch. If the mapped object is linked at specific address, we must obey it. Before, only in-kernel ELF image activator needed to keep the mapping address, since only binaries are linked at the fixed address, and binaries were always mapped by kernel.

For rtld.c bits, check if passed phdr is actually phdr of the interpreter itself, and decide that this is the case of direct execution. In this case, the binary to activate is specified in the argv[1]. After opening it, shift down on-stack structure with argv, env and aux vectors to emulate execution of the binary and not of the interpreter.

Right now, the calculation of execpath in the direct execution mode is too naive, it affects $ORIGIN. Might be, ld-elf.so.1 should search the binary in $PATH, but I am not sure. Anyway, this patch is a good initial step, I believe.

Test Plan

Booted multiuser on amd64, tested manually i386 ld-elf.so.1 with compat32.

[emaste]

libexec/rtld-elf/rtld.c

342–353

Perhaps commit the re-sorting independently first?

487

As an aside it seems O_VERIFY is not documented in open.2. I see that this came from https://github.com/freebsd/freebsd/pull/27 - paging @stevek

Also, what happens if we try to ld-elf.so.1 foo where foo has ---x--x--x perms?

[kib]

pooma% ls -l ~/build/bsd/DEV/stuff/tests/args
---x--x--x  1 kostik  kostik  6234 Apr 18  2010 /usr/home/kostik/build/bsd/DEV/stuff/tests/args
pooma% LD_LIBRARY_PATH=/usr/lib32 obj-i386/i386.i386/usr/home/kostik/work/build/bsd/DEV/src/libexec/rtld-elf/ld-elf.so.1 ~/build/bsd/DEV/stuff/tests/args 1 2 3
Fatal error

I.e. as expected, userspace needs to read executing binary.

libexec/rtld-elf/rtld.c

342–353

This is not just resorting, I moved locals from the nested blocks and added some variables as well.

I can produce separate patch to reorder existing local section, but the bits in the patch are not extractable.

487

It is something internal from J that got committed but never get the flesh in our repository.

[emaste]

libexec/rtld-elf/rtld.c

342–353

Right, I mean there is both resorting and other changes and I think the resorting part could be committed by itself - it would make it more clear what the other changes are.

[kib]

Re-merge after the locals reordering changes were committed.

[emaste]

libexec/rtld-elf/rtld.c

481

extra space before aux_info?

494–495

I think this comment could be expanded a bit to separately explain main_arg[cv] and the argv/envp/auxp manipulation that follows it.

[kib]

Update comment, remove stray space after copy.
Correct (remove) re-calculation of main_argv, it was shifted right too much.
Noted by emaste.

[emaste]

This looks good to me.

[emaste]

libexec/rtld-elf/rtld.c

381

With the patch this assert triggers via

% /libexec/ld-elf.so.1 /libexec/ld-elf.so.1
ld-elf.so.1: assert failed: /usr/home/emaste/src/freebsd/libexec/rtld-elf/rtld.c:381

[kib]

Fixes.

Move the shift before first getenv() call in the rtld, to not depend on details of getenv() implementation.
Do not forget to reset 'env' after environment is moved,not only environ (this caused env -i /libexec/ld-elf.so.1 /bin/ls segfault in jemalloc initializer).
Fix off-by-half in copying the auxv (this caused the assertion reported by Ed).

[emaste]

libexec/rtld-elf/rtld.c

444

!= NULL?

also envp below

[kib]

s/0/NULL/

[kan]

very unlikely I'll look at this any time soon.

[stevek]

libexec/rtld-elf/rtld.c

487

See D2902, which has the changes. I guess after all this time people have had enough time to comment. I'll see about committing that set of changes.

[emaste]

libexec/rtld-elf/rtld.c

487

Ah. I added a comment in there - I'd suggest we should go ahead with the open.2 changes right now, since the corresponding source changes were made some time ago, and follow up with a D2902 rebase.

[stevek]

libexec/rtld-elf/rtld.c

487

Sure, makes sense. I committed just the open.2 change and will rebase D2902.