Created static pic libc with no SSP to be used by rtld

[lffpires_ruabrasil.org]

Summary

Created libc_nossp_pic.a to be used when linking rtld.

[kib]

This looks fine to me, in particular the rtld changes. But I am not competent to review the build changes.

lib/libc/Makefile

204

I suggest to make this a separate change.

[bdrewery]

share/mk/bsd.lib.mk

312

I don't think we need the MK_TOOLCHAIN check here.

[brooks]

This seems fine though I'm not super excited about compiling all of libc yet another time.

share/mk/bsd.lib.mk

312

It looks like it's just a copy of the case above.

[bdrewery]

In D15283#322252, @brooks wrote:

This seems fine though I'm not super excited about compiling all of libc yet another time.

Same. The benefit seems low so far if it is just removing a small hack from rtld. Perhaps a limited libc without all of the stuff rtld obviously doesn't need.

Also I suspect this is happening because lib/libc/secure/stack_protector.c was recently updated. We could easily use some ifdefs in that file and directly build it again from rtld with a different -DRTLD define to use a different implementation. Thus fixing having the logic in 2 separate files like before.

Is there something more coming where libc_nossp_pic will be useful?

[kib]

In D15283#322290, @bdrewery wrote:

In D15283#322252, @brooks wrote:

This seems fine though I'm not super excited about compiling all of libc yet another time.

Same. The benefit seems low so far if it is just removing a small hack from rtld. Perhaps a limited libc without all of the stuff rtld obviously doesn't need.

Also I suspect this is happening because lib/libc/secure/stack_protector.c was recently updated. We could easily use some ifdefs in that file and directly build it again from rtld with a different -DRTLD define to use a different implementation. Thus fixing having the logic in 2 separate files like before.

Is there something more coming where libc_nossp_pic will be useful?

I diagree with the characterization of the issue as small or as a hack. Rtld is both very important and very different from the normal runtime environment so that its correctness must be maintained properly. I would rather argue that the removed code is the hack (which I am guilty of). I do think that libc_nossp_pic is worth it, at least because that it provides the fix which does not depend on the implementation details of the stack protector.

Also, I suggest to splut the change into two parts. One should be the modifications to the build infrastructure, and the other the rtld changes.

[lffpires_ruabrasil.org]

• Separated libc restriction- and rtld-related changes
• Removed unnecessary MK_TOOLCHAIN check

[lffpires_ruabrasil.org]

Made changes according to review comments.

If the reviewers feel that this is a good approach, this revision is now ready to be committed.

I'll post 2 separate revisions for the following 2 changes:
Removed -fstack-protector-all restriction from libc (introduced in rS180012).

Removed rtld-private implementations of stack_chk_guard,
stack_chk_fail() and __chk_fail() symbols (introduced in rS232861), since rtld
is now using libc_nossp_pic.a, which doesn't have SSP.

[lffpires_ruabrasil.org]

2 follow up revisions created:
D15340 (Removed -fstack-protector-all restriction from libc)
D15341 (Updated rtld to use libc_nossp_pic.a)

[brooks]

LGTM.