preload.c is, or should be, mostly unnecessary

[stephenrkell]

I believe I recently "fixed" #132. But the underlying problem was that preload.c is a very confusing file. It dates from a long time ago, before we had systrapping, when we relied on preloading to catch mmap and the like. But since that did not catch all mmap calls, we added systrapping. Yet we kept this file because it was doing some other things too, including relating to malloc/alloca instrumentation. Why do its functions exist?

• __assert_fail (protected) -- reentrancy avoidance (this is an mmap-free implementation)
• abort (protected; debug-only) -- for debugging, delay termination of the process
• malloc_or_alloca_usable_size -- obsolete, but retained for debugging
• malloc_usable_size -- receives cross-DSO calls to the global malloc_usable_size (global malloc et al are in user2hook.o, but that does not generate a malloc_usable_size hook)
• mmap -- trap-free hooking with nudging
• mmap64 -- trap-free hooking with nudging
• munmap -- trap-free hooking
• mremap -- trap-free hooking with nudging
• signal -- prevent SIGILL handler clobbering -- should be systrap
• sigaction -- prevent SIGILL handler clobbering -- should be systrap
• memcpy -- to support shadow memory (bulk copy notification)
• memmove -- to support shadow memory (bulk copy notification)

Now that we have systrapping, preload.c exists mainly to provide a no-trap path to these system calls. How important is that? We could just call the libc and let it trap.

The bug was that we were doing that already, in addition! Our preloaded mremap was calling the real libc, not a raw syscall. So it was being trapped, and hooked there. But also, our preload wrapper was running the hook. The hook running twice was not good for mremap. It is harmless for munmap but was happening for that too.

[stephenrkell]

I think the right solution is:

• keep __assert_fail and abort
• move the signal and sigaction stuff to the systrapping side
• drop memcpy and memmove -- we leave this to the actual shadow-memory client (libcrunch)
• provide a new name for our internal mmap (raw_mmap_with_nudge)?
• have libmallochooks hook malloc_usable_size

[stephenrkell]

The __assert_fail and abort calls make the rump of preload.c; so for clarity, rename to nomalloc-libc.c? We could morally move fake_dlsym (et al?) here too, but that/tehy may be best left in relf.h with just a comment here.