Add test for patching a system call instruction

Add test for patching a system call instruction

This tests patching a specific, fixed instruction sequence, whereas
the existing tests just test whatever is in the version of glibc on
the system.

Change library.cc to make the patching code easier to test: Split out
a patchSystemCallsInRange() method, since the existing methods assume
they are operating on a whole dynamically-loaded ELF object.  Change
maps_ to be per-instance so that creating Library objects with
different Maps objects works.

This doesn't try to cover the various corner cases in library.cc yet.

BUG=http://code.google.com/p/seccompsandbox/issues/detail?id=17
TEST=test_patching_syscall

Committed: http://code.google.com/p/seccompsandbox/source/detail?r=177

[Markus (顧孟勤), 2011-11-18 18:16:46]

lgtm

http://codereview.chromium.org/8596009/diff/3001/tests/test_patching.cc
File tests/test_patching.cc (right):

http://codereview.chromium.org/8596009/diff/3001/tests/test_patching.cc#newco...
tests/test_patching.cc:23: char *page_start = (char *) ((uintptr_t) start &
~(getpagesize() - 1));
You should probably round "end" up to the next page boundary. That way, you do
the right thing if your function straddles a page boundary. And you also make
sure that mprotect always gets a "len" parameter that is a multiple of the page
size. I am actually surprised the kernel doesn't check this already.

[Mark Seaborn, 2011-11-18 18:28:49]

On 18 November 2011 10:16, <markus@chromium.org> wrote:

> lgtm
>
>
> http://codereview.chromium.**org/8596009/diff/3001/tests/**
>
test_patching.cc<http://codereview.chromium.org/8596009/diff/3001/tests/test_patching.cc>
> File tests/test_patching.cc (right):
>
> http://codereview.chromium.**org/8596009/diff/3001/tests/**
>
test_patching.cc#newcode23<http://codereview.chromium.org/8596009/diff/3001/tests/test_patching.cc#newcode23>
> tests/test_patching.cc:23: char *page_start = (char *) ((uintptr_t)
> start & ~(getpagesize() - 1));
> You should probably round "end" up to the next page boundary. That way,
> you do the right thing if your function straddles a page boundary. And
> you also make sure that mprotect always gets a "len" parameter that is a
> multiple of the page size. I am actually surprised the kernel doesn't
> check this already.
>

I believe the kernel rounds the size up for you.  The Linux man page
specifically says "addr must be aligned to a page boundary" but does not
say the same about len.  The same goes for
http://pubs.opengroup.org/onlinepubs/7908799/xsh/mprotect.html.

I'm using "end - page_start" not "end - start" so it should do the right
thing if the function cross a page boundary.

It's the same with the args for mmap() and munmap().  I observe that ld.so
passes unaligned sizes to mmap() sometimes.

Cheers,
Mark