When a binary has a PT_GNU_RELRO segment, the elfhack injected code
uses mprotect to add the writable flag to relocated pages before
applying relocations, removing it afterwards. To do so, the elfhack
program uses the location and size of the PT_GNU_RELRO segment, and
adjusts it to be aligned according to the PT_LOAD alignment.
The problem here is that the PT_LOAD alignment doesn't necessarily match
the actual page alignment, and the resulting mprotect may end up not
covering the full extent of what the dynamic linker has protected
read-only according to the PT_GNU_RELRO segment. In turn, this can lead
to a crash on startup when trying to apply relocations to the still
read-only locations.
Practically speaking, this doesn't end up being a problem on x86, where
the PT_LOAD alignment is usually 4096, which happens to be the page
size, but on Debian armhf, it is 64k, while the run time page size can be
4k.
lld is being too smart for its own good, and places non-relocatable data
right after the program headers, which prevents the program headers from
growing. But elfhack wasn't checking for that, so happily placed the
non-relocatable data at its non-relocated location, overwriting the last
item of the program headers.
Bug 1385783 changed things such that the two elfhack sections are not
adjacent anymore. They can even be in different segments in some cases,
but the undo code doesn't know how to actually handle that case.
So for now, allow non adjacent sections, but still verify that they are
in the same segment.
In bug 635961, elfhack was made to (ab)use the bss section as a
temporary space for a pointer. To find it, it scanned writable PT_LOAD
segments to find one that has a different file and memory size,
indicating the presence of .bss. This usually works fine, but when
the binary is linked with lld and relro is enabled, the end of the
file-backed part of the PT_LOAD segment containing the .bss section
ends up in the RELRO segment, making that location read-only and
subsequently making the elfhacked binary crash when it tries to restore
the .bss to a clean state, because it's not actually writing in the .bss
section: lld page aligns it after the RELRO segment.
So instead of scanning PT_LOAD segments, we scan for SHT_NOBITS
sections that are not SHF_TLS (i.e. not .tbss).
The lld linker creates separate segments for purely executable sections
(such as .text) and sections preceding those (such as .rel.dyn). Neither
gold nor bfd ld do that, and just put all those sections in the same
executable segment.
Since elfhack is putting its executable code between the two relocation
sections, it ends up in a non-executable segment, leading to a crash
when it's time to run that code.
We thus insert the elfhack code before the first executable section
instead of between the two relocation sections (which is where the
elfhack data lies, and stays).
Since bug 635961, building with relro makes elfhack try to use the bss
data for a temporary function pointer. If there is not enough space for
a pointer in the bss, elfhack will complain it couldn't find the bss.
In normal circumstances, this is most likely fine. Libraries with a bss
so small that it can't fit a pointer are already too small to be
elfhacked anyways. In Firefox, the two libraries with the smallest bss
have enough space for two pointers, and aren't elfhacked (libmozgtk.so
and libplds4.so).
However, the testcase that is used during the build to validate that
elfhack works doesn't have a large enough bss on x86-64, making elfhack
bail out, and the build fail as a consequence.
This, in turn, is due to the only non-thread-local zeroed data being an
int, which is not enough to fit a pointer on x86-64. We thus make it a
size_t.
The PT_PHDR segment is optional, but the Android toolchain decides to
create one in some cases, and places it first. When that happens, the
work around for bug 1233963 fails, because the fake phdr section has not
been adjusted yet (it only happens when we see a PT_LOAD).
So we adjust the fake phdr section when we see a PT_PHDR segment (and
avoid re-updating it when we see a subsequent PT_LOAD).
Somehow, with the Android toolchain, we end up with
non-empty-but-really-empty DT_INIT_ARRAYs.
In practical terms, they are arrays with no relocations, and content
that is meaningless:
$ objdump -s -j .init_array libnss3.so
libnss3.so: file format elf32-little
Contents of section .init_array:
1086e0 00000000 ....
$ readelf -r libnss3.so | grep 1086e0
$ objdump -s -j .init_array libplugin-container-pie.so
libplugin-container-pie.so: file format elf32-little
Contents of section .init_array:
4479c ffffffff 00000000 ffffffff 00000000 ................
$ readelf -r libplugin-container-pie.so | grep 4479c
Because so far, elfhack expected meaningful DT_INIT_ARRAYs, it bailed out
early in that case.
When the clang plugin is used, building something during export needs to
happen after the plugin is built. But there is no dependency ensuring
this happens.
OTOH, these sources in elfhack/inject don't need to be built that early,
so we'll just leave to the build system to build it at a proper time.
This removes the unnecessary setting of c-basic-offset from all
python-mode files.
This was automatically generated using
perl -pi -e 's/; *c-basic-offset: *[0-9]+//'
... on the affected files.
The bulk of these files are moz.build files but there a few others as
well.
MozReview-Commit-ID: 2pPf3DEiZqx
We have very few directories where we have SOURCES declared that are not
part of a library or program in some way. In fact, there is only one
where it is legitimate because we only use the object file
(build/unix/elfhack/inject). Others are the result of moz.build control
flow (see e.g. netwerk/standalone), and we end up building more objects
than we need to.
There are other cases where we need objects without actually linking
them anywhere, but there are other sources in the same directory, and a
corresponding Linkable is emitted. And in fact, the only case I knew
about (media/libvpx), doesn't use such objects since bug 1151175.
Its only purpose is to disable PGO. Where that was not already explicitly done,
or irrelevant (because the directory only contains python), I disabled it in
moz.build.
LTO changes in GCC5 make it break the assumptions made in the test-ctors
source. The simplest and more correct thing to do is to just force the
test files never to be built with LTO. They are meant to have a particular
layout that LTO might break at any time anyways.
There are, sadly, many combinations of linkage in use throughout the tree.
The main differentiator, though, is between program/libraries related to
Gecko or not. Kind of. Some need mozglue, some don't. Some need dependent
linkage, some standalone.
Anyways, these new templates remove the need to manually define the
right dependencies against xpcomglue, nspr, mozalloc and mozglue
in most cases.
Places that build programs and were resetting MOZ_GLUE_PROGRAM_LDFLAGS
or that build libraries and were resetting MOZ_GLUE_LDFLAGS can now
just not use those Gecko-specific templates.
