Reducing code size of Position Independent Executables (PIE) by shrinking the size of dynamic relocations section

["Sriraman Tallam via binutils" <binutils at sourceware dot org>]

We identified a problem with PIE executables, more than 5% code size
bloat compared to non-PIE and we have a few proposals to reduce the
bloat.  Please take a look and let us know what you think.

* What is the problem?

PIE is a security hardening feature that enables ASLR (Address Space
Layout Randomization) and enables the executable to be loaded at a
random virtual address upon every execution instance. On an average, a
binary when built as PIE is larger by 5% to 9%, as measured on a suite
of benchmarks used at Google where the average text size is ~100MB,
when compared to the one built without PIE.  This is also independent
of the target architecture and we found this to be true for x86_64,
arm64 and power.  We noticed that the primary reason for this code
size bloat is due to the extra dynamic relocations that are generated
in order to make the binary position independent.  This proposal
introduces new ways to represent these dynamic relocations that can
reduce the code size bloat to just a few percent.

As an example,  to show the bloat in code size, here is the data from
one of our larger  binaries,

Without PIE, the binary’s code size in bytes is this as displayed by
the ‘size’ command:

 text             data            bss           dec
504663285 16242884 9130248 530036417

With PIE, the binary’s code size in bytes is this as displayed by the
‘size’ command:

 text            data           bss           dec
539781977 16242900 9130248 565155125

The text size of the binary grew by 7% and the total size by 6.6%.
Our experiments have shown that the binary sizes grow anywhere from 5%
to 9%  with PIE on almost all benchmarks we looked at.  Notice that
almost all the code bloat comes from the “text” segment of the binary,
which contains the executable code of the application and any
read-only data.  We looked into this segment to see why this is
happening and found that the size of the section that contains the
dynamic relocations for a binary explodes with PIE.  For instance,
without PIE, for the above binary the dynamic relocation section
contains 46 entries whereas with PIE, the same section contains
1463325 entries.  It takes 24 bytes to store one entry, that is 3
integer values each of size 8 bytes.  So, the dynamic relocations
alone need an extra space of (1463325 - 46) * 8 bytes which is 35
million bytes which is almost all the bloat incurred!.

* What are these extra dynamic relocations that are created for PIE executables?

We noticed that these extra relocations for PIE binaries have a common
pattern and are needed for the reason that it is not known until
run-time where the binary will be loaded.  All of these extra dynamic
relocations are of the ELF type R_X86_64_RELATIVE.   Let us show using
an example what these relocations do.
Let us take an example of a program that stores the address of a global:

#include <stdio.h>

const int a = 10;

const int *b = &a;

int main() {

 printf (“b = %p\n”, b);

}

First, let us look at the binary built without PIE.  Let’s look at the
data section where ‘b’ and ‘a’ are allocated.

00000000004007d0 <a>:
 4007d0:       0a 00


0000000000401b10 <b>:
 401b10:       d0 07
 401b12:       40 00 00

Variable ‘a’ is allocated at address 0x4007d0 which matches the output
when running the binary.  ‘b’ is allocated at address 0x401b10 and its
contents in little-endian byte order is the address of ‘a’.

Now, lets us examine the contents of the PIE binary:

00000000000008d8 <a>:
8d8:   0a 00

0000000000001c50 <b>:
   1c50:       d8 08
                    1c50: R_X86_64_RELATIVE *ABS*+0x8d8
   1c52:       00 00
   1c54:       00 00


Notice there is a dynamic relocation here which tells the dynamic
linker that this value needs to be fixed at run-time.  This is needed
because ASLR can load this binary anywhere in the address space and
this relocation fixes the address after it is loaded.


* More details about R_X86_64_RELATIVE relocations

This relocation is worth 24 bytes  and has three fields

Offset

Type - here it is R_X86_64_RELATIVE

Addend (what extra value needs to be added)

The offset field of this relocation is the address offset from the
start where this relocation applies.  The type field indicates the
type of the dynamic relocation but we are interested in particularly
one type of dynamic relocation, R_X86_64_RELATIVE.   This is important
because in the motivating example that we presented above, all the
extra dynamic relocations were of this type!


* We have these proposals to reduce the size of the dynamic relocations section:


Idea A: Shrinking the size of the dynamic relocations

1. Use the offset of the relocation to store the addend

2. Create a separate section for storing R_X86_64_RELATIVE type
relocations, “.pie.dyn”.

3. Create a new relocation type, R_X86_64_RELATIVE_OFFSET that has
exactly one field, the offset.

4. The type field is not necessary as these is only one type of
relocation in this new section.

5. Store the addend at the offset, which has enough space to store 8
byte addends.

6. The dynamic linker can be thought to read the value from this
offset for the addend rather than reading this from the entry.

7. This new section is just a sequence of offsets which are
R_X86_64_RELATIVE relocations.  The dynamic linker needs to be taught
to apply these relocations correctly.

8. Figure out a way to either prevent this from running or warning at
startup when trying to run this program with old dynamic linkers.

9. Eliminating two-thirds of the relocation reduces the size of these
sections by 66%.  Additionally, this section can also be compressed
and a compression factor of 10 is possible which makes the total size
of these sections 3% to 4% of the original.

10. The decompression by the dynamic linker can be done in chunks to
keep the memory overhead fixed and small.  Notice that the
decompressed chunk’s contents can be discarded after the relocation is
applied.

Idea B: Compressing the dynamic relocation section into .zrela.dyn

1. Simpler to implement.

2. The dynamic linker needs to decompress and apply these relocations
without  memory overhead, which otherwise defeats the purpose of doing
this, by decompressing in chunks.

3. The decompressed chunks can be discarded as and when the
relocations are applied so a fixed size buffer for decompression
should work well.

4. Figure out a way to either prevent this from running or warning at
startup when trying to run this program with unsupported dynamic
linkers.

5. Compression factors that can be achieved with gzip are a factor of
10 which makes the total size 10% of the original.


Thanks to Sterling Augustine, Ian Lance Taylor, Paul Pluzhnikov, and
David Li for the numerous  inputs!  Please let us know what you think
and what would be suitable for implementation. Also, attaching the
notes from having this discussion on binutils before I sent it here:

Notes from Peter Collingbourne:


* You can make the observation that relative relocations are normally
aligned to the machine's pointer size, and they normally pertain to
consecutive entries in a table of addresses, such as a vtable. That
hints at a more efficient scheme where instead of storing offsets you
store the difference between the offset and the previous one, divided
by the machine's pointer size. That, combined with a simple run-length
encoding scheme (see below) should be enough. Any unaligned
relocations can be represented using regular R_*_RELATIVE relocations.

* What I had in mind for backwards compatibility was that you could
have the linker arrange for the binary to contain code that applies
the relocations when the image is loaded, but only if the dynamic
loader hasn't applied them itself.

i.e. you could have something like

static void *foo = &foo;

void apply_relocs() {
  if (foo != &foo) {
    //apply relocations
  }
}

such that apply_relocs would be the first thing called by the dynamic
loader before the user's initialization code.

Of course, if you didn't want to support old dynamic loaders at all,
you could exit() inside the body of the if statement instead of
applying relocations.

* I don't think you need any sort of standard compression for this.
The run-length encoding scheme I had in mind would represent the first
group of consecutive relocations in the file as a pair of values
(offset of the relocation / machine pointer size, number of
consecutive relocations), and subsequent groups as the pair (relative
offset from previous group of consecutive relocations / machine
pointer size, number of consecutive relocations) all using the uleb128
encoding. So for example, the relocation list:

0xc8
0xd0
0xd8
0xe0
0x100
0x108
0x110

would be encoded as: uleb128(0xc8 / 8) uleb128(4) uleb128((0x100-0xe0)
/ 8) uleb128(3)

i.e. 4 bytes.

In one experiment, I extracted the list of relative relocations from a
Chromium binary and compressed them using this scheme. There were
90302 relocations in total, and this scheme compressed them down to
65488 bytes, which would be 3% of the original for a rela section
(this experiment was actually conducted using an ARM binary, but I
would expect similar results for x86_64).

Notes from Rui Ueyama:

* Windows DLLs have the same problem and already solves it in their
own way, so let me share my knowledge about how DLLs work. It might be
useful for you.

* On Windows, each DLL is linked against a fixed base address and
usually loaded to that address. However, if there's already another
DLL that overlaps in the address space, the Windows loader has to
relocate it. To do that, Windows DLLs contain ".reloc" sections which
contain offsets that need to be fixed up at runtime. If the Windows
loader find that a DLL cannot be loaded to its desired base address,
it loads it to somewhere else, and add <actual base address> -
<desired base address> to each offset that is specified by the .reloc
section.

* In ELF terms, .reloc sections contains arrays of relocation offsets.
All these relocations in the section are implicitly R_X86_64_RELATIVE,
and addends are read from section contents (so it is REL as opposed to
RELA).

* This already reduce the size of relocations to 1/3, but Windows does
more to reduce .reloc size (probably because it was invented for late
'80s or '90s PCs.) Offsets in .reloc are grouped by page where page
size is wordsize minus 16 bits, and offsets sharing the same page
address are stored consecutively to represent them with less space.
This is a very similar technique as the page table which is grouped by
(multiple stages of) pages.

* For example, let's say we have 0x00030, 0x00500, 0x01000, 0x01100,
0x20004, and 0x20008 in a .reloc section. In the section, they are
represented like this:

  0x00000  -- takes 6 bytes
    0x0030 -- each takes 2 bytes
    0x0500
    0x1000
    0x1100
 0x20000 -- takes 6 bytes
    0x0004 -- each takes 2 bytes
    0x0008

* In total, the .reloc section is 24 bytes. If you store each address
as 8 byte integer, it takes 48 bytes, so this technique indeed reduces
the size of the .reloc section.