Bug 439046 - valgrind is unusably large when linked with lld
Summary: valgrind is unusably large when linked with lld
Status: RESOLVED FIXED
Alias: None
Product: valgrind
Classification: Developer tools
Component: general (show other bugs)
Version: unspecified
Platform: Other Linux
: NOR normal
Target Milestone: ---
Assignee: Julian Seward
URL:
Keywords:
: 426135 (view as bug list)
Depends on:
Blocks:
 
Reported: 2021-06-23 05:26 UTC by Matt Turner
Modified: 2021-10-09 13:06 UTC (History)
4 users (show)

See Also:
Latest Commit:
Version Fixed In:
Sentry Crash Report:


Attachments

Note You need to log in before you can comment on or make changes to this bug.
Description Matt Turner 2021-06-23 05:26:32 UTC
clang/lld pads out the .text section by the amount given by the -Ttext=... option. This produces enormous executables in valgrind (memcheck-amd64-linux is ~1.4GB). This is the result of valgrind's configure.ac testing for linker options -Ttext-segment=... which lld does not support and falling back to -Ttext=... which was subject to a breaking change in behavior in lld-10 (see [1]).

The patch at https://github.com/paulfloyd/freebsd_valgrind/commit/9748df5f6a0cefd5b4a4beeeffe1d3668611e561 modifies configure.ac to test for lld's --image-base=... option which provides the desired behavior.

I dug into this when I found valgrind to be unusably large (~14GB) on ChromeOS. See [2]. Searching for relevant linker arguments with "valgrind" and "lld" lead me to the GitHub patch.

It would be great to have this upstream sooner rather than later, because that greatly simplifies the justification for cherry-picking the patch in downstream distributions.

[1] https://releases.llvm.org/10.0.0/tools/lld/docs/ReleaseNotes.html
[2] https://issuetracker.google.com/issues/191520718
Comment 1 Mark Wielaard 2021-06-23 10:09:58 UTC
(In reply to Matt Turner from comment #0)
> [2] https://issuetracker.google.com/issues/191520718

That requires some kind of google account to view.
Could you post your analysis somewhere public or in this bug report?
Comment 2 Paul Floyd 2021-06-23 19:25:02 UTC
Here is a copy/paste of the text.

Created issue.

tl;dr
ChromeOS's clang/lld pads out the .text section by the amount given by the -Ttext=... option. This produces enormous executables in the dev-util/valgrind package. gcc/ld does not do this on ChromeOS, nor does clang/lld on vanilla Gentoo.

Reproduction steps:

$ cat boom.c
int _start () { return 0; }
$ x86_64-cros-linux-gnu-gcc -static -nodefaultlibs -nostartfiles -Wl,-Ttext=0x58000000 boom.c -o boom
$ ls -lh boom
-rwxrwxr-x 1 msturner primarygroup 9.0K Jun 19 02:20 boom
$ x86_64-cros-linux-gnu-clang -static -nodefaultlibs -nostartfiles -Wl,-Ttext=0x58000000 boom.c -o boom
$ ls -lh boom
-rwxrwxr-x 1 msturner primarygroup 1.4G Jun 19 02:20 boom
Background
When built on ChromeOS with clang/lld, dev-util/valgrind takes far too much disk space, making it too large to deploy a DUT:

emerge-lakitu valgrind
[...]
 * Final size of build directory: 14706448 KiB (14.0 GiB)
 * Final size of installed tree:  14514812 KiB (13.8 GiB)
whereas when built with gcc/ld it takes a much more expected amount of space:

CC="x86_64-cros-linux-gnu-gcc" CXX="x86_64-cros-linux-gnu-g++" emerge-lakitu valgrind
[...]
 * Final size of build directory: 401664 KiB (392.2 MiB)                                                                                                                                                           
 * Final size of installed tree:  181876 KiB (177.6 MiB)                                                                                                                                                           
This is because a handful of static executables (e.g., memcheck-amd64-linux) end up being ~1.4GB each in the clang/lld build and only a few MB when built with gcc/ld.

Those static executables are linked with a linker option that positions the .text section start address at 0x58000000 (1.375GB). See this for rationale.

The configure script tests for and chooses -Ttext-segment=... if available and falls back to -Ttext=... otherwise. LLD does not support -Ttext-segment=...:

$ x86_64-cros-linux-gnu-ld.lld -Ttext-segment=0x58000000
ld.lld: error: -Ttext-segment is not supported. Use --image-base if you intend to set the base address
so it uses -Ttext=... instead.

readelf -a memcheck-amd64-linux on one of the clang/lld-produced 1.4GB files shows what's happening:

$ ls -lh memcheck-amd64-linux 
-rwxrwxr-x 1 msturner portage 1.4G Jun 19 01:58 memcheck-amd64-linux
$ readelf -a memcheck-amd64-linux
[...]
  Entry point address:               0x580ff894
  Start of program headers:          64 (bytes into file)
  Start of section headers:          1481854064 (bytes into file)
[...]
Section Headers:
  [Nr] Name              Type             Address           Offset
       Size              EntSize          Flags  Link  Info  Align
  [ 0]                   NULL             0000000000000000  00000000
       0000000000000000  0000000000000000           0     0     0
  [ 1] .text             PROGBITS         0000000058000000  57e00000
       0000000000196db3  0000000000000000  AX       0     0     8
  [ 2] .rodata           PROGBITS         0000000058196dc0  57f96dc0
       00000000000796d8  0000000000000000 AMS       0     0     16
whereas with gcc/ld.bfd:

$ ls -lh memcheck-amd64-linux 
-rwxrwxr-x 1 msturner portage 12M Jun 19 01:53 memcheck-amd64-linux
$ readelf -a memcheck-amd64-linux
[...]
  Entry point address:               0x580846c5
  Start of program headers:          64 (bytes into file)
  Start of section headers:          12045048 (bytes into file)
[...]
Section Headers:
  [Nr] Name              Type             Address           Offset
       Size              EntSize          Flags  Link  Info  Align
  [ 0]                   NULL             0000000000000000  00000000
       0000000000000000  0000000000000000           0     0     0
  [ 1] .note.gnu.bu[...] NOTE             00000000580001c8  000001c8
       0000000000000024  0000000000000000   A       0     0     4
  [ 2] .text             PROGBITS         0000000058001000  00001000
       000000000016901a  0000000000000000  AX       0     0     16
  [ 3] .rodata           PROGBITS         000000005816b000  0016b000
       000000000006c1e8  0000000000000000   A       0     0     32
That is, using -Ttext=0x58000000 with ChromeOS's lld seems to actually pad out the section headers by roughly 0x58000000 bytes! (See the Offset of the .text section and the Start of section headers)

Notably, using -Ttext=0x58000000 with gcc/ld.bfd does not do this and instead produces a reasonably-sized binary.

I tried reproducing this clang/lld behavior on a plain Gentoo install, and I could not, which leads me to wonder if ChromeOS's clang/lld is configured with some hardening option that causes this...

Modifying valgrind's configure to use --image-base=... instead of -Ttext=... seems to work and produce a reasonably-sized executables (even smaller than gcc/ld).

diff --git a/configure.ac b/configure.ac
index 4582fb5d0..86cd9176b 100755
--- a/configure.ac
+++ b/configure.ac
@@ -2657,7 +2657,7 @@ AC_LINK_IFELSE(
   AC_MSG_RESULT([yes])
 ], [
   linker_using_t_text="yes"
-  AC_SUBST([FLAG_T_TEXT], ["-Ttext"])
+  AC_SUBST([FLAG_T_TEXT], ["--image-base"])
   AC_MSG_RESULT([no])
 ])
 CFLAGS=$safe_CFLAGS
---
 * Final size of build directory: 308972 KiB (301.7 MiB)                                                 
 * Final size of installed tree:  117356 KiB (114.6 MiB)                                                 
$ ls -lh memcheck-amd64-linux 
-rwxrwxr-x 1 msturner portage 7.3M Jun 19 02:12 memcheck-amd64-linux
$ readelf -a memcheck-amd64-linux
[...]
  Entry point address:               0x581792c4
  Start of program headers:          64 (bytes into file)
  Start of section headers:          7556672 (bytes into file)
[...]
Section Headers:$ readelf -a memcheck-amd64-linux

  [Nr] Name              Type             Address           Offset
       Size              EntSize          Flags  Link  Info  Align
  [ 0]                   NULL             0000000000000000  00000000
       0000000000000000  0000000000000000           0     0     0
  [ 1] .rodata           PROGBITS         00000000580001d0  000001d0
       00000000000796d8  0000000000000000 AMS       0     0     16
  [ 2] .eh_frame_hdr     PROGBITS         00000000580798a8  000798a8
       0000000000000034  0000000000000000   A       0     0     4
  [ 3] .eh_frame         PROGBITS         00000000580798e0  000798e0
       000000000000014c  0000000000000000   A       0     0     8
  [ 4] .text             PROGBITS         0000000058079a30  00079a30
       0000000000196db3  0000000000000000  AX       0     0     8
Questions
Why does x86_64-cros-linux-gnu-clang behave this way?
Should valgrind's configure script test for and use --image-base=... if available?
ms...@google.com<ms...@google.com> #2Jun 21, 2021 10:52PM
The FreeBSD port of valgrind contains a patch that looks relevant and fixes the issue I see: https://github.com/paulfloyd/freebsd_valgrind/commit/9748df5f6a0cefd5b4a4beeeffe1d3668611e561

It contains a comment that seems to describe the problem somewhat:

+# - LLVM's ld.lld, for at least versions 8.0 (shipping with FreeBSD 12.1)
+#   and 9.0 support the -Ttext option and behave as desired.  As of
+#   LLVM ld.lld version 10.0 a breaking change made -Ttext unusable,
+#   however the --image-base option has the desired semantics.
+#   It turns out that ld.lld has supported --image-base since at least
+#   as far back as version 8.0.
I'll try to connect with the patch authors and see if I can help get this upstreamed.
Comment 3 Matt Turner 2021-06-23 19:27:15 UTC
(In reply to Mark Wielaard from comment #1)
> (In reply to Matt Turner from comment #0)
> > [2] https://issuetracker.google.com/issues/191520718
> 
> That requires some kind of google account to view.
> Could you post your analysis somewhere public or in this bug report?

Sorry about that, and thanks Paul for reposting. I wrote it in markdown, so here's a copy with the formatting: https://gist.github.com/mattst88/be75c31d0ec33d5cc51124625e410799
Comment 4 Tom Hughes 2021-06-29 08:29:14 UTC
*** Bug 426135 has been marked as a duplicate of this bug. ***
Comment 5 Paul Floyd 2021-10-09 13:06:15 UTC
This should now be fixed with my recent series of commits to merge FreeBSD support.