Patchwork insane.bbclass: Fix RPATH warning in the face of funny path strings

login
register
mail settings
Submitter Andy Ross
Date Aug. 15, 2012, 10:46 p.m.
Message ID <1345070782-4518-2-git-send-email-andy.ross@windriver.com>
Download mbox | patch
Permalink /patch/34675/
State New
Headers show

Comments

Andy Ross - Aug. 15, 2012, 10:46 p.m.
In toolchain edge cases it's possible for the RPATH of a library to be
set to something like "/usr/lib/../lib".  This should be detected as
"/usr/lib" and generate a warning.  Also clarify the warning text to
indicate potential link-time pollution from the host libraries.

Signed-off-by: Andy Ross <andy.ross@windriver.com>
---
 meta/classes/insane.bbclass | 22 ++++++++++++++++++----
 1 file changed, 18 insertions(+), 4 deletions(-)
Chris Larson - Aug. 16, 2012, 12:14 a.m.
On Wed, Aug 15, 2012 at 3:46 PM, Andy Ross <andy.ross@windriver.com> wrote:
> diff --git a/meta/classes/insane.bbclass b/meta/classes/insane.bbclass
> index 556a176..ade0616 100644
> --- a/meta/classes/insane.bbclass
> +++ b/meta/classes/insane.bbclass
> @@ -161,6 +161,17 @@ def package_qa_check_rpath(file,name, d, elf, messages):
>              if dir in line:
>                  messages.append("package %s contains bad RPATH %s in file %s" % (name, line, file))
>
> +def rpath_norm(s):
> +    import re
> +    s = re.sub('[^/]+/\.\.(/|$)', '/', s) # snip ".." components
> +    s = re.sub('/\.(/|$)', '/', s)        # snip "." components
> +    s = re.sub('/+', '/', s)              # snip repeated slashes
> +    s = re.sub('/
> , '', s)               # snip trailing slash
> +    return s
> +
> +def rpath_eq(a, b):
> +    return rpath_norm(a) == rpath_norm(b)
> +

Please just use os.path.normpath() rather than reinventing the wheel here.
Phil Blundell - Aug. 16, 2012, 8:39 a.m.
On Wed, 2012-08-15 at 15:46 -0700, Andy Ross wrote:
> In toolchain edge cases it's possible for the RPATH of a library to be
> set to something like "/usr/lib/../lib".  This should be detected as
> "/usr/lib" and generate a warning.  Also clarify the warning text to
> indicate potential link-time pollution from the host libraries.

If these RPATHs are causing host pollution then that sounds like there
is another bug elsewhere.  They ought to be resolved relative to the
sysroot during link edit.

It's also tempting to say that we should, unconditionally, normalise any
RPATHs which contain .. or . components before insane.bbclass runs.
Having extra, redundant, components in the pathname will just cause
extra work for ld.so (and hence run-time slowness) and bring no benefit.
Doing that would mean that the existing check would catch things like
the case you mentioned automatically.

p.
Andy Ross - Aug. 16, 2012, 4:10 p.m.
Chris Larson wrote:
> Please just use os.path.normpath() rather than reinventing the wheel here.

Learn something new every day.  Fixed.

Andy
Andy Ross - Aug. 16, 2012, 5:43 p.m.
On 08/16/2012 01:39 AM, Phil Blundell wrote:
> If these RPATHs are causing host pollution then that sounds like there
> is another bug elsewhere.  They ought to be resolved relative to the
> sysroot during link edit.

Indeed, this is turning out to be a deeper issue than I wanted it to
be.  What seems to be is happening is this: an RPATH in the ELF header
is interpreted relative to sysroot normally.  But when linking, a
-rpath argument to the ld is interpreted relative to the *host*
filesystem even when there is a --sysroot argument.  See the quick
test script below.

As it happens, both of those are ultimately produced by libtool, and
it's only the second one that is fatal.  The RPATH itself is probably
still a warning condition, but it's not a build breaker.  But neither
is needed, they are happening in the case I'm looking at only because
libtool (I think) is confused by the "/../" syntax in the link path
into trying to add an RPATH where one isn't needed.

The specific case I'm looking at (with our internal tree, I'm working
on reproducing vs. poky right now) is with a pulseaudio build, where
an x86-64 build on an x86_64 host hits a RPATH in libgdk-x11-2.0 that
pulls in the host libXranr and fails due to version skew.  I was just
pointed at a discussion from last week on owl_video which looks all
but identical.

At least for the moment I'm going to try to track down the libtool
issue (maybe sanify the path before it sees if if possible) instead of
trying to fix a linker bug.

Andy

Reproducer for the underlying linker issue:

#!/bin/sh
set -ex

SYSROOT=/home/andy/oe/poky/build/tmp/sysroots/qemux86-64
CC=/home/andy/oe/poky/build/tmp/sysroots/x86_64-linux/usr/bin/x86_64-poky-linux/x86_64-poky-linux-gcc

# A library:
echo 'int foo(){return 0;}' > foo.c
$CC -shared -fPIC -o libfoo.so foo.c

# Another library that references the first via RPATH=`pwd`
echo 'int foo(); int bar(){return foo();}' > bar.c
$CC -shared -fPIC -o libbar.so bar.c -L. -lfoo -Wl,-rpath -Wl,`pwd`

# A program that uses the second library:
echo 'int bar(); int main(){return bar();}' > main.c

# Works (incorrectly!) because the "-rpath `pwd" argument here is
# *not* interpreted relative to sysroot, so the linker sees
# ./libfoo.so as a potential library.
$CC --sysroot=$SYSROOT -L$SYSROOT/usr/lib64 -Wl,-rpath -Wl,`pwd` -o main2 main.c libbar.so

echo THAT LINK SHOULD HAVE FAILED

# Fails (correctly) because nothing on the link line tells libbar.so
# how to find libfoo.so, nor does libfoo.so exist in the
# sysroot-relative RPATH.
$CC --sysroot=$SYSROOT -o main main.c libbar.so
Richard Purdie - Aug. 17, 2012, 10:28 a.m.
On Thu, 2012-08-16 at 10:43 -0700, Andy Ross wrote:
> On 08/16/2012 01:39 AM, Phil Blundell wrote:
> > If these RPATHs are causing host pollution then that sounds like there
> > is another bug elsewhere.  They ought to be resolved relative to the
> > sysroot during link edit.
> 
> Indeed, this is turning out to be a deeper issue than I wanted it to
> be.  What seems to be is happening is this: an RPATH in the ELF header
> is interpreted relative to sysroot normally.  But when linking, a
> -rpath argument to the ld is interpreted relative to the *host*
> filesystem even when there is a --sysroot argument.  See the quick
> test script below.
> 
> As it happens, both of those are ultimately produced by libtool, and
> it's only the second one that is fatal.  The RPATH itself is probably
> still a warning condition, but it's not a build breaker.  But neither
> is needed, they are happening in the case I'm looking at only because
> libtool (I think) is confused by the "/../" syntax in the link path
> into trying to add an RPATH where one isn't needed.
> 
> The specific case I'm looking at (with our internal tree, I'm working
> on reproducing vs. poky right now) is with a pulseaudio build, where
> an x86-64 build on an x86_64 host hits a RPATH in libgdk-x11-2.0 that
> pulls in the host libXranr and fails due to version skew.  I was just
> pointed at a discussion from last week on owl_video which looks all
> but identical.
> 
> At least for the moment I'm going to try to track down the libtool
> issue (maybe sanify the path before it sees if if possible) instead of
> trying to fix a linker bug.

I suspect you need to look somewhere around:

http://git.yoctoproject.org/cgit.cgi/poky/tree/meta/recipes-devtools/libtool/libtool/fix-rpath.patch

and normalise the rpaths being used by libtool. I think it has some kind
of path normalisation function somewhere...

Cheers,

Richard
Andy Ross - Aug. 17, 2012, 3:02 p.m.
On 08/17/2012 03:28 AM, Richard Purdie wrote:
> I suspect you need to look somewhere around:
>
> http://git.yoctoproject.org/cgit.cgi/poky/tree/meta/recipes-devtools/libtool/libtool/fix-rpath.patch

Indeed.  Found that yesterday and am currently testing a fix.
Should be inbound within the hour.

Andy

Patch

diff --git a/meta/classes/insane.bbclass b/meta/classes/insane.bbclass
index 556a176..ade0616 100644
--- a/meta/classes/insane.bbclass
+++ b/meta/classes/insane.bbclass
@@ -161,6 +161,17 @@  def package_qa_check_rpath(file,name, d, elf, messages):
             if dir in line:
                 messages.append("package %s contains bad RPATH %s in file %s" % (name, line, file))
 
+def rpath_norm(s):
+    import re
+    s = re.sub('[^/]+/\.\.(/|$)', '/', s) # snip ".." components
+    s = re.sub('/\.(/|$)', '/', s)        # snip "." components
+    s = re.sub('/+', '/', s)              # snip repeated slashes
+    s = re.sub('/$', '', s)               # snip trailing slash
+    return s
+
+def rpath_eq(a, b):
+    return rpath_norm(a) == rpath_norm(b)
+
 QAPATHTEST[useless-rpaths] = "package_qa_check_useless_rpaths"
 def package_qa_check_useless_rpaths(file, name, d, elf, messages):
     """
@@ -181,10 +192,13 @@  def package_qa_check_useless_rpaths(file, name, d, elf, messages):
     	m = rpath_re.match(line)
 	if m:
 	   rpath = m.group(1)
-	   if rpath == libdir or rpath == base_libdir:
-	      # The dynamic linker searches both these places anyway.  There is no point in
-	      # looking there again.
-	      messages.append("%s: %s contains probably-redundant RPATH %s" % (name, package_qa_clean_path(file, d), rpath))
+	   if rpath_eq(rpath, libdir) or rpath_eq(rpath, base_libdir):
+	      # The dynamic linker searches both these places anyway.
+	      # There is no point in looking there again.  And it may
+	      # be harmful: the RPATH may point to host directories
+	      # (e.g. /usr/lib) which will be searched at link time,
+	      # creating build issues.
+	      messages.append("%s: %s contains probably-redundant, possibly host-polluted RPATH %s" % (name, package_qa_clean_path(file, d), rpath))
 
 QAPATHTEST[dev-so] = "package_qa_check_dev"
 def package_qa_check_dev(path, name, d, elf, messages):