workaround for debugedit segv
cover workaround for debugedit segv
[ YOCTO #4089 ]
This is an attempt at a workaround for the debugedit segv.
The problem is that during the creation of the build-id hash, the system iterates over the various ELF sections passing a point to the loaded data and size of each section to a hash function.
When it gets to the hash for the .plt and .bss section, the size is non-zero, while the data is point to 0. This immediately causes a segfault on access of data at address 0.
See the bug and patch comments for more diagnostic information.
RFC PATCH 1/2 rpm: Add workaround for debugedit-segv
[ YOCTO #4089 ]
On PPC and MIPS, there appears to be a condition that causes debugedit to segfault. The segfault is related to a call into the md5hash algorithm, an address of '0', and a size > 0 is passed causing the access of the address to segv.
This workaround may prove to be the final fix, but it's currently unclear what the actual cause of the 0 address is.
--- meta/recipes-devtools/rpm/rpm/debugedit-segv.patch | 35 ++++++++++++++++++++++ meta/recipes-devtools/rpm/rpm_5.4.9.bb | 3 +- 2 files changed, 37 insertions(+), 1 deletion(-) create mode 100644 meta/recipes-devtools/rpm/rpm/debugedit-segv.patch
meta/recipes-devtools/rpm/rpm/debugedit-segv.patch new file +35 -0
@@ -0,0 +1,35 @@ +There are cases, especially on PPC and MIPS, where the data address +returned is 0, but the size is not 0. + +It appears to happen when the sections headers are similar to: + + [21] .data PROGBITS 000239c0 0139c0 000010 00 WA 0 0 8 + [22] .got PROGBITS 000239d0 0139d0 000014 04 WAX 0 0 4 + [23] .plt NOBITS 000239e4 0139e4 000234 00 WAX 0 0 4 + [24] .bss NOBITS 00023c18 0139e4 0001c8 00 WA 0 0 8 + [25] .comment PROGBITS 00000000 0139e4 000011 01 MS 0 0 1 + [26] .debug_aranges PROGBITS 00000000 0139f8 000d68 00 0 0 8 + +Sections 23 and 24 (.plt and .bss) which are NOBITS have a loaded data address +of 0, but a size != 0. + +This could be a bug in libelf... + +Upstream-status: Pending + +Signed-off-by: Mark Hatle <[email protected]> + +Index: rpm-5.4.9/tools/debugedit.c +=================================================================== +--- rpm-5.4.9.orig/tools/debugedit.c ++++ rpm-5.4.9/tools/debugedit.c +@@ -1434,7 +1434,8 @@ handle_build_id (DSO *dso, Elf_Data *bui + auto inline void process (const void *data, size_t size) + { + memchunk chunk = { .data = (void *) data, .size = size }; +- hashFunctionContextUpdateMC (&ctx, &chunk); ++ if (data != NULL && size != 0) ++ hashFunctionContextUpdateMC (&ctx, &chunk); + } + union + {
meta/recipes-devtools/rpm/rpm_5.4.9.bb +2 -1
@@ -84,7 +84,8 @@ SRC_URI = "http://www.rpm5.org/files/rpm/rpm-5.4/rpm-5.4.9-0.20120508.src.rpm;ex file://python-rpm-rpmsense.patch \ file://rpm-reloc-macros.patch \ file://rpm-platform2.patch \ - file://rpm-remove-sykcparse-decl.patch \ + file://rpm-remove-sykcparse-decl.patch \ + file://debugedit-segv.patch \ " # Uncomment the following line to enable platform score debugging
Signature · 1 line
1.8.1.2.545.g2f19ada
Re: [PATCH 1/2 RFC] rpm: Add workaround for debugedit-segv
meta/recipes-devtools/rpm/rpm/debugedit-segv.patch line 14
Quoted 2 lines · Mark Hatle
meta/recipes-devtools/rpm/rpm/debugedit-segv.patch new file +2
-0+Sections 23 and 24 (.plt and .bss) which are NOBITS have a loaded data address +of 0, but a size != 0.
That doesn't seem like totally unreasonable behaviour for a NOBITS section. What were you expecting libelf to do in that case?
++ if (data != NULL && size != 0)
++ hashFunctionContextUpdateMC (&ctx, &chunk);
I suppose one could argue that allocating a chunk of zero-filled memory of the right size and then hashing that would be a slightly better fix. Whether it matters in practice or not would depend on what exactly is going into this hash and what it's being used for.
p.
Re: [PATCH 1/2 RFC] rpm: Add workaround for debugedit-segv
Quoted 7 lines · Phil Blundell
meta/recipes-devtools/rpm/rpm/debugedit-segv.patch line 14
Quoted 2 lines · Mark Hatle
meta/recipes-devtools/rpm/rpm/debugedit-segv.patch new file +2
-0+Sections 23 and 24 (.plt and .bss) which are NOBITS have a loaded data address +of 0, but a size != 0.That doesn't seem like totally unreasonable behaviour for a NOBITS section. What were you expecting libelf to do in that case?
Quoted 7 lines
++ if (data != NULL && size != 0)
++ hashFunctionContextUpdateMC (&ctx, &chunk);I suppose one could argue that allocating a chunk of zero-filled memory of the right size and then hashing that would be a slightly better fix. Whether it matters in practice or not would depend on what exactly is going into this hash and what it's being used for.
It appears in the past it either didn't load the section at all, or the size was set to 0.
It's a combination of the data pointer set to NULL and the size != 0 that is causing the segfault. This doesn't appear to happen outside of PPC and MIPS.
I'm going to look into identifying if the section is a NOBITS and skipping the whole operation if it is.
--Mark
Re: [PATCH 1/2 RFC] rpm: Add workaround for debugedit-segv
Quoted 2 lines · Mark Hatle
I'm going to look into identifying if the section is a NOBITS and skipping the whole operation if it is.
That would mean that a change in the size of the .bss wouldn't have any impact on the hash. Maybe that's fine for your application though, I dunno.
p.
Re: [PATCH 1/2 RFC] rpm: Add workaround for debugedit-segv
Quoted 7 lines · Phil Blundell
Quoted 2 lines · Mark Hatle
I'm going to look into identifying if the section is a NOBITS and skipping the whole operation if it is.
That would mean that a change in the size of the .bss wouldn't have any impact on the hash. Maybe that's fine for your application though, I dunno.
I'm not completely familiar with how the buildid is calculated, other then an md5 hash over the sections themselves. I can't think of a case where the contents of the sections wouldn't change along with the .bss and .plt size.
(buildid is supposed to be used by gdb to find/verify that the debuginfo matches the binary actually being debugged...)
Through my google search, I never found a full spec of what the buildid was supposed to contain, only a high-level description from an old Fedora 8 work item. (Note, as long the buildid is consistent it really doesn't appear to matter how it was generated, just that both the app and debuginfo contain it.)
--Mark
Quoted 3 lines
p.
Re: [PATCH 1/2 RFC] rpm: Add workaround for debugedit-segv
Quoted 13 lines · Mark Hatle
Quoted 7 lines · Phil Blundell
Quoted 2 lines · Mark Hatle
I'm going to look into identifying if the section is a NOBITS and skipping the whole operation if it is.
That would mean that a change in the size of the .bss wouldn't have any impact on the hash. Maybe that's fine for your application though, I dunno.
I'm not completely familiar with how the buildid is calculated, other then an md5 hash over the sections themselves. I can't think of a case where the contents of the sections wouldn't change along with the .bss and .plt size.
Quoted 7 lines
(buildid is supposed to be used by gdb to find/verify that the debuginfo matches the binary actually being debugged...)
Through my google search, I never found a full spec of what the buildid was supposed to contain, only a high-level description from an old Fedora 8 work item. (Note, as long the buildid is consistent it really doesn't appear to matter how it was generated, just that both the app and debuginfo contain it.)
I've looked at the code some more. It does checksum the header itself and then if it's got contents, it also adds the contents to the checksum.. That is where the failure appears to be happening:
if (u.shdr.sh_type != SHT_NOBITS)
{
Elf_Data *d = elf_rawdata (dso->scn[i], NULL);
if (d == NULL)
goto bad;
process (d->d_buf, d->d_size);
}So it's specifically checking for SHT_NOBITS, but it's matching so it falls through and d->d_buf == 0, causing the failure. I'll keep investigating, but somehow that value (u.shdr.sh_type) is wrong [or at least unexpected!].
--Mark
Quoted 13 lines
--Mark
Quoted 3 lines
p.
_______________________________________________
Openembedded-core mailing list
[email protected]
http://lists.linuxtogo.org/cgi-bin/mailman/listinfo/openembedded-core
Re: [PATCH 1/2 RFC] rpm: Add workaround for debugedit-segv
Quoted 12 lines · Mark Hatle
I've looked at the code some more. It does checksum the header itself and then if it's got contents, it also adds the contents to the checksum.. That is where the failure appears to be happening:
if (u.shdr.sh_type != SHT_NOBITS) { Elf_Data *d = elf_rawdata (dso->scn[i], NULL); if (d == NULL) goto bad; process (d->d_buf, d->d_size); }
Quoted 3 lines
So it's specifically checking for
SHT_NOBITS, but it's matching so it falls through andd->d_buf== 0, causing the failure. I'll keep investigating, but somehow that value (u.shdr.sh_type) is wrong [or at least unexpected!].
Ah. If a NOBITS section is getting reported as something else then it seems that this must clearly be a bug in libelf and ought to be fixed there rather than working around it in rpm. What do you actually get as u.shdr.sh_type?
p.
Re: [PATCH 1/2 RFC] rpm: Add workaround for debugedit-segv
Quoted 13 lines · Phil Blundell
Quoted 12 lines · Mark Hatle
I've looked at the code some more. It does checksum the header itself and then if it's got contents, it also adds the contents to the checksum.. That is where the failure appears to be happening:
if (u.shdr.sh_type != SHT_NOBITS) { Elf_Data *d = elf_rawdata (dso->scn[i], NULL); if (d == NULL) goto bad; process (d->d_buf, d->d_size); }
Quoted 8 lines
Quoted 3 lines
So it's specifically checking for
SHT_NOBITS, but it's matching so it falls through andd->d_buf== 0, causing the failure. I'll keep investigating, but somehow that value (u.shdr.sh_type) is wrong [or at least unexpected!].Ah. If a NOBITS section is getting reported as something else then it seems that this must clearly be a bug in libelf and ought to be fixed there rather than working around it in rpm. What do you actually get as u.shdr.sh_type?
I posted an updated patch last night. debugedit was re-translating (byte swapping) the header elements during the buildid calculations. This was causing the value of sh_type to be in the ELF binaries native endian during the check. So big endian binaries failed to match the SHT_NOBITS, thus the problem being discovered on PPC and MIPS.
(Fix was to not translate the original values, but instead only translate a copy. I verified the produced buildid was expected, as I now have a better understanding of how the buildid is generated.)
--Mark
Quoted 2 lines
p.
RFC PATCH 2/2 package.bbclass: Trigger a bb.error if split/strip fails
In reply to Mark Hatle[ YOCTO #4089 ]
If debugedit, objcopy calls fail, capture the error and return a failure to the user.
--- meta/classes/package.bbclass | 24 ++++++++++++++++++++---- 1 file changed, 20 insertions(+), 4 deletions(-)
meta/classes/package.bbclass +20 -4
@@ -254,14 +254,26 @@ def splitdebuginfo(file, debugfile, debugsrcdir, d): # We need to extract the debug src information here... if debugsrcdir: - subprocess.call("'%s' -b '%s' -d '%s' -i -l '%s' '%s'" % (debugedit, workparentdir, debugsrcdir, sourcefile, file), shell=True) + cmd = "'%s' -b '%s' -d '%s' -i -l '%s' '%s'" % (debugedit, workparentdir, debugsrcdir, sourcefile, file) + ret = subprocess.call(cmd, shell=True) + if ret != 0: + bb.error("debugedit failed: %s - %d" % (cmd, ret)) + return ret bb.utils.mkdirhier(os.path.dirname(debugfile)) - subprocess.call("'%s' --only-keep-debug '%s' '%s'" % (objcopy, file, debugfile), shell=True) + cmd = "'%s' --only-keep-debug '%s' '%s'" % (objcopy, file, debugfile) + ret = subprocess.call(cmd, shell=True) + if ret != 0: + bb.error("debugedit failed: %s - %d" % (cmd, ret)) + return ret # Set the debuglink to have the view of the file path on the target - subprocess.call("'%s' --add-gnu-debuglink='%s' '%s'" % (objcopy, debugfile, file), shell=True) + cmd = "'%s' --add-gnu-debuglink='%s' '%s'" % (objcopy, debugfile, file) + ret = subprocess.call(cmd, shell=True) + if ret != 0: + bb.error("debugedit failed: %s - %d" % (cmd, ret)) + return ret if newmode: os.chmod(file, origmode)
@@ -806,7 +818,9 @@ python split_and_strip_files () { bb.utils.mkdirhier(os.path.dirname(fpath)) #bb.note("Split %s -> %s" % (file, fpath)) # Only store off the hard link reference if we successfully split! - splitdebuginfo(file, fpath, debugsrcdir, d) + ret = splitdebuginfo(file, fpath, debugsrcdir, d) + if ret != 0: + return ret # Hardlink our debug symbols to the other hardlink copies for file in hardlinks:
@@ -878,6 +892,8 @@ python split_and_strip_files () { # # End of strip # + + return 0 } python populate_packages () {
Signature · 1 line
1.8.1.2.545.g2f19ada
Quoted 8 lines · Mark Hatle
[ YOCTO #4089 ]
This is an attempt at a workaround for the debugedit segv.
The problem is that during the creation of the build-id hash, the system iterates over the various ELF sections passing a point to the loaded data and size of each section to a hash function.
Quoted 5 lines
When it gets to the hash for the .plt and .bss section, the size is non-zero, while the data is point to 0. This immediately causes a segfault on access of data at address 0.
See the bug and patch comments for more diagnostic information.
I really want to figure out how widespread issues are so I merged the workaround and my version of the error handling patch so we start seeing where the remaining problems are (if any).
I'm hoping to test this out on the autobuilder, see how we stand.
If further improvements are found to the workaround we can merge those on top of the existing tree.
Cheers,
Richard