Skip to contentSkip to conversation

uninative: call patchelf-uninative only when needed

PATCH In review Martin Jansa 2 participants 3 messages Jun 23, 2023 – Jun 28, 2023

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 uninative: call patchelf-uninative only when needed

mke2fs.real, mkfs.ext2.real, mkfs.ext3.real, mkfs.ext4.real are indentical binary with multiple hardlinks and we end calling patchelf-uninative 4 times even when the interpreter is already set correctly from the build

To avoid corrupted binaries created by patchelf-0.18.0 when set-interpreter is called multiple times (on some systems like ubuntu-18.04 this leads to segfaults elsewhere just ldd complains that the executable is no longer dynamically linked, but doesn't fail when executed).

The issue was reported upstream with mkfs.ext4.real as possible reproducer:
https://github.com/NixOS/patchelf/issues/492#issuecomment-1602862272
alternatively we can revert
https://github.com/NixOS/patchelf/commit/65cdee904431d16668f95d816a495bc35a05a192
and create new uninative release with update patchelf-uninative, but better to wait a bit for upstream to have a look and possibly backport proper fix later, until then this change fixes the mkfs.ext4 issues I was seeing in kirkstone, mickledore, nanbield since uninative-3.9 upgrade, as reported in:
https://lists.openembedded.org/g/openembedded-core/message/182862

---
 meta/classes-global/uninative.bbclass | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)
meta/classes-global/uninative.bbclass +8 -3
index 366f7ac793..b54acdd542 100644
@@ -175,7 +175,12 @@ python uninative_changeinterp () {
             if not elf.isDynamic():
                 continue
 
-            os.chmod(f, s[stat.ST_MODE] | stat.S_IWUSR)
-            subprocess.check_output(("patchelf-uninative", "--set-interpreter", d.getVar("UNINATIVE_LOADER"), f), stderr=subprocess.STDOUT)
-            os.chmod(f, s[stat.ST_MODE])
+            current = subprocess.check_output(("patchelf-uninative", "--print-interpreter", f), stderr=subprocess.STDOUT).decode('utf-8').split('\n')[0]
+            if current != d.getVar("UNINATIVE_LOADER"):
+                bb.debug(2, "Changing interpreter from %s to %s with %s" % (current, d.getVar("UNINATIVE_LOADER"), ("patchelf-uninative", "--set-interpreter", d.getVar("UNINATIVE_LOADER"), f)))
+                os.chmod(f, s[stat.ST_MODE] | stat.S_IWUSR)
+                subprocess.check_output(("patchelf-uninative", "--set-interpreter", d.getVar("UNINATIVE_LOADER"), f), stderr=subprocess.STDOUT)
+                os.chmod(f, s[stat.ST_MODE])
+            else:
+                bb.debug(2, "Interpreter was already set to %s in %s" % (d.getVar("UNINATIVE_LOADER"), f))
 }
Signature · 1 line
2.41.0
#1 lore raw
Quoted 35 lines · Martin Jansa

mke2fs.real, mkfs.ext2.real, mkfs.ext3.real, mkfs.ext4.real are indentical binary with multiple hardlinks and we end calling patchelf-uninative 4 times even when the interpreter is already set correctly from the build

To avoid corrupted binaries created by patchelf-0.18.0 when set-interpreter is called multiple times (on some systems like ubuntu-18.04 this leads to segfaults elsewhere just ldd complains that the executable is no longer dynamically linked, but doesn't fail when executed).

The issue was reported upstream with mkfs.ext4.real as possible reproducer:
https://github.com/NixOS/patchelf/issues/492#issuecomment-1602862272
alternatively we can revert
https://github.com/NixOS/patchelf/commit/65cdee904431d16668f95d816a495bc35a05a192
and create new uninative release with update patchelf-uninative, but better to wait a bit for upstream to have a look and possibly backport proper fix later, until then this change fixes the mkfs.ext4 issues I was seeing in kirkstone, mickledore, nanbield since uninative-3.9 upgrade, as reported in:
https://lists.openembedded.org/g/openembedded-core/message/182862

---
 meta/classes-global/uninative.bbclass | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)
meta/classes-global/uninative.bbclass +0 -2
index 366f7ac793..b54acdd542 100644
@@ -175,7 +175,12 @@ python uninative_changeinterp () {
             if not elf.isDynamic():
                 continue
 
-            os.chmod(f, s[stat.ST_MODE] | stat.S_IWUSR)
-            subprocess.check_output(("patchelf-uninative", "--set-interpreter", d.getVar("UNINATIVE_LOADER"), f), stderr=subprocess.STDOUT)

meta/classes-global/uninative.bbclass python uninative_changeinterp () { · line 186

Quoted 10 lines
meta/classes-global/uninative.bbclass +8 -1
-            os.chmod(f, s[stat.ST_MODE])
+            current = subprocess.check_output(("patchelf-uninative", "--print-interpreter", f), stderr=subprocess.STDOUT).decode('utf-8').split('\n')[0]
+            if current != d.getVar("UNINATIVE_LOADER"):
+                bb.debug(2, "Changing interpreter from %s to %s with %s" % (current, d.getVar("UNINATIVE_LOADER"), ("patchelf-uninative", "--set-interpreter", d.getVar("UNINATIVE_LOADER"), f)))
+                os.chmod(f, s[stat.ST_MODE] | stat.S_IWUSR)
+                subprocess.check_output(("patchelf-uninative", "--set-interpreter", d.getVar("UNINATIVE_LOADER"), f), stderr=subprocess.STDOUT)
+                os.chmod(f, s[stat.ST_MODE])
+            else:
+                bb.debug(2, "Interpreter was already set to %s in %s" % (d.getVar("UNINATIVE_LOADER"), f))
 }

I know why you've done it this way but ideally we should make patchelf handle this internally? It would be nice to avoid fork calls if we can in the long run.

Cheers,

Richard

#2 lore raw
Quoted 58 lines · Richard Purdie
Quoted 1 line · Martin Jansa

mke2fs.real, mkfs.ext2.real, mkfs.ext3.real, mkfs.ext4.real are

indentical

Quoted 4 lines

binary with multiple hardlinks and we end calling patchelf-uninative 4 times even when the interpreter is already set correctly from the build

To avoid corrupted binaries created by patchelf-0.18.0 when

set-interpreter

Quoted 5 lines

is called multiple times (on some systems like ubuntu-18.04 this leads to segfaults elsewhere just ldd complains that the executable is no longer dynamically linked, but doesn't fail when executed).

The issue was reported upstream with mkfs.ext4.real as possible

reproducer:

Quoted 3 lines

https://github.com/NixOS/patchelf/issues/492#issuecomment-1602862272
alternatively we can revert

https://github.com/NixOS/patchelf/commit/65cdee904431d16668f95d816a495bc35a05a192

Quoted 7 lines

and create new uninative release with update patchelf-uninative, but better to wait a bit for upstream to have a look and possibly backport proper fix later, until then this change fixes the mkfs.ext4 issues I was seeing in kirkstone, mickledore, nanbield since uninative-3.9 upgrade, as reported in:
https://lists.openembedded.org/g/openembedded-core/message/182862

Review

Quoted 6 lines
---
 meta/classes-global/uninative.bbclass | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)
(quoted diff) +0 -0

b/meta/classes-global/uninative.bbclass

meta/classes-global/uninative.bbclass python uninative_changeinterp () { · line 179

Quoted 9 lines

index 366f7ac793..b54acdd542 100644

meta/classes-global/uninative.bbclass +0 -2
@@ -175,7 +175,12 @@ python uninative_changeinterp () {
             if not elf.isDynamic():
                 continue
 
-            os.chmod(f, s[stat.ST_MODE] | stat.S_IWUSR)
-            subprocess.check_output(("patchelf-uninative",

"--set-interpreter", d.getVar("UNINATIVE_LOADER"), f), stderr=subprocess.STDOUT)

meta/classes-global/uninative.bbclass python uninative_changeinterp () { · line 178

Quoted 2 lines
meta/classes-global/uninative.bbclass +1 -1
-            os.chmod(f, s[stat.ST_MODE])
+            current = subprocess.check_output(("patchelf-uninative",

"--print-interpreter", f), stderr=subprocess.STDOUT).decode('utf-8').split('\n')[0]

meta/classes-global/uninative.bbclass python uninative_changeinterp () { · line 180

Quoted 2 lines
meta/classes-global/uninative.bbclass +2 -0
+            if current != d.getVar("UNINATIVE_LOADER"):
+                bb.debug(2, "Changing interpreter from %s to %s with

%s" % (current, d.getVar("UNINATIVE_LOADER"), ("patchelf-uninative", "--set-interpreter", d.getVar("UNINATIVE_LOADER"), f)))

meta/classes-global/uninative.bbclass python uninative_changeinterp () { · line 182

Quoted 2 lines
meta/classes-global/uninative.bbclass +2 -0
+                os.chmod(f, s[stat.ST_MODE] | stat.S_IWUSR)
+                subprocess.check_output(("patchelf-uninative",

"--set-interpreter", d.getVar("UNINATIVE_LOADER"), f), stderr=subprocess.STDOUT)

meta/classes-global/uninative.bbclass python uninative_changeinterp () {

Quoted 3 lines
meta/classes-global/uninative.bbclass +3 -0
+                os.chmod(f, s[stat.ST_MODE])
+            else:
+                bb.debug(2, "Interpreter was already set to %s in %s" %
Quoted 7 lines

(d.getVar("UNINATIVE_LOADER"), f))

meta/classes-global/uninative.bbclass python uninative_changeinterp () { · line 186

Quoted 1 line
meta/classes-global/uninative.bbclass +0 -0
 }

I know why you've done it this way but ideally we should make patchelf handle this internally? It would be nice to avoid fork calls if we can in the long run.

yes, but that would require new uninative tarball to be built and applied in master, mickledore, kirkstone, dunfell, so I was waiting for feedback from upstream first before changing patchelf and using this as work around to be able to use uninative 3.9 and 4.0 until patchelf is fixed properly.