sip-workaround / trace mode no longer works on arm64 macOS ≥ 13 due to new security features

[reneeotten (Renee Otten)]

Trace-mode breaks the extract phase under Ventura (never seen this on my previous OS, Catalina). As an example I've attached the main.log when running the command sudo port -dvt extract py310-setuptools. Anyone else seeing this issue or is it something specific to my installation?

[as an aside: is there an indication when there will be a buildbot for Ventura available?]

[ryandesign (Ryan Carsten Schmidt)]

Replying to reneeotten:

[as an aside: is there an indication when there will be a buildbot for Ventura available?]

I have no information for you about that.

[kencu (Ken)]

if anyone knows a workaround for this, I'd like to hear it.

As much as possible, I try to use trace mode to verify builds, but this makes it difficult.

Extracting normally and then building in trace mode fails.

[ryandesign (Ryan Carsten Schmidt)]

Has duplicates #66421, #66638, #66698.

[ryandesign (Ryan Carsten Schmidt)]

In #66421 it was suggested to run xcodebuild -runFirstLaunch and that appeared to solve it. Please confirm if that works for you too.

[jmroot (Joshua Root)]

Replying to ryandesign:

In #66421 it was suggested to run xcodebuild -runFirstLaunch and that appeared to solve it.

AIUI, that fixed another error that that user was getting after they stopped using -t.

[linuxgemini (İlteriş Eroğlu)]

Replying to ryandesign:

In #66421 it was suggested to run xcodebuild -runFirstLaunch and that appeared to solve it. Please confirm if that works for you too.

This did not fix the issue for me, even after a full uninstall and reinstall of MacPorts didn't help.

[reneeotten (Renee Otten)]

no, that indeed doesn't help

[linuxgemini (İlteriş Eroğlu)]

After looking at system console logs, I have noticed that SIP is kicking in to block sandbox-exec:

default	11:02:43.231827+0300	kernel	AMFI: Launch Constraint Violation (enforcing), error info: c[1]p[1]m[1]e[3], (Constraint not matched) launching proc[vc: 1 pid: 38987]: /opt/local/var/macports/sip-workaround/503/usr/bin/sandbox-exec, launch type 0, failure proc [vc: 1 pid: 38987]: /opt/local/var/macports/sip-workaround/503/usr/bin/sandbox-exec
default	11:02:43.231920+0300	kernel	ASP: Security policy would not allow process: 38987, /opt/local/var/macports/sip-workaround/503/usr/bin/sandbox-exec
default	11:02:43.238965+0300	tccd	AUTHREQ_ATTRIBUTION: msgID=180.536, attribution={responsible={TCCDProcess: identifier=com.googlecode.iterm2, pid=38987, auid=502, euid=503, responsible_path=/Applications/iTerm.app/Contents/MacOS/iTerm2, binary_path=/opt/local/var/macports/sip-workaround/503/usr/bin/sandbox-exec}, accessing={TCCDProcess: identifier=com.apple.sandbox-exec, pid=38987, auid=502, euid=503, binary_path=/opt/local/var/macports/sip-workaround/503/usr/bin/sandbox-exec}, requesting={TCCDProcess: identifier=com.apple.syspolicyd, pid=180, auid=0, euid=0, binary_path=/usr/libexec/syspolicyd}, },
default	11:02:43.261656+0300	kernel	sandbox-exec[38987] Corpse allowed 1 of 5
default	11:02:46.863394+0300	ReportCrash	Formulating fatal 309 report for corpse[38987] sandbox-exec
default	11:02:46.869412+0300	osanalyticshelper	creating type 309 as /Users/ilteris.eroglu/Library/Logs/DiagnosticReports/.sandbox-exec-2023-02-03-110246.ips
default	11:02:47.033565+0300	osanalyticshelper	Saved type '309(<private>)' report (10 of max 25) at /Users/ilteris.eroglu/Library/Logs/DiagnosticReports/sandbox-exec-2023-02-03-110246.ips
default	11:02:47.033731+0300	osanalyticshelper	xpc log creation type 309 result success: /Users/ilteris.eroglu/Library/Logs/DiagnosticReports/sandbox-exec-2023-02-03-110246.ips
default	11:02:47.033818+0300	ReportCrash	client log create type 309 result success: /Users/ilteris.eroglu/Library/Logs/DiagnosticReports/sandbox-exec-2023-02-03-110246.ips
default	11:02:47.033899+0300	ReportCrash	no MetricKit for process sandbox-exec type 309 bundleId (null)
default	11:02:47.034039+0300	ReportCrash	Sending event: com.apple.stability.crash {"coalitionName":"com.googlecode.iterm2","exceptionCodes":"0x0000000000000000, 0x0000000000000000(\n    0,\n    0\n)EXC_CRASHSIGKILL (Code Signature Invalid)","incidentID":"CDC1A1CF-829A-48C5-9BD2-6C8D471F2D6A","logwritten":1,"process":"sandbox-exec","terminationReasonExceptionCode":"0x4","terminationReasonNamespace":"CODESIGNING"}
default	11:02:47.034143+0300	analyticsd	Received event: com.apple.stability.crash {"coalitionName":"com.googlecode.iterm2","exceptionCodes":"0x0000000000000000, 0x0000000000000000(\n    0,\n    0\n)EXC_CRASHSIGKILL (Code Signature Invalid)","incidentID":"CDC1A1CF-829A-48C5-9BD2-6C8D471F2D6A","logwritten":1,"process":"sandbox-exec","terminationReasonExceptionCode":"0x4","terminationReasonNamespace":"CODESIGNING"}
default	11:02:47.034575+0300	analyticsd	Aggregated. Transform: StabilityCrashNumerator3WithBundleVersion Dirty: 1 Event: com.apple.stability.crash {"coalitionName":"com.googlecode.iterm2","exceptionCodes":"0x0000000000000000, 0x0000000000000000(\n    0,\n    0\n)EXC_CRASHSIGKILL (Code Signature Invalid)","incidentID":"CDC1A1CF-829A-48C5-9BD2-6C8D471F2D6A","logwritten":1,"process":"sandbox-exec","terminationReasonExceptionCode":"0x4","terminationReasonNamespace":"CODESIGNING","timestamp":1675411367033982}
default	11:02:47.035773+0300	analyticsd	Aggregated. Transform: StabilityCrashNumerator3WithIncidentID Dirty: 1 Event: com.apple.stability.crash {"coalitionName":"com.googlecode.iterm2","exceptionCodes":"0x0000000000000000, 0x0000000000000000(\n    0,\n    0\n)EXC_CRASHSIGKILL (Code Signature Invalid)","incidentID":"CDC1A1CF-829A-48C5-9BD2-6C8D471F2D6A","logwritten":1,"process":"sandbox-exec","terminationReasonExceptionCode":"0x4","terminationReasonNamespace":"CODESIGNING","timestamp":1675411367033982}
default	11:02:47.035988+0300	analyticsd	Aggregated. Transform: StabilityCrashNumerator3 Dirty: 1 Event: com.apple.stability.crash {"coalitionName":"com.googlecode.iterm2","exceptionCodes":"0x0000000000000000, 0x0000000000000000(\n    0,\n    0\n)EXC_CRASHSIGKILL (Code Signature Invalid)","incidentID":"CDC1A1CF-829A-48C5-9BD2-6C8D471F2D6A","logwritten":1,"process":"sandbox-exec","terminationReasonExceptionCode":"0x4","terminationReasonNamespace":"CODESIGNING","timestamp":1675411367033982}

Looks like there is an entitlement issue, not quite sure.

[linuxgemini (İlteriş Eroğlu)]

I have been informed by a friend that this is caused by a new feature at AppleMobileFileIntegrity called Launch Constraints(1) which basically blocks Apple system binaries under the sip-workaround folder(2).

(1): ​https://arstechnica.com/gadgets/2022/10/macos-13-ventura-the-ars-technica-review/17/#launch-constraints

(2): ​https://theevilbit.github.io/posts/amfi_launch_constraints/

[kencu (Ken)]

perhaps then the process of copying the system binaries and running the copies to avoid SIP might need to be revisited:

[de1977a709f86b2e663ffc1f43ae70b075fc4e9a/macports-base]

[kencu (Ken)]

might as well bump this up as it looks like it is going to need some thinking from the team

[neverpanic (Clemens Lang)]

Chances are this happens because the binary has a signature from Apple. I'm guessing we will end up having to strip the Apple signature when copying the binary, and then run the unsigned copy. This, of course, comes with the risk of breaking some binaries that would otherwise have required entitlements, so we probably have to skip the copy for binaries with entitlements and just run the original binary at the expense of not having trace mode work on those binaries.

Could anybody test this for me on a Ventura system by copying a binary from /usr/bin to a different place, re-signing it using codesign -f -i -, and then running it?

Trace mode also does not work on aarch64 Monterey either, although I don't know why. It doesn't cause failures, but also does not work as expected.

[kencu (Ken)]

well you can't run the binary if you move it, that is for sure:

cp /usr/bin/make ./make

 % /usr/bin/make --version
GNU Make 3.81
Copyright (C) 2006  Free Software Foundation, Inc.
This is free software; see the source for copying conditions.
There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A
PARTICULAR PURPOSE.

This program built for i386-apple-darwin11.3.0

% ./make --version
zsh: killed     ./make --version

but I can't (so far) figure out how to codesign it:

 1008  codesign -f -i - make
 1009  codesign -f -i make
 1010  codesign -f -i make
 1011  man codesign
 1012  codesign -f make
 1013  codesign -d make

nothing works (yet).

[kencu (Ken)]

% codesign -f -i - ./make
Usage: codesign -s identity [-fv*] [-o flags] [-r reqs] [-i ident] path ... # sign
       codesign -v [-v*] [-R=<req string>|-R <req file path>] path|[+]pid ... # verify
       codesign -d [options] path ... # display contents
       codesign -h pid ... # display hosting paths

[neverpanic (Clemens Lang)]

Sorry, should have checked the manpage. It's codesign -s - -f $path.

[kencu (Ken)]

thanks.

no happiness yet:

% codesign -s - -f ./make
./make: replacing existing signature

% ./make
zsh: killed     ./make

codesign -s - -f /Users/cunningh/kentest/make
/Users/cunningh/kentest/make: replacing existing signature

% ./make --version       
zsh: killed     ./make --version

% /Users/cunningh/kentest/make --version
zsh: killed     /Users/cunningh/kentest/make --version

[neverpanic (Clemens Lang)]

Ugh, so Apple thought of some other mechanism to identify what binaries were produced by them and shouldn't be run from copies.

Until somebody figures out how they did it, we're stuck. Maybe it's a new stat flag, an entitlement, something in the resource fork, or some other metadata? It could be a database of hashes, although the signature modification should have bypassed that already.

[linuxgemini (İlteriş Eroğlu)]

If I preserve the entitlements already present and sign with my Apple Development certificate, it works:

~ ❯ /usr/bin/make --version
GNU Make 3.81
Copyright (C) 2006  Free Software Foundation, Inc.
This is free software; see the source for copying conditions.
There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A
PARTICULAR PURPOSE.

This program built for i386-apple-darwin11.3.0

~ ❯ ./make --version
[1]    39270 killed     ./make --version

~ ❯ codesign --preserve-metadata=entitlements --force --verbose --sign "Apple Development" ./make
./make: replacing existing signature
./make: signed Mach-O universal (x86_64 arm64e) [com.apple.dt.xcode_select.tool-shim]

~ ❯ ./make --version
GNU Make 3.81
Copyright (C) 2006  Free Software Foundation, Inc.
This is free software; see the source for copying conditions.
There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A
PARTICULAR PURPOSE.

This program built for i386-apple-darwin11.3.0

[kurthindenburg (Kurt Hindenburg)]

In case this is helpful ​https://mjtsai.com/blog/2023/03/16/ventura-adds-com-apple-provenance/

[neverpanic (Clemens Lang)]

Hm, can somebody check on Ventura or later whether a copy of a binary from /usr/bin has this xattr, and if it does, whether it can be removed, and if it can be removed, whether removing it allows the copied binary to run on Ventura? I'd be surprised if that worked, honestly, but it's worth checking.

[kencu (Ken)]

if it is there, it is hidden in some way:

cp /usr/bin/make ./
xattr -l ./make
<nothing>


% xattr -px com.apple.provenance  ./make
xattr: ./make: No such xattr: com.apple.provenance

[kencu (Ken)]

so I tried making a new codesigning certificate using this as a template:

​https://github.com/llvm/llvm-project/blob/main/lldb/scripts/macos-setup-codesign.sh

and renamed it slightly to macports_codesign instead of lldb_codesign. The certificate is generated and is there:

% ./macports-codesign.sh                                                                       
Certificate has already been generated and installed

but no bueno:

% rm make
% cp /usr/bin/make ./
% codesign --preserve-metadata=entitlements --force --verbose --sign "macports_codesign" ./make
./make: replacing existing signature
./make: signed Mach-O universal (x86_64 arm64e) [com.apple.dt.xcode_select.tool-shim]
% ./make --version
zsh: killed     ./make --version

[kencu (Ken)]

my slightly edited copy of lldb's codesign-certificate-generatings-script to generate a macports signing certificate

[neverpanic (Clemens Lang)]

The issue is in dyld. Apple has added new restrictions that depend on the result of a syscall called amfi_check_dyld_policy_self(). See

• ​https://github.com/apple-oss-distributions/dyld/blob/c8a445f88f9fc1713db34674e79b00e30723e79d/dyld/DyldProcessConfig.cpp#L672-L683
• ​https://github.com/apple-oss-distributions/dyld/blob/c8a445f88f9fc1713db34674e79b00e30723e79d/dyld/DyldProcessConfig.cpp#L702-L724
• ​https://github.com/apple-oss-distributions/dyld/blob/c8a445f88f9fc1713db34674e79b00e30723e79d/dyld/DyldDelegates.cpp#L120-L142
• ​https://github.com/apple-oss-distributions/dyld/blob/c8a445f88f9fc1713db34674e79b00e30723e79d/dyld/glue.c#L971-L977

The only two inputs into this syscall are whether the process is restricted or fairPlayEncrypted. Fair Play encryption would be visible as a load command LC_ENCRYPTION_INFO or LC_ENCRYPTION_INFO_64, which isn't present on the binaries in /usr/bin as far as I can see. Restriction is enabled if the binary has a __RESTRICT,__restrict section according to ​https://github.com/apple-oss-distributions/dyld/blob/c8a445f88f9fc1713db34674e79b00e30723e79d/common/MachOFile.cpp#L1588-L1598. A quick check in /usr/bin suggests that only the fontrestore binary has such a section, so it's also not the reason for our problems.

However, since the system call is being performed from within the process in question, the kernel probably knows what binary is being executed, so we must also treat the process and its executable file as inputs, and I believe that's where the magic happens: If I try to run a copy of make, it is immediately killed:

$ cd /tmp
$ cp /usr/bin/make make
$ ./make
Killed: 9

If I, however, remove the code signature on that binary using codesign --remove-signature, it runs and DYLD_* variables seem to work:

$ codesign --remove-signature make
$ ./make
make: *** No targets specified and no makefile found.  Stop.
$ DYLD_INSERT_LIBRARIES=foo ./make
dyld[58717]: terminating because inserted dylib 'foo' could not be loaded: tried: 'foo' (no such file), '/System/Volumes/Preboot/Cryptexes/OSfoo' (no such file), 'foo' (no such file), '/private/tmp/foo' (no such file), '/System/Volumes/Preboot/Cryptexes/OS/private/tmp/foo' (no such file), '/private/tmp/foo' (no such file)
dyld[58717]: tried: 'foo' (no such file), '/System/Volumes/Preboot/Cryptexes/OSfoo' (no such file), 'foo' (no such file), '/private/tmp/foo' (no such file), '/System/Volumes/Preboot/Cryptexes/OS/private/tmp/foo' (no such file), '/private/tmp/foo' (no such file)
Abort trap: 6

It seems to me we need to modify sip_copy_proc to strip Apple's signature from those binaries. Signing the binaries with an ad-hoc signature also seems to work, but that's probably a lot more complicated to implement (although we may have to for arm64).

Additionally, those binaries use the arm64e architecture, so we'll probably also have to start building an arm64e slice of darwintrace.dylib.

[neverpanic (Clemens Lang)]

Just for completeness, there are also solutions out there that use an Endpoint Security entitlement to interpose processes before they are started to patch the amfi_check_dyld_policy_self syscall from the binary before the loader has a chance to call it (see ​https://gist.github.com/saagarjha/a70d44951cb72f82efee3317d80ac07f), but that probably requires disabling SIP. If all else fails, that's probably something we could do on a copy of the binaries from /usr/bin, too, although it would look a bit different for us since we wouldn't be patching in memory but essentially doing what would be called a GOT hijack on Linux.

[neverpanic (Clemens Lang)]

For those where removing the signature, or replacing it with an ad-hoc signature did not work, can you check whether the output of sysctl security | grep launch_constraints suggests that launch constraints disallow 3rd party binaries?

[jrabinow]

macOS 13.3 22E252 arm64
Xcode 14.3 14E222b

As far as I can tell, amfi_check_dyld_policy_self isn't being issued as per

• /usr/bin/syscallbypid.d
• /usr/bin/syscallbyproc.d
• /usr/bin/syscallbysysc.d

and no new process is being executed as per /usr/bin/newproc.d. execsnoop does see make but when I tried tracing it with dtruss -W I got failed to grab pid 19387: (null) which makes me think it's too short-lived and/or doesn't get initalized properly. So somewhere between execsnoop and newproc.d the process gets killed off.

---

I tried to strip the signature from make as described above, it's still getting killed:

$ codesign --remove-signature make
$ codesign -vv make
make: code object is not signed at all
In architecture: arm64e
[Exit 1]
$ ./make
Killed: 9
[Exit 137 (KILL)]

I also tried signing make with my own custom signature and it still didn't launch.

---

Moving onto provenance: I can see the provenance on make. Worth noting it didn't appear immediately (less than 1h of playing around with it, digging into this and a few other things). When I tried checking and copying a make2 to tmp, the provenance appeared right away though, so go figure if I didn't mess up somehow.

$ /bin/ls -l@ /tmp/make
-rwxr-xr-x@ 1 julien  wheel  132017 Jul  4 16:09 /tmp/make
        com.apple.provenance        11
$ xattr -px com.apple.provenance ./make
01 00 00 5D 19 F7 5D 19 6C 9E A9

From: ​https://eclecticlight.co/2023/03/13/ventura-has-changed-app-quarantine-with-a-new-xattr/ The xattr is protected by SIP, so can’t be removed unless the app is copied to another volume. I haven't tried removing it.

Also for what it's worth, I asked chatgpt for a summary of what are cdhashes and checked it against apple's docs (links below). It had this to say:

A CDHash, short for Code Directory Hash, is a cryptographic hash value associated with a binary executable file on macOS systems. It is used as a security measure to verify the integrity and authenticity of the code before it is executed.

When a binary is signed with an Apple Developer certificate, the signing process generates a CDHash for the code. This hash is derived from the code directory, which contains a list of the file offsets and sizes for each segment of the executable. By calculating the CDHash, the system can ensure that the code has not been tampered with or modified since it was signed.

The CDHash is stored in the code signature of the binary and is checked by the Gatekeeper service on macOS when an application is launched. If the CDHash does not match the expected value, indicating that the binary has been altered, Gatekeeper will prevent the application from running, providing an additional layer of security against malicious software.

It's worth noting that the CDHash is just one component of the code signing process on macOS, which includes other checks.

Also see:

• ​https://developer.apple.com/documentation/endpointsecurity/es_process_t/3228976-cdhash
• ​https://developer.apple.com/documentation/technotes/tn3126-inside-code-signing-hashes

If this is what is causing issues, I'd assume the following approximate setup:

• the provenance is stored in the binary's extended attributes and can't be removed because SIP.
• this provenance points to an entry in the ExecPolicy database in provenance_tracking table. I tried finding it but couldn't.
• The execpolicy database would allow the OS (going by opensnoop's output, most likely syspolicyd) to compare the calculated cdhash against the cdhash stored in the provenance_tracking table.
• The cdhash is used to seal the binary. If the calculated values don't match the cdhash, the OS (again, most likely syspolicyd) sends kill -9

Worth noting that I tried finding make in /var/db/SystemPolicyConfiguration/ExecPolicy provenance_tracking table but couldn't find anything that seemed relevant. I rebooted my computer in case the db was only flushed to disk on process restart, still no dice. I didn't see anything relating to /usr/bin in there either, so I'm still missing something and I'm not sure if this ExecPolicy database is relevant. Maybe provenance is used by something else? Maybe, as chatgpt suggests, it's another mechanism entirely?

I'd suggest the way forward on this one would be to open up execsnoop script, repurpose it to have it attach to a new process named make, gatekeeper (assuming it runs in its own process) or syspolicyd and have it observe everything that happens to it. I've barely looked at gatekeeper, and ​https://eclecticlight.co/2023/03/16/what-is-macos-ventura-doing-tracking-provenance/ suggests to me that gatekeeper might run even before syspolicyd gets involved at all.

Also I'd be curious if anyone can repro the copying of a binary out of /usr/bin without the 'provenance' xattr, like I initially got and like kencu's comment above suggests. Can make run at that time? If you wait for some time, does the provenance appear on its own? If so, does the binary start getting killed then?

Question from a complete newbie, why do we need to use Apple's binaries at all? Could macports not ship its own binaries and bootstrap from there, reminescent of the gcc bootstrapping process?

[jrabinow]

I forgot:

$ sysctl security | grep launch_constraints
security.mac.amfi.launch_constraints_enforced: 1
security.mac.amfi.launch_constraints_3rd_party_allowed: 1

so 3rd-party sources doesn't seem to be related either.

[neverpanic (Clemens Lang)]

This is all very weird. On the one machine that I have upgraded to Ventura so far (which is x86_64), I implemented a change that re-signs all binaries before executing them with an ad-hoc signature, and trace mode works just fine with that.

I cannot reproduce the killed process when re-signing with an ad-hoc signature. Did Apple just implement additional security for arm64 machines, and left x86_64 with less security?

On the other hand, there were people above who reported getting killed binaries on x86_64, which I never saw. I don't know what's going on. I can try pushing the re-signing code for somebody else to test, if that helps?

[linuxgemini (İlteriş Eroğlu)]

I'm the initial reporter for proxmark3-iceman, which the issue happens on an x86_64 machine.

Haven't checked the trace mode for a while; I'll check later tomorrow.

[linuxgemini (İlteriş Eroğlu)]

It is tomorrow, so I tried trace mode again with proxmark3-iceman; still no dice:

:info:extract Executing:  cd "/opt/local/var/macports/build/_Users_ilteris.eroglu_Projects_macports-ports_science_proxmark3-iceman/proxmark3-iceman/work" && /usr/bin/gzip -dc '/opt/local/var/macports/distfiles/proxmark3-iceman/proxmark3-4.16717.tar.gz' | /usr/bin/tar -xf -
:debug:extract system:  cd "/opt/local/var/macports/build/_Users_ilteris.eroglu_Projects_macports-ports_science_proxmark3-iceman/proxmark3-iceman/work" && /usr/bin/gzip -dc '/opt/local/var/macports/distfiles/proxmark3-iceman/proxmark3-4.16717.tar.gz' | /usr/bin/tar -xf -
:info:extract Command failed:  cd "/opt/local/var/macports/build/_Users_ilteris.eroglu_Projects_macports-ports_science_proxmark3-iceman/proxmark3-iceman/work" && /usr/bin/gzip -dc '/opt/local/var/macports/distfiles/proxmark3-iceman/proxmark3-4.16717.tar.gz' | /usr/bin/tar -xf -
:info:extract Killed by signal: 9
:error:extract Failed to extract proxmark3-iceman: command execution failed
:debug:extract Error code: NONE
:debug:extract Backtrace: command execution failed
:debug:extract     while executing
:debug:extract "$procedure $targetname"
:error:extract See /opt/local/var/macports/logs/_Users_ilteris.eroglu_Projects_macports-ports_science_proxmark3-iceman/proxmark3-iceman/main.log for details.

default	09:56:59.075368+0300	kernel	AMFI: Launch Constraint Violation (enforcing), error info: c[1]p[1]m[1]e[3], (Constraint not matched) launching proc[vc: 1 pid: 80195]: /opt/local/var/macports/sip-workaround/503/usr/bin/sandbox-exec, launch type 0, failure proc [vc: 1 pid: 80195]: /opt/local/var/macports/sip-workaround/503/usr/bin/sandbox-exec
default	09:56:59.075443+0300	kernel	ASP: Security policy would not allow process: 80195, /opt/local/var/macports/sip-workaround/503/usr/bin/sandbox-exec
default	09:56:59.075544+0300	kernel	sandbox-exec[80195] Corpse allowed 1 of 5
default	09:56:59.093028+0300	ReportCrash	Formulating fatal 309 report for corpse[80195] sandbox-exec
default	09:56:59.113785+0300	osanalyticshelper	creating type 309 as /Users/ilteris.eroglu/Library/Logs/DiagnosticReports/.sandbox-exec-2023-07-07-095659.ips
default	09:56:59.138890+0300	osanalyticshelper	Saved type '309(<private>)' report (2 of max 25) at /Users/ilteris.eroglu/Library/Logs/DiagnosticReports/sandbox-exec-2023-07-07-095659.ips
default	09:56:59.139312+0300	ReportCrash	client log create type 309 result success: /Users/ilteris.eroglu/Library/Logs/DiagnosticReports/sandbox-exec-2023-07-07-095659.ips
default	09:56:59.139370+0300	ReportCrash	no MetricKit for process sandbox-exec type 309 bundleId (null)
default	09:56:59.139173+0300	osanalyticshelper	xpc log creation type 309 result success: /Users/ilteris.eroglu/Library/Logs/DiagnosticReports/sandbox-exec-2023-07-07-095659.ips

---

Using an ad-hoc certificate to re-sign still works:

~ ❯ cp /usr/bin/make ./make
~ ❯
~ ❯ /usr/bin/make --version
GNU Make 3.81
Copyright (C) 2006  Free Software Foundation, Inc.
This is free software; see the source for copying conditions.
There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A
PARTICULAR PURPOSE.

This program built for i386-apple-darwin11.3.0
~ ❯
~ ❯ ./make --version
[1]    80407 killed     ./make --version
~ ❯
~ ❯ codesign --preserve-metadata=entitlements --force --verbose --sign "test_codesign" ./make
./make: replacing existing signature
./make: signed Mach-O universal (x86_64 arm64e) [com.apple.dt.xcode_select.tool-shim]
~ ❯
~ ❯ ./make --version
GNU Make 3.81
Copyright (C) 2006  Free Software Foundation, Inc.
This is free software; see the source for copying conditions.
There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A
PARTICULAR PURPOSE.

This program built for i386-apple-darwin11.3.0
~ ❯

This is my work Mac, so SIP is always enabled and there's multiple EDR and DLP solutions installed.

[jrabinow]

I don't know what's going on. I can try pushing the re-signing code for somebody else to test, if that helps?

Assuming I have step-by-step instructions to follow, I'd love to test on my M1 arm64. Also of note is some parts of my SIP are disabled.

[ryandesign (Ryan Carsten Schmidt)]

Has duplicate #68010.

[neverpanic (Clemens Lang)]

On arm64, apple uses a different ABI, arm64e, which includes pointer authentication. The macOS kernel refuses to run arm64e binaries unless they are signed by Apple. A workaround is to run

sudo nvram boot-args=-arm64e_preview_abi

which disables this check. I think what happens here is twofold: On x86_64, resigning the binaries is sufficient, but on arm64, this doesn't work, and the additional boot argument may be required as well.

[neverpanic (Clemens Lang)]

On x86_64, I'd welcome feedback for ​https://github.com/neverpanic/macports-base/commits/cal-tracemode-optimizations. I'm assuming those work, since they do for me.

On arm64, it seems one cannot enable the arm64e preview ABI without disabling system integrity protection, which is unfortunate. If somebody wants to test the branch above with that change, that would be helpful.

[Clemens Lang <neverpanic@…>]

In 4a1b0c70c41473882f9fc1cb994487ac3855d884/macports-base (master):

darwintrace: Resign with codesign when copying

macOS Ventura seems to have broken trace mode because it kills signed
processes when preload libraries are present. Fix this by re-signing
binaries when copying them.

I tested this on x86_64 Ventura, where it works as expected.

Closes: #66358

[neverpanic (Clemens Lang)]

In 5c43ff0458dbcd8703f044a44efd28e2e6ec0cf1/macports-base (master):

darwintrace: Disable broken tests on arm64

Let's not break CI on arm64, even though this test failure points to an
actual problem.

See: #66358#comment:39

[neverpanic (Clemens Lang)]

Re-opening since the problem still exists on arm64.

[mrdomino (Jōshin)]

Sorry if this is the wrong place, but I haven't found answers anywhere else and I wound up here, so asking for posterity: is this the root cause for there not being any port builders for Sonoma arm64?

[reneeotten (Renee Otten)]

Replying to mrdomino:

Sorry if this is the wrong place, but I haven't found answers anywhere else and I wound up here, so asking for posterity: is this the root cause for there not being any port builders for Sonoma arm64?

If so, this seems like a rather severe issue, and the verbiage should be clearer about it in e.g. SonomaProblems and Migration. I didn't find this out until after upgrading, and it's at times very painful to not have binary packages.

No, this is about trace-mode not working with the released version of MacPorts - it has nothing to do with the fact that there is no buildbot yet for Sonoma. Please do not hijack tickets for unrelated comments.

[mrdomino (Jōshin)]

Ok, sorry. Where is the right place for that issue? Whatever the resource that is short, I might be able to contribute some of it and would like to.

[reneeotten (Renee Otten)]

Replying to mrdomino:

Ok, sorry. Where is the right place for that issue? Whatever the resource that is short, I might be able to contribute some of it and would like to.

there is a ​build-bot for Sonoma but it will take a while before it will have build all the ports so that there are binaries available. Feel free to open a ticket to offer resources setting the "component" to "buildbot/mpbb"; Ryan is running the buildbots so perhaps you tag him (i.e., ryandesign) in that ticket.

[kencu (Ken)]

if MacPorts wanted to use it’s own binaries instead of Apple’s binaries to make trace mode work again on arm, what kind of list would we need?

We probably already have most of them. And older systems would often prefer to use them too.

[neverpanic (Clemens Lang)]

Replying to kencu:

if MacPorts wanted to use it’s own binaries instead of Apple’s binaries to make trace mode work again on arm, what kind of list would we need?

Xcode, probably. There are a bunch of ports that use it to build GUI software, and I'm not sure there are open source alternatives for those.

You can get an approximation by collecting the contents of $prefix/var/macports/sip-workaround on a machine where trace mode is supported. Everything in there had system integrity protection enabled and was thus copied and executed from a copy in trace mode. On one of the x86_64 systems I own where I haven't done a huge amount of compiling, this list is:

# cd /opt/local/var/macports/sip-workaround && find . -type f | sed -E 's/^\.\/[0-9]+\///g' | sort -u
System/Library/Frameworks/Ruby.framework/Versions/Current/usr/bin/ruby
bin/bash
bin/cat
bin/chmod
bin/cp
bin/date
bin/dd
bin/echo
bin/expr
bin/hostname
bin/launchctl
bin/ln
bin/ls
bin/mkdir
bin/mv
bin/pwd
bin/rm
bin/rmdir
bin/sh
bin/sleep
usr/bin/ar
usr/bin/arch
usr/bin/awk
usr/bin/basename
usr/bin/bison
usr/bin/clang
usr/bin/clang++
usr/bin/cmp
usr/bin/codesign
usr/bin/cpio
usr/bin/ctags
usr/bin/cut
usr/bin/diff
usr/bin/dirname
usr/bin/egrep
usr/bin/env
usr/bin/file
usr/bin/find
usr/bin/flex
usr/bin/git
usr/bin/gm4
usr/bin/grep
usr/bin/gzip
usr/bin/head
usr/bin/hostinfo
usr/bin/id
usr/bin/install
usr/bin/install_name_tool
usr/bin/ld
usr/bin/lipo
usr/bin/m4
usr/bin/make
usr/bin/mktemp
usr/bin/nm
usr/bin/otool
usr/bin/patch
usr/bin/perl
usr/bin/perl5.30
usr/bin/python3
usr/bin/ranlib
usr/bin/ruby
usr/bin/sandbox-exec
usr/bin/sed
usr/bin/sort
usr/bin/sqlite3
usr/bin/strip
usr/bin/sw_vers
usr/bin/tail
usr/bin/tar
usr/bin/tclsh
usr/bin/touch
usr/bin/tr
usr/bin/true
usr/bin/uname
usr/bin/uniq
usr/bin/unzip
usr/bin/wc
usr/bin/which
usr/bin/xcode-select
usr/bin/xcrun
usr/bin/xsltproc
usr/libexec/PlistBuddy
usr/sbin/chown
usr/sbin/sysctl

launchctl, codesign, hostinfo, install_name_tool, lipo, sandbox-exec, sw_vers, xcode-select, xcrun, PlistBuddy are probably specific enough that we don't yet have them all. Note that this is also just the subset that ports I compiled on my machine use.

We probably already have most of them. And older systems would often prefer to use them too.

I'm not sure that's worth the effort it would be, but feel free to beat me to doing that.

[kencu (Ken)]

OK... that is a list.

Comes to mind we don't really care about tracing the things in /usr/bin or /bin anyway...

What we really care about are opportunistically found ports in ${prefix}.

Even just having trace mode work only on the things in ${prefix} would be a huge step forward...

[neverpanic (Clemens Lang)]

Replying to kencu:

Comes to mind we don't really care about tracing the things in /usr/bin or /bin anyway...

That isn't correct, unless you want to allow binaries in /usr/bin or binaries executed through a binary in /usr/bin to allow arbitrary unfiltered access to the filesystem. Those include /usr/bin/clang (which we really want to trace) as well as /usr/bin/make, which will execute most of our build steps, or /bin/sh, which will run essentially all build scripts.

That's required because running any binary with system integrity protection will remove all DYLD_* variables, including the DYLD_INSERT_LIBRARIES we rely on for trace mode. In other words, the moment we run /usr/bin/make or /bin/sh, everything started by those will also automatically be untraced.

What we really care about are opportunistically found ports in ${prefix}.

Yes, but those aren't found by programs in $prefix.

Even just having trace mode work only on the things in ${prefix} would be a huge step forward...

No, that will likely just lead to build failures, because the view of the filesystem is suddenly no longer consistent. The same binary would behave different depending on whether it is run directly or through /bin/sh.

[kencu (Ken)]

I was hoping that instead of needing to modify the now-unavailable binaries, there might instead be a way to put a file system trace on ${prefix}, and only allowing access to files that have been allowed to be accessed.

picture the equivalent of making a virtual /opt/local populated by symlinks to the contents of ports that have been allowed prior to the build.

then you would have the equivalent of trace mode, leaving the binaries alone.

But I don't know enough about how this is done. chroot, etc ... and I haven't explored any of the trace mode code.

[neverpanic (Clemens Lang)]

Replying to kencu:

I was hoping that instead of needing to modify the now-unavailable binaries, there might instead be a way to put a file system trace on ${prefix}, and only allowing access to files that have been allowed to be accessed.

macOS does not have this functionality, at least not in the fashion we'd need. There is a sandboxing mechanism, but it only allows denying access to files, not hiding them. We've tried that before, and most build systems will fail when you deny access to a file that they would like to use and know exists.

That's why trace mode emulates this by intercepting system calls related to file system access, but that requires doing this interception on all binaries, regardless of their file system location.

picture the equivalent of making a virtual /opt/local populated by symlinks to the contents of ports that have been allowed prior to the build.

then you would have the equivalent of trace mode, leaving the binaries alone.

But I don't know enough about how this is done. chroot, etc ... and I haven't explored any of the trace mode code.

Yes, that would be the equivalent of Linux mount namespaces, or chroots. macOS does not have the former, and while it does have the latter, they require root access, and are known to break Xcode and other macOS core functionality such as DNS lookups, which is why we're not using them.

Unless you can convince Apple to provide a mechanism to selectively hide files using sandboxes, or provide a container-like mount namespace mechanism, library preloading is the only viable option, and that doesn't work on arm64 macOS at the moment, unless you're willing to disable SIP.

[kencu (Ken)]

OK. Thanks for taking the time to explain all that.

If I understand the current options, for now, folks can install macports-base from master use trace mode on an Intel system, or on an arm system install macports-base from master, disable SIP and use:

sudo nvram boot-args=-arm64e_preview_abi

which probably nobody will really do, but there it is.

[markmentovai (Mark Mentovai)]

In many cases, it ought to be possible to reset the mach_header_64::cpusubtype from CPU_SUBTYPE_ARM64E to CPU_SUBTYPE_ARM64_ALL (0), and then re-sign. PAC instructions essentially become no-op under arm64 (as opposed to arm64e).

This may be feasible if you’re not trying to run anything with a restricted entitlement.

mark@arm-and-hammer zsh% sw_vers
ProductName:		macOS
ProductVersion:		15.0
BuildVersion:		24A335
mark@arm-and-hammer zsh% uname -orm
Darwin 24.0.0 arm64

mark@arm-and-hammer zsh% lipo /bin/ls -thin arm64e -output /tmp/ls
mark@arm-and-hammer zsh% otool -hv /tmp/ls
/tmp/ls:
Mach header
      magic  cputype cpusubtype  caps    filetype ncmds sizeofcmds      flags
MH_MAGIC_64    ARM64          E USR00     EXECUTE    20       1712   NOUNDEFS DYLDLINK TWOLEVEL PIE
mark@arm-and-hammer zsh% python3
Python 3.12.6 (main, Sep  7 2024, 05:43:35) [Clang 16.0.0 (clang-1600.0.26.3)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import struct
>>> f = open('/tmp/ls', 'r+b')
>>> f.seek(8)
8
>>> hex(struct.unpack('<I', f.read(4))[0])
'0x80000002'
>>> f.seek(8)
8
>>> f.write(struct.pack('<I', 0))
4
>>> f.close()
>>> ^D
mark@arm-and-hammer zsh% otool -hv /tmp/ls
/tmp/ls:
Mach header
      magic  cputype cpusubtype  caps    filetype ncmds sizeofcmds      flags
MH_MAGIC_64    ARM64        ALL  0x00     EXECUTE    20       1712   NOUNDEFS DYLDLINK TWOLEVEL PIE
mark@arm-and-hammer zsh% codesign --sign=- --force --preserve-metadata /tmp/ls
warning: default usage of --preserve-metadata implies "resource-rules" (deprecated in Mac OS X >= 10.10)!
/tmp/ls: replacing existing signature
mark@arm-and-hammer zsh% /tmp/ls -dl /tmp/ls /tmp/ /tmp /
drwxr-xr-x  20 root  wheel    640 Sep  5 16:54 /
lrwxr-xr-x@  1 root  wheel     11 Sep  5 16:54 /tmp -> private/tmp
drwxrwxrwt  35 root  wheel   1120 Sep 25 15:45 /tmp/
-rwxr-xr-x   1 mark  wheel  89088 Sep 25 15:45 /tmp/ls

[markmentovai (Mark Mentovai)]

I should probably show something with the full end-to-end context demonstrating that DYLD_INSERT_LIBRARIES now works.

mark@arm-and-hammer zsh% cat /tmp/preload.cc 
#include <stdio.h>

static struct Preload {
 public:
  Preload() {
    fprintf(stderr, "preloaded!\n");
  }
} preload;
mark@arm-and-hammer zsh% clang++ -dynamiclib -o /tmp/libpreload.dylib /tmp/preload.cc
mark@arm-and-hammer zsh% DYLD_INSERT_LIBRARIES=/tmp/libpreload.dylib /tmp/ls -dl /tmp/ls /tmp/ /tmp /
preloaded!
drwxr-xr-x  20 root  wheel    640 Sep  5 16:54 /
lrwxr-xr-x@  1 root  wheel     11 Sep  5 16:54 /tmp -> private/tmp
drwxrwxrwt  37 root  wheel   1184 Sep 25 16:04 /tmp/
-rwxr-xr-x   1 mark  wheel  89088 Sep 25 15:54 /tmp/ls

[neverpanic (Clemens Lang)]

Replying to markmentovai:

In many cases, it ought to be possible to reset the mach_header_64::cpusubtype from CPU_SUBTYPE_ARM64E to CPU_SUBTYPE_ARM64_ALL (0), and then re-sign. PAC instructions essentially become no-op under arm64 (as opposed to arm64e).

That's actually a great idea. We already have code that copies and re-signs binaries, seeking to an offset of 8 and writing 4 0-bytes should not be all that complicated. I didn't know that the binaries would just run with the changed header, I had expected them to start failing.

This may be feasible if you’re not trying to run anything with a restricted entitlement.

Our mechanism did already break binaries with entitlements, and it hasn't been a problem in practice, so I think we can ignore this.

​https://github.com/macports/macports-base/blob/master/src/pextlib1.0/sip_copy_proc.c#L485-L488 is the place where we currently re-sign binaries, if we add something right before that that adjusts the mach-o header if required, trace mode might actually start working again. Note that this does not go the easy way of using a thin binary, so the function would have to understand universal binaries.

[neverpanic (Clemens Lang)]

​https://github.com/macports/macports-base/pull/354

[neverpanic (Clemens Lang)]

Looking for testers of this PR on macOS 14; it works fine for me on macOS 15.1, but in CI on macOS 14, I see

dyld[20831]: bad bind opcode 0x1E
…
Killed by signal: 6

Now, that error message is completely unhelpful, because the code that prints it (​https://github.com/opensource-apple/dyld/blob/master/src/ImageLoaderMachOCompressed.cpp#L1027-L1028) does so after moving the pointer that it prints (​https://github.com/opensource-apple/dyld/blob/master/src/ImageLoaderMachOCompressed.cpp#L955). I'd still like to know if this happens in regular usage on macOS 14, or whether it's just a corner case I am triggering with my test.

[markmentovai (Mark Mentovai)]

Replying to neverpanic:

Looking for testers of this PR on macOS 14; it works fine for me on macOS 15.1, but in CI on macOS 14, I see

dyld[20831]: bad bind opcode 0x1E
…
Killed by signal: 6

Now, that error message is completely unhelpful, because the code that prints it (​https://github.com/opensource-apple/dyld/blob/master/src/ImageLoaderMachOCompressed.cpp#L1027-L1028) does so after moving the pointer that it prints (​https://github.com/opensource-apple/dyld/blob/master/src/ImageLoaderMachOCompressed.cpp#L955). I'd still like to know if this happens in regular usage on macOS 14, or whether it's just a corner case I am triggering with my test.

The message you’re quoting from the source (bad bind opcode %d in bind info) doesn’t even match the error you’re seeing (dyld[20831]: bad bind opcode 0x1E). You’re looking at super-old dyld source from an unofficial dump of Apple’s old opensource site (​github opensource-apple), from before Apple pushed source to GitHub directly. It’s showing dyld-360.18 from 10.11.2 (2015-12-08), and it hasn’t been current since 10.11.3 (2016-01-19). I recommend that everyone purge that unmaintained opensource-apple dump from their bookmarks and workflow, and never consult it for any purpose. It’s outdated and misleading.

Current Apple open-source drops are at (​github apple-oss-distributions), and dyld in particular is ​https://github.com/apple-oss-distributions/dyld. You can look at source specific to the macOS version you’re interested in, ranging from ​dyld-1222.1 from 14.0 (2023-09-26) through ​from 14.6 (2024-07-29).

Now, that doesn’t change the fact that, even 9 years later, the loop is still structured such that “bad bind opcode” still references the opcode that follows the one it’s actually complaining about. But given the proper current source, and the fact that there are 8 different loops that might result in bad bind opcode 0x%02X, you may find a loop that’s used for non-arm64e that’s missing handling for an opcode that the arm64e equivalent does handle.

• ​common/MachOAnalyzer.cpp:1535 in dyld3::MachOAnalyzer::forEachBind, handles 13 cases
• ​common/MachOAnalyzer.cpp:1700 in dyld3::MachOAnalyzer::forEachBind, handles 13 cases
• ​common/MachOAnalyzer.cpp:2170 in dyld3::MachOAnalyzer::parseOrgArm64eChainedFixups, handles 10 cases
• ​common/MachOAnalyzer.cpp:5070 in dyld3::MachOAnalyzer::forEachBind_OpcodesWeak, handles 13 cases
• ​common/MachOAnalyzer.cpp:5179 in dyld3::MachOAnalyzer::forEachBind_OpcodesRegular, handles 13 cases
• ​common/MachOLayout.cpp:603 in mach_o::Fixups::parseOrgArm64eChainedFixups, handles 10 cases
• ​common/MachOLayout.cpp:1062 in mach_o::Fixups::forEachBind_OpcodesWeak, handles 13 cases
• ​common/MachOLayout.cpp:1171 in mach_o::Fixups::forEachBind_OpcodesRegular, handles 13 cases

​The full suite of opcodes is defined in <mach-o/loader.h>, the 14 #defines beginning with the terminator BIND_OPCODE_DONE:

#define BIND_OPCODE_DONE					0x00
#define BIND_OPCODE_SET_DYLIB_ORDINAL_IMM			0x10
#define BIND_OPCODE_SET_DYLIB_ORDINAL_ULEB			0x20
#define BIND_OPCODE_SET_DYLIB_SPECIAL_IMM			0x30
#define BIND_OPCODE_SET_SYMBOL_TRAILING_FLAGS_IMM		0x40
#define BIND_OPCODE_SET_TYPE_IMM				0x50
#define BIND_OPCODE_SET_ADDEND_SLEB				0x60
#define BIND_OPCODE_SET_SEGMENT_AND_OFFSET_ULEB			0x70
#define BIND_OPCODE_ADD_ADDR_ULEB				0x80
#define BIND_OPCODE_DO_BIND					0x90
#define BIND_OPCODE_DO_BIND_ADD_ADDR_ULEB			0xA0
#define BIND_OPCODE_DO_BIND_ADD_ADDR_IMM_SCALED			0xB0
#define BIND_OPCODE_DO_BIND_ULEB_TIMES_SKIPPING_ULEB		0xC0
#define	BIND_OPCODE_THREADED					0xD0

The ones that handle 13 cases omit BIND_OPCODE_THREADED. The ones that handle 10 cases omit BIND_OPCODE_ADD_ADDR_ULEB, BIND_OPCODE_DO_BIND_ADD_ADDR_ULEB, BIND_OPCODE_DO_BIND_ADD_ADDR_IMM_SCALED, and BIND_OPCODE_DO_BIND_ULEB_TIMES_SKIPPING_ULEB.

Given that we’re potentially putting an arm64e executable through a non-arm64e path, I’d first suspect a use of BIND_OPCODE_THREADED.

BIND_OPCODE_THREADED was for an earlier form of chained fixups (“old arm64e”), while the newer form just uses LC_DYLD_CHAINED_FIXUPS and is available beyond just arm64e. With additional effort, it may be possible to apply further translations to chained fixups in this format to make them work more broadly.

Your log message is too truncated to dig deeper. Do you have a link? What executable is your test operating on when you observe this failure? What specific OS version?

[neverpanic (Clemens Lang)]

Replying to markmentovai:

The message you’re quoting from the source (bad bind opcode %d in bind info) doesn’t even match the error you’re seeing (dyld[20831]: bad bind opcode 0x1E). You’re looking at super-old dyld source from an unofficial dump of Apple’s old opensource site (​github opensource-apple), from before Apple pushed source to GitHub directly. It’s showing dyld-360.18 from 10.11.2 (2015-12-08), and it hasn’t been current since 10.11.3 (2016-01-19). I recommend that everyone purge that unmaintained opensource-apple dump from their bookmarks and workflow, and never consult it for any purpose. It’s outdated and misleading.

I noticed that after posting the comment, noticed that the correct version still had the same bug, and thus didn't bother to correct it here to avoid adding more noise to the signal.

The ones that handle 13 cases omit BIND_OPCODE_THREADED. The ones that handle 10 cases omit BIND_OPCODE_ADD_ADDR_ULEB, BIND_OPCODE_DO_BIND_ADD_ADDR_ULEB, BIND_OPCODE_DO_BIND_ADD_ADDR_IMM_SCALED, and BIND_OPCODE_DO_BIND_ULEB_TIMES_SKIPPING_ULEB.

Given that we’re potentially putting an arm64e executable through a non-arm64e path, I’d first suspect a use of BIND_OPCODE_THREADED.

Thanks, that information is very helpful. Is there some documentation that I'm not aware of that explains what BIND_OPCODE_THREADED actually does, and whether just stripping it from the binaries is a safe operation?

BIND_OPCODE_THREADED was for an earlier form of chained fixups (“old arm64e”), while the newer form just uses LC_DYLD_CHAINED_FIXUPS and is available beyond just arm64e. With additional effort, it may be possible to apply further translations to chained fixups in this format to make them work more broadly.

Your log message is too truncated to dig deeper. Do you have a link? What executable is your test operating on when you observe this failure? What specific OS version?

I'm seeing this in the macOS 14 CI run for the PR above, which is at ​https://github.com/macports/macports-base/actions/runs/11673868984/job/32505429280?pr=354. The OS version is macOS 14.7 23H124. The specific failure occurs close to the end of the "Test MacPorts Base" foldable section with the /bin/ln binary. For reasons that I haven't yet investigated this does not actually fail the run, even though the test definitely failed and it should have marked the build as failing.

One instance is, for example:

system: /bin/ln -s ../../../../../../../../../..//Users/runner/work/macports-base/macports-base/src/darwintracelib1.0/tests//./stat symlink
dyld[20334]: bad bind opcode 0x1E
Command failed: /bin/ln -s ../../../../../../../../../..//Users/runner/work/macports-base/macports-base/src/darwintracelib1.0/tests//./stat symlink
Killed by signal: 6
Expected violation '/Users/runner/work/macports-base/macports-base/src/darwintracelib1.0/tests/stat' did not occur
SANDBOX
  /Users/runner/work/macports-base/macports-base/src/darwintracelib1.0/tests/symlink=+
  /private/var/select/sh=+
  /bin/bash=+
  /bin=+
  /usr=+
  /dev=+
  /opt/local/var/macports/sip-workaround=+


==== darwintrace_relative_symlinks Test that resolution of relative symlinks works as expected FAILED
==== Contents of test case:
exec -ignorestderr -- ./stat "$cwd/symlink" 2>@1
---- Test setup failed:
command execution failed
---- errorInfo(setup): command execution failed
    while executing
"system "/bin/ln -s $path symlink""
    ("uplevel" body line 6)
    invoked from within
"uplevel 1 $setup"
---- errorCode(setup): CHILDKILLED 20334 SIGABRT SIGABRT
==== darwintrace_relative_symlinks FAILED

All invocations that fail use /bin/ln.

[markmentovai (Mark Mentovai)]

Replying to neverpanic:

Thanks, that information is very helpful. Is there some documentation that I'm not aware of that explains what BIND_OPCODE_THREADED actually does, and whether just stripping it from the binaries is a safe operation?

There’s no documentation aside from dyld source. There have been some WWDC sessions that discuss chained fixups, but only at a high level, not at this low implementation level. And in any case BIND_OPCODE_THREADED was an older form of chained fixups, before the more recent LC_DYLD_CHAINED_FIXUPS that Apple would have showcased at WWDC. Limited to arm64e, BIND_OPCODE_THREADED was only ever Apple-private.

It’s definitely not safe to strip in general. You’ll lose (at least some) relocations, which will leave you with a broken executable (except in the very unlikely event that it only covered some relocations, all of which are irrelevant to execution). You’d need to interpret BIND_OPCODE_THREADED and translate to a different form.

Without resorting to disabling SIP, I do have another recipe to put Apple’s executables more under user control: ​exec_via_dlopen. I don’t like it as much because it’s not as “pure” or transparent an environment as the recipe I shared in comment:60, but in light of comment:67, it has the property that it leaves arm64e as arm64e, and allows dyld to process BIND_OPCODE_THREADED wherever it may validly be found in an arm64e image. It has another advantage: you don’t have to rewrite Apple’s executables. There’s no code signature stripping or Mach-O tweaking involved at all.

mark@arm-and-hammer zsh% clang -Wall -Wextra -Werror -o /tmp/exec_via_dlopen exec_via_dlopen.c
mark@arm-and-hammer zsh% DYLD_INSERT_LIBRARIES=/tmp/libpreload.dylib /tmp/exec_via_dlopen /bin/ls -dl /bin/ls /bin /
preloaded!
drwxr-xr-x  20 root  wheel     640 Oct 22 03:49 /
drwxr-xr-x@ 39 root  wheel    1248 Oct 22 03:49 /bin
-rwxr-xr-x   1 root  wheel  154624 Oct 22 03:49 /bin/ls