General: DevForums subtopic: App & System Services > Core OS Core OS is a catch-all subtopic for low-level APIs that don’t fall into one of these more specific areas: Processes & Concurrency Resources Files and Storage Resources Networking Resources Network Extension Resources Security Resources Virtualization Resources Share and Enjoy — Quinn “The Eskimo!” @ Developer Technical Support @ Apple let myEmail = "eskimo" + "1" + "@" + "apple.com"

We have a group of Macs that mount SMB shares over 10 GbE with jumbo frames (MTU 9000). Since late June, they have been kernel panicking several times a day with the same assertion: panic(cpu N caller ...): assertion failed: m->m_flags & M_PKTHDR, file: .../xnu/bsd/kern/uipc_mbuf.c, line: 4839 @uipc_socket.c:8260 Panicked thread: dlil_input_en0 Last started kext: com.apple.filesystems.smbfs 6.0.1 Environment Clients: Mac Studio (M1 Max and M1 Ultra) and Mac Pro (2019, Intel with T2), using the built-in 10GBASE-T at MTU 9000 macOS 26.5.1 (25F80), 26.6.2 (25G83) and 26.7 (25G229); it panics on all three Servers: Samba-based NAS, SMB 3.1.1, signing on, encryption off Filed as FB24912731 What we've found It still panics with our third-party EDR fully uninstalled. The Mac that panics needs an active SMB session. A Mac left on the network without a share mounted stayed up through several events that took down the others. Panics are often simultaneous across machines: two to six Macs, with different hardware and different macOS builds, within the same minute. It doesn't need sustained heavy throughput. Some panics came within minutes of reconnecting, during light editing. Setting kern.skywalk.flowswitch.rx_agg_tcp_host=0 did not help. The switch and server links stay up, and spanning tree doesn't change during these events. Only the Macs' ports drop. In one server-side capture, the client stopped sending within about 0.2 ms of receiving a READ response made of 8948-byte frames. That fits the panicked thread being dlil_input. Two existing threads look related Kernel panic using Vagrant synced folders via NFS beginning with macOS 15.4 (FB17853906). A DTS reply there said the issue is in kernel mbuf management and that SMB "is very likely to experience a similar panic." Incorrect packet handling in SMBClient MacOS 26, which describes a race in SMBClient under heavy load above about 10 Gbps. Questions Is this the same underlying issue as FB17853906, and is a fix planned for macOS 26? Our 2019 Mac Pros can't move to a later major release. Is there a known workaround, such as a sysctl, an nsmb.conf option, or a change to MTU or offload settings? Is there logging or a diagnostic we can leave enabled to capture more state at panic time? We can't reproduce this on demand, but between several machines we see it multiple times a day. We can provide full panic reports, sysdiagnoses, and packet captures from both client and server sides.

Subject: Virtualization.framework VM execution ownership and crash reclamation on Intel macOS Monterey Hello, I’m investigating the lifecycle guarantees of Virtualization.framework on Intel macOS Monterey 12.7.x. The specific scenario is a VZVirtualMachine running a Linux guest. I need to understand the ownership and reclamation behavior when the process holding the VZVirtualMachine is abruptly terminated without calling stop() or performing normal cleanup. The key questions are: For a specific VZVirtualMachine on Intel macOS Monterey, which userspace task/process actually owns the Hypervisor VM and the vCPU threads backing guest execution? Is Hypervisor execution owned directly by the calling process, or by a separate process such as: com.apple.Virtualization.VirtualMachine or another Virtualization.framework backend? If the process holding the VZVirtualMachine is terminated with SIGKILL or crashes without executing cleanup code, is the underlying guest execution context necessarily destroyed? More specifically: Can guest vCPU execution continue after the client process has died? If a separate backend process owns the VM, is that backend guaranteed to terminate or destroy the VM when the client dies? Does this behavior apply to Intel macOS Monterey 12.7.x, or only to newer macOS releases? Is there a supported diagnostic on Monterey that can map one specific VZVirtualMachine instance to the task/process that actually owns its Hypervisor VM/vCPU execution? For example, would a diagnostic showing Hypervisor execution frames such as hv_vcpu_run in a process, combined with a reliable process-exit notification, be sufficient to establish that ownership relationship? If the Virtualization backend can survive the client process, what supported VM-specific recovery or termination mechanism is available to another process? The security property I need to establish is intentionally narrow: If the userspace owner of a VM is abruptly destroyed, guest computation must not be able to continue indefinitely as an independent execution domain. Persistent disk files or other inert VM artifacts are not the concern; the question is specifically about live guest/vCPU execution and its ownership lifecycle. I’m looking for the supported architectural contract or diagnostic approach, not undocumented implementation details. Target environment: macOS Monterey 12.7.x Intel x86_64 Virtualization.framework Hypervisor.framework Hardware virtualization available No private APIs or privileged/kernel extensions Thank you.

There have been 2 serious regressions in the hypervisor framework since developer beta 6 of macOS 27 that have continued into the final release, and the first beta of 27.2 The first is that since developer beta 6 of macOS 27, virtual machines that have an Apple ID with iCloud+ signed in fail to route traffic in Safari through iCloud Private Relay despite it being on. Parallels, UTM, VirtualBuddy have all been tested and the issue applies to all of them, exposing the host machine's IP address. I have reported the issue since I discovered it and there hasn't been any communication that Apple even knows its an issue to my open report in Feedback Assistant. The second issue is a newer one, and it affects macOS 26.6.2 and earlier virtual machines. Attempting to install 26.7 through the built-in software update causes the virtual machine to black screen upon reboot during the installation. Forcing the machine off and back on causes the virtual machine to revert back to macOS 26.6.2. There has been no available IPSW file to test if a clean install of 26.7 in a virtual machine is a viable workaround, or to see if there is a bug in the updating mechanism or bug in the 26.7 release itself in virtual machines. The Parallels Desktop forum is beginning to get reports from users of that software of the same black screen issue trying to update their own 26.6.2 VMs to 26.7. These issues have also not been corrected in either 27.2 Beta 1 nor 26.7.1 Has anyone found a workaround to either of these 2 issues, or submitted similar reports and got any kind of response from Apple? The feedback reports about these issues are FB24828992 and FB24791716

macOS guest (Virtualization.framework) freezes with all vCPUs halted right after a macOS update reboot - 3/3 on a macOS 26.7.1 (25G309) M4 host, seen with both UTM and Parallels Summary On a macOS 26.7.1 (25G309, beta/seed build) host with an Apple M4, every macOS guest that runs an in-place macOS update freezes at the same point: the in-OS phase of the update completes and reports success, the guest requests a reboot, records a shutdown stall 9 seconds later, reboots twice, shows the update progress screen for about a minute, and then stops. All four vCPUs enter WFI and never wake, paravirtualized graphics stops submitting, no I/O is pending on the host side, and no host service logs an error. The VM never recovers and has to be killed. Reproduced 3 out of 3 times, under two different front-ends (UTM and Parallels Desktop) and two guest versions (15.7.x and 26.6.2). Environment Host: iMac (Mac16,3), Apple M4, 16 GB RAM, macOS 26.7.1 build 25G309 (seed channel), installed the evening before the first failure. Primary guest: macOS 15.7.9 (24G830), hardware model VirtualMac2,1, 4 vCPUs, 6 GB RAM, 90 GB raw disk image (virtio-blk) stored on an external USB SSD. Networking bridged to the built-in Ethernet port. Devices enabled: memory balloon, audio, entropy, clipboard sharing; display 1920x1200 with dynamic resolution. Front-end for the primary case: UTM [version], Apple Virtualization backend. Guest was the only VM running, cold-booted, window open. Update being applied: MSU_UPDATE_24H23_patch_15.8_minor (15.7.9 -> 15.8). Other cases: a macOS 26.6.2 guest under UTM; a macOS 15.7.7 -> 15.8 guest under Parallels Desktop 26.4.2 (57518). Steps to reproduce Cold-boot a macOS 15.7.9 guest under Virtualization.framework, as the only VM on the host. In the guest: System Settings > Software Update > install macOS 15.8. Let it reboot. Expected The guest installs the update and boots into 15.8. Actual — timeline of the primary case (R = the moment Software Update requested the reboot; absolute timestamps are in the attached logs) R-23 min to R-5 min: UpdateBrainService prepares the update inside the guest, writing about 25 GB (two "disk writes" resource reports: 16.5 GB then 8.5 GB). No errors. R-1:56: post-logout install configured; disk space check passes (5.7 GB required, 40.7 GB free). R-1:44: "SUOSUPostLogoutInstallOperation: Applying MSU update". R: "Applied MSU update"; FileVault stash committed ("kAppleFDEKeyStore_commitStash success"); "Rebooting (success = 1, displayAsleep = 0, shutdown = 0)". This is the last line the guest ever writes to install.log. R+9 s: the guest writes a shutdown_stall diagnostic report. R+10 s: host sees ParavirtualizedGraphics "Device reset" / "PGDisplayNub[0]: Destroyed" (reboot #1). R+1:20: second device reset (reboot #2), 70 s after the first. On the host, the vmnet interface is torn down and recreated with no error, the AppleVirtualPlatformIdentity service completes boot attestation with no error, and the guest's graphics driver renegotiates ("Guest requested binary version: 209"). R+1:38: host logs "PGDisplay[0]: Change display mode to 3606x2254" — the guest is on the update progress screen. R+1:41 to R+2:25: the guest's six PGFifoThreads go idle one at a time. The progress bar stops partway. No further activity of any kind. R+11:35: spindump of the VM host process (com.apple.Virtualization.VirtualMachine): 0.042 s of CPU over a 5 s sample all four com.apple.virtualization.thread.cpu-N threads in Hv::Vcpu::run() -> HvCore::Hypervisor::VcpuStateManager::wait_for_interrupt() -> __psynch_cvwait cpu-0/cpu-1 wake on a ~20 ms timer tick and return to WFI; cpu-2/cpu-3 had not run for seconds no thread in any file read, write or fsync (no host I/O outstanding) PGFifoThreads last ran 550–614 s earlier process state Ss (sleeping), not U Host kernel log for the window: no USB, APFS, I/O error, timeout or reset entries. No hardware video decoder (AppleAVD) errors. The VM was left for over an hour with no change, then killed. Second case (same host, same day, UTM, guest macOS 26.6.2) Identical signature: two graphics device resets 69 s apart, display mode set 3 s after the second one, last graphics activity about a second later, then all vCPUs idle in WFI for ~6.5 hours until the VM was killed. Third case (same host, same day, Parallels Desktop 26.4.2, guest 15.7.7 -> 15.8) The guest was suspended in the middle of its update and resumed later. On resume ("-[_PGDevice willResumeWithSuspendState:error:]: Begin resume", preceded by "[VirtualMachineParameterBuilder] Failed to get auxiliary file identifier"), paravirtualized graphics never came back — its FIFO threads ran once and never again — and the guest's vCPUs sat in WFI. This may be a separate save/restore defect, but the end state is the same. What I believe is ruled out The in-OS phase of the update: it completed and reported success. Guest kernel panic: vCPUs are halted, not spinning, and there is no panic report on the guest's data volume. Disk space: 40.7 GB free in the guest at install time. Host storage stall: no uninterruptible wait, no I/O frames in the spindump, no kernel storage errors. Host hardware video decoder: no AppleAVD errors in the primary case. Host services: vmnet and AppleVirtualPlatformIdentity completed normally seconds before the hang. Guest memory: 6 GB allocated. Not yet isolated The VM images live on an external USB SSD; not yet reproduced from internal storage. The memory balloon, audio and clipboard-sharing devices were enabled; not yet reproduced with them disabled. A third-party VPN client was running on the host (guest networking is bridged, so guest traffic bypasses the host tunnel, but host firewall rules could still affect bridged frames). I don't have a confirmed-good in-place guest update on this machine from before 26.7.1, so I can't state with certainty that this is a regression. Frequency 3 of 3 attempts on this host. Attachments / available on request spindumps of the hung VM host process (primary case and second case) host unified-log excerpts for both hang windows the guest's full install.log the guest's shutdown_stall report and the two UpdateBrainService disk-writes reports sysdiagnose captured while hung, if obtained Has anyone seen macOS guests stop at this point on the 26.7.x seeds? If you can reproduce, your host build, hypervisor, and whether the VM image is on internal or external storage would be useful to compare.

The language in the new EULA seems to indicate we can get permission to run more than 2 VMs on a single host. "(iii) except as otherwise provided in writing, signed, or issued by an authorized representative of Apple, to install, use and run up to two (2) additional copies or instances of the Apple Software, or any prior macOS or OS X operating system software or subsequent release of the Apple Software, within virtual operating system environments on each Apple-branded computer you own or control that is already running the Apple Software, for purposes of: (a) software development; (b) testing during software development; (c) using macOS Server; or (d) personal, non-commercial use." This is important because we often have use-cases that require running docker containers and other Virtualization tools along-side of two VMs on the host and obviously cant. What's the steps we can take to get written permission for this? Thank you!

Hi, we have a bug report in Cholesky implementation of Accelerate on the LAPACK GitHub. See: https://github.com/Reference-LAPACK/lapack/issues/1408 This is to let you know. This might be a user error but I thought I'd let you know. I did not try out for myself. Julien.

Hi Quinn — following up from DTS case 22070584. I'm working on a Windows compatibility runtime (Wine plus a CPU translator) that runs natively on Apple Silicon. 64-bit x86 Windows programs work fine. 32-bit ones don't, because they need address space in the low 4 GB: guest pointers are 32-bit, and some Windows structures sit at fixed addresses like 0x7ffe0000 that programs read directly. On arm64 I can't get anything down there: task_info(TASK_VM_INFO) -> min_address 0x100ea0000 mmap(0x7ffe0000, MAP_FIXED) -> ENOMEM mach_vm_allocate(0x7ffe0000, VM_FLAGS_FIXED) -> KERN_INVALID_ADDRESS There's also nothing below 4 GB to remove: mach_vm_region finds no entry there at all, and mach_vm_deallocate(0, 4 GB) returns KERN_SUCCESS without changing anything. Building with a smaller __PAGEZERO doesn't help either — every size I tried (0x1000, 0x4000, 0x10000, 0x100000, 0x1000000, 0x10000000, 0x80000000) gets SIGKILLed before main, with no crash report. Ad-hoc signing, the hardened runtime and -no_pie made no difference. I did notice /usr/libexec/rosetta/runtime is arm64 with no __PAGEZERO segment at all and __TEXT at vmaddr 0, so the kernel can clearly do this, at least for platform binaries. Is there a supported way for a third-party arm64 process to map below 4 GB — an entitlement, a spawn attribute, something I've missed? If the answer is no, that's fine, I'd just like to know so I can stop looking and plan around it. I have two small test programs that print all of the above if they'd be useful.

In the LAPACK interface of Accelerate (the default, without ACCELERATE_NEW_LAPACK), zheev with JOBZ='V', UPLO='U' returns eigenvectors that are neither orthonormal nor eigenvectors, while INFO=0. The eigenvalues are correct. zhegv with UPLO='U' fails the same way (tested with B = identity). It happens when: the matrix is complex Hermitian of order N >= 130 (N = 129 is correct, N = 130 already fails, N = 300 fails), and its last two or more rows and columns are exactly zero (one zero row is fine). Zero rows at the start or in the middle are fine. Banded and dense matrices both fail. On the same matrices the following are correct to rounding (1e-15): zheev and zhegv with UPLO='L', zheevd with UPLO='U', dsyev with UPLO='U' on real symmetric matrices of the same shape, zheev with UPLO='U' through the new LAPACK interface (-DACCELERATE_NEW_LAPACK). Since only the eigenvectors are wrong, and only above N=129 (presumably where zhetrd switches to its blocked code), the fault seems to be in the blocked Householder reduction or in the back-transformation (zhetrd/zlatrd or zungtr/zungql) of the legacy library, for columns that are zero. Asking LAPACK for the optimal LWORK (LWORK = -1 query) does not change the result. We found this in a physics code (Quanty, quanty.org), where a matrix of this shape arises naturally: a block tridiagonal (block Lanczos) matrix whose last block is padded with zeros after deflation. The wrong eigenvectors silently degraded results by 1e-4 relative, instead of 1e-15. STEPS TO REPRODUCE Save the attached zheev_upper_bug.c. clang -O2 zheev_upper_bug.c -o zheev_upper_bug -framework Accelerate ./zheev_upper_bug For comparison: clang -O2 -DACCELERATE_NEW_LAPACK zheev_upper_bug.c -o zheev_upper_bug_new -framework Accelerate && ./zheev_upper_bug_new The program builds a random Hermitian matrix (fixed seed, entries in [-0.5,0.5), bandwidth 11 or dense), zeroes its last NZERO rows and columns, calls zheev/zhegv/zheevd and prints max|V^H V - 1| and max|A V - V diag(W)|. EXPECTED RESULT Both numbers of order 1e-15 for every line, as with the new LAPACK interface. ACTUAL RESULT (legacy interface) zheev UPLO=U N=129 zero trailing rows=2 INFO=0 max|V^H V - 1| = 4.0e-15 max|A V - V W| = 4.4e-15 ok zheev UPLO=U N=130 zero trailing rows=2 INFO=0 max|V^H V - 1| = 9.8e-01 max|A V - V W| = 1.1e+00 <-- WRONG zheev UPLO=U N=130 zero trailing rows=1 INFO=0 max|V^H V - 1| = 4.7e-15 max|A V - V W| = 3.3e-15 ok zheev UPLO=U N=138 zero trailing rows=5 INFO=0 max|V^H V - 1| = 1.1e+00 max|A V - V W| = 1.2e+00 <-- WRONG zheev UPLO=U N=138 zero trailing rows=5 INFO=0 max|V^H V - 1| = 2.1e-01 max|A V - V W| = 2.2e+00 <-- WRONG zheev UPLO=U N=300 zero trailing rows=2 INFO=0 max|V^H V - 1| = 7.0e-01 max|A V - V W| = 1.5e+00 <-- WRONG zhegv UPLO=U N=138 zero trailing rows=5 INFO=0 max|V^H V - 1| = 1.1e+00 max|A V - V W| = 1.2e+00 <-- WRONG the same matrices with UPLO = 'L' or zheevd: zheev UPLO=L N=138 zero trailing rows=5 INFO=0 max|V^H V - 1| = 5.9e-15 max|A V - V W| = 3.6e-15 ok zhegv UPLO=L N=138 zero trailing rows=5 INFO=0 max|V^H V - 1| = 4.4e-15 max|A V - V W| = 3.7e-15 ok zheevd UPLO=U N=138 zero trailing rows=5 INFO=0 max|V^H V - 1| = 1.9e-15 max|A V - V W| = 1.4e-15 ok With -DACCELERATE_NEW_LAPACK every line is "ok" (1e-15). Code to reproduce the error: zheev_upper_bug.c.txt

My custom filesystem kernel extension stores ACLs as an extended attribute, com.apple.system.Security. Sanboxed apps such as TextEdit, Pages, etc., running as a non-privileged process, fail to save modified contents when permissive ACLs are in use. Running them as a privileged process, does allow for file changes to be saved though. Non-sandboxed apps, such as VSCode, and command line programs are not susceptible to this behaviour. APFS, on the other hand, seems to handle ACLs as an ATTR_CMN_EXTENDED_SECURITY filesystem attribute, rather than as an EA. In this case, sandboxed apps have no trouble accessing the ACL data. I implemented a minimal PoC within my custom kext to verify this. I construct an ACL in memory allowing a given user to write,append,delete file contents, and return it that via vnop_getattr. This allows the file contents to be modified and saved by sandboxed apps. Can you please confirm if my findings are accurate and sandboxed apps fail to read the com.apple.system.Security EA by design? Also, Is it an accurate assumption, that ACLs should be handled either as an EA, or an ATTR_CMN_EXTENDED_SECURITY, but not both? Thanks.

I am seeing a repeatable interaction between audio activity or audio-device discovery and a Bluetooth LE HID mouse on a Mac mini M2 Pro. Opening the detailed Sound panel in Control Center, or changing the HDMI audio output, makes the pointer stutter. Opening Apple’s Music.app has also caused a temporary complete loss of the mouse. In the worst cases, the mouse becomes completely unavailable and does not recover without Bluetooth troubleshooting or re-pairing. Configuration Mac mini Mac14,12, Apple M2 Pro, 32 GB RAM macOS Tahoe 26.7 (25G229) Also reproduced on macOS Tahoe 26.1 (25B78) Logitech MX Anywhere 3 Mac over Bluetooth LE VID 0x046D, PID 0xB025 Firmware MPM24.01_0015 Logitech MX Keys M Mac over Bluetooth LE PL4071UH display and audio output over HDMI Logi Options+ installed Trigger Connect the MX Anywhere 3 Mac and MX Keys over Bluetooth. Open Control Center from the menu bar. Open the detailed Sound panel, or change the HDMI audio output. Move the Bluetooth mouse continuously while audio-output discovery is taking place. The issue is intermittent, but it has been reproduced several times on both macOS 26.1 and 26.7. Pointer stuttering can continue after the Sound panel is closed. A later incident occurred when I opened Apple’s Music.app. The Bluetooth mouse became unavailable again, but this time it recovered automatically after approximately ten seconds. The Sony WH-1000XM6 headset was connected over Bluetooth, while the 512 GB USB flash drive was not connected. This shows that the USB drive is not required for a temporary mouse loss and makes concurrent Bluetooth audio activity a more relevant condition to investigate. The additional trigger suggests that the problem may not be limited to the Control Center interface itself and may also be associated with audio-subsystem initialization or audio-device activity. This is an observation, not yet a confirmed causal link. Additional condition in the permanent-loss cases In every incident where the mouse did not recover on its own and became effectively unusable, the following two devices were also present: a Sony WH-1000XM6 wireless headset connected over Bluetooth; a 512 GB SanDisk USB flash drive connected to the Mac. I am reporting this as a correlation, not as a confirmed cause. The headset adds simultaneous Bluetooth audio activity. The USB drive may or may not be relevant depending on its USB mode, activity, shielding, physical location, and possible 2.4 GHz interference. The 512 GB capacity is included only to identify the device. Less severe cursor stuttering has also been observed around Sound-panel discovery; the two devices above are specifically the common condition I recorded during the permanent mouse-loss incidents. Observed result The pointer starts stuttering and raw mouse HID reports develop multi-second gaps. In severe incidents, the Bluetooth mouse stops working entirely. Opening Music.app has also caused a temporary complete loss of the mouse; during that incident the Sony WH-1000XM6 was connected over Bluetooth, the USB flash drive was not connected, and the mouse recovered automatically after approximately ten seconds. Logi Options+ then displays the mouse as inactive. The mouse disappears from the IOHID tree, while the Bluetooth keyboard can remain present. A wired USB mouse remains responsive, so the Mac and pointer rendering are not completely frozen. The failure can persist after the Sound panel is closed. Relevant logs On macOS 26.7, bluetoothd repeatedly recorded the following sequence for the MX Anywhere 3 Mac: MX Anywhere 3 Mac needs low latency current connection parameters: interval 30.00 ms, latency 0 connection changed to: interval 7.50 ms, latency 44 Device "MX Anywhere 3 Mac" is Incompatible LE HID Device "MX Anywhere 3 Mac" does not have _HID_SERVICE_ Connection timed out disconnect result: 307 BLEPaired HIDBadBehavior Central The sequence repeated during automatic reconnection attempts. The mouse later recovered on some occasions, but other incidents required turning Bluetooth off for about 15 seconds, turning it back on, forgetting the mouse, and pairing it again. Instrumented reproduction During a monitored reproduction on macOS 26.1: detailed audio discovery started at +21.333 s; a 2.400 s gap in raw HID reports from the mouse occurred 0.899 s before that event; bluetoothd then reported a HID lag issue for VID 0x046D, PID 0xB025 at +1.897 s; no display-frame gaps were detected; there was no sustained CPU saturation. This indicates that the interruption is already present in the raw HID input stream. It does not appear to be only a cursor animation or WindowServer rendering delay. Expected result Discovering or switching audio outputs should not alter a separate BLE HID mouse connection, cause multi-second HID report gaps, remove _HID_SERVICE_, or disconnect the mouse. Technical points for investigation Could this be a macOS Bluetooth/HID connection-parameter or service-discovery problem triggered by concurrent audio discovery? In particular, why does macOS change the mouse from 30 ms / latency 0 to 7.5 ms / latency 44, then classify it as Incompatible LE HID, lose _HID_SERVICE_, and eventually disconnect it with result 307? Has anyone reproduced the same sequence with a Logitech BLE mouse, a Bluetooth audio headset, or a USB 3 storage device connected? The issue was also submitted through Apple’s public macOS feedback form on 24 September 2026. That form confirmed receipt but did not provide an FB case number. Possibly related report Another Developer Forums report describes a different BLE HID device with Incompatible LE HID and disconnect result 307 on macOS 27. The hardware and trigger are different, but the Bluetooth/HID failure signature may be related: https://developer.apple.com/forums/thread/845857

To whomever this may concern, We are investigating a BLE firmware-update compatibility issue affecting a legacy Nordic nRF51 DFU bootloader on a subset of iPads. We are seeking guidance on whether there are known CoreBluetooth or Bluetooth-controller differences across recent iPad hardware that could affect reconnection to Bluetooth LE connectable directed advertisements. Background Our accessory uses a legacy Nordic nRF51 DFU flow: The iPad is connected to the accessory’s normal application firmware. The app requests entry into DFU mode. The accessory disconnects normally. The accessory restarts into its bootloader. The bootloader sends connectable directed advertisements targeted at the iPad that initiated DFU. The iPad is expected to reconnect, after which the firmware transfer begins. Observed behavior On working iPads, the BLE trace shows: Application disconnect → accessory begins connectable directed advertising → iPad sends BLE connection indication → bootloader connection succeeds → DFU begins On affected iPads, the trace shows: Application disconnect → accessory begins the same connectable directed advertising → no connection indication from the iPad → DFU cannot proceed The accessory is advertising to the current resolvable address of the connected iPad. The bootloader starts directed advertising at nearly identical timing in both working and failing traces, approximately 456 ms after disconnect. The initial disconnect is normal in both cases: LL_TERMINATE_IND / Remote User Terminated Connection / reason 0x13 The problem occurs only after the accessory begins directed advertising. It seems like the iPad might be able to receive and listen to directed advertisement, however it seems to be unable to send back a connection indication. Hardware correlation observed so far We observed a correlation with Bluetooth controller vendor information visible in our BLE captures: FORMAT FOR FOLLOWING LINES Result -> Example device -> Observed controller vendor DFU succeeds -> iPad Air, 5th generation -> Broadcom DFU succeeds -> Other tested iPads, including one iPad with an A16-class platform -> Broadcom DFU fails -> One affected iPad with an A16-class platform -> MediaTek We want to be careful not to overstate this result: We have tested only one affected MediaTek-based iPad so far. The model string is MD3Y4LL/A and its current iOS version 26.6.1. The functioning Broadcomm iPads A16 we have tested that did not fail DFU process have model string MD7F4KN/A. The other iPad models we tested DFU on (which were all successful, albeit none of them were MediaTek chips, they were all Broadcomm) were: MRJN2KN/A, MM9C3KN/A, MQ6J3KN/A, MM6T3KN/A (the latter with iOS 26.6.1 as well, suggesting it is not necessarily an iPad OS version issue) We have not found a second MediaTek-based iPad to determine whether the behavior reproduces across that hardware. We initially suspected that recent non-cellular iPads might consistently use MediaTek, but this was disproved by the iPad Air, 5th generation we tested (MM9C3KN/A): it had no cellular connection and used Broadcom, and DFU worked normally. Therefore, we do not conclude that the issue is caused by MediaTek hardware. We only see a correlation worth investigating. What we have ruled out The target firmware image is not involved; failure occurs before any firmware data is transferred. The accessory reaches bootloader mode and sends directed advertisements at the expected time. The iOS Nordic DFU library already calls CoreBluetooth’s normal reconnect API: centralManager.connect(peripheral, options: nil) iOS does not expose an API allowing an app to create or send a raw Link Layer CONNECT_IND; therefore the application cannot directly work around controller-level connection behavior. The affected iPad does not provide a normal scanner/discovery callback for this directed bootloader advertisement, so matching by name or service UUID is not available as an application-layer workaround. Questions . Are there known CoreBluetooth, iPadOS, or Bluetooth-controller differences that can affect handling of BLE connectable directed advertisements, particularly when an accessory directs advertisements to a central using a resolvable private address? . Is there a recommended CoreBluetooth pattern for reconnecting to a known peripheral after it restarts into a legacy bootloader that uses directed advertising? . Does CBCentralManager.connect(_:options:) queue a pending connection request that should be serviced when a matching directed advertisement is received? If so, are there circumstances in which this behavior differs by iPad hardware or controller implementation? . Is there any public diagnostic information, entitlement, or supported API that could help us determine why a pending CoreBluetooth connection request does not result in a connection attempt in this scenario? . Does Apple recommend a specific compatibility approach for legacy BLE DFU bootloaders that use directed advertising, beyond updating the accessory firmware to use ordinary undirected advertising? . We can provide packet captures from both the working and failing cases, including timestamps, directed-advertisement target addresses, and the absence/presence of the subsequent connection indication. Thank you for any guidance on whether this is expected behavior, a known compatibility limitation, or if there is anything we could try on our end to hopefully resolve this issue.

Subject: Hard SMC and Kernel bug on 14" MacBook Pro M5 Pro (Mac17,9) in Build 26B5091g - High Power Mode completely unresponsive Hello CoreOS and Apple Silicon Performance Engineering Teams, I am writing to report a severe, reproducible system-level regression regarding High Power Mode and SMC power distribution on the new 14-inch MacBook Pro with the M5 Pro chip (Model Identifier: Mac17,9), currently running macOS 27.2 Beta 2 (Build 26B5091g). The issue prevents the M5 Pro hardware from scaling its cooling and power envelope correctly, rendering High Power Mode completely broken regardless of the power source (tested with both official 96W and 140W Apple power adapters). Key technical diagnostic findings from my device: Kernel vs UI State Mismatch: Running pmset -g via Terminal returns powermode 2, proving that the kernel registers High Power Mode as active. However, the System Information profile reader (Hardware -> Power) shows a direct conflict: "High Power Mode: No" and "Low Power Mode: Yes" (Режим энергосбережения: Да). Unresponsive SMC & Fan Curves: The cooling hardware remains completely unresponsive to the high-performance toggle. The fans fail to engage the pre-emptive 4500-5000 RPM cooling curve under heavy compute or gaming loads, remaining locked in the silent "Automatic" profile, which causes artificial thermal constraints. Powermetrics Telemetry Anomalies: A raw powermetrics trace captured during this state reveals a severe thread management bug: The P0-Cluster (High-Performance cores) shows a down residency of 98.88%, meaning the heavy performance cores are effectively kept asleep by the OS governor during high-power configuration. Combined Power (CPU + GPU + ANE) drops to a mere 429 mW. The thermal sampler entirely fails to pull hardware RPM telemetry from the SMC fan loops, providing only a generic "Current pressure level: Nominal" string. This profile behavior confirms that the macOS 27.2 power governor layer is suffering from a hard breakdown in SMC communication specific to the J614/Mac17,9 platform. The operating system handles the software flag but fails to pass the physical instruction downstream to the hardware control registers. Please prioritize this issue for the next macOS 27.2 beta cycle to ensure full thermal and compute optimization for the M5 Pro architecture prior to the final public release. I have attached the complete sysdiagnose archive, terminal configuration text files, and raw powermetrics logs. Thank you for your dedication to Apple Silicon optimization. FB24929043

I'm using the macOS 26 vmnet network API with Virtualization framework (vmnet_network_create + VZVmnetNetworkDeviceAttachment) and giving each VM a stable address with vmnet_network_configuration_add_dhcp_reservation, plus creation-time rules from vmnet_network_configuration_add_port_forwarding_rule. Both work on the network's first run. But if the network's last interface leaves while I still hold the vmnet_network_ref, then the next interface start brings the network back without its reservations or forwarding rules. The subnet and gateway are kept. How it shows up in practice: a VM that is the only guest on its network gets restarted from inside the guest. Virtualization framework removes and re-adds the VM's vmnet interface about 1 s apart, with no delegate callback. That's enough to stop and restart the network, and the guest comes back on a dynamic lease with its forwarded ports refused. Minimal repro, no VM needed: Create a configuration in VMNET_SHARED_MODE with set_ipv4_subnet, one add_dhcp_reservation and one add_port_forwarding_rule, then call vmnet_network_create. vmnet_interface_start_with_network. InternetSharing logs port forwarding enabled …, and /etc/bootptab contains the reservation. vmnet_stop_interface. InternetSharing logs no internal interface left, stopping network → reset to idle, and /etc/bootptab is emptied. vmnet_interface_start_with_network again on the same, still-retained ref. InternetSharing logs has been started with no port forwarding enabled line, and /etc/bootptab is never rewritten. The forwarded port is refused. Control: release the ref, call vmnet_network_create again from the same configuration, and start an interface. The reservation and the rule are both back. Reproduced on macOS 27.0 (26A428) and (with a minimal vmnet only test, no actual VM) on macOS 26.6.2 (25G83). This seems to contradict the documentation for vmnet_network_create: "The lifetime of such reservation is the same as that of vmnet_network_ref." A related issue: vmnet_interface_add_ip_port_forwarding_rule and its remove and get counterparts return VMNET_FAILURE synchronously on an interface started with vmnet_interface_start_with_network. The same calls work on a vmnet_start_interface interface. The add_port_forwarding_rule documentation points to those calls for managing rules after start, so there's no way to put a rule back after the restart. The same "torn down on last detach while the ref lives on" lifecycle is also reported in apple/container#2051 (https://github.com/apple/container/issues/2051), with a different symptom (two networks sharing a bridge, so tearing one down breaks the other's egress). It proposes the same mitigation: an anchor interface held with vmnet_interface_start_with_network. Filed as: FB24895271: reservations and rules discarded when the last interface is removed FB24895264: the runtime forwarding calls fail on network-attached interfaces FB24895282: suggestion for a network state-change callback and a DHCP lease query by MAC Questions: Is losing reservations and rules when the network goes idle intended? Is holding an app-owned interface on the network (via vmnet_interface_start_with_network) so it never reaches zero interfaces a supported workaround, or does it risk other side effects?

After scouring the XNU kernel source code, I came across restartable ranges [1]. They seem to be a task-level version of the the part of Linux's restartable sequences [2] for user requested synchronization [3] (akin to MEMBARRIER_CMD_RSEQ). Because of their differences, I would like to know more about how they work and what they are used for. From what I have gathered [4], they seem to be closer to RCU [5] that waits for eviction (as opposed to completion). But the fact that they are barely mentioned anywhere (see [4]) in the XNU source code I'd be really happy if you can point me to more information about them and how they compare to rseq (or rcu). What are the XNU devs planning on doing with them? [1] osfmk/kern/restartable.c: https://github.com/apple-oss-distributions/xnu/blob/f6217f8/osfmk/kern/restartable.c [2] https://criu.org/Restartable_Sequences [3] as opposed to (what I think is what makes rseq really cool) automatic eviction on preemption/CPU migration [4] there seems to be very little information online (including this forum) and the only places inside the XNU kernel source trees that aren't tests are in osfmk/kern/{thread,thread,sched_prim}.{c,h}, which doesn't really give much more information. [5] Read-Copy-Update: https://www.kernel.org/doc/html/latest/RCU/whatisRCU.html

Technology: Service Management (SMAppService), App Sandbox, launchd, Foundation file URLs. This is a documentation/support question based on a standalone signed probe. It is not an assertion of a confirmed Apple defect, and it is not a duplicate of the launchd enhancement FB24726121 referenced in DTS thread 844906. Request Please clarify the supported identity and persistent-storage architecture for this combination of requirements. We need a sandboxed Mac App Store app to install its bundled, sandboxed SMAppService LaunchDaemon through one in-app action and the required native approval. The LAN listener must be unprivileged and remain available after logout and at boot. On macOS 27.2 beta 1 (26B5086k), our signed app-like probe launches and accesses its private container as the default root identity. We have read DTS thread 844906’s single-process privilege-drop guidance; this request concerns the additional sandbox, storage, and distribution limits. Observations • UserName=nobody exits during sandbox initialization with /var/empty home and secinitd UID errors. • An existing non-root account control also traps in libsecinit before main. • SessionCreate=true with that existing non-root account still terminates in libsecinit before main. • Creating a new audit session did not resolve sandbox initialization in this probe. • A retained directory descriptor works with openat after dropping privileges, but ordinary container paths and /dev/fd/directory/child do not. • All diagnostic registrations were removed for these tests; no product data or network listener was involved. Questions Is a sandboxed bootstrap that permanently drops groups, GID, and UID before opening a LAN listener a supported architecture for the Mac App Store distribution lane? How should such a LaunchDaemon obtain an isolated service identity and persistent container with Foundation/MLX-compatible file URLs without an external installer, shared writable storage, manual account/ACL setup, or unsupported sandbox exceptions? Please identify the supported public-API path, and clearly distinguish Service Management support from App Review guideline 2.4.5 constraints. If these requirements cannot be combined, please state which ones are mutually exclusive. Reference: https://developer.apple.com/forums/thread/844906 Expected outcome: documented guidance on a supported arrangement, or a clear statement of which requirements cannot be combined. We are withholding the product’s Release host installation until the identity/storage design is qualified.

we are currently found the vulnerability about Python 3, but it is located under "/usr/bin/python3" does anyone know how to update this? we try to update python app, but it doesn't fix this, since it is under /usr/local/bin....

I have a watchOS app communicating with a BLE peripheral during an active HKWorkoutSession. The workout provides sufficient runtime: once connected, the Watch can receive a notifying GATT characteristic approximately once per minute for hours, including while the display is off. In normal conditions the connection is very stable. The difficult case is outdoor use with marginal RF conditions. The peripheral and application protocol otherwise work reliably, but outdoors there are fewer reflected RF paths and normal body movement can intermittently obstruct the Watch-to-peripheral path. In those situations the link occasionally terminates with CBError.connectionTimeout. Reconnection is particularly important because this peripheral advertises for reconnection only approximately once per minute. We also know approximately which second of each minute the peripheral becomes active, so missing a single advertising opportunity can cause a relatively long data gap. I’m trying to better understand the supported CoreBluetooth/watchOS behavior in this situation. Maintaining a marginal existing connection Once the BLE connection is established, are there any public CoreBluetooth mechanisms or recommended practices for making that connection more robust against CBError.connectionTimeout? In particular, can application-level GATT activity have any influence on connection scheduling or link robustness, or is this entirely handled by the Bluetooth stack below CoreBluetooth? CBConnectPeripheralOptionEnableAutoReconnect with sparse advertising When CBConnectPeripheralOptionEnableAutoReconnect is enabled and the connection is lost, how does watchOS handle a peripheral that advertises only very infrequently? Does CoreBluetooth simply keep the reconnect pending and wait for subsequent advertisements indefinitely, subject to system policy? Known peripheral activity timing If the app knows approximately when the peripheral becomes active and advertises each minute, is there any supported way to make use of that timing during reconnection? Or should the app simply leave AutoReconnect pending and let CoreBluetooth handle reception? WWDC22 reduced reconnection range WWDC22 session 10135 says: “If your device is on the edge of Bluetooth range and repeatedly disconnects while in Background BLE connection, the reconnection range will be reduced. This means only devices close to the Apple Watch will reconnect.” The session also says these limits are counted over a rolling 24-hour window and are reset when the user interacts with the app. I would particularly appreciate clarification of how this works in practice: What conditions cause a disconnect to count toward this policy? Does this policy apply while an app has an active HKWorkoutSession and continues executing in the background? Is the state maintained per peripheral, per app, or more globally? What exactly constitutes the user interaction that resets it? Does bringing an already-running workout app back to the foreground reset it, or does the app need to be relaunched? Is there any public API or diagnostic indication that reduced reconnection range is currently active? Is there any supported way to explicitly clear/reset that state? Connection parameters Does watchOS expose any supported mechanism for a CoreBluetooth central to influence parameters relevant to connection robustness, such as connection interval, supervision timeout, PHY, transmit power, or similar link parameters? Diagnosing CBError.connectionTimeout When CoreBluetooth reports CBError.connectionTimeout, is there any supported diagnostic method to determine whether the Watch stopped receiving the peripheral, the peripheral stopped receiving the Watch, or whether this should simply be treated as a link-layer timeout without attribution to either side? For context, this is not primarily a background-runtime problem. The active workout keeps the app running and the normal persistent BLE connection can operate flawlessly for many hours. I’m specifically trying to understand the best supported strategy for preserving and recovering the BLE link when RF conditions temporarily become marginal, especially with a peripheral that advertises very sparsely after a disconnect. DTS asked me to post this question here for review.

On Apple silicon with macOS 26.6.2 (25G83), I am preparing a small synthetic VM using Swift and public Virtualization.framework APIs. Before invoking VZMacOSInstaller, I need to establish a hard aggregate allocation bound covering its temporary extraction data and relevant helper caches, not just the supplied restore image, guest disk and auxiliary-storage URLs. The application-owned artifacts would be placed inside a fixed-size, capped filesystem. Unknown installer scratch remains part of the byte budget. Sampling free space or cancelling after a threshold is crossed is not a substitute for filesystem enforcement in this design. No installer has been started for this trial; this is an API-design question, not a reproduced installation failure. Is there a supported mechanism to select or constrain the filesystem for all VZMacOSInstaller and relevant helper temporary writes? In particular, is there a documented binding between a caller's temporary directory and independently managed installer-service storage, or another supported aggregate quota design? If the public API cannot provide this guarantee, an explicit statement of that limitation would help. I am not seeking private parameters, managed-container relocation, disabled system protections, or undocumented sandbox overrides.

I’m using Virtualization.framework on Apple silicon with a Linux guest (VZGenericPlatformConfiguration + VZLinuxBootLoader). The VM is intentionally minimal: 2 vCPUs, 2 GiB RAM 1 virtio entropy device 2 virtio block devices (base read-only, scratch read-write) 1 virtio console with 2 ports, both isConsole = false no serial, network, sharing, socket, USB, audio, graphics, keyboard, pointing, balloon, or custom virtio devices no EFI variable store nested virtualization disabled On the normal success path the host does not call requestStop(). A destructive host stop is tracked separately and treated as failure. I need a supported way for the host to distinguish: a clean Linux guest shutdown intentionally issued by the guest after its application protocol and cleanup have completed, from an abnormal or independent shutdown path such as kernel panic, watchdog, thermal / hardware-protection shutdown, emergency shutdown, or another kernel/platform-triggered stop. guestDidStopVirtualMachine tells me that the guest stopped, but I cannot find a public contract that says which Linux/kernel/platform histories can produce that callback, nor a public shutdown reason/initiator value. My specific questions are: For VZGenericPlatformConfiguration + VZLinuxBootLoader, what is the documented complete guest-visible shutdown/reset event surface, including implicit platform events not represented by explicitly configured device arrays? What Linux-facing mechanism does VZVirtualMachine.requestStop() use in this configuration? Can guestDidStopVirtualMachine also be emitted after panic, watchdog, thermal/hardware-protection shutdown, emergency shutdown, or another guest-kernel/platform shutdown source? Are those abnormal cases guaranteed to arrive through virtualMachine(_:didStopWithError:) instead? If guestDidStopVirtualMachine can represent multiple terminal histories, is there any supported public API or documented guarantee that lets the host distinguish a clean guest system-off from the abnormal/platform-triggered cases? If not, is it correct to treat this distinction as unspecified by the public Virtualization.framework contract? I do not need private implementation details. A public/supported contract describing which terminal histories can produce each delegate callback would be enough. This matters because the host is fail-closed: it must accept PASS only after an application-level success condition and a clean guest shutdown. A successful runtime observation alone is not enough for the qualification. Environment: Apple silicon / arm64 macOS 26.6.2 (25G83) public Virtualization.framework APIs

In macOS Tahoe 26.7 or Sequoia 15.7.5 (and probably other prior versions), when you install a .pkg, symlinks are followed: i.e. if your payload install a file to /Users/Shared/my-folder and, on the target volume, the /Users/Shared/my-folder item is a symbolic link to /Users/Shared/actual-folder, the file will be installed inside /Users/Shared/actual-folder. It used to be possible to tell the installation framework/mechanism not to follow symbolic links by using the followSymLinks attribute in the raw package PackageInfo file. And in this case, the symbolic link would be replaced by a real directory and the file would be installed where it was supposed to be installed. This attribute is apparently not taken into account anymore by the framework/mechanism. [Q] Is this on purpose or is it a bug (in shove for instance)? If it is on purpose, why was it considered a good idea to always follow symbolic links considering the security issues this can create?