Repository navigation
Install
Download the archive for your system, unpack it and plug in the USB stick. Then run one of these in the unpacked folder.
Make a new stick (this erases it):
sudo ./stela
Update an existing Stela stick to v0.5.0:
sudo ./stela update
On Windows, open Terminal with Run as administrator and write .\stela.exe in place of sudo ./stela.
On macOS, stela is not notarized. If macOS says it cannot be opened, run this once in the unpacked folder:
xattr -d com.apple.quarantine stela
Then boot the PC from the stick.
Changes
Boot
- [#176] A PC whose firmware marks the running boot processor unhealthy uses every enabled processor. All but one were left out and the boot stopped at the T8112 roster check.
- [#106, 829375, 512036] COM1 counts as present only when it passes Linux's 8250 register test, and a port whose transmitter never empties gets no more bytes. On a laptop without a UART each log line after the firmware exit could take seconds.
- [#106] A boot log partition that Stela's drivers cannot find ends the stick's log with its reason and one notice on screen. The processor step names the firmware's exit on screen and logs its time.
- [565491] Each call into the firmware's AML is bounded to 10 s. A method that runs past it is named in the log and in a step watchdog stop. AML that never finished hung the boot at the ACPI tables step.
- An AML semaphore wait that nothing can signal ends at once. It spun for up to 65 s.
- [#150] Hybrid laptops: the firmware's AML loads before the PCI scan and powers on each PCI port the firmware left in D3cold, as Linux does, so a discrete GPU behind such a port is scanned. A powered-off port that Stela does not power on is named.
- [743840, 628968] A PCI defect refuses only the function it concerns, its possible VFs or a failed bridge's subtree, and is named at the refused claim. One defect left every function on its segment unverified, the GPU too.
- A function or bus reset puts each Resizable BAR back at the size it had, as Linux does. The BAR came back at its default size while its driver kept the firmware's.
- [041024] The engine's timer runs at a vector above every device vector, as Linux orders them. A device vector left in service with no handler is ended and named once. Such a vector stopped the timer and then the boot.
- [#119, 041024] The first stop's reason stays on screen and in the record. The step watchdog of another processor replaced it with a 30-second step timeout.
- NMI, double fault and machine check each run on their own interrupt stack on every processor, as on Linux.
- Every stop writes the macOS console lines the log lacks before its reason.
- A processor that does not acknowledge a TLB shootdown within 250 us gets an NMI that does it, and a halted processor counts as having acknowledged. A processor busy in host work could hold every retirement for seconds.
- [#126] A processor that spins in a macOS kernel lock wait counts as running, so macOS's clock holds while the host keeps that processor away. After 25 ms of such a spin the clock runs on only in 2 ms steps. macOS panicked with
Ticket spinlock timeout. - The native clock's samples are updated under a lock. A read from another processor could stop the clock as corrupt.
- [#121] Reads of the RTC and PMU counts never go back while the system counter holds, and an RTC alarm fires on time. macOS reported the RTC running backwards.
- [#97] The SPI NOR controller is served under its own lock. A restart's NVRAM sync took the board-wide lock for each register access and could run past macOS's 30 s halt timer.
- [#161] The startup menu counts down with the PC's other systems listed when the firmware has an entry for a system on the PC's disks. After an install macOS started at once and Windows or Linux could no longer be chosen.
- Small heap blocks come from slabs of one size that go back whole when empty. Long sessions held about a million free pieces.
- [#151 by 59c84486t9-alt] pci_set_master reads the command register back and warns when bus mastering stayed off, as Linux does.
- Linux driver calls that Stela does not provide report once and answer as Linux does without that facility. A sleep where the caller may not sleep is reported once and spins. Some of these calls stopped the machine.
Disks and USB
- [#112] A USB stick that returns a short MODE SENSE reply is read as not write protected, as Linux does. It was refused and the install files on it were not found.
- [#110] Intel VMD: the root port, NVMe SSD and RAID-mode SATA controller that the firmware leaves behind a VMD are admitted, and a refused VMD resource names its step.
- NVMe: CC is written with EN clear and CAP read again before EN is set, as Linux does.
- NVMe rescans and event logs, SATA error handling and disks without a command queue run in bounded steps. A hard reset or an admin command could hold a processor for seconds.
- [#136 by G-Nug85] While a flush waits, up to 16 discard commands stay on the disk and each slot is refilled as one completes. One poll starts at most 16.
- macOS's shutdown flush of its disk runs on the host disk outside the devices' lock.
- USB 3 UAS disks run up to 16 commands at once, as Linux does. Every UAS disk ran one command at a time.
- [#165] A USB storage command that a transport error ended runs again after the device's reset, up to Linux sd's five retries, also for a stick behind hubs. The first transaction error failed the read.
- [#127] The disk report names USB units whose medium changed and the sense that said so.
- [#159] A phone that offers USB tethering is reported once as unsupported, with the reason.
Network
- [#103 by meuhland] macOS gets its network as a USB NCM adapter with its own receive interrupt.
- [#166] macOS's Ethernet link follows the host adapter's cable state, and each host link change is logged. macOS always saw a connected cable and kept a failed DHCP after a late link.
- [#166] RTL8125 and RTL8126: a PHY register read of all ones no longer ends link polling for the rest of the boot, as on Linux. The guest network report shows the host adapter's state when it gives macOS no frames.
- [#130, #40, 458416, 631120] Realtek: a received frame is dropped only when its error summary bit is set, and the PHY advertises pause frames, as Linux's r8169 does. An adapter that gets no address logs its registers and rings.
- [#130, 458416] A network with only a unique local IPv6 address waits for its DHCPv4 lease and looks names up over IPv4. The install took the local IPv6 address and every download range timed out.
- An unanswered DHCPDISCOVER is sent again after 4 s, as RFC 2131 advises. It was sent again after 10 s.
- Each DNS server gets an 11 s turn at a query. The third server was never asked.
- [236556] The install's network tries every host adapter, those with a link first. With two adapters, the cable in the second one failed the install.
- The USB input line no longer prints a network adapter's address.
GPU
- [#85, #118, #138, #119, #155 by sagarkothapalli] RTX 50 series (GB20x) runs GSP-RM 570.211.01, which the CLI takes from NVIDIA's open kernel module sources and driver installer. 570.144 did not start the RTX 5060 and 5060 Ti (
gsp: init failed, -22). Booted on an RTX 5080 here; the contributor booted an RTX 5060. - The GB20x bootloader and GSP-FMC carry the MIT notice of NVIDIA's open kernel modules, and the GSP-RM image the NVIDIA Driver License. The notices are on the stick, and the CLI names the licence once when it writes the image.
- An update removes an older per-PC firmware set that the bundle replaces. A set with GSP-RM 570.144 for an RTX 50 GPU stayed and was read at every boot.
- When GSP-RM ends its own start with an error, the stop names the step.
- The GPU's P-state and PCIe link are logged. A clock query that GSP-RM 570.144 refused every second is no longer sent.
- [#74, #138] A new GPU buffer no longer waits for every job on its address space, and Metal compiles run in short slices that any processor resumes. A GPU call could run into the 10 s stop (
a GPU call (0xc0306480) still ran after 10000 ms) or fail as out of device memory. - [346297] A GPU fence wait from a macOS processor ends after 1 s, and one exit holds a processor at most 200 ms in it. A hung channel is then handled as a lost device. Such a wait ran to the 10 s stop (
a GPU call (0xc02864c3) still ran after 10000 ms). - [#141, #148, #132, 061892] A texture whose rows are not a multiple of 32 bytes, as on a 3440 or 1680 pixel wide screen, is used through an aligned copy. WindowServer stopped on its plane textures.
- [#141, #145 by takoyakiwhite] Display modes are ranked without the GPU's row alignment, so a panel's native 3440, 1366 or 1680 wide mode is taken.
- [#167] Shared depth, stencil and multisample textures are made. Chess did not start.
- Buffers, textures, heaps and device memory are freed only after every command buffer that may read them completes, also when another connection committed it. Memory was freed under running work and the GPU channel faulted.
- A destroyed two-plane IOSurface texture frees only its own descriptors.
- [#157 by bassrock] A destroyed event's semaphore is forgotten by the submissions still in flight.
- A fragment shader's buffer read in helper invocations no longer reads below the buffer. Core Animation's buffer image blit faulted the GPU channel and lost the device after sign-in.
- Device fences and acquire or release atomics in a translated shader keep their memory barrier.
- A shader PHI of an array or nested struct compiles, so MPSGraph inference in mediaanalysisd runs on the GPU.
- [#156 by bassrock] The Metal plugin builds with a Clang that emits objc_msgSend selector stubs (Apple clang, LLVM 23). WindowServer could not load a plugin built with one.
- A GPU claim that stops before the GPU's reset gives the firmware's screen back.
Display
- [#124, #144, 574357, 824165] Hybrid laptops with the panel on the processor's GPU: the firmware's screen is kept when the NVIDIA GPU does not show it, and macOS's frames are copied to it at the panel's refresh rate. The panel stayed at Starting the GPU or went dark. The log says when frames are copied.
- [#167] nouveau's eDP backlight support is built in: a panel's DPCD backlight is enabled at each mode set. A built-in panel driven by the GPU has its ACPI brightness set to the level it reads, else the highest listed. The log names the connectors a display shows on, such as eDP-1 (built-in panel).
- [#133] A PC whose firmware shows no screen, as after a warm reset, starts in the GPU display's mode, else 1920x1080. The boot bundle was refused (
dispext0 framebuffer-width). - A monitor runs its fastest progressive mode at the chosen size. 2560x1440 at 240 Hz and 3840x2160 at 120 to 240 Hz reach macOS through reduced-blanking timings. 3840x2160 was held to 60 Hz.
- macOS is offered the monitor's other refresh rates at the same size and can switch between them.
- A panel whose EDID has its timing only in a DisplayID extension runs that timing. It kept the firmware's mode at 60 Hz.
- Vblanks follow the monitor's own, and each vsync carries the vblank's time and the refresh period. Core Animation's display timeline sat after the vblanks and assumed 60 Hz.
- A vsync no longer waits behind swap completions and mode callbacks. A display that is off or asleep sends no vsync.
- macOS learns the display's real frame interval and gets frame information for each displayed swap.
- A second monitor that lists no mode of the picture's size, and is larger, shows the picture scaled into its native mode. It stayed dark.
- A connected output left dark, as one read before its EDID, is tried again up to three times. It stayed dark until it was unplugged.
- Each monitor on a DisplayPort hub or daisy chain shows the picture. Only the first one did.
- [256643] A display found after macOS's first frame shows that frame at once. It stayed dark until macOS drew again.
Input
- [#135 by G-Nug85] Laptops: the PS/2 keyboard starts without a controller self-test, reset or scan-set query, as Linux does on x86. An I2C controller's registers map inside the window its ACPI memory registered (
native mapping: ResourceConflict). - The PS/2 keyboard is proved as Linux's atkbd does it, and a command it refuses is logged as unsupported, not fatal. A failing port no longer removes the other one, and a running controller resynchronizes broken sequences and packets.
- [496223, 642083] Touchpads move the pointer by finger travel at their own resolution and scroll with two fingers: I2C and SMBus pads, PS/2 Synaptics and Elantech pads, and I2C-HID precision touchpads, which are switched to touchpad mode. PS/2 pads mapped onto the whole screen and I2C pads moved in coarse steps.
macOS
- [#139] Admin authorization prompts accept the correct password. The SEP credential manager keeps ACM contexts and their credentials, answers DeviceOwner and biometric policies and stores ACM environment variables. Tested in a VM.
- The SEP keystore takes an immediate lock request (MKBLockDeviceNow). It was refused.
- [#147 by 59c84486t9-alt] The automatic GPU bundle repair creates a missing GPUBundles directory. The repair failed with
StelaGPU.bundle is still not installed. - Personalized cryptex grafts no longer fail on a
cnchmismatch. The prepared device tree leaves out entangle-nonce. - [#174] The boot log says once that macOS gets no Bluetooth controller, so its Bluetooth daemon restarts, when the PC has none to lend.
- A floating-point multiply that rounds to the smallest normal number sets Arm's underflow flag, and flushes to zero under FZ, as on Arm.
Install
- [#128, #165, #168, #170, 371139, 784802, 076859] An offline install no longer reads the stick again from the start at
load sep os. Each bounded turn goes on where the last one stopped, and the bar stood still for hours before. An answer still waiting after 30 s is logged. - [#91, #144, #127] The drivers' clock reads one device counter that never goes back on any processor, at the boot's calibrated TSC rate. On AMD PCs whose processors keep a software TSC offset, the restore stopped with
Clock(Regressed). - [#142, #146, #133, #132] The disk picker measures each disk's reads and cautions on one too slow for macOS, USB disks too. A kernelmanagerd panic on a macOS disk slower than 2 ms a command names that disk on the stop screen. Such a disk still cannot run macOS.
CLI
- Each step is said in a few plain words. The ECID, signing details, disk serials, partition GUIDs and internal paths are no longer printed.
- A run whose output is not a terminal prints one line at the start and one at the end of the firmware download.
- A firmware file that moved or was replaced where it is published is not asked for again, and the error says so in plain words.
- The CLI reads NVIDIA's firmware archives only within their pinned sizes, and refuses an entry larger than the pinned image.
Performance
- A full translation pack replaces records that no boot used since with the boot's new ones, in up to two more merges a boot. The store shed every new entry after the first merge.
- A block compiled again soon after it left the engine's code ring is relocated from its kept form, not compiled again. A ring retirement step gives the code lock to a waiting processor sooner.
- Exits that transfer most often get stamped cells and tails, and a link through a cell is filled for a block that another address space compiled. Those transfers took the table probe.
- On hosts with AES-NI, translated code computes pointer authentication itself with AES rounds.
- A guest call and its return run as a host call and return. Exception returns within EL2 and timer writes that lower the timer complete in translated code.
- One page walk fills the ITLB for every 4 KiB page of a 16 KiB page and is shared by all address spaces for global pages. A shadow fill also maps the other pages of its 64 KiB window.
- Translated code keeps flags in host flags across more branches, uses shorter constants, keeps SHA-256 vectors in registers and uses AVX-512 where every processor has it.
- Each processor that runs macOS asks for full performance through Intel HWP or AMD CPPC, as Linux's performance policy does.
- Device registers are served under the device's own lock, and a repeated access at one device site skips its decode and page walk.
- Any running processor serves due device DMA and runs the GPU's Linux work while the boot processor is busy. A deferred fast IPI fires at its time from its own processor.
- Frames from Intel and Realtek adapters wake the host service on the network vector.
- Shader compiles, pipeline creation and binary archive lookups run in background jobs, not in the frame's turn, and a waiting job wakes a sleeping processor.
- [#154] The pipeline store keeps the costliest compiles first and grows with the stick, 64 to 128 MiB. Core Animation's shaders were compiled again at every boot.
- A GPU client mapping fills 2 MiB at a time. Plain shader stores are weak, and small thread arrays live in registers.
- Large two-plane IOSurface textures, such as video, are kept in device memory and copied only when the surface changed. Every sample read guest RAM over the bus.
- The boot screen, the serial log and the profile reports are written in short steps.
Boot log
If Stela does not work, save the boot log and include log.txt in your report:
sudo ./stela log --output log.txt