Skip to content

Stela v0.4.1

Choose a tag to compare

@github-actions github-actions released this 14 Nov 08:51
· 937 commits to main since this release

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 v0.4.1:

sudo ./stela update

On Windows, open Terminal with Run as administrator and write .\stela.exe in place of sudo ./stela.

Then boot the PC from the stick.

Changes

CLI

  • stela setup is replaced by stela (menu), create, update, repair, status, settings and log. The stick is an optional argument.
  • A v0.3 command line (setup, boot-log, configure, --usb) is not run; it prints its replacement.
  • A stick records the Stela version that made it. An older stela asks before it changes a newer stick.
  • A download over 1 GB is asked about once, with its size. --yes accepts it.
  • The stick menu ends with Quit, and Enter takes it.
  • [#64, 119484] 0.4.1: Erase and create in the stela menu makes the stick. 0.4.0 stopped there with diskN is not a mounted volume; stela create worked.
  • [390518, 153198, 816394, #39] api.ipsw.me is asked again, up to four times, when it answers without its list of macOS releases.
  • [645273] An Offline installer needs 26 GiB, so it fits on sticks sold as 32 GB.
  • [#57] Firmware downloads use an idle timeout, resume with Range and check the pinned hash.
  • [180142] Linux: a composefs root (Fedora Silverblue) is resolved through /sysroot.
  • The Wi-Fi country comes from the computer's region.

Boot

  • [#13, #23, #34, #37, #48, #59, 271204, 027142, 910090, 220613, 822814, 923096] An ISA/LPC or host bridge with a hard-wired Bus Master bit no longer fails PCI containment ("GPU refused").
  • [#26, #33, #53, #55, #61, 983140, 847572] A vendor PCI capability shorter than its header is accepted (PCI capability at 0xb0 is malformed, Intel PCH xHCI).
  • [#24, 271204, 815932] A conventional PCI function that aliases past 0xff has no extended capability list.
  • [#23] An MSI capability with Extended Message Data is accepted.
  • [#56] Unset root bridge windows are left out; a window whose maximum is below its minimum ends at its length.
  • [#33] A masked IOAPIC entry with an Extended Destination ID is kept.
  • [117112] PCI bridges are reopened before any device starts; an NVMe controller with CFS set at ready is started as Linux starts it.
  • PCI bridges, IOAPIC pins and IOMMU units (AMD-Vi, VT-d) are taken from the firmware per function, as Linux does. Only the function that cannot be stopped is refused. Intel VMD is attached with the NVMe drives behind it.
  • [#36] Each xHCI controller is claimed from the firmware before its bus mastering is turned off.
  • [#62, 245578, 782959] Processors are placed by APIC ID when the firmware lists them at one location. The stop names its reason.
  • [#56] First-generation Ryzen: processors are placed by core from CPUID 0x8000001e.
  • [#39, #41, 606738] AMD: each processor keeps its own fixed DRAM routing bits during the processor takeover.
  • [#41] AMD without IA32_TSC_ADJUST: each processor's TSC is written to agree with the first (Non-monotonic time panic).
  • [#44] Memory is taken by address from the memory map when the firmware refuses an allocation ("firmware has no memory for Stela").
  • [#36, #38, #62] A PC with no NVIDIA card, or only cards Stela does not drive, is told so before an install or a start.
  • [#35, #42, #52, 589608, 790145] A device whose start did not finish is skipped on the next boot and named.
  • [#26, #32] An exception between ExitBootServices and the engine's tables is shown. Firmware drivers are connected only to the devices the boot needs.
  • [#47] A PC with no UART at COM1 is detected and the port is not read.
  • [#56] A stop inside the GPU start is saved; the next start shows that the last one stopped, and why.
  • A stop the USB log cannot hold is kept in firmware variables and copied into the log at the next start.
  • A panic or exception is shown on screen and saved in the boot log. A heap allocation failure names its size and callers.
  • The firmware's 5-minute watchdog is turned off at start.
  • At a stop screen, the power button turns the PC off and Ctrl+Alt+Delete restarts it.
  • A restart that macOS asks for returns to Stela when the firmware's boot order lists another system first.
  • The takeover of devices and processors is shown step by step on screen.
  • The log holds one line for each unsupported operation.

Disks and USB

  • [#34, #46, #49, #54, 407261, 555315, 897418] The host Bluetooth controller is dropped in the loader's root (page fault after the disks start in an install).
  • [#45, #59] An xHCI command no longer ends at once when the device start budget has passed, and the budget is renewed before a controller is attached again ("xHCI host controller died").
  • [#45, #59] An xHCI controller that stops answering while its devices start is reset once and attached again. The log names the command it did not answer and its status registers.
  • [#45] xHCI waits 50 ms after a root port reset and 10 ms after SET_ADDRESS, as Linux does.
  • [043433, 999396] A USB storage device whose first scan times out is reset once and probed again. A Bulk-Only command gets 10 s during the first scan.
  • [910090] A root port stays usable after a refused device.
  • [271204] A card reader with no card is logged as empty.
  • A USB device with one configuration value at two indices is kept.
  • The install's disk picker no longer writes the disk through the firmware; Stela's drivers write the partition table after the firmware has exited. The NVRAM save is written the same way.
  • NVMe: no interrupt mask access under MSI-X, CSTS.RDY alone ends the enable wait, the queue count is set before the queues are made, a refused Identify Namespace keeps the drive. A failure names the command.
  • SATA disks signal completions by interrupt.
  • USB controller, input and network polls no longer wait for another processor inside one service turn.

Ethernet

  • [#20, #26, #28, #52, 723416] RTL8125 and RTL8126: the whole register range is mapped before the start.
  • [#40, #60] RTL8168H: the multiple-descriptor fetch is left off; received frames were written into the wrong buffers.
  • [#40] The network setup waits 60 s for a DHCP lease. A failure lists the frames sent and received.
  • Broadcom Tigon3: BCM5717, 5718, 5719, 5784, 57761, 57762, 57765, 57766, 57781, 57782, 57785, 57786, 57791 and 57795 are driven.
  • Firmware patches for USB RTL8153 and RTL8156 are on the stick.
  • Every Ethernet driver stops its adapter and frees its memory when dropped.

Wi-Fi and Bluetooth do not work in this release.

Graphics and display

  • [#11, #27, #33] GTX 16 and RTX 20: DMA pages for the GSP booter are allocated.
  • [#23] The RTX 50 audio function is no longer refused over its MSI layout.
  • A GPU without Function Level Reset is reset through its power state or a Secondary Bus Reset on its port.
  • A monitor on the NVIDIA card gets a picture: the scanout buffer's address in the upper half of the address space was read as an error.
  • A monitor on the NVIDIA card runs its own preferred mode, at the fastest refresh a detailed timing can describe (pixel clock up to 655 MHz). The firmware's size was kept before.
  • A mode with a line or frame rate past 255 is described with the EDID's rate offsets (1920x1080 at 240 Hz gave a dispext0 busy timeout).
  • A monitor that reports late after the GPU reset is waited for; one plugged in later is taken.
  • The firmware mode chosen is the largest of the monitor's own shape.
  • More Metal shader forms compile: void pointer kernel arguments, bfloat globals, constant data from C++ templates, pointers made from 64-bit addresses. A function that cannot be compiled returns Metal's compile error.

Input

  • macOS on a stick or disk gets the PC's keyboards and pointers.
  • macOS is given no Bluetooth controller. It never configured the one it was given, and Setup's Language Chooser then stayed on a black screen.

Install

  • [#25, #60, 212638, 334469] The ASR sender survives a read that only ran out of wait time ("ASR already failed").
  • The first start after an install creates the folders for the Metal plugin and restarts once, not twice.
  • A restart or shutdown just after an install no longer ends in a "Halt/Restart Timed Out" panic.
  • The menu has one Install macOS entry, which opens the disk picker.
  • [#66] 0.4.1: An install reads its inventory only from the stick's own partition. A target disk that was once a Stela stick stopped it with restore inventory path appears on multiple volumes.

Performance

  • More floating-point instructions stay in translated code: half precision, FMA, square root, conversions.
  • WKdm memory compression is faster.
  • Disk commands are started and completed with less waiting between processors.
  • Many device register accesses complete without leaving translated code.
  • The translation cache on disk is bounded and merged while macOS runs.

Boot log

If Stela does not work, save the boot log and include log.txt in your report:

sudo ./stela log --output log.txt