• Re: [OSDev] How to switch to long mode in x86 CPUs?

    From vantix@3:633/10 to All on Friday, November 20, 2026 23:50:53
    Dan Cross <cross@spitfire.i.gajendra.net> wrote:
    I feel like this is where Ron Minnich's LinuxBoot stuff works
    well. Replace the UEFI DXE layer with a copy of Linux and a
    small userspace written in a type- and memory-safe language.
    You've got a real IP stack and can do DHCP and use TCP to load
    and kexec your real OS.

    That approach makes a lot of sense, especially compared to maintaining
    a custom, half-baked network stack inside early firmware. Having a minimal kernel handle hardware init and hand off via kexec (or a clean initrd-to-rootfs transition) avoids duplicating driver complexity in two places.

    Out of curiosity: in setups where you bypass UEFI/AGESA and initialize
    the fabric or early IO from your own host OS, how painful is the lack of documentation for the hidden management cores (like SMU/MPIO)? Even with openSIL shedding some light on the init flow, it still feels like a lot
    of black-box state machines under the hood.

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Dan Cross@3:633/10 to All on Saturday, November 21, 2026 13:48:20
    In article <11dqmct$73j6$1@solani.org>, vantix <vantix@veloos.invalid> wrote: >Dan Cross <cross@spitfire.i.gajendra.net> wrote:
    I feel like this is where Ron Minnich's LinuxBoot stuff works
    well. Replace the UEFI DXE layer with a copy of Linux and a
    small userspace written in a type- and memory-safe language.
    You've got a real IP stack and can do DHCP and use TCP to load
    and kexec your real OS.

    That approach makes a lot of sense, especially compared to maintaining
    a custom, half-baked network stack inside early firmware. Having a minimal >kernel handle hardware init and hand off via kexec (or a clean initrd-to-rootfs
    transition) avoids duplicating driver complexity in two places.

    I'm not sure that follows. The kernel that does the hardware
    initialization is not necessarily the same as the one that you
    actually run in steady state, thus there is no guarantee that
    the drivers are the same.

    The critical observation with e.g. LinuxBoot is that having a
    Real(TM) kernel in the DXE layer means that you are out of the
    world of buggy proprietary UEFI stacks. It's also a lot faster
    because the overall archictecture of UEFI is almost deliberately
    slow. Regardless, it is still reliant on UEFI's PEI layer to
    set up the fabric.

    Out of curiosity: in setups where you bypass UEFI/AGESA and initialize
    the fabric or early IO from your own host OS, how painful is the lack of >documentation for the hidden management cores (like SMU/MPIO)? Even with >openSIL shedding some light on the init flow, it still feels like a lot
    of black-box state machines under the hood.

    Extremely. For AMD specifically, you really need to interact
    with the SMU and MPIO to do a lot of the low-level fabirc init.
    For example, you have to send an RPC to MPIO pointing to the
    configuration of how you'd like to set up PHYs and PCIe lane
    pairs; MPIO then does the timing training and so forth. If you
    don't have some idea of what the RPCs are (and last time I
    checked, AMD had replaced the relevant constants with
    `SIL_RESERVED_12345`-type names in OpenSIL) then it's hard to
    figure out exactly what is going on.

    As far as guides go, our code is reasonably well documented, and
    is open source. We start from the (x86) reset vector in the
    pico bootloader, https://github.com/oxidecomputer/phbl/. Then
    we invoke our Unix entry point, which does the early OS startup
    and fabric initialization: https://github.com/oxidecomputer/illumos-gate/blob/stlouis/usr/src/uts/oxide/os/fakebop.c#L679
    The actual fabric initialization starts from `zen_fabric_init`: https://github.com/oxidecomputer/illumos-gate/blob/stlouis/usr/src/uts/oxide/zen/zen_fabric.c#L3262

    An interesting consequence of our design is that the kernel
    image is treated exactly like a "normal" ELF binary, and is
    loaded at its linked addresses; we invoke `_start` directly in
    64-bit mode (with virtual memory enabled) via a normal `CALLQ`
    instruction.

    - Dan C.


    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From vantix@3:633/10 to All on Saturday, November 21, 2026 18:18:56
    Dan Cross <cross@spitfire.i.gajendra.net> wrote:
    I'm not sure that follows. The kernel that does the hardware
    initialization is not necessarily the same as the one that you
    actually run in steady state, thus there is no guarantee that
    the drivers are the same.

    The critical observation with e.g. LinuxBoot is that having a
    Real(TM) kernel in the DXE layer means that you are out of the
    world of buggy proprietary UEFI stacks. It's also a lot faster
    because the overall archictecture of UEFI is almost deliberately
    slow. Regardless, it is still reliant on UEFI's PEI layer to
    set up the fabric.

    Fair point -- especially on still being tied to PEI for early silicon
    and memory init, which only pushes the proprietary black box one stage
    earlier rather than eliminating it.

    For example, you have to send an RPC to MPIO pointing to the
    configuration of how you'd like to set up PHYs and PCIe lane
    pairs; MPIO then does the timing training and so forth. If you
    don't have some idea of what the RPCs are (and last time I
    checked, AMD had replaced the relevant constants with `SIL_RESERVED_12345`-type names in OpenSIL) then it's hard to
    figure out exactly what is going on.

    Seeing `SIL_RESERVED_*` all over OpenSIL was exactly what made me wonder
    how viable clean-room or hobbyist early init actually is on modern Zen platforms without vendor NDA docs. Turning PHY and lane configuration
    into opaque firmware RPCs effectively shifts the old BIOS complexity
    into the management cores.

    An interesting consequence of our design is that the kernel
    image is treated exactly like a "normal" ELF binary, and is
    loaded at its linked addresses; we invoke `_start` directly in
    64-bit mode (with virtual memory enabled) via a normal `CALLQ`
    instruction.

    Thanks for the pointers into phbl and `zen_fabric.c` -- jumping straight
    into a linked 64-bit ELF `_start` with virtual memory already up is a
    really clean cut compared to carrying legacy boot protocols around.

    Since phbl already sets up virtual memory before that `CALLQ`, how
    minimal do you keep the initial page tables in the loader before the
    kernel takes over and rebuilds its own mappings?

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From Dan Cross@3:633/10 to All on Saturday, November 21, 2026 19:34:05
    In article <11dsnag$8mf6$1@solani.org>, vantix <vantix@veloos.invalid> wrote: >Dan Cross <cross@spitfire.i.gajendra.net> wrote:
    I'm not sure that follows. The kernel that does the hardware
    initialization is not necessarily the same as the one that you
    actually run in steady state, thus there is no guarantee that
    the drivers are the same.

    The critical observation with e.g. LinuxBoot is that having a
    Real(TM) kernel in the DXE layer means that you are out of the
    world of buggy proprietary UEFI stacks. It's also a lot faster
    because the overall archictecture of UEFI is almost deliberately
    slow. Regardless, it is still reliant on UEFI's PEI layer to
    set up the fabric.

    Fair point -- especially on still being tied to PEI for early silicon
    and memory init, which only pushes the proprietary black box one stage >earlier rather than eliminating it.

    For example, you have to send an RPC to MPIO pointing to the
    configuration of how you'd like to set up PHYs and PCIe lane
    pairs; MPIO then does the timing training and so forth. If you
    don't have some idea of what the RPCs are (and last time I
    checked, AMD had replaced the relevant constants with
    `SIL_RESERVED_12345`-type names in OpenSIL) then it's hard to
    figure out exactly what is going on.

    Seeing `SIL_RESERVED_*` all over OpenSIL was exactly what made me wonder
    how viable clean-room or hobbyist early init actually is on modern Zen >platforms without vendor NDA docs. Turning PHY and lane configuration
    into opaque firmware RPCs effectively shifts the old BIOS complexity
    into the management cores.

    An interesting consequence of our design is that the kernel
    image is treated exactly like a "normal" ELF binary, and is
    loaded at its linked addresses; we invoke `_start` directly in
    64-bit mode (with virtual memory enabled) via a normal `CALLQ`
    instruction.

    Thanks for the pointers into phbl and `zen_fabric.c` -- jumping straight
    into a linked 64-bit ELF `_start` with virtual memory already up is a
    really clean cut compared to carrying legacy boot protocols around.

    Since phbl already sets up virtual memory before that `CALLQ`, how
    minimal do you keep the initial page tables in the loader before the
    kernel takes over and rebuilds its own mappings?

    Sorry, was this generated by an LLM? It sounds almost entirely
    artificial.

    - Dan C.


    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)
  • From vantix@3:633/10 to All on Saturday, November 21, 2026 19:40:44
    Dan Cross <cross@spitfire.i.gajendra.net> schrieb:


    Sorry, was this generated by an LLM? It sounds almost entirely
    artificial.

    - Dan C.


    I know it sounds like I have generated that with LLM, but my english is not that good so I used Deepl (deepl.com) to translate what I want to write. Deepl use in the core an LLM for specified translations.
    Because of that it can look like AI generated.
    - Vantix

    --- PyGate Linux v1.5.19
    * Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)