On 2026-01-19 15:47, Ludwig Rydberg wrote:
> Hi all,
>
> This series adds support for the clone3 system call to the SPARC{32|64}
> architectures and also adds a related patch for clone/fork/vfork that fix an
> issue previously reported[1] that could result in -EFAULT for no good reason.
> Without this patch, the clone3 system call would need the same mitigation as
> introduced in glibc[2] for the clone system call.
>
> About "sparc: Synchronize user stack on fork and clone"
> ---------------------------------------------------------
>
> The clone3 implementation is developed on top of a fix for an issue reported
> by Adrian Glaubitz[1], where a clone call could return -EFAULT. This problem
> has since been mitigated in glibc[2] by synchronizing the user stack before
> calling clone.
>
> The root cause analysis of the kernel side when running the program in [1]
> shows that the window spill handler routine on both SPARC{32|64} is unable to
> flush a user window to the stack (due to MMU related faults) when flushing all
> windows before handling the syscall. This then results in a -EFAULT when
> copy_thread() fails to clone the uncommitted stackframe of the parent.
>
> For SPARC32:
> Prior to calling the syscall wrappers for clone/fork/vork all windows are
> flushed by a macro (FLUSH_ALL_KERNEL_WINDOWS).
> In the window spill trap handler, MMU fault-handling is temporarily
> disabled while storing the window. If the window can't be stored
> (which normally would have triggered a fault trap) the routine
> backups the user window and increments a thread counter (wsaved).
>
> For SPARC64:
> Prior to calling the syscall wrappers for clone/fork/vork all windows are
> flushed by issuing the flushw instruction.
> In the window spill trap handler, if an exception triggers, then the user
> window is added to the thread's user window buffer (in kernel memory) and
> a thread counter (wsaved) counter is incremented.
>
> Both SPARC{32|64}:
> Eventually copy_thread will be called, which then will fail to clone the
> parent
> stackframe to the child as the user window has not been flushed to the stack.
>
> Fixed by adding a call to synchronize_user_stack() prior to calling
> kernel_clone(). The patch has been tested both with and without the mitigation
> in glibc by running the program mentioned in [1].
>
> SPARC32:
> - Tested in QEMU emulating sun4m using Buildroot 2025.02
> (qemu_sparc_ss10_defconfig).
> - Tested on LEON using a GR-CPCI-GR740 development board from
> Frontgrade Gaisler.
>
> SPARC64:
> - Tested in QEMU emulating sun4u using Buildroot 2025.02
> (qemu_sparc64_sun4u_defconfig).
>
> About the clone3 implementation:
> --------------------------------
>
> The implementation in the architectural port follows the same pattern as for
> the
> original clone syscall. But instead of explicitly calling kernel_clone (as in
> sparc_clone) the clone3 handler calls the generic sys_clone3 handler
> (in kernel/fork). To get this to work without a user provided stack,
> the copy_thread functions had to be updated to handle cl_args.stack == NULL.
> In this case the stack of the parent is re-used.
>
> The return value convention for clone3 follows the regular kernel return
> value convention (in contrast to the original clone/fork on SPARC).
>
> In sparc_{clone|vfork|fork} there is a workaround due to the SunOS
> return value convention which is not needed for clone3.
>
> When applying the patch series on top of v6.19-rc1 the relevant clone3 tests
> of
> kselftest pass:
>
> # /usr/lib/kselftests/run_kselftest.sh -c clone3 -s
> kselftest: Running tests in clone3
> TAP version 13
> 1..4
> # selftests: clone3: clone3
> ok 1 selftests: clone3: clone3
> # selftests: clone3: clone3_clear_sighand
> ok 2 selftests: clone3: clone3_clear_sighand
> # selftests: clone3: clone3_set_tid
> ok 3 selftests: clone3: clone3_set_tid
> # selftests: clone3: clone3_cap_checkpoint_restore
> ok 4 selftests: clone3: clone3_cap_checkpoint_restore
>
> Note that the clone3_cap_checkpoint test failed in the same way as mentioned
> in
> [3] (due to incompatibility with the libcap version on my system).
> When applying the patch from [4] or by downgrading libcap to 2.59 the test
> pass.
>
> SPARC32:
> - Tested in QEMU emulating sun4m using Buildroot 2025.02
> (qemu_sparc_ss10_defconfig).
> - Tested on LEON using a GR-CPCI-GR740 development board from
> Frontgrade Gaisler.
>
> SPARC64:
> - Tested in QEMU emulating sun4u using Buildroot 2025.02
> (qemu_sparc64_sun4u_defconfig).
> - Tested on UltraSparc T4
>
> [1]:
> https://lore.kernel.org/sparclinux/3ae4130c-c5aa-428e-b819-44cf2daf2...@mkarcher.dialup.fu-berlin.de/
> [2]: https://sourceware.org/bugzilla/show_bug.cgi?id=31394
> [3]:
> https://lore.kernel.org/all/20250901-nios2-implement-clone3-v2-0-53fcf5577...@siemens-energy.com/
> [4]:
> https://lore.kernel.org/all/[email protected]/
>
> ---
> Changes in v2:
> - Added Tested-by tag from Adrian Glaubitz
> - Implement generic Linux return value convention for clone3
> - Drop SPARC handling for sys_clone3 in selftests
> - Remove legacy ABI return value handling in sparc_clone3
> - Link to v1:
> https://lore.kernel.org/sparclinux/[email protected]/
> ---
>
> Andreas Larsson (1):
> sparc: Synchronize user stack on fork and clone
>
> Ludwig Rydberg (1):
> sparc: Add architecture support for clone3
>
> arch/sparc/include/asm/syscalls.h | 1 +
> arch/sparc/include/asm/unistd.h | 2 -
> arch/sparc/kernel/entry.S | 15 ++++++++
> arch/sparc/kernel/kernel.h | 1 +
> arch/sparc/kernel/process.c | 52 +++++++++++++++++++-------
> arch/sparc/kernel/process_32.c | 23 +++++++++---
> arch/sparc/kernel/process_64.c | 27 ++++++++++---
> arch/sparc/kernel/syscalls.S | 8 ++++
> arch/sparc/kernel/syscalls/syscall.tbl | 2 +-
> 9 files changed, 102 insertions(+), 29 deletions(-)
>
>
> base-commit: 8f0b4cce4481fb22653697cced8d0d04027cb1e8
Picking the series up to my for-next.
Thanks,
Andreas
