Mike Galbraith reported the following against an old fork of preempt-rt
but the same issue also applies to the current preempt-rt tree.

   BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:46
   in_atomic(): 1, irqs_disabled(): 0, non_block: 0, pid: 1, name: systemd
   preempt_count: 1, expected: 0
   RCU nest depth: 0, expected: 0
   Preemption disabled at:
   fpu_clone
   CPU: 6 PID: 1 Comm: systemd Tainted: G            E       (unreleased)
   Call Trace:
    <TASK>
    dump_stack_lvl
    ? fpu_clone
    __might_resched
    rt_spin_lock
    fpu_clone
    ? copy_thread
    ? copy_process
    ? shmem_alloc_inode
    ? kmem_cache_alloc
    ? kernel_clone
    ? __do_sys_clone
    ? do_syscall_64
    ? __x64_sys_rt_sigprocmask
    ? syscall_exit_to_user_mode
    ? do_syscall_64
    ? syscall_exit_to_user_mode
    ? do_syscall_64
    ? syscall_exit_to_user_mode
    ? do_syscall_64
    ? exc_page_fault
    ? entry_SYSCALL_64_after_hwframe
    </TASK>

Mike says:

  The splat comes from fpu_inherit_perms() being called under fpregs_lock(),
  and us reaching the spin_lock_irq() therein due to fpu_state_size_dynamic()
  returning true despite static key __fpu_state_size_dynamic having never
  been enabled.

Mike's assessment looks correct. fpregs_lock on a PREEMPT_RT kernel disables
preemption so calling spin_lock_irq() in fpu_inherit_perms() is unsafe. This
problem exists since commit

  9e798e9a ("x86/fpu: Prepare fpu_clone() for dynamically enabled features").

Even though the original bug report should not have enabled the paths at
all, the bug still exists.

fpregs_lock is necessary when editing the FPU registers or a task's FP
state but it is not necessary for fpu_inherit_perms(). The only write
of any FP state in fpu_inherit_perms() is for the new child which is
not running yet and cannot context switch or be borrowed by a kernel
thread yet. Hence, fpregs_lock is not protecting anything in the new
child until clone() completes and can be dropped earlier. The siglock
still needs to be acquired by fpu_inherit_perms() as the read of the
parent's permissions has to be serialised.

  [ bp: Cleanup splat. ]

Fixes: 9e798e9a ("x86/fpu: Prepare fpu_clone() for dynamically enabled features")
Reported-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Cc: <stable@vger.kernel.org>
Link: https://lore.kernel.org/r/20221110124400.zgymc2lnwqjukgfh@techsingularity.net