[RFC] x86: Drop fpregs lock before inheriting FPU permissions during clone

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

   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+0xfa/0x480
   CPU: 6 PID: 1 Comm: systemd Tainted: G            E       (unreleased)
   Call Trace:
    <TASK>
    dump_stack_lvl+0x45/0x5b
    ? fpu_clone+0xfa/0x480
    __might_resched+0x165/0x200
    rt_spin_lock+0x2d/0x70
    fpu_clone+0x32a/0x480
    ? copy_thread+0xef/0x270
    ? copy_process+0xd2c/0x1c00
    ? shmem_alloc_inode+0x16/0x30
    ? kmem_cache_alloc+0x120/0x2a0
    ? kernel_clone+0x9b/0x460
    ? __do_sys_clone+0x72/0xa0
    ? do_syscall_64+0x58/0x80
    ? __x64_sys_rt_sigprocmask+0x93/0xd0
    ? syscall_exit_to_user_mode+0x18/0x40
    ? do_syscall_64+0x67/0x80
    ? syscall_exit_to_user_mode+0x18/0x40
    ? do_syscall_64+0x67/0x80
    ? syscall_exit_to_user_mode+0x18/0x40
    ? do_syscall_64+0x67/0x80
    ? exc_page_fault+0x6a/0x190
    ? entry_SYSCALL_64_after_hwframe+0x61/0xcb
    </TASK>

  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 PREEMPT_RT disables
preemption only so the spin_lock_irq() in fpu_inherit_perms is unsafe
and converting siglock to raw spinlock would be an unwelcome change.
This problem exists since commit 9e798e9aa14c ("x86/fpu: Prepare fpu_clone()
for dynamically enabled features"). While the bug triggering is probably a
mistake for the affected machine and due to a bug that is not in mainline,
spin_lock_irq within a preempt_disable section on PREEMPT_RT is problematic.

In this specific context, it may not be necessary to hold fpregs_lock at
all. The lock is necessary when editing the FPU registers or a tasks fpstate
but in this case, the only write of any FP state in fpu_inherit_perms is
for the new child which is not running yet so it 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. The siglock still needs
to be acquired by fpu_inherit_perms as the read of the parents permissions
has to be serialised.

This is not tested as I did not access to a machine with Intel's
eXtended Feature Disable (XFD) feature that enables the relevant path
in fpu_inherit_perms and the bug is against a non-mainline kernel.

Reported-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
---
 arch/x86/kernel/fpu/core.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
  

On Wed, Nov 09 2022 at 11:30, Mel Gorman wrote:
>    BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:46
...
>   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 PREEMPT_RT disables
> preemption only so the spin_lock_irq() in fpu_inherit_perms is unsafe
> and converting siglock to raw spinlock would be an unwelcome change.
> This problem exists since commit 9e798e9aa14c ("x86/fpu: Prepare fpu_clone()
> for dynamically enabled features"). While the bug triggering is probably a
> mistake for the affected machine and due to a bug that is not in mainline,
> spin_lock_irq within a preempt_disable section on PREEMPT_RT is problematic.
>
> In this specific context, it may not be necessary to hold fpregs_lock at
> all. The lock is necessary when editing the FPU registers or a tasks fpstate
> but in this case, the only write of any FP state in fpu_inherit_perms is
> for the new child which is not running yet so it 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. The siglock still needs
> to be acquired by fpu_inherit_perms as the read of the parents permissions
> has to be serialised.

That's correct and siglock is the real protection for the permissions.

> This is not tested as I did not access to a machine with Intel's
> eXtended Feature Disable (XFD) feature that enables the relevant path
> in fpu_inherit_perms and the bug is against a non-mainline kernel.

It's still entirely correct on mainline as there is no requirement to
hold fpregs_lock in this case

> Reported-by: Mike Galbraith <efault@gmx.de>
> Signed-off-by: Mel Gorman <mgorman@techsingularity.net>

Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
  

On Wed, Nov 09, 2022 at 05:25:47PM +0100, Thomas Gleixner wrote:
> On Wed, Nov 09 2022 at 11:30, Mel Gorman wrote:
> >    BUG: sleeping function called from invalid context at kernel/locking/spinlock_rt.c:46
> ...
> >   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 PREEMPT_RT disables
> > preemption only so the spin_lock_irq() in fpu_inherit_perms is unsafe
> > and converting siglock to raw spinlock would be an unwelcome change.
> > This problem exists since commit 9e798e9aa14c ("x86/fpu: Prepare fpu_clone()
> > for dynamically enabled features"). While the bug triggering is probably a
> > mistake for the affected machine and due to a bug that is not in mainline,
> > spin_lock_irq within a preempt_disable section on PREEMPT_RT is problematic.
> >
> > In this specific context, it may not be necessary to hold fpregs_lock at
> > all. The lock is necessary when editing the FPU registers or a tasks fpstate
> > but in this case, the only write of any FP state in fpu_inherit_perms is
> > for the new child which is not running yet so it 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. The siglock still needs
> > to be acquired by fpu_inherit_perms as the read of the parents permissions
> > has to be serialised.
> 
> That's correct and siglock is the real protection for the permissions.
> 
> > This is not tested as I did not access to a machine with Intel's
> > eXtended Feature Disable (XFD) feature that enables the relevant path
> > in fpu_inherit_perms and the bug is against a non-mainline kernel.
> 
> It's still entirely correct on mainline as there is no requirement to
> hold fpregs_lock in this case
> 
> > Reported-by: Mike Galbraith <efault@gmx.de>
> > Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
> 
> Reviewed-by: Thomas Gleixner <tglx@linutronix.de>

Perfect, I'll rephase the changelog slightly and resend without RFC and
all the x86 maintainers cc'd. Thanks Thomas!