x86/fpu/xstate: Restore supervisor states for signal return

The signal return fast path directly restores user states from the user
buffer. Once that succeeds, restore supervisor states (but only when
they are not yet restored).

For the slow path, save supervisor states to preserve them across context
switches, and restore after the user states are restored.

The previous version has the overhead of an XSAVES in both the fast and the
slow paths.  It is addressed as the following:

- In the fast path, only do an XRSTORS.
- In the slow path, do a supervisor-state-only XSAVES, and relocate the
  buffer contents.

Some thoughts in the implementation:

- In the slow path, can any supervisor state become stale between
  save/restore?

  Answer: set_thread_flag(TIF_NEED_FPU_LOAD) protects the xstate buffer.

- In the slow path, can any code reference a stale supervisor state
  register between save/restore?

  Answer: In the current lazy-restore scheme, any reference to xstate
  registers needs fpregs_lock()/fpregs_unlock() and __fpregs_load_activate().

- Are there other options?

  One other option is eagerly restoring all supervisor states.

  Currently, CET user-mode states and ENQCMD's PASID do not need to be
  eagerly restored.  The upcoming CET kernel-mode states (24 bytes) need
  to be eagerly restored.  To me, eagerly restoring all supervisor states
  adds more overhead then benefit at this point.

Signed-off-by: Yu-cheng Yu <yu-cheng.yu@intel.com>
Signed-off-by: Borislav Petkov <bp@suse.de>
Reviewed-by: Dave Hansen <dave.hansen@linux.intel.com>
Link: https://lkml.kernel.org/r/20200512145444.15483-11-yu-cheng.yu@intel.com

diff --git a/arch/x86/kernel/fpu/signal.c b/arch/x86/kernel/fpu/signal.c
index 6184fe70a9e1f0a2eee6044bbc9905112887d184..9393a445d73c81f97b912969fc6a6ce2b9952e27 100644 (file)
--- a/arch/x86/kernel/fpu/signal.c
+++ b/arch/x86/kernel/fpu/signal.c
@@ -347,6 +347,23 @@ static int __fpu__restore_sig(void __user *buf, void __user *buf_fx, int size)
                ret = copy_user_to_fpregs_zeroing(buf_fx, user_xfeatures, fx_only);
                pagefault_enable();
                if (!ret) {
+
+                       /*
+                        * Restore supervisor states: previous context switch
+                        * etc has done XSAVES and saved the supervisor states
+                        * in the kernel buffer from which they can be restored
+                        * now.
+                        *
+                        * We cannot do a single XRSTORS here - which would
+                        * be nice - because the rest of the FPU registers are
+                        * being restored from a user buffer directly. The
+                        * single XRSTORS happens below, when the user buffer
+                        * has been copied to the kernel one.
+                        */
+                       if (test_thread_flag(TIF_NEED_FPU_LOAD) &&
+                           xfeatures_mask_supervisor())
+                               copy_kernel_to_xregs(&fpu->state.xsave,
+                                                    xfeatures_mask_supervisor());
                        fpregs_mark_activate();
                        fpregs_unlock();
                        return 0;
@@ -364,14 +381,25 @@ static int __fpu__restore_sig(void __user *buf, void __user *buf_fx, int size)
        }
 
        /*
-        * The current state of the FPU registers does not matter. By setting
-        * TIF_NEED_FPU_LOAD unconditionally it is ensured that the our xstate
-        * is not modified on context switch and that the xstate is considered
+        * By setting TIF_NEED_FPU_LOAD it is ensured that our xstate is
+        * not modified on context switch and that the xstate is considered
         * to be loaded again on return to userland (overriding last_cpu avoids
         * the optimisation).
         */
-       set_thread_flag(TIF_NEED_FPU_LOAD);
+       fpregs_lock();
+
+       if (!test_thread_flag(TIF_NEED_FPU_LOAD)) {
+
+               /*
+                * Supervisor states are not modified by user space input.  Save
+                * current supervisor states first and invalidate the FPU regs.
+                */
+               if (xfeatures_mask_supervisor())
+                       copy_supervisor_to_kernel(&fpu->state.xsave);
+               set_thread_flag(TIF_NEED_FPU_LOAD);
+       }
        __fpu_invalidate_fpregs_state(fpu);
+       fpregs_unlock();
 
        if (use_xsave() && !fx_only) {
                u64 init_bv = xfeatures_mask_user() & ~user_xfeatures;
@@ -393,7 +421,13 @@ static int __fpu__restore_sig(void __user *buf, void __user *buf_fx, int size)
                fpregs_lock();
                if (unlikely(init_bv))
                        copy_kernel_to_xregs(&init_fpstate.xsave, init_bv);
-               ret = copy_kernel_to_xregs_err(&fpu->state.xsave, user_xfeatures);
+
+               /*
+                * Restore previously saved supervisor xstates along with
+                * copied-in user xstates.
+                */
+               ret = copy_kernel_to_xregs_err(&fpu->state.xsave,
+                                              user_xfeatures | xfeatures_mask_supervisor());
 
        } else if (use_fxsr()) {
                ret = __copy_from_user(&fpu->state.fxsave, buf_fx, state_size);