0-Day found a 34.6% regression in stress-ng's 'af-alg' test case, and
bisected it to commit b81fac906a8f ("x86/fpu: Move FPU initialization into
arch_cpu_finalize_init()"), which optimizes the FPU init order, and moves
the CR4_OSXSAVE enabling into a later place:

   arch_cpu_finalize_init
       identify_boot_cpu
	   identify_cpu
	       generic_identify
                   get_cpu_cap --> setup cpu capability
       ...
       fpu__init_cpu
           fpu__init_cpu_xstate
               cr4_set_bits(X86_CR4_OSXSAVE);

As the FPU is not yet initialized the CPU capability setup fails to set
X86_FEATURE_OSXSAVE. Many security module like 'camellia_aesni_avx_x86_64'
depend on this feature and therefore fail to load, causing the regression.

Cure this by setting X86_FEATURE_OSXSAVE feature right after OSXSAVE
enabling.

[ tglx: Moved it into the actual BSP FPU initialization code and added a comment ]

Fixes: b81fac906a8f ("x86/fpu: Move FPU initialization into arch_cpu_finalize_init()")
Reported-by: kernel test robot <oliver.sang@intel.com>
Signed-off-by: Feng Tang <feng.tang@intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/lkml/202307192135.203ac24e-oliver.sang@intel.com
Link: https://lore.kernel.org/lkml/20230823065747.92257-1-feng.tang@intel.com

The thread flag TIF_NEED_FPU_LOAD indicates that the FPU saved state is
valid and should be reloaded when returning to userspace. However, the
kernel will skip doing this if the FPU registers are already valid as
determined by fpregs_state_valid(). The logic embedded there considers
the state valid if two cases are both true:

  1: fpu_fpregs_owner_ctx points to the current tasks FPU state
  2: the last CPU the registers were live in was the current CPU.

This is usually correct logic. A CPU’s fpu_fpregs_owner_ctx is set to
the current FPU during the fpregs_restore_userregs() operation, so it
indicates that the registers have been restored on this CPU. But this
alone doesn’t preclude that the task hasn’t been rescheduled to a
different CPU, where the registers were modified, and then back to the
current CPU. To verify that this was not the case the logic relies on the
second condition. So the assumption is that if the registers have been
restored, AND they haven’t had the chance to be modified (by being
loaded on another CPU), then they MUST be valid on the current CPU.

Besides the lazy FPU optimizations, the other cases where the FPU
registers might not be valid are when the kernel modifies the FPU register
state or the FPU saved buffer. In this case the operation modifying the
FPU state needs to let the kernel know the correspondence has been
broken. The comment in “arch/x86/kernel/fpu/context.h” has:
/*
...
 * If the FPU register state is valid, the kernel can skip restoring the
 * FPU state from memory.
 *
 * Any code that clobbers the FPU registers or updates the in-memory
 * FPU state for a task MUST let the rest of the kernel know that the
 * FPU registers are no longer valid for this task.
 *
 * Either one of these invalidation functions is enough. Invalidate
 * a resource you control: CPU if using the CPU for something else
 * (with preemption disabled), FPU for the current task, or a task that
 * is prevented from running by the current task.
 */

However, this is not completely true. When the kernel modifies the
registers or saved FPU state, it can only rely on
__fpu_invalidate_fpregs_state(), which wipes the FPU’s last_cpu
tracking. The exec path instead relies on fpregs_deactivate(), which sets
the CPU’s FPU context to NULL. This was observed to fail to restore the
reset FPU state to the registers when returning to userspace in the
following scenario:

1. A task is executing in userspace on CPU0
	- CPU0’s FPU context points to tasks
	- fpu->last_cpu=CPU0

2. The task exec()’s

3. While in the kernel the task is preempted
	- CPU0 gets a thread executing in the kernel (such that no other
		FPU context is activated)
	- Scheduler sets task’s fpu->last_cpu=CPU0 when scheduling out

4. Task is migrated to CPU1

5. Continuing the exec(), the task gets to
   fpu_flush_thread()->fpu_reset_fpregs()
	- Sets CPU1’s fpu context to NULL
	- Copies the init state to the task’s FPU buffer
	- Sets TIF_NEED_FPU_LOAD on the task

6. The task reschedules back to CPU0 before completing the exec() and
   returning to userspace
	- During the reschedule, scheduler finds TIF_NEED_FPU_LOAD is set
	- Skips saving the registers and updating task’s fpu→last_cpu,
	  because TIF_NEED_FPU_LOAD is the canonical source.

7. Now CPU0’s FPU context is still pointing to the task’s, and
   fpu->last_cpu is still CPU0. So fpregs_state_valid() returns true even
   though the reset FPU state has not been restored.

So the root cause is that exec() is doing the wrong kind of invalidate. It
should reset fpu->last_cpu via __fpu_invalidate_fpregs_state(). Further,
fpu__drop() doesn't really seem appropriate as the task (and FPU) are not
going away, they are just getting reset as part of an exec. So switch to
__fpu_invalidate_fpregs_state().

Also, delete the misleading comment that says that either kind of
invalidate will be enough, because it’s not always the case.

Fixes: 33344368cb08 ("x86/fpu: Clean up the fpu__clear() variants")
Reported-by: Lei Wang <lei4.wang@intel.com>
Signed-off-by: Rick Edgecombe <rick.p.edgecombe@intel.com>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Lijun Pan <lijun.pan@intel.com>
Reviewed-by: Sohil Mehta <sohil.mehta@intel.com>
Acked-by: Lijun Pan <lijun.pan@intel.com>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20230818170305.502891-1-rick.p.edgecombe@intel.com

Specify how is SRSO mitigated when SMT is disabled. Also, correct the
SMT check for that.

Fixes: e9fbc47b818b ("x86/srso: Disable the mitigation on unaffected configurations")
Suggested-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Acked-by: Josh Poimboeuf <jpoimboe@kernel.org>
Link: https://lore.kernel.org/r/20230814200813.p5czl47zssuej7nv@treble

Christian reported spurious module load crashes after some of Song's
module memory layout patches.

Turns out that if the very last instruction on the very last page of the
module is a 'JMP __x86_return_thunk' then __static_call_fixup() will
trip a fault and die.

And while the module rework made this slightly more likely to happen,
it's always been possible.

Fixes: ee88d363d156 ("x86,static_call: Use alternative RET encoding")
Reported-by: Christian Bricart <christian@bricart.de>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Josh Poimboeuf <jpoimboe@kernel.org>
Link: https://lkml.kernel.org/r/20230816104419.GA982867@hirez.programming.kicks-ass.net

The goal is to eventually have a proper documentation about all this.

Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Link: https://lore.kernel.org/r/20230814164447.GFZNpZ/64H4lENIe94@fat_crate.local

Similar to how it doesn't make sense to have UNTRAIN_RET have two
untrain calls, it also doesn't make sense for VMEXIT to have an extra
IBPB call.

This cures VMEXIT doing potentially unret+IBPB or double IBPB.
Also, the (SEV) VMEXIT case seems to have been overlooked.

Redefine the meaning of the synthetic IBPB flags to:

 - ENTRY_IBPB     -- issue IBPB on entry  (was: entry + VMEXIT)
 - IBPB_ON_VMEXIT -- issue IBPB on VMEXIT

And have 'retbleed=ibpb' set *BOTH* feature flags to ensure it retains
the previous behaviour and issues IBPB on entry+VMEXIT.

The new 'srso=ibpb_vmexit' option only sets IBPB_ON_VMEXIT.

Create UNTRAIN_RET_VM specifically for the VMEXIT case, and have that
check IBPB_ON_VMEXIT.

All this avoids having the VMEXIT case having to check both ENTRY_IBPB
and IBPB_ON_VMEXIT and simplifies the alternatives.

Fixes: fb3bd914b3ec ("x86/srso: Add a Speculative RAS Overflow mitigation")
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Link: https://lore.kernel.org/r/20230814121149.109557833@infradead.org

Since there can only be one active return_thunk, there only needs be
one (matching) untrain_ret. It fundamentally doesn't make sense to
allow multiple untrain_ret at the same time.

Fold all the 3 different untrain methods into a single (temporary)
helper stub.

Fixes: fb3bd914b3ec ("x86/srso: Add a Speculative RAS Overflow mitigation")
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Link: https://lore.kernel.org/r/20230814121149.042774962@infradead.org

For a more consistent namespace.

  [ bp: Fixup names in the doc too. ]

Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Link: https://lore.kernel.org/r/20230814121148.976236447@infradead.org

Rename the original retbleed return thunk and untrain_ret to
retbleed_return_thunk() and retbleed_untrain_ret().

No functional changes.

Suggested-by: Josh Poimboeuf <jpoimboe@kernel.org>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Link: https://lore.kernel.org/r/20230814121148.909378169@infradead.org

Use the existing configurable return thunk. There is absolute no
justification for having created this __x86_return_thunk alternative.

To clarify, the whole thing looks like:

Zen3/4 does:

  srso_alias_untrain_ret:
	  nop2
	  lfence
	  jmp srso_alias_return_thunk
	  int3

  srso_alias_safe_ret: // aliasses srso_alias_untrain_ret just so
	  add $8, %rsp
	  ret
	  int3

  srso_alias_return_thunk:
	  call srso_alias_safe_ret
	  ud2

While Zen1/2 does:

  srso_untrain_ret:
	  movabs $foo, %rax
	  lfence
	  call srso_safe_ret           (jmp srso_return_thunk ?)
	  int3

  srso_safe_ret: // embedded in movabs instruction
	  add $8,%rsp
          ret
          int3

  srso_return_thunk:
	  call srso_safe_ret
	  ud2

While retbleed does:

  zen_untrain_ret:
	  test $0xcc, %bl
	  lfence
	  jmp zen_return_thunk
          int3

  zen_return_thunk: // embedded in the test instruction
	  ret
          int3

Where Zen1/2 flush the BTB entry using the instruction decoder trick
(test,movabs) Zen3/4 use BTB aliasing. SRSO adds a return sequence
(srso_safe_ret()) which forces the function return instruction to
speculate into a trap (UD2).  This RET will then mispredict and
execution will continue at the return site read from the top of the
stack.

Pick one of three options at boot (evey function can only ever return
once).

  [ bp: Fixup commit message uarch details and add them in a comment in
    the code too. Add a comment about the srso_select_mitigation()
    dependency on retbleed_select_mitigation(). Add moar ifdeffery for
    32-bit builds. Add a dummy srso_untrain_ret_alias() definition for
    32-bit alternatives needing the symbol. ]

Fixes: fb3bd914b3ec ("x86/srso: Add a Speculative RAS Overflow mitigation")
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Link: https://lore.kernel.org/r/20230814121148.842775684@infradead.org