Skip to contentSkip to conversation

x86/speculation: Update TIF_SPEC_IB before ibpb barrier

PATCH In review Zhenzhong Duan 2 participants 7 messages Jan 18, 2019 – Jan 28, 2019

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86/speculation: Update TIF_SPEC_IB before ibpb barrier

When a task is set for updating TIF_SPEC_IB throuth SECCOMP by others and it's scheduled in the first time, a stale TIF_SPEC_IB value is picked in cond_ibpb(). This is due to TIF_SPEC_IB is updated later at __switch_to_xtra().

Add an extra call to speculation_ctrl_update_tif() to update it before IBPB barrier.

---
 arch/x86/include/asm/spec-ctrl.h | 1 +
 arch/x86/kernel/process.c        | 2 +-
 arch/x86/mm/tlb.c                | 4 +++-
 3 files changed, 5 insertions(+), 2 deletions(-)
arch/x86/include/asm/spec-ctrl.h +1 -0
index 5393bab..8b2814a 100644
@@ -82,6 +82,7 @@ static inline u64 ssbd_tif_to_amd_ls_cfg(u64 tifn)
 static inline void speculative_store_bypass_ht_init(void) { }
 #endif
 
+extern unsigned long speculation_ctrl_update_tif(struct task_struct *tsk);
 extern void speculation_ctrl_update(unsigned long tif);
 extern void speculation_ctrl_update_current(void);
 
arch/x86/kernel/process.c +1 -1
index 90ae0ca..454e71d 100644
@@ -446,7 +446,7 @@ static __always_inline void __speculation_ctrl_update(unsigned long tifp,
 		wrmsrl(MSR_IA32_SPEC_CTRL, msr);
 }
 
-static unsigned long speculation_ctrl_update_tif(struct task_struct *tsk)
+unsigned long speculation_ctrl_update_tif(struct task_struct *tsk)
 {
 	if (test_and_clear_tsk_thread_flag(tsk, TIF_SPEC_FORCE_UPDATE)) {
 		if (task_spec_ssb_disable(tsk))
arch/x86/mm/tlb.c +3 -1
index 999d6d8..c0f3fcf 100644
@@ -8,6 +8,7 @@
 #include <linux/cpu.h>
 #include <linux/debugfs.h>
 
+#include <asm/spec-ctrl.h>
 #include <asm/tlbflush.h>
 #include <asm/mmu_context.h>
 #include <asm/nospec-branch.h>
@@ -190,7 +191,8 @@ static void sync_current_stack_to_mm(struct mm_struct *mm)
 
 static inline unsigned long mm_mangle_tif_spec_ib(struct task_struct *next)
 {
-	unsigned long next_tif = task_thread_info(next)->flags;
+	/* Update the flag bits to newest value actively */
+	unsigned long next_tif = speculation_ctrl_update_tif(next);
 	unsigned long ibpb = (next_tif >> TIF_SPEC_IB) & LAST_USER_MM_IBPB;
 
 	return (unsigned long)next->mm | ibpb;
Signature · 1 line
1.8.3.1
#1 lore raw
Quoted 7 lines · Zhenzhong Duan

When a task is set for updating TIF_SPEC_IB throuth SECCOMP by others and it's scheduled in the first time, a stale TIF_SPEC_IB value is picked in cond_ibpb(). This is due to TIF_SPEC_IB is updated later at __switch_to_xtra().

Add an extra call to speculation_ctrl_update_tif() to update it before IBPB barrier.

Errm. No. It adds that call to speculation_ctrl_update_tif() for every mm switch, most of the time for nothing.

If at all, and we discussed that before and decided not to worry about it (because it gets fixed up on the next context switch), then you want to handle ibpb() from there:

        if (likely(!((tifp | tifn) & _TIF_SPEC_FORCE_UPDATE))) {
                __speculation_ctrl_update(tifp, tifn);
        } else {
                speculation_ctrl_update_tif(prev_p);
                tifn = speculation_ctrl_update_tif(next_p);

                /* Enforce MSR update to ensure consistent state */
                __speculation_ctrl_update(~tifn, tifn);

------> Force update IBPB

        }

Thanks,

tglx

#2 lore raw
Quoted 10 lines · Thomas Gleixner
Quoted 7 lines · Zhenzhong Duan

When a task is set for updating TIF_SPEC_IB throuth SECCOMP by others and it's scheduled in the first time, a stale TIF_SPEC_IB value is picked in cond_ibpb(). This is due to TIF_SPEC_IB is updated later at __switch_to_xtra().

Add an extra call to speculation_ctrl_update_tif() to update it before IBPB barrier.

Quoted 6 lines

Errm. No. It adds that call to speculation_ctrl_update_tif() for every mm switch, most of the time for nothing.

If at all, and we discussed that before and decided not to worry about it (because it gets fixed up on the next context switch), then you want to handle ibpb() from there:

Actually we need to do that. It's not only the scheduled in first problem. That whole thing might become completely stale in either direction. Care to whip up a patch?

Thanks,

tglx

#3 lore raw
Quoted 15 lines · Thomas Gleixner
Quoted 13 lines · Thomas Gleixner
Quoted 7 lines · Zhenzhong Duan

When a task is set for updating TIF_SPEC_IB throuth SECCOMP by others and it's scheduled in the first time, a stale TIF_SPEC_IB value is picked in cond_ibpb(). This is due to TIF_SPEC_IB is updated later at __switch_to_xtra().

Add an extra call to speculation_ctrl_update_tif() to update it before IBPB barrier.

Errm. No. It adds that call to speculation_ctrl_update_tif() for every mm switch, most of the time for nothing.

Quoted 7 lines
Quoted 3 lines

If at all, and we discussed that before and decided not to worry about it (because it gets fixed up on the next context switch), then you want to handle ibpb() from there:

Actually we need to do that. It's not only the scheduled in first problem. That whole thing might become completely stale in either direction. Care to whip up a patch?

Bah, nonsense. Brain was clearly still out for lunch and I confused IBPB and STIBP for a moment. cond_ibpb() is the thing issues in switch_mm() and that is not leaving a stale MSR around because we only write to it when we need the barrier. The bit is not stale because the barrier is only issued with the write. The bit has not to be cleared.

So the only 'issue' what happens is that switch_to() either issues a barrier too much or misses one. That's really not a problem.

Thanks,

tglx

#4 lore raw
Quoted 24 lines · Thomas Gleixner
Quoted 15 lines · Thomas Gleixner
Quoted 13 lines · Thomas Gleixner
Quoted 7 lines · Zhenzhong Duan

When a task is set for updating TIF_SPEC_IB throuth SECCOMP by others and it's scheduled in the first time, a stale TIF_SPEC_IB value is picked in cond_ibpb(). This is due to TIF_SPEC_IB is updated later at __switch_to_xtra().

Add an extra call to speculation_ctrl_update_tif() to update it before IBPB barrier.

Errm. No. It adds that call to speculation_ctrl_update_tif() for every mm switch, most of the time for nothing.

Quoted 7 lines
Quoted 3 lines

If at all, and we discussed that before and decided not to worry about it (because it gets fixed up on the next context switch), then you want to handle ibpb() from there:

Actually we need to do that. It's not only the scheduled in first problem. That whole thing might become completely stale in either direction. Care to whip up a patch?

Quoted 8 lines

Bah, nonsense. Brain was clearly still out for lunch and I confused IBPB and STIBP for a moment. cond_ibpb() is the thing issues in switch_mm() and that is not leaving a stale MSR around because we only write to it when we need the barrier. The bit is not stale because the barrier is only issued with the write. The bit has not to be cleared.

So the only 'issue' what happens is that switch_to() either issues a barrier too much or misses one. That's really not a problem.

Ok, yes, the purpose of this patch is to avoid the one missed barrier. Thanks for your reply.

Zhenzhong

#5 lore raw
Quoted 11 lines · Zhenzhong Duan
Quoted 8 lines · Thomas Gleixner

Bah, nonsense. Brain was clearly still out for lunch and I confused IBPB and STIBP for a moment. cond_ibpb() is the thing issues in switch_mm() and that is not leaving a stale MSR around because we only write to it when we need the barrier. The bit is not stale because the barrier is only issued with the write. The bit has not to be cleared.

So the only 'issue' what happens is that switch_to() either issues a barrier too much or misses one. That's really not a problem.

Ok, yes, the purpose of this patch is to avoid the one missed barrier.

And that missed barrier is not worth it to do extra work in switch_to/mm simply because it's a one off event and there is no way to exploit that reliably.

Thanks,

tglx

#6 lore raw
Quoted 8 lines · Thomas Gleixner
Quoted 7 lines · Zhenzhong Duan
Quoted 6 lines · Thomas Gleixner

Bah, nonsense. Brain was clearly still out for lunch and I confused IBPB and STIBP for a moment. cond_ibpb() is the thing issues in switch_mm() and that is not leaving a stale MSR around because we only write to it when we need the barrier. The bit is not stale because the barrier is only issued with the write. The bit has not to be cleared.

Quoted 8 lines
Quoted 4 lines
Quoted 2 lines

So the only 'issue' what happens is that switch_to() either issues a barrier too much or misses one. That's really not a problem.

Ok, yes, the purpose of this patch is to avoid the one missed barrier.

And that missed barrier is not worth it to do extra work in switch_to/mm simply because it's a one off event and there is no way to exploit that reliably.

Got it.

Thanks
Zhenzhong