Re: [PATCH v1] LoongArch: Fix __smp_mb__{before,after}_atomic()

[Tiezhu Yang]

On 2026/8/18 下午10:05, Kumar Kartikeya Dwivedi wrote:
> On Tue Aug 18, 2026 at 3:50 PM CEST, Tiezhu Yang wrote:
> > On 2026/8/18 下午9:35, Kumar Kartikeya Dwivedi wrote:
> > > On Tue Aug 18, 2026 at 3:28 PM CEST, Tiezhu Yang wrote:
> > > > On 2026/8/17 上午11:59, Tiezhu Yang wrote:
> > > > > When testing the BPF selftest "sudo ./test_progs -t timer_lockup", there
> > > > > is a kernel lockup and panic:

...

> > > There is a full barrier in both cases. I don't know what performance improvement
> > > will be achieved by changing this.
> >
> > We performed benchmarking using UnixBench on a physical LoongArch
> > machine, the UnixBench score is different (amadd.w + dbar < amadd_db.w).
>
> I mean, this function is already quite heavy. We have multiple fully ordered
> atomics (refcount bumps/drops, xchg(), etc.) spread across the operation. Do you
> observe any measurable speedup in the throughput of this function?  The second
> question is whether timer cancellation is really frequent. In practice, I don't
> think that is the case.  The actual hrtimer_cancel() in itself is heavy and
> waits synchronously for the callback to finish.
>
> I'm not opposed to it or anything, I just don't think it's worth it in this
> case. If the latency of cancel is a problem there are other bigger opportunities
> to pursue than this.

Hi Kumar,

Thanks for your analysis. I totally agree that bpf_timer_cancel() is
heavy and cancellation is not a frequent fast path.

However, there is an interesting compiler and architecture background
here:

During our internal review, toolchain developers noted that executing
two separate instructions (amadd.w + dbar) causes a noticeable pipeline
bubble, and they are currently working on a compiler-level optimization
to automatically fuse them into a single ordered instruction amadd_db.w.

Since toolchain-level fusion into a single instruction is the clear
goal, explicitly using atomic_fetch_add() in the BPF helper achieves
this exact single-instruction benefit immediately, without waiting
for compiler updates.

More importantly, it hardens defensive programming, making the BPF
core logic immune to any architecture-level macro omissions.

To provide a concrete micro-architectural perspective, here is the
actual objdump comparison on a physical LoongArch machine between
the two approaches:

Approach 1 (Architecture fix: atomic_inc + smp_mb__after_atomic):
    2610:  amadd.w     $zero, $t2, $t3  /* Relaxed atomic increment */
    2614:  dbar        0x10             /* Explicit data memory barrier */

Approach 2 (Our BPF core change: atomic_fetch_add):
    2610:  amadd_db.w  $t3, $t1, $t4    /* Consolidated ordered atomic */

As the assembly shows, Approach 2 successfully eliminates a heavy
standalone hardware fence instruction (dbar 0x10).

What are your thoughts on this logic?

---
Appendix: Code Diff & Objdump for Reference

[1] Code Diff & Objdump for Approach 1:
diff --git a/arch/loongarch/include/asm/barrier.h 
b/arch/loongarch/include/asm/barrier.h
index 4b663f197706..adfe343dfa65 100644
--- a/arch/loongarch/include/asm/barrier.h
+++ b/arch/loongarch/include/asm/barrier.h
@@ -57,8 +57,8 @@
 #define __WEAK_LLSC_MB         "                       \n"
 #endif

-#define __smp_mb__before_atomic()      barrier()
-#define __smp_mb__after_atomic()       barrier()
+#define __smp_mb__before_atomic()      __smp_mb()
+#define __smp_mb__after_atomic()       __smp_mb()

$ objdump -d -S kernel/bpf/helpers.o
ATOMIC_OPS(add, i, add, +)
    2608:	0280040e 	li.w        	$t2, 1
    260c:	02c2e2ef 	addi.d      	$t3, $s0, 184
    2610:	386139e0 	amadd.w     	$zero, $t2, $t3
	smp_mb__after_atomic();
    2614:	38720010 	dbar        	0x10

[2] Code Diff & Objdump for Approach 2:
diff --git a/kernel/bpf/helpers.c b/kernel/bpf/helpers.c
index c18f1e16edee..ee2b3a4dcc05 100644
--- a/kernel/bpf/helpers.c
+++ b/kernel/bpf/helpers.c
@@ -1591,9 +1591,7 @@ BPF_CALL_1(bpf_timer_cancel, struct bpf_async_kern 
*, async)
         */
        if (!cur_t)
                goto drop;
-       atomic_inc(&t->cancelling);
-       /* Need full barrier after relaxed atomic_inc */
-       smp_mb__after_atomic();
+       atomic_fetch_add(1, &t->cancelling);
        inc = true;
        if (atomic_read(&cur_t->cancelling)) {

$ objdump -d -S kernel/bpf/helpers.o
ATOMIC_OPS(add, i, add, +)
    2608:	0280040d 	li.w        	$t1, 1
    260c:	02c2e2f0 	addi.d      	$t4, $s0, 184
    2610:	386a360f 	amadd_db.w  	$t3, $t1, $t4

Thanks,
Tiezhu