On Fri, Sep 10, 2021 at 05:25:15PM +0000, Michael Kelley wrote:

From: Wei Liu <wei.liu@kernel.org> Sent: Wednesday, September 8, 2021 12:43 PM

quoted

[...]

quoted

-static bool __send_ipi_mask(const struct cpumask *mask, int vector)
+static bool __send_ipi_mask(const struct cpumask *mask, int vector,
+		bool exclude_self)
 {
-	int cur_cpu, vcpu;
+	int cur_cpu, vcpu, this_cpu = smp_processor_id();
 	struct hv_send_ipi ipi_arg;
 	u64 status;

@@ -172,6 +177,8 @@ static bool __send_ipi_mask(const struct cpumask *mask, int vector)
 	ipi_arg.cpu_mask = 0;

 	for_each_cpu(cur_cpu, mask) {
+		if (exclude_self && cur_cpu == this_cpu)
+			continue;
 		vcpu = hv_cpu_number_to_vp_number(cur_cpu);
 		if (vcpu == VP_INVAL)
 			return false;

@@ -191,7 +198,7 @@ static bool __send_ipi_mask(const struct cpumask *mask, int vector)
 	return hv_result_success(status);

 do_ex_hypercall:
-	return __send_ipi_mask_ex(mask, vector);
+	return __send_ipi_mask_ex(mask, vector, exclude_self);
 }

This all looks correct to me, except for one difference compared with the
current code.  In the current code, if the cpumask passed to
hv_send_ipi_mask_allbutself() indicates only a single CPU that is "self",
__send_ipi_mask() will detect that the cpumask is now empty, and 
correctly return success without making the hypercall.  But
the new code will make the hypercall with an empty input mask (both
in the SEND_IPI and SEND_IPI_EX cases).   The Hyper-V TLFS is silent
on whether such a hypercall is a no-op that returns success or is an
error.  We'll have a problem if it is an error.  I think the safest thing
is to enhance the cpumask_empty() test at the beginning of
__send_ipi_mask() to also detect the case where only a single CPU
is specified, and it is "self".   This could be done using cpumask_weight()
and checking for zero as the "empty" case.   Then check for "1", and if
exclude_self is set, check if it is the "self" CPU.

Sure. Making this change should not be too difficult.

Wei.