The recent adding of multicall debug mixed up the referencing of
the debug data. A __percpu tagged pointer can't be initialized with a
plain pointer, so use another percpu variable for the pointer and set
it on each new cpu via a function.

Fixes: 942d917cb92a ("xen: make multicall debug boot time selectable")
Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/oe-kbuild-all/202407151106.5s7Mnfpz-lkp@intel.com/
Signed-off-by: Juergen Gross <jgross@suse.com>
---
 arch/x86/xen/multicalls.c | 19 ++++++++++++-------
 arch/x86/xen/smp_pv.c     |  1 +
 arch/x86/xen/xen-ops.h    |  3 +++
 3 files changed, 16 insertions(+), 7 deletions(-)

diff --git a/arch/x86/xen/multicalls.c b/arch/x86/xen/multicalls.c
index d4cefd8a9af4..10c660fae8b3 100644
--- a/arch/x86/xen/multicalls.c
+++ b/arch/x86/xen/multicalls.c
@@ -54,8 +54,9 @@ struct mc_debug_data {
 
 static DEFINE_PER_CPU(struct mc_buffer, mc_buffer);
 static struct mc_debug_data mc_debug_data_early __initdata;
-static struct mc_debug_data __percpu *mc_debug_data __refdata =
+static DEFINE_PER_CPU(struct mc_debug_data *, mc_debug_data) =
 	&mc_debug_data_early;
+static struct mc_debug_data __percpu *mc_debug_data_ptr;
 DEFINE_PER_CPU(unsigned long, xen_mc_irq_flags);
 
 static struct static_key mc_debug __ro_after_init;
@@ -70,16 +71,20 @@ static int __init xen_parse_mc_debug(char *arg)
 }
 early_param("xen_mc_debug", xen_parse_mc_debug);
 
+void mc_percpu_init(unsigned int cpu)
+{
+	per_cpu(mc_debug_data, cpu) = per_cpu_ptr(mc_debug_data_ptr, cpu);
+}
+
 static int __init mc_debug_enable(void)
 {
-	struct mc_debug_data __percpu *mcdb;
 	unsigned long flags;
 
 	if (!mc_debug_enabled)
 		return 0;
 
-	mcdb = alloc_percpu(struct mc_debug_data);
-	if (!mcdb) {
+	mc_debug_data_ptr = alloc_percpu(struct mc_debug_data);
+	if (!mc_debug_data_ptr) {
 		pr_err("xen_mc_debug inactive\n");
 		static_key_slow_dec(&mc_debug);
 		return -ENOMEM;
@@ -88,7 +93,7 @@ static int __init mc_debug_enable(void)
 	/* Be careful when switching to percpu debug data. */
 	local_irq_save(flags);
 	xen_mc_flush();
-	mc_debug_data = mcdb;
+	mc_percpu_init(0);
 	local_irq_restore(flags);
 
 	pr_info("xen_mc_debug active\n");
@@ -150,7 +155,7 @@ void xen_mc_flush(void)
 	trace_xen_mc_flush(b->mcidx, b->argidx, b->cbidx);
 
 	if (static_key_false(&mc_debug)) {
-		mcdb = this_cpu_ptr(mc_debug_data);
+		mcdb = __this_cpu_read(mc_debug_data);
 		memcpy(mcdb->entries, b->entries,
 		       b->mcidx * sizeof(struct multicall_entry));
 	}
@@ -230,7 +235,7 @@ struct multicall_space __xen_mc_entry(size_t args)
 
 	ret.mc = &b->entries[b->mcidx];
 	if (static_key_false(&mc_debug)) {
-		struct mc_debug_data *mcdb = this_cpu_ptr(mc_debug_data);
+		struct mc_debug_data *mcdb = __this_cpu_read(mc_debug_data);
 
 		mcdb->caller[b->mcidx] = __builtin_return_address(0);
 		mcdb->argsz[b->mcidx] = args;
diff --git a/arch/x86/xen/smp_pv.c b/arch/x86/xen/smp_pv.c
index 7ea57f728b89..6863d3da7dec 100644
--- a/arch/x86/xen/smp_pv.c
+++ b/arch/x86/xen/smp_pv.c
@@ -305,6 +305,7 @@ static int xen_pv_kick_ap(unsigned int cpu, struct task_struct *idle)
 		return rc;
 
 	xen_pmu_init(cpu);
+	mc_percpu_init(cpu);
 
 	/*
 	 * Why is this a BUG? If the hypercall fails then everything can be
diff --git a/arch/x86/xen/xen-ops.h b/arch/x86/xen/xen-ops.h
index e7775dff9452..0cf16fc79e0b 100644
--- a/arch/x86/xen/xen-ops.h
+++ b/arch/x86/xen/xen-ops.h
@@ -257,6 +257,9 @@ void xen_mc_callback(void (*fn)(void *), void *data);
  */
 struct multicall_space xen_mc_extend_args(unsigned long op, size_t arg_size);
 
+/* Do percpu data initialization for multicalls. */
+void mc_percpu_init(unsigned int cpu);
+
 extern bool is_xen_pmu;
 
 irqreturn_t xen_pmu_irq_handler(int irq, void *dev_id);
-- 
2.43.0

On 7/18/24 2:21 AM, Juergen Gross wrote:
> The recent adding of multicall debug mixed up the referencing of
> the debug data. A __percpu tagged pointer can't be initialized with a
> plain pointer, so use another percpu variable for the pointer and set
> it on each new cpu via a function.
>
> Fixes: 942d917cb92a ("xen: make multicall debug boot time selectable")
> Reported-by: kernel test robot <lkp@intel.com>
> Closes: https://lore.kernel.org/oe-kbuild-all/202407151106.5s7Mnfpz-lkp@intel.com/
> Signed-off-by: Juergen Gross <jgross@suse.com>


IIUIC we only need data until SMP is initialized, so setting per-cpu 
pointer just on cpu0 may be sufficient. Not sure where to do this though.

But this works as well.

Reviewed-by: Boris Ostrovsky <boris.ostrovsky@oracle.com>