[PATCH] powerpc/kvm-hv-pmu: don't allocate under lock_l0_stats in kvmppc_init_hostwide

[PATCH] powerpc/kvm-hv-pmu: don't allocate under lock_l0_stats in kvmppc_init_hostwide

[Hyunjung Ko]

kvmppc_init_hostwide() builds the hostwide L0-stats guest-state message and
buffer with kvmppc_gsm_new() and kvmppc_gsb_new() while holding
lock_l0_stats, which is taken with spin_lock_irqsave(). Both helpers
allocate with GFP_KERNEL (and kvmppc_gsb_new() does a further GFP_KERNEL
allocation internally), so they can sleep in atomic context, triggering
"BUG: sleeping function called from invalid context" under
CONFIG_DEBUG_ATOMIC_SLEEP.

lock_l0_stats is a spinlock because it also serializes the PMU ->read path
(kvmppc_pmu_event_update()), so it cannot be turned into a mutex. Build the
message and buffer in local variables outside the lock and only take the
lock to publish them, rechecking for a concurrent registration.

Fixes: f0c9c49c5067 ("powerpc/kvm-hv-pmu: Implement GSB message-ops for hostwide counters")
Cc: stable@vger.kernel.org
Signed-off-by: Hyunjung Ko <hj351016@gmail.com>
---
 arch/powerpc/perf/kvm-hv-pmu.c | 70 ++++++++++++++++++++++++++----------------
 1 file changed, 42 insertions(+), 28 deletions(-)

diff --git a/arch/powerpc/perf/kvm-hv-pmu.c b/arch/powerpc/perf/kvm-hv-pmu.c
index aa72b96b5a8c..deba59cbe234 100644
--- a/arch/powerpc/perf/kvm-hv-pmu.c
+++ b/arch/powerpc/perf/kvm-hv-pmu.c
@@ -314,53 +314,67 @@ static struct kvmppc_gs_msg_ops gsb_ops_l0_stats = {
 
 static int kvmppc_init_hostwide(void)
 {
-	int rc = 0;
+	struct kvmppc_gs_msg *gsm;
+	struct kvmppc_gs_buff *gsb;
 	unsigned long flags;
+	int rc;
 
 	spin_lock_irqsave(&lock_l0_stats, flags);
-
 	/* already registered ? */
 	if (gsm_l0_stats) {
-		rc = 0;
-		goto out;
+		spin_unlock_irqrestore(&lock_l0_stats, flags);
+		return 0;
 	}
+	spin_unlock_irqrestore(&lock_l0_stats, flags);
+
+	/*
+	 * kvmppc_gsm_new() and kvmppc_gsb_new() allocate with GFP_KERNEL and
+	 * may sleep, so build the message and buffer outside lock_l0_stats,
+	 * which is a spinlock (it also serializes the PMU ->read path), and
+	 * only take the lock to publish them.
+	 */
 
 	/* setup the Guest state message/buffer to talk to L0 */
-	gsm_l0_stats = kvmppc_gsm_new(&gsb_ops_l0_stats, &l0_stats,
-				      GSM_SEND, GFP_KERNEL);
-	if (!gsm_l0_stats) {
-		rc = -ENOMEM;
-		goto out;
-	}
+	gsm = kvmppc_gsm_new(&gsb_ops_l0_stats, &l0_stats, GSM_SEND, GFP_KERNEL);
+	if (!gsm)
+		return -ENOMEM;
 
 	/* Populate the Idents */
-	kvmppc_gsm_include(gsm_l0_stats, KVMPPC_GSID_L0_GUEST_HEAP);
-	kvmppc_gsm_include(gsm_l0_stats, KVMPPC_GSID_L0_GUEST_HEAP_MAX);
-	kvmppc_gsm_include(gsm_l0_stats, KVMPPC_GSID_L0_GUEST_PGTABLE_SIZE);
-	kvmppc_gsm_include(gsm_l0_stats, KVMPPC_GSID_L0_GUEST_PGTABLE_SIZE_MAX);
-	kvmppc_gsm_include(gsm_l0_stats, KVMPPC_GSID_L0_GUEST_PGTABLE_RECLAIM);
+	kvmppc_gsm_include(gsm, KVMPPC_GSID_L0_GUEST_HEAP);
+	kvmppc_gsm_include(gsm, KVMPPC_GSID_L0_GUEST_HEAP_MAX);
+	kvmppc_gsm_include(gsm, KVMPPC_GSID_L0_GUEST_PGTABLE_SIZE);
+	kvmppc_gsm_include(gsm, KVMPPC_GSID_L0_GUEST_PGTABLE_SIZE_MAX);
+	kvmppc_gsm_include(gsm, KVMPPC_GSID_L0_GUEST_PGTABLE_RECLAIM);
 
 	/* allocate GSB. Guest/Vcpu Id is ignored */
-	gsb_l0_stats = kvmppc_gsb_new(kvmppc_gsm_size(gsm_l0_stats), 0, 0,
-				      GFP_KERNEL);
-	if (!gsb_l0_stats) {
+	gsb = kvmppc_gsb_new(kvmppc_gsm_size(gsm), 0, 0, GFP_KERNEL);
+	if (!gsb) {
 		rc = -ENOMEM;
-		goto out;
+		goto free_gsm;
 	}
 
 	/* ask the ops to fill in the info */
-	rc = kvmppc_gsm_fill_info(gsm_l0_stats, gsb_l0_stats);
+	rc = kvmppc_gsm_fill_info(gsm, gsb);
+	if (rc)
+		goto free_gsb;
 
-out:
-	if (rc) {
-		if (gsm_l0_stats)
-			kvmppc_gsm_free(gsm_l0_stats);
-		if (gsb_l0_stats)
-			kvmppc_gsb_free(gsb_l0_stats);
-		gsm_l0_stats = NULL;
-		gsb_l0_stats = NULL;
+	spin_lock_irqsave(&lock_l0_stats, flags);
+	if (gsm_l0_stats) {
+		/* raced with another CPU; use the existing registration */
+		spin_unlock_irqrestore(&lock_l0_stats, flags);
+		rc = 0;
+		goto free_gsb;
 	}
+	gsm_l0_stats = gsm;
+	gsb_l0_stats = gsb;
 	spin_unlock_irqrestore(&lock_l0_stats, flags);
+
+	return 0;
+
+free_gsb:
+	kvmppc_gsb_free(gsb);
+free_gsm:
+	kvmppc_gsm_free(gsm);
 	return rc;
 }
 
-- 
2.43.0