iommu/arm-smmu-v3: Avoid back-to-back CMD_SYNC operations

Putting adjacent CMD_SYNCs into the command queue is nonsensical, but
can happen when multiple CPUs are inserting commands. Rather than leave
the poor old hardware to chew through these operations, we can instead
drop the subsequent SYNCs and poll for completion of the first. This
has been shown to improve IO performance under pressure, where the
number of SYNC operations reduces by about a third:

	CMD_SYNCs reduced:	19542181
	CMD_SYNCs total:	58098548	(include reduced)
	CMDs total:		116197099	(TLBI:SYNC about 1:1)

Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
Signed-off-by: Will Deacon <will.deacon@arm.com>

drivers/iommu/arm-smmu-v3.c

@@ -568,6 +568,7 @@ struct arm_smmu_device {
 	int				gerr_irq;
 	int				combined_irq;
 	u32				sync_nr;
+	u8				prev_cmd_opcode;
 
 	unsigned long			ias; /* IPA */
 	unsigned long			oas; /* PA */
@@ -901,6 +902,8 @@ static void arm_smmu_cmdq_insert_cmd(struct arm_smmu_device *smmu, u64 *cmd)
 	struct arm_smmu_queue *q = &smmu->cmdq.q;
 	bool wfe = !!(smmu->features & ARM_SMMU_FEAT_SEV);
 
+	smmu->prev_cmd_opcode = FIELD_GET(CMDQ_0_OP, cmd[0]);
+
 	while (queue_insert_raw(q, cmd) == -ENOSPC) {
 		if (queue_poll_cons(q, false, wfe))
 			dev_err_ratelimited(smmu->dev, "CMDQ timeout\n");
@@ -953,9 +956,16 @@ static int __arm_smmu_cmdq_issue_sync_msi(struct arm_smmu_device *smmu)
 	};
 
 	spin_lock_irqsave(&smmu->cmdq.lock, flags);
-	ent.sync.msidata = ++smmu->sync_nr;
-	arm_smmu_cmdq_build_cmd(cmd, &ent);
-	arm_smmu_cmdq_insert_cmd(smmu, cmd);
+
+	/* Piggy-back on the previous command if it's a SYNC */
+	if (smmu->prev_cmd_opcode == CMDQ_OP_CMD_SYNC) {
+		ent.sync.msidata = smmu->sync_nr;
+	} else {
+		ent.sync.msidata = ++smmu->sync_nr;
+		arm_smmu_cmdq_build_cmd(cmd, &ent);
+		arm_smmu_cmdq_insert_cmd(smmu, cmd);
+	}
+
 	spin_unlock_irqrestore(&smmu->cmdq.lock, flags);
 
 	return __arm_smmu_sync_poll_msi(smmu, ent.sync.msidata);