[PATCH v4 0/2] bugfix and optimization about CMD_SYNC

[PATCH v4 0/2] bugfix and optimization about CMD_SYNC

[Zhen Lei]

v3->v4:
1. create a new function arm_smmu_cmdq_build_sync_msi_cmd, it's only used to
build CMD_SYNC for CS=SIG_IRQ mode.
2. In order to observe the optimization effect, I conducted 5 tests for each
case. Although the test result is volatility, but we can still get which case
is good or bad.

Test command: fio -numjobs=8 -rw=randread -runtime=30 ... -bs=4k
Test Result: IOPS

Case 1: (without these patches)
675480
672055
665275
648610
661146

Case 2: (only apply the variant of patch 1, move arm_smmu_cmdq_build_cmd into lock)
688714
697355
632951
700540
678459

Case 3: (only apply patch 1)
721582
729226
689574
679710
727770

Case 4: (apply both patch 1 and patch 2)
734077
742868
738194
682544
740586

v2 -> v3:
Although I have no data to show how many performance will be impacted
because of arm_smmu_cmdq_build_cmd is protected by spinlock. But it's
clear that the performance is bound to drop, a memset operation and 
a complicate switch..case in the function arm_smmu_cmdq_build_cmd.

v1 -> v2:
1. move the call to arm_smmu_cmdq_build_cmd into the critical section,
   and keep itself unchange.
2. Although patch2 can make sure no two CMD_SYNCs will be adjacent,
but patch1 is still needed, see below:

cpu0			cpu1			cpu2
msidata=0
			msidata=1
			insert cmd1
						insert a TLBI command
insert cmd0
			smmu execute cmd1
						smmu execute TLBI
smmu execute cmd0
			poll timeout, because msidata=1 is overridden by
			cmd0, that means VAL=0, sync_idx=1.

Zhen Lei (2):
  iommu/arm-smmu-v3: fix unexpected CMD_SYNC timeout
  iommu/arm-smmu-v3: avoid redundant CMD_SYNCs if possible

 drivers/iommu/arm-smmu-v3.c | 44 ++++++++++++++++++++++++++++++++------------
 1 file changed, 32 insertions(+), 12 deletions(-)

-- 
1.8.3

[PATCH v4 1/2] iommu/arm-smmu-v3: fix unexpected CMD_SYNC timeout

[Zhen Lei]

The condition "(int)(VAL - sync_idx) >= 0" to break loop in function
__arm_smmu_sync_poll_msi requires that sync_idx must be increased
monotonously according to the sequence of the CMDs in the cmdq.

But ".msidata = atomic_inc_return_relaxed(&smmu->sync_nr)" is not protected
by spinlock, so the following scenarios may appear:
cpu0			cpu1
msidata=0
			msidata=1
			insert cmd1
insert cmd0
			smmu execute cmd1
smmu execute cmd0
			poll timeout, because msidata=1 is overridden by
			cmd0, that means VAL=0, sync_idx=1.

This is not a functional problem, just make the caller wait for a long
time until TIMEOUT. It's rare to happen, because any other CMD_SYNCs
during the waiting period will break it.

Split the building of CMD_SYNC for SIG_IRQ mode into a single function,
to ensure better performance and good coding style.

Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com>
---
 drivers/iommu/arm-smmu-v3.c | 26 +++++++++++++++-----------
 1 file changed, 15 insertions(+), 11 deletions(-)

diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c
index 1d64710..ac6d6df 100644
--- a/drivers/iommu/arm-smmu-v3.c
+++ b/drivers/iommu/arm-smmu-v3.c
@@ -566,7 +566,7 @@ struct arm_smmu_device {

 	int				gerr_irq;
 	int				combined_irq;
-	atomic_t			sync_nr;
+	u32				sync_nr;

 	unsigned long			ias; /* IPA */
 	unsigned long			oas; /* PA */
@@ -775,6 +775,17 @@ static int queue_remove_raw(struct arm_smmu_queue *q, u64 *ent)
 	return 0;
 }

+static inline
+void arm_smmu_cmdq_build_sync_msi_cmd(u64 *cmd, struct arm_smmu_cmdq_ent *ent)
+{
+	cmd[0]  = FIELD_PREP(CMDQ_0_OP, ent->opcode);
+	cmd[0] |= FIELD_PREP(CMDQ_SYNC_0_CS, CMDQ_SYNC_0_CS_IRQ);
+	cmd[0] |= FIELD_PREP(CMDQ_SYNC_0_MSH, ARM_SMMU_SH_ISH);
+	cmd[0] |= FIELD_PREP(CMDQ_SYNC_0_MSIATTR, ARM_SMMU_MEMATTR_OIWB);
+	cmd[0] |= FIELD_PREP(CMDQ_SYNC_0_MSIDATA, ent->sync.msidata);
+	cmd[1]  = ent->sync.msiaddr & CMDQ_SYNC_1_MSIADDR_MASK;
+}
+
 /* High-level queue accessors */
 static int arm_smmu_cmdq_build_cmd(u64 *cmd, struct arm_smmu_cmdq_ent *ent)
 {
@@ -830,14 +841,9 @@ static int arm_smmu_cmdq_build_cmd(u64 *cmd, struct arm_smmu_cmdq_ent *ent)
 		cmd[1] |= FIELD_PREP(CMDQ_PRI_1_RESP, ent->pri.resp);
 		break;
 	case CMDQ_OP_CMD_SYNC:
-		if (ent->sync.msiaddr)
-			cmd[0] |= FIELD_PREP(CMDQ_SYNC_0_CS, CMDQ_SYNC_0_CS_IRQ);
-		else
-			cmd[0] |= FIELD_PREP(CMDQ_SYNC_0_CS, CMDQ_SYNC_0_CS_SEV);
+		cmd[0] |= FIELD_PREP(CMDQ_SYNC_0_CS, CMDQ_SYNC_0_CS_SEV);
 		cmd[0] |= FIELD_PREP(CMDQ_SYNC_0_MSH, ARM_SMMU_SH_ISH);
 		cmd[0] |= FIELD_PREP(CMDQ_SYNC_0_MSIATTR, ARM_SMMU_MEMATTR_OIWB);
-		cmd[0] |= FIELD_PREP(CMDQ_SYNC_0_MSIDATA, ent->sync.msidata);
-		cmd[1] |= ent->sync.msiaddr & CMDQ_SYNC_1_MSIADDR_MASK;
 		break;
 	default:
 		return -ENOENT;
@@ -947,14 +953,13 @@ static int __arm_smmu_cmdq_issue_sync_msi(struct arm_smmu_device *smmu)
 	struct arm_smmu_cmdq_ent ent = {
 		.opcode = CMDQ_OP_CMD_SYNC,
 		.sync	= {
-			.msidata = atomic_inc_return_relaxed(&smmu->sync_nr),
 			.msiaddr = virt_to_phys(&smmu->sync_count),
 		},
 	};

-	arm_smmu_cmdq_build_cmd(cmd, &ent);
-
 	spin_lock_irqsave(&smmu->cmdq.lock, flags);
+	ent.sync.msidata = ++smmu->sync_nr;
+	arm_smmu_cmdq_build_sync_msi_cmd(cmd, &ent);
 	arm_smmu_cmdq_insert_cmd(smmu, cmd);
 	spin_unlock_irqrestore(&smmu->cmdq.lock, flags);

@@ -2179,7 +2184,6 @@ static int arm_smmu_init_structures(struct arm_smmu_device *smmu)
 {
 	int ret;

-	atomic_set(&smmu->sync_nr, 0);
 	ret = arm_smmu_init_queues(smmu);
 	if (ret)
 		return ret;
--
1.8.3

[PATCH v4 2/2] iommu/arm-smmu-v3: avoid redundant CMD_SYNCs if possible

[Zhen Lei]

More than two CMD_SYNCs maybe adjacent in the command queue, and the first
one has done what others want to do. Drop the redundant CMD_SYNCs can
improve IO performance especially under the pressure scene.

I did the statistics in my test environment, the number of CMD_SYNCs can
be reduced about 1/3. See below:
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>
---
 drivers/iommu/arm-smmu-v3.c | 22 +++++++++++++++++++---
 1 file changed, 19 insertions(+), 3 deletions(-)

diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c
index ac6d6df..f3a56e1 100644
--- a/drivers/iommu/arm-smmu-v3.c
+++ b/drivers/iommu/arm-smmu-v3.c
@@ -567,6 +567,7 @@ 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 */
@@ -786,6 +787,11 @@ void arm_smmu_cmdq_build_sync_msi_cmd(u64 *cmd, struct arm_smmu_cmdq_ent *ent)
 	cmd[1]  = ent->sync.msiaddr & CMDQ_SYNC_1_MSIADDR_MASK;
 }

+static inline u8 arm_smmu_cmd_opcode_get(u64 *cmd)
+{
+	return cmd[0] & CMDQ_0_OP;
+}
+
 /* High-level queue accessors */
 static int arm_smmu_cmdq_build_cmd(u64 *cmd, struct arm_smmu_cmdq_ent *ent)
 {
@@ -906,6 +912,8 @@ 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 = arm_smmu_cmd_opcode_get(cmd);
+
 	while (queue_insert_raw(q, cmd) == -ENOSPC) {
 		if (queue_poll_cons(q, false, wfe))
 			dev_err_ratelimited(smmu->dev, "CMDQ timeout\n");
@@ -958,9 +966,17 @@ 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_sync_msi_cmd(cmd, &ent);
-	arm_smmu_cmdq_insert_cmd(smmu, cmd);
+	if (smmu->prev_cmd_opcode == CMDQ_OP_CMD_SYNC) {
+		/*
+		 * Previous command is CMD_SYNC also, there is no need to add
+		 * one more. Just poll it.
+		 */
+		ent.sync.msidata = smmu->sync_nr;
+	} else {
+		ent.sync.msidata = ++smmu->sync_nr;
+		arm_smmu_cmdq_build_sync_msi_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);
--
1.8.3

Re: [PATCH v4 2/2] iommu/arm-smmu-v3: avoid redundant CMD_SYNCs if possible

[John Garry]

On 19/08/2018 08:51, Zhen Lei wrote:
> More than two CMD_SYNCs maybe adjacent in the command queue, and the first
> one has done what others want to do. Drop the redundant CMD_SYNCs can
> improve IO performance especially under the pressure scene.
>
> I did the statistics in my test environment, the number of CMD_SYNCs can
> be reduced about 1/3. See below:
> 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>
> ---
>  drivers/iommu/arm-smmu-v3.c | 22 +++++++++++++++++++---
>  1 file changed, 19 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/iommu/arm-smmu-v3.c b/drivers/iommu/arm-smmu-v3.c
> index ac6d6df..f3a56e1 100644
> --- a/drivers/iommu/arm-smmu-v3.c
> +++ b/drivers/iommu/arm-smmu-v3.c
> @@ -567,6 +567,7 @@ 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 */
> @@ -786,6 +787,11 @@ void arm_smmu_cmdq_build_sync_msi_cmd(u64 *cmd, struct arm_smmu_cmdq_ent *ent)
>  	cmd[1]  = ent->sync.msiaddr & CMDQ_SYNC_1_MSIADDR_MASK;
>  }
>
> +static inline u8 arm_smmu_cmd_opcode_get(u64 *cmd)
> +{
> +	return cmd[0] & CMDQ_0_OP;
> +}
> +
>  /* High-level queue accessors */
>  static int arm_smmu_cmdq_build_cmd(u64 *cmd, struct arm_smmu_cmdq_ent *ent)
>  {
> @@ -906,6 +912,8 @@ 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 = arm_smmu_cmd_opcode_get(cmd);
> +
>  	while (queue_insert_raw(q, cmd) == -ENOSPC) {
>  		if (queue_poll_cons(q, false, wfe))
>  			dev_err_ratelimited(smmu->dev, "CMDQ timeout\n");
> @@ -958,9 +966,17 @@ 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_sync_msi_cmd(cmd, &ent);
> -	arm_smmu_cmdq_insert_cmd(smmu, cmd);
> +	if (smmu->prev_cmd_opcode == CMDQ_OP_CMD_SYNC) {
> +		/*
> +		 * Previous command is CMD_SYNC also, there is no need to add
> +		 * one more. Just poll it.
> +		 */
> +		ent.sync.msidata = smmu->sync_nr;
> +	} else {
> +		ent.sync.msidata = ++smmu->sync_nr;
> +		arm_smmu_cmdq_build_sync_msi_cmd(cmd, &ent);
> +		arm_smmu_cmdq_insert_cmd(smmu, cmd);
> +	}
>  	spin_unlock_irqrestore(&smmu->cmdq.lock, flags);

I find something like this adds support for combining CMD_SYNC commands 
for regular polling mode:

@@ -569,6 +569,7 @@ struct arm_smmu_device {
         int                             combined_irq;
         u32                             sync_nr;
         u8                              prev_cmd_opcode;
+       int                             prev_cmd_sync_res;

         unsigned long                   ias; /* IPA */
         unsigned long                   oas; /* PA */
@@ -985,17 +986,33 @@ static int __arm_smmu_cmdq_issue_sync_msi(struct 
arm_smmu_device *smmu)

  static int __arm_smmu_cmdq_issue_sync(struct arm_smmu_device *smmu)
  {
-       u64 cmd[CMDQ_ENT_DWORDS];
+       static u64 cmd[CMDQ_ENT_DWORDS] = {
+               _FIELD_PREP(CMDQ_0_OP, CMDQ_OP_CMD_SYNC) |
+               _FIELD_PREP(CMDQ_SYNC_0_CS, CMDQ_SYNC_0_CS_SEV) |
+               _FIELD_PREP(CMDQ_SYNC_0_MSH, ARM_SMMU_SH_ISH) |
+               _FIELD_PREP(CMDQ_SYNC_0_MSIATTR, ARM_SMMU_MEMATTR_OIWB)
+       };
         unsigned long flags;
         bool wfe = !!(smmu->features & ARM_SMMU_FEAT_SEV);
-       struct arm_smmu_cmdq_ent ent = { .opcode = CMDQ_OP_CMD_SYNC };
-       int ret;
+       int ret = 0;

-       arm_smmu_cmdq_build_cmd(cmd, &ent);

         spin_lock_irqsave(&smmu->cmdq.lock, flags);
-       arm_smmu_cmdq_insert_cmd(smmu, cmd);
-       ret = queue_poll_cons(&smmu->cmdq.q, true, wfe);
+       if (smmu->prev_cmd_opcode != CMDQ_OP_CMD_SYNC ||
+               smmu->prev_cmd_sync_res != 0) {
+               arm_smmu_cmdq_insert_cmd(smmu, cmd);
+               smmu->prev_cmd_sync_res = ret =
+                       queue_poll_cons(&smmu->cmdq.q, true, wfe);
+       }

I tested iperf on a 1G network link and was seeing 6-10% CMD_SYNC 
commands combined. I would really need to test this on a faster 
connection to see any throughout difference.

 From the above figures, I think leizhen was seeing 25% combine rate, right?

As for this code, it could be neatened...

Cheers,
John

>
>  	return __arm_smmu_sync_poll_msi(smmu, ent.sync.msidata);
> --
> 1.8.3
>
>
>
> .
>

Re: [PATCH v4 2/2] iommu/arm-smmu-v3: avoid redundant CMD_SYNCs if possible

[Leizhen (ThunderTown)]

On 2018/8/30 19:18, John Garry wrote:
> On 19/08/2018 08:51, Zhen Lei wrote:
>>      spin_unlock_irqrestore(&smmu->cmdq.lock, flags);
> 
> I find something like this adds support for combining CMD_SYNC commands for regular polling mode:
> 
> @@ -569,6 +569,7 @@ struct arm_smmu_device {
>         int                             combined_irq;
>         u32                             sync_nr;
>         u8                              prev_cmd_opcode;
> +       int                             prev_cmd_sync_res;
> 
>         unsigned long                   ias; /* IPA */
>         unsigned long                   oas; /* PA */
> @@ -985,17 +986,33 @@ static int __arm_smmu_cmdq_issue_sync_msi(struct arm_smmu_device *smmu)
> 
>  static int __arm_smmu_cmdq_issue_sync(struct arm_smmu_device *smmu)
>  {
> -       u64 cmd[CMDQ_ENT_DWORDS];
> +       static u64 cmd[CMDQ_ENT_DWORDS] = {
> +               _FIELD_PREP(CMDQ_0_OP, CMDQ_OP_CMD_SYNC) |
> +               _FIELD_PREP(CMDQ_SYNC_0_CS, CMDQ_SYNC_0_CS_SEV) |
> +               _FIELD_PREP(CMDQ_SYNC_0_MSH, ARM_SMMU_SH_ISH) |
> +               _FIELD_PREP(CMDQ_SYNC_0_MSIATTR, ARM_SMMU_MEMATTR_OIWB)
> +       };
>         unsigned long flags;
>         bool wfe = !!(smmu->features & ARM_SMMU_FEAT_SEV);
> -       struct arm_smmu_cmdq_ent ent = { .opcode = CMDQ_OP_CMD_SYNC };
> -       int ret;
> +       int ret = 0;
> 
> -       arm_smmu_cmdq_build_cmd(cmd, &ent);
> 
>         spin_lock_irqsave(&smmu->cmdq.lock, flags);
> -       arm_smmu_cmdq_insert_cmd(smmu, cmd);
> -       ret = queue_poll_cons(&smmu->cmdq.q, true, wfe);
> +       if (smmu->prev_cmd_opcode != CMDQ_OP_CMD_SYNC ||
> +               smmu->prev_cmd_sync_res != 0) {
> +               arm_smmu_cmdq_insert_cmd(smmu, cmd);
> +               smmu->prev_cmd_sync_res = ret =
> +                       queue_poll_cons(&smmu->cmdq.q, true, wfe);
> +       }
> 
> I tested iperf on a 1G network link and was seeing 6-10% CMD_SYNC commands combined. I would really need to test this on a faster connection to see any throughout difference.
> 
> From the above figures, I think leizhen was seeing 25% combine rate, right?

Yes. In my test case, the size of unmap are almost one page, that means 1 TLBI follows 1 SYNC,
so the probability that two CMD_SYNCs next to each other will be greater.

> 
> As for this code, it could be neatened...
> 
> Cheers,
> John
> 
>>
>>      return __arm_smmu_sync_poll_msi(smmu, ent.sync.msidata);
>> -- 
>> 1.8.3
>>
>>
>>
>> .
>>
> 
> 
> 
> .
> 

-- 
Thanks!
BestRegards