Skip to contentSkip to conversation

raw/ntb: fix write memory barrier issue

PATCH Superseded Xiaoyun Li 4 participants 13 messages Dec 4, 2019 – Jan 20, 2020

lore
Share & tools
lore

Xiaoyun Li

author
#0 lore raw

PATCH 1/1 raw/ntb: fix write memory barrier issue

All buffers and ring info should be written before tail register update. This patch relocates the write memory barrier before updating tail register to avoid potential issues.

Fixes 11b5c7daf019 ("raw/ntb: add enqueue and dequeue functions")
---
 drivers/raw/ntb/ntb.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/raw/ntb/ntb.c +2 -2
index ad7f6abfd..dd0b72f8c 100644
@@ -683,8 +683,8 @@ ntb_enqueue_bufs(struct rte_rawdev *dev,
 			   sizeof(struct ntb_used) * nb1);
 		rte_memcpy(txq->tx_used_ring, tx_used + nb1,
 			   sizeof(struct ntb_used) * nb2);
-		*txq->used_cnt = txq->last_used;
 		rte_wmb();
+		*txq->used_cnt = txq->last_used;
 
 		/* update queue stats */
 		hw->ntb_xstats[NTB_TX_BYTES_ID + off] += bytes;
@@ -789,8 +789,8 @@ ntb_dequeue_bufs(struct rte_rawdev *dev,
 			   sizeof(struct ntb_desc) * nb1);
 		rte_memcpy(rxq->rx_desc_ring, rx_desc + nb1,
 			   sizeof(struct ntb_desc) * nb2);
-		*rxq->avail_cnt = rxq->last_avail;
 		rte_wmb();
+		*rxq->avail_cnt = rxq->last_avail;
 
 		/* update queue stats */
 		off = NTB_XSTATS_NUM * ((size_t)context + 1);
Signature · 1 line
2.17.1
#1 lore raw

Hi Xiaoyun,

Quoted 24 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Xiaoyun Li
Sent: Wednesday, December 4, 2019 11:19 PM
To: [email protected]
Cc: [email protected]; Xiaoyun Li <[email protected]>; [email protected]
Subject: [dpdk-dev] [PATCH] raw/ntb: fix write memory barrier issue

All buffers and ring info should be written before tail register update. This patch relocates the write memory barrier before updating tail register to avoid potential issues.

Fixes 11b5c7daf019 ("raw/ntb: add enqueue and dequeue functions")
---
 drivers/raw/ntb/ntb.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/raw/ntb/ntb.c +0 -0
index ad7f6abfd..dd0b72f8c 100644

drivers/raw/ntb/ntb.c ntb_enqueue_bufs(struct rte_rawdev *dev, · line 687

Quoted 6 lines
drivers/raw/ntb/ntb.c +1 -1
@@ -683,8 +683,8 @@ ntb_enqueue_bufs(struct rte_rawdev *dev,
 			   sizeof(struct ntb_used) * nb1);
 		rte_memcpy(txq->tx_used_ring, tx_used + nb1,
 			   sizeof(struct ntb_used) * nb2);
-		*txq->used_cnt = txq->last_used;
 		rte_wmb();
+		*txq->used_cnt = txq->last_used;

I am ok with the re-location of the barrier, but why not the rte_io_wmb instead of rte_wmb? Rte_io_wmb is sufficient to guarantee the preceding stores are visible to the device, rte_wmb is overkill.
https://code.dpdk.org/dpdk/latest/source/lib/librte_eal/common/include/generic/rte_atomic.h#L92

Quoted 15 lines
drivers/raw/ntb/ntb.c +1 -2
 		/* update queue stats */
 		hw->ntb_xstats[NTB_TX_BYTES_ID + off] += bytes;
@@ -789,8 +789,8 @@ ntb_dequeue_bufs(struct rte_rawdev *dev,
 			   sizeof(struct ntb_desc) * nb1);
 		rte_memcpy(rxq->rx_desc_ring, rx_desc + nb1,
 			   sizeof(struct ntb_desc) * nb2);
-		*rxq->avail_cnt = rxq->last_avail;
 		rte_wmb();
+		*rxq->avail_cnt = rxq->last_avail;
 
 		/* update queue stats */
 		off = NTB_XSTATS_NUM * ((size_t)context + 1);
--

2.17.1

#2 lore raw

Didn't notice that. Will fix it in v2. Thanks!

Quoted 60 lines

-----Original Message-----
From: Gavin Hu (Arm Technology China) [mailto:[email protected]]
Sent: Saturday, December 14, 2019 23:30
To: Li, Xiaoyun <[email protected]>; Wu, Jingjing <[email protected]>
Cc: [email protected]; [email protected]; nd <[email protected]>
Subject: RE: [dpdk-dev] [PATCH] raw/ntb: fix write memory barrier issue

Hi Xiaoyun,

Quoted 24 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Xiaoyun Li
Sent: Wednesday, December 4, 2019 11:19 PM
To: [email protected]
Cc: [email protected]; Xiaoyun Li <[email protected]>; [email protected]
Subject: [dpdk-dev] [PATCH] raw/ntb: fix write memory barrier issue

All buffers and ring info should be written before tail register update. This patch relocates the write memory barrier before updating tail register to avoid potential issues.

Fixes 11b5c7daf019 ("raw/ntb: add enqueue and dequeue functions")
---
 drivers/raw/ntb/ntb.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/raw/ntb/ntb.c +0 -0
index ad7f6abfd..dd0b72f8c 100644

drivers/raw/ntb/ntb.c ntb_enqueue_bufs(struct rte_rawdev *dev, · line 687

Quoted 6 lines
drivers/raw/ntb/ntb.c +1 -1
@@ -683,8 +683,8 @@ ntb_enqueue_bufs(struct rte_rawdev *dev,
 			   sizeof(struct ntb_used) * nb1);
 		rte_memcpy(txq->tx_used_ring, tx_used + nb1,
 			   sizeof(struct ntb_used) * nb2);
-		*txq->used_cnt = txq->last_used;
 		rte_wmb();
+		*txq->used_cnt = txq->last_used;

I am ok with the re-location of the barrier, but why not the rte_io_wmb instead of rte_wmb?
Rte_io_wmb is sufficient to guarantee the preceding stores are visible to the device, rte_wmb is overkill.
https://code.dpdk.org/dpdk/latest/source/lib/librte_eal/common/include/gener
ic/rte_atomic.h#L92

Quoted 15 lines
drivers/raw/ntb/ntb.c +1 -2
 		/* update queue stats */
 		hw->ntb_xstats[NTB_TX_BYTES_ID + off] += bytes;
@@ -789,8 +789,8 @@ ntb_dequeue_bufs(struct rte_rawdev *dev,
 			   sizeof(struct ntb_desc) * nb1);
 		rte_memcpy(rxq->rx_desc_ring, rx_desc + nb1,
 			   sizeof(struct ntb_desc) * nb2);
-		*rxq->avail_cnt = rxq->last_avail;
 		rte_wmb();
+		*rxq->avail_cnt = rxq->last_avail;
 
 		/* update queue stats */
 		off = NTB_XSTATS_NUM * ((size_t)context + 1);
--

2.17.1

author
#4 lore raw

PATCH v2 1/1 raw/ntb: fix write memory barrier issue

In reply to Xiaoyun Li

All buffers and ring info should be written before tail register update. This patch relocates the write memory barrier before updating tail register to avoid potential issues.

Fixes 11b5c7daf019 ("raw/ntb: add enqueue and dequeue functions")

---
v2:
* Replaced rte_wmb with rte_io_wmb since rte_io_wmb is enough.

---
 drivers/raw/ntb/ntb.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/raw/ntb/ntb.c +2 -2
index ad7f6abfd..c7de86f36 100644
@@ -683,8 +683,8 @@ ntb_enqueue_bufs(struct rte_rawdev *dev,
 			   sizeof(struct ntb_used) * nb1);
 		rte_memcpy(txq->tx_used_ring, tx_used + nb1,
 			   sizeof(struct ntb_used) * nb2);
+		rte_io_wmb();
 		*txq->used_cnt = txq->last_used;
-		rte_wmb();
 
 		/* update queue stats */
 		hw->ntb_xstats[NTB_TX_BYTES_ID + off] += bytes;
@@ -789,8 +789,8 @@ ntb_dequeue_bufs(struct rte_rawdev *dev,
 			   sizeof(struct ntb_desc) * nb1);
 		rte_memcpy(rxq->rx_desc_ring, rx_desc + nb1,
 			   sizeof(struct ntb_desc) * nb2);
+		rte_io_wmb();
 		*rxq->avail_cnt = rxq->last_avail;
-		rte_wmb();
 
 		/* update queue stats */
 		off = NTB_XSTATS_NUM * ((size_t)context + 1);
Signature · 1 line
2.17.1
Reviewed-by on v2 1/1
#5 lore raw
Quoted 30 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Xiaoyun Li
Sent: Monday, December 16, 2019 9:59 AM
To: [email protected]
Cc: [email protected]; [email protected]; Xiaoyun Li <[email protected]>; [email protected]
Subject: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier issue

All buffers and ring info should be written before tail register update. This patch relocates the write memory barrier before updating tail register to avoid potential issues.

Fixes 11b5c7daf019 ("raw/ntb: add enqueue and dequeue functions")

---
v2:
* Replaced rte_wmb with rte_io_wmb since rte_io_wmb is enough.

---
 drivers/raw/ntb/ntb.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/raw/ntb/ntb.c +0 -0
index ad7f6abfd..c7de86f36 100644

drivers/raw/ntb/ntb.c ntb_enqueue_bufs(struct rte_rawdev *dev, · line 686

Quoted 4 lines
drivers/raw/ntb/ntb.c +1 -0
@@ -683,8 +683,8 @@ ntb_enqueue_bufs(struct rte_rawdev *dev,
 			   sizeof(struct ntb_used) * nb1);
 		rte_memcpy(txq->tx_used_ring, tx_used + nb1,
 			   sizeof(struct ntb_used) * nb2);
+		rte_io_wmb();

As both txq->tx_used_ring and *txq->used_cnt are physically reside in the PCI device side, rte_io_wmb is correct to ensure the ordering.

Quoted 16 lines
drivers/raw/ntb/ntb.c +1 -3
 		*txq->used_cnt = txq->last_used;
-		rte_wmb();
 
 		/* update queue stats */
 		hw->ntb_xstats[NTB_TX_BYTES_ID + off] += bytes;
@@ -789,8 +789,8 @@ ntb_dequeue_bufs(struct rte_rawdev *dev,
 			   sizeof(struct ntb_desc) * nb1);
 		rte_memcpy(rxq->rx_desc_ring, rx_desc + nb1,
 			   sizeof(struct ntb_desc) * nb2);
+		rte_io_wmb();
 		*rxq->avail_cnt = rxq->last_avail;
-		rte_wmb();
 
 		/* update queue stats */
 		off = NTB_XSTATS_NUM * ((size_t)context + 1);
--
Quoted 1 line

2.17.1

#6 lore raw

Hi
I reconsidered and retested about this issue.
I still need to use rte_wmb instead of using rte_io_wmb.

Because to achieve high performance, ntb needs to turn on WC(write combining) feature. The perf difference with and without WC enabled is more than 20X. And when WC enabled, rte_io_wmb cannot make sure the instructions are in order only rte_wmb can make sure that.

And in my retest, when sending 64 bytes packets, using rte_io_wmb will cause out-of-order issue and cause memory corruption on rx side. And using rte_wmb is fine.

So I can only use v1 patch and suspend v2 patch in patchwork.

Best Regards
Xiaoyun Li

Quoted 65 lines

-----Original Message-----
From: Gavin Hu (Arm Technology China) [mailto:[email protected]]
Sent: Monday, December 16, 2019 18:50
To: Li, Xiaoyun <[email protected]>; Wu, Jingjing <[email protected]>
Cc: [email protected]; Maslekar, Omkar <[email protected]>; [email protected]; nd <[email protected]>
Subject: RE: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier issue

Quoted 28 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Xiaoyun Li
Sent: Monday, December 16, 2019 9:59 AM
To: [email protected]
Cc: [email protected]; [email protected]; Xiaoyun Li <[email protected]>; [email protected]
Subject: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier issue

All buffers and ring info should be written before tail register update. This patch relocates the write memory barrier before updating tail register to avoid potential issues.

Fixes 11b5c7daf019 ("raw/ntb: add enqueue and dequeue functions")

---
v2:
* Replaced rte_wmb with rte_io_wmb since rte_io_wmb is enough.

---
 drivers/raw/ntb/ntb.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/raw/ntb/ntb.c index renamed from drivers/raw/ntb/ntb.c +0 -0

ad7f6abfd..c7de86f36 100644

drivers/raw/ntb/ntb.c ntb_enqueue_bufs(struct rte_rawdev *dev, · line 686

Quoted 4 lines
drivers/raw/ntb/ntb.c +1 -0
@@ -683,8 +683,8 @@ ntb_enqueue_bufs(struct rte_rawdev *dev,
 			   sizeof(struct ntb_used) * nb1);
 		rte_memcpy(txq->tx_used_ring, tx_used + nb1,
 			   sizeof(struct ntb_used) * nb2);
+		rte_io_wmb();

As both txq->tx_used_ring and *txq->used_cnt are physically reside in the PCI device side, rte_io_wmb is correct to ensure the ordering.

drivers/raw/ntb/ntb.c ntb_enqueue_bufs(struct rte_rawdev *dev,

Quoted 5 lines
drivers/raw/ntb/ntb.c +0 -1
 		*txq->used_cnt = txq->last_used;
-		rte_wmb();
 
 		/* update queue stats */
 		hw->ntb_xstats[NTB_TX_BYTES_ID + off] += bytes; @@ -789,8

+789,8 @@

Quoted 11 lines

ntb_dequeue_bufs(struct rte_rawdev *dev,

drivers/raw/ntb/ntb.c +1 -2
 			   sizeof(struct ntb_desc) * nb1);
 		rte_memcpy(rxq->rx_desc_ring, rx_desc + nb1,
 			   sizeof(struct ntb_desc) * nb2);
+		rte_io_wmb();
 		*rxq->avail_cnt = rxq->last_avail;
-		rte_wmb();
 
 		/* update queue stats */
 		off = NTB_XSTATS_NUM * ((size_t)context + 1);
--
Quoted 1 line

2.17.1

#7 lore raw

Hi Xiaoyun,

Quoted 18 lines

-----Original Message-----
From: Li, Xiaoyun <[email protected]>
Sent: Monday, December 23, 2019 3:52 PM
To: Gavin Hu <[email protected]>; Wu, Jingjing <[email protected]>
Cc: [email protected]; Maslekar, Omkar <[email protected]>; [email protected]; nd <[email protected]>
Subject: RE: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier issue

Hi
I reconsidered and retested about this issue.
I still need to use rte_wmb instead of using rte_io_wmb.

Because to achieve high performance, ntb needs to turn on WC(write combining) feature. The perf difference with and without WC enabled is more than 20X.
And when WC enabled, rte_io_wmb cannot make sure the instructions are in order only rte_wmb can make sure that.

Quoted 3 lines

And in my retest, when sending 64 bytes packets, using rte_io_wmb will cause out-of-order issue and cause memory corruption on rx side. And using rte_wmb is fine.

That's true, as it is declared as 'write combine' region, even x86 is known as strong ordered, it is the interconnect or PCI RC may do the reordering, 'write combine', 'write coalescing', which caused this problem. IMO, rte_io_*mb barriers on x86 should be promoted to stronger is WC is involved(but that will sap performance for non-WC memories?).
https://code.dpdk.org/dpdk/latest/source/lib/librte_eal/common/include/arch/x86/rte_atomic.h#L78

Using rte_wmb will hurt performance for aarch64 also, as pci device memory accesses to a single device are strongly ordered therefore the strongest rte_wmb is not necessary.

Quoted 74 lines

So I can only use v1 patch and suspend v2 patch in patchwork.

Best Regards
Xiaoyun Li

Quoted 4 lines

-----Original Message-----
From: Gavin Hu (Arm Technology China) [mailto:[email protected]]
Sent: Monday, December 16, 2019 18:50
To: Li, Xiaoyun <[email protected]>; Wu, Jingjing

<[email protected]>

Quoted 3 lines

Cc: [email protected]; Maslekar, Omkar <[email protected]>; [email protected]; nd <[email protected]>
Subject: RE: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier

issue

Quoted 26 lines
Quoted 26 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Xiaoyun Li
Sent: Monday, December 16, 2019 9:59 AM
To: [email protected]
Cc: [email protected]; [email protected]; Xiaoyun Li <[email protected]>; [email protected]
Subject: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier issue

All buffers and ring info should be written before tail register update. This patch relocates the write memory barrier before updating tail register to avoid potential issues.

Fixes 11b5c7daf019 ("raw/ntb: add enqueue and dequeue functions")

---
v2:
* Replaced rte_wmb with rte_io_wmb since rte_io_wmb is enough.

---
 drivers/raw/ntb/ntb.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
Quoted 10 lines

drivers/raw/ntb/ntb.c ntb_enqueue_bufs(struct rte_rawdev *dev, · line 686

Quoted 9 lines
drivers/raw/ntb/ntb.c index renamed from drivers/raw/ntb/ntb.c +0 -0

ad7f6abfd..c7de86f36 100644

drivers/raw/ntb/ntb.c +1 -0
@@ -683,8 +683,8 @@ ntb_enqueue_bufs(struct rte_rawdev *dev,
 			   sizeof(struct ntb_used) * nb1);
 		rte_memcpy(txq->tx_used_ring, tx_used + nb1,
 			   sizeof(struct ntb_used) * nb2);
+		rte_io_wmb();

As both txq->tx_used_ring and *txq->used_cnt are physically reside in the

PCI

Quoted 7 lines

device side, rte_io_wmb is correct to ensure the ordering.

drivers/raw/ntb/ntb.c ntb_enqueue_bufs(struct rte_rawdev *dev,

Quoted 5 lines
drivers/raw/ntb/ntb.c +0 -1
 		*txq->used_cnt = txq->last_used;
-		rte_wmb();
 
 		/* update queue stats */
 		hw->ntb_xstats[NTB_TX_BYTES_ID + off] += bytes; @@ -

789,8

Quoted 15 lines

+789,8 @@

Quoted 11 lines

ntb_dequeue_bufs(struct rte_rawdev *dev,

drivers/raw/ntb/ntb.c +1 -2
 			   sizeof(struct ntb_desc) * nb1);
 		rte_memcpy(rxq->rx_desc_ring, rx_desc + nb1,
 			   sizeof(struct ntb_desc) * nb2);
+		rte_io_wmb();
 		*rxq->avail_cnt = rxq->last_avail;
-		rte_wmb();
 
 		/* update queue stats */
 		off = NTB_XSTATS_NUM * ((size_t)context + 1);
--
Quoted 1 line

2.17.1

#8 lore raw

Hi

Still, stability and correctness are much more important than performance.
As I said, with WC can benefit more than 20X perf. Comparing to this benefit, the difference between rte_wmb and rte_io_wmb is not that important. And in my test, the performance is not that bad with rte_wmb especially with large packets which are the normal use cases.

BTW, I've searched linux kernel codes and don't see any NTB device on arm platform. So I don't think you need to consider the perf hurt to arm.

Best Regards
Xiaoyun Li

Quoted 121 lines

-----Original Message-----
From: Gavin Hu [mailto:[email protected]]
Sent: Monday, December 23, 2019 16:38
To: Li, Xiaoyun <[email protected]>; Wu, Jingjing <[email protected]>
Cc: [email protected]; Maslekar, Omkar <[email protected]>; [email protected]; nd <[email protected]>; [email protected]; Honnappa Nagarahalli <[email protected]>; Richardson, Bruce <[email protected]>; nd <[email protected]>
Subject: RE: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier issue

Hi Xiaoyun,

Quoted 19 lines

-----Original Message-----
From: Li, Xiaoyun <[email protected]>
Sent: Monday, December 23, 2019 3:52 PM
To: Gavin Hu <[email protected]>; Wu, Jingjing <[email protected]>
Cc: [email protected]; Maslekar, Omkar <[email protected]>; [email protected]; nd <[email protected]>
Subject: RE: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier issue

Hi
I reconsidered and retested about this issue.
I still need to use rte_wmb instead of using rte_io_wmb.

Because to achieve high performance, ntb needs to turn on WC(write combining) feature. The perf difference with and without WC enabled is more than 20X.
And when WC enabled, rte_io_wmb cannot make sure the instructions are in order only rte_wmb can make sure that.

Quoted 3 lines

And in my retest, when sending 64 bytes packets, using rte_io_wmb will cause out-of-order issue and cause memory corruption on rx side. And using rte_wmb is fine.

That's true, as it is declared as 'write combine' region, even x86 is known as strong ordered, it is the interconnect or PCI RC may do the reordering, 'write combine', 'write coalescing', which caused this problem.
IMO, rte_io_*mb barriers on x86 should be promoted to stronger is WC is involved(but that will sap performance for non-WC memories?).
https://code.dpdk.org/dpdk/latest/source/lib/librte_eal/common/include/arch/
x86/rte_atomic.h#L78

Using rte_wmb will hurt performance for aarch64 also, as pci device memory accesses to a single device are strongly ordered therefore the strongest rte_wmb is not necessary.

Quoted 76 lines

So I can only use v1 patch and suspend v2 patch in patchwork.

Best Regards
Xiaoyun Li

Quoted 4 lines

-----Original Message-----
From: Gavin Hu (Arm Technology China) [mailto:[email protected]]
Sent: Monday, December 16, 2019 18:50
To: Li, Xiaoyun <[email protected]>; Wu, Jingjing

<[email protected]>

Quoted 3 lines

Cc: [email protected]; Maslekar, Omkar <[email protected]>; [email protected]; nd <[email protected]>
Subject: RE: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier

issue

Quoted 27 lines
Quoted 27 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Xiaoyun Li
Sent: Monday, December 16, 2019 9:59 AM
To: [email protected]
Cc: [email protected]; [email protected]; Xiaoyun Li <[email protected]>; [email protected]
Subject: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier issue

All buffers and ring info should be written before tail register update. This patch relocates the write memory barrier before updating tail register to avoid potential issues.

Fixes 11b5c7daf019 ("raw/ntb: add enqueue and dequeue functions")

---
v2:
* Replaced rte_wmb with rte_io_wmb since rte_io_wmb is enough.

---
 drivers/raw/ntb/ntb.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
Quoted 11 lines

drivers/raw/ntb/ntb.c ntb_enqueue_bufs(struct rte_rawdev *dev, · line 686

Quoted 9 lines
drivers/raw/ntb/ntb.c index renamed from drivers/raw/ntb/ntb.c +0 -0

ad7f6abfd..c7de86f36 100644

drivers/raw/ntb/ntb.c +1 -0
@@ -683,8 +683,8 @@ ntb_enqueue_bufs(struct rte_rawdev *dev,
 			   sizeof(struct ntb_used) * nb1);
 		rte_memcpy(txq->tx_used_ring, tx_used + nb1,
 			   sizeof(struct ntb_used) * nb2);
+		rte_io_wmb();

As both txq->tx_used_ring and *txq->used_cnt are physically reside in the

PCI

Quoted 7 lines

device side, rte_io_wmb is correct to ensure the ordering.

drivers/raw/ntb/ntb.c ntb_enqueue_bufs(struct rte_rawdev *dev,

Quoted 5 lines
drivers/raw/ntb/ntb.c +0 -1
 		*txq->used_cnt = txq->last_used;
-		rte_wmb();
 
 		/* update queue stats */
 		hw->ntb_xstats[NTB_TX_BYTES_ID + off] += bytes; @@ -

789,8

Quoted 15 lines

+789,8 @@

Quoted 11 lines

ntb_dequeue_bufs(struct rte_rawdev *dev,

drivers/raw/ntb/ntb.c +1 -2
 			   sizeof(struct ntb_desc) * nb1);
 		rte_memcpy(rxq->rx_desc_ring, rx_desc + nb1,
 			   sizeof(struct ntb_desc) * nb2);
+		rte_io_wmb();
 		*rxq->avail_cnt = rxq->last_avail;
-		rte_wmb();
 
 		/* update queue stats */
 		off = NTB_XSTATS_NUM * ((size_t)context + 1);
--
Quoted 1 line

2.17.1

#9 lore raw

Hi Xiaoyun,

Quoted 10 lines

-----Original Message-----
From: Li, Xiaoyun <[email protected]>
Sent: Monday, December 23, 2019 5:36 PM
To: Gavin Hu <[email protected]>; Wu, Jingjing <[email protected]>
Cc: [email protected]; Maslekar, Omkar <[email protected]>; [email protected]; nd <[email protected]>; [email protected]; Honnappa Nagarahalli <[email protected]>; Richardson, Bruce <[email protected]>; nd <[email protected]>
Subject: RE: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier issue

Quoted 7 lines

Hi

Still, stability and correctness are much more important than performance. As I said, with WC can benefit more than 20X perf. Comparing to this benefit, the difference between rte_wmb and rte_io_wmb is not that important. And in my test, the performance is not that bad with rte_wmb especially with large packets which are the normal use cases.

I agree 'sfence' is the correct barrier for WC, as it is a weak memory model. Rte_io on x86 is a compiler barrier, that is not strong enough.

Quoted 4 lines

BTW, I've searched linux kernel codes and don't see any NTB device on arm platform.
So I don't think you need to consider the perf hurt to arm.

Limited the discussion to NTB, I am fine, and since WC in not used for any other NICs, that's ok.

Quoted 130 lines

Best Regards
Xiaoyun Li

Quoted 4 lines

-----Original Message-----
From: Gavin Hu [mailto:[email protected]]
Sent: Monday, December 23, 2019 16:38
To: Li, Xiaoyun <[email protected]>; Wu, Jingjing

<[email protected]>

Quoted 6 lines

Cc: [email protected]; Maslekar, Omkar <[email protected]>; [email protected]; nd <[email protected]>; [email protected]; Honnappa Nagarahalli <[email protected]>; Richardson, Bruce <[email protected]>; nd <[email protected]>
Subject: RE: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier issue

Quoted 19 lines

Hi Xiaoyun,

Quoted 17 lines

-----Original Message-----
From: Li, Xiaoyun <[email protected]>
Sent: Monday, December 23, 2019 3:52 PM
To: Gavin Hu <[email protected]>; Wu, Jingjing <[email protected]>
Cc: [email protected]; Maslekar, Omkar <[email protected]>; [email protected]; nd <[email protected]>
Subject: RE: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier issue

Hi
I reconsidered and retested about this issue.
I still need to use rte_wmb instead of using rte_io_wmb.

Because to achieve high performance, ntb needs to turn on WC(write combining) feature. The perf difference with and without WC enabled is more than 20X.
And when WC enabled, rte_io_wmb cannot make sure the instructions

are

Quoted 11 lines
Quoted 5 lines

in order only rte_wmb can make sure that.

And in my retest, when sending 64 bytes packets, using rte_io_wmb will cause out-of-order issue and cause memory corruption on rx side. And using rte_wmb is fine.

That's true, as it is declared as 'write combine' region, even x86 is known as strong ordered, it is the interconnect or PCI RC may do the reordering, 'write combine', 'write coalescing', which caused this problem.
IMO, rte_io_*mb barriers on x86 should be promoted to stronger is WC is involved(but that will sap performance for non-WC memories?).

https://code.dpdk.org/dpdk/latest/source/lib/librte_eal/common/include/ar
ch/

Quoted 3 lines

x86/rte_atomic.h#L78

Using rte_wmb will hurt performance for aarch64 also, as pci device

memory

Quoted 78 lines

accesses to a single device are strongly ordered therefore the strongest rte_wmb is not necessary.

Quoted 76 lines

So I can only use v1 patch and suspend v2 patch in patchwork.

Best Regards
Xiaoyun Li

Quoted 4 lines

-----Original Message-----
From: Gavin Hu (Arm Technology China) [mailto:[email protected]]
Sent: Monday, December 16, 2019 18:50
To: Li, Xiaoyun <[email protected]>; Wu, Jingjing

<[email protected]>

Quoted 3 lines

Cc: [email protected]; Maslekar, Omkar <[email protected]>; [email protected]; nd <[email protected]>
Subject: RE: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier

issue

Quoted 27 lines
Quoted 27 lines

-----Original Message-----
From: dev <[email protected]> On Behalf Of Xiaoyun Li
Sent: Monday, December 16, 2019 9:59 AM
To: [email protected]
Cc: [email protected]; [email protected]; Xiaoyun Li <[email protected]>; [email protected]
Subject: [dpdk-dev] [PATCH v2] raw/ntb: fix write memory barrier issue

All buffers and ring info should be written before tail register update. This patch relocates the write memory barrier before updating tail register to avoid potential issues.

Fixes 11b5c7daf019 ("raw/ntb: add enqueue and dequeue functions")

---
v2:
* Replaced rte_wmb with rte_io_wmb since rte_io_wmb is enough.

---
 drivers/raw/ntb/ntb.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
Quoted 11 lines

drivers/raw/ntb/ntb.c ntb_enqueue_bufs(struct rte_rawdev *dev, · line 686

Quoted 9 lines
drivers/raw/ntb/ntb.c index renamed from drivers/raw/ntb/ntb.c +0 -0

ad7f6abfd..c7de86f36 100644

drivers/raw/ntb/ntb.c +1 -0
@@ -683,8 +683,8 @@ ntb_enqueue_bufs(struct rte_rawdev *dev,
 			   sizeof(struct ntb_used) * nb1);
 		rte_memcpy(txq->tx_used_ring, tx_used + nb1,
 			   sizeof(struct ntb_used) * nb2);
+		rte_io_wmb();

As both txq->tx_used_ring and *txq->used_cnt are physically reside in the

PCI

Quoted 7 lines

device side, rte_io_wmb is correct to ensure the ordering.

drivers/raw/ntb/ntb.c ntb_enqueue_bufs(struct rte_rawdev *dev,

Quoted 5 lines
drivers/raw/ntb/ntb.c +0 -1
 		*txq->used_cnt = txq->last_used;
-		rte_wmb();
 
 		/* update queue stats */
 		hw->ntb_xstats[NTB_TX_BYTES_ID + off] += bytes; @@ -

789,8

Quoted 15 lines

+789,8 @@

Quoted 11 lines

ntb_dequeue_bufs(struct rte_rawdev *dev,

drivers/raw/ntb/ntb.c +1 -2
 			   sizeof(struct ntb_desc) * nb1);
 		rte_memcpy(rxq->rx_desc_ring, rx_desc + nb1,
 			   sizeof(struct ntb_desc) * nb2);
+		rte_io_wmb();
 		*rxq->avail_cnt = rxq->last_avail;
-		rte_wmb();
 
 		/* update queue stats */
 		off = NTB_XSTATS_NUM * ((size_t)context + 1);
--
Quoted 1 line

2.17.1

Acked-by on v2 1/1
#10 lore raw
In reply to Xiaoyun Li
Quoted 10 lines

-----Original Message-----
From: Li, Xiaoyun <[email protected]>
Sent: Monday, December 16, 2019 9:59 AM
To: Wu, Jingjing <[email protected]>
Cc: [email protected]; Maslekar, Omkar <[email protected]>; Li, Xiaoyun <[email protected]>; [email protected]
Subject: [PATCH v2] raw/ntb: fix write memory barrier issue

Quoted 8 lines

All buffers and ring info should be written before tail register update. This patch relocates the write memory barrier before updating tail register to avoid potential issues.

Fixes 11b5c7daf019 ("raw/ntb: add enqueue and dequeue functions")
Quoted 33 lines

---
v2:
* Replaced rte_wmb with rte_io_wmb since rte_io_wmb is enough.

---
 drivers/raw/ntb/ntb.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
drivers/raw/ntb/ntb.c +2 -3
index ad7f6abfd..c7de86f36 100644
@@ -683,8 +683,8 @@ ntb_enqueue_bufs(struct rte_rawdev *dev,
 			   sizeof(struct ntb_used) * nb1);
 		rte_memcpy(txq->tx_used_ring, tx_used + nb1,
 			   sizeof(struct ntb_used) * nb2);
+		rte_io_wmb();
 		*txq->used_cnt = txq->last_used;
-		rte_wmb();
 
 		/* update queue stats */
 		hw->ntb_xstats[NTB_TX_BYTES_ID + off] += bytes;
@@ -789,8 +789,8 @@ ntb_dequeue_bufs(struct rte_rawdev *dev,
 			   sizeof(struct ntb_desc) * nb1);
 		rte_memcpy(rxq->rx_desc_ring, rx_desc + nb1,
 			   sizeof(struct ntb_desc) * nb2);
+		rte_io_wmb();
 		*rxq->avail_cnt = rxq->last_avail;
-		rte_wmb();
 
 		/* update queue stats */
 		off = NTB_XSTATS_NUM * ((size_t)context + 1);
--

2.17.1

Acked-by on v1 1/1
#11 lore raw
In reply to Xiaoyun Li
Quoted 9 lines

-----Original Message-----
From: Li, Xiaoyun <[email protected]>
Sent: Wednesday, December 4, 2019 11:19 PM
To: Wu, Jingjing <[email protected]>
Cc: [email protected]; Li, Xiaoyun <[email protected]>; [email protected]
Subject: [PATCH] raw/ntb: fix write memory barrier issue

Quoted 8 lines

All buffers and ring info should be written before tail register update. This patch relocates the write memory barrier before updating tail register to avoid potential issues.

Fixes 11b5c7daf019 ("raw/ntb: add enqueue and dequeue functions")
#12 lore raw

26/12/2019 02:46, Wu, Jingjing:

Quoted 10 lines

From: Li, Xiaoyun <[email protected]>

Quoted 8 lines

All buffers and ring info should be written before tail register update. This patch relocates the write memory barrier before updating tail register to avoid potential issues.

Fixes 11b5c7daf019 ("raw/ntb: add enqueue and dequeue functions")

v1 applied, thanks