Skip to contentSkip to conversation

net/mlx5: fix the eCPRI common header endianness

PATCH In review Bing Zhao 3 participants 6 messages Nov 3, 2020 – Nov 9, 2020 Network drivers

lore
Share & tools
lore

Bing Zhao

author
#0 lore raw

PATCH 1/1 net/mlx5: fix the eCPRI common header endianness

The input header of a RTE flow item is with network byte order. In the host with little endian, the bit field order are the same as the byte order.
When checking the an eCPRI message type, the wrong field will be selected. Right now, since the whole u32 is being checked and for all types, the following implementation is unique. There is no functional risk but it is still an error to fix.

Fixes daa38a8924a0 ("net/mlx5: add flow translation of eCPRI header")
---
 drivers/net/mlx5/mlx5_flow_dv.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)
drivers/net/mlx5/mlx5_flow_dv.c +3 -1
index 01b6e7c..7af01e9 100644
@@ -7798,6 +7798,7 @@ struct mlx5_hlist_entry *
 	struct mlx5_priv *priv = dev->data->dev_private;
 	const struct rte_flow_item_ecpri *ecpri_m = item->mask;
 	const struct rte_flow_item_ecpri *ecpri_v = item->spec;
+	struct rte_ecpri_common_hdr common;
 	void *misc4_m = MLX5_ADDR_OF(fte_match_param, matcher,
 				     misc_parameters_4);
 	void *misc4_v = MLX5_ADDR_OF(fte_match_param, key, misc_parameters_4);
@@ -7838,7 +7839,8 @@ struct mlx5_hlist_entry *
 	 * Some wildcard rules only matching type field should be supported.
 	 */
 	if (ecpri_m->hdr.dummy[0]) {
-		switch (ecpri_v->hdr.common.type) {
+		common.u32 = rte_be_to_cpu_32(ecpri_v->hdr.common.u32);
+		switch (common.type) {
 		case RTE_ECPRI_MSG_TYPE_IQ_DATA:
 		case RTE_ECPRI_MSG_TYPE_RTC_CTRL:
 		case RTE_ECPRI_MSG_TYPE_DLY_MSR:
Signature · 1 line
1.8.3.1
#1 lore raw

Hi,

Quoted 19 lines

-----Original Message-----
From: Bing Zhao <[email protected]>
Sent: Tuesday, November 3, 2020 7:42 AM
To: Slava Ovsiienko <[email protected]>; Matan Azrad <[email protected]>
Cc: [email protected]; Ori Kam <[email protected]>; Raslan Darawsheh <[email protected]>; [email protected]
Subject: [PATCH] net/mlx5: fix the eCPRI common header endianness

The input header of a RTE flow item is with network byte order. In the host with little endian, the bit field order are the same as the byte order.
When checking the an eCPRI message type, the wrong field will be selected. Right now, since the whole u32 is being checked and for all types, the following implementation is unique. There is no functional risk but it is still an error to fix.

Fixes daa38a8924a0 ("net/mlx5: add flow translation of eCPRI header")
Quoted 8 lines
---
 drivers/net/mlx5/mlx5_flow_dv.c | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

Patch applied to next-net-mlx,

Kindest regards,
Raslan Darawsheh

#2 lore raw
In reply to Bing Zhao
Quoted 8 lines · Bing Zhao

The input header of a RTE flow item is with network byte order. In the host with little endian, the bit field order are the same as the byte order.
When checking the an eCPRI message type, the wrong field will be selected. Right now, since the whole u32 is being checked and for all types, the following implementation is unique. There is no functional risk but it is still an error to fix.

Isn't the 'ecpri_v' filled by application (CPU), why there is an assumption that it is big endian?

And even if it is big endian, it should be broken previously right? Since it was checking wrong field as 'type' as you said, why there were no functional risk?

Quoted 34 lines
Fixes daa38a8924a0 ("net/mlx5: add flow translation of eCPRI header")
---
  drivers/net/mlx5/mlx5_flow_dv.c | 4 +++-
  1 file changed, 3 insertions(+), 1 deletion(-)
drivers/net/mlx5/mlx5_flow_dv.c +3 -1
index 01b6e7c..7af01e9 100644
@@ -7798,6 +7798,7 @@ struct mlx5_hlist_entry *
  	struct mlx5_priv *priv = dev->data->dev_private;
  	const struct rte_flow_item_ecpri *ecpri_m = item->mask;
  	const struct rte_flow_item_ecpri *ecpri_v = item->spec;
+	struct rte_ecpri_common_hdr common;
  	void *misc4_m = MLX5_ADDR_OF(fte_match_param, matcher,
  				     misc_parameters_4);
  	void *misc4_v = MLX5_ADDR_OF(fte_match_param, key, misc_parameters_4);
@@ -7838,7 +7839,8 @@ struct mlx5_hlist_entry *
  	 * Some wildcard rules only matching type field should be supported.
  	 */
  	if (ecpri_m->hdr.dummy[0]) {
-		switch (ecpri_v->hdr.common.type) {
+		common.u32 = rte_be_to_cpu_32(ecpri_v->hdr.common.u32);
+		switch (common.type) {
  		case RTE_ECPRI_MSG_TYPE_IQ_DATA:
  		case RTE_ECPRI_MSG_TYPE_RTC_CTRL:
  		case RTE_ECPRI_MSG_TYPE_DLY_MSR:
author
#3 lore raw

Hi Ferruh,

Thanks for your review and comments, PSB.

Quoted 21 lines

-----Original Message-----
From: Ferruh Yigit <[email protected]>
Sent: Friday, November 6, 2020 7:20 PM
To: Bing Zhao <[email protected]>; Slava Ovsiienko <[email protected]>; Matan Azrad <[email protected]>
Cc: [email protected]; Ori Kam <[email protected]>; Raslan Darawsheh <[email protected]>; [email protected]
Subject: Re: [dpdk-dev] [PATCH] net/mlx5: fix the eCPRI common header endianness

External email: Use caution opening links or attachments

Quoted 1 line · Bing Zhao

The input header of a RTE flow item is with network byte order. In

the

Quoted 1 line

host with little endian, the bit field order are the same as the

byte

Quoted 3 lines

order.
When checking the an eCPRI message type, the wrong field will be selected. Right now, since the whole u32 is being checked and for

Quoted 9 lines

all

Quoted 1 line

types, the following implementation is unique. There is no

functional

Quoted 2 lines

risk but it is still an error to fix.

Isn't the 'ecpri_v' filled by application (CPU), why there is an assumption that it is big endian?

Yes, this is filled by the application SW. I checked the current code of other headers and current implementation.
1. In the testpmd flow example, all the headers of network stack like IPv4 are translated into to be in BE, "ARGS_ENTRY_HTON"
2. All fields are with "rte_be*_t"type, even though only a typedef, it should be considered in BE.
3. RTE flow will just pass the flow items pointer to the PMD drivers, so in the PMD part, the headers should be treated as in BE. So, to my understanding, this is not an assumption but some constraint.
Correct me if my understanding is wrong.

Quoted 3 lines

And even if it is big endian, it should be broken previously right? Since it was checking wrong field as 'type' as you said, why there were no functional risk?

In the PMD driver, the first u32 *value containing this type is already used for matching. And a checking is used for the following "sub-headers" matching. Indeed, the first u32 *mask is used to confirm if the following sub-header need to be matched.
Since different types will lead to different variants of the packets, the switch-of-type is used.
But all the 3 types supported in PMD now almost have the same results (part of the second u32, remaining will be all 0s). So even if the type of the value is always "0" by mistake, the cases are the same and it still works by coincidence.
If more types are supported in the future, this will be problematic.

Quoted 39 lines
Quoted 3 lines
Fixes daa38a8924a0 ("net/mlx5: add flow translation of eCPRI

header")

Quoted 21 lines
---
  drivers/net/mlx5/mlx5_flow_dv.c | 4 +++-
  1 file changed, 3 insertions(+), 1 deletion(-)
(quoted diff) +0 -0

b/drivers/net/mlx5/mlx5_flow_dv.c index 01b6e7c..7af01e9 100644

drivers/net/mlx5/mlx5_flow_dv.c +1 -0
@@ -7798,6 +7798,7 @@ struct mlx5_hlist_entry *
      struct mlx5_priv *priv = dev->data->dev_private;
      const struct rte_flow_item_ecpri *ecpri_m = item->mask;
      const struct rte_flow_item_ecpri *ecpri_v = item->spec;
+     struct rte_ecpri_common_hdr common;
      void *misc4_m = MLX5_ADDR_OF(fte_match_param, matcher,
                                   misc_parameters_4);
      void *misc4_v = MLX5_ADDR_OF(fte_match_param, key,
Quoted 1 line

misc_parameters_4); @@ -7838,7 +7839,8 @@ struct mlx5_hlist_entry

*

Quoted 1 line

* Some wildcard rules only matching type field should be

supported.

Quoted 10 lines
       */
      if (ecpri_m->hdr.dummy[0]) {
drivers/net/mlx5/mlx5_flow_dv.c +1 -1
-             switch (ecpri_v->hdr.common.type) {
+             common.u32 = rte_be_to_cpu_32(ecpri_v-

hdr.common.u32);
+ switch (common.type) {
case RTE_ECPRI_MSG_TYPE_IQ_DATA:
case RTE_ECPRI_MSG_TYPE_RTC_CTRL:
case RTE_ECPRI_MSG_TYPE_DLY_MSR:

#4 lore raw
Quoted 35 lines · Bing Zhao

Hi Ferruh,

Thanks for your review and comments, PSB.

Quoted 21 lines

-----Original Message-----
From: Ferruh Yigit <[email protected]>
Sent: Friday, November 6, 2020 7:20 PM
To: Bing Zhao <[email protected]>; Slava Ovsiienko <[email protected]>; Matan Azrad <[email protected]>
Cc: [email protected]; Ori Kam <[email protected]>; Raslan Darawsheh <[email protected]>; [email protected]
Subject: Re: [dpdk-dev] [PATCH] net/mlx5: fix the eCPRI common header endianness

External email: Use caution opening links or attachments

Quoted 1 line · Bing Zhao

The input header of a RTE flow item is with network byte order. In

the

Quoted 1 line

host with little endian, the bit field order are the same as the

byte

Quoted 3 lines

order.
When checking the an eCPRI message type, the wrong field will be selected. Right now, since the whole u32 is being checked and for

Quoted 9 lines

all

Quoted 1 line

types, the following implementation is unique. There is no

functional

Quoted 2 lines

risk but it is still an error to fix.

Isn't the 'ecpri_v' filled by application (CPU), why there is an assumption that it is big endian?

Quoted 6 lines

Yes, this is filled by the application SW. I checked the current code of other headers and current implementation.
1. In the testpmd flow example, all the headers of network stack like IPv4 are translated into to be in BE, "ARGS_ENTRY_HTON"
2. All fields are with "rte_be*_t"type, even though only a typedef, it should be considered in BE.
3. RTE flow will just pass the flow items pointer to the PMD drivers, so in the PMD part, the headers should be treated as in BE. So, to my understanding, this is not an assumption but some constraint.
Correct me if my understanding is wrong.

I missed in 'cmdline_flow.c' big endian conversion done via 'arg->hton' check, so yes PMD getting big endian values makes sense, thanks for clarification.

Quoted 10 lines
Quoted 4 lines

And even if it is big endian, it should be broken previously right? Since it was checking wrong field as 'type' as you said, why there were no functional risk?

In the PMD driver, the first u32 *value containing this type is already used for matching. And a checking is used for the following "sub-headers" matching. Indeed, the first u32 *mask is used to confirm if the following sub-header need to be matched.
Since different types will lead to different variants of the packets, the switch-of-type is used.
But all the 3 types supported in PMD now almost have the same results (part of the second u32, remaining will be all 0s). So even if the type of the value is always "0" by mistake, the cases are the same and it still works by coincidence.

if 'type' is '0' I see it works, but what if the 'type' is something other than values covered in 'switch', than it may fail matching 'sub-headers' and I guess that can happen based on value of 'size' field.

Anyway, since you are already fixing it, will you be OK if I drop last two sentences from the commit log and proceed?

author
#5 lore raw

Hi Ferruh,

Quoted 83 lines

-----Original Message-----
From: Ferruh Yigit <[email protected]>
Sent: Saturday, November 7, 2020 1:41 AM
To: Bing Zhao <[email protected]>; Slava Ovsiienko <[email protected]>; Matan Azrad <[email protected]>
Cc: [email protected]; Ori Kam <[email protected]>; Raslan Darawsheh <[email protected]>; [email protected]
Subject: Re: [dpdk-dev] [PATCH] net/mlx5: fix the eCPRI common header endianness

External email: Use caution opening links or attachments

Quoted 12 lines · Bing Zhao

Hi Ferruh,

Thanks for your review and comments, PSB.

Quoted 8 lines

-----Original Message-----
From: Ferruh Yigit <[email protected]>
Sent: Friday, November 6, 2020 7:20 PM
To: Bing Zhao <[email protected]>; Slava Ovsiienko <[email protected]>; Matan Azrad <[email protected]>
Cc: [email protected]; Ori Kam <[email protected]>; Raslan Darawsheh <[email protected]>; [email protected]
Subject: Re: [dpdk-dev] [PATCH] net/mlx5: fix the eCPRI common

header

Quoted 7 lines
Quoted 7 lines

endianness

External email: Use caution opening links or attachments

Quoted 1 line · Bing Zhao

The input header of a RTE flow item is with network byte order.

In

Quoted 6 lines
Quoted 6 lines

the

Quoted 1 line

host with little endian, the bit field order are the same as the

byte

Quoted 3 lines

order.
When checking the an eCPRI message type, the wrong field will be selected. Right now, since the whole u32 is being checked and

for

Quoted 11 lines
Quoted 9 lines

all

Quoted 1 line

types, the following implementation is unique. There is no

functional

Quoted 2 lines

risk but it is still an error to fix.

Isn't the 'ecpri_v' filled by application (CPU), why there is an assumption that it is big endian?

Yes, this is filled by the application SW. I checked the current

code of other headers and current implementation.

Quoted 1 line

1. In the testpmd flow example, all the headers of network stack

like IPv4 are translated into to be in BE, "ARGS_ENTRY_HTON"

Quoted 1 line

2. All fields are with "rte_be*_t"type, even though only a typedef,

it should be considered in BE.

Quoted 1 line

3. RTE flow will just pass the flow items pointer to the PMD

drivers, so in the PMD part, the headers should be treated as in BE.

Quoted 1 line

So, to my understanding, this is not an assumption but some

constraint.

Quoted 1 line

Correct me if my understanding is wrong.

I missed in 'cmdline_flow.c' big endian conversion done via 'arg-

Quoted 1 line

hton' check, so yes PMD getting big endian values makes sense,

thanks for clarification.

Quoted 2 lines
Quoted 2 lines

And even if it is big endian, it should be broken previously

right?

Quoted 1 line
Quoted 1 line

Since it was checking wrong field as 'type' as you said, why

there

Quoted 3 lines
Quoted 1 line

were no functional risk?

In the PMD driver, the first u32 *value containing this type is

already used for matching. And a checking is used for the following "sub-headers"

Quoted 1 line

matching. Indeed, the first u32 *mask is used to confirm if the

following sub-header need to be matched.

Quoted 1 line

Since different types will lead to different variants of the

packets, the switch-of-type is used.

Quoted 1 line

But all the 3 types supported in PMD now almost have the same

Quoted 8 lines

results (part of the second u32, remaining will be all 0s).

Quoted 1 line

So even if the type of the value is always "0" by mistake, the

cases are the same and it still works by coincidence.

if 'type' is '0' I see it works, but what if the 'type' is something other than values covered in 'switch', than it may fail matching 'sub-headers' and I guess that can happen based on value of 'size' field.

Yes, it has the risk. In the past, the packet's length tested didn't expose this issue. So when it change the size upper byte to other value instead of 0/2/5, then the flow will get a failure when being created. Thanks for catching this point.

Quoted 3 lines

Anyway, since you are already fixing it, will you be OK if I drop last two sentences from the commit log and proceed?

Sure, thanks a lot for your help.

BR. Bing