Skip to contentSkip to conversation

doc: announce ABI change for cryptodev config

PATCH Superseded Anoob Joseph 5 participants 13 messages Jan 17, 2019 – Mar 29, 2019

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 doc: announce ABI change for cryptodev config

Add new field ff_enable in rte_cryptodev_config. This enables applications to control the features enabled on the crypto device.

Proposed new layout:

/** Crypto device configuration structure */

struct rte_cryptodev_config {
    int socket_id;            /**< Socket to allocate resources on */
    uint16_t nb_queue_pairs;
    /**< Number of queue pairs to configure on device */

+ uint64_t ff_enable;
+ /**< Feature flags to be enabled on the device. Only the features set
+ * on rte_cryptodev_info.feature_flags are allowed to be set.
+ */
};

For eth devices, rte_eth_conf.rx_mode.offloads and rte_eth_conf.tx_mode.offloads fields are used by applications to control the offloads enabled on the eth device. This proposal adds a similar ability for the crypto device.

---
 doc/guides/rel_notes/deprecation.rst | 10 ++++++++++
 1 file changed, 10 insertions(+)
doc/guides/rel_notes/deprecation.rst +10 -0
index 5f03443..e289c2d 100644
@@ -66,3 +66,13 @@ Deprecation Notices
 
 * crypto/aesni_mb: the minimum supported intel-ipsec-mb library version will be
   changed from 0.49.0 to 0.52.0.
+
+* cryptodev: New member in ``rte_cryptodev_config`` to allow applications to
+  specify the features to be enabled on the crypto device. For eth devices,
+  applications can use ``rte_eth_conf.rxmode.offloads`` and
+  ``rte_eth_conf.txmode.offloads`` to control the offloads enabled. Adding
+  a similar field to facilitate efficient usage of HW/SW offloads.
+
+  - Member ``uint64_t ff_enable`` in ``rte_cryptodev_config``
+
+  The field would be added in v19.05.
Signature · 1 line
2.7.4
#1 lore raw

Hi Joseph,

Quoted 25 lines

-----Original Message-----
From: Anoob Joseph [mailto:[email protected]]
Sent: Thursday, January 17, 2019 9:40 AM
To: Akhil Goyal <[email protected]>; De Lara Guarch, Pablo <[email protected]>; Trahe, Fiona <[email protected]>
Cc: Anoob Joseph <[email protected]>; Jerin Jacob Kollanukkaran <[email protected]>; Narayana Prasad Raju Athreya <[email protected]>; Shally Verma <[email protected]>; [email protected]
Subject: [PATCH] doc: announce ABI change for cryptodev config

Add new field ff_enable in rte_cryptodev_config. This enables applications to control the features enabled on the crypto device.

Proposed new layout:

/** Crypto device configuration structure */

struct rte_cryptodev_config {
    int socket_id;            /**< Socket to allocate resources on */
    uint16_t nb_queue_pairs;
    /**< Number of queue pairs to configure on device */
(quoted diff) +4 -0
+   uint64_t ff_enable;
+   /**< Feature flags to be enabled on the device. Only the features set
+    * on rte_cryptodev_info.feature_flags are allowed to be set.
+    */

};

Quoted 4 lines

For eth devices, rte_eth_conf.rx_mode.offloads and rte_eth_conf.tx_mode.offloads fields are used by applications to control the offloads enabled on the eth device. This proposal adds a similar ability for the crypto device.

[Fiona] I'm unfamiliar with eth config so can you explain a bit more how this works? e.g. if a crypto device's info says it supports
RTE_CRYPTODEV_FF_SYMMETRIC_CRYPTO
RTE_CRYPTODEV_FF_SYM_OPERATION_CHAINING
RTE_CRYPTODEV_FF_CPU_AESNI
RTE_CRYPTODEV_FF_SECURITY
RTE_CRYPTODEV_FF_OOP_LB_IN_LB_OUT
these things are all already enabled.
Is the param a way to disable some of them?
If so, it would be better to call it ff_disable.
And to limit it to the subset of features that it makes sense to disable.
i.e. why would an application disable chaining or any of the SGL flags? It would just add extra cycles in the PMD to error check on these cases, instead the appl can just not send such commands. And it doesn't make sense to disable AESNI or RTE_CRYPTODEV_FF_HW_ACCELERATED. Actually I can't see what an application would want to achieve by disabling any flag?

#2 lore raw

Hi Fiona,

Please see inline.

Thanks,
Anoob

Quoted 39 lines

-----Original Message-----
From: Trahe, Fiona <[email protected]>
Sent: 17 January 2019 17:07
To: Anoob Joseph <[email protected]>; Akhil Goyal <[email protected]>; De Lara Guarch, Pablo <[email protected]>
Cc: Jerin Jacob Kollanukkaran <[email protected]>; Narayana Prasad Raju Athreya <[email protected]>; Shally Verma <[email protected]>; [email protected]; Trahe, Fiona <[email protected]>
Subject: RE: [PATCH] doc: announce ABI change for cryptodev config

Hi Joseph,

Quoted 27 lines

-----Original Message-----
From: Anoob Joseph [mailto:[email protected]]
Sent: Thursday, January 17, 2019 9:40 AM
To: Akhil Goyal <[email protected]>; De Lara Guarch, Pablo <[email protected]>; Trahe, Fiona <[email protected]>
Cc: Anoob Joseph <[email protected]>; Jerin Jacob Kollanukkaran <[email protected]>; Narayana Prasad Raju Athreya <[email protected]>; Shally Verma <[email protected]>; [email protected]
Subject: [PATCH] doc: announce ABI change for cryptodev config

Add new field ff_enable in rte_cryptodev_config. This enables applications to control the features enabled on the crypto device.

Proposed new layout:

/** Crypto device configuration structure */ struct

rte_cryptodev_config {
    int socket_id;            /**< Socket to allocate resources on */
    uint16_t nb_queue_pairs;
    /**< Number of queue pairs to configure on device */
(quoted diff) +4 -0
+   uint64_t ff_enable;
+   /**< Feature flags to be enabled on the device. Only the features set
+    * on rte_cryptodev_info.feature_flags are allowed to be set.
+    */

};

Quoted 6 lines
Quoted 4 lines

For eth devices, rte_eth_conf.rx_mode.offloads and rte_eth_conf.tx_mode.offloads fields are used by applications to control the offloads enabled on the eth device. This proposal adds a similar ability for the crypto device.

[Fiona] I'm unfamiliar with eth config so can you explain a bit more how this works?

[Anoob] For eth devices, application would first do rte_eth_dev_info_get() and gets the capabilities. The device would expose it's capabilities using dev_info.rx_offload_capa & dev_info.tx_offload_capa. The application can enable/disable these features while configuring the eth ports.

Quoted 1 line

From ipsec-secgw:

https://git.dpdk.org/dpdk/tree/examples/ipsec-secgw/ipsec-secgw.c#n1866

if (frame_size) {
		local_port_conf.rxmode.max_rx_pkt_len = frame_size;
		local_port_conf.rxmode.offloads |= DEV_RX_OFFLOAD_JUMBO_FRAME;
	}

<snip>

ret = rte_eth_dev_configure(portid, nb_rx_queue, nb_tx_queue,

			&local_port_conf);

<snip>

This way application can choose to disable unwanted offloads.

Port init in ipsec-secgw: https://git.dpdk.org/dpdk/tree/examples/ipsec-secgw/ipsec-secgw.c#n1826

Quoted 8 lines

e.g. if a crypto device's info says it supports
RTE_CRYPTODEV_FF_SYMMETRIC_CRYPTO
RTE_CRYPTODEV_FF_SYM_OPERATION_CHAINING
RTE_CRYPTODEV_FF_CPU_AESNI
RTE_CRYPTODEV_FF_SECURITY
RTE_CRYPTODEV_FF_OOP_LB_IN_LB_OUT
these things are all already enabled.
Is the param a way to disable some of them?

[Anoob] Yes. There are few other flags in addition to the above. Like RTE_CRYPTODEV_FF_ASYMMETRIC_CRYPTO.

Quoted 1 line

If so, it would be better to call it ff_disable.

[Anoob] Calling it ff_enable is to make it similar to how it's done with eth devices. Either way should be fine.

Quoted 8 lines

And to limit it to the subset of features that it makes sense to disable. i.e. why would an application disable chaining or any of the SGL flags? It would just add extra cycles in the PMD to error check on these cases, instead the appl can just not send such commands.
And it doesn't make sense to disable AESNI or RTE_CRYPTODEV_FF_HW_ACCELERATED.
Actually I can't see what an application would want to achieve by disabling any flag?

[Anoob] Features like ASYMMETRIC or SECURITY is not required for every application. SECURITY is added for eth devices also. But since the feature can be disabled while configuring the port, applications are given a choice to disable it. That way apps like l2fwd doesn't enable SECURITY. With crypto this option is not available.

If the unused offloads can be communicated to the PMD before initialization, the PMD will be allowed to optimize hardware usage. Otherwise, supporting more features would affect performance, even if application is not making use of those features.

Ex: test-crypto-perf doesn't use ASYM/SECURITY. Now adding these features would affect the performance of this app.

#3 lore raw

HI Fiona, Anoob

Quoted 94 lines

-----Original Message-----
From: Anoob Joseph
Sent: 17 January 2019 19:17
To: Trahe, Fiona <[email protected]>; Akhil Goyal <[email protected]>; De Lara Guarch, Pablo <[email protected]>
Cc: Jerin Jacob Kollanukkaran <[email protected]>; Narayana Prasad Raju Athreya <[email protected]>; Shally Verma <[email protected]>; [email protected]
Subject: RE: [PATCH] doc: announce ABI change for cryptodev config

Hi Fiona,

Please see inline.

Thanks,
Anoob

Quoted 39 lines

-----Original Message-----
From: Trahe, Fiona <[email protected]>
Sent: 17 January 2019 17:07
To: Anoob Joseph <[email protected]>; Akhil Goyal <[email protected]>; De Lara Guarch, Pablo <[email protected]>
Cc: Jerin Jacob Kollanukkaran <[email protected]>; Narayana Prasad Raju Athreya <[email protected]>; Shally Verma <[email protected]>; [email protected]; Trahe, Fiona <[email protected]>
Subject: RE: [PATCH] doc: announce ABI change for cryptodev config

Hi Joseph,

Quoted 27 lines

-----Original Message-----
From: Anoob Joseph [mailto:[email protected]]
Sent: Thursday, January 17, 2019 9:40 AM
To: Akhil Goyal <[email protected]>; De Lara Guarch, Pablo <[email protected]>; Trahe, Fiona <[email protected]>
Cc: Anoob Joseph <[email protected]>; Jerin Jacob Kollanukkaran <[email protected]>; Narayana Prasad Raju Athreya <[email protected]>; Shally Verma <[email protected]>; [email protected]
Subject: [PATCH] doc: announce ABI change for cryptodev config

Add new field ff_enable in rte_cryptodev_config. This enables applications to control the features enabled on the crypto device.

Proposed new layout:

/** Crypto device configuration structure */ struct

rte_cryptodev_config {
    int socket_id;            /**< Socket to allocate resources on */
    uint16_t nb_queue_pairs;
    /**< Number of queue pairs to configure on device */
(quoted diff) +4 -0
+   uint64_t ff_enable;
+   /**< Feature flags to be enabled on the device. Only the features set
+    * on rte_cryptodev_info.feature_flags are allowed to be set.
+    */

};

Quoted 6 lines
Quoted 4 lines

For eth devices, rte_eth_conf.rx_mode.offloads and rte_eth_conf.tx_mode.offloads fields are used by applications to control the offloads enabled on the eth device. This proposal adds a similar ability for the crypto device.

[Fiona] I'm unfamiliar with eth config so can you explain a bit more how this works?

[Anoob] For eth devices, application would first do rte_eth_dev_info_get() and gets the capabilities. The device would expose it's capabilities using dev_info.rx_offload_capa & dev_info.tx_offload_capa. The application can enable/disable these features while configuring the eth ports.

From ipsec-secgw:
https://git.dpdk.org/dpdk/tree/examples/ipsec-secgw/ipsec-secgw.c#n1866

if (frame_size) {
		local_port_conf.rxmode.max_rx_pkt_len = frame_size;
		local_port_conf.rxmode.offloads |= DEV_RX_OFFLOAD_JUMBO_FRAME;
	}

<snip>

ret = rte_eth_dev_configure(portid, nb_rx_queue, nb_tx_queue,

			&local_port_conf);

<snip>

This way application can choose to disable unwanted offloads.

Port init in ipsec-secgw: https://git.dpdk.org/dpdk/tree/examples/ipsec-secgw/ipsec-secgw.c#n1826

Quoted 8 lines

e.g. if a crypto device's info says it supports
RTE_CRYPTODEV_FF_SYMMETRIC_CRYPTO
RTE_CRYPTODEV_FF_SYM_OPERATION_CHAINING
RTE_CRYPTODEV_FF_CPU_AESNI
RTE_CRYPTODEV_FF_SECURITY
RTE_CRYPTODEV_FF_OOP_LB_IN_LB_OUT
these things are all already enabled.
Is the param a way to disable some of them?

Quoted 5 lines

[Anoob] Yes. There are few other flags in addition to the above. Like RTE_CRYPTODEV_FF_ASYMMETRIC_CRYPTO.

Quoted 1 line

If so, it would be better to call it ff_disable.

[Anoob] Calling it ff_enable is to make it similar to how it's done with eth devices. Either way should be fine.

[Shally] keeping it as "ff_enable" will require application to set these flags to configure PMD. If we define it other way around, then it would be mean to mask out unwanted features which can be quite many. Though purpose can be achieved both ways but keeping it as "enable" looks more easy to use, readable and inline with how ethdev handle multi feature support.
So I would prefer to keep it as "ff_enable"

Thanks
Shally

Quoted 18 lines
Quoted 9 lines

And to limit it to the subset of features that it makes sense to disable. i.e. why would an application disable chaining or any of the SGL flags? It would just add extra cycles in the PMD to error check on these cases, instead the appl can just not send such commands.
And it doesn't make sense to disable AESNI or RTE_CRYPTODEV_FF_HW_ACCELERATED.
Actually I can't see what an application would want to achieve by disabling any flag?

[Anoob] Features like ASYMMETRIC or SECURITY is not required for every application. SECURITY is added for eth devices also. But since the feature can be disabled while configuring the port, applications are given a choice to disable it. That way apps like l2fwd doesn't enable SECURITY. With crypto this option is not available.

If the unused offloads can be communicated to the PMD before initialization, the PMD will be allowed to optimize hardware usage. Otherwise, supporting more features would affect performance, even if application is not making use of those features.

Ex: test-crypto-perf doesn't use ASYM/SECURITY. Now adding these features would affect the performance of this app.

#4 lore raw

Hi Fiona,

Any more comments on this?

@ Akhil, Pablo

Can you review this change and share your thoughts?

Thanks,
Anoob

Quoted 156 lines

-----Original Message-----
From: Shally Verma
Sent: 18 January 2019 12:29
To: Anoob Joseph <[email protected]>; Trahe, Fiona <[email protected]>; Akhil Goyal <[email protected]>; De Lara Guarch, Pablo <[email protected]>
Cc: Jerin Jacob Kollanukkaran <[email protected]>; Narayana Prasad Raju Athreya <[email protected]>; [email protected]
Subject: RE: [PATCH] doc: announce ABI change for cryptodev config

HI Fiona, Anoob

Quoted 64 lines

-----Original Message-----
From: Anoob Joseph
Sent: 17 January 2019 19:17
To: Trahe, Fiona <[email protected]>; Akhil Goyal <[email protected]>; De Lara Guarch, Pablo <[email protected]>
Cc: Jerin Jacob Kollanukkaran <[email protected]>; Narayana Prasad Raju Athreya <[email protected]>; Shally Verma <[email protected]>; [email protected]
Subject: RE: [PATCH] doc: announce ABI change for cryptodev config

Hi Fiona,

Please see inline.

Thanks,
Anoob

Quoted 46 lines

-----Original Message-----
From: Trahe, Fiona <[email protected]>
Sent: 17 January 2019 17:07
To: Anoob Joseph <[email protected]>; Akhil Goyal <[email protected]>; De Lara Guarch, Pablo <[email protected]>
Cc: Jerin Jacob Kollanukkaran <[email protected]>; Narayana Prasad Raju Athreya <[email protected]>; Shally Verma <[email protected]>; [email protected]; Trahe, Fiona <[email protected]>
Subject: RE: [PATCH] doc: announce ABI change for cryptodev config

Hi Joseph,

Quoted 32 lines

-----Original Message-----
From: Anoob Joseph [mailto:[email protected]]
Sent: Thursday, January 17, 2019 9:40 AM
To: Akhil Goyal <[email protected]>; De Lara Guarch, Pablo <[email protected]>; Trahe, Fiona <[email protected]>
Cc: Anoob Joseph <[email protected]>; Jerin Jacob Kollanukkaran <[email protected]>; Narayana Prasad Raju Athreya <[email protected]>; Shally Verma <[email protected]>; [email protected]
Subject: [PATCH] doc: announce ABI change for cryptodev config

Add new field ff_enable in rte_cryptodev_config. This enables applications to control the features enabled on the crypto device.

Proposed new layout:

/** Crypto device configuration structure */ struct

rte_cryptodev_config {
    int socket_id;            /**< Socket to allocate resources on */
    uint16_t nb_queue_pairs;
    /**< Number of queue pairs to configure on device */
(quoted diff) +4 -0
+   uint64_t ff_enable;
+   /**< Feature flags to be enabled on the device. Only the features set
+    * on rte_cryptodev_info.feature_flags are allowed to be set.
+    */

};

For eth devices, rte_eth_conf.rx_mode.offloads and rte_eth_conf.tx_mode.offloads fields are used by applications to control the offloads enabled on the eth device. This proposal adds a similar ability for the crypto device.

Quoted 6 lines
Quoted 2 lines

[Fiona] I'm unfamiliar with eth config so can you explain a bit more how this works?

[Anoob] For eth devices, application would first do rte_eth_dev_info_get() and gets the capabilities. The device would expose it's capabilities using dev_info.rx_offload_capa &

dev_info.tx_offload_capa. The application can enable/disable these features while configuring the eth ports.

Quoted 7 lines

From ipsec-secgw:
https://git.dpdk.org/dpdk/tree/examples/ipsec-secgw/ipsec-secgw.c#n1866

if (frame_size) {
		local_port_conf.rxmode.max_rx_pkt_len = frame_size;
		local_port_conf.rxmode.offloads |=

DEV_RX_OFFLOAD_JUMBO_FRAME;

Quoted 23 lines
	}

<snip>

ret = rte_eth_dev_configure(portid, nb_rx_queue, nb_tx_queue,

			&local_port_conf);

<snip>

This way application can choose to disable unwanted offloads.

Port init in ipsec-secgw:
https://git.dpdk.org/dpdk/tree/examples/ipsec-secgw/ipsec-secgw.c#n1826

Quoted 8 lines

e.g. if a crypto device's info says it supports
RTE_CRYPTODEV_FF_SYMMETRIC_CRYPTO
RTE_CRYPTODEV_FF_SYM_OPERATION_CHAINING
RTE_CRYPTODEV_FF_CPU_AESNI
RTE_CRYPTODEV_FF_SECURITY
RTE_CRYPTODEV_FF_OOP_LB_IN_LB_OUT
these things are all already enabled.
Is the param a way to disable some of them?

Quoted 1 line

[Anoob] Yes. There are few other flags in addition to the above. Like

RTE_CRYPTODEV_FF_ASYMMETRIC_CRYPTO.

Quoted 4 lines
Quoted 2 lines

If so, it would be better to call it ff_disable.

[Anoob] Calling it ff_enable is to make it similar to how it's done with eth

devices. Either way should be fine.
[Shally] keeping it as "ff_enable" will require application to set these flags to configure PMD. If we define it other way around, then it would be mean to mask out unwanted features which can be quite many.
Though purpose can be achieved both ways but keeping it as "enable" looks more easy to use, readable and inline with how ethdev handle multi feature support.
So I would prefer to keep it as "ff_enable"

Thanks
Shally

Quoted 10 lines
Quoted 9 lines

And to limit it to the subset of features that it makes sense to disable. i.e. why would an application disable chaining or any of the SGL flags? It would just add extra cycles in the PMD to error check on these cases, instead the appl can just not send such commands. And it doesn't make sense to disable AESNI or RTE_CRYPTODEV_FF_HW_ACCELERATED.
Actually I can't see what an application would want to achieve by disabling any flag?

Quoted 3 lines

[Anoob] Features like ASYMMETRIC or SECURITY is not required for every application. SECURITY is added for eth devices also. But since the feature can be disabled while configuring the port, applications are given a

choice to disable it. That way apps like l2fwd doesn't enable SECURITY. With crypto this option is not available.

Quoted 2 lines

If the unused offloads can be communicated to the PMD before initialization,

the PMD will be allowed to optimize hardware usage.

Quoted 1 line

Otherwise, supporting more features would affect performance, even if

application is not making use of those features.

Quoted 2 lines

Ex: test-crypto-perf doesn't use ASYM/SECURITY. Now adding these features

would affect the performance of this app.

Acked-by on v1 1/1
#5 lore raw
In reply to Anoob Joseph
Quoted 16 lines · Anoob Joseph

Add new field ff_enable in rte_cryptodev_config. This enables applications to control the features enabled on the crypto device.

Proposed new layout:

/** Crypto device configuration structure */

struct rte_cryptodev_config {
     int socket_id;            /**< Socket to allocate resources on */
     uint16_t nb_queue_pairs;
     /**< Number of queue pairs to configure on device */
(quoted diff) +4 -0
+   uint64_t ff_enable;
+   /**< Feature flags to be enabled on the device. Only the features set
+    * on rte_cryptodev_info.feature_flags are allowed to be set.
+    */

};

Quoted 7 lines

For eth devices, rte_eth_conf.rx_mode.offloads and rte_eth_conf.tx_mode.offloads fields are used by applications to control the offloads enabled on the eth device. This proposal adds a similar ability for the crypto device.

#6 lore raw

There is only one ack for this change.
A deprecation requires more agreement (3 valuable acks). Other opinions?

31/01/2019 10:53, Akhil Goyal:

Quoted 25 lines · Anoob Joseph
Quoted 16 lines

Add new field ff_enable in rte_cryptodev_config. This enables applications to control the features enabled on the crypto device.

Proposed new layout:

/** Crypto device configuration structure */

struct rte_cryptodev_config {
     int socket_id;            /**< Socket to allocate resources on */
     uint16_t nb_queue_pairs;
     /**< Number of queue pairs to configure on device */
(quoted diff) +4 -0
+   uint64_t ff_enable;
+   /**< Feature flags to be enabled on the device. Only the features set
+    * on rte_cryptodev_info.feature_flags are allowed to be set.
+    */

};

Quoted 7 lines

For eth devices, rte_eth_conf.rx_mode.offloads and rte_eth_conf.tx_mode.offloads fields are used by applications to control the offloads enabled on the eth device. This proposal adds a similar ability for the crypto device.

#7 lore raw

Hi Thomas, Akhil, Anoob,

Quoted 10 lines

-----Original Message-----
From: Thomas Monjalon [mailto:[email protected]]
Sent: Friday, February 1, 2019 11:14 AM
To: [email protected]
Cc: Akhil Goyal <[email protected]>; Anoob Joseph <[email protected]>; De Lara Guarch, Pablo <[email protected]>; Trahe, Fiona <[email protected]>; Jerin Jacob Kollanukkaran <[email protected]>; Narayana Prasad Raju Athreya <[email protected]>; Shally Verma <[email protected]>; Doherty, Declan <[email protected]>
Subject: Re: [dpdk-dev] [PATCH] doc: announce ABI change for cryptodev config

Quoted 31 lines

There is only one ack for this change.
A deprecation requires more agreement (3 valuable acks). Other opinions?

31/01/2019 10:53, Akhil Goyal:

Quoted 25 lines · Anoob Joseph
Quoted 16 lines

Add new field ff_enable in rte_cryptodev_config. This enables applications to control the features enabled on the crypto device.

Proposed new layout:

/** Crypto device configuration structure */

struct rte_cryptodev_config {
     int socket_id;            /**< Socket to allocate resources on */
     uint16_t nb_queue_pairs;
     /**< Number of queue pairs to configure on device */
(quoted diff) +4 -0
+   uint64_t ff_enable;
+   /**< Feature flags to be enabled on the device. Only the features set
+    * on rte_cryptodev_info.feature_flags are allowed to be set.
+    */

};

Quoted 7 lines

For eth devices, rte_eth_conf.rx_mode.offloads and rte_eth_conf.tx_mode.offloads fields are used by applications to control the offloads enabled on the eth device. This proposal adds a similar ability for the crypto device.

[Fiona] Keeping consistent with ethdev is a lower priority that adding a
param that works well for crypto. As proposed this ff_enable is problematic for crypto
as it makes no sense to enable/disable most of the flags.
For some there's no sensible action a PMD can take, e.g. RTE_CRYPTODEV_FF_HW_ACCELERATED.
For some, PMDs would need to add performance impacting forks on the data path to check for disabled features. E.g. if an applic disables the RTE_CRYPTODEV_FF_CPU_AESNI flag, what does it expect the PMD to do? Not use the aesni capability of the CPU? Fork and do a less performant implementation? Or if this flag is set, RTE_CRYPTODEV_FF_OOP_LB_IN_LB_OUT or RTE_CRYPTODEV_FF_SYM_OPERATION_CHAINING, does the PMD need to check for operations like these and reject them?
It seems there are only 3 of the 16 flags that it's desirable to disable, based on the earlier thread
RTE_CRYPTODEV_FF_ASYMMETRIC_CRYPTO
RTE_CRYPTODEV_FF_SYMMETRIC_CRYPTO
RTE_CRYPTODEV_FF_SECURITY
So I propose this param should be called ff_disable.
It should documented - and maybe provide a mask for - the flags the application is allowed to disable - only the above 3. Then PMDs only need to implement enable/disable functionality for that subset of flags.

#8 lore raw

Hi Anoob,

Quoted 43 lines · Fiona

Hi Thomas, Akhil, Anoob,

Quoted 10 lines

-----Original Message-----
From: Thomas Monjalon [mailto:[email protected]]
Sent: Friday, February 1, 2019 11:14 AM
To: [email protected]
Cc: Akhil Goyal <[email protected]>; Anoob Joseph <[email protected]>; De Lara Guarch, Pablo <[email protected]>; Trahe, Fiona <[email protected]>; Jerin Jacob Kollanukkaran <[email protected]>; Narayana Prasad Raju Athreya <[email protected]>; Shally Verma <[email protected]>; Doherty, Declan <[email protected]>
Subject: Re: [dpdk-dev] [PATCH] doc: announce ABI change for cryptodev config

Quoted 31 lines

There is only one ack for this change.
A deprecation requires more agreement (3 valuable acks). Other opinions?

31/01/2019 10:53, Akhil Goyal:

Quoted 25 lines · Anoob Joseph
Quoted 16 lines

Add new field ff_enable in rte_cryptodev_config. This enables applications to control the features enabled on the crypto device.

Proposed new layout:

/** Crypto device configuration structure */

struct rte_cryptodev_config {
      int socket_id;            /**< Socket to allocate resources on */
      uint16_t nb_queue_pairs;
      /**< Number of queue pairs to configure on device */
(quoted diff) +4 -0
+   uint64_t ff_enable;
+   /**< Feature flags to be enabled on the device. Only the features set
+    * on rte_cryptodev_info.feature_flags are allowed to be set.
+    */

};

Quoted 7 lines

For eth devices, rte_eth_conf.rx_mode.offloads and rte_eth_conf.tx_mode.offloads fields are used by applications to control the offloads enabled on the eth device. This proposal adds a similar ability for the crypto device.

Quoted 12 lines

[Fiona] Keeping consistent with ethdev is a lower priority that adding a
param that works well for crypto. As proposed this ff_enable is problematic for crypto
as it makes no sense to enable/disable most of the flags.
For some there's no sensible action a PMD can take, e.g. RTE_CRYPTODEV_FF_HW_ACCELERATED.
For some, PMDs would need to add performance impacting forks on the data path to check for disabled features. E.g. if an applic disables the RTE_CRYPTODEV_FF_CPU_AESNI flag, what does it expect the PMD to do? Not use the aesni capability of the CPU? Fork and do a less performant implementation? Or if this flag is set, RTE_CRYPTODEV_FF_OOP_LB_IN_LB_OUT or RTE_CRYPTODEV_FF_SYM_OPERATION_CHAINING, does the PMD need to check for operations like these and reject them?
It seems there are only 3 of the 16 flags that it's desirable to disable, based on the earlier thread
RTE_CRYPTODEV_FF_ASYMMETRIC_CRYPTO
RTE_CRYPTODEV_FF_SYMMETRIC_CRYPTO
RTE_CRYPTODEV_FF_SECURITY
So I propose this param should be called ff_disable.
It should documented - and maybe provide a mask for - the flags the application is allowed to disable - only the above 3. Then PMDs only need to implement enable/disable functionality for that subset of flags.

could you send a new version of this patch as per the comments from Fiona.

Thanks,
Akhil

#9 lore raw

Hi Akhil,

I'll send a v2 incorporating Fiona's comments.

Thanks,
Anoob

Quoted 93 lines

-----Original Message-----
From: Akhil Goyal <[email protected]>
Sent: Thursday, February 28, 2019 3:32 PM
To: Trahe, Fiona <[email protected]>; Thomas Monjalon <[email protected]>; [email protected]
Cc: Anoob Joseph <[email protected]>; De Lara Guarch, Pablo <[email protected]>; Jerin Jacob Kollanukkaran <[email protected]>; Narayana Prasad Raju Athreya <[email protected]>; Shally Verma <[email protected]>; Doherty, Declan <[email protected]>
Subject: Re: [dpdk-dev] [PATCH] doc: announce ABI change for cryptodev config

Hi Anoob,

Quoted 12 lines · Fiona

Hi Thomas, Akhil, Anoob,

Quoted 10 lines

-----Original Message-----
From: Thomas Monjalon [mailto:[email protected]]
Sent: Friday, February 1, 2019 11:14 AM
To: [email protected]
Cc: Akhil Goyal <[email protected]>; Anoob Joseph <[email protected]>; De Lara Guarch, Pablo <[email protected]>; Trahe, Fiona <[email protected]>; Jerin Jacob Kollanukkaran <[email protected]>; Narayana Prasad Raju Athreya <[email protected]>; Shally Verma <[email protected]>;

Doherty,

Quoted 16 lines
Quoted 16 lines

Declan <[email protected]>
Subject: Re: [dpdk-dev] [PATCH] doc: announce ABI change for cryptodev config

There is only one ack for this change.
A deprecation requires more agreement (3 valuable acks). Other opinions?

31/01/2019 10:53, Akhil Goyal:

Quoted 6 lines · Anoob Joseph
Quoted 5 lines

Add new field ff_enable in rte_cryptodev_config. This enables applications to control the features enabled on the crypto device.

Proposed new layout:

Quoted 7 lines
Quoted 7 lines
Quoted 7 lines

Review

Quoted 7 lines

/** Crypto device configuration structure */ struct

rte_cryptodev_config {
      int socket_id;            /**< Socket to allocate resources on */
      uint16_t nb_queue_pairs;
      /**< Number of queue pairs to configure on device */
(quoted diff) +2 -0
+   uint64_t ff_enable;
+   /**< Feature flags to be enabled on the device. Only the features

set

Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
Quoted 9 lines
(quoted diff) +2 -0
+    * on rte_cryptodev_info.feature_flags are allowed to be set.
+    */

};

For eth devices, rte_eth_conf.rx_mode.offloads and rte_eth_conf.tx_mode.offloads fields are used by applications to control the offloads enabled on the eth device. This proposal adds a similar ability for the crypto device.

Quoted 6 lines
Quoted 3 lines
Quoted 3 lines
Quoted 2 lines

[Fiona] Keeping consistent with ethdev is a lower priority that adding a param that works well for crypto. As proposed this ff_enable is problematic for crypto as it makes no sense to enable/disable most of the

flags.

Quoted 1 line

For some there's no sensible action a PMD can take, e.g.

RTE_CRYPTODEV_FF_HW_ACCELERATED.

Quoted 1 line

For some, PMDs would need to add performance impacting forks on the

data path to check for disabled features. E.g. if an applic disables the RTE_CRYPTODEV_FF_CPU_AESNI flag, what does it expect the PMD to do? Not use the aesni capability of the CPU? Fork and do a less performant implementation?

Quoted 1 line

Or if this flag is set, RTE_CRYPTODEV_FF_OOP_LB_IN_LB_OUT or

RTE_CRYPTODEV_FF_SYM_OPERATION_CHAINING, does the PMD need to check for operations like these and reject them?

Quoted 7 lines

It seems there are only 3 of the 16 flags that it's desirable to disable, based on the earlier thread RTE_CRYPTODEV_FF_ASYMMETRIC_CRYPTO RTE_CRYPTODEV_FF_SYMMETRIC_CRYPTO RTE_CRYPTODEV_FF_SECURITY
So I propose this param should be called ff_disable.
It should documented - and maybe provide a mask for - the flags the

application is allowed to disable - only the above 3. Then PMDs only need to implement enable/disable functionality for that subset of flags.

could you send a new version of this patch as per the comments from Fiona.

Thanks,
Akhil

#10 lore raw

PATCH v2 1/1 doc: announce ABI change for cryptodev config

In reply to Anoob Joseph

Add new field ff_disable in rte_cryptodev_config. This enables applications to control the features enabled on the crypto device.

Proposed new layout:

/** Crypto device configuration structure */

struct rte_cryptodev_config {
    int socket_id;            /**< Socket to allocate resources on */
    uint16_t nb_queue_pairs;
    /**< Number of queue pairs to configure on device */
+   uint64_t ff_disable;
+   /**< Feature flags to be disabled. Only the following features are
+    * allowed to be disabled,
+    *  - RTE_CRYPTODEV_FF_SYMMETRIC_CRYPTO
+    *  - RTE_CRYPTODEV_FF_ASYMMETRIC_CRYPTO
+    *  - RTE_CRYTPODEV_FF_SECURITY
+    */

};

For eth devices, rte_eth_conf.rx_mode.offloads and rte_eth_conf.tx_mode.offloads fields are used by applications to control the offloads enabled on the eth device. This proposal adds a similar ability for the crypto device.

---
v2:
* Renamed 'ff_enable' to 'ff_disable'

 doc/guides/rel_notes/deprecation.rst | 14 ++++++++++++++
 1 file changed, 14 insertions(+)
doc/guides/rel_notes/deprecation.rst +14 -0
index 1b4fcb7..d0a60f9 100644
@@ -75,3 +75,17 @@ Deprecation Notices
 
 * crypto/aesni_mb: the minimum supported intel-ipsec-mb library version will be
   changed from 0.49.0 to 0.52.0.
+
+* cryptodev: New member in ``rte_cryptodev_config`` to allow applications to
+  disable features supported by the crypto device. Only the following features
+  would be allowed to be disabled this way,
+
+  - ``RTE_CRYPTODEV_FF_SYMMETRIC_CRYPTO``
+  - ``RTE_CRYPTODEV_FF_ASYMMETRIC_CRYPTO``
+  - ``RTE_CRYPTODEV_FF_SECURITY``
+
+  Disabling unused features would facilitate efficient usage of HW/SW offload.
+
+  - Member ``uint64_t ff_disable`` in ``rte_cryptodev_config``
+
+  The field would be added in v19.08.
Signature · 1 line
2.7.4
Acked-by on v2 1/1
#11 lore raw
Quoted 31 lines

-----Original Message-----
From: Anoob Joseph [mailto:[email protected]]
Sent: Thursday, March 7, 2019 10:40 AM
To: Akhil Goyal <[email protected]>; Trahe, Fiona <[email protected]>; De Lara Guarch, Pablo <[email protected]>
Cc: Anoob Joseph <[email protected]>; Jerin Jacob Kollanukkaran <[email protected]>; Narayana Prasad Raju Athreya <[email protected]>; Shally Verma <[email protected]>; Suheil Chandran <[email protected]>; [email protected]
Subject: [PATCH v2] doc: announce ABI change for cryptodev config

Add new field ff_disable in rte_cryptodev_config. This enables applications to control the features enabled on the crypto device.

Proposed new layout:

/** Crypto device configuration structure */

struct rte_cryptodev_config {
    int socket_id;            /**< Socket to allocate resources on */
    uint16_t nb_queue_pairs;
    /**< Number of queue pairs to configure on device */
(quoted diff) +7 -0
+   uint64_t ff_disable;
+   /**< Feature flags to be disabled. Only the following features are
+    * allowed to be disabled,
+    *  - RTE_CRYPTODEV_FF_SYMMETRIC_CRYPTO
+    *  - RTE_CRYPTODEV_FF_ASYMMETRIC_CRYPTO
+    *  - RTE_CRYTPODEV_FF_SECURITY
+    */

};

Quoted 6 lines

For eth devices, rte_eth_conf.rx_mode.offloads and rte_eth_conf.tx_mode.offloads fields are used by applications to control the offloads enabled on the eth device. This proposal adds a similar ability for the crypto device.

Acked-by Applied
#12 lore raw
Quoted 30 lines · Fiona
Quoted 30 lines

-----Original Message-----
From: Anoob Joseph [mailto:[email protected]]
Sent: Thursday, March 7, 2019 10:40 AM
To: Akhil Goyal <[email protected]>; Trahe, Fiona <[email protected]>; De Lara Guarch, Pablo <[email protected]>
Cc: Anoob Joseph <[email protected]>; Jerin Jacob Kollanukkaran <[email protected]>; Narayana Prasad Raju Athreya <[email protected]>; Shally Verma <[email protected]>; Suheil Chandran <[email protected]>; [email protected]
Subject: [PATCH v2] doc: announce ABI change for cryptodev config

Add new field ff_disable in rte_cryptodev_config. This enables applications to control the features enabled on the crypto device.

Proposed new layout:

/** Crypto device configuration structure */

struct rte_cryptodev_config {
     int socket_id;            /**< Socket to allocate resources on */
     uint16_t nb_queue_pairs;
     /**< Number of queue pairs to configure on device */
(quoted diff) +7 -0
+   uint64_t ff_disable;
+   /**< Feature flags to be disabled. Only the following features are
+    * allowed to be disabled,
+    *  - RTE_CRYPTODEV_FF_SYMMETRIC_CRYPTO
+    *  - RTE_CRYPTODEV_FF_ASYMMETRIC_CRYPTO
+    *  - RTE_CRYTPODEV_FF_SECURITY
+    */

};

Quoted 8 lines
Quoted 6 lines

For eth devices, rte_eth_conf.rx_mode.offloads and rte_eth_conf.tx_mode.offloads fields are used by applications to control the offloads enabled on the eth device. This proposal adds a similar ability for the crypto device.

Applied to dpdk-next-crypto

Thanks.