security: fix crash at accessing non-implemented ops
PATCH 1/1 security: fix crash at accessing non-implemented ops
Valid checks for optional function pointers inside dev-ops were disabled by undefined macro.
--- lib/librte_security/rte_security.c | 4 ---- 1 file changed, 4 deletions(-)
lib/librte_security/rte_security.c +0 -4
@@ -107,11 +107,9 @@ rte_security_set_pkt_metadata(struct rte_security_ctx *instance,struct rte_security_session *sess, struct rte_mbuf *m, void *params) { -#ifdef RTE_DEBUG RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, set_pkt_metadata, -EINVAL, -ENOTSUP);RTE_PTR_OR_ERR_RET (sess, -EINVAL ); -#endif return instance->ops->set_pkt_metadata (instance->device, sess, m, params); }
@@ -121,9 +119,7 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md) {void *userdata =NULL ; -#ifdef RTE_DEBUG RTE_PTR_CHAIN3_OR_ERR_RET (instance, ops, get_userdata,NULL ,NULL ); -#endif if (instance->ops->get_userdata (instance->device, md, &userdata))return NULL ;
Signature · 1 line
2.17.1
Actually looking at app/test/test_security.c
I also see a few '#ifdef RTE_DEBUG's.
Let say:
+static int
+test_get_userdata_inv_context(void)
+{
+#ifdef RTE_DEBUG
+ uint64_t md = 0xDEADBEEF;
+
+ void *ret = rte_security_get_userdata(NULL, md);
+ TEST_ASSERT_MOCK_FUNCTION_CALL_RET(rte_security_get_userdata,
+ ret, NULL, "%p");
+ TEST_ASSERT_MOCK_CALLS(mock_get_userdata_exp, 0);
+
+ return TEST_SUCCESS;
+#else
+ return TEST_SKIPPED;
+#endif
+}
What is the point?
Why not always run the test unconditionally?
Quoted 45 lines
-----Original Message-----
From: Ananyev, Konstantin <[email protected]>
Sent: Thursday, April 23, 2020 12:52 AM
To: [email protected]
Cc: [email protected]; Doherty, Declan <[email protected]>; Ananyev, Konstantin <[email protected]>
Subject: [PATCH] security: fix crash at accessing non-implemented opsValid checks for optional function pointers inside dev-ops were disabled by undefined macro.
Fixes b6ee98547847 ("security: fix verification of parameters")Signed-off-by Konstantin Ananyev <[email protected]>--- lib/librte_security/rte_security.c | 4 ---- 1 file changed, 4 deletions(-)lib/librte_security/rte_security.c +0
-5@@ -107,11 +107,9 @@ rte_security_set_pkt_metadata(struct rte_security_ctx *instance,struct rte_security_session *sess, struct rte_mbuf *m, void *params) { -#ifdef RTE_DEBUG RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, set_pkt_metadata, -EINVAL, -ENOTSUP);RTE_PTR_OR_ERR_RET (sess, -EINVAL ); -#endif return instance->ops->set_pkt_metadata (instance->device, sess, m, params); }@@ -121,9 +119,7 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md) {void *userdata =NULL ; -#ifdef RTE_DEBUG RTE_PTR_CHAIN3_OR_ERR_RET (instance, ops, get_userdata,NULL ,NULL ); -#endif if (instance->ops->get_userdata (instance->device, md, &userdata))return NULL ; --2.17.1
W dniu 23.04.2020 o 02:11, Ananyev, Konstantin pisze:
Quoted 21 lines
Actually looking at
app/test/test_security.c
I also see a few '#ifdef RTE_DEBUG's.
Let say:(quoted diff) +16
-0+static int +test_get_userdata_inv_context(void) +{ +#ifdef RTE_DEBUG + uint64_t md = 0xDEADBEEF; + + void *ret = rte_security_get_userdata(NULL, md); + TEST_ASSERT_MOCK_FUNCTION_CALL_RET(rte_security_get_userdata, + ret, NULL, "%p"); + TEST_ASSERT_MOCK_CALLS(mock_get_userdata_exp, 0); + + return TEST_SUCCESS; +#else + return TEST_SKIPPED; +#endif +}
Quoted 2 lines
What is the point?
Why not always run the test unconditionally?
If there is no RTE_DEBUG defined, the tested functionality is not compiled, so the tests won't work.
They must be wrapped with same macro as library code.
Quoted 47 lines
Quoted 47 lines
-----Original Message-----
From: Ananyev, Konstantin <[email protected]>
Sent: Thursday, April 23, 2020 12:52 AM
To: [email protected]
Cc: [email protected]; Doherty, Declan <[email protected]>; Ananyev, Konstantin <[email protected]>
Subject: [PATCH] security: fix crash at accessing non-implemented opsValid checks for optional function pointers inside dev-ops were disabled by undefined macro.
Fixes b6ee98547847 ("security: fix verification of parameters")Signed-off-by Konstantin Ananyev <[email protected]>--- lib/librte_security/rte_security.c | 4 ---- 1 file changed, 4 deletions(-)lib/librte_security/rte_security.c +0
-5@@ -107,11 +107,9 @@ rte_security_set_pkt_metadata(struct rte_security_ctx *instance,struct rte_security_session *sess, struct rte_mbuf *m, void *params) { -#ifdef RTE_DEBUG RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, set_pkt_metadata, -EINVAL, -ENOTSUP);RTE_PTR_OR_ERR_RET (sess, -EINVAL ); -#endif return instance->ops->set_pkt_metadata (instance->device, sess, m, params); }@@ -121,9 +119,7 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md) {void *userdata =NULL ; -#ifdef RTE_DEBUG RTE_PTR_CHAIN3_OR_ERR_RET (instance, ops, get_userdata,NULL ,NULL ); -#endif if (instance->ops->get_userdata (instance->device, md, &userdata))return NULL ; --2.17.1
Signature · 7 lines
Lukasz Wojciechowski Principal Software Engineer Samsung R&D Institute Poland Samsung Electronics Office +48 22 377 88 25 [email protected]
Hi Konstantin,
These are data path ops and so it will be better if we can avoid such checks in the datapath. The same is done in ethdev also.
http://code.dpdk.org/dpdk/v20.02/source/lib/librte_ethdev/rte_ethdev.h#L4372
Datapath functions in cryptodev (enqueue/dequeue) doesn't even have such checks.
http://code.dpdk.org/dpdk/v20.02/source/lib/librte_cryptodev/rte_cryptodev.h#L962
Thanks,
Anoob
Quoted 48 lines
-----Original Message-----
From: dev <[email protected]> On Behalf Of Konstantin Ananyev
Sent: Thursday, April 23, 2020 5:22 AM
To: [email protected]
Cc: [email protected]; [email protected]; Konstantin Ananyev <[email protected]>
Subject: [dpdk-dev] [PATCH] security: fix crash at accessing non-implemented opsValid checks for optional function pointers inside dev-ops were disabled by undefined macro.
Fixes b6ee98547847 ("security: fix verification of parameters")Signed-off-by Konstantin Ananyev <[email protected]>--- lib/librte_security/rte_security.c | 4 ---- 1 file changed, 4 deletions(-)lib/librte_security/rte_security.c +0
-0@@ -107,11 +107,9 @@ rte_security_set_pkt_metadata(struct rte_security_ctx *instance, struct rte_security_session *sess, struct rte_mbuf *m, void *params) { -#ifdef
RTE_DEBUG
RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops,set_pkt_metadata, -
EINVAL,lib/librte_security/rte_security.c +0
-2-ENOTSUP);RTE_PTR_OR_ERR_RET (sess, -EINVAL ); -#endif return instance->ops->set_pkt_metadata (instance->device, sess, m, params); }@@ -121,9 +119,7 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md) {void *userdata =NULL ; -#ifdef RTE_DEBUG RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, get_userdata, NULL,NULL); -#endif
lib/librte_security/rte_security.c +0
-1if (instance->ops->get_userdata (instance->device, md, &userdata))return NULL ; --2.17.1
Quoted 4 lines
Hi Konstantin,
These are data path ops and so it will be better if we can avoid such checks in the datapath. The same is done in ethdev also.
AFAIK, get_userdata is an *optional* dev-ops function that can be used by data-path. So far there was no strict requirement for the rte_security PMDs to *always* implement it. So what you guys did is a silent change of public API behaviour.
As result ixgbe, (and probably some others rte_security PMDs) stopped working properly. I don't see any point in these changes, but if you'd like to do that, at least our usual procedure has to be followed:
1. Send and RFC to get an agreement with rte_security PMDs maintainers (one release ahead)
2. send a deprecation note (one release ahead)
3. change the behaviour of the public API
4. update release notes
AFAIK 1), 2), 4) wasn't done.
So I think right now we need to revert original behaviour.
Quoted 5 lines
http://code.dpdk.org/dpdk/v20.02/source/lib/librte_ethdev/rte_ethdev.h#L4372
Datapath functions in cryptodev (enqueue/dequeue) doesn't even have such checks.
http://code.dpdk.org/dpdk/v20.02/source/lib/librte_cryptodev/rte_cryptodev.h#L962
That's a different story:
rx_burst/tx_burst, enqueue/dequeue are mandatory dev-ops functions that have to be implemented by each ethdev/cryptodev API.
Quoted 53 lines
Thanks,
AnoobQuoted 48 lines
-----Original Message-----
From: dev <[email protected]> On Behalf Of Konstantin Ananyev
Sent: Thursday, April 23, 2020 5:22 AM
To: [email protected]
Cc: [email protected]; [email protected]; Konstantin Ananyev <[email protected]>
Subject: [dpdk-dev] [PATCH] security: fix crash at accessing non-implemented opsValid checks for optional function pointers inside dev-ops were disabled by undefined macro.
Fixes b6ee98547847 ("security: fix verification of parameters")Signed-off-by Konstantin Ananyev <[email protected]>--- lib/librte_security/rte_security.c | 4 ---- 1 file changed, 4 deletions(-)lib/librte_security/rte_security.c +0
-0@@ -107,11 +107,9 @@ rte_security_set_pkt_metadata(struct rte_security_ctx *instance, struct rte_security_session *sess, struct rte_mbuf *m, void *params) { -#ifdef
RTE_DEBUG
RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops,set_pkt_metadata, -
EINVAL,lib/librte_security/rte_security.c +0
-2-ENOTSUP);RTE_PTR_OR_ERR_RET (sess, -EINVAL ); -#endif return instance->ops->set_pkt_metadata (instance->device, sess, m, params); }@@ -121,9 +119,7 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md) {void *userdata =NULL ; -#ifdef RTE_DEBUG RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, get_userdata, NULL,NULL); -#endif
lib/librte_security/rte_security.c +0
-1if (instance->ops->get_userdata (instance->device, md, &userdata))return NULL ; --2.17.1
W dniu 23.04.2020 o 09:54, Ananyev, Konstantin pisze:
Quoted 14 lines
Quoted 3 lines
Hi Konstantin,
These are data path ops and so it will be better if we can avoid such checks in the datapath. The same is done in ethdev also.
AFAIK,
get_userdatais an *optional* dev-ops function that can be used by data-path. So far there was no strict requirement for therte_securityPMDs to *always* implement it. So what you guys did is a silent change of public API behaviour.
As result ixgbe, (and probably some othersrte_securityPMDs) stopped working properly. I don't see any point in these changes, but if you'd like to do that, at least our usual procedure has to be followed:
1. Send and RFC to get an agreement withrte_securityPMDs maintainers (one release ahead)
2. send a deprecation note (one release ahead)
3. change the behaviour of the public API
4. update release notes
Quoted 2 lines
AFAIK 1), 2), 4) wasn't done.
So I think right now we need to revert original behaviour.
The current changes were made in patch: b6ee9854784 security: fix verification of parameters
<snip>
@@ -91,7 +119,9 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md)
{
void *userdata = NULL;
- RTE_FUNC_PTR_OR_ERR_RET(*instance->ops->get_userdata, NULL); +#ifdef RTE_DEBUG
+ RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, get_userdata, NULL, NULL); +#endif
if (instance->ops->get_userdata(instance->device, md, &userdata)) return NULL;
<snip>
So as you can see, the checks were already there. They've just been wrapped up with RTE_DEBUG macro for disabling them in non-debug compilation mode and the validation of paramter was change to avoid possible segmentation fault if instance lub ops would be NULL
Quoted 61 lines
Quoted 4 lines
Datapath functions in cryptodev (enqueue/dequeue) doesn't even have such checks.
https://protect2.fireeye.com/url?k=79d7974a-244939d0-79d61c05-0cc47a336fae-19f540008a9467cf&q=1&u=http%3A%2F%2Fcode.dpdk.org%2Fdpdk%2Fv20.02%2Fsource%2Flib%2Flibrte_cryptodev%2Frte_cryptodev.h%23L962That's a different story:
rx_burst/tx_burst, enqueue/dequeue are mandatory dev-ops functions that have to be implemented by each ethdev/cryptodev API.Quoted 52 lines
Thanks,
AnoobQuoted 48 lines
-----Original Message-----
From: dev <[email protected]> On Behalf Of Konstantin Ananyev
Sent: Thursday, April 23, 2020 5:22 AM
To: [email protected]
Cc: [email protected]; [email protected]; Konstantin Ananyev <[email protected]>
Subject: [dpdk-dev] [PATCH] security: fix crash at accessing non-implemented opsValid checks for optional function pointers inside dev-ops were disabled by undefined macro.
Fixes b6ee98547847 ("security: fix verification of parameters")Signed-off-by Konstantin Ananyev <[email protected]>--- lib/librte_security/rte_security.c | 4 ---- 1 file changed, 4 deletions(-)lib/librte_security/rte_security.c +0
-0@@ -107,11 +107,9 @@ rte_security_set_pkt_metadata(struct rte_security_ctx *instance, struct rte_security_session *sess, struct rte_mbuf *m, void *params) { -#ifdef
RTE_DEBUG
RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops,set_pkt_metadata, -
EINVAL,
-ENOTSUP);
RTE_PTR_OR_ERR_RET(sess, -EINVAL);lib/librte_security/rte_security.c +0
-2-#endif return instance->ops->set_pkt_metadata (instance->device, sess, m, params); }@@ -121,9 +119,7 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md) {void *userdata =NULL ; -#ifdef RTE_DEBUG RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, get_userdata, NULL,NULL); -#endif
if (instance->ops->get_userdata(instance->device, md, &userdata))
return NULL;--
2.17.1
Signature · 7 lines
Lukasz Wojciechowski Principal Software Engineer Samsung R&D Institute Poland Samsung Electronics Office +48 22 377 88 25 [email protected]
Quoted 124 lines
-----Original Message-----
From: Lukasz Wojciechowski <[email protected]>
Sent: Thursday, April 23, 2020 9:06 AM
To: Ananyev, Konstantin <[email protected]>; Anoob Joseph <[email protected]>; [email protected]
Cc: [email protected]; Doherty, Declan <[email protected]>; [email protected]
Subject: Re: [dpdk-dev] [PATCH] security: fix crash at accessing non-implemented opsW dniu 23.04.2020 o 09:54, Ananyev, Konstantin pisze:
Quoted 14 lines
Quoted 3 lines
Hi Konstantin,
These are data path ops and so it will be better if we can avoid such checks in the datapath. The same is done in ethdev also.
AFAIK,
get_userdatais an *optional* dev-ops function that can be used by data-path. So far there was no strict requirement for therte_securityPMDs to *always* implement it. So what you guys did is a silent change of public API behaviour.
As result ixgbe, (and probably some othersrte_securityPMDs) stopped working properly. I don't see any point in these changes, but if you'd like to do that, at least our usual procedure has to be followed:
1. Send and RFC to get an agreement withrte_securityPMDs maintainers (one release ahead)
2. send a deprecation note (one release ahead)
3. change the behaviour of the public API
4. update release notesQuoted 2 lines
AFAIK 1), 2), 4) wasn't done.
So I think right now we need to revert original behaviour.The current changes were made in patch:
b6ee9854784security: fix verification of parameters<snip>
lib/librte_security/rte_security.c +3
-1@@ -91,7 +119,9 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md) { void *userdata = NULL; -RTE_FUNC_PTR_OR_ERR_RET (*instance->ops->get_userdata,NULL ); +#ifdef RTE_DEBUG + RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, get_userdata, NULL, NULL); +#endif if (instance->ops->get_userdata(instance->device, md, &userdata)) return NULL;So as you can see, the checks were already there. They've just been wrapped up with
RTE_DEBUGmacro for disabling them in non-debug compilation mode and the validation of paramter was change to avoid possible segmentation fault if instance lub ops would be NULLQuoted 1 line
55cc35a7b94c97c0&q=1&u=http%3A%2F%2Fcode.dpdk.org%2Fdpdk%2Fv20.02%2Fsource%2Flib%2Flibrte_ethdev%2Frte_ethdev.h%23L43 72
Quoted 3 lines
Quoted 3 lines
Datapath functions in cryptodev (enqueue/dequeue) doesn't even have such checks.
https://protect2.fireeye.com/url?k=79d7974a-244939d0-79d61c05-0cc47a336fae-19f540008a9467cf&q=1&u=http%3A%2F%2Fcode.dpdk.org%2Fdpdk%2Fv20.02%2Fsource%2Flib%2Flibrte_cryptodev%2Frte_cryptodev.h% 23L962
Quoted 48 lines
That's a different story:
rx_burst/tx_burst, enqueue/dequeue are mandatory dev-ops functions that have to be implemented by each ethdev/cryptodev API.Quoted 44 lines
Thanks,
AnoobQuoted 40 lines
-----Original Message-----
From: dev <[email protected]> On Behalf Of Konstantin Ananyev
Sent: Thursday, April 23, 2020 5:22 AM
To: [email protected]
Cc: [email protected]; [email protected]; Konstantin Ananyev <[email protected]>
Subject: [dpdk-dev] [PATCH] security: fix crash at accessing non-implemented opsValid checks for optional function pointers inside dev-ops were disabled by undefined macro.
Fixes b6ee98547847 ("security: fix verification of parameters")Signed-off-by Konstantin Ananyev <[email protected]>--- lib/librte_security/rte_security.c | 4 ---- 1 file changed, 4 deletions(-)lib/librte_security/rte_security.c +0
-0@@ -107,11 +107,9 @@ rte_security_set_pkt_metadata(struct rte_security_ctx *instance, struct rte_security_session *sess, struct rte_mbuf *m, void *params) { -#ifdef
RTE_DEBUG
RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops,set_pkt_metadata, -
EINVAL,
-ENOTSUP);
RTE_PTR_OR_ERR_RET(sess, -EINVAL);lib/librte_security/rte_security.c +0
-1-#endif return instance->ops->set_pkt_metadata (instance->device, sess, m, params); }@@ -121,9 +119,7 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md) {void *userdata =NULL ;Quoted 9 lines
Quoted 8 lines
Quoted 8 lines
lib/librte_security/rte_security.c +0
-1-#ifdef RTE_DEBUG RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, get_userdata, NULL,NULL); -#endif
if (instance->ops->get_userdata(instance->device, md, &userdata))
return NULL;--
2.17.1--
Lukasz Wojciechowski
Principal Software EngineerSamsung R&D Institute Poland
Samsung Electronics
Office +48 22 377 88 25
[email protected]
Quoted 35 lines
Quoted 15 lines
Quoted 4 lines
Hi Konstantin,
These are data path ops and so it will be better if we can avoid such checks in the datapath. The same is done in ethdev also.
AFAIK,
get_userdatais an *optional* dev-ops function that can be used by data-path. So far there was no strict requirement for therte_securityPMDs to *always* implement it. So what you guys did is a silent change of public API behaviour.
As result ixgbe, (and probably some othersrte_securityPMDs) stopped working properly. I don't see any point in these changes, but if you'd like to do that, at least our usual procedure has to be followed:
1. Send and RFC to get an agreement withrte_securityPMDs maintainers (one release ahead)
2. send a deprecation note (one release ahead)
3. change the behaviour of the public API
4. update release notesQuoted 2 lines
AFAIK 1), 2), 4) wasn't done.
So I think right now we need to revert original behaviour.The current changes were made in patch:
b6ee9854784security: fix verification of parameters<snip>
lib/librte_security/rte_security.c +3
-1@@ -91,7 +119,9 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md) { void *userdata = NULL; -RTE_FUNC_PTR_OR_ERR_RET (*instance->ops->get_userdata,NULL ); +#ifdef RTE_DEBUG + RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, get_userdata, NULL, NULL); +#endif if (instance->ops->get_userdata(instance->device, md, &userdata)) return NULL;
Quoted 5 lines
So as you can see, the checks were already there. They've just been
wrapped up withRTE_DEBUGmacro for disabling them in non-debug compilation mode and the validation of paramter was change to avoid possible segmentation fault if instance lub ops would be NULL
Sigh, that's what I am talking about:
you effectively complied out valid checks for non-debug mode. Yes, these checks have been there and they *should* stay there for *ANY* mode (both debug and non-debug).
This is an *optional* dev-ops function.
PMD has a freedom not to implement optional function.
It is a rte_security framework responsibility to check that these functions are implemented or not.
If you like to change that - the procedure described above has to be followed.
Konstantin
Quoted 75 lines
Quoted 2 lines
55cc35a7b94c97c0&q=1&u=http%3A%2F%2Fcode.dpdk.org%2Fdpdk%2Fv20.02%2Fsource%2Flib%2Flibrte_ethdev%2Frte_ethdev.h%23L43 72
Quoted 3 lines
Quoted 3 lines
Datapath functions in cryptodev (enqueue/dequeue) doesn't even have such checks.
https://protect2.fireeye.com/url?k=79d7974a-244939d0-79d61c05-0cc47a336fae-19f540008a9467cf&q=1&u=http%3A%2F%2Fcode.dpdk.org%2Fdpdk%2Fv20.02%2Fsource%2Flib%2Flibrte_cryptodev%2Frte_cryptodev.h% 23L962
Quoted 48 lines
That's a different story:
rx_burst/tx_burst, enqueue/dequeue are mandatory dev-ops functions that have to be implemented by each ethdev/cryptodev API.Quoted 44 lines
Thanks,
AnoobQuoted 40 lines
-----Original Message-----
From: dev <[email protected]> On Behalf Of Konstantin Ananyev
Sent: Thursday, April 23, 2020 5:22 AM
To: [email protected]
Cc: [email protected]; [email protected]; Konstantin Ananyev <[email protected]>
Subject: [dpdk-dev] [PATCH] security: fix crash at accessing non-implemented opsValid checks for optional function pointers inside dev-ops were disabled by undefined macro.
Fixes b6ee98547847 ("security: fix verification of parameters")Signed-off-by Konstantin Ananyev <[email protected]>--- lib/librte_security/rte_security.c | 4 ---- 1 file changed, 4 deletions(-)lib/librte_security/rte_security.c +0
-0@@ -107,11 +107,9 @@ rte_security_set_pkt_metadata(struct rte_security_ctx *instance, struct rte_security_session *sess, struct rte_mbuf *m, void *params) { -#ifdef
RTE_DEBUG
RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops,set_pkt_metadata, -
EINVAL,
-ENOTSUP);
RTE_PTR_OR_ERR_RET(sess, -EINVAL);lib/librte_security/rte_security.c +0
-1-#endif return instance->ops->set_pkt_metadata (instance->device, sess, m, params); }@@ -121,9 +119,7 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md) {void *userdata =NULL ;Quoted 9 lines
Quoted 8 lines
Quoted 8 lines
lib/librte_security/rte_security.c +0
-1-#ifdef RTE_DEBUG RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, get_userdata, NULL,NULL); -#endif
if (instance->ops->get_userdata(instance->device, md, &userdata))
return NULL;--
2.17.1--
Lukasz Wojciechowski
Principal Software EngineerSamsung R&D Institute Poland
Samsung Electronics
Office +48 22 377 88 25
[email protected]
Hi Konstantin,
Please see inline.
Thanks,
Anoob
Quoted 15 lines
-----Original Message-----
From: Ananyev, Konstantin <[email protected]>
Sent: Thursday, April 23, 2020 1:24 PM
To: Anoob Joseph <[email protected]>; [email protected]
Cc: [email protected]; Doherty, Declan <[email protected]>; [email protected]
Subject: [EXT] RE: [dpdk-dev] [PATCH] security: fix crash at accessing non- implemented opsExternal Email
----------------------------------------------------------------------
Quoted 3 lines
Hi Konstantin,
Quoted 7 lines
Quoted 1 line
These are data path ops and so it will be better if we can avoid such checks in
the datapath. The same is done in ethdev also.
AFAIK,
get_userdatais an *optional* dev-ops function that can be used by data- path.
So far there was no strict requirement for therte_securityPMDs to *always* implement it.
[Anoob] I don't think DPDK categorizes dev-ops as *optional* and *always*. If yes, can you point me?
My understanding is, all ops are optional. For example, I could implement a crypto PMD which is doing packet delivery only via event device (using crypto adapter). So dequeue op will not be implemented in that case and DPDK spec allows it.
Quoted 1 line
So what you guys did is a silent change of public API behaviour.
[Anoob] I believe Lukasz had submitted 3 or 4 revisions and it was all in the ML. RTE_DEBUG was suggested by Thomas I guess.
Quoted 2 lines
As result ixgbe, (and probably some others
rte_securityPMDs) stopped working properly.
[Anoob] set_pkt_metadata() is the only one of interest to IXGBE. And I believe the function is implemented as well. So what exactly is the concern?
Quoted 23 lines
I don't see any point in these changes, but if you'd like to do that, at least our usual procedure has to be followed:
1. Send and RFC to get an agreement withrte_securityPMDs maintainers (one release ahead) 2. send a deprecation note (one release ahead) 3. change the behaviour of the public API 4. update release notesAFAIK 1), 2), 4) wasn't done.
So I think right now we need to revert original behaviour.Quoted 4 lines
https://urldefense.proofpoint.com/v2/url?u=http-3A__code.dpdk.org_dpdk
_v20.02_source_lib_librte-5Fethdev_rte-5Fethdev.h-23L4372&d=DwIFAg&c=n KjWec2b6R0mOyPaz7xtfQ&r=jPfB8rwwviRSxyLWs2n6B-WYLn1v9SyTMrT5EQqh2TU&m=
Quoted 1 line
6ObfSanVVuHOsiqVlWxXsFWi-
2XNp76HCOX0vbUfma4&s=jDVyDDEILmgY1Yb9ZBswBVbn
Quoted 3 lines
8FpZuQI5ukH_osmtUiI&e=
Datapath functions in cryptodev (enqueue/dequeue) doesn't even have such
checks.
Quoted 3 lines
https://urldefense.proofpoint.com/v2/url?u=http-3A__code.dpdk.org_dpdk
_v20.02_source_lib_librte-5Fcryptodev_rte-5Fcryptodev.h-23L962&d=DwIFA g&c=nKjWec2b6R0mOyPaz7xtfQ&r=jPfB8rwwviRSxyLWs2n6B-
Quoted 8 lines
WYLn1v9SyTMrT5EQqh2
Quoted 1 line
TU&m=6ObfSanVVuHOsiqVlWxXsFWi-
2XNp76HCOX0vbUfma4&s=LEWQOKs0r2Im_zL95VI
Quoted 1 line
df4kQ2Pu0iRHV9Co2J1gsNBE&e=
That's a different story:
rx_burst/tx_burst, enqueue/dequeue are mandatory dev-ops functions that have to be implemented by each ethdev/cryptodev API.
[Anoob] I couldn't find any reference stating that way. If you can point me, I can update that to include datapath ops required for inline protocol processing.
Quoted 54 lines
Quoted 25 lines
Thanks,
AnoobQuoted 19 lines
-----Original Message-----
From: dev <[email protected]> On Behalf Of Konstantin Ananyev
Sent: Thursday, April 23, 2020 5:22 AM
To: [email protected]
Cc: [email protected]; [email protected]; Konstantin Ananyev <[email protected]>
Subject: [dpdk-dev] [PATCH] security: fix crash at accessing non-implemented opsValid checks for optional function pointers inside dev-ops were disabled by undefined macro.
Fixes b6ee98547847 ("security: fix verification of parameters")Signed-off-by Konstantin Ananyev <[email protected]>--- lib/librte_security/rte_security.c | 4 ---- 1 file changed, 4 deletions(-)Quoted 9 lines
lib/librte_security/rte_security.c rte_security_set_pkt_metadata(struct · line 109
Quoted 9 lines
lib/librte_security/rte_security.c +0
-0@@ -107,11 +107,9 @@ rte_security_set_pkt_metadata(structrte_security_ctx *instance, struct rte_security_session *sess, struct rte_mbuf *m, void *params) { -#ifdef
RTE_DEBUGQuoted 19 lines
Quoted 19 lines
lib/librte_security/rte_security.c +0
-1RTE_PTR_CHAIN3_OR_ERR_RET (instance, ops, set_pkt_metadata, -EINVAL , -ENOTSUP ); RTE_PTR_OR_ERR_RET(sess, -EINVAL); -#endifreturn instance->ops->set_pkt_metadata (instance->device, sess, m, params); }@@ -121,9 +119,7 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md) { void *userdata = NULL; -#ifdef RTE_DEBUG RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, get_userdata, NULL,NULL); -#endif
lib/librte_security/rte_security.c +0
-1if (instance->ops->get_userdata (instance->device, md, &userdata))return NULL ; --2.17.1
Quoted 10 lines
Quoted 9 lines
Quoted 3 lines
These are data path ops and so it will be better if we can avoid such checks in
the datapath. The same is done in ethdev also.
AFAIK,
get_userdatais an *optional* dev-ops function that can be used by data- path.
So far there was no strict requirement for therte_securityPMDs to *always* implement it.
Quoted 4 lines
[Anoob] I don't think DPDK categorizes dev-ops as *optional* and *always*. If yes, can you point me?
My understanding is, all ops are optional. For example, I could implement a crypto PMD which is doing packet delivery only via event device (using crypto adapter). So dequeue op will not be implemented in that case and DPDK spec allows it.
Your PMD can have enqueue_burst/dequeue_burst as NOP,
but you still have to provide valid function pointers:
they are stored inside crypto_dev structure itself and will be called unconditionally (without any extra checking) by rte_cryptodev_enqueue_burst/rte_cryptodev_dequeue_burst. For all other calls (both data and control path) there is a check that actual function pointer is a valid one.
Same story for eth dev: pkt_rx_burst/pkt_tx_burst and rest of dev-ops.
Quoted 3 lines
Quoted 1 line
So what you guys did is a silent change of public API behaviour.
[Anoob] I believe Lukasz had submitted 3 or 4 revisions and it was all in the ML.
RTE_DEBUGwas suggested by Thomas I guess.
I believe it is not a right procedure to change existing behaviour of rte_security framework. I think you have to communicate clear and loudly in advance (at least one release in advance). Plus RTE_DEBUG has nothing to do with changing non-debug behaviour.
Quoted 5 lines
Quoted 2 lines
As result ixgbe, (and probably some others
rte_securityPMDs) stopped working properly.[Anoob]
set_pkt_metadata()is the only one of interest to IXGBE. And I believe the function is implemented as well. So what exactly is the concern?
Check that ops->get_userdata is a valid function pointer will be compiled out. So PMDs that don't implement this function will crash in rte_security_get_userdata(). In our particular case - ixgbe.
Same story with rte_security_set_pkt_metadata() - see the patch.
Quoted 27 lines
Quoted 27 lines
I don't see any point in these changes, but if you'd like to do that, at least our usual procedure has to be followed:
1. Send and RFC to get an agreement withrte_securityPMDs maintainers (one release ahead) 2. send a deprecation note (one release ahead) 3. change the behaviour of the public API 4. update release notesAFAIK 1), 2), 4) wasn't done.
So I think right now we need to revert original behaviour.Quoted 4 lines
https://urldefense.proofpoint.com/v2/url?u=http-3A__code.dpdk.org_dpdk
_v20.02_source_lib_librte-5Fethdev_rte-5Fethdev.h-23L4372&d=DwIFAg&c=n KjWec2b6R0mOyPaz7xtfQ&r=jPfB8rwwviRSxyLWs2n6B-WYLn1v9SyTMrT5EQqh2TU&m=
Quoted 1 line
6ObfSanVVuHOsiqVlWxXsFWi-
2XNp76HCOX0vbUfma4&s=jDVyDDEILmgY1Yb9ZBswBVbn
Quoted 3 lines
8FpZuQI5ukH_osmtUiI&e=
Datapath functions in cryptodev (enqueue/dequeue) doesn't even have such
checks.
Quoted 3 lines
https://urldefense.proofpoint.com/v2/url?u=http-3A__code.dpdk.org_dpdk
_v20.02_source_lib_librte-5Fcryptodev_rte-5Fcryptodev.h-23L962&d=DwIFA g&c=nKjWec2b6R0mOyPaz7xtfQ&r=jPfB8rwwviRSxyLWs2n6B-WYLn1v9SyTMrT5EQqh2
Quoted 1 line
TU&m=6ObfSanVVuHOsiqVlWxXsFWi-
2XNp76HCOX0vbUfma4&s=LEWQOKs0r2Im_zL95VI
Quoted 8 lines
Quoted 5 lines
Quoted 1 line
df4kQ2Pu0iRHV9Co2J1gsNBE&e=
That's a different story:
rx_burst/tx_burst, enqueue/dequeue are mandatory dev-ops functions that have to be implemented by each ethdev/cryptodev API.[Anoob] I couldn't find any reference stating that way. If you can point me, I can update that to include datapath ops required for inline protocol processing.
Look at the code.
Quoted 55 lines
Quoted 55 lines
Quoted 26 lines
Thanks,
AnoobQuoted 19 lines
-----Original Message-----
From: dev <[email protected]> On Behalf Of Konstantin Ananyev
Sent: Thursday, April 23, 2020 5:22 AM
To: [email protected]
Cc: [email protected]; [email protected]; Konstantin Ananyev <[email protected]>
Subject: [dpdk-dev] [PATCH] security: fix crash at accessing non-implemented opsValid checks for optional function pointers inside dev-ops were disabled by undefined macro.
Fixes b6ee98547847 ("security: fix verification of parameters")Signed-off-by Konstantin Ananyev <[email protected]>--- lib/librte_security/rte_security.c | 4 ---- 1 file changed, 4 deletions(-)Quoted 9 lines
lib/librte_security/rte_security.c rte_security_set_pkt_metadata(struct · line 109
Quoted 9 lines
lib/librte_security/rte_security.c +0
-0@@ -107,11 +107,9 @@ rte_security_set_pkt_metadata(structrte_security_ctx *instance, struct rte_security_session *sess, struct rte_mbuf *m, void *params) { -#ifdef
RTE_DEBUGQuoted 19 lines
Quoted 19 lines
lib/librte_security/rte_security.c +0
-1RTE_PTR_CHAIN3_OR_ERR_RET (instance, ops, set_pkt_metadata, -EINVAL , -ENOTSUP ); RTE_PTR_OR_ERR_RET(sess, -EINVAL); -#endifreturn instance->ops->set_pkt_metadata (instance->device, sess, m, params); }@@ -121,9 +119,7 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md) { void *userdata = NULL; -#ifdef RTE_DEBUG RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, get_userdata, NULL,NULL); -#endif
lib/librte_security/rte_security.c +0
-1if (instance->ops->get_userdata (instance->device, md, &userdata))return NULL ; --2.17.1
Hi Konstantin,
Please see inline.
Thanks,
Anoob
Quoted 28 lines
-----Original Message-----
From: Ananyev, Konstantin <[email protected]>
Sent: Thursday, April 23, 2020 4:25 PM
To: Anoob Joseph <[email protected]>; [email protected]; Lukasz Wojciechowski <[email protected]>
Cc: [email protected]; Doherty, Declan <[email protected]>
Subject: [EXT] RE: [dpdk-dev] [PATCH] security: fix crash at accessing non- implemented opsExternal Email
----------------------------------------------------------------------
Quoted 11 lines
Quoted 9 lines
Quoted 3 lines
These are data path ops and so it will be better if we can avoid such checks in
the datapath. The same is done in ethdev also.
AFAIK,
get_userdatais an *optional* dev-ops function that can be used by data- path.
So far there was no strict requirement for therte_securityPMDs to *always* implement it.[Anoob] I don't think DPDK categorizes dev-ops as *optional* and *always*. If
yes, can you point me?
Quoted 2 lines
My understanding is, all ops are optional. For example, I could implement a crypto PMD which is doing packet delivery only via event device
Quoted 8 lines
(using crypto adapter). So dequeue op will not be implemented in that case and DPDK spec allows it.
Your PMD can have enqueue_burst/dequeue_burst as NOP, but you still have to provide valid function pointers:
they are stored insidecrypto_devstructure itself and will be called unconditionally (without any extra checking) by rte_cryptodev_enqueue_burst/rte_cryptodev_dequeue_burst.
[Anoob] I think there is a basic misunderstanding here. You are treating unconditional calls as mandatory implementations. If that is the case rte_eth_tx_burst() & rte_eth_rx_burst() shouldn't check for function pointers even when DEBUG is enabled.
static inline uint16_t
rte_eth_rx_burst(uint16_t port_id, uint16_t queue_id,
struct rte_mbuf **rx_pkts, const uint16_t nb_pkts)
{
struct rte_eth_dev *dev = &rte_eth_devices[port_id];
uint16_t nb_rx;#ifdef RTE_LIBRTE_ETHDEV_DEBUG
RTE_ETH_VALID_PORTID_OR_ERR_RET(port_id, 0);
RTE_FUNC_PTR_OR_ERR_RET(*dev->rx_pkt_burst, 0);
if (queue_id >= dev->data->nb_rx_queues) {
RTE_ETHDEV_LOG(ERR, "Invalid RX queue_id=%u\n", queue_id);
return 0;
}
#endif
nb_rx = (*dev->rx_pkt_burst)(dev->data->rx_queues[queue_id],
rx_pkts, nb_pkts);From my view point, function pointer checks and argument checks are required in every API for stability. But having such checks in the datapath adversely affects the performance. And for cases where function pointers are not set, application would get one crash in the first run. And that can be debugged after having the required options enabled.
Quoted 18 lines
For all other calls (both data and control path) there is a check that actual function pointer is a valid one.
Same story for eth dev: pkt_rx_burst/pkt_tx_burst and rest of dev-ops.Quoted 3 lines
Quoted 1 line
So what you guys did is a silent change of public API behaviour.
[Anoob] I believe Lukasz had submitted 3 or 4 revisions and it was all in the ML.
RTE_DEBUGwas suggested by Thomas I guess.I believe it is not a right procedure to change existing behaviour of
rte_securityframework.
I think you have to communicate clear and loudly in advance (at least one release in advance).
PlusRTE_DEBUGhas nothing to do with changing non-debug behaviour.Quoted 3 lines
Quoted 2 lines
As result ixgbe, (and probably some others
rte_securityPMDs) stopped working properly.
Quoted 8 lines
Quoted 2 lines
[Anoob]
set_pkt_metadata()is the only one of interest to IXGBE. And I believe the function is implemented as well. So what exactly is the concern?Check that
ops->get_userdatais a valid function pointer will be compiled out. So PMDs that don't implement this function will crash inrte_security_get_userdata().
In our particular case - ixgbe.
Same story withrte_security_set_pkt_metadata()- see the patch.
[Anoob] But ixgbe doesn't implement inline protocol which is the primary consumer of this API (rte_security_get_userdata()). So what is the trouble?
Also, application is expected to call rte_security_set_pkt_metadata() only on devices with offload flag RTE_SECURITY_TX_OLOAD_NEED_MDATA. If a PMD states it needs MDATA but fails to register a function pointer for doing the same, it is a control path problem. Checking for that in the datapath is an overkill.
Quoted 40 lines
Quoted 40 lines
Quoted 39 lines
I don't see any point in these changes, but if you'd like to do that, at least our usual procedure has to be followed:
1. Send and RFC to get an agreement withrte_securityPMDs maintainers (one release ahead) 2. send a deprecation note (one release ahead) 3. change the behaviour of the public API 4. update release notesAFAIK 1), 2), 4) wasn't done.
So I think right now we need to revert original behaviour.Quoted 6 lines
https://urldefense.proofpoint.com/v2/url?u=http-3A__code.dpdk.org_
dpdk _v20.02_source_lib_librte-5Fethdev_rte-5Fethdev.h-23L4372&d=DwIFAg &c=n
KjWec2b6R0mOyPaz7xtfQ&r=jPfB8rwwviRSxyLWs2n6B-WYLn1v9SyTMrT5EQqh2TU&m=
Quoted 1 line
6ObfSanVVuHOsiqVlWxXsFWi-
2XNp76HCOX0vbUfma4&s=jDVyDDEILmgY1Yb9ZBswBVbn
Quoted 4 lines
8FpZuQI5ukH_osmtUiI&e=
Datapath functions in cryptodev (enqueue/dequeue) doesn't even have such
checks.
Quoted 5 lines
https://urldefense.proofpoint.com/v2/url?u=http-3A__code.dpdk.org_
dpdk _v20.02_source_lib_librte-5Fcryptodev_rte-5Fcryptodev.h-23L962&d=D wIFA g&c=nKjWec2b6R0mOyPaz7xtfQ&r=jPfB8rwwviRSxyLWs2n6B-WYLn1v9SyTMrT5EQqh2
Quoted 1 line
TU&m=6ObfSanVVuHOsiqVlWxXsFWi-
2XNp76HCOX0vbUfma4&s=LEWQOKs0r2Im_zL95VI
Quoted 1 line
df4kQ2Pu0iRHV9Co2J1gsNBE&e=
That's a different story:
rx_burst/tx_burst, enqueue/dequeue are mandatory dev-ops functions that have to be implemented by each ethdev/cryptodev API.
Quoted 5 lines
Quoted 2 lines
[Anoob] I couldn't find any reference stating that way. If you can point me, I can update that to include datapath ops required for inline protocol
processing.
Look at the code.
[Anoob] Code is not conclusive for this as I've explained above for rte_eth_rx/tx_burst() case.
Quoted 56 lines
Quoted 56 lines
Quoted 56 lines
Quoted 27 lines
Thanks,
AnoobQuoted 19 lines
-----Original Message-----
From: dev <[email protected]> On Behalf Of Konstantin Ananyev
Sent: Thursday, April 23, 2020 5:22 AM
To: [email protected]
Cc: [email protected]; [email protected]; Konstantin Ananyev <[email protected]>
Subject: [dpdk-dev] [PATCH] security: fix crash at accessing non-implemented opsValid checks for optional function pointers inside dev-ops were disabled by undefined macro.
Fixes b6ee98547847 ("security: fix verification of parameters")Signed-off-by Konstantin Ananyev <[email protected]>--- lib/librte_security/rte_security.c | 4 ---- 1 file changed, 4 deletions(-)Quoted 9 lines
lib/librte_security/rte_security.c rte_security_set_pkt_metadata(struct · line 109
Quoted 9 lines
lib/librte_security/rte_security.c +0
-0@@ -107,11 +107,9 @@ rte_security_set_pkt_metadata(structrte_security_ctx *instance, struct rte_security_session *sess, struct rte_mbuf *m, void *params) { -#ifdef
RTE_DEBUGQuoted 19 lines
Quoted 19 lines
lib/librte_security/rte_security.c +0
-1RTE_PTR_CHAIN3_OR_ERR_RET (instance, ops, set_pkt_metadata, -EINVAL , -ENOTSUP ); RTE_PTR_OR_ERR_RET(sess, -EINVAL); -#endifreturn instance->ops->set_pkt_metadata (instance->device, sess, m, params); }@@ -121,9 +119,7 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md) { void *userdata = NULL; -#ifdef RTE_DEBUG RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, get_userdata, NULL,NULL); -#endif
lib/librte_security/rte_security.c +0
-1if (instance->ops->get_userdata (instance->device, md, &userdata))return NULL ; --2.17.1
Hi Anoob/Konstantin,
Quoted 7 lines
Quoted 6 lines
Check that
ops->get_userdatais a valid function pointer will be compiled out. So PMDs that don't implement this function will crash inrte_security_get_userdata().
In our particular case - ixgbe.
Same story withrte_security_set_pkt_metadata()- see the patch.
Quoted 8 lines
[Anoob] But ixgbe doesn't implement inline protocol which is the primary consumer of this API (
rte_security_get_userdata()). So what is the trouble?Also, application is expected to call
rte_security_set_pkt_metadata()only on devices with offload flagRTE_SECURITY_TX_OLOAD_NEED_MDATA. If a PMD states it needs MDATA but fails to register a function pointer for doing the same, it is a control path problem. Checking for that in the datapath is an overkill.
Whatever your concern is, we can resolve it later, but for now we should have the same Unconditional checks that were there earlier. We need to make RC1 today/tomorrow. And this cannot go as an issue.
These are optional APIs and every PMD may not have supported that.
Konstantin,
Please send an update to your patch reverting the original patch for these 2 functions. Currently it is adding 2 extra checks.
Regards,
Akhil
W dniu 23.04.2020 o 14:55, Akhil Goyal pisze:
Quoted 18 lines
Hi Anoob/Konstantin,
Quoted 5 lines
Quoted 5 lines
Check that
ops->get_userdatais a valid function pointer will be compiled out. So PMDs that don't implement this function will crash inrte_security_get_userdata().
In our particular case - ixgbe.
Same story withrte_security_set_pkt_metadata()- see the patch.Quoted 8 lines
[Anoob] But ixgbe doesn't implement inline protocol which is the primary consumer of this API (
rte_security_get_userdata()). So what is the trouble?Also, application is expected to call
rte_security_set_pkt_metadata()only on devices with offload flagRTE_SECURITY_TX_OLOAD_NEED_MDATA. If a PMD states it needs MDATA but fails to register a function pointer for doing the same, it is a control path problem. Checking for that in the datapath is an overkill.Whatever your concern is, we can resolve it later, but for now we should have the same Unconditional checks that were there earlier. We need to make RC1 today/tomorrow. And this cannot go as an issue.
Quoted 9 lines
These are optional APIs and every PMD may not have supported that.
Konstantin,
Please send an update to your patch reverting the original patch for these 2 functions. Currently it is adding 2 extra checks.Regards,
Akhil
Please remember also about updating app/test.
I will be glad to help with this matter.
Signature · 7 lines
Lukasz Wojciechowski Principal Software Engineer Samsung R&D Institute Poland Samsung Electronics Office +48 22 377 88 25 [email protected]
Hi Akhil,
Quoted 21 lines
Hi Anoob/Konstantin,
Quoted 7 lines
Quoted 6 lines
Check that
ops->get_userdatais a valid function pointer will be compiled out. So PMDs that don't implement this function will crash inrte_security_get_userdata().
In our particular case - ixgbe.
Same story withrte_security_set_pkt_metadata()- see the patch.Quoted 8 lines
[Anoob] But ixgbe doesn't implement inline protocol which is the primary consumer of this API (
rte_security_get_userdata()). So what is the trouble?Also, application is expected to call
rte_security_set_pkt_metadata()only on devices with offload flagRTE_SECURITY_TX_OLOAD_NEED_MDATA. If a PMD states it needs MDATA but fails to register a function pointer for doing the same, it is a control path problem. Checking for that in the datapath is an overkill.Whatever your concern is, we can resolve it later, but for now we should have the same Unconditional checks that were there earlier. We need to make RC1 today/tomorrow. And this cannot go as an issue.
Quoted 6 lines
These are optional APIs and every PMD may not have supported that.
Konstantin,
Please send an update to your patch reverting the original patch for these 2 functions. Currently it is adding 2 extra checks.
I am afraid we can't do just that.
As in that case /app/test/test_security.c build wih -DRE_DEBUG will start crashing.
I think we have 3 alternative how to fix it:
1. Keep all these 3 checks for debug and non-debug mode (that what my current patch does).
2. Have both: existed 1 check in non-debug mode, plus new checks in debug mode, i.e.:rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md)
{
void *userdata = NULL;
+#ifdef RTE_DEBUG
+ RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, get_userdata, NULL, NULL);
+#else
RTE_FUNC_PTR_OR_ERR_RET(*instance->ops->get_userdata, NULL);
+#endif
....
3. Keep only 1 existed check in non-debug mode and remove cases in app/test/test_security.c that would crash with -DRTE_DEBUG.
My preference is 1), I don't think these 2 extra checks will affect performance greatly. Also with 1) we can make these new test-case to be executed for non-debug mode too.
2) is probably also ok - but I think RTE_DEBUG concept should be a separate patch series, and I don't want to mix things.
What is your opinion here?
Konstantin
Quoted 65 lines
Hi Akhil,
Quoted 4 lines
Hi Anoob/Konstantin,
Quoted 2 lines
Quoted 2 lines
Check that
ops->get_userdatais a valid function pointer will be compiledout.
Quoted 10 lines
Quoted 10 lines
Quoted 4 lines
So PMDs that don't implement this function will crash in
rte_security_get_userdata().
In our particular case - ixgbe.
Same story withrte_security_set_pkt_metadata()- see the patch.[Anoob] But ixgbe doesn't implement inline protocol which is the primary consumer of this API (
rte_security_get_userdata()). So what is the trouble?Also, application is expected to call
rte_security_set_pkt_metadata()only on devices with offload flagRTE_SECURITY_TX_OLOAD_NEED_MDATA. If aPMD
Quoted 1 line
Quoted 1 line
states it needs MDATA but fails to register a function pointer for doing the
same,
Quoted 3 lines
Quoted 2 lines
it is a control path problem. Checking for that in the datapath is an overkill.
Whatever your concern is, we can resolve it later, but for now we should have
the same
Quoted 1 line
Unconditional checks that were there earlier. We need to make RC1
today/tomorrow.
Quoted 6 lines
And this cannot go as an issue.
These are optional APIs and every PMD may not have supported that.
Konstantin,
Please send an update to your patch reverting the original patch for these 2functions.
Quoted 2 lines
Currently it is adding 2 extra checks.
I am afraid we can't do just that.
As in that case /app/test/test_security.c build wih -DRE_DEBUG will start crashing.I think we have 3 alternative how to fix it:
1. Keep all these 3 checks for debug and non-debug mode (that what my current patch does).
2. Have both: existed 1 check in non-debug mode, plus new checks in debug mode, i.e.:rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md)
{(quoted diff) +2
-0void *userdata = NULL; +#ifdef RTE_DEBUG + RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, get_userdata, NULL,NULL);
+#elseRTE_FUNC_PTR_OR_ERR_RET(*instance->ops->get_userdata, NULL);+#endif
....
3. Keep only 1 existed check in non-debug mode and remove cases in
app/test/test_security.c
that would crash with -DRTE_DEBUG.
Quoted 9 lines
My preference is 1), I don't think these 2 extra checks will affect performance greatly.
Also with 1) we can make these new test-case to be executed for non-debug mode too.
2) is probably also ok - but I thinkRTE_DEBUGconcept should be a separate patch series,
and I don't want to mix things.
What is your opinion here?
I am OK with both 1 and 2.
Anoob may be concerned about the performance.
But if we go with 2, it would be better to have rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md)
{
void *userdata = NULL;
+#ifdef RTE_DEBUG
+ RTE_PTR_OR_ERR_RET(instance, NULL);
+ RTE_PTR_OR_ERR_RET(instance->ops, NULL);
+#endif
RTE_FUNC_PTR_OR_ERR_RET(*instance->ops->get_userdata, NULL);
....
}
And for security test, we can have a separate patch. Lukasz or you can send that later if not now.
Quoted 89 lines
Quoted 65 lines
Hi Akhil,
Quoted 4 lines
Hi Anoob/Konstantin,
Quoted 2 lines
Quoted 2 lines
Check that
ops->get_userdatais a valid function pointer will be compiledout.
Quoted 10 lines
Quoted 10 lines
Quoted 4 lines
So PMDs that don't implement this function will crash in
rte_security_get_userdata().
In our particular case - ixgbe.
Same story withrte_security_set_pkt_metadata()- see the patch.[Anoob] But ixgbe doesn't implement inline protocol which is the primary consumer of this API (
rte_security_get_userdata()). So what is the trouble?Also, application is expected to call
rte_security_set_pkt_metadata()only on devices with offload flagRTE_SECURITY_TX_OLOAD_NEED_MDATA. If aPMD
Quoted 1 line
Quoted 1 line
states it needs MDATA but fails to register a function pointer for doing the
same,
Quoted 3 lines
Quoted 2 lines
it is a control path problem. Checking for that in the datapath is an overkill.
Whatever your concern is, we can resolve it later, but for now we should have
the same
Quoted 1 line
Unconditional checks that were there earlier. We need to make RC1
today/tomorrow.
Quoted 6 lines
And this cannot go as an issue.
These are optional APIs and every PMD may not have supported that.
Konstantin,
Please send an update to your patch reverting the original patch for these 2functions.
Quoted 2 lines
Currently it is adding 2 extra checks.
I am afraid we can't do just that.
As in that case /app/test/test_security.c build wih -DRE_DEBUG will start crashing.I think we have 3 alternative how to fix it:
1. Keep all these 3 checks for debug and non-debug mode (that what my current patch does).
2. Have both: existed 1 check in non-debug mode, plus new checks in debug mode, i.e.:rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md)
{(quoted diff) +2
-0void *userdata = NULL; +#ifdef RTE_DEBUG + RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, get_userdata, NULL,NULL);
+#elseRTE_FUNC_PTR_OR_ERR_RET(*instance->ops->get_userdata, NULL);+#endif
....
3. Keep only 1 existed check in non-debug mode and remove cases in
app/test/test_security.c
that would crash with -DRTE_DEBUG.Quoted 9 lines
My preference is 1), I don't think these 2 extra checks will affect performance greatly.
Also with 1) we can make these new test-case to be executed for non-debug mode too.
2) is probably also ok - but I thinkRTE_DEBUGconcept should be a separate patch series,
and I don't want to mix things.
What is your opinion here?I am OK with both 1 and 2.
Anoob may be concerned about the performance.
But if we go with 2, it would be better to haverte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md)
{
void *userdata = NULL;+#ifdef
RTE_DEBUG
+RTE_PTR_OR_ERR_RET(instance, NULL);
+RTE_PTR_OR_ERR_RET(instance->ops, NULL);
+#endif
RTE_FUNC_PTR_OR_ERR_RET(*instance->ops->get_userdata, NULL);
....
}
Quoted 1 line
And for security test, we can have a separate patch. Lukasz or you can send that later if not now.
Ok, then to keep everyone happy will go with your code snippet above.
W dniu 23.04.2020 o 06:07, Anoob Joseph pisze:
Quoted 4 lines
Hi Konstantin,
These are data path ops and so it will be better if we can avoid such checks in the datapath. The same is done in ethdev also.
Quoted 8 lines
Datapath functions in cryptodev (enqueue/dequeue) doesn't even have such checks.
https://protect2.fireeye.com/url?k=51324200-0ca9be63-5133c94f-0cc47a31cdbc-11f88758fc12c996&q=1&u=http%3A%2F%2Fcode.dpdk.org%2Fdpdk%2Fv20.02%2Fsource%2Flib%2Flibrte_cryptodev%2Frte_cryptodev.h%23L962Thanks,
Anoob
Hi Konstantine,
It's my fault. Sorry.
These checks need to be disabled in non-debug code, so they should be wrapped in a macro. It's just not the valid macro.
The discussion about rte_debug mode is ongoing
(https://patchwork.dpdk.org/patch/68815/)
and currently the v2 version of patches is prepared to gather maintainers opinion.
After the rte_debug is introduced the proper macro to use will be RTE_DEBUG_SECURITY.
Until then, the RTE_DEBUG macro can stay as like Anoob mentioned the checks will have impact on dataplane performance.
If you want to enable this code, please use CFLAGS="-DRTE_DEBUG"
Quoted 50 lines
Quoted 49 lines
-----Original Message-----
From: dev <[email protected]> On Behalf Of Konstantin Ananyev
Sent: Thursday, April 23, 2020 5:22 AM
To: [email protected]
Cc: [email protected]; [email protected]; Konstantin Ananyev <[email protected]>
Subject: [dpdk-dev] [PATCH] security: fix crash at accessing non-implemented opsValid checks for optional function pointers inside dev-ops were disabled by undefined macro.
Fixes b6ee98547847 ("security: fix verification of parameters")Signed-off-by Konstantin Ananyev <[email protected]>--- lib/librte_security/rte_security.c | 4 ---- 1 file changed, 4 deletions(-)lib/librte_security/rte_security.c +0
-0@@ -107,11 +107,9 @@ rte_security_set_pkt_metadata(struct rte_security_ctx *instance, struct rte_security_session *sess, struct rte_mbuf *m, void *params) { -#ifdef
RTE_DEBUG
RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops,set_pkt_metadata, -
EINVAL,
-ENOTSUP);
RTE_PTR_OR_ERR_RET(sess, -EINVAL);lib/librte_security/rte_security.c +0
-2-#endif return instance->ops->set_pkt_metadata (instance->device, sess, m, params); }@@ -121,9 +119,7 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md) {void *userdata =NULL ; -#ifdef RTE_DEBUG RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, get_userdata, NULL,NULL); -#endif
if (instance->ops->get_userdata(instance->device, md, &userdata))
return NULL;--
2.17.1
Signature · 7 lines
Lukasz Wojciechowski Principal Software Engineer Samsung R&D Institute Poland Samsung Electronics Office +48 22 377 88 25 [email protected]
Quoted 30 lines
W dniu 23.04.2020 o 06:07, Anoob Joseph pisze:
Quoted 5 lines
Hi Konstantin,
These are data path ops and so it will be better if we can avoid such checks in the datapath. The same is done in ethdev also.
https://protect2.fireeye.com/url?k=d44931cf-89d2cdac-d448ba80-0cc47a31cdbc-
8281a62b4c91d848&q=1&u=http%3A%2F%2Fcode.dpdk.org%2Fdpdk%2Fv20.02%2Fsource%2Flib%2Flibrte_ethdev%2Frte_ethdev.h%23L43 72
Quoted 3 lines
Datapath functions in cryptodev (enqueue/dequeue) doesn't even have such checks.
https://protect2.fireeye.com/url?k=51324200-0ca9be63-5133c94f-0cc47a31cdbc-11f88758fc12c996&q=1&u=http%3A%2F%2Fcode.dpdk.org%2Fdpdk%2Fv20.02%2Fsource%2Flib%2Flibrte_cryptodev%2Frte_cryptodev.h% 23L962
Quoted 4 lines
Thanks,
AnoobHi Konstantine,
It's my fault. Sorry.
These checks need to be disabled in non-debug code, so they should be wrapped in a macro. It's just not the valid macro.
The discussion aboutrte_debugmode is ongoing
(https://patchwork.dpdk.org/patch/68815/)
and currently the v2 version of patches is prepared to gather maintainers opinion.
Quoted 7 lines
After the
rte_debugis introduced the proper macro to use will beRTE_DEBUG_SECURITY.Until then, the
RTE_DEBUGmacro can stay as like Anoob mentioned the checks will have impact on dataplane performance.If you want to enable this code, please use CFLAGS="-DRTE_DEBUG"
Really? So what we have to tell now to our customers? "Yes, rte_security is broken and can easily crash your app. But we might fix it in future versions... or maybe not. For now just recompile our source with that flag enabled?" Obviously this is not an option.
It is a bug and it is a stopper for 20.05 release. It has to be fixed asap.
Quoted 60 lines
Quoted 42 lines
Quoted 42 lines
-----Original Message-----
From: dev <[email protected]> On Behalf Of Konstantin Ananyev
Sent: Thursday, April 23, 2020 5:22 AM
To: [email protected]
Cc: [email protected]; [email protected]; Konstantin Ananyev <[email protected]>
Subject: [dpdk-dev] [PATCH] security: fix crash at accessing non-implemented opsValid checks for optional function pointers inside dev-ops were disabled by undefined macro.
Fixes b6ee98547847 ("security: fix verification of parameters")Signed-off-by Konstantin Ananyev <[email protected]>--- lib/librte_security/rte_security.c | 4 ---- 1 file changed, 4 deletions(-)lib/librte_security/rte_security.c +0
-0@@ -107,11 +107,9 @@ rte_security_set_pkt_metadata(struct rte_security_ctx *instance, struct rte_security_session *sess, struct rte_mbuf *m, void *params) { -#ifdef
RTE_DEBUG
RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops,set_pkt_metadata, -
EINVAL,
-ENOTSUP);
RTE_PTR_OR_ERR_RET(sess, -EINVAL);lib/librte_security/rte_security.c +0
-1-#endif return instance->ops->set_pkt_metadata (instance->device, sess, m, params); }@@ -121,9 +119,7 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md) {void *userdata =NULL ;Quoted 9 lines
Quoted 8 lines
lib/librte_security/rte_security.c +0
-1-#ifdef RTE_DEBUG RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, get_userdata, NULL,NULL); -#endif
if (instance->ops->get_userdata(instance->device, md, &userdata))
return NULL;--
2.17.1--
Lukasz Wojciechowski
Principal Software EngineerSamsung R&D Institute Poland
Samsung Electronics
Office +48 22 377 88 25
[email protected]
PATCH v2 1/1 security: fix crash at accessing non-implemented ops
In reply to Konstantin AnanyevValid checks for optional function pointers inside dev-ops were disabled by undefined macro.
--- lib/librte_security/rte_security.c | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-)
lib/librte_security/rte_security.c +6 -3
@@ -108,10 +108,11 @@ rte_security_set_pkt_metadata(struct rte_security_ctx *instance, struct rte_mbuf *m, void *params) { #ifdef RTE_DEBUG - RTE_PTR_CHAIN3_OR_ERR_RET(instance, ops, set_pkt_metadata, -EINVAL, - -ENOTSUP); RTE_PTR_OR_ERR_RET(sess, -EINVAL); +RTE_PTR_OR_ERR_RET (instance, -EINVAL ); +RTE_PTR_OR_ERR_RET (instance->ops, -EINVAL );#endif +RTE_FUNC_PTR_OR_ERR_RET (*instance->ops->set_pkt_metadata, -ENOTSUP );return instance->ops->set_pkt_metadata (instance->device, sess, m, params); }
@@ -122,8 +123,10 @@ rte_security_get_userdata(struct rte_security_ctx *instance, uint64_t md)void *userdata =NULL ;#ifdef RTE_DEBUG -RTE_PTR_CHAIN3_OR_ERR_RET (instance, ops, get_userdata,NULL ,NULL ); +RTE_PTR_OR_ERR_RET (instance,NULL ); +RTE_PTR_OR_ERR_RET (instance->ops,NULL );#endif +RTE_FUNC_PTR_OR_ERR_RET (*instance->ops->get_userdata,NULL );if (instance->ops->get_userdata (instance->device, md, &userdata))return NULL ;
Signature · 1 line
2.17.1
Quoted 9 lines
Valid checks for optional function pointers inside dev-ops were disabled by undefined macro.
Fixes b6ee98547847 ("security: fix verification of parameters")Signed-off-by Konstantin Ananyev <[email protected]>---
Anoob,
Do you have any concerns over this patch?
Regards,
Akhil
Hi Akhil,
I have my concerns over unwanted checks in the datapath. Something that crypto enqueue/dequeue APIs are not doing is being enforced on other APIs. As Konstantin had suggested, PMDs (IXGBE here) could define a function which returns -ENOTSUP and it would have been win-win for everyone.
Anyway, I don't have any objections to this.
Thanks,
Anoob
Quoted 30 lines
-----Original Message-----
From: Akhil Goyal <[email protected]>
Sent: Thursday, April 23, 2020 9:22 PM
To: Konstantin Ananyev <[email protected]>; [email protected]; Anoob Joseph <[email protected]>
Cc: [email protected]; [email protected]
Subject: [EXT] RE: [PATCH v2] security: fix crash at accessing non-implemented opsExternal Email
----------------------------------------------------------------------
Quoted 8 lines
Valid checks for optional function pointers inside dev-ops were disabled by undefined macro.
Fixes b6ee98547847 ("security: fix verification of parameters")Signed-off-by Konstantin Ananyev <[email protected]>---
Acked-by Akhil Goyal <[email protected]>Anoob,
Do you have any concerns over this patch?
Regards,
Akhil
Hi Anoob,
Quoted 10 lines
Hi Akhil,
I have my concerns over unwanted checks in the datapath. Something that crypto enqueue/dequeue APIs are not doing is being enforced on other APIs. As Konstantin had suggested, PMDs (IXGBE here) could define a function which returns
-ENOTSUPand it would have been win-win for everyone.Anyway, I don't have any objections to this.
Your concerns are valid and those can be handled separately. We will apply this patch.
W dniu 23.04.2020 o 18:14, Akhil Goyal pisze:
Quoted 8 lines
Hi Anoob,
Quoted 7 lines
Hi Akhil,
I have my concerns over unwanted checks in the datapath. Something that crypto enqueue/dequeue APIs are not doing is being enforced on other APIs. As Konstantin had suggested, PMDs (IXGBE here) could define a function which returns
-ENOTSUPand it would have been win-win for everyone.
Quoted 4 lines
Quoted 2 lines
Anyway, I don't have any objections to this.
Your concerns are valid and those can be handled separately. We will apply this patch.
I just pushed a patch enabling 2 tests in non-debug build mode.
Sorry for the trouble I caused
Best regards
Lukasz
Signature · 7 lines
Lukasz Wojciechowski Principal Software Engineer Samsung R&D Institute Poland Samsung Electronics Office +48 22 377 88 25 [email protected]
PATCH 1/1 test/security: enable tests for non-implemented ops
In reply to Konstantin AnanyevAfter re-enabling checks for non-implemneted ops in non-debug mode in librte_security set_pkt_metadata and get_userdata functions, tests verifying proper work of tests can be enabled also.
--- app/test/test_security.c | 8 -------- 1 file changed, 8 deletions(-)
app/test/test_security.c +0 -8
@@ -1474,7 +1474,6 @@ test_set_pkt_metadata_inv_context_ops(void)static int test_set_pkt_metadata_inv_context_ops_fun (void ) { -#ifdef RTE_DEBUG struct security_unittest_params *ut_params = &unittest_params;struct rte_mbuf m;int params;
@@ -1487,9 +1486,6 @@ test_set_pkt_metadata_inv_context_ops_fun(void)TEST_ASSERT_MOCK_CALLS (mock_set_pkt_metadata_exp,0 );return TEST_SUCCESS ; -#else -return TEST_SKIPPED ; -#endif }/**
@@ -1621,7 +1617,6 @@ test_get_userdata_inv_context_ops(void)static int test_get_userdata_inv_context_ops_fun (void ) { -#ifdef RTE_DEBUG struct security_unittest_params *ut_params = &unittest_params;uint64_t md =0xDEADBEEF ; ut_params->ctx.ops = &empty_ops;
@@ -1632,9 +1627,6 @@ test_get_userdata_inv_context_ops_fun(void)TEST_ASSERT_MOCK_CALLS (mock_get_userdata_exp,0 );return TEST_SUCCESS ; -#else -return TEST_SKIPPED ; -#endif }/**
Signature · 1 line
2.17.1
Re: [dpdk-dev] [PATCH] test/security: enable tests for non-implemented ops
Quoted 7 lines
After re-enabling checks for non-implemneted ops in non-debug mode in
librte_securityset_pkt_metadataandget_userdatafunctions, tests verifying proper work of tests can be enabled also.Signed-off-by Lukasz Wojciechowski <[email protected]>---
Applied to dpdk-next-crypto
Thanks.
Re: [dpdk-dev] [PATCH] test/security: enable tests for non-implemented ops
W dniu 09.05.2020 o 23:47, Akhil Goyal pisze:
Quoted 11 lines
Quoted 6 lines
After re-enabling checks for non-implemneted ops in non-debug mode in
librte_securityset_pkt_metadataandget_userdatafunctions, tests verifying proper work of tests can be enabled also.Signed-off-by Lukasz Wojciechowski <[email protected]>---
Acked-by Akhil Goyal <[email protected]>Applied to dpdk-next-crypto
Thanks.
Thank you
Signature · 7 lines
Lukasz Wojciechowski Principal Software Engineer Samsung R&D Institute Poland Samsung Electronics Office +48 22 377 88 25 [email protected]