[PATCH] crypto: af_alg - Fix regression on empty requests

On Tue, Jun 23, 2020 at 10:02:17AM -0700, Eric Biggers wrote:

The source code for the two failing AF_ALG tests is here:

https://github.com/linux-test-project/ltp/blob/master/testcases/kernel/crypto/af_alg02.c
https://github.com/linux-test-project/ltp/blob/master/testcases/kernel/crypto/af_alg05.c

They use read() and write(), not send() and recv().

af_alg02 uses read() to read from a "salsa20" request socket without writing
anything to it.  It is expected that this returns 0, i.e. that behaves like
encrypting an empty message.

af_alg05 uses write() to write 15 bytes to a "cbc(aes-generic)" request socket,
then read() to read 15 bytes.  It is expected that this fails with EINVAL, since
the length is not aligned to the AES block size (16 bytes).

This patch should fix the regression:

---8<---
Some user-space programs rely on crypto requests that have no
control metadata.  This broke when a check was added to require
the presence of control metadata with the ctx->init flag.

This patch fixes the regression by removing the ctx->init flag.

This means that we do not distinguish the case of no metadata
as opposed to an empty request.  IOW it is always assumed that
if you call recv(2) before sending metadata that you are working
with an empty request.

Reported-by: Sachin Sant <redacted>
Reported-by: Naresh Kamboju <redacted>
Fixes: f3c802a1f300 ("crypto: algif_aead - Only wake up when...")
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

diff --git a/crypto/af_alg.c b/crypto/af_alg.c
index 9fcb91ea10c4..2d391117c020 100644
--- a/crypto/af_alg.c
+++ b/crypto/af_alg.c

@@ -635,7 +635,6 @@ void af_alg_pull_tsgl(struct sock *sk, size_t used, struct scatterlist *dst,
 
 	if (!ctx->used)
 		ctx->merge = 0;
-	ctx->init = ctx->more;
 }
 EXPORT_SYMBOL_GPL(af_alg_pull_tsgl);
 

@@ -757,8 +756,7 @@ int af_alg_wait_for_data(struct sock *sk, unsigned flags, unsigned min)
 			break;
 		timeout = MAX_SCHEDULE_TIMEOUT;
 		if (sk_wait_event(sk, &timeout,
-				  ctx->init && (!ctx->more ||
-						(min && ctx->used >= min)),
+				  !ctx->more || (min && ctx->used >= min),
 				  &wait)) {
 			err = 0;
 			break;

@@ -847,7 +845,7 @@ int af_alg_sendmsg(struct socket *sock, struct msghdr *msg, size_t size,
 	}
 
 	lock_sock(sk);
-	if (ctx->init && (init || !ctx->more)) {
+	if (!ctx->more && ctx->used) {
 		err = -EINVAL;
 		goto unlock;
 	}

@@ -858,7 +856,6 @@ int af_alg_sendmsg(struct socket *sock, struct msghdr *msg, size_t size,
 			memcpy(ctx->iv, con.iv->iv, ivsize);
 
 		ctx->aead_assoclen = con.aead_assoclen;
-		ctx->init = true;
 	}
 
 	while (size) {

diff --git a/crypto/algif_aead.c b/crypto/algif_aead.c
index d48d2156e621..749fe42315be 100644
--- a/crypto/algif_aead.c
+++ b/crypto/algif_aead.c

@@ -106,7 +106,7 @@ static int _aead_recvmsg(struct socket *sock, struct msghdr *msg,
 	size_t usedpages = 0;		/* [in]  RX bufs to be used from user */
 	size_t processed = 0;		/* [in]  TX bufs to be consumed */
 
-	if (!ctx->init || ctx->more) {
+	if (ctx->more) {
 		err = af_alg_wait_for_data(sk, flags, 0);
 		if (err)
 			return err;

diff --git a/crypto/algif_skcipher.c b/crypto/algif_skcipher.c
index a51ba22fef58..5b6fa5e8c00d 100644
--- a/crypto/algif_skcipher.c
+++ b/crypto/algif_skcipher.c

@@ -61,7 +61,7 @@ static int _skcipher_recvmsg(struct socket *sock, struct msghdr *msg,
 	int err = 0;
 	size_t len = 0;
 
-	if (!ctx->init || (ctx->more && ctx->used < bs)) {
+	if (ctx->more && ctx->used < bs) {
 		err = af_alg_wait_for_data(sk, flags, bs);
 		if (err)
 			return err;

diff --git a/include/crypto/if_alg.h b/include/crypto/if_alg.h
index ee6412314f8f..08c087cc89d6 100644
--- a/include/crypto/if_alg.h
+++ b/include/crypto/if_alg.h

@@ -135,7 +135,6 @@ struct af_alg_async_req {
  *			SG?
  * @enc:		Cryptographic operation to be performed when
  *			recvmsg is invoked.
- * @init:		True if metadata has been sent.
  * @len:		Length of memory allocated for this data structure.
  */
 struct af_alg_ctx {

@@ -152,7 +151,6 @@ struct af_alg_ctx {
 	bool more;
 	bool merge;
 	bool enc;
-	bool init;
 
 	unsigned int len;
 };

-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

On Fri, 26 Jun 2020 at 12:00, Herbert Xu [off-list ref] wrote:

On Tue, Jun 23, 2020 at 10:02:17AM -0700, Eric Biggers wrote:

quoted

The source code for the two failing AF_ALG tests is here:

https://github.com/linux-test-project/ltp/blob/master/testcases/kernel/crypto/af_alg02.c
https://github.com/linux-test-project/ltp/blob/master/testcases/kernel/crypto/af_alg05.c

They use read() and write(), not send() and recv().

af_alg02 uses read() to read from a "salsa20" request socket without writing
anything to it.  It is expected that this returns 0, i.e. that behaves like
encrypting an empty message.

Since we are on this subject,
LTP af_alg02  test case fails on stable 4.9 and stable 4.4
This is not a regression because the test case has been failing from
the beginning.

Is this test case expected to fail on stable 4.9 and 4.4 ?
or any chance to fix this on these older branches ?

Test output:
af_alg02.c:52: BROK: Timed out while reading from request socket.

ref:
https://qa-reports.linaro.org/lkft/linux-stable-rc-4.9-oe/build/v4.9.228-191-g082e807235d7/testrun/2884917/suite/ltp-crypto-tests/test/af_alg02/history/
https://qa-reports.linaro.org/lkft/linux-stable-rc-4.9-oe/build/v4.9.228-191-g082e807235d7/testrun/2884606/suite/ltp-crypto-tests/test/af_alg02/log

- Naresh

On Tue, Jun 30, 2020 at 02:18:11PM +0530, Naresh Kamboju wrote:

Since we are on this subject,
LTP af_alg02  test case fails on stable 4.9 and stable 4.4
This is not a regression because the test case has been failing from
the beginning.

Is this test case expected to fail on stable 4.9 and 4.4 ?
or any chance to fix this on these older branches ?

Test output:
af_alg02.c:52: BROK: Timed out while reading from request socket.

ref:
https://qa-reports.linaro.org/lkft/linux-stable-rc-4.9-oe/build/v4.9.228-191-g082e807235d7/testrun/2884917/suite/ltp-crypto-tests/test/af_alg02/history/
https://qa-reports.linaro.org/lkft/linux-stable-rc-4.9-oe/build/v4.9.228-191-g082e807235d7/testrun/2884606/suite/ltp-crypto-tests/test/af_alg02/log

Actually this test really is broken.  Even though empty requests
are legal, they should never be done with no write(2) at all.
Because this fundamentally breaks the use of a blocking read(2)
to wait for more data.

Granted this has been broken since 2017 but I'm not going to
reintroduce this just because of a broken test case.

So please either remove af_alg02 or fix it by adding a control
message through sendmsg(2).

Thanks,

---8<---
Some user-space programs rely on crypto requests that have no
control metadata.  This broke when a check was added to require
the presence of control metadata with the ctx->init flag.

This patch fixes the regression by setting ctx->init as long as
one sendmsg(2) has been made, with or without a control message.

Reported-by: Sachin Sant <redacted>
Reported-by: Naresh Kamboju <redacted>
Fixes: f3c802a1f300 ("crypto: algif_aead - Only wake up when...")
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>

diff --git a/crypto/af_alg.c b/crypto/af_alg.c
index 9fcb91ea10c41..5882ed46f1adb 100644
--- a/crypto/af_alg.c
+++ b/crypto/af_alg.c

@@ -851,6 +851,7 @@ int af_alg_sendmsg(struct socket *sock, struct msghdr *msg, size_t size,
 		err = -EINVAL;
 		goto unlock;
 	}
+	ctx->init = true;
 
 	if (init) {
 		ctx->enc = enc;

@@ -858,7 +859,6 @@ int af_alg_sendmsg(struct socket *sock, struct msghdr *msg, size_t size,
 			memcpy(ctx->iv, con.iv->iv, ivsize);
 
 		ctx->aead_assoclen = con.aead_assoclen;
-		ctx->init = true;
 	}
 
 	while (size) {

-- 
Email: Herbert Xu <herbert@gondor.apana.org.au>
Home Page: http://gondor.apana.org.au/~herbert/
PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt

On Thu, Jul 02, 2020 at 01:32:21PM +1000, Herbert Xu wrote:

On Tue, Jun 30, 2020 at 02:18:11PM +0530, Naresh Kamboju wrote:

quoted

Since we are on this subject,
LTP af_alg02  test case fails on stable 4.9 and stable 4.4
This is not a regression because the test case has been failing from
the beginning.

Is this test case expected to fail on stable 4.9 and 4.4 ?
or any chance to fix this on these older branches ?

Test output:
af_alg02.c:52: BROK: Timed out while reading from request socket.

ref:
https://qa-reports.linaro.org/lkft/linux-stable-rc-4.9-oe/build/v4.9.228-191-g082e807235d7/testrun/2884917/suite/ltp-crypto-tests/test/af_alg02/history/
https://qa-reports.linaro.org/lkft/linux-stable-rc-4.9-oe/build/v4.9.228-191-g082e807235d7/testrun/2884606/suite/ltp-crypto-tests/test/af_alg02/log

Actually this test really is broken.

FWIW the patch "umh: fix processed error when UMH_WAIT_PROC is used" was
dropped from linux-next for now as it was missing checking for signals.
I'll be open coding iall checks for each UMH_WAIT_PROC callers next. Its
not clear if this was the issue with this test case, but figured I'd let
you know.

  Luis