Skip to contentSkip to conversation

KEYS: asymmetric: fix OOB read in KEYCTL_PKEY_DECRYPT on zero-length message

PATCH In review azraelxuemo 3 participants 3 messages Jun 22 – Jun 24

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 KEYS: asymmetric: fix OOB read in KEYCTL_PKEY_DECRYPT on zero-length message

From: HanQuan <[email protected]>

In software_key_eds_op(), the condition

    if (!issig && ret == 0)
        ret = crypto_akcipher_maxsize(tfm);

replaces ret with the key size when crypto_akcipher_sync_decrypt() returns 0. This was added as a workaround for encrypt, where crypto_akcipher_sync_encrypt() returns 0 on success (it lacks the "?:" data.dlen" tail that sync_decrypt has). However, for decrypt, ret == 0 is legitimate: pkcs1pad_decrypt_complete() sets req->dst_len = 0 for a zero-length PKCS#1 message, causing crypto_akcipher_sync_decrypt() to return 0 via "0 ?: data.dlen".

When ret is replaced with maxsize, the caller keyctl_pkey_e_d_s() does copy_to_user(_out, out, ret) with ret = key_size (e.g. 256 for RSA-2048) on a buffer allocated with kmalloc(params.out_len), which can be as small as 1 byte. This reads key_size - out_len bytes beyond the allocation.

Restrict the maxsize substitution to the encrypt operation only, where ret == 0 genuinely indicates success and the full key size is the correct output length.

Fixes 63ba4d67594a ("KEYS: asymmetric: Use new crypto interface without scatterlists")
---
 crypto/asymmetric_keys/public_key.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
crypto/asymmetric_keys/public_key.c +4 -1
index 09a0b83d5d77..2c3ac75ec92b 100644
@@ -358,7 +358,10 @@ static int software_key_eds_op(struct kernel_pkey_params *params,
 		BUG();
 	}
 
-	if (!issig && ret == 0)
+	/* Decrypt may legitimately return 0 (zero-length message); only
+	 * replace ret with maxsize for encrypt, which returns 0 on success.
+	 */
+	if (!issig && ret == 0 && params->op == kernel_pkey_encrypt)
 		ret = crypto_akcipher_maxsize(tfm);
 
 error_free_tfm:
Signature · 1 line
2.43.0
maintainer
#1 lore raw

[cc += Ignat, Jarkko, keyrings; start of thread is here:
https://lore.kernel.org/r/[email protected]
]

Quoted 5 lines · azraelxuemo

When ret is replaced with maxsize, the caller keyctl_pkey_e_d_s() does copy_to_user(_out, out, ret) with ret = key_size (e.g. 256 for RSA-2048) on a buffer allocated with kmalloc(params.out_len), which can be as small as 1 byte. This reads key_size - out_len bytes beyond the allocation.

It would probably make sense to tighten security in keyctl_pkey_e_d_s() by using kzalloc() instead of kmalloc() and by capping the amount of data copied with min(ret, params.out_len).

Quoted 2 lines
Fixes 63ba4d67594a ("KEYS: asymmetric: Use new crypto interface without scatterlists")

Please add:

You don't need to cc that address when submitting the patch, but including the tag in the commit message helps stable maintainers identify patches that need backporting.

crypto/asymmetric_keys/public_key.c static int software_key_eds_op(struct kernel_pkey_params *params, · line 365

Quoted 11 lines

+++ b/crypto/asymmetric_keys/public_key.c

crypto/asymmetric_keys/public_key.c +4 -1
@@ -358,7 +358,10 @@ static int software_key_eds_op(struct kernel_pkey_params *params,
 		BUG();
 	}
 
-	if (!issig && ret == 0)
+	/* Decrypt may legitimately return 0 (zero-length message); only
+	 * replace ret with maxsize for encrypt, which returns 0 on success.
+	 */
+	if (!issig && ret == 0 && params->op == kernel_pkey_encrypt)
 		ret = crypto_akcipher_maxsize(tfm);

Given that out of 3 operations (encrypt, decrypt, sign),
2 already return the size, I think a better approach would be to let crypto_akcipher_sync_encrypt() return crypto_akcipher_maxsize() on success, i.e.:

return crypto_akcipher_sync_prep(&data) ?:
crypto_akcipher_sync_post(&data,
- crypto_akcipher_encrypt(data.req));
+ crypto_akcipher_encrypt(data.req)) ?:
+ crypto_akcipher_maxsize(tfm);

and then remove the if-clause in software_key_eds_op() altogether which overwrites ret with the maxsize.

Do you agree?

Your patch wasn't cc'ed to all maintainers of this file. Please double-check that you've checked out Linus' current master when running scripts/get_maintainer.pl.

Thanks,

Lukas

maintainer
#2 lore raw
Quoted 54 lines · Lukas Wunner

[cc += Ignat, Jarkko, keyrings; start of thread is here:
https://lore.kernel.org/r/[email protected]
]

Quoted 5 lines · azraelxuemo

When ret is replaced with maxsize, the caller keyctl_pkey_e_d_s() does copy_to_user(_out, out, ret) with ret = key_size (e.g. 256 for RSA-2048) on a buffer allocated with kmalloc(params.out_len), which can be as small as 1 byte. This reads key_size - out_len bytes beyond the allocation.

It would probably make sense to tighten security in keyctl_pkey_e_d_s() by using kzalloc() instead of kmalloc() and by capping the amount of data copied with min(ret, params.out_len).

Quoted 2 lines
Fixes 63ba4d67594a ("KEYS: asymmetric: Use new crypto interface without scatterlists")

Please add:

You don't need to cc that address when submitting the patch, but including the tag in the commit message helps stable maintainers identify patches that need backporting.

crypto/asymmetric_keys/public_key.c static int software_key_eds_op(struct kernel_pkey_params *params, · line 365

Quoted 11 lines

+++ b/crypto/asymmetric_keys/public_key.c

crypto/asymmetric_keys/public_key.c +4 -1
@@ -358,7 +358,10 @@ static int software_key_eds_op(struct kernel_pkey_params *params,
 		BUG();
 	}
 
-	if (!issig && ret == 0)
+	/* Decrypt may legitimately return 0 (zero-length message); only
+	 * replace ret with maxsize for encrypt, which returns 0 on success.
+	 */
+	if (!issig && ret == 0 && params->op == kernel_pkey_encrypt)
 		ret = crypto_akcipher_maxsize(tfm);

Given that out of 3 operations (encrypt, decrypt, sign),
2 already return the size, I think a better approach would be to let crypto_akcipher_sync_encrypt() return crypto_akcipher_maxsize() on success, i.e.:

return crypto_akcipher_sync_prep(&data) ?:
crypto_akcipher_sync_post(&data,

(quoted diff) +2 -1
-					 crypto_akcipher_encrypt(data.req));
+					 crypto_akcipher_encrypt(data.req)) ?:
+	       crypto_akcipher_maxsize(tfm);

and then remove the if-clause in software_key_eds_op() altogether which overwrites ret with the maxsize.

Do you agree?

Quoted 7 lines

Your patch wasn't cc'ed to all maintainers of this file. Please double-check that you've checked out Linus' current master when running scripts/get_maintainer.pl.

Thanks,

Lukas

Lukas, thank you for forwarding this, and with a quick sim I do.

Can this be next time be also CC'd to [email protected], in addition of me and rest? What I can then do is to look this in detail, test the change, apply and eventually forward to Linus...

BR, Jarkko