[PATCH v4] mac80211: move extra crypto data off the stack

[PATCH v4] mac80211: move extra crypto data off the stack

[Johannes Berg]

From: Johannes Berg <johannes.berg@intel.com>

As the stack can (on x86-64) now be virtually mapped rather than
using "normal" kernel memory, Sergey noticed mac80211 isn't using
the SG APIs correctly by putting on-stack buffers into SG tables.
This leads to kernel crashes.

Fix this by allocating the extra fields dynamically on the fly as
needed, using a kmem cache.

I used per-CPU memory in a previous iteration of this patch, but
Ard Biesheuvel pointed out that was also vmalloc'ed on some
architectures.

Reported-by: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com>
Signed-off-by: Johannes Berg <johannes.berg@intel.com>
---
 net/mac80211/aes_cmac.c    |   5 +-
 net/mac80211/aes_cmac.h    |   2 +
 net/mac80211/aes_gmac.c    |   9 ++-
 net/mac80211/aes_gmac.h    |   5 +-
 net/mac80211/ieee80211_i.h |   7 ++
 net/mac80211/main.c        |   8 +++
 net/mac80211/wpa.c         | 173 ++++++++++++++++++++++++++++++++++++---------
 7 files changed, 166 insertions(+), 43 deletions(-)

diff --git a/net/mac80211/aes_cmac.c b/net/mac80211/aes_cmac.c
index bdf0790d89cc..ebb8c2dc9928 100644
--- a/net/mac80211/aes_cmac.c
+++ b/net/mac80211/aes_cmac.c
@@ -20,7 +20,6 @@
 
 #define CMAC_TLEN 8 /* CMAC TLen = 64 bits (8 octets) */
 #define CMAC_TLEN_256 16 /* CMAC TLen = 128 bits (16 octets) */
-#define AAD_LEN 20
 
 
 static void gf_mulx(u8 *pad)
@@ -101,7 +100,7 @@ void ieee80211_aes_cmac(struct crypto_cipher *tfm, const u8 *aad,
 
 	memset(zero, 0, CMAC_TLEN);
 	addr[0] = aad;
-	len[0] = AAD_LEN;
+	len[0] = CMAC_AAD_LEN;
 	addr[1] = data;
 	len[1] = data_len - CMAC_TLEN;
 	addr[2] = zero;
@@ -119,7 +118,7 @@ void ieee80211_aes_cmac_256(struct crypto_cipher *tfm, const u8 *aad,
 
 	memset(zero, 0, CMAC_TLEN_256);
 	addr[0] = aad;
-	len[0] = AAD_LEN;
+	len[0] = CMAC_AAD_LEN;
 	addr[1] = data;
 	len[1] = data_len - CMAC_TLEN_256;
 	addr[2] = zero;
diff --git a/net/mac80211/aes_cmac.h b/net/mac80211/aes_cmac.h
index 3702041f44fd..6645f8963278 100644
--- a/net/mac80211/aes_cmac.h
+++ b/net/mac80211/aes_cmac.h
@@ -11,6 +11,8 @@
 
 #include <linux/crypto.h>
 
+#define CMAC_AAD_LEN 20
+
 struct crypto_cipher *ieee80211_aes_cmac_key_setup(const u8 key[],
 						   size_t key_len);
 void ieee80211_aes_cmac(struct crypto_cipher *tfm, const u8 *aad,
diff --git a/net/mac80211/aes_gmac.c b/net/mac80211/aes_gmac.c
index 6951af9715c0..86892e2e3c8c 100644
--- a/net/mac80211/aes_gmac.c
+++ b/net/mac80211/aes_gmac.c
@@ -19,13 +19,12 @@
 
 #define GMAC_MIC_LEN 16
 #define GMAC_NONCE_LEN 12
-#define AAD_LEN 20
 
 int ieee80211_aes_gmac(struct crypto_aead *tfm, const u8 *aad, u8 *nonce,
-		       const u8 *data, size_t data_len, u8 *mic)
+		       const u8 *data, size_t data_len, u8 *mic, u8 *zero)
 {
 	struct scatterlist sg[4];
-	u8 zero[GMAC_MIC_LEN], iv[AES_BLOCK_SIZE];
+	u8 iv[AES_BLOCK_SIZE];
 	struct aead_request *aead_req;
 
 	if (data_len < GMAC_MIC_LEN)
@@ -37,7 +36,7 @@ int ieee80211_aes_gmac(struct crypto_aead *tfm, const u8 *aad, u8 *nonce,
 
 	memset(zero, 0, GMAC_MIC_LEN);
 	sg_init_table(sg, 4);
-	sg_set_buf(&sg[0], aad, AAD_LEN);
+	sg_set_buf(&sg[0], aad, GMAC_AAD_LEN);
 	sg_set_buf(&sg[1], data, data_len - GMAC_MIC_LEN);
 	sg_set_buf(&sg[2], zero, GMAC_MIC_LEN);
 	sg_set_buf(&sg[3], mic, GMAC_MIC_LEN);
@@ -47,7 +46,7 @@ int ieee80211_aes_gmac(struct crypto_aead *tfm, const u8 *aad, u8 *nonce,
 	iv[AES_BLOCK_SIZE - 1] = 0x01;
 
 	aead_request_set_crypt(aead_req, sg, sg, 0, iv);
-	aead_request_set_ad(aead_req, AAD_LEN + data_len);
+	aead_request_set_ad(aead_req, GMAC_AAD_LEN + data_len);
 
 	crypto_aead_encrypt(aead_req);
 	aead_request_free(aead_req);
diff --git a/net/mac80211/aes_gmac.h b/net/mac80211/aes_gmac.h
index d328204d73a8..f06833c9095f 100644
--- a/net/mac80211/aes_gmac.h
+++ b/net/mac80211/aes_gmac.h
@@ -11,10 +11,13 @@
 
 #include <linux/crypto.h>
 
+#define GMAC_MIC_LEN 16
+#define GMAC_AAD_LEN 20
+
 struct crypto_aead *ieee80211_aes_gmac_key_setup(const u8 key[],
 						 size_t key_len);
 int ieee80211_aes_gmac(struct crypto_aead *tfm, const u8 *aad, u8 *nonce,
-		       const u8 *data, size_t data_len, u8 *mic);
+		       const u8 *data, size_t data_len, u8 *mic, u8 *zero);
 void ieee80211_aes_gmac_key_free(struct crypto_aead *tfm);
 
 #endif /* AES_GMAC_H */
diff --git a/net/mac80211/ieee80211_i.h b/net/mac80211/ieee80211_i.h
index 34c2add2c455..a63593f6b645 100644
--- a/net/mac80211/ieee80211_i.h
+++ b/net/mac80211/ieee80211_i.h
@@ -1128,6 +1128,13 @@ enum mac80211_scan_state {
 	SCAN_ABORT,
 };
 
+struct ieee80211_crypto_bufs {
+	u8 buf1[32];
+	u8 buf2[16];
+} ____cacheline_aligned_in_smp;
+
+extern struct kmem_cache *ieee80211_crypto_bufs_cache;
+
 struct ieee80211_local {
 	/* embed the driver visible part.
 	 * don't cast (use the static inlines below), but we keep
diff --git a/net/mac80211/main.c b/net/mac80211/main.c
index 1075ac24c8c5..c6303a8a12d2 100644
--- a/net/mac80211/main.c
+++ b/net/mac80211/main.c
@@ -33,6 +33,8 @@
 #include "led.h"
 #include "debugfs.h"
 
+struct kmem_cache *ieee80211_crypto_bufs_cache;
+
 void ieee80211_configure_filter(struct ieee80211_local *local)
 {
 	u64 mc;
@@ -1234,6 +1236,10 @@ static int __init ieee80211_init(void)
 	BUILD_BUG_ON(offsetof(struct ieee80211_tx_info, driver_data) +
 		     IEEE80211_TX_INFO_DRIVER_DATA_SIZE > sizeof(skb->cb));
 
+	ieee80211_crypto_bufs_cache = KMEM_CACHE(ieee80211_crypto_bufs, 0);
+	if (!ieee80211_crypto_bufs_cache)
+		return -ENOMEM;
+
 	ret = rc80211_minstrel_init();
 	if (ret)
 		return ret;
@@ -1264,6 +1270,8 @@ static void __exit ieee80211_exit(void)
 
 	ieee80211_iface_exit();
 
+	kmem_cache_destroy(ieee80211_crypto_bufs_cache);
+
 	rcu_barrier();
 }
 
diff --git a/net/mac80211/wpa.c b/net/mac80211/wpa.c
index 2e366438f8ef..ee9105d57545 100644
--- a/net/mac80211/wpa.c
+++ b/net/mac80211/wpa.c
@@ -405,8 +405,6 @@ static int ccmp_encrypt_skb(struct ieee80211_tx_data *tx, struct sk_buff *skb,
 	u8 *pos;
 	u8 pn[6];
 	u64 pn64;
-	u8 aad[2 * AES_BLOCK_SIZE];
-	u8 b_0[AES_BLOCK_SIZE];
 
 	if (info->control.hw_key &&
 	    !(info->control.hw_key->flags & IEEE80211_KEY_FLAG_GENERATE_IV) &&
@@ -456,13 +454,34 @@ static int ccmp_encrypt_skb(struct ieee80211_tx_data *tx, struct sk_buff *skb,
 	ccmp_pn2hdr(pos, pn, key->conf.keyidx);
 
 	/* hwaccel - with software CCMP header */
-	if (info->control.hw_key)
-		return 0;
+	if (unlikely(!info->control.hw_key)) {
+		struct ieee80211_crypto_bufs *bufs;
+		int err;
+		u8 *aad;
+		u8 *b_0;
+
+		bufs = kmem_cache_alloc(ieee80211_crypto_bufs_cache,
+					GFP_ATOMIC);
+		if (!bufs)
+			return -ENOMEM;
+
+		BUILD_BUG_ON(sizeof(bufs->buf1) < 2 * AES_BLOCK_SIZE);
+		BUILD_BUG_ON(sizeof(bufs->buf2) < AES_BLOCK_SIZE);
+
+		aad = bufs->buf1;
+		b_0 = bufs->buf2;
+
+		pos += IEEE80211_CCMP_HDR_LEN;
+		ccmp_special_blocks(skb, pn, b_0, aad);
+		err = ieee80211_aes_ccm_encrypt(key->u.ccmp.tfm, b_0, aad,
+						pos, len,
+						skb_put(skb, mic_len),
+						mic_len);
+		kmem_cache_free(ieee80211_crypto_bufs_cache, bufs);
+		return err;
+	}
 
-	pos += IEEE80211_CCMP_HDR_LEN;
-	ccmp_special_blocks(skb, pn, b_0, aad);
-	return ieee80211_aes_ccm_encrypt(key->u.ccmp.tfm, b_0, aad, pos, len,
-					 skb_put(skb, mic_len), mic_len);
+	return 0;
 }
 
 
@@ -532,16 +551,33 @@ ieee80211_crypto_ccmp_decrypt(struct ieee80211_rx_data *rx,
 		}
 
 		if (!(status->flag & RX_FLAG_DECRYPTED)) {
-			u8 aad[2 * AES_BLOCK_SIZE];
-			u8 b_0[AES_BLOCK_SIZE];
+			struct ieee80211_crypto_bufs *bufs;
+			int err;
+			u8 *aad;
+			u8 *b_0;
+
+			bufs = kmem_cache_alloc(ieee80211_crypto_bufs_cache,
+						GFP_ATOMIC);
+			if (!bufs)
+				return -ENOMEM;
+
+			BUILD_BUG_ON(sizeof(bufs->buf1) < 2 * AES_BLOCK_SIZE);
+			BUILD_BUG_ON(sizeof(bufs->buf2) < AES_BLOCK_SIZE);
+
+			aad = bufs->buf1;
+			b_0 = bufs->buf2;
+
 			/* hardware didn't decrypt/verify MIC */
 			ccmp_special_blocks(skb, pn, b_0, aad);
 
-			if (ieee80211_aes_ccm_decrypt(
+			err = ieee80211_aes_ccm_decrypt(
 				    key->u.ccmp.tfm, b_0, aad,
 				    skb->data + hdrlen + IEEE80211_CCMP_HDR_LEN,
 				    data_len,
-				    skb->data + skb->len - mic_len, mic_len))
+				    skb->data + skb->len - mic_len, mic_len);
+			kmem_cache_free(ieee80211_crypto_bufs_cache, bufs);
+
+			if (err)
 				return RX_DROP_UNUSABLE;
 		}
 
@@ -637,8 +673,6 @@ static int gcmp_encrypt_skb(struct ieee80211_tx_data *tx, struct sk_buff *skb)
 	u8 *pos;
 	u8 pn[6];
 	u64 pn64;
-	u8 aad[2 * AES_BLOCK_SIZE];
-	u8 j_0[AES_BLOCK_SIZE];
 
 	if (info->control.hw_key &&
 	    !(info->control.hw_key->flags & IEEE80211_KEY_FLAG_GENERATE_IV) &&
@@ -689,13 +723,34 @@ static int gcmp_encrypt_skb(struct ieee80211_tx_data *tx, struct sk_buff *skb)
 	gcmp_pn2hdr(pos, pn, key->conf.keyidx);
 
 	/* hwaccel - with software GCMP header */
-	if (info->control.hw_key)
-		return 0;
+	if (unlikely(!info->control.hw_key)) {
+		struct ieee80211_crypto_bufs *bufs;
+		int err;
+		u8 *aad;
+		u8 *j_0;
+		u8 *mic = skb_put(skb, IEEE80211_GCMP_MIC_LEN);
+
+		bufs = kmem_cache_alloc(ieee80211_crypto_bufs_cache,
+					GFP_ATOMIC);
+		if (!bufs)
+			return -ENOMEM;
+
+		BUILD_BUG_ON(sizeof(bufs->buf1) < 2 * AES_BLOCK_SIZE);
+		BUILD_BUG_ON(sizeof(bufs->buf2) < AES_BLOCK_SIZE);
+
+		aad = bufs->buf1;
+		j_0 = bufs->buf2;
+
+		pos += IEEE80211_GCMP_HDR_LEN;
+		gcmp_special_blocks(skb, pn, j_0, aad);
+		err = ieee80211_aes_gcm_encrypt(key->u.gcmp.tfm, j_0, aad,
+						pos, len, mic);
+		kmem_cache_free(ieee80211_crypto_bufs_cache, bufs);
+
+		return err;
+	}
 
-	pos += IEEE80211_GCMP_HDR_LEN;
-	gcmp_special_blocks(skb, pn, j_0, aad);
-	return ieee80211_aes_gcm_encrypt(key->u.gcmp.tfm, j_0, aad, pos, len,
-					 skb_put(skb, IEEE80211_GCMP_MIC_LEN));
+	return 0;
 }
 
 ieee80211_tx_result
@@ -760,17 +815,34 @@ ieee80211_crypto_gcmp_decrypt(struct ieee80211_rx_data *rx)
 		}
 
 		if (!(status->flag & RX_FLAG_DECRYPTED)) {
-			u8 aad[2 * AES_BLOCK_SIZE];
-			u8 j_0[AES_BLOCK_SIZE];
+			struct ieee80211_crypto_bufs *bufs;
+			int err;
+			u8 *aad;
+			u8 *j_0;
+
+			bufs = kmem_cache_alloc(ieee80211_crypto_bufs_cache,
+						GFP_ATOMIC);
+			if (!bufs)
+				return -ENOMEM;
+
+			BUILD_BUG_ON(sizeof(bufs->buf1) < 2 * AES_BLOCK_SIZE);
+			BUILD_BUG_ON(sizeof(bufs->buf2) < AES_BLOCK_SIZE);
+
+			aad = bufs->buf1;
+			j_0 = bufs->buf2;
+
 			/* hardware didn't decrypt/verify MIC */
 			gcmp_special_blocks(skb, pn, j_0, aad);
 
-			if (ieee80211_aes_gcm_decrypt(
+			err = ieee80211_aes_gcm_decrypt(
 				    key->u.gcmp.tfm, j_0, aad,
 				    skb->data + hdrlen + IEEE80211_GCMP_HDR_LEN,
 				    data_len,
 				    skb->data + skb->len -
-				    IEEE80211_GCMP_MIC_LEN))
+				    IEEE80211_GCMP_MIC_LEN);
+			kmem_cache_free(ieee80211_crypto_bufs_cache, bufs);
+
+			if (err)
 				return RX_DROP_UNUSABLE;
 		}
 
@@ -1119,9 +1191,12 @@ ieee80211_crypto_aes_gmac_encrypt(struct ieee80211_tx_data *tx)
 	struct ieee80211_key *key = tx->key;
 	struct ieee80211_mmie_16 *mmie;
 	struct ieee80211_hdr *hdr;
-	u8 aad[20];
 	u64 pn64;
 	u8 nonce[12];
+	struct ieee80211_crypto_bufs *bufs;
+	int err;
+	u8 *aad;
+	u8 *zero;
 
 	if (WARN_ON(skb_queue_len(&tx->skbs) != 1))
 		return TX_DROP;
@@ -1136,6 +1211,16 @@ ieee80211_crypto_aes_gmac_encrypt(struct ieee80211_tx_data *tx)
 	if (WARN_ON(skb_tailroom(skb) < sizeof(*mmie)))
 		return TX_DROP;
 
+	bufs = kmem_cache_alloc(ieee80211_crypto_bufs_cache, GFP_ATOMIC);
+	if (!bufs)
+		return TX_DROP;
+
+	BUILD_BUG_ON(sizeof(bufs->buf1) < GMAC_AAD_LEN);
+	BUILD_BUG_ON(sizeof(bufs->buf2) < GMAC_MIC_LEN);
+
+	aad = bufs->buf1;
+	zero = bufs->buf2;
+
 	mmie = (struct ieee80211_mmie_16 *)skb_put(skb, sizeof(*mmie));
 	mmie->element_id = WLAN_EID_MMIE;
 	mmie->length = sizeof(*mmie) - 2;
@@ -1153,11 +1238,13 @@ ieee80211_crypto_aes_gmac_encrypt(struct ieee80211_tx_data *tx)
 	bip_ipn_swap(nonce + ETH_ALEN, mmie->sequence_number);
 
 	/* MIC = AES-GMAC(IGTK, AAD || Management Frame Body || MMIE, 128) */
-	if (ieee80211_aes_gmac(key->u.aes_gmac.tfm, aad, nonce,
-			       skb->data + 24, skb->len - 24, mmie->mic) < 0)
-		return TX_DROP;
+	err = ieee80211_aes_gmac(key->u.aes_gmac.tfm, aad, nonce,
+				 skb->data + 24, skb->len - 24, mmie->mic,
+				 zero);
 
-	return TX_CONTINUE;
+	kmem_cache_free(ieee80211_crypto_bufs_cache, bufs);
+
+	return err < 0 ? TX_DROP : TX_CONTINUE;
 }
 
 ieee80211_rx_result
@@ -1167,7 +1254,7 @@ ieee80211_crypto_aes_gmac_decrypt(struct ieee80211_rx_data *rx)
 	struct ieee80211_rx_status *status = IEEE80211_SKB_RXCB(skb);
 	struct ieee80211_key *key = rx->key;
 	struct ieee80211_mmie_16 *mmie;
-	u8 aad[20], mic[16], ipn[6], nonce[12];
+	u8 mic[16], ipn[6], nonce[12];
 	struct ieee80211_hdr *hdr = (struct ieee80211_hdr *)skb->data;
 
 	if (!ieee80211_is_mgmt(hdr->frame_control))
@@ -1192,16 +1279,34 @@ ieee80211_crypto_aes_gmac_decrypt(struct ieee80211_rx_data *rx)
 	}
 
 	if (!(status->flag & RX_FLAG_DECRYPTED)) {
+		struct ieee80211_crypto_bufs *bufs;
+		int err;
+		u8 *aad;
+		u8 *zero;
+
+		bufs = kmem_cache_alloc(ieee80211_crypto_bufs_cache,
+					GFP_ATOMIC);
+		if (!bufs)
+			return RX_DROP_UNUSABLE;
+
+		BUILD_BUG_ON(sizeof(bufs->buf1) < GMAC_AAD_LEN);
+		BUILD_BUG_ON(sizeof(bufs->buf2) < GMAC_MIC_LEN);
+
+		aad = bufs->buf1;
+		zero = bufs->buf2;
+
 		/* hardware didn't decrypt/verify MIC */
 		bip_aad(skb, aad);
 
 		memcpy(nonce, hdr->addr2, ETH_ALEN);
 		memcpy(nonce + ETH_ALEN, ipn, 6);
 
-		if (ieee80211_aes_gmac(key->u.aes_gmac.tfm, aad, nonce,
-				       skb->data + 24, skb->len - 24,
-				       mic) < 0 ||
-		    memcmp(mic, mmie->mic, sizeof(mmie->mic)) != 0) {
+		err = ieee80211_aes_gmac(key->u.aes_gmac.tfm, aad, nonce,
+					 skb->data + 24, skb->len - 24,
+					 mic, zero);
+		kmem_cache_free(ieee80211_crypto_bufs_cache, bufs);
+
+		if (err < 0 || memcmp(mic, mmie->mic, sizeof(mmie->mic)) != 0) {
 			key->u.aes_gmac.icverrors++;
 			return RX_DROP_UNUSABLE;
 		}
-- 
2.8.1

Re: [PATCH v4] mac80211: move extra crypto data off the stack

[Ard Biesheuvel]

On 17 October 2016 at 09:33, Johannes Berg <johannes@sipsolutions.net> wrote:
> From: Johannes Berg <johannes.berg@intel.com>
>
> As the stack can (on x86-64) now be virtually mapped rather than
> using "normal" kernel memory, Sergey noticed mac80211 isn't using
> the SG APIs correctly by putting on-stack buffers into SG tables.
> This leads to kernel crashes.
>
> Fix this by allocating the extra fields dynamically on the fly as
> needed, using a kmem cache.
>
> I used per-CPU memory in a previous iteration of this patch, but
> Ard Biesheuvel pointed out that was also vmalloc'ed on some
> architectures.
>
> Reported-by: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com>
> Signed-off-by: Johannes Berg <johannes.berg@intel.com>

Apologies for going back and forth on this, but it appears there may
be another way to deal with this.

First of all, we only need this handling for the authenticated data,
and only for CCM and GCM, not CMAC (which does not use scatterlists at
all, it simply calls the AES cipher directly)

So that leaves a fixed 20 bytes for GCM and fixed 32 bytes for CCM,
which we could allocate along with the AEAD request, e..g.,

"""
diff --git a/net/mac80211/aes_ccm.c b/net/mac80211/aes_ccm.c
index 8e898a6e8de8..c0c33e6ad94e 100644
--- a/net/mac80211/aes_ccm.c
+++ b/net/mac80211/aes_ccm.c
@@ -24,13 +24,17 @@ int ieee80211_aes_ccm_encrypt(struct crypto_aead
*tfm, u8 *b_0, u8 *aad,
 {
        struct scatterlist sg[3];
        struct aead_request *aead_req;
+       u8 *__aad;

        aead_req = aead_request_alloc(tfm, GFP_ATOMIC);
        if (!aead_req)
                return -ENOMEM;

+       __aad = (u8 *)aead_req + crypto_aead_reqsize(tfm);
+       memcpy(__aad, aad, 2 * AES_BLOCK_SIZE);
+
        sg_init_table(sg, 3);
-       sg_set_buf(&sg[0], &aad[2], be16_to_cpup((__be16 *)aad));
+       sg_set_buf(&sg[0], &__aad[2], be16_to_cpup((__be16 *)__aad));
        sg_set_buf(&sg[1], data, data_len);
        sg_set_buf(&sg[2], mic, mic_len);

@@ -49,6 +53,7 @@ int ieee80211_aes_ccm_decrypt(struct crypto_aead
*tfm, u8 *b_0, u8 *aad,
 {
        struct scatterlist sg[3];
        struct aead_request *aead_req;
+       u8 *__aad;
        int err;

        if (data_len == 0)
@@ -58,8 +63,11 @@ int ieee80211_aes_ccm_decrypt(struct crypto_aead
*tfm, u8 *b_0, u8 *aad,
        if (!aead_req)
                return -ENOMEM;

+       __aad = (u8 *)aead_req + crypto_aead_reqsize(tfm);
+       memcpy(__aad, aad, 2 * AES_BLOCK_SIZE);
+
        sg_init_table(sg, 3);
-       sg_set_buf(&sg[0], &aad[2], be16_to_cpup((__be16 *)aad));
+       sg_set_buf(&sg[0], &__aad[2], be16_to_cpup((__be16 *)__aad));
        sg_set_buf(&sg[1], data, data_len);
        sg_set_buf(&sg[2], mic, mic_len);

@@ -90,6 +98,8 @@ struct crypto_aead
*ieee80211_aes_key_setup_encrypt(const u8 key[],
        if (err)
                goto free_aead;

+       crypto_aead_set_reqsize(tfm,
+                               crypto_aead_reqsize(tfm) + 2 * AES_BLOCK_SIZE));
        return tfm;

 free_aead:
"""

Re: [PATCH v4] mac80211: move extra crypto data off the stack

[Ard Biesheuvel]

On 17 October 2016 at 10:14, Ard Biesheuvel <ard.biesheuvel@linaro.org> wrote:
> On 17 October 2016 at 09:33, Johannes Berg <johannes@sipsolutions.net> wrote:
>> From: Johannes Berg <johannes.berg@intel.com>
>>
>> As the stack can (on x86-64) now be virtually mapped rather than
>> using "normal" kernel memory, Sergey noticed mac80211 isn't using
>> the SG APIs correctly by putting on-stack buffers into SG tables.
>> This leads to kernel crashes.
>>
>> Fix this by allocating the extra fields dynamically on the fly as
>> needed, using a kmem cache.
>>
>> I used per-CPU memory in a previous iteration of this patch, but
>> Ard Biesheuvel pointed out that was also vmalloc'ed on some
>> architectures.
>>
>> Reported-by: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com>
>> Signed-off-by: Johannes Berg <johannes.berg@intel.com>
>
> Apologies for going back and forth on this, but it appears there may
> be another way to deal with this.
>
> First of all, we only need this handling for the authenticated data,
> and only for CCM and GCM, not CMAC (which does not use scatterlists at
> all, it simply calls the AES cipher directly)
>
> So that leaves a fixed 20 bytes for GCM and fixed 32 bytes for CCM,
> which we could allocate along with the AEAD request, e..g.,
>
> """
> diff --git a/net/mac80211/aes_ccm.c b/net/mac80211/aes_ccm.c
> index 8e898a6e8de8..c0c33e6ad94e 100644
> --- a/net/mac80211/aes_ccm.c
> +++ b/net/mac80211/aes_ccm.c
> @@ -24,13 +24,17 @@ int ieee80211_aes_ccm_encrypt(struct crypto_aead
> *tfm, u8 *b_0, u8 *aad,
>  {
>         struct scatterlist sg[3];
>         struct aead_request *aead_req;
> +       u8 *__aad;
>
>         aead_req = aead_request_alloc(tfm, GFP_ATOMIC);
>         if (!aead_req)
>                 return -ENOMEM;
>
> +       __aad = (u8 *)aead_req + crypto_aead_reqsize(tfm);
> +       memcpy(__aad, aad, 2 * AES_BLOCK_SIZE);
> +
>         sg_init_table(sg, 3);
> -       sg_set_buf(&sg[0], &aad[2], be16_to_cpup((__be16 *)aad));
> +       sg_set_buf(&sg[0], &__aad[2], be16_to_cpup((__be16 *)__aad));
>         sg_set_buf(&sg[1], data, data_len);
>         sg_set_buf(&sg[2], mic, mic_len);
>
> @@ -49,6 +53,7 @@ int ieee80211_aes_ccm_decrypt(struct crypto_aead
> *tfm, u8 *b_0, u8 *aad,
>  {
>         struct scatterlist sg[3];
>         struct aead_request *aead_req;
> +       u8 *__aad;
>         int err;
>
>         if (data_len == 0)
> @@ -58,8 +63,11 @@ int ieee80211_aes_ccm_decrypt(struct crypto_aead
> *tfm, u8 *b_0, u8 *aad,
>         if (!aead_req)
>                 return -ENOMEM;
>
> +       __aad = (u8 *)aead_req + crypto_aead_reqsize(tfm);
> +       memcpy(__aad, aad, 2 * AES_BLOCK_SIZE);
> +
>         sg_init_table(sg, 3);
> -       sg_set_buf(&sg[0], &aad[2], be16_to_cpup((__be16 *)aad));
> +       sg_set_buf(&sg[0], &__aad[2], be16_to_cpup((__be16 *)__aad));
>         sg_set_buf(&sg[1], data, data_len);
>         sg_set_buf(&sg[2], mic, mic_len);
>
> @@ -90,6 +98,8 @@ struct crypto_aead
> *ieee80211_aes_key_setup_encrypt(const u8 key[],
>         if (err)
>                 goto free_aead;
>
> +       crypto_aead_set_reqsize(tfm,
> +                               crypto_aead_reqsize(tfm) + 2 * AES_BLOCK_SIZE));
>         return tfm;
>

Darn, it seems crypto_aead_set_reqsize() is internal to the crypto API ... :-(

Re: [PATCH v4] mac80211: move extra crypto data off the stack

[Johannes Berg]

> Apologies for going back and forth on this, but it appears there may
> be another way to deal with this.
> 
> First of all, we only need this handling for the authenticated data,

Are you sure b_0/j_0 aren't needed? We pass those
to aead_request_set_crypt(), and I wasn't sure what that really did
internally, perhaps like the internal data.

Testing with that on the stack does seem to work, in fact.

Surely we need zero for GMAC though, since we also put that into the sg
list. Thus for GMAC we definitely need 20+16 bytes, and since I round
up to a cacheline (at least on SMP) it doesn't really matter that we
could get 36 instead of the 48 I have now.

> and only for CCM and GCM, not CMAC (which does not use scatterlists
> at all, it simply calls the AES cipher directly)

I didn't modify CMAC, I think, only GMAC, which also uses scatterlists.

> So that leaves a fixed 20 bytes for GCM and fixed 32 bytes for CCM,

and 36 for GMAC :)

johannes

Re: [PATCH v4] mac80211: move extra crypto data off the stack

[Ard Biesheuvel]

On 17 October 2016 at 10:23, Johannes Berg <johannes@sipsolutions.net> wrote:
>
>> Apologies for going back and forth on this, but it appears there may
>> be another way to deal with this.
>>
>> First of all, we only need this handling for the authenticated data,
>
> Are you sure b_0/j_0 aren't needed? We pass those
> to aead_request_set_crypt(), and I wasn't sure what that really did
> internally, perhaps like the internal data.
>

They are the IV[], which is a fixed length parameter of the algorithm.
In contrast, the AAD[] could be of arbitrary length (from the POV of
the crypto API) so it uses scatterlists.

> Testing with that on the stack does seem to work, in fact.
>
> Surely we need zero for GMAC though, since we also put that into the sg
> list. Thus for GMAC we definitely need 20+16 bytes, and since I round
> up to a cacheline (at least on SMP) it doesn't really matter that we
> could get 36 instead of the 48 I have now.
>
>> and only for CCM and GCM, not CMAC (which does not use scatterlists
>> at all, it simply calls the AES cipher directly)
>
> I didn't modify CMAC, I think, only GMAC, which also uses scatterlists.
>

Ah ok, I misread the patch.

>> So that leaves a fixed 20 bytes for GCM and fixed 32 bytes for CCM,
>
> and 36 for GMAC :)

Yes. But as I replied, setting the req size is not supported atm,
although it is reasonable to demand a way to allocate additional data
in the request specifically for this issue. So let's proceed with the
aead_request_alloc/free patch, but I would like to propose something
on the API side to address this particular issue

Re: [PATCH v4] mac80211: move extra crypto data off the stack

[Johannes Berg]

On Mon, 2016-10-17 at 10:30 +0100, Ard Biesheuvel wrote:

> Yes. But as I replied, setting the req size is not supported atm,
> although it is reasonable to demand a way to allocate additional data
> in the request specifically for this issue. So let's proceed with the
> aead_request_alloc/free patch, but I would like to propose something
> on the API side to address this particular issue

Well, if your other patch to make it OK to be on-stack would be applied
instead, this wouldn't make much sense either :)

In this particular patch, we could reduce the size of the struct, but I
don't actually think it'll make a difference to go from 48 to 36 bytes,
given alignment etc., so I think I'll just leave it as is.

johannes

Re: [PATCH v4] mac80211: move extra crypto data off the stack

[Ard Biesheuvel]

> On 17 Oct 2016, at 10:35, Johannes Berg <johannes@sipsolutions.net> wrote:
> 
>> On Mon, 2016-10-17 at 10:30 +0100, Ard Biesheuvel wrote:
>> 
>> Yes. But as I replied, setting the req size is not supported atm,
>> although it is reasonable to demand a way to allocate additional data
>> in the request specifically for this issue. So let's proceed with the
>> aead_request_alloc/free patch, but I would like to propose something
>> on the API side to address this particular issue
> 
> Well, if your other patch to make it OK to be on-stack would be applied
> instead, this wouldn't make much sense either :)
> 
> In this particular patch, we could reduce the size of the struct, but I
> don't actually think it'll make a difference to go from 48 to 36 bytes,
> given alignment etc., so I think I'll just leave it as is.
> 
> johannes

Re: [PATCH v4] mac80211: move extra crypto data off the stack

[Ard Biesheuvel]

> On 17 Oct 2016, at 10:35, Johannes Berg <johannes@sipsolutions.net> wrote:=

>=20
>> On Mon, 2016-10-17 at 10:30 +0100, Ard Biesheuvel wrote:
>>=20
>> Yes. But as I replied, setting the req size is not supported atm,
>> although it is reasonable to demand a way to allocate additional data
>> in the request specifically for this issue. So let's proceed with the
>> aead_request_alloc/free patch, but I would like to propose something
>> on the API side to address this particular issue
>=20
> Well, if your other patch to make it OK to be on-stack would be applied
> instead, this wouldn't make much sense either :)
>=20

Yes but that one only fixes ccm not gcm

> In this particular patch, we could reduce the size of the struct, but I
> don't actually think it'll make a difference to go from 48 to 36 bytes,
> given alignment etc., so I think I'll just leave it as is.
>=20

I understand you are in a hurry, but this is unlikely to be the only such is=
sue. I will propose an 'auxdata' feature for the crypto api that can be used=
 here, but also for any other occurrence where client data assoiciated with t=
he request can no longer be allocated on the stack=

Re: [PATCH v4] mac80211: move extra crypto data off the stack

[Johannes Berg]

> > Well, if your other patch to make it OK to be on-stack would be
> > applied instead, this wouldn't make much sense either :)
> > 
> 
> Yes but that one only fixes ccm not gcm

Yes, but we can do the same for GCM, no?

> > In this particular patch, we could reduce the size of the struct,
> > but I
> > don't actually think it'll make a difference to go from 48 to 36
> > bytes,
> > given alignment etc., so I think I'll just leave it as is.
> > 
> 
> I understand you are in a hurry, but this is unlikely to be the only
> such issue. I will propose an 'auxdata' feature for the crypto api
> that can be used here, but also for any other occurrence where client
> data assoiciated with the request can no longer be allocated on the
> stack

No objections. I'll merge this anyway today I think, reverting is easy
later.

johannes

Re: [PATCH v4] mac80211: move extra crypto data off the stack

[Ard Biesheuvel]

> On 17 Oct 2016, at 10:54, Johannes Berg <johannes@sipsolutions.net> wrote:=

>=20
>=20
>>> Well, if your other patch to make it OK to be on-stack would be
>>> applied instead, this wouldn't make much sense either :)
>>>=20
>>=20
>> Yes but that one only fixes ccm not gcm
>=20
> Yes, but we can do the same for GCM, no?
>=20

No, not really. ccm happens to use aes with the same key for the mac and the=
 encryption. gcm uses an another algo entirely for the mac

>>> In this particular patch, we could reduce the size of the struct,
>>> but I
>>> don't actually think it'll make a difference to go from 48 to 36
>>> bytes,
>>> given alignment etc., so I think I'll just leave it as is.
>>>=20
>>=20
>> I understand you are in a hurry, but this is unlikely to be the only
>> such issue. I will propose an 'auxdata' feature for the crypto api
>> that can be used here, but also for any other occurrence where client
>> data assoiciated with the request can no longer be allocated on the
>> stack
>=20
> No objections. I'll merge this anyway today I think, reverting is easy
> later.
>=20

ok fair enough=

Re: [PATCH v4] mac80211: move extra crypto data off the stack

[Ard Biesheuvel]

On 17 October 2016 at 11:02, Ard Biesheuvel <ard.biesheuvel@linaro.org> wrote:
>
>
>> On 17 Oct 2016, at 10:54, Johannes Berg <johannes@sipsolutions.net> wrote:
>>
>>
>>>> Well, if your other patch to make it OK to be on-stack would be
>>>> applied instead, this wouldn't make much sense either :)
>>>>
>>>
>>> Yes but that one only fixes ccm not gcm
>>
>> Yes, but we can do the same for GCM, no?
>>
>
> No, not really. ccm happens to use aes with the same key for the mac and the encryption. gcm uses an another algo entirely for the mac
>
>>>> In this particular patch, we could reduce the size of the struct,
>>>> but I
>>>> don't actually think it'll make a difference to go from 48 to 36
>>>> bytes,
>>>> given alignment etc., so I think I'll just leave it as is.
>>>>
>>>
>>> I understand you are in a hurry, but this is unlikely to be the only
>>> such issue. I will propose an 'auxdata' feature for the crypto api
>>> that can be used here, but also for any other occurrence where client
>>> data assoiciated with the request can no longer be allocated on the
>>> stack
>>
>> No objections. I'll merge this anyway today I think, reverting is easy
>> later.
>>
>
> ok fair enough

Actually, while I think it will be worthwhile going forward to
implement such an 'auxiliary data' feature in a generic way, I still
think we should address the issue at hand without too much
complication.

If we pedal back to the version of 'mac80211: move struct aead_req off
the stack' that uses kzalloc() instead of aead_request_alloc(), we can
simply add some space for aad[] and/or zero[], and get rid of the kmem
cache entirely.

If you're past this point already, i won't bother but otherwise I can
rework 'mac80211: move struct aead_req off the stack' so that the
other patch is no longer required (and IIRC, this is actually
something you proposed yourself a couple of iterations ago?)

Re: [PATCH v4] mac80211: move extra crypto data off the stack

[Johannes Berg]

On Mon, 2016-10-17 at 14:06 +0100, Ard Biesheuvel wrote:
> 
> Actually, while I think it will be worthwhile going forward to
> implement such an 'auxiliary data' feature in a generic way, I still
> think we should address the issue at hand without too much
> complication.
> 
> If we pedal back to the version of 'mac80211: move struct aead_req
> off the stack' that uses kzalloc() instead of aead_request_alloc(),
> we can simply add some space for aad[] and/or zero[], and get rid of
> the kmem cache entirely.
> 
> If you're past this point already, i won't bother but otherwise I can
> rework 'mac80211: move struct aead_req off the stack' so that the
> other patch is no longer required (and IIRC, this is actually
> something you proposed yourself a couple of iterations ago?)

Yes, I did consider that.

It makes some sense, and I guess the extra memcpy() would be cheaper
than the extra alloc?

I'd happily use that instead of the combination of my two patches. The
aead_request_alloc() is just a simple inline anyway, so no real problem
not using it.

johannes

Re: [PATCH v4] mac80211: move extra crypto data off the stack

[Ard Biesheuvel]

On 17 October 2016 at 14:16, Johannes Berg <johannes@sipsolutions.net> wrote:
> On Mon, 2016-10-17 at 14:06 +0100, Ard Biesheuvel wrote:
>>
>> Actually, while I think it will be worthwhile going forward to
>> implement such an 'auxiliary data' feature in a generic way, I still
>> think we should address the issue at hand without too much
>> complication.
>>
>> If we pedal back to the version of 'mac80211: move struct aead_req
>> off the stack' that uses kzalloc() instead of aead_request_alloc(),
>> we can simply add some space for aad[] and/or zero[], and get rid of
>> the kmem cache entirely.
>>
>> If you're past this point already, i won't bother but otherwise I can
>> rework 'mac80211: move struct aead_req off the stack' so that the
>> other patch is no longer required (and IIRC, this is actually
>> something you proposed yourself a couple of iterations ago?)
>
> Yes, I did consider that.
>
> It makes some sense, and I guess the extra memcpy() would be cheaper
> than the extra alloc?
>
> I'd happily use that instead of the combination of my two patches. The
> aead_request_alloc() is just a simple inline anyway, so no real problem
> not using it.
>

Indeed. And it keeps the clutter inside the aes_xxx.c files, which
could easily be updated in the future to use some auxdata feature if
it ever materializes.

I think it would help this code, but also the ESP code you pointed
out, to have some kind of 'ordered synchronous' CRYPTO_xxx flag, where
the crypto API could manage the kmem cache and percpu pointers to
allocations. This goes well beyond what we can do as a fix, though, so
we need an intermediate solution in any case.

Shall I propose the patch?

Re: [PATCH v4] mac80211: move extra crypto data off the stack

[Johannes Berg]

> Indeed. And it keeps the clutter inside the aes_xxx.c files, which
> could easily be updated in the future to use some auxdata feature if
> it ever materializes.
> 
> I think it would help this code, but also the ESP code you pointed
> out, to have some kind of 'ordered synchronous' CRYPTO_xxx flag,
> where the crypto API could manage the kmem cache and percpu pointers
> to allocations.

Yeah, could be useful to have that more generally.

>  This goes well beyond what we can do as a fix, though, so we need an
> intermediate solution in any case.
> 
> Shall I propose the patch?

I assume you mean a mac80211 patch - sure, I'll take that instead of
the two I have now.

johannes