Hi Ard, On Fri, Jun 14, 2019 at 11:29:22AM +0200, Ard Biesheuvel wrote: > -static void ccmp_init_blocks(struct crypto_cipher *tfm, > - struct ieee80211_hdr *hdr, > - u8 * pn, size_t dlen, u8 * b0, u8 * auth, u8 * s0) > +static void ccmp_init_blocks(struct ieee80211_hdr *hdr, > + u8 * pn, size_t dlen, u8 * b0, u8 * aad) > { > u8 *pos, qc = 0; > size_t aad_len; > int a4_included, qc_included; > - u8 aad[2 * AES_BLOCK_LEN]; > > a4_included = ieee80211_has_a4(hdr->frame_control); > qc_included = ieee80211_is_data_qos(hdr->frame_control); > @@ -131,17 +123,19 @@ static void ccmp_init_blocks(struct crypto_cipher *tfm, > aad_len += 2; > } > > - /* CCM Initial Block: > - * Flag (Include authentication header, M=3 (8-octet MIC), > - * L=1 (2-octet Dlen)) > - * Nonce: 0x00 | A2 | PN > - * Dlen */ > - b0[0] = 0x59; > + /* In CCM, the initial vectors (IV) used for CTR mode encryption and CBC > + * mode authentication are not allowed to collide, yet both are derived > + * from this vector b0. We only set L := 1 here to indicate that the > + * data size can be represented in (L+1) bytes. The CCM layer will take > + * care of storing the data length in the top (L+1) bytes and setting > + * and clearing the other bits as is required to derive the two IVs. > + */ > + b0[0] = 0x1; > + > + /* Nonce: QC | A2 | PN */ > b0[1] = qc; > memcpy(b0 + 2, hdr->addr2, ETH_ALEN); > memcpy(b0 + 8, pn, CCMP_PN_LEN); > - b0[14] = (dlen >> 8) & 0xff; > - b0[15] = dlen & 0xff; > > /* AAD: > * FC with bits 4..6 and 11..13 masked to zero; 14 is always one > @@ -166,16 +160,6 @@ static void ccmp_init_blocks(struct crypto_cipher *tfm, > aad[a4_included ? 30 : 24] = qc; > /* rest of QC masked */ > } > - > - /* Start with the first block and AAD */ > - lib80211_ccmp_aes_encrypt(tfm, b0, auth); > - xor_block(auth, aad, AES_BLOCK_LEN); > - lib80211_ccmp_aes_encrypt(tfm, auth, auth); > - xor_block(auth, &aad[AES_BLOCK_LEN], AES_BLOCK_LEN); > - lib80211_ccmp_aes_encrypt(tfm, auth, auth); > - b0[0] &= 0x07; > - b0[14] = b0[15] = 0; > - lib80211_ccmp_aes_encrypt(tfm, b0, s0); > } How about shifting the contents of aad over by 2 bytes and returning the AAD length from this function instead? It's confusing to still manually format the AAD length for CCM mode, when actually it's ignored now. Also I suggest fixing up the naming: ccmp_init_blocks() => ccmp_init_iv_and_aad() b0 => iv > > static int lib80211_ccmp_hdr(struct sk_buff *skb, int hdr_len, > @@ -218,13 +202,13 @@ static int lib80211_ccmp_hdr(struct sk_buff *skb, int hdr_len, > static int lib80211_ccmp_encrypt(struct sk_buff *skb, int hdr_len, void *priv) > { > struct lib80211_ccmp_data *key = priv; > - int data_len, i, blocks, last, len; > - u8 *pos, *mic; > struct ieee80211_hdr *hdr; > - u8 *b0 = key->tx_b0; > - u8 *b = key->tx_b; > - u8 *e = key->tx_e; > - u8 *s0 = key->tx_s0; > + struct aead_request *req; > + struct scatterlist sg[2]; > + u8 *aad = key->tx_aad; > + u8 b0[AES_BLOCK_LEN]; > + int len, data_len; > + int ret; > > if (skb_tailroom(skb) < CCMP_MIC_LEN || skb->len < hdr_len) > return -1; > @@ -234,31 +218,29 @@ static int lib80211_ccmp_encrypt(struct sk_buff *skb, int hdr_len, void *priv) > if (len < 0) > return -1; > > - pos = skb->data + hdr_len + CCMP_HDR_LEN; > + req = kzalloc(sizeof(*req) + crypto_aead_reqsize(key->tfm), GFP_ATOMIC); > + if (!req) > + return -ENOMEM; > + Why not kzalloc() and kzfree() instead of aead_request_alloc() and aead_request_free()? Same in lib80211_ccmp_decrypt(). Otherwise this patch looks good, though I'd like for someone to test it. Thanks for doing this! - Eric