]> CyberLeo.Net >> Repos - FreeBSD/FreeBSD.git/blob - sys/dev/cxgbe/tom/t4_tls.c
MFV r353141 (by phillip):
[FreeBSD/FreeBSD.git] / sys / dev / cxgbe / tom / t4_tls.c
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2017-2018 Chelsio Communications, Inc.
5  * All rights reserved.
6  * Written by: John Baldwin <jhb@FreeBSD.org>
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29
30 #include "opt_inet.h"
31 #include "opt_kern_tls.h"
32
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
35
36 #include <sys/param.h>
37 #include <sys/ktr.h>
38 #ifdef KERN_TLS
39 #include <sys/ktls.h>
40 #endif
41 #include <sys/sglist.h>
42 #include <sys/socket.h>
43 #include <sys/socketvar.h>
44 #include <sys/systm.h>
45 #include <netinet/in.h>
46 #include <netinet/in_pcb.h>
47 #include <netinet/tcp_var.h>
48 #include <netinet/toecore.h>
49 #ifdef KERN_TLS
50 #include <opencrypto/cryptodev.h>
51 #include <opencrypto/xform.h>
52 #endif
53
54 #ifdef TCP_OFFLOAD
55 #include "common/common.h"
56 #include "common/t4_tcb.h"
57 #include "crypto/t4_crypto.h"
58 #include "tom/t4_tom_l2t.h"
59 #include "tom/t4_tom.h"
60
61 /*
62  * The TCP sequence number of a CPL_TLS_DATA mbuf is saved here while
63  * the mbuf is in the ulp_pdu_reclaimq.
64  */
65 #define tls_tcp_seq     PH_loc.thirtytwo[0]
66
67 /*
68  * Handshake lock used for the handshake timer.  Having a global lock
69  * is perhaps not ideal, but it avoids having to use callout_drain()
70  * in tls_uninit_toep() which can't block.  Also, the timer shouldn't
71  * actually fire for most connections.
72  */
73 static struct mtx tls_handshake_lock;
74
75 static void
76 t4_set_tls_tcb_field(struct toepcb *toep, uint16_t word, uint64_t mask,
77     uint64_t val)
78 {
79         struct adapter *sc = td_adapter(toep->td);
80
81         t4_set_tcb_field(sc, toep->ofld_txq, toep, word, mask, val, 0, 0);
82 }
83
84 /* TLS and DTLS common routines */
85 bool
86 can_tls_offload(struct adapter *sc)
87 {
88
89         return (sc->tt.tls && sc->cryptocaps & FW_CAPS_CONFIG_TLSKEYS);
90 }
91
92 int
93 tls_tx_key(struct toepcb *toep)
94 {
95         struct tls_ofld_info *tls_ofld = &toep->tls;
96
97         return (tls_ofld->tx_key_addr >= 0);
98 }
99
100 int
101 tls_rx_key(struct toepcb *toep)
102 {
103         struct tls_ofld_info *tls_ofld = &toep->tls;
104
105         return (tls_ofld->rx_key_addr >= 0);
106 }
107
108 static int
109 key_size(struct toepcb *toep)
110 {
111         struct tls_ofld_info *tls_ofld = &toep->tls;
112
113         return ((tls_ofld->key_location == TLS_SFO_WR_CONTEXTLOC_IMMEDIATE) ?
114                 tls_ofld->k_ctx.tx_key_info_size : KEY_IN_DDR_SIZE);
115 }
116
117 /* Set TLS Key-Id in TCB */
118 static void
119 t4_set_tls_keyid(struct toepcb *toep, unsigned int key_id)
120 {
121
122         t4_set_tls_tcb_field(toep, W_TCB_RX_TLS_KEY_TAG,
123                          V_TCB_RX_TLS_KEY_TAG(M_TCB_RX_TLS_BUF_TAG),
124                          V_TCB_RX_TLS_KEY_TAG(key_id));
125 }
126
127 /* Clear TF_RX_QUIESCE to re-enable receive. */
128 static void
129 t4_clear_rx_quiesce(struct toepcb *toep)
130 {
131
132         t4_set_tls_tcb_field(toep, W_TCB_T_FLAGS, V_TF_RX_QUIESCE(1), 0);
133 }
134
135 static void
136 tls_clr_ofld_mode(struct toepcb *toep)
137 {
138
139         tls_stop_handshake_timer(toep);
140
141         /* Operate in PDU extraction mode only. */
142         t4_set_tls_tcb_field(toep, W_TCB_ULP_RAW,
143             V_TCB_ULP_RAW(M_TCB_ULP_RAW),
144             V_TCB_ULP_RAW(V_TF_TLS_ENABLE(1)));
145         t4_clear_rx_quiesce(toep);
146 }
147
148 static void
149 tls_clr_quiesce(struct toepcb *toep)
150 {
151
152         tls_stop_handshake_timer(toep);
153         t4_clear_rx_quiesce(toep);
154 }
155
156 /*
157  * Calculate the TLS data expansion size
158  */
159 static int
160 tls_expansion_size(struct toepcb *toep, int data_len, int full_pdus_only,
161     unsigned short *pdus_per_ulp)
162 {
163         struct tls_ofld_info *tls_ofld = &toep->tls;
164         struct tls_scmd *scmd = &tls_ofld->scmd0;
165         int expn_size = 0, frag_count = 0, pad_per_pdu = 0,
166             pad_last_pdu = 0, last_frag_size = 0, max_frag_size = 0;
167         int exp_per_pdu = 0;
168         int hdr_len = TLS_HEADER_LENGTH;
169
170         do {
171                 max_frag_size = tls_ofld->k_ctx.frag_size;
172                 if (G_SCMD_CIPH_MODE(scmd->seqno_numivs) ==
173                    SCMD_CIPH_MODE_AES_GCM) {
174                         frag_count = (data_len / max_frag_size);
175                         exp_per_pdu = GCM_TAG_SIZE + AEAD_EXPLICIT_DATA_SIZE +
176                                 hdr_len;
177                         expn_size =  frag_count * exp_per_pdu;
178                         if (full_pdus_only) {
179                                 *pdus_per_ulp = data_len / (exp_per_pdu +
180                                         max_frag_size);
181                                 if (*pdus_per_ulp > 32)
182                                         *pdus_per_ulp = 32;
183                                 else if(!*pdus_per_ulp)
184                                         *pdus_per_ulp = 1;
185                                 expn_size = (*pdus_per_ulp) * exp_per_pdu;
186                                 break;
187                         }
188                         if ((last_frag_size = data_len % max_frag_size) > 0) {
189                                 frag_count += 1;
190                                 expn_size += exp_per_pdu;
191                         }
192                         break;
193                 } else if (G_SCMD_CIPH_MODE(scmd->seqno_numivs) !=
194                            SCMD_CIPH_MODE_NOP) {
195                         /* Calculate the number of fragments we can make */
196                         frag_count  = (data_len / max_frag_size);
197                         if (frag_count > 0) {
198                                 pad_per_pdu = (((howmany((max_frag_size +
199                                                        tls_ofld->mac_length),
200                                                       CIPHER_BLOCK_SIZE)) *
201                                                 CIPHER_BLOCK_SIZE) -
202                                                (max_frag_size +
203                                                 tls_ofld->mac_length));
204                                 if (!pad_per_pdu)
205                                         pad_per_pdu = CIPHER_BLOCK_SIZE;
206                                 exp_per_pdu = pad_per_pdu +
207                                         tls_ofld->mac_length +
208                                         hdr_len + CIPHER_BLOCK_SIZE;
209                                 expn_size = frag_count * exp_per_pdu;
210                         }
211                         if (full_pdus_only) {
212                                 *pdus_per_ulp = data_len / (exp_per_pdu +
213                                         max_frag_size);
214                                 if (*pdus_per_ulp > 32)
215                                         *pdus_per_ulp = 32;
216                                 else if (!*pdus_per_ulp)
217                                         *pdus_per_ulp = 1;
218                                 expn_size = (*pdus_per_ulp) * exp_per_pdu;
219                                 break;
220                         }
221                         /* Consider the last fragment */
222                         if ((last_frag_size = data_len % max_frag_size) > 0) {
223                                 pad_last_pdu = (((howmany((last_frag_size +
224                                                         tls_ofld->mac_length),
225                                                        CIPHER_BLOCK_SIZE)) *
226                                                  CIPHER_BLOCK_SIZE) -
227                                                 (last_frag_size +
228                                                  tls_ofld->mac_length));
229                                 if (!pad_last_pdu)
230                                         pad_last_pdu = CIPHER_BLOCK_SIZE;
231                                 expn_size += (pad_last_pdu +
232                                               tls_ofld->mac_length + hdr_len +
233                                               CIPHER_BLOCK_SIZE);
234                         }
235                 }
236         } while (0);
237
238         return (expn_size);
239 }
240
241 /* Copy Key to WR */
242 static void
243 tls_copy_tx_key(struct toepcb *toep, void *dst)
244 {
245         struct tls_ofld_info *tls_ofld = &toep->tls;
246         struct ulptx_sc_memrd *sc_memrd;
247         struct ulptx_idata *sc;
248
249         if (tls_ofld->k_ctx.tx_key_info_size <= 0)
250                 return;
251
252         if (tls_ofld->key_location == TLS_SFO_WR_CONTEXTLOC_DDR) {
253                 sc = dst;
254                 sc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP));
255                 sc->len = htobe32(0);
256                 sc_memrd = (struct ulptx_sc_memrd *)(sc + 1);
257                 sc_memrd->cmd_to_len = htobe32(V_ULPTX_CMD(ULP_TX_SC_MEMRD) |
258                     V_ULP_TX_SC_MORE(1) |
259                     V_ULPTX_LEN16(tls_ofld->k_ctx.tx_key_info_size >> 4));
260                 sc_memrd->addr = htobe32(tls_ofld->tx_key_addr >> 5);
261         } else if (tls_ofld->key_location == TLS_SFO_WR_CONTEXTLOC_IMMEDIATE) {
262                 memcpy(dst, &tls_ofld->k_ctx.tx,
263                     tls_ofld->k_ctx.tx_key_info_size);
264         }
265 }
266
267 /* TLS/DTLS content type  for CPL SFO */
268 static inline unsigned char
269 tls_content_type(unsigned char content_type)
270 {
271         /*
272          * XXX: Shouldn't this map CONTENT_TYPE_APP_DATA to DATA and
273          * default to "CUSTOM" for all other types including
274          * heartbeat?
275          */
276         switch (content_type) {
277         case CONTENT_TYPE_CCS:
278                 return CPL_TX_TLS_SFO_TYPE_CCS;
279         case CONTENT_TYPE_ALERT:
280                 return CPL_TX_TLS_SFO_TYPE_ALERT;
281         case CONTENT_TYPE_HANDSHAKE:
282                 return CPL_TX_TLS_SFO_TYPE_HANDSHAKE;
283         case CONTENT_TYPE_HEARTBEAT:
284                 return CPL_TX_TLS_SFO_TYPE_HEARTBEAT;
285         }
286         return CPL_TX_TLS_SFO_TYPE_DATA;
287 }
288
289 static unsigned char
290 get_cipher_key_size(unsigned int ck_size)
291 {
292         switch (ck_size) {
293         case AES_NOP: /* NOP */
294                 return 15;
295         case AES_128: /* AES128 */
296                 return CH_CK_SIZE_128;
297         case AES_192: /* AES192 */
298                 return CH_CK_SIZE_192;
299         case AES_256: /* AES256 */
300                 return CH_CK_SIZE_256;
301         default:
302                 return CH_CK_SIZE_256;
303         }
304 }
305
306 static unsigned char
307 get_mac_key_size(unsigned int mk_size)
308 {
309         switch (mk_size) {
310         case SHA_NOP: /* NOP */
311                 return CH_MK_SIZE_128;
312         case SHA_GHASH: /* GHASH */
313         case SHA_512: /* SHA512 */
314                 return CH_MK_SIZE_512;
315         case SHA_224: /* SHA2-224 */
316                 return CH_MK_SIZE_192;
317         case SHA_256: /* SHA2-256*/
318                 return CH_MK_SIZE_256;
319         case SHA_384: /* SHA384 */
320                 return CH_MK_SIZE_512;
321         case SHA1: /* SHA1 */
322         default:
323                 return CH_MK_SIZE_160;
324         }
325 }
326
327 static unsigned int
328 get_proto_ver(int proto_ver)
329 {
330         switch (proto_ver) {
331         case TLS1_2_VERSION:
332                 return TLS_1_2_VERSION;
333         case TLS1_1_VERSION:
334                 return TLS_1_1_VERSION;
335         case DTLS1_2_VERSION:
336                 return DTLS_1_2_VERSION;
337         default:
338                 return TLS_VERSION_MAX;
339         }
340 }
341
342 static void
343 tls_rxkey_flit1(struct tls_keyctx *kwr, struct tls_key_context *kctx)
344 {
345
346         if (kctx->state.enc_mode == CH_EVP_CIPH_GCM_MODE) {
347                 kwr->u.rxhdr.ivinsert_to_authinsrt =
348                     htobe64(V_TLS_KEYCTX_TX_WR_IVINSERT(6ULL) |
349                         V_TLS_KEYCTX_TX_WR_AADSTRTOFST(1ULL) |
350                         V_TLS_KEYCTX_TX_WR_AADSTOPOFST(5ULL) |
351                         V_TLS_KEYCTX_TX_WR_AUTHSRTOFST(14ULL) |
352                         V_TLS_KEYCTX_TX_WR_AUTHSTOPOFST(16ULL) |
353                         V_TLS_KEYCTX_TX_WR_CIPHERSRTOFST(14ULL) |
354                         V_TLS_KEYCTX_TX_WR_CIPHERSTOPOFST(0ULL) |
355                         V_TLS_KEYCTX_TX_WR_AUTHINSRT(16ULL));
356                 kwr->u.rxhdr.ivpresent_to_rxmk_size &=
357                         ~(V_TLS_KEYCTX_TX_WR_RXOPAD_PRESENT(1));
358                 kwr->u.rxhdr.authmode_to_rxvalid &=
359                         ~(V_TLS_KEYCTX_TX_WR_CIPHAUTHSEQCTRL(1));
360         } else {
361                 kwr->u.rxhdr.ivinsert_to_authinsrt =
362                     htobe64(V_TLS_KEYCTX_TX_WR_IVINSERT(6ULL) |
363                         V_TLS_KEYCTX_TX_WR_AADSTRTOFST(1ULL) |
364                         V_TLS_KEYCTX_TX_WR_AADSTOPOFST(5ULL) |
365                         V_TLS_KEYCTX_TX_WR_AUTHSRTOFST(22ULL) |
366                         V_TLS_KEYCTX_TX_WR_AUTHSTOPOFST(0ULL) |
367                         V_TLS_KEYCTX_TX_WR_CIPHERSRTOFST(22ULL) |
368                         V_TLS_KEYCTX_TX_WR_CIPHERSTOPOFST(0ULL) |
369                         V_TLS_KEYCTX_TX_WR_AUTHINSRT(0ULL));
370         }
371 }
372
373 /* Rx key */
374 static void
375 prepare_rxkey_wr(struct tls_keyctx *kwr, struct tls_key_context *kctx)
376 {
377         unsigned int ck_size = kctx->cipher_secret_size;
378         unsigned int mk_size = kctx->mac_secret_size;
379         int proto_ver = kctx->proto_ver;
380
381         kwr->u.rxhdr.flitcnt_hmacctrl =
382                 ((kctx->tx_key_info_size >> 4) << 3) | kctx->hmac_ctrl;
383
384         kwr->u.rxhdr.protover_ciphmode =
385                 V_TLS_KEYCTX_TX_WR_PROTOVER(get_proto_ver(proto_ver)) |
386                 V_TLS_KEYCTX_TX_WR_CIPHMODE(kctx->state.enc_mode);
387
388         kwr->u.rxhdr.authmode_to_rxvalid =
389                 V_TLS_KEYCTX_TX_WR_AUTHMODE(kctx->state.auth_mode) |
390                 V_TLS_KEYCTX_TX_WR_CIPHAUTHSEQCTRL(1) |
391                 V_TLS_KEYCTX_TX_WR_SEQNUMCTRL(3) |
392                 V_TLS_KEYCTX_TX_WR_RXVALID(1);
393
394         kwr->u.rxhdr.ivpresent_to_rxmk_size =
395                 V_TLS_KEYCTX_TX_WR_IVPRESENT(0) |
396                 V_TLS_KEYCTX_TX_WR_RXOPAD_PRESENT(1) |
397                 V_TLS_KEYCTX_TX_WR_RXCK_SIZE(get_cipher_key_size(ck_size)) |
398                 V_TLS_KEYCTX_TX_WR_RXMK_SIZE(get_mac_key_size(mk_size));
399
400         tls_rxkey_flit1(kwr, kctx);
401
402         /* No key reversal for GCM */
403         if (kctx->state.enc_mode != CH_EVP_CIPH_GCM_MODE) {
404                 t4_aes_getdeckey(kwr->keys.edkey, kctx->rx.key,
405                                  (kctx->cipher_secret_size << 3));
406                 memcpy(kwr->keys.edkey + kctx->cipher_secret_size,
407                        kctx->rx.key + kctx->cipher_secret_size,
408                        (IPAD_SIZE + OPAD_SIZE));
409         } else {
410                 memcpy(kwr->keys.edkey, kctx->rx.key,
411                        (kctx->tx_key_info_size - SALT_SIZE));
412                 memcpy(kwr->u.rxhdr.rxsalt, kctx->rx.salt, SALT_SIZE);
413         }
414 }
415
416 /* Tx key */
417 static void
418 prepare_txkey_wr(struct tls_keyctx *kwr, struct tls_key_context *kctx)
419 {
420         unsigned int ck_size = kctx->cipher_secret_size;
421         unsigned int mk_size = kctx->mac_secret_size;
422
423         kwr->u.txhdr.ctxlen =
424                 (kctx->tx_key_info_size >> 4);
425         kwr->u.txhdr.dualck_to_txvalid =
426                 V_TLS_KEYCTX_TX_WR_TXOPAD_PRESENT(1) |
427                 V_TLS_KEYCTX_TX_WR_SALT_PRESENT(1) |
428                 V_TLS_KEYCTX_TX_WR_TXCK_SIZE(get_cipher_key_size(ck_size)) |
429                 V_TLS_KEYCTX_TX_WR_TXMK_SIZE(get_mac_key_size(mk_size)) |
430                 V_TLS_KEYCTX_TX_WR_TXVALID(1);
431
432         memcpy(kwr->keys.edkey, kctx->tx.key, HDR_KCTX_SIZE);
433         if (kctx->state.enc_mode == CH_EVP_CIPH_GCM_MODE) {
434                 memcpy(kwr->u.txhdr.txsalt, kctx->tx.salt, SALT_SIZE);
435                 kwr->u.txhdr.dualck_to_txvalid &=
436                         ~(V_TLS_KEYCTX_TX_WR_TXOPAD_PRESENT(1));
437         }
438         kwr->u.txhdr.dualck_to_txvalid = htons(kwr->u.txhdr.dualck_to_txvalid);
439 }
440
441 /* TLS Key memory management */
442 static int
443 get_new_keyid(struct toepcb *toep)
444 {
445         struct adapter *sc = td_adapter(toep->td);
446         vmem_addr_t addr;
447
448         if (vmem_alloc(sc->key_map, TLS_KEY_CONTEXT_SZ, M_NOWAIT | M_FIRSTFIT,
449             &addr) != 0)
450                 return (-1);
451
452         return (addr);
453 }
454
455 static void
456 free_keyid(struct toepcb *toep, int keyid)
457 {
458         struct adapter *sc = td_adapter(toep->td);
459
460         vmem_free(sc->key_map, keyid, TLS_KEY_CONTEXT_SZ);
461 }
462
463 static void
464 clear_tls_keyid(struct toepcb *toep)
465 {
466         struct tls_ofld_info *tls_ofld = &toep->tls;
467
468         if (tls_ofld->rx_key_addr >= 0) {
469                 free_keyid(toep, tls_ofld->rx_key_addr);
470                 tls_ofld->rx_key_addr = -1;
471         }
472         if (tls_ofld->tx_key_addr >= 0) {
473                 free_keyid(toep, tls_ofld->tx_key_addr);
474                 tls_ofld->tx_key_addr = -1;
475         }
476 }
477
478 static int
479 get_keyid(struct tls_ofld_info *tls_ofld, unsigned int ops)
480 {
481         return (ops & KEY_WRITE_RX ? tls_ofld->rx_key_addr :
482                 ((ops & KEY_WRITE_TX) ? tls_ofld->tx_key_addr : -1));
483 }
484
485 static int
486 get_tp_plen_max(struct tls_ofld_info *tls_ofld)
487 {
488         int plen = ((min(3*4096, TP_TX_PG_SZ))/1448) * 1448;
489
490         return (tls_ofld->k_ctx.frag_size <= 8192 ? plen : FC_TP_PLEN_MAX);
491 }
492
493 /* Send request to get the key-id */
494 static int
495 tls_program_key_id(struct toepcb *toep, struct tls_key_context *k_ctx)
496 {
497         struct tls_ofld_info *tls_ofld = &toep->tls;
498         struct adapter *sc = td_adapter(toep->td);
499         struct ofld_tx_sdesc *txsd;
500         int kwrlen, kctxlen, keyid, len;
501         struct wrqe *wr;
502         struct tls_key_req *kwr;
503         struct tls_keyctx *kctx;
504
505         kwrlen = sizeof(*kwr);
506         kctxlen = roundup2(sizeof(*kctx), 32);
507         len = roundup2(kwrlen + kctxlen, 16);
508
509         if (toep->txsd_avail == 0)
510                 return (EAGAIN);
511
512         /* Dont initialize key for re-neg */
513         if (!G_KEY_CLR_LOC(k_ctx->l_p_key)) {
514                 if ((keyid = get_new_keyid(toep)) < 0) {
515                         return (ENOSPC);
516                 }
517         } else {
518                 keyid = get_keyid(tls_ofld, k_ctx->l_p_key);
519         }
520
521         wr = alloc_wrqe(len, toep->ofld_txq);
522         if (wr == NULL) {
523                 free_keyid(toep, keyid);
524                 return (ENOMEM);
525         }
526         kwr = wrtod(wr);
527         memset(kwr, 0, kwrlen);
528
529         kwr->wr_hi = htobe32(V_FW_WR_OP(FW_ULPTX_WR) | F_FW_WR_COMPL |
530             F_FW_WR_ATOMIC);
531         kwr->wr_mid = htobe32(V_FW_WR_LEN16(DIV_ROUND_UP(len, 16)) |
532             V_FW_WR_FLOWID(toep->tid));
533         kwr->protocol = get_proto_ver(k_ctx->proto_ver);
534         kwr->mfs = htons(k_ctx->frag_size);
535         kwr->reneg_to_write_rx = k_ctx->l_p_key;
536
537         /* master command */
538         kwr->cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE) |
539             V_T5_ULP_MEMIO_ORDER(1) | V_T5_ULP_MEMIO_IMM(1));
540         kwr->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(kctxlen >> 5));
541         kwr->len16 = htobe32((toep->tid << 8) |
542             DIV_ROUND_UP(len - sizeof(struct work_request_hdr), 16));
543         kwr->kaddr = htobe32(V_ULP_MEMIO_ADDR(keyid >> 5));
544
545         /* sub command */
546         kwr->sc_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM));
547         kwr->sc_len = htobe32(kctxlen);
548
549         kctx = (struct tls_keyctx *)(kwr + 1);
550         memset(kctx, 0, kctxlen);
551
552         if (G_KEY_GET_LOC(k_ctx->l_p_key) == KEY_WRITE_TX) {
553                 tls_ofld->tx_key_addr = keyid;
554                 prepare_txkey_wr(kctx, k_ctx);
555         } else if (G_KEY_GET_LOC(k_ctx->l_p_key) == KEY_WRITE_RX) {
556                 tls_ofld->rx_key_addr = keyid;
557                 prepare_rxkey_wr(kctx, k_ctx);
558         }
559
560         txsd = &toep->txsd[toep->txsd_pidx];
561         txsd->tx_credits = DIV_ROUND_UP(len, 16);
562         txsd->plen = 0;
563         toep->tx_credits -= txsd->tx_credits;
564         if (__predict_false(++toep->txsd_pidx == toep->txsd_total))
565                 toep->txsd_pidx = 0;
566         toep->txsd_avail--;
567
568         t4_wrq_tx(sc, wr);
569
570         return (0);
571 }
572
573 /* Store a key received from SSL in DDR. */
574 static int
575 program_key_context(struct tcpcb *tp, struct toepcb *toep,
576     struct tls_key_context *uk_ctx)
577 {
578         struct adapter *sc = td_adapter(toep->td);
579         struct tls_ofld_info *tls_ofld = &toep->tls;
580         struct tls_key_context *k_ctx;
581         int error, key_offset;
582
583         if (tp->t_state != TCPS_ESTABLISHED) {
584                 /*
585                  * XXX: Matches Linux driver, but not sure this is a
586                  * very appropriate error.
587                  */
588                 return (ENOENT);
589         }
590
591         /* Stop timer on handshake completion */
592         tls_stop_handshake_timer(toep);
593
594         toep->flags &= ~TPF_FORCE_CREDITS;
595
596         CTR4(KTR_CXGBE, "%s: tid %d %s proto_ver %#x", __func__, toep->tid,
597             G_KEY_GET_LOC(uk_ctx->l_p_key) == KEY_WRITE_RX ? "KEY_WRITE_RX" :
598             "KEY_WRITE_TX", uk_ctx->proto_ver);
599
600         if (G_KEY_GET_LOC(uk_ctx->l_p_key) == KEY_WRITE_RX &&
601             ulp_mode(toep) != ULP_MODE_TLS)
602                 return (EOPNOTSUPP);
603
604         /* Don't copy the 'tx' and 'rx' fields. */
605         k_ctx = &tls_ofld->k_ctx;
606         memcpy(&k_ctx->l_p_key, &uk_ctx->l_p_key,
607             sizeof(*k_ctx) - offsetof(struct tls_key_context, l_p_key));
608
609         /* TLS version != 1.1 and !1.2 OR DTLS != 1.2 */
610         if (get_proto_ver(k_ctx->proto_ver) > DTLS_1_2_VERSION) {
611                 if (G_KEY_GET_LOC(k_ctx->l_p_key) == KEY_WRITE_RX) {
612                         tls_ofld->rx_key_addr = -1;
613                         t4_clear_rx_quiesce(toep);
614                 } else {
615                         tls_ofld->tx_key_addr = -1;
616                 }
617                 return (0);
618         }
619
620         if (k_ctx->state.enc_mode == CH_EVP_CIPH_GCM_MODE) {
621                 k_ctx->iv_size = 4;
622                 k_ctx->mac_first = 0;
623                 k_ctx->hmac_ctrl = 0;
624         } else {
625                 k_ctx->iv_size = 8; /* for CBC, iv is 16B, unit of 2B */
626                 k_ctx->mac_first = 1;
627         }
628
629         tls_ofld->scmd0.seqno_numivs =
630                 (V_SCMD_SEQ_NO_CTRL(3) |
631                  V_SCMD_PROTO_VERSION(get_proto_ver(k_ctx->proto_ver)) |
632                  V_SCMD_ENC_DEC_CTRL(SCMD_ENCDECCTRL_ENCRYPT) |
633                  V_SCMD_CIPH_AUTH_SEQ_CTRL((k_ctx->mac_first == 0)) |
634                  V_SCMD_CIPH_MODE(k_ctx->state.enc_mode) |
635                  V_SCMD_AUTH_MODE(k_ctx->state.auth_mode) |
636                  V_SCMD_HMAC_CTRL(k_ctx->hmac_ctrl) |
637                  V_SCMD_IV_SIZE(k_ctx->iv_size));
638
639         tls_ofld->scmd0.ivgen_hdrlen =
640                 (V_SCMD_IV_GEN_CTRL(k_ctx->iv_ctrl) |
641                  V_SCMD_KEY_CTX_INLINE(0) |
642                  V_SCMD_TLS_FRAG_ENABLE(1));
643
644         tls_ofld->mac_length = k_ctx->mac_secret_size;
645
646         if (G_KEY_GET_LOC(k_ctx->l_p_key) == KEY_WRITE_RX) {
647                 k_ctx->rx = uk_ctx->rx;
648                 /* Dont initialize key for re-neg */
649                 if (!G_KEY_CLR_LOC(k_ctx->l_p_key))
650                         tls_ofld->rx_key_addr = -1;
651         } else {
652                 k_ctx->tx = uk_ctx->tx;
653                 /* Dont initialize key for re-neg */
654                 if (!G_KEY_CLR_LOC(k_ctx->l_p_key))
655                         tls_ofld->tx_key_addr = -1;
656         }
657
658         /* Flush pending data before new Tx key becomes active */
659         if (G_KEY_GET_LOC(k_ctx->l_p_key) == KEY_WRITE_TX) {
660                 struct sockbuf *sb;
661
662                 /* XXX: This might not drain everything. */
663                 t4_push_frames(sc, toep, 0);
664                 sb = &toep->inp->inp_socket->so_snd;
665                 SOCKBUF_LOCK(sb);
666
667                 /* XXX: This asserts that everything has been pushed. */
668                 MPASS(sb->sb_sndptr == NULL || sb->sb_sndptr->m_next == NULL);
669                 sb->sb_sndptr = NULL;
670                 tls_ofld->sb_off = sbavail(sb);
671                 SOCKBUF_UNLOCK(sb);
672                 tls_ofld->tx_seq_no = 0;
673         }
674
675         if ((G_KEY_GET_LOC(k_ctx->l_p_key) == KEY_WRITE_RX) ||
676             (tls_ofld->key_location == TLS_SFO_WR_CONTEXTLOC_DDR)) {
677                 error = tls_program_key_id(toep, k_ctx);
678                 if (error) {
679                         /* XXX: Only clear quiesce for KEY_WRITE_RX? */
680                         t4_clear_rx_quiesce(toep);
681                         return (error);
682                 }
683         }
684
685         if (G_KEY_GET_LOC(k_ctx->l_p_key) == KEY_WRITE_RX) {
686                 /*
687                  * RX key tags are an index into the key portion of MA
688                  * memory stored as an offset from the base address in
689                  * units of 64 bytes.
690                  */
691                 key_offset = tls_ofld->rx_key_addr - sc->vres.key.start;
692                 t4_set_tls_keyid(toep, key_offset / 64);
693                 t4_set_tls_tcb_field(toep, W_TCB_ULP_RAW,
694                                  V_TCB_ULP_RAW(M_TCB_ULP_RAW),
695                                  V_TCB_ULP_RAW((V_TF_TLS_KEY_SIZE(3) |
696                                                 V_TF_TLS_CONTROL(1) |
697                                                 V_TF_TLS_ACTIVE(1) |
698                                                 V_TF_TLS_ENABLE(1))));
699                 t4_set_tls_tcb_field(toep, W_TCB_TLS_SEQ,
700                                  V_TCB_TLS_SEQ(M_TCB_TLS_SEQ),
701                                  V_TCB_TLS_SEQ(0));
702                 t4_clear_rx_quiesce(toep);
703         } else {
704                 unsigned short pdus_per_ulp;
705
706                 if (tls_ofld->key_location == TLS_SFO_WR_CONTEXTLOC_IMMEDIATE)
707                         tls_ofld->tx_key_addr = 1;
708
709                 tls_ofld->fcplenmax = get_tp_plen_max(tls_ofld);
710                 tls_ofld->expn_per_ulp = tls_expansion_size(toep,
711                                 tls_ofld->fcplenmax, 1, &pdus_per_ulp);
712                 tls_ofld->pdus_per_ulp = pdus_per_ulp;
713                 tls_ofld->adjusted_plen = tls_ofld->pdus_per_ulp *
714                         ((tls_ofld->expn_per_ulp/tls_ofld->pdus_per_ulp) +
715                          tls_ofld->k_ctx.frag_size);
716         }
717
718         return (0);
719 }
720
721 /*
722  * In some cases a client connection can hang without sending the
723  * ServerHelloDone message from the NIC to the host.  Send a dummy
724  * RX_DATA_ACK with RX_MODULATE to unstick the connection.
725  */
726 static void
727 tls_send_handshake_ack(void *arg)
728 {
729         struct toepcb *toep = arg;
730         struct tls_ofld_info *tls_ofld = &toep->tls;
731         struct adapter *sc = td_adapter(toep->td);
732
733         /*
734          * XXX: Does not have the t4_get_tcb() checks to refine the
735          * workaround.
736          */
737         callout_schedule(&tls_ofld->handshake_timer, TLS_SRV_HELLO_RD_TM * hz);
738
739         CTR2(KTR_CXGBE, "%s: tid %d sending RX_DATA_ACK", __func__, toep->tid);
740         send_rx_modulate(sc, toep);
741 }
742
743 static void
744 tls_start_handshake_timer(struct toepcb *toep)
745 {
746         struct tls_ofld_info *tls_ofld = &toep->tls;
747
748         mtx_lock(&tls_handshake_lock);
749         callout_reset(&tls_ofld->handshake_timer, TLS_SRV_HELLO_BKOFF_TM * hz,
750             tls_send_handshake_ack, toep);
751         mtx_unlock(&tls_handshake_lock);
752 }
753
754 void
755 tls_stop_handshake_timer(struct toepcb *toep)
756 {
757         struct tls_ofld_info *tls_ofld = &toep->tls;
758
759         mtx_lock(&tls_handshake_lock);
760         callout_stop(&tls_ofld->handshake_timer);
761         mtx_unlock(&tls_handshake_lock);
762 }
763
764 int
765 t4_ctloutput_tls(struct socket *so, struct sockopt *sopt)
766 {
767         struct tls_key_context uk_ctx;
768         struct inpcb *inp;
769         struct tcpcb *tp;
770         struct toepcb *toep;
771         int error, optval;
772
773         error = 0;
774         if (sopt->sopt_dir == SOPT_SET &&
775             sopt->sopt_name == TCP_TLSOM_SET_TLS_CONTEXT) {
776                 error = sooptcopyin(sopt, &uk_ctx, sizeof(uk_ctx),
777                     sizeof(uk_ctx));
778                 if (error)
779                         return (error);
780         }
781
782         inp = sotoinpcb(so);
783         KASSERT(inp != NULL, ("tcp_ctloutput: inp == NULL"));
784         INP_WLOCK(inp);
785         if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) {
786                 INP_WUNLOCK(inp);
787                 return (ECONNRESET);
788         }
789         tp = intotcpcb(inp);
790         toep = tp->t_toe;
791         switch (sopt->sopt_dir) {
792         case SOPT_SET:
793                 switch (sopt->sopt_name) {
794                 case TCP_TLSOM_SET_TLS_CONTEXT:
795                         if (toep->tls.mode == TLS_MODE_KTLS)
796                                 error = EINVAL;
797                         else {
798                                 error = program_key_context(tp, toep, &uk_ctx);
799                                 if (error == 0)
800                                         toep->tls.mode = TLS_MODE_TLSOM;
801                         }
802                         INP_WUNLOCK(inp);
803                         break;
804                 case TCP_TLSOM_CLR_TLS_TOM:
805                         if (toep->tls.mode == TLS_MODE_KTLS)
806                                 error = EINVAL;
807                         else if (ulp_mode(toep) == ULP_MODE_TLS) {
808                                 CTR2(KTR_CXGBE, "%s: tid %d CLR_TLS_TOM",
809                                     __func__, toep->tid);
810                                 tls_clr_ofld_mode(toep);
811                         } else
812                                 error = EOPNOTSUPP;
813                         INP_WUNLOCK(inp);
814                         break;
815                 case TCP_TLSOM_CLR_QUIES:
816                         if (toep->tls.mode == TLS_MODE_KTLS)
817                                 error = EINVAL;
818                         else if (ulp_mode(toep) == ULP_MODE_TLS) {
819                                 CTR2(KTR_CXGBE, "%s: tid %d CLR_QUIES",
820                                     __func__, toep->tid);
821                                 tls_clr_quiesce(toep);
822                         } else
823                                 error = EOPNOTSUPP;
824                         INP_WUNLOCK(inp);
825                         break;
826                 default:
827                         INP_WUNLOCK(inp);
828                         error = EOPNOTSUPP;
829                         break;
830                 }
831                 break;
832         case SOPT_GET:
833                 switch (sopt->sopt_name) {
834                 case TCP_TLSOM_GET_TLS_TOM:
835                         /*
836                          * TLS TX is permitted on any TOE socket, but
837                          * TLS RX requires a TLS ULP mode.
838                          */
839                         optval = TLS_TOM_NONE;
840                         if (can_tls_offload(td_adapter(toep->td)) &&
841                             toep->tls.mode != TLS_MODE_KTLS) {
842                                 switch (ulp_mode(toep)) {
843                                 case ULP_MODE_NONE:
844                                 case ULP_MODE_TCPDDP:
845                                         optval = TLS_TOM_TXONLY;
846                                         break;
847                                 case ULP_MODE_TLS:
848                                         optval = TLS_TOM_BOTH;
849                                         break;
850                                 }
851                         }
852                         CTR3(KTR_CXGBE, "%s: tid %d GET_TLS_TOM = %d",
853                             __func__, toep->tid, optval);
854                         INP_WUNLOCK(inp);
855                         error = sooptcopyout(sopt, &optval, sizeof(optval));
856                         break;
857                 default:
858                         INP_WUNLOCK(inp);
859                         error = EOPNOTSUPP;
860                         break;
861                 }
862                 break;
863         }
864         return (error);
865 }
866
867 #ifdef KERN_TLS
868 static void
869 init_ktls_key_context(struct ktls_session *tls, struct tls_key_context *k_ctx)
870 {
871         struct auth_hash *axf;
872         u_int mac_key_size;
873         char *hash;
874
875         k_ctx->l_p_key = V_KEY_GET_LOC(KEY_WRITE_TX);
876         if (tls->params.tls_vminor == TLS_MINOR_VER_ONE)
877                 k_ctx->proto_ver = SCMD_PROTO_VERSION_TLS_1_1;
878         else
879                 k_ctx->proto_ver = SCMD_PROTO_VERSION_TLS_1_2;
880         k_ctx->cipher_secret_size = tls->params.cipher_key_len;
881         k_ctx->tx_key_info_size = sizeof(struct tx_keyctx_hdr) +
882             k_ctx->cipher_secret_size;
883         memcpy(k_ctx->tx.key, tls->params.cipher_key,
884             tls->params.cipher_key_len);
885         hash = k_ctx->tx.key + tls->params.cipher_key_len;
886         if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16) {
887                 k_ctx->state.auth_mode = SCMD_AUTH_MODE_GHASH;
888                 k_ctx->state.enc_mode = SCMD_CIPH_MODE_AES_GCM;
889                 k_ctx->iv_size = 4;
890                 k_ctx->mac_first = 0;
891                 k_ctx->hmac_ctrl = SCMD_HMAC_CTRL_NOP;
892                 k_ctx->tx_key_info_size += GMAC_BLOCK_LEN;
893                 memcpy(k_ctx->tx.salt, tls->params.iv, SALT_SIZE);
894                 t4_init_gmac_hash(tls->params.cipher_key,
895                     tls->params.cipher_key_len * 8, hash);
896         } else {
897                 switch (tls->params.auth_algorithm) {
898                 case CRYPTO_SHA1_HMAC:
899                         axf = &auth_hash_hmac_sha1;
900                         mac_key_size = SHA1_HASH_LEN;
901                         k_ctx->state.auth_mode = SCMD_AUTH_MODE_SHA1;
902                         break;
903                 case CRYPTO_SHA2_256_HMAC:
904                         axf = &auth_hash_hmac_sha2_256;
905                         mac_key_size = SHA2_256_HASH_LEN;
906                         k_ctx->state.auth_mode = SCMD_AUTH_MODE_SHA256;
907                         break;
908                 case CRYPTO_SHA2_384_HMAC:
909                         axf = &auth_hash_hmac_sha2_384;
910                         mac_key_size = SHA2_512_HASH_LEN;
911                         k_ctx->state.auth_mode = SCMD_AUTH_MODE_SHA512_384;
912                         break;
913                 default:
914                         panic("bad auth mode");
915                 }
916                 k_ctx->state.enc_mode = SCMD_CIPH_MODE_AES_CBC;
917                 k_ctx->iv_size = 8; /* for CBC, iv is 16B, unit of 2B */
918                 k_ctx->mac_first = 1;
919                 k_ctx->hmac_ctrl = SCMD_HMAC_CTRL_NO_TRUNC;
920                 k_ctx->tx_key_info_size += roundup2(mac_key_size, 16) * 2;
921                 k_ctx->mac_secret_size = mac_key_size;
922                 t4_init_hmac_digest(axf, mac_key_size, tls->params.auth_key,
923                     tls->params.auth_key_len * 8, hash);
924         }
925
926         k_ctx->frag_size = tls->params.max_frame_len;
927         k_ctx->iv_ctrl = 1;
928 }
929
930 int
931 tls_alloc_ktls(struct toepcb *toep, struct ktls_session *tls)
932 {
933         struct tls_key_context *k_ctx;
934         int error;
935
936         if (toep->tls.mode == TLS_MODE_TLSOM)
937                 return (EINVAL);
938         if (!can_tls_offload(td_adapter(toep->td)))
939                 return (EINVAL);
940         switch (ulp_mode(toep)) {
941         case ULP_MODE_NONE:
942         case ULP_MODE_TCPDDP:
943                 break;
944         default:
945                 return (EINVAL);
946         }
947
948         switch (tls->params.cipher_algorithm) {
949         case CRYPTO_AES_CBC:
950                 /* XXX: Explicitly ignore any provided IV. */
951                 switch (tls->params.cipher_key_len) {
952                 case 128 / 8:
953                 case 192 / 8:
954                 case 256 / 8:
955                         break;
956                 default:
957                         return (EINVAL);
958                 }
959                 switch (tls->params.auth_algorithm) {
960                 case CRYPTO_SHA1_HMAC:
961                 case CRYPTO_SHA2_256_HMAC:
962                 case CRYPTO_SHA2_384_HMAC:
963                         break;
964                 default:
965                         return (EPROTONOSUPPORT);
966                 }
967                 break;
968         case CRYPTO_AES_NIST_GCM_16:
969                 if (tls->params.iv_len != SALT_SIZE)
970                         return (EINVAL);
971                 switch (tls->params.cipher_key_len) {
972                 case 128 / 8:
973                 case 192 / 8:
974                 case 256 / 8:
975                         break;
976                 default:
977                         return (EINVAL);
978                 }
979                 break;
980         default:
981                 return (EPROTONOSUPPORT);
982         }
983
984         /* Only TLS 1.1 and TLS 1.2 are currently supported. */
985         if (tls->params.tls_vmajor != TLS_MAJOR_VER_ONE ||
986             tls->params.tls_vminor < TLS_MINOR_VER_ONE ||
987             tls->params.tls_vminor > TLS_MINOR_VER_TWO)
988                 return (EPROTONOSUPPORT);
989
990         /*
991          * XXX: This assumes no key renegotation.  If KTLS ever supports
992          * that we will want to allocate TLS sessions dynamically rather
993          * than as a static member of toep.
994          */
995         k_ctx = &toep->tls.k_ctx;
996         init_ktls_key_context(tls, k_ctx);
997
998         toep->tls.scmd0.seqno_numivs =
999                 (V_SCMD_SEQ_NO_CTRL(3) |
1000                  V_SCMD_PROTO_VERSION(k_ctx->proto_ver) |
1001                  V_SCMD_ENC_DEC_CTRL(SCMD_ENCDECCTRL_ENCRYPT) |
1002                  V_SCMD_CIPH_AUTH_SEQ_CTRL((k_ctx->mac_first == 0)) |
1003                  V_SCMD_CIPH_MODE(k_ctx->state.enc_mode) |
1004                  V_SCMD_AUTH_MODE(k_ctx->state.auth_mode) |
1005                  V_SCMD_HMAC_CTRL(k_ctx->hmac_ctrl) |
1006                  V_SCMD_IV_SIZE(k_ctx->iv_size));
1007
1008         toep->tls.scmd0.ivgen_hdrlen =
1009                 (V_SCMD_IV_GEN_CTRL(k_ctx->iv_ctrl) |
1010                  V_SCMD_KEY_CTX_INLINE(0) |
1011                  V_SCMD_TLS_FRAG_ENABLE(1));
1012
1013         if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16)
1014                 toep->tls.iv_len = 8;
1015         else
1016                 toep->tls.iv_len = AES_BLOCK_LEN;
1017
1018         toep->tls.mac_length = k_ctx->mac_secret_size;
1019
1020         toep->tls.tx_key_addr = -1;
1021
1022         error = tls_program_key_id(toep, k_ctx);
1023         if (error)
1024                 return (error);
1025
1026         toep->tls.fcplenmax = get_tp_plen_max(&toep->tls);
1027         toep->tls.expn_per_ulp = tls->params.tls_hlen + tls->params.tls_tlen;
1028         toep->tls.pdus_per_ulp = 1;
1029         toep->tls.adjusted_plen = toep->tls.expn_per_ulp +
1030             toep->tls.k_ctx.frag_size;
1031
1032         toep->tls.mode = TLS_MODE_KTLS;
1033
1034         return (0);
1035 }
1036 #endif
1037
1038 void
1039 tls_init_toep(struct toepcb *toep)
1040 {
1041         struct tls_ofld_info *tls_ofld = &toep->tls;
1042
1043         tls_ofld->mode = TLS_MODE_OFF;
1044         tls_ofld->key_location = TLS_SFO_WR_CONTEXTLOC_DDR;
1045         tls_ofld->rx_key_addr = -1;
1046         tls_ofld->tx_key_addr = -1;
1047         if (ulp_mode(toep) == ULP_MODE_TLS)
1048                 callout_init_mtx(&tls_ofld->handshake_timer,
1049                     &tls_handshake_lock, 0);
1050 }
1051
1052 void
1053 tls_establish(struct toepcb *toep)
1054 {
1055
1056         /*
1057          * Enable PDU extraction.
1058          *
1059          * XXX: Supposedly this should be done by the firmware when
1060          * the ULP_MODE FLOWC parameter is set in send_flowc_wr(), but
1061          * in practice this seems to be required.
1062          */
1063         CTR2(KTR_CXGBE, "%s: tid %d setting TLS_ENABLE", __func__, toep->tid);
1064         t4_set_tls_tcb_field(toep, W_TCB_ULP_RAW, V_TCB_ULP_RAW(M_TCB_ULP_RAW),
1065             V_TCB_ULP_RAW(V_TF_TLS_ENABLE(1)));
1066
1067         toep->flags |= TPF_FORCE_CREDITS;
1068
1069         tls_start_handshake_timer(toep);
1070 }
1071
1072 void
1073 tls_uninit_toep(struct toepcb *toep)
1074 {
1075
1076         if (ulp_mode(toep) == ULP_MODE_TLS)
1077                 tls_stop_handshake_timer(toep);
1078         clear_tls_keyid(toep);
1079 }
1080
1081 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16)
1082 #define MIN_OFLD_TLSTX_CREDITS(toep)                                    \
1083         (howmany(sizeof(struct fw_tlstx_data_wr) +                      \
1084             sizeof(struct cpl_tx_tls_sfo) + key_size((toep)) +          \
1085             CIPHER_BLOCK_SIZE + 1, 16))
1086
1087 static inline u_int
1088 max_imm_tls_space(int tx_credits)
1089 {
1090         const int n = 2;        /* Use only up to 2 desc for imm. data WR */
1091         int space;
1092
1093         KASSERT(tx_credits >= 0 &&
1094                 tx_credits <= MAX_OFLD_TX_CREDITS,
1095                 ("%s: %d credits", __func__, tx_credits));
1096
1097         if (tx_credits >= (n * EQ_ESIZE) / 16)
1098                 space = (n * EQ_ESIZE);
1099         else
1100                 space = tx_credits * 16;
1101         return (space);
1102 }
1103
1104 static int
1105 count_mbuf_segs(struct mbuf *m, int skip, int len, int *max_nsegs_1mbufp)
1106 {
1107         int max_nsegs_1mbuf, n, nsegs;
1108
1109         while (skip >= m->m_len) {
1110                 skip -= m->m_len;
1111                 m = m->m_next;
1112         }
1113
1114         nsegs = 0;
1115         max_nsegs_1mbuf = 0;
1116         while (len > 0) {
1117                 n = sglist_count(mtod(m, char *) + skip, m->m_len - skip);
1118                 if (n > max_nsegs_1mbuf)
1119                         max_nsegs_1mbuf = n;
1120                 nsegs += n;
1121                 len -= m->m_len - skip;
1122                 skip = 0;
1123                 m = m->m_next;
1124         }
1125         *max_nsegs_1mbufp = max_nsegs_1mbuf;
1126         return (nsegs);
1127 }
1128
1129 static void
1130 write_tlstx_wr(struct fw_tlstx_data_wr *txwr, struct toepcb *toep,
1131     unsigned int immdlen, unsigned int plen, unsigned int expn,
1132     unsigned int pdus, uint8_t credits, int shove, int imm_ivs)
1133 {
1134         struct tls_ofld_info *tls_ofld = &toep->tls;
1135         unsigned int len = plen + expn;
1136
1137         txwr->op_to_immdlen = htobe32(V_WR_OP(FW_TLSTX_DATA_WR) |
1138             V_FW_TLSTX_DATA_WR_COMPL(1) |
1139             V_FW_TLSTX_DATA_WR_IMMDLEN(immdlen));
1140         txwr->flowid_len16 = htobe32(V_FW_TLSTX_DATA_WR_FLOWID(toep->tid) |
1141             V_FW_TLSTX_DATA_WR_LEN16(credits));
1142         txwr->plen = htobe32(len);
1143         txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(ULP_MODE_TLS) |
1144             V_TX_URG(0) | /* F_T6_TX_FORCE | */ V_TX_SHOVE(shove));
1145         txwr->ctxloc_to_exp = htobe32(V_FW_TLSTX_DATA_WR_NUMIVS(pdus) |
1146             V_FW_TLSTX_DATA_WR_EXP(expn) |
1147             V_FW_TLSTX_DATA_WR_CTXLOC(tls_ofld->key_location) |
1148             V_FW_TLSTX_DATA_WR_IVDSGL(!imm_ivs) |
1149             V_FW_TLSTX_DATA_WR_KEYSIZE(tls_ofld->k_ctx.tx_key_info_size >> 4));
1150         txwr->mfs = htobe16(tls_ofld->k_ctx.frag_size);
1151         txwr->adjustedplen_pkd = htobe16(
1152             V_FW_TLSTX_DATA_WR_ADJUSTEDPLEN(tls_ofld->adjusted_plen));
1153         txwr->expinplenmax_pkd = htobe16(
1154             V_FW_TLSTX_DATA_WR_EXPINPLENMAX(tls_ofld->expn_per_ulp));
1155         txwr->pdusinplenmax_pkd = 
1156             V_FW_TLSTX_DATA_WR_PDUSINPLENMAX(tls_ofld->pdus_per_ulp);
1157 }
1158
1159 static void
1160 write_tlstx_cpl(struct cpl_tx_tls_sfo *cpl, struct toepcb *toep,
1161     struct tls_hdr *tls_hdr, unsigned int plen, unsigned int pdus)
1162 {
1163         struct tls_ofld_info *tls_ofld = &toep->tls;
1164         int data_type, seglen;
1165
1166         if (plen < tls_ofld->k_ctx.frag_size)
1167                 seglen = plen;
1168         else
1169                 seglen = tls_ofld->k_ctx.frag_size;
1170         data_type = tls_content_type(tls_hdr->type);
1171         cpl->op_to_seg_len = htobe32(V_CPL_TX_TLS_SFO_OPCODE(CPL_TX_TLS_SFO) |
1172             V_CPL_TX_TLS_SFO_DATA_TYPE(data_type) |
1173             V_CPL_TX_TLS_SFO_CPL_LEN(2) | V_CPL_TX_TLS_SFO_SEG_LEN(seglen));
1174         cpl->pld_len = htobe32(plen);
1175         if (data_type == CPL_TX_TLS_SFO_TYPE_HEARTBEAT)
1176                 cpl->type_protover = htobe32(
1177                     V_CPL_TX_TLS_SFO_TYPE(tls_hdr->type));
1178         cpl->seqno_numivs = htobe32(tls_ofld->scmd0.seqno_numivs |
1179             V_SCMD_NUM_IVS(pdus));
1180         cpl->ivgen_hdrlen = htobe32(tls_ofld->scmd0.ivgen_hdrlen);
1181         cpl->scmd1 = htobe64(tls_ofld->tx_seq_no);
1182         tls_ofld->tx_seq_no += pdus;
1183 }
1184
1185 /*
1186  * Similar to write_tx_sgl() except that it accepts an optional
1187  * trailer buffer for IVs.
1188  */
1189 static void
1190 write_tlstx_sgl(void *dst, struct mbuf *start, int skip, int plen,
1191     void *iv_buffer, int iv_len, int nsegs, int n)
1192 {
1193         struct mbuf *m;
1194         struct ulptx_sgl *usgl = dst;
1195         int i, j, rc;
1196         struct sglist sg;
1197         struct sglist_seg segs[n];
1198
1199         KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
1200
1201         sglist_init(&sg, n, segs);
1202         usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
1203             V_ULPTX_NSGE(nsegs));
1204
1205         for (m = start; skip >= m->m_len; m = m->m_next)
1206                 skip -= m->m_len;
1207
1208         i = -1;
1209         for (m = start; plen > 0; m = m->m_next) {
1210                 rc = sglist_append(&sg, mtod(m, char *) + skip,
1211                     m->m_len - skip);
1212                 if (__predict_false(rc != 0))
1213                         panic("%s: sglist_append %d", __func__, rc);
1214                 plen -= m->m_len - skip;
1215                 skip = 0;
1216
1217                 for (j = 0; j < sg.sg_nseg; i++, j++) {
1218                         if (i < 0) {
1219                                 usgl->len0 = htobe32(segs[j].ss_len);
1220                                 usgl->addr0 = htobe64(segs[j].ss_paddr);
1221                         } else {
1222                                 usgl->sge[i / 2].len[i & 1] =
1223                                     htobe32(segs[j].ss_len);
1224                                 usgl->sge[i / 2].addr[i & 1] =
1225                                     htobe64(segs[j].ss_paddr);
1226                         }
1227 #ifdef INVARIANTS
1228                         nsegs--;
1229 #endif
1230                 }
1231                 sglist_reset(&sg);
1232         }
1233         if (iv_buffer != NULL) {
1234                 rc = sglist_append(&sg, iv_buffer, iv_len);
1235                 if (__predict_false(rc != 0))
1236                         panic("%s: sglist_append %d", __func__, rc);
1237
1238                 for (j = 0; j < sg.sg_nseg; i++, j++) {
1239                         if (i < 0) {
1240                                 usgl->len0 = htobe32(segs[j].ss_len);
1241                                 usgl->addr0 = htobe64(segs[j].ss_paddr);
1242                         } else {
1243                                 usgl->sge[i / 2].len[i & 1] =
1244                                     htobe32(segs[j].ss_len);
1245                                 usgl->sge[i / 2].addr[i & 1] =
1246                                     htobe64(segs[j].ss_paddr);
1247                         }
1248 #ifdef INVARIANTS
1249                         nsegs--;
1250 #endif
1251                 }
1252         }
1253         if (i & 1)
1254                 usgl->sge[i / 2].len[1] = htobe32(0);
1255         KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, iv_buffer %p",
1256             __func__, nsegs, start, iv_buffer));
1257 }
1258
1259 /*
1260  * Similar to t4_push_frames() but handles TLS sockets when TLS offload
1261  * is enabled.  Rather than transmitting bulk data, the socket buffer
1262  * contains TLS records.  The work request requires a full TLS record,
1263  * so batch mbufs up until a full TLS record is seen.  This requires
1264  * reading the TLS header out of the start of each record to determine
1265  * its length.
1266  */
1267 void
1268 t4_push_tls_records(struct adapter *sc, struct toepcb *toep, int drop)
1269 {
1270         struct tls_hdr thdr;
1271         struct mbuf *sndptr;
1272         struct fw_tlstx_data_wr *txwr;
1273         struct cpl_tx_tls_sfo *cpl;
1274         struct wrqe *wr;
1275         u_int plen, nsegs, credits, space, max_nsegs_1mbuf, wr_len;
1276         u_int expn_size, iv_len, pdus, sndptroff;
1277         struct tls_ofld_info *tls_ofld = &toep->tls;
1278         struct inpcb *inp = toep->inp;
1279         struct tcpcb *tp = intotcpcb(inp);
1280         struct socket *so = inp->inp_socket;
1281         struct sockbuf *sb = &so->so_snd;
1282         int tls_size, tx_credits, shove, /* compl,*/ sowwakeup;
1283         struct ofld_tx_sdesc *txsd;
1284         bool imm_ivs, imm_payload;
1285         void *iv_buffer, *iv_dst, *buf;
1286
1287         INP_WLOCK_ASSERT(inp);
1288         KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1289             ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
1290
1291         KASSERT(ulp_mode(toep) == ULP_MODE_NONE ||
1292             ulp_mode(toep) == ULP_MODE_TCPDDP || ulp_mode(toep) == ULP_MODE_TLS,
1293             ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep));
1294         KASSERT(tls_tx_key(toep),
1295             ("%s: TX key not set for toep %p", __func__, toep));
1296
1297 #ifdef VERBOSE_TRACES
1298         CTR4(KTR_CXGBE, "%s: tid %d toep flags %#x tp flags %#x drop %d",
1299             __func__, toep->tid, toep->flags, tp->t_flags);
1300 #endif
1301         if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
1302                 return;
1303
1304 #ifdef RATELIMIT
1305         if (__predict_false(inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) &&
1306             (update_tx_rate_limit(sc, toep, so->so_max_pacing_rate) == 0)) {
1307                 inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED;
1308         }
1309 #endif
1310
1311         /*
1312          * This function doesn't resume by itself.  Someone else must clear the
1313          * flag and call this function.
1314          */
1315         if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
1316                 KASSERT(drop == 0,
1317                     ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
1318                 return;
1319         }
1320
1321         txsd = &toep->txsd[toep->txsd_pidx];
1322         for (;;) {
1323                 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
1324                 space = max_imm_tls_space(tx_credits);
1325                 wr_len = sizeof(struct fw_tlstx_data_wr) +
1326                     sizeof(struct cpl_tx_tls_sfo) + key_size(toep);
1327                 if (wr_len + CIPHER_BLOCK_SIZE + 1 > space) {
1328 #ifdef VERBOSE_TRACES
1329                         CTR5(KTR_CXGBE,
1330                             "%s: tid %d tx_credits %d min_wr %d space %d",
1331                             __func__, toep->tid, tx_credits, wr_len +
1332                             CIPHER_BLOCK_SIZE + 1, space);
1333 #endif
1334                         return;
1335                 }
1336
1337                 SOCKBUF_LOCK(sb);
1338                 sowwakeup = drop;
1339                 if (drop) {
1340                         sbdrop_locked(sb, drop);
1341                         MPASS(tls_ofld->sb_off >= drop);
1342                         tls_ofld->sb_off -= drop;
1343                         drop = 0;
1344                 }
1345
1346                 /*
1347                  * Send a FIN if requested, but only if there's no
1348                  * more data to send.
1349                  */
1350                 if (sbavail(sb) == tls_ofld->sb_off &&
1351                     toep->flags & TPF_SEND_FIN) {
1352                         if (sowwakeup)
1353                                 sowwakeup_locked(so);
1354                         else
1355                                 SOCKBUF_UNLOCK(sb);
1356                         SOCKBUF_UNLOCK_ASSERT(sb);
1357                         t4_close_conn(sc, toep);
1358                         return;
1359                 }
1360
1361                 if (sbavail(sb) < tls_ofld->sb_off + TLS_HEADER_LENGTH) {
1362                         /*
1363                          * A full TLS header is not yet queued, stop
1364                          * for now until more data is added to the
1365                          * socket buffer.  However, if the connection
1366                          * has been closed, we will never get the rest
1367                          * of the header so just discard the partial
1368                          * header and close the connection.
1369                          */
1370 #ifdef VERBOSE_TRACES
1371                         CTR5(KTR_CXGBE, "%s: tid %d sbavail %d sb_off %d%s",
1372                             __func__, toep->tid, sbavail(sb), tls_ofld->sb_off,
1373                             toep->flags & TPF_SEND_FIN ? "" : " SEND_FIN");
1374 #endif
1375                         if (sowwakeup)
1376                                 sowwakeup_locked(so);
1377                         else
1378                                 SOCKBUF_UNLOCK(sb);
1379                         SOCKBUF_UNLOCK_ASSERT(sb);
1380                         if (toep->flags & TPF_SEND_FIN)
1381                                 t4_close_conn(sc, toep);
1382                         return;
1383                 }
1384
1385                 /* Read the header of the next TLS record. */
1386                 sndptr = sbsndmbuf(sb, tls_ofld->sb_off, &sndptroff);
1387                 m_copydata(sndptr, sndptroff, sizeof(thdr), (caddr_t)&thdr);
1388                 tls_size = htons(thdr.length);
1389                 plen = TLS_HEADER_LENGTH + tls_size;
1390                 pdus = howmany(tls_size, tls_ofld->k_ctx.frag_size);
1391                 iv_len = pdus * CIPHER_BLOCK_SIZE;
1392
1393                 if (sbavail(sb) < tls_ofld->sb_off + plen) {
1394                         /*
1395                          * The full TLS record is not yet queued, stop
1396                          * for now until more data is added to the
1397                          * socket buffer.  However, if the connection
1398                          * has been closed, we will never get the rest
1399                          * of the record so just discard the partial
1400                          * record and close the connection.
1401                          */
1402 #ifdef VERBOSE_TRACES
1403                         CTR6(KTR_CXGBE,
1404                             "%s: tid %d sbavail %d sb_off %d plen %d%s",
1405                             __func__, toep->tid, sbavail(sb), tls_ofld->sb_off,
1406                             plen, toep->flags & TPF_SEND_FIN ? "" :
1407                             " SEND_FIN");
1408 #endif
1409                         if (sowwakeup)
1410                                 sowwakeup_locked(so);
1411                         else
1412                                 SOCKBUF_UNLOCK(sb);
1413                         SOCKBUF_UNLOCK_ASSERT(sb);
1414                         if (toep->flags & TPF_SEND_FIN)
1415                                 t4_close_conn(sc, toep);
1416                         return;
1417                 }
1418
1419                 /* Shove if there is no additional data pending. */
1420                 shove = (sbavail(sb) == tls_ofld->sb_off + plen) &&
1421                     !(tp->t_flags & TF_MORETOCOME);
1422
1423                 if (sb->sb_flags & SB_AUTOSIZE &&
1424                     V_tcp_do_autosndbuf &&
1425                     sb->sb_hiwat < V_tcp_autosndbuf_max &&
1426                     sbused(sb) >= sb->sb_hiwat * 7 / 8) {
1427                         int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
1428                             V_tcp_autosndbuf_max);
1429
1430                         if (!sbreserve_locked(sb, newsize, so, NULL))
1431                                 sb->sb_flags &= ~SB_AUTOSIZE;
1432                         else
1433                                 sowwakeup = 1;  /* room available */
1434                 }
1435                 if (sowwakeup)
1436                         sowwakeup_locked(so);
1437                 else
1438                         SOCKBUF_UNLOCK(sb);
1439                 SOCKBUF_UNLOCK_ASSERT(sb);
1440
1441                 if (__predict_false(toep->flags & TPF_FIN_SENT))
1442                         panic("%s: excess tx.", __func__);
1443
1444                 /* Determine whether to use immediate vs SGL. */
1445                 imm_payload = false;
1446                 imm_ivs = false;
1447                 if (wr_len + iv_len <= space) {
1448                         imm_ivs = true;
1449                         wr_len += iv_len;
1450                         if (wr_len + tls_size <= space) {
1451                                 wr_len += tls_size;
1452                                 imm_payload = true;
1453                         }
1454                 }
1455
1456                 /* Allocate space for IVs if needed. */
1457                 if (!imm_ivs) {
1458                         iv_buffer = malloc(iv_len, M_CXGBE, M_NOWAIT);
1459                         if (iv_buffer == NULL) {
1460                                 /*
1461                                  * XXX: How to restart this?
1462                                  */
1463                                 if (sowwakeup)
1464                                         sowwakeup_locked(so);
1465                                 else
1466                                         SOCKBUF_UNLOCK(sb);
1467                                 SOCKBUF_UNLOCK_ASSERT(sb);
1468                                 CTR3(KTR_CXGBE,
1469                             "%s: tid %d failed to alloc IV space len %d",
1470                                     __func__, toep->tid, iv_len);
1471                                 return;
1472                         }
1473                 } else
1474                         iv_buffer = NULL;
1475
1476                 /* Determine size of SGL. */
1477                 nsegs = 0;
1478                 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */
1479                 if (!imm_payload) {
1480                         nsegs = count_mbuf_segs(sndptr, sndptroff +
1481                             TLS_HEADER_LENGTH, tls_size, &max_nsegs_1mbuf);
1482                         if (!imm_ivs) {
1483                                 int n = sglist_count(iv_buffer, iv_len);
1484                                 nsegs += n;
1485                                 if (n > max_nsegs_1mbuf)
1486                                         max_nsegs_1mbuf = n;
1487                         }
1488
1489                         /* Account for SGL in work request length. */
1490                         wr_len += sizeof(struct ulptx_sgl) +
1491                             ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
1492                 }
1493
1494                 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
1495                 if (wr == NULL) {
1496                         /* XXX: how will we recover from this? */
1497                         toep->flags |= TPF_TX_SUSPENDED;
1498                         return;
1499                 }
1500
1501 #ifdef VERBOSE_TRACES
1502                 CTR5(KTR_CXGBE, "%s: tid %d TLS record %d len %#x pdus %d",
1503                     __func__, toep->tid, thdr.type, tls_size, pdus);
1504 #endif
1505                 txwr = wrtod(wr);
1506                 cpl = (struct cpl_tx_tls_sfo *)(txwr + 1);
1507                 memset(txwr, 0, roundup2(wr_len, 16));
1508                 credits = howmany(wr_len, 16);
1509                 expn_size = tls_expansion_size(toep, tls_size, 0, NULL);
1510                 write_tlstx_wr(txwr, toep, imm_payload ? tls_size : 0,
1511                     tls_size, expn_size, pdus, credits, shove, imm_ivs ? 1 : 0);
1512                 write_tlstx_cpl(cpl, toep, &thdr, tls_size, pdus);
1513                 tls_copy_tx_key(toep, cpl + 1);
1514
1515                 /* Generate random IVs */
1516                 buf = (char *)(cpl + 1) + key_size(toep);
1517                 if (imm_ivs) {
1518                         MPASS(iv_buffer == NULL);
1519                         iv_dst = buf;
1520                         buf = (char *)iv_dst + iv_len;
1521                 } else
1522                         iv_dst = iv_buffer;
1523                 arc4rand(iv_dst, iv_len, 0);
1524
1525                 if (imm_payload) {
1526                         m_copydata(sndptr, sndptroff + TLS_HEADER_LENGTH,
1527                             tls_size, buf);
1528                 } else {
1529                         write_tlstx_sgl(buf, sndptr,
1530                             sndptroff + TLS_HEADER_LENGTH, tls_size, iv_buffer,
1531                             iv_len, nsegs, max_nsegs_1mbuf);
1532                 }
1533
1534                 KASSERT(toep->tx_credits >= credits,
1535                         ("%s: not enough credits", __func__));
1536
1537                 toep->tx_credits -= credits;
1538
1539                 tp->snd_nxt += plen;
1540                 tp->snd_max += plen;
1541
1542                 SOCKBUF_LOCK(sb);
1543                 sbsndptr_adv(sb, sb->sb_sndptr, plen);
1544                 tls_ofld->sb_off += plen;
1545                 SOCKBUF_UNLOCK(sb);
1546
1547                 toep->flags |= TPF_TX_DATA_SENT;
1548                 if (toep->tx_credits < MIN_OFLD_TLSTX_CREDITS(toep))
1549                         toep->flags |= TPF_TX_SUSPENDED;
1550
1551                 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
1552                 txsd->plen = plen;
1553                 txsd->tx_credits = credits;
1554                 txsd->iv_buffer = iv_buffer;
1555                 txsd++;
1556                 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
1557                         toep->txsd_pidx = 0;
1558                         txsd = &toep->txsd[0];
1559                 }
1560                 toep->txsd_avail--;
1561
1562                 atomic_add_long(&toep->vi->pi->tx_tls_records, 1);
1563                 atomic_add_long(&toep->vi->pi->tx_tls_octets, plen);
1564
1565                 t4_l2t_send(sc, wr, toep->l2te);
1566         }
1567 }
1568
1569 #ifdef KERN_TLS
1570 static int
1571 count_ext_pgs_segs(struct mbuf_ext_pgs *ext_pgs)
1572 {
1573         vm_paddr_t nextpa;
1574         u_int i, nsegs;
1575
1576         MPASS(ext_pgs->npgs > 0);
1577         nsegs = 1;
1578         nextpa = ext_pgs->pa[0] + PAGE_SIZE;
1579         for (i = 1; i < ext_pgs->npgs; i++) {
1580                 if (nextpa != ext_pgs->pa[i])
1581                         nsegs++;
1582                 nextpa = ext_pgs->pa[i] + PAGE_SIZE;
1583         }
1584         return (nsegs);
1585 }
1586
1587 static void
1588 write_ktlstx_sgl(void *dst, struct mbuf_ext_pgs *ext_pgs, int nsegs)
1589 {
1590         struct ulptx_sgl *usgl = dst;
1591         vm_paddr_t pa;
1592         uint32_t len;
1593         int i, j;
1594
1595         KASSERT(nsegs > 0, ("%s: nsegs 0", __func__));
1596
1597         usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) |
1598             V_ULPTX_NSGE(nsegs));
1599
1600         /* Figure out the first S/G length. */
1601         pa = ext_pgs->pa[0] + ext_pgs->first_pg_off;
1602         usgl->addr0 = htobe64(pa);
1603         len = mbuf_ext_pg_len(ext_pgs, 0, ext_pgs->first_pg_off);
1604         pa += len;
1605         for (i = 1; i < ext_pgs->npgs; i++) {
1606                 if (ext_pgs->pa[i] != pa)
1607                         break;
1608                 len += mbuf_ext_pg_len(ext_pgs, i, 0);
1609                 pa += mbuf_ext_pg_len(ext_pgs, i, 0);
1610         }
1611         usgl->len0 = htobe32(len);
1612 #ifdef INVARIANTS
1613         nsegs--;
1614 #endif
1615
1616         j = -1;
1617         for (; i < ext_pgs->npgs; i++) {
1618                 if (j == -1 || ext_pgs->pa[i] != pa) {
1619                         if (j >= 0)
1620                                 usgl->sge[j / 2].len[j & 1] = htobe32(len);
1621                         j++;
1622 #ifdef INVARIANTS
1623                         nsegs--;
1624 #endif
1625                         pa = ext_pgs->pa[i];
1626                         usgl->sge[j / 2].addr[j & 1] = htobe64(pa);
1627                         len = mbuf_ext_pg_len(ext_pgs, i, 0);
1628                         pa += len;
1629                 } else {
1630                         len += mbuf_ext_pg_len(ext_pgs, i, 0);
1631                         pa += mbuf_ext_pg_len(ext_pgs, i, 0);
1632                 }
1633         }
1634         if (j >= 0) {
1635                 usgl->sge[j / 2].len[j & 1] = htobe32(len);
1636
1637                 if ((j & 1) == 0)
1638                         usgl->sge[j / 2].len[1] = htobe32(0);
1639         }
1640         KASSERT(nsegs == 0, ("%s: nsegs %d, ext_pgs %p", __func__, nsegs,
1641             ext_pgs));
1642 }
1643
1644 /*
1645  * Similar to t4_push_frames() but handles sockets that contain TLS
1646  * record mbufs.  Unlike TLSOM, each mbuf is a complete TLS record and
1647  * corresponds to a single work request.
1648  */
1649 void
1650 t4_push_ktls(struct adapter *sc, struct toepcb *toep, int drop)
1651 {
1652         struct tls_hdr *thdr;
1653         struct fw_tlstx_data_wr *txwr;
1654         struct cpl_tx_tls_sfo *cpl;
1655         struct wrqe *wr;
1656         struct mbuf *m;
1657         u_int nsegs, credits, wr_len;
1658         u_int expn_size;
1659         struct inpcb *inp = toep->inp;
1660         struct tcpcb *tp = intotcpcb(inp);
1661         struct socket *so = inp->inp_socket;
1662         struct sockbuf *sb = &so->so_snd;
1663         int tls_size, tx_credits, shove, sowwakeup;
1664         struct ofld_tx_sdesc *txsd;
1665         char *buf;
1666
1667         INP_WLOCK_ASSERT(inp);
1668         KASSERT(toep->flags & TPF_FLOWC_WR_SENT,
1669             ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid));
1670
1671         KASSERT(ulp_mode(toep) == ULP_MODE_NONE ||
1672             ulp_mode(toep) == ULP_MODE_TCPDDP,
1673             ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep));
1674         KASSERT(tls_tx_key(toep),
1675             ("%s: TX key not set for toep %p", __func__, toep));
1676
1677 #ifdef VERBOSE_TRACES
1678         CTR4(KTR_CXGBE, "%s: tid %d toep flags %#x tp flags %#x drop %d",
1679             __func__, toep->tid, toep->flags, tp->t_flags);
1680 #endif
1681         if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN))
1682                 return;
1683
1684 #ifdef RATELIMIT
1685         if (__predict_false(inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) &&
1686             (update_tx_rate_limit(sc, toep, so->so_max_pacing_rate) == 0)) {
1687                 inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED;
1688         }
1689 #endif
1690
1691         /*
1692          * This function doesn't resume by itself.  Someone else must clear the
1693          * flag and call this function.
1694          */
1695         if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) {
1696                 KASSERT(drop == 0,
1697                     ("%s: drop (%d) != 0 but tx is suspended", __func__, drop));
1698                 return;
1699         }
1700
1701         txsd = &toep->txsd[toep->txsd_pidx];
1702         for (;;) {
1703                 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS);
1704
1705                 SOCKBUF_LOCK(sb);
1706                 sowwakeup = drop;
1707                 if (drop) {
1708                         sbdrop_locked(sb, drop);
1709                         drop = 0;
1710                 }
1711
1712                 m = sb->sb_sndptr != NULL ? sb->sb_sndptr->m_next : sb->sb_mb;
1713
1714                 /*
1715                  * Send a FIN if requested, but only if there's no
1716                  * more data to send.
1717                  */
1718                 if (m == NULL && toep->flags & TPF_SEND_FIN) {
1719                         if (sowwakeup)
1720                                 sowwakeup_locked(so);
1721                         else
1722                                 SOCKBUF_UNLOCK(sb);
1723                         SOCKBUF_UNLOCK_ASSERT(sb);
1724                         t4_close_conn(sc, toep);
1725                         return;
1726                 }
1727
1728                 /*
1729                  * If there is no ready data to send, wait until more
1730                  * data arrives.
1731                  */
1732                 if (m == NULL || (m->m_flags & M_NOTAVAIL) != 0) {
1733                         if (sowwakeup)
1734                                 sowwakeup_locked(so);
1735                         else
1736                                 SOCKBUF_UNLOCK(sb);
1737                         SOCKBUF_UNLOCK_ASSERT(sb);
1738 #ifdef VERBOSE_TRACES
1739                         CTR2(KTR_CXGBE, "%s: tid %d no ready data to send",
1740                             __func__, toep->tid);
1741 #endif
1742                         return;
1743                 }
1744
1745                 KASSERT(m->m_flags & M_NOMAP, ("%s: mbuf %p is not NOMAP",
1746                     __func__, m));
1747                 KASSERT(m->m_ext.ext_pgs->tls != NULL,
1748                     ("%s: mbuf %p doesn't have TLS session", __func__, m));
1749
1750                 /* Calculate WR length. */
1751                 wr_len = sizeof(struct fw_tlstx_data_wr) +
1752                     sizeof(struct cpl_tx_tls_sfo) + key_size(toep);
1753
1754                 /* Explicit IVs for AES-CBC and AES-GCM are <= 16. */
1755                 MPASS(toep->tls.iv_len <= AES_BLOCK_LEN);
1756                 wr_len += AES_BLOCK_LEN;
1757
1758                 /* Account for SGL in work request length. */
1759                 nsegs = count_ext_pgs_segs(m->m_ext.ext_pgs);
1760                 wr_len += sizeof(struct ulptx_sgl) +
1761                     ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8;
1762
1763                 /* Not enough credits for this work request. */
1764                 if (howmany(wr_len, 16) > tx_credits) {
1765                         if (sowwakeup)
1766                                 sowwakeup_locked(so);
1767                         else
1768                                 SOCKBUF_UNLOCK(sb);
1769                         SOCKBUF_UNLOCK_ASSERT(sb);
1770 #ifdef VERBOSE_TRACES
1771                         CTR5(KTR_CXGBE,
1772             "%s: tid %d mbuf %p requires %d credits, but only %d available",
1773                             __func__, toep->tid, m, howmany(wr_len, 16),
1774                             tx_credits);
1775 #endif
1776                         toep->flags |= TPF_TX_SUSPENDED;
1777                         return;
1778                 }
1779         
1780                 /* Shove if there is no additional data pending. */
1781                 shove = ((m->m_next == NULL ||
1782                     (m->m_next->m_flags & M_NOTAVAIL) != 0)) &&
1783                     (tp->t_flags & TF_MORETOCOME) == 0;
1784
1785                 if (sb->sb_flags & SB_AUTOSIZE &&
1786                     V_tcp_do_autosndbuf &&
1787                     sb->sb_hiwat < V_tcp_autosndbuf_max &&
1788                     sbused(sb) >= sb->sb_hiwat * 7 / 8) {
1789                         int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc,
1790                             V_tcp_autosndbuf_max);
1791
1792                         if (!sbreserve_locked(sb, newsize, so, NULL))
1793                                 sb->sb_flags &= ~SB_AUTOSIZE;
1794                         else
1795                                 sowwakeup = 1;  /* room available */
1796                 }
1797                 if (sowwakeup)
1798                         sowwakeup_locked(so);
1799                 else
1800                         SOCKBUF_UNLOCK(sb);
1801                 SOCKBUF_UNLOCK_ASSERT(sb);
1802
1803                 if (__predict_false(toep->flags & TPF_FIN_SENT))
1804                         panic("%s: excess tx.", __func__);
1805
1806                 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq);
1807                 if (wr == NULL) {
1808                         /* XXX: how will we recover from this? */
1809                         toep->flags |= TPF_TX_SUSPENDED;
1810                         return;
1811                 }
1812
1813                 thdr = (struct tls_hdr *)m->m_ext.ext_pgs->hdr;
1814 #ifdef VERBOSE_TRACES
1815                 CTR5(KTR_CXGBE, "%s: tid %d TLS record %ju type %d len %#x",
1816                     __func__, toep->tid, m->m_ext.ext_pgs->seqno, thdr->type,
1817                     m->m_len);
1818 #endif
1819                 txwr = wrtod(wr);
1820                 cpl = (struct cpl_tx_tls_sfo *)(txwr + 1);
1821                 memset(txwr, 0, roundup2(wr_len, 16));
1822                 credits = howmany(wr_len, 16);
1823                 expn_size = m->m_ext.ext_pgs->hdr_len +
1824                     m->m_ext.ext_pgs->trail_len;
1825                 tls_size = m->m_len - expn_size;
1826                 write_tlstx_wr(txwr, toep, 0,
1827                     tls_size, expn_size, 1, credits, shove, 1);
1828                 toep->tls.tx_seq_no = m->m_ext.ext_pgs->seqno;
1829                 write_tlstx_cpl(cpl, toep, thdr, tls_size, 1);
1830                 tls_copy_tx_key(toep, cpl + 1);
1831
1832                 /* Copy IV. */
1833                 buf = (char *)(cpl + 1) + key_size(toep);
1834                 memcpy(buf, thdr + 1, toep->tls.iv_len);
1835                 buf += AES_BLOCK_LEN;
1836
1837                 write_ktlstx_sgl(buf, m->m_ext.ext_pgs, nsegs);
1838
1839                 KASSERT(toep->tx_credits >= credits,
1840                         ("%s: not enough credits", __func__));
1841
1842                 toep->tx_credits -= credits;
1843
1844                 tp->snd_nxt += m->m_len;
1845                 tp->snd_max += m->m_len;
1846
1847                 SOCKBUF_LOCK(sb);
1848                 sb->sb_sndptr = m;
1849                 SOCKBUF_UNLOCK(sb);
1850
1851                 toep->flags |= TPF_TX_DATA_SENT;
1852                 if (toep->tx_credits < MIN_OFLD_TLSTX_CREDITS(toep))
1853                         toep->flags |= TPF_TX_SUSPENDED;
1854
1855                 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__));
1856                 txsd->plen = m->m_len;
1857                 txsd->tx_credits = credits;
1858                 txsd++;
1859                 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) {
1860                         toep->txsd_pidx = 0;
1861                         txsd = &toep->txsd[0];
1862                 }
1863                 toep->txsd_avail--;
1864
1865                 atomic_add_long(&toep->vi->pi->tx_tls_records, 1);
1866                 atomic_add_long(&toep->vi->pi->tx_tls_octets, m->m_len);
1867
1868                 t4_l2t_send(sc, wr, toep->l2te);
1869         }
1870 }
1871 #endif
1872
1873 /*
1874  * For TLS data we place received mbufs received via CPL_TLS_DATA into
1875  * an mbufq in the TLS offload state.  When CPL_RX_TLS_CMP is
1876  * received, the completed PDUs are placed into the socket receive
1877  * buffer.
1878  *
1879  * The TLS code reuses the ulp_pdu_reclaimq to hold the pending mbufs.
1880  */
1881 static int
1882 do_tls_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1883 {
1884         struct adapter *sc = iq->adapter;
1885         const struct cpl_tls_data *cpl = mtod(m, const void *);
1886         unsigned int tid = GET_TID(cpl);
1887         struct toepcb *toep = lookup_tid(sc, tid);
1888         struct inpcb *inp = toep->inp;
1889         struct tcpcb *tp;
1890         int len;
1891
1892         /* XXX: Should this match do_rx_data instead? */
1893         KASSERT(!(toep->flags & TPF_SYNQE),
1894             ("%s: toep %p claims to be a synq entry", __func__, toep));
1895
1896         KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
1897
1898         /* strip off CPL header */
1899         m_adj(m, sizeof(*cpl));
1900         len = m->m_pkthdr.len;
1901
1902         atomic_add_long(&toep->vi->pi->rx_tls_octets, len);
1903
1904         KASSERT(len == G_CPL_TLS_DATA_LENGTH(be32toh(cpl->length_pkd)),
1905             ("%s: payload length mismatch", __func__));
1906
1907         INP_WLOCK(inp);
1908         if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1909                 CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1910                     __func__, tid, len, inp->inp_flags);
1911                 INP_WUNLOCK(inp);
1912                 m_freem(m);
1913                 return (0);
1914         }
1915
1916         /* Save TCP sequence number. */
1917         m->m_pkthdr.tls_tcp_seq = be32toh(cpl->seq);
1918
1919         if (mbufq_enqueue(&toep->ulp_pdu_reclaimq, m)) {
1920 #ifdef INVARIANTS
1921                 panic("Failed to queue TLS data packet");
1922 #else
1923                 printf("%s: Failed to queue TLS data packet\n", __func__);
1924                 INP_WUNLOCK(inp);
1925                 m_freem(m);
1926                 return (0);
1927 #endif
1928         }
1929
1930         tp = intotcpcb(inp);
1931         tp->t_rcvtime = ticks;
1932
1933 #ifdef VERBOSE_TRACES
1934         CTR4(KTR_CXGBE, "%s: tid %u len %d seq %u", __func__, tid, len,
1935             be32toh(cpl->seq));
1936 #endif
1937
1938         INP_WUNLOCK(inp);
1939         return (0);
1940 }
1941
1942 static int
1943 do_rx_tls_cmp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m)
1944 {
1945         struct adapter *sc = iq->adapter;
1946         const struct cpl_rx_tls_cmp *cpl = mtod(m, const void *);
1947         struct tlsrx_hdr_pkt *tls_hdr_pkt;
1948         unsigned int tid = GET_TID(cpl);
1949         struct toepcb *toep = lookup_tid(sc, tid);
1950         struct inpcb *inp = toep->inp;
1951         struct tcpcb *tp;
1952         struct socket *so;
1953         struct sockbuf *sb;
1954         struct mbuf *tls_data;
1955         int len, pdu_length, rx_credits;
1956
1957         KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__));
1958         KASSERT(!(toep->flags & TPF_SYNQE),
1959             ("%s: toep %p claims to be a synq entry", __func__, toep));
1960
1961         /* strip off CPL header */
1962         m_adj(m, sizeof(*cpl));
1963         len = m->m_pkthdr.len;
1964
1965         atomic_add_long(&toep->vi->pi->rx_tls_records, 1);
1966
1967         KASSERT(len == G_CPL_RX_TLS_CMP_LENGTH(be32toh(cpl->pdulength_length)),
1968             ("%s: payload length mismatch", __func__));
1969
1970         INP_WLOCK(inp);
1971         if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) {
1972                 CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x",
1973                     __func__, tid, len, inp->inp_flags);
1974                 INP_WUNLOCK(inp);
1975                 m_freem(m);
1976                 return (0);
1977         }
1978
1979         pdu_length = G_CPL_RX_TLS_CMP_PDULENGTH(be32toh(cpl->pdulength_length));
1980
1981         tp = intotcpcb(inp);
1982
1983 #ifdef VERBOSE_TRACES
1984         CTR6(KTR_CXGBE, "%s: tid %u PDU len %d len %d seq %u, rcv_nxt %u",
1985             __func__, tid, pdu_length, len, be32toh(cpl->seq), tp->rcv_nxt);
1986 #endif
1987
1988         tp->rcv_nxt += pdu_length;
1989         if (tp->rcv_wnd < pdu_length) {
1990                 toep->tls.rcv_over += pdu_length - tp->rcv_wnd;
1991                 tp->rcv_wnd = 0;
1992         } else
1993                 tp->rcv_wnd -= pdu_length;
1994
1995         /* XXX: Not sure what to do about urgent data. */
1996
1997         /*
1998          * The payload of this CPL is the TLS header followed by
1999          * additional fields.
2000          */
2001         KASSERT(m->m_len >= sizeof(*tls_hdr_pkt),
2002             ("%s: payload too small", __func__));
2003         tls_hdr_pkt = mtod(m, void *);
2004
2005         /*
2006          * Only the TLS header is sent to OpenSSL, so report errors by
2007          * altering the record type.
2008          */
2009         if ((tls_hdr_pkt->res_to_mac_error & M_TLSRX_HDR_PKT_ERROR) != 0)
2010                 tls_hdr_pkt->type = CONTENT_TYPE_ERROR;
2011
2012         /* Trim this CPL's mbuf to only include the TLS header. */
2013         KASSERT(m->m_len == len && m->m_next == NULL,
2014             ("%s: CPL spans multiple mbufs", __func__));
2015         m->m_len = TLS_HEADER_LENGTH;
2016         m->m_pkthdr.len = TLS_HEADER_LENGTH;
2017
2018         tls_data = mbufq_dequeue(&toep->ulp_pdu_reclaimq);
2019         if (tls_data != NULL) {
2020                 KASSERT(be32toh(cpl->seq) == tls_data->m_pkthdr.tls_tcp_seq,
2021                     ("%s: sequence mismatch", __func__));
2022
2023                 /*
2024                  * Update the TLS header length to be the length of
2025                  * the payload data.
2026                  */
2027                 tls_hdr_pkt->length = htobe16(tls_data->m_pkthdr.len);
2028
2029                 m->m_next = tls_data;
2030                 m->m_pkthdr.len += tls_data->m_len;
2031         }
2032
2033         so = inp_inpcbtosocket(inp);
2034         sb = &so->so_rcv;
2035         SOCKBUF_LOCK(sb);
2036
2037         if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) {
2038                 struct epoch_tracker et;
2039
2040                 CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)",
2041                     __func__, tid, pdu_length);
2042                 m_freem(m);
2043                 SOCKBUF_UNLOCK(sb);
2044                 INP_WUNLOCK(inp);
2045
2046                 CURVNET_SET(toep->vnet);
2047                 NET_EPOCH_ENTER(et);
2048                 INP_WLOCK(inp);
2049                 tp = tcp_drop(tp, ECONNRESET);
2050                 if (tp)
2051                         INP_WUNLOCK(inp);
2052                 NET_EPOCH_EXIT(et);
2053                 CURVNET_RESTORE();
2054
2055                 return (0);
2056         }
2057
2058         /*
2059          * Not all of the bytes on the wire are included in the socket buffer
2060          * (e.g. the MAC of the TLS record).  However, those bytes are included
2061          * in the TCP sequence space.
2062          */
2063
2064         /* receive buffer autosize */
2065         MPASS(toep->vnet == so->so_vnet);
2066         CURVNET_SET(toep->vnet);
2067         if (sb->sb_flags & SB_AUTOSIZE &&
2068             V_tcp_do_autorcvbuf &&
2069             sb->sb_hiwat < V_tcp_autorcvbuf_max &&
2070             m->m_pkthdr.len > (sbspace(sb) / 8 * 7)) {
2071                 unsigned int hiwat = sb->sb_hiwat;
2072                 unsigned int newsize = min(hiwat + sc->tt.autorcvbuf_inc,
2073                     V_tcp_autorcvbuf_max);
2074
2075                 if (!sbreserve_locked(sb, newsize, so, NULL))
2076                         sb->sb_flags &= ~SB_AUTOSIZE;
2077         }
2078
2079         sbappendstream_locked(sb, m, 0);
2080         rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0;
2081 #ifdef VERBOSE_TRACES
2082         CTR4(KTR_CXGBE, "%s: tid %u rx_credits %u rcv_wnd %u",
2083             __func__, tid, rx_credits, tp->rcv_wnd);
2084 #endif
2085         if (rx_credits > 0 && sbused(sb) + tp->rcv_wnd < sb->sb_lowat) {
2086                 rx_credits = send_rx_credits(sc, toep, rx_credits);
2087                 tp->rcv_wnd += rx_credits;
2088                 tp->rcv_adv += rx_credits;
2089         }
2090
2091         sorwakeup_locked(so);
2092         SOCKBUF_UNLOCK_ASSERT(sb);
2093
2094         INP_WUNLOCK(inp);
2095         CURVNET_RESTORE();
2096         return (0);
2097 }
2098
2099 void
2100 t4_tls_mod_load(void)
2101 {
2102
2103         mtx_init(&tls_handshake_lock, "t4tls handshake", NULL, MTX_DEF);
2104         t4_register_cpl_handler(CPL_TLS_DATA, do_tls_data);
2105         t4_register_cpl_handler(CPL_RX_TLS_CMP, do_rx_tls_cmp);
2106 }
2107
2108 void
2109 t4_tls_mod_unload(void)
2110 {
2111
2112         t4_register_cpl_handler(CPL_TLS_DATA, NULL);
2113         t4_register_cpl_handler(CPL_RX_TLS_CMP, NULL);
2114         mtx_destroy(&tls_handshake_lock);
2115 }
2116 #endif  /* TCP_OFFLOAD */