git.oblomov.eu Git - linux-2.6/blob - drivers/scsi/cxgb3i/cxgb3i_offload.c

   1 /*
   2  * cxgb3i_offload.c: Chelsio S3xx iscsi offloaded tcp connection management
   3  *
   4  * Copyright (C) 2003-2008 Chelsio Communications.  All rights reserved.
   5  *
   6  * This program is distributed in the hope that it will be useful, but WITHOUT
   7  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
   8  * FITNESS FOR A PARTICULAR PURPOSE.  See the LICENSE file included in this
   9  * release for licensing terms and conditions.
  10  *
  11  * Written by:  Dimitris Michailidis (dm@chelsio.com)
  12  *              Karen Xie (kxie@chelsio.com)
  13  */
  14
  15 #include <linux/if_vlan.h>
  16 #include <linux/version.h>
  17
  18 #include "cxgb3_defs.h"
  19 #include "cxgb3_ctl_defs.h"
  20 #include "firmware_exports.h"
  21 #include "cxgb3i_offload.h"
  22 #include "cxgb3i_pdu.h"
  23 #include "cxgb3i_ddp.h"
  24
  25 #ifdef __DEBUG_C3CN_CONN__
  26 #define c3cn_conn_debug         cxgb3i_log_debug
  27 #else
  28 #define c3cn_conn_debug(fmt...)
  29 #endif
  30
  31 #ifdef __DEBUG_C3CN_TX__
  32 #define c3cn_tx_debug           cxgb3i_log_debug
  33 #else
  34 #define c3cn_tx_debug(fmt...)
  35 #endif
  36
  37 #ifdef __DEBUG_C3CN_RX__
  38 #define c3cn_rx_debug           cxgb3i_log_debug
  39 #else
  40 #define c3cn_rx_debug(fmt...)
  41 #endif
  42
  43 /*
  44  * module parameters releated to offloaded iscsi connection
  45  */
  46 static int cxgb3_rcv_win = 256 * 1024;
  47 module_param(cxgb3_rcv_win, int, 0644);
  48 MODULE_PARM_DESC(cxgb3_rcv_win, "TCP receive window in bytes (default=256KB)");
  49
  50 static int cxgb3_snd_win = 128 * 1024;
  51 module_param(cxgb3_snd_win, int, 0644);
  52 MODULE_PARM_DESC(cxgb3_snd_win, "TCP send window in bytes (default=128KB)");
  53
  54 static int cxgb3_rx_credit_thres = 10 * 1024;
  55 module_param(cxgb3_rx_credit_thres, int, 0644);
  56 MODULE_PARM_DESC(rx_credit_thres,
  57                  "RX credits return threshold in bytes (default=10KB)");
  58
  59 static unsigned int cxgb3_max_connect = 8 * 1024;
  60 module_param(cxgb3_max_connect, uint, 0644);
  61 MODULE_PARM_DESC(cxgb3_max_connect, "Max. # of connections (default=8092)");
  62
  63 static unsigned int cxgb3_sport_base = 20000;
  64 module_param(cxgb3_sport_base, uint, 0644);
  65 MODULE_PARM_DESC(cxgb3_sport_base, "starting port number (default=20000)");
  66
  67 /*
  68  * cxgb3i tcp connection data(per adapter) list
  69  */
  70 static LIST_HEAD(cdata_list);
  71 static DEFINE_RWLOCK(cdata_rwlock);
  72
  73 static int c3cn_push_tx_frames(struct s3_conn *c3cn, int req_completion);
  74 static void c3cn_release_offload_resources(struct s3_conn *c3cn);
  75
  76 /*
  77  * iscsi source port management
  78  *
  79  * Find a free source port in the port allocation map. We use a very simple
  80  * rotor scheme to look for the next free port.
  81  *
  82  * If a source port has been specified make sure that it doesn't collide with
  83  * our normal source port allocation map.  If it's outside the range of our
  84  * allocation/deallocation scheme just let them use it.
  85  *
  86  * If the source port is outside our allocation range, the caller is
  87  * responsible for keeping track of their port usage.
  88  */
  89 static int c3cn_get_port(struct s3_conn *c3cn, struct cxgb3i_sdev_data *cdata)
  90 {
  91         unsigned int start;
  92         int idx;
  93
  94         if (!cdata)
  95                 goto error_out;
  96
  97         if (c3cn->saddr.sin_port) {
  98                 cxgb3i_log_error("connect, sin_port NON-ZERO %u.\n",
  99                                  c3cn->saddr.sin_port);
 100                 return -EADDRINUSE;
 101         }
 102
 103         spin_lock_bh(&cdata->lock);
 104         start = idx = cdata->sport_next;
 105         do {
 106                 if (++idx >= cxgb3_max_connect)
 107                         idx = 0;
 108                 if (!cdata->sport_conn[idx]) {
 109                         c3cn->saddr.sin_port = htons(cxgb3_sport_base + idx);
 110                         cdata->sport_next = idx;
 111                         cdata->sport_conn[idx] = c3cn;
 112                         spin_unlock_bh(&cdata->lock);
 113
 114                         c3cn_conn_debug("%s reserve port %u.\n",
 115                                         cdata->cdev->name,
 116                                         cxgb3_sport_base + idx);
 117                         return 0;
 118                 }
 119         } while (idx != start);
 120         spin_unlock_bh(&cdata->lock);
 121
 122 error_out:
 123         return -EADDRNOTAVAIL;
 124 }
 125
 126 static void c3cn_put_port(struct s3_conn *c3cn)
 127 {
 128         if (!c3cn->cdev)
 129                 return;
 130
 131         if (c3cn->saddr.sin_port) {
 132                 struct cxgb3i_sdev_data *cdata = CXGB3_SDEV_DATA(c3cn->cdev);
 133                 int idx = ntohs(c3cn->saddr.sin_port) - cxgb3_sport_base;
 134
 135                 c3cn->saddr.sin_port = 0;
 136                 if (idx < 0 || idx >= cxgb3_max_connect)
 137                         return;
 138                 spin_lock_bh(&cdata->lock);
 139                 cdata->sport_conn[idx] = NULL;
 140                 spin_unlock_bh(&cdata->lock);
 141                 c3cn_conn_debug("%s, release port %u.\n",
 142                                 cdata->cdev->name, cxgb3_sport_base + idx);
 143         }
 144 }
 145
 146 static inline void c3cn_set_flag(struct s3_conn *c3cn, enum c3cn_flags flag)
 147 {
 148         __set_bit(flag, &c3cn->flags);
 149         c3cn_conn_debug("c3cn 0x%p, set %d, s %u, f 0x%lx.\n",
 150                         c3cn, flag, c3cn->state, c3cn->flags);
 151 }
 152
 153 static inline void c3cn_clear_flag(struct s3_conn *c3cn, enum c3cn_flags flag)
 154 {
 155         __clear_bit(flag, &c3cn->flags);
 156         c3cn_conn_debug("c3cn 0x%p, clear %d, s %u, f 0x%lx.\n",
 157                         c3cn, flag, c3cn->state, c3cn->flags);
 158 }
 159
 160 static inline int c3cn_flag(struct s3_conn *c3cn, enum c3cn_flags flag)
 161 {
 162         if (c3cn == NULL)
 163                 return 0;
 164         return test_bit(flag, &c3cn->flags);
 165 }
 166
 167 static void c3cn_set_state(struct s3_conn *c3cn, int state)
 168 {
 169         c3cn_conn_debug("c3cn 0x%p state -> %u.\n", c3cn, state);
 170         c3cn->state = state;
 171 }
 172
 173 static inline void c3cn_hold(struct s3_conn *c3cn)
 174 {
 175         atomic_inc(&c3cn->refcnt);
 176 }
 177
 178 static inline void c3cn_put(struct s3_conn *c3cn)
 179 {
 180         if (atomic_dec_and_test(&c3cn->refcnt)) {
 181                 c3cn_conn_debug("free c3cn 0x%p, s %u, f 0x%lx.\n",
 182                                 c3cn, c3cn->state, c3cn->flags);
 183                 kfree(c3cn);
 184         }
 185 }
 186
 187 static void c3cn_closed(struct s3_conn *c3cn)
 188 {
 189         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
 190                          c3cn, c3cn->state, c3cn->flags);
 191
 192         c3cn_put_port(c3cn);
 193         c3cn_release_offload_resources(c3cn);
 194         c3cn_set_state(c3cn, C3CN_STATE_CLOSED);
 195         cxgb3i_conn_closing(c3cn);
 196 }
 197
 198 /*
 199  * CPL (Chelsio Protocol Language) defines a message passing interface between
 200  * the host driver and T3 asic.
 201  * The section below implments CPLs that related to iscsi tcp connection
 202  * open/close/abort and data send/receive.
 203  */
 204
 205 /*
 206  * CPL connection active open request: host ->
 207  */
 208 static unsigned int find_best_mtu(const struct t3c_data *d, unsigned short mtu)
 209 {
 210         int i = 0;
 211
 212         while (i < d->nmtus - 1 && d->mtus[i + 1] <= mtu)
 213                 ++i;
 214         return i;
 215 }
 216
 217 static unsigned int select_mss(struct s3_conn *c3cn, unsigned int pmtu)
 218 {
 219         unsigned int idx;
 220         struct dst_entry *dst = c3cn->dst_cache;
 221         struct t3cdev *cdev = c3cn->cdev;
 222         const struct t3c_data *td = T3C_DATA(cdev);
 223         u16 advmss = dst_metric(dst, RTAX_ADVMSS);
 224
 225         if (advmss > pmtu - 40)
 226                 advmss = pmtu - 40;
 227         if (advmss < td->mtus[0] - 40)
 228                 advmss = td->mtus[0] - 40;
 229         idx = find_best_mtu(td, advmss + 40);
 230         return idx;
 231 }
 232
 233 static inline int compute_wscale(int win)
 234 {
 235         int wscale = 0;
 236         while (wscale < 14 && (65535<<wscale) < win)
 237                 wscale++;
 238         return wscale;
 239 }
 240
 241 static inline unsigned int calc_opt0h(struct s3_conn *c3cn)
 242 {
 243         int wscale = compute_wscale(cxgb3_rcv_win);
 244         return  V_KEEP_ALIVE(1) |
 245                 F_TCAM_BYPASS |
 246                 V_WND_SCALE(wscale) |
 247                 V_MSS_IDX(c3cn->mss_idx);
 248 }
 249
 250 static inline unsigned int calc_opt0l(struct s3_conn *c3cn)
 251 {
 252         return  V_ULP_MODE(ULP_MODE_ISCSI) |
 253                 V_RCV_BUFSIZ(cxgb3_rcv_win>>10);
 254 }
 255
 256 static void make_act_open_req(struct s3_conn *c3cn, struct sk_buff *skb,
 257                               unsigned int atid, const struct l2t_entry *e)
 258 {
 259         struct cpl_act_open_req *req;
 260
 261         c3cn_conn_debug("c3cn 0x%p, atid 0x%x.\n", c3cn, atid);
 262
 263         skb->priority = CPL_PRIORITY_SETUP;
 264         req = (struct cpl_act_open_req *)__skb_put(skb, sizeof(*req));
 265         req->wr.wr_hi = htonl(V_WR_OP(FW_WROPCODE_FORWARD));
 266         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_ACT_OPEN_REQ, atid));
 267         req->local_port = c3cn->saddr.sin_port;
 268         req->peer_port = c3cn->daddr.sin_port;
 269         req->local_ip = c3cn->saddr.sin_addr.s_addr;
 270         req->peer_ip = c3cn->daddr.sin_addr.s_addr;
 271         req->opt0h = htonl(calc_opt0h(c3cn) | V_L2T_IDX(e->idx) |
 272                            V_TX_CHANNEL(e->smt_idx));
 273         req->opt0l = htonl(calc_opt0l(c3cn));
 274         req->params = 0;
 275 }
 276
 277 static void fail_act_open(struct s3_conn *c3cn, int errno)
 278 {
 279         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
 280                         c3cn, c3cn->state, c3cn->flags);
 281         c3cn->err = errno;
 282         c3cn_closed(c3cn);
 283 }
 284
 285 static void act_open_req_arp_failure(struct t3cdev *dev, struct sk_buff *skb)
 286 {
 287         struct s3_conn *c3cn = (struct s3_conn *)skb->sk;
 288
 289         c3cn_conn_debug("c3cn 0x%p, state %u.\n", c3cn, c3cn->state);
 290
 291         c3cn_hold(c3cn);
 292         spin_lock_bh(&c3cn->lock);
 293         if (c3cn->state == C3CN_STATE_CONNECTING)
 294                 fail_act_open(c3cn, EHOSTUNREACH);
 295         spin_unlock_bh(&c3cn->lock);
 296         c3cn_put(c3cn);
 297         __kfree_skb(skb);
 298 }
 299
 300 /*
 301  * CPL connection close request: host ->
 302  *
 303  * Close a connection by sending a CPL_CLOSE_CON_REQ message and queue it to
 304  * the write queue (i.e., after any unsent txt data).
 305  */
 306 static void skb_entail(struct s3_conn *c3cn, struct sk_buff *skb,
 307                        int flags)
 308 {
 309         skb_tcp_seq(skb) = c3cn->write_seq;
 310         skb_flags(skb) = flags;
 311         __skb_queue_tail(&c3cn->write_queue, skb);
 312 }
 313
 314 static void send_close_req(struct s3_conn *c3cn)
 315 {
 316         struct sk_buff *skb = c3cn->cpl_close;
 317         struct cpl_close_con_req *req = (struct cpl_close_con_req *)skb->head;
 318         unsigned int tid = c3cn->tid;
 319
 320         c3cn_conn_debug("c3cn 0x%p, state 0x%x, flag 0x%lx.\n",
 321                         c3cn, c3cn->state, c3cn->flags);
 322
 323         c3cn->cpl_close = NULL;
 324
 325         req->wr.wr_hi = htonl(V_WR_OP(FW_WROPCODE_OFLD_CLOSE_CON));
 326         req->wr.wr_lo = htonl(V_WR_TID(tid));
 327         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
 328         req->rsvd = htonl(c3cn->write_seq);
 329
 330         skb_entail(c3cn, skb, C3CB_FLAG_NO_APPEND);
 331         if (c3cn->state != C3CN_STATE_CONNECTING)
 332                 c3cn_push_tx_frames(c3cn, 1);
 333 }
 334
 335 /*
 336  * CPL connection abort request: host ->
 337  *
 338  * Send an ABORT_REQ message. Makes sure we do not send multiple ABORT_REQs
 339  * for the same connection and also that we do not try to send a message
 340  * after the connection has closed.
 341  */
 342 static void abort_arp_failure(struct t3cdev *cdev, struct sk_buff *skb)
 343 {
 344         struct cpl_abort_req *req = cplhdr(skb);
 345
 346         c3cn_conn_debug("tdev 0x%p.\n", cdev);
 347
 348         req->cmd = CPL_ABORT_NO_RST;
 349         cxgb3_ofld_send(cdev, skb);
 350 }
 351
 352 static inline void c3cn_purge_write_queue(struct s3_conn *c3cn)
 353 {
 354         struct sk_buff *skb;
 355
 356         while ((skb = __skb_dequeue(&c3cn->write_queue)))
 357                 __kfree_skb(skb);
 358 }
 359
 360 static void send_abort_req(struct s3_conn *c3cn)
 361 {
 362         struct sk_buff *skb = c3cn->cpl_abort_req;
 363         struct cpl_abort_req *req;
 364         unsigned int tid = c3cn->tid;
 365
 366         if (unlikely(c3cn->state == C3CN_STATE_ABORTING) || !skb ||
 367                      !c3cn->cdev)
 368                 return;
 369
 370         c3cn_set_state(c3cn, C3CN_STATE_ABORTING);
 371
 372         c3cn_conn_debug("c3cn 0x%p, flag ABORT_RPL + ABORT_SHUT.\n", c3cn);
 373
 374         c3cn_set_flag(c3cn, C3CN_ABORT_RPL_PENDING);
 375
 376         /* Purge the send queue so we don't send anything after an abort. */
 377         c3cn_purge_write_queue(c3cn);
 378
 379         c3cn->cpl_abort_req = NULL;
 380         req = (struct cpl_abort_req *)skb->head;
 381
 382         skb->priority = CPL_PRIORITY_DATA;
 383         set_arp_failure_handler(skb, abort_arp_failure);
 384
 385         req->wr.wr_hi = htonl(V_WR_OP(FW_WROPCODE_OFLD_HOST_ABORT_CON_REQ));
 386         req->wr.wr_lo = htonl(V_WR_TID(tid));
 387         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_ABORT_REQ, tid));
 388         req->rsvd0 = htonl(c3cn->snd_nxt);
 389         req->rsvd1 = !c3cn_flag(c3cn, C3CN_TX_DATA_SENT);
 390         req->cmd = CPL_ABORT_SEND_RST;
 391
 392         l2t_send(c3cn->cdev, skb, c3cn->l2t);
 393 }
 394
 395 /*
 396  * CPL connection abort reply: host ->
 397  *
 398  * Send an ABORT_RPL message in response of the ABORT_REQ received.
 399  */
 400 static void send_abort_rpl(struct s3_conn *c3cn, int rst_status)
 401 {
 402         struct sk_buff *skb = c3cn->cpl_abort_rpl;
 403         struct cpl_abort_rpl *rpl = (struct cpl_abort_rpl *)skb->head;
 404
 405         c3cn->cpl_abort_rpl = NULL;
 406
 407         skb->priority = CPL_PRIORITY_DATA;
 408         rpl->wr.wr_hi = htonl(V_WR_OP(FW_WROPCODE_OFLD_HOST_ABORT_CON_RPL));
 409         rpl->wr.wr_lo = htonl(V_WR_TID(c3cn->tid));
 410         OPCODE_TID(rpl) = htonl(MK_OPCODE_TID(CPL_ABORT_RPL, c3cn->tid));
 411         rpl->cmd = rst_status;
 412
 413         cxgb3_ofld_send(c3cn->cdev, skb);
 414 }
 415
 416 /*
 417  * CPL connection rx data ack: host ->
 418  * Send RX credits through an RX_DATA_ACK CPL message. Returns the number of
 419  * credits sent.
 420  */
 421 static u32 send_rx_credits(struct s3_conn *c3cn, u32 credits, u32 dack)
 422 {
 423         struct sk_buff *skb;
 424         struct cpl_rx_data_ack *req;
 425
 426         skb = alloc_skb(sizeof(*req), GFP_ATOMIC);
 427         if (!skb)
 428                 return 0;
 429
 430         req = (struct cpl_rx_data_ack *)__skb_put(skb, sizeof(*req));
 431         req->wr.wr_hi = htonl(V_WR_OP(FW_WROPCODE_FORWARD));
 432         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_RX_DATA_ACK, c3cn->tid));
 433         req->credit_dack = htonl(dack | V_RX_CREDITS(credits));
 434         skb->priority = CPL_PRIORITY_ACK;
 435         cxgb3_ofld_send(c3cn->cdev, skb);
 436         return credits;
 437 }
 438
 439 /*
 440  * CPL connection tx data: host ->
 441  *
 442  * Send iscsi PDU via TX_DATA CPL message. Returns the number of
 443  * credits sent.
 444  * Each TX_DATA consumes work request credit (wrs), so we need to keep track of
 445  * how many we've used so far and how many are pending (i.e., yet ack'ed by T3).
 446  */
 447
 448 /*
 449  * For ULP connections HW may inserts digest bytes into the pdu. Those digest
 450  * bytes are not sent by the host but are part of the TCP payload and therefore
 451  * consume TCP sequence space.
 452  */
 453 static const unsigned int cxgb3_ulp_extra_len[] = { 0, 4, 4, 8 };
 454 static inline unsigned int ulp_extra_len(const struct sk_buff *skb)
 455 {
 456         return cxgb3_ulp_extra_len[skb_ulp_mode(skb) & 3];
 457 }
 458
 459 static unsigned int wrlen __read_mostly;
 460
 461 /*
 462  * The number of WRs needed for an skb depends on the number of fragments
 463  * in the skb and whether it has any payload in its main body.  This maps the
 464  * length of the gather list represented by an skb into the # of necessary WRs.
 465  * The extra two fragments are for iscsi bhs and payload padding.
 466  */
 467 #define SKB_WR_LIST_SIZE        (MAX_SKB_FRAGS + 2)
 468 static unsigned int skb_wrs[SKB_WR_LIST_SIZE] __read_mostly;
 469
 470 static void s3_init_wr_tab(unsigned int wr_len)
 471 {
 472         int i;
 473
 474         if (skb_wrs[1])         /* already initialized */
 475                 return;
 476
 477         for (i = 1; i < SKB_WR_LIST_SIZE; i++) {
 478                 int sgl_len = (3 * i) / 2 + (i & 1);
 479
 480                 sgl_len += 3;
 481                 skb_wrs[i] = (sgl_len <= wr_len
 482                               ? 1 : 1 + (sgl_len - 2) / (wr_len - 1));
 483         }
 484
 485         wrlen = wr_len * 8;
 486 }
 487
 488 static inline void reset_wr_list(struct s3_conn *c3cn)
 489 {
 490         c3cn->wr_pending_head = c3cn->wr_pending_tail = NULL;
 491 }
 492
 493 /*
 494  * Add a WR to a connections's list of pending WRs.  This is a singly-linked
 495  * list of sk_buffs operating as a FIFO.  The head is kept in wr_pending_head
 496  * and the tail in wr_pending_tail.
 497  */
 498 static inline void enqueue_wr(struct s3_conn *c3cn,
 499                               struct sk_buff *skb)
 500 {
 501         skb_tx_wr_next(skb) = NULL;
 502
 503         /*
 504          * We want to take an extra reference since both us and the driver
 505          * need to free the packet before it's really freed. We know there's
 506          * just one user currently so we use atomic_set rather than skb_get
 507          * to avoid the atomic op.
 508          */
 509         atomic_set(&skb->users, 2);
 510
 511         if (!c3cn->wr_pending_head)
 512                 c3cn->wr_pending_head = skb;
 513         else
 514                 skb_tx_wr_next(c3cn->wr_pending_tail) = skb;
 515         c3cn->wr_pending_tail = skb;
 516 }
 517
 518 static int count_pending_wrs(struct s3_conn *c3cn)
 519 {
 520         int n = 0;
 521         const struct sk_buff *skb = c3cn->wr_pending_head;
 522
 523         while (skb) {
 524                 n += skb->csum;
 525                 skb = skb_tx_wr_next(skb);
 526         }
 527         return n;
 528 }
 529
 530 static inline struct sk_buff *peek_wr(const struct s3_conn *c3cn)
 531 {
 532         return c3cn->wr_pending_head;
 533 }
 534
 535 static inline void free_wr_skb(struct sk_buff *skb)
 536 {
 537         kfree_skb(skb);
 538 }
 539
 540 static inline struct sk_buff *dequeue_wr(struct s3_conn *c3cn)
 541 {
 542         struct sk_buff *skb = c3cn->wr_pending_head;
 543
 544         if (likely(skb)) {
 545                 /* Don't bother clearing the tail */
 546                 c3cn->wr_pending_head = skb_tx_wr_next(skb);
 547                 skb_tx_wr_next(skb) = NULL;
 548         }
 549         return skb;
 550 }
 551
 552 static void purge_wr_queue(struct s3_conn *c3cn)
 553 {
 554         struct sk_buff *skb;
 555         while ((skb = dequeue_wr(c3cn)) != NULL)
 556                 free_wr_skb(skb);
 557 }
 558
 559 static inline void make_tx_data_wr(struct s3_conn *c3cn, struct sk_buff *skb,
 560                                    int len, int req_completion)
 561 {
 562         struct tx_data_wr *req;
 563
 564         skb_reset_transport_header(skb);
 565         req = (struct tx_data_wr *)__skb_push(skb, sizeof(*req));
 566         req->wr_hi = htonl(V_WR_OP(FW_WROPCODE_OFLD_TX_DATA) |
 567                         (req_completion ? F_WR_COMPL : 0));
 568         req->wr_lo = htonl(V_WR_TID(c3cn->tid));
 569         req->sndseq = htonl(c3cn->snd_nxt);
 570         /* len includes the length of any HW ULP additions */
 571         req->len = htonl(len);
 572         req->param = htonl(V_TX_PORT(c3cn->l2t->smt_idx));
 573         /* V_TX_ULP_SUBMODE sets both the mode and submode */
 574         req->flags = htonl(V_TX_ULP_SUBMODE(skb_ulp_mode(skb)) |
 575                            V_TX_SHOVE((skb_peek(&c3cn->write_queue) ? 0 : 1)));
 576
 577         if (!c3cn_flag(c3cn, C3CN_TX_DATA_SENT)) {
 578                 req->flags |= htonl(V_TX_ACK_PAGES(2) | F_TX_INIT |
 579                                     V_TX_CPU_IDX(c3cn->qset));
 580                 /* Sendbuffer is in units of 32KB. */
 581                 req->param |= htonl(V_TX_SNDBUF(cxgb3_snd_win >> 15));
 582                 c3cn_set_flag(c3cn, C3CN_TX_DATA_SENT);
 583         }
 584 }
 585
 586 /**
 587  * c3cn_push_tx_frames -- start transmit
 588  * @c3cn: the offloaded connection
 589  * @req_completion: request wr_ack or not
 590  *
 591  * Prepends TX_DATA_WR or CPL_CLOSE_CON_REQ headers to buffers waiting in a
 592  * connection's send queue and sends them on to T3.  Must be called with the
 593  * connection's lock held.  Returns the amount of send buffer space that was
 594  * freed as a result of sending queued data to T3.
 595  */
 596 static void arp_failure_discard(struct t3cdev *cdev, struct sk_buff *skb)
 597 {
 598         kfree_skb(skb);
 599 }
 600
 601 static int c3cn_push_tx_frames(struct s3_conn *c3cn, int req_completion)
 602 {
 603         int total_size = 0;
 604         struct sk_buff *skb;
 605         struct t3cdev *cdev;
 606         struct cxgb3i_sdev_data *cdata;
 607
 608         if (unlikely(c3cn->state == C3CN_STATE_CONNECTING ||
 609                      c3cn->state == C3CN_STATE_CLOSE_WAIT_1 ||
 610                      c3cn->state >= C3CN_STATE_ABORTING)) {
 611                 c3cn_tx_debug("c3cn 0x%p, in closing state %u.\n",
 612                               c3cn, c3cn->state);
 613                 return 0;
 614         }
 615
 616         cdev = c3cn->cdev;
 617         cdata = CXGB3_SDEV_DATA(cdev);
 618
 619         while (c3cn->wr_avail
 620                && (skb = skb_peek(&c3cn->write_queue)) != NULL) {
 621                 int len = skb->len;     /* length before skb_push */
 622                 int frags = skb_shinfo(skb)->nr_frags + (len != skb->data_len);
 623                 int wrs_needed = skb_wrs[frags];
 624
 625                 if (wrs_needed > 1 && len + sizeof(struct tx_data_wr) <= wrlen)
 626                         wrs_needed = 1;
 627
 628                 WARN_ON(frags >= SKB_WR_LIST_SIZE || wrs_needed < 1);
 629
 630                 if (c3cn->wr_avail < wrs_needed) {
 631                         c3cn_tx_debug("c3cn 0x%p, skb len %u/%u, frag %u, "
 632                                       "wr %d < %u.\n",
 633                                       c3cn, skb->len, skb->data_len, frags,
 634                                       wrs_needed, c3cn->wr_avail);
 635                         break;
 636                 }
 637
 638                 __skb_unlink(skb, &c3cn->write_queue);
 639                 skb->priority = CPL_PRIORITY_DATA;
 640                 skb->csum = wrs_needed; /* remember this until the WR_ACK */
 641                 c3cn->wr_avail -= wrs_needed;
 642                 c3cn->wr_unacked += wrs_needed;
 643                 enqueue_wr(c3cn, skb);
 644
 645                 c3cn_tx_debug("c3cn 0x%p, enqueue, skb len %u/%u, frag %u, "
 646                                 "wr %d, left %u, unack %u.\n",
 647                                 c3cn, skb->len, skb->data_len, frags,
 648                                 wrs_needed, c3cn->wr_avail, c3cn->wr_unacked);
 649
 650
 651                 if (likely(skb_flags(skb) & C3CB_FLAG_NEED_HDR)) {
 652                         if ((req_completion &&
 653                                 c3cn->wr_unacked == wrs_needed) ||
 654                             (skb_flags(skb) & C3CB_FLAG_COMPL) ||
 655                             c3cn->wr_unacked >= c3cn->wr_max / 2) {
 656                                 req_completion = 1;
 657                                 c3cn->wr_unacked = 0;
 658                         }
 659                         len += ulp_extra_len(skb);
 660                         make_tx_data_wr(c3cn, skb, len, req_completion);
 661                         c3cn->snd_nxt += len;
 662                         skb_flags(skb) &= ~C3CB_FLAG_NEED_HDR;
 663                 }
 664
 665                 total_size += skb->truesize;
 666                 set_arp_failure_handler(skb, arp_failure_discard);
 667                 l2t_send(cdev, skb, c3cn->l2t);
 668         }
 669         return total_size;
 670 }
 671
 672 /*
 673  * process_cpl_msg: -> host
 674  * Top-level CPL message processing used by most CPL messages that
 675  * pertain to connections.
 676  */
 677 static inline void process_cpl_msg(void (*fn)(struct s3_conn *,
 678                                               struct sk_buff *),
 679                                    struct s3_conn *c3cn,
 680                                    struct sk_buff *skb)
 681 {
 682         spin_lock_bh(&c3cn->lock);
 683         fn(c3cn, skb);
 684         spin_unlock_bh(&c3cn->lock);
 685 }
 686
 687 /*
 688  * process_cpl_msg_ref: -> host
 689  * Similar to process_cpl_msg() but takes an extra connection reference around
 690  * the call to the handler.  Should be used if the handler may drop a
 691  * connection reference.
 692  */
 693 static inline void process_cpl_msg_ref(void (*fn) (struct s3_conn *,
 694                                                    struct sk_buff *),
 695                                        struct s3_conn *c3cn,
 696                                        struct sk_buff *skb)
 697 {
 698         c3cn_hold(c3cn);
 699         process_cpl_msg(fn, c3cn, skb);
 700         c3cn_put(c3cn);
 701 }
 702
 703 /*
 704  * Process a CPL_ACT_ESTABLISH message: -> host
 705  * Updates connection state from an active establish CPL message.  Runs with
 706  * the connection lock held.
 707  */
 708
 709 static inline void s3_free_atid(struct t3cdev *cdev, unsigned int tid)
 710 {
 711         struct s3_conn *c3cn = cxgb3_free_atid(cdev, tid);
 712         if (c3cn)
 713                 c3cn_put(c3cn);
 714 }
 715
 716 static void c3cn_established(struct s3_conn *c3cn, u32 snd_isn,
 717                              unsigned int opt)
 718 {
 719         c3cn_conn_debug("c3cn 0x%p, state %u.\n", c3cn, c3cn->state);
 720
 721         c3cn->write_seq = c3cn->snd_nxt = c3cn->snd_una = snd_isn;
 722
 723         /*
 724          * Causes the first RX_DATA_ACK to supply any Rx credits we couldn't
 725          * pass through opt0.
 726          */
 727         if (cxgb3_rcv_win > (M_RCV_BUFSIZ << 10))
 728                 c3cn->rcv_wup -= cxgb3_rcv_win - (M_RCV_BUFSIZ << 10);
 729
 730         dst_confirm(c3cn->dst_cache);
 731
 732         smp_mb();
 733
 734         c3cn_set_state(c3cn, C3CN_STATE_ESTABLISHED);
 735 }
 736
 737 static void process_act_establish(struct s3_conn *c3cn, struct sk_buff *skb)
 738 {
 739         struct cpl_act_establish *req = cplhdr(skb);
 740         u32 rcv_isn = ntohl(req->rcv_isn);      /* real RCV_ISN + 1 */
 741
 742         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
 743                         c3cn, c3cn->state, c3cn->flags);
 744
 745         if (unlikely(c3cn->state != C3CN_STATE_CONNECTING))
 746                 cxgb3i_log_error("TID %u expected SYN_SENT, got EST., s %u\n",
 747                                  c3cn->tid, c3cn->state);
 748
 749         c3cn->copied_seq = c3cn->rcv_wup = c3cn->rcv_nxt = rcv_isn;
 750         c3cn_established(c3cn, ntohl(req->snd_isn), ntohs(req->tcp_opt));
 751
 752         __kfree_skb(skb);
 753
 754         if (unlikely(c3cn_flag(c3cn, C3CN_ACTIVE_CLOSE_NEEDED)))
 755                 /* upper layer has requested closing */
 756                 send_abort_req(c3cn);
 757         else {
 758                 if (skb_queue_len(&c3cn->write_queue))
 759                         c3cn_push_tx_frames(c3cn, 1);
 760                 cxgb3i_conn_tx_open(c3cn);
 761         }
 762 }
 763
 764 static int do_act_establish(struct t3cdev *cdev, struct sk_buff *skb,
 765                             void *ctx)
 766 {
 767         struct cpl_act_establish *req = cplhdr(skb);
 768         unsigned int tid = GET_TID(req);
 769         unsigned int atid = G_PASS_OPEN_TID(ntohl(req->tos_tid));
 770         struct s3_conn *c3cn = ctx;
 771         struct cxgb3i_sdev_data *cdata = CXGB3_SDEV_DATA(cdev);
 772
 773         c3cn_conn_debug("rcv, tid 0x%x, c3cn 0x%p, s %u, f 0x%lx.\n",
 774                         tid, c3cn, c3cn->state, c3cn->flags);
 775
 776         c3cn->tid = tid;
 777         c3cn_hold(c3cn);
 778         cxgb3_insert_tid(cdata->cdev, cdata->client, c3cn, tid);
 779         s3_free_atid(cdev, atid);
 780
 781         c3cn->qset = G_QNUM(ntohl(skb->csum));
 782
 783         process_cpl_msg(process_act_establish, c3cn, skb);
 784         return 0;
 785 }
 786
 787 /*
 788  * Process a CPL_ACT_OPEN_RPL message: -> host
 789  * Handle active open failures.
 790  */
 791 static int act_open_rpl_status_to_errno(int status)
 792 {
 793         switch (status) {
 794         case CPL_ERR_CONN_RESET:
 795                 return ECONNREFUSED;
 796         case CPL_ERR_ARP_MISS:
 797                 return EHOSTUNREACH;
 798         case CPL_ERR_CONN_TIMEDOUT:
 799                 return ETIMEDOUT;
 800         case CPL_ERR_TCAM_FULL:
 801                 return ENOMEM;
 802         case CPL_ERR_CONN_EXIST:
 803                 cxgb3i_log_error("ACTIVE_OPEN_RPL: 4-tuple in use\n");
 804                 return EADDRINUSE;
 805         default:
 806                 return EIO;
 807         }
 808 }
 809
 810 static void act_open_retry_timer(unsigned long data)
 811 {
 812         struct sk_buff *skb;
 813         struct s3_conn *c3cn = (struct s3_conn *)data;
 814
 815         c3cn_conn_debug("c3cn 0x%p, state %u.\n", c3cn, c3cn->state);
 816
 817         spin_lock_bh(&c3cn->lock);
 818         skb = alloc_skb(sizeof(struct cpl_act_open_req), GFP_ATOMIC);
 819         if (!skb)
 820                 fail_act_open(c3cn, ENOMEM);
 821         else {
 822                 skb->sk = (struct sock *)c3cn;
 823                 set_arp_failure_handler(skb, act_open_req_arp_failure);
 824                 make_act_open_req(c3cn, skb, c3cn->tid, c3cn->l2t);
 825                 l2t_send(c3cn->cdev, skb, c3cn->l2t);
 826         }
 827         spin_unlock_bh(&c3cn->lock);
 828         c3cn_put(c3cn);
 829 }
 830
 831 static void process_act_open_rpl(struct s3_conn *c3cn, struct sk_buff *skb)
 832 {
 833         struct cpl_act_open_rpl *rpl = cplhdr(skb);
 834
 835         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
 836                         c3cn, c3cn->state, c3cn->flags);
 837
 838         if (rpl->status == CPL_ERR_CONN_EXIST &&
 839             c3cn->retry_timer.function != act_open_retry_timer) {
 840                 c3cn->retry_timer.function = act_open_retry_timer;
 841                 if (!mod_timer(&c3cn->retry_timer, jiffies + HZ / 2))
 842                         c3cn_hold(c3cn);
 843         } else
 844                 fail_act_open(c3cn, act_open_rpl_status_to_errno(rpl->status));
 845         __kfree_skb(skb);
 846 }
 847
 848 static int do_act_open_rpl(struct t3cdev *cdev, struct sk_buff *skb, void *ctx)
 849 {
 850         struct s3_conn *c3cn = ctx;
 851         struct cpl_act_open_rpl *rpl = cplhdr(skb);
 852
 853         c3cn_conn_debug("rcv, status 0x%x, c3cn 0x%p, s %u, f 0x%lx.\n",
 854                         rpl->status, c3cn, c3cn->state, c3cn->flags);
 855
 856         if (rpl->status != CPL_ERR_TCAM_FULL &&
 857             rpl->status != CPL_ERR_CONN_EXIST &&
 858             rpl->status != CPL_ERR_ARP_MISS)
 859                 cxgb3_queue_tid_release(cdev, GET_TID(rpl));
 860
 861         process_cpl_msg_ref(process_act_open_rpl, c3cn, skb);
 862         return 0;
 863 }
 864
 865 /*
 866  * Process PEER_CLOSE CPL messages: -> host
 867  * Handle peer FIN.
 868  */
 869 static void process_peer_close(struct s3_conn *c3cn, struct sk_buff *skb)
 870 {
 871         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
 872                         c3cn, c3cn->state, c3cn->flags);
 873
 874         if (c3cn_flag(c3cn, C3CN_ABORT_RPL_PENDING))
 875                 goto out;
 876
 877         switch (c3cn->state) {
 878         case C3CN_STATE_ESTABLISHED:
 879                 c3cn_set_state(c3cn, C3CN_STATE_PASSIVE_CLOSE);
 880                 break;
 881         case C3CN_STATE_ACTIVE_CLOSE:
 882                 c3cn_set_state(c3cn, C3CN_STATE_CLOSE_WAIT_2);
 883                 break;
 884         case C3CN_STATE_CLOSE_WAIT_1:
 885                 c3cn_closed(c3cn);
 886                 break;
 887         case C3CN_STATE_ABORTING:
 888                 break;
 889         default:
 890                 cxgb3i_log_error("%s: peer close, TID %u in bad state %u\n",
 891                                  c3cn->cdev->name, c3cn->tid, c3cn->state);
 892         }
 893
 894         cxgb3i_conn_closing(c3cn);
 895 out:
 896         __kfree_skb(skb);
 897 }
 898
 899 static int do_peer_close(struct t3cdev *cdev, struct sk_buff *skb, void *ctx)
 900 {
 901         struct s3_conn *c3cn = ctx;
 902
 903         c3cn_conn_debug("rcv, c3cn 0x%p, s %u, f 0x%lx.\n",
 904                         c3cn, c3cn->state, c3cn->flags);
 905         process_cpl_msg_ref(process_peer_close, c3cn, skb);
 906         return 0;
 907 }
 908
 909 /*
 910  * Process CLOSE_CONN_RPL CPL message: -> host
 911  * Process a peer ACK to our FIN.
 912  */
 913 static void process_close_con_rpl(struct s3_conn *c3cn, struct sk_buff *skb)
 914 {
 915         struct cpl_close_con_rpl *rpl = cplhdr(skb);
 916
 917         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
 918                         c3cn, c3cn->state, c3cn->flags);
 919
 920         c3cn->snd_una = ntohl(rpl->snd_nxt) - 1;        /* exclude FIN */
 921
 922         if (c3cn_flag(c3cn, C3CN_ABORT_RPL_PENDING))
 923                 goto out;
 924
 925         switch (c3cn->state) {
 926         case C3CN_STATE_ACTIVE_CLOSE:
 927                 c3cn_set_state(c3cn, C3CN_STATE_CLOSE_WAIT_1);
 928                 break;
 929         case C3CN_STATE_CLOSE_WAIT_1:
 930         case C3CN_STATE_CLOSE_WAIT_2:
 931                 c3cn_closed(c3cn);
 932                 break;
 933         case C3CN_STATE_ABORTING:
 934                 break;
 935         default:
 936                 cxgb3i_log_error("%s: close_rpl, TID %u in bad state %u\n",
 937                                  c3cn->cdev->name, c3cn->tid, c3cn->state);
 938         }
 939
 940 out:
 941         kfree_skb(skb);
 942 }
 943
 944 static int do_close_con_rpl(struct t3cdev *cdev, struct sk_buff *skb,
 945                             void *ctx)
 946 {
 947         struct s3_conn *c3cn = ctx;
 948
 949         c3cn_conn_debug("rcv, c3cn 0x%p, s %u, f 0x%lx.\n",
 950                          c3cn, c3cn->state, c3cn->flags);
 951
 952         process_cpl_msg_ref(process_close_con_rpl, c3cn, skb);
 953         return 0;
 954 }
 955
 956 /*
 957  * Process ABORT_REQ_RSS CPL message: -> host
 958  * Process abort requests.  If we are waiting for an ABORT_RPL we ignore this
 959  * request except that we need to reply to it.
 960  */
 961
 962 static int abort_status_to_errno(struct s3_conn *c3cn, int abort_reason,
 963                                  int *need_rst)
 964 {
 965         switch (abort_reason) {
 966         case CPL_ERR_BAD_SYN: /* fall through */
 967         case CPL_ERR_CONN_RESET:
 968                 return c3cn->state > C3CN_STATE_ESTABLISHED ?
 969                         EPIPE : ECONNRESET;
 970         case CPL_ERR_XMIT_TIMEDOUT:
 971         case CPL_ERR_PERSIST_TIMEDOUT:
 972         case CPL_ERR_FINWAIT2_TIMEDOUT:
 973         case CPL_ERR_KEEPALIVE_TIMEDOUT:
 974                 return ETIMEDOUT;
 975         default:
 976                 return EIO;
 977         }
 978 }
 979
 980 static void process_abort_req(struct s3_conn *c3cn, struct sk_buff *skb)
 981 {
 982         int rst_status = CPL_ABORT_NO_RST;
 983         const struct cpl_abort_req_rss *req = cplhdr(skb);
 984
 985         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
 986                         c3cn, c3cn->state, c3cn->flags);
 987
 988         if (!c3cn_flag(c3cn, C3CN_ABORT_REQ_RCVD)) {
 989                 c3cn_set_flag(c3cn, C3CN_ABORT_REQ_RCVD);
 990                 c3cn_set_state(c3cn, C3CN_STATE_ABORTING);
 991                 __kfree_skb(skb);
 992                 return;
 993         }
 994
 995         c3cn_clear_flag(c3cn, C3CN_ABORT_REQ_RCVD);
 996         send_abort_rpl(c3cn, rst_status);
 997
 998         if (!c3cn_flag(c3cn, C3CN_ABORT_RPL_PENDING)) {
 999                 c3cn->err =
1000                     abort_status_to_errno(c3cn, req->status, &rst_status);
1001                 c3cn_closed(c3cn);
1002         }
1003 }
1004
1005 static int do_abort_req(struct t3cdev *cdev, struct sk_buff *skb, void *ctx)
1006 {
1007         const struct cpl_abort_req_rss *req = cplhdr(skb);
1008         struct s3_conn *c3cn = ctx;
1009
1010         c3cn_conn_debug("rcv, c3cn 0x%p, s 0x%x, f 0x%lx.\n",
1011                         c3cn, c3cn->state, c3cn->flags);
1012
1013         if (req->status == CPL_ERR_RTX_NEG_ADVICE ||
1014             req->status == CPL_ERR_PERSIST_NEG_ADVICE) {
1015                 __kfree_skb(skb);
1016                 return 0;
1017         }
1018
1019         process_cpl_msg_ref(process_abort_req, c3cn, skb);
1020         return 0;
1021 }
1022
1023 /*
1024  * Process ABORT_RPL_RSS CPL message: -> host
1025  * Process abort replies.  We only process these messages if we anticipate
1026  * them as the coordination between SW and HW in this area is somewhat lacking
1027  * and sometimes we get ABORT_RPLs after we are done with the connection that
1028  * originated the ABORT_REQ.
1029  */
1030 static void process_abort_rpl(struct s3_conn *c3cn, struct sk_buff *skb)
1031 {
1032         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
1033                         c3cn, c3cn->state, c3cn->flags);
1034
1035         if (c3cn_flag(c3cn, C3CN_ABORT_RPL_PENDING)) {
1036                 if (!c3cn_flag(c3cn, C3CN_ABORT_RPL_RCVD))
1037                         c3cn_set_flag(c3cn, C3CN_ABORT_RPL_RCVD);
1038                 else {
1039                         c3cn_clear_flag(c3cn, C3CN_ABORT_RPL_RCVD);
1040                         c3cn_clear_flag(c3cn, C3CN_ABORT_RPL_PENDING);
1041                         if (c3cn_flag(c3cn, C3CN_ABORT_REQ_RCVD))
1042                                 cxgb3i_log_error("%s tid %u, ABORT_RPL_RSS\n",
1043                                                  c3cn->cdev->name, c3cn->tid);
1044                         c3cn_closed(c3cn);
1045                 }
1046         }
1047         __kfree_skb(skb);
1048 }
1049
1050 static int do_abort_rpl(struct t3cdev *cdev, struct sk_buff *skb, void *ctx)
1051 {
1052         struct cpl_abort_rpl_rss *rpl = cplhdr(skb);
1053         struct s3_conn *c3cn = ctx;
1054
1055         c3cn_conn_debug("rcv, status 0x%x, c3cn 0x%p, s %u, 0x%lx.\n",
1056                         rpl->status, c3cn, c3cn ? c3cn->state : 0,
1057                         c3cn ? c3cn->flags : 0UL);
1058
1059         /*
1060          * Ignore replies to post-close aborts indicating that the abort was
1061          * requested too late.  These connections are terminated when we get
1062          * PEER_CLOSE or CLOSE_CON_RPL and by the time the abort_rpl_rss
1063          * arrives the TID is either no longer used or it has been recycled.
1064          */
1065         if (rpl->status == CPL_ERR_ABORT_FAILED)
1066                 goto discard;
1067
1068         /*
1069          * Sometimes we've already closed the connection, e.g., a post-close
1070          * abort races with ABORT_REQ_RSS, the latter frees the connection
1071          * expecting the ABORT_REQ will fail with CPL_ERR_ABORT_FAILED,
1072          * but FW turns the ABORT_REQ into a regular one and so we get
1073          * ABORT_RPL_RSS with status 0 and no connection.
1074          */
1075         if (!c3cn)
1076                 goto discard;
1077
1078         process_cpl_msg_ref(process_abort_rpl, c3cn, skb);
1079         return 0;
1080
1081 discard:
1082         __kfree_skb(skb);
1083         return 0;
1084 }
1085
1086 /*
1087  * Process RX_ISCSI_HDR CPL message: -> host
1088  * Handle received PDUs, the payload could be DDP'ed. If not, the payload
1089  * follow after the bhs.
1090  */
1091 static void process_rx_iscsi_hdr(struct s3_conn *c3cn, struct sk_buff *skb)
1092 {
1093         struct cpl_iscsi_hdr *hdr_cpl = cplhdr(skb);
1094         struct cpl_iscsi_hdr_norss data_cpl;
1095         struct cpl_rx_data_ddp_norss ddp_cpl;
1096         unsigned int hdr_len, data_len, status;
1097         unsigned int len;
1098         int err;
1099
1100         if (unlikely(c3cn->state >= C3CN_STATE_PASSIVE_CLOSE)) {
1101                 if (c3cn->state != C3CN_STATE_ABORTING)
1102                         send_abort_req(c3cn);
1103                 __kfree_skb(skb);
1104                 return;
1105         }
1106
1107         skb_tcp_seq(skb) = ntohl(hdr_cpl->seq);
1108         skb_flags(skb) = 0;
1109
1110         skb_reset_transport_header(skb);
1111         __skb_pull(skb, sizeof(struct cpl_iscsi_hdr));
1112
1113         len = hdr_len = ntohs(hdr_cpl->len);
1114         /* msg coalesce is off or not enough data received */
1115         if (skb->len <= hdr_len) {
1116                 cxgb3i_log_error("%s: TID %u, ISCSI_HDR, skb len %u < %u.\n",
1117                                  c3cn->cdev->name, c3cn->tid,
1118                                  skb->len, hdr_len);
1119                 goto abort_conn;
1120         }
1121
1122         err = skb_copy_bits(skb, skb->len - sizeof(ddp_cpl), &ddp_cpl,
1123                             sizeof(ddp_cpl));
1124         if (err < 0)
1125                 goto abort_conn;
1126
1127         skb_ulp_mode(skb) = ULP2_FLAG_DATA_READY;
1128         skb_rx_pdulen(skb) = ntohs(ddp_cpl.len);
1129         skb_rx_ddigest(skb) = ntohl(ddp_cpl.ulp_crc);
1130         status = ntohl(ddp_cpl.ddp_status);
1131
1132         c3cn_rx_debug("rx skb 0x%p, len %u, pdulen %u, ddp status 0x%x.\n",
1133                       skb, skb->len, skb_rx_pdulen(skb), status);
1134
1135         if (status & (1 << RX_DDP_STATUS_HCRC_SHIFT))
1136                 skb_ulp_mode(skb) |= ULP2_FLAG_HCRC_ERROR;
1137         if (status & (1 << RX_DDP_STATUS_DCRC_SHIFT))
1138                 skb_ulp_mode(skb) |= ULP2_FLAG_DCRC_ERROR;
1139         if (status & (1 << RX_DDP_STATUS_PAD_SHIFT))
1140                 skb_ulp_mode(skb) |= ULP2_FLAG_PAD_ERROR;
1141
1142         if (skb->len > (hdr_len + sizeof(ddp_cpl))) {
1143                 err = skb_copy_bits(skb, hdr_len, &data_cpl, sizeof(data_cpl));
1144                 if (err < 0)
1145                         goto abort_conn;
1146                 data_len = ntohs(data_cpl.len);
1147                 len += sizeof(data_cpl) + data_len;
1148         } else if (status & (1 << RX_DDP_STATUS_DDP_SHIFT))
1149                 skb_ulp_mode(skb) |= ULP2_FLAG_DATA_DDPED;
1150
1151         c3cn->rcv_nxt = ntohl(ddp_cpl.seq) + skb_rx_pdulen(skb);
1152         __pskb_trim(skb, len);
1153         __skb_queue_tail(&c3cn->receive_queue, skb);
1154         cxgb3i_conn_pdu_ready(c3cn);
1155
1156         return;
1157
1158 abort_conn:
1159         send_abort_req(c3cn);
1160         __kfree_skb(skb);
1161 }
1162
1163 static int do_iscsi_hdr(struct t3cdev *t3dev, struct sk_buff *skb, void *ctx)
1164 {
1165         struct s3_conn *c3cn = ctx;
1166
1167         process_cpl_msg(process_rx_iscsi_hdr, c3cn, skb);
1168         return 0;
1169 }
1170
1171 /*
1172  * Process TX_DATA_ACK CPL messages: -> host
1173  * Process an acknowledgment of WR completion.  Advance snd_una and send the
1174  * next batch of work requests from the write queue.
1175  */
1176 static void check_wr_invariants(struct s3_conn *c3cn)
1177 {
1178         int pending = count_pending_wrs(c3cn);
1179
1180         if (unlikely(c3cn->wr_avail + pending != c3cn->wr_max))
1181                 cxgb3i_log_error("TID %u: credit imbalance: avail %u, "
1182                                 "pending %u, total should be %u\n",
1183                                 c3cn->tid, c3cn->wr_avail, pending,
1184                                 c3cn->wr_max);
1185 }
1186
1187 static void process_wr_ack(struct s3_conn *c3cn, struct sk_buff *skb)
1188 {
1189         struct cpl_wr_ack *hdr = cplhdr(skb);
1190         unsigned int credits = ntohs(hdr->credits);
1191         u32 snd_una = ntohl(hdr->snd_una);
1192
1193         c3cn_tx_debug("%u WR credits, avail %u, unack %u, TID %u, state %u.\n",
1194                         credits, c3cn->wr_avail, c3cn->wr_unacked,
1195                         c3cn->tid, c3cn->state);
1196
1197         c3cn->wr_avail += credits;
1198         if (c3cn->wr_unacked > c3cn->wr_max - c3cn->wr_avail)
1199                 c3cn->wr_unacked = c3cn->wr_max - c3cn->wr_avail;
1200
1201         while (credits) {
1202                 struct sk_buff *p = peek_wr(c3cn);
1203
1204                 if (unlikely(!p)) {
1205                         cxgb3i_log_error("%u WR_ACK credits for TID %u with "
1206                                          "nothing pending, state %u\n",
1207                                          credits, c3cn->tid, c3cn->state);
1208                         break;
1209                 }
1210                 if (unlikely(credits < p->csum)) {
1211                         struct tx_data_wr *w = cplhdr(p);
1212                         cxgb3i_log_error("TID %u got %u WR credits need %u, "
1213                                          "len %u, main body %u, frags %u, "
1214                                          "seq # %u, ACK una %u, ACK nxt %u, "
1215                                          "WR_AVAIL %u, WRs pending %u\n",
1216                                          c3cn->tid, credits, p->csum, p->len,
1217                                          p->len - p->data_len,
1218                                          skb_shinfo(p)->nr_frags,
1219                                          ntohl(w->sndseq), snd_una,
1220                                          ntohl(hdr->snd_nxt), c3cn->wr_avail,
1221                                          count_pending_wrs(c3cn) - credits);
1222                         p->csum -= credits;
1223                         break;
1224                 } else {
1225                         dequeue_wr(c3cn);
1226                         credits -= p->csum;
1227                         free_wr_skb(p);
1228                 }
1229         }
1230
1231         check_wr_invariants(c3cn);
1232
1233         if (unlikely(before(snd_una, c3cn->snd_una))) {
1234                 cxgb3i_log_error("TID %u, unexpected sequence # %u in WR_ACK "
1235                                  "snd_una %u\n",
1236                                  c3cn->tid, snd_una, c3cn->snd_una);
1237                 goto out_free;
1238         }
1239
1240         if (c3cn->snd_una != snd_una) {
1241                 c3cn->snd_una = snd_una;
1242                 dst_confirm(c3cn->dst_cache);
1243         }
1244
1245         if (skb_queue_len(&c3cn->write_queue)) {
1246                 if (c3cn_push_tx_frames(c3cn, 0))
1247                         cxgb3i_conn_tx_open(c3cn);
1248         } else
1249                 cxgb3i_conn_tx_open(c3cn);
1250 out_free:
1251         __kfree_skb(skb);
1252 }
1253
1254 static int do_wr_ack(struct t3cdev *cdev, struct sk_buff *skb, void *ctx)
1255 {
1256         struct s3_conn *c3cn = ctx;
1257
1258         process_cpl_msg(process_wr_ack, c3cn, skb);
1259         return 0;
1260 }
1261
1262 /*
1263  * for each connection, pre-allocate skbs needed for close/abort requests. So
1264  * that we can service the request right away.
1265  */
1266 static void c3cn_free_cpl_skbs(struct s3_conn *c3cn)
1267 {
1268         if (c3cn->cpl_close)
1269                 kfree_skb(c3cn->cpl_close);
1270         if (c3cn->cpl_abort_req)
1271                 kfree_skb(c3cn->cpl_abort_req);
1272         if (c3cn->cpl_abort_rpl)
1273                 kfree_skb(c3cn->cpl_abort_rpl);
1274 }
1275
1276 static int c3cn_alloc_cpl_skbs(struct s3_conn *c3cn)
1277 {
1278         c3cn->cpl_close = alloc_skb(sizeof(struct cpl_close_con_req),
1279                                    GFP_KERNEL);
1280         if (!c3cn->cpl_close)
1281                 return -ENOMEM;
1282         skb_put(c3cn->cpl_close, sizeof(struct cpl_close_con_req));
1283
1284         c3cn->cpl_abort_req = alloc_skb(sizeof(struct cpl_abort_req),
1285                                         GFP_KERNEL);
1286         if (!c3cn->cpl_abort_req)
1287                 goto free_cpl_skbs;
1288         skb_put(c3cn->cpl_abort_req, sizeof(struct cpl_abort_req));
1289
1290         c3cn->cpl_abort_rpl = alloc_skb(sizeof(struct cpl_abort_rpl),
1291                                         GFP_KERNEL);
1292         if (!c3cn->cpl_abort_rpl)
1293                 goto free_cpl_skbs;
1294         skb_put(c3cn->cpl_abort_rpl, sizeof(struct cpl_abort_rpl));
1295
1296         return 0;
1297
1298 free_cpl_skbs:
1299         c3cn_free_cpl_skbs(c3cn);
1300         return -ENOMEM;
1301 }
1302
1303 /**
1304  * c3cn_release_offload_resources - release offload resource
1305  * @c3cn: the offloaded iscsi tcp connection.
1306  * Release resources held by an offload connection (TID, L2T entry, etc.)
1307  */
1308 static void c3cn_release_offload_resources(struct s3_conn *c3cn)
1309 {
1310         struct t3cdev *cdev = c3cn->cdev;
1311         unsigned int tid = c3cn->tid;
1312
1313         c3cn->qset = 0;
1314         c3cn_free_cpl_skbs(c3cn);
1315
1316         if (c3cn->wr_avail != c3cn->wr_max) {
1317                 purge_wr_queue(c3cn);
1318                 reset_wr_list(c3cn);
1319         }
1320
1321         if (cdev) {
1322                 if (c3cn->l2t) {
1323                         l2t_release(L2DATA(cdev), c3cn->l2t);
1324                         c3cn->l2t = NULL;
1325                 }
1326                 if (c3cn->state == C3CN_STATE_CONNECTING)
1327                         /* we have ATID */
1328                         s3_free_atid(cdev, tid);
1329                 else {
1330                         /* we have TID */
1331                         cxgb3_remove_tid(cdev, (void *)c3cn, tid);
1332                         c3cn_put(c3cn);
1333                 }
1334         }
1335
1336         c3cn->dst_cache = NULL;
1337         c3cn->cdev = NULL;
1338 }
1339
1340 /**
1341  * cxgb3i_c3cn_create - allocate and initialize an s3_conn structure
1342  * returns the s3_conn structure allocated.
1343  */
1344 struct s3_conn *cxgb3i_c3cn_create(void)
1345 {
1346         struct s3_conn *c3cn;
1347
1348         c3cn = kzalloc(sizeof(*c3cn), GFP_KERNEL);
1349         if (!c3cn)
1350                 return NULL;
1351
1352         /* pre-allocate close/abort cpl, so we don't need to wait for memory
1353            when close/abort is requested. */
1354         if (c3cn_alloc_cpl_skbs(c3cn) < 0)
1355                 goto free_c3cn;
1356
1357         c3cn_conn_debug("alloc c3cn 0x%p.\n", c3cn);
1358
1359         c3cn->flags = 0;
1360         spin_lock_init(&c3cn->lock);
1361         atomic_set(&c3cn->refcnt, 1);
1362         skb_queue_head_init(&c3cn->receive_queue);
1363         skb_queue_head_init(&c3cn->write_queue);
1364         setup_timer(&c3cn->retry_timer, NULL, (unsigned long)c3cn);
1365         rwlock_init(&c3cn->callback_lock);
1366
1367         return c3cn;
1368
1369 free_c3cn:
1370         kfree(c3cn);
1371         return NULL;
1372 }
1373
1374 static void c3cn_active_close(struct s3_conn *c3cn)
1375 {
1376         int data_lost;
1377         int close_req = 0;
1378
1379         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
1380                          c3cn, c3cn->state, c3cn->flags);
1381
1382         dst_confirm(c3cn->dst_cache);
1383
1384         c3cn_hold(c3cn);
1385         spin_lock_bh(&c3cn->lock);
1386
1387         data_lost = skb_queue_len(&c3cn->receive_queue);
1388         __skb_queue_purge(&c3cn->receive_queue);
1389
1390         switch (c3cn->state) {
1391         case C3CN_STATE_CLOSED:
1392         case C3CN_STATE_ACTIVE_CLOSE:
1393         case C3CN_STATE_CLOSE_WAIT_1:
1394         case C3CN_STATE_CLOSE_WAIT_2:
1395         case C3CN_STATE_ABORTING:
1396                 /* nothing need to be done */
1397                 break;
1398         case C3CN_STATE_CONNECTING:
1399                 /* defer until cpl_act_open_rpl or cpl_act_establish */
1400                 c3cn_set_flag(c3cn, C3CN_ACTIVE_CLOSE_NEEDED);
1401                 break;
1402         case C3CN_STATE_ESTABLISHED:
1403                 close_req = 1;
1404                 c3cn_set_state(c3cn, C3CN_STATE_ACTIVE_CLOSE);
1405                 break;
1406         case C3CN_STATE_PASSIVE_CLOSE:
1407                 close_req = 1;
1408                 c3cn_set_state(c3cn, C3CN_STATE_CLOSE_WAIT_2);
1409                 break;
1410         }
1411
1412         if (close_req) {
1413                 if (data_lost)
1414                         /* Unread data was tossed, zap the connection. */
1415                         send_abort_req(c3cn);
1416                 else
1417                         send_close_req(c3cn);
1418         }
1419
1420         spin_unlock_bh(&c3cn->lock);
1421         c3cn_put(c3cn);
1422 }
1423
1424 /**
1425  * cxgb3i_c3cn_release - close and release an iscsi tcp connection and any
1426  *      resource held
1427  * @c3cn: the iscsi tcp connection
1428  */
1429 void cxgb3i_c3cn_release(struct s3_conn *c3cn)
1430 {
1431         c3cn_conn_debug("c3cn 0x%p, s %u, f 0x%lx.\n",
1432                         c3cn, c3cn->state, c3cn->flags);
1433         if (unlikely(c3cn->state == C3CN_STATE_CONNECTING))
1434                 c3cn_set_flag(c3cn, C3CN_ACTIVE_CLOSE_NEEDED);
1435         else if (likely(c3cn->state != C3CN_STATE_CLOSED))
1436                 c3cn_active_close(c3cn);
1437         c3cn_put(c3cn);
1438 }
1439
1440 static int is_cxgb3_dev(struct net_device *dev)
1441 {
1442         struct cxgb3i_sdev_data *cdata;
1443
1444         write_lock(&cdata_rwlock);
1445         list_for_each_entry(cdata, &cdata_list, list) {
1446                 struct adap_ports *ports = &cdata->ports;
1447                 int i;
1448
1449                 for (i = 0; i < ports->nports; i++)
1450                         if (dev == ports->lldevs[i]) {
1451                                 write_unlock(&cdata_rwlock);
1452                                 return 1;
1453                         }
1454         }
1455         write_unlock(&cdata_rwlock);
1456         return 0;
1457 }
1458
1459 /**
1460  * cxgb3_egress_dev - return the cxgb3 egress device
1461  * @root_dev: the root device anchoring the search
1462  * @c3cn: the connection used to determine egress port in bonding mode
1463  * @context: in bonding mode, indicates a connection set up or failover
1464  *
1465  * Return egress device or NULL if the egress device isn't one of our ports.
1466  */
1467 static struct net_device *cxgb3_egress_dev(struct net_device *root_dev,
1468                                            struct s3_conn *c3cn,
1469                                            int context)
1470 {
1471         while (root_dev) {
1472                 if (root_dev->priv_flags & IFF_802_1Q_VLAN)
1473                         root_dev = vlan_dev_real_dev(root_dev);
1474                 else if (is_cxgb3_dev(root_dev))
1475                         return root_dev;
1476                 else
1477                         return NULL;
1478         }
1479         return NULL;
1480 }
1481
1482 static struct rtable *find_route(__be32 saddr, __be32 daddr,
1483                                  __be16 sport, __be16 dport)
1484 {
1485         struct rtable *rt;
1486         struct flowi fl = {
1487                 .oif = 0,
1488                 .nl_u = {
1489                          .ip4_u = {
1490                                    .daddr = daddr,
1491                                    .saddr = saddr,
1492                                    .tos = 0 } },
1493                 .proto = IPPROTO_TCP,
1494                 .uli_u = {
1495                           .ports = {
1496                                     .sport = sport,
1497                                     .dport = dport } } };
1498
1499         if (ip_route_output_flow(&init_net, &rt, &fl, NULL, 0))
1500                 return NULL;
1501         return rt;
1502 }
1503
1504 /*
1505  * Assign offload parameters to some connection fields.
1506  */
1507 static void init_offload_conn(struct s3_conn *c3cn,
1508                               struct t3cdev *cdev,
1509                               struct dst_entry *dst)
1510 {
1511         BUG_ON(c3cn->cdev != cdev);
1512         c3cn->wr_max = c3cn->wr_avail = T3C_DATA(cdev)->max_wrs - 1;
1513         c3cn->wr_unacked = 0;
1514         c3cn->mss_idx = select_mss(c3cn, dst_mtu(dst));
1515
1516         reset_wr_list(c3cn);
1517 }
1518
1519 static int initiate_act_open(struct s3_conn *c3cn, struct net_device *dev)
1520 {
1521         struct cxgb3i_sdev_data *cdata = NDEV2CDATA(dev);
1522         struct t3cdev *cdev = cdata->cdev;
1523         struct dst_entry *dst = c3cn->dst_cache;
1524         struct sk_buff *skb;
1525
1526         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
1527                         c3cn, c3cn->state, c3cn->flags);
1528         /*
1529          * Initialize connection data.  Note that the flags and ULP mode are
1530          * initialized higher up ...
1531          */
1532         c3cn->dev = dev;
1533         c3cn->cdev = cdev;
1534         c3cn->tid = cxgb3_alloc_atid(cdev, cdata->client, c3cn);
1535         if (c3cn->tid < 0)
1536                 goto out_err;
1537
1538         c3cn->qset = 0;
1539         c3cn->l2t = t3_l2t_get(cdev, dst->neighbour, dev);
1540         if (!c3cn->l2t)
1541                 goto free_tid;
1542
1543         skb = alloc_skb(sizeof(struct cpl_act_open_req), GFP_KERNEL);
1544         if (!skb)
1545                 goto free_l2t;
1546
1547         skb->sk = (struct sock *)c3cn;
1548         set_arp_failure_handler(skb, act_open_req_arp_failure);
1549
1550         c3cn_hold(c3cn);
1551
1552         init_offload_conn(c3cn, cdev, dst);
1553         c3cn->err = 0;
1554
1555         make_act_open_req(c3cn, skb, c3cn->tid, c3cn->l2t);
1556         l2t_send(cdev, skb, c3cn->l2t);
1557         return 0;
1558
1559 free_l2t:
1560         l2t_release(L2DATA(cdev), c3cn->l2t);
1561 free_tid:
1562         s3_free_atid(cdev, c3cn->tid);
1563         c3cn->tid = 0;
1564 out_err:
1565         return -1;
1566 }
1567
1568
1569 /**
1570  * cxgb3i_c3cn_connect - initiates an iscsi tcp connection to a given address
1571  * @c3cn: the iscsi tcp connection
1572  * @usin: destination address
1573  *
1574  * return 0 if active open request is sent, < 0 otherwise.
1575  */
1576 int cxgb3i_c3cn_connect(struct s3_conn *c3cn, struct sockaddr_in *usin)
1577 {
1578         struct rtable *rt;
1579         struct net_device *dev;
1580         struct cxgb3i_sdev_data *cdata;
1581         struct t3cdev *cdev;
1582         __be32 sipv4;
1583         int err;
1584
1585         if (usin->sin_family != AF_INET)
1586                 return -EAFNOSUPPORT;
1587
1588         c3cn->daddr.sin_port = usin->sin_port;
1589         c3cn->daddr.sin_addr.s_addr = usin->sin_addr.s_addr;
1590
1591         rt = find_route(c3cn->saddr.sin_addr.s_addr,
1592                         c3cn->daddr.sin_addr.s_addr,
1593                         c3cn->saddr.sin_port,
1594                         c3cn->daddr.sin_port);
1595         if (rt == NULL) {
1596                 c3cn_conn_debug("NO route to 0x%x, port %u.\n",
1597                                 c3cn->daddr.sin_addr.s_addr,
1598                                 ntohs(c3cn->daddr.sin_port));
1599                 return -ENETUNREACH;
1600         }
1601
1602         if (rt->rt_flags & (RTCF_MULTICAST | RTCF_BROADCAST)) {
1603                 c3cn_conn_debug("multi-cast route to 0x%x, port %u.\n",
1604                                 c3cn->daddr.sin_addr.s_addr,
1605                                 ntohs(c3cn->daddr.sin_port));
1606                 ip_rt_put(rt);
1607                 return -ENETUNREACH;
1608         }
1609
1610         if (!c3cn->saddr.sin_addr.s_addr)
1611                 c3cn->saddr.sin_addr.s_addr = rt->rt_src;
1612
1613         /* now commit destination to connection */
1614         c3cn->dst_cache = &rt->u.dst;
1615
1616         /* try to establish an offloaded connection */
1617         dev = cxgb3_egress_dev(c3cn->dst_cache->dev, c3cn, 0);
1618         if (dev == NULL) {
1619                 c3cn_conn_debug("c3cn 0x%p, egress dev NULL.\n", c3cn);
1620                 return -ENETUNREACH;
1621         }
1622         cdata = NDEV2CDATA(dev);
1623         cdev = cdata->cdev;
1624
1625         /* get a source port if one hasn't been provided */
1626         err = c3cn_get_port(c3cn, cdata);
1627         if (err)
1628                 return err;
1629
1630         c3cn_conn_debug("c3cn 0x%p get port %u.\n",
1631                         c3cn, ntohs(c3cn->saddr.sin_port));
1632
1633         sipv4 = cxgb3i_get_private_ipv4addr(dev);
1634         if (!sipv4) {
1635                 c3cn_conn_debug("c3cn 0x%p, iscsi ip not configured.\n", c3cn);
1636                 sipv4 = c3cn->saddr.sin_addr.s_addr;
1637                 cxgb3i_set_private_ipv4addr(dev, sipv4);
1638         } else
1639                 c3cn->saddr.sin_addr.s_addr = sipv4;
1640
1641         c3cn_conn_debug("c3cn 0x%p, %u.%u.%u.%u,%u-%u.%u.%u.%u,%u SYN_SENT.\n",
1642                         c3cn, NIPQUAD(c3cn->saddr.sin_addr.s_addr),
1643                         ntohs(c3cn->saddr.sin_port),
1644                         NIPQUAD(c3cn->daddr.sin_addr.s_addr),
1645                         ntohs(c3cn->daddr.sin_port));
1646
1647         c3cn_set_state(c3cn, C3CN_STATE_CONNECTING);
1648         if (!initiate_act_open(c3cn, dev))
1649                 return 0;
1650
1651         /*
1652          * If we get here, we don't have an offload connection so simply
1653          * return a failure.
1654          */
1655         err = -ENOTSUPP;
1656
1657         /*
1658          * This trashes the connection and releases the local port,
1659          * if necessary.
1660          */
1661         c3cn_conn_debug("c3cn 0x%p -> CLOSED.\n", c3cn);
1662         c3cn_set_state(c3cn, C3CN_STATE_CLOSED);
1663         ip_rt_put(rt);
1664         c3cn_put_port(c3cn);
1665         return err;
1666 }
1667
1668 /**
1669  * cxgb3i_c3cn_rx_credits - ack received tcp data.
1670  * @c3cn: iscsi tcp connection
1671  * @copied: # of bytes processed
1672  *
1673  * Called after some received data has been read.  It returns RX credits
1674  * to the HW for the amount of data processed.
1675  */
1676 void cxgb3i_c3cn_rx_credits(struct s3_conn *c3cn, int copied)
1677 {
1678         struct t3cdev *cdev;
1679         int must_send;
1680         u32 credits, dack = 0;
1681
1682         if (c3cn->state != C3CN_STATE_ESTABLISHED)
1683                 return;
1684
1685         credits = c3cn->copied_seq - c3cn->rcv_wup;
1686         if (unlikely(!credits))
1687                 return;
1688
1689         cdev = c3cn->cdev;
1690
1691         if (unlikely(cxgb3_rx_credit_thres == 0))
1692                 return;
1693
1694         dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
1695
1696         /*
1697          * For coalescing to work effectively ensure the receive window has
1698          * at least 16KB left.
1699          */
1700         must_send = credits + 16384 >= cxgb3_rcv_win;
1701
1702         if (must_send || credits >= cxgb3_rx_credit_thres)
1703                 c3cn->rcv_wup += send_rx_credits(c3cn, credits, dack);
1704 }
1705
1706 /**
1707  * cxgb3i_c3cn_send_pdus - send the skbs containing iscsi pdus
1708  * @c3cn: iscsi tcp connection
1709  * @skb: skb contains the iscsi pdu
1710  *
1711  * Add a list of skbs to a connection send queue. The skbs must comply with
1712  * the max size limit of the device and have a headroom of at least
1713  * TX_HEADER_LEN bytes.
1714  * Return # of bytes queued.
1715  */
1716 int cxgb3i_c3cn_send_pdus(struct s3_conn *c3cn, struct sk_buff *skb)
1717 {
1718         struct sk_buff *next;
1719         int err, copied = 0;
1720
1721         spin_lock_bh(&c3cn->lock);
1722
1723         if (c3cn->state != C3CN_STATE_ESTABLISHED) {
1724                 c3cn_tx_debug("c3cn 0x%p, not in est. state %u.\n",
1725                               c3cn, c3cn->state);
1726                 err = -EAGAIN;
1727                 goto out_err;
1728         }
1729
1730         if (c3cn->err) {
1731                 c3cn_tx_debug("c3cn 0x%p, err %d.\n", c3cn, c3cn->err);
1732                 err = -EPIPE;
1733                 goto out_err;
1734         }
1735
1736         if (c3cn->write_seq - c3cn->snd_una >= cxgb3_snd_win) {
1737                 c3cn_tx_debug("c3cn 0x%p, snd %u - %u > %u.\n",
1738                                 c3cn, c3cn->write_seq, c3cn->snd_una,
1739                                 cxgb3_snd_win);
1740                 err = -EAGAIN;
1741                 goto out_err;
1742         }
1743
1744         while (skb) {
1745                 int frags = skb_shinfo(skb)->nr_frags +
1746                                 (skb->len != skb->data_len);
1747
1748                 if (unlikely(skb_headroom(skb) < TX_HEADER_LEN)) {
1749                         c3cn_tx_debug("c3cn 0x%p, skb head.\n", c3cn);
1750                         err = -EINVAL;
1751                         goto out_err;
1752                 }
1753
1754                 if (frags >= SKB_WR_LIST_SIZE) {
1755                         cxgb3i_log_error("c3cn 0x%p, tx frags %d, len %u,%u.\n",
1756                                          c3cn, skb_shinfo(skb)->nr_frags,
1757                                          skb->len, skb->data_len);
1758                         err = -EINVAL;
1759                         goto out_err;
1760                 }
1761
1762                 next = skb->next;
1763                 skb->next = NULL;
1764                 skb_entail(c3cn, skb, C3CB_FLAG_NO_APPEND | C3CB_FLAG_NEED_HDR);
1765                 copied += skb->len;
1766                 c3cn->write_seq += skb->len + ulp_extra_len(skb);
1767                 skb = next;
1768         }
1769 done:
1770         if (likely(skb_queue_len(&c3cn->write_queue)))
1771                 c3cn_push_tx_frames(c3cn, 1);
1772         spin_unlock_bh(&c3cn->lock);
1773         return copied;
1774
1775 out_err:
1776         if (copied == 0 && err == -EPIPE)
1777                 copied = c3cn->err ? c3cn->err : -EPIPE;
1778         goto done;
1779 }
1780
1781 static void sdev_data_cleanup(struct cxgb3i_sdev_data *cdata)
1782 {
1783         struct adap_ports *ports = &cdata->ports;
1784         struct s3_conn *c3cn;
1785         int i;
1786
1787         for (i = 0; i < cxgb3_max_connect; i++) {
1788                 if (cdata->sport_conn[i]) {
1789                         c3cn = cdata->sport_conn[i];
1790                         cdata->sport_conn[i] = NULL;
1791
1792                         spin_lock_bh(&c3cn->lock);
1793                         c3cn->cdev = NULL;
1794                         c3cn_set_flag(c3cn, C3CN_OFFLOAD_DOWN);
1795                         c3cn_closed(c3cn);
1796                         spin_unlock_bh(&c3cn->lock);
1797                 }
1798         }
1799
1800         for (i = 0; i < ports->nports; i++)
1801                 NDEV2CDATA(ports->lldevs[i]) = NULL;
1802
1803         cxgb3i_free_big_mem(cdata);
1804 }
1805
1806 void cxgb3i_sdev_cleanup(void)
1807 {
1808         struct cxgb3i_sdev_data *cdata;
1809
1810         write_lock(&cdata_rwlock);
1811         list_for_each_entry(cdata, &cdata_list, list) {
1812                 list_del(&cdata->list);
1813                 sdev_data_cleanup(cdata);
1814         }
1815         write_unlock(&cdata_rwlock);
1816 }
1817
1818 int cxgb3i_sdev_init(cxgb3_cpl_handler_func *cpl_handlers)
1819 {
1820         cpl_handlers[CPL_ACT_ESTABLISH] = do_act_establish;
1821         cpl_handlers[CPL_ACT_OPEN_RPL] = do_act_open_rpl;
1822         cpl_handlers[CPL_PEER_CLOSE] = do_peer_close;
1823         cpl_handlers[CPL_ABORT_REQ_RSS] = do_abort_req;
1824         cpl_handlers[CPL_ABORT_RPL_RSS] = do_abort_rpl;
1825         cpl_handlers[CPL_CLOSE_CON_RPL] = do_close_con_rpl;
1826         cpl_handlers[CPL_TX_DMA_ACK] = do_wr_ack;
1827         cpl_handlers[CPL_ISCSI_HDR] = do_iscsi_hdr;
1828
1829         if (cxgb3_max_connect > CXGB3I_MAX_CONN)
1830                 cxgb3_max_connect = CXGB3I_MAX_CONN;
1831         return 0;
1832 }
1833
1834 /**
1835  * cxgb3i_sdev_add - allocate and initialize resources for each adapter found
1836  * @cdev:       t3cdev adapter
1837  * @client:     cxgb3 driver client
1838  */
1839 void cxgb3i_sdev_add(struct t3cdev *cdev, struct cxgb3_client *client)
1840 {
1841         struct cxgb3i_sdev_data *cdata;
1842         struct ofld_page_info rx_page_info;
1843         unsigned int wr_len;
1844         int mapsize = cxgb3_max_connect * sizeof(struct s3_conn *);
1845         int i;
1846
1847         cdata =  cxgb3i_alloc_big_mem(sizeof(*cdata) + mapsize, GFP_KERNEL);
1848         if (!cdata) {
1849                 cxgb3i_log_warn("t3dev 0x%p, offload up, OOM %d.\n",
1850                                 cdev, mapsize);
1851                 return;
1852         }
1853
1854         if (cdev->ctl(cdev, GET_WR_LEN, &wr_len) < 0 ||
1855             cdev->ctl(cdev, GET_PORTS, &cdata->ports) < 0 ||
1856             cdev->ctl(cdev, GET_RX_PAGE_INFO, &rx_page_info) < 0) {
1857                 cxgb3i_log_warn("t3dev 0x%p, offload up, ioctl failed.\n",
1858                                 cdev);
1859                 goto free_cdata;
1860         }
1861
1862         s3_init_wr_tab(wr_len);
1863
1864         spin_lock_init(&cdata->lock);
1865         INIT_LIST_HEAD(&cdata->list);
1866         cdata->cdev = cdev;
1867         cdata->client = client;
1868
1869         for (i = 0; i < cdata->ports.nports; i++)
1870                 NDEV2CDATA(cdata->ports.lldevs[i]) = cdata;
1871
1872         write_lock(&cdata_rwlock);
1873         list_add_tail(&cdata->list, &cdata_list);
1874         write_unlock(&cdata_rwlock);
1875
1876         cxgb3i_log_info("t3dev 0x%p, offload up, added.\n", cdev);
1877         return;
1878
1879 free_cdata:
1880         cxgb3i_free_big_mem(cdata);
1881 }
1882
1883 /**
1884  * cxgb3i_sdev_remove - free the allocated resources for the adapter
1885  * @cdev:       t3cdev adapter
1886  */
1887 void cxgb3i_sdev_remove(struct t3cdev *cdev)
1888 {
1889         struct cxgb3i_sdev_data *cdata = CXGB3_SDEV_DATA(cdev);
1890
1891         cxgb3i_log_info("t3dev 0x%p, offload down, remove.\n", cdev);
1892
1893         write_lock(&cdata_rwlock);
1894         list_del(&cdata->list);
1895         write_unlock(&cdata_rwlock);
1896
1897         sdev_data_cleanup(cdata);
1898 }