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/slab.h>
  17 #include <linux/version.h>
  18
  19 #include "cxgb3_defs.h"
  20 #include "cxgb3_ctl_defs.h"
  21 #include "firmware_exports.h"
  22 #include "cxgb3i_offload.h"
  23 #include "cxgb3i_pdu.h"
  24 #include "cxgb3i_ddp.h"
  25
  26 #ifdef __DEBUG_C3CN_CONN__
  27 #define c3cn_conn_debug         cxgb3i_log_debug
  28 #else
  29 #define c3cn_conn_debug(fmt...)
  30 #endif
  31
  32 #ifdef __DEBUG_C3CN_TX__
  33 #define c3cn_tx_debug           cxgb3i_log_debug
  34 #else
  35 #define c3cn_tx_debug(fmt...)
  36 #endif
  37
  38 #ifdef __DEBUG_C3CN_RX__
  39 #define c3cn_rx_debug           cxgb3i_log_debug
  40 #else
  41 #define c3cn_rx_debug(fmt...)
  42 #endif
  43
  44 /*
  45  * module parameters releated to offloaded iscsi connection
  46  */
  47 static int cxgb3_rcv_win = 256 * 1024;
  48 module_param(cxgb3_rcv_win, int, 0644);
  49 MODULE_PARM_DESC(cxgb3_rcv_win, "TCP receive window in bytes (default=256KB)");
  50
  51 static int cxgb3_snd_win = 128 * 1024;
  52 module_param(cxgb3_snd_win, int, 0644);
  53 MODULE_PARM_DESC(cxgb3_snd_win, "TCP send window in bytes (default=128KB)");
  54
  55 static int cxgb3_rx_credit_thres = 10 * 1024;
  56 module_param(cxgb3_rx_credit_thres, int, 0644);
  57 MODULE_PARM_DESC(rx_credit_thres,
  58                  "RX credits return threshold in bytes (default=10KB)");
  59
  60 static unsigned int cxgb3_max_connect = 8 * 1024;
  61 module_param(cxgb3_max_connect, uint, 0644);
  62 MODULE_PARM_DESC(cxgb3_max_connect, "Max. # of connections (default=8092)");
  63
  64 static unsigned int cxgb3_sport_base = 20000;
  65 module_param(cxgb3_sport_base, uint, 0644);
  66 MODULE_PARM_DESC(cxgb3_sport_base, "starting port number (default=20000)");
  67
  68 /*
  69  * cxgb3i tcp connection data(per adapter) list
  70  */
  71 static LIST_HEAD(cdata_list);
  72 static DEFINE_RWLOCK(cdata_rwlock);
  73
  74 static int c3cn_push_tx_frames(struct s3_conn *c3cn, int req_completion);
  75 static void c3cn_release_offload_resources(struct s3_conn *c3cn);
  76
  77 /*
  78  * iscsi source port management
  79  *
  80  * Find a free source port in the port allocation map. We use a very simple
  81  * rotor scheme to look for the next free port.
  82  *
  83  * If a source port has been specified make sure that it doesn't collide with
  84  * our normal source port allocation map.  If it's outside the range of our
  85  * allocation/deallocation scheme just let them use it.
  86  *
  87  * If the source port is outside our allocation range, the caller is
  88  * responsible for keeping track of their port usage.
  89  */
  90 static int c3cn_get_port(struct s3_conn *c3cn, struct cxgb3i_sdev_data *cdata)
  91 {
  92         unsigned int start;
  93         int idx;
  94
  95         if (!cdata)
  96                 goto error_out;
  97
  98         if (c3cn->saddr.sin_port) {
  99                 cxgb3i_log_error("connect, sin_port NON-ZERO %u.\n",
 100                                  c3cn->saddr.sin_port);
 101                 return -EADDRINUSE;
 102         }
 103
 104         spin_lock_bh(&cdata->lock);
 105         start = idx = cdata->sport_next;
 106         do {
 107                 if (++idx >= cxgb3_max_connect)
 108                         idx = 0;
 109                 if (!cdata->sport_conn[idx]) {
 110                         c3cn->saddr.sin_port = htons(cxgb3_sport_base + idx);
 111                         cdata->sport_next = idx;
 112                         cdata->sport_conn[idx] = c3cn;
 113                         spin_unlock_bh(&cdata->lock);
 114
 115                         c3cn_conn_debug("%s reserve port %u.\n",
 116                                         cdata->cdev->name,
 117                                         cxgb3_sport_base + idx);
 118                         return 0;
 119                 }
 120         } while (idx != start);
 121         spin_unlock_bh(&cdata->lock);
 122
 123 error_out:
 124         return -EADDRNOTAVAIL;
 125 }
 126
 127 static void c3cn_put_port(struct s3_conn *c3cn)
 128 {
 129         if (!c3cn->cdev)
 130                 return;
 131
 132         if (c3cn->saddr.sin_port) {
 133                 struct cxgb3i_sdev_data *cdata = CXGB3_SDEV_DATA(c3cn->cdev);
 134                 int idx = ntohs(c3cn->saddr.sin_port) - cxgb3_sport_base;
 135
 136                 c3cn->saddr.sin_port = 0;
 137                 if (idx < 0 || idx >= cxgb3_max_connect)
 138                         return;
 139                 spin_lock_bh(&cdata->lock);
 140                 cdata->sport_conn[idx] = NULL;
 141                 spin_unlock_bh(&cdata->lock);
 142                 c3cn_conn_debug("%s, release port %u.\n",
 143                                 cdata->cdev->name, cxgb3_sport_base + idx);
 144         }
 145 }
 146
 147 static inline void c3cn_set_flag(struct s3_conn *c3cn, enum c3cn_flags flag)
 148 {
 149         __set_bit(flag, &c3cn->flags);
 150         c3cn_conn_debug("c3cn 0x%p, set %d, s %u, f 0x%lx.\n",
 151                         c3cn, flag, c3cn->state, c3cn->flags);
 152 }
 153
 154 static inline void c3cn_clear_flag(struct s3_conn *c3cn, enum c3cn_flags flag)
 155 {
 156         __clear_bit(flag, &c3cn->flags);
 157         c3cn_conn_debug("c3cn 0x%p, clear %d, s %u, f 0x%lx.\n",
 158                         c3cn, flag, c3cn->state, c3cn->flags);
 159 }
 160
 161 static inline int c3cn_flag(struct s3_conn *c3cn, enum c3cn_flags flag)
 162 {
 163         if (c3cn == NULL)
 164                 return 0;
 165         return test_bit(flag, &c3cn->flags);
 166 }
 167
 168 static void c3cn_set_state(struct s3_conn *c3cn, int state)
 169 {
 170         c3cn_conn_debug("c3cn 0x%p state -> %u.\n", c3cn, state);
 171         c3cn->state = state;
 172 }
 173
 174 static inline void c3cn_hold(struct s3_conn *c3cn)
 175 {
 176         atomic_inc(&c3cn->refcnt);
 177 }
 178
 179 static inline void c3cn_put(struct s3_conn *c3cn)
 180 {
 181         if (atomic_dec_and_test(&c3cn->refcnt)) {
 182                 c3cn_conn_debug("free c3cn 0x%p, s %u, f 0x%lx.\n",
 183                                 c3cn, c3cn->state, c3cn->flags);
 184                 kfree(c3cn);
 185         }
 186 }
 187
 188 static void c3cn_closed(struct s3_conn *c3cn)
 189 {
 190         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
 191                          c3cn, c3cn->state, c3cn->flags);
 192
 193         c3cn_put_port(c3cn);
 194         c3cn_release_offload_resources(c3cn);
 195         c3cn_set_state(c3cn, C3CN_STATE_CLOSED);
 196         cxgb3i_conn_closing(c3cn);
 197 }
 198
 199 /*
 200  * CPL (Chelsio Protocol Language) defines a message passing interface between
 201  * the host driver and T3 asic.
 202  * The section below implments CPLs that related to iscsi tcp connection
 203  * open/close/abort and data send/receive.
 204  */
 205
 206 /*
 207  * CPL connection active open request: host ->
 208  */
 209 static unsigned int find_best_mtu(const struct t3c_data *d, unsigned short mtu)
 210 {
 211         int i = 0;
 212
 213         while (i < d->nmtus - 1 && d->mtus[i + 1] <= mtu)
 214                 ++i;
 215         return i;
 216 }
 217
 218 static unsigned int select_mss(struct s3_conn *c3cn, unsigned int pmtu)
 219 {
 220         unsigned int idx;
 221         struct dst_entry *dst = c3cn->dst_cache;
 222         struct t3cdev *cdev = c3cn->cdev;
 223         const struct t3c_data *td = T3C_DATA(cdev);
 224         u16 advmss = dst_metric(dst, RTAX_ADVMSS);
 225
 226         if (advmss > pmtu - 40)
 227                 advmss = pmtu - 40;
 228         if (advmss < td->mtus[0] - 40)
 229                 advmss = td->mtus[0] - 40;
 230         idx = find_best_mtu(td, advmss + 40);
 231         return idx;
 232 }
 233
 234 static inline int compute_wscale(int win)
 235 {
 236         int wscale = 0;
 237         while (wscale < 14 && (65535<<wscale) < win)
 238                 wscale++;
 239         return wscale;
 240 }
 241
 242 static inline unsigned int calc_opt0h(struct s3_conn *c3cn)
 243 {
 244         int wscale = compute_wscale(cxgb3_rcv_win);
 245         return  V_KEEP_ALIVE(1) |
 246                 F_TCAM_BYPASS |
 247                 V_WND_SCALE(wscale) |
 248                 V_MSS_IDX(c3cn->mss_idx);
 249 }
 250
 251 static inline unsigned int calc_opt0l(struct s3_conn *c3cn)
 252 {
 253         return  V_ULP_MODE(ULP_MODE_ISCSI) |
 254                 V_RCV_BUFSIZ(cxgb3_rcv_win>>10);
 255 }
 256
 257 static void make_act_open_req(struct s3_conn *c3cn, struct sk_buff *skb,
 258                               unsigned int atid, const struct l2t_entry *e)
 259 {
 260         struct cpl_act_open_req *req;
 261
 262         c3cn_conn_debug("c3cn 0x%p, atid 0x%x.\n", c3cn, atid);
 263
 264         skb->priority = CPL_PRIORITY_SETUP;
 265         req = (struct cpl_act_open_req *)__skb_put(skb, sizeof(*req));
 266         req->wr.wr_hi = htonl(V_WR_OP(FW_WROPCODE_FORWARD));
 267         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_ACT_OPEN_REQ, atid));
 268         req->local_port = c3cn->saddr.sin_port;
 269         req->peer_port = c3cn->daddr.sin_port;
 270         req->local_ip = c3cn->saddr.sin_addr.s_addr;
 271         req->peer_ip = c3cn->daddr.sin_addr.s_addr;
 272         req->opt0h = htonl(calc_opt0h(c3cn) | V_L2T_IDX(e->idx) |
 273                            V_TX_CHANNEL(e->smt_idx));
 274         req->opt0l = htonl(calc_opt0l(c3cn));
 275         req->params = 0;
 276 }
 277
 278 static void fail_act_open(struct s3_conn *c3cn, int errno)
 279 {
 280         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
 281                         c3cn, c3cn->state, c3cn->flags);
 282         c3cn->err = errno;
 283         c3cn_closed(c3cn);
 284 }
 285
 286 static void act_open_req_arp_failure(struct t3cdev *dev, struct sk_buff *skb)
 287 {
 288         struct s3_conn *c3cn = (struct s3_conn *)skb->sk;
 289
 290         c3cn_conn_debug("c3cn 0x%p, state %u.\n", c3cn, c3cn->state);
 291
 292         c3cn_hold(c3cn);
 293         spin_lock_bh(&c3cn->lock);
 294         if (c3cn->state == C3CN_STATE_CONNECTING)
 295                 fail_act_open(c3cn, -EHOSTUNREACH);
 296         spin_unlock_bh(&c3cn->lock);
 297         c3cn_put(c3cn);
 298         __kfree_skb(skb);
 299 }
 300
 301 /*
 302  * CPL connection close request: host ->
 303  *
 304  * Close a connection by sending a CPL_CLOSE_CON_REQ message and queue it to
 305  * the write queue (i.e., after any unsent txt data).
 306  */
 307 static void skb_entail(struct s3_conn *c3cn, struct sk_buff *skb,
 308                        int flags)
 309 {
 310         skb_tcp_seq(skb) = c3cn->write_seq;
 311         skb_flags(skb) = flags;
 312         __skb_queue_tail(&c3cn->write_queue, skb);
 313 }
 314
 315 static void send_close_req(struct s3_conn *c3cn)
 316 {
 317         struct sk_buff *skb = c3cn->cpl_close;
 318         struct cpl_close_con_req *req = (struct cpl_close_con_req *)skb->head;
 319         unsigned int tid = c3cn->tid;
 320
 321         c3cn_conn_debug("c3cn 0x%p, state 0x%x, flag 0x%lx.\n",
 322                         c3cn, c3cn->state, c3cn->flags);
 323
 324         c3cn->cpl_close = NULL;
 325
 326         req->wr.wr_hi = htonl(V_WR_OP(FW_WROPCODE_OFLD_CLOSE_CON));
 327         req->wr.wr_lo = htonl(V_WR_TID(tid));
 328         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid));
 329         req->rsvd = htonl(c3cn->write_seq);
 330
 331         skb_entail(c3cn, skb, C3CB_FLAG_NO_APPEND);
 332         if (c3cn->state != C3CN_STATE_CONNECTING)
 333                 c3cn_push_tx_frames(c3cn, 1);
 334 }
 335
 336 /*
 337  * CPL connection abort request: host ->
 338  *
 339  * Send an ABORT_REQ message. Makes sure we do not send multiple ABORT_REQs
 340  * for the same connection and also that we do not try to send a message
 341  * after the connection has closed.
 342  */
 343 static void abort_arp_failure(struct t3cdev *cdev, struct sk_buff *skb)
 344 {
 345         struct cpl_abort_req *req = cplhdr(skb);
 346
 347         c3cn_conn_debug("tdev 0x%p.\n", cdev);
 348
 349         req->cmd = CPL_ABORT_NO_RST;
 350         cxgb3_ofld_send(cdev, skb);
 351 }
 352
 353 static inline void c3cn_purge_write_queue(struct s3_conn *c3cn)
 354 {
 355         struct sk_buff *skb;
 356
 357         while ((skb = __skb_dequeue(&c3cn->write_queue)))
 358                 __kfree_skb(skb);
 359 }
 360
 361 static void send_abort_req(struct s3_conn *c3cn)
 362 {
 363         struct sk_buff *skb = c3cn->cpl_abort_req;
 364         struct cpl_abort_req *req;
 365         unsigned int tid = c3cn->tid;
 366
 367         if (unlikely(c3cn->state == C3CN_STATE_ABORTING) || !skb ||
 368                      !c3cn->cdev)
 369                 return;
 370
 371         c3cn_set_state(c3cn, C3CN_STATE_ABORTING);
 372
 373         c3cn_conn_debug("c3cn 0x%p, flag ABORT_RPL + ABORT_SHUT.\n", c3cn);
 374
 375         c3cn_set_flag(c3cn, C3CN_ABORT_RPL_PENDING);
 376
 377         /* Purge the send queue so we don't send anything after an abort. */
 378         c3cn_purge_write_queue(c3cn);
 379
 380         c3cn->cpl_abort_req = NULL;
 381         req = (struct cpl_abort_req *)skb->head;
 382
 383         skb->priority = CPL_PRIORITY_DATA;
 384         set_arp_failure_handler(skb, abort_arp_failure);
 385
 386         req->wr.wr_hi = htonl(V_WR_OP(FW_WROPCODE_OFLD_HOST_ABORT_CON_REQ));
 387         req->wr.wr_lo = htonl(V_WR_TID(tid));
 388         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_ABORT_REQ, tid));
 389         req->rsvd0 = htonl(c3cn->snd_nxt);
 390         req->rsvd1 = !c3cn_flag(c3cn, C3CN_TX_DATA_SENT);
 391         req->cmd = CPL_ABORT_SEND_RST;
 392
 393         l2t_send(c3cn->cdev, skb, c3cn->l2t);
 394 }
 395
 396 /*
 397  * CPL connection abort reply: host ->
 398  *
 399  * Send an ABORT_RPL message in response of the ABORT_REQ received.
 400  */
 401 static void send_abort_rpl(struct s3_conn *c3cn, int rst_status)
 402 {
 403         struct sk_buff *skb = c3cn->cpl_abort_rpl;
 404         struct cpl_abort_rpl *rpl = (struct cpl_abort_rpl *)skb->head;
 405
 406         c3cn->cpl_abort_rpl = NULL;
 407
 408         skb->priority = CPL_PRIORITY_DATA;
 409         rpl->wr.wr_hi = htonl(V_WR_OP(FW_WROPCODE_OFLD_HOST_ABORT_CON_RPL));
 410         rpl->wr.wr_lo = htonl(V_WR_TID(c3cn->tid));
 411         OPCODE_TID(rpl) = htonl(MK_OPCODE_TID(CPL_ABORT_RPL, c3cn->tid));
 412         rpl->cmd = rst_status;
 413
 414         cxgb3_ofld_send(c3cn->cdev, skb);
 415 }
 416
 417 /*
 418  * CPL connection rx data ack: host ->
 419  * Send RX credits through an RX_DATA_ACK CPL message. Returns the number of
 420  * credits sent.
 421  */
 422 static u32 send_rx_credits(struct s3_conn *c3cn, u32 credits, u32 dack)
 423 {
 424         struct sk_buff *skb;
 425         struct cpl_rx_data_ack *req;
 426
 427         skb = alloc_skb(sizeof(*req), GFP_ATOMIC);
 428         if (!skb)
 429                 return 0;
 430
 431         req = (struct cpl_rx_data_ack *)__skb_put(skb, sizeof(*req));
 432         req->wr.wr_hi = htonl(V_WR_OP(FW_WROPCODE_FORWARD));
 433         OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_RX_DATA_ACK, c3cn->tid));
 434         req->credit_dack = htonl(dack | V_RX_CREDITS(credits));
 435         skb->priority = CPL_PRIORITY_ACK;
 436         cxgb3_ofld_send(c3cn->cdev, skb);
 437         return credits;
 438 }
 439
 440 /*
 441  * CPL connection tx data: host ->
 442  *
 443  * Send iscsi PDU via TX_DATA CPL message. Returns the number of
 444  * credits sent.
 445  * Each TX_DATA consumes work request credit (wrs), so we need to keep track of
 446  * how many we've used so far and how many are pending (i.e., yet ack'ed by T3).
 447  */
 448
 449 /*
 450  * For ULP connections HW may inserts digest bytes into the pdu. Those digest
 451  * bytes are not sent by the host but are part of the TCP payload and therefore
 452  * consume TCP sequence space.
 453  */
 454 static const unsigned int cxgb3_ulp_extra_len[] = { 0, 4, 4, 8 };
 455 static inline unsigned int ulp_extra_len(const struct sk_buff *skb)
 456 {
 457         return cxgb3_ulp_extra_len[skb_ulp_mode(skb) & 3];
 458 }
 459
 460 static unsigned int wrlen __read_mostly;
 461
 462 /*
 463  * The number of WRs needed for an skb depends on the number of fragments
 464  * in the skb and whether it has any payload in its main body.  This maps the
 465  * length of the gather list represented by an skb into the # of necessary WRs.
 466  * The extra two fragments are for iscsi bhs and payload padding.
 467  */
 468 #define SKB_WR_LIST_SIZE        (MAX_SKB_FRAGS + 2)
 469 static unsigned int skb_wrs[SKB_WR_LIST_SIZE] __read_mostly;
 470
 471 static void s3_init_wr_tab(unsigned int wr_len)
 472 {
 473         int i;
 474
 475         if (skb_wrs[1])         /* already initialized */
 476                 return;
 477
 478         for (i = 1; i < SKB_WR_LIST_SIZE; i++) {
 479                 int sgl_len = (3 * i) / 2 + (i & 1);
 480
 481                 sgl_len += 3;
 482                 skb_wrs[i] = (sgl_len <= wr_len
 483                               ? 1 : 1 + (sgl_len - 2) / (wr_len - 1));
 484         }
 485
 486         wrlen = wr_len * 8;
 487 }
 488
 489 static inline void reset_wr_list(struct s3_conn *c3cn)
 490 {
 491         c3cn->wr_pending_head = c3cn->wr_pending_tail = NULL;
 492 }
 493
 494 /*
 495  * Add a WR to a connections's list of pending WRs.  This is a singly-linked
 496  * list of sk_buffs operating as a FIFO.  The head is kept in wr_pending_head
 497  * and the tail in wr_pending_tail.
 498  */
 499 static inline void enqueue_wr(struct s3_conn *c3cn,
 500                               struct sk_buff *skb)
 501 {
 502         skb_tx_wr_next(skb) = NULL;
 503
 504         /*
 505          * We want to take an extra reference since both us and the driver
 506          * need to free the packet before it's really freed. We know there's
 507          * just one user currently so we use atomic_set rather than skb_get
 508          * to avoid the atomic op.
 509          */
 510         atomic_set(&skb->users, 2);
 511
 512         if (!c3cn->wr_pending_head)
 513                 c3cn->wr_pending_head = skb;
 514         else
 515                 skb_tx_wr_next(c3cn->wr_pending_tail) = skb;
 516         c3cn->wr_pending_tail = skb;
 517 }
 518
 519 static int count_pending_wrs(struct s3_conn *c3cn)
 520 {
 521         int n = 0;
 522         const struct sk_buff *skb = c3cn->wr_pending_head;
 523
 524         while (skb) {
 525                 n += skb->csum;
 526                 skb = skb_tx_wr_next(skb);
 527         }
 528         return n;
 529 }
 530
 531 static inline struct sk_buff *peek_wr(const struct s3_conn *c3cn)
 532 {
 533         return c3cn->wr_pending_head;
 534 }
 535
 536 static inline void free_wr_skb(struct sk_buff *skb)
 537 {
 538         kfree_skb(skb);
 539 }
 540
 541 static inline struct sk_buff *dequeue_wr(struct s3_conn *c3cn)
 542 {
 543         struct sk_buff *skb = c3cn->wr_pending_head;
 544
 545         if (likely(skb)) {
 546                 /* Don't bother clearing the tail */
 547                 c3cn->wr_pending_head = skb_tx_wr_next(skb);
 548                 skb_tx_wr_next(skb) = NULL;
 549         }
 550         return skb;
 551 }
 552
 553 static void purge_wr_queue(struct s3_conn *c3cn)
 554 {
 555         struct sk_buff *skb;
 556         while ((skb = dequeue_wr(c3cn)) != NULL)
 557                 free_wr_skb(skb);
 558 }
 559
 560 static inline void make_tx_data_wr(struct s3_conn *c3cn, struct sk_buff *skb,
 561                                    int len, int req_completion)
 562 {
 563         struct tx_data_wr *req;
 564
 565         skb_reset_transport_header(skb);
 566         req = (struct tx_data_wr *)__skb_push(skb, sizeof(*req));
 567         req->wr_hi = htonl(V_WR_OP(FW_WROPCODE_OFLD_TX_DATA) |
 568                         (req_completion ? F_WR_COMPL : 0));
 569         req->wr_lo = htonl(V_WR_TID(c3cn->tid));
 570         req->sndseq = htonl(c3cn->snd_nxt);
 571         /* len includes the length of any HW ULP additions */
 572         req->len = htonl(len);
 573         req->param = htonl(V_TX_PORT(c3cn->l2t->smt_idx));
 574         /* V_TX_ULP_SUBMODE sets both the mode and submode */
 575         req->flags = htonl(V_TX_ULP_SUBMODE(skb_ulp_mode(skb)) |
 576                            V_TX_SHOVE((skb_peek(&c3cn->write_queue) ? 0 : 1)));
 577
 578         if (!c3cn_flag(c3cn, C3CN_TX_DATA_SENT)) {
 579                 req->flags |= htonl(V_TX_ACK_PAGES(2) | F_TX_INIT |
 580                                     V_TX_CPU_IDX(c3cn->qset));
 581                 /* Sendbuffer is in units of 32KB. */
 582                 req->param |= htonl(V_TX_SNDBUF(cxgb3_snd_win >> 15));
 583                 c3cn_set_flag(c3cn, C3CN_TX_DATA_SENT);
 584         }
 585 }
 586
 587 /**
 588  * c3cn_push_tx_frames -- start transmit
 589  * @c3cn: the offloaded connection
 590  * @req_completion: request wr_ack or not
 591  *
 592  * Prepends TX_DATA_WR or CPL_CLOSE_CON_REQ headers to buffers waiting in a
 593  * connection's send queue and sends them on to T3.  Must be called with the
 594  * connection's lock held.  Returns the amount of send buffer space that was
 595  * freed as a result of sending queued data to T3.
 596  */
 597 static void arp_failure_discard(struct t3cdev *cdev, struct sk_buff *skb)
 598 {
 599         kfree_skb(skb);
 600 }
 601
 602 static int c3cn_push_tx_frames(struct s3_conn *c3cn, int req_completion)
 603 {
 604         int total_size = 0;
 605         struct sk_buff *skb;
 606         struct t3cdev *cdev;
 607         struct cxgb3i_sdev_data *cdata;
 608
 609         if (unlikely(c3cn->state == C3CN_STATE_CONNECTING ||
 610                      c3cn->state == C3CN_STATE_CLOSE_WAIT_1 ||
 611                      c3cn->state >= C3CN_STATE_ABORTING)) {
 612                 c3cn_tx_debug("c3cn 0x%p, in closing state %u.\n",
 613                               c3cn, c3cn->state);
 614                 return 0;
 615         }
 616
 617         cdev = c3cn->cdev;
 618         cdata = CXGB3_SDEV_DATA(cdev);
 619
 620         while (c3cn->wr_avail
 621                && (skb = skb_peek(&c3cn->write_queue)) != NULL) {
 622                 int len = skb->len;     /* length before skb_push */
 623                 int frags = skb_shinfo(skb)->nr_frags + (len != skb->data_len);
 624                 int wrs_needed = skb_wrs[frags];
 625
 626                 if (wrs_needed > 1 && len + sizeof(struct tx_data_wr) <= wrlen)
 627                         wrs_needed = 1;
 628
 629                 WARN_ON(frags >= SKB_WR_LIST_SIZE || wrs_needed < 1);
 630
 631                 if (c3cn->wr_avail < wrs_needed) {
 632                         c3cn_tx_debug("c3cn 0x%p, skb len %u/%u, frag %u, "
 633                                       "wr %d < %u.\n",
 634                                       c3cn, skb->len, skb->data_len, frags,
 635                                       wrs_needed, c3cn->wr_avail);
 636                         break;
 637                 }
 638
 639                 __skb_unlink(skb, &c3cn->write_queue);
 640                 skb->priority = CPL_PRIORITY_DATA;
 641                 skb->csum = wrs_needed; /* remember this until the WR_ACK */
 642                 c3cn->wr_avail -= wrs_needed;
 643                 c3cn->wr_unacked += wrs_needed;
 644                 enqueue_wr(c3cn, skb);
 645
 646                 c3cn_tx_debug("c3cn 0x%p, enqueue, skb len %u/%u, frag %u, "
 647                                 "wr %d, left %u, unack %u.\n",
 648                                 c3cn, skb->len, skb->data_len, frags,
 649                                 wrs_needed, c3cn->wr_avail, c3cn->wr_unacked);
 650
 651
 652                 if (likely(skb_flags(skb) & C3CB_FLAG_NEED_HDR)) {
 653                         if ((req_completion &&
 654                                 c3cn->wr_unacked == wrs_needed) ||
 655                             (skb_flags(skb) & C3CB_FLAG_COMPL) ||
 656                             c3cn->wr_unacked >= c3cn->wr_max / 2) {
 657                                 req_completion = 1;
 658                                 c3cn->wr_unacked = 0;
 659                         }
 660                         len += ulp_extra_len(skb);
 661                         make_tx_data_wr(c3cn, skb, len, req_completion);
 662                         c3cn->snd_nxt += len;
 663                         skb_flags(skb) &= ~C3CB_FLAG_NEED_HDR;
 664                 }
 665
 666                 total_size += skb->truesize;
 667                 set_arp_failure_handler(skb, arp_failure_discard);
 668                 l2t_send(cdev, skb, c3cn->l2t);
 669         }
 670         return total_size;
 671 }
 672
 673 /*
 674  * process_cpl_msg: -> host
 675  * Top-level CPL message processing used by most CPL messages that
 676  * pertain to connections.
 677  */
 678 static inline void process_cpl_msg(void (*fn)(struct s3_conn *,
 679                                               struct sk_buff *),
 680                                    struct s3_conn *c3cn,
 681                                    struct sk_buff *skb)
 682 {
 683         spin_lock_bh(&c3cn->lock);
 684         fn(c3cn, skb);
 685         spin_unlock_bh(&c3cn->lock);
 686 }
 687
 688 /*
 689  * process_cpl_msg_ref: -> host
 690  * Similar to process_cpl_msg() but takes an extra connection reference around
 691  * the call to the handler.  Should be used if the handler may drop a
 692  * connection reference.
 693  */
 694 static inline void process_cpl_msg_ref(void (*fn) (struct s3_conn *,
 695                                                    struct sk_buff *),
 696                                        struct s3_conn *c3cn,
 697                                        struct sk_buff *skb)
 698 {
 699         c3cn_hold(c3cn);
 700         process_cpl_msg(fn, c3cn, skb);
 701         c3cn_put(c3cn);
 702 }
 703
 704 /*
 705  * Process a CPL_ACT_ESTABLISH message: -> host
 706  * Updates connection state from an active establish CPL message.  Runs with
 707  * the connection lock held.
 708  */
 709
 710 static inline void s3_free_atid(struct t3cdev *cdev, unsigned int tid)
 711 {
 712         struct s3_conn *c3cn = cxgb3_free_atid(cdev, tid);
 713         if (c3cn)
 714                 c3cn_put(c3cn);
 715 }
 716
 717 static void c3cn_established(struct s3_conn *c3cn, u32 snd_isn,
 718                              unsigned int opt)
 719 {
 720         c3cn_conn_debug("c3cn 0x%p, state %u.\n", c3cn, c3cn->state);
 721
 722         c3cn->write_seq = c3cn->snd_nxt = c3cn->snd_una = snd_isn;
 723
 724         /*
 725          * Causes the first RX_DATA_ACK to supply any Rx credits we couldn't
 726          * pass through opt0.
 727          */
 728         if (cxgb3_rcv_win > (M_RCV_BUFSIZ << 10))
 729                 c3cn->rcv_wup -= cxgb3_rcv_win - (M_RCV_BUFSIZ << 10);
 730
 731         dst_confirm(c3cn->dst_cache);
 732
 733         smp_mb();
 734
 735         c3cn_set_state(c3cn, C3CN_STATE_ESTABLISHED);
 736 }
 737
 738 static void process_act_establish(struct s3_conn *c3cn, struct sk_buff *skb)
 739 {
 740         struct cpl_act_establish *req = cplhdr(skb);
 741         u32 rcv_isn = ntohl(req->rcv_isn);      /* real RCV_ISN + 1 */
 742
 743         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
 744                         c3cn, c3cn->state, c3cn->flags);
 745
 746         if (unlikely(c3cn->state != C3CN_STATE_CONNECTING))
 747                 cxgb3i_log_error("TID %u expected SYN_SENT, got EST., s %u\n",
 748                                  c3cn->tid, c3cn->state);
 749
 750         c3cn->copied_seq = c3cn->rcv_wup = c3cn->rcv_nxt = rcv_isn;
 751         c3cn_established(c3cn, ntohl(req->snd_isn), ntohs(req->tcp_opt));
 752
 753         __kfree_skb(skb);
 754
 755         if (unlikely(c3cn_flag(c3cn, C3CN_ACTIVE_CLOSE_NEEDED)))
 756                 /* upper layer has requested closing */
 757                 send_abort_req(c3cn);
 758         else {
 759                 if (skb_queue_len(&c3cn->write_queue))
 760                         c3cn_push_tx_frames(c3cn, 1);
 761                 cxgb3i_conn_tx_open(c3cn);
 762         }
 763 }
 764
 765 static int do_act_establish(struct t3cdev *cdev, struct sk_buff *skb,
 766                             void *ctx)
 767 {
 768         struct cpl_act_establish *req = cplhdr(skb);
 769         unsigned int tid = GET_TID(req);
 770         unsigned int atid = G_PASS_OPEN_TID(ntohl(req->tos_tid));
 771         struct s3_conn *c3cn = ctx;
 772         struct cxgb3i_sdev_data *cdata = CXGB3_SDEV_DATA(cdev);
 773
 774         c3cn_conn_debug("rcv, tid 0x%x, c3cn 0x%p, s %u, f 0x%lx.\n",
 775                         tid, c3cn, c3cn->state, c3cn->flags);
 776
 777         c3cn->tid = tid;
 778         c3cn_hold(c3cn);
 779         cxgb3_insert_tid(cdata->cdev, cdata->client, c3cn, tid);
 780         s3_free_atid(cdev, atid);
 781
 782         c3cn->qset = G_QNUM(ntohl(skb->csum));
 783
 784         process_cpl_msg(process_act_establish, c3cn, skb);
 785         return 0;
 786 }
 787
 788 /*
 789  * Process a CPL_ACT_OPEN_RPL message: -> host
 790  * Handle active open failures.
 791  */
 792 static int act_open_rpl_status_to_errno(int status)
 793 {
 794         switch (status) {
 795         case CPL_ERR_CONN_RESET:
 796                 return -ECONNREFUSED;
 797         case CPL_ERR_ARP_MISS:
 798                 return -EHOSTUNREACH;
 799         case CPL_ERR_CONN_TIMEDOUT:
 800                 return -ETIMEDOUT;
 801         case CPL_ERR_TCAM_FULL:
 802                 return -ENOMEM;
 803         case CPL_ERR_CONN_EXIST:
 804                 cxgb3i_log_error("ACTIVE_OPEN_RPL: 4-tuple in use\n");
 805                 return -EADDRINUSE;
 806         default:
 807                 return -EIO;
 808         }
 809 }
 810
 811 static void act_open_retry_timer(unsigned long data)
 812 {
 813         struct sk_buff *skb;
 814         struct s3_conn *c3cn = (struct s3_conn *)data;
 815
 816         c3cn_conn_debug("c3cn 0x%p, state %u.\n", c3cn, c3cn->state);
 817
 818         spin_lock_bh(&c3cn->lock);
 819         skb = alloc_skb(sizeof(struct cpl_act_open_req), GFP_ATOMIC);
 820         if (!skb)
 821                 fail_act_open(c3cn, -ENOMEM);
 822         else {
 823                 skb->sk = (struct sock *)c3cn;
 824                 set_arp_failure_handler(skb, act_open_req_arp_failure);
 825                 make_act_open_req(c3cn, skb, c3cn->tid, c3cn->l2t);
 826                 l2t_send(c3cn->cdev, skb, c3cn->l2t);
 827         }
 828         spin_unlock_bh(&c3cn->lock);
 829         c3cn_put(c3cn);
 830 }
 831
 832 static void process_act_open_rpl(struct s3_conn *c3cn, struct sk_buff *skb)
 833 {
 834         struct cpl_act_open_rpl *rpl = cplhdr(skb);
 835
 836         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
 837                         c3cn, c3cn->state, c3cn->flags);
 838
 839         if (rpl->status == CPL_ERR_CONN_EXIST &&
 840             c3cn->retry_timer.function != act_open_retry_timer) {
 841                 c3cn->retry_timer.function = act_open_retry_timer;
 842                 if (!mod_timer(&c3cn->retry_timer, jiffies + HZ / 2))
 843                         c3cn_hold(c3cn);
 844         } else
 845                 fail_act_open(c3cn, act_open_rpl_status_to_errno(rpl->status));
 846         __kfree_skb(skb);
 847 }
 848
 849 static int do_act_open_rpl(struct t3cdev *cdev, struct sk_buff *skb, void *ctx)
 850 {
 851         struct s3_conn *c3cn = ctx;
 852         struct cpl_act_open_rpl *rpl = cplhdr(skb);
 853
 854         c3cn_conn_debug("rcv, status 0x%x, c3cn 0x%p, s %u, f 0x%lx.\n",
 855                         rpl->status, c3cn, c3cn->state, c3cn->flags);
 856
 857         if (rpl->status != CPL_ERR_TCAM_FULL &&
 858             rpl->status != CPL_ERR_CONN_EXIST &&
 859             rpl->status != CPL_ERR_ARP_MISS)
 860                 cxgb3_queue_tid_release(cdev, GET_TID(rpl));
 861
 862         process_cpl_msg_ref(process_act_open_rpl, c3cn, skb);
 863         return 0;
 864 }
 865
 866 /*
 867  * Process PEER_CLOSE CPL messages: -> host
 868  * Handle peer FIN.
 869  */
 870 static void process_peer_close(struct s3_conn *c3cn, struct sk_buff *skb)
 871 {
 872         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
 873                         c3cn, c3cn->state, c3cn->flags);
 874
 875         if (c3cn_flag(c3cn, C3CN_ABORT_RPL_PENDING))
 876                 goto out;
 877
 878         switch (c3cn->state) {
 879         case C3CN_STATE_ESTABLISHED:
 880                 c3cn_set_state(c3cn, C3CN_STATE_PASSIVE_CLOSE);
 881                 break;
 882         case C3CN_STATE_ACTIVE_CLOSE:
 883                 c3cn_set_state(c3cn, C3CN_STATE_CLOSE_WAIT_2);
 884                 break;
 885         case C3CN_STATE_CLOSE_WAIT_1:
 886                 c3cn_closed(c3cn);
 887                 break;
 888         case C3CN_STATE_ABORTING:
 889                 break;
 890         default:
 891                 cxgb3i_log_error("%s: peer close, TID %u in bad state %u\n",
 892                                  c3cn->cdev->name, c3cn->tid, c3cn->state);
 893         }
 894
 895         cxgb3i_conn_closing(c3cn);
 896 out:
 897         __kfree_skb(skb);
 898 }
 899
 900 static int do_peer_close(struct t3cdev *cdev, struct sk_buff *skb, void *ctx)
 901 {
 902         struct s3_conn *c3cn = ctx;
 903
 904         c3cn_conn_debug("rcv, c3cn 0x%p, s %u, f 0x%lx.\n",
 905                         c3cn, c3cn->state, c3cn->flags);
 906         process_cpl_msg_ref(process_peer_close, c3cn, skb);
 907         return 0;
 908 }
 909
 910 /*
 911  * Process CLOSE_CONN_RPL CPL message: -> host
 912  * Process a peer ACK to our FIN.
 913  */
 914 static void process_close_con_rpl(struct s3_conn *c3cn, struct sk_buff *skb)
 915 {
 916         struct cpl_close_con_rpl *rpl = cplhdr(skb);
 917
 918         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
 919                         c3cn, c3cn->state, c3cn->flags);
 920
 921         c3cn->snd_una = ntohl(rpl->snd_nxt) - 1;        /* exclude FIN */
 922
 923         if (c3cn_flag(c3cn, C3CN_ABORT_RPL_PENDING))
 924                 goto out;
 925
 926         switch (c3cn->state) {
 927         case C3CN_STATE_ACTIVE_CLOSE:
 928                 c3cn_set_state(c3cn, C3CN_STATE_CLOSE_WAIT_1);
 929                 break;
 930         case C3CN_STATE_CLOSE_WAIT_1:
 931         case C3CN_STATE_CLOSE_WAIT_2:
 932                 c3cn_closed(c3cn);
 933                 break;
 934         case C3CN_STATE_ABORTING:
 935                 break;
 936         default:
 937                 cxgb3i_log_error("%s: close_rpl, TID %u in bad state %u\n",
 938                                  c3cn->cdev->name, c3cn->tid, c3cn->state);
 939         }
 940
 941 out:
 942         kfree_skb(skb);
 943 }
 944
 945 static int do_close_con_rpl(struct t3cdev *cdev, struct sk_buff *skb,
 946                             void *ctx)
 947 {
 948         struct s3_conn *c3cn = ctx;
 949
 950         c3cn_conn_debug("rcv, c3cn 0x%p, s %u, f 0x%lx.\n",
 951                          c3cn, c3cn->state, c3cn->flags);
 952
 953         process_cpl_msg_ref(process_close_con_rpl, c3cn, skb);
 954         return 0;
 955 }
 956
 957 /*
 958  * Process ABORT_REQ_RSS CPL message: -> host
 959  * Process abort requests.  If we are waiting for an ABORT_RPL we ignore this
 960  * request except that we need to reply to it.
 961  */
 962
 963 static int abort_status_to_errno(struct s3_conn *c3cn, int abort_reason,
 964                                  int *need_rst)
 965 {
 966         switch (abort_reason) {
 967         case CPL_ERR_BAD_SYN: /* fall through */
 968         case CPL_ERR_CONN_RESET:
 969                 return c3cn->state > C3CN_STATE_ESTABLISHED ?
 970                         -EPIPE : -ECONNRESET;
 971         case CPL_ERR_XMIT_TIMEDOUT:
 972         case CPL_ERR_PERSIST_TIMEDOUT:
 973         case CPL_ERR_FINWAIT2_TIMEDOUT:
 974         case CPL_ERR_KEEPALIVE_TIMEDOUT:
 975                 return -ETIMEDOUT;
 976         default:
 977                 return -EIO;
 978         }
 979 }
 980
 981 static void process_abort_req(struct s3_conn *c3cn, struct sk_buff *skb)
 982 {
 983         int rst_status = CPL_ABORT_NO_RST;
 984         const struct cpl_abort_req_rss *req = cplhdr(skb);
 985
 986         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
 987                         c3cn, c3cn->state, c3cn->flags);
 988
 989         if (!c3cn_flag(c3cn, C3CN_ABORT_REQ_RCVD)) {
 990                 c3cn_set_flag(c3cn, C3CN_ABORT_REQ_RCVD);
 991                 c3cn_set_state(c3cn, C3CN_STATE_ABORTING);
 992                 __kfree_skb(skb);
 993                 return;
 994         }
 995
 996         c3cn_clear_flag(c3cn, C3CN_ABORT_REQ_RCVD);
 997         send_abort_rpl(c3cn, rst_status);
 998
 999         if (!c3cn_flag(c3cn, C3CN_ABORT_RPL_PENDING)) {
1000                 c3cn->err =
1001                     abort_status_to_errno(c3cn, req->status, &rst_status);
1002                 c3cn_closed(c3cn);
1003         }
1004 }
1005
1006 static int do_abort_req(struct t3cdev *cdev, struct sk_buff *skb, void *ctx)
1007 {
1008         const struct cpl_abort_req_rss *req = cplhdr(skb);
1009         struct s3_conn *c3cn = ctx;
1010
1011         c3cn_conn_debug("rcv, c3cn 0x%p, s 0x%x, f 0x%lx.\n",
1012                         c3cn, c3cn->state, c3cn->flags);
1013
1014         if (req->status == CPL_ERR_RTX_NEG_ADVICE ||
1015             req->status == CPL_ERR_PERSIST_NEG_ADVICE) {
1016                 __kfree_skb(skb);
1017                 return 0;
1018         }
1019
1020         process_cpl_msg_ref(process_abort_req, c3cn, skb);
1021         return 0;
1022 }
1023
1024 /*
1025  * Process ABORT_RPL_RSS CPL message: -> host
1026  * Process abort replies.  We only process these messages if we anticipate
1027  * them as the coordination between SW and HW in this area is somewhat lacking
1028  * and sometimes we get ABORT_RPLs after we are done with the connection that
1029  * originated the ABORT_REQ.
1030  */
1031 static void process_abort_rpl(struct s3_conn *c3cn, struct sk_buff *skb)
1032 {
1033         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
1034                         c3cn, c3cn->state, c3cn->flags);
1035
1036         if (c3cn_flag(c3cn, C3CN_ABORT_RPL_PENDING)) {
1037                 if (!c3cn_flag(c3cn, C3CN_ABORT_RPL_RCVD))
1038                         c3cn_set_flag(c3cn, C3CN_ABORT_RPL_RCVD);
1039                 else {
1040                         c3cn_clear_flag(c3cn, C3CN_ABORT_RPL_RCVD);
1041                         c3cn_clear_flag(c3cn, C3CN_ABORT_RPL_PENDING);
1042                         if (c3cn_flag(c3cn, C3CN_ABORT_REQ_RCVD))
1043                                 cxgb3i_log_error("%s tid %u, ABORT_RPL_RSS\n",
1044                                                  c3cn->cdev->name, c3cn->tid);
1045                         c3cn_closed(c3cn);
1046                 }
1047         }
1048         __kfree_skb(skb);
1049 }
1050
1051 static int do_abort_rpl(struct t3cdev *cdev, struct sk_buff *skb, void *ctx)
1052 {
1053         struct cpl_abort_rpl_rss *rpl = cplhdr(skb);
1054         struct s3_conn *c3cn = ctx;
1055
1056         c3cn_conn_debug("rcv, status 0x%x, c3cn 0x%p, s %u, 0x%lx.\n",
1057                         rpl->status, c3cn, c3cn ? c3cn->state : 0,
1058                         c3cn ? c3cn->flags : 0UL);
1059
1060         /*
1061          * Ignore replies to post-close aborts indicating that the abort was
1062          * requested too late.  These connections are terminated when we get
1063          * PEER_CLOSE or CLOSE_CON_RPL and by the time the abort_rpl_rss
1064          * arrives the TID is either no longer used or it has been recycled.
1065          */
1066         if (rpl->status == CPL_ERR_ABORT_FAILED)
1067                 goto discard;
1068
1069         /*
1070          * Sometimes we've already closed the connection, e.g., a post-close
1071          * abort races with ABORT_REQ_RSS, the latter frees the connection
1072          * expecting the ABORT_REQ will fail with CPL_ERR_ABORT_FAILED,
1073          * but FW turns the ABORT_REQ into a regular one and so we get
1074          * ABORT_RPL_RSS with status 0 and no connection.
1075          */
1076         if (!c3cn)
1077                 goto discard;
1078
1079         process_cpl_msg_ref(process_abort_rpl, c3cn, skb);
1080         return 0;
1081
1082 discard:
1083         __kfree_skb(skb);
1084         return 0;
1085 }
1086
1087 /*
1088  * Process RX_ISCSI_HDR CPL message: -> host
1089  * Handle received PDUs, the payload could be DDP'ed. If not, the payload
1090  * follow after the bhs.
1091  */
1092 static void process_rx_iscsi_hdr(struct s3_conn *c3cn, struct sk_buff *skb)
1093 {
1094         struct cpl_iscsi_hdr *hdr_cpl = cplhdr(skb);
1095         struct cpl_iscsi_hdr_norss data_cpl;
1096         struct cpl_rx_data_ddp_norss ddp_cpl;
1097         unsigned int hdr_len, data_len, status;
1098         unsigned int len;
1099         int err;
1100
1101         if (unlikely(c3cn->state >= C3CN_STATE_PASSIVE_CLOSE)) {
1102                 if (c3cn->state != C3CN_STATE_ABORTING)
1103                         send_abort_req(c3cn);
1104                 __kfree_skb(skb);
1105                 return;
1106         }
1107
1108         skb_tcp_seq(skb) = ntohl(hdr_cpl->seq);
1109         skb_flags(skb) = 0;
1110
1111         skb_reset_transport_header(skb);
1112         __skb_pull(skb, sizeof(struct cpl_iscsi_hdr));
1113
1114         len = hdr_len = ntohs(hdr_cpl->len);
1115         /* msg coalesce is off or not enough data received */
1116         if (skb->len <= hdr_len) {
1117                 cxgb3i_log_error("%s: TID %u, ISCSI_HDR, skb len %u < %u.\n",
1118                                  c3cn->cdev->name, c3cn->tid,
1119                                  skb->len, hdr_len);
1120                 goto abort_conn;
1121         }
1122
1123         err = skb_copy_bits(skb, skb->len - sizeof(ddp_cpl), &ddp_cpl,
1124                             sizeof(ddp_cpl));
1125         if (err < 0)
1126                 goto abort_conn;
1127
1128         skb_ulp_mode(skb) = ULP2_FLAG_DATA_READY;
1129         skb_rx_pdulen(skb) = ntohs(ddp_cpl.len);
1130         skb_rx_ddigest(skb) = ntohl(ddp_cpl.ulp_crc);
1131         status = ntohl(ddp_cpl.ddp_status);
1132
1133         c3cn_rx_debug("rx skb 0x%p, len %u, pdulen %u, ddp status 0x%x.\n",
1134                       skb, skb->len, skb_rx_pdulen(skb), status);
1135
1136         if (status & (1 << RX_DDP_STATUS_HCRC_SHIFT))
1137                 skb_ulp_mode(skb) |= ULP2_FLAG_HCRC_ERROR;
1138         if (status & (1 << RX_DDP_STATUS_DCRC_SHIFT))
1139                 skb_ulp_mode(skb) |= ULP2_FLAG_DCRC_ERROR;
1140         if (status & (1 << RX_DDP_STATUS_PAD_SHIFT))
1141                 skb_ulp_mode(skb) |= ULP2_FLAG_PAD_ERROR;
1142
1143         if (skb->len > (hdr_len + sizeof(ddp_cpl))) {
1144                 err = skb_copy_bits(skb, hdr_len, &data_cpl, sizeof(data_cpl));
1145                 if (err < 0)
1146                         goto abort_conn;
1147                 data_len = ntohs(data_cpl.len);
1148                 len += sizeof(data_cpl) + data_len;
1149         } else if (status & (1 << RX_DDP_STATUS_DDP_SHIFT))
1150                 skb_ulp_mode(skb) |= ULP2_FLAG_DATA_DDPED;
1151
1152         c3cn->rcv_nxt = ntohl(ddp_cpl.seq) + skb_rx_pdulen(skb);
1153         __pskb_trim(skb, len);
1154         __skb_queue_tail(&c3cn->receive_queue, skb);
1155         cxgb3i_conn_pdu_ready(c3cn);
1156
1157         return;
1158
1159 abort_conn:
1160         send_abort_req(c3cn);
1161         __kfree_skb(skb);
1162 }
1163
1164 static int do_iscsi_hdr(struct t3cdev *t3dev, struct sk_buff *skb, void *ctx)
1165 {
1166         struct s3_conn *c3cn = ctx;
1167
1168         process_cpl_msg(process_rx_iscsi_hdr, c3cn, skb);
1169         return 0;
1170 }
1171
1172 /*
1173  * Process TX_DATA_ACK CPL messages: -> host
1174  * Process an acknowledgment of WR completion.  Advance snd_una and send the
1175  * next batch of work requests from the write queue.
1176  */
1177 static void check_wr_invariants(struct s3_conn *c3cn)
1178 {
1179         int pending = count_pending_wrs(c3cn);
1180
1181         if (unlikely(c3cn->wr_avail + pending != c3cn->wr_max))
1182                 cxgb3i_log_error("TID %u: credit imbalance: avail %u, "
1183                                 "pending %u, total should be %u\n",
1184                                 c3cn->tid, c3cn->wr_avail, pending,
1185                                 c3cn->wr_max);
1186 }
1187
1188 static void process_wr_ack(struct s3_conn *c3cn, struct sk_buff *skb)
1189 {
1190         struct cpl_wr_ack *hdr = cplhdr(skb);
1191         unsigned int credits = ntohs(hdr->credits);
1192         u32 snd_una = ntohl(hdr->snd_una);
1193
1194         c3cn_tx_debug("%u WR credits, avail %u, unack %u, TID %u, state %u.\n",
1195                         credits, c3cn->wr_avail, c3cn->wr_unacked,
1196                         c3cn->tid, c3cn->state);
1197
1198         c3cn->wr_avail += credits;
1199         if (c3cn->wr_unacked > c3cn->wr_max - c3cn->wr_avail)
1200                 c3cn->wr_unacked = c3cn->wr_max - c3cn->wr_avail;
1201
1202         while (credits) {
1203                 struct sk_buff *p = peek_wr(c3cn);
1204
1205                 if (unlikely(!p)) {
1206                         cxgb3i_log_error("%u WR_ACK credits for TID %u with "
1207                                          "nothing pending, state %u\n",
1208                                          credits, c3cn->tid, c3cn->state);
1209                         break;
1210                 }
1211                 if (unlikely(credits < p->csum)) {
1212                         struct tx_data_wr *w = cplhdr(p);
1213                         cxgb3i_log_error("TID %u got %u WR credits need %u, "
1214                                          "len %u, main body %u, frags %u, "
1215                                          "seq # %u, ACK una %u, ACK nxt %u, "
1216                                          "WR_AVAIL %u, WRs pending %u\n",
1217                                          c3cn->tid, credits, p->csum, p->len,
1218                                          p->len - p->data_len,
1219                                          skb_shinfo(p)->nr_frags,
1220                                          ntohl(w->sndseq), snd_una,
1221                                          ntohl(hdr->snd_nxt), c3cn->wr_avail,
1222                                          count_pending_wrs(c3cn) - credits);
1223                         p->csum -= credits;
1224                         break;
1225                 } else {
1226                         dequeue_wr(c3cn);
1227                         credits -= p->csum;
1228                         free_wr_skb(p);
1229                 }
1230         }
1231
1232         check_wr_invariants(c3cn);
1233
1234         if (unlikely(before(snd_una, c3cn->snd_una))) {
1235                 cxgb3i_log_error("TID %u, unexpected sequence # %u in WR_ACK "
1236                                  "snd_una %u\n",
1237                                  c3cn->tid, snd_una, c3cn->snd_una);
1238                 goto out_free;
1239         }
1240
1241         if (c3cn->snd_una != snd_una) {
1242                 c3cn->snd_una = snd_una;
1243                 dst_confirm(c3cn->dst_cache);
1244         }
1245
1246         if (skb_queue_len(&c3cn->write_queue)) {
1247                 if (c3cn_push_tx_frames(c3cn, 0))
1248                         cxgb3i_conn_tx_open(c3cn);
1249         } else
1250                 cxgb3i_conn_tx_open(c3cn);
1251 out_free:
1252         __kfree_skb(skb);
1253 }
1254
1255 static int do_wr_ack(struct t3cdev *cdev, struct sk_buff *skb, void *ctx)
1256 {
1257         struct s3_conn *c3cn = ctx;
1258
1259         process_cpl_msg(process_wr_ack, c3cn, skb);
1260         return 0;
1261 }
1262
1263 /*
1264  * for each connection, pre-allocate skbs needed for close/abort requests. So
1265  * that we can service the request right away.
1266  */
1267 static void c3cn_free_cpl_skbs(struct s3_conn *c3cn)
1268 {
1269         if (c3cn->cpl_close)
1270                 kfree_skb(c3cn->cpl_close);
1271         if (c3cn->cpl_abort_req)
1272                 kfree_skb(c3cn->cpl_abort_req);
1273         if (c3cn->cpl_abort_rpl)
1274                 kfree_skb(c3cn->cpl_abort_rpl);
1275 }
1276
1277 static int c3cn_alloc_cpl_skbs(struct s3_conn *c3cn)
1278 {
1279         c3cn->cpl_close = alloc_skb(sizeof(struct cpl_close_con_req),
1280                                    GFP_KERNEL);
1281         if (!c3cn->cpl_close)
1282                 return -ENOMEM;
1283         skb_put(c3cn->cpl_close, sizeof(struct cpl_close_con_req));
1284
1285         c3cn->cpl_abort_req = alloc_skb(sizeof(struct cpl_abort_req),
1286                                         GFP_KERNEL);
1287         if (!c3cn->cpl_abort_req)
1288                 goto free_cpl_skbs;
1289         skb_put(c3cn->cpl_abort_req, sizeof(struct cpl_abort_req));
1290
1291         c3cn->cpl_abort_rpl = alloc_skb(sizeof(struct cpl_abort_rpl),
1292                                         GFP_KERNEL);
1293         if (!c3cn->cpl_abort_rpl)
1294                 goto free_cpl_skbs;
1295         skb_put(c3cn->cpl_abort_rpl, sizeof(struct cpl_abort_rpl));
1296
1297         return 0;
1298
1299 free_cpl_skbs:
1300         c3cn_free_cpl_skbs(c3cn);
1301         return -ENOMEM;
1302 }
1303
1304 /**
1305  * c3cn_release_offload_resources - release offload resource
1306  * @c3cn: the offloaded iscsi tcp connection.
1307  * Release resources held by an offload connection (TID, L2T entry, etc.)
1308  */
1309 static void c3cn_release_offload_resources(struct s3_conn *c3cn)
1310 {
1311         struct t3cdev *cdev = c3cn->cdev;
1312         unsigned int tid = c3cn->tid;
1313
1314         c3cn->qset = 0;
1315         c3cn_free_cpl_skbs(c3cn);
1316
1317         if (c3cn->wr_avail != c3cn->wr_max) {
1318                 purge_wr_queue(c3cn);
1319                 reset_wr_list(c3cn);
1320         }
1321
1322         if (cdev) {
1323                 if (c3cn->l2t) {
1324                         l2t_release(L2DATA(cdev), c3cn->l2t);
1325                         c3cn->l2t = NULL;
1326                 }
1327                 if (c3cn->state == C3CN_STATE_CONNECTING)
1328                         /* we have ATID */
1329                         s3_free_atid(cdev, tid);
1330                 else {
1331                         /* we have TID */
1332                         cxgb3_remove_tid(cdev, (void *)c3cn, tid);
1333                         c3cn_put(c3cn);
1334                 }
1335         }
1336
1337         c3cn->dst_cache = NULL;
1338         c3cn->cdev = NULL;
1339 }
1340
1341 /**
1342  * cxgb3i_c3cn_create - allocate and initialize an s3_conn structure
1343  * returns the s3_conn structure allocated.
1344  */
1345 struct s3_conn *cxgb3i_c3cn_create(void)
1346 {
1347         struct s3_conn *c3cn;
1348
1349         c3cn = kzalloc(sizeof(*c3cn), GFP_KERNEL);
1350         if (!c3cn)
1351                 return NULL;
1352
1353         /* pre-allocate close/abort cpl, so we don't need to wait for memory
1354            when close/abort is requested. */
1355         if (c3cn_alloc_cpl_skbs(c3cn) < 0)
1356                 goto free_c3cn;
1357
1358         c3cn_conn_debug("alloc c3cn 0x%p.\n", c3cn);
1359
1360         c3cn->flags = 0;
1361         spin_lock_init(&c3cn->lock);
1362         atomic_set(&c3cn->refcnt, 1);
1363         skb_queue_head_init(&c3cn->receive_queue);
1364         skb_queue_head_init(&c3cn->write_queue);
1365         setup_timer(&c3cn->retry_timer, NULL, (unsigned long)c3cn);
1366         rwlock_init(&c3cn->callback_lock);
1367
1368         return c3cn;
1369
1370 free_c3cn:
1371         kfree(c3cn);
1372         return NULL;
1373 }
1374
1375 static void c3cn_active_close(struct s3_conn *c3cn)
1376 {
1377         int data_lost;
1378         int close_req = 0;
1379
1380         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
1381                          c3cn, c3cn->state, c3cn->flags);
1382
1383         dst_confirm(c3cn->dst_cache);
1384
1385         c3cn_hold(c3cn);
1386         spin_lock_bh(&c3cn->lock);
1387
1388         data_lost = skb_queue_len(&c3cn->receive_queue);
1389         __skb_queue_purge(&c3cn->receive_queue);
1390
1391         switch (c3cn->state) {
1392         case C3CN_STATE_CLOSED:
1393         case C3CN_STATE_ACTIVE_CLOSE:
1394         case C3CN_STATE_CLOSE_WAIT_1:
1395         case C3CN_STATE_CLOSE_WAIT_2:
1396         case C3CN_STATE_ABORTING:
1397                 /* nothing need to be done */
1398                 break;
1399         case C3CN_STATE_CONNECTING:
1400                 /* defer until cpl_act_open_rpl or cpl_act_establish */
1401                 c3cn_set_flag(c3cn, C3CN_ACTIVE_CLOSE_NEEDED);
1402                 break;
1403         case C3CN_STATE_ESTABLISHED:
1404                 close_req = 1;
1405                 c3cn_set_state(c3cn, C3CN_STATE_ACTIVE_CLOSE);
1406                 break;
1407         case C3CN_STATE_PASSIVE_CLOSE:
1408                 close_req = 1;
1409                 c3cn_set_state(c3cn, C3CN_STATE_CLOSE_WAIT_2);
1410                 break;
1411         }
1412
1413         if (close_req) {
1414                 if (data_lost)
1415                         /* Unread data was tossed, zap the connection. */
1416                         send_abort_req(c3cn);
1417                 else
1418                         send_close_req(c3cn);
1419         }
1420
1421         spin_unlock_bh(&c3cn->lock);
1422         c3cn_put(c3cn);
1423 }
1424
1425 /**
1426  * cxgb3i_c3cn_release - close and release an iscsi tcp connection and any
1427  *      resource held
1428  * @c3cn: the iscsi tcp connection
1429  */
1430 void cxgb3i_c3cn_release(struct s3_conn *c3cn)
1431 {
1432         c3cn_conn_debug("c3cn 0x%p, s %u, f 0x%lx.\n",
1433                         c3cn, c3cn->state, c3cn->flags);
1434         if (unlikely(c3cn->state == C3CN_STATE_CONNECTING))
1435                 c3cn_set_flag(c3cn, C3CN_ACTIVE_CLOSE_NEEDED);
1436         else if (likely(c3cn->state != C3CN_STATE_CLOSED))
1437                 c3cn_active_close(c3cn);
1438         c3cn_put(c3cn);
1439 }
1440
1441 static int is_cxgb3_dev(struct net_device *dev)
1442 {
1443         struct cxgb3i_sdev_data *cdata;
1444         struct net_device *ndev = dev;
1445
1446         if (dev->priv_flags & IFF_802_1Q_VLAN)
1447                 ndev = vlan_dev_real_dev(dev);
1448
1449         write_lock(&cdata_rwlock);
1450         list_for_each_entry(cdata, &cdata_list, list) {
1451                 struct adap_ports *ports = &cdata->ports;
1452                 int i;
1453
1454                 for (i = 0; i < ports->nports; i++)
1455                         if (ndev == ports->lldevs[i]) {
1456                                 write_unlock(&cdata_rwlock);
1457                                 return 1;
1458                         }
1459         }
1460         write_unlock(&cdata_rwlock);
1461         return 0;
1462 }
1463
1464 /**
1465  * cxgb3_egress_dev - return the cxgb3 egress device
1466  * @root_dev: the root device anchoring the search
1467  * @c3cn: the connection used to determine egress port in bonding mode
1468  * @context: in bonding mode, indicates a connection set up or failover
1469  *
1470  * Return egress device or NULL if the egress device isn't one of our ports.
1471  */
1472 static struct net_device *cxgb3_egress_dev(struct net_device *root_dev,
1473                                            struct s3_conn *c3cn,
1474                                            int context)
1475 {
1476         while (root_dev) {
1477                 if (root_dev->priv_flags & IFF_802_1Q_VLAN)
1478                         root_dev = vlan_dev_real_dev(root_dev);
1479                 else if (is_cxgb3_dev(root_dev))
1480                         return root_dev;
1481                 else
1482                         return NULL;
1483         }
1484         return NULL;
1485 }
1486
1487 static struct rtable *find_route(struct net_device *dev,
1488                                  __be32 saddr, __be32 daddr,
1489                                  __be16 sport, __be16 dport)
1490 {
1491         struct rtable *rt;
1492         struct flowi fl = {
1493                 .oif = dev ? dev->ifindex : 0,
1494                 .nl_u = {
1495                          .ip4_u = {
1496                                    .daddr = daddr,
1497                                    .saddr = saddr,
1498                                    .tos = 0 } },
1499                 .proto = IPPROTO_TCP,
1500                 .uli_u = {
1501                           .ports = {
1502                                     .sport = sport,
1503                                     .dport = dport } } };
1504
1505         if (ip_route_output_flow(&init_net, &rt, &fl, NULL, 0))
1506                 return NULL;
1507         return rt;
1508 }
1509
1510 /*
1511  * Assign offload parameters to some connection fields.
1512  */
1513 static void init_offload_conn(struct s3_conn *c3cn,
1514                               struct t3cdev *cdev,
1515                               struct dst_entry *dst)
1516 {
1517         BUG_ON(c3cn->cdev != cdev);
1518         c3cn->wr_max = c3cn->wr_avail = T3C_DATA(cdev)->max_wrs - 1;
1519         c3cn->wr_unacked = 0;
1520         c3cn->mss_idx = select_mss(c3cn, dst_mtu(dst));
1521
1522         reset_wr_list(c3cn);
1523 }
1524
1525 static int initiate_act_open(struct s3_conn *c3cn, struct net_device *dev)
1526 {
1527         struct cxgb3i_sdev_data *cdata = NDEV2CDATA(dev);
1528         struct t3cdev *cdev = cdata->cdev;
1529         struct dst_entry *dst = c3cn->dst_cache;
1530         struct sk_buff *skb;
1531
1532         c3cn_conn_debug("c3cn 0x%p, state %u, flag 0x%lx.\n",
1533                         c3cn, c3cn->state, c3cn->flags);
1534         /*
1535          * Initialize connection data.  Note that the flags and ULP mode are
1536          * initialized higher up ...
1537          */
1538         c3cn->dev = dev;
1539         c3cn->cdev = cdev;
1540         c3cn->tid = cxgb3_alloc_atid(cdev, cdata->client, c3cn);
1541         if (c3cn->tid < 0)
1542                 goto out_err;
1543
1544         c3cn->qset = 0;
1545         c3cn->l2t = t3_l2t_get(cdev, dst->neighbour, dev);
1546         if (!c3cn->l2t)
1547                 goto free_tid;
1548
1549         skb = alloc_skb(sizeof(struct cpl_act_open_req), GFP_KERNEL);
1550         if (!skb)
1551                 goto free_l2t;
1552
1553         skb->sk = (struct sock *)c3cn;
1554         set_arp_failure_handler(skb, act_open_req_arp_failure);
1555
1556         c3cn_hold(c3cn);
1557
1558         init_offload_conn(c3cn, cdev, dst);
1559         c3cn->err = 0;
1560
1561         make_act_open_req(c3cn, skb, c3cn->tid, c3cn->l2t);
1562         l2t_send(cdev, skb, c3cn->l2t);
1563         return 0;
1564
1565 free_l2t:
1566         l2t_release(L2DATA(cdev), c3cn->l2t);
1567 free_tid:
1568         s3_free_atid(cdev, c3cn->tid);
1569         c3cn->tid = 0;
1570 out_err:
1571         return -EINVAL;
1572 }
1573
1574 /**
1575  * cxgb3i_find_dev - find the interface associated with the given address
1576  * @ipaddr: ip address
1577  */
1578 static struct net_device *
1579 cxgb3i_find_dev(struct net_device *dev, __be32 ipaddr)
1580 {
1581         struct flowi fl;
1582         int err;
1583         struct rtable *rt;
1584
1585         memset(&fl, 0, sizeof(fl));
1586         fl.nl_u.ip4_u.daddr = ipaddr;
1587
1588         err = ip_route_output_key(dev ? dev_net(dev) : &init_net, &rt, &fl);
1589         if (!err)
1590                 return (&rt->u.dst)->dev;
1591
1592         return NULL;
1593 }
1594
1595 /**
1596  * cxgb3i_c3cn_connect - initiates an iscsi tcp connection to a given address
1597  * @c3cn: the iscsi tcp connection
1598  * @usin: destination address
1599  *
1600  * return 0 if active open request is sent, < 0 otherwise.
1601  */
1602 int cxgb3i_c3cn_connect(struct net_device *dev, struct s3_conn *c3cn,
1603                         struct sockaddr_in *usin)
1604 {
1605         struct rtable *rt;
1606         struct cxgb3i_sdev_data *cdata;
1607         struct t3cdev *cdev;
1608         __be32 sipv4;
1609         struct net_device *dstdev;
1610         int err;
1611
1612         c3cn_conn_debug("c3cn 0x%p, dev 0x%p.\n", c3cn, dev);
1613
1614         if (usin->sin_family != AF_INET)
1615                 return -EAFNOSUPPORT;
1616
1617         c3cn->daddr.sin_port = usin->sin_port;
1618         c3cn->daddr.sin_addr.s_addr = usin->sin_addr.s_addr;
1619
1620         dstdev = cxgb3i_find_dev(dev, usin->sin_addr.s_addr);
1621         if (!dstdev || !is_cxgb3_dev(dstdev))
1622                 return -ENETUNREACH;
1623
1624         if (dstdev->priv_flags & IFF_802_1Q_VLAN)
1625                 dev = dstdev;
1626
1627         rt = find_route(dev, c3cn->saddr.sin_addr.s_addr,
1628                         c3cn->daddr.sin_addr.s_addr,
1629                         c3cn->saddr.sin_port,
1630                         c3cn->daddr.sin_port);
1631         if (rt == NULL) {
1632                 c3cn_conn_debug("NO route to 0x%x, port %u, dev %s.\n",
1633                                 c3cn->daddr.sin_addr.s_addr,
1634                                 ntohs(c3cn->daddr.sin_port),
1635                                 dev ? dev->name : "any");
1636                 return -ENETUNREACH;
1637         }
1638
1639         if (rt->rt_flags & (RTCF_MULTICAST | RTCF_BROADCAST)) {
1640                 c3cn_conn_debug("multi-cast route to 0x%x, port %u, dev %s.\n",
1641                                 c3cn->daddr.sin_addr.s_addr,
1642                                 ntohs(c3cn->daddr.sin_port),
1643                                 dev ? dev->name : "any");
1644                 ip_rt_put(rt);
1645                 return -ENETUNREACH;
1646         }
1647
1648         if (!c3cn->saddr.sin_addr.s_addr)
1649                 c3cn->saddr.sin_addr.s_addr = rt->rt_src;
1650
1651         /* now commit destination to connection */
1652         c3cn->dst_cache = &rt->u.dst;
1653
1654         /* try to establish an offloaded connection */
1655         dev = cxgb3_egress_dev(c3cn->dst_cache->dev, c3cn, 0);
1656         if (dev == NULL) {
1657                 c3cn_conn_debug("c3cn 0x%p, egress dev NULL.\n", c3cn);
1658                 return -ENETUNREACH;
1659         }
1660         cdata = NDEV2CDATA(dev);
1661         cdev = cdata->cdev;
1662
1663         /* get a source port if one hasn't been provided */
1664         err = c3cn_get_port(c3cn, cdata);
1665         if (err)
1666                 return err;
1667
1668         c3cn_conn_debug("c3cn 0x%p get port %u.\n",
1669                         c3cn, ntohs(c3cn->saddr.sin_port));
1670
1671         sipv4 = cxgb3i_get_private_ipv4addr(dev);
1672         if (!sipv4) {
1673                 c3cn_conn_debug("c3cn 0x%p, iscsi ip not configured.\n", c3cn);
1674                 sipv4 = c3cn->saddr.sin_addr.s_addr;
1675                 cxgb3i_set_private_ipv4addr(dev, sipv4);
1676         } else
1677                 c3cn->saddr.sin_addr.s_addr = sipv4;
1678
1679         c3cn_conn_debug("c3cn 0x%p, %pI4,%u-%pI4,%u SYN_SENT.\n",
1680                         c3cn,
1681                         &c3cn->saddr.sin_addr.s_addr,
1682                         ntohs(c3cn->saddr.sin_port),
1683                         &c3cn->daddr.sin_addr.s_addr,
1684                         ntohs(c3cn->daddr.sin_port));
1685
1686         c3cn_set_state(c3cn, C3CN_STATE_CONNECTING);
1687         if (!initiate_act_open(c3cn, dev))
1688                 return 0;
1689
1690         /*
1691          * If we get here, we don't have an offload connection so simply
1692          * return a failure.
1693          */
1694         err = -ENOTSUPP;
1695
1696         /*
1697          * This trashes the connection and releases the local port,
1698          * if necessary.
1699          */
1700         c3cn_conn_debug("c3cn 0x%p -> CLOSED.\n", c3cn);
1701         c3cn_set_state(c3cn, C3CN_STATE_CLOSED);
1702         ip_rt_put(rt);
1703         c3cn_put_port(c3cn);
1704         return err;
1705 }
1706
1707 /**
1708  * cxgb3i_c3cn_rx_credits - ack received tcp data.
1709  * @c3cn: iscsi tcp connection
1710  * @copied: # of bytes processed
1711  *
1712  * Called after some received data has been read.  It returns RX credits
1713  * to the HW for the amount of data processed.
1714  */
1715 void cxgb3i_c3cn_rx_credits(struct s3_conn *c3cn, int copied)
1716 {
1717         struct t3cdev *cdev;
1718         int must_send;
1719         u32 credits, dack = 0;
1720
1721         if (c3cn->state != C3CN_STATE_ESTABLISHED)
1722                 return;
1723
1724         credits = c3cn->copied_seq - c3cn->rcv_wup;
1725         if (unlikely(!credits))
1726                 return;
1727
1728         cdev = c3cn->cdev;
1729
1730         if (unlikely(cxgb3_rx_credit_thres == 0))
1731                 return;
1732
1733         dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1);
1734
1735         /*
1736          * For coalescing to work effectively ensure the receive window has
1737          * at least 16KB left.
1738          */
1739         must_send = credits + 16384 >= cxgb3_rcv_win;
1740
1741         if (must_send || credits >= cxgb3_rx_credit_thres)
1742                 c3cn->rcv_wup += send_rx_credits(c3cn, credits, dack);
1743 }
1744
1745 /**
1746  * cxgb3i_c3cn_send_pdus - send the skbs containing iscsi pdus
1747  * @c3cn: iscsi tcp connection
1748  * @skb: skb contains the iscsi pdu
1749  *
1750  * Add a list of skbs to a connection send queue. The skbs must comply with
1751  * the max size limit of the device and have a headroom of at least
1752  * TX_HEADER_LEN bytes.
1753  * Return # of bytes queued.
1754  */
1755 int cxgb3i_c3cn_send_pdus(struct s3_conn *c3cn, struct sk_buff *skb)
1756 {
1757         struct sk_buff *next;
1758         int err, copied = 0;
1759
1760         spin_lock_bh(&c3cn->lock);
1761
1762         if (c3cn->state != C3CN_STATE_ESTABLISHED) {
1763                 c3cn_tx_debug("c3cn 0x%p, not in est. state %u.\n",
1764                               c3cn, c3cn->state);
1765                 err = -EAGAIN;
1766                 goto out_err;
1767         }
1768
1769         if (c3cn->err) {
1770                 c3cn_tx_debug("c3cn 0x%p, err %d.\n", c3cn, c3cn->err);
1771                 err = -EPIPE;
1772                 goto out_err;
1773         }
1774
1775         if (c3cn->write_seq - c3cn->snd_una >= cxgb3_snd_win) {
1776                 c3cn_tx_debug("c3cn 0x%p, snd %u - %u > %u.\n",
1777                                 c3cn, c3cn->write_seq, c3cn->snd_una,
1778                                 cxgb3_snd_win);
1779                 err = -ENOBUFS;
1780                 goto out_err;
1781         }
1782
1783         while (skb) {
1784                 int frags = skb_shinfo(skb)->nr_frags +
1785                                 (skb->len != skb->data_len);
1786
1787                 if (unlikely(skb_headroom(skb) < TX_HEADER_LEN)) {
1788                         c3cn_tx_debug("c3cn 0x%p, skb head.\n", c3cn);
1789                         err = -EINVAL;
1790                         goto out_err;
1791                 }
1792
1793                 if (frags >= SKB_WR_LIST_SIZE) {
1794                         cxgb3i_log_error("c3cn 0x%p, tx frags %d, len %u,%u.\n",
1795                                          c3cn, skb_shinfo(skb)->nr_frags,
1796                                          skb->len, skb->data_len);
1797                         err = -EINVAL;
1798                         goto out_err;
1799                 }
1800
1801                 next = skb->next;
1802                 skb->next = NULL;
1803                 skb_entail(c3cn, skb, C3CB_FLAG_NO_APPEND | C3CB_FLAG_NEED_HDR);
1804                 copied += skb->len;
1805                 c3cn->write_seq += skb->len + ulp_extra_len(skb);
1806                 skb = next;
1807         }
1808 done:
1809         if (likely(skb_queue_len(&c3cn->write_queue)))
1810                 c3cn_push_tx_frames(c3cn, 1);
1811         spin_unlock_bh(&c3cn->lock);
1812         return copied;
1813
1814 out_err:
1815         if (copied == 0 && err == -EPIPE)
1816                 copied = c3cn->err ? c3cn->err : -EPIPE;
1817         else
1818                 copied = err;
1819         goto done;
1820 }
1821
1822 static void sdev_data_cleanup(struct cxgb3i_sdev_data *cdata)
1823 {
1824         struct adap_ports *ports = &cdata->ports;
1825         struct s3_conn *c3cn;
1826         int i;
1827
1828         for (i = 0; i < cxgb3_max_connect; i++) {
1829                 if (cdata->sport_conn[i]) {
1830                         c3cn = cdata->sport_conn[i];
1831                         cdata->sport_conn[i] = NULL;
1832
1833                         spin_lock_bh(&c3cn->lock);
1834                         c3cn->cdev = NULL;
1835                         c3cn_set_flag(c3cn, C3CN_OFFLOAD_DOWN);
1836                         c3cn_closed(c3cn);
1837                         spin_unlock_bh(&c3cn->lock);
1838                 }
1839         }
1840
1841         for (i = 0; i < ports->nports; i++)
1842                 NDEV2CDATA(ports->lldevs[i]) = NULL;
1843
1844         cxgb3i_free_big_mem(cdata);
1845 }
1846
1847 void cxgb3i_sdev_cleanup(void)
1848 {
1849         struct cxgb3i_sdev_data *cdata;
1850
1851         write_lock(&cdata_rwlock);
1852         list_for_each_entry(cdata, &cdata_list, list) {
1853                 list_del(&cdata->list);
1854                 sdev_data_cleanup(cdata);
1855         }
1856         write_unlock(&cdata_rwlock);
1857 }
1858
1859 int cxgb3i_sdev_init(cxgb3_cpl_handler_func *cpl_handlers)
1860 {
1861         cpl_handlers[CPL_ACT_ESTABLISH] = do_act_establish;
1862         cpl_handlers[CPL_ACT_OPEN_RPL] = do_act_open_rpl;
1863         cpl_handlers[CPL_PEER_CLOSE] = do_peer_close;
1864         cpl_handlers[CPL_ABORT_REQ_RSS] = do_abort_req;
1865         cpl_handlers[CPL_ABORT_RPL_RSS] = do_abort_rpl;
1866         cpl_handlers[CPL_CLOSE_CON_RPL] = do_close_con_rpl;
1867         cpl_handlers[CPL_TX_DMA_ACK] = do_wr_ack;
1868         cpl_handlers[CPL_ISCSI_HDR] = do_iscsi_hdr;
1869
1870         if (cxgb3_max_connect > CXGB3I_MAX_CONN)
1871                 cxgb3_max_connect = CXGB3I_MAX_CONN;
1872         return 0;
1873 }
1874
1875 /**
1876  * cxgb3i_sdev_add - allocate and initialize resources for each adapter found
1877  * @cdev:       t3cdev adapter
1878  * @client:     cxgb3 driver client
1879  */
1880 void cxgb3i_sdev_add(struct t3cdev *cdev, struct cxgb3_client *client)
1881 {
1882         struct cxgb3i_sdev_data *cdata;
1883         struct ofld_page_info rx_page_info;
1884         unsigned int wr_len;
1885         int mapsize = cxgb3_max_connect * sizeof(struct s3_conn *);
1886         int i;
1887
1888         cdata =  cxgb3i_alloc_big_mem(sizeof(*cdata) + mapsize, GFP_KERNEL);
1889         if (!cdata) {
1890                 cxgb3i_log_warn("t3dev 0x%p, offload up, OOM %d.\n",
1891                                 cdev, mapsize);
1892                 return;
1893         }
1894
1895         if (cdev->ctl(cdev, GET_WR_LEN, &wr_len) < 0 ||
1896             cdev->ctl(cdev, GET_PORTS, &cdata->ports) < 0 ||
1897             cdev->ctl(cdev, GET_RX_PAGE_INFO, &rx_page_info) < 0) {
1898                 cxgb3i_log_warn("t3dev 0x%p, offload up, ioctl failed.\n",
1899                                 cdev);
1900                 goto free_cdata;
1901         }
1902
1903         s3_init_wr_tab(wr_len);
1904
1905         spin_lock_init(&cdata->lock);
1906         INIT_LIST_HEAD(&cdata->list);
1907         cdata->cdev = cdev;
1908         cdata->client = client;
1909
1910         for (i = 0; i < cdata->ports.nports; i++)
1911                 NDEV2CDATA(cdata->ports.lldevs[i]) = cdata;
1912
1913         write_lock(&cdata_rwlock);
1914         list_add_tail(&cdata->list, &cdata_list);
1915         write_unlock(&cdata_rwlock);
1916
1917         cxgb3i_log_info("t3dev 0x%p, offload up, added.\n", cdev);
1918         return;
1919
1920 free_cdata:
1921         cxgb3i_free_big_mem(cdata);
1922 }
1923
1924 /**
1925  * cxgb3i_sdev_remove - free the allocated resources for the adapter
1926  * @cdev:       t3cdev adapter
1927  */
1928 void cxgb3i_sdev_remove(struct t3cdev *cdev)
1929 {
1930         struct cxgb3i_sdev_data *cdata = CXGB3_SDEV_DATA(cdev);
1931
1932         cxgb3i_log_info("t3dev 0x%p, offload down, remove.\n", cdev);
1933
1934         write_lock(&cdata_rwlock);
1935         list_del(&cdata->list);
1936         write_unlock(&cdata_rwlock);
1937
1938         sdev_data_cleanup(cdata);
1939 }