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

   1 /*
   2  * iSCSI Initiator over TCP/IP Data-Path
   3  *
   4  * Copyright (C) 2004 Dmitry Yusupov
   5  * Copyright (C) 2004 Alex Aizman
   6  * Copyright (C) 2005 - 2006 Mike Christie
   7  * Copyright (C) 2006 Red Hat, Inc.  All rights reserved.
   8  * maintained by open-iscsi@googlegroups.com
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License as published
  12  * by the Free Software Foundation; either version 2 of the License, or
  13  * (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful, but
  16  * WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  18  * General Public License for more details.
  19  *
  20  * See the file COPYING included with this distribution for more details.
  21  *
  22  * Credits:
  23  *      Christoph Hellwig
  24  *      FUJITA Tomonori
  25  *      Arne Redlich
  26  *      Zhenyu Wang
  27  */
  28
  29 #include <linux/types.h>
  30 #include <linux/list.h>
  31 #include <linux/inet.h>
  32 #include <linux/blkdev.h>
  33 #include <linux/crypto.h>
  34 #include <linux/delay.h>
  35 #include <linux/kfifo.h>
  36 #include <linux/scatterlist.h>
  37 #include <linux/mutex.h>
  38 #include <net/tcp.h>
  39 #include <scsi/scsi_cmnd.h>
  40 #include <scsi/scsi_host.h>
  41 #include <scsi/scsi.h>
  42 #include <scsi/scsi_transport_iscsi.h>
  43
  44 #include "iscsi_tcp.h"
  45
  46 MODULE_AUTHOR("Dmitry Yusupov <dmitry_yus@yahoo.com>, "
  47               "Alex Aizman <itn780@yahoo.com>");
  48 MODULE_DESCRIPTION("iSCSI/TCP data-path");
  49 MODULE_LICENSE("GPL");
  50 MODULE_VERSION("0:4.445");
  51 /* #define DEBUG_TCP */
  52 #define DEBUG_ASSERT
  53
  54 #ifdef DEBUG_TCP
  55 #define debug_tcp(fmt...) printk(KERN_INFO "tcp: " fmt)
  56 #else
  57 #define debug_tcp(fmt...)
  58 #endif
  59
  60 #ifndef DEBUG_ASSERT
  61 #ifdef BUG_ON
  62 #undef BUG_ON
  63 #endif
  64 #define BUG_ON(expr)
  65 #endif
  66
  67 static unsigned int iscsi_max_lun = 512;
  68 module_param_named(max_lun, iscsi_max_lun, uint, S_IRUGO);
  69
  70 /* global data */
  71 static kmem_cache_t *taskcache;
  72
  73 static inline void
  74 iscsi_buf_init_virt(struct iscsi_buf *ibuf, char *vbuf, int size)
  75 {
  76         sg_init_one(&ibuf->sg, (u8 *)vbuf, size);
  77         ibuf->sent = 0;
  78         ibuf->use_sendmsg = 0;
  79 }
  80
  81 static inline void
  82 iscsi_buf_init_iov(struct iscsi_buf *ibuf, char *vbuf, int size)
  83 {
  84         ibuf->sg.page = virt_to_page(vbuf);
  85         ibuf->sg.offset = offset_in_page(vbuf);
  86         ibuf->sg.length = size;
  87         ibuf->sent = 0;
  88         ibuf->use_sendmsg = 1;
  89 }
  90
  91 static inline void
  92 iscsi_buf_init_sg(struct iscsi_buf *ibuf, struct scatterlist *sg)
  93 {
  94         ibuf->sg.page = sg->page;
  95         ibuf->sg.offset = sg->offset;
  96         ibuf->sg.length = sg->length;
  97         /*
  98          * Fastpath: sg element fits into single page
  99          */
 100         if (sg->length + sg->offset <= PAGE_SIZE && !PageSlab(sg->page))
 101                 ibuf->use_sendmsg = 0;
 102         else
 103                 ibuf->use_sendmsg = 1;
 104         ibuf->sent = 0;
 105 }
 106
 107 static inline int
 108 iscsi_buf_left(struct iscsi_buf *ibuf)
 109 {
 110         int rc;
 111
 112         rc = ibuf->sg.length - ibuf->sent;
 113         BUG_ON(rc < 0);
 114         return rc;
 115 }
 116
 117 static inline void
 118 iscsi_hdr_digest(struct iscsi_conn *conn, struct iscsi_buf *buf,
 119                  u8* crc)
 120 {
 121         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 122
 123         crypto_digest_digest(tcp_conn->tx_tfm, &buf->sg, 1, crc);
 124         buf->sg.length += sizeof(uint32_t);
 125 }
 126
 127 static inline int
 128 iscsi_hdr_extract(struct iscsi_tcp_conn *tcp_conn)
 129 {
 130         struct sk_buff *skb = tcp_conn->in.skb;
 131
 132         tcp_conn->in.zero_copy_hdr = 0;
 133
 134         if (tcp_conn->in.copy >= tcp_conn->hdr_size &&
 135             tcp_conn->in_progress == IN_PROGRESS_WAIT_HEADER) {
 136                 /*
 137                  * Zero-copy PDU Header: using connection context
 138                  * to store header pointer.
 139                  */
 140                 if (skb_shinfo(skb)->frag_list == NULL &&
 141                     !skb_shinfo(skb)->nr_frags) {
 142                         tcp_conn->in.hdr = (struct iscsi_hdr *)
 143                                 ((char*)skb->data + tcp_conn->in.offset);
 144                         tcp_conn->in.zero_copy_hdr = 1;
 145                 } else {
 146                         /* ignoring return code since we checked
 147                          * in.copy before */
 148                         skb_copy_bits(skb, tcp_conn->in.offset,
 149                                 &tcp_conn->hdr, tcp_conn->hdr_size);
 150                         tcp_conn->in.hdr = &tcp_conn->hdr;
 151                 }
 152                 tcp_conn->in.offset += tcp_conn->hdr_size;
 153                 tcp_conn->in.copy -= tcp_conn->hdr_size;
 154         } else {
 155                 int hdr_remains;
 156                 int copylen;
 157
 158                 /*
 159                  * PDU header scattered across SKB's,
 160                  * copying it... This'll happen quite rarely.
 161                  */
 162
 163                 if (tcp_conn->in_progress == IN_PROGRESS_WAIT_HEADER)
 164                         tcp_conn->in.hdr_offset = 0;
 165
 166                 hdr_remains = tcp_conn->hdr_size - tcp_conn->in.hdr_offset;
 167                 BUG_ON(hdr_remains <= 0);
 168
 169                 copylen = min(tcp_conn->in.copy, hdr_remains);
 170                 skb_copy_bits(skb, tcp_conn->in.offset,
 171                         (char*)&tcp_conn->hdr + tcp_conn->in.hdr_offset,
 172                         copylen);
 173
 174                 debug_tcp("PDU gather offset %d bytes %d in.offset %d "
 175                        "in.copy %d\n", tcp_conn->in.hdr_offset, copylen,
 176                        tcp_conn->in.offset, tcp_conn->in.copy);
 177
 178                 tcp_conn->in.offset += copylen;
 179                 tcp_conn->in.copy -= copylen;
 180                 if (copylen < hdr_remains)  {
 181                         tcp_conn->in_progress = IN_PROGRESS_HEADER_GATHER;
 182                         tcp_conn->in.hdr_offset += copylen;
 183                         return -EAGAIN;
 184                 }
 185                 tcp_conn->in.hdr = &tcp_conn->hdr;
 186                 tcp_conn->discontiguous_hdr_cnt++;
 187                 tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
 188         }
 189
 190         return 0;
 191 }
 192
 193 /*
 194  * must be called with session lock
 195  */
 196 static void
 197 __iscsi_ctask_cleanup(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
 198 {
 199         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 200         struct scsi_cmnd *sc;
 201
 202         sc = ctask->sc;
 203         if (unlikely(!sc))
 204                 return;
 205
 206         if (sc->sc_data_direction == DMA_TO_DEVICE) {
 207                 struct iscsi_data_task *dtask, *n;
 208
 209                 /* WRITE: cleanup Data-Out's if any */
 210                 list_for_each_entry_safe(dtask, n, &tcp_ctask->dataqueue,
 211                                          item) {
 212                         list_del(&dtask->item);
 213                         mempool_free(dtask, tcp_ctask->datapool);
 214                 }
 215         }
 216         tcp_ctask->xmstate = XMSTATE_IDLE;
 217         tcp_ctask->r2t = NULL;
 218 }
 219
 220 /**
 221  * iscsi_data_rsp - SCSI Data-In Response processing
 222  * @conn: iscsi connection
 223  * @ctask: scsi command task
 224  **/
 225 static int
 226 iscsi_data_rsp(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
 227 {
 228         int rc;
 229         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 230         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 231         struct iscsi_data_rsp *rhdr = (struct iscsi_data_rsp *)tcp_conn->in.hdr;
 232         struct iscsi_session *session = conn->session;
 233         int datasn = be32_to_cpu(rhdr->datasn);
 234
 235         rc = iscsi_check_assign_cmdsn(session, (struct iscsi_nopin*)rhdr);
 236         if (rc)
 237                 return rc;
 238         /*
 239          * setup Data-In byte counter (gets decremented..)
 240          */
 241         ctask->data_count = tcp_conn->in.datalen;
 242
 243         if (tcp_conn->in.datalen == 0)
 244                 return 0;
 245
 246         if (ctask->datasn != datasn)
 247                 return ISCSI_ERR_DATASN;
 248
 249         ctask->datasn++;
 250
 251         tcp_ctask->data_offset = be32_to_cpu(rhdr->offset);
 252         if (tcp_ctask->data_offset + tcp_conn->in.datalen > ctask->total_length)
 253                 return ISCSI_ERR_DATA_OFFSET;
 254
 255         if (rhdr->flags & ISCSI_FLAG_DATA_STATUS) {
 256                 struct scsi_cmnd *sc = ctask->sc;
 257
 258                 conn->exp_statsn = be32_to_cpu(rhdr->statsn) + 1;
 259                 if (rhdr->flags & ISCSI_FLAG_DATA_UNDERFLOW) {
 260                         int res_count = be32_to_cpu(rhdr->residual_count);
 261
 262                         if (res_count > 0 &&
 263                             res_count <= sc->request_bufflen) {
 264                                 sc->resid = res_count;
 265                                 sc->result = (DID_OK << 16) | rhdr->cmd_status;
 266                         } else
 267                                 sc->result = (DID_BAD_TARGET << 16) |
 268                                         rhdr->cmd_status;
 269                 } else if (rhdr->flags & ISCSI_FLAG_DATA_OVERFLOW) {
 270                         sc->resid = be32_to_cpu(rhdr->residual_count);
 271                         sc->result = (DID_OK << 16) | rhdr->cmd_status;
 272                 } else
 273                         sc->result = (DID_OK << 16) | rhdr->cmd_status;
 274         }
 275
 276         conn->datain_pdus_cnt++;
 277         return 0;
 278 }
 279
 280 /**
 281  * iscsi_solicit_data_init - initialize first Data-Out
 282  * @conn: iscsi connection
 283  * @ctask: scsi command task
 284  * @r2t: R2T info
 285  *
 286  * Notes:
 287  *      Initialize first Data-Out within this R2T sequence and finds
 288  *      proper data_offset within this SCSI command.
 289  *
 290  *      This function is called with connection lock taken.
 291  **/
 292 static void
 293 iscsi_solicit_data_init(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask,
 294                         struct iscsi_r2t_info *r2t)
 295 {
 296         struct iscsi_data *hdr;
 297         struct iscsi_data_task *dtask;
 298         struct scsi_cmnd *sc = ctask->sc;
 299         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 300
 301         dtask = mempool_alloc(tcp_ctask->datapool, GFP_ATOMIC);
 302         BUG_ON(!dtask);
 303         INIT_LIST_HEAD(&dtask->item);
 304         hdr = &dtask->hdr;
 305         memset(hdr, 0, sizeof(struct iscsi_data));
 306         hdr->ttt = r2t->ttt;
 307         hdr->datasn = cpu_to_be32(r2t->solicit_datasn);
 308         r2t->solicit_datasn++;
 309         hdr->opcode = ISCSI_OP_SCSI_DATA_OUT;
 310         memcpy(hdr->lun, ctask->hdr->lun, sizeof(hdr->lun));
 311         hdr->itt = ctask->hdr->itt;
 312         hdr->exp_statsn = r2t->exp_statsn;
 313         hdr->offset = cpu_to_be32(r2t->data_offset);
 314         if (r2t->data_length > conn->max_xmit_dlength) {
 315                 hton24(hdr->dlength, conn->max_xmit_dlength);
 316                 r2t->data_count = conn->max_xmit_dlength;
 317                 hdr->flags = 0;
 318         } else {
 319                 hton24(hdr->dlength, r2t->data_length);
 320                 r2t->data_count = r2t->data_length;
 321                 hdr->flags = ISCSI_FLAG_CMD_FINAL;
 322         }
 323         conn->dataout_pdus_cnt++;
 324
 325         r2t->sent = 0;
 326
 327         iscsi_buf_init_virt(&r2t->headbuf, (char*)hdr,
 328                            sizeof(struct iscsi_hdr));
 329
 330         r2t->dtask = dtask;
 331
 332         if (sc->use_sg) {
 333                 int i, sg_count = 0;
 334                 struct scatterlist *sg = sc->request_buffer;
 335
 336                 r2t->sg = NULL;
 337                 for (i = 0; i < sc->use_sg; i++, sg += 1) {
 338                         /* FIXME: prefetch ? */
 339                         if (sg_count + sg->length > r2t->data_offset) {
 340                                 int page_offset;
 341
 342                                 /* sg page found! */
 343
 344                                 /* offset within this page */
 345                                 page_offset = r2t->data_offset - sg_count;
 346
 347                                 /* fill in this buffer */
 348                                 iscsi_buf_init_sg(&r2t->sendbuf, sg);
 349                                 r2t->sendbuf.sg.offset += page_offset;
 350                                 r2t->sendbuf.sg.length -= page_offset;
 351
 352                                 /* xmit logic will continue with next one */
 353                                 r2t->sg = sg + 1;
 354                                 break;
 355                         }
 356                         sg_count += sg->length;
 357                 }
 358                 BUG_ON(r2t->sg == NULL);
 359         } else
 360                 iscsi_buf_init_iov(&tcp_ctask->sendbuf,
 361                             (char*)sc->request_buffer + r2t->data_offset,
 362                             r2t->data_count);
 363
 364         list_add(&dtask->item, &tcp_ctask->dataqueue);
 365 }
 366
 367 /**
 368  * iscsi_r2t_rsp - iSCSI R2T Response processing
 369  * @conn: iscsi connection
 370  * @ctask: scsi command task
 371  **/
 372 static int
 373 iscsi_r2t_rsp(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
 374 {
 375         struct iscsi_r2t_info *r2t;
 376         struct iscsi_session *session = conn->session;
 377         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 378         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 379         struct iscsi_r2t_rsp *rhdr = (struct iscsi_r2t_rsp *)tcp_conn->in.hdr;
 380         int r2tsn = be32_to_cpu(rhdr->r2tsn);
 381         int rc;
 382
 383         if (tcp_conn->in.datalen)
 384                 return ISCSI_ERR_DATALEN;
 385
 386         if (tcp_ctask->exp_r2tsn && tcp_ctask->exp_r2tsn != r2tsn)
 387                 return ISCSI_ERR_R2TSN;
 388
 389         rc = iscsi_check_assign_cmdsn(session, (struct iscsi_nopin*)rhdr);
 390         if (rc)
 391                 return rc;
 392
 393         /* FIXME: use R2TSN to detect missing R2T */
 394
 395         /* fill-in new R2T associated with the task */
 396         spin_lock(&session->lock);
 397         if (!ctask->sc || ctask->mtask ||
 398              session->state != ISCSI_STATE_LOGGED_IN) {
 399                 printk(KERN_INFO "iscsi_tcp: dropping R2T itt %d in "
 400                        "recovery...\n", ctask->itt);
 401                 spin_unlock(&session->lock);
 402                 return 0;
 403         }
 404         rc = __kfifo_get(tcp_ctask->r2tpool.queue, (void*)&r2t, sizeof(void*));
 405         BUG_ON(!rc);
 406
 407         r2t->exp_statsn = rhdr->statsn;
 408         r2t->data_length = be32_to_cpu(rhdr->data_length);
 409         if (r2t->data_length == 0 ||
 410             r2t->data_length > session->max_burst) {
 411                 spin_unlock(&session->lock);
 412                 return ISCSI_ERR_DATALEN;
 413         }
 414
 415         r2t->data_offset = be32_to_cpu(rhdr->data_offset);
 416         if (r2t->data_offset + r2t->data_length > ctask->total_length) {
 417                 spin_unlock(&session->lock);
 418                 return ISCSI_ERR_DATALEN;
 419         }
 420
 421         r2t->ttt = rhdr->ttt; /* no flip */
 422         r2t->solicit_datasn = 0;
 423
 424         iscsi_solicit_data_init(conn, ctask, r2t);
 425
 426         tcp_ctask->exp_r2tsn = r2tsn + 1;
 427         tcp_ctask->xmstate |= XMSTATE_SOL_HDR;
 428         __kfifo_put(tcp_ctask->r2tqueue, (void*)&r2t, sizeof(void*));
 429         __kfifo_put(conn->xmitqueue, (void*)&ctask, sizeof(void*));
 430
 431         scsi_queue_work(session->host, &conn->xmitwork);
 432         conn->r2t_pdus_cnt++;
 433         spin_unlock(&session->lock);
 434
 435         return 0;
 436 }
 437
 438 static int
 439 iscsi_tcp_hdr_recv(struct iscsi_conn *conn)
 440 {
 441         int rc = 0, opcode, ahslen;
 442         struct iscsi_hdr *hdr;
 443         struct iscsi_session *session = conn->session;
 444         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 445         uint32_t cdgst, rdgst = 0, itt;
 446
 447         hdr = tcp_conn->in.hdr;
 448
 449         /* verify PDU length */
 450         tcp_conn->in.datalen = ntoh24(hdr->dlength);
 451         if (tcp_conn->in.datalen > conn->max_recv_dlength) {
 452                 printk(KERN_ERR "iscsi_tcp: datalen %d > %d\n",
 453                        tcp_conn->in.datalen, conn->max_recv_dlength);
 454                 return ISCSI_ERR_DATALEN;
 455         }
 456         tcp_conn->data_copied = 0;
 457
 458         /* read AHS */
 459         ahslen = hdr->hlength << 2;
 460         tcp_conn->in.offset += ahslen;
 461         tcp_conn->in.copy -= ahslen;
 462         if (tcp_conn->in.copy < 0) {
 463                 printk(KERN_ERR "iscsi_tcp: can't handle AHS with length "
 464                        "%d bytes\n", ahslen);
 465                 return ISCSI_ERR_AHSLEN;
 466         }
 467
 468         /* calculate read padding */
 469         tcp_conn->in.padding = tcp_conn->in.datalen & (ISCSI_PAD_LEN-1);
 470         if (tcp_conn->in.padding) {
 471                 tcp_conn->in.padding = ISCSI_PAD_LEN - tcp_conn->in.padding;
 472                 debug_scsi("read padding %d bytes\n", tcp_conn->in.padding);
 473         }
 474
 475         if (conn->hdrdgst_en) {
 476                 struct scatterlist sg;
 477
 478                 sg_init_one(&sg, (u8 *)hdr,
 479                             sizeof(struct iscsi_hdr) + ahslen);
 480                 crypto_digest_digest(tcp_conn->rx_tfm, &sg, 1, (u8 *)&cdgst);
 481                 rdgst = *(uint32_t*)((char*)hdr + sizeof(struct iscsi_hdr) +
 482                                      ahslen);
 483                 if (cdgst != rdgst) {
 484                         printk(KERN_ERR "iscsi_tcp: hdrdgst error "
 485                                "recv 0x%x calc 0x%x\n", rdgst, cdgst);
 486                         return ISCSI_ERR_HDR_DGST;
 487                 }
 488         }
 489
 490         opcode = hdr->opcode & ISCSI_OPCODE_MASK;
 491         /* verify itt (itt encoding: age+cid+itt) */
 492         rc = iscsi_verify_itt(conn, hdr, &itt);
 493         if (rc == ISCSI_ERR_NO_SCSI_CMD) {
 494                 tcp_conn->in.datalen = 0; /* force drop */
 495                 return 0;
 496         } else if (rc)
 497                 return rc;
 498
 499         debug_tcp("opcode 0x%x offset %d copy %d ahslen %d datalen %d\n",
 500                   opcode, tcp_conn->in.offset, tcp_conn->in.copy,
 501                   ahslen, tcp_conn->in.datalen);
 502
 503         switch(opcode) {
 504         case ISCSI_OP_SCSI_DATA_IN:
 505                 tcp_conn->in.ctask = session->cmds[itt];
 506                 rc = iscsi_data_rsp(conn, tcp_conn->in.ctask);
 507                 /* fall through */
 508         case ISCSI_OP_SCSI_CMD_RSP:
 509                 tcp_conn->in.ctask = session->cmds[itt];
 510                 if (tcp_conn->in.datalen)
 511                         goto copy_hdr;
 512
 513                 spin_lock(&session->lock);
 514                 __iscsi_ctask_cleanup(conn, tcp_conn->in.ctask);
 515                 rc = __iscsi_complete_pdu(conn, hdr, NULL, 0);
 516                 spin_unlock(&session->lock);
 517                 break;
 518         case ISCSI_OP_R2T:
 519                 tcp_conn->in.ctask = session->cmds[itt];
 520                 if (ahslen)
 521                         rc = ISCSI_ERR_AHSLEN;
 522                 else if (tcp_conn->in.ctask->sc->sc_data_direction ==
 523                                                                 DMA_TO_DEVICE)
 524                         rc = iscsi_r2t_rsp(conn, tcp_conn->in.ctask);
 525                 else
 526                         rc = ISCSI_ERR_PROTO;
 527                 break;
 528         case ISCSI_OP_LOGIN_RSP:
 529         case ISCSI_OP_TEXT_RSP:
 530         case ISCSI_OP_LOGOUT_RSP:
 531         case ISCSI_OP_NOOP_IN:
 532         case ISCSI_OP_REJECT:
 533         case ISCSI_OP_ASYNC_EVENT:
 534                 if (tcp_conn->in.datalen)
 535                         goto copy_hdr;
 536         /* fall through */
 537         case ISCSI_OP_SCSI_TMFUNC_RSP:
 538                 rc = iscsi_complete_pdu(conn, hdr, NULL, 0);
 539                 break;
 540         default:
 541                 rc = ISCSI_ERR_BAD_OPCODE;
 542                 break;
 543         }
 544
 545         return rc;
 546
 547 copy_hdr:
 548         /*
 549          * if we did zero copy for the header but we will need multiple
 550          * skbs to complete the command then we have to copy the header
 551          * for later use
 552          */
 553         if (tcp_conn->in.zero_copy_hdr && tcp_conn->in.copy <
 554            (tcp_conn->in.datalen + tcp_conn->in.padding +
 555             (conn->datadgst_en ? 4 : 0))) {
 556                 debug_tcp("Copying header for later use. in.copy %d in.datalen"
 557                           " %d\n", tcp_conn->in.copy, tcp_conn->in.datalen);
 558                 memcpy(&tcp_conn->hdr, tcp_conn->in.hdr,
 559                        sizeof(struct iscsi_hdr));
 560                 tcp_conn->in.hdr = &tcp_conn->hdr;
 561                 tcp_conn->in.zero_copy_hdr = 0;
 562         }
 563         return 0;
 564 }
 565
 566 /**
 567  * iscsi_ctask_copy - copy skb bits to the destanation cmd task
 568  * @conn: iscsi tcp connection
 569  * @ctask: scsi command task
 570  * @buf: buffer to copy to
 571  * @buf_size: size of buffer
 572  * @offset: offset within the buffer
 573  *
 574  * Notes:
 575  *      The function calls skb_copy_bits() and updates per-connection and
 576  *      per-cmd byte counters.
 577  *
 578  *      Read counters (in bytes):
 579  *
 580  *      conn->in.offset         offset within in progress SKB
 581  *      conn->in.copy           left to copy from in progress SKB
 582  *                              including padding
 583  *      conn->in.copied         copied already from in progress SKB
 584  *      conn->data_copied       copied already from in progress buffer
 585  *      ctask->sent             total bytes sent up to the MidLayer
 586  *      ctask->data_count       left to copy from in progress Data-In
 587  *      buf_left                left to copy from in progress buffer
 588  **/
 589 static inline int
 590 iscsi_ctask_copy(struct iscsi_tcp_conn *tcp_conn, struct iscsi_cmd_task *ctask,
 591                 void *buf, int buf_size, int offset)
 592 {
 593         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 594         int buf_left = buf_size - (tcp_conn->data_copied + offset);
 595         int size = min(tcp_conn->in.copy, buf_left);
 596         int rc;
 597
 598         size = min(size, ctask->data_count);
 599
 600         debug_tcp("ctask_copy %d bytes at offset %d copied %d\n",
 601                size, tcp_conn->in.offset, tcp_conn->in.copied);
 602
 603         BUG_ON(size <= 0);
 604         BUG_ON(tcp_ctask->sent + size > ctask->total_length);
 605
 606         rc = skb_copy_bits(tcp_conn->in.skb, tcp_conn->in.offset,
 607                            (char*)buf + (offset + tcp_conn->data_copied), size);
 608         /* must fit into skb->len */
 609         BUG_ON(rc);
 610
 611         tcp_conn->in.offset += size;
 612         tcp_conn->in.copy -= size;
 613         tcp_conn->in.copied += size;
 614         tcp_conn->data_copied += size;
 615         tcp_ctask->sent += size;
 616         ctask->data_count -= size;
 617
 618         BUG_ON(tcp_conn->in.copy < 0);
 619         BUG_ON(ctask->data_count < 0);
 620
 621         if (buf_size != (tcp_conn->data_copied + offset)) {
 622                 if (!ctask->data_count) {
 623                         BUG_ON(buf_size - tcp_conn->data_copied < 0);
 624                         /* done with this PDU */
 625                         return buf_size - tcp_conn->data_copied;
 626                 }
 627                 return -EAGAIN;
 628         }
 629
 630         /* done with this buffer or with both - PDU and buffer */
 631         tcp_conn->data_copied = 0;
 632         return 0;
 633 }
 634
 635 /**
 636  * iscsi_tcp_copy - copy skb bits to the destanation buffer
 637  * @conn: iscsi tcp connection
 638  *
 639  * Notes:
 640  *      The function calls skb_copy_bits() and updates per-connection
 641  *      byte counters.
 642  **/
 643 static inline int
 644 iscsi_tcp_copy(struct iscsi_tcp_conn *tcp_conn)
 645 {
 646         void *buf = tcp_conn->data;
 647         int buf_size = tcp_conn->in.datalen;
 648         int buf_left = buf_size - tcp_conn->data_copied;
 649         int size = min(tcp_conn->in.copy, buf_left);
 650         int rc;
 651
 652         debug_tcp("tcp_copy %d bytes at offset %d copied %d\n",
 653                size, tcp_conn->in.offset, tcp_conn->data_copied);
 654         BUG_ON(size <= 0);
 655
 656         rc = skb_copy_bits(tcp_conn->in.skb, tcp_conn->in.offset,
 657                            (char*)buf + tcp_conn->data_copied, size);
 658         BUG_ON(rc);
 659
 660         tcp_conn->in.offset += size;
 661         tcp_conn->in.copy -= size;
 662         tcp_conn->in.copied += size;
 663         tcp_conn->data_copied += size;
 664
 665         if (buf_size != tcp_conn->data_copied)
 666                 return -EAGAIN;
 667
 668         return 0;
 669 }
 670
 671 static inline void
 672 partial_sg_digest_update(struct iscsi_tcp_conn *tcp_conn,
 673                          struct scatterlist *sg, int offset, int length)
 674 {
 675         struct scatterlist temp;
 676
 677         memcpy(&temp, sg, sizeof(struct scatterlist));
 678         temp.offset = offset;
 679         temp.length = length;
 680         crypto_digest_update(tcp_conn->data_rx_tfm, &temp, 1);
 681 }
 682
 683 static void
 684 iscsi_recv_digest_update(struct iscsi_tcp_conn *tcp_conn, char* buf, int len)
 685 {
 686         struct scatterlist tmp;
 687
 688         sg_init_one(&tmp, buf, len);
 689         crypto_digest_update(tcp_conn->data_rx_tfm, &tmp, 1);
 690 }
 691
 692 static int iscsi_scsi_data_in(struct iscsi_conn *conn)
 693 {
 694         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 695         struct iscsi_cmd_task *ctask = tcp_conn->in.ctask;
 696         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 697         struct scsi_cmnd *sc = ctask->sc;
 698         struct scatterlist *sg;
 699         int i, offset, rc = 0;
 700
 701         BUG_ON((void*)ctask != sc->SCp.ptr);
 702
 703         /*
 704          * copying Data-In into the Scsi_Cmnd
 705          */
 706         if (!sc->use_sg) {
 707                 i = ctask->data_count;
 708                 rc = iscsi_ctask_copy(tcp_conn, ctask, sc->request_buffer,
 709                                       sc->request_bufflen,
 710                                       tcp_ctask->data_offset);
 711                 if (rc == -EAGAIN)
 712                         return rc;
 713                 if (conn->datadgst_en)
 714                         iscsi_recv_digest_update(tcp_conn, sc->request_buffer,
 715                                                  i);
 716                 rc = 0;
 717                 goto done;
 718         }
 719
 720         offset = tcp_ctask->data_offset;
 721         sg = sc->request_buffer;
 722
 723         if (tcp_ctask->data_offset)
 724                 for (i = 0; i < tcp_ctask->sg_count; i++)
 725                         offset -= sg[i].length;
 726         /* we've passed through partial sg*/
 727         if (offset < 0)
 728                 offset = 0;
 729
 730         for (i = tcp_ctask->sg_count; i < sc->use_sg; i++) {
 731                 char *dest;
 732
 733                 dest = kmap_atomic(sg[i].page, KM_SOFTIRQ0);
 734                 rc = iscsi_ctask_copy(tcp_conn, ctask, dest + sg[i].offset,
 735                                       sg[i].length, offset);
 736                 kunmap_atomic(dest, KM_SOFTIRQ0);
 737                 if (rc == -EAGAIN)
 738                         /* continue with the next SKB/PDU */
 739                         return rc;
 740                 if (!rc) {
 741                         if (conn->datadgst_en) {
 742                                 if (!offset)
 743                                         crypto_digest_update(
 744                                                         tcp_conn->data_rx_tfm,
 745                                                         &sg[i], 1);
 746                                 else
 747                                         partial_sg_digest_update(tcp_conn,
 748                                                         &sg[i],
 749                                                         sg[i].offset + offset,
 750                                                         sg[i].length - offset);
 751                         }
 752                         offset = 0;
 753                         tcp_ctask->sg_count++;
 754                 }
 755
 756                 if (!ctask->data_count) {
 757                         if (rc && conn->datadgst_en)
 758                                 /*
 759                                  * data-in is complete, but buffer not...
 760                                  */
 761                                 partial_sg_digest_update(tcp_conn, &sg[i],
 762                                                 sg[i].offset, sg[i].length-rc);
 763                         rc = 0;
 764                         break;
 765                 }
 766
 767                 if (!tcp_conn->in.copy)
 768                         return -EAGAIN;
 769         }
 770         BUG_ON(ctask->data_count);
 771
 772 done:
 773         /* check for non-exceptional status */
 774         if (tcp_conn->in.hdr->flags & ISCSI_FLAG_DATA_STATUS) {
 775                 debug_scsi("done [sc %lx res %d itt 0x%x]\n",
 776                            (long)sc, sc->result, ctask->itt);
 777                 spin_lock(&conn->session->lock);
 778                 __iscsi_ctask_cleanup(conn, ctask);
 779                 __iscsi_complete_pdu(conn, tcp_conn->in.hdr, NULL, 0);
 780                 spin_unlock(&conn->session->lock);
 781         }
 782
 783         return rc;
 784 }
 785
 786 static int
 787 iscsi_data_recv(struct iscsi_conn *conn)
 788 {
 789         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 790         int rc = 0, opcode;
 791
 792         opcode = tcp_conn->in.hdr->opcode & ISCSI_OPCODE_MASK;
 793         switch (opcode) {
 794         case ISCSI_OP_SCSI_DATA_IN:
 795                 rc = iscsi_scsi_data_in(conn);
 796                 break;
 797         case ISCSI_OP_SCSI_CMD_RSP:
 798                 spin_lock(&conn->session->lock);
 799                 __iscsi_ctask_cleanup(conn, tcp_conn->in.ctask);
 800                 spin_unlock(&conn->session->lock);
 801         case ISCSI_OP_TEXT_RSP:
 802         case ISCSI_OP_LOGIN_RSP:
 803         case ISCSI_OP_NOOP_IN:
 804         case ISCSI_OP_ASYNC_EVENT:
 805         case ISCSI_OP_REJECT:
 806                 /*
 807                  * Collect data segment to the connection's data
 808                  * placeholder
 809                  */
 810                 if (iscsi_tcp_copy(tcp_conn)) {
 811                         rc = -EAGAIN;
 812                         goto exit;
 813                 }
 814
 815                 rc = iscsi_complete_pdu(conn, tcp_conn->in.hdr, tcp_conn->data,
 816                                         tcp_conn->in.datalen);
 817                 if (!rc && conn->datadgst_en && opcode != ISCSI_OP_LOGIN_RSP)
 818                         iscsi_recv_digest_update(tcp_conn, tcp_conn->data,
 819                                                 tcp_conn->in.datalen);
 820                 break;
 821         default:
 822                 BUG_ON(1);
 823         }
 824 exit:
 825         return rc;
 826 }
 827
 828 /**
 829  * iscsi_tcp_data_recv - TCP receive in sendfile fashion
 830  * @rd_desc: read descriptor
 831  * @skb: socket buffer
 832  * @offset: offset in skb
 833  * @len: skb->len - offset
 834  **/
 835 static int
 836 iscsi_tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 837                 unsigned int offset, size_t len)
 838 {
 839         int rc;
 840         struct iscsi_conn *conn = rd_desc->arg.data;
 841         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 842         int processed;
 843         char pad[ISCSI_PAD_LEN];
 844         struct scatterlist sg;
 845
 846         /*
 847          * Save current SKB and its offset in the corresponding
 848          * connection context.
 849          */
 850         tcp_conn->in.copy = skb->len - offset;
 851         tcp_conn->in.offset = offset;
 852         tcp_conn->in.skb = skb;
 853         tcp_conn->in.len = tcp_conn->in.copy;
 854         BUG_ON(tcp_conn->in.copy <= 0);
 855         debug_tcp("in %d bytes\n", tcp_conn->in.copy);
 856
 857 more:
 858         tcp_conn->in.copied = 0;
 859         rc = 0;
 860
 861         if (unlikely(conn->suspend_rx)) {
 862                 debug_tcp("conn %d Rx suspended!\n", conn->id);
 863                 return 0;
 864         }
 865
 866         if (tcp_conn->in_progress == IN_PROGRESS_WAIT_HEADER ||
 867             tcp_conn->in_progress == IN_PROGRESS_HEADER_GATHER) {
 868                 rc = iscsi_hdr_extract(tcp_conn);
 869                 if (rc) {
 870                        if (rc == -EAGAIN)
 871                                 goto nomore;
 872                        else {
 873                                 iscsi_conn_failure(conn, rc);
 874                                 return 0;
 875                        }
 876                 }
 877
 878                 /*
 879                  * Verify and process incoming PDU header.
 880                  */
 881                 rc = iscsi_tcp_hdr_recv(conn);
 882                 if (!rc && tcp_conn->in.datalen) {
 883                         if (conn->datadgst_en) {
 884                                 BUG_ON(!tcp_conn->data_rx_tfm);
 885                                 crypto_digest_init(tcp_conn->data_rx_tfm);
 886                         }
 887                         tcp_conn->in_progress = IN_PROGRESS_DATA_RECV;
 888                 } else if (rc) {
 889                         iscsi_conn_failure(conn, rc);
 890                         return 0;
 891                 }
 892         }
 893
 894         if (tcp_conn->in_progress == IN_PROGRESS_DDIGEST_RECV) {
 895                 uint32_t recv_digest;
 896
 897                 debug_tcp("extra data_recv offset %d copy %d\n",
 898                           tcp_conn->in.offset, tcp_conn->in.copy);
 899                 skb_copy_bits(tcp_conn->in.skb, tcp_conn->in.offset,
 900                                 &recv_digest, 4);
 901                 tcp_conn->in.offset += 4;
 902                 tcp_conn->in.copy -= 4;
 903                 if (recv_digest != tcp_conn->in.datadgst) {
 904                         debug_tcp("iscsi_tcp: data digest error!"
 905                                   "0x%x != 0x%x\n", recv_digest,
 906                                   tcp_conn->in.datadgst);
 907                         iscsi_conn_failure(conn, ISCSI_ERR_DATA_DGST);
 908                         return 0;
 909                 } else {
 910                         debug_tcp("iscsi_tcp: data digest match!"
 911                                   "0x%x == 0x%x\n", recv_digest,
 912                                   tcp_conn->in.datadgst);
 913                         tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
 914                 }
 915         }
 916
 917         if (tcp_conn->in_progress == IN_PROGRESS_DATA_RECV &&
 918            tcp_conn->in.copy) {
 919
 920                 debug_tcp("data_recv offset %d copy %d\n",
 921                        tcp_conn->in.offset, tcp_conn->in.copy);
 922
 923                 rc = iscsi_data_recv(conn);
 924                 if (rc) {
 925                         if (rc == -EAGAIN) {
 926                                 rd_desc->count = tcp_conn->in.datalen -
 927                                                 tcp_conn->in.ctask->data_count;
 928                                 goto again;
 929                         }
 930                         iscsi_conn_failure(conn, rc);
 931                         return 0;
 932                 }
 933                 tcp_conn->in.copy -= tcp_conn->in.padding;
 934                 tcp_conn->in.offset += tcp_conn->in.padding;
 935                 if (conn->datadgst_en) {
 936                         if (tcp_conn->in.padding) {
 937                                 debug_tcp("padding -> %d\n",
 938                                           tcp_conn->in.padding);
 939                                 memset(pad, 0, tcp_conn->in.padding);
 940                                 sg_init_one(&sg, pad, tcp_conn->in.padding);
 941                                 crypto_digest_update(tcp_conn->data_rx_tfm,
 942                                                      &sg, 1);
 943                         }
 944                         crypto_digest_final(tcp_conn->data_rx_tfm,
 945                                             (u8 *) & tcp_conn->in.datadgst);
 946                         debug_tcp("rx digest 0x%x\n", tcp_conn->in.datadgst);
 947                         tcp_conn->in_progress = IN_PROGRESS_DDIGEST_RECV;
 948                 } else
 949                         tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
 950         }
 951
 952         debug_tcp("f, processed %d from out of %d padding %d\n",
 953                tcp_conn->in.offset - offset, (int)len, tcp_conn->in.padding);
 954         BUG_ON(tcp_conn->in.offset - offset > len);
 955
 956         if (tcp_conn->in.offset - offset != len) {
 957                 debug_tcp("continue to process %d bytes\n",
 958                        (int)len - (tcp_conn->in.offset - offset));
 959                 goto more;
 960         }
 961
 962 nomore:
 963         processed = tcp_conn->in.offset - offset;
 964         BUG_ON(processed == 0);
 965         return processed;
 966
 967 again:
 968         processed = tcp_conn->in.offset - offset;
 969         debug_tcp("c, processed %d from out of %d rd_desc_cnt %d\n",
 970                   processed, (int)len, (int)rd_desc->count);
 971         BUG_ON(processed == 0);
 972         BUG_ON(processed > len);
 973
 974         conn->rxdata_octets += processed;
 975         return processed;
 976 }
 977
 978 static void
 979 iscsi_tcp_data_ready(struct sock *sk, int flag)
 980 {
 981         struct iscsi_conn *conn = sk->sk_user_data;
 982         read_descriptor_t rd_desc;
 983
 984         read_lock(&sk->sk_callback_lock);
 985
 986         /* use rd_desc to pass 'conn' to iscsi_tcp_data_recv */
 987         rd_desc.arg.data = conn;
 988         rd_desc.count = 0;
 989         tcp_read_sock(sk, &rd_desc, iscsi_tcp_data_recv);
 990
 991         read_unlock(&sk->sk_callback_lock);
 992 }
 993
 994 static void
 995 iscsi_tcp_state_change(struct sock *sk)
 996 {
 997         struct iscsi_tcp_conn *tcp_conn;
 998         struct iscsi_conn *conn;
 999         struct iscsi_session *session;
1000         void (*old_state_change)(struct sock *);
1001
1002         read_lock(&sk->sk_callback_lock);
1003
1004         conn = (struct iscsi_conn*)sk->sk_user_data;
1005         session = conn->session;
1006
1007         if ((sk->sk_state == TCP_CLOSE_WAIT ||
1008              sk->sk_state == TCP_CLOSE) &&
1009             !atomic_read(&sk->sk_rmem_alloc)) {
1010                 debug_tcp("iscsi_tcp_state_change: TCP_CLOSE|TCP_CLOSE_WAIT\n");
1011                 iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
1012         }
1013
1014         tcp_conn = conn->dd_data;
1015         old_state_change = tcp_conn->old_state_change;
1016
1017         read_unlock(&sk->sk_callback_lock);
1018
1019         old_state_change(sk);
1020 }
1021
1022 /**
1023  * iscsi_write_space - Called when more output buffer space is available
1024  * @sk: socket space is available for
1025  **/
1026 static void
1027 iscsi_write_space(struct sock *sk)
1028 {
1029         struct iscsi_conn *conn = (struct iscsi_conn*)sk->sk_user_data;
1030         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1031
1032         tcp_conn->old_write_space(sk);
1033         debug_tcp("iscsi_write_space: cid %d\n", conn->id);
1034         clear_bit(ISCSI_SUSPEND_BIT, &conn->suspend_tx);
1035         scsi_queue_work(conn->session->host, &conn->xmitwork);
1036 }
1037
1038 static void
1039 iscsi_conn_set_callbacks(struct iscsi_conn *conn)
1040 {
1041         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1042         struct sock *sk = tcp_conn->sock->sk;
1043
1044         /* assign new callbacks */
1045         write_lock_bh(&sk->sk_callback_lock);
1046         sk->sk_user_data = conn;
1047         tcp_conn->old_data_ready = sk->sk_data_ready;
1048         tcp_conn->old_state_change = sk->sk_state_change;
1049         tcp_conn->old_write_space = sk->sk_write_space;
1050         sk->sk_data_ready = iscsi_tcp_data_ready;
1051         sk->sk_state_change = iscsi_tcp_state_change;
1052         sk->sk_write_space = iscsi_write_space;
1053         write_unlock_bh(&sk->sk_callback_lock);
1054 }
1055
1056 static void
1057 iscsi_conn_restore_callbacks(struct iscsi_conn *conn)
1058 {
1059         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1060         struct sock *sk = tcp_conn->sock->sk;
1061
1062         /* restore socket callbacks, see also: iscsi_conn_set_callbacks() */
1063         write_lock_bh(&sk->sk_callback_lock);
1064         sk->sk_user_data    = NULL;
1065         sk->sk_data_ready   = tcp_conn->old_data_ready;
1066         sk->sk_state_change = tcp_conn->old_state_change;
1067         sk->sk_write_space  = tcp_conn->old_write_space;
1068         sk->sk_no_check  = 0;
1069         write_unlock_bh(&sk->sk_callback_lock);
1070 }
1071
1072 /**
1073  * iscsi_send - generic send routine
1074  * @sk: kernel's socket
1075  * @buf: buffer to write from
1076  * @size: actual size to write
1077  * @flags: socket's flags
1078  */
1079 static inline int
1080 iscsi_send(struct iscsi_conn *conn, struct iscsi_buf *buf, int size, int flags)
1081 {
1082         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1083         struct socket *sk = tcp_conn->sock;
1084         int offset = buf->sg.offset + buf->sent;
1085
1086         /*
1087          * if we got use_sg=0 or are sending something we kmallocd
1088          * then we did not have to do kmap (kmap returns page_address)
1089          *
1090          * if we got use_sg > 0, but had to drop down, we do not
1091          * set clustering so this should only happen for that
1092          * slab case.
1093          */
1094         if (buf->use_sendmsg)
1095                 return sock_no_sendpage(sk, buf->sg.page, offset, size, flags);
1096         else
1097                 return tcp_conn->sendpage(sk, buf->sg.page, offset, size,
1098                                           flags);
1099 }
1100
1101 /**
1102  * iscsi_sendhdr - send PDU Header via tcp_sendpage()
1103  * @conn: iscsi connection
1104  * @buf: buffer to write from
1105  * @datalen: lenght of data to be sent after the header
1106  *
1107  * Notes:
1108  *      (Tx, Fast Path)
1109  **/
1110 static inline int
1111 iscsi_sendhdr(struct iscsi_conn *conn, struct iscsi_buf *buf, int datalen)
1112 {
1113         struct iscsi_tcp_conn *tcp_conn;
1114         int flags = 0; /* MSG_DONTWAIT; */
1115         int res, size;
1116
1117         size = buf->sg.length - buf->sent;
1118         BUG_ON(buf->sent + size > buf->sg.length);
1119         if (buf->sent + size != buf->sg.length || datalen)
1120                 flags |= MSG_MORE;
1121
1122         res = iscsi_send(conn, buf, size, flags);
1123         debug_tcp("sendhdr %d bytes, sent %d res %d\n", size, buf->sent, res);
1124         if (res >= 0) {
1125                 conn->txdata_octets += res;
1126                 buf->sent += res;
1127                 if (size != res)
1128                         return -EAGAIN;
1129                 return 0;
1130         } else if (res == -EAGAIN) {
1131                 tcp_conn = conn->dd_data;
1132                 tcp_conn->sendpage_failures_cnt++;
1133                 set_bit(ISCSI_SUSPEND_BIT, &conn->suspend_tx);
1134         } else if (res == -EPIPE)
1135                 iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
1136
1137         return res;
1138 }
1139
1140 /**
1141  * iscsi_sendpage - send one page of iSCSI Data-Out.
1142  * @conn: iscsi connection
1143  * @buf: buffer to write from
1144  * @count: remaining data
1145  * @sent: number of bytes sent
1146  *
1147  * Notes:
1148  *      (Tx, Fast Path)
1149  **/
1150 static inline int
1151 iscsi_sendpage(struct iscsi_conn *conn, struct iscsi_buf *buf,
1152                int *count, int *sent)
1153 {
1154         struct iscsi_tcp_conn *tcp_conn;
1155         int flags = 0; /* MSG_DONTWAIT; */
1156         int res, size;
1157
1158         size = buf->sg.length - buf->sent;
1159         BUG_ON(buf->sent + size > buf->sg.length);
1160         if (size > *count)
1161                 size = *count;
1162         if (buf->sent + size != buf->sg.length || *count != size)
1163                 flags |= MSG_MORE;
1164
1165         res = iscsi_send(conn, buf, size, flags);
1166         debug_tcp("sendpage: %d bytes, sent %d left %d sent %d res %d\n",
1167                   size, buf->sent, *count, *sent, res);
1168         if (res >= 0) {
1169                 conn->txdata_octets += res;
1170                 buf->sent += res;
1171                 *count -= res;
1172                 *sent += res;
1173                 if (size != res)
1174                         return -EAGAIN;
1175                 return 0;
1176         } else if (res == -EAGAIN) {
1177                 tcp_conn = conn->dd_data;
1178                 tcp_conn->sendpage_failures_cnt++;
1179                 set_bit(ISCSI_SUSPEND_BIT, &conn->suspend_tx);
1180         } else if (res == -EPIPE)
1181                 iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
1182
1183         return res;
1184 }
1185
1186 static inline void
1187 iscsi_data_digest_init(struct iscsi_tcp_conn *tcp_conn,
1188                       struct iscsi_cmd_task *ctask)
1189 {
1190         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1191
1192         BUG_ON(!tcp_conn->data_tx_tfm);
1193         crypto_digest_init(tcp_conn->data_tx_tfm);
1194         tcp_ctask->digest_count = 4;
1195 }
1196
1197 static int
1198 iscsi_digest_final_send(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask,
1199                         struct iscsi_buf *buf, uint32_t *digest, int final)
1200 {
1201         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1202         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1203         int rc = 0;
1204         int sent = 0;
1205
1206         if (final)
1207                 crypto_digest_final(tcp_conn->data_tx_tfm, (u8*)digest);
1208
1209         iscsi_buf_init_virt(buf, (char*)digest, 4);
1210         rc = iscsi_sendpage(conn, buf, &tcp_ctask->digest_count, &sent);
1211         if (rc) {
1212                 tcp_ctask->datadigest = *digest;
1213                 tcp_ctask->xmstate |= XMSTATE_DATA_DIGEST;
1214         } else
1215                 tcp_ctask->digest_count = 4;
1216         return rc;
1217 }
1218
1219 /**
1220  * iscsi_solicit_data_cont - initialize next Data-Out
1221  * @conn: iscsi connection
1222  * @ctask: scsi command task
1223  * @r2t: R2T info
1224  * @left: bytes left to transfer
1225  *
1226  * Notes:
1227  *      Initialize next Data-Out within this R2T sequence and continue
1228  *      to process next Scatter-Gather element(if any) of this SCSI command.
1229  *
1230  *      Called under connection lock.
1231  **/
1232 static void
1233 iscsi_solicit_data_cont(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask,
1234                         struct iscsi_r2t_info *r2t, int left)
1235 {
1236         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1237         struct iscsi_data *hdr;
1238         struct iscsi_data_task *dtask;
1239         struct scsi_cmnd *sc = ctask->sc;
1240         int new_offset;
1241
1242         dtask = mempool_alloc(tcp_ctask->datapool, GFP_ATOMIC);
1243         BUG_ON(!dtask);
1244         INIT_LIST_HEAD(&dtask->item);
1245         hdr = &dtask->hdr;
1246         memset(hdr, 0, sizeof(struct iscsi_data));
1247         hdr->ttt = r2t->ttt;
1248         hdr->datasn = cpu_to_be32(r2t->solicit_datasn);
1249         r2t->solicit_datasn++;
1250         hdr->opcode = ISCSI_OP_SCSI_DATA_OUT;
1251         memcpy(hdr->lun, ctask->hdr->lun, sizeof(hdr->lun));
1252         hdr->itt = ctask->hdr->itt;
1253         hdr->exp_statsn = r2t->exp_statsn;
1254         new_offset = r2t->data_offset + r2t->sent;
1255         hdr->offset = cpu_to_be32(new_offset);
1256         if (left > conn->max_xmit_dlength) {
1257                 hton24(hdr->dlength, conn->max_xmit_dlength);
1258                 r2t->data_count = conn->max_xmit_dlength;
1259         } else {
1260                 hton24(hdr->dlength, left);
1261                 r2t->data_count = left;
1262                 hdr->flags = ISCSI_FLAG_CMD_FINAL;
1263         }
1264         conn->dataout_pdus_cnt++;
1265
1266         iscsi_buf_init_virt(&r2t->headbuf, (char*)hdr,
1267                            sizeof(struct iscsi_hdr));
1268
1269         r2t->dtask = dtask;
1270
1271         if (sc->use_sg && !iscsi_buf_left(&r2t->sendbuf)) {
1272                 BUG_ON(tcp_ctask->bad_sg == r2t->sg);
1273                 iscsi_buf_init_sg(&r2t->sendbuf, r2t->sg);
1274                 r2t->sg += 1;
1275         } else
1276                 iscsi_buf_init_iov(&tcp_ctask->sendbuf,
1277                             (char*)sc->request_buffer + new_offset,
1278                             r2t->data_count);
1279
1280         list_add(&dtask->item, &tcp_ctask->dataqueue);
1281 }
1282
1283 static void
1284 iscsi_unsolicit_data_init(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1285 {
1286         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1287         struct iscsi_data_task *dtask;
1288
1289         dtask = mempool_alloc(tcp_ctask->datapool, GFP_ATOMIC);
1290         BUG_ON(!dtask);
1291         INIT_LIST_HEAD(&dtask->item);
1292
1293         iscsi_prep_unsolicit_data_pdu(ctask, &dtask->hdr,
1294                                       tcp_ctask->r2t_data_count);
1295         iscsi_buf_init_virt(&tcp_ctask->headbuf, (char*)&dtask->hdr,
1296                            sizeof(struct iscsi_hdr));
1297
1298         list_add(&dtask->item, &tcp_ctask->dataqueue);
1299         tcp_ctask->dtask = dtask;
1300 }
1301
1302 /**
1303  * iscsi_tcp_cmd_init - Initialize iSCSI SCSI_READ or SCSI_WRITE commands
1304  * @conn: iscsi connection
1305  * @ctask: scsi command task
1306  * @sc: scsi command
1307  **/
1308 static void
1309 iscsi_tcp_cmd_init(struct iscsi_cmd_task *ctask)
1310 {
1311         struct scsi_cmnd *sc = ctask->sc;
1312         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1313
1314         BUG_ON(__kfifo_len(tcp_ctask->r2tqueue));
1315
1316         tcp_ctask->sent = 0;
1317         tcp_ctask->sg_count = 0;
1318
1319         if (sc->sc_data_direction == DMA_TO_DEVICE) {
1320                 tcp_ctask->xmstate = XMSTATE_W_HDR;
1321                 tcp_ctask->exp_r2tsn = 0;
1322                 BUG_ON(ctask->total_length == 0);
1323
1324                 if (sc->use_sg) {
1325                         struct scatterlist *sg = sc->request_buffer;
1326
1327                         iscsi_buf_init_sg(&tcp_ctask->sendbuf,
1328                                           &sg[tcp_ctask->sg_count++]);
1329                         tcp_ctask->sg = sg;
1330                         tcp_ctask->bad_sg = sg + sc->use_sg;
1331                 } else
1332                         iscsi_buf_init_iov(&tcp_ctask->sendbuf,
1333                                            sc->request_buffer,
1334                                            sc->request_bufflen);
1335
1336                 if (ctask->imm_count)
1337                         tcp_ctask->xmstate |= XMSTATE_IMM_DATA;
1338
1339                 tcp_ctask->pad_count = ctask->total_length & (ISCSI_PAD_LEN-1);
1340                 if (tcp_ctask->pad_count) {
1341                         tcp_ctask->pad_count = ISCSI_PAD_LEN -
1342                                                         tcp_ctask->pad_count;
1343                         debug_scsi("write padding %d bytes\n",
1344                                    tcp_ctask->pad_count);
1345                         tcp_ctask->xmstate |= XMSTATE_W_PAD;
1346                 }
1347
1348                 if (ctask->unsol_count)
1349                         tcp_ctask->xmstate |= XMSTATE_UNS_HDR |
1350                                                 XMSTATE_UNS_INIT;
1351                 tcp_ctask->r2t_data_count = ctask->total_length -
1352                                     ctask->imm_count -
1353                                     ctask->unsol_count;
1354
1355                 debug_scsi("cmd [itt %x total %d imm %d imm_data %d "
1356                            "r2t_data %d]\n",
1357                            ctask->itt, ctask->total_length, ctask->imm_count,
1358                            ctask->unsol_count, tcp_ctask->r2t_data_count);
1359         } else
1360                 tcp_ctask->xmstate = XMSTATE_R_HDR;
1361
1362         iscsi_buf_init_virt(&tcp_ctask->headbuf, (char*)ctask->hdr,
1363                             sizeof(struct iscsi_hdr));
1364 }
1365
1366 /**
1367  * iscsi_tcp_mtask_xmit - xmit management(immediate) task
1368  * @conn: iscsi connection
1369  * @mtask: task management task
1370  *
1371  * Notes:
1372  *      The function can return -EAGAIN in which case caller must
1373  *      call it again later, or recover. '0' return code means successful
1374  *      xmit.
1375  *
1376  *      Management xmit state machine consists of two states:
1377  *              IN_PROGRESS_IMM_HEAD - PDU Header xmit in progress
1378  *              IN_PROGRESS_IMM_DATA - PDU Data xmit in progress
1379  **/
1380 static int
1381 iscsi_tcp_mtask_xmit(struct iscsi_conn *conn, struct iscsi_mgmt_task *mtask)
1382 {
1383         struct iscsi_tcp_mgmt_task *tcp_mtask = mtask->dd_data;
1384
1385         debug_scsi("mtask deq [cid %d state %x itt 0x%x]\n",
1386                 conn->id, tcp_mtask->xmstate, mtask->itt);
1387
1388         if (tcp_mtask->xmstate & XMSTATE_IMM_HDR) {
1389                 tcp_mtask->xmstate &= ~XMSTATE_IMM_HDR;
1390                 if (mtask->data_count)
1391                         tcp_mtask->xmstate |= XMSTATE_IMM_DATA;
1392                 if (conn->c_stage != ISCSI_CONN_INITIAL_STAGE &&
1393                     conn->stop_stage != STOP_CONN_RECOVER &&
1394                     conn->hdrdgst_en)
1395                         iscsi_hdr_digest(conn, &tcp_mtask->headbuf,
1396                                         (u8*)tcp_mtask->hdrext);
1397                 if (iscsi_sendhdr(conn, &tcp_mtask->headbuf,
1398                                   mtask->data_count)) {
1399                         tcp_mtask->xmstate |= XMSTATE_IMM_HDR;
1400                         if (mtask->data_count)
1401                                 tcp_mtask->xmstate &= ~XMSTATE_IMM_DATA;
1402                         return -EAGAIN;
1403                 }
1404         }
1405
1406         if (tcp_mtask->xmstate & XMSTATE_IMM_DATA) {
1407                 BUG_ON(!mtask->data_count);
1408                 tcp_mtask->xmstate &= ~XMSTATE_IMM_DATA;
1409                 /* FIXME: implement.
1410                  * Virtual buffer could be spreaded across multiple pages...
1411                  */
1412                 do {
1413                         if (iscsi_sendpage(conn, &tcp_mtask->sendbuf,
1414                                    &mtask->data_count, &tcp_mtask->sent)) {
1415                                 tcp_mtask->xmstate |= XMSTATE_IMM_DATA;
1416                                 return -EAGAIN;
1417                         }
1418                 } while (mtask->data_count);
1419         }
1420
1421         BUG_ON(tcp_mtask->xmstate != XMSTATE_IDLE);
1422         if (mtask->hdr->itt == cpu_to_be32(ISCSI_RESERVED_TAG)) {
1423                 struct iscsi_session *session = conn->session;
1424
1425                 spin_lock_bh(&session->lock);
1426                 list_del(&conn->mtask->running);
1427                 __kfifo_put(session->mgmtpool.queue, (void*)&conn->mtask,
1428                             sizeof(void*));
1429                 spin_unlock_bh(&session->lock);
1430         }
1431         return 0;
1432 }
1433
1434 static inline int
1435 handle_xmstate_r_hdr(struct iscsi_conn *conn,
1436                      struct iscsi_tcp_cmd_task *tcp_ctask)
1437 {
1438         tcp_ctask->xmstate &= ~XMSTATE_R_HDR;
1439         if (conn->hdrdgst_en)
1440                 iscsi_hdr_digest(conn, &tcp_ctask->headbuf,
1441                                  (u8*)tcp_ctask->hdrext);
1442         if (!iscsi_sendhdr(conn, &tcp_ctask->headbuf, 0)) {
1443                 BUG_ON(tcp_ctask->xmstate != XMSTATE_IDLE);
1444                 return 0; /* wait for Data-In */
1445         }
1446         tcp_ctask->xmstate |= XMSTATE_R_HDR;
1447         return -EAGAIN;
1448 }
1449
1450 static inline int
1451 handle_xmstate_w_hdr(struct iscsi_conn *conn,
1452                      struct iscsi_cmd_task *ctask)
1453 {
1454         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1455
1456         tcp_ctask->xmstate &= ~XMSTATE_W_HDR;
1457         if (conn->hdrdgst_en)
1458                 iscsi_hdr_digest(conn, &tcp_ctask->headbuf,
1459                                  (u8*)tcp_ctask->hdrext);
1460         if (iscsi_sendhdr(conn, &tcp_ctask->headbuf, ctask->imm_count)) {
1461                 tcp_ctask->xmstate |= XMSTATE_W_HDR;
1462                 return -EAGAIN;
1463         }
1464         return 0;
1465 }
1466
1467 static inline int
1468 handle_xmstate_data_digest(struct iscsi_conn *conn,
1469                            struct iscsi_cmd_task *ctask)
1470 {
1471         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1472
1473         tcp_ctask->xmstate &= ~XMSTATE_DATA_DIGEST;
1474         debug_tcp("resent data digest 0x%x\n", tcp_ctask->datadigest);
1475         if (iscsi_digest_final_send(conn, ctask, &tcp_ctask->immbuf,
1476                                     &tcp_ctask->datadigest, 0)) {
1477                 tcp_ctask->xmstate |= XMSTATE_DATA_DIGEST;
1478                 debug_tcp("resent data digest 0x%x fail!\n",
1479                           tcp_ctask->datadigest);
1480                 return -EAGAIN;
1481         }
1482         return 0;
1483 }
1484
1485 static inline int
1486 handle_xmstate_imm_data(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1487 {
1488         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1489         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1490
1491         BUG_ON(!ctask->imm_count);
1492         tcp_ctask->xmstate &= ~XMSTATE_IMM_DATA;
1493
1494         if (conn->datadgst_en) {
1495                 iscsi_data_digest_init(tcp_conn, ctask);
1496                 tcp_ctask->immdigest = 0;
1497         }
1498
1499         for (;;) {
1500                 if (iscsi_sendpage(conn, &tcp_ctask->sendbuf, &ctask->imm_count,
1501                                    &tcp_ctask->sent)) {
1502                         tcp_ctask->xmstate |= XMSTATE_IMM_DATA;
1503                         if (conn->datadgst_en) {
1504                                 crypto_digest_final(tcp_conn->data_tx_tfm,
1505                                                 (u8*)&tcp_ctask->immdigest);
1506                                 debug_tcp("tx imm sendpage fail 0x%x\n",
1507                                           tcp_ctask->datadigest);
1508                         }
1509                         return -EAGAIN;
1510                 }
1511                 if (conn->datadgst_en)
1512                         crypto_digest_update(tcp_conn->data_tx_tfm,
1513                                              &tcp_ctask->sendbuf.sg, 1);
1514
1515                 if (!ctask->imm_count)
1516                         break;
1517                 iscsi_buf_init_sg(&tcp_ctask->sendbuf,
1518                                   &tcp_ctask->sg[tcp_ctask->sg_count++]);
1519         }
1520
1521         if (conn->datadgst_en && !(tcp_ctask->xmstate & XMSTATE_W_PAD)) {
1522                 if (iscsi_digest_final_send(conn, ctask, &tcp_ctask->immbuf,
1523                                             &tcp_ctask->immdigest, 1)) {
1524                         debug_tcp("sending imm digest 0x%x fail!\n",
1525                                   tcp_ctask->immdigest);
1526                         return -EAGAIN;
1527                 }
1528                 debug_tcp("sending imm digest 0x%x\n", tcp_ctask->immdigest);
1529         }
1530
1531         return 0;
1532 }
1533
1534 static inline int
1535 handle_xmstate_uns_hdr(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1536 {
1537         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1538         struct iscsi_data_task *dtask;
1539
1540         tcp_ctask->xmstate |= XMSTATE_UNS_DATA;
1541         if (tcp_ctask->xmstate & XMSTATE_UNS_INIT) {
1542                 iscsi_unsolicit_data_init(conn, ctask);
1543                 BUG_ON(!tcp_ctask->dtask);
1544                 dtask = tcp_ctask->dtask;
1545                 if (conn->hdrdgst_en)
1546                         iscsi_hdr_digest(conn, &tcp_ctask->headbuf,
1547                                         (u8*)dtask->hdrext);
1548                 tcp_ctask->xmstate &= ~XMSTATE_UNS_INIT;
1549         }
1550         if (iscsi_sendhdr(conn, &tcp_ctask->headbuf, ctask->data_count)) {
1551                 tcp_ctask->xmstate &= ~XMSTATE_UNS_DATA;
1552                 tcp_ctask->xmstate |= XMSTATE_UNS_HDR;
1553                 return -EAGAIN;
1554         }
1555
1556         debug_scsi("uns dout [itt 0x%x dlen %d sent %d]\n",
1557                    ctask->itt, ctask->unsol_count, tcp_ctask->sent);
1558         return 0;
1559 }
1560
1561 static inline int
1562 handle_xmstate_uns_data(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1563 {
1564         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1565         struct iscsi_data_task *dtask = tcp_ctask->dtask;
1566         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1567
1568         BUG_ON(!ctask->data_count);
1569         tcp_ctask->xmstate &= ~XMSTATE_UNS_DATA;
1570
1571         if (conn->datadgst_en) {
1572                 iscsi_data_digest_init(tcp_conn, ctask);
1573                 dtask->digest = 0;
1574         }
1575
1576         for (;;) {
1577                 int start = tcp_ctask->sent;
1578
1579                 if (iscsi_sendpage(conn, &tcp_ctask->sendbuf,
1580                                    &ctask->data_count, &tcp_ctask->sent)) {
1581                         ctask->unsol_count -= tcp_ctask->sent - start;
1582                         tcp_ctask->xmstate |= XMSTATE_UNS_DATA;
1583                         /* will continue with this ctask later.. */
1584                         if (conn->datadgst_en) {
1585                                 crypto_digest_final(tcp_conn->data_tx_tfm,
1586                                                 (u8 *)&dtask->digest);
1587                                 debug_tcp("tx uns data fail 0x%x\n",
1588                                           dtask->digest);
1589                         }
1590                         return -EAGAIN;
1591                 }
1592
1593                 BUG_ON(tcp_ctask->sent > ctask->total_length);
1594                 ctask->unsol_count -= tcp_ctask->sent - start;
1595
1596                 /*
1597                  * XXX:we may run here with un-initial sendbuf.
1598                  * so pass it
1599                  */
1600                 if (conn->datadgst_en && tcp_ctask->sent - start > 0)
1601                         crypto_digest_update(tcp_conn->data_tx_tfm,
1602                                              &tcp_ctask->sendbuf.sg, 1);
1603
1604                 if (!ctask->data_count)
1605                         break;
1606                 iscsi_buf_init_sg(&tcp_ctask->sendbuf,
1607                                   &tcp_ctask->sg[tcp_ctask->sg_count++]);
1608         }
1609         BUG_ON(ctask->unsol_count < 0);
1610
1611         /*
1612          * Done with the Data-Out. Next, check if we need
1613          * to send another unsolicited Data-Out.
1614          */
1615         if (ctask->unsol_count) {
1616                 if (conn->datadgst_en) {
1617                         if (iscsi_digest_final_send(conn, ctask,
1618                                                     &dtask->digestbuf,
1619                                                     &dtask->digest, 1)) {
1620                                 debug_tcp("send uns digest 0x%x fail\n",
1621                                           dtask->digest);
1622                                 return -EAGAIN;
1623                         }
1624                         debug_tcp("sending uns digest 0x%x, more uns\n",
1625                                   dtask->digest);
1626                 }
1627                 tcp_ctask->xmstate |= XMSTATE_UNS_INIT;
1628                 return 1;
1629         }
1630
1631         if (conn->datadgst_en && !(tcp_ctask->xmstate & XMSTATE_W_PAD)) {
1632                 if (iscsi_digest_final_send(conn, ctask,
1633                                             &dtask->digestbuf,
1634                                             &dtask->digest, 1)) {
1635                         debug_tcp("send last uns digest 0x%x fail\n",
1636                                    dtask->digest);
1637                         return -EAGAIN;
1638                 }
1639                 debug_tcp("sending uns digest 0x%x\n",dtask->digest);
1640         }
1641
1642         return 0;
1643 }
1644
1645 static inline int
1646 handle_xmstate_sol_data(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1647 {
1648         struct iscsi_session *session = conn->session;
1649         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1650         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1651         struct iscsi_r2t_info *r2t = tcp_ctask->r2t;
1652         struct iscsi_data_task *dtask = r2t->dtask;
1653         int left;
1654
1655         tcp_ctask->xmstate &= ~XMSTATE_SOL_DATA;
1656         tcp_ctask->dtask = dtask;
1657
1658         if (conn->datadgst_en) {
1659                 iscsi_data_digest_init(tcp_conn, ctask);
1660                 dtask->digest = 0;
1661         }
1662 solicit_again:
1663         /*
1664          * send Data-Out whitnin this R2T sequence.
1665          */
1666         if (!r2t->data_count)
1667                 goto data_out_done;
1668
1669         if (iscsi_sendpage(conn, &r2t->sendbuf, &r2t->data_count, &r2t->sent)) {
1670                 tcp_ctask->xmstate |= XMSTATE_SOL_DATA;
1671                 /* will continue with this ctask later.. */
1672                 if (conn->datadgst_en) {
1673                         crypto_digest_final(tcp_conn->data_tx_tfm,
1674                                           (u8 *)&dtask->digest);
1675                         debug_tcp("r2t data send fail 0x%x\n", dtask->digest);
1676                 }
1677                 return -EAGAIN;
1678         }
1679
1680         BUG_ON(r2t->data_count < 0);
1681         if (conn->datadgst_en)
1682                 crypto_digest_update(tcp_conn->data_tx_tfm, &r2t->sendbuf.sg,
1683                                      1);
1684
1685         if (r2t->data_count) {
1686                 BUG_ON(ctask->sc->use_sg == 0);
1687                 if (!iscsi_buf_left(&r2t->sendbuf)) {
1688                         BUG_ON(tcp_ctask->bad_sg == r2t->sg);
1689                         iscsi_buf_init_sg(&r2t->sendbuf, r2t->sg);
1690                         r2t->sg += 1;
1691                 }
1692                 goto solicit_again;
1693         }
1694
1695 data_out_done:
1696         /*
1697          * Done with this Data-Out. Next, check if we have
1698          * to send another Data-Out for this R2T.
1699          */
1700         BUG_ON(r2t->data_length - r2t->sent < 0);
1701         left = r2t->data_length - r2t->sent;
1702         if (left) {
1703                 if (conn->datadgst_en) {
1704                         if (iscsi_digest_final_send(conn, ctask,
1705                                                     &dtask->digestbuf,
1706                                                     &dtask->digest, 1)) {
1707                                 debug_tcp("send r2t data digest 0x%x"
1708                                           "fail\n", dtask->digest);
1709                                 return -EAGAIN;
1710                         }
1711                         debug_tcp("r2t data send digest 0x%x\n",
1712                                   dtask->digest);
1713                 }
1714                 iscsi_solicit_data_cont(conn, ctask, r2t, left);
1715                 tcp_ctask->xmstate |= XMSTATE_SOL_DATA;
1716                 tcp_ctask->xmstate &= ~XMSTATE_SOL_HDR;
1717                 return 1;
1718         }
1719
1720         /*
1721          * Done with this R2T. Check if there are more
1722          * outstanding R2Ts ready to be processed.
1723          */
1724         BUG_ON(tcp_ctask->r2t_data_count - r2t->data_length < 0);
1725         if (conn->datadgst_en) {
1726                 if (iscsi_digest_final_send(conn, ctask, &dtask->digestbuf,
1727                                             &dtask->digest, 1)) {
1728                         debug_tcp("send last r2t data digest 0x%x"
1729                                   "fail\n", dtask->digest);
1730                         return -EAGAIN;
1731                 }
1732                 debug_tcp("r2t done dout digest 0x%x\n", dtask->digest);
1733         }
1734
1735         tcp_ctask->r2t_data_count -= r2t->data_length;
1736         tcp_ctask->r2t = NULL;
1737         spin_lock_bh(&session->lock);
1738         __kfifo_put(tcp_ctask->r2tpool.queue, (void*)&r2t, sizeof(void*));
1739         spin_unlock_bh(&session->lock);
1740         if (__kfifo_get(tcp_ctask->r2tqueue, (void*)&r2t, sizeof(void*))) {
1741                 tcp_ctask->r2t = r2t;
1742                 tcp_ctask->xmstate |= XMSTATE_SOL_DATA;
1743                 tcp_ctask->xmstate &= ~XMSTATE_SOL_HDR;
1744                 return 1;
1745         }
1746
1747         return 0;
1748 }
1749
1750 static inline int
1751 handle_xmstate_w_pad(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1752 {
1753         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1754         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1755         struct iscsi_data_task *dtask = tcp_ctask->dtask;
1756         int sent;
1757
1758         tcp_ctask->xmstate &= ~XMSTATE_W_PAD;
1759         iscsi_buf_init_virt(&tcp_ctask->sendbuf, (char*)&tcp_ctask->pad,
1760                             tcp_ctask->pad_count);
1761         if (iscsi_sendpage(conn, &tcp_ctask->sendbuf, &tcp_ctask->pad_count,
1762                            &sent)) {
1763                 tcp_ctask->xmstate |= XMSTATE_W_PAD;
1764                 return -EAGAIN;
1765         }
1766
1767         if (conn->datadgst_en) {
1768                 crypto_digest_update(tcp_conn->data_tx_tfm,
1769                                      &tcp_ctask->sendbuf.sg, 1);
1770                 /* imm data? */
1771                 if (!dtask) {
1772                         if (iscsi_digest_final_send(conn, ctask,
1773                                                     &tcp_ctask->immbuf,
1774                                                     &tcp_ctask->immdigest, 1)) {
1775                                 debug_tcp("send padding digest 0x%x"
1776                                           "fail!\n", tcp_ctask->immdigest);
1777                                 return -EAGAIN;
1778                         }
1779                         debug_tcp("done with padding, digest 0x%x\n",
1780                                   tcp_ctask->datadigest);
1781                 } else {
1782                         if (iscsi_digest_final_send(conn, ctask,
1783                                                     &dtask->digestbuf,
1784                                                     &dtask->digest, 1)) {
1785                                 debug_tcp("send padding digest 0x%x"
1786                                           "fail\n", dtask->digest);
1787                                 return -EAGAIN;
1788                         }
1789                         debug_tcp("done with padding, digest 0x%x\n",
1790                                   dtask->digest);
1791                 }
1792         }
1793
1794         return 0;
1795 }
1796
1797 static int
1798 iscsi_tcp_ctask_xmit(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1799 {
1800         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1801         int rc = 0;
1802
1803         debug_scsi("ctask deq [cid %d xmstate %x itt 0x%x]\n",
1804                 conn->id, tcp_ctask->xmstate, ctask->itt);
1805
1806         /*
1807          * serialize with TMF AbortTask
1808          */
1809         if (ctask->mtask)
1810                 return rc;
1811
1812         if (tcp_ctask->xmstate & XMSTATE_R_HDR) {
1813                 rc = handle_xmstate_r_hdr(conn, tcp_ctask);
1814                 return rc;
1815         }
1816
1817         if (tcp_ctask->xmstate & XMSTATE_W_HDR) {
1818                 rc = handle_xmstate_w_hdr(conn, ctask);
1819                 if (rc)
1820                         return rc;
1821         }
1822
1823         /* XXX: for data digest xmit recover */
1824         if (tcp_ctask->xmstate & XMSTATE_DATA_DIGEST) {
1825                 rc = handle_xmstate_data_digest(conn, ctask);
1826                 if (rc)
1827                         return rc;
1828         }
1829
1830         if (tcp_ctask->xmstate & XMSTATE_IMM_DATA) {
1831                 rc = handle_xmstate_imm_data(conn, ctask);
1832                 if (rc)
1833                         return rc;
1834         }
1835
1836         if (tcp_ctask->xmstate & XMSTATE_UNS_HDR) {
1837                 BUG_ON(!ctask->unsol_count);
1838                 tcp_ctask->xmstate &= ~XMSTATE_UNS_HDR;
1839 unsolicit_head_again:
1840                 rc = handle_xmstate_uns_hdr(conn, ctask);
1841                 if (rc)
1842                         return rc;
1843         }
1844
1845         if (tcp_ctask->xmstate & XMSTATE_UNS_DATA) {
1846                 rc = handle_xmstate_uns_data(conn, ctask);
1847                 if (rc == 1)
1848                         goto unsolicit_head_again;
1849                 else if (rc)
1850                         return rc;
1851                 goto done;
1852         }
1853
1854         if (tcp_ctask->xmstate & XMSTATE_SOL_HDR) {
1855                 struct iscsi_r2t_info *r2t;
1856
1857                 tcp_ctask->xmstate &= ~XMSTATE_SOL_HDR;
1858                 tcp_ctask->xmstate |= XMSTATE_SOL_DATA;
1859                 if (!tcp_ctask->r2t)
1860                         __kfifo_get(tcp_ctask->r2tqueue, (void*)&tcp_ctask->r2t,
1861                                     sizeof(void*));
1862 solicit_head_again:
1863                 r2t = tcp_ctask->r2t;
1864                 if (conn->hdrdgst_en)
1865                         iscsi_hdr_digest(conn, &r2t->headbuf,
1866                                         (u8*)r2t->dtask->hdrext);
1867                 if (iscsi_sendhdr(conn, &r2t->headbuf, r2t->data_count)) {
1868                         tcp_ctask->xmstate &= ~XMSTATE_SOL_DATA;
1869                         tcp_ctask->xmstate |= XMSTATE_SOL_HDR;
1870                         return -EAGAIN;
1871                 }
1872
1873                 debug_scsi("sol dout [dsn %d itt 0x%x dlen %d sent %d]\n",
1874                         r2t->solicit_datasn - 1, ctask->itt, r2t->data_count,
1875                         r2t->sent);
1876         }
1877
1878         if (tcp_ctask->xmstate & XMSTATE_SOL_DATA) {
1879                 rc = handle_xmstate_sol_data(conn, ctask);
1880                 if (rc == 1)
1881                         goto solicit_head_again;
1882                 if (rc)
1883                         return rc;
1884         }
1885
1886 done:
1887         /*
1888          * Last thing to check is whether we need to send write
1889          * padding. Note that we check for xmstate equality, not just the bit.
1890          */
1891         if (tcp_ctask->xmstate == XMSTATE_W_PAD)
1892                 rc = handle_xmstate_w_pad(conn, ctask);
1893
1894         return rc;
1895 }
1896
1897 static struct iscsi_cls_conn *
1898 iscsi_tcp_conn_create(struct iscsi_cls_session *cls_session, uint32_t conn_idx)
1899 {
1900         struct iscsi_conn *conn;
1901         struct iscsi_cls_conn *cls_conn;
1902         struct iscsi_tcp_conn *tcp_conn;
1903
1904         cls_conn = iscsi_conn_setup(cls_session, conn_idx);
1905         if (!cls_conn)
1906                 return NULL;
1907         conn = cls_conn->dd_data;
1908         /*
1909          * due to strange issues with iser these are not set
1910          * in iscsi_conn_setup
1911          */
1912         conn->max_recv_dlength = DEFAULT_MAX_RECV_DATA_SEGMENT_LENGTH;
1913
1914         tcp_conn = kzalloc(sizeof(*tcp_conn), GFP_KERNEL);
1915         if (!tcp_conn)
1916                 goto tcp_conn_alloc_fail;
1917
1918         conn->dd_data = tcp_conn;
1919         tcp_conn->iscsi_conn = conn;
1920         tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
1921         /* initial operational parameters */
1922         tcp_conn->hdr_size = sizeof(struct iscsi_hdr);
1923         tcp_conn->data_size = DEFAULT_MAX_RECV_DATA_SEGMENT_LENGTH;
1924
1925         /* allocate initial PDU receive place holder */
1926         if (tcp_conn->data_size <= PAGE_SIZE)
1927                 tcp_conn->data = kmalloc(tcp_conn->data_size, GFP_KERNEL);
1928         else
1929                 tcp_conn->data = (void*)__get_free_pages(GFP_KERNEL,
1930                                         get_order(tcp_conn->data_size));
1931         if (!tcp_conn->data)
1932                 goto max_recv_dlenght_alloc_fail;
1933
1934         return cls_conn;
1935
1936 max_recv_dlenght_alloc_fail:
1937         kfree(tcp_conn);
1938 tcp_conn_alloc_fail:
1939         iscsi_conn_teardown(cls_conn);
1940         return NULL;
1941 }
1942
1943 static void
1944 iscsi_tcp_conn_destroy(struct iscsi_cls_conn *cls_conn)
1945 {
1946         struct iscsi_conn *conn = cls_conn->dd_data;
1947         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1948         int digest = 0;
1949
1950         if (conn->hdrdgst_en || conn->datadgst_en)
1951                 digest = 1;
1952
1953         iscsi_conn_teardown(cls_conn);
1954
1955         /* now free tcp_conn */
1956         if (digest) {
1957                 if (tcp_conn->tx_tfm)
1958                         crypto_free_tfm(tcp_conn->tx_tfm);
1959                 if (tcp_conn->rx_tfm)
1960                         crypto_free_tfm(tcp_conn->rx_tfm);
1961                 if (tcp_conn->data_tx_tfm)
1962                         crypto_free_tfm(tcp_conn->data_tx_tfm);
1963                 if (tcp_conn->data_rx_tfm)
1964                         crypto_free_tfm(tcp_conn->data_rx_tfm);
1965         }
1966
1967         /* free conn->data, size = MaxRecvDataSegmentLength */
1968         if (tcp_conn->data_size <= PAGE_SIZE)
1969                 kfree(tcp_conn->data);
1970         else
1971                 free_pages((unsigned long)tcp_conn->data,
1972                            get_order(tcp_conn->data_size));
1973         kfree(tcp_conn);
1974 }
1975
1976 static int
1977 iscsi_tcp_conn_bind(struct iscsi_cls_session *cls_session,
1978                     struct iscsi_cls_conn *cls_conn, uint32_t transport_fd,
1979                     int is_leading)
1980 {
1981         struct iscsi_conn *conn = cls_conn->dd_data;
1982         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1983         struct sock *sk;
1984         struct socket *sock;
1985         int err;
1986
1987         /* lookup for existing socket */
1988         sock = sockfd_lookup(transport_fd, &err);
1989         if (!sock) {
1990                 printk(KERN_ERR "iscsi_tcp: sockfd_lookup failed %d\n", err);
1991                 return -EEXIST;
1992         }
1993
1994         err = iscsi_conn_bind(cls_session, cls_conn, is_leading);
1995         if (err)
1996                 return err;
1997
1998         if (conn->stop_stage != STOP_CONN_SUSPEND) {
1999                 /* bind iSCSI connection and socket */
2000                 tcp_conn->sock = sock;
2001
2002                 /* setup Socket parameters */
2003                 sk = sock->sk;
2004                 sk->sk_reuse = 1;
2005                 sk->sk_sndtimeo = 15 * HZ; /* FIXME: make it configurable */
2006                 sk->sk_allocation = GFP_ATOMIC;
2007
2008                 /* FIXME: disable Nagle's algorithm */
2009
2010                 /*
2011                  * Intercept TCP callbacks for sendfile like receive
2012                  * processing.
2013                  */
2014                 conn->recv_lock = &sk->sk_callback_lock;
2015                 iscsi_conn_set_callbacks(conn);
2016                 tcp_conn->sendpage = tcp_conn->sock->ops->sendpage;
2017                 /*
2018                  * set receive state machine into initial state
2019                  */
2020                 tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
2021         }
2022
2023         return 0;
2024 }
2025
2026 static void
2027 iscsi_tcp_cleanup_ctask(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
2028 {
2029         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
2030         struct iscsi_r2t_info *r2t;
2031
2032         /* flush ctask's r2t queues */
2033         while (__kfifo_get(tcp_ctask->r2tqueue, (void*)&r2t, sizeof(void*)))
2034                 __kfifo_put(tcp_ctask->r2tpool.queue, (void*)&r2t,
2035                             sizeof(void*));
2036
2037         __iscsi_ctask_cleanup(conn, ctask);
2038 }
2039
2040 static void
2041 iscsi_tcp_suspend_conn_rx(struct iscsi_conn *conn)
2042 {
2043         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
2044         struct sock *sk;
2045
2046         if (!tcp_conn->sock)
2047                 return;
2048
2049         sk = tcp_conn->sock->sk;
2050         write_lock_bh(&sk->sk_callback_lock);
2051         set_bit(ISCSI_SUSPEND_BIT, &conn->suspend_rx);
2052         write_unlock_bh(&sk->sk_callback_lock);
2053 }
2054
2055 static void
2056 iscsi_tcp_terminate_conn(struct iscsi_conn *conn)
2057 {
2058         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
2059
2060         if (!tcp_conn->sock)
2061                 return;
2062
2063         sock_hold(tcp_conn->sock->sk);
2064         iscsi_conn_restore_callbacks(conn);
2065         sock_put(tcp_conn->sock->sk);
2066
2067         sock_release(tcp_conn->sock);
2068         tcp_conn->sock = NULL;
2069         conn->recv_lock = NULL;
2070 }
2071
2072 /* called with host lock */
2073 static void
2074 iscsi_tcp_mgmt_init(struct iscsi_conn *conn, struct iscsi_mgmt_task *mtask,
2075                     char *data, uint32_t data_size)
2076 {
2077         struct iscsi_tcp_mgmt_task *tcp_mtask = mtask->dd_data;
2078
2079         iscsi_buf_init_virt(&tcp_mtask->headbuf, (char*)mtask->hdr,
2080                                     sizeof(struct iscsi_hdr));
2081         tcp_mtask->xmstate = XMSTATE_IMM_HDR;
2082
2083         if (mtask->data_count)
2084                 iscsi_buf_init_iov(&tcp_mtask->sendbuf, (char*)mtask->data,
2085                                     mtask->data_count);
2086 }
2087
2088 static int
2089 iscsi_r2tpool_alloc(struct iscsi_session *session)
2090 {
2091         int i;
2092         int cmd_i;
2093
2094         /*
2095          * initialize per-task: R2T pool and xmit queue
2096          */
2097         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
2098                 struct iscsi_cmd_task *ctask = session->cmds[cmd_i];
2099                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
2100
2101                 /*
2102                  * pre-allocated x4 as much r2ts to handle race when
2103                  * target acks DataOut faster than we data_xmit() queues
2104                  * could replenish r2tqueue.
2105                  */
2106
2107                 /* R2T pool */
2108                 if (iscsi_pool_init(&tcp_ctask->r2tpool, session->max_r2t * 4,
2109                                     (void***)&tcp_ctask->r2ts,
2110                                     sizeof(struct iscsi_r2t_info))) {
2111                         goto r2t_alloc_fail;
2112                 }
2113
2114                 /* R2T xmit queue */
2115                 tcp_ctask->r2tqueue = kfifo_alloc(
2116                       session->max_r2t * 4 * sizeof(void*), GFP_KERNEL, NULL);
2117                 if (tcp_ctask->r2tqueue == ERR_PTR(-ENOMEM)) {
2118                         iscsi_pool_free(&tcp_ctask->r2tpool,
2119                                         (void**)tcp_ctask->r2ts);
2120                         goto r2t_alloc_fail;
2121                 }
2122
2123                 /*
2124                  * number of
2125                  * Data-Out PDU's within R2T-sequence can be quite big;
2126                  * using mempool
2127                  */
2128                 tcp_ctask->datapool = mempool_create_slab_pool(ISCSI_DTASK_DEFAULT_MAX,
2129                                                                taskcache);
2130                 if (tcp_ctask->datapool == NULL) {
2131                         kfifo_free(tcp_ctask->r2tqueue);
2132                         iscsi_pool_free(&tcp_ctask->r2tpool,
2133                                         (void**)tcp_ctask->r2ts);
2134                         goto r2t_alloc_fail;
2135                 }
2136                 INIT_LIST_HEAD(&tcp_ctask->dataqueue);
2137         }
2138
2139         return 0;
2140
2141 r2t_alloc_fail:
2142         for (i = 0; i < cmd_i; i++) {
2143                 struct iscsi_cmd_task *ctask = session->cmds[i];
2144                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
2145
2146                 mempool_destroy(tcp_ctask->datapool);
2147                 kfifo_free(tcp_ctask->r2tqueue);
2148                 iscsi_pool_free(&tcp_ctask->r2tpool,
2149                                 (void**)tcp_ctask->r2ts);
2150         }
2151         return -ENOMEM;
2152 }
2153
2154 static void
2155 iscsi_r2tpool_free(struct iscsi_session *session)
2156 {
2157         int i;
2158
2159         for (i = 0; i < session->cmds_max; i++) {
2160                 struct iscsi_cmd_task *ctask = session->cmds[i];
2161                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
2162
2163                 mempool_destroy(tcp_ctask->datapool);
2164                 kfifo_free(tcp_ctask->r2tqueue);
2165                 iscsi_pool_free(&tcp_ctask->r2tpool,
2166                                 (void**)tcp_ctask->r2ts);
2167         }
2168 }
2169
2170 static int
2171 iscsi_conn_set_param(struct iscsi_cls_conn *cls_conn, enum iscsi_param param,
2172                      uint32_t value)
2173 {
2174         struct iscsi_conn *conn = cls_conn->dd_data;
2175         struct iscsi_session *session = conn->session;
2176         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
2177
2178         switch(param) {
2179         case ISCSI_PARAM_MAX_RECV_DLENGTH: {
2180                 char *saveptr = tcp_conn->data;
2181                 gfp_t flags = GFP_KERNEL;
2182
2183                 if (tcp_conn->data_size >= value) {
2184                         conn->max_recv_dlength = value;
2185                         break;
2186                 }
2187
2188                 spin_lock_bh(&session->lock);
2189                 if (conn->stop_stage == STOP_CONN_RECOVER)
2190                         flags = GFP_ATOMIC;
2191                 spin_unlock_bh(&session->lock);
2192
2193                 if (value <= PAGE_SIZE)
2194                         tcp_conn->data = kmalloc(value, flags);
2195                 else
2196                         tcp_conn->data = (void*)__get_free_pages(flags,
2197                                                              get_order(value));
2198                 if (tcp_conn->data == NULL) {
2199                         tcp_conn->data = saveptr;
2200                         return -ENOMEM;
2201                 }
2202                 if (tcp_conn->data_size <= PAGE_SIZE)
2203                         kfree(saveptr);
2204                 else
2205                         free_pages((unsigned long)saveptr,
2206                                    get_order(tcp_conn->data_size));
2207                 conn->max_recv_dlength = value;
2208                 tcp_conn->data_size = value;
2209                 }
2210                 break;
2211         case ISCSI_PARAM_MAX_XMIT_DLENGTH:
2212                 conn->max_xmit_dlength =  value;
2213                 break;
2214         case ISCSI_PARAM_HDRDGST_EN:
2215                 conn->hdrdgst_en = value;
2216                 tcp_conn->hdr_size = sizeof(struct iscsi_hdr);
2217                 if (conn->hdrdgst_en) {
2218                         tcp_conn->hdr_size += sizeof(__u32);
2219                         if (!tcp_conn->tx_tfm)
2220                                 tcp_conn->tx_tfm = crypto_alloc_tfm("crc32c",
2221                                                                     0);
2222                         if (!tcp_conn->tx_tfm)
2223                                 return -ENOMEM;
2224                         if (!tcp_conn->rx_tfm)
2225                                 tcp_conn->rx_tfm = crypto_alloc_tfm("crc32c",
2226                                                                     0);
2227                         if (!tcp_conn->rx_tfm) {
2228                                 crypto_free_tfm(tcp_conn->tx_tfm);
2229                                 return -ENOMEM;
2230                         }
2231                 } else {
2232                         if (tcp_conn->tx_tfm)
2233                                 crypto_free_tfm(tcp_conn->tx_tfm);
2234                         if (tcp_conn->rx_tfm)
2235                                 crypto_free_tfm(tcp_conn->rx_tfm);
2236                 }
2237                 break;
2238         case ISCSI_PARAM_DATADGST_EN:
2239                 conn->datadgst_en = value;
2240                 if (conn->datadgst_en) {
2241                         if (!tcp_conn->data_tx_tfm)
2242                                 tcp_conn->data_tx_tfm =
2243                                     crypto_alloc_tfm("crc32c", 0);
2244                         if (!tcp_conn->data_tx_tfm)
2245                                 return -ENOMEM;
2246                         if (!tcp_conn->data_rx_tfm)
2247                                 tcp_conn->data_rx_tfm =
2248                                     crypto_alloc_tfm("crc32c", 0);
2249                         if (!tcp_conn->data_rx_tfm) {
2250                                 crypto_free_tfm(tcp_conn->data_tx_tfm);
2251                                 return -ENOMEM;
2252                         }
2253                 } else {
2254                         if (tcp_conn->data_tx_tfm)
2255                                 crypto_free_tfm(tcp_conn->data_tx_tfm);
2256                         if (tcp_conn->data_rx_tfm)
2257                                 crypto_free_tfm(tcp_conn->data_rx_tfm);
2258                 }
2259                 tcp_conn->sendpage = conn->datadgst_en ?
2260                         sock_no_sendpage : tcp_conn->sock->ops->sendpage;
2261                 break;
2262         case ISCSI_PARAM_INITIAL_R2T_EN:
2263                 session->initial_r2t_en = value;
2264                 break;
2265         case ISCSI_PARAM_MAX_R2T:
2266                 if (session->max_r2t == roundup_pow_of_two(value))
2267                         break;
2268                 iscsi_r2tpool_free(session);
2269                 session->max_r2t = value;
2270                 if (session->max_r2t & (session->max_r2t - 1))
2271                         session->max_r2t = roundup_pow_of_two(session->max_r2t);
2272                 if (iscsi_r2tpool_alloc(session))
2273                         return -ENOMEM;
2274                 break;
2275         case ISCSI_PARAM_IMM_DATA_EN:
2276                 session->imm_data_en = value;
2277                 break;
2278         case ISCSI_PARAM_FIRST_BURST:
2279                 session->first_burst = value;
2280                 break;
2281         case ISCSI_PARAM_MAX_BURST:
2282                 session->max_burst = value;
2283                 break;
2284         case ISCSI_PARAM_PDU_INORDER_EN:
2285                 session->pdu_inorder_en = value;
2286                 break;
2287         case ISCSI_PARAM_DATASEQ_INORDER_EN:
2288                 session->dataseq_inorder_en = value;
2289                 break;
2290         case ISCSI_PARAM_ERL:
2291                 session->erl = value;
2292                 break;
2293         case ISCSI_PARAM_IFMARKER_EN:
2294                 BUG_ON(value);
2295                 session->ifmarker_en = value;
2296                 break;
2297         case ISCSI_PARAM_OFMARKER_EN:
2298                 BUG_ON(value);
2299                 session->ofmarker_en = value;
2300                 break;
2301         default:
2302                 break;
2303         }
2304
2305         return 0;
2306 }
2307
2308 static int
2309 iscsi_session_get_param(struct iscsi_cls_session *cls_session,
2310                         enum iscsi_param param, uint32_t *value)
2311 {
2312         struct Scsi_Host *shost = iscsi_session_to_shost(cls_session);
2313         struct iscsi_session *session = iscsi_hostdata(shost->hostdata);
2314
2315         switch(param) {
2316         case ISCSI_PARAM_INITIAL_R2T_EN:
2317                 *value = session->initial_r2t_en;
2318                 break;
2319         case ISCSI_PARAM_MAX_R2T:
2320                 *value = session->max_r2t;
2321                 break;
2322         case ISCSI_PARAM_IMM_DATA_EN:
2323                 *value = session->imm_data_en;
2324                 break;
2325         case ISCSI_PARAM_FIRST_BURST:
2326                 *value = session->first_burst;
2327                 break;
2328         case ISCSI_PARAM_MAX_BURST:
2329                 *value = session->max_burst;
2330                 break;
2331         case ISCSI_PARAM_PDU_INORDER_EN:
2332                 *value = session->pdu_inorder_en;
2333                 break;
2334         case ISCSI_PARAM_DATASEQ_INORDER_EN:
2335                 *value = session->dataseq_inorder_en;
2336                 break;
2337         case ISCSI_PARAM_ERL:
2338                 *value = session->erl;
2339                 break;
2340         case ISCSI_PARAM_IFMARKER_EN:
2341                 *value = session->ifmarker_en;
2342                 break;
2343         case ISCSI_PARAM_OFMARKER_EN:
2344                 *value = session->ofmarker_en;
2345                 break;
2346         default:
2347                 return -EINVAL;
2348         }
2349
2350         return 0;
2351 }
2352
2353 static int
2354 iscsi_conn_get_param(struct iscsi_cls_conn *cls_conn,
2355                      enum iscsi_param param, uint32_t *value)
2356 {
2357         struct iscsi_conn *conn = cls_conn->dd_data;
2358         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
2359         struct inet_sock *inet;
2360
2361         switch(param) {
2362         case ISCSI_PARAM_MAX_RECV_DLENGTH:
2363                 *value = conn->max_recv_dlength;
2364                 break;
2365         case ISCSI_PARAM_MAX_XMIT_DLENGTH:
2366                 *value = conn->max_xmit_dlength;
2367                 break;
2368         case ISCSI_PARAM_HDRDGST_EN:
2369                 *value = conn->hdrdgst_en;
2370                 break;
2371         case ISCSI_PARAM_DATADGST_EN:
2372                 *value = conn->datadgst_en;
2373                 break;
2374         case ISCSI_PARAM_CONN_PORT:
2375                 mutex_lock(&conn->xmitmutex);
2376                 if (!tcp_conn->sock) {
2377                         mutex_unlock(&conn->xmitmutex);
2378                         return -EINVAL;
2379                 }
2380
2381                 inet = inet_sk(tcp_conn->sock->sk);
2382                 *value = be16_to_cpu(inet->dport);
2383                 mutex_unlock(&conn->xmitmutex);
2384         default:
2385                 return -EINVAL;
2386         }
2387
2388         return 0;
2389 }
2390
2391 static int
2392 iscsi_conn_get_str_param(struct iscsi_cls_conn *cls_conn,
2393                          enum iscsi_param param, char *buf)
2394 {
2395         struct iscsi_conn *conn = cls_conn->dd_data;
2396         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
2397         struct sock *sk;
2398         struct inet_sock *inet;
2399         struct ipv6_pinfo *np;
2400         int len = 0;
2401
2402         switch (param) {
2403         case ISCSI_PARAM_CONN_ADDRESS:
2404                 mutex_lock(&conn->xmitmutex);
2405                 if (!tcp_conn->sock) {
2406                         mutex_unlock(&conn->xmitmutex);
2407                         return -EINVAL;
2408                 }
2409
2410                 sk = tcp_conn->sock->sk;
2411                 if (sk->sk_family == PF_INET) {
2412                         inet = inet_sk(sk);
2413                         len = sprintf(buf, "%u.%u.%u.%u\n",
2414                                       NIPQUAD(inet->daddr));
2415                 } else {
2416                         np = inet6_sk(sk);
2417                         len = sprintf(buf,
2418                                 "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
2419                                 NIP6(np->daddr));
2420                 }
2421                 mutex_unlock(&conn->xmitmutex);
2422                 break;
2423         default:
2424                 return -EINVAL;
2425         }
2426
2427         return len;
2428 }
2429
2430 static void
2431 iscsi_conn_get_stats(struct iscsi_cls_conn *cls_conn, struct iscsi_stats *stats)
2432 {
2433         struct iscsi_conn *conn = cls_conn->dd_data;
2434         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
2435
2436         stats->txdata_octets = conn->txdata_octets;
2437         stats->rxdata_octets = conn->rxdata_octets;
2438         stats->scsicmd_pdus = conn->scsicmd_pdus_cnt;
2439         stats->dataout_pdus = conn->dataout_pdus_cnt;
2440         stats->scsirsp_pdus = conn->scsirsp_pdus_cnt;
2441         stats->datain_pdus = conn->datain_pdus_cnt;
2442         stats->r2t_pdus = conn->r2t_pdus_cnt;
2443         stats->tmfcmd_pdus = conn->tmfcmd_pdus_cnt;
2444         stats->tmfrsp_pdus = conn->tmfrsp_pdus_cnt;
2445         stats->custom_length = 3;
2446         strcpy(stats->custom[0].desc, "tx_sendpage_failures");
2447         stats->custom[0].value = tcp_conn->sendpage_failures_cnt;
2448         strcpy(stats->custom[1].desc, "rx_discontiguous_hdr");
2449         stats->custom[1].value = tcp_conn->discontiguous_hdr_cnt;
2450         strcpy(stats->custom[2].desc, "eh_abort_cnt");
2451         stats->custom[2].value = conn->eh_abort_cnt;
2452 }
2453
2454 static struct iscsi_cls_session *
2455 iscsi_tcp_session_create(struct iscsi_transport *iscsit,
2456                          struct scsi_transport_template *scsit,
2457                          uint32_t initial_cmdsn, uint32_t *hostno)
2458 {
2459         struct iscsi_cls_session *cls_session;
2460         struct iscsi_session *session;
2461         uint32_t hn;
2462         int cmd_i;
2463
2464         cls_session = iscsi_session_setup(iscsit, scsit,
2465                                          sizeof(struct iscsi_tcp_cmd_task),
2466                                          sizeof(struct iscsi_tcp_mgmt_task),
2467                                          initial_cmdsn, &hn);
2468         if (!cls_session)
2469                 return NULL;
2470         *hostno = hn;
2471
2472         session = class_to_transport_session(cls_session);
2473         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
2474                 struct iscsi_cmd_task *ctask = session->cmds[cmd_i];
2475                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
2476
2477                 ctask->hdr = &tcp_ctask->hdr;
2478         }
2479
2480         for (cmd_i = 0; cmd_i < session->mgmtpool_max; cmd_i++) {
2481                 struct iscsi_mgmt_task *mtask = session->mgmt_cmds[cmd_i];
2482                 struct iscsi_tcp_mgmt_task *tcp_mtask = mtask->dd_data;
2483
2484                 mtask->hdr = &tcp_mtask->hdr;
2485         }
2486
2487         if (iscsi_r2tpool_alloc(class_to_transport_session(cls_session)))
2488                 goto r2tpool_alloc_fail;
2489
2490         return cls_session;
2491
2492 r2tpool_alloc_fail:
2493         iscsi_session_teardown(cls_session);
2494         return NULL;
2495 }
2496
2497 static void iscsi_tcp_session_destroy(struct iscsi_cls_session *cls_session)
2498 {
2499         struct iscsi_session *session = class_to_transport_session(cls_session);
2500         struct iscsi_data_task *dtask, *n;
2501         int cmd_i;
2502
2503         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
2504                 struct iscsi_cmd_task *ctask = session->cmds[cmd_i];
2505                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
2506
2507                 list_for_each_entry_safe(dtask, n, &tcp_ctask->dataqueue,
2508                                          item) {
2509                         list_del(&dtask->item);
2510                         mempool_free(dtask, tcp_ctask->datapool);
2511                 }
2512         }
2513
2514         iscsi_r2tpool_free(class_to_transport_session(cls_session));
2515         iscsi_session_teardown(cls_session);
2516 }
2517
2518 static struct scsi_host_template iscsi_sht = {
2519         .name                   = "iSCSI Initiator over TCP/IP, v."
2520                                   ISCSI_VERSION_STR,
2521         .queuecommand           = iscsi_queuecommand,
2522         .change_queue_depth     = iscsi_change_queue_depth,
2523         .can_queue              = ISCSI_XMIT_CMDS_MAX - 1,
2524         .sg_tablesize           = ISCSI_SG_TABLESIZE,
2525         .cmd_per_lun            = ISCSI_DEF_CMD_PER_LUN,
2526         .eh_abort_handler       = iscsi_eh_abort,
2527         .eh_host_reset_handler  = iscsi_eh_host_reset,
2528         .use_clustering         = DISABLE_CLUSTERING,
2529         .proc_name              = "iscsi_tcp",
2530         .this_id                = -1,
2531 };
2532
2533 static struct iscsi_transport iscsi_tcp_transport = {
2534         .owner                  = THIS_MODULE,
2535         .name                   = "tcp",
2536         .caps                   = CAP_RECOVERY_L0 | CAP_MULTI_R2T | CAP_HDRDGST
2537                                   | CAP_DATADGST,
2538         .param_mask             = ISCSI_MAX_RECV_DLENGTH |
2539                                   ISCSI_MAX_XMIT_DLENGTH |
2540                                   ISCSI_HDRDGST_EN |
2541                                   ISCSI_DATADGST_EN |
2542                                   ISCSI_INITIAL_R2T_EN |
2543                                   ISCSI_MAX_R2T |
2544                                   ISCSI_IMM_DATA_EN |
2545                                   ISCSI_FIRST_BURST |
2546                                   ISCSI_MAX_BURST |
2547                                   ISCSI_PDU_INORDER_EN |
2548                                   ISCSI_DATASEQ_INORDER_EN |
2549                                   ISCSI_ERL |
2550                                   ISCSI_CONN_PORT |
2551                                   ISCSI_CONN_ADDRESS,
2552         .host_template          = &iscsi_sht,
2553         .conndata_size          = sizeof(struct iscsi_conn),
2554         .max_conn               = 1,
2555         .max_cmd_len            = ISCSI_TCP_MAX_CMD_LEN,
2556         /* session management */
2557         .create_session         = iscsi_tcp_session_create,
2558         .destroy_session        = iscsi_tcp_session_destroy,
2559         /* connection management */
2560         .create_conn            = iscsi_tcp_conn_create,
2561         .bind_conn              = iscsi_tcp_conn_bind,
2562         .destroy_conn           = iscsi_tcp_conn_destroy,
2563         .set_param              = iscsi_conn_set_param,
2564         .get_conn_param         = iscsi_conn_get_param,
2565         .get_conn_str_param     = iscsi_conn_get_str_param,
2566         .get_session_param      = iscsi_session_get_param,
2567         .start_conn             = iscsi_conn_start,
2568         .stop_conn              = iscsi_conn_stop,
2569         /* these are called as part of conn recovery */
2570         .suspend_conn_recv      = iscsi_tcp_suspend_conn_rx,
2571         .terminate_conn         = iscsi_tcp_terminate_conn,
2572         /* IO */
2573         .send_pdu               = iscsi_conn_send_pdu,
2574         .get_stats              = iscsi_conn_get_stats,
2575         .init_cmd_task          = iscsi_tcp_cmd_init,
2576         .init_mgmt_task         = iscsi_tcp_mgmt_init,
2577         .xmit_cmd_task          = iscsi_tcp_ctask_xmit,
2578         .xmit_mgmt_task         = iscsi_tcp_mtask_xmit,
2579         .cleanup_cmd_task       = iscsi_tcp_cleanup_ctask,
2580         /* recovery */
2581         .session_recovery_timedout = iscsi_session_recovery_timedout,
2582 };
2583
2584 static int __init
2585 iscsi_tcp_init(void)
2586 {
2587         if (iscsi_max_lun < 1) {
2588                 printk(KERN_ERR "Invalid max_lun value of %u\n", iscsi_max_lun);
2589                 return -EINVAL;
2590         }
2591         iscsi_tcp_transport.max_lun = iscsi_max_lun;
2592
2593         taskcache = kmem_cache_create("iscsi_taskcache",
2594                         sizeof(struct iscsi_data_task), 0,
2595                         SLAB_HWCACHE_ALIGN, NULL, NULL);
2596         if (!taskcache)
2597                 return -ENOMEM;
2598
2599         if (!iscsi_register_transport(&iscsi_tcp_transport))
2600                 kmem_cache_destroy(taskcache);
2601
2602         return 0;
2603 }
2604
2605 static void __exit
2606 iscsi_tcp_exit(void)
2607 {
2608         iscsi_unregister_transport(&iscsi_tcp_transport);
2609         kmem_cache_destroy(taskcache);
2610 }
2611
2612 module_init(iscsi_tcp_init);
2613 module_exit(iscsi_tcp_exit);