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 /* #define DEBUG_TCP */
  51 #define DEBUG_ASSERT
  52
  53 #ifdef DEBUG_TCP
  54 #define debug_tcp(fmt...) printk(KERN_INFO "tcp: " fmt)
  55 #else
  56 #define debug_tcp(fmt...)
  57 #endif
  58
  59 #ifndef DEBUG_ASSERT
  60 #ifdef BUG_ON
  61 #undef BUG_ON
  62 #endif
  63 #define BUG_ON(expr)
  64 #endif
  65
  66 static unsigned int iscsi_max_lun = 512;
  67 module_param_named(max_lun, iscsi_max_lun, uint, S_IRUGO);
  68
  69 static inline void
  70 iscsi_buf_init_iov(struct iscsi_buf *ibuf, char *vbuf, int size)
  71 {
  72         ibuf->sg.page = virt_to_page(vbuf);
  73         ibuf->sg.offset = offset_in_page(vbuf);
  74         ibuf->sg.length = size;
  75         ibuf->sent = 0;
  76         ibuf->use_sendmsg = 1;
  77 }
  78
  79 static inline void
  80 iscsi_buf_init_sg(struct iscsi_buf *ibuf, struct scatterlist *sg)
  81 {
  82         ibuf->sg.page = sg->page;
  83         ibuf->sg.offset = sg->offset;
  84         ibuf->sg.length = sg->length;
  85         /*
  86          * Fastpath: sg element fits into single page
  87          */
  88         if (sg->length + sg->offset <= PAGE_SIZE && !PageSlab(sg->page))
  89                 ibuf->use_sendmsg = 0;
  90         else
  91                 ibuf->use_sendmsg = 1;
  92         ibuf->sent = 0;
  93 }
  94
  95 static inline int
  96 iscsi_buf_left(struct iscsi_buf *ibuf)
  97 {
  98         int rc;
  99
 100         rc = ibuf->sg.length - ibuf->sent;
 101         BUG_ON(rc < 0);
 102         return rc;
 103 }
 104
 105 static inline void
 106 iscsi_hdr_digest(struct iscsi_conn *conn, struct iscsi_buf *buf,
 107                  u8* crc)
 108 {
 109         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 110
 111         crypto_hash_digest(&tcp_conn->tx_hash, &buf->sg, buf->sg.length, crc);
 112         buf->sg.length = tcp_conn->hdr_size;
 113 }
 114
 115 static inline int
 116 iscsi_hdr_extract(struct iscsi_tcp_conn *tcp_conn)
 117 {
 118         struct sk_buff *skb = tcp_conn->in.skb;
 119
 120         tcp_conn->in.zero_copy_hdr = 0;
 121
 122         if (tcp_conn->in.copy >= tcp_conn->hdr_size &&
 123             tcp_conn->in_progress == IN_PROGRESS_WAIT_HEADER) {
 124                 /*
 125                  * Zero-copy PDU Header: using connection context
 126                  * to store header pointer.
 127                  */
 128                 if (skb_shinfo(skb)->frag_list == NULL &&
 129                     !skb_shinfo(skb)->nr_frags) {
 130                         tcp_conn->in.hdr = (struct iscsi_hdr *)
 131                                 ((char*)skb->data + tcp_conn->in.offset);
 132                         tcp_conn->in.zero_copy_hdr = 1;
 133                 } else {
 134                         /* ignoring return code since we checked
 135                          * in.copy before */
 136                         skb_copy_bits(skb, tcp_conn->in.offset,
 137                                 &tcp_conn->hdr, tcp_conn->hdr_size);
 138                         tcp_conn->in.hdr = &tcp_conn->hdr;
 139                 }
 140                 tcp_conn->in.offset += tcp_conn->hdr_size;
 141                 tcp_conn->in.copy -= tcp_conn->hdr_size;
 142         } else {
 143                 int hdr_remains;
 144                 int copylen;
 145
 146                 /*
 147                  * PDU header scattered across SKB's,
 148                  * copying it... This'll happen quite rarely.
 149                  */
 150
 151                 if (tcp_conn->in_progress == IN_PROGRESS_WAIT_HEADER)
 152                         tcp_conn->in.hdr_offset = 0;
 153
 154                 hdr_remains = tcp_conn->hdr_size - tcp_conn->in.hdr_offset;
 155                 BUG_ON(hdr_remains <= 0);
 156
 157                 copylen = min(tcp_conn->in.copy, hdr_remains);
 158                 skb_copy_bits(skb, tcp_conn->in.offset,
 159                         (char*)&tcp_conn->hdr + tcp_conn->in.hdr_offset,
 160                         copylen);
 161
 162                 debug_tcp("PDU gather offset %d bytes %d in.offset %d "
 163                        "in.copy %d\n", tcp_conn->in.hdr_offset, copylen,
 164                        tcp_conn->in.offset, tcp_conn->in.copy);
 165
 166                 tcp_conn->in.offset += copylen;
 167                 tcp_conn->in.copy -= copylen;
 168                 if (copylen < hdr_remains)  {
 169                         tcp_conn->in_progress = IN_PROGRESS_HEADER_GATHER;
 170                         tcp_conn->in.hdr_offset += copylen;
 171                         return -EAGAIN;
 172                 }
 173                 tcp_conn->in.hdr = &tcp_conn->hdr;
 174                 tcp_conn->discontiguous_hdr_cnt++;
 175                 tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
 176         }
 177
 178         return 0;
 179 }
 180
 181 /*
 182  * must be called with session lock
 183  */
 184 static void
 185 iscsi_tcp_cleanup_ctask(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
 186 {
 187         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 188         struct iscsi_r2t_info *r2t;
 189         struct scsi_cmnd *sc;
 190
 191         /* flush ctask's r2t queues */
 192         while (__kfifo_get(tcp_ctask->r2tqueue, (void*)&r2t, sizeof(void*))) {
 193                 __kfifo_put(tcp_ctask->r2tpool.queue, (void*)&r2t,
 194                             sizeof(void*));
 195                 debug_scsi("iscsi_tcp_cleanup_ctask pending r2t dropped\n");
 196         }
 197
 198         sc = ctask->sc;
 199         if (unlikely(!sc))
 200                 return;
 201
 202         tcp_ctask->xmstate = XMSTATE_IDLE;
 203         tcp_ctask->r2t = NULL;
 204 }
 205
 206 /**
 207  * iscsi_data_rsp - SCSI Data-In Response processing
 208  * @conn: iscsi connection
 209  * @ctask: scsi command task
 210  **/
 211 static int
 212 iscsi_data_rsp(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
 213 {
 214         int rc;
 215         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 216         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 217         struct iscsi_data_rsp *rhdr = (struct iscsi_data_rsp *)tcp_conn->in.hdr;
 218         struct iscsi_session *session = conn->session;
 219         int datasn = be32_to_cpu(rhdr->datasn);
 220
 221         rc = iscsi_check_assign_cmdsn(session, (struct iscsi_nopin*)rhdr);
 222         if (rc)
 223                 return rc;
 224         /*
 225          * setup Data-In byte counter (gets decremented..)
 226          */
 227         ctask->data_count = tcp_conn->in.datalen;
 228
 229         if (tcp_conn->in.datalen == 0)
 230                 return 0;
 231
 232         if (ctask->datasn != datasn)
 233                 return ISCSI_ERR_DATASN;
 234
 235         ctask->datasn++;
 236
 237         tcp_ctask->data_offset = be32_to_cpu(rhdr->offset);
 238         if (tcp_ctask->data_offset + tcp_conn->in.datalen > ctask->total_length)
 239                 return ISCSI_ERR_DATA_OFFSET;
 240
 241         if (rhdr->flags & ISCSI_FLAG_DATA_STATUS) {
 242                 struct scsi_cmnd *sc = ctask->sc;
 243
 244                 conn->exp_statsn = be32_to_cpu(rhdr->statsn) + 1;
 245                 if (rhdr->flags & ISCSI_FLAG_DATA_UNDERFLOW) {
 246                         int res_count = be32_to_cpu(rhdr->residual_count);
 247
 248                         if (res_count > 0 &&
 249                             res_count <= sc->request_bufflen) {
 250                                 sc->resid = res_count;
 251                                 sc->result = (DID_OK << 16) | rhdr->cmd_status;
 252                         } else
 253                                 sc->result = (DID_BAD_TARGET << 16) |
 254                                         rhdr->cmd_status;
 255                 } else if (rhdr->flags & ISCSI_FLAG_DATA_OVERFLOW) {
 256                         sc->resid = be32_to_cpu(rhdr->residual_count);
 257                         sc->result = (DID_OK << 16) | rhdr->cmd_status;
 258                 } else
 259                         sc->result = (DID_OK << 16) | rhdr->cmd_status;
 260         }
 261
 262         conn->datain_pdus_cnt++;
 263         return 0;
 264 }
 265
 266 /**
 267  * iscsi_solicit_data_init - initialize first Data-Out
 268  * @conn: iscsi connection
 269  * @ctask: scsi command task
 270  * @r2t: R2T info
 271  *
 272  * Notes:
 273  *      Initialize first Data-Out within this R2T sequence and finds
 274  *      proper data_offset within this SCSI command.
 275  *
 276  *      This function is called with connection lock taken.
 277  **/
 278 static void
 279 iscsi_solicit_data_init(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask,
 280                         struct iscsi_r2t_info *r2t)
 281 {
 282         struct iscsi_data *hdr;
 283         struct scsi_cmnd *sc = ctask->sc;
 284
 285         hdr = &r2t->dtask.hdr;
 286         memset(hdr, 0, sizeof(struct iscsi_data));
 287         hdr->ttt = r2t->ttt;
 288         hdr->datasn = cpu_to_be32(r2t->solicit_datasn);
 289         r2t->solicit_datasn++;
 290         hdr->opcode = ISCSI_OP_SCSI_DATA_OUT;
 291         memcpy(hdr->lun, ctask->hdr->lun, sizeof(hdr->lun));
 292         hdr->itt = ctask->hdr->itt;
 293         hdr->exp_statsn = r2t->exp_statsn;
 294         hdr->offset = cpu_to_be32(r2t->data_offset);
 295         if (r2t->data_length > conn->max_xmit_dlength) {
 296                 hton24(hdr->dlength, conn->max_xmit_dlength);
 297                 r2t->data_count = conn->max_xmit_dlength;
 298                 hdr->flags = 0;
 299         } else {
 300                 hton24(hdr->dlength, r2t->data_length);
 301                 r2t->data_count = r2t->data_length;
 302                 hdr->flags = ISCSI_FLAG_CMD_FINAL;
 303         }
 304         conn->dataout_pdus_cnt++;
 305
 306         r2t->sent = 0;
 307
 308         iscsi_buf_init_iov(&r2t->headbuf, (char*)hdr,
 309                            sizeof(struct iscsi_hdr));
 310
 311         if (sc->use_sg) {
 312                 int i, sg_count = 0;
 313                 struct scatterlist *sg = sc->request_buffer;
 314
 315                 r2t->sg = NULL;
 316                 for (i = 0; i < sc->use_sg; i++, sg += 1) {
 317                         /* FIXME: prefetch ? */
 318                         if (sg_count + sg->length > r2t->data_offset) {
 319                                 int page_offset;
 320
 321                                 /* sg page found! */
 322
 323                                 /* offset within this page */
 324                                 page_offset = r2t->data_offset - sg_count;
 325
 326                                 /* fill in this buffer */
 327                                 iscsi_buf_init_sg(&r2t->sendbuf, sg);
 328                                 r2t->sendbuf.sg.offset += page_offset;
 329                                 r2t->sendbuf.sg.length -= page_offset;
 330
 331                                 /* xmit logic will continue with next one */
 332                                 r2t->sg = sg + 1;
 333                                 break;
 334                         }
 335                         sg_count += sg->length;
 336                 }
 337                 BUG_ON(r2t->sg == NULL);
 338         } else {
 339                 iscsi_buf_init_iov(&r2t->sendbuf,
 340                             (char*)sc->request_buffer + r2t->data_offset,
 341                             r2t->data_count);
 342                 r2t->sg = NULL;
 343         }
 344 }
 345
 346 /**
 347  * iscsi_r2t_rsp - iSCSI R2T Response processing
 348  * @conn: iscsi connection
 349  * @ctask: scsi command task
 350  **/
 351 static int
 352 iscsi_r2t_rsp(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
 353 {
 354         struct iscsi_r2t_info *r2t;
 355         struct iscsi_session *session = conn->session;
 356         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 357         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 358         struct iscsi_r2t_rsp *rhdr = (struct iscsi_r2t_rsp *)tcp_conn->in.hdr;
 359         int r2tsn = be32_to_cpu(rhdr->r2tsn);
 360         int rc;
 361
 362         if (tcp_conn->in.datalen) {
 363                 printk(KERN_ERR "iscsi_tcp: invalid R2t with datalen %d\n",
 364                        tcp_conn->in.datalen);
 365                 return ISCSI_ERR_DATALEN;
 366         }
 367
 368         if (tcp_ctask->exp_r2tsn && tcp_ctask->exp_r2tsn != r2tsn)
 369                 return ISCSI_ERR_R2TSN;
 370
 371         rc = iscsi_check_assign_cmdsn(session, (struct iscsi_nopin*)rhdr);
 372         if (rc)
 373                 return rc;
 374
 375         /* FIXME: use R2TSN to detect missing R2T */
 376
 377         /* fill-in new R2T associated with the task */
 378         spin_lock(&session->lock);
 379         if (!ctask->sc || ctask->mtask ||
 380              session->state != ISCSI_STATE_LOGGED_IN) {
 381                 printk(KERN_INFO "iscsi_tcp: dropping R2T itt %d in "
 382                        "recovery...\n", ctask->itt);
 383                 spin_unlock(&session->lock);
 384                 return 0;
 385         }
 386
 387         rc = __kfifo_get(tcp_ctask->r2tpool.queue, (void*)&r2t, sizeof(void*));
 388         BUG_ON(!rc);
 389
 390         r2t->exp_statsn = rhdr->statsn;
 391         r2t->data_length = be32_to_cpu(rhdr->data_length);
 392         if (r2t->data_length == 0) {
 393                 printk(KERN_ERR "iscsi_tcp: invalid R2T with zero data len\n");
 394                 spin_unlock(&session->lock);
 395                 return ISCSI_ERR_DATALEN;
 396         }
 397
 398         if (r2t->data_length > session->max_burst)
 399                 debug_scsi("invalid R2T with data len %u and max burst %u."
 400                            "Attempting to execute request.\n",
 401                             r2t->data_length, session->max_burst);
 402
 403         r2t->data_offset = be32_to_cpu(rhdr->data_offset);
 404         if (r2t->data_offset + r2t->data_length > ctask->total_length) {
 405                 spin_unlock(&session->lock);
 406                 printk(KERN_ERR "iscsi_tcp: invalid R2T with data len %u at "
 407                        "offset %u and total length %d\n", r2t->data_length,
 408                        r2t->data_offset, ctask->total_length);
 409                 return ISCSI_ERR_DATALEN;
 410         }
 411
 412         r2t->ttt = rhdr->ttt; /* no flip */
 413         r2t->solicit_datasn = 0;
 414
 415         iscsi_solicit_data_init(conn, ctask, r2t);
 416
 417         tcp_ctask->exp_r2tsn = r2tsn + 1;
 418         __kfifo_put(tcp_ctask->r2tqueue, (void*)&r2t, sizeof(void*));
 419         tcp_ctask->xmstate |= XMSTATE_SOL_HDR;
 420         list_move_tail(&ctask->running, &conn->xmitqueue);
 421
 422         scsi_queue_work(session->host, &conn->xmitwork);
 423         conn->r2t_pdus_cnt++;
 424         spin_unlock(&session->lock);
 425
 426         return 0;
 427 }
 428
 429 static int
 430 iscsi_tcp_hdr_recv(struct iscsi_conn *conn)
 431 {
 432         int rc = 0, opcode, ahslen;
 433         struct iscsi_hdr *hdr;
 434         struct iscsi_session *session = conn->session;
 435         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 436         uint32_t cdgst, rdgst = 0, itt;
 437
 438         hdr = tcp_conn->in.hdr;
 439
 440         /* verify PDU length */
 441         tcp_conn->in.datalen = ntoh24(hdr->dlength);
 442         if (tcp_conn->in.datalen > conn->max_recv_dlength) {
 443                 printk(KERN_ERR "iscsi_tcp: datalen %d > %d\n",
 444                        tcp_conn->in.datalen, conn->max_recv_dlength);
 445                 return ISCSI_ERR_DATALEN;
 446         }
 447         tcp_conn->data_copied = 0;
 448
 449         /* read AHS */
 450         ahslen = hdr->hlength << 2;
 451         tcp_conn->in.offset += ahslen;
 452         tcp_conn->in.copy -= ahslen;
 453         if (tcp_conn->in.copy < 0) {
 454                 printk(KERN_ERR "iscsi_tcp: can't handle AHS with length "
 455                        "%d bytes\n", ahslen);
 456                 return ISCSI_ERR_AHSLEN;
 457         }
 458
 459         /* calculate read padding */
 460         tcp_conn->in.padding = tcp_conn->in.datalen & (ISCSI_PAD_LEN-1);
 461         if (tcp_conn->in.padding) {
 462                 tcp_conn->in.padding = ISCSI_PAD_LEN - tcp_conn->in.padding;
 463                 debug_scsi("read padding %d bytes\n", tcp_conn->in.padding);
 464         }
 465
 466         if (conn->hdrdgst_en) {
 467                 struct scatterlist sg;
 468
 469                 sg_init_one(&sg, (u8 *)hdr,
 470                             sizeof(struct iscsi_hdr) + ahslen);
 471                 crypto_hash_digest(&tcp_conn->rx_hash, &sg, sg.length,
 472                                    (u8 *)&cdgst);
 473                 rdgst = *(uint32_t*)((char*)hdr + sizeof(struct iscsi_hdr) +
 474                                      ahslen);
 475                 if (cdgst != rdgst) {
 476                         printk(KERN_ERR "iscsi_tcp: hdrdgst error "
 477                                "recv 0x%x calc 0x%x\n", rdgst, cdgst);
 478                         return ISCSI_ERR_HDR_DGST;
 479                 }
 480         }
 481
 482         opcode = hdr->opcode & ISCSI_OPCODE_MASK;
 483         /* verify itt (itt encoding: age+cid+itt) */
 484         rc = iscsi_verify_itt(conn, hdr, &itt);
 485         if (rc == ISCSI_ERR_NO_SCSI_CMD) {
 486                 tcp_conn->in.datalen = 0; /* force drop */
 487                 return 0;
 488         } else if (rc)
 489                 return rc;
 490
 491         debug_tcp("opcode 0x%x offset %d copy %d ahslen %d datalen %d\n",
 492                   opcode, tcp_conn->in.offset, tcp_conn->in.copy,
 493                   ahslen, tcp_conn->in.datalen);
 494
 495         switch(opcode) {
 496         case ISCSI_OP_SCSI_DATA_IN:
 497                 tcp_conn->in.ctask = session->cmds[itt];
 498                 rc = iscsi_data_rsp(conn, tcp_conn->in.ctask);
 499                 if (rc)
 500                         return rc;
 501                 /* fall through */
 502         case ISCSI_OP_SCSI_CMD_RSP:
 503                 tcp_conn->in.ctask = session->cmds[itt];
 504                 if (tcp_conn->in.datalen)
 505                         goto copy_hdr;
 506
 507                 spin_lock(&session->lock);
 508                 rc = __iscsi_complete_pdu(conn, hdr, NULL, 0);
 509                 spin_unlock(&session->lock);
 510                 break;
 511         case ISCSI_OP_R2T:
 512                 tcp_conn->in.ctask = session->cmds[itt];
 513                 if (ahslen)
 514                         rc = ISCSI_ERR_AHSLEN;
 515                 else if (tcp_conn->in.ctask->sc->sc_data_direction ==
 516                                                                 DMA_TO_DEVICE)
 517                         rc = iscsi_r2t_rsp(conn, tcp_conn->in.ctask);
 518                 else
 519                         rc = ISCSI_ERR_PROTO;
 520                 break;
 521         case ISCSI_OP_LOGIN_RSP:
 522         case ISCSI_OP_TEXT_RSP:
 523         case ISCSI_OP_REJECT:
 524         case ISCSI_OP_ASYNC_EVENT:
 525                 /*
 526                  * It is possible that we could get a PDU with a buffer larger
 527                  * than 8K, but there are no targets that currently do this.
 528                  * For now we fail until we find a vendor that needs it
 529                  */
 530                 if (ISCSI_DEF_MAX_RECV_SEG_LEN <
 531                     tcp_conn->in.datalen) {
 532                         printk(KERN_ERR "iscsi_tcp: received buffer of len %u "
 533                               "but conn buffer is only %u (opcode %0x)\n",
 534                               tcp_conn->in.datalen,
 535                               ISCSI_DEF_MAX_RECV_SEG_LEN, opcode);
 536                         rc = ISCSI_ERR_PROTO;
 537                         break;
 538                 }
 539
 540                 if (tcp_conn->in.datalen)
 541                         goto copy_hdr;
 542         /* fall through */
 543         case ISCSI_OP_LOGOUT_RSP:
 544         case ISCSI_OP_NOOP_IN:
 545         case ISCSI_OP_SCSI_TMFUNC_RSP:
 546                 rc = iscsi_complete_pdu(conn, hdr, NULL, 0);
 547                 break;
 548         default:
 549                 rc = ISCSI_ERR_BAD_OPCODE;
 550                 break;
 551         }
 552
 553         return rc;
 554
 555 copy_hdr:
 556         /*
 557          * if we did zero copy for the header but we will need multiple
 558          * skbs to complete the command then we have to copy the header
 559          * for later use
 560          */
 561         if (tcp_conn->in.zero_copy_hdr && tcp_conn->in.copy <=
 562            (tcp_conn->in.datalen + tcp_conn->in.padding +
 563             (conn->datadgst_en ? 4 : 0))) {
 564                 debug_tcp("Copying header for later use. in.copy %d in.datalen"
 565                           " %d\n", tcp_conn->in.copy, tcp_conn->in.datalen);
 566                 memcpy(&tcp_conn->hdr, tcp_conn->in.hdr,
 567                        sizeof(struct iscsi_hdr));
 568                 tcp_conn->in.hdr = &tcp_conn->hdr;
 569                 tcp_conn->in.zero_copy_hdr = 0;
 570         }
 571         return 0;
 572 }
 573
 574 /**
 575  * iscsi_ctask_copy - copy skb bits to the destanation cmd task
 576  * @conn: iscsi tcp connection
 577  * @ctask: scsi command task
 578  * @buf: buffer to copy to
 579  * @buf_size: size of buffer
 580  * @offset: offset within the buffer
 581  *
 582  * Notes:
 583  *      The function calls skb_copy_bits() and updates per-connection and
 584  *      per-cmd byte counters.
 585  *
 586  *      Read counters (in bytes):
 587  *
 588  *      conn->in.offset         offset within in progress SKB
 589  *      conn->in.copy           left to copy from in progress SKB
 590  *                              including padding
 591  *      conn->in.copied         copied already from in progress SKB
 592  *      conn->data_copied       copied already from in progress buffer
 593  *      ctask->sent             total bytes sent up to the MidLayer
 594  *      ctask->data_count       left to copy from in progress Data-In
 595  *      buf_left                left to copy from in progress buffer
 596  **/
 597 static inline int
 598 iscsi_ctask_copy(struct iscsi_tcp_conn *tcp_conn, struct iscsi_cmd_task *ctask,
 599                 void *buf, int buf_size, int offset)
 600 {
 601         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 602         int buf_left = buf_size - (tcp_conn->data_copied + offset);
 603         int size = min(tcp_conn->in.copy, buf_left);
 604         int rc;
 605
 606         size = min(size, ctask->data_count);
 607
 608         debug_tcp("ctask_copy %d bytes at offset %d copied %d\n",
 609                size, tcp_conn->in.offset, tcp_conn->in.copied);
 610
 611         BUG_ON(size <= 0);
 612         BUG_ON(tcp_ctask->sent + size > ctask->total_length);
 613
 614         rc = skb_copy_bits(tcp_conn->in.skb, tcp_conn->in.offset,
 615                            (char*)buf + (offset + tcp_conn->data_copied), size);
 616         /* must fit into skb->len */
 617         BUG_ON(rc);
 618
 619         tcp_conn->in.offset += size;
 620         tcp_conn->in.copy -= size;
 621         tcp_conn->in.copied += size;
 622         tcp_conn->data_copied += size;
 623         tcp_ctask->sent += size;
 624         ctask->data_count -= size;
 625
 626         BUG_ON(tcp_conn->in.copy < 0);
 627         BUG_ON(ctask->data_count < 0);
 628
 629         if (buf_size != (tcp_conn->data_copied + offset)) {
 630                 if (!ctask->data_count) {
 631                         BUG_ON(buf_size - tcp_conn->data_copied < 0);
 632                         /* done with this PDU */
 633                         return buf_size - tcp_conn->data_copied;
 634                 }
 635                 return -EAGAIN;
 636         }
 637
 638         /* done with this buffer or with both - PDU and buffer */
 639         tcp_conn->data_copied = 0;
 640         return 0;
 641 }
 642
 643 /**
 644  * iscsi_tcp_copy - copy skb bits to the destanation buffer
 645  * @conn: iscsi tcp connection
 646  *
 647  * Notes:
 648  *      The function calls skb_copy_bits() and updates per-connection
 649  *      byte counters.
 650  **/
 651 static inline int
 652 iscsi_tcp_copy(struct iscsi_conn *conn, int buf_size)
 653 {
 654         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 655         int buf_left = buf_size - tcp_conn->data_copied;
 656         int size = min(tcp_conn->in.copy, buf_left);
 657         int rc;
 658
 659         debug_tcp("tcp_copy %d bytes at offset %d copied %d\n",
 660                size, tcp_conn->in.offset, tcp_conn->data_copied);
 661         BUG_ON(size <= 0);
 662
 663         rc = skb_copy_bits(tcp_conn->in.skb, tcp_conn->in.offset,
 664                            (char*)conn->data + tcp_conn->data_copied, size);
 665         BUG_ON(rc);
 666
 667         tcp_conn->in.offset += size;
 668         tcp_conn->in.copy -= size;
 669         tcp_conn->in.copied += size;
 670         tcp_conn->data_copied += size;
 671
 672         if (buf_size != tcp_conn->data_copied)
 673                 return -EAGAIN;
 674
 675         return 0;
 676 }
 677
 678 static inline void
 679 partial_sg_digest_update(struct hash_desc *desc, struct scatterlist *sg,
 680                          int offset, int length)
 681 {
 682         struct scatterlist temp;
 683
 684         memcpy(&temp, sg, sizeof(struct scatterlist));
 685         temp.offset = offset;
 686         temp.length = length;
 687         crypto_hash_update(desc, &temp, length);
 688 }
 689
 690 static void
 691 iscsi_recv_digest_update(struct iscsi_tcp_conn *tcp_conn, char* buf, int len)
 692 {
 693         struct scatterlist tmp;
 694
 695         sg_init_one(&tmp, buf, len);
 696         crypto_hash_update(&tcp_conn->rx_hash, &tmp, len);
 697 }
 698
 699 static int iscsi_scsi_data_in(struct iscsi_conn *conn)
 700 {
 701         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 702         struct iscsi_cmd_task *ctask = tcp_conn->in.ctask;
 703         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 704         struct scsi_cmnd *sc = ctask->sc;
 705         struct scatterlist *sg;
 706         int i, offset, rc = 0;
 707
 708         BUG_ON((void*)ctask != sc->SCp.ptr);
 709
 710         /*
 711          * copying Data-In into the Scsi_Cmnd
 712          */
 713         if (!sc->use_sg) {
 714                 i = ctask->data_count;
 715                 rc = iscsi_ctask_copy(tcp_conn, ctask, sc->request_buffer,
 716                                       sc->request_bufflen,
 717                                       tcp_ctask->data_offset);
 718                 if (rc == -EAGAIN)
 719                         return rc;
 720                 if (conn->datadgst_en)
 721                         iscsi_recv_digest_update(tcp_conn, sc->request_buffer,
 722                                                  i);
 723                 rc = 0;
 724                 goto done;
 725         }
 726
 727         offset = tcp_ctask->data_offset;
 728         sg = sc->request_buffer;
 729
 730         if (tcp_ctask->data_offset)
 731                 for (i = 0; i < tcp_ctask->sg_count; i++)
 732                         offset -= sg[i].length;
 733         /* we've passed through partial sg*/
 734         if (offset < 0)
 735                 offset = 0;
 736
 737         for (i = tcp_ctask->sg_count; i < sc->use_sg; i++) {
 738                 char *dest;
 739
 740                 dest = kmap_atomic(sg[i].page, KM_SOFTIRQ0);
 741                 rc = iscsi_ctask_copy(tcp_conn, ctask, dest + sg[i].offset,
 742                                       sg[i].length, offset);
 743                 kunmap_atomic(dest, KM_SOFTIRQ0);
 744                 if (rc == -EAGAIN)
 745                         /* continue with the next SKB/PDU */
 746                         return rc;
 747                 if (!rc) {
 748                         if (conn->datadgst_en) {
 749                                 if (!offset)
 750                                         crypto_hash_update(
 751                                                         &tcp_conn->rx_hash,
 752                                                         &sg[i], sg[i].length);
 753                                 else
 754                                         partial_sg_digest_update(
 755                                                         &tcp_conn->rx_hash,
 756                                                         &sg[i],
 757                                                         sg[i].offset + offset,
 758                                                         sg[i].length - offset);
 759                         }
 760                         offset = 0;
 761                         tcp_ctask->sg_count++;
 762                 }
 763
 764                 if (!ctask->data_count) {
 765                         if (rc && conn->datadgst_en)
 766                                 /*
 767                                  * data-in is complete, but buffer not...
 768                                  */
 769                                 partial_sg_digest_update(&tcp_conn->rx_hash,
 770                                                          &sg[i],
 771                                                          sg[i].offset,
 772                                                          sg[i].length-rc);
 773                         rc = 0;
 774                         break;
 775                 }
 776
 777                 if (!tcp_conn->in.copy)
 778                         return -EAGAIN;
 779         }
 780         BUG_ON(ctask->data_count);
 781
 782 done:
 783         /* check for non-exceptional status */
 784         if (tcp_conn->in.hdr->flags & ISCSI_FLAG_DATA_STATUS) {
 785                 debug_scsi("done [sc %lx res %d itt 0x%x flags 0x%x]\n",
 786                            (long)sc, sc->result, ctask->itt,
 787                            tcp_conn->in.hdr->flags);
 788                 spin_lock(&conn->session->lock);
 789                 __iscsi_complete_pdu(conn, tcp_conn->in.hdr, NULL, 0);
 790                 spin_unlock(&conn->session->lock);
 791         }
 792
 793         return rc;
 794 }
 795
 796 static int
 797 iscsi_data_recv(struct iscsi_conn *conn)
 798 {
 799         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 800         int rc = 0, opcode;
 801
 802         opcode = tcp_conn->in.hdr->opcode & ISCSI_OPCODE_MASK;
 803         switch (opcode) {
 804         case ISCSI_OP_SCSI_DATA_IN:
 805                 rc = iscsi_scsi_data_in(conn);
 806                 break;
 807         case ISCSI_OP_SCSI_CMD_RSP:
 808         case ISCSI_OP_TEXT_RSP:
 809         case ISCSI_OP_LOGIN_RSP:
 810         case ISCSI_OP_ASYNC_EVENT:
 811         case ISCSI_OP_REJECT:
 812                 /*
 813                  * Collect data segment to the connection's data
 814                  * placeholder
 815                  */
 816                 if (iscsi_tcp_copy(conn, tcp_conn->in.datalen)) {
 817                         rc = -EAGAIN;
 818                         goto exit;
 819                 }
 820
 821                 rc = iscsi_complete_pdu(conn, tcp_conn->in.hdr, conn->data,
 822                                         tcp_conn->in.datalen);
 823                 if (!rc && conn->datadgst_en && opcode != ISCSI_OP_LOGIN_RSP)
 824                         iscsi_recv_digest_update(tcp_conn, conn->data,
 825                                                 tcp_conn->in.datalen);
 826                 break;
 827         default:
 828                 BUG_ON(1);
 829         }
 830 exit:
 831         return rc;
 832 }
 833
 834 /**
 835  * iscsi_tcp_data_recv - TCP receive in sendfile fashion
 836  * @rd_desc: read descriptor
 837  * @skb: socket buffer
 838  * @offset: offset in skb
 839  * @len: skb->len - offset
 840  **/
 841 static int
 842 iscsi_tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 843                 unsigned int offset, size_t len)
 844 {
 845         int rc;
 846         struct iscsi_conn *conn = rd_desc->arg.data;
 847         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 848         int processed;
 849         char pad[ISCSI_PAD_LEN];
 850         struct scatterlist sg;
 851
 852         /*
 853          * Save current SKB and its offset in the corresponding
 854          * connection context.
 855          */
 856         tcp_conn->in.copy = skb->len - offset;
 857         tcp_conn->in.offset = offset;
 858         tcp_conn->in.skb = skb;
 859         tcp_conn->in.len = tcp_conn->in.copy;
 860         BUG_ON(tcp_conn->in.copy <= 0);
 861         debug_tcp("in %d bytes\n", tcp_conn->in.copy);
 862
 863 more:
 864         tcp_conn->in.copied = 0;
 865         rc = 0;
 866
 867         if (unlikely(conn->suspend_rx)) {
 868                 debug_tcp("conn %d Rx suspended!\n", conn->id);
 869                 return 0;
 870         }
 871
 872         if (tcp_conn->in_progress == IN_PROGRESS_WAIT_HEADER ||
 873             tcp_conn->in_progress == IN_PROGRESS_HEADER_GATHER) {
 874                 rc = iscsi_hdr_extract(tcp_conn);
 875                 if (rc) {
 876                        if (rc == -EAGAIN)
 877                                 goto nomore;
 878                        else {
 879                                 iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
 880                                 return 0;
 881                        }
 882                 }
 883
 884                 /*
 885                  * Verify and process incoming PDU header.
 886                  */
 887                 rc = iscsi_tcp_hdr_recv(conn);
 888                 if (!rc && tcp_conn->in.datalen) {
 889                         if (conn->datadgst_en)
 890                                 crypto_hash_init(&tcp_conn->rx_hash);
 891                         tcp_conn->in_progress = IN_PROGRESS_DATA_RECV;
 892                 } else if (rc) {
 893                         iscsi_conn_failure(conn, rc);
 894                         return 0;
 895                 }
 896         }
 897
 898         if (tcp_conn->in_progress == IN_PROGRESS_DDIGEST_RECV) {
 899                 uint32_t recv_digest;
 900
 901                 debug_tcp("extra data_recv offset %d copy %d\n",
 902                           tcp_conn->in.offset, tcp_conn->in.copy);
 903                 rc = iscsi_tcp_copy(conn, sizeof(uint32_t));
 904                 if (rc) {
 905                         if (rc == -EAGAIN)
 906                                 goto again;
 907                         iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
 908                         return 0;
 909                 }
 910
 911                 memcpy(&recv_digest, conn->data, sizeof(uint32_t));
 912                 if (recv_digest != tcp_conn->in.datadgst) {
 913                         debug_tcp("iscsi_tcp: data digest error!"
 914                                   "0x%x != 0x%x\n", recv_digest,
 915                                   tcp_conn->in.datadgst);
 916                         iscsi_conn_failure(conn, ISCSI_ERR_DATA_DGST);
 917                         return 0;
 918                 } else {
 919                         debug_tcp("iscsi_tcp: data digest match!"
 920                                   "0x%x == 0x%x\n", recv_digest,
 921                                   tcp_conn->in.datadgst);
 922                         tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
 923                 }
 924         }
 925
 926         if (tcp_conn->in_progress == IN_PROGRESS_DATA_RECV &&
 927            tcp_conn->in.copy) {
 928
 929                 debug_tcp("data_recv offset %d copy %d\n",
 930                        tcp_conn->in.offset, tcp_conn->in.copy);
 931
 932                 rc = iscsi_data_recv(conn);
 933                 if (rc) {
 934                         if (rc == -EAGAIN)
 935                                 goto again;
 936                         iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
 937                         return 0;
 938                 }
 939                 tcp_conn->in.copy -= tcp_conn->in.padding;
 940                 tcp_conn->in.offset += tcp_conn->in.padding;
 941                 if (conn->datadgst_en) {
 942                         if (tcp_conn->in.padding) {
 943                                 debug_tcp("padding -> %d\n",
 944                                           tcp_conn->in.padding);
 945                                 memset(pad, 0, tcp_conn->in.padding);
 946                                 sg_init_one(&sg, pad, tcp_conn->in.padding);
 947                                 crypto_hash_update(&tcp_conn->rx_hash,
 948                                                    &sg, sg.length);
 949                         }
 950                         crypto_hash_final(&tcp_conn->rx_hash,
 951                                           (u8 *) &tcp_conn->in.datadgst);
 952                         debug_tcp("rx digest 0x%x\n", tcp_conn->in.datadgst);
 953                         tcp_conn->in_progress = IN_PROGRESS_DDIGEST_RECV;
 954                         tcp_conn->data_copied = 0;
 955                 } else
 956                         tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
 957         }
 958
 959         debug_tcp("f, processed %d from out of %d padding %d\n",
 960                tcp_conn->in.offset - offset, (int)len, tcp_conn->in.padding);
 961         BUG_ON(tcp_conn->in.offset - offset > len);
 962
 963         if (tcp_conn->in.offset - offset != len) {
 964                 debug_tcp("continue to process %d bytes\n",
 965                        (int)len - (tcp_conn->in.offset - offset));
 966                 goto more;
 967         }
 968
 969 nomore:
 970         processed = tcp_conn->in.offset - offset;
 971         BUG_ON(processed == 0);
 972         return processed;
 973
 974 again:
 975         processed = tcp_conn->in.offset - offset;
 976         debug_tcp("c, processed %d from out of %d rd_desc_cnt %d\n",
 977                   processed, (int)len, (int)rd_desc->count);
 978         BUG_ON(processed == 0);
 979         BUG_ON(processed > len);
 980
 981         conn->rxdata_octets += processed;
 982         return processed;
 983 }
 984
 985 static void
 986 iscsi_tcp_data_ready(struct sock *sk, int flag)
 987 {
 988         struct iscsi_conn *conn = sk->sk_user_data;
 989         read_descriptor_t rd_desc;
 990
 991         read_lock(&sk->sk_callback_lock);
 992
 993         /*
 994          * Use rd_desc to pass 'conn' to iscsi_tcp_data_recv.
 995          * We set count to 1 because we want the network layer to
 996          * hand us all the skbs that are available. iscsi_tcp_data_recv
 997          * handled pdus that cross buffers or pdus that still need data.
 998          */
 999         rd_desc.arg.data = conn;
1000         rd_desc.count = 1;
1001         tcp_read_sock(sk, &rd_desc, iscsi_tcp_data_recv);
1002
1003         read_unlock(&sk->sk_callback_lock);
1004 }
1005
1006 static void
1007 iscsi_tcp_state_change(struct sock *sk)
1008 {
1009         struct iscsi_tcp_conn *tcp_conn;
1010         struct iscsi_conn *conn;
1011         struct iscsi_session *session;
1012         void (*old_state_change)(struct sock *);
1013
1014         read_lock(&sk->sk_callback_lock);
1015
1016         conn = (struct iscsi_conn*)sk->sk_user_data;
1017         session = conn->session;
1018
1019         if ((sk->sk_state == TCP_CLOSE_WAIT ||
1020              sk->sk_state == TCP_CLOSE) &&
1021             !atomic_read(&sk->sk_rmem_alloc)) {
1022                 debug_tcp("iscsi_tcp_state_change: TCP_CLOSE|TCP_CLOSE_WAIT\n");
1023                 iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
1024         }
1025
1026         tcp_conn = conn->dd_data;
1027         old_state_change = tcp_conn->old_state_change;
1028
1029         read_unlock(&sk->sk_callback_lock);
1030
1031         old_state_change(sk);
1032 }
1033
1034 /**
1035  * iscsi_write_space - Called when more output buffer space is available
1036  * @sk: socket space is available for
1037  **/
1038 static void
1039 iscsi_write_space(struct sock *sk)
1040 {
1041         struct iscsi_conn *conn = (struct iscsi_conn*)sk->sk_user_data;
1042         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1043
1044         tcp_conn->old_write_space(sk);
1045         debug_tcp("iscsi_write_space: cid %d\n", conn->id);
1046         scsi_queue_work(conn->session->host, &conn->xmitwork);
1047 }
1048
1049 static void
1050 iscsi_conn_set_callbacks(struct iscsi_conn *conn)
1051 {
1052         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1053         struct sock *sk = tcp_conn->sock->sk;
1054
1055         /* assign new callbacks */
1056         write_lock_bh(&sk->sk_callback_lock);
1057         sk->sk_user_data = conn;
1058         tcp_conn->old_data_ready = sk->sk_data_ready;
1059         tcp_conn->old_state_change = sk->sk_state_change;
1060         tcp_conn->old_write_space = sk->sk_write_space;
1061         sk->sk_data_ready = iscsi_tcp_data_ready;
1062         sk->sk_state_change = iscsi_tcp_state_change;
1063         sk->sk_write_space = iscsi_write_space;
1064         write_unlock_bh(&sk->sk_callback_lock);
1065 }
1066
1067 static void
1068 iscsi_conn_restore_callbacks(struct iscsi_tcp_conn *tcp_conn)
1069 {
1070         struct sock *sk = tcp_conn->sock->sk;
1071
1072         /* restore socket callbacks, see also: iscsi_conn_set_callbacks() */
1073         write_lock_bh(&sk->sk_callback_lock);
1074         sk->sk_user_data    = NULL;
1075         sk->sk_data_ready   = tcp_conn->old_data_ready;
1076         sk->sk_state_change = tcp_conn->old_state_change;
1077         sk->sk_write_space  = tcp_conn->old_write_space;
1078         sk->sk_no_check  = 0;
1079         write_unlock_bh(&sk->sk_callback_lock);
1080 }
1081
1082 /**
1083  * iscsi_send - generic send routine
1084  * @sk: kernel's socket
1085  * @buf: buffer to write from
1086  * @size: actual size to write
1087  * @flags: socket's flags
1088  */
1089 static inline int
1090 iscsi_send(struct iscsi_conn *conn, struct iscsi_buf *buf, int size, int flags)
1091 {
1092         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1093         struct socket *sk = tcp_conn->sock;
1094         int offset = buf->sg.offset + buf->sent, res;
1095
1096         /*
1097          * if we got use_sg=0 or are sending something we kmallocd
1098          * then we did not have to do kmap (kmap returns page_address)
1099          *
1100          * if we got use_sg > 0, but had to drop down, we do not
1101          * set clustering so this should only happen for that
1102          * slab case.
1103          */
1104         if (buf->use_sendmsg)
1105                 res = sock_no_sendpage(sk, buf->sg.page, offset, size, flags);
1106         else
1107                 res = tcp_conn->sendpage(sk, buf->sg.page, offset, size, flags);
1108
1109         if (res >= 0) {
1110                 conn->txdata_octets += res;
1111                 buf->sent += res;
1112                 return res;
1113         }
1114
1115         tcp_conn->sendpage_failures_cnt++;
1116         if (res == -EAGAIN)
1117                 res = -ENOBUFS;
1118         else
1119                 iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
1120         return res;
1121 }
1122
1123 /**
1124  * iscsi_sendhdr - send PDU Header via tcp_sendpage()
1125  * @conn: iscsi connection
1126  * @buf: buffer to write from
1127  * @datalen: lenght of data to be sent after the header
1128  *
1129  * Notes:
1130  *      (Tx, Fast Path)
1131  **/
1132 static inline int
1133 iscsi_sendhdr(struct iscsi_conn *conn, struct iscsi_buf *buf, int datalen)
1134 {
1135         int flags = 0; /* MSG_DONTWAIT; */
1136         int res, size;
1137
1138         size = buf->sg.length - buf->sent;
1139         BUG_ON(buf->sent + size > buf->sg.length);
1140         if (buf->sent + size != buf->sg.length || datalen)
1141                 flags |= MSG_MORE;
1142
1143         res = iscsi_send(conn, buf, size, flags);
1144         debug_tcp("sendhdr %d bytes, sent %d res %d\n", size, buf->sent, res);
1145         if (res >= 0) {
1146                 if (size != res)
1147                         return -EAGAIN;
1148                 return 0;
1149         }
1150
1151         return res;
1152 }
1153
1154 /**
1155  * iscsi_sendpage - send one page of iSCSI Data-Out.
1156  * @conn: iscsi connection
1157  * @buf: buffer to write from
1158  * @count: remaining data
1159  * @sent: number of bytes sent
1160  *
1161  * Notes:
1162  *      (Tx, Fast Path)
1163  **/
1164 static inline int
1165 iscsi_sendpage(struct iscsi_conn *conn, struct iscsi_buf *buf,
1166                int *count, int *sent)
1167 {
1168         int flags = 0; /* MSG_DONTWAIT; */
1169         int res, size;
1170
1171         size = buf->sg.length - buf->sent;
1172         BUG_ON(buf->sent + size > buf->sg.length);
1173         if (size > *count)
1174                 size = *count;
1175         if (buf->sent + size != buf->sg.length || *count != size)
1176                 flags |= MSG_MORE;
1177
1178         res = iscsi_send(conn, buf, size, flags);
1179         debug_tcp("sendpage: %d bytes, sent %d left %d sent %d res %d\n",
1180                   size, buf->sent, *count, *sent, res);
1181         if (res >= 0) {
1182                 *count -= res;
1183                 *sent += res;
1184                 if (size != res)
1185                         return -EAGAIN;
1186                 return 0;
1187         }
1188
1189         return res;
1190 }
1191
1192 static inline void
1193 iscsi_data_digest_init(struct iscsi_tcp_conn *tcp_conn,
1194                       struct iscsi_tcp_cmd_task *tcp_ctask)
1195 {
1196         crypto_hash_init(&tcp_conn->tx_hash);
1197         tcp_ctask->digest_count = 4;
1198 }
1199
1200 /**
1201  * iscsi_solicit_data_cont - initialize next Data-Out
1202  * @conn: iscsi connection
1203  * @ctask: scsi command task
1204  * @r2t: R2T info
1205  * @left: bytes left to transfer
1206  *
1207  * Notes:
1208  *      Initialize next Data-Out within this R2T sequence and continue
1209  *      to process next Scatter-Gather element(if any) of this SCSI command.
1210  *
1211  *      Called under connection lock.
1212  **/
1213 static void
1214 iscsi_solicit_data_cont(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask,
1215                         struct iscsi_r2t_info *r2t, int left)
1216 {
1217         struct iscsi_data *hdr;
1218         struct scsi_cmnd *sc = ctask->sc;
1219         int new_offset;
1220
1221         hdr = &r2t->dtask.hdr;
1222         memset(hdr, 0, sizeof(struct iscsi_data));
1223         hdr->ttt = r2t->ttt;
1224         hdr->datasn = cpu_to_be32(r2t->solicit_datasn);
1225         r2t->solicit_datasn++;
1226         hdr->opcode = ISCSI_OP_SCSI_DATA_OUT;
1227         memcpy(hdr->lun, ctask->hdr->lun, sizeof(hdr->lun));
1228         hdr->itt = ctask->hdr->itt;
1229         hdr->exp_statsn = r2t->exp_statsn;
1230         new_offset = r2t->data_offset + r2t->sent;
1231         hdr->offset = cpu_to_be32(new_offset);
1232         if (left > conn->max_xmit_dlength) {
1233                 hton24(hdr->dlength, conn->max_xmit_dlength);
1234                 r2t->data_count = conn->max_xmit_dlength;
1235         } else {
1236                 hton24(hdr->dlength, left);
1237                 r2t->data_count = left;
1238                 hdr->flags = ISCSI_FLAG_CMD_FINAL;
1239         }
1240         conn->dataout_pdus_cnt++;
1241
1242         iscsi_buf_init_iov(&r2t->headbuf, (char*)hdr,
1243                            sizeof(struct iscsi_hdr));
1244
1245         if (iscsi_buf_left(&r2t->sendbuf))
1246                 return;
1247
1248         if (sc->use_sg) {
1249                 iscsi_buf_init_sg(&r2t->sendbuf, r2t->sg);
1250                 r2t->sg += 1;
1251         } else {
1252                 iscsi_buf_init_iov(&r2t->sendbuf,
1253                             (char*)sc->request_buffer + new_offset,
1254                             r2t->data_count);
1255                 r2t->sg = NULL;
1256         }
1257 }
1258
1259 static void iscsi_set_padding(struct iscsi_tcp_cmd_task *tcp_ctask,
1260                               unsigned long len)
1261 {
1262         tcp_ctask->pad_count = len & (ISCSI_PAD_LEN - 1);
1263         if (!tcp_ctask->pad_count)
1264                 return;
1265
1266         tcp_ctask->pad_count = ISCSI_PAD_LEN - tcp_ctask->pad_count;
1267         debug_scsi("write padding %d bytes\n", tcp_ctask->pad_count);
1268         tcp_ctask->xmstate |= XMSTATE_W_PAD;
1269 }
1270
1271 /**
1272  * iscsi_tcp_cmd_init - Initialize iSCSI SCSI_READ or SCSI_WRITE commands
1273  * @conn: iscsi connection
1274  * @ctask: scsi command task
1275  * @sc: scsi command
1276  **/
1277 static void
1278 iscsi_tcp_cmd_init(struct iscsi_cmd_task *ctask)
1279 {
1280         struct scsi_cmnd *sc = ctask->sc;
1281         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1282
1283         BUG_ON(__kfifo_len(tcp_ctask->r2tqueue));
1284
1285         tcp_ctask->sent = 0;
1286         tcp_ctask->sg_count = 0;
1287
1288         if (sc->sc_data_direction == DMA_TO_DEVICE) {
1289                 tcp_ctask->xmstate = XMSTATE_W_HDR;
1290                 tcp_ctask->exp_r2tsn = 0;
1291                 BUG_ON(ctask->total_length == 0);
1292
1293                 if (sc->use_sg) {
1294                         struct scatterlist *sg = sc->request_buffer;
1295
1296                         iscsi_buf_init_sg(&tcp_ctask->sendbuf, sg);
1297                         tcp_ctask->sg = sg + 1;
1298                         tcp_ctask->bad_sg = sg + sc->use_sg;
1299                 } else {
1300                         iscsi_buf_init_iov(&tcp_ctask->sendbuf,
1301                                            sc->request_buffer,
1302                                            sc->request_bufflen);
1303                         tcp_ctask->sg = NULL;
1304                         tcp_ctask->bad_sg = NULL;
1305                 }
1306                 debug_scsi("cmd [itt 0x%x total %d imm_data %d "
1307                            "unsol count %d, unsol offset %d]\n",
1308                            ctask->itt, ctask->total_length, ctask->imm_count,
1309                            ctask->unsol_count, ctask->unsol_offset);
1310         } else
1311                 tcp_ctask->xmstate = XMSTATE_R_HDR;
1312
1313         iscsi_buf_init_iov(&tcp_ctask->headbuf, (char*)ctask->hdr,
1314                             sizeof(struct iscsi_hdr));
1315 }
1316
1317 /**
1318  * iscsi_tcp_mtask_xmit - xmit management(immediate) task
1319  * @conn: iscsi connection
1320  * @mtask: task management task
1321  *
1322  * Notes:
1323  *      The function can return -EAGAIN in which case caller must
1324  *      call it again later, or recover. '0' return code means successful
1325  *      xmit.
1326  *
1327  *      Management xmit state machine consists of two states:
1328  *              IN_PROGRESS_IMM_HEAD - PDU Header xmit in progress
1329  *              IN_PROGRESS_IMM_DATA - PDU Data xmit in progress
1330  **/
1331 static int
1332 iscsi_tcp_mtask_xmit(struct iscsi_conn *conn, struct iscsi_mgmt_task *mtask)
1333 {
1334         struct iscsi_tcp_mgmt_task *tcp_mtask = mtask->dd_data;
1335         int rc;
1336
1337         debug_scsi("mtask deq [cid %d state %x itt 0x%x]\n",
1338                 conn->id, tcp_mtask->xmstate, mtask->itt);
1339
1340         if (tcp_mtask->xmstate & XMSTATE_IMM_HDR) {
1341                 tcp_mtask->xmstate &= ~XMSTATE_IMM_HDR;
1342                 if (mtask->data_count)
1343                         tcp_mtask->xmstate |= XMSTATE_IMM_DATA;
1344                 if (conn->c_stage != ISCSI_CONN_INITIAL_STAGE &&
1345                     conn->stop_stage != STOP_CONN_RECOVER &&
1346                     conn->hdrdgst_en)
1347                         iscsi_hdr_digest(conn, &tcp_mtask->headbuf,
1348                                         (u8*)tcp_mtask->hdrext);
1349                 rc = iscsi_sendhdr(conn, &tcp_mtask->headbuf,
1350                                    mtask->data_count);
1351                 if (rc) {
1352                         tcp_mtask->xmstate |= XMSTATE_IMM_HDR;
1353                         if (mtask->data_count)
1354                                 tcp_mtask->xmstate &= ~XMSTATE_IMM_DATA;
1355                         return rc;
1356                 }
1357         }
1358
1359         if (tcp_mtask->xmstate & XMSTATE_IMM_DATA) {
1360                 BUG_ON(!mtask->data_count);
1361                 tcp_mtask->xmstate &= ~XMSTATE_IMM_DATA;
1362                 /* FIXME: implement.
1363                  * Virtual buffer could be spreaded across multiple pages...
1364                  */
1365                 do {
1366                         int rc;
1367
1368                         rc = iscsi_sendpage(conn, &tcp_mtask->sendbuf,
1369                                         &mtask->data_count, &tcp_mtask->sent);
1370                         if (rc) {
1371                                 tcp_mtask->xmstate |= XMSTATE_IMM_DATA;
1372                                 return rc;
1373                         }
1374                 } while (mtask->data_count);
1375         }
1376
1377         BUG_ON(tcp_mtask->xmstate != XMSTATE_IDLE);
1378         if (mtask->hdr->itt == RESERVED_ITT) {
1379                 struct iscsi_session *session = conn->session;
1380
1381                 spin_lock_bh(&session->lock);
1382                 list_del(&conn->mtask->running);
1383                 __kfifo_put(session->mgmtpool.queue, (void*)&conn->mtask,
1384                             sizeof(void*));
1385                 spin_unlock_bh(&session->lock);
1386         }
1387         return 0;
1388 }
1389
1390 static inline int
1391 iscsi_send_read_hdr(struct iscsi_conn *conn,
1392                     struct iscsi_tcp_cmd_task *tcp_ctask)
1393 {
1394         int rc;
1395
1396         tcp_ctask->xmstate &= ~XMSTATE_R_HDR;
1397         if (conn->hdrdgst_en)
1398                 iscsi_hdr_digest(conn, &tcp_ctask->headbuf,
1399                                  (u8*)tcp_ctask->hdrext);
1400         rc = iscsi_sendhdr(conn, &tcp_ctask->headbuf, 0);
1401         if (!rc) {
1402                 BUG_ON(tcp_ctask->xmstate != XMSTATE_IDLE);
1403                 return 0; /* wait for Data-In */
1404         }
1405         tcp_ctask->xmstate |= XMSTATE_R_HDR;
1406         return rc;
1407 }
1408
1409 static inline int
1410 iscsi_send_write_hdr(struct iscsi_conn *conn,
1411                      struct iscsi_cmd_task *ctask)
1412 {
1413         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1414         int rc;
1415
1416         tcp_ctask->xmstate &= ~XMSTATE_W_HDR;
1417         if (conn->hdrdgst_en)
1418                 iscsi_hdr_digest(conn, &tcp_ctask->headbuf,
1419                                  (u8*)tcp_ctask->hdrext);
1420         rc = iscsi_sendhdr(conn, &tcp_ctask->headbuf, ctask->imm_count);
1421         if (rc) {
1422                 tcp_ctask->xmstate |= XMSTATE_W_HDR;
1423                 return rc;
1424         }
1425
1426         if (ctask->imm_count) {
1427                 tcp_ctask->xmstate |= XMSTATE_IMM_DATA;
1428                 iscsi_set_padding(tcp_ctask, ctask->imm_count);
1429
1430                 if (ctask->conn->datadgst_en) {
1431                         iscsi_data_digest_init(ctask->conn->dd_data, tcp_ctask);
1432                         tcp_ctask->immdigest = 0;
1433                 }
1434         }
1435
1436         if (ctask->unsol_count)
1437                 tcp_ctask->xmstate |= XMSTATE_UNS_HDR | XMSTATE_UNS_INIT;
1438         return 0;
1439 }
1440
1441 static int
1442 iscsi_send_padding(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1443 {
1444         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1445         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1446         int sent = 0, rc;
1447
1448         if (tcp_ctask->xmstate & XMSTATE_W_PAD) {
1449                 iscsi_buf_init_iov(&tcp_ctask->sendbuf, (char*)&tcp_ctask->pad,
1450                                    tcp_ctask->pad_count);
1451                 if (conn->datadgst_en)
1452                         crypto_hash_update(&tcp_conn->tx_hash,
1453                                            &tcp_ctask->sendbuf.sg,
1454                                            tcp_ctask->sendbuf.sg.length);
1455         } else if (!(tcp_ctask->xmstate & XMSTATE_W_RESEND_PAD))
1456                 return 0;
1457
1458         tcp_ctask->xmstate &= ~XMSTATE_W_PAD;
1459         tcp_ctask->xmstate &= ~XMSTATE_W_RESEND_PAD;
1460         debug_scsi("sending %d pad bytes for itt 0x%x\n",
1461                    tcp_ctask->pad_count, ctask->itt);
1462         rc = iscsi_sendpage(conn, &tcp_ctask->sendbuf, &tcp_ctask->pad_count,
1463                            &sent);
1464         if (rc) {
1465                 debug_scsi("padding send failed %d\n", rc);
1466                 tcp_ctask->xmstate |= XMSTATE_W_RESEND_PAD;
1467         }
1468         return rc;
1469 }
1470
1471 static int
1472 iscsi_send_digest(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask,
1473                         struct iscsi_buf *buf, uint32_t *digest)
1474 {
1475         struct iscsi_tcp_cmd_task *tcp_ctask;
1476         struct iscsi_tcp_conn *tcp_conn;
1477         int rc, sent = 0;
1478
1479         if (!conn->datadgst_en)
1480                 return 0;
1481
1482         tcp_ctask = ctask->dd_data;
1483         tcp_conn = conn->dd_data;
1484
1485         if (!(tcp_ctask->xmstate & XMSTATE_W_RESEND_DATA_DIGEST)) {
1486                 crypto_hash_final(&tcp_conn->tx_hash, (u8*)digest);
1487                 iscsi_buf_init_iov(buf, (char*)digest, 4);
1488         }
1489         tcp_ctask->xmstate &= ~XMSTATE_W_RESEND_DATA_DIGEST;
1490
1491         rc = iscsi_sendpage(conn, buf, &tcp_ctask->digest_count, &sent);
1492         if (!rc)
1493                 debug_scsi("sent digest 0x%x for itt 0x%x\n", *digest,
1494                           ctask->itt);
1495         else {
1496                 debug_scsi("sending digest 0x%x failed for itt 0x%x!\n",
1497                           *digest, ctask->itt);
1498                 tcp_ctask->xmstate |= XMSTATE_W_RESEND_DATA_DIGEST;
1499         }
1500         return rc;
1501 }
1502
1503 static int
1504 iscsi_send_data(struct iscsi_cmd_task *ctask, struct iscsi_buf *sendbuf,
1505                 struct scatterlist **sg, int *sent, int *count,
1506                 struct iscsi_buf *digestbuf, uint32_t *digest)
1507 {
1508         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1509         struct iscsi_conn *conn = ctask->conn;
1510         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1511         int rc, buf_sent, offset;
1512
1513         while (*count) {
1514                 buf_sent = 0;
1515                 offset = sendbuf->sent;
1516
1517                 rc = iscsi_sendpage(conn, sendbuf, count, &buf_sent);
1518                 *sent = *sent + buf_sent;
1519                 if (buf_sent && conn->datadgst_en)
1520                         partial_sg_digest_update(&tcp_conn->tx_hash,
1521                                 &sendbuf->sg, sendbuf->sg.offset + offset,
1522                                 buf_sent);
1523                 if (!iscsi_buf_left(sendbuf) && *sg != tcp_ctask->bad_sg) {
1524                         iscsi_buf_init_sg(sendbuf, *sg);
1525                         *sg = *sg + 1;
1526                 }
1527
1528                 if (rc)
1529                         return rc;
1530         }
1531
1532         rc = iscsi_send_padding(conn, ctask);
1533         if (rc)
1534                 return rc;
1535
1536         return iscsi_send_digest(conn, ctask, digestbuf, digest);
1537 }
1538
1539 static int
1540 iscsi_send_unsol_hdr(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1541 {
1542         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1543         struct iscsi_data_task *dtask;
1544         int rc;
1545
1546         tcp_ctask->xmstate |= XMSTATE_UNS_DATA;
1547         if (tcp_ctask->xmstate & XMSTATE_UNS_INIT) {
1548                 dtask = &tcp_ctask->unsol_dtask;
1549
1550                 iscsi_prep_unsolicit_data_pdu(ctask, &dtask->hdr);
1551                 iscsi_buf_init_iov(&tcp_ctask->headbuf, (char*)&dtask->hdr,
1552                                    sizeof(struct iscsi_hdr));
1553                 if (conn->hdrdgst_en)
1554                         iscsi_hdr_digest(conn, &tcp_ctask->headbuf,
1555                                         (u8*)dtask->hdrext);
1556
1557                 tcp_ctask->xmstate &= ~XMSTATE_UNS_INIT;
1558                 iscsi_set_padding(tcp_ctask, ctask->data_count);
1559         }
1560
1561         rc = iscsi_sendhdr(conn, &tcp_ctask->headbuf, ctask->data_count);
1562         if (rc) {
1563                 tcp_ctask->xmstate &= ~XMSTATE_UNS_DATA;
1564                 tcp_ctask->xmstate |= XMSTATE_UNS_HDR;
1565                 return rc;
1566         }
1567
1568         if (conn->datadgst_en) {
1569                 dtask = &tcp_ctask->unsol_dtask;
1570                 iscsi_data_digest_init(ctask->conn->dd_data, tcp_ctask);
1571                 dtask->digest = 0;
1572         }
1573
1574         debug_scsi("uns dout [itt 0x%x dlen %d sent %d]\n",
1575                    ctask->itt, ctask->unsol_count, tcp_ctask->sent);
1576         return 0;
1577 }
1578
1579 static int
1580 iscsi_send_unsol_pdu(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1581 {
1582         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1583         int rc;
1584
1585         if (tcp_ctask->xmstate & XMSTATE_UNS_HDR) {
1586                 BUG_ON(!ctask->unsol_count);
1587                 tcp_ctask->xmstate &= ~XMSTATE_UNS_HDR;
1588 send_hdr:
1589                 rc = iscsi_send_unsol_hdr(conn, ctask);
1590                 if (rc)
1591                         return rc;
1592         }
1593
1594         if (tcp_ctask->xmstate & XMSTATE_UNS_DATA) {
1595                 struct iscsi_data_task *dtask = &tcp_ctask->unsol_dtask;
1596                 int start = tcp_ctask->sent;
1597
1598                 rc = iscsi_send_data(ctask, &tcp_ctask->sendbuf, &tcp_ctask->sg,
1599                                      &tcp_ctask->sent, &ctask->data_count,
1600                                      &dtask->digestbuf, &dtask->digest);
1601                 ctask->unsol_count -= tcp_ctask->sent - start;
1602                 if (rc)
1603                         return rc;
1604                 tcp_ctask->xmstate &= ~XMSTATE_UNS_DATA;
1605                 /*
1606                  * Done with the Data-Out. Next, check if we need
1607                  * to send another unsolicited Data-Out.
1608                  */
1609                 if (ctask->unsol_count) {
1610                         debug_scsi("sending more uns\n");
1611                         tcp_ctask->xmstate |= XMSTATE_UNS_INIT;
1612                         goto send_hdr;
1613                 }
1614         }
1615         return 0;
1616 }
1617
1618 static int iscsi_send_sol_pdu(struct iscsi_conn *conn,
1619                               struct iscsi_cmd_task *ctask)
1620 {
1621         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1622         struct iscsi_session *session = conn->session;
1623         struct iscsi_r2t_info *r2t;
1624         struct iscsi_data_task *dtask;
1625         int left, rc;
1626
1627         if (tcp_ctask->xmstate & XMSTATE_SOL_HDR) {
1628                 tcp_ctask->xmstate &= ~XMSTATE_SOL_HDR;
1629                 tcp_ctask->xmstate |= XMSTATE_SOL_DATA;
1630                 if (!tcp_ctask->r2t) {
1631                         spin_lock_bh(&session->lock);
1632                         __kfifo_get(tcp_ctask->r2tqueue, (void*)&tcp_ctask->r2t,
1633                                     sizeof(void*));
1634                         spin_unlock_bh(&session->lock);
1635                 }
1636 send_hdr:
1637                 r2t = tcp_ctask->r2t;
1638                 dtask = &r2t->dtask;
1639
1640                 if (conn->hdrdgst_en)
1641                         iscsi_hdr_digest(conn, &r2t->headbuf,
1642                                         (u8*)dtask->hdrext);
1643                 rc = iscsi_sendhdr(conn, &r2t->headbuf, r2t->data_count);
1644                 if (rc) {
1645                         tcp_ctask->xmstate &= ~XMSTATE_SOL_DATA;
1646                         tcp_ctask->xmstate |= XMSTATE_SOL_HDR;
1647                         return rc;
1648                 }
1649
1650                 if (conn->datadgst_en) {
1651                         iscsi_data_digest_init(conn->dd_data, tcp_ctask);
1652                         dtask->digest = 0;
1653                 }
1654
1655                 iscsi_set_padding(tcp_ctask, r2t->data_count);
1656                 debug_scsi("sol dout [dsn %d itt 0x%x dlen %d sent %d]\n",
1657                         r2t->solicit_datasn - 1, ctask->itt, r2t->data_count,
1658                         r2t->sent);
1659         }
1660
1661         if (tcp_ctask->xmstate & XMSTATE_SOL_DATA) {
1662                 r2t = tcp_ctask->r2t;
1663                 dtask = &r2t->dtask;
1664
1665                 rc = iscsi_send_data(ctask, &r2t->sendbuf, &r2t->sg,
1666                                      &r2t->sent, &r2t->data_count,
1667                                      &dtask->digestbuf, &dtask->digest);
1668                 if (rc)
1669                         return rc;
1670                 tcp_ctask->xmstate &= ~XMSTATE_SOL_DATA;
1671
1672                 /*
1673                  * Done with this Data-Out. Next, check if we have
1674                  * to send another Data-Out for this R2T.
1675                  */
1676                 BUG_ON(r2t->data_length - r2t->sent < 0);
1677                 left = r2t->data_length - r2t->sent;
1678                 if (left) {
1679                         iscsi_solicit_data_cont(conn, ctask, r2t, left);
1680                         tcp_ctask->xmstate |= XMSTATE_SOL_DATA;
1681                         tcp_ctask->xmstate &= ~XMSTATE_SOL_HDR;
1682                         goto send_hdr;
1683                 }
1684
1685                 /*
1686                  * Done with this R2T. Check if there are more
1687                  * outstanding R2Ts ready to be processed.
1688                  */
1689                 spin_lock_bh(&session->lock);
1690                 tcp_ctask->r2t = NULL;
1691                 __kfifo_put(tcp_ctask->r2tpool.queue, (void*)&r2t,
1692                             sizeof(void*));
1693                 if (__kfifo_get(tcp_ctask->r2tqueue, (void*)&r2t,
1694                                 sizeof(void*))) {
1695                         tcp_ctask->r2t = r2t;
1696                         tcp_ctask->xmstate |= XMSTATE_SOL_DATA;
1697                         tcp_ctask->xmstate &= ~XMSTATE_SOL_HDR;
1698                         spin_unlock_bh(&session->lock);
1699                         goto send_hdr;
1700                 }
1701                 spin_unlock_bh(&session->lock);
1702         }
1703         return 0;
1704 }
1705
1706 static int
1707 iscsi_tcp_ctask_xmit(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1708 {
1709         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1710         int rc = 0;
1711
1712         debug_scsi("ctask deq [cid %d xmstate %x itt 0x%x]\n",
1713                 conn->id, tcp_ctask->xmstate, ctask->itt);
1714
1715         /*
1716          * serialize with TMF AbortTask
1717          */
1718         if (ctask->mtask)
1719                 return rc;
1720
1721         if (tcp_ctask->xmstate & XMSTATE_R_HDR)
1722                 return iscsi_send_read_hdr(conn, tcp_ctask);
1723
1724         if (tcp_ctask->xmstate & XMSTATE_W_HDR) {
1725                 rc = iscsi_send_write_hdr(conn, ctask);
1726                 if (rc)
1727                         return rc;
1728         }
1729
1730         if (tcp_ctask->xmstate & XMSTATE_IMM_DATA) {
1731                 rc = iscsi_send_data(ctask, &tcp_ctask->sendbuf, &tcp_ctask->sg,
1732                                      &tcp_ctask->sent, &ctask->imm_count,
1733                                      &tcp_ctask->immbuf, &tcp_ctask->immdigest);
1734                 if (rc)
1735                         return rc;
1736                 tcp_ctask->xmstate &= ~XMSTATE_IMM_DATA;
1737         }
1738
1739         rc = iscsi_send_unsol_pdu(conn, ctask);
1740         if (rc)
1741                 return rc;
1742
1743         rc = iscsi_send_sol_pdu(conn, ctask);
1744         if (rc)
1745                 return rc;
1746
1747         return rc;
1748 }
1749
1750 static struct iscsi_cls_conn *
1751 iscsi_tcp_conn_create(struct iscsi_cls_session *cls_session, uint32_t conn_idx)
1752 {
1753         struct iscsi_conn *conn;
1754         struct iscsi_cls_conn *cls_conn;
1755         struct iscsi_tcp_conn *tcp_conn;
1756
1757         cls_conn = iscsi_conn_setup(cls_session, conn_idx);
1758         if (!cls_conn)
1759                 return NULL;
1760         conn = cls_conn->dd_data;
1761         /*
1762          * due to strange issues with iser these are not set
1763          * in iscsi_conn_setup
1764          */
1765         conn->max_recv_dlength = ISCSI_DEF_MAX_RECV_SEG_LEN;
1766
1767         tcp_conn = kzalloc(sizeof(*tcp_conn), GFP_KERNEL);
1768         if (!tcp_conn)
1769                 goto tcp_conn_alloc_fail;
1770
1771         conn->dd_data = tcp_conn;
1772         tcp_conn->iscsi_conn = conn;
1773         tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
1774         /* initial operational parameters */
1775         tcp_conn->hdr_size = sizeof(struct iscsi_hdr);
1776
1777         tcp_conn->tx_hash.tfm = crypto_alloc_hash("crc32c", 0,
1778                                                   CRYPTO_ALG_ASYNC);
1779         tcp_conn->tx_hash.flags = 0;
1780         if (IS_ERR(tcp_conn->tx_hash.tfm))
1781                 goto free_tcp_conn;
1782
1783         tcp_conn->rx_hash.tfm = crypto_alloc_hash("crc32c", 0,
1784                                                   CRYPTO_ALG_ASYNC);
1785         tcp_conn->rx_hash.flags = 0;
1786         if (IS_ERR(tcp_conn->rx_hash.tfm))
1787                 goto free_tx_tfm;
1788
1789         return cls_conn;
1790
1791 free_tx_tfm:
1792         crypto_free_hash(tcp_conn->tx_hash.tfm);
1793 free_tcp_conn:
1794         kfree(tcp_conn);
1795 tcp_conn_alloc_fail:
1796         iscsi_conn_teardown(cls_conn);
1797         return NULL;
1798 }
1799
1800 static void
1801 iscsi_tcp_release_conn(struct iscsi_conn *conn)
1802 {
1803         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1804
1805         if (!tcp_conn->sock)
1806                 return;
1807
1808         sock_hold(tcp_conn->sock->sk);
1809         iscsi_conn_restore_callbacks(tcp_conn);
1810         sock_put(tcp_conn->sock->sk);
1811
1812         sock_release(tcp_conn->sock);
1813         tcp_conn->sock = NULL;
1814         conn->recv_lock = NULL;
1815 }
1816
1817 static void
1818 iscsi_tcp_conn_destroy(struct iscsi_cls_conn *cls_conn)
1819 {
1820         struct iscsi_conn *conn = cls_conn->dd_data;
1821         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1822
1823         iscsi_tcp_release_conn(conn);
1824         iscsi_conn_teardown(cls_conn);
1825
1826         if (tcp_conn->tx_hash.tfm)
1827                 crypto_free_hash(tcp_conn->tx_hash.tfm);
1828         if (tcp_conn->rx_hash.tfm)
1829                 crypto_free_hash(tcp_conn->rx_hash.tfm);
1830
1831         kfree(tcp_conn);
1832 }
1833
1834 static void
1835 iscsi_tcp_conn_stop(struct iscsi_cls_conn *cls_conn, int flag)
1836 {
1837         struct iscsi_conn *conn = cls_conn->dd_data;
1838         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1839
1840         iscsi_conn_stop(cls_conn, flag);
1841         iscsi_tcp_release_conn(conn);
1842         tcp_conn->hdr_size = sizeof(struct iscsi_hdr);
1843 }
1844
1845 static int
1846 iscsi_tcp_conn_bind(struct iscsi_cls_session *cls_session,
1847                     struct iscsi_cls_conn *cls_conn, uint64_t transport_eph,
1848                     int is_leading)
1849 {
1850         struct iscsi_conn *conn = cls_conn->dd_data;
1851         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1852         struct sock *sk;
1853         struct socket *sock;
1854         int err;
1855
1856         /* lookup for existing socket */
1857         sock = sockfd_lookup((int)transport_eph, &err);
1858         if (!sock) {
1859                 printk(KERN_ERR "iscsi_tcp: sockfd_lookup failed %d\n", err);
1860                 return -EEXIST;
1861         }
1862
1863         err = iscsi_conn_bind(cls_session, cls_conn, is_leading);
1864         if (err)
1865                 return err;
1866
1867         /* bind iSCSI connection and socket */
1868         tcp_conn->sock = sock;
1869
1870         /* setup Socket parameters */
1871         sk = sock->sk;
1872         sk->sk_reuse = 1;
1873         sk->sk_sndtimeo = 15 * HZ; /* FIXME: make it configurable */
1874         sk->sk_allocation = GFP_ATOMIC;
1875
1876         /* FIXME: disable Nagle's algorithm */
1877
1878         /*
1879          * Intercept TCP callbacks for sendfile like receive
1880          * processing.
1881          */
1882         conn->recv_lock = &sk->sk_callback_lock;
1883         iscsi_conn_set_callbacks(conn);
1884         tcp_conn->sendpage = tcp_conn->sock->ops->sendpage;
1885         /*
1886          * set receive state machine into initial state
1887          */
1888         tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
1889
1890         return 0;
1891 }
1892
1893 /* called with host lock */
1894 static void
1895 iscsi_tcp_mgmt_init(struct iscsi_conn *conn, struct iscsi_mgmt_task *mtask,
1896                     char *data, uint32_t data_size)
1897 {
1898         struct iscsi_tcp_mgmt_task *tcp_mtask = mtask->dd_data;
1899
1900         iscsi_buf_init_iov(&tcp_mtask->headbuf, (char*)mtask->hdr,
1901                            sizeof(struct iscsi_hdr));
1902         tcp_mtask->xmstate = XMSTATE_IMM_HDR;
1903         tcp_mtask->sent = 0;
1904
1905         if (mtask->data_count)
1906                 iscsi_buf_init_iov(&tcp_mtask->sendbuf, (char*)mtask->data,
1907                                     mtask->data_count);
1908 }
1909
1910 static int
1911 iscsi_r2tpool_alloc(struct iscsi_session *session)
1912 {
1913         int i;
1914         int cmd_i;
1915
1916         /*
1917          * initialize per-task: R2T pool and xmit queue
1918          */
1919         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
1920                 struct iscsi_cmd_task *ctask = session->cmds[cmd_i];
1921                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1922
1923                 /*
1924                  * pre-allocated x4 as much r2ts to handle race when
1925                  * target acks DataOut faster than we data_xmit() queues
1926                  * could replenish r2tqueue.
1927                  */
1928
1929                 /* R2T pool */
1930                 if (iscsi_pool_init(&tcp_ctask->r2tpool, session->max_r2t * 4,
1931                                     (void***)&tcp_ctask->r2ts,
1932                                     sizeof(struct iscsi_r2t_info))) {
1933                         goto r2t_alloc_fail;
1934                 }
1935
1936                 /* R2T xmit queue */
1937                 tcp_ctask->r2tqueue = kfifo_alloc(
1938                       session->max_r2t * 4 * sizeof(void*), GFP_KERNEL, NULL);
1939                 if (tcp_ctask->r2tqueue == ERR_PTR(-ENOMEM)) {
1940                         iscsi_pool_free(&tcp_ctask->r2tpool,
1941                                         (void**)tcp_ctask->r2ts);
1942                         goto r2t_alloc_fail;
1943                 }
1944         }
1945
1946         return 0;
1947
1948 r2t_alloc_fail:
1949         for (i = 0; i < cmd_i; i++) {
1950                 struct iscsi_cmd_task *ctask = session->cmds[i];
1951                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1952
1953                 kfifo_free(tcp_ctask->r2tqueue);
1954                 iscsi_pool_free(&tcp_ctask->r2tpool,
1955                                 (void**)tcp_ctask->r2ts);
1956         }
1957         return -ENOMEM;
1958 }
1959
1960 static void
1961 iscsi_r2tpool_free(struct iscsi_session *session)
1962 {
1963         int i;
1964
1965         for (i = 0; i < session->cmds_max; i++) {
1966                 struct iscsi_cmd_task *ctask = session->cmds[i];
1967                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1968
1969                 kfifo_free(tcp_ctask->r2tqueue);
1970                 iscsi_pool_free(&tcp_ctask->r2tpool,
1971                                 (void**)tcp_ctask->r2ts);
1972         }
1973 }
1974
1975 static int
1976 iscsi_conn_set_param(struct iscsi_cls_conn *cls_conn, enum iscsi_param param,
1977                      char *buf, int buflen)
1978 {
1979         struct iscsi_conn *conn = cls_conn->dd_data;
1980         struct iscsi_session *session = conn->session;
1981         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1982         int value;
1983
1984         switch(param) {
1985         case ISCSI_PARAM_HDRDGST_EN:
1986                 iscsi_set_param(cls_conn, param, buf, buflen);
1987                 tcp_conn->hdr_size = sizeof(struct iscsi_hdr);
1988                 if (conn->hdrdgst_en)
1989                         tcp_conn->hdr_size += sizeof(__u32);
1990                 break;
1991         case ISCSI_PARAM_DATADGST_EN:
1992                 iscsi_set_param(cls_conn, param, buf, buflen);
1993                 tcp_conn->sendpage = conn->datadgst_en ?
1994                         sock_no_sendpage : tcp_conn->sock->ops->sendpage;
1995                 break;
1996         case ISCSI_PARAM_MAX_R2T:
1997                 sscanf(buf, "%d", &value);
1998                 if (session->max_r2t == roundup_pow_of_two(value))
1999                         break;
2000                 iscsi_r2tpool_free(session);
2001                 iscsi_set_param(cls_conn, param, buf, buflen);
2002                 if (session->max_r2t & (session->max_r2t - 1))
2003                         session->max_r2t = roundup_pow_of_two(session->max_r2t);
2004                 if (iscsi_r2tpool_alloc(session))
2005                         return -ENOMEM;
2006                 break;
2007         default:
2008                 return iscsi_set_param(cls_conn, param, buf, buflen);
2009         }
2010
2011         return 0;
2012 }
2013
2014 static int
2015 iscsi_tcp_conn_get_param(struct iscsi_cls_conn *cls_conn,
2016                          enum iscsi_param param, char *buf)
2017 {
2018         struct iscsi_conn *conn = cls_conn->dd_data;
2019         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
2020         struct inet_sock *inet;
2021         struct ipv6_pinfo *np;
2022         struct sock *sk;
2023         int len;
2024
2025         switch(param) {
2026         case ISCSI_PARAM_CONN_PORT:
2027                 mutex_lock(&conn->xmitmutex);
2028                 if (!tcp_conn->sock) {
2029                         mutex_unlock(&conn->xmitmutex);
2030                         return -EINVAL;
2031                 }
2032
2033                 inet = inet_sk(tcp_conn->sock->sk);
2034                 len = sprintf(buf, "%hu\n", be16_to_cpu(inet->dport));
2035                 mutex_unlock(&conn->xmitmutex);
2036                 break;
2037         case ISCSI_PARAM_CONN_ADDRESS:
2038                 mutex_lock(&conn->xmitmutex);
2039                 if (!tcp_conn->sock) {
2040                         mutex_unlock(&conn->xmitmutex);
2041                         return -EINVAL;
2042                 }
2043
2044                 sk = tcp_conn->sock->sk;
2045                 if (sk->sk_family == PF_INET) {
2046                         inet = inet_sk(sk);
2047                         len = sprintf(buf, NIPQUAD_FMT "\n",
2048                                       NIPQUAD(inet->daddr));
2049                 } else {
2050                         np = inet6_sk(sk);
2051                         len = sprintf(buf, NIP6_FMT "\n", NIP6(np->daddr));
2052                 }
2053                 mutex_unlock(&conn->xmitmutex);
2054                 break;
2055         default:
2056                 return iscsi_conn_get_param(cls_conn, param, buf);
2057         }
2058
2059         return len;
2060 }
2061
2062 static void
2063 iscsi_conn_get_stats(struct iscsi_cls_conn *cls_conn, struct iscsi_stats *stats)
2064 {
2065         struct iscsi_conn *conn = cls_conn->dd_data;
2066         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
2067
2068         stats->txdata_octets = conn->txdata_octets;
2069         stats->rxdata_octets = conn->rxdata_octets;
2070         stats->scsicmd_pdus = conn->scsicmd_pdus_cnt;
2071         stats->dataout_pdus = conn->dataout_pdus_cnt;
2072         stats->scsirsp_pdus = conn->scsirsp_pdus_cnt;
2073         stats->datain_pdus = conn->datain_pdus_cnt;
2074         stats->r2t_pdus = conn->r2t_pdus_cnt;
2075         stats->tmfcmd_pdus = conn->tmfcmd_pdus_cnt;
2076         stats->tmfrsp_pdus = conn->tmfrsp_pdus_cnt;
2077         stats->custom_length = 3;
2078         strcpy(stats->custom[0].desc, "tx_sendpage_failures");
2079         stats->custom[0].value = tcp_conn->sendpage_failures_cnt;
2080         strcpy(stats->custom[1].desc, "rx_discontiguous_hdr");
2081         stats->custom[1].value = tcp_conn->discontiguous_hdr_cnt;
2082         strcpy(stats->custom[2].desc, "eh_abort_cnt");
2083         stats->custom[2].value = conn->eh_abort_cnt;
2084 }
2085
2086 static struct iscsi_cls_session *
2087 iscsi_tcp_session_create(struct iscsi_transport *iscsit,
2088                          struct scsi_transport_template *scsit,
2089                          uint32_t initial_cmdsn, uint32_t *hostno)
2090 {
2091         struct iscsi_cls_session *cls_session;
2092         struct iscsi_session *session;
2093         uint32_t hn;
2094         int cmd_i;
2095
2096         cls_session = iscsi_session_setup(iscsit, scsit,
2097                                          sizeof(struct iscsi_tcp_cmd_task),
2098                                          sizeof(struct iscsi_tcp_mgmt_task),
2099                                          initial_cmdsn, &hn);
2100         if (!cls_session)
2101                 return NULL;
2102         *hostno = hn;
2103
2104         session = class_to_transport_session(cls_session);
2105         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
2106                 struct iscsi_cmd_task *ctask = session->cmds[cmd_i];
2107                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
2108
2109                 ctask->hdr = &tcp_ctask->hdr;
2110         }
2111
2112         for (cmd_i = 0; cmd_i < session->mgmtpool_max; cmd_i++) {
2113                 struct iscsi_mgmt_task *mtask = session->mgmt_cmds[cmd_i];
2114                 struct iscsi_tcp_mgmt_task *tcp_mtask = mtask->dd_data;
2115
2116                 mtask->hdr = &tcp_mtask->hdr;
2117         }
2118
2119         if (iscsi_r2tpool_alloc(class_to_transport_session(cls_session)))
2120                 goto r2tpool_alloc_fail;
2121
2122         return cls_session;
2123
2124 r2tpool_alloc_fail:
2125         iscsi_session_teardown(cls_session);
2126         return NULL;
2127 }
2128
2129 static void iscsi_tcp_session_destroy(struct iscsi_cls_session *cls_session)
2130 {
2131         iscsi_r2tpool_free(class_to_transport_session(cls_session));
2132         iscsi_session_teardown(cls_session);
2133 }
2134
2135 static struct scsi_host_template iscsi_sht = {
2136         .name                   = "iSCSI Initiator over TCP/IP",
2137         .queuecommand           = iscsi_queuecommand,
2138         .change_queue_depth     = iscsi_change_queue_depth,
2139         .can_queue              = ISCSI_XMIT_CMDS_MAX - 1,
2140         .sg_tablesize           = ISCSI_SG_TABLESIZE,
2141         .max_sectors            = 0xFFFF,
2142         .cmd_per_lun            = ISCSI_DEF_CMD_PER_LUN,
2143         .eh_abort_handler       = iscsi_eh_abort,
2144         .eh_host_reset_handler  = iscsi_eh_host_reset,
2145         .use_clustering         = DISABLE_CLUSTERING,
2146         .proc_name              = "iscsi_tcp",
2147         .this_id                = -1,
2148 };
2149
2150 static struct iscsi_transport iscsi_tcp_transport = {
2151         .owner                  = THIS_MODULE,
2152         .name                   = "tcp",
2153         .caps                   = CAP_RECOVERY_L0 | CAP_MULTI_R2T | CAP_HDRDGST
2154                                   | CAP_DATADGST,
2155         .param_mask             = ISCSI_MAX_RECV_DLENGTH |
2156                                   ISCSI_MAX_XMIT_DLENGTH |
2157                                   ISCSI_HDRDGST_EN |
2158                                   ISCSI_DATADGST_EN |
2159                                   ISCSI_INITIAL_R2T_EN |
2160                                   ISCSI_MAX_R2T |
2161                                   ISCSI_IMM_DATA_EN |
2162                                   ISCSI_FIRST_BURST |
2163                                   ISCSI_MAX_BURST |
2164                                   ISCSI_PDU_INORDER_EN |
2165                                   ISCSI_DATASEQ_INORDER_EN |
2166                                   ISCSI_ERL |
2167                                   ISCSI_CONN_PORT |
2168                                   ISCSI_CONN_ADDRESS |
2169                                   ISCSI_EXP_STATSN |
2170                                   ISCSI_PERSISTENT_PORT |
2171                                   ISCSI_PERSISTENT_ADDRESS |
2172                                   ISCSI_TARGET_NAME |
2173                                   ISCSI_TPGT,
2174         .host_template          = &iscsi_sht,
2175         .conndata_size          = sizeof(struct iscsi_conn),
2176         .max_conn               = 1,
2177         .max_cmd_len            = ISCSI_TCP_MAX_CMD_LEN,
2178         /* session management */
2179         .create_session         = iscsi_tcp_session_create,
2180         .destroy_session        = iscsi_tcp_session_destroy,
2181         /* connection management */
2182         .create_conn            = iscsi_tcp_conn_create,
2183         .bind_conn              = iscsi_tcp_conn_bind,
2184         .destroy_conn           = iscsi_tcp_conn_destroy,
2185         .set_param              = iscsi_conn_set_param,
2186         .get_conn_param         = iscsi_tcp_conn_get_param,
2187         .get_session_param      = iscsi_session_get_param,
2188         .start_conn             = iscsi_conn_start,
2189         .stop_conn              = iscsi_tcp_conn_stop,
2190         /* IO */
2191         .send_pdu               = iscsi_conn_send_pdu,
2192         .get_stats              = iscsi_conn_get_stats,
2193         .init_cmd_task          = iscsi_tcp_cmd_init,
2194         .init_mgmt_task         = iscsi_tcp_mgmt_init,
2195         .xmit_cmd_task          = iscsi_tcp_ctask_xmit,
2196         .xmit_mgmt_task         = iscsi_tcp_mtask_xmit,
2197         .cleanup_cmd_task       = iscsi_tcp_cleanup_ctask,
2198         /* recovery */
2199         .session_recovery_timedout = iscsi_session_recovery_timedout,
2200 };
2201
2202 static int __init
2203 iscsi_tcp_init(void)
2204 {
2205         if (iscsi_max_lun < 1) {
2206                 printk(KERN_ERR "iscsi_tcp: Invalid max_lun value of %u\n",
2207                        iscsi_max_lun);
2208                 return -EINVAL;
2209         }
2210         iscsi_tcp_transport.max_lun = iscsi_max_lun;
2211
2212         if (!iscsi_register_transport(&iscsi_tcp_transport))
2213                 return -ENODEV;
2214
2215         return 0;
2216 }
2217
2218 static void __exit
2219 iscsi_tcp_exit(void)
2220 {
2221         iscsi_unregister_transport(&iscsi_tcp_transport);
2222 }
2223
2224 module_init(iscsi_tcp_init);
2225 module_exit(iscsi_tcp_exit);