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/file.h>
  33 #include <linux/blkdev.h>
  34 #include <linux/crypto.h>
  35 #include <linux/delay.h>
  36 #include <linux/kfifo.h>
  37 #include <linux/scatterlist.h>
  38 #include <net/tcp.h>
  39 #include <scsi/scsi_cmnd.h>
  40 #include <scsi/scsi_device.h>
  41 #include <scsi/scsi_host.h>
  42 #include <scsi/scsi.h>
  43 #include <scsi/scsi_transport_iscsi.h>
  44
  45 #include "iscsi_tcp.h"
  46
  47 MODULE_AUTHOR("Dmitry Yusupov <dmitry_yus@yahoo.com>, "
  48               "Alex Aizman <itn780@yahoo.com>");
  49 MODULE_DESCRIPTION("iSCSI/TCP data-path");
  50 MODULE_LICENSE("GPL");
  51 #undef 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 struct scsi_transport_template *iscsi_tcp_scsi_transport;
  68 static struct scsi_host_template iscsi_sht;
  69 static struct iscsi_transport iscsi_tcp_transport;
  70
  71 static unsigned int iscsi_max_lun = 512;
  72 module_param_named(max_lun, iscsi_max_lun, uint, S_IRUGO);
  73
  74 static int iscsi_tcp_hdr_recv_done(struct iscsi_tcp_conn *tcp_conn,
  75                                    struct iscsi_segment *segment);
  76
  77 /*
  78  * Scatterlist handling: inside the iscsi_segment, we
  79  * remember an index into the scatterlist, and set data/size
  80  * to the current scatterlist entry. For highmem pages, we
  81  * kmap as needed.
  82  *
  83  * Note that the page is unmapped when we return from
  84  * TCP's data_ready handler, so we may end up mapping and
  85  * unmapping the same page repeatedly. The whole reason
  86  * for this is that we shouldn't keep the page mapped
  87  * outside the softirq.
  88  */
  89
  90 /**
  91  * iscsi_tcp_segment_init_sg - init indicated scatterlist entry
  92  * @segment: the buffer object
  93  * @sg: scatterlist
  94  * @offset: byte offset into that sg entry
  95  *
  96  * This function sets up the segment so that subsequent
  97  * data is copied to the indicated sg entry, at the given
  98  * offset.
  99  */
 100 static inline void
 101 iscsi_tcp_segment_init_sg(struct iscsi_segment *segment,
 102                           struct scatterlist *sg, unsigned int offset)
 103 {
 104         segment->sg = sg;
 105         segment->sg_offset = offset;
 106         segment->size = min(sg->length - offset,
 107                             segment->total_size - segment->total_copied);
 108         segment->data = NULL;
 109 }
 110
 111 /**
 112  * iscsi_tcp_segment_map - map the current S/G page
 113  * @segment: iscsi_segment
 114  * @recv: 1 if called from recv path
 115  *
 116  * We only need to possibly kmap data if scatter lists are being used,
 117  * because the iscsi passthrough and internal IO paths will never use high
 118  * mem pages.
 119  */
 120 static inline void
 121 iscsi_tcp_segment_map(struct iscsi_segment *segment, int recv)
 122 {
 123         struct scatterlist *sg;
 124
 125         if (segment->data != NULL || !segment->sg)
 126                 return;
 127
 128         sg = segment->sg;
 129         BUG_ON(segment->sg_mapped);
 130         BUG_ON(sg->length == 0);
 131
 132         /*
 133          * If the page count is greater than one it is ok to send
 134          * to the network layer's zero copy send path. If not we
 135          * have to go the slow sendmsg path. We always map for the
 136          * recv path.
 137          */
 138         if (page_count(sg_page(sg)) >= 1 && !recv)
 139                 return;
 140
 141         debug_tcp("iscsi_tcp_segment_map %s %p\n", recv ? "recv" : "xmit",
 142                   segment);
 143         segment->sg_mapped = kmap_atomic(sg_page(sg), KM_SOFTIRQ0);
 144         segment->data = segment->sg_mapped + sg->offset + segment->sg_offset;
 145 }
 146
 147 static inline void
 148 iscsi_tcp_segment_unmap(struct iscsi_segment *segment)
 149 {
 150         debug_tcp("iscsi_tcp_segment_unmap %p\n", segment);
 151
 152         if (segment->sg_mapped) {
 153                 debug_tcp("iscsi_tcp_segment_unmap valid\n");
 154                 kunmap_atomic(segment->sg_mapped, KM_SOFTIRQ0);
 155                 segment->sg_mapped = NULL;
 156                 segment->data = NULL;
 157         }
 158 }
 159
 160 /*
 161  * Splice the digest buffer into the buffer
 162  */
 163 static inline void
 164 iscsi_tcp_segment_splice_digest(struct iscsi_segment *segment, void *digest)
 165 {
 166         segment->data = digest;
 167         segment->digest_len = ISCSI_DIGEST_SIZE;
 168         segment->total_size += ISCSI_DIGEST_SIZE;
 169         segment->size = ISCSI_DIGEST_SIZE;
 170         segment->copied = 0;
 171         segment->sg = NULL;
 172         segment->hash = NULL;
 173 }
 174
 175 /**
 176  * iscsi_tcp_segment_done - check whether the segment is complete
 177  * @segment: iscsi segment to check
 178  * @recv: set to one of this is called from the recv path
 179  * @copied: number of bytes copied
 180  *
 181  * Check if we're done receiving this segment. If the receive
 182  * buffer is full but we expect more data, move on to the
 183  * next entry in the scatterlist.
 184  *
 185  * If the amount of data we received isn't a multiple of 4,
 186  * we will transparently receive the pad bytes, too.
 187  *
 188  * This function must be re-entrant.
 189  */
 190 static inline int
 191 iscsi_tcp_segment_done(struct iscsi_segment *segment, int recv, unsigned copied)
 192 {
 193         static unsigned char padbuf[ISCSI_PAD_LEN];
 194         struct scatterlist sg;
 195         unsigned int pad;
 196
 197         debug_tcp("copied %u %u size %u %s\n", segment->copied, copied,
 198                   segment->size, recv ? "recv" : "xmit");
 199         if (segment->hash && copied) {
 200                 /*
 201                  * If a segment is kmapd we must unmap it before sending
 202                  * to the crypto layer since that will try to kmap it again.
 203                  */
 204                 iscsi_tcp_segment_unmap(segment);
 205
 206                 if (!segment->data) {
 207                         sg_init_table(&sg, 1);
 208                         sg_set_page(&sg, sg_page(segment->sg), copied,
 209                                     segment->copied + segment->sg_offset +
 210                                                         segment->sg->offset);
 211                 } else
 212                         sg_init_one(&sg, segment->data + segment->copied,
 213                                     copied);
 214                 crypto_hash_update(segment->hash, &sg, copied);
 215         }
 216
 217         segment->copied += copied;
 218         if (segment->copied < segment->size) {
 219                 iscsi_tcp_segment_map(segment, recv);
 220                 return 0;
 221         }
 222
 223         segment->total_copied += segment->copied;
 224         segment->copied = 0;
 225         segment->size = 0;
 226
 227         /* Unmap the current scatterlist page, if there is one. */
 228         iscsi_tcp_segment_unmap(segment);
 229
 230         /* Do we have more scatterlist entries? */
 231         debug_tcp("total copied %u total size %u\n", segment->total_copied,
 232                    segment->total_size);
 233         if (segment->total_copied < segment->total_size) {
 234                 /* Proceed to the next entry in the scatterlist. */
 235                 iscsi_tcp_segment_init_sg(segment, sg_next(segment->sg),
 236                                           0);
 237                 iscsi_tcp_segment_map(segment, recv);
 238                 BUG_ON(segment->size == 0);
 239                 return 0;
 240         }
 241
 242         /* Do we need to handle padding? */
 243         pad = iscsi_padding(segment->total_copied);
 244         if (pad != 0) {
 245                 debug_tcp("consume %d pad bytes\n", pad);
 246                 segment->total_size += pad;
 247                 segment->size = pad;
 248                 segment->data = padbuf;
 249                 return 0;
 250         }
 251
 252         /*
 253          * Set us up for transferring the data digest. hdr digest
 254          * is completely handled in hdr done function.
 255          */
 256         if (segment->hash) {
 257                 crypto_hash_final(segment->hash, segment->digest);
 258                 iscsi_tcp_segment_splice_digest(segment,
 259                                  recv ? segment->recv_digest : segment->digest);
 260                 return 0;
 261         }
 262
 263         return 1;
 264 }
 265
 266 /**
 267  * iscsi_tcp_xmit_segment - transmit segment
 268  * @tcp_conn: the iSCSI TCP connection
 269  * @segment: the buffer to transmnit
 270  *
 271  * This function transmits as much of the buffer as
 272  * the network layer will accept, and returns the number of
 273  * bytes transmitted.
 274  *
 275  * If CRC hashing is enabled, the function will compute the
 276  * hash as it goes. When the entire segment has been transmitted,
 277  * it will retrieve the hash value and send it as well.
 278  */
 279 static int
 280 iscsi_tcp_xmit_segment(struct iscsi_tcp_conn *tcp_conn,
 281                        struct iscsi_segment *segment)
 282 {
 283         struct socket *sk = tcp_conn->sock;
 284         unsigned int copied = 0;
 285         int r = 0;
 286
 287         while (!iscsi_tcp_segment_done(segment, 0, r)) {
 288                 struct scatterlist *sg;
 289                 unsigned int offset, copy;
 290                 int flags = 0;
 291
 292                 r = 0;
 293                 offset = segment->copied;
 294                 copy = segment->size - offset;
 295
 296                 if (segment->total_copied + segment->size < segment->total_size)
 297                         flags |= MSG_MORE;
 298
 299                 /* Use sendpage if we can; else fall back to sendmsg */
 300                 if (!segment->data) {
 301                         sg = segment->sg;
 302                         offset += segment->sg_offset + sg->offset;
 303                         r = tcp_conn->sendpage(sk, sg_page(sg), offset, copy,
 304                                                flags);
 305                 } else {
 306                         struct msghdr msg = { .msg_flags = flags };
 307                         struct kvec iov = {
 308                                 .iov_base = segment->data + offset,
 309                                 .iov_len = copy
 310                         };
 311
 312                         r = kernel_sendmsg(sk, &msg, &iov, 1, copy);
 313                 }
 314
 315                 if (r < 0) {
 316                         iscsi_tcp_segment_unmap(segment);
 317                         if (copied || r == -EAGAIN)
 318                                 break;
 319                         return r;
 320                 }
 321                 copied += r;
 322         }
 323         return copied;
 324 }
 325
 326 /**
 327  * iscsi_tcp_segment_recv - copy data to segment
 328  * @tcp_conn: the iSCSI TCP connection
 329  * @segment: the buffer to copy to
 330  * @ptr: data pointer
 331  * @len: amount of data available
 332  *
 333  * This function copies up to @len bytes to the
 334  * given buffer, and returns the number of bytes
 335  * consumed, which can actually be less than @len.
 336  *
 337  * If hash digest is enabled, the function will update the
 338  * hash while copying.
 339  * Combining these two operations doesn't buy us a lot (yet),
 340  * but in the future we could implement combined copy+crc,
 341  * just way we do for network layer checksums.
 342  */
 343 static int
 344 iscsi_tcp_segment_recv(struct iscsi_tcp_conn *tcp_conn,
 345                        struct iscsi_segment *segment, const void *ptr,
 346                        unsigned int len)
 347 {
 348         unsigned int copy = 0, copied = 0;
 349
 350         while (!iscsi_tcp_segment_done(segment, 1, copy)) {
 351                 if (copied == len) {
 352                         debug_tcp("iscsi_tcp_segment_recv copied %d bytes\n",
 353                                   len);
 354                         break;
 355                 }
 356
 357                 copy = min(len - copied, segment->size - segment->copied);
 358                 debug_tcp("iscsi_tcp_segment_recv copying %d\n", copy);
 359                 memcpy(segment->data + segment->copied, ptr + copied, copy);
 360                 copied += copy;
 361         }
 362         return copied;
 363 }
 364
 365 static inline void
 366 iscsi_tcp_dgst_header(struct hash_desc *hash, const void *hdr, size_t hdrlen,
 367                       unsigned char digest[ISCSI_DIGEST_SIZE])
 368 {
 369         struct scatterlist sg;
 370
 371         sg_init_one(&sg, hdr, hdrlen);
 372         crypto_hash_digest(hash, &sg, hdrlen, digest);
 373 }
 374
 375 static inline int
 376 iscsi_tcp_dgst_verify(struct iscsi_tcp_conn *tcp_conn,
 377                       struct iscsi_segment *segment)
 378 {
 379         if (!segment->digest_len)
 380                 return 1;
 381
 382         if (memcmp(segment->recv_digest, segment->digest,
 383                    segment->digest_len)) {
 384                 debug_scsi("digest mismatch\n");
 385                 return 0;
 386         }
 387
 388         return 1;
 389 }
 390
 391 /*
 392  * Helper function to set up segment buffer
 393  */
 394 static inline void
 395 __iscsi_segment_init(struct iscsi_segment *segment, size_t size,
 396                      iscsi_segment_done_fn_t *done, struct hash_desc *hash)
 397 {
 398         memset(segment, 0, sizeof(*segment));
 399         segment->total_size = size;
 400         segment->done = done;
 401
 402         if (hash) {
 403                 segment->hash = hash;
 404                 crypto_hash_init(hash);
 405         }
 406 }
 407
 408 static inline void
 409 iscsi_segment_init_linear(struct iscsi_segment *segment, void *data,
 410                           size_t size, iscsi_segment_done_fn_t *done,
 411                           struct hash_desc *hash)
 412 {
 413         __iscsi_segment_init(segment, size, done, hash);
 414         segment->data = data;
 415         segment->size = size;
 416 }
 417
 418 static inline int
 419 iscsi_segment_seek_sg(struct iscsi_segment *segment,
 420                       struct scatterlist *sg_list, unsigned int sg_count,
 421                       unsigned int offset, size_t size,
 422                       iscsi_segment_done_fn_t *done, struct hash_desc *hash)
 423 {
 424         struct scatterlist *sg;
 425         unsigned int i;
 426
 427         debug_scsi("iscsi_segment_seek_sg offset %u size %llu\n",
 428                   offset, size);
 429         __iscsi_segment_init(segment, size, done, hash);
 430         for_each_sg(sg_list, sg, sg_count, i) {
 431                 debug_scsi("sg %d, len %u offset %u\n", i, sg->length,
 432                            sg->offset);
 433                 if (offset < sg->length) {
 434                         iscsi_tcp_segment_init_sg(segment, sg, offset);
 435                         return 0;
 436                 }
 437                 offset -= sg->length;
 438         }
 439
 440         return ISCSI_ERR_DATA_OFFSET;
 441 }
 442
 443 /**
 444  * iscsi_tcp_hdr_recv_prep - prep segment for hdr reception
 445  * @tcp_conn: iscsi connection to prep for
 446  *
 447  * This function always passes NULL for the hash argument, because when this
 448  * function is called we do not yet know the final size of the header and want
 449  * to delay the digest processing until we know that.
 450  */
 451 static void
 452 iscsi_tcp_hdr_recv_prep(struct iscsi_tcp_conn *tcp_conn)
 453 {
 454         debug_tcp("iscsi_tcp_hdr_recv_prep(%p%s)\n", tcp_conn,
 455                   tcp_conn->iscsi_conn->hdrdgst_en ? ", digest enabled" : "");
 456         iscsi_segment_init_linear(&tcp_conn->in.segment,
 457                                 tcp_conn->in.hdr_buf, sizeof(struct iscsi_hdr),
 458                                 iscsi_tcp_hdr_recv_done, NULL);
 459 }
 460
 461 /*
 462  * Handle incoming reply to any other type of command
 463  */
 464 static int
 465 iscsi_tcp_data_recv_done(struct iscsi_tcp_conn *tcp_conn,
 466                          struct iscsi_segment *segment)
 467 {
 468         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 469         int rc = 0;
 470
 471         if (!iscsi_tcp_dgst_verify(tcp_conn, segment))
 472                 return ISCSI_ERR_DATA_DGST;
 473
 474         rc = iscsi_complete_pdu(conn, tcp_conn->in.hdr,
 475                         conn->data, tcp_conn->in.datalen);
 476         if (rc)
 477                 return rc;
 478
 479         iscsi_tcp_hdr_recv_prep(tcp_conn);
 480         return 0;
 481 }
 482
 483 static void
 484 iscsi_tcp_data_recv_prep(struct iscsi_tcp_conn *tcp_conn)
 485 {
 486         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 487         struct hash_desc *rx_hash = NULL;
 488
 489         if (conn->datadgst_en)
 490                 rx_hash = &tcp_conn->rx_hash;
 491
 492         iscsi_segment_init_linear(&tcp_conn->in.segment,
 493                                 conn->data, tcp_conn->in.datalen,
 494                                 iscsi_tcp_data_recv_done, rx_hash);
 495 }
 496
 497 /*
 498  * must be called with session lock
 499  */
 500 static void
 501 iscsi_tcp_cleanup_task(struct iscsi_conn *conn, struct iscsi_task *task)
 502 {
 503         struct iscsi_tcp_task *tcp_task = task->dd_data;
 504         struct iscsi_r2t_info *r2t;
 505
 506         /* nothing to do for mgmt tasks */
 507         if (!task->sc)
 508                 return;
 509
 510         /* flush task's r2t queues */
 511         while (__kfifo_get(tcp_task->r2tqueue, (void*)&r2t, sizeof(void*))) {
 512                 __kfifo_put(tcp_task->r2tpool.queue, (void*)&r2t,
 513                             sizeof(void*));
 514                 debug_scsi("iscsi_tcp_cleanup_task pending r2t dropped\n");
 515         }
 516
 517         r2t = tcp_task->r2t;
 518         if (r2t != NULL) {
 519                 __kfifo_put(tcp_task->r2tpool.queue, (void*)&r2t,
 520                             sizeof(void*));
 521                 tcp_task->r2t = NULL;
 522         }
 523 }
 524
 525 /**
 526  * iscsi_data_in - SCSI Data-In Response processing
 527  * @conn: iscsi connection
 528  * @task: scsi command task
 529  **/
 530 static int
 531 iscsi_data_in(struct iscsi_conn *conn, struct iscsi_task *task)
 532 {
 533         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 534         struct iscsi_tcp_task *tcp_task = task->dd_data;
 535         struct iscsi_data_rsp *rhdr = (struct iscsi_data_rsp *)tcp_conn->in.hdr;
 536         int datasn = be32_to_cpu(rhdr->datasn);
 537         unsigned total_in_length = scsi_in(task->sc)->length;
 538
 539         iscsi_update_cmdsn(conn->session, (struct iscsi_nopin*)rhdr);
 540         if (tcp_conn->in.datalen == 0)
 541                 return 0;
 542
 543         if (tcp_task->exp_datasn != datasn) {
 544                 debug_tcp("%s: task->exp_datasn(%d) != rhdr->datasn(%d)\n",
 545                           __func__, tcp_task->exp_datasn, datasn);
 546                 return ISCSI_ERR_DATASN;
 547         }
 548
 549         tcp_task->exp_datasn++;
 550
 551         tcp_task->data_offset = be32_to_cpu(rhdr->offset);
 552         if (tcp_task->data_offset + tcp_conn->in.datalen > total_in_length) {
 553                 debug_tcp("%s: data_offset(%d) + data_len(%d) > total_length_in(%d)\n",
 554                           __func__, tcp_task->data_offset,
 555                           tcp_conn->in.datalen, total_in_length);
 556                 return ISCSI_ERR_DATA_OFFSET;
 557         }
 558
 559         conn->datain_pdus_cnt++;
 560         return 0;
 561 }
 562
 563 /**
 564  * iscsi_solicit_data_init - initialize first Data-Out
 565  * @conn: iscsi connection
 566  * @task: scsi command task
 567  * @r2t: R2T info
 568  *
 569  * Notes:
 570  *      Initialize first Data-Out within this R2T sequence and finds
 571  *      proper data_offset within this SCSI command.
 572  *
 573  *      This function is called with connection lock taken.
 574  **/
 575 static void
 576 iscsi_solicit_data_init(struct iscsi_conn *conn, struct iscsi_task *task,
 577                         struct iscsi_r2t_info *r2t)
 578 {
 579         struct iscsi_data *hdr;
 580
 581         hdr = &r2t->dtask.hdr;
 582         memset(hdr, 0, sizeof(struct iscsi_data));
 583         hdr->ttt = r2t->ttt;
 584         hdr->datasn = cpu_to_be32(r2t->solicit_datasn);
 585         r2t->solicit_datasn++;
 586         hdr->opcode = ISCSI_OP_SCSI_DATA_OUT;
 587         memcpy(hdr->lun, task->hdr->lun, sizeof(hdr->lun));
 588         hdr->itt = task->hdr->itt;
 589         hdr->exp_statsn = r2t->exp_statsn;
 590         hdr->offset = cpu_to_be32(r2t->data_offset);
 591         if (r2t->data_length > conn->max_xmit_dlength) {
 592                 hton24(hdr->dlength, conn->max_xmit_dlength);
 593                 r2t->data_count = conn->max_xmit_dlength;
 594                 hdr->flags = 0;
 595         } else {
 596                 hton24(hdr->dlength, r2t->data_length);
 597                 r2t->data_count = r2t->data_length;
 598                 hdr->flags = ISCSI_FLAG_CMD_FINAL;
 599         }
 600         conn->dataout_pdus_cnt++;
 601
 602         r2t->sent = 0;
 603 }
 604
 605 /**
 606  * iscsi_r2t_rsp - iSCSI R2T Response processing
 607  * @conn: iscsi connection
 608  * @task: scsi command task
 609  **/
 610 static int
 611 iscsi_r2t_rsp(struct iscsi_conn *conn, struct iscsi_task *task)
 612 {
 613         struct iscsi_r2t_info *r2t;
 614         struct iscsi_session *session = conn->session;
 615         struct iscsi_tcp_task *tcp_task = task->dd_data;
 616         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 617         struct iscsi_r2t_rsp *rhdr = (struct iscsi_r2t_rsp *)tcp_conn->in.hdr;
 618         int r2tsn = be32_to_cpu(rhdr->r2tsn);
 619         int rc;
 620
 621         if (tcp_conn->in.datalen) {
 622                 iscsi_conn_printk(KERN_ERR, conn,
 623                                   "invalid R2t with datalen %d\n",
 624                                   tcp_conn->in.datalen);
 625                 return ISCSI_ERR_DATALEN;
 626         }
 627
 628         if (tcp_task->exp_datasn != r2tsn){
 629                 debug_tcp("%s: task->exp_datasn(%d) != rhdr->r2tsn(%d)\n",
 630                           __func__, tcp_task->exp_datasn, r2tsn);
 631                 return ISCSI_ERR_R2TSN;
 632         }
 633
 634         /* fill-in new R2T associated with the task */
 635         iscsi_update_cmdsn(session, (struct iscsi_nopin*)rhdr);
 636
 637         if (!task->sc || session->state != ISCSI_STATE_LOGGED_IN) {
 638                 iscsi_conn_printk(KERN_INFO, conn,
 639                                   "dropping R2T itt %d in recovery.\n",
 640                                   task->itt);
 641                 return 0;
 642         }
 643
 644         rc = __kfifo_get(tcp_task->r2tpool.queue, (void*)&r2t, sizeof(void*));
 645         BUG_ON(!rc);
 646
 647         r2t->exp_statsn = rhdr->statsn;
 648         r2t->data_length = be32_to_cpu(rhdr->data_length);
 649         if (r2t->data_length == 0) {
 650                 iscsi_conn_printk(KERN_ERR, conn,
 651                                   "invalid R2T with zero data len\n");
 652                 __kfifo_put(tcp_task->r2tpool.queue, (void*)&r2t,
 653                             sizeof(void*));
 654                 return ISCSI_ERR_DATALEN;
 655         }
 656
 657         if (r2t->data_length > session->max_burst)
 658                 debug_scsi("invalid R2T with data len %u and max burst %u."
 659                            "Attempting to execute request.\n",
 660                             r2t->data_length, session->max_burst);
 661
 662         r2t->data_offset = be32_to_cpu(rhdr->data_offset);
 663         if (r2t->data_offset + r2t->data_length > scsi_out(task->sc)->length) {
 664                 iscsi_conn_printk(KERN_ERR, conn,
 665                                   "invalid R2T with data len %u at offset %u "
 666                                   "and total length %d\n", r2t->data_length,
 667                                   r2t->data_offset, scsi_out(task->sc)->length);
 668                 __kfifo_put(tcp_task->r2tpool.queue, (void*)&r2t,
 669                             sizeof(void*));
 670                 return ISCSI_ERR_DATALEN;
 671         }
 672
 673         r2t->ttt = rhdr->ttt; /* no flip */
 674         r2t->solicit_datasn = 0;
 675
 676         iscsi_solicit_data_init(conn, task, r2t);
 677
 678         tcp_task->exp_datasn = r2tsn + 1;
 679         __kfifo_put(tcp_task->r2tqueue, (void*)&r2t, sizeof(void*));
 680         conn->r2t_pdus_cnt++;
 681
 682         iscsi_requeue_task(task);
 683         return 0;
 684 }
 685
 686 /*
 687  * Handle incoming reply to DataIn command
 688  */
 689 static int
 690 iscsi_tcp_process_data_in(struct iscsi_tcp_conn *tcp_conn,
 691                           struct iscsi_segment *segment)
 692 {
 693         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 694         struct iscsi_hdr *hdr = tcp_conn->in.hdr;
 695         int rc;
 696
 697         if (!iscsi_tcp_dgst_verify(tcp_conn, segment))
 698                 return ISCSI_ERR_DATA_DGST;
 699
 700         /* check for non-exceptional status */
 701         if (hdr->flags & ISCSI_FLAG_DATA_STATUS) {
 702                 rc = iscsi_complete_pdu(conn, tcp_conn->in.hdr, NULL, 0);
 703                 if (rc)
 704                         return rc;
 705         }
 706
 707         iscsi_tcp_hdr_recv_prep(tcp_conn);
 708         return 0;
 709 }
 710
 711 /**
 712  * iscsi_tcp_hdr_dissect - process PDU header
 713  * @conn: iSCSI connection
 714  * @hdr: PDU header
 715  *
 716  * This function analyzes the header of the PDU received,
 717  * and performs several sanity checks. If the PDU is accompanied
 718  * by data, the receive buffer is set up to copy the incoming data
 719  * to the correct location.
 720  */
 721 static int
 722 iscsi_tcp_hdr_dissect(struct iscsi_conn *conn, struct iscsi_hdr *hdr)
 723 {
 724         int rc = 0, opcode, ahslen;
 725         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 726         struct iscsi_task *task;
 727
 728         /* verify PDU length */
 729         tcp_conn->in.datalen = ntoh24(hdr->dlength);
 730         if (tcp_conn->in.datalen > conn->max_recv_dlength) {
 731                 iscsi_conn_printk(KERN_ERR, conn,
 732                                   "iscsi_tcp: datalen %d > %d\n",
 733                                   tcp_conn->in.datalen, conn->max_recv_dlength);
 734                 return ISCSI_ERR_DATALEN;
 735         }
 736
 737         /* Additional header segments. So far, we don't
 738          * process additional headers.
 739          */
 740         ahslen = hdr->hlength << 2;
 741
 742         opcode = hdr->opcode & ISCSI_OPCODE_MASK;
 743         /* verify itt (itt encoding: age+cid+itt) */
 744         rc = iscsi_verify_itt(conn, hdr->itt);
 745         if (rc)
 746                 return rc;
 747
 748         debug_tcp("opcode 0x%x ahslen %d datalen %d\n",
 749                   opcode, ahslen, tcp_conn->in.datalen);
 750
 751         switch(opcode) {
 752         case ISCSI_OP_SCSI_DATA_IN:
 753                 spin_lock(&conn->session->lock);
 754                 task = iscsi_itt_to_ctask(conn, hdr->itt);
 755                 if (!task)
 756                         rc = ISCSI_ERR_BAD_ITT;
 757                 else
 758                         rc = iscsi_data_in(conn, task);
 759                 if (rc) {
 760                         spin_unlock(&conn->session->lock);
 761                         break;
 762                 }
 763
 764                 if (tcp_conn->in.datalen) {
 765                         struct iscsi_tcp_task *tcp_task = task->dd_data;
 766                         struct hash_desc *rx_hash = NULL;
 767                         struct scsi_data_buffer *sdb = scsi_in(task->sc);
 768
 769                         /*
 770                          * Setup copy of Data-In into the Scsi_Cmnd
 771                          * Scatterlist case:
 772                          * We set up the iscsi_segment to point to the next
 773                          * scatterlist entry to copy to. As we go along,
 774                          * we move on to the next scatterlist entry and
 775                          * update the digest per-entry.
 776                          */
 777                         if (conn->datadgst_en)
 778                                 rx_hash = &tcp_conn->rx_hash;
 779
 780                         debug_tcp("iscsi_tcp_begin_data_in(%p, offset=%d, "
 781                                   "datalen=%d)\n", tcp_conn,
 782                                   tcp_task->data_offset,
 783                                   tcp_conn->in.datalen);
 784                         rc = iscsi_segment_seek_sg(&tcp_conn->in.segment,
 785                                                    sdb->table.sgl,
 786                                                    sdb->table.nents,
 787                                                    tcp_task->data_offset,
 788                                                    tcp_conn->in.datalen,
 789                                                    iscsi_tcp_process_data_in,
 790                                                    rx_hash);
 791                         spin_unlock(&conn->session->lock);
 792                         return rc;
 793                 }
 794                 rc = __iscsi_complete_pdu(conn, hdr, NULL, 0);
 795                 spin_unlock(&conn->session->lock);
 796                 break;
 797         case ISCSI_OP_SCSI_CMD_RSP:
 798                 if (tcp_conn->in.datalen) {
 799                         iscsi_tcp_data_recv_prep(tcp_conn);
 800                         return 0;
 801                 }
 802                 rc = iscsi_complete_pdu(conn, hdr, NULL, 0);
 803                 break;
 804         case ISCSI_OP_R2T:
 805                 spin_lock(&conn->session->lock);
 806                 task = iscsi_itt_to_ctask(conn, hdr->itt);
 807                 if (!task)
 808                         rc = ISCSI_ERR_BAD_ITT;
 809                 else if (ahslen)
 810                         rc = ISCSI_ERR_AHSLEN;
 811                 else if (task->sc->sc_data_direction == DMA_TO_DEVICE)
 812                         rc = iscsi_r2t_rsp(conn, task);
 813                 else
 814                         rc = ISCSI_ERR_PROTO;
 815                 spin_unlock(&conn->session->lock);
 816                 break;
 817         case ISCSI_OP_LOGIN_RSP:
 818         case ISCSI_OP_TEXT_RSP:
 819         case ISCSI_OP_REJECT:
 820         case ISCSI_OP_ASYNC_EVENT:
 821                 /*
 822                  * It is possible that we could get a PDU with a buffer larger
 823                  * than 8K, but there are no targets that currently do this.
 824                  * For now we fail until we find a vendor that needs it
 825                  */
 826                 if (ISCSI_DEF_MAX_RECV_SEG_LEN < tcp_conn->in.datalen) {
 827                         iscsi_conn_printk(KERN_ERR, conn,
 828                                           "iscsi_tcp: received buffer of "
 829                                           "len %u but conn buffer is only %u "
 830                                           "(opcode %0x)\n",
 831                                           tcp_conn->in.datalen,
 832                                           ISCSI_DEF_MAX_RECV_SEG_LEN, opcode);
 833                         rc = ISCSI_ERR_PROTO;
 834                         break;
 835                 }
 836
 837                 /* If there's data coming in with the response,
 838                  * receive it to the connection's buffer.
 839                  */
 840                 if (tcp_conn->in.datalen) {
 841                         iscsi_tcp_data_recv_prep(tcp_conn);
 842                         return 0;
 843                 }
 844         /* fall through */
 845         case ISCSI_OP_LOGOUT_RSP:
 846         case ISCSI_OP_NOOP_IN:
 847         case ISCSI_OP_SCSI_TMFUNC_RSP:
 848                 rc = iscsi_complete_pdu(conn, hdr, NULL, 0);
 849                 break;
 850         default:
 851                 rc = ISCSI_ERR_BAD_OPCODE;
 852                 break;
 853         }
 854
 855         if (rc == 0) {
 856                 /* Anything that comes with data should have
 857                  * been handled above. */
 858                 if (tcp_conn->in.datalen)
 859                         return ISCSI_ERR_PROTO;
 860                 iscsi_tcp_hdr_recv_prep(tcp_conn);
 861         }
 862
 863         return rc;
 864 }
 865
 866 /**
 867  * iscsi_tcp_hdr_recv_done - process PDU header
 868  *
 869  * This is the callback invoked when the PDU header has
 870  * been received. If the header is followed by additional
 871  * header segments, we go back for more data.
 872  */
 873 static int
 874 iscsi_tcp_hdr_recv_done(struct iscsi_tcp_conn *tcp_conn,
 875                         struct iscsi_segment *segment)
 876 {
 877         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 878         struct iscsi_hdr *hdr;
 879
 880         /* Check if there are additional header segments
 881          * *prior* to computing the digest, because we
 882          * may need to go back to the caller for more.
 883          */
 884         hdr = (struct iscsi_hdr *) tcp_conn->in.hdr_buf;
 885         if (segment->copied == sizeof(struct iscsi_hdr) && hdr->hlength) {
 886                 /* Bump the header length - the caller will
 887                  * just loop around and get the AHS for us, and
 888                  * call again. */
 889                 unsigned int ahslen = hdr->hlength << 2;
 890
 891                 /* Make sure we don't overflow */
 892                 if (sizeof(*hdr) + ahslen > sizeof(tcp_conn->in.hdr_buf))
 893                         return ISCSI_ERR_AHSLEN;
 894
 895                 segment->total_size += ahslen;
 896                 segment->size += ahslen;
 897                 return 0;
 898         }
 899
 900         /* We're done processing the header. See if we're doing
 901          * header digests; if so, set up the recv_digest buffer
 902          * and go back for more. */
 903         if (conn->hdrdgst_en) {
 904                 if (segment->digest_len == 0) {
 905                         iscsi_tcp_segment_splice_digest(segment,
 906                                                         segment->recv_digest);
 907                         return 0;
 908                 }
 909                 iscsi_tcp_dgst_header(&tcp_conn->rx_hash, hdr,
 910                                       segment->total_copied - ISCSI_DIGEST_SIZE,
 911                                       segment->digest);
 912
 913                 if (!iscsi_tcp_dgst_verify(tcp_conn, segment))
 914                         return ISCSI_ERR_HDR_DGST;
 915         }
 916
 917         tcp_conn->in.hdr = hdr;
 918         return iscsi_tcp_hdr_dissect(conn, hdr);
 919 }
 920
 921 /**
 922  * iscsi_tcp_recv - TCP receive in sendfile fashion
 923  * @rd_desc: read descriptor
 924  * @skb: socket buffer
 925  * @offset: offset in skb
 926  * @len: skb->len - offset
 927  **/
 928 static int
 929 iscsi_tcp_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 930                unsigned int offset, size_t len)
 931 {
 932         struct iscsi_conn *conn = rd_desc->arg.data;
 933         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 934         struct iscsi_segment *segment = &tcp_conn->in.segment;
 935         struct skb_seq_state seq;
 936         unsigned int consumed = 0;
 937         int rc = 0;
 938
 939         debug_tcp("in %d bytes\n", skb->len - offset);
 940
 941         if (unlikely(conn->suspend_rx)) {
 942                 debug_tcp("conn %d Rx suspended!\n", conn->id);
 943                 return 0;
 944         }
 945
 946         skb_prepare_seq_read(skb, offset, skb->len, &seq);
 947         while (1) {
 948                 unsigned int avail;
 949                 const u8 *ptr;
 950
 951                 avail = skb_seq_read(consumed, &ptr, &seq);
 952                 if (avail == 0) {
 953                         debug_tcp("no more data avail. Consumed %d\n",
 954                                   consumed);
 955                         break;
 956                 }
 957                 BUG_ON(segment->copied >= segment->size);
 958
 959                 debug_tcp("skb %p ptr=%p avail=%u\n", skb, ptr, avail);
 960                 rc = iscsi_tcp_segment_recv(tcp_conn, segment, ptr, avail);
 961                 BUG_ON(rc == 0);
 962                 consumed += rc;
 963
 964                 if (segment->total_copied >= segment->total_size) {
 965                         debug_tcp("segment done\n");
 966                         rc = segment->done(tcp_conn, segment);
 967                         if (rc != 0) {
 968                                 skb_abort_seq_read(&seq);
 969                                 goto error;
 970                         }
 971
 972                         /* The done() functions sets up the
 973                          * next segment. */
 974                 }
 975         }
 976         skb_abort_seq_read(&seq);
 977         conn->rxdata_octets += consumed;
 978         return consumed;
 979
 980 error:
 981         debug_tcp("Error receiving PDU, errno=%d\n", rc);
 982         iscsi_conn_failure(conn, rc);
 983         return 0;
 984 }
 985
 986 static void
 987 iscsi_tcp_data_ready(struct sock *sk, int flag)
 988 {
 989         struct iscsi_conn *conn = sk->sk_user_data;
 990         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 991         read_descriptor_t rd_desc;
 992
 993         read_lock(&sk->sk_callback_lock);
 994
 995         /*
 996          * Use rd_desc to pass 'conn' to iscsi_tcp_recv.
 997          * We set count to 1 because we want the network layer to
 998          * hand us all the skbs that are available. iscsi_tcp_recv
 999          * handled pdus that cross buffers or pdus that still need data.
1000          */
1001         rd_desc.arg.data = conn;
1002         rd_desc.count = 1;
1003         tcp_read_sock(sk, &rd_desc, iscsi_tcp_recv);
1004
1005         read_unlock(&sk->sk_callback_lock);
1006
1007         /* If we had to (atomically) map a highmem page,
1008          * unmap it now. */
1009         iscsi_tcp_segment_unmap(&tcp_conn->in.segment);
1010 }
1011
1012 static void
1013 iscsi_tcp_state_change(struct sock *sk)
1014 {
1015         struct iscsi_tcp_conn *tcp_conn;
1016         struct iscsi_conn *conn;
1017         struct iscsi_session *session;
1018         void (*old_state_change)(struct sock *);
1019
1020         read_lock(&sk->sk_callback_lock);
1021
1022         conn = (struct iscsi_conn*)sk->sk_user_data;
1023         session = conn->session;
1024
1025         if ((sk->sk_state == TCP_CLOSE_WAIT ||
1026              sk->sk_state == TCP_CLOSE) &&
1027             !atomic_read(&sk->sk_rmem_alloc)) {
1028                 debug_tcp("iscsi_tcp_state_change: TCP_CLOSE|TCP_CLOSE_WAIT\n");
1029                 iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
1030         }
1031
1032         tcp_conn = conn->dd_data;
1033         old_state_change = tcp_conn->old_state_change;
1034
1035         read_unlock(&sk->sk_callback_lock);
1036
1037         old_state_change(sk);
1038 }
1039
1040 /**
1041  * iscsi_write_space - Called when more output buffer space is available
1042  * @sk: socket space is available for
1043  **/
1044 static void
1045 iscsi_write_space(struct sock *sk)
1046 {
1047         struct iscsi_conn *conn = (struct iscsi_conn*)sk->sk_user_data;
1048         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1049
1050         tcp_conn->old_write_space(sk);
1051         debug_tcp("iscsi_write_space: cid %d\n", conn->id);
1052         scsi_queue_work(conn->session->host, &conn->xmitwork);
1053 }
1054
1055 static void
1056 iscsi_conn_set_callbacks(struct iscsi_conn *conn)
1057 {
1058         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1059         struct sock *sk = tcp_conn->sock->sk;
1060
1061         /* assign new callbacks */
1062         write_lock_bh(&sk->sk_callback_lock);
1063         sk->sk_user_data = conn;
1064         tcp_conn->old_data_ready = sk->sk_data_ready;
1065         tcp_conn->old_state_change = sk->sk_state_change;
1066         tcp_conn->old_write_space = sk->sk_write_space;
1067         sk->sk_data_ready = iscsi_tcp_data_ready;
1068         sk->sk_state_change = iscsi_tcp_state_change;
1069         sk->sk_write_space = iscsi_write_space;
1070         write_unlock_bh(&sk->sk_callback_lock);
1071 }
1072
1073 static void
1074 iscsi_conn_restore_callbacks(struct iscsi_tcp_conn *tcp_conn)
1075 {
1076         struct sock *sk = tcp_conn->sock->sk;
1077
1078         /* restore socket callbacks, see also: iscsi_conn_set_callbacks() */
1079         write_lock_bh(&sk->sk_callback_lock);
1080         sk->sk_user_data    = NULL;
1081         sk->sk_data_ready   = tcp_conn->old_data_ready;
1082         sk->sk_state_change = tcp_conn->old_state_change;
1083         sk->sk_write_space  = tcp_conn->old_write_space;
1084         sk->sk_no_check  = 0;
1085         write_unlock_bh(&sk->sk_callback_lock);
1086 }
1087
1088 /**
1089  * iscsi_xmit - TCP transmit
1090  **/
1091 static int
1092 iscsi_xmit(struct iscsi_conn *conn)
1093 {
1094         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1095         struct iscsi_segment *segment = &tcp_conn->out.segment;
1096         unsigned int consumed = 0;
1097         int rc = 0;
1098
1099         while (1) {
1100                 rc = iscsi_tcp_xmit_segment(tcp_conn, segment);
1101                 if (rc < 0) {
1102                         rc = ISCSI_ERR_XMIT_FAILED;
1103                         goto error;
1104                 }
1105                 if (rc == 0)
1106                         break;
1107
1108                 consumed += rc;
1109
1110                 if (segment->total_copied >= segment->total_size) {
1111                         if (segment->done != NULL) {
1112                                 rc = segment->done(tcp_conn, segment);
1113                                 if (rc != 0)
1114                                         goto error;
1115                         }
1116                 }
1117         }
1118
1119         debug_tcp("xmit %d bytes\n", consumed);
1120
1121         conn->txdata_octets += consumed;
1122         return consumed;
1123
1124 error:
1125         /* Transmit error. We could initiate error recovery
1126          * here. */
1127         debug_tcp("Error sending PDU, errno=%d\n", rc);
1128         iscsi_conn_failure(conn, rc);
1129         return -EIO;
1130 }
1131
1132 /**
1133  * iscsi_tcp_xmit_qlen - return the number of bytes queued for xmit
1134  */
1135 static inline int
1136 iscsi_tcp_xmit_qlen(struct iscsi_conn *conn)
1137 {
1138         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1139         struct iscsi_segment *segment = &tcp_conn->out.segment;
1140
1141         return segment->total_copied - segment->total_size;
1142 }
1143
1144 static inline int
1145 iscsi_tcp_flush(struct iscsi_conn *conn)
1146 {
1147         int rc;
1148
1149         while (iscsi_tcp_xmit_qlen(conn)) {
1150                 rc = iscsi_xmit(conn);
1151                 if (rc == 0)
1152                         return -EAGAIN;
1153                 if (rc < 0)
1154                         return rc;
1155         }
1156
1157         return 0;
1158 }
1159
1160 /*
1161  * This is called when we're done sending the header.
1162  * Simply copy the data_segment to the send segment, and return.
1163  */
1164 static int
1165 iscsi_tcp_send_hdr_done(struct iscsi_tcp_conn *tcp_conn,
1166                         struct iscsi_segment *segment)
1167 {
1168         tcp_conn->out.segment = tcp_conn->out.data_segment;
1169         debug_tcp("Header done. Next segment size %u total_size %u\n",
1170                   tcp_conn->out.segment.size, tcp_conn->out.segment.total_size);
1171         return 0;
1172 }
1173
1174 static void
1175 iscsi_tcp_send_hdr_prep(struct iscsi_conn *conn, void *hdr, size_t hdrlen)
1176 {
1177         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1178
1179         debug_tcp("%s(%p%s)\n", __func__, tcp_conn,
1180                         conn->hdrdgst_en? ", digest enabled" : "");
1181
1182         /* Clear the data segment - needs to be filled in by the
1183          * caller using iscsi_tcp_send_data_prep() */
1184         memset(&tcp_conn->out.data_segment, 0, sizeof(struct iscsi_segment));
1185
1186         /* If header digest is enabled, compute the CRC and
1187          * place the digest into the same buffer. We make
1188          * sure that both iscsi_tcp_task and mtask have
1189          * sufficient room.
1190          */
1191         if (conn->hdrdgst_en) {
1192                 iscsi_tcp_dgst_header(&tcp_conn->tx_hash, hdr, hdrlen,
1193                                       hdr + hdrlen);
1194                 hdrlen += ISCSI_DIGEST_SIZE;
1195         }
1196
1197         /* Remember header pointer for later, when we need
1198          * to decide whether there's a payload to go along
1199          * with the header. */
1200         tcp_conn->out.hdr = hdr;
1201
1202         iscsi_segment_init_linear(&tcp_conn->out.segment, hdr, hdrlen,
1203                                 iscsi_tcp_send_hdr_done, NULL);
1204 }
1205
1206 /*
1207  * Prepare the send buffer for the payload data.
1208  * Padding and checksumming will all be taken care
1209  * of by the iscsi_segment routines.
1210  */
1211 static int
1212 iscsi_tcp_send_data_prep(struct iscsi_conn *conn, struct scatterlist *sg,
1213                          unsigned int count, unsigned int offset,
1214                          unsigned int len)
1215 {
1216         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1217         struct hash_desc *tx_hash = NULL;
1218         unsigned int hdr_spec_len;
1219
1220         debug_tcp("%s(%p, offset=%d, datalen=%d%s)\n", __func__,
1221                         tcp_conn, offset, len,
1222                         conn->datadgst_en? ", digest enabled" : "");
1223
1224         /* Make sure the datalen matches what the caller
1225            said he would send. */
1226         hdr_spec_len = ntoh24(tcp_conn->out.hdr->dlength);
1227         WARN_ON(iscsi_padded(len) != iscsi_padded(hdr_spec_len));
1228
1229         if (conn->datadgst_en)
1230                 tx_hash = &tcp_conn->tx_hash;
1231
1232         return iscsi_segment_seek_sg(&tcp_conn->out.data_segment,
1233                                    sg, count, offset, len,
1234                                    NULL, tx_hash);
1235 }
1236
1237 static void
1238 iscsi_tcp_send_linear_data_prepare(struct iscsi_conn *conn, void *data,
1239                                    size_t len)
1240 {
1241         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1242         struct hash_desc *tx_hash = NULL;
1243         unsigned int hdr_spec_len;
1244
1245         debug_tcp("%s(%p, datalen=%d%s)\n", __func__, tcp_conn, len,
1246                   conn->datadgst_en? ", digest enabled" : "");
1247
1248         /* Make sure the datalen matches what the caller
1249            said he would send. */
1250         hdr_spec_len = ntoh24(tcp_conn->out.hdr->dlength);
1251         WARN_ON(iscsi_padded(len) != iscsi_padded(hdr_spec_len));
1252
1253         if (conn->datadgst_en)
1254                 tx_hash = &tcp_conn->tx_hash;
1255
1256         iscsi_segment_init_linear(&tcp_conn->out.data_segment,
1257                                 data, len, NULL, tx_hash);
1258 }
1259
1260 /**
1261  * iscsi_solicit_data_cont - initialize next Data-Out
1262  * @conn: iscsi connection
1263  * @task: scsi command task
1264  * @r2t: R2T info
1265  * @left: bytes left to transfer
1266  *
1267  * Notes:
1268  *      Initialize next Data-Out within this R2T sequence and continue
1269  *      to process next Scatter-Gather element(if any) of this SCSI command.
1270  *
1271  *      Called under connection lock.
1272  **/
1273 static int
1274 iscsi_solicit_data_cont(struct iscsi_conn *conn, struct iscsi_task *task,
1275                         struct iscsi_r2t_info *r2t)
1276 {
1277         struct iscsi_data *hdr;
1278         int new_offset, left;
1279
1280         BUG_ON(r2t->data_length - r2t->sent < 0);
1281         left = r2t->data_length - r2t->sent;
1282         if (left == 0)
1283                 return 0;
1284
1285         hdr = &r2t->dtask.hdr;
1286         memset(hdr, 0, sizeof(struct iscsi_data));
1287         hdr->ttt = r2t->ttt;
1288         hdr->datasn = cpu_to_be32(r2t->solicit_datasn);
1289         r2t->solicit_datasn++;
1290         hdr->opcode = ISCSI_OP_SCSI_DATA_OUT;
1291         memcpy(hdr->lun, task->hdr->lun, sizeof(hdr->lun));
1292         hdr->itt = task->hdr->itt;
1293         hdr->exp_statsn = r2t->exp_statsn;
1294         new_offset = r2t->data_offset + r2t->sent;
1295         hdr->offset = cpu_to_be32(new_offset);
1296         if (left > conn->max_xmit_dlength) {
1297                 hton24(hdr->dlength, conn->max_xmit_dlength);
1298                 r2t->data_count = conn->max_xmit_dlength;
1299         } else {
1300                 hton24(hdr->dlength, left);
1301                 r2t->data_count = left;
1302                 hdr->flags = ISCSI_FLAG_CMD_FINAL;
1303         }
1304
1305         conn->dataout_pdus_cnt++;
1306         return 1;
1307 }
1308
1309 /**
1310  * iscsi_tcp_task - Initialize iSCSI SCSI_READ or SCSI_WRITE commands
1311  * @conn: iscsi connection
1312  * @task: scsi command task
1313  * @sc: scsi command
1314  **/
1315 static int
1316 iscsi_tcp_task_init(struct iscsi_task *task)
1317 {
1318         struct iscsi_tcp_task *tcp_task = task->dd_data;
1319         struct iscsi_conn *conn = task->conn;
1320         struct scsi_cmnd *sc = task->sc;
1321         int err;
1322
1323         if (!sc) {
1324                 /*
1325                  * mgmt tasks do not have a scatterlist since they come
1326                  * in from the iscsi interface.
1327                  */
1328                 debug_scsi("mtask deq [cid %d itt 0x%x]\n", conn->id,
1329                            task->itt);
1330
1331                 /* Prepare PDU, optionally w/ immediate data */
1332                 iscsi_tcp_send_hdr_prep(conn, task->hdr, sizeof(*task->hdr));
1333
1334                 /* If we have immediate data, attach a payload */
1335                 if (task->data_count)
1336                         iscsi_tcp_send_linear_data_prepare(conn, task->data,
1337                                                            task->data_count);
1338                 return 0;
1339         }
1340
1341         BUG_ON(__kfifo_len(tcp_task->r2tqueue));
1342         tcp_task->sent = 0;
1343         tcp_task->exp_datasn = 0;
1344
1345         /* Prepare PDU, optionally w/ immediate data */
1346         debug_scsi("task deq [cid %d itt 0x%x imm %d unsol %d]\n",
1347                     conn->id, task->itt, task->imm_count,
1348                     task->unsol_count);
1349         iscsi_tcp_send_hdr_prep(conn, task->hdr, task->hdr_len);
1350
1351         if (!task->imm_count)
1352                 return 0;
1353
1354         /* If we have immediate data, attach a payload */
1355         err = iscsi_tcp_send_data_prep(conn, scsi_out(sc)->table.sgl,
1356                                        scsi_out(sc)->table.nents,
1357                                        0, task->imm_count);
1358         if (err)
1359                 return err;
1360         tcp_task->sent += task->imm_count;
1361         task->imm_count = 0;
1362         return 0;
1363 }
1364
1365 /*
1366  * iscsi_tcp_task_xmit - xmit normal PDU task
1367  * @task: iscsi command task
1368  *
1369  * We're expected to return 0 when everything was transmitted succesfully,
1370  * -EAGAIN if there's still data in the queue, or != 0 for any other kind
1371  * of error.
1372  */
1373 static int
1374 iscsi_tcp_task_xmit(struct iscsi_task *task)
1375 {
1376         struct iscsi_conn *conn = task->conn;
1377         struct iscsi_tcp_task *tcp_task = task->dd_data;
1378         struct scsi_cmnd *sc = task->sc;
1379         struct scsi_data_buffer *sdb;
1380         int rc = 0;
1381
1382 flush:
1383         /* Flush any pending data first. */
1384         rc = iscsi_tcp_flush(conn);
1385         if (rc < 0)
1386                 return rc;
1387
1388         /* mgmt command */
1389         if (!sc) {
1390                 if (task->hdr->itt == RESERVED_ITT)
1391                         iscsi_put_task(task);
1392                 return 0;
1393         }
1394
1395         /* Are we done already? */
1396         if (sc->sc_data_direction != DMA_TO_DEVICE)
1397                 return 0;
1398
1399         sdb = scsi_out(sc);
1400         if (task->unsol_count != 0) {
1401                 struct iscsi_data *hdr = &tcp_task->unsol_dtask.hdr;
1402
1403                 /* Prepare a header for the unsolicited PDU.
1404                  * The amount of data we want to send will be
1405                  * in task->data_count.
1406                  * FIXME: return the data count instead.
1407                  */
1408                 iscsi_prep_unsolicit_data_pdu(task, hdr);
1409
1410                 debug_tcp("unsol dout [itt 0x%x doff %d dlen %d]\n",
1411                                 task->itt, tcp_task->sent, task->data_count);
1412
1413                 iscsi_tcp_send_hdr_prep(conn, hdr, sizeof(*hdr));
1414                 rc = iscsi_tcp_send_data_prep(conn, sdb->table.sgl,
1415                                               sdb->table.nents, tcp_task->sent,
1416                                               task->data_count);
1417                 if (rc)
1418                         goto fail;
1419                 tcp_task->sent += task->data_count;
1420                 task->unsol_count -= task->data_count;
1421                 goto flush;
1422         } else {
1423                 struct iscsi_session *session = conn->session;
1424                 struct iscsi_r2t_info *r2t;
1425
1426                 /* All unsolicited PDUs sent. Check for solicited PDUs.
1427                  */
1428                 spin_lock_bh(&session->lock);
1429                 r2t = tcp_task->r2t;
1430                 if (r2t != NULL) {
1431                         /* Continue with this R2T? */
1432                         if (!iscsi_solicit_data_cont(conn, task, r2t)) {
1433                                 debug_scsi("  done with r2t %p\n", r2t);
1434
1435                                 __kfifo_put(tcp_task->r2tpool.queue,
1436                                             (void*)&r2t, sizeof(void*));
1437                                 tcp_task->r2t = r2t = NULL;
1438                         }
1439                 }
1440
1441                 if (r2t == NULL) {
1442                         __kfifo_get(tcp_task->r2tqueue, (void*)&tcp_task->r2t,
1443                                     sizeof(void*));
1444                         r2t = tcp_task->r2t;
1445                 }
1446                 spin_unlock_bh(&session->lock);
1447
1448                 /* Waiting for more R2Ts to arrive. */
1449                 if (r2t == NULL) {
1450                         debug_tcp("no R2Ts yet\n");
1451                         return 0;
1452                 }
1453
1454                 debug_scsi("sol dout %p [dsn %d itt 0x%x doff %d dlen %d]\n",
1455                         r2t, r2t->solicit_datasn - 1, task->itt,
1456                         r2t->data_offset + r2t->sent, r2t->data_count);
1457
1458                 iscsi_tcp_send_hdr_prep(conn, &r2t->dtask.hdr,
1459                                         sizeof(struct iscsi_hdr));
1460
1461                 rc = iscsi_tcp_send_data_prep(conn, sdb->table.sgl,
1462                                               sdb->table.nents,
1463                                               r2t->data_offset + r2t->sent,
1464                                               r2t->data_count);
1465                 if (rc)
1466                         goto fail;
1467                 tcp_task->sent += r2t->data_count;
1468                 r2t->sent += r2t->data_count;
1469                 goto flush;
1470         }
1471         return 0;
1472 fail:
1473         iscsi_conn_failure(conn, rc);
1474         return -EIO;
1475 }
1476
1477 static struct iscsi_cls_conn *
1478 iscsi_tcp_conn_create(struct iscsi_cls_session *cls_session, uint32_t conn_idx)
1479 {
1480         struct iscsi_conn *conn;
1481         struct iscsi_cls_conn *cls_conn;
1482         struct iscsi_tcp_conn *tcp_conn;
1483
1484         cls_conn = iscsi_conn_setup(cls_session, sizeof(*tcp_conn), conn_idx);
1485         if (!cls_conn)
1486                 return NULL;
1487         conn = cls_conn->dd_data;
1488         /*
1489          * due to strange issues with iser these are not set
1490          * in iscsi_conn_setup
1491          */
1492         conn->max_recv_dlength = ISCSI_DEF_MAX_RECV_SEG_LEN;
1493
1494         tcp_conn = conn->dd_data;
1495         tcp_conn->iscsi_conn = conn;
1496
1497         tcp_conn->tx_hash.tfm = crypto_alloc_hash("crc32c", 0,
1498                                                   CRYPTO_ALG_ASYNC);
1499         tcp_conn->tx_hash.flags = 0;
1500         if (IS_ERR(tcp_conn->tx_hash.tfm))
1501                 goto free_conn;
1502
1503         tcp_conn->rx_hash.tfm = crypto_alloc_hash("crc32c", 0,
1504                                                   CRYPTO_ALG_ASYNC);
1505         tcp_conn->rx_hash.flags = 0;
1506         if (IS_ERR(tcp_conn->rx_hash.tfm))
1507                 goto free_tx_tfm;
1508
1509         return cls_conn;
1510
1511 free_tx_tfm:
1512         crypto_free_hash(tcp_conn->tx_hash.tfm);
1513 free_conn:
1514         iscsi_conn_printk(KERN_ERR, conn,
1515                           "Could not create connection due to crc32c "
1516                           "loading error. Make sure the crc32c "
1517                           "module is built as a module or into the "
1518                           "kernel\n");
1519         iscsi_conn_teardown(cls_conn);
1520         return NULL;
1521 }
1522
1523 static void
1524 iscsi_tcp_release_conn(struct iscsi_conn *conn)
1525 {
1526         struct iscsi_session *session = conn->session;
1527         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1528         struct socket *sock = tcp_conn->sock;
1529
1530         if (!sock)
1531                 return;
1532
1533         sock_hold(sock->sk);
1534         iscsi_conn_restore_callbacks(tcp_conn);
1535         sock_put(sock->sk);
1536
1537         spin_lock_bh(&session->lock);
1538         tcp_conn->sock = NULL;
1539         spin_unlock_bh(&session->lock);
1540         sockfd_put(sock);
1541 }
1542
1543 static void
1544 iscsi_tcp_conn_destroy(struct iscsi_cls_conn *cls_conn)
1545 {
1546         struct iscsi_conn *conn = cls_conn->dd_data;
1547         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1548
1549         iscsi_tcp_release_conn(conn);
1550
1551         if (tcp_conn->tx_hash.tfm)
1552                 crypto_free_hash(tcp_conn->tx_hash.tfm);
1553         if (tcp_conn->rx_hash.tfm)
1554                 crypto_free_hash(tcp_conn->rx_hash.tfm);
1555
1556         iscsi_conn_teardown(cls_conn);
1557 }
1558
1559 static void
1560 iscsi_tcp_conn_stop(struct iscsi_cls_conn *cls_conn, int flag)
1561 {
1562         struct iscsi_conn *conn = cls_conn->dd_data;
1563         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1564
1565         /* userspace may have goofed up and not bound us */
1566         if (!tcp_conn->sock)
1567                 return;
1568         /*
1569          * Make sure our recv side is stopped.
1570          * Older tools called conn stop before ep_disconnect
1571          * so IO could still be coming in.
1572          */
1573         write_lock_bh(&tcp_conn->sock->sk->sk_callback_lock);
1574         set_bit(ISCSI_SUSPEND_BIT, &conn->suspend_rx);
1575         write_unlock_bh(&tcp_conn->sock->sk->sk_callback_lock);
1576
1577         iscsi_conn_stop(cls_conn, flag);
1578         iscsi_tcp_release_conn(conn);
1579 }
1580
1581 static int iscsi_tcp_get_addr(struct iscsi_conn *conn, struct socket *sock,
1582                               char *buf, int *port,
1583                               int (*getname)(struct socket *, struct sockaddr *,
1584                                         int *addrlen))
1585 {
1586         struct sockaddr_storage *addr;
1587         struct sockaddr_in6 *sin6;
1588         struct sockaddr_in *sin;
1589         int rc = 0, len;
1590
1591         addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1592         if (!addr)
1593                 return -ENOMEM;
1594
1595         if (getname(sock, (struct sockaddr *) addr, &len)) {
1596                 rc = -ENODEV;
1597                 goto free_addr;
1598         }
1599
1600         switch (addr->ss_family) {
1601         case AF_INET:
1602                 sin = (struct sockaddr_in *)addr;
1603                 spin_lock_bh(&conn->session->lock);
1604                 sprintf(buf, NIPQUAD_FMT, NIPQUAD(sin->sin_addr.s_addr));
1605                 *port = be16_to_cpu(sin->sin_port);
1606                 spin_unlock_bh(&conn->session->lock);
1607                 break;
1608         case AF_INET6:
1609                 sin6 = (struct sockaddr_in6 *)addr;
1610                 spin_lock_bh(&conn->session->lock);
1611                 sprintf(buf, NIP6_FMT, NIP6(sin6->sin6_addr));
1612                 *port = be16_to_cpu(sin6->sin6_port);
1613                 spin_unlock_bh(&conn->session->lock);
1614                 break;
1615         }
1616 free_addr:
1617         kfree(addr);
1618         return rc;
1619 }
1620
1621 static int
1622 iscsi_tcp_conn_bind(struct iscsi_cls_session *cls_session,
1623                     struct iscsi_cls_conn *cls_conn, uint64_t transport_eph,
1624                     int is_leading)
1625 {
1626         struct Scsi_Host *shost = iscsi_session_to_shost(cls_session);
1627         struct iscsi_host *ihost = shost_priv(shost);
1628         struct iscsi_conn *conn = cls_conn->dd_data;
1629         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1630         struct sock *sk;
1631         struct socket *sock;
1632         int err;
1633
1634         /* lookup for existing socket */
1635         sock = sockfd_lookup((int)transport_eph, &err);
1636         if (!sock) {
1637                 iscsi_conn_printk(KERN_ERR, conn,
1638                                   "sockfd_lookup failed %d\n", err);
1639                 return -EEXIST;
1640         }
1641         /*
1642          * copy these values now because if we drop the session
1643          * userspace may still want to query the values since we will
1644          * be using them for the reconnect
1645          */
1646         err = iscsi_tcp_get_addr(conn, sock, conn->portal_address,
1647                                  &conn->portal_port, kernel_getpeername);
1648         if (err)
1649                 goto free_socket;
1650
1651         err = iscsi_tcp_get_addr(conn, sock, ihost->local_address,
1652                                 &ihost->local_port, kernel_getsockname);
1653         if (err)
1654                 goto free_socket;
1655
1656         err = iscsi_conn_bind(cls_session, cls_conn, is_leading);
1657         if (err)
1658                 goto free_socket;
1659
1660         /* bind iSCSI connection and socket */
1661         tcp_conn->sock = sock;
1662
1663         /* setup Socket parameters */
1664         sk = sock->sk;
1665         sk->sk_reuse = 1;
1666         sk->sk_sndtimeo = 15 * HZ; /* FIXME: make it configurable */
1667         sk->sk_allocation = GFP_ATOMIC;
1668
1669         iscsi_conn_set_callbacks(conn);
1670         tcp_conn->sendpage = tcp_conn->sock->ops->sendpage;
1671         /*
1672          * set receive state machine into initial state
1673          */
1674         iscsi_tcp_hdr_recv_prep(tcp_conn);
1675         return 0;
1676
1677 free_socket:
1678         sockfd_put(sock);
1679         return err;
1680 }
1681
1682 static int
1683 iscsi_r2tpool_alloc(struct iscsi_session *session)
1684 {
1685         int i;
1686         int cmd_i;
1687
1688         /*
1689          * initialize per-task: R2T pool and xmit queue
1690          */
1691         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
1692                 struct iscsi_task *task = session->cmds[cmd_i];
1693                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1694
1695                 /*
1696                  * pre-allocated x4 as much r2ts to handle race when
1697                  * target acks DataOut faster than we data_xmit() queues
1698                  * could replenish r2tqueue.
1699                  */
1700
1701                 /* R2T pool */
1702                 if (iscsi_pool_init(&tcp_task->r2tpool, session->max_r2t * 4, NULL,
1703                                     sizeof(struct iscsi_r2t_info))) {
1704                         goto r2t_alloc_fail;
1705                 }
1706
1707                 /* R2T xmit queue */
1708                 tcp_task->r2tqueue = kfifo_alloc(
1709                       session->max_r2t * 4 * sizeof(void*), GFP_KERNEL, NULL);
1710                 if (tcp_task->r2tqueue == ERR_PTR(-ENOMEM)) {
1711                         iscsi_pool_free(&tcp_task->r2tpool);
1712                         goto r2t_alloc_fail;
1713                 }
1714         }
1715
1716         return 0;
1717
1718 r2t_alloc_fail:
1719         for (i = 0; i < cmd_i; i++) {
1720                 struct iscsi_task *task = session->cmds[i];
1721                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1722
1723                 kfifo_free(tcp_task->r2tqueue);
1724                 iscsi_pool_free(&tcp_task->r2tpool);
1725         }
1726         return -ENOMEM;
1727 }
1728
1729 static void
1730 iscsi_r2tpool_free(struct iscsi_session *session)
1731 {
1732         int i;
1733
1734         for (i = 0; i < session->cmds_max; i++) {
1735                 struct iscsi_task *task = session->cmds[i];
1736                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1737
1738                 kfifo_free(tcp_task->r2tqueue);
1739                 iscsi_pool_free(&tcp_task->r2tpool);
1740         }
1741 }
1742
1743 static int
1744 iscsi_conn_set_param(struct iscsi_cls_conn *cls_conn, enum iscsi_param param,
1745                      char *buf, int buflen)
1746 {
1747         struct iscsi_conn *conn = cls_conn->dd_data;
1748         struct iscsi_session *session = conn->session;
1749         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1750         int value;
1751
1752         switch(param) {
1753         case ISCSI_PARAM_HDRDGST_EN:
1754                 iscsi_set_param(cls_conn, param, buf, buflen);
1755                 break;
1756         case ISCSI_PARAM_DATADGST_EN:
1757                 iscsi_set_param(cls_conn, param, buf, buflen);
1758                 tcp_conn->sendpage = conn->datadgst_en ?
1759                         sock_no_sendpage : tcp_conn->sock->ops->sendpage;
1760                 break;
1761         case ISCSI_PARAM_MAX_R2T:
1762                 sscanf(buf, "%d", &value);
1763                 if (value <= 0 || !is_power_of_2(value))
1764                         return -EINVAL;
1765                 if (session->max_r2t == value)
1766                         break;
1767                 iscsi_r2tpool_free(session);
1768                 iscsi_set_param(cls_conn, param, buf, buflen);
1769                 if (iscsi_r2tpool_alloc(session))
1770                         return -ENOMEM;
1771                 break;
1772         default:
1773                 return iscsi_set_param(cls_conn, param, buf, buflen);
1774         }
1775
1776         return 0;
1777 }
1778
1779 static int
1780 iscsi_tcp_conn_get_param(struct iscsi_cls_conn *cls_conn,
1781                          enum iscsi_param param, char *buf)
1782 {
1783         struct iscsi_conn *conn = cls_conn->dd_data;
1784         int len;
1785
1786         switch(param) {
1787         case ISCSI_PARAM_CONN_PORT:
1788                 spin_lock_bh(&conn->session->lock);
1789                 len = sprintf(buf, "%hu\n", conn->portal_port);
1790                 spin_unlock_bh(&conn->session->lock);
1791                 break;
1792         case ISCSI_PARAM_CONN_ADDRESS:
1793                 spin_lock_bh(&conn->session->lock);
1794                 len = sprintf(buf, "%s\n", conn->portal_address);
1795                 spin_unlock_bh(&conn->session->lock);
1796                 break;
1797         default:
1798                 return iscsi_conn_get_param(cls_conn, param, buf);
1799         }
1800
1801         return len;
1802 }
1803
1804 static void
1805 iscsi_conn_get_stats(struct iscsi_cls_conn *cls_conn, struct iscsi_stats *stats)
1806 {
1807         struct iscsi_conn *conn = cls_conn->dd_data;
1808         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1809
1810         stats->txdata_octets = conn->txdata_octets;
1811         stats->rxdata_octets = conn->rxdata_octets;
1812         stats->scsicmd_pdus = conn->scsicmd_pdus_cnt;
1813         stats->dataout_pdus = conn->dataout_pdus_cnt;
1814         stats->scsirsp_pdus = conn->scsirsp_pdus_cnt;
1815         stats->datain_pdus = conn->datain_pdus_cnt;
1816         stats->r2t_pdus = conn->r2t_pdus_cnt;
1817         stats->tmfcmd_pdus = conn->tmfcmd_pdus_cnt;
1818         stats->tmfrsp_pdus = conn->tmfrsp_pdus_cnt;
1819         stats->custom_length = 3;
1820         strcpy(stats->custom[0].desc, "tx_sendpage_failures");
1821         stats->custom[0].value = tcp_conn->sendpage_failures_cnt;
1822         strcpy(stats->custom[1].desc, "rx_discontiguous_hdr");
1823         stats->custom[1].value = tcp_conn->discontiguous_hdr_cnt;
1824         strcpy(stats->custom[2].desc, "eh_abort_cnt");
1825         stats->custom[2].value = conn->eh_abort_cnt;
1826 }
1827
1828 static struct iscsi_cls_session *
1829 iscsi_tcp_session_create(struct iscsi_endpoint *ep, uint16_t cmds_max,
1830                          uint16_t qdepth, uint32_t initial_cmdsn,
1831                          uint32_t *hostno)
1832 {
1833         struct iscsi_cls_session *cls_session;
1834         struct iscsi_session *session;
1835         struct Scsi_Host *shost;
1836         int cmd_i;
1837
1838         if (ep) {
1839                 printk(KERN_ERR "iscsi_tcp: invalid ep %p.\n", ep);
1840                 return NULL;
1841         }
1842
1843         shost = iscsi_host_alloc(&iscsi_sht, 0, qdepth);
1844         if (!shost)
1845                 return NULL;
1846         shost->transportt = iscsi_tcp_scsi_transport;
1847         shost->max_lun = iscsi_max_lun;
1848         shost->max_id = 0;
1849         shost->max_channel = 0;
1850         shost->max_cmd_len = SCSI_MAX_VARLEN_CDB_SIZE;
1851
1852         if (iscsi_host_add(shost, NULL))
1853                 goto free_host;
1854         *hostno = shost->host_no;
1855
1856         cls_session = iscsi_session_setup(&iscsi_tcp_transport, shost, cmds_max,
1857                                           sizeof(struct iscsi_tcp_task),
1858                                           initial_cmdsn, 0);
1859         if (!cls_session)
1860                 goto remove_host;
1861         session = cls_session->dd_data;
1862
1863         shost->can_queue = session->scsi_cmds_max;
1864         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
1865                 struct iscsi_task *task = session->cmds[cmd_i];
1866                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1867
1868                 task->hdr = &tcp_task->hdr.cmd_hdr;
1869                 task->hdr_max = sizeof(tcp_task->hdr) - ISCSI_DIGEST_SIZE;
1870         }
1871
1872         if (iscsi_r2tpool_alloc(session))
1873                 goto remove_session;
1874         return cls_session;
1875
1876 remove_session:
1877         iscsi_session_teardown(cls_session);
1878 remove_host:
1879         iscsi_host_remove(shost);
1880 free_host:
1881         iscsi_host_free(shost);
1882         return NULL;
1883 }
1884
1885 static void iscsi_tcp_session_destroy(struct iscsi_cls_session *cls_session)
1886 {
1887         struct Scsi_Host *shost = iscsi_session_to_shost(cls_session);
1888
1889         iscsi_r2tpool_free(cls_session->dd_data);
1890         iscsi_session_teardown(cls_session);
1891
1892         iscsi_host_remove(shost);
1893         iscsi_host_free(shost);
1894 }
1895
1896 static int iscsi_tcp_slave_configure(struct scsi_device *sdev)
1897 {
1898         blk_queue_bounce_limit(sdev->request_queue, BLK_BOUNCE_ANY);
1899         blk_queue_dma_alignment(sdev->request_queue, 0);
1900         return 0;
1901 }
1902
1903 static struct scsi_host_template iscsi_sht = {
1904         .module                 = THIS_MODULE,
1905         .name                   = "iSCSI Initiator over TCP/IP",
1906         .queuecommand           = iscsi_queuecommand,
1907         .change_queue_depth     = iscsi_change_queue_depth,
1908         .can_queue              = ISCSI_DEF_XMIT_CMDS_MAX - 1,
1909         .sg_tablesize           = 4096,
1910         .max_sectors            = 0xFFFF,
1911         .cmd_per_lun            = ISCSI_DEF_CMD_PER_LUN,
1912         .eh_abort_handler       = iscsi_eh_abort,
1913         .eh_device_reset_handler= iscsi_eh_device_reset,
1914         .eh_target_reset_handler= iscsi_eh_target_reset,
1915         .use_clustering         = DISABLE_CLUSTERING,
1916         .slave_configure        = iscsi_tcp_slave_configure,
1917         .proc_name              = "iscsi_tcp",
1918         .this_id                = -1,
1919 };
1920
1921 static struct iscsi_transport iscsi_tcp_transport = {
1922         .owner                  = THIS_MODULE,
1923         .name                   = "tcp",
1924         .caps                   = CAP_RECOVERY_L0 | CAP_MULTI_R2T | CAP_HDRDGST
1925                                   | CAP_DATADGST,
1926         .param_mask             = ISCSI_MAX_RECV_DLENGTH |
1927                                   ISCSI_MAX_XMIT_DLENGTH |
1928                                   ISCSI_HDRDGST_EN |
1929                                   ISCSI_DATADGST_EN |
1930                                   ISCSI_INITIAL_R2T_EN |
1931                                   ISCSI_MAX_R2T |
1932                                   ISCSI_IMM_DATA_EN |
1933                                   ISCSI_FIRST_BURST |
1934                                   ISCSI_MAX_BURST |
1935                                   ISCSI_PDU_INORDER_EN |
1936                                   ISCSI_DATASEQ_INORDER_EN |
1937                                   ISCSI_ERL |
1938                                   ISCSI_CONN_PORT |
1939                                   ISCSI_CONN_ADDRESS |
1940                                   ISCSI_EXP_STATSN |
1941                                   ISCSI_PERSISTENT_PORT |
1942                                   ISCSI_PERSISTENT_ADDRESS |
1943                                   ISCSI_TARGET_NAME | ISCSI_TPGT |
1944                                   ISCSI_USERNAME | ISCSI_PASSWORD |
1945                                   ISCSI_USERNAME_IN | ISCSI_PASSWORD_IN |
1946                                   ISCSI_FAST_ABORT | ISCSI_ABORT_TMO |
1947                                   ISCSI_LU_RESET_TMO |
1948                                   ISCSI_PING_TMO | ISCSI_RECV_TMO |
1949                                   ISCSI_IFACE_NAME | ISCSI_INITIATOR_NAME,
1950         .host_param_mask        = ISCSI_HOST_HWADDRESS | ISCSI_HOST_IPADDRESS |
1951                                   ISCSI_HOST_INITIATOR_NAME |
1952                                   ISCSI_HOST_NETDEV_NAME,
1953         /* session management */
1954         .create_session         = iscsi_tcp_session_create,
1955         .destroy_session        = iscsi_tcp_session_destroy,
1956         /* connection management */
1957         .create_conn            = iscsi_tcp_conn_create,
1958         .bind_conn              = iscsi_tcp_conn_bind,
1959         .destroy_conn           = iscsi_tcp_conn_destroy,
1960         .set_param              = iscsi_conn_set_param,
1961         .get_conn_param         = iscsi_tcp_conn_get_param,
1962         .get_session_param      = iscsi_session_get_param,
1963         .start_conn             = iscsi_conn_start,
1964         .stop_conn              = iscsi_tcp_conn_stop,
1965         /* iscsi host params */
1966         .get_host_param         = iscsi_host_get_param,
1967         .set_host_param         = iscsi_host_set_param,
1968         /* IO */
1969         .send_pdu               = iscsi_conn_send_pdu,
1970         .get_stats              = iscsi_conn_get_stats,
1971         .init_task              = iscsi_tcp_task_init,
1972         .xmit_task              = iscsi_tcp_task_xmit,
1973         .cleanup_task           = iscsi_tcp_cleanup_task,
1974         /* recovery */
1975         .session_recovery_timedout = iscsi_session_recovery_timedout,
1976 };
1977
1978 static int __init
1979 iscsi_tcp_init(void)
1980 {
1981         if (iscsi_max_lun < 1) {
1982                 printk(KERN_ERR "iscsi_tcp: Invalid max_lun value of %u\n",
1983                        iscsi_max_lun);
1984                 return -EINVAL;
1985         }
1986
1987         iscsi_tcp_scsi_transport = iscsi_register_transport(
1988                                                         &iscsi_tcp_transport);
1989         if (!iscsi_tcp_scsi_transport)
1990                 return -ENODEV;
1991
1992         return 0;
1993 }
1994
1995 static void __exit
1996 iscsi_tcp_exit(void)
1997 {
1998         iscsi_unregister_transport(&iscsi_tcp_transport);
1999 }
2000
2001 module_init(iscsi_tcp_init);
2002 module_exit(iscsi_tcp_exit);