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_rsp - SCSI Data-In Response processing
 527  * @conn: iscsi connection
 528  * @task: scsi command task
 529  **/
 530 static int
 531 iscsi_data_rsp(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         struct iscsi_session *session = conn->session;
 537         struct scsi_cmnd *sc = task->sc;
 538         int datasn = be32_to_cpu(rhdr->datasn);
 539         unsigned total_in_length = scsi_in(sc)->length;
 540
 541         iscsi_update_cmdsn(session, (struct iscsi_nopin*)rhdr);
 542         if (tcp_conn->in.datalen == 0)
 543                 return 0;
 544
 545         if (tcp_task->exp_datasn != datasn) {
 546                 debug_tcp("%s: task->exp_datasn(%d) != rhdr->datasn(%d)\n",
 547                           __func__, tcp_task->exp_datasn, datasn);
 548                 return ISCSI_ERR_DATASN;
 549         }
 550
 551         tcp_task->exp_datasn++;
 552
 553         tcp_task->data_offset = be32_to_cpu(rhdr->offset);
 554         if (tcp_task->data_offset + tcp_conn->in.datalen > total_in_length) {
 555                 debug_tcp("%s: data_offset(%d) + data_len(%d) > total_length_in(%d)\n",
 556                           __func__, tcp_task->data_offset,
 557                           tcp_conn->in.datalen, total_in_length);
 558                 return ISCSI_ERR_DATA_OFFSET;
 559         }
 560
 561         if (rhdr->flags & ISCSI_FLAG_DATA_STATUS) {
 562                 sc->result = (DID_OK << 16) | rhdr->cmd_status;
 563                 conn->exp_statsn = be32_to_cpu(rhdr->statsn) + 1;
 564                 if (rhdr->flags & (ISCSI_FLAG_DATA_UNDERFLOW |
 565                                    ISCSI_FLAG_DATA_OVERFLOW)) {
 566                         int res_count = be32_to_cpu(rhdr->residual_count);
 567
 568                         if (res_count > 0 &&
 569                             (rhdr->flags & ISCSI_FLAG_CMD_OVERFLOW ||
 570                              res_count <= total_in_length))
 571                                 scsi_in(sc)->resid = res_count;
 572                         else
 573                                 sc->result = (DID_BAD_TARGET << 16) |
 574                                         rhdr->cmd_status;
 575                 }
 576         }
 577
 578         conn->datain_pdus_cnt++;
 579         return 0;
 580 }
 581
 582 /**
 583  * iscsi_solicit_data_init - initialize first Data-Out
 584  * @conn: iscsi connection
 585  * @task: scsi command task
 586  * @r2t: R2T info
 587  *
 588  * Notes:
 589  *      Initialize first Data-Out within this R2T sequence and finds
 590  *      proper data_offset within this SCSI command.
 591  *
 592  *      This function is called with connection lock taken.
 593  **/
 594 static void
 595 iscsi_solicit_data_init(struct iscsi_conn *conn, struct iscsi_task *task,
 596                         struct iscsi_r2t_info *r2t)
 597 {
 598         struct iscsi_data *hdr;
 599
 600         hdr = &r2t->dtask.hdr;
 601         memset(hdr, 0, sizeof(struct iscsi_data));
 602         hdr->ttt = r2t->ttt;
 603         hdr->datasn = cpu_to_be32(r2t->solicit_datasn);
 604         r2t->solicit_datasn++;
 605         hdr->opcode = ISCSI_OP_SCSI_DATA_OUT;
 606         memcpy(hdr->lun, task->hdr->lun, sizeof(hdr->lun));
 607         hdr->itt = task->hdr->itt;
 608         hdr->exp_statsn = r2t->exp_statsn;
 609         hdr->offset = cpu_to_be32(r2t->data_offset);
 610         if (r2t->data_length > conn->max_xmit_dlength) {
 611                 hton24(hdr->dlength, conn->max_xmit_dlength);
 612                 r2t->data_count = conn->max_xmit_dlength;
 613                 hdr->flags = 0;
 614         } else {
 615                 hton24(hdr->dlength, r2t->data_length);
 616                 r2t->data_count = r2t->data_length;
 617                 hdr->flags = ISCSI_FLAG_CMD_FINAL;
 618         }
 619         conn->dataout_pdus_cnt++;
 620
 621         r2t->sent = 0;
 622 }
 623
 624 /**
 625  * iscsi_r2t_rsp - iSCSI R2T Response processing
 626  * @conn: iscsi connection
 627  * @task: scsi command task
 628  **/
 629 static int
 630 iscsi_r2t_rsp(struct iscsi_conn *conn, struct iscsi_task *task)
 631 {
 632         struct iscsi_r2t_info *r2t;
 633         struct iscsi_session *session = conn->session;
 634         struct iscsi_tcp_task *tcp_task = task->dd_data;
 635         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 636         struct iscsi_r2t_rsp *rhdr = (struct iscsi_r2t_rsp *)tcp_conn->in.hdr;
 637         int r2tsn = be32_to_cpu(rhdr->r2tsn);
 638         int rc;
 639
 640         if (tcp_conn->in.datalen) {
 641                 iscsi_conn_printk(KERN_ERR, conn,
 642                                   "invalid R2t with datalen %d\n",
 643                                   tcp_conn->in.datalen);
 644                 return ISCSI_ERR_DATALEN;
 645         }
 646
 647         if (tcp_task->exp_datasn != r2tsn){
 648                 debug_tcp("%s: task->exp_datasn(%d) != rhdr->r2tsn(%d)\n",
 649                           __func__, tcp_task->exp_datasn, r2tsn);
 650                 return ISCSI_ERR_R2TSN;
 651         }
 652
 653         /* fill-in new R2T associated with the task */
 654         iscsi_update_cmdsn(session, (struct iscsi_nopin*)rhdr);
 655
 656         if (!task->sc || session->state != ISCSI_STATE_LOGGED_IN) {
 657                 iscsi_conn_printk(KERN_INFO, conn,
 658                                   "dropping R2T itt %d in recovery.\n",
 659                                   task->itt);
 660                 return 0;
 661         }
 662
 663         rc = __kfifo_get(tcp_task->r2tpool.queue, (void*)&r2t, sizeof(void*));
 664         BUG_ON(!rc);
 665
 666         r2t->exp_statsn = rhdr->statsn;
 667         r2t->data_length = be32_to_cpu(rhdr->data_length);
 668         if (r2t->data_length == 0) {
 669                 iscsi_conn_printk(KERN_ERR, conn,
 670                                   "invalid R2T with zero data len\n");
 671                 __kfifo_put(tcp_task->r2tpool.queue, (void*)&r2t,
 672                             sizeof(void*));
 673                 return ISCSI_ERR_DATALEN;
 674         }
 675
 676         if (r2t->data_length > session->max_burst)
 677                 debug_scsi("invalid R2T with data len %u and max burst %u."
 678                            "Attempting to execute request.\n",
 679                             r2t->data_length, session->max_burst);
 680
 681         r2t->data_offset = be32_to_cpu(rhdr->data_offset);
 682         if (r2t->data_offset + r2t->data_length > scsi_out(task->sc)->length) {
 683                 iscsi_conn_printk(KERN_ERR, conn,
 684                                   "invalid R2T with data len %u at offset %u "
 685                                   "and total length %d\n", r2t->data_length,
 686                                   r2t->data_offset, scsi_out(task->sc)->length);
 687                 __kfifo_put(tcp_task->r2tpool.queue, (void*)&r2t,
 688                             sizeof(void*));
 689                 return ISCSI_ERR_DATALEN;
 690         }
 691
 692         r2t->ttt = rhdr->ttt; /* no flip */
 693         r2t->solicit_datasn = 0;
 694
 695         iscsi_solicit_data_init(conn, task, r2t);
 696
 697         tcp_task->exp_datasn = r2tsn + 1;
 698         __kfifo_put(tcp_task->r2tqueue, (void*)&r2t, sizeof(void*));
 699         conn->r2t_pdus_cnt++;
 700
 701         iscsi_requeue_task(task);
 702         return 0;
 703 }
 704
 705 /*
 706  * Handle incoming reply to DataIn command
 707  */
 708 static int
 709 iscsi_tcp_process_data_in(struct iscsi_tcp_conn *tcp_conn,
 710                           struct iscsi_segment *segment)
 711 {
 712         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 713         struct iscsi_hdr *hdr = tcp_conn->in.hdr;
 714         int rc;
 715
 716         if (!iscsi_tcp_dgst_verify(tcp_conn, segment))
 717                 return ISCSI_ERR_DATA_DGST;
 718
 719         /* check for non-exceptional status */
 720         if (hdr->flags & ISCSI_FLAG_DATA_STATUS) {
 721                 rc = iscsi_complete_pdu(conn, tcp_conn->in.hdr, NULL, 0);
 722                 if (rc)
 723                         return rc;
 724         }
 725
 726         iscsi_tcp_hdr_recv_prep(tcp_conn);
 727         return 0;
 728 }
 729
 730 /**
 731  * iscsi_tcp_hdr_dissect - process PDU header
 732  * @conn: iSCSI connection
 733  * @hdr: PDU header
 734  *
 735  * This function analyzes the header of the PDU received,
 736  * and performs several sanity checks. If the PDU is accompanied
 737  * by data, the receive buffer is set up to copy the incoming data
 738  * to the correct location.
 739  */
 740 static int
 741 iscsi_tcp_hdr_dissect(struct iscsi_conn *conn, struct iscsi_hdr *hdr)
 742 {
 743         int rc = 0, opcode, ahslen;
 744         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 745         struct iscsi_task *task;
 746
 747         /* verify PDU length */
 748         tcp_conn->in.datalen = ntoh24(hdr->dlength);
 749         if (tcp_conn->in.datalen > conn->max_recv_dlength) {
 750                 iscsi_conn_printk(KERN_ERR, conn,
 751                                   "iscsi_tcp: datalen %d > %d\n",
 752                                   tcp_conn->in.datalen, conn->max_recv_dlength);
 753                 return ISCSI_ERR_DATALEN;
 754         }
 755
 756         /* Additional header segments. So far, we don't
 757          * process additional headers.
 758          */
 759         ahslen = hdr->hlength << 2;
 760
 761         opcode = hdr->opcode & ISCSI_OPCODE_MASK;
 762         /* verify itt (itt encoding: age+cid+itt) */
 763         rc = iscsi_verify_itt(conn, hdr->itt);
 764         if (rc)
 765                 return rc;
 766
 767         debug_tcp("opcode 0x%x ahslen %d datalen %d\n",
 768                   opcode, ahslen, tcp_conn->in.datalen);
 769
 770         switch(opcode) {
 771         case ISCSI_OP_SCSI_DATA_IN:
 772                 spin_lock(&conn->session->lock);
 773                 task = iscsi_itt_to_ctask(conn, hdr->itt);
 774                 if (!task)
 775                         rc = ISCSI_ERR_BAD_ITT;
 776                 else
 777                         rc = iscsi_data_rsp(conn, task);
 778                 if (rc) {
 779                         spin_unlock(&conn->session->lock);
 780                         break;
 781                 }
 782
 783                 if (tcp_conn->in.datalen) {
 784                         struct iscsi_tcp_task *tcp_task = task->dd_data;
 785                         struct hash_desc *rx_hash = NULL;
 786                         struct scsi_data_buffer *sdb = scsi_in(task->sc);
 787
 788                         /*
 789                          * Setup copy of Data-In into the Scsi_Cmnd
 790                          * Scatterlist case:
 791                          * We set up the iscsi_segment to point to the next
 792                          * scatterlist entry to copy to. As we go along,
 793                          * we move on to the next scatterlist entry and
 794                          * update the digest per-entry.
 795                          */
 796                         if (conn->datadgst_en)
 797                                 rx_hash = &tcp_conn->rx_hash;
 798
 799                         debug_tcp("iscsi_tcp_begin_data_in(%p, offset=%d, "
 800                                   "datalen=%d)\n", tcp_conn,
 801                                   tcp_task->data_offset,
 802                                   tcp_conn->in.datalen);
 803                         rc = iscsi_segment_seek_sg(&tcp_conn->in.segment,
 804                                                    sdb->table.sgl,
 805                                                    sdb->table.nents,
 806                                                    tcp_task->data_offset,
 807                                                    tcp_conn->in.datalen,
 808                                                    iscsi_tcp_process_data_in,
 809                                                    rx_hash);
 810                         spin_unlock(&conn->session->lock);
 811                         return rc;
 812                 }
 813                 rc = __iscsi_complete_pdu(conn, hdr, NULL, 0);
 814                 spin_unlock(&conn->session->lock);
 815                 break;
 816         case ISCSI_OP_SCSI_CMD_RSP:
 817                 if (tcp_conn->in.datalen) {
 818                         iscsi_tcp_data_recv_prep(tcp_conn);
 819                         return 0;
 820                 }
 821                 rc = iscsi_complete_pdu(conn, hdr, NULL, 0);
 822                 break;
 823         case ISCSI_OP_R2T:
 824                 spin_lock(&conn->session->lock);
 825                 task = iscsi_itt_to_ctask(conn, hdr->itt);
 826                 if (!task)
 827                         rc = ISCSI_ERR_BAD_ITT;
 828                 else if (ahslen)
 829                         rc = ISCSI_ERR_AHSLEN;
 830                 else if (task->sc->sc_data_direction == DMA_TO_DEVICE)
 831                         rc = iscsi_r2t_rsp(conn, task);
 832                 else
 833                         rc = ISCSI_ERR_PROTO;
 834                 spin_unlock(&conn->session->lock);
 835                 break;
 836         case ISCSI_OP_LOGIN_RSP:
 837         case ISCSI_OP_TEXT_RSP:
 838         case ISCSI_OP_REJECT:
 839         case ISCSI_OP_ASYNC_EVENT:
 840                 /*
 841                  * It is possible that we could get a PDU with a buffer larger
 842                  * than 8K, but there are no targets that currently do this.
 843                  * For now we fail until we find a vendor that needs it
 844                  */
 845                 if (ISCSI_DEF_MAX_RECV_SEG_LEN < tcp_conn->in.datalen) {
 846                         iscsi_conn_printk(KERN_ERR, conn,
 847                                           "iscsi_tcp: received buffer of "
 848                                           "len %u but conn buffer is only %u "
 849                                           "(opcode %0x)\n",
 850                                           tcp_conn->in.datalen,
 851                                           ISCSI_DEF_MAX_RECV_SEG_LEN, opcode);
 852                         rc = ISCSI_ERR_PROTO;
 853                         break;
 854                 }
 855
 856                 /* If there's data coming in with the response,
 857                  * receive it to the connection's buffer.
 858                  */
 859                 if (tcp_conn->in.datalen) {
 860                         iscsi_tcp_data_recv_prep(tcp_conn);
 861                         return 0;
 862                 }
 863         /* fall through */
 864         case ISCSI_OP_LOGOUT_RSP:
 865         case ISCSI_OP_NOOP_IN:
 866         case ISCSI_OP_SCSI_TMFUNC_RSP:
 867                 rc = iscsi_complete_pdu(conn, hdr, NULL, 0);
 868                 break;
 869         default:
 870                 rc = ISCSI_ERR_BAD_OPCODE;
 871                 break;
 872         }
 873
 874         if (rc == 0) {
 875                 /* Anything that comes with data should have
 876                  * been handled above. */
 877                 if (tcp_conn->in.datalen)
 878                         return ISCSI_ERR_PROTO;
 879                 iscsi_tcp_hdr_recv_prep(tcp_conn);
 880         }
 881
 882         return rc;
 883 }
 884
 885 /**
 886  * iscsi_tcp_hdr_recv_done - process PDU header
 887  *
 888  * This is the callback invoked when the PDU header has
 889  * been received. If the header is followed by additional
 890  * header segments, we go back for more data.
 891  */
 892 static int
 893 iscsi_tcp_hdr_recv_done(struct iscsi_tcp_conn *tcp_conn,
 894                         struct iscsi_segment *segment)
 895 {
 896         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 897         struct iscsi_hdr *hdr;
 898
 899         /* Check if there are additional header segments
 900          * *prior* to computing the digest, because we
 901          * may need to go back to the caller for more.
 902          */
 903         hdr = (struct iscsi_hdr *) tcp_conn->in.hdr_buf;
 904         if (segment->copied == sizeof(struct iscsi_hdr) && hdr->hlength) {
 905                 /* Bump the header length - the caller will
 906                  * just loop around and get the AHS for us, and
 907                  * call again. */
 908                 unsigned int ahslen = hdr->hlength << 2;
 909
 910                 /* Make sure we don't overflow */
 911                 if (sizeof(*hdr) + ahslen > sizeof(tcp_conn->in.hdr_buf))
 912                         return ISCSI_ERR_AHSLEN;
 913
 914                 segment->total_size += ahslen;
 915                 segment->size += ahslen;
 916                 return 0;
 917         }
 918
 919         /* We're done processing the header. See if we're doing
 920          * header digests; if so, set up the recv_digest buffer
 921          * and go back for more. */
 922         if (conn->hdrdgst_en) {
 923                 if (segment->digest_len == 0) {
 924                         iscsi_tcp_segment_splice_digest(segment,
 925                                                         segment->recv_digest);
 926                         return 0;
 927                 }
 928                 iscsi_tcp_dgst_header(&tcp_conn->rx_hash, hdr,
 929                                       segment->total_copied - ISCSI_DIGEST_SIZE,
 930                                       segment->digest);
 931
 932                 if (!iscsi_tcp_dgst_verify(tcp_conn, segment))
 933                         return ISCSI_ERR_HDR_DGST;
 934         }
 935
 936         tcp_conn->in.hdr = hdr;
 937         return iscsi_tcp_hdr_dissect(conn, hdr);
 938 }
 939
 940 /**
 941  * iscsi_tcp_recv - TCP receive in sendfile fashion
 942  * @rd_desc: read descriptor
 943  * @skb: socket buffer
 944  * @offset: offset in skb
 945  * @len: skb->len - offset
 946  **/
 947 static int
 948 iscsi_tcp_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 949                unsigned int offset, size_t len)
 950 {
 951         struct iscsi_conn *conn = rd_desc->arg.data;
 952         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 953         struct iscsi_segment *segment = &tcp_conn->in.segment;
 954         struct skb_seq_state seq;
 955         unsigned int consumed = 0;
 956         int rc = 0;
 957
 958         debug_tcp("in %d bytes\n", skb->len - offset);
 959
 960         if (unlikely(conn->suspend_rx)) {
 961                 debug_tcp("conn %d Rx suspended!\n", conn->id);
 962                 return 0;
 963         }
 964
 965         skb_prepare_seq_read(skb, offset, skb->len, &seq);
 966         while (1) {
 967                 unsigned int avail;
 968                 const u8 *ptr;
 969
 970                 avail = skb_seq_read(consumed, &ptr, &seq);
 971                 if (avail == 0) {
 972                         debug_tcp("no more data avail. Consumed %d\n",
 973                                   consumed);
 974                         break;
 975                 }
 976                 BUG_ON(segment->copied >= segment->size);
 977
 978                 debug_tcp("skb %p ptr=%p avail=%u\n", skb, ptr, avail);
 979                 rc = iscsi_tcp_segment_recv(tcp_conn, segment, ptr, avail);
 980                 BUG_ON(rc == 0);
 981                 consumed += rc;
 982
 983                 if (segment->total_copied >= segment->total_size) {
 984                         debug_tcp("segment done\n");
 985                         rc = segment->done(tcp_conn, segment);
 986                         if (rc != 0) {
 987                                 skb_abort_seq_read(&seq);
 988                                 goto error;
 989                         }
 990
 991                         /* The done() functions sets up the
 992                          * next segment. */
 993                 }
 994         }
 995         skb_abort_seq_read(&seq);
 996         conn->rxdata_octets += consumed;
 997         return consumed;
 998
 999 error:
1000         debug_tcp("Error receiving PDU, errno=%d\n", rc);
1001         iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
1002         return 0;
1003 }
1004
1005 static void
1006 iscsi_tcp_data_ready(struct sock *sk, int flag)
1007 {
1008         struct iscsi_conn *conn = sk->sk_user_data;
1009         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1010         read_descriptor_t rd_desc;
1011
1012         read_lock(&sk->sk_callback_lock);
1013
1014         /*
1015          * Use rd_desc to pass 'conn' to iscsi_tcp_recv.
1016          * We set count to 1 because we want the network layer to
1017          * hand us all the skbs that are available. iscsi_tcp_recv
1018          * handled pdus that cross buffers or pdus that still need data.
1019          */
1020         rd_desc.arg.data = conn;
1021         rd_desc.count = 1;
1022         tcp_read_sock(sk, &rd_desc, iscsi_tcp_recv);
1023
1024         read_unlock(&sk->sk_callback_lock);
1025
1026         /* If we had to (atomically) map a highmem page,
1027          * unmap it now. */
1028         iscsi_tcp_segment_unmap(&tcp_conn->in.segment);
1029 }
1030
1031 static void
1032 iscsi_tcp_state_change(struct sock *sk)
1033 {
1034         struct iscsi_tcp_conn *tcp_conn;
1035         struct iscsi_conn *conn;
1036         struct iscsi_session *session;
1037         void (*old_state_change)(struct sock *);
1038
1039         read_lock(&sk->sk_callback_lock);
1040
1041         conn = (struct iscsi_conn*)sk->sk_user_data;
1042         session = conn->session;
1043
1044         if ((sk->sk_state == TCP_CLOSE_WAIT ||
1045              sk->sk_state == TCP_CLOSE) &&
1046             !atomic_read(&sk->sk_rmem_alloc)) {
1047                 debug_tcp("iscsi_tcp_state_change: TCP_CLOSE|TCP_CLOSE_WAIT\n");
1048                 iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
1049         }
1050
1051         tcp_conn = conn->dd_data;
1052         old_state_change = tcp_conn->old_state_change;
1053
1054         read_unlock(&sk->sk_callback_lock);
1055
1056         old_state_change(sk);
1057 }
1058
1059 /**
1060  * iscsi_write_space - Called when more output buffer space is available
1061  * @sk: socket space is available for
1062  **/
1063 static void
1064 iscsi_write_space(struct sock *sk)
1065 {
1066         struct iscsi_conn *conn = (struct iscsi_conn*)sk->sk_user_data;
1067         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1068
1069         tcp_conn->old_write_space(sk);
1070         debug_tcp("iscsi_write_space: cid %d\n", conn->id);
1071         scsi_queue_work(conn->session->host, &conn->xmitwork);
1072 }
1073
1074 static void
1075 iscsi_conn_set_callbacks(struct iscsi_conn *conn)
1076 {
1077         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1078         struct sock *sk = tcp_conn->sock->sk;
1079
1080         /* assign new callbacks */
1081         write_lock_bh(&sk->sk_callback_lock);
1082         sk->sk_user_data = conn;
1083         tcp_conn->old_data_ready = sk->sk_data_ready;
1084         tcp_conn->old_state_change = sk->sk_state_change;
1085         tcp_conn->old_write_space = sk->sk_write_space;
1086         sk->sk_data_ready = iscsi_tcp_data_ready;
1087         sk->sk_state_change = iscsi_tcp_state_change;
1088         sk->sk_write_space = iscsi_write_space;
1089         write_unlock_bh(&sk->sk_callback_lock);
1090 }
1091
1092 static void
1093 iscsi_conn_restore_callbacks(struct iscsi_tcp_conn *tcp_conn)
1094 {
1095         struct sock *sk = tcp_conn->sock->sk;
1096
1097         /* restore socket callbacks, see also: iscsi_conn_set_callbacks() */
1098         write_lock_bh(&sk->sk_callback_lock);
1099         sk->sk_user_data    = NULL;
1100         sk->sk_data_ready   = tcp_conn->old_data_ready;
1101         sk->sk_state_change = tcp_conn->old_state_change;
1102         sk->sk_write_space  = tcp_conn->old_write_space;
1103         sk->sk_no_check  = 0;
1104         write_unlock_bh(&sk->sk_callback_lock);
1105 }
1106
1107 /**
1108  * iscsi_xmit - TCP transmit
1109  **/
1110 static int
1111 iscsi_xmit(struct iscsi_conn *conn)
1112 {
1113         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1114         struct iscsi_segment *segment = &tcp_conn->out.segment;
1115         unsigned int consumed = 0;
1116         int rc = 0;
1117
1118         while (1) {
1119                 rc = iscsi_tcp_xmit_segment(tcp_conn, segment);
1120                 if (rc < 0)
1121                         goto error;
1122                 if (rc == 0)
1123                         break;
1124
1125                 consumed += rc;
1126
1127                 if (segment->total_copied >= segment->total_size) {
1128                         if (segment->done != NULL) {
1129                                 rc = segment->done(tcp_conn, segment);
1130                                 if (rc < 0)
1131                                         goto error;
1132                         }
1133                 }
1134         }
1135
1136         debug_tcp("xmit %d bytes\n", consumed);
1137
1138         conn->txdata_octets += consumed;
1139         return consumed;
1140
1141 error:
1142         /* Transmit error. We could initiate error recovery
1143          * here. */
1144         debug_tcp("Error sending PDU, errno=%d\n", rc);
1145         iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
1146         return rc;
1147 }
1148
1149 /**
1150  * iscsi_tcp_xmit_qlen - return the number of bytes queued for xmit
1151  */
1152 static inline int
1153 iscsi_tcp_xmit_qlen(struct iscsi_conn *conn)
1154 {
1155         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1156         struct iscsi_segment *segment = &tcp_conn->out.segment;
1157
1158         return segment->total_copied - segment->total_size;
1159 }
1160
1161 static inline int
1162 iscsi_tcp_flush(struct iscsi_conn *conn)
1163 {
1164         int rc;
1165
1166         while (iscsi_tcp_xmit_qlen(conn)) {
1167                 rc = iscsi_xmit(conn);
1168                 if (rc == 0)
1169                         return -EAGAIN;
1170                 if (rc < 0)
1171                         return rc;
1172         }
1173
1174         return 0;
1175 }
1176
1177 /*
1178  * This is called when we're done sending the header.
1179  * Simply copy the data_segment to the send segment, and return.
1180  */
1181 static int
1182 iscsi_tcp_send_hdr_done(struct iscsi_tcp_conn *tcp_conn,
1183                         struct iscsi_segment *segment)
1184 {
1185         tcp_conn->out.segment = tcp_conn->out.data_segment;
1186         debug_tcp("Header done. Next segment size %u total_size %u\n",
1187                   tcp_conn->out.segment.size, tcp_conn->out.segment.total_size);
1188         return 0;
1189 }
1190
1191 static void
1192 iscsi_tcp_send_hdr_prep(struct iscsi_conn *conn, void *hdr, size_t hdrlen)
1193 {
1194         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1195
1196         debug_tcp("%s(%p%s)\n", __func__, tcp_conn,
1197                         conn->hdrdgst_en? ", digest enabled" : "");
1198
1199         /* Clear the data segment - needs to be filled in by the
1200          * caller using iscsi_tcp_send_data_prep() */
1201         memset(&tcp_conn->out.data_segment, 0, sizeof(struct iscsi_segment));
1202
1203         /* If header digest is enabled, compute the CRC and
1204          * place the digest into the same buffer. We make
1205          * sure that both iscsi_tcp_task and mtask have
1206          * sufficient room.
1207          */
1208         if (conn->hdrdgst_en) {
1209                 iscsi_tcp_dgst_header(&tcp_conn->tx_hash, hdr, hdrlen,
1210                                       hdr + hdrlen);
1211                 hdrlen += ISCSI_DIGEST_SIZE;
1212         }
1213
1214         /* Remember header pointer for later, when we need
1215          * to decide whether there's a payload to go along
1216          * with the header. */
1217         tcp_conn->out.hdr = hdr;
1218
1219         iscsi_segment_init_linear(&tcp_conn->out.segment, hdr, hdrlen,
1220                                 iscsi_tcp_send_hdr_done, NULL);
1221 }
1222
1223 /*
1224  * Prepare the send buffer for the payload data.
1225  * Padding and checksumming will all be taken care
1226  * of by the iscsi_segment routines.
1227  */
1228 static int
1229 iscsi_tcp_send_data_prep(struct iscsi_conn *conn, struct scatterlist *sg,
1230                          unsigned int count, unsigned int offset,
1231                          unsigned int len)
1232 {
1233         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1234         struct hash_desc *tx_hash = NULL;
1235         unsigned int hdr_spec_len;
1236
1237         debug_tcp("%s(%p, offset=%d, datalen=%d%s)\n", __func__,
1238                         tcp_conn, offset, len,
1239                         conn->datadgst_en? ", digest enabled" : "");
1240
1241         /* Make sure the datalen matches what the caller
1242            said he would send. */
1243         hdr_spec_len = ntoh24(tcp_conn->out.hdr->dlength);
1244         WARN_ON(iscsi_padded(len) != iscsi_padded(hdr_spec_len));
1245
1246         if (conn->datadgst_en)
1247                 tx_hash = &tcp_conn->tx_hash;
1248
1249         return iscsi_segment_seek_sg(&tcp_conn->out.data_segment,
1250                                    sg, count, offset, len,
1251                                    NULL, tx_hash);
1252 }
1253
1254 static void
1255 iscsi_tcp_send_linear_data_prepare(struct iscsi_conn *conn, void *data,
1256                                    size_t len)
1257 {
1258         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1259         struct hash_desc *tx_hash = NULL;
1260         unsigned int hdr_spec_len;
1261
1262         debug_tcp("%s(%p, datalen=%d%s)\n", __func__, tcp_conn, len,
1263                   conn->datadgst_en? ", digest enabled" : "");
1264
1265         /* Make sure the datalen matches what the caller
1266            said he would send. */
1267         hdr_spec_len = ntoh24(tcp_conn->out.hdr->dlength);
1268         WARN_ON(iscsi_padded(len) != iscsi_padded(hdr_spec_len));
1269
1270         if (conn->datadgst_en)
1271                 tx_hash = &tcp_conn->tx_hash;
1272
1273         iscsi_segment_init_linear(&tcp_conn->out.data_segment,
1274                                 data, len, NULL, tx_hash);
1275 }
1276
1277 /**
1278  * iscsi_solicit_data_cont - initialize next Data-Out
1279  * @conn: iscsi connection
1280  * @task: scsi command task
1281  * @r2t: R2T info
1282  * @left: bytes left to transfer
1283  *
1284  * Notes:
1285  *      Initialize next Data-Out within this R2T sequence and continue
1286  *      to process next Scatter-Gather element(if any) of this SCSI command.
1287  *
1288  *      Called under connection lock.
1289  **/
1290 static int
1291 iscsi_solicit_data_cont(struct iscsi_conn *conn, struct iscsi_task *task,
1292                         struct iscsi_r2t_info *r2t)
1293 {
1294         struct iscsi_data *hdr;
1295         int new_offset, left;
1296
1297         BUG_ON(r2t->data_length - r2t->sent < 0);
1298         left = r2t->data_length - r2t->sent;
1299         if (left == 0)
1300                 return 0;
1301
1302         hdr = &r2t->dtask.hdr;
1303         memset(hdr, 0, sizeof(struct iscsi_data));
1304         hdr->ttt = r2t->ttt;
1305         hdr->datasn = cpu_to_be32(r2t->solicit_datasn);
1306         r2t->solicit_datasn++;
1307         hdr->opcode = ISCSI_OP_SCSI_DATA_OUT;
1308         memcpy(hdr->lun, task->hdr->lun, sizeof(hdr->lun));
1309         hdr->itt = task->hdr->itt;
1310         hdr->exp_statsn = r2t->exp_statsn;
1311         new_offset = r2t->data_offset + r2t->sent;
1312         hdr->offset = cpu_to_be32(new_offset);
1313         if (left > conn->max_xmit_dlength) {
1314                 hton24(hdr->dlength, conn->max_xmit_dlength);
1315                 r2t->data_count = conn->max_xmit_dlength;
1316         } else {
1317                 hton24(hdr->dlength, left);
1318                 r2t->data_count = left;
1319                 hdr->flags = ISCSI_FLAG_CMD_FINAL;
1320         }
1321
1322         conn->dataout_pdus_cnt++;
1323         return 1;
1324 }
1325
1326 /**
1327  * iscsi_tcp_task - Initialize iSCSI SCSI_READ or SCSI_WRITE commands
1328  * @conn: iscsi connection
1329  * @task: scsi command task
1330  * @sc: scsi command
1331  **/
1332 static int
1333 iscsi_tcp_task_init(struct iscsi_task *task)
1334 {
1335         struct iscsi_tcp_task *tcp_task = task->dd_data;
1336         struct iscsi_conn *conn = task->conn;
1337         struct scsi_cmnd *sc = task->sc;
1338         int err;
1339
1340         if (!sc) {
1341                 /*
1342                  * mgmt tasks do not have a scatterlist since they come
1343                  * in from the iscsi interface.
1344                  */
1345                 debug_scsi("mtask deq [cid %d itt 0x%x]\n", conn->id,
1346                            task->itt);
1347
1348                 /* Prepare PDU, optionally w/ immediate data */
1349                 iscsi_tcp_send_hdr_prep(conn, task->hdr, sizeof(*task->hdr));
1350
1351                 /* If we have immediate data, attach a payload */
1352                 if (task->data_count)
1353                         iscsi_tcp_send_linear_data_prepare(conn, task->data,
1354                                                            task->data_count);
1355                 return 0;
1356         }
1357
1358         BUG_ON(__kfifo_len(tcp_task->r2tqueue));
1359         tcp_task->sent = 0;
1360         tcp_task->exp_datasn = 0;
1361
1362         /* Prepare PDU, optionally w/ immediate data */
1363         debug_scsi("task deq [cid %d itt 0x%x imm %d unsol %d]\n",
1364                     conn->id, task->itt, task->imm_count,
1365                     task->unsol_count);
1366         iscsi_tcp_send_hdr_prep(conn, task->hdr, task->hdr_len);
1367
1368         if (!task->imm_count)
1369                 return 0;
1370
1371         /* If we have immediate data, attach a payload */
1372         err = iscsi_tcp_send_data_prep(conn, scsi_out(sc)->table.sgl,
1373                                        scsi_out(sc)->table.nents,
1374                                        0, task->imm_count);
1375         if (err)
1376                 return err;
1377         tcp_task->sent += task->imm_count;
1378         task->imm_count = 0;
1379         return 0;
1380 }
1381
1382 /*
1383  * iscsi_tcp_task_xmit - xmit normal PDU task
1384  * @task: iscsi command task
1385  *
1386  * We're expected to return 0 when everything was transmitted succesfully,
1387  * -EAGAIN if there's still data in the queue, or != 0 for any other kind
1388  * of error.
1389  */
1390 static int
1391 iscsi_tcp_task_xmit(struct iscsi_task *task)
1392 {
1393         struct iscsi_conn *conn = task->conn;
1394         struct iscsi_tcp_task *tcp_task = task->dd_data;
1395         struct scsi_cmnd *sc = task->sc;
1396         struct scsi_data_buffer *sdb;
1397         int rc = 0;
1398
1399 flush:
1400         /* Flush any pending data first. */
1401         rc = iscsi_tcp_flush(conn);
1402         if (rc < 0)
1403                 return rc;
1404
1405         /* mgmt command */
1406         if (!sc) {
1407                 if (task->hdr->itt == RESERVED_ITT)
1408                         iscsi_put_task(task);
1409                 return 0;
1410         }
1411
1412         /* Are we done already? */
1413         if (sc->sc_data_direction != DMA_TO_DEVICE)
1414                 return 0;
1415
1416         sdb = scsi_out(sc);
1417         if (task->unsol_count != 0) {
1418                 struct iscsi_data *hdr = &tcp_task->unsol_dtask.hdr;
1419
1420                 /* Prepare a header for the unsolicited PDU.
1421                  * The amount of data we want to send will be
1422                  * in task->data_count.
1423                  * FIXME: return the data count instead.
1424                  */
1425                 iscsi_prep_unsolicit_data_pdu(task, hdr);
1426
1427                 debug_tcp("unsol dout [itt 0x%x doff %d dlen %d]\n",
1428                                 task->itt, tcp_task->sent, task->data_count);
1429
1430                 iscsi_tcp_send_hdr_prep(conn, hdr, sizeof(*hdr));
1431                 rc = iscsi_tcp_send_data_prep(conn, sdb->table.sgl,
1432                                               sdb->table.nents, tcp_task->sent,
1433                                               task->data_count);
1434                 if (rc)
1435                         goto fail;
1436                 tcp_task->sent += task->data_count;
1437                 task->unsol_count -= task->data_count;
1438                 goto flush;
1439         } else {
1440                 struct iscsi_session *session = conn->session;
1441                 struct iscsi_r2t_info *r2t;
1442
1443                 /* All unsolicited PDUs sent. Check for solicited PDUs.
1444                  */
1445                 spin_lock_bh(&session->lock);
1446                 r2t = tcp_task->r2t;
1447                 if (r2t != NULL) {
1448                         /* Continue with this R2T? */
1449                         if (!iscsi_solicit_data_cont(conn, task, r2t)) {
1450                                 debug_scsi("  done with r2t %p\n", r2t);
1451
1452                                 __kfifo_put(tcp_task->r2tpool.queue,
1453                                             (void*)&r2t, sizeof(void*));
1454                                 tcp_task->r2t = r2t = NULL;
1455                         }
1456                 }
1457
1458                 if (r2t == NULL) {
1459                         __kfifo_get(tcp_task->r2tqueue, (void*)&tcp_task->r2t,
1460                                     sizeof(void*));
1461                         r2t = tcp_task->r2t;
1462                 }
1463                 spin_unlock_bh(&session->lock);
1464
1465                 /* Waiting for more R2Ts to arrive. */
1466                 if (r2t == NULL) {
1467                         debug_tcp("no R2Ts yet\n");
1468                         return 0;
1469                 }
1470
1471                 debug_scsi("sol dout %p [dsn %d itt 0x%x doff %d dlen %d]\n",
1472                         r2t, r2t->solicit_datasn - 1, task->itt,
1473                         r2t->data_offset + r2t->sent, r2t->data_count);
1474
1475                 iscsi_tcp_send_hdr_prep(conn, &r2t->dtask.hdr,
1476                                         sizeof(struct iscsi_hdr));
1477
1478                 rc = iscsi_tcp_send_data_prep(conn, sdb->table.sgl,
1479                                               sdb->table.nents,
1480                                               r2t->data_offset + r2t->sent,
1481                                               r2t->data_count);
1482                 if (rc)
1483                         goto fail;
1484                 tcp_task->sent += r2t->data_count;
1485                 r2t->sent += r2t->data_count;
1486                 goto flush;
1487         }
1488         return 0;
1489 fail:
1490         iscsi_conn_failure(conn, rc);
1491         return -EIO;
1492 }
1493
1494 static struct iscsi_cls_conn *
1495 iscsi_tcp_conn_create(struct iscsi_cls_session *cls_session, uint32_t conn_idx)
1496 {
1497         struct iscsi_conn *conn;
1498         struct iscsi_cls_conn *cls_conn;
1499         struct iscsi_tcp_conn *tcp_conn;
1500
1501         cls_conn = iscsi_conn_setup(cls_session, sizeof(*tcp_conn), conn_idx);
1502         if (!cls_conn)
1503                 return NULL;
1504         conn = cls_conn->dd_data;
1505         /*
1506          * due to strange issues with iser these are not set
1507          * in iscsi_conn_setup
1508          */
1509         conn->max_recv_dlength = ISCSI_DEF_MAX_RECV_SEG_LEN;
1510
1511         tcp_conn = conn->dd_data;
1512         tcp_conn->iscsi_conn = conn;
1513
1514         tcp_conn->tx_hash.tfm = crypto_alloc_hash("crc32c", 0,
1515                                                   CRYPTO_ALG_ASYNC);
1516         tcp_conn->tx_hash.flags = 0;
1517         if (IS_ERR(tcp_conn->tx_hash.tfm))
1518                 goto free_conn;
1519
1520         tcp_conn->rx_hash.tfm = crypto_alloc_hash("crc32c", 0,
1521                                                   CRYPTO_ALG_ASYNC);
1522         tcp_conn->rx_hash.flags = 0;
1523         if (IS_ERR(tcp_conn->rx_hash.tfm))
1524                 goto free_tx_tfm;
1525
1526         return cls_conn;
1527
1528 free_tx_tfm:
1529         crypto_free_hash(tcp_conn->tx_hash.tfm);
1530 free_conn:
1531         iscsi_conn_printk(KERN_ERR, conn,
1532                           "Could not create connection due to crc32c "
1533                           "loading error. Make sure the crc32c "
1534                           "module is built as a module or into the "
1535                           "kernel\n");
1536         iscsi_conn_teardown(cls_conn);
1537         return NULL;
1538 }
1539
1540 static void
1541 iscsi_tcp_release_conn(struct iscsi_conn *conn)
1542 {
1543         struct iscsi_session *session = conn->session;
1544         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1545         struct socket *sock = tcp_conn->sock;
1546
1547         if (!sock)
1548                 return;
1549
1550         sock_hold(sock->sk);
1551         iscsi_conn_restore_callbacks(tcp_conn);
1552         sock_put(sock->sk);
1553
1554         spin_lock_bh(&session->lock);
1555         tcp_conn->sock = NULL;
1556         spin_unlock_bh(&session->lock);
1557         sockfd_put(sock);
1558 }
1559
1560 static void
1561 iscsi_tcp_conn_destroy(struct iscsi_cls_conn *cls_conn)
1562 {
1563         struct iscsi_conn *conn = cls_conn->dd_data;
1564         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1565
1566         iscsi_tcp_release_conn(conn);
1567
1568         if (tcp_conn->tx_hash.tfm)
1569                 crypto_free_hash(tcp_conn->tx_hash.tfm);
1570         if (tcp_conn->rx_hash.tfm)
1571                 crypto_free_hash(tcp_conn->rx_hash.tfm);
1572
1573         iscsi_conn_teardown(cls_conn);
1574 }
1575
1576 static void
1577 iscsi_tcp_conn_stop(struct iscsi_cls_conn *cls_conn, int flag)
1578 {
1579         struct iscsi_conn *conn = cls_conn->dd_data;
1580         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1581
1582         /* userspace may have goofed up and not bound us */
1583         if (!tcp_conn->sock)
1584                 return;
1585         /*
1586          * Make sure our recv side is stopped.
1587          * Older tools called conn stop before ep_disconnect
1588          * so IO could still be coming in.
1589          */
1590         write_lock_bh(&tcp_conn->sock->sk->sk_callback_lock);
1591         set_bit(ISCSI_SUSPEND_BIT, &conn->suspend_rx);
1592         write_unlock_bh(&tcp_conn->sock->sk->sk_callback_lock);
1593
1594         iscsi_conn_stop(cls_conn, flag);
1595         iscsi_tcp_release_conn(conn);
1596 }
1597
1598 static int iscsi_tcp_get_addr(struct iscsi_conn *conn, struct socket *sock,
1599                               char *buf, int *port,
1600                               int (*getname)(struct socket *, struct sockaddr *,
1601                                         int *addrlen))
1602 {
1603         struct sockaddr_storage *addr;
1604         struct sockaddr_in6 *sin6;
1605         struct sockaddr_in *sin;
1606         int rc = 0, len;
1607
1608         addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1609         if (!addr)
1610                 return -ENOMEM;
1611
1612         if (getname(sock, (struct sockaddr *) addr, &len)) {
1613                 rc = -ENODEV;
1614                 goto free_addr;
1615         }
1616
1617         switch (addr->ss_family) {
1618         case AF_INET:
1619                 sin = (struct sockaddr_in *)addr;
1620                 spin_lock_bh(&conn->session->lock);
1621                 sprintf(buf, NIPQUAD_FMT, NIPQUAD(sin->sin_addr.s_addr));
1622                 *port = be16_to_cpu(sin->sin_port);
1623                 spin_unlock_bh(&conn->session->lock);
1624                 break;
1625         case AF_INET6:
1626                 sin6 = (struct sockaddr_in6 *)addr;
1627                 spin_lock_bh(&conn->session->lock);
1628                 sprintf(buf, NIP6_FMT, NIP6(sin6->sin6_addr));
1629                 *port = be16_to_cpu(sin6->sin6_port);
1630                 spin_unlock_bh(&conn->session->lock);
1631                 break;
1632         }
1633 free_addr:
1634         kfree(addr);
1635         return rc;
1636 }
1637
1638 static int
1639 iscsi_tcp_conn_bind(struct iscsi_cls_session *cls_session,
1640                     struct iscsi_cls_conn *cls_conn, uint64_t transport_eph,
1641                     int is_leading)
1642 {
1643         struct Scsi_Host *shost = iscsi_session_to_shost(cls_session);
1644         struct iscsi_host *ihost = shost_priv(shost);
1645         struct iscsi_conn *conn = cls_conn->dd_data;
1646         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1647         struct sock *sk;
1648         struct socket *sock;
1649         int err;
1650
1651         /* lookup for existing socket */
1652         sock = sockfd_lookup((int)transport_eph, &err);
1653         if (!sock) {
1654                 iscsi_conn_printk(KERN_ERR, conn,
1655                                   "sockfd_lookup failed %d\n", err);
1656                 return -EEXIST;
1657         }
1658         /*
1659          * copy these values now because if we drop the session
1660          * userspace may still want to query the values since we will
1661          * be using them for the reconnect
1662          */
1663         err = iscsi_tcp_get_addr(conn, sock, conn->portal_address,
1664                                  &conn->portal_port, kernel_getpeername);
1665         if (err)
1666                 goto free_socket;
1667
1668         err = iscsi_tcp_get_addr(conn, sock, ihost->local_address,
1669                                 &ihost->local_port, kernel_getsockname);
1670         if (err)
1671                 goto free_socket;
1672
1673         err = iscsi_conn_bind(cls_session, cls_conn, is_leading);
1674         if (err)
1675                 goto free_socket;
1676
1677         /* bind iSCSI connection and socket */
1678         tcp_conn->sock = sock;
1679
1680         /* setup Socket parameters */
1681         sk = sock->sk;
1682         sk->sk_reuse = 1;
1683         sk->sk_sndtimeo = 15 * HZ; /* FIXME: make it configurable */
1684         sk->sk_allocation = GFP_ATOMIC;
1685
1686         iscsi_conn_set_callbacks(conn);
1687         tcp_conn->sendpage = tcp_conn->sock->ops->sendpage;
1688         /*
1689          * set receive state machine into initial state
1690          */
1691         iscsi_tcp_hdr_recv_prep(tcp_conn);
1692         return 0;
1693
1694 free_socket:
1695         sockfd_put(sock);
1696         return err;
1697 }
1698
1699 static int
1700 iscsi_r2tpool_alloc(struct iscsi_session *session)
1701 {
1702         int i;
1703         int cmd_i;
1704
1705         /*
1706          * initialize per-task: R2T pool and xmit queue
1707          */
1708         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
1709                 struct iscsi_task *task = session->cmds[cmd_i];
1710                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1711
1712                 /*
1713                  * pre-allocated x4 as much r2ts to handle race when
1714                  * target acks DataOut faster than we data_xmit() queues
1715                  * could replenish r2tqueue.
1716                  */
1717
1718                 /* R2T pool */
1719                 if (iscsi_pool_init(&tcp_task->r2tpool, session->max_r2t * 4, NULL,
1720                                     sizeof(struct iscsi_r2t_info))) {
1721                         goto r2t_alloc_fail;
1722                 }
1723
1724                 /* R2T xmit queue */
1725                 tcp_task->r2tqueue = kfifo_alloc(
1726                       session->max_r2t * 4 * sizeof(void*), GFP_KERNEL, NULL);
1727                 if (tcp_task->r2tqueue == ERR_PTR(-ENOMEM)) {
1728                         iscsi_pool_free(&tcp_task->r2tpool);
1729                         goto r2t_alloc_fail;
1730                 }
1731         }
1732
1733         return 0;
1734
1735 r2t_alloc_fail:
1736         for (i = 0; i < cmd_i; i++) {
1737                 struct iscsi_task *task = session->cmds[i];
1738                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1739
1740                 kfifo_free(tcp_task->r2tqueue);
1741                 iscsi_pool_free(&tcp_task->r2tpool);
1742         }
1743         return -ENOMEM;
1744 }
1745
1746 static void
1747 iscsi_r2tpool_free(struct iscsi_session *session)
1748 {
1749         int i;
1750
1751         for (i = 0; i < session->cmds_max; i++) {
1752                 struct iscsi_task *task = session->cmds[i];
1753                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1754
1755                 kfifo_free(tcp_task->r2tqueue);
1756                 iscsi_pool_free(&tcp_task->r2tpool);
1757         }
1758 }
1759
1760 static int
1761 iscsi_conn_set_param(struct iscsi_cls_conn *cls_conn, enum iscsi_param param,
1762                      char *buf, int buflen)
1763 {
1764         struct iscsi_conn *conn = cls_conn->dd_data;
1765         struct iscsi_session *session = conn->session;
1766         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1767         int value;
1768
1769         switch(param) {
1770         case ISCSI_PARAM_HDRDGST_EN:
1771                 iscsi_set_param(cls_conn, param, buf, buflen);
1772                 break;
1773         case ISCSI_PARAM_DATADGST_EN:
1774                 iscsi_set_param(cls_conn, param, buf, buflen);
1775                 tcp_conn->sendpage = conn->datadgst_en ?
1776                         sock_no_sendpage : tcp_conn->sock->ops->sendpage;
1777                 break;
1778         case ISCSI_PARAM_MAX_R2T:
1779                 sscanf(buf, "%d", &value);
1780                 if (value <= 0 || !is_power_of_2(value))
1781                         return -EINVAL;
1782                 if (session->max_r2t == value)
1783                         break;
1784                 iscsi_r2tpool_free(session);
1785                 iscsi_set_param(cls_conn, param, buf, buflen);
1786                 if (iscsi_r2tpool_alloc(session))
1787                         return -ENOMEM;
1788                 break;
1789         default:
1790                 return iscsi_set_param(cls_conn, param, buf, buflen);
1791         }
1792
1793         return 0;
1794 }
1795
1796 static int
1797 iscsi_tcp_conn_get_param(struct iscsi_cls_conn *cls_conn,
1798                          enum iscsi_param param, char *buf)
1799 {
1800         struct iscsi_conn *conn = cls_conn->dd_data;
1801         int len;
1802
1803         switch(param) {
1804         case ISCSI_PARAM_CONN_PORT:
1805                 spin_lock_bh(&conn->session->lock);
1806                 len = sprintf(buf, "%hu\n", conn->portal_port);
1807                 spin_unlock_bh(&conn->session->lock);
1808                 break;
1809         case ISCSI_PARAM_CONN_ADDRESS:
1810                 spin_lock_bh(&conn->session->lock);
1811                 len = sprintf(buf, "%s\n", conn->portal_address);
1812                 spin_unlock_bh(&conn->session->lock);
1813                 break;
1814         default:
1815                 return iscsi_conn_get_param(cls_conn, param, buf);
1816         }
1817
1818         return len;
1819 }
1820
1821 static void
1822 iscsi_conn_get_stats(struct iscsi_cls_conn *cls_conn, struct iscsi_stats *stats)
1823 {
1824         struct iscsi_conn *conn = cls_conn->dd_data;
1825         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1826
1827         stats->txdata_octets = conn->txdata_octets;
1828         stats->rxdata_octets = conn->rxdata_octets;
1829         stats->scsicmd_pdus = conn->scsicmd_pdus_cnt;
1830         stats->dataout_pdus = conn->dataout_pdus_cnt;
1831         stats->scsirsp_pdus = conn->scsirsp_pdus_cnt;
1832         stats->datain_pdus = conn->datain_pdus_cnt;
1833         stats->r2t_pdus = conn->r2t_pdus_cnt;
1834         stats->tmfcmd_pdus = conn->tmfcmd_pdus_cnt;
1835         stats->tmfrsp_pdus = conn->tmfrsp_pdus_cnt;
1836         stats->custom_length = 3;
1837         strcpy(stats->custom[0].desc, "tx_sendpage_failures");
1838         stats->custom[0].value = tcp_conn->sendpage_failures_cnt;
1839         strcpy(stats->custom[1].desc, "rx_discontiguous_hdr");
1840         stats->custom[1].value = tcp_conn->discontiguous_hdr_cnt;
1841         strcpy(stats->custom[2].desc, "eh_abort_cnt");
1842         stats->custom[2].value = conn->eh_abort_cnt;
1843 }
1844
1845 static struct iscsi_cls_session *
1846 iscsi_tcp_session_create(struct iscsi_endpoint *ep, uint16_t cmds_max,
1847                          uint16_t qdepth, uint32_t initial_cmdsn,
1848                          uint32_t *hostno)
1849 {
1850         struct iscsi_cls_session *cls_session;
1851         struct iscsi_session *session;
1852         struct Scsi_Host *shost;
1853         int cmd_i;
1854
1855         if (ep) {
1856                 printk(KERN_ERR "iscsi_tcp: invalid ep %p.\n", ep);
1857                 return NULL;
1858         }
1859
1860         shost = iscsi_host_alloc(&iscsi_sht, 0, qdepth);
1861         if (!shost)
1862                 return NULL;
1863         shost->transportt = iscsi_tcp_scsi_transport;
1864         shost->max_lun = iscsi_max_lun;
1865         shost->max_id = 0;
1866         shost->max_channel = 0;
1867         shost->max_cmd_len = SCSI_MAX_VARLEN_CDB_SIZE;
1868
1869         if (iscsi_host_add(shost, NULL))
1870                 goto free_host;
1871         *hostno = shost->host_no;
1872
1873         cls_session = iscsi_session_setup(&iscsi_tcp_transport, shost, cmds_max,
1874                                           sizeof(struct iscsi_tcp_task),
1875                                           initial_cmdsn, 0);
1876         if (!cls_session)
1877                 goto remove_host;
1878         session = cls_session->dd_data;
1879
1880         shost->can_queue = session->scsi_cmds_max;
1881         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
1882                 struct iscsi_task *task = session->cmds[cmd_i];
1883                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1884
1885                 task->hdr = &tcp_task->hdr.cmd_hdr;
1886                 task->hdr_max = sizeof(tcp_task->hdr) - ISCSI_DIGEST_SIZE;
1887         }
1888
1889         if (iscsi_r2tpool_alloc(session))
1890                 goto remove_session;
1891         return cls_session;
1892
1893 remove_session:
1894         iscsi_session_teardown(cls_session);
1895 remove_host:
1896         iscsi_host_remove(shost);
1897 free_host:
1898         iscsi_host_free(shost);
1899         return NULL;
1900 }
1901
1902 static void iscsi_tcp_session_destroy(struct iscsi_cls_session *cls_session)
1903 {
1904         struct Scsi_Host *shost = iscsi_session_to_shost(cls_session);
1905
1906         iscsi_r2tpool_free(cls_session->dd_data);
1907
1908         iscsi_host_remove(shost);
1909         iscsi_host_free(shost);
1910 }
1911
1912 static int iscsi_tcp_slave_configure(struct scsi_device *sdev)
1913 {
1914         blk_queue_bounce_limit(sdev->request_queue, BLK_BOUNCE_ANY);
1915         blk_queue_dma_alignment(sdev->request_queue, 0);
1916         return 0;
1917 }
1918
1919 static struct scsi_host_template iscsi_sht = {
1920         .module                 = THIS_MODULE,
1921         .name                   = "iSCSI Initiator over TCP/IP",
1922         .queuecommand           = iscsi_queuecommand,
1923         .change_queue_depth     = iscsi_change_queue_depth,
1924         .can_queue              = ISCSI_DEF_XMIT_CMDS_MAX - 1,
1925         .sg_tablesize           = 4096,
1926         .max_sectors            = 0xFFFF,
1927         .cmd_per_lun            = ISCSI_DEF_CMD_PER_LUN,
1928         .eh_abort_handler       = iscsi_eh_abort,
1929         .eh_device_reset_handler= iscsi_eh_device_reset,
1930         .eh_host_reset_handler  = iscsi_eh_host_reset,
1931         .use_clustering         = DISABLE_CLUSTERING,
1932         .slave_configure        = iscsi_tcp_slave_configure,
1933         .proc_name              = "iscsi_tcp",
1934         .this_id                = -1,
1935 };
1936
1937 static struct iscsi_transport iscsi_tcp_transport = {
1938         .owner                  = THIS_MODULE,
1939         .name                   = "tcp",
1940         .caps                   = CAP_RECOVERY_L0 | CAP_MULTI_R2T | CAP_HDRDGST
1941                                   | CAP_DATADGST,
1942         .param_mask             = ISCSI_MAX_RECV_DLENGTH |
1943                                   ISCSI_MAX_XMIT_DLENGTH |
1944                                   ISCSI_HDRDGST_EN |
1945                                   ISCSI_DATADGST_EN |
1946                                   ISCSI_INITIAL_R2T_EN |
1947                                   ISCSI_MAX_R2T |
1948                                   ISCSI_IMM_DATA_EN |
1949                                   ISCSI_FIRST_BURST |
1950                                   ISCSI_MAX_BURST |
1951                                   ISCSI_PDU_INORDER_EN |
1952                                   ISCSI_DATASEQ_INORDER_EN |
1953                                   ISCSI_ERL |
1954                                   ISCSI_CONN_PORT |
1955                                   ISCSI_CONN_ADDRESS |
1956                                   ISCSI_EXP_STATSN |
1957                                   ISCSI_PERSISTENT_PORT |
1958                                   ISCSI_PERSISTENT_ADDRESS |
1959                                   ISCSI_TARGET_NAME | ISCSI_TPGT |
1960                                   ISCSI_USERNAME | ISCSI_PASSWORD |
1961                                   ISCSI_USERNAME_IN | ISCSI_PASSWORD_IN |
1962                                   ISCSI_FAST_ABORT | ISCSI_ABORT_TMO |
1963                                   ISCSI_LU_RESET_TMO |
1964                                   ISCSI_PING_TMO | ISCSI_RECV_TMO |
1965                                   ISCSI_IFACE_NAME | ISCSI_INITIATOR_NAME,
1966         .host_param_mask        = ISCSI_HOST_HWADDRESS | ISCSI_HOST_IPADDRESS |
1967                                   ISCSI_HOST_INITIATOR_NAME |
1968                                   ISCSI_HOST_NETDEV_NAME,
1969         /* session management */
1970         .create_session         = iscsi_tcp_session_create,
1971         .destroy_session        = iscsi_tcp_session_destroy,
1972         /* connection management */
1973         .create_conn            = iscsi_tcp_conn_create,
1974         .bind_conn              = iscsi_tcp_conn_bind,
1975         .destroy_conn           = iscsi_tcp_conn_destroy,
1976         .set_param              = iscsi_conn_set_param,
1977         .get_conn_param         = iscsi_tcp_conn_get_param,
1978         .get_session_param      = iscsi_session_get_param,
1979         .start_conn             = iscsi_conn_start,
1980         .stop_conn              = iscsi_tcp_conn_stop,
1981         /* iscsi host params */
1982         .get_host_param         = iscsi_host_get_param,
1983         .set_host_param         = iscsi_host_set_param,
1984         /* IO */
1985         .send_pdu               = iscsi_conn_send_pdu,
1986         .get_stats              = iscsi_conn_get_stats,
1987         .init_task              = iscsi_tcp_task_init,
1988         .xmit_task              = iscsi_tcp_task_xmit,
1989         .cleanup_task           = iscsi_tcp_cleanup_task,
1990         /* recovery */
1991         .session_recovery_timedout = iscsi_session_recovery_timedout,
1992 };
1993
1994 static int __init
1995 iscsi_tcp_init(void)
1996 {
1997         if (iscsi_max_lun < 1) {
1998                 printk(KERN_ERR "iscsi_tcp: Invalid max_lun value of %u\n",
1999                        iscsi_max_lun);
2000                 return -EINVAL;
2001         }
2002
2003         iscsi_tcp_scsi_transport = iscsi_register_transport(
2004                                                         &iscsi_tcp_transport);
2005         if (!iscsi_tcp_scsi_transport)
2006                 return -ENODEV;
2007
2008         return 0;
2009 }
2010
2011 static void __exit
2012 iscsi_tcp_exit(void)
2013 {
2014         iscsi_unregister_transport(&iscsi_tcp_transport);
2015 }
2016
2017 module_init(iscsi_tcp_init);
2018 module_exit(iscsi_tcp_exit);