SAFE public projects git trees. - safe/jmp/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_ctask(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
 502 {
 503         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 504         struct iscsi_r2t_info *r2t;
 505
 506         /* flush ctask's r2t queues */
 507         while (__kfifo_get(tcp_ctask->r2tqueue, (void*)&r2t, sizeof(void*))) {
 508                 __kfifo_put(tcp_ctask->r2tpool.queue, (void*)&r2t,
 509                             sizeof(void*));
 510                 debug_scsi("iscsi_tcp_cleanup_ctask pending r2t dropped\n");
 511         }
 512
 513         r2t = tcp_ctask->r2t;
 514         if (r2t != NULL) {
 515                 __kfifo_put(tcp_ctask->r2tpool.queue, (void*)&r2t,
 516                             sizeof(void*));
 517                 tcp_ctask->r2t = NULL;
 518         }
 519 }
 520
 521 /**
 522  * iscsi_data_rsp - SCSI Data-In Response processing
 523  * @conn: iscsi connection
 524  * @ctask: scsi command task
 525  **/
 526 static int
 527 iscsi_data_rsp(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
 528 {
 529         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 530         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 531         struct iscsi_data_rsp *rhdr = (struct iscsi_data_rsp *)tcp_conn->in.hdr;
 532         struct iscsi_session *session = conn->session;
 533         struct scsi_cmnd *sc = ctask->sc;
 534         int datasn = be32_to_cpu(rhdr->datasn);
 535         unsigned total_in_length = scsi_in(sc)->length;
 536
 537         iscsi_update_cmdsn(session, (struct iscsi_nopin*)rhdr);
 538         if (tcp_conn->in.datalen == 0)
 539                 return 0;
 540
 541         if (tcp_ctask->exp_datasn != datasn) {
 542                 debug_tcp("%s: ctask->exp_datasn(%d) != rhdr->datasn(%d)\n",
 543                           __FUNCTION__, tcp_ctask->exp_datasn, datasn);
 544                 return ISCSI_ERR_DATASN;
 545         }
 546
 547         tcp_ctask->exp_datasn++;
 548
 549         tcp_ctask->data_offset = be32_to_cpu(rhdr->offset);
 550         if (tcp_ctask->data_offset + tcp_conn->in.datalen > total_in_length) {
 551                 debug_tcp("%s: data_offset(%d) + data_len(%d) > total_length_in(%d)\n",
 552                           __FUNCTION__, tcp_ctask->data_offset,
 553                           tcp_conn->in.datalen, total_in_length);
 554                 return ISCSI_ERR_DATA_OFFSET;
 555         }
 556
 557         if (rhdr->flags & ISCSI_FLAG_DATA_STATUS) {
 558                 sc->result = (DID_OK << 16) | rhdr->cmd_status;
 559                 conn->exp_statsn = be32_to_cpu(rhdr->statsn) + 1;
 560                 if (rhdr->flags & (ISCSI_FLAG_DATA_UNDERFLOW |
 561                                    ISCSI_FLAG_DATA_OVERFLOW)) {
 562                         int res_count = be32_to_cpu(rhdr->residual_count);
 563
 564                         if (res_count > 0 &&
 565                             (rhdr->flags & ISCSI_FLAG_CMD_OVERFLOW ||
 566                              res_count <= total_in_length))
 567                                 scsi_in(sc)->resid = res_count;
 568                         else
 569                                 sc->result = (DID_BAD_TARGET << 16) |
 570                                         rhdr->cmd_status;
 571                 }
 572         }
 573
 574         conn->datain_pdus_cnt++;
 575         return 0;
 576 }
 577
 578 /**
 579  * iscsi_solicit_data_init - initialize first Data-Out
 580  * @conn: iscsi connection
 581  * @ctask: scsi command task
 582  * @r2t: R2T info
 583  *
 584  * Notes:
 585  *      Initialize first Data-Out within this R2T sequence and finds
 586  *      proper data_offset within this SCSI command.
 587  *
 588  *      This function is called with connection lock taken.
 589  **/
 590 static void
 591 iscsi_solicit_data_init(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask,
 592                         struct iscsi_r2t_info *r2t)
 593 {
 594         struct iscsi_data *hdr;
 595
 596         hdr = &r2t->dtask.hdr;
 597         memset(hdr, 0, sizeof(struct iscsi_data));
 598         hdr->ttt = r2t->ttt;
 599         hdr->datasn = cpu_to_be32(r2t->solicit_datasn);
 600         r2t->solicit_datasn++;
 601         hdr->opcode = ISCSI_OP_SCSI_DATA_OUT;
 602         memcpy(hdr->lun, ctask->hdr->lun, sizeof(hdr->lun));
 603         hdr->itt = ctask->hdr->itt;
 604         hdr->exp_statsn = r2t->exp_statsn;
 605         hdr->offset = cpu_to_be32(r2t->data_offset);
 606         if (r2t->data_length > conn->max_xmit_dlength) {
 607                 hton24(hdr->dlength, conn->max_xmit_dlength);
 608                 r2t->data_count = conn->max_xmit_dlength;
 609                 hdr->flags = 0;
 610         } else {
 611                 hton24(hdr->dlength, r2t->data_length);
 612                 r2t->data_count = r2t->data_length;
 613                 hdr->flags = ISCSI_FLAG_CMD_FINAL;
 614         }
 615         conn->dataout_pdus_cnt++;
 616
 617         r2t->sent = 0;
 618 }
 619
 620 /**
 621  * iscsi_r2t_rsp - iSCSI R2T Response processing
 622  * @conn: iscsi connection
 623  * @ctask: scsi command task
 624  **/
 625 static int
 626 iscsi_r2t_rsp(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
 627 {
 628         struct iscsi_r2t_info *r2t;
 629         struct iscsi_session *session = conn->session;
 630         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 631         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 632         struct iscsi_r2t_rsp *rhdr = (struct iscsi_r2t_rsp *)tcp_conn->in.hdr;
 633         int r2tsn = be32_to_cpu(rhdr->r2tsn);
 634         int rc;
 635
 636         if (tcp_conn->in.datalen) {
 637                 iscsi_conn_printk(KERN_ERR, conn,
 638                                   "invalid R2t with datalen %d\n",
 639                                   tcp_conn->in.datalen);
 640                 return ISCSI_ERR_DATALEN;
 641         }
 642
 643         if (tcp_ctask->exp_datasn != r2tsn){
 644                 debug_tcp("%s: ctask->exp_datasn(%d) != rhdr->r2tsn(%d)\n",
 645                           __FUNCTION__, tcp_ctask->exp_datasn, r2tsn);
 646                 return ISCSI_ERR_R2TSN;
 647         }
 648
 649         /* fill-in new R2T associated with the task */
 650         iscsi_update_cmdsn(session, (struct iscsi_nopin*)rhdr);
 651
 652         if (!ctask->sc || session->state != ISCSI_STATE_LOGGED_IN) {
 653                 iscsi_conn_printk(KERN_INFO, conn,
 654                                   "dropping R2T itt %d in recovery.\n",
 655                                   ctask->itt);
 656                 return 0;
 657         }
 658
 659         rc = __kfifo_get(tcp_ctask->r2tpool.queue, (void*)&r2t, sizeof(void*));
 660         BUG_ON(!rc);
 661
 662         r2t->exp_statsn = rhdr->statsn;
 663         r2t->data_length = be32_to_cpu(rhdr->data_length);
 664         if (r2t->data_length == 0) {
 665                 iscsi_conn_printk(KERN_ERR, conn,
 666                                   "invalid R2T with zero data len\n");
 667                 __kfifo_put(tcp_ctask->r2tpool.queue, (void*)&r2t,
 668                             sizeof(void*));
 669                 return ISCSI_ERR_DATALEN;
 670         }
 671
 672         if (r2t->data_length > session->max_burst)
 673                 debug_scsi("invalid R2T with data len %u and max burst %u."
 674                            "Attempting to execute request.\n",
 675                             r2t->data_length, session->max_burst);
 676
 677         r2t->data_offset = be32_to_cpu(rhdr->data_offset);
 678         if (r2t->data_offset + r2t->data_length > scsi_out(ctask->sc)->length) {
 679                 iscsi_conn_printk(KERN_ERR, conn,
 680                                   "invalid R2T with data len %u at offset %u "
 681                                   "and total length %d\n", r2t->data_length,
 682                                   r2t->data_offset, scsi_out(ctask->sc)->length);
 683                 __kfifo_put(tcp_ctask->r2tpool.queue, (void*)&r2t,
 684                             sizeof(void*));
 685                 return ISCSI_ERR_DATALEN;
 686         }
 687
 688         r2t->ttt = rhdr->ttt; /* no flip */
 689         r2t->solicit_datasn = 0;
 690
 691         iscsi_solicit_data_init(conn, ctask, r2t);
 692
 693         tcp_ctask->exp_datasn = r2tsn + 1;
 694         __kfifo_put(tcp_ctask->r2tqueue, (void*)&r2t, sizeof(void*));
 695         conn->r2t_pdus_cnt++;
 696
 697         iscsi_requeue_ctask(ctask);
 698         return 0;
 699 }
 700
 701 /*
 702  * Handle incoming reply to DataIn command
 703  */
 704 static int
 705 iscsi_tcp_process_data_in(struct iscsi_tcp_conn *tcp_conn,
 706                           struct iscsi_segment *segment)
 707 {
 708         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 709         struct iscsi_hdr *hdr = tcp_conn->in.hdr;
 710         int rc;
 711
 712         if (!iscsi_tcp_dgst_verify(tcp_conn, segment))
 713                 return ISCSI_ERR_DATA_DGST;
 714
 715         /* check for non-exceptional status */
 716         if (hdr->flags & ISCSI_FLAG_DATA_STATUS) {
 717                 rc = iscsi_complete_pdu(conn, tcp_conn->in.hdr, NULL, 0);
 718                 if (rc)
 719                         return rc;
 720         }
 721
 722         iscsi_tcp_hdr_recv_prep(tcp_conn);
 723         return 0;
 724 }
 725
 726 /**
 727  * iscsi_tcp_hdr_dissect - process PDU header
 728  * @conn: iSCSI connection
 729  * @hdr: PDU header
 730  *
 731  * This function analyzes the header of the PDU received,
 732  * and performs several sanity checks. If the PDU is accompanied
 733  * by data, the receive buffer is set up to copy the incoming data
 734  * to the correct location.
 735  */
 736 static int
 737 iscsi_tcp_hdr_dissect(struct iscsi_conn *conn, struct iscsi_hdr *hdr)
 738 {
 739         int rc = 0, opcode, ahslen;
 740         struct iscsi_session *session = conn->session;
 741         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 742         struct iscsi_cmd_task *ctask;
 743         uint32_t itt;
 744
 745         /* verify PDU length */
 746         tcp_conn->in.datalen = ntoh24(hdr->dlength);
 747         if (tcp_conn->in.datalen > conn->max_recv_dlength) {
 748                 iscsi_conn_printk(KERN_ERR, conn,
 749                                   "iscsi_tcp: datalen %d > %d\n",
 750                                   tcp_conn->in.datalen, conn->max_recv_dlength);
 751                 return ISCSI_ERR_DATALEN;
 752         }
 753
 754         /* Additional header segments. So far, we don't
 755          * process additional headers.
 756          */
 757         ahslen = hdr->hlength << 2;
 758
 759         opcode = hdr->opcode & ISCSI_OPCODE_MASK;
 760         /* verify itt (itt encoding: age+cid+itt) */
 761         rc = iscsi_verify_itt(conn, hdr, &itt);
 762         if (rc)
 763                 return rc;
 764
 765         debug_tcp("opcode 0x%x ahslen %d datalen %d\n",
 766                   opcode, ahslen, tcp_conn->in.datalen);
 767
 768         switch(opcode) {
 769         case ISCSI_OP_SCSI_DATA_IN:
 770                 ctask = session->cmds[itt];
 771                 spin_lock(&conn->session->lock);
 772                 rc = iscsi_data_rsp(conn, ctask);
 773                 spin_unlock(&conn->session->lock);
 774                 if (rc)
 775                         return rc;
 776                 if (tcp_conn->in.datalen) {
 777                         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 778                         struct hash_desc *rx_hash = NULL;
 779                         struct scsi_data_buffer *sdb = scsi_in(ctask->sc);
 780
 781                         /*
 782                          * Setup copy of Data-In into the Scsi_Cmnd
 783                          * Scatterlist case:
 784                          * We set up the iscsi_segment to point to the next
 785                          * scatterlist entry to copy to. As we go along,
 786                          * we move on to the next scatterlist entry and
 787                          * update the digest per-entry.
 788                          */
 789                         if (conn->datadgst_en)
 790                                 rx_hash = &tcp_conn->rx_hash;
 791
 792                         debug_tcp("iscsi_tcp_begin_data_in(%p, offset=%d, "
 793                                   "datalen=%d)\n", tcp_conn,
 794                                   tcp_ctask->data_offset,
 795                                   tcp_conn->in.datalen);
 796                         return iscsi_segment_seek_sg(&tcp_conn->in.segment,
 797                                                      sdb->table.sgl,
 798                                                      sdb->table.nents,
 799                                                      tcp_ctask->data_offset,
 800                                                      tcp_conn->in.datalen,
 801                                                      iscsi_tcp_process_data_in,
 802                                                      rx_hash);
 803                 }
 804                 /* fall through */
 805         case ISCSI_OP_SCSI_CMD_RSP:
 806                 if (tcp_conn->in.datalen) {
 807                         iscsi_tcp_data_recv_prep(tcp_conn);
 808                         return 0;
 809                 }
 810                 rc = iscsi_complete_pdu(conn, hdr, NULL, 0);
 811                 break;
 812         case ISCSI_OP_R2T:
 813                 ctask = session->cmds[itt];
 814                 if (ahslen)
 815                         rc = ISCSI_ERR_AHSLEN;
 816                 else if (ctask->sc->sc_data_direction == DMA_TO_DEVICE) {
 817                         spin_lock(&session->lock);
 818                         rc = iscsi_r2t_rsp(conn, ctask);
 819                         spin_unlock(&session->lock);
 820                 } else
 821                         rc = ISCSI_ERR_PROTO;
 822                 break;
 823         case ISCSI_OP_LOGIN_RSP:
 824         case ISCSI_OP_TEXT_RSP:
 825         case ISCSI_OP_REJECT:
 826         case ISCSI_OP_ASYNC_EVENT:
 827                 /*
 828                  * It is possible that we could get a PDU with a buffer larger
 829                  * than 8K, but there are no targets that currently do this.
 830                  * For now we fail until we find a vendor that needs it
 831                  */
 832                 if (ISCSI_DEF_MAX_RECV_SEG_LEN < tcp_conn->in.datalen) {
 833                         iscsi_conn_printk(KERN_ERR, conn,
 834                                           "iscsi_tcp: received buffer of "
 835                                           "len %u but conn buffer is only %u "
 836                                           "(opcode %0x)\n",
 837                                           tcp_conn->in.datalen,
 838                                           ISCSI_DEF_MAX_RECV_SEG_LEN, opcode);
 839                         rc = ISCSI_ERR_PROTO;
 840                         break;
 841                 }
 842
 843                 /* If there's data coming in with the response,
 844                  * receive it to the connection's buffer.
 845                  */
 846                 if (tcp_conn->in.datalen) {
 847                         iscsi_tcp_data_recv_prep(tcp_conn);
 848                         return 0;
 849                 }
 850         /* fall through */
 851         case ISCSI_OP_LOGOUT_RSP:
 852         case ISCSI_OP_NOOP_IN:
 853         case ISCSI_OP_SCSI_TMFUNC_RSP:
 854                 rc = iscsi_complete_pdu(conn, hdr, NULL, 0);
 855                 break;
 856         default:
 857                 rc = ISCSI_ERR_BAD_OPCODE;
 858                 break;
 859         }
 860
 861         if (rc == 0) {
 862                 /* Anything that comes with data should have
 863                  * been handled above. */
 864                 if (tcp_conn->in.datalen)
 865                         return ISCSI_ERR_PROTO;
 866                 iscsi_tcp_hdr_recv_prep(tcp_conn);
 867         }
 868
 869         return rc;
 870 }
 871
 872 /**
 873  * iscsi_tcp_hdr_recv_done - process PDU header
 874  *
 875  * This is the callback invoked when the PDU header has
 876  * been received. If the header is followed by additional
 877  * header segments, we go back for more data.
 878  */
 879 static int
 880 iscsi_tcp_hdr_recv_done(struct iscsi_tcp_conn *tcp_conn,
 881                         struct iscsi_segment *segment)
 882 {
 883         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 884         struct iscsi_hdr *hdr;
 885
 886         /* Check if there are additional header segments
 887          * *prior* to computing the digest, because we
 888          * may need to go back to the caller for more.
 889          */
 890         hdr = (struct iscsi_hdr *) tcp_conn->in.hdr_buf;
 891         if (segment->copied == sizeof(struct iscsi_hdr) && hdr->hlength) {
 892                 /* Bump the header length - the caller will
 893                  * just loop around and get the AHS for us, and
 894                  * call again. */
 895                 unsigned int ahslen = hdr->hlength << 2;
 896
 897                 /* Make sure we don't overflow */
 898                 if (sizeof(*hdr) + ahslen > sizeof(tcp_conn->in.hdr_buf))
 899                         return ISCSI_ERR_AHSLEN;
 900
 901                 segment->total_size += ahslen;
 902                 segment->size += ahslen;
 903                 return 0;
 904         }
 905
 906         /* We're done processing the header. See if we're doing
 907          * header digests; if so, set up the recv_digest buffer
 908          * and go back for more. */
 909         if (conn->hdrdgst_en) {
 910                 if (segment->digest_len == 0) {
 911                         iscsi_tcp_segment_splice_digest(segment,
 912                                                         segment->recv_digest);
 913                         return 0;
 914                 }
 915                 iscsi_tcp_dgst_header(&tcp_conn->rx_hash, hdr,
 916                                       segment->total_copied - ISCSI_DIGEST_SIZE,
 917                                       segment->digest);
 918
 919                 if (!iscsi_tcp_dgst_verify(tcp_conn, segment))
 920                         return ISCSI_ERR_HDR_DGST;
 921         }
 922
 923         tcp_conn->in.hdr = hdr;
 924         return iscsi_tcp_hdr_dissect(conn, hdr);
 925 }
 926
 927 /**
 928  * iscsi_tcp_recv - TCP receive in sendfile fashion
 929  * @rd_desc: read descriptor
 930  * @skb: socket buffer
 931  * @offset: offset in skb
 932  * @len: skb->len - offset
 933  **/
 934 static int
 935 iscsi_tcp_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 936                unsigned int offset, size_t len)
 937 {
 938         struct iscsi_conn *conn = rd_desc->arg.data;
 939         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 940         struct iscsi_segment *segment = &tcp_conn->in.segment;
 941         struct skb_seq_state seq;
 942         unsigned int consumed = 0;
 943         int rc = 0;
 944
 945         debug_tcp("in %d bytes\n", skb->len - offset);
 946
 947         if (unlikely(conn->suspend_rx)) {
 948                 debug_tcp("conn %d Rx suspended!\n", conn->id);
 949                 return 0;
 950         }
 951
 952         skb_prepare_seq_read(skb, offset, skb->len, &seq);
 953         while (1) {
 954                 unsigned int avail;
 955                 const u8 *ptr;
 956
 957                 avail = skb_seq_read(consumed, &ptr, &seq);
 958                 if (avail == 0) {
 959                         debug_tcp("no more data avail. Consumed %d\n",
 960                                   consumed);
 961                         break;
 962                 }
 963                 BUG_ON(segment->copied >= segment->size);
 964
 965                 debug_tcp("skb %p ptr=%p avail=%u\n", skb, ptr, avail);
 966                 rc = iscsi_tcp_segment_recv(tcp_conn, segment, ptr, avail);
 967                 BUG_ON(rc == 0);
 968                 consumed += rc;
 969
 970                 if (segment->total_copied >= segment->total_size) {
 971                         debug_tcp("segment done\n");
 972                         rc = segment->done(tcp_conn, segment);
 973                         if (rc != 0) {
 974                                 skb_abort_seq_read(&seq);
 975                                 goto error;
 976                         }
 977
 978                         /* The done() functions sets up the
 979                          * next segment. */
 980                 }
 981         }
 982         skb_abort_seq_read(&seq);
 983         conn->rxdata_octets += consumed;
 984         return consumed;
 985
 986 error:
 987         debug_tcp("Error receiving PDU, errno=%d\n", rc);
 988         iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
 989         return 0;
 990 }
 991
 992 static void
 993 iscsi_tcp_data_ready(struct sock *sk, int flag)
 994 {
 995         struct iscsi_conn *conn = sk->sk_user_data;
 996         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 997         read_descriptor_t rd_desc;
 998
 999         read_lock(&sk->sk_callback_lock);
1000
1001         /*
1002          * Use rd_desc to pass 'conn' to iscsi_tcp_recv.
1003          * We set count to 1 because we want the network layer to
1004          * hand us all the skbs that are available. iscsi_tcp_recv
1005          * handled pdus that cross buffers or pdus that still need data.
1006          */
1007         rd_desc.arg.data = conn;
1008         rd_desc.count = 1;
1009         tcp_read_sock(sk, &rd_desc, iscsi_tcp_recv);
1010
1011         read_unlock(&sk->sk_callback_lock);
1012
1013         /* If we had to (atomically) map a highmem page,
1014          * unmap it now. */
1015         iscsi_tcp_segment_unmap(&tcp_conn->in.segment);
1016 }
1017
1018 static void
1019 iscsi_tcp_state_change(struct sock *sk)
1020 {
1021         struct iscsi_tcp_conn *tcp_conn;
1022         struct iscsi_conn *conn;
1023         struct iscsi_session *session;
1024         void (*old_state_change)(struct sock *);
1025
1026         read_lock(&sk->sk_callback_lock);
1027
1028         conn = (struct iscsi_conn*)sk->sk_user_data;
1029         session = conn->session;
1030
1031         if ((sk->sk_state == TCP_CLOSE_WAIT ||
1032              sk->sk_state == TCP_CLOSE) &&
1033             !atomic_read(&sk->sk_rmem_alloc)) {
1034                 debug_tcp("iscsi_tcp_state_change: TCP_CLOSE|TCP_CLOSE_WAIT\n");
1035                 iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
1036         }
1037
1038         tcp_conn = conn->dd_data;
1039         old_state_change = tcp_conn->old_state_change;
1040
1041         read_unlock(&sk->sk_callback_lock);
1042
1043         old_state_change(sk);
1044 }
1045
1046 /**
1047  * iscsi_write_space - Called when more output buffer space is available
1048  * @sk: socket space is available for
1049  **/
1050 static void
1051 iscsi_write_space(struct sock *sk)
1052 {
1053         struct iscsi_conn *conn = (struct iscsi_conn*)sk->sk_user_data;
1054         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1055
1056         tcp_conn->old_write_space(sk);
1057         debug_tcp("iscsi_write_space: cid %d\n", conn->id);
1058         scsi_queue_work(conn->session->host, &conn->xmitwork);
1059 }
1060
1061 static void
1062 iscsi_conn_set_callbacks(struct iscsi_conn *conn)
1063 {
1064         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1065         struct sock *sk = tcp_conn->sock->sk;
1066
1067         /* assign new callbacks */
1068         write_lock_bh(&sk->sk_callback_lock);
1069         sk->sk_user_data = conn;
1070         tcp_conn->old_data_ready = sk->sk_data_ready;
1071         tcp_conn->old_state_change = sk->sk_state_change;
1072         tcp_conn->old_write_space = sk->sk_write_space;
1073         sk->sk_data_ready = iscsi_tcp_data_ready;
1074         sk->sk_state_change = iscsi_tcp_state_change;
1075         sk->sk_write_space = iscsi_write_space;
1076         write_unlock_bh(&sk->sk_callback_lock);
1077 }
1078
1079 static void
1080 iscsi_conn_restore_callbacks(struct iscsi_tcp_conn *tcp_conn)
1081 {
1082         struct sock *sk = tcp_conn->sock->sk;
1083
1084         /* restore socket callbacks, see also: iscsi_conn_set_callbacks() */
1085         write_lock_bh(&sk->sk_callback_lock);
1086         sk->sk_user_data    = NULL;
1087         sk->sk_data_ready   = tcp_conn->old_data_ready;
1088         sk->sk_state_change = tcp_conn->old_state_change;
1089         sk->sk_write_space  = tcp_conn->old_write_space;
1090         sk->sk_no_check  = 0;
1091         write_unlock_bh(&sk->sk_callback_lock);
1092 }
1093
1094 /**
1095  * iscsi_xmit - TCP transmit
1096  **/
1097 static int
1098 iscsi_xmit(struct iscsi_conn *conn)
1099 {
1100         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1101         struct iscsi_segment *segment = &tcp_conn->out.segment;
1102         unsigned int consumed = 0;
1103         int rc = 0;
1104
1105         while (1) {
1106                 rc = iscsi_tcp_xmit_segment(tcp_conn, segment);
1107                 if (rc < 0)
1108                         goto error;
1109                 if (rc == 0)
1110                         break;
1111
1112                 consumed += rc;
1113
1114                 if (segment->total_copied >= segment->total_size) {
1115                         if (segment->done != NULL) {
1116                                 rc = segment->done(tcp_conn, segment);
1117                                 if (rc < 0)
1118                                         goto error;
1119                         }
1120                 }
1121         }
1122
1123         debug_tcp("xmit %d bytes\n", consumed);
1124
1125         conn->txdata_octets += consumed;
1126         return consumed;
1127
1128 error:
1129         /* Transmit error. We could initiate error recovery
1130          * here. */
1131         debug_tcp("Error sending PDU, errno=%d\n", rc);
1132         iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
1133         return rc;
1134 }
1135
1136 /**
1137  * iscsi_tcp_xmit_qlen - return the number of bytes queued for xmit
1138  */
1139 static inline int
1140 iscsi_tcp_xmit_qlen(struct iscsi_conn *conn)
1141 {
1142         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1143         struct iscsi_segment *segment = &tcp_conn->out.segment;
1144
1145         return segment->total_copied - segment->total_size;
1146 }
1147
1148 static inline int
1149 iscsi_tcp_flush(struct iscsi_conn *conn)
1150 {
1151         int rc;
1152
1153         while (iscsi_tcp_xmit_qlen(conn)) {
1154                 rc = iscsi_xmit(conn);
1155                 if (rc == 0)
1156                         return -EAGAIN;
1157                 if (rc < 0)
1158                         return rc;
1159         }
1160
1161         return 0;
1162 }
1163
1164 /*
1165  * This is called when we're done sending the header.
1166  * Simply copy the data_segment to the send segment, and return.
1167  */
1168 static int
1169 iscsi_tcp_send_hdr_done(struct iscsi_tcp_conn *tcp_conn,
1170                         struct iscsi_segment *segment)
1171 {
1172         tcp_conn->out.segment = tcp_conn->out.data_segment;
1173         debug_tcp("Header done. Next segment size %u total_size %u\n",
1174                   tcp_conn->out.segment.size, tcp_conn->out.segment.total_size);
1175         return 0;
1176 }
1177
1178 static void
1179 iscsi_tcp_send_hdr_prep(struct iscsi_conn *conn, void *hdr, size_t hdrlen)
1180 {
1181         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1182
1183         debug_tcp("%s(%p%s)\n", __FUNCTION__, tcp_conn,
1184                         conn->hdrdgst_en? ", digest enabled" : "");
1185
1186         /* Clear the data segment - needs to be filled in by the
1187          * caller using iscsi_tcp_send_data_prep() */
1188         memset(&tcp_conn->out.data_segment, 0, sizeof(struct iscsi_segment));
1189
1190         /* If header digest is enabled, compute the CRC and
1191          * place the digest into the same buffer. We make
1192          * sure that both iscsi_tcp_ctask and mtask have
1193          * sufficient room.
1194          */
1195         if (conn->hdrdgst_en) {
1196                 iscsi_tcp_dgst_header(&tcp_conn->tx_hash, hdr, hdrlen,
1197                                       hdr + hdrlen);
1198                 hdrlen += ISCSI_DIGEST_SIZE;
1199         }
1200
1201         /* Remember header pointer for later, when we need
1202          * to decide whether there's a payload to go along
1203          * with the header. */
1204         tcp_conn->out.hdr = hdr;
1205
1206         iscsi_segment_init_linear(&tcp_conn->out.segment, hdr, hdrlen,
1207                                 iscsi_tcp_send_hdr_done, NULL);
1208 }
1209
1210 /*
1211  * Prepare the send buffer for the payload data.
1212  * Padding and checksumming will all be taken care
1213  * of by the iscsi_segment routines.
1214  */
1215 static int
1216 iscsi_tcp_send_data_prep(struct iscsi_conn *conn, struct scatterlist *sg,
1217                          unsigned int count, unsigned int offset,
1218                          unsigned int len)
1219 {
1220         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1221         struct hash_desc *tx_hash = NULL;
1222         unsigned int hdr_spec_len;
1223
1224         debug_tcp("%s(%p, offset=%d, datalen=%d%s)\n", __FUNCTION__,
1225                         tcp_conn, offset, len,
1226                         conn->datadgst_en? ", digest enabled" : "");
1227
1228         /* Make sure the datalen matches what the caller
1229            said he would send. */
1230         hdr_spec_len = ntoh24(tcp_conn->out.hdr->dlength);
1231         WARN_ON(iscsi_padded(len) != iscsi_padded(hdr_spec_len));
1232
1233         if (conn->datadgst_en)
1234                 tx_hash = &tcp_conn->tx_hash;
1235
1236         return iscsi_segment_seek_sg(&tcp_conn->out.data_segment,
1237                                    sg, count, offset, len,
1238                                    NULL, tx_hash);
1239 }
1240
1241 static void
1242 iscsi_tcp_send_linear_data_prepare(struct iscsi_conn *conn, void *data,
1243                                    size_t len)
1244 {
1245         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1246         struct hash_desc *tx_hash = NULL;
1247         unsigned int hdr_spec_len;
1248
1249         debug_tcp("%s(%p, datalen=%d%s)\n", __FUNCTION__, tcp_conn, len,
1250                   conn->datadgst_en? ", digest enabled" : "");
1251
1252         /* Make sure the datalen matches what the caller
1253            said he would send. */
1254         hdr_spec_len = ntoh24(tcp_conn->out.hdr->dlength);
1255         WARN_ON(iscsi_padded(len) != iscsi_padded(hdr_spec_len));
1256
1257         if (conn->datadgst_en)
1258                 tx_hash = &tcp_conn->tx_hash;
1259
1260         iscsi_segment_init_linear(&tcp_conn->out.data_segment,
1261                                 data, len, NULL, tx_hash);
1262 }
1263
1264 /**
1265  * iscsi_solicit_data_cont - initialize next Data-Out
1266  * @conn: iscsi connection
1267  * @ctask: scsi command task
1268  * @r2t: R2T info
1269  * @left: bytes left to transfer
1270  *
1271  * Notes:
1272  *      Initialize next Data-Out within this R2T sequence and continue
1273  *      to process next Scatter-Gather element(if any) of this SCSI command.
1274  *
1275  *      Called under connection lock.
1276  **/
1277 static int
1278 iscsi_solicit_data_cont(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask,
1279                         struct iscsi_r2t_info *r2t)
1280 {
1281         struct iscsi_data *hdr;
1282         int new_offset, left;
1283
1284         BUG_ON(r2t->data_length - r2t->sent < 0);
1285         left = r2t->data_length - r2t->sent;
1286         if (left == 0)
1287                 return 0;
1288
1289         hdr = &r2t->dtask.hdr;
1290         memset(hdr, 0, sizeof(struct iscsi_data));
1291         hdr->ttt = r2t->ttt;
1292         hdr->datasn = cpu_to_be32(r2t->solicit_datasn);
1293         r2t->solicit_datasn++;
1294         hdr->opcode = ISCSI_OP_SCSI_DATA_OUT;
1295         memcpy(hdr->lun, ctask->hdr->lun, sizeof(hdr->lun));
1296         hdr->itt = ctask->hdr->itt;
1297         hdr->exp_statsn = r2t->exp_statsn;
1298         new_offset = r2t->data_offset + r2t->sent;
1299         hdr->offset = cpu_to_be32(new_offset);
1300         if (left > conn->max_xmit_dlength) {
1301                 hton24(hdr->dlength, conn->max_xmit_dlength);
1302                 r2t->data_count = conn->max_xmit_dlength;
1303         } else {
1304                 hton24(hdr->dlength, left);
1305                 r2t->data_count = left;
1306                 hdr->flags = ISCSI_FLAG_CMD_FINAL;
1307         }
1308
1309         conn->dataout_pdus_cnt++;
1310         return 1;
1311 }
1312
1313 /**
1314  * iscsi_tcp_ctask - Initialize iSCSI SCSI_READ or SCSI_WRITE commands
1315  * @conn: iscsi connection
1316  * @ctask: scsi command task
1317  * @sc: scsi command
1318  **/
1319 static int
1320 iscsi_tcp_ctask_init(struct iscsi_cmd_task *ctask)
1321 {
1322         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1323         struct iscsi_conn *conn = ctask->conn;
1324         struct scsi_cmnd *sc = ctask->sc;
1325         int err;
1326
1327         BUG_ON(__kfifo_len(tcp_ctask->r2tqueue));
1328         tcp_ctask->sent = 0;
1329         tcp_ctask->exp_datasn = 0;
1330
1331         /* Prepare PDU, optionally w/ immediate data */
1332         debug_scsi("ctask deq [cid %d itt 0x%x imm %d unsol %d]\n",
1333                     conn->id, ctask->itt, ctask->imm_count,
1334                     ctask->unsol_count);
1335         iscsi_tcp_send_hdr_prep(conn, ctask->hdr, ctask->hdr_len);
1336
1337         if (!ctask->imm_count)
1338                 return 0;
1339
1340         /* If we have immediate data, attach a payload */
1341         err = iscsi_tcp_send_data_prep(conn, scsi_out(sc)->table.sgl,
1342                                        scsi_out(sc)->table.nents,
1343                                        0, ctask->imm_count);
1344         if (err)
1345                 return err;
1346         tcp_ctask->sent += ctask->imm_count;
1347         ctask->imm_count = 0;
1348         return 0;
1349 }
1350
1351 /**
1352  * iscsi_tcp_mtask_xmit - xmit management(immediate) task
1353  * @conn: iscsi connection
1354  * @mtask: task management task
1355  *
1356  * Notes:
1357  *      The function can return -EAGAIN in which case caller must
1358  *      call it again later, or recover. '0' return code means successful
1359  *      xmit.
1360  **/
1361 static int
1362 iscsi_tcp_mtask_xmit(struct iscsi_conn *conn, struct iscsi_mgmt_task *mtask)
1363 {
1364         int rc;
1365
1366         /* Flush any pending data first. */
1367         rc = iscsi_tcp_flush(conn);
1368         if (rc < 0)
1369                 return rc;
1370
1371         if (mtask->hdr->itt == RESERVED_ITT) {
1372                 struct iscsi_session *session = conn->session;
1373
1374                 spin_lock_bh(&session->lock);
1375                 iscsi_free_mgmt_task(conn, mtask);
1376                 spin_unlock_bh(&session->lock);
1377         }
1378
1379         return 0;
1380 }
1381
1382 /*
1383  * iscsi_tcp_ctask_xmit - xmit normal PDU task
1384  * @conn: iscsi connection
1385  * @ctask: iscsi command task
1386  *
1387  * We're expected to return 0 when everything was transmitted succesfully,
1388  * -EAGAIN if there's still data in the queue, or != 0 for any other kind
1389  * of error.
1390  */
1391 static int
1392 iscsi_tcp_ctask_xmit(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1393 {
1394         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1395         struct scsi_cmnd *sc = ctask->sc;
1396         struct scsi_data_buffer *sdb = scsi_out(sc);
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         /* Are we done already? */
1406         if (sc->sc_data_direction != DMA_TO_DEVICE)
1407                 return 0;
1408
1409         if (ctask->unsol_count != 0) {
1410                 struct iscsi_data *hdr = &tcp_ctask->unsol_dtask.hdr;
1411
1412                 /* Prepare a header for the unsolicited PDU.
1413                  * The amount of data we want to send will be
1414                  * in ctask->data_count.
1415                  * FIXME: return the data count instead.
1416                  */
1417                 iscsi_prep_unsolicit_data_pdu(ctask, hdr);
1418
1419                 debug_tcp("unsol dout [itt 0x%x doff %d dlen %d]\n",
1420                                 ctask->itt, tcp_ctask->sent, ctask->data_count);
1421
1422                 iscsi_tcp_send_hdr_prep(conn, hdr, sizeof(*hdr));
1423                 rc = iscsi_tcp_send_data_prep(conn, sdb->table.sgl,
1424                                               sdb->table.nents, tcp_ctask->sent,
1425                                               ctask->data_count);
1426                 if (rc)
1427                         goto fail;
1428                 tcp_ctask->sent += ctask->data_count;
1429                 ctask->unsol_count -= ctask->data_count;
1430                 goto flush;
1431         } else {
1432                 struct iscsi_session *session = conn->session;
1433                 struct iscsi_r2t_info *r2t;
1434
1435                 /* All unsolicited PDUs sent. Check for solicited PDUs.
1436                  */
1437                 spin_lock_bh(&session->lock);
1438                 r2t = tcp_ctask->r2t;
1439                 if (r2t != NULL) {
1440                         /* Continue with this R2T? */
1441                         if (!iscsi_solicit_data_cont(conn, ctask, r2t)) {
1442                                 debug_scsi("  done with r2t %p\n", r2t);
1443
1444                                 __kfifo_put(tcp_ctask->r2tpool.queue,
1445                                             (void*)&r2t, sizeof(void*));
1446                                 tcp_ctask->r2t = r2t = NULL;
1447                         }
1448                 }
1449
1450                 if (r2t == NULL) {
1451                         __kfifo_get(tcp_ctask->r2tqueue, (void*)&tcp_ctask->r2t,
1452                                     sizeof(void*));
1453                         r2t = tcp_ctask->r2t;
1454                 }
1455                 spin_unlock_bh(&session->lock);
1456
1457                 /* Waiting for more R2Ts to arrive. */
1458                 if (r2t == NULL) {
1459                         debug_tcp("no R2Ts yet\n");
1460                         return 0;
1461                 }
1462
1463                 debug_scsi("sol dout %p [dsn %d itt 0x%x doff %d dlen %d]\n",
1464                         r2t, r2t->solicit_datasn - 1, ctask->itt,
1465                         r2t->data_offset + r2t->sent, r2t->data_count);
1466
1467                 iscsi_tcp_send_hdr_prep(conn, &r2t->dtask.hdr,
1468                                         sizeof(struct iscsi_hdr));
1469
1470                 rc = iscsi_tcp_send_data_prep(conn, sdb->table.sgl,
1471                                               sdb->table.nents,
1472                                               r2t->data_offset + r2t->sent,
1473                                               r2t->data_count);
1474                 if (rc)
1475                         goto fail;
1476                 tcp_ctask->sent += r2t->data_count;
1477                 r2t->sent += r2t->data_count;
1478                 goto flush;
1479         }
1480         return 0;
1481 fail:
1482         iscsi_conn_failure(conn, rc);
1483         return -EIO;
1484 }
1485
1486 static struct iscsi_cls_conn *
1487 iscsi_tcp_conn_create(struct iscsi_cls_session *cls_session, uint32_t conn_idx)
1488 {
1489         struct iscsi_conn *conn;
1490         struct iscsi_cls_conn *cls_conn;
1491         struct iscsi_tcp_conn *tcp_conn;
1492
1493         cls_conn = iscsi_conn_setup(cls_session, conn_idx);
1494         if (!cls_conn)
1495                 return NULL;
1496         conn = cls_conn->dd_data;
1497         /*
1498          * due to strange issues with iser these are not set
1499          * in iscsi_conn_setup
1500          */
1501         conn->max_recv_dlength = ISCSI_DEF_MAX_RECV_SEG_LEN;
1502
1503         tcp_conn = kzalloc(sizeof(*tcp_conn), GFP_KERNEL);
1504         if (!tcp_conn)
1505                 goto tcp_conn_alloc_fail;
1506
1507         conn->dd_data = tcp_conn;
1508         tcp_conn->iscsi_conn = conn;
1509
1510         tcp_conn->tx_hash.tfm = crypto_alloc_hash("crc32c", 0,
1511                                                   CRYPTO_ALG_ASYNC);
1512         tcp_conn->tx_hash.flags = 0;
1513         if (IS_ERR(tcp_conn->tx_hash.tfm))
1514                 goto free_tcp_conn;
1515
1516         tcp_conn->rx_hash.tfm = crypto_alloc_hash("crc32c", 0,
1517                                                   CRYPTO_ALG_ASYNC);
1518         tcp_conn->rx_hash.flags = 0;
1519         if (IS_ERR(tcp_conn->rx_hash.tfm))
1520                 goto free_tx_tfm;
1521
1522         return cls_conn;
1523
1524 free_tx_tfm:
1525         crypto_free_hash(tcp_conn->tx_hash.tfm);
1526 free_tcp_conn:
1527         iscsi_conn_printk(KERN_ERR, conn,
1528                           "Could not create connection due to crc32c "
1529                           "loading error. Make sure the crc32c "
1530                           "module is built as a module or into the "
1531                           "kernel\n");
1532         kfree(tcp_conn);
1533 tcp_conn_alloc_fail:
1534         iscsi_conn_teardown(cls_conn);
1535         return NULL;
1536 }
1537
1538 static void
1539 iscsi_tcp_release_conn(struct iscsi_conn *conn)
1540 {
1541         struct iscsi_session *session = conn->session;
1542         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1543         struct socket *sock = tcp_conn->sock;
1544
1545         if (!sock)
1546                 return;
1547
1548         sock_hold(sock->sk);
1549         iscsi_conn_restore_callbacks(tcp_conn);
1550         sock_put(sock->sk);
1551
1552         spin_lock_bh(&session->lock);
1553         tcp_conn->sock = NULL;
1554         conn->recv_lock = NULL;
1555         spin_unlock_bh(&session->lock);
1556         sockfd_put(sock);
1557 }
1558
1559 static void
1560 iscsi_tcp_conn_destroy(struct iscsi_cls_conn *cls_conn)
1561 {
1562         struct iscsi_conn *conn = cls_conn->dd_data;
1563         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1564
1565         iscsi_tcp_release_conn(conn);
1566         iscsi_conn_teardown(cls_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         kfree(tcp_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
1581         iscsi_conn_stop(cls_conn, flag);
1582         iscsi_tcp_release_conn(conn);
1583 }
1584
1585 static int iscsi_tcp_get_addr(struct iscsi_conn *conn, struct socket *sock,
1586                               char *buf, int *port,
1587                               int (*getname)(struct socket *, struct sockaddr *,
1588                                         int *addrlen))
1589 {
1590         struct sockaddr_storage *addr;
1591         struct sockaddr_in6 *sin6;
1592         struct sockaddr_in *sin;
1593         int rc = 0, len;
1594
1595         addr = kmalloc(sizeof(*addr), GFP_KERNEL);
1596         if (!addr)
1597                 return -ENOMEM;
1598
1599         if (getname(sock, (struct sockaddr *) addr, &len)) {
1600                 rc = -ENODEV;
1601                 goto free_addr;
1602         }
1603
1604         switch (addr->ss_family) {
1605         case AF_INET:
1606                 sin = (struct sockaddr_in *)addr;
1607                 spin_lock_bh(&conn->session->lock);
1608                 sprintf(buf, NIPQUAD_FMT, NIPQUAD(sin->sin_addr.s_addr));
1609                 *port = be16_to_cpu(sin->sin_port);
1610                 spin_unlock_bh(&conn->session->lock);
1611                 break;
1612         case AF_INET6:
1613                 sin6 = (struct sockaddr_in6 *)addr;
1614                 spin_lock_bh(&conn->session->lock);
1615                 sprintf(buf, NIP6_FMT, NIP6(sin6->sin6_addr));
1616                 *port = be16_to_cpu(sin6->sin6_port);
1617                 spin_unlock_bh(&conn->session->lock);
1618                 break;
1619         }
1620 free_addr:
1621         kfree(addr);
1622         return rc;
1623 }
1624
1625 static int
1626 iscsi_tcp_conn_bind(struct iscsi_cls_session *cls_session,
1627                     struct iscsi_cls_conn *cls_conn, uint64_t transport_eph,
1628                     int is_leading)
1629 {
1630         struct Scsi_Host *shost = iscsi_session_to_shost(cls_session);
1631         struct iscsi_host *ihost = shost_priv(shost);
1632         struct iscsi_conn *conn = cls_conn->dd_data;
1633         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1634         struct sock *sk;
1635         struct socket *sock;
1636         int err;
1637
1638         /* lookup for existing socket */
1639         sock = sockfd_lookup((int)transport_eph, &err);
1640         if (!sock) {
1641                 iscsi_conn_printk(KERN_ERR, conn,
1642                                   "sockfd_lookup failed %d\n", err);
1643                 return -EEXIST;
1644         }
1645         /*
1646          * copy these values now because if we drop the session
1647          * userspace may still want to query the values since we will
1648          * be using them for the reconnect
1649          */
1650         err = iscsi_tcp_get_addr(conn, sock, conn->portal_address,
1651                                  &conn->portal_port, kernel_getpeername);
1652         if (err)
1653                 goto free_socket;
1654
1655         err = iscsi_tcp_get_addr(conn, sock, ihost->local_address,
1656                                 &ihost->local_port, kernel_getsockname);
1657         if (err)
1658                 goto free_socket;
1659
1660         err = iscsi_conn_bind(cls_session, cls_conn, is_leading);
1661         if (err)
1662                 goto free_socket;
1663
1664         /* bind iSCSI connection and socket */
1665         tcp_conn->sock = sock;
1666
1667         /* setup Socket parameters */
1668         sk = sock->sk;
1669         sk->sk_reuse = 1;
1670         sk->sk_sndtimeo = 15 * HZ; /* FIXME: make it configurable */
1671         sk->sk_allocation = GFP_ATOMIC;
1672
1673         /* FIXME: disable Nagle's algorithm */
1674
1675         /*
1676          * Intercept TCP callbacks for sendfile like receive
1677          * processing.
1678          */
1679         conn->recv_lock = &sk->sk_callback_lock;
1680         iscsi_conn_set_callbacks(conn);
1681         tcp_conn->sendpage = tcp_conn->sock->ops->sendpage;
1682         /*
1683          * set receive state machine into initial state
1684          */
1685         iscsi_tcp_hdr_recv_prep(tcp_conn);
1686         return 0;
1687
1688 free_socket:
1689         sockfd_put(sock);
1690         return err;
1691 }
1692
1693 /* called with host lock */
1694 static void
1695 iscsi_tcp_mtask_init(struct iscsi_conn *conn, struct iscsi_mgmt_task *mtask)
1696 {
1697         debug_scsi("mtask deq [cid %d itt 0x%x]\n", conn->id, mtask->itt);
1698
1699         /* Prepare PDU, optionally w/ immediate data */
1700         iscsi_tcp_send_hdr_prep(conn, mtask->hdr, sizeof(*mtask->hdr));
1701
1702         /* If we have immediate data, attach a payload */
1703         if (mtask->data_count)
1704                 iscsi_tcp_send_linear_data_prepare(conn, mtask->data,
1705                                                    mtask->data_count);
1706 }
1707
1708 static int
1709 iscsi_r2tpool_alloc(struct iscsi_session *session)
1710 {
1711         int i;
1712         int cmd_i;
1713
1714         /*
1715          * initialize per-task: R2T pool and xmit queue
1716          */
1717         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
1718                 struct iscsi_cmd_task *ctask = session->cmds[cmd_i];
1719                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1720
1721                 /*
1722                  * pre-allocated x4 as much r2ts to handle race when
1723                  * target acks DataOut faster than we data_xmit() queues
1724                  * could replenish r2tqueue.
1725                  */
1726
1727                 /* R2T pool */
1728                 if (iscsi_pool_init(&tcp_ctask->r2tpool, session->max_r2t * 4, NULL,
1729                                     sizeof(struct iscsi_r2t_info))) {
1730                         goto r2t_alloc_fail;
1731                 }
1732
1733                 /* R2T xmit queue */
1734                 tcp_ctask->r2tqueue = kfifo_alloc(
1735                       session->max_r2t * 4 * sizeof(void*), GFP_KERNEL, NULL);
1736                 if (tcp_ctask->r2tqueue == ERR_PTR(-ENOMEM)) {
1737                         iscsi_pool_free(&tcp_ctask->r2tpool);
1738                         goto r2t_alloc_fail;
1739                 }
1740         }
1741
1742         return 0;
1743
1744 r2t_alloc_fail:
1745         for (i = 0; i < cmd_i; i++) {
1746                 struct iscsi_cmd_task *ctask = session->cmds[i];
1747                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1748
1749                 kfifo_free(tcp_ctask->r2tqueue);
1750                 iscsi_pool_free(&tcp_ctask->r2tpool);
1751         }
1752         return -ENOMEM;
1753 }
1754
1755 static void
1756 iscsi_r2tpool_free(struct iscsi_session *session)
1757 {
1758         int i;
1759
1760         for (i = 0; i < session->cmds_max; i++) {
1761                 struct iscsi_cmd_task *ctask = session->cmds[i];
1762                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1763
1764                 kfifo_free(tcp_ctask->r2tqueue);
1765                 iscsi_pool_free(&tcp_ctask->r2tpool);
1766         }
1767 }
1768
1769 static int
1770 iscsi_conn_set_param(struct iscsi_cls_conn *cls_conn, enum iscsi_param param,
1771                      char *buf, int buflen)
1772 {
1773         struct iscsi_conn *conn = cls_conn->dd_data;
1774         struct iscsi_session *session = conn->session;
1775         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1776         int value;
1777
1778         switch(param) {
1779         case ISCSI_PARAM_HDRDGST_EN:
1780                 iscsi_set_param(cls_conn, param, buf, buflen);
1781                 break;
1782         case ISCSI_PARAM_DATADGST_EN:
1783                 iscsi_set_param(cls_conn, param, buf, buflen);
1784                 tcp_conn->sendpage = conn->datadgst_en ?
1785                         sock_no_sendpage : tcp_conn->sock->ops->sendpage;
1786                 break;
1787         case ISCSI_PARAM_MAX_R2T:
1788                 sscanf(buf, "%d", &value);
1789                 if (value <= 0 || !is_power_of_2(value))
1790                         return -EINVAL;
1791                 if (session->max_r2t == value)
1792                         break;
1793                 iscsi_r2tpool_free(session);
1794                 iscsi_set_param(cls_conn, param, buf, buflen);
1795                 if (iscsi_r2tpool_alloc(session))
1796                         return -ENOMEM;
1797                 break;
1798         default:
1799                 return iscsi_set_param(cls_conn, param, buf, buflen);
1800         }
1801
1802         return 0;
1803 }
1804
1805 static int
1806 iscsi_tcp_conn_get_param(struct iscsi_cls_conn *cls_conn,
1807                          enum iscsi_param param, char *buf)
1808 {
1809         struct iscsi_conn *conn = cls_conn->dd_data;
1810         int len;
1811
1812         switch(param) {
1813         case ISCSI_PARAM_CONN_PORT:
1814                 spin_lock_bh(&conn->session->lock);
1815                 len = sprintf(buf, "%hu\n", conn->portal_port);
1816                 spin_unlock_bh(&conn->session->lock);
1817                 break;
1818         case ISCSI_PARAM_CONN_ADDRESS:
1819                 spin_lock_bh(&conn->session->lock);
1820                 len = sprintf(buf, "%s\n", conn->portal_address);
1821                 spin_unlock_bh(&conn->session->lock);
1822                 break;
1823         default:
1824                 return iscsi_conn_get_param(cls_conn, param, buf);
1825         }
1826
1827         return len;
1828 }
1829
1830 static void
1831 iscsi_conn_get_stats(struct iscsi_cls_conn *cls_conn, struct iscsi_stats *stats)
1832 {
1833         struct iscsi_conn *conn = cls_conn->dd_data;
1834         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1835
1836         stats->txdata_octets = conn->txdata_octets;
1837         stats->rxdata_octets = conn->rxdata_octets;
1838         stats->scsicmd_pdus = conn->scsicmd_pdus_cnt;
1839         stats->dataout_pdus = conn->dataout_pdus_cnt;
1840         stats->scsirsp_pdus = conn->scsirsp_pdus_cnt;
1841         stats->datain_pdus = conn->datain_pdus_cnt;
1842         stats->r2t_pdus = conn->r2t_pdus_cnt;
1843         stats->tmfcmd_pdus = conn->tmfcmd_pdus_cnt;
1844         stats->tmfrsp_pdus = conn->tmfrsp_pdus_cnt;
1845         stats->custom_length = 3;
1846         strcpy(stats->custom[0].desc, "tx_sendpage_failures");
1847         stats->custom[0].value = tcp_conn->sendpage_failures_cnt;
1848         strcpy(stats->custom[1].desc, "rx_discontiguous_hdr");
1849         stats->custom[1].value = tcp_conn->discontiguous_hdr_cnt;
1850         strcpy(stats->custom[2].desc, "eh_abort_cnt");
1851         stats->custom[2].value = conn->eh_abort_cnt;
1852 }
1853
1854 static struct iscsi_cls_session *
1855 iscsi_tcp_session_create(struct Scsi_Host *shost, uint16_t cmds_max,
1856                          uint16_t qdepth, uint32_t initial_cmdsn,
1857                          uint32_t *hostno)
1858 {
1859         struct iscsi_cls_session *cls_session;
1860         struct iscsi_session *session;
1861         int cmd_i;
1862
1863         if (shost) {
1864                 printk(KERN_ERR "iscsi_tcp: invalid shost %d.\n",
1865                        shost->host_no);
1866                 return NULL;
1867         }
1868
1869         shost = iscsi_host_alloc(&iscsi_sht, 0, qdepth);
1870         if (!shost)
1871                 return NULL;
1872         shost->transportt = iscsi_tcp_scsi_transport;
1873         shost->max_lun = iscsi_max_lun;
1874         shost->max_id = 0;
1875         shost->max_channel = 0;
1876         shost->max_cmd_len = 16;
1877         shost->can_queue = cmds_max;
1878
1879         if (iscsi_host_add(shost, NULL))
1880                 goto free_host;
1881         *hostno = shost->host_no;
1882
1883         cls_session = iscsi_session_setup(&iscsi_tcp_transport, shost, cmds_max,
1884                                           sizeof(struct iscsi_tcp_cmd_task),
1885                                           sizeof(struct iscsi_tcp_mgmt_task),
1886                                           initial_cmdsn);
1887         if (!cls_session)
1888                 goto remove_host;
1889         session = cls_session->dd_data;
1890
1891         shost->can_queue = session->cmds_max;
1892         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
1893                 struct iscsi_cmd_task *ctask = session->cmds[cmd_i];
1894                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1895
1896                 ctask->hdr = &tcp_ctask->hdr.cmd_hdr;
1897                 ctask->hdr_max = sizeof(tcp_ctask->hdr) - ISCSI_DIGEST_SIZE;
1898         }
1899
1900         for (cmd_i = 0; cmd_i < session->mgmtpool_max; cmd_i++) {
1901                 struct iscsi_mgmt_task *mtask = session->mgmt_cmds[cmd_i];
1902                 struct iscsi_tcp_mgmt_task *tcp_mtask = mtask->dd_data;
1903
1904                 mtask->hdr = (struct iscsi_hdr *) &tcp_mtask->hdr;
1905         }
1906
1907         if (iscsi_r2tpool_alloc(session))
1908                 goto remove_session;
1909         return cls_session;
1910
1911 remove_session:
1912         iscsi_session_teardown(cls_session);
1913 remove_host:
1914         iscsi_host_remove(shost);
1915 free_host:
1916         iscsi_host_free(shost);
1917         return NULL;
1918 }
1919
1920 static void iscsi_tcp_session_destroy(struct iscsi_cls_session *cls_session)
1921 {
1922         struct Scsi_Host *shost = iscsi_session_to_shost(cls_session);
1923
1924         iscsi_r2tpool_free(cls_session->dd_data);
1925
1926         iscsi_host_remove(shost);
1927         iscsi_host_free(shost);
1928 }
1929
1930 static int iscsi_tcp_slave_configure(struct scsi_device *sdev)
1931 {
1932         blk_queue_bounce_limit(sdev->request_queue, BLK_BOUNCE_ANY);
1933         blk_queue_dma_alignment(sdev->request_queue, 0);
1934         return 0;
1935 }
1936
1937 static struct scsi_host_template iscsi_sht = {
1938         .module                 = THIS_MODULE,
1939         .name                   = "iSCSI Initiator over TCP/IP",
1940         .queuecommand           = iscsi_queuecommand,
1941         .change_queue_depth     = iscsi_change_queue_depth,
1942         .can_queue              = ISCSI_DEF_XMIT_CMDS_MAX - 1,
1943         .sg_tablesize           = 4096,
1944         .max_sectors            = 0xFFFF,
1945         .cmd_per_lun            = ISCSI_DEF_CMD_PER_LUN,
1946         .eh_abort_handler       = iscsi_eh_abort,
1947         .eh_device_reset_handler= iscsi_eh_device_reset,
1948         .eh_host_reset_handler  = iscsi_eh_host_reset,
1949         .use_clustering         = DISABLE_CLUSTERING,
1950         .slave_configure        = iscsi_tcp_slave_configure,
1951         .proc_name              = "iscsi_tcp",
1952         .this_id                = -1,
1953 };
1954
1955 static struct iscsi_transport iscsi_tcp_transport = {
1956         .owner                  = THIS_MODULE,
1957         .name                   = "tcp",
1958         .caps                   = CAP_RECOVERY_L0 | CAP_MULTI_R2T | CAP_HDRDGST
1959                                   | CAP_DATADGST,
1960         .param_mask             = ISCSI_MAX_RECV_DLENGTH |
1961                                   ISCSI_MAX_XMIT_DLENGTH |
1962                                   ISCSI_HDRDGST_EN |
1963                                   ISCSI_DATADGST_EN |
1964                                   ISCSI_INITIAL_R2T_EN |
1965                                   ISCSI_MAX_R2T |
1966                                   ISCSI_IMM_DATA_EN |
1967                                   ISCSI_FIRST_BURST |
1968                                   ISCSI_MAX_BURST |
1969                                   ISCSI_PDU_INORDER_EN |
1970                                   ISCSI_DATASEQ_INORDER_EN |
1971                                   ISCSI_ERL |
1972                                   ISCSI_CONN_PORT |
1973                                   ISCSI_CONN_ADDRESS |
1974                                   ISCSI_EXP_STATSN |
1975                                   ISCSI_PERSISTENT_PORT |
1976                                   ISCSI_PERSISTENT_ADDRESS |
1977                                   ISCSI_TARGET_NAME | ISCSI_TPGT |
1978                                   ISCSI_USERNAME | ISCSI_PASSWORD |
1979                                   ISCSI_USERNAME_IN | ISCSI_PASSWORD_IN |
1980                                   ISCSI_FAST_ABORT | ISCSI_ABORT_TMO |
1981                                   ISCSI_LU_RESET_TMO |
1982                                   ISCSI_PING_TMO | ISCSI_RECV_TMO,
1983         .host_param_mask        = ISCSI_HOST_HWADDRESS | ISCSI_HOST_IPADDRESS |
1984                                   ISCSI_HOST_INITIATOR_NAME |
1985                                   ISCSI_HOST_NETDEV_NAME,
1986         .conndata_size          = sizeof(struct iscsi_conn),
1987         .sessiondata_size       = sizeof(struct iscsi_session),
1988         /* session management */
1989         .create_session         = iscsi_tcp_session_create,
1990         .destroy_session        = iscsi_tcp_session_destroy,
1991         /* connection management */
1992         .create_conn            = iscsi_tcp_conn_create,
1993         .bind_conn              = iscsi_tcp_conn_bind,
1994         .destroy_conn           = iscsi_tcp_conn_destroy,
1995         .set_param              = iscsi_conn_set_param,
1996         .get_conn_param         = iscsi_tcp_conn_get_param,
1997         .get_session_param      = iscsi_session_get_param,
1998         .start_conn             = iscsi_conn_start,
1999         .stop_conn              = iscsi_tcp_conn_stop,
2000         /* iscsi host params */
2001         .get_host_param         = iscsi_host_get_param,
2002         .set_host_param         = iscsi_host_set_param,
2003         /* IO */
2004         .send_pdu               = iscsi_conn_send_pdu,
2005         .get_stats              = iscsi_conn_get_stats,
2006         .init_cmd_task          = iscsi_tcp_ctask_init,
2007         .init_mgmt_task         = iscsi_tcp_mtask_init,
2008         .xmit_cmd_task          = iscsi_tcp_ctask_xmit,
2009         .xmit_mgmt_task         = iscsi_tcp_mtask_xmit,
2010         .cleanup_cmd_task       = iscsi_tcp_cleanup_ctask,
2011         /* recovery */
2012         .session_recovery_timedout = iscsi_session_recovery_timedout,
2013 };
2014
2015 static int __init
2016 iscsi_tcp_init(void)
2017 {
2018         if (iscsi_max_lun < 1) {
2019                 printk(KERN_ERR "iscsi_tcp: Invalid max_lun value of %u\n",
2020                        iscsi_max_lun);
2021                 return -EINVAL;
2022         }
2023
2024         iscsi_tcp_scsi_transport = iscsi_register_transport(
2025                                                         &iscsi_tcp_transport);
2026         if (!iscsi_tcp_scsi_transport)
2027                 return -ENODEV;
2028
2029         return 0;
2030 }
2031
2032 static void __exit
2033 iscsi_tcp_exit(void)
2034 {
2035         iscsi_unregister_transport(&iscsi_tcp_transport);
2036 }
2037
2038 module_init(iscsi_tcp_init);
2039 module_exit(iscsi_tcp_exit);