drivers/scsi/libiscsi_tcp.c

   1 /*
   2  * iSCSI over TCP/IP Data-Path lib
   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("Mike Christie <michaelc@cs.wisc.edu>, "
  48               "Dmitry Yusupov <dmitry_yus@yahoo.com>, "
  49               "Alex Aizman <itn780@yahoo.com>");
  50 MODULE_DESCRIPTION("iSCSI/TCP data-path");
  51 MODULE_LICENSE("GPL");
  52
  53 static int iscsi_dbg_libtcp;
  54 module_param_named(debug_libiscsi_tcp, iscsi_dbg_libtcp, int,
  55                    S_IRUGO | S_IWUSR);
  56 MODULE_PARM_DESC(debug_libiscsi_tcp, "Turn on debugging for libiscsi_tcp "
  57                  "module. Set to 1 to turn on, and zero to turn off. Default "
  58                  "is off.");
  59
  60 #define ISCSI_DBG_TCP(_conn, dbg_fmt, arg...)                   \
  61         do {                                                    \
  62                 if (iscsi_dbg_libtcp)                           \
  63                         iscsi_conn_printk(KERN_INFO, _conn,     \
  64                                              "%s " dbg_fmt,     \
  65                                              __func__, ##arg);  \
  66         } while (0);
  67
  68 static int iscsi_tcp_hdr_recv_done(struct iscsi_tcp_conn *tcp_conn,
  69                                    struct iscsi_segment *segment);
  70
  71 /*
  72  * Scatterlist handling: inside the iscsi_segment, we
  73  * remember an index into the scatterlist, and set data/size
  74  * to the current scatterlist entry. For highmem pages, we
  75  * kmap as needed.
  76  *
  77  * Note that the page is unmapped when we return from
  78  * TCP's data_ready handler, so we may end up mapping and
  79  * unmapping the same page repeatedly. The whole reason
  80  * for this is that we shouldn't keep the page mapped
  81  * outside the softirq.
  82  */
  83
  84 /**
  85  * iscsi_tcp_segment_init_sg - init indicated scatterlist entry
  86  * @segment: the buffer object
  87  * @sg: scatterlist
  88  * @offset: byte offset into that sg entry
  89  *
  90  * This function sets up the segment so that subsequent
  91  * data is copied to the indicated sg entry, at the given
  92  * offset.
  93  */
  94 static inline void
  95 iscsi_tcp_segment_init_sg(struct iscsi_segment *segment,
  96                           struct scatterlist *sg, unsigned int offset)
  97 {
  98         segment->sg = sg;
  99         segment->sg_offset = offset;
 100         segment->size = min(sg->length - offset,
 101                             segment->total_size - segment->total_copied);
 102         segment->data = NULL;
 103 }
 104
 105 /**
 106  * iscsi_tcp_segment_map - map the current S/G page
 107  * @segment: iscsi_segment
 108  * @recv: 1 if called from recv path
 109  *
 110  * We only need to possibly kmap data if scatter lists are being used,
 111  * because the iscsi passthrough and internal IO paths will never use high
 112  * mem pages.
 113  */
 114 static void iscsi_tcp_segment_map(struct iscsi_segment *segment, int recv)
 115 {
 116         struct scatterlist *sg;
 117
 118         if (segment->data != NULL || !segment->sg)
 119                 return;
 120
 121         sg = segment->sg;
 122         BUG_ON(segment->sg_mapped);
 123         BUG_ON(sg->length == 0);
 124
 125         /*
 126          * If the page count is greater than one it is ok to send
 127          * to the network layer's zero copy send path. If not we
 128          * have to go the slow sendmsg path. We always map for the
 129          * recv path.
 130          */
 131         if (page_count(sg_page(sg)) >= 1 && !recv)
 132                 return;
 133
 134         segment->sg_mapped = kmap_atomic(sg_page(sg), KM_SOFTIRQ0);
 135         segment->data = segment->sg_mapped + sg->offset + segment->sg_offset;
 136 }
 137
 138 void iscsi_tcp_segment_unmap(struct iscsi_segment *segment)
 139 {
 140         if (segment->sg_mapped) {
 141                 kunmap_atomic(segment->sg_mapped, KM_SOFTIRQ0);
 142                 segment->sg_mapped = NULL;
 143                 segment->data = NULL;
 144         }
 145 }
 146 EXPORT_SYMBOL_GPL(iscsi_tcp_segment_unmap);
 147
 148 /*
 149  * Splice the digest buffer into the buffer
 150  */
 151 static inline void
 152 iscsi_tcp_segment_splice_digest(struct iscsi_segment *segment, void *digest)
 153 {
 154         segment->data = digest;
 155         segment->digest_len = ISCSI_DIGEST_SIZE;
 156         segment->total_size += ISCSI_DIGEST_SIZE;
 157         segment->size = ISCSI_DIGEST_SIZE;
 158         segment->copied = 0;
 159         segment->sg = NULL;
 160         segment->hash = NULL;
 161 }
 162
 163 /**
 164  * iscsi_tcp_segment_done - check whether the segment is complete
 165  * @tcp_conn: iscsi tcp connection
 166  * @segment: iscsi segment to check
 167  * @recv: set to one of this is called from the recv path
 168  * @copied: number of bytes copied
 169  *
 170  * Check if we're done receiving this segment. If the receive
 171  * buffer is full but we expect more data, move on to the
 172  * next entry in the scatterlist.
 173  *
 174  * If the amount of data we received isn't a multiple of 4,
 175  * we will transparently receive the pad bytes, too.
 176  *
 177  * This function must be re-entrant.
 178  */
 179 int iscsi_tcp_segment_done(struct iscsi_tcp_conn *tcp_conn,
 180                            struct iscsi_segment *segment, int recv,
 181                            unsigned copied)
 182 {
 183         struct scatterlist sg;
 184         unsigned int pad;
 185
 186         ISCSI_DBG_TCP(tcp_conn->iscsi_conn, "copied %u %u size %u %s\n",
 187                       segment->copied, copied, segment->size,
 188                       recv ? "recv" : "xmit");
 189         if (segment->hash && copied) {
 190                 /*
 191                  * If a segment is kmapd we must unmap it before sending
 192                  * to the crypto layer since that will try to kmap it again.
 193                  */
 194                 iscsi_tcp_segment_unmap(segment);
 195
 196                 if (!segment->data) {
 197                         sg_init_table(&sg, 1);
 198                         sg_set_page(&sg, sg_page(segment->sg), copied,
 199                                     segment->copied + segment->sg_offset +
 200                                                         segment->sg->offset);
 201                 } else
 202                         sg_init_one(&sg, segment->data + segment->copied,
 203                                     copied);
 204                 crypto_hash_update(segment->hash, &sg, copied);
 205         }
 206
 207         segment->copied += copied;
 208         if (segment->copied < segment->size) {
 209                 iscsi_tcp_segment_map(segment, recv);
 210                 return 0;
 211         }
 212
 213         segment->total_copied += segment->copied;
 214         segment->copied = 0;
 215         segment->size = 0;
 216
 217         /* Unmap the current scatterlist page, if there is one. */
 218         iscsi_tcp_segment_unmap(segment);
 219
 220         /* Do we have more scatterlist entries? */
 221         ISCSI_DBG_TCP(tcp_conn->iscsi_conn, "total copied %u total size %u\n",
 222                       segment->total_copied, segment->total_size);
 223         if (segment->total_copied < segment->total_size) {
 224                 /* Proceed to the next entry in the scatterlist. */
 225                 iscsi_tcp_segment_init_sg(segment, sg_next(segment->sg),
 226                                           0);
 227                 iscsi_tcp_segment_map(segment, recv);
 228                 BUG_ON(segment->size == 0);
 229                 return 0;
 230         }
 231
 232         /* Do we need to handle padding? */
 233         if (!(tcp_conn->iscsi_conn->session->tt->caps & CAP_PADDING_OFFLOAD)) {
 234                 pad = iscsi_padding(segment->total_copied);
 235                 if (pad != 0) {
 236                         ISCSI_DBG_TCP(tcp_conn->iscsi_conn,
 237                                       "consume %d pad bytes\n", pad);
 238                         segment->total_size += pad;
 239                         segment->size = pad;
 240                         segment->data = segment->padbuf;
 241                         return 0;
 242                 }
 243         }
 244
 245         /*
 246          * Set us up for transferring the data digest. hdr digest
 247          * is completely handled in hdr done function.
 248          */
 249         if (segment->hash) {
 250                 crypto_hash_final(segment->hash, segment->digest);
 251                 iscsi_tcp_segment_splice_digest(segment,
 252                                  recv ? segment->recv_digest : segment->digest);
 253                 return 0;
 254         }
 255
 256         return 1;
 257 }
 258 EXPORT_SYMBOL_GPL(iscsi_tcp_segment_done);
 259
 260 /**
 261  * iscsi_tcp_segment_recv - copy data to segment
 262  * @tcp_conn: the iSCSI TCP connection
 263  * @segment: the buffer to copy to
 264  * @ptr: data pointer
 265  * @len: amount of data available
 266  *
 267  * This function copies up to @len bytes to the
 268  * given buffer, and returns the number of bytes
 269  * consumed, which can actually be less than @len.
 270  *
 271  * If hash digest is enabled, the function will update the
 272  * hash while copying.
 273  * Combining these two operations doesn't buy us a lot (yet),
 274  * but in the future we could implement combined copy+crc,
 275  * just way we do for network layer checksums.
 276  */
 277 static int
 278 iscsi_tcp_segment_recv(struct iscsi_tcp_conn *tcp_conn,
 279                        struct iscsi_segment *segment, const void *ptr,
 280                        unsigned int len)
 281 {
 282         unsigned int copy = 0, copied = 0;
 283
 284         while (!iscsi_tcp_segment_done(tcp_conn, segment, 1, copy)) {
 285                 if (copied == len) {
 286                         ISCSI_DBG_TCP(tcp_conn->iscsi_conn,
 287                                       "copied %d bytes\n", len);
 288                         break;
 289                 }
 290
 291                 copy = min(len - copied, segment->size - segment->copied);
 292                 ISCSI_DBG_TCP(tcp_conn->iscsi_conn, "copying %d\n", copy);
 293                 memcpy(segment->data + segment->copied, ptr + copied, copy);
 294                 copied += copy;
 295         }
 296         return copied;
 297 }
 298
 299 inline void
 300 iscsi_tcp_dgst_header(struct hash_desc *hash, const void *hdr, size_t hdrlen,
 301                       unsigned char digest[ISCSI_DIGEST_SIZE])
 302 {
 303         struct scatterlist sg;
 304
 305         sg_init_one(&sg, hdr, hdrlen);
 306         crypto_hash_digest(hash, &sg, hdrlen, digest);
 307 }
 308 EXPORT_SYMBOL_GPL(iscsi_tcp_dgst_header);
 309
 310 static inline int
 311 iscsi_tcp_dgst_verify(struct iscsi_tcp_conn *tcp_conn,
 312                       struct iscsi_segment *segment)
 313 {
 314         if (!segment->digest_len)
 315                 return 1;
 316
 317         if (memcmp(segment->recv_digest, segment->digest,
 318                    segment->digest_len)) {
 319                 ISCSI_DBG_TCP(tcp_conn->iscsi_conn, "digest mismatch\n");
 320                 return 0;
 321         }
 322
 323         return 1;
 324 }
 325
 326 /*
 327  * Helper function to set up segment buffer
 328  */
 329 static inline void
 330 __iscsi_segment_init(struct iscsi_segment *segment, size_t size,
 331                      iscsi_segment_done_fn_t *done, struct hash_desc *hash)
 332 {
 333         memset(segment, 0, sizeof(*segment));
 334         segment->total_size = size;
 335         segment->done = done;
 336
 337         if (hash) {
 338                 segment->hash = hash;
 339                 crypto_hash_init(hash);
 340         }
 341 }
 342
 343 inline void
 344 iscsi_segment_init_linear(struct iscsi_segment *segment, void *data,
 345                           size_t size, iscsi_segment_done_fn_t *done,
 346                           struct hash_desc *hash)
 347 {
 348         __iscsi_segment_init(segment, size, done, hash);
 349         segment->data = data;
 350         segment->size = size;
 351 }
 352 EXPORT_SYMBOL_GPL(iscsi_segment_init_linear);
 353
 354 inline int
 355 iscsi_segment_seek_sg(struct iscsi_segment *segment,
 356                       struct scatterlist *sg_list, unsigned int sg_count,
 357                       unsigned int offset, size_t size,
 358                       iscsi_segment_done_fn_t *done, struct hash_desc *hash)
 359 {
 360         struct scatterlist *sg;
 361         unsigned int i;
 362
 363         __iscsi_segment_init(segment, size, done, hash);
 364         for_each_sg(sg_list, sg, sg_count, i) {
 365                 if (offset < sg->length) {
 366                         iscsi_tcp_segment_init_sg(segment, sg, offset);
 367                         return 0;
 368                 }
 369                 offset -= sg->length;
 370         }
 371
 372         return ISCSI_ERR_DATA_OFFSET;
 373 }
 374 EXPORT_SYMBOL_GPL(iscsi_segment_seek_sg);
 375
 376 /**
 377  * iscsi_tcp_hdr_recv_prep - prep segment for hdr reception
 378  * @tcp_conn: iscsi connection to prep for
 379  *
 380  * This function always passes NULL for the hash argument, because when this
 381  * function is called we do not yet know the final size of the header and want
 382  * to delay the digest processing until we know that.
 383  */
 384 void iscsi_tcp_hdr_recv_prep(struct iscsi_tcp_conn *tcp_conn)
 385 {
 386         ISCSI_DBG_TCP(tcp_conn->iscsi_conn,
 387                       "(%s)\n", tcp_conn->iscsi_conn->hdrdgst_en ?
 388                       "digest enabled" : "digest disabled");
 389         iscsi_segment_init_linear(&tcp_conn->in.segment,
 390                                 tcp_conn->in.hdr_buf, sizeof(struct iscsi_hdr),
 391                                 iscsi_tcp_hdr_recv_done, NULL);
 392 }
 393 EXPORT_SYMBOL_GPL(iscsi_tcp_hdr_recv_prep);
 394
 395 /*
 396  * Handle incoming reply to any other type of command
 397  */
 398 static int
 399 iscsi_tcp_data_recv_done(struct iscsi_tcp_conn *tcp_conn,
 400                          struct iscsi_segment *segment)
 401 {
 402         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 403         int rc = 0;
 404
 405         if (!iscsi_tcp_dgst_verify(tcp_conn, segment))
 406                 return ISCSI_ERR_DATA_DGST;
 407
 408         rc = iscsi_complete_pdu(conn, tcp_conn->in.hdr,
 409                         conn->data, tcp_conn->in.datalen);
 410         if (rc)
 411                 return rc;
 412
 413         iscsi_tcp_hdr_recv_prep(tcp_conn);
 414         return 0;
 415 }
 416
 417 static void
 418 iscsi_tcp_data_recv_prep(struct iscsi_tcp_conn *tcp_conn)
 419 {
 420         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 421         struct hash_desc *rx_hash = NULL;
 422
 423         if (conn->datadgst_en &
 424             !(conn->session->tt->caps & CAP_DIGEST_OFFLOAD))
 425                 rx_hash = tcp_conn->rx_hash;
 426
 427         iscsi_segment_init_linear(&tcp_conn->in.segment,
 428                                 conn->data, tcp_conn->in.datalen,
 429                                 iscsi_tcp_data_recv_done, rx_hash);
 430 }
 431
 432 /**
 433  * iscsi_tcp_cleanup_task - free tcp_task resources
 434  * @task: iscsi task
 435  *
 436  * must be called with session lock
 437  */
 438 void iscsi_tcp_cleanup_task(struct iscsi_task *task)
 439 {
 440         struct iscsi_tcp_task *tcp_task = task->dd_data;
 441         struct iscsi_r2t_info *r2t;
 442
 443         /* nothing to do for mgmt */
 444         if (!task->sc)
 445                 return;
 446
 447         /* flush task's r2t queues */
 448         while (__kfifo_get(tcp_task->r2tqueue, (void*)&r2t, sizeof(void*))) {
 449                 __kfifo_put(tcp_task->r2tpool.queue, (void*)&r2t,
 450                             sizeof(void*));
 451                 ISCSI_DBG_TCP(task->conn, "pending r2t dropped\n");
 452         }
 453
 454         r2t = tcp_task->r2t;
 455         if (r2t != NULL) {
 456                 __kfifo_put(tcp_task->r2tpool.queue, (void*)&r2t,
 457                             sizeof(void*));
 458                 tcp_task->r2t = NULL;
 459         }
 460 }
 461 EXPORT_SYMBOL_GPL(iscsi_tcp_cleanup_task);
 462
 463 /**
 464  * iscsi_tcp_data_in - SCSI Data-In Response processing
 465  * @conn: iscsi connection
 466  * @task: scsi command task
 467  */
 468 static int iscsi_tcp_data_in(struct iscsi_conn *conn, struct iscsi_task *task)
 469 {
 470         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 471         struct iscsi_tcp_task *tcp_task = task->dd_data;
 472         struct iscsi_data_rsp *rhdr = (struct iscsi_data_rsp *)tcp_conn->in.hdr;
 473         int datasn = be32_to_cpu(rhdr->datasn);
 474         unsigned total_in_length = scsi_in(task->sc)->length;
 475
 476         /*
 477          * lib iscsi will update this in the completion handling if there
 478          * is status.
 479          */
 480         if (!(rhdr->flags & ISCSI_FLAG_DATA_STATUS))
 481                 iscsi_update_cmdsn(conn->session, (struct iscsi_nopin*)rhdr);
 482
 483         if (tcp_conn->in.datalen == 0)
 484                 return 0;
 485
 486         if (tcp_task->exp_datasn != datasn) {
 487                 ISCSI_DBG_TCP(conn, "task->exp_datasn(%d) != rhdr->datasn(%d)"
 488                               "\n", tcp_task->exp_datasn, datasn);
 489                 return ISCSI_ERR_DATASN;
 490         }
 491
 492         tcp_task->exp_datasn++;
 493
 494         tcp_task->data_offset = be32_to_cpu(rhdr->offset);
 495         if (tcp_task->data_offset + tcp_conn->in.datalen > total_in_length) {
 496                 ISCSI_DBG_TCP(conn, "data_offset(%d) + data_len(%d) > "
 497                               "total_length_in(%d)\n", tcp_task->data_offset,
 498                               tcp_conn->in.datalen, total_in_length);
 499                 return ISCSI_ERR_DATA_OFFSET;
 500         }
 501
 502         conn->datain_pdus_cnt++;
 503         return 0;
 504 }
 505
 506 /**
 507  * iscsi_tcp_r2t_rsp - iSCSI R2T Response processing
 508  * @conn: iscsi connection
 509  * @task: scsi command task
 510  */
 511 static int iscsi_tcp_r2t_rsp(struct iscsi_conn *conn, struct iscsi_task *task)
 512 {
 513         struct iscsi_session *session = conn->session;
 514         struct iscsi_tcp_task *tcp_task = task->dd_data;
 515         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 516         struct iscsi_r2t_rsp *rhdr = (struct iscsi_r2t_rsp *)tcp_conn->in.hdr;
 517         struct iscsi_r2t_info *r2t;
 518         int r2tsn = be32_to_cpu(rhdr->r2tsn);
 519         int rc;
 520
 521         if (tcp_conn->in.datalen) {
 522                 iscsi_conn_printk(KERN_ERR, conn,
 523                                   "invalid R2t with datalen %d\n",
 524                                   tcp_conn->in.datalen);
 525                 return ISCSI_ERR_DATALEN;
 526         }
 527
 528         if (tcp_task->exp_datasn != r2tsn){
 529                 ISCSI_DBG_TCP(conn, "task->exp_datasn(%d) != rhdr->r2tsn(%d)\n",
 530                               tcp_task->exp_datasn, r2tsn);
 531                 return ISCSI_ERR_R2TSN;
 532         }
 533
 534         /* fill-in new R2T associated with the task */
 535         iscsi_update_cmdsn(session, (struct iscsi_nopin*)rhdr);
 536
 537         if (!task->sc || session->state != ISCSI_STATE_LOGGED_IN) {
 538                 iscsi_conn_printk(KERN_INFO, conn,
 539                                   "dropping R2T itt %d in recovery.\n",
 540                                   task->itt);
 541                 return 0;
 542         }
 543
 544         rc = __kfifo_get(tcp_task->r2tpool.queue, (void*)&r2t, sizeof(void*));
 545         if (!rc) {
 546                 iscsi_conn_printk(KERN_ERR, conn, "Could not allocate R2T. "
 547                                   "Target has sent more R2Ts than it "
 548                                   "negotiated for or driver has has leaked.\n");
 549                 return ISCSI_ERR_PROTO;
 550         }
 551
 552         r2t->exp_statsn = rhdr->statsn;
 553         r2t->data_length = be32_to_cpu(rhdr->data_length);
 554         if (r2t->data_length == 0) {
 555                 iscsi_conn_printk(KERN_ERR, conn,
 556                                   "invalid R2T with zero data len\n");
 557                 __kfifo_put(tcp_task->r2tpool.queue, (void*)&r2t,
 558                             sizeof(void*));
 559                 return ISCSI_ERR_DATALEN;
 560         }
 561
 562         if (r2t->data_length > session->max_burst)
 563                 ISCSI_DBG_TCP(conn, "invalid R2T with data len %u and max "
 564                               "burst %u. Attempting to execute request.\n",
 565                               r2t->data_length, session->max_burst);
 566
 567         r2t->data_offset = be32_to_cpu(rhdr->data_offset);
 568         if (r2t->data_offset + r2t->data_length > scsi_out(task->sc)->length) {
 569                 iscsi_conn_printk(KERN_ERR, conn,
 570                                   "invalid R2T with data len %u at offset %u "
 571                                   "and total length %d\n", r2t->data_length,
 572                                   r2t->data_offset, scsi_out(task->sc)->length);
 573                 __kfifo_put(tcp_task->r2tpool.queue, (void*)&r2t,
 574                             sizeof(void*));
 575                 return ISCSI_ERR_DATALEN;
 576         }
 577
 578         r2t->ttt = rhdr->ttt; /* no flip */
 579         r2t->datasn = 0;
 580         r2t->sent = 0;
 581
 582         tcp_task->exp_datasn = r2tsn + 1;
 583         __kfifo_put(tcp_task->r2tqueue, (void*)&r2t, sizeof(void*));
 584         conn->r2t_pdus_cnt++;
 585
 586         iscsi_requeue_task(task);
 587         return 0;
 588 }
 589
 590 /*
 591  * Handle incoming reply to DataIn command
 592  */
 593 static int
 594 iscsi_tcp_process_data_in(struct iscsi_tcp_conn *tcp_conn,
 595                           struct iscsi_segment *segment)
 596 {
 597         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 598         struct iscsi_hdr *hdr = tcp_conn->in.hdr;
 599         int rc;
 600
 601         if (!iscsi_tcp_dgst_verify(tcp_conn, segment))
 602                 return ISCSI_ERR_DATA_DGST;
 603
 604         /* check for non-exceptional status */
 605         if (hdr->flags & ISCSI_FLAG_DATA_STATUS) {
 606                 rc = iscsi_complete_pdu(conn, tcp_conn->in.hdr, NULL, 0);
 607                 if (rc)
 608                         return rc;
 609         }
 610
 611         iscsi_tcp_hdr_recv_prep(tcp_conn);
 612         return 0;
 613 }
 614
 615 /**
 616  * iscsi_tcp_hdr_dissect - process PDU header
 617  * @conn: iSCSI connection
 618  * @hdr: PDU header
 619  *
 620  * This function analyzes the header of the PDU received,
 621  * and performs several sanity checks. If the PDU is accompanied
 622  * by data, the receive buffer is set up to copy the incoming data
 623  * to the correct location.
 624  */
 625 static int
 626 iscsi_tcp_hdr_dissect(struct iscsi_conn *conn, struct iscsi_hdr *hdr)
 627 {
 628         int rc = 0, opcode, ahslen;
 629         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 630         struct iscsi_task *task;
 631
 632         /* verify PDU length */
 633         tcp_conn->in.datalen = ntoh24(hdr->dlength);
 634         if (tcp_conn->in.datalen > conn->max_recv_dlength) {
 635                 iscsi_conn_printk(KERN_ERR, conn,
 636                                   "iscsi_tcp: datalen %d > %d\n",
 637                                   tcp_conn->in.datalen, conn->max_recv_dlength);
 638                 return ISCSI_ERR_DATALEN;
 639         }
 640
 641         /* Additional header segments. So far, we don't
 642          * process additional headers.
 643          */
 644         ahslen = hdr->hlength << 2;
 645
 646         opcode = hdr->opcode & ISCSI_OPCODE_MASK;
 647         /* verify itt (itt encoding: age+cid+itt) */
 648         rc = iscsi_verify_itt(conn, hdr->itt);
 649         if (rc)
 650                 return rc;
 651
 652         ISCSI_DBG_TCP(conn, "opcode 0x%x ahslen %d datalen %d\n",
 653                       opcode, ahslen, tcp_conn->in.datalen);
 654
 655         switch(opcode) {
 656         case ISCSI_OP_SCSI_DATA_IN:
 657                 spin_lock(&conn->session->lock);
 658                 task = iscsi_itt_to_ctask(conn, hdr->itt);
 659                 if (!task)
 660                         rc = ISCSI_ERR_BAD_ITT;
 661                 else
 662                         rc = iscsi_tcp_data_in(conn, task);
 663                 if (rc) {
 664                         spin_unlock(&conn->session->lock);
 665                         break;
 666                 }
 667
 668                 if (tcp_conn->in.datalen) {
 669                         struct iscsi_tcp_task *tcp_task = task->dd_data;
 670                         struct hash_desc *rx_hash = NULL;
 671                         struct scsi_data_buffer *sdb = scsi_in(task->sc);
 672
 673                         /*
 674                          * Setup copy of Data-In into the Scsi_Cmnd
 675                          * Scatterlist case:
 676                          * We set up the iscsi_segment to point to the next
 677                          * scatterlist entry to copy to. As we go along,
 678                          * we move on to the next scatterlist entry and
 679                          * update the digest per-entry.
 680                          */
 681                         if (conn->datadgst_en &&
 682                             !(conn->session->tt->caps & CAP_DIGEST_OFFLOAD))
 683                                 rx_hash = tcp_conn->rx_hash;
 684
 685                         ISCSI_DBG_TCP(conn, "iscsi_tcp_begin_data_in( "
 686                                      "offset=%d, datalen=%d)\n",
 687                                       tcp_task->data_offset,
 688                                       tcp_conn->in.datalen);
 689                         task->last_xfer = jiffies;
 690                         rc = iscsi_segment_seek_sg(&tcp_conn->in.segment,
 691                                                    sdb->table.sgl,
 692                                                    sdb->table.nents,
 693                                                    tcp_task->data_offset,
 694                                                    tcp_conn->in.datalen,
 695                                                    iscsi_tcp_process_data_in,
 696                                                    rx_hash);
 697                         spin_unlock(&conn->session->lock);
 698                         return rc;
 699                 }
 700                 rc = __iscsi_complete_pdu(conn, hdr, NULL, 0);
 701                 spin_unlock(&conn->session->lock);
 702                 break;
 703         case ISCSI_OP_SCSI_CMD_RSP:
 704                 if (tcp_conn->in.datalen) {
 705                         iscsi_tcp_data_recv_prep(tcp_conn);
 706                         return 0;
 707                 }
 708                 rc = iscsi_complete_pdu(conn, hdr, NULL, 0);
 709                 break;
 710         case ISCSI_OP_R2T:
 711                 spin_lock(&conn->session->lock);
 712                 task = iscsi_itt_to_ctask(conn, hdr->itt);
 713                 if (!task)
 714                         rc = ISCSI_ERR_BAD_ITT;
 715                 else if (ahslen)
 716                         rc = ISCSI_ERR_AHSLEN;
 717                 else if (task->sc->sc_data_direction == DMA_TO_DEVICE) {
 718                         task->last_xfer = jiffies;
 719                         rc = iscsi_tcp_r2t_rsp(conn, task);
 720                 } else
 721                         rc = ISCSI_ERR_PROTO;
 722                 spin_unlock(&conn->session->lock);
 723                 break;
 724         case ISCSI_OP_LOGIN_RSP:
 725         case ISCSI_OP_TEXT_RSP:
 726         case ISCSI_OP_REJECT:
 727         case ISCSI_OP_ASYNC_EVENT:
 728                 /*
 729                  * It is possible that we could get a PDU with a buffer larger
 730                  * than 8K, but there are no targets that currently do this.
 731                  * For now we fail until we find a vendor that needs it
 732                  */
 733                 if (ISCSI_DEF_MAX_RECV_SEG_LEN < tcp_conn->in.datalen) {
 734                         iscsi_conn_printk(KERN_ERR, conn,
 735                                           "iscsi_tcp: received buffer of "
 736                                           "len %u but conn buffer is only %u "
 737                                           "(opcode %0x)\n",
 738                                           tcp_conn->in.datalen,
 739                                           ISCSI_DEF_MAX_RECV_SEG_LEN, opcode);
 740                         rc = ISCSI_ERR_PROTO;
 741                         break;
 742                 }
 743
 744                 /* If there's data coming in with the response,
 745                  * receive it to the connection's buffer.
 746                  */
 747                 if (tcp_conn->in.datalen) {
 748                         iscsi_tcp_data_recv_prep(tcp_conn);
 749                         return 0;
 750                 }
 751         /* fall through */
 752         case ISCSI_OP_LOGOUT_RSP:
 753         case ISCSI_OP_NOOP_IN:
 754         case ISCSI_OP_SCSI_TMFUNC_RSP:
 755                 rc = iscsi_complete_pdu(conn, hdr, NULL, 0);
 756                 break;
 757         default:
 758                 rc = ISCSI_ERR_BAD_OPCODE;
 759                 break;
 760         }
 761
 762         if (rc == 0) {
 763                 /* Anything that comes with data should have
 764                  * been handled above. */
 765                 if (tcp_conn->in.datalen)
 766                         return ISCSI_ERR_PROTO;
 767                 iscsi_tcp_hdr_recv_prep(tcp_conn);
 768         }
 769
 770         return rc;
 771 }
 772
 773 /**
 774  * iscsi_tcp_hdr_recv_done - process PDU header
 775  *
 776  * This is the callback invoked when the PDU header has
 777  * been received. If the header is followed by additional
 778  * header segments, we go back for more data.
 779  */
 780 static int
 781 iscsi_tcp_hdr_recv_done(struct iscsi_tcp_conn *tcp_conn,
 782                         struct iscsi_segment *segment)
 783 {
 784         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 785         struct iscsi_hdr *hdr;
 786
 787         /* Check if there are additional header segments
 788          * *prior* to computing the digest, because we
 789          * may need to go back to the caller for more.
 790          */
 791         hdr = (struct iscsi_hdr *) tcp_conn->in.hdr_buf;
 792         if (segment->copied == sizeof(struct iscsi_hdr) && hdr->hlength) {
 793                 /* Bump the header length - the caller will
 794                  * just loop around and get the AHS for us, and
 795                  * call again. */
 796                 unsigned int ahslen = hdr->hlength << 2;
 797
 798                 /* Make sure we don't overflow */
 799                 if (sizeof(*hdr) + ahslen > sizeof(tcp_conn->in.hdr_buf))
 800                         return ISCSI_ERR_AHSLEN;
 801
 802                 segment->total_size += ahslen;
 803                 segment->size += ahslen;
 804                 return 0;
 805         }
 806
 807         /* We're done processing the header. See if we're doing
 808          * header digests; if so, set up the recv_digest buffer
 809          * and go back for more. */
 810         if (conn->hdrdgst_en &&
 811             !(conn->session->tt->caps & CAP_DIGEST_OFFLOAD)) {
 812                 if (segment->digest_len == 0) {
 813                         /*
 814                          * Even if we offload the digest processing we
 815                          * splice it in so we can increment the skb/segment
 816                          * counters in preparation for the data segment.
 817                          */
 818                         iscsi_tcp_segment_splice_digest(segment,
 819                                                         segment->recv_digest);
 820                         return 0;
 821                 }
 822
 823                 iscsi_tcp_dgst_header(tcp_conn->rx_hash, hdr,
 824                                       segment->total_copied - ISCSI_DIGEST_SIZE,
 825                                       segment->digest);
 826
 827                 if (!iscsi_tcp_dgst_verify(tcp_conn, segment))
 828                         return ISCSI_ERR_HDR_DGST;
 829         }
 830
 831         tcp_conn->in.hdr = hdr;
 832         return iscsi_tcp_hdr_dissect(conn, hdr);
 833 }
 834
 835 /**
 836  * iscsi_tcp_recv_segment_is_hdr - tests if we are reading in a header
 837  * @tcp_conn: iscsi tcp conn
 838  *
 839  * returns non zero if we are currently processing or setup to process
 840  * a header.
 841  */
 842 inline int iscsi_tcp_recv_segment_is_hdr(struct iscsi_tcp_conn *tcp_conn)
 843 {
 844         return tcp_conn->in.segment.done == iscsi_tcp_hdr_recv_done;
 845 }
 846 EXPORT_SYMBOL_GPL(iscsi_tcp_recv_segment_is_hdr);
 847
 848 /**
 849  * iscsi_tcp_recv_skb - Process skb
 850  * @conn: iscsi connection
 851  * @skb: network buffer with header and/or data segment
 852  * @offset: offset in skb
 853  * @offload: bool indicating if transfer was offloaded
 854  *
 855  * Will return status of transfer in status. And will return
 856  * number of bytes copied.
 857  */
 858 int iscsi_tcp_recv_skb(struct iscsi_conn *conn, struct sk_buff *skb,
 859                        unsigned int offset, bool offloaded, int *status)
 860 {
 861         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 862         struct iscsi_segment *segment = &tcp_conn->in.segment;
 863         struct skb_seq_state seq;
 864         unsigned int consumed = 0;
 865         int rc = 0;
 866
 867         ISCSI_DBG_TCP(conn, "in %d bytes\n", skb->len - offset);
 868         /*
 869          * Update for each skb instead of pdu, because over slow networks a
 870          * data_in's data could take a while to read in. We also want to
 871          * account for r2ts.
 872          */
 873         conn->last_recv = jiffies;
 874
 875         if (unlikely(conn->suspend_rx)) {
 876                 ISCSI_DBG_TCP(conn, "Rx suspended!\n");
 877                 *status = ISCSI_TCP_SUSPENDED;
 878                 return 0;
 879         }
 880
 881         if (offloaded) {
 882                 segment->total_copied = segment->total_size;
 883                 goto segment_done;
 884         }
 885
 886         skb_prepare_seq_read(skb, offset, skb->len, &seq);
 887         while (1) {
 888                 unsigned int avail;
 889                 const u8 *ptr;
 890
 891                 avail = skb_seq_read(consumed, &ptr, &seq);
 892                 if (avail == 0) {
 893                         ISCSI_DBG_TCP(conn, "no more data avail. Consumed %d\n",
 894                                       consumed);
 895                         *status = ISCSI_TCP_SKB_DONE;
 896                         skb_abort_seq_read(&seq);
 897                         goto skb_done;
 898                 }
 899                 BUG_ON(segment->copied >= segment->size);
 900
 901                 ISCSI_DBG_TCP(conn, "skb %p ptr=%p avail=%u\n", skb, ptr,
 902                               avail);
 903                 rc = iscsi_tcp_segment_recv(tcp_conn, segment, ptr, avail);
 904                 BUG_ON(rc == 0);
 905                 consumed += rc;
 906
 907                 if (segment->total_copied >= segment->total_size) {
 908                         skb_abort_seq_read(&seq);
 909                         goto segment_done;
 910                 }
 911         }
 912
 913 segment_done:
 914         *status = ISCSI_TCP_SEGMENT_DONE;
 915         ISCSI_DBG_TCP(conn, "segment done\n");
 916         rc = segment->done(tcp_conn, segment);
 917         if (rc != 0) {
 918                 *status = ISCSI_TCP_CONN_ERR;
 919                 ISCSI_DBG_TCP(conn, "Error receiving PDU, errno=%d\n", rc);
 920                 iscsi_conn_failure(conn, rc);
 921                 return 0;
 922         }
 923         /* The done() functions sets up the next segment. */
 924
 925 skb_done:
 926         conn->rxdata_octets += consumed;
 927         return consumed;
 928 }
 929 EXPORT_SYMBOL_GPL(iscsi_tcp_recv_skb);
 930
 931 /**
 932  * iscsi_tcp_task_init - Initialize iSCSI SCSI_READ or SCSI_WRITE commands
 933  * @conn: iscsi connection
 934  * @task: scsi command task
 935  * @sc: scsi command
 936  */
 937 int iscsi_tcp_task_init(struct iscsi_task *task)
 938 {
 939         struct iscsi_tcp_task *tcp_task = task->dd_data;
 940         struct iscsi_conn *conn = task->conn;
 941         struct scsi_cmnd *sc = task->sc;
 942         int err;
 943
 944         if (!sc) {
 945                 /*
 946                  * mgmt tasks do not have a scatterlist since they come
 947                  * in from the iscsi interface.
 948                  */
 949                 ISCSI_DBG_TCP(conn, "mtask deq [itt 0x%x]\n", task->itt);
 950
 951                 return conn->session->tt->init_pdu(task, 0, task->data_count);
 952         }
 953
 954         BUG_ON(__kfifo_len(tcp_task->r2tqueue));
 955         tcp_task->exp_datasn = 0;
 956
 957         /* Prepare PDU, optionally w/ immediate data */
 958         ISCSI_DBG_TCP(conn, "task deq [itt 0x%x imm %d unsol %d]\n",
 959                       task->itt, task->imm_count, task->unsol_r2t.data_length);
 960
 961         err = conn->session->tt->init_pdu(task, 0, task->imm_count);
 962         if (err)
 963                 return err;
 964         task->imm_count = 0;
 965         return 0;
 966 }
 967 EXPORT_SYMBOL_GPL(iscsi_tcp_task_init);
 968
 969 static struct iscsi_r2t_info *iscsi_tcp_get_curr_r2t(struct iscsi_task *task)
 970 {
 971         struct iscsi_session *session = task->conn->session;
 972         struct iscsi_tcp_task *tcp_task = task->dd_data;
 973         struct iscsi_r2t_info *r2t = NULL;
 974
 975         if (iscsi_task_has_unsol_data(task))
 976                 r2t = &task->unsol_r2t;
 977         else {
 978                 spin_lock_bh(&session->lock);
 979                 if (tcp_task->r2t) {
 980                         r2t = tcp_task->r2t;
 981                         /* Continue with this R2T? */
 982                         if (r2t->data_length <= r2t->sent) {
 983                                 ISCSI_DBG_TCP(task->conn,
 984                                               "  done with r2t %p\n", r2t);
 985                                 __kfifo_put(tcp_task->r2tpool.queue,
 986                                             (void *)&tcp_task->r2t,
 987                                             sizeof(void *));
 988                                 tcp_task->r2t = r2t = NULL;
 989                         }
 990                 }
 991
 992                 if (r2t == NULL) {
 993                         __kfifo_get(tcp_task->r2tqueue,
 994                                     (void *)&tcp_task->r2t, sizeof(void *));
 995                         r2t = tcp_task->r2t;
 996                 }
 997                 spin_unlock_bh(&session->lock);
 998         }
 999
1000         return r2t;
1001 }
1002
1003 /**
1004  * iscsi_tcp_task_xmit - xmit normal PDU task
1005  * @task: iscsi command task
1006  *
1007  * We're expected to return 0 when everything was transmitted succesfully,
1008  * -EAGAIN if there's still data in the queue, or != 0 for any other kind
1009  * of error.
1010  */
1011 int iscsi_tcp_task_xmit(struct iscsi_task *task)
1012 {
1013         struct iscsi_conn *conn = task->conn;
1014         struct iscsi_session *session = conn->session;
1015         struct iscsi_r2t_info *r2t;
1016         int rc = 0;
1017
1018 flush:
1019         /* Flush any pending data first. */
1020         rc = session->tt->xmit_pdu(task);
1021         if (rc < 0)
1022                 return rc;
1023
1024         /* mgmt command */
1025         if (!task->sc) {
1026                 if (task->hdr->itt == RESERVED_ITT)
1027                         iscsi_put_task(task);
1028                 return 0;
1029         }
1030
1031         /* Are we done already? */
1032         if (task->sc->sc_data_direction != DMA_TO_DEVICE)
1033                 return 0;
1034
1035         r2t = iscsi_tcp_get_curr_r2t(task);
1036         if (r2t == NULL) {
1037                 /* Waiting for more R2Ts to arrive. */
1038                 ISCSI_DBG_TCP(conn, "no R2Ts yet\n");
1039                 return 0;
1040         }
1041
1042         rc = conn->session->tt->alloc_pdu(task, ISCSI_OP_SCSI_DATA_OUT);
1043         if (rc)
1044                 return rc;
1045         iscsi_prep_data_out_pdu(task, r2t, (struct iscsi_data *) task->hdr);
1046
1047         ISCSI_DBG_TCP(conn, "sol dout %p [dsn %d itt 0x%x doff %d dlen %d]\n",
1048                       r2t, r2t->datasn - 1, task->hdr->itt,
1049                       r2t->data_offset + r2t->sent, r2t->data_count);
1050
1051         rc = conn->session->tt->init_pdu(task, r2t->data_offset + r2t->sent,
1052                                          r2t->data_count);
1053         if (rc) {
1054                 iscsi_conn_failure(conn, ISCSI_ERR_XMIT_FAILED);
1055                 return rc;
1056         }
1057
1058         r2t->sent += r2t->data_count;
1059         goto flush;
1060 }
1061 EXPORT_SYMBOL_GPL(iscsi_tcp_task_xmit);
1062
1063 struct iscsi_cls_conn *
1064 iscsi_tcp_conn_setup(struct iscsi_cls_session *cls_session, int dd_data_size,
1065                       uint32_t conn_idx)
1066
1067 {
1068         struct iscsi_conn *conn;
1069         struct iscsi_cls_conn *cls_conn;
1070         struct iscsi_tcp_conn *tcp_conn;
1071
1072         cls_conn = iscsi_conn_setup(cls_session, sizeof(*tcp_conn), conn_idx);
1073         if (!cls_conn)
1074                 return NULL;
1075         conn = cls_conn->dd_data;
1076         /*
1077          * due to strange issues with iser these are not set
1078          * in iscsi_conn_setup
1079          */
1080         conn->max_recv_dlength = ISCSI_DEF_MAX_RECV_SEG_LEN;
1081
1082         tcp_conn = conn->dd_data;
1083         tcp_conn->iscsi_conn = conn;
1084
1085         tcp_conn->dd_data = kzalloc(dd_data_size, GFP_KERNEL);
1086         if (!tcp_conn->dd_data) {
1087                 iscsi_conn_teardown(cls_conn);
1088                 return NULL;
1089         }
1090         return cls_conn;
1091 }
1092 EXPORT_SYMBOL_GPL(iscsi_tcp_conn_setup);
1093
1094 void iscsi_tcp_conn_teardown(struct iscsi_cls_conn *cls_conn)
1095 {
1096         struct iscsi_conn *conn = cls_conn->dd_data;
1097         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1098
1099         kfree(tcp_conn->dd_data);
1100         iscsi_conn_teardown(cls_conn);
1101 }
1102 EXPORT_SYMBOL_GPL(iscsi_tcp_conn_teardown);
1103
1104 int iscsi_tcp_r2tpool_alloc(struct iscsi_session *session)
1105 {
1106         int i;
1107         int cmd_i;
1108
1109         /*
1110          * initialize per-task: R2T pool and xmit queue
1111          */
1112         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
1113                 struct iscsi_task *task = session->cmds[cmd_i];
1114                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1115
1116                 /*
1117                  * pre-allocated x2 as much r2ts to handle race when
1118                  * target acks DataOut faster than we data_xmit() queues
1119                  * could replenish r2tqueue.
1120                  */
1121
1122                 /* R2T pool */
1123                 if (iscsi_pool_init(&tcp_task->r2tpool,
1124                                     session->max_r2t * 2, NULL,
1125                                     sizeof(struct iscsi_r2t_info))) {
1126                         goto r2t_alloc_fail;
1127                 }
1128
1129                 /* R2T xmit queue */
1130                 tcp_task->r2tqueue = kfifo_alloc(
1131                       session->max_r2t * 4 * sizeof(void*), GFP_KERNEL, NULL);
1132                 if (tcp_task->r2tqueue == ERR_PTR(-ENOMEM)) {
1133                         iscsi_pool_free(&tcp_task->r2tpool);
1134                         goto r2t_alloc_fail;
1135                 }
1136         }
1137
1138         return 0;
1139
1140 r2t_alloc_fail:
1141         for (i = 0; i < cmd_i; i++) {
1142                 struct iscsi_task *task = session->cmds[i];
1143                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1144
1145                 kfifo_free(tcp_task->r2tqueue);
1146                 iscsi_pool_free(&tcp_task->r2tpool);
1147         }
1148         return -ENOMEM;
1149 }
1150 EXPORT_SYMBOL_GPL(iscsi_tcp_r2tpool_alloc);
1151
1152 void iscsi_tcp_r2tpool_free(struct iscsi_session *session)
1153 {
1154         int i;
1155
1156         for (i = 0; i < session->cmds_max; i++) {
1157                 struct iscsi_task *task = session->cmds[i];
1158                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1159
1160                 kfifo_free(tcp_task->r2tqueue);
1161                 iscsi_pool_free(&tcp_task->r2tpool);
1162         }
1163 }
1164 EXPORT_SYMBOL_GPL(iscsi_tcp_r2tpool_free);
1165
1166 void iscsi_tcp_conn_get_stats(struct iscsi_cls_conn *cls_conn,
1167                               struct iscsi_stats *stats)
1168 {
1169         struct iscsi_conn *conn = cls_conn->dd_data;
1170
1171         stats->txdata_octets = conn->txdata_octets;
1172         stats->rxdata_octets = conn->rxdata_octets;
1173         stats->scsicmd_pdus = conn->scsicmd_pdus_cnt;
1174         stats->dataout_pdus = conn->dataout_pdus_cnt;
1175         stats->scsirsp_pdus = conn->scsirsp_pdus_cnt;
1176         stats->datain_pdus = conn->datain_pdus_cnt;
1177         stats->r2t_pdus = conn->r2t_pdus_cnt;
1178         stats->tmfcmd_pdus = conn->tmfcmd_pdus_cnt;
1179         stats->tmfrsp_pdus = conn->tmfrsp_pdus_cnt;
1180 }
1181 EXPORT_SYMBOL_GPL(iscsi_tcp_conn_get_stats);