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/slab.h>
  33 #include <linux/file.h>
  34 #include <linux/blkdev.h>
  35 #include <linux/crypto.h>
  36 #include <linux/delay.h>
  37 #include <linux/kfifo.h>
  38 #include <linux/scatterlist.h>
  39 #include <net/tcp.h>
  40 #include <scsi/scsi_cmnd.h>
  41 #include <scsi/scsi_device.h>
  42 #include <scsi/scsi_host.h>
  43 #include <scsi/scsi.h>
  44 #include <scsi/scsi_transport_iscsi.h>
  45
  46 #include "iscsi_tcp.h"
  47
  48 MODULE_AUTHOR("Mike Christie <michaelc@cs.wisc.edu>, "
  49               "Dmitry Yusupov <dmitry_yus@yahoo.com>, "
  50               "Alex Aizman <itn780@yahoo.com>");
  51 MODULE_DESCRIPTION("iSCSI/TCP data-path");
  52 MODULE_LICENSE("GPL");
  53
  54 static int iscsi_dbg_libtcp;
  55 module_param_named(debug_libiscsi_tcp, iscsi_dbg_libtcp, int,
  56                    S_IRUGO | S_IWUSR);
  57 MODULE_PARM_DESC(debug_libiscsi_tcp, "Turn on debugging for libiscsi_tcp "
  58                  "module. Set to 1 to turn on, and zero to turn off. Default "
  59                  "is off.");
  60
  61 #define ISCSI_DBG_TCP(_conn, dbg_fmt, arg...)                   \
  62         do {                                                    \
  63                 if (iscsi_dbg_libtcp)                           \
  64                         iscsi_conn_printk(KERN_INFO, _conn,     \
  65                                              "%s " dbg_fmt,     \
  66                                              __func__, ##arg);  \
  67         } while (0);
  68
  69 static int iscsi_tcp_hdr_recv_done(struct iscsi_tcp_conn *tcp_conn,
  70                                    struct iscsi_segment *segment);
  71
  72 /*
  73  * Scatterlist handling: inside the iscsi_segment, we
  74  * remember an index into the scatterlist, and set data/size
  75  * to the current scatterlist entry. For highmem pages, we
  76  * kmap as needed.
  77  *
  78  * Note that the page is unmapped when we return from
  79  * TCP's data_ready handler, so we may end up mapping and
  80  * unmapping the same page repeatedly. The whole reason
  81  * for this is that we shouldn't keep the page mapped
  82  * outside the softirq.
  83  */
  84
  85 /**
  86  * iscsi_tcp_segment_init_sg - init indicated scatterlist entry
  87  * @segment: the buffer object
  88  * @sg: scatterlist
  89  * @offset: byte offset into that sg entry
  90  *
  91  * This function sets up the segment so that subsequent
  92  * data is copied to the indicated sg entry, at the given
  93  * offset.
  94  */
  95 static inline void
  96 iscsi_tcp_segment_init_sg(struct iscsi_segment *segment,
  97                           struct scatterlist *sg, unsigned int offset)
  98 {
  99         segment->sg = sg;
 100         segment->sg_offset = offset;
 101         segment->size = min(sg->length - offset,
 102                             segment->total_size - segment->total_copied);
 103         segment->data = NULL;
 104 }
 105
 106 /**
 107  * iscsi_tcp_segment_map - map the current S/G page
 108  * @segment: iscsi_segment
 109  * @recv: 1 if called from recv path
 110  *
 111  * We only need to possibly kmap data if scatter lists are being used,
 112  * because the iscsi passthrough and internal IO paths will never use high
 113  * mem pages.
 114  */
 115 static void iscsi_tcp_segment_map(struct iscsi_segment *segment, int recv)
 116 {
 117         struct scatterlist *sg;
 118
 119         if (segment->data != NULL || !segment->sg)
 120                 return;
 121
 122         sg = segment->sg;
 123         BUG_ON(segment->sg_mapped);
 124         BUG_ON(sg->length == 0);
 125
 126         /*
 127          * If the page count is greater than one it is ok to send
 128          * to the network layer's zero copy send path. If not we
 129          * have to go the slow sendmsg path. We always map for the
 130          * recv path.
 131          */
 132         if (page_count(sg_page(sg)) >= 1 && !recv)
 133                 return;
 134
 135         segment->sg_mapped = kmap_atomic(sg_page(sg), KM_SOFTIRQ0);
 136         segment->data = segment->sg_mapped + sg->offset + segment->sg_offset;
 137 }
 138
 139 void iscsi_tcp_segment_unmap(struct iscsi_segment *segment)
 140 {
 141         if (segment->sg_mapped) {
 142                 kunmap_atomic(segment->sg_mapped, KM_SOFTIRQ0);
 143                 segment->sg_mapped = NULL;
 144                 segment->data = NULL;
 145         }
 146 }
 147 EXPORT_SYMBOL_GPL(iscsi_tcp_segment_unmap);
 148
 149 /*
 150  * Splice the digest buffer into the buffer
 151  */
 152 static inline void
 153 iscsi_tcp_segment_splice_digest(struct iscsi_segment *segment, void *digest)
 154 {
 155         segment->data = digest;
 156         segment->digest_len = ISCSI_DIGEST_SIZE;
 157         segment->total_size += ISCSI_DIGEST_SIZE;
 158         segment->size = ISCSI_DIGEST_SIZE;
 159         segment->copied = 0;
 160         segment->sg = NULL;
 161         segment->hash = NULL;
 162 }
 163
 164 /**
 165  * iscsi_tcp_segment_done - check whether the segment is complete
 166  * @tcp_conn: iscsi tcp connection
 167  * @segment: iscsi segment to check
 168  * @recv: set to one of this is called from the recv path
 169  * @copied: number of bytes copied
 170  *
 171  * Check if we're done receiving this segment. If the receive
 172  * buffer is full but we expect more data, move on to the
 173  * next entry in the scatterlist.
 174  *
 175  * If the amount of data we received isn't a multiple of 4,
 176  * we will transparently receive the pad bytes, too.
 177  *
 178  * This function must be re-entrant.
 179  */
 180 int iscsi_tcp_segment_done(struct iscsi_tcp_conn *tcp_conn,
 181                            struct iscsi_segment *segment, int recv,
 182                            unsigned copied)
 183 {
 184         struct scatterlist sg;
 185         unsigned int pad;
 186
 187         ISCSI_DBG_TCP(tcp_conn->iscsi_conn, "copied %u %u size %u %s\n",
 188                       segment->copied, copied, segment->size,
 189                       recv ? "recv" : "xmit");
 190         if (segment->hash && copied) {
 191                 /*
 192                  * If a segment is kmapd we must unmap it before sending
 193                  * to the crypto layer since that will try to kmap it again.
 194                  */
 195                 iscsi_tcp_segment_unmap(segment);
 196
 197                 if (!segment->data) {
 198                         sg_init_table(&sg, 1);
 199                         sg_set_page(&sg, sg_page(segment->sg), copied,
 200                                     segment->copied + segment->sg_offset +
 201                                                         segment->sg->offset);
 202                 } else
 203                         sg_init_one(&sg, segment->data + segment->copied,
 204                                     copied);
 205                 crypto_hash_update(segment->hash, &sg, copied);
 206         }
 207
 208         segment->copied += copied;
 209         if (segment->copied < segment->size) {
 210                 iscsi_tcp_segment_map(segment, recv);
 211                 return 0;
 212         }
 213
 214         segment->total_copied += segment->copied;
 215         segment->copied = 0;
 216         segment->size = 0;
 217
 218         /* Unmap the current scatterlist page, if there is one. */
 219         iscsi_tcp_segment_unmap(segment);
 220
 221         /* Do we have more scatterlist entries? */
 222         ISCSI_DBG_TCP(tcp_conn->iscsi_conn, "total copied %u total size %u\n",
 223                       segment->total_copied, segment->total_size);
 224         if (segment->total_copied < segment->total_size) {
 225                 /* Proceed to the next entry in the scatterlist. */
 226                 iscsi_tcp_segment_init_sg(segment, sg_next(segment->sg),
 227                                           0);
 228                 iscsi_tcp_segment_map(segment, recv);
 229                 BUG_ON(segment->size == 0);
 230                 return 0;
 231         }
 232
 233         /* Do we need to handle padding? */
 234         if (!(tcp_conn->iscsi_conn->session->tt->caps & CAP_PADDING_OFFLOAD)) {
 235                 pad = iscsi_padding(segment->total_copied);
 236                 if (pad != 0) {
 237                         ISCSI_DBG_TCP(tcp_conn->iscsi_conn,
 238                                       "consume %d pad bytes\n", pad);
 239                         segment->total_size += pad;
 240                         segment->size = pad;
 241                         segment->data = segment->padbuf;
 242                         return 0;
 243                 }
 244         }
 245
 246         /*
 247          * Set us up for transferring the data digest. hdr digest
 248          * is completely handled in hdr done function.
 249          */
 250         if (segment->hash) {
 251                 crypto_hash_final(segment->hash, segment->digest);
 252                 iscsi_tcp_segment_splice_digest(segment,
 253                                  recv ? segment->recv_digest : segment->digest);
 254                 return 0;
 255         }
 256
 257         return 1;
 258 }
 259 EXPORT_SYMBOL_GPL(iscsi_tcp_segment_done);
 260
 261 /**
 262  * iscsi_tcp_segment_recv - copy data to segment
 263  * @tcp_conn: the iSCSI TCP connection
 264  * @segment: the buffer to copy to
 265  * @ptr: data pointer
 266  * @len: amount of data available
 267  *
 268  * This function copies up to @len bytes to the
 269  * given buffer, and returns the number of bytes
 270  * consumed, which can actually be less than @len.
 271  *
 272  * If hash digest is enabled, the function will update the
 273  * hash while copying.
 274  * Combining these two operations doesn't buy us a lot (yet),
 275  * but in the future we could implement combined copy+crc,
 276  * just way we do for network layer checksums.
 277  */
 278 static int
 279 iscsi_tcp_segment_recv(struct iscsi_tcp_conn *tcp_conn,
 280                        struct iscsi_segment *segment, const void *ptr,
 281                        unsigned int len)
 282 {
 283         unsigned int copy = 0, copied = 0;
 284
 285         while (!iscsi_tcp_segment_done(tcp_conn, segment, 1, copy)) {
 286                 if (copied == len) {
 287                         ISCSI_DBG_TCP(tcp_conn->iscsi_conn,
 288                                       "copied %d bytes\n", len);
 289                         break;
 290                 }
 291
 292                 copy = min(len - copied, segment->size - segment->copied);
 293                 ISCSI_DBG_TCP(tcp_conn->iscsi_conn, "copying %d\n", copy);
 294                 memcpy(segment->data + segment->copied, ptr + copied, copy);
 295                 copied += copy;
 296         }
 297         return copied;
 298 }
 299
 300 inline void
 301 iscsi_tcp_dgst_header(struct hash_desc *hash, const void *hdr, size_t hdrlen,
 302                       unsigned char digest[ISCSI_DIGEST_SIZE])
 303 {
 304         struct scatterlist sg;
 305
 306         sg_init_one(&sg, hdr, hdrlen);
 307         crypto_hash_digest(hash, &sg, hdrlen, digest);
 308 }
 309 EXPORT_SYMBOL_GPL(iscsi_tcp_dgst_header);
 310
 311 static inline int
 312 iscsi_tcp_dgst_verify(struct iscsi_tcp_conn *tcp_conn,
 313                       struct iscsi_segment *segment)
 314 {
 315         if (!segment->digest_len)
 316                 return 1;
 317
 318         if (memcmp(segment->recv_digest, segment->digest,
 319                    segment->digest_len)) {
 320                 ISCSI_DBG_TCP(tcp_conn->iscsi_conn, "digest mismatch\n");
 321                 return 0;
 322         }
 323
 324         return 1;
 325 }
 326
 327 /*
 328  * Helper function to set up segment buffer
 329  */
 330 static inline void
 331 __iscsi_segment_init(struct iscsi_segment *segment, size_t size,
 332                      iscsi_segment_done_fn_t *done, struct hash_desc *hash)
 333 {
 334         memset(segment, 0, sizeof(*segment));
 335         segment->total_size = size;
 336         segment->done = done;
 337
 338         if (hash) {
 339                 segment->hash = hash;
 340                 crypto_hash_init(hash);
 341         }
 342 }
 343
 344 inline void
 345 iscsi_segment_init_linear(struct iscsi_segment *segment, void *data,
 346                           size_t size, iscsi_segment_done_fn_t *done,
 347                           struct hash_desc *hash)
 348 {
 349         __iscsi_segment_init(segment, size, done, hash);
 350         segment->data = data;
 351         segment->size = size;
 352 }
 353 EXPORT_SYMBOL_GPL(iscsi_segment_init_linear);
 354
 355 inline int
 356 iscsi_segment_seek_sg(struct iscsi_segment *segment,
 357                       struct scatterlist *sg_list, unsigned int sg_count,
 358                       unsigned int offset, size_t size,
 359                       iscsi_segment_done_fn_t *done, struct hash_desc *hash)
 360 {
 361         struct scatterlist *sg;
 362         unsigned int i;
 363
 364         __iscsi_segment_init(segment, size, done, hash);
 365         for_each_sg(sg_list, sg, sg_count, i) {
 366                 if (offset < sg->length) {
 367                         iscsi_tcp_segment_init_sg(segment, sg, offset);
 368                         return 0;
 369                 }
 370                 offset -= sg->length;
 371         }
 372
 373         return ISCSI_ERR_DATA_OFFSET;
 374 }
 375 EXPORT_SYMBOL_GPL(iscsi_segment_seek_sg);
 376
 377 /**
 378  * iscsi_tcp_hdr_recv_prep - prep segment for hdr reception
 379  * @tcp_conn: iscsi connection to prep for
 380  *
 381  * This function always passes NULL for the hash argument, because when this
 382  * function is called we do not yet know the final size of the header and want
 383  * to delay the digest processing until we know that.
 384  */
 385 void iscsi_tcp_hdr_recv_prep(struct iscsi_tcp_conn *tcp_conn)
 386 {
 387         ISCSI_DBG_TCP(tcp_conn->iscsi_conn,
 388                       "(%s)\n", tcp_conn->iscsi_conn->hdrdgst_en ?
 389                       "digest enabled" : "digest disabled");
 390         iscsi_segment_init_linear(&tcp_conn->in.segment,
 391                                 tcp_conn->in.hdr_buf, sizeof(struct iscsi_hdr),
 392                                 iscsi_tcp_hdr_recv_done, NULL);
 393 }
 394 EXPORT_SYMBOL_GPL(iscsi_tcp_hdr_recv_prep);
 395
 396 /*
 397  * Handle incoming reply to any other type of command
 398  */
 399 static int
 400 iscsi_tcp_data_recv_done(struct iscsi_tcp_conn *tcp_conn,
 401                          struct iscsi_segment *segment)
 402 {
 403         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 404         int rc = 0;
 405
 406         if (!iscsi_tcp_dgst_verify(tcp_conn, segment))
 407                 return ISCSI_ERR_DATA_DGST;
 408
 409         rc = iscsi_complete_pdu(conn, tcp_conn->in.hdr,
 410                         conn->data, tcp_conn->in.datalen);
 411         if (rc)
 412                 return rc;
 413
 414         iscsi_tcp_hdr_recv_prep(tcp_conn);
 415         return 0;
 416 }
 417
 418 static void
 419 iscsi_tcp_data_recv_prep(struct iscsi_tcp_conn *tcp_conn)
 420 {
 421         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 422         struct hash_desc *rx_hash = NULL;
 423
 424         if (conn->datadgst_en &
 425             !(conn->session->tt->caps & CAP_DIGEST_OFFLOAD))
 426                 rx_hash = tcp_conn->rx_hash;
 427
 428         iscsi_segment_init_linear(&tcp_conn->in.segment,
 429                                 conn->data, tcp_conn->in.datalen,
 430                                 iscsi_tcp_data_recv_done, rx_hash);
 431 }
 432
 433 /**
 434  * iscsi_tcp_cleanup_task - free tcp_task resources
 435  * @task: iscsi task
 436  *
 437  * must be called with session lock
 438  */
 439 void iscsi_tcp_cleanup_task(struct iscsi_task *task)
 440 {
 441         struct iscsi_tcp_task *tcp_task = task->dd_data;
 442         struct iscsi_r2t_info *r2t;
 443
 444         /* nothing to do for mgmt */
 445         if (!task->sc)
 446                 return;
 447
 448         /* flush task's r2t queues */
 449         while (kfifo_out(&tcp_task->r2tqueue, (void*)&r2t, sizeof(void*))) {
 450                 kfifo_in(&tcp_task->r2tpool.queue, (void*)&r2t,
 451                             sizeof(void*));
 452                 ISCSI_DBG_TCP(task->conn, "pending r2t dropped\n");
 453         }
 454
 455         r2t = tcp_task->r2t;
 456         if (r2t != NULL) {
 457                 kfifo_in(&tcp_task->r2tpool.queue, (void*)&r2t,
 458                             sizeof(void*));
 459                 tcp_task->r2t = NULL;
 460         }
 461 }
 462 EXPORT_SYMBOL_GPL(iscsi_tcp_cleanup_task);
 463
 464 /**
 465  * iscsi_tcp_data_in - SCSI Data-In Response processing
 466  * @conn: iscsi connection
 467  * @task: scsi command task
 468  */
 469 static int iscsi_tcp_data_in(struct iscsi_conn *conn, struct iscsi_task *task)
 470 {
 471         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 472         struct iscsi_tcp_task *tcp_task = task->dd_data;
 473         struct iscsi_data_rsp *rhdr = (struct iscsi_data_rsp *)tcp_conn->in.hdr;
 474         int datasn = be32_to_cpu(rhdr->datasn);
 475         unsigned total_in_length = scsi_in(task->sc)->length;
 476
 477         /*
 478          * lib iscsi will update this in the completion handling if there
 479          * is status.
 480          */
 481         if (!(rhdr->flags & ISCSI_FLAG_DATA_STATUS))
 482                 iscsi_update_cmdsn(conn->session, (struct iscsi_nopin*)rhdr);
 483
 484         if (tcp_conn->in.datalen == 0)
 485                 return 0;
 486
 487         if (tcp_task->exp_datasn != datasn) {
 488                 ISCSI_DBG_TCP(conn, "task->exp_datasn(%d) != rhdr->datasn(%d)"
 489                               "\n", tcp_task->exp_datasn, datasn);
 490                 return ISCSI_ERR_DATASN;
 491         }
 492
 493         tcp_task->exp_datasn++;
 494
 495         tcp_task->data_offset = be32_to_cpu(rhdr->offset);
 496         if (tcp_task->data_offset + tcp_conn->in.datalen > total_in_length) {
 497                 ISCSI_DBG_TCP(conn, "data_offset(%d) + data_len(%d) > "
 498                               "total_length_in(%d)\n", tcp_task->data_offset,
 499                               tcp_conn->in.datalen, total_in_length);
 500                 return ISCSI_ERR_DATA_OFFSET;
 501         }
 502
 503         conn->datain_pdus_cnt++;
 504         return 0;
 505 }
 506
 507 /**
 508  * iscsi_tcp_r2t_rsp - iSCSI R2T Response processing
 509  * @conn: iscsi connection
 510  * @task: scsi command task
 511  */
 512 static int iscsi_tcp_r2t_rsp(struct iscsi_conn *conn, struct iscsi_task *task)
 513 {
 514         struct iscsi_session *session = conn->session;
 515         struct iscsi_tcp_task *tcp_task = task->dd_data;
 516         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 517         struct iscsi_r2t_rsp *rhdr = (struct iscsi_r2t_rsp *)tcp_conn->in.hdr;
 518         struct iscsi_r2t_info *r2t;
 519         int r2tsn = be32_to_cpu(rhdr->r2tsn);
 520         int rc;
 521
 522         if (tcp_conn->in.datalen) {
 523                 iscsi_conn_printk(KERN_ERR, conn,
 524                                   "invalid R2t with datalen %d\n",
 525                                   tcp_conn->in.datalen);
 526                 return ISCSI_ERR_DATALEN;
 527         }
 528
 529         if (tcp_task->exp_datasn != r2tsn){
 530                 ISCSI_DBG_TCP(conn, "task->exp_datasn(%d) != rhdr->r2tsn(%d)\n",
 531                               tcp_task->exp_datasn, r2tsn);
 532                 return ISCSI_ERR_R2TSN;
 533         }
 534
 535         /* fill-in new R2T associated with the task */
 536         iscsi_update_cmdsn(session, (struct iscsi_nopin*)rhdr);
 537
 538         if (!task->sc || session->state != ISCSI_STATE_LOGGED_IN) {
 539                 iscsi_conn_printk(KERN_INFO, conn,
 540                                   "dropping R2T itt %d in recovery.\n",
 541                                   task->itt);
 542                 return 0;
 543         }
 544
 545         rc = kfifo_out(&tcp_task->r2tpool.queue, (void*)&r2t, sizeof(void*));
 546         if (!rc) {
 547                 iscsi_conn_printk(KERN_ERR, conn, "Could not allocate R2T. "
 548                                   "Target has sent more R2Ts than it "
 549                                   "negotiated for or driver has has leaked.\n");
 550                 return ISCSI_ERR_PROTO;
 551         }
 552
 553         r2t->exp_statsn = rhdr->statsn;
 554         r2t->data_length = be32_to_cpu(rhdr->data_length);
 555         if (r2t->data_length == 0) {
 556                 iscsi_conn_printk(KERN_ERR, conn,
 557                                   "invalid R2T with zero data len\n");
 558                 kfifo_in(&tcp_task->r2tpool.queue, (void*)&r2t,
 559                             sizeof(void*));
 560                 return ISCSI_ERR_DATALEN;
 561         }
 562
 563         if (r2t->data_length > session->max_burst)
 564                 ISCSI_DBG_TCP(conn, "invalid R2T with data len %u and max "
 565                               "burst %u. Attempting to execute request.\n",
 566                               r2t->data_length, session->max_burst);
 567
 568         r2t->data_offset = be32_to_cpu(rhdr->data_offset);
 569         if (r2t->data_offset + r2t->data_length > scsi_out(task->sc)->length) {
 570                 iscsi_conn_printk(KERN_ERR, conn,
 571                                   "invalid R2T with data len %u at offset %u "
 572                                   "and total length %d\n", r2t->data_length,
 573                                   r2t->data_offset, scsi_out(task->sc)->length);
 574                 kfifo_in(&tcp_task->r2tpool.queue, (void*)&r2t,
 575                             sizeof(void*));
 576                 return ISCSI_ERR_DATALEN;
 577         }
 578
 579         r2t->ttt = rhdr->ttt; /* no flip */
 580         r2t->datasn = 0;
 581         r2t->sent = 0;
 582
 583         tcp_task->exp_datasn = r2tsn + 1;
 584         kfifo_in(&tcp_task->r2tqueue, (void*)&r2t, sizeof(void*));
 585         conn->r2t_pdus_cnt++;
 586
 587         iscsi_requeue_task(task);
 588         return 0;
 589 }
 590
 591 /*
 592  * Handle incoming reply to DataIn command
 593  */
 594 static int
 595 iscsi_tcp_process_data_in(struct iscsi_tcp_conn *tcp_conn,
 596                           struct iscsi_segment *segment)
 597 {
 598         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 599         struct iscsi_hdr *hdr = tcp_conn->in.hdr;
 600         int rc;
 601
 602         if (!iscsi_tcp_dgst_verify(tcp_conn, segment))
 603                 return ISCSI_ERR_DATA_DGST;
 604
 605         /* check for non-exceptional status */
 606         if (hdr->flags & ISCSI_FLAG_DATA_STATUS) {
 607                 rc = iscsi_complete_pdu(conn, tcp_conn->in.hdr, NULL, 0);
 608                 if (rc)
 609                         return rc;
 610         }
 611
 612         iscsi_tcp_hdr_recv_prep(tcp_conn);
 613         return 0;
 614 }
 615
 616 /**
 617  * iscsi_tcp_hdr_dissect - process PDU header
 618  * @conn: iSCSI connection
 619  * @hdr: PDU header
 620  *
 621  * This function analyzes the header of the PDU received,
 622  * and performs several sanity checks. If the PDU is accompanied
 623  * by data, the receive buffer is set up to copy the incoming data
 624  * to the correct location.
 625  */
 626 static int
 627 iscsi_tcp_hdr_dissect(struct iscsi_conn *conn, struct iscsi_hdr *hdr)
 628 {
 629         int rc = 0, opcode, ahslen;
 630         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 631         struct iscsi_task *task;
 632
 633         /* verify PDU length */
 634         tcp_conn->in.datalen = ntoh24(hdr->dlength);
 635         if (tcp_conn->in.datalen > conn->max_recv_dlength) {
 636                 iscsi_conn_printk(KERN_ERR, conn,
 637                                   "iscsi_tcp: datalen %d > %d\n",
 638                                   tcp_conn->in.datalen, conn->max_recv_dlength);
 639                 return ISCSI_ERR_DATALEN;
 640         }
 641
 642         /* Additional header segments. So far, we don't
 643          * process additional headers.
 644          */
 645         ahslen = hdr->hlength << 2;
 646
 647         opcode = hdr->opcode & ISCSI_OPCODE_MASK;
 648         /* verify itt (itt encoding: age+cid+itt) */
 649         rc = iscsi_verify_itt(conn, hdr->itt);
 650         if (rc)
 651                 return rc;
 652
 653         ISCSI_DBG_TCP(conn, "opcode 0x%x ahslen %d datalen %d\n",
 654                       opcode, ahslen, tcp_conn->in.datalen);
 655
 656         switch(opcode) {
 657         case ISCSI_OP_SCSI_DATA_IN:
 658                 spin_lock(&conn->session->lock);
 659                 task = iscsi_itt_to_ctask(conn, hdr->itt);
 660                 if (!task)
 661                         rc = ISCSI_ERR_BAD_ITT;
 662                 else
 663                         rc = iscsi_tcp_data_in(conn, task);
 664                 if (rc) {
 665                         spin_unlock(&conn->session->lock);
 666                         break;
 667                 }
 668
 669                 if (tcp_conn->in.datalen) {
 670                         struct iscsi_tcp_task *tcp_task = task->dd_data;
 671                         struct hash_desc *rx_hash = NULL;
 672                         struct scsi_data_buffer *sdb = scsi_in(task->sc);
 673
 674                         /*
 675                          * Setup copy of Data-In into the Scsi_Cmnd
 676                          * Scatterlist case:
 677                          * We set up the iscsi_segment to point to the next
 678                          * scatterlist entry to copy to. As we go along,
 679                          * we move on to the next scatterlist entry and
 680                          * update the digest per-entry.
 681                          */
 682                         if (conn->datadgst_en &&
 683                             !(conn->session->tt->caps & CAP_DIGEST_OFFLOAD))
 684                                 rx_hash = tcp_conn->rx_hash;
 685
 686                         ISCSI_DBG_TCP(conn, "iscsi_tcp_begin_data_in( "
 687                                      "offset=%d, datalen=%d)\n",
 688                                       tcp_task->data_offset,
 689                                       tcp_conn->in.datalen);
 690                         task->last_xfer = jiffies;
 691                         rc = iscsi_segment_seek_sg(&tcp_conn->in.segment,
 692                                                    sdb->table.sgl,
 693                                                    sdb->table.nents,
 694                                                    tcp_task->data_offset,
 695                                                    tcp_conn->in.datalen,
 696                                                    iscsi_tcp_process_data_in,
 697                                                    rx_hash);
 698                         spin_unlock(&conn->session->lock);
 699                         return rc;
 700                 }
 701                 rc = __iscsi_complete_pdu(conn, hdr, NULL, 0);
 702                 spin_unlock(&conn->session->lock);
 703                 break;
 704         case ISCSI_OP_SCSI_CMD_RSP:
 705                 if (tcp_conn->in.datalen) {
 706                         iscsi_tcp_data_recv_prep(tcp_conn);
 707                         return 0;
 708                 }
 709                 rc = iscsi_complete_pdu(conn, hdr, NULL, 0);
 710                 break;
 711         case ISCSI_OP_R2T:
 712                 spin_lock(&conn->session->lock);
 713                 task = iscsi_itt_to_ctask(conn, hdr->itt);
 714                 if (!task)
 715                         rc = ISCSI_ERR_BAD_ITT;
 716                 else if (ahslen)
 717                         rc = ISCSI_ERR_AHSLEN;
 718                 else if (task->sc->sc_data_direction == DMA_TO_DEVICE) {
 719                         task->last_xfer = jiffies;
 720                         rc = iscsi_tcp_r2t_rsp(conn, task);
 721                 } else
 722                         rc = ISCSI_ERR_PROTO;
 723                 spin_unlock(&conn->session->lock);
 724                 break;
 725         case ISCSI_OP_LOGIN_RSP:
 726         case ISCSI_OP_TEXT_RSP:
 727         case ISCSI_OP_REJECT:
 728         case ISCSI_OP_ASYNC_EVENT:
 729                 /*
 730                  * It is possible that we could get a PDU with a buffer larger
 731                  * than 8K, but there are no targets that currently do this.
 732                  * For now we fail until we find a vendor that needs it
 733                  */
 734                 if (ISCSI_DEF_MAX_RECV_SEG_LEN < tcp_conn->in.datalen) {
 735                         iscsi_conn_printk(KERN_ERR, conn,
 736                                           "iscsi_tcp: received buffer of "
 737                                           "len %u but conn buffer is only %u "
 738                                           "(opcode %0x)\n",
 739                                           tcp_conn->in.datalen,
 740                                           ISCSI_DEF_MAX_RECV_SEG_LEN, opcode);
 741                         rc = ISCSI_ERR_PROTO;
 742                         break;
 743                 }
 744
 745                 /* If there's data coming in with the response,
 746                  * receive it to the connection's buffer.
 747                  */
 748                 if (tcp_conn->in.datalen) {
 749                         iscsi_tcp_data_recv_prep(tcp_conn);
 750                         return 0;
 751                 }
 752         /* fall through */
 753         case ISCSI_OP_LOGOUT_RSP:
 754         case ISCSI_OP_NOOP_IN:
 755         case ISCSI_OP_SCSI_TMFUNC_RSP:
 756                 rc = iscsi_complete_pdu(conn, hdr, NULL, 0);
 757                 break;
 758         default:
 759                 rc = ISCSI_ERR_BAD_OPCODE;
 760                 break;
 761         }
 762
 763         if (rc == 0) {
 764                 /* Anything that comes with data should have
 765                  * been handled above. */
 766                 if (tcp_conn->in.datalen)
 767                         return ISCSI_ERR_PROTO;
 768                 iscsi_tcp_hdr_recv_prep(tcp_conn);
 769         }
 770
 771         return rc;
 772 }
 773
 774 /**
 775  * iscsi_tcp_hdr_recv_done - process PDU header
 776  *
 777  * This is the callback invoked when the PDU header has
 778  * been received. If the header is followed by additional
 779  * header segments, we go back for more data.
 780  */
 781 static int
 782 iscsi_tcp_hdr_recv_done(struct iscsi_tcp_conn *tcp_conn,
 783                         struct iscsi_segment *segment)
 784 {
 785         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 786         struct iscsi_hdr *hdr;
 787
 788         /* Check if there are additional header segments
 789          * *prior* to computing the digest, because we
 790          * may need to go back to the caller for more.
 791          */
 792         hdr = (struct iscsi_hdr *) tcp_conn->in.hdr_buf;
 793         if (segment->copied == sizeof(struct iscsi_hdr) && hdr->hlength) {
 794                 /* Bump the header length - the caller will
 795                  * just loop around and get the AHS for us, and
 796                  * call again. */
 797                 unsigned int ahslen = hdr->hlength << 2;
 798
 799                 /* Make sure we don't overflow */
 800                 if (sizeof(*hdr) + ahslen > sizeof(tcp_conn->in.hdr_buf))
 801                         return ISCSI_ERR_AHSLEN;
 802
 803                 segment->total_size += ahslen;
 804                 segment->size += ahslen;
 805                 return 0;
 806         }
 807
 808         /* We're done processing the header. See if we're doing
 809          * header digests; if so, set up the recv_digest buffer
 810          * and go back for more. */
 811         if (conn->hdrdgst_en &&
 812             !(conn->session->tt->caps & CAP_DIGEST_OFFLOAD)) {
 813                 if (segment->digest_len == 0) {
 814                         /*
 815                          * Even if we offload the digest processing we
 816                          * splice it in so we can increment the skb/segment
 817                          * counters in preparation for the data segment.
 818                          */
 819                         iscsi_tcp_segment_splice_digest(segment,
 820                                                         segment->recv_digest);
 821                         return 0;
 822                 }
 823
 824                 iscsi_tcp_dgst_header(tcp_conn->rx_hash, hdr,
 825                                       segment->total_copied - ISCSI_DIGEST_SIZE,
 826                                       segment->digest);
 827
 828                 if (!iscsi_tcp_dgst_verify(tcp_conn, segment))
 829                         return ISCSI_ERR_HDR_DGST;
 830         }
 831
 832         tcp_conn->in.hdr = hdr;
 833         return iscsi_tcp_hdr_dissect(conn, hdr);
 834 }
 835
 836 /**
 837  * iscsi_tcp_recv_segment_is_hdr - tests if we are reading in a header
 838  * @tcp_conn: iscsi tcp conn
 839  *
 840  * returns non zero if we are currently processing or setup to process
 841  * a header.
 842  */
 843 inline int iscsi_tcp_recv_segment_is_hdr(struct iscsi_tcp_conn *tcp_conn)
 844 {
 845         return tcp_conn->in.segment.done == iscsi_tcp_hdr_recv_done;
 846 }
 847 EXPORT_SYMBOL_GPL(iscsi_tcp_recv_segment_is_hdr);
 848
 849 /**
 850  * iscsi_tcp_recv_skb - Process skb
 851  * @conn: iscsi connection
 852  * @skb: network buffer with header and/or data segment
 853  * @offset: offset in skb
 854  * @offload: bool indicating if transfer was offloaded
 855  *
 856  * Will return status of transfer in status. And will return
 857  * number of bytes copied.
 858  */
 859 int iscsi_tcp_recv_skb(struct iscsi_conn *conn, struct sk_buff *skb,
 860                        unsigned int offset, bool offloaded, int *status)
 861 {
 862         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 863         struct iscsi_segment *segment = &tcp_conn->in.segment;
 864         struct skb_seq_state seq;
 865         unsigned int consumed = 0;
 866         int rc = 0;
 867
 868         ISCSI_DBG_TCP(conn, "in %d bytes\n", skb->len - offset);
 869         /*
 870          * Update for each skb instead of pdu, because over slow networks a
 871          * data_in's data could take a while to read in. We also want to
 872          * account for r2ts.
 873          */
 874         conn->last_recv = jiffies;
 875
 876         if (unlikely(conn->suspend_rx)) {
 877                 ISCSI_DBG_TCP(conn, "Rx suspended!\n");
 878                 *status = ISCSI_TCP_SUSPENDED;
 879                 return 0;
 880         }
 881
 882         if (offloaded) {
 883                 segment->total_copied = segment->total_size;
 884                 goto segment_done;
 885         }
 886
 887         skb_prepare_seq_read(skb, offset, skb->len, &seq);
 888         while (1) {
 889                 unsigned int avail;
 890                 const u8 *ptr;
 891
 892                 avail = skb_seq_read(consumed, &ptr, &seq);
 893                 if (avail == 0) {
 894                         ISCSI_DBG_TCP(conn, "no more data avail. Consumed %d\n",
 895                                       consumed);
 896                         *status = ISCSI_TCP_SKB_DONE;
 897                         skb_abort_seq_read(&seq);
 898                         goto skb_done;
 899                 }
 900                 BUG_ON(segment->copied >= segment->size);
 901
 902                 ISCSI_DBG_TCP(conn, "skb %p ptr=%p avail=%u\n", skb, ptr,
 903                               avail);
 904                 rc = iscsi_tcp_segment_recv(tcp_conn, segment, ptr, avail);
 905                 BUG_ON(rc == 0);
 906                 consumed += rc;
 907
 908                 if (segment->total_copied >= segment->total_size) {
 909                         skb_abort_seq_read(&seq);
 910                         goto segment_done;
 911                 }
 912         }
 913
 914 segment_done:
 915         *status = ISCSI_TCP_SEGMENT_DONE;
 916         ISCSI_DBG_TCP(conn, "segment done\n");
 917         rc = segment->done(tcp_conn, segment);
 918         if (rc != 0) {
 919                 *status = ISCSI_TCP_CONN_ERR;
 920                 ISCSI_DBG_TCP(conn, "Error receiving PDU, errno=%d\n", rc);
 921                 iscsi_conn_failure(conn, rc);
 922                 return 0;
 923         }
 924         /* The done() functions sets up the next segment. */
 925
 926 skb_done:
 927         conn->rxdata_octets += consumed;
 928         return consumed;
 929 }
 930 EXPORT_SYMBOL_GPL(iscsi_tcp_recv_skb);
 931
 932 /**
 933  * iscsi_tcp_task_init - Initialize iSCSI SCSI_READ or SCSI_WRITE commands
 934  * @conn: iscsi connection
 935  * @task: scsi command task
 936  * @sc: scsi command
 937  */
 938 int iscsi_tcp_task_init(struct iscsi_task *task)
 939 {
 940         struct iscsi_tcp_task *tcp_task = task->dd_data;
 941         struct iscsi_conn *conn = task->conn;
 942         struct scsi_cmnd *sc = task->sc;
 943         int err;
 944
 945         if (!sc) {
 946                 /*
 947                  * mgmt tasks do not have a scatterlist since they come
 948                  * in from the iscsi interface.
 949                  */
 950                 ISCSI_DBG_TCP(conn, "mtask deq [itt 0x%x]\n", task->itt);
 951
 952                 return conn->session->tt->init_pdu(task, 0, task->data_count);
 953         }
 954
 955         BUG_ON(kfifo_len(&tcp_task->r2tqueue));
 956         tcp_task->exp_datasn = 0;
 957
 958         /* Prepare PDU, optionally w/ immediate data */
 959         ISCSI_DBG_TCP(conn, "task deq [itt 0x%x imm %d unsol %d]\n",
 960                       task->itt, task->imm_count, task->unsol_r2t.data_length);
 961
 962         err = conn->session->tt->init_pdu(task, 0, task->imm_count);
 963         if (err)
 964                 return err;
 965         task->imm_count = 0;
 966         return 0;
 967 }
 968 EXPORT_SYMBOL_GPL(iscsi_tcp_task_init);
 969
 970 static struct iscsi_r2t_info *iscsi_tcp_get_curr_r2t(struct iscsi_task *task)
 971 {
 972         struct iscsi_session *session = task->conn->session;
 973         struct iscsi_tcp_task *tcp_task = task->dd_data;
 974         struct iscsi_r2t_info *r2t = NULL;
 975
 976         if (iscsi_task_has_unsol_data(task))
 977                 r2t = &task->unsol_r2t;
 978         else {
 979                 spin_lock_bh(&session->lock);
 980                 if (tcp_task->r2t) {
 981                         r2t = tcp_task->r2t;
 982                         /* Continue with this R2T? */
 983                         if (r2t->data_length <= r2t->sent) {
 984                                 ISCSI_DBG_TCP(task->conn,
 985                                               "  done with r2t %p\n", r2t);
 986                                 kfifo_in(&tcp_task->r2tpool.queue,
 987                                             (void *)&tcp_task->r2t,
 988                                             sizeof(void *));
 989                                 tcp_task->r2t = r2t = NULL;
 990                         }
 991                 }
 992
 993                 if (r2t == NULL) {
 994                         if (kfifo_out(&tcp_task->r2tqueue,
 995                             (void *)&tcp_task->r2t, sizeof(void *)) !=
 996                             sizeof(void *))
 997                                 r2t = NULL;
 998                         else
 999                                 r2t = tcp_task->r2t;
1000                 }
1001                 spin_unlock_bh(&session->lock);
1002         }
1003
1004         return r2t;
1005 }
1006
1007 /**
1008  * iscsi_tcp_task_xmit - xmit normal PDU task
1009  * @task: iscsi command task
1010  *
1011  * We're expected to return 0 when everything was transmitted successfully,
1012  * -EAGAIN if there's still data in the queue, or != 0 for any other kind
1013  * of error.
1014  */
1015 int iscsi_tcp_task_xmit(struct iscsi_task *task)
1016 {
1017         struct iscsi_conn *conn = task->conn;
1018         struct iscsi_session *session = conn->session;
1019         struct iscsi_r2t_info *r2t;
1020         int rc = 0;
1021
1022 flush:
1023         /* Flush any pending data first. */
1024         rc = session->tt->xmit_pdu(task);
1025         if (rc < 0)
1026                 return rc;
1027
1028         /* mgmt command */
1029         if (!task->sc) {
1030                 if (task->hdr->itt == RESERVED_ITT)
1031                         iscsi_put_task(task);
1032                 return 0;
1033         }
1034
1035         /* Are we done already? */
1036         if (task->sc->sc_data_direction != DMA_TO_DEVICE)
1037                 return 0;
1038
1039         r2t = iscsi_tcp_get_curr_r2t(task);
1040         if (r2t == NULL) {
1041                 /* Waiting for more R2Ts to arrive. */
1042                 ISCSI_DBG_TCP(conn, "no R2Ts yet\n");
1043                 return 0;
1044         }
1045
1046         rc = conn->session->tt->alloc_pdu(task, ISCSI_OP_SCSI_DATA_OUT);
1047         if (rc)
1048                 return rc;
1049         iscsi_prep_data_out_pdu(task, r2t, (struct iscsi_data *) task->hdr);
1050
1051         ISCSI_DBG_TCP(conn, "sol dout %p [dsn %d itt 0x%x doff %d dlen %d]\n",
1052                       r2t, r2t->datasn - 1, task->hdr->itt,
1053                       r2t->data_offset + r2t->sent, r2t->data_count);
1054
1055         rc = conn->session->tt->init_pdu(task, r2t->data_offset + r2t->sent,
1056                                          r2t->data_count);
1057         if (rc) {
1058                 iscsi_conn_failure(conn, ISCSI_ERR_XMIT_FAILED);
1059                 return rc;
1060         }
1061
1062         r2t->sent += r2t->data_count;
1063         goto flush;
1064 }
1065 EXPORT_SYMBOL_GPL(iscsi_tcp_task_xmit);
1066
1067 struct iscsi_cls_conn *
1068 iscsi_tcp_conn_setup(struct iscsi_cls_session *cls_session, int dd_data_size,
1069                       uint32_t conn_idx)
1070
1071 {
1072         struct iscsi_conn *conn;
1073         struct iscsi_cls_conn *cls_conn;
1074         struct iscsi_tcp_conn *tcp_conn;
1075
1076         cls_conn = iscsi_conn_setup(cls_session, sizeof(*tcp_conn), conn_idx);
1077         if (!cls_conn)
1078                 return NULL;
1079         conn = cls_conn->dd_data;
1080         /*
1081          * due to strange issues with iser these are not set
1082          * in iscsi_conn_setup
1083          */
1084         conn->max_recv_dlength = ISCSI_DEF_MAX_RECV_SEG_LEN;
1085
1086         tcp_conn = conn->dd_data;
1087         tcp_conn->iscsi_conn = conn;
1088
1089         tcp_conn->dd_data = kzalloc(dd_data_size, GFP_KERNEL);
1090         if (!tcp_conn->dd_data) {
1091                 iscsi_conn_teardown(cls_conn);
1092                 return NULL;
1093         }
1094         return cls_conn;
1095 }
1096 EXPORT_SYMBOL_GPL(iscsi_tcp_conn_setup);
1097
1098 void iscsi_tcp_conn_teardown(struct iscsi_cls_conn *cls_conn)
1099 {
1100         struct iscsi_conn *conn = cls_conn->dd_data;
1101         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1102
1103         kfree(tcp_conn->dd_data);
1104         iscsi_conn_teardown(cls_conn);
1105 }
1106 EXPORT_SYMBOL_GPL(iscsi_tcp_conn_teardown);
1107
1108 int iscsi_tcp_r2tpool_alloc(struct iscsi_session *session)
1109 {
1110         int i;
1111         int cmd_i;
1112
1113         /*
1114          * initialize per-task: R2T pool and xmit queue
1115          */
1116         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
1117                 struct iscsi_task *task = session->cmds[cmd_i];
1118                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1119
1120                 /*
1121                  * pre-allocated x2 as much r2ts to handle race when
1122                  * target acks DataOut faster than we data_xmit() queues
1123                  * could replenish r2tqueue.
1124                  */
1125
1126                 /* R2T pool */
1127                 if (iscsi_pool_init(&tcp_task->r2tpool,
1128                                     session->max_r2t * 2, NULL,
1129                                     sizeof(struct iscsi_r2t_info))) {
1130                         goto r2t_alloc_fail;
1131                 }
1132
1133                 /* R2T xmit queue */
1134                 if (kfifo_alloc(&tcp_task->r2tqueue,
1135                       session->max_r2t * 4 * sizeof(void*), GFP_KERNEL)) {
1136                         iscsi_pool_free(&tcp_task->r2tpool);
1137                         goto r2t_alloc_fail;
1138                 }
1139         }
1140
1141         return 0;
1142
1143 r2t_alloc_fail:
1144         for (i = 0; i < cmd_i; i++) {
1145                 struct iscsi_task *task = session->cmds[i];
1146                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1147
1148                 kfifo_free(&tcp_task->r2tqueue);
1149                 iscsi_pool_free(&tcp_task->r2tpool);
1150         }
1151         return -ENOMEM;
1152 }
1153 EXPORT_SYMBOL_GPL(iscsi_tcp_r2tpool_alloc);
1154
1155 void iscsi_tcp_r2tpool_free(struct iscsi_session *session)
1156 {
1157         int i;
1158
1159         for (i = 0; i < session->cmds_max; i++) {
1160                 struct iscsi_task *task = session->cmds[i];
1161                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1162
1163                 kfifo_free(&tcp_task->r2tqueue);
1164                 iscsi_pool_free(&tcp_task->r2tpool);
1165         }
1166 }
1167 EXPORT_SYMBOL_GPL(iscsi_tcp_r2tpool_free);
1168
1169 void iscsi_tcp_conn_get_stats(struct iscsi_cls_conn *cls_conn,
1170                               struct iscsi_stats *stats)
1171 {
1172         struct iscsi_conn *conn = cls_conn->dd_data;
1173
1174         stats->txdata_octets = conn->txdata_octets;
1175         stats->rxdata_octets = conn->rxdata_octets;
1176         stats->scsicmd_pdus = conn->scsicmd_pdus_cnt;
1177         stats->dataout_pdus = conn->dataout_pdus_cnt;
1178         stats->scsirsp_pdus = conn->scsirsp_pdus_cnt;
1179         stats->datain_pdus = conn->datain_pdus_cnt;
1180         stats->r2t_pdus = conn->r2t_pdus_cnt;
1181         stats->tmfcmd_pdus = conn->tmfcmd_pdus_cnt;
1182         stats->tmfrsp_pdus = conn->tmfrsp_pdus_cnt;
1183 }
1184 EXPORT_SYMBOL_GPL(iscsi_tcp_conn_get_stats);