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