SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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         static unsigned char padbuf[ISCSI_PAD_LEN];
 181         struct scatterlist sg;
 182         unsigned int pad;
 183
 184         debug_tcp("copied %u %u size %u %s\n", segment->copied, copied,
 185                   segment->size, recv ? "recv" : "xmit");
 186         if (segment->hash && copied) {
 187                 /*
 188                  * If a segment is kmapd we must unmap it before sending
 189                  * to the crypto layer since that will try to kmap it again.
 190                  */
 191                 iscsi_tcp_segment_unmap(segment);
 192
 193                 if (!segment->data) {
 194                         sg_init_table(&sg, 1);
 195                         sg_set_page(&sg, sg_page(segment->sg), copied,
 196                                     segment->copied + segment->sg_offset +
 197                                                         segment->sg->offset);
 198                 } else
 199                         sg_init_one(&sg, segment->data + segment->copied,
 200                                     copied);
 201                 crypto_hash_update(segment->hash, &sg, copied);
 202         }
 203
 204         segment->copied += copied;
 205         if (segment->copied < segment->size) {
 206                 iscsi_tcp_segment_map(segment, recv);
 207                 return 0;
 208         }
 209
 210         segment->total_copied += segment->copied;
 211         segment->copied = 0;
 212         segment->size = 0;
 213
 214         /* Unmap the current scatterlist page, if there is one. */
 215         iscsi_tcp_segment_unmap(segment);
 216
 217         /* Do we have more scatterlist entries? */
 218         debug_tcp("total copied %u total size %u\n", segment->total_copied,
 219                    segment->total_size);
 220         if (segment->total_copied < segment->total_size) {
 221                 /* Proceed to the next entry in the scatterlist. */
 222                 iscsi_tcp_segment_init_sg(segment, sg_next(segment->sg),
 223                                           0);
 224                 iscsi_tcp_segment_map(segment, recv);
 225                 BUG_ON(segment->size == 0);
 226                 return 0;
 227         }
 228
 229         /* Do we need to handle padding? */
 230         if (!(tcp_conn->iscsi_conn->session->tt->caps & CAP_PADDING_OFFLOAD)) {
 231                 pad = iscsi_padding(segment->total_copied);
 232                 if (pad != 0) {
 233                         debug_tcp("consume %d pad bytes\n", pad);
 234                         segment->total_size += pad;
 235                         segment->size = pad;
 236                         segment->data = padbuf;
 237                         return 0;
 238                 }
 239         }
 240
 241         /*
 242          * Set us up for transferring the data digest. hdr digest
 243          * is completely handled in hdr done function.
 244          */
 245         if (segment->hash) {
 246                 crypto_hash_final(segment->hash, segment->digest);
 247                 iscsi_tcp_segment_splice_digest(segment,
 248                                  recv ? segment->recv_digest : segment->digest);
 249                 return 0;
 250         }
 251
 252         return 1;
 253 }
 254 EXPORT_SYMBOL_GPL(iscsi_tcp_segment_done);
 255
 256 /**
 257  * iscsi_tcp_segment_recv - copy data to segment
 258  * @tcp_conn: the iSCSI TCP connection
 259  * @segment: the buffer to copy to
 260  * @ptr: data pointer
 261  * @len: amount of data available
 262  *
 263  * This function copies up to @len bytes to the
 264  * given buffer, and returns the number of bytes
 265  * consumed, which can actually be less than @len.
 266  *
 267  * If hash digest is enabled, the function will update the
 268  * hash while copying.
 269  * Combining these two operations doesn't buy us a lot (yet),
 270  * but in the future we could implement combined copy+crc,
 271  * just way we do for network layer checksums.
 272  */
 273 static int
 274 iscsi_tcp_segment_recv(struct iscsi_tcp_conn *tcp_conn,
 275                        struct iscsi_segment *segment, const void *ptr,
 276                        unsigned int len)
 277 {
 278         unsigned int copy = 0, copied = 0;
 279
 280         while (!iscsi_tcp_segment_done(tcp_conn, segment, 1, copy)) {
 281                 if (copied == len) {
 282                         debug_tcp("iscsi_tcp_segment_recv copied %d bytes\n",
 283                                   len);
 284                         break;
 285                 }
 286
 287                 copy = min(len - copied, segment->size - segment->copied);
 288                 debug_tcp("iscsi_tcp_segment_recv copying %d\n", copy);
 289                 memcpy(segment->data + segment->copied, ptr + copied, copy);
 290                 copied += copy;
 291         }
 292         return copied;
 293 }
 294
 295 inline void
 296 iscsi_tcp_dgst_header(struct hash_desc *hash, const void *hdr, size_t hdrlen,
 297                       unsigned char digest[ISCSI_DIGEST_SIZE])
 298 {
 299         struct scatterlist sg;
 300
 301         sg_init_one(&sg, hdr, hdrlen);
 302         crypto_hash_digest(hash, &sg, hdrlen, digest);
 303 }
 304 EXPORT_SYMBOL_GPL(iscsi_tcp_dgst_header);
 305
 306 static inline int
 307 iscsi_tcp_dgst_verify(struct iscsi_tcp_conn *tcp_conn,
 308                       struct iscsi_segment *segment)
 309 {
 310         if (!segment->digest_len)
 311                 return 1;
 312
 313         if (memcmp(segment->recv_digest, segment->digest,
 314                    segment->digest_len)) {
 315                 debug_scsi("digest mismatch\n");
 316                 return 0;
 317         }
 318
 319         return 1;
 320 }
 321
 322 /*
 323  * Helper function to set up segment buffer
 324  */
 325 static inline void
 326 __iscsi_segment_init(struct iscsi_segment *segment, size_t size,
 327                      iscsi_segment_done_fn_t *done, struct hash_desc *hash)
 328 {
 329         memset(segment, 0, sizeof(*segment));
 330         segment->total_size = size;
 331         segment->done = done;
 332
 333         if (hash) {
 334                 segment->hash = hash;
 335                 crypto_hash_init(hash);
 336         }
 337 }
 338
 339 inline void
 340 iscsi_segment_init_linear(struct iscsi_segment *segment, void *data,
 341                           size_t size, iscsi_segment_done_fn_t *done,
 342                           struct hash_desc *hash)
 343 {
 344         __iscsi_segment_init(segment, size, done, hash);
 345         segment->data = data;
 346         segment->size = size;
 347 }
 348 EXPORT_SYMBOL_GPL(iscsi_segment_init_linear);
 349
 350 inline int
 351 iscsi_segment_seek_sg(struct iscsi_segment *segment,
 352                       struct scatterlist *sg_list, unsigned int sg_count,
 353                       unsigned int offset, size_t size,
 354                       iscsi_segment_done_fn_t *done, struct hash_desc *hash)
 355 {
 356         struct scatterlist *sg;
 357         unsigned int i;
 358
 359         debug_scsi("iscsi_segment_seek_sg offset %u size %llu\n",
 360                   offset, size);
 361         __iscsi_segment_init(segment, size, done, hash);
 362         for_each_sg(sg_list, sg, sg_count, i) {
 363                 debug_scsi("sg %d, len %u offset %u\n", i, sg->length,
 364                            sg->offset);
 365                 if (offset < sg->length) {
 366                         iscsi_tcp_segment_init_sg(segment, sg, offset);
 367                         return 0;
 368                 }
 369                 offset -= sg->length;
 370         }
 371
 372         return ISCSI_ERR_DATA_OFFSET;
 373 }
 374 EXPORT_SYMBOL_GPL(iscsi_segment_seek_sg);
 375
 376 /**
 377  * iscsi_tcp_hdr_recv_prep - prep segment for hdr reception
 378  * @tcp_conn: iscsi connection to prep for
 379  *
 380  * This function always passes NULL for the hash argument, because when this
 381  * function is called we do not yet know the final size of the header and want
 382  * to delay the digest processing until we know that.
 383  */
 384 void iscsi_tcp_hdr_recv_prep(struct iscsi_tcp_conn *tcp_conn)
 385 {
 386         debug_tcp("iscsi_tcp_hdr_recv_prep(%p%s)\n", tcp_conn,
 387                   tcp_conn->iscsi_conn->hdrdgst_en ? ", digest enabled" : "");
 388         iscsi_segment_init_linear(&tcp_conn->in.segment,
 389                                 tcp_conn->in.hdr_buf, sizeof(struct iscsi_hdr),
 390                                 iscsi_tcp_hdr_recv_done, NULL);
 391 }
 392 EXPORT_SYMBOL_GPL(iscsi_tcp_hdr_recv_prep);
 393
 394 /*
 395  * Handle incoming reply to any other type of command
 396  */
 397 static int
 398 iscsi_tcp_data_recv_done(struct iscsi_tcp_conn *tcp_conn,
 399                          struct iscsi_segment *segment)
 400 {
 401         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 402         int rc = 0;
 403
 404         if (!iscsi_tcp_dgst_verify(tcp_conn, segment))
 405                 return ISCSI_ERR_DATA_DGST;
 406
 407         rc = iscsi_complete_pdu(conn, tcp_conn->in.hdr,
 408                         conn->data, tcp_conn->in.datalen);
 409         if (rc)
 410                 return rc;
 411
 412         iscsi_tcp_hdr_recv_prep(tcp_conn);
 413         return 0;
 414 }
 415
 416 static void
 417 iscsi_tcp_data_recv_prep(struct iscsi_tcp_conn *tcp_conn)
 418 {
 419         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 420         struct hash_desc *rx_hash = NULL;
 421
 422         if (conn->datadgst_en &
 423             !(conn->session->tt->caps & CAP_DIGEST_OFFLOAD))
 424                 rx_hash = tcp_conn->rx_hash;
 425
 426         iscsi_segment_init_linear(&tcp_conn->in.segment,
 427                                 conn->data, tcp_conn->in.datalen,
 428                                 iscsi_tcp_data_recv_done, rx_hash);
 429 }
 430
 431 /**
 432  * iscsi_tcp_cleanup_task - free tcp_task resources
 433  * @task: iscsi task
 434  *
 435  * must be called with session lock
 436  */
 437 void iscsi_tcp_cleanup_task(struct iscsi_task *task)
 438 {
 439         struct iscsi_tcp_task *tcp_task = task->dd_data;
 440         struct iscsi_r2t_info *r2t;
 441
 442         /* nothing to do for mgmt or pending tasks */
 443         if (!task->sc || task->state == ISCSI_TASK_PENDING)
 444                 return;
 445
 446         /* flush task's r2t queues */
 447         while (__kfifo_get(tcp_task->r2tqueue, (void*)&r2t, sizeof(void*))) {
 448                 __kfifo_put(tcp_task->r2tpool.queue, (void*)&r2t,
 449                             sizeof(void*));
 450                 debug_scsi("iscsi_tcp_cleanup_task pending r2t dropped\n");
 451         }
 452
 453         r2t = tcp_task->r2t;
 454         if (r2t != NULL) {
 455                 __kfifo_put(tcp_task->r2tpool.queue, (void*)&r2t,
 456                             sizeof(void*));
 457                 tcp_task->r2t = NULL;
 458         }
 459 }
 460 EXPORT_SYMBOL_GPL(iscsi_tcp_cleanup_task);
 461
 462 /**
 463  * iscsi_tcp_data_in - SCSI Data-In Response processing
 464  * @conn: iscsi connection
 465  * @task: scsi command task
 466  */
 467 static int iscsi_tcp_data_in(struct iscsi_conn *conn, struct iscsi_task *task)
 468 {
 469         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 470         struct iscsi_tcp_task *tcp_task = task->dd_data;
 471         struct iscsi_data_rsp *rhdr = (struct iscsi_data_rsp *)tcp_conn->in.hdr;
 472         int datasn = be32_to_cpu(rhdr->datasn);
 473         unsigned total_in_length = scsi_in(task->sc)->length;
 474
 475         iscsi_update_cmdsn(conn->session, (struct iscsi_nopin*)rhdr);
 476         if (tcp_conn->in.datalen == 0)
 477                 return 0;
 478
 479         if (tcp_task->exp_datasn != datasn) {
 480                 debug_tcp("%s: task->exp_datasn(%d) != rhdr->datasn(%d)\n",
 481                           __func__, tcp_task->exp_datasn, datasn);
 482                 return ISCSI_ERR_DATASN;
 483         }
 484
 485         tcp_task->exp_datasn++;
 486
 487         tcp_task->data_offset = be32_to_cpu(rhdr->offset);
 488         if (tcp_task->data_offset + tcp_conn->in.datalen > total_in_length) {
 489                 debug_tcp("%s: data_offset(%d) + data_len(%d) > total_length_in(%d)\n",
 490                           __func__, tcp_task->data_offset,
 491                           tcp_conn->in.datalen, total_in_length);
 492                 return ISCSI_ERR_DATA_OFFSET;
 493         }
 494
 495         conn->datain_pdus_cnt++;
 496         return 0;
 497 }
 498
 499 /**
 500  * iscsi_tcp_r2t_rsp - iSCSI R2T Response processing
 501  * @conn: iscsi connection
 502  * @task: scsi command task
 503  */
 504 static int iscsi_tcp_r2t_rsp(struct iscsi_conn *conn, struct iscsi_task *task)
 505 {
 506         struct iscsi_session *session = conn->session;
 507         struct iscsi_tcp_task *tcp_task = task->dd_data;
 508         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 509         struct iscsi_r2t_rsp *rhdr = (struct iscsi_r2t_rsp *)tcp_conn->in.hdr;
 510         struct iscsi_r2t_info *r2t;
 511         int r2tsn = be32_to_cpu(rhdr->r2tsn);
 512         int rc;
 513
 514         if (tcp_conn->in.datalen) {
 515                 iscsi_conn_printk(KERN_ERR, conn,
 516                                   "invalid R2t with datalen %d\n",
 517                                   tcp_conn->in.datalen);
 518                 return ISCSI_ERR_DATALEN;
 519         }
 520
 521         if (tcp_task->exp_datasn != r2tsn){
 522                 debug_tcp("%s: task->exp_datasn(%d) != rhdr->r2tsn(%d)\n",
 523                           __func__, tcp_task->exp_datasn, r2tsn);
 524                 return ISCSI_ERR_R2TSN;
 525         }
 526
 527         /* fill-in new R2T associated with the task */
 528         iscsi_update_cmdsn(session, (struct iscsi_nopin*)rhdr);
 529
 530         if (!task->sc || session->state != ISCSI_STATE_LOGGED_IN) {
 531                 iscsi_conn_printk(KERN_INFO, conn,
 532                                   "dropping R2T itt %d in recovery.\n",
 533                                   task->itt);
 534                 return 0;
 535         }
 536
 537         rc = __kfifo_get(tcp_task->r2tpool.queue, (void*)&r2t, sizeof(void*));
 538         if (!rc) {
 539                 iscsi_conn_printk(KERN_ERR, conn, "Could not allocate R2T. "
 540                                   "Target has sent more R2Ts than it "
 541                                   "negotiated for or driver has has leaked.\n");
 542                 return ISCSI_ERR_PROTO;
 543         }
 544
 545         r2t->exp_statsn = rhdr->statsn;
 546         r2t->data_length = be32_to_cpu(rhdr->data_length);
 547         if (r2t->data_length == 0) {
 548                 iscsi_conn_printk(KERN_ERR, conn,
 549                                   "invalid R2T with zero data len\n");
 550                 __kfifo_put(tcp_task->r2tpool.queue, (void*)&r2t,
 551                             sizeof(void*));
 552                 return ISCSI_ERR_DATALEN;
 553         }
 554
 555         if (r2t->data_length > session->max_burst)
 556                 debug_scsi("invalid R2T with data len %u and max burst %u."
 557                            "Attempting to execute request.\n",
 558                             r2t->data_length, session->max_burst);
 559
 560         r2t->data_offset = be32_to_cpu(rhdr->data_offset);
 561         if (r2t->data_offset + r2t->data_length > scsi_out(task->sc)->length) {
 562                 iscsi_conn_printk(KERN_ERR, conn,
 563                                   "invalid R2T with data len %u at offset %u "
 564                                   "and total length %d\n", r2t->data_length,
 565                                   r2t->data_offset, scsi_out(task->sc)->length);
 566                 __kfifo_put(tcp_task->r2tpool.queue, (void*)&r2t,
 567                             sizeof(void*));
 568                 return ISCSI_ERR_DATALEN;
 569         }
 570
 571         r2t->ttt = rhdr->ttt; /* no flip */
 572         r2t->datasn = 0;
 573         r2t->sent = 0;
 574
 575         tcp_task->exp_datasn = r2tsn + 1;
 576         __kfifo_put(tcp_task->r2tqueue, (void*)&r2t, sizeof(void*));
 577         conn->r2t_pdus_cnt++;
 578
 579         iscsi_requeue_task(task);
 580         return 0;
 581 }
 582
 583 /*
 584  * Handle incoming reply to DataIn command
 585  */
 586 static int
 587 iscsi_tcp_process_data_in(struct iscsi_tcp_conn *tcp_conn,
 588                           struct iscsi_segment *segment)
 589 {
 590         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 591         struct iscsi_hdr *hdr = tcp_conn->in.hdr;
 592         int rc;
 593
 594         if (!iscsi_tcp_dgst_verify(tcp_conn, segment))
 595                 return ISCSI_ERR_DATA_DGST;
 596
 597         /* check for non-exceptional status */
 598         if (hdr->flags & ISCSI_FLAG_DATA_STATUS) {
 599                 rc = iscsi_complete_pdu(conn, tcp_conn->in.hdr, NULL, 0);
 600                 if (rc)
 601                         return rc;
 602         }
 603
 604         iscsi_tcp_hdr_recv_prep(tcp_conn);
 605         return 0;
 606 }
 607
 608 /**
 609  * iscsi_tcp_hdr_dissect - process PDU header
 610  * @conn: iSCSI connection
 611  * @hdr: PDU header
 612  *
 613  * This function analyzes the header of the PDU received,
 614  * and performs several sanity checks. If the PDU is accompanied
 615  * by data, the receive buffer is set up to copy the incoming data
 616  * to the correct location.
 617  */
 618 static int
 619 iscsi_tcp_hdr_dissect(struct iscsi_conn *conn, struct iscsi_hdr *hdr)
 620 {
 621         int rc = 0, opcode, ahslen;
 622         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 623         struct iscsi_task *task;
 624
 625         /* verify PDU length */
 626         tcp_conn->in.datalen = ntoh24(hdr->dlength);
 627         if (tcp_conn->in.datalen > conn->max_recv_dlength) {
 628                 iscsi_conn_printk(KERN_ERR, conn,
 629                                   "iscsi_tcp: datalen %d > %d\n",
 630                                   tcp_conn->in.datalen, conn->max_recv_dlength);
 631                 return ISCSI_ERR_DATALEN;
 632         }
 633
 634         /* Additional header segments. So far, we don't
 635          * process additional headers.
 636          */
 637         ahslen = hdr->hlength << 2;
 638
 639         opcode = hdr->opcode & ISCSI_OPCODE_MASK;
 640         /* verify itt (itt encoding: age+cid+itt) */
 641         rc = iscsi_verify_itt(conn, hdr->itt);
 642         if (rc)
 643                 return rc;
 644
 645         debug_tcp("opcode 0x%x ahslen %d datalen %d\n",
 646                   opcode, ahslen, tcp_conn->in.datalen);
 647
 648         switch(opcode) {
 649         case ISCSI_OP_SCSI_DATA_IN:
 650                 spin_lock(&conn->session->lock);
 651                 task = iscsi_itt_to_ctask(conn, hdr->itt);
 652                 if (!task)
 653                         rc = ISCSI_ERR_BAD_ITT;
 654                 else
 655                         rc = iscsi_tcp_data_in(conn, task);
 656                 if (rc) {
 657                         spin_unlock(&conn->session->lock);
 658                         break;
 659                 }
 660
 661                 if (tcp_conn->in.datalen) {
 662                         struct iscsi_tcp_task *tcp_task = task->dd_data;
 663                         struct hash_desc *rx_hash = NULL;
 664                         struct scsi_data_buffer *sdb = scsi_in(task->sc);
 665
 666                         /*
 667                          * Setup copy of Data-In into the Scsi_Cmnd
 668                          * Scatterlist case:
 669                          * We set up the iscsi_segment to point to the next
 670                          * scatterlist entry to copy to. As we go along,
 671                          * we move on to the next scatterlist entry and
 672                          * update the digest per-entry.
 673                          */
 674                         if (conn->datadgst_en &&
 675                             !(conn->session->tt->caps & CAP_DIGEST_OFFLOAD))
 676                                 rx_hash = tcp_conn->rx_hash;
 677
 678                         debug_tcp("iscsi_tcp_begin_data_in(%p, offset=%d, "
 679                                   "datalen=%d)\n", tcp_conn,
 680                                   tcp_task->data_offset,
 681                                   tcp_conn->in.datalen);
 682                         rc = iscsi_segment_seek_sg(&tcp_conn->in.segment,
 683                                                    sdb->table.sgl,
 684                                                    sdb->table.nents,
 685                                                    tcp_task->data_offset,
 686                                                    tcp_conn->in.datalen,
 687                                                    iscsi_tcp_process_data_in,
 688                                                    rx_hash);
 689                         spin_unlock(&conn->session->lock);
 690                         return rc;
 691                 }
 692                 rc = __iscsi_complete_pdu(conn, hdr, NULL, 0);
 693                 spin_unlock(&conn->session->lock);
 694                 break;
 695         case ISCSI_OP_SCSI_CMD_RSP:
 696                 if (tcp_conn->in.datalen) {
 697                         iscsi_tcp_data_recv_prep(tcp_conn);
 698                         return 0;
 699                 }
 700                 rc = iscsi_complete_pdu(conn, hdr, NULL, 0);
 701                 break;
 702         case ISCSI_OP_R2T:
 703                 spin_lock(&conn->session->lock);
 704                 task = iscsi_itt_to_ctask(conn, hdr->itt);
 705                 if (!task)
 706                         rc = ISCSI_ERR_BAD_ITT;
 707                 else if (ahslen)
 708                         rc = ISCSI_ERR_AHSLEN;
 709                 else if (task->sc->sc_data_direction == DMA_TO_DEVICE)
 710                         rc = iscsi_tcp_r2t_rsp(conn, task);
 711                 else
 712                         rc = ISCSI_ERR_PROTO;
 713                 spin_unlock(&conn->session->lock);
 714                 break;
 715         case ISCSI_OP_LOGIN_RSP:
 716         case ISCSI_OP_TEXT_RSP:
 717         case ISCSI_OP_REJECT:
 718         case ISCSI_OP_ASYNC_EVENT:
 719                 /*
 720                  * It is possible that we could get a PDU with a buffer larger
 721                  * than 8K, but there are no targets that currently do this.
 722                  * For now we fail until we find a vendor that needs it
 723                  */
 724                 if (ISCSI_DEF_MAX_RECV_SEG_LEN < tcp_conn->in.datalen) {
 725                         iscsi_conn_printk(KERN_ERR, conn,
 726                                           "iscsi_tcp: received buffer of "
 727                                           "len %u but conn buffer is only %u "
 728                                           "(opcode %0x)\n",
 729                                           tcp_conn->in.datalen,
 730                                           ISCSI_DEF_MAX_RECV_SEG_LEN, opcode);
 731                         rc = ISCSI_ERR_PROTO;
 732                         break;
 733                 }
 734
 735                 /* If there's data coming in with the response,
 736                  * receive it to the connection's buffer.
 737                  */
 738                 if (tcp_conn->in.datalen) {
 739                         iscsi_tcp_data_recv_prep(tcp_conn);
 740                         return 0;
 741                 }
 742         /* fall through */
 743         case ISCSI_OP_LOGOUT_RSP:
 744         case ISCSI_OP_NOOP_IN:
 745         case ISCSI_OP_SCSI_TMFUNC_RSP:
 746                 rc = iscsi_complete_pdu(conn, hdr, NULL, 0);
 747                 break;
 748         default:
 749                 rc = ISCSI_ERR_BAD_OPCODE;
 750                 break;
 751         }
 752
 753         if (rc == 0) {
 754                 /* Anything that comes with data should have
 755                  * been handled above. */
 756                 if (tcp_conn->in.datalen)
 757                         return ISCSI_ERR_PROTO;
 758                 iscsi_tcp_hdr_recv_prep(tcp_conn);
 759         }
 760
 761         return rc;
 762 }
 763
 764 /**
 765  * iscsi_tcp_hdr_recv_done - process PDU header
 766  *
 767  * This is the callback invoked when the PDU header has
 768  * been received. If the header is followed by additional
 769  * header segments, we go back for more data.
 770  */
 771 static int
 772 iscsi_tcp_hdr_recv_done(struct iscsi_tcp_conn *tcp_conn,
 773                         struct iscsi_segment *segment)
 774 {
 775         struct iscsi_conn *conn = tcp_conn->iscsi_conn;
 776         struct iscsi_hdr *hdr;
 777
 778         /* Check if there are additional header segments
 779          * *prior* to computing the digest, because we
 780          * may need to go back to the caller for more.
 781          */
 782         hdr = (struct iscsi_hdr *) tcp_conn->in.hdr_buf;
 783         if (segment->copied == sizeof(struct iscsi_hdr) && hdr->hlength) {
 784                 /* Bump the header length - the caller will
 785                  * just loop around and get the AHS for us, and
 786                  * call again. */
 787                 unsigned int ahslen = hdr->hlength << 2;
 788
 789                 /* Make sure we don't overflow */
 790                 if (sizeof(*hdr) + ahslen > sizeof(tcp_conn->in.hdr_buf))
 791                         return ISCSI_ERR_AHSLEN;
 792
 793                 segment->total_size += ahslen;
 794                 segment->size += ahslen;
 795                 return 0;
 796         }
 797
 798         /* We're done processing the header. See if we're doing
 799          * header digests; if so, set up the recv_digest buffer
 800          * and go back for more. */
 801         if (conn->hdrdgst_en &&
 802             !(conn->session->tt->caps & CAP_DIGEST_OFFLOAD)) {
 803                 if (segment->digest_len == 0) {
 804                         /*
 805                          * Even if we offload the digest processing we
 806                          * splice it in so we can increment the skb/segment
 807                          * counters in preparation for the data segment.
 808                          */
 809                         iscsi_tcp_segment_splice_digest(segment,
 810                                                         segment->recv_digest);
 811                         return 0;
 812                 }
 813
 814                 iscsi_tcp_dgst_header(tcp_conn->rx_hash, hdr,
 815                                       segment->total_copied - ISCSI_DIGEST_SIZE,
 816                                       segment->digest);
 817
 818                 if (!iscsi_tcp_dgst_verify(tcp_conn, segment))
 819                         return ISCSI_ERR_HDR_DGST;
 820         }
 821
 822         tcp_conn->in.hdr = hdr;
 823         return iscsi_tcp_hdr_dissect(conn, hdr);
 824 }
 825
 826 /**
 827  * iscsi_tcp_recv_segment_is_hdr - tests if we are reading in a header
 828  * @tcp_conn: iscsi tcp conn
 829  *
 830  * returns non zero if we are currently processing or setup to process
 831  * a header.
 832  */
 833 inline int iscsi_tcp_recv_segment_is_hdr(struct iscsi_tcp_conn *tcp_conn)
 834 {
 835         return tcp_conn->in.segment.done == iscsi_tcp_hdr_recv_done;
 836 }
 837 EXPORT_SYMBOL_GPL(iscsi_tcp_recv_segment_is_hdr);
 838
 839 /**
 840  * iscsi_tcp_recv_skb - Process skb
 841  * @conn: iscsi connection
 842  * @skb: network buffer with header and/or data segment
 843  * @offset: offset in skb
 844  * @offload: bool indicating if transfer was offloaded
 845  *
 846  * Will return status of transfer in status. And will return
 847  * number of bytes copied.
 848  */
 849 int iscsi_tcp_recv_skb(struct iscsi_conn *conn, struct sk_buff *skb,
 850                        unsigned int offset, bool offloaded, int *status)
 851 {
 852         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 853         struct iscsi_segment *segment = &tcp_conn->in.segment;
 854         struct skb_seq_state seq;
 855         unsigned int consumed = 0;
 856         int rc = 0;
 857
 858         debug_tcp("in %d bytes\n", skb->len - offset);
 859
 860         if (unlikely(conn->suspend_rx)) {
 861                 debug_tcp("conn %d Rx suspended!\n", conn->id);
 862                 *status = ISCSI_TCP_SUSPENDED;
 863                 return 0;
 864         }
 865
 866         if (offloaded) {
 867                 segment->total_copied = segment->total_size;
 868                 goto segment_done;
 869         }
 870
 871         skb_prepare_seq_read(skb, offset, skb->len, &seq);
 872         while (1) {
 873                 unsigned int avail;
 874                 const u8 *ptr;
 875
 876                 avail = skb_seq_read(consumed, &ptr, &seq);
 877                 if (avail == 0) {
 878                         debug_tcp("no more data avail. Consumed %d\n",
 879                                   consumed);
 880                         *status = ISCSI_TCP_SKB_DONE;
 881                         skb_abort_seq_read(&seq);
 882                         goto skb_done;
 883                 }
 884                 BUG_ON(segment->copied >= segment->size);
 885
 886                 debug_tcp("skb %p ptr=%p avail=%u\n", skb, ptr, avail);
 887                 rc = iscsi_tcp_segment_recv(tcp_conn, segment, ptr, avail);
 888                 BUG_ON(rc == 0);
 889                 consumed += rc;
 890
 891                 if (segment->total_copied >= segment->total_size) {
 892                         skb_abort_seq_read(&seq);
 893                         goto segment_done;
 894                 }
 895         }
 896
 897 segment_done:
 898         *status = ISCSI_TCP_SEGMENT_DONE;
 899         debug_tcp("segment done\n");
 900         rc = segment->done(tcp_conn, segment);
 901         if (rc != 0) {
 902                 *status = ISCSI_TCP_CONN_ERR;
 903                 debug_tcp("Error receiving PDU, errno=%d\n", rc);
 904                 iscsi_conn_failure(conn, rc);
 905                 return 0;
 906         }
 907         /* The done() functions sets up the next segment. */
 908
 909 skb_done:
 910         conn->rxdata_octets += consumed;
 911         return consumed;
 912 }
 913 EXPORT_SYMBOL_GPL(iscsi_tcp_recv_skb);
 914
 915 /**
 916  * iscsi_tcp_task_init - Initialize iSCSI SCSI_READ or SCSI_WRITE commands
 917  * @conn: iscsi connection
 918  * @task: scsi command task
 919  * @sc: scsi command
 920  */
 921 int iscsi_tcp_task_init(struct iscsi_task *task)
 922 {
 923         struct iscsi_tcp_task *tcp_task = task->dd_data;
 924         struct iscsi_conn *conn = task->conn;
 925         struct scsi_cmnd *sc = task->sc;
 926         int err;
 927
 928         if (!sc) {
 929                 /*
 930                  * mgmt tasks do not have a scatterlist since they come
 931                  * in from the iscsi interface.
 932                  */
 933                 debug_scsi("mtask deq [cid %d itt 0x%x]\n", conn->id,
 934                            task->itt);
 935
 936                 return conn->session->tt->init_pdu(task, 0, task->data_count);
 937         }
 938
 939         BUG_ON(__kfifo_len(tcp_task->r2tqueue));
 940         tcp_task->exp_datasn = 0;
 941
 942         /* Prepare PDU, optionally w/ immediate data */
 943         debug_scsi("task deq [cid %d itt 0x%x imm %d unsol %d]\n",
 944                     conn->id, task->itt, task->imm_count,
 945                     task->unsol_r2t.data_length);
 946
 947         err = conn->session->tt->init_pdu(task, 0, task->imm_count);
 948         if (err)
 949                 return err;
 950         task->imm_count = 0;
 951         return 0;
 952 }
 953 EXPORT_SYMBOL_GPL(iscsi_tcp_task_init);
 954
 955 static struct iscsi_r2t_info *iscsi_tcp_get_curr_r2t(struct iscsi_task *task)
 956 {
 957         struct iscsi_session *session = task->conn->session;
 958         struct iscsi_tcp_task *tcp_task = task->dd_data;
 959         struct iscsi_r2t_info *r2t = NULL;
 960
 961         if (iscsi_task_has_unsol_data(task))
 962                 r2t = &task->unsol_r2t;
 963         else {
 964                 spin_lock_bh(&session->lock);
 965                 if (tcp_task->r2t) {
 966                         r2t = tcp_task->r2t;
 967                         /* Continue with this R2T? */
 968                         if (r2t->data_length <= r2t->sent) {
 969                                 debug_scsi("  done with r2t %p\n", r2t);
 970                                 __kfifo_put(tcp_task->r2tpool.queue,
 971                                             (void *)&tcp_task->r2t,
 972                                             sizeof(void *));
 973                                 tcp_task->r2t = r2t = NULL;
 974                         }
 975                 }
 976
 977                 if (r2t == NULL) {
 978                         __kfifo_get(tcp_task->r2tqueue,
 979                                     (void *)&tcp_task->r2t, sizeof(void *));
 980                         r2t = tcp_task->r2t;
 981                 }
 982                 spin_unlock_bh(&session->lock);
 983         }
 984
 985         return r2t;
 986 }
 987
 988 /**
 989  * iscsi_tcp_task_xmit - xmit normal PDU task
 990  * @task: iscsi command task
 991  *
 992  * We're expected to return 0 when everything was transmitted succesfully,
 993  * -EAGAIN if there's still data in the queue, or != 0 for any other kind
 994  * of error.
 995  */
 996 int iscsi_tcp_task_xmit(struct iscsi_task *task)
 997 {
 998         struct iscsi_conn *conn = task->conn;
 999         struct iscsi_session *session = conn->session;
1000         struct iscsi_r2t_info *r2t;
1001         int rc = 0;
1002
1003 flush:
1004         /* Flush any pending data first. */
1005         rc = session->tt->xmit_pdu(task);
1006         if (rc < 0)
1007                 return rc;
1008
1009         /* mgmt command */
1010         if (!task->sc) {
1011                 if (task->hdr->itt == RESERVED_ITT)
1012                         iscsi_put_task(task);
1013                 return 0;
1014         }
1015
1016         /* Are we done already? */
1017         if (task->sc->sc_data_direction != DMA_TO_DEVICE)
1018                 return 0;
1019
1020         r2t = iscsi_tcp_get_curr_r2t(task);
1021         if (r2t == NULL) {
1022                 /* Waiting for more R2Ts to arrive. */
1023                 debug_tcp("no R2Ts yet\n");
1024                 return 0;
1025         }
1026
1027         rc = conn->session->tt->alloc_pdu(task, ISCSI_OP_SCSI_DATA_OUT);
1028         if (rc)
1029                 return rc;
1030         iscsi_prep_data_out_pdu(task, r2t, (struct iscsi_data *) task->hdr);
1031
1032         debug_scsi("sol dout %p [dsn %d itt 0x%x doff %d dlen %d]\n",
1033                    r2t, r2t->datasn - 1, task->hdr->itt,
1034                    r2t->data_offset + r2t->sent, r2t->data_count);
1035
1036         rc = conn->session->tt->init_pdu(task, r2t->data_offset + r2t->sent,
1037                                          r2t->data_count);
1038         if (rc)
1039                 return rc;
1040         r2t->sent += r2t->data_count;
1041         goto flush;
1042 }
1043 EXPORT_SYMBOL_GPL(iscsi_tcp_task_xmit);
1044
1045 struct iscsi_cls_conn *
1046 iscsi_tcp_conn_setup(struct iscsi_cls_session *cls_session, int dd_data_size,
1047                       uint32_t conn_idx)
1048
1049 {
1050         struct iscsi_conn *conn;
1051         struct iscsi_cls_conn *cls_conn;
1052         struct iscsi_tcp_conn *tcp_conn;
1053
1054         cls_conn = iscsi_conn_setup(cls_session, sizeof(*tcp_conn), conn_idx);
1055         if (!cls_conn)
1056                 return NULL;
1057         conn = cls_conn->dd_data;
1058         /*
1059          * due to strange issues with iser these are not set
1060          * in iscsi_conn_setup
1061          */
1062         conn->max_recv_dlength = ISCSI_DEF_MAX_RECV_SEG_LEN;
1063
1064         tcp_conn = conn->dd_data;
1065         tcp_conn->iscsi_conn = conn;
1066
1067         tcp_conn->dd_data = kzalloc(dd_data_size, GFP_KERNEL);
1068         if (!tcp_conn->dd_data) {
1069                 iscsi_conn_teardown(cls_conn);
1070                 return NULL;
1071         }
1072         return cls_conn;
1073 }
1074 EXPORT_SYMBOL_GPL(iscsi_tcp_conn_setup);
1075
1076 void iscsi_tcp_conn_teardown(struct iscsi_cls_conn *cls_conn)
1077 {
1078         struct iscsi_conn *conn = cls_conn->dd_data;
1079         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1080
1081         kfree(tcp_conn->dd_data);
1082         iscsi_conn_teardown(cls_conn);
1083 }
1084 EXPORT_SYMBOL_GPL(iscsi_tcp_conn_teardown);
1085
1086 int iscsi_tcp_r2tpool_alloc(struct iscsi_session *session)
1087 {
1088         int i;
1089         int cmd_i;
1090
1091         /*
1092          * initialize per-task: R2T pool and xmit queue
1093          */
1094         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
1095                 struct iscsi_task *task = session->cmds[cmd_i];
1096                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1097
1098                 /*
1099                  * pre-allocated x2 as much r2ts to handle race when
1100                  * target acks DataOut faster than we data_xmit() queues
1101                  * could replenish r2tqueue.
1102                  */
1103
1104                 /* R2T pool */
1105                 if (iscsi_pool_init(&tcp_task->r2tpool,
1106                                     session->max_r2t * 2, NULL,
1107                                     sizeof(struct iscsi_r2t_info))) {
1108                         goto r2t_alloc_fail;
1109                 }
1110
1111                 /* R2T xmit queue */
1112                 tcp_task->r2tqueue = kfifo_alloc(
1113                       session->max_r2t * 4 * sizeof(void*), GFP_KERNEL, NULL);
1114                 if (tcp_task->r2tqueue == ERR_PTR(-ENOMEM)) {
1115                         iscsi_pool_free(&tcp_task->r2tpool);
1116                         goto r2t_alloc_fail;
1117                 }
1118         }
1119
1120         return 0;
1121
1122 r2t_alloc_fail:
1123         for (i = 0; i < cmd_i; i++) {
1124                 struct iscsi_task *task = session->cmds[i];
1125                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1126
1127                 kfifo_free(tcp_task->r2tqueue);
1128                 iscsi_pool_free(&tcp_task->r2tpool);
1129         }
1130         return -ENOMEM;
1131 }
1132 EXPORT_SYMBOL_GPL(iscsi_tcp_r2tpool_alloc);
1133
1134 void iscsi_tcp_r2tpool_free(struct iscsi_session *session)
1135 {
1136         int i;
1137
1138         for (i = 0; i < session->cmds_max; i++) {
1139                 struct iscsi_task *task = session->cmds[i];
1140                 struct iscsi_tcp_task *tcp_task = task->dd_data;
1141
1142                 kfifo_free(tcp_task->r2tqueue);
1143                 iscsi_pool_free(&tcp_task->r2tpool);
1144         }
1145 }
1146 EXPORT_SYMBOL_GPL(iscsi_tcp_r2tpool_free);
1147
1148 void iscsi_tcp_conn_get_stats(struct iscsi_cls_conn *cls_conn,
1149                               struct iscsi_stats *stats)
1150 {
1151         struct iscsi_conn *conn = cls_conn->dd_data;
1152
1153         stats->txdata_octets = conn->txdata_octets;
1154         stats->rxdata_octets = conn->rxdata_octets;
1155         stats->scsicmd_pdus = conn->scsicmd_pdus_cnt;
1156         stats->dataout_pdus = conn->dataout_pdus_cnt;
1157         stats->scsirsp_pdus = conn->scsirsp_pdus_cnt;
1158         stats->datain_pdus = conn->datain_pdus_cnt;
1159         stats->r2t_pdus = conn->r2t_pdus_cnt;
1160         stats->tmfcmd_pdus = conn->tmfcmd_pdus_cnt;
1161         stats->tmfrsp_pdus = conn->tmfrsp_pdus_cnt;
1162 }
1163 EXPORT_SYMBOL_GPL(iscsi_tcp_conn_get_stats);