SAFE public projects git trees. - safe/jmp/linux-2.6/blob - drivers/scsi/iscsi_tcp.c

   1 /*
   2  * iSCSI Initiator over TCP/IP Data-Path
   3  *
   4  * Copyright (C) 2004 Dmitry Yusupov
   5  * Copyright (C) 2004 Alex Aizman
   6  * Copyright (C) 2005 - 2006 Mike Christie
   7  * Copyright (C) 2006 Red Hat, Inc.  All rights reserved.
   8  * maintained by open-iscsi@googlegroups.com
   9  *
  10  * This program is free software; you can redistribute it and/or modify
  11  * it under the terms of the GNU General Public License as published
  12  * by the Free Software Foundation; either version 2 of the License, or
  13  * (at your option) any later version.
  14  *
  15  * This program is distributed in the hope that it will be useful, but
  16  * WITHOUT ANY WARRANTY; without even the implied warranty of
  17  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  18  * General Public License for more details.
  19  *
  20  * See the file COPYING included with this distribution for more details.
  21  *
  22  * Credits:
  23  *      Christoph Hellwig
  24  *      FUJITA Tomonori
  25  *      Arne Redlich
  26  *      Zhenyu Wang
  27  */
  28
  29 #include <linux/types.h>
  30 #include <linux/list.h>
  31 #include <linux/inet.h>
  32 #include <linux/blkdev.h>
  33 #include <linux/crypto.h>
  34 #include <linux/delay.h>
  35 #include <linux/kfifo.h>
  36 #include <linux/scatterlist.h>
  37 #include <linux/mutex.h>
  38 #include <net/tcp.h>
  39 #include <scsi/scsi_cmnd.h>
  40 #include <scsi/scsi_host.h>
  41 #include <scsi/scsi.h>
  42 #include <scsi/scsi_transport_iscsi.h>
  43
  44 #include "iscsi_tcp.h"
  45
  46 MODULE_AUTHOR("Dmitry Yusupov <dmitry_yus@yahoo.com>, "
  47               "Alex Aizman <itn780@yahoo.com>");
  48 MODULE_DESCRIPTION("iSCSI/TCP data-path");
  49 MODULE_LICENSE("GPL");
  50 /* #define DEBUG_TCP */
  51 #define DEBUG_ASSERT
  52
  53 #ifdef DEBUG_TCP
  54 #define debug_tcp(fmt...) printk(KERN_INFO "tcp: " fmt)
  55 #else
  56 #define debug_tcp(fmt...)
  57 #endif
  58
  59 #ifndef DEBUG_ASSERT
  60 #ifdef BUG_ON
  61 #undef BUG_ON
  62 #endif
  63 #define BUG_ON(expr)
  64 #endif
  65
  66 static unsigned int iscsi_max_lun = 512;
  67 module_param_named(max_lun, iscsi_max_lun, uint, S_IRUGO);
  68
  69 static inline void
  70 iscsi_buf_init_iov(struct iscsi_buf *ibuf, char *vbuf, int size)
  71 {
  72         ibuf->sg.page = virt_to_page(vbuf);
  73         ibuf->sg.offset = offset_in_page(vbuf);
  74         ibuf->sg.length = size;
  75         ibuf->sent = 0;
  76         ibuf->use_sendmsg = 1;
  77 }
  78
  79 static inline void
  80 iscsi_buf_init_sg(struct iscsi_buf *ibuf, struct scatterlist *sg)
  81 {
  82         ibuf->sg.page = sg->page;
  83         ibuf->sg.offset = sg->offset;
  84         ibuf->sg.length = sg->length;
  85         /*
  86          * Fastpath: sg element fits into single page
  87          */
  88         if (sg->length + sg->offset <= PAGE_SIZE && !PageSlab(sg->page))
  89                 ibuf->use_sendmsg = 0;
  90         else
  91                 ibuf->use_sendmsg = 1;
  92         ibuf->sent = 0;
  93 }
  94
  95 static inline int
  96 iscsi_buf_left(struct iscsi_buf *ibuf)
  97 {
  98         int rc;
  99
 100         rc = ibuf->sg.length - ibuf->sent;
 101         BUG_ON(rc < 0);
 102         return rc;
 103 }
 104
 105 static inline void
 106 iscsi_hdr_digest(struct iscsi_conn *conn, struct iscsi_buf *buf,
 107                  u8* crc)
 108 {
 109         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 110
 111         crypto_hash_digest(&tcp_conn->tx_hash, &buf->sg, buf->sg.length, crc);
 112         buf->sg.length = tcp_conn->hdr_size;
 113 }
 114
 115 static inline int
 116 iscsi_hdr_extract(struct iscsi_tcp_conn *tcp_conn)
 117 {
 118         struct sk_buff *skb = tcp_conn->in.skb;
 119
 120         tcp_conn->in.zero_copy_hdr = 0;
 121
 122         if (tcp_conn->in.copy >= tcp_conn->hdr_size &&
 123             tcp_conn->in_progress == IN_PROGRESS_WAIT_HEADER) {
 124                 /*
 125                  * Zero-copy PDU Header: using connection context
 126                  * to store header pointer.
 127                  */
 128                 if (skb_shinfo(skb)->frag_list == NULL &&
 129                     !skb_shinfo(skb)->nr_frags) {
 130                         tcp_conn->in.hdr = (struct iscsi_hdr *)
 131                                 ((char*)skb->data + tcp_conn->in.offset);
 132                         tcp_conn->in.zero_copy_hdr = 1;
 133                 } else {
 134                         /* ignoring return code since we checked
 135                          * in.copy before */
 136                         skb_copy_bits(skb, tcp_conn->in.offset,
 137                                 &tcp_conn->hdr, tcp_conn->hdr_size);
 138                         tcp_conn->in.hdr = &tcp_conn->hdr;
 139                 }
 140                 tcp_conn->in.offset += tcp_conn->hdr_size;
 141                 tcp_conn->in.copy -= tcp_conn->hdr_size;
 142         } else {
 143                 int hdr_remains;
 144                 int copylen;
 145
 146                 /*
 147                  * PDU header scattered across SKB's,
 148                  * copying it... This'll happen quite rarely.
 149                  */
 150
 151                 if (tcp_conn->in_progress == IN_PROGRESS_WAIT_HEADER)
 152                         tcp_conn->in.hdr_offset = 0;
 153
 154                 hdr_remains = tcp_conn->hdr_size - tcp_conn->in.hdr_offset;
 155                 BUG_ON(hdr_remains <= 0);
 156
 157                 copylen = min(tcp_conn->in.copy, hdr_remains);
 158                 skb_copy_bits(skb, tcp_conn->in.offset,
 159                         (char*)&tcp_conn->hdr + tcp_conn->in.hdr_offset,
 160                         copylen);
 161
 162                 debug_tcp("PDU gather offset %d bytes %d in.offset %d "
 163                        "in.copy %d\n", tcp_conn->in.hdr_offset, copylen,
 164                        tcp_conn->in.offset, tcp_conn->in.copy);
 165
 166                 tcp_conn->in.offset += copylen;
 167                 tcp_conn->in.copy -= copylen;
 168                 if (copylen < hdr_remains)  {
 169                         tcp_conn->in_progress = IN_PROGRESS_HEADER_GATHER;
 170                         tcp_conn->in.hdr_offset += copylen;
 171                         return -EAGAIN;
 172                 }
 173                 tcp_conn->in.hdr = &tcp_conn->hdr;
 174                 tcp_conn->discontiguous_hdr_cnt++;
 175                 tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
 176         }
 177
 178         return 0;
 179 }
 180
 181 /*
 182  * must be called with session lock
 183  */
 184 static void
 185 iscsi_tcp_cleanup_ctask(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
 186 {
 187         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 188         struct iscsi_r2t_info *r2t;
 189         struct scsi_cmnd *sc;
 190
 191         /* flush ctask's r2t queues */
 192         while (__kfifo_get(tcp_ctask->r2tqueue, (void*)&r2t, sizeof(void*))) {
 193                 __kfifo_put(tcp_ctask->r2tpool.queue, (void*)&r2t,
 194                             sizeof(void*));
 195                 debug_scsi("iscsi_tcp_cleanup_ctask pending r2t dropped\n");
 196         }
 197
 198         sc = ctask->sc;
 199         if (unlikely(!sc))
 200                 return;
 201
 202         tcp_ctask->xmstate = XMSTATE_IDLE;
 203         tcp_ctask->r2t = NULL;
 204 }
 205
 206 /**
 207  * iscsi_data_rsp - SCSI Data-In Response processing
 208  * @conn: iscsi connection
 209  * @ctask: scsi command task
 210  **/
 211 static int
 212 iscsi_data_rsp(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
 213 {
 214         int rc;
 215         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 216         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 217         struct iscsi_data_rsp *rhdr = (struct iscsi_data_rsp *)tcp_conn->in.hdr;
 218         struct iscsi_session *session = conn->session;
 219         struct scsi_cmnd *sc = ctask->sc;
 220         int datasn = be32_to_cpu(rhdr->datasn);
 221
 222         rc = iscsi_check_assign_cmdsn(session, (struct iscsi_nopin*)rhdr);
 223         if (rc)
 224                 return rc;
 225         /*
 226          * setup Data-In byte counter (gets decremented..)
 227          */
 228         ctask->data_count = tcp_conn->in.datalen;
 229
 230         if (tcp_conn->in.datalen == 0)
 231                 return 0;
 232
 233         if (tcp_ctask->exp_datasn != datasn) {
 234                 debug_tcp("%s: ctask->exp_datasn(%d) != rhdr->datasn(%d)\n",
 235                           __FUNCTION__, tcp_ctask->exp_datasn, datasn);
 236                 return ISCSI_ERR_DATASN;
 237         }
 238
 239         tcp_ctask->exp_datasn++;
 240
 241         tcp_ctask->data_offset = be32_to_cpu(rhdr->offset);
 242         if (tcp_ctask->data_offset + tcp_conn->in.datalen > sc->request_bufflen) {
 243                 debug_tcp("%s: data_offset(%d) + data_len(%d) > total_length_in(%d)\n",
 244                           __FUNCTION__, tcp_ctask->data_offset,
 245                           tcp_conn->in.datalen, sc->request_bufflen);
 246                 return ISCSI_ERR_DATA_OFFSET;
 247         }
 248
 249         if (rhdr->flags & ISCSI_FLAG_DATA_STATUS) {
 250                 conn->exp_statsn = be32_to_cpu(rhdr->statsn) + 1;
 251                 if (rhdr->flags & ISCSI_FLAG_DATA_UNDERFLOW) {
 252                         int res_count = be32_to_cpu(rhdr->residual_count);
 253
 254                         if (res_count > 0 &&
 255                             res_count <= sc->request_bufflen) {
 256                                 sc->resid = res_count;
 257                                 sc->result = (DID_OK << 16) | rhdr->cmd_status;
 258                         } else
 259                                 sc->result = (DID_BAD_TARGET << 16) |
 260                                         rhdr->cmd_status;
 261                 } else if (rhdr->flags & ISCSI_FLAG_DATA_OVERFLOW) {
 262                         sc->resid = be32_to_cpu(rhdr->residual_count);
 263                         sc->result = (DID_OK << 16) | rhdr->cmd_status;
 264                 } else
 265                         sc->result = (DID_OK << 16) | rhdr->cmd_status;
 266         }
 267
 268         conn->datain_pdus_cnt++;
 269         return 0;
 270 }
 271
 272 /**
 273  * iscsi_solicit_data_init - initialize first Data-Out
 274  * @conn: iscsi connection
 275  * @ctask: scsi command task
 276  * @r2t: R2T info
 277  *
 278  * Notes:
 279  *      Initialize first Data-Out within this R2T sequence and finds
 280  *      proper data_offset within this SCSI command.
 281  *
 282  *      This function is called with connection lock taken.
 283  **/
 284 static void
 285 iscsi_solicit_data_init(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask,
 286                         struct iscsi_r2t_info *r2t)
 287 {
 288         struct iscsi_data *hdr;
 289         struct scsi_cmnd *sc = ctask->sc;
 290
 291         hdr = &r2t->dtask.hdr;
 292         memset(hdr, 0, sizeof(struct iscsi_data));
 293         hdr->ttt = r2t->ttt;
 294         hdr->datasn = cpu_to_be32(r2t->solicit_datasn);
 295         r2t->solicit_datasn++;
 296         hdr->opcode = ISCSI_OP_SCSI_DATA_OUT;
 297         memcpy(hdr->lun, ctask->hdr->lun, sizeof(hdr->lun));
 298         hdr->itt = ctask->hdr->itt;
 299         hdr->exp_statsn = r2t->exp_statsn;
 300         hdr->offset = cpu_to_be32(r2t->data_offset);
 301         if (r2t->data_length > conn->max_xmit_dlength) {
 302                 hton24(hdr->dlength, conn->max_xmit_dlength);
 303                 r2t->data_count = conn->max_xmit_dlength;
 304                 hdr->flags = 0;
 305         } else {
 306                 hton24(hdr->dlength, r2t->data_length);
 307                 r2t->data_count = r2t->data_length;
 308                 hdr->flags = ISCSI_FLAG_CMD_FINAL;
 309         }
 310         conn->dataout_pdus_cnt++;
 311
 312         r2t->sent = 0;
 313
 314         iscsi_buf_init_iov(&r2t->headbuf, (char*)hdr,
 315                            sizeof(struct iscsi_hdr));
 316
 317         if (sc->use_sg) {
 318                 int i, sg_count = 0;
 319                 struct scatterlist *sg = sc->request_buffer;
 320
 321                 r2t->sg = NULL;
 322                 for (i = 0; i < sc->use_sg; i++, sg += 1) {
 323                         /* FIXME: prefetch ? */
 324                         if (sg_count + sg->length > r2t->data_offset) {
 325                                 int page_offset;
 326
 327                                 /* sg page found! */
 328
 329                                 /* offset within this page */
 330                                 page_offset = r2t->data_offset - sg_count;
 331
 332                                 /* fill in this buffer */
 333                                 iscsi_buf_init_sg(&r2t->sendbuf, sg);
 334                                 r2t->sendbuf.sg.offset += page_offset;
 335                                 r2t->sendbuf.sg.length -= page_offset;
 336
 337                                 /* xmit logic will continue with next one */
 338                                 r2t->sg = sg + 1;
 339                                 break;
 340                         }
 341                         sg_count += sg->length;
 342                 }
 343                 BUG_ON(r2t->sg == NULL);
 344         } else {
 345                 iscsi_buf_init_iov(&r2t->sendbuf,
 346                             (char*)sc->request_buffer + r2t->data_offset,
 347                             r2t->data_count);
 348                 r2t->sg = NULL;
 349         }
 350 }
 351
 352 /**
 353  * iscsi_r2t_rsp - iSCSI R2T Response processing
 354  * @conn: iscsi connection
 355  * @ctask: scsi command task
 356  **/
 357 static int
 358 iscsi_r2t_rsp(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
 359 {
 360         struct iscsi_r2t_info *r2t;
 361         struct iscsi_session *session = conn->session;
 362         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 363         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 364         struct iscsi_r2t_rsp *rhdr = (struct iscsi_r2t_rsp *)tcp_conn->in.hdr;
 365         int r2tsn = be32_to_cpu(rhdr->r2tsn);
 366         int rc;
 367
 368         if (tcp_conn->in.datalen) {
 369                 printk(KERN_ERR "iscsi_tcp: invalid R2t with datalen %d\n",
 370                        tcp_conn->in.datalen);
 371                 return ISCSI_ERR_DATALEN;
 372         }
 373
 374         if (tcp_ctask->exp_datasn != r2tsn){
 375                 debug_tcp("%s: ctask->exp_datasn(%d) != rhdr->r2tsn(%d)\n",
 376                           __FUNCTION__, tcp_ctask->exp_datasn, r2tsn);
 377                 return ISCSI_ERR_R2TSN;
 378         }
 379
 380         rc = iscsi_check_assign_cmdsn(session, (struct iscsi_nopin*)rhdr);
 381         if (rc)
 382                 return rc;
 383
 384         /* fill-in new R2T associated with the task */
 385         spin_lock(&session->lock);
 386         if (!ctask->sc || ctask->mtask ||
 387              session->state != ISCSI_STATE_LOGGED_IN) {
 388                 printk(KERN_INFO "iscsi_tcp: dropping R2T itt %d in "
 389                        "recovery...\n", ctask->itt);
 390                 spin_unlock(&session->lock);
 391                 return 0;
 392         }
 393
 394         rc = __kfifo_get(tcp_ctask->r2tpool.queue, (void*)&r2t, sizeof(void*));
 395         BUG_ON(!rc);
 396
 397         r2t->exp_statsn = rhdr->statsn;
 398         r2t->data_length = be32_to_cpu(rhdr->data_length);
 399         if (r2t->data_length == 0) {
 400                 printk(KERN_ERR "iscsi_tcp: invalid R2T with zero data len\n");
 401                 spin_unlock(&session->lock);
 402                 return ISCSI_ERR_DATALEN;
 403         }
 404
 405         if (r2t->data_length > session->max_burst)
 406                 debug_scsi("invalid R2T with data len %u and max burst %u."
 407                            "Attempting to execute request.\n",
 408                             r2t->data_length, session->max_burst);
 409
 410         r2t->data_offset = be32_to_cpu(rhdr->data_offset);
 411         if (r2t->data_offset + r2t->data_length > ctask->sc->request_bufflen) {
 412                 spin_unlock(&session->lock);
 413                 printk(KERN_ERR "iscsi_tcp: invalid R2T with data len %u at "
 414                        "offset %u and total length %d\n", r2t->data_length,
 415                        r2t->data_offset, ctask->sc->request_bufflen);
 416                 return ISCSI_ERR_DATALEN;
 417         }
 418
 419         r2t->ttt = rhdr->ttt; /* no flip */
 420         r2t->solicit_datasn = 0;
 421
 422         iscsi_solicit_data_init(conn, ctask, r2t);
 423
 424         tcp_ctask->exp_datasn = r2tsn + 1;
 425         __kfifo_put(tcp_ctask->r2tqueue, (void*)&r2t, sizeof(void*));
 426         tcp_ctask->xmstate |= XMSTATE_SOL_HDR;
 427         list_move_tail(&ctask->running, &conn->xmitqueue);
 428
 429         scsi_queue_work(session->host, &conn->xmitwork);
 430         conn->r2t_pdus_cnt++;
 431         spin_unlock(&session->lock);
 432
 433         return 0;
 434 }
 435
 436 static int
 437 iscsi_tcp_hdr_recv(struct iscsi_conn *conn)
 438 {
 439         int rc = 0, opcode, ahslen;
 440         struct iscsi_hdr *hdr;
 441         struct iscsi_session *session = conn->session;
 442         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 443         uint32_t cdgst, rdgst = 0, itt;
 444
 445         hdr = tcp_conn->in.hdr;
 446
 447         /* verify PDU length */
 448         tcp_conn->in.datalen = ntoh24(hdr->dlength);
 449         if (tcp_conn->in.datalen > conn->max_recv_dlength) {
 450                 printk(KERN_ERR "iscsi_tcp: datalen %d > %d\n",
 451                        tcp_conn->in.datalen, conn->max_recv_dlength);
 452                 return ISCSI_ERR_DATALEN;
 453         }
 454         tcp_conn->data_copied = 0;
 455
 456         /* read AHS */
 457         ahslen = hdr->hlength << 2;
 458         tcp_conn->in.offset += ahslen;
 459         tcp_conn->in.copy -= ahslen;
 460         if (tcp_conn->in.copy < 0) {
 461                 printk(KERN_ERR "iscsi_tcp: can't handle AHS with length "
 462                        "%d bytes\n", ahslen);
 463                 return ISCSI_ERR_AHSLEN;
 464         }
 465
 466         /* calculate read padding */
 467         tcp_conn->in.padding = tcp_conn->in.datalen & (ISCSI_PAD_LEN-1);
 468         if (tcp_conn->in.padding) {
 469                 tcp_conn->in.padding = ISCSI_PAD_LEN - tcp_conn->in.padding;
 470                 debug_scsi("read padding %d bytes\n", tcp_conn->in.padding);
 471         }
 472
 473         if (conn->hdrdgst_en) {
 474                 struct scatterlist sg;
 475
 476                 sg_init_one(&sg, (u8 *)hdr,
 477                             sizeof(struct iscsi_hdr) + ahslen);
 478                 crypto_hash_digest(&tcp_conn->rx_hash, &sg, sg.length,
 479                                    (u8 *)&cdgst);
 480                 rdgst = *(uint32_t*)((char*)hdr + sizeof(struct iscsi_hdr) +
 481                                      ahslen);
 482                 if (cdgst != rdgst) {
 483                         printk(KERN_ERR "iscsi_tcp: hdrdgst error "
 484                                "recv 0x%x calc 0x%x\n", rdgst, cdgst);
 485                         return ISCSI_ERR_HDR_DGST;
 486                 }
 487         }
 488
 489         opcode = hdr->opcode & ISCSI_OPCODE_MASK;
 490         /* verify itt (itt encoding: age+cid+itt) */
 491         rc = iscsi_verify_itt(conn, hdr, &itt);
 492         if (rc == ISCSI_ERR_NO_SCSI_CMD) {
 493                 tcp_conn->in.datalen = 0; /* force drop */
 494                 return 0;
 495         } else if (rc)
 496                 return rc;
 497
 498         debug_tcp("opcode 0x%x offset %d copy %d ahslen %d datalen %d\n",
 499                   opcode, tcp_conn->in.offset, tcp_conn->in.copy,
 500                   ahslen, tcp_conn->in.datalen);
 501
 502         switch(opcode) {
 503         case ISCSI_OP_SCSI_DATA_IN:
 504                 tcp_conn->in.ctask = session->cmds[itt];
 505                 rc = iscsi_data_rsp(conn, tcp_conn->in.ctask);
 506                 if (rc)
 507                         return rc;
 508                 /* fall through */
 509         case ISCSI_OP_SCSI_CMD_RSP:
 510                 tcp_conn->in.ctask = session->cmds[itt];
 511                 if (tcp_conn->in.datalen)
 512                         goto copy_hdr;
 513
 514                 spin_lock(&session->lock);
 515                 rc = __iscsi_complete_pdu(conn, hdr, NULL, 0);
 516                 spin_unlock(&session->lock);
 517                 break;
 518         case ISCSI_OP_R2T:
 519                 tcp_conn->in.ctask = session->cmds[itt];
 520                 if (ahslen)
 521                         rc = ISCSI_ERR_AHSLEN;
 522                 else if (tcp_conn->in.ctask->sc->sc_data_direction ==
 523                                                                 DMA_TO_DEVICE)
 524                         rc = iscsi_r2t_rsp(conn, tcp_conn->in.ctask);
 525                 else
 526                         rc = ISCSI_ERR_PROTO;
 527                 break;
 528         case ISCSI_OP_LOGIN_RSP:
 529         case ISCSI_OP_TEXT_RSP:
 530         case ISCSI_OP_REJECT:
 531         case ISCSI_OP_ASYNC_EVENT:
 532                 /*
 533                  * It is possible that we could get a PDU with a buffer larger
 534                  * than 8K, but there are no targets that currently do this.
 535                  * For now we fail until we find a vendor that needs it
 536                  */
 537                 if (ISCSI_DEF_MAX_RECV_SEG_LEN <
 538                     tcp_conn->in.datalen) {
 539                         printk(KERN_ERR "iscsi_tcp: received buffer of len %u "
 540                               "but conn buffer is only %u (opcode %0x)\n",
 541                               tcp_conn->in.datalen,
 542                               ISCSI_DEF_MAX_RECV_SEG_LEN, opcode);
 543                         rc = ISCSI_ERR_PROTO;
 544                         break;
 545                 }
 546
 547                 if (tcp_conn->in.datalen)
 548                         goto copy_hdr;
 549         /* fall through */
 550         case ISCSI_OP_LOGOUT_RSP:
 551         case ISCSI_OP_NOOP_IN:
 552         case ISCSI_OP_SCSI_TMFUNC_RSP:
 553                 rc = iscsi_complete_pdu(conn, hdr, NULL, 0);
 554                 break;
 555         default:
 556                 rc = ISCSI_ERR_BAD_OPCODE;
 557                 break;
 558         }
 559
 560         return rc;
 561
 562 copy_hdr:
 563         /*
 564          * if we did zero copy for the header but we will need multiple
 565          * skbs to complete the command then we have to copy the header
 566          * for later use
 567          */
 568         if (tcp_conn->in.zero_copy_hdr && tcp_conn->in.copy <=
 569            (tcp_conn->in.datalen + tcp_conn->in.padding +
 570             (conn->datadgst_en ? 4 : 0))) {
 571                 debug_tcp("Copying header for later use. in.copy %d in.datalen"
 572                           " %d\n", tcp_conn->in.copy, tcp_conn->in.datalen);
 573                 memcpy(&tcp_conn->hdr, tcp_conn->in.hdr,
 574                        sizeof(struct iscsi_hdr));
 575                 tcp_conn->in.hdr = &tcp_conn->hdr;
 576                 tcp_conn->in.zero_copy_hdr = 0;
 577         }
 578         return 0;
 579 }
 580
 581 /**
 582  * iscsi_ctask_copy - copy skb bits to the destanation cmd task
 583  * @conn: iscsi tcp connection
 584  * @ctask: scsi command task
 585  * @buf: buffer to copy to
 586  * @buf_size: size of buffer
 587  * @offset: offset within the buffer
 588  *
 589  * Notes:
 590  *      The function calls skb_copy_bits() and updates per-connection and
 591  *      per-cmd byte counters.
 592  *
 593  *      Read counters (in bytes):
 594  *
 595  *      conn->in.offset         offset within in progress SKB
 596  *      conn->in.copy           left to copy from in progress SKB
 597  *                              including padding
 598  *      conn->in.copied         copied already from in progress SKB
 599  *      conn->data_copied       copied already from in progress buffer
 600  *      ctask->sent             total bytes sent up to the MidLayer
 601  *      ctask->data_count       left to copy from in progress Data-In
 602  *      buf_left                left to copy from in progress buffer
 603  **/
 604 static inline int
 605 iscsi_ctask_copy(struct iscsi_tcp_conn *tcp_conn, struct iscsi_cmd_task *ctask,
 606                 void *buf, int buf_size, int offset)
 607 {
 608         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 609         int buf_left = buf_size - (tcp_conn->data_copied + offset);
 610         unsigned size = min(tcp_conn->in.copy, buf_left);
 611         int rc;
 612
 613         size = min(size, ctask->data_count);
 614
 615         debug_tcp("ctask_copy %d bytes at offset %d copied %d\n",
 616                size, tcp_conn->in.offset, tcp_conn->in.copied);
 617
 618         BUG_ON(size <= 0);
 619         BUG_ON(tcp_ctask->sent + size > ctask->sc->request_bufflen);
 620
 621         rc = skb_copy_bits(tcp_conn->in.skb, tcp_conn->in.offset,
 622                            (char*)buf + (offset + tcp_conn->data_copied), size);
 623         /* must fit into skb->len */
 624         BUG_ON(rc);
 625
 626         tcp_conn->in.offset += size;
 627         tcp_conn->in.copy -= size;
 628         tcp_conn->in.copied += size;
 629         tcp_conn->data_copied += size;
 630         tcp_ctask->sent += size;
 631         ctask->data_count -= size;
 632
 633         BUG_ON(tcp_conn->in.copy < 0);
 634         BUG_ON(ctask->data_count < 0);
 635
 636         if (buf_size != (tcp_conn->data_copied + offset)) {
 637                 if (!ctask->data_count) {
 638                         BUG_ON(buf_size - tcp_conn->data_copied < 0);
 639                         /* done with this PDU */
 640                         return buf_size - tcp_conn->data_copied;
 641                 }
 642                 return -EAGAIN;
 643         }
 644
 645         /* done with this buffer or with both - PDU and buffer */
 646         tcp_conn->data_copied = 0;
 647         return 0;
 648 }
 649
 650 /**
 651  * iscsi_tcp_copy - copy skb bits to the destanation buffer
 652  * @conn: iscsi tcp connection
 653  *
 654  * Notes:
 655  *      The function calls skb_copy_bits() and updates per-connection
 656  *      byte counters.
 657  **/
 658 static inline int
 659 iscsi_tcp_copy(struct iscsi_conn *conn, int buf_size)
 660 {
 661         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 662         int buf_left = buf_size - tcp_conn->data_copied;
 663         int size = min(tcp_conn->in.copy, buf_left);
 664         int rc;
 665
 666         debug_tcp("tcp_copy %d bytes at offset %d copied %d\n",
 667                size, tcp_conn->in.offset, tcp_conn->data_copied);
 668         BUG_ON(size <= 0);
 669
 670         rc = skb_copy_bits(tcp_conn->in.skb, tcp_conn->in.offset,
 671                            (char*)conn->data + tcp_conn->data_copied, size);
 672         BUG_ON(rc);
 673
 674         tcp_conn->in.offset += size;
 675         tcp_conn->in.copy -= size;
 676         tcp_conn->in.copied += size;
 677         tcp_conn->data_copied += size;
 678
 679         if (buf_size != tcp_conn->data_copied)
 680                 return -EAGAIN;
 681
 682         return 0;
 683 }
 684
 685 static inline void
 686 partial_sg_digest_update(struct hash_desc *desc, struct scatterlist *sg,
 687                          int offset, int length)
 688 {
 689         struct scatterlist temp;
 690
 691         memcpy(&temp, sg, sizeof(struct scatterlist));
 692         temp.offset = offset;
 693         temp.length = length;
 694         crypto_hash_update(desc, &temp, length);
 695 }
 696
 697 static void
 698 iscsi_recv_digest_update(struct iscsi_tcp_conn *tcp_conn, char* buf, int len)
 699 {
 700         struct scatterlist tmp;
 701
 702         sg_init_one(&tmp, buf, len);
 703         crypto_hash_update(&tcp_conn->rx_hash, &tmp, len);
 704 }
 705
 706 static int iscsi_scsi_data_in(struct iscsi_conn *conn)
 707 {
 708         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 709         struct iscsi_cmd_task *ctask = tcp_conn->in.ctask;
 710         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 711         struct scsi_cmnd *sc = ctask->sc;
 712         struct scatterlist *sg;
 713         int i, offset, rc = 0;
 714
 715         BUG_ON((void*)ctask != sc->SCp.ptr);
 716
 717         /*
 718          * copying Data-In into the Scsi_Cmnd
 719          */
 720         if (!sc->use_sg) {
 721                 i = ctask->data_count;
 722                 rc = iscsi_ctask_copy(tcp_conn, ctask, sc->request_buffer,
 723                                       sc->request_bufflen,
 724                                       tcp_ctask->data_offset);
 725                 if (rc == -EAGAIN)
 726                         return rc;
 727                 if (conn->datadgst_en)
 728                         iscsi_recv_digest_update(tcp_conn, sc->request_buffer,
 729                                                  i);
 730                 rc = 0;
 731                 goto done;
 732         }
 733
 734         offset = tcp_ctask->data_offset;
 735         sg = sc->request_buffer;
 736
 737         if (tcp_ctask->data_offset)
 738                 for (i = 0; i < tcp_ctask->sg_count; i++)
 739                         offset -= sg[i].length;
 740         /* we've passed through partial sg*/
 741         if (offset < 0)
 742                 offset = 0;
 743
 744         for (i = tcp_ctask->sg_count; i < sc->use_sg; i++) {
 745                 char *dest;
 746
 747                 dest = kmap_atomic(sg[i].page, KM_SOFTIRQ0);
 748                 rc = iscsi_ctask_copy(tcp_conn, ctask, dest + sg[i].offset,
 749                                       sg[i].length, offset);
 750                 kunmap_atomic(dest, KM_SOFTIRQ0);
 751                 if (rc == -EAGAIN)
 752                         /* continue with the next SKB/PDU */
 753                         return rc;
 754                 if (!rc) {
 755                         if (conn->datadgst_en) {
 756                                 if (!offset)
 757                                         crypto_hash_update(
 758                                                         &tcp_conn->rx_hash,
 759                                                         &sg[i], sg[i].length);
 760                                 else
 761                                         partial_sg_digest_update(
 762                                                         &tcp_conn->rx_hash,
 763                                                         &sg[i],
 764                                                         sg[i].offset + offset,
 765                                                         sg[i].length - offset);
 766                         }
 767                         offset = 0;
 768                         tcp_ctask->sg_count++;
 769                 }
 770
 771                 if (!ctask->data_count) {
 772                         if (rc && conn->datadgst_en)
 773                                 /*
 774                                  * data-in is complete, but buffer not...
 775                                  */
 776                                 partial_sg_digest_update(&tcp_conn->rx_hash,
 777                                                          &sg[i],
 778                                                          sg[i].offset,
 779                                                          sg[i].length-rc);
 780                         rc = 0;
 781                         break;
 782                 }
 783
 784                 if (!tcp_conn->in.copy)
 785                         return -EAGAIN;
 786         }
 787         BUG_ON(ctask->data_count);
 788
 789 done:
 790         /* check for non-exceptional status */
 791         if (tcp_conn->in.hdr->flags & ISCSI_FLAG_DATA_STATUS) {
 792                 debug_scsi("done [sc %lx res %d itt 0x%x flags 0x%x]\n",
 793                            (long)sc, sc->result, ctask->itt,
 794                            tcp_conn->in.hdr->flags);
 795                 spin_lock(&conn->session->lock);
 796                 __iscsi_complete_pdu(conn, tcp_conn->in.hdr, NULL, 0);
 797                 spin_unlock(&conn->session->lock);
 798         }
 799
 800         return rc;
 801 }
 802
 803 static int
 804 iscsi_data_recv(struct iscsi_conn *conn)
 805 {
 806         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 807         int rc = 0, opcode;
 808
 809         opcode = tcp_conn->in.hdr->opcode & ISCSI_OPCODE_MASK;
 810         switch (opcode) {
 811         case ISCSI_OP_SCSI_DATA_IN:
 812                 rc = iscsi_scsi_data_in(conn);
 813                 break;
 814         case ISCSI_OP_SCSI_CMD_RSP:
 815         case ISCSI_OP_TEXT_RSP:
 816         case ISCSI_OP_LOGIN_RSP:
 817         case ISCSI_OP_ASYNC_EVENT:
 818         case ISCSI_OP_REJECT:
 819                 /*
 820                  * Collect data segment to the connection's data
 821                  * placeholder
 822                  */
 823                 if (iscsi_tcp_copy(conn, tcp_conn->in.datalen)) {
 824                         rc = -EAGAIN;
 825                         goto exit;
 826                 }
 827
 828                 rc = iscsi_complete_pdu(conn, tcp_conn->in.hdr, conn->data,
 829                                         tcp_conn->in.datalen);
 830                 if (!rc && conn->datadgst_en && opcode != ISCSI_OP_LOGIN_RSP)
 831                         iscsi_recv_digest_update(tcp_conn, conn->data,
 832                                                 tcp_conn->in.datalen);
 833                 break;
 834         default:
 835                 BUG_ON(1);
 836         }
 837 exit:
 838         return rc;
 839 }
 840
 841 /**
 842  * iscsi_tcp_data_recv - TCP receive in sendfile fashion
 843  * @rd_desc: read descriptor
 844  * @skb: socket buffer
 845  * @offset: offset in skb
 846  * @len: skb->len - offset
 847  **/
 848 static int
 849 iscsi_tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 850                 unsigned int offset, size_t len)
 851 {
 852         int rc;
 853         struct iscsi_conn *conn = rd_desc->arg.data;
 854         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 855         int processed;
 856         char pad[ISCSI_PAD_LEN];
 857         struct scatterlist sg;
 858
 859         /*
 860          * Save current SKB and its offset in the corresponding
 861          * connection context.
 862          */
 863         tcp_conn->in.copy = skb->len - offset;
 864         tcp_conn->in.offset = offset;
 865         tcp_conn->in.skb = skb;
 866         tcp_conn->in.len = tcp_conn->in.copy;
 867         BUG_ON(tcp_conn->in.copy <= 0);
 868         debug_tcp("in %d bytes\n", tcp_conn->in.copy);
 869
 870 more:
 871         tcp_conn->in.copied = 0;
 872         rc = 0;
 873
 874         if (unlikely(conn->suspend_rx)) {
 875                 debug_tcp("conn %d Rx suspended!\n", conn->id);
 876                 return 0;
 877         }
 878
 879         if (tcp_conn->in_progress == IN_PROGRESS_WAIT_HEADER ||
 880             tcp_conn->in_progress == IN_PROGRESS_HEADER_GATHER) {
 881                 rc = iscsi_hdr_extract(tcp_conn);
 882                 if (rc) {
 883                        if (rc == -EAGAIN)
 884                                 goto nomore;
 885                        else {
 886                                 iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
 887                                 return 0;
 888                        }
 889                 }
 890
 891                 /*
 892                  * Verify and process incoming PDU header.
 893                  */
 894                 rc = iscsi_tcp_hdr_recv(conn);
 895                 if (!rc && tcp_conn->in.datalen) {
 896                         if (conn->datadgst_en)
 897                                 crypto_hash_init(&tcp_conn->rx_hash);
 898                         tcp_conn->in_progress = IN_PROGRESS_DATA_RECV;
 899                 } else if (rc) {
 900                         iscsi_conn_failure(conn, rc);
 901                         return 0;
 902                 }
 903         }
 904
 905         if (tcp_conn->in_progress == IN_PROGRESS_DDIGEST_RECV) {
 906                 uint32_t recv_digest;
 907
 908                 debug_tcp("extra data_recv offset %d copy %d\n",
 909                           tcp_conn->in.offset, tcp_conn->in.copy);
 910                 rc = iscsi_tcp_copy(conn, sizeof(uint32_t));
 911                 if (rc) {
 912                         if (rc == -EAGAIN)
 913                                 goto again;
 914                         iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
 915                         return 0;
 916                 }
 917
 918                 memcpy(&recv_digest, conn->data, sizeof(uint32_t));
 919                 if (recv_digest != tcp_conn->in.datadgst) {
 920                         debug_tcp("iscsi_tcp: data digest error!"
 921                                   "0x%x != 0x%x\n", recv_digest,
 922                                   tcp_conn->in.datadgst);
 923                         iscsi_conn_failure(conn, ISCSI_ERR_DATA_DGST);
 924                         return 0;
 925                 } else {
 926                         debug_tcp("iscsi_tcp: data digest match!"
 927                                   "0x%x == 0x%x\n", recv_digest,
 928                                   tcp_conn->in.datadgst);
 929                         tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
 930                 }
 931         }
 932
 933         if (tcp_conn->in_progress == IN_PROGRESS_DATA_RECV &&
 934            tcp_conn->in.copy) {
 935
 936                 debug_tcp("data_recv offset %d copy %d\n",
 937                        tcp_conn->in.offset, tcp_conn->in.copy);
 938
 939                 rc = iscsi_data_recv(conn);
 940                 if (rc) {
 941                         if (rc == -EAGAIN)
 942                                 goto again;
 943                         iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
 944                         return 0;
 945                 }
 946                 tcp_conn->in.copy -= tcp_conn->in.padding;
 947                 tcp_conn->in.offset += tcp_conn->in.padding;
 948                 if (conn->datadgst_en) {
 949                         if (tcp_conn->in.padding) {
 950                                 debug_tcp("padding -> %d\n",
 951                                           tcp_conn->in.padding);
 952                                 memset(pad, 0, tcp_conn->in.padding);
 953                                 sg_init_one(&sg, pad, tcp_conn->in.padding);
 954                                 crypto_hash_update(&tcp_conn->rx_hash,
 955                                                    &sg, sg.length);
 956                         }
 957                         crypto_hash_final(&tcp_conn->rx_hash,
 958                                           (u8 *) &tcp_conn->in.datadgst);
 959                         debug_tcp("rx digest 0x%x\n", tcp_conn->in.datadgst);
 960                         tcp_conn->in_progress = IN_PROGRESS_DDIGEST_RECV;
 961                         tcp_conn->data_copied = 0;
 962                 } else
 963                         tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
 964         }
 965
 966         debug_tcp("f, processed %d from out of %d padding %d\n",
 967                tcp_conn->in.offset - offset, (int)len, tcp_conn->in.padding);
 968         BUG_ON(tcp_conn->in.offset - offset > len);
 969
 970         if (tcp_conn->in.offset - offset != len) {
 971                 debug_tcp("continue to process %d bytes\n",
 972                        (int)len - (tcp_conn->in.offset - offset));
 973                 goto more;
 974         }
 975
 976 nomore:
 977         processed = tcp_conn->in.offset - offset;
 978         BUG_ON(processed == 0);
 979         return processed;
 980
 981 again:
 982         processed = tcp_conn->in.offset - offset;
 983         debug_tcp("c, processed %d from out of %d rd_desc_cnt %d\n",
 984                   processed, (int)len, (int)rd_desc->count);
 985         BUG_ON(processed == 0);
 986         BUG_ON(processed > len);
 987
 988         conn->rxdata_octets += processed;
 989         return processed;
 990 }
 991
 992 static void
 993 iscsi_tcp_data_ready(struct sock *sk, int flag)
 994 {
 995         struct iscsi_conn *conn = sk->sk_user_data;
 996         read_descriptor_t rd_desc;
 997
 998         read_lock(&sk->sk_callback_lock);
 999
1000         /*
1001          * Use rd_desc to pass 'conn' to iscsi_tcp_data_recv.
1002          * We set count to 1 because we want the network layer to
1003          * hand us all the skbs that are available. iscsi_tcp_data_recv
1004          * handled pdus that cross buffers or pdus that still need data.
1005          */
1006         rd_desc.arg.data = conn;
1007         rd_desc.count = 1;
1008         tcp_read_sock(sk, &rd_desc, iscsi_tcp_data_recv);
1009
1010         read_unlock(&sk->sk_callback_lock);
1011 }
1012
1013 static void
1014 iscsi_tcp_state_change(struct sock *sk)
1015 {
1016         struct iscsi_tcp_conn *tcp_conn;
1017         struct iscsi_conn *conn;
1018         struct iscsi_session *session;
1019         void (*old_state_change)(struct sock *);
1020
1021         read_lock(&sk->sk_callback_lock);
1022
1023         conn = (struct iscsi_conn*)sk->sk_user_data;
1024         session = conn->session;
1025
1026         if ((sk->sk_state == TCP_CLOSE_WAIT ||
1027              sk->sk_state == TCP_CLOSE) &&
1028             !atomic_read(&sk->sk_rmem_alloc)) {
1029                 debug_tcp("iscsi_tcp_state_change: TCP_CLOSE|TCP_CLOSE_WAIT\n");
1030                 iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
1031         }
1032
1033         tcp_conn = conn->dd_data;
1034         old_state_change = tcp_conn->old_state_change;
1035
1036         read_unlock(&sk->sk_callback_lock);
1037
1038         old_state_change(sk);
1039 }
1040
1041 /**
1042  * iscsi_write_space - Called when more output buffer space is available
1043  * @sk: socket space is available for
1044  **/
1045 static void
1046 iscsi_write_space(struct sock *sk)
1047 {
1048         struct iscsi_conn *conn = (struct iscsi_conn*)sk->sk_user_data;
1049         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1050
1051         tcp_conn->old_write_space(sk);
1052         debug_tcp("iscsi_write_space: cid %d\n", conn->id);
1053         scsi_queue_work(conn->session->host, &conn->xmitwork);
1054 }
1055
1056 static void
1057 iscsi_conn_set_callbacks(struct iscsi_conn *conn)
1058 {
1059         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1060         struct sock *sk = tcp_conn->sock->sk;
1061
1062         /* assign new callbacks */
1063         write_lock_bh(&sk->sk_callback_lock);
1064         sk->sk_user_data = conn;
1065         tcp_conn->old_data_ready = sk->sk_data_ready;
1066         tcp_conn->old_state_change = sk->sk_state_change;
1067         tcp_conn->old_write_space = sk->sk_write_space;
1068         sk->sk_data_ready = iscsi_tcp_data_ready;
1069         sk->sk_state_change = iscsi_tcp_state_change;
1070         sk->sk_write_space = iscsi_write_space;
1071         write_unlock_bh(&sk->sk_callback_lock);
1072 }
1073
1074 static void
1075 iscsi_conn_restore_callbacks(struct iscsi_tcp_conn *tcp_conn)
1076 {
1077         struct sock *sk = tcp_conn->sock->sk;
1078
1079         /* restore socket callbacks, see also: iscsi_conn_set_callbacks() */
1080         write_lock_bh(&sk->sk_callback_lock);
1081         sk->sk_user_data    = NULL;
1082         sk->sk_data_ready   = tcp_conn->old_data_ready;
1083         sk->sk_state_change = tcp_conn->old_state_change;
1084         sk->sk_write_space  = tcp_conn->old_write_space;
1085         sk->sk_no_check  = 0;
1086         write_unlock_bh(&sk->sk_callback_lock);
1087 }
1088
1089 /**
1090  * iscsi_send - generic send routine
1091  * @sk: kernel's socket
1092  * @buf: buffer to write from
1093  * @size: actual size to write
1094  * @flags: socket's flags
1095  */
1096 static inline int
1097 iscsi_send(struct iscsi_conn *conn, struct iscsi_buf *buf, int size, int flags)
1098 {
1099         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1100         struct socket *sk = tcp_conn->sock;
1101         int offset = buf->sg.offset + buf->sent, res;
1102
1103         /*
1104          * if we got use_sg=0 or are sending something we kmallocd
1105          * then we did not have to do kmap (kmap returns page_address)
1106          *
1107          * if we got use_sg > 0, but had to drop down, we do not
1108          * set clustering so this should only happen for that
1109          * slab case.
1110          */
1111         if (buf->use_sendmsg)
1112                 res = sock_no_sendpage(sk, buf->sg.page, offset, size, flags);
1113         else
1114                 res = tcp_conn->sendpage(sk, buf->sg.page, offset, size, flags);
1115
1116         if (res >= 0) {
1117                 conn->txdata_octets += res;
1118                 buf->sent += res;
1119                 return res;
1120         }
1121
1122         tcp_conn->sendpage_failures_cnt++;
1123         if (res == -EAGAIN)
1124                 res = -ENOBUFS;
1125         else
1126                 iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
1127         return res;
1128 }
1129
1130 /**
1131  * iscsi_sendhdr - send PDU Header via tcp_sendpage()
1132  * @conn: iscsi connection
1133  * @buf: buffer to write from
1134  * @datalen: lenght of data to be sent after the header
1135  *
1136  * Notes:
1137  *      (Tx, Fast Path)
1138  **/
1139 static inline int
1140 iscsi_sendhdr(struct iscsi_conn *conn, struct iscsi_buf *buf, int datalen)
1141 {
1142         int flags = 0; /* MSG_DONTWAIT; */
1143         int res, size;
1144
1145         size = buf->sg.length - buf->sent;
1146         BUG_ON(buf->sent + size > buf->sg.length);
1147         if (buf->sent + size != buf->sg.length || datalen)
1148                 flags |= MSG_MORE;
1149
1150         res = iscsi_send(conn, buf, size, flags);
1151         debug_tcp("sendhdr %d bytes, sent %d res %d\n", size, buf->sent, res);
1152         if (res >= 0) {
1153                 if (size != res)
1154                         return -EAGAIN;
1155                 return 0;
1156         }
1157
1158         return res;
1159 }
1160
1161 /**
1162  * iscsi_sendpage - send one page of iSCSI Data-Out.
1163  * @conn: iscsi connection
1164  * @buf: buffer to write from
1165  * @count: remaining data
1166  * @sent: number of bytes sent
1167  *
1168  * Notes:
1169  *      (Tx, Fast Path)
1170  **/
1171 static inline int
1172 iscsi_sendpage(struct iscsi_conn *conn, struct iscsi_buf *buf,
1173                int *count, int *sent)
1174 {
1175         int flags = 0; /* MSG_DONTWAIT; */
1176         int res, size;
1177
1178         size = buf->sg.length - buf->sent;
1179         BUG_ON(buf->sent + size > buf->sg.length);
1180         if (size > *count)
1181                 size = *count;
1182         if (buf->sent + size != buf->sg.length || *count != size)
1183                 flags |= MSG_MORE;
1184
1185         res = iscsi_send(conn, buf, size, flags);
1186         debug_tcp("sendpage: %d bytes, sent %d left %d sent %d res %d\n",
1187                   size, buf->sent, *count, *sent, res);
1188         if (res >= 0) {
1189                 *count -= res;
1190                 *sent += res;
1191                 if (size != res)
1192                         return -EAGAIN;
1193                 return 0;
1194         }
1195
1196         return res;
1197 }
1198
1199 static inline void
1200 iscsi_data_digest_init(struct iscsi_tcp_conn *tcp_conn,
1201                       struct iscsi_tcp_cmd_task *tcp_ctask)
1202 {
1203         crypto_hash_init(&tcp_conn->tx_hash);
1204         tcp_ctask->digest_count = 4;
1205 }
1206
1207 /**
1208  * iscsi_solicit_data_cont - initialize next Data-Out
1209  * @conn: iscsi connection
1210  * @ctask: scsi command task
1211  * @r2t: R2T info
1212  * @left: bytes left to transfer
1213  *
1214  * Notes:
1215  *      Initialize next Data-Out within this R2T sequence and continue
1216  *      to process next Scatter-Gather element(if any) of this SCSI command.
1217  *
1218  *      Called under connection lock.
1219  **/
1220 static void
1221 iscsi_solicit_data_cont(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask,
1222                         struct iscsi_r2t_info *r2t, int left)
1223 {
1224         struct iscsi_data *hdr;
1225         struct scsi_cmnd *sc = ctask->sc;
1226         int new_offset;
1227
1228         hdr = &r2t->dtask.hdr;
1229         memset(hdr, 0, sizeof(struct iscsi_data));
1230         hdr->ttt = r2t->ttt;
1231         hdr->datasn = cpu_to_be32(r2t->solicit_datasn);
1232         r2t->solicit_datasn++;
1233         hdr->opcode = ISCSI_OP_SCSI_DATA_OUT;
1234         memcpy(hdr->lun, ctask->hdr->lun, sizeof(hdr->lun));
1235         hdr->itt = ctask->hdr->itt;
1236         hdr->exp_statsn = r2t->exp_statsn;
1237         new_offset = r2t->data_offset + r2t->sent;
1238         hdr->offset = cpu_to_be32(new_offset);
1239         if (left > conn->max_xmit_dlength) {
1240                 hton24(hdr->dlength, conn->max_xmit_dlength);
1241                 r2t->data_count = conn->max_xmit_dlength;
1242         } else {
1243                 hton24(hdr->dlength, left);
1244                 r2t->data_count = left;
1245                 hdr->flags = ISCSI_FLAG_CMD_FINAL;
1246         }
1247         conn->dataout_pdus_cnt++;
1248
1249         iscsi_buf_init_iov(&r2t->headbuf, (char*)hdr,
1250                            sizeof(struct iscsi_hdr));
1251
1252         if (iscsi_buf_left(&r2t->sendbuf))
1253                 return;
1254
1255         if (sc->use_sg) {
1256                 iscsi_buf_init_sg(&r2t->sendbuf, r2t->sg);
1257                 r2t->sg += 1;
1258         } else {
1259                 iscsi_buf_init_iov(&r2t->sendbuf,
1260                             (char*)sc->request_buffer + new_offset,
1261                             r2t->data_count);
1262                 r2t->sg = NULL;
1263         }
1264 }
1265
1266 static void iscsi_set_padding(struct iscsi_tcp_cmd_task *tcp_ctask,
1267                               unsigned long len)
1268 {
1269         tcp_ctask->pad_count = len & (ISCSI_PAD_LEN - 1);
1270         if (!tcp_ctask->pad_count)
1271                 return;
1272
1273         tcp_ctask->pad_count = ISCSI_PAD_LEN - tcp_ctask->pad_count;
1274         debug_scsi("write padding %d bytes\n", tcp_ctask->pad_count);
1275         tcp_ctask->xmstate |= XMSTATE_W_PAD;
1276 }
1277
1278 /**
1279  * iscsi_tcp_cmd_init - Initialize iSCSI SCSI_READ or SCSI_WRITE commands
1280  * @conn: iscsi connection
1281  * @ctask: scsi command task
1282  * @sc: scsi command
1283  **/
1284 static void
1285 iscsi_tcp_cmd_init(struct iscsi_cmd_task *ctask)
1286 {
1287         struct scsi_cmnd *sc = ctask->sc;
1288         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1289
1290         BUG_ON(__kfifo_len(tcp_ctask->r2tqueue));
1291
1292         tcp_ctask->sent = 0;
1293         tcp_ctask->sg_count = 0;
1294         tcp_ctask->exp_datasn = 0;
1295
1296         if (sc->sc_data_direction == DMA_TO_DEVICE) {
1297                 tcp_ctask->xmstate = XMSTATE_W_HDR;
1298                 BUG_ON(sc->request_bufflen == 0);
1299
1300                 if (sc->use_sg) {
1301                         struct scatterlist *sg = sc->request_buffer;
1302
1303                         iscsi_buf_init_sg(&tcp_ctask->sendbuf, sg);
1304                         tcp_ctask->sg = sg + 1;
1305                         tcp_ctask->bad_sg = sg + sc->use_sg;
1306                 } else {
1307                         iscsi_buf_init_iov(&tcp_ctask->sendbuf,
1308                                            sc->request_buffer,
1309                                            sc->request_bufflen);
1310                         tcp_ctask->sg = NULL;
1311                         tcp_ctask->bad_sg = NULL;
1312                 }
1313                 debug_scsi("cmd [itt 0x%x total %d imm_data %d "
1314                            "unsol count %d, unsol offset %d]\n",
1315                            ctask->itt, sc->request_bufflen, ctask->imm_count,
1316                            ctask->unsol_count, ctask->unsol_offset);
1317         } else
1318                 tcp_ctask->xmstate = XMSTATE_R_HDR;
1319
1320         iscsi_buf_init_iov(&tcp_ctask->headbuf, (char*)ctask->hdr,
1321                             sizeof(struct iscsi_hdr));
1322 }
1323
1324 /**
1325  * iscsi_tcp_mtask_xmit - xmit management(immediate) task
1326  * @conn: iscsi connection
1327  * @mtask: task management task
1328  *
1329  * Notes:
1330  *      The function can return -EAGAIN in which case caller must
1331  *      call it again later, or recover. '0' return code means successful
1332  *      xmit.
1333  *
1334  *      Management xmit state machine consists of two states:
1335  *              IN_PROGRESS_IMM_HEAD - PDU Header xmit in progress
1336  *              IN_PROGRESS_IMM_DATA - PDU Data xmit in progress
1337  **/
1338 static int
1339 iscsi_tcp_mtask_xmit(struct iscsi_conn *conn, struct iscsi_mgmt_task *mtask)
1340 {
1341         struct iscsi_tcp_mgmt_task *tcp_mtask = mtask->dd_data;
1342         int rc;
1343
1344         debug_scsi("mtask deq [cid %d state %x itt 0x%x]\n",
1345                 conn->id, tcp_mtask->xmstate, mtask->itt);
1346
1347         if (tcp_mtask->xmstate & XMSTATE_IMM_HDR) {
1348                 tcp_mtask->xmstate &= ~XMSTATE_IMM_HDR;
1349                 if (mtask->data_count)
1350                         tcp_mtask->xmstate |= XMSTATE_IMM_DATA;
1351                 if (conn->c_stage != ISCSI_CONN_INITIAL_STAGE &&
1352                     conn->stop_stage != STOP_CONN_RECOVER &&
1353                     conn->hdrdgst_en)
1354                         iscsi_hdr_digest(conn, &tcp_mtask->headbuf,
1355                                         (u8*)tcp_mtask->hdrext);
1356                 rc = iscsi_sendhdr(conn, &tcp_mtask->headbuf,
1357                                    mtask->data_count);
1358                 if (rc) {
1359                         tcp_mtask->xmstate |= XMSTATE_IMM_HDR;
1360                         if (mtask->data_count)
1361                                 tcp_mtask->xmstate &= ~XMSTATE_IMM_DATA;
1362                         return rc;
1363                 }
1364         }
1365
1366         if (tcp_mtask->xmstate & XMSTATE_IMM_DATA) {
1367                 BUG_ON(!mtask->data_count);
1368                 tcp_mtask->xmstate &= ~XMSTATE_IMM_DATA;
1369                 /* FIXME: implement.
1370                  * Virtual buffer could be spreaded across multiple pages...
1371                  */
1372                 do {
1373                         int rc;
1374
1375                         rc = iscsi_sendpage(conn, &tcp_mtask->sendbuf,
1376                                         &mtask->data_count, &tcp_mtask->sent);
1377                         if (rc) {
1378                                 tcp_mtask->xmstate |= XMSTATE_IMM_DATA;
1379                                 return rc;
1380                         }
1381                 } while (mtask->data_count);
1382         }
1383
1384         BUG_ON(tcp_mtask->xmstate != XMSTATE_IDLE);
1385         if (mtask->hdr->itt == RESERVED_ITT) {
1386                 struct iscsi_session *session = conn->session;
1387
1388                 spin_lock_bh(&session->lock);
1389                 list_del(&conn->mtask->running);
1390                 __kfifo_put(session->mgmtpool.queue, (void*)&conn->mtask,
1391                             sizeof(void*));
1392                 spin_unlock_bh(&session->lock);
1393         }
1394         return 0;
1395 }
1396
1397 static inline int
1398 iscsi_send_read_hdr(struct iscsi_conn *conn,
1399                     struct iscsi_tcp_cmd_task *tcp_ctask)
1400 {
1401         int rc;
1402
1403         tcp_ctask->xmstate &= ~XMSTATE_R_HDR;
1404         if (conn->hdrdgst_en)
1405                 iscsi_hdr_digest(conn, &tcp_ctask->headbuf,
1406                                  (u8*)tcp_ctask->hdrext);
1407         rc = iscsi_sendhdr(conn, &tcp_ctask->headbuf, 0);
1408         if (!rc) {
1409                 BUG_ON(tcp_ctask->xmstate != XMSTATE_IDLE);
1410                 return 0; /* wait for Data-In */
1411         }
1412         tcp_ctask->xmstate |= XMSTATE_R_HDR;
1413         return rc;
1414 }
1415
1416 static inline int
1417 iscsi_send_write_hdr(struct iscsi_conn *conn,
1418                      struct iscsi_cmd_task *ctask)
1419 {
1420         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1421         int rc;
1422
1423         tcp_ctask->xmstate &= ~XMSTATE_W_HDR;
1424         if (conn->hdrdgst_en)
1425                 iscsi_hdr_digest(conn, &tcp_ctask->headbuf,
1426                                  (u8*)tcp_ctask->hdrext);
1427         rc = iscsi_sendhdr(conn, &tcp_ctask->headbuf, ctask->imm_count);
1428         if (rc) {
1429                 tcp_ctask->xmstate |= XMSTATE_W_HDR;
1430                 return rc;
1431         }
1432
1433         if (ctask->imm_count) {
1434                 tcp_ctask->xmstate |= XMSTATE_IMM_DATA;
1435                 iscsi_set_padding(tcp_ctask, ctask->imm_count);
1436
1437                 if (ctask->conn->datadgst_en) {
1438                         iscsi_data_digest_init(ctask->conn->dd_data, tcp_ctask);
1439                         tcp_ctask->immdigest = 0;
1440                 }
1441         }
1442
1443         if (ctask->unsol_count)
1444                 tcp_ctask->xmstate |= XMSTATE_UNS_HDR | XMSTATE_UNS_INIT;
1445         return 0;
1446 }
1447
1448 static int
1449 iscsi_send_padding(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1450 {
1451         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1452         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1453         int sent = 0, rc;
1454
1455         if (tcp_ctask->xmstate & XMSTATE_W_PAD) {
1456                 iscsi_buf_init_iov(&tcp_ctask->sendbuf, (char*)&tcp_ctask->pad,
1457                                    tcp_ctask->pad_count);
1458                 if (conn->datadgst_en)
1459                         crypto_hash_update(&tcp_conn->tx_hash,
1460                                            &tcp_ctask->sendbuf.sg,
1461                                            tcp_ctask->sendbuf.sg.length);
1462         } else if (!(tcp_ctask->xmstate & XMSTATE_W_RESEND_PAD))
1463                 return 0;
1464
1465         tcp_ctask->xmstate &= ~XMSTATE_W_PAD;
1466         tcp_ctask->xmstate &= ~XMSTATE_W_RESEND_PAD;
1467         debug_scsi("sending %d pad bytes for itt 0x%x\n",
1468                    tcp_ctask->pad_count, ctask->itt);
1469         rc = iscsi_sendpage(conn, &tcp_ctask->sendbuf, &tcp_ctask->pad_count,
1470                            &sent);
1471         if (rc) {
1472                 debug_scsi("padding send failed %d\n", rc);
1473                 tcp_ctask->xmstate |= XMSTATE_W_RESEND_PAD;
1474         }
1475         return rc;
1476 }
1477
1478 static int
1479 iscsi_send_digest(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask,
1480                         struct iscsi_buf *buf, uint32_t *digest)
1481 {
1482         struct iscsi_tcp_cmd_task *tcp_ctask;
1483         struct iscsi_tcp_conn *tcp_conn;
1484         int rc, sent = 0;
1485
1486         if (!conn->datadgst_en)
1487                 return 0;
1488
1489         tcp_ctask = ctask->dd_data;
1490         tcp_conn = conn->dd_data;
1491
1492         if (!(tcp_ctask->xmstate & XMSTATE_W_RESEND_DATA_DIGEST)) {
1493                 crypto_hash_final(&tcp_conn->tx_hash, (u8*)digest);
1494                 iscsi_buf_init_iov(buf, (char*)digest, 4);
1495         }
1496         tcp_ctask->xmstate &= ~XMSTATE_W_RESEND_DATA_DIGEST;
1497
1498         rc = iscsi_sendpage(conn, buf, &tcp_ctask->digest_count, &sent);
1499         if (!rc)
1500                 debug_scsi("sent digest 0x%x for itt 0x%x\n", *digest,
1501                           ctask->itt);
1502         else {
1503                 debug_scsi("sending digest 0x%x failed for itt 0x%x!\n",
1504                           *digest, ctask->itt);
1505                 tcp_ctask->xmstate |= XMSTATE_W_RESEND_DATA_DIGEST;
1506         }
1507         return rc;
1508 }
1509
1510 static int
1511 iscsi_send_data(struct iscsi_cmd_task *ctask, struct iscsi_buf *sendbuf,
1512                 struct scatterlist **sg, int *sent, int *count,
1513                 struct iscsi_buf *digestbuf, uint32_t *digest)
1514 {
1515         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1516         struct iscsi_conn *conn = ctask->conn;
1517         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1518         int rc, buf_sent, offset;
1519
1520         while (*count) {
1521                 buf_sent = 0;
1522                 offset = sendbuf->sent;
1523
1524                 rc = iscsi_sendpage(conn, sendbuf, count, &buf_sent);
1525                 *sent = *sent + buf_sent;
1526                 if (buf_sent && conn->datadgst_en)
1527                         partial_sg_digest_update(&tcp_conn->tx_hash,
1528                                 &sendbuf->sg, sendbuf->sg.offset + offset,
1529                                 buf_sent);
1530                 if (!iscsi_buf_left(sendbuf) && *sg != tcp_ctask->bad_sg) {
1531                         iscsi_buf_init_sg(sendbuf, *sg);
1532                         *sg = *sg + 1;
1533                 }
1534
1535                 if (rc)
1536                         return rc;
1537         }
1538
1539         rc = iscsi_send_padding(conn, ctask);
1540         if (rc)
1541                 return rc;
1542
1543         return iscsi_send_digest(conn, ctask, digestbuf, digest);
1544 }
1545
1546 static int
1547 iscsi_send_unsol_hdr(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1548 {
1549         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1550         struct iscsi_data_task *dtask;
1551         int rc;
1552
1553         tcp_ctask->xmstate |= XMSTATE_UNS_DATA;
1554         if (tcp_ctask->xmstate & XMSTATE_UNS_INIT) {
1555                 dtask = &tcp_ctask->unsol_dtask;
1556
1557                 iscsi_prep_unsolicit_data_pdu(ctask, &dtask->hdr);
1558                 iscsi_buf_init_iov(&tcp_ctask->headbuf, (char*)&dtask->hdr,
1559                                    sizeof(struct iscsi_hdr));
1560                 if (conn->hdrdgst_en)
1561                         iscsi_hdr_digest(conn, &tcp_ctask->headbuf,
1562                                         (u8*)dtask->hdrext);
1563
1564                 tcp_ctask->xmstate &= ~XMSTATE_UNS_INIT;
1565                 iscsi_set_padding(tcp_ctask, ctask->data_count);
1566         }
1567
1568         rc = iscsi_sendhdr(conn, &tcp_ctask->headbuf, ctask->data_count);
1569         if (rc) {
1570                 tcp_ctask->xmstate &= ~XMSTATE_UNS_DATA;
1571                 tcp_ctask->xmstate |= XMSTATE_UNS_HDR;
1572                 return rc;
1573         }
1574
1575         if (conn->datadgst_en) {
1576                 dtask = &tcp_ctask->unsol_dtask;
1577                 iscsi_data_digest_init(ctask->conn->dd_data, tcp_ctask);
1578                 dtask->digest = 0;
1579         }
1580
1581         debug_scsi("uns dout [itt 0x%x dlen %d sent %d]\n",
1582                    ctask->itt, ctask->unsol_count, tcp_ctask->sent);
1583         return 0;
1584 }
1585
1586 static int
1587 iscsi_send_unsol_pdu(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1588 {
1589         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1590         int rc;
1591
1592         if (tcp_ctask->xmstate & XMSTATE_UNS_HDR) {
1593                 BUG_ON(!ctask->unsol_count);
1594                 tcp_ctask->xmstate &= ~XMSTATE_UNS_HDR;
1595 send_hdr:
1596                 rc = iscsi_send_unsol_hdr(conn, ctask);
1597                 if (rc)
1598                         return rc;
1599         }
1600
1601         if (tcp_ctask->xmstate & XMSTATE_UNS_DATA) {
1602                 struct iscsi_data_task *dtask = &tcp_ctask->unsol_dtask;
1603                 int start = tcp_ctask->sent;
1604
1605                 rc = iscsi_send_data(ctask, &tcp_ctask->sendbuf, &tcp_ctask->sg,
1606                                      &tcp_ctask->sent, &ctask->data_count,
1607                                      &dtask->digestbuf, &dtask->digest);
1608                 ctask->unsol_count -= tcp_ctask->sent - start;
1609                 if (rc)
1610                         return rc;
1611                 tcp_ctask->xmstate &= ~XMSTATE_UNS_DATA;
1612                 /*
1613                  * Done with the Data-Out. Next, check if we need
1614                  * to send another unsolicited Data-Out.
1615                  */
1616                 if (ctask->unsol_count) {
1617                         debug_scsi("sending more uns\n");
1618                         tcp_ctask->xmstate |= XMSTATE_UNS_INIT;
1619                         goto send_hdr;
1620                 }
1621         }
1622         return 0;
1623 }
1624
1625 static int iscsi_send_sol_pdu(struct iscsi_conn *conn,
1626                               struct iscsi_cmd_task *ctask)
1627 {
1628         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1629         struct iscsi_session *session = conn->session;
1630         struct iscsi_r2t_info *r2t;
1631         struct iscsi_data_task *dtask;
1632         int left, rc;
1633
1634         if (tcp_ctask->xmstate & XMSTATE_SOL_HDR) {
1635                 tcp_ctask->xmstate &= ~XMSTATE_SOL_HDR;
1636                 tcp_ctask->xmstate |= XMSTATE_SOL_DATA;
1637                 if (!tcp_ctask->r2t) {
1638                         spin_lock_bh(&session->lock);
1639                         __kfifo_get(tcp_ctask->r2tqueue, (void*)&tcp_ctask->r2t,
1640                                     sizeof(void*));
1641                         spin_unlock_bh(&session->lock);
1642                 }
1643 send_hdr:
1644                 r2t = tcp_ctask->r2t;
1645                 dtask = &r2t->dtask;
1646
1647                 if (conn->hdrdgst_en)
1648                         iscsi_hdr_digest(conn, &r2t->headbuf,
1649                                         (u8*)dtask->hdrext);
1650                 rc = iscsi_sendhdr(conn, &r2t->headbuf, r2t->data_count);
1651                 if (rc) {
1652                         tcp_ctask->xmstate &= ~XMSTATE_SOL_DATA;
1653                         tcp_ctask->xmstate |= XMSTATE_SOL_HDR;
1654                         return rc;
1655                 }
1656
1657                 if (conn->datadgst_en) {
1658                         iscsi_data_digest_init(conn->dd_data, tcp_ctask);
1659                         dtask->digest = 0;
1660                 }
1661
1662                 iscsi_set_padding(tcp_ctask, r2t->data_count);
1663                 debug_scsi("sol dout [dsn %d itt 0x%x dlen %d sent %d]\n",
1664                         r2t->solicit_datasn - 1, ctask->itt, r2t->data_count,
1665                         r2t->sent);
1666         }
1667
1668         if (tcp_ctask->xmstate & XMSTATE_SOL_DATA) {
1669                 r2t = tcp_ctask->r2t;
1670                 dtask = &r2t->dtask;
1671
1672                 rc = iscsi_send_data(ctask, &r2t->sendbuf, &r2t->sg,
1673                                      &r2t->sent, &r2t->data_count,
1674                                      &dtask->digestbuf, &dtask->digest);
1675                 if (rc)
1676                         return rc;
1677                 tcp_ctask->xmstate &= ~XMSTATE_SOL_DATA;
1678
1679                 /*
1680                  * Done with this Data-Out. Next, check if we have
1681                  * to send another Data-Out for this R2T.
1682                  */
1683                 BUG_ON(r2t->data_length - r2t->sent < 0);
1684                 left = r2t->data_length - r2t->sent;
1685                 if (left) {
1686                         iscsi_solicit_data_cont(conn, ctask, r2t, left);
1687                         tcp_ctask->xmstate |= XMSTATE_SOL_DATA;
1688                         tcp_ctask->xmstate &= ~XMSTATE_SOL_HDR;
1689                         goto send_hdr;
1690                 }
1691
1692                 /*
1693                  * Done with this R2T. Check if there are more
1694                  * outstanding R2Ts ready to be processed.
1695                  */
1696                 spin_lock_bh(&session->lock);
1697                 tcp_ctask->r2t = NULL;
1698                 __kfifo_put(tcp_ctask->r2tpool.queue, (void*)&r2t,
1699                             sizeof(void*));
1700                 if (__kfifo_get(tcp_ctask->r2tqueue, (void*)&r2t,
1701                                 sizeof(void*))) {
1702                         tcp_ctask->r2t = r2t;
1703                         tcp_ctask->xmstate |= XMSTATE_SOL_DATA;
1704                         tcp_ctask->xmstate &= ~XMSTATE_SOL_HDR;
1705                         spin_unlock_bh(&session->lock);
1706                         goto send_hdr;
1707                 }
1708                 spin_unlock_bh(&session->lock);
1709         }
1710         return 0;
1711 }
1712
1713 static int
1714 iscsi_tcp_ctask_xmit(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1715 {
1716         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1717         int rc = 0;
1718
1719         debug_scsi("ctask deq [cid %d xmstate %x itt 0x%x]\n",
1720                 conn->id, tcp_ctask->xmstate, ctask->itt);
1721
1722         /*
1723          * serialize with TMF AbortTask
1724          */
1725         if (ctask->mtask)
1726                 return rc;
1727
1728         if (tcp_ctask->xmstate & XMSTATE_R_HDR)
1729                 return iscsi_send_read_hdr(conn, tcp_ctask);
1730
1731         if (tcp_ctask->xmstate & XMSTATE_W_HDR) {
1732                 rc = iscsi_send_write_hdr(conn, ctask);
1733                 if (rc)
1734                         return rc;
1735         }
1736
1737         if (tcp_ctask->xmstate & XMSTATE_IMM_DATA) {
1738                 rc = iscsi_send_data(ctask, &tcp_ctask->sendbuf, &tcp_ctask->sg,
1739                                      &tcp_ctask->sent, &ctask->imm_count,
1740                                      &tcp_ctask->immbuf, &tcp_ctask->immdigest);
1741                 if (rc)
1742                         return rc;
1743                 tcp_ctask->xmstate &= ~XMSTATE_IMM_DATA;
1744         }
1745
1746         rc = iscsi_send_unsol_pdu(conn, ctask);
1747         if (rc)
1748                 return rc;
1749
1750         rc = iscsi_send_sol_pdu(conn, ctask);
1751         if (rc)
1752                 return rc;
1753
1754         return rc;
1755 }
1756
1757 static struct iscsi_cls_conn *
1758 iscsi_tcp_conn_create(struct iscsi_cls_session *cls_session, uint32_t conn_idx)
1759 {
1760         struct iscsi_conn *conn;
1761         struct iscsi_cls_conn *cls_conn;
1762         struct iscsi_tcp_conn *tcp_conn;
1763
1764         cls_conn = iscsi_conn_setup(cls_session, conn_idx);
1765         if (!cls_conn)
1766                 return NULL;
1767         conn = cls_conn->dd_data;
1768         /*
1769          * due to strange issues with iser these are not set
1770          * in iscsi_conn_setup
1771          */
1772         conn->max_recv_dlength = ISCSI_DEF_MAX_RECV_SEG_LEN;
1773
1774         tcp_conn = kzalloc(sizeof(*tcp_conn), GFP_KERNEL);
1775         if (!tcp_conn)
1776                 goto tcp_conn_alloc_fail;
1777
1778         conn->dd_data = tcp_conn;
1779         tcp_conn->iscsi_conn = conn;
1780         tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
1781         /* initial operational parameters */
1782         tcp_conn->hdr_size = sizeof(struct iscsi_hdr);
1783
1784         tcp_conn->tx_hash.tfm = crypto_alloc_hash("crc32c", 0,
1785                                                   CRYPTO_ALG_ASYNC);
1786         tcp_conn->tx_hash.flags = 0;
1787         if (IS_ERR(tcp_conn->tx_hash.tfm)) {
1788                 printk(KERN_ERR "Could not create connection due to crc32c "
1789                        "loading error %ld. Make sure the crc32c module is "
1790                        "built as a module or into the kernel\n",
1791                         PTR_ERR(tcp_conn->tx_hash.tfm));
1792                 goto free_tcp_conn;
1793         }
1794
1795         tcp_conn->rx_hash.tfm = crypto_alloc_hash("crc32c", 0,
1796                                                   CRYPTO_ALG_ASYNC);
1797         tcp_conn->rx_hash.flags = 0;
1798         if (IS_ERR(tcp_conn->rx_hash.tfm)) {
1799                 printk(KERN_ERR "Could not create connection due to crc32c "
1800                        "loading error %ld. Make sure the crc32c module is "
1801                        "built as a module or into the kernel\n",
1802                         PTR_ERR(tcp_conn->rx_hash.tfm));
1803                 goto free_tx_tfm;
1804         }
1805
1806         return cls_conn;
1807
1808 free_tx_tfm:
1809         crypto_free_hash(tcp_conn->tx_hash.tfm);
1810 free_tcp_conn:
1811         kfree(tcp_conn);
1812 tcp_conn_alloc_fail:
1813         iscsi_conn_teardown(cls_conn);
1814         return NULL;
1815 }
1816
1817 static void
1818 iscsi_tcp_release_conn(struct iscsi_conn *conn)
1819 {
1820         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1821
1822         if (!tcp_conn->sock)
1823                 return;
1824
1825         sock_hold(tcp_conn->sock->sk);
1826         iscsi_conn_restore_callbacks(tcp_conn);
1827         sock_put(tcp_conn->sock->sk);
1828
1829         sock_release(tcp_conn->sock);
1830         tcp_conn->sock = NULL;
1831         conn->recv_lock = NULL;
1832 }
1833
1834 static void
1835 iscsi_tcp_conn_destroy(struct iscsi_cls_conn *cls_conn)
1836 {
1837         struct iscsi_conn *conn = cls_conn->dd_data;
1838         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1839
1840         iscsi_tcp_release_conn(conn);
1841         iscsi_conn_teardown(cls_conn);
1842
1843         if (tcp_conn->tx_hash.tfm)
1844                 crypto_free_hash(tcp_conn->tx_hash.tfm);
1845         if (tcp_conn->rx_hash.tfm)
1846                 crypto_free_hash(tcp_conn->rx_hash.tfm);
1847
1848         kfree(tcp_conn);
1849 }
1850
1851 static void
1852 iscsi_tcp_conn_stop(struct iscsi_cls_conn *cls_conn, int flag)
1853 {
1854         struct iscsi_conn *conn = cls_conn->dd_data;
1855         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1856
1857         iscsi_conn_stop(cls_conn, flag);
1858         iscsi_tcp_release_conn(conn);
1859         tcp_conn->hdr_size = sizeof(struct iscsi_hdr);
1860 }
1861
1862 static int
1863 iscsi_tcp_conn_bind(struct iscsi_cls_session *cls_session,
1864                     struct iscsi_cls_conn *cls_conn, uint64_t transport_eph,
1865                     int is_leading)
1866 {
1867         struct iscsi_conn *conn = cls_conn->dd_data;
1868         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1869         struct sock *sk;
1870         struct socket *sock;
1871         int err;
1872
1873         /* lookup for existing socket */
1874         sock = sockfd_lookup((int)transport_eph, &err);
1875         if (!sock) {
1876                 printk(KERN_ERR "iscsi_tcp: sockfd_lookup failed %d\n", err);
1877                 return -EEXIST;
1878         }
1879
1880         err = iscsi_conn_bind(cls_session, cls_conn, is_leading);
1881         if (err)
1882                 return err;
1883
1884         /* bind iSCSI connection and socket */
1885         tcp_conn->sock = sock;
1886
1887         /* setup Socket parameters */
1888         sk = sock->sk;
1889         sk->sk_reuse = 1;
1890         sk->sk_sndtimeo = 15 * HZ; /* FIXME: make it configurable */
1891         sk->sk_allocation = GFP_ATOMIC;
1892
1893         /* FIXME: disable Nagle's algorithm */
1894
1895         /*
1896          * Intercept TCP callbacks for sendfile like receive
1897          * processing.
1898          */
1899         conn->recv_lock = &sk->sk_callback_lock;
1900         iscsi_conn_set_callbacks(conn);
1901         tcp_conn->sendpage = tcp_conn->sock->ops->sendpage;
1902         /*
1903          * set receive state machine into initial state
1904          */
1905         tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
1906
1907         return 0;
1908 }
1909
1910 /* called with host lock */
1911 static void
1912 iscsi_tcp_mgmt_init(struct iscsi_conn *conn, struct iscsi_mgmt_task *mtask,
1913                     char *data, uint32_t data_size)
1914 {
1915         struct iscsi_tcp_mgmt_task *tcp_mtask = mtask->dd_data;
1916
1917         iscsi_buf_init_iov(&tcp_mtask->headbuf, (char*)mtask->hdr,
1918                            sizeof(struct iscsi_hdr));
1919         tcp_mtask->xmstate = XMSTATE_IMM_HDR;
1920         tcp_mtask->sent = 0;
1921
1922         if (mtask->data_count)
1923                 iscsi_buf_init_iov(&tcp_mtask->sendbuf, (char*)mtask->data,
1924                                     mtask->data_count);
1925 }
1926
1927 static int
1928 iscsi_r2tpool_alloc(struct iscsi_session *session)
1929 {
1930         int i;
1931         int cmd_i;
1932
1933         /*
1934          * initialize per-task: R2T pool and xmit queue
1935          */
1936         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
1937                 struct iscsi_cmd_task *ctask = session->cmds[cmd_i];
1938                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1939
1940                 /*
1941                  * pre-allocated x4 as much r2ts to handle race when
1942                  * target acks DataOut faster than we data_xmit() queues
1943                  * could replenish r2tqueue.
1944                  */
1945
1946                 /* R2T pool */
1947                 if (iscsi_pool_init(&tcp_ctask->r2tpool, session->max_r2t * 4,
1948                                     (void***)&tcp_ctask->r2ts,
1949                                     sizeof(struct iscsi_r2t_info))) {
1950                         goto r2t_alloc_fail;
1951                 }
1952
1953                 /* R2T xmit queue */
1954                 tcp_ctask->r2tqueue = kfifo_alloc(
1955                       session->max_r2t * 4 * sizeof(void*), GFP_KERNEL, NULL);
1956                 if (tcp_ctask->r2tqueue == ERR_PTR(-ENOMEM)) {
1957                         iscsi_pool_free(&tcp_ctask->r2tpool,
1958                                         (void**)tcp_ctask->r2ts);
1959                         goto r2t_alloc_fail;
1960                 }
1961         }
1962
1963         return 0;
1964
1965 r2t_alloc_fail:
1966         for (i = 0; i < cmd_i; i++) {
1967                 struct iscsi_cmd_task *ctask = session->cmds[i];
1968                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1969
1970                 kfifo_free(tcp_ctask->r2tqueue);
1971                 iscsi_pool_free(&tcp_ctask->r2tpool,
1972                                 (void**)tcp_ctask->r2ts);
1973         }
1974         return -ENOMEM;
1975 }
1976
1977 static void
1978 iscsi_r2tpool_free(struct iscsi_session *session)
1979 {
1980         int i;
1981
1982         for (i = 0; i < session->cmds_max; i++) {
1983                 struct iscsi_cmd_task *ctask = session->cmds[i];
1984                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1985
1986                 kfifo_free(tcp_ctask->r2tqueue);
1987                 iscsi_pool_free(&tcp_ctask->r2tpool,
1988                                 (void**)tcp_ctask->r2ts);
1989         }
1990 }
1991
1992 static int
1993 iscsi_conn_set_param(struct iscsi_cls_conn *cls_conn, enum iscsi_param param,
1994                      char *buf, int buflen)
1995 {
1996         struct iscsi_conn *conn = cls_conn->dd_data;
1997         struct iscsi_session *session = conn->session;
1998         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1999         int value;
2000
2001         switch(param) {
2002         case ISCSI_PARAM_HDRDGST_EN:
2003                 iscsi_set_param(cls_conn, param, buf, buflen);
2004                 tcp_conn->hdr_size = sizeof(struct iscsi_hdr);
2005                 if (conn->hdrdgst_en)
2006                         tcp_conn->hdr_size += sizeof(__u32);
2007                 break;
2008         case ISCSI_PARAM_DATADGST_EN:
2009                 iscsi_set_param(cls_conn, param, buf, buflen);
2010                 tcp_conn->sendpage = conn->datadgst_en ?
2011                         sock_no_sendpage : tcp_conn->sock->ops->sendpage;
2012                 break;
2013         case ISCSI_PARAM_MAX_R2T:
2014                 sscanf(buf, "%d", &value);
2015                 if (session->max_r2t == roundup_pow_of_two(value))
2016                         break;
2017                 iscsi_r2tpool_free(session);
2018                 iscsi_set_param(cls_conn, param, buf, buflen);
2019                 if (session->max_r2t & (session->max_r2t - 1))
2020                         session->max_r2t = roundup_pow_of_two(session->max_r2t);
2021                 if (iscsi_r2tpool_alloc(session))
2022                         return -ENOMEM;
2023                 break;
2024         default:
2025                 return iscsi_set_param(cls_conn, param, buf, buflen);
2026         }
2027
2028         return 0;
2029 }
2030
2031 static int
2032 iscsi_tcp_conn_get_param(struct iscsi_cls_conn *cls_conn,
2033                          enum iscsi_param param, char *buf)
2034 {
2035         struct iscsi_conn *conn = cls_conn->dd_data;
2036         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
2037         struct inet_sock *inet;
2038         struct ipv6_pinfo *np;
2039         struct sock *sk;
2040         int len;
2041
2042         switch(param) {
2043         case ISCSI_PARAM_CONN_PORT:
2044                 mutex_lock(&conn->xmitmutex);
2045                 if (!tcp_conn->sock) {
2046                         mutex_unlock(&conn->xmitmutex);
2047                         return -EINVAL;
2048                 }
2049
2050                 inet = inet_sk(tcp_conn->sock->sk);
2051                 len = sprintf(buf, "%hu\n", be16_to_cpu(inet->dport));
2052                 mutex_unlock(&conn->xmitmutex);
2053                 break;
2054         case ISCSI_PARAM_CONN_ADDRESS:
2055                 mutex_lock(&conn->xmitmutex);
2056                 if (!tcp_conn->sock) {
2057                         mutex_unlock(&conn->xmitmutex);
2058                         return -EINVAL;
2059                 }
2060
2061                 sk = tcp_conn->sock->sk;
2062                 if (sk->sk_family == PF_INET) {
2063                         inet = inet_sk(sk);
2064                         len = sprintf(buf, NIPQUAD_FMT "\n",
2065                                       NIPQUAD(inet->daddr));
2066                 } else {
2067                         np = inet6_sk(sk);
2068                         len = sprintf(buf, NIP6_FMT "\n", NIP6(np->daddr));
2069                 }
2070                 mutex_unlock(&conn->xmitmutex);
2071                 break;
2072         default:
2073                 return iscsi_conn_get_param(cls_conn, param, buf);
2074         }
2075
2076         return len;
2077 }
2078
2079 static void
2080 iscsi_conn_get_stats(struct iscsi_cls_conn *cls_conn, struct iscsi_stats *stats)
2081 {
2082         struct iscsi_conn *conn = cls_conn->dd_data;
2083         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
2084
2085         stats->txdata_octets = conn->txdata_octets;
2086         stats->rxdata_octets = conn->rxdata_octets;
2087         stats->scsicmd_pdus = conn->scsicmd_pdus_cnt;
2088         stats->dataout_pdus = conn->dataout_pdus_cnt;
2089         stats->scsirsp_pdus = conn->scsirsp_pdus_cnt;
2090         stats->datain_pdus = conn->datain_pdus_cnt;
2091         stats->r2t_pdus = conn->r2t_pdus_cnt;
2092         stats->tmfcmd_pdus = conn->tmfcmd_pdus_cnt;
2093         stats->tmfrsp_pdus = conn->tmfrsp_pdus_cnt;
2094         stats->custom_length = 3;
2095         strcpy(stats->custom[0].desc, "tx_sendpage_failures");
2096         stats->custom[0].value = tcp_conn->sendpage_failures_cnt;
2097         strcpy(stats->custom[1].desc, "rx_discontiguous_hdr");
2098         stats->custom[1].value = tcp_conn->discontiguous_hdr_cnt;
2099         strcpy(stats->custom[2].desc, "eh_abort_cnt");
2100         stats->custom[2].value = conn->eh_abort_cnt;
2101 }
2102
2103 static struct iscsi_cls_session *
2104 iscsi_tcp_session_create(struct iscsi_transport *iscsit,
2105                          struct scsi_transport_template *scsit,
2106                          uint32_t initial_cmdsn, uint32_t *hostno)
2107 {
2108         struct iscsi_cls_session *cls_session;
2109         struct iscsi_session *session;
2110         uint32_t hn;
2111         int cmd_i;
2112
2113         cls_session = iscsi_session_setup(iscsit, scsit,
2114                                          sizeof(struct iscsi_tcp_cmd_task),
2115                                          sizeof(struct iscsi_tcp_mgmt_task),
2116                                          initial_cmdsn, &hn);
2117         if (!cls_session)
2118                 return NULL;
2119         *hostno = hn;
2120
2121         session = class_to_transport_session(cls_session);
2122         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
2123                 struct iscsi_cmd_task *ctask = session->cmds[cmd_i];
2124                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
2125
2126                 ctask->hdr = &tcp_ctask->hdr;
2127         }
2128
2129         for (cmd_i = 0; cmd_i < session->mgmtpool_max; cmd_i++) {
2130                 struct iscsi_mgmt_task *mtask = session->mgmt_cmds[cmd_i];
2131                 struct iscsi_tcp_mgmt_task *tcp_mtask = mtask->dd_data;
2132
2133                 mtask->hdr = &tcp_mtask->hdr;
2134         }
2135
2136         if (iscsi_r2tpool_alloc(class_to_transport_session(cls_session)))
2137                 goto r2tpool_alloc_fail;
2138
2139         return cls_session;
2140
2141 r2tpool_alloc_fail:
2142         iscsi_session_teardown(cls_session);
2143         return NULL;
2144 }
2145
2146 static void iscsi_tcp_session_destroy(struct iscsi_cls_session *cls_session)
2147 {
2148         iscsi_r2tpool_free(class_to_transport_session(cls_session));
2149         iscsi_session_teardown(cls_session);
2150 }
2151
2152 static struct scsi_host_template iscsi_sht = {
2153         .name                   = "iSCSI Initiator over TCP/IP",
2154         .queuecommand           = iscsi_queuecommand,
2155         .change_queue_depth     = iscsi_change_queue_depth,
2156         .can_queue              = ISCSI_XMIT_CMDS_MAX - 1,
2157         .sg_tablesize           = ISCSI_SG_TABLESIZE,
2158         .max_sectors            = 0xFFFF,
2159         .cmd_per_lun            = ISCSI_DEF_CMD_PER_LUN,
2160         .eh_abort_handler       = iscsi_eh_abort,
2161         .eh_host_reset_handler  = iscsi_eh_host_reset,
2162         .use_clustering         = DISABLE_CLUSTERING,
2163         .proc_name              = "iscsi_tcp",
2164         .this_id                = -1,
2165 };
2166
2167 static struct iscsi_transport iscsi_tcp_transport = {
2168         .owner                  = THIS_MODULE,
2169         .name                   = "tcp",
2170         .caps                   = CAP_RECOVERY_L0 | CAP_MULTI_R2T | CAP_HDRDGST
2171                                   | CAP_DATADGST,
2172         .param_mask             = ISCSI_MAX_RECV_DLENGTH |
2173                                   ISCSI_MAX_XMIT_DLENGTH |
2174                                   ISCSI_HDRDGST_EN |
2175                                   ISCSI_DATADGST_EN |
2176                                   ISCSI_INITIAL_R2T_EN |
2177                                   ISCSI_MAX_R2T |
2178                                   ISCSI_IMM_DATA_EN |
2179                                   ISCSI_FIRST_BURST |
2180                                   ISCSI_MAX_BURST |
2181                                   ISCSI_PDU_INORDER_EN |
2182                                   ISCSI_DATASEQ_INORDER_EN |
2183                                   ISCSI_ERL |
2184                                   ISCSI_CONN_PORT |
2185                                   ISCSI_CONN_ADDRESS |
2186                                   ISCSI_EXP_STATSN |
2187                                   ISCSI_PERSISTENT_PORT |
2188                                   ISCSI_PERSISTENT_ADDRESS |
2189                                   ISCSI_TARGET_NAME |
2190                                   ISCSI_TPGT,
2191         .host_param_mask        = ISCSI_HOST_HWADDRESS |
2192                                   ISCSI_HOST_INITIATOR_NAME,
2193         .host_template          = &iscsi_sht,
2194         .conndata_size          = sizeof(struct iscsi_conn),
2195         .max_conn               = 1,
2196         .max_cmd_len            = ISCSI_TCP_MAX_CMD_LEN,
2197         /* session management */
2198         .create_session         = iscsi_tcp_session_create,
2199         .destroy_session        = iscsi_tcp_session_destroy,
2200         /* connection management */
2201         .create_conn            = iscsi_tcp_conn_create,
2202         .bind_conn              = iscsi_tcp_conn_bind,
2203         .destroy_conn           = iscsi_tcp_conn_destroy,
2204         .set_param              = iscsi_conn_set_param,
2205         .get_conn_param         = iscsi_tcp_conn_get_param,
2206         .get_session_param      = iscsi_session_get_param,
2207         .start_conn             = iscsi_conn_start,
2208         .stop_conn              = iscsi_tcp_conn_stop,
2209         /* iscsi host params */
2210         .get_host_param         = iscsi_host_get_param,
2211         .set_host_param         = iscsi_host_set_param,
2212         /* IO */
2213         .send_pdu               = iscsi_conn_send_pdu,
2214         .get_stats              = iscsi_conn_get_stats,
2215         .init_cmd_task          = iscsi_tcp_cmd_init,
2216         .init_mgmt_task         = iscsi_tcp_mgmt_init,
2217         .xmit_cmd_task          = iscsi_tcp_ctask_xmit,
2218         .xmit_mgmt_task         = iscsi_tcp_mtask_xmit,
2219         .cleanup_cmd_task       = iscsi_tcp_cleanup_ctask,
2220         /* recovery */
2221         .session_recovery_timedout = iscsi_session_recovery_timedout,
2222 };
2223
2224 static int __init
2225 iscsi_tcp_init(void)
2226 {
2227         if (iscsi_max_lun < 1) {
2228                 printk(KERN_ERR "iscsi_tcp: Invalid max_lun value of %u\n",
2229                        iscsi_max_lun);
2230                 return -EINVAL;
2231         }
2232         iscsi_tcp_transport.max_lun = iscsi_max_lun;
2233
2234         if (!iscsi_register_transport(&iscsi_tcp_transport))
2235                 return -ENODEV;
2236
2237         return 0;
2238 }
2239
2240 static void __exit
2241 iscsi_tcp_exit(void)
2242 {
2243         iscsi_unregister_transport(&iscsi_tcp_transport);
2244 }
2245
2246 module_init(iscsi_tcp_init);
2247 module_exit(iscsi_tcp_exit);