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 <net/tcp.h>
  38 #include <scsi/scsi_cmnd.h>
  39 #include <scsi/scsi_host.h>
  40 #include <scsi/scsi.h>
  41 #include <scsi/scsi_transport_iscsi.h>
  42
  43 #include "iscsi_tcp.h"
  44
  45 MODULE_AUTHOR("Dmitry Yusupov <dmitry_yus@yahoo.com>, "
  46               "Alex Aizman <itn780@yahoo.com>");
  47 MODULE_DESCRIPTION("iSCSI/TCP data-path");
  48 MODULE_LICENSE("GPL");
  49 /* #define DEBUG_TCP */
  50 #define DEBUG_ASSERT
  51
  52 #ifdef DEBUG_TCP
  53 #define debug_tcp(fmt...) printk(KERN_INFO "tcp: " fmt)
  54 #else
  55 #define debug_tcp(fmt...)
  56 #endif
  57
  58 #ifndef DEBUG_ASSERT
  59 #ifdef BUG_ON
  60 #undef BUG_ON
  61 #endif
  62 #define BUG_ON(expr)
  63 #endif
  64
  65 static unsigned int iscsi_max_lun = 512;
  66 module_param_named(max_lun, iscsi_max_lun, uint, S_IRUGO);
  67
  68 static inline void
  69 iscsi_buf_init_iov(struct iscsi_buf *ibuf, char *vbuf, int size)
  70 {
  71         ibuf->sg.page = virt_to_page(vbuf);
  72         ibuf->sg.offset = offset_in_page(vbuf);
  73         ibuf->sg.length = size;
  74         ibuf->sent = 0;
  75         ibuf->use_sendmsg = 1;
  76 }
  77
  78 static inline void
  79 iscsi_buf_init_sg(struct iscsi_buf *ibuf, struct scatterlist *sg)
  80 {
  81         ibuf->sg.page = sg->page;
  82         ibuf->sg.offset = sg->offset;
  83         ibuf->sg.length = sg->length;
  84         /*
  85          * Fastpath: sg element fits into single page
  86          */
  87         if (sg->length + sg->offset <= PAGE_SIZE && !PageSlab(sg->page))
  88                 ibuf->use_sendmsg = 0;
  89         else
  90                 ibuf->use_sendmsg = 1;
  91         ibuf->sent = 0;
  92 }
  93
  94 static inline int
  95 iscsi_buf_left(struct iscsi_buf *ibuf)
  96 {
  97         int rc;
  98
  99         rc = ibuf->sg.length - ibuf->sent;
 100         BUG_ON(rc < 0);
 101         return rc;
 102 }
 103
 104 static inline void
 105 iscsi_hdr_digest(struct iscsi_conn *conn, struct iscsi_buf *buf,
 106                  u8* crc)
 107 {
 108         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 109
 110         crypto_hash_digest(&tcp_conn->tx_hash, &buf->sg, buf->sg.length, crc);
 111         buf->sg.length += sizeof(u32);
 112 }
 113
 114 static inline int
 115 iscsi_hdr_extract(struct iscsi_tcp_conn *tcp_conn)
 116 {
 117         struct sk_buff *skb = tcp_conn->in.skb;
 118
 119         tcp_conn->in.zero_copy_hdr = 0;
 120
 121         if (tcp_conn->in.copy >= tcp_conn->hdr_size &&
 122             tcp_conn->in_progress == IN_PROGRESS_WAIT_HEADER) {
 123                 /*
 124                  * Zero-copy PDU Header: using connection context
 125                  * to store header pointer.
 126                  */
 127                 if (skb_shinfo(skb)->frag_list == NULL &&
 128                     !skb_shinfo(skb)->nr_frags) {
 129                         tcp_conn->in.hdr = (struct iscsi_hdr *)
 130                                 ((char*)skb->data + tcp_conn->in.offset);
 131                         tcp_conn->in.zero_copy_hdr = 1;
 132                 } else {
 133                         /* ignoring return code since we checked
 134                          * in.copy before */
 135                         skb_copy_bits(skb, tcp_conn->in.offset,
 136                                 &tcp_conn->hdr, tcp_conn->hdr_size);
 137                         tcp_conn->in.hdr = &tcp_conn->hdr;
 138                 }
 139                 tcp_conn->in.offset += tcp_conn->hdr_size;
 140                 tcp_conn->in.copy -= tcp_conn->hdr_size;
 141         } else {
 142                 int hdr_remains;
 143                 int copylen;
 144
 145                 /*
 146                  * PDU header scattered across SKB's,
 147                  * copying it... This'll happen quite rarely.
 148                  */
 149
 150                 if (tcp_conn->in_progress == IN_PROGRESS_WAIT_HEADER)
 151                         tcp_conn->in.hdr_offset = 0;
 152
 153                 hdr_remains = tcp_conn->hdr_size - tcp_conn->in.hdr_offset;
 154                 BUG_ON(hdr_remains <= 0);
 155
 156                 copylen = min(tcp_conn->in.copy, hdr_remains);
 157                 skb_copy_bits(skb, tcp_conn->in.offset,
 158                         (char*)&tcp_conn->hdr + tcp_conn->in.hdr_offset,
 159                         copylen);
 160
 161                 debug_tcp("PDU gather offset %d bytes %d in.offset %d "
 162                        "in.copy %d\n", tcp_conn->in.hdr_offset, copylen,
 163                        tcp_conn->in.offset, tcp_conn->in.copy);
 164
 165                 tcp_conn->in.offset += copylen;
 166                 tcp_conn->in.copy -= copylen;
 167                 if (copylen < hdr_remains)  {
 168                         tcp_conn->in_progress = IN_PROGRESS_HEADER_GATHER;
 169                         tcp_conn->in.hdr_offset += copylen;
 170                         return -EAGAIN;
 171                 }
 172                 tcp_conn->in.hdr = &tcp_conn->hdr;
 173                 tcp_conn->discontiguous_hdr_cnt++;
 174                 tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
 175         }
 176
 177         return 0;
 178 }
 179
 180 /*
 181  * must be called with session lock
 182  */
 183 static void
 184 iscsi_tcp_cleanup_ctask(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
 185 {
 186         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 187         struct iscsi_r2t_info *r2t;
 188         struct scsi_cmnd *sc;
 189
 190         /* flush ctask's r2t queues */
 191         while (__kfifo_get(tcp_ctask->r2tqueue, (void*)&r2t, sizeof(void*))) {
 192                 __kfifo_put(tcp_ctask->r2tpool.queue, (void*)&r2t,
 193                             sizeof(void*));
 194                 debug_scsi("iscsi_tcp_cleanup_ctask pending r2t dropped\n");
 195         }
 196
 197         sc = ctask->sc;
 198         if (unlikely(!sc))
 199                 return;
 200
 201         tcp_ctask->xmstate = XMSTATE_IDLE;
 202         tcp_ctask->r2t = NULL;
 203 }
 204
 205 /**
 206  * iscsi_data_rsp - SCSI Data-In Response processing
 207  * @conn: iscsi connection
 208  * @ctask: scsi command task
 209  **/
 210 static int
 211 iscsi_data_rsp(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
 212 {
 213         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 214         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 215         struct iscsi_data_rsp *rhdr = (struct iscsi_data_rsp *)tcp_conn->in.hdr;
 216         struct iscsi_session *session = conn->session;
 217         struct scsi_cmnd *sc = ctask->sc;
 218         int datasn = be32_to_cpu(rhdr->datasn);
 219
 220         iscsi_update_cmdsn(session, (struct iscsi_nopin*)rhdr);
 221         /*
 222          * setup Data-In byte counter (gets decremented..)
 223          */
 224         ctask->data_count = tcp_conn->in.datalen;
 225
 226         if (tcp_conn->in.datalen == 0)
 227                 return 0;
 228
 229         if (tcp_ctask->exp_datasn != datasn) {
 230                 debug_tcp("%s: ctask->exp_datasn(%d) != rhdr->datasn(%d)\n",
 231                           __FUNCTION__, tcp_ctask->exp_datasn, datasn);
 232                 return ISCSI_ERR_DATASN;
 233         }
 234
 235         tcp_ctask->exp_datasn++;
 236
 237         tcp_ctask->data_offset = be32_to_cpu(rhdr->offset);
 238         if (tcp_ctask->data_offset + tcp_conn->in.datalen > sc->request_bufflen) {
 239                 debug_tcp("%s: data_offset(%d) + data_len(%d) > total_length_in(%d)\n",
 240                           __FUNCTION__, tcp_ctask->data_offset,
 241                           tcp_conn->in.datalen, sc->request_bufflen);
 242                 return ISCSI_ERR_DATA_OFFSET;
 243         }
 244
 245         if (rhdr->flags & ISCSI_FLAG_DATA_STATUS) {
 246                 conn->exp_statsn = be32_to_cpu(rhdr->statsn) + 1;
 247                 if (rhdr->flags & ISCSI_FLAG_DATA_UNDERFLOW) {
 248                         int res_count = be32_to_cpu(rhdr->residual_count);
 249
 250                         if (res_count > 0 &&
 251                             res_count <= sc->request_bufflen) {
 252                                 sc->resid = res_count;
 253                                 sc->result = (DID_OK << 16) | rhdr->cmd_status;
 254                         } else
 255                                 sc->result = (DID_BAD_TARGET << 16) |
 256                                         rhdr->cmd_status;
 257                 } else if (rhdr->flags & ISCSI_FLAG_DATA_OVERFLOW) {
 258                         sc->resid = be32_to_cpu(rhdr->residual_count);
 259                         sc->result = (DID_OK << 16) | rhdr->cmd_status;
 260                 } else
 261                         sc->result = (DID_OK << 16) | rhdr->cmd_status;
 262         }
 263
 264         conn->datain_pdus_cnt++;
 265         return 0;
 266 }
 267
 268 /**
 269  * iscsi_solicit_data_init - initialize first Data-Out
 270  * @conn: iscsi connection
 271  * @ctask: scsi command task
 272  * @r2t: R2T info
 273  *
 274  * Notes:
 275  *      Initialize first Data-Out within this R2T sequence and finds
 276  *      proper data_offset within this SCSI command.
 277  *
 278  *      This function is called with connection lock taken.
 279  **/
 280 static void
 281 iscsi_solicit_data_init(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask,
 282                         struct iscsi_r2t_info *r2t)
 283 {
 284         struct iscsi_data *hdr;
 285         struct scsi_cmnd *sc = ctask->sc;
 286
 287         hdr = &r2t->dtask.hdr;
 288         memset(hdr, 0, sizeof(struct iscsi_data));
 289         hdr->ttt = r2t->ttt;
 290         hdr->datasn = cpu_to_be32(r2t->solicit_datasn);
 291         r2t->solicit_datasn++;
 292         hdr->opcode = ISCSI_OP_SCSI_DATA_OUT;
 293         memcpy(hdr->lun, ctask->hdr->lun, sizeof(hdr->lun));
 294         hdr->itt = ctask->hdr->itt;
 295         hdr->exp_statsn = r2t->exp_statsn;
 296         hdr->offset = cpu_to_be32(r2t->data_offset);
 297         if (r2t->data_length > conn->max_xmit_dlength) {
 298                 hton24(hdr->dlength, conn->max_xmit_dlength);
 299                 r2t->data_count = conn->max_xmit_dlength;
 300                 hdr->flags = 0;
 301         } else {
 302                 hton24(hdr->dlength, r2t->data_length);
 303                 r2t->data_count = r2t->data_length;
 304                 hdr->flags = ISCSI_FLAG_CMD_FINAL;
 305         }
 306         conn->dataout_pdus_cnt++;
 307
 308         r2t->sent = 0;
 309
 310         iscsi_buf_init_iov(&r2t->headbuf, (char*)hdr,
 311                            sizeof(struct iscsi_hdr));
 312
 313         if (sc->use_sg) {
 314                 int i, sg_count = 0;
 315                 struct scatterlist *sg = sc->request_buffer;
 316
 317                 r2t->sg = NULL;
 318                 for (i = 0; i < sc->use_sg; i++, sg += 1) {
 319                         /* FIXME: prefetch ? */
 320                         if (sg_count + sg->length > r2t->data_offset) {
 321                                 int page_offset;
 322
 323                                 /* sg page found! */
 324
 325                                 /* offset within this page */
 326                                 page_offset = r2t->data_offset - sg_count;
 327
 328                                 /* fill in this buffer */
 329                                 iscsi_buf_init_sg(&r2t->sendbuf, sg);
 330                                 r2t->sendbuf.sg.offset += page_offset;
 331                                 r2t->sendbuf.sg.length -= page_offset;
 332
 333                                 /* xmit logic will continue with next one */
 334                                 r2t->sg = sg + 1;
 335                                 break;
 336                         }
 337                         sg_count += sg->length;
 338                 }
 339                 BUG_ON(r2t->sg == NULL);
 340         } else {
 341                 iscsi_buf_init_iov(&r2t->sendbuf,
 342                             (char*)sc->request_buffer + r2t->data_offset,
 343                             r2t->data_count);
 344                 r2t->sg = NULL;
 345         }
 346 }
 347
 348 /**
 349  * iscsi_r2t_rsp - iSCSI R2T Response processing
 350  * @conn: iscsi connection
 351  * @ctask: scsi command task
 352  **/
 353 static int
 354 iscsi_r2t_rsp(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
 355 {
 356         struct iscsi_r2t_info *r2t;
 357         struct iscsi_session *session = conn->session;
 358         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 359         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 360         struct iscsi_r2t_rsp *rhdr = (struct iscsi_r2t_rsp *)tcp_conn->in.hdr;
 361         int r2tsn = be32_to_cpu(rhdr->r2tsn);
 362         int rc;
 363
 364         if (tcp_conn->in.datalen) {
 365                 printk(KERN_ERR "iscsi_tcp: invalid R2t with datalen %d\n",
 366                        tcp_conn->in.datalen);
 367                 return ISCSI_ERR_DATALEN;
 368         }
 369
 370         if (tcp_ctask->exp_datasn != r2tsn){
 371                 debug_tcp("%s: ctask->exp_datasn(%d) != rhdr->r2tsn(%d)\n",
 372                           __FUNCTION__, tcp_ctask->exp_datasn, r2tsn);
 373                 return ISCSI_ERR_R2TSN;
 374         }
 375
 376         /* fill-in new R2T associated with the task */
 377         spin_lock(&session->lock);
 378         iscsi_update_cmdsn(session, (struct iscsi_nopin*)rhdr);
 379
 380         if (!ctask->sc || ctask->mtask ||
 381              session->state != ISCSI_STATE_LOGGED_IN) {
 382                 printk(KERN_INFO "iscsi_tcp: dropping R2T itt %d in "
 383                        "recovery...\n", ctask->itt);
 384                 spin_unlock(&session->lock);
 385                 return 0;
 386         }
 387
 388         rc = __kfifo_get(tcp_ctask->r2tpool.queue, (void*)&r2t, sizeof(void*));
 389         BUG_ON(!rc);
 390
 391         r2t->exp_statsn = rhdr->statsn;
 392         r2t->data_length = be32_to_cpu(rhdr->data_length);
 393         if (r2t->data_length == 0) {
 394                 printk(KERN_ERR "iscsi_tcp: invalid R2T with zero data len\n");
 395                 spin_unlock(&session->lock);
 396                 return ISCSI_ERR_DATALEN;
 397         }
 398
 399         if (r2t->data_length > session->max_burst)
 400                 debug_scsi("invalid R2T with data len %u and max burst %u."
 401                            "Attempting to execute request.\n",
 402                             r2t->data_length, session->max_burst);
 403
 404         r2t->data_offset = be32_to_cpu(rhdr->data_offset);
 405         if (r2t->data_offset + r2t->data_length > ctask->sc->request_bufflen) {
 406                 spin_unlock(&session->lock);
 407                 printk(KERN_ERR "iscsi_tcp: invalid R2T with data len %u at "
 408                        "offset %u and total length %d\n", r2t->data_length,
 409                        r2t->data_offset, ctask->sc->request_bufflen);
 410                 return ISCSI_ERR_DATALEN;
 411         }
 412
 413         r2t->ttt = rhdr->ttt; /* no flip */
 414         r2t->solicit_datasn = 0;
 415
 416         iscsi_solicit_data_init(conn, ctask, r2t);
 417
 418         tcp_ctask->exp_datasn = r2tsn + 1;
 419         __kfifo_put(tcp_ctask->r2tqueue, (void*)&r2t, sizeof(void*));
 420         tcp_ctask->xmstate |= XMSTATE_SOL_HDR_INIT;
 421         list_move_tail(&ctask->running, &conn->xmitqueue);
 422
 423         scsi_queue_work(session->host, &conn->xmitwork);
 424         conn->r2t_pdus_cnt++;
 425         spin_unlock(&session->lock);
 426
 427         return 0;
 428 }
 429
 430 static int
 431 iscsi_tcp_hdr_recv(struct iscsi_conn *conn)
 432 {
 433         int rc = 0, opcode, ahslen;
 434         struct iscsi_hdr *hdr;
 435         struct iscsi_session *session = conn->session;
 436         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 437         uint32_t cdgst, rdgst = 0, itt;
 438
 439         hdr = tcp_conn->in.hdr;
 440
 441         /* verify PDU length */
 442         tcp_conn->in.datalen = ntoh24(hdr->dlength);
 443         if (tcp_conn->in.datalen > conn->max_recv_dlength) {
 444                 printk(KERN_ERR "iscsi_tcp: datalen %d > %d\n",
 445                        tcp_conn->in.datalen, conn->max_recv_dlength);
 446                 return ISCSI_ERR_DATALEN;
 447         }
 448         tcp_conn->data_copied = 0;
 449
 450         /* read AHS */
 451         ahslen = hdr->hlength << 2;
 452         tcp_conn->in.offset += ahslen;
 453         tcp_conn->in.copy -= ahslen;
 454         if (tcp_conn->in.copy < 0) {
 455                 printk(KERN_ERR "iscsi_tcp: can't handle AHS with length "
 456                        "%d bytes\n", ahslen);
 457                 return ISCSI_ERR_AHSLEN;
 458         }
 459
 460         /* calculate read padding */
 461         tcp_conn->in.padding = tcp_conn->in.datalen & (ISCSI_PAD_LEN-1);
 462         if (tcp_conn->in.padding) {
 463                 tcp_conn->in.padding = ISCSI_PAD_LEN - tcp_conn->in.padding;
 464                 debug_scsi("read padding %d bytes\n", tcp_conn->in.padding);
 465         }
 466
 467         if (conn->hdrdgst_en) {
 468                 struct scatterlist sg;
 469
 470                 sg_init_one(&sg, (u8 *)hdr,
 471                             sizeof(struct iscsi_hdr) + ahslen);
 472                 crypto_hash_digest(&tcp_conn->rx_hash, &sg, sg.length,
 473                                    (u8 *)&cdgst);
 474                 rdgst = *(uint32_t*)((char*)hdr + sizeof(struct iscsi_hdr) +
 475                                      ahslen);
 476                 if (cdgst != rdgst) {
 477                         printk(KERN_ERR "iscsi_tcp: hdrdgst error "
 478                                "recv 0x%x calc 0x%x\n", rdgst, cdgst);
 479                         return ISCSI_ERR_HDR_DGST;
 480                 }
 481         }
 482
 483         opcode = hdr->opcode & ISCSI_OPCODE_MASK;
 484         /* verify itt (itt encoding: age+cid+itt) */
 485         rc = iscsi_verify_itt(conn, hdr, &itt);
 486         if (rc == ISCSI_ERR_NO_SCSI_CMD) {
 487                 tcp_conn->in.datalen = 0; /* force drop */
 488                 return 0;
 489         } else if (rc)
 490                 return rc;
 491
 492         debug_tcp("opcode 0x%x offset %d copy %d ahslen %d datalen %d\n",
 493                   opcode, tcp_conn->in.offset, tcp_conn->in.copy,
 494                   ahslen, tcp_conn->in.datalen);
 495
 496         switch(opcode) {
 497         case ISCSI_OP_SCSI_DATA_IN:
 498                 tcp_conn->in.ctask = session->cmds[itt];
 499                 rc = iscsi_data_rsp(conn, tcp_conn->in.ctask);
 500                 if (rc)
 501                         return rc;
 502                 /* fall through */
 503         case ISCSI_OP_SCSI_CMD_RSP:
 504                 tcp_conn->in.ctask = session->cmds[itt];
 505                 if (tcp_conn->in.datalen)
 506                         goto copy_hdr;
 507
 508                 spin_lock(&session->lock);
 509                 rc = __iscsi_complete_pdu(conn, hdr, NULL, 0);
 510                 spin_unlock(&session->lock);
 511                 break;
 512         case ISCSI_OP_R2T:
 513                 tcp_conn->in.ctask = session->cmds[itt];
 514                 if (ahslen)
 515                         rc = ISCSI_ERR_AHSLEN;
 516                 else if (tcp_conn->in.ctask->sc->sc_data_direction ==
 517                                                                 DMA_TO_DEVICE)
 518                         rc = iscsi_r2t_rsp(conn, tcp_conn->in.ctask);
 519                 else
 520                         rc = ISCSI_ERR_PROTO;
 521                 break;
 522         case ISCSI_OP_LOGIN_RSP:
 523         case ISCSI_OP_TEXT_RSP:
 524         case ISCSI_OP_REJECT:
 525         case ISCSI_OP_ASYNC_EVENT:
 526                 /*
 527                  * It is possible that we could get a PDU with a buffer larger
 528                  * than 8K, but there are no targets that currently do this.
 529                  * For now we fail until we find a vendor that needs it
 530                  */
 531                 if (ISCSI_DEF_MAX_RECV_SEG_LEN <
 532                     tcp_conn->in.datalen) {
 533                         printk(KERN_ERR "iscsi_tcp: received buffer of len %u "
 534                               "but conn buffer is only %u (opcode %0x)\n",
 535                               tcp_conn->in.datalen,
 536                               ISCSI_DEF_MAX_RECV_SEG_LEN, opcode);
 537                         rc = ISCSI_ERR_PROTO;
 538                         break;
 539                 }
 540
 541                 if (tcp_conn->in.datalen)
 542                         goto copy_hdr;
 543         /* fall through */
 544         case ISCSI_OP_LOGOUT_RSP:
 545         case ISCSI_OP_NOOP_IN:
 546         case ISCSI_OP_SCSI_TMFUNC_RSP:
 547                 rc = iscsi_complete_pdu(conn, hdr, NULL, 0);
 548                 break;
 549         default:
 550                 rc = ISCSI_ERR_BAD_OPCODE;
 551                 break;
 552         }
 553
 554         return rc;
 555
 556 copy_hdr:
 557         /*
 558          * if we did zero copy for the header but we will need multiple
 559          * skbs to complete the command then we have to copy the header
 560          * for later use
 561          */
 562         if (tcp_conn->in.zero_copy_hdr && tcp_conn->in.copy <=
 563            (tcp_conn->in.datalen + tcp_conn->in.padding +
 564             (conn->datadgst_en ? 4 : 0))) {
 565                 debug_tcp("Copying header for later use. in.copy %d in.datalen"
 566                           " %d\n", tcp_conn->in.copy, tcp_conn->in.datalen);
 567                 memcpy(&tcp_conn->hdr, tcp_conn->in.hdr,
 568                        sizeof(struct iscsi_hdr));
 569                 tcp_conn->in.hdr = &tcp_conn->hdr;
 570                 tcp_conn->in.zero_copy_hdr = 0;
 571         }
 572         return 0;
 573 }
 574
 575 /**
 576  * iscsi_ctask_copy - copy skb bits to the destanation cmd task
 577  * @conn: iscsi tcp connection
 578  * @ctask: scsi command task
 579  * @buf: buffer to copy to
 580  * @buf_size: size of buffer
 581  * @offset: offset within the buffer
 582  *
 583  * Notes:
 584  *      The function calls skb_copy_bits() and updates per-connection and
 585  *      per-cmd byte counters.
 586  *
 587  *      Read counters (in bytes):
 588  *
 589  *      conn->in.offset         offset within in progress SKB
 590  *      conn->in.copy           left to copy from in progress SKB
 591  *                              including padding
 592  *      conn->in.copied         copied already from in progress SKB
 593  *      conn->data_copied       copied already from in progress buffer
 594  *      ctask->sent             total bytes sent up to the MidLayer
 595  *      ctask->data_count       left to copy from in progress Data-In
 596  *      buf_left                left to copy from in progress buffer
 597  **/
 598 static inline int
 599 iscsi_ctask_copy(struct iscsi_tcp_conn *tcp_conn, struct iscsi_cmd_task *ctask,
 600                 void *buf, int buf_size, int offset)
 601 {
 602         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 603         int buf_left = buf_size - (tcp_conn->data_copied + offset);
 604         unsigned size = min(tcp_conn->in.copy, buf_left);
 605         int rc;
 606
 607         size = min(size, ctask->data_count);
 608
 609         debug_tcp("ctask_copy %d bytes at offset %d copied %d\n",
 610                size, tcp_conn->in.offset, tcp_conn->in.copied);
 611
 612         BUG_ON(size <= 0);
 613         BUG_ON(tcp_ctask->sent + size > ctask->sc->request_bufflen);
 614
 615         rc = skb_copy_bits(tcp_conn->in.skb, tcp_conn->in.offset,
 616                            (char*)buf + (offset + tcp_conn->data_copied), size);
 617         /* must fit into skb->len */
 618         BUG_ON(rc);
 619
 620         tcp_conn->in.offset += size;
 621         tcp_conn->in.copy -= size;
 622         tcp_conn->in.copied += size;
 623         tcp_conn->data_copied += size;
 624         tcp_ctask->sent += size;
 625         ctask->data_count -= size;
 626
 627         BUG_ON(tcp_conn->in.copy < 0);
 628         BUG_ON(ctask->data_count < 0);
 629
 630         if (buf_size != (tcp_conn->data_copied + offset)) {
 631                 if (!ctask->data_count) {
 632                         BUG_ON(buf_size - tcp_conn->data_copied < 0);
 633                         /* done with this PDU */
 634                         return buf_size - tcp_conn->data_copied;
 635                 }
 636                 return -EAGAIN;
 637         }
 638
 639         /* done with this buffer or with both - PDU and buffer */
 640         tcp_conn->data_copied = 0;
 641         return 0;
 642 }
 643
 644 /**
 645  * iscsi_tcp_copy - copy skb bits to the destanation buffer
 646  * @conn: iscsi tcp connection
 647  *
 648  * Notes:
 649  *      The function calls skb_copy_bits() and updates per-connection
 650  *      byte counters.
 651  **/
 652 static inline int
 653 iscsi_tcp_copy(struct iscsi_conn *conn, int buf_size)
 654 {
 655         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 656         int buf_left = buf_size - tcp_conn->data_copied;
 657         int size = min(tcp_conn->in.copy, buf_left);
 658         int rc;
 659
 660         debug_tcp("tcp_copy %d bytes at offset %d copied %d\n",
 661                size, tcp_conn->in.offset, tcp_conn->data_copied);
 662         BUG_ON(size <= 0);
 663
 664         rc = skb_copy_bits(tcp_conn->in.skb, tcp_conn->in.offset,
 665                            (char*)conn->data + tcp_conn->data_copied, size);
 666         BUG_ON(rc);
 667
 668         tcp_conn->in.offset += size;
 669         tcp_conn->in.copy -= size;
 670         tcp_conn->in.copied += size;
 671         tcp_conn->data_copied += size;
 672
 673         if (buf_size != tcp_conn->data_copied)
 674                 return -EAGAIN;
 675
 676         return 0;
 677 }
 678
 679 static inline void
 680 partial_sg_digest_update(struct hash_desc *desc, struct scatterlist *sg,
 681                          int offset, int length)
 682 {
 683         struct scatterlist temp;
 684
 685         memcpy(&temp, sg, sizeof(struct scatterlist));
 686         temp.offset = offset;
 687         temp.length = length;
 688         crypto_hash_update(desc, &temp, length);
 689 }
 690
 691 static void
 692 iscsi_recv_digest_update(struct iscsi_tcp_conn *tcp_conn, char* buf, int len)
 693 {
 694         struct scatterlist tmp;
 695
 696         sg_init_one(&tmp, buf, len);
 697         crypto_hash_update(&tcp_conn->rx_hash, &tmp, len);
 698 }
 699
 700 static int iscsi_scsi_data_in(struct iscsi_conn *conn)
 701 {
 702         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 703         struct iscsi_cmd_task *ctask = tcp_conn->in.ctask;
 704         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
 705         struct scsi_cmnd *sc = ctask->sc;
 706         struct scatterlist *sg;
 707         int i, offset, rc = 0;
 708
 709         BUG_ON((void*)ctask != sc->SCp.ptr);
 710
 711         /*
 712          * copying Data-In into the Scsi_Cmnd
 713          */
 714         if (!sc->use_sg) {
 715                 i = ctask->data_count;
 716                 rc = iscsi_ctask_copy(tcp_conn, ctask, sc->request_buffer,
 717                                       sc->request_bufflen,
 718                                       tcp_ctask->data_offset);
 719                 if (rc == -EAGAIN)
 720                         return rc;
 721                 if (conn->datadgst_en)
 722                         iscsi_recv_digest_update(tcp_conn, sc->request_buffer,
 723                                                  i);
 724                 rc = 0;
 725                 goto done;
 726         }
 727
 728         offset = tcp_ctask->data_offset;
 729         sg = sc->request_buffer;
 730
 731         if (tcp_ctask->data_offset)
 732                 for (i = 0; i < tcp_ctask->sg_count; i++)
 733                         offset -= sg[i].length;
 734         /* we've passed through partial sg*/
 735         if (offset < 0)
 736                 offset = 0;
 737
 738         for (i = tcp_ctask->sg_count; i < sc->use_sg; i++) {
 739                 char *dest;
 740
 741                 dest = kmap_atomic(sg[i].page, KM_SOFTIRQ0);
 742                 rc = iscsi_ctask_copy(tcp_conn, ctask, dest + sg[i].offset,
 743                                       sg[i].length, offset);
 744                 kunmap_atomic(dest, KM_SOFTIRQ0);
 745                 if (rc == -EAGAIN)
 746                         /* continue with the next SKB/PDU */
 747                         return rc;
 748                 if (!rc) {
 749                         if (conn->datadgst_en) {
 750                                 if (!offset)
 751                                         crypto_hash_update(
 752                                                         &tcp_conn->rx_hash,
 753                                                         &sg[i], sg[i].length);
 754                                 else
 755                                         partial_sg_digest_update(
 756                                                         &tcp_conn->rx_hash,
 757                                                         &sg[i],
 758                                                         sg[i].offset + offset,
 759                                                         sg[i].length - offset);
 760                         }
 761                         offset = 0;
 762                         tcp_ctask->sg_count++;
 763                 }
 764
 765                 if (!ctask->data_count) {
 766                         if (rc && conn->datadgst_en)
 767                                 /*
 768                                  * data-in is complete, but buffer not...
 769                                  */
 770                                 partial_sg_digest_update(&tcp_conn->rx_hash,
 771                                                          &sg[i],
 772                                                          sg[i].offset,
 773                                                          sg[i].length-rc);
 774                         rc = 0;
 775                         break;
 776                 }
 777
 778                 if (!tcp_conn->in.copy)
 779                         return -EAGAIN;
 780         }
 781         BUG_ON(ctask->data_count);
 782
 783 done:
 784         /* check for non-exceptional status */
 785         if (tcp_conn->in.hdr->flags & ISCSI_FLAG_DATA_STATUS) {
 786                 debug_scsi("done [sc %lx res %d itt 0x%x flags 0x%x]\n",
 787                            (long)sc, sc->result, ctask->itt,
 788                            tcp_conn->in.hdr->flags);
 789                 spin_lock(&conn->session->lock);
 790                 __iscsi_complete_pdu(conn, tcp_conn->in.hdr, NULL, 0);
 791                 spin_unlock(&conn->session->lock);
 792         }
 793
 794         return rc;
 795 }
 796
 797 static int
 798 iscsi_data_recv(struct iscsi_conn *conn)
 799 {
 800         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 801         int rc = 0, opcode;
 802
 803         opcode = tcp_conn->in.hdr->opcode & ISCSI_OPCODE_MASK;
 804         switch (opcode) {
 805         case ISCSI_OP_SCSI_DATA_IN:
 806                 rc = iscsi_scsi_data_in(conn);
 807                 break;
 808         case ISCSI_OP_SCSI_CMD_RSP:
 809         case ISCSI_OP_TEXT_RSP:
 810         case ISCSI_OP_LOGIN_RSP:
 811         case ISCSI_OP_ASYNC_EVENT:
 812         case ISCSI_OP_REJECT:
 813                 /*
 814                  * Collect data segment to the connection's data
 815                  * placeholder
 816                  */
 817                 if (iscsi_tcp_copy(conn, tcp_conn->in.datalen)) {
 818                         rc = -EAGAIN;
 819                         goto exit;
 820                 }
 821
 822                 rc = iscsi_complete_pdu(conn, tcp_conn->in.hdr, conn->data,
 823                                         tcp_conn->in.datalen);
 824                 if (!rc && conn->datadgst_en && opcode != ISCSI_OP_LOGIN_RSP)
 825                         iscsi_recv_digest_update(tcp_conn, conn->data,
 826                                                 tcp_conn->in.datalen);
 827                 break;
 828         default:
 829                 BUG_ON(1);
 830         }
 831 exit:
 832         return rc;
 833 }
 834
 835 /**
 836  * iscsi_tcp_data_recv - TCP receive in sendfile fashion
 837  * @rd_desc: read descriptor
 838  * @skb: socket buffer
 839  * @offset: offset in skb
 840  * @len: skb->len - offset
 841  **/
 842 static int
 843 iscsi_tcp_data_recv(read_descriptor_t *rd_desc, struct sk_buff *skb,
 844                 unsigned int offset, size_t len)
 845 {
 846         int rc;
 847         struct iscsi_conn *conn = rd_desc->arg.data;
 848         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
 849         int processed;
 850         char pad[ISCSI_PAD_LEN];
 851         struct scatterlist sg;
 852
 853         /*
 854          * Save current SKB and its offset in the corresponding
 855          * connection context.
 856          */
 857         tcp_conn->in.copy = skb->len - offset;
 858         tcp_conn->in.offset = offset;
 859         tcp_conn->in.skb = skb;
 860         tcp_conn->in.len = tcp_conn->in.copy;
 861         BUG_ON(tcp_conn->in.copy <= 0);
 862         debug_tcp("in %d bytes\n", tcp_conn->in.copy);
 863
 864 more:
 865         tcp_conn->in.copied = 0;
 866         rc = 0;
 867
 868         if (unlikely(conn->suspend_rx)) {
 869                 debug_tcp("conn %d Rx suspended!\n", conn->id);
 870                 return 0;
 871         }
 872
 873         if (tcp_conn->in_progress == IN_PROGRESS_WAIT_HEADER ||
 874             tcp_conn->in_progress == IN_PROGRESS_HEADER_GATHER) {
 875                 rc = iscsi_hdr_extract(tcp_conn);
 876                 if (rc) {
 877                        if (rc == -EAGAIN)
 878                                 goto nomore;
 879                        else {
 880                                 iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
 881                                 return 0;
 882                        }
 883                 }
 884
 885                 /*
 886                  * Verify and process incoming PDU header.
 887                  */
 888                 rc = iscsi_tcp_hdr_recv(conn);
 889                 if (!rc && tcp_conn->in.datalen) {
 890                         if (conn->datadgst_en)
 891                                 crypto_hash_init(&tcp_conn->rx_hash);
 892                         tcp_conn->in_progress = IN_PROGRESS_DATA_RECV;
 893                 } else if (rc) {
 894                         iscsi_conn_failure(conn, rc);
 895                         return 0;
 896                 }
 897         }
 898
 899         if (tcp_conn->in_progress == IN_PROGRESS_DDIGEST_RECV) {
 900                 uint32_t recv_digest;
 901
 902                 debug_tcp("extra data_recv offset %d copy %d\n",
 903                           tcp_conn->in.offset, tcp_conn->in.copy);
 904                 rc = iscsi_tcp_copy(conn, sizeof(uint32_t));
 905                 if (rc) {
 906                         if (rc == -EAGAIN)
 907                                 goto again;
 908                         iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
 909                         return 0;
 910                 }
 911
 912                 memcpy(&recv_digest, conn->data, sizeof(uint32_t));
 913                 if (recv_digest != tcp_conn->in.datadgst) {
 914                         debug_tcp("iscsi_tcp: data digest error!"
 915                                   "0x%x != 0x%x\n", recv_digest,
 916                                   tcp_conn->in.datadgst);
 917                         iscsi_conn_failure(conn, ISCSI_ERR_DATA_DGST);
 918                         return 0;
 919                 } else {
 920                         debug_tcp("iscsi_tcp: data digest match!"
 921                                   "0x%x == 0x%x\n", recv_digest,
 922                                   tcp_conn->in.datadgst);
 923                         tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
 924                 }
 925         }
 926
 927         if (tcp_conn->in_progress == IN_PROGRESS_DATA_RECV &&
 928            tcp_conn->in.copy) {
 929
 930                 debug_tcp("data_recv offset %d copy %d\n",
 931                        tcp_conn->in.offset, tcp_conn->in.copy);
 932
 933                 rc = iscsi_data_recv(conn);
 934                 if (rc) {
 935                         if (rc == -EAGAIN)
 936                                 goto again;
 937                         iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
 938                         return 0;
 939                 }
 940                 tcp_conn->in.copy -= tcp_conn->in.padding;
 941                 tcp_conn->in.offset += tcp_conn->in.padding;
 942                 if (conn->datadgst_en) {
 943                         if (tcp_conn->in.padding) {
 944                                 debug_tcp("padding -> %d\n",
 945                                           tcp_conn->in.padding);
 946                                 memset(pad, 0, tcp_conn->in.padding);
 947                                 sg_init_one(&sg, pad, tcp_conn->in.padding);
 948                                 crypto_hash_update(&tcp_conn->rx_hash,
 949                                                    &sg, sg.length);
 950                         }
 951                         crypto_hash_final(&tcp_conn->rx_hash,
 952                                           (u8 *) &tcp_conn->in.datadgst);
 953                         debug_tcp("rx digest 0x%x\n", tcp_conn->in.datadgst);
 954                         tcp_conn->in_progress = IN_PROGRESS_DDIGEST_RECV;
 955                         tcp_conn->data_copied = 0;
 956                 } else
 957                         tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
 958         }
 959
 960         debug_tcp("f, processed %d from out of %d padding %d\n",
 961                tcp_conn->in.offset - offset, (int)len, tcp_conn->in.padding);
 962         BUG_ON(tcp_conn->in.offset - offset > len);
 963
 964         if (tcp_conn->in.offset - offset != len) {
 965                 debug_tcp("continue to process %d bytes\n",
 966                        (int)len - (tcp_conn->in.offset - offset));
 967                 goto more;
 968         }
 969
 970 nomore:
 971         processed = tcp_conn->in.offset - offset;
 972         BUG_ON(processed == 0);
 973         return processed;
 974
 975 again:
 976         processed = tcp_conn->in.offset - offset;
 977         debug_tcp("c, processed %d from out of %d rd_desc_cnt %d\n",
 978                   processed, (int)len, (int)rd_desc->count);
 979         BUG_ON(processed == 0);
 980         BUG_ON(processed > len);
 981
 982         conn->rxdata_octets += processed;
 983         return processed;
 984 }
 985
 986 static void
 987 iscsi_tcp_data_ready(struct sock *sk, int flag)
 988 {
 989         struct iscsi_conn *conn = sk->sk_user_data;
 990         read_descriptor_t rd_desc;
 991
 992         read_lock(&sk->sk_callback_lock);
 993
 994         /*
 995          * Use rd_desc to pass 'conn' to iscsi_tcp_data_recv.
 996          * We set count to 1 because we want the network layer to
 997          * hand us all the skbs that are available. iscsi_tcp_data_recv
 998          * handled pdus that cross buffers or pdus that still need data.
 999          */
1000         rd_desc.arg.data = conn;
1001         rd_desc.count = 1;
1002         tcp_read_sock(sk, &rd_desc, iscsi_tcp_data_recv);
1003
1004         read_unlock(&sk->sk_callback_lock);
1005 }
1006
1007 static void
1008 iscsi_tcp_state_change(struct sock *sk)
1009 {
1010         struct iscsi_tcp_conn *tcp_conn;
1011         struct iscsi_conn *conn;
1012         struct iscsi_session *session;
1013         void (*old_state_change)(struct sock *);
1014
1015         read_lock(&sk->sk_callback_lock);
1016
1017         conn = (struct iscsi_conn*)sk->sk_user_data;
1018         session = conn->session;
1019
1020         if ((sk->sk_state == TCP_CLOSE_WAIT ||
1021              sk->sk_state == TCP_CLOSE) &&
1022             !atomic_read(&sk->sk_rmem_alloc)) {
1023                 debug_tcp("iscsi_tcp_state_change: TCP_CLOSE|TCP_CLOSE_WAIT\n");
1024                 iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
1025         }
1026
1027         tcp_conn = conn->dd_data;
1028         old_state_change = tcp_conn->old_state_change;
1029
1030         read_unlock(&sk->sk_callback_lock);
1031
1032         old_state_change(sk);
1033 }
1034
1035 /**
1036  * iscsi_write_space - Called when more output buffer space is available
1037  * @sk: socket space is available for
1038  **/
1039 static void
1040 iscsi_write_space(struct sock *sk)
1041 {
1042         struct iscsi_conn *conn = (struct iscsi_conn*)sk->sk_user_data;
1043         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1044
1045         tcp_conn->old_write_space(sk);
1046         debug_tcp("iscsi_write_space: cid %d\n", conn->id);
1047         scsi_queue_work(conn->session->host, &conn->xmitwork);
1048 }
1049
1050 static void
1051 iscsi_conn_set_callbacks(struct iscsi_conn *conn)
1052 {
1053         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1054         struct sock *sk = tcp_conn->sock->sk;
1055
1056         /* assign new callbacks */
1057         write_lock_bh(&sk->sk_callback_lock);
1058         sk->sk_user_data = conn;
1059         tcp_conn->old_data_ready = sk->sk_data_ready;
1060         tcp_conn->old_state_change = sk->sk_state_change;
1061         tcp_conn->old_write_space = sk->sk_write_space;
1062         sk->sk_data_ready = iscsi_tcp_data_ready;
1063         sk->sk_state_change = iscsi_tcp_state_change;
1064         sk->sk_write_space = iscsi_write_space;
1065         write_unlock_bh(&sk->sk_callback_lock);
1066 }
1067
1068 static void
1069 iscsi_conn_restore_callbacks(struct iscsi_tcp_conn *tcp_conn)
1070 {
1071         struct sock *sk = tcp_conn->sock->sk;
1072
1073         /* restore socket callbacks, see also: iscsi_conn_set_callbacks() */
1074         write_lock_bh(&sk->sk_callback_lock);
1075         sk->sk_user_data    = NULL;
1076         sk->sk_data_ready   = tcp_conn->old_data_ready;
1077         sk->sk_state_change = tcp_conn->old_state_change;
1078         sk->sk_write_space  = tcp_conn->old_write_space;
1079         sk->sk_no_check  = 0;
1080         write_unlock_bh(&sk->sk_callback_lock);
1081 }
1082
1083 /**
1084  * iscsi_send - generic send routine
1085  * @sk: kernel's socket
1086  * @buf: buffer to write from
1087  * @size: actual size to write
1088  * @flags: socket's flags
1089  */
1090 static inline int
1091 iscsi_send(struct iscsi_conn *conn, struct iscsi_buf *buf, int size, int flags)
1092 {
1093         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1094         struct socket *sk = tcp_conn->sock;
1095         int offset = buf->sg.offset + buf->sent, res;
1096
1097         /*
1098          * if we got use_sg=0 or are sending something we kmallocd
1099          * then we did not have to do kmap (kmap returns page_address)
1100          *
1101          * if we got use_sg > 0, but had to drop down, we do not
1102          * set clustering so this should only happen for that
1103          * slab case.
1104          */
1105         if (buf->use_sendmsg)
1106                 res = sock_no_sendpage(sk, buf->sg.page, offset, size, flags);
1107         else
1108                 res = tcp_conn->sendpage(sk, buf->sg.page, offset, size, flags);
1109
1110         if (res >= 0) {
1111                 conn->txdata_octets += res;
1112                 buf->sent += res;
1113                 return res;
1114         }
1115
1116         tcp_conn->sendpage_failures_cnt++;
1117         if (res == -EAGAIN)
1118                 res = -ENOBUFS;
1119         else
1120                 iscsi_conn_failure(conn, ISCSI_ERR_CONN_FAILED);
1121         return res;
1122 }
1123
1124 /**
1125  * iscsi_sendhdr - send PDU Header via tcp_sendpage()
1126  * @conn: iscsi connection
1127  * @buf: buffer to write from
1128  * @datalen: lenght of data to be sent after the header
1129  *
1130  * Notes:
1131  *      (Tx, Fast Path)
1132  **/
1133 static inline int
1134 iscsi_sendhdr(struct iscsi_conn *conn, struct iscsi_buf *buf, int datalen)
1135 {
1136         int flags = 0; /* MSG_DONTWAIT; */
1137         int res, size;
1138
1139         size = buf->sg.length - buf->sent;
1140         BUG_ON(buf->sent + size > buf->sg.length);
1141         if (buf->sent + size != buf->sg.length || datalen)
1142                 flags |= MSG_MORE;
1143
1144         res = iscsi_send(conn, buf, size, flags);
1145         debug_tcp("sendhdr %d bytes, sent %d res %d\n", size, buf->sent, res);
1146         if (res >= 0) {
1147                 if (size != res)
1148                         return -EAGAIN;
1149                 return 0;
1150         }
1151
1152         return res;
1153 }
1154
1155 /**
1156  * iscsi_sendpage - send one page of iSCSI Data-Out.
1157  * @conn: iscsi connection
1158  * @buf: buffer to write from
1159  * @count: remaining data
1160  * @sent: number of bytes sent
1161  *
1162  * Notes:
1163  *      (Tx, Fast Path)
1164  **/
1165 static inline int
1166 iscsi_sendpage(struct iscsi_conn *conn, struct iscsi_buf *buf,
1167                int *count, int *sent)
1168 {
1169         int flags = 0; /* MSG_DONTWAIT; */
1170         int res, size;
1171
1172         size = buf->sg.length - buf->sent;
1173         BUG_ON(buf->sent + size > buf->sg.length);
1174         if (size > *count)
1175                 size = *count;
1176         if (buf->sent + size != buf->sg.length || *count != size)
1177                 flags |= MSG_MORE;
1178
1179         res = iscsi_send(conn, buf, size, flags);
1180         debug_tcp("sendpage: %d bytes, sent %d left %d sent %d res %d\n",
1181                   size, buf->sent, *count, *sent, res);
1182         if (res >= 0) {
1183                 *count -= res;
1184                 *sent += res;
1185                 if (size != res)
1186                         return -EAGAIN;
1187                 return 0;
1188         }
1189
1190         return res;
1191 }
1192
1193 static inline void
1194 iscsi_data_digest_init(struct iscsi_tcp_conn *tcp_conn,
1195                       struct iscsi_tcp_cmd_task *tcp_ctask)
1196 {
1197         crypto_hash_init(&tcp_conn->tx_hash);
1198         tcp_ctask->digest_count = 4;
1199 }
1200
1201 /**
1202  * iscsi_solicit_data_cont - initialize next Data-Out
1203  * @conn: iscsi connection
1204  * @ctask: scsi command task
1205  * @r2t: R2T info
1206  * @left: bytes left to transfer
1207  *
1208  * Notes:
1209  *      Initialize next Data-Out within this R2T sequence and continue
1210  *      to process next Scatter-Gather element(if any) of this SCSI command.
1211  *
1212  *      Called under connection lock.
1213  **/
1214 static void
1215 iscsi_solicit_data_cont(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask,
1216                         struct iscsi_r2t_info *r2t, int left)
1217 {
1218         struct iscsi_data *hdr;
1219         struct scsi_cmnd *sc = ctask->sc;
1220         int new_offset;
1221
1222         hdr = &r2t->dtask.hdr;
1223         memset(hdr, 0, sizeof(struct iscsi_data));
1224         hdr->ttt = r2t->ttt;
1225         hdr->datasn = cpu_to_be32(r2t->solicit_datasn);
1226         r2t->solicit_datasn++;
1227         hdr->opcode = ISCSI_OP_SCSI_DATA_OUT;
1228         memcpy(hdr->lun, ctask->hdr->lun, sizeof(hdr->lun));
1229         hdr->itt = ctask->hdr->itt;
1230         hdr->exp_statsn = r2t->exp_statsn;
1231         new_offset = r2t->data_offset + r2t->sent;
1232         hdr->offset = cpu_to_be32(new_offset);
1233         if (left > conn->max_xmit_dlength) {
1234                 hton24(hdr->dlength, conn->max_xmit_dlength);
1235                 r2t->data_count = conn->max_xmit_dlength;
1236         } else {
1237                 hton24(hdr->dlength, left);
1238                 r2t->data_count = left;
1239                 hdr->flags = ISCSI_FLAG_CMD_FINAL;
1240         }
1241         conn->dataout_pdus_cnt++;
1242
1243         iscsi_buf_init_iov(&r2t->headbuf, (char*)hdr,
1244                            sizeof(struct iscsi_hdr));
1245
1246         if (iscsi_buf_left(&r2t->sendbuf))
1247                 return;
1248
1249         if (sc->use_sg) {
1250                 iscsi_buf_init_sg(&r2t->sendbuf, r2t->sg);
1251                 r2t->sg += 1;
1252         } else {
1253                 iscsi_buf_init_iov(&r2t->sendbuf,
1254                             (char*)sc->request_buffer + new_offset,
1255                             r2t->data_count);
1256                 r2t->sg = NULL;
1257         }
1258 }
1259
1260 static void iscsi_set_padding(struct iscsi_tcp_cmd_task *tcp_ctask,
1261                               unsigned long len)
1262 {
1263         tcp_ctask->pad_count = len & (ISCSI_PAD_LEN - 1);
1264         if (!tcp_ctask->pad_count)
1265                 return;
1266
1267         tcp_ctask->pad_count = ISCSI_PAD_LEN - tcp_ctask->pad_count;
1268         debug_scsi("write padding %d bytes\n", tcp_ctask->pad_count);
1269         tcp_ctask->xmstate |= XMSTATE_W_PAD;
1270 }
1271
1272 /**
1273  * iscsi_tcp_cmd_init - Initialize iSCSI SCSI_READ or SCSI_WRITE commands
1274  * @conn: iscsi connection
1275  * @ctask: scsi command task
1276  * @sc: scsi command
1277  **/
1278 static void
1279 iscsi_tcp_cmd_init(struct iscsi_cmd_task *ctask)
1280 {
1281         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1282
1283         BUG_ON(__kfifo_len(tcp_ctask->r2tqueue));
1284         tcp_ctask->xmstate = XMSTATE_CMD_HDR_INIT;
1285 }
1286
1287 /**
1288  * iscsi_tcp_mtask_xmit - xmit management(immediate) task
1289  * @conn: iscsi connection
1290  * @mtask: task management task
1291  *
1292  * Notes:
1293  *      The function can return -EAGAIN in which case caller must
1294  *      call it again later, or recover. '0' return code means successful
1295  *      xmit.
1296  *
1297  *      Management xmit state machine consists of these states:
1298  *              XMSTATE_IMM_HDR_INIT    - calculate digest of PDU Header
1299  *              XMSTATE_IMM_HDR         - PDU Header xmit in progress
1300  *              XMSTATE_IMM_DATA        - PDU Data xmit in progress
1301  *              XMSTATE_IDLE            - management PDU is done
1302  **/
1303 static int
1304 iscsi_tcp_mtask_xmit(struct iscsi_conn *conn, struct iscsi_mgmt_task *mtask)
1305 {
1306         struct iscsi_tcp_mgmt_task *tcp_mtask = mtask->dd_data;
1307         int rc;
1308
1309         debug_scsi("mtask deq [cid %d state %x itt 0x%x]\n",
1310                 conn->id, tcp_mtask->xmstate, mtask->itt);
1311
1312         if (tcp_mtask->xmstate & XMSTATE_IMM_HDR_INIT) {
1313                 iscsi_buf_init_iov(&tcp_mtask->headbuf, (char*)mtask->hdr,
1314                                    sizeof(struct iscsi_hdr));
1315
1316                 if (mtask->data_count) {
1317                         tcp_mtask->xmstate |= XMSTATE_IMM_DATA;
1318                         iscsi_buf_init_iov(&tcp_mtask->sendbuf,
1319                                            (char*)mtask->data,
1320                                            mtask->data_count);
1321                 }
1322
1323                 if (conn->c_stage != ISCSI_CONN_INITIAL_STAGE &&
1324                     conn->stop_stage != STOP_CONN_RECOVER &&
1325                     conn->hdrdgst_en)
1326                         iscsi_hdr_digest(conn, &tcp_mtask->headbuf,
1327                                         (u8*)tcp_mtask->hdrext);
1328
1329                 tcp_mtask->sent = 0;
1330                 tcp_mtask->xmstate &= ~XMSTATE_IMM_HDR_INIT;
1331                 tcp_mtask->xmstate |= XMSTATE_IMM_HDR;
1332         }
1333
1334         if (tcp_mtask->xmstate & XMSTATE_IMM_HDR) {
1335                 rc = iscsi_sendhdr(conn, &tcp_mtask->headbuf,
1336                                    mtask->data_count);
1337                 if (rc)
1338                         return rc;
1339                 tcp_mtask->xmstate &= ~XMSTATE_IMM_HDR;
1340         }
1341
1342         if (tcp_mtask->xmstate & XMSTATE_IMM_DATA) {
1343                 BUG_ON(!mtask->data_count);
1344                 tcp_mtask->xmstate &= ~XMSTATE_IMM_DATA;
1345                 /* FIXME: implement.
1346                  * Virtual buffer could be spreaded across multiple pages...
1347                  */
1348                 do {
1349                         int rc;
1350
1351                         rc = iscsi_sendpage(conn, &tcp_mtask->sendbuf,
1352                                         &mtask->data_count, &tcp_mtask->sent);
1353                         if (rc) {
1354                                 tcp_mtask->xmstate |= XMSTATE_IMM_DATA;
1355                                 return rc;
1356                         }
1357                 } while (mtask->data_count);
1358         }
1359
1360         BUG_ON(tcp_mtask->xmstate != XMSTATE_IDLE);
1361         if (mtask->hdr->itt == RESERVED_ITT) {
1362                 struct iscsi_session *session = conn->session;
1363
1364                 spin_lock_bh(&session->lock);
1365                 list_del(&conn->mtask->running);
1366                 __kfifo_put(session->mgmtpool.queue, (void*)&conn->mtask,
1367                             sizeof(void*));
1368                 spin_unlock_bh(&session->lock);
1369         }
1370         return 0;
1371 }
1372
1373 static int
1374 iscsi_send_cmd_hdr(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1375 {
1376         struct scsi_cmnd *sc = ctask->sc;
1377         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1378         int rc = 0;
1379
1380         if (tcp_ctask->xmstate & XMSTATE_CMD_HDR_INIT) {
1381                 tcp_ctask->sent = 0;
1382                 tcp_ctask->sg_count = 0;
1383                 tcp_ctask->exp_datasn = 0;
1384
1385                 if (sc->sc_data_direction == DMA_TO_DEVICE) {
1386                         if (sc->use_sg) {
1387                                 struct scatterlist *sg = sc->request_buffer;
1388
1389                                 iscsi_buf_init_sg(&tcp_ctask->sendbuf, sg);
1390                                 tcp_ctask->sg = sg + 1;
1391                                 tcp_ctask->bad_sg = sg + sc->use_sg;
1392                         } else {
1393                                 iscsi_buf_init_iov(&tcp_ctask->sendbuf,
1394                                                    sc->request_buffer,
1395                                                    sc->request_bufflen);
1396                                 tcp_ctask->sg = NULL;
1397                                 tcp_ctask->bad_sg = NULL;
1398                         }
1399
1400                         debug_scsi("cmd [itt 0x%x total %d imm_data %d "
1401                                    "unsol count %d, unsol offset %d]\n",
1402                                    ctask->itt, sc->request_bufflen,
1403                                    ctask->imm_count, ctask->unsol_count,
1404                                    ctask->unsol_offset);
1405                 }
1406
1407                 iscsi_buf_init_iov(&tcp_ctask->headbuf, (char*)ctask->hdr,
1408                                   sizeof(struct iscsi_hdr));
1409
1410                 if (conn->hdrdgst_en)
1411                         iscsi_hdr_digest(conn, &tcp_ctask->headbuf,
1412                                          (u8*)tcp_ctask->hdrext);
1413                 tcp_ctask->xmstate &= ~XMSTATE_CMD_HDR_INIT;
1414                 tcp_ctask->xmstate |= XMSTATE_CMD_HDR_XMIT;
1415         }
1416
1417         if (tcp_ctask->xmstate & XMSTATE_CMD_HDR_XMIT) {
1418                 rc = iscsi_sendhdr(conn, &tcp_ctask->headbuf, ctask->imm_count);
1419                 if (rc)
1420                         return rc;
1421                 tcp_ctask->xmstate &= ~XMSTATE_CMD_HDR_XMIT;
1422
1423                 if (sc->sc_data_direction != DMA_TO_DEVICE)
1424                         return 0;
1425
1426                 if (ctask->imm_count) {
1427                         tcp_ctask->xmstate |= XMSTATE_IMM_DATA;
1428                         iscsi_set_padding(tcp_ctask, ctask->imm_count);
1429
1430                         if (ctask->conn->datadgst_en) {
1431                                 iscsi_data_digest_init(ctask->conn->dd_data,
1432                                                        tcp_ctask);
1433                                 tcp_ctask->immdigest = 0;
1434                         }
1435                 }
1436
1437                 if (ctask->unsol_count)
1438                         tcp_ctask->xmstate |=
1439                                         XMSTATE_UNS_HDR | XMSTATE_UNS_INIT;
1440         }
1441         return rc;
1442 }
1443
1444 static int
1445 iscsi_send_padding(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1446 {
1447         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1448         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1449         int sent = 0, rc;
1450
1451         if (tcp_ctask->xmstate & XMSTATE_W_PAD) {
1452                 iscsi_buf_init_iov(&tcp_ctask->sendbuf, (char*)&tcp_ctask->pad,
1453                                    tcp_ctask->pad_count);
1454                 if (conn->datadgst_en)
1455                         crypto_hash_update(&tcp_conn->tx_hash,
1456                                            &tcp_ctask->sendbuf.sg,
1457                                            tcp_ctask->sendbuf.sg.length);
1458         } else if (!(tcp_ctask->xmstate & XMSTATE_W_RESEND_PAD))
1459                 return 0;
1460
1461         tcp_ctask->xmstate &= ~XMSTATE_W_PAD;
1462         tcp_ctask->xmstate &= ~XMSTATE_W_RESEND_PAD;
1463         debug_scsi("sending %d pad bytes for itt 0x%x\n",
1464                    tcp_ctask->pad_count, ctask->itt);
1465         rc = iscsi_sendpage(conn, &tcp_ctask->sendbuf, &tcp_ctask->pad_count,
1466                            &sent);
1467         if (rc) {
1468                 debug_scsi("padding send failed %d\n", rc);
1469                 tcp_ctask->xmstate |= XMSTATE_W_RESEND_PAD;
1470         }
1471         return rc;
1472 }
1473
1474 static int
1475 iscsi_send_digest(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask,
1476                         struct iscsi_buf *buf, uint32_t *digest)
1477 {
1478         struct iscsi_tcp_cmd_task *tcp_ctask;
1479         struct iscsi_tcp_conn *tcp_conn;
1480         int rc, sent = 0;
1481
1482         if (!conn->datadgst_en)
1483                 return 0;
1484
1485         tcp_ctask = ctask->dd_data;
1486         tcp_conn = conn->dd_data;
1487
1488         if (!(tcp_ctask->xmstate & XMSTATE_W_RESEND_DATA_DIGEST)) {
1489                 crypto_hash_final(&tcp_conn->tx_hash, (u8*)digest);
1490                 iscsi_buf_init_iov(buf, (char*)digest, 4);
1491         }
1492         tcp_ctask->xmstate &= ~XMSTATE_W_RESEND_DATA_DIGEST;
1493
1494         rc = iscsi_sendpage(conn, buf, &tcp_ctask->digest_count, &sent);
1495         if (!rc)
1496                 debug_scsi("sent digest 0x%x for itt 0x%x\n", *digest,
1497                           ctask->itt);
1498         else {
1499                 debug_scsi("sending digest 0x%x failed for itt 0x%x!\n",
1500                           *digest, ctask->itt);
1501                 tcp_ctask->xmstate |= XMSTATE_W_RESEND_DATA_DIGEST;
1502         }
1503         return rc;
1504 }
1505
1506 static int
1507 iscsi_send_data(struct iscsi_cmd_task *ctask, struct iscsi_buf *sendbuf,
1508                 struct scatterlist **sg, int *sent, int *count,
1509                 struct iscsi_buf *digestbuf, uint32_t *digest)
1510 {
1511         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1512         struct iscsi_conn *conn = ctask->conn;
1513         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1514         int rc, buf_sent, offset;
1515
1516         while (*count) {
1517                 buf_sent = 0;
1518                 offset = sendbuf->sent;
1519
1520                 rc = iscsi_sendpage(conn, sendbuf, count, &buf_sent);
1521                 *sent = *sent + buf_sent;
1522                 if (buf_sent && conn->datadgst_en)
1523                         partial_sg_digest_update(&tcp_conn->tx_hash,
1524                                 &sendbuf->sg, sendbuf->sg.offset + offset,
1525                                 buf_sent);
1526                 if (!iscsi_buf_left(sendbuf) && *sg != tcp_ctask->bad_sg) {
1527                         iscsi_buf_init_sg(sendbuf, *sg);
1528                         *sg = *sg + 1;
1529                 }
1530
1531                 if (rc)
1532                         return rc;
1533         }
1534
1535         rc = iscsi_send_padding(conn, ctask);
1536         if (rc)
1537                 return rc;
1538
1539         return iscsi_send_digest(conn, ctask, digestbuf, digest);
1540 }
1541
1542 static int
1543 iscsi_send_unsol_hdr(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1544 {
1545         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1546         struct iscsi_data_task *dtask;
1547         int rc;
1548
1549         tcp_ctask->xmstate |= XMSTATE_UNS_DATA;
1550         if (tcp_ctask->xmstate & XMSTATE_UNS_INIT) {
1551                 dtask = &tcp_ctask->unsol_dtask;
1552
1553                 iscsi_prep_unsolicit_data_pdu(ctask, &dtask->hdr);
1554                 iscsi_buf_init_iov(&tcp_ctask->headbuf, (char*)&dtask->hdr,
1555                                    sizeof(struct iscsi_hdr));
1556                 if (conn->hdrdgst_en)
1557                         iscsi_hdr_digest(conn, &tcp_ctask->headbuf,
1558                                         (u8*)dtask->hdrext);
1559
1560                 tcp_ctask->xmstate &= ~XMSTATE_UNS_INIT;
1561                 iscsi_set_padding(tcp_ctask, ctask->data_count);
1562         }
1563
1564         rc = iscsi_sendhdr(conn, &tcp_ctask->headbuf, ctask->data_count);
1565         if (rc) {
1566                 tcp_ctask->xmstate &= ~XMSTATE_UNS_DATA;
1567                 tcp_ctask->xmstate |= XMSTATE_UNS_HDR;
1568                 return rc;
1569         }
1570
1571         if (conn->datadgst_en) {
1572                 dtask = &tcp_ctask->unsol_dtask;
1573                 iscsi_data_digest_init(ctask->conn->dd_data, tcp_ctask);
1574                 dtask->digest = 0;
1575         }
1576
1577         debug_scsi("uns dout [itt 0x%x dlen %d sent %d]\n",
1578                    ctask->itt, ctask->unsol_count, tcp_ctask->sent);
1579         return 0;
1580 }
1581
1582 static int
1583 iscsi_send_unsol_pdu(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1584 {
1585         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1586         int rc;
1587
1588         if (tcp_ctask->xmstate & XMSTATE_UNS_HDR) {
1589                 BUG_ON(!ctask->unsol_count);
1590                 tcp_ctask->xmstate &= ~XMSTATE_UNS_HDR;
1591 send_hdr:
1592                 rc = iscsi_send_unsol_hdr(conn, ctask);
1593                 if (rc)
1594                         return rc;
1595         }
1596
1597         if (tcp_ctask->xmstate & XMSTATE_UNS_DATA) {
1598                 struct iscsi_data_task *dtask = &tcp_ctask->unsol_dtask;
1599                 int start = tcp_ctask->sent;
1600
1601                 rc = iscsi_send_data(ctask, &tcp_ctask->sendbuf, &tcp_ctask->sg,
1602                                      &tcp_ctask->sent, &ctask->data_count,
1603                                      &dtask->digestbuf, &dtask->digest);
1604                 ctask->unsol_count -= tcp_ctask->sent - start;
1605                 if (rc)
1606                         return rc;
1607                 tcp_ctask->xmstate &= ~XMSTATE_UNS_DATA;
1608                 /*
1609                  * Done with the Data-Out. Next, check if we need
1610                  * to send another unsolicited Data-Out.
1611                  */
1612                 if (ctask->unsol_count) {
1613                         debug_scsi("sending more uns\n");
1614                         tcp_ctask->xmstate |= XMSTATE_UNS_INIT;
1615                         goto send_hdr;
1616                 }
1617         }
1618         return 0;
1619 }
1620
1621 static int iscsi_send_sol_pdu(struct iscsi_conn *conn,
1622                               struct iscsi_cmd_task *ctask)
1623 {
1624         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1625         struct iscsi_session *session = conn->session;
1626         struct iscsi_r2t_info *r2t;
1627         struct iscsi_data_task *dtask;
1628         int left, rc;
1629
1630         if (tcp_ctask->xmstate & XMSTATE_SOL_HDR_INIT) {
1631                 if (!tcp_ctask->r2t) {
1632                         spin_lock_bh(&session->lock);
1633                         __kfifo_get(tcp_ctask->r2tqueue, (void*)&tcp_ctask->r2t,
1634                                     sizeof(void*));
1635                         spin_unlock_bh(&session->lock);
1636                 }
1637 send_hdr:
1638                 r2t = tcp_ctask->r2t;
1639                 dtask = &r2t->dtask;
1640
1641                 if (conn->hdrdgst_en)
1642                         iscsi_hdr_digest(conn, &r2t->headbuf,
1643                                         (u8*)dtask->hdrext);
1644                 tcp_ctask->xmstate &= ~XMSTATE_SOL_HDR_INIT;
1645                 tcp_ctask->xmstate |= XMSTATE_SOL_HDR;
1646         }
1647
1648         if (tcp_ctask->xmstate & XMSTATE_SOL_HDR) {
1649                 r2t = tcp_ctask->r2t;
1650                 dtask = &r2t->dtask;
1651
1652                 rc = iscsi_sendhdr(conn, &r2t->headbuf, r2t->data_count);
1653                 if (rc)
1654                         return rc;
1655                 tcp_ctask->xmstate &= ~XMSTATE_SOL_HDR;
1656                 tcp_ctask->xmstate |= XMSTATE_SOL_DATA;
1657
1658                 if (conn->datadgst_en) {
1659                         iscsi_data_digest_init(conn->dd_data, tcp_ctask);
1660                         dtask->digest = 0;
1661                 }
1662
1663                 iscsi_set_padding(tcp_ctask, r2t->data_count);
1664                 debug_scsi("sol dout [dsn %d itt 0x%x dlen %d sent %d]\n",
1665                         r2t->solicit_datasn - 1, ctask->itt, r2t->data_count,
1666                         r2t->sent);
1667         }
1668
1669         if (tcp_ctask->xmstate & XMSTATE_SOL_DATA) {
1670                 r2t = tcp_ctask->r2t;
1671                 dtask = &r2t->dtask;
1672
1673                 rc = iscsi_send_data(ctask, &r2t->sendbuf, &r2t->sg,
1674                                      &r2t->sent, &r2t->data_count,
1675                                      &dtask->digestbuf, &dtask->digest);
1676                 if (rc)
1677                         return rc;
1678                 tcp_ctask->xmstate &= ~XMSTATE_SOL_DATA;
1679
1680                 /*
1681                  * Done with this Data-Out. Next, check if we have
1682                  * to send another Data-Out for this R2T.
1683                  */
1684                 BUG_ON(r2t->data_length - r2t->sent < 0);
1685                 left = r2t->data_length - r2t->sent;
1686                 if (left) {
1687                         iscsi_solicit_data_cont(conn, ctask, r2t, left);
1688                         goto send_hdr;
1689                 }
1690
1691                 /*
1692                  * Done with this R2T. Check if there are more
1693                  * outstanding R2Ts ready to be processed.
1694                  */
1695                 spin_lock_bh(&session->lock);
1696                 tcp_ctask->r2t = NULL;
1697                 __kfifo_put(tcp_ctask->r2tpool.queue, (void*)&r2t,
1698                             sizeof(void*));
1699                 if (__kfifo_get(tcp_ctask->r2tqueue, (void*)&r2t,
1700                                 sizeof(void*))) {
1701                         tcp_ctask->r2t = r2t;
1702                         spin_unlock_bh(&session->lock);
1703                         goto send_hdr;
1704                 }
1705                 spin_unlock_bh(&session->lock);
1706         }
1707         return 0;
1708 }
1709
1710 /**
1711  * iscsi_tcp_ctask_xmit - xmit normal PDU task
1712  * @conn: iscsi connection
1713  * @ctask: iscsi command task
1714  *
1715  * Notes:
1716  *      The function can return -EAGAIN in which case caller must
1717  *      call it again later, or recover. '0' return code means successful
1718  *      xmit.
1719  *      The function is devided to logical helpers (above) for the different
1720  *      xmit stages.
1721  *
1722  *iscsi_send_cmd_hdr()
1723  *      XMSTATE_CMD_HDR_INIT - prepare Header and Data buffers Calculate
1724  *                             Header Digest
1725  *      XMSTATE_CMD_HDR_XMIT - Transmit header in progress
1726  *
1727  *iscsi_send_padding
1728  *      XMSTATE_W_PAD        - Prepare and send pading
1729  *      XMSTATE_W_RESEND_PAD - retry send pading
1730  *
1731  *iscsi_send_digest
1732  *      XMSTATE_W_RESEND_DATA_DIGEST - Finalize and send Data Digest
1733  *      XMSTATE_W_RESEND_DATA_DIGEST - retry sending digest
1734  *
1735  *iscsi_send_unsol_hdr
1736  *      XMSTATE_UNS_INIT     - prepare un-solicit data header and digest
1737  *      XMSTATE_UNS_HDR      - send un-solicit header
1738  *
1739  *iscsi_send_unsol_pdu
1740  *      XMSTATE_UNS_DATA     - send un-solicit data in progress
1741  *
1742  *iscsi_send_sol_pdu
1743  *      XMSTATE_SOL_HDR_INIT - solicit data header and digest initialize
1744  *      XMSTATE_SOL_HDR      - send solicit header
1745  *      XMSTATE_SOL_DATA     - send solicit data
1746  *
1747  *iscsi_tcp_ctask_xmit
1748  *      XMSTATE_IMM_DATA     - xmit managment data (??)
1749  **/
1750 static int
1751 iscsi_tcp_ctask_xmit(struct iscsi_conn *conn, struct iscsi_cmd_task *ctask)
1752 {
1753         struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1754         int rc = 0;
1755
1756         debug_scsi("ctask deq [cid %d xmstate %x itt 0x%x]\n",
1757                 conn->id, tcp_ctask->xmstate, ctask->itt);
1758
1759         rc = iscsi_send_cmd_hdr(conn, ctask);
1760         if (rc)
1761                 return rc;
1762         if (ctask->sc->sc_data_direction != DMA_TO_DEVICE)
1763                 return 0;
1764
1765         if (tcp_ctask->xmstate & XMSTATE_IMM_DATA) {
1766                 rc = iscsi_send_data(ctask, &tcp_ctask->sendbuf, &tcp_ctask->sg,
1767                                      &tcp_ctask->sent, &ctask->imm_count,
1768                                      &tcp_ctask->immbuf, &tcp_ctask->immdigest);
1769                 if (rc)
1770                         return rc;
1771                 tcp_ctask->xmstate &= ~XMSTATE_IMM_DATA;
1772         }
1773
1774         rc = iscsi_send_unsol_pdu(conn, ctask);
1775         if (rc)
1776                 return rc;
1777
1778         rc = iscsi_send_sol_pdu(conn, ctask);
1779         if (rc)
1780                 return rc;
1781
1782         return rc;
1783 }
1784
1785 static struct iscsi_cls_conn *
1786 iscsi_tcp_conn_create(struct iscsi_cls_session *cls_session, uint32_t conn_idx)
1787 {
1788         struct iscsi_conn *conn;
1789         struct iscsi_cls_conn *cls_conn;
1790         struct iscsi_tcp_conn *tcp_conn;
1791
1792         cls_conn = iscsi_conn_setup(cls_session, conn_idx);
1793         if (!cls_conn)
1794                 return NULL;
1795         conn = cls_conn->dd_data;
1796         /*
1797          * due to strange issues with iser these are not set
1798          * in iscsi_conn_setup
1799          */
1800         conn->max_recv_dlength = ISCSI_DEF_MAX_RECV_SEG_LEN;
1801
1802         tcp_conn = kzalloc(sizeof(*tcp_conn), GFP_KERNEL);
1803         if (!tcp_conn)
1804                 goto tcp_conn_alloc_fail;
1805
1806         conn->dd_data = tcp_conn;
1807         tcp_conn->iscsi_conn = conn;
1808         tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
1809         /* initial operational parameters */
1810         tcp_conn->hdr_size = sizeof(struct iscsi_hdr);
1811
1812         tcp_conn->tx_hash.tfm = crypto_alloc_hash("crc32c", 0,
1813                                                   CRYPTO_ALG_ASYNC);
1814         tcp_conn->tx_hash.flags = 0;
1815         if (IS_ERR(tcp_conn->tx_hash.tfm)) {
1816                 printk(KERN_ERR "Could not create connection due to crc32c "
1817                        "loading error %ld. Make sure the crc32c module is "
1818                        "built as a module or into the kernel\n",
1819                         PTR_ERR(tcp_conn->tx_hash.tfm));
1820                 goto free_tcp_conn;
1821         }
1822
1823         tcp_conn->rx_hash.tfm = crypto_alloc_hash("crc32c", 0,
1824                                                   CRYPTO_ALG_ASYNC);
1825         tcp_conn->rx_hash.flags = 0;
1826         if (IS_ERR(tcp_conn->rx_hash.tfm)) {
1827                 printk(KERN_ERR "Could not create connection due to crc32c "
1828                        "loading error %ld. Make sure the crc32c module is "
1829                        "built as a module or into the kernel\n",
1830                         PTR_ERR(tcp_conn->rx_hash.tfm));
1831                 goto free_tx_tfm;
1832         }
1833
1834         return cls_conn;
1835
1836 free_tx_tfm:
1837         crypto_free_hash(tcp_conn->tx_hash.tfm);
1838 free_tcp_conn:
1839         kfree(tcp_conn);
1840 tcp_conn_alloc_fail:
1841         iscsi_conn_teardown(cls_conn);
1842         return NULL;
1843 }
1844
1845 static void
1846 iscsi_tcp_release_conn(struct iscsi_conn *conn)
1847 {
1848         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1849
1850         if (!tcp_conn->sock)
1851                 return;
1852
1853         sock_hold(tcp_conn->sock->sk);
1854         iscsi_conn_restore_callbacks(tcp_conn);
1855         sock_put(tcp_conn->sock->sk);
1856
1857         sock_release(tcp_conn->sock);
1858         tcp_conn->sock = NULL;
1859         conn->recv_lock = NULL;
1860 }
1861
1862 static void
1863 iscsi_tcp_conn_destroy(struct iscsi_cls_conn *cls_conn)
1864 {
1865         struct iscsi_conn *conn = cls_conn->dd_data;
1866         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1867
1868         iscsi_tcp_release_conn(conn);
1869         iscsi_conn_teardown(cls_conn);
1870
1871         if (tcp_conn->tx_hash.tfm)
1872                 crypto_free_hash(tcp_conn->tx_hash.tfm);
1873         if (tcp_conn->rx_hash.tfm)
1874                 crypto_free_hash(tcp_conn->rx_hash.tfm);
1875
1876         kfree(tcp_conn);
1877 }
1878
1879 static void
1880 iscsi_tcp_conn_stop(struct iscsi_cls_conn *cls_conn, int flag)
1881 {
1882         struct iscsi_conn *conn = cls_conn->dd_data;
1883         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1884
1885         iscsi_conn_stop(cls_conn, flag);
1886         iscsi_tcp_release_conn(conn);
1887         tcp_conn->hdr_size = sizeof(struct iscsi_hdr);
1888 }
1889
1890 static int
1891 iscsi_tcp_conn_bind(struct iscsi_cls_session *cls_session,
1892                     struct iscsi_cls_conn *cls_conn, uint64_t transport_eph,
1893                     int is_leading)
1894 {
1895         struct iscsi_conn *conn = cls_conn->dd_data;
1896         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
1897         struct sock *sk;
1898         struct socket *sock;
1899         int err;
1900
1901         /* lookup for existing socket */
1902         sock = sockfd_lookup((int)transport_eph, &err);
1903         if (!sock) {
1904                 printk(KERN_ERR "iscsi_tcp: sockfd_lookup failed %d\n", err);
1905                 return -EEXIST;
1906         }
1907
1908         err = iscsi_conn_bind(cls_session, cls_conn, is_leading);
1909         if (err)
1910                 return err;
1911
1912         /* bind iSCSI connection and socket */
1913         tcp_conn->sock = sock;
1914
1915         /* setup Socket parameters */
1916         sk = sock->sk;
1917         sk->sk_reuse = 1;
1918         sk->sk_sndtimeo = 15 * HZ; /* FIXME: make it configurable */
1919         sk->sk_allocation = GFP_ATOMIC;
1920
1921         /* FIXME: disable Nagle's algorithm */
1922
1923         /*
1924          * Intercept TCP callbacks for sendfile like receive
1925          * processing.
1926          */
1927         conn->recv_lock = &sk->sk_callback_lock;
1928         iscsi_conn_set_callbacks(conn);
1929         tcp_conn->sendpage = tcp_conn->sock->ops->sendpage;
1930         /*
1931          * set receive state machine into initial state
1932          */
1933         tcp_conn->in_progress = IN_PROGRESS_WAIT_HEADER;
1934
1935         return 0;
1936 }
1937
1938 /* called with host lock */
1939 static void
1940 iscsi_tcp_mgmt_init(struct iscsi_conn *conn, struct iscsi_mgmt_task *mtask)
1941 {
1942         struct iscsi_tcp_mgmt_task *tcp_mtask = mtask->dd_data;
1943         tcp_mtask->xmstate = XMSTATE_IMM_HDR_INIT;
1944 }
1945
1946 static int
1947 iscsi_r2tpool_alloc(struct iscsi_session *session)
1948 {
1949         int i;
1950         int cmd_i;
1951
1952         /*
1953          * initialize per-task: R2T pool and xmit queue
1954          */
1955         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
1956                 struct iscsi_cmd_task *ctask = session->cmds[cmd_i];
1957                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1958
1959                 /*
1960                  * pre-allocated x4 as much r2ts to handle race when
1961                  * target acks DataOut faster than we data_xmit() queues
1962                  * could replenish r2tqueue.
1963                  */
1964
1965                 /* R2T pool */
1966                 if (iscsi_pool_init(&tcp_ctask->r2tpool, session->max_r2t * 4,
1967                                     (void***)&tcp_ctask->r2ts,
1968                                     sizeof(struct iscsi_r2t_info))) {
1969                         goto r2t_alloc_fail;
1970                 }
1971
1972                 /* R2T xmit queue */
1973                 tcp_ctask->r2tqueue = kfifo_alloc(
1974                       session->max_r2t * 4 * sizeof(void*), GFP_KERNEL, NULL);
1975                 if (tcp_ctask->r2tqueue == ERR_PTR(-ENOMEM)) {
1976                         iscsi_pool_free(&tcp_ctask->r2tpool,
1977                                         (void**)tcp_ctask->r2ts);
1978                         goto r2t_alloc_fail;
1979                 }
1980         }
1981
1982         return 0;
1983
1984 r2t_alloc_fail:
1985         for (i = 0; i < cmd_i; i++) {
1986                 struct iscsi_cmd_task *ctask = session->cmds[i];
1987                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
1988
1989                 kfifo_free(tcp_ctask->r2tqueue);
1990                 iscsi_pool_free(&tcp_ctask->r2tpool,
1991                                 (void**)tcp_ctask->r2ts);
1992         }
1993         return -ENOMEM;
1994 }
1995
1996 static void
1997 iscsi_r2tpool_free(struct iscsi_session *session)
1998 {
1999         int i;
2000
2001         for (i = 0; i < session->cmds_max; i++) {
2002                 struct iscsi_cmd_task *ctask = session->cmds[i];
2003                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
2004
2005                 kfifo_free(tcp_ctask->r2tqueue);
2006                 iscsi_pool_free(&tcp_ctask->r2tpool,
2007                                 (void**)tcp_ctask->r2ts);
2008         }
2009 }
2010
2011 static int
2012 iscsi_conn_set_param(struct iscsi_cls_conn *cls_conn, enum iscsi_param param,
2013                      char *buf, int buflen)
2014 {
2015         struct iscsi_conn *conn = cls_conn->dd_data;
2016         struct iscsi_session *session = conn->session;
2017         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
2018         int value;
2019
2020         switch(param) {
2021         case ISCSI_PARAM_HDRDGST_EN:
2022                 iscsi_set_param(cls_conn, param, buf, buflen);
2023                 tcp_conn->hdr_size = sizeof(struct iscsi_hdr);
2024                 if (conn->hdrdgst_en)
2025                         tcp_conn->hdr_size += sizeof(__u32);
2026                 break;
2027         case ISCSI_PARAM_DATADGST_EN:
2028                 iscsi_set_param(cls_conn, param, buf, buflen);
2029                 tcp_conn->sendpage = conn->datadgst_en ?
2030                         sock_no_sendpage : tcp_conn->sock->ops->sendpage;
2031                 break;
2032         case ISCSI_PARAM_MAX_R2T:
2033                 sscanf(buf, "%d", &value);
2034                 if (session->max_r2t == roundup_pow_of_two(value))
2035                         break;
2036                 iscsi_r2tpool_free(session);
2037                 iscsi_set_param(cls_conn, param, buf, buflen);
2038                 if (session->max_r2t & (session->max_r2t - 1))
2039                         session->max_r2t = roundup_pow_of_two(session->max_r2t);
2040                 if (iscsi_r2tpool_alloc(session))
2041                         return -ENOMEM;
2042                 break;
2043         default:
2044                 return iscsi_set_param(cls_conn, param, buf, buflen);
2045         }
2046
2047         return 0;
2048 }
2049
2050 static int
2051 iscsi_tcp_conn_get_param(struct iscsi_cls_conn *cls_conn,
2052                          enum iscsi_param param, char *buf)
2053 {
2054         struct iscsi_conn *conn = cls_conn->dd_data;
2055         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
2056         struct inet_sock *inet;
2057         struct ipv6_pinfo *np;
2058         struct sock *sk;
2059         int len;
2060
2061         switch(param) {
2062         case ISCSI_PARAM_CONN_PORT:
2063                 if (!tcp_conn->sock)
2064                         return -EINVAL;
2065
2066                 inet = inet_sk(tcp_conn->sock->sk);
2067                 len = sprintf(buf, "%hu\n", be16_to_cpu(inet->dport));
2068                 break;
2069         case ISCSI_PARAM_CONN_ADDRESS:
2070                 if (!tcp_conn->sock)
2071                         return -EINVAL;
2072
2073                 sk = tcp_conn->sock->sk;
2074                 if (sk->sk_family == PF_INET) {
2075                         inet = inet_sk(sk);
2076                         len = sprintf(buf, NIPQUAD_FMT "\n",
2077                                       NIPQUAD(inet->daddr));
2078                 } else {
2079                         np = inet6_sk(sk);
2080                         len = sprintf(buf, NIP6_FMT "\n", NIP6(np->daddr));
2081                 }
2082                 break;
2083         default:
2084                 return iscsi_conn_get_param(cls_conn, param, buf);
2085         }
2086
2087         return len;
2088 }
2089
2090 static void
2091 iscsi_conn_get_stats(struct iscsi_cls_conn *cls_conn, struct iscsi_stats *stats)
2092 {
2093         struct iscsi_conn *conn = cls_conn->dd_data;
2094         struct iscsi_tcp_conn *tcp_conn = conn->dd_data;
2095
2096         stats->txdata_octets = conn->txdata_octets;
2097         stats->rxdata_octets = conn->rxdata_octets;
2098         stats->scsicmd_pdus = conn->scsicmd_pdus_cnt;
2099         stats->dataout_pdus = conn->dataout_pdus_cnt;
2100         stats->scsirsp_pdus = conn->scsirsp_pdus_cnt;
2101         stats->datain_pdus = conn->datain_pdus_cnt;
2102         stats->r2t_pdus = conn->r2t_pdus_cnt;
2103         stats->tmfcmd_pdus = conn->tmfcmd_pdus_cnt;
2104         stats->tmfrsp_pdus = conn->tmfrsp_pdus_cnt;
2105         stats->custom_length = 3;
2106         strcpy(stats->custom[0].desc, "tx_sendpage_failures");
2107         stats->custom[0].value = tcp_conn->sendpage_failures_cnt;
2108         strcpy(stats->custom[1].desc, "rx_discontiguous_hdr");
2109         stats->custom[1].value = tcp_conn->discontiguous_hdr_cnt;
2110         strcpy(stats->custom[2].desc, "eh_abort_cnt");
2111         stats->custom[2].value = conn->eh_abort_cnt;
2112 }
2113
2114 static struct iscsi_cls_session *
2115 iscsi_tcp_session_create(struct iscsi_transport *iscsit,
2116                          struct scsi_transport_template *scsit,
2117                          uint16_t cmds_max, uint16_t qdepth,
2118                          uint32_t initial_cmdsn, uint32_t *hostno)
2119 {
2120         struct iscsi_cls_session *cls_session;
2121         struct iscsi_session *session;
2122         uint32_t hn;
2123         int cmd_i;
2124
2125         cls_session = iscsi_session_setup(iscsit, scsit, cmds_max, qdepth,
2126                                          sizeof(struct iscsi_tcp_cmd_task),
2127                                          sizeof(struct iscsi_tcp_mgmt_task),
2128                                          initial_cmdsn, &hn);
2129         if (!cls_session)
2130                 return NULL;
2131         *hostno = hn;
2132
2133         session = class_to_transport_session(cls_session);
2134         for (cmd_i = 0; cmd_i < session->cmds_max; cmd_i++) {
2135                 struct iscsi_cmd_task *ctask = session->cmds[cmd_i];
2136                 struct iscsi_tcp_cmd_task *tcp_ctask = ctask->dd_data;
2137
2138                 ctask->hdr = &tcp_ctask->hdr;
2139         }
2140
2141         for (cmd_i = 0; cmd_i < session->mgmtpool_max; cmd_i++) {
2142                 struct iscsi_mgmt_task *mtask = session->mgmt_cmds[cmd_i];
2143                 struct iscsi_tcp_mgmt_task *tcp_mtask = mtask->dd_data;
2144
2145                 mtask->hdr = &tcp_mtask->hdr;
2146         }
2147
2148         if (iscsi_r2tpool_alloc(class_to_transport_session(cls_session)))
2149                 goto r2tpool_alloc_fail;
2150
2151         return cls_session;
2152
2153 r2tpool_alloc_fail:
2154         iscsi_session_teardown(cls_session);
2155         return NULL;
2156 }
2157
2158 static void iscsi_tcp_session_destroy(struct iscsi_cls_session *cls_session)
2159 {
2160         iscsi_r2tpool_free(class_to_transport_session(cls_session));
2161         iscsi_session_teardown(cls_session);
2162 }
2163
2164 static struct scsi_host_template iscsi_sht = {
2165         .name                   = "iSCSI Initiator over TCP/IP",
2166         .queuecommand           = iscsi_queuecommand,
2167         .change_queue_depth     = iscsi_change_queue_depth,
2168         .can_queue              = ISCSI_DEF_XMIT_CMDS_MAX - 1,
2169         .sg_tablesize           = ISCSI_SG_TABLESIZE,
2170         .max_sectors            = 0xFFFF,
2171         .cmd_per_lun            = ISCSI_DEF_CMD_PER_LUN,
2172         .eh_abort_handler       = iscsi_eh_abort,
2173         .eh_host_reset_handler  = iscsi_eh_host_reset,
2174         .use_clustering         = DISABLE_CLUSTERING,
2175         .proc_name              = "iscsi_tcp",
2176         .this_id                = -1,
2177 };
2178
2179 static struct iscsi_transport iscsi_tcp_transport = {
2180         .owner                  = THIS_MODULE,
2181         .name                   = "tcp",
2182         .caps                   = CAP_RECOVERY_L0 | CAP_MULTI_R2T | CAP_HDRDGST
2183                                   | CAP_DATADGST,
2184         .param_mask             = ISCSI_MAX_RECV_DLENGTH |
2185                                   ISCSI_MAX_XMIT_DLENGTH |
2186                                   ISCSI_HDRDGST_EN |
2187                                   ISCSI_DATADGST_EN |
2188                                   ISCSI_INITIAL_R2T_EN |
2189                                   ISCSI_MAX_R2T |
2190                                   ISCSI_IMM_DATA_EN |
2191                                   ISCSI_FIRST_BURST |
2192                                   ISCSI_MAX_BURST |
2193                                   ISCSI_PDU_INORDER_EN |
2194                                   ISCSI_DATASEQ_INORDER_EN |
2195                                   ISCSI_ERL |
2196                                   ISCSI_CONN_PORT |
2197                                   ISCSI_CONN_ADDRESS |
2198                                   ISCSI_EXP_STATSN |
2199                                   ISCSI_PERSISTENT_PORT |
2200                                   ISCSI_PERSISTENT_ADDRESS |
2201                                   ISCSI_TARGET_NAME | ISCSI_TPGT |
2202                                   ISCSI_USERNAME | ISCSI_PASSWORD |
2203                                   ISCSI_USERNAME_IN | ISCSI_PASSWORD_IN,
2204         .host_param_mask        = ISCSI_HOST_HWADDRESS |
2205                                   ISCSI_HOST_INITIATOR_NAME,
2206         .host_template          = &iscsi_sht,
2207         .conndata_size          = sizeof(struct iscsi_conn),
2208         .max_conn               = 1,
2209         .max_cmd_len            = ISCSI_TCP_MAX_CMD_LEN,
2210         /* session management */
2211         .create_session         = iscsi_tcp_session_create,
2212         .destroy_session        = iscsi_tcp_session_destroy,
2213         /* connection management */
2214         .create_conn            = iscsi_tcp_conn_create,
2215         .bind_conn              = iscsi_tcp_conn_bind,
2216         .destroy_conn           = iscsi_tcp_conn_destroy,
2217         .set_param              = iscsi_conn_set_param,
2218         .get_conn_param         = iscsi_tcp_conn_get_param,
2219         .get_session_param      = iscsi_session_get_param,
2220         .start_conn             = iscsi_conn_start,
2221         .stop_conn              = iscsi_tcp_conn_stop,
2222         /* iscsi host params */
2223         .get_host_param         = iscsi_host_get_param,
2224         .set_host_param         = iscsi_host_set_param,
2225         /* IO */
2226         .send_pdu               = iscsi_conn_send_pdu,
2227         .get_stats              = iscsi_conn_get_stats,
2228         .init_cmd_task          = iscsi_tcp_cmd_init,
2229         .init_mgmt_task         = iscsi_tcp_mgmt_init,
2230         .xmit_cmd_task          = iscsi_tcp_ctask_xmit,
2231         .xmit_mgmt_task         = iscsi_tcp_mtask_xmit,
2232         .cleanup_cmd_task       = iscsi_tcp_cleanup_ctask,
2233         /* recovery */
2234         .session_recovery_timedout = iscsi_session_recovery_timedout,
2235 };
2236
2237 static int __init
2238 iscsi_tcp_init(void)
2239 {
2240         if (iscsi_max_lun < 1) {
2241                 printk(KERN_ERR "iscsi_tcp: Invalid max_lun value of %u\n",
2242                        iscsi_max_lun);
2243                 return -EINVAL;
2244         }
2245         iscsi_tcp_transport.max_lun = iscsi_max_lun;
2246
2247         if (!iscsi_register_transport(&iscsi_tcp_transport))
2248                 return -ENODEV;
2249
2250         return 0;
2251 }
2252
2253 static void __exit
2254 iscsi_tcp_exit(void)
2255 {
2256         iscsi_unregister_transport(&iscsi_tcp_transport);
2257 }
2258
2259 module_init(iscsi_tcp_init);
2260 module_exit(iscsi_tcp_exit);