sys/dev/iscsi/iscsi.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2012 The FreeBSD Foundation
   5  *
   6  * This software was developed by Edward Tomasz Napierala under sponsorship
   7  * from the FreeBSD Foundation.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  28  * SUCH DAMAGE.
  29  *
  30  */
  31
  32 #include <sys/cdefs.h>
  33 __FBSDID("$FreeBSD$");
  34
  35 #include <sys/param.h>
  36 #include <sys/condvar.h>
  37 #include <sys/conf.h>
  38 #include <sys/endian.h>
  39 #include <sys/eventhandler.h>
  40 #include <sys/file.h>
  41 #include <sys/kernel.h>
  42 #include <sys/kthread.h>
  43 #include <sys/lock.h>
  44 #include <sys/malloc.h>
  45 #include <sys/mutex.h>
  46 #include <sys/module.h>
  47 #include <sys/socket.h>
  48 #include <sys/sysctl.h>
  49 #include <sys/systm.h>
  50 #include <sys/sx.h>
  51 #include <vm/uma.h>
  52
  53 #include <cam/cam.h>
  54 #include <cam/cam_ccb.h>
  55 #include <cam/cam_xpt.h>
  56 #include <cam/cam_debug.h>
  57 #include <cam/cam_sim.h>
  58 #include <cam/cam_xpt_sim.h>
  59 #include <cam/cam_xpt_periph.h>
  60 #include <cam/cam_periph.h>
  61 #include <cam/scsi/scsi_all.h>
  62 #include <cam/scsi/scsi_message.h>
  63
  64 #include <dev/iscsi/icl.h>
  65 #include <dev/iscsi/icl_wrappers.h>
  66 #include <dev/iscsi/iscsi_ioctl.h>
  67 #include <dev/iscsi/iscsi_proto.h>
  68 #include <dev/iscsi/iscsi.h>
  69
  70 #ifdef ICL_KERNEL_PROXY
  71 #include <sys/socketvar.h>
  72 #endif
  73
  74 #ifdef ICL_KERNEL_PROXY
  75 FEATURE(iscsi_kernel_proxy, "iSCSI initiator built with ICL_KERNEL_PROXY");
  76 #endif
  77
  78 /*
  79  * XXX: This is global so the iscsi_unload() can access it.
  80  *      Think about how to do this properly.
  81  */
  82 static struct iscsi_softc       *sc;
  83
  84 SYSCTL_NODE(_kern, OID_AUTO, iscsi, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
  85     "iSCSI initiator");
  86 static int debug = 1;
  87 SYSCTL_INT(_kern_iscsi, OID_AUTO, debug, CTLFLAG_RWTUN,
  88     &debug, 0, "Enable debug messages");
  89 static int ping_timeout = 5;
  90 SYSCTL_INT(_kern_iscsi, OID_AUTO, ping_timeout, CTLFLAG_RWTUN, &ping_timeout,
  91     0, "Timeout for ping (NOP-Out) requests, in seconds");
  92 static int iscsid_timeout = 60;
  93 SYSCTL_INT(_kern_iscsi, OID_AUTO, iscsid_timeout, CTLFLAG_RWTUN, &iscsid_timeout,
  94     0, "Time to wait for iscsid(8) to handle reconnection, in seconds");
  95 static int login_timeout = 60;
  96 SYSCTL_INT(_kern_iscsi, OID_AUTO, login_timeout, CTLFLAG_RWTUN, &login_timeout,
  97     0, "Time to wait for iscsid(8) to finish Login Phase, in seconds");
  98 static int maxtags = 255;
  99 SYSCTL_INT(_kern_iscsi, OID_AUTO, maxtags, CTLFLAG_RWTUN, &maxtags,
 100     0, "Max number of IO requests queued");
 101 static int fail_on_disconnection = 0;
 102 SYSCTL_INT(_kern_iscsi, OID_AUTO, fail_on_disconnection, CTLFLAG_RWTUN,
 103     &fail_on_disconnection, 0, "Destroy CAM SIM on connection failure");
 104 static int fail_on_shutdown = 1;
 105 SYSCTL_INT(_kern_iscsi, OID_AUTO, fail_on_shutdown, CTLFLAG_RWTUN,
 106     &fail_on_shutdown, 0, "Fail disconnected sessions on shutdown");
 107
 108 static MALLOC_DEFINE(M_ISCSI, "iSCSI", "iSCSI initiator");
 109 static uma_zone_t iscsi_outstanding_zone;
 110
 111 #define CONN_SESSION(X) ((struct iscsi_session *)X->ic_prv0)
 112 #define PDU_SESSION(X)  (CONN_SESSION(X->ip_conn))
 113
 114 #define ISCSI_DEBUG(X, ...)                                             \
 115         do {                                                            \
 116                 if (debug > 1)                                          \
 117                         printf("%s: " X "\n", __func__, ## __VA_ARGS__);\
 118         } while (0)
 119
 120 #define ISCSI_WARN(X, ...)                                              \
 121         do {                                                            \
 122                 if (debug > 0) {                                        \
 123                         printf("WARNING: %s: " X "\n",                  \
 124                             __func__, ## __VA_ARGS__);                  \
 125                 }                                                       \
 126         } while (0)
 127
 128 #define ISCSI_SESSION_DEBUG(S, X, ...)                                  \
 129         do {                                                            \
 130                 if (debug > 1) {                                        \
 131                         printf("%s: %s (%s): " X "\n",                  \
 132                             __func__, S->is_conf.isc_target_addr,       \
 133                             S->is_conf.isc_target, ## __VA_ARGS__);     \
 134                 }                                                       \
 135         } while (0)
 136
 137 #define ISCSI_SESSION_WARN(S, X, ...)                                   \
 138         do {                                                            \
 139                 if (debug > 0) {                                        \
 140                         printf("WARNING: %s (%s): " X "\n",             \
 141                             S->is_conf.isc_target_addr,                 \
 142                             S->is_conf.isc_target, ## __VA_ARGS__);     \
 143                 }                                                       \
 144         } while (0)
 145
 146 #define ISCSI_SESSION_LOCK(X)           mtx_lock(&X->is_lock)
 147 #define ISCSI_SESSION_UNLOCK(X)         mtx_unlock(&X->is_lock)
 148 #define ISCSI_SESSION_LOCK_ASSERT(X)    mtx_assert(&X->is_lock, MA_OWNED)
 149 #define ISCSI_SESSION_LOCK_ASSERT_NOT(X) mtx_assert(&X->is_lock, MA_NOTOWNED)
 150
 151 static int      iscsi_ioctl(struct cdev *dev, u_long cmd, caddr_t arg,
 152                     int mode, struct thread *td);
 153
 154 static struct cdevsw iscsi_cdevsw = {
 155      .d_version = D_VERSION,
 156      .d_ioctl   = iscsi_ioctl,
 157      .d_name    = "iscsi",
 158 };
 159
 160 static void     iscsi_pdu_queue_locked(struct icl_pdu *request);
 161 static void     iscsi_pdu_queue(struct icl_pdu *request);
 162 static void     iscsi_pdu_update_statsn(const struct icl_pdu *response);
 163 static void     iscsi_pdu_handle_nop_in(struct icl_pdu *response);
 164 static void     iscsi_pdu_handle_scsi_response(struct icl_pdu *response);
 165 static void     iscsi_pdu_handle_task_response(struct icl_pdu *response);
 166 static void     iscsi_pdu_handle_data_in(struct icl_pdu *response);
 167 static void     iscsi_pdu_handle_logout_response(struct icl_pdu *response);
 168 static void     iscsi_pdu_handle_r2t(struct icl_pdu *response);
 169 static void     iscsi_pdu_handle_async_message(struct icl_pdu *response);
 170 static void     iscsi_pdu_handle_reject(struct icl_pdu *response);
 171 static void     iscsi_session_reconnect(struct iscsi_session *is);
 172 static void     iscsi_session_terminate(struct iscsi_session *is);
 173 static void     iscsi_action(struct cam_sim *sim, union ccb *ccb);
 174 static struct iscsi_outstanding *iscsi_outstanding_find(struct iscsi_session *is,
 175                     uint32_t initiator_task_tag);
 176 static struct iscsi_outstanding *iscsi_outstanding_add(struct iscsi_session *is,
 177                     struct icl_pdu *request, union ccb *ccb,
 178                     uint32_t *initiator_task_tagp);
 179 static void     iscsi_outstanding_remove(struct iscsi_session *is,
 180                     struct iscsi_outstanding *io);
 181
 182 static bool
 183 iscsi_pdu_prepare(struct icl_pdu *request)
 184 {
 185         struct iscsi_session *is;
 186         struct iscsi_bhs_scsi_command *bhssc;
 187
 188         is = PDU_SESSION(request);
 189
 190         ISCSI_SESSION_LOCK_ASSERT(is);
 191
 192         /*
 193          * We're only using fields common for all the request
 194          * (initiator -> target) PDUs.
 195          */
 196         bhssc = (struct iscsi_bhs_scsi_command *)request->ip_bhs;
 197
 198         /*
 199          * Data-Out PDU does not contain CmdSN.
 200          */
 201         if (bhssc->bhssc_opcode != ISCSI_BHS_OPCODE_SCSI_DATA_OUT) {
 202                 if (ISCSI_SNGT(is->is_cmdsn, is->is_maxcmdsn) &&
 203                     (bhssc->bhssc_opcode & ISCSI_BHS_OPCODE_IMMEDIATE) == 0) {
 204                         /*
 205                          * Current MaxCmdSN prevents us from sending any more
 206                          * SCSI Command PDUs to the target; postpone the PDU.
 207                          * It will get resent by either iscsi_pdu_queue(),
 208                          * or by maintenance thread.
 209                          */
 210 #if 0
 211                         ISCSI_SESSION_DEBUG(is, "postponing send, CmdSN %u, "
 212                             "ExpCmdSN %u, MaxCmdSN %u, opcode 0x%x",
 213                             is->is_cmdsn, is->is_expcmdsn, is->is_maxcmdsn,
 214                             bhssc->bhssc_opcode);
 215 #endif
 216                         return (true);
 217                 }
 218                 bhssc->bhssc_cmdsn = htonl(is->is_cmdsn);
 219                 if ((bhssc->bhssc_opcode & ISCSI_BHS_OPCODE_IMMEDIATE) == 0)
 220                         is->is_cmdsn++;
 221         }
 222         bhssc->bhssc_expstatsn = htonl(is->is_statsn + 1);
 223
 224         return (false);
 225 }
 226
 227 static void
 228 iscsi_session_send_postponed(struct iscsi_session *is)
 229 {
 230         struct icl_pdu *request;
 231         bool postpone;
 232
 233         ISCSI_SESSION_LOCK_ASSERT(is);
 234
 235         if (STAILQ_EMPTY(&is->is_postponed))
 236                 return;
 237         while ((request = STAILQ_FIRST(&is->is_postponed)) != NULL) {
 238                 postpone = iscsi_pdu_prepare(request);
 239                 if (postpone)
 240                         return;
 241                 STAILQ_REMOVE_HEAD(&is->is_postponed, ip_next);
 242                 icl_pdu_queue(request);
 243         }
 244         xpt_release_simq(is->is_sim, 1);
 245 }
 246
 247 static void
 248 iscsi_pdu_queue_locked(struct icl_pdu *request)
 249 {
 250         struct iscsi_session *is;
 251         bool postpone;
 252
 253         is = PDU_SESSION(request);
 254         ISCSI_SESSION_LOCK_ASSERT(is);
 255         iscsi_session_send_postponed(is);
 256         postpone = iscsi_pdu_prepare(request);
 257         if (postpone) {
 258                 if (STAILQ_EMPTY(&is->is_postponed))
 259                         xpt_freeze_simq(is->is_sim, 1);
 260                 STAILQ_INSERT_TAIL(&is->is_postponed, request, ip_next);
 261                 return;
 262         }
 263         icl_pdu_queue(request);
 264 }
 265
 266 static void
 267 iscsi_pdu_queue(struct icl_pdu *request)
 268 {
 269         struct iscsi_session *is;
 270
 271         is = PDU_SESSION(request);
 272         ISCSI_SESSION_LOCK(is);
 273         iscsi_pdu_queue_locked(request);
 274         ISCSI_SESSION_UNLOCK(is);
 275 }
 276
 277 static void
 278 iscsi_session_logout(struct iscsi_session *is)
 279 {
 280         struct icl_pdu *request;
 281         struct iscsi_bhs_logout_request *bhslr;
 282
 283         request = icl_pdu_new(is->is_conn, M_NOWAIT);
 284         if (request == NULL)
 285                 return;
 286
 287         bhslr = (struct iscsi_bhs_logout_request *)request->ip_bhs;
 288         bhslr->bhslr_opcode = ISCSI_BHS_OPCODE_LOGOUT_REQUEST;
 289         bhslr->bhslr_reason = BHSLR_REASON_CLOSE_SESSION;
 290         iscsi_pdu_queue_locked(request);
 291 }
 292
 293 static void
 294 iscsi_session_terminate_task(struct iscsi_session *is,
 295     struct iscsi_outstanding *io, cam_status status)
 296 {
 297
 298         ISCSI_SESSION_LOCK_ASSERT(is);
 299
 300         if (io->io_ccb != NULL) {
 301                 io->io_ccb->ccb_h.status &= ~(CAM_SIM_QUEUED | CAM_STATUS_MASK);
 302                 io->io_ccb->ccb_h.status |= status;
 303                 if ((io->io_ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
 304                         io->io_ccb->ccb_h.status |= CAM_DEV_QFRZN;
 305                         xpt_freeze_devq(io->io_ccb->ccb_h.path, 1);
 306                         ISCSI_SESSION_DEBUG(is, "freezing devq");
 307                 }
 308                 xpt_done(io->io_ccb);
 309         }
 310         iscsi_outstanding_remove(is, io);
 311 }
 312
 313 static void
 314 iscsi_session_terminate_tasks(struct iscsi_session *is, cam_status status)
 315 {
 316         struct iscsi_outstanding *io, *tmp;
 317
 318         ISCSI_SESSION_LOCK_ASSERT(is);
 319
 320         TAILQ_FOREACH_SAFE(io, &is->is_outstanding, io_next, tmp) {
 321                 iscsi_session_terminate_task(is, io, status);
 322         }
 323 }
 324
 325 static void
 326 iscsi_session_cleanup(struct iscsi_session *is, bool destroy_sim)
 327 {
 328         struct icl_pdu *pdu;
 329
 330         ISCSI_SESSION_LOCK_ASSERT(is);
 331
 332         /*
 333          * Don't queue any new PDUs.
 334          */
 335         if (is->is_sim != NULL && is->is_simq_frozen == false) {
 336                 ISCSI_SESSION_DEBUG(is, "freezing");
 337                 xpt_freeze_simq(is->is_sim, 1);
 338                 is->is_simq_frozen = true;
 339         }
 340
 341         /*
 342          * Remove postponed PDUs.
 343          */
 344         if (!STAILQ_EMPTY(&is->is_postponed))
 345                 xpt_release_simq(is->is_sim, 1);
 346         while ((pdu = STAILQ_FIRST(&is->is_postponed)) != NULL) {
 347                 STAILQ_REMOVE_HEAD(&is->is_postponed, ip_next);
 348                 icl_pdu_free(pdu);
 349         }
 350
 351         if (destroy_sim == false) {
 352                 /*
 353                  * Terminate SCSI tasks, asking CAM to requeue them.
 354                  */
 355                 iscsi_session_terminate_tasks(is, CAM_REQUEUE_REQ);
 356                 return;
 357         }
 358
 359         iscsi_session_terminate_tasks(is, CAM_DEV_NOT_THERE);
 360
 361         if (is->is_sim == NULL)
 362                 return;
 363
 364         ISCSI_SESSION_DEBUG(is, "deregistering SIM");
 365         xpt_async(AC_LOST_DEVICE, is->is_path, NULL);
 366
 367         if (is->is_simq_frozen) {
 368                 is->is_simq_frozen = false;
 369                 xpt_release_simq(is->is_sim, 1);
 370         }
 371
 372         xpt_free_path(is->is_path);
 373         is->is_path = NULL;
 374         xpt_bus_deregister(cam_sim_path(is->is_sim));
 375         cam_sim_free(is->is_sim, TRUE /*free_devq*/);
 376         is->is_sim = NULL;
 377         is->is_devq = NULL;
 378 }
 379
 380 static void
 381 iscsi_maintenance_thread_reconnect(struct iscsi_session *is)
 382 {
 383
 384         icl_conn_close(is->is_conn);
 385
 386         ISCSI_SESSION_LOCK(is);
 387
 388         is->is_connected = false;
 389         is->is_reconnecting = false;
 390         is->is_login_phase = false;
 391
 392 #ifdef ICL_KERNEL_PROXY
 393         if (is->is_login_pdu != NULL) {
 394                 icl_pdu_free(is->is_login_pdu);
 395                 is->is_login_pdu = NULL;
 396         }
 397         cv_signal(&is->is_login_cv);
 398 #endif
 399
 400         if (fail_on_disconnection) {
 401                 ISCSI_SESSION_DEBUG(is, "connection failed, destroying devices");
 402                 iscsi_session_cleanup(is, true);
 403         } else {
 404                 iscsi_session_cleanup(is, false);
 405         }
 406
 407         KASSERT(TAILQ_EMPTY(&is->is_outstanding),
 408             ("destroying session with active tasks"));
 409         KASSERT(STAILQ_EMPTY(&is->is_postponed),
 410             ("destroying session with postponed PDUs"));
 411
 412         if (is->is_conf.isc_enable == 0 && is->is_conf.isc_discovery == 0) {
 413                 ISCSI_SESSION_UNLOCK(is);
 414                 return;
 415         }
 416
 417         /*
 418          * Request immediate reconnection from iscsid(8).
 419          */
 420         //ISCSI_SESSION_DEBUG(is, "waking up iscsid(8)");
 421         is->is_waiting_for_iscsid = true;
 422         strlcpy(is->is_reason, "Waiting for iscsid(8)", sizeof(is->is_reason));
 423         is->is_timeout = 0;
 424         ISCSI_SESSION_UNLOCK(is);
 425         cv_signal(&is->is_softc->sc_cv);
 426 }
 427
 428 static void
 429 iscsi_maintenance_thread_terminate(struct iscsi_session *is)
 430 {
 431         struct iscsi_softc *sc;
 432
 433         sc = is->is_softc;
 434         sx_xlock(&sc->sc_lock);
 435         TAILQ_REMOVE(&sc->sc_sessions, is, is_next);
 436         sx_xunlock(&sc->sc_lock);
 437
 438         icl_conn_close(is->is_conn);
 439         callout_drain(&is->is_callout);
 440
 441         ISCSI_SESSION_LOCK(is);
 442
 443         KASSERT(is->is_terminating, ("is_terminating == false"));
 444
 445 #ifdef ICL_KERNEL_PROXY
 446         if (is->is_login_pdu != NULL) {
 447                 icl_pdu_free(is->is_login_pdu);
 448                 is->is_login_pdu = NULL;
 449         }
 450         cv_signal(&is->is_login_cv);
 451 #endif
 452
 453         iscsi_session_cleanup(is, true);
 454
 455         KASSERT(TAILQ_EMPTY(&is->is_outstanding),
 456             ("destroying session with active tasks"));
 457         KASSERT(STAILQ_EMPTY(&is->is_postponed),
 458             ("destroying session with postponed PDUs"));
 459
 460         ISCSI_SESSION_UNLOCK(is);
 461
 462         icl_conn_free(is->is_conn);
 463         mtx_destroy(&is->is_lock);
 464         cv_destroy(&is->is_maintenance_cv);
 465 #ifdef ICL_KERNEL_PROXY
 466         cv_destroy(&is->is_login_cv);
 467 #endif
 468
 469         ISCSI_SESSION_DEBUG(is, "terminated");
 470         free(is, M_ISCSI);
 471
 472         /*
 473          * The iscsi_unload() routine might be waiting.
 474          */
 475         cv_signal(&sc->sc_cv);
 476 }
 477
 478 static void
 479 iscsi_maintenance_thread(void *arg)
 480 {
 481         struct iscsi_session *is = arg;
 482
 483         ISCSI_SESSION_LOCK(is);
 484         for (;;) {
 485                 if (is->is_reconnecting == false &&
 486                     is->is_terminating == false &&
 487                     (STAILQ_EMPTY(&is->is_postponed) ||
 488                      ISCSI_SNGT(is->is_cmdsn, is->is_maxcmdsn)))
 489                         cv_wait(&is->is_maintenance_cv, &is->is_lock);
 490
 491                 /* Terminate supersedes reconnect. */
 492                 if (is->is_terminating) {
 493                         ISCSI_SESSION_UNLOCK(is);
 494                         iscsi_maintenance_thread_terminate(is);
 495                         kthread_exit();
 496                         return;
 497                 }
 498
 499                 if (is->is_reconnecting) {
 500                         ISCSI_SESSION_UNLOCK(is);
 501                         iscsi_maintenance_thread_reconnect(is);
 502                         ISCSI_SESSION_LOCK(is);
 503                         continue;
 504                 }
 505
 506                 iscsi_session_send_postponed(is);
 507         }
 508         ISCSI_SESSION_UNLOCK(is);
 509 }
 510
 511 static void
 512 iscsi_session_reconnect(struct iscsi_session *is)
 513 {
 514
 515         /*
 516          * XXX: We can't use locking here, because
 517          *      it's being called from various contexts.
 518          *      Hope it doesn't break anything.
 519          */
 520         if (is->is_reconnecting)
 521                 return;
 522
 523         is->is_reconnecting = true;
 524         cv_signal(&is->is_maintenance_cv);
 525 }
 526
 527 static void
 528 iscsi_session_terminate(struct iscsi_session *is)
 529 {
 530
 531         if (is->is_terminating)
 532                 return;
 533
 534         is->is_terminating = true;
 535
 536 #if 0
 537         iscsi_session_logout(is);
 538 #endif
 539         cv_signal(&is->is_maintenance_cv);
 540 }
 541
 542 static void
 543 iscsi_callout(void *context)
 544 {
 545         struct icl_pdu *request;
 546         struct iscsi_bhs_nop_out *bhsno;
 547         struct iscsi_session *is;
 548         bool reconnect_needed = false;
 549
 550         is = context;
 551
 552         ISCSI_SESSION_LOCK(is);
 553         if (is->is_terminating) {
 554                 ISCSI_SESSION_UNLOCK(is);
 555                 return;
 556         }
 557
 558         callout_schedule(&is->is_callout, 1 * hz);
 559
 560         if (is->is_conf.isc_enable == 0)
 561                 goto out;
 562
 563         is->is_timeout++;
 564
 565         if (is->is_waiting_for_iscsid) {
 566                 if (iscsid_timeout > 0 && is->is_timeout > iscsid_timeout) {
 567                         ISCSI_SESSION_WARN(is, "timed out waiting for iscsid(8) "
 568                             "for %d seconds; reconnecting",
 569                             is->is_timeout);
 570                         reconnect_needed = true;
 571                 }
 572                 goto out;
 573         }
 574
 575         if (is->is_login_phase) {
 576                 if (login_timeout > 0 && is->is_timeout > login_timeout) {
 577                         ISCSI_SESSION_WARN(is, "login timed out after %d seconds; "
 578                             "reconnecting", is->is_timeout);
 579                         reconnect_needed = true;
 580                 }
 581                 goto out;
 582         }
 583
 584         if (ping_timeout <= 0) {
 585                 /*
 586                  * Pings are disabled.  Don't send NOP-Out in this case.
 587                  * Reset the timeout, to avoid triggering reconnection,
 588                  * should the user decide to reenable them.
 589                  */
 590                 is->is_timeout = 0;
 591                 goto out;
 592         }
 593
 594         if (is->is_timeout >= ping_timeout) {
 595                 ISCSI_SESSION_WARN(is, "no ping reply (NOP-In) after %d seconds; "
 596                     "reconnecting", ping_timeout);
 597                 reconnect_needed = true;
 598                 goto out;
 599         }
 600
 601         ISCSI_SESSION_UNLOCK(is);
 602
 603         /*
 604          * If the ping was reset less than one second ago - which means
 605          * that we've received some PDU during the last second - assume
 606          * the traffic flows correctly and don't bother sending a NOP-Out.
 607          *
 608          * (It's 2 - one for one second, and one for incrementing is_timeout
 609          * earlier in this routine.)
 610          */
 611         if (is->is_timeout < 2)
 612                 return;
 613
 614         request = icl_pdu_new(is->is_conn, M_NOWAIT);
 615         if (request == NULL) {
 616                 ISCSI_SESSION_WARN(is, "failed to allocate PDU");
 617                 return;
 618         }
 619         bhsno = (struct iscsi_bhs_nop_out *)request->ip_bhs;
 620         bhsno->bhsno_opcode = ISCSI_BHS_OPCODE_NOP_OUT |
 621             ISCSI_BHS_OPCODE_IMMEDIATE;
 622         bhsno->bhsno_flags = 0x80;
 623         bhsno->bhsno_target_transfer_tag = 0xffffffff;
 624         iscsi_pdu_queue(request);
 625         return;
 626
 627 out:
 628         if (is->is_terminating) {
 629                 ISCSI_SESSION_UNLOCK(is);
 630                 return;
 631         }
 632
 633         ISCSI_SESSION_UNLOCK(is);
 634
 635         if (reconnect_needed)
 636                 iscsi_session_reconnect(is);
 637 }
 638
 639 static void
 640 iscsi_pdu_update_statsn(const struct icl_pdu *response)
 641 {
 642         const struct iscsi_bhs_data_in *bhsdi;
 643         struct iscsi_session *is;
 644         uint32_t expcmdsn, maxcmdsn, statsn;
 645
 646         is = PDU_SESSION(response);
 647
 648         ISCSI_SESSION_LOCK_ASSERT(is);
 649
 650         /*
 651          * We're only using fields common for all the response
 652          * (target -> initiator) PDUs.
 653          */
 654         bhsdi = (const struct iscsi_bhs_data_in *)response->ip_bhs;
 655         /*
 656          * Ok, I lied.  In case of Data-In, "The fields StatSN, Status,
 657          * and Residual Count only have meaningful content if the S bit
 658          * is set to 1", so we also need to check the bit specific for
 659          * Data-In PDU.
 660          */
 661         if (bhsdi->bhsdi_opcode != ISCSI_BHS_OPCODE_SCSI_DATA_IN ||
 662             (bhsdi->bhsdi_flags & BHSDI_FLAGS_S) != 0) {
 663                 statsn = ntohl(bhsdi->bhsdi_statsn);
 664                 if (statsn != is->is_statsn && statsn != (is->is_statsn + 1)) {
 665                         /* XXX: This is normal situation for MCS */
 666                         ISCSI_SESSION_WARN(is, "PDU 0x%x StatSN %u != "
 667                             "session ExpStatSN %u (or + 1); reconnecting",
 668                             bhsdi->bhsdi_opcode, statsn, is->is_statsn);
 669                         iscsi_session_reconnect(is);
 670                 }
 671                 if (ISCSI_SNGT(statsn, is->is_statsn))
 672                         is->is_statsn = statsn;
 673         }
 674
 675         expcmdsn = ntohl(bhsdi->bhsdi_expcmdsn);
 676         maxcmdsn = ntohl(bhsdi->bhsdi_maxcmdsn);
 677
 678         if (ISCSI_SNLT(maxcmdsn + 1, expcmdsn)) {
 679                 ISCSI_SESSION_DEBUG(is,
 680                     "PDU MaxCmdSN %u + 1 < PDU ExpCmdSN %u; ignoring",
 681                     maxcmdsn, expcmdsn);
 682         } else {
 683                 if (ISCSI_SNGT(maxcmdsn, is->is_maxcmdsn)) {
 684                         is->is_maxcmdsn = maxcmdsn;
 685
 686                         /*
 687                          * Command window increased; kick the maintanance thread
 688                          * to send out postponed commands.
 689                          */
 690                         if (!STAILQ_EMPTY(&is->is_postponed))
 691                                 cv_signal(&is->is_maintenance_cv);
 692                 } else if (ISCSI_SNLT(maxcmdsn, is->is_maxcmdsn)) {
 693                         /* XXX: This is normal situation for MCS */
 694                         ISCSI_SESSION_DEBUG(is,
 695                             "PDU MaxCmdSN %u < session MaxCmdSN %u; ignoring",
 696                             maxcmdsn, is->is_maxcmdsn);
 697                 }
 698
 699                 if (ISCSI_SNGT(expcmdsn, is->is_expcmdsn)) {
 700                         is->is_expcmdsn = expcmdsn;
 701                 } else if (ISCSI_SNLT(expcmdsn, is->is_expcmdsn)) {
 702                         /* XXX: This is normal situation for MCS */
 703                         ISCSI_SESSION_DEBUG(is,
 704                             "PDU ExpCmdSN %u < session ExpCmdSN %u; ignoring",
 705                             expcmdsn, is->is_expcmdsn);
 706                 }
 707         }
 708
 709         /*
 710          * Every incoming PDU - not just NOP-In - resets the ping timer.
 711          * The purpose of the timeout is to reset the connection when it stalls;
 712          * we don't want this to happen when NOP-In or NOP-Out ends up delayed
 713          * in some queue.
 714          */
 715         is->is_timeout = 0;
 716 }
 717
 718 static void
 719 iscsi_receive_callback(struct icl_pdu *response)
 720 {
 721         struct iscsi_session *is;
 722
 723         is = PDU_SESSION(response);
 724
 725         ISCSI_SESSION_LOCK(is);
 726
 727         iscsi_pdu_update_statsn(response);
 728
 729 #ifdef ICL_KERNEL_PROXY
 730         if (is->is_login_phase) {
 731                 if (is->is_login_pdu == NULL)
 732                         is->is_login_pdu = response;
 733                 else
 734                         icl_pdu_free(response);
 735                 ISCSI_SESSION_UNLOCK(is);
 736                 cv_signal(&is->is_login_cv);
 737                 return;
 738         }
 739 #endif
 740
 741         /*
 742          * The handling routine is responsible for freeing the PDU
 743          * when it's no longer needed.
 744          */
 745         switch (response->ip_bhs->bhs_opcode) {
 746         case ISCSI_BHS_OPCODE_NOP_IN:
 747                 iscsi_pdu_handle_nop_in(response);
 748                 ISCSI_SESSION_UNLOCK(is);
 749                 break;
 750         case ISCSI_BHS_OPCODE_SCSI_RESPONSE:
 751                 iscsi_pdu_handle_scsi_response(response);
 752                 /* Session lock dropped inside. */
 753                 ISCSI_SESSION_LOCK_ASSERT_NOT(is);
 754                 break;
 755         case ISCSI_BHS_OPCODE_TASK_RESPONSE:
 756                 iscsi_pdu_handle_task_response(response);
 757                 ISCSI_SESSION_UNLOCK(is);
 758                 break;
 759         case ISCSI_BHS_OPCODE_SCSI_DATA_IN:
 760                 iscsi_pdu_handle_data_in(response);
 761                 /* Session lock dropped inside. */
 762                 ISCSI_SESSION_LOCK_ASSERT_NOT(is);
 763                 break;
 764         case ISCSI_BHS_OPCODE_LOGOUT_RESPONSE:
 765                 iscsi_pdu_handle_logout_response(response);
 766                 ISCSI_SESSION_UNLOCK(is);
 767                 break;
 768         case ISCSI_BHS_OPCODE_R2T:
 769                 iscsi_pdu_handle_r2t(response);
 770                 ISCSI_SESSION_UNLOCK(is);
 771                 break;
 772         case ISCSI_BHS_OPCODE_ASYNC_MESSAGE:
 773                 iscsi_pdu_handle_async_message(response);
 774                 ISCSI_SESSION_UNLOCK(is);
 775                 break;
 776         case ISCSI_BHS_OPCODE_REJECT:
 777                 iscsi_pdu_handle_reject(response);
 778                 ISCSI_SESSION_UNLOCK(is);
 779                 break;
 780         default:
 781                 ISCSI_SESSION_WARN(is, "received PDU with unsupported "
 782                     "opcode 0x%x; reconnecting",
 783                     response->ip_bhs->bhs_opcode);
 784                 iscsi_session_reconnect(is);
 785                 ISCSI_SESSION_UNLOCK(is);
 786                 icl_pdu_free(response);
 787         }
 788 }
 789
 790 static void
 791 iscsi_error_callback(struct icl_conn *ic)
 792 {
 793         struct iscsi_session *is;
 794
 795         is = CONN_SESSION(ic);
 796
 797         ISCSI_SESSION_WARN(is, "connection error; reconnecting");
 798         iscsi_session_reconnect(is);
 799 }
 800
 801 static void
 802 iscsi_pdu_handle_nop_in(struct icl_pdu *response)
 803 {
 804         struct iscsi_session *is;
 805         struct iscsi_bhs_nop_out *bhsno;
 806         struct iscsi_bhs_nop_in *bhsni;
 807         struct icl_pdu *request;
 808         void *data = NULL;
 809         size_t datasize;
 810         int error;
 811
 812         is = PDU_SESSION(response);
 813         bhsni = (struct iscsi_bhs_nop_in *)response->ip_bhs;
 814
 815         if (bhsni->bhsni_target_transfer_tag == 0xffffffff) {
 816                 /*
 817                  * Nothing to do; iscsi_pdu_update_statsn() already
 818                  * zeroed the timeout.
 819                  */
 820                 icl_pdu_free(response);
 821                 return;
 822         }
 823
 824         datasize = icl_pdu_data_segment_length(response);
 825         if (datasize > 0) {
 826                 data = malloc(datasize, M_ISCSI, M_NOWAIT | M_ZERO);
 827                 if (data == NULL) {
 828                         ISCSI_SESSION_WARN(is, "failed to allocate memory; "
 829                             "reconnecting");
 830                         icl_pdu_free(response);
 831                         iscsi_session_reconnect(is);
 832                         return;
 833                 }
 834                 icl_pdu_get_data(response, 0, data, datasize);
 835         }
 836
 837         request = icl_pdu_new(response->ip_conn, M_NOWAIT);
 838         if (request == NULL) {
 839                 ISCSI_SESSION_WARN(is, "failed to allocate memory; "
 840                     "reconnecting");
 841                 free(data, M_ISCSI);
 842                 icl_pdu_free(response);
 843                 iscsi_session_reconnect(is);
 844                 return;
 845         }
 846         bhsno = (struct iscsi_bhs_nop_out *)request->ip_bhs;
 847         bhsno->bhsno_opcode = ISCSI_BHS_OPCODE_NOP_OUT |
 848             ISCSI_BHS_OPCODE_IMMEDIATE;
 849         bhsno->bhsno_flags = 0x80;
 850         bhsno->bhsno_initiator_task_tag = 0xffffffff;
 851         bhsno->bhsno_target_transfer_tag = bhsni->bhsni_target_transfer_tag;
 852         if (datasize > 0) {
 853                 error = icl_pdu_append_data(request, data, datasize, M_NOWAIT);
 854                 if (error != 0) {
 855                         ISCSI_SESSION_WARN(is, "failed to allocate memory; "
 856                             "reconnecting");
 857                         free(data, M_ISCSI);
 858                         icl_pdu_free(request);
 859                         icl_pdu_free(response);
 860                         iscsi_session_reconnect(is);
 861                         return;
 862                 }
 863                 free(data, M_ISCSI);
 864         }
 865
 866         icl_pdu_free(response);
 867         iscsi_pdu_queue_locked(request);
 868 }
 869
 870 static void
 871 iscsi_pdu_handle_scsi_response(struct icl_pdu *response)
 872 {
 873         struct iscsi_bhs_scsi_response *bhssr;
 874         struct iscsi_outstanding *io;
 875         struct iscsi_session *is;
 876         union ccb *ccb;
 877         struct ccb_scsiio *csio;
 878         size_t data_segment_len, received;
 879         uint16_t sense_len;
 880         uint32_t resid;
 881
 882         is = PDU_SESSION(response);
 883
 884         bhssr = (struct iscsi_bhs_scsi_response *)response->ip_bhs;
 885         io = iscsi_outstanding_find(is, bhssr->bhssr_initiator_task_tag);
 886         if (io == NULL || io->io_ccb == NULL) {
 887                 ISCSI_SESSION_WARN(is, "bad itt 0x%x", bhssr->bhssr_initiator_task_tag);
 888                 icl_pdu_free(response);
 889                 iscsi_session_reconnect(is);
 890                 ISCSI_SESSION_UNLOCK(is);
 891                 return;
 892         }
 893
 894         ccb = io->io_ccb;
 895
 896         /*
 897          * With iSER, after getting good response we can be sure
 898          * that all the data has been successfully transferred.
 899          */
 900         if (is->is_conn->ic_iser) {
 901                 resid = ntohl(bhssr->bhssr_residual_count);
 902                 if (bhssr->bhssr_flags & BHSSR_FLAGS_RESIDUAL_UNDERFLOW) {
 903                         io->io_received = ccb->csio.dxfer_len - resid;
 904                 } else if (bhssr->bhssr_flags & BHSSR_FLAGS_RESIDUAL_OVERFLOW) {
 905                         ISCSI_SESSION_WARN(is, "overflow: target indicates %d", resid);
 906                 } else {
 907                         io->io_received = ccb->csio.dxfer_len;
 908                 }
 909         }
 910
 911         received = io->io_received;
 912         iscsi_outstanding_remove(is, io);
 913         ISCSI_SESSION_UNLOCK(is);
 914
 915         if (bhssr->bhssr_response != BHSSR_RESPONSE_COMMAND_COMPLETED) {
 916                 ISCSI_SESSION_WARN(is, "service response 0x%x", bhssr->bhssr_response);
 917                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
 918                         xpt_freeze_devq(ccb->ccb_h.path, 1);
 919                         ISCSI_SESSION_DEBUG(is, "freezing devq");
 920                 }
 921                 ccb->ccb_h.status = CAM_REQ_CMP_ERR | CAM_DEV_QFRZN;
 922         } else if (bhssr->bhssr_status == 0) {
 923                 ccb->ccb_h.status = CAM_REQ_CMP;
 924         } else {
 925                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
 926                         xpt_freeze_devq(ccb->ccb_h.path, 1);
 927                         ISCSI_SESSION_DEBUG(is, "freezing devq");
 928                 }
 929                 ccb->ccb_h.status = CAM_SCSI_STATUS_ERROR | CAM_DEV_QFRZN;
 930                 ccb->csio.scsi_status = bhssr->bhssr_status;
 931         }
 932
 933         csio = &ccb->csio;
 934         data_segment_len = icl_pdu_data_segment_length(response);
 935         if (data_segment_len > 0) {
 936                 if (data_segment_len < sizeof(sense_len)) {
 937                         ISCSI_SESSION_WARN(is, "truncated data segment (%zd bytes)",
 938                             data_segment_len);
 939                         if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
 940                                 xpt_freeze_devq(ccb->ccb_h.path, 1);
 941                                 ISCSI_SESSION_DEBUG(is, "freezing devq");
 942                         }
 943                         ccb->ccb_h.status = CAM_REQ_CMP_ERR | CAM_DEV_QFRZN;
 944                         goto out;
 945                 }
 946                 icl_pdu_get_data(response, 0, &sense_len, sizeof(sense_len));
 947                 sense_len = ntohs(sense_len);
 948 #if 0
 949                 ISCSI_SESSION_DEBUG(is, "sense_len %d, data len %zd",
 950                     sense_len, data_segment_len);
 951 #endif
 952                 if (sizeof(sense_len) + sense_len > data_segment_len) {
 953                         ISCSI_SESSION_WARN(is, "truncated data segment "
 954                             "(%zd bytes, should be %zd)",
 955                             data_segment_len, sizeof(sense_len) + sense_len);
 956                         if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
 957                                 xpt_freeze_devq(ccb->ccb_h.path, 1);
 958                                 ISCSI_SESSION_DEBUG(is, "freezing devq");
 959                         }
 960                         ccb->ccb_h.status = CAM_REQ_CMP_ERR | CAM_DEV_QFRZN;
 961                         goto out;
 962                 } else if (sizeof(sense_len) + sense_len < data_segment_len)
 963                         ISCSI_SESSION_WARN(is, "oversize data segment "
 964                             "(%zd bytes, should be %zd)",
 965                             data_segment_len, sizeof(sense_len) + sense_len);
 966                 if (sense_len > csio->sense_len) {
 967                         ISCSI_SESSION_DEBUG(is, "truncating sense from %d to %d",
 968                             sense_len, csio->sense_len);
 969                         sense_len = csio->sense_len;
 970                 }
 971                 icl_pdu_get_data(response, sizeof(sense_len), &csio->sense_data, sense_len);
 972                 csio->sense_resid = csio->sense_len - sense_len;
 973                 ccb->ccb_h.status |= CAM_AUTOSNS_VALID;
 974         }
 975
 976 out:
 977         if (bhssr->bhssr_flags & BHSSR_FLAGS_RESIDUAL_UNDERFLOW)
 978                 csio->resid = ntohl(bhssr->bhssr_residual_count);
 979
 980         if ((csio->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_IN) {
 981                 KASSERT(received <= csio->dxfer_len,
 982                     ("received > csio->dxfer_len"));
 983                 if (received < csio->dxfer_len) {
 984                         if (csio->resid != csio->dxfer_len - received) {
 985                                 ISCSI_SESSION_WARN(is, "underflow mismatch: "
 986                                     "target indicates %d, we calculated %zd",
 987                                     csio->resid, csio->dxfer_len - received);
 988                         }
 989                         csio->resid = csio->dxfer_len - received;
 990                 }
 991         }
 992
 993         xpt_done(ccb);
 994         icl_pdu_free(response);
 995 }
 996
 997 static void
 998 iscsi_pdu_handle_task_response(struct icl_pdu *response)
 999 {
1000         struct iscsi_bhs_task_management_response *bhstmr;
1001         struct iscsi_outstanding *io, *aio;
1002         struct iscsi_session *is;
1003
1004         is = PDU_SESSION(response);
1005
1006         bhstmr = (struct iscsi_bhs_task_management_response *)response->ip_bhs;
1007         io = iscsi_outstanding_find(is, bhstmr->bhstmr_initiator_task_tag);
1008         if (io == NULL || io->io_ccb != NULL) {
1009                 ISCSI_SESSION_WARN(is, "bad itt 0x%x",
1010                     bhstmr->bhstmr_initiator_task_tag);
1011                 icl_pdu_free(response);
1012                 iscsi_session_reconnect(is);
1013                 return;
1014         }
1015
1016         if (bhstmr->bhstmr_response != BHSTMR_RESPONSE_FUNCTION_COMPLETE) {
1017                 ISCSI_SESSION_WARN(is, "task response 0x%x",
1018                     bhstmr->bhstmr_response);
1019         } else {
1020                 aio = iscsi_outstanding_find(is, io->io_referenced_task_tag);
1021                 if (aio != NULL && aio->io_ccb != NULL)
1022                         iscsi_session_terminate_task(is, aio, CAM_REQ_ABORTED);
1023         }
1024
1025         iscsi_outstanding_remove(is, io);
1026         icl_pdu_free(response);
1027 }
1028
1029 static void
1030 iscsi_pdu_handle_data_in(struct icl_pdu *response)
1031 {
1032         struct iscsi_bhs_data_in *bhsdi;
1033         struct iscsi_outstanding *io;
1034         struct iscsi_session *is;
1035         union ccb *ccb;
1036         struct ccb_scsiio *csio;
1037         size_t data_segment_len, received, oreceived;
1038
1039         is = PDU_SESSION(response);
1040         bhsdi = (struct iscsi_bhs_data_in *)response->ip_bhs;
1041         io = iscsi_outstanding_find(is, bhsdi->bhsdi_initiator_task_tag);
1042         if (io == NULL || io->io_ccb == NULL) {
1043                 ISCSI_SESSION_WARN(is, "bad itt 0x%x", bhsdi->bhsdi_initiator_task_tag);
1044                 icl_pdu_free(response);
1045                 iscsi_session_reconnect(is);
1046                 ISCSI_SESSION_UNLOCK(is);
1047                 return;
1048         }
1049
1050         data_segment_len = icl_pdu_data_segment_length(response);
1051         if (data_segment_len == 0) {
1052                 /*
1053                  * "The sending of 0 length data segments should be avoided,
1054                  * but initiators and targets MUST be able to properly receive
1055                  * 0 length data segments."
1056                  */
1057                 ISCSI_SESSION_UNLOCK(is);
1058                 icl_pdu_free(response);
1059                 return;
1060         }
1061
1062         /*
1063          * We need to track this for security reasons - without it, malicious target
1064          * could respond to SCSI READ without sending Data-In PDUs, which would result
1065          * in read operation on the initiator side returning random kernel data.
1066          */
1067         if (ntohl(bhsdi->bhsdi_buffer_offset) != io->io_received) {
1068                 ISCSI_SESSION_WARN(is, "data out of order; expected offset %zd, got %zd",
1069                     io->io_received, (size_t)ntohl(bhsdi->bhsdi_buffer_offset));
1070                 icl_pdu_free(response);
1071                 iscsi_session_reconnect(is);
1072                 ISCSI_SESSION_UNLOCK(is);
1073                 return;
1074         }
1075
1076         ccb = io->io_ccb;
1077         csio = &ccb->csio;
1078
1079         if (io->io_received + data_segment_len > csio->dxfer_len) {
1080                 ISCSI_SESSION_WARN(is, "oversize data segment (%zd bytes "
1081                     "at offset %zd, buffer is %d)",
1082                     data_segment_len, io->io_received, csio->dxfer_len);
1083                 icl_pdu_free(response);
1084                 iscsi_session_reconnect(is);
1085                 ISCSI_SESSION_UNLOCK(is);
1086                 return;
1087         }
1088
1089         oreceived = io->io_received;
1090         io->io_received += data_segment_len;
1091         received = io->io_received;
1092         if ((bhsdi->bhsdi_flags & BHSDI_FLAGS_S) != 0)
1093                 iscsi_outstanding_remove(is, io);
1094         ISCSI_SESSION_UNLOCK(is);
1095
1096         icl_pdu_get_data(response, 0, csio->data_ptr + oreceived, data_segment_len);
1097
1098         /*
1099          * XXX: Check DataSN.
1100          * XXX: Check F.
1101          */
1102         if ((bhsdi->bhsdi_flags & BHSDI_FLAGS_S) == 0) {
1103                 /*
1104                  * Nothing more to do.
1105                  */
1106                 icl_pdu_free(response);
1107                 return;
1108         }
1109
1110         //ISCSI_SESSION_DEBUG(is, "got S flag; status 0x%x", bhsdi->bhsdi_status);
1111         if (bhsdi->bhsdi_status == 0) {
1112                 ccb->ccb_h.status = CAM_REQ_CMP;
1113         } else {
1114                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
1115                         xpt_freeze_devq(ccb->ccb_h.path, 1);
1116                         ISCSI_SESSION_DEBUG(is, "freezing devq");
1117                 }
1118                 ccb->ccb_h.status = CAM_SCSI_STATUS_ERROR | CAM_DEV_QFRZN;
1119                 csio->scsi_status = bhsdi->bhsdi_status;
1120         }
1121
1122         if ((csio->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_IN) {
1123                 KASSERT(received <= csio->dxfer_len,
1124                     ("received > csio->dxfer_len"));
1125                 if (received < csio->dxfer_len) {
1126                         csio->resid = ntohl(bhsdi->bhsdi_residual_count);
1127                         if (csio->resid != csio->dxfer_len - received) {
1128                                 ISCSI_SESSION_WARN(is, "underflow mismatch: "
1129                                     "target indicates %d, we calculated %zd",
1130                                     csio->resid, csio->dxfer_len - received);
1131                         }
1132                         csio->resid = csio->dxfer_len - received;
1133                 }
1134         }
1135
1136         xpt_done(ccb);
1137         icl_pdu_free(response);
1138 }
1139
1140 static void
1141 iscsi_pdu_handle_logout_response(struct icl_pdu *response)
1142 {
1143
1144         ISCSI_SESSION_DEBUG(PDU_SESSION(response), "logout response");
1145         icl_pdu_free(response);
1146 }
1147
1148 static void
1149 iscsi_pdu_handle_r2t(struct icl_pdu *response)
1150 {
1151         struct icl_pdu *request;
1152         struct iscsi_session *is;
1153         struct iscsi_bhs_r2t *bhsr2t;
1154         struct iscsi_bhs_data_out *bhsdo;
1155         struct iscsi_outstanding *io;
1156         struct ccb_scsiio *csio;
1157         size_t off, len, total_len;
1158         int error;
1159         uint32_t datasn = 0;
1160
1161         is = PDU_SESSION(response);
1162
1163         bhsr2t = (struct iscsi_bhs_r2t *)response->ip_bhs;
1164         io = iscsi_outstanding_find(is, bhsr2t->bhsr2t_initiator_task_tag);
1165         if (io == NULL || io->io_ccb == NULL) {
1166                 ISCSI_SESSION_WARN(is, "bad itt 0x%x; reconnecting",
1167                     bhsr2t->bhsr2t_initiator_task_tag);
1168                 icl_pdu_free(response);
1169                 iscsi_session_reconnect(is);
1170                 return;
1171         }
1172
1173         csio = &io->io_ccb->csio;
1174
1175         if ((csio->ccb_h.flags & CAM_DIR_MASK) != CAM_DIR_OUT) {
1176                 ISCSI_SESSION_WARN(is, "received R2T for read command; reconnecting");
1177                 icl_pdu_free(response);
1178                 iscsi_session_reconnect(is);
1179                 return;
1180         }
1181
1182         /*
1183          * XXX: Verify R2TSN.
1184          */
1185
1186         off = ntohl(bhsr2t->bhsr2t_buffer_offset);
1187         if (off > csio->dxfer_len) {
1188                 ISCSI_SESSION_WARN(is, "target requested invalid offset "
1189                     "%zd, buffer is is %d; reconnecting", off, csio->dxfer_len);
1190                 icl_pdu_free(response);
1191                 iscsi_session_reconnect(is);
1192                 return;
1193         }
1194
1195         total_len = ntohl(bhsr2t->bhsr2t_desired_data_transfer_length);
1196         if (total_len == 0 || total_len > csio->dxfer_len) {
1197                 ISCSI_SESSION_WARN(is, "target requested invalid length "
1198                     "%zd, buffer is %d; reconnecting", total_len, csio->dxfer_len);
1199                 icl_pdu_free(response);
1200                 iscsi_session_reconnect(is);
1201                 return;
1202         }
1203
1204         //ISCSI_SESSION_DEBUG(is, "r2t; off %zd, len %zd", off, total_len);
1205
1206         for (;;) {
1207                 len = total_len;
1208
1209                 if (len > is->is_conn->ic_max_send_data_segment_length)
1210                         len = is->is_conn->ic_max_send_data_segment_length;
1211
1212                 if (off + len > csio->dxfer_len) {
1213                         ISCSI_SESSION_WARN(is, "target requested invalid "
1214                             "length/offset %zd, buffer is %d; reconnecting",
1215                             off + len, csio->dxfer_len);
1216                         icl_pdu_free(response);
1217                         iscsi_session_reconnect(is);
1218                         return;
1219                 }
1220
1221                 request = icl_pdu_new(response->ip_conn, M_NOWAIT);
1222                 if (request == NULL) {
1223                         icl_pdu_free(response);
1224                         iscsi_session_reconnect(is);
1225                         return;
1226                 }
1227
1228                 bhsdo = (struct iscsi_bhs_data_out *)request->ip_bhs;
1229                 bhsdo->bhsdo_opcode = ISCSI_BHS_OPCODE_SCSI_DATA_OUT;
1230                 bhsdo->bhsdo_lun = bhsr2t->bhsr2t_lun;
1231                 bhsdo->bhsdo_initiator_task_tag =
1232                     bhsr2t->bhsr2t_initiator_task_tag;
1233                 bhsdo->bhsdo_target_transfer_tag =
1234                     bhsr2t->bhsr2t_target_transfer_tag;
1235                 bhsdo->bhsdo_datasn = htonl(datasn++);
1236                 bhsdo->bhsdo_buffer_offset = htonl(off);
1237                 error = icl_pdu_append_data(request, csio->data_ptr + off, len,
1238                     M_NOWAIT);
1239                 if (error != 0) {
1240                         ISCSI_SESSION_WARN(is, "failed to allocate memory; "
1241                             "reconnecting");
1242                         icl_pdu_free(request);
1243                         icl_pdu_free(response);
1244                         iscsi_session_reconnect(is);
1245                         return;
1246                 }
1247
1248                 off += len;
1249                 total_len -= len;
1250
1251                 if (total_len == 0) {
1252                         bhsdo->bhsdo_flags |= BHSDO_FLAGS_F;
1253                         //ISCSI_SESSION_DEBUG(is, "setting F, off %zd", off);
1254                 } else {
1255                         //ISCSI_SESSION_DEBUG(is, "not finished, off %zd", off);
1256                 }
1257
1258                 iscsi_pdu_queue_locked(request);
1259
1260                 if (total_len == 0)
1261                         break;
1262         }
1263
1264         icl_pdu_free(response);
1265 }
1266
1267 static void
1268 iscsi_pdu_handle_async_message(struct icl_pdu *response)
1269 {
1270         struct iscsi_bhs_asynchronous_message *bhsam;
1271         struct iscsi_session *is;
1272
1273         is = PDU_SESSION(response);
1274         bhsam = (struct iscsi_bhs_asynchronous_message *)response->ip_bhs;
1275         switch (bhsam->bhsam_async_event) {
1276         case BHSAM_EVENT_TARGET_REQUESTS_LOGOUT:
1277                 ISCSI_SESSION_WARN(is, "target requests logout; removing session");
1278                 iscsi_session_logout(is);
1279                 iscsi_session_terminate(is);
1280                 break;
1281         case BHSAM_EVENT_TARGET_TERMINATES_CONNECTION:
1282                 ISCSI_SESSION_WARN(is, "target indicates it will drop the connection");
1283                 break;
1284         case BHSAM_EVENT_TARGET_TERMINATES_SESSION:
1285                 ISCSI_SESSION_WARN(is, "target indicates it will drop the session");
1286                 break;
1287         default:
1288                 /*
1289                  * XXX: Technically, we're obligated to also handle
1290                  *      parameter renegotiation.
1291                  */
1292                 ISCSI_SESSION_WARN(is, "ignoring AsyncEvent %d", bhsam->bhsam_async_event);
1293                 break;
1294         }
1295
1296         icl_pdu_free(response);
1297 }
1298
1299 static void
1300 iscsi_pdu_handle_reject(struct icl_pdu *response)
1301 {
1302         struct iscsi_bhs_reject *bhsr;
1303         struct iscsi_session *is;
1304
1305         is = PDU_SESSION(response);
1306         bhsr = (struct iscsi_bhs_reject *)response->ip_bhs;
1307         ISCSI_SESSION_WARN(is, "received Reject PDU, reason 0x%x; protocol error?",
1308             bhsr->bhsr_reason);
1309
1310         icl_pdu_free(response);
1311 }
1312
1313 static int
1314 iscsi_ioctl_daemon_wait(struct iscsi_softc *sc,
1315     struct iscsi_daemon_request *request)
1316 {
1317         struct iscsi_session *is;
1318         struct icl_drv_limits idl;
1319         int error;
1320
1321         sx_slock(&sc->sc_lock);
1322         for (;;) {
1323                 TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1324                         ISCSI_SESSION_LOCK(is);
1325                         if (is->is_conf.isc_enable == 0 &&
1326                             is->is_conf.isc_discovery == 0) {
1327                                 ISCSI_SESSION_UNLOCK(is);
1328                                 continue;
1329                         }
1330                         if (is->is_waiting_for_iscsid)
1331                                 break;
1332                         ISCSI_SESSION_UNLOCK(is);
1333                 }
1334
1335                 if (is == NULL) {
1336                         if (sc->sc_unloading) {
1337                                 sx_sunlock(&sc->sc_lock);
1338                                 return (ENXIO);
1339                         }
1340
1341                         /*
1342                          * No session requires attention from iscsid(8); wait.
1343                          */
1344                         error = cv_wait_sig(&sc->sc_cv, &sc->sc_lock);
1345                         if (error != 0) {
1346                                 sx_sunlock(&sc->sc_lock);
1347                                 return (error);
1348                         }
1349                         continue;
1350                 }
1351
1352                 is->is_waiting_for_iscsid = false;
1353                 is->is_login_phase = true;
1354                 is->is_reason[0] = '\0';
1355                 ISCSI_SESSION_UNLOCK(is);
1356
1357                 request->idr_session_id = is->is_id;
1358                 memcpy(&request->idr_isid, &is->is_isid,
1359                     sizeof(request->idr_isid));
1360                 request->idr_tsih = 0;  /* New or reinstated session. */
1361                 memcpy(&request->idr_conf, &is->is_conf,
1362                     sizeof(request->idr_conf));
1363
1364                 error = icl_limits(is->is_conf.isc_offload,
1365                     is->is_conf.isc_iser, &idl);
1366                 if (error != 0) {
1367                         ISCSI_SESSION_WARN(is, "icl_limits for offload \"%s\" "
1368                             "failed with error %d", is->is_conf.isc_offload,
1369                             error);
1370                         sx_sunlock(&sc->sc_lock);
1371                         return (error);
1372                 }
1373                 request->idr_limits.isl_max_recv_data_segment_length =
1374                     idl.idl_max_recv_data_segment_length;
1375                 request->idr_limits.isl_max_send_data_segment_length =
1376                     idl.idl_max_send_data_segment_length;
1377                 request->idr_limits.isl_max_burst_length =
1378                     idl.idl_max_burst_length;
1379                 request->idr_limits.isl_first_burst_length =
1380                     idl.idl_first_burst_length;
1381
1382                 sx_sunlock(&sc->sc_lock);
1383                 return (0);
1384         }
1385 }
1386
1387 static int
1388 iscsi_ioctl_daemon_handoff(struct iscsi_softc *sc,
1389     struct iscsi_daemon_handoff *handoff)
1390 {
1391         struct iscsi_session *is;
1392         struct icl_conn *ic;
1393         int error;
1394
1395         sx_slock(&sc->sc_lock);
1396
1397         /*
1398          * Find the session to hand off socket to.
1399          */
1400         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1401                 if (is->is_id == handoff->idh_session_id)
1402                         break;
1403         }
1404         if (is == NULL) {
1405                 sx_sunlock(&sc->sc_lock);
1406                 return (ESRCH);
1407         }
1408         ISCSI_SESSION_LOCK(is);
1409         ic = is->is_conn;
1410         if (is->is_conf.isc_discovery || is->is_terminating) {
1411                 ISCSI_SESSION_UNLOCK(is);
1412                 sx_sunlock(&sc->sc_lock);
1413                 return (EINVAL);
1414         }
1415         if (is->is_connected) {
1416                 /*
1417                  * This might have happened because another iscsid(8)
1418                  * instance handed off the connection in the meantime.
1419                  * Just return.
1420                  */
1421                 ISCSI_SESSION_WARN(is, "handoff on already connected "
1422                     "session");
1423                 ISCSI_SESSION_UNLOCK(is);
1424                 sx_sunlock(&sc->sc_lock);
1425                 return (EBUSY);
1426         }
1427
1428         strlcpy(is->is_target_alias, handoff->idh_target_alias,
1429             sizeof(is->is_target_alias));
1430         is->is_tsih = handoff->idh_tsih;
1431         is->is_statsn = handoff->idh_statsn;
1432         is->is_protocol_level = handoff->idh_protocol_level;
1433         is->is_initial_r2t = handoff->idh_initial_r2t;
1434         is->is_immediate_data = handoff->idh_immediate_data;
1435
1436         ic->ic_max_recv_data_segment_length =
1437             handoff->idh_max_recv_data_segment_length;
1438         ic->ic_max_send_data_segment_length =
1439             handoff->idh_max_send_data_segment_length;
1440         is->is_max_burst_length = handoff->idh_max_burst_length;
1441         is->is_first_burst_length = handoff->idh_first_burst_length;
1442
1443         if (handoff->idh_header_digest == ISCSI_DIGEST_CRC32C)
1444                 ic->ic_header_crc32c = true;
1445         else
1446                 ic->ic_header_crc32c = false;
1447         if (handoff->idh_data_digest == ISCSI_DIGEST_CRC32C)
1448                 ic->ic_data_crc32c = true;
1449         else
1450                 ic->ic_data_crc32c = false;
1451         ic->ic_maxtags = maxtags;
1452
1453         is->is_cmdsn = 0;
1454         is->is_expcmdsn = 0;
1455         is->is_maxcmdsn = 0;
1456         is->is_waiting_for_iscsid = false;
1457         is->is_login_phase = false;
1458         is->is_timeout = 0;
1459         is->is_connected = true;
1460         is->is_reason[0] = '\0';
1461
1462         ISCSI_SESSION_UNLOCK(is);
1463
1464         /*
1465          * If we're going through the proxy, the idh_socket will be 0,
1466          * and the ICL module can simply ignore this call.  It can also
1467          * use it to determine it's no longer in the Login phase.
1468          */
1469         error = icl_conn_handoff(ic, handoff->idh_socket);
1470         if (error != 0) {
1471                 sx_sunlock(&sc->sc_lock);
1472                 iscsi_session_terminate(is);
1473                 return (error);
1474         }
1475
1476         sx_sunlock(&sc->sc_lock);
1477
1478         if (is->is_sim != NULL) {
1479                 /*
1480                  * When reconnecting, there already is SIM allocated for the session.
1481                  */
1482                 KASSERT(is->is_simq_frozen, ("reconnect without frozen simq"));
1483                 ISCSI_SESSION_LOCK(is);
1484                 ISCSI_SESSION_DEBUG(is, "releasing");
1485                 is->is_simq_frozen = false;
1486                 xpt_release_simq(is->is_sim, 1);
1487                 ISCSI_SESSION_UNLOCK(is);
1488
1489         } else {
1490                 ISCSI_SESSION_LOCK(is);
1491                 is->is_devq = cam_simq_alloc(ic->ic_maxtags);
1492                 if (is->is_devq == NULL) {
1493                         ISCSI_SESSION_WARN(is, "failed to allocate simq");
1494                         iscsi_session_terminate(is);
1495                         return (ENOMEM);
1496                 }
1497
1498                 is->is_sim = cam_sim_alloc(iscsi_action, NULL, "iscsi",
1499                     is, is->is_id /* unit */, &is->is_lock,
1500                     1, ic->ic_maxtags, is->is_devq);
1501                 if (is->is_sim == NULL) {
1502                         ISCSI_SESSION_UNLOCK(is);
1503                         ISCSI_SESSION_WARN(is, "failed to allocate SIM");
1504                         cam_simq_free(is->is_devq);
1505                         iscsi_session_terminate(is);
1506                         return (ENOMEM);
1507                 }
1508
1509                 error = xpt_bus_register(is->is_sim, NULL, 0);
1510                 if (error != 0) {
1511                         ISCSI_SESSION_UNLOCK(is);
1512                         ISCSI_SESSION_WARN(is, "failed to register bus");
1513                         iscsi_session_terminate(is);
1514                         return (ENOMEM);
1515                 }
1516
1517                 error = xpt_create_path(&is->is_path, /*periph*/NULL,
1518                     cam_sim_path(is->is_sim), CAM_TARGET_WILDCARD,
1519                     CAM_LUN_WILDCARD);
1520                 if (error != CAM_REQ_CMP) {
1521                         ISCSI_SESSION_UNLOCK(is);
1522                         ISCSI_SESSION_WARN(is, "failed to create path");
1523                         iscsi_session_terminate(is);
1524                         return (ENOMEM);
1525                 }
1526                 ISCSI_SESSION_UNLOCK(is);
1527         }
1528
1529         return (0);
1530 }
1531
1532 static int
1533 iscsi_ioctl_daemon_fail(struct iscsi_softc *sc,
1534     struct iscsi_daemon_fail *fail)
1535 {
1536         struct iscsi_session *is;
1537
1538         sx_slock(&sc->sc_lock);
1539
1540         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1541                 if (is->is_id == fail->idf_session_id)
1542                         break;
1543         }
1544         if (is == NULL) {
1545                 sx_sunlock(&sc->sc_lock);
1546                 return (ESRCH);
1547         }
1548         ISCSI_SESSION_LOCK(is);
1549         ISCSI_SESSION_DEBUG(is, "iscsid(8) failed: %s",
1550             fail->idf_reason);
1551         strlcpy(is->is_reason, fail->idf_reason, sizeof(is->is_reason));
1552         //is->is_waiting_for_iscsid = false;
1553         //is->is_login_phase = true;
1554         //iscsi_session_reconnect(is);
1555         ISCSI_SESSION_UNLOCK(is);
1556         sx_sunlock(&sc->sc_lock);
1557
1558         return (0);
1559 }
1560
1561 #ifdef ICL_KERNEL_PROXY
1562 static int
1563 iscsi_ioctl_daemon_connect(struct iscsi_softc *sc,
1564     struct iscsi_daemon_connect *idc)
1565 {
1566         struct iscsi_session *is;
1567         struct sockaddr *from_sa, *to_sa;
1568         int error;
1569
1570         sx_slock(&sc->sc_lock);
1571         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1572                 if (is->is_id == idc->idc_session_id)
1573                         break;
1574         }
1575         if (is == NULL) {
1576                 sx_sunlock(&sc->sc_lock);
1577                 return (ESRCH);
1578         }
1579         sx_sunlock(&sc->sc_lock);
1580
1581         if (idc->idc_from_addrlen > 0) {
1582                 error = getsockaddr(&from_sa, (void *)idc->idc_from_addr, idc->idc_from_addrlen);
1583                 if (error != 0) {
1584                         ISCSI_SESSION_WARN(is,
1585                             "getsockaddr failed with error %d", error);
1586                         return (error);
1587                 }
1588         } else {
1589                 from_sa = NULL;
1590         }
1591         error = getsockaddr(&to_sa, (void *)idc->idc_to_addr, idc->idc_to_addrlen);
1592         if (error != 0) {
1593                 ISCSI_SESSION_WARN(is, "getsockaddr failed with error %d",
1594                     error);
1595                 free(from_sa, M_SONAME);
1596                 return (error);
1597         }
1598
1599         ISCSI_SESSION_LOCK(is);
1600         is->is_statsn = 0;
1601         is->is_cmdsn = 0;
1602         is->is_expcmdsn = 0;
1603         is->is_maxcmdsn = 0;
1604         is->is_waiting_for_iscsid = false;
1605         is->is_login_phase = true;
1606         is->is_timeout = 0;
1607         ISCSI_SESSION_UNLOCK(is);
1608
1609         error = icl_conn_connect(is->is_conn, idc->idc_domain,
1610             idc->idc_socktype, idc->idc_protocol, from_sa, to_sa);
1611         free(from_sa, M_SONAME);
1612         free(to_sa, M_SONAME);
1613
1614         /*
1615          * Digests are always disabled during login phase.
1616          */
1617         is->is_conn->ic_header_crc32c = false;
1618         is->is_conn->ic_data_crc32c = false;
1619
1620         return (error);
1621 }
1622
1623 static int
1624 iscsi_ioctl_daemon_send(struct iscsi_softc *sc,
1625     struct iscsi_daemon_send *ids)
1626 {
1627         struct iscsi_session *is;
1628         struct icl_pdu *ip;
1629         size_t datalen;
1630         void *data;
1631         int error;
1632
1633         sx_slock(&sc->sc_lock);
1634         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1635                 if (is->is_id == ids->ids_session_id)
1636                         break;
1637         }
1638         if (is == NULL) {
1639                 sx_sunlock(&sc->sc_lock);
1640                 return (ESRCH);
1641         }
1642         sx_sunlock(&sc->sc_lock);
1643
1644         if (is->is_login_phase == false)
1645                 return (EBUSY);
1646
1647         if (is->is_terminating || is->is_reconnecting)
1648                 return (EIO);
1649
1650         datalen = ids->ids_data_segment_len;
1651         if (datalen > is->is_conn->ic_max_send_data_segment_length)
1652                 return (EINVAL);
1653         if (datalen > 0) {
1654                 data = malloc(datalen, M_ISCSI, M_WAITOK);
1655                 error = copyin(ids->ids_data_segment, data, datalen);
1656                 if (error != 0) {
1657                         free(data, M_ISCSI);
1658                         return (error);
1659                 }
1660         }
1661
1662         ip = icl_pdu_new(is->is_conn, M_WAITOK);
1663         memcpy(ip->ip_bhs, ids->ids_bhs, sizeof(*ip->ip_bhs));
1664         if (datalen > 0) {
1665                 error = icl_pdu_append_data(ip, data, datalen, M_WAITOK);
1666                 KASSERT(error == 0, ("icl_pdu_append_data(..., M_WAITOK) failed"));
1667                 free(data, M_ISCSI);
1668         }
1669         iscsi_pdu_queue(ip);
1670
1671         return (0);
1672 }
1673
1674 static int
1675 iscsi_ioctl_daemon_receive(struct iscsi_softc *sc,
1676     struct iscsi_daemon_receive *idr)
1677 {
1678         struct iscsi_session *is;
1679         struct icl_pdu *ip;
1680         void *data;
1681         int error;
1682
1683         sx_slock(&sc->sc_lock);
1684         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1685                 if (is->is_id == idr->idr_session_id)
1686                         break;
1687         }
1688         if (is == NULL) {
1689                 sx_sunlock(&sc->sc_lock);
1690                 return (ESRCH);
1691         }
1692         sx_sunlock(&sc->sc_lock);
1693
1694         if (is->is_login_phase == false)
1695                 return (EBUSY);
1696
1697         ISCSI_SESSION_LOCK(is);
1698         while (is->is_login_pdu == NULL &&
1699             is->is_terminating == false &&
1700             is->is_reconnecting == false) {
1701                 error = cv_wait_sig(&is->is_login_cv, &is->is_lock);
1702                 if (error != 0) {
1703                         ISCSI_SESSION_UNLOCK(is);
1704                         return (error);
1705                 }
1706         }
1707         if (is->is_terminating || is->is_reconnecting) {
1708                 ISCSI_SESSION_UNLOCK(is);
1709                 return (EIO);
1710         }
1711         ip = is->is_login_pdu;
1712         is->is_login_pdu = NULL;
1713         ISCSI_SESSION_UNLOCK(is);
1714
1715         if (ip->ip_data_len > idr->idr_data_segment_len) {
1716                 icl_pdu_free(ip);
1717                 return (EMSGSIZE);
1718         }
1719
1720         copyout(ip->ip_bhs, idr->idr_bhs, sizeof(*ip->ip_bhs));
1721         if (ip->ip_data_len > 0) {
1722                 data = malloc(ip->ip_data_len, M_ISCSI, M_WAITOK);
1723                 icl_pdu_get_data(ip, 0, data, ip->ip_data_len);
1724                 copyout(data, idr->idr_data_segment, ip->ip_data_len);
1725                 free(data, M_ISCSI);
1726         }
1727
1728         icl_pdu_free(ip);
1729
1730         return (0);
1731 }
1732 #endif /* ICL_KERNEL_PROXY */
1733
1734 static void
1735 iscsi_sanitize_session_conf(struct iscsi_session_conf *isc)
1736 {
1737         /*
1738          * Just make sure all the fields are null-terminated.
1739          *
1740          * XXX: This is not particularly secure.  We should
1741          *      create our own conf and then copy in relevant
1742          *      fields.
1743          */
1744         isc->isc_initiator[ISCSI_NAME_LEN - 1] = '\0';
1745         isc->isc_initiator_addr[ISCSI_ADDR_LEN - 1] = '\0';
1746         isc->isc_initiator_alias[ISCSI_ALIAS_LEN - 1] = '\0';
1747         isc->isc_target[ISCSI_NAME_LEN - 1] = '\0';
1748         isc->isc_target_addr[ISCSI_ADDR_LEN - 1] = '\0';
1749         isc->isc_user[ISCSI_NAME_LEN - 1] = '\0';
1750         isc->isc_secret[ISCSI_SECRET_LEN - 1] = '\0';
1751         isc->isc_mutual_user[ISCSI_NAME_LEN - 1] = '\0';
1752         isc->isc_mutual_secret[ISCSI_SECRET_LEN - 1] = '\0';
1753 }
1754
1755 static bool
1756 iscsi_valid_session_conf(const struct iscsi_session_conf *isc)
1757 {
1758
1759         if (isc->isc_initiator[0] == '\0') {
1760                 ISCSI_DEBUG("empty isc_initiator");
1761                 return (false);
1762         }
1763
1764         if (isc->isc_target_addr[0] == '\0') {
1765                 ISCSI_DEBUG("empty isc_target_addr");
1766                 return (false);
1767         }
1768
1769         if (isc->isc_discovery != 0 && isc->isc_target[0] != 0) {
1770                 ISCSI_DEBUG("non-empty isc_target for discovery session");
1771                 return (false);
1772         }
1773
1774         if (isc->isc_discovery == 0 && isc->isc_target[0] == 0) {
1775                 ISCSI_DEBUG("empty isc_target for non-discovery session");
1776                 return (false);
1777         }
1778
1779         return (true);
1780 }
1781
1782 static int
1783 iscsi_ioctl_session_add(struct iscsi_softc *sc, struct iscsi_session_add *isa)
1784 {
1785         struct iscsi_session *is;
1786         const struct iscsi_session *is2;
1787         int error;
1788
1789         iscsi_sanitize_session_conf(&isa->isa_conf);
1790         if (iscsi_valid_session_conf(&isa->isa_conf) == false)
1791                 return (EINVAL);
1792
1793         is = malloc(sizeof(*is), M_ISCSI, M_ZERO | M_WAITOK);
1794         memcpy(&is->is_conf, &isa->isa_conf, sizeof(is->is_conf));
1795
1796         sx_xlock(&sc->sc_lock);
1797
1798         /*
1799          * Prevent duplicates.
1800          */
1801         TAILQ_FOREACH(is2, &sc->sc_sessions, is_next) {
1802                 if (!!is->is_conf.isc_discovery !=
1803                     !!is2->is_conf.isc_discovery)
1804                         continue;
1805
1806                 if (strcmp(is->is_conf.isc_target_addr,
1807                     is2->is_conf.isc_target_addr) != 0)
1808                         continue;
1809
1810                 if (is->is_conf.isc_discovery == 0 &&
1811                     strcmp(is->is_conf.isc_target,
1812                     is2->is_conf.isc_target) != 0)
1813                         continue;
1814
1815                 sx_xunlock(&sc->sc_lock);
1816                 free(is, M_ISCSI);
1817                 return (EBUSY);
1818         }
1819
1820         is->is_conn = icl_new_conn(is->is_conf.isc_offload,
1821             is->is_conf.isc_iser, "iscsi", &is->is_lock);
1822         if (is->is_conn == NULL) {
1823                 sx_xunlock(&sc->sc_lock);
1824                 free(is, M_ISCSI);
1825                 return (EINVAL);
1826         }
1827         is->is_conn->ic_receive = iscsi_receive_callback;
1828         is->is_conn->ic_error = iscsi_error_callback;
1829         is->is_conn->ic_prv0 = is;
1830         TAILQ_INIT(&is->is_outstanding);
1831         STAILQ_INIT(&is->is_postponed);
1832         mtx_init(&is->is_lock, "iscsi_lock", NULL, MTX_DEF);
1833         cv_init(&is->is_maintenance_cv, "iscsi_mt");
1834 #ifdef ICL_KERNEL_PROXY
1835         cv_init(&is->is_login_cv, "iscsi_login");
1836 #endif
1837
1838         /*
1839          * Set some default values, from RFC 3720, section 12.
1840          *
1841          * These values are updated by the handoff IOCTL, but are
1842          * needed prior to the handoff to support sending the ISER
1843          * login PDU.
1844          */
1845         is->is_conn->ic_max_recv_data_segment_length = 8192;
1846         is->is_conn->ic_max_send_data_segment_length = 8192;
1847         is->is_max_burst_length = 262144;
1848         is->is_first_burst_length = 65536;
1849
1850         is->is_softc = sc;
1851         sc->sc_last_session_id++;
1852         is->is_id = sc->sc_last_session_id;
1853         is->is_isid[0] = 0x80; /* RFC 3720, 10.12.5: 10b, "Random" ISID. */
1854         arc4rand(&is->is_isid[1], 5, 0);
1855         is->is_tsih = 0;
1856         callout_init(&is->is_callout, 1);
1857
1858         error = kthread_add(iscsi_maintenance_thread, is, NULL, NULL, 0, 0, "iscsimt");
1859         if (error != 0) {
1860                 ISCSI_SESSION_WARN(is, "kthread_add(9) failed with error %d", error);
1861                 sx_xunlock(&sc->sc_lock);
1862                 return (error);
1863         }
1864
1865         callout_reset(&is->is_callout, 1 * hz, iscsi_callout, is);
1866         TAILQ_INSERT_TAIL(&sc->sc_sessions, is, is_next);
1867
1868         ISCSI_SESSION_LOCK(is);
1869         /*
1870          * Don't notify iscsid(8) if the session is disabled and it's not
1871          * a discovery session,
1872          */
1873         if (is->is_conf.isc_enable == 0 && is->is_conf.isc_discovery == 0) {
1874                 ISCSI_SESSION_UNLOCK(is);
1875                 sx_xunlock(&sc->sc_lock);
1876                 return (0);
1877         }
1878
1879         is->is_waiting_for_iscsid = true;
1880         strlcpy(is->is_reason, "Waiting for iscsid(8)", sizeof(is->is_reason));
1881         ISCSI_SESSION_UNLOCK(is);
1882         cv_signal(&sc->sc_cv);
1883         sx_xunlock(&sc->sc_lock);
1884         return (0);
1885 }
1886
1887 static bool
1888 iscsi_session_conf_matches(unsigned int id1, const struct iscsi_session_conf *c1,
1889     unsigned int id2, const struct iscsi_session_conf *c2)
1890 {
1891
1892         if (id2 != 0 && id2 != id1)
1893                 return (false);
1894         if (c2->isc_target[0] != '\0' &&
1895             strcmp(c1->isc_target, c2->isc_target) != 0)
1896                 return (false);
1897         if (c2->isc_target_addr[0] != '\0' &&
1898             strcmp(c1->isc_target_addr, c2->isc_target_addr) != 0)
1899                 return (false);
1900         return (true);
1901 }
1902
1903 static int
1904 iscsi_ioctl_session_remove(struct iscsi_softc *sc,
1905     struct iscsi_session_remove *isr)
1906 {
1907         struct iscsi_session *is, *tmp;
1908         bool found = false;
1909
1910         iscsi_sanitize_session_conf(&isr->isr_conf);
1911
1912         sx_xlock(&sc->sc_lock);
1913         TAILQ_FOREACH_SAFE(is, &sc->sc_sessions, is_next, tmp) {
1914                 ISCSI_SESSION_LOCK(is);
1915                 if (iscsi_session_conf_matches(is->is_id, &is->is_conf,
1916                     isr->isr_session_id, &isr->isr_conf)) {
1917                         found = true;
1918                         iscsi_session_logout(is);
1919                         iscsi_session_terminate(is);
1920                 }
1921                 ISCSI_SESSION_UNLOCK(is);
1922         }
1923         sx_xunlock(&sc->sc_lock);
1924
1925         if (!found)
1926                 return (ESRCH);
1927
1928         return (0);
1929 }
1930
1931 static int
1932 iscsi_ioctl_session_list(struct iscsi_softc *sc, struct iscsi_session_list *isl)
1933 {
1934         int error;
1935         unsigned int i = 0;
1936         struct iscsi_session *is;
1937         struct iscsi_session_state iss;
1938
1939         sx_slock(&sc->sc_lock);
1940         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1941                 if (i >= isl->isl_nentries) {
1942                         sx_sunlock(&sc->sc_lock);
1943                         return (EMSGSIZE);
1944                 }
1945                 memset(&iss, 0, sizeof(iss));
1946                 memcpy(&iss.iss_conf, &is->is_conf, sizeof(iss.iss_conf));
1947                 iss.iss_id = is->is_id;
1948                 strlcpy(iss.iss_target_alias, is->is_target_alias, sizeof(iss.iss_target_alias));
1949                 strlcpy(iss.iss_reason, is->is_reason, sizeof(iss.iss_reason));
1950                 strlcpy(iss.iss_offload, is->is_conn->ic_offload, sizeof(iss.iss_offload));
1951
1952                 if (is->is_conn->ic_header_crc32c)
1953                         iss.iss_header_digest = ISCSI_DIGEST_CRC32C;
1954                 else
1955                         iss.iss_header_digest = ISCSI_DIGEST_NONE;
1956
1957                 if (is->is_conn->ic_data_crc32c)
1958                         iss.iss_data_digest = ISCSI_DIGEST_CRC32C;
1959                 else
1960                         iss.iss_data_digest = ISCSI_DIGEST_NONE;
1961
1962                 iss.iss_max_send_data_segment_length =
1963                     is->is_conn->ic_max_send_data_segment_length;
1964                 iss.iss_max_recv_data_segment_length =
1965                     is->is_conn->ic_max_recv_data_segment_length;
1966                 iss.iss_max_burst_length = is->is_max_burst_length;
1967                 iss.iss_first_burst_length = is->is_first_burst_length;
1968                 iss.iss_immediate_data = is->is_immediate_data;
1969                 iss.iss_connected = is->is_connected;
1970
1971                 error = copyout(&iss, isl->isl_pstates + i, sizeof(iss));
1972                 if (error != 0) {
1973                         sx_sunlock(&sc->sc_lock);
1974                         return (error);
1975                 }
1976                 i++;
1977         }
1978         sx_sunlock(&sc->sc_lock);
1979
1980         isl->isl_nentries = i;
1981
1982         return (0);
1983 }
1984
1985 static int
1986 iscsi_ioctl_session_modify(struct iscsi_softc *sc,
1987     struct iscsi_session_modify *ism)
1988 {
1989         struct iscsi_session *is;
1990         const struct iscsi_session *is2;
1991
1992         iscsi_sanitize_session_conf(&ism->ism_conf);
1993         if (iscsi_valid_session_conf(&ism->ism_conf) == false)
1994                 return (EINVAL);
1995
1996         sx_xlock(&sc->sc_lock);
1997         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1998                 ISCSI_SESSION_LOCK(is);
1999                 if (is->is_id == ism->ism_session_id) {
2000                         /* Note that the session remains locked. */
2001                         break;
2002                 }
2003                 ISCSI_SESSION_UNLOCK(is);
2004         }
2005         if (is == NULL) {
2006                 sx_xunlock(&sc->sc_lock);
2007                 return (ESRCH);
2008         }
2009
2010         /*
2011          * Prevent duplicates.
2012          */
2013         TAILQ_FOREACH(is2, &sc->sc_sessions, is_next) {
2014                 if (is == is2)
2015                         continue;
2016
2017                 if (!!ism->ism_conf.isc_discovery !=
2018                     !!is2->is_conf.isc_discovery)
2019                         continue;
2020
2021                 if (strcmp(ism->ism_conf.isc_target_addr,
2022                     is2->is_conf.isc_target_addr) != 0)
2023                         continue;
2024
2025                 if (ism->ism_conf.isc_discovery == 0 &&
2026                     strcmp(ism->ism_conf.isc_target,
2027                     is2->is_conf.isc_target) != 0)
2028                         continue;
2029
2030                 ISCSI_SESSION_UNLOCK(is);
2031                 sx_xunlock(&sc->sc_lock);
2032                 return (EBUSY);
2033         }
2034
2035         sx_xunlock(&sc->sc_lock);
2036
2037         memcpy(&is->is_conf, &ism->ism_conf, sizeof(is->is_conf));
2038         ISCSI_SESSION_UNLOCK(is);
2039
2040         iscsi_session_reconnect(is);
2041
2042         return (0);
2043 }
2044
2045 static int
2046 iscsi_ioctl(struct cdev *dev, u_long cmd, caddr_t arg, int mode,
2047     struct thread *td)
2048 {
2049         struct iscsi_softc *sc;
2050
2051         sc = dev->si_drv1;
2052
2053         switch (cmd) {
2054         case ISCSIDWAIT:
2055                 return (iscsi_ioctl_daemon_wait(sc,
2056                     (struct iscsi_daemon_request *)arg));
2057         case ISCSIDHANDOFF:
2058                 return (iscsi_ioctl_daemon_handoff(sc,
2059                     (struct iscsi_daemon_handoff *)arg));
2060         case ISCSIDFAIL:
2061                 return (iscsi_ioctl_daemon_fail(sc,
2062                     (struct iscsi_daemon_fail *)arg));
2063 #ifdef ICL_KERNEL_PROXY
2064         case ISCSIDCONNECT:
2065                 return (iscsi_ioctl_daemon_connect(sc,
2066                     (struct iscsi_daemon_connect *)arg));
2067         case ISCSIDSEND:
2068                 return (iscsi_ioctl_daemon_send(sc,
2069                     (struct iscsi_daemon_send *)arg));
2070         case ISCSIDRECEIVE:
2071                 return (iscsi_ioctl_daemon_receive(sc,
2072                     (struct iscsi_daemon_receive *)arg));
2073 #endif /* ICL_KERNEL_PROXY */
2074         case ISCSISADD:
2075                 return (iscsi_ioctl_session_add(sc,
2076                     (struct iscsi_session_add *)arg));
2077         case ISCSISREMOVE:
2078                 return (iscsi_ioctl_session_remove(sc,
2079                     (struct iscsi_session_remove *)arg));
2080         case ISCSISLIST:
2081                 return (iscsi_ioctl_session_list(sc,
2082                     (struct iscsi_session_list *)arg));
2083         case ISCSISMODIFY:
2084                 return (iscsi_ioctl_session_modify(sc,
2085                     (struct iscsi_session_modify *)arg));
2086         default:
2087                 return (EINVAL);
2088         }
2089 }
2090
2091 static struct iscsi_outstanding *
2092 iscsi_outstanding_find(struct iscsi_session *is, uint32_t initiator_task_tag)
2093 {
2094         struct iscsi_outstanding *io;
2095
2096         ISCSI_SESSION_LOCK_ASSERT(is);
2097
2098         TAILQ_FOREACH(io, &is->is_outstanding, io_next) {
2099                 if (io->io_initiator_task_tag == initiator_task_tag)
2100                         return (io);
2101         }
2102         return (NULL);
2103 }
2104
2105 static struct iscsi_outstanding *
2106 iscsi_outstanding_find_ccb(struct iscsi_session *is, union ccb *ccb)
2107 {
2108         struct iscsi_outstanding *io;
2109
2110         ISCSI_SESSION_LOCK_ASSERT(is);
2111
2112         TAILQ_FOREACH(io, &is->is_outstanding, io_next) {
2113                 if (io->io_ccb == ccb)
2114                         return (io);
2115         }
2116         return (NULL);
2117 }
2118
2119 static struct iscsi_outstanding *
2120 iscsi_outstanding_add(struct iscsi_session *is, struct icl_pdu *request,
2121     union ccb *ccb, uint32_t *initiator_task_tagp)
2122 {
2123         struct iscsi_outstanding *io;
2124         int error;
2125
2126         ISCSI_SESSION_LOCK_ASSERT(is);
2127
2128         io = uma_zalloc(iscsi_outstanding_zone, M_NOWAIT | M_ZERO);
2129         if (io == NULL) {
2130                 ISCSI_SESSION_WARN(is, "failed to allocate %zd bytes",
2131                     sizeof(*io));
2132                 return (NULL);
2133         }
2134
2135         error = icl_conn_task_setup(is->is_conn, request, &ccb->csio,
2136             initiator_task_tagp, &io->io_icl_prv);
2137         if (error != 0) {
2138                 ISCSI_SESSION_WARN(is,
2139                     "icl_conn_task_setup() failed with error %d", error);
2140                 uma_zfree(iscsi_outstanding_zone, io);
2141                 return (NULL);
2142         }
2143
2144         KASSERT(iscsi_outstanding_find(is, *initiator_task_tagp) == NULL,
2145             ("initiator_task_tag 0x%x already added", *initiator_task_tagp));
2146
2147         io->io_initiator_task_tag = *initiator_task_tagp;
2148         io->io_ccb = ccb;
2149         TAILQ_INSERT_TAIL(&is->is_outstanding, io, io_next);
2150         return (io);
2151 }
2152
2153 static void
2154 iscsi_outstanding_remove(struct iscsi_session *is, struct iscsi_outstanding *io)
2155 {
2156
2157         ISCSI_SESSION_LOCK_ASSERT(is);
2158
2159         icl_conn_task_done(is->is_conn, io->io_icl_prv);
2160         TAILQ_REMOVE(&is->is_outstanding, io, io_next);
2161         uma_zfree(iscsi_outstanding_zone, io);
2162 }
2163
2164 static void
2165 iscsi_action_abort(struct iscsi_session *is, union ccb *ccb)
2166 {
2167         struct icl_pdu *request;
2168         struct iscsi_bhs_task_management_request *bhstmr;
2169         struct ccb_abort *cab = &ccb->cab;
2170         struct iscsi_outstanding *io, *aio;
2171         uint32_t initiator_task_tag;
2172
2173         ISCSI_SESSION_LOCK_ASSERT(is);
2174
2175 #if 0
2176         KASSERT(is->is_login_phase == false, ("%s called during Login Phase", __func__));
2177 #else
2178         if (is->is_login_phase) {
2179                 ccb->ccb_h.status = CAM_REQ_ABORTED;
2180                 xpt_done(ccb);
2181                 return;
2182         }
2183 #endif
2184
2185         aio = iscsi_outstanding_find_ccb(is, cab->abort_ccb);
2186         if (aio == NULL) {
2187                 ccb->ccb_h.status = CAM_REQ_CMP;
2188                 xpt_done(ccb);
2189                 return;
2190         }
2191
2192         request = icl_pdu_new(is->is_conn, M_NOWAIT);
2193         if (request == NULL) {
2194                 ccb->ccb_h.status = CAM_RESRC_UNAVAIL;
2195                 xpt_done(ccb);
2196                 return;
2197         }
2198
2199         initiator_task_tag = is->is_initiator_task_tag++;
2200         if (initiator_task_tag == 0xffffffff)
2201                 initiator_task_tag = is->is_initiator_task_tag++;
2202
2203         io = iscsi_outstanding_add(is, request, NULL, &initiator_task_tag);
2204         if (io == NULL) {
2205                 icl_pdu_free(request);
2206                 ccb->ccb_h.status = CAM_RESRC_UNAVAIL;
2207                 xpt_done(ccb);
2208                 return;
2209         }
2210         io->io_referenced_task_tag = aio->io_initiator_task_tag;
2211
2212         bhstmr = (struct iscsi_bhs_task_management_request *)request->ip_bhs;
2213         bhstmr->bhstmr_opcode = ISCSI_BHS_OPCODE_TASK_REQUEST;
2214         bhstmr->bhstmr_function = 0x80 | BHSTMR_FUNCTION_ABORT_TASK;
2215         bhstmr->bhstmr_lun = htobe64(CAM_EXTLUN_BYTE_SWIZZLE(ccb->ccb_h.target_lun));
2216         bhstmr->bhstmr_initiator_task_tag = initiator_task_tag;
2217         bhstmr->bhstmr_referenced_task_tag = aio->io_initiator_task_tag;
2218
2219         iscsi_pdu_queue_locked(request);
2220 }
2221
2222 static void
2223 iscsi_action_scsiio(struct iscsi_session *is, union ccb *ccb)
2224 {
2225         struct icl_pdu *request;
2226         struct iscsi_bhs_scsi_command *bhssc;
2227         struct ccb_scsiio *csio;
2228         struct iscsi_outstanding *io;
2229         size_t len;
2230         uint32_t initiator_task_tag;
2231         int error;
2232
2233         ISCSI_SESSION_LOCK_ASSERT(is);
2234
2235 #if 0
2236         KASSERT(is->is_login_phase == false, ("%s called during Login Phase", __func__));
2237 #else
2238         if (is->is_login_phase) {
2239                 ISCSI_SESSION_DEBUG(is, "called during login phase");
2240                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
2241                         xpt_freeze_devq(ccb->ccb_h.path, 1);
2242                         ISCSI_SESSION_DEBUG(is, "freezing devq");
2243                 }
2244                 ccb->ccb_h.status = CAM_REQ_ABORTED | CAM_DEV_QFRZN;
2245                 xpt_done(ccb);
2246                 return;
2247         }
2248 #endif
2249
2250         request = icl_pdu_new(is->is_conn, M_NOWAIT);
2251         if (request == NULL) {
2252                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
2253                         xpt_freeze_devq(ccb->ccb_h.path, 1);
2254                         ISCSI_SESSION_DEBUG(is, "freezing devq");
2255                 }
2256                 ccb->ccb_h.status = CAM_RESRC_UNAVAIL | CAM_DEV_QFRZN;
2257                 xpt_done(ccb);
2258                 return;
2259         }
2260
2261         initiator_task_tag = is->is_initiator_task_tag++;
2262         if (initiator_task_tag == 0xffffffff)
2263                 initiator_task_tag = is->is_initiator_task_tag++;
2264
2265         io = iscsi_outstanding_add(is, request, ccb, &initiator_task_tag);
2266         if (io == NULL) {
2267                 icl_pdu_free(request);
2268                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
2269                         xpt_freeze_devq(ccb->ccb_h.path, 1);
2270                         ISCSI_SESSION_DEBUG(is, "freezing devq");
2271                 }
2272                 ccb->ccb_h.status = CAM_RESRC_UNAVAIL | CAM_DEV_QFRZN;
2273                 xpt_done(ccb);
2274                 return;
2275         }
2276
2277         csio = &ccb->csio;
2278         bhssc = (struct iscsi_bhs_scsi_command *)request->ip_bhs;
2279         bhssc->bhssc_opcode = ISCSI_BHS_OPCODE_SCSI_COMMAND;
2280         bhssc->bhssc_flags |= BHSSC_FLAGS_F;
2281         switch (csio->ccb_h.flags & CAM_DIR_MASK) {
2282         case CAM_DIR_IN:
2283                 bhssc->bhssc_flags |= BHSSC_FLAGS_R;
2284                 break;
2285         case CAM_DIR_OUT:
2286                 bhssc->bhssc_flags |= BHSSC_FLAGS_W;
2287                 break;
2288         }
2289
2290         if ((ccb->ccb_h.flags & CAM_TAG_ACTION_VALID) != 0) {
2291                 switch (csio->tag_action) {
2292                 case MSG_HEAD_OF_Q_TAG:
2293                         bhssc->bhssc_flags |= BHSSC_FLAGS_ATTR_HOQ;
2294                         break;
2295                 case MSG_ORDERED_Q_TAG:
2296                         bhssc->bhssc_flags |= BHSSC_FLAGS_ATTR_ORDERED;
2297                         break;
2298                 case MSG_ACA_TASK:
2299                         bhssc->bhssc_flags |= BHSSC_FLAGS_ATTR_ACA;
2300                         break;
2301                 case MSG_SIMPLE_Q_TAG:
2302                 default:
2303                         bhssc->bhssc_flags |= BHSSC_FLAGS_ATTR_SIMPLE;
2304                         break;
2305                 }
2306         } else
2307                 bhssc->bhssc_flags |= BHSSC_FLAGS_ATTR_UNTAGGED;
2308
2309         if (is->is_protocol_level >= 2) {
2310                 bhssc->bhssc_pri = (csio->priority << BHSSC_PRI_SHIFT) &
2311                     BHSSC_PRI_MASK;
2312         }
2313
2314         bhssc->bhssc_lun = htobe64(CAM_EXTLUN_BYTE_SWIZZLE(ccb->ccb_h.target_lun));
2315         bhssc->bhssc_initiator_task_tag = initiator_task_tag;
2316         bhssc->bhssc_expected_data_transfer_length = htonl(csio->dxfer_len);
2317         KASSERT(csio->cdb_len <= sizeof(bhssc->bhssc_cdb),
2318             ("unsupported CDB size %zd", (size_t)csio->cdb_len));
2319
2320         if (csio->ccb_h.flags & CAM_CDB_POINTER)
2321                 memcpy(&bhssc->bhssc_cdb, csio->cdb_io.cdb_ptr, csio->cdb_len);
2322         else
2323                 memcpy(&bhssc->bhssc_cdb, csio->cdb_io.cdb_bytes, csio->cdb_len);
2324
2325         if (is->is_immediate_data &&
2326             (csio->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_OUT) {
2327                 len = csio->dxfer_len;
2328                 //ISCSI_SESSION_DEBUG(is, "adding %zd of immediate data", len);
2329                 if (len > is->is_first_burst_length) {
2330                         ISCSI_SESSION_DEBUG(is, "len %zd -> %d", len, is->is_first_burst_length);
2331                         len = is->is_first_burst_length;
2332                 }
2333                 if (len > is->is_conn->ic_max_send_data_segment_length) {
2334                         ISCSI_SESSION_DEBUG(is, "len %zd -> %d", len,
2335                             is->is_conn->ic_max_send_data_segment_length);
2336                         len = is->is_conn->ic_max_send_data_segment_length;
2337                 }
2338
2339                 error = icl_pdu_append_data(request, csio->data_ptr, len, M_NOWAIT);
2340                 if (error != 0) {
2341                         iscsi_outstanding_remove(is, io);
2342                         icl_pdu_free(request);
2343                         if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
2344                                 xpt_freeze_devq(ccb->ccb_h.path, 1);
2345                                 ISCSI_SESSION_DEBUG(is, "freezing devq");
2346                         }
2347                         ccb->ccb_h.status = CAM_RESRC_UNAVAIL | CAM_DEV_QFRZN;
2348                         xpt_done(ccb);
2349                         return;
2350                 }
2351         }
2352         iscsi_pdu_queue_locked(request);
2353 }
2354
2355 static void
2356 iscsi_action(struct cam_sim *sim, union ccb *ccb)
2357 {
2358         struct iscsi_session *is;
2359
2360         is = cam_sim_softc(sim);
2361
2362         ISCSI_SESSION_LOCK_ASSERT(is);
2363
2364         if (is->is_terminating ||
2365             (is->is_connected == false && fail_on_disconnection)) {
2366                 ccb->ccb_h.status = CAM_DEV_NOT_THERE;
2367                 xpt_done(ccb);
2368                 return;
2369         }
2370
2371         /*
2372          * Make sure CAM doesn't sneak in a CCB just after freezing the queue.
2373          */
2374         if (is->is_simq_frozen == true) {
2375                 ccb->ccb_h.status &= ~(CAM_SIM_QUEUED | CAM_STATUS_MASK);
2376                 ccb->ccb_h.status |= CAM_REQUEUE_REQ;
2377                 /* Don't freeze the devq - the SIM queue is already frozen. */
2378                 xpt_done(ccb);
2379                 return;
2380         }
2381
2382         switch (ccb->ccb_h.func_code) {
2383         case XPT_PATH_INQ:
2384         {
2385                 struct ccb_pathinq *cpi = &ccb->cpi;
2386
2387                 cpi->version_num = 1;
2388                 cpi->hba_inquiry = PI_TAG_ABLE;
2389                 cpi->target_sprt = 0;
2390                 cpi->hba_misc = PIM_EXTLUNS;
2391                 /*
2392                  * XXX: It shouldn't ever be NULL; this could be turned
2393                  *      into a KASSERT eventually.
2394                  */
2395                 if (is->is_conn == NULL)
2396                         ISCSI_WARN("NULL conn");
2397                 else if (is->is_conn->ic_unmapped)
2398                         cpi->hba_misc |= PIM_UNMAPPED;
2399                 cpi->hba_eng_cnt = 0;
2400                 cpi->max_target = 0;
2401                 /*
2402                  * Note that the variable below is only relevant for targets
2403                  * that don't claim compliance with anything above SPC2, which
2404                  * means they don't support REPORT_LUNS.
2405                  */
2406                 cpi->max_lun = 255;
2407                 cpi->initiator_id = ~0;
2408                 strlcpy(cpi->sim_vid, "FreeBSD", SIM_IDLEN);
2409                 strlcpy(cpi->hba_vid, "iSCSI", HBA_IDLEN);
2410                 strlcpy(cpi->dev_name, cam_sim_name(sim), DEV_IDLEN);
2411                 cpi->unit_number = cam_sim_unit(sim);
2412                 cpi->bus_id = cam_sim_bus(sim);
2413                 cpi->base_transfer_speed = 150000; /* XXX */
2414                 cpi->transport = XPORT_ISCSI;
2415                 cpi->transport_version = 0;
2416                 cpi->protocol = PROTO_SCSI;
2417                 cpi->protocol_version = SCSI_REV_SPC3;
2418                 cpi->maxio = maxphys;
2419                 cpi->ccb_h.status = CAM_REQ_CMP;
2420                 break;
2421         }
2422         case XPT_GET_TRAN_SETTINGS:
2423         {
2424                 struct ccb_trans_settings       *cts;
2425                 struct ccb_trans_settings_scsi  *scsi;
2426
2427                 cts = &ccb->cts;
2428                 scsi = &cts->proto_specific.scsi;
2429
2430                 cts->protocol = PROTO_SCSI;
2431                 cts->protocol_version = SCSI_REV_SPC3;
2432                 cts->transport = XPORT_ISCSI;
2433                 cts->transport_version = 0;
2434                 scsi->valid = CTS_SCSI_VALID_TQ;
2435                 scsi->flags = CTS_SCSI_FLAGS_TAG_ENB;
2436                 cts->ccb_h.status = CAM_REQ_CMP;
2437                 break;
2438         }
2439         case XPT_CALC_GEOMETRY:
2440                 cam_calc_geometry(&ccb->ccg, /*extended*/1);
2441                 ccb->ccb_h.status = CAM_REQ_CMP;
2442                 break;
2443 #if 0
2444         /*
2445          * XXX: What's the point?
2446          */
2447         case XPT_RESET_BUS:
2448         case XPT_TERM_IO:
2449                 ISCSI_SESSION_DEBUG(is, "faking success for reset, abort, or term_io");
2450                 ccb->ccb_h.status = CAM_REQ_CMP;
2451                 break;
2452 #endif
2453         case XPT_ABORT:
2454                 iscsi_action_abort(is, ccb);
2455                 return;
2456         case XPT_SCSI_IO:
2457                 iscsi_action_scsiio(is, ccb);
2458                 return;
2459         default:
2460 #if 0
2461                 ISCSI_SESSION_DEBUG(is, "got unsupported code 0x%x", ccb->ccb_h.func_code);
2462 #endif
2463                 ccb->ccb_h.status = CAM_FUNC_NOTAVAIL;
2464                 break;
2465         }
2466         xpt_done(ccb);
2467 }
2468
2469 static void
2470 iscsi_terminate_sessions(struct iscsi_softc *sc)
2471 {
2472         struct iscsi_session *is;
2473
2474         sx_slock(&sc->sc_lock);
2475         TAILQ_FOREACH(is, &sc->sc_sessions, is_next)
2476                 iscsi_session_terminate(is);
2477         while(!TAILQ_EMPTY(&sc->sc_sessions)) {
2478                 ISCSI_DEBUG("waiting for sessions to terminate");
2479                 cv_wait(&sc->sc_cv, &sc->sc_lock);
2480         }
2481         ISCSI_DEBUG("all sessions terminated");
2482         sx_sunlock(&sc->sc_lock);
2483 }
2484
2485 static void
2486 iscsi_shutdown_pre(struct iscsi_softc *sc)
2487 {
2488         struct iscsi_session *is;
2489
2490         if (!fail_on_shutdown)
2491                 return;
2492
2493         /*
2494          * If we have any sessions waiting for reconnection, request
2495          * maintenance thread to fail them immediately instead of waiting
2496          * for reconnect timeout.
2497          *
2498          * This prevents LUNs with mounted filesystems that are supported
2499          * by disconnected iSCSI sessions from hanging, however it will
2500          * fail all queued BIOs.
2501          */
2502         ISCSI_DEBUG("forcing failing all disconnected sessions due to shutdown");
2503
2504         fail_on_disconnection = 1;
2505
2506         sx_slock(&sc->sc_lock);
2507         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
2508                 ISCSI_SESSION_LOCK(is);
2509                 if (!is->is_connected) {
2510                         ISCSI_SESSION_DEBUG(is, "force failing disconnected session early");
2511                         iscsi_session_reconnect(is);
2512                 }
2513                 ISCSI_SESSION_UNLOCK(is);
2514         }
2515         sx_sunlock(&sc->sc_lock);
2516 }
2517
2518 static void
2519 iscsi_shutdown_post(struct iscsi_softc *sc)
2520 {
2521
2522         if (!KERNEL_PANICKED()) {
2523                 ISCSI_DEBUG("removing all sessions due to shutdown");
2524                 iscsi_terminate_sessions(sc);
2525         }
2526 }
2527
2528 static int
2529 iscsi_load(void)
2530 {
2531         int error;
2532
2533         sc = malloc(sizeof(*sc), M_ISCSI, M_ZERO | M_WAITOK);
2534         sx_init(&sc->sc_lock, "iscsi");
2535         TAILQ_INIT(&sc->sc_sessions);
2536         cv_init(&sc->sc_cv, "iscsi_cv");
2537
2538         iscsi_outstanding_zone = uma_zcreate("iscsi_outstanding",
2539             sizeof(struct iscsi_outstanding), NULL, NULL, NULL, NULL,
2540             UMA_ALIGN_PTR, 0);
2541
2542         error = make_dev_p(MAKEDEV_CHECKNAME, &sc->sc_cdev, &iscsi_cdevsw,
2543             NULL, UID_ROOT, GID_WHEEL, 0600, "iscsi");
2544         if (error != 0) {
2545                 ISCSI_WARN("failed to create device node, error %d", error);
2546                 return (error);
2547         }
2548         sc->sc_cdev->si_drv1 = sc;
2549
2550         sc->sc_shutdown_pre_eh = EVENTHANDLER_REGISTER(shutdown_pre_sync,
2551             iscsi_shutdown_pre, sc, SHUTDOWN_PRI_FIRST);
2552         /*
2553          * shutdown_post_sync needs to run after filesystem shutdown and before
2554          * CAM shutdown - otherwise when rebooting with an iSCSI session that is
2555          * disconnected but has outstanding requests, dashutdown() will hang on
2556          * cam_periph_runccb().
2557          */
2558         sc->sc_shutdown_post_eh = EVENTHANDLER_REGISTER(shutdown_post_sync,
2559             iscsi_shutdown_post, sc, SHUTDOWN_PRI_DEFAULT - 1);
2560
2561         return (0);
2562 }
2563
2564 static int
2565 iscsi_unload(void)
2566 {
2567
2568         /* Awaken any threads asleep in iscsi_ioctl(). */
2569         sx_xlock(&sc->sc_lock);
2570         sc->sc_unloading = true;
2571         cv_signal(&sc->sc_cv);
2572         sx_xunlock(&sc->sc_lock);
2573
2574         if (sc->sc_cdev != NULL) {
2575                 ISCSI_DEBUG("removing device node");
2576                 destroy_dev(sc->sc_cdev);
2577                 ISCSI_DEBUG("device node removed");
2578         }
2579
2580         if (sc->sc_shutdown_pre_eh != NULL)
2581                 EVENTHANDLER_DEREGISTER(shutdown_pre_sync, sc->sc_shutdown_pre_eh);
2582         if (sc->sc_shutdown_post_eh != NULL)
2583                 EVENTHANDLER_DEREGISTER(shutdown_post_sync, sc->sc_shutdown_post_eh);
2584
2585         iscsi_terminate_sessions(sc);
2586
2587         uma_zdestroy(iscsi_outstanding_zone);
2588         sx_destroy(&sc->sc_lock);
2589         cv_destroy(&sc->sc_cv);
2590         free(sc, M_ISCSI);
2591         return (0);
2592 }
2593
2594 static int
2595 iscsi_quiesce(void)
2596 {
2597         sx_slock(&sc->sc_lock);
2598         if (!TAILQ_EMPTY(&sc->sc_sessions)) {
2599                 sx_sunlock(&sc->sc_lock);
2600                 return (EBUSY);
2601         }
2602         sx_sunlock(&sc->sc_lock);
2603         return (0);
2604 }
2605
2606 static int
2607 iscsi_modevent(module_t mod, int what, void *arg)
2608 {
2609         int error;
2610
2611         switch (what) {
2612         case MOD_LOAD:
2613                 error = iscsi_load();
2614                 break;
2615         case MOD_UNLOAD:
2616                 error = iscsi_unload();
2617                 break;
2618         case MOD_QUIESCE:
2619                 error = iscsi_quiesce();
2620                 break;
2621         default:
2622                 error = EINVAL;
2623                 break;
2624         }
2625         return (error);
2626 }
2627
2628 moduledata_t iscsi_data = {
2629         "iscsi",
2630         iscsi_modevent,
2631         0
2632 };
2633
2634 DECLARE_MODULE(iscsi, iscsi_data, SI_SUB_DRIVERS, SI_ORDER_MIDDLE);
2635 MODULE_DEPEND(iscsi, cam, 1, 1, 1);
2636 MODULE_DEPEND(iscsi, icl, 1, 1, 1);