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         sbintime_t sbt, pr;
 550
 551         is = context;
 552
 553         ISCSI_SESSION_LOCK(is);
 554         if (is->is_terminating) {
 555                 ISCSI_SESSION_UNLOCK(is);
 556                 return;
 557         }
 558
 559         sbt = mstosbt(995);
 560         pr  = mstosbt(10);
 561         callout_schedule_sbt(&is->is_callout, sbt, pr, 0);
 562
 563         if (is->is_conf.isc_enable == 0)
 564                 goto out;
 565
 566         is->is_timeout++;
 567
 568         if (is->is_waiting_for_iscsid) {
 569                 if (iscsid_timeout > 0 && is->is_timeout > iscsid_timeout) {
 570                         ISCSI_SESSION_WARN(is, "timed out waiting for iscsid(8) "
 571                             "for %d seconds; reconnecting",
 572                             is->is_timeout);
 573                         reconnect_needed = true;
 574                 }
 575                 goto out;
 576         }
 577
 578         if (is->is_login_phase) {
 579                 if (login_timeout > 0 && is->is_timeout > login_timeout) {
 580                         ISCSI_SESSION_WARN(is, "login timed out after %d seconds; "
 581                             "reconnecting", is->is_timeout);
 582                         reconnect_needed = true;
 583                 }
 584                 goto out;
 585         }
 586
 587         if (ping_timeout <= 0) {
 588                 /*
 589                  * Pings are disabled.  Don't send NOP-Out in this case.
 590                  * Reset the timeout, to avoid triggering reconnection,
 591                  * should the user decide to reenable them.
 592                  */
 593                 is->is_timeout = 0;
 594                 goto out;
 595         }
 596
 597         if (is->is_timeout >= ping_timeout) {
 598                 ISCSI_SESSION_WARN(is, "no ping reply (NOP-In) after %d seconds; "
 599                     "reconnecting", ping_timeout);
 600                 reconnect_needed = true;
 601                 goto out;
 602         }
 603
 604         ISCSI_SESSION_UNLOCK(is);
 605
 606         /*
 607          * If the ping was reset less than one second ago - which means
 608          * that we've received some PDU during the last second - assume
 609          * the traffic flows correctly and don't bother sending a NOP-Out.
 610          *
 611          * (It's 2 - one for one second, and one for incrementing is_timeout
 612          * earlier in this routine.)
 613          */
 614         if (is->is_timeout < 2)
 615                 return;
 616
 617         request = icl_pdu_new(is->is_conn, M_NOWAIT);
 618         if (request == NULL) {
 619                 ISCSI_SESSION_WARN(is, "failed to allocate PDU");
 620                 return;
 621         }
 622         bhsno = (struct iscsi_bhs_nop_out *)request->ip_bhs;
 623         bhsno->bhsno_opcode = ISCSI_BHS_OPCODE_NOP_OUT |
 624             ISCSI_BHS_OPCODE_IMMEDIATE;
 625         bhsno->bhsno_flags = 0x80;
 626         bhsno->bhsno_target_transfer_tag = 0xffffffff;
 627         iscsi_pdu_queue(request);
 628         return;
 629
 630 out:
 631         if (is->is_terminating) {
 632                 ISCSI_SESSION_UNLOCK(is);
 633                 return;
 634         }
 635
 636         ISCSI_SESSION_UNLOCK(is);
 637
 638         if (reconnect_needed)
 639                 iscsi_session_reconnect(is);
 640 }
 641
 642 static void
 643 iscsi_pdu_update_statsn(const struct icl_pdu *response)
 644 {
 645         const struct iscsi_bhs_data_in *bhsdi;
 646         struct iscsi_session *is;
 647         uint32_t expcmdsn, maxcmdsn, statsn;
 648
 649         is = PDU_SESSION(response);
 650
 651         ISCSI_SESSION_LOCK_ASSERT(is);
 652
 653         /*
 654          * We're only using fields common for all the response
 655          * (target -> initiator) PDUs.
 656          */
 657         bhsdi = (const struct iscsi_bhs_data_in *)response->ip_bhs;
 658         /*
 659          * Ok, I lied.  In case of Data-In, "The fields StatSN, Status,
 660          * and Residual Count only have meaningful content if the S bit
 661          * is set to 1", so we also need to check the bit specific for
 662          * Data-In PDU.
 663          */
 664         if (bhsdi->bhsdi_opcode != ISCSI_BHS_OPCODE_SCSI_DATA_IN ||
 665             (bhsdi->bhsdi_flags & BHSDI_FLAGS_S) != 0) {
 666                 statsn = ntohl(bhsdi->bhsdi_statsn);
 667                 if (statsn != is->is_statsn && statsn != (is->is_statsn + 1)) {
 668                         /* XXX: This is normal situation for MCS */
 669                         ISCSI_SESSION_WARN(is, "PDU 0x%x StatSN %u != "
 670                             "session ExpStatSN %u (or + 1); reconnecting",
 671                             bhsdi->bhsdi_opcode, statsn, is->is_statsn);
 672                         iscsi_session_reconnect(is);
 673                 }
 674                 if (ISCSI_SNGT(statsn, is->is_statsn))
 675                         is->is_statsn = statsn;
 676         }
 677
 678         expcmdsn = ntohl(bhsdi->bhsdi_expcmdsn);
 679         maxcmdsn = ntohl(bhsdi->bhsdi_maxcmdsn);
 680
 681         if (ISCSI_SNLT(maxcmdsn + 1, expcmdsn)) {
 682                 ISCSI_SESSION_DEBUG(is,
 683                     "PDU MaxCmdSN %u + 1 < PDU ExpCmdSN %u; ignoring",
 684                     maxcmdsn, expcmdsn);
 685         } else {
 686                 if (ISCSI_SNGT(maxcmdsn, is->is_maxcmdsn)) {
 687                         is->is_maxcmdsn = maxcmdsn;
 688
 689                         /*
 690                          * Command window increased; kick the maintanance thread
 691                          * to send out postponed commands.
 692                          */
 693                         if (!STAILQ_EMPTY(&is->is_postponed))
 694                                 cv_signal(&is->is_maintenance_cv);
 695                 } else if (ISCSI_SNLT(maxcmdsn, is->is_maxcmdsn)) {
 696                         /* XXX: This is normal situation for MCS */
 697                         ISCSI_SESSION_DEBUG(is,
 698                             "PDU MaxCmdSN %u < session MaxCmdSN %u; ignoring",
 699                             maxcmdsn, is->is_maxcmdsn);
 700                 }
 701
 702                 if (ISCSI_SNGT(expcmdsn, is->is_expcmdsn)) {
 703                         is->is_expcmdsn = expcmdsn;
 704                 } else if (ISCSI_SNLT(expcmdsn, is->is_expcmdsn)) {
 705                         /* XXX: This is normal situation for MCS */
 706                         ISCSI_SESSION_DEBUG(is,
 707                             "PDU ExpCmdSN %u < session ExpCmdSN %u; ignoring",
 708                             expcmdsn, is->is_expcmdsn);
 709                 }
 710         }
 711
 712         /*
 713          * Every incoming PDU - not just NOP-In - resets the ping timer.
 714          * The purpose of the timeout is to reset the connection when it stalls;
 715          * we don't want this to happen when NOP-In or NOP-Out ends up delayed
 716          * in some queue.
 717          */
 718         is->is_timeout = 0;
 719 }
 720
 721 static void
 722 iscsi_receive_callback(struct icl_pdu *response)
 723 {
 724         struct iscsi_session *is;
 725
 726         is = PDU_SESSION(response);
 727
 728         ISCSI_SESSION_LOCK(is);
 729
 730         iscsi_pdu_update_statsn(response);
 731
 732 #ifdef ICL_KERNEL_PROXY
 733         if (is->is_login_phase) {
 734                 if (is->is_login_pdu == NULL)
 735                         is->is_login_pdu = response;
 736                 else
 737                         icl_pdu_free(response);
 738                 ISCSI_SESSION_UNLOCK(is);
 739                 cv_signal(&is->is_login_cv);
 740                 return;
 741         }
 742 #endif
 743
 744         /*
 745          * The handling routine is responsible for freeing the PDU
 746          * when it's no longer needed.
 747          */
 748         switch (response->ip_bhs->bhs_opcode) {
 749         case ISCSI_BHS_OPCODE_NOP_IN:
 750                 iscsi_pdu_handle_nop_in(response);
 751                 ISCSI_SESSION_UNLOCK(is);
 752                 break;
 753         case ISCSI_BHS_OPCODE_SCSI_RESPONSE:
 754                 iscsi_pdu_handle_scsi_response(response);
 755                 /* Session lock dropped inside. */
 756                 ISCSI_SESSION_LOCK_ASSERT_NOT(is);
 757                 break;
 758         case ISCSI_BHS_OPCODE_TASK_RESPONSE:
 759                 iscsi_pdu_handle_task_response(response);
 760                 ISCSI_SESSION_UNLOCK(is);
 761                 break;
 762         case ISCSI_BHS_OPCODE_SCSI_DATA_IN:
 763                 iscsi_pdu_handle_data_in(response);
 764                 /* Session lock dropped inside. */
 765                 ISCSI_SESSION_LOCK_ASSERT_NOT(is);
 766                 break;
 767         case ISCSI_BHS_OPCODE_LOGOUT_RESPONSE:
 768                 iscsi_pdu_handle_logout_response(response);
 769                 ISCSI_SESSION_UNLOCK(is);
 770                 break;
 771         case ISCSI_BHS_OPCODE_R2T:
 772                 iscsi_pdu_handle_r2t(response);
 773                 ISCSI_SESSION_UNLOCK(is);
 774                 break;
 775         case ISCSI_BHS_OPCODE_ASYNC_MESSAGE:
 776                 iscsi_pdu_handle_async_message(response);
 777                 ISCSI_SESSION_UNLOCK(is);
 778                 break;
 779         case ISCSI_BHS_OPCODE_REJECT:
 780                 iscsi_pdu_handle_reject(response);
 781                 ISCSI_SESSION_UNLOCK(is);
 782                 break;
 783         default:
 784                 ISCSI_SESSION_WARN(is, "received PDU with unsupported "
 785                     "opcode 0x%x; reconnecting",
 786                     response->ip_bhs->bhs_opcode);
 787                 iscsi_session_reconnect(is);
 788                 ISCSI_SESSION_UNLOCK(is);
 789                 icl_pdu_free(response);
 790         }
 791 }
 792
 793 static void
 794 iscsi_error_callback(struct icl_conn *ic)
 795 {
 796         struct iscsi_session *is;
 797
 798         is = CONN_SESSION(ic);
 799
 800         ISCSI_SESSION_WARN(is, "connection error; reconnecting");
 801         iscsi_session_reconnect(is);
 802 }
 803
 804 static void
 805 iscsi_pdu_handle_nop_in(struct icl_pdu *response)
 806 {
 807         struct iscsi_session *is;
 808         struct iscsi_bhs_nop_out *bhsno;
 809         struct iscsi_bhs_nop_in *bhsni;
 810         struct icl_pdu *request;
 811         void *data = NULL;
 812         size_t datasize;
 813         int error;
 814
 815         is = PDU_SESSION(response);
 816         bhsni = (struct iscsi_bhs_nop_in *)response->ip_bhs;
 817
 818         if (bhsni->bhsni_target_transfer_tag == 0xffffffff) {
 819                 /*
 820                  * Nothing to do; iscsi_pdu_update_statsn() already
 821                  * zeroed the timeout.
 822                  */
 823                 icl_pdu_free(response);
 824                 return;
 825         }
 826
 827         datasize = icl_pdu_data_segment_length(response);
 828         if (datasize > 0) {
 829                 data = malloc(datasize, M_ISCSI, M_NOWAIT | M_ZERO);
 830                 if (data == NULL) {
 831                         ISCSI_SESSION_WARN(is, "failed to allocate memory; "
 832                             "reconnecting");
 833                         icl_pdu_free(response);
 834                         iscsi_session_reconnect(is);
 835                         return;
 836                 }
 837                 icl_pdu_get_data(response, 0, data, datasize);
 838         }
 839
 840         request = icl_pdu_new(response->ip_conn, M_NOWAIT);
 841         if (request == NULL) {
 842                 ISCSI_SESSION_WARN(is, "failed to allocate memory; "
 843                     "reconnecting");
 844                 free(data, M_ISCSI);
 845                 icl_pdu_free(response);
 846                 iscsi_session_reconnect(is);
 847                 return;
 848         }
 849         bhsno = (struct iscsi_bhs_nop_out *)request->ip_bhs;
 850         bhsno->bhsno_opcode = ISCSI_BHS_OPCODE_NOP_OUT |
 851             ISCSI_BHS_OPCODE_IMMEDIATE;
 852         bhsno->bhsno_flags = 0x80;
 853         bhsno->bhsno_initiator_task_tag = 0xffffffff;
 854         bhsno->bhsno_target_transfer_tag = bhsni->bhsni_target_transfer_tag;
 855         if (datasize > 0) {
 856                 error = icl_pdu_append_data(request, data, datasize, M_NOWAIT);
 857                 if (error != 0) {
 858                         ISCSI_SESSION_WARN(is, "failed to allocate memory; "
 859                             "reconnecting");
 860                         free(data, M_ISCSI);
 861                         icl_pdu_free(request);
 862                         icl_pdu_free(response);
 863                         iscsi_session_reconnect(is);
 864                         return;
 865                 }
 866                 free(data, M_ISCSI);
 867         }
 868
 869         icl_pdu_free(response);
 870         iscsi_pdu_queue_locked(request);
 871 }
 872
 873 static void
 874 iscsi_pdu_handle_scsi_response(struct icl_pdu *response)
 875 {
 876         struct iscsi_bhs_scsi_response *bhssr;
 877         struct iscsi_outstanding *io;
 878         struct iscsi_session *is;
 879         union ccb *ccb;
 880         struct ccb_scsiio *csio;
 881         size_t data_segment_len, received;
 882         uint16_t sense_len;
 883         uint32_t resid;
 884
 885         is = PDU_SESSION(response);
 886
 887         bhssr = (struct iscsi_bhs_scsi_response *)response->ip_bhs;
 888         io = iscsi_outstanding_find(is, bhssr->bhssr_initiator_task_tag);
 889         if (io == NULL || io->io_ccb == NULL) {
 890                 ISCSI_SESSION_WARN(is, "bad itt 0x%x", bhssr->bhssr_initiator_task_tag);
 891                 icl_pdu_free(response);
 892                 iscsi_session_reconnect(is);
 893                 ISCSI_SESSION_UNLOCK(is);
 894                 return;
 895         }
 896
 897         ccb = io->io_ccb;
 898         if (bhssr->bhssr_response == BHSSR_RESPONSE_COMMAND_COMPLETED) {
 899                 if (ntohl(bhssr->bhssr_expdatasn) != io->io_datasn) {
 900                         ISCSI_SESSION_WARN(is,
 901                             "ExpDataSN mismatch in SCSI Response (%u vs %u)",
 902                             ntohl(bhssr->bhssr_expdatasn), io->io_datasn);
 903
 904                         /*
 905                          * XXX: Permit an ExpDataSN of zero for errors.
 906                          *
 907                          * This doesn't conform to RFC 7143, but some
 908                          * targets seem to do this.
 909                          */
 910                         if (bhssr->bhssr_status != 0 &&
 911                             bhssr->bhssr_expdatasn == htonl(0))
 912                                 goto skip_expdatasn;
 913
 914                         icl_pdu_free(response);
 915                         iscsi_session_reconnect(is);
 916                         ISCSI_SESSION_UNLOCK(is);
 917                         return;
 918                 }
 919         } else {
 920                 if (bhssr->bhssr_expdatasn != htonl(0)) {
 921                         ISCSI_SESSION_WARN(is,
 922                             "ExpDataSN mismatch in SCSI Response (%u vs 0)",
 923                             ntohl(bhssr->bhssr_expdatasn));
 924                         icl_pdu_free(response);
 925                         iscsi_session_reconnect(is);
 926                         ISCSI_SESSION_UNLOCK(is);
 927                         return;
 928                 }
 929         }
 930 skip_expdatasn:
 931
 932         /*
 933          * With iSER, after getting good response we can be sure
 934          * that all the data has been successfully transferred.
 935          */
 936         if (is->is_conn->ic_iser) {
 937                 resid = ntohl(bhssr->bhssr_residual_count);
 938                 if (bhssr->bhssr_flags & BHSSR_FLAGS_RESIDUAL_UNDERFLOW) {
 939                         io->io_received = ccb->csio.dxfer_len - resid;
 940                 } else if (bhssr->bhssr_flags & BHSSR_FLAGS_RESIDUAL_OVERFLOW) {
 941                         ISCSI_SESSION_WARN(is, "overflow: target indicates %d", resid);
 942                 } else {
 943                         io->io_received = ccb->csio.dxfer_len;
 944                 }
 945         }
 946
 947         received = io->io_received;
 948         iscsi_outstanding_remove(is, io);
 949         ISCSI_SESSION_UNLOCK(is);
 950
 951         if (bhssr->bhssr_response != BHSSR_RESPONSE_COMMAND_COMPLETED) {
 952                 ISCSI_SESSION_WARN(is, "service response 0x%x", bhssr->bhssr_response);
 953                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
 954                         xpt_freeze_devq(ccb->ccb_h.path, 1);
 955                         ISCSI_SESSION_DEBUG(is, "freezing devq");
 956                 }
 957                 ccb->ccb_h.status = CAM_REQ_CMP_ERR | CAM_DEV_QFRZN;
 958         } else if (bhssr->bhssr_status == 0) {
 959                 ccb->ccb_h.status = CAM_REQ_CMP;
 960         } else {
 961                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
 962                         xpt_freeze_devq(ccb->ccb_h.path, 1);
 963                         ISCSI_SESSION_DEBUG(is, "freezing devq");
 964                 }
 965                 ccb->ccb_h.status = CAM_SCSI_STATUS_ERROR | CAM_DEV_QFRZN;
 966                 ccb->csio.scsi_status = bhssr->bhssr_status;
 967         }
 968
 969         csio = &ccb->csio;
 970         data_segment_len = icl_pdu_data_segment_length(response);
 971         if (data_segment_len > 0) {
 972                 if (data_segment_len < sizeof(sense_len)) {
 973                         ISCSI_SESSION_WARN(is, "truncated data segment (%zd bytes)",
 974                             data_segment_len);
 975                         if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
 976                                 xpt_freeze_devq(ccb->ccb_h.path, 1);
 977                                 ISCSI_SESSION_DEBUG(is, "freezing devq");
 978                         }
 979                         ccb->ccb_h.status = CAM_REQ_CMP_ERR | CAM_DEV_QFRZN;
 980                         goto out;
 981                 }
 982                 icl_pdu_get_data(response, 0, &sense_len, sizeof(sense_len));
 983                 sense_len = ntohs(sense_len);
 984 #if 0
 985                 ISCSI_SESSION_DEBUG(is, "sense_len %d, data len %zd",
 986                     sense_len, data_segment_len);
 987 #endif
 988                 if (sizeof(sense_len) + sense_len > data_segment_len) {
 989                         ISCSI_SESSION_WARN(is, "truncated data segment "
 990                             "(%zd bytes, should be %zd)",
 991                             data_segment_len, sizeof(sense_len) + sense_len);
 992                         if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
 993                                 xpt_freeze_devq(ccb->ccb_h.path, 1);
 994                                 ISCSI_SESSION_DEBUG(is, "freezing devq");
 995                         }
 996                         ccb->ccb_h.status = CAM_REQ_CMP_ERR | CAM_DEV_QFRZN;
 997                         goto out;
 998                 } else if (sizeof(sense_len) + sense_len < data_segment_len)
 999                         ISCSI_SESSION_WARN(is, "oversize data segment "
1000                             "(%zd bytes, should be %zd)",
1001                             data_segment_len, sizeof(sense_len) + sense_len);
1002                 if (sense_len > csio->sense_len) {
1003                         ISCSI_SESSION_DEBUG(is, "truncating sense from %d to %d",
1004                             sense_len, csio->sense_len);
1005                         sense_len = csio->sense_len;
1006                 }
1007                 icl_pdu_get_data(response, sizeof(sense_len), &csio->sense_data, sense_len);
1008                 csio->sense_resid = csio->sense_len - sense_len;
1009                 ccb->ccb_h.status |= CAM_AUTOSNS_VALID;
1010         }
1011
1012 out:
1013         if (bhssr->bhssr_flags & BHSSR_FLAGS_RESIDUAL_UNDERFLOW)
1014                 csio->resid = ntohl(bhssr->bhssr_residual_count);
1015
1016         if ((csio->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_IN) {
1017                 KASSERT(received <= csio->dxfer_len,
1018                     ("received > csio->dxfer_len"));
1019                 if (received < csio->dxfer_len) {
1020                         if (csio->resid != csio->dxfer_len - received) {
1021                                 ISCSI_SESSION_WARN(is, "underflow mismatch: "
1022                                     "target indicates %d, we calculated %zd",
1023                                     csio->resid, csio->dxfer_len - received);
1024                         }
1025                         csio->resid = csio->dxfer_len - received;
1026                 }
1027         }
1028
1029         xpt_done(ccb);
1030         icl_pdu_free(response);
1031 }
1032
1033 static void
1034 iscsi_pdu_handle_task_response(struct icl_pdu *response)
1035 {
1036         struct iscsi_bhs_task_management_response *bhstmr;
1037         struct iscsi_outstanding *io, *aio;
1038         struct iscsi_session *is;
1039
1040         is = PDU_SESSION(response);
1041
1042         bhstmr = (struct iscsi_bhs_task_management_response *)response->ip_bhs;
1043         io = iscsi_outstanding_find(is, bhstmr->bhstmr_initiator_task_tag);
1044         if (io == NULL || io->io_ccb != NULL) {
1045                 ISCSI_SESSION_WARN(is, "bad itt 0x%x",
1046                     bhstmr->bhstmr_initiator_task_tag);
1047                 icl_pdu_free(response);
1048                 iscsi_session_reconnect(is);
1049                 return;
1050         }
1051
1052         if (bhstmr->bhstmr_response != BHSTMR_RESPONSE_FUNCTION_COMPLETE) {
1053                 ISCSI_SESSION_WARN(is, "task response 0x%x",
1054                     bhstmr->bhstmr_response);
1055         } else {
1056                 aio = iscsi_outstanding_find(is, io->io_referenced_task_tag);
1057                 if (aio != NULL && aio->io_ccb != NULL)
1058                         iscsi_session_terminate_task(is, aio, CAM_REQ_ABORTED);
1059         }
1060
1061         iscsi_outstanding_remove(is, io);
1062         icl_pdu_free(response);
1063 }
1064
1065 static void
1066 iscsi_pdu_handle_data_in(struct icl_pdu *response)
1067 {
1068         struct iscsi_bhs_data_in *bhsdi;
1069         struct iscsi_outstanding *io;
1070         struct iscsi_session *is;
1071         union ccb *ccb;
1072         struct ccb_scsiio *csio;
1073         size_t data_segment_len, received, oreceived;
1074
1075         is = PDU_SESSION(response);
1076         bhsdi = (struct iscsi_bhs_data_in *)response->ip_bhs;
1077         io = iscsi_outstanding_find(is, bhsdi->bhsdi_initiator_task_tag);
1078         if (io == NULL || io->io_ccb == NULL) {
1079                 ISCSI_SESSION_WARN(is, "bad itt 0x%x", bhsdi->bhsdi_initiator_task_tag);
1080                 icl_pdu_free(response);
1081                 iscsi_session_reconnect(is);
1082                 ISCSI_SESSION_UNLOCK(is);
1083                 return;
1084         }
1085
1086         if (io->io_datasn != ntohl(bhsdi->bhsdi_datasn)) {
1087                 ISCSI_SESSION_WARN(is, "received Data-In PDU with "
1088                     "DataSN %u, while expected %u; dropping connection",
1089                     ntohl(bhsdi->bhsdi_datasn), io->io_datasn);
1090                 icl_pdu_free(response);
1091                 iscsi_session_reconnect(is);
1092                 ISCSI_SESSION_UNLOCK(is);
1093                 return;
1094         }
1095         io->io_datasn += response->ip_additional_pdus + 1;
1096
1097         data_segment_len = icl_pdu_data_segment_length(response);
1098         if (data_segment_len == 0) {
1099                 /*
1100                  * "The sending of 0 length data segments should be avoided,
1101                  * but initiators and targets MUST be able to properly receive
1102                  * 0 length data segments."
1103                  */
1104                 ISCSI_SESSION_UNLOCK(is);
1105                 icl_pdu_free(response);
1106                 return;
1107         }
1108
1109         /*
1110          * We need to track this for security reasons - without it, malicious target
1111          * could respond to SCSI READ without sending Data-In PDUs, which would result
1112          * in read operation on the initiator side returning random kernel data.
1113          */
1114         if (ntohl(bhsdi->bhsdi_buffer_offset) != io->io_received) {
1115                 ISCSI_SESSION_WARN(is, "data out of order; expected offset %zd, got %zd",
1116                     io->io_received, (size_t)ntohl(bhsdi->bhsdi_buffer_offset));
1117                 icl_pdu_free(response);
1118                 iscsi_session_reconnect(is);
1119                 ISCSI_SESSION_UNLOCK(is);
1120                 return;
1121         }
1122
1123         ccb = io->io_ccb;
1124         csio = &ccb->csio;
1125
1126         if (io->io_received + data_segment_len > csio->dxfer_len) {
1127                 ISCSI_SESSION_WARN(is, "oversize data segment (%zd bytes "
1128                     "at offset %zd, buffer is %d)",
1129                     data_segment_len, io->io_received, csio->dxfer_len);
1130                 icl_pdu_free(response);
1131                 iscsi_session_reconnect(is);
1132                 ISCSI_SESSION_UNLOCK(is);
1133                 return;
1134         }
1135
1136         oreceived = io->io_received;
1137         io->io_received += data_segment_len;
1138         received = io->io_received;
1139         if ((bhsdi->bhsdi_flags & BHSDI_FLAGS_S) != 0)
1140                 iscsi_outstanding_remove(is, io);
1141         ISCSI_SESSION_UNLOCK(is);
1142
1143         icl_pdu_get_data(response, 0, csio->data_ptr + oreceived, data_segment_len);
1144
1145         /*
1146          * XXX: Check F.
1147          */
1148         if ((bhsdi->bhsdi_flags & BHSDI_FLAGS_S) == 0) {
1149                 /*
1150                  * Nothing more to do.
1151                  */
1152                 icl_pdu_free(response);
1153                 return;
1154         }
1155
1156         //ISCSI_SESSION_DEBUG(is, "got S flag; status 0x%x", bhsdi->bhsdi_status);
1157         if (bhsdi->bhsdi_status == 0) {
1158                 ccb->ccb_h.status = CAM_REQ_CMP;
1159         } else {
1160                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
1161                         xpt_freeze_devq(ccb->ccb_h.path, 1);
1162                         ISCSI_SESSION_DEBUG(is, "freezing devq");
1163                 }
1164                 ccb->ccb_h.status = CAM_SCSI_STATUS_ERROR | CAM_DEV_QFRZN;
1165                 csio->scsi_status = bhsdi->bhsdi_status;
1166         }
1167
1168         if ((csio->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_IN) {
1169                 KASSERT(received <= csio->dxfer_len,
1170                     ("received > csio->dxfer_len"));
1171                 if (received < csio->dxfer_len) {
1172                         csio->resid = ntohl(bhsdi->bhsdi_residual_count);
1173                         if (csio->resid != csio->dxfer_len - received) {
1174                                 ISCSI_SESSION_WARN(is, "underflow mismatch: "
1175                                     "target indicates %d, we calculated %zd",
1176                                     csio->resid, csio->dxfer_len - received);
1177                         }
1178                         csio->resid = csio->dxfer_len - received;
1179                 }
1180         }
1181
1182         xpt_done(ccb);
1183         icl_pdu_free(response);
1184 }
1185
1186 static void
1187 iscsi_pdu_handle_logout_response(struct icl_pdu *response)
1188 {
1189
1190         ISCSI_SESSION_DEBUG(PDU_SESSION(response), "logout response");
1191         icl_pdu_free(response);
1192 }
1193
1194 static void
1195 iscsi_pdu_handle_r2t(struct icl_pdu *response)
1196 {
1197         struct icl_pdu *request;
1198         struct iscsi_session *is;
1199         struct iscsi_bhs_r2t *bhsr2t;
1200         struct iscsi_bhs_data_out *bhsdo;
1201         struct iscsi_outstanding *io;
1202         struct ccb_scsiio *csio;
1203         size_t off, len, max_send_data_segment_length, total_len;
1204         int error;
1205         uint32_t datasn = 0;
1206
1207         is = PDU_SESSION(response);
1208
1209         bhsr2t = (struct iscsi_bhs_r2t *)response->ip_bhs;
1210         io = iscsi_outstanding_find(is, bhsr2t->bhsr2t_initiator_task_tag);
1211         if (io == NULL || io->io_ccb == NULL) {
1212                 ISCSI_SESSION_WARN(is, "bad itt 0x%x; reconnecting",
1213                     bhsr2t->bhsr2t_initiator_task_tag);
1214                 icl_pdu_free(response);
1215                 iscsi_session_reconnect(is);
1216                 return;
1217         }
1218
1219         csio = &io->io_ccb->csio;
1220
1221         if ((csio->ccb_h.flags & CAM_DIR_MASK) != CAM_DIR_OUT) {
1222                 ISCSI_SESSION_WARN(is, "received R2T for read command; reconnecting");
1223                 icl_pdu_free(response);
1224                 iscsi_session_reconnect(is);
1225                 return;
1226         }
1227
1228         /*
1229          * XXX: Verify R2TSN.
1230          */
1231
1232         off = ntohl(bhsr2t->bhsr2t_buffer_offset);
1233         if (off > csio->dxfer_len) {
1234                 ISCSI_SESSION_WARN(is, "target requested invalid offset "
1235                     "%zd, buffer is is %d; reconnecting", off, csio->dxfer_len);
1236                 icl_pdu_free(response);
1237                 iscsi_session_reconnect(is);
1238                 return;
1239         }
1240
1241         total_len = ntohl(bhsr2t->bhsr2t_desired_data_transfer_length);
1242         if (total_len == 0 || total_len > csio->dxfer_len) {
1243                 ISCSI_SESSION_WARN(is, "target requested invalid length "
1244                     "%zd, buffer is %d; reconnecting", total_len, csio->dxfer_len);
1245                 icl_pdu_free(response);
1246                 iscsi_session_reconnect(is);
1247                 return;
1248         }
1249
1250         //ISCSI_SESSION_DEBUG(is, "r2t; off %zd, len %zd", off, total_len);
1251
1252         if (is->is_conn->ic_hw_isomax != 0)
1253                 max_send_data_segment_length = is->is_conn->ic_hw_isomax;
1254         else
1255                 max_send_data_segment_length =
1256                     is->is_conn->ic_max_send_data_segment_length;
1257         for (;;) {
1258                 len = total_len;
1259
1260                 if (len > max_send_data_segment_length)
1261                         len = max_send_data_segment_length;
1262
1263                 if (off + len > csio->dxfer_len) {
1264                         ISCSI_SESSION_WARN(is, "target requested invalid "
1265                             "length/offset %zd, buffer is %d; reconnecting",
1266                             off + len, csio->dxfer_len);
1267                         icl_pdu_free(response);
1268                         iscsi_session_reconnect(is);
1269                         return;
1270                 }
1271
1272                 request = icl_pdu_new(response->ip_conn, M_NOWAIT);
1273                 if (request == NULL) {
1274                         icl_pdu_free(response);
1275                         iscsi_session_reconnect(is);
1276                         return;
1277                 }
1278
1279                 bhsdo = (struct iscsi_bhs_data_out *)request->ip_bhs;
1280                 bhsdo->bhsdo_opcode = ISCSI_BHS_OPCODE_SCSI_DATA_OUT;
1281                 bhsdo->bhsdo_lun = bhsr2t->bhsr2t_lun;
1282                 bhsdo->bhsdo_initiator_task_tag =
1283                     bhsr2t->bhsr2t_initiator_task_tag;
1284                 bhsdo->bhsdo_target_transfer_tag =
1285                     bhsr2t->bhsr2t_target_transfer_tag;
1286                 bhsdo->bhsdo_datasn = htonl(datasn);
1287                 bhsdo->bhsdo_buffer_offset = htonl(off);
1288                 error = icl_pdu_append_data(request, csio->data_ptr + off, len,
1289                     M_NOWAIT);
1290                 if (error != 0) {
1291                         ISCSI_SESSION_WARN(is, "failed to allocate memory; "
1292                             "reconnecting");
1293                         icl_pdu_free(request);
1294                         icl_pdu_free(response);
1295                         iscsi_session_reconnect(is);
1296                         return;
1297                 }
1298
1299                 datasn += howmany(len,
1300                     is->is_conn->ic_max_send_data_segment_length);
1301                 off += len;
1302                 total_len -= len;
1303
1304                 if (total_len == 0) {
1305                         bhsdo->bhsdo_flags |= BHSDO_FLAGS_F;
1306                         //ISCSI_SESSION_DEBUG(is, "setting F, off %zd", off);
1307                 } else {
1308                         //ISCSI_SESSION_DEBUG(is, "not finished, off %zd", off);
1309                 }
1310
1311                 iscsi_pdu_queue_locked(request);
1312
1313                 if (total_len == 0)
1314                         break;
1315         }
1316
1317         icl_pdu_free(response);
1318 }
1319
1320 static void
1321 iscsi_pdu_handle_async_message(struct icl_pdu *response)
1322 {
1323         struct iscsi_bhs_asynchronous_message *bhsam;
1324         struct iscsi_session *is;
1325
1326         is = PDU_SESSION(response);
1327         bhsam = (struct iscsi_bhs_asynchronous_message *)response->ip_bhs;
1328         switch (bhsam->bhsam_async_event) {
1329         case BHSAM_EVENT_TARGET_REQUESTS_LOGOUT:
1330                 ISCSI_SESSION_WARN(is, "target requests logout; removing session");
1331                 iscsi_session_logout(is);
1332                 iscsi_session_terminate(is);
1333                 break;
1334         case BHSAM_EVENT_TARGET_TERMINATES_CONNECTION:
1335                 ISCSI_SESSION_WARN(is, "target indicates it will drop the connection");
1336                 break;
1337         case BHSAM_EVENT_TARGET_TERMINATES_SESSION:
1338                 ISCSI_SESSION_WARN(is, "target indicates it will drop the session");
1339                 break;
1340         default:
1341                 /*
1342                  * XXX: Technically, we're obligated to also handle
1343                  *      parameter renegotiation.
1344                  */
1345                 ISCSI_SESSION_WARN(is, "ignoring AsyncEvent %d", bhsam->bhsam_async_event);
1346                 break;
1347         }
1348
1349         icl_pdu_free(response);
1350 }
1351
1352 static void
1353 iscsi_pdu_handle_reject(struct icl_pdu *response)
1354 {
1355         struct iscsi_bhs_reject *bhsr;
1356         struct iscsi_session *is;
1357
1358         is = PDU_SESSION(response);
1359         bhsr = (struct iscsi_bhs_reject *)response->ip_bhs;
1360         ISCSI_SESSION_WARN(is, "received Reject PDU, reason 0x%x; protocol error?",
1361             bhsr->bhsr_reason);
1362
1363         icl_pdu_free(response);
1364 }
1365
1366 static int
1367 iscsi_ioctl_daemon_wait(struct iscsi_softc *sc,
1368     struct iscsi_daemon_request *request)
1369 {
1370         struct iscsi_session *is;
1371         struct icl_drv_limits idl;
1372         int error;
1373
1374         sx_slock(&sc->sc_lock);
1375         for (;;) {
1376                 TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1377                         ISCSI_SESSION_LOCK(is);
1378                         if (is->is_conf.isc_enable == 0 &&
1379                             is->is_conf.isc_discovery == 0) {
1380                                 ISCSI_SESSION_UNLOCK(is);
1381                                 continue;
1382                         }
1383                         if (is->is_waiting_for_iscsid)
1384                                 break;
1385                         ISCSI_SESSION_UNLOCK(is);
1386                 }
1387
1388                 if (is == NULL) {
1389                         if (sc->sc_unloading) {
1390                                 sx_sunlock(&sc->sc_lock);
1391                                 return (ENXIO);
1392                         }
1393
1394                         /*
1395                          * No session requires attention from iscsid(8); wait.
1396                          */
1397                         error = cv_wait_sig(&sc->sc_cv, &sc->sc_lock);
1398                         if (error != 0) {
1399                                 sx_sunlock(&sc->sc_lock);
1400                                 return (error);
1401                         }
1402                         continue;
1403                 }
1404
1405                 is->is_waiting_for_iscsid = false;
1406                 is->is_login_phase = true;
1407                 is->is_reason[0] = '\0';
1408                 ISCSI_SESSION_UNLOCK(is);
1409
1410                 request->idr_session_id = is->is_id;
1411                 memcpy(&request->idr_isid, &is->is_isid,
1412                     sizeof(request->idr_isid));
1413                 request->idr_tsih = 0;  /* New or reinstated session. */
1414                 memcpy(&request->idr_conf, &is->is_conf,
1415                     sizeof(request->idr_conf));
1416
1417                 error = icl_limits(is->is_conf.isc_offload,
1418                     is->is_conf.isc_iser, &idl);
1419                 if (error != 0) {
1420                         ISCSI_SESSION_WARN(is, "icl_limits for offload \"%s\" "
1421                             "failed with error %d", is->is_conf.isc_offload,
1422                             error);
1423                         sx_sunlock(&sc->sc_lock);
1424                         return (error);
1425                 }
1426                 request->idr_limits.isl_max_recv_data_segment_length =
1427                     idl.idl_max_recv_data_segment_length;
1428                 request->idr_limits.isl_max_send_data_segment_length =
1429                     idl.idl_max_send_data_segment_length;
1430                 request->idr_limits.isl_max_burst_length =
1431                     idl.idl_max_burst_length;
1432                 request->idr_limits.isl_first_burst_length =
1433                     idl.idl_first_burst_length;
1434
1435                 sx_sunlock(&sc->sc_lock);
1436                 return (0);
1437         }
1438 }
1439
1440 static int
1441 iscsi_ioctl_daemon_handoff(struct iscsi_softc *sc,
1442     struct iscsi_daemon_handoff *handoff)
1443 {
1444         struct iscsi_session *is;
1445         struct icl_conn *ic;
1446         int error;
1447
1448         sx_slock(&sc->sc_lock);
1449
1450         /*
1451          * Find the session to hand off socket to.
1452          */
1453         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1454                 if (is->is_id == handoff->idh_session_id)
1455                         break;
1456         }
1457         if (is == NULL) {
1458                 sx_sunlock(&sc->sc_lock);
1459                 return (ESRCH);
1460         }
1461         ISCSI_SESSION_LOCK(is);
1462         ic = is->is_conn;
1463         if (is->is_conf.isc_discovery || is->is_terminating) {
1464                 ISCSI_SESSION_UNLOCK(is);
1465                 sx_sunlock(&sc->sc_lock);
1466                 return (EINVAL);
1467         }
1468         if (is->is_connected) {
1469                 /*
1470                  * This might have happened because another iscsid(8)
1471                  * instance handed off the connection in the meantime.
1472                  * Just return.
1473                  */
1474                 ISCSI_SESSION_WARN(is, "handoff on already connected "
1475                     "session");
1476                 ISCSI_SESSION_UNLOCK(is);
1477                 sx_sunlock(&sc->sc_lock);
1478                 return (EBUSY);
1479         }
1480
1481         strlcpy(is->is_target_alias, handoff->idh_target_alias,
1482             sizeof(is->is_target_alias));
1483         is->is_tsih = handoff->idh_tsih;
1484         is->is_statsn = handoff->idh_statsn;
1485         is->is_protocol_level = handoff->idh_protocol_level;
1486         is->is_initial_r2t = handoff->idh_initial_r2t;
1487         is->is_immediate_data = handoff->idh_immediate_data;
1488
1489         ic->ic_max_recv_data_segment_length =
1490             handoff->idh_max_recv_data_segment_length;
1491         ic->ic_max_send_data_segment_length =
1492             handoff->idh_max_send_data_segment_length;
1493         is->is_max_burst_length = handoff->idh_max_burst_length;
1494         is->is_first_burst_length = handoff->idh_first_burst_length;
1495
1496         if (handoff->idh_header_digest == ISCSI_DIGEST_CRC32C)
1497                 ic->ic_header_crc32c = true;
1498         else
1499                 ic->ic_header_crc32c = false;
1500         if (handoff->idh_data_digest == ISCSI_DIGEST_CRC32C)
1501                 ic->ic_data_crc32c = true;
1502         else
1503                 ic->ic_data_crc32c = false;
1504         ic->ic_maxtags = maxtags;
1505
1506         is->is_cmdsn = 0;
1507         is->is_expcmdsn = 0;
1508         is->is_maxcmdsn = 0;
1509         is->is_waiting_for_iscsid = false;
1510         is->is_login_phase = false;
1511         is->is_timeout = 0;
1512         is->is_connected = true;
1513         is->is_reason[0] = '\0';
1514
1515         ISCSI_SESSION_UNLOCK(is);
1516
1517         /*
1518          * If we're going through the proxy, the idh_socket will be 0,
1519          * and the ICL module can simply ignore this call.  It can also
1520          * use it to determine it's no longer in the Login phase.
1521          */
1522         error = icl_conn_handoff(ic, handoff->idh_socket);
1523         if (error != 0) {
1524                 sx_sunlock(&sc->sc_lock);
1525                 iscsi_session_terminate(is);
1526                 return (error);
1527         }
1528
1529         sx_sunlock(&sc->sc_lock);
1530
1531         if (is->is_sim != NULL) {
1532                 /*
1533                  * When reconnecting, there already is SIM allocated for the session.
1534                  */
1535                 KASSERT(is->is_simq_frozen, ("reconnect without frozen simq"));
1536                 ISCSI_SESSION_LOCK(is);
1537                 ISCSI_SESSION_DEBUG(is, "releasing");
1538                 is->is_simq_frozen = false;
1539                 xpt_release_simq(is->is_sim, 1);
1540                 ISCSI_SESSION_UNLOCK(is);
1541
1542         } else {
1543                 ISCSI_SESSION_LOCK(is);
1544                 is->is_devq = cam_simq_alloc(ic->ic_maxtags);
1545                 if (is->is_devq == NULL) {
1546                         ISCSI_SESSION_UNLOCK(is);
1547                         ISCSI_SESSION_WARN(is, "failed to allocate simq");
1548                         iscsi_session_terminate(is);
1549                         return (ENOMEM);
1550                 }
1551
1552                 is->is_sim = cam_sim_alloc(iscsi_action, NULL, "iscsi",
1553                     is, is->is_id /* unit */, &is->is_lock,
1554                     1, ic->ic_maxtags, is->is_devq);
1555                 if (is->is_sim == NULL) {
1556                         ISCSI_SESSION_UNLOCK(is);
1557                         ISCSI_SESSION_WARN(is, "failed to allocate SIM");
1558                         cam_simq_free(is->is_devq);
1559                         iscsi_session_terminate(is);
1560                         return (ENOMEM);
1561                 }
1562
1563                 if (xpt_bus_register(is->is_sim, NULL, 0) != 0) {
1564                         ISCSI_SESSION_UNLOCK(is);
1565                         ISCSI_SESSION_WARN(is, "failed to register bus");
1566                         iscsi_session_terminate(is);
1567                         return (ENOMEM);
1568                 }
1569
1570                 error = xpt_create_path(&is->is_path, /*periph*/NULL,
1571                     cam_sim_path(is->is_sim), CAM_TARGET_WILDCARD,
1572                     CAM_LUN_WILDCARD);
1573                 if (error != CAM_REQ_CMP) {
1574                         ISCSI_SESSION_UNLOCK(is);
1575                         ISCSI_SESSION_WARN(is, "failed to create path");
1576                         iscsi_session_terminate(is);
1577                         return (ENOMEM);
1578                 }
1579                 ISCSI_SESSION_UNLOCK(is);
1580         }
1581
1582         return (0);
1583 }
1584
1585 static int
1586 iscsi_ioctl_daemon_fail(struct iscsi_softc *sc,
1587     struct iscsi_daemon_fail *fail)
1588 {
1589         struct iscsi_session *is;
1590
1591         sx_slock(&sc->sc_lock);
1592
1593         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1594                 if (is->is_id == fail->idf_session_id)
1595                         break;
1596         }
1597         if (is == NULL) {
1598                 sx_sunlock(&sc->sc_lock);
1599                 return (ESRCH);
1600         }
1601         ISCSI_SESSION_LOCK(is);
1602         ISCSI_SESSION_DEBUG(is, "iscsid(8) failed: %s",
1603             fail->idf_reason);
1604         strlcpy(is->is_reason, fail->idf_reason, sizeof(is->is_reason));
1605         //is->is_waiting_for_iscsid = false;
1606         //is->is_login_phase = true;
1607         //iscsi_session_reconnect(is);
1608         ISCSI_SESSION_UNLOCK(is);
1609         sx_sunlock(&sc->sc_lock);
1610
1611         return (0);
1612 }
1613
1614 #ifdef ICL_KERNEL_PROXY
1615 static int
1616 iscsi_ioctl_daemon_connect(struct iscsi_softc *sc,
1617     struct iscsi_daemon_connect *idc)
1618 {
1619         struct iscsi_session *is;
1620         struct sockaddr *from_sa, *to_sa;
1621         int error;
1622
1623         sx_slock(&sc->sc_lock);
1624         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1625                 if (is->is_id == idc->idc_session_id)
1626                         break;
1627         }
1628         if (is == NULL) {
1629                 sx_sunlock(&sc->sc_lock);
1630                 return (ESRCH);
1631         }
1632         sx_sunlock(&sc->sc_lock);
1633
1634         if (idc->idc_from_addrlen > 0) {
1635                 error = getsockaddr(&from_sa, (void *)idc->idc_from_addr, idc->idc_from_addrlen);
1636                 if (error != 0) {
1637                         ISCSI_SESSION_WARN(is,
1638                             "getsockaddr failed with error %d", error);
1639                         return (error);
1640                 }
1641         } else {
1642                 from_sa = NULL;
1643         }
1644         error = getsockaddr(&to_sa, (void *)idc->idc_to_addr, idc->idc_to_addrlen);
1645         if (error != 0) {
1646                 ISCSI_SESSION_WARN(is, "getsockaddr failed with error %d",
1647                     error);
1648                 free(from_sa, M_SONAME);
1649                 return (error);
1650         }
1651
1652         ISCSI_SESSION_LOCK(is);
1653         is->is_statsn = 0;
1654         is->is_cmdsn = 0;
1655         is->is_expcmdsn = 0;
1656         is->is_maxcmdsn = 0;
1657         is->is_waiting_for_iscsid = false;
1658         is->is_login_phase = true;
1659         is->is_timeout = 0;
1660         ISCSI_SESSION_UNLOCK(is);
1661
1662         error = icl_conn_connect(is->is_conn, idc->idc_domain,
1663             idc->idc_socktype, idc->idc_protocol, from_sa, to_sa);
1664         free(from_sa, M_SONAME);
1665         free(to_sa, M_SONAME);
1666
1667         /*
1668          * Digests are always disabled during login phase.
1669          */
1670         is->is_conn->ic_header_crc32c = false;
1671         is->is_conn->ic_data_crc32c = false;
1672
1673         return (error);
1674 }
1675
1676 static int
1677 iscsi_ioctl_daemon_send(struct iscsi_softc *sc,
1678     struct iscsi_daemon_send *ids)
1679 {
1680         struct iscsi_session *is;
1681         struct icl_pdu *ip;
1682         size_t datalen;
1683         void *data;
1684         int error;
1685
1686         sx_slock(&sc->sc_lock);
1687         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1688                 if (is->is_id == ids->ids_session_id)
1689                         break;
1690         }
1691         if (is == NULL) {
1692                 sx_sunlock(&sc->sc_lock);
1693                 return (ESRCH);
1694         }
1695         sx_sunlock(&sc->sc_lock);
1696
1697         if (is->is_login_phase == false)
1698                 return (EBUSY);
1699
1700         if (is->is_terminating || is->is_reconnecting)
1701                 return (EIO);
1702
1703         datalen = ids->ids_data_segment_len;
1704         if (datalen > is->is_conn->ic_max_send_data_segment_length)
1705                 return (EINVAL);
1706         if (datalen > 0) {
1707                 data = malloc(datalen, M_ISCSI, M_WAITOK);
1708                 error = copyin(ids->ids_data_segment, data, datalen);
1709                 if (error != 0) {
1710                         free(data, M_ISCSI);
1711                         return (error);
1712                 }
1713         }
1714
1715         ip = icl_pdu_new(is->is_conn, M_WAITOK);
1716         memcpy(ip->ip_bhs, ids->ids_bhs, sizeof(*ip->ip_bhs));
1717         if (datalen > 0) {
1718                 error = icl_pdu_append_data(ip, data, datalen, M_WAITOK);
1719                 KASSERT(error == 0, ("icl_pdu_append_data(..., M_WAITOK) failed"));
1720                 free(data, M_ISCSI);
1721         }
1722         iscsi_pdu_queue(ip);
1723
1724         return (0);
1725 }
1726
1727 static int
1728 iscsi_ioctl_daemon_receive(struct iscsi_softc *sc,
1729     struct iscsi_daemon_receive *idr)
1730 {
1731         struct iscsi_session *is;
1732         struct icl_pdu *ip;
1733         void *data;
1734         int error;
1735
1736         sx_slock(&sc->sc_lock);
1737         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1738                 if (is->is_id == idr->idr_session_id)
1739                         break;
1740         }
1741         if (is == NULL) {
1742                 sx_sunlock(&sc->sc_lock);
1743                 return (ESRCH);
1744         }
1745         sx_sunlock(&sc->sc_lock);
1746
1747         if (is->is_login_phase == false)
1748                 return (EBUSY);
1749
1750         ISCSI_SESSION_LOCK(is);
1751         while (is->is_login_pdu == NULL &&
1752             is->is_terminating == false &&
1753             is->is_reconnecting == false) {
1754                 error = cv_wait_sig(&is->is_login_cv, &is->is_lock);
1755                 if (error != 0) {
1756                         ISCSI_SESSION_UNLOCK(is);
1757                         return (error);
1758                 }
1759         }
1760         if (is->is_terminating || is->is_reconnecting) {
1761                 ISCSI_SESSION_UNLOCK(is);
1762                 return (EIO);
1763         }
1764         ip = is->is_login_pdu;
1765         is->is_login_pdu = NULL;
1766         ISCSI_SESSION_UNLOCK(is);
1767
1768         if (ip->ip_data_len > idr->idr_data_segment_len) {
1769                 icl_pdu_free(ip);
1770                 return (EMSGSIZE);
1771         }
1772
1773         copyout(ip->ip_bhs, idr->idr_bhs, sizeof(*ip->ip_bhs));
1774         if (ip->ip_data_len > 0) {
1775                 data = malloc(ip->ip_data_len, M_ISCSI, M_WAITOK);
1776                 icl_pdu_get_data(ip, 0, data, ip->ip_data_len);
1777                 copyout(data, idr->idr_data_segment, ip->ip_data_len);
1778                 free(data, M_ISCSI);
1779         }
1780
1781         icl_pdu_free(ip);
1782
1783         return (0);
1784 }
1785 #endif /* ICL_KERNEL_PROXY */
1786
1787 static void
1788 iscsi_sanitize_session_conf(struct iscsi_session_conf *isc)
1789 {
1790         /*
1791          * Just make sure all the fields are null-terminated.
1792          *
1793          * XXX: This is not particularly secure.  We should
1794          *      create our own conf and then copy in relevant
1795          *      fields.
1796          */
1797         isc->isc_initiator[ISCSI_NAME_LEN - 1] = '\0';
1798         isc->isc_initiator_addr[ISCSI_ADDR_LEN - 1] = '\0';
1799         isc->isc_initiator_alias[ISCSI_ALIAS_LEN - 1] = '\0';
1800         isc->isc_target[ISCSI_NAME_LEN - 1] = '\0';
1801         isc->isc_target_addr[ISCSI_ADDR_LEN - 1] = '\0';
1802         isc->isc_user[ISCSI_NAME_LEN - 1] = '\0';
1803         isc->isc_secret[ISCSI_SECRET_LEN - 1] = '\0';
1804         isc->isc_mutual_user[ISCSI_NAME_LEN - 1] = '\0';
1805         isc->isc_mutual_secret[ISCSI_SECRET_LEN - 1] = '\0';
1806 }
1807
1808 static bool
1809 iscsi_valid_session_conf(const struct iscsi_session_conf *isc)
1810 {
1811
1812         if (isc->isc_initiator[0] == '\0') {
1813                 ISCSI_DEBUG("empty isc_initiator");
1814                 return (false);
1815         }
1816
1817         if (isc->isc_target_addr[0] == '\0') {
1818                 ISCSI_DEBUG("empty isc_target_addr");
1819                 return (false);
1820         }
1821
1822         if (isc->isc_discovery != 0 && isc->isc_target[0] != 0) {
1823                 ISCSI_DEBUG("non-empty isc_target for discovery session");
1824                 return (false);
1825         }
1826
1827         if (isc->isc_discovery == 0 && isc->isc_target[0] == 0) {
1828                 ISCSI_DEBUG("empty isc_target for non-discovery session");
1829                 return (false);
1830         }
1831
1832         return (true);
1833 }
1834
1835 static int
1836 iscsi_ioctl_session_add(struct iscsi_softc *sc, struct iscsi_session_add *isa)
1837 {
1838         struct iscsi_session *is;
1839         const struct iscsi_session *is2;
1840         int error;
1841         sbintime_t sbt, pr;
1842
1843         iscsi_sanitize_session_conf(&isa->isa_conf);
1844         if (iscsi_valid_session_conf(&isa->isa_conf) == false)
1845                 return (EINVAL);
1846
1847         is = malloc(sizeof(*is), M_ISCSI, M_ZERO | M_WAITOK);
1848         memcpy(&is->is_conf, &isa->isa_conf, sizeof(is->is_conf));
1849
1850         sx_xlock(&sc->sc_lock);
1851
1852         /*
1853          * Prevent duplicates.
1854          */
1855         TAILQ_FOREACH(is2, &sc->sc_sessions, is_next) {
1856                 if (!!is->is_conf.isc_discovery !=
1857                     !!is2->is_conf.isc_discovery)
1858                         continue;
1859
1860                 if (strcmp(is->is_conf.isc_target_addr,
1861                     is2->is_conf.isc_target_addr) != 0)
1862                         continue;
1863
1864                 if (is->is_conf.isc_discovery == 0 &&
1865                     strcmp(is->is_conf.isc_target,
1866                     is2->is_conf.isc_target) != 0)
1867                         continue;
1868
1869                 sx_xunlock(&sc->sc_lock);
1870                 free(is, M_ISCSI);
1871                 return (EBUSY);
1872         }
1873
1874         is->is_conn = icl_new_conn(is->is_conf.isc_offload,
1875             is->is_conf.isc_iser, "iscsi", &is->is_lock);
1876         if (is->is_conn == NULL) {
1877                 sx_xunlock(&sc->sc_lock);
1878                 free(is, M_ISCSI);
1879                 return (EINVAL);
1880         }
1881         is->is_conn->ic_receive = iscsi_receive_callback;
1882         is->is_conn->ic_error = iscsi_error_callback;
1883         is->is_conn->ic_prv0 = is;
1884         TAILQ_INIT(&is->is_outstanding);
1885         STAILQ_INIT(&is->is_postponed);
1886         mtx_init(&is->is_lock, "iscsi_lock", NULL, MTX_DEF);
1887         cv_init(&is->is_maintenance_cv, "iscsi_mt");
1888 #ifdef ICL_KERNEL_PROXY
1889         cv_init(&is->is_login_cv, "iscsi_login");
1890 #endif
1891
1892         /*
1893          * Set some default values, from RFC 3720, section 12.
1894          *
1895          * These values are updated by the handoff IOCTL, but are
1896          * needed prior to the handoff to support sending the ISER
1897          * login PDU.
1898          */
1899         is->is_conn->ic_max_recv_data_segment_length = 8192;
1900         is->is_conn->ic_max_send_data_segment_length = 8192;
1901         is->is_max_burst_length = 262144;
1902         is->is_first_burst_length = 65536;
1903
1904         is->is_softc = sc;
1905         sc->sc_last_session_id++;
1906         is->is_id = sc->sc_last_session_id;
1907         is->is_isid[0] = 0x80; /* RFC 3720, 10.12.5: 10b, "Random" ISID. */
1908         arc4rand(&is->is_isid[1], 5, 0);
1909         is->is_tsih = 0;
1910         callout_init(&is->is_callout, 1);
1911
1912         error = kthread_add(iscsi_maintenance_thread, is, NULL, NULL, 0, 0, "iscsimt");
1913         if (error != 0) {
1914                 ISCSI_SESSION_WARN(is, "kthread_add(9) failed with error %d", error);
1915                 sx_xunlock(&sc->sc_lock);
1916                 return (error);
1917         }
1918
1919         sbt = mstosbt(995);
1920         pr = mstosbt(10);
1921         callout_reset_sbt(&is->is_callout, sbt, pr, iscsi_callout, is, 0);
1922         TAILQ_INSERT_TAIL(&sc->sc_sessions, is, is_next);
1923
1924         ISCSI_SESSION_LOCK(is);
1925         /*
1926          * Don't notify iscsid(8) if the session is disabled and it's not
1927          * a discovery session,
1928          */
1929         if (is->is_conf.isc_enable == 0 && is->is_conf.isc_discovery == 0) {
1930                 ISCSI_SESSION_UNLOCK(is);
1931                 sx_xunlock(&sc->sc_lock);
1932                 return (0);
1933         }
1934
1935         is->is_waiting_for_iscsid = true;
1936         strlcpy(is->is_reason, "Waiting for iscsid(8)", sizeof(is->is_reason));
1937         ISCSI_SESSION_UNLOCK(is);
1938         cv_signal(&sc->sc_cv);
1939         sx_xunlock(&sc->sc_lock);
1940         return (0);
1941 }
1942
1943 static bool
1944 iscsi_session_conf_matches(unsigned int id1, const struct iscsi_session_conf *c1,
1945     unsigned int id2, const struct iscsi_session_conf *c2)
1946 {
1947
1948         if (id2 != 0 && id2 != id1)
1949                 return (false);
1950         if (c2->isc_target[0] != '\0' &&
1951             strcmp(c1->isc_target, c2->isc_target) != 0)
1952                 return (false);
1953         if (c2->isc_target_addr[0] != '\0' &&
1954             strcmp(c1->isc_target_addr, c2->isc_target_addr) != 0)
1955                 return (false);
1956         return (true);
1957 }
1958
1959 static int
1960 iscsi_ioctl_session_remove(struct iscsi_softc *sc,
1961     struct iscsi_session_remove *isr)
1962 {
1963         struct iscsi_session *is, *tmp;
1964         bool found = false;
1965
1966         iscsi_sanitize_session_conf(&isr->isr_conf);
1967
1968         sx_xlock(&sc->sc_lock);
1969         TAILQ_FOREACH_SAFE(is, &sc->sc_sessions, is_next, tmp) {
1970                 ISCSI_SESSION_LOCK(is);
1971                 if (iscsi_session_conf_matches(is->is_id, &is->is_conf,
1972                     isr->isr_session_id, &isr->isr_conf)) {
1973                         found = true;
1974                         iscsi_session_logout(is);
1975                         iscsi_session_terminate(is);
1976                 }
1977                 ISCSI_SESSION_UNLOCK(is);
1978         }
1979         sx_xunlock(&sc->sc_lock);
1980
1981         if (!found)
1982                 return (ESRCH);
1983
1984         return (0);
1985 }
1986
1987 static int
1988 iscsi_ioctl_session_list(struct iscsi_softc *sc, struct iscsi_session_list *isl)
1989 {
1990         int error;
1991         unsigned int i = 0;
1992         struct iscsi_session *is;
1993         struct iscsi_session_state iss;
1994
1995         sx_slock(&sc->sc_lock);
1996         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1997                 if (i >= isl->isl_nentries) {
1998                         sx_sunlock(&sc->sc_lock);
1999                         return (EMSGSIZE);
2000                 }
2001                 memset(&iss, 0, sizeof(iss));
2002                 memcpy(&iss.iss_conf, &is->is_conf, sizeof(iss.iss_conf));
2003                 iss.iss_id = is->is_id;
2004                 strlcpy(iss.iss_target_alias, is->is_target_alias, sizeof(iss.iss_target_alias));
2005                 strlcpy(iss.iss_reason, is->is_reason, sizeof(iss.iss_reason));
2006                 strlcpy(iss.iss_offload, is->is_conn->ic_offload, sizeof(iss.iss_offload));
2007
2008                 if (is->is_conn->ic_header_crc32c)
2009                         iss.iss_header_digest = ISCSI_DIGEST_CRC32C;
2010                 else
2011                         iss.iss_header_digest = ISCSI_DIGEST_NONE;
2012
2013                 if (is->is_conn->ic_data_crc32c)
2014                         iss.iss_data_digest = ISCSI_DIGEST_CRC32C;
2015                 else
2016                         iss.iss_data_digest = ISCSI_DIGEST_NONE;
2017
2018                 iss.iss_max_send_data_segment_length =
2019                     is->is_conn->ic_max_send_data_segment_length;
2020                 iss.iss_max_recv_data_segment_length =
2021                     is->is_conn->ic_max_recv_data_segment_length;
2022                 iss.iss_max_burst_length = is->is_max_burst_length;
2023                 iss.iss_first_burst_length = is->is_first_burst_length;
2024                 iss.iss_immediate_data = is->is_immediate_data;
2025                 iss.iss_connected = is->is_connected;
2026
2027                 error = copyout(&iss, isl->isl_pstates + i, sizeof(iss));
2028                 if (error != 0) {
2029                         sx_sunlock(&sc->sc_lock);
2030                         return (error);
2031                 }
2032                 i++;
2033         }
2034         sx_sunlock(&sc->sc_lock);
2035
2036         isl->isl_nentries = i;
2037
2038         return (0);
2039 }
2040
2041 static int
2042 iscsi_ioctl_session_modify(struct iscsi_softc *sc,
2043     struct iscsi_session_modify *ism)
2044 {
2045         struct iscsi_session *is;
2046         const struct iscsi_session *is2;
2047
2048         iscsi_sanitize_session_conf(&ism->ism_conf);
2049         if (iscsi_valid_session_conf(&ism->ism_conf) == false)
2050                 return (EINVAL);
2051
2052         sx_xlock(&sc->sc_lock);
2053         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
2054                 ISCSI_SESSION_LOCK(is);
2055                 if (is->is_id == ism->ism_session_id) {
2056                         /* Note that the session remains locked. */
2057                         break;
2058                 }
2059                 ISCSI_SESSION_UNLOCK(is);
2060         }
2061         if (is == NULL) {
2062                 sx_xunlock(&sc->sc_lock);
2063                 return (ESRCH);
2064         }
2065
2066         /*
2067          * Prevent duplicates.
2068          */
2069         TAILQ_FOREACH(is2, &sc->sc_sessions, is_next) {
2070                 if (is == is2)
2071                         continue;
2072
2073                 if (!!ism->ism_conf.isc_discovery !=
2074                     !!is2->is_conf.isc_discovery)
2075                         continue;
2076
2077                 if (strcmp(ism->ism_conf.isc_target_addr,
2078                     is2->is_conf.isc_target_addr) != 0)
2079                         continue;
2080
2081                 if (ism->ism_conf.isc_discovery == 0 &&
2082                     strcmp(ism->ism_conf.isc_target,
2083                     is2->is_conf.isc_target) != 0)
2084                         continue;
2085
2086                 ISCSI_SESSION_UNLOCK(is);
2087                 sx_xunlock(&sc->sc_lock);
2088                 return (EBUSY);
2089         }
2090
2091         sx_xunlock(&sc->sc_lock);
2092
2093         memcpy(&is->is_conf, &ism->ism_conf, sizeof(is->is_conf));
2094         ISCSI_SESSION_UNLOCK(is);
2095
2096         iscsi_session_reconnect(is);
2097
2098         return (0);
2099 }
2100
2101 static int
2102 iscsi_ioctl(struct cdev *dev, u_long cmd, caddr_t arg, int mode,
2103     struct thread *td)
2104 {
2105         struct iscsi_softc *sc;
2106
2107         sc = dev->si_drv1;
2108
2109         switch (cmd) {
2110         case ISCSIDWAIT:
2111                 return (iscsi_ioctl_daemon_wait(sc,
2112                     (struct iscsi_daemon_request *)arg));
2113         case ISCSIDHANDOFF:
2114                 return (iscsi_ioctl_daemon_handoff(sc,
2115                     (struct iscsi_daemon_handoff *)arg));
2116         case ISCSIDFAIL:
2117                 return (iscsi_ioctl_daemon_fail(sc,
2118                     (struct iscsi_daemon_fail *)arg));
2119 #ifdef ICL_KERNEL_PROXY
2120         case ISCSIDCONNECT:
2121                 return (iscsi_ioctl_daemon_connect(sc,
2122                     (struct iscsi_daemon_connect *)arg));
2123         case ISCSIDSEND:
2124                 return (iscsi_ioctl_daemon_send(sc,
2125                     (struct iscsi_daemon_send *)arg));
2126         case ISCSIDRECEIVE:
2127                 return (iscsi_ioctl_daemon_receive(sc,
2128                     (struct iscsi_daemon_receive *)arg));
2129 #endif /* ICL_KERNEL_PROXY */
2130         case ISCSISADD:
2131                 return (iscsi_ioctl_session_add(sc,
2132                     (struct iscsi_session_add *)arg));
2133         case ISCSISREMOVE:
2134                 return (iscsi_ioctl_session_remove(sc,
2135                     (struct iscsi_session_remove *)arg));
2136         case ISCSISLIST:
2137                 return (iscsi_ioctl_session_list(sc,
2138                     (struct iscsi_session_list *)arg));
2139         case ISCSISMODIFY:
2140                 return (iscsi_ioctl_session_modify(sc,
2141                     (struct iscsi_session_modify *)arg));
2142         default:
2143                 return (EINVAL);
2144         }
2145 }
2146
2147 static struct iscsi_outstanding *
2148 iscsi_outstanding_find(struct iscsi_session *is, uint32_t initiator_task_tag)
2149 {
2150         struct iscsi_outstanding *io;
2151
2152         ISCSI_SESSION_LOCK_ASSERT(is);
2153
2154         TAILQ_FOREACH(io, &is->is_outstanding, io_next) {
2155                 if (io->io_initiator_task_tag == initiator_task_tag)
2156                         return (io);
2157         }
2158         return (NULL);
2159 }
2160
2161 static struct iscsi_outstanding *
2162 iscsi_outstanding_find_ccb(struct iscsi_session *is, union ccb *ccb)
2163 {
2164         struct iscsi_outstanding *io;
2165
2166         ISCSI_SESSION_LOCK_ASSERT(is);
2167
2168         TAILQ_FOREACH(io, &is->is_outstanding, io_next) {
2169                 if (io->io_ccb == ccb)
2170                         return (io);
2171         }
2172         return (NULL);
2173 }
2174
2175 static struct iscsi_outstanding *
2176 iscsi_outstanding_add(struct iscsi_session *is, struct icl_pdu *request,
2177     union ccb *ccb, uint32_t *initiator_task_tagp)
2178 {
2179         struct iscsi_outstanding *io;
2180         int error;
2181
2182         ISCSI_SESSION_LOCK_ASSERT(is);
2183
2184         io = uma_zalloc(iscsi_outstanding_zone, M_NOWAIT | M_ZERO);
2185         if (io == NULL) {
2186                 ISCSI_SESSION_WARN(is, "failed to allocate %zd bytes",
2187                     sizeof(*io));
2188                 return (NULL);
2189         }
2190
2191         error = icl_conn_task_setup(is->is_conn, request, &ccb->csio,
2192             initiator_task_tagp, &io->io_icl_prv);
2193         if (error != 0) {
2194                 ISCSI_SESSION_WARN(is,
2195                     "icl_conn_task_setup() failed with error %d", error);
2196                 uma_zfree(iscsi_outstanding_zone, io);
2197                 return (NULL);
2198         }
2199
2200         KASSERT(iscsi_outstanding_find(is, *initiator_task_tagp) == NULL,
2201             ("initiator_task_tag 0x%x already added", *initiator_task_tagp));
2202
2203         io->io_initiator_task_tag = *initiator_task_tagp;
2204         io->io_ccb = ccb;
2205         TAILQ_INSERT_TAIL(&is->is_outstanding, io, io_next);
2206         return (io);
2207 }
2208
2209 static void
2210 iscsi_outstanding_remove(struct iscsi_session *is, struct iscsi_outstanding *io)
2211 {
2212
2213         ISCSI_SESSION_LOCK_ASSERT(is);
2214
2215         icl_conn_task_done(is->is_conn, io->io_icl_prv);
2216         TAILQ_REMOVE(&is->is_outstanding, io, io_next);
2217         uma_zfree(iscsi_outstanding_zone, io);
2218 }
2219
2220 static void
2221 iscsi_action_abort(struct iscsi_session *is, union ccb *ccb)
2222 {
2223         struct icl_pdu *request;
2224         struct iscsi_bhs_task_management_request *bhstmr;
2225         struct ccb_abort *cab = &ccb->cab;
2226         struct iscsi_outstanding *io, *aio;
2227         uint32_t initiator_task_tag;
2228
2229         ISCSI_SESSION_LOCK_ASSERT(is);
2230
2231 #if 0
2232         KASSERT(is->is_login_phase == false, ("%s called during Login Phase", __func__));
2233 #else
2234         if (is->is_login_phase) {
2235                 ccb->ccb_h.status = CAM_REQ_ABORTED;
2236                 xpt_done(ccb);
2237                 return;
2238         }
2239 #endif
2240
2241         aio = iscsi_outstanding_find_ccb(is, cab->abort_ccb);
2242         if (aio == NULL) {
2243                 ccb->ccb_h.status = CAM_REQ_CMP;
2244                 xpt_done(ccb);
2245                 return;
2246         }
2247
2248         request = icl_pdu_new(is->is_conn, M_NOWAIT);
2249         if (request == NULL) {
2250                 ccb->ccb_h.status = CAM_RESRC_UNAVAIL;
2251                 xpt_done(ccb);
2252                 return;
2253         }
2254
2255         initiator_task_tag = is->is_initiator_task_tag++;
2256         if (initiator_task_tag == 0xffffffff)
2257                 initiator_task_tag = is->is_initiator_task_tag++;
2258
2259         io = iscsi_outstanding_add(is, request, NULL, &initiator_task_tag);
2260         if (io == NULL) {
2261                 icl_pdu_free(request);
2262                 ccb->ccb_h.status = CAM_RESRC_UNAVAIL;
2263                 xpt_done(ccb);
2264                 return;
2265         }
2266         io->io_referenced_task_tag = aio->io_initiator_task_tag;
2267
2268         bhstmr = (struct iscsi_bhs_task_management_request *)request->ip_bhs;
2269         bhstmr->bhstmr_opcode = ISCSI_BHS_OPCODE_TASK_REQUEST;
2270         bhstmr->bhstmr_function = 0x80 | BHSTMR_FUNCTION_ABORT_TASK;
2271         bhstmr->bhstmr_lun = htobe64(CAM_EXTLUN_BYTE_SWIZZLE(ccb->ccb_h.target_lun));
2272         bhstmr->bhstmr_initiator_task_tag = initiator_task_tag;
2273         bhstmr->bhstmr_referenced_task_tag = aio->io_initiator_task_tag;
2274
2275         iscsi_pdu_queue_locked(request);
2276 }
2277
2278 static void
2279 iscsi_action_scsiio(struct iscsi_session *is, union ccb *ccb)
2280 {
2281         struct icl_pdu *request;
2282         struct iscsi_bhs_scsi_command *bhssc;
2283         struct ccb_scsiio *csio;
2284         struct iscsi_outstanding *io;
2285         size_t len;
2286         uint32_t initiator_task_tag;
2287         int error;
2288
2289         ISCSI_SESSION_LOCK_ASSERT(is);
2290
2291 #if 0
2292         KASSERT(is->is_login_phase == false, ("%s called during Login Phase", __func__));
2293 #else
2294         if (is->is_login_phase) {
2295                 ISCSI_SESSION_DEBUG(is, "called during login phase");
2296                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
2297                         xpt_freeze_devq(ccb->ccb_h.path, 1);
2298                         ISCSI_SESSION_DEBUG(is, "freezing devq");
2299                 }
2300                 ccb->ccb_h.status = CAM_REQ_ABORTED | CAM_DEV_QFRZN;
2301                 xpt_done(ccb);
2302                 return;
2303         }
2304 #endif
2305
2306         request = icl_pdu_new(is->is_conn, M_NOWAIT);
2307         if (request == NULL) {
2308                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
2309                         xpt_freeze_devq(ccb->ccb_h.path, 1);
2310                         ISCSI_SESSION_DEBUG(is, "freezing devq");
2311                 }
2312                 ccb->ccb_h.status = CAM_RESRC_UNAVAIL | CAM_DEV_QFRZN;
2313                 xpt_done(ccb);
2314                 return;
2315         }
2316
2317         initiator_task_tag = is->is_initiator_task_tag++;
2318         if (initiator_task_tag == 0xffffffff)
2319                 initiator_task_tag = is->is_initiator_task_tag++;
2320
2321         io = iscsi_outstanding_add(is, request, ccb, &initiator_task_tag);
2322         if (io == NULL) {
2323                 icl_pdu_free(request);
2324                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
2325                         xpt_freeze_devq(ccb->ccb_h.path, 1);
2326                         ISCSI_SESSION_DEBUG(is, "freezing devq");
2327                 }
2328                 ccb->ccb_h.status = CAM_RESRC_UNAVAIL | CAM_DEV_QFRZN;
2329                 xpt_done(ccb);
2330                 return;
2331         }
2332
2333         csio = &ccb->csio;
2334         bhssc = (struct iscsi_bhs_scsi_command *)request->ip_bhs;
2335         bhssc->bhssc_opcode = ISCSI_BHS_OPCODE_SCSI_COMMAND;
2336         bhssc->bhssc_flags |= BHSSC_FLAGS_F;
2337         switch (csio->ccb_h.flags & CAM_DIR_MASK) {
2338         case CAM_DIR_IN:
2339                 bhssc->bhssc_flags |= BHSSC_FLAGS_R;
2340                 break;
2341         case CAM_DIR_OUT:
2342                 bhssc->bhssc_flags |= BHSSC_FLAGS_W;
2343                 break;
2344         }
2345
2346         if ((ccb->ccb_h.flags & CAM_TAG_ACTION_VALID) != 0) {
2347                 switch (csio->tag_action) {
2348                 case MSG_HEAD_OF_Q_TAG:
2349                         bhssc->bhssc_flags |= BHSSC_FLAGS_ATTR_HOQ;
2350                         break;
2351                 case MSG_ORDERED_Q_TAG:
2352                         bhssc->bhssc_flags |= BHSSC_FLAGS_ATTR_ORDERED;
2353                         break;
2354                 case MSG_ACA_TASK:
2355                         bhssc->bhssc_flags |= BHSSC_FLAGS_ATTR_ACA;
2356                         break;
2357                 case MSG_SIMPLE_Q_TAG:
2358                 default:
2359                         bhssc->bhssc_flags |= BHSSC_FLAGS_ATTR_SIMPLE;
2360                         break;
2361                 }
2362         } else
2363                 bhssc->bhssc_flags |= BHSSC_FLAGS_ATTR_UNTAGGED;
2364
2365         if (is->is_protocol_level >= 2) {
2366                 bhssc->bhssc_pri = (csio->priority << BHSSC_PRI_SHIFT) &
2367                     BHSSC_PRI_MASK;
2368         }
2369
2370         bhssc->bhssc_lun = htobe64(CAM_EXTLUN_BYTE_SWIZZLE(ccb->ccb_h.target_lun));
2371         bhssc->bhssc_initiator_task_tag = initiator_task_tag;
2372         bhssc->bhssc_expected_data_transfer_length = htonl(csio->dxfer_len);
2373         KASSERT(csio->cdb_len <= sizeof(bhssc->bhssc_cdb),
2374             ("unsupported CDB size %zd", (size_t)csio->cdb_len));
2375
2376         if (csio->ccb_h.flags & CAM_CDB_POINTER)
2377                 memcpy(&bhssc->bhssc_cdb, csio->cdb_io.cdb_ptr, csio->cdb_len);
2378         else
2379                 memcpy(&bhssc->bhssc_cdb, csio->cdb_io.cdb_bytes, csio->cdb_len);
2380
2381         if (is->is_immediate_data &&
2382             (csio->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_OUT) {
2383                 len = csio->dxfer_len;
2384                 //ISCSI_SESSION_DEBUG(is, "adding %zd of immediate data", len);
2385                 if (len > is->is_first_burst_length) {
2386                         ISCSI_SESSION_DEBUG(is, "len %zd -> %d", len, is->is_first_burst_length);
2387                         len = is->is_first_burst_length;
2388                 }
2389                 if (len > is->is_conn->ic_max_send_data_segment_length) {
2390                         ISCSI_SESSION_DEBUG(is, "len %zd -> %d", len,
2391                             is->is_conn->ic_max_send_data_segment_length);
2392                         len = is->is_conn->ic_max_send_data_segment_length;
2393                 }
2394
2395                 error = icl_pdu_append_data(request, csio->data_ptr, len, M_NOWAIT);
2396                 if (error != 0) {
2397                         iscsi_outstanding_remove(is, io);
2398                         icl_pdu_free(request);
2399                         if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
2400                                 xpt_freeze_devq(ccb->ccb_h.path, 1);
2401                                 ISCSI_SESSION_DEBUG(is, "freezing devq");
2402                         }
2403                         ccb->ccb_h.status = CAM_RESRC_UNAVAIL | CAM_DEV_QFRZN;
2404                         xpt_done(ccb);
2405                         return;
2406                 }
2407         }
2408         iscsi_pdu_queue_locked(request);
2409 }
2410
2411 static void
2412 iscsi_action(struct cam_sim *sim, union ccb *ccb)
2413 {
2414         struct iscsi_session *is;
2415
2416         is = cam_sim_softc(sim);
2417
2418         ISCSI_SESSION_LOCK_ASSERT(is);
2419
2420         if (is->is_terminating ||
2421             (is->is_connected == false && fail_on_disconnection)) {
2422                 ccb->ccb_h.status = CAM_DEV_NOT_THERE;
2423                 xpt_done(ccb);
2424                 return;
2425         }
2426
2427         /*
2428          * Make sure CAM doesn't sneak in a CCB just after freezing the queue.
2429          */
2430         if (is->is_simq_frozen == true) {
2431                 ccb->ccb_h.status &= ~(CAM_SIM_QUEUED | CAM_STATUS_MASK);
2432                 ccb->ccb_h.status |= CAM_REQUEUE_REQ;
2433                 /* Don't freeze the devq - the SIM queue is already frozen. */
2434                 xpt_done(ccb);
2435                 return;
2436         }
2437
2438         switch (ccb->ccb_h.func_code) {
2439         case XPT_PATH_INQ:
2440         {
2441                 struct ccb_pathinq *cpi = &ccb->cpi;
2442
2443                 cpi->version_num = 1;
2444                 cpi->hba_inquiry = PI_TAG_ABLE;
2445                 cpi->target_sprt = 0;
2446                 cpi->hba_misc = PIM_EXTLUNS;
2447                 /*
2448                  * XXX: It shouldn't ever be NULL; this could be turned
2449                  *      into a KASSERT eventually.
2450                  */
2451                 if (is->is_conn == NULL)
2452                         ISCSI_WARN("NULL conn");
2453                 else if (is->is_conn->ic_unmapped)
2454                         cpi->hba_misc |= PIM_UNMAPPED;
2455                 cpi->hba_eng_cnt = 0;
2456                 cpi->max_target = 0;
2457                 /*
2458                  * Note that the variable below is only relevant for targets
2459                  * that don't claim compliance with anything above SPC2, which
2460                  * means they don't support REPORT_LUNS.
2461                  */
2462                 cpi->max_lun = 255;
2463                 cpi->initiator_id = ~0;
2464                 strlcpy(cpi->sim_vid, "FreeBSD", SIM_IDLEN);
2465                 strlcpy(cpi->hba_vid, "iSCSI", HBA_IDLEN);
2466                 strlcpy(cpi->dev_name, cam_sim_name(sim), DEV_IDLEN);
2467                 cpi->unit_number = cam_sim_unit(sim);
2468                 cpi->bus_id = cam_sim_bus(sim);
2469                 cpi->base_transfer_speed = 150000; /* XXX */
2470                 cpi->transport = XPORT_ISCSI;
2471                 cpi->transport_version = 0;
2472                 cpi->protocol = PROTO_SCSI;
2473                 cpi->protocol_version = SCSI_REV_SPC3;
2474                 cpi->maxio = maxphys;
2475                 cpi->ccb_h.status = CAM_REQ_CMP;
2476                 break;
2477         }
2478         case XPT_GET_TRAN_SETTINGS:
2479         {
2480                 struct ccb_trans_settings       *cts;
2481                 struct ccb_trans_settings_scsi  *scsi;
2482
2483                 cts = &ccb->cts;
2484                 scsi = &cts->proto_specific.scsi;
2485
2486                 cts->protocol = PROTO_SCSI;
2487                 cts->protocol_version = SCSI_REV_SPC3;
2488                 cts->transport = XPORT_ISCSI;
2489                 cts->transport_version = 0;
2490                 scsi->valid = CTS_SCSI_VALID_TQ;
2491                 scsi->flags = CTS_SCSI_FLAGS_TAG_ENB;
2492                 cts->ccb_h.status = CAM_REQ_CMP;
2493                 break;
2494         }
2495         case XPT_CALC_GEOMETRY:
2496                 cam_calc_geometry(&ccb->ccg, /*extended*/1);
2497                 ccb->ccb_h.status = CAM_REQ_CMP;
2498                 break;
2499 #if 0
2500         /*
2501          * XXX: What's the point?
2502          */
2503         case XPT_RESET_BUS:
2504         case XPT_TERM_IO:
2505                 ISCSI_SESSION_DEBUG(is, "faking success for reset, abort, or term_io");
2506                 ccb->ccb_h.status = CAM_REQ_CMP;
2507                 break;
2508 #endif
2509         case XPT_ABORT:
2510                 iscsi_action_abort(is, ccb);
2511                 return;
2512         case XPT_SCSI_IO:
2513                 iscsi_action_scsiio(is, ccb);
2514                 return;
2515         default:
2516 #if 0
2517                 ISCSI_SESSION_DEBUG(is, "got unsupported code 0x%x", ccb->ccb_h.func_code);
2518 #endif
2519                 ccb->ccb_h.status = CAM_FUNC_NOTAVAIL;
2520                 break;
2521         }
2522         xpt_done(ccb);
2523 }
2524
2525 static void
2526 iscsi_terminate_sessions(struct iscsi_softc *sc)
2527 {
2528         struct iscsi_session *is;
2529
2530         sx_slock(&sc->sc_lock);
2531         TAILQ_FOREACH(is, &sc->sc_sessions, is_next)
2532                 iscsi_session_terminate(is);
2533         while(!TAILQ_EMPTY(&sc->sc_sessions)) {
2534                 ISCSI_DEBUG("waiting for sessions to terminate");
2535                 cv_wait(&sc->sc_cv, &sc->sc_lock);
2536         }
2537         ISCSI_DEBUG("all sessions terminated");
2538         sx_sunlock(&sc->sc_lock);
2539 }
2540
2541 static void
2542 iscsi_shutdown_pre(struct iscsi_softc *sc)
2543 {
2544         struct iscsi_session *is;
2545
2546         if (!fail_on_shutdown)
2547                 return;
2548
2549         /*
2550          * If we have any sessions waiting for reconnection, request
2551          * maintenance thread to fail them immediately instead of waiting
2552          * for reconnect timeout.
2553          *
2554          * This prevents LUNs with mounted filesystems that are supported
2555          * by disconnected iSCSI sessions from hanging, however it will
2556          * fail all queued BIOs.
2557          */
2558         ISCSI_DEBUG("forcing failing all disconnected sessions due to shutdown");
2559
2560         fail_on_disconnection = 1;
2561
2562         sx_slock(&sc->sc_lock);
2563         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
2564                 ISCSI_SESSION_LOCK(is);
2565                 if (!is->is_connected) {
2566                         ISCSI_SESSION_DEBUG(is, "force failing disconnected session early");
2567                         iscsi_session_reconnect(is);
2568                 }
2569                 ISCSI_SESSION_UNLOCK(is);
2570         }
2571         sx_sunlock(&sc->sc_lock);
2572 }
2573
2574 static void
2575 iscsi_shutdown_post(struct iscsi_softc *sc)
2576 {
2577
2578         if (!KERNEL_PANICKED()) {
2579                 ISCSI_DEBUG("removing all sessions due to shutdown");
2580                 iscsi_terminate_sessions(sc);
2581         }
2582 }
2583
2584 static int
2585 iscsi_load(void)
2586 {
2587         int error;
2588
2589         sc = malloc(sizeof(*sc), M_ISCSI, M_ZERO | M_WAITOK);
2590         sx_init(&sc->sc_lock, "iscsi");
2591         TAILQ_INIT(&sc->sc_sessions);
2592         cv_init(&sc->sc_cv, "iscsi_cv");
2593
2594         iscsi_outstanding_zone = uma_zcreate("iscsi_outstanding",
2595             sizeof(struct iscsi_outstanding), NULL, NULL, NULL, NULL,
2596             UMA_ALIGN_PTR, 0);
2597
2598         error = make_dev_p(MAKEDEV_CHECKNAME, &sc->sc_cdev, &iscsi_cdevsw,
2599             NULL, UID_ROOT, GID_WHEEL, 0600, "iscsi");
2600         if (error != 0) {
2601                 ISCSI_WARN("failed to create device node, error %d", error);
2602                 return (error);
2603         }
2604         sc->sc_cdev->si_drv1 = sc;
2605
2606         sc->sc_shutdown_pre_eh = EVENTHANDLER_REGISTER(shutdown_pre_sync,
2607             iscsi_shutdown_pre, sc, SHUTDOWN_PRI_FIRST);
2608         /*
2609          * shutdown_post_sync needs to run after filesystem shutdown and before
2610          * CAM shutdown - otherwise when rebooting with an iSCSI session that is
2611          * disconnected but has outstanding requests, dashutdown() will hang on
2612          * cam_periph_runccb().
2613          */
2614         sc->sc_shutdown_post_eh = EVENTHANDLER_REGISTER(shutdown_post_sync,
2615             iscsi_shutdown_post, sc, SHUTDOWN_PRI_DEFAULT - 1);
2616
2617         return (0);
2618 }
2619
2620 static int
2621 iscsi_unload(void)
2622 {
2623
2624         /* Awaken any threads asleep in iscsi_ioctl(). */
2625         sx_xlock(&sc->sc_lock);
2626         sc->sc_unloading = true;
2627         cv_signal(&sc->sc_cv);
2628         sx_xunlock(&sc->sc_lock);
2629
2630         if (sc->sc_cdev != NULL) {
2631                 ISCSI_DEBUG("removing device node");
2632                 destroy_dev(sc->sc_cdev);
2633                 ISCSI_DEBUG("device node removed");
2634         }
2635
2636         if (sc->sc_shutdown_pre_eh != NULL)
2637                 EVENTHANDLER_DEREGISTER(shutdown_pre_sync, sc->sc_shutdown_pre_eh);
2638         if (sc->sc_shutdown_post_eh != NULL)
2639                 EVENTHANDLER_DEREGISTER(shutdown_post_sync, sc->sc_shutdown_post_eh);
2640
2641         iscsi_terminate_sessions(sc);
2642
2643         uma_zdestroy(iscsi_outstanding_zone);
2644         sx_destroy(&sc->sc_lock);
2645         cv_destroy(&sc->sc_cv);
2646         free(sc, M_ISCSI);
2647         return (0);
2648 }
2649
2650 static int
2651 iscsi_quiesce(void)
2652 {
2653         sx_slock(&sc->sc_lock);
2654         if (!TAILQ_EMPTY(&sc->sc_sessions)) {
2655                 sx_sunlock(&sc->sc_lock);
2656                 return (EBUSY);
2657         }
2658         sx_sunlock(&sc->sc_lock);
2659         return (0);
2660 }
2661
2662 static int
2663 iscsi_modevent(module_t mod, int what, void *arg)
2664 {
2665         int error;
2666
2667         switch (what) {
2668         case MOD_LOAD:
2669                 error = iscsi_load();
2670                 break;
2671         case MOD_UNLOAD:
2672                 error = iscsi_unload();
2673                 break;
2674         case MOD_QUIESCE:
2675                 error = iscsi_quiesce();
2676                 break;
2677         default:
2678                 error = EINVAL;
2679                 break;
2680         }
2681         return (error);
2682 }
2683
2684 moduledata_t iscsi_data = {
2685         "iscsi",
2686         iscsi_modevent,
2687         0
2688 };
2689
2690 DECLARE_MODULE(iscsi, iscsi_data, SI_SUB_DRIVERS, SI_ORDER_MIDDLE);
2691 MODULE_DEPEND(iscsi, cam, 1, 1, 1);
2692 MODULE_DEPEND(iscsi, icl, 1, 1, 1);