sys/dev/iscsi/iscsi.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2012 The FreeBSD Foundation
   5  *
   6  * This software was developed by Edward Tomasz Napierala under sponsorship
   7  * from the FreeBSD Foundation.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  28  * SUCH DAMAGE.
  29  *
  30  */
  31
  32 #include <sys/cdefs.h>
  33 __FBSDID("$FreeBSD$");
  34
  35 #include <sys/param.h>
  36 #include <sys/condvar.h>
  37 #include <sys/conf.h>
  38 #include <sys/endian.h>
  39 #include <sys/eventhandler.h>
  40 #include <sys/file.h>
  41 #include <sys/kernel.h>
  42 #include <sys/kthread.h>
  43 #include <sys/lock.h>
  44 #include <sys/malloc.h>
  45 #include <sys/mutex.h>
  46 #include <sys/module.h>
  47 #include <sys/socket.h>
  48 #include <sys/sysctl.h>
  49 #include <sys/systm.h>
  50 #include <sys/sx.h>
  51 #include <vm/uma.h>
  52
  53 #include <cam/cam.h>
  54 #include <cam/cam_ccb.h>
  55 #include <cam/cam_xpt.h>
  56 #include <cam/cam_debug.h>
  57 #include <cam/cam_sim.h>
  58 #include <cam/cam_xpt_sim.h>
  59 #include <cam/cam_xpt_periph.h>
  60 #include <cam/cam_periph.h>
  61 #include <cam/scsi/scsi_all.h>
  62 #include <cam/scsi/scsi_message.h>
  63
  64 #include <dev/iscsi/icl.h>
  65 #include <dev/iscsi/icl_wrappers.h>
  66 #include <dev/iscsi/iscsi_ioctl.h>
  67 #include <dev/iscsi/iscsi_proto.h>
  68 #include <dev/iscsi/iscsi.h>
  69
  70 #ifdef ICL_KERNEL_PROXY
  71 #include <sys/socketvar.h>
  72 #endif
  73
  74 #ifdef ICL_KERNEL_PROXY
  75 FEATURE(iscsi_kernel_proxy, "iSCSI initiator built with ICL_KERNEL_PROXY");
  76 #endif
  77
  78 /*
  79  * XXX: This is global so the iscsi_unload() can access it.
  80  *      Think about how to do this properly.
  81  */
  82 static struct iscsi_softc       *sc;
  83
  84 SYSCTL_NODE(_kern, OID_AUTO, iscsi, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
  85     "iSCSI initiator");
  86 static int debug = 1;
  87 SYSCTL_INT(_kern_iscsi, OID_AUTO, debug, CTLFLAG_RWTUN,
  88     &debug, 0, "Enable debug messages");
  89 static int ping_timeout = 5;
  90 SYSCTL_INT(_kern_iscsi, OID_AUTO, ping_timeout, CTLFLAG_RWTUN, &ping_timeout,
  91     0, "Timeout for ping (NOP-Out) requests, in seconds");
  92 static int iscsid_timeout = 60;
  93 SYSCTL_INT(_kern_iscsi, OID_AUTO, iscsid_timeout, CTLFLAG_RWTUN, &iscsid_timeout,
  94     0, "Time to wait for iscsid(8) to handle reconnection, in seconds");
  95 static int login_timeout = 60;
  96 SYSCTL_INT(_kern_iscsi, OID_AUTO, login_timeout, CTLFLAG_RWTUN, &login_timeout,
  97     0, "Time to wait for iscsid(8) to finish Login Phase, in seconds");
  98 static int maxtags = 255;
  99 SYSCTL_INT(_kern_iscsi, OID_AUTO, maxtags, CTLFLAG_RWTUN, &maxtags,
 100     0, "Max number of IO requests queued");
 101 static int fail_on_disconnection = 0;
 102 SYSCTL_INT(_kern_iscsi, OID_AUTO, fail_on_disconnection, CTLFLAG_RWTUN,
 103     &fail_on_disconnection, 0, "Destroy CAM SIM on connection failure");
 104 static int fail_on_shutdown = 1;
 105 SYSCTL_INT(_kern_iscsi, OID_AUTO, fail_on_shutdown, CTLFLAG_RWTUN,
 106     &fail_on_shutdown, 0, "Fail disconnected sessions on shutdown");
 107
 108 static MALLOC_DEFINE(M_ISCSI, "iSCSI", "iSCSI initiator");
 109 static uma_zone_t iscsi_outstanding_zone;
 110
 111 #define CONN_SESSION(X) ((struct iscsi_session *)X->ic_prv0)
 112 #define PDU_SESSION(X)  (CONN_SESSION(X->ip_conn))
 113
 114 #define ISCSI_DEBUG(X, ...)                                             \
 115         do {                                                            \
 116                 if (debug > 1)                                          \
 117                         printf("%s: " X "\n", __func__, ## __VA_ARGS__);\
 118         } while (0)
 119
 120 #define ISCSI_WARN(X, ...)                                              \
 121         do {                                                            \
 122                 if (debug > 0) {                                        \
 123                         printf("WARNING: %s: " X "\n",                  \
 124                             __func__, ## __VA_ARGS__);                  \
 125                 }                                                       \
 126         } while (0)
 127
 128 #define ISCSI_SESSION_DEBUG(S, X, ...)                                  \
 129         do {                                                            \
 130                 if (debug > 1) {                                        \
 131                         printf("%s: %s (%s): " X "\n",                  \
 132                             __func__, S->is_conf.isc_target_addr,       \
 133                             S->is_conf.isc_target, ## __VA_ARGS__);     \
 134                 }                                                       \
 135         } while (0)
 136
 137 #define ISCSI_SESSION_WARN(S, X, ...)                                   \
 138         do {                                                            \
 139                 if (debug > 0) {                                        \
 140                         printf("WARNING: %s (%s): " X "\n",             \
 141                             S->is_conf.isc_target_addr,                 \
 142                             S->is_conf.isc_target, ## __VA_ARGS__);     \
 143                 }                                                       \
 144         } while (0)
 145
 146 #define ISCSI_SESSION_LOCK(X)           mtx_lock(&X->is_lock)
 147 #define ISCSI_SESSION_UNLOCK(X)         mtx_unlock(&X->is_lock)
 148 #define ISCSI_SESSION_LOCK_ASSERT(X)    mtx_assert(&X->is_lock, MA_OWNED)
 149 #define ISCSI_SESSION_LOCK_ASSERT_NOT(X) mtx_assert(&X->is_lock, MA_NOTOWNED)
 150
 151 static int      iscsi_ioctl(struct cdev *dev, u_long cmd, caddr_t arg,
 152                     int mode, struct thread *td);
 153
 154 static struct cdevsw iscsi_cdevsw = {
 155      .d_version = D_VERSION,
 156      .d_ioctl   = iscsi_ioctl,
 157      .d_name    = "iscsi",
 158 };
 159
 160 static void     iscsi_pdu_queue_locked(struct icl_pdu *request);
 161 static void     iscsi_pdu_queue(struct icl_pdu *request);
 162 static void     iscsi_pdu_update_statsn(const struct icl_pdu *response);
 163 static void     iscsi_pdu_handle_nop_in(struct icl_pdu *response);
 164 static void     iscsi_pdu_handle_scsi_response(struct icl_pdu *response);
 165 static void     iscsi_pdu_handle_task_response(struct icl_pdu *response);
 166 static void     iscsi_pdu_handle_data_in(struct icl_pdu *response);
 167 static void     iscsi_pdu_handle_logout_response(struct icl_pdu *response);
 168 static void     iscsi_pdu_handle_r2t(struct icl_pdu *response);
 169 static void     iscsi_pdu_handle_async_message(struct icl_pdu *response);
 170 static void     iscsi_pdu_handle_reject(struct icl_pdu *response);
 171 static void     iscsi_session_reconnect(struct iscsi_session *is);
 172 static void     iscsi_session_terminate(struct iscsi_session *is);
 173 static void     iscsi_action(struct cam_sim *sim, union ccb *ccb);
 174 static struct iscsi_outstanding *iscsi_outstanding_find(struct iscsi_session *is,
 175                     uint32_t initiator_task_tag);
 176 static struct iscsi_outstanding *iscsi_outstanding_add(struct iscsi_session *is,
 177                     struct icl_pdu *request, union ccb *ccb,
 178                     uint32_t *initiator_task_tagp);
 179 static void     iscsi_outstanding_remove(struct iscsi_session *is,
 180                     struct iscsi_outstanding *io);
 181
 182 static bool
 183 iscsi_pdu_prepare(struct icl_pdu *request)
 184 {
 185         struct iscsi_session *is;
 186         struct iscsi_bhs_scsi_command *bhssc;
 187
 188         is = PDU_SESSION(request);
 189
 190         ISCSI_SESSION_LOCK_ASSERT(is);
 191
 192         /*
 193          * We're only using fields common for all the request
 194          * (initiator -> target) PDUs.
 195          */
 196         bhssc = (struct iscsi_bhs_scsi_command *)request->ip_bhs;
 197
 198         /*
 199          * Data-Out PDU does not contain CmdSN.
 200          */
 201         if (bhssc->bhssc_opcode != ISCSI_BHS_OPCODE_SCSI_DATA_OUT) {
 202                 if (ISCSI_SNGT(is->is_cmdsn, is->is_maxcmdsn) &&
 203                     (bhssc->bhssc_opcode & ISCSI_BHS_OPCODE_IMMEDIATE) == 0) {
 204                         /*
 205                          * Current MaxCmdSN prevents us from sending any more
 206                          * SCSI Command PDUs to the target; postpone the PDU.
 207                          * It will get resent by either iscsi_pdu_queue(),
 208                          * or by maintenance thread.
 209                          */
 210 #if 0
 211                         ISCSI_SESSION_DEBUG(is, "postponing send, CmdSN %u, "
 212                             "ExpCmdSN %u, MaxCmdSN %u, opcode 0x%x",
 213                             is->is_cmdsn, is->is_expcmdsn, is->is_maxcmdsn,
 214                             bhssc->bhssc_opcode);
 215 #endif
 216                         return (true);
 217                 }
 218                 bhssc->bhssc_cmdsn = htonl(is->is_cmdsn);
 219                 if ((bhssc->bhssc_opcode & ISCSI_BHS_OPCODE_IMMEDIATE) == 0)
 220                         is->is_cmdsn++;
 221         }
 222         bhssc->bhssc_expstatsn = htonl(is->is_statsn + 1);
 223
 224         return (false);
 225 }
 226
 227 static void
 228 iscsi_session_send_postponed(struct iscsi_session *is)
 229 {
 230         struct icl_pdu *request;
 231         bool postpone;
 232
 233         ISCSI_SESSION_LOCK_ASSERT(is);
 234
 235         if (STAILQ_EMPTY(&is->is_postponed))
 236                 return;
 237         while ((request = STAILQ_FIRST(&is->is_postponed)) != NULL) {
 238                 postpone = iscsi_pdu_prepare(request);
 239                 if (postpone)
 240                         return;
 241                 STAILQ_REMOVE_HEAD(&is->is_postponed, ip_next);
 242                 icl_pdu_queue(request);
 243         }
 244         xpt_release_simq(is->is_sim, 1);
 245 }
 246
 247 static void
 248 iscsi_pdu_queue_locked(struct icl_pdu *request)
 249 {
 250         struct iscsi_session *is;
 251         bool postpone;
 252
 253         is = PDU_SESSION(request);
 254         ISCSI_SESSION_LOCK_ASSERT(is);
 255         iscsi_session_send_postponed(is);
 256         postpone = iscsi_pdu_prepare(request);
 257         if (postpone) {
 258                 if (STAILQ_EMPTY(&is->is_postponed))
 259                         xpt_freeze_simq(is->is_sim, 1);
 260                 STAILQ_INSERT_TAIL(&is->is_postponed, request, ip_next);
 261                 return;
 262         }
 263         icl_pdu_queue(request);
 264 }
 265
 266 static void
 267 iscsi_pdu_queue(struct icl_pdu *request)
 268 {
 269         struct iscsi_session *is;
 270
 271         is = PDU_SESSION(request);
 272         ISCSI_SESSION_LOCK(is);
 273         iscsi_pdu_queue_locked(request);
 274         ISCSI_SESSION_UNLOCK(is);
 275 }
 276
 277 static void
 278 iscsi_session_logout(struct iscsi_session *is)
 279 {
 280         struct icl_pdu *request;
 281         struct iscsi_bhs_logout_request *bhslr;
 282
 283         request = icl_pdu_new(is->is_conn, M_NOWAIT);
 284         if (request == NULL)
 285                 return;
 286
 287         bhslr = (struct iscsi_bhs_logout_request *)request->ip_bhs;
 288         bhslr->bhslr_opcode = ISCSI_BHS_OPCODE_LOGOUT_REQUEST;
 289         bhslr->bhslr_reason = BHSLR_REASON_CLOSE_SESSION;
 290         iscsi_pdu_queue_locked(request);
 291 }
 292
 293 static void
 294 iscsi_session_terminate_task(struct iscsi_session *is,
 295     struct iscsi_outstanding *io, cam_status status)
 296 {
 297
 298         ISCSI_SESSION_LOCK_ASSERT(is);
 299
 300         if (io->io_ccb != NULL) {
 301                 io->io_ccb->ccb_h.status &= ~(CAM_SIM_QUEUED | CAM_STATUS_MASK);
 302                 io->io_ccb->ccb_h.status |= status;
 303                 if ((io->io_ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
 304                         io->io_ccb->ccb_h.status |= CAM_DEV_QFRZN;
 305                         xpt_freeze_devq(io->io_ccb->ccb_h.path, 1);
 306                         ISCSI_SESSION_DEBUG(is, "freezing devq");
 307                 }
 308                 xpt_done(io->io_ccb);
 309         }
 310         iscsi_outstanding_remove(is, io);
 311 }
 312
 313 static void
 314 iscsi_session_terminate_tasks(struct iscsi_session *is, cam_status status)
 315 {
 316         struct iscsi_outstanding *io, *tmp;
 317
 318         ISCSI_SESSION_LOCK_ASSERT(is);
 319
 320         TAILQ_FOREACH_SAFE(io, &is->is_outstanding, io_next, tmp) {
 321                 iscsi_session_terminate_task(is, io, status);
 322         }
 323 }
 324
 325 static void
 326 iscsi_session_cleanup(struct iscsi_session *is, bool destroy_sim)
 327 {
 328         struct icl_pdu *pdu;
 329
 330         ISCSI_SESSION_LOCK_ASSERT(is);
 331
 332         /*
 333          * Don't queue any new PDUs.
 334          */
 335         if (is->is_sim != NULL && is->is_simq_frozen == false) {
 336                 ISCSI_SESSION_DEBUG(is, "freezing");
 337                 xpt_freeze_simq(is->is_sim, 1);
 338                 is->is_simq_frozen = true;
 339         }
 340
 341         /*
 342          * Remove postponed PDUs.
 343          */
 344         if (!STAILQ_EMPTY(&is->is_postponed))
 345                 xpt_release_simq(is->is_sim, 1);
 346         while ((pdu = STAILQ_FIRST(&is->is_postponed)) != NULL) {
 347                 STAILQ_REMOVE_HEAD(&is->is_postponed, ip_next);
 348                 icl_pdu_free(pdu);
 349         }
 350
 351         if (destroy_sim == false) {
 352                 /*
 353                  * Terminate SCSI tasks, asking CAM to requeue them.
 354                  */
 355                 iscsi_session_terminate_tasks(is, CAM_REQUEUE_REQ);
 356                 return;
 357         }
 358
 359         iscsi_session_terminate_tasks(is, CAM_DEV_NOT_THERE);
 360
 361         if (is->is_sim == NULL)
 362                 return;
 363
 364         ISCSI_SESSION_DEBUG(is, "deregistering SIM");
 365         xpt_async(AC_LOST_DEVICE, is->is_path, NULL);
 366
 367         if (is->is_simq_frozen) {
 368                 is->is_simq_frozen = false;
 369                 xpt_release_simq(is->is_sim, 1);
 370         }
 371
 372         xpt_free_path(is->is_path);
 373         is->is_path = NULL;
 374         xpt_bus_deregister(cam_sim_path(is->is_sim));
 375         cam_sim_free(is->is_sim, TRUE /*free_devq*/);
 376         is->is_sim = NULL;
 377         is->is_devq = NULL;
 378 }
 379
 380 static void
 381 iscsi_maintenance_thread_reconnect(struct iscsi_session *is)
 382 {
 383
 384         icl_conn_close(is->is_conn);
 385
 386         ISCSI_SESSION_LOCK(is);
 387
 388         is->is_connected = false;
 389         is->is_reconnecting = false;
 390         is->is_login_phase = false;
 391
 392 #ifdef ICL_KERNEL_PROXY
 393         if (is->is_login_pdu != NULL) {
 394                 icl_pdu_free(is->is_login_pdu);
 395                 is->is_login_pdu = NULL;
 396         }
 397         cv_signal(&is->is_login_cv);
 398 #endif
 399
 400         if (fail_on_disconnection) {
 401                 ISCSI_SESSION_DEBUG(is, "connection failed, destroying devices");
 402                 iscsi_session_cleanup(is, true);
 403         } else {
 404                 iscsi_session_cleanup(is, false);
 405         }
 406
 407         KASSERT(TAILQ_EMPTY(&is->is_outstanding),
 408             ("destroying session with active tasks"));
 409         KASSERT(STAILQ_EMPTY(&is->is_postponed),
 410             ("destroying session with postponed PDUs"));
 411
 412         if (is->is_conf.isc_enable == 0 && is->is_conf.isc_discovery == 0) {
 413                 ISCSI_SESSION_UNLOCK(is);
 414                 return;
 415         }
 416
 417         /*
 418          * Request immediate reconnection from iscsid(8).
 419          */
 420         //ISCSI_SESSION_DEBUG(is, "waking up iscsid(8)");
 421         is->is_waiting_for_iscsid = true;
 422         strlcpy(is->is_reason, "Waiting for iscsid(8)", sizeof(is->is_reason));
 423         is->is_timeout = 0;
 424         ISCSI_SESSION_UNLOCK(is);
 425         cv_signal(&is->is_softc->sc_cv);
 426 }
 427
 428 static void
 429 iscsi_maintenance_thread_terminate(struct iscsi_session *is)
 430 {
 431         struct iscsi_softc *sc;
 432
 433         sc = is->is_softc;
 434         sx_xlock(&sc->sc_lock);
 435         TAILQ_REMOVE(&sc->sc_sessions, is, is_next);
 436         sx_xunlock(&sc->sc_lock);
 437
 438         icl_conn_close(is->is_conn);
 439         callout_drain(&is->is_callout);
 440
 441         ISCSI_SESSION_LOCK(is);
 442
 443         KASSERT(is->is_terminating, ("is_terminating == false"));
 444
 445 #ifdef ICL_KERNEL_PROXY
 446         if (is->is_login_pdu != NULL) {
 447                 icl_pdu_free(is->is_login_pdu);
 448                 is->is_login_pdu = NULL;
 449         }
 450         cv_signal(&is->is_login_cv);
 451 #endif
 452
 453         iscsi_session_cleanup(is, true);
 454
 455         KASSERT(TAILQ_EMPTY(&is->is_outstanding),
 456             ("destroying session with active tasks"));
 457         KASSERT(STAILQ_EMPTY(&is->is_postponed),
 458             ("destroying session with postponed PDUs"));
 459
 460         ISCSI_SESSION_UNLOCK(is);
 461
 462         icl_conn_free(is->is_conn);
 463         mtx_destroy(&is->is_lock);
 464         cv_destroy(&is->is_maintenance_cv);
 465 #ifdef ICL_KERNEL_PROXY
 466         cv_destroy(&is->is_login_cv);
 467 #endif
 468
 469         ISCSI_SESSION_DEBUG(is, "terminated");
 470         free(is, M_ISCSI);
 471
 472         /*
 473          * The iscsi_unload() routine might be waiting.
 474          */
 475         cv_signal(&sc->sc_cv);
 476 }
 477
 478 static void
 479 iscsi_maintenance_thread(void *arg)
 480 {
 481         struct iscsi_session *is = arg;
 482
 483         ISCSI_SESSION_LOCK(is);
 484         for (;;) {
 485                 if (is->is_reconnecting == false &&
 486                     is->is_terminating == false &&
 487                     (STAILQ_EMPTY(&is->is_postponed) ||
 488                      ISCSI_SNGT(is->is_cmdsn, is->is_maxcmdsn)))
 489                         cv_wait(&is->is_maintenance_cv, &is->is_lock);
 490
 491                 /* Terminate supersedes reconnect. */
 492                 if (is->is_terminating) {
 493                         ISCSI_SESSION_UNLOCK(is);
 494                         iscsi_maintenance_thread_terminate(is);
 495                         kthread_exit();
 496                         return;
 497                 }
 498
 499                 if (is->is_reconnecting) {
 500                         ISCSI_SESSION_UNLOCK(is);
 501                         iscsi_maintenance_thread_reconnect(is);
 502                         ISCSI_SESSION_LOCK(is);
 503                         continue;
 504                 }
 505
 506                 iscsi_session_send_postponed(is);
 507         }
 508         ISCSI_SESSION_UNLOCK(is);
 509 }
 510
 511 static void
 512 iscsi_session_reconnect(struct iscsi_session *is)
 513 {
 514
 515         /*
 516          * XXX: We can't use locking here, because
 517          *      it's being called from various contexts.
 518          *      Hope it doesn't break anything.
 519          */
 520         if (is->is_reconnecting)
 521                 return;
 522
 523         is->is_reconnecting = true;
 524         cv_signal(&is->is_maintenance_cv);
 525 }
 526
 527 static void
 528 iscsi_session_terminate(struct iscsi_session *is)
 529 {
 530
 531         if (is->is_terminating)
 532                 return;
 533
 534         is->is_terminating = true;
 535
 536 #if 0
 537         iscsi_session_logout(is);
 538 #endif
 539         cv_signal(&is->is_maintenance_cv);
 540 }
 541
 542 static void
 543 iscsi_callout(void *context)
 544 {
 545         struct icl_pdu *request;
 546         struct iscsi_bhs_nop_out *bhsno;
 547         struct iscsi_session *is;
 548         bool reconnect_needed = false;
 549
 550         is = context;
 551
 552         ISCSI_SESSION_LOCK(is);
 553         if (is->is_terminating) {
 554                 ISCSI_SESSION_UNLOCK(is);
 555                 return;
 556         }
 557
 558         callout_schedule(&is->is_callout, 1 * hz);
 559
 560         if (is->is_conf.isc_enable == 0)
 561                 goto out;
 562
 563         is->is_timeout++;
 564
 565         if (is->is_waiting_for_iscsid) {
 566                 if (iscsid_timeout > 0 && is->is_timeout > iscsid_timeout) {
 567                         ISCSI_SESSION_WARN(is, "timed out waiting for iscsid(8) "
 568                             "for %d seconds; reconnecting",
 569                             is->is_timeout);
 570                         reconnect_needed = true;
 571                 }
 572                 goto out;
 573         }
 574
 575         if (is->is_login_phase) {
 576                 if (login_timeout > 0 && is->is_timeout > login_timeout) {
 577                         ISCSI_SESSION_WARN(is, "login timed out after %d seconds; "
 578                             "reconnecting", is->is_timeout);
 579                         reconnect_needed = true;
 580                 }
 581                 goto out;
 582         }
 583
 584         if (ping_timeout <= 0) {
 585                 /*
 586                  * Pings are disabled.  Don't send NOP-Out in this case.
 587                  * Reset the timeout, to avoid triggering reconnection,
 588                  * should the user decide to reenable them.
 589                  */
 590                 is->is_timeout = 0;
 591                 goto out;
 592         }
 593
 594         if (is->is_timeout >= ping_timeout) {
 595                 ISCSI_SESSION_WARN(is, "no ping reply (NOP-In) after %d seconds; "
 596                     "reconnecting", ping_timeout);
 597                 reconnect_needed = true;
 598                 goto out;
 599         }
 600
 601         ISCSI_SESSION_UNLOCK(is);
 602
 603         /*
 604          * If the ping was reset less than one second ago - which means
 605          * that we've received some PDU during the last second - assume
 606          * the traffic flows correctly and don't bother sending a NOP-Out.
 607          *
 608          * (It's 2 - one for one second, and one for incrementing is_timeout
 609          * earlier in this routine.)
 610          */
 611         if (is->is_timeout < 2)
 612                 return;
 613
 614         request = icl_pdu_new(is->is_conn, M_NOWAIT);
 615         if (request == NULL) {
 616                 ISCSI_SESSION_WARN(is, "failed to allocate PDU");
 617                 return;
 618         }
 619         bhsno = (struct iscsi_bhs_nop_out *)request->ip_bhs;
 620         bhsno->bhsno_opcode = ISCSI_BHS_OPCODE_NOP_OUT |
 621             ISCSI_BHS_OPCODE_IMMEDIATE;
 622         bhsno->bhsno_flags = 0x80;
 623         bhsno->bhsno_target_transfer_tag = 0xffffffff;
 624         iscsi_pdu_queue(request);
 625         return;
 626
 627 out:
 628         if (is->is_terminating) {
 629                 ISCSI_SESSION_UNLOCK(is);
 630                 return;
 631         }
 632
 633         ISCSI_SESSION_UNLOCK(is);
 634
 635         if (reconnect_needed)
 636                 iscsi_session_reconnect(is);
 637 }
 638
 639 static void
 640 iscsi_pdu_update_statsn(const struct icl_pdu *response)
 641 {
 642         const struct iscsi_bhs_data_in *bhsdi;
 643         struct iscsi_session *is;
 644         uint32_t expcmdsn, maxcmdsn, statsn;
 645
 646         is = PDU_SESSION(response);
 647
 648         ISCSI_SESSION_LOCK_ASSERT(is);
 649
 650         /*
 651          * We're only using fields common for all the response
 652          * (target -> initiator) PDUs.
 653          */
 654         bhsdi = (const struct iscsi_bhs_data_in *)response->ip_bhs;
 655         /*
 656          * Ok, I lied.  In case of Data-In, "The fields StatSN, Status,
 657          * and Residual Count only have meaningful content if the S bit
 658          * is set to 1", so we also need to check the bit specific for
 659          * Data-In PDU.
 660          */
 661         if (bhsdi->bhsdi_opcode != ISCSI_BHS_OPCODE_SCSI_DATA_IN ||
 662             (bhsdi->bhsdi_flags & BHSDI_FLAGS_S) != 0) {
 663                 statsn = ntohl(bhsdi->bhsdi_statsn);
 664                 if (statsn != is->is_statsn && statsn != (is->is_statsn + 1)) {
 665                         /* XXX: This is normal situation for MCS */
 666                         ISCSI_SESSION_WARN(is, "PDU 0x%x StatSN %u != "
 667                             "session ExpStatSN %u (or + 1); reconnecting",
 668                             bhsdi->bhsdi_opcode, statsn, is->is_statsn);
 669                         iscsi_session_reconnect(is);
 670                 }
 671                 if (ISCSI_SNGT(statsn, is->is_statsn))
 672                         is->is_statsn = statsn;
 673         }
 674
 675         expcmdsn = ntohl(bhsdi->bhsdi_expcmdsn);
 676         maxcmdsn = ntohl(bhsdi->bhsdi_maxcmdsn);
 677
 678         if (ISCSI_SNLT(maxcmdsn + 1, expcmdsn)) {
 679                 ISCSI_SESSION_DEBUG(is,
 680                     "PDU MaxCmdSN %u + 1 < PDU ExpCmdSN %u; ignoring",
 681                     maxcmdsn, expcmdsn);
 682         } else {
 683                 if (ISCSI_SNGT(maxcmdsn, is->is_maxcmdsn)) {
 684                         is->is_maxcmdsn = maxcmdsn;
 685
 686                         /*
 687                          * Command window increased; kick the maintanance thread
 688                          * to send out postponed commands.
 689                          */
 690                         if (!STAILQ_EMPTY(&is->is_postponed))
 691                                 cv_signal(&is->is_maintenance_cv);
 692                 } else if (ISCSI_SNLT(maxcmdsn, is->is_maxcmdsn)) {
 693                         /* XXX: This is normal situation for MCS */
 694                         ISCSI_SESSION_DEBUG(is,
 695                             "PDU MaxCmdSN %u < session MaxCmdSN %u; ignoring",
 696                             maxcmdsn, is->is_maxcmdsn);
 697                 }
 698
 699                 if (ISCSI_SNGT(expcmdsn, is->is_expcmdsn)) {
 700                         is->is_expcmdsn = expcmdsn;
 701                 } else if (ISCSI_SNLT(expcmdsn, is->is_expcmdsn)) {
 702                         /* XXX: This is normal situation for MCS */
 703                         ISCSI_SESSION_DEBUG(is,
 704                             "PDU ExpCmdSN %u < session ExpCmdSN %u; ignoring",
 705                             expcmdsn, is->is_expcmdsn);
 706                 }
 707         }
 708
 709         /*
 710          * Every incoming PDU - not just NOP-In - resets the ping timer.
 711          * The purpose of the timeout is to reset the connection when it stalls;
 712          * we don't want this to happen when NOP-In or NOP-Out ends up delayed
 713          * in some queue.
 714          */
 715         is->is_timeout = 0;
 716 }
 717
 718 static void
 719 iscsi_receive_callback(struct icl_pdu *response)
 720 {
 721         struct iscsi_session *is;
 722
 723         is = PDU_SESSION(response);
 724
 725         ISCSI_SESSION_LOCK(is);
 726
 727         iscsi_pdu_update_statsn(response);
 728
 729 #ifdef ICL_KERNEL_PROXY
 730         if (is->is_login_phase) {
 731                 if (is->is_login_pdu == NULL)
 732                         is->is_login_pdu = response;
 733                 else
 734                         icl_pdu_free(response);
 735                 ISCSI_SESSION_UNLOCK(is);
 736                 cv_signal(&is->is_login_cv);
 737                 return;
 738         }
 739 #endif
 740
 741         /*
 742          * The handling routine is responsible for freeing the PDU
 743          * when it's no longer needed.
 744          */
 745         switch (response->ip_bhs->bhs_opcode) {
 746         case ISCSI_BHS_OPCODE_NOP_IN:
 747                 iscsi_pdu_handle_nop_in(response);
 748                 ISCSI_SESSION_UNLOCK(is);
 749                 break;
 750         case ISCSI_BHS_OPCODE_SCSI_RESPONSE:
 751                 iscsi_pdu_handle_scsi_response(response);
 752                 /* Session lock dropped inside. */
 753                 ISCSI_SESSION_LOCK_ASSERT_NOT(is);
 754                 break;
 755         case ISCSI_BHS_OPCODE_TASK_RESPONSE:
 756                 iscsi_pdu_handle_task_response(response);
 757                 ISCSI_SESSION_UNLOCK(is);
 758                 break;
 759         case ISCSI_BHS_OPCODE_SCSI_DATA_IN:
 760                 iscsi_pdu_handle_data_in(response);
 761                 /* Session lock dropped inside. */
 762                 ISCSI_SESSION_LOCK_ASSERT_NOT(is);
 763                 break;
 764         case ISCSI_BHS_OPCODE_LOGOUT_RESPONSE:
 765                 iscsi_pdu_handle_logout_response(response);
 766                 ISCSI_SESSION_UNLOCK(is);
 767                 break;
 768         case ISCSI_BHS_OPCODE_R2T:
 769                 iscsi_pdu_handle_r2t(response);
 770                 ISCSI_SESSION_UNLOCK(is);
 771                 break;
 772         case ISCSI_BHS_OPCODE_ASYNC_MESSAGE:
 773                 iscsi_pdu_handle_async_message(response);
 774                 ISCSI_SESSION_UNLOCK(is);
 775                 break;
 776         case ISCSI_BHS_OPCODE_REJECT:
 777                 iscsi_pdu_handle_reject(response);
 778                 ISCSI_SESSION_UNLOCK(is);
 779                 break;
 780         default:
 781                 ISCSI_SESSION_WARN(is, "received PDU with unsupported "
 782                     "opcode 0x%x; reconnecting",
 783                     response->ip_bhs->bhs_opcode);
 784                 iscsi_session_reconnect(is);
 785                 ISCSI_SESSION_UNLOCK(is);
 786                 icl_pdu_free(response);
 787         }
 788 }
 789
 790 static void
 791 iscsi_error_callback(struct icl_conn *ic)
 792 {
 793         struct iscsi_session *is;
 794
 795         is = CONN_SESSION(ic);
 796
 797         ISCSI_SESSION_WARN(is, "connection error; reconnecting");
 798         iscsi_session_reconnect(is);
 799 }
 800
 801 static void
 802 iscsi_pdu_handle_nop_in(struct icl_pdu *response)
 803 {
 804         struct iscsi_session *is;
 805         struct iscsi_bhs_nop_out *bhsno;
 806         struct iscsi_bhs_nop_in *bhsni;
 807         struct icl_pdu *request;
 808         void *data = NULL;
 809         size_t datasize;
 810         int error;
 811
 812         is = PDU_SESSION(response);
 813         bhsni = (struct iscsi_bhs_nop_in *)response->ip_bhs;
 814
 815         if (bhsni->bhsni_target_transfer_tag == 0xffffffff) {
 816                 /*
 817                  * Nothing to do; iscsi_pdu_update_statsn() already
 818                  * zeroed the timeout.
 819                  */
 820                 icl_pdu_free(response);
 821                 return;
 822         }
 823
 824         datasize = icl_pdu_data_segment_length(response);
 825         if (datasize > 0) {
 826                 data = malloc(datasize, M_ISCSI, M_NOWAIT | M_ZERO);
 827                 if (data == NULL) {
 828                         ISCSI_SESSION_WARN(is, "failed to allocate memory; "
 829                             "reconnecting");
 830                         icl_pdu_free(response);
 831                         iscsi_session_reconnect(is);
 832                         return;
 833                 }
 834                 icl_pdu_get_data(response, 0, data, datasize);
 835         }
 836
 837         request = icl_pdu_new(response->ip_conn, M_NOWAIT);
 838         if (request == NULL) {
 839                 ISCSI_SESSION_WARN(is, "failed to allocate memory; "
 840                     "reconnecting");
 841                 free(data, M_ISCSI);
 842                 icl_pdu_free(response);
 843                 iscsi_session_reconnect(is);
 844                 return;
 845         }
 846         bhsno = (struct iscsi_bhs_nop_out *)request->ip_bhs;
 847         bhsno->bhsno_opcode = ISCSI_BHS_OPCODE_NOP_OUT |
 848             ISCSI_BHS_OPCODE_IMMEDIATE;
 849         bhsno->bhsno_flags = 0x80;
 850         bhsno->bhsno_initiator_task_tag = 0xffffffff;
 851         bhsno->bhsno_target_transfer_tag = bhsni->bhsni_target_transfer_tag;
 852         if (datasize > 0) {
 853                 error = icl_pdu_append_data(request, data, datasize, M_NOWAIT);
 854                 if (error != 0) {
 855                         ISCSI_SESSION_WARN(is, "failed to allocate memory; "
 856                             "reconnecting");
 857                         free(data, M_ISCSI);
 858                         icl_pdu_free(request);
 859                         icl_pdu_free(response);
 860                         iscsi_session_reconnect(is);
 861                         return;
 862                 }
 863                 free(data, M_ISCSI);
 864         }
 865
 866         icl_pdu_free(response);
 867         iscsi_pdu_queue_locked(request);
 868 }
 869
 870 static void
 871 iscsi_pdu_handle_scsi_response(struct icl_pdu *response)
 872 {
 873         struct iscsi_bhs_scsi_response *bhssr;
 874         struct iscsi_outstanding *io;
 875         struct iscsi_session *is;
 876         union ccb *ccb;
 877         struct ccb_scsiio *csio;
 878         size_t data_segment_len, received;
 879         uint16_t sense_len;
 880         uint32_t resid;
 881
 882         is = PDU_SESSION(response);
 883
 884         bhssr = (struct iscsi_bhs_scsi_response *)response->ip_bhs;
 885         io = iscsi_outstanding_find(is, bhssr->bhssr_initiator_task_tag);
 886         if (io == NULL || io->io_ccb == NULL) {
 887                 ISCSI_SESSION_WARN(is, "bad itt 0x%x", bhssr->bhssr_initiator_task_tag);
 888                 icl_pdu_free(response);
 889                 iscsi_session_reconnect(is);
 890                 ISCSI_SESSION_UNLOCK(is);
 891                 return;
 892         }
 893
 894         ccb = io->io_ccb;
 895         if (bhssr->bhssr_response == BHSSR_RESPONSE_COMMAND_COMPLETED) {
 896                 if (ntohl(bhssr->bhssr_expdatasn) != io->io_datasn) {
 897                         ISCSI_SESSION_WARN(is,
 898                             "ExpDataSN mismatch in SCSI Response (%u vs %u)",
 899                             ntohl(bhssr->bhssr_expdatasn), io->io_datasn);
 900
 901                         /*
 902                          * XXX: Permit an ExpDataSN of zero for errors.
 903                          *
 904                          * This doesn't conform to RFC 7143, but some
 905                          * targets seem to do this.
 906                          */
 907                         if (bhssr->bhssr_status != 0 &&
 908                             bhssr->bhssr_expdatasn == htonl(0))
 909                                 goto skip_expdatasn;
 910
 911                         icl_pdu_free(response);
 912                         iscsi_session_reconnect(is);
 913                         ISCSI_SESSION_UNLOCK(is);
 914                         return;
 915                 }
 916         } else {
 917                 if (bhssr->bhssr_expdatasn != htonl(0)) {
 918                         ISCSI_SESSION_WARN(is,
 919                             "ExpDataSN mismatch in SCSI Response (%u vs 0)",
 920                             ntohl(bhssr->bhssr_expdatasn));
 921                         icl_pdu_free(response);
 922                         iscsi_session_reconnect(is);
 923                         ISCSI_SESSION_UNLOCK(is);
 924                         return;
 925                 }
 926         }
 927 skip_expdatasn:
 928
 929         /*
 930          * With iSER, after getting good response we can be sure
 931          * that all the data has been successfully transferred.
 932          */
 933         if (is->is_conn->ic_iser) {
 934                 resid = ntohl(bhssr->bhssr_residual_count);
 935                 if (bhssr->bhssr_flags & BHSSR_FLAGS_RESIDUAL_UNDERFLOW) {
 936                         io->io_received = ccb->csio.dxfer_len - resid;
 937                 } else if (bhssr->bhssr_flags & BHSSR_FLAGS_RESIDUAL_OVERFLOW) {
 938                         ISCSI_SESSION_WARN(is, "overflow: target indicates %d", resid);
 939                 } else {
 940                         io->io_received = ccb->csio.dxfer_len;
 941                 }
 942         }
 943
 944         received = io->io_received;
 945         iscsi_outstanding_remove(is, io);
 946         ISCSI_SESSION_UNLOCK(is);
 947
 948         if (bhssr->bhssr_response != BHSSR_RESPONSE_COMMAND_COMPLETED) {
 949                 ISCSI_SESSION_WARN(is, "service response 0x%x", bhssr->bhssr_response);
 950                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
 951                         xpt_freeze_devq(ccb->ccb_h.path, 1);
 952                         ISCSI_SESSION_DEBUG(is, "freezing devq");
 953                 }
 954                 ccb->ccb_h.status = CAM_REQ_CMP_ERR | CAM_DEV_QFRZN;
 955         } else if (bhssr->bhssr_status == 0) {
 956                 ccb->ccb_h.status = CAM_REQ_CMP;
 957         } else {
 958                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
 959                         xpt_freeze_devq(ccb->ccb_h.path, 1);
 960                         ISCSI_SESSION_DEBUG(is, "freezing devq");
 961                 }
 962                 ccb->ccb_h.status = CAM_SCSI_STATUS_ERROR | CAM_DEV_QFRZN;
 963                 ccb->csio.scsi_status = bhssr->bhssr_status;
 964         }
 965
 966         csio = &ccb->csio;
 967         data_segment_len = icl_pdu_data_segment_length(response);
 968         if (data_segment_len > 0) {
 969                 if (data_segment_len < sizeof(sense_len)) {
 970                         ISCSI_SESSION_WARN(is, "truncated data segment (%zd bytes)",
 971                             data_segment_len);
 972                         if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
 973                                 xpt_freeze_devq(ccb->ccb_h.path, 1);
 974                                 ISCSI_SESSION_DEBUG(is, "freezing devq");
 975                         }
 976                         ccb->ccb_h.status = CAM_REQ_CMP_ERR | CAM_DEV_QFRZN;
 977                         goto out;
 978                 }
 979                 icl_pdu_get_data(response, 0, &sense_len, sizeof(sense_len));
 980                 sense_len = ntohs(sense_len);
 981 #if 0
 982                 ISCSI_SESSION_DEBUG(is, "sense_len %d, data len %zd",
 983                     sense_len, data_segment_len);
 984 #endif
 985                 if (sizeof(sense_len) + sense_len > data_segment_len) {
 986                         ISCSI_SESSION_WARN(is, "truncated data segment "
 987                             "(%zd bytes, should be %zd)",
 988                             data_segment_len, sizeof(sense_len) + sense_len);
 989                         if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
 990                                 xpt_freeze_devq(ccb->ccb_h.path, 1);
 991                                 ISCSI_SESSION_DEBUG(is, "freezing devq");
 992                         }
 993                         ccb->ccb_h.status = CAM_REQ_CMP_ERR | CAM_DEV_QFRZN;
 994                         goto out;
 995                 } else if (sizeof(sense_len) + sense_len < data_segment_len)
 996                         ISCSI_SESSION_WARN(is, "oversize data segment "
 997                             "(%zd bytes, should be %zd)",
 998                             data_segment_len, sizeof(sense_len) + sense_len);
 999                 if (sense_len > csio->sense_len) {
1000                         ISCSI_SESSION_DEBUG(is, "truncating sense from %d to %d",
1001                             sense_len, csio->sense_len);
1002                         sense_len = csio->sense_len;
1003                 }
1004                 icl_pdu_get_data(response, sizeof(sense_len), &csio->sense_data, sense_len);
1005                 csio->sense_resid = csio->sense_len - sense_len;
1006                 ccb->ccb_h.status |= CAM_AUTOSNS_VALID;
1007         }
1008
1009 out:
1010         if (bhssr->bhssr_flags & BHSSR_FLAGS_RESIDUAL_UNDERFLOW)
1011                 csio->resid = ntohl(bhssr->bhssr_residual_count);
1012
1013         if ((csio->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_IN) {
1014                 KASSERT(received <= csio->dxfer_len,
1015                     ("received > csio->dxfer_len"));
1016                 if (received < csio->dxfer_len) {
1017                         if (csio->resid != csio->dxfer_len - received) {
1018                                 ISCSI_SESSION_WARN(is, "underflow mismatch: "
1019                                     "target indicates %d, we calculated %zd",
1020                                     csio->resid, csio->dxfer_len - received);
1021                         }
1022                         csio->resid = csio->dxfer_len - received;
1023                 }
1024         }
1025
1026         xpt_done(ccb);
1027         icl_pdu_free(response);
1028 }
1029
1030 static void
1031 iscsi_pdu_handle_task_response(struct icl_pdu *response)
1032 {
1033         struct iscsi_bhs_task_management_response *bhstmr;
1034         struct iscsi_outstanding *io, *aio;
1035         struct iscsi_session *is;
1036
1037         is = PDU_SESSION(response);
1038
1039         bhstmr = (struct iscsi_bhs_task_management_response *)response->ip_bhs;
1040         io = iscsi_outstanding_find(is, bhstmr->bhstmr_initiator_task_tag);
1041         if (io == NULL || io->io_ccb != NULL) {
1042                 ISCSI_SESSION_WARN(is, "bad itt 0x%x",
1043                     bhstmr->bhstmr_initiator_task_tag);
1044                 icl_pdu_free(response);
1045                 iscsi_session_reconnect(is);
1046                 return;
1047         }
1048
1049         if (bhstmr->bhstmr_response != BHSTMR_RESPONSE_FUNCTION_COMPLETE) {
1050                 ISCSI_SESSION_WARN(is, "task response 0x%x",
1051                     bhstmr->bhstmr_response);
1052         } else {
1053                 aio = iscsi_outstanding_find(is, io->io_referenced_task_tag);
1054                 if (aio != NULL && aio->io_ccb != NULL)
1055                         iscsi_session_terminate_task(is, aio, CAM_REQ_ABORTED);
1056         }
1057
1058         iscsi_outstanding_remove(is, io);
1059         icl_pdu_free(response);
1060 }
1061
1062 static void
1063 iscsi_pdu_handle_data_in(struct icl_pdu *response)
1064 {
1065         struct iscsi_bhs_data_in *bhsdi;
1066         struct iscsi_outstanding *io;
1067         struct iscsi_session *is;
1068         union ccb *ccb;
1069         struct ccb_scsiio *csio;
1070         size_t data_segment_len, received, oreceived;
1071
1072         is = PDU_SESSION(response);
1073         bhsdi = (struct iscsi_bhs_data_in *)response->ip_bhs;
1074         io = iscsi_outstanding_find(is, bhsdi->bhsdi_initiator_task_tag);
1075         if (io == NULL || io->io_ccb == NULL) {
1076                 ISCSI_SESSION_WARN(is, "bad itt 0x%x", bhsdi->bhsdi_initiator_task_tag);
1077                 icl_pdu_free(response);
1078                 iscsi_session_reconnect(is);
1079                 ISCSI_SESSION_UNLOCK(is);
1080                 return;
1081         }
1082
1083         if (io->io_datasn != ntohl(bhsdi->bhsdi_datasn)) {
1084                 ISCSI_SESSION_WARN(is, "received Data-In PDU with "
1085                     "DataSN %u, while expected %u; dropping connection",
1086                     ntohl(bhsdi->bhsdi_datasn), io->io_datasn);
1087                 icl_pdu_free(response);
1088                 iscsi_session_reconnect(is);
1089                 ISCSI_SESSION_UNLOCK(is);
1090                 return;
1091         }
1092         io->io_datasn += response->ip_additional_pdus + 1;
1093
1094         data_segment_len = icl_pdu_data_segment_length(response);
1095         if (data_segment_len == 0) {
1096                 /*
1097                  * "The sending of 0 length data segments should be avoided,
1098                  * but initiators and targets MUST be able to properly receive
1099                  * 0 length data segments."
1100                  */
1101                 ISCSI_SESSION_UNLOCK(is);
1102                 icl_pdu_free(response);
1103                 return;
1104         }
1105
1106         /*
1107          * We need to track this for security reasons - without it, malicious target
1108          * could respond to SCSI READ without sending Data-In PDUs, which would result
1109          * in read operation on the initiator side returning random kernel data.
1110          */
1111         if (ntohl(bhsdi->bhsdi_buffer_offset) != io->io_received) {
1112                 ISCSI_SESSION_WARN(is, "data out of order; expected offset %zd, got %zd",
1113                     io->io_received, (size_t)ntohl(bhsdi->bhsdi_buffer_offset));
1114                 icl_pdu_free(response);
1115                 iscsi_session_reconnect(is);
1116                 ISCSI_SESSION_UNLOCK(is);
1117                 return;
1118         }
1119
1120         ccb = io->io_ccb;
1121         csio = &ccb->csio;
1122
1123         if (io->io_received + data_segment_len > csio->dxfer_len) {
1124                 ISCSI_SESSION_WARN(is, "oversize data segment (%zd bytes "
1125                     "at offset %zd, buffer is %d)",
1126                     data_segment_len, io->io_received, csio->dxfer_len);
1127                 icl_pdu_free(response);
1128                 iscsi_session_reconnect(is);
1129                 ISCSI_SESSION_UNLOCK(is);
1130                 return;
1131         }
1132
1133         oreceived = io->io_received;
1134         io->io_received += data_segment_len;
1135         received = io->io_received;
1136         if ((bhsdi->bhsdi_flags & BHSDI_FLAGS_S) != 0)
1137                 iscsi_outstanding_remove(is, io);
1138         ISCSI_SESSION_UNLOCK(is);
1139
1140         icl_pdu_get_data(response, 0, csio->data_ptr + oreceived, data_segment_len);
1141
1142         /*
1143          * XXX: Check F.
1144          */
1145         if ((bhsdi->bhsdi_flags & BHSDI_FLAGS_S) == 0) {
1146                 /*
1147                  * Nothing more to do.
1148                  */
1149                 icl_pdu_free(response);
1150                 return;
1151         }
1152
1153         //ISCSI_SESSION_DEBUG(is, "got S flag; status 0x%x", bhsdi->bhsdi_status);
1154         if (bhsdi->bhsdi_status == 0) {
1155                 ccb->ccb_h.status = CAM_REQ_CMP;
1156         } else {
1157                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
1158                         xpt_freeze_devq(ccb->ccb_h.path, 1);
1159                         ISCSI_SESSION_DEBUG(is, "freezing devq");
1160                 }
1161                 ccb->ccb_h.status = CAM_SCSI_STATUS_ERROR | CAM_DEV_QFRZN;
1162                 csio->scsi_status = bhsdi->bhsdi_status;
1163         }
1164
1165         if ((csio->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_IN) {
1166                 KASSERT(received <= csio->dxfer_len,
1167                     ("received > csio->dxfer_len"));
1168                 if (received < csio->dxfer_len) {
1169                         csio->resid = ntohl(bhsdi->bhsdi_residual_count);
1170                         if (csio->resid != csio->dxfer_len - received) {
1171                                 ISCSI_SESSION_WARN(is, "underflow mismatch: "
1172                                     "target indicates %d, we calculated %zd",
1173                                     csio->resid, csio->dxfer_len - received);
1174                         }
1175                         csio->resid = csio->dxfer_len - received;
1176                 }
1177         }
1178
1179         xpt_done(ccb);
1180         icl_pdu_free(response);
1181 }
1182
1183 static void
1184 iscsi_pdu_handle_logout_response(struct icl_pdu *response)
1185 {
1186
1187         ISCSI_SESSION_DEBUG(PDU_SESSION(response), "logout response");
1188         icl_pdu_free(response);
1189 }
1190
1191 static void
1192 iscsi_pdu_handle_r2t(struct icl_pdu *response)
1193 {
1194         struct icl_pdu *request;
1195         struct iscsi_session *is;
1196         struct iscsi_bhs_r2t *bhsr2t;
1197         struct iscsi_bhs_data_out *bhsdo;
1198         struct iscsi_outstanding *io;
1199         struct ccb_scsiio *csio;
1200         size_t off, len, max_send_data_segment_length, total_len;
1201         int error;
1202         uint32_t datasn = 0;
1203
1204         is = PDU_SESSION(response);
1205
1206         bhsr2t = (struct iscsi_bhs_r2t *)response->ip_bhs;
1207         io = iscsi_outstanding_find(is, bhsr2t->bhsr2t_initiator_task_tag);
1208         if (io == NULL || io->io_ccb == NULL) {
1209                 ISCSI_SESSION_WARN(is, "bad itt 0x%x; reconnecting",
1210                     bhsr2t->bhsr2t_initiator_task_tag);
1211                 icl_pdu_free(response);
1212                 iscsi_session_reconnect(is);
1213                 return;
1214         }
1215
1216         csio = &io->io_ccb->csio;
1217
1218         if ((csio->ccb_h.flags & CAM_DIR_MASK) != CAM_DIR_OUT) {
1219                 ISCSI_SESSION_WARN(is, "received R2T for read command; reconnecting");
1220                 icl_pdu_free(response);
1221                 iscsi_session_reconnect(is);
1222                 return;
1223         }
1224
1225         /*
1226          * XXX: Verify R2TSN.
1227          */
1228
1229         off = ntohl(bhsr2t->bhsr2t_buffer_offset);
1230         if (off > csio->dxfer_len) {
1231                 ISCSI_SESSION_WARN(is, "target requested invalid offset "
1232                     "%zd, buffer is is %d; reconnecting", off, csio->dxfer_len);
1233                 icl_pdu_free(response);
1234                 iscsi_session_reconnect(is);
1235                 return;
1236         }
1237
1238         total_len = ntohl(bhsr2t->bhsr2t_desired_data_transfer_length);
1239         if (total_len == 0 || total_len > csio->dxfer_len) {
1240                 ISCSI_SESSION_WARN(is, "target requested invalid length "
1241                     "%zd, buffer is %d; reconnecting", total_len, csio->dxfer_len);
1242                 icl_pdu_free(response);
1243                 iscsi_session_reconnect(is);
1244                 return;
1245         }
1246
1247         //ISCSI_SESSION_DEBUG(is, "r2t; off %zd, len %zd", off, total_len);
1248
1249         if (is->is_conn->ic_hw_isomax != 0)
1250                 max_send_data_segment_length = is->is_conn->ic_hw_isomax;
1251         else
1252                 max_send_data_segment_length =
1253                     is->is_conn->ic_max_send_data_segment_length;
1254         for (;;) {
1255                 len = total_len;
1256
1257                 if (len > max_send_data_segment_length)
1258                         len = max_send_data_segment_length;
1259
1260                 if (off + len > csio->dxfer_len) {
1261                         ISCSI_SESSION_WARN(is, "target requested invalid "
1262                             "length/offset %zd, buffer is %d; reconnecting",
1263                             off + len, csio->dxfer_len);
1264                         icl_pdu_free(response);
1265                         iscsi_session_reconnect(is);
1266                         return;
1267                 }
1268
1269                 request = icl_pdu_new(response->ip_conn, M_NOWAIT);
1270                 if (request == NULL) {
1271                         icl_pdu_free(response);
1272                         iscsi_session_reconnect(is);
1273                         return;
1274                 }
1275
1276                 bhsdo = (struct iscsi_bhs_data_out *)request->ip_bhs;
1277                 bhsdo->bhsdo_opcode = ISCSI_BHS_OPCODE_SCSI_DATA_OUT;
1278                 bhsdo->bhsdo_lun = bhsr2t->bhsr2t_lun;
1279                 bhsdo->bhsdo_initiator_task_tag =
1280                     bhsr2t->bhsr2t_initiator_task_tag;
1281                 bhsdo->bhsdo_target_transfer_tag =
1282                     bhsr2t->bhsr2t_target_transfer_tag;
1283                 bhsdo->bhsdo_datasn = htonl(datasn);
1284                 bhsdo->bhsdo_buffer_offset = htonl(off);
1285                 error = icl_pdu_append_data(request, csio->data_ptr + off, len,
1286                     M_NOWAIT);
1287                 if (error != 0) {
1288                         ISCSI_SESSION_WARN(is, "failed to allocate memory; "
1289                             "reconnecting");
1290                         icl_pdu_free(request);
1291                         icl_pdu_free(response);
1292                         iscsi_session_reconnect(is);
1293                         return;
1294                 }
1295
1296                 datasn += howmany(len,
1297                     is->is_conn->ic_max_send_data_segment_length);
1298                 off += len;
1299                 total_len -= len;
1300
1301                 if (total_len == 0) {
1302                         bhsdo->bhsdo_flags |= BHSDO_FLAGS_F;
1303                         //ISCSI_SESSION_DEBUG(is, "setting F, off %zd", off);
1304                 } else {
1305                         //ISCSI_SESSION_DEBUG(is, "not finished, off %zd", off);
1306                 }
1307
1308                 iscsi_pdu_queue_locked(request);
1309
1310                 if (total_len == 0)
1311                         break;
1312         }
1313
1314         icl_pdu_free(response);
1315 }
1316
1317 static void
1318 iscsi_pdu_handle_async_message(struct icl_pdu *response)
1319 {
1320         struct iscsi_bhs_asynchronous_message *bhsam;
1321         struct iscsi_session *is;
1322
1323         is = PDU_SESSION(response);
1324         bhsam = (struct iscsi_bhs_asynchronous_message *)response->ip_bhs;
1325         switch (bhsam->bhsam_async_event) {
1326         case BHSAM_EVENT_TARGET_REQUESTS_LOGOUT:
1327                 ISCSI_SESSION_WARN(is, "target requests logout; removing session");
1328                 iscsi_session_logout(is);
1329                 iscsi_session_terminate(is);
1330                 break;
1331         case BHSAM_EVENT_TARGET_TERMINATES_CONNECTION:
1332                 ISCSI_SESSION_WARN(is, "target indicates it will drop the connection");
1333                 break;
1334         case BHSAM_EVENT_TARGET_TERMINATES_SESSION:
1335                 ISCSI_SESSION_WARN(is, "target indicates it will drop the session");
1336                 break;
1337         default:
1338                 /*
1339                  * XXX: Technically, we're obligated to also handle
1340                  *      parameter renegotiation.
1341                  */
1342                 ISCSI_SESSION_WARN(is, "ignoring AsyncEvent %d", bhsam->bhsam_async_event);
1343                 break;
1344         }
1345
1346         icl_pdu_free(response);
1347 }
1348
1349 static void
1350 iscsi_pdu_handle_reject(struct icl_pdu *response)
1351 {
1352         struct iscsi_bhs_reject *bhsr;
1353         struct iscsi_session *is;
1354
1355         is = PDU_SESSION(response);
1356         bhsr = (struct iscsi_bhs_reject *)response->ip_bhs;
1357         ISCSI_SESSION_WARN(is, "received Reject PDU, reason 0x%x; protocol error?",
1358             bhsr->bhsr_reason);
1359
1360         icl_pdu_free(response);
1361 }
1362
1363 static int
1364 iscsi_ioctl_daemon_wait(struct iscsi_softc *sc,
1365     struct iscsi_daemon_request *request)
1366 {
1367         struct iscsi_session *is;
1368         struct icl_drv_limits idl;
1369         int error;
1370
1371         sx_slock(&sc->sc_lock);
1372         for (;;) {
1373                 TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1374                         ISCSI_SESSION_LOCK(is);
1375                         if (is->is_conf.isc_enable == 0 &&
1376                             is->is_conf.isc_discovery == 0) {
1377                                 ISCSI_SESSION_UNLOCK(is);
1378                                 continue;
1379                         }
1380                         if (is->is_waiting_for_iscsid)
1381                                 break;
1382                         ISCSI_SESSION_UNLOCK(is);
1383                 }
1384
1385                 if (is == NULL) {
1386                         if (sc->sc_unloading) {
1387                                 sx_sunlock(&sc->sc_lock);
1388                                 return (ENXIO);
1389                         }
1390
1391                         /*
1392                          * No session requires attention from iscsid(8); wait.
1393                          */
1394                         error = cv_wait_sig(&sc->sc_cv, &sc->sc_lock);
1395                         if (error != 0) {
1396                                 sx_sunlock(&sc->sc_lock);
1397                                 return (error);
1398                         }
1399                         continue;
1400                 }
1401
1402                 is->is_waiting_for_iscsid = false;
1403                 is->is_login_phase = true;
1404                 is->is_reason[0] = '\0';
1405                 ISCSI_SESSION_UNLOCK(is);
1406
1407                 request->idr_session_id = is->is_id;
1408                 memcpy(&request->idr_isid, &is->is_isid,
1409                     sizeof(request->idr_isid));
1410                 request->idr_tsih = 0;  /* New or reinstated session. */
1411                 memcpy(&request->idr_conf, &is->is_conf,
1412                     sizeof(request->idr_conf));
1413
1414                 error = icl_limits(is->is_conf.isc_offload,
1415                     is->is_conf.isc_iser, &idl);
1416                 if (error != 0) {
1417                         ISCSI_SESSION_WARN(is, "icl_limits for offload \"%s\" "
1418                             "failed with error %d", is->is_conf.isc_offload,
1419                             error);
1420                         sx_sunlock(&sc->sc_lock);
1421                         return (error);
1422                 }
1423                 request->idr_limits.isl_max_recv_data_segment_length =
1424                     idl.idl_max_recv_data_segment_length;
1425                 request->idr_limits.isl_max_send_data_segment_length =
1426                     idl.idl_max_send_data_segment_length;
1427                 request->idr_limits.isl_max_burst_length =
1428                     idl.idl_max_burst_length;
1429                 request->idr_limits.isl_first_burst_length =
1430                     idl.idl_first_burst_length;
1431
1432                 sx_sunlock(&sc->sc_lock);
1433                 return (0);
1434         }
1435 }
1436
1437 static int
1438 iscsi_ioctl_daemon_handoff(struct iscsi_softc *sc,
1439     struct iscsi_daemon_handoff *handoff)
1440 {
1441         struct iscsi_session *is;
1442         struct icl_conn *ic;
1443         int error;
1444
1445         sx_slock(&sc->sc_lock);
1446
1447         /*
1448          * Find the session to hand off socket to.
1449          */
1450         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1451                 if (is->is_id == handoff->idh_session_id)
1452                         break;
1453         }
1454         if (is == NULL) {
1455                 sx_sunlock(&sc->sc_lock);
1456                 return (ESRCH);
1457         }
1458         ISCSI_SESSION_LOCK(is);
1459         ic = is->is_conn;
1460         if (is->is_conf.isc_discovery || is->is_terminating) {
1461                 ISCSI_SESSION_UNLOCK(is);
1462                 sx_sunlock(&sc->sc_lock);
1463                 return (EINVAL);
1464         }
1465         if (is->is_connected) {
1466                 /*
1467                  * This might have happened because another iscsid(8)
1468                  * instance handed off the connection in the meantime.
1469                  * Just return.
1470                  */
1471                 ISCSI_SESSION_WARN(is, "handoff on already connected "
1472                     "session");
1473                 ISCSI_SESSION_UNLOCK(is);
1474                 sx_sunlock(&sc->sc_lock);
1475                 return (EBUSY);
1476         }
1477
1478         strlcpy(is->is_target_alias, handoff->idh_target_alias,
1479             sizeof(is->is_target_alias));
1480         is->is_tsih = handoff->idh_tsih;
1481         is->is_statsn = handoff->idh_statsn;
1482         is->is_protocol_level = handoff->idh_protocol_level;
1483         is->is_initial_r2t = handoff->idh_initial_r2t;
1484         is->is_immediate_data = handoff->idh_immediate_data;
1485
1486         ic->ic_max_recv_data_segment_length =
1487             handoff->idh_max_recv_data_segment_length;
1488         ic->ic_max_send_data_segment_length =
1489             handoff->idh_max_send_data_segment_length;
1490         is->is_max_burst_length = handoff->idh_max_burst_length;
1491         is->is_first_burst_length = handoff->idh_first_burst_length;
1492
1493         if (handoff->idh_header_digest == ISCSI_DIGEST_CRC32C)
1494                 ic->ic_header_crc32c = true;
1495         else
1496                 ic->ic_header_crc32c = false;
1497         if (handoff->idh_data_digest == ISCSI_DIGEST_CRC32C)
1498                 ic->ic_data_crc32c = true;
1499         else
1500                 ic->ic_data_crc32c = false;
1501         ic->ic_maxtags = maxtags;
1502
1503         is->is_cmdsn = 0;
1504         is->is_expcmdsn = 0;
1505         is->is_maxcmdsn = 0;
1506         is->is_waiting_for_iscsid = false;
1507         is->is_login_phase = false;
1508         is->is_timeout = 0;
1509         is->is_connected = true;
1510         is->is_reason[0] = '\0';
1511
1512         ISCSI_SESSION_UNLOCK(is);
1513
1514         /*
1515          * If we're going through the proxy, the idh_socket will be 0,
1516          * and the ICL module can simply ignore this call.  It can also
1517          * use it to determine it's no longer in the Login phase.
1518          */
1519         error = icl_conn_handoff(ic, handoff->idh_socket);
1520         if (error != 0) {
1521                 sx_sunlock(&sc->sc_lock);
1522                 iscsi_session_terminate(is);
1523                 return (error);
1524         }
1525
1526         sx_sunlock(&sc->sc_lock);
1527
1528         if (is->is_sim != NULL) {
1529                 /*
1530                  * When reconnecting, there already is SIM allocated for the session.
1531                  */
1532                 KASSERT(is->is_simq_frozen, ("reconnect without frozen simq"));
1533                 ISCSI_SESSION_LOCK(is);
1534                 ISCSI_SESSION_DEBUG(is, "releasing");
1535                 is->is_simq_frozen = false;
1536                 xpt_release_simq(is->is_sim, 1);
1537                 ISCSI_SESSION_UNLOCK(is);
1538
1539         } else {
1540                 ISCSI_SESSION_LOCK(is);
1541                 is->is_devq = cam_simq_alloc(ic->ic_maxtags);
1542                 if (is->is_devq == NULL) {
1543                         ISCSI_SESSION_WARN(is, "failed to allocate simq");
1544                         iscsi_session_terminate(is);
1545                         return (ENOMEM);
1546                 }
1547
1548                 is->is_sim = cam_sim_alloc(iscsi_action, NULL, "iscsi",
1549                     is, is->is_id /* unit */, &is->is_lock,
1550                     1, ic->ic_maxtags, is->is_devq);
1551                 if (is->is_sim == NULL) {
1552                         ISCSI_SESSION_UNLOCK(is);
1553                         ISCSI_SESSION_WARN(is, "failed to allocate SIM");
1554                         cam_simq_free(is->is_devq);
1555                         iscsi_session_terminate(is);
1556                         return (ENOMEM);
1557                 }
1558
1559                 if (xpt_bus_register(is->is_sim, NULL, 0) != 0) {
1560                         ISCSI_SESSION_UNLOCK(is);
1561                         ISCSI_SESSION_WARN(is, "failed to register bus");
1562                         iscsi_session_terminate(is);
1563                         return (ENOMEM);
1564                 }
1565
1566                 error = xpt_create_path(&is->is_path, /*periph*/NULL,
1567                     cam_sim_path(is->is_sim), CAM_TARGET_WILDCARD,
1568                     CAM_LUN_WILDCARD);
1569                 if (error != CAM_REQ_CMP) {
1570                         ISCSI_SESSION_UNLOCK(is);
1571                         ISCSI_SESSION_WARN(is, "failed to create path");
1572                         iscsi_session_terminate(is);
1573                         return (ENOMEM);
1574                 }
1575                 ISCSI_SESSION_UNLOCK(is);
1576         }
1577
1578         return (0);
1579 }
1580
1581 static int
1582 iscsi_ioctl_daemon_fail(struct iscsi_softc *sc,
1583     struct iscsi_daemon_fail *fail)
1584 {
1585         struct iscsi_session *is;
1586
1587         sx_slock(&sc->sc_lock);
1588
1589         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1590                 if (is->is_id == fail->idf_session_id)
1591                         break;
1592         }
1593         if (is == NULL) {
1594                 sx_sunlock(&sc->sc_lock);
1595                 return (ESRCH);
1596         }
1597         ISCSI_SESSION_LOCK(is);
1598         ISCSI_SESSION_DEBUG(is, "iscsid(8) failed: %s",
1599             fail->idf_reason);
1600         strlcpy(is->is_reason, fail->idf_reason, sizeof(is->is_reason));
1601         //is->is_waiting_for_iscsid = false;
1602         //is->is_login_phase = true;
1603         //iscsi_session_reconnect(is);
1604         ISCSI_SESSION_UNLOCK(is);
1605         sx_sunlock(&sc->sc_lock);
1606
1607         return (0);
1608 }
1609
1610 #ifdef ICL_KERNEL_PROXY
1611 static int
1612 iscsi_ioctl_daemon_connect(struct iscsi_softc *sc,
1613     struct iscsi_daemon_connect *idc)
1614 {
1615         struct iscsi_session *is;
1616         struct sockaddr *from_sa, *to_sa;
1617         int error;
1618
1619         sx_slock(&sc->sc_lock);
1620         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1621                 if (is->is_id == idc->idc_session_id)
1622                         break;
1623         }
1624         if (is == NULL) {
1625                 sx_sunlock(&sc->sc_lock);
1626                 return (ESRCH);
1627         }
1628         sx_sunlock(&sc->sc_lock);
1629
1630         if (idc->idc_from_addrlen > 0) {
1631                 error = getsockaddr(&from_sa, (void *)idc->idc_from_addr, idc->idc_from_addrlen);
1632                 if (error != 0) {
1633                         ISCSI_SESSION_WARN(is,
1634                             "getsockaddr failed with error %d", error);
1635                         return (error);
1636                 }
1637         } else {
1638                 from_sa = NULL;
1639         }
1640         error = getsockaddr(&to_sa, (void *)idc->idc_to_addr, idc->idc_to_addrlen);
1641         if (error != 0) {
1642                 ISCSI_SESSION_WARN(is, "getsockaddr failed with error %d",
1643                     error);
1644                 free(from_sa, M_SONAME);
1645                 return (error);
1646         }
1647
1648         ISCSI_SESSION_LOCK(is);
1649         is->is_statsn = 0;
1650         is->is_cmdsn = 0;
1651         is->is_expcmdsn = 0;
1652         is->is_maxcmdsn = 0;
1653         is->is_waiting_for_iscsid = false;
1654         is->is_login_phase = true;
1655         is->is_timeout = 0;
1656         ISCSI_SESSION_UNLOCK(is);
1657
1658         error = icl_conn_connect(is->is_conn, idc->idc_domain,
1659             idc->idc_socktype, idc->idc_protocol, from_sa, to_sa);
1660         free(from_sa, M_SONAME);
1661         free(to_sa, M_SONAME);
1662
1663         /*
1664          * Digests are always disabled during login phase.
1665          */
1666         is->is_conn->ic_header_crc32c = false;
1667         is->is_conn->ic_data_crc32c = false;
1668
1669         return (error);
1670 }
1671
1672 static int
1673 iscsi_ioctl_daemon_send(struct iscsi_softc *sc,
1674     struct iscsi_daemon_send *ids)
1675 {
1676         struct iscsi_session *is;
1677         struct icl_pdu *ip;
1678         size_t datalen;
1679         void *data;
1680         int error;
1681
1682         sx_slock(&sc->sc_lock);
1683         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1684                 if (is->is_id == ids->ids_session_id)
1685                         break;
1686         }
1687         if (is == NULL) {
1688                 sx_sunlock(&sc->sc_lock);
1689                 return (ESRCH);
1690         }
1691         sx_sunlock(&sc->sc_lock);
1692
1693         if (is->is_login_phase == false)
1694                 return (EBUSY);
1695
1696         if (is->is_terminating || is->is_reconnecting)
1697                 return (EIO);
1698
1699         datalen = ids->ids_data_segment_len;
1700         if (datalen > is->is_conn->ic_max_send_data_segment_length)
1701                 return (EINVAL);
1702         if (datalen > 0) {
1703                 data = malloc(datalen, M_ISCSI, M_WAITOK);
1704                 error = copyin(ids->ids_data_segment, data, datalen);
1705                 if (error != 0) {
1706                         free(data, M_ISCSI);
1707                         return (error);
1708                 }
1709         }
1710
1711         ip = icl_pdu_new(is->is_conn, M_WAITOK);
1712         memcpy(ip->ip_bhs, ids->ids_bhs, sizeof(*ip->ip_bhs));
1713         if (datalen > 0) {
1714                 error = icl_pdu_append_data(ip, data, datalen, M_WAITOK);
1715                 KASSERT(error == 0, ("icl_pdu_append_data(..., M_WAITOK) failed"));
1716                 free(data, M_ISCSI);
1717         }
1718         iscsi_pdu_queue(ip);
1719
1720         return (0);
1721 }
1722
1723 static int
1724 iscsi_ioctl_daemon_receive(struct iscsi_softc *sc,
1725     struct iscsi_daemon_receive *idr)
1726 {
1727         struct iscsi_session *is;
1728         struct icl_pdu *ip;
1729         void *data;
1730         int error;
1731
1732         sx_slock(&sc->sc_lock);
1733         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1734                 if (is->is_id == idr->idr_session_id)
1735                         break;
1736         }
1737         if (is == NULL) {
1738                 sx_sunlock(&sc->sc_lock);
1739                 return (ESRCH);
1740         }
1741         sx_sunlock(&sc->sc_lock);
1742
1743         if (is->is_login_phase == false)
1744                 return (EBUSY);
1745
1746         ISCSI_SESSION_LOCK(is);
1747         while (is->is_login_pdu == NULL &&
1748             is->is_terminating == false &&
1749             is->is_reconnecting == false) {
1750                 error = cv_wait_sig(&is->is_login_cv, &is->is_lock);
1751                 if (error != 0) {
1752                         ISCSI_SESSION_UNLOCK(is);
1753                         return (error);
1754                 }
1755         }
1756         if (is->is_terminating || is->is_reconnecting) {
1757                 ISCSI_SESSION_UNLOCK(is);
1758                 return (EIO);
1759         }
1760         ip = is->is_login_pdu;
1761         is->is_login_pdu = NULL;
1762         ISCSI_SESSION_UNLOCK(is);
1763
1764         if (ip->ip_data_len > idr->idr_data_segment_len) {
1765                 icl_pdu_free(ip);
1766                 return (EMSGSIZE);
1767         }
1768
1769         copyout(ip->ip_bhs, idr->idr_bhs, sizeof(*ip->ip_bhs));
1770         if (ip->ip_data_len > 0) {
1771                 data = malloc(ip->ip_data_len, M_ISCSI, M_WAITOK);
1772                 icl_pdu_get_data(ip, 0, data, ip->ip_data_len);
1773                 copyout(data, idr->idr_data_segment, ip->ip_data_len);
1774                 free(data, M_ISCSI);
1775         }
1776
1777         icl_pdu_free(ip);
1778
1779         return (0);
1780 }
1781 #endif /* ICL_KERNEL_PROXY */
1782
1783 static void
1784 iscsi_sanitize_session_conf(struct iscsi_session_conf *isc)
1785 {
1786         /*
1787          * Just make sure all the fields are null-terminated.
1788          *
1789          * XXX: This is not particularly secure.  We should
1790          *      create our own conf and then copy in relevant
1791          *      fields.
1792          */
1793         isc->isc_initiator[ISCSI_NAME_LEN - 1] = '\0';
1794         isc->isc_initiator_addr[ISCSI_ADDR_LEN - 1] = '\0';
1795         isc->isc_initiator_alias[ISCSI_ALIAS_LEN - 1] = '\0';
1796         isc->isc_target[ISCSI_NAME_LEN - 1] = '\0';
1797         isc->isc_target_addr[ISCSI_ADDR_LEN - 1] = '\0';
1798         isc->isc_user[ISCSI_NAME_LEN - 1] = '\0';
1799         isc->isc_secret[ISCSI_SECRET_LEN - 1] = '\0';
1800         isc->isc_mutual_user[ISCSI_NAME_LEN - 1] = '\0';
1801         isc->isc_mutual_secret[ISCSI_SECRET_LEN - 1] = '\0';
1802 }
1803
1804 static bool
1805 iscsi_valid_session_conf(const struct iscsi_session_conf *isc)
1806 {
1807
1808         if (isc->isc_initiator[0] == '\0') {
1809                 ISCSI_DEBUG("empty isc_initiator");
1810                 return (false);
1811         }
1812
1813         if (isc->isc_target_addr[0] == '\0') {
1814                 ISCSI_DEBUG("empty isc_target_addr");
1815                 return (false);
1816         }
1817
1818         if (isc->isc_discovery != 0 && isc->isc_target[0] != 0) {
1819                 ISCSI_DEBUG("non-empty isc_target for discovery session");
1820                 return (false);
1821         }
1822
1823         if (isc->isc_discovery == 0 && isc->isc_target[0] == 0) {
1824                 ISCSI_DEBUG("empty isc_target for non-discovery session");
1825                 return (false);
1826         }
1827
1828         return (true);
1829 }
1830
1831 static int
1832 iscsi_ioctl_session_add(struct iscsi_softc *sc, struct iscsi_session_add *isa)
1833 {
1834         struct iscsi_session *is;
1835         const struct iscsi_session *is2;
1836         int error;
1837
1838         iscsi_sanitize_session_conf(&isa->isa_conf);
1839         if (iscsi_valid_session_conf(&isa->isa_conf) == false)
1840                 return (EINVAL);
1841
1842         is = malloc(sizeof(*is), M_ISCSI, M_ZERO | M_WAITOK);
1843         memcpy(&is->is_conf, &isa->isa_conf, sizeof(is->is_conf));
1844
1845         sx_xlock(&sc->sc_lock);
1846
1847         /*
1848          * Prevent duplicates.
1849          */
1850         TAILQ_FOREACH(is2, &sc->sc_sessions, is_next) {
1851                 if (!!is->is_conf.isc_discovery !=
1852                     !!is2->is_conf.isc_discovery)
1853                         continue;
1854
1855                 if (strcmp(is->is_conf.isc_target_addr,
1856                     is2->is_conf.isc_target_addr) != 0)
1857                         continue;
1858
1859                 if (is->is_conf.isc_discovery == 0 &&
1860                     strcmp(is->is_conf.isc_target,
1861                     is2->is_conf.isc_target) != 0)
1862                         continue;
1863
1864                 sx_xunlock(&sc->sc_lock);
1865                 free(is, M_ISCSI);
1866                 return (EBUSY);
1867         }
1868
1869         is->is_conn = icl_new_conn(is->is_conf.isc_offload,
1870             is->is_conf.isc_iser, "iscsi", &is->is_lock);
1871         if (is->is_conn == NULL) {
1872                 sx_xunlock(&sc->sc_lock);
1873                 free(is, M_ISCSI);
1874                 return (EINVAL);
1875         }
1876         is->is_conn->ic_receive = iscsi_receive_callback;
1877         is->is_conn->ic_error = iscsi_error_callback;
1878         is->is_conn->ic_prv0 = is;
1879         TAILQ_INIT(&is->is_outstanding);
1880         STAILQ_INIT(&is->is_postponed);
1881         mtx_init(&is->is_lock, "iscsi_lock", NULL, MTX_DEF);
1882         cv_init(&is->is_maintenance_cv, "iscsi_mt");
1883 #ifdef ICL_KERNEL_PROXY
1884         cv_init(&is->is_login_cv, "iscsi_login");
1885 #endif
1886
1887         /*
1888          * Set some default values, from RFC 3720, section 12.
1889          *
1890          * These values are updated by the handoff IOCTL, but are
1891          * needed prior to the handoff to support sending the ISER
1892          * login PDU.
1893          */
1894         is->is_conn->ic_max_recv_data_segment_length = 8192;
1895         is->is_conn->ic_max_send_data_segment_length = 8192;
1896         is->is_max_burst_length = 262144;
1897         is->is_first_burst_length = 65536;
1898
1899         is->is_softc = sc;
1900         sc->sc_last_session_id++;
1901         is->is_id = sc->sc_last_session_id;
1902         is->is_isid[0] = 0x80; /* RFC 3720, 10.12.5: 10b, "Random" ISID. */
1903         arc4rand(&is->is_isid[1], 5, 0);
1904         is->is_tsih = 0;
1905         callout_init(&is->is_callout, 1);
1906
1907         error = kthread_add(iscsi_maintenance_thread, is, NULL, NULL, 0, 0, "iscsimt");
1908         if (error != 0) {
1909                 ISCSI_SESSION_WARN(is, "kthread_add(9) failed with error %d", error);
1910                 sx_xunlock(&sc->sc_lock);
1911                 return (error);
1912         }
1913
1914         callout_reset(&is->is_callout, 1 * hz, iscsi_callout, is);
1915         TAILQ_INSERT_TAIL(&sc->sc_sessions, is, is_next);
1916
1917         ISCSI_SESSION_LOCK(is);
1918         /*
1919          * Don't notify iscsid(8) if the session is disabled and it's not
1920          * a discovery session,
1921          */
1922         if (is->is_conf.isc_enable == 0 && is->is_conf.isc_discovery == 0) {
1923                 ISCSI_SESSION_UNLOCK(is);
1924                 sx_xunlock(&sc->sc_lock);
1925                 return (0);
1926         }
1927
1928         is->is_waiting_for_iscsid = true;
1929         strlcpy(is->is_reason, "Waiting for iscsid(8)", sizeof(is->is_reason));
1930         ISCSI_SESSION_UNLOCK(is);
1931         cv_signal(&sc->sc_cv);
1932         sx_xunlock(&sc->sc_lock);
1933         return (0);
1934 }
1935
1936 static bool
1937 iscsi_session_conf_matches(unsigned int id1, const struct iscsi_session_conf *c1,
1938     unsigned int id2, const struct iscsi_session_conf *c2)
1939 {
1940
1941         if (id2 != 0 && id2 != id1)
1942                 return (false);
1943         if (c2->isc_target[0] != '\0' &&
1944             strcmp(c1->isc_target, c2->isc_target) != 0)
1945                 return (false);
1946         if (c2->isc_target_addr[0] != '\0' &&
1947             strcmp(c1->isc_target_addr, c2->isc_target_addr) != 0)
1948                 return (false);
1949         return (true);
1950 }
1951
1952 static int
1953 iscsi_ioctl_session_remove(struct iscsi_softc *sc,
1954     struct iscsi_session_remove *isr)
1955 {
1956         struct iscsi_session *is, *tmp;
1957         bool found = false;
1958
1959         iscsi_sanitize_session_conf(&isr->isr_conf);
1960
1961         sx_xlock(&sc->sc_lock);
1962         TAILQ_FOREACH_SAFE(is, &sc->sc_sessions, is_next, tmp) {
1963                 ISCSI_SESSION_LOCK(is);
1964                 if (iscsi_session_conf_matches(is->is_id, &is->is_conf,
1965                     isr->isr_session_id, &isr->isr_conf)) {
1966                         found = true;
1967                         iscsi_session_logout(is);
1968                         iscsi_session_terminate(is);
1969                 }
1970                 ISCSI_SESSION_UNLOCK(is);
1971         }
1972         sx_xunlock(&sc->sc_lock);
1973
1974         if (!found)
1975                 return (ESRCH);
1976
1977         return (0);
1978 }
1979
1980 static int
1981 iscsi_ioctl_session_list(struct iscsi_softc *sc, struct iscsi_session_list *isl)
1982 {
1983         int error;
1984         unsigned int i = 0;
1985         struct iscsi_session *is;
1986         struct iscsi_session_state iss;
1987
1988         sx_slock(&sc->sc_lock);
1989         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
1990                 if (i >= isl->isl_nentries) {
1991                         sx_sunlock(&sc->sc_lock);
1992                         return (EMSGSIZE);
1993                 }
1994                 memset(&iss, 0, sizeof(iss));
1995                 memcpy(&iss.iss_conf, &is->is_conf, sizeof(iss.iss_conf));
1996                 iss.iss_id = is->is_id;
1997                 strlcpy(iss.iss_target_alias, is->is_target_alias, sizeof(iss.iss_target_alias));
1998                 strlcpy(iss.iss_reason, is->is_reason, sizeof(iss.iss_reason));
1999                 strlcpy(iss.iss_offload, is->is_conn->ic_offload, sizeof(iss.iss_offload));
2000
2001                 if (is->is_conn->ic_header_crc32c)
2002                         iss.iss_header_digest = ISCSI_DIGEST_CRC32C;
2003                 else
2004                         iss.iss_header_digest = ISCSI_DIGEST_NONE;
2005
2006                 if (is->is_conn->ic_data_crc32c)
2007                         iss.iss_data_digest = ISCSI_DIGEST_CRC32C;
2008                 else
2009                         iss.iss_data_digest = ISCSI_DIGEST_NONE;
2010
2011                 iss.iss_max_send_data_segment_length =
2012                     is->is_conn->ic_max_send_data_segment_length;
2013                 iss.iss_max_recv_data_segment_length =
2014                     is->is_conn->ic_max_recv_data_segment_length;
2015                 iss.iss_max_burst_length = is->is_max_burst_length;
2016                 iss.iss_first_burst_length = is->is_first_burst_length;
2017                 iss.iss_immediate_data = is->is_immediate_data;
2018                 iss.iss_connected = is->is_connected;
2019
2020                 error = copyout(&iss, isl->isl_pstates + i, sizeof(iss));
2021                 if (error != 0) {
2022                         sx_sunlock(&sc->sc_lock);
2023                         return (error);
2024                 }
2025                 i++;
2026         }
2027         sx_sunlock(&sc->sc_lock);
2028
2029         isl->isl_nentries = i;
2030
2031         return (0);
2032 }
2033
2034 static int
2035 iscsi_ioctl_session_modify(struct iscsi_softc *sc,
2036     struct iscsi_session_modify *ism)
2037 {
2038         struct iscsi_session *is;
2039         const struct iscsi_session *is2;
2040
2041         iscsi_sanitize_session_conf(&ism->ism_conf);
2042         if (iscsi_valid_session_conf(&ism->ism_conf) == false)
2043                 return (EINVAL);
2044
2045         sx_xlock(&sc->sc_lock);
2046         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
2047                 ISCSI_SESSION_LOCK(is);
2048                 if (is->is_id == ism->ism_session_id) {
2049                         /* Note that the session remains locked. */
2050                         break;
2051                 }
2052                 ISCSI_SESSION_UNLOCK(is);
2053         }
2054         if (is == NULL) {
2055                 sx_xunlock(&sc->sc_lock);
2056                 return (ESRCH);
2057         }
2058
2059         /*
2060          * Prevent duplicates.
2061          */
2062         TAILQ_FOREACH(is2, &sc->sc_sessions, is_next) {
2063                 if (is == is2)
2064                         continue;
2065
2066                 if (!!ism->ism_conf.isc_discovery !=
2067                     !!is2->is_conf.isc_discovery)
2068                         continue;
2069
2070                 if (strcmp(ism->ism_conf.isc_target_addr,
2071                     is2->is_conf.isc_target_addr) != 0)
2072                         continue;
2073
2074                 if (ism->ism_conf.isc_discovery == 0 &&
2075                     strcmp(ism->ism_conf.isc_target,
2076                     is2->is_conf.isc_target) != 0)
2077                         continue;
2078
2079                 ISCSI_SESSION_UNLOCK(is);
2080                 sx_xunlock(&sc->sc_lock);
2081                 return (EBUSY);
2082         }
2083
2084         sx_xunlock(&sc->sc_lock);
2085
2086         memcpy(&is->is_conf, &ism->ism_conf, sizeof(is->is_conf));
2087         ISCSI_SESSION_UNLOCK(is);
2088
2089         iscsi_session_reconnect(is);
2090
2091         return (0);
2092 }
2093
2094 static int
2095 iscsi_ioctl(struct cdev *dev, u_long cmd, caddr_t arg, int mode,
2096     struct thread *td)
2097 {
2098         struct iscsi_softc *sc;
2099
2100         sc = dev->si_drv1;
2101
2102         switch (cmd) {
2103         case ISCSIDWAIT:
2104                 return (iscsi_ioctl_daemon_wait(sc,
2105                     (struct iscsi_daemon_request *)arg));
2106         case ISCSIDHANDOFF:
2107                 return (iscsi_ioctl_daemon_handoff(sc,
2108                     (struct iscsi_daemon_handoff *)arg));
2109         case ISCSIDFAIL:
2110                 return (iscsi_ioctl_daemon_fail(sc,
2111                     (struct iscsi_daemon_fail *)arg));
2112 #ifdef ICL_KERNEL_PROXY
2113         case ISCSIDCONNECT:
2114                 return (iscsi_ioctl_daemon_connect(sc,
2115                     (struct iscsi_daemon_connect *)arg));
2116         case ISCSIDSEND:
2117                 return (iscsi_ioctl_daemon_send(sc,
2118                     (struct iscsi_daemon_send *)arg));
2119         case ISCSIDRECEIVE:
2120                 return (iscsi_ioctl_daemon_receive(sc,
2121                     (struct iscsi_daemon_receive *)arg));
2122 #endif /* ICL_KERNEL_PROXY */
2123         case ISCSISADD:
2124                 return (iscsi_ioctl_session_add(sc,
2125                     (struct iscsi_session_add *)arg));
2126         case ISCSISREMOVE:
2127                 return (iscsi_ioctl_session_remove(sc,
2128                     (struct iscsi_session_remove *)arg));
2129         case ISCSISLIST:
2130                 return (iscsi_ioctl_session_list(sc,
2131                     (struct iscsi_session_list *)arg));
2132         case ISCSISMODIFY:
2133                 return (iscsi_ioctl_session_modify(sc,
2134                     (struct iscsi_session_modify *)arg));
2135         default:
2136                 return (EINVAL);
2137         }
2138 }
2139
2140 static struct iscsi_outstanding *
2141 iscsi_outstanding_find(struct iscsi_session *is, uint32_t initiator_task_tag)
2142 {
2143         struct iscsi_outstanding *io;
2144
2145         ISCSI_SESSION_LOCK_ASSERT(is);
2146
2147         TAILQ_FOREACH(io, &is->is_outstanding, io_next) {
2148                 if (io->io_initiator_task_tag == initiator_task_tag)
2149                         return (io);
2150         }
2151         return (NULL);
2152 }
2153
2154 static struct iscsi_outstanding *
2155 iscsi_outstanding_find_ccb(struct iscsi_session *is, union ccb *ccb)
2156 {
2157         struct iscsi_outstanding *io;
2158
2159         ISCSI_SESSION_LOCK_ASSERT(is);
2160
2161         TAILQ_FOREACH(io, &is->is_outstanding, io_next) {
2162                 if (io->io_ccb == ccb)
2163                         return (io);
2164         }
2165         return (NULL);
2166 }
2167
2168 static struct iscsi_outstanding *
2169 iscsi_outstanding_add(struct iscsi_session *is, struct icl_pdu *request,
2170     union ccb *ccb, uint32_t *initiator_task_tagp)
2171 {
2172         struct iscsi_outstanding *io;
2173         int error;
2174
2175         ISCSI_SESSION_LOCK_ASSERT(is);
2176
2177         io = uma_zalloc(iscsi_outstanding_zone, M_NOWAIT | M_ZERO);
2178         if (io == NULL) {
2179                 ISCSI_SESSION_WARN(is, "failed to allocate %zd bytes",
2180                     sizeof(*io));
2181                 return (NULL);
2182         }
2183
2184         error = icl_conn_task_setup(is->is_conn, request, &ccb->csio,
2185             initiator_task_tagp, &io->io_icl_prv);
2186         if (error != 0) {
2187                 ISCSI_SESSION_WARN(is,
2188                     "icl_conn_task_setup() failed with error %d", error);
2189                 uma_zfree(iscsi_outstanding_zone, io);
2190                 return (NULL);
2191         }
2192
2193         KASSERT(iscsi_outstanding_find(is, *initiator_task_tagp) == NULL,
2194             ("initiator_task_tag 0x%x already added", *initiator_task_tagp));
2195
2196         io->io_initiator_task_tag = *initiator_task_tagp;
2197         io->io_ccb = ccb;
2198         TAILQ_INSERT_TAIL(&is->is_outstanding, io, io_next);
2199         return (io);
2200 }
2201
2202 static void
2203 iscsi_outstanding_remove(struct iscsi_session *is, struct iscsi_outstanding *io)
2204 {
2205
2206         ISCSI_SESSION_LOCK_ASSERT(is);
2207
2208         icl_conn_task_done(is->is_conn, io->io_icl_prv);
2209         TAILQ_REMOVE(&is->is_outstanding, io, io_next);
2210         uma_zfree(iscsi_outstanding_zone, io);
2211 }
2212
2213 static void
2214 iscsi_action_abort(struct iscsi_session *is, union ccb *ccb)
2215 {
2216         struct icl_pdu *request;
2217         struct iscsi_bhs_task_management_request *bhstmr;
2218         struct ccb_abort *cab = &ccb->cab;
2219         struct iscsi_outstanding *io, *aio;
2220         uint32_t initiator_task_tag;
2221
2222         ISCSI_SESSION_LOCK_ASSERT(is);
2223
2224 #if 0
2225         KASSERT(is->is_login_phase == false, ("%s called during Login Phase", __func__));
2226 #else
2227         if (is->is_login_phase) {
2228                 ccb->ccb_h.status = CAM_REQ_ABORTED;
2229                 xpt_done(ccb);
2230                 return;
2231         }
2232 #endif
2233
2234         aio = iscsi_outstanding_find_ccb(is, cab->abort_ccb);
2235         if (aio == NULL) {
2236                 ccb->ccb_h.status = CAM_REQ_CMP;
2237                 xpt_done(ccb);
2238                 return;
2239         }
2240
2241         request = icl_pdu_new(is->is_conn, M_NOWAIT);
2242         if (request == NULL) {
2243                 ccb->ccb_h.status = CAM_RESRC_UNAVAIL;
2244                 xpt_done(ccb);
2245                 return;
2246         }
2247
2248         initiator_task_tag = is->is_initiator_task_tag++;
2249         if (initiator_task_tag == 0xffffffff)
2250                 initiator_task_tag = is->is_initiator_task_tag++;
2251
2252         io = iscsi_outstanding_add(is, request, NULL, &initiator_task_tag);
2253         if (io == NULL) {
2254                 icl_pdu_free(request);
2255                 ccb->ccb_h.status = CAM_RESRC_UNAVAIL;
2256                 xpt_done(ccb);
2257                 return;
2258         }
2259         io->io_referenced_task_tag = aio->io_initiator_task_tag;
2260
2261         bhstmr = (struct iscsi_bhs_task_management_request *)request->ip_bhs;
2262         bhstmr->bhstmr_opcode = ISCSI_BHS_OPCODE_TASK_REQUEST;
2263         bhstmr->bhstmr_function = 0x80 | BHSTMR_FUNCTION_ABORT_TASK;
2264         bhstmr->bhstmr_lun = htobe64(CAM_EXTLUN_BYTE_SWIZZLE(ccb->ccb_h.target_lun));
2265         bhstmr->bhstmr_initiator_task_tag = initiator_task_tag;
2266         bhstmr->bhstmr_referenced_task_tag = aio->io_initiator_task_tag;
2267
2268         iscsi_pdu_queue_locked(request);
2269 }
2270
2271 static void
2272 iscsi_action_scsiio(struct iscsi_session *is, union ccb *ccb)
2273 {
2274         struct icl_pdu *request;
2275         struct iscsi_bhs_scsi_command *bhssc;
2276         struct ccb_scsiio *csio;
2277         struct iscsi_outstanding *io;
2278         size_t len;
2279         uint32_t initiator_task_tag;
2280         int error;
2281
2282         ISCSI_SESSION_LOCK_ASSERT(is);
2283
2284 #if 0
2285         KASSERT(is->is_login_phase == false, ("%s called during Login Phase", __func__));
2286 #else
2287         if (is->is_login_phase) {
2288                 ISCSI_SESSION_DEBUG(is, "called during login phase");
2289                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
2290                         xpt_freeze_devq(ccb->ccb_h.path, 1);
2291                         ISCSI_SESSION_DEBUG(is, "freezing devq");
2292                 }
2293                 ccb->ccb_h.status = CAM_REQ_ABORTED | CAM_DEV_QFRZN;
2294                 xpt_done(ccb);
2295                 return;
2296         }
2297 #endif
2298
2299         request = icl_pdu_new(is->is_conn, M_NOWAIT);
2300         if (request == NULL) {
2301                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
2302                         xpt_freeze_devq(ccb->ccb_h.path, 1);
2303                         ISCSI_SESSION_DEBUG(is, "freezing devq");
2304                 }
2305                 ccb->ccb_h.status = CAM_RESRC_UNAVAIL | CAM_DEV_QFRZN;
2306                 xpt_done(ccb);
2307                 return;
2308         }
2309
2310         initiator_task_tag = is->is_initiator_task_tag++;
2311         if (initiator_task_tag == 0xffffffff)
2312                 initiator_task_tag = is->is_initiator_task_tag++;
2313
2314         io = iscsi_outstanding_add(is, request, ccb, &initiator_task_tag);
2315         if (io == NULL) {
2316                 icl_pdu_free(request);
2317                 if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
2318                         xpt_freeze_devq(ccb->ccb_h.path, 1);
2319                         ISCSI_SESSION_DEBUG(is, "freezing devq");
2320                 }
2321                 ccb->ccb_h.status = CAM_RESRC_UNAVAIL | CAM_DEV_QFRZN;
2322                 xpt_done(ccb);
2323                 return;
2324         }
2325
2326         csio = &ccb->csio;
2327         bhssc = (struct iscsi_bhs_scsi_command *)request->ip_bhs;
2328         bhssc->bhssc_opcode = ISCSI_BHS_OPCODE_SCSI_COMMAND;
2329         bhssc->bhssc_flags |= BHSSC_FLAGS_F;
2330         switch (csio->ccb_h.flags & CAM_DIR_MASK) {
2331         case CAM_DIR_IN:
2332                 bhssc->bhssc_flags |= BHSSC_FLAGS_R;
2333                 break;
2334         case CAM_DIR_OUT:
2335                 bhssc->bhssc_flags |= BHSSC_FLAGS_W;
2336                 break;
2337         }
2338
2339         if ((ccb->ccb_h.flags & CAM_TAG_ACTION_VALID) != 0) {
2340                 switch (csio->tag_action) {
2341                 case MSG_HEAD_OF_Q_TAG:
2342                         bhssc->bhssc_flags |= BHSSC_FLAGS_ATTR_HOQ;
2343                         break;
2344                 case MSG_ORDERED_Q_TAG:
2345                         bhssc->bhssc_flags |= BHSSC_FLAGS_ATTR_ORDERED;
2346                         break;
2347                 case MSG_ACA_TASK:
2348                         bhssc->bhssc_flags |= BHSSC_FLAGS_ATTR_ACA;
2349                         break;
2350                 case MSG_SIMPLE_Q_TAG:
2351                 default:
2352                         bhssc->bhssc_flags |= BHSSC_FLAGS_ATTR_SIMPLE;
2353                         break;
2354                 }
2355         } else
2356                 bhssc->bhssc_flags |= BHSSC_FLAGS_ATTR_UNTAGGED;
2357
2358         if (is->is_protocol_level >= 2) {
2359                 bhssc->bhssc_pri = (csio->priority << BHSSC_PRI_SHIFT) &
2360                     BHSSC_PRI_MASK;
2361         }
2362
2363         bhssc->bhssc_lun = htobe64(CAM_EXTLUN_BYTE_SWIZZLE(ccb->ccb_h.target_lun));
2364         bhssc->bhssc_initiator_task_tag = initiator_task_tag;
2365         bhssc->bhssc_expected_data_transfer_length = htonl(csio->dxfer_len);
2366         KASSERT(csio->cdb_len <= sizeof(bhssc->bhssc_cdb),
2367             ("unsupported CDB size %zd", (size_t)csio->cdb_len));
2368
2369         if (csio->ccb_h.flags & CAM_CDB_POINTER)
2370                 memcpy(&bhssc->bhssc_cdb, csio->cdb_io.cdb_ptr, csio->cdb_len);
2371         else
2372                 memcpy(&bhssc->bhssc_cdb, csio->cdb_io.cdb_bytes, csio->cdb_len);
2373
2374         if (is->is_immediate_data &&
2375             (csio->ccb_h.flags & CAM_DIR_MASK) == CAM_DIR_OUT) {
2376                 len = csio->dxfer_len;
2377                 //ISCSI_SESSION_DEBUG(is, "adding %zd of immediate data", len);
2378                 if (len > is->is_first_burst_length) {
2379                         ISCSI_SESSION_DEBUG(is, "len %zd -> %d", len, is->is_first_burst_length);
2380                         len = is->is_first_burst_length;
2381                 }
2382                 if (len > is->is_conn->ic_max_send_data_segment_length) {
2383                         ISCSI_SESSION_DEBUG(is, "len %zd -> %d", len,
2384                             is->is_conn->ic_max_send_data_segment_length);
2385                         len = is->is_conn->ic_max_send_data_segment_length;
2386                 }
2387
2388                 error = icl_pdu_append_data(request, csio->data_ptr, len, M_NOWAIT);
2389                 if (error != 0) {
2390                         iscsi_outstanding_remove(is, io);
2391                         icl_pdu_free(request);
2392                         if ((ccb->ccb_h.status & CAM_DEV_QFRZN) == 0) {
2393                                 xpt_freeze_devq(ccb->ccb_h.path, 1);
2394                                 ISCSI_SESSION_DEBUG(is, "freezing devq");
2395                         }
2396                         ccb->ccb_h.status = CAM_RESRC_UNAVAIL | CAM_DEV_QFRZN;
2397                         xpt_done(ccb);
2398                         return;
2399                 }
2400         }
2401         iscsi_pdu_queue_locked(request);
2402 }
2403
2404 static void
2405 iscsi_action(struct cam_sim *sim, union ccb *ccb)
2406 {
2407         struct iscsi_session *is;
2408
2409         is = cam_sim_softc(sim);
2410
2411         ISCSI_SESSION_LOCK_ASSERT(is);
2412
2413         if (is->is_terminating ||
2414             (is->is_connected == false && fail_on_disconnection)) {
2415                 ccb->ccb_h.status = CAM_DEV_NOT_THERE;
2416                 xpt_done(ccb);
2417                 return;
2418         }
2419
2420         /*
2421          * Make sure CAM doesn't sneak in a CCB just after freezing the queue.
2422          */
2423         if (is->is_simq_frozen == true) {
2424                 ccb->ccb_h.status &= ~(CAM_SIM_QUEUED | CAM_STATUS_MASK);
2425                 ccb->ccb_h.status |= CAM_REQUEUE_REQ;
2426                 /* Don't freeze the devq - the SIM queue is already frozen. */
2427                 xpt_done(ccb);
2428                 return;
2429         }
2430
2431         switch (ccb->ccb_h.func_code) {
2432         case XPT_PATH_INQ:
2433         {
2434                 struct ccb_pathinq *cpi = &ccb->cpi;
2435
2436                 cpi->version_num = 1;
2437                 cpi->hba_inquiry = PI_TAG_ABLE;
2438                 cpi->target_sprt = 0;
2439                 cpi->hba_misc = PIM_EXTLUNS;
2440                 /*
2441                  * XXX: It shouldn't ever be NULL; this could be turned
2442                  *      into a KASSERT eventually.
2443                  */
2444                 if (is->is_conn == NULL)
2445                         ISCSI_WARN("NULL conn");
2446                 else if (is->is_conn->ic_unmapped)
2447                         cpi->hba_misc |= PIM_UNMAPPED;
2448                 cpi->hba_eng_cnt = 0;
2449                 cpi->max_target = 0;
2450                 /*
2451                  * Note that the variable below is only relevant for targets
2452                  * that don't claim compliance with anything above SPC2, which
2453                  * means they don't support REPORT_LUNS.
2454                  */
2455                 cpi->max_lun = 255;
2456                 cpi->initiator_id = ~0;
2457                 strlcpy(cpi->sim_vid, "FreeBSD", SIM_IDLEN);
2458                 strlcpy(cpi->hba_vid, "iSCSI", HBA_IDLEN);
2459                 strlcpy(cpi->dev_name, cam_sim_name(sim), DEV_IDLEN);
2460                 cpi->unit_number = cam_sim_unit(sim);
2461                 cpi->bus_id = cam_sim_bus(sim);
2462                 cpi->base_transfer_speed = 150000; /* XXX */
2463                 cpi->transport = XPORT_ISCSI;
2464                 cpi->transport_version = 0;
2465                 cpi->protocol = PROTO_SCSI;
2466                 cpi->protocol_version = SCSI_REV_SPC3;
2467                 cpi->maxio = maxphys;
2468                 cpi->ccb_h.status = CAM_REQ_CMP;
2469                 break;
2470         }
2471         case XPT_GET_TRAN_SETTINGS:
2472         {
2473                 struct ccb_trans_settings       *cts;
2474                 struct ccb_trans_settings_scsi  *scsi;
2475
2476                 cts = &ccb->cts;
2477                 scsi = &cts->proto_specific.scsi;
2478
2479                 cts->protocol = PROTO_SCSI;
2480                 cts->protocol_version = SCSI_REV_SPC3;
2481                 cts->transport = XPORT_ISCSI;
2482                 cts->transport_version = 0;
2483                 scsi->valid = CTS_SCSI_VALID_TQ;
2484                 scsi->flags = CTS_SCSI_FLAGS_TAG_ENB;
2485                 cts->ccb_h.status = CAM_REQ_CMP;
2486                 break;
2487         }
2488         case XPT_CALC_GEOMETRY:
2489                 cam_calc_geometry(&ccb->ccg, /*extended*/1);
2490                 ccb->ccb_h.status = CAM_REQ_CMP;
2491                 break;
2492 #if 0
2493         /*
2494          * XXX: What's the point?
2495          */
2496         case XPT_RESET_BUS:
2497         case XPT_TERM_IO:
2498                 ISCSI_SESSION_DEBUG(is, "faking success for reset, abort, or term_io");
2499                 ccb->ccb_h.status = CAM_REQ_CMP;
2500                 break;
2501 #endif
2502         case XPT_ABORT:
2503                 iscsi_action_abort(is, ccb);
2504                 return;
2505         case XPT_SCSI_IO:
2506                 iscsi_action_scsiio(is, ccb);
2507                 return;
2508         default:
2509 #if 0
2510                 ISCSI_SESSION_DEBUG(is, "got unsupported code 0x%x", ccb->ccb_h.func_code);
2511 #endif
2512                 ccb->ccb_h.status = CAM_FUNC_NOTAVAIL;
2513                 break;
2514         }
2515         xpt_done(ccb);
2516 }
2517
2518 static void
2519 iscsi_terminate_sessions(struct iscsi_softc *sc)
2520 {
2521         struct iscsi_session *is;
2522
2523         sx_slock(&sc->sc_lock);
2524         TAILQ_FOREACH(is, &sc->sc_sessions, is_next)
2525                 iscsi_session_terminate(is);
2526         while(!TAILQ_EMPTY(&sc->sc_sessions)) {
2527                 ISCSI_DEBUG("waiting for sessions to terminate");
2528                 cv_wait(&sc->sc_cv, &sc->sc_lock);
2529         }
2530         ISCSI_DEBUG("all sessions terminated");
2531         sx_sunlock(&sc->sc_lock);
2532 }
2533
2534 static void
2535 iscsi_shutdown_pre(struct iscsi_softc *sc)
2536 {
2537         struct iscsi_session *is;
2538
2539         if (!fail_on_shutdown)
2540                 return;
2541
2542         /*
2543          * If we have any sessions waiting for reconnection, request
2544          * maintenance thread to fail them immediately instead of waiting
2545          * for reconnect timeout.
2546          *
2547          * This prevents LUNs with mounted filesystems that are supported
2548          * by disconnected iSCSI sessions from hanging, however it will
2549          * fail all queued BIOs.
2550          */
2551         ISCSI_DEBUG("forcing failing all disconnected sessions due to shutdown");
2552
2553         fail_on_disconnection = 1;
2554
2555         sx_slock(&sc->sc_lock);
2556         TAILQ_FOREACH(is, &sc->sc_sessions, is_next) {
2557                 ISCSI_SESSION_LOCK(is);
2558                 if (!is->is_connected) {
2559                         ISCSI_SESSION_DEBUG(is, "force failing disconnected session early");
2560                         iscsi_session_reconnect(is);
2561                 }
2562                 ISCSI_SESSION_UNLOCK(is);
2563         }
2564         sx_sunlock(&sc->sc_lock);
2565 }
2566
2567 static void
2568 iscsi_shutdown_post(struct iscsi_softc *sc)
2569 {
2570
2571         if (!KERNEL_PANICKED()) {
2572                 ISCSI_DEBUG("removing all sessions due to shutdown");
2573                 iscsi_terminate_sessions(sc);
2574         }
2575 }
2576
2577 static int
2578 iscsi_load(void)
2579 {
2580         int error;
2581
2582         sc = malloc(sizeof(*sc), M_ISCSI, M_ZERO | M_WAITOK);
2583         sx_init(&sc->sc_lock, "iscsi");
2584         TAILQ_INIT(&sc->sc_sessions);
2585         cv_init(&sc->sc_cv, "iscsi_cv");
2586
2587         iscsi_outstanding_zone = uma_zcreate("iscsi_outstanding",
2588             sizeof(struct iscsi_outstanding), NULL, NULL, NULL, NULL,
2589             UMA_ALIGN_PTR, 0);
2590
2591         error = make_dev_p(MAKEDEV_CHECKNAME, &sc->sc_cdev, &iscsi_cdevsw,
2592             NULL, UID_ROOT, GID_WHEEL, 0600, "iscsi");
2593         if (error != 0) {
2594                 ISCSI_WARN("failed to create device node, error %d", error);
2595                 return (error);
2596         }
2597         sc->sc_cdev->si_drv1 = sc;
2598
2599         sc->sc_shutdown_pre_eh = EVENTHANDLER_REGISTER(shutdown_pre_sync,
2600             iscsi_shutdown_pre, sc, SHUTDOWN_PRI_FIRST);
2601         /*
2602          * shutdown_post_sync needs to run after filesystem shutdown and before
2603          * CAM shutdown - otherwise when rebooting with an iSCSI session that is
2604          * disconnected but has outstanding requests, dashutdown() will hang on
2605          * cam_periph_runccb().
2606          */
2607         sc->sc_shutdown_post_eh = EVENTHANDLER_REGISTER(shutdown_post_sync,
2608             iscsi_shutdown_post, sc, SHUTDOWN_PRI_DEFAULT - 1);
2609
2610         return (0);
2611 }
2612
2613 static int
2614 iscsi_unload(void)
2615 {
2616
2617         /* Awaken any threads asleep in iscsi_ioctl(). */
2618         sx_xlock(&sc->sc_lock);
2619         sc->sc_unloading = true;
2620         cv_signal(&sc->sc_cv);
2621         sx_xunlock(&sc->sc_lock);
2622
2623         if (sc->sc_cdev != NULL) {
2624                 ISCSI_DEBUG("removing device node");
2625                 destroy_dev(sc->sc_cdev);
2626                 ISCSI_DEBUG("device node removed");
2627         }
2628
2629         if (sc->sc_shutdown_pre_eh != NULL)
2630                 EVENTHANDLER_DEREGISTER(shutdown_pre_sync, sc->sc_shutdown_pre_eh);
2631         if (sc->sc_shutdown_post_eh != NULL)
2632                 EVENTHANDLER_DEREGISTER(shutdown_post_sync, sc->sc_shutdown_post_eh);
2633
2634         iscsi_terminate_sessions(sc);
2635
2636         uma_zdestroy(iscsi_outstanding_zone);
2637         sx_destroy(&sc->sc_lock);
2638         cv_destroy(&sc->sc_cv);
2639         free(sc, M_ISCSI);
2640         return (0);
2641 }
2642
2643 static int
2644 iscsi_quiesce(void)
2645 {
2646         sx_slock(&sc->sc_lock);
2647         if (!TAILQ_EMPTY(&sc->sc_sessions)) {
2648                 sx_sunlock(&sc->sc_lock);
2649                 return (EBUSY);
2650         }
2651         sx_sunlock(&sc->sc_lock);
2652         return (0);
2653 }
2654
2655 static int
2656 iscsi_modevent(module_t mod, int what, void *arg)
2657 {
2658         int error;
2659
2660         switch (what) {
2661         case MOD_LOAD:
2662                 error = iscsi_load();
2663                 break;
2664         case MOD_UNLOAD:
2665                 error = iscsi_unload();
2666                 break;
2667         case MOD_QUIESCE:
2668                 error = iscsi_quiesce();
2669                 break;
2670         default:
2671                 error = EINVAL;
2672                 break;
2673         }
2674         return (error);
2675 }
2676
2677 moduledata_t iscsi_data = {
2678         "iscsi",
2679         iscsi_modevent,
2680         0
2681 };
2682
2683 DECLARE_MODULE(iscsi, iscsi_data, SI_SUB_DRIVERS, SI_ORDER_MIDDLE);
2684 MODULE_DEPEND(iscsi, cam, 1, 1, 1);
2685 MODULE_DEPEND(iscsi, icl, 1, 1, 1);