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