1 /*- 2 * Copyright (c) 2012 The DragonFly Project. All rights reserved. 3 * 4 * This code is derived from software contributed to The DragonFly Project 5 * by Matthew Dillon <dillon@backplane.com> 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in 15 * the documentation and/or other materials provided with the 16 * distribution. 17 * 3. Neither the name of The DragonFly Project nor the names of its 18 * contributors may be used to endorse or promote products derived 19 * from this software without specific, prior written permission. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 22 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 23 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS 24 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE 25 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, 26 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING, 27 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; 28 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED 29 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, 30 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT 31 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 */ 34 /* 35 * TODO: txcmd CREATE state is deferred by tx msgq, need to calculate 36 * a streaming response. See subr_diskiocom()'s diskiodone(). 37 */ 38 #include <sys/param.h> 39 #include <sys/types.h> 40 #include <sys/kernel.h> 41 #include <sys/conf.h> 42 #include <sys/systm.h> 43 #include <sys/queue.h> 44 #include <sys/tree.h> 45 #include <sys/malloc.h> 46 #include <sys/mount.h> 47 #include <sys/socket.h> 48 #include <sys/vnode.h> 49 #include <sys/sysctl.h> 50 #include <sys/file.h> 51 #include <sys/proc.h> 52 #include <sys/priv.h> 53 #include <sys/thread.h> 54 #include <sys/globaldata.h> 55 #include <sys/limits.h> 56 57 #include <sys/dmsg.h> 58 59 RB_GENERATE(kdmsg_state_tree, kdmsg_state, rbnode, kdmsg_state_cmp); 60 61 SYSCTL_NODE(, OID_AUTO, kdmsg, CTLFLAG_RW, 0, "kdmsg"); 62 static int kdmsg_debug = 1; 63 SYSCTL_INT(_kdmsg, OID_AUTO, debug, CTLFLAG_RW, &kdmsg_debug, 0, 64 "Set debug level for kernel dmsg layer"); 65 66 #define kd_printf(level, ctl, ...) \ 67 if (kdmsg_debug >= (level)) kprintf("kdmsg: " ctl, __VA_ARGS__) 68 69 #define kdio_printf(iocom, level, ctl, ...) \ 70 if (kdmsg_debug >= (level)) kprintf("kdmsg: " ctl, __VA_ARGS__) 71 72 static int kdmsg_msg_receive_handling(kdmsg_msg_t *msg); 73 static int kdmsg_state_msgrx(kdmsg_msg_t *msg); 74 static int kdmsg_state_msgtx(kdmsg_msg_t *msg); 75 static void kdmsg_msg_write_locked(kdmsg_iocom_t *iocom, kdmsg_msg_t *msg); 76 static void kdmsg_state_cleanuprx(kdmsg_msg_t *msg); 77 static void kdmsg_state_cleanuptx(kdmsg_msg_t *msg); 78 static void kdmsg_subq_delete(kdmsg_state_t *state); 79 static void kdmsg_simulate_failure(kdmsg_state_t *state, int meto, int error); 80 static void kdmsg_state_abort(kdmsg_state_t *state); 81 static void kdmsg_state_dying(kdmsg_state_t *state); 82 static void kdmsg_state_free(kdmsg_state_t *state); 83 84 #ifdef KDMSG_DEBUG 85 #define KDMSG_DEBUG_ARGS , const char *file, int line 86 #define kdmsg_state_hold(state) _kdmsg_state_hold(state, __FILE__, __LINE__) 87 #define kdmsg_state_drop(state) _kdmsg_state_drop(state, __FILE__, __LINE__) 88 #else 89 #define KDMSG_DEBUG_ARGS 90 #define kdmsg_state_hold(state) _kdmsg_state_hold(state) 91 #define kdmsg_state_drop(state) _kdmsg_state_drop(state) 92 #endif 93 static void _kdmsg_state_hold(kdmsg_state_t *state KDMSG_DEBUG_ARGS); 94 static void _kdmsg_state_drop(kdmsg_state_t *state KDMSG_DEBUG_ARGS); 95 96 static void kdmsg_iocom_thread_rd(void *arg); 97 static void kdmsg_iocom_thread_wr(void *arg); 98 static int kdmsg_autorxmsg(kdmsg_msg_t *msg); 99 100 /*static struct lwkt_token kdmsg_token = LWKT_TOKEN_INITIALIZER(kdmsg_token);*/ 101 102 /* 103 * Initialize the roll-up communications structure for a network 104 * messaging session. This function does not install the socket. 105 */ 106 void 107 kdmsg_iocom_init(kdmsg_iocom_t *iocom, void *handle, uint32_t flags, 108 struct malloc_type *mmsg, 109 int (*rcvmsg)(kdmsg_msg_t *msg)) 110 { 111 bzero(iocom, sizeof(*iocom)); 112 iocom->handle = handle; 113 iocom->mmsg = mmsg; 114 iocom->rcvmsg = rcvmsg; 115 iocom->flags = flags; 116 lockinit(&iocom->msglk, "h2msg", 0, 0); 117 TAILQ_INIT(&iocom->msgq); 118 RB_INIT(&iocom->staterd_tree); 119 RB_INIT(&iocom->statewr_tree); 120 121 iocom->state0.iocom = iocom; 122 iocom->state0.parent = &iocom->state0; 123 TAILQ_INIT(&iocom->state0.subq); 124 } 125 126 /* 127 * [Re]connect using the passed file pointer. The caller must ref the 128 * fp for us. We own that ref now. 129 */ 130 void 131 kdmsg_iocom_reconnect(kdmsg_iocom_t *iocom, struct file *fp, 132 const char *subsysname) 133 { 134 /* 135 * Destroy the current connection 136 */ 137 lockmgr(&iocom->msglk, LK_EXCLUSIVE); 138 atomic_set_int(&iocom->msg_ctl, KDMSG_CLUSTERCTL_KILLRX); 139 while (iocom->msgrd_td || iocom->msgwr_td) { 140 wakeup(&iocom->msg_ctl); 141 lksleep(iocom, &iocom->msglk, 0, "clstrkl", hz); 142 } 143 144 /* 145 * Drop communications descriptor 146 */ 147 if (iocom->msg_fp) { 148 fdrop(iocom->msg_fp); 149 iocom->msg_fp = NULL; 150 } 151 152 /* 153 * Setup new communications descriptor 154 */ 155 iocom->msg_ctl = 0; 156 iocom->msg_fp = fp; 157 iocom->msg_seq = 0; 158 iocom->flags &= ~KDMSG_IOCOMF_EXITNOACC; 159 160 lwkt_create(kdmsg_iocom_thread_rd, iocom, &iocom->msgrd_td, 161 NULL, 0, -1, "%s-msgrd", subsysname); 162 lwkt_create(kdmsg_iocom_thread_wr, iocom, &iocom->msgwr_td, 163 NULL, 0, -1, "%s-msgwr", subsysname); 164 lockmgr(&iocom->msglk, LK_RELEASE); 165 } 166 167 /* 168 * Caller sets up iocom->auto_lnk_conn and iocom->auto_lnk_span, then calls 169 * this function to handle the state machine for LNK_CONN and LNK_SPAN. 170 */ 171 static int kdmsg_lnk_conn_reply(kdmsg_state_t *state, kdmsg_msg_t *msg); 172 static int kdmsg_lnk_span_reply(kdmsg_state_t *state, kdmsg_msg_t *msg); 173 174 void 175 kdmsg_iocom_autoinitiate(kdmsg_iocom_t *iocom, 176 void (*auto_callback)(kdmsg_msg_t *msg)) 177 { 178 kdmsg_msg_t *msg; 179 180 iocom->auto_callback = auto_callback; 181 182 msg = kdmsg_msg_alloc(&iocom->state0, 183 DMSG_LNK_CONN | DMSGF_CREATE, 184 kdmsg_lnk_conn_reply, NULL); 185 iocom->auto_lnk_conn.head = msg->any.head; 186 msg->any.lnk_conn = iocom->auto_lnk_conn; 187 iocom->conn_state = msg->state; 188 kdmsg_state_hold(msg->state); /* iocom->conn_state */ 189 kdmsg_msg_write(msg); 190 } 191 192 static 193 int 194 kdmsg_lnk_conn_reply(kdmsg_state_t *state, kdmsg_msg_t *msg) 195 { 196 kdmsg_iocom_t *iocom = state->iocom; 197 kdmsg_msg_t *rmsg; 198 199 /* 200 * Upon receipt of the LNK_CONN acknowledgement initiate an 201 * automatic SPAN if we were asked to. Used by e.g. xdisk, but 202 * not used by HAMMER2 which must manage more than one transmitted 203 * SPAN. 204 */ 205 if ((msg->any.head.cmd & DMSGF_CREATE) && 206 (iocom->flags & KDMSG_IOCOMF_AUTOTXSPAN)) { 207 rmsg = kdmsg_msg_alloc(&iocom->state0, 208 DMSG_LNK_SPAN | DMSGF_CREATE, 209 kdmsg_lnk_span_reply, NULL); 210 iocom->auto_lnk_span.head = rmsg->any.head; 211 rmsg->any.lnk_span = iocom->auto_lnk_span; 212 kdmsg_msg_write(rmsg); 213 } 214 215 /* 216 * Process shim after the CONN is acknowledged and before the CONN 217 * transaction is deleted. For deletions this gives device drivers 218 * the ability to interlock new operations on the circuit before 219 * it becomes illegal and panics. 220 */ 221 if (iocom->auto_callback) 222 iocom->auto_callback(msg); 223 224 if ((state->txcmd & DMSGF_DELETE) == 0 && 225 (msg->any.head.cmd & DMSGF_DELETE)) { 226 /* 227 * iocom->conn_state has a state ref, drop it when clearing. 228 */ 229 if (iocom->conn_state) 230 kdmsg_state_drop(iocom->conn_state); 231 iocom->conn_state = NULL; 232 kdmsg_msg_reply(msg, 0); 233 } 234 235 return (0); 236 } 237 238 static 239 int 240 kdmsg_lnk_span_reply(kdmsg_state_t *state, kdmsg_msg_t *msg) 241 { 242 /* 243 * Be sure to process shim before terminating the SPAN 244 * transaction. Gives device drivers the ability to 245 * interlock new operations on the circuit before it 246 * becomes illegal and panics. 247 */ 248 if (state->iocom->auto_callback) 249 state->iocom->auto_callback(msg); 250 251 if ((state->txcmd & DMSGF_DELETE) == 0 && 252 (msg->any.head.cmd & DMSGF_DELETE)) { 253 kdmsg_msg_reply(msg, 0); 254 } 255 return (0); 256 } 257 258 /* 259 * Disconnect and clean up 260 */ 261 void 262 kdmsg_iocom_uninit(kdmsg_iocom_t *iocom) 263 { 264 kdmsg_state_t *state; 265 kdmsg_msg_t *msg; 266 int retries; 267 268 /* 269 * Ask the cluster controller to go away by setting 270 * KILLRX. Send a PING to get a response to unstick reading 271 * from the pipe. 272 * 273 * After 10 seconds shitcan the pipe and do an unclean shutdown. 274 */ 275 lockmgr(&iocom->msglk, LK_EXCLUSIVE); 276 277 atomic_set_int(&iocom->msg_ctl, KDMSG_CLUSTERCTL_KILLRX); 278 msg = kdmsg_msg_alloc(&iocom->state0, DMSG_LNK_PING, NULL, NULL); 279 kdmsg_msg_write_locked(iocom, msg); 280 281 retries = 10; 282 while (iocom->msgrd_td || iocom->msgwr_td) { 283 wakeup(&iocom->msg_ctl); 284 lksleep(iocom, &iocom->msglk, 0, "clstrkl", hz); 285 if (--retries == 0 && iocom->msg_fp) { 286 kdio_printf(iocom, 0, "%s\n", 287 "iocom_uninit: " 288 "shitcanning unresponsive pipe"); 289 fp_shutdown(iocom->msg_fp, SHUT_RDWR); 290 /* retries allowed to go negative, keep looping */ 291 } 292 } 293 294 /* 295 * Cleanup caches 296 */ 297 if ((state = iocom->freerd_state) != NULL) { 298 iocom->freerd_state = NULL; 299 kdmsg_state_drop(state); 300 } 301 302 if ((state = iocom->freewr_state) != NULL) { 303 iocom->freewr_state = NULL; 304 kdmsg_state_drop(state); 305 } 306 307 /* 308 * Drop communications descriptor 309 */ 310 if (iocom->msg_fp) { 311 fdrop(iocom->msg_fp); 312 iocom->msg_fp = NULL; 313 } 314 lockmgr(&iocom->msglk, LK_RELEASE); 315 } 316 317 /* 318 * Cluster controller thread. Perform messaging functions. We have one 319 * thread for the reader and one for the writer. The writer handles 320 * shutdown requests (which should break the reader thread). 321 */ 322 static 323 void 324 kdmsg_iocom_thread_rd(void *arg) 325 { 326 kdmsg_iocom_t *iocom = arg; 327 dmsg_hdr_t hdr; 328 kdmsg_msg_t *msg = NULL; 329 size_t hbytes; 330 size_t abytes; 331 int error = 0; 332 333 while ((iocom->msg_ctl & KDMSG_CLUSTERCTL_KILLRX) == 0) { 334 /* 335 * Retrieve the message from the pipe or socket. 336 */ 337 error = fp_read(iocom->msg_fp, &hdr, sizeof(hdr), 338 NULL, 1, UIO_SYSSPACE); 339 if (error) 340 break; 341 if (hdr.magic != DMSG_HDR_MAGIC) { 342 kdio_printf(iocom, 1, "bad magic: %04x\n", hdr.magic); 343 error = EINVAL; 344 break; 345 } 346 hbytes = (hdr.cmd & DMSGF_SIZE) * DMSG_ALIGN; 347 if (hbytes < sizeof(hdr) || hbytes > DMSG_HDR_MAX) { 348 kdio_printf(iocom, 1, "bad header size %zd\n", hbytes); 349 error = EINVAL; 350 break; 351 } 352 353 /* XXX messy: mask cmd to avoid allocating state */ 354 msg = kdmsg_msg_alloc(&iocom->state0, 355 hdr.cmd & DMSGF_BASECMDMASK, 356 NULL, NULL); 357 msg->any.head = hdr; 358 msg->hdr_size = hbytes; 359 if (hbytes > sizeof(hdr)) { 360 error = fp_read(iocom->msg_fp, &msg->any.head + 1, 361 hbytes - sizeof(hdr), 362 NULL, 1, UIO_SYSSPACE); 363 if (error) { 364 kdio_printf(iocom, 1, "%s\n", 365 "short msg received"); 366 error = EINVAL; 367 break; 368 } 369 } 370 msg->aux_size = hdr.aux_bytes; 371 if (msg->aux_size > DMSG_AUX_MAX) { 372 kdio_printf(iocom, 1, 373 "illegal msg payload size %zd\n", 374 msg->aux_size); 375 error = EINVAL; 376 break; 377 } 378 if (msg->aux_size) { 379 abytes = DMSG_DOALIGN(msg->aux_size); 380 msg->aux_data = kmalloc(abytes, iocom->mmsg, M_WAITOK); 381 msg->flags |= KDMSG_FLAG_AUXALLOC; 382 error = fp_read(iocom->msg_fp, msg->aux_data, 383 abytes, NULL, 1, UIO_SYSSPACE); 384 if (error) { 385 kdio_printf(iocom, 1, "%s\n", 386 "short msg payload received"); 387 break; 388 } 389 } 390 391 error = kdmsg_msg_receive_handling(msg); 392 msg = NULL; 393 } 394 395 kdio_printf(iocom, 1, "read thread terminating error=%d\n", error); 396 397 lockmgr(&iocom->msglk, LK_EXCLUSIVE); 398 if (msg) 399 kdmsg_msg_free(msg); 400 401 /* 402 * Shutdown the socket and set KILLRX for consistency in case the 403 * shutdown was not commanded. Signal the transmit side to shutdown 404 * by setting KILLTX and waking it up. 405 */ 406 fp_shutdown(iocom->msg_fp, SHUT_RDWR); 407 atomic_set_int(&iocom->msg_ctl, KDMSG_CLUSTERCTL_KILLRX | 408 KDMSG_CLUSTERCTL_KILLTX); 409 iocom->msgrd_td = NULL; 410 lockmgr(&iocom->msglk, LK_RELEASE); 411 wakeup(&iocom->msg_ctl); 412 413 /* 414 * iocom can be ripped out at any time once the lock is 415 * released with msgrd_td set to NULL. The wakeup()s are safe but 416 * that is all. 417 */ 418 wakeup(iocom); 419 lwkt_exit(); 420 } 421 422 static 423 void 424 kdmsg_iocom_thread_wr(void *arg) 425 { 426 kdmsg_iocom_t *iocom = arg; 427 kdmsg_msg_t *msg; 428 ssize_t res; 429 size_t abytes; 430 int error = 0; 431 int save_ticks; 432 int didwarn; 433 434 /* 435 * Transmit loop 436 */ 437 msg = NULL; 438 lockmgr(&iocom->msglk, LK_EXCLUSIVE); 439 440 while ((iocom->msg_ctl & KDMSG_CLUSTERCTL_KILLTX) == 0 && error == 0) { 441 /* 442 * Sleep if no messages pending. Interlock with flag while 443 * holding msglk. 444 */ 445 if (TAILQ_EMPTY(&iocom->msgq)) { 446 atomic_set_int(&iocom->msg_ctl, 447 KDMSG_CLUSTERCTL_SLEEPING); 448 lksleep(&iocom->msg_ctl, &iocom->msglk, 0, "msgwr", hz); 449 atomic_clear_int(&iocom->msg_ctl, 450 KDMSG_CLUSTERCTL_SLEEPING); 451 } 452 453 while ((msg = TAILQ_FIRST(&iocom->msgq)) != NULL) { 454 /* 455 * Remove msg from the transmit queue and do 456 * persist and half-closed state handling. 457 */ 458 TAILQ_REMOVE(&iocom->msgq, msg, qentry); 459 460 error = kdmsg_state_msgtx(msg); 461 if (error == EALREADY) { 462 error = 0; 463 kdmsg_msg_free(msg); 464 continue; 465 } 466 if (error) { 467 kdmsg_msg_free(msg); 468 break; 469 } 470 471 /* 472 * Dump the message to the pipe or socket. 473 * 474 * We have to clean up the message as if the transmit 475 * succeeded even if it failed. 476 */ 477 lockmgr(&iocom->msglk, LK_RELEASE); 478 error = fp_write(iocom->msg_fp, &msg->any, 479 msg->hdr_size, &res, UIO_SYSSPACE); 480 if (error || res != msg->hdr_size) { 481 if (error == 0) 482 error = EINVAL; 483 lockmgr(&iocom->msglk, LK_EXCLUSIVE); 484 kdmsg_state_cleanuptx(msg); 485 break; 486 } 487 if (msg->aux_size) { 488 abytes = DMSG_DOALIGN(msg->aux_size); 489 error = fp_write(iocom->msg_fp, 490 msg->aux_data, abytes, 491 &res, UIO_SYSSPACE); 492 if (error || res != abytes) { 493 if (error == 0) 494 error = EINVAL; 495 lockmgr(&iocom->msglk, LK_EXCLUSIVE); 496 kdmsg_state_cleanuptx(msg); 497 break; 498 } 499 } 500 lockmgr(&iocom->msglk, LK_EXCLUSIVE); 501 kdmsg_state_cleanuptx(msg); 502 } 503 } 504 505 kdio_printf(iocom, 1, "write thread terminating error=%d\n", error); 506 507 /* 508 * Shutdown the socket and set KILLTX for consistency in case the 509 * shutdown was not commanded. Signal the receive side to shutdown 510 * by setting KILLRX and waking it up. 511 */ 512 fp_shutdown(iocom->msg_fp, SHUT_RDWR); 513 atomic_set_int(&iocom->msg_ctl, KDMSG_CLUSTERCTL_KILLRX | 514 KDMSG_CLUSTERCTL_KILLTX); 515 wakeup(&iocom->msg_ctl); 516 517 /* 518 * The transmit thread is responsible for final cleanups, wait 519 * for the receive side to terminate to prevent new received 520 * states from interfering with our cleanup. 521 * 522 * Do not set msgwr_td to NULL until we actually exit. 523 */ 524 while (iocom->msgrd_td) { 525 wakeup(&iocom->msg_ctl); 526 lksleep(iocom, &iocom->msglk, 0, "clstrkt", hz); 527 } 528 529 /* 530 * We can no longer receive new messages. We must drain the transmit 531 * message queue and simulate received messages to close anay remaining 532 * states. 533 * 534 * Loop until all the states are gone and there are no messages 535 * pending transmit. 536 */ 537 save_ticks = ticks; 538 didwarn = 0; 539 540 while (TAILQ_FIRST(&iocom->msgq) || 541 RB_ROOT(&iocom->staterd_tree) || 542 RB_ROOT(&iocom->statewr_tree)) { 543 /* 544 * Simulate failure for all sub-states of state0. 545 */ 546 kdmsg_drain_msgq(iocom); 547 kdio_printf(iocom, 2, "%s\n", 548 "simulate failure for all substates of state0"); 549 kdmsg_simulate_failure(&iocom->state0, 0, DMSG_ERR_LOSTLINK); 550 551 lksleep(iocom, &iocom->msglk, 0, "clstrtk", hz / 2); 552 553 if ((int)(ticks - save_ticks) > hz*2 && didwarn == 0) { 554 didwarn = 1; 555 kdio_printf(iocom, 0, 556 "Warning, write thread on %p " 557 "still terminating\n", 558 iocom); 559 } 560 if ((int)(ticks - save_ticks) > hz*15 && didwarn == 1) { 561 didwarn = 2; 562 kdio_printf(iocom, 0, 563 "Warning, write thread on %p " 564 "still terminating\n", 565 iocom); 566 } 567 if ((int)(ticks - save_ticks) > hz*60) { 568 kdio_printf(iocom, 0, 569 "Can't terminate: msgq %p " 570 "rd_tree %p wr_tree %p\n", 571 TAILQ_FIRST(&iocom->msgq), 572 RB_ROOT(&iocom->staterd_tree), 573 RB_ROOT(&iocom->statewr_tree)); 574 lksleep(iocom, &iocom->msglk, 0, "clstrtk", hz * 10); 575 } 576 } 577 578 /* 579 * Exit handling is done by the write thread. 580 */ 581 iocom->flags |= KDMSG_IOCOMF_EXITNOACC; 582 lockmgr(&iocom->msglk, LK_RELEASE); 583 584 /* 585 * The state trees had better be empty now 586 */ 587 KKASSERT(RB_EMPTY(&iocom->staterd_tree)); 588 KKASSERT(RB_EMPTY(&iocom->statewr_tree)); 589 KKASSERT(iocom->conn_state == NULL); 590 591 if (iocom->exit_func) { 592 /* 593 * iocom is invalid after we call the exit function. 594 */ 595 iocom->msgwr_td = NULL; 596 iocom->exit_func(iocom); 597 } else { 598 /* 599 * iocom can be ripped out from under us once msgwr_td is 600 * set to NULL. The wakeup is safe. 601 */ 602 iocom->msgwr_td = NULL; 603 wakeup(iocom); 604 } 605 lwkt_exit(); 606 } 607 608 /* 609 * This cleans out the pending transmit message queue, adjusting any 610 * persistent states properly in the process. 611 * 612 * Called with iocom locked. 613 */ 614 void 615 kdmsg_drain_msgq(kdmsg_iocom_t *iocom) 616 { 617 kdmsg_msg_t *msg; 618 619 /* 620 * Clean out our pending transmit queue, executing the 621 * appropriate state adjustments. If this tries to open 622 * any new outgoing transactions we have to loop up and 623 * clean them out. 624 */ 625 while ((msg = TAILQ_FIRST(&iocom->msgq)) != NULL) { 626 TAILQ_REMOVE(&iocom->msgq, msg, qentry); 627 if (kdmsg_state_msgtx(msg)) 628 kdmsg_msg_free(msg); 629 else 630 kdmsg_state_cleanuptx(msg); 631 } 632 } 633 634 /* 635 * Do all processing required to handle a freshly received message 636 * after its low level header has been validated. 637 * 638 * iocom is not locked. 639 */ 640 static 641 int 642 kdmsg_msg_receive_handling(kdmsg_msg_t *msg) 643 { 644 kdmsg_iocom_t *iocom = msg->state->iocom; 645 int error; 646 647 /* 648 * State machine tracking, state assignment for msg, 649 * returns error and discard status. Errors are fatal 650 * to the connection except for EALREADY which forces 651 * a discard without execution. 652 */ 653 error = kdmsg_state_msgrx(msg); 654 if (msg->state->flags & KDMSG_STATE_ABORTING) { 655 kdio_printf(iocom, 5, 656 "kdmsg_state_abort(b): state %p rxcmd=%08x " 657 "txcmd=%08x msgrx error %d\n", 658 msg->state, msg->state->rxcmd, 659 msg->state->txcmd, error); 660 } 661 if (error) { 662 /* 663 * Raw protocol or connection error 664 */ 665 if (msg->state->flags & KDMSG_STATE_ABORTING) 666 kdio_printf(iocom, 5, 667 "X1 state %p error %d\n", 668 msg->state, error); 669 kdmsg_msg_free(msg); 670 if (error == EALREADY) 671 error = 0; 672 } else if (msg->state && msg->state->func) { 673 /* 674 * Message related to state which already has a 675 * handling function installed for it. 676 */ 677 if (msg->state->flags & KDMSG_STATE_ABORTING) 678 kdio_printf(iocom, 5, 679 "X2 state %p func %p\n", 680 msg->state, msg->state->func); 681 error = msg->state->func(msg->state, msg); 682 kdmsg_state_cleanuprx(msg); 683 } else if (iocom->flags & KDMSG_IOCOMF_AUTOANY) { 684 if (msg->state->flags & KDMSG_STATE_ABORTING) 685 kdio_printf(iocom, 5, 686 "X3 state %p\n", msg->state); 687 error = kdmsg_autorxmsg(msg); 688 kdmsg_state_cleanuprx(msg); 689 } else { 690 if (msg->state->flags & KDMSG_STATE_ABORTING) 691 kdio_printf(iocom, 5, 692 "X4 state %p\n", msg->state); 693 error = iocom->rcvmsg(msg); 694 kdmsg_state_cleanuprx(msg); 695 } 696 return error; 697 } 698 699 /* 700 * Process state tracking for a message after reception and dequeueing, 701 * prior to execution of the state callback. The state is updated and 702 * will be removed from the RBTREE if completely closed, but the state->parent 703 * and subq linkage is not cleaned up until after the callback (see 704 * cleanuprx()). 705 * 706 * msglk is not held. 707 * 708 * NOTE: A message transaction can consist of several messages in either 709 * direction. 710 * 711 * NOTE: The msgid is unique to the initiator, not necessarily unique for 712 * us or for any relay or for the return direction for that matter. 713 * That is, two sides sending a new message can use the same msgid 714 * without colliding. 715 * 716 * -- 717 * 718 * ABORT sequences work by setting the ABORT flag along with normal message 719 * state. However, ABORTs can also be sent on half-closed messages, that is 720 * even if the command or reply side has already sent a DELETE, as long as 721 * the message has not been fully closed it can still send an ABORT+DELETE 722 * to terminate the half-closed message state. 723 * 724 * Since ABORT+DELETEs can race we silently discard ABORT's for message 725 * state which has already been fully closed. REPLY+ABORT+DELETEs can 726 * also race, and in this situation the other side might have already 727 * initiated a new unrelated command with the same message id. Since 728 * the abort has not set the CREATE flag the situation can be detected 729 * and the message will also be discarded. 730 * 731 * Non-blocking requests can be initiated with ABORT+CREATE[+DELETE]. 732 * The ABORT request is essentially integrated into the command instead 733 * of being sent later on. In this situation the command implementation 734 * detects that CREATE and ABORT are both set (vs ABORT alone) and can 735 * special-case non-blocking operation for the command. 736 * 737 * NOTE! Messages with ABORT set without CREATE or DELETE are considered 738 * to be mid-stream aborts for command/reply sequences. ABORTs on 739 * one-way messages are not supported. 740 * 741 * NOTE! If a command sequence does not support aborts the ABORT flag is 742 * simply ignored. 743 * 744 * -- 745 * 746 * One-off messages (no reply expected) are sent with neither CREATE or DELETE 747 * set. One-off messages cannot be aborted and typically aren't processed 748 * by these routines. The REPLY bit can be used to distinguish whether a 749 * one-off message is a command or reply. For example, one-off replies 750 * will typically just contain status updates. 751 */ 752 static 753 int 754 kdmsg_state_msgrx(kdmsg_msg_t *msg) 755 { 756 kdmsg_iocom_t *iocom = msg->state->iocom; 757 kdmsg_state_t *state; 758 kdmsg_state_t *pstate; 759 kdmsg_state_t sdummy; 760 int error; 761 762 /* 763 * Make sure a state structure is ready to go in case we need a new 764 * one. This is the only routine which uses freerd_state so no 765 * races are possible. 766 */ 767 if ((state = iocom->freerd_state) == NULL) { 768 state = kmalloc(sizeof(*state), iocom->mmsg, M_WAITOK | M_ZERO); 769 state->flags = KDMSG_STATE_DYNAMIC; 770 state->iocom = iocom; 771 state->refs = 1; 772 TAILQ_INIT(&state->subq); 773 iocom->freerd_state = state; 774 } 775 state = NULL; /* safety */ 776 777 /* 778 * Lock RB tree and locate existing persistent state, if any. 779 * 780 * If received msg is a command state is on staterd_tree. 781 * If received msg is a reply state is on statewr_tree. 782 */ 783 lockmgr(&iocom->msglk, LK_EXCLUSIVE); 784 785 again: 786 if (msg->state == &iocom->state0) { 787 sdummy.msgid = msg->any.head.msgid; 788 sdummy.iocom = iocom; 789 if (msg->any.head.cmd & DMSGF_REVTRANS) { 790 state = RB_FIND(kdmsg_state_tree, &iocom->statewr_tree, 791 &sdummy); 792 } else { 793 state = RB_FIND(kdmsg_state_tree, &iocom->staterd_tree, 794 &sdummy); 795 } 796 797 /* 798 * Set message state unconditionally. If this is a CREATE 799 * message this state will become the parent state and new 800 * state will be allocated for the message state. 801 */ 802 if (state == NULL) 803 state = &iocom->state0; 804 if (state->flags & KDMSG_STATE_INTERLOCK) { 805 state->flags |= KDMSG_STATE_SIGNAL; 806 lksleep(state, &iocom->msglk, 0, "dmrace", hz); 807 goto again; 808 } 809 kdmsg_state_hold(state); 810 kdmsg_state_drop(msg->state); /* iocom->state0 */ 811 msg->state = state; 812 } else { 813 state = msg->state; 814 } 815 816 /* 817 * Short-cut one-off or mid-stream messages. 818 */ 819 if ((msg->any.head.cmd & (DMSGF_CREATE | DMSGF_DELETE | 820 DMSGF_ABORT)) == 0) { 821 error = 0; 822 goto done; 823 } 824 825 /* 826 * Switch on CREATE, DELETE, REPLY, and also handle ABORT from 827 * inside the case statements. 828 */ 829 switch(msg->any.head.cmd & (DMSGF_CREATE|DMSGF_DELETE|DMSGF_REPLY)) { 830 case DMSGF_CREATE: 831 case DMSGF_CREATE | DMSGF_DELETE: 832 /* 833 * New persistant command received. 834 */ 835 if (state != &iocom->state0) { 836 kdio_printf(iocom, 1, "%s\n", 837 "duplicate transaction"); 838 error = EINVAL; 839 break; 840 } 841 842 /* 843 * Lookup the circuit. The circuit is an open transaction. 844 * the REVCIRC bit in the message tells us which side 845 * initiated the transaction representing the circuit. 846 */ 847 if (msg->any.head.circuit) { 848 sdummy.msgid = msg->any.head.circuit; 849 850 if (msg->any.head.cmd & DMSGF_REVCIRC) { 851 pstate = RB_FIND(kdmsg_state_tree, 852 &iocom->statewr_tree, 853 &sdummy); 854 } else { 855 pstate = RB_FIND(kdmsg_state_tree, 856 &iocom->staterd_tree, 857 &sdummy); 858 } 859 if (pstate == NULL) { 860 kdio_printf(iocom, 1, "%s\n", 861 "missing parent in " 862 "stacked trans"); 863 error = EINVAL; 864 break; 865 } 866 } else { 867 pstate = &iocom->state0; 868 } 869 870 /* 871 * Allocate new state. 872 * 873 * msg->state becomes the owner of the ref we inherit from 874 * freerd_stae. 875 */ 876 kdmsg_state_drop(state); 877 state = iocom->freerd_state; 878 iocom->freerd_state = NULL; 879 880 msg->state = state; /* inherits freerd ref */ 881 state->parent = pstate; 882 KKASSERT(state->iocom == iocom); 883 state->flags |= KDMSG_STATE_RBINSERTED | 884 KDMSG_STATE_SUBINSERTED | 885 KDMSG_STATE_OPPOSITE; 886 if (TAILQ_EMPTY(&pstate->subq)) 887 kdmsg_state_hold(pstate);/* states on pstate->subq */ 888 kdmsg_state_hold(state); /* state on pstate->subq */ 889 kdmsg_state_hold(state); /* state on rbtree */ 890 state->icmd = msg->any.head.cmd & DMSGF_BASECMDMASK; 891 state->rxcmd = msg->any.head.cmd & ~DMSGF_DELETE; 892 state->txcmd = DMSGF_REPLY; 893 state->msgid = msg->any.head.msgid; 894 state->flags &= ~KDMSG_STATE_NEW; 895 RB_INSERT(kdmsg_state_tree, &iocom->staterd_tree, state); 896 TAILQ_INSERT_TAIL(&pstate->subq, state, entry); 897 error = 0; 898 break; 899 case DMSGF_DELETE: 900 /* 901 * Persistent state is expected but might not exist if an 902 * ABORT+DELETE races the close. 903 */ 904 if (state == &iocom->state0) { 905 if (msg->any.head.cmd & DMSGF_ABORT) { 906 kdio_printf(iocom, 1, "%s\n", 907 "msgrx: " 908 "state already A"); 909 error = EALREADY; 910 } else { 911 kdio_printf(iocom, 1, "%s\n", 912 "msgrx: no state for DELETE"); 913 error = EINVAL; 914 } 915 break; 916 } 917 918 /* 919 * Handle another ABORT+DELETE case if the msgid has already 920 * been reused. 921 */ 922 if ((state->rxcmd & DMSGF_CREATE) == 0) { 923 if (msg->any.head.cmd & DMSGF_ABORT) { 924 kdio_printf(iocom, 1, "%s\n", 925 "msgrx: state already B"); 926 error = EALREADY; 927 } else { 928 kdio_printf(iocom, 1, "%s\n", 929 "msgrx: state reused for DELETE"); 930 error = EINVAL; 931 } 932 break; 933 } 934 error = 0; 935 break; 936 default: 937 /* 938 * Check for mid-stream ABORT command received, otherwise 939 * allow. 940 */ 941 if (msg->any.head.cmd & DMSGF_ABORT) { 942 if (state == &iocom->state0 || 943 (state->rxcmd & DMSGF_CREATE) == 0) { 944 error = EALREADY; 945 break; 946 } 947 } 948 error = 0; 949 break; 950 case DMSGF_REPLY | DMSGF_CREATE: 951 case DMSGF_REPLY | DMSGF_CREATE | DMSGF_DELETE: 952 /* 953 * When receiving a reply with CREATE set the original 954 * persistent state message should already exist. 955 */ 956 if (state == &iocom->state0) { 957 kdio_printf(iocom, 1, 958 "msgrx: no state match for " 959 "REPLY cmd=%08x msgid=%016jx\n", 960 msg->any.head.cmd, 961 (intmax_t)msg->any.head.msgid); 962 error = EINVAL; 963 break; 964 } 965 state->rxcmd = msg->any.head.cmd & ~DMSGF_DELETE; 966 error = 0; 967 break; 968 case DMSGF_REPLY | DMSGF_DELETE: 969 /* 970 * Received REPLY+ABORT+DELETE in case where msgid has 971 * already been fully closed, ignore the message. 972 */ 973 if (state == &iocom->state0) { 974 if (msg->any.head.cmd & DMSGF_ABORT) { 975 error = EALREADY; 976 } else { 977 kdio_printf(iocom, 1, "%s\n", 978 "msgrx: no state match " 979 "for REPLY|DELETE"); 980 error = EINVAL; 981 } 982 break; 983 } 984 985 /* 986 * Received REPLY+ABORT+DELETE in case where msgid has 987 * already been reused for an unrelated message, 988 * ignore the message. 989 */ 990 if ((state->rxcmd & DMSGF_CREATE) == 0) { 991 if (msg->any.head.cmd & DMSGF_ABORT) { 992 error = EALREADY; 993 } else { 994 kdio_printf(iocom, 1, "%s\n", 995 "msgrx: state reused " 996 "for REPLY|DELETE"); 997 error = EINVAL; 998 } 999 break; 1000 } 1001 error = 0; 1002 break; 1003 case DMSGF_REPLY: 1004 /* 1005 * Check for mid-stream ABORT reply received to sent command. 1006 */ 1007 if (msg->any.head.cmd & DMSGF_ABORT) { 1008 if (state == &iocom->state0 || 1009 (state->rxcmd & DMSGF_CREATE) == 0) { 1010 error = EALREADY; 1011 break; 1012 } 1013 } 1014 error = 0; 1015 break; 1016 } 1017 1018 /* 1019 * Calculate the easy-switch() transactional command. Represents 1020 * the outer-transaction command for any transaction-create or 1021 * transaction-delete, and the inner message command for any 1022 * non-transaction or inside-transaction command. tcmd will be 1023 * set to 0 if the message state is illegal. 1024 * 1025 * The two can be told apart because outer-transaction commands 1026 * always have a DMSGF_CREATE and/or DMSGF_DELETE flag. 1027 */ 1028 done: 1029 if (msg->any.head.cmd & (DMSGF_CREATE | DMSGF_DELETE)) { 1030 if (state != &iocom->state0) { 1031 msg->tcmd = (msg->state->icmd & DMSGF_BASECMDMASK) | 1032 (msg->any.head.cmd & (DMSGF_CREATE | 1033 DMSGF_DELETE | 1034 DMSGF_REPLY)); 1035 } else { 1036 msg->tcmd = 0; 1037 } 1038 } else { 1039 msg->tcmd = msg->any.head.cmd & DMSGF_CMDSWMASK; 1040 } 1041 1042 /* 1043 * Adjust the state for DELETE handling now, before making the 1044 * callback so we are atomic with other state updates. 1045 * 1046 * Subq/parent linkages are cleaned up after the callback. 1047 * If an error occurred the message is ignored and state is not 1048 * updated. 1049 */ 1050 if ((state = msg->state) == NULL || error != 0) { 1051 kdio_printf(iocom, 1, 1052 "msgrx: state=%p error %d\n", 1053 state, error); 1054 } else if (msg->any.head.cmd & DMSGF_DELETE) { 1055 KKASSERT((state->rxcmd & DMSGF_DELETE) == 0); 1056 state->rxcmd |= DMSGF_DELETE; 1057 if (state->txcmd & DMSGF_DELETE) { 1058 KKASSERT(state->flags & KDMSG_STATE_RBINSERTED); 1059 if (state->rxcmd & DMSGF_REPLY) { 1060 KKASSERT(msg->any.head.cmd & 1061 DMSGF_REPLY); 1062 RB_REMOVE(kdmsg_state_tree, 1063 &iocom->statewr_tree, state); 1064 } else { 1065 KKASSERT((msg->any.head.cmd & 1066 DMSGF_REPLY) == 0); 1067 RB_REMOVE(kdmsg_state_tree, 1068 &iocom->staterd_tree, state); 1069 } 1070 state->flags &= ~KDMSG_STATE_RBINSERTED; 1071 kdmsg_state_drop(state); /* state on rbtree */ 1072 } 1073 } 1074 lockmgr(&iocom->msglk, LK_RELEASE); 1075 1076 return (error); 1077 } 1078 1079 /* 1080 * Called instead of iocom->rcvmsg() if any of the AUTO flags are set. 1081 * This routine must call iocom->rcvmsg() for anything not automatically 1082 * handled. 1083 */ 1084 static int 1085 kdmsg_autorxmsg(kdmsg_msg_t *msg) 1086 { 1087 kdmsg_iocom_t *iocom = msg->state->iocom; 1088 kdmsg_msg_t *rep; 1089 int error = 0; 1090 uint32_t cmd; 1091 1092 /* 1093 * Main switch processes transaction create/delete sequences only. 1094 * Use icmd (DELETEs use DMSG_LNK_ERROR 1095 * 1096 * NOTE: If processing in-transaction messages you generally want 1097 * an inner switch on msg->any.head.cmd. 1098 */ 1099 if (msg->state) { 1100 cmd = (msg->state->icmd & DMSGF_BASECMDMASK) | 1101 (msg->any.head.cmd & (DMSGF_CREATE | 1102 DMSGF_DELETE | 1103 DMSGF_REPLY)); 1104 } else { 1105 cmd = 0; 1106 } 1107 1108 switch(cmd) { 1109 case DMSG_LNK_PING: 1110 /* 1111 * Received ping, send reply 1112 */ 1113 rep = kdmsg_msg_alloc(msg->state, DMSG_LNK_PING | DMSGF_REPLY, 1114 NULL, NULL); 1115 kdmsg_msg_write(rep); 1116 break; 1117 case DMSG_LNK_PING | DMSGF_REPLY: 1118 /* ignore replies */ 1119 break; 1120 case DMSG_LNK_CONN | DMSGF_CREATE: 1121 case DMSG_LNK_CONN | DMSGF_CREATE | DMSGF_DELETE: 1122 /* 1123 * Received LNK_CONN transaction. Transmit response and 1124 * leave transaction open, which allows the other end to 1125 * start to the SPAN protocol. 1126 * 1127 * Handle shim after acknowledging the CONN. 1128 */ 1129 if ((msg->any.head.cmd & DMSGF_DELETE) == 0) { 1130 if (iocom->flags & KDMSG_IOCOMF_AUTOCONN) { 1131 kdmsg_msg_result(msg, 0); 1132 if (iocom->auto_callback) 1133 iocom->auto_callback(msg); 1134 } else { 1135 error = iocom->rcvmsg(msg); 1136 } 1137 break; 1138 } 1139 /* fall through */ 1140 case DMSG_LNK_CONN | DMSGF_DELETE: 1141 /* 1142 * This message is usually simulated after a link is lost 1143 * to clean up the transaction. 1144 */ 1145 if (iocom->flags & KDMSG_IOCOMF_AUTOCONN) { 1146 if (iocom->auto_callback) 1147 iocom->auto_callback(msg); 1148 kdmsg_msg_reply(msg, 0); 1149 } else { 1150 error = iocom->rcvmsg(msg); 1151 } 1152 break; 1153 case DMSG_LNK_SPAN | DMSGF_CREATE: 1154 case DMSG_LNK_SPAN | DMSGF_CREATE | DMSGF_DELETE: 1155 /* 1156 * Received LNK_SPAN transaction. We do not have to respond 1157 * (except on termination), but we must leave the transaction 1158 * open. 1159 * 1160 * Handle shim after acknowledging the SPAN. 1161 */ 1162 if (iocom->flags & KDMSG_IOCOMF_AUTORXSPAN) { 1163 if ((msg->any.head.cmd & DMSGF_DELETE) == 0) { 1164 if (iocom->auto_callback) 1165 iocom->auto_callback(msg); 1166 break; 1167 } 1168 /* fall through */ 1169 } else { 1170 error = iocom->rcvmsg(msg); 1171 break; 1172 } 1173 /* fall through */ 1174 case DMSG_LNK_SPAN | DMSGF_DELETE: 1175 /* 1176 * Process shims (auto_callback) before cleaning up the 1177 * circuit structure and closing the transactions. Device 1178 * driver should ensure that the circuit is not used after 1179 * the auto_callback() returns. 1180 * 1181 * Handle shim before closing the SPAN transaction. 1182 */ 1183 if (iocom->flags & KDMSG_IOCOMF_AUTORXSPAN) { 1184 if (iocom->auto_callback) 1185 iocom->auto_callback(msg); 1186 kdmsg_msg_reply(msg, 0); 1187 } else { 1188 error = iocom->rcvmsg(msg); 1189 } 1190 break; 1191 default: 1192 /* 1193 * Anything unhandled goes into rcvmsg. 1194 * 1195 * NOTE: Replies to link-level messages initiated by our side 1196 * are handled by the state callback, they are NOT 1197 * handled here. 1198 */ 1199 error = iocom->rcvmsg(msg); 1200 break; 1201 } 1202 return (error); 1203 } 1204 1205 /* 1206 * Post-receive-handling message and state cleanup. This routine is called 1207 * after the state function handling/callback to properly dispose of the 1208 * message and unlink the state's parent/subq linkage if the state is 1209 * completely closed. 1210 * 1211 * msglk is not held. 1212 */ 1213 static 1214 void 1215 kdmsg_state_cleanuprx(kdmsg_msg_t *msg) 1216 { 1217 kdmsg_state_t *state = msg->state; 1218 kdmsg_iocom_t *iocom = state->iocom; 1219 1220 lockmgr(&iocom->msglk, LK_EXCLUSIVE); 1221 if (state != &iocom->state0) { 1222 /* 1223 * When terminating a transaction (in either direction), all 1224 * sub-states are aborted. 1225 */ 1226 if ((msg->any.head.cmd & DMSGF_DELETE) && 1227 TAILQ_FIRST(&msg->state->subq)) { 1228 kdio_printf(iocom, 2, 1229 "simulate failure for substates of " 1230 "state %p cmd %08x/%08x\n", 1231 msg->state, 1232 msg->state->rxcmd, 1233 msg->state->txcmd); 1234 kdmsg_simulate_failure(msg->state, 1235 0, DMSG_ERR_LOSTLINK); 1236 } 1237 1238 /* 1239 * Once the state is fully closed we can (try to) remove it 1240 * from the subq topology. 1241 */ 1242 if ((state->flags & KDMSG_STATE_SUBINSERTED) && 1243 (state->rxcmd & DMSGF_DELETE) && 1244 (state->txcmd & DMSGF_DELETE)) { 1245 /* 1246 * Remove parent linkage if state is completely closed. 1247 */ 1248 kdmsg_subq_delete(state); 1249 } 1250 } 1251 kdmsg_msg_free(msg); 1252 1253 lockmgr(&iocom->msglk, LK_RELEASE); 1254 } 1255 1256 /* 1257 * Remove state from its parent's subq. This can wind up recursively 1258 * dropping the parent upward. 1259 * 1260 * NOTE: Once we drop the parent, our pstate pointer may become invalid. 1261 */ 1262 static 1263 void 1264 kdmsg_subq_delete(kdmsg_state_t *state) 1265 { 1266 kdmsg_state_t *pstate; 1267 1268 if (state->flags & KDMSG_STATE_SUBINSERTED) { 1269 pstate = state->parent; 1270 KKASSERT(pstate); 1271 if (pstate->scan == state) 1272 pstate->scan = NULL; 1273 TAILQ_REMOVE(&pstate->subq, state, entry); 1274 state->flags &= ~KDMSG_STATE_SUBINSERTED; 1275 state->parent = NULL; 1276 if (TAILQ_EMPTY(&pstate->subq)) { 1277 kdmsg_state_drop(pstate);/* pstate->subq */ 1278 } 1279 pstate = NULL; /* safety */ 1280 kdmsg_state_drop(state); /* pstate->subq */ 1281 } else { 1282 KKASSERT(state->parent == NULL); 1283 } 1284 } 1285 1286 /* 1287 * Simulate receiving a message which terminates an active transaction 1288 * state. Our simulated received message must set DELETE and may also 1289 * have to set CREATE. It must also ensure that all fields are set such 1290 * that the receive handling code can find the state (kdmsg_state_msgrx()) 1291 * or an endless loop will ensue. 1292 * 1293 * This is used when the other end of the link is dead so the device driver 1294 * gets a completed transaction for all pending states. 1295 * 1296 * Called with iocom locked. 1297 */ 1298 static 1299 void 1300 kdmsg_simulate_failure(kdmsg_state_t *state, int meto, int error) 1301 { 1302 kdmsg_state_t *substate; 1303 1304 kdmsg_state_hold(state); /* aborting */ 1305 1306 /* 1307 * Abort parent state first. Parent will not actually disappear 1308 * until children are gone. Device drivers must handle the situation. 1309 * The advantage of this is that device drivers can flag the situation 1310 * as an interlock against new operations on dying states. And since 1311 * device operations are often asynchronous anyway, this sequence of 1312 * events works out better. 1313 */ 1314 if (meto) 1315 kdmsg_state_abort(state); 1316 1317 /* 1318 * Recurse through any children. 1319 */ 1320 again: 1321 TAILQ_FOREACH(substate, &state->subq, entry) { 1322 if (substate->flags & KDMSG_STATE_ABORTING) 1323 continue; 1324 state->scan = substate; 1325 kdmsg_simulate_failure(substate, 1, error); 1326 if (state->scan != substate) 1327 goto again; 1328 } 1329 kdmsg_state_drop(state); /* aborting */ 1330 } 1331 1332 static 1333 void 1334 kdmsg_state_abort(kdmsg_state_t *state) 1335 { 1336 kdmsg_msg_t *msg; 1337 1338 /* 1339 * Set ABORTING and DYING, return if already set. If the state was 1340 * just allocated we defer the abort operation until the related 1341 * message is processed. 1342 */ 1343 KKASSERT((state->flags & KDMSG_STATE_ABORTING) == 0); 1344 if (state->flags & KDMSG_STATE_ABORTING) 1345 return; 1346 state->flags |= KDMSG_STATE_ABORTING; 1347 kdmsg_state_dying(state); 1348 if (state->flags & KDMSG_STATE_NEW) { 1349 kdio_printf(iocom, 5, 1350 "kdmsg_state_abort(0): state %p rxcmd %08x " 1351 "txcmd %08x flags %08x - in NEW state\n", 1352 state, state->rxcmd, 1353 state->txcmd, state->flags); 1354 return; 1355 } 1356 1357 /* 1358 * NOTE: The DELETE flag might already be set due to an early 1359 * termination. 1360 * 1361 * NOTE: Args to kdmsg_msg_alloc() to avoid dynamic state allocation. 1362 * 1363 * NOTE: We are simulating a received message using our state 1364 * (vs a message generated by the other side using its state), 1365 * so we must invert DMSGF_REVTRANS and DMSGF_REVCIRC. 1366 */ 1367 kdio_printf(iocom, 5, 1368 "kdmsg_state_abort(1): state %p rxcmd %08x txcmd %08x\n", 1369 state, state->rxcmd, state->txcmd); 1370 if ((state->rxcmd & DMSGF_DELETE) == 0) { 1371 msg = kdmsg_msg_alloc(state, DMSG_LNK_ERROR, NULL, NULL); 1372 if ((state->rxcmd & DMSGF_CREATE) == 0) 1373 msg->any.head.cmd |= DMSGF_CREATE; 1374 msg->any.head.cmd |= DMSGF_DELETE | 1375 (state->rxcmd & DMSGF_REPLY); 1376 msg->any.head.cmd ^= (DMSGF_REVTRANS | DMSGF_REVCIRC); 1377 msg->any.head.error = DMSG_ERR_LOSTLINK; 1378 kdio_printf(iocom, 5, 1379 "kdmsg_state_abort(a): state %p msgcmd %08x\n", 1380 state, msg->any.head.cmd); 1381 /* circuit not initialized */ 1382 lockmgr(&state->iocom->msglk, LK_RELEASE); 1383 kdmsg_msg_receive_handling(msg); 1384 lockmgr(&state->iocom->msglk, LK_EXCLUSIVE); 1385 msg = NULL; 1386 } 1387 kdio_printf(iocom, 5, 1388 "kdmsg_state_abort(2): state %p rxcmd %08x txcmd %08x\n", 1389 state, state->rxcmd, state->txcmd); 1390 } 1391 1392 /* 1393 * Recursively sets KDMSG_STATE_DYING on state and all sub-states, preventing 1394 * the transmission of any new messages on these states. This is done 1395 * atomically when parent state is terminating, whereas setting ABORTING is 1396 * not atomic and can leak races. 1397 */ 1398 static 1399 void 1400 kdmsg_state_dying(kdmsg_state_t *state) 1401 { 1402 kdmsg_state_t *scan; 1403 1404 if ((state->flags & KDMSG_STATE_DYING) == 0) { 1405 state->flags |= KDMSG_STATE_DYING; 1406 TAILQ_FOREACH(scan, &state->subq, entry) 1407 kdmsg_state_dying(scan); 1408 } 1409 } 1410 1411 /* 1412 * Process state tracking for a message prior to transmission. 1413 * 1414 * Called with msglk held and the msg dequeued. Returns non-zero if 1415 * the message is bad and should be deleted by the caller. 1416 * 1417 * One-off messages are usually with dummy state and msg->state may be NULL 1418 * in this situation. 1419 * 1420 * New transactions (when CREATE is set) will insert the state. 1421 * 1422 * May request that caller discard the message by setting *discardp to 1. 1423 * A NULL state may be returned in this case. 1424 */ 1425 static 1426 int 1427 kdmsg_state_msgtx(kdmsg_msg_t *msg) 1428 { 1429 kdmsg_iocom_t *iocom = msg->state->iocom; 1430 kdmsg_state_t *state; 1431 int error; 1432 1433 /* 1434 * Make sure a state structure is ready to go in case we need a new 1435 * one. This is the only routine which uses freewr_state so no 1436 * races are possible. 1437 */ 1438 if ((state = iocom->freewr_state) == NULL) { 1439 state = kmalloc(sizeof(*state), iocom->mmsg, M_WAITOK | M_ZERO); 1440 state->flags = KDMSG_STATE_DYNAMIC; 1441 state->iocom = iocom; 1442 state->refs = 1; 1443 TAILQ_INIT(&state->subq); 1444 iocom->freewr_state = state; 1445 } 1446 1447 /* 1448 * Lock RB tree. If persistent state is present it will have already 1449 * been assigned to msg. 1450 */ 1451 state = msg->state; 1452 1453 /* 1454 * Short-cut one-off or mid-stream messages (state may be NULL). 1455 */ 1456 if ((msg->any.head.cmd & (DMSGF_CREATE | DMSGF_DELETE | 1457 DMSGF_ABORT)) == 0) { 1458 return(0); 1459 } 1460 1461 1462 /* 1463 * Switch on CREATE, DELETE, REPLY, and also handle ABORT from 1464 * inside the case statements. 1465 */ 1466 switch(msg->any.head.cmd & (DMSGF_CREATE | DMSGF_DELETE | 1467 DMSGF_REPLY)) { 1468 case DMSGF_CREATE: 1469 case DMSGF_CREATE | DMSGF_DELETE: 1470 /* 1471 * Insert the new persistent message state and mark 1472 * half-closed if DELETE is set. Since this is a new 1473 * message it isn't possible to transition into the fully 1474 * closed state here. 1475 * 1476 * XXX state must be assigned and inserted by 1477 * kdmsg_msg_write(). txcmd is assigned by us 1478 * on-transmit. 1479 */ 1480 KKASSERT(state != NULL); 1481 state->icmd = msg->any.head.cmd & DMSGF_BASECMDMASK; 1482 state->txcmd = msg->any.head.cmd & ~DMSGF_DELETE; 1483 state->rxcmd = DMSGF_REPLY; 1484 state->flags &= ~KDMSG_STATE_NEW; 1485 error = 0; 1486 break; 1487 case DMSGF_DELETE: 1488 /* 1489 * Sent ABORT+DELETE in case where msgid has already 1490 * been fully closed, ignore the message. 1491 */ 1492 if (state == &iocom->state0) { 1493 if (msg->any.head.cmd & DMSGF_ABORT) { 1494 error = EALREADY; 1495 } else { 1496 kdio_printf(iocom, 1, 1497 "msgtx: no state match " 1498 "for DELETE cmd=%08x msgid=%016jx\n", 1499 msg->any.head.cmd, 1500 (intmax_t)msg->any.head.msgid); 1501 error = EINVAL; 1502 } 1503 break; 1504 } 1505 1506 /* 1507 * Sent ABORT+DELETE in case where msgid has 1508 * already been reused for an unrelated message, 1509 * ignore the message. 1510 */ 1511 if ((state->txcmd & DMSGF_CREATE) == 0) { 1512 if (msg->any.head.cmd & DMSGF_ABORT) { 1513 error = EALREADY; 1514 } else { 1515 kdio_printf(iocom, 1, "%s\n", 1516 "msgtx: state reused " 1517 "for DELETE"); 1518 error = EINVAL; 1519 } 1520 break; 1521 } 1522 error = 0; 1523 break; 1524 default: 1525 /* 1526 * Check for mid-stream ABORT command sent 1527 */ 1528 if (msg->any.head.cmd & DMSGF_ABORT) { 1529 if (state == &state->iocom->state0 || 1530 (state->txcmd & DMSGF_CREATE) == 0) { 1531 error = EALREADY; 1532 break; 1533 } 1534 } 1535 error = 0; 1536 break; 1537 case DMSGF_REPLY | DMSGF_CREATE: 1538 case DMSGF_REPLY | DMSGF_CREATE | DMSGF_DELETE: 1539 /* 1540 * When transmitting a reply with CREATE set the original 1541 * persistent state message should already exist. 1542 */ 1543 if (state == &state->iocom->state0) { 1544 kdio_printf(iocom, 1, "%s\n", 1545 "msgtx: no state match " 1546 "for REPLY | CREATE"); 1547 error = EINVAL; 1548 break; 1549 } 1550 state->txcmd = msg->any.head.cmd & ~DMSGF_DELETE; 1551 error = 0; 1552 break; 1553 case DMSGF_REPLY | DMSGF_DELETE: 1554 /* 1555 * When transmitting a reply with DELETE set the original 1556 * persistent state message should already exist. 1557 * 1558 * This is very similar to the REPLY|CREATE|* case except 1559 * txcmd is already stored, so we just add the DELETE flag. 1560 * 1561 * Sent REPLY+ABORT+DELETE in case where msgid has 1562 * already been fully closed, ignore the message. 1563 */ 1564 if (state == &state->iocom->state0) { 1565 if (msg->any.head.cmd & DMSGF_ABORT) { 1566 error = EALREADY; 1567 } else { 1568 kdio_printf(iocom, 1, "%s\n", 1569 "msgtx: no state match " 1570 "for REPLY | DELETE"); 1571 error = EINVAL; 1572 } 1573 break; 1574 } 1575 1576 /* 1577 * Sent REPLY+ABORT+DELETE in case where msgid has already 1578 * been reused for an unrelated message, ignore the message. 1579 */ 1580 if ((state->txcmd & DMSGF_CREATE) == 0) { 1581 if (msg->any.head.cmd & DMSGF_ABORT) { 1582 error = EALREADY; 1583 } else { 1584 kdio_printf(iocom, 1, "%s\n", 1585 "msgtx: state reused " 1586 "for REPLY | DELETE"); 1587 error = EINVAL; 1588 } 1589 break; 1590 } 1591 error = 0; 1592 break; 1593 case DMSGF_REPLY: 1594 /* 1595 * Check for mid-stream ABORT reply sent. 1596 * 1597 * One-off REPLY messages are allowed for e.g. status updates. 1598 */ 1599 if (msg->any.head.cmd & DMSGF_ABORT) { 1600 if (state == &state->iocom->state0 || 1601 (state->txcmd & DMSGF_CREATE) == 0) { 1602 error = EALREADY; 1603 break; 1604 } 1605 } 1606 error = 0; 1607 break; 1608 } 1609 1610 /* 1611 * Set interlock (XXX hack) in case the send side blocks and a 1612 * response is returned before kdmsg_state_cleanuptx() can be 1613 * run. 1614 */ 1615 if (state && error == 0) 1616 state->flags |= KDMSG_STATE_INTERLOCK; 1617 1618 return (error); 1619 } 1620 1621 /* 1622 * Called with iocom locked. 1623 */ 1624 static 1625 void 1626 kdmsg_state_cleanuptx(kdmsg_msg_t *msg) 1627 { 1628 kdmsg_iocom_t *iocom = msg->state->iocom; 1629 kdmsg_state_t *state; 1630 1631 if ((state = msg->state) == NULL) { 1632 kdmsg_msg_free(msg); 1633 return; 1634 } 1635 1636 /* 1637 * Clear interlock (XXX hack) in case the send side blocks and a 1638 * response is returned in the other thread before 1639 * kdmsg_state_cleanuptx() can be run. We maintain our hold on 1640 * iocom->msglk so we can do this before completing our task. 1641 */ 1642 if (state->flags & KDMSG_STATE_SIGNAL) { 1643 kdio_printf(iocom, 1, "state %p interlock!\n", state); 1644 wakeup(state); 1645 } 1646 state->flags &= ~(KDMSG_STATE_INTERLOCK | KDMSG_STATE_SIGNAL); 1647 kdmsg_state_hold(state); 1648 1649 if (msg->any.head.cmd & DMSGF_DELETE) { 1650 KKASSERT((state->txcmd & DMSGF_DELETE) == 0); 1651 state->txcmd |= DMSGF_DELETE; 1652 if (state->rxcmd & DMSGF_DELETE) { 1653 KKASSERT(state->flags & KDMSG_STATE_RBINSERTED); 1654 if (state->txcmd & DMSGF_REPLY) { 1655 KKASSERT(msg->any.head.cmd & 1656 DMSGF_REPLY); 1657 RB_REMOVE(kdmsg_state_tree, 1658 &iocom->staterd_tree, state); 1659 } else { 1660 KKASSERT((msg->any.head.cmd & 1661 DMSGF_REPLY) == 0); 1662 RB_REMOVE(kdmsg_state_tree, 1663 &iocom->statewr_tree, state); 1664 } 1665 state->flags &= ~KDMSG_STATE_RBINSERTED; 1666 1667 /* 1668 * The subq recursion is used for parent linking and 1669 * scanning the topology for aborts, we can only 1670 * remove leafs. The circuit is effectively dead now, 1671 * but topology won't be torn down until all of its 1672 * children have finished/aborted. 1673 * 1674 * This is particularly important for end-point 1675 * devices which might need to access private data 1676 * in parent states. Out of order disconnects can 1677 * occur if an end-point device is processing a 1678 * message transaction asynchronously because abort 1679 * requests are basically synchronous and it probably 1680 * isn't convenient (or possible) for the end-point 1681 * to abort an asynchronous operation. 1682 */ 1683 if (TAILQ_EMPTY(&state->subq)) 1684 kdmsg_subq_delete(state); 1685 kdmsg_msg_free(msg); 1686 kdmsg_state_drop(state); /* state on rbtree */ 1687 } else { 1688 kdmsg_msg_free(msg); 1689 } 1690 } else { 1691 kdmsg_msg_free(msg); 1692 } 1693 1694 /* 1695 * Deferred abort after transmission. 1696 */ 1697 if ((state->flags & (KDMSG_STATE_ABORTING | KDMSG_STATE_DYING)) && 1698 (state->rxcmd & DMSGF_DELETE) == 0) { 1699 kdio_printf(iocom, 5, 1700 "kdmsg_state_cleanuptx: state=%p " 1701 "executing deferred abort\n", 1702 state); 1703 state->flags &= ~KDMSG_STATE_ABORTING; 1704 kdmsg_state_abort(state); 1705 } 1706 kdmsg_state_drop(state); 1707 } 1708 1709 static 1710 void 1711 _kdmsg_state_hold(kdmsg_state_t *state KDMSG_DEBUG_ARGS) 1712 { 1713 atomic_add_int(&state->refs, 1); 1714 #if KDMSG_DEBUG 1715 kd_printf(4, "state %p +%d\t%s:%d\n", state, state->refs, file, line); 1716 #endif 1717 } 1718 1719 static 1720 void 1721 _kdmsg_state_drop(kdmsg_state_t *state KDMSG_DEBUG_ARGS) 1722 { 1723 KKASSERT(state->refs > 0); 1724 #if KDMSG_DEBUG 1725 kd_printf(4, "state %p -%d\t%s:%d\n", state, state->refs, file, line); 1726 #endif 1727 if (atomic_fetchadd_int(&state->refs, -1) == 1) 1728 kdmsg_state_free(state); 1729 } 1730 1731 static 1732 void 1733 kdmsg_state_free(kdmsg_state_t *state) 1734 { 1735 kdmsg_iocom_t *iocom = state->iocom; 1736 1737 KKASSERT((state->flags & KDMSG_STATE_RBINSERTED) == 0); 1738 KKASSERT((state->flags & KDMSG_STATE_SUBINSERTED) == 0); 1739 KKASSERT(TAILQ_EMPTY(&state->subq)); 1740 1741 if (state != &state->iocom->state0) 1742 kfree(state, iocom->mmsg); 1743 } 1744 1745 kdmsg_msg_t * 1746 kdmsg_msg_alloc(kdmsg_state_t *state, uint32_t cmd, 1747 int (*func)(kdmsg_state_t *, kdmsg_msg_t *), void *data) 1748 { 1749 kdmsg_iocom_t *iocom = state->iocom; 1750 kdmsg_state_t *pstate; 1751 kdmsg_msg_t *msg; 1752 size_t hbytes; 1753 1754 KKASSERT(iocom != NULL); 1755 hbytes = (cmd & DMSGF_SIZE) * DMSG_ALIGN; 1756 msg = kmalloc(offsetof(struct kdmsg_msg, any) + hbytes, 1757 iocom->mmsg, M_WAITOK | M_ZERO); 1758 msg->hdr_size = hbytes; 1759 1760 if ((cmd & (DMSGF_CREATE | DMSGF_REPLY)) == DMSGF_CREATE) { 1761 /* 1762 * New transaction, requires tracking state and a unique 1763 * msgid to be allocated. 1764 * 1765 * It is possible to race a circuit failure, inherit the 1766 * parent's STATE_DYING flag to trigger an abort sequence 1767 * in the transmit path. By not inheriting ABORTING the 1768 * abort sequence can recurse. 1769 * 1770 * NOTE: The transactions has not yet been initiated so we 1771 * cannot set DMSGF_CREATE/DELETE bits in txcmd or rxcmd. 1772 * We have to properly setup DMSGF_REPLY, however. 1773 */ 1774 pstate = state; 1775 state = kmalloc(sizeof(*state), iocom->mmsg, M_WAITOK | M_ZERO); 1776 TAILQ_INIT(&state->subq); 1777 state->iocom = iocom; 1778 state->parent = pstate; 1779 state->flags = KDMSG_STATE_DYNAMIC | 1780 KDMSG_STATE_NEW; 1781 state->func = func; 1782 state->any.any = data; 1783 state->msgid = (uint64_t)(uintptr_t)state; 1784 /*msg->any.head.msgid = state->msgid;XXX*/ 1785 1786 lockmgr(&iocom->msglk, LK_EXCLUSIVE); 1787 if (RB_INSERT(kdmsg_state_tree, &iocom->statewr_tree, state)) 1788 panic("duplicate msgid allocated"); 1789 if (TAILQ_EMPTY(&pstate->subq)) 1790 kdmsg_state_hold(pstate);/* pstate->subq */ 1791 TAILQ_INSERT_TAIL(&pstate->subq, state, entry); 1792 state->flags |= KDMSG_STATE_RBINSERTED | 1793 KDMSG_STATE_SUBINSERTED; 1794 state->flags |= pstate->flags & KDMSG_STATE_DYING; 1795 kdmsg_state_hold(state); /* pstate->subq */ 1796 kdmsg_state_hold(state); /* state on rbtree */ 1797 kdmsg_state_hold(state); /* msg->state */ 1798 lockmgr(&iocom->msglk, LK_RELEASE); 1799 } else { 1800 pstate = state->parent; 1801 KKASSERT(pstate != NULL); 1802 kdmsg_state_hold(state); /* msg->state */ 1803 } 1804 1805 if (state->flags & KDMSG_STATE_OPPOSITE) 1806 cmd |= DMSGF_REVTRANS; 1807 if (pstate->flags & KDMSG_STATE_OPPOSITE) 1808 cmd |= DMSGF_REVCIRC; 1809 1810 msg->any.head.magic = DMSG_HDR_MAGIC; 1811 msg->any.head.cmd = cmd; 1812 msg->any.head.msgid = state->msgid; 1813 msg->any.head.circuit = pstate->msgid; 1814 msg->state = state; 1815 1816 return (msg); 1817 } 1818 1819 void 1820 kdmsg_msg_free(kdmsg_msg_t *msg) 1821 { 1822 kdmsg_iocom_t *iocom = msg->state->iocom; 1823 kdmsg_state_t *state; 1824 1825 if ((msg->flags & KDMSG_FLAG_AUXALLOC) && 1826 msg->aux_data && msg->aux_size) { 1827 kfree(msg->aux_data, iocom->mmsg); 1828 msg->flags &= ~KDMSG_FLAG_AUXALLOC; 1829 } 1830 if ((state = msg->state) != NULL) { 1831 msg->state = NULL; 1832 kdmsg_state_drop(state); /* msg->state */ 1833 } 1834 msg->aux_data = NULL; 1835 msg->aux_size = 0; 1836 1837 kfree(msg, iocom->mmsg); 1838 } 1839 1840 void 1841 kdmsg_detach_aux_data(kdmsg_msg_t *msg, kdmsg_data_t *data) 1842 { 1843 if (msg->flags & KDMSG_FLAG_AUXALLOC) { 1844 data->aux_data = msg->aux_data; 1845 data->aux_size = msg->aux_size; 1846 data->iocom = msg->state->iocom; 1847 msg->flags &= ~KDMSG_FLAG_AUXALLOC; 1848 } else { 1849 data->aux_data = NULL; 1850 data->aux_size = 0; 1851 data->iocom = msg->state->iocom; 1852 } 1853 } 1854 1855 void 1856 kdmsg_free_aux_data(kdmsg_data_t *data) 1857 { 1858 if (data->aux_data) 1859 kfree(data->aux_data, data->iocom->mmsg); 1860 } 1861 1862 /* 1863 * Indexed messages are stored in a red-black tree indexed by their 1864 * msgid. Only persistent messages are indexed. 1865 */ 1866 int 1867 kdmsg_state_cmp(kdmsg_state_t *state1, kdmsg_state_t *state2) 1868 { 1869 if (state1->iocom < state2->iocom) 1870 return(-1); 1871 if (state1->iocom > state2->iocom) 1872 return(1); 1873 if (state1->msgid < state2->msgid) 1874 return(-1); 1875 if (state1->msgid > state2->msgid) 1876 return(1); 1877 return(0); 1878 } 1879 1880 /* 1881 * Write a message. All requisit command flags have been set. 1882 * 1883 * If msg->state is non-NULL the message is written to the existing 1884 * transaction. msgid will be set accordingly. 1885 * 1886 * If msg->state is NULL and CREATE is set new state is allocated and 1887 * (func, data) is installed. A msgid is assigned. 1888 * 1889 * If msg->state is NULL and CREATE is not set the message is assumed 1890 * to be a one-way message. The originator must assign the msgid 1891 * (or leave it 0, which is typical. 1892 * 1893 * This function merely queues the message to the management thread, it 1894 * does not write to the message socket/pipe. 1895 */ 1896 void 1897 kdmsg_msg_write(kdmsg_msg_t *msg) 1898 { 1899 kdmsg_iocom_t *iocom = msg->state->iocom; 1900 1901 lockmgr(&iocom->msglk, LK_EXCLUSIVE); 1902 kdmsg_msg_write_locked(iocom, msg); 1903 lockmgr(&iocom->msglk, LK_RELEASE); 1904 } 1905 1906 static void 1907 kdmsg_msg_write_locked(kdmsg_iocom_t *iocom, kdmsg_msg_t *msg) 1908 { 1909 kdmsg_state_t *state; 1910 1911 if (msg->state) { 1912 /* 1913 * Continuance or termination of existing transaction. 1914 * The transaction could have been initiated by either end. 1915 * 1916 * (Function callback and aux data for the receive side can 1917 * be replaced or left alone). 1918 */ 1919 state = msg->state; 1920 msg->any.head.msgid = state->msgid; 1921 } else { 1922 /* 1923 * One-off message (always uses msgid 0 to distinguish 1924 * between a possibly lost in-transaction message due to 1925 * competing aborts and a real one-off message?) 1926 */ 1927 state = NULL; 1928 msg->any.head.msgid = 0; 1929 } 1930 1931 #if 0 1932 /* 1933 * XXX removed - don't make this a panic, allow the state checks 1934 * below to catch the situation. 1935 * 1936 * This flag is not set until after the tx thread has drained 1937 * the tx msgq and simulated responses. After that point the 1938 * txthread is dead and can no longer simulate responses. 1939 * 1940 * Device drivers should never try to send a message once this 1941 * flag is set. They should have detected (through the state 1942 * closures) that the link is in trouble. 1943 */ 1944 if (iocom->flags & KDMSG_IOCOMF_EXITNOACC) { 1945 lockmgr(&iocom->msglk, LK_RELEASE); 1946 panic("kdmsg_msg_write: Attempt to write message to " 1947 "terminated iocom\n"); 1948 } 1949 #endif 1950 1951 /* 1952 * For stateful messages, if the circuit is dead or dying we have 1953 * to abort the potentially newly-created state and discard the 1954 * message. 1955 * 1956 * - We must discard the message because the other end will not 1957 * be expecting any more messages over the dead or dying circuit 1958 * and might not be able to receive them. 1959 * 1960 * - We abort the state by simulating a failure to generate a fake 1961 * incoming DELETE. This will trigger the state callback and allow 1962 * the device to clean things up and reply, closing the outgoing 1963 * direction and allowing the state to be freed. 1964 * 1965 * This situation occurs quite often, particularly as SPANs stabilize. 1966 * End-points must do the right thing. 1967 */ 1968 if (state) { 1969 KKASSERT((state->txcmd & DMSGF_DELETE) == 0); 1970 if (state->flags & KDMSG_STATE_DYING) { 1971 #if 0 1972 if ((state->flags & KDMSG_STATE_DYING) || 1973 (state->parent->txcmd & DMSGF_DELETE) || 1974 (state->parent->flags & KDMSG_STATE_DYING)) { 1975 #endif 1976 kdio_printf(iocom, 4, 1977 "kdmsg_msg_write: Write to dying circuit " 1978 "state=%p " 1979 "ptxcmd=%08x prxcmd=%08x flags=%08x\n", 1980 state, 1981 state->parent->rxcmd, 1982 state->parent->txcmd, 1983 state->parent->flags); 1984 kdmsg_state_hold(state); 1985 kdmsg_state_msgtx(msg); 1986 kdmsg_state_cleanuptx(msg); 1987 kdmsg_state_drop(state); 1988 return; 1989 } 1990 } 1991 1992 /* 1993 * Finish up the msg fields. Note that msg->aux_size and the 1994 * aux_bytes stored in the message header represent the unaligned 1995 * (actual) bytes of data, but the buffer is sized to an aligned 1996 * size and the CRC is generated over the aligned length. 1997 */ 1998 msg->any.head.salt = /* (random << 8) | */ (iocom->msg_seq & 255); 1999 ++iocom->msg_seq; 2000 2001 if (msg->aux_data && msg->aux_size) { 2002 uint32_t abytes = DMSG_DOALIGN(msg->aux_size); 2003 2004 msg->any.head.aux_bytes = msg->aux_size; 2005 msg->any.head.aux_crc = iscsi_crc32(msg->aux_data, abytes); 2006 } 2007 msg->any.head.hdr_crc = 0; 2008 msg->any.head.hdr_crc = iscsi_crc32(msg->any.buf, msg->hdr_size); 2009 2010 TAILQ_INSERT_TAIL(&iocom->msgq, msg, qentry); 2011 2012 if (iocom->msg_ctl & KDMSG_CLUSTERCTL_SLEEPING) { 2013 atomic_clear_int(&iocom->msg_ctl, 2014 KDMSG_CLUSTERCTL_SLEEPING); 2015 wakeup(&iocom->msg_ctl); 2016 } 2017 } 2018 2019 /* 2020 * Reply to a message and terminate our side of the transaction. 2021 * 2022 * If msg->state is non-NULL we are replying to a one-way message. 2023 */ 2024 void 2025 kdmsg_msg_reply(kdmsg_msg_t *msg, uint32_t error) 2026 { 2027 kdmsg_state_t *state = msg->state; 2028 kdmsg_msg_t *nmsg; 2029 uint32_t cmd; 2030 2031 /* 2032 * Reply with a simple error code and terminate the transaction. 2033 */ 2034 cmd = DMSG_LNK_ERROR; 2035 2036 /* 2037 * Check if our direction has even been initiated yet, set CREATE. 2038 * 2039 * Check what direction this is (command or reply direction). Note 2040 * that txcmd might not have been initiated yet. 2041 * 2042 * If our direction has already been closed we just return without 2043 * doing anything. 2044 */ 2045 if (state != &state->iocom->state0) { 2046 if (state->txcmd & DMSGF_DELETE) 2047 return; 2048 if ((state->txcmd & DMSGF_CREATE) == 0) 2049 cmd |= DMSGF_CREATE; 2050 if (state->txcmd & DMSGF_REPLY) 2051 cmd |= DMSGF_REPLY; 2052 cmd |= DMSGF_DELETE; 2053 } else { 2054 if ((msg->any.head.cmd & DMSGF_REPLY) == 0) 2055 cmd |= DMSGF_REPLY; 2056 } 2057 2058 nmsg = kdmsg_msg_alloc(state, cmd, NULL, NULL); 2059 nmsg->any.head.error = error; 2060 kdmsg_msg_write(nmsg); 2061 } 2062 2063 /* 2064 * Reply to a message and continue our side of the transaction. 2065 * 2066 * If msg->state is non-NULL we are replying to a one-way message and this 2067 * function degenerates into the same as kdmsg_msg_reply(). 2068 */ 2069 void 2070 kdmsg_msg_result(kdmsg_msg_t *msg, uint32_t error) 2071 { 2072 kdmsg_state_t *state = msg->state; 2073 kdmsg_msg_t *nmsg; 2074 uint32_t cmd; 2075 2076 /* 2077 * Return a simple result code, do NOT terminate the transaction. 2078 */ 2079 cmd = DMSG_LNK_ERROR; 2080 2081 /* 2082 * Check if our direction has even been initiated yet, set CREATE. 2083 * 2084 * Check what direction this is (command or reply direction). Note 2085 * that txcmd might not have been initiated yet. 2086 * 2087 * If our direction has already been closed we just return without 2088 * doing anything. 2089 */ 2090 if (state != &state->iocom->state0) { 2091 if (state->txcmd & DMSGF_DELETE) 2092 return; 2093 if ((state->txcmd & DMSGF_CREATE) == 0) 2094 cmd |= DMSGF_CREATE; 2095 if (state->txcmd & DMSGF_REPLY) 2096 cmd |= DMSGF_REPLY; 2097 /* continuing transaction, do not set MSGF_DELETE */ 2098 } else { 2099 if ((msg->any.head.cmd & DMSGF_REPLY) == 0) 2100 cmd |= DMSGF_REPLY; 2101 } 2102 2103 nmsg = kdmsg_msg_alloc(state, cmd, NULL, NULL); 2104 nmsg->any.head.error = error; 2105 kdmsg_msg_write(nmsg); 2106 } 2107 2108 /* 2109 * Reply to a message and terminate our side of the transaction. 2110 * 2111 * If msg->state is non-NULL we are replying to a one-way message. 2112 */ 2113 void 2114 kdmsg_state_reply(kdmsg_state_t *state, uint32_t error) 2115 { 2116 kdmsg_msg_t *nmsg; 2117 uint32_t cmd; 2118 2119 /* 2120 * Reply with a simple error code and terminate the transaction. 2121 */ 2122 cmd = DMSG_LNK_ERROR; 2123 2124 /* 2125 * Check if our direction has even been initiated yet, set CREATE. 2126 * 2127 * Check what direction this is (command or reply direction). Note 2128 * that txcmd might not have been initiated yet. 2129 * 2130 * If our direction has already been closed we just return without 2131 * doing anything. 2132 */ 2133 KKASSERT(state); 2134 if (state->txcmd & DMSGF_DELETE) 2135 return; 2136 if ((state->txcmd & DMSGF_CREATE) == 0) 2137 cmd |= DMSGF_CREATE; 2138 if (state->txcmd & DMSGF_REPLY) 2139 cmd |= DMSGF_REPLY; 2140 cmd |= DMSGF_DELETE; 2141 2142 nmsg = kdmsg_msg_alloc(state, cmd, NULL, NULL); 2143 nmsg->any.head.error = error; 2144 kdmsg_msg_write(nmsg); 2145 } 2146 2147 /* 2148 * Reply to a message and continue our side of the transaction. 2149 * 2150 * If msg->state is non-NULL we are replying to a one-way message and this 2151 * function degenerates into the same as kdmsg_msg_reply(). 2152 */ 2153 void 2154 kdmsg_state_result(kdmsg_state_t *state, uint32_t error) 2155 { 2156 kdmsg_msg_t *nmsg; 2157 uint32_t cmd; 2158 2159 /* 2160 * Return a simple result code, do NOT terminate the transaction. 2161 */ 2162 cmd = DMSG_LNK_ERROR; 2163 2164 /* 2165 * Check if our direction has even been initiated yet, set CREATE. 2166 * 2167 * Check what direction this is (command or reply direction). Note 2168 * that txcmd might not have been initiated yet. 2169 * 2170 * If our direction has already been closed we just return without 2171 * doing anything. 2172 */ 2173 KKASSERT(state); 2174 if (state->txcmd & DMSGF_DELETE) 2175 return; 2176 if ((state->txcmd & DMSGF_CREATE) == 0) 2177 cmd |= DMSGF_CREATE; 2178 if (state->txcmd & DMSGF_REPLY) 2179 cmd |= DMSGF_REPLY; 2180 /* continuing transaction, do not set MSGF_DELETE */ 2181 2182 nmsg = kdmsg_msg_alloc(state, cmd, NULL, NULL); 2183 nmsg->any.head.error = error; 2184 kdmsg_msg_write(nmsg); 2185 } 2186