1/* 2 * This file is subject to the terms and conditions of the GNU General Public 3 * License. See the file "COPYING" in the main directory of this archive 4 * for more details. 5 * 6 * (C) Copyright 2020 Hewlett Packard Enterprise Development LP 7 * Copyright (c) 2004-2009 Silicon Graphics, Inc. All Rights Reserved. 8 */ 9 10/* 11 * Cross Partition Communication (XPC) support - standard version. 12 * 13 * XPC provides a message passing capability that crosses partition 14 * boundaries. This module is made up of two parts: 15 * 16 * partition This part detects the presence/absence of other 17 * partitions. It provides a heartbeat and monitors 18 * the heartbeats of other partitions. 19 * 20 * channel This part manages the channels and sends/receives 21 * messages across them to/from other partitions. 22 * 23 * There are a couple of additional functions residing in XP, which 24 * provide an interface to XPC for its users. 25 * 26 * 27 * Caveats: 28 * 29 * . Currently on sn2, we have no way to determine which nasid an IRQ 30 * came from. Thus, xpc_send_IRQ_sn2() does a remote amo write 31 * followed by an IPI. The amo indicates where data is to be pulled 32 * from, so after the IPI arrives, the remote partition checks the amo 33 * word. The IPI can actually arrive before the amo however, so other 34 * code must periodically check for this case. Also, remote amo 35 * operations do not reliably time out. Thus we do a remote PIO read 36 * solely to know whether the remote partition is down and whether we 37 * should stop sending IPIs to it. This remote PIO read operation is 38 * set up in a special nofault region so SAL knows to ignore (and 39 * cleanup) any errors due to the remote amo write, PIO read, and/or 40 * PIO write operations. 41 * 42 * If/when new hardware solves this IPI problem, we should abandon 43 * the current approach. 44 * 45 */ 46 47#include <linux/module.h> 48#include <linux/slab.h> 49#include <linux/sysctl.h> 50#include <linux/device.h> 51#include <linux/delay.h> 52#include <linux/reboot.h> 53#include <linux/kdebug.h> 54#include <linux/kthread.h> 55#include "xpc.h" 56 57#ifdef CONFIG_X86_64 58#include <asm/traps.h> 59#endif 60 61/* define two XPC debug device structures to be used with dev_dbg() et al */ 62 63static struct device_driver xpc_dbg_name = { 64 .name = "xpc" 65}; 66 67static struct device xpc_part_dbg_subname = { 68 .init_name = "", /* set to "part" at xpc_init() time */ 69 .driver = &xpc_dbg_name 70}; 71 72static struct device xpc_chan_dbg_subname = { 73 .init_name = "", /* set to "chan" at xpc_init() time */ 74 .driver = &xpc_dbg_name 75}; 76 77struct device *xpc_part = &xpc_part_dbg_subname; 78struct device *xpc_chan = &xpc_chan_dbg_subname; 79 80static int xpc_kdebug_ignore; 81 82/* systune related variables for /proc/sys directories */ 83 84static int xpc_hb_interval = XPC_HB_DEFAULT_INTERVAL; 85static int xpc_hb_min_interval = 1; 86static int xpc_hb_max_interval = 10; 87 88static int xpc_hb_check_interval = XPC_HB_CHECK_DEFAULT_INTERVAL; 89static int xpc_hb_check_min_interval = 10; 90static int xpc_hb_check_max_interval = 120; 91 92int xpc_disengage_timelimit = XPC_DISENGAGE_DEFAULT_TIMELIMIT; 93static int xpc_disengage_min_timelimit; /* = 0 */ 94static int xpc_disengage_max_timelimit = 120; 95 96static struct ctl_table xpc_sys_xpc_hb_dir[] = { 97 { 98 .procname = "hb_interval", 99 .data = &xpc_hb_interval, 100 .maxlen = sizeof(int), 101 .mode = 0644, 102 .proc_handler = proc_dointvec_minmax, 103 .extra1 = &xpc_hb_min_interval, 104 .extra2 = &xpc_hb_max_interval}, 105 { 106 .procname = "hb_check_interval", 107 .data = &xpc_hb_check_interval, 108 .maxlen = sizeof(int), 109 .mode = 0644, 110 .proc_handler = proc_dointvec_minmax, 111 .extra1 = &xpc_hb_check_min_interval, 112 .extra2 = &xpc_hb_check_max_interval}, 113 {} 114}; 115static struct ctl_table xpc_sys_xpc_dir[] = { 116 { 117 .procname = "hb", 118 .mode = 0555, 119 .child = xpc_sys_xpc_hb_dir}, 120 { 121 .procname = "disengage_timelimit", 122 .data = &xpc_disengage_timelimit, 123 .maxlen = sizeof(int), 124 .mode = 0644, 125 .proc_handler = proc_dointvec_minmax, 126 .extra1 = &xpc_disengage_min_timelimit, 127 .extra2 = &xpc_disengage_max_timelimit}, 128 {} 129}; 130static struct ctl_table xpc_sys_dir[] = { 131 { 132 .procname = "xpc", 133 .mode = 0555, 134 .child = xpc_sys_xpc_dir}, 135 {} 136}; 137static struct ctl_table_header *xpc_sysctl; 138 139/* non-zero if any remote partition disengage was timed out */ 140int xpc_disengage_timedout; 141 142/* #of activate IRQs received and not yet processed */ 143int xpc_activate_IRQ_rcvd; 144DEFINE_SPINLOCK(xpc_activate_IRQ_rcvd_lock); 145 146/* IRQ handler notifies this wait queue on receipt of an IRQ */ 147DECLARE_WAIT_QUEUE_HEAD(xpc_activate_IRQ_wq); 148 149static unsigned long xpc_hb_check_timeout; 150static struct timer_list xpc_hb_timer; 151 152/* notification that the xpc_hb_checker thread has exited */ 153static DECLARE_COMPLETION(xpc_hb_checker_exited); 154 155/* notification that the xpc_discovery thread has exited */ 156static DECLARE_COMPLETION(xpc_discovery_exited); 157 158static void xpc_kthread_waitmsgs(struct xpc_partition *, struct xpc_channel *); 159 160static int xpc_system_reboot(struct notifier_block *, unsigned long, void *); 161static struct notifier_block xpc_reboot_notifier = { 162 .notifier_call = xpc_system_reboot, 163}; 164 165static int xpc_system_die(struct notifier_block *, unsigned long, void *); 166static struct notifier_block xpc_die_notifier = { 167 .notifier_call = xpc_system_die, 168}; 169 170struct xpc_arch_operations xpc_arch_ops; 171 172/* 173 * Timer function to enforce the timelimit on the partition disengage. 174 */ 175static void 176xpc_timeout_partition_disengage(struct timer_list *t) 177{ 178 struct xpc_partition *part = from_timer(part, t, disengage_timer); 179 180 DBUG_ON(time_is_after_jiffies(part->disengage_timeout)); 181 182 (void)xpc_partition_disengaged(part); 183 184 DBUG_ON(part->disengage_timeout != 0); 185 DBUG_ON(xpc_arch_ops.partition_engaged(XPC_PARTID(part))); 186} 187 188/* 189 * Timer to produce the heartbeat. The timer structures function is 190 * already set when this is initially called. A tunable is used to 191 * specify when the next timeout should occur. 192 */ 193static void 194xpc_hb_beater(struct timer_list *unused) 195{ 196 xpc_arch_ops.increment_heartbeat(); 197 198 if (time_is_before_eq_jiffies(xpc_hb_check_timeout)) 199 wake_up_interruptible(&xpc_activate_IRQ_wq); 200 201 xpc_hb_timer.expires = jiffies + (xpc_hb_interval * HZ); 202 add_timer(&xpc_hb_timer); 203} 204 205static void 206xpc_start_hb_beater(void) 207{ 208 xpc_arch_ops.heartbeat_init(); 209 timer_setup(&xpc_hb_timer, xpc_hb_beater, 0); 210 xpc_hb_beater(0); 211} 212 213static void 214xpc_stop_hb_beater(void) 215{ 216 del_timer_sync(&xpc_hb_timer); 217 xpc_arch_ops.heartbeat_exit(); 218} 219 220/* 221 * At periodic intervals, scan through all active partitions and ensure 222 * their heartbeat is still active. If not, the partition is deactivated. 223 */ 224static void 225xpc_check_remote_hb(void) 226{ 227 struct xpc_partition *part; 228 short partid; 229 enum xp_retval ret; 230 231 for (partid = 0; partid < xp_max_npartitions; partid++) { 232 233 if (xpc_exiting) 234 break; 235 236 if (partid == xp_partition_id) 237 continue; 238 239 part = &xpc_partitions[partid]; 240 241 if (part->act_state == XPC_P_AS_INACTIVE || 242 part->act_state == XPC_P_AS_DEACTIVATING) { 243 continue; 244 } 245 246 ret = xpc_arch_ops.get_remote_heartbeat(part); 247 if (ret != xpSuccess) 248 XPC_DEACTIVATE_PARTITION(part, ret); 249 } 250} 251 252/* 253 * This thread is responsible for nearly all of the partition 254 * activation/deactivation. 255 */ 256static int 257xpc_hb_checker(void *ignore) 258{ 259 int force_IRQ = 0; 260 261 /* this thread was marked active by xpc_hb_init() */ 262 263 set_cpus_allowed_ptr(current, cpumask_of(XPC_HB_CHECK_CPU)); 264 265 /* set our heartbeating to other partitions into motion */ 266 xpc_hb_check_timeout = jiffies + (xpc_hb_check_interval * HZ); 267 xpc_start_hb_beater(); 268 269 while (!xpc_exiting) { 270 271 dev_dbg(xpc_part, "woke up with %d ticks rem; %d IRQs have " 272 "been received\n", 273 (int)(xpc_hb_check_timeout - jiffies), 274 xpc_activate_IRQ_rcvd); 275 276 /* checking of remote heartbeats is skewed by IRQ handling */ 277 if (time_is_before_eq_jiffies(xpc_hb_check_timeout)) { 278 xpc_hb_check_timeout = jiffies + 279 (xpc_hb_check_interval * HZ); 280 281 dev_dbg(xpc_part, "checking remote heartbeats\n"); 282 xpc_check_remote_hb(); 283 } 284 285 /* check for outstanding IRQs */ 286 if (xpc_activate_IRQ_rcvd > 0 || force_IRQ != 0) { 287 force_IRQ = 0; 288 dev_dbg(xpc_part, "processing activate IRQs " 289 "received\n"); 290 xpc_arch_ops.process_activate_IRQ_rcvd(); 291 } 292 293 /* wait for IRQ or timeout */ 294 (void)wait_event_interruptible(xpc_activate_IRQ_wq, 295 (time_is_before_eq_jiffies( 296 xpc_hb_check_timeout) || 297 xpc_activate_IRQ_rcvd > 0 || 298 xpc_exiting)); 299 } 300 301 xpc_stop_hb_beater(); 302 303 dev_dbg(xpc_part, "heartbeat checker is exiting\n"); 304 305 /* mark this thread as having exited */ 306 complete(&xpc_hb_checker_exited); 307 return 0; 308} 309 310/* 311 * This thread will attempt to discover other partitions to activate 312 * based on info provided by SAL. This new thread is short lived and 313 * will exit once discovery is complete. 314 */ 315static int 316xpc_initiate_discovery(void *ignore) 317{ 318 xpc_discovery(); 319 320 dev_dbg(xpc_part, "discovery thread is exiting\n"); 321 322 /* mark this thread as having exited */ 323 complete(&xpc_discovery_exited); 324 return 0; 325} 326 327/* 328 * The first kthread assigned to a newly activated partition is the one 329 * created by XPC HB with which it calls xpc_activating(). XPC hangs on to 330 * that kthread until the partition is brought down, at which time that kthread 331 * returns back to XPC HB. (The return of that kthread will signify to XPC HB 332 * that XPC has dismantled all communication infrastructure for the associated 333 * partition.) This kthread becomes the channel manager for that partition. 334 * 335 * Each active partition has a channel manager, who, besides connecting and 336 * disconnecting channels, will ensure that each of the partition's connected 337 * channels has the required number of assigned kthreads to get the work done. 338 */ 339static void 340xpc_channel_mgr(struct xpc_partition *part) 341{ 342 while (part->act_state != XPC_P_AS_DEACTIVATING || 343 atomic_read(&part->nchannels_active) > 0 || 344 !xpc_partition_disengaged(part)) { 345 346 xpc_process_sent_chctl_flags(part); 347 348 /* 349 * Wait until we've been requested to activate kthreads or 350 * all of the channel's message queues have been torn down or 351 * a signal is pending. 352 * 353 * The channel_mgr_requests is set to 1 after being awakened, 354 * This is done to prevent the channel mgr from making one pass 355 * through the loop for each request, since he will 356 * be servicing all the requests in one pass. The reason it's 357 * set to 1 instead of 0 is so that other kthreads will know 358 * that the channel mgr is running and won't bother trying to 359 * wake him up. 360 */ 361 atomic_dec(&part->channel_mgr_requests); 362 (void)wait_event_interruptible(part->channel_mgr_wq, 363 (atomic_read(&part->channel_mgr_requests) > 0 || 364 part->chctl.all_flags != 0 || 365 (part->act_state == XPC_P_AS_DEACTIVATING && 366 atomic_read(&part->nchannels_active) == 0 && 367 xpc_partition_disengaged(part)))); 368 atomic_set(&part->channel_mgr_requests, 1); 369 } 370} 371 372/* 373 * Guarantee that the kzalloc'd memory is cacheline aligned. 374 */ 375void * 376xpc_kzalloc_cacheline_aligned(size_t size, gfp_t flags, void **base) 377{ 378 /* see if kzalloc will give us cachline aligned memory by default */ 379 *base = kzalloc(size, flags); 380 if (*base == NULL) 381 return NULL; 382 383 if ((u64)*base == L1_CACHE_ALIGN((u64)*base)) 384 return *base; 385 386 kfree(*base); 387 388 /* nope, we'll have to do it ourselves */ 389 *base = kzalloc(size + L1_CACHE_BYTES, flags); 390 if (*base == NULL) 391 return NULL; 392 393 return (void *)L1_CACHE_ALIGN((u64)*base); 394} 395 396/* 397 * Setup the channel structures necessary to support XPartition Communication 398 * between the specified remote partition and the local one. 399 */ 400static enum xp_retval 401xpc_setup_ch_structures(struct xpc_partition *part) 402{ 403 enum xp_retval ret; 404 int ch_number; 405 struct xpc_channel *ch; 406 short partid = XPC_PARTID(part); 407 408 /* 409 * Allocate all of the channel structures as a contiguous chunk of 410 * memory. 411 */ 412 DBUG_ON(part->channels != NULL); 413 part->channels = kcalloc(XPC_MAX_NCHANNELS, 414 sizeof(struct xpc_channel), 415 GFP_KERNEL); 416 if (part->channels == NULL) { 417 dev_err(xpc_chan, "can't get memory for channels\n"); 418 return xpNoMemory; 419 } 420 421 /* allocate the remote open and close args */ 422 423 part->remote_openclose_args = 424 xpc_kzalloc_cacheline_aligned(XPC_OPENCLOSE_ARGS_SIZE, 425 GFP_KERNEL, &part-> 426 remote_openclose_args_base); 427 if (part->remote_openclose_args == NULL) { 428 dev_err(xpc_chan, "can't get memory for remote connect args\n"); 429 ret = xpNoMemory; 430 goto out_1; 431 } 432 433 part->chctl.all_flags = 0; 434 spin_lock_init(&part->chctl_lock); 435 436 atomic_set(&part->channel_mgr_requests, 1); 437 init_waitqueue_head(&part->channel_mgr_wq); 438 439 part->nchannels = XPC_MAX_NCHANNELS; 440 441 atomic_set(&part->nchannels_active, 0); 442 atomic_set(&part->nchannels_engaged, 0); 443 444 for (ch_number = 0; ch_number < part->nchannels; ch_number++) { 445 ch = &part->channels[ch_number]; 446 447 ch->partid = partid; 448 ch->number = ch_number; 449 ch->flags = XPC_C_DISCONNECTED; 450 451 atomic_set(&ch->kthreads_assigned, 0); 452 atomic_set(&ch->kthreads_idle, 0); 453 atomic_set(&ch->kthreads_active, 0); 454 455 atomic_set(&ch->references, 0); 456 atomic_set(&ch->n_to_notify, 0); 457 458 spin_lock_init(&ch->lock); 459 init_completion(&ch->wdisconnect_wait); 460 461 atomic_set(&ch->n_on_msg_allocate_wq, 0); 462 init_waitqueue_head(&ch->msg_allocate_wq); 463 init_waitqueue_head(&ch->idle_wq); 464 } 465 466 ret = xpc_arch_ops.setup_ch_structures(part); 467 if (ret != xpSuccess) 468 goto out_2; 469 470 /* 471 * With the setting of the partition setup_state to XPC_P_SS_SETUP, 472 * we're declaring that this partition is ready to go. 473 */ 474 part->setup_state = XPC_P_SS_SETUP; 475 476 return xpSuccess; 477 478 /* setup of ch structures failed */ 479out_2: 480 kfree(part->remote_openclose_args_base); 481 part->remote_openclose_args = NULL; 482out_1: 483 kfree(part->channels); 484 part->channels = NULL; 485 return ret; 486} 487 488/* 489 * Teardown the channel structures necessary to support XPartition Communication 490 * between the specified remote partition and the local one. 491 */ 492static void 493xpc_teardown_ch_structures(struct xpc_partition *part) 494{ 495 DBUG_ON(atomic_read(&part->nchannels_engaged) != 0); 496 DBUG_ON(atomic_read(&part->nchannels_active) != 0); 497 498 /* 499 * Make this partition inaccessible to local processes by marking it 500 * as no longer setup. Then wait before proceeding with the teardown 501 * until all existing references cease. 502 */ 503 DBUG_ON(part->setup_state != XPC_P_SS_SETUP); 504 part->setup_state = XPC_P_SS_WTEARDOWN; 505 506 wait_event(part->teardown_wq, (atomic_read(&part->references) == 0)); 507 508 /* now we can begin tearing down the infrastructure */ 509 510 xpc_arch_ops.teardown_ch_structures(part); 511 512 kfree(part->remote_openclose_args_base); 513 part->remote_openclose_args = NULL; 514 kfree(part->channels); 515 part->channels = NULL; 516 517 part->setup_state = XPC_P_SS_TORNDOWN; 518} 519 520/* 521 * When XPC HB determines that a partition has come up, it will create a new 522 * kthread and that kthread will call this function to attempt to set up the 523 * basic infrastructure used for Cross Partition Communication with the newly 524 * upped partition. 525 * 526 * The kthread that was created by XPC HB and which setup the XPC 527 * infrastructure will remain assigned to the partition becoming the channel 528 * manager for that partition until the partition is deactivating, at which 529 * time the kthread will teardown the XPC infrastructure and then exit. 530 */ 531static int 532xpc_activating(void *__partid) 533{ 534 short partid = (u64)__partid; 535 struct xpc_partition *part = &xpc_partitions[partid]; 536 unsigned long irq_flags; 537 538 DBUG_ON(partid < 0 || partid >= xp_max_npartitions); 539 540 spin_lock_irqsave(&part->act_lock, irq_flags); 541 542 if (part->act_state == XPC_P_AS_DEACTIVATING) { 543 part->act_state = XPC_P_AS_INACTIVE; 544 spin_unlock_irqrestore(&part->act_lock, irq_flags); 545 part->remote_rp_pa = 0; 546 return 0; 547 } 548 549 /* indicate the thread is activating */ 550 DBUG_ON(part->act_state != XPC_P_AS_ACTIVATION_REQ); 551 part->act_state = XPC_P_AS_ACTIVATING; 552 553 XPC_SET_REASON(part, 0, 0); 554 spin_unlock_irqrestore(&part->act_lock, irq_flags); 555 556 dev_dbg(xpc_part, "activating partition %d\n", partid); 557 558 xpc_arch_ops.allow_hb(partid); 559 560 if (xpc_setup_ch_structures(part) == xpSuccess) { 561 (void)xpc_part_ref(part); /* this will always succeed */ 562 563 if (xpc_arch_ops.make_first_contact(part) == xpSuccess) { 564 xpc_mark_partition_active(part); 565 xpc_channel_mgr(part); 566 /* won't return until partition is deactivating */ 567 } 568 569 xpc_part_deref(part); 570 xpc_teardown_ch_structures(part); 571 } 572 573 xpc_arch_ops.disallow_hb(partid); 574 xpc_mark_partition_inactive(part); 575 576 if (part->reason == xpReactivating) { 577 /* interrupting ourselves results in activating partition */ 578 xpc_arch_ops.request_partition_reactivation(part); 579 } 580 581 return 0; 582} 583 584void 585xpc_activate_partition(struct xpc_partition *part) 586{ 587 short partid = XPC_PARTID(part); 588 unsigned long irq_flags; 589 struct task_struct *kthread; 590 591 spin_lock_irqsave(&part->act_lock, irq_flags); 592 593 DBUG_ON(part->act_state != XPC_P_AS_INACTIVE); 594 595 part->act_state = XPC_P_AS_ACTIVATION_REQ; 596 XPC_SET_REASON(part, xpCloneKThread, __LINE__); 597 598 spin_unlock_irqrestore(&part->act_lock, irq_flags); 599 600 kthread = kthread_run(xpc_activating, (void *)((u64)partid), "xpc%02d", 601 partid); 602 if (IS_ERR(kthread)) { 603 spin_lock_irqsave(&part->act_lock, irq_flags); 604 part->act_state = XPC_P_AS_INACTIVE; 605 XPC_SET_REASON(part, xpCloneKThreadFailed, __LINE__); 606 spin_unlock_irqrestore(&part->act_lock, irq_flags); 607 } 608} 609 610void 611xpc_activate_kthreads(struct xpc_channel *ch, int needed) 612{ 613 int idle = atomic_read(&ch->kthreads_idle); 614 int assigned = atomic_read(&ch->kthreads_assigned); 615 int wakeup; 616 617 DBUG_ON(needed <= 0); 618 619 if (idle > 0) { 620 wakeup = (needed > idle) ? idle : needed; 621 needed -= wakeup; 622 623 dev_dbg(xpc_chan, "wakeup %d idle kthreads, partid=%d, " 624 "channel=%d\n", wakeup, ch->partid, ch->number); 625 626 /* only wakeup the requested number of kthreads */ 627 wake_up_nr(&ch->idle_wq, wakeup); 628 } 629 630 if (needed <= 0) 631 return; 632 633 if (needed + assigned > ch->kthreads_assigned_limit) { 634 needed = ch->kthreads_assigned_limit - assigned; 635 if (needed <= 0) 636 return; 637 } 638 639 dev_dbg(xpc_chan, "create %d new kthreads, partid=%d, channel=%d\n", 640 needed, ch->partid, ch->number); 641 642 xpc_create_kthreads(ch, needed, 0); 643} 644 645/* 646 * This function is where XPC's kthreads wait for messages to deliver. 647 */ 648static void 649xpc_kthread_waitmsgs(struct xpc_partition *part, struct xpc_channel *ch) 650{ 651 int (*n_of_deliverable_payloads) (struct xpc_channel *) = 652 xpc_arch_ops.n_of_deliverable_payloads; 653 654 do { 655 /* deliver messages to their intended recipients */ 656 657 while (n_of_deliverable_payloads(ch) > 0 && 658 !(ch->flags & XPC_C_DISCONNECTING)) { 659 xpc_deliver_payload(ch); 660 } 661 662 if (atomic_inc_return(&ch->kthreads_idle) > 663 ch->kthreads_idle_limit) { 664 /* too many idle kthreads on this channel */ 665 atomic_dec(&ch->kthreads_idle); 666 break; 667 } 668 669 dev_dbg(xpc_chan, "idle kthread calling " 670 "wait_event_interruptible_exclusive()\n"); 671 672 (void)wait_event_interruptible_exclusive(ch->idle_wq, 673 (n_of_deliverable_payloads(ch) > 0 || 674 (ch->flags & XPC_C_DISCONNECTING))); 675 676 atomic_dec(&ch->kthreads_idle); 677 678 } while (!(ch->flags & XPC_C_DISCONNECTING)); 679} 680 681static int 682xpc_kthread_start(void *args) 683{ 684 short partid = XPC_UNPACK_ARG1(args); 685 u16 ch_number = XPC_UNPACK_ARG2(args); 686 struct xpc_partition *part = &xpc_partitions[partid]; 687 struct xpc_channel *ch; 688 int n_needed; 689 unsigned long irq_flags; 690 int (*n_of_deliverable_payloads) (struct xpc_channel *) = 691 xpc_arch_ops.n_of_deliverable_payloads; 692 693 dev_dbg(xpc_chan, "kthread starting, partid=%d, channel=%d\n", 694 partid, ch_number); 695 696 ch = &part->channels[ch_number]; 697 698 if (!(ch->flags & XPC_C_DISCONNECTING)) { 699 700 /* let registerer know that connection has been established */ 701 702 spin_lock_irqsave(&ch->lock, irq_flags); 703 if (!(ch->flags & XPC_C_CONNECTEDCALLOUT)) { 704 ch->flags |= XPC_C_CONNECTEDCALLOUT; 705 spin_unlock_irqrestore(&ch->lock, irq_flags); 706 707 xpc_connected_callout(ch); 708 709 spin_lock_irqsave(&ch->lock, irq_flags); 710 ch->flags |= XPC_C_CONNECTEDCALLOUT_MADE; 711 spin_unlock_irqrestore(&ch->lock, irq_flags); 712 713 /* 714 * It is possible that while the callout was being 715 * made that the remote partition sent some messages. 716 * If that is the case, we may need to activate 717 * additional kthreads to help deliver them. We only 718 * need one less than total #of messages to deliver. 719 */ 720 n_needed = n_of_deliverable_payloads(ch) - 1; 721 if (n_needed > 0 && !(ch->flags & XPC_C_DISCONNECTING)) 722 xpc_activate_kthreads(ch, n_needed); 723 724 } else { 725 spin_unlock_irqrestore(&ch->lock, irq_flags); 726 } 727 728 xpc_kthread_waitmsgs(part, ch); 729 } 730 731 /* let registerer know that connection is disconnecting */ 732 733 spin_lock_irqsave(&ch->lock, irq_flags); 734 if ((ch->flags & XPC_C_CONNECTEDCALLOUT_MADE) && 735 !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) { 736 ch->flags |= XPC_C_DISCONNECTINGCALLOUT; 737 spin_unlock_irqrestore(&ch->lock, irq_flags); 738 739 xpc_disconnect_callout(ch, xpDisconnecting); 740 741 spin_lock_irqsave(&ch->lock, irq_flags); 742 ch->flags |= XPC_C_DISCONNECTINGCALLOUT_MADE; 743 } 744 spin_unlock_irqrestore(&ch->lock, irq_flags); 745 746 if (atomic_dec_return(&ch->kthreads_assigned) == 0 && 747 atomic_dec_return(&part->nchannels_engaged) == 0) { 748 xpc_arch_ops.indicate_partition_disengaged(part); 749 } 750 751 xpc_msgqueue_deref(ch); 752 753 dev_dbg(xpc_chan, "kthread exiting, partid=%d, channel=%d\n", 754 partid, ch_number); 755 756 xpc_part_deref(part); 757 return 0; 758} 759 760/* 761 * For each partition that XPC has established communications with, there is 762 * a minimum of one kernel thread assigned to perform any operation that 763 * may potentially sleep or block (basically the callouts to the asynchronous 764 * functions registered via xpc_connect()). 765 * 766 * Additional kthreads are created and destroyed by XPC as the workload 767 * demands. 768 * 769 * A kthread is assigned to one of the active channels that exists for a given 770 * partition. 771 */ 772void 773xpc_create_kthreads(struct xpc_channel *ch, int needed, 774 int ignore_disconnecting) 775{ 776 unsigned long irq_flags; 777 u64 args = XPC_PACK_ARGS(ch->partid, ch->number); 778 struct xpc_partition *part = &xpc_partitions[ch->partid]; 779 struct task_struct *kthread; 780 void (*indicate_partition_disengaged) (struct xpc_partition *) = 781 xpc_arch_ops.indicate_partition_disengaged; 782 783 while (needed-- > 0) { 784 785 /* 786 * The following is done on behalf of the newly created 787 * kthread. That kthread is responsible for doing the 788 * counterpart to the following before it exits. 789 */ 790 if (ignore_disconnecting) { 791 if (!atomic_inc_not_zero(&ch->kthreads_assigned)) { 792 /* kthreads assigned had gone to zero */ 793 BUG_ON(!(ch->flags & 794 XPC_C_DISCONNECTINGCALLOUT_MADE)); 795 break; 796 } 797 798 } else if (ch->flags & XPC_C_DISCONNECTING) { 799 break; 800 801 } else if (atomic_inc_return(&ch->kthreads_assigned) == 1 && 802 atomic_inc_return(&part->nchannels_engaged) == 1) { 803 xpc_arch_ops.indicate_partition_engaged(part); 804 } 805 (void)xpc_part_ref(part); 806 xpc_msgqueue_ref(ch); 807 808 kthread = kthread_run(xpc_kthread_start, (void *)args, 809 "xpc%02dc%d", ch->partid, ch->number); 810 if (IS_ERR(kthread)) { 811 /* the fork failed */ 812 813 /* 814 * NOTE: if (ignore_disconnecting && 815 * !(ch->flags & XPC_C_DISCONNECTINGCALLOUT)) is true, 816 * then we'll deadlock if all other kthreads assigned 817 * to this channel are blocked in the channel's 818 * registerer, because the only thing that will unblock 819 * them is the xpDisconnecting callout that this 820 * failed kthread_run() would have made. 821 */ 822 823 if (atomic_dec_return(&ch->kthreads_assigned) == 0 && 824 atomic_dec_return(&part->nchannels_engaged) == 0) { 825 indicate_partition_disengaged(part); 826 } 827 xpc_msgqueue_deref(ch); 828 xpc_part_deref(part); 829 830 if (atomic_read(&ch->kthreads_assigned) < 831 ch->kthreads_idle_limit) { 832 /* 833 * Flag this as an error only if we have an 834 * insufficient #of kthreads for the channel 835 * to function. 836 */ 837 spin_lock_irqsave(&ch->lock, irq_flags); 838 XPC_DISCONNECT_CHANNEL(ch, xpLackOfResources, 839 &irq_flags); 840 spin_unlock_irqrestore(&ch->lock, irq_flags); 841 } 842 break; 843 } 844 } 845} 846 847void 848xpc_disconnect_wait(int ch_number) 849{ 850 unsigned long irq_flags; 851 short partid; 852 struct xpc_partition *part; 853 struct xpc_channel *ch; 854 int wakeup_channel_mgr; 855 856 /* now wait for all callouts to the caller's function to cease */ 857 for (partid = 0; partid < xp_max_npartitions; partid++) { 858 part = &xpc_partitions[partid]; 859 860 if (!xpc_part_ref(part)) 861 continue; 862 863 ch = &part->channels[ch_number]; 864 865 if (!(ch->flags & XPC_C_WDISCONNECT)) { 866 xpc_part_deref(part); 867 continue; 868 } 869 870 wait_for_completion(&ch->wdisconnect_wait); 871 872 spin_lock_irqsave(&ch->lock, irq_flags); 873 DBUG_ON(!(ch->flags & XPC_C_DISCONNECTED)); 874 wakeup_channel_mgr = 0; 875 876 if (ch->delayed_chctl_flags) { 877 if (part->act_state != XPC_P_AS_DEACTIVATING) { 878 spin_lock(&part->chctl_lock); 879 part->chctl.flags[ch->number] |= 880 ch->delayed_chctl_flags; 881 spin_unlock(&part->chctl_lock); 882 wakeup_channel_mgr = 1; 883 } 884 ch->delayed_chctl_flags = 0; 885 } 886 887 ch->flags &= ~XPC_C_WDISCONNECT; 888 spin_unlock_irqrestore(&ch->lock, irq_flags); 889 890 if (wakeup_channel_mgr) 891 xpc_wakeup_channel_mgr(part); 892 893 xpc_part_deref(part); 894 } 895} 896 897static int 898xpc_setup_partitions(void) 899{ 900 short partid; 901 struct xpc_partition *part; 902 903 xpc_partitions = kcalloc(xp_max_npartitions, 904 sizeof(struct xpc_partition), 905 GFP_KERNEL); 906 if (xpc_partitions == NULL) { 907 dev_err(xpc_part, "can't get memory for partition structure\n"); 908 return -ENOMEM; 909 } 910 911 /* 912 * The first few fields of each entry of xpc_partitions[] need to 913 * be initialized now so that calls to xpc_connect() and 914 * xpc_disconnect() can be made prior to the activation of any remote 915 * partition. NOTE THAT NONE OF THE OTHER FIELDS BELONGING TO THESE 916 * ENTRIES ARE MEANINGFUL UNTIL AFTER AN ENTRY'S CORRESPONDING 917 * PARTITION HAS BEEN ACTIVATED. 918 */ 919 for (partid = 0; partid < xp_max_npartitions; partid++) { 920 part = &xpc_partitions[partid]; 921 922 DBUG_ON((u64)part != L1_CACHE_ALIGN((u64)part)); 923 924 part->activate_IRQ_rcvd = 0; 925 spin_lock_init(&part->act_lock); 926 part->act_state = XPC_P_AS_INACTIVE; 927 XPC_SET_REASON(part, 0, 0); 928 929 timer_setup(&part->disengage_timer, 930 xpc_timeout_partition_disengage, 0); 931 932 part->setup_state = XPC_P_SS_UNSET; 933 init_waitqueue_head(&part->teardown_wq); 934 atomic_set(&part->references, 0); 935 } 936 937 return xpc_arch_ops.setup_partitions(); 938} 939 940static void 941xpc_teardown_partitions(void) 942{ 943 xpc_arch_ops.teardown_partitions(); 944 kfree(xpc_partitions); 945} 946 947static void 948xpc_do_exit(enum xp_retval reason) 949{ 950 short partid; 951 int active_part_count, printed_waiting_msg = 0; 952 struct xpc_partition *part; 953 unsigned long printmsg_time, disengage_timeout = 0; 954 955 /* a 'rmmod XPC' and a 'reboot' cannot both end up here together */ 956 DBUG_ON(xpc_exiting == 1); 957 958 /* 959 * Let the heartbeat checker thread and the discovery thread 960 * (if one is running) know that they should exit. Also wake up 961 * the heartbeat checker thread in case it's sleeping. 962 */ 963 xpc_exiting = 1; 964 wake_up_interruptible(&xpc_activate_IRQ_wq); 965 966 /* wait for the discovery thread to exit */ 967 wait_for_completion(&xpc_discovery_exited); 968 969 /* wait for the heartbeat checker thread to exit */ 970 wait_for_completion(&xpc_hb_checker_exited); 971 972 /* sleep for a 1/3 of a second or so */ 973 (void)msleep_interruptible(300); 974 975 /* wait for all partitions to become inactive */ 976 977 printmsg_time = jiffies + (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ); 978 xpc_disengage_timedout = 0; 979 980 do { 981 active_part_count = 0; 982 983 for (partid = 0; partid < xp_max_npartitions; partid++) { 984 part = &xpc_partitions[partid]; 985 986 if (xpc_partition_disengaged(part) && 987 part->act_state == XPC_P_AS_INACTIVE) { 988 continue; 989 } 990 991 active_part_count++; 992 993 XPC_DEACTIVATE_PARTITION(part, reason); 994 995 if (part->disengage_timeout > disengage_timeout) 996 disengage_timeout = part->disengage_timeout; 997 } 998 999 if (xpc_arch_ops.any_partition_engaged()) { 1000 if (time_is_before_jiffies(printmsg_time)) { 1001 dev_info(xpc_part, "waiting for remote " 1002 "partitions to deactivate, timeout in " 1003 "%ld seconds\n", (disengage_timeout - 1004 jiffies) / HZ); 1005 printmsg_time = jiffies + 1006 (XPC_DEACTIVATE_PRINTMSG_INTERVAL * HZ); 1007 printed_waiting_msg = 1; 1008 } 1009 1010 } else if (active_part_count > 0) { 1011 if (printed_waiting_msg) { 1012 dev_info(xpc_part, "waiting for local partition" 1013 " to deactivate\n"); 1014 printed_waiting_msg = 0; 1015 } 1016 1017 } else { 1018 if (!xpc_disengage_timedout) { 1019 dev_info(xpc_part, "all partitions have " 1020 "deactivated\n"); 1021 } 1022 break; 1023 } 1024 1025 /* sleep for a 1/3 of a second or so */ 1026 (void)msleep_interruptible(300); 1027 1028 } while (1); 1029 1030 DBUG_ON(xpc_arch_ops.any_partition_engaged()); 1031 1032 xpc_teardown_rsvd_page(); 1033 1034 if (reason == xpUnloading) { 1035 (void)unregister_die_notifier(&xpc_die_notifier); 1036 (void)unregister_reboot_notifier(&xpc_reboot_notifier); 1037 } 1038 1039 /* clear the interface to XPC's functions */ 1040 xpc_clear_interface(); 1041 1042 if (xpc_sysctl) 1043 unregister_sysctl_table(xpc_sysctl); 1044 1045 xpc_teardown_partitions(); 1046 1047 if (is_uv_system()) 1048 xpc_exit_uv(); 1049} 1050 1051/* 1052 * This function is called when the system is being rebooted. 1053 */ 1054static int 1055xpc_system_reboot(struct notifier_block *nb, unsigned long event, void *unused) 1056{ 1057 enum xp_retval reason; 1058 1059 switch (event) { 1060 case SYS_RESTART: 1061 reason = xpSystemReboot; 1062 break; 1063 case SYS_HALT: 1064 reason = xpSystemHalt; 1065 break; 1066 case SYS_POWER_OFF: 1067 reason = xpSystemPoweroff; 1068 break; 1069 default: 1070 reason = xpSystemGoingDown; 1071 } 1072 1073 xpc_do_exit(reason); 1074 return NOTIFY_DONE; 1075} 1076 1077/* Used to only allow one cpu to complete disconnect */ 1078static unsigned int xpc_die_disconnecting; 1079 1080/* 1081 * Notify other partitions to deactivate from us by first disengaging from all 1082 * references to our memory. 1083 */ 1084static void 1085xpc_die_deactivate(void) 1086{ 1087 struct xpc_partition *part; 1088 short partid; 1089 int any_engaged; 1090 long keep_waiting; 1091 long wait_to_print; 1092 1093 if (cmpxchg(&xpc_die_disconnecting, 0, 1)) 1094 return; 1095 1096 /* keep xpc_hb_checker thread from doing anything (just in case) */ 1097 xpc_exiting = 1; 1098 1099 xpc_arch_ops.disallow_all_hbs(); /*indicate we're deactivated */ 1100 1101 for (partid = 0; partid < xp_max_npartitions; partid++) { 1102 part = &xpc_partitions[partid]; 1103 1104 if (xpc_arch_ops.partition_engaged(partid) || 1105 part->act_state != XPC_P_AS_INACTIVE) { 1106 xpc_arch_ops.request_partition_deactivation(part); 1107 xpc_arch_ops.indicate_partition_disengaged(part); 1108 } 1109 } 1110 1111 /* 1112 * Though we requested that all other partitions deactivate from us, 1113 * we only wait until they've all disengaged or we've reached the 1114 * defined timelimit. 1115 * 1116 * Given that one iteration through the following while-loop takes 1117 * approximately 200 microseconds, calculate the #of loops to take 1118 * before bailing and the #of loops before printing a waiting message. 1119 */ 1120 keep_waiting = xpc_disengage_timelimit * 1000 * 5; 1121 wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL * 1000 * 5; 1122 1123 while (1) { 1124 any_engaged = xpc_arch_ops.any_partition_engaged(); 1125 if (!any_engaged) { 1126 dev_info(xpc_part, "all partitions have deactivated\n"); 1127 break; 1128 } 1129 1130 if (!keep_waiting--) { 1131 for (partid = 0; partid < xp_max_npartitions; 1132 partid++) { 1133 if (xpc_arch_ops.partition_engaged(partid)) { 1134 dev_info(xpc_part, "deactivate from " 1135 "remote partition %d timed " 1136 "out\n", partid); 1137 } 1138 } 1139 break; 1140 } 1141 1142 if (!wait_to_print--) { 1143 dev_info(xpc_part, "waiting for remote partitions to " 1144 "deactivate, timeout in %ld seconds\n", 1145 keep_waiting / (1000 * 5)); 1146 wait_to_print = XPC_DEACTIVATE_PRINTMSG_INTERVAL * 1147 1000 * 5; 1148 } 1149 1150 udelay(200); 1151 } 1152} 1153 1154/* 1155 * This function is called when the system is being restarted or halted due 1156 * to some sort of system failure. If this is the case we need to notify the 1157 * other partitions to disengage from all references to our memory. 1158 * This function can also be called when our heartbeater could be offlined 1159 * for a time. In this case we need to notify other partitions to not worry 1160 * about the lack of a heartbeat. 1161 */ 1162static int 1163xpc_system_die(struct notifier_block *nb, unsigned long event, void *_die_args) 1164{ 1165#ifdef CONFIG_IA64 /* !!! temporary kludge */ 1166 switch (event) { 1167 case DIE_MACHINE_RESTART: 1168 case DIE_MACHINE_HALT: 1169 xpc_die_deactivate(); 1170 break; 1171 1172 case DIE_KDEBUG_ENTER: 1173 /* Should lack of heartbeat be ignored by other partitions? */ 1174 if (!xpc_kdebug_ignore) 1175 break; 1176 1177 fallthrough; 1178 case DIE_MCA_MONARCH_ENTER: 1179 case DIE_INIT_MONARCH_ENTER: 1180 xpc_arch_ops.offline_heartbeat(); 1181 break; 1182 1183 case DIE_KDEBUG_LEAVE: 1184 /* Is lack of heartbeat being ignored by other partitions? */ 1185 if (!xpc_kdebug_ignore) 1186 break; 1187 1188 fallthrough; 1189 case DIE_MCA_MONARCH_LEAVE: 1190 case DIE_INIT_MONARCH_LEAVE: 1191 xpc_arch_ops.online_heartbeat(); 1192 break; 1193 } 1194#else 1195 struct die_args *die_args = _die_args; 1196 1197 switch (event) { 1198 case DIE_TRAP: 1199 if (die_args->trapnr == X86_TRAP_DF) 1200 xpc_die_deactivate(); 1201 1202 if (((die_args->trapnr == X86_TRAP_MF) || 1203 (die_args->trapnr == X86_TRAP_XF)) && 1204 !user_mode(die_args->regs)) 1205 xpc_die_deactivate(); 1206 1207 break; 1208 case DIE_INT3: 1209 case DIE_DEBUG: 1210 break; 1211 case DIE_OOPS: 1212 case DIE_GPF: 1213 default: 1214 xpc_die_deactivate(); 1215 } 1216#endif 1217 1218 return NOTIFY_DONE; 1219} 1220 1221static int __init 1222xpc_init(void) 1223{ 1224 int ret; 1225 struct task_struct *kthread; 1226 1227 dev_set_name(xpc_part, "part"); 1228 dev_set_name(xpc_chan, "chan"); 1229 1230 if (is_uv_system()) { 1231 ret = xpc_init_uv(); 1232 1233 } else { 1234 ret = -ENODEV; 1235 } 1236 1237 if (ret != 0) 1238 return ret; 1239 1240 ret = xpc_setup_partitions(); 1241 if (ret != 0) { 1242 dev_err(xpc_part, "can't get memory for partition structure\n"); 1243 goto out_1; 1244 } 1245 1246 xpc_sysctl = register_sysctl_table(xpc_sys_dir); 1247 1248 /* 1249 * Fill the partition reserved page with the information needed by 1250 * other partitions to discover we are alive and establish initial 1251 * communications. 1252 */ 1253 ret = xpc_setup_rsvd_page(); 1254 if (ret != 0) { 1255 dev_err(xpc_part, "can't setup our reserved page\n"); 1256 goto out_2; 1257 } 1258 1259 /* add ourselves to the reboot_notifier_list */ 1260 ret = register_reboot_notifier(&xpc_reboot_notifier); 1261 if (ret != 0) 1262 dev_warn(xpc_part, "can't register reboot notifier\n"); 1263 1264 /* add ourselves to the die_notifier list */ 1265 ret = register_die_notifier(&xpc_die_notifier); 1266 if (ret != 0) 1267 dev_warn(xpc_part, "can't register die notifier\n"); 1268 1269 /* 1270 * The real work-horse behind xpc. This processes incoming 1271 * interrupts and monitors remote heartbeats. 1272 */ 1273 kthread = kthread_run(xpc_hb_checker, NULL, XPC_HB_CHECK_THREAD_NAME); 1274 if (IS_ERR(kthread)) { 1275 dev_err(xpc_part, "failed while forking hb check thread\n"); 1276 ret = -EBUSY; 1277 goto out_3; 1278 } 1279 1280 /* 1281 * Startup a thread that will attempt to discover other partitions to 1282 * activate based on info provided by SAL. This new thread is short 1283 * lived and will exit once discovery is complete. 1284 */ 1285 kthread = kthread_run(xpc_initiate_discovery, NULL, 1286 XPC_DISCOVERY_THREAD_NAME); 1287 if (IS_ERR(kthread)) { 1288 dev_err(xpc_part, "failed while forking discovery thread\n"); 1289 1290 /* mark this new thread as a non-starter */ 1291 complete(&xpc_discovery_exited); 1292 1293 xpc_do_exit(xpUnloading); 1294 return -EBUSY; 1295 } 1296 1297 /* set the interface to point at XPC's functions */ 1298 xpc_set_interface(xpc_initiate_connect, xpc_initiate_disconnect, 1299 xpc_initiate_send, xpc_initiate_send_notify, 1300 xpc_initiate_received, xpc_initiate_partid_to_nasids); 1301 1302 return 0; 1303 1304 /* initialization was not successful */ 1305out_3: 1306 xpc_teardown_rsvd_page(); 1307 1308 (void)unregister_die_notifier(&xpc_die_notifier); 1309 (void)unregister_reboot_notifier(&xpc_reboot_notifier); 1310out_2: 1311 if (xpc_sysctl) 1312 unregister_sysctl_table(xpc_sysctl); 1313 1314 xpc_teardown_partitions(); 1315out_1: 1316 if (is_uv_system()) 1317 xpc_exit_uv(); 1318 return ret; 1319} 1320 1321module_init(xpc_init); 1322 1323static void __exit 1324xpc_exit(void) 1325{ 1326 xpc_do_exit(xpUnloading); 1327} 1328 1329module_exit(xpc_exit); 1330 1331MODULE_AUTHOR("Silicon Graphics, Inc."); 1332MODULE_DESCRIPTION("Cross Partition Communication (XPC) support"); 1333MODULE_LICENSE("GPL"); 1334 1335module_param(xpc_hb_interval, int, 0); 1336MODULE_PARM_DESC(xpc_hb_interval, "Number of seconds between " 1337 "heartbeat increments."); 1338 1339module_param(xpc_hb_check_interval, int, 0); 1340MODULE_PARM_DESC(xpc_hb_check_interval, "Number of seconds between " 1341 "heartbeat checks."); 1342 1343module_param(xpc_disengage_timelimit, int, 0); 1344MODULE_PARM_DESC(xpc_disengage_timelimit, "Number of seconds to wait " 1345 "for disengage to complete."); 1346 1347module_param(xpc_kdebug_ignore, int, 0); 1348MODULE_PARM_DESC(xpc_kdebug_ignore, "Should lack of heartbeat be ignored by " 1349 "other partitions when dropping into kdebug."); 1350