Linux Audio

Check our new training course

Loading...
v3.1
   1/*
   2 * Copyright (C) 2003 Sistina Software Limited.
   3 * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4 *
   5 * This file is released under the GPL.
   6 */
   7
   8#include <linux/device-mapper.h>
   9
 
 
  10#include "dm-path-selector.h"
  11#include "dm-uevent.h"
  12
 
  13#include <linux/ctype.h>
  14#include <linux/init.h>
  15#include <linux/mempool.h>
  16#include <linux/module.h>
  17#include <linux/pagemap.h>
  18#include <linux/slab.h>
  19#include <linux/time.h>
 
  20#include <linux/workqueue.h>
 
  21#include <scsi/scsi_dh.h>
  22#include <linux/atomic.h>
 
  23
  24#define DM_MSG_PREFIX "multipath"
  25#define DM_PG_INIT_DELAY_MSECS 2000
  26#define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
 
 
 
  27
  28/* Path properties */
  29struct pgpath {
  30	struct list_head list;
  31
  32	struct priority_group *pg;	/* Owning PG */
  33	unsigned is_active;		/* Path status */
  34	unsigned fail_count;		/* Cumulative failure count */
  35
  36	struct dm_path path;
  37	struct delayed_work activate_path;
 
 
  38};
  39
  40#define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  41
  42/*
  43 * Paths are grouped into Priority Groups and numbered from 1 upwards.
  44 * Each has a path selector which controls which path gets used.
  45 */
  46struct priority_group {
  47	struct list_head list;
  48
  49	struct multipath *m;		/* Owning multipath instance */
  50	struct path_selector ps;
  51
  52	unsigned pg_num;		/* Reference number */
  53	unsigned bypassed;		/* Temporarily bypass this PG? */
  54
  55	unsigned nr_pgpaths;		/* Number of paths in PG */
  56	struct list_head pgpaths;
 
 
  57};
  58
  59/* Multipath context */
  60struct multipath {
  61	struct list_head list;
  62	struct dm_target *ti;
  63
  64	spinlock_t lock;
 
  65
  66	const char *hw_handler_name;
  67	char *hw_handler_params;
 
  68
 
  69	unsigned nr_priority_groups;
  70	struct list_head priority_groups;
  71
 
 
  72	wait_queue_head_t pg_init_wait;	/* Wait for pg_init completion */
  73
  74	unsigned pg_init_required;	/* pg_init needs calling? */
  75	unsigned pg_init_in_progress;	/* Only one pg_init allowed at once */
  76	unsigned pg_init_delay_retry;	/* Delay pg_init retry? */
  77
  78	unsigned nr_valid_paths;	/* Total number of usable paths */
  79	struct pgpath *current_pgpath;
  80	struct priority_group *current_pg;
  81	struct priority_group *next_pg;	/* Switch to this PG if set */
  82	unsigned repeat_count;		/* I/Os left before calling PS again */
  83
  84	unsigned queue_io;		/* Must we queue all I/O? */
  85	unsigned queue_if_no_path;	/* Queue I/O if last path fails? */
  86	unsigned saved_queue_if_no_path;/* Saved state during suspension */
  87	unsigned pg_init_retries;	/* Number of times to retry pg_init */
  88	unsigned pg_init_count;		/* Number of times pg_init called */
  89	unsigned pg_init_delay_msecs;	/* Number of msecs before pg_init retry */
 
 
  90
  91	struct work_struct process_queued_ios;
  92	struct list_head queued_ios;
  93	unsigned queue_size;
  94
  95	struct work_struct trigger_event;
 
  96
  97	/*
  98	 * We must use a mempool of dm_mpath_io structs so that we
  99	 * can resubmit bios on error.
 100	 */
 101	mempool_t *mpio_pool;
 102
 103	struct mutex work_mutex;
 104};
 105
 106/*
 107 * Context information attached to each bio we process.
 108 */
 109struct dm_mpath_io {
 110	struct pgpath *pgpath;
 111	size_t nr_bytes;
 112};
 113
 114typedef int (*action_fn) (struct pgpath *pgpath);
 115
 116#define MIN_IOS 256	/* Mempool size */
 117
 118static struct kmem_cache *_mpio_cache;
 119
 120static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 121static void process_queued_ios(struct work_struct *work);
 122static void trigger_event(struct work_struct *work);
 123static void activate_path(struct work_struct *work);
 
 
 
 124
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 125
 126/*-----------------------------------------------
 127 * Allocation routines
 128 *-----------------------------------------------*/
 129
 130static struct pgpath *alloc_pgpath(void)
 131{
 132	struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 133
 134	if (pgpath) {
 135		pgpath->is_active = 1;
 136		INIT_DELAYED_WORK(&pgpath->activate_path, activate_path);
 137	}
 138
 139	return pgpath;
 140}
 141
 142static void free_pgpath(struct pgpath *pgpath)
 143{
 144	kfree(pgpath);
 145}
 146
 147static struct priority_group *alloc_priority_group(void)
 148{
 149	struct priority_group *pg;
 150
 151	pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 152
 153	if (pg)
 154		INIT_LIST_HEAD(&pg->pgpaths);
 155
 156	return pg;
 157}
 158
 159static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 160{
 161	struct pgpath *pgpath, *tmp;
 162	struct multipath *m = ti->private;
 163
 164	list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 165		list_del(&pgpath->list);
 166		if (m->hw_handler_name)
 167			scsi_dh_detach(bdev_get_queue(pgpath->path.dev->bdev));
 168		dm_put_device(ti, pgpath->path.dev);
 169		free_pgpath(pgpath);
 170	}
 171}
 172
 173static void free_priority_group(struct priority_group *pg,
 174				struct dm_target *ti)
 175{
 176	struct path_selector *ps = &pg->ps;
 177
 178	if (ps->type) {
 179		ps->type->destroy(ps);
 180		dm_put_path_selector(ps->type);
 181	}
 182
 183	free_pgpaths(&pg->pgpaths, ti);
 184	kfree(pg);
 185}
 186
 187static struct multipath *alloc_multipath(struct dm_target *ti)
 188{
 189	struct multipath *m;
 190
 191	m = kzalloc(sizeof(*m), GFP_KERNEL);
 192	if (m) {
 193		INIT_LIST_HEAD(&m->priority_groups);
 194		INIT_LIST_HEAD(&m->queued_ios);
 195		spin_lock_init(&m->lock);
 196		m->queue_io = 1;
 197		m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 198		INIT_WORK(&m->process_queued_ios, process_queued_ios);
 199		INIT_WORK(&m->trigger_event, trigger_event);
 200		init_waitqueue_head(&m->pg_init_wait);
 201		mutex_init(&m->work_mutex);
 202		m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 203		if (!m->mpio_pool) {
 204			kfree(m);
 205			return NULL;
 206		}
 207		m->ti = ti;
 208		ti->private = m;
 
 
 209	}
 210
 211	return m;
 212}
 213
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 214static void free_multipath(struct multipath *m)
 215{
 216	struct priority_group *pg, *tmp;
 217
 218	list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 219		list_del(&pg->list);
 220		free_priority_group(pg, m->ti);
 221	}
 222
 223	kfree(m->hw_handler_name);
 224	kfree(m->hw_handler_params);
 225	mempool_destroy(m->mpio_pool);
 226	kfree(m);
 227}
 228
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 229
 230/*-----------------------------------------------
 231 * Path selection
 232 *-----------------------------------------------*/
 233
 234static void __pg_init_all_paths(struct multipath *m)
 235{
 236	struct pgpath *pgpath;
 237	unsigned long pg_init_delay = 0;
 238
 239	m->pg_init_count++;
 240	m->pg_init_required = 0;
 241	if (m->pg_init_delay_retry)
 
 
 
 
 
 
 
 
 
 
 242		pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 243						 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 244	list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 245		/* Skip failed paths */
 246		if (!pgpath->is_active)
 247			continue;
 248		if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 249				       pg_init_delay))
 250			m->pg_init_in_progress++;
 251	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 252}
 253
 254static void __switch_pg(struct multipath *m, struct pgpath *pgpath)
 255{
 256	m->current_pg = pgpath->pg;
 
 
 257
 258	/* Must we initialise the PG first, and queue I/O till it's ready? */
 259	if (m->hw_handler_name) {
 260		m->pg_init_required = 1;
 261		m->queue_io = 1;
 262	} else {
 263		m->pg_init_required = 0;
 264		m->queue_io = 0;
 265	}
 266
 267	m->pg_init_count = 0;
 268}
 269
 270static int __choose_path_in_pg(struct multipath *m, struct priority_group *pg,
 271			       size_t nr_bytes)
 
 272{
 
 273	struct dm_path *path;
 
 274
 275	path = pg->ps.type->select_path(&pg->ps, &m->repeat_count, nr_bytes);
 276	if (!path)
 277		return -ENXIO;
 278
 279	m->current_pgpath = path_to_pgpath(path);
 280
 281	if (m->current_pg != pg)
 282		__switch_pg(m, m->current_pgpath);
 
 
 
 
 
 283
 284	return 0;
 285}
 286
 287static void __choose_pgpath(struct multipath *m, size_t nr_bytes)
 288{
 
 289	struct priority_group *pg;
 
 290	unsigned bypassed = 1;
 291
 292	if (!m->nr_valid_paths)
 
 
 
 293		goto failed;
 
 294
 295	/* Were we instructed to switch PG? */
 296	if (m->next_pg) {
 
 297		pg = m->next_pg;
 
 
 
 
 298		m->next_pg = NULL;
 299		if (!__choose_path_in_pg(m, pg, nr_bytes))
 300			return;
 
 
 301	}
 302
 303	/* Don't change PG until it has no remaining paths */
 304	if (m->current_pg && !__choose_path_in_pg(m, m->current_pg, nr_bytes))
 305		return;
 
 
 
 
 
 306
 307	/*
 308	 * Loop through priority groups until we find a valid path.
 309	 * First time we skip PGs marked 'bypassed'.
 310	 * Second time we only try the ones we skipped.
 
 311	 */
 312	do {
 313		list_for_each_entry(pg, &m->priority_groups, list) {
 314			if (pg->bypassed == bypassed)
 315				continue;
 316			if (!__choose_path_in_pg(m, pg, nr_bytes))
 317				return;
 
 
 
 
 
 
 
 318		}
 319	} while (bypassed--);
 320
 321failed:
 
 322	m->current_pgpath = NULL;
 323	m->current_pg = NULL;
 
 
 
 324}
 325
 326/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 327 * Check whether bios must be queued in the device-mapper core rather
 328 * than here in the target.
 329 *
 330 * m->lock must be held on entry.
 331 *
 332 * If m->queue_if_no_path and m->saved_queue_if_no_path hold the
 333 * same value then we are not between multipath_presuspend()
 334 * and multipath_resume() calls and we have no need to check
 335 * for the DMF_NOFLUSH_SUSPENDING flag.
 336 */
 337static int __must_push_back(struct multipath *m)
 338{
 339	return (m->queue_if_no_path != m->saved_queue_if_no_path &&
 340		dm_noflush_suspending(m->ti));
 341}
 342
 343static int map_io(struct multipath *m, struct request *clone,
 344		  struct dm_mpath_io *mpio, unsigned was_queued)
 345{
 346	int r = DM_MAPIO_REMAPPED;
 347	size_t nr_bytes = blk_rq_bytes(clone);
 348	unsigned long flags;
 349	struct pgpath *pgpath;
 350	struct block_device *bdev;
 351
 352	spin_lock_irqsave(&m->lock, flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 353
 354	/* Do we need to select a new pgpath? */
 355	if (!m->current_pgpath ||
 356	    (!m->queue_io && (m->repeat_count && --m->repeat_count == 0)))
 357		__choose_pgpath(m, nr_bytes);
 358
 359	pgpath = m->current_pgpath;
 360
 361	if (was_queued)
 362		m->queue_size--;
 363
 364	if ((pgpath && m->queue_io) ||
 365	    (!pgpath && m->queue_if_no_path)) {
 366		/* Queue for the daemon to resubmit */
 367		list_add_tail(&clone->queuelist, &m->queued_ios);
 368		m->queue_size++;
 369		if ((m->pg_init_required && !m->pg_init_in_progress) ||
 370		    !m->queue_io)
 371			queue_work(kmultipathd, &m->process_queued_ios);
 372		pgpath = NULL;
 373		r = DM_MAPIO_SUBMITTED;
 374	} else if (pgpath) {
 375		bdev = pgpath->path.dev->bdev;
 376		clone->q = bdev_get_queue(bdev);
 377		clone->rq_disk = bdev->bd_disk;
 378	} else if (__must_push_back(m))
 379		r = DM_MAPIO_REQUEUE;
 380	else
 381		r = -EIO;	/* Failed */
 382
 383	mpio->pgpath = pgpath;
 384	mpio->nr_bytes = nr_bytes;
 385
 386	if (r == DM_MAPIO_REMAPPED && pgpath->pg->ps.type->start_io)
 387		pgpath->pg->ps.type->start_io(&pgpath->pg->ps, &pgpath->path,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 388					      nr_bytes);
 
 
 389
 390	spin_unlock_irqrestore(&m->lock, flags);
 
 
 
 
 
 
 
 
 
 391
 392	return r;
 
 
 
 
 
 
 
 393}
 394
 395/*
 396 * If we run out of usable paths, should we queue I/O or error it?
 397 */
 398static int queue_if_no_path(struct multipath *m, unsigned queue_if_no_path,
 399			    unsigned save_old_value)
 
 
 
 
 
 
 
 
 400{
 401	unsigned long flags;
 402
 403	spin_lock_irqsave(&m->lock, flags);
 
 
 
 404
 405	if (save_old_value)
 406		m->saved_queue_if_no_path = m->queue_if_no_path;
 407	else
 408		m->saved_queue_if_no_path = queue_if_no_path;
 409	m->queue_if_no_path = queue_if_no_path;
 410	if (!m->queue_if_no_path && m->queue_size)
 411		queue_work(kmultipathd, &m->process_queued_ios);
 412
 413	spin_unlock_irqrestore(&m->lock, flags);
 
 
 
 414
 415	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 416}
 417
 418/*-----------------------------------------------------------------
 419 * The multipath daemon is responsible for resubmitting queued ios.
 420 *---------------------------------------------------------------*/
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 421
 422static void dispatch_queued_ios(struct multipath *m)
 423{
 424	int r;
 425	unsigned long flags;
 426	struct dm_mpath_io *mpio;
 427	union map_info *info;
 428	struct request *clone, *n;
 429	LIST_HEAD(cl);
 
 
 
 430
 431	spin_lock_irqsave(&m->lock, flags);
 432	list_splice_init(&m->queued_ios, &cl);
 433	spin_unlock_irqrestore(&m->lock, flags);
 434
 435	list_for_each_entry_safe(clone, n, &cl, queuelist) {
 436		list_del_init(&clone->queuelist);
 
 
 437
 438		info = dm_get_rq_mapinfo(clone);
 439		mpio = info->ptr;
 440
 441		r = map_io(m, clone, mpio, 1);
 442		if (r < 0) {
 443			mempool_free(mpio, m->mpio_pool);
 444			dm_kill_unmapped_request(clone, r);
 445		} else if (r == DM_MAPIO_REMAPPED)
 446			dm_dispatch_request(clone);
 447		else if (r == DM_MAPIO_REQUEUE) {
 448			mempool_free(mpio, m->mpio_pool);
 449			dm_requeue_unmapped_request(clone);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 450		}
 451	}
 
 452}
 453
 454static void process_queued_ios(struct work_struct *work)
 
 
 
 
 455{
 456	struct multipath *m =
 457		container_of(work, struct multipath, process_queued_ios);
 458	struct pgpath *pgpath = NULL;
 459	unsigned must_queue = 1;
 460	unsigned long flags;
 
 
 
 
 
 461
 462	spin_lock_irqsave(&m->lock, flags);
 463
 464	if (!m->queue_size)
 465		goto out;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 466
 467	if (!m->current_pgpath)
 468		__choose_pgpath(m, 0);
 469
 470	pgpath = m->current_pgpath;
 
 
 
 471
 472	if ((pgpath && !m->queue_io) ||
 473	    (!pgpath && !m->queue_if_no_path))
 474		must_queue = 0;
 475
 476	if (m->pg_init_required && !m->pg_init_in_progress && pgpath)
 477		__pg_init_all_paths(m);
 
 
 
 
 
 478
 479out:
 480	spin_unlock_irqrestore(&m->lock, flags);
 481	if (!must_queue)
 482		dispatch_queued_ios(m);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 483}
 484
 485/*
 486 * An event is triggered whenever a path is taken out of use.
 487 * Includes path failure and PG bypass.
 488 */
 489static void trigger_event(struct work_struct *work)
 490{
 491	struct multipath *m =
 492		container_of(work, struct multipath, trigger_event);
 493
 494	dm_table_event(m->ti->table);
 495}
 496
 497/*-----------------------------------------------------------------
 498 * Constructor/argument parsing:
 499 * <#multipath feature args> [<arg>]*
 500 * <#hw_handler args> [hw_handler [<arg>]*]
 501 * <#priority groups>
 502 * <initial priority group>
 503 *     [<selector> <#selector args> [<arg>]*
 504 *      <#paths> <#per-path selector args>
 505 *         [<path> [<arg>]* ]+ ]+
 506 *---------------------------------------------------------------*/
 507static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 508			       struct dm_target *ti)
 509{
 510	int r;
 511	struct path_selector_type *pst;
 512	unsigned ps_argc;
 513
 514	static struct dm_arg _args[] = {
 515		{0, 1024, "invalid number of path selector args"},
 516	};
 517
 518	pst = dm_get_path_selector(dm_shift_arg(as));
 519	if (!pst) {
 520		ti->error = "unknown path selector type";
 521		return -EINVAL;
 522	}
 523
 524	r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 525	if (r) {
 526		dm_put_path_selector(pst);
 527		return -EINVAL;
 528	}
 529
 530	r = pst->create(&pg->ps, ps_argc, as->argv);
 531	if (r) {
 532		dm_put_path_selector(pst);
 533		ti->error = "path selector constructor failed";
 534		return r;
 535	}
 536
 537	pg->ps.type = pst;
 538	dm_consume_args(as, ps_argc);
 539
 540	return 0;
 541}
 542
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 543static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 544			       struct dm_target *ti)
 545{
 546	int r;
 547	struct pgpath *p;
 548	struct multipath *m = ti->private;
 
 
 549
 550	/* we need at least a path arg */
 551	if (as->argc < 1) {
 552		ti->error = "no device given";
 553		return ERR_PTR(-EINVAL);
 554	}
 555
 556	p = alloc_pgpath();
 557	if (!p)
 558		return ERR_PTR(-ENOMEM);
 559
 560	r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 561			  &p->path.dev);
 562	if (r) {
 563		ti->error = "error getting device";
 564		goto bad;
 565	}
 566
 567	if (m->hw_handler_name) {
 568		struct request_queue *q = bdev_get_queue(p->path.dev->bdev);
 569
 570		r = scsi_dh_attach(q, m->hw_handler_name);
 571		if (r == -EBUSY) {
 572			/*
 573			 * Already attached to different hw_handler,
 574			 * try to reattach with correct one.
 575			 */
 576			scsi_dh_detach(q);
 577			r = scsi_dh_attach(q, m->hw_handler_name);
 578		}
 579
 580		if (r < 0) {
 581			ti->error = "error attaching hardware handler";
 582			dm_put_device(ti, p->path.dev);
 583			goto bad;
 584		}
 585
 586		if (m->hw_handler_params) {
 587			r = scsi_dh_set_params(q, m->hw_handler_params);
 588			if (r < 0) {
 589				ti->error = "unable to set hardware "
 590							"handler parameters";
 591				scsi_dh_detach(q);
 592				dm_put_device(ti, p->path.dev);
 593				goto bad;
 594			}
 595		}
 596	}
 597
 598	r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 599	if (r) {
 600		dm_put_device(ti, p->path.dev);
 601		goto bad;
 602	}
 603
 604	return p;
 605
 606 bad:
 607	free_pgpath(p);
 608	return ERR_PTR(r);
 609}
 610
 611static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 612						   struct multipath *m)
 613{
 614	static struct dm_arg _args[] = {
 615		{1, 1024, "invalid number of paths"},
 616		{0, 1024, "invalid number of selector args"}
 617	};
 618
 619	int r;
 620	unsigned i, nr_selector_args, nr_args;
 621	struct priority_group *pg;
 622	struct dm_target *ti = m->ti;
 623
 624	if (as->argc < 2) {
 625		as->argc = 0;
 626		ti->error = "not enough priority group arguments";
 627		return ERR_PTR(-EINVAL);
 628	}
 629
 630	pg = alloc_priority_group();
 631	if (!pg) {
 632		ti->error = "couldn't allocate priority group";
 633		return ERR_PTR(-ENOMEM);
 634	}
 635	pg->m = m;
 636
 637	r = parse_path_selector(as, pg, ti);
 638	if (r)
 639		goto bad;
 640
 641	/*
 642	 * read the paths
 643	 */
 644	r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
 645	if (r)
 646		goto bad;
 647
 648	r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
 649	if (r)
 650		goto bad;
 651
 652	nr_args = 1 + nr_selector_args;
 653	for (i = 0; i < pg->nr_pgpaths; i++) {
 654		struct pgpath *pgpath;
 655		struct dm_arg_set path_args;
 656
 657		if (as->argc < nr_args) {
 658			ti->error = "not enough path parameters";
 659			r = -EINVAL;
 660			goto bad;
 661		}
 662
 663		path_args.argc = nr_args;
 664		path_args.argv = as->argv;
 665
 666		pgpath = parse_path(&path_args, &pg->ps, ti);
 667		if (IS_ERR(pgpath)) {
 668			r = PTR_ERR(pgpath);
 669			goto bad;
 670		}
 671
 672		pgpath->pg = pg;
 673		list_add_tail(&pgpath->list, &pg->pgpaths);
 674		dm_consume_args(as, nr_args);
 675	}
 676
 677	return pg;
 678
 679 bad:
 680	free_priority_group(pg, ti);
 681	return ERR_PTR(r);
 682}
 683
 684static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
 685{
 686	unsigned hw_argc;
 687	int ret;
 688	struct dm_target *ti = m->ti;
 689
 690	static struct dm_arg _args[] = {
 691		{0, 1024, "invalid number of hardware handler args"},
 692	};
 693
 694	if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
 695		return -EINVAL;
 696
 697	if (!hw_argc)
 698		return 0;
 699
 700	m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
 701	request_module("scsi_dh_%s", m->hw_handler_name);
 702	if (scsi_dh_handler_exist(m->hw_handler_name) == 0) {
 703		ti->error = "unknown hardware handler type";
 704		ret = -EINVAL;
 705		goto fail;
 706	}
 707
 
 
 
 
 708	if (hw_argc > 1) {
 709		char *p;
 710		int i, j, len = 4;
 711
 712		for (i = 0; i <= hw_argc - 2; i++)
 713			len += strlen(as->argv[i]) + 1;
 714		p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
 715		if (!p) {
 716			ti->error = "memory allocation failed";
 717			ret = -ENOMEM;
 718			goto fail;
 719		}
 720		j = sprintf(p, "%d", hw_argc - 1);
 721		for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
 722			j = sprintf(p, "%s", as->argv[i]);
 723	}
 724	dm_consume_args(as, hw_argc - 1);
 725
 726	return 0;
 727fail:
 728	kfree(m->hw_handler_name);
 729	m->hw_handler_name = NULL;
 730	return ret;
 731}
 732
 733static int parse_features(struct dm_arg_set *as, struct multipath *m)
 734{
 735	int r;
 736	unsigned argc;
 737	struct dm_target *ti = m->ti;
 738	const char *arg_name;
 739
 740	static struct dm_arg _args[] = {
 741		{0, 5, "invalid number of feature args"},
 742		{1, 50, "pg_init_retries must be between 1 and 50"},
 743		{0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
 744	};
 745
 746	r = dm_read_arg_group(_args, as, &argc, &ti->error);
 747	if (r)
 748		return -EINVAL;
 749
 750	if (!argc)
 751		return 0;
 752
 753	do {
 754		arg_name = dm_shift_arg(as);
 755		argc--;
 756
 757		if (!strcasecmp(arg_name, "queue_if_no_path")) {
 758			r = queue_if_no_path(m, 1, 0);
 
 
 
 
 
 759			continue;
 760		}
 761
 762		if (!strcasecmp(arg_name, "pg_init_retries") &&
 763		    (argc >= 1)) {
 764			r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
 765			argc--;
 766			continue;
 767		}
 768
 769		if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
 770		    (argc >= 1)) {
 771			r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
 772			argc--;
 773			continue;
 774		}
 775
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 776		ti->error = "Unrecognised multipath feature request";
 777		r = -EINVAL;
 778	} while (argc && !r);
 779
 780	return r;
 781}
 782
 783static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 784			 char **argv)
 785{
 786	/* target arguments */
 787	static struct dm_arg _args[] = {
 788		{0, 1024, "invalid number of priority groups"},
 789		{0, 1024, "invalid initial priority group number"},
 790	};
 791
 792	int r;
 793	struct multipath *m;
 794	struct dm_arg_set as;
 795	unsigned pg_count = 0;
 796	unsigned next_pg_num;
 
 797
 798	as.argc = argc;
 799	as.argv = argv;
 800
 801	m = alloc_multipath(ti);
 802	if (!m) {
 803		ti->error = "can't allocate multipath";
 804		return -EINVAL;
 805	}
 806
 807	r = parse_features(&as, m);
 808	if (r)
 809		goto bad;
 810
 
 
 
 
 811	r = parse_hw_handler(&as, m);
 812	if (r)
 813		goto bad;
 814
 815	r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
 816	if (r)
 817		goto bad;
 818
 819	r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
 820	if (r)
 821		goto bad;
 822
 823	if ((!m->nr_priority_groups && next_pg_num) ||
 824	    (m->nr_priority_groups && !next_pg_num)) {
 825		ti->error = "invalid initial priority group";
 826		r = -EINVAL;
 827		goto bad;
 828	}
 829
 830	/* parse the priority groups */
 831	while (as.argc) {
 832		struct priority_group *pg;
 
 833
 834		pg = parse_priority_group(&as, m);
 835		if (IS_ERR(pg)) {
 836			r = PTR_ERR(pg);
 837			goto bad;
 838		}
 839
 840		m->nr_valid_paths += pg->nr_pgpaths;
 
 
 841		list_add_tail(&pg->list, &m->priority_groups);
 842		pg_count++;
 843		pg->pg_num = pg_count;
 844		if (!--next_pg_num)
 845			m->next_pg = pg;
 846	}
 847
 848	if (pg_count != m->nr_priority_groups) {
 849		ti->error = "priority group count mismatch";
 850		r = -EINVAL;
 851		goto bad;
 852	}
 853
 854	ti->num_flush_requests = 1;
 855	ti->num_discard_requests = 1;
 
 
 
 
 
 
 
 
 
 
 856
 857	return 0;
 858
 859 bad:
 860	free_multipath(m);
 861	return r;
 862}
 863
 864static void multipath_wait_for_pg_init_completion(struct multipath *m)
 865{
 866	DECLARE_WAITQUEUE(wait, current);
 867	unsigned long flags;
 868
 869	add_wait_queue(&m->pg_init_wait, &wait);
 870
 871	while (1) {
 872		set_current_state(TASK_UNINTERRUPTIBLE);
 873
 874		spin_lock_irqsave(&m->lock, flags);
 875		if (!m->pg_init_in_progress) {
 876			spin_unlock_irqrestore(&m->lock, flags);
 877			break;
 878		}
 879		spin_unlock_irqrestore(&m->lock, flags);
 880
 881		io_schedule();
 882	}
 883	set_current_state(TASK_RUNNING);
 884
 885	remove_wait_queue(&m->pg_init_wait, &wait);
 886}
 887
 888static void flush_multipath_work(struct multipath *m)
 889{
 890	flush_workqueue(kmpath_handlerd);
 891	multipath_wait_for_pg_init_completion(m);
 892	flush_workqueue(kmultipathd);
 893	flush_work_sync(&m->trigger_event);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 894}
 895
 896static void multipath_dtr(struct dm_target *ti)
 897{
 898	struct multipath *m = ti->private;
 899
 
 900	flush_multipath_work(m);
 901	free_multipath(m);
 902}
 903
 904/*
 905 * Map cloned requests
 906 */
 907static int multipath_map(struct dm_target *ti, struct request *clone,
 908			 union map_info *map_context)
 909{
 910	int r;
 911	struct dm_mpath_io *mpio;
 912	struct multipath *m = (struct multipath *) ti->private;
 913
 914	mpio = mempool_alloc(m->mpio_pool, GFP_ATOMIC);
 915	if (!mpio)
 916		/* ENOMEM, requeue */
 917		return DM_MAPIO_REQUEUE;
 918	memset(mpio, 0, sizeof(*mpio));
 919
 920	map_context->ptr = mpio;
 921	clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 922	r = map_io(m, clone, mpio, 0);
 923	if (r < 0 || r == DM_MAPIO_REQUEUE)
 924		mempool_free(mpio, m->mpio_pool);
 925
 926	return r;
 927}
 928
 929/*
 930 * Take a path out of use.
 931 */
 932static int fail_path(struct pgpath *pgpath)
 933{
 934	unsigned long flags;
 935	struct multipath *m = pgpath->pg->m;
 936
 937	spin_lock_irqsave(&m->lock, flags);
 938
 939	if (!pgpath->is_active)
 940		goto out;
 941
 942	DMWARN("Failing path %s.", pgpath->path.dev->name);
 
 
 943
 944	pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
 945	pgpath->is_active = 0;
 946	pgpath->fail_count++;
 947
 948	m->nr_valid_paths--;
 949
 950	if (pgpath == m->current_pgpath)
 951		m->current_pgpath = NULL;
 952
 953	dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
 954		      pgpath->path.dev->name, m->nr_valid_paths);
 955
 956	schedule_work(&m->trigger_event);
 957
 
 
 958out:
 959	spin_unlock_irqrestore(&m->lock, flags);
 960
 961	return 0;
 962}
 963
 964/*
 965 * Reinstate a previously-failed path
 966 */
 967static int reinstate_path(struct pgpath *pgpath)
 968{
 969	int r = 0;
 970	unsigned long flags;
 971	struct multipath *m = pgpath->pg->m;
 
 972
 973	spin_lock_irqsave(&m->lock, flags);
 974
 975	if (pgpath->is_active)
 976		goto out;
 977
 978	if (!pgpath->pg->ps.type->reinstate_path) {
 979		DMWARN("Reinstate path not supported by path selector %s",
 980		       pgpath->pg->ps.type->name);
 981		r = -EINVAL;
 982		goto out;
 983	}
 984
 985	r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
 986	if (r)
 987		goto out;
 988
 989	pgpath->is_active = 1;
 990
 991	if (!m->nr_valid_paths++ && m->queue_size) {
 
 992		m->current_pgpath = NULL;
 993		queue_work(kmultipathd, &m->process_queued_ios);
 994	} else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
 995		if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
 996			m->pg_init_in_progress++;
 997	}
 998
 999	dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1000		      pgpath->path.dev->name, m->nr_valid_paths);
1001
1002	schedule_work(&m->trigger_event);
1003
1004out:
1005	spin_unlock_irqrestore(&m->lock, flags);
 
 
 
 
 
 
 
1006
1007	return r;
1008}
1009
1010/*
1011 * Fail or reinstate all paths that match the provided struct dm_dev.
1012 */
1013static int action_dev(struct multipath *m, struct dm_dev *dev,
1014		      action_fn action)
1015{
1016	int r = -EINVAL;
1017	struct pgpath *pgpath;
1018	struct priority_group *pg;
1019
1020	list_for_each_entry(pg, &m->priority_groups, list) {
1021		list_for_each_entry(pgpath, &pg->pgpaths, list) {
1022			if (pgpath->path.dev == dev)
1023				r = action(pgpath);
1024		}
1025	}
1026
1027	return r;
1028}
1029
1030/*
1031 * Temporarily try to avoid having to use the specified PG
1032 */
1033static void bypass_pg(struct multipath *m, struct priority_group *pg,
1034		      int bypassed)
1035{
1036	unsigned long flags;
1037
1038	spin_lock_irqsave(&m->lock, flags);
1039
1040	pg->bypassed = bypassed;
1041	m->current_pgpath = NULL;
1042	m->current_pg = NULL;
1043
1044	spin_unlock_irqrestore(&m->lock, flags);
1045
1046	schedule_work(&m->trigger_event);
1047}
1048
1049/*
1050 * Switch to using the specified PG from the next I/O that gets mapped
1051 */
1052static int switch_pg_num(struct multipath *m, const char *pgstr)
1053{
1054	struct priority_group *pg;
1055	unsigned pgnum;
1056	unsigned long flags;
 
1057
1058	if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1059	    (pgnum > m->nr_priority_groups)) {
1060		DMWARN("invalid PG number supplied to switch_pg_num");
1061		return -EINVAL;
1062	}
1063
1064	spin_lock_irqsave(&m->lock, flags);
1065	list_for_each_entry(pg, &m->priority_groups, list) {
1066		pg->bypassed = 0;
1067		if (--pgnum)
1068			continue;
1069
1070		m->current_pgpath = NULL;
1071		m->current_pg = NULL;
1072		m->next_pg = pg;
1073	}
1074	spin_unlock_irqrestore(&m->lock, flags);
1075
1076	schedule_work(&m->trigger_event);
1077	return 0;
1078}
1079
1080/*
1081 * Set/clear bypassed status of a PG.
1082 * PGs are numbered upwards from 1 in the order they were declared.
1083 */
1084static int bypass_pg_num(struct multipath *m, const char *pgstr, int bypassed)
1085{
1086	struct priority_group *pg;
1087	unsigned pgnum;
 
1088
1089	if (!pgstr || (sscanf(pgstr, "%u", &pgnum) != 1) || !pgnum ||
1090	    (pgnum > m->nr_priority_groups)) {
1091		DMWARN("invalid PG number supplied to bypass_pg");
1092		return -EINVAL;
1093	}
1094
1095	list_for_each_entry(pg, &m->priority_groups, list) {
1096		if (!--pgnum)
1097			break;
1098	}
1099
1100	bypass_pg(m, pg, bypassed);
1101	return 0;
1102}
1103
1104/*
1105 * Should we retry pg_init immediately?
1106 */
1107static int pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1108{
1109	unsigned long flags;
1110	int limit_reached = 0;
1111
1112	spin_lock_irqsave(&m->lock, flags);
1113
1114	if (m->pg_init_count <= m->pg_init_retries)
1115		m->pg_init_required = 1;
 
1116	else
1117		limit_reached = 1;
1118
1119	spin_unlock_irqrestore(&m->lock, flags);
1120
1121	return limit_reached;
1122}
1123
1124static void pg_init_done(void *data, int errors)
1125{
1126	struct pgpath *pgpath = data;
1127	struct priority_group *pg = pgpath->pg;
1128	struct multipath *m = pg->m;
1129	unsigned long flags;
1130	unsigned delay_retry = 0;
1131
1132	/* device or driver problems */
1133	switch (errors) {
1134	case SCSI_DH_OK:
1135		break;
1136	case SCSI_DH_NOSYS:
1137		if (!m->hw_handler_name) {
1138			errors = 0;
1139			break;
1140		}
1141		DMERR("Could not failover the device: Handler scsi_dh_%s "
1142		      "Error %d.", m->hw_handler_name, errors);
1143		/*
1144		 * Fail path for now, so we do not ping pong
1145		 */
1146		fail_path(pgpath);
1147		break;
1148	case SCSI_DH_DEV_TEMP_BUSY:
1149		/*
1150		 * Probably doing something like FW upgrade on the
1151		 * controller so try the other pg.
1152		 */
1153		bypass_pg(m, pg, 1);
1154		break;
1155	case SCSI_DH_RETRY:
1156		/* Wait before retrying. */
1157		delay_retry = 1;
 
1158	case SCSI_DH_IMM_RETRY:
1159	case SCSI_DH_RES_TEMP_UNAVAIL:
1160		if (pg_init_limit_reached(m, pgpath))
1161			fail_path(pgpath);
1162		errors = 0;
1163		break;
 
1164	default:
1165		/*
1166		 * We probably do not want to fail the path for a device
1167		 * error, but this is what the old dm did. In future
1168		 * patches we can do more advanced handling.
1169		 */
1170		fail_path(pgpath);
1171	}
1172
1173	spin_lock_irqsave(&m->lock, flags);
1174	if (errors) {
1175		if (pgpath == m->current_pgpath) {
1176			DMERR("Could not failover device. Error %d.", errors);
1177			m->current_pgpath = NULL;
1178			m->current_pg = NULL;
1179		}
1180	} else if (!m->pg_init_required)
1181		pg->bypassed = 0;
1182
1183	if (--m->pg_init_in_progress)
1184		/* Activations of other paths are still on going */
1185		goto out;
1186
1187	if (!m->pg_init_required)
1188		m->queue_io = 0;
 
 
 
1189
1190	m->pg_init_delay_retry = delay_retry;
1191	queue_work(kmultipathd, &m->process_queued_ios);
 
 
 
 
1192
1193	/*
1194	 * Wake up any thread waiting to suspend.
1195	 */
1196	wake_up(&m->pg_init_wait);
1197
1198out:
1199	spin_unlock_irqrestore(&m->lock, flags);
1200}
1201
1202static void activate_path(struct work_struct *work)
 
 
 
 
 
 
 
 
 
 
1203{
1204	struct pgpath *pgpath =
1205		container_of(work, struct pgpath, activate_path.work);
1206
1207	scsi_dh_activate(bdev_get_queue(pgpath->path.dev->bdev),
1208				pg_init_done, pgpath);
1209}
1210
1211/*
1212 * end_io handling
1213 */
1214static int do_end_io(struct multipath *m, struct request *clone,
1215		     int error, struct dm_mpath_io *mpio)
1216{
 
 
 
 
1217	/*
1218	 * We don't queue any clone request inside the multipath target
1219	 * during end I/O handling, since those clone requests don't have
1220	 * bio clones.  If we queue them inside the multipath target,
1221	 * we need to make bio clones, that requires memory allocation.
1222	 * (See drivers/md/dm.c:end_clone_bio() about why the clone requests
1223	 *  don't have bio clones.)
1224	 * Instead of queueing the clone request here, we queue the original
1225	 * request into dm core, which will remake a clone request and
1226	 * clone bios for it and resubmit it later.
1227	 */
1228	int r = DM_ENDIO_REQUEUE;
1229	unsigned long flags;
1230
1231	if (!error && !clone->errors)
1232		return 0;	/* I/O complete */
1233
1234	if (error == -EOPNOTSUPP || error == -EREMOTEIO || error == -EILSEQ)
1235		return error;
 
 
1236
1237	if (mpio->pgpath)
1238		fail_path(mpio->pgpath);
1239
1240	spin_lock_irqsave(&m->lock, flags);
1241	if (!m->nr_valid_paths) {
1242		if (!m->queue_if_no_path) {
1243			if (!__must_push_back(m))
1244				r = -EIO;
1245		} else {
1246			if (error == -EBADE)
1247				r = error;
1248		}
1249	}
1250	spin_unlock_irqrestore(&m->lock, flags);
 
 
 
 
 
 
 
1251
1252	return r;
1253}
1254
1255static int multipath_end_io(struct dm_target *ti, struct request *clone,
1256			    int error, union map_info *map_context)
1257{
1258	struct multipath *m = ti->private;
1259	struct dm_mpath_io *mpio = map_context->ptr;
1260	struct pgpath *pgpath = mpio->pgpath;
1261	struct path_selector *ps;
1262	int r;
1263
1264	r  = do_end_io(m, clone, error, mpio);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1265	if (pgpath) {
1266		ps = &pgpath->pg->ps;
 
1267		if (ps->type->end_io)
1268			ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes);
 
1269	}
1270	mempool_free(mpio, m->mpio_pool);
1271
1272	return r;
1273}
1274
1275/*
1276 * Suspend can't complete until all the I/O is processed so if
1277 * the last path fails we must error any remaining I/O.
1278 * Note that if the freeze_bdev fails while suspending, the
1279 * queue_if_no_path state is lost - userspace should reset it.
 
1280 */
1281static void multipath_presuspend(struct dm_target *ti)
1282{
1283	struct multipath *m = (struct multipath *) ti->private;
1284
1285	queue_if_no_path(m, 0, 1);
 
 
1286}
1287
1288static void multipath_postsuspend(struct dm_target *ti)
1289{
1290	struct multipath *m = ti->private;
1291
1292	mutex_lock(&m->work_mutex);
1293	flush_multipath_work(m);
1294	mutex_unlock(&m->work_mutex);
1295}
1296
1297/*
1298 * Restore the queue_if_no_path setting.
1299 */
1300static void multipath_resume(struct dm_target *ti)
1301{
1302	struct multipath *m = (struct multipath *) ti->private;
1303	unsigned long flags;
1304
1305	spin_lock_irqsave(&m->lock, flags);
1306	m->queue_if_no_path = m->saved_queue_if_no_path;
 
 
 
 
 
 
 
 
 
1307	spin_unlock_irqrestore(&m->lock, flags);
1308}
1309
1310/*
1311 * Info output has the following format:
1312 * num_multipath_feature_args [multipath_feature_args]*
1313 * num_handler_status_args [handler_status_args]*
1314 * num_groups init_group_number
1315 *            [A|D|E num_ps_status_args [ps_status_args]*
1316 *             num_paths num_selector_args
1317 *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1318 *
1319 * Table output has the following format (identical to the constructor string):
1320 * num_feature_args [features_args]*
1321 * num_handler_args hw_handler [hw_handler_args]*
1322 * num_groups init_group_number
1323 *     [priority selector-name num_ps_args [ps_args]*
1324 *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1325 */
1326static int multipath_status(struct dm_target *ti, status_type_t type,
1327			    char *result, unsigned int maxlen)
1328{
1329	int sz = 0;
1330	unsigned long flags;
1331	struct multipath *m = (struct multipath *) ti->private;
1332	struct priority_group *pg;
1333	struct pgpath *p;
1334	unsigned pg_num;
1335	char state;
1336
1337	spin_lock_irqsave(&m->lock, flags);
1338
1339	/* Features */
1340	if (type == STATUSTYPE_INFO)
1341		DMEMIT("2 %u %u ", m->queue_size, m->pg_init_count);
 
1342	else {
1343		DMEMIT("%u ", m->queue_if_no_path +
1344			      (m->pg_init_retries > 0) * 2 +
1345			      (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2);
1346		if (m->queue_if_no_path)
 
 
 
1347			DMEMIT("queue_if_no_path ");
1348		if (m->pg_init_retries)
1349			DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1350		if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1351			DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
 
 
 
 
 
 
 
 
 
 
 
 
1352	}
1353
1354	if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1355		DMEMIT("0 ");
1356	else
1357		DMEMIT("1 %s ", m->hw_handler_name);
1358
1359	DMEMIT("%u ", m->nr_priority_groups);
1360
1361	if (m->next_pg)
1362		pg_num = m->next_pg->pg_num;
1363	else if (m->current_pg)
1364		pg_num = m->current_pg->pg_num;
1365	else
1366		pg_num = (m->nr_priority_groups ? 1 : 0);
1367
1368	DMEMIT("%u ", pg_num);
1369
1370	switch (type) {
1371	case STATUSTYPE_INFO:
1372		list_for_each_entry(pg, &m->priority_groups, list) {
1373			if (pg->bypassed)
1374				state = 'D';	/* Disabled */
1375			else if (pg == m->current_pg)
1376				state = 'A';	/* Currently Active */
1377			else
1378				state = 'E';	/* Enabled */
1379
1380			DMEMIT("%c ", state);
1381
1382			if (pg->ps.type->status)
1383				sz += pg->ps.type->status(&pg->ps, NULL, type,
1384							  result + sz,
1385							  maxlen - sz);
1386			else
1387				DMEMIT("0 ");
1388
1389			DMEMIT("%u %u ", pg->nr_pgpaths,
1390			       pg->ps.type->info_args);
1391
1392			list_for_each_entry(p, &pg->pgpaths, list) {
1393				DMEMIT("%s %s %u ", p->path.dev->name,
1394				       p->is_active ? "A" : "F",
1395				       p->fail_count);
1396				if (pg->ps.type->status)
1397					sz += pg->ps.type->status(&pg->ps,
1398					      &p->path, type, result + sz,
1399					      maxlen - sz);
1400			}
1401		}
1402		break;
1403
1404	case STATUSTYPE_TABLE:
1405		list_for_each_entry(pg, &m->priority_groups, list) {
1406			DMEMIT("%s ", pg->ps.type->name);
1407
1408			if (pg->ps.type->status)
1409				sz += pg->ps.type->status(&pg->ps, NULL, type,
1410							  result + sz,
1411							  maxlen - sz);
1412			else
1413				DMEMIT("0 ");
1414
1415			DMEMIT("%u %u ", pg->nr_pgpaths,
1416			       pg->ps.type->table_args);
1417
1418			list_for_each_entry(p, &pg->pgpaths, list) {
1419				DMEMIT("%s ", p->path.dev->name);
1420				if (pg->ps.type->status)
1421					sz += pg->ps.type->status(&pg->ps,
1422					      &p->path, type, result + sz,
1423					      maxlen - sz);
1424			}
1425		}
1426		break;
1427	}
1428
1429	spin_unlock_irqrestore(&m->lock, flags);
1430
1431	return 0;
1432}
1433
1434static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
 
1435{
1436	int r = -EINVAL;
1437	struct dm_dev *dev;
1438	struct multipath *m = (struct multipath *) ti->private;
1439	action_fn action;
 
1440
1441	mutex_lock(&m->work_mutex);
1442
1443	if (dm_suspended(ti)) {
1444		r = -EBUSY;
1445		goto out;
1446	}
1447
1448	if (argc == 1) {
1449		if (!strcasecmp(argv[0], "queue_if_no_path")) {
1450			r = queue_if_no_path(m, 1, 0);
 
 
 
1451			goto out;
1452		} else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1453			r = queue_if_no_path(m, 0, 0);
 
1454			goto out;
1455		}
1456	}
1457
1458	if (argc != 2) {
1459		DMWARN("Unrecognised multipath message received.");
1460		goto out;
1461	}
1462
1463	if (!strcasecmp(argv[0], "disable_group")) {
1464		r = bypass_pg_num(m, argv[1], 1);
1465		goto out;
1466	} else if (!strcasecmp(argv[0], "enable_group")) {
1467		r = bypass_pg_num(m, argv[1], 0);
1468		goto out;
1469	} else if (!strcasecmp(argv[0], "switch_group")) {
1470		r = switch_pg_num(m, argv[1]);
1471		goto out;
1472	} else if (!strcasecmp(argv[0], "reinstate_path"))
1473		action = reinstate_path;
1474	else if (!strcasecmp(argv[0], "fail_path"))
1475		action = fail_path;
1476	else {
1477		DMWARN("Unrecognised multipath message received.");
1478		goto out;
1479	}
1480
1481	r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1482	if (r) {
1483		DMWARN("message: error getting device %s",
1484		       argv[1]);
1485		goto out;
1486	}
1487
1488	r = action_dev(m, dev, action);
1489
1490	dm_put_device(ti, dev);
1491
1492out:
1493	mutex_unlock(&m->work_mutex);
1494	return r;
1495}
1496
1497static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
1498			   unsigned long arg)
1499{
1500	struct multipath *m = (struct multipath *) ti->private;
1501	struct block_device *bdev = NULL;
1502	fmode_t mode = 0;
1503	unsigned long flags;
1504	int r = 0;
1505
1506	spin_lock_irqsave(&m->lock, flags);
1507
1508	if (!m->current_pgpath)
1509		__choose_pgpath(m, 0);
1510
1511	if (m->current_pgpath) {
1512		bdev = m->current_pgpath->path.dev->bdev;
1513		mode = m->current_pgpath->path.dev->mode;
1514	}
1515
1516	if (m->queue_io)
1517		r = -EAGAIN;
1518	else if (!bdev)
 
 
 
 
 
 
 
1519		r = -EIO;
 
 
 
 
 
1520
1521	spin_unlock_irqrestore(&m->lock, flags);
 
 
 
 
 
 
 
 
 
 
 
1522
1523	return r ? : __blkdev_driver_ioctl(bdev, mode, cmd, arg);
 
 
 
 
 
1524}
1525
1526static int multipath_iterate_devices(struct dm_target *ti,
1527				     iterate_devices_callout_fn fn, void *data)
1528{
1529	struct multipath *m = ti->private;
1530	struct priority_group *pg;
1531	struct pgpath *p;
1532	int ret = 0;
1533
1534	list_for_each_entry(pg, &m->priority_groups, list) {
1535		list_for_each_entry(p, &pg->pgpaths, list) {
1536			ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
1537			if (ret)
1538				goto out;
1539		}
1540	}
1541
1542out:
1543	return ret;
1544}
1545
1546static int __pgpath_busy(struct pgpath *pgpath)
1547{
1548	struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1549
1550	return dm_underlying_device_busy(q);
1551}
1552
1553/*
1554 * We return "busy", only when we can map I/Os but underlying devices
1555 * are busy (so even if we map I/Os now, the I/Os will wait on
1556 * the underlying queue).
1557 * In other words, if we want to kill I/Os or queue them inside us
1558 * due to map unavailability, we don't return "busy".  Otherwise,
1559 * dm core won't give us the I/Os and we can't do what we want.
1560 */
1561static int multipath_busy(struct dm_target *ti)
1562{
1563	int busy = 0, has_active = 0;
1564	struct multipath *m = ti->private;
1565	struct priority_group *pg;
1566	struct pgpath *pgpath;
1567	unsigned long flags;
1568
1569	spin_lock_irqsave(&m->lock, flags);
 
 
 
 
 
 
 
 
 
 
 
 
 
1570
1571	/* Guess which priority_group will be used at next mapping time */
1572	if (unlikely(!m->current_pgpath && m->next_pg))
1573		pg = m->next_pg;
1574	else if (likely(m->current_pg))
1575		pg = m->current_pg;
1576	else
 
1577		/*
1578		 * We don't know which pg will be used at next mapping time.
1579		 * We don't call __choose_pgpath() here to avoid to trigger
1580		 * pg_init just by busy checking.
1581		 * So we don't know whether underlying devices we will be using
1582		 * at next mapping time are busy or not. Just try mapping.
1583		 */
1584		goto out;
 
1585
1586	/*
1587	 * If there is one non-busy active path at least, the path selector
1588	 * will be able to select it. So we consider such a pg as not busy.
1589	 */
1590	busy = 1;
1591	list_for_each_entry(pgpath, &pg->pgpaths, list)
1592		if (pgpath->is_active) {
1593			has_active = 1;
1594
1595			if (!__pgpath_busy(pgpath)) {
1596				busy = 0;
1597				break;
1598			}
1599		}
 
1600
1601	if (!has_active)
1602		/*
1603		 * No active path in this pg, so this pg won't be used and
1604		 * the current_pg will be changed at next mapping time.
1605		 * We need to try mapping to determine it.
1606		 */
1607		busy = 0;
1608
1609out:
1610	spin_unlock_irqrestore(&m->lock, flags);
1611
1612	return busy;
1613}
1614
1615/*-----------------------------------------------------------------
1616 * Module setup
1617 *---------------------------------------------------------------*/
1618static struct target_type multipath_target = {
1619	.name = "multipath",
1620	.version = {1, 3, 0},
 
 
1621	.module = THIS_MODULE,
1622	.ctr = multipath_ctr,
1623	.dtr = multipath_dtr,
1624	.map_rq = multipath_map,
 
1625	.rq_end_io = multipath_end_io,
 
 
1626	.presuspend = multipath_presuspend,
1627	.postsuspend = multipath_postsuspend,
1628	.resume = multipath_resume,
1629	.status = multipath_status,
1630	.message = multipath_message,
1631	.ioctl  = multipath_ioctl,
1632	.iterate_devices = multipath_iterate_devices,
1633	.busy = multipath_busy,
1634};
1635
1636static int __init dm_multipath_init(void)
1637{
1638	int r;
1639
1640	/* allocate a slab for the dm_ios */
1641	_mpio_cache = KMEM_CACHE(dm_mpath_io, 0);
1642	if (!_mpio_cache)
1643		return -ENOMEM;
1644
1645	r = dm_register_target(&multipath_target);
1646	if (r < 0) {
1647		DMERR("register failed %d", r);
1648		kmem_cache_destroy(_mpio_cache);
1649		return -EINVAL;
1650	}
1651
1652	kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
1653	if (!kmultipathd) {
1654		DMERR("failed to create workqueue kmpathd");
1655		dm_unregister_target(&multipath_target);
1656		kmem_cache_destroy(_mpio_cache);
1657		return -ENOMEM;
1658	}
1659
1660	/*
1661	 * A separate workqueue is used to handle the device handlers
1662	 * to avoid overloading existing workqueue. Overloading the
1663	 * old workqueue would also create a bottleneck in the
1664	 * path of the storage hardware device activation.
1665	 */
1666	kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
1667						  WQ_MEM_RECLAIM);
1668	if (!kmpath_handlerd) {
1669		DMERR("failed to create workqueue kmpath_handlerd");
1670		destroy_workqueue(kmultipathd);
1671		dm_unregister_target(&multipath_target);
1672		kmem_cache_destroy(_mpio_cache);
1673		return -ENOMEM;
 
 
 
 
 
1674	}
1675
1676	DMINFO("version %u.%u.%u loaded",
1677	       multipath_target.version[0], multipath_target.version[1],
1678	       multipath_target.version[2]);
1679
 
 
 
 
 
1680	return r;
1681}
1682
1683static void __exit dm_multipath_exit(void)
1684{
1685	destroy_workqueue(kmpath_handlerd);
1686	destroy_workqueue(kmultipathd);
1687
1688	dm_unregister_target(&multipath_target);
1689	kmem_cache_destroy(_mpio_cache);
1690}
1691
1692module_init(dm_multipath_init);
1693module_exit(dm_multipath_exit);
 
 
 
 
1694
1695MODULE_DESCRIPTION(DM_NAME " multipath target");
1696MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
1697MODULE_LICENSE("GPL");
v5.14.15
   1/*
   2 * Copyright (C) 2003 Sistina Software Limited.
   3 * Copyright (C) 2004-2005 Red Hat, Inc. All rights reserved.
   4 *
   5 * This file is released under the GPL.
   6 */
   7
   8#include <linux/device-mapper.h>
   9
  10#include "dm-rq.h"
  11#include "dm-bio-record.h"
  12#include "dm-path-selector.h"
  13#include "dm-uevent.h"
  14
  15#include <linux/blkdev.h>
  16#include <linux/ctype.h>
  17#include <linux/init.h>
  18#include <linux/mempool.h>
  19#include <linux/module.h>
  20#include <linux/pagemap.h>
  21#include <linux/slab.h>
  22#include <linux/time.h>
  23#include <linux/timer.h>
  24#include <linux/workqueue.h>
  25#include <linux/delay.h>
  26#include <scsi/scsi_dh.h>
  27#include <linux/atomic.h>
  28#include <linux/blk-mq.h>
  29
  30#define DM_MSG_PREFIX "multipath"
  31#define DM_PG_INIT_DELAY_MSECS 2000
  32#define DM_PG_INIT_DELAY_DEFAULT ((unsigned) -1)
  33#define QUEUE_IF_NO_PATH_TIMEOUT_DEFAULT 0
  34
  35static unsigned long queue_if_no_path_timeout_secs = QUEUE_IF_NO_PATH_TIMEOUT_DEFAULT;
  36
  37/* Path properties */
  38struct pgpath {
  39	struct list_head list;
  40
  41	struct priority_group *pg;	/* Owning PG */
 
  42	unsigned fail_count;		/* Cumulative failure count */
  43
  44	struct dm_path path;
  45	struct delayed_work activate_path;
  46
  47	bool is_active:1;		/* Path status */
  48};
  49
  50#define path_to_pgpath(__pgp) container_of((__pgp), struct pgpath, path)
  51
  52/*
  53 * Paths are grouped into Priority Groups and numbered from 1 upwards.
  54 * Each has a path selector which controls which path gets used.
  55 */
  56struct priority_group {
  57	struct list_head list;
  58
  59	struct multipath *m;		/* Owning multipath instance */
  60	struct path_selector ps;
  61
  62	unsigned pg_num;		/* Reference number */
 
 
  63	unsigned nr_pgpaths;		/* Number of paths in PG */
  64	struct list_head pgpaths;
  65
  66	bool bypassed:1;		/* Temporarily bypass this PG? */
  67};
  68
  69/* Multipath context */
  70struct multipath {
  71	unsigned long flags;		/* Multipath state flags */
 
  72
  73	spinlock_t lock;
  74	enum dm_queue_mode queue_mode;
  75
  76	struct pgpath *current_pgpath;
  77	struct priority_group *current_pg;
  78	struct priority_group *next_pg;	/* Switch to this PG if set */
  79
  80	atomic_t nr_valid_paths;	/* Total number of usable paths */
  81	unsigned nr_priority_groups;
  82	struct list_head priority_groups;
  83
  84	const char *hw_handler_name;
  85	char *hw_handler_params;
  86	wait_queue_head_t pg_init_wait;	/* Wait for pg_init completion */
 
 
 
 
 
 
 
 
 
 
 
 
 
 
  87	unsigned pg_init_retries;	/* Number of times to retry pg_init */
 
  88	unsigned pg_init_delay_msecs;	/* Number of msecs before pg_init retry */
  89	atomic_t pg_init_in_progress;	/* Only one pg_init allowed at once */
  90	atomic_t pg_init_count;		/* Number of times pg_init called */
  91
  92	struct mutex work_mutex;
 
 
 
  93	struct work_struct trigger_event;
  94	struct dm_target *ti;
  95
  96	struct work_struct process_queued_bios;
  97	struct bio_list queued_bios;
 
 
 
  98
  99	struct timer_list nopath_timer;	/* Timeout for queue_if_no_path */
 100};
 101
 102/*
 103 * Context information attached to each io we process.
 104 */
 105struct dm_mpath_io {
 106	struct pgpath *pgpath;
 107	size_t nr_bytes;
 108};
 109
 110typedef int (*action_fn) (struct pgpath *pgpath);
 111
 
 
 
 
 112static struct workqueue_struct *kmultipathd, *kmpath_handlerd;
 
 113static void trigger_event(struct work_struct *work);
 114static void activate_or_offline_path(struct pgpath *pgpath);
 115static void activate_path_work(struct work_struct *work);
 116static void process_queued_bios(struct work_struct *work);
 117static void queue_if_no_path_timeout_work(struct timer_list *t);
 118
 119/*-----------------------------------------------
 120 * Multipath state flags.
 121 *-----------------------------------------------*/
 122
 123#define MPATHF_QUEUE_IO 0			/* Must we queue all I/O? */
 124#define MPATHF_QUEUE_IF_NO_PATH 1		/* Queue I/O if last path fails? */
 125#define MPATHF_SAVED_QUEUE_IF_NO_PATH 2		/* Saved state during suspension */
 126#define MPATHF_RETAIN_ATTACHED_HW_HANDLER 3	/* If there's already a hw_handler present, don't change it. */
 127#define MPATHF_PG_INIT_DISABLED 4		/* pg_init is not currently allowed */
 128#define MPATHF_PG_INIT_REQUIRED 5		/* pg_init needs calling? */
 129#define MPATHF_PG_INIT_DELAY_RETRY 6		/* Delay pg_init retry? */
 130
 131static bool mpath_double_check_test_bit(int MPATHF_bit, struct multipath *m)
 132{
 133	bool r = test_bit(MPATHF_bit, &m->flags);
 134
 135	if (r) {
 136		unsigned long flags;
 137		spin_lock_irqsave(&m->lock, flags);
 138		r = test_bit(MPATHF_bit, &m->flags);
 139		spin_unlock_irqrestore(&m->lock, flags);
 140	}
 141
 142	return r;
 143}
 144
 145/*-----------------------------------------------
 146 * Allocation routines
 147 *-----------------------------------------------*/
 148
 149static struct pgpath *alloc_pgpath(void)
 150{
 151	struct pgpath *pgpath = kzalloc(sizeof(*pgpath), GFP_KERNEL);
 152
 153	if (!pgpath)
 154		return NULL;
 155
 156	pgpath->is_active = true;
 157
 158	return pgpath;
 159}
 160
 161static void free_pgpath(struct pgpath *pgpath)
 162{
 163	kfree(pgpath);
 164}
 165
 166static struct priority_group *alloc_priority_group(void)
 167{
 168	struct priority_group *pg;
 169
 170	pg = kzalloc(sizeof(*pg), GFP_KERNEL);
 171
 172	if (pg)
 173		INIT_LIST_HEAD(&pg->pgpaths);
 174
 175	return pg;
 176}
 177
 178static void free_pgpaths(struct list_head *pgpaths, struct dm_target *ti)
 179{
 180	struct pgpath *pgpath, *tmp;
 
 181
 182	list_for_each_entry_safe(pgpath, tmp, pgpaths, list) {
 183		list_del(&pgpath->list);
 
 
 184		dm_put_device(ti, pgpath->path.dev);
 185		free_pgpath(pgpath);
 186	}
 187}
 188
 189static void free_priority_group(struct priority_group *pg,
 190				struct dm_target *ti)
 191{
 192	struct path_selector *ps = &pg->ps;
 193
 194	if (ps->type) {
 195		ps->type->destroy(ps);
 196		dm_put_path_selector(ps->type);
 197	}
 198
 199	free_pgpaths(&pg->pgpaths, ti);
 200	kfree(pg);
 201}
 202
 203static struct multipath *alloc_multipath(struct dm_target *ti)
 204{
 205	struct multipath *m;
 206
 207	m = kzalloc(sizeof(*m), GFP_KERNEL);
 208	if (m) {
 209		INIT_LIST_HEAD(&m->priority_groups);
 
 210		spin_lock_init(&m->lock);
 211		atomic_set(&m->nr_valid_paths, 0);
 
 
 212		INIT_WORK(&m->trigger_event, trigger_event);
 
 213		mutex_init(&m->work_mutex);
 214
 215		m->queue_mode = DM_TYPE_NONE;
 216
 
 
 217		m->ti = ti;
 218		ti->private = m;
 219
 220		timer_setup(&m->nopath_timer, queue_if_no_path_timeout_work, 0);
 221	}
 222
 223	return m;
 224}
 225
 226static int alloc_multipath_stage2(struct dm_target *ti, struct multipath *m)
 227{
 228	if (m->queue_mode == DM_TYPE_NONE) {
 229		m->queue_mode = DM_TYPE_REQUEST_BASED;
 230	} else if (m->queue_mode == DM_TYPE_BIO_BASED) {
 231		INIT_WORK(&m->process_queued_bios, process_queued_bios);
 232		/*
 233		 * bio-based doesn't support any direct scsi_dh management;
 234		 * it just discovers if a scsi_dh is attached.
 235		 */
 236		set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
 237	}
 238
 239	dm_table_set_type(ti->table, m->queue_mode);
 240
 241	/*
 242	 * Init fields that are only used when a scsi_dh is attached
 243	 * - must do this unconditionally (really doesn't hurt non-SCSI uses)
 244	 */
 245	set_bit(MPATHF_QUEUE_IO, &m->flags);
 246	atomic_set(&m->pg_init_in_progress, 0);
 247	atomic_set(&m->pg_init_count, 0);
 248	m->pg_init_delay_msecs = DM_PG_INIT_DELAY_DEFAULT;
 249	init_waitqueue_head(&m->pg_init_wait);
 250
 251	return 0;
 252}
 253
 254static void free_multipath(struct multipath *m)
 255{
 256	struct priority_group *pg, *tmp;
 257
 258	list_for_each_entry_safe(pg, tmp, &m->priority_groups, list) {
 259		list_del(&pg->list);
 260		free_priority_group(pg, m->ti);
 261	}
 262
 263	kfree(m->hw_handler_name);
 264	kfree(m->hw_handler_params);
 265	mutex_destroy(&m->work_mutex);
 266	kfree(m);
 267}
 268
 269static struct dm_mpath_io *get_mpio(union map_info *info)
 270{
 271	return info->ptr;
 272}
 273
 274static size_t multipath_per_bio_data_size(void)
 275{
 276	return sizeof(struct dm_mpath_io) + sizeof(struct dm_bio_details);
 277}
 278
 279static struct dm_mpath_io *get_mpio_from_bio(struct bio *bio)
 280{
 281	return dm_per_bio_data(bio, multipath_per_bio_data_size());
 282}
 283
 284static struct dm_bio_details *get_bio_details_from_mpio(struct dm_mpath_io *mpio)
 285{
 286	/* dm_bio_details is immediately after the dm_mpath_io in bio's per-bio-data */
 287	void *bio_details = mpio + 1;
 288	return bio_details;
 289}
 290
 291static void multipath_init_per_bio_data(struct bio *bio, struct dm_mpath_io **mpio_p)
 292{
 293	struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 294	struct dm_bio_details *bio_details = get_bio_details_from_mpio(mpio);
 295
 296	mpio->nr_bytes = bio->bi_iter.bi_size;
 297	mpio->pgpath = NULL;
 298	*mpio_p = mpio;
 299
 300	dm_bio_record(bio_details, bio);
 301}
 302
 303/*-----------------------------------------------
 304 * Path selection
 305 *-----------------------------------------------*/
 306
 307static int __pg_init_all_paths(struct multipath *m)
 308{
 309	struct pgpath *pgpath;
 310	unsigned long pg_init_delay = 0;
 311
 312	lockdep_assert_held(&m->lock);
 313
 314	if (atomic_read(&m->pg_init_in_progress) || test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
 315		return 0;
 316
 317	atomic_inc(&m->pg_init_count);
 318	clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 319
 320	/* Check here to reset pg_init_required */
 321	if (!m->current_pg)
 322		return 0;
 323
 324	if (test_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags))
 325		pg_init_delay = msecs_to_jiffies(m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT ?
 326						 m->pg_init_delay_msecs : DM_PG_INIT_DELAY_MSECS);
 327	list_for_each_entry(pgpath, &m->current_pg->pgpaths, list) {
 328		/* Skip failed paths */
 329		if (!pgpath->is_active)
 330			continue;
 331		if (queue_delayed_work(kmpath_handlerd, &pgpath->activate_path,
 332				       pg_init_delay))
 333			atomic_inc(&m->pg_init_in_progress);
 334	}
 335	return atomic_read(&m->pg_init_in_progress);
 336}
 337
 338static int pg_init_all_paths(struct multipath *m)
 339{
 340	int ret;
 341	unsigned long flags;
 342
 343	spin_lock_irqsave(&m->lock, flags);
 344	ret = __pg_init_all_paths(m);
 345	spin_unlock_irqrestore(&m->lock, flags);
 346
 347	return ret;
 348}
 349
 350static void __switch_pg(struct multipath *m, struct priority_group *pg)
 351{
 352	lockdep_assert_held(&m->lock);
 353
 354	m->current_pg = pg;
 355
 356	/* Must we initialise the PG first, and queue I/O till it's ready? */
 357	if (m->hw_handler_name) {
 358		set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 359		set_bit(MPATHF_QUEUE_IO, &m->flags);
 360	} else {
 361		clear_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
 362		clear_bit(MPATHF_QUEUE_IO, &m->flags);
 363	}
 364
 365	atomic_set(&m->pg_init_count, 0);
 366}
 367
 368static struct pgpath *choose_path_in_pg(struct multipath *m,
 369					struct priority_group *pg,
 370					size_t nr_bytes)
 371{
 372	unsigned long flags;
 373	struct dm_path *path;
 374	struct pgpath *pgpath;
 375
 376	path = pg->ps.type->select_path(&pg->ps, nr_bytes);
 377	if (!path)
 378		return ERR_PTR(-ENXIO);
 379
 380	pgpath = path_to_pgpath(path);
 381
 382	if (unlikely(READ_ONCE(m->current_pg) != pg)) {
 383		/* Only update current_pgpath if pg changed */
 384		spin_lock_irqsave(&m->lock, flags);
 385		m->current_pgpath = pgpath;
 386		__switch_pg(m, pg);
 387		spin_unlock_irqrestore(&m->lock, flags);
 388	}
 389
 390	return pgpath;
 391}
 392
 393static struct pgpath *choose_pgpath(struct multipath *m, size_t nr_bytes)
 394{
 395	unsigned long flags;
 396	struct priority_group *pg;
 397	struct pgpath *pgpath;
 398	unsigned bypassed = 1;
 399
 400	if (!atomic_read(&m->nr_valid_paths)) {
 401		spin_lock_irqsave(&m->lock, flags);
 402		clear_bit(MPATHF_QUEUE_IO, &m->flags);
 403		spin_unlock_irqrestore(&m->lock, flags);
 404		goto failed;
 405	}
 406
 407	/* Were we instructed to switch PG? */
 408	if (READ_ONCE(m->next_pg)) {
 409		spin_lock_irqsave(&m->lock, flags);
 410		pg = m->next_pg;
 411		if (!pg) {
 412			spin_unlock_irqrestore(&m->lock, flags);
 413			goto check_current_pg;
 414		}
 415		m->next_pg = NULL;
 416		spin_unlock_irqrestore(&m->lock, flags);
 417		pgpath = choose_path_in_pg(m, pg, nr_bytes);
 418		if (!IS_ERR_OR_NULL(pgpath))
 419			return pgpath;
 420	}
 421
 422	/* Don't change PG until it has no remaining paths */
 423check_current_pg:
 424	pg = READ_ONCE(m->current_pg);
 425	if (pg) {
 426		pgpath = choose_path_in_pg(m, pg, nr_bytes);
 427		if (!IS_ERR_OR_NULL(pgpath))
 428			return pgpath;
 429	}
 430
 431	/*
 432	 * Loop through priority groups until we find a valid path.
 433	 * First time we skip PGs marked 'bypassed'.
 434	 * Second time we only try the ones we skipped, but set
 435	 * pg_init_delay_retry so we do not hammer controllers.
 436	 */
 437	do {
 438		list_for_each_entry(pg, &m->priority_groups, list) {
 439			if (pg->bypassed == !!bypassed)
 440				continue;
 441			pgpath = choose_path_in_pg(m, pg, nr_bytes);
 442			if (!IS_ERR_OR_NULL(pgpath)) {
 443				if (!bypassed) {
 444					spin_lock_irqsave(&m->lock, flags);
 445					set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
 446					spin_unlock_irqrestore(&m->lock, flags);
 447				}
 448				return pgpath;
 449			}
 450		}
 451	} while (bypassed--);
 452
 453failed:
 454	spin_lock_irqsave(&m->lock, flags);
 455	m->current_pgpath = NULL;
 456	m->current_pg = NULL;
 457	spin_unlock_irqrestore(&m->lock, flags);
 458
 459	return NULL;
 460}
 461
 462/*
 463 * dm_report_EIO() is a macro instead of a function to make pr_debug_ratelimited()
 464 * report the function name and line number of the function from which
 465 * it has been invoked.
 466 */
 467#define dm_report_EIO(m)						\
 468do {									\
 469	DMDEBUG_LIMIT("%s: returning EIO; QIFNP = %d; SQIFNP = %d; DNFS = %d", \
 470		      dm_table_device_name((m)->ti->table),		\
 471		      test_bit(MPATHF_QUEUE_IF_NO_PATH, &(m)->flags),	\
 472		      test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &(m)->flags), \
 473		      dm_noflush_suspending((m)->ti));			\
 474} while (0)
 475
 476/*
 477 * Check whether bios must be queued in the device-mapper core rather
 478 * than here in the target.
 
 
 
 
 
 
 
 479 */
 480static bool __must_push_back(struct multipath *m)
 481{
 482	return dm_noflush_suspending(m->ti);
 
 483}
 484
 485static bool must_push_back_rq(struct multipath *m)
 
 486{
 
 
 487	unsigned long flags;
 488	bool ret;
 
 489
 490	spin_lock_irqsave(&m->lock, flags);
 491	ret = (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) || __must_push_back(m));
 492	spin_unlock_irqrestore(&m->lock, flags);
 493
 494	return ret;
 495}
 496
 497/*
 498 * Map cloned requests (request-based multipath)
 499 */
 500static int multipath_clone_and_map(struct dm_target *ti, struct request *rq,
 501				   union map_info *map_context,
 502				   struct request **__clone)
 503{
 504	struct multipath *m = ti->private;
 505	size_t nr_bytes = blk_rq_bytes(rq);
 506	struct pgpath *pgpath;
 507	struct block_device *bdev;
 508	struct dm_mpath_io *mpio = get_mpio(map_context);
 509	struct request_queue *q;
 510	struct request *clone;
 511
 512	/* Do we need to select a new pgpath? */
 513	pgpath = READ_ONCE(m->current_pgpath);
 514	if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
 515		pgpath = choose_pgpath(m, nr_bytes);
 516
 517	if (!pgpath) {
 518		if (must_push_back_rq(m))
 519			return DM_MAPIO_DELAY_REQUEUE;
 520		dm_report_EIO(m);	/* Failed */
 521		return DM_MAPIO_KILL;
 522	} else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
 523		   mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
 524		pg_init_all_paths(m);
 525		return DM_MAPIO_DELAY_REQUEUE;
 526	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 527
 528	mpio->pgpath = pgpath;
 529	mpio->nr_bytes = nr_bytes;
 530
 531	bdev = pgpath->path.dev->bdev;
 532	q = bdev_get_queue(bdev);
 533	clone = blk_get_request(q, rq->cmd_flags | REQ_NOMERGE,
 534			BLK_MQ_REQ_NOWAIT);
 535	if (IS_ERR(clone)) {
 536		/* EBUSY, ENODEV or EWOULDBLOCK: requeue */
 537		if (blk_queue_dying(q)) {
 538			atomic_inc(&m->pg_init_in_progress);
 539			activate_or_offline_path(pgpath);
 540			return DM_MAPIO_DELAY_REQUEUE;
 541		}
 542
 543		/*
 544		 * blk-mq's SCHED_RESTART can cover this requeue, so we
 545		 * needn't deal with it by DELAY_REQUEUE. More importantly,
 546		 * we have to return DM_MAPIO_REQUEUE so that blk-mq can
 547		 * get the queue busy feedback (via BLK_STS_RESOURCE),
 548		 * otherwise I/O merging can suffer.
 549		 */
 550		return DM_MAPIO_REQUEUE;
 551	}
 552	clone->bio = clone->biotail = NULL;
 553	clone->rq_disk = bdev->bd_disk;
 554	clone->cmd_flags |= REQ_FAILFAST_TRANSPORT;
 555	*__clone = clone;
 556
 557	if (pgpath->pg->ps.type->start_io)
 558		pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 559					      &pgpath->path,
 560					      nr_bytes);
 561	return DM_MAPIO_REMAPPED;
 562}
 563
 564static void multipath_release_clone(struct request *clone,
 565				    union map_info *map_context)
 566{
 567	if (unlikely(map_context)) {
 568		/*
 569		 * non-NULL map_context means caller is still map
 570		 * method; must undo multipath_clone_and_map()
 571		 */
 572		struct dm_mpath_io *mpio = get_mpio(map_context);
 573		struct pgpath *pgpath = mpio->pgpath;
 574
 575		if (pgpath && pgpath->pg->ps.type->end_io)
 576			pgpath->pg->ps.type->end_io(&pgpath->pg->ps,
 577						    &pgpath->path,
 578						    mpio->nr_bytes,
 579						    clone->io_start_time_ns);
 580	}
 581
 582	blk_put_request(clone);
 583}
 584
 585/*
 586 * Map cloned bios (bio-based multipath)
 587 */
 588
 589static void __multipath_queue_bio(struct multipath *m, struct bio *bio)
 590{
 591	/* Queue for the daemon to resubmit */
 592	bio_list_add(&m->queued_bios, bio);
 593	if (!test_bit(MPATHF_QUEUE_IO, &m->flags))
 594		queue_work(kmultipathd, &m->process_queued_bios);
 595}
 596
 597static void multipath_queue_bio(struct multipath *m, struct bio *bio)
 598{
 599	unsigned long flags;
 600
 601	spin_lock_irqsave(&m->lock, flags);
 602	__multipath_queue_bio(m, bio);
 603	spin_unlock_irqrestore(&m->lock, flags);
 604}
 605
 606static struct pgpath *__map_bio(struct multipath *m, struct bio *bio)
 607{
 608	struct pgpath *pgpath;
 609	unsigned long flags;
 
 
 
 610
 611	/* Do we need to select a new pgpath? */
 612	pgpath = READ_ONCE(m->current_pgpath);
 613	if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
 614		pgpath = choose_pgpath(m, bio->bi_iter.bi_size);
 615
 616	if (!pgpath) {
 617		spin_lock_irqsave(&m->lock, flags);
 618		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
 619			__multipath_queue_bio(m, bio);
 620			pgpath = ERR_PTR(-EAGAIN);
 621		}
 622		spin_unlock_irqrestore(&m->lock, flags);
 623
 624	} else if (mpath_double_check_test_bit(MPATHF_QUEUE_IO, m) ||
 625		   mpath_double_check_test_bit(MPATHF_PG_INIT_REQUIRED, m)) {
 626		multipath_queue_bio(m, bio);
 627		pg_init_all_paths(m);
 628		return ERR_PTR(-EAGAIN);
 629	}
 630
 631	return pgpath;
 632}
 633
 634static int __multipath_map_bio(struct multipath *m, struct bio *bio,
 635			       struct dm_mpath_io *mpio)
 636{
 637	struct pgpath *pgpath = __map_bio(m, bio);
 638
 639	if (IS_ERR(pgpath))
 640		return DM_MAPIO_SUBMITTED;
 641
 642	if (!pgpath) {
 643		if (__must_push_back(m))
 644			return DM_MAPIO_REQUEUE;
 645		dm_report_EIO(m);
 646		return DM_MAPIO_KILL;
 647	}
 648
 649	mpio->pgpath = pgpath;
 650
 651	bio->bi_status = 0;
 652	bio_set_dev(bio, pgpath->path.dev->bdev);
 653	bio->bi_opf |= REQ_FAILFAST_TRANSPORT;
 654
 655	if (pgpath->pg->ps.type->start_io)
 656		pgpath->pg->ps.type->start_io(&pgpath->pg->ps,
 657					      &pgpath->path,
 658					      mpio->nr_bytes);
 659	return DM_MAPIO_REMAPPED;
 660}
 661
 662static int multipath_map_bio(struct dm_target *ti, struct bio *bio)
 663{
 664	struct multipath *m = ti->private;
 665	struct dm_mpath_io *mpio = NULL;
 666
 667	multipath_init_per_bio_data(bio, &mpio);
 668	return __multipath_map_bio(m, bio, mpio);
 669}
 670
 671static void process_queued_io_list(struct multipath *m)
 672{
 673	if (m->queue_mode == DM_TYPE_REQUEST_BASED)
 674		dm_mq_kick_requeue_list(dm_table_get_md(m->ti->table));
 675	else if (m->queue_mode == DM_TYPE_BIO_BASED)
 676		queue_work(kmultipathd, &m->process_queued_bios);
 677}
 678
 679static void process_queued_bios(struct work_struct *work)
 680{
 681	int r;
 682	unsigned long flags;
 683	struct bio *bio;
 684	struct bio_list bios;
 685	struct blk_plug plug;
 686	struct multipath *m =
 687		container_of(work, struct multipath, process_queued_bios);
 688
 689	bio_list_init(&bios);
 690
 691	spin_lock_irqsave(&m->lock, flags);
 
 
 692
 693	if (bio_list_empty(&m->queued_bios)) {
 694		spin_unlock_irqrestore(&m->lock, flags);
 695		return;
 696	}
 697
 698	bio_list_merge(&bios, &m->queued_bios);
 699	bio_list_init(&m->queued_bios);
 700
 701	spin_unlock_irqrestore(&m->lock, flags);
 702
 703	blk_start_plug(&plug);
 704	while ((bio = bio_list_pop(&bios))) {
 705		struct dm_mpath_io *mpio = get_mpio_from_bio(bio);
 706		dm_bio_restore(get_bio_details_from_mpio(mpio), bio);
 707		r = __multipath_map_bio(m, bio, mpio);
 708		switch (r) {
 709		case DM_MAPIO_KILL:
 710			bio->bi_status = BLK_STS_IOERR;
 711			bio_endio(bio);
 712			break;
 713		case DM_MAPIO_REQUEUE:
 714			bio->bi_status = BLK_STS_DM_REQUEUE;
 715			bio_endio(bio);
 716			break;
 717		case DM_MAPIO_REMAPPED:
 718			submit_bio_noacct(bio);
 719			break;
 720		case DM_MAPIO_SUBMITTED:
 721			break;
 722		default:
 723			WARN_ONCE(true, "__multipath_map_bio() returned %d\n", r);
 724		}
 725	}
 726	blk_finish_plug(&plug);
 727}
 728
 729/*
 730 * If we run out of usable paths, should we queue I/O or error it?
 731 */
 732static int queue_if_no_path(struct multipath *m, bool queue_if_no_path,
 733			    bool save_old_value, const char *caller)
 734{
 
 
 
 
 735	unsigned long flags;
 736	bool queue_if_no_path_bit, saved_queue_if_no_path_bit;
 737	const char *dm_dev_name = dm_table_device_name(m->ti->table);
 738
 739	DMDEBUG("%s: %s caller=%s queue_if_no_path=%d save_old_value=%d",
 740		dm_dev_name, __func__, caller, queue_if_no_path, save_old_value);
 741
 742	spin_lock_irqsave(&m->lock, flags);
 743
 744	queue_if_no_path_bit = test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
 745	saved_queue_if_no_path_bit = test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 746
 747	if (save_old_value) {
 748		if (unlikely(!queue_if_no_path_bit && saved_queue_if_no_path_bit)) {
 749			DMERR("%s: QIFNP disabled but saved as enabled, saving again loses state, not saving!",
 750			      dm_dev_name);
 751		} else
 752			assign_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path_bit);
 753	} else if (!queue_if_no_path && saved_queue_if_no_path_bit) {
 754		/* due to "fail_if_no_path" message, need to honor it. */
 755		clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
 756	}
 757	assign_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags, queue_if_no_path);
 758
 759	DMDEBUG("%s: after %s changes; QIFNP = %d; SQIFNP = %d; DNFS = %d",
 760		dm_dev_name, __func__,
 761		test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
 762		test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags),
 763		dm_noflush_suspending(m->ti));
 764
 765	spin_unlock_irqrestore(&m->lock, flags);
 
 766
 767	if (!queue_if_no_path) {
 768		dm_table_run_md_queue_async(m->ti->table);
 769		process_queued_io_list(m);
 770	}
 771
 772	return 0;
 773}
 
 774
 775/*
 776 * If the queue_if_no_path timeout fires, turn off queue_if_no_path and
 777 * process any queued I/O.
 778 */
 779static void queue_if_no_path_timeout_work(struct timer_list *t)
 780{
 781	struct multipath *m = from_timer(m, t, nopath_timer);
 782
 783	DMWARN("queue_if_no_path timeout on %s, failing queued IO",
 784	       dm_table_device_name(m->ti->table));
 785	queue_if_no_path(m, false, false, __func__);
 786}
 787
 788/*
 789 * Enable the queue_if_no_path timeout if necessary.
 790 * Called with m->lock held.
 791 */
 792static void enable_nopath_timeout(struct multipath *m)
 793{
 794	unsigned long queue_if_no_path_timeout =
 795		READ_ONCE(queue_if_no_path_timeout_secs) * HZ;
 796
 797	lockdep_assert_held(&m->lock);
 798
 799	if (queue_if_no_path_timeout > 0 &&
 800	    atomic_read(&m->nr_valid_paths) == 0 &&
 801	    test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
 802		mod_timer(&m->nopath_timer,
 803			  jiffies + queue_if_no_path_timeout);
 804	}
 805}
 806
 807static void disable_nopath_timeout(struct multipath *m)
 808{
 809	del_timer_sync(&m->nopath_timer);
 810}
 811
 812/*
 813 * An event is triggered whenever a path is taken out of use.
 814 * Includes path failure and PG bypass.
 815 */
 816static void trigger_event(struct work_struct *work)
 817{
 818	struct multipath *m =
 819		container_of(work, struct multipath, trigger_event);
 820
 821	dm_table_event(m->ti->table);
 822}
 823
 824/*-----------------------------------------------------------------
 825 * Constructor/argument parsing:
 826 * <#multipath feature args> [<arg>]*
 827 * <#hw_handler args> [hw_handler [<arg>]*]
 828 * <#priority groups>
 829 * <initial priority group>
 830 *     [<selector> <#selector args> [<arg>]*
 831 *      <#paths> <#per-path selector args>
 832 *         [<path> [<arg>]* ]+ ]+
 833 *---------------------------------------------------------------*/
 834static int parse_path_selector(struct dm_arg_set *as, struct priority_group *pg,
 835			       struct dm_target *ti)
 836{
 837	int r;
 838	struct path_selector_type *pst;
 839	unsigned ps_argc;
 840
 841	static const struct dm_arg _args[] = {
 842		{0, 1024, "invalid number of path selector args"},
 843	};
 844
 845	pst = dm_get_path_selector(dm_shift_arg(as));
 846	if (!pst) {
 847		ti->error = "unknown path selector type";
 848		return -EINVAL;
 849	}
 850
 851	r = dm_read_arg_group(_args, as, &ps_argc, &ti->error);
 852	if (r) {
 853		dm_put_path_selector(pst);
 854		return -EINVAL;
 855	}
 856
 857	r = pst->create(&pg->ps, ps_argc, as->argv);
 858	if (r) {
 859		dm_put_path_selector(pst);
 860		ti->error = "path selector constructor failed";
 861		return r;
 862	}
 863
 864	pg->ps.type = pst;
 865	dm_consume_args(as, ps_argc);
 866
 867	return 0;
 868}
 869
 870static int setup_scsi_dh(struct block_device *bdev, struct multipath *m,
 871			 const char **attached_handler_name, char **error)
 872{
 873	struct request_queue *q = bdev_get_queue(bdev);
 874	int r;
 875
 876	if (mpath_double_check_test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, m)) {
 877retain:
 878		if (*attached_handler_name) {
 879			/*
 880			 * Clear any hw_handler_params associated with a
 881			 * handler that isn't already attached.
 882			 */
 883			if (m->hw_handler_name && strcmp(*attached_handler_name, m->hw_handler_name)) {
 884				kfree(m->hw_handler_params);
 885				m->hw_handler_params = NULL;
 886			}
 887
 888			/*
 889			 * Reset hw_handler_name to match the attached handler
 890			 *
 891			 * NB. This modifies the table line to show the actual
 892			 * handler instead of the original table passed in.
 893			 */
 894			kfree(m->hw_handler_name);
 895			m->hw_handler_name = *attached_handler_name;
 896			*attached_handler_name = NULL;
 897		}
 898	}
 899
 900	if (m->hw_handler_name) {
 901		r = scsi_dh_attach(q, m->hw_handler_name);
 902		if (r == -EBUSY) {
 903			char b[BDEVNAME_SIZE];
 904
 905			printk(KERN_INFO "dm-mpath: retaining handler on device %s\n",
 906			       bdevname(bdev, b));
 907			goto retain;
 908		}
 909		if (r < 0) {
 910			*error = "error attaching hardware handler";
 911			return r;
 912		}
 913
 914		if (m->hw_handler_params) {
 915			r = scsi_dh_set_params(q, m->hw_handler_params);
 916			if (r < 0) {
 917				*error = "unable to set hardware handler parameters";
 918				return r;
 919			}
 920		}
 921	}
 922
 923	return 0;
 924}
 925
 926static struct pgpath *parse_path(struct dm_arg_set *as, struct path_selector *ps,
 927				 struct dm_target *ti)
 928{
 929	int r;
 930	struct pgpath *p;
 931	struct multipath *m = ti->private;
 932	struct request_queue *q;
 933	const char *attached_handler_name = NULL;
 934
 935	/* we need at least a path arg */
 936	if (as->argc < 1) {
 937		ti->error = "no device given";
 938		return ERR_PTR(-EINVAL);
 939	}
 940
 941	p = alloc_pgpath();
 942	if (!p)
 943		return ERR_PTR(-ENOMEM);
 944
 945	r = dm_get_device(ti, dm_shift_arg(as), dm_table_get_mode(ti->table),
 946			  &p->path.dev);
 947	if (r) {
 948		ti->error = "error getting device";
 949		goto bad;
 950	}
 951
 952	q = bdev_get_queue(p->path.dev->bdev);
 953	attached_handler_name = scsi_dh_attached_handler_name(q, GFP_KERNEL);
 954	if (attached_handler_name || m->hw_handler_name) {
 955		INIT_DELAYED_WORK(&p->activate_path, activate_path_work);
 956		r = setup_scsi_dh(p->path.dev->bdev, m, &attached_handler_name, &ti->error);
 957		kfree(attached_handler_name);
 958		if (r) {
 
 
 
 
 
 
 
 
 959			dm_put_device(ti, p->path.dev);
 960			goto bad;
 961		}
 
 
 
 
 
 
 
 
 
 
 
 962	}
 963
 964	r = ps->type->add_path(ps, &p->path, as->argc, as->argv, &ti->error);
 965	if (r) {
 966		dm_put_device(ti, p->path.dev);
 967		goto bad;
 968	}
 969
 970	return p;
 
 971 bad:
 972	free_pgpath(p);
 973	return ERR_PTR(r);
 974}
 975
 976static struct priority_group *parse_priority_group(struct dm_arg_set *as,
 977						   struct multipath *m)
 978{
 979	static const struct dm_arg _args[] = {
 980		{1, 1024, "invalid number of paths"},
 981		{0, 1024, "invalid number of selector args"}
 982	};
 983
 984	int r;
 985	unsigned i, nr_selector_args, nr_args;
 986	struct priority_group *pg;
 987	struct dm_target *ti = m->ti;
 988
 989	if (as->argc < 2) {
 990		as->argc = 0;
 991		ti->error = "not enough priority group arguments";
 992		return ERR_PTR(-EINVAL);
 993	}
 994
 995	pg = alloc_priority_group();
 996	if (!pg) {
 997		ti->error = "couldn't allocate priority group";
 998		return ERR_PTR(-ENOMEM);
 999	}
1000	pg->m = m;
1001
1002	r = parse_path_selector(as, pg, ti);
1003	if (r)
1004		goto bad;
1005
1006	/*
1007	 * read the paths
1008	 */
1009	r = dm_read_arg(_args, as, &pg->nr_pgpaths, &ti->error);
1010	if (r)
1011		goto bad;
1012
1013	r = dm_read_arg(_args + 1, as, &nr_selector_args, &ti->error);
1014	if (r)
1015		goto bad;
1016
1017	nr_args = 1 + nr_selector_args;
1018	for (i = 0; i < pg->nr_pgpaths; i++) {
1019		struct pgpath *pgpath;
1020		struct dm_arg_set path_args;
1021
1022		if (as->argc < nr_args) {
1023			ti->error = "not enough path parameters";
1024			r = -EINVAL;
1025			goto bad;
1026		}
1027
1028		path_args.argc = nr_args;
1029		path_args.argv = as->argv;
1030
1031		pgpath = parse_path(&path_args, &pg->ps, ti);
1032		if (IS_ERR(pgpath)) {
1033			r = PTR_ERR(pgpath);
1034			goto bad;
1035		}
1036
1037		pgpath->pg = pg;
1038		list_add_tail(&pgpath->list, &pg->pgpaths);
1039		dm_consume_args(as, nr_args);
1040	}
1041
1042	return pg;
1043
1044 bad:
1045	free_priority_group(pg, ti);
1046	return ERR_PTR(r);
1047}
1048
1049static int parse_hw_handler(struct dm_arg_set *as, struct multipath *m)
1050{
1051	unsigned hw_argc;
1052	int ret;
1053	struct dm_target *ti = m->ti;
1054
1055	static const struct dm_arg _args[] = {
1056		{0, 1024, "invalid number of hardware handler args"},
1057	};
1058
1059	if (dm_read_arg_group(_args, as, &hw_argc, &ti->error))
1060		return -EINVAL;
1061
1062	if (!hw_argc)
1063		return 0;
1064
1065	if (m->queue_mode == DM_TYPE_BIO_BASED) {
1066		dm_consume_args(as, hw_argc);
1067		DMERR("bio-based multipath doesn't allow hardware handler args");
1068		return 0;
 
 
1069	}
1070
1071	m->hw_handler_name = kstrdup(dm_shift_arg(as), GFP_KERNEL);
1072	if (!m->hw_handler_name)
1073		return -EINVAL;
1074
1075	if (hw_argc > 1) {
1076		char *p;
1077		int i, j, len = 4;
1078
1079		for (i = 0; i <= hw_argc - 2; i++)
1080			len += strlen(as->argv[i]) + 1;
1081		p = m->hw_handler_params = kzalloc(len, GFP_KERNEL);
1082		if (!p) {
1083			ti->error = "memory allocation failed";
1084			ret = -ENOMEM;
1085			goto fail;
1086		}
1087		j = sprintf(p, "%d", hw_argc - 1);
1088		for (i = 0, p+=j+1; i <= hw_argc - 2; i++, p+=j+1)
1089			j = sprintf(p, "%s", as->argv[i]);
1090	}
1091	dm_consume_args(as, hw_argc - 1);
1092
1093	return 0;
1094fail:
1095	kfree(m->hw_handler_name);
1096	m->hw_handler_name = NULL;
1097	return ret;
1098}
1099
1100static int parse_features(struct dm_arg_set *as, struct multipath *m)
1101{
1102	int r;
1103	unsigned argc;
1104	struct dm_target *ti = m->ti;
1105	const char *arg_name;
1106
1107	static const struct dm_arg _args[] = {
1108		{0, 8, "invalid number of feature args"},
1109		{1, 50, "pg_init_retries must be between 1 and 50"},
1110		{0, 60000, "pg_init_delay_msecs must be between 0 and 60000"},
1111	};
1112
1113	r = dm_read_arg_group(_args, as, &argc, &ti->error);
1114	if (r)
1115		return -EINVAL;
1116
1117	if (!argc)
1118		return 0;
1119
1120	do {
1121		arg_name = dm_shift_arg(as);
1122		argc--;
1123
1124		if (!strcasecmp(arg_name, "queue_if_no_path")) {
1125			r = queue_if_no_path(m, true, false, __func__);
1126			continue;
1127		}
1128
1129		if (!strcasecmp(arg_name, "retain_attached_hw_handler")) {
1130			set_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags);
1131			continue;
1132		}
1133
1134		if (!strcasecmp(arg_name, "pg_init_retries") &&
1135		    (argc >= 1)) {
1136			r = dm_read_arg(_args + 1, as, &m->pg_init_retries, &ti->error);
1137			argc--;
1138			continue;
1139		}
1140
1141		if (!strcasecmp(arg_name, "pg_init_delay_msecs") &&
1142		    (argc >= 1)) {
1143			r = dm_read_arg(_args + 2, as, &m->pg_init_delay_msecs, &ti->error);
1144			argc--;
1145			continue;
1146		}
1147
1148		if (!strcasecmp(arg_name, "queue_mode") &&
1149		    (argc >= 1)) {
1150			const char *queue_mode_name = dm_shift_arg(as);
1151
1152			if (!strcasecmp(queue_mode_name, "bio"))
1153				m->queue_mode = DM_TYPE_BIO_BASED;
1154			else if (!strcasecmp(queue_mode_name, "rq") ||
1155				 !strcasecmp(queue_mode_name, "mq"))
1156				m->queue_mode = DM_TYPE_REQUEST_BASED;
1157			else {
1158				ti->error = "Unknown 'queue_mode' requested";
1159				r = -EINVAL;
1160			}
1161			argc--;
1162			continue;
1163		}
1164
1165		ti->error = "Unrecognised multipath feature request";
1166		r = -EINVAL;
1167	} while (argc && !r);
1168
1169	return r;
1170}
1171
1172static int multipath_ctr(struct dm_target *ti, unsigned argc, char **argv)
 
1173{
1174	/* target arguments */
1175	static const struct dm_arg _args[] = {
1176		{0, 1024, "invalid number of priority groups"},
1177		{0, 1024, "invalid initial priority group number"},
1178	};
1179
1180	int r;
1181	struct multipath *m;
1182	struct dm_arg_set as;
1183	unsigned pg_count = 0;
1184	unsigned next_pg_num;
1185	unsigned long flags;
1186
1187	as.argc = argc;
1188	as.argv = argv;
1189
1190	m = alloc_multipath(ti);
1191	if (!m) {
1192		ti->error = "can't allocate multipath";
1193		return -EINVAL;
1194	}
1195
1196	r = parse_features(&as, m);
1197	if (r)
1198		goto bad;
1199
1200	r = alloc_multipath_stage2(ti, m);
1201	if (r)
1202		goto bad;
1203
1204	r = parse_hw_handler(&as, m);
1205	if (r)
1206		goto bad;
1207
1208	r = dm_read_arg(_args, &as, &m->nr_priority_groups, &ti->error);
1209	if (r)
1210		goto bad;
1211
1212	r = dm_read_arg(_args + 1, &as, &next_pg_num, &ti->error);
1213	if (r)
1214		goto bad;
1215
1216	if ((!m->nr_priority_groups && next_pg_num) ||
1217	    (m->nr_priority_groups && !next_pg_num)) {
1218		ti->error = "invalid initial priority group";
1219		r = -EINVAL;
1220		goto bad;
1221	}
1222
1223	/* parse the priority groups */
1224	while (as.argc) {
1225		struct priority_group *pg;
1226		unsigned nr_valid_paths = atomic_read(&m->nr_valid_paths);
1227
1228		pg = parse_priority_group(&as, m);
1229		if (IS_ERR(pg)) {
1230			r = PTR_ERR(pg);
1231			goto bad;
1232		}
1233
1234		nr_valid_paths += pg->nr_pgpaths;
1235		atomic_set(&m->nr_valid_paths, nr_valid_paths);
1236
1237		list_add_tail(&pg->list, &m->priority_groups);
1238		pg_count++;
1239		pg->pg_num = pg_count;
1240		if (!--next_pg_num)
1241			m->next_pg = pg;
1242	}
1243
1244	if (pg_count != m->nr_priority_groups) {
1245		ti->error = "priority group count mismatch";
1246		r = -EINVAL;
1247		goto bad;
1248	}
1249
1250	spin_lock_irqsave(&m->lock, flags);
1251	enable_nopath_timeout(m);
1252	spin_unlock_irqrestore(&m->lock, flags);
1253
1254	ti->num_flush_bios = 1;
1255	ti->num_discard_bios = 1;
1256	ti->num_write_same_bios = 1;
1257	ti->num_write_zeroes_bios = 1;
1258	if (m->queue_mode == DM_TYPE_BIO_BASED)
1259		ti->per_io_data_size = multipath_per_bio_data_size();
1260	else
1261		ti->per_io_data_size = sizeof(struct dm_mpath_io);
1262
1263	return 0;
1264
1265 bad:
1266	free_multipath(m);
1267	return r;
1268}
1269
1270static void multipath_wait_for_pg_init_completion(struct multipath *m)
1271{
1272	DEFINE_WAIT(wait);
 
 
 
1273
1274	while (1) {
1275		prepare_to_wait(&m->pg_init_wait, &wait, TASK_UNINTERRUPTIBLE);
1276
1277		if (!atomic_read(&m->pg_init_in_progress))
 
 
1278			break;
 
 
1279
1280		io_schedule();
1281	}
1282	finish_wait(&m->pg_init_wait, &wait);
 
 
1283}
1284
1285static void flush_multipath_work(struct multipath *m)
1286{
1287	if (m->hw_handler_name) {
1288		unsigned long flags;
1289
1290		if (!atomic_read(&m->pg_init_in_progress))
1291			goto skip;
1292
1293		spin_lock_irqsave(&m->lock, flags);
1294		if (atomic_read(&m->pg_init_in_progress) &&
1295		    !test_and_set_bit(MPATHF_PG_INIT_DISABLED, &m->flags)) {
1296			spin_unlock_irqrestore(&m->lock, flags);
1297
1298			flush_workqueue(kmpath_handlerd);
1299			multipath_wait_for_pg_init_completion(m);
1300
1301			spin_lock_irqsave(&m->lock, flags);
1302			clear_bit(MPATHF_PG_INIT_DISABLED, &m->flags);
1303		}
1304		spin_unlock_irqrestore(&m->lock, flags);
1305	}
1306skip:
1307	if (m->queue_mode == DM_TYPE_BIO_BASED)
1308		flush_work(&m->process_queued_bios);
1309	flush_work(&m->trigger_event);
1310}
1311
1312static void multipath_dtr(struct dm_target *ti)
1313{
1314	struct multipath *m = ti->private;
1315
1316	disable_nopath_timeout(m);
1317	flush_multipath_work(m);
1318	free_multipath(m);
1319}
1320
1321/*
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1322 * Take a path out of use.
1323 */
1324static int fail_path(struct pgpath *pgpath)
1325{
1326	unsigned long flags;
1327	struct multipath *m = pgpath->pg->m;
1328
1329	spin_lock_irqsave(&m->lock, flags);
1330
1331	if (!pgpath->is_active)
1332		goto out;
1333
1334	DMWARN("%s: Failing path %s.",
1335	       dm_table_device_name(m->ti->table),
1336	       pgpath->path.dev->name);
1337
1338	pgpath->pg->ps.type->fail_path(&pgpath->pg->ps, &pgpath->path);
1339	pgpath->is_active = false;
1340	pgpath->fail_count++;
1341
1342	atomic_dec(&m->nr_valid_paths);
1343
1344	if (pgpath == m->current_pgpath)
1345		m->current_pgpath = NULL;
1346
1347	dm_path_uevent(DM_UEVENT_PATH_FAILED, m->ti,
1348		       pgpath->path.dev->name, atomic_read(&m->nr_valid_paths));
1349
1350	schedule_work(&m->trigger_event);
1351
1352	enable_nopath_timeout(m);
1353
1354out:
1355	spin_unlock_irqrestore(&m->lock, flags);
1356
1357	return 0;
1358}
1359
1360/*
1361 * Reinstate a previously-failed path
1362 */
1363static int reinstate_path(struct pgpath *pgpath)
1364{
1365	int r = 0, run_queue = 0;
1366	unsigned long flags;
1367	struct multipath *m = pgpath->pg->m;
1368	unsigned nr_valid_paths;
1369
1370	spin_lock_irqsave(&m->lock, flags);
1371
1372	if (pgpath->is_active)
1373		goto out;
1374
1375	DMWARN("%s: Reinstating path %s.",
1376	       dm_table_device_name(m->ti->table),
1377	       pgpath->path.dev->name);
 
 
 
1378
1379	r = pgpath->pg->ps.type->reinstate_path(&pgpath->pg->ps, &pgpath->path);
1380	if (r)
1381		goto out;
1382
1383	pgpath->is_active = true;
1384
1385	nr_valid_paths = atomic_inc_return(&m->nr_valid_paths);
1386	if (nr_valid_paths == 1) {
1387		m->current_pgpath = NULL;
1388		run_queue = 1;
1389	} else if (m->hw_handler_name && (m->current_pg == pgpath->pg)) {
1390		if (queue_work(kmpath_handlerd, &pgpath->activate_path.work))
1391			atomic_inc(&m->pg_init_in_progress);
1392	}
1393
1394	dm_path_uevent(DM_UEVENT_PATH_REINSTATED, m->ti,
1395		       pgpath->path.dev->name, nr_valid_paths);
1396
1397	schedule_work(&m->trigger_event);
1398
1399out:
1400	spin_unlock_irqrestore(&m->lock, flags);
1401	if (run_queue) {
1402		dm_table_run_md_queue_async(m->ti->table);
1403		process_queued_io_list(m);
1404	}
1405
1406	if (pgpath->is_active)
1407		disable_nopath_timeout(m);
1408
1409	return r;
1410}
1411
1412/*
1413 * Fail or reinstate all paths that match the provided struct dm_dev.
1414 */
1415static int action_dev(struct multipath *m, struct dm_dev *dev,
1416		      action_fn action)
1417{
1418	int r = -EINVAL;
1419	struct pgpath *pgpath;
1420	struct priority_group *pg;
1421
1422	list_for_each_entry(pg, &m->priority_groups, list) {
1423		list_for_each_entry(pgpath, &pg->pgpaths, list) {
1424			if (pgpath->path.dev == dev)
1425				r = action(pgpath);
1426		}
1427	}
1428
1429	return r;
1430}
1431
1432/*
1433 * Temporarily try to avoid having to use the specified PG
1434 */
1435static void bypass_pg(struct multipath *m, struct priority_group *pg,
1436		      bool bypassed)
1437{
1438	unsigned long flags;
1439
1440	spin_lock_irqsave(&m->lock, flags);
1441
1442	pg->bypassed = bypassed;
1443	m->current_pgpath = NULL;
1444	m->current_pg = NULL;
1445
1446	spin_unlock_irqrestore(&m->lock, flags);
1447
1448	schedule_work(&m->trigger_event);
1449}
1450
1451/*
1452 * Switch to using the specified PG from the next I/O that gets mapped
1453 */
1454static int switch_pg_num(struct multipath *m, const char *pgstr)
1455{
1456	struct priority_group *pg;
1457	unsigned pgnum;
1458	unsigned long flags;
1459	char dummy;
1460
1461	if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1462	    !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1463		DMWARN("invalid PG number supplied to switch_pg_num");
1464		return -EINVAL;
1465	}
1466
1467	spin_lock_irqsave(&m->lock, flags);
1468	list_for_each_entry(pg, &m->priority_groups, list) {
1469		pg->bypassed = false;
1470		if (--pgnum)
1471			continue;
1472
1473		m->current_pgpath = NULL;
1474		m->current_pg = NULL;
1475		m->next_pg = pg;
1476	}
1477	spin_unlock_irqrestore(&m->lock, flags);
1478
1479	schedule_work(&m->trigger_event);
1480	return 0;
1481}
1482
1483/*
1484 * Set/clear bypassed status of a PG.
1485 * PGs are numbered upwards from 1 in the order they were declared.
1486 */
1487static int bypass_pg_num(struct multipath *m, const char *pgstr, bool bypassed)
1488{
1489	struct priority_group *pg;
1490	unsigned pgnum;
1491	char dummy;
1492
1493	if (!pgstr || (sscanf(pgstr, "%u%c", &pgnum, &dummy) != 1) || !pgnum ||
1494	    !m->nr_priority_groups || (pgnum > m->nr_priority_groups)) {
1495		DMWARN("invalid PG number supplied to bypass_pg");
1496		return -EINVAL;
1497	}
1498
1499	list_for_each_entry(pg, &m->priority_groups, list) {
1500		if (!--pgnum)
1501			break;
1502	}
1503
1504	bypass_pg(m, pg, bypassed);
1505	return 0;
1506}
1507
1508/*
1509 * Should we retry pg_init immediately?
1510 */
1511static bool pg_init_limit_reached(struct multipath *m, struct pgpath *pgpath)
1512{
1513	unsigned long flags;
1514	bool limit_reached = false;
1515
1516	spin_lock_irqsave(&m->lock, flags);
1517
1518	if (atomic_read(&m->pg_init_count) <= m->pg_init_retries &&
1519	    !test_bit(MPATHF_PG_INIT_DISABLED, &m->flags))
1520		set_bit(MPATHF_PG_INIT_REQUIRED, &m->flags);
1521	else
1522		limit_reached = true;
1523
1524	spin_unlock_irqrestore(&m->lock, flags);
1525
1526	return limit_reached;
1527}
1528
1529static void pg_init_done(void *data, int errors)
1530{
1531	struct pgpath *pgpath = data;
1532	struct priority_group *pg = pgpath->pg;
1533	struct multipath *m = pg->m;
1534	unsigned long flags;
1535	bool delay_retry = false;
1536
1537	/* device or driver problems */
1538	switch (errors) {
1539	case SCSI_DH_OK:
1540		break;
1541	case SCSI_DH_NOSYS:
1542		if (!m->hw_handler_name) {
1543			errors = 0;
1544			break;
1545		}
1546		DMERR("Could not failover the device: Handler scsi_dh_%s "
1547		      "Error %d.", m->hw_handler_name, errors);
1548		/*
1549		 * Fail path for now, so we do not ping pong
1550		 */
1551		fail_path(pgpath);
1552		break;
1553	case SCSI_DH_DEV_TEMP_BUSY:
1554		/*
1555		 * Probably doing something like FW upgrade on the
1556		 * controller so try the other pg.
1557		 */
1558		bypass_pg(m, pg, true);
1559		break;
1560	case SCSI_DH_RETRY:
1561		/* Wait before retrying. */
1562		delay_retry = true;
1563		fallthrough;
1564	case SCSI_DH_IMM_RETRY:
1565	case SCSI_DH_RES_TEMP_UNAVAIL:
1566		if (pg_init_limit_reached(m, pgpath))
1567			fail_path(pgpath);
1568		errors = 0;
1569		break;
1570	case SCSI_DH_DEV_OFFLINED:
1571	default:
1572		/*
1573		 * We probably do not want to fail the path for a device
1574		 * error, but this is what the old dm did. In future
1575		 * patches we can do more advanced handling.
1576		 */
1577		fail_path(pgpath);
1578	}
1579
1580	spin_lock_irqsave(&m->lock, flags);
1581	if (errors) {
1582		if (pgpath == m->current_pgpath) {
1583			DMERR("Could not failover device. Error %d.", errors);
1584			m->current_pgpath = NULL;
1585			m->current_pg = NULL;
1586		}
1587	} else if (!test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
1588		pg->bypassed = false;
1589
1590	if (atomic_dec_return(&m->pg_init_in_progress) > 0)
1591		/* Activations of other paths are still on going */
1592		goto out;
1593
1594	if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags)) {
1595		if (delay_retry)
1596			set_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1597		else
1598			clear_bit(MPATHF_PG_INIT_DELAY_RETRY, &m->flags);
1599
1600		if (__pg_init_all_paths(m))
1601			goto out;
1602	}
1603	clear_bit(MPATHF_QUEUE_IO, &m->flags);
1604
1605	process_queued_io_list(m);
1606
1607	/*
1608	 * Wake up any thread waiting to suspend.
1609	 */
1610	wake_up(&m->pg_init_wait);
1611
1612out:
1613	spin_unlock_irqrestore(&m->lock, flags);
1614}
1615
1616static void activate_or_offline_path(struct pgpath *pgpath)
1617{
1618	struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
1619
1620	if (pgpath->is_active && !blk_queue_dying(q))
1621		scsi_dh_activate(q, pg_init_done, pgpath);
1622	else
1623		pg_init_done(pgpath, SCSI_DH_DEV_OFFLINED);
1624}
1625
1626static void activate_path_work(struct work_struct *work)
1627{
1628	struct pgpath *pgpath =
1629		container_of(work, struct pgpath, activate_path.work);
1630
1631	activate_or_offline_path(pgpath);
 
1632}
1633
1634static int multipath_end_io(struct dm_target *ti, struct request *clone,
1635			    blk_status_t error, union map_info *map_context)
 
 
 
1636{
1637	struct dm_mpath_io *mpio = get_mpio(map_context);
1638	struct pgpath *pgpath = mpio->pgpath;
1639	int r = DM_ENDIO_DONE;
1640
1641	/*
1642	 * We don't queue any clone request inside the multipath target
1643	 * during end I/O handling, since those clone requests don't have
1644	 * bio clones.  If we queue them inside the multipath target,
1645	 * we need to make bio clones, that requires memory allocation.
1646	 * (See drivers/md/dm-rq.c:end_clone_bio() about why the clone requests
1647	 *  don't have bio clones.)
1648	 * Instead of queueing the clone request here, we queue the original
1649	 * request into dm core, which will remake a clone request and
1650	 * clone bios for it and resubmit it later.
1651	 */
1652	if (error && blk_path_error(error)) {
1653		struct multipath *m = ti->private;
 
 
 
1654
1655		if (error == BLK_STS_RESOURCE)
1656			r = DM_ENDIO_DELAY_REQUEUE;
1657		else
1658			r = DM_ENDIO_REQUEUE;
1659
1660		if (pgpath)
1661			fail_path(pgpath);
1662
1663		if (!atomic_read(&m->nr_valid_paths) &&
1664		    !must_push_back_rq(m)) {
1665			if (error == BLK_STS_IOERR)
1666				dm_report_EIO(m);
1667			/* complete with the original error */
1668			r = DM_ENDIO_DONE;
 
 
1669		}
1670	}
1671
1672	if (pgpath) {
1673		struct path_selector *ps = &pgpath->pg->ps;
1674
1675		if (ps->type->end_io)
1676			ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1677					 clone->io_start_time_ns);
1678	}
1679
1680	return r;
1681}
1682
1683static int multipath_end_io_bio(struct dm_target *ti, struct bio *clone,
1684				blk_status_t *error)
1685{
1686	struct multipath *m = ti->private;
1687	struct dm_mpath_io *mpio = get_mpio_from_bio(clone);
1688	struct pgpath *pgpath = mpio->pgpath;
1689	unsigned long flags;
1690	int r = DM_ENDIO_DONE;
1691
1692	if (!*error || !blk_path_error(*error))
1693		goto done;
1694
1695	if (pgpath)
1696		fail_path(pgpath);
1697
1698	if (!atomic_read(&m->nr_valid_paths)) {
1699		spin_lock_irqsave(&m->lock, flags);
1700		if (!test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
1701			if (__must_push_back(m)) {
1702				r = DM_ENDIO_REQUEUE;
1703			} else {
1704				dm_report_EIO(m);
1705				*error = BLK_STS_IOERR;
1706			}
1707			spin_unlock_irqrestore(&m->lock, flags);
1708			goto done;
1709		}
1710		spin_unlock_irqrestore(&m->lock, flags);
1711	}
1712
1713	multipath_queue_bio(m, clone);
1714	r = DM_ENDIO_INCOMPLETE;
1715done:
1716	if (pgpath) {
1717		struct path_selector *ps = &pgpath->pg->ps;
1718
1719		if (ps->type->end_io)
1720			ps->type->end_io(ps, &pgpath->path, mpio->nr_bytes,
1721					 dm_start_time_ns_from_clone(clone));
1722	}
 
1723
1724	return r;
1725}
1726
1727/*
1728 * Suspend with flush can't complete until all the I/O is processed
1729 * so if the last path fails we must error any remaining I/O.
1730 * - Note that if the freeze_bdev fails while suspending, the
1731 *   queue_if_no_path state is lost - userspace should reset it.
1732 * Otherwise, during noflush suspend, queue_if_no_path will not change.
1733 */
1734static void multipath_presuspend(struct dm_target *ti)
1735{
1736	struct multipath *m = ti->private;
1737
1738	/* FIXME: bio-based shouldn't need to always disable queue_if_no_path */
1739	if (m->queue_mode == DM_TYPE_BIO_BASED || !dm_noflush_suspending(m->ti))
1740		queue_if_no_path(m, false, true, __func__);
1741}
1742
1743static void multipath_postsuspend(struct dm_target *ti)
1744{
1745	struct multipath *m = ti->private;
1746
1747	mutex_lock(&m->work_mutex);
1748	flush_multipath_work(m);
1749	mutex_unlock(&m->work_mutex);
1750}
1751
1752/*
1753 * Restore the queue_if_no_path setting.
1754 */
1755static void multipath_resume(struct dm_target *ti)
1756{
1757	struct multipath *m = ti->private;
1758	unsigned long flags;
1759
1760	spin_lock_irqsave(&m->lock, flags);
1761	if (test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags)) {
1762		set_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags);
1763		clear_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags);
1764	}
1765
1766	DMDEBUG("%s: %s finished; QIFNP = %d; SQIFNP = %d",
1767		dm_table_device_name(m->ti->table), __func__,
1768		test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags),
1769		test_bit(MPATHF_SAVED_QUEUE_IF_NO_PATH, &m->flags));
1770
1771	spin_unlock_irqrestore(&m->lock, flags);
1772}
1773
1774/*
1775 * Info output has the following format:
1776 * num_multipath_feature_args [multipath_feature_args]*
1777 * num_handler_status_args [handler_status_args]*
1778 * num_groups init_group_number
1779 *            [A|D|E num_ps_status_args [ps_status_args]*
1780 *             num_paths num_selector_args
1781 *             [path_dev A|F fail_count [selector_args]* ]+ ]+
1782 *
1783 * Table output has the following format (identical to the constructor string):
1784 * num_feature_args [features_args]*
1785 * num_handler_args hw_handler [hw_handler_args]*
1786 * num_groups init_group_number
1787 *     [priority selector-name num_ps_args [ps_args]*
1788 *      num_paths num_selector_args [path_dev [selector_args]* ]+ ]+
1789 */
1790static void multipath_status(struct dm_target *ti, status_type_t type,
1791			     unsigned status_flags, char *result, unsigned maxlen)
1792{
1793	int sz = 0;
1794	unsigned long flags;
1795	struct multipath *m = ti->private;
1796	struct priority_group *pg;
1797	struct pgpath *p;
1798	unsigned pg_num;
1799	char state;
1800
1801	spin_lock_irqsave(&m->lock, flags);
1802
1803	/* Features */
1804	if (type == STATUSTYPE_INFO)
1805		DMEMIT("2 %u %u ", test_bit(MPATHF_QUEUE_IO, &m->flags),
1806		       atomic_read(&m->pg_init_count));
1807	else {
1808		DMEMIT("%u ", test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags) +
1809			      (m->pg_init_retries > 0) * 2 +
1810			      (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT) * 2 +
1811			      test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags) +
1812			      (m->queue_mode != DM_TYPE_REQUEST_BASED) * 2);
1813
1814		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
1815			DMEMIT("queue_if_no_path ");
1816		if (m->pg_init_retries)
1817			DMEMIT("pg_init_retries %u ", m->pg_init_retries);
1818		if (m->pg_init_delay_msecs != DM_PG_INIT_DELAY_DEFAULT)
1819			DMEMIT("pg_init_delay_msecs %u ", m->pg_init_delay_msecs);
1820		if (test_bit(MPATHF_RETAIN_ATTACHED_HW_HANDLER, &m->flags))
1821			DMEMIT("retain_attached_hw_handler ");
1822		if (m->queue_mode != DM_TYPE_REQUEST_BASED) {
1823			switch(m->queue_mode) {
1824			case DM_TYPE_BIO_BASED:
1825				DMEMIT("queue_mode bio ");
1826				break;
1827			default:
1828				WARN_ON_ONCE(true);
1829				break;
1830			}
1831		}
1832	}
1833
1834	if (!m->hw_handler_name || type == STATUSTYPE_INFO)
1835		DMEMIT("0 ");
1836	else
1837		DMEMIT("1 %s ", m->hw_handler_name);
1838
1839	DMEMIT("%u ", m->nr_priority_groups);
1840
1841	if (m->next_pg)
1842		pg_num = m->next_pg->pg_num;
1843	else if (m->current_pg)
1844		pg_num = m->current_pg->pg_num;
1845	else
1846		pg_num = (m->nr_priority_groups ? 1 : 0);
1847
1848	DMEMIT("%u ", pg_num);
1849
1850	switch (type) {
1851	case STATUSTYPE_INFO:
1852		list_for_each_entry(pg, &m->priority_groups, list) {
1853			if (pg->bypassed)
1854				state = 'D';	/* Disabled */
1855			else if (pg == m->current_pg)
1856				state = 'A';	/* Currently Active */
1857			else
1858				state = 'E';	/* Enabled */
1859
1860			DMEMIT("%c ", state);
1861
1862			if (pg->ps.type->status)
1863				sz += pg->ps.type->status(&pg->ps, NULL, type,
1864							  result + sz,
1865							  maxlen - sz);
1866			else
1867				DMEMIT("0 ");
1868
1869			DMEMIT("%u %u ", pg->nr_pgpaths,
1870			       pg->ps.type->info_args);
1871
1872			list_for_each_entry(p, &pg->pgpaths, list) {
1873				DMEMIT("%s %s %u ", p->path.dev->name,
1874				       p->is_active ? "A" : "F",
1875				       p->fail_count);
1876				if (pg->ps.type->status)
1877					sz += pg->ps.type->status(&pg->ps,
1878					      &p->path, type, result + sz,
1879					      maxlen - sz);
1880			}
1881		}
1882		break;
1883
1884	case STATUSTYPE_TABLE:
1885		list_for_each_entry(pg, &m->priority_groups, list) {
1886			DMEMIT("%s ", pg->ps.type->name);
1887
1888			if (pg->ps.type->status)
1889				sz += pg->ps.type->status(&pg->ps, NULL, type,
1890							  result + sz,
1891							  maxlen - sz);
1892			else
1893				DMEMIT("0 ");
1894
1895			DMEMIT("%u %u ", pg->nr_pgpaths,
1896			       pg->ps.type->table_args);
1897
1898			list_for_each_entry(p, &pg->pgpaths, list) {
1899				DMEMIT("%s ", p->path.dev->name);
1900				if (pg->ps.type->status)
1901					sz += pg->ps.type->status(&pg->ps,
1902					      &p->path, type, result + sz,
1903					      maxlen - sz);
1904			}
1905		}
1906		break;
1907	}
1908
1909	spin_unlock_irqrestore(&m->lock, flags);
 
 
1910}
1911
1912static int multipath_message(struct dm_target *ti, unsigned argc, char **argv,
1913			     char *result, unsigned maxlen)
1914{
1915	int r = -EINVAL;
1916	struct dm_dev *dev;
1917	struct multipath *m = ti->private;
1918	action_fn action;
1919	unsigned long flags;
1920
1921	mutex_lock(&m->work_mutex);
1922
1923	if (dm_suspended(ti)) {
1924		r = -EBUSY;
1925		goto out;
1926	}
1927
1928	if (argc == 1) {
1929		if (!strcasecmp(argv[0], "queue_if_no_path")) {
1930			r = queue_if_no_path(m, true, false, __func__);
1931			spin_lock_irqsave(&m->lock, flags);
1932			enable_nopath_timeout(m);
1933			spin_unlock_irqrestore(&m->lock, flags);
1934			goto out;
1935		} else if (!strcasecmp(argv[0], "fail_if_no_path")) {
1936			r = queue_if_no_path(m, false, false, __func__);
1937			disable_nopath_timeout(m);
1938			goto out;
1939		}
1940	}
1941
1942	if (argc != 2) {
1943		DMWARN("Invalid multipath message arguments. Expected 2 arguments, got %d.", argc);
1944		goto out;
1945	}
1946
1947	if (!strcasecmp(argv[0], "disable_group")) {
1948		r = bypass_pg_num(m, argv[1], true);
1949		goto out;
1950	} else if (!strcasecmp(argv[0], "enable_group")) {
1951		r = bypass_pg_num(m, argv[1], false);
1952		goto out;
1953	} else if (!strcasecmp(argv[0], "switch_group")) {
1954		r = switch_pg_num(m, argv[1]);
1955		goto out;
1956	} else if (!strcasecmp(argv[0], "reinstate_path"))
1957		action = reinstate_path;
1958	else if (!strcasecmp(argv[0], "fail_path"))
1959		action = fail_path;
1960	else {
1961		DMWARN("Unrecognised multipath message received: %s", argv[0]);
1962		goto out;
1963	}
1964
1965	r = dm_get_device(ti, argv[1], dm_table_get_mode(ti->table), &dev);
1966	if (r) {
1967		DMWARN("message: error getting device %s",
1968		       argv[1]);
1969		goto out;
1970	}
1971
1972	r = action_dev(m, dev, action);
1973
1974	dm_put_device(ti, dev);
1975
1976out:
1977	mutex_unlock(&m->work_mutex);
1978	return r;
1979}
1980
1981static int multipath_prepare_ioctl(struct dm_target *ti,
1982				   struct block_device **bdev)
1983{
1984	struct multipath *m = ti->private;
1985	struct pgpath *pgpath;
 
1986	unsigned long flags;
1987	int r;
 
 
 
 
 
1988
1989	pgpath = READ_ONCE(m->current_pgpath);
1990	if (!pgpath || !mpath_double_check_test_bit(MPATHF_QUEUE_IO, m))
1991		pgpath = choose_pgpath(m, 0);
 
1992
1993	if (pgpath) {
1994		if (!mpath_double_check_test_bit(MPATHF_QUEUE_IO, m)) {
1995			*bdev = pgpath->path.dev->bdev;
1996			r = 0;
1997		} else {
1998			/* pg_init has not started or completed */
1999			r = -ENOTCONN;
2000		}
2001	} else {
2002		/* No path is available */
2003		r = -EIO;
2004		spin_lock_irqsave(&m->lock, flags);
2005		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags))
2006			r = -ENOTCONN;
2007		spin_unlock_irqrestore(&m->lock, flags);
2008	}
2009
2010	if (r == -ENOTCONN) {
2011		if (!READ_ONCE(m->current_pg)) {
2012			/* Path status changed, redo selection */
2013			(void) choose_pgpath(m, 0);
2014		}
2015		spin_lock_irqsave(&m->lock, flags);
2016		if (test_bit(MPATHF_PG_INIT_REQUIRED, &m->flags))
2017			(void) __pg_init_all_paths(m);
2018		spin_unlock_irqrestore(&m->lock, flags);
2019		dm_table_run_md_queue_async(m->ti->table);
2020		process_queued_io_list(m);
2021	}
2022
2023	/*
2024	 * Only pass ioctls through if the device sizes match exactly.
2025	 */
2026	if (!r && ti->len != i_size_read((*bdev)->bd_inode) >> SECTOR_SHIFT)
2027		return 1;
2028	return r;
2029}
2030
2031static int multipath_iterate_devices(struct dm_target *ti,
2032				     iterate_devices_callout_fn fn, void *data)
2033{
2034	struct multipath *m = ti->private;
2035	struct priority_group *pg;
2036	struct pgpath *p;
2037	int ret = 0;
2038
2039	list_for_each_entry(pg, &m->priority_groups, list) {
2040		list_for_each_entry(p, &pg->pgpaths, list) {
2041			ret = fn(ti, p->path.dev, ti->begin, ti->len, data);
2042			if (ret)
2043				goto out;
2044		}
2045	}
2046
2047out:
2048	return ret;
2049}
2050
2051static int pgpath_busy(struct pgpath *pgpath)
2052{
2053	struct request_queue *q = bdev_get_queue(pgpath->path.dev->bdev);
2054
2055	return blk_lld_busy(q);
2056}
2057
2058/*
2059 * We return "busy", only when we can map I/Os but underlying devices
2060 * are busy (so even if we map I/Os now, the I/Os will wait on
2061 * the underlying queue).
2062 * In other words, if we want to kill I/Os or queue them inside us
2063 * due to map unavailability, we don't return "busy".  Otherwise,
2064 * dm core won't give us the I/Os and we can't do what we want.
2065 */
2066static int multipath_busy(struct dm_target *ti)
2067{
2068	bool busy = false, has_active = false;
2069	struct multipath *m = ti->private;
2070	struct priority_group *pg, *next_pg;
2071	struct pgpath *pgpath;
 
2072
2073	/* pg_init in progress */
2074	if (atomic_read(&m->pg_init_in_progress))
2075		return true;
2076
2077	/* no paths available, for blk-mq: rely on IO mapping to delay requeue */
2078	if (!atomic_read(&m->nr_valid_paths)) {
2079		unsigned long flags;
2080		spin_lock_irqsave(&m->lock, flags);
2081		if (test_bit(MPATHF_QUEUE_IF_NO_PATH, &m->flags)) {
2082			spin_unlock_irqrestore(&m->lock, flags);
2083			return (m->queue_mode != DM_TYPE_REQUEST_BASED);
2084		}
2085		spin_unlock_irqrestore(&m->lock, flags);
2086	}
2087
2088	/* Guess which priority_group will be used at next mapping time */
2089	pg = READ_ONCE(m->current_pg);
2090	next_pg = READ_ONCE(m->next_pg);
2091	if (unlikely(!READ_ONCE(m->current_pgpath) && next_pg))
2092		pg = next_pg;
2093
2094	if (!pg) {
2095		/*
2096		 * We don't know which pg will be used at next mapping time.
2097		 * We don't call choose_pgpath() here to avoid to trigger
2098		 * pg_init just by busy checking.
2099		 * So we don't know whether underlying devices we will be using
2100		 * at next mapping time are busy or not. Just try mapping.
2101		 */
2102		return busy;
2103	}
2104
2105	/*
2106	 * If there is one non-busy active path at least, the path selector
2107	 * will be able to select it. So we consider such a pg as not busy.
2108	 */
2109	busy = true;
2110	list_for_each_entry(pgpath, &pg->pgpaths, list) {
2111		if (pgpath->is_active) {
2112			has_active = true;
2113			if (!pgpath_busy(pgpath)) {
2114				busy = false;
 
2115				break;
2116			}
2117		}
2118	}
2119
2120	if (!has_active) {
2121		/*
2122		 * No active path in this pg, so this pg won't be used and
2123		 * the current_pg will be changed at next mapping time.
2124		 * We need to try mapping to determine it.
2125		 */
2126		busy = false;
2127	}
 
 
2128
2129	return busy;
2130}
2131
2132/*-----------------------------------------------------------------
2133 * Module setup
2134 *---------------------------------------------------------------*/
2135static struct target_type multipath_target = {
2136	.name = "multipath",
2137	.version = {1, 14, 0},
2138	.features = DM_TARGET_SINGLETON | DM_TARGET_IMMUTABLE |
2139		    DM_TARGET_PASSES_INTEGRITY,
2140	.module = THIS_MODULE,
2141	.ctr = multipath_ctr,
2142	.dtr = multipath_dtr,
2143	.clone_and_map_rq = multipath_clone_and_map,
2144	.release_clone_rq = multipath_release_clone,
2145	.rq_end_io = multipath_end_io,
2146	.map = multipath_map_bio,
2147	.end_io = multipath_end_io_bio,
2148	.presuspend = multipath_presuspend,
2149	.postsuspend = multipath_postsuspend,
2150	.resume = multipath_resume,
2151	.status = multipath_status,
2152	.message = multipath_message,
2153	.prepare_ioctl = multipath_prepare_ioctl,
2154	.iterate_devices = multipath_iterate_devices,
2155	.busy = multipath_busy,
2156};
2157
2158static int __init dm_multipath_init(void)
2159{
2160	int r;
2161
 
 
 
 
 
 
 
 
 
 
 
 
2162	kmultipathd = alloc_workqueue("kmpathd", WQ_MEM_RECLAIM, 0);
2163	if (!kmultipathd) {
2164		DMERR("failed to create workqueue kmpathd");
2165		r = -ENOMEM;
2166		goto bad_alloc_kmultipathd;
 
2167	}
2168
2169	/*
2170	 * A separate workqueue is used to handle the device handlers
2171	 * to avoid overloading existing workqueue. Overloading the
2172	 * old workqueue would also create a bottleneck in the
2173	 * path of the storage hardware device activation.
2174	 */
2175	kmpath_handlerd = alloc_ordered_workqueue("kmpath_handlerd",
2176						  WQ_MEM_RECLAIM);
2177	if (!kmpath_handlerd) {
2178		DMERR("failed to create workqueue kmpath_handlerd");
2179		r = -ENOMEM;
2180		goto bad_alloc_kmpath_handlerd;
2181	}
2182
2183	r = dm_register_target(&multipath_target);
2184	if (r < 0) {
2185		DMERR("request-based register failed %d", r);
2186		r = -EINVAL;
2187		goto bad_register_target;
2188	}
2189
2190	return 0;
 
 
2191
2192bad_register_target:
2193	destroy_workqueue(kmpath_handlerd);
2194bad_alloc_kmpath_handlerd:
2195	destroy_workqueue(kmultipathd);
2196bad_alloc_kmultipathd:
2197	return r;
2198}
2199
2200static void __exit dm_multipath_exit(void)
2201{
2202	destroy_workqueue(kmpath_handlerd);
2203	destroy_workqueue(kmultipathd);
2204
2205	dm_unregister_target(&multipath_target);
 
2206}
2207
2208module_init(dm_multipath_init);
2209module_exit(dm_multipath_exit);
2210
2211module_param_named(queue_if_no_path_timeout_secs,
2212		   queue_if_no_path_timeout_secs, ulong, S_IRUGO | S_IWUSR);
2213MODULE_PARM_DESC(queue_if_no_path_timeout_secs, "No available paths queue IO timeout in seconds");
2214
2215MODULE_DESCRIPTION(DM_NAME " multipath target");
2216MODULE_AUTHOR("Sistina Software <dm-devel@redhat.com>");
2217MODULE_LICENSE("GPL");