1 /*
2  * Generic SCSI-3 ALUA SCSI Device Handler
3  *
4  * Copyright (C) 2007-2010 Hannes Reinecke, SUSE Linux Products GmbH.
5  * All rights reserved.
6  *
7  * This program is free software; you can redistribute it and/or modify
8  * it under the terms of the GNU General Public License as published by
9  * the Free Software Foundation; either version 2 of the License, or
10  * (at your option) any later version.
11  *
12  * This program is distributed in the hope that it will be useful,
13  * but WITHOUT ANY WARRANTY; without even the implied warranty of
14  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
15  * GNU General Public License for more details.
16  *
17  * You should have received a copy of the GNU General Public License
18  * along with this program; if not, write to the Free Software
19  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
20  *
21  */
22 #include <linux/slab.h>
23 #include <linux/delay.h>
24 #include <linux/module.h>
25 #include <asm/unaligned.h>
26 #include <scsi/scsi.h>
27 #include <scsi/scsi_proto.h>
28 #include <scsi/scsi_dbg.h>
29 #include <scsi/scsi_eh.h>
30 #include <scsi/scsi_dh.h>
31 
32 #define ALUA_DH_NAME "alua"
33 #define ALUA_DH_VER "2.0"
34 
35 #define TPGS_SUPPORT_NONE		0x00
36 #define TPGS_SUPPORT_OPTIMIZED		0x01
37 #define TPGS_SUPPORT_NONOPTIMIZED	0x02
38 #define TPGS_SUPPORT_STANDBY		0x04
39 #define TPGS_SUPPORT_UNAVAILABLE	0x08
40 #define TPGS_SUPPORT_LBA_DEPENDENT	0x10
41 #define TPGS_SUPPORT_OFFLINE		0x40
42 #define TPGS_SUPPORT_TRANSITION		0x80
43 #define TPGS_SUPPORT_ALL		0xdf
44 
45 #define RTPG_FMT_MASK			0x70
46 #define RTPG_FMT_EXT_HDR		0x10
47 
48 #define TPGS_MODE_UNINITIALIZED		 -1
49 #define TPGS_MODE_NONE			0x0
50 #define TPGS_MODE_IMPLICIT		0x1
51 #define TPGS_MODE_EXPLICIT		0x2
52 
53 #define ALUA_RTPG_SIZE			128
54 #define ALUA_FAILOVER_TIMEOUT		60
55 #define ALUA_FAILOVER_RETRIES		5
56 #define ALUA_RTPG_DELAY_MSECS		5
57 
58 /* device handler flags */
59 #define ALUA_OPTIMIZE_STPG		0x01
60 #define ALUA_RTPG_EXT_HDR_UNSUPP	0x02
61 /* State machine flags */
62 #define ALUA_PG_RUN_RTPG		0x10
63 #define ALUA_PG_RUN_STPG		0x20
64 #define ALUA_PG_RUNNING			0x40
65 
66 static uint optimize_stpg;
67 module_param(optimize_stpg, uint, S_IRUGO|S_IWUSR);
68 MODULE_PARM_DESC(optimize_stpg, "Allow use of a non-optimized path, rather than sending a STPG, when implicit TPGS is supported (0=No,1=Yes). Default is 0.");
69 
70 static LIST_HEAD(port_group_list);
71 static DEFINE_SPINLOCK(port_group_lock);
72 static struct workqueue_struct *kaluad_wq;
73 
74 struct alua_port_group {
75 	struct kref		kref;
76 	struct rcu_head		rcu;
77 	struct list_head	node;
78 	struct list_head	dh_list;
79 	unsigned char		device_id_str[256];
80 	int			device_id_len;
81 	int			group_id;
82 	int			tpgs;
83 	int			state;
84 	int			pref;
85 	int			valid_states;
86 	unsigned		flags; /* used for optimizing STPG */
87 	unsigned char		transition_tmo;
88 	unsigned long		expiry;
89 	unsigned long		interval;
90 	struct delayed_work	rtpg_work;
91 	spinlock_t		lock;
92 	struct list_head	rtpg_list;
93 	struct scsi_device	*rtpg_sdev;
94 };
95 
96 struct alua_dh_data {
97 	struct list_head	node;
98 	struct alua_port_group __rcu *pg;
99 	int			group_id;
100 	spinlock_t		pg_lock;
101 	struct scsi_device	*sdev;
102 	int			init_error;
103 	struct mutex		init_mutex;
104 };
105 
106 struct alua_queue_data {
107 	struct list_head	entry;
108 	activate_complete	callback_fn;
109 	void			*callback_data;
110 };
111 
112 #define ALUA_POLICY_SWITCH_CURRENT	0
113 #define ALUA_POLICY_SWITCH_ALL		1
114 
115 static void alua_rtpg_work(struct work_struct *work);
116 static bool alua_rtpg_queue(struct alua_port_group *pg,
117 			    struct scsi_device *sdev,
118 			    struct alua_queue_data *qdata, bool force);
119 static void alua_check(struct scsi_device *sdev, bool force);
120 
121 static void release_port_group(struct kref *kref)
122 {
123 	struct alua_port_group *pg;
124 
125 	pg = container_of(kref, struct alua_port_group, kref);
126 	if (pg->rtpg_sdev)
127 		flush_delayed_work(&pg->rtpg_work);
128 	spin_lock(&port_group_lock);
129 	list_del(&pg->node);
130 	spin_unlock(&port_group_lock);
131 	kfree_rcu(pg, rcu);
132 }
133 
134 /*
135  * submit_rtpg - Issue a REPORT TARGET GROUP STATES command
136  * @sdev: sdev the command should be sent to
137  */
138 static int submit_rtpg(struct scsi_device *sdev, unsigned char *buff,
139 		       int bufflen, struct scsi_sense_hdr *sshdr, int flags)
140 {
141 	u8 cdb[MAX_COMMAND_SIZE];
142 	int req_flags = REQ_FAILFAST_DEV | REQ_FAILFAST_TRANSPORT |
143 		REQ_FAILFAST_DRIVER;
144 
145 	/* Prepare the command. */
146 	memset(cdb, 0x0, MAX_COMMAND_SIZE);
147 	cdb[0] = MAINTENANCE_IN;
148 	if (!(flags & ALUA_RTPG_EXT_HDR_UNSUPP))
149 		cdb[1] = MI_REPORT_TARGET_PGS | MI_EXT_HDR_PARAM_FMT;
150 	else
151 		cdb[1] = MI_REPORT_TARGET_PGS;
152 	put_unaligned_be32(bufflen, &cdb[6]);
153 
154 	return scsi_execute(sdev, cdb, DMA_FROM_DEVICE, buff, bufflen, NULL,
155 			sshdr, ALUA_FAILOVER_TIMEOUT * HZ,
156 			ALUA_FAILOVER_RETRIES, req_flags, 0, NULL);
157 }
158 
159 /*
160  * submit_stpg - Issue a SET TARGET PORT GROUP command
161  *
162  * Currently we're only setting the current target port group state
163  * to 'active/optimized' and let the array firmware figure out
164  * the states of the remaining groups.
165  */
166 static int submit_stpg(struct scsi_device *sdev, int group_id,
167 		       struct scsi_sense_hdr *sshdr)
168 {
169 	u8 cdb[MAX_COMMAND_SIZE];
170 	unsigned char stpg_data[8];
171 	int stpg_len = 8;
172 	int req_flags = REQ_FAILFAST_DEV | REQ_FAILFAST_TRANSPORT |
173 		REQ_FAILFAST_DRIVER;
174 
175 	/* Prepare the data buffer */
176 	memset(stpg_data, 0, stpg_len);
177 	stpg_data[4] = SCSI_ACCESS_STATE_OPTIMAL;
178 	put_unaligned_be16(group_id, &stpg_data[6]);
179 
180 	/* Prepare the command. */
181 	memset(cdb, 0x0, MAX_COMMAND_SIZE);
182 	cdb[0] = MAINTENANCE_OUT;
183 	cdb[1] = MO_SET_TARGET_PGS;
184 	put_unaligned_be32(stpg_len, &cdb[6]);
185 
186 	return scsi_execute(sdev, cdb, DMA_TO_DEVICE, stpg_data, stpg_len, NULL,
187 			sshdr, ALUA_FAILOVER_TIMEOUT * HZ,
188 			ALUA_FAILOVER_RETRIES, req_flags, 0, NULL);
189 }
190 
191 static struct alua_port_group *alua_find_get_pg(char *id_str, size_t id_size,
192 						int group_id)
193 {
194 	struct alua_port_group *pg;
195 
196 	if (!id_str || !id_size || !strlen(id_str))
197 		return NULL;
198 
199 	list_for_each_entry(pg, &port_group_list, node) {
200 		if (pg->group_id != group_id)
201 			continue;
202 		if (!pg->device_id_len || pg->device_id_len != id_size)
203 			continue;
204 		if (strncmp(pg->device_id_str, id_str, id_size))
205 			continue;
206 		if (!kref_get_unless_zero(&pg->kref))
207 			continue;
208 		return pg;
209 	}
210 
211 	return NULL;
212 }
213 
214 /*
215  * alua_alloc_pg - Allocate a new port_group structure
216  * @sdev: scsi device
217  * @group_id: port group id
218  * @tpgs: target port group settings
219  *
220  * Allocate a new port_group structure for a given
221  * device.
222  */
223 static struct alua_port_group *alua_alloc_pg(struct scsi_device *sdev,
224 					     int group_id, int tpgs)
225 {
226 	struct alua_port_group *pg, *tmp_pg;
227 
228 	pg = kzalloc(sizeof(struct alua_port_group), GFP_KERNEL);
229 	if (!pg)
230 		return ERR_PTR(-ENOMEM);
231 
232 	pg->device_id_len = scsi_vpd_lun_id(sdev, pg->device_id_str,
233 					    sizeof(pg->device_id_str));
234 	if (pg->device_id_len <= 0) {
235 		/*
236 		 * TPGS supported but no device identification found.
237 		 * Generate private device identification.
238 		 */
239 		sdev_printk(KERN_INFO, sdev,
240 			    "%s: No device descriptors found\n",
241 			    ALUA_DH_NAME);
242 		pg->device_id_str[0] = '\0';
243 		pg->device_id_len = 0;
244 	}
245 	pg->group_id = group_id;
246 	pg->tpgs = tpgs;
247 	pg->state = SCSI_ACCESS_STATE_OPTIMAL;
248 	pg->valid_states = TPGS_SUPPORT_ALL;
249 	if (optimize_stpg)
250 		pg->flags |= ALUA_OPTIMIZE_STPG;
251 	kref_init(&pg->kref);
252 	INIT_DELAYED_WORK(&pg->rtpg_work, alua_rtpg_work);
253 	INIT_LIST_HEAD(&pg->rtpg_list);
254 	INIT_LIST_HEAD(&pg->node);
255 	INIT_LIST_HEAD(&pg->dh_list);
256 	spin_lock_init(&pg->lock);
257 
258 	spin_lock(&port_group_lock);
259 	tmp_pg = alua_find_get_pg(pg->device_id_str, pg->device_id_len,
260 				  group_id);
261 	if (tmp_pg) {
262 		spin_unlock(&port_group_lock);
263 		kfree(pg);
264 		return tmp_pg;
265 	}
266 
267 	list_add(&pg->node, &port_group_list);
268 	spin_unlock(&port_group_lock);
269 
270 	return pg;
271 }
272 
273 /*
274  * alua_check_tpgs - Evaluate TPGS setting
275  * @sdev: device to be checked
276  *
277  * Examine the TPGS setting of the sdev to find out if ALUA
278  * is supported.
279  */
280 static int alua_check_tpgs(struct scsi_device *sdev)
281 {
282 	int tpgs = TPGS_MODE_NONE;
283 
284 	/*
285 	 * ALUA support for non-disk devices is fraught with
286 	 * difficulties, so disable it for now.
287 	 */
288 	if (sdev->type != TYPE_DISK) {
289 		sdev_printk(KERN_INFO, sdev,
290 			    "%s: disable for non-disk devices\n",
291 			    ALUA_DH_NAME);
292 		return tpgs;
293 	}
294 
295 	tpgs = scsi_device_tpgs(sdev);
296 	switch (tpgs) {
297 	case TPGS_MODE_EXPLICIT|TPGS_MODE_IMPLICIT:
298 		sdev_printk(KERN_INFO, sdev,
299 			    "%s: supports implicit and explicit TPGS\n",
300 			    ALUA_DH_NAME);
301 		break;
302 	case TPGS_MODE_EXPLICIT:
303 		sdev_printk(KERN_INFO, sdev, "%s: supports explicit TPGS\n",
304 			    ALUA_DH_NAME);
305 		break;
306 	case TPGS_MODE_IMPLICIT:
307 		sdev_printk(KERN_INFO, sdev, "%s: supports implicit TPGS\n",
308 			    ALUA_DH_NAME);
309 		break;
310 	case TPGS_MODE_NONE:
311 		sdev_printk(KERN_INFO, sdev, "%s: not supported\n",
312 			    ALUA_DH_NAME);
313 		break;
314 	default:
315 		sdev_printk(KERN_INFO, sdev,
316 			    "%s: unsupported TPGS setting %d\n",
317 			    ALUA_DH_NAME, tpgs);
318 		tpgs = TPGS_MODE_NONE;
319 		break;
320 	}
321 
322 	return tpgs;
323 }
324 
325 /*
326  * alua_check_vpd - Evaluate INQUIRY vpd page 0x83
327  * @sdev: device to be checked
328  *
329  * Extract the relative target port and the target port group
330  * descriptor from the list of identificators.
331  */
332 static int alua_check_vpd(struct scsi_device *sdev, struct alua_dh_data *h,
333 			  int tpgs)
334 {
335 	int rel_port = -1, group_id;
336 	struct alua_port_group *pg, *old_pg = NULL;
337 	bool pg_updated = false;
338 	unsigned long flags;
339 
340 	group_id = scsi_vpd_tpg_id(sdev, &rel_port);
341 	if (group_id < 0) {
342 		/*
343 		 * Internal error; TPGS supported but required
344 		 * VPD identification descriptors not present.
345 		 * Disable ALUA support
346 		 */
347 		sdev_printk(KERN_INFO, sdev,
348 			    "%s: No target port descriptors found\n",
349 			    ALUA_DH_NAME);
350 		return SCSI_DH_DEV_UNSUPP;
351 	}
352 
353 	pg = alua_alloc_pg(sdev, group_id, tpgs);
354 	if (IS_ERR(pg)) {
355 		if (PTR_ERR(pg) == -ENOMEM)
356 			return SCSI_DH_NOMEM;
357 		return SCSI_DH_DEV_UNSUPP;
358 	}
359 	if (pg->device_id_len)
360 		sdev_printk(KERN_INFO, sdev,
361 			    "%s: device %s port group %x rel port %x\n",
362 			    ALUA_DH_NAME, pg->device_id_str,
363 			    group_id, rel_port);
364 	else
365 		sdev_printk(KERN_INFO, sdev,
366 			    "%s: port group %x rel port %x\n",
367 			    ALUA_DH_NAME, group_id, rel_port);
368 
369 	/* Check for existing port group references */
370 	spin_lock(&h->pg_lock);
371 	old_pg = rcu_dereference_protected(h->pg, lockdep_is_held(&h->pg_lock));
372 	if (old_pg != pg) {
373 		/* port group has changed. Update to new port group */
374 		if (h->pg) {
375 			spin_lock_irqsave(&old_pg->lock, flags);
376 			list_del_rcu(&h->node);
377 			spin_unlock_irqrestore(&old_pg->lock, flags);
378 		}
379 		rcu_assign_pointer(h->pg, pg);
380 		pg_updated = true;
381 	}
382 
383 	spin_lock_irqsave(&pg->lock, flags);
384 	if (pg_updated)
385 		list_add_rcu(&h->node, &pg->dh_list);
386 	spin_unlock_irqrestore(&pg->lock, flags);
387 
388 	alua_rtpg_queue(rcu_dereference_protected(h->pg,
389 						  lockdep_is_held(&h->pg_lock)),
390 			sdev, NULL, true);
391 	spin_unlock(&h->pg_lock);
392 
393 	if (old_pg)
394 		kref_put(&old_pg->kref, release_port_group);
395 
396 	return SCSI_DH_OK;
397 }
398 
399 static char print_alua_state(unsigned char state)
400 {
401 	switch (state) {
402 	case SCSI_ACCESS_STATE_OPTIMAL:
403 		return 'A';
404 	case SCSI_ACCESS_STATE_ACTIVE:
405 		return 'N';
406 	case SCSI_ACCESS_STATE_STANDBY:
407 		return 'S';
408 	case SCSI_ACCESS_STATE_UNAVAILABLE:
409 		return 'U';
410 	case SCSI_ACCESS_STATE_LBA:
411 		return 'L';
412 	case SCSI_ACCESS_STATE_OFFLINE:
413 		return 'O';
414 	case SCSI_ACCESS_STATE_TRANSITIONING:
415 		return 'T';
416 	default:
417 		return 'X';
418 	}
419 }
420 
421 static int alua_check_sense(struct scsi_device *sdev,
422 			    struct scsi_sense_hdr *sense_hdr)
423 {
424 	switch (sense_hdr->sense_key) {
425 	case NOT_READY:
426 		if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0a) {
427 			/*
428 			 * LUN Not Accessible - ALUA state transition
429 			 */
430 			alua_check(sdev, false);
431 			return NEEDS_RETRY;
432 		}
433 		break;
434 	case UNIT_ATTENTION:
435 		if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x00) {
436 			/*
437 			 * Power On, Reset, or Bus Device Reset.
438 			 * Might have obscured a state transition,
439 			 * so schedule a recheck.
440 			 */
441 			alua_check(sdev, true);
442 			return ADD_TO_MLQUEUE;
443 		}
444 		if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x04)
445 			/*
446 			 * Device internal reset
447 			 */
448 			return ADD_TO_MLQUEUE;
449 		if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x01)
450 			/*
451 			 * Mode Parameters Changed
452 			 */
453 			return ADD_TO_MLQUEUE;
454 		if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x06) {
455 			/*
456 			 * ALUA state changed
457 			 */
458 			alua_check(sdev, true);
459 			return ADD_TO_MLQUEUE;
460 		}
461 		if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x07) {
462 			/*
463 			 * Implicit ALUA state transition failed
464 			 */
465 			alua_check(sdev, true);
466 			return ADD_TO_MLQUEUE;
467 		}
468 		if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x03)
469 			/*
470 			 * Inquiry data has changed
471 			 */
472 			return ADD_TO_MLQUEUE;
473 		if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x0e)
474 			/*
475 			 * REPORTED_LUNS_DATA_HAS_CHANGED is reported
476 			 * when switching controllers on targets like
477 			 * Intel Multi-Flex. We can just retry.
478 			 */
479 			return ADD_TO_MLQUEUE;
480 		break;
481 	}
482 
483 	return SCSI_RETURN_NOT_HANDLED;
484 }
485 
486 /*
487  * alua_tur - Send a TEST UNIT READY
488  * @sdev: device to which the TEST UNIT READY command should be send
489  *
490  * Send a TEST UNIT READY to @sdev to figure out the device state
491  * Returns SCSI_DH_RETRY if the sense code is NOT READY/ALUA TRANSITIONING,
492  * SCSI_DH_OK if no error occurred, and SCSI_DH_IO otherwise.
493  */
494 static int alua_tur(struct scsi_device *sdev)
495 {
496 	struct scsi_sense_hdr sense_hdr;
497 	int retval;
498 
499 	retval = scsi_test_unit_ready(sdev, ALUA_FAILOVER_TIMEOUT * HZ,
500 				      ALUA_FAILOVER_RETRIES, &sense_hdr);
501 	if (sense_hdr.sense_key == NOT_READY &&
502 	    sense_hdr.asc == 0x04 && sense_hdr.ascq == 0x0a)
503 		return SCSI_DH_RETRY;
504 	else if (retval)
505 		return SCSI_DH_IO;
506 	else
507 		return SCSI_DH_OK;
508 }
509 
510 /*
511  * alua_rtpg - Evaluate REPORT TARGET GROUP STATES
512  * @sdev: the device to be evaluated.
513  *
514  * Evaluate the Target Port Group State.
515  * Returns SCSI_DH_DEV_OFFLINED if the path is
516  * found to be unusable.
517  */
518 static int alua_rtpg(struct scsi_device *sdev, struct alua_port_group *pg)
519 {
520 	struct scsi_sense_hdr sense_hdr;
521 	struct alua_port_group *tmp_pg;
522 	int len, k, off, bufflen = ALUA_RTPG_SIZE;
523 	unsigned char *desc, *buff;
524 	unsigned err, retval;
525 	unsigned int tpg_desc_tbl_off;
526 	unsigned char orig_transition_tmo;
527 	unsigned long flags;
528 
529 	if (!pg->expiry) {
530 		unsigned long transition_tmo = ALUA_FAILOVER_TIMEOUT * HZ;
531 
532 		if (pg->transition_tmo)
533 			transition_tmo = pg->transition_tmo * HZ;
534 
535 		pg->expiry = round_jiffies_up(jiffies + transition_tmo);
536 	}
537 
538 	buff = kzalloc(bufflen, GFP_KERNEL);
539 	if (!buff)
540 		return SCSI_DH_DEV_TEMP_BUSY;
541 
542  retry:
543 	err = 0;
544 	retval = submit_rtpg(sdev, buff, bufflen, &sense_hdr, pg->flags);
545 
546 	if (retval) {
547 		/*
548 		 * Some (broken) implementations have a habit of returning
549 		 * an error during things like firmware update etc.
550 		 * But if the target only supports active/optimized there's
551 		 * not much we can do; it's not that we can switch paths
552 		 * or anything.
553 		 * So ignore any errors to avoid spurious failures during
554 		 * path failover.
555 		 */
556 		if ((pg->valid_states & ~TPGS_SUPPORT_OPTIMIZED) == 0) {
557 			sdev_printk(KERN_INFO, sdev,
558 				    "%s: ignoring rtpg result %d\n",
559 				    ALUA_DH_NAME, retval);
560 			kfree(buff);
561 			return SCSI_DH_OK;
562 		}
563 		if (!scsi_sense_valid(&sense_hdr)) {
564 			sdev_printk(KERN_INFO, sdev,
565 				    "%s: rtpg failed, result %d\n",
566 				    ALUA_DH_NAME, retval);
567 			kfree(buff);
568 			if (driver_byte(retval) == DRIVER_ERROR)
569 				return SCSI_DH_DEV_TEMP_BUSY;
570 			return SCSI_DH_IO;
571 		}
572 
573 		/*
574 		 * submit_rtpg() has failed on existing arrays
575 		 * when requesting extended header info, and
576 		 * the array doesn't support extended headers,
577 		 * even though it shouldn't according to T10.
578 		 * The retry without rtpg_ext_hdr_req set
579 		 * handles this.
580 		 */
581 		if (!(pg->flags & ALUA_RTPG_EXT_HDR_UNSUPP) &&
582 		    sense_hdr.sense_key == ILLEGAL_REQUEST &&
583 		    sense_hdr.asc == 0x24 && sense_hdr.ascq == 0) {
584 			pg->flags |= ALUA_RTPG_EXT_HDR_UNSUPP;
585 			goto retry;
586 		}
587 		/*
588 		 * Retry on ALUA state transition or if any
589 		 * UNIT ATTENTION occurred.
590 		 */
591 		if (sense_hdr.sense_key == NOT_READY &&
592 		    sense_hdr.asc == 0x04 && sense_hdr.ascq == 0x0a)
593 			err = SCSI_DH_RETRY;
594 		else if (sense_hdr.sense_key == UNIT_ATTENTION)
595 			err = SCSI_DH_RETRY;
596 		if (err == SCSI_DH_RETRY &&
597 		    pg->expiry != 0 && time_before(jiffies, pg->expiry)) {
598 			sdev_printk(KERN_ERR, sdev, "%s: rtpg retry\n",
599 				    ALUA_DH_NAME);
600 			scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
601 			kfree(buff);
602 			return err;
603 		}
604 		sdev_printk(KERN_ERR, sdev, "%s: rtpg failed\n",
605 			    ALUA_DH_NAME);
606 		scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
607 		kfree(buff);
608 		pg->expiry = 0;
609 		return SCSI_DH_IO;
610 	}
611 
612 	len = get_unaligned_be32(&buff[0]) + 4;
613 
614 	if (len > bufflen) {
615 		/* Resubmit with the correct length */
616 		kfree(buff);
617 		bufflen = len;
618 		buff = kmalloc(bufflen, GFP_KERNEL);
619 		if (!buff) {
620 			sdev_printk(KERN_WARNING, sdev,
621 				    "%s: kmalloc buffer failed\n",__func__);
622 			/* Temporary failure, bypass */
623 			pg->expiry = 0;
624 			return SCSI_DH_DEV_TEMP_BUSY;
625 		}
626 		goto retry;
627 	}
628 
629 	orig_transition_tmo = pg->transition_tmo;
630 	if ((buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR && buff[5] != 0)
631 		pg->transition_tmo = buff[5];
632 	else
633 		pg->transition_tmo = ALUA_FAILOVER_TIMEOUT;
634 
635 	if (orig_transition_tmo != pg->transition_tmo) {
636 		sdev_printk(KERN_INFO, sdev,
637 			    "%s: transition timeout set to %d seconds\n",
638 			    ALUA_DH_NAME, pg->transition_tmo);
639 		pg->expiry = jiffies + pg->transition_tmo * HZ;
640 	}
641 
642 	if ((buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR)
643 		tpg_desc_tbl_off = 8;
644 	else
645 		tpg_desc_tbl_off = 4;
646 
647 	for (k = tpg_desc_tbl_off, desc = buff + tpg_desc_tbl_off;
648 	     k < len;
649 	     k += off, desc += off) {
650 		u16 group_id = get_unaligned_be16(&desc[2]);
651 
652 		spin_lock_irqsave(&port_group_lock, flags);
653 		tmp_pg = alua_find_get_pg(pg->device_id_str, pg->device_id_len,
654 					  group_id);
655 		spin_unlock_irqrestore(&port_group_lock, flags);
656 		if (tmp_pg) {
657 			if (spin_trylock_irqsave(&tmp_pg->lock, flags)) {
658 				if ((tmp_pg == pg) ||
659 				    !(tmp_pg->flags & ALUA_PG_RUNNING)) {
660 					struct alua_dh_data *h;
661 
662 					tmp_pg->state = desc[0] & 0x0f;
663 					tmp_pg->pref = desc[0] >> 7;
664 					rcu_read_lock();
665 					list_for_each_entry_rcu(h,
666 						&tmp_pg->dh_list, node) {
667 						/* h->sdev should always be valid */
668 						BUG_ON(!h->sdev);
669 						h->sdev->access_state = desc[0];
670 					}
671 					rcu_read_unlock();
672 				}
673 				if (tmp_pg == pg)
674 					tmp_pg->valid_states = desc[1];
675 				spin_unlock_irqrestore(&tmp_pg->lock, flags);
676 			}
677 			kref_put(&tmp_pg->kref, release_port_group);
678 		}
679 		off = 8 + (desc[7] * 4);
680 	}
681 
682 	spin_lock_irqsave(&pg->lock, flags);
683 	sdev_printk(KERN_INFO, sdev,
684 		    "%s: port group %02x state %c %s supports %c%c%c%c%c%c%c\n",
685 		    ALUA_DH_NAME, pg->group_id, print_alua_state(pg->state),
686 		    pg->pref ? "preferred" : "non-preferred",
687 		    pg->valid_states&TPGS_SUPPORT_TRANSITION?'T':'t',
688 		    pg->valid_states&TPGS_SUPPORT_OFFLINE?'O':'o',
689 		    pg->valid_states&TPGS_SUPPORT_LBA_DEPENDENT?'L':'l',
690 		    pg->valid_states&TPGS_SUPPORT_UNAVAILABLE?'U':'u',
691 		    pg->valid_states&TPGS_SUPPORT_STANDBY?'S':'s',
692 		    pg->valid_states&TPGS_SUPPORT_NONOPTIMIZED?'N':'n',
693 		    pg->valid_states&TPGS_SUPPORT_OPTIMIZED?'A':'a');
694 
695 	switch (pg->state) {
696 	case SCSI_ACCESS_STATE_TRANSITIONING:
697 		if (time_before(jiffies, pg->expiry)) {
698 			/* State transition, retry */
699 			pg->interval = 2;
700 			err = SCSI_DH_RETRY;
701 		} else {
702 			struct alua_dh_data *h;
703 
704 			/* Transitioning time exceeded, set port to standby */
705 			err = SCSI_DH_IO;
706 			pg->state = SCSI_ACCESS_STATE_STANDBY;
707 			pg->expiry = 0;
708 			rcu_read_lock();
709 			list_for_each_entry_rcu(h, &pg->dh_list, node) {
710 				BUG_ON(!h->sdev);
711 				h->sdev->access_state =
712 					(pg->state & SCSI_ACCESS_STATE_MASK);
713 				if (pg->pref)
714 					h->sdev->access_state |=
715 						SCSI_ACCESS_STATE_PREFERRED;
716 			}
717 			rcu_read_unlock();
718 		}
719 		break;
720 	case SCSI_ACCESS_STATE_OFFLINE:
721 		/* Path unusable */
722 		err = SCSI_DH_DEV_OFFLINED;
723 		pg->expiry = 0;
724 		break;
725 	default:
726 		/* Useable path if active */
727 		err = SCSI_DH_OK;
728 		pg->expiry = 0;
729 		break;
730 	}
731 	spin_unlock_irqrestore(&pg->lock, flags);
732 	kfree(buff);
733 	return err;
734 }
735 
736 /*
737  * alua_stpg - Issue a SET TARGET PORT GROUP command
738  *
739  * Issue a SET TARGET PORT GROUP command and evaluate the
740  * response. Returns SCSI_DH_RETRY per default to trigger
741  * a re-evaluation of the target group state or SCSI_DH_OK
742  * if no further action needs to be taken.
743  */
744 static unsigned alua_stpg(struct scsi_device *sdev, struct alua_port_group *pg)
745 {
746 	int retval;
747 	struct scsi_sense_hdr sense_hdr;
748 
749 	if (!(pg->tpgs & TPGS_MODE_EXPLICIT)) {
750 		/* Only implicit ALUA supported, retry */
751 		return SCSI_DH_RETRY;
752 	}
753 	switch (pg->state) {
754 	case SCSI_ACCESS_STATE_OPTIMAL:
755 		return SCSI_DH_OK;
756 	case SCSI_ACCESS_STATE_ACTIVE:
757 		if ((pg->flags & ALUA_OPTIMIZE_STPG) &&
758 		    !pg->pref &&
759 		    (pg->tpgs & TPGS_MODE_IMPLICIT))
760 			return SCSI_DH_OK;
761 		break;
762 	case SCSI_ACCESS_STATE_STANDBY:
763 	case SCSI_ACCESS_STATE_UNAVAILABLE:
764 		break;
765 	case SCSI_ACCESS_STATE_OFFLINE:
766 		return SCSI_DH_IO;
767 	case SCSI_ACCESS_STATE_TRANSITIONING:
768 		break;
769 	default:
770 		sdev_printk(KERN_INFO, sdev,
771 			    "%s: stpg failed, unhandled TPGS state %d",
772 			    ALUA_DH_NAME, pg->state);
773 		return SCSI_DH_NOSYS;
774 	}
775 	retval = submit_stpg(sdev, pg->group_id, &sense_hdr);
776 
777 	if (retval) {
778 		if (!scsi_sense_valid(&sense_hdr)) {
779 			sdev_printk(KERN_INFO, sdev,
780 				    "%s: stpg failed, result %d",
781 				    ALUA_DH_NAME, retval);
782 			if (driver_byte(retval) == DRIVER_ERROR)
783 				return SCSI_DH_DEV_TEMP_BUSY;
784 		} else {
785 			sdev_printk(KERN_INFO, sdev, "%s: stpg failed\n",
786 				    ALUA_DH_NAME);
787 			scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
788 		}
789 	}
790 	/* Retry RTPG */
791 	return SCSI_DH_RETRY;
792 }
793 
794 static void alua_rtpg_work(struct work_struct *work)
795 {
796 	struct alua_port_group *pg =
797 		container_of(work, struct alua_port_group, rtpg_work.work);
798 	struct scsi_device *sdev;
799 	LIST_HEAD(qdata_list);
800 	int err = SCSI_DH_OK;
801 	struct alua_queue_data *qdata, *tmp;
802 	unsigned long flags;
803 
804 	spin_lock_irqsave(&pg->lock, flags);
805 	sdev = pg->rtpg_sdev;
806 	if (!sdev) {
807 		WARN_ON(pg->flags & ALUA_PG_RUN_RTPG);
808 		WARN_ON(pg->flags & ALUA_PG_RUN_STPG);
809 		spin_unlock_irqrestore(&pg->lock, flags);
810 		kref_put(&pg->kref, release_port_group);
811 		return;
812 	}
813 	pg->flags |= ALUA_PG_RUNNING;
814 	if (pg->flags & ALUA_PG_RUN_RTPG) {
815 		int state = pg->state;
816 
817 		pg->flags &= ~ALUA_PG_RUN_RTPG;
818 		spin_unlock_irqrestore(&pg->lock, flags);
819 		if (state == SCSI_ACCESS_STATE_TRANSITIONING) {
820 			if (alua_tur(sdev) == SCSI_DH_RETRY) {
821 				spin_lock_irqsave(&pg->lock, flags);
822 				pg->flags &= ~ALUA_PG_RUNNING;
823 				pg->flags |= ALUA_PG_RUN_RTPG;
824 				spin_unlock_irqrestore(&pg->lock, flags);
825 				queue_delayed_work(kaluad_wq, &pg->rtpg_work,
826 						   pg->interval * HZ);
827 				return;
828 			}
829 			/* Send RTPG on failure or if TUR indicates SUCCESS */
830 		}
831 		err = alua_rtpg(sdev, pg);
832 		spin_lock_irqsave(&pg->lock, flags);
833 		if (err == SCSI_DH_RETRY || pg->flags & ALUA_PG_RUN_RTPG) {
834 			pg->flags &= ~ALUA_PG_RUNNING;
835 			pg->flags |= ALUA_PG_RUN_RTPG;
836 			spin_unlock_irqrestore(&pg->lock, flags);
837 			queue_delayed_work(kaluad_wq, &pg->rtpg_work,
838 					   pg->interval * HZ);
839 			return;
840 		}
841 		if (err != SCSI_DH_OK)
842 			pg->flags &= ~ALUA_PG_RUN_STPG;
843 	}
844 	if (pg->flags & ALUA_PG_RUN_STPG) {
845 		pg->flags &= ~ALUA_PG_RUN_STPG;
846 		spin_unlock_irqrestore(&pg->lock, flags);
847 		err = alua_stpg(sdev, pg);
848 		spin_lock_irqsave(&pg->lock, flags);
849 		if (err == SCSI_DH_RETRY || pg->flags & ALUA_PG_RUN_RTPG) {
850 			pg->flags |= ALUA_PG_RUN_RTPG;
851 			pg->interval = 0;
852 			pg->flags &= ~ALUA_PG_RUNNING;
853 			spin_unlock_irqrestore(&pg->lock, flags);
854 			queue_delayed_work(kaluad_wq, &pg->rtpg_work,
855 					   pg->interval * HZ);
856 			return;
857 		}
858 	}
859 
860 	list_splice_init(&pg->rtpg_list, &qdata_list);
861 	pg->rtpg_sdev = NULL;
862 	spin_unlock_irqrestore(&pg->lock, flags);
863 
864 	list_for_each_entry_safe(qdata, tmp, &qdata_list, entry) {
865 		list_del(&qdata->entry);
866 		if (qdata->callback_fn)
867 			qdata->callback_fn(qdata->callback_data, err);
868 		kfree(qdata);
869 	}
870 	spin_lock_irqsave(&pg->lock, flags);
871 	pg->flags &= ~ALUA_PG_RUNNING;
872 	spin_unlock_irqrestore(&pg->lock, flags);
873 	scsi_device_put(sdev);
874 	kref_put(&pg->kref, release_port_group);
875 }
876 
877 /**
878  * alua_rtpg_queue() - cause RTPG to be submitted asynchronously
879  * @pg: ALUA port group associated with @sdev.
880  * @sdev: SCSI device for which to submit an RTPG.
881  * @qdata: Information about the callback to invoke after the RTPG.
882  * @force: Whether or not to submit an RTPG if a work item that will submit an
883  *         RTPG already has been scheduled.
884  *
885  * Returns true if and only if alua_rtpg_work() will be called asynchronously.
886  * That function is responsible for calling @qdata->fn().
887  */
888 static bool alua_rtpg_queue(struct alua_port_group *pg,
889 			    struct scsi_device *sdev,
890 			    struct alua_queue_data *qdata, bool force)
891 {
892 	int start_queue = 0;
893 	unsigned long flags;
894 	if (WARN_ON_ONCE(!pg) || scsi_device_get(sdev))
895 		return false;
896 
897 	spin_lock_irqsave(&pg->lock, flags);
898 	if (qdata) {
899 		list_add_tail(&qdata->entry, &pg->rtpg_list);
900 		pg->flags |= ALUA_PG_RUN_STPG;
901 		force = true;
902 	}
903 	if (pg->rtpg_sdev == NULL) {
904 		pg->interval = 0;
905 		pg->flags |= ALUA_PG_RUN_RTPG;
906 		kref_get(&pg->kref);
907 		pg->rtpg_sdev = sdev;
908 		start_queue = 1;
909 	} else if (!(pg->flags & ALUA_PG_RUN_RTPG) && force) {
910 		pg->flags |= ALUA_PG_RUN_RTPG;
911 		/* Do not queue if the worker is already running */
912 		if (!(pg->flags & ALUA_PG_RUNNING)) {
913 			kref_get(&pg->kref);
914 			start_queue = 1;
915 		}
916 	}
917 
918 	spin_unlock_irqrestore(&pg->lock, flags);
919 
920 	if (start_queue) {
921 		if (queue_delayed_work(kaluad_wq, &pg->rtpg_work,
922 				msecs_to_jiffies(ALUA_RTPG_DELAY_MSECS)))
923 			sdev = NULL;
924 		else
925 			kref_put(&pg->kref, release_port_group);
926 	}
927 	if (sdev)
928 		scsi_device_put(sdev);
929 
930 	return true;
931 }
932 
933 /*
934  * alua_initialize - Initialize ALUA state
935  * @sdev: the device to be initialized
936  *
937  * For the prep_fn to work correctly we have
938  * to initialize the ALUA state for the device.
939  */
940 static int alua_initialize(struct scsi_device *sdev, struct alua_dh_data *h)
941 {
942 	int err = SCSI_DH_DEV_UNSUPP, tpgs;
943 
944 	mutex_lock(&h->init_mutex);
945 	tpgs = alua_check_tpgs(sdev);
946 	if (tpgs != TPGS_MODE_NONE)
947 		err = alua_check_vpd(sdev, h, tpgs);
948 	h->init_error = err;
949 	mutex_unlock(&h->init_mutex);
950 	return err;
951 }
952 /*
953  * alua_set_params - set/unset the optimize flag
954  * @sdev: device on the path to be activated
955  * params - parameters in the following format
956  *      "no_of_params\0param1\0param2\0param3\0...\0"
957  * For example, to set the flag pass the following parameters
958  * from multipath.conf
959  *     hardware_handler        "2 alua 1"
960  */
961 static int alua_set_params(struct scsi_device *sdev, const char *params)
962 {
963 	struct alua_dh_data *h = sdev->handler_data;
964 	struct alua_port_group *pg = NULL;
965 	unsigned int optimize = 0, argc;
966 	const char *p = params;
967 	int result = SCSI_DH_OK;
968 	unsigned long flags;
969 
970 	if ((sscanf(params, "%u", &argc) != 1) || (argc != 1))
971 		return -EINVAL;
972 
973 	while (*p++)
974 		;
975 	if ((sscanf(p, "%u", &optimize) != 1) || (optimize > 1))
976 		return -EINVAL;
977 
978 	rcu_read_lock();
979 	pg = rcu_dereference(h->pg);
980 	if (!pg) {
981 		rcu_read_unlock();
982 		return -ENXIO;
983 	}
984 	spin_lock_irqsave(&pg->lock, flags);
985 	if (optimize)
986 		pg->flags |= ALUA_OPTIMIZE_STPG;
987 	else
988 		pg->flags &= ~ALUA_OPTIMIZE_STPG;
989 	spin_unlock_irqrestore(&pg->lock, flags);
990 	rcu_read_unlock();
991 
992 	return result;
993 }
994 
995 /*
996  * alua_activate - activate a path
997  * @sdev: device on the path to be activated
998  *
999  * We're currently switching the port group to be activated only and
1000  * let the array figure out the rest.
1001  * There may be other arrays which require us to switch all port groups
1002  * based on a certain policy. But until we actually encounter them it
1003  * should be okay.
1004  */
1005 static int alua_activate(struct scsi_device *sdev,
1006 			activate_complete fn, void *data)
1007 {
1008 	struct alua_dh_data *h = sdev->handler_data;
1009 	int err = SCSI_DH_OK;
1010 	struct alua_queue_data *qdata;
1011 	struct alua_port_group *pg;
1012 
1013 	qdata = kzalloc(sizeof(*qdata), GFP_KERNEL);
1014 	if (!qdata) {
1015 		err = SCSI_DH_RES_TEMP_UNAVAIL;
1016 		goto out;
1017 	}
1018 	qdata->callback_fn = fn;
1019 	qdata->callback_data = data;
1020 
1021 	mutex_lock(&h->init_mutex);
1022 	rcu_read_lock();
1023 	pg = rcu_dereference(h->pg);
1024 	if (!pg || !kref_get_unless_zero(&pg->kref)) {
1025 		rcu_read_unlock();
1026 		kfree(qdata);
1027 		err = h->init_error;
1028 		mutex_unlock(&h->init_mutex);
1029 		goto out;
1030 	}
1031 	rcu_read_unlock();
1032 	mutex_unlock(&h->init_mutex);
1033 
1034 	if (alua_rtpg_queue(pg, sdev, qdata, true))
1035 		fn = NULL;
1036 	else
1037 		err = SCSI_DH_DEV_OFFLINED;
1038 	kref_put(&pg->kref, release_port_group);
1039 out:
1040 	if (fn)
1041 		fn(data, err);
1042 	return 0;
1043 }
1044 
1045 /*
1046  * alua_check - check path status
1047  * @sdev: device on the path to be checked
1048  *
1049  * Check the device status
1050  */
1051 static void alua_check(struct scsi_device *sdev, bool force)
1052 {
1053 	struct alua_dh_data *h = sdev->handler_data;
1054 	struct alua_port_group *pg;
1055 
1056 	rcu_read_lock();
1057 	pg = rcu_dereference(h->pg);
1058 	if (!pg || !kref_get_unless_zero(&pg->kref)) {
1059 		rcu_read_unlock();
1060 		return;
1061 	}
1062 	rcu_read_unlock();
1063 
1064 	alua_rtpg_queue(pg, sdev, NULL, force);
1065 	kref_put(&pg->kref, release_port_group);
1066 }
1067 
1068 /*
1069  * alua_prep_fn - request callback
1070  *
1071  * Fail I/O to all paths not in state
1072  * active/optimized or active/non-optimized.
1073  */
1074 static int alua_prep_fn(struct scsi_device *sdev, struct request *req)
1075 {
1076 	struct alua_dh_data *h = sdev->handler_data;
1077 	struct alua_port_group *pg;
1078 	unsigned char state = SCSI_ACCESS_STATE_OPTIMAL;
1079 	int ret = BLKPREP_OK;
1080 
1081 	rcu_read_lock();
1082 	pg = rcu_dereference(h->pg);
1083 	if (pg)
1084 		state = pg->state;
1085 	rcu_read_unlock();
1086 	if (state == SCSI_ACCESS_STATE_TRANSITIONING)
1087 		ret = BLKPREP_DEFER;
1088 	else if (state != SCSI_ACCESS_STATE_OPTIMAL &&
1089 		 state != SCSI_ACCESS_STATE_ACTIVE &&
1090 		 state != SCSI_ACCESS_STATE_LBA) {
1091 		ret = BLKPREP_KILL;
1092 		req->rq_flags |= RQF_QUIET;
1093 	}
1094 	return ret;
1095 
1096 }
1097 
1098 static void alua_rescan(struct scsi_device *sdev)
1099 {
1100 	struct alua_dh_data *h = sdev->handler_data;
1101 
1102 	alua_initialize(sdev, h);
1103 }
1104 
1105 /*
1106  * alua_bus_attach - Attach device handler
1107  * @sdev: device to be attached to
1108  */
1109 static int alua_bus_attach(struct scsi_device *sdev)
1110 {
1111 	struct alua_dh_data *h;
1112 	int err;
1113 
1114 	h = kzalloc(sizeof(*h) , GFP_KERNEL);
1115 	if (!h)
1116 		return SCSI_DH_NOMEM;
1117 	spin_lock_init(&h->pg_lock);
1118 	rcu_assign_pointer(h->pg, NULL);
1119 	h->init_error = SCSI_DH_OK;
1120 	h->sdev = sdev;
1121 	INIT_LIST_HEAD(&h->node);
1122 
1123 	mutex_init(&h->init_mutex);
1124 	err = alua_initialize(sdev, h);
1125 	if (err != SCSI_DH_OK && err != SCSI_DH_DEV_OFFLINED)
1126 		goto failed;
1127 
1128 	sdev->handler_data = h;
1129 	return SCSI_DH_OK;
1130 failed:
1131 	kfree(h);
1132 	return err;
1133 }
1134 
1135 /*
1136  * alua_bus_detach - Detach device handler
1137  * @sdev: device to be detached from
1138  */
1139 static void alua_bus_detach(struct scsi_device *sdev)
1140 {
1141 	struct alua_dh_data *h = sdev->handler_data;
1142 	struct alua_port_group *pg;
1143 
1144 	spin_lock(&h->pg_lock);
1145 	pg = rcu_dereference_protected(h->pg, lockdep_is_held(&h->pg_lock));
1146 	rcu_assign_pointer(h->pg, NULL);
1147 	h->sdev = NULL;
1148 	spin_unlock(&h->pg_lock);
1149 	if (pg) {
1150 		spin_lock_irq(&pg->lock);
1151 		list_del_rcu(&h->node);
1152 		spin_unlock_irq(&pg->lock);
1153 		kref_put(&pg->kref, release_port_group);
1154 	}
1155 	sdev->handler_data = NULL;
1156 	kfree(h);
1157 }
1158 
1159 static struct scsi_device_handler alua_dh = {
1160 	.name = ALUA_DH_NAME,
1161 	.module = THIS_MODULE,
1162 	.attach = alua_bus_attach,
1163 	.detach = alua_bus_detach,
1164 	.prep_fn = alua_prep_fn,
1165 	.check_sense = alua_check_sense,
1166 	.activate = alua_activate,
1167 	.rescan = alua_rescan,
1168 	.set_params = alua_set_params,
1169 };
1170 
1171 static int __init alua_init(void)
1172 {
1173 	int r;
1174 
1175 	kaluad_wq = alloc_workqueue("kaluad", WQ_MEM_RECLAIM, 0);
1176 	if (!kaluad_wq) {
1177 		/* Temporary failure, bypass */
1178 		return SCSI_DH_DEV_TEMP_BUSY;
1179 	}
1180 
1181 	r = scsi_register_device_handler(&alua_dh);
1182 	if (r != 0) {
1183 		printk(KERN_ERR "%s: Failed to register scsi device handler",
1184 			ALUA_DH_NAME);
1185 		destroy_workqueue(kaluad_wq);
1186 	}
1187 	return r;
1188 }
1189 
1190 static void __exit alua_exit(void)
1191 {
1192 	scsi_unregister_device_handler(&alua_dh);
1193 	destroy_workqueue(kaluad_wq);
1194 }
1195 
1196 module_init(alua_init);
1197 module_exit(alua_exit);
1198 
1199 MODULE_DESCRIPTION("DM Multipath ALUA support");
1200 MODULE_AUTHOR("Hannes Reinecke <hare@suse.de>");
1201 MODULE_LICENSE("GPL");
1202 MODULE_VERSION(ALUA_DH_VER);
1203