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