GNU Linux-libre 4.14.303-gnu1
[releases.git] / drivers / scsi / device_handler / scsi_dh_alua.c
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 #define ALUA_RTPG_RETRY_DELAY           2
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         unsigned                flags; /* used for optimizing STPG */
86         unsigned char           transition_tmo;
87         unsigned long           expiry;
88         unsigned long           interval;
89         struct delayed_work     rtpg_work;
90         spinlock_t              lock;
91         struct list_head        rtpg_list;
92         struct scsi_device      *rtpg_sdev;
93 };
94
95 struct alua_dh_data {
96         struct list_head        node;
97         struct alua_port_group __rcu *pg;
98         int                     group_id;
99         spinlock_t              pg_lock;
100         struct scsi_device      *sdev;
101         int                     init_error;
102         struct mutex            init_mutex;
103 };
104
105 struct alua_queue_data {
106         struct list_head        entry;
107         activate_complete       callback_fn;
108         void                    *callback_data;
109 };
110
111 #define ALUA_POLICY_SWITCH_CURRENT      0
112 #define ALUA_POLICY_SWITCH_ALL          1
113
114 static void alua_rtpg_work(struct work_struct *work);
115 static bool alua_rtpg_queue(struct alua_port_group *pg,
116                             struct scsi_device *sdev,
117                             struct alua_queue_data *qdata, bool force);
118 static void alua_check(struct scsi_device *sdev, bool force);
119
120 static void release_port_group(struct kref *kref)
121 {
122         struct alua_port_group *pg;
123
124         pg = container_of(kref, struct alua_port_group, kref);
125         if (pg->rtpg_sdev)
126                 flush_delayed_work(&pg->rtpg_work);
127         spin_lock(&port_group_lock);
128         list_del(&pg->node);
129         spin_unlock(&port_group_lock);
130         kfree_rcu(pg, rcu);
131 }
132
133 /*
134  * submit_rtpg - Issue a REPORT TARGET GROUP STATES command
135  * @sdev: sdev the command should be sent to
136  */
137 static int submit_rtpg(struct scsi_device *sdev, unsigned char *buff,
138                        int bufflen, struct scsi_sense_hdr *sshdr, int flags)
139 {
140         u8 cdb[COMMAND_SIZE(MAINTENANCE_IN)];
141         int req_flags = REQ_FAILFAST_DEV | REQ_FAILFAST_TRANSPORT |
142                 REQ_FAILFAST_DRIVER;
143
144         /* Prepare the command. */
145         memset(cdb, 0x0, COMMAND_SIZE(MAINTENANCE_IN));
146         cdb[0] = MAINTENANCE_IN;
147         if (!(flags & ALUA_RTPG_EXT_HDR_UNSUPP))
148                 cdb[1] = MI_REPORT_TARGET_PGS | MI_EXT_HDR_PARAM_FMT;
149         else
150                 cdb[1] = MI_REPORT_TARGET_PGS;
151         put_unaligned_be32(bufflen, &cdb[6]);
152
153         return scsi_execute(sdev, cdb, DMA_FROM_DEVICE, buff, bufflen, NULL,
154                         sshdr, ALUA_FAILOVER_TIMEOUT * HZ,
155                         ALUA_FAILOVER_RETRIES, req_flags, 0, NULL);
156 }
157
158 /*
159  * submit_stpg - Issue a SET TARGET PORT GROUP command
160  *
161  * Currently we're only setting the current target port group state
162  * to 'active/optimized' and let the array firmware figure out
163  * the states of the remaining groups.
164  */
165 static int submit_stpg(struct scsi_device *sdev, int group_id,
166                        struct scsi_sense_hdr *sshdr)
167 {
168         u8 cdb[COMMAND_SIZE(MAINTENANCE_OUT)];
169         unsigned char stpg_data[8];
170         int stpg_len = 8;
171         int req_flags = REQ_FAILFAST_DEV | REQ_FAILFAST_TRANSPORT |
172                 REQ_FAILFAST_DRIVER;
173
174         /* Prepare the data buffer */
175         memset(stpg_data, 0, stpg_len);
176         stpg_data[4] = SCSI_ACCESS_STATE_OPTIMAL;
177         put_unaligned_be16(group_id, &stpg_data[6]);
178
179         /* Prepare the command. */
180         memset(cdb, 0x0, COMMAND_SIZE(MAINTENANCE_OUT));
181         cdb[0] = MAINTENANCE_OUT;
182         cdb[1] = MO_SET_TARGET_PGS;
183         put_unaligned_be32(stpg_len, &cdb[6]);
184
185         return scsi_execute(sdev, cdb, DMA_TO_DEVICE, stpg_data, stpg_len, NULL,
186                         sshdr, ALUA_FAILOVER_TIMEOUT * HZ,
187                         ALUA_FAILOVER_RETRIES, req_flags, 0, NULL);
188 }
189
190 static struct alua_port_group *alua_find_get_pg(char *id_str, size_t id_size,
191                                                 int group_id)
192 {
193         struct alua_port_group *pg;
194
195         if (!id_str || !id_size || !strlen(id_str))
196                 return NULL;
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 || 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 static 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                  * TPGS supported but no device identification found.
236                  * Generate private device identification.
237                  */
238                 sdev_printk(KERN_INFO, sdev,
239                             "%s: No device descriptors found\n",
240                             ALUA_DH_NAME);
241                 pg->device_id_str[0] = '\0';
242                 pg->device_id_len = 0;
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         if (pg->device_id_len)
358                 sdev_printk(KERN_INFO, sdev,
359                             "%s: device %s port group %x rel port %x\n",
360                             ALUA_DH_NAME, pg->device_id_str,
361                             group_id, rel_port);
362         else
363                 sdev_printk(KERN_INFO, sdev,
364                             "%s: port group %x rel port %x\n",
365                             ALUA_DH_NAME, group_id, rel_port);
366
367         /* Check for existing port group references */
368         spin_lock(&h->pg_lock);
369         old_pg = rcu_dereference_protected(h->pg, lockdep_is_held(&h->pg_lock));
370         if (old_pg != pg) {
371                 /* port group has changed. Update to new port group */
372                 if (h->pg) {
373                         spin_lock_irqsave(&old_pg->lock, flags);
374                         list_del_rcu(&h->node);
375                         spin_unlock_irqrestore(&old_pg->lock, flags);
376                 }
377                 rcu_assign_pointer(h->pg, pg);
378                 pg_updated = true;
379         }
380
381         spin_lock_irqsave(&pg->lock, flags);
382         if (pg_updated)
383                 list_add_rcu(&h->node, &pg->dh_list);
384         spin_unlock_irqrestore(&pg->lock, flags);
385
386         alua_rtpg_queue(rcu_dereference_protected(h->pg,
387                                                   lockdep_is_held(&h->pg_lock)),
388                         sdev, NULL, true);
389         spin_unlock(&h->pg_lock);
390
391         if (old_pg)
392                 kref_put(&old_pg->kref, release_port_group);
393
394         return SCSI_DH_OK;
395 }
396
397 static char print_alua_state(unsigned char state)
398 {
399         switch (state) {
400         case SCSI_ACCESS_STATE_OPTIMAL:
401                 return 'A';
402         case SCSI_ACCESS_STATE_ACTIVE:
403                 return 'N';
404         case SCSI_ACCESS_STATE_STANDBY:
405                 return 'S';
406         case SCSI_ACCESS_STATE_UNAVAILABLE:
407                 return 'U';
408         case SCSI_ACCESS_STATE_LBA:
409                 return 'L';
410         case SCSI_ACCESS_STATE_OFFLINE:
411                 return 'O';
412         case SCSI_ACCESS_STATE_TRANSITIONING:
413                 return 'T';
414         default:
415                 return 'X';
416         }
417 }
418
419 static int alua_check_sense(struct scsi_device *sdev,
420                             struct scsi_sense_hdr *sense_hdr)
421 {
422         switch (sense_hdr->sense_key) {
423         case NOT_READY:
424                 if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0a) {
425                         /*
426                          * LUN Not Accessible - ALUA state transition
427                          */
428                         alua_check(sdev, false);
429                         return NEEDS_RETRY;
430                 }
431                 break;
432         case UNIT_ATTENTION:
433                 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x00) {
434                         /*
435                          * Power On, Reset, or Bus Device Reset.
436                          * Might have obscured a state transition,
437                          * so schedule a recheck.
438                          */
439                         alua_check(sdev, true);
440                         return ADD_TO_MLQUEUE;
441                 }
442                 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x04)
443                         /*
444                          * Device internal reset
445                          */
446                         return ADD_TO_MLQUEUE;
447                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x01)
448                         /*
449                          * Mode Parameters Changed
450                          */
451                         return ADD_TO_MLQUEUE;
452                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x06) {
453                         /*
454                          * ALUA state changed
455                          */
456                         alua_check(sdev, true);
457                         return ADD_TO_MLQUEUE;
458                 }
459                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x07) {
460                         /*
461                          * Implicit ALUA state transition failed
462                          */
463                         alua_check(sdev, true);
464                         return ADD_TO_MLQUEUE;
465                 }
466                 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x03)
467                         /*
468                          * Inquiry data has changed
469                          */
470                         return ADD_TO_MLQUEUE;
471                 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x0e)
472                         /*
473                          * REPORTED_LUNS_DATA_HAS_CHANGED is reported
474                          * when switching controllers on targets like
475                          * Intel Multi-Flex. We can just retry.
476                          */
477                         return ADD_TO_MLQUEUE;
478                 break;
479         }
480
481         return SCSI_RETURN_NOT_HANDLED;
482 }
483
484 /*
485  * alua_tur - Send a TEST UNIT READY
486  * @sdev: device to which the TEST UNIT READY command should be send
487  *
488  * Send a TEST UNIT READY to @sdev to figure out the device state
489  * Returns SCSI_DH_RETRY if the sense code is NOT READY/ALUA TRANSITIONING,
490  * SCSI_DH_OK if no error occurred, and SCSI_DH_IO otherwise.
491  */
492 static int alua_tur(struct scsi_device *sdev)
493 {
494         struct scsi_sense_hdr sense_hdr;
495         int retval;
496
497         retval = scsi_test_unit_ready(sdev, ALUA_FAILOVER_TIMEOUT * HZ,
498                                       ALUA_FAILOVER_RETRIES, &sense_hdr);
499         if (sense_hdr.sense_key == NOT_READY &&
500             sense_hdr.asc == 0x04 && sense_hdr.ascq == 0x0a)
501                 return SCSI_DH_RETRY;
502         else if (retval)
503                 return SCSI_DH_IO;
504         else
505                 return SCSI_DH_OK;
506 }
507
508 /*
509  * alua_rtpg - Evaluate REPORT TARGET GROUP STATES
510  * @sdev: the device to be evaluated.
511  *
512  * Evaluate the Target Port Group State.
513  * Returns SCSI_DH_DEV_OFFLINED if the path is
514  * found to be unusable.
515  */
516 static int alua_rtpg(struct scsi_device *sdev, struct alua_port_group *pg)
517 {
518         struct scsi_sense_hdr sense_hdr;
519         struct alua_port_group *tmp_pg;
520         int len, k, off, valid_states = 0, bufflen = ALUA_RTPG_SIZE;
521         unsigned char *desc, *buff;
522         unsigned err, retval;
523         unsigned int tpg_desc_tbl_off;
524         unsigned char orig_transition_tmo;
525         unsigned long flags;
526         bool transitioning_sense = false;
527
528         if (!pg->expiry) {
529                 unsigned long transition_tmo = ALUA_FAILOVER_TIMEOUT * HZ;
530
531                 if (pg->transition_tmo)
532                         transition_tmo = pg->transition_tmo * HZ;
533
534                 pg->expiry = round_jiffies_up(jiffies + transition_tmo);
535         }
536
537         buff = kzalloc(bufflen, GFP_KERNEL);
538         if (!buff)
539                 return SCSI_DH_DEV_TEMP_BUSY;
540
541  retry:
542         err = 0;
543         retval = submit_rtpg(sdev, buff, bufflen, &sense_hdr, pg->flags);
544
545         if (retval) {
546                 if (!scsi_sense_valid(&sense_hdr)) {
547                         sdev_printk(KERN_INFO, sdev,
548                                     "%s: rtpg failed, result %d\n",
549                                     ALUA_DH_NAME, retval);
550                         kfree(buff);
551                         if (driver_byte(retval) == DRIVER_ERROR)
552                                 return SCSI_DH_DEV_TEMP_BUSY;
553                         return SCSI_DH_IO;
554                 }
555
556                 /*
557                  * submit_rtpg() has failed on existing arrays
558                  * when requesting extended header info, and
559                  * the array doesn't support extended headers,
560                  * even though it shouldn't according to T10.
561                  * The retry without rtpg_ext_hdr_req set
562                  * handles this.
563                  * Note:  some arrays return a sense key of ILLEGAL_REQUEST
564                  * with ASC 00h if they don't support the extended header.
565                  */
566                 if (!(pg->flags & ALUA_RTPG_EXT_HDR_UNSUPP) &&
567                     sense_hdr.sense_key == ILLEGAL_REQUEST) {
568                         pg->flags |= ALUA_RTPG_EXT_HDR_UNSUPP;
569                         goto retry;
570                 }
571                 /*
572                  * If the array returns with 'ALUA state transition'
573                  * sense code here it cannot return RTPG data during
574                  * transition. So set the state to 'transitioning' directly.
575                  */
576                 if (sense_hdr.sense_key == NOT_READY &&
577                     sense_hdr.asc == 0x04 && sense_hdr.ascq == 0x0a) {
578                         transitioning_sense = true;
579                         goto skip_rtpg;
580                 }
581                 /*
582                  * Retry on any other UNIT ATTENTION occurred.
583                  */
584                 if (sense_hdr.sense_key == UNIT_ATTENTION)
585                         err = SCSI_DH_RETRY;
586                 if (err == SCSI_DH_RETRY &&
587                     pg->expiry != 0 && time_before(jiffies, pg->expiry)) {
588                         sdev_printk(KERN_ERR, sdev, "%s: rtpg retry\n",
589                                     ALUA_DH_NAME);
590                         scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
591                         kfree(buff);
592                         return err;
593                 }
594                 sdev_printk(KERN_ERR, sdev, "%s: rtpg failed\n",
595                             ALUA_DH_NAME);
596                 scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
597                 kfree(buff);
598                 pg->expiry = 0;
599                 return SCSI_DH_IO;
600         }
601
602         len = get_unaligned_be32(&buff[0]) + 4;
603
604         if (len > bufflen) {
605                 /* Resubmit with the correct length */
606                 kfree(buff);
607                 bufflen = len;
608                 buff = kmalloc(bufflen, GFP_KERNEL);
609                 if (!buff) {
610                         sdev_printk(KERN_WARNING, sdev,
611                                     "%s: kmalloc buffer failed\n",__func__);
612                         /* Temporary failure, bypass */
613                         pg->expiry = 0;
614                         return SCSI_DH_DEV_TEMP_BUSY;
615                 }
616                 goto retry;
617         }
618
619         orig_transition_tmo = pg->transition_tmo;
620         if ((buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR && buff[5] != 0)
621                 pg->transition_tmo = buff[5];
622         else
623                 pg->transition_tmo = ALUA_FAILOVER_TIMEOUT;
624
625         if (orig_transition_tmo != pg->transition_tmo) {
626                 sdev_printk(KERN_INFO, sdev,
627                             "%s: transition timeout set to %d seconds\n",
628                             ALUA_DH_NAME, pg->transition_tmo);
629                 pg->expiry = jiffies + pg->transition_tmo * HZ;
630         }
631
632         if ((buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR)
633                 tpg_desc_tbl_off = 8;
634         else
635                 tpg_desc_tbl_off = 4;
636
637         for (k = tpg_desc_tbl_off, desc = buff + tpg_desc_tbl_off;
638              k < len;
639              k += off, desc += off) {
640                 u16 group_id = get_unaligned_be16(&desc[2]);
641
642                 spin_lock_irqsave(&port_group_lock, flags);
643                 tmp_pg = alua_find_get_pg(pg->device_id_str, pg->device_id_len,
644                                           group_id);
645                 spin_unlock_irqrestore(&port_group_lock, flags);
646                 if (tmp_pg) {
647                         if (spin_trylock_irqsave(&tmp_pg->lock, flags)) {
648                                 if ((tmp_pg == pg) ||
649                                     !(tmp_pg->flags & ALUA_PG_RUNNING)) {
650                                         struct alua_dh_data *h;
651
652                                         tmp_pg->state = desc[0] & 0x0f;
653                                         tmp_pg->pref = desc[0] >> 7;
654                                         rcu_read_lock();
655                                         list_for_each_entry_rcu(h,
656                                                 &tmp_pg->dh_list, node) {
657                                                 if (!h->sdev)
658                                                         continue;
659                                                 h->sdev->access_state = desc[0];
660                                         }
661                                         rcu_read_unlock();
662                                 }
663                                 if (tmp_pg == pg)
664                                         valid_states = desc[1];
665                                 spin_unlock_irqrestore(&tmp_pg->lock, flags);
666                         }
667                         kref_put(&tmp_pg->kref, release_port_group);
668                 }
669                 off = 8 + (desc[7] * 4);
670         }
671
672  skip_rtpg:
673         spin_lock_irqsave(&pg->lock, flags);
674         if (transitioning_sense)
675                 pg->state = SCSI_ACCESS_STATE_TRANSITIONING;
676
677         sdev_printk(KERN_INFO, sdev,
678                     "%s: port group %02x state %c %s supports %c%c%c%c%c%c%c\n",
679                     ALUA_DH_NAME, pg->group_id, print_alua_state(pg->state),
680                     pg->pref ? "preferred" : "non-preferred",
681                     valid_states&TPGS_SUPPORT_TRANSITION?'T':'t',
682                     valid_states&TPGS_SUPPORT_OFFLINE?'O':'o',
683                     valid_states&TPGS_SUPPORT_LBA_DEPENDENT?'L':'l',
684                     valid_states&TPGS_SUPPORT_UNAVAILABLE?'U':'u',
685                     valid_states&TPGS_SUPPORT_STANDBY?'S':'s',
686                     valid_states&TPGS_SUPPORT_NONOPTIMIZED?'N':'n',
687                     valid_states&TPGS_SUPPORT_OPTIMIZED?'A':'a');
688
689         switch (pg->state) {
690         case SCSI_ACCESS_STATE_TRANSITIONING:
691                 if (time_before(jiffies, pg->expiry)) {
692                         /* State transition, retry */
693                         pg->interval = ALUA_RTPG_RETRY_DELAY;
694                         err = SCSI_DH_RETRY;
695                 } else {
696                         struct alua_dh_data *h;
697
698                         /* Transitioning time exceeded, set port to standby */
699                         err = SCSI_DH_IO;
700                         pg->state = SCSI_ACCESS_STATE_STANDBY;
701                         pg->expiry = 0;
702                         rcu_read_lock();
703                         list_for_each_entry_rcu(h, &pg->dh_list, node) {
704                                 if (!h->sdev)
705                                         continue;
706                                 h->sdev->access_state =
707                                         (pg->state & SCSI_ACCESS_STATE_MASK);
708                                 if (pg->pref)
709                                         h->sdev->access_state |=
710                                                 SCSI_ACCESS_STATE_PREFERRED;
711                         }
712                         rcu_read_unlock();
713                 }
714                 break;
715         case SCSI_ACCESS_STATE_OFFLINE:
716                 /* Path unusable */
717                 err = SCSI_DH_DEV_OFFLINED;
718                 pg->expiry = 0;
719                 break;
720         default:
721                 /* Useable path if active */
722                 err = SCSI_DH_OK;
723                 pg->expiry = 0;
724                 break;
725         }
726         spin_unlock_irqrestore(&pg->lock, flags);
727         kfree(buff);
728         return err;
729 }
730
731 /*
732  * alua_stpg - Issue a SET TARGET PORT GROUP command
733  *
734  * Issue a SET TARGET PORT GROUP command and evaluate the
735  * response. Returns SCSI_DH_RETRY per default to trigger
736  * a re-evaluation of the target group state or SCSI_DH_OK
737  * if no further action needs to be taken.
738  */
739 static unsigned alua_stpg(struct scsi_device *sdev, struct alua_port_group *pg)
740 {
741         int retval;
742         struct scsi_sense_hdr sense_hdr;
743
744         if (!(pg->tpgs & TPGS_MODE_EXPLICIT)) {
745                 /* Only implicit ALUA supported, retry */
746                 return SCSI_DH_RETRY;
747         }
748         switch (pg->state) {
749         case SCSI_ACCESS_STATE_OPTIMAL:
750                 return SCSI_DH_OK;
751         case SCSI_ACCESS_STATE_ACTIVE:
752                 if ((pg->flags & ALUA_OPTIMIZE_STPG) &&
753                     !pg->pref &&
754                     (pg->tpgs & TPGS_MODE_IMPLICIT))
755                         return SCSI_DH_OK;
756                 break;
757         case SCSI_ACCESS_STATE_STANDBY:
758         case SCSI_ACCESS_STATE_UNAVAILABLE:
759                 break;
760         case SCSI_ACCESS_STATE_OFFLINE:
761                 return SCSI_DH_IO;
762         case SCSI_ACCESS_STATE_TRANSITIONING:
763                 break;
764         default:
765                 sdev_printk(KERN_INFO, sdev,
766                             "%s: stpg failed, unhandled TPGS state %d",
767                             ALUA_DH_NAME, pg->state);
768                 return SCSI_DH_NOSYS;
769         }
770         retval = submit_stpg(sdev, pg->group_id, &sense_hdr);
771
772         if (retval) {
773                 if (!scsi_sense_valid(&sense_hdr)) {
774                         sdev_printk(KERN_INFO, sdev,
775                                     "%s: stpg failed, result %d",
776                                     ALUA_DH_NAME, retval);
777                         if (driver_byte(retval) == DRIVER_ERROR)
778                                 return SCSI_DH_DEV_TEMP_BUSY;
779                 } else {
780                         sdev_printk(KERN_INFO, sdev, "%s: stpg failed\n",
781                                     ALUA_DH_NAME);
782                         scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
783                 }
784         }
785         /* Retry RTPG */
786         return SCSI_DH_RETRY;
787 }
788
789 static void alua_rtpg_work(struct work_struct *work)
790 {
791         struct alua_port_group *pg =
792                 container_of(work, struct alua_port_group, rtpg_work.work);
793         struct scsi_device *sdev;
794         LIST_HEAD(qdata_list);
795         int err = SCSI_DH_OK;
796         struct alua_queue_data *qdata, *tmp;
797         unsigned long flags;
798
799         spin_lock_irqsave(&pg->lock, flags);
800         sdev = pg->rtpg_sdev;
801         if (!sdev) {
802                 WARN_ON(pg->flags & ALUA_PG_RUN_RTPG);
803                 WARN_ON(pg->flags & ALUA_PG_RUN_STPG);
804                 spin_unlock_irqrestore(&pg->lock, flags);
805                 kref_put(&pg->kref, release_port_group);
806                 return;
807         }
808         pg->flags |= ALUA_PG_RUNNING;
809         if (pg->flags & ALUA_PG_RUN_RTPG) {
810                 int state = pg->state;
811
812                 pg->flags &= ~ALUA_PG_RUN_RTPG;
813                 spin_unlock_irqrestore(&pg->lock, flags);
814                 if (state == SCSI_ACCESS_STATE_TRANSITIONING) {
815                         if (alua_tur(sdev) == SCSI_DH_RETRY) {
816                                 spin_lock_irqsave(&pg->lock, flags);
817                                 pg->flags &= ~ALUA_PG_RUNNING;
818                                 pg->flags |= ALUA_PG_RUN_RTPG;
819                                 if (!pg->interval)
820                                         pg->interval = ALUA_RTPG_RETRY_DELAY;
821                                 spin_unlock_irqrestore(&pg->lock, flags);
822                                 queue_delayed_work(kaluad_wq, &pg->rtpg_work,
823                                                    pg->interval * HZ);
824                                 return;
825                         }
826                         /* Send RTPG on failure or if TUR indicates SUCCESS */
827                 }
828                 err = alua_rtpg(sdev, pg);
829                 spin_lock_irqsave(&pg->lock, flags);
830                 if (err == SCSI_DH_RETRY || pg->flags & ALUA_PG_RUN_RTPG) {
831                         pg->flags &= ~ALUA_PG_RUNNING;
832                         if (!pg->interval && !(pg->flags & ALUA_PG_RUN_RTPG))
833                                 pg->interval = ALUA_RTPG_RETRY_DELAY;
834                         pg->flags |= ALUA_PG_RUN_RTPG;
835                         spin_unlock_irqrestore(&pg->lock, flags);
836                         queue_delayed_work(kaluad_wq, &pg->rtpg_work,
837                                            pg->interval * HZ);
838                         return;
839                 }
840                 if (err != SCSI_DH_OK)
841                         pg->flags &= ~ALUA_PG_RUN_STPG;
842         }
843         if (pg->flags & ALUA_PG_RUN_STPG) {
844                 pg->flags &= ~ALUA_PG_RUN_STPG;
845                 spin_unlock_irqrestore(&pg->lock, flags);
846                 err = alua_stpg(sdev, pg);
847                 spin_lock_irqsave(&pg->lock, flags);
848                 if (err == SCSI_DH_RETRY || pg->flags & ALUA_PG_RUN_RTPG) {
849                         pg->flags |= ALUA_PG_RUN_RTPG;
850                         pg->interval = 0;
851                         pg->flags &= ~ALUA_PG_RUNNING;
852                         spin_unlock_irqrestore(&pg->lock, flags);
853                         queue_delayed_work(kaluad_wq, &pg->rtpg_work,
854                                            pg->interval * HZ);
855                         return;
856                 }
857         }
858
859         list_splice_init(&pg->rtpg_list, &qdata_list);
860         pg->rtpg_sdev = NULL;
861         spin_unlock_irqrestore(&pg->lock, flags);
862
863         list_for_each_entry_safe(qdata, tmp, &qdata_list, entry) {
864                 list_del(&qdata->entry);
865                 if (qdata->callback_fn)
866                         qdata->callback_fn(qdata->callback_data, err);
867                 kfree(qdata);
868         }
869         spin_lock_irqsave(&pg->lock, flags);
870         pg->flags &= ~ALUA_PG_RUNNING;
871         spin_unlock_irqrestore(&pg->lock, flags);
872         scsi_device_put(sdev);
873         kref_put(&pg->kref, release_port_group);
874 }
875
876 /**
877  * alua_rtpg_queue() - cause RTPG to be submitted asynchronously
878  *
879  * Returns true if and only if alua_rtpg_work() will be called asynchronously.
880  * That function is responsible for calling @qdata->fn().
881  */
882 static bool alua_rtpg_queue(struct alua_port_group *pg,
883                             struct scsi_device *sdev,
884                             struct alua_queue_data *qdata, bool force)
885 {
886         int start_queue = 0;
887         unsigned long flags;
888         if (WARN_ON_ONCE(!pg) || scsi_device_get(sdev))
889                 return false;
890
891         spin_lock_irqsave(&pg->lock, flags);
892         if (qdata) {
893                 list_add_tail(&qdata->entry, &pg->rtpg_list);
894                 pg->flags |= ALUA_PG_RUN_STPG;
895                 force = true;
896         }
897         if (pg->rtpg_sdev == NULL) {
898                 pg->interval = 0;
899                 pg->flags |= ALUA_PG_RUN_RTPG;
900                 kref_get(&pg->kref);
901                 pg->rtpg_sdev = sdev;
902                 start_queue = 1;
903         } else if (!(pg->flags & ALUA_PG_RUN_RTPG) && force) {
904                 pg->flags |= ALUA_PG_RUN_RTPG;
905                 /* Do not queue if the worker is already running */
906                 if (!(pg->flags & ALUA_PG_RUNNING)) {
907                         kref_get(&pg->kref);
908                         start_queue = 1;
909                 }
910         }
911
912         spin_unlock_irqrestore(&pg->lock, flags);
913
914         if (start_queue) {
915                 if (queue_delayed_work(kaluad_wq, &pg->rtpg_work,
916                                 msecs_to_jiffies(ALUA_RTPG_DELAY_MSECS)))
917                         sdev = NULL;
918                 else
919                         kref_put(&pg->kref, release_port_group);
920         }
921         if (sdev)
922                 scsi_device_put(sdev);
923
924         return true;
925 }
926
927 /*
928  * alua_initialize - Initialize ALUA state
929  * @sdev: the device to be initialized
930  *
931  * For the prep_fn to work correctly we have
932  * to initialize the ALUA state for the device.
933  */
934 static int alua_initialize(struct scsi_device *sdev, struct alua_dh_data *h)
935 {
936         int err = SCSI_DH_DEV_UNSUPP, tpgs;
937
938         mutex_lock(&h->init_mutex);
939         tpgs = alua_check_tpgs(sdev);
940         if (tpgs != TPGS_MODE_NONE)
941                 err = alua_check_vpd(sdev, h, tpgs);
942         h->init_error = err;
943         mutex_unlock(&h->init_mutex);
944         return err;
945 }
946 /*
947  * alua_set_params - set/unset the optimize flag
948  * @sdev: device on the path to be activated
949  * params - parameters in the following format
950  *      "no_of_params\0param1\0param2\0param3\0...\0"
951  * For example, to set the flag pass the following parameters
952  * from multipath.conf
953  *     hardware_handler        "2 alua 1"
954  */
955 static int alua_set_params(struct scsi_device *sdev, const char *params)
956 {
957         struct alua_dh_data *h = sdev->handler_data;
958         struct alua_port_group *pg = NULL;
959         unsigned int optimize = 0, argc;
960         const char *p = params;
961         int result = SCSI_DH_OK;
962         unsigned long flags;
963
964         if ((sscanf(params, "%u", &argc) != 1) || (argc != 1))
965                 return -EINVAL;
966
967         while (*p++)
968                 ;
969         if ((sscanf(p, "%u", &optimize) != 1) || (optimize > 1))
970                 return -EINVAL;
971
972         rcu_read_lock();
973         pg = rcu_dereference(h->pg);
974         if (!pg) {
975                 rcu_read_unlock();
976                 return -ENXIO;
977         }
978         spin_lock_irqsave(&pg->lock, flags);
979         if (optimize)
980                 pg->flags |= ALUA_OPTIMIZE_STPG;
981         else
982                 pg->flags &= ~ALUA_OPTIMIZE_STPG;
983         spin_unlock_irqrestore(&pg->lock, flags);
984         rcu_read_unlock();
985
986         return result;
987 }
988
989 /*
990  * alua_activate - activate a path
991  * @sdev: device on the path to be activated
992  *
993  * We're currently switching the port group to be activated only and
994  * let the array figure out the rest.
995  * There may be other arrays which require us to switch all port groups
996  * based on a certain policy. But until we actually encounter them it
997  * should be okay.
998  */
999 static int alua_activate(struct scsi_device *sdev,
1000                         activate_complete fn, void *data)
1001 {
1002         struct alua_dh_data *h = sdev->handler_data;
1003         int err = SCSI_DH_OK;
1004         struct alua_queue_data *qdata;
1005         struct alua_port_group *pg;
1006
1007         qdata = kzalloc(sizeof(*qdata), GFP_KERNEL);
1008         if (!qdata) {
1009                 err = SCSI_DH_RES_TEMP_UNAVAIL;
1010                 goto out;
1011         }
1012         qdata->callback_fn = fn;
1013         qdata->callback_data = data;
1014
1015         mutex_lock(&h->init_mutex);
1016         rcu_read_lock();
1017         pg = rcu_dereference(h->pg);
1018         if (!pg || !kref_get_unless_zero(&pg->kref)) {
1019                 rcu_read_unlock();
1020                 kfree(qdata);
1021                 err = h->init_error;
1022                 mutex_unlock(&h->init_mutex);
1023                 goto out;
1024         }
1025         rcu_read_unlock();
1026         mutex_unlock(&h->init_mutex);
1027
1028         if (alua_rtpg_queue(pg, sdev, qdata, true))
1029                 fn = NULL;
1030         else
1031                 err = SCSI_DH_DEV_OFFLINED;
1032         kref_put(&pg->kref, release_port_group);
1033 out:
1034         if (fn)
1035                 fn(data, err);
1036         return 0;
1037 }
1038
1039 /*
1040  * alua_check - check path status
1041  * @sdev: device on the path to be checked
1042  *
1043  * Check the device status
1044  */
1045 static void alua_check(struct scsi_device *sdev, bool force)
1046 {
1047         struct alua_dh_data *h = sdev->handler_data;
1048         struct alua_port_group *pg;
1049
1050         rcu_read_lock();
1051         pg = rcu_dereference(h->pg);
1052         if (!pg || !kref_get_unless_zero(&pg->kref)) {
1053                 rcu_read_unlock();
1054                 return;
1055         }
1056         rcu_read_unlock();
1057
1058         alua_rtpg_queue(pg, sdev, NULL, force);
1059         kref_put(&pg->kref, release_port_group);
1060 }
1061
1062 /*
1063  * alua_prep_fn - request callback
1064  *
1065  * Fail I/O to all paths not in state
1066  * active/optimized or active/non-optimized.
1067  */
1068 static int alua_prep_fn(struct scsi_device *sdev, struct request *req)
1069 {
1070         struct alua_dh_data *h = sdev->handler_data;
1071         struct alua_port_group *pg;
1072         unsigned char state = SCSI_ACCESS_STATE_OPTIMAL;
1073         int ret = BLKPREP_OK;
1074
1075         rcu_read_lock();
1076         pg = rcu_dereference(h->pg);
1077         if (pg)
1078                 state = pg->state;
1079         rcu_read_unlock();
1080         if (state == SCSI_ACCESS_STATE_TRANSITIONING)
1081                 ret = BLKPREP_DEFER;
1082         else if (state != SCSI_ACCESS_STATE_OPTIMAL &&
1083                  state != SCSI_ACCESS_STATE_ACTIVE &&
1084                  state != SCSI_ACCESS_STATE_LBA) {
1085                 ret = BLKPREP_KILL;
1086                 req->rq_flags |= RQF_QUIET;
1087         }
1088         return ret;
1089
1090 }
1091
1092 static void alua_rescan(struct scsi_device *sdev)
1093 {
1094         struct alua_dh_data *h = sdev->handler_data;
1095
1096         alua_initialize(sdev, h);
1097 }
1098
1099 /*
1100  * alua_bus_attach - Attach device handler
1101  * @sdev: device to be attached to
1102  */
1103 static int alua_bus_attach(struct scsi_device *sdev)
1104 {
1105         struct alua_dh_data *h;
1106         int err, ret = -EINVAL;
1107
1108         h = kzalloc(sizeof(*h) , GFP_KERNEL);
1109         if (!h)
1110                 return -ENOMEM;
1111         spin_lock_init(&h->pg_lock);
1112         rcu_assign_pointer(h->pg, NULL);
1113         h->init_error = SCSI_DH_OK;
1114         h->sdev = sdev;
1115         INIT_LIST_HEAD(&h->node);
1116
1117         mutex_init(&h->init_mutex);
1118         err = alua_initialize(sdev, h);
1119         if (err == SCSI_DH_NOMEM)
1120                 ret = -ENOMEM;
1121         if (err != SCSI_DH_OK && err != SCSI_DH_DEV_OFFLINED)
1122                 goto failed;
1123
1124         sdev->handler_data = h;
1125         return 0;
1126 failed:
1127         kfree(h);
1128         return ret;
1129 }
1130
1131 /*
1132  * alua_bus_detach - Detach device handler
1133  * @sdev: device to be detached from
1134  */
1135 static void alua_bus_detach(struct scsi_device *sdev)
1136 {
1137         struct alua_dh_data *h = sdev->handler_data;
1138         struct alua_port_group *pg;
1139
1140         spin_lock(&h->pg_lock);
1141         pg = rcu_dereference_protected(h->pg, lockdep_is_held(&h->pg_lock));
1142         rcu_assign_pointer(h->pg, NULL);
1143         spin_unlock(&h->pg_lock);
1144         if (pg) {
1145                 spin_lock_irq(&pg->lock);
1146                 list_del_rcu(&h->node);
1147                 spin_unlock_irq(&pg->lock);
1148                 kref_put(&pg->kref, release_port_group);
1149         }
1150         sdev->handler_data = NULL;
1151         synchronize_rcu();
1152         kfree(h);
1153 }
1154
1155 static struct scsi_device_handler alua_dh = {
1156         .name = ALUA_DH_NAME,
1157         .module = THIS_MODULE,
1158         .attach = alua_bus_attach,
1159         .detach = alua_bus_detach,
1160         .prep_fn = alua_prep_fn,
1161         .check_sense = alua_check_sense,
1162         .activate = alua_activate,
1163         .rescan = alua_rescan,
1164         .set_params = alua_set_params,
1165 };
1166
1167 static int __init alua_init(void)
1168 {
1169         int r;
1170
1171         kaluad_wq = alloc_workqueue("kaluad", WQ_MEM_RECLAIM, 0);
1172         if (!kaluad_wq)
1173                 return -ENOMEM;
1174
1175         r = scsi_register_device_handler(&alua_dh);
1176         if (r != 0) {
1177                 printk(KERN_ERR "%s: Failed to register scsi device handler",
1178                         ALUA_DH_NAME);
1179                 destroy_workqueue(kaluad_wq);
1180         }
1181         return r;
1182 }
1183
1184 static void __exit alua_exit(void)
1185 {
1186         scsi_unregister_device_handler(&alua_dh);
1187         destroy_workqueue(kaluad_wq);
1188 }
1189
1190 module_init(alua_init);
1191 module_exit(alua_exit);
1192
1193 MODULE_DESCRIPTION("DM Multipath ALUA support");
1194 MODULE_AUTHOR("Hannes Reinecke <hare@suse.de>");
1195 MODULE_LICENSE("GPL");
1196 MODULE_VERSION(ALUA_DH_VER);