GNU Linux-libre 4.19.295-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 #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 #define ALUA_RTPG_RETRY_DELAY           2
58
59 /* device handler flags */
60 #define ALUA_OPTIMIZE_STPG              0x01
61 #define ALUA_RTPG_EXT_HDR_UNSUPP        0x02
62 /* State machine flags */
63 #define ALUA_PG_RUN_RTPG                0x10
64 #define ALUA_PG_RUN_STPG                0x20
65 #define ALUA_PG_RUNNING                 0x40
66
67 static uint optimize_stpg;
68 module_param(optimize_stpg, uint, S_IRUGO|S_IWUSR);
69 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.");
70
71 static LIST_HEAD(port_group_list);
72 static DEFINE_SPINLOCK(port_group_lock);
73 static struct workqueue_struct *kaluad_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         int                     valid_states;
87         unsigned                flags; /* used for optimizing STPG */
88         unsigned char           transition_tmo;
89         unsigned long           expiry;
90         unsigned long           interval;
91         struct delayed_work     rtpg_work;
92         spinlock_t              lock;
93         struct list_head        rtpg_list;
94         struct scsi_device      *rtpg_sdev;
95 };
96
97 struct alua_dh_data {
98         struct list_head        node;
99         struct alua_port_group __rcu *pg;
100         int                     group_id;
101         spinlock_t              pg_lock;
102         struct scsi_device      *sdev;
103         int                     init_error;
104         struct mutex            init_mutex;
105 };
106
107 struct alua_queue_data {
108         struct list_head        entry;
109         activate_complete       callback_fn;
110         void                    *callback_data;
111 };
112
113 #define ALUA_POLICY_SWITCH_CURRENT      0
114 #define ALUA_POLICY_SWITCH_ALL          1
115
116 static void alua_rtpg_work(struct work_struct *work);
117 static bool alua_rtpg_queue(struct alua_port_group *pg,
118                             struct scsi_device *sdev,
119                             struct alua_queue_data *qdata, bool force);
120 static void alua_check(struct scsi_device *sdev, bool force);
121
122 static void release_port_group(struct kref *kref)
123 {
124         struct alua_port_group *pg;
125
126         pg = container_of(kref, struct alua_port_group, kref);
127         if (pg->rtpg_sdev)
128                 flush_delayed_work(&pg->rtpg_work);
129         spin_lock(&port_group_lock);
130         list_del(&pg->node);
131         spin_unlock(&port_group_lock);
132         kfree_rcu(pg, rcu);
133 }
134
135 /*
136  * submit_rtpg - Issue a REPORT TARGET GROUP STATES command
137  * @sdev: sdev the command should be sent to
138  */
139 static int submit_rtpg(struct scsi_device *sdev, unsigned char *buff,
140                        int bufflen, struct scsi_sense_hdr *sshdr, int flags)
141 {
142         u8 cdb[MAX_COMMAND_SIZE];
143         int req_flags = REQ_FAILFAST_DEV | REQ_FAILFAST_TRANSPORT |
144                 REQ_FAILFAST_DRIVER;
145
146         /* Prepare the command. */
147         memset(cdb, 0x0, MAX_COMMAND_SIZE);
148         cdb[0] = MAINTENANCE_IN;
149         if (!(flags & ALUA_RTPG_EXT_HDR_UNSUPP))
150                 cdb[1] = MI_REPORT_TARGET_PGS | MI_EXT_HDR_PARAM_FMT;
151         else
152                 cdb[1] = MI_REPORT_TARGET_PGS;
153         put_unaligned_be32(bufflen, &cdb[6]);
154
155         return scsi_execute(sdev, cdb, DMA_FROM_DEVICE, buff, bufflen, NULL,
156                         sshdr, ALUA_FAILOVER_TIMEOUT * HZ,
157                         ALUA_FAILOVER_RETRIES, req_flags, 0, NULL);
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[MAX_COMMAND_SIZE];
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, MAX_COMMAND_SIZE);
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(sdev, cdb, DMA_TO_DEVICE, stpg_data, stpg_len, NULL,
188                         sshdr, ALUA_FAILOVER_TIMEOUT * HZ,
189                         ALUA_FAILOVER_RETRIES, req_flags, 0, NULL);
190 }
191
192 static struct alua_port_group *alua_find_get_pg(char *id_str, size_t id_size,
193                                                 int group_id)
194 {
195         struct alua_port_group *pg;
196
197         if (!id_str || !id_size || !strlen(id_str))
198                 return NULL;
199
200         list_for_each_entry(pg, &port_group_list, node) {
201                 if (pg->group_id != group_id)
202                         continue;
203                 if (!pg->device_id_len || pg->device_id_len != id_size)
204                         continue;
205                 if (strncmp(pg->device_id_str, id_str, id_size))
206                         continue;
207                 if (!kref_get_unless_zero(&pg->kref))
208                         continue;
209                 return pg;
210         }
211
212         return NULL;
213 }
214
215 /*
216  * alua_alloc_pg - Allocate a new port_group structure
217  * @sdev: scsi device
218  * @group_id: port group id
219  * @tpgs: target port group settings
220  *
221  * Allocate a new port_group structure for a given
222  * device.
223  */
224 static struct alua_port_group *alua_alloc_pg(struct scsi_device *sdev,
225                                              int group_id, int tpgs)
226 {
227         struct alua_port_group *pg, *tmp_pg;
228
229         pg = kzalloc(sizeof(struct alua_port_group), GFP_KERNEL);
230         if (!pg)
231                 return ERR_PTR(-ENOMEM);
232
233         pg->device_id_len = scsi_vpd_lun_id(sdev, pg->device_id_str,
234                                             sizeof(pg->device_id_str));
235         if (pg->device_id_len <= 0) {
236                 /*
237                  * TPGS supported but no device identification found.
238                  * Generate private device identification.
239                  */
240                 sdev_printk(KERN_INFO, sdev,
241                             "%s: No device descriptors found\n",
242                             ALUA_DH_NAME);
243                 pg->device_id_str[0] = '\0';
244                 pg->device_id_len = 0;
245         }
246         pg->group_id = group_id;
247         pg->tpgs = tpgs;
248         pg->state = SCSI_ACCESS_STATE_OPTIMAL;
249         pg->valid_states = TPGS_SUPPORT_ALL;
250         if (optimize_stpg)
251                 pg->flags |= ALUA_OPTIMIZE_STPG;
252         kref_init(&pg->kref);
253         INIT_DELAYED_WORK(&pg->rtpg_work, alua_rtpg_work);
254         INIT_LIST_HEAD(&pg->rtpg_list);
255         INIT_LIST_HEAD(&pg->node);
256         INIT_LIST_HEAD(&pg->dh_list);
257         spin_lock_init(&pg->lock);
258
259         spin_lock(&port_group_lock);
260         tmp_pg = alua_find_get_pg(pg->device_id_str, pg->device_id_len,
261                                   group_id);
262         if (tmp_pg) {
263                 spin_unlock(&port_group_lock);
264                 kfree(pg);
265                 return tmp_pg;
266         }
267
268         list_add(&pg->node, &port_group_list);
269         spin_unlock(&port_group_lock);
270
271         return pg;
272 }
273
274 /*
275  * alua_check_tpgs - Evaluate TPGS setting
276  * @sdev: device to be checked
277  *
278  * Examine the TPGS setting of the sdev to find out if ALUA
279  * is supported.
280  */
281 static int alua_check_tpgs(struct scsi_device *sdev)
282 {
283         int tpgs = TPGS_MODE_NONE;
284
285         /*
286          * ALUA support for non-disk devices is fraught with
287          * difficulties, so disable it for now.
288          */
289         if (sdev->type != TYPE_DISK) {
290                 sdev_printk(KERN_INFO, sdev,
291                             "%s: disable for non-disk devices\n",
292                             ALUA_DH_NAME);
293                 return tpgs;
294         }
295
296         tpgs = scsi_device_tpgs(sdev);
297         switch (tpgs) {
298         case TPGS_MODE_EXPLICIT|TPGS_MODE_IMPLICIT:
299                 sdev_printk(KERN_INFO, sdev,
300                             "%s: supports implicit and explicit TPGS\n",
301                             ALUA_DH_NAME);
302                 break;
303         case TPGS_MODE_EXPLICIT:
304                 sdev_printk(KERN_INFO, sdev, "%s: supports explicit TPGS\n",
305                             ALUA_DH_NAME);
306                 break;
307         case TPGS_MODE_IMPLICIT:
308                 sdev_printk(KERN_INFO, sdev, "%s: supports implicit TPGS\n",
309                             ALUA_DH_NAME);
310                 break;
311         case TPGS_MODE_NONE:
312                 sdev_printk(KERN_INFO, sdev, "%s: not supported\n",
313                             ALUA_DH_NAME);
314                 break;
315         default:
316                 sdev_printk(KERN_INFO, sdev,
317                             "%s: unsupported TPGS setting %d\n",
318                             ALUA_DH_NAME, tpgs);
319                 tpgs = TPGS_MODE_NONE;
320                 break;
321         }
322
323         return tpgs;
324 }
325
326 /*
327  * alua_check_vpd - Evaluate INQUIRY vpd page 0x83
328  * @sdev: device to be checked
329  *
330  * Extract the relative target port and the target port group
331  * descriptor from the list of identificators.
332  */
333 static int alua_check_vpd(struct scsi_device *sdev, struct alua_dh_data *h,
334                           int tpgs)
335 {
336         int rel_port = -1, group_id;
337         struct alua_port_group *pg, *old_pg = NULL;
338         bool pg_updated = false;
339         unsigned long flags;
340
341         group_id = scsi_vpd_tpg_id(sdev, &rel_port);
342         if (group_id < 0) {
343                 /*
344                  * Internal error; TPGS supported but required
345                  * VPD identification descriptors not present.
346                  * Disable ALUA support
347                  */
348                 sdev_printk(KERN_INFO, sdev,
349                             "%s: No target port descriptors found\n",
350                             ALUA_DH_NAME);
351                 return SCSI_DH_DEV_UNSUPP;
352         }
353
354         pg = alua_alloc_pg(sdev, group_id, tpgs);
355         if (IS_ERR(pg)) {
356                 if (PTR_ERR(pg) == -ENOMEM)
357                         return SCSI_DH_NOMEM;
358                 return SCSI_DH_DEV_UNSUPP;
359         }
360         if (pg->device_id_len)
361                 sdev_printk(KERN_INFO, sdev,
362                             "%s: device %s port group %x rel port %x\n",
363                             ALUA_DH_NAME, pg->device_id_str,
364                             group_id, rel_port);
365         else
366                 sdev_printk(KERN_INFO, sdev,
367                             "%s: port group %x rel port %x\n",
368                             ALUA_DH_NAME, group_id, rel_port);
369
370         /* Check for existing port group references */
371         spin_lock(&h->pg_lock);
372         old_pg = rcu_dereference_protected(h->pg, lockdep_is_held(&h->pg_lock));
373         if (old_pg != pg) {
374                 /* port group has changed. Update to new port group */
375                 if (h->pg) {
376                         spin_lock_irqsave(&old_pg->lock, flags);
377                         list_del_rcu(&h->node);
378                         spin_unlock_irqrestore(&old_pg->lock, flags);
379                 }
380                 rcu_assign_pointer(h->pg, pg);
381                 pg_updated = true;
382         }
383
384         spin_lock_irqsave(&pg->lock, flags);
385         if (pg_updated)
386                 list_add_rcu(&h->node, &pg->dh_list);
387         spin_unlock_irqrestore(&pg->lock, flags);
388
389         alua_rtpg_queue(rcu_dereference_protected(h->pg,
390                                                   lockdep_is_held(&h->pg_lock)),
391                         sdev, NULL, true);
392         spin_unlock(&h->pg_lock);
393
394         if (old_pg)
395                 kref_put(&old_pg->kref, release_port_group);
396
397         return SCSI_DH_OK;
398 }
399
400 static char print_alua_state(unsigned char state)
401 {
402         switch (state) {
403         case SCSI_ACCESS_STATE_OPTIMAL:
404                 return 'A';
405         case SCSI_ACCESS_STATE_ACTIVE:
406                 return 'N';
407         case SCSI_ACCESS_STATE_STANDBY:
408                 return 'S';
409         case SCSI_ACCESS_STATE_UNAVAILABLE:
410                 return 'U';
411         case SCSI_ACCESS_STATE_LBA:
412                 return 'L';
413         case SCSI_ACCESS_STATE_OFFLINE:
414                 return 'O';
415         case SCSI_ACCESS_STATE_TRANSITIONING:
416                 return 'T';
417         default:
418                 return 'X';
419         }
420 }
421
422 static int alua_check_sense(struct scsi_device *sdev,
423                             struct scsi_sense_hdr *sense_hdr)
424 {
425         switch (sense_hdr->sense_key) {
426         case NOT_READY:
427                 if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0a) {
428                         /*
429                          * LUN Not Accessible - ALUA state transition
430                          */
431                         alua_check(sdev, false);
432                         return NEEDS_RETRY;
433                 }
434                 break;
435         case UNIT_ATTENTION:
436                 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x00) {
437                         /*
438                          * Power On, Reset, or Bus Device Reset.
439                          * Might have obscured a state transition,
440                          * so schedule a recheck.
441                          */
442                         alua_check(sdev, true);
443                         return ADD_TO_MLQUEUE;
444                 }
445                 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x04)
446                         /*
447                          * Device internal reset
448                          */
449                         return ADD_TO_MLQUEUE;
450                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x01)
451                         /*
452                          * Mode Parameters Changed
453                          */
454                         return ADD_TO_MLQUEUE;
455                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x06) {
456                         /*
457                          * ALUA state changed
458                          */
459                         alua_check(sdev, true);
460                         return ADD_TO_MLQUEUE;
461                 }
462                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x07) {
463                         /*
464                          * Implicit ALUA state transition failed
465                          */
466                         alua_check(sdev, true);
467                         return ADD_TO_MLQUEUE;
468                 }
469                 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x03)
470                         /*
471                          * Inquiry data has changed
472                          */
473                         return ADD_TO_MLQUEUE;
474                 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x0e)
475                         /*
476                          * REPORTED_LUNS_DATA_HAS_CHANGED is reported
477                          * when switching controllers on targets like
478                          * Intel Multi-Flex. We can just retry.
479                          */
480                         return ADD_TO_MLQUEUE;
481                 break;
482         }
483
484         return SCSI_RETURN_NOT_HANDLED;
485 }
486
487 /*
488  * alua_tur - Send a TEST UNIT READY
489  * @sdev: device to which the TEST UNIT READY command should be send
490  *
491  * Send a TEST UNIT READY to @sdev to figure out the device state
492  * Returns SCSI_DH_RETRY if the sense code is NOT READY/ALUA TRANSITIONING,
493  * SCSI_DH_OK if no error occurred, and SCSI_DH_IO otherwise.
494  */
495 static int alua_tur(struct scsi_device *sdev)
496 {
497         struct scsi_sense_hdr sense_hdr;
498         int retval;
499
500         retval = scsi_test_unit_ready(sdev, ALUA_FAILOVER_TIMEOUT * HZ,
501                                       ALUA_FAILOVER_RETRIES, &sense_hdr);
502         if (sense_hdr.sense_key == NOT_READY &&
503             sense_hdr.asc == 0x04 && sense_hdr.ascq == 0x0a)
504                 return SCSI_DH_RETRY;
505         else if (retval)
506                 return SCSI_DH_IO;
507         else
508                 return SCSI_DH_OK;
509 }
510
511 /*
512  * alua_rtpg - Evaluate REPORT TARGET GROUP STATES
513  * @sdev: the device to be evaluated.
514  *
515  * Evaluate the Target Port Group State.
516  * Returns SCSI_DH_DEV_OFFLINED if the path is
517  * found to be unusable.
518  */
519 static int alua_rtpg(struct scsi_device *sdev, struct alua_port_group *pg)
520 {
521         struct scsi_sense_hdr sense_hdr;
522         struct alua_port_group *tmp_pg;
523         int len, k, off, bufflen = ALUA_RTPG_SIZE;
524         unsigned char *desc, *buff;
525         unsigned err;
526         int retval;
527         unsigned int tpg_desc_tbl_off;
528         unsigned char orig_transition_tmo;
529         unsigned long flags;
530         bool transitioning_sense = false;
531
532         if (!pg->expiry) {
533                 unsigned long transition_tmo = ALUA_FAILOVER_TIMEOUT * HZ;
534
535                 if (pg->transition_tmo)
536                         transition_tmo = pg->transition_tmo * HZ;
537
538                 pg->expiry = round_jiffies_up(jiffies + transition_tmo);
539         }
540
541         buff = kzalloc(bufflen, GFP_KERNEL);
542         if (!buff)
543                 return SCSI_DH_DEV_TEMP_BUSY;
544
545  retry:
546         err = 0;
547         retval = submit_rtpg(sdev, buff, bufflen, &sense_hdr, pg->flags);
548
549         if (retval) {
550                 /*
551                  * Some (broken) implementations have a habit of returning
552                  * an error during things like firmware update etc.
553                  * But if the target only supports active/optimized there's
554                  * not much we can do; it's not that we can switch paths
555                  * or anything.
556                  * So ignore any errors to avoid spurious failures during
557                  * path failover.
558                  */
559                 if ((pg->valid_states & ~TPGS_SUPPORT_OPTIMIZED) == 0) {
560                         sdev_printk(KERN_INFO, sdev,
561                                     "%s: ignoring rtpg result %d\n",
562                                     ALUA_DH_NAME, retval);
563                         kfree(buff);
564                         return SCSI_DH_OK;
565                 }
566                 if (retval < 0 || !scsi_sense_valid(&sense_hdr)) {
567                         sdev_printk(KERN_INFO, sdev,
568                                     "%s: rtpg failed, result %d\n",
569                                     ALUA_DH_NAME, retval);
570                         kfree(buff);
571                         if (retval < 0)
572                                 return SCSI_DH_DEV_TEMP_BUSY;
573                         return SCSI_DH_IO;
574                 }
575
576                 /*
577                  * submit_rtpg() has failed on existing arrays
578                  * when requesting extended header info, and
579                  * the array doesn't support extended headers,
580                  * even though it shouldn't according to T10.
581                  * The retry without rtpg_ext_hdr_req set
582                  * handles this.
583                  * Note:  some arrays return a sense key of ILLEGAL_REQUEST
584                  * with ASC 00h if they don't support the extended header.
585                  */
586                 if (!(pg->flags & ALUA_RTPG_EXT_HDR_UNSUPP) &&
587                     sense_hdr.sense_key == ILLEGAL_REQUEST) {
588                         pg->flags |= ALUA_RTPG_EXT_HDR_UNSUPP;
589                         goto retry;
590                 }
591                 /*
592                  * If the array returns with 'ALUA state transition'
593                  * sense code here it cannot return RTPG data during
594                  * transition. So set the state to 'transitioning' directly.
595                  */
596                 if (sense_hdr.sense_key == NOT_READY &&
597                     sense_hdr.asc == 0x04 && sense_hdr.ascq == 0x0a) {
598                         transitioning_sense = true;
599                         goto skip_rtpg;
600                 }
601                 /*
602                  * Retry on any other UNIT ATTENTION occurred.
603                  */
604                 if (sense_hdr.sense_key == UNIT_ATTENTION)
605                         err = SCSI_DH_RETRY;
606                 if (err == SCSI_DH_RETRY &&
607                     pg->expiry != 0 && time_before(jiffies, pg->expiry)) {
608                         sdev_printk(KERN_ERR, sdev, "%s: rtpg retry\n",
609                                     ALUA_DH_NAME);
610                         scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
611                         kfree(buff);
612                         return err;
613                 }
614                 sdev_printk(KERN_ERR, sdev, "%s: rtpg failed\n",
615                             ALUA_DH_NAME);
616                 scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
617                 kfree(buff);
618                 pg->expiry = 0;
619                 return SCSI_DH_IO;
620         }
621
622         len = get_unaligned_be32(&buff[0]) + 4;
623
624         if (len > bufflen) {
625                 /* Resubmit with the correct length */
626                 kfree(buff);
627                 bufflen = len;
628                 buff = kmalloc(bufflen, GFP_KERNEL);
629                 if (!buff) {
630                         sdev_printk(KERN_WARNING, sdev,
631                                     "%s: kmalloc buffer failed\n",__func__);
632                         /* Temporary failure, bypass */
633                         pg->expiry = 0;
634                         return SCSI_DH_DEV_TEMP_BUSY;
635                 }
636                 goto retry;
637         }
638
639         orig_transition_tmo = pg->transition_tmo;
640         if ((buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR && buff[5] != 0)
641                 pg->transition_tmo = buff[5];
642         else
643                 pg->transition_tmo = ALUA_FAILOVER_TIMEOUT;
644
645         if (orig_transition_tmo != pg->transition_tmo) {
646                 sdev_printk(KERN_INFO, sdev,
647                             "%s: transition timeout set to %d seconds\n",
648                             ALUA_DH_NAME, pg->transition_tmo);
649                 pg->expiry = jiffies + pg->transition_tmo * HZ;
650         }
651
652         if ((buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR)
653                 tpg_desc_tbl_off = 8;
654         else
655                 tpg_desc_tbl_off = 4;
656
657         for (k = tpg_desc_tbl_off, desc = buff + tpg_desc_tbl_off;
658              k < len;
659              k += off, desc += off) {
660                 u16 group_id = get_unaligned_be16(&desc[2]);
661
662                 spin_lock_irqsave(&port_group_lock, flags);
663                 tmp_pg = alua_find_get_pg(pg->device_id_str, pg->device_id_len,
664                                           group_id);
665                 spin_unlock_irqrestore(&port_group_lock, flags);
666                 if (tmp_pg) {
667                         if (spin_trylock_irqsave(&tmp_pg->lock, flags)) {
668                                 if ((tmp_pg == pg) ||
669                                     !(tmp_pg->flags & ALUA_PG_RUNNING)) {
670                                         struct alua_dh_data *h;
671
672                                         tmp_pg->state = desc[0] & 0x0f;
673                                         tmp_pg->pref = desc[0] >> 7;
674                                         rcu_read_lock();
675                                         list_for_each_entry_rcu(h,
676                                                 &tmp_pg->dh_list, node) {
677                                                 if (!h->sdev)
678                                                         continue;
679                                                 h->sdev->access_state = desc[0];
680                                         }
681                                         rcu_read_unlock();
682                                 }
683                                 if (tmp_pg == pg)
684                                         tmp_pg->valid_states = desc[1];
685                                 spin_unlock_irqrestore(&tmp_pg->lock, flags);
686                         }
687                         kref_put(&tmp_pg->kref, release_port_group);
688                 }
689                 off = 8 + (desc[7] * 4);
690         }
691
692  skip_rtpg:
693         spin_lock_irqsave(&pg->lock, flags);
694         if (transitioning_sense)
695                 pg->state = SCSI_ACCESS_STATE_TRANSITIONING;
696
697         sdev_printk(KERN_INFO, sdev,
698                     "%s: port group %02x state %c %s supports %c%c%c%c%c%c%c\n",
699                     ALUA_DH_NAME, pg->group_id, print_alua_state(pg->state),
700                     pg->pref ? "preferred" : "non-preferred",
701                     pg->valid_states&TPGS_SUPPORT_TRANSITION?'T':'t',
702                     pg->valid_states&TPGS_SUPPORT_OFFLINE?'O':'o',
703                     pg->valid_states&TPGS_SUPPORT_LBA_DEPENDENT?'L':'l',
704                     pg->valid_states&TPGS_SUPPORT_UNAVAILABLE?'U':'u',
705                     pg->valid_states&TPGS_SUPPORT_STANDBY?'S':'s',
706                     pg->valid_states&TPGS_SUPPORT_NONOPTIMIZED?'N':'n',
707                     pg->valid_states&TPGS_SUPPORT_OPTIMIZED?'A':'a');
708
709         switch (pg->state) {
710         case SCSI_ACCESS_STATE_TRANSITIONING:
711                 if (time_before(jiffies, pg->expiry)) {
712                         /* State transition, retry */
713                         pg->interval = ALUA_RTPG_RETRY_DELAY;
714                         err = SCSI_DH_RETRY;
715                 } else {
716                         struct alua_dh_data *h;
717
718                         /* Transitioning time exceeded, set port to standby */
719                         err = SCSI_DH_IO;
720                         pg->state = SCSI_ACCESS_STATE_STANDBY;
721                         pg->expiry = 0;
722                         rcu_read_lock();
723                         list_for_each_entry_rcu(h, &pg->dh_list, node) {
724                                 if (!h->sdev)
725                                         continue;
726                                 h->sdev->access_state =
727                                         (pg->state & SCSI_ACCESS_STATE_MASK);
728                                 if (pg->pref)
729                                         h->sdev->access_state |=
730                                                 SCSI_ACCESS_STATE_PREFERRED;
731                         }
732                         rcu_read_unlock();
733                 }
734                 break;
735         case SCSI_ACCESS_STATE_OFFLINE:
736                 /* Path unusable */
737                 err = SCSI_DH_DEV_OFFLINED;
738                 pg->expiry = 0;
739                 break;
740         default:
741                 /* Useable path if active */
742                 err = SCSI_DH_OK;
743                 pg->expiry = 0;
744                 break;
745         }
746         spin_unlock_irqrestore(&pg->lock, flags);
747         kfree(buff);
748         return err;
749 }
750
751 /*
752  * alua_stpg - Issue a SET TARGET PORT GROUP command
753  *
754  * Issue a SET TARGET PORT GROUP command and evaluate the
755  * response. Returns SCSI_DH_RETRY per default to trigger
756  * a re-evaluation of the target group state or SCSI_DH_OK
757  * if no further action needs to be taken.
758  */
759 static unsigned alua_stpg(struct scsi_device *sdev, struct alua_port_group *pg)
760 {
761         int retval;
762         struct scsi_sense_hdr sense_hdr;
763
764         if (!(pg->tpgs & TPGS_MODE_EXPLICIT)) {
765                 /* Only implicit ALUA supported, retry */
766                 return SCSI_DH_RETRY;
767         }
768         switch (pg->state) {
769         case SCSI_ACCESS_STATE_OPTIMAL:
770                 return SCSI_DH_OK;
771         case SCSI_ACCESS_STATE_ACTIVE:
772                 if ((pg->flags & ALUA_OPTIMIZE_STPG) &&
773                     !pg->pref &&
774                     (pg->tpgs & TPGS_MODE_IMPLICIT))
775                         return SCSI_DH_OK;
776                 break;
777         case SCSI_ACCESS_STATE_STANDBY:
778         case SCSI_ACCESS_STATE_UNAVAILABLE:
779                 break;
780         case SCSI_ACCESS_STATE_OFFLINE:
781                 return SCSI_DH_IO;
782         case SCSI_ACCESS_STATE_TRANSITIONING:
783                 break;
784         default:
785                 sdev_printk(KERN_INFO, sdev,
786                             "%s: stpg failed, unhandled TPGS state %d",
787                             ALUA_DH_NAME, pg->state);
788                 return SCSI_DH_NOSYS;
789         }
790         retval = submit_stpg(sdev, pg->group_id, &sense_hdr);
791
792         if (retval) {
793                 if (retval < 0 || !scsi_sense_valid(&sense_hdr)) {
794                         sdev_printk(KERN_INFO, sdev,
795                                     "%s: stpg failed, result %d",
796                                     ALUA_DH_NAME, retval);
797                         if (retval < 0)
798                                 return SCSI_DH_DEV_TEMP_BUSY;
799                 } else {
800                         sdev_printk(KERN_INFO, sdev, "%s: stpg failed\n",
801                                     ALUA_DH_NAME);
802                         scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
803                 }
804         }
805         /* Retry RTPG */
806         return SCSI_DH_RETRY;
807 }
808
809 static void alua_rtpg_work(struct work_struct *work)
810 {
811         struct alua_port_group *pg =
812                 container_of(work, struct alua_port_group, rtpg_work.work);
813         struct scsi_device *sdev;
814         LIST_HEAD(qdata_list);
815         int err = SCSI_DH_OK;
816         struct alua_queue_data *qdata, *tmp;
817         unsigned long flags;
818
819         spin_lock_irqsave(&pg->lock, flags);
820         sdev = pg->rtpg_sdev;
821         if (!sdev) {
822                 WARN_ON(pg->flags & ALUA_PG_RUN_RTPG);
823                 WARN_ON(pg->flags & ALUA_PG_RUN_STPG);
824                 spin_unlock_irqrestore(&pg->lock, flags);
825                 kref_put(&pg->kref, release_port_group);
826                 return;
827         }
828         pg->flags |= ALUA_PG_RUNNING;
829         if (pg->flags & ALUA_PG_RUN_RTPG) {
830                 int state = pg->state;
831
832                 pg->flags &= ~ALUA_PG_RUN_RTPG;
833                 spin_unlock_irqrestore(&pg->lock, flags);
834                 if (state == SCSI_ACCESS_STATE_TRANSITIONING) {
835                         if (alua_tur(sdev) == SCSI_DH_RETRY) {
836                                 spin_lock_irqsave(&pg->lock, flags);
837                                 pg->flags &= ~ALUA_PG_RUNNING;
838                                 pg->flags |= ALUA_PG_RUN_RTPG;
839                                 if (!pg->interval)
840                                         pg->interval = ALUA_RTPG_RETRY_DELAY;
841                                 spin_unlock_irqrestore(&pg->lock, flags);
842                                 queue_delayed_work(kaluad_wq, &pg->rtpg_work,
843                                                    pg->interval * HZ);
844                                 return;
845                         }
846                         /* Send RTPG on failure or if TUR indicates SUCCESS */
847                 }
848                 err = alua_rtpg(sdev, pg);
849                 spin_lock_irqsave(&pg->lock, flags);
850                 if (err == SCSI_DH_RETRY || pg->flags & ALUA_PG_RUN_RTPG) {
851                         pg->flags &= ~ALUA_PG_RUNNING;
852                         if (!pg->interval && !(pg->flags & ALUA_PG_RUN_RTPG))
853                                 pg->interval = ALUA_RTPG_RETRY_DELAY;
854                         pg->flags |= ALUA_PG_RUN_RTPG;
855                         spin_unlock_irqrestore(&pg->lock, flags);
856                         queue_delayed_work(kaluad_wq, &pg->rtpg_work,
857                                            pg->interval * HZ);
858                         return;
859                 }
860                 if (err != SCSI_DH_OK)
861                         pg->flags &= ~ALUA_PG_RUN_STPG;
862         }
863         if (pg->flags & ALUA_PG_RUN_STPG) {
864                 pg->flags &= ~ALUA_PG_RUN_STPG;
865                 spin_unlock_irqrestore(&pg->lock, flags);
866                 err = alua_stpg(sdev, pg);
867                 spin_lock_irqsave(&pg->lock, flags);
868                 if (err == SCSI_DH_RETRY || pg->flags & ALUA_PG_RUN_RTPG) {
869                         pg->flags |= ALUA_PG_RUN_RTPG;
870                         pg->interval = 0;
871                         pg->flags &= ~ALUA_PG_RUNNING;
872                         spin_unlock_irqrestore(&pg->lock, flags);
873                         queue_delayed_work(kaluad_wq, &pg->rtpg_work,
874                                            pg->interval * HZ);
875                         return;
876                 }
877         }
878
879         list_splice_init(&pg->rtpg_list, &qdata_list);
880         pg->rtpg_sdev = NULL;
881         spin_unlock_irqrestore(&pg->lock, flags);
882
883         list_for_each_entry_safe(qdata, tmp, &qdata_list, entry) {
884                 list_del(&qdata->entry);
885                 if (qdata->callback_fn)
886                         qdata->callback_fn(qdata->callback_data, err);
887                 kfree(qdata);
888         }
889         spin_lock_irqsave(&pg->lock, flags);
890         pg->flags &= ~ALUA_PG_RUNNING;
891         spin_unlock_irqrestore(&pg->lock, flags);
892         scsi_device_put(sdev);
893         kref_put(&pg->kref, release_port_group);
894 }
895
896 /**
897  * alua_rtpg_queue() - cause RTPG to be submitted asynchronously
898  * @pg: ALUA port group associated with @sdev.
899  * @sdev: SCSI device for which to submit an RTPG.
900  * @qdata: Information about the callback to invoke after the RTPG.
901  * @force: Whether or not to submit an RTPG if a work item that will submit an
902  *         RTPG already has been scheduled.
903  *
904  * Returns true if and only if alua_rtpg_work() will be called asynchronously.
905  * That function is responsible for calling @qdata->fn().
906  */
907 static bool alua_rtpg_queue(struct alua_port_group *pg,
908                             struct scsi_device *sdev,
909                             struct alua_queue_data *qdata, bool force)
910 {
911         int start_queue = 0;
912         unsigned long flags;
913         if (WARN_ON_ONCE(!pg) || scsi_device_get(sdev))
914                 return false;
915
916         spin_lock_irqsave(&pg->lock, flags);
917         if (qdata) {
918                 list_add_tail(&qdata->entry, &pg->rtpg_list);
919                 pg->flags |= ALUA_PG_RUN_STPG;
920                 force = true;
921         }
922         if (pg->rtpg_sdev == NULL) {
923                 pg->interval = 0;
924                 pg->flags |= ALUA_PG_RUN_RTPG;
925                 kref_get(&pg->kref);
926                 pg->rtpg_sdev = sdev;
927                 start_queue = 1;
928         } else if (!(pg->flags & ALUA_PG_RUN_RTPG) && force) {
929                 pg->flags |= ALUA_PG_RUN_RTPG;
930                 /* Do not queue if the worker is already running */
931                 if (!(pg->flags & ALUA_PG_RUNNING)) {
932                         kref_get(&pg->kref);
933                         start_queue = 1;
934                 }
935         }
936
937         spin_unlock_irqrestore(&pg->lock, flags);
938
939         if (start_queue) {
940                 if (queue_delayed_work(kaluad_wq, &pg->rtpg_work,
941                                 msecs_to_jiffies(ALUA_RTPG_DELAY_MSECS)))
942                         sdev = NULL;
943                 else
944                         kref_put(&pg->kref, release_port_group);
945         }
946         if (sdev)
947                 scsi_device_put(sdev);
948
949         return true;
950 }
951
952 /*
953  * alua_initialize - Initialize ALUA state
954  * @sdev: the device to be initialized
955  *
956  * For the prep_fn to work correctly we have
957  * to initialize the ALUA state for the device.
958  */
959 static int alua_initialize(struct scsi_device *sdev, struct alua_dh_data *h)
960 {
961         int err = SCSI_DH_DEV_UNSUPP, tpgs;
962
963         mutex_lock(&h->init_mutex);
964         tpgs = alua_check_tpgs(sdev);
965         if (tpgs != TPGS_MODE_NONE)
966                 err = alua_check_vpd(sdev, h, tpgs);
967         h->init_error = err;
968         mutex_unlock(&h->init_mutex);
969         return err;
970 }
971 /*
972  * alua_set_params - set/unset the optimize flag
973  * @sdev: device on the path to be activated
974  * params - parameters in the following format
975  *      "no_of_params\0param1\0param2\0param3\0...\0"
976  * For example, to set the flag pass the following parameters
977  * from multipath.conf
978  *     hardware_handler        "2 alua 1"
979  */
980 static int alua_set_params(struct scsi_device *sdev, const char *params)
981 {
982         struct alua_dh_data *h = sdev->handler_data;
983         struct alua_port_group *pg = NULL;
984         unsigned int optimize = 0, argc;
985         const char *p = params;
986         int result = SCSI_DH_OK;
987         unsigned long flags;
988
989         if ((sscanf(params, "%u", &argc) != 1) || (argc != 1))
990                 return -EINVAL;
991
992         while (*p++)
993                 ;
994         if ((sscanf(p, "%u", &optimize) != 1) || (optimize > 1))
995                 return -EINVAL;
996
997         rcu_read_lock();
998         pg = rcu_dereference(h->pg);
999         if (!pg) {
1000                 rcu_read_unlock();
1001                 return -ENXIO;
1002         }
1003         spin_lock_irqsave(&pg->lock, flags);
1004         if (optimize)
1005                 pg->flags |= ALUA_OPTIMIZE_STPG;
1006         else
1007                 pg->flags &= ~ALUA_OPTIMIZE_STPG;
1008         spin_unlock_irqrestore(&pg->lock, flags);
1009         rcu_read_unlock();
1010
1011         return result;
1012 }
1013
1014 /*
1015  * alua_activate - activate a path
1016  * @sdev: device on the path to be activated
1017  *
1018  * We're currently switching the port group to be activated only and
1019  * let the array figure out the rest.
1020  * There may be other arrays which require us to switch all port groups
1021  * based on a certain policy. But until we actually encounter them it
1022  * should be okay.
1023  */
1024 static int alua_activate(struct scsi_device *sdev,
1025                         activate_complete fn, void *data)
1026 {
1027         struct alua_dh_data *h = sdev->handler_data;
1028         int err = SCSI_DH_OK;
1029         struct alua_queue_data *qdata;
1030         struct alua_port_group *pg;
1031
1032         qdata = kzalloc(sizeof(*qdata), GFP_KERNEL);
1033         if (!qdata) {
1034                 err = SCSI_DH_RES_TEMP_UNAVAIL;
1035                 goto out;
1036         }
1037         qdata->callback_fn = fn;
1038         qdata->callback_data = data;
1039
1040         mutex_lock(&h->init_mutex);
1041         rcu_read_lock();
1042         pg = rcu_dereference(h->pg);
1043         if (!pg || !kref_get_unless_zero(&pg->kref)) {
1044                 rcu_read_unlock();
1045                 kfree(qdata);
1046                 err = h->init_error;
1047                 mutex_unlock(&h->init_mutex);
1048                 goto out;
1049         }
1050         rcu_read_unlock();
1051         mutex_unlock(&h->init_mutex);
1052
1053         if (alua_rtpg_queue(pg, sdev, qdata, true)) {
1054                 fn = NULL;
1055         } else {
1056                 kfree(qdata);
1057                 err = SCSI_DH_DEV_OFFLINED;
1058         }
1059         kref_put(&pg->kref, release_port_group);
1060 out:
1061         if (fn)
1062                 fn(data, err);
1063         return 0;
1064 }
1065
1066 /*
1067  * alua_check - check path status
1068  * @sdev: device on the path to be checked
1069  *
1070  * Check the device status
1071  */
1072 static void alua_check(struct scsi_device *sdev, bool force)
1073 {
1074         struct alua_dh_data *h = sdev->handler_data;
1075         struct alua_port_group *pg;
1076
1077         rcu_read_lock();
1078         pg = rcu_dereference(h->pg);
1079         if (!pg || !kref_get_unless_zero(&pg->kref)) {
1080                 rcu_read_unlock();
1081                 return;
1082         }
1083         rcu_read_unlock();
1084
1085         alua_rtpg_queue(pg, sdev, NULL, force);
1086         kref_put(&pg->kref, release_port_group);
1087 }
1088
1089 /*
1090  * alua_prep_fn - request callback
1091  *
1092  * Fail I/O to all paths not in state
1093  * active/optimized or active/non-optimized.
1094  */
1095 static int alua_prep_fn(struct scsi_device *sdev, struct request *req)
1096 {
1097         struct alua_dh_data *h = sdev->handler_data;
1098         struct alua_port_group *pg;
1099         unsigned char state = SCSI_ACCESS_STATE_OPTIMAL;
1100         int ret = BLKPREP_OK;
1101
1102         rcu_read_lock();
1103         pg = rcu_dereference(h->pg);
1104         if (pg)
1105                 state = pg->state;
1106         rcu_read_unlock();
1107         if (state == SCSI_ACCESS_STATE_TRANSITIONING)
1108                 ret = BLKPREP_DEFER;
1109         else if (state != SCSI_ACCESS_STATE_OPTIMAL &&
1110                  state != SCSI_ACCESS_STATE_ACTIVE &&
1111                  state != SCSI_ACCESS_STATE_LBA) {
1112                 ret = BLKPREP_KILL;
1113                 req->rq_flags |= RQF_QUIET;
1114         }
1115         return ret;
1116
1117 }
1118
1119 static void alua_rescan(struct scsi_device *sdev)
1120 {
1121         struct alua_dh_data *h = sdev->handler_data;
1122
1123         alua_initialize(sdev, h);
1124 }
1125
1126 /*
1127  * alua_bus_attach - Attach device handler
1128  * @sdev: device to be attached to
1129  */
1130 static int alua_bus_attach(struct scsi_device *sdev)
1131 {
1132         struct alua_dh_data *h;
1133         int err;
1134
1135         h = kzalloc(sizeof(*h) , GFP_KERNEL);
1136         if (!h)
1137                 return SCSI_DH_NOMEM;
1138         spin_lock_init(&h->pg_lock);
1139         rcu_assign_pointer(h->pg, NULL);
1140         h->init_error = SCSI_DH_OK;
1141         h->sdev = sdev;
1142         INIT_LIST_HEAD(&h->node);
1143
1144         mutex_init(&h->init_mutex);
1145         err = alua_initialize(sdev, h);
1146         if (err != SCSI_DH_OK && err != SCSI_DH_DEV_OFFLINED)
1147                 goto failed;
1148
1149         sdev->handler_data = h;
1150         return SCSI_DH_OK;
1151 failed:
1152         kfree(h);
1153         return err;
1154 }
1155
1156 /*
1157  * alua_bus_detach - Detach device handler
1158  * @sdev: device to be detached from
1159  */
1160 static void alua_bus_detach(struct scsi_device *sdev)
1161 {
1162         struct alua_dh_data *h = sdev->handler_data;
1163         struct alua_port_group *pg;
1164
1165         spin_lock(&h->pg_lock);
1166         pg = rcu_dereference_protected(h->pg, lockdep_is_held(&h->pg_lock));
1167         rcu_assign_pointer(h->pg, NULL);
1168         spin_unlock(&h->pg_lock);
1169         if (pg) {
1170                 spin_lock_irq(&pg->lock);
1171                 list_del_rcu(&h->node);
1172                 spin_unlock_irq(&pg->lock);
1173                 kref_put(&pg->kref, release_port_group);
1174         }
1175         sdev->handler_data = NULL;
1176         synchronize_rcu();
1177         kfree(h);
1178 }
1179
1180 static struct scsi_device_handler alua_dh = {
1181         .name = ALUA_DH_NAME,
1182         .module = THIS_MODULE,
1183         .attach = alua_bus_attach,
1184         .detach = alua_bus_detach,
1185         .prep_fn = alua_prep_fn,
1186         .check_sense = alua_check_sense,
1187         .activate = alua_activate,
1188         .rescan = alua_rescan,
1189         .set_params = alua_set_params,
1190 };
1191
1192 static int __init alua_init(void)
1193 {
1194         int r;
1195
1196         kaluad_wq = alloc_workqueue("kaluad", WQ_MEM_RECLAIM, 0);
1197         if (!kaluad_wq)
1198                 return -ENOMEM;
1199
1200         r = scsi_register_device_handler(&alua_dh);
1201         if (r != 0) {
1202                 printk(KERN_ERR "%s: Failed to register scsi device handler",
1203                         ALUA_DH_NAME);
1204                 destroy_workqueue(kaluad_wq);
1205         }
1206         return r;
1207 }
1208
1209 static void __exit alua_exit(void)
1210 {
1211         scsi_unregister_device_handler(&alua_dh);
1212         destroy_workqueue(kaluad_wq);
1213 }
1214
1215 module_init(alua_init);
1216 module_exit(alua_exit);
1217
1218 MODULE_DESCRIPTION("DM Multipath ALUA support");
1219 MODULE_AUTHOR("Hannes Reinecke <hare@suse.de>");
1220 MODULE_LICENSE("GPL");
1221 MODULE_VERSION(ALUA_DH_VER);