]> git.karo-electronics.de Git - karo-tx-linux.git/blob - drivers/scsi/device_handler/scsi_dh_alua.c
Merge tag 'perf-urgent-for-mingo-4.11-20170317' of git://git.kernel.org/pub/scm/linux...
[karo-tx-linux.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
57 /* device handler flags */
58 #define ALUA_OPTIMIZE_STPG              0x01
59 #define ALUA_RTPG_EXT_HDR_UNSUPP        0x02
60 #define ALUA_SYNC_STPG                  0x04
61 /* State machine flags */
62 #define ALUA_PG_RUN_RTPG                0x10
63 #define ALUA_PG_RUN_STPG                0x20
64 #define ALUA_PG_RUNNING                 0x40
65
66 static uint optimize_stpg;
67 module_param(optimize_stpg, uint, S_IRUGO|S_IWUSR);
68 MODULE_PARM_DESC(optimize_stpg, "Allow use of a non-optimized path, rather than sending a STPG, when implicit TPGS is supported (0=No,1=Yes). Default is 0.");
69
70 static LIST_HEAD(port_group_list);
71 static DEFINE_SPINLOCK(port_group_lock);
72 static struct workqueue_struct *kaluad_wq;
73 static struct workqueue_struct *kaluad_sync_wq;
74
75 struct alua_port_group {
76         struct kref             kref;
77         struct rcu_head         rcu;
78         struct list_head        node;
79         struct list_head        dh_list;
80         unsigned char           device_id_str[256];
81         int                     device_id_len;
82         int                     group_id;
83         int                     tpgs;
84         int                     state;
85         int                     pref;
86         unsigned                flags; /* used for optimizing STPG */
87         unsigned char           transition_tmo;
88         unsigned long           expiry;
89         unsigned long           interval;
90         struct delayed_work     rtpg_work;
91         spinlock_t              lock;
92         struct list_head        rtpg_list;
93         struct scsi_device      *rtpg_sdev;
94 };
95
96 struct alua_dh_data {
97         struct list_head        node;
98         struct alua_port_group __rcu *pg;
99         int                     group_id;
100         spinlock_t              pg_lock;
101         struct scsi_device      *sdev;
102         int                     init_error;
103         struct mutex            init_mutex;
104 };
105
106 struct alua_queue_data {
107         struct list_head        entry;
108         activate_complete       callback_fn;
109         void                    *callback_data;
110 };
111
112 #define ALUA_POLICY_SWITCH_CURRENT      0
113 #define ALUA_POLICY_SWITCH_ALL          1
114
115 static void alua_rtpg_work(struct work_struct *work);
116 static void alua_rtpg_queue(struct alua_port_group *pg,
117                             struct scsi_device *sdev,
118                             struct alua_queue_data *qdata, bool force);
119 static void alua_check(struct scsi_device *sdev, bool force);
120
121 static void release_port_group(struct kref *kref)
122 {
123         struct alua_port_group *pg;
124
125         pg = container_of(kref, struct alua_port_group, kref);
126         if (pg->rtpg_sdev)
127                 flush_delayed_work(&pg->rtpg_work);
128         spin_lock(&port_group_lock);
129         list_del(&pg->node);
130         spin_unlock(&port_group_lock);
131         kfree_rcu(pg, rcu);
132 }
133
134 /*
135  * submit_rtpg - Issue a REPORT TARGET GROUP STATES command
136  * @sdev: sdev the command should be sent to
137  */
138 static int submit_rtpg(struct scsi_device *sdev, unsigned char *buff,
139                        int bufflen, struct scsi_sense_hdr *sshdr, int flags)
140 {
141         u8 cdb[COMMAND_SIZE(MAINTENANCE_IN)];
142         int req_flags = REQ_FAILFAST_DEV | REQ_FAILFAST_TRANSPORT |
143                 REQ_FAILFAST_DRIVER;
144
145         /* Prepare the command. */
146         memset(cdb, 0x0, COMMAND_SIZE(MAINTENANCE_IN));
147         cdb[0] = MAINTENANCE_IN;
148         if (!(flags & ALUA_RTPG_EXT_HDR_UNSUPP))
149                 cdb[1] = MI_REPORT_TARGET_PGS | MI_EXT_HDR_PARAM_FMT;
150         else
151                 cdb[1] = MI_REPORT_TARGET_PGS;
152         put_unaligned_be32(bufflen, &cdb[6]);
153
154         return scsi_execute_req_flags(sdev, cdb, DMA_FROM_DEVICE,
155                                       buff, bufflen, sshdr,
156                                       ALUA_FAILOVER_TIMEOUT * HZ,
157                                       ALUA_FAILOVER_RETRIES, NULL,
158                                       req_flags, 0);
159 }
160
161 /*
162  * submit_stpg - Issue a SET TARGET PORT GROUP command
163  *
164  * Currently we're only setting the current target port group state
165  * to 'active/optimized' and let the array firmware figure out
166  * the states of the remaining groups.
167  */
168 static int submit_stpg(struct scsi_device *sdev, int group_id,
169                        struct scsi_sense_hdr *sshdr)
170 {
171         u8 cdb[COMMAND_SIZE(MAINTENANCE_OUT)];
172         unsigned char stpg_data[8];
173         int stpg_len = 8;
174         int req_flags = REQ_FAILFAST_DEV | REQ_FAILFAST_TRANSPORT |
175                 REQ_FAILFAST_DRIVER;
176
177         /* Prepare the data buffer */
178         memset(stpg_data, 0, stpg_len);
179         stpg_data[4] = SCSI_ACCESS_STATE_OPTIMAL;
180         put_unaligned_be16(group_id, &stpg_data[6]);
181
182         /* Prepare the command. */
183         memset(cdb, 0x0, COMMAND_SIZE(MAINTENANCE_OUT));
184         cdb[0] = MAINTENANCE_OUT;
185         cdb[1] = MO_SET_TARGET_PGS;
186         put_unaligned_be32(stpg_len, &cdb[6]);
187
188         return scsi_execute_req_flags(sdev, cdb, DMA_TO_DEVICE,
189                                       stpg_data, stpg_len,
190                                       sshdr, ALUA_FAILOVER_TIMEOUT * HZ,
191                                       ALUA_FAILOVER_RETRIES, NULL,
192                                       req_flags, 0);
193 }
194
195 static struct alua_port_group *alua_find_get_pg(char *id_str, size_t id_size,
196                                                 int group_id)
197 {
198         struct alua_port_group *pg;
199
200         if (!id_str || !id_size || !strlen(id_str))
201                 return NULL;
202
203         list_for_each_entry(pg, &port_group_list, node) {
204                 if (pg->group_id != group_id)
205                         continue;
206                 if (!pg->device_id_len || pg->device_id_len != id_size)
207                         continue;
208                 if (strncmp(pg->device_id_str, id_str, id_size))
209                         continue;
210                 if (!kref_get_unless_zero(&pg->kref))
211                         continue;
212                 return pg;
213         }
214
215         return NULL;
216 }
217
218 /*
219  * alua_alloc_pg - Allocate a new port_group structure
220  * @sdev: scsi device
221  * @h: alua device_handler data
222  * @group_id: port group id
223  *
224  * Allocate a new port_group structure for a given
225  * device.
226  */
227 static struct alua_port_group *alua_alloc_pg(struct scsi_device *sdev,
228                                              int group_id, int tpgs)
229 {
230         struct alua_port_group *pg, *tmp_pg;
231
232         pg = kzalloc(sizeof(struct alua_port_group), GFP_KERNEL);
233         if (!pg)
234                 return ERR_PTR(-ENOMEM);
235
236         pg->device_id_len = scsi_vpd_lun_id(sdev, pg->device_id_str,
237                                             sizeof(pg->device_id_str));
238         if (pg->device_id_len <= 0) {
239                 /*
240                  * TPGS supported but no device identification found.
241                  * Generate private device identification.
242                  */
243                 sdev_printk(KERN_INFO, sdev,
244                             "%s: No device descriptors found\n",
245                             ALUA_DH_NAME);
246                 pg->device_id_str[0] = '\0';
247                 pg->device_id_len = 0;
248         }
249         pg->group_id = group_id;
250         pg->tpgs = tpgs;
251         pg->state = SCSI_ACCESS_STATE_OPTIMAL;
252         if (optimize_stpg)
253                 pg->flags |= ALUA_OPTIMIZE_STPG;
254         kref_init(&pg->kref);
255         INIT_DELAYED_WORK(&pg->rtpg_work, alua_rtpg_work);
256         INIT_LIST_HEAD(&pg->rtpg_list);
257         INIT_LIST_HEAD(&pg->node);
258         INIT_LIST_HEAD(&pg->dh_list);
259         spin_lock_init(&pg->lock);
260
261         spin_lock(&port_group_lock);
262         tmp_pg = alua_find_get_pg(pg->device_id_str, pg->device_id_len,
263                                   group_id);
264         if (tmp_pg) {
265                 spin_unlock(&port_group_lock);
266                 kfree(pg);
267                 return tmp_pg;
268         }
269
270         list_add(&pg->node, &port_group_list);
271         spin_unlock(&port_group_lock);
272
273         return pg;
274 }
275
276 /*
277  * alua_check_tpgs - Evaluate TPGS setting
278  * @sdev: device to be checked
279  *
280  * Examine the TPGS setting of the sdev to find out if ALUA
281  * is supported.
282  */
283 static int alua_check_tpgs(struct scsi_device *sdev)
284 {
285         int tpgs = TPGS_MODE_NONE;
286
287         /*
288          * ALUA support for non-disk devices is fraught with
289          * difficulties, so disable it for now.
290          */
291         if (sdev->type != TYPE_DISK) {
292                 sdev_printk(KERN_INFO, sdev,
293                             "%s: disable for non-disk devices\n",
294                             ALUA_DH_NAME);
295                 return tpgs;
296         }
297
298         tpgs = scsi_device_tpgs(sdev);
299         switch (tpgs) {
300         case TPGS_MODE_EXPLICIT|TPGS_MODE_IMPLICIT:
301                 sdev_printk(KERN_INFO, sdev,
302                             "%s: supports implicit and explicit TPGS\n",
303                             ALUA_DH_NAME);
304                 break;
305         case TPGS_MODE_EXPLICIT:
306                 sdev_printk(KERN_INFO, sdev, "%s: supports explicit TPGS\n",
307                             ALUA_DH_NAME);
308                 break;
309         case TPGS_MODE_IMPLICIT:
310                 sdev_printk(KERN_INFO, sdev, "%s: supports implicit TPGS\n",
311                             ALUA_DH_NAME);
312                 break;
313         case TPGS_MODE_NONE:
314                 sdev_printk(KERN_INFO, sdev, "%s: not supported\n",
315                             ALUA_DH_NAME);
316                 break;
317         default:
318                 sdev_printk(KERN_INFO, sdev,
319                             "%s: unsupported TPGS setting %d\n",
320                             ALUA_DH_NAME, tpgs);
321                 tpgs = TPGS_MODE_NONE;
322                 break;
323         }
324
325         return tpgs;
326 }
327
328 /*
329  * alua_check_vpd - Evaluate INQUIRY vpd page 0x83
330  * @sdev: device to be checked
331  *
332  * Extract the relative target port and the target port group
333  * descriptor from the list of identificators.
334  */
335 static int alua_check_vpd(struct scsi_device *sdev, struct alua_dh_data *h,
336                           int tpgs)
337 {
338         int rel_port = -1, group_id;
339         struct alua_port_group *pg, *old_pg = NULL;
340         bool pg_updated = false;
341         unsigned long flags;
342
343         group_id = scsi_vpd_tpg_id(sdev, &rel_port);
344         if (group_id < 0) {
345                 /*
346                  * Internal error; TPGS supported but required
347                  * VPD identification descriptors not present.
348                  * Disable ALUA support
349                  */
350                 sdev_printk(KERN_INFO, sdev,
351                             "%s: No target port descriptors found\n",
352                             ALUA_DH_NAME);
353                 return SCSI_DH_DEV_UNSUPP;
354         }
355
356         pg = alua_alloc_pg(sdev, group_id, tpgs);
357         if (IS_ERR(pg)) {
358                 if (PTR_ERR(pg) == -ENOMEM)
359                         return SCSI_DH_NOMEM;
360                 return SCSI_DH_DEV_UNSUPP;
361         }
362         if (pg->device_id_len)
363                 sdev_printk(KERN_INFO, sdev,
364                             "%s: device %s port group %x rel port %x\n",
365                             ALUA_DH_NAME, pg->device_id_str,
366                             group_id, rel_port);
367         else
368                 sdev_printk(KERN_INFO, sdev,
369                             "%s: port group %x rel port %x\n",
370                             ALUA_DH_NAME, group_id, rel_port);
371
372         /* Check for existing port group references */
373         spin_lock(&h->pg_lock);
374         old_pg = rcu_dereference_protected(h->pg, lockdep_is_held(&h->pg_lock));
375         if (old_pg != pg) {
376                 /* port group has changed. Update to new port group */
377                 if (h->pg) {
378                         spin_lock_irqsave(&old_pg->lock, flags);
379                         list_del_rcu(&h->node);
380                         spin_unlock_irqrestore(&old_pg->lock, flags);
381                 }
382                 rcu_assign_pointer(h->pg, pg);
383                 pg_updated = true;
384         }
385
386         spin_lock_irqsave(&pg->lock, flags);
387         if (sdev->synchronous_alua)
388                 pg->flags |= ALUA_SYNC_STPG;
389         if (pg_updated)
390                 list_add_rcu(&h->node, &pg->dh_list);
391         spin_unlock_irqrestore(&pg->lock, flags);
392
393         alua_rtpg_queue(rcu_dereference_protected(h->pg,
394                                                   lockdep_is_held(&h->pg_lock)),
395                         sdev, NULL, true);
396         spin_unlock(&h->pg_lock);
397
398         if (old_pg)
399                 kref_put(&old_pg->kref, release_port_group);
400
401         return SCSI_DH_OK;
402 }
403
404 static char print_alua_state(unsigned char state)
405 {
406         switch (state) {
407         case SCSI_ACCESS_STATE_OPTIMAL:
408                 return 'A';
409         case SCSI_ACCESS_STATE_ACTIVE:
410                 return 'N';
411         case SCSI_ACCESS_STATE_STANDBY:
412                 return 'S';
413         case SCSI_ACCESS_STATE_UNAVAILABLE:
414                 return 'U';
415         case SCSI_ACCESS_STATE_LBA:
416                 return 'L';
417         case SCSI_ACCESS_STATE_OFFLINE:
418                 return 'O';
419         case SCSI_ACCESS_STATE_TRANSITIONING:
420                 return 'T';
421         default:
422                 return 'X';
423         }
424 }
425
426 static int alua_check_sense(struct scsi_device *sdev,
427                             struct scsi_sense_hdr *sense_hdr)
428 {
429         switch (sense_hdr->sense_key) {
430         case NOT_READY:
431                 if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0a) {
432                         /*
433                          * LUN Not Accessible - ALUA state transition
434                          */
435                         alua_check(sdev, false);
436                         return NEEDS_RETRY;
437                 }
438                 break;
439         case UNIT_ATTENTION:
440                 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x00) {
441                         /*
442                          * Power On, Reset, or Bus Device Reset.
443                          * Might have obscured a state transition,
444                          * so schedule a recheck.
445                          */
446                         alua_check(sdev, true);
447                         return ADD_TO_MLQUEUE;
448                 }
449                 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x04)
450                         /*
451                          * Device internal reset
452                          */
453                         return ADD_TO_MLQUEUE;
454                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x01)
455                         /*
456                          * Mode Parameters Changed
457                          */
458                         return ADD_TO_MLQUEUE;
459                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x06) {
460                         /*
461                          * ALUA state changed
462                          */
463                         alua_check(sdev, true);
464                         return ADD_TO_MLQUEUE;
465                 }
466                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x07) {
467                         /*
468                          * Implicit ALUA state transition failed
469                          */
470                         alua_check(sdev, true);
471                         return ADD_TO_MLQUEUE;
472                 }
473                 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x03)
474                         /*
475                          * Inquiry data has changed
476                          */
477                         return ADD_TO_MLQUEUE;
478                 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x0e)
479                         /*
480                          * REPORTED_LUNS_DATA_HAS_CHANGED is reported
481                          * when switching controllers on targets like
482                          * Intel Multi-Flex. We can just retry.
483                          */
484                         return ADD_TO_MLQUEUE;
485                 break;
486         }
487
488         return SCSI_RETURN_NOT_HANDLED;
489 }
490
491 /*
492  * alua_tur - Send a TEST UNIT READY
493  * @sdev: device to which the TEST UNIT READY command should be send
494  *
495  * Send a TEST UNIT READY to @sdev to figure out the device state
496  * Returns SCSI_DH_RETRY if the sense code is NOT READY/ALUA TRANSITIONING,
497  * SCSI_DH_OK if no error occurred, and SCSI_DH_IO otherwise.
498  */
499 static int alua_tur(struct scsi_device *sdev)
500 {
501         struct scsi_sense_hdr sense_hdr;
502         int retval;
503
504         retval = scsi_test_unit_ready(sdev, ALUA_FAILOVER_TIMEOUT * HZ,
505                                       ALUA_FAILOVER_RETRIES, &sense_hdr);
506         if (sense_hdr.sense_key == NOT_READY &&
507             sense_hdr.asc == 0x04 && sense_hdr.ascq == 0x0a)
508                 return SCSI_DH_RETRY;
509         else if (retval)
510                 return SCSI_DH_IO;
511         else
512                 return SCSI_DH_OK;
513 }
514
515 /*
516  * alua_rtpg - Evaluate REPORT TARGET GROUP STATES
517  * @sdev: the device to be evaluated.
518  *
519  * Evaluate the Target Port Group State.
520  * Returns SCSI_DH_DEV_OFFLINED if the path is
521  * found to be unusable.
522  */
523 static int alua_rtpg(struct scsi_device *sdev, struct alua_port_group *pg)
524 {
525         struct scsi_sense_hdr sense_hdr;
526         struct alua_port_group *tmp_pg;
527         int len, k, off, valid_states = 0, bufflen = ALUA_RTPG_SIZE;
528         unsigned char *desc, *buff;
529         unsigned err, retval;
530         unsigned int tpg_desc_tbl_off;
531         unsigned char orig_transition_tmo;
532         unsigned long flags;
533
534         if (!pg->expiry) {
535                 unsigned long transition_tmo = ALUA_FAILOVER_TIMEOUT * HZ;
536
537                 if (pg->transition_tmo)
538                         transition_tmo = pg->transition_tmo * HZ;
539
540                 pg->expiry = round_jiffies_up(jiffies + transition_tmo);
541         }
542
543         buff = kzalloc(bufflen, GFP_KERNEL);
544         if (!buff)
545                 return SCSI_DH_DEV_TEMP_BUSY;
546
547  retry:
548         err = 0;
549         retval = submit_rtpg(sdev, buff, bufflen, &sense_hdr, pg->flags);
550
551         if (retval) {
552                 if (!scsi_sense_valid(&sense_hdr)) {
553                         sdev_printk(KERN_INFO, sdev,
554                                     "%s: rtpg failed, result %d\n",
555                                     ALUA_DH_NAME, retval);
556                         kfree(buff);
557                         if (driver_byte(retval) == DRIVER_ERROR)
558                                 return SCSI_DH_DEV_TEMP_BUSY;
559                         return SCSI_DH_IO;
560                 }
561
562                 /*
563                  * submit_rtpg() has failed on existing arrays
564                  * when requesting extended header info, and
565                  * the array doesn't support extended headers,
566                  * even though it shouldn't according to T10.
567                  * The retry without rtpg_ext_hdr_req set
568                  * handles this.
569                  */
570                 if (!(pg->flags & ALUA_RTPG_EXT_HDR_UNSUPP) &&
571                     sense_hdr.sense_key == ILLEGAL_REQUEST &&
572                     sense_hdr.asc == 0x24 && sense_hdr.ascq == 0) {
573                         pg->flags |= ALUA_RTPG_EXT_HDR_UNSUPP;
574                         goto retry;
575                 }
576                 /*
577                  * Retry on ALUA state transition or if any
578                  * UNIT ATTENTION occurred.
579                  */
580                 if (sense_hdr.sense_key == NOT_READY &&
581                     sense_hdr.asc == 0x04 && sense_hdr.ascq == 0x0a)
582                         err = SCSI_DH_RETRY;
583                 else if (sense_hdr.sense_key == UNIT_ATTENTION)
584                         err = SCSI_DH_RETRY;
585                 if (err == SCSI_DH_RETRY &&
586                     pg->expiry != 0 && time_before(jiffies, pg->expiry)) {
587                         sdev_printk(KERN_ERR, sdev, "%s: rtpg retry\n",
588                                     ALUA_DH_NAME);
589                         scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
590                         kfree(buff);
591                         return err;
592                 }
593                 sdev_printk(KERN_ERR, sdev, "%s: rtpg failed\n",
594                             ALUA_DH_NAME);
595                 scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
596                 kfree(buff);
597                 pg->expiry = 0;
598                 return SCSI_DH_IO;
599         }
600
601         len = get_unaligned_be32(&buff[0]) + 4;
602
603         if (len > bufflen) {
604                 /* Resubmit with the correct length */
605                 kfree(buff);
606                 bufflen = len;
607                 buff = kmalloc(bufflen, GFP_KERNEL);
608                 if (!buff) {
609                         sdev_printk(KERN_WARNING, sdev,
610                                     "%s: kmalloc buffer failed\n",__func__);
611                         /* Temporary failure, bypass */
612                         pg->expiry = 0;
613                         return SCSI_DH_DEV_TEMP_BUSY;
614                 }
615                 goto retry;
616         }
617
618         orig_transition_tmo = pg->transition_tmo;
619         if ((buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR && buff[5] != 0)
620                 pg->transition_tmo = buff[5];
621         else
622                 pg->transition_tmo = ALUA_FAILOVER_TIMEOUT;
623
624         if (orig_transition_tmo != pg->transition_tmo) {
625                 sdev_printk(KERN_INFO, sdev,
626                             "%s: transition timeout set to %d seconds\n",
627                             ALUA_DH_NAME, pg->transition_tmo);
628                 pg->expiry = jiffies + pg->transition_tmo * HZ;
629         }
630
631         if ((buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR)
632                 tpg_desc_tbl_off = 8;
633         else
634                 tpg_desc_tbl_off = 4;
635
636         for (k = tpg_desc_tbl_off, desc = buff + tpg_desc_tbl_off;
637              k < len;
638              k += off, desc += off) {
639                 u16 group_id = get_unaligned_be16(&desc[2]);
640
641                 spin_lock_irqsave(&port_group_lock, flags);
642                 tmp_pg = alua_find_get_pg(pg->device_id_str, pg->device_id_len,
643                                           group_id);
644                 spin_unlock_irqrestore(&port_group_lock, flags);
645                 if (tmp_pg) {
646                         if (spin_trylock_irqsave(&tmp_pg->lock, flags)) {
647                                 if ((tmp_pg == pg) ||
648                                     !(tmp_pg->flags & ALUA_PG_RUNNING)) {
649                                         struct alua_dh_data *h;
650
651                                         tmp_pg->state = desc[0] & 0x0f;
652                                         tmp_pg->pref = desc[0] >> 7;
653                                         rcu_read_lock();
654                                         list_for_each_entry_rcu(h,
655                                                 &tmp_pg->dh_list, node) {
656                                                 /* h->sdev should always be valid */
657                                                 BUG_ON(!h->sdev);
658                                                 h->sdev->access_state = desc[0];
659                                         }
660                                         rcu_read_unlock();
661                                 }
662                                 if (tmp_pg == pg)
663                                         valid_states = desc[1];
664                                 spin_unlock_irqrestore(&tmp_pg->lock, flags);
665                         }
666                         kref_put(&tmp_pg->kref, release_port_group);
667                 }
668                 off = 8 + (desc[7] * 4);
669         }
670
671         spin_lock_irqsave(&pg->lock, flags);
672         sdev_printk(KERN_INFO, sdev,
673                     "%s: port group %02x state %c %s supports %c%c%c%c%c%c%c\n",
674                     ALUA_DH_NAME, pg->group_id, print_alua_state(pg->state),
675                     pg->pref ? "preferred" : "non-preferred",
676                     valid_states&TPGS_SUPPORT_TRANSITION?'T':'t',
677                     valid_states&TPGS_SUPPORT_OFFLINE?'O':'o',
678                     valid_states&TPGS_SUPPORT_LBA_DEPENDENT?'L':'l',
679                     valid_states&TPGS_SUPPORT_UNAVAILABLE?'U':'u',
680                     valid_states&TPGS_SUPPORT_STANDBY?'S':'s',
681                     valid_states&TPGS_SUPPORT_NONOPTIMIZED?'N':'n',
682                     valid_states&TPGS_SUPPORT_OPTIMIZED?'A':'a');
683
684         switch (pg->state) {
685         case SCSI_ACCESS_STATE_TRANSITIONING:
686                 if (time_before(jiffies, pg->expiry)) {
687                         /* State transition, retry */
688                         pg->interval = 2;
689                         err = SCSI_DH_RETRY;
690                 } else {
691                         struct alua_dh_data *h;
692
693                         /* Transitioning time exceeded, set port to standby */
694                         err = SCSI_DH_IO;
695                         pg->state = SCSI_ACCESS_STATE_STANDBY;
696                         pg->expiry = 0;
697                         rcu_read_lock();
698                         list_for_each_entry_rcu(h, &pg->dh_list, node) {
699                                 BUG_ON(!h->sdev);
700                                 h->sdev->access_state =
701                                         (pg->state & SCSI_ACCESS_STATE_MASK);
702                                 if (pg->pref)
703                                         h->sdev->access_state |=
704                                                 SCSI_ACCESS_STATE_PREFERRED;
705                         }
706                         rcu_read_unlock();
707                 }
708                 break;
709         case SCSI_ACCESS_STATE_OFFLINE:
710                 /* Path unusable */
711                 err = SCSI_DH_DEV_OFFLINED;
712                 pg->expiry = 0;
713                 break;
714         default:
715                 /* Useable path if active */
716                 err = SCSI_DH_OK;
717                 pg->expiry = 0;
718                 break;
719         }
720         spin_unlock_irqrestore(&pg->lock, flags);
721         kfree(buff);
722         return err;
723 }
724
725 /*
726  * alua_stpg - Issue a SET TARGET PORT GROUP command
727  *
728  * Issue a SET TARGET PORT GROUP command and evaluate the
729  * response. Returns SCSI_DH_RETRY per default to trigger
730  * a re-evaluation of the target group state or SCSI_DH_OK
731  * if no further action needs to be taken.
732  */
733 static unsigned alua_stpg(struct scsi_device *sdev, struct alua_port_group *pg)
734 {
735         int retval;
736         struct scsi_sense_hdr sense_hdr;
737
738         if (!(pg->tpgs & TPGS_MODE_EXPLICIT)) {
739                 /* Only implicit ALUA supported, retry */
740                 return SCSI_DH_RETRY;
741         }
742         switch (pg->state) {
743         case SCSI_ACCESS_STATE_OPTIMAL:
744                 return SCSI_DH_OK;
745         case SCSI_ACCESS_STATE_ACTIVE:
746                 if ((pg->flags & ALUA_OPTIMIZE_STPG) &&
747                     !pg->pref &&
748                     (pg->tpgs & TPGS_MODE_IMPLICIT))
749                         return SCSI_DH_OK;
750                 break;
751         case SCSI_ACCESS_STATE_STANDBY:
752         case SCSI_ACCESS_STATE_UNAVAILABLE:
753                 break;
754         case SCSI_ACCESS_STATE_OFFLINE:
755                 return SCSI_DH_IO;
756         case SCSI_ACCESS_STATE_TRANSITIONING:
757                 break;
758         default:
759                 sdev_printk(KERN_INFO, sdev,
760                             "%s: stpg failed, unhandled TPGS state %d",
761                             ALUA_DH_NAME, pg->state);
762                 return SCSI_DH_NOSYS;
763         }
764         retval = submit_stpg(sdev, pg->group_id, &sense_hdr);
765
766         if (retval) {
767                 if (!scsi_sense_valid(&sense_hdr)) {
768                         sdev_printk(KERN_INFO, sdev,
769                                     "%s: stpg failed, result %d",
770                                     ALUA_DH_NAME, retval);
771                         if (driver_byte(retval) == DRIVER_ERROR)
772                                 return SCSI_DH_DEV_TEMP_BUSY;
773                 } else {
774                         sdev_printk(KERN_INFO, sdev, "%s: stpg failed\n",
775                                     ALUA_DH_NAME);
776                         scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
777                 }
778         }
779         /* Retry RTPG */
780         return SCSI_DH_RETRY;
781 }
782
783 static void alua_rtpg_work(struct work_struct *work)
784 {
785         struct alua_port_group *pg =
786                 container_of(work, struct alua_port_group, rtpg_work.work);
787         struct scsi_device *sdev;
788         LIST_HEAD(qdata_list);
789         int err = SCSI_DH_OK;
790         struct alua_queue_data *qdata, *tmp;
791         unsigned long flags;
792         struct workqueue_struct *alua_wq = kaluad_wq;
793
794         spin_lock_irqsave(&pg->lock, flags);
795         sdev = pg->rtpg_sdev;
796         if (!sdev) {
797                 WARN_ON(pg->flags & ALUA_PG_RUN_RTPG);
798                 WARN_ON(pg->flags & ALUA_PG_RUN_STPG);
799                 spin_unlock_irqrestore(&pg->lock, flags);
800                 kref_put(&pg->kref, release_port_group);
801                 return;
802         }
803         if (pg->flags & ALUA_SYNC_STPG)
804                 alua_wq = kaluad_sync_wq;
805         pg->flags |= ALUA_PG_RUNNING;
806         if (pg->flags & ALUA_PG_RUN_RTPG) {
807                 int state = pg->state;
808
809                 pg->flags &= ~ALUA_PG_RUN_RTPG;
810                 spin_unlock_irqrestore(&pg->lock, flags);
811                 if (state == SCSI_ACCESS_STATE_TRANSITIONING) {
812                         if (alua_tur(sdev) == SCSI_DH_RETRY) {
813                                 spin_lock_irqsave(&pg->lock, flags);
814                                 pg->flags &= ~ALUA_PG_RUNNING;
815                                 pg->flags |= ALUA_PG_RUN_RTPG;
816                                 spin_unlock_irqrestore(&pg->lock, flags);
817                                 queue_delayed_work(alua_wq, &pg->rtpg_work,
818                                                    pg->interval * HZ);
819                                 return;
820                         }
821                         /* Send RTPG on failure or if TUR indicates SUCCESS */
822                 }
823                 err = alua_rtpg(sdev, pg);
824                 spin_lock_irqsave(&pg->lock, flags);
825                 if (err == SCSI_DH_RETRY || pg->flags & ALUA_PG_RUN_RTPG) {
826                         pg->flags &= ~ALUA_PG_RUNNING;
827                         pg->flags |= ALUA_PG_RUN_RTPG;
828                         spin_unlock_irqrestore(&pg->lock, flags);
829                         queue_delayed_work(alua_wq, &pg->rtpg_work,
830                                            pg->interval * HZ);
831                         return;
832                 }
833                 if (err != SCSI_DH_OK)
834                         pg->flags &= ~ALUA_PG_RUN_STPG;
835         }
836         if (pg->flags & ALUA_PG_RUN_STPG) {
837                 pg->flags &= ~ALUA_PG_RUN_STPG;
838                 spin_unlock_irqrestore(&pg->lock, flags);
839                 err = alua_stpg(sdev, pg);
840                 spin_lock_irqsave(&pg->lock, flags);
841                 if (err == SCSI_DH_RETRY || pg->flags & ALUA_PG_RUN_RTPG) {
842                         pg->flags |= ALUA_PG_RUN_RTPG;
843                         pg->interval = 0;
844                         pg->flags &= ~ALUA_PG_RUNNING;
845                         spin_unlock_irqrestore(&pg->lock, flags);
846                         queue_delayed_work(alua_wq, &pg->rtpg_work,
847                                            pg->interval * HZ);
848                         return;
849                 }
850         }
851
852         list_splice_init(&pg->rtpg_list, &qdata_list);
853         pg->rtpg_sdev = NULL;
854         spin_unlock_irqrestore(&pg->lock, flags);
855
856         list_for_each_entry_safe(qdata, tmp, &qdata_list, entry) {
857                 list_del(&qdata->entry);
858                 if (qdata->callback_fn)
859                         qdata->callback_fn(qdata->callback_data, err);
860                 kfree(qdata);
861         }
862         spin_lock_irqsave(&pg->lock, flags);
863         pg->flags &= ~ALUA_PG_RUNNING;
864         spin_unlock_irqrestore(&pg->lock, flags);
865         scsi_device_put(sdev);
866         kref_put(&pg->kref, release_port_group);
867 }
868
869 static void alua_rtpg_queue(struct alua_port_group *pg,
870                             struct scsi_device *sdev,
871                             struct alua_queue_data *qdata, bool force)
872 {
873         int start_queue = 0;
874         unsigned long flags;
875         struct workqueue_struct *alua_wq = kaluad_wq;
876
877         if (!pg)
878                 return;
879
880         spin_lock_irqsave(&pg->lock, flags);
881         if (qdata) {
882                 list_add_tail(&qdata->entry, &pg->rtpg_list);
883                 pg->flags |= ALUA_PG_RUN_STPG;
884                 force = true;
885         }
886         if (pg->rtpg_sdev == NULL) {
887                 pg->interval = 0;
888                 pg->flags |= ALUA_PG_RUN_RTPG;
889                 kref_get(&pg->kref);
890                 pg->rtpg_sdev = sdev;
891                 scsi_device_get(sdev);
892                 start_queue = 1;
893         } else if (!(pg->flags & ALUA_PG_RUN_RTPG) && force) {
894                 pg->flags |= ALUA_PG_RUN_RTPG;
895                 /* Do not queue if the worker is already running */
896                 if (!(pg->flags & ALUA_PG_RUNNING)) {
897                         kref_get(&pg->kref);
898                         sdev = NULL;
899                         start_queue = 1;
900                 }
901         }
902
903         if (pg->flags & ALUA_SYNC_STPG)
904                 alua_wq = kaluad_sync_wq;
905         spin_unlock_irqrestore(&pg->lock, flags);
906
907         if (start_queue &&
908             !queue_delayed_work(alua_wq, &pg->rtpg_work,
909                                 msecs_to_jiffies(ALUA_RTPG_DELAY_MSECS))) {
910                 if (sdev)
911                         scsi_device_put(sdev);
912                 kref_put(&pg->kref, release_port_group);
913         }
914 }
915
916 /*
917  * alua_initialize - Initialize ALUA state
918  * @sdev: the device to be initialized
919  *
920  * For the prep_fn to work correctly we have
921  * to initialize the ALUA state for the device.
922  */
923 static int alua_initialize(struct scsi_device *sdev, struct alua_dh_data *h)
924 {
925         int err = SCSI_DH_DEV_UNSUPP, tpgs;
926
927         mutex_lock(&h->init_mutex);
928         tpgs = alua_check_tpgs(sdev);
929         if (tpgs != TPGS_MODE_NONE)
930                 err = alua_check_vpd(sdev, h, tpgs);
931         h->init_error = err;
932         mutex_unlock(&h->init_mutex);
933         return err;
934 }
935 /*
936  * alua_set_params - set/unset the optimize flag
937  * @sdev: device on the path to be activated
938  * params - parameters in the following format
939  *      "no_of_params\0param1\0param2\0param3\0...\0"
940  * For example, to set the flag pass the following parameters
941  * from multipath.conf
942  *     hardware_handler        "2 alua 1"
943  */
944 static int alua_set_params(struct scsi_device *sdev, const char *params)
945 {
946         struct alua_dh_data *h = sdev->handler_data;
947         struct alua_port_group *pg = NULL;
948         unsigned int optimize = 0, argc;
949         const char *p = params;
950         int result = SCSI_DH_OK;
951         unsigned long flags;
952
953         if ((sscanf(params, "%u", &argc) != 1) || (argc != 1))
954                 return -EINVAL;
955
956         while (*p++)
957                 ;
958         if ((sscanf(p, "%u", &optimize) != 1) || (optimize > 1))
959                 return -EINVAL;
960
961         rcu_read_lock();
962         pg = rcu_dereference(h->pg);
963         if (!pg) {
964                 rcu_read_unlock();
965                 return -ENXIO;
966         }
967         spin_lock_irqsave(&pg->lock, flags);
968         if (optimize)
969                 pg->flags |= ALUA_OPTIMIZE_STPG;
970         else
971                 pg->flags &= ~ALUA_OPTIMIZE_STPG;
972         spin_unlock_irqrestore(&pg->lock, flags);
973         rcu_read_unlock();
974
975         return result;
976 }
977
978 /*
979  * alua_activate - activate a path
980  * @sdev: device on the path to be activated
981  *
982  * We're currently switching the port group to be activated only and
983  * let the array figure out the rest.
984  * There may be other arrays which require us to switch all port groups
985  * based on a certain policy. But until we actually encounter them it
986  * should be okay.
987  */
988 static int alua_activate(struct scsi_device *sdev,
989                         activate_complete fn, void *data)
990 {
991         struct alua_dh_data *h = sdev->handler_data;
992         int err = SCSI_DH_OK;
993         struct alua_queue_data *qdata;
994         struct alua_port_group *pg;
995
996         qdata = kzalloc(sizeof(*qdata), GFP_KERNEL);
997         if (!qdata) {
998                 err = SCSI_DH_RES_TEMP_UNAVAIL;
999                 goto out;
1000         }
1001         qdata->callback_fn = fn;
1002         qdata->callback_data = data;
1003
1004         mutex_lock(&h->init_mutex);
1005         rcu_read_lock();
1006         pg = rcu_dereference(h->pg);
1007         if (!pg || !kref_get_unless_zero(&pg->kref)) {
1008                 rcu_read_unlock();
1009                 kfree(qdata);
1010                 err = h->init_error;
1011                 mutex_unlock(&h->init_mutex);
1012                 goto out;
1013         }
1014         fn = NULL;
1015         rcu_read_unlock();
1016         mutex_unlock(&h->init_mutex);
1017
1018         alua_rtpg_queue(pg, sdev, qdata, true);
1019         kref_put(&pg->kref, release_port_group);
1020 out:
1021         if (fn)
1022                 fn(data, err);
1023         return 0;
1024 }
1025
1026 /*
1027  * alua_check - check path status
1028  * @sdev: device on the path to be checked
1029  *
1030  * Check the device status
1031  */
1032 static void alua_check(struct scsi_device *sdev, bool force)
1033 {
1034         struct alua_dh_data *h = sdev->handler_data;
1035         struct alua_port_group *pg;
1036
1037         rcu_read_lock();
1038         pg = rcu_dereference(h->pg);
1039         if (!pg || !kref_get_unless_zero(&pg->kref)) {
1040                 rcu_read_unlock();
1041                 return;
1042         }
1043         rcu_read_unlock();
1044
1045         alua_rtpg_queue(pg, sdev, NULL, force);
1046         kref_put(&pg->kref, release_port_group);
1047 }
1048
1049 /*
1050  * alua_prep_fn - request callback
1051  *
1052  * Fail I/O to all paths not in state
1053  * active/optimized or active/non-optimized.
1054  */
1055 static int alua_prep_fn(struct scsi_device *sdev, struct request *req)
1056 {
1057         struct alua_dh_data *h = sdev->handler_data;
1058         struct alua_port_group *pg;
1059         unsigned char state = SCSI_ACCESS_STATE_OPTIMAL;
1060         int ret = BLKPREP_OK;
1061
1062         rcu_read_lock();
1063         pg = rcu_dereference(h->pg);
1064         if (pg)
1065                 state = pg->state;
1066         rcu_read_unlock();
1067         if (state == SCSI_ACCESS_STATE_TRANSITIONING)
1068                 ret = BLKPREP_DEFER;
1069         else if (state != SCSI_ACCESS_STATE_OPTIMAL &&
1070                  state != SCSI_ACCESS_STATE_ACTIVE &&
1071                  state != SCSI_ACCESS_STATE_LBA) {
1072                 ret = BLKPREP_KILL;
1073                 req->rq_flags |= RQF_QUIET;
1074         }
1075         return ret;
1076
1077 }
1078
1079 static void alua_rescan(struct scsi_device *sdev)
1080 {
1081         struct alua_dh_data *h = sdev->handler_data;
1082
1083         alua_initialize(sdev, h);
1084 }
1085
1086 /*
1087  * alua_bus_attach - Attach device handler
1088  * @sdev: device to be attached to
1089  */
1090 static int alua_bus_attach(struct scsi_device *sdev)
1091 {
1092         struct alua_dh_data *h;
1093         int err, ret = -EINVAL;
1094
1095         h = kzalloc(sizeof(*h) , GFP_KERNEL);
1096         if (!h)
1097                 return -ENOMEM;
1098         spin_lock_init(&h->pg_lock);
1099         rcu_assign_pointer(h->pg, NULL);
1100         h->init_error = SCSI_DH_OK;
1101         h->sdev = sdev;
1102         INIT_LIST_HEAD(&h->node);
1103
1104         mutex_init(&h->init_mutex);
1105         err = alua_initialize(sdev, h);
1106         if (err == SCSI_DH_NOMEM)
1107                 ret = -ENOMEM;
1108         if (err != SCSI_DH_OK && err != SCSI_DH_DEV_OFFLINED)
1109                 goto failed;
1110
1111         sdev->handler_data = h;
1112         return 0;
1113 failed:
1114         kfree(h);
1115         return ret;
1116 }
1117
1118 /*
1119  * alua_bus_detach - Detach device handler
1120  * @sdev: device to be detached from
1121  */
1122 static void alua_bus_detach(struct scsi_device *sdev)
1123 {
1124         struct alua_dh_data *h = sdev->handler_data;
1125         struct alua_port_group *pg;
1126
1127         spin_lock(&h->pg_lock);
1128         pg = rcu_dereference_protected(h->pg, lockdep_is_held(&h->pg_lock));
1129         rcu_assign_pointer(h->pg, NULL);
1130         h->sdev = NULL;
1131         spin_unlock(&h->pg_lock);
1132         if (pg) {
1133                 spin_lock_irq(&pg->lock);
1134                 list_del_rcu(&h->node);
1135                 spin_unlock_irq(&pg->lock);
1136                 kref_put(&pg->kref, release_port_group);
1137         }
1138         sdev->handler_data = NULL;
1139         kfree(h);
1140 }
1141
1142 static struct scsi_device_handler alua_dh = {
1143         .name = ALUA_DH_NAME,
1144         .module = THIS_MODULE,
1145         .attach = alua_bus_attach,
1146         .detach = alua_bus_detach,
1147         .prep_fn = alua_prep_fn,
1148         .check_sense = alua_check_sense,
1149         .activate = alua_activate,
1150         .rescan = alua_rescan,
1151         .set_params = alua_set_params,
1152 };
1153
1154 static int __init alua_init(void)
1155 {
1156         int r;
1157
1158         kaluad_wq = alloc_workqueue("kaluad", WQ_MEM_RECLAIM, 0);
1159         if (!kaluad_wq) {
1160                 /* Temporary failure, bypass */
1161                 return SCSI_DH_DEV_TEMP_BUSY;
1162         }
1163         kaluad_sync_wq = create_workqueue("kaluad_sync");
1164         if (!kaluad_sync_wq) {
1165                 destroy_workqueue(kaluad_wq);
1166                 return SCSI_DH_DEV_TEMP_BUSY;
1167         }
1168         r = scsi_register_device_handler(&alua_dh);
1169         if (r != 0) {
1170                 printk(KERN_ERR "%s: Failed to register scsi device handler",
1171                         ALUA_DH_NAME);
1172                 destroy_workqueue(kaluad_sync_wq);
1173                 destroy_workqueue(kaluad_wq);
1174         }
1175         return r;
1176 }
1177
1178 static void __exit alua_exit(void)
1179 {
1180         scsi_unregister_device_handler(&alua_dh);
1181         destroy_workqueue(kaluad_sync_wq);
1182         destroy_workqueue(kaluad_wq);
1183 }
1184
1185 module_init(alua_init);
1186 module_exit(alua_exit);
1187
1188 MODULE_DESCRIPTION("DM Multipath ALUA support");
1189 MODULE_AUTHOR("Hannes Reinecke <hare@suse.de>");
1190 MODULE_LICENSE("GPL");
1191 MODULE_VERSION(ALUA_DH_VER);