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