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_dbg.h>
  28 #include <scsi/scsi_eh.h>
  29 #include <scsi/scsi_dh.h>
  30
  31 #define ALUA_DH_NAME "alua"
  32 #define ALUA_DH_VER "1.3"
  33
  34 #define TPGS_STATE_OPTIMIZED            0x0
  35 #define TPGS_STATE_NONOPTIMIZED         0x1
  36 #define TPGS_STATE_STANDBY              0x2
  37 #define TPGS_STATE_UNAVAILABLE          0x3
  38 #define TPGS_STATE_LBA_DEPENDENT        0x4
  39 #define TPGS_STATE_OFFLINE              0xe
  40 #define TPGS_STATE_TRANSITIONING        0xf
  41
  42 #define TPGS_SUPPORT_NONE               0x00
  43 #define TPGS_SUPPORT_OPTIMIZED          0x01
  44 #define TPGS_SUPPORT_NONOPTIMIZED       0x02
  45 #define TPGS_SUPPORT_STANDBY            0x04
  46 #define TPGS_SUPPORT_UNAVAILABLE        0x08
  47 #define TPGS_SUPPORT_LBA_DEPENDENT      0x10
  48 #define TPGS_SUPPORT_OFFLINE            0x40
  49 #define TPGS_SUPPORT_TRANSITION         0x80
  50
  51 #define RTPG_FMT_MASK                   0x70
  52 #define RTPG_FMT_EXT_HDR                0x10
  53
  54 #define TPGS_MODE_UNINITIALIZED          -1
  55 #define TPGS_MODE_NONE                  0x0
  56 #define TPGS_MODE_IMPLICIT              0x1
  57 #define TPGS_MODE_EXPLICIT              0x2
  58
  59 #define ALUA_RTPG_SIZE                  128
  60 #define ALUA_FAILOVER_TIMEOUT           60
  61 #define ALUA_FAILOVER_RETRIES           5
  62 #define ALUA_RTPG_DELAY_MSECS           5
  63
  64 /* device handler flags */
  65 #define ALUA_OPTIMIZE_STPG              0x01
  66 #define ALUA_RTPG_EXT_HDR_UNSUPP        0x02
  67 #define ALUA_SYNC_STPG                  0x04
  68 /* State machine flags */
  69 #define ALUA_PG_RUN_RTPG                0x10
  70 #define ALUA_PG_RUN_STPG                0x20
  71 #define ALUA_PG_RUNNING                 0x40
  72
  73 static uint optimize_stpg;
  74 module_param(optimize_stpg, uint, S_IRUGO|S_IWUSR);
  75 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.");
  76
  77 static LIST_HEAD(port_group_list);
  78 static DEFINE_SPINLOCK(port_group_lock);
  79 static struct workqueue_struct *kaluad_wq;
  80 static struct workqueue_struct *kaluad_sync_wq;
  81
  82 struct alua_port_group {
  83         struct kref             kref;
  84         struct rcu_head         rcu;
  85         struct list_head        node;
  86         unsigned char           device_id_str[256];
  87         int                     device_id_len;
  88         int                     group_id;
  89         int                     tpgs;
  90         int                     state;
  91         int                     pref;
  92         unsigned                flags; /* used for optimizing STPG */
  93         unsigned char           transition_tmo;
  94         unsigned long           expiry;
  95         unsigned long           interval;
  96         struct delayed_work     rtpg_work;
  97         spinlock_t              lock;
  98         struct list_head        rtpg_list;
  99         struct scsi_device      *rtpg_sdev;
 100 };
 101
 102 struct alua_dh_data {
 103         struct alua_port_group  *pg;
 104         int                     group_id;
 105         spinlock_t              pg_lock;
 106         struct scsi_device      *sdev;
 107         int                     init_error;
 108         struct mutex            init_mutex;
 109 };
 110
 111 struct alua_queue_data {
 112         struct list_head        entry;
 113         activate_complete       callback_fn;
 114         void                    *callback_data;
 115 };
 116
 117 #define ALUA_POLICY_SWITCH_CURRENT      0
 118 #define ALUA_POLICY_SWITCH_ALL          1
 119
 120 static void alua_rtpg_work(struct work_struct *work);
 121 static void alua_rtpg_queue(struct alua_port_group *pg,
 122                             struct scsi_device *sdev,
 123                             struct alua_queue_data *qdata, bool force);
 124 static void alua_check(struct scsi_device *sdev, bool force);
 125
 126 static void release_port_group(struct kref *kref)
 127 {
 128         struct alua_port_group *pg;
 129
 130         pg = container_of(kref, struct alua_port_group, kref);
 131         if (pg->rtpg_sdev)
 132                 flush_delayed_work(&pg->rtpg_work);
 133         spin_lock(&port_group_lock);
 134         list_del(&pg->node);
 135         spin_unlock(&port_group_lock);
 136         kfree_rcu(pg, rcu);
 137 }
 138
 139 /*
 140  * submit_rtpg - Issue a REPORT TARGET GROUP STATES command
 141  * @sdev: sdev the command should be sent to
 142  */
 143 static int submit_rtpg(struct scsi_device *sdev, unsigned char *buff,
 144                        int bufflen, struct scsi_sense_hdr *sshdr, int flags)
 145 {
 146         u8 cdb[COMMAND_SIZE(MAINTENANCE_IN)];
 147         int req_flags = REQ_FAILFAST_DEV | REQ_FAILFAST_TRANSPORT |
 148                 REQ_FAILFAST_DRIVER;
 149
 150         /* Prepare the command. */
 151         memset(cdb, 0x0, COMMAND_SIZE(MAINTENANCE_IN));
 152         cdb[0] = MAINTENANCE_IN;
 153         if (!(flags & ALUA_RTPG_EXT_HDR_UNSUPP))
 154                 cdb[1] = MI_REPORT_TARGET_PGS | MI_EXT_HDR_PARAM_FMT;
 155         else
 156                 cdb[1] = MI_REPORT_TARGET_PGS;
 157         put_unaligned_be32(bufflen, &cdb[6]);
 158
 159         return scsi_execute_req_flags(sdev, cdb, DMA_FROM_DEVICE,
 160                                       buff, bufflen, sshdr,
 161                                       ALUA_FAILOVER_TIMEOUT * HZ,
 162                                       ALUA_FAILOVER_RETRIES, NULL, req_flags);
 163 }
 164
 165 /*
 166  * submit_stpg - Issue a SET TARGET PORT GROUP command
 167  *
 168  * Currently we're only setting the current target port group state
 169  * to 'active/optimized' and let the array firmware figure out
 170  * the states of the remaining groups.
 171  */
 172 static int submit_stpg(struct scsi_device *sdev, int group_id,
 173                        struct scsi_sense_hdr *sshdr)
 174 {
 175         u8 cdb[COMMAND_SIZE(MAINTENANCE_OUT)];
 176         unsigned char stpg_data[8];
 177         int stpg_len = 8;
 178         int req_flags = REQ_FAILFAST_DEV | REQ_FAILFAST_TRANSPORT |
 179                 REQ_FAILFAST_DRIVER;
 180
 181         /* Prepare the data buffer */
 182         memset(stpg_data, 0, stpg_len);
 183         stpg_data[4] = TPGS_STATE_OPTIMIZED & 0x0f;
 184         put_unaligned_be16(group_id, &stpg_data[6]);
 185
 186         /* Prepare the command. */
 187         memset(cdb, 0x0, COMMAND_SIZE(MAINTENANCE_OUT));
 188         cdb[0] = MAINTENANCE_OUT;
 189         cdb[1] = MO_SET_TARGET_PGS;
 190         put_unaligned_be32(stpg_len, &cdb[6]);
 191
 192         return scsi_execute_req_flags(sdev, cdb, DMA_TO_DEVICE,
 193                                       stpg_data, stpg_len,
 194                                       sshdr, ALUA_FAILOVER_TIMEOUT * HZ,
 195                                       ALUA_FAILOVER_RETRIES, NULL, req_flags);
 196 }
 197
 198 struct alua_port_group *alua_find_get_pg(char *id_str, size_t id_size,
 199                                          int group_id)
 200 {
 201         struct alua_port_group *pg;
 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 != 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 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                  * Internal error: TPGS supported but no device
 241                  * identifcation found. Disable ALUA support.
 242                  */
 243                 kfree(pg);
 244                 sdev_printk(KERN_INFO, sdev,
 245                             "%s: No device descriptors found\n",
 246                             ALUA_DH_NAME);
 247                 return ERR_PTR(-ENXIO);
 248         }
 249         pg->group_id = group_id;
 250         pg->tpgs = tpgs;
 251         pg->state = TPGS_STATE_OPTIMIZED;
 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         spin_lock_init(&pg->lock);
 259
 260         spin_lock(&port_group_lock);
 261         tmp_pg = alua_find_get_pg(pg->device_id_str, pg->device_id_len,
 262                                   group_id);
 263         if (tmp_pg) {
 264                 spin_unlock(&port_group_lock);
 265                 kfree(pg);
 266                 return tmp_pg;
 267         }
 268
 269         list_add(&pg->node, &port_group_list);
 270         spin_unlock(&port_group_lock);
 271
 272         return pg;
 273 }
 274
 275 /*
 276  * alua_check_tpgs - Evaluate TPGS setting
 277  * @sdev: device to be checked
 278  *
 279  * Examine the TPGS setting of the sdev to find out if ALUA
 280  * is supported.
 281  */
 282 static int alua_check_tpgs(struct scsi_device *sdev)
 283 {
 284         int tpgs = TPGS_MODE_NONE;
 285
 286         /*
 287          * ALUA support for non-disk devices is fraught with
 288          * difficulties, so disable it for now.
 289          */
 290         if (sdev->type != TYPE_DISK) {
 291                 sdev_printk(KERN_INFO, sdev,
 292                             "%s: disable for non-disk devices\n",
 293                             ALUA_DH_NAME);
 294                 return tpgs;
 295         }
 296
 297         tpgs = scsi_device_tpgs(sdev);
 298         switch (tpgs) {
 299         case TPGS_MODE_EXPLICIT|TPGS_MODE_IMPLICIT:
 300                 sdev_printk(KERN_INFO, sdev,
 301                             "%s: supports implicit and explicit TPGS\n",
 302                             ALUA_DH_NAME);
 303                 break;
 304         case TPGS_MODE_EXPLICIT:
 305                 sdev_printk(KERN_INFO, sdev, "%s: supports explicit TPGS\n",
 306                             ALUA_DH_NAME);
 307                 break;
 308         case TPGS_MODE_IMPLICIT:
 309                 sdev_printk(KERN_INFO, sdev, "%s: supports implicit TPGS\n",
 310                             ALUA_DH_NAME);
 311                 break;
 312         case TPGS_MODE_NONE:
 313                 sdev_printk(KERN_INFO, sdev, "%s: not supported\n",
 314                             ALUA_DH_NAME);
 315                 break;
 316         default:
 317                 sdev_printk(KERN_INFO, sdev,
 318                             "%s: unsupported TPGS setting %d\n",
 319                             ALUA_DH_NAME, tpgs);
 320                 tpgs = TPGS_MODE_NONE;
 321                 break;
 322         }
 323
 324         return tpgs;
 325 }
 326
 327 /*
 328  * alua_check_vpd - Evaluate INQUIRY vpd page 0x83
 329  * @sdev: device to be checked
 330  *
 331  * Extract the relative target port and the target port group
 332  * descriptor from the list of identificators.
 333  */
 334 static int alua_check_vpd(struct scsi_device *sdev, struct alua_dh_data *h,
 335                           int tpgs)
 336 {
 337         int rel_port = -1, group_id;
 338         struct alua_port_group *pg, *old_pg = NULL;
 339
 340         group_id = scsi_vpd_tpg_id(sdev, &rel_port);
 341         if (group_id < 0) {
 342                 /*
 343                  * Internal error; TPGS supported but required
 344                  * VPD identification descriptors not present.
 345                  * Disable ALUA support
 346                  */
 347                 sdev_printk(KERN_INFO, sdev,
 348                             "%s: No target port descriptors found\n",
 349                             ALUA_DH_NAME);
 350                 return SCSI_DH_DEV_UNSUPP;
 351         }
 352
 353         pg = alua_alloc_pg(sdev, group_id, tpgs);
 354         if (IS_ERR(pg)) {
 355                 if (PTR_ERR(pg) == -ENOMEM)
 356                         return SCSI_DH_NOMEM;
 357                 return SCSI_DH_DEV_UNSUPP;
 358         }
 359         sdev_printk(KERN_INFO, sdev,
 360                     "%s: device %s port group %x rel port %x\n",
 361                     ALUA_DH_NAME, pg->device_id_str, group_id, rel_port);
 362
 363         /* Check for existing port group references */
 364         spin_lock(&h->pg_lock);
 365         old_pg = h->pg;
 366         if (old_pg != pg) {
 367                 /* port group has changed. Update to new port group */
 368                 rcu_assign_pointer(h->pg, pg);
 369         }
 370         if (sdev->synchronous_alua)
 371                 pg->flags |= ALUA_SYNC_STPG;
 372         alua_rtpg_queue(h->pg, sdev, NULL, true);
 373         spin_unlock(&h->pg_lock);
 374
 375         if (old_pg)
 376                 kref_put(&old_pg->kref, release_port_group);
 377
 378         return SCSI_DH_OK;
 379 }
 380
 381 static char print_alua_state(int state)
 382 {
 383         switch (state) {
 384         case TPGS_STATE_OPTIMIZED:
 385                 return 'A';
 386         case TPGS_STATE_NONOPTIMIZED:
 387                 return 'N';
 388         case TPGS_STATE_STANDBY:
 389                 return 'S';
 390         case TPGS_STATE_UNAVAILABLE:
 391                 return 'U';
 392         case TPGS_STATE_LBA_DEPENDENT:
 393                 return 'L';
 394         case TPGS_STATE_OFFLINE:
 395                 return 'O';
 396         case TPGS_STATE_TRANSITIONING:
 397                 return 'T';
 398         default:
 399                 return 'X';
 400         }
 401 }
 402
 403 static int alua_check_sense(struct scsi_device *sdev,
 404                             struct scsi_sense_hdr *sense_hdr)
 405 {
 406         switch (sense_hdr->sense_key) {
 407         case NOT_READY:
 408                 if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0a) {
 409                         /*
 410                          * LUN Not Accessible - ALUA state transition
 411                          */
 412                         alua_check(sdev, false);
 413                         return NEEDS_RETRY;
 414                 }
 415                 break;
 416         case UNIT_ATTENTION:
 417                 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x00) {
 418                         /*
 419                          * Power On, Reset, or Bus Device Reset.
 420                          * Might have obscured a state transition,
 421                          * so schedule a recheck.
 422                          */
 423                         alua_check(sdev, true);
 424                         return ADD_TO_MLQUEUE;
 425                 }
 426                 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x04)
 427                         /*
 428                          * Device internal reset
 429                          */
 430                         return ADD_TO_MLQUEUE;
 431                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x01)
 432                         /*
 433                          * Mode Parameters Changed
 434                          */
 435                         return ADD_TO_MLQUEUE;
 436                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x06) {
 437                         /*
 438                          * ALUA state changed
 439                          */
 440                         alua_check(sdev, true);
 441                         return ADD_TO_MLQUEUE;
 442                 }
 443                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x07) {
 444                         /*
 445                          * Implicit ALUA state transition failed
 446                          */
 447                         alua_check(sdev, true);
 448                         return ADD_TO_MLQUEUE;
 449                 }
 450                 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x03)
 451                         /*
 452                          * Inquiry data has changed
 453                          */
 454                         return ADD_TO_MLQUEUE;
 455                 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x0e)
 456                         /*
 457                          * REPORTED_LUNS_DATA_HAS_CHANGED is reported
 458                          * when switching controllers on targets like
 459                          * Intel Multi-Flex. We can just retry.
 460                          */
 461                         return ADD_TO_MLQUEUE;
 462                 break;
 463         }
 464
 465         return SCSI_RETURN_NOT_HANDLED;
 466 }
 467
 468 /*
 469  * alua_tur - Send a TEST UNIT READY
 470  * @sdev: device to which the TEST UNIT READY command should be send
 471  *
 472  * Send a TEST UNIT READY to @sdev to figure out the device state
 473  * Returns SCSI_DH_RETRY if the sense code is NOT READY/ALUA TRANSITIONING,
 474  * SCSI_DH_OK if no error occurred, and SCSI_DH_IO otherwise.
 475  */
 476 static int alua_tur(struct scsi_device *sdev)
 477 {
 478         struct scsi_sense_hdr sense_hdr;
 479         int retval;
 480
 481         retval = scsi_test_unit_ready(sdev, ALUA_FAILOVER_TIMEOUT * HZ,
 482                                       ALUA_FAILOVER_RETRIES, &sense_hdr);
 483         if (sense_hdr.sense_key == NOT_READY &&
 484             sense_hdr.asc == 0x04 && sense_hdr.ascq == 0x0a)
 485                 return SCSI_DH_RETRY;
 486         else if (retval)
 487                 return SCSI_DH_IO;
 488         else
 489                 return SCSI_DH_OK;
 490 }
 491
 492 /*
 493  * alua_rtpg - Evaluate REPORT TARGET GROUP STATES
 494  * @sdev: the device to be evaluated.
 495  *
 496  * Evaluate the Target Port Group State.
 497  * Returns SCSI_DH_DEV_OFFLINED if the path is
 498  * found to be unusable.
 499  */
 500 static int alua_rtpg(struct scsi_device *sdev, struct alua_port_group *pg)
 501 {
 502         struct scsi_sense_hdr sense_hdr;
 503         struct alua_port_group *tmp_pg;
 504         int len, k, off, valid_states = 0, bufflen = ALUA_RTPG_SIZE;
 505         unsigned char *desc, *buff;
 506         unsigned err, retval;
 507         unsigned int tpg_desc_tbl_off;
 508         unsigned char orig_transition_tmo;
 509         unsigned long flags;
 510
 511         if (!pg->expiry) {
 512                 unsigned long transition_tmo = ALUA_FAILOVER_TIMEOUT * HZ;
 513
 514                 if (pg->transition_tmo)
 515                         transition_tmo = pg->transition_tmo * HZ;
 516
 517                 pg->expiry = round_jiffies_up(jiffies + transition_tmo);
 518         }
 519
 520         buff = kzalloc(bufflen, GFP_KERNEL);
 521         if (!buff)
 522                 return SCSI_DH_DEV_TEMP_BUSY;
 523
 524  retry:
 525         retval = submit_rtpg(sdev, buff, bufflen, &sense_hdr, pg->flags);
 526
 527         if (retval) {
 528                 if (!scsi_sense_valid(&sense_hdr)) {
 529                         sdev_printk(KERN_INFO, sdev,
 530                                     "%s: rtpg failed, result %d\n",
 531                                     ALUA_DH_NAME, retval);
 532                         kfree(buff);
 533                         if (driver_byte(retval) == DRIVER_ERROR)
 534                                 return SCSI_DH_DEV_TEMP_BUSY;
 535                         return SCSI_DH_IO;
 536                 }
 537
 538                 /*
 539                  * submit_rtpg() has failed on existing arrays
 540                  * when requesting extended header info, and
 541                  * the array doesn't support extended headers,
 542                  * even though it shouldn't according to T10.
 543                  * The retry without rtpg_ext_hdr_req set
 544                  * handles this.
 545                  */
 546                 if (!(pg->flags & ALUA_RTPG_EXT_HDR_UNSUPP) &&
 547                     sense_hdr.sense_key == ILLEGAL_REQUEST &&
 548                     sense_hdr.asc == 0x24 && sense_hdr.ascq == 0) {
 549                         pg->flags |= ALUA_RTPG_EXT_HDR_UNSUPP;
 550                         goto retry;
 551                 }
 552                 /*
 553                  * Retry on ALUA state transition or if any
 554                  * UNIT ATTENTION occurred.
 555                  */
 556                 if (sense_hdr.sense_key == NOT_READY &&
 557                     sense_hdr.asc == 0x04 && sense_hdr.ascq == 0x0a)
 558                         err = SCSI_DH_RETRY;
 559                 else if (sense_hdr.sense_key == UNIT_ATTENTION)
 560                         err = SCSI_DH_RETRY;
 561                 if (err == SCSI_DH_RETRY &&
 562                     pg->expiry != 0 && time_before(jiffies, pg->expiry)) {
 563                         sdev_printk(KERN_ERR, sdev, "%s: rtpg retry\n",
 564                                     ALUA_DH_NAME);
 565                         scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
 566                         return err;
 567                 }
 568                 sdev_printk(KERN_ERR, sdev, "%s: rtpg failed\n",
 569                             ALUA_DH_NAME);
 570                 scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
 571                 kfree(buff);
 572                 pg->expiry = 0;
 573                 return SCSI_DH_IO;
 574         }
 575
 576         len = get_unaligned_be32(&buff[0]) + 4;
 577
 578         if (len > bufflen) {
 579                 /* Resubmit with the correct length */
 580                 kfree(buff);
 581                 bufflen = len;
 582                 buff = kmalloc(bufflen, GFP_KERNEL);
 583                 if (!buff) {
 584                         sdev_printk(KERN_WARNING, sdev,
 585                                     "%s: kmalloc buffer failed\n",__func__);
 586                         /* Temporary failure, bypass */
 587                         pg->expiry = 0;
 588                         return SCSI_DH_DEV_TEMP_BUSY;
 589                 }
 590                 goto retry;
 591         }
 592
 593         orig_transition_tmo = pg->transition_tmo;
 594         if ((buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR && buff[5] != 0)
 595                 pg->transition_tmo = buff[5];
 596         else
 597                 pg->transition_tmo = ALUA_FAILOVER_TIMEOUT;
 598
 599         if (orig_transition_tmo != pg->transition_tmo) {
 600                 sdev_printk(KERN_INFO, sdev,
 601                             "%s: transition timeout set to %d seconds\n",
 602                             ALUA_DH_NAME, pg->transition_tmo);
 603                 pg->expiry = jiffies + pg->transition_tmo * HZ;
 604         }
 605
 606         if ((buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR)
 607                 tpg_desc_tbl_off = 8;
 608         else
 609                 tpg_desc_tbl_off = 4;
 610
 611         for (k = tpg_desc_tbl_off, desc = buff + tpg_desc_tbl_off;
 612              k < len;
 613              k += off, desc += off) {
 614                 u16 group_id = get_unaligned_be16(&desc[2]);
 615
 616                 spin_lock_irqsave(&port_group_lock, flags);
 617                 tmp_pg = alua_find_get_pg(pg->device_id_str, pg->device_id_len,
 618                                           group_id);
 619                 spin_unlock_irqrestore(&port_group_lock, flags);
 620                 if (tmp_pg) {
 621                         if (spin_trylock_irqsave(&tmp_pg->lock, flags)) {
 622                                 if ((tmp_pg == pg) ||
 623                                     !(tmp_pg->flags & ALUA_PG_RUNNING)) {
 624                                         tmp_pg->state = desc[0] & 0x0f;
 625                                         tmp_pg->pref = desc[0] >> 7;
 626                                 }
 627                                 if (tmp_pg == pg)
 628                                         valid_states = desc[1];
 629                                 spin_unlock_irqrestore(&tmp_pg->lock, flags);
 630                         }
 631                         kref_put(&tmp_pg->kref, release_port_group);
 632                 }
 633                 off = 8 + (desc[7] * 4);
 634         }
 635
 636         spin_lock_irqsave(&pg->lock, flags);
 637         sdev_printk(KERN_INFO, sdev,
 638                     "%s: port group %02x state %c %s supports %c%c%c%c%c%c%c\n",
 639                     ALUA_DH_NAME, pg->group_id, print_alua_state(pg->state),
 640                     pg->pref ? "preferred" : "non-preferred",
 641                     valid_states&TPGS_SUPPORT_TRANSITION?'T':'t',
 642                     valid_states&TPGS_SUPPORT_OFFLINE?'O':'o',
 643                     valid_states&TPGS_SUPPORT_LBA_DEPENDENT?'L':'l',
 644                     valid_states&TPGS_SUPPORT_UNAVAILABLE?'U':'u',
 645                     valid_states&TPGS_SUPPORT_STANDBY?'S':'s',
 646                     valid_states&TPGS_SUPPORT_NONOPTIMIZED?'N':'n',
 647                     valid_states&TPGS_SUPPORT_OPTIMIZED?'A':'a');
 648
 649         switch (pg->state) {
 650         case TPGS_STATE_TRANSITIONING:
 651                 if (time_before(jiffies, pg->expiry)) {
 652                         /* State transition, retry */
 653                         pg->interval = 2;
 654                         err = SCSI_DH_RETRY;
 655                 } else {
 656                         /* Transitioning time exceeded, set port to standby */
 657                         err = SCSI_DH_IO;
 658                         pg->state = TPGS_STATE_STANDBY;
 659                         pg->expiry = 0;
 660                 }
 661                 break;
 662         case TPGS_STATE_OFFLINE:
 663                 /* Path unusable */
 664                 err = SCSI_DH_DEV_OFFLINED;
 665                 pg->expiry = 0;
 666                 break;
 667         default:
 668                 /* Useable path if active */
 669                 err = SCSI_DH_OK;
 670                 pg->expiry = 0;
 671                 break;
 672         }
 673         spin_unlock_irqrestore(&pg->lock, flags);
 674         kfree(buff);
 675         return err;
 676 }
 677
 678 /*
 679  * alua_stpg - Issue a SET TARGET PORT GROUP command
 680  *
 681  * Issue a SET TARGET PORT GROUP command and evaluate the
 682  * response. Returns SCSI_DH_RETRY per default to trigger
 683  * a re-evaluation of the target group state or SCSI_DH_OK
 684  * if no further action needs to be taken.
 685  */
 686 static unsigned alua_stpg(struct scsi_device *sdev, struct alua_port_group *pg)
 687 {
 688         int retval;
 689         struct scsi_sense_hdr sense_hdr;
 690
 691         if (!(pg->tpgs & TPGS_MODE_EXPLICIT)) {
 692                 /* Only implicit ALUA supported, retry */
 693                 return SCSI_DH_RETRY;
 694         }
 695         switch (pg->state) {
 696         case TPGS_STATE_OPTIMIZED:
 697                 return SCSI_DH_OK;
 698         case TPGS_STATE_NONOPTIMIZED:
 699                 if ((pg->flags & ALUA_OPTIMIZE_STPG) &&
 700                     !pg->pref &&
 701                     (pg->tpgs & TPGS_MODE_IMPLICIT))
 702                         return SCSI_DH_OK;
 703                 break;
 704         case TPGS_STATE_STANDBY:
 705         case TPGS_STATE_UNAVAILABLE:
 706                 break;
 707         case TPGS_STATE_OFFLINE:
 708                 return SCSI_DH_IO;
 709         case TPGS_STATE_TRANSITIONING:
 710                 break;
 711         default:
 712                 sdev_printk(KERN_INFO, sdev,
 713                             "%s: stpg failed, unhandled TPGS state %d",
 714                             ALUA_DH_NAME, pg->state);
 715                 return SCSI_DH_NOSYS;
 716         }
 717         retval = submit_stpg(sdev, pg->group_id, &sense_hdr);
 718
 719         if (retval) {
 720                 if (!scsi_sense_valid(&sense_hdr)) {
 721                         sdev_printk(KERN_INFO, sdev,
 722                                     "%s: stpg failed, result %d",
 723                                     ALUA_DH_NAME, retval);
 724                         if (driver_byte(retval) == DRIVER_ERROR)
 725                                 return SCSI_DH_DEV_TEMP_BUSY;
 726                 } else {
 727                         sdev_printk(KERN_INFO, sdev, "%s: stpg failed\n",
 728                                     ALUA_DH_NAME);
 729                         scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
 730                 }
 731         }
 732         /* Retry RTPG */
 733         return SCSI_DH_RETRY;
 734 }
 735
 736 static void alua_rtpg_work(struct work_struct *work)
 737 {
 738         struct alua_port_group *pg =
 739                 container_of(work, struct alua_port_group, rtpg_work.work);
 740         struct scsi_device *sdev;
 741         LIST_HEAD(qdata_list);
 742         int err = SCSI_DH_OK;
 743         struct alua_queue_data *qdata, *tmp;
 744         unsigned long flags;
 745         struct workqueue_struct *alua_wq = kaluad_wq;
 746
 747         spin_lock_irqsave(&pg->lock, flags);
 748         sdev = pg->rtpg_sdev;
 749         if (!sdev) {
 750                 WARN_ON(pg->flags & ALUA_PG_RUN_RTPG);
 751                 WARN_ON(pg->flags & ALUA_PG_RUN_STPG);
 752                 spin_unlock_irqrestore(&pg->lock, flags);
 753                 return;
 754         }
 755         if (pg->flags & ALUA_SYNC_STPG)
 756                 alua_wq = kaluad_sync_wq;
 757         pg->flags |= ALUA_PG_RUNNING;
 758         if (pg->flags & ALUA_PG_RUN_RTPG) {
 759                 int state = pg->state;
 760
 761                 pg->flags &= ~ALUA_PG_RUN_RTPG;
 762                 spin_unlock_irqrestore(&pg->lock, flags);
 763                 if (state == TPGS_STATE_TRANSITIONING) {
 764                         if (alua_tur(sdev) == SCSI_DH_RETRY) {
 765                                 spin_lock_irqsave(&pg->lock, flags);
 766                                 pg->flags &= ~ALUA_PG_RUNNING;
 767                                 pg->flags |= ALUA_PG_RUN_RTPG;
 768                                 spin_unlock_irqrestore(&pg->lock, flags);
 769                                 queue_delayed_work(alua_wq, &pg->rtpg_work,
 770                                                    pg->interval * HZ);
 771                                 return;
 772                         }
 773                         /* Send RTPG on failure or if TUR indicates SUCCESS */
 774                 }
 775                 err = alua_rtpg(sdev, pg);
 776                 spin_lock_irqsave(&pg->lock, flags);
 777                 if (err == SCSI_DH_RETRY || pg->flags & ALUA_PG_RUN_RTPG) {
 778                         pg->flags &= ~ALUA_PG_RUNNING;
 779                         pg->flags |= ALUA_PG_RUN_RTPG;
 780                         spin_unlock_irqrestore(&pg->lock, flags);
 781                         queue_delayed_work(alua_wq, &pg->rtpg_work,
 782                                            pg->interval * HZ);
 783                         return;
 784                 }
 785                 if (err != SCSI_DH_OK)
 786                         pg->flags &= ~ALUA_PG_RUN_STPG;
 787         }
 788         if (pg->flags & ALUA_PG_RUN_STPG) {
 789                 pg->flags &= ~ALUA_PG_RUN_STPG;
 790                 spin_unlock_irqrestore(&pg->lock, flags);
 791                 err = alua_stpg(sdev, pg);
 792                 spin_lock_irqsave(&pg->lock, flags);
 793                 if (err == SCSI_DH_RETRY || pg->flags & ALUA_PG_RUN_RTPG) {
 794                         pg->flags |= ALUA_PG_RUN_RTPG;
 795                         pg->interval = 0;
 796                         pg->flags &= ~ALUA_PG_RUNNING;
 797                         spin_unlock_irqrestore(&pg->lock, flags);
 798                         queue_delayed_work(alua_wq, &pg->rtpg_work,
 799                                            pg->interval * HZ);
 800                         return;
 801                 }
 802         }
 803
 804         list_splice_init(&pg->rtpg_list, &qdata_list);
 805         pg->rtpg_sdev = NULL;
 806         spin_unlock_irqrestore(&pg->lock, flags);
 807
 808         list_for_each_entry_safe(qdata, tmp, &qdata_list, entry) {
 809                 list_del(&qdata->entry);
 810                 if (qdata->callback_fn)
 811                         qdata->callback_fn(qdata->callback_data, err);
 812                 kfree(qdata);
 813         }
 814         spin_lock_irqsave(&pg->lock, flags);
 815         pg->flags &= ~ALUA_PG_RUNNING;
 816         spin_unlock_irqrestore(&pg->lock, flags);
 817         scsi_device_put(sdev);
 818         kref_put(&pg->kref, release_port_group);
 819 }
 820
 821 static void alua_rtpg_queue(struct alua_port_group *pg,
 822                             struct scsi_device *sdev,
 823                             struct alua_queue_data *qdata, bool force)
 824 {
 825         int start_queue = 0;
 826         unsigned long flags;
 827         struct workqueue_struct *alua_wq = kaluad_wq;
 828
 829         if (!pg)
 830                 return;
 831
 832         spin_lock_irqsave(&pg->lock, flags);
 833         if (qdata) {
 834                 list_add_tail(&qdata->entry, &pg->rtpg_list);
 835                 pg->flags |= ALUA_PG_RUN_STPG;
 836                 force = true;
 837         }
 838         if (pg->rtpg_sdev == NULL) {
 839                 pg->interval = 0;
 840                 pg->flags |= ALUA_PG_RUN_RTPG;
 841                 kref_get(&pg->kref);
 842                 pg->rtpg_sdev = sdev;
 843                 scsi_device_get(sdev);
 844                 start_queue = 1;
 845         } else if (!(pg->flags & ALUA_PG_RUN_RTPG) && force) {
 846                 pg->flags |= ALUA_PG_RUN_RTPG;
 847                 /* Do not queue if the worker is already running */
 848                 if (!(pg->flags & ALUA_PG_RUNNING)) {
 849                         kref_get(&pg->kref);
 850                         start_queue = 1;
 851                 }
 852         }
 853
 854         if (pg->flags & ALUA_SYNC_STPG)
 855                 alua_wq = kaluad_sync_wq;
 856         spin_unlock_irqrestore(&pg->lock, flags);
 857
 858         if (start_queue &&
 859             !queue_delayed_work(alua_wq, &pg->rtpg_work,
 860                                 msecs_to_jiffies(ALUA_RTPG_DELAY_MSECS))) {
 861                 scsi_device_put(sdev);
 862                 kref_put(&pg->kref, release_port_group);
 863         }
 864 }
 865
 866 /*
 867  * alua_initialize - Initialize ALUA state
 868  * @sdev: the device to be initialized
 869  *
 870  * For the prep_fn to work correctly we have
 871  * to initialize the ALUA state for the device.
 872  */
 873 static int alua_initialize(struct scsi_device *sdev, struct alua_dh_data *h)
 874 {
 875         int err = SCSI_DH_DEV_UNSUPP, tpgs;
 876
 877         mutex_lock(&h->init_mutex);
 878         tpgs = alua_check_tpgs(sdev);
 879         if (tpgs != TPGS_MODE_NONE)
 880                 err = alua_check_vpd(sdev, h, tpgs);
 881         h->init_error = err;
 882         mutex_unlock(&h->init_mutex);
 883         return err;
 884 }
 885 /*
 886  * alua_set_params - set/unset the optimize flag
 887  * @sdev: device on the path to be activated
 888  * params - parameters in the following format
 889  *      "no_of_params\0param1\0param2\0param3\0...\0"
 890  * For example, to set the flag pass the following parameters
 891  * from multipath.conf
 892  *     hardware_handler        "2 alua 1"
 893  */
 894 static int alua_set_params(struct scsi_device *sdev, const char *params)
 895 {
 896         struct alua_dh_data *h = sdev->handler_data;
 897         struct alua_port_group __rcu *pg = NULL;
 898         unsigned int optimize = 0, argc;
 899         const char *p = params;
 900         int result = SCSI_DH_OK;
 901         unsigned long flags;
 902
 903         if ((sscanf(params, "%u", &argc) != 1) || (argc != 1))
 904                 return -EINVAL;
 905
 906         while (*p++)
 907                 ;
 908         if ((sscanf(p, "%u", &optimize) != 1) || (optimize > 1))
 909                 return -EINVAL;
 910
 911         rcu_read_lock();
 912         pg = rcu_dereference(h->pg);
 913         if (!pg) {
 914                 rcu_read_unlock();
 915                 return -ENXIO;
 916         }
 917         spin_lock_irqsave(&pg->lock, flags);
 918         if (optimize)
 919                 pg->flags |= ALUA_OPTIMIZE_STPG;
 920         else
 921                 pg->flags &= ~ALUA_OPTIMIZE_STPG;
 922         spin_unlock_irqrestore(&pg->lock, flags);
 923         rcu_read_unlock();
 924
 925         return result;
 926 }
 927
 928 /*
 929  * alua_activate - activate a path
 930  * @sdev: device on the path to be activated
 931  *
 932  * We're currently switching the port group to be activated only and
 933  * let the array figure out the rest.
 934  * There may be other arrays which require us to switch all port groups
 935  * based on a certain policy. But until we actually encounter them it
 936  * should be okay.
 937  */
 938 static int alua_activate(struct scsi_device *sdev,
 939                         activate_complete fn, void *data)
 940 {
 941         struct alua_dh_data *h = sdev->handler_data;
 942         int err = SCSI_DH_OK;
 943         struct alua_queue_data *qdata;
 944         struct alua_port_group __rcu *pg;
 945
 946         qdata = kzalloc(sizeof(*qdata), GFP_KERNEL);
 947         if (!qdata) {
 948                 err = SCSI_DH_RES_TEMP_UNAVAIL;
 949                 goto out;
 950         }
 951         qdata->callback_fn = fn;
 952         qdata->callback_data = data;
 953
 954         mutex_lock(&h->init_mutex);
 955         rcu_read_lock();
 956         pg = rcu_dereference(h->pg);
 957         if (!pg || !kref_get_unless_zero(&pg->kref)) {
 958                 rcu_read_unlock();
 959                 kfree(qdata);
 960                 err = h->init_error;
 961                 mutex_unlock(&h->init_mutex);
 962                 goto out;
 963         }
 964         fn = NULL;
 965         rcu_read_unlock();
 966         mutex_unlock(&h->init_mutex);
 967
 968         alua_rtpg_queue(pg, sdev, qdata, true);
 969         kref_put(&pg->kref, release_port_group);
 970 out:
 971         if (fn)
 972                 fn(data, err);
 973         return 0;
 974 }
 975
 976 /*
 977  * alua_check - check path status
 978  * @sdev: device on the path to be checked
 979  *
 980  * Check the device status
 981  */
 982 static void alua_check(struct scsi_device *sdev, bool force)
 983 {
 984         struct alua_dh_data *h = sdev->handler_data;
 985         struct alua_port_group *pg;
 986
 987         rcu_read_lock();
 988         pg = rcu_dereference(h->pg);
 989         if (!pg || !kref_get_unless_zero(&pg->kref)) {
 990                 rcu_read_unlock();
 991                 return;
 992         }
 993         rcu_read_unlock();
 994
 995         alua_rtpg_queue(pg, sdev, NULL, force);
 996         kref_put(&pg->kref, release_port_group);
 997 }
 998
 999 /*
1000  * alua_prep_fn - request callback
1001  *
1002  * Fail I/O to all paths not in state
1003  * active/optimized or active/non-optimized.
1004  */
1005 static int alua_prep_fn(struct scsi_device *sdev, struct request *req)
1006 {
1007         struct alua_dh_data *h = sdev->handler_data;
1008         struct alua_port_group __rcu *pg;
1009         int state = TPGS_STATE_OPTIMIZED;
1010         int ret = BLKPREP_OK;
1011
1012         rcu_read_lock();
1013         pg = rcu_dereference(h->pg);
1014         if (pg)
1015                 state = pg->state;
1016         rcu_read_unlock();
1017         if (state == TPGS_STATE_TRANSITIONING)
1018                 ret = BLKPREP_DEFER;
1019         else if (state != TPGS_STATE_OPTIMIZED &&
1020                  state != TPGS_STATE_NONOPTIMIZED &&
1021                  state != TPGS_STATE_LBA_DEPENDENT) {
1022                 ret = BLKPREP_KILL;
1023                 req->cmd_flags |= REQ_QUIET;
1024         }
1025         return ret;
1026
1027 }
1028
1029 static void alua_rescan(struct scsi_device *sdev)
1030 {
1031         struct alua_dh_data *h = sdev->handler_data;
1032
1033         alua_initialize(sdev, h);
1034 }
1035
1036 /*
1037  * alua_bus_attach - Attach device handler
1038  * @sdev: device to be attached to
1039  */
1040 static int alua_bus_attach(struct scsi_device *sdev)
1041 {
1042         struct alua_dh_data *h;
1043         int err, ret = -EINVAL;
1044
1045         h = kzalloc(sizeof(*h) , GFP_KERNEL);
1046         if (!h)
1047                 return -ENOMEM;
1048         spin_lock_init(&h->pg_lock);
1049         rcu_assign_pointer(h->pg, NULL);
1050         h->init_error = SCSI_DH_OK;
1051         h->sdev = sdev;
1052
1053         mutex_init(&h->init_mutex);
1054         err = alua_initialize(sdev, h);
1055         if (err == SCSI_DH_NOMEM)
1056                 ret = -ENOMEM;
1057         if (err != SCSI_DH_OK && err != SCSI_DH_DEV_OFFLINED)
1058                 goto failed;
1059
1060         sdev->handler_data = h;
1061         return 0;
1062 failed:
1063         kfree(h);
1064         return ret;
1065 }
1066
1067 /*
1068  * alua_bus_detach - Detach device handler
1069  * @sdev: device to be detached from
1070  */
1071 static void alua_bus_detach(struct scsi_device *sdev)
1072 {
1073         struct alua_dh_data *h = sdev->handler_data;
1074         struct alua_port_group *pg;
1075
1076         spin_lock(&h->pg_lock);
1077         pg = h->pg;
1078         rcu_assign_pointer(h->pg, NULL);
1079         h->sdev = NULL;
1080         spin_unlock(&h->pg_lock);
1081         if (pg)
1082                 kref_put(&pg->kref, release_port_group);
1083
1084         sdev->handler_data = NULL;
1085         kfree(h);
1086 }
1087
1088 static struct scsi_device_handler alua_dh = {
1089         .name = ALUA_DH_NAME,
1090         .module = THIS_MODULE,
1091         .attach = alua_bus_attach,
1092         .detach = alua_bus_detach,
1093         .prep_fn = alua_prep_fn,
1094         .check_sense = alua_check_sense,
1095         .activate = alua_activate,
1096         .rescan = alua_rescan,
1097         .set_params = alua_set_params,
1098 };
1099
1100 static int __init alua_init(void)
1101 {
1102         int r;
1103
1104         kaluad_wq = alloc_workqueue("kaluad", WQ_MEM_RECLAIM, 0);
1105         if (!kaluad_wq) {
1106                 /* Temporary failure, bypass */
1107                 return SCSI_DH_DEV_TEMP_BUSY;
1108         }
1109         kaluad_sync_wq = create_workqueue("kaluad_sync");
1110         if (!kaluad_sync_wq) {
1111                 destroy_workqueue(kaluad_wq);
1112                 return SCSI_DH_DEV_TEMP_BUSY;
1113         }
1114         r = scsi_register_device_handler(&alua_dh);
1115         if (r != 0) {
1116                 printk(KERN_ERR "%s: Failed to register scsi device handler",
1117                         ALUA_DH_NAME);
1118                 destroy_workqueue(kaluad_sync_wq);
1119                 destroy_workqueue(kaluad_wq);
1120         }
1121         return r;
1122 }
1123
1124 static void __exit alua_exit(void)
1125 {
1126         scsi_unregister_device_handler(&alua_dh);
1127         destroy_workqueue(kaluad_sync_wq);
1128         destroy_workqueue(kaluad_wq);
1129 }
1130
1131 module_init(alua_init);
1132 module_exit(alua_exit);
1133
1134 MODULE_DESCRIPTION("DM Multipath ALUA support");
1135 MODULE_AUTHOR("Hannes Reinecke <hare@suse.de>");
1136 MODULE_LICENSE("GPL");
1137 MODULE_VERSION(ALUA_DH_VER);