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_INQUIRY_SIZE               36
  60 #define ALUA_FAILOVER_TIMEOUT           60
  61 #define ALUA_FAILOVER_RETRIES           5
  62
  63 /* device handler flags */
  64 #define ALUA_OPTIMIZE_STPG              1
  65 #define ALUA_RTPG_EXT_HDR_UNSUPP        2
  66
  67 struct alua_dh_data {
  68         int                     group_id;
  69         int                     rel_port;
  70         int                     tpgs;
  71         int                     state;
  72         int                     pref;
  73         unsigned                flags; /* used for optimizing STPG */
  74         unsigned char           inq[ALUA_INQUIRY_SIZE];
  75         unsigned char           *buff;
  76         int                     bufflen;
  77         unsigned char           transition_tmo;
  78         unsigned char           sense[SCSI_SENSE_BUFFERSIZE];
  79         struct scsi_device      *sdev;
  80         activate_complete       callback_fn;
  81         void                    *callback_data;
  82 };
  83
  84 #define ALUA_POLICY_SWITCH_CURRENT      0
  85 #define ALUA_POLICY_SWITCH_ALL          1
  86
  87 static char print_alua_state(int);
  88
  89 static int realloc_buffer(struct alua_dh_data *h, unsigned len)
  90 {
  91         if (h->buff && h->buff != h->inq)
  92                 kfree(h->buff);
  93
  94         h->buff = kmalloc(len, GFP_NOIO);
  95         if (!h->buff) {
  96                 h->buff = h->inq;
  97                 h->bufflen = ALUA_INQUIRY_SIZE;
  98                 return 1;
  99         }
 100         h->bufflen = len;
 101         return 0;
 102 }
 103
 104 static struct request *get_alua_req(struct scsi_device *sdev,
 105                                     void *buffer, unsigned buflen, int rw)
 106 {
 107         struct request *rq;
 108         struct request_queue *q = sdev->request_queue;
 109
 110         rq = blk_get_request(q, rw, GFP_NOIO);
 111
 112         if (IS_ERR(rq)) {
 113                 sdev_printk(KERN_INFO, sdev,
 114                             "%s: blk_get_request failed\n", __func__);
 115                 return NULL;
 116         }
 117         blk_rq_set_block_pc(rq);
 118
 119         if (buflen && blk_rq_map_kern(q, rq, buffer, buflen, GFP_NOIO)) {
 120                 blk_put_request(rq);
 121                 sdev_printk(KERN_INFO, sdev,
 122                             "%s: blk_rq_map_kern failed\n", __func__);
 123                 return NULL;
 124         }
 125
 126         rq->cmd_flags |= REQ_FAILFAST_DEV | REQ_FAILFAST_TRANSPORT |
 127                          REQ_FAILFAST_DRIVER;
 128         rq->retries = ALUA_FAILOVER_RETRIES;
 129         rq->timeout = ALUA_FAILOVER_TIMEOUT * HZ;
 130
 131         return rq;
 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 unsigned submit_rtpg(struct scsi_device *sdev, struct alua_dh_data *h)
 139 {
 140         struct request *rq;
 141         int err = 0;
 142
 143         rq = get_alua_req(sdev, h->buff, h->bufflen, READ);
 144         if (!rq) {
 145                 err = DRIVER_BUSY << 24;
 146                 goto done;
 147         }
 148
 149         /* Prepare the command. */
 150         rq->cmd[0] = MAINTENANCE_IN;
 151         if (!(h->flags & ALUA_RTPG_EXT_HDR_UNSUPP))
 152                 rq->cmd[1] = MI_REPORT_TARGET_PGS | MI_EXT_HDR_PARAM_FMT;
 153         else
 154                 rq->cmd[1] = MI_REPORT_TARGET_PGS;
 155         put_unaligned_be32(h->bufflen, &rq->cmd[6]);
 156         rq->cmd_len = COMMAND_SIZE(MAINTENANCE_IN);
 157
 158         rq->sense = h->sense;
 159         memset(rq->sense, 0, SCSI_SENSE_BUFFERSIZE);
 160         rq->sense_len = 0;
 161
 162         blk_execute_rq(rq->q, NULL, rq, 1);
 163         if (rq->errors)
 164                 err = rq->errors;
 165         blk_put_request(rq);
 166 done:
 167         return err;
 168 }
 169
 170 /*
 171  * stpg_endio - Evaluate SET TARGET GROUP STATES
 172  * @sdev: the device to be evaluated
 173  * @state: the new target group state
 174  *
 175  * Evaluate a SET TARGET GROUP STATES command response.
 176  */
 177 static void stpg_endio(struct request *req, int error)
 178 {
 179         struct alua_dh_data *h = req->end_io_data;
 180         struct scsi_sense_hdr sense_hdr;
 181         unsigned err = SCSI_DH_OK;
 182
 183         if (host_byte(req->errors) != DID_OK ||
 184             msg_byte(req->errors) != COMMAND_COMPLETE) {
 185                 err = SCSI_DH_IO;
 186                 goto done;
 187         }
 188
 189         if (scsi_normalize_sense(h->sense, SCSI_SENSE_BUFFERSIZE,
 190                                  &sense_hdr)) {
 191                 if (sense_hdr.sense_key == NOT_READY &&
 192                     sense_hdr.asc == 0x04 && sense_hdr.ascq == 0x0a) {
 193                         /* ALUA state transition already in progress */
 194                         err = SCSI_DH_OK;
 195                         goto done;
 196                 }
 197                 if (sense_hdr.sense_key == UNIT_ATTENTION) {
 198                         err = SCSI_DH_RETRY;
 199                         goto done;
 200                 }
 201                 sdev_printk(KERN_INFO, h->sdev, "%s: stpg failed\n",
 202                             ALUA_DH_NAME);
 203                 scsi_print_sense_hdr(h->sdev, ALUA_DH_NAME, &sense_hdr);
 204                 err = SCSI_DH_IO;
 205         } else if (error)
 206                 err = SCSI_DH_IO;
 207
 208         if (err == SCSI_DH_OK) {
 209                 h->state = TPGS_STATE_OPTIMIZED;
 210                 sdev_printk(KERN_INFO, h->sdev,
 211                             "%s: port group %02x switched to state %c\n",
 212                             ALUA_DH_NAME, h->group_id,
 213                             print_alua_state(h->state));
 214         }
 215 done:
 216         req->end_io_data = NULL;
 217         __blk_put_request(req->q, req);
 218         if (h->callback_fn) {
 219                 h->callback_fn(h->callback_data, err);
 220                 h->callback_fn = h->callback_data = NULL;
 221         }
 222         return;
 223 }
 224
 225 /*
 226  * submit_stpg - Issue a SET TARGET GROUP STATES command
 227  *
 228  * Currently we're only setting the current target port group state
 229  * to 'active/optimized' and let the array firmware figure out
 230  * the states of the remaining groups.
 231  */
 232 static unsigned submit_stpg(struct alua_dh_data *h)
 233 {
 234         struct request *rq;
 235         int stpg_len = 8;
 236         struct scsi_device *sdev = h->sdev;
 237
 238         /* Prepare the data buffer */
 239         memset(h->buff, 0, stpg_len);
 240         h->buff[4] = TPGS_STATE_OPTIMIZED & 0x0f;
 241         put_unaligned_be16(h->group_id, &h->buff[6]);
 242
 243         rq = get_alua_req(sdev, h->buff, stpg_len, WRITE);
 244         if (!rq)
 245                 return SCSI_DH_RES_TEMP_UNAVAIL;
 246
 247         /* Prepare the command. */
 248         rq->cmd[0] = MAINTENANCE_OUT;
 249         rq->cmd[1] = MO_SET_TARGET_PGS;
 250         put_unaligned_be32(stpg_len, &rq->cmd[6]);
 251         rq->cmd_len = COMMAND_SIZE(MAINTENANCE_OUT);
 252
 253         rq->sense = h->sense;
 254         memset(rq->sense, 0, SCSI_SENSE_BUFFERSIZE);
 255         rq->sense_len = 0;
 256         rq->end_io_data = h;
 257
 258         blk_execute_rq_nowait(rq->q, NULL, rq, 1, stpg_endio);
 259         return SCSI_DH_OK;
 260 }
 261
 262 /*
 263  * alua_check_tpgs - Evaluate TPGS setting
 264  * @sdev: device to be checked
 265  *
 266  * Examine the TPGS setting of the sdev to find out if ALUA
 267  * is supported.
 268  */
 269 static int alua_check_tpgs(struct scsi_device *sdev)
 270 {
 271         int tpgs = TPGS_MODE_NONE;
 272
 273         /*
 274          * ALUA support for non-disk devices is fraught with
 275          * difficulties, so disable it for now.
 276          */
 277         if (sdev->type != TYPE_DISK) {
 278                 sdev_printk(KERN_INFO, sdev,
 279                             "%s: disable for non-disk devices\n",
 280                             ALUA_DH_NAME);
 281                 return tpgs;
 282         }
 283
 284         tpgs = scsi_device_tpgs(sdev);
 285         switch (tpgs) {
 286         case TPGS_MODE_EXPLICIT|TPGS_MODE_IMPLICIT:
 287                 sdev_printk(KERN_INFO, sdev,
 288                             "%s: supports implicit and explicit TPGS\n",
 289                             ALUA_DH_NAME);
 290                 break;
 291         case TPGS_MODE_EXPLICIT:
 292                 sdev_printk(KERN_INFO, sdev, "%s: supports explicit TPGS\n",
 293                             ALUA_DH_NAME);
 294                 break;
 295         case TPGS_MODE_IMPLICIT:
 296                 sdev_printk(KERN_INFO, sdev, "%s: supports implicit TPGS\n",
 297                             ALUA_DH_NAME);
 298                 break;
 299         case TPGS_MODE_NONE:
 300                 sdev_printk(KERN_INFO, sdev, "%s: not supported\n",
 301                             ALUA_DH_NAME);
 302                 break;
 303         default:
 304                 sdev_printk(KERN_INFO, sdev,
 305                             "%s: unsupported TPGS setting %d\n",
 306                             ALUA_DH_NAME, tpgs);
 307                 tpgs = TPGS_MODE_NONE;
 308                 break;
 309         }
 310
 311         return tpgs;
 312 }
 313
 314 /*
 315  * alua_check_vpd - Evaluate INQUIRY vpd page 0x83
 316  * @sdev: device to be checked
 317  *
 318  * Extract the relative target port and the target port group
 319  * descriptor from the list of identificators.
 320  */
 321 static int alua_check_vpd(struct scsi_device *sdev, struct alua_dh_data *h)
 322 {
 323         int rel_port = -1, group_id;
 324
 325         group_id = scsi_vpd_tpg_id(sdev, &rel_port);
 326         if (group_id < 0) {
 327                 /*
 328                  * Internal error; TPGS supported but required
 329                  * VPD identification descriptors not present.
 330                  * Disable ALUA support
 331                  */
 332                 sdev_printk(KERN_INFO, sdev,
 333                             "%s: No target port descriptors found\n",
 334                             ALUA_DH_NAME);
 335                 return SCSI_DH_DEV_UNSUPP;
 336         }
 337         h->state = TPGS_STATE_OPTIMIZED;
 338         h->group_id = group_id;
 339
 340         sdev_printk(KERN_INFO, sdev,
 341                     "%s: port group %02x rel port %02x\n",
 342                     ALUA_DH_NAME, h->group_id, h->rel_port);
 343
 344         return 0;
 345 }
 346
 347 static char print_alua_state(int state)
 348 {
 349         switch (state) {
 350         case TPGS_STATE_OPTIMIZED:
 351                 return 'A';
 352         case TPGS_STATE_NONOPTIMIZED:
 353                 return 'N';
 354         case TPGS_STATE_STANDBY:
 355                 return 'S';
 356         case TPGS_STATE_UNAVAILABLE:
 357                 return 'U';
 358         case TPGS_STATE_LBA_DEPENDENT:
 359                 return 'L';
 360         case TPGS_STATE_OFFLINE:
 361                 return 'O';
 362         case TPGS_STATE_TRANSITIONING:
 363                 return 'T';
 364         default:
 365                 return 'X';
 366         }
 367 }
 368
 369 static int alua_check_sense(struct scsi_device *sdev,
 370                             struct scsi_sense_hdr *sense_hdr)
 371 {
 372         switch (sense_hdr->sense_key) {
 373         case NOT_READY:
 374                 if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0a)
 375                         /*
 376                          * LUN Not Accessible - ALUA state transition
 377                          */
 378                         return ADD_TO_MLQUEUE;
 379                 break;
 380         case UNIT_ATTENTION:
 381                 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x00)
 382                         /*
 383                          * Power On, Reset, or Bus Device Reset, just retry.
 384                          */
 385                         return ADD_TO_MLQUEUE;
 386                 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x04)
 387                         /*
 388                          * Device internal reset
 389                          */
 390                         return ADD_TO_MLQUEUE;
 391                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x01)
 392                         /*
 393                          * Mode Parameters Changed
 394                          */
 395                         return ADD_TO_MLQUEUE;
 396                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x06)
 397                         /*
 398                          * ALUA state changed
 399                          */
 400                         return ADD_TO_MLQUEUE;
 401                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x07)
 402                         /*
 403                          * Implicit ALUA state transition failed
 404                          */
 405                         return ADD_TO_MLQUEUE;
 406                 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x03)
 407                         /*
 408                          * Inquiry data has changed
 409                          */
 410                         return ADD_TO_MLQUEUE;
 411                 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x0e)
 412                         /*
 413                          * REPORTED_LUNS_DATA_HAS_CHANGED is reported
 414                          * when switching controllers on targets like
 415                          * Intel Multi-Flex. We can just retry.
 416                          */
 417                         return ADD_TO_MLQUEUE;
 418                 break;
 419         }
 420
 421         return SCSI_RETURN_NOT_HANDLED;
 422 }
 423
 424 /*
 425  * alua_rtpg - Evaluate REPORT TARGET GROUP STATES
 426  * @sdev: the device to be evaluated.
 427  * @wait_for_transition: if nonzero, wait ALUA_FAILOVER_TIMEOUT seconds for device to exit transitioning state
 428  *
 429  * Evaluate the Target Port Group State.
 430  * Returns SCSI_DH_DEV_OFFLINED if the path is
 431  * found to be unusable.
 432  */
 433 static int alua_rtpg(struct scsi_device *sdev, struct alua_dh_data *h, int wait_for_transition)
 434 {
 435         struct scsi_sense_hdr sense_hdr;
 436         int len, k, off, valid_states = 0;
 437         unsigned char *ucp;
 438         unsigned err, retval;
 439         unsigned long expiry, interval = 0;
 440         unsigned int tpg_desc_tbl_off;
 441         unsigned char orig_transition_tmo;
 442
 443         if (!h->transition_tmo)
 444                 expiry = round_jiffies_up(jiffies + ALUA_FAILOVER_TIMEOUT * HZ);
 445         else
 446                 expiry = round_jiffies_up(jiffies + h->transition_tmo * HZ);
 447
 448  retry:
 449         retval = submit_rtpg(sdev, h);
 450         if (retval) {
 451                 if (!scsi_normalize_sense(h->sense, SCSI_SENSE_BUFFERSIZE,
 452                                           &sense_hdr)) {
 453                         sdev_printk(KERN_INFO, sdev,
 454                                     "%s: rtpg failed, result %d\n",
 455                                     ALUA_DH_NAME, retval);
 456                         if (driver_byte(retval) == DRIVER_BUSY)
 457                                 return SCSI_DH_DEV_TEMP_BUSY;
 458                         return SCSI_DH_IO;
 459                 }
 460
 461                 /*
 462                  * submit_rtpg() has failed on existing arrays
 463                  * when requesting extended header info, and
 464                  * the array doesn't support extended headers,
 465                  * even though it shouldn't according to T10.
 466                  * The retry without rtpg_ext_hdr_req set
 467                  * handles this.
 468                  */
 469                 if (!(h->flags & ALUA_RTPG_EXT_HDR_UNSUPP) &&
 470                     sense_hdr.sense_key == ILLEGAL_REQUEST &&
 471                     sense_hdr.asc == 0x24 && sense_hdr.ascq == 0) {
 472                         h->flags |= ALUA_RTPG_EXT_HDR_UNSUPP;
 473                         goto retry;
 474                 }
 475                 /*
 476                  * Retry on ALUA state transition or if any
 477                  * UNIT ATTENTION occurred.
 478                  */
 479                 if (sense_hdr.sense_key == NOT_READY &&
 480                     sense_hdr.asc == 0x04 && sense_hdr.ascq == 0x0a)
 481                         err = SCSI_DH_RETRY;
 482                 else if (sense_hdr.sense_key == UNIT_ATTENTION)
 483                         err = SCSI_DH_RETRY;
 484                 if (err == SCSI_DH_RETRY && time_before(jiffies, expiry)) {
 485                         sdev_printk(KERN_ERR, sdev, "%s: rtpg retry\n",
 486                                     ALUA_DH_NAME);
 487                         scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
 488                         goto retry;
 489                 }
 490                 sdev_printk(KERN_ERR, sdev, "%s: rtpg failed\n",
 491                             ALUA_DH_NAME);
 492                 scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);
 493                 return SCSI_DH_IO;
 494         }
 495
 496         len = get_unaligned_be32(&h->buff[0]) + 4;
 497
 498         if (len > h->bufflen) {
 499                 /* Resubmit with the correct length */
 500                 if (realloc_buffer(h, len)) {
 501                         sdev_printk(KERN_WARNING, sdev,
 502                                     "%s: kmalloc buffer failed\n",__func__);
 503                         /* Temporary failure, bypass */
 504                         return SCSI_DH_DEV_TEMP_BUSY;
 505                 }
 506                 goto retry;
 507         }
 508
 509         orig_transition_tmo = h->transition_tmo;
 510         if ((h->buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR && h->buff[5] != 0)
 511                 h->transition_tmo = h->buff[5];
 512         else
 513                 h->transition_tmo = ALUA_FAILOVER_TIMEOUT;
 514
 515         if (wait_for_transition && (orig_transition_tmo != h->transition_tmo)) {
 516                 sdev_printk(KERN_INFO, sdev,
 517                             "%s: transition timeout set to %d seconds\n",
 518                             ALUA_DH_NAME, h->transition_tmo);
 519                 expiry = jiffies + h->transition_tmo * HZ;
 520         }
 521
 522         if ((h->buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR)
 523                 tpg_desc_tbl_off = 8;
 524         else
 525                 tpg_desc_tbl_off = 4;
 526
 527         for (k = tpg_desc_tbl_off, ucp = h->buff + tpg_desc_tbl_off;
 528              k < len;
 529              k += off, ucp += off) {
 530
 531                 if (h->group_id == get_unaligned_be16(&ucp[2])) {
 532                         h->state = ucp[0] & 0x0f;
 533                         h->pref = ucp[0] >> 7;
 534                         valid_states = ucp[1];
 535                 }
 536                 off = 8 + (ucp[7] * 4);
 537         }
 538
 539         sdev_printk(KERN_INFO, sdev,
 540                     "%s: port group %02x state %c %s supports %c%c%c%c%c%c%c\n",
 541                     ALUA_DH_NAME, h->group_id, print_alua_state(h->state),
 542                     h->pref ? "preferred" : "non-preferred",
 543                     valid_states&TPGS_SUPPORT_TRANSITION?'T':'t',
 544                     valid_states&TPGS_SUPPORT_OFFLINE?'O':'o',
 545                     valid_states&TPGS_SUPPORT_LBA_DEPENDENT?'L':'l',
 546                     valid_states&TPGS_SUPPORT_UNAVAILABLE?'U':'u',
 547                     valid_states&TPGS_SUPPORT_STANDBY?'S':'s',
 548                     valid_states&TPGS_SUPPORT_NONOPTIMIZED?'N':'n',
 549                     valid_states&TPGS_SUPPORT_OPTIMIZED?'A':'a');
 550
 551         switch (h->state) {
 552         case TPGS_STATE_TRANSITIONING:
 553                 if (wait_for_transition) {
 554                         if (time_before(jiffies, expiry)) {
 555                                 /* State transition, retry */
 556                                 interval += 2000;
 557                                 msleep(interval);
 558                                 goto retry;
 559                         }
 560                         err = SCSI_DH_RETRY;
 561                 } else {
 562                         err = SCSI_DH_OK;
 563                 }
 564
 565                 /* Transitioning time exceeded, set port to standby */
 566                 h->state = TPGS_STATE_STANDBY;
 567                 break;
 568         case TPGS_STATE_OFFLINE:
 569                 /* Path unusable */
 570                 err = SCSI_DH_DEV_OFFLINED;
 571                 break;
 572         default:
 573                 /* Useable path if active */
 574                 err = SCSI_DH_OK;
 575                 break;
 576         }
 577         return err;
 578 }
 579
 580 /*
 581  * alua_initialize - Initialize ALUA state
 582  * @sdev: the device to be initialized
 583  *
 584  * For the prep_fn to work correctly we have
 585  * to initialize the ALUA state for the device.
 586  */
 587 static int alua_initialize(struct scsi_device *sdev, struct alua_dh_data *h)
 588 {
 589         int err = SCSI_DH_DEV_UNSUPP;
 590
 591         h->tpgs = alua_check_tpgs(sdev);
 592         if (h->tpgs == TPGS_MODE_NONE)
 593                 goto out;
 594
 595         err = alua_check_vpd(sdev, h);
 596         if (err != SCSI_DH_OK)
 597                 goto out;
 598
 599         err = alua_rtpg(sdev, h, 0);
 600         if (err != SCSI_DH_OK)
 601                 goto out;
 602
 603 out:
 604         return err;
 605 }
 606 /*
 607  * alua_set_params - set/unset the optimize flag
 608  * @sdev: device on the path to be activated
 609  * params - parameters in the following format
 610  *      "no_of_params\0param1\0param2\0param3\0...\0"
 611  * For example, to set the flag pass the following parameters
 612  * from multipath.conf
 613  *     hardware_handler        "2 alua 1"
 614  */
 615 static int alua_set_params(struct scsi_device *sdev, const char *params)
 616 {
 617         struct alua_dh_data *h = sdev->handler_data;
 618         unsigned int optimize = 0, argc;
 619         const char *p = params;
 620         int result = SCSI_DH_OK;
 621
 622         if ((sscanf(params, "%u", &argc) != 1) || (argc != 1))
 623                 return -EINVAL;
 624
 625         while (*p++)
 626                 ;
 627         if ((sscanf(p, "%u", &optimize) != 1) || (optimize > 1))
 628                 return -EINVAL;
 629
 630         if (optimize)
 631                 h->flags |= ALUA_OPTIMIZE_STPG;
 632         else
 633                 h->flags &= ~ALUA_OPTIMIZE_STPG;
 634
 635         return result;
 636 }
 637
 638 static uint optimize_stpg;
 639 module_param(optimize_stpg, uint, S_IRUGO|S_IWUSR);
 640 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.");
 641
 642 /*
 643  * alua_activate - activate a path
 644  * @sdev: device on the path to be activated
 645  *
 646  * We're currently switching the port group to be activated only and
 647  * let the array figure out the rest.
 648  * There may be other arrays which require us to switch all port groups
 649  * based on a certain policy. But until we actually encounter them it
 650  * should be okay.
 651  */
 652 static int alua_activate(struct scsi_device *sdev,
 653                         activate_complete fn, void *data)
 654 {
 655         struct alua_dh_data *h = sdev->handler_data;
 656         int err = SCSI_DH_OK;
 657         int stpg = 0;
 658
 659         err = alua_rtpg(sdev, h, 1);
 660         if (err != SCSI_DH_OK)
 661                 goto out;
 662
 663         if (optimize_stpg)
 664                 h->flags |= ALUA_OPTIMIZE_STPG;
 665
 666         if (h->tpgs & TPGS_MODE_EXPLICIT) {
 667                 switch (h->state) {
 668                 case TPGS_STATE_NONOPTIMIZED:
 669                         stpg = 1;
 670                         if ((h->flags & ALUA_OPTIMIZE_STPG) &&
 671                             (!h->pref) &&
 672                             (h->tpgs & TPGS_MODE_IMPLICIT))
 673                                 stpg = 0;
 674                         break;
 675                 case TPGS_STATE_STANDBY:
 676                 case TPGS_STATE_UNAVAILABLE:
 677                         stpg = 1;
 678                         break;
 679                 case TPGS_STATE_OFFLINE:
 680                         err = SCSI_DH_IO;
 681                         break;
 682                 case TPGS_STATE_TRANSITIONING:
 683                         err = SCSI_DH_RETRY;
 684                         break;
 685                 default:
 686                         break;
 687                 }
 688         }
 689
 690         if (stpg) {
 691                 h->callback_fn = fn;
 692                 h->callback_data = data;
 693                 err = submit_stpg(h);
 694                 if (err == SCSI_DH_OK)
 695                         return 0;
 696                 h->callback_fn = h->callback_data = NULL;
 697         }
 698
 699 out:
 700         if (fn)
 701                 fn(data, err);
 702         return 0;
 703 }
 704
 705 /*
 706  * alua_prep_fn - request callback
 707  *
 708  * Fail I/O to all paths not in state
 709  * active/optimized or active/non-optimized.
 710  */
 711 static int alua_prep_fn(struct scsi_device *sdev, struct request *req)
 712 {
 713         struct alua_dh_data *h = sdev->handler_data;
 714         int ret = BLKPREP_OK;
 715
 716         if (h->state == TPGS_STATE_TRANSITIONING)
 717                 ret = BLKPREP_DEFER;
 718         else if (h->state != TPGS_STATE_OPTIMIZED &&
 719                  h->state != TPGS_STATE_NONOPTIMIZED &&
 720                  h->state != TPGS_STATE_LBA_DEPENDENT) {
 721                 ret = BLKPREP_KILL;
 722                 req->cmd_flags |= REQ_QUIET;
 723         }
 724         return ret;
 725
 726 }
 727
 728 /*
 729  * alua_bus_attach - Attach device handler
 730  * @sdev: device to be attached to
 731  */
 732 static int alua_bus_attach(struct scsi_device *sdev)
 733 {
 734         struct alua_dh_data *h;
 735         int err;
 736
 737         h = kzalloc(sizeof(*h) , GFP_KERNEL);
 738         if (!h)
 739                 return -ENOMEM;
 740         h->tpgs = TPGS_MODE_UNINITIALIZED;
 741         h->state = TPGS_STATE_OPTIMIZED;
 742         h->group_id = -1;
 743         h->rel_port = -1;
 744         h->buff = h->inq;
 745         h->bufflen = ALUA_INQUIRY_SIZE;
 746         h->sdev = sdev;
 747
 748         err = alua_initialize(sdev, h);
 749         if (err != SCSI_DH_OK && err != SCSI_DH_DEV_OFFLINED)
 750                 goto failed;
 751
 752         sdev->handler_data = h;
 753         return 0;
 754 failed:
 755         kfree(h);
 756         return -EINVAL;
 757 }
 758
 759 /*
 760  * alua_bus_detach - Detach device handler
 761  * @sdev: device to be detached from
 762  */
 763 static void alua_bus_detach(struct scsi_device *sdev)
 764 {
 765         struct alua_dh_data *h = sdev->handler_data;
 766
 767         if (h->buff && h->inq != h->buff)
 768                 kfree(h->buff);
 769         sdev->handler_data = NULL;
 770         kfree(h);
 771 }
 772
 773 static struct scsi_device_handler alua_dh = {
 774         .name = ALUA_DH_NAME,
 775         .module = THIS_MODULE,
 776         .attach = alua_bus_attach,
 777         .detach = alua_bus_detach,
 778         .prep_fn = alua_prep_fn,
 779         .check_sense = alua_check_sense,
 780         .activate = alua_activate,
 781         .set_params = alua_set_params,
 782 };
 783
 784 static int __init alua_init(void)
 785 {
 786         int r;
 787
 788         r = scsi_register_device_handler(&alua_dh);
 789         if (r != 0)
 790                 printk(KERN_ERR "%s: Failed to register scsi device handler",
 791                         ALUA_DH_NAME);
 792         return r;
 793 }
 794
 795 static void __exit alua_exit(void)
 796 {
 797         scsi_unregister_device_handler(&alua_dh);
 798 }
 799
 800 module_init(alua_init);
 801 module_exit(alua_exit);
 802
 803 MODULE_DESCRIPTION("DM Multipath ALUA support");
 804 MODULE_AUTHOR("Hannes Reinecke <hare@suse.de>");
 805 MODULE_LICENSE("GPL");
 806 MODULE_VERSION(ALUA_DH_VER);