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 <scsi/scsi.h>
  26 #include <scsi/scsi_eh.h>
  27 #include <scsi/scsi_dh.h>
  28
  29 #define ALUA_DH_NAME "alua"
  30 #define ALUA_DH_VER "1.3"
  31
  32 #define TPGS_STATE_OPTIMIZED            0x0
  33 #define TPGS_STATE_NONOPTIMIZED         0x1
  34 #define TPGS_STATE_STANDBY              0x2
  35 #define TPGS_STATE_UNAVAILABLE          0x3
  36 #define TPGS_STATE_LBA_DEPENDENT        0x4
  37 #define TPGS_STATE_OFFLINE              0xe
  38 #define TPGS_STATE_TRANSITIONING        0xf
  39
  40 #define TPGS_SUPPORT_NONE               0x00
  41 #define TPGS_SUPPORT_OPTIMIZED          0x01
  42 #define TPGS_SUPPORT_NONOPTIMIZED       0x02
  43 #define TPGS_SUPPORT_STANDBY            0x04
  44 #define TPGS_SUPPORT_UNAVAILABLE        0x08
  45 #define TPGS_SUPPORT_LBA_DEPENDENT      0x10
  46 #define TPGS_SUPPORT_OFFLINE            0x40
  47 #define TPGS_SUPPORT_TRANSITION         0x80
  48
  49 #define RTPG_FMT_MASK                   0x70
  50 #define RTPG_FMT_EXT_HDR                0x10
  51
  52 #define TPGS_MODE_UNINITIALIZED          -1
  53 #define TPGS_MODE_NONE                  0x0
  54 #define TPGS_MODE_IMPLICIT              0x1
  55 #define TPGS_MODE_EXPLICIT              0x2
  56
  57 #define ALUA_INQUIRY_SIZE               36
  58 #define ALUA_FAILOVER_TIMEOUT           60
  59 #define ALUA_FAILOVER_RETRIES           5
  60
  61 /* flags passed from user level */
  62 #define ALUA_OPTIMIZE_STPG              1
  63
  64 struct alua_dh_data {
  65         int                     group_id;
  66         int                     rel_port;
  67         int                     tpgs;
  68         int                     state;
  69         int                     pref;
  70         unsigned                flags; /* used for optimizing STPG */
  71         unsigned char           inq[ALUA_INQUIRY_SIZE];
  72         unsigned char           *buff;
  73         int                     bufflen;
  74         unsigned char           transition_tmo;
  75         unsigned char           sense[SCSI_SENSE_BUFFERSIZE];
  76         int                     senselen;
  77         struct scsi_device      *sdev;
  78         activate_complete       callback_fn;
  79         void                    *callback_data;
  80 };
  81
  82 #define ALUA_POLICY_SWITCH_CURRENT      0
  83 #define ALUA_POLICY_SWITCH_ALL          1
  84
  85 static char print_alua_state(int);
  86 static int alua_check_sense(struct scsi_device *, struct scsi_sense_hdr *);
  87
  88 static inline struct alua_dh_data *get_alua_data(struct scsi_device *sdev)
  89 {
  90         struct scsi_dh_data *scsi_dh_data = sdev->scsi_dh_data;
  91         BUG_ON(scsi_dh_data == NULL);
  92         return ((struct alua_dh_data *) scsi_dh_data->buf);
  93 }
  94
  95 static int realloc_buffer(struct alua_dh_data *h, unsigned len)
  96 {
  97         if (h->buff && h->buff != h->inq)
  98                 kfree(h->buff);
  99
 100         h->buff = kmalloc(len, GFP_NOIO);
 101         if (!h->buff) {
 102                 h->buff = h->inq;
 103                 h->bufflen = ALUA_INQUIRY_SIZE;
 104                 return 1;
 105         }
 106         h->bufflen = len;
 107         return 0;
 108 }
 109
 110 static struct request *get_alua_req(struct scsi_device *sdev,
 111                                     void *buffer, unsigned buflen, int rw)
 112 {
 113         struct request *rq;
 114         struct request_queue *q = sdev->request_queue;
 115
 116         rq = blk_get_request(q, rw, GFP_NOIO);
 117
 118         if (IS_ERR(rq)) {
 119                 sdev_printk(KERN_INFO, sdev,
 120                             "%s: blk_get_request failed\n", __func__);
 121                 return NULL;
 122         }
 123         blk_rq_set_block_pc(rq);
 124
 125         if (buflen && blk_rq_map_kern(q, rq, buffer, buflen, GFP_NOIO)) {
 126                 blk_put_request(rq);
 127                 sdev_printk(KERN_INFO, sdev,
 128                             "%s: blk_rq_map_kern failed\n", __func__);
 129                 return NULL;
 130         }
 131
 132         rq->cmd_flags |= REQ_FAILFAST_DEV | REQ_FAILFAST_TRANSPORT |
 133                          REQ_FAILFAST_DRIVER;
 134         rq->retries = ALUA_FAILOVER_RETRIES;
 135         rq->timeout = ALUA_FAILOVER_TIMEOUT * HZ;
 136
 137         return rq;
 138 }
 139
 140 /*
 141  * submit_vpd_inquiry - Issue an INQUIRY VPD page 0x83 command
 142  * @sdev: sdev the command should be sent to
 143  */
 144 static int submit_vpd_inquiry(struct scsi_device *sdev, struct alua_dh_data *h)
 145 {
 146         struct request *rq;
 147         int err = SCSI_DH_RES_TEMP_UNAVAIL;
 148
 149         rq = get_alua_req(sdev, h->buff, h->bufflen, READ);
 150         if (!rq)
 151                 goto done;
 152
 153         /* Prepare the command. */
 154         rq->cmd[0] = INQUIRY;
 155         rq->cmd[1] = 1;
 156         rq->cmd[2] = 0x83;
 157         rq->cmd[4] = h->bufflen;
 158         rq->cmd_len = COMMAND_SIZE(INQUIRY);
 159
 160         rq->sense = h->sense;
 161         memset(rq->sense, 0, SCSI_SENSE_BUFFERSIZE);
 162         rq->sense_len = h->senselen = 0;
 163
 164         err = blk_execute_rq(rq->q, NULL, rq, 1);
 165         if (err == -EIO) {
 166                 sdev_printk(KERN_INFO, sdev,
 167                             "%s: evpd inquiry failed with %x\n",
 168                             ALUA_DH_NAME, rq->errors);
 169                 h->senselen = rq->sense_len;
 170                 err = SCSI_DH_IO;
 171         }
 172         blk_put_request(rq);
 173 done:
 174         return err;
 175 }
 176
 177 /*
 178  * submit_rtpg - Issue a REPORT TARGET GROUP STATES command
 179  * @sdev: sdev the command should be sent to
 180  */
 181 static unsigned submit_rtpg(struct scsi_device *sdev, struct alua_dh_data *h,
 182                             bool rtpg_ext_hdr_req)
 183 {
 184         struct request *rq;
 185         int err = SCSI_DH_RES_TEMP_UNAVAIL;
 186
 187         rq = get_alua_req(sdev, h->buff, h->bufflen, READ);
 188         if (!rq)
 189                 goto done;
 190
 191         /* Prepare the command. */
 192         rq->cmd[0] = MAINTENANCE_IN;
 193         if (rtpg_ext_hdr_req)
 194                 rq->cmd[1] = MI_REPORT_TARGET_PGS | MI_EXT_HDR_PARAM_FMT;
 195         else
 196                 rq->cmd[1] = MI_REPORT_TARGET_PGS;
 197         rq->cmd[6] = (h->bufflen >> 24) & 0xff;
 198         rq->cmd[7] = (h->bufflen >> 16) & 0xff;
 199         rq->cmd[8] = (h->bufflen >>  8) & 0xff;
 200         rq->cmd[9] = h->bufflen & 0xff;
 201         rq->cmd_len = COMMAND_SIZE(MAINTENANCE_IN);
 202
 203         rq->sense = h->sense;
 204         memset(rq->sense, 0, SCSI_SENSE_BUFFERSIZE);
 205         rq->sense_len = h->senselen = 0;
 206
 207         err = blk_execute_rq(rq->q, NULL, rq, 1);
 208         if (err == -EIO) {
 209                 sdev_printk(KERN_INFO, sdev,
 210                             "%s: rtpg failed with %x\n",
 211                             ALUA_DH_NAME, rq->errors);
 212                 h->senselen = rq->sense_len;
 213                 err = SCSI_DH_IO;
 214         }
 215         blk_put_request(rq);
 216 done:
 217         return err;
 218 }
 219
 220 /*
 221  * alua_stpg - Evaluate SET TARGET GROUP STATES
 222  * @sdev: the device to be evaluated
 223  * @state: the new target group state
 224  *
 225  * Send a SET TARGET GROUP STATES command to the device.
 226  * We only have to test here if we should resubmit the command;
 227  * any other error is assumed as a failure.
 228  */
 229 static void stpg_endio(struct request *req, int error)
 230 {
 231         struct alua_dh_data *h = req->end_io_data;
 232         struct scsi_sense_hdr sense_hdr;
 233         unsigned err = SCSI_DH_OK;
 234
 235         if (host_byte(req->errors) != DID_OK ||
 236             msg_byte(req->errors) != COMMAND_COMPLETE) {
 237                 err = SCSI_DH_IO;
 238                 goto done;
 239         }
 240
 241         if (req->sense_len > 0) {
 242                 err = scsi_normalize_sense(h->sense, SCSI_SENSE_BUFFERSIZE,
 243                                            &sense_hdr);
 244                 if (!err) {
 245                         err = SCSI_DH_IO;
 246                         goto done;
 247                 }
 248                 err = alua_check_sense(h->sdev, &sense_hdr);
 249                 if (err == ADD_TO_MLQUEUE) {
 250                         err = SCSI_DH_RETRY;
 251                         goto done;
 252                 }
 253                 sdev_printk(KERN_INFO, h->sdev,
 254                             "%s: stpg sense code: %02x/%02x/%02x\n",
 255                             ALUA_DH_NAME, sense_hdr.sense_key,
 256                             sense_hdr.asc, sense_hdr.ascq);
 257                 err = SCSI_DH_IO;
 258         } else if (error)
 259                 err = SCSI_DH_IO;
 260
 261         if (err == SCSI_DH_OK) {
 262                 h->state = TPGS_STATE_OPTIMIZED;
 263                 sdev_printk(KERN_INFO, h->sdev,
 264                             "%s: port group %02x switched to state %c\n",
 265                             ALUA_DH_NAME, h->group_id,
 266                             print_alua_state(h->state));
 267         }
 268 done:
 269         req->end_io_data = NULL;
 270         __blk_put_request(req->q, req);
 271         if (h->callback_fn) {
 272                 h->callback_fn(h->callback_data, err);
 273                 h->callback_fn = h->callback_data = NULL;
 274         }
 275         return;
 276 }
 277
 278 /*
 279  * submit_stpg - Issue a SET TARGET GROUP STATES command
 280  *
 281  * Currently we're only setting the current target port group state
 282  * to 'active/optimized' and let the array firmware figure out
 283  * the states of the remaining groups.
 284  */
 285 static unsigned submit_stpg(struct alua_dh_data *h)
 286 {
 287         struct request *rq;
 288         int stpg_len = 8;
 289         struct scsi_device *sdev = h->sdev;
 290
 291         /* Prepare the data buffer */
 292         memset(h->buff, 0, stpg_len);
 293         h->buff[4] = TPGS_STATE_OPTIMIZED & 0x0f;
 294         h->buff[6] = (h->group_id >> 8) & 0xff;
 295         h->buff[7] = h->group_id & 0xff;
 296
 297         rq = get_alua_req(sdev, h->buff, stpg_len, WRITE);
 298         if (!rq)
 299                 return SCSI_DH_RES_TEMP_UNAVAIL;
 300
 301         /* Prepare the command. */
 302         rq->cmd[0] = MAINTENANCE_OUT;
 303         rq->cmd[1] = MO_SET_TARGET_PGS;
 304         rq->cmd[6] = (stpg_len >> 24) & 0xff;
 305         rq->cmd[7] = (stpg_len >> 16) & 0xff;
 306         rq->cmd[8] = (stpg_len >>  8) & 0xff;
 307         rq->cmd[9] = stpg_len & 0xff;
 308         rq->cmd_len = COMMAND_SIZE(MAINTENANCE_OUT);
 309
 310         rq->sense = h->sense;
 311         memset(rq->sense, 0, SCSI_SENSE_BUFFERSIZE);
 312         rq->sense_len = h->senselen = 0;
 313         rq->end_io_data = h;
 314
 315         blk_execute_rq_nowait(rq->q, NULL, rq, 1, stpg_endio);
 316         return SCSI_DH_OK;
 317 }
 318
 319 /*
 320  * alua_check_tpgs - Evaluate TPGS setting
 321  * @sdev: device to be checked
 322  *
 323  * Examine the TPGS setting of the sdev to find out if ALUA
 324  * is supported.
 325  */
 326 static int alua_check_tpgs(struct scsi_device *sdev, struct alua_dh_data *h)
 327 {
 328         int err = SCSI_DH_OK;
 329
 330         h->tpgs = scsi_device_tpgs(sdev);
 331         switch (h->tpgs) {
 332         case TPGS_MODE_EXPLICIT|TPGS_MODE_IMPLICIT:
 333                 sdev_printk(KERN_INFO, sdev,
 334                             "%s: supports implicit and explicit TPGS\n",
 335                             ALUA_DH_NAME);
 336                 break;
 337         case TPGS_MODE_EXPLICIT:
 338                 sdev_printk(KERN_INFO, sdev, "%s: supports explicit TPGS\n",
 339                             ALUA_DH_NAME);
 340                 break;
 341         case TPGS_MODE_IMPLICIT:
 342                 sdev_printk(KERN_INFO, sdev, "%s: supports implicit TPGS\n",
 343                             ALUA_DH_NAME);
 344                 break;
 345         default:
 346                 h->tpgs = TPGS_MODE_NONE;
 347                 sdev_printk(KERN_INFO, sdev, "%s: not supported\n",
 348                             ALUA_DH_NAME);
 349                 err = SCSI_DH_DEV_UNSUPP;
 350                 break;
 351         }
 352
 353         return err;
 354 }
 355
 356 /*
 357  * alua_vpd_inquiry - Evaluate INQUIRY vpd page 0x83
 358  * @sdev: device to be checked
 359  *
 360  * Extract the relative target port and the target port group
 361  * descriptor from the list of identificators.
 362  */
 363 static int alua_vpd_inquiry(struct scsi_device *sdev, struct alua_dh_data *h)
 364 {
 365         int len;
 366         unsigned err;
 367         unsigned char *d;
 368
 369  retry:
 370         err = submit_vpd_inquiry(sdev, h);
 371
 372         if (err != SCSI_DH_OK)
 373                 return err;
 374
 375         /* Check if vpd page exceeds initial buffer */
 376         len = (h->buff[2] << 8) + h->buff[3] + 4;
 377         if (len > h->bufflen) {
 378                 /* Resubmit with the correct length */
 379                 if (realloc_buffer(h, len)) {
 380                         sdev_printk(KERN_WARNING, sdev,
 381                                     "%s: kmalloc buffer failed\n",
 382                                     ALUA_DH_NAME);
 383                         /* Temporary failure, bypass */
 384                         return SCSI_DH_DEV_TEMP_BUSY;
 385                 }
 386                 goto retry;
 387         }
 388
 389         /*
 390          * Now look for the correct descriptor.
 391          */
 392         d = h->buff + 4;
 393         while (d < h->buff + len) {
 394                 switch (d[1] & 0xf) {
 395                 case 0x4:
 396                         /* Relative target port */
 397                         h->rel_port = (d[6] << 8) + d[7];
 398                         break;
 399                 case 0x5:
 400                         /* Target port group */
 401                         h->group_id = (d[6] << 8) + d[7];
 402                         break;
 403                 default:
 404                         break;
 405                 }
 406                 d += d[3] + 4;
 407         }
 408
 409         if (h->group_id == -1) {
 410                 /*
 411                  * Internal error; TPGS supported but required
 412                  * VPD identification descriptors not present.
 413                  * Disable ALUA support
 414                  */
 415                 sdev_printk(KERN_INFO, sdev,
 416                             "%s: No target port descriptors found\n",
 417                             ALUA_DH_NAME);
 418                 h->state = TPGS_STATE_OPTIMIZED;
 419                 h->tpgs = TPGS_MODE_NONE;
 420                 err = SCSI_DH_DEV_UNSUPP;
 421         } else {
 422                 sdev_printk(KERN_INFO, sdev,
 423                             "%s: port group %02x rel port %02x\n",
 424                             ALUA_DH_NAME, h->group_id, h->rel_port);
 425         }
 426
 427         return err;
 428 }
 429
 430 static char print_alua_state(int state)
 431 {
 432         switch (state) {
 433         case TPGS_STATE_OPTIMIZED:
 434                 return 'A';
 435         case TPGS_STATE_NONOPTIMIZED:
 436                 return 'N';
 437         case TPGS_STATE_STANDBY:
 438                 return 'S';
 439         case TPGS_STATE_UNAVAILABLE:
 440                 return 'U';
 441         case TPGS_STATE_LBA_DEPENDENT:
 442                 return 'L';
 443         case TPGS_STATE_OFFLINE:
 444                 return 'O';
 445         case TPGS_STATE_TRANSITIONING:
 446                 return 'T';
 447         default:
 448                 return 'X';
 449         }
 450 }
 451
 452 static int alua_check_sense(struct scsi_device *sdev,
 453                             struct scsi_sense_hdr *sense_hdr)
 454 {
 455         switch (sense_hdr->sense_key) {
 456         case NOT_READY:
 457                 if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0a)
 458                         /*
 459                          * LUN Not Accessible - ALUA state transition
 460                          */
 461                         return ADD_TO_MLQUEUE;
 462                 if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0b)
 463                         /*
 464                          * LUN Not Accessible -- Target port in standby state
 465                          */
 466                         return SUCCESS;
 467                 if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0c)
 468                         /*
 469                          * LUN Not Accessible -- Target port in unavailable state
 470                          */
 471                         return SUCCESS;
 472                 if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x12)
 473                         /*
 474                          * LUN Not Ready -- Offline
 475                          */
 476                         return SUCCESS;
 477                 break;
 478         case UNIT_ATTENTION:
 479                 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x00)
 480                         /*
 481                          * Power On, Reset, or Bus Device Reset, just retry.
 482                          */
 483                         return ADD_TO_MLQUEUE;
 484                 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x04)
 485                         /*
 486                          * Device internal reset
 487                          */
 488                         return ADD_TO_MLQUEUE;
 489                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x01)
 490                         /*
 491                          * Mode Parameters Changed
 492                          */
 493                         return ADD_TO_MLQUEUE;
 494                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x06)
 495                         /*
 496                          * ALUA state changed
 497                          */
 498                         return ADD_TO_MLQUEUE;
 499                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x07)
 500                         /*
 501                          * Implicit ALUA state transition failed
 502                          */
 503                         return ADD_TO_MLQUEUE;
 504                 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x03)
 505                         /*
 506                          * Inquiry data has changed
 507                          */
 508                         return ADD_TO_MLQUEUE;
 509                 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x0e)
 510                         /*
 511                          * REPORTED_LUNS_DATA_HAS_CHANGED is reported
 512                          * when switching controllers on targets like
 513                          * Intel Multi-Flex. We can just retry.
 514                          */
 515                         return ADD_TO_MLQUEUE;
 516                 break;
 517         }
 518
 519         return SCSI_RETURN_NOT_HANDLED;
 520 }
 521
 522 /*
 523  * alua_rtpg - Evaluate REPORT TARGET GROUP STATES
 524  * @sdev: the device to be evaluated.
 525  * @wait_for_transition: if nonzero, wait ALUA_FAILOVER_TIMEOUT seconds for device to exit transitioning state
 526  *
 527  * Evaluate the Target Port Group State.
 528  * Returns SCSI_DH_DEV_OFFLINED if the path is
 529  * found to be unusable.
 530  */
 531 static int alua_rtpg(struct scsi_device *sdev, struct alua_dh_data *h, int wait_for_transition)
 532 {
 533         struct scsi_sense_hdr sense_hdr;
 534         int len, k, off, valid_states = 0;
 535         unsigned char *ucp;
 536         unsigned err;
 537         bool rtpg_ext_hdr_req = 1;
 538         unsigned long expiry, interval = 0;
 539         unsigned int tpg_desc_tbl_off;
 540         unsigned char orig_transition_tmo;
 541
 542         if (!h->transition_tmo)
 543                 expiry = round_jiffies_up(jiffies + ALUA_FAILOVER_TIMEOUT * HZ);
 544         else
 545                 expiry = round_jiffies_up(jiffies + h->transition_tmo * HZ);
 546
 547  retry:
 548         err = submit_rtpg(sdev, h, rtpg_ext_hdr_req);
 549
 550         if (err == SCSI_DH_IO && h->senselen > 0) {
 551                 err = scsi_normalize_sense(h->sense, SCSI_SENSE_BUFFERSIZE,
 552                                            &sense_hdr);
 553                 if (!err)
 554                         return SCSI_DH_IO;
 555
 556                 /*
 557                  * submit_rtpg() has failed on existing arrays
 558                  * when requesting extended header info, and
 559                  * the array doesn't support extended headers,
 560                  * even though it shouldn't according to T10.
 561                  * The retry without rtpg_ext_hdr_req set
 562                  * handles this.
 563                  */
 564                 if (rtpg_ext_hdr_req == 1 &&
 565                     sense_hdr.sense_key == ILLEGAL_REQUEST &&
 566                     sense_hdr.asc == 0x24 && sense_hdr.ascq == 0) {
 567                         rtpg_ext_hdr_req = 0;
 568                         goto retry;
 569                 }
 570
 571                 err = alua_check_sense(sdev, &sense_hdr);
 572                 if (err == ADD_TO_MLQUEUE && time_before(jiffies, expiry))
 573                         goto retry;
 574                 sdev_printk(KERN_INFO, sdev,
 575                             "%s: rtpg sense code %02x/%02x/%02x\n",
 576                             ALUA_DH_NAME, sense_hdr.sense_key,
 577                             sense_hdr.asc, sense_hdr.ascq);
 578                 err = SCSI_DH_IO;
 579         }
 580         if (err != SCSI_DH_OK)
 581                 return err;
 582
 583         len = (h->buff[0] << 24) + (h->buff[1] << 16) +
 584                 (h->buff[2] << 8) + h->buff[3] + 4;
 585
 586         if (len > h->bufflen) {
 587                 /* Resubmit with the correct length */
 588                 if (realloc_buffer(h, len)) {
 589                         sdev_printk(KERN_WARNING, sdev,
 590                                     "%s: kmalloc buffer failed\n",__func__);
 591                         /* Temporary failure, bypass */
 592                         return SCSI_DH_DEV_TEMP_BUSY;
 593                 }
 594                 goto retry;
 595         }
 596
 597         orig_transition_tmo = h->transition_tmo;
 598         if ((h->buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR && h->buff[5] != 0)
 599                 h->transition_tmo = h->buff[5];
 600         else
 601                 h->transition_tmo = ALUA_FAILOVER_TIMEOUT;
 602
 603         if (wait_for_transition && (orig_transition_tmo != h->transition_tmo)) {
 604                 sdev_printk(KERN_INFO, sdev,
 605                             "%s: transition timeout set to %d seconds\n",
 606                             ALUA_DH_NAME, h->transition_tmo);
 607                 expiry = jiffies + h->transition_tmo * HZ;
 608         }
 609
 610         if ((h->buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR)
 611                 tpg_desc_tbl_off = 8;
 612         else
 613                 tpg_desc_tbl_off = 4;
 614
 615         for (k = tpg_desc_tbl_off, ucp = h->buff + tpg_desc_tbl_off;
 616              k < len;
 617              k += off, ucp += off) {
 618
 619                 if (h->group_id == (ucp[2] << 8) + ucp[3]) {
 620                         h->state = ucp[0] & 0x0f;
 621                         h->pref = ucp[0] >> 7;
 622                         valid_states = ucp[1];
 623                 }
 624                 off = 8 + (ucp[7] * 4);
 625         }
 626
 627         sdev_printk(KERN_INFO, sdev,
 628                     "%s: port group %02x state %c %s supports %c%c%c%c%c%c%c\n",
 629                     ALUA_DH_NAME, h->group_id, print_alua_state(h->state),
 630                     h->pref ? "preferred" : "non-preferred",
 631                     valid_states&TPGS_SUPPORT_TRANSITION?'T':'t',
 632                     valid_states&TPGS_SUPPORT_OFFLINE?'O':'o',
 633                     valid_states&TPGS_SUPPORT_LBA_DEPENDENT?'L':'l',
 634                     valid_states&TPGS_SUPPORT_UNAVAILABLE?'U':'u',
 635                     valid_states&TPGS_SUPPORT_STANDBY?'S':'s',
 636                     valid_states&TPGS_SUPPORT_NONOPTIMIZED?'N':'n',
 637                     valid_states&TPGS_SUPPORT_OPTIMIZED?'A':'a');
 638
 639         switch (h->state) {
 640         case TPGS_STATE_TRANSITIONING:
 641                 if (wait_for_transition) {
 642                         if (time_before(jiffies, expiry)) {
 643                                 /* State transition, retry */
 644                                 interval += 2000;
 645                                 msleep(interval);
 646                                 goto retry;
 647                         }
 648                         err = SCSI_DH_RETRY;
 649                 } else {
 650                         err = SCSI_DH_OK;
 651                 }
 652
 653                 /* Transitioning time exceeded, set port to standby */
 654                 h->state = TPGS_STATE_STANDBY;
 655                 break;
 656         case TPGS_STATE_OFFLINE:
 657                 /* Path unusable */
 658                 err = SCSI_DH_DEV_OFFLINED;
 659                 break;
 660         default:
 661                 /* Useable path if active */
 662                 err = SCSI_DH_OK;
 663                 break;
 664         }
 665         return err;
 666 }
 667
 668 /*
 669  * alua_initialize - Initialize ALUA state
 670  * @sdev: the device to be initialized
 671  *
 672  * For the prep_fn to work correctly we have
 673  * to initialize the ALUA state for the device.
 674  */
 675 static int alua_initialize(struct scsi_device *sdev, struct alua_dh_data *h)
 676 {
 677         int err;
 678
 679         err = alua_check_tpgs(sdev, h);
 680         if (err != SCSI_DH_OK)
 681                 goto out;
 682
 683         err = alua_vpd_inquiry(sdev, h);
 684         if (err != SCSI_DH_OK)
 685                 goto out;
 686
 687         err = alua_rtpg(sdev, h, 0);
 688         if (err != SCSI_DH_OK)
 689                 goto out;
 690
 691 out:
 692         return err;
 693 }
 694 /*
 695  * alua_set_params - set/unset the optimize flag
 696  * @sdev: device on the path to be activated
 697  * params - parameters in the following format
 698  *      "no_of_params\0param1\0param2\0param3\0...\0"
 699  * For example, to set the flag pass the following parameters
 700  * from multipath.conf
 701  *     hardware_handler        "2 alua 1"
 702  */
 703 static int alua_set_params(struct scsi_device *sdev, const char *params)
 704 {
 705         struct alua_dh_data *h = get_alua_data(sdev);
 706         unsigned int optimize = 0, argc;
 707         const char *p = params;
 708         int result = SCSI_DH_OK;
 709
 710         if ((sscanf(params, "%u", &argc) != 1) || (argc != 1))
 711                 return -EINVAL;
 712
 713         while (*p++)
 714                 ;
 715         if ((sscanf(p, "%u", &optimize) != 1) || (optimize > 1))
 716                 return -EINVAL;
 717
 718         if (optimize)
 719                 h->flags |= ALUA_OPTIMIZE_STPG;
 720         else
 721                 h->flags &= ~ALUA_OPTIMIZE_STPG;
 722
 723         return result;
 724 }
 725
 726 static uint optimize_stpg;
 727 module_param(optimize_stpg, uint, S_IRUGO|S_IWUSR);
 728 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.");
 729
 730 /*
 731  * alua_activate - activate a path
 732  * @sdev: device on the path to be activated
 733  *
 734  * We're currently switching the port group to be activated only and
 735  * let the array figure out the rest.
 736  * There may be other arrays which require us to switch all port groups
 737  * based on a certain policy. But until we actually encounter them it
 738  * should be okay.
 739  */
 740 static int alua_activate(struct scsi_device *sdev,
 741                         activate_complete fn, void *data)
 742 {
 743         struct alua_dh_data *h = get_alua_data(sdev);
 744         int err = SCSI_DH_OK;
 745         int stpg = 0;
 746
 747         err = alua_rtpg(sdev, h, 1);
 748         if (err != SCSI_DH_OK)
 749                 goto out;
 750
 751         if (optimize_stpg)
 752                 h->flags |= ALUA_OPTIMIZE_STPG;
 753
 754         if (h->tpgs & TPGS_MODE_EXPLICIT) {
 755                 switch (h->state) {
 756                 case TPGS_STATE_NONOPTIMIZED:
 757                         stpg = 1;
 758                         if ((h->flags & ALUA_OPTIMIZE_STPG) &&
 759                             (!h->pref) &&
 760                             (h->tpgs & TPGS_MODE_IMPLICIT))
 761                                 stpg = 0;
 762                         break;
 763                 case TPGS_STATE_STANDBY:
 764                 case TPGS_STATE_UNAVAILABLE:
 765                         stpg = 1;
 766                         break;
 767                 case TPGS_STATE_OFFLINE:
 768                         err = SCSI_DH_IO;
 769                         break;
 770                 case TPGS_STATE_TRANSITIONING:
 771                         err = SCSI_DH_RETRY;
 772                         break;
 773                 default:
 774                         break;
 775                 }
 776         }
 777
 778         if (stpg) {
 779                 h->callback_fn = fn;
 780                 h->callback_data = data;
 781                 err = submit_stpg(h);
 782                 if (err == SCSI_DH_OK)
 783                         return 0;
 784                 h->callback_fn = h->callback_data = NULL;
 785         }
 786
 787 out:
 788         if (fn)
 789                 fn(data, err);
 790         return 0;
 791 }
 792
 793 /*
 794  * alua_prep_fn - request callback
 795  *
 796  * Fail I/O to all paths not in state
 797  * active/optimized or active/non-optimized.
 798  */
 799 static int alua_prep_fn(struct scsi_device *sdev, struct request *req)
 800 {
 801         struct alua_dh_data *h = get_alua_data(sdev);
 802         int ret = BLKPREP_OK;
 803
 804         if (h->state == TPGS_STATE_TRANSITIONING)
 805                 ret = BLKPREP_DEFER;
 806         else if (h->state != TPGS_STATE_OPTIMIZED &&
 807                  h->state != TPGS_STATE_NONOPTIMIZED &&
 808                  h->state != TPGS_STATE_LBA_DEPENDENT) {
 809                 ret = BLKPREP_KILL;
 810                 req->cmd_flags |= REQ_QUIET;
 811         }
 812         return ret;
 813
 814 }
 815
 816 static bool alua_match(struct scsi_device *sdev)
 817 {
 818         return (scsi_device_tpgs(sdev) != 0);
 819 }
 820
 821 static int alua_bus_attach(struct scsi_device *sdev);
 822 static void alua_bus_detach(struct scsi_device *sdev);
 823
 824 static struct scsi_device_handler alua_dh = {
 825         .name = ALUA_DH_NAME,
 826         .module = THIS_MODULE,
 827         .attach = alua_bus_attach,
 828         .detach = alua_bus_detach,
 829         .prep_fn = alua_prep_fn,
 830         .check_sense = alua_check_sense,
 831         .activate = alua_activate,
 832         .set_params = alua_set_params,
 833         .match = alua_match,
 834 };
 835
 836 /*
 837  * alua_bus_attach - Attach device handler
 838  * @sdev: device to be attached to
 839  */
 840 static int alua_bus_attach(struct scsi_device *sdev)
 841 {
 842         struct scsi_dh_data *scsi_dh_data;
 843         struct alua_dh_data *h;
 844         unsigned long flags;
 845         int err = SCSI_DH_OK;
 846
 847         scsi_dh_data = kzalloc(sizeof(*scsi_dh_data)
 848                                + sizeof(*h) , GFP_KERNEL);
 849         if (!scsi_dh_data) {
 850                 sdev_printk(KERN_ERR, sdev, "%s: Attach failed\n",
 851                             ALUA_DH_NAME);
 852                 return -ENOMEM;
 853         }
 854
 855         scsi_dh_data->scsi_dh = &alua_dh;
 856         h = (struct alua_dh_data *) scsi_dh_data->buf;
 857         h->tpgs = TPGS_MODE_UNINITIALIZED;
 858         h->state = TPGS_STATE_OPTIMIZED;
 859         h->group_id = -1;
 860         h->rel_port = -1;
 861         h->buff = h->inq;
 862         h->bufflen = ALUA_INQUIRY_SIZE;
 863         h->sdev = sdev;
 864
 865         err = alua_initialize(sdev, h);
 866         if ((err != SCSI_DH_OK) && (err != SCSI_DH_DEV_OFFLINED))
 867                 goto failed;
 868
 869         if (!try_module_get(THIS_MODULE))
 870                 goto failed;
 871
 872         spin_lock_irqsave(sdev->request_queue->queue_lock, flags);
 873         sdev->scsi_dh_data = scsi_dh_data;
 874         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
 875         sdev_printk(KERN_NOTICE, sdev, "%s: Attached\n", ALUA_DH_NAME);
 876
 877         return 0;
 878
 879 failed:
 880         kfree(scsi_dh_data);
 881         sdev_printk(KERN_ERR, sdev, "%s: not attached\n", ALUA_DH_NAME);
 882         return -EINVAL;
 883 }
 884
 885 /*
 886  * alua_bus_detach - Detach device handler
 887  * @sdev: device to be detached from
 888  */
 889 static void alua_bus_detach(struct scsi_device *sdev)
 890 {
 891         struct scsi_dh_data *scsi_dh_data;
 892         struct alua_dh_data *h;
 893         unsigned long flags;
 894
 895         spin_lock_irqsave(sdev->request_queue->queue_lock, flags);
 896         scsi_dh_data = sdev->scsi_dh_data;
 897         sdev->scsi_dh_data = NULL;
 898         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
 899
 900         h = (struct alua_dh_data *) scsi_dh_data->buf;
 901         if (h->buff && h->inq != h->buff)
 902                 kfree(h->buff);
 903         kfree(scsi_dh_data);
 904         module_put(THIS_MODULE);
 905         sdev_printk(KERN_NOTICE, sdev, "%s: Detached\n", ALUA_DH_NAME);
 906 }
 907
 908 static int __init alua_init(void)
 909 {
 910         int r;
 911
 912         r = scsi_register_device_handler(&alua_dh);
 913         if (r != 0)
 914                 printk(KERN_ERR "%s: Failed to register scsi device handler",
 915                         ALUA_DH_NAME);
 916         return r;
 917 }
 918
 919 static void __exit alua_exit(void)
 920 {
 921         scsi_unregister_device_handler(&alua_dh);
 922 }
 923
 924 module_init(alua_init);
 925 module_exit(alua_exit);
 926
 927 MODULE_DESCRIPTION("DM Multipath ALUA support");
 928 MODULE_AUTHOR("Hannes Reinecke <hare@suse.de>");
 929 MODULE_LICENSE("GPL");
 930 MODULE_VERSION(ALUA_DH_VER);