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                 if (sdev->allow_restart &&
 478                     sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x02)
 479                         /*
 480                          * if the device is not started, we need to wake
 481                          * the error handler to start the motor
 482                          */
 483                         return FAILED;
 484                 break;
 485         case UNIT_ATTENTION:
 486                 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x00)
 487                         /*
 488                          * Power On, Reset, or Bus Device Reset, just retry.
 489                          */
 490                         return ADD_TO_MLQUEUE;
 491                 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x04)
 492                         /*
 493                          * Device internal reset
 494                          */
 495                         return ADD_TO_MLQUEUE;
 496                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x01)
 497                         /*
 498                          * Mode Parameters Changed
 499                          */
 500                         return ADD_TO_MLQUEUE;
 501                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x06)
 502                         /*
 503                          * ALUA state changed
 504                          */
 505                         return ADD_TO_MLQUEUE;
 506                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x07)
 507                         /*
 508                          * Implicit ALUA state transition failed
 509                          */
 510                         return ADD_TO_MLQUEUE;
 511                 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x03)
 512                         /*
 513                          * Inquiry data has changed
 514                          */
 515                         return ADD_TO_MLQUEUE;
 516                 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x0e)
 517                         /*
 518                          * REPORTED_LUNS_DATA_HAS_CHANGED is reported
 519                          * when switching controllers on targets like
 520                          * Intel Multi-Flex. We can just retry.
 521                          */
 522                         return ADD_TO_MLQUEUE;
 523                 break;
 524         }
 525
 526         return SCSI_RETURN_NOT_HANDLED;
 527 }
 528
 529 /*
 530  * alua_rtpg - Evaluate REPORT TARGET GROUP STATES
 531  * @sdev: the device to be evaluated.
 532  * @wait_for_transition: if nonzero, wait ALUA_FAILOVER_TIMEOUT seconds for device to exit transitioning state
 533  *
 534  * Evaluate the Target Port Group State.
 535  * Returns SCSI_DH_DEV_OFFLINED if the path is
 536  * found to be unusable.
 537  */
 538 static int alua_rtpg(struct scsi_device *sdev, struct alua_dh_data *h, int wait_for_transition)
 539 {
 540         struct scsi_sense_hdr sense_hdr;
 541         int len, k, off, valid_states = 0;
 542         unsigned char *ucp;
 543         unsigned err;
 544         bool rtpg_ext_hdr_req = 1;
 545         unsigned long expiry, interval = 0;
 546         unsigned int tpg_desc_tbl_off;
 547         unsigned char orig_transition_tmo;
 548
 549         if (!h->transition_tmo)
 550                 expiry = round_jiffies_up(jiffies + ALUA_FAILOVER_TIMEOUT * HZ);
 551         else
 552                 expiry = round_jiffies_up(jiffies + h->transition_tmo * HZ);
 553
 554  retry:
 555         err = submit_rtpg(sdev, h, rtpg_ext_hdr_req);
 556
 557         if (err == SCSI_DH_IO && h->senselen > 0) {
 558                 err = scsi_normalize_sense(h->sense, SCSI_SENSE_BUFFERSIZE,
 559                                            &sense_hdr);
 560                 if (!err)
 561                         return SCSI_DH_IO;
 562
 563                 /*
 564                  * submit_rtpg() has failed on existing arrays
 565                  * when requesting extended header info, and
 566                  * the array doesn't support extended headers,
 567                  * even though it shouldn't according to T10.
 568                  * The retry without rtpg_ext_hdr_req set
 569                  * handles this.
 570                  */
 571                 if (rtpg_ext_hdr_req == 1 &&
 572                     sense_hdr.sense_key == ILLEGAL_REQUEST &&
 573                     sense_hdr.asc == 0x24 && sense_hdr.ascq == 0) {
 574                         rtpg_ext_hdr_req = 0;
 575                         goto retry;
 576                 }
 577
 578                 err = alua_check_sense(sdev, &sense_hdr);
 579                 if (err == ADD_TO_MLQUEUE && time_before(jiffies, expiry))
 580                         goto retry;
 581                 sdev_printk(KERN_INFO, sdev,
 582                             "%s: rtpg sense code %02x/%02x/%02x\n",
 583                             ALUA_DH_NAME, sense_hdr.sense_key,
 584                             sense_hdr.asc, sense_hdr.ascq);
 585                 err = SCSI_DH_IO;
 586         }
 587         if (err != SCSI_DH_OK)
 588                 return err;
 589
 590         len = (h->buff[0] << 24) + (h->buff[1] << 16) +
 591                 (h->buff[2] << 8) + h->buff[3] + 4;
 592
 593         if (len > h->bufflen) {
 594                 /* Resubmit with the correct length */
 595                 if (realloc_buffer(h, len)) {
 596                         sdev_printk(KERN_WARNING, sdev,
 597                                     "%s: kmalloc buffer failed\n",__func__);
 598                         /* Temporary failure, bypass */
 599                         return SCSI_DH_DEV_TEMP_BUSY;
 600                 }
 601                 goto retry;
 602         }
 603
 604         orig_transition_tmo = h->transition_tmo;
 605         if ((h->buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR && h->buff[5] != 0)
 606                 h->transition_tmo = h->buff[5];
 607         else
 608                 h->transition_tmo = ALUA_FAILOVER_TIMEOUT;
 609
 610         if (wait_for_transition && (orig_transition_tmo != h->transition_tmo)) {
 611                 sdev_printk(KERN_INFO, sdev,
 612                             "%s: transition timeout set to %d seconds\n",
 613                             ALUA_DH_NAME, h->transition_tmo);
 614                 expiry = jiffies + h->transition_tmo * HZ;
 615         }
 616
 617         if ((h->buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR)
 618                 tpg_desc_tbl_off = 8;
 619         else
 620                 tpg_desc_tbl_off = 4;
 621
 622         for (k = tpg_desc_tbl_off, ucp = h->buff + tpg_desc_tbl_off;
 623              k < len;
 624              k += off, ucp += off) {
 625
 626                 if (h->group_id == (ucp[2] << 8) + ucp[3]) {
 627                         h->state = ucp[0] & 0x0f;
 628                         h->pref = ucp[0] >> 7;
 629                         valid_states = ucp[1];
 630                 }
 631                 off = 8 + (ucp[7] * 4);
 632         }
 633
 634         sdev_printk(KERN_INFO, sdev,
 635                     "%s: port group %02x state %c %s supports %c%c%c%c%c%c%c\n",
 636                     ALUA_DH_NAME, h->group_id, print_alua_state(h->state),
 637                     h->pref ? "preferred" : "non-preferred",
 638                     valid_states&TPGS_SUPPORT_TRANSITION?'T':'t',
 639                     valid_states&TPGS_SUPPORT_OFFLINE?'O':'o',
 640                     valid_states&TPGS_SUPPORT_LBA_DEPENDENT?'L':'l',
 641                     valid_states&TPGS_SUPPORT_UNAVAILABLE?'U':'u',
 642                     valid_states&TPGS_SUPPORT_STANDBY?'S':'s',
 643                     valid_states&TPGS_SUPPORT_NONOPTIMIZED?'N':'n',
 644                     valid_states&TPGS_SUPPORT_OPTIMIZED?'A':'a');
 645
 646         switch (h->state) {
 647         case TPGS_STATE_TRANSITIONING:
 648                 if (wait_for_transition) {
 649                         if (time_before(jiffies, expiry)) {
 650                                 /* State transition, retry */
 651                                 interval += 2000;
 652                                 msleep(interval);
 653                                 goto retry;
 654                         }
 655                         err = SCSI_DH_RETRY;
 656                 } else {
 657                         err = SCSI_DH_OK;
 658                 }
 659
 660                 /* Transitioning time exceeded, set port to standby */
 661                 h->state = TPGS_STATE_STANDBY;
 662                 break;
 663         case TPGS_STATE_OFFLINE:
 664                 /* Path unusable */
 665                 err = SCSI_DH_DEV_OFFLINED;
 666                 break;
 667         default:
 668                 /* Useable path if active */
 669                 err = SCSI_DH_OK;
 670                 break;
 671         }
 672         return err;
 673 }
 674
 675 /*
 676  * alua_initialize - Initialize ALUA state
 677  * @sdev: the device to be initialized
 678  *
 679  * For the prep_fn to work correctly we have
 680  * to initialize the ALUA state for the device.
 681  */
 682 static int alua_initialize(struct scsi_device *sdev, struct alua_dh_data *h)
 683 {
 684         int err;
 685
 686         err = alua_check_tpgs(sdev, h);
 687         if (err != SCSI_DH_OK)
 688                 goto out;
 689
 690         err = alua_vpd_inquiry(sdev, h);
 691         if (err != SCSI_DH_OK)
 692                 goto out;
 693
 694         err = alua_rtpg(sdev, h, 0);
 695         if (err != SCSI_DH_OK)
 696                 goto out;
 697
 698 out:
 699         return err;
 700 }
 701 /*
 702  * alua_set_params - set/unset the optimize flag
 703  * @sdev: device on the path to be activated
 704  * params - parameters in the following format
 705  *      "no_of_params\0param1\0param2\0param3\0...\0"
 706  * For example, to set the flag pass the following parameters
 707  * from multipath.conf
 708  *     hardware_handler        "2 alua 1"
 709  */
 710 static int alua_set_params(struct scsi_device *sdev, const char *params)
 711 {
 712         struct alua_dh_data *h = get_alua_data(sdev);
 713         unsigned int optimize = 0, argc;
 714         const char *p = params;
 715         int result = SCSI_DH_OK;
 716
 717         if ((sscanf(params, "%u", &argc) != 1) || (argc != 1))
 718                 return -EINVAL;
 719
 720         while (*p++)
 721                 ;
 722         if ((sscanf(p, "%u", &optimize) != 1) || (optimize > 1))
 723                 return -EINVAL;
 724
 725         if (optimize)
 726                 h->flags |= ALUA_OPTIMIZE_STPG;
 727         else
 728                 h->flags &= ~ALUA_OPTIMIZE_STPG;
 729
 730         return result;
 731 }
 732
 733 static uint optimize_stpg;
 734 module_param(optimize_stpg, uint, S_IRUGO|S_IWUSR);
 735 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.");
 736
 737 /*
 738  * alua_activate - activate a path
 739  * @sdev: device on the path to be activated
 740  *
 741  * We're currently switching the port group to be activated only and
 742  * let the array figure out the rest.
 743  * There may be other arrays which require us to switch all port groups
 744  * based on a certain policy. But until we actually encounter them it
 745  * should be okay.
 746  */
 747 static int alua_activate(struct scsi_device *sdev,
 748                         activate_complete fn, void *data)
 749 {
 750         struct alua_dh_data *h = get_alua_data(sdev);
 751         int err = SCSI_DH_OK;
 752         int stpg = 0;
 753
 754         err = alua_rtpg(sdev, h, 1);
 755         if (err != SCSI_DH_OK)
 756                 goto out;
 757
 758         if (optimize_stpg)
 759                 h->flags |= ALUA_OPTIMIZE_STPG;
 760
 761         if (h->tpgs & TPGS_MODE_EXPLICIT) {
 762                 switch (h->state) {
 763                 case TPGS_STATE_NONOPTIMIZED:
 764                         stpg = 1;
 765                         if ((h->flags & ALUA_OPTIMIZE_STPG) &&
 766                             (!h->pref) &&
 767                             (h->tpgs & TPGS_MODE_IMPLICIT))
 768                                 stpg = 0;
 769                         break;
 770                 case TPGS_STATE_STANDBY:
 771                 case TPGS_STATE_UNAVAILABLE:
 772                         stpg = 1;
 773                         break;
 774                 case TPGS_STATE_OFFLINE:
 775                         err = SCSI_DH_IO;
 776                         break;
 777                 case TPGS_STATE_TRANSITIONING:
 778                         err = SCSI_DH_RETRY;
 779                         break;
 780                 default:
 781                         break;
 782                 }
 783         }
 784
 785         if (stpg) {
 786                 h->callback_fn = fn;
 787                 h->callback_data = data;
 788                 err = submit_stpg(h);
 789                 if (err == SCSI_DH_OK)
 790                         return 0;
 791                 h->callback_fn = h->callback_data = NULL;
 792         }
 793
 794 out:
 795         if (fn)
 796                 fn(data, err);
 797         return 0;
 798 }
 799
 800 /*
 801  * alua_prep_fn - request callback
 802  *
 803  * Fail I/O to all paths not in state
 804  * active/optimized or active/non-optimized.
 805  */
 806 static int alua_prep_fn(struct scsi_device *sdev, struct request *req)
 807 {
 808         struct alua_dh_data *h = get_alua_data(sdev);
 809         int ret = BLKPREP_OK;
 810
 811         if (h->state == TPGS_STATE_TRANSITIONING)
 812                 ret = BLKPREP_DEFER;
 813         else if (h->state != TPGS_STATE_OPTIMIZED &&
 814                  h->state != TPGS_STATE_NONOPTIMIZED &&
 815                  h->state != TPGS_STATE_LBA_DEPENDENT) {
 816                 ret = BLKPREP_KILL;
 817                 req->cmd_flags |= REQ_QUIET;
 818         }
 819         return ret;
 820
 821 }
 822
 823 static bool alua_match(struct scsi_device *sdev)
 824 {
 825         return (scsi_device_tpgs(sdev) != 0);
 826 }
 827
 828 static int alua_bus_attach(struct scsi_device *sdev);
 829 static void alua_bus_detach(struct scsi_device *sdev);
 830
 831 static struct scsi_device_handler alua_dh = {
 832         .name = ALUA_DH_NAME,
 833         .module = THIS_MODULE,
 834         .attach = alua_bus_attach,
 835         .detach = alua_bus_detach,
 836         .prep_fn = alua_prep_fn,
 837         .check_sense = alua_check_sense,
 838         .activate = alua_activate,
 839         .set_params = alua_set_params,
 840         .match = alua_match,
 841 };
 842
 843 /*
 844  * alua_bus_attach - Attach device handler
 845  * @sdev: device to be attached to
 846  */
 847 static int alua_bus_attach(struct scsi_device *sdev)
 848 {
 849         struct scsi_dh_data *scsi_dh_data;
 850         struct alua_dh_data *h;
 851         unsigned long flags;
 852         int err = SCSI_DH_OK;
 853
 854         scsi_dh_data = kzalloc(sizeof(*scsi_dh_data)
 855                                + sizeof(*h) , GFP_KERNEL);
 856         if (!scsi_dh_data) {
 857                 sdev_printk(KERN_ERR, sdev, "%s: Attach failed\n",
 858                             ALUA_DH_NAME);
 859                 return -ENOMEM;
 860         }
 861
 862         scsi_dh_data->scsi_dh = &alua_dh;
 863         h = (struct alua_dh_data *) scsi_dh_data->buf;
 864         h->tpgs = TPGS_MODE_UNINITIALIZED;
 865         h->state = TPGS_STATE_OPTIMIZED;
 866         h->group_id = -1;
 867         h->rel_port = -1;
 868         h->buff = h->inq;
 869         h->bufflen = ALUA_INQUIRY_SIZE;
 870         h->sdev = sdev;
 871
 872         err = alua_initialize(sdev, h);
 873         if ((err != SCSI_DH_OK) && (err != SCSI_DH_DEV_OFFLINED))
 874                 goto failed;
 875
 876         if (!try_module_get(THIS_MODULE))
 877                 goto failed;
 878
 879         spin_lock_irqsave(sdev->request_queue->queue_lock, flags);
 880         sdev->scsi_dh_data = scsi_dh_data;
 881         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
 882         sdev_printk(KERN_NOTICE, sdev, "%s: Attached\n", ALUA_DH_NAME);
 883
 884         return 0;
 885
 886 failed:
 887         kfree(scsi_dh_data);
 888         sdev_printk(KERN_ERR, sdev, "%s: not attached\n", ALUA_DH_NAME);
 889         return -EINVAL;
 890 }
 891
 892 /*
 893  * alua_bus_detach - Detach device handler
 894  * @sdev: device to be detached from
 895  */
 896 static void alua_bus_detach(struct scsi_device *sdev)
 897 {
 898         struct scsi_dh_data *scsi_dh_data;
 899         struct alua_dh_data *h;
 900         unsigned long flags;
 901
 902         spin_lock_irqsave(sdev->request_queue->queue_lock, flags);
 903         scsi_dh_data = sdev->scsi_dh_data;
 904         sdev->scsi_dh_data = NULL;
 905         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
 906
 907         h = (struct alua_dh_data *) scsi_dh_data->buf;
 908         if (h->buff && h->inq != h->buff)
 909                 kfree(h->buff);
 910         kfree(scsi_dh_data);
 911         module_put(THIS_MODULE);
 912         sdev_printk(KERN_NOTICE, sdev, "%s: Detached\n", ALUA_DH_NAME);
 913 }
 914
 915 static int __init alua_init(void)
 916 {
 917         int r;
 918
 919         r = scsi_register_device_handler(&alua_dh);
 920         if (r != 0)
 921                 printk(KERN_ERR "%s: Failed to register scsi device handler",
 922                         ALUA_DH_NAME);
 923         return r;
 924 }
 925
 926 static void __exit alua_exit(void)
 927 {
 928         scsi_unregister_device_handler(&alua_dh);
 929 }
 930
 931 module_init(alua_init);
 932 module_exit(alua_exit);
 933
 934 MODULE_DESCRIPTION("DM Multipath ALUA support");
 935 MODULE_AUTHOR("Hannes Reinecke <hare@suse.de>");
 936 MODULE_LICENSE("GPL");
 937 MODULE_VERSION(ALUA_DH_VER);