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