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 (!rq) {
 119                 sdev_printk(KERN_INFO, sdev,
 120                             "%s: blk_get_request failed\n", __func__);
 121                 return NULL;
 122         }
 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_type = REQ_TYPE_BLOCK_PC;
 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 (error || 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 (h->senselen > 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         }
 259         if (err == SCSI_DH_OK) {
 260                 h->state = TPGS_STATE_OPTIMIZED;
 261                 sdev_printk(KERN_INFO, h->sdev,
 262                             "%s: port group %02x switched to state %c\n",
 263                             ALUA_DH_NAME, h->group_id,
 264                             print_alua_state(h->state));
 265         }
 266 done:
 267         req->end_io_data = NULL;
 268         __blk_put_request(req->q, req);
 269         if (h->callback_fn) {
 270                 h->callback_fn(h->callback_data, err);
 271                 h->callback_fn = h->callback_data = NULL;
 272         }
 273         return;
 274 }
 275
 276 /*
 277  * submit_stpg - Issue a SET TARGET GROUP STATES command
 278  *
 279  * Currently we're only setting the current target port group state
 280  * to 'active/optimized' and let the array firmware figure out
 281  * the states of the remaining groups.
 282  */
 283 static unsigned submit_stpg(struct alua_dh_data *h)
 284 {
 285         struct request *rq;
 286         int stpg_len = 8;
 287         struct scsi_device *sdev = h->sdev;
 288
 289         /* Prepare the data buffer */
 290         memset(h->buff, 0, stpg_len);
 291         h->buff[4] = TPGS_STATE_OPTIMIZED & 0x0f;
 292         h->buff[6] = (h->group_id >> 8) & 0xff;
 293         h->buff[7] = h->group_id & 0xff;
 294
 295         rq = get_alua_req(sdev, h->buff, stpg_len, WRITE);
 296         if (!rq)
 297                 return SCSI_DH_RES_TEMP_UNAVAIL;
 298
 299         /* Prepare the command. */
 300         rq->cmd[0] = MAINTENANCE_OUT;
 301         rq->cmd[1] = MO_SET_TARGET_PGS;
 302         rq->cmd[6] = (stpg_len >> 24) & 0xff;
 303         rq->cmd[7] = (stpg_len >> 16) & 0xff;
 304         rq->cmd[8] = (stpg_len >>  8) & 0xff;
 305         rq->cmd[9] = stpg_len & 0xff;
 306         rq->cmd_len = COMMAND_SIZE(MAINTENANCE_OUT);
 307
 308         rq->sense = h->sense;
 309         memset(rq->sense, 0, SCSI_SENSE_BUFFERSIZE);
 310         rq->sense_len = h->senselen = 0;
 311         rq->end_io_data = h;
 312
 313         blk_execute_rq_nowait(rq->q, NULL, rq, 1, stpg_endio);
 314         return SCSI_DH_OK;
 315 }
 316
 317 /*
 318  * alua_check_tpgs - Evaluate TPGS setting
 319  * @sdev: device to be checked
 320  *
 321  * Examine the TPGS setting of the sdev to find out if ALUA
 322  * is supported.
 323  */
 324 static int alua_check_tpgs(struct scsi_device *sdev, struct alua_dh_data *h)
 325 {
 326         int err = SCSI_DH_OK;
 327
 328         h->tpgs = scsi_device_tpgs(sdev);
 329         switch (h->tpgs) {
 330         case TPGS_MODE_EXPLICIT|TPGS_MODE_IMPLICIT:
 331                 sdev_printk(KERN_INFO, sdev,
 332                             "%s: supports implicit and explicit TPGS\n",
 333                             ALUA_DH_NAME);
 334                 break;
 335         case TPGS_MODE_EXPLICIT:
 336                 sdev_printk(KERN_INFO, sdev, "%s: supports explicit TPGS\n",
 337                             ALUA_DH_NAME);
 338                 break;
 339         case TPGS_MODE_IMPLICIT:
 340                 sdev_printk(KERN_INFO, sdev, "%s: supports implicit TPGS\n",
 341                             ALUA_DH_NAME);
 342                 break;
 343         default:
 344                 h->tpgs = TPGS_MODE_NONE;
 345                 sdev_printk(KERN_INFO, sdev, "%s: not supported\n",
 346                             ALUA_DH_NAME);
 347                 err = SCSI_DH_DEV_UNSUPP;
 348                 break;
 349         }
 350
 351         return err;
 352 }
 353
 354 /*
 355  * alua_vpd_inquiry - Evaluate INQUIRY vpd page 0x83
 356  * @sdev: device to be checked
 357  *
 358  * Extract the relative target port and the target port group
 359  * descriptor from the list of identificators.
 360  */
 361 static int alua_vpd_inquiry(struct scsi_device *sdev, struct alua_dh_data *h)
 362 {
 363         int len;
 364         unsigned err;
 365         unsigned char *d;
 366
 367  retry:
 368         err = submit_vpd_inquiry(sdev, h);
 369
 370         if (err != SCSI_DH_OK)
 371                 return err;
 372
 373         /* Check if vpd page exceeds initial buffer */
 374         len = (h->buff[2] << 8) + h->buff[3] + 4;
 375         if (len > h->bufflen) {
 376                 /* Resubmit with the correct length */
 377                 if (realloc_buffer(h, len)) {
 378                         sdev_printk(KERN_WARNING, sdev,
 379                                     "%s: kmalloc buffer failed\n",
 380                                     ALUA_DH_NAME);
 381                         /* Temporary failure, bypass */
 382                         return SCSI_DH_DEV_TEMP_BUSY;
 383                 }
 384                 goto retry;
 385         }
 386
 387         /*
 388          * Now look for the correct descriptor.
 389          */
 390         d = h->buff + 4;
 391         while (d < h->buff + len) {
 392                 switch (d[1] & 0xf) {
 393                 case 0x4:
 394                         /* Relative target port */
 395                         h->rel_port = (d[6] << 8) + d[7];
 396                         break;
 397                 case 0x5:
 398                         /* Target port group */
 399                         h->group_id = (d[6] << 8) + d[7];
 400                         break;
 401                 default:
 402                         break;
 403                 }
 404                 d += d[3] + 4;
 405         }
 406
 407         if (h->group_id == -1) {
 408                 /*
 409                  * Internal error; TPGS supported but required
 410                  * VPD identification descriptors not present.
 411                  * Disable ALUA support
 412                  */
 413                 sdev_printk(KERN_INFO, sdev,
 414                             "%s: No target port descriptors found\n",
 415                             ALUA_DH_NAME);
 416                 h->state = TPGS_STATE_OPTIMIZED;
 417                 h->tpgs = TPGS_MODE_NONE;
 418                 err = SCSI_DH_DEV_UNSUPP;
 419         } else {
 420                 sdev_printk(KERN_INFO, sdev,
 421                             "%s: port group %02x rel port %02x\n",
 422                             ALUA_DH_NAME, h->group_id, h->rel_port);
 423         }
 424
 425         return err;
 426 }
 427
 428 static char print_alua_state(int state)
 429 {
 430         switch (state) {
 431         case TPGS_STATE_OPTIMIZED:
 432                 return 'A';
 433         case TPGS_STATE_NONOPTIMIZED:
 434                 return 'N';
 435         case TPGS_STATE_STANDBY:
 436                 return 'S';
 437         case TPGS_STATE_UNAVAILABLE:
 438                 return 'U';
 439         case TPGS_STATE_LBA_DEPENDENT:
 440                 return 'L';
 441         case TPGS_STATE_OFFLINE:
 442                 return 'O';
 443         case TPGS_STATE_TRANSITIONING:
 444                 return 'T';
 445         default:
 446                 return 'X';
 447         }
 448 }
 449
 450 static int alua_check_sense(struct scsi_device *sdev,
 451                             struct scsi_sense_hdr *sense_hdr)
 452 {
 453         switch (sense_hdr->sense_key) {
 454         case NOT_READY:
 455                 if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0a)
 456                         /*
 457                          * LUN Not Accessible - ALUA state transition
 458                          */
 459                         return ADD_TO_MLQUEUE;
 460                 if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0b)
 461                         /*
 462                          * LUN Not Accessible -- Target port in standby state
 463                          */
 464                         return SUCCESS;
 465                 if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0c)
 466                         /*
 467                          * LUN Not Accessible -- Target port in unavailable state
 468                          */
 469                         return SUCCESS;
 470                 if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x12)
 471                         /*
 472                          * LUN Not Ready -- Offline
 473                          */
 474                         return SUCCESS;
 475                 break;
 476         case UNIT_ATTENTION:
 477                 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x00)
 478                         /*
 479                          * Power On, Reset, or Bus Device Reset, just retry.
 480                          */
 481                         return ADD_TO_MLQUEUE;
 482                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x01)
 483                         /*
 484                          * Mode Parameters Changed
 485                          */
 486                         return ADD_TO_MLQUEUE;
 487                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x06)
 488                         /*
 489                          * ALUA state changed
 490                          */
 491                         return ADD_TO_MLQUEUE;
 492                 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x07)
 493                         /*
 494                          * Implicit ALUA state transition failed
 495                          */
 496                         return ADD_TO_MLQUEUE;
 497                 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x03)
 498                         /*
 499                          * Inquiry data has changed
 500                          */
 501                         return ADD_TO_MLQUEUE;
 502                 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x0e)
 503                         /*
 504                          * REPORTED_LUNS_DATA_HAS_CHANGED is reported
 505                          * when switching controllers on targets like
 506                          * Intel Multi-Flex. We can just retry.
 507                          */
 508                         return ADD_TO_MLQUEUE;
 509                 break;
 510         }
 511
 512         return SCSI_RETURN_NOT_HANDLED;
 513 }
 514
 515 /*
 516  * alua_rtpg - Evaluate REPORT TARGET GROUP STATES
 517  * @sdev: the device to be evaluated.
 518  *
 519  * Evaluate the Target Port Group State.
 520  * Returns SCSI_DH_DEV_OFFLINED if the path is
 521  * found to be unusable.
 522  */
 523 static int alua_rtpg(struct scsi_device *sdev, struct alua_dh_data *h)
 524 {
 525         struct scsi_sense_hdr sense_hdr;
 526         int len, k, off, valid_states = 0;
 527         unsigned char *ucp;
 528         unsigned err;
 529         bool rtpg_ext_hdr_req = 1;
 530         unsigned long expiry, interval = 0;
 531         unsigned int tpg_desc_tbl_off;
 532         unsigned char orig_transition_tmo;
 533
 534         if (!h->transition_tmo)
 535                 expiry = round_jiffies_up(jiffies + ALUA_FAILOVER_TIMEOUT * HZ);
 536         else
 537                 expiry = round_jiffies_up(jiffies + h->transition_tmo * HZ);
 538
 539  retry:
 540         err = submit_rtpg(sdev, h, rtpg_ext_hdr_req);
 541
 542         if (err == SCSI_DH_IO && h->senselen > 0) {
 543                 err = scsi_normalize_sense(h->sense, SCSI_SENSE_BUFFERSIZE,
 544                                            &sense_hdr);
 545                 if (!err)
 546                         return SCSI_DH_IO;
 547
 548                 /*
 549                  * submit_rtpg() has failed on existing arrays
 550                  * when requesting extended header info, and
 551                  * the array doesn't support extended headers,
 552                  * even though it shouldn't according to T10.
 553                  * The retry without rtpg_ext_hdr_req set
 554                  * handles this.
 555                  */
 556                 if (rtpg_ext_hdr_req == 1 &&
 557                     sense_hdr.sense_key == ILLEGAL_REQUEST &&
 558                     sense_hdr.asc == 0x24 && sense_hdr.ascq == 0) {
 559                         rtpg_ext_hdr_req = 0;
 560                         goto retry;
 561                 }
 562
 563                 err = alua_check_sense(sdev, &sense_hdr);
 564                 if (err == ADD_TO_MLQUEUE && time_before(jiffies, expiry))
 565                         goto retry;
 566                 sdev_printk(KERN_INFO, sdev,
 567                             "%s: rtpg sense code %02x/%02x/%02x\n",
 568                             ALUA_DH_NAME, sense_hdr.sense_key,
 569                             sense_hdr.asc, sense_hdr.ascq);
 570                 err = SCSI_DH_IO;
 571         }
 572         if (err != SCSI_DH_OK)
 573                 return err;
 574
 575         len = (h->buff[0] << 24) + (h->buff[1] << 16) +
 576                 (h->buff[2] << 8) + h->buff[3] + 4;
 577
 578         if (len > h->bufflen) {
 579                 /* Resubmit with the correct length */
 580                 if (realloc_buffer(h, len)) {
 581                         sdev_printk(KERN_WARNING, sdev,
 582                                     "%s: kmalloc buffer failed\n",__func__);
 583                         /* Temporary failure, bypass */
 584                         return SCSI_DH_DEV_TEMP_BUSY;
 585                 }
 586                 goto retry;
 587         }
 588
 589         orig_transition_tmo = h->transition_tmo;
 590         if ((h->buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR && h->buff[5] != 0)
 591                 h->transition_tmo = h->buff[5];
 592         else
 593                 h->transition_tmo = ALUA_FAILOVER_TIMEOUT;
 594
 595         if (orig_transition_tmo != h->transition_tmo) {
 596                 sdev_printk(KERN_INFO, sdev,
 597                             "%s: transition timeout set to %d seconds\n",
 598                             ALUA_DH_NAME, h->transition_tmo);
 599                 expiry = jiffies + h->transition_tmo * HZ;
 600         }
 601
 602         if ((h->buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR)
 603                 tpg_desc_tbl_off = 8;
 604         else
 605                 tpg_desc_tbl_off = 4;
 606
 607         for (k = tpg_desc_tbl_off, ucp = h->buff + tpg_desc_tbl_off;
 608              k < len;
 609              k += off, ucp += off) {
 610
 611                 if (h->group_id == (ucp[2] << 8) + ucp[3]) {
 612                         h->state = ucp[0] & 0x0f;
 613                         h->pref = ucp[0] >> 7;
 614                         valid_states = ucp[1];
 615                 }
 616                 off = 8 + (ucp[7] * 4);
 617         }
 618
 619         sdev_printk(KERN_INFO, sdev,
 620                     "%s: port group %02x state %c %s supports %c%c%c%c%c%c%c\n",
 621                     ALUA_DH_NAME, h->group_id, print_alua_state(h->state),
 622                     h->pref ? "preferred" : "non-preferred",
 623                     valid_states&TPGS_SUPPORT_TRANSITION?'T':'t',
 624                     valid_states&TPGS_SUPPORT_OFFLINE?'O':'o',
 625                     valid_states&TPGS_SUPPORT_LBA_DEPENDENT?'L':'l',
 626                     valid_states&TPGS_SUPPORT_UNAVAILABLE?'U':'u',
 627                     valid_states&TPGS_SUPPORT_STANDBY?'S':'s',
 628                     valid_states&TPGS_SUPPORT_NONOPTIMIZED?'N':'n',
 629                     valid_states&TPGS_SUPPORT_OPTIMIZED?'A':'a');
 630
 631         switch (h->state) {
 632         case TPGS_STATE_TRANSITIONING:
 633                 if (time_before(jiffies, expiry)) {
 634                         /* State transition, retry */
 635                         interval += 2000;
 636                         msleep(interval);
 637                         goto retry;
 638                 }
 639                 /* Transitioning time exceeded, set port to standby */
 640                 err = SCSI_DH_RETRY;
 641                 h->state = TPGS_STATE_STANDBY;
 642                 break;
 643         case TPGS_STATE_OFFLINE:
 644         case TPGS_STATE_UNAVAILABLE:
 645                 /* Path unusable for unavailable/offline */
 646                 err = SCSI_DH_DEV_OFFLINED;
 647                 break;
 648         default:
 649                 /* Useable path if active */
 650                 err = SCSI_DH_OK;
 651                 break;
 652         }
 653         return err;
 654 }
 655
 656 /*
 657  * alua_initialize - Initialize ALUA state
 658  * @sdev: the device to be initialized
 659  *
 660  * For the prep_fn to work correctly we have
 661  * to initialize the ALUA state for the device.
 662  */
 663 static int alua_initialize(struct scsi_device *sdev, struct alua_dh_data *h)
 664 {
 665         int err;
 666
 667         err = alua_check_tpgs(sdev, h);
 668         if (err != SCSI_DH_OK)
 669                 goto out;
 670
 671         err = alua_vpd_inquiry(sdev, h);
 672         if (err != SCSI_DH_OK)
 673                 goto out;
 674
 675         err = alua_rtpg(sdev, h);
 676         if (err != SCSI_DH_OK)
 677                 goto out;
 678
 679 out:
 680         return err;
 681 }
 682 /*
 683  * alua_set_params - set/unset the optimize flag
 684  * @sdev: device on the path to be activated
 685  * params - parameters in the following format
 686  *      "no_of_params\0param1\0param2\0param3\0...\0"
 687  * For example, to set the flag pass the following parameters
 688  * from multipath.conf
 689  *     hardware_handler        "2 alua 1"
 690  */
 691 static int alua_set_params(struct scsi_device *sdev, const char *params)
 692 {
 693         struct alua_dh_data *h = get_alua_data(sdev);
 694         unsigned int optimize = 0, argc;
 695         const char *p = params;
 696         int result = SCSI_DH_OK;
 697
 698         if ((sscanf(params, "%u", &argc) != 1) || (argc != 1))
 699                 return -EINVAL;
 700
 701         while (*p++)
 702                 ;
 703         if ((sscanf(p, "%u", &optimize) != 1) || (optimize > 1))
 704                 return -EINVAL;
 705
 706         if (optimize)
 707                 h->flags |= ALUA_OPTIMIZE_STPG;
 708         else
 709                 h->flags &= ~ALUA_OPTIMIZE_STPG;
 710
 711         return result;
 712 }
 713
 714 /*
 715  * alua_activate - activate a path
 716  * @sdev: device on the path to be activated
 717  *
 718  * We're currently switching the port group to be activated only and
 719  * let the array figure out the rest.
 720  * There may be other arrays which require us to switch all port groups
 721  * based on a certain policy. But until we actually encounter them it
 722  * should be okay.
 723  */
 724 static int alua_activate(struct scsi_device *sdev,
 725                         activate_complete fn, void *data)
 726 {
 727         struct alua_dh_data *h = get_alua_data(sdev);
 728         int err = SCSI_DH_OK;
 729         int stpg = 0;
 730
 731         err = alua_rtpg(sdev, h);
 732         if (err != SCSI_DH_OK)
 733                 goto out;
 734
 735         if (h->tpgs & TPGS_MODE_EXPLICIT) {
 736                 switch (h->state) {
 737                 case TPGS_STATE_NONOPTIMIZED:
 738                         stpg = 1;
 739                         if ((h->flags & ALUA_OPTIMIZE_STPG) &&
 740                             (!h->pref) &&
 741                             (h->tpgs & TPGS_MODE_IMPLICIT))
 742                                 stpg = 0;
 743                         break;
 744                 case TPGS_STATE_STANDBY:
 745                 case TPGS_STATE_UNAVAILABLE:
 746                         stpg = 1;
 747                         break;
 748                 case TPGS_STATE_OFFLINE:
 749                         err = SCSI_DH_IO;
 750                         break;
 751                 case TPGS_STATE_TRANSITIONING:
 752                         err = SCSI_DH_RETRY;
 753                         break;
 754                 default:
 755                         break;
 756                 }
 757         }
 758
 759         if (stpg) {
 760                 h->callback_fn = fn;
 761                 h->callback_data = data;
 762                 err = submit_stpg(h);
 763                 if (err == SCSI_DH_OK)
 764                         return 0;
 765                 h->callback_fn = h->callback_data = NULL;
 766         }
 767
 768 out:
 769         if (fn)
 770                 fn(data, err);
 771         return 0;
 772 }
 773
 774 /*
 775  * alua_prep_fn - request callback
 776  *
 777  * Fail I/O to all paths not in state
 778  * active/optimized or active/non-optimized.
 779  */
 780 static int alua_prep_fn(struct scsi_device *sdev, struct request *req)
 781 {
 782         struct alua_dh_data *h = get_alua_data(sdev);
 783         int ret = BLKPREP_OK;
 784
 785         if (h->state == TPGS_STATE_TRANSITIONING)
 786                 ret = BLKPREP_DEFER;
 787         else if (h->state != TPGS_STATE_OPTIMIZED &&
 788                  h->state != TPGS_STATE_NONOPTIMIZED &&
 789                  h->state != TPGS_STATE_LBA_DEPENDENT) {
 790                 ret = BLKPREP_KILL;
 791                 req->cmd_flags |= REQ_QUIET;
 792         }
 793         return ret;
 794
 795 }
 796
 797 static bool alua_match(struct scsi_device *sdev)
 798 {
 799         return (scsi_device_tpgs(sdev) != 0);
 800 }
 801
 802 static int alua_bus_attach(struct scsi_device *sdev);
 803 static void alua_bus_detach(struct scsi_device *sdev);
 804
 805 static struct scsi_device_handler alua_dh = {
 806         .name = ALUA_DH_NAME,
 807         .module = THIS_MODULE,
 808         .attach = alua_bus_attach,
 809         .detach = alua_bus_detach,
 810         .prep_fn = alua_prep_fn,
 811         .check_sense = alua_check_sense,
 812         .activate = alua_activate,
 813         .set_params = alua_set_params,
 814         .match = alua_match,
 815 };
 816
 817 /*
 818  * alua_bus_attach - Attach device handler
 819  * @sdev: device to be attached to
 820  */
 821 static int alua_bus_attach(struct scsi_device *sdev)
 822 {
 823         struct scsi_dh_data *scsi_dh_data;
 824         struct alua_dh_data *h;
 825         unsigned long flags;
 826         int err = SCSI_DH_OK;
 827
 828         scsi_dh_data = kzalloc(sizeof(*scsi_dh_data)
 829                                + sizeof(*h) , GFP_KERNEL);
 830         if (!scsi_dh_data) {
 831                 sdev_printk(KERN_ERR, sdev, "%s: Attach failed\n",
 832                             ALUA_DH_NAME);
 833                 return -ENOMEM;
 834         }
 835
 836         scsi_dh_data->scsi_dh = &alua_dh;
 837         h = (struct alua_dh_data *) scsi_dh_data->buf;
 838         h->tpgs = TPGS_MODE_UNINITIALIZED;
 839         h->state = TPGS_STATE_OPTIMIZED;
 840         h->group_id = -1;
 841         h->rel_port = -1;
 842         h->buff = h->inq;
 843         h->bufflen = ALUA_INQUIRY_SIZE;
 844         h->sdev = sdev;
 845
 846         err = alua_initialize(sdev, h);
 847         if ((err != SCSI_DH_OK) && (err != SCSI_DH_DEV_OFFLINED))
 848                 goto failed;
 849
 850         if (!try_module_get(THIS_MODULE))
 851                 goto failed;
 852
 853         spin_lock_irqsave(sdev->request_queue->queue_lock, flags);
 854         sdev->scsi_dh_data = scsi_dh_data;
 855         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
 856         sdev_printk(KERN_NOTICE, sdev, "%s: Attached\n", ALUA_DH_NAME);
 857
 858         return 0;
 859
 860 failed:
 861         kfree(scsi_dh_data);
 862         sdev_printk(KERN_ERR, sdev, "%s: not attached\n", ALUA_DH_NAME);
 863         return -EINVAL;
 864 }
 865
 866 /*
 867  * alua_bus_detach - Detach device handler
 868  * @sdev: device to be detached from
 869  */
 870 static void alua_bus_detach(struct scsi_device *sdev)
 871 {
 872         struct scsi_dh_data *scsi_dh_data;
 873         struct alua_dh_data *h;
 874         unsigned long flags;
 875
 876         spin_lock_irqsave(sdev->request_queue->queue_lock, flags);
 877         scsi_dh_data = sdev->scsi_dh_data;
 878         sdev->scsi_dh_data = NULL;
 879         spin_unlock_irqrestore(sdev->request_queue->queue_lock, flags);
 880
 881         h = (struct alua_dh_data *) scsi_dh_data->buf;
 882         if (h->buff && h->inq != h->buff)
 883                 kfree(h->buff);
 884         kfree(scsi_dh_data);
 885         module_put(THIS_MODULE);
 886         sdev_printk(KERN_NOTICE, sdev, "%s: Detached\n", ALUA_DH_NAME);
 887 }
 888
 889 static int __init alua_init(void)
 890 {
 891         int r;
 892
 893         r = scsi_register_device_handler(&alua_dh);
 894         if (r != 0)
 895                 printk(KERN_ERR "%s: Failed to register scsi device handler",
 896                         ALUA_DH_NAME);
 897         return r;
 898 }
 899
 900 static void __exit alua_exit(void)
 901 {
 902         scsi_unregister_device_handler(&alua_dh);
 903 }
 904
 905 module_init(alua_init);
 906 module_exit(alua_exit);
 907
 908 MODULE_DESCRIPTION("DM Multipath ALUA support");
 909 MODULE_AUTHOR("Hannes Reinecke <hare@suse.de>");
 910 MODULE_LICENSE("GPL");
 911 MODULE_VERSION(ALUA_DH_VER);