1329 lines · c
1// SPDX-License-Identifier: GPL-2.0-or-later2/*3 * Generic SCSI-3 ALUA SCSI Device Handler4 *5 * Copyright (C) 2007-2010 Hannes Reinecke, SUSE Linux Products GmbH.6 * All rights reserved.7 */8#include <linux/slab.h>9#include <linux/delay.h>10#include <linux/module.h>11#include <linux/unaligned.h>12#include <scsi/scsi.h>13#include <scsi/scsi_proto.h>14#include <scsi/scsi_dbg.h>15#include <scsi/scsi_eh.h>16#include <scsi/scsi_dh.h>17 18#define ALUA_DH_NAME "alua"19#define ALUA_DH_VER "2.0"20 21#define TPGS_SUPPORT_NONE 0x0022#define TPGS_SUPPORT_OPTIMIZED 0x0123#define TPGS_SUPPORT_NONOPTIMIZED 0x0224#define TPGS_SUPPORT_STANDBY 0x0425#define TPGS_SUPPORT_UNAVAILABLE 0x0826#define TPGS_SUPPORT_LBA_DEPENDENT 0x1027#define TPGS_SUPPORT_OFFLINE 0x4028#define TPGS_SUPPORT_TRANSITION 0x8029#define TPGS_SUPPORT_ALL 0xdf30 31#define RTPG_FMT_MASK 0x7032#define RTPG_FMT_EXT_HDR 0x1033 34#define TPGS_MODE_UNINITIALIZED -135#define TPGS_MODE_NONE 0x036#define TPGS_MODE_IMPLICIT 0x137#define TPGS_MODE_EXPLICIT 0x238 39#define ALUA_RTPG_SIZE 12840#define ALUA_FAILOVER_TIMEOUT 6041#define ALUA_FAILOVER_RETRIES 542#define ALUA_RTPG_DELAY_MSECS 543#define ALUA_RTPG_RETRY_DELAY 244 45/* device handler flags */46#define ALUA_OPTIMIZE_STPG 0x0147#define ALUA_RTPG_EXT_HDR_UNSUPP 0x0248/* State machine flags */49#define ALUA_PG_RUN_RTPG 0x1050#define ALUA_PG_RUN_STPG 0x2051#define ALUA_PG_RUNNING 0x4052 53static uint optimize_stpg;54module_param(optimize_stpg, uint, S_IRUGO|S_IWUSR);55MODULE_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.");56 57static LIST_HEAD(port_group_list);58static DEFINE_SPINLOCK(port_group_lock);59static struct workqueue_struct *kaluad_wq;60 61struct alua_port_group {62 struct kref kref;63 struct rcu_head rcu;64 struct list_head node;65 struct list_head dh_list;66 unsigned char device_id_str[256];67 int device_id_len;68 int group_id;69 int tpgs;70 int state;71 int pref;72 int valid_states;73 unsigned flags; /* used for optimizing STPG */74 unsigned char transition_tmo;75 unsigned long expiry;76 unsigned long interval;77 struct delayed_work rtpg_work;78 spinlock_t lock;79 struct list_head rtpg_list;80 struct scsi_device *rtpg_sdev;81};82 83struct alua_dh_data {84 struct list_head node;85 struct alua_port_group __rcu *pg;86 int group_id;87 spinlock_t pg_lock;88 struct scsi_device *sdev;89 int init_error;90 struct mutex init_mutex;91 bool disabled;92};93 94struct alua_queue_data {95 struct list_head entry;96 activate_complete callback_fn;97 void *callback_data;98};99 100#define ALUA_POLICY_SWITCH_CURRENT 0101#define ALUA_POLICY_SWITCH_ALL 1102 103static void alua_rtpg_work(struct work_struct *work);104static bool alua_rtpg_queue(struct alua_port_group *pg,105 struct scsi_device *sdev,106 struct alua_queue_data *qdata, bool force);107static void alua_check(struct scsi_device *sdev, bool force);108 109static void release_port_group(struct kref *kref)110{111 struct alua_port_group *pg;112 113 pg = container_of(kref, struct alua_port_group, kref);114 if (pg->rtpg_sdev)115 flush_delayed_work(&pg->rtpg_work);116 spin_lock(&port_group_lock);117 list_del(&pg->node);118 spin_unlock(&port_group_lock);119 kfree_rcu(pg, rcu);120}121 122/*123 * submit_rtpg - Issue a REPORT TARGET GROUP STATES command124 * @sdev: sdev the command should be sent to125 */126static int submit_rtpg(struct scsi_device *sdev, unsigned char *buff,127 int bufflen, struct scsi_sense_hdr *sshdr, int flags)128{129 u8 cdb[MAX_COMMAND_SIZE];130 blk_opf_t opf = REQ_OP_DRV_IN | REQ_FAILFAST_DEV |131 REQ_FAILFAST_TRANSPORT | REQ_FAILFAST_DRIVER;132 const struct scsi_exec_args exec_args = {133 .sshdr = sshdr,134 };135 136 /* Prepare the command. */137 memset(cdb, 0x0, MAX_COMMAND_SIZE);138 cdb[0] = MAINTENANCE_IN;139 if (!(flags & ALUA_RTPG_EXT_HDR_UNSUPP))140 cdb[1] = MI_REPORT_TARGET_PGS | MI_EXT_HDR_PARAM_FMT;141 else142 cdb[1] = MI_REPORT_TARGET_PGS;143 put_unaligned_be32(bufflen, &cdb[6]);144 145 return scsi_execute_cmd(sdev, cdb, opf, buff, bufflen,146 ALUA_FAILOVER_TIMEOUT * HZ,147 ALUA_FAILOVER_RETRIES, &exec_args);148}149 150/*151 * submit_stpg - Issue a SET TARGET PORT GROUP command152 *153 * Currently we're only setting the current target port group state154 * to 'active/optimized' and let the array firmware figure out155 * the states of the remaining groups.156 */157static int submit_stpg(struct scsi_device *sdev, int group_id,158 struct scsi_sense_hdr *sshdr)159{160 u8 cdb[MAX_COMMAND_SIZE];161 unsigned char stpg_data[8];162 int stpg_len = 8;163 blk_opf_t opf = REQ_OP_DRV_OUT | REQ_FAILFAST_DEV |164 REQ_FAILFAST_TRANSPORT | REQ_FAILFAST_DRIVER;165 const struct scsi_exec_args exec_args = {166 .sshdr = sshdr,167 };168 169 /* Prepare the data buffer */170 memset(stpg_data, 0, stpg_len);171 stpg_data[4] = SCSI_ACCESS_STATE_OPTIMAL;172 put_unaligned_be16(group_id, &stpg_data[6]);173 174 /* Prepare the command. */175 memset(cdb, 0x0, MAX_COMMAND_SIZE);176 cdb[0] = MAINTENANCE_OUT;177 cdb[1] = MO_SET_TARGET_PGS;178 put_unaligned_be32(stpg_len, &cdb[6]);179 180 return scsi_execute_cmd(sdev, cdb, opf, stpg_data,181 stpg_len, ALUA_FAILOVER_TIMEOUT * HZ,182 ALUA_FAILOVER_RETRIES, &exec_args);183}184 185static struct alua_port_group *alua_find_get_pg(char *id_str, size_t id_size,186 int group_id)187{188 struct alua_port_group *pg;189 190 if (!id_str || !id_size || !strlen(id_str))191 return NULL;192 193 list_for_each_entry(pg, &port_group_list, node) {194 if (pg->group_id != group_id)195 continue;196 if (!pg->device_id_len || pg->device_id_len != id_size)197 continue;198 if (strncmp(pg->device_id_str, id_str, id_size))199 continue;200 if (!kref_get_unless_zero(&pg->kref))201 continue;202 return pg;203 }204 205 return NULL;206}207 208/*209 * alua_alloc_pg - Allocate a new port_group structure210 * @sdev: scsi device211 * @group_id: port group id212 * @tpgs: target port group settings213 *214 * Allocate a new port_group structure for a given215 * device.216 */217static struct alua_port_group *alua_alloc_pg(struct scsi_device *sdev,218 int group_id, int tpgs)219{220 struct alua_port_group *pg, *tmp_pg;221 222 pg = kzalloc(sizeof(struct alua_port_group), GFP_KERNEL);223 if (!pg)224 return ERR_PTR(-ENOMEM);225 226 pg->device_id_len = scsi_vpd_lun_id(sdev, pg->device_id_str,227 sizeof(pg->device_id_str));228 if (pg->device_id_len <= 0) {229 /*230 * TPGS supported but no device identification found.231 * Generate private device identification.232 */233 sdev_printk(KERN_INFO, sdev,234 "%s: No device descriptors found\n",235 ALUA_DH_NAME);236 pg->device_id_str[0] = '\0';237 pg->device_id_len = 0;238 }239 pg->group_id = group_id;240 pg->tpgs = tpgs;241 pg->state = SCSI_ACCESS_STATE_OPTIMAL;242 pg->valid_states = TPGS_SUPPORT_ALL;243 if (optimize_stpg)244 pg->flags |= ALUA_OPTIMIZE_STPG;245 kref_init(&pg->kref);246 INIT_DELAYED_WORK(&pg->rtpg_work, alua_rtpg_work);247 INIT_LIST_HEAD(&pg->rtpg_list);248 INIT_LIST_HEAD(&pg->node);249 INIT_LIST_HEAD(&pg->dh_list);250 spin_lock_init(&pg->lock);251 252 spin_lock(&port_group_lock);253 tmp_pg = alua_find_get_pg(pg->device_id_str, pg->device_id_len,254 group_id);255 if (tmp_pg) {256 spin_unlock(&port_group_lock);257 kfree(pg);258 return tmp_pg;259 }260 261 list_add(&pg->node, &port_group_list);262 spin_unlock(&port_group_lock);263 264 return pg;265}266 267/*268 * alua_check_tpgs - Evaluate TPGS setting269 * @sdev: device to be checked270 *271 * Examine the TPGS setting of the sdev to find out if ALUA272 * is supported.273 */274static int alua_check_tpgs(struct scsi_device *sdev)275{276 int tpgs = TPGS_MODE_NONE;277 278 /*279 * ALUA support for non-disk devices is fraught with280 * difficulties, so disable it for now.281 */282 if (sdev->type != TYPE_DISK) {283 sdev_printk(KERN_INFO, sdev,284 "%s: disable for non-disk devices\n",285 ALUA_DH_NAME);286 return tpgs;287 }288 289 tpgs = scsi_device_tpgs(sdev);290 switch (tpgs) {291 case TPGS_MODE_EXPLICIT|TPGS_MODE_IMPLICIT:292 sdev_printk(KERN_INFO, sdev,293 "%s: supports implicit and explicit TPGS\n",294 ALUA_DH_NAME);295 break;296 case TPGS_MODE_EXPLICIT:297 sdev_printk(KERN_INFO, sdev, "%s: supports explicit TPGS\n",298 ALUA_DH_NAME);299 break;300 case TPGS_MODE_IMPLICIT:301 sdev_printk(KERN_INFO, sdev, "%s: supports implicit TPGS\n",302 ALUA_DH_NAME);303 break;304 case TPGS_MODE_NONE:305 sdev_printk(KERN_INFO, sdev, "%s: not supported\n",306 ALUA_DH_NAME);307 break;308 default:309 sdev_printk(KERN_INFO, sdev,310 "%s: unsupported TPGS setting %d\n",311 ALUA_DH_NAME, tpgs);312 tpgs = TPGS_MODE_NONE;313 break;314 }315 316 return tpgs;317}318 319/*320 * alua_check_vpd - Evaluate INQUIRY vpd page 0x83321 * @sdev: device to be checked322 *323 * Extract the relative target port and the target port group324 * descriptor from the list of identificators.325 */326static int alua_check_vpd(struct scsi_device *sdev, struct alua_dh_data *h,327 int tpgs)328{329 int rel_port = -1, group_id;330 struct alua_port_group *pg, *old_pg = NULL;331 bool pg_updated = false;332 unsigned long flags;333 334 group_id = scsi_vpd_tpg_id(sdev, &rel_port);335 if (group_id < 0) {336 /*337 * Internal error; TPGS supported but required338 * VPD identification descriptors not present.339 * Disable ALUA support340 */341 sdev_printk(KERN_INFO, sdev,342 "%s: No target port descriptors found\n",343 ALUA_DH_NAME);344 return SCSI_DH_DEV_UNSUPP;345 }346 347 pg = alua_alloc_pg(sdev, group_id, tpgs);348 if (IS_ERR(pg)) {349 if (PTR_ERR(pg) == -ENOMEM)350 return SCSI_DH_NOMEM;351 return SCSI_DH_DEV_UNSUPP;352 }353 if (pg->device_id_len)354 sdev_printk(KERN_INFO, sdev,355 "%s: device %s port group %x rel port %x\n",356 ALUA_DH_NAME, pg->device_id_str,357 group_id, rel_port);358 else359 sdev_printk(KERN_INFO, sdev,360 "%s: port group %x rel port %x\n",361 ALUA_DH_NAME, group_id, rel_port);362 363 kref_get(&pg->kref);364 365 /* Check for existing port group references */366 spin_lock(&h->pg_lock);367 old_pg = rcu_dereference_protected(h->pg, lockdep_is_held(&h->pg_lock));368 if (old_pg != pg) {369 /* port group has changed. Update to new port group */370 if (h->pg) {371 spin_lock_irqsave(&old_pg->lock, flags);372 list_del_rcu(&h->node);373 spin_unlock_irqrestore(&old_pg->lock, flags);374 }375 rcu_assign_pointer(h->pg, pg);376 pg_updated = true;377 }378 379 spin_lock_irqsave(&pg->lock, flags);380 if (pg_updated)381 list_add_rcu(&h->node, &pg->dh_list);382 spin_unlock_irqrestore(&pg->lock, flags);383 384 spin_unlock(&h->pg_lock);385 386 alua_rtpg_queue(pg, sdev, NULL, true);387 kref_put(&pg->kref, release_port_group);388 389 if (old_pg)390 kref_put(&old_pg->kref, release_port_group);391 392 return SCSI_DH_OK;393}394 395static char print_alua_state(unsigned char state)396{397 switch (state) {398 case SCSI_ACCESS_STATE_OPTIMAL:399 return 'A';400 case SCSI_ACCESS_STATE_ACTIVE:401 return 'N';402 case SCSI_ACCESS_STATE_STANDBY:403 return 'S';404 case SCSI_ACCESS_STATE_UNAVAILABLE:405 return 'U';406 case SCSI_ACCESS_STATE_LBA:407 return 'L';408 case SCSI_ACCESS_STATE_OFFLINE:409 return 'O';410 case SCSI_ACCESS_STATE_TRANSITIONING:411 return 'T';412 default:413 return 'X';414 }415}416 417static void alua_handle_state_transition(struct scsi_device *sdev)418{419 struct alua_dh_data *h = sdev->handler_data;420 struct alua_port_group *pg;421 422 rcu_read_lock();423 pg = rcu_dereference(h->pg);424 if (pg)425 pg->state = SCSI_ACCESS_STATE_TRANSITIONING;426 rcu_read_unlock();427 alua_check(sdev, false);428}429 430static enum scsi_disposition alua_check_sense(struct scsi_device *sdev,431 struct scsi_sense_hdr *sense_hdr)432{433 switch (sense_hdr->sense_key) {434 case NOT_READY:435 if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0a) {436 /*437 * LUN Not Accessible - ALUA state transition438 */439 alua_handle_state_transition(sdev);440 return NEEDS_RETRY;441 }442 break;443 case UNIT_ATTENTION:444 if (sense_hdr->asc == 0x04 && sense_hdr->ascq == 0x0a) {445 /*446 * LUN Not Accessible - ALUA state transition447 */448 alua_handle_state_transition(sdev);449 return NEEDS_RETRY;450 }451 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x00) {452 /*453 * Power On, Reset, or Bus Device Reset.454 * Might have obscured a state transition,455 * so schedule a recheck.456 */457 alua_check(sdev, true);458 return ADD_TO_MLQUEUE;459 }460 if (sense_hdr->asc == 0x29 && sense_hdr->ascq == 0x04)461 /*462 * Device internal reset463 */464 return ADD_TO_MLQUEUE;465 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x01)466 /*467 * Mode Parameters Changed468 */469 return ADD_TO_MLQUEUE;470 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x06) {471 /*472 * ALUA state changed473 */474 alua_check(sdev, true);475 return ADD_TO_MLQUEUE;476 }477 if (sense_hdr->asc == 0x2a && sense_hdr->ascq == 0x07) {478 /*479 * Implicit ALUA state transition failed480 */481 alua_check(sdev, true);482 return ADD_TO_MLQUEUE;483 }484 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x03)485 /*486 * Inquiry data has changed487 */488 return ADD_TO_MLQUEUE;489 if (sense_hdr->asc == 0x3f && sense_hdr->ascq == 0x0e)490 /*491 * REPORTED_LUNS_DATA_HAS_CHANGED is reported492 * when switching controllers on targets like493 * Intel Multi-Flex. We can just retry.494 */495 return ADD_TO_MLQUEUE;496 break;497 }498 499 return SCSI_RETURN_NOT_HANDLED;500}501 502/*503 * alua_tur - Send a TEST UNIT READY504 * @sdev: device to which the TEST UNIT READY command should be send505 *506 * Send a TEST UNIT READY to @sdev to figure out the device state507 * Returns SCSI_DH_RETRY if the sense code is NOT READY/ALUA TRANSITIONING,508 * SCSI_DH_OK if no error occurred, and SCSI_DH_IO otherwise.509 */510static int alua_tur(struct scsi_device *sdev)511{512 struct scsi_sense_hdr sense_hdr;513 int retval;514 515 retval = scsi_test_unit_ready(sdev, ALUA_FAILOVER_TIMEOUT * HZ,516 ALUA_FAILOVER_RETRIES, &sense_hdr);517 if ((sense_hdr.sense_key == NOT_READY ||518 sense_hdr.sense_key == UNIT_ATTENTION) &&519 sense_hdr.asc == 0x04 && sense_hdr.ascq == 0x0a)520 return SCSI_DH_RETRY;521 else if (retval)522 return SCSI_DH_IO;523 else524 return SCSI_DH_OK;525}526 527/*528 * alua_rtpg - Evaluate REPORT TARGET GROUP STATES529 * @sdev: the device to be evaluated.530 *531 * Evaluate the Target Port Group State.532 * Returns SCSI_DH_DEV_OFFLINED if the path is533 * found to be unusable.534 */535static int alua_rtpg(struct scsi_device *sdev, struct alua_port_group *pg)536{537 struct scsi_sense_hdr sense_hdr;538 struct alua_port_group *tmp_pg;539 int len, k, off, bufflen = ALUA_RTPG_SIZE;540 int group_id_old, state_old, pref_old, valid_states_old;541 unsigned char *desc, *buff;542 unsigned err;543 int retval;544 unsigned int tpg_desc_tbl_off;545 unsigned char orig_transition_tmo;546 unsigned long flags;547 bool transitioning_sense = false;548 549 group_id_old = pg->group_id;550 state_old = pg->state;551 pref_old = pg->pref;552 valid_states_old = pg->valid_states;553 554 if (!pg->expiry) {555 unsigned long transition_tmo = ALUA_FAILOVER_TIMEOUT * HZ;556 557 if (pg->transition_tmo)558 transition_tmo = pg->transition_tmo * HZ;559 560 pg->expiry = round_jiffies_up(jiffies + transition_tmo);561 }562 563 buff = kzalloc(bufflen, GFP_KERNEL);564 if (!buff)565 return SCSI_DH_DEV_TEMP_BUSY;566 567 retry:568 err = 0;569 retval = submit_rtpg(sdev, buff, bufflen, &sense_hdr, pg->flags);570 571 if (retval) {572 /*573 * Some (broken) implementations have a habit of returning574 * an error during things like firmware update etc.575 * But if the target only supports active/optimized there's576 * not much we can do; it's not that we can switch paths577 * or anything.578 * So ignore any errors to avoid spurious failures during579 * path failover.580 */581 if ((pg->valid_states & ~TPGS_SUPPORT_OPTIMIZED) == 0) {582 sdev_printk(KERN_INFO, sdev,583 "%s: ignoring rtpg result %d\n",584 ALUA_DH_NAME, retval);585 kfree(buff);586 return SCSI_DH_OK;587 }588 if (retval < 0 || !scsi_sense_valid(&sense_hdr)) {589 sdev_printk(KERN_INFO, sdev,590 "%s: rtpg failed, result %d\n",591 ALUA_DH_NAME, retval);592 kfree(buff);593 if (retval < 0)594 return SCSI_DH_DEV_TEMP_BUSY;595 if (host_byte(retval) == DID_NO_CONNECT)596 return SCSI_DH_RES_TEMP_UNAVAIL;597 return SCSI_DH_IO;598 }599 600 /*601 * submit_rtpg() has failed on existing arrays602 * when requesting extended header info, and603 * the array doesn't support extended headers,604 * even though it shouldn't according to T10.605 * The retry without rtpg_ext_hdr_req set606 * handles this.607 * Note: some arrays return a sense key of ILLEGAL_REQUEST608 * with ASC 00h if they don't support the extended header.609 */610 if (!(pg->flags & ALUA_RTPG_EXT_HDR_UNSUPP) &&611 sense_hdr.sense_key == ILLEGAL_REQUEST) {612 pg->flags |= ALUA_RTPG_EXT_HDR_UNSUPP;613 goto retry;614 }615 /*616 * If the array returns with 'ALUA state transition'617 * sense code here it cannot return RTPG data during618 * transition. So set the state to 'transitioning' directly.619 */620 if (sense_hdr.sense_key == NOT_READY &&621 sense_hdr.asc == 0x04 && sense_hdr.ascq == 0x0a) {622 transitioning_sense = true;623 goto skip_rtpg;624 }625 /*626 * Retry on any other UNIT ATTENTION occurred.627 */628 if (sense_hdr.sense_key == UNIT_ATTENTION)629 err = SCSI_DH_RETRY;630 if (err == SCSI_DH_RETRY &&631 pg->expiry != 0 && time_before(jiffies, pg->expiry)) {632 sdev_printk(KERN_ERR, sdev, "%s: rtpg retry\n",633 ALUA_DH_NAME);634 scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);635 kfree(buff);636 return err;637 }638 sdev_printk(KERN_ERR, sdev, "%s: rtpg failed\n",639 ALUA_DH_NAME);640 scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);641 kfree(buff);642 pg->expiry = 0;643 return SCSI_DH_IO;644 }645 646 len = get_unaligned_be32(&buff[0]) + 4;647 648 if (len > bufflen) {649 /* Resubmit with the correct length */650 kfree(buff);651 bufflen = len;652 buff = kmalloc(bufflen, GFP_KERNEL);653 if (!buff) {654 sdev_printk(KERN_WARNING, sdev,655 "%s: kmalloc buffer failed\n",__func__);656 /* Temporary failure, bypass */657 pg->expiry = 0;658 return SCSI_DH_DEV_TEMP_BUSY;659 }660 goto retry;661 }662 663 orig_transition_tmo = pg->transition_tmo;664 if ((buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR && buff[5] != 0)665 pg->transition_tmo = buff[5];666 else667 pg->transition_tmo = ALUA_FAILOVER_TIMEOUT;668 669 if (orig_transition_tmo != pg->transition_tmo) {670 sdev_printk(KERN_INFO, sdev,671 "%s: transition timeout set to %d seconds\n",672 ALUA_DH_NAME, pg->transition_tmo);673 pg->expiry = jiffies + pg->transition_tmo * HZ;674 }675 676 if ((buff[4] & RTPG_FMT_MASK) == RTPG_FMT_EXT_HDR)677 tpg_desc_tbl_off = 8;678 else679 tpg_desc_tbl_off = 4;680 681 for (k = tpg_desc_tbl_off, desc = buff + tpg_desc_tbl_off;682 k < len;683 k += off, desc += off) {684 u16 group_id = get_unaligned_be16(&desc[2]);685 686 spin_lock_irqsave(&port_group_lock, flags);687 tmp_pg = alua_find_get_pg(pg->device_id_str, pg->device_id_len,688 group_id);689 spin_unlock_irqrestore(&port_group_lock, flags);690 if (tmp_pg) {691 if (spin_trylock_irqsave(&tmp_pg->lock, flags)) {692 if ((tmp_pg == pg) ||693 !(tmp_pg->flags & ALUA_PG_RUNNING)) {694 struct alua_dh_data *h;695 696 tmp_pg->state = desc[0] & 0x0f;697 tmp_pg->pref = desc[0] >> 7;698 rcu_read_lock();699 list_for_each_entry_rcu(h,700 &tmp_pg->dh_list, node) {701 if (!h->sdev)702 continue;703 h->sdev->access_state = desc[0];704 }705 rcu_read_unlock();706 }707 if (tmp_pg == pg)708 tmp_pg->valid_states = desc[1];709 spin_unlock_irqrestore(&tmp_pg->lock, flags);710 }711 kref_put(&tmp_pg->kref, release_port_group);712 }713 off = 8 + (desc[7] * 4);714 }715 716 skip_rtpg:717 spin_lock_irqsave(&pg->lock, flags);718 if (transitioning_sense)719 pg->state = SCSI_ACCESS_STATE_TRANSITIONING;720 721 if (group_id_old != pg->group_id || state_old != pg->state ||722 pref_old != pg->pref || valid_states_old != pg->valid_states)723 sdev_printk(KERN_INFO, sdev,724 "%s: port group %02x state %c %s supports %c%c%c%c%c%c%c\n",725 ALUA_DH_NAME, pg->group_id, print_alua_state(pg->state),726 pg->pref ? "preferred" : "non-preferred",727 pg->valid_states&TPGS_SUPPORT_TRANSITION?'T':'t',728 pg->valid_states&TPGS_SUPPORT_OFFLINE?'O':'o',729 pg->valid_states&TPGS_SUPPORT_LBA_DEPENDENT?'L':'l',730 pg->valid_states&TPGS_SUPPORT_UNAVAILABLE?'U':'u',731 pg->valid_states&TPGS_SUPPORT_STANDBY?'S':'s',732 pg->valid_states&TPGS_SUPPORT_NONOPTIMIZED?'N':'n',733 pg->valid_states&TPGS_SUPPORT_OPTIMIZED?'A':'a');734 735 switch (pg->state) {736 case SCSI_ACCESS_STATE_TRANSITIONING:737 if (time_before(jiffies, pg->expiry)) {738 /* State transition, retry */739 pg->interval = ALUA_RTPG_RETRY_DELAY;740 err = SCSI_DH_RETRY;741 } else {742 struct alua_dh_data *h;743 744 /* Transitioning time exceeded, set port to standby */745 err = SCSI_DH_IO;746 pg->state = SCSI_ACCESS_STATE_STANDBY;747 pg->expiry = 0;748 rcu_read_lock();749 list_for_each_entry_rcu(h, &pg->dh_list, node) {750 if (!h->sdev)751 continue;752 h->sdev->access_state =753 (pg->state & SCSI_ACCESS_STATE_MASK);754 if (pg->pref)755 h->sdev->access_state |=756 SCSI_ACCESS_STATE_PREFERRED;757 }758 rcu_read_unlock();759 }760 break;761 case SCSI_ACCESS_STATE_OFFLINE:762 /* Path unusable */763 err = SCSI_DH_DEV_OFFLINED;764 pg->expiry = 0;765 break;766 default:767 /* Useable path if active */768 err = SCSI_DH_OK;769 pg->expiry = 0;770 break;771 }772 spin_unlock_irqrestore(&pg->lock, flags);773 kfree(buff);774 return err;775}776 777/*778 * alua_stpg - Issue a SET TARGET PORT GROUP command779 *780 * Issue a SET TARGET PORT GROUP command and evaluate the781 * response. Returns SCSI_DH_RETRY per default to trigger782 * a re-evaluation of the target group state or SCSI_DH_OK783 * if no further action needs to be taken.784 */785static unsigned alua_stpg(struct scsi_device *sdev, struct alua_port_group *pg)786{787 int retval;788 struct scsi_sense_hdr sense_hdr;789 790 if (!(pg->tpgs & TPGS_MODE_EXPLICIT)) {791 /* Only implicit ALUA supported, retry */792 return SCSI_DH_RETRY;793 }794 switch (pg->state) {795 case SCSI_ACCESS_STATE_OPTIMAL:796 return SCSI_DH_OK;797 case SCSI_ACCESS_STATE_ACTIVE:798 if ((pg->flags & ALUA_OPTIMIZE_STPG) &&799 !pg->pref &&800 (pg->tpgs & TPGS_MODE_IMPLICIT))801 return SCSI_DH_OK;802 break;803 case SCSI_ACCESS_STATE_STANDBY:804 case SCSI_ACCESS_STATE_UNAVAILABLE:805 break;806 case SCSI_ACCESS_STATE_OFFLINE:807 return SCSI_DH_IO;808 case SCSI_ACCESS_STATE_TRANSITIONING:809 break;810 default:811 sdev_printk(KERN_INFO, sdev,812 "%s: stpg failed, unhandled TPGS state %d",813 ALUA_DH_NAME, pg->state);814 return SCSI_DH_NOSYS;815 }816 retval = submit_stpg(sdev, pg->group_id, &sense_hdr);817 818 if (retval) {819 if (retval < 0 || !scsi_sense_valid(&sense_hdr)) {820 sdev_printk(KERN_INFO, sdev,821 "%s: stpg failed, result %d",822 ALUA_DH_NAME, retval);823 if (retval < 0)824 return SCSI_DH_DEV_TEMP_BUSY;825 } else {826 sdev_printk(KERN_INFO, sdev, "%s: stpg failed\n",827 ALUA_DH_NAME);828 scsi_print_sense_hdr(sdev, ALUA_DH_NAME, &sense_hdr);829 }830 }831 /* Retry RTPG */832 return SCSI_DH_RETRY;833}834 835/*836 * The caller must call scsi_device_put() on the returned pointer if it is not837 * NULL.838 */839static struct scsi_device * __must_check840alua_rtpg_select_sdev(struct alua_port_group *pg)841{842 struct alua_dh_data *h;843 struct scsi_device *sdev = NULL, *prev_sdev;844 845 lockdep_assert_held(&pg->lock);846 if (WARN_ON(!pg->rtpg_sdev))847 return NULL;848 849 /*850 * RCU protection isn't necessary for dh_list here851 * as we hold pg->lock, but for access to h->pg.852 */853 rcu_read_lock();854 list_for_each_entry_rcu(h, &pg->dh_list, node) {855 if (!h->sdev)856 continue;857 if (h->sdev == pg->rtpg_sdev) {858 h->disabled = true;859 continue;860 }861 if (rcu_dereference(h->pg) == pg &&862 !h->disabled &&863 !scsi_device_get(h->sdev)) {864 sdev = h->sdev;865 break;866 }867 }868 rcu_read_unlock();869 870 if (!sdev) {871 pr_warn("%s: no device found for rtpg\n",872 (pg->device_id_len ?873 (char *)pg->device_id_str : "(nameless PG)"));874 return NULL;875 }876 877 sdev_printk(KERN_INFO, sdev, "rtpg retry on different device\n");878 879 prev_sdev = pg->rtpg_sdev;880 pg->rtpg_sdev = sdev;881 882 return prev_sdev;883}884 885static void alua_rtpg_work(struct work_struct *work)886{887 struct alua_port_group *pg =888 container_of(work, struct alua_port_group, rtpg_work.work);889 struct scsi_device *sdev, *prev_sdev = NULL;890 LIST_HEAD(qdata_list);891 int err = SCSI_DH_OK;892 struct alua_queue_data *qdata, *tmp;893 struct alua_dh_data *h;894 unsigned long flags;895 896 spin_lock_irqsave(&pg->lock, flags);897 sdev = pg->rtpg_sdev;898 if (!sdev) {899 WARN_ON(pg->flags & ALUA_PG_RUN_RTPG);900 WARN_ON(pg->flags & ALUA_PG_RUN_STPG);901 spin_unlock_irqrestore(&pg->lock, flags);902 kref_put(&pg->kref, release_port_group);903 return;904 }905 pg->flags |= ALUA_PG_RUNNING;906 if (pg->flags & ALUA_PG_RUN_RTPG) {907 int state = pg->state;908 909 pg->flags &= ~ALUA_PG_RUN_RTPG;910 spin_unlock_irqrestore(&pg->lock, flags);911 if (state == SCSI_ACCESS_STATE_TRANSITIONING) {912 if (alua_tur(sdev) == SCSI_DH_RETRY) {913 spin_lock_irqsave(&pg->lock, flags);914 pg->flags &= ~ALUA_PG_RUNNING;915 pg->flags |= ALUA_PG_RUN_RTPG;916 if (!pg->interval)917 pg->interval = ALUA_RTPG_RETRY_DELAY;918 spin_unlock_irqrestore(&pg->lock, flags);919 queue_delayed_work(kaluad_wq, &pg->rtpg_work,920 pg->interval * HZ);921 return;922 }923 /* Send RTPG on failure or if TUR indicates SUCCESS */924 }925 err = alua_rtpg(sdev, pg);926 spin_lock_irqsave(&pg->lock, flags);927 928 /* If RTPG failed on the current device, try using another */929 if (err == SCSI_DH_RES_TEMP_UNAVAIL &&930 (prev_sdev = alua_rtpg_select_sdev(pg)))931 err = SCSI_DH_IMM_RETRY;932 933 if (err == SCSI_DH_RETRY || err == SCSI_DH_IMM_RETRY ||934 pg->flags & ALUA_PG_RUN_RTPG) {935 pg->flags &= ~ALUA_PG_RUNNING;936 if (err == SCSI_DH_IMM_RETRY)937 pg->interval = 0;938 else if (!pg->interval && !(pg->flags & ALUA_PG_RUN_RTPG))939 pg->interval = ALUA_RTPG_RETRY_DELAY;940 pg->flags |= ALUA_PG_RUN_RTPG;941 spin_unlock_irqrestore(&pg->lock, flags);942 goto queue_rtpg;943 }944 if (err != SCSI_DH_OK)945 pg->flags &= ~ALUA_PG_RUN_STPG;946 }947 if (pg->flags & ALUA_PG_RUN_STPG) {948 pg->flags &= ~ALUA_PG_RUN_STPG;949 spin_unlock_irqrestore(&pg->lock, flags);950 err = alua_stpg(sdev, pg);951 spin_lock_irqsave(&pg->lock, flags);952 if (err == SCSI_DH_RETRY || pg->flags & ALUA_PG_RUN_RTPG) {953 pg->flags |= ALUA_PG_RUN_RTPG;954 pg->interval = 0;955 pg->flags &= ~ALUA_PG_RUNNING;956 spin_unlock_irqrestore(&pg->lock, flags);957 goto queue_rtpg;958 }959 }960 961 list_splice_init(&pg->rtpg_list, &qdata_list);962 /*963 * We went through an RTPG, for good or bad.964 * Re-enable all devices for the next attempt.965 */966 list_for_each_entry(h, &pg->dh_list, node)967 h->disabled = false;968 pg->rtpg_sdev = NULL;969 spin_unlock_irqrestore(&pg->lock, flags);970 971 if (prev_sdev)972 scsi_device_put(prev_sdev);973 974 list_for_each_entry_safe(qdata, tmp, &qdata_list, entry) {975 list_del(&qdata->entry);976 if (qdata->callback_fn)977 qdata->callback_fn(qdata->callback_data, err);978 kfree(qdata);979 }980 spin_lock_irqsave(&pg->lock, flags);981 pg->flags &= ~ALUA_PG_RUNNING;982 spin_unlock_irqrestore(&pg->lock, flags);983 scsi_device_put(sdev);984 kref_put(&pg->kref, release_port_group);985 return;986 987queue_rtpg:988 if (prev_sdev)989 scsi_device_put(prev_sdev);990 queue_delayed_work(kaluad_wq, &pg->rtpg_work, pg->interval * HZ);991}992 993/**994 * alua_rtpg_queue() - cause RTPG to be submitted asynchronously995 * @pg: ALUA port group associated with @sdev.996 * @sdev: SCSI device for which to submit an RTPG.997 * @qdata: Information about the callback to invoke after the RTPG.998 * @force: Whether or not to submit an RTPG if a work item that will submit an999 * RTPG already has been scheduled.1000 *1001 * Returns true if and only if alua_rtpg_work() will be called asynchronously.1002 * That function is responsible for calling @qdata->fn().1003 *1004 * Context: may be called from atomic context (alua_check()) only if the caller1005 * holds an sdev reference.1006 */1007static bool alua_rtpg_queue(struct alua_port_group *pg,1008 struct scsi_device *sdev,1009 struct alua_queue_data *qdata, bool force)1010{1011 int start_queue = 0;1012 unsigned long flags;1013 1014 if (WARN_ON_ONCE(!pg) || scsi_device_get(sdev))1015 return false;1016 1017 spin_lock_irqsave(&pg->lock, flags);1018 if (qdata) {1019 list_add_tail(&qdata->entry, &pg->rtpg_list);1020 pg->flags |= ALUA_PG_RUN_STPG;1021 force = true;1022 }1023 if (pg->rtpg_sdev == NULL) {1024 struct alua_dh_data *h = sdev->handler_data;1025 1026 rcu_read_lock();1027 if (h && rcu_dereference(h->pg) == pg) {1028 pg->interval = 0;1029 pg->flags |= ALUA_PG_RUN_RTPG;1030 kref_get(&pg->kref);1031 pg->rtpg_sdev = sdev;1032 start_queue = 1;1033 }1034 rcu_read_unlock();1035 } else if (!(pg->flags & ALUA_PG_RUN_RTPG) && force) {1036 pg->flags |= ALUA_PG_RUN_RTPG;1037 /* Do not queue if the worker is already running */1038 if (!(pg->flags & ALUA_PG_RUNNING)) {1039 kref_get(&pg->kref);1040 start_queue = 1;1041 }1042 }1043 1044 spin_unlock_irqrestore(&pg->lock, flags);1045 1046 if (start_queue) {1047 if (queue_delayed_work(kaluad_wq, &pg->rtpg_work,1048 msecs_to_jiffies(ALUA_RTPG_DELAY_MSECS)))1049 sdev = NULL;1050 else1051 kref_put(&pg->kref, release_port_group);1052 }1053 if (sdev)1054 scsi_device_put(sdev);1055 1056 return true;1057}1058 1059/*1060 * alua_initialize - Initialize ALUA state1061 * @sdev: the device to be initialized1062 *1063 * For the prep_fn to work correctly we have1064 * to initialize the ALUA state for the device.1065 */1066static int alua_initialize(struct scsi_device *sdev, struct alua_dh_data *h)1067{1068 int err = SCSI_DH_DEV_UNSUPP, tpgs;1069 1070 mutex_lock(&h->init_mutex);1071 h->disabled = false;1072 tpgs = alua_check_tpgs(sdev);1073 if (tpgs != TPGS_MODE_NONE)1074 err = alua_check_vpd(sdev, h, tpgs);1075 h->init_error = err;1076 mutex_unlock(&h->init_mutex);1077 return err;1078}1079/*1080 * alua_set_params - set/unset the optimize flag1081 * @sdev: device on the path to be activated1082 * params - parameters in the following format1083 * "no_of_params\0param1\0param2\0param3\0...\0"1084 * For example, to set the flag pass the following parameters1085 * from multipath.conf1086 * hardware_handler "2 alua 1"1087 */1088static int alua_set_params(struct scsi_device *sdev, const char *params)1089{1090 struct alua_dh_data *h = sdev->handler_data;1091 struct alua_port_group *pg = NULL;1092 unsigned int optimize = 0, argc;1093 const char *p = params;1094 int result = SCSI_DH_OK;1095 unsigned long flags;1096 1097 if ((sscanf(params, "%u", &argc) != 1) || (argc != 1))1098 return -EINVAL;1099 1100 while (*p++)1101 ;1102 if ((sscanf(p, "%u", &optimize) != 1) || (optimize > 1))1103 return -EINVAL;1104 1105 rcu_read_lock();1106 pg = rcu_dereference(h->pg);1107 if (!pg) {1108 rcu_read_unlock();1109 return -ENXIO;1110 }1111 spin_lock_irqsave(&pg->lock, flags);1112 if (optimize)1113 pg->flags |= ALUA_OPTIMIZE_STPG;1114 else1115 pg->flags &= ~ALUA_OPTIMIZE_STPG;1116 spin_unlock_irqrestore(&pg->lock, flags);1117 rcu_read_unlock();1118 1119 return result;1120}1121 1122/*1123 * alua_activate - activate a path1124 * @sdev: device on the path to be activated1125 *1126 * We're currently switching the port group to be activated only and1127 * let the array figure out the rest.1128 * There may be other arrays which require us to switch all port groups1129 * based on a certain policy. But until we actually encounter them it1130 * should be okay.1131 */1132static int alua_activate(struct scsi_device *sdev,1133 activate_complete fn, void *data)1134{1135 struct alua_dh_data *h = sdev->handler_data;1136 int err = SCSI_DH_OK;1137 struct alua_queue_data *qdata;1138 struct alua_port_group *pg;1139 1140 qdata = kzalloc(sizeof(*qdata), GFP_KERNEL);1141 if (!qdata) {1142 err = SCSI_DH_RES_TEMP_UNAVAIL;1143 goto out;1144 }1145 qdata->callback_fn = fn;1146 qdata->callback_data = data;1147 1148 mutex_lock(&h->init_mutex);1149 rcu_read_lock();1150 pg = rcu_dereference(h->pg);1151 if (!pg || !kref_get_unless_zero(&pg->kref)) {1152 rcu_read_unlock();1153 kfree(qdata);1154 err = h->init_error;1155 mutex_unlock(&h->init_mutex);1156 goto out;1157 }1158 rcu_read_unlock();1159 mutex_unlock(&h->init_mutex);1160 1161 if (alua_rtpg_queue(pg, sdev, qdata, true)) {1162 fn = NULL;1163 } else {1164 kfree(qdata);1165 err = SCSI_DH_DEV_OFFLINED;1166 }1167 kref_put(&pg->kref, release_port_group);1168out:1169 if (fn)1170 fn(data, err);1171 return 0;1172}1173 1174/*1175 * alua_check - check path status1176 * @sdev: device on the path to be checked1177 *1178 * Check the device status1179 */1180static void alua_check(struct scsi_device *sdev, bool force)1181{1182 struct alua_dh_data *h = sdev->handler_data;1183 struct alua_port_group *pg;1184 1185 rcu_read_lock();1186 pg = rcu_dereference(h->pg);1187 if (!pg || !kref_get_unless_zero(&pg->kref)) {1188 rcu_read_unlock();1189 return;1190 }1191 rcu_read_unlock();1192 alua_rtpg_queue(pg, sdev, NULL, force);1193 kref_put(&pg->kref, release_port_group);1194}1195 1196/*1197 * alua_prep_fn - request callback1198 *1199 * Fail I/O to all paths not in state1200 * active/optimized or active/non-optimized.1201 */1202static blk_status_t alua_prep_fn(struct scsi_device *sdev, struct request *req)1203{1204 struct alua_dh_data *h = sdev->handler_data;1205 struct alua_port_group *pg;1206 unsigned char state = SCSI_ACCESS_STATE_OPTIMAL;1207 1208 rcu_read_lock();1209 pg = rcu_dereference(h->pg);1210 if (pg)1211 state = pg->state;1212 rcu_read_unlock();1213 1214 switch (state) {1215 case SCSI_ACCESS_STATE_OPTIMAL:1216 case SCSI_ACCESS_STATE_ACTIVE:1217 case SCSI_ACCESS_STATE_LBA:1218 case SCSI_ACCESS_STATE_TRANSITIONING:1219 return BLK_STS_OK;1220 default:1221 req->rq_flags |= RQF_QUIET;1222 return BLK_STS_IOERR;1223 }1224}1225 1226static void alua_rescan(struct scsi_device *sdev)1227{1228 struct alua_dh_data *h = sdev->handler_data;1229 1230 alua_initialize(sdev, h);1231}1232 1233/*1234 * alua_bus_attach - Attach device handler1235 * @sdev: device to be attached to1236 */1237static int alua_bus_attach(struct scsi_device *sdev)1238{1239 struct alua_dh_data *h;1240 int err;1241 1242 h = kzalloc(sizeof(*h) , GFP_KERNEL);1243 if (!h)1244 return SCSI_DH_NOMEM;1245 spin_lock_init(&h->pg_lock);1246 rcu_assign_pointer(h->pg, NULL);1247 h->init_error = SCSI_DH_OK;1248 h->sdev = sdev;1249 INIT_LIST_HEAD(&h->node);1250 1251 mutex_init(&h->init_mutex);1252 err = alua_initialize(sdev, h);1253 if (err != SCSI_DH_OK && err != SCSI_DH_DEV_OFFLINED)1254 goto failed;1255 1256 sdev->handler_data = h;1257 return SCSI_DH_OK;1258failed:1259 kfree(h);1260 return err;1261}1262 1263/*1264 * alua_bus_detach - Detach device handler1265 * @sdev: device to be detached from1266 */1267static void alua_bus_detach(struct scsi_device *sdev)1268{1269 struct alua_dh_data *h = sdev->handler_data;1270 struct alua_port_group *pg;1271 1272 spin_lock(&h->pg_lock);1273 pg = rcu_dereference_protected(h->pg, lockdep_is_held(&h->pg_lock));1274 rcu_assign_pointer(h->pg, NULL);1275 spin_unlock(&h->pg_lock);1276 if (pg) {1277 spin_lock_irq(&pg->lock);1278 list_del_rcu(&h->node);1279 spin_unlock_irq(&pg->lock);1280 kref_put(&pg->kref, release_port_group);1281 }1282 sdev->handler_data = NULL;1283 synchronize_rcu();1284 kfree(h);1285}1286 1287static struct scsi_device_handler alua_dh = {1288 .name = ALUA_DH_NAME,1289 .module = THIS_MODULE,1290 .attach = alua_bus_attach,1291 .detach = alua_bus_detach,1292 .prep_fn = alua_prep_fn,1293 .check_sense = alua_check_sense,1294 .activate = alua_activate,1295 .rescan = alua_rescan,1296 .set_params = alua_set_params,1297};1298 1299static int __init alua_init(void)1300{1301 int r;1302 1303 kaluad_wq = alloc_workqueue("kaluad", WQ_MEM_RECLAIM, 0);1304 if (!kaluad_wq)1305 return -ENOMEM;1306 1307 r = scsi_register_device_handler(&alua_dh);1308 if (r != 0) {1309 printk(KERN_ERR "%s: Failed to register scsi device handler",1310 ALUA_DH_NAME);1311 destroy_workqueue(kaluad_wq);1312 }1313 return r;1314}1315 1316static void __exit alua_exit(void)1317{1318 scsi_unregister_device_handler(&alua_dh);1319 destroy_workqueue(kaluad_wq);1320}1321 1322module_init(alua_init);1323module_exit(alua_exit);1324 1325MODULE_DESCRIPTION("DM Multipath ALUA support");1326MODULE_AUTHOR("Hannes Reinecke <hare@suse.de>");1327MODULE_LICENSE("GPL");1328MODULE_VERSION(ALUA_DH_VER);1329