drivers/block/rnbd/rnbd-clt.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * RDMA Network Block Driver
   4  *
   5  * Copyright (c) 2014 - 2018 ProfitBricks GmbH. All rights reserved.
   6  * Copyright (c) 2018 - 2019 1&1 IONOS Cloud GmbH. All rights reserved.
   7  * Copyright (c) 2019 - 2020 1&1 IONOS SE. All rights reserved.
   8  */
   9
  10 #undef pr_fmt
  11 #define pr_fmt(fmt) KBUILD_MODNAME " L" __stringify(__LINE__) ": " fmt
  12
  13 #include <linux/module.h>
  14 #include <linux/blkdev.h>
  15 #include <linux/hdreg.h>
  16 #include <linux/scatterlist.h>
  17 #include <linux/idr.h>
  18
  19 #include "rnbd-clt.h"
  20
  21 MODULE_DESCRIPTION("RDMA Network Block Device Client");
  22 MODULE_LICENSE("GPL");
  23
  24 static int rnbd_client_major;
  25 static DEFINE_IDA(index_ida);
  26 static DEFINE_MUTEX(ida_lock);
  27 static DEFINE_MUTEX(sess_lock);
  28 static LIST_HEAD(sess_list);
  29
  30 /*
  31  * Maximum number of partitions an instance can have.
  32  * 6 bits = 64 minors = 63 partitions (one minor is used for the device itself)
  33  */
  34 #define RNBD_PART_BITS          6
  35
  36 static inline bool rnbd_clt_get_sess(struct rnbd_clt_session *sess)
  37 {
  38         return refcount_inc_not_zero(&sess->refcount);
  39 }
  40
  41 static void free_sess(struct rnbd_clt_session *sess);
  42
  43 static void rnbd_clt_put_sess(struct rnbd_clt_session *sess)
  44 {
  45         might_sleep();
  46
  47         if (refcount_dec_and_test(&sess->refcount))
  48                 free_sess(sess);
  49 }
  50
  51 static void rnbd_clt_put_dev(struct rnbd_clt_dev *dev)
  52 {
  53         might_sleep();
  54
  55         if (!refcount_dec_and_test(&dev->refcount))
  56                 return;
  57
  58         mutex_lock(&ida_lock);
  59         ida_simple_remove(&index_ida, dev->clt_device_id);
  60         mutex_unlock(&ida_lock);
  61         kfree(dev->hw_queues);
  62         kfree(dev->pathname);
  63         rnbd_clt_put_sess(dev->sess);
  64         mutex_destroy(&dev->lock);
  65         kfree(dev);
  66 }
  67
  68 static inline bool rnbd_clt_get_dev(struct rnbd_clt_dev *dev)
  69 {
  70         return refcount_inc_not_zero(&dev->refcount);
  71 }
  72
  73 static int rnbd_clt_set_dev_attr(struct rnbd_clt_dev *dev,
  74                                  const struct rnbd_msg_open_rsp *rsp)
  75 {
  76         struct rnbd_clt_session *sess = dev->sess;
  77
  78         if (!rsp->logical_block_size)
  79                 return -EINVAL;
  80
  81         dev->device_id              = le32_to_cpu(rsp->device_id);
  82         dev->nsectors               = le64_to_cpu(rsp->nsectors);
  83         dev->logical_block_size     = le16_to_cpu(rsp->logical_block_size);
  84         dev->physical_block_size    = le16_to_cpu(rsp->physical_block_size);
  85         dev->max_write_same_sectors = le32_to_cpu(rsp->max_write_same_sectors);
  86         dev->max_discard_sectors    = le32_to_cpu(rsp->max_discard_sectors);
  87         dev->discard_granularity    = le32_to_cpu(rsp->discard_granularity);
  88         dev->discard_alignment      = le32_to_cpu(rsp->discard_alignment);
  89         dev->secure_discard         = le16_to_cpu(rsp->secure_discard);
  90         dev->rotational             = rsp->rotational;
  91         dev->wc                     = !!(rsp->cache_policy & RNBD_WRITEBACK);
  92         dev->fua                    = !!(rsp->cache_policy & RNBD_FUA);
  93
  94         dev->max_hw_sectors = sess->max_io_size / SECTOR_SIZE;
  95         dev->max_segments = sess->max_segments;
  96
  97         return 0;
  98 }
  99
 100 static int rnbd_clt_change_capacity(struct rnbd_clt_dev *dev,
 101                                     size_t new_nsectors)
 102 {
 103         rnbd_clt_info(dev, "Device size changed from %zu to %zu sectors\n",
 104                        dev->nsectors, new_nsectors);
 105         dev->nsectors = new_nsectors;
 106         set_capacity_and_notify(dev->gd, dev->nsectors);
 107         return 0;
 108 }
 109
 110 static int process_msg_open_rsp(struct rnbd_clt_dev *dev,
 111                                 struct rnbd_msg_open_rsp *rsp)
 112 {
 113         struct kobject *gd_kobj;
 114         int err = 0;
 115
 116         mutex_lock(&dev->lock);
 117         if (dev->dev_state == DEV_STATE_UNMAPPED) {
 118                 rnbd_clt_info(dev,
 119                                "Ignoring Open-Response message from server for  unmapped device\n");
 120                 err = -ENOENT;
 121                 goto out;
 122         }
 123         if (dev->dev_state == DEV_STATE_MAPPED_DISCONNECTED) {
 124                 u64 nsectors = le64_to_cpu(rsp->nsectors);
 125
 126                 /*
 127                  * If the device was remapped and the size changed in the
 128                  * meantime we need to revalidate it
 129                  */
 130                 if (dev->nsectors != nsectors)
 131                         rnbd_clt_change_capacity(dev, nsectors);
 132                 gd_kobj = &disk_to_dev(dev->gd)->kobj;
 133                 kobject_uevent(gd_kobj, KOBJ_ONLINE);
 134                 rnbd_clt_info(dev, "Device online, device remapped successfully\n");
 135         }
 136         err = rnbd_clt_set_dev_attr(dev, rsp);
 137         if (err)
 138                 goto out;
 139         dev->dev_state = DEV_STATE_MAPPED;
 140
 141 out:
 142         mutex_unlock(&dev->lock);
 143
 144         return err;
 145 }
 146
 147 int rnbd_clt_resize_disk(struct rnbd_clt_dev *dev, size_t newsize)
 148 {
 149         int ret = 0;
 150
 151         mutex_lock(&dev->lock);
 152         if (dev->dev_state != DEV_STATE_MAPPED) {
 153                 pr_err("Failed to set new size of the device, device is not opened\n");
 154                 ret = -ENOENT;
 155                 goto out;
 156         }
 157         ret = rnbd_clt_change_capacity(dev, newsize);
 158
 159 out:
 160         mutex_unlock(&dev->lock);
 161
 162         return ret;
 163 }
 164
 165 static inline void rnbd_clt_dev_requeue(struct rnbd_queue *q)
 166 {
 167         if (WARN_ON(!q->hctx))
 168                 return;
 169
 170         /* We can come here from interrupt, thus async=true */
 171         blk_mq_run_hw_queue(q->hctx, true);
 172 }
 173
 174 enum {
 175         RNBD_DELAY_IFBUSY = -1,
 176 };
 177
 178 /**
 179  * rnbd_get_cpu_qlist() - finds a list with HW queues to be rerun
 180  * @sess:       Session to find a queue for
 181  * @cpu:        Cpu to start the search from
 182  *
 183  * Description:
 184  *     Each CPU has a list of HW queues, which needs to be rerun.  If a list
 185  *     is not empty - it is marked with a bit.  This function finds first
 186  *     set bit in a bitmap and returns corresponding CPU list.
 187  */
 188 static struct rnbd_cpu_qlist *
 189 rnbd_get_cpu_qlist(struct rnbd_clt_session *sess, int cpu)
 190 {
 191         int bit;
 192
 193         /* Search from cpu to nr_cpu_ids */
 194         bit = find_next_bit(sess->cpu_queues_bm, nr_cpu_ids, cpu);
 195         if (bit < nr_cpu_ids) {
 196                 return per_cpu_ptr(sess->cpu_queues, bit);
 197         } else if (cpu != 0) {
 198                 /* Search from 0 to cpu */
 199                 bit = find_next_bit(sess->cpu_queues_bm, cpu, 0);
 200                 if (bit < cpu)
 201                         return per_cpu_ptr(sess->cpu_queues, bit);
 202         }
 203
 204         return NULL;
 205 }
 206
 207 static inline int nxt_cpu(int cpu)
 208 {
 209         return (cpu + 1) % nr_cpu_ids;
 210 }
 211
 212 /**
 213  * rnbd_rerun_if_needed() - rerun next queue marked as stopped
 214  * @sess:       Session to rerun a queue on
 215  *
 216  * Description:
 217  *     Each CPU has it's own list of HW queues, which should be rerun.
 218  *     Function finds such list with HW queues, takes a list lock, picks up
 219  *     the first HW queue out of the list and requeues it.
 220  *
 221  * Return:
 222  *     True if the queue was requeued, false otherwise.
 223  *
 224  * Context:
 225  *     Does not matter.
 226  */
 227 static bool rnbd_rerun_if_needed(struct rnbd_clt_session *sess)
 228 {
 229         struct rnbd_queue *q = NULL;
 230         struct rnbd_cpu_qlist *cpu_q;
 231         unsigned long flags;
 232         int *cpup;
 233
 234         /*
 235          * To keep fairness and not to let other queues starve we always
 236          * try to wake up someone else in round-robin manner.  That of course
 237          * increases latency but queues always have a chance to be executed.
 238          */
 239         cpup = get_cpu_ptr(sess->cpu_rr);
 240         for (cpu_q = rnbd_get_cpu_qlist(sess, nxt_cpu(*cpup)); cpu_q;
 241              cpu_q = rnbd_get_cpu_qlist(sess, nxt_cpu(cpu_q->cpu))) {
 242                 if (!spin_trylock_irqsave(&cpu_q->requeue_lock, flags))
 243                         continue;
 244                 if (!test_bit(cpu_q->cpu, sess->cpu_queues_bm))
 245                         goto unlock;
 246                 q = list_first_entry_or_null(&cpu_q->requeue_list,
 247                                              typeof(*q), requeue_list);
 248                 if (WARN_ON(!q))
 249                         goto clear_bit;
 250                 list_del_init(&q->requeue_list);
 251                 clear_bit_unlock(0, &q->in_list);
 252
 253                 if (list_empty(&cpu_q->requeue_list)) {
 254                         /* Clear bit if nothing is left */
 255 clear_bit:
 256                         clear_bit(cpu_q->cpu, sess->cpu_queues_bm);
 257                 }
 258 unlock:
 259                 spin_unlock_irqrestore(&cpu_q->requeue_lock, flags);
 260
 261                 if (q)
 262                         break;
 263         }
 264
 265         /**
 266          * Saves the CPU that is going to be requeued on the per-cpu var. Just
 267          * incrementing it doesn't work because rnbd_get_cpu_qlist() will
 268          * always return the first CPU with something on the queue list when the
 269          * value stored on the var is greater than the last CPU with something
 270          * on the list.
 271          */
 272         if (cpu_q)
 273                 *cpup = cpu_q->cpu;
 274         put_cpu_ptr(sess->cpu_rr);
 275
 276         if (q)
 277                 rnbd_clt_dev_requeue(q);
 278
 279         return q;
 280 }
 281
 282 /**
 283  * rnbd_rerun_all_if_idle() - rerun all queues left in the list if
 284  *                               session is idling (there are no requests
 285  *                               in-flight).
 286  * @sess:       Session to rerun the queues on
 287  *
 288  * Description:
 289  *     This function tries to rerun all stopped queues if there are no
 290  *     requests in-flight anymore.  This function tries to solve an obvious
 291  *     problem, when number of tags < than number of queues (hctx), which
 292  *     are stopped and put to sleep.  If last permit, which has been just put,
 293  *     does not wake up all left queues (hctxs), IO requests hang forever.
 294  *
 295  *     That can happen when all number of permits, say N, have been exhausted
 296  *     from one CPU, and we have many block devices per session, say M.
 297  *     Each block device has it's own queue (hctx) for each CPU, so eventually
 298  *     we can put that number of queues (hctxs) to sleep: M x nr_cpu_ids.
 299  *     If number of permits N < M x nr_cpu_ids finally we will get an IO hang.
 300  *
 301  *     To avoid this hang last caller of rnbd_put_permit() (last caller is the
 302  *     one who observes sess->busy == 0) must wake up all remaining queues.
 303  *
 304  * Context:
 305  *     Does not matter.
 306  */
 307 static void rnbd_rerun_all_if_idle(struct rnbd_clt_session *sess)
 308 {
 309         bool requeued;
 310
 311         do {
 312                 requeued = rnbd_rerun_if_needed(sess);
 313         } while (atomic_read(&sess->busy) == 0 && requeued);
 314 }
 315
 316 static struct rtrs_permit *rnbd_get_permit(struct rnbd_clt_session *sess,
 317                                              enum rtrs_clt_con_type con_type,
 318                                              enum wait_type wait)
 319 {
 320         struct rtrs_permit *permit;
 321
 322         permit = rtrs_clt_get_permit(sess->rtrs, con_type, wait);
 323         if (permit)
 324                 /* We have a subtle rare case here, when all permits can be
 325                  * consumed before busy counter increased.  This is safe,
 326                  * because loser will get NULL as a permit, observe 0 busy
 327                  * counter and immediately restart the queue himself.
 328                  */
 329                 atomic_inc(&sess->busy);
 330
 331         return permit;
 332 }
 333
 334 static void rnbd_put_permit(struct rnbd_clt_session *sess,
 335                              struct rtrs_permit *permit)
 336 {
 337         rtrs_clt_put_permit(sess->rtrs, permit);
 338         atomic_dec(&sess->busy);
 339         /* Paired with rnbd_clt_dev_add_to_requeue().  Decrement first
 340          * and then check queue bits.
 341          */
 342         smp_mb__after_atomic();
 343         rnbd_rerun_all_if_idle(sess);
 344 }
 345
 346 static struct rnbd_iu *rnbd_get_iu(struct rnbd_clt_session *sess,
 347                                      enum rtrs_clt_con_type con_type,
 348                                      enum wait_type wait)
 349 {
 350         struct rnbd_iu *iu;
 351         struct rtrs_permit *permit;
 352
 353         iu = kzalloc(sizeof(*iu), GFP_KERNEL);
 354         if (!iu)
 355                 return NULL;
 356
 357         permit = rnbd_get_permit(sess, con_type, wait);
 358         if (!permit) {
 359                 kfree(iu);
 360                 return NULL;
 361         }
 362
 363         iu->permit = permit;
 364         /*
 365          * 1st reference is dropped after finishing sending a "user" message,
 366          * 2nd reference is dropped after confirmation with the response is
 367          * returned.
 368          * 1st and 2nd can happen in any order, so the rnbd_iu should be
 369          * released (rtrs_permit returned to rtrs) only after both
 370          * are finished.
 371          */
 372         atomic_set(&iu->refcount, 2);
 373         init_waitqueue_head(&iu->comp.wait);
 374         iu->comp.errno = INT_MAX;
 375
 376         if (sg_alloc_table(&iu->sgt, 1, GFP_KERNEL)) {
 377                 rnbd_put_permit(sess, permit);
 378                 kfree(iu);
 379                 return NULL;
 380         }
 381
 382         return iu;
 383 }
 384
 385 static void rnbd_put_iu(struct rnbd_clt_session *sess, struct rnbd_iu *iu)
 386 {
 387         if (atomic_dec_and_test(&iu->refcount)) {
 388                 sg_free_table(&iu->sgt);
 389                 rnbd_put_permit(sess, iu->permit);
 390                 kfree(iu);
 391         }
 392 }
 393
 394 static void rnbd_softirq_done_fn(struct request *rq)
 395 {
 396         struct rnbd_clt_dev *dev        = rq->rq_disk->private_data;
 397         struct rnbd_clt_session *sess   = dev->sess;
 398         struct rnbd_iu *iu;
 399
 400         iu = blk_mq_rq_to_pdu(rq);
 401         sg_free_table_chained(&iu->sgt, RNBD_INLINE_SG_CNT);
 402         rnbd_put_permit(sess, iu->permit);
 403         blk_mq_end_request(rq, errno_to_blk_status(iu->errno));
 404 }
 405
 406 static void msg_io_conf(void *priv, int errno)
 407 {
 408         struct rnbd_iu *iu = priv;
 409         struct rnbd_clt_dev *dev = iu->dev;
 410         struct request *rq = iu->rq;
 411         int rw = rq_data_dir(rq);
 412
 413         iu->errno = errno;
 414
 415         blk_mq_complete_request(rq);
 416
 417         if (errno)
 418                 rnbd_clt_info_rl(dev, "%s I/O failed with err: %d\n",
 419                                  rw == READ ? "read" : "write", errno);
 420 }
 421
 422 static void wake_up_iu_comp(struct rnbd_iu *iu, int errno)
 423 {
 424         iu->comp.errno = errno;
 425         wake_up(&iu->comp.wait);
 426 }
 427
 428 static void msg_conf(void *priv, int errno)
 429 {
 430         struct rnbd_iu *iu = priv;
 431
 432         iu->errno = errno;
 433         schedule_work(&iu->work);
 434 }
 435
 436 static int send_usr_msg(struct rtrs_clt *rtrs, int dir,
 437                         struct rnbd_iu *iu, struct kvec *vec,
 438                         size_t len, struct scatterlist *sg, unsigned int sg_len,
 439                         void (*conf)(struct work_struct *work),
 440                         int *errno, int wait)
 441 {
 442         int err;
 443         struct rtrs_clt_req_ops req_ops;
 444
 445         INIT_WORK(&iu->work, conf);
 446         req_ops = (struct rtrs_clt_req_ops) {
 447                 .priv = iu,
 448                 .conf_fn = msg_conf,
 449         };
 450         err = rtrs_clt_request(dir, &req_ops, rtrs, iu->permit,
 451                                 vec, 1, len, sg, sg_len);
 452         if (!err && wait) {
 453                 wait_event(iu->comp.wait, iu->comp.errno != INT_MAX);
 454                 *errno = iu->comp.errno;
 455         } else {
 456                 *errno = 0;
 457         }
 458
 459         return err;
 460 }
 461
 462 static void msg_close_conf(struct work_struct *work)
 463 {
 464         struct rnbd_iu *iu = container_of(work, struct rnbd_iu, work);
 465         struct rnbd_clt_dev *dev = iu->dev;
 466
 467         wake_up_iu_comp(iu, iu->errno);
 468         rnbd_put_iu(dev->sess, iu);
 469         rnbd_clt_put_dev(dev);
 470 }
 471
 472 static int send_msg_close(struct rnbd_clt_dev *dev, u32 device_id,
 473                           enum wait_type wait)
 474 {
 475         struct rnbd_clt_session *sess = dev->sess;
 476         struct rnbd_msg_close msg;
 477         struct rnbd_iu *iu;
 478         struct kvec vec = {
 479                 .iov_base = &msg,
 480                 .iov_len  = sizeof(msg)
 481         };
 482         int err, errno;
 483
 484         iu = rnbd_get_iu(sess, RTRS_ADMIN_CON, RTRS_PERMIT_WAIT);
 485         if (!iu)
 486                 return -ENOMEM;
 487
 488         iu->buf = NULL;
 489         iu->dev = dev;
 490
 491         msg.hdr.type    = cpu_to_le16(RNBD_MSG_CLOSE);
 492         msg.device_id   = cpu_to_le32(device_id);
 493
 494         WARN_ON(!rnbd_clt_get_dev(dev));
 495         err = send_usr_msg(sess->rtrs, WRITE, iu, &vec, 0, NULL, 0,
 496                            msg_close_conf, &errno, wait);
 497         if (err) {
 498                 rnbd_clt_put_dev(dev);
 499                 rnbd_put_iu(sess, iu);
 500         } else {
 501                 err = errno;
 502         }
 503
 504         rnbd_put_iu(sess, iu);
 505         return err;
 506 }
 507
 508 static void msg_open_conf(struct work_struct *work)
 509 {
 510         struct rnbd_iu *iu = container_of(work, struct rnbd_iu, work);
 511         struct rnbd_msg_open_rsp *rsp = iu->buf;
 512         struct rnbd_clt_dev *dev = iu->dev;
 513         int errno = iu->errno;
 514
 515         if (errno) {
 516                 rnbd_clt_err(dev,
 517                               "Opening failed, server responded: %d\n",
 518                               errno);
 519         } else {
 520                 errno = process_msg_open_rsp(dev, rsp);
 521                 if (errno) {
 522                         u32 device_id = le32_to_cpu(rsp->device_id);
 523                         /*
 524                          * If server thinks its fine, but we fail to process
 525                          * then be nice and send a close to server.
 526                          */
 527                         send_msg_close(dev, device_id, RTRS_PERMIT_NOWAIT);
 528                 }
 529         }
 530         kfree(rsp);
 531         wake_up_iu_comp(iu, errno);
 532         rnbd_put_iu(dev->sess, iu);
 533         rnbd_clt_put_dev(dev);
 534 }
 535
 536 static void msg_sess_info_conf(struct work_struct *work)
 537 {
 538         struct rnbd_iu *iu = container_of(work, struct rnbd_iu, work);
 539         struct rnbd_msg_sess_info_rsp *rsp = iu->buf;
 540         struct rnbd_clt_session *sess = iu->sess;
 541
 542         if (!iu->errno)
 543                 sess->ver = min_t(u8, rsp->ver, RNBD_PROTO_VER_MAJOR);
 544
 545         kfree(rsp);
 546         wake_up_iu_comp(iu, iu->errno);
 547         rnbd_put_iu(sess, iu);
 548         rnbd_clt_put_sess(sess);
 549 }
 550
 551 static int send_msg_open(struct rnbd_clt_dev *dev, enum wait_type wait)
 552 {
 553         struct rnbd_clt_session *sess = dev->sess;
 554         struct rnbd_msg_open_rsp *rsp;
 555         struct rnbd_msg_open msg;
 556         struct rnbd_iu *iu;
 557         struct kvec vec = {
 558                 .iov_base = &msg,
 559                 .iov_len  = sizeof(msg)
 560         };
 561         int err, errno;
 562
 563         rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
 564         if (!rsp)
 565                 return -ENOMEM;
 566
 567         iu = rnbd_get_iu(sess, RTRS_ADMIN_CON, RTRS_PERMIT_WAIT);
 568         if (!iu) {
 569                 kfree(rsp);
 570                 return -ENOMEM;
 571         }
 572
 573         iu->buf = rsp;
 574         iu->dev = dev;
 575
 576         sg_init_one(iu->sgt.sgl, rsp, sizeof(*rsp));
 577
 578         msg.hdr.type    = cpu_to_le16(RNBD_MSG_OPEN);
 579         msg.access_mode = dev->access_mode;
 580         strscpy(msg.dev_name, dev->pathname, sizeof(msg.dev_name));
 581
 582         WARN_ON(!rnbd_clt_get_dev(dev));
 583         err = send_usr_msg(sess->rtrs, READ, iu,
 584                            &vec, sizeof(*rsp), iu->sgt.sgl, 1,
 585                            msg_open_conf, &errno, wait);
 586         if (err) {
 587                 rnbd_clt_put_dev(dev);
 588                 rnbd_put_iu(sess, iu);
 589                 kfree(rsp);
 590         } else {
 591                 err = errno;
 592         }
 593
 594         rnbd_put_iu(sess, iu);
 595         return err;
 596 }
 597
 598 static int send_msg_sess_info(struct rnbd_clt_session *sess, enum wait_type wait)
 599 {
 600         struct rnbd_msg_sess_info_rsp *rsp;
 601         struct rnbd_msg_sess_info msg;
 602         struct rnbd_iu *iu;
 603         struct kvec vec = {
 604                 .iov_base = &msg,
 605                 .iov_len  = sizeof(msg)
 606         };
 607         int err, errno;
 608
 609         rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
 610         if (!rsp)
 611                 return -ENOMEM;
 612
 613         iu = rnbd_get_iu(sess, RTRS_ADMIN_CON, RTRS_PERMIT_WAIT);
 614         if (!iu) {
 615                 kfree(rsp);
 616                 return -ENOMEM;
 617         }
 618
 619         iu->buf = rsp;
 620         iu->sess = sess;
 621         sg_init_one(iu->sgt.sgl, rsp, sizeof(*rsp));
 622
 623         msg.hdr.type = cpu_to_le16(RNBD_MSG_SESS_INFO);
 624         msg.ver      = RNBD_PROTO_VER_MAJOR;
 625
 626         if (!rnbd_clt_get_sess(sess)) {
 627                 /*
 628                  * That can happen only in one case, when RTRS has restablished
 629                  * the connection and link_ev() is called, but session is almost
 630                  * dead, last reference on session is put and caller is waiting
 631                  * for RTRS to close everything.
 632                  */
 633                 err = -ENODEV;
 634                 goto put_iu;
 635         }
 636         err = send_usr_msg(sess->rtrs, READ, iu,
 637                            &vec, sizeof(*rsp), iu->sgt.sgl, 1,
 638                            msg_sess_info_conf, &errno, wait);
 639         if (err) {
 640                 rnbd_clt_put_sess(sess);
 641 put_iu:
 642                 rnbd_put_iu(sess, iu);
 643                 kfree(rsp);
 644         } else {
 645                 err = errno;
 646         }
 647         rnbd_put_iu(sess, iu);
 648         return err;
 649 }
 650
 651 static void set_dev_states_to_disconnected(struct rnbd_clt_session *sess)
 652 {
 653         struct rnbd_clt_dev *dev;
 654         struct kobject *gd_kobj;
 655
 656         mutex_lock(&sess->lock);
 657         list_for_each_entry(dev, &sess->devs_list, list) {
 658                 rnbd_clt_err(dev, "Device disconnected.\n");
 659
 660                 mutex_lock(&dev->lock);
 661                 if (dev->dev_state == DEV_STATE_MAPPED) {
 662                         dev->dev_state = DEV_STATE_MAPPED_DISCONNECTED;
 663                         gd_kobj = &disk_to_dev(dev->gd)->kobj;
 664                         kobject_uevent(gd_kobj, KOBJ_OFFLINE);
 665                 }
 666                 mutex_unlock(&dev->lock);
 667         }
 668         mutex_unlock(&sess->lock);
 669 }
 670
 671 static void remap_devs(struct rnbd_clt_session *sess)
 672 {
 673         struct rnbd_clt_dev *dev;
 674         struct rtrs_attrs attrs;
 675         int err;
 676
 677         /*
 678          * Careful here: we are called from RTRS link event directly,
 679          * thus we can't send any RTRS request and wait for response
 680          * or RTRS will not be able to complete request with failure
 681          * if something goes wrong (failing of outstanding requests
 682          * happens exactly from the context where we are blocking now).
 683          *
 684          * So to avoid deadlocks each usr message sent from here must
 685          * be asynchronous.
 686          */
 687
 688         err = send_msg_sess_info(sess, RTRS_PERMIT_NOWAIT);
 689         if (err) {
 690                 pr_err("send_msg_sess_info(\"%s\"): %d\n", sess->sessname, err);
 691                 return;
 692         }
 693
 694         err = rtrs_clt_query(sess->rtrs, &attrs);
 695         if (err) {
 696                 pr_err("rtrs_clt_query(\"%s\"): %d\n", sess->sessname, err);
 697                 return;
 698         }
 699         mutex_lock(&sess->lock);
 700         sess->max_io_size = attrs.max_io_size;
 701
 702         list_for_each_entry(dev, &sess->devs_list, list) {
 703                 bool skip;
 704
 705                 mutex_lock(&dev->lock);
 706                 skip = (dev->dev_state == DEV_STATE_INIT);
 707                 mutex_unlock(&dev->lock);
 708                 if (skip)
 709                         /*
 710                          * When device is establishing connection for the first
 711                          * time - do not remap, it will be closed soon.
 712                          */
 713                         continue;
 714
 715                 rnbd_clt_info(dev, "session reconnected, remapping device\n");
 716                 err = send_msg_open(dev, RTRS_PERMIT_NOWAIT);
 717                 if (err) {
 718                         rnbd_clt_err(dev, "send_msg_open(): %d\n", err);
 719                         break;
 720                 }
 721         }
 722         mutex_unlock(&sess->lock);
 723 }
 724
 725 static void rnbd_clt_link_ev(void *priv, enum rtrs_clt_link_ev ev)
 726 {
 727         struct rnbd_clt_session *sess = priv;
 728
 729         switch (ev) {
 730         case RTRS_CLT_LINK_EV_DISCONNECTED:
 731                 set_dev_states_to_disconnected(sess);
 732                 break;
 733         case RTRS_CLT_LINK_EV_RECONNECTED:
 734                 remap_devs(sess);
 735                 break;
 736         default:
 737                 pr_err("Unknown session event received (%d), session: %s\n",
 738                        ev, sess->sessname);
 739         }
 740 }
 741
 742 static void rnbd_init_cpu_qlists(struct rnbd_cpu_qlist __percpu *cpu_queues)
 743 {
 744         unsigned int cpu;
 745         struct rnbd_cpu_qlist *cpu_q;
 746
 747         for_each_possible_cpu(cpu) {
 748                 cpu_q = per_cpu_ptr(cpu_queues, cpu);
 749
 750                 cpu_q->cpu = cpu;
 751                 INIT_LIST_HEAD(&cpu_q->requeue_list);
 752                 spin_lock_init(&cpu_q->requeue_lock);
 753         }
 754 }
 755
 756 static void destroy_mq_tags(struct rnbd_clt_session *sess)
 757 {
 758         if (sess->tag_set.tags)
 759                 blk_mq_free_tag_set(&sess->tag_set);
 760 }
 761
 762 static inline void wake_up_rtrs_waiters(struct rnbd_clt_session *sess)
 763 {
 764         sess->rtrs_ready = true;
 765         wake_up_all(&sess->rtrs_waitq);
 766 }
 767
 768 static void close_rtrs(struct rnbd_clt_session *sess)
 769 {
 770         might_sleep();
 771
 772         if (!IS_ERR_OR_NULL(sess->rtrs)) {
 773                 rtrs_clt_close(sess->rtrs);
 774                 sess->rtrs = NULL;
 775                 wake_up_rtrs_waiters(sess);
 776         }
 777 }
 778
 779 static void free_sess(struct rnbd_clt_session *sess)
 780 {
 781         WARN_ON(!list_empty(&sess->devs_list));
 782
 783         might_sleep();
 784
 785         close_rtrs(sess);
 786         destroy_mq_tags(sess);
 787         if (!list_empty(&sess->list)) {
 788                 mutex_lock(&sess_lock);
 789                 list_del(&sess->list);
 790                 mutex_unlock(&sess_lock);
 791         }
 792         free_percpu(sess->cpu_queues);
 793         free_percpu(sess->cpu_rr);
 794         mutex_destroy(&sess->lock);
 795         kfree(sess);
 796 }
 797
 798 static struct rnbd_clt_session *alloc_sess(const char *sessname)
 799 {
 800         struct rnbd_clt_session *sess;
 801         int err, cpu;
 802
 803         sess = kzalloc_node(sizeof(*sess), GFP_KERNEL, NUMA_NO_NODE);
 804         if (!sess)
 805                 return ERR_PTR(-ENOMEM);
 806         strscpy(sess->sessname, sessname, sizeof(sess->sessname));
 807         atomic_set(&sess->busy, 0);
 808         mutex_init(&sess->lock);
 809         INIT_LIST_HEAD(&sess->devs_list);
 810         INIT_LIST_HEAD(&sess->list);
 811         bitmap_zero(sess->cpu_queues_bm, num_possible_cpus());
 812         init_waitqueue_head(&sess->rtrs_waitq);
 813         refcount_set(&sess->refcount, 1);
 814
 815         sess->cpu_queues = alloc_percpu(struct rnbd_cpu_qlist);
 816         if (!sess->cpu_queues) {
 817                 err = -ENOMEM;
 818                 goto err;
 819         }
 820         rnbd_init_cpu_qlists(sess->cpu_queues);
 821
 822         /*
 823          * That is simple percpu variable which stores cpu indices, which are
 824          * incremented on each access.  We need that for the sake of fairness
 825          * to wake up queues in a round-robin manner.
 826          */
 827         sess->cpu_rr = alloc_percpu(int);
 828         if (!sess->cpu_rr) {
 829                 err = -ENOMEM;
 830                 goto err;
 831         }
 832         for_each_possible_cpu(cpu)
 833                 * per_cpu_ptr(sess->cpu_rr, cpu) = cpu;
 834
 835         return sess;
 836
 837 err:
 838         free_sess(sess);
 839
 840         return ERR_PTR(err);
 841 }
 842
 843 static int wait_for_rtrs_connection(struct rnbd_clt_session *sess)
 844 {
 845         wait_event(sess->rtrs_waitq, sess->rtrs_ready);
 846         if (IS_ERR_OR_NULL(sess->rtrs))
 847                 return -ECONNRESET;
 848
 849         return 0;
 850 }
 851
 852 static void wait_for_rtrs_disconnection(struct rnbd_clt_session *sess)
 853         __releases(&sess_lock)
 854         __acquires(&sess_lock)
 855 {
 856         DEFINE_WAIT(wait);
 857
 858         prepare_to_wait(&sess->rtrs_waitq, &wait, TASK_UNINTERRUPTIBLE);
 859         if (IS_ERR_OR_NULL(sess->rtrs)) {
 860                 finish_wait(&sess->rtrs_waitq, &wait);
 861                 return;
 862         }
 863         mutex_unlock(&sess_lock);
 864         /* loop in caller, see __find_and_get_sess().
 865          * You can't leave mutex locked and call schedule(), you will catch a
 866          * deadlock with a caller of free_sess(), which has just put the last
 867          * reference and is about to take the sess_lock in order to delete
 868          * the session from the list.
 869          */
 870         schedule();
 871         mutex_lock(&sess_lock);
 872 }
 873
 874 static struct rnbd_clt_session *__find_and_get_sess(const char *sessname)
 875         __releases(&sess_lock)
 876         __acquires(&sess_lock)
 877 {
 878         struct rnbd_clt_session *sess, *sn;
 879         int err;
 880
 881 again:
 882         list_for_each_entry_safe(sess, sn, &sess_list, list) {
 883                 if (strcmp(sessname, sess->sessname))
 884                         continue;
 885
 886                 if (sess->rtrs_ready && IS_ERR_OR_NULL(sess->rtrs))
 887                         /*
 888                          * No RTRS connection, session is dying.
 889                          */
 890                         continue;
 891
 892                 if (rnbd_clt_get_sess(sess)) {
 893                         /*
 894                          * Alive session is found, wait for RTRS connection.
 895                          */
 896                         mutex_unlock(&sess_lock);
 897                         err = wait_for_rtrs_connection(sess);
 898                         if (err)
 899                                 rnbd_clt_put_sess(sess);
 900                         mutex_lock(&sess_lock);
 901
 902                         if (err)
 903                                 /* Session is dying, repeat the loop */
 904                                 goto again;
 905
 906                         return sess;
 907                 }
 908                 /*
 909                  * Ref is 0, session is dying, wait for RTRS disconnect
 910                  * in order to avoid session names clashes.
 911                  */
 912                 wait_for_rtrs_disconnection(sess);
 913                 /*
 914                  * RTRS is disconnected and soon session will be freed,
 915                  * so repeat a loop.
 916                  */
 917                 goto again;
 918         }
 919
 920         return NULL;
 921 }
 922
 923 /* caller is responsible for initializing 'first' to false */
 924 static struct
 925 rnbd_clt_session *find_or_create_sess(const char *sessname, bool *first)
 926 {
 927         struct rnbd_clt_session *sess = NULL;
 928
 929         mutex_lock(&sess_lock);
 930         sess = __find_and_get_sess(sessname);
 931         if (!sess) {
 932                 sess = alloc_sess(sessname);
 933                 if (IS_ERR(sess)) {
 934                         mutex_unlock(&sess_lock);
 935                         return sess;
 936                 }
 937                 list_add(&sess->list, &sess_list);
 938                 *first = true;
 939         }
 940         mutex_unlock(&sess_lock);
 941
 942         return sess;
 943 }
 944
 945 static int rnbd_client_open(struct block_device *block_device, fmode_t mode)
 946 {
 947         struct rnbd_clt_dev *dev = block_device->bd_disk->private_data;
 948
 949         if (dev->read_only && (mode & FMODE_WRITE))
 950                 return -EPERM;
 951
 952         if (dev->dev_state == DEV_STATE_UNMAPPED ||
 953             !rnbd_clt_get_dev(dev))
 954                 return -EIO;
 955
 956         return 0;
 957 }
 958
 959 static void rnbd_client_release(struct gendisk *gen, fmode_t mode)
 960 {
 961         struct rnbd_clt_dev *dev = gen->private_data;
 962
 963         rnbd_clt_put_dev(dev);
 964 }
 965
 966 static int rnbd_client_getgeo(struct block_device *block_device,
 967                               struct hd_geometry *geo)
 968 {
 969         u64 size;
 970         struct rnbd_clt_dev *dev;
 971
 972         dev = block_device->bd_disk->private_data;
 973         size = dev->size * (dev->logical_block_size / SECTOR_SIZE);
 974         geo->cylinders  = size >> 6;    /* size/64 */
 975         geo->heads      = 4;
 976         geo->sectors    = 16;
 977         geo->start      = 0;
 978
 979         return 0;
 980 }
 981
 982 static const struct block_device_operations rnbd_client_ops = {
 983         .owner          = THIS_MODULE,
 984         .open           = rnbd_client_open,
 985         .release        = rnbd_client_release,
 986         .getgeo         = rnbd_client_getgeo
 987 };
 988
 989 /* The amount of data that belongs to an I/O and the amount of data that
 990  * should be read or written to the disk (bi_size) can differ.
 991  *
 992  * E.g. When WRITE_SAME is used, only a small amount of data is
 993  * transferred that is then written repeatedly over a lot of sectors.
 994  *
 995  * Get the size of data to be transferred via RTRS by summing up the size
 996  * of the scather-gather list entries.
 997  */
 998 static size_t rnbd_clt_get_sg_size(struct scatterlist *sglist, u32 len)
 999 {
1000         struct scatterlist *sg;
1001         size_t tsize = 0;
1002         int i;
1003
1004         for_each_sg(sglist, sg, len, i)
1005                 tsize += sg->length;
1006         return tsize;
1007 }
1008
1009 static int rnbd_client_xfer_request(struct rnbd_clt_dev *dev,
1010                                      struct request *rq,
1011                                      struct rnbd_iu *iu)
1012 {
1013         struct rtrs_clt *rtrs = dev->sess->rtrs;
1014         struct rtrs_permit *permit = iu->permit;
1015         struct rnbd_msg_io msg;
1016         struct rtrs_clt_req_ops req_ops;
1017         unsigned int sg_cnt = 0;
1018         struct kvec vec;
1019         size_t size;
1020         int err;
1021
1022         iu->rq          = rq;
1023         iu->dev         = dev;
1024         msg.sector      = cpu_to_le64(blk_rq_pos(rq));
1025         msg.bi_size     = cpu_to_le32(blk_rq_bytes(rq));
1026         msg.rw          = cpu_to_le32(rq_to_rnbd_flags(rq));
1027         msg.prio        = cpu_to_le16(req_get_ioprio(rq));
1028
1029         /*
1030          * We only support discards with single segment for now.
1031          * See queue limits.
1032          */
1033         if (req_op(rq) != REQ_OP_DISCARD)
1034                 sg_cnt = blk_rq_map_sg(dev->queue, rq, iu->sgt.sgl);
1035
1036         if (sg_cnt == 0)
1037                 sg_mark_end(&iu->sgt.sgl[0]);
1038
1039         msg.hdr.type    = cpu_to_le16(RNBD_MSG_IO);
1040         msg.device_id   = cpu_to_le32(dev->device_id);
1041
1042         vec = (struct kvec) {
1043                 .iov_base = &msg,
1044                 .iov_len  = sizeof(msg)
1045         };
1046         size = rnbd_clt_get_sg_size(iu->sgt.sgl, sg_cnt);
1047         req_ops = (struct rtrs_clt_req_ops) {
1048                 .priv = iu,
1049                 .conf_fn = msg_io_conf,
1050         };
1051         err = rtrs_clt_request(rq_data_dir(rq), &req_ops, rtrs, permit,
1052                                &vec, 1, size, iu->sgt.sgl, sg_cnt);
1053         if (err) {
1054                 rnbd_clt_err_rl(dev, "RTRS failed to transfer IO, err: %d\n",
1055                                  err);
1056                 return err;
1057         }
1058
1059         return 0;
1060 }
1061
1062 /**
1063  * rnbd_clt_dev_add_to_requeue() - add device to requeue if session is busy
1064  * @dev:        Device to be checked
1065  * @q:          Queue to be added to the requeue list if required
1066  *
1067  * Description:
1068  *     If session is busy, that means someone will requeue us when resources
1069  *     are freed.  If session is not doing anything - device is not added to
1070  *     the list and @false is returned.
1071  */
1072 static bool rnbd_clt_dev_add_to_requeue(struct rnbd_clt_dev *dev,
1073                                                 struct rnbd_queue *q)
1074 {
1075         struct rnbd_clt_session *sess = dev->sess;
1076         struct rnbd_cpu_qlist *cpu_q;
1077         unsigned long flags;
1078         bool added = true;
1079         bool need_set;
1080
1081         cpu_q = get_cpu_ptr(sess->cpu_queues);
1082         spin_lock_irqsave(&cpu_q->requeue_lock, flags);
1083
1084         if (!test_and_set_bit_lock(0, &q->in_list)) {
1085                 if (WARN_ON(!list_empty(&q->requeue_list)))
1086                         goto unlock;
1087
1088                 need_set = !test_bit(cpu_q->cpu, sess->cpu_queues_bm);
1089                 if (need_set) {
1090                         set_bit(cpu_q->cpu, sess->cpu_queues_bm);
1091                         /* Paired with rnbd_put_permit(). Set a bit first
1092                          * and then observe the busy counter.
1093                          */
1094                         smp_mb__before_atomic();
1095                 }
1096                 if (atomic_read(&sess->busy)) {
1097                         list_add_tail(&q->requeue_list, &cpu_q->requeue_list);
1098                 } else {
1099                         /* Very unlikely, but possible: busy counter was
1100                          * observed as zero.  Drop all bits and return
1101                          * false to restart the queue by ourselves.
1102                          */
1103                         if (need_set)
1104                                 clear_bit(cpu_q->cpu, sess->cpu_queues_bm);
1105                         clear_bit_unlock(0, &q->in_list);
1106                         added = false;
1107                 }
1108         }
1109 unlock:
1110         spin_unlock_irqrestore(&cpu_q->requeue_lock, flags);
1111         put_cpu_ptr(sess->cpu_queues);
1112
1113         return added;
1114 }
1115
1116 static void rnbd_clt_dev_kick_mq_queue(struct rnbd_clt_dev *dev,
1117                                         struct blk_mq_hw_ctx *hctx,
1118                                         int delay)
1119 {
1120         struct rnbd_queue *q = hctx->driver_data;
1121
1122         if (delay != RNBD_DELAY_IFBUSY)
1123                 blk_mq_delay_run_hw_queue(hctx, delay);
1124         else if (!rnbd_clt_dev_add_to_requeue(dev, q))
1125                 /*
1126                  * If session is not busy we have to restart
1127                  * the queue ourselves.
1128                  */
1129                 blk_mq_delay_run_hw_queue(hctx, 10/*ms*/);
1130 }
1131
1132 static blk_status_t rnbd_queue_rq(struct blk_mq_hw_ctx *hctx,
1133                                    const struct blk_mq_queue_data *bd)
1134 {
1135         struct request *rq = bd->rq;
1136         struct rnbd_clt_dev *dev = rq->rq_disk->private_data;
1137         struct rnbd_iu *iu = blk_mq_rq_to_pdu(rq);
1138         int err;
1139         blk_status_t ret = BLK_STS_IOERR;
1140
1141         if (dev->dev_state != DEV_STATE_MAPPED)
1142                 return BLK_STS_IOERR;
1143
1144         iu->permit = rnbd_get_permit(dev->sess, RTRS_IO_CON,
1145                                       RTRS_PERMIT_NOWAIT);
1146         if (!iu->permit) {
1147                 rnbd_clt_dev_kick_mq_queue(dev, hctx, RNBD_DELAY_IFBUSY);
1148                 return BLK_STS_RESOURCE;
1149         }
1150
1151         iu->sgt.sgl = iu->first_sgl;
1152         err = sg_alloc_table_chained(&iu->sgt,
1153                                      /* Even-if the request has no segment,
1154                                       * sglist must have one entry at least.
1155                                       */
1156                                      blk_rq_nr_phys_segments(rq) ? : 1,
1157                                      iu->sgt.sgl,
1158                                      RNBD_INLINE_SG_CNT);
1159         if (err) {
1160                 rnbd_clt_err_rl(dev, "sg_alloc_table_chained ret=%d\n", err);
1161                 rnbd_clt_dev_kick_mq_queue(dev, hctx, 10/*ms*/);
1162                 rnbd_put_permit(dev->sess, iu->permit);
1163                 return BLK_STS_RESOURCE;
1164         }
1165
1166         blk_mq_start_request(rq);
1167         err = rnbd_client_xfer_request(dev, rq, iu);
1168         if (err == 0)
1169                 return BLK_STS_OK;
1170         if (err == -EAGAIN || err == -ENOMEM) {
1171                 rnbd_clt_dev_kick_mq_queue(dev, hctx, 10/*ms*/);
1172                 ret = BLK_STS_RESOURCE;
1173         }
1174         sg_free_table_chained(&iu->sgt, RNBD_INLINE_SG_CNT);
1175         rnbd_put_permit(dev->sess, iu->permit);
1176         return ret;
1177 }
1178
1179 static int rnbd_rdma_poll(struct blk_mq_hw_ctx *hctx)
1180 {
1181         struct rnbd_queue *q = hctx->driver_data;
1182         struct rnbd_clt_dev *dev = q->dev;
1183         int cnt;
1184
1185         cnt = rtrs_clt_rdma_cq_direct(dev->sess->rtrs, hctx->queue_num);
1186         return cnt;
1187 }
1188
1189 static int rnbd_rdma_map_queues(struct blk_mq_tag_set *set)
1190 {
1191         struct rnbd_clt_session *sess = set->driver_data;
1192
1193         /* shared read/write queues */
1194         set->map[HCTX_TYPE_DEFAULT].nr_queues = num_online_cpus();
1195         set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
1196         set->map[HCTX_TYPE_READ].nr_queues = num_online_cpus();
1197         set->map[HCTX_TYPE_READ].queue_offset = 0;
1198         blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
1199         blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
1200
1201         if (sess->nr_poll_queues) {
1202                 /* dedicated queue for poll */
1203                 set->map[HCTX_TYPE_POLL].nr_queues = sess->nr_poll_queues;
1204                 set->map[HCTX_TYPE_POLL].queue_offset = set->map[HCTX_TYPE_READ].queue_offset +
1205                         set->map[HCTX_TYPE_READ].nr_queues;
1206                 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
1207                 pr_info("[session=%s] mapped %d/%d/%d default/read/poll queues.\n",
1208                         sess->sessname,
1209                         set->map[HCTX_TYPE_DEFAULT].nr_queues,
1210                         set->map[HCTX_TYPE_READ].nr_queues,
1211                         set->map[HCTX_TYPE_POLL].nr_queues);
1212         } else {
1213                 pr_info("[session=%s] mapped %d/%d default/read queues.\n",
1214                         sess->sessname,
1215                         set->map[HCTX_TYPE_DEFAULT].nr_queues,
1216                         set->map[HCTX_TYPE_READ].nr_queues);
1217         }
1218
1219         return 0;
1220 }
1221
1222 static struct blk_mq_ops rnbd_mq_ops = {
1223         .queue_rq       = rnbd_queue_rq,
1224         .complete       = rnbd_softirq_done_fn,
1225         .map_queues     = rnbd_rdma_map_queues,
1226         .poll           = rnbd_rdma_poll,
1227 };
1228
1229 static int setup_mq_tags(struct rnbd_clt_session *sess)
1230 {
1231         struct blk_mq_tag_set *tag_set = &sess->tag_set;
1232
1233         memset(tag_set, 0, sizeof(*tag_set));
1234         tag_set->ops            = &rnbd_mq_ops;
1235         tag_set->queue_depth    = sess->queue_depth;
1236         tag_set->numa_node              = NUMA_NO_NODE;
1237         tag_set->flags          = BLK_MQ_F_SHOULD_MERGE |
1238                                   BLK_MQ_F_TAG_QUEUE_SHARED;
1239         tag_set->cmd_size       = sizeof(struct rnbd_iu) + RNBD_RDMA_SGL_SIZE;
1240
1241         /* for HCTX_TYPE_DEFAULT, HCTX_TYPE_READ, HCTX_TYPE_POLL */
1242         tag_set->nr_maps        = sess->nr_poll_queues ? HCTX_MAX_TYPES : 2;
1243         /*
1244          * HCTX_TYPE_DEFAULT and HCTX_TYPE_READ share one set of queues
1245          * others are for HCTX_TYPE_POLL
1246          */
1247         tag_set->nr_hw_queues   = num_online_cpus() + sess->nr_poll_queues;
1248         tag_set->driver_data    = sess;
1249
1250         return blk_mq_alloc_tag_set(tag_set);
1251 }
1252
1253 static struct rnbd_clt_session *
1254 find_and_get_or_create_sess(const char *sessname,
1255                             const struct rtrs_addr *paths,
1256                             size_t path_cnt, u16 port_nr, u32 nr_poll_queues)
1257 {
1258         struct rnbd_clt_session *sess;
1259         struct rtrs_attrs attrs;
1260         int err;
1261         bool first = false;
1262         struct rtrs_clt_ops rtrs_ops;
1263
1264         sess = find_or_create_sess(sessname, &first);
1265         if (sess == ERR_PTR(-ENOMEM))
1266                 return ERR_PTR(-ENOMEM);
1267         else if ((nr_poll_queues && !first) ||  (!nr_poll_queues && sess->nr_poll_queues)) {
1268                 /*
1269                  * A device MUST have its own session to use the polling-mode.
1270                  * It must fail to map new device with the same session.
1271                  */
1272                 err = -EINVAL;
1273                 goto put_sess;
1274         }
1275
1276         if (!first)
1277                 return sess;
1278
1279         if (!path_cnt) {
1280                 pr_err("Session %s not found, and path parameter not given", sessname);
1281                 err = -ENXIO;
1282                 goto put_sess;
1283         }
1284
1285         rtrs_ops = (struct rtrs_clt_ops) {
1286                 .priv = sess,
1287                 .link_ev = rnbd_clt_link_ev,
1288         };
1289         /*
1290          * Nothing was found, establish rtrs connection and proceed further.
1291          */
1292         sess->rtrs = rtrs_clt_open(&rtrs_ops, sessname,
1293                                    paths, path_cnt, port_nr,
1294                                    0, /* Do not use pdu of rtrs */
1295                                    RECONNECT_DELAY,
1296                                    MAX_RECONNECTS, nr_poll_queues);
1297         if (IS_ERR(sess->rtrs)) {
1298                 err = PTR_ERR(sess->rtrs);
1299                 goto wake_up_and_put;
1300         }
1301
1302         err = rtrs_clt_query(sess->rtrs, &attrs);
1303         if (err)
1304                 goto close_rtrs;
1305
1306         sess->max_io_size = attrs.max_io_size;
1307         sess->queue_depth = attrs.queue_depth;
1308         sess->nr_poll_queues = nr_poll_queues;
1309         sess->max_segments = attrs.max_segments;
1310
1311         err = setup_mq_tags(sess);
1312         if (err)
1313                 goto close_rtrs;
1314
1315         err = send_msg_sess_info(sess, RTRS_PERMIT_WAIT);
1316         if (err)
1317                 goto close_rtrs;
1318
1319         wake_up_rtrs_waiters(sess);
1320
1321         return sess;
1322
1323 close_rtrs:
1324         close_rtrs(sess);
1325 put_sess:
1326         rnbd_clt_put_sess(sess);
1327
1328         return ERR_PTR(err);
1329
1330 wake_up_and_put:
1331         wake_up_rtrs_waiters(sess);
1332         goto put_sess;
1333 }
1334
1335 static inline void rnbd_init_hw_queue(struct rnbd_clt_dev *dev,
1336                                        struct rnbd_queue *q,
1337                                        struct blk_mq_hw_ctx *hctx)
1338 {
1339         INIT_LIST_HEAD(&q->requeue_list);
1340         q->dev  = dev;
1341         q->hctx = hctx;
1342 }
1343
1344 static void rnbd_init_mq_hw_queues(struct rnbd_clt_dev *dev)
1345 {
1346         int i;
1347         struct blk_mq_hw_ctx *hctx;
1348         struct rnbd_queue *q;
1349
1350         queue_for_each_hw_ctx(dev->queue, hctx, i) {
1351                 q = &dev->hw_queues[i];
1352                 rnbd_init_hw_queue(dev, q, hctx);
1353                 hctx->driver_data = q;
1354         }
1355 }
1356
1357 static void setup_request_queue(struct rnbd_clt_dev *dev)
1358 {
1359         blk_queue_logical_block_size(dev->queue, dev->logical_block_size);
1360         blk_queue_physical_block_size(dev->queue, dev->physical_block_size);
1361         blk_queue_max_hw_sectors(dev->queue, dev->max_hw_sectors);
1362         blk_queue_max_write_same_sectors(dev->queue,
1363                                          dev->max_write_same_sectors);
1364
1365         /*
1366          * we don't support discards to "discontiguous" segments
1367          * in on request
1368          */
1369         blk_queue_max_discard_segments(dev->queue, 1);
1370
1371         blk_queue_max_discard_sectors(dev->queue, dev->max_discard_sectors);
1372         dev->queue->limits.discard_granularity  = dev->discard_granularity;
1373         dev->queue->limits.discard_alignment    = dev->discard_alignment;
1374         if (dev->max_discard_sectors)
1375                 blk_queue_flag_set(QUEUE_FLAG_DISCARD, dev->queue);
1376         if (dev->secure_discard)
1377                 blk_queue_flag_set(QUEUE_FLAG_SECERASE, dev->queue);
1378
1379         blk_queue_flag_set(QUEUE_FLAG_SAME_COMP, dev->queue);
1380         blk_queue_flag_set(QUEUE_FLAG_SAME_FORCE, dev->queue);
1381         blk_queue_max_segments(dev->queue, dev->max_segments);
1382         blk_queue_io_opt(dev->queue, dev->sess->max_io_size);
1383         blk_queue_virt_boundary(dev->queue, SZ_4K - 1);
1384         blk_queue_write_cache(dev->queue, dev->wc, dev->fua);
1385 }
1386
1387 static void rnbd_clt_setup_gen_disk(struct rnbd_clt_dev *dev, int idx)
1388 {
1389         dev->gd->major          = rnbd_client_major;
1390         dev->gd->first_minor    = idx << RNBD_PART_BITS;
1391         dev->gd->minors         = 1 << RNBD_PART_BITS;
1392         dev->gd->fops           = &rnbd_client_ops;
1393         dev->gd->queue          = dev->queue;
1394         dev->gd->private_data   = dev;
1395         snprintf(dev->gd->disk_name, sizeof(dev->gd->disk_name), "rnbd%d",
1396                  idx);
1397         pr_debug("disk_name=%s, capacity=%zu\n",
1398                  dev->gd->disk_name,
1399                  dev->nsectors * (dev->logical_block_size / SECTOR_SIZE)
1400                  );
1401
1402         set_capacity(dev->gd, dev->nsectors);
1403
1404         if (dev->access_mode == RNBD_ACCESS_RO) {
1405                 dev->read_only = true;
1406                 set_disk_ro(dev->gd, true);
1407         } else {
1408                 dev->read_only = false;
1409         }
1410
1411         if (!dev->rotational)
1412                 blk_queue_flag_set(QUEUE_FLAG_NONROT, dev->queue);
1413         add_disk(dev->gd);
1414 }
1415
1416 static int rnbd_client_setup_device(struct rnbd_clt_dev *dev)
1417 {
1418         int idx = dev->clt_device_id;
1419
1420         dev->size = dev->nsectors * dev->logical_block_size;
1421
1422         dev->gd = blk_mq_alloc_disk(&dev->sess->tag_set, dev);
1423         if (IS_ERR(dev->gd))
1424                 return PTR_ERR(dev->gd);
1425         dev->queue = dev->gd->queue;
1426         rnbd_init_mq_hw_queues(dev);
1427
1428         setup_request_queue(dev);
1429         rnbd_clt_setup_gen_disk(dev, idx);
1430         return 0;
1431 }
1432
1433 static struct rnbd_clt_dev *init_dev(struct rnbd_clt_session *sess,
1434                                       enum rnbd_access_mode access_mode,
1435                                       const char *pathname,
1436                                       u32 nr_poll_queues)
1437 {
1438         struct rnbd_clt_dev *dev;
1439         int ret;
1440
1441         dev = kzalloc_node(sizeof(*dev), GFP_KERNEL, NUMA_NO_NODE);
1442         if (!dev)
1443                 return ERR_PTR(-ENOMEM);
1444
1445         /*
1446          * nr_cpu_ids: the number of softirq queues
1447          * nr_poll_queues: the number of polling queues
1448          */
1449         dev->hw_queues = kcalloc(nr_cpu_ids + nr_poll_queues,
1450                                  sizeof(*dev->hw_queues),
1451                                  GFP_KERNEL);
1452         if (!dev->hw_queues) {
1453                 ret = -ENOMEM;
1454                 goto out_alloc;
1455         }
1456
1457         mutex_lock(&ida_lock);
1458         ret = ida_simple_get(&index_ida, 0, 1 << (MINORBITS - RNBD_PART_BITS),
1459                              GFP_KERNEL);
1460         mutex_unlock(&ida_lock);
1461         if (ret < 0) {
1462                 pr_err("Failed to initialize device '%s' from session %s, allocating idr failed, err: %d\n",
1463                        pathname, sess->sessname, ret);
1464                 goto out_queues;
1465         }
1466
1467         dev->pathname = kstrdup(pathname, GFP_KERNEL);
1468         if (!dev->pathname) {
1469                 ret = -ENOMEM;
1470                 goto out_queues;
1471         }
1472
1473         dev->clt_device_id      = ret;
1474         dev->sess               = sess;
1475         dev->access_mode        = access_mode;
1476         dev->nr_poll_queues     = nr_poll_queues;
1477         mutex_init(&dev->lock);
1478         refcount_set(&dev->refcount, 1);
1479         dev->dev_state = DEV_STATE_INIT;
1480
1481         /*
1482          * Here we called from sysfs entry, thus clt-sysfs is
1483          * responsible that session will not disappear.
1484          */
1485         WARN_ON(!rnbd_clt_get_sess(sess));
1486
1487         return dev;
1488
1489 out_queues:
1490         kfree(dev->hw_queues);
1491 out_alloc:
1492         kfree(dev);
1493         return ERR_PTR(ret);
1494 }
1495
1496 static bool __exists_dev(const char *pathname, const char *sessname)
1497 {
1498         struct rnbd_clt_session *sess;
1499         struct rnbd_clt_dev *dev;
1500         bool found = false;
1501
1502         list_for_each_entry(sess, &sess_list, list) {
1503                 if (sessname && strncmp(sess->sessname, sessname,
1504                                         sizeof(sess->sessname)))
1505                         continue;
1506                 mutex_lock(&sess->lock);
1507                 list_for_each_entry(dev, &sess->devs_list, list) {
1508                         if (strlen(dev->pathname) == strlen(pathname) &&
1509                             !strcmp(dev->pathname, pathname)) {
1510                                 found = true;
1511                                 break;
1512                         }
1513                 }
1514                 mutex_unlock(&sess->lock);
1515                 if (found)
1516                         break;
1517         }
1518
1519         return found;
1520 }
1521
1522 static bool exists_devpath(const char *pathname, const char *sessname)
1523 {
1524         bool found;
1525
1526         mutex_lock(&sess_lock);
1527         found = __exists_dev(pathname, sessname);
1528         mutex_unlock(&sess_lock);
1529
1530         return found;
1531 }
1532
1533 static bool insert_dev_if_not_exists_devpath(struct rnbd_clt_dev *dev)
1534 {
1535         bool found;
1536         struct rnbd_clt_session *sess = dev->sess;
1537
1538         mutex_lock(&sess_lock);
1539         found = __exists_dev(dev->pathname, sess->sessname);
1540         if (!found) {
1541                 mutex_lock(&sess->lock);
1542                 list_add_tail(&dev->list, &sess->devs_list);
1543                 mutex_unlock(&sess->lock);
1544         }
1545         mutex_unlock(&sess_lock);
1546
1547         return found;
1548 }
1549
1550 static void delete_dev(struct rnbd_clt_dev *dev)
1551 {
1552         struct rnbd_clt_session *sess = dev->sess;
1553
1554         mutex_lock(&sess->lock);
1555         list_del(&dev->list);
1556         mutex_unlock(&sess->lock);
1557 }
1558
1559 struct rnbd_clt_dev *rnbd_clt_map_device(const char *sessname,
1560                                            struct rtrs_addr *paths,
1561                                            size_t path_cnt, u16 port_nr,
1562                                            const char *pathname,
1563                                            enum rnbd_access_mode access_mode,
1564                                            u32 nr_poll_queues)
1565 {
1566         struct rnbd_clt_session *sess;
1567         struct rnbd_clt_dev *dev;
1568         int ret;
1569
1570         if (exists_devpath(pathname, sessname))
1571                 return ERR_PTR(-EEXIST);
1572
1573         sess = find_and_get_or_create_sess(sessname, paths, path_cnt, port_nr, nr_poll_queues);
1574         if (IS_ERR(sess))
1575                 return ERR_CAST(sess);
1576
1577         dev = init_dev(sess, access_mode, pathname, nr_poll_queues);
1578         if (IS_ERR(dev)) {
1579                 pr_err("map_device: failed to map device '%s' from session %s, can't initialize device, err: %ld\n",
1580                        pathname, sess->sessname, PTR_ERR(dev));
1581                 ret = PTR_ERR(dev);
1582                 goto put_sess;
1583         }
1584         if (insert_dev_if_not_exists_devpath(dev)) {
1585                 ret = -EEXIST;
1586                 goto put_dev;
1587         }
1588         ret = send_msg_open(dev, RTRS_PERMIT_WAIT);
1589         if (ret) {
1590                 rnbd_clt_err(dev,
1591                               "map_device: failed, can't open remote device, err: %d\n",
1592                               ret);
1593                 goto del_dev;
1594         }
1595         mutex_lock(&dev->lock);
1596         pr_debug("Opened remote device: session=%s, path='%s'\n",
1597                  sess->sessname, pathname);
1598         ret = rnbd_client_setup_device(dev);
1599         if (ret) {
1600                 rnbd_clt_err(dev,
1601                               "map_device: Failed to configure device, err: %d\n",
1602                               ret);
1603                 mutex_unlock(&dev->lock);
1604                 goto send_close;
1605         }
1606
1607         rnbd_clt_info(dev,
1608                        "map_device: Device mapped as %s (nsectors: %zu, logical_block_size: %d, physical_block_size: %d, max_write_same_sectors: %d, max_discard_sectors: %d, discard_granularity: %d, discard_alignment: %d, secure_discard: %d, max_segments: %d, max_hw_sectors: %d, rotational: %d, wc: %d, fua: %d)\n",
1609                        dev->gd->disk_name, dev->nsectors,
1610                        dev->logical_block_size, dev->physical_block_size,
1611                        dev->max_write_same_sectors, dev->max_discard_sectors,
1612                        dev->discard_granularity, dev->discard_alignment,
1613                        dev->secure_discard, dev->max_segments,
1614                        dev->max_hw_sectors, dev->rotational, dev->wc, dev->fua);
1615
1616         mutex_unlock(&dev->lock);
1617         rnbd_clt_put_sess(sess);
1618
1619         return dev;
1620
1621 send_close:
1622         send_msg_close(dev, dev->device_id, RTRS_PERMIT_WAIT);
1623 del_dev:
1624         delete_dev(dev);
1625 put_dev:
1626         rnbd_clt_put_dev(dev);
1627 put_sess:
1628         rnbd_clt_put_sess(sess);
1629
1630         return ERR_PTR(ret);
1631 }
1632
1633 static void destroy_gen_disk(struct rnbd_clt_dev *dev)
1634 {
1635         del_gendisk(dev->gd);
1636         blk_cleanup_disk(dev->gd);
1637 }
1638
1639 static void destroy_sysfs(struct rnbd_clt_dev *dev,
1640                           const struct attribute *sysfs_self)
1641 {
1642         rnbd_clt_remove_dev_symlink(dev);
1643         if (dev->kobj.state_initialized) {
1644                 if (sysfs_self)
1645                         /* To avoid deadlock firstly remove itself */
1646                         sysfs_remove_file_self(&dev->kobj, sysfs_self);
1647                 kobject_del(&dev->kobj);
1648                 kobject_put(&dev->kobj);
1649         }
1650 }
1651
1652 int rnbd_clt_unmap_device(struct rnbd_clt_dev *dev, bool force,
1653                            const struct attribute *sysfs_self)
1654 {
1655         struct rnbd_clt_session *sess = dev->sess;
1656         int refcount, ret = 0;
1657         bool was_mapped;
1658
1659         mutex_lock(&dev->lock);
1660         if (dev->dev_state == DEV_STATE_UNMAPPED) {
1661                 rnbd_clt_info(dev, "Device is already being unmapped\n");
1662                 ret = -EALREADY;
1663                 goto err;
1664         }
1665         refcount = refcount_read(&dev->refcount);
1666         if (!force && refcount > 1) {
1667                 rnbd_clt_err(dev,
1668                               "Closing device failed, device is in use, (%d device users)\n",
1669                               refcount - 1);
1670                 ret = -EBUSY;
1671                 goto err;
1672         }
1673         was_mapped = (dev->dev_state == DEV_STATE_MAPPED);
1674         dev->dev_state = DEV_STATE_UNMAPPED;
1675         mutex_unlock(&dev->lock);
1676
1677         delete_dev(dev);
1678         destroy_sysfs(dev, sysfs_self);
1679         destroy_gen_disk(dev);
1680         if (was_mapped && sess->rtrs)
1681                 send_msg_close(dev, dev->device_id, RTRS_PERMIT_WAIT);
1682
1683         rnbd_clt_info(dev, "Device is unmapped\n");
1684
1685         /* Likely last reference put */
1686         rnbd_clt_put_dev(dev);
1687
1688         /*
1689          * Here device and session can be vanished!
1690          */
1691
1692         return 0;
1693 err:
1694         mutex_unlock(&dev->lock);
1695
1696         return ret;
1697 }
1698
1699 int rnbd_clt_remap_device(struct rnbd_clt_dev *dev)
1700 {
1701         int err;
1702
1703         mutex_lock(&dev->lock);
1704         if (dev->dev_state == DEV_STATE_MAPPED_DISCONNECTED)
1705                 err = 0;
1706         else if (dev->dev_state == DEV_STATE_UNMAPPED)
1707                 err = -ENODEV;
1708         else if (dev->dev_state == DEV_STATE_MAPPED)
1709                 err = -EALREADY;
1710         else
1711                 err = -EBUSY;
1712         mutex_unlock(&dev->lock);
1713         if (!err) {
1714                 rnbd_clt_info(dev, "Remapping device.\n");
1715                 err = send_msg_open(dev, RTRS_PERMIT_WAIT);
1716                 if (err)
1717                         rnbd_clt_err(dev, "remap_device: %d\n", err);
1718         }
1719
1720         return err;
1721 }
1722
1723 static void unmap_device_work(struct work_struct *work)
1724 {
1725         struct rnbd_clt_dev *dev;
1726
1727         dev = container_of(work, typeof(*dev), unmap_on_rmmod_work);
1728         rnbd_clt_unmap_device(dev, true, NULL);
1729 }
1730
1731 static void rnbd_destroy_sessions(void)
1732 {
1733         struct rnbd_clt_session *sess, *sn;
1734         struct rnbd_clt_dev *dev, *tn;
1735
1736         /* Firstly forbid access through sysfs interface */
1737         rnbd_clt_destroy_sysfs_files();
1738
1739         /*
1740          * Here at this point there is no any concurrent access to sessions
1741          * list and devices list:
1742          *   1. New session or device can't be created - session sysfs files
1743          *      are removed.
1744          *   2. Device or session can't be removed - module reference is taken
1745          *      into account in unmap device sysfs callback.
1746          *   3. No IO requests inflight - each file open of block_dev increases
1747          *      module reference in get_disk().
1748          *
1749          * But still there can be user requests inflights, which are sent by
1750          * asynchronous send_msg_*() functions, thus before unmapping devices
1751          * RTRS session must be explicitly closed.
1752          */
1753
1754         list_for_each_entry_safe(sess, sn, &sess_list, list) {
1755                 if (!rnbd_clt_get_sess(sess))
1756                         continue;
1757                 close_rtrs(sess);
1758                 list_for_each_entry_safe(dev, tn, &sess->devs_list, list) {
1759                         /*
1760                          * Here unmap happens in parallel for only one reason:
1761                          * blk_cleanup_queue() takes around half a second, so
1762                          * on huge amount of devices the whole module unload
1763                          * procedure takes minutes.
1764                          */
1765                         INIT_WORK(&dev->unmap_on_rmmod_work, unmap_device_work);
1766                         queue_work(system_long_wq, &dev->unmap_on_rmmod_work);
1767                 }
1768                 rnbd_clt_put_sess(sess);
1769         }
1770         /* Wait for all scheduled unmap works */
1771         flush_workqueue(system_long_wq);
1772         WARN_ON(!list_empty(&sess_list));
1773 }
1774
1775 static int __init rnbd_client_init(void)
1776 {
1777         int err = 0;
1778
1779         BUILD_BUG_ON(sizeof(struct rnbd_msg_hdr) != 4);
1780         BUILD_BUG_ON(sizeof(struct rnbd_msg_sess_info) != 36);
1781         BUILD_BUG_ON(sizeof(struct rnbd_msg_sess_info_rsp) != 36);
1782         BUILD_BUG_ON(sizeof(struct rnbd_msg_open) != 264);
1783         BUILD_BUG_ON(sizeof(struct rnbd_msg_close) != 8);
1784         BUILD_BUG_ON(sizeof(struct rnbd_msg_open_rsp) != 56);
1785         rnbd_client_major = register_blkdev(rnbd_client_major, "rnbd");
1786         if (rnbd_client_major <= 0) {
1787                 pr_err("Failed to load module, block device registration failed\n");
1788                 return -EBUSY;
1789         }
1790
1791         err = rnbd_clt_create_sysfs_files();
1792         if (err) {
1793                 pr_err("Failed to load module, creating sysfs device files failed, err: %d\n",
1794                        err);
1795                 unregister_blkdev(rnbd_client_major, "rnbd");
1796         }
1797
1798         return err;
1799 }
1800
1801 static void __exit rnbd_client_exit(void)
1802 {
1803         rnbd_destroy_sessions();
1804         unregister_blkdev(rnbd_client_major, "rnbd");
1805         ida_destroy(&index_ida);
1806 }
1807
1808 module_init(rnbd_client_init);
1809 module_exit(rnbd_client_exit);