drivers/nvme/host/rdma.c

   1 /*
   2  * NVMe over Fabrics RDMA host code.
   3  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  */
  14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  15 #include <linux/module.h>
  16 #include <linux/init.h>
  17 #include <linux/slab.h>
  18 #include <linux/err.h>
  19 #include <linux/string.h>
  20 #include <linux/atomic.h>
  21 #include <linux/blk-mq.h>
  22 #include <linux/types.h>
  23 #include <linux/list.h>
  24 #include <linux/mutex.h>
  25 #include <linux/scatterlist.h>
  26 #include <linux/nvme.h>
  27 #include <asm/unaligned.h>
  28
  29 #include <rdma/ib_verbs.h>
  30 #include <rdma/rdma_cm.h>
  31 #include <rdma/ib_cm.h>
  32 #include <linux/nvme-rdma.h>
  33
  34 #include "nvme.h"
  35 #include "fabrics.h"
  36
  37
  38 #define NVME_RDMA_CONNECT_TIMEOUT_MS    1000            /* 1 second */
  39
  40 #define NVME_RDMA_MAX_SEGMENT_SIZE      0xffffff        /* 24-bit SGL field */
  41
  42 #define NVME_RDMA_MAX_SEGMENTS          256
  43
  44 #define NVME_RDMA_MAX_INLINE_SEGMENTS   1
  45
  46 /*
  47  * We handle AEN commands ourselves and don't even let the
  48  * block layer know about them.
  49  */
  50 #define NVME_RDMA_NR_AEN_COMMANDS      1
  51 #define NVME_RDMA_AQ_BLKMQ_DEPTH       \
  52         (NVMF_AQ_DEPTH - NVME_RDMA_NR_AEN_COMMANDS)
  53
  54 struct nvme_rdma_device {
  55         struct ib_device       *dev;
  56         struct ib_pd           *pd;
  57         struct kref             ref;
  58         struct list_head        entry;
  59 };
  60
  61 struct nvme_rdma_qe {
  62         struct ib_cqe           cqe;
  63         void                    *data;
  64         u64                     dma;
  65 };
  66
  67 struct nvme_rdma_queue;
  68 struct nvme_rdma_request {
  69         struct ib_mr            *mr;
  70         struct nvme_rdma_qe     sqe;
  71         struct ib_sge           sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
  72         u32                     num_sge;
  73         int                     nents;
  74         bool                    inline_data;
  75         struct ib_reg_wr        reg_wr;
  76         struct ib_cqe           reg_cqe;
  77         struct nvme_rdma_queue  *queue;
  78         struct sg_table         sg_table;
  79         struct scatterlist      first_sgl[];
  80 };
  81
  82 enum nvme_rdma_queue_flags {
  83         NVME_RDMA_Q_CONNECTED = (1 << 0),
  84         NVME_RDMA_IB_QUEUE_ALLOCATED = (1 << 1),
  85         NVME_RDMA_Q_DELETING = (1 << 2),
  86         NVME_RDMA_Q_LIVE = (1 << 3),
  87 };
  88
  89 struct nvme_rdma_queue {
  90         struct nvme_rdma_qe     *rsp_ring;
  91         atomic_t                sig_count;
  92         int                     queue_size;
  93         size_t                  cmnd_capsule_len;
  94         struct nvme_rdma_ctrl   *ctrl;
  95         struct nvme_rdma_device *device;
  96         struct ib_cq            *ib_cq;
  97         struct ib_qp            *qp;
  98
  99         unsigned long           flags;
 100         struct rdma_cm_id       *cm_id;
 101         int                     cm_error;
 102         struct completion       cm_done;
 103 };
 104
 105 struct nvme_rdma_ctrl {
 106         /* read and written in the hot path */
 107         spinlock_t              lock;
 108
 109         /* read only in the hot path */
 110         struct nvme_rdma_queue  *queues;
 111         u32                     queue_count;
 112
 113         /* other member variables */
 114         struct blk_mq_tag_set   tag_set;
 115         struct work_struct      delete_work;
 116         struct work_struct      reset_work;
 117         struct work_struct      err_work;
 118
 119         struct nvme_rdma_qe     async_event_sqe;
 120
 121         int                     reconnect_delay;
 122         struct delayed_work     reconnect_work;
 123
 124         struct list_head        list;
 125
 126         struct blk_mq_tag_set   admin_tag_set;
 127         struct nvme_rdma_device *device;
 128
 129         u64                     cap;
 130         u32                     max_fr_pages;
 131
 132         union {
 133                 struct sockaddr addr;
 134                 struct sockaddr_in addr_in;
 135         };
 136
 137         struct nvme_ctrl        ctrl;
 138 };
 139
 140 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
 141 {
 142         return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
 143 }
 144
 145 static LIST_HEAD(device_list);
 146 static DEFINE_MUTEX(device_list_mutex);
 147
 148 static LIST_HEAD(nvme_rdma_ctrl_list);
 149 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
 150
 151 static struct workqueue_struct *nvme_rdma_wq;
 152
 153 /*
 154  * Disabling this option makes small I/O goes faster, but is fundamentally
 155  * unsafe.  With it turned off we will have to register a global rkey that
 156  * allows read and write access to all physical memory.
 157  */
 158 static bool register_always = true;
 159 module_param(register_always, bool, 0444);
 160 MODULE_PARM_DESC(register_always,
 161          "Use memory registration even for contiguous memory regions");
 162
 163 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
 164                 struct rdma_cm_event *event);
 165 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 166
 167 /* XXX: really should move to a generic header sooner or later.. */
 168 static inline void put_unaligned_le24(u32 val, u8 *p)
 169 {
 170         *p++ = val;
 171         *p++ = val >> 8;
 172         *p++ = val >> 16;
 173 }
 174
 175 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
 176 {
 177         return queue - queue->ctrl->queues;
 178 }
 179
 180 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
 181 {
 182         return queue->cmnd_capsule_len - sizeof(struct nvme_command);
 183 }
 184
 185 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 186                 size_t capsule_size, enum dma_data_direction dir)
 187 {
 188         ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
 189         kfree(qe->data);
 190 }
 191
 192 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
 193                 size_t capsule_size, enum dma_data_direction dir)
 194 {
 195         qe->data = kzalloc(capsule_size, GFP_KERNEL);
 196         if (!qe->data)
 197                 return -ENOMEM;
 198
 199         qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
 200         if (ib_dma_mapping_error(ibdev, qe->dma)) {
 201                 kfree(qe->data);
 202                 return -ENOMEM;
 203         }
 204
 205         return 0;
 206 }
 207
 208 static void nvme_rdma_free_ring(struct ib_device *ibdev,
 209                 struct nvme_rdma_qe *ring, size_t ib_queue_size,
 210                 size_t capsule_size, enum dma_data_direction dir)
 211 {
 212         int i;
 213
 214         for (i = 0; i < ib_queue_size; i++)
 215                 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
 216         kfree(ring);
 217 }
 218
 219 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
 220                 size_t ib_queue_size, size_t capsule_size,
 221                 enum dma_data_direction dir)
 222 {
 223         struct nvme_rdma_qe *ring;
 224         int i;
 225
 226         ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
 227         if (!ring)
 228                 return NULL;
 229
 230         for (i = 0; i < ib_queue_size; i++) {
 231                 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
 232                         goto out_free_ring;
 233         }
 234
 235         return ring;
 236
 237 out_free_ring:
 238         nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
 239         return NULL;
 240 }
 241
 242 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
 243 {
 244         pr_debug("QP event %d\n", event->event);
 245 }
 246
 247 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
 248 {
 249         wait_for_completion_interruptible_timeout(&queue->cm_done,
 250                         msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
 251         return queue->cm_error;
 252 }
 253
 254 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
 255 {
 256         struct nvme_rdma_device *dev = queue->device;
 257         struct ib_qp_init_attr init_attr;
 258         int ret;
 259
 260         memset(&init_attr, 0, sizeof(init_attr));
 261         init_attr.event_handler = nvme_rdma_qp_event;
 262         /* +1 for drain */
 263         init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
 264         /* +1 for drain */
 265         init_attr.cap.max_recv_wr = queue->queue_size + 1;
 266         init_attr.cap.max_recv_sge = 1;
 267         init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
 268         init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 269         init_attr.qp_type = IB_QPT_RC;
 270         init_attr.send_cq = queue->ib_cq;
 271         init_attr.recv_cq = queue->ib_cq;
 272
 273         ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
 274
 275         queue->qp = queue->cm_id->qp;
 276         return ret;
 277 }
 278
 279 static int nvme_rdma_reinit_request(void *data, struct request *rq)
 280 {
 281         struct nvme_rdma_ctrl *ctrl = data;
 282         struct nvme_rdma_device *dev = ctrl->device;
 283         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 284         int ret = 0;
 285
 286         if (!req->mr->need_inval)
 287                 goto out;
 288
 289         ib_dereg_mr(req->mr);
 290
 291         req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
 292                         ctrl->max_fr_pages);
 293         if (IS_ERR(req->mr)) {
 294                 ret = PTR_ERR(req->mr);
 295                 req->mr = NULL;
 296                 goto out;
 297         }
 298
 299         req->mr->need_inval = false;
 300
 301 out:
 302         return ret;
 303 }
 304
 305 static void __nvme_rdma_exit_request(struct nvme_rdma_ctrl *ctrl,
 306                 struct request *rq, unsigned int queue_idx)
 307 {
 308         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 309         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 310         struct nvme_rdma_device *dev = queue->device;
 311
 312         if (req->mr)
 313                 ib_dereg_mr(req->mr);
 314
 315         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 316                         DMA_TO_DEVICE);
 317 }
 318
 319 static void nvme_rdma_exit_request(void *data, struct request *rq,
 320                                 unsigned int hctx_idx, unsigned int rq_idx)
 321 {
 322         return __nvme_rdma_exit_request(data, rq, hctx_idx + 1);
 323 }
 324
 325 static void nvme_rdma_exit_admin_request(void *data, struct request *rq,
 326                                 unsigned int hctx_idx, unsigned int rq_idx)
 327 {
 328         return __nvme_rdma_exit_request(data, rq, 0);
 329 }
 330
 331 static int __nvme_rdma_init_request(struct nvme_rdma_ctrl *ctrl,
 332                 struct request *rq, unsigned int queue_idx)
 333 {
 334         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 335         struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
 336         struct nvme_rdma_device *dev = queue->device;
 337         struct ib_device *ibdev = dev->dev;
 338         int ret;
 339
 340         ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
 341                         DMA_TO_DEVICE);
 342         if (ret)
 343                 return ret;
 344
 345         req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
 346                         ctrl->max_fr_pages);
 347         if (IS_ERR(req->mr)) {
 348                 ret = PTR_ERR(req->mr);
 349                 goto out_free_qe;
 350         }
 351
 352         req->queue = queue;
 353
 354         return 0;
 355
 356 out_free_qe:
 357         nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
 358                         DMA_TO_DEVICE);
 359         return -ENOMEM;
 360 }
 361
 362 static int nvme_rdma_init_request(void *data, struct request *rq,
 363                                 unsigned int hctx_idx, unsigned int rq_idx,
 364                                 unsigned int numa_node)
 365 {
 366         return __nvme_rdma_init_request(data, rq, hctx_idx + 1);
 367 }
 368
 369 static int nvme_rdma_init_admin_request(void *data, struct request *rq,
 370                                 unsigned int hctx_idx, unsigned int rq_idx,
 371                                 unsigned int numa_node)
 372 {
 373         return __nvme_rdma_init_request(data, rq, 0);
 374 }
 375
 376 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 377                 unsigned int hctx_idx)
 378 {
 379         struct nvme_rdma_ctrl *ctrl = data;
 380         struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
 381
 382         BUG_ON(hctx_idx >= ctrl->queue_count);
 383
 384         hctx->driver_data = queue;
 385         return 0;
 386 }
 387
 388 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
 389                 unsigned int hctx_idx)
 390 {
 391         struct nvme_rdma_ctrl *ctrl = data;
 392         struct nvme_rdma_queue *queue = &ctrl->queues[0];
 393
 394         BUG_ON(hctx_idx != 0);
 395
 396         hctx->driver_data = queue;
 397         return 0;
 398 }
 399
 400 static void nvme_rdma_free_dev(struct kref *ref)
 401 {
 402         struct nvme_rdma_device *ndev =
 403                 container_of(ref, struct nvme_rdma_device, ref);
 404
 405         mutex_lock(&device_list_mutex);
 406         list_del(&ndev->entry);
 407         mutex_unlock(&device_list_mutex);
 408
 409         ib_dealloc_pd(ndev->pd);
 410         kfree(ndev);
 411 }
 412
 413 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
 414 {
 415         kref_put(&dev->ref, nvme_rdma_free_dev);
 416 }
 417
 418 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
 419 {
 420         return kref_get_unless_zero(&dev->ref);
 421 }
 422
 423 static struct nvme_rdma_device *
 424 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
 425 {
 426         struct nvme_rdma_device *ndev;
 427
 428         mutex_lock(&device_list_mutex);
 429         list_for_each_entry(ndev, &device_list, entry) {
 430                 if (ndev->dev->node_guid == cm_id->device->node_guid &&
 431                     nvme_rdma_dev_get(ndev))
 432                         goto out_unlock;
 433         }
 434
 435         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 436         if (!ndev)
 437                 goto out_err;
 438
 439         ndev->dev = cm_id->device;
 440         kref_init(&ndev->ref);
 441
 442         ndev->pd = ib_alloc_pd(ndev->dev,
 443                 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
 444         if (IS_ERR(ndev->pd))
 445                 goto out_free_dev;
 446
 447         if (!(ndev->dev->attrs.device_cap_flags &
 448               IB_DEVICE_MEM_MGT_EXTENSIONS)) {
 449                 dev_err(&ndev->dev->dev,
 450                         "Memory registrations not supported.\n");
 451                 goto out_free_pd;
 452         }
 453
 454         list_add(&ndev->entry, &device_list);
 455 out_unlock:
 456         mutex_unlock(&device_list_mutex);
 457         return ndev;
 458
 459 out_free_pd:
 460         ib_dealloc_pd(ndev->pd);
 461 out_free_dev:
 462         kfree(ndev);
 463 out_err:
 464         mutex_unlock(&device_list_mutex);
 465         return NULL;
 466 }
 467
 468 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
 469 {
 470         struct nvme_rdma_device *dev;
 471         struct ib_device *ibdev;
 472
 473         if (!test_and_clear_bit(NVME_RDMA_IB_QUEUE_ALLOCATED, &queue->flags))
 474                 return;
 475
 476         dev = queue->device;
 477         ibdev = dev->dev;
 478         rdma_destroy_qp(queue->cm_id);
 479         ib_free_cq(queue->ib_cq);
 480
 481         nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
 482                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 483
 484         nvme_rdma_dev_put(dev);
 485 }
 486
 487 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue,
 488                 struct nvme_rdma_device *dev)
 489 {
 490         struct ib_device *ibdev = dev->dev;
 491         const int send_wr_factor = 3;                   /* MR, SEND, INV */
 492         const int cq_factor = send_wr_factor + 1;       /* + RECV */
 493         int comp_vector, idx = nvme_rdma_queue_idx(queue);
 494
 495         int ret;
 496
 497         queue->device = dev;
 498
 499         /*
 500          * The admin queue is barely used once the controller is live, so don't
 501          * bother to spread it out.
 502          */
 503         if (idx == 0)
 504                 comp_vector = 0;
 505         else
 506                 comp_vector = idx % ibdev->num_comp_vectors;
 507
 508
 509         /* +1 for ib_stop_cq */
 510         queue->ib_cq = ib_alloc_cq(dev->dev, queue,
 511                                 cq_factor * queue->queue_size + 1, comp_vector,
 512                                 IB_POLL_SOFTIRQ);
 513         if (IS_ERR(queue->ib_cq)) {
 514                 ret = PTR_ERR(queue->ib_cq);
 515                 goto out;
 516         }
 517
 518         ret = nvme_rdma_create_qp(queue, send_wr_factor);
 519         if (ret)
 520                 goto out_destroy_ib_cq;
 521
 522         queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
 523                         sizeof(struct nvme_completion), DMA_FROM_DEVICE);
 524         if (!queue->rsp_ring) {
 525                 ret = -ENOMEM;
 526                 goto out_destroy_qp;
 527         }
 528         set_bit(NVME_RDMA_IB_QUEUE_ALLOCATED, &queue->flags);
 529
 530         return 0;
 531
 532 out_destroy_qp:
 533         ib_destroy_qp(queue->qp);
 534 out_destroy_ib_cq:
 535         ib_free_cq(queue->ib_cq);
 536 out:
 537         return ret;
 538 }
 539
 540 static int nvme_rdma_init_queue(struct nvme_rdma_ctrl *ctrl,
 541                 int idx, size_t queue_size)
 542 {
 543         struct nvme_rdma_queue *queue;
 544         int ret;
 545
 546         queue = &ctrl->queues[idx];
 547         queue->ctrl = ctrl;
 548         init_completion(&queue->cm_done);
 549
 550         if (idx > 0)
 551                 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
 552         else
 553                 queue->cmnd_capsule_len = sizeof(struct nvme_command);
 554
 555         queue->queue_size = queue_size;
 556         atomic_set(&queue->sig_count, 0);
 557
 558         queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
 559                         RDMA_PS_TCP, IB_QPT_RC);
 560         if (IS_ERR(queue->cm_id)) {
 561                 dev_info(ctrl->ctrl.device,
 562                         "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
 563                 return PTR_ERR(queue->cm_id);
 564         }
 565
 566         queue->cm_error = -ETIMEDOUT;
 567         ret = rdma_resolve_addr(queue->cm_id, NULL, &ctrl->addr,
 568                         NVME_RDMA_CONNECT_TIMEOUT_MS);
 569         if (ret) {
 570                 dev_info(ctrl->ctrl.device,
 571                         "rdma_resolve_addr failed (%d).\n", ret);
 572                 goto out_destroy_cm_id;
 573         }
 574
 575         ret = nvme_rdma_wait_for_cm(queue);
 576         if (ret) {
 577                 dev_info(ctrl->ctrl.device,
 578                         "rdma_resolve_addr wait failed (%d).\n", ret);
 579                 goto out_destroy_cm_id;
 580         }
 581
 582         clear_bit(NVME_RDMA_Q_DELETING, &queue->flags);
 583         set_bit(NVME_RDMA_Q_CONNECTED, &queue->flags);
 584
 585         return 0;
 586
 587 out_destroy_cm_id:
 588         nvme_rdma_destroy_queue_ib(queue);
 589         rdma_destroy_id(queue->cm_id);
 590         return ret;
 591 }
 592
 593 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
 594 {
 595         rdma_disconnect(queue->cm_id);
 596         ib_drain_qp(queue->qp);
 597 }
 598
 599 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
 600 {
 601         nvme_rdma_destroy_queue_ib(queue);
 602         rdma_destroy_id(queue->cm_id);
 603 }
 604
 605 static void nvme_rdma_stop_and_free_queue(struct nvme_rdma_queue *queue)
 606 {
 607         if (test_and_set_bit(NVME_RDMA_Q_DELETING, &queue->flags))
 608                 return;
 609         nvme_rdma_stop_queue(queue);
 610         nvme_rdma_free_queue(queue);
 611 }
 612
 613 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
 614 {
 615         int i;
 616
 617         for (i = 1; i < ctrl->queue_count; i++)
 618                 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
 619 }
 620
 621 static int nvme_rdma_connect_io_queues(struct nvme_rdma_ctrl *ctrl)
 622 {
 623         int i, ret = 0;
 624
 625         for (i = 1; i < ctrl->queue_count; i++) {
 626                 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
 627                 if (ret) {
 628                         dev_info(ctrl->ctrl.device,
 629                                 "failed to connect i/o queue: %d\n", ret);
 630                         goto out_free_queues;
 631                 }
 632                 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[i].flags);
 633         }
 634
 635         return 0;
 636
 637 out_free_queues:
 638         nvme_rdma_free_io_queues(ctrl);
 639         return ret;
 640 }
 641
 642 static int nvme_rdma_init_io_queues(struct nvme_rdma_ctrl *ctrl)
 643 {
 644         struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
 645         unsigned int nr_io_queues;
 646         int i, ret;
 647
 648         nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
 649         ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
 650         if (ret)
 651                 return ret;
 652
 653         ctrl->queue_count = nr_io_queues + 1;
 654         if (ctrl->queue_count < 2)
 655                 return 0;
 656
 657         dev_info(ctrl->ctrl.device,
 658                 "creating %d I/O queues.\n", nr_io_queues);
 659
 660         for (i = 1; i < ctrl->queue_count; i++) {
 661                 ret = nvme_rdma_init_queue(ctrl, i,
 662                                            ctrl->ctrl.opts->queue_size);
 663                 if (ret) {
 664                         dev_info(ctrl->ctrl.device,
 665                                 "failed to initialize i/o queue: %d\n", ret);
 666                         goto out_free_queues;
 667                 }
 668         }
 669
 670         return 0;
 671
 672 out_free_queues:
 673         for (i--; i >= 1; i--)
 674                 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
 675
 676         return ret;
 677 }
 678
 679 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl)
 680 {
 681         nvme_rdma_free_qe(ctrl->queues[0].device->dev, &ctrl->async_event_sqe,
 682                         sizeof(struct nvme_command), DMA_TO_DEVICE);
 683         nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
 684         blk_cleanup_queue(ctrl->ctrl.admin_q);
 685         blk_mq_free_tag_set(&ctrl->admin_tag_set);
 686         nvme_rdma_dev_put(ctrl->device);
 687 }
 688
 689 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
 690 {
 691         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
 692
 693         if (list_empty(&ctrl->list))
 694                 goto free_ctrl;
 695
 696         mutex_lock(&nvme_rdma_ctrl_mutex);
 697         list_del(&ctrl->list);
 698         mutex_unlock(&nvme_rdma_ctrl_mutex);
 699
 700         kfree(ctrl->queues);
 701         nvmf_free_options(nctrl->opts);
 702 free_ctrl:
 703         kfree(ctrl);
 704 }
 705
 706 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
 707 {
 708         struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
 709                         struct nvme_rdma_ctrl, reconnect_work);
 710         bool changed;
 711         int ret;
 712
 713         if (ctrl->queue_count > 1) {
 714                 nvme_rdma_free_io_queues(ctrl);
 715
 716                 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
 717                 if (ret)
 718                         goto requeue;
 719         }
 720
 721         nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
 722
 723         ret = blk_mq_reinit_tagset(&ctrl->admin_tag_set);
 724         if (ret)
 725                 goto requeue;
 726
 727         ret = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
 728         if (ret)
 729                 goto requeue;
 730
 731         blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true);
 732
 733         ret = nvmf_connect_admin_queue(&ctrl->ctrl);
 734         if (ret)
 735                 goto stop_admin_q;
 736
 737         set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags);
 738
 739         ret = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
 740         if (ret)
 741                 goto stop_admin_q;
 742
 743         nvme_start_keep_alive(&ctrl->ctrl);
 744
 745         if (ctrl->queue_count > 1) {
 746                 ret = nvme_rdma_init_io_queues(ctrl);
 747                 if (ret)
 748                         goto stop_admin_q;
 749
 750                 ret = nvme_rdma_connect_io_queues(ctrl);
 751                 if (ret)
 752                         goto stop_admin_q;
 753         }
 754
 755         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
 756         WARN_ON_ONCE(!changed);
 757
 758         if (ctrl->queue_count > 1) {
 759                 nvme_start_queues(&ctrl->ctrl);
 760                 nvme_queue_scan(&ctrl->ctrl);
 761                 nvme_queue_async_events(&ctrl->ctrl);
 762         }
 763
 764         dev_info(ctrl->ctrl.device, "Successfully reconnected\n");
 765
 766         return;
 767
 768 stop_admin_q:
 769         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
 770 requeue:
 771         /* Make sure we are not resetting/deleting */
 772         if (ctrl->ctrl.state == NVME_CTRL_RECONNECTING) {
 773                 dev_info(ctrl->ctrl.device,
 774                         "Failed reconnect attempt, requeueing...\n");
 775                 queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work,
 776                                         ctrl->reconnect_delay * HZ);
 777         }
 778 }
 779
 780 static void nvme_rdma_error_recovery_work(struct work_struct *work)
 781 {
 782         struct nvme_rdma_ctrl *ctrl = container_of(work,
 783                         struct nvme_rdma_ctrl, err_work);
 784         int i;
 785
 786         nvme_stop_keep_alive(&ctrl->ctrl);
 787
 788         for (i = 0; i < ctrl->queue_count; i++) {
 789                 clear_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[i].flags);
 790                 clear_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[i].flags);
 791         }
 792
 793         if (ctrl->queue_count > 1)
 794                 nvme_stop_queues(&ctrl->ctrl);
 795         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
 796
 797         /* We must take care of fastfail/requeue all our inflight requests */
 798         if (ctrl->queue_count > 1)
 799                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
 800                                         nvme_cancel_request, &ctrl->ctrl);
 801         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
 802                                 nvme_cancel_request, &ctrl->ctrl);
 803
 804         dev_info(ctrl->ctrl.device, "reconnecting in %d seconds\n",
 805                 ctrl->reconnect_delay);
 806
 807         queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work,
 808                                 ctrl->reconnect_delay * HZ);
 809 }
 810
 811 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
 812 {
 813         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING))
 814                 return;
 815
 816         queue_work(nvme_rdma_wq, &ctrl->err_work);
 817 }
 818
 819 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
 820                 const char *op)
 821 {
 822         struct nvme_rdma_queue *queue = cq->cq_context;
 823         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
 824
 825         if (ctrl->ctrl.state == NVME_CTRL_LIVE)
 826                 dev_info(ctrl->ctrl.device,
 827                              "%s for CQE 0x%p failed with status %s (%d)\n",
 828                              op, wc->wr_cqe,
 829                              ib_wc_status_msg(wc->status), wc->status);
 830         nvme_rdma_error_recovery(ctrl);
 831 }
 832
 833 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
 834 {
 835         if (unlikely(wc->status != IB_WC_SUCCESS))
 836                 nvme_rdma_wr_error(cq, wc, "MEMREG");
 837 }
 838
 839 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
 840 {
 841         if (unlikely(wc->status != IB_WC_SUCCESS))
 842                 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
 843 }
 844
 845 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
 846                 struct nvme_rdma_request *req)
 847 {
 848         struct ib_send_wr *bad_wr;
 849         struct ib_send_wr wr = {
 850                 .opcode             = IB_WR_LOCAL_INV,
 851                 .next               = NULL,
 852                 .num_sge            = 0,
 853                 .send_flags         = 0,
 854                 .ex.invalidate_rkey = req->mr->rkey,
 855         };
 856
 857         req->reg_cqe.done = nvme_rdma_inv_rkey_done;
 858         wr.wr_cqe = &req->reg_cqe;
 859
 860         return ib_post_send(queue->qp, &wr, &bad_wr);
 861 }
 862
 863 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
 864                 struct request *rq)
 865 {
 866         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 867         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
 868         struct nvme_rdma_device *dev = queue->device;
 869         struct ib_device *ibdev = dev->dev;
 870         int res;
 871
 872         if (!blk_rq_bytes(rq))
 873                 return;
 874
 875         if (req->mr->need_inval) {
 876                 res = nvme_rdma_inv_rkey(queue, req);
 877                 if (res < 0) {
 878                         dev_err(ctrl->ctrl.device,
 879                                 "Queueing INV WR for rkey %#x failed (%d)\n",
 880                                 req->mr->rkey, res);
 881                         nvme_rdma_error_recovery(queue->ctrl);
 882                 }
 883         }
 884
 885         ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
 886                         req->nents, rq_data_dir(rq) ==
 887                                     WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
 888
 889         nvme_cleanup_cmd(rq);
 890         sg_free_table_chained(&req->sg_table, true);
 891 }
 892
 893 static int nvme_rdma_set_sg_null(struct nvme_command *c)
 894 {
 895         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 896
 897         sg->addr = 0;
 898         put_unaligned_le24(0, sg->length);
 899         put_unaligned_le32(0, sg->key);
 900         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
 901         return 0;
 902 }
 903
 904 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
 905                 struct nvme_rdma_request *req, struct nvme_command *c)
 906 {
 907         struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
 908
 909         req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
 910         req->sge[1].length = sg_dma_len(req->sg_table.sgl);
 911         req->sge[1].lkey = queue->device->pd->local_dma_lkey;
 912
 913         sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
 914         sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
 915         sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
 916
 917         req->inline_data = true;
 918         req->num_sge++;
 919         return 0;
 920 }
 921
 922 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
 923                 struct nvme_rdma_request *req, struct nvme_command *c)
 924 {
 925         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 926
 927         sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
 928         put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
 929         put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
 930         sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
 931         return 0;
 932 }
 933
 934 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
 935                 struct nvme_rdma_request *req, struct nvme_command *c,
 936                 int count)
 937 {
 938         struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
 939         int nr;
 940
 941         nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, PAGE_SIZE);
 942         if (nr < count) {
 943                 if (nr < 0)
 944                         return nr;
 945                 return -EINVAL;
 946         }
 947
 948         ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
 949
 950         req->reg_cqe.done = nvme_rdma_memreg_done;
 951         memset(&req->reg_wr, 0, sizeof(req->reg_wr));
 952         req->reg_wr.wr.opcode = IB_WR_REG_MR;
 953         req->reg_wr.wr.wr_cqe = &req->reg_cqe;
 954         req->reg_wr.wr.num_sge = 0;
 955         req->reg_wr.mr = req->mr;
 956         req->reg_wr.key = req->mr->rkey;
 957         req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
 958                              IB_ACCESS_REMOTE_READ |
 959                              IB_ACCESS_REMOTE_WRITE;
 960
 961         req->mr->need_inval = true;
 962
 963         sg->addr = cpu_to_le64(req->mr->iova);
 964         put_unaligned_le24(req->mr->length, sg->length);
 965         put_unaligned_le32(req->mr->rkey, sg->key);
 966         sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
 967                         NVME_SGL_FMT_INVALIDATE;
 968
 969         return 0;
 970 }
 971
 972 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
 973                 struct request *rq, unsigned int map_len,
 974                 struct nvme_command *c)
 975 {
 976         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
 977         struct nvme_rdma_device *dev = queue->device;
 978         struct ib_device *ibdev = dev->dev;
 979         int nents, count;
 980         int ret;
 981
 982         req->num_sge = 1;
 983         req->inline_data = false;
 984         req->mr->need_inval = false;
 985
 986         c->common.flags |= NVME_CMD_SGL_METABUF;
 987
 988         if (!blk_rq_bytes(rq))
 989                 return nvme_rdma_set_sg_null(c);
 990
 991         req->sg_table.sgl = req->first_sgl;
 992         ret = sg_alloc_table_chained(&req->sg_table, rq->nr_phys_segments,
 993                                 req->sg_table.sgl);
 994         if (ret)
 995                 return -ENOMEM;
 996
 997         nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
 998         BUG_ON(nents > rq->nr_phys_segments);
 999         req->nents = nents;
1000
1001         count = ib_dma_map_sg(ibdev, req->sg_table.sgl, nents,
1002                     rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1003         if (unlikely(count <= 0)) {
1004                 sg_free_table_chained(&req->sg_table, true);
1005                 return -EIO;
1006         }
1007
1008         if (count == 1) {
1009                 if (rq_data_dir(rq) == WRITE &&
1010                     map_len <= nvme_rdma_inline_data_size(queue) &&
1011                     nvme_rdma_queue_idx(queue))
1012                         return nvme_rdma_map_sg_inline(queue, req, c);
1013
1014                 if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY)
1015                         return nvme_rdma_map_sg_single(queue, req, c);
1016         }
1017
1018         return nvme_rdma_map_sg_fr(queue, req, c, count);
1019 }
1020
1021 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1022 {
1023         if (unlikely(wc->status != IB_WC_SUCCESS))
1024                 nvme_rdma_wr_error(cq, wc, "SEND");
1025 }
1026
1027 /*
1028  * We want to signal completion at least every queue depth/2.  This returns the
1029  * largest power of two that is not above half of (queue size + 1) to optimize
1030  * (avoid divisions).
1031  */
1032 static inline bool nvme_rdma_queue_sig_limit(struct nvme_rdma_queue *queue)
1033 {
1034         int limit = 1 << ilog2((queue->queue_size + 1) / 2);
1035
1036         return (atomic_inc_return(&queue->sig_count) & (limit - 1)) == 0;
1037 }
1038
1039 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1040                 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1041                 struct ib_send_wr *first, bool flush)
1042 {
1043         struct ib_send_wr wr, *bad_wr;
1044         int ret;
1045
1046         sge->addr   = qe->dma;
1047         sge->length = sizeof(struct nvme_command),
1048         sge->lkey   = queue->device->pd->local_dma_lkey;
1049
1050         qe->cqe.done = nvme_rdma_send_done;
1051
1052         wr.next       = NULL;
1053         wr.wr_cqe     = &qe->cqe;
1054         wr.sg_list    = sge;
1055         wr.num_sge    = num_sge;
1056         wr.opcode     = IB_WR_SEND;
1057         wr.send_flags = 0;
1058
1059         /*
1060          * Unsignalled send completions are another giant desaster in the
1061          * IB Verbs spec:  If we don't regularly post signalled sends
1062          * the send queue will fill up and only a QP reset will rescue us.
1063          * Would have been way to obvious to handle this in hardware or
1064          * at least the RDMA stack..
1065          *
1066          * Always signal the flushes. The magic request used for the flush
1067          * sequencer is not allocated in our driver's tagset and it's
1068          * triggered to be freed by blk_cleanup_queue(). So we need to
1069          * always mark it as signaled to ensure that the "wr_cqe", which is
1070          * embeded in request's payload, is not freed when __ib_process_cq()
1071          * calls wr_cqe->done().
1072          */
1073         if (nvme_rdma_queue_sig_limit(queue) || flush)
1074                 wr.send_flags |= IB_SEND_SIGNALED;
1075
1076         if (first)
1077                 first->next = &wr;
1078         else
1079                 first = &wr;
1080
1081         ret = ib_post_send(queue->qp, first, &bad_wr);
1082         if (ret) {
1083                 dev_err(queue->ctrl->ctrl.device,
1084                              "%s failed with error code %d\n", __func__, ret);
1085         }
1086         return ret;
1087 }
1088
1089 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1090                 struct nvme_rdma_qe *qe)
1091 {
1092         struct ib_recv_wr wr, *bad_wr;
1093         struct ib_sge list;
1094         int ret;
1095
1096         list.addr   = qe->dma;
1097         list.length = sizeof(struct nvme_completion);
1098         list.lkey   = queue->device->pd->local_dma_lkey;
1099
1100         qe->cqe.done = nvme_rdma_recv_done;
1101
1102         wr.next     = NULL;
1103         wr.wr_cqe   = &qe->cqe;
1104         wr.sg_list  = &list;
1105         wr.num_sge  = 1;
1106
1107         ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1108         if (ret) {
1109                 dev_err(queue->ctrl->ctrl.device,
1110                         "%s failed with error code %d\n", __func__, ret);
1111         }
1112         return ret;
1113 }
1114
1115 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1116 {
1117         u32 queue_idx = nvme_rdma_queue_idx(queue);
1118
1119         if (queue_idx == 0)
1120                 return queue->ctrl->admin_tag_set.tags[queue_idx];
1121         return queue->ctrl->tag_set.tags[queue_idx - 1];
1122 }
1123
1124 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
1125 {
1126         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1127         struct nvme_rdma_queue *queue = &ctrl->queues[0];
1128         struct ib_device *dev = queue->device->dev;
1129         struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1130         struct nvme_command *cmd = sqe->data;
1131         struct ib_sge sge;
1132         int ret;
1133
1134         if (WARN_ON_ONCE(aer_idx != 0))
1135                 return;
1136
1137         ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1138
1139         memset(cmd, 0, sizeof(*cmd));
1140         cmd->common.opcode = nvme_admin_async_event;
1141         cmd->common.command_id = NVME_RDMA_AQ_BLKMQ_DEPTH;
1142         cmd->common.flags |= NVME_CMD_SGL_METABUF;
1143         nvme_rdma_set_sg_null(cmd);
1144
1145         ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1146                         DMA_TO_DEVICE);
1147
1148         ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL, false);
1149         WARN_ON_ONCE(ret);
1150 }
1151
1152 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1153                 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1154 {
1155         u16 status = le16_to_cpu(cqe->status);
1156         struct request *rq;
1157         struct nvme_rdma_request *req;
1158         int ret = 0;
1159
1160         status >>= 1;
1161
1162         rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1163         if (!rq) {
1164                 dev_err(queue->ctrl->ctrl.device,
1165                         "tag 0x%x on QP %#x not found\n",
1166                         cqe->command_id, queue->qp->qp_num);
1167                 nvme_rdma_error_recovery(queue->ctrl);
1168                 return ret;
1169         }
1170         req = blk_mq_rq_to_pdu(rq);
1171
1172         if (rq->cmd_type == REQ_TYPE_DRV_PRIV && rq->special)
1173                 memcpy(rq->special, cqe, sizeof(*cqe));
1174
1175         if (rq->tag == tag)
1176                 ret = 1;
1177
1178         if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
1179             wc->ex.invalidate_rkey == req->mr->rkey)
1180                 req->mr->need_inval = false;
1181
1182         blk_mq_complete_request(rq, status);
1183
1184         return ret;
1185 }
1186
1187 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1188 {
1189         struct nvme_rdma_qe *qe =
1190                 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1191         struct nvme_rdma_queue *queue = cq->cq_context;
1192         struct ib_device *ibdev = queue->device->dev;
1193         struct nvme_completion *cqe = qe->data;
1194         const size_t len = sizeof(struct nvme_completion);
1195         int ret = 0;
1196
1197         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1198                 nvme_rdma_wr_error(cq, wc, "RECV");
1199                 return 0;
1200         }
1201
1202         ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1203         /*
1204          * AEN requests are special as they don't time out and can
1205          * survive any kind of queue freeze and often don't respond to
1206          * aborts.  We don't even bother to allocate a struct request
1207          * for them but rather special case them here.
1208          */
1209         if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1210                         cqe->command_id >= NVME_RDMA_AQ_BLKMQ_DEPTH))
1211                 nvme_complete_async_event(&queue->ctrl->ctrl, cqe);
1212         else
1213                 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1214         ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1215
1216         nvme_rdma_post_recv(queue, qe);
1217         return ret;
1218 }
1219
1220 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1221 {
1222         __nvme_rdma_recv_done(cq, wc, -1);
1223 }
1224
1225 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1226 {
1227         int ret, i;
1228
1229         for (i = 0; i < queue->queue_size; i++) {
1230                 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1231                 if (ret)
1232                         goto out_destroy_queue_ib;
1233         }
1234
1235         return 0;
1236
1237 out_destroy_queue_ib:
1238         nvme_rdma_destroy_queue_ib(queue);
1239         return ret;
1240 }
1241
1242 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1243                 struct rdma_cm_event *ev)
1244 {
1245         if (ev->param.conn.private_data_len) {
1246                 struct nvme_rdma_cm_rej *rej =
1247                         (struct nvme_rdma_cm_rej *)ev->param.conn.private_data;
1248
1249                 dev_err(queue->ctrl->ctrl.device,
1250                         "Connect rejected, status %d.", le16_to_cpu(rej->sts));
1251                 /* XXX: Think of something clever to do here... */
1252         } else {
1253                 dev_err(queue->ctrl->ctrl.device,
1254                         "Connect rejected, no private data.\n");
1255         }
1256
1257         return -ECONNRESET;
1258 }
1259
1260 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1261 {
1262         struct nvme_rdma_device *dev;
1263         int ret;
1264
1265         dev = nvme_rdma_find_get_device(queue->cm_id);
1266         if (!dev) {
1267                 dev_err(queue->cm_id->device->dma_device,
1268                         "no client data found!\n");
1269                 return -ECONNREFUSED;
1270         }
1271
1272         ret = nvme_rdma_create_queue_ib(queue, dev);
1273         if (ret) {
1274                 nvme_rdma_dev_put(dev);
1275                 goto out;
1276         }
1277
1278         ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1279         if (ret) {
1280                 dev_err(queue->ctrl->ctrl.device,
1281                         "rdma_resolve_route failed (%d).\n",
1282                         queue->cm_error);
1283                 goto out_destroy_queue;
1284         }
1285
1286         return 0;
1287
1288 out_destroy_queue:
1289         nvme_rdma_destroy_queue_ib(queue);
1290 out:
1291         return ret;
1292 }
1293
1294 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1295 {
1296         struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1297         struct rdma_conn_param param = { };
1298         struct nvme_rdma_cm_req priv = { };
1299         int ret;
1300
1301         param.qp_num = queue->qp->qp_num;
1302         param.flow_control = 1;
1303
1304         param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1305         /* maximum retry count */
1306         param.retry_count = 7;
1307         param.rnr_retry_count = 7;
1308         param.private_data = &priv;
1309         param.private_data_len = sizeof(priv);
1310
1311         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1312         priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1313         /*
1314          * set the admin queue depth to the minimum size
1315          * specified by the Fabrics standard.
1316          */
1317         if (priv.qid == 0) {
1318                 priv.hrqsize = cpu_to_le16(NVMF_AQ_DEPTH);
1319                 priv.hsqsize = cpu_to_le16(NVMF_AQ_DEPTH - 1);
1320         } else {
1321                 /*
1322                  * current interpretation of the fabrics spec
1323                  * is at minimum you make hrqsize sqsize+1, or a
1324                  * 1's based representation of sqsize.
1325                  */
1326                 priv.hrqsize = cpu_to_le16(queue->queue_size);
1327                 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1328         }
1329
1330         ret = rdma_connect(queue->cm_id, &param);
1331         if (ret) {
1332                 dev_err(ctrl->ctrl.device,
1333                         "rdma_connect failed (%d).\n", ret);
1334                 goto out_destroy_queue_ib;
1335         }
1336
1337         return 0;
1338
1339 out_destroy_queue_ib:
1340         nvme_rdma_destroy_queue_ib(queue);
1341         return ret;
1342 }
1343
1344 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1345                 struct rdma_cm_event *ev)
1346 {
1347         struct nvme_rdma_queue *queue = cm_id->context;
1348         int cm_error = 0;
1349
1350         dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1351                 rdma_event_msg(ev->event), ev->event,
1352                 ev->status, cm_id);
1353
1354         switch (ev->event) {
1355         case RDMA_CM_EVENT_ADDR_RESOLVED:
1356                 cm_error = nvme_rdma_addr_resolved(queue);
1357                 break;
1358         case RDMA_CM_EVENT_ROUTE_RESOLVED:
1359                 cm_error = nvme_rdma_route_resolved(queue);
1360                 break;
1361         case RDMA_CM_EVENT_ESTABLISHED:
1362                 queue->cm_error = nvme_rdma_conn_established(queue);
1363                 /* complete cm_done regardless of success/failure */
1364                 complete(&queue->cm_done);
1365                 return 0;
1366         case RDMA_CM_EVENT_REJECTED:
1367                 cm_error = nvme_rdma_conn_rejected(queue, ev);
1368                 break;
1369         case RDMA_CM_EVENT_ADDR_ERROR:
1370         case RDMA_CM_EVENT_ROUTE_ERROR:
1371         case RDMA_CM_EVENT_CONNECT_ERROR:
1372         case RDMA_CM_EVENT_UNREACHABLE:
1373                 dev_dbg(queue->ctrl->ctrl.device,
1374                         "CM error event %d\n", ev->event);
1375                 cm_error = -ECONNRESET;
1376                 break;
1377         case RDMA_CM_EVENT_DISCONNECTED:
1378         case RDMA_CM_EVENT_ADDR_CHANGE:
1379         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1380                 dev_dbg(queue->ctrl->ctrl.device,
1381                         "disconnect received - connection closed\n");
1382                 nvme_rdma_error_recovery(queue->ctrl);
1383                 break;
1384         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1385                 /* device removal is handled via the ib_client API */
1386                 break;
1387         default:
1388                 dev_err(queue->ctrl->ctrl.device,
1389                         "Unexpected RDMA CM event (%d)\n", ev->event);
1390                 nvme_rdma_error_recovery(queue->ctrl);
1391                 break;
1392         }
1393
1394         if (cm_error) {
1395                 queue->cm_error = cm_error;
1396                 complete(&queue->cm_done);
1397         }
1398
1399         return 0;
1400 }
1401
1402 static enum blk_eh_timer_return
1403 nvme_rdma_timeout(struct request *rq, bool reserved)
1404 {
1405         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1406
1407         /* queue error recovery */
1408         nvme_rdma_error_recovery(req->queue->ctrl);
1409
1410         /* fail with DNR on cmd timeout */
1411         rq->errors = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1412
1413         return BLK_EH_HANDLED;
1414 }
1415
1416 /*
1417  * We cannot accept any other command until the Connect command has completed.
1418  */
1419 static inline bool nvme_rdma_queue_is_ready(struct nvme_rdma_queue *queue,
1420                 struct request *rq)
1421 {
1422         if (unlikely(!test_bit(NVME_RDMA_Q_LIVE, &queue->flags))) {
1423                 struct nvme_command *cmd = (struct nvme_command *)rq->cmd;
1424
1425                 if (rq->cmd_type != REQ_TYPE_DRV_PRIV ||
1426                     cmd->common.opcode != nvme_fabrics_command ||
1427                     cmd->fabrics.fctype != nvme_fabrics_type_connect)
1428                         return false;
1429         }
1430
1431         return true;
1432 }
1433
1434 static int nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1435                 const struct blk_mq_queue_data *bd)
1436 {
1437         struct nvme_ns *ns = hctx->queue->queuedata;
1438         struct nvme_rdma_queue *queue = hctx->driver_data;
1439         struct request *rq = bd->rq;
1440         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1441         struct nvme_rdma_qe *sqe = &req->sqe;
1442         struct nvme_command *c = sqe->data;
1443         bool flush = false;
1444         struct ib_device *dev;
1445         unsigned int map_len;
1446         int ret;
1447
1448         WARN_ON_ONCE(rq->tag < 0);
1449
1450         if (!nvme_rdma_queue_is_ready(queue, rq))
1451                 return BLK_MQ_RQ_QUEUE_BUSY;
1452
1453         dev = queue->device->dev;
1454         ib_dma_sync_single_for_cpu(dev, sqe->dma,
1455                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1456
1457         ret = nvme_setup_cmd(ns, rq, c);
1458         if (ret)
1459                 return ret;
1460
1461         c->common.command_id = rq->tag;
1462         blk_mq_start_request(rq);
1463
1464         map_len = nvme_map_len(rq);
1465         ret = nvme_rdma_map_data(queue, rq, map_len, c);
1466         if (ret < 0) {
1467                 dev_err(queue->ctrl->ctrl.device,
1468                              "Failed to map data (%d)\n", ret);
1469                 nvme_cleanup_cmd(rq);
1470                 goto err;
1471         }
1472
1473         ib_dma_sync_single_for_device(dev, sqe->dma,
1474                         sizeof(struct nvme_command), DMA_TO_DEVICE);
1475
1476         if (rq->cmd_type == REQ_TYPE_FS && req_op(rq) == REQ_OP_FLUSH)
1477                 flush = true;
1478         ret = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1479                         req->mr->need_inval ? &req->reg_wr.wr : NULL, flush);
1480         if (ret) {
1481                 nvme_rdma_unmap_data(queue, rq);
1482                 goto err;
1483         }
1484
1485         return BLK_MQ_RQ_QUEUE_OK;
1486 err:
1487         return (ret == -ENOMEM || ret == -EAGAIN) ?
1488                 BLK_MQ_RQ_QUEUE_BUSY : BLK_MQ_RQ_QUEUE_ERROR;
1489 }
1490
1491 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1492 {
1493         struct nvme_rdma_queue *queue = hctx->driver_data;
1494         struct ib_cq *cq = queue->ib_cq;
1495         struct ib_wc wc;
1496         int found = 0;
1497
1498         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1499         while (ib_poll_cq(cq, 1, &wc) > 0) {
1500                 struct ib_cqe *cqe = wc.wr_cqe;
1501
1502                 if (cqe) {
1503                         if (cqe->done == nvme_rdma_recv_done)
1504                                 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1505                         else
1506                                 cqe->done(cq, &wc);
1507                 }
1508         }
1509
1510         return found;
1511 }
1512
1513 static void nvme_rdma_complete_rq(struct request *rq)
1514 {
1515         struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1516         struct nvme_rdma_queue *queue = req->queue;
1517         int error = 0;
1518
1519         nvme_rdma_unmap_data(queue, rq);
1520
1521         if (unlikely(rq->errors)) {
1522                 if (nvme_req_needs_retry(rq, rq->errors)) {
1523                         nvme_requeue_req(rq);
1524                         return;
1525                 }
1526
1527                 if (rq->cmd_type == REQ_TYPE_DRV_PRIV)
1528                         error = rq->errors;
1529                 else
1530                         error = nvme_error_status(rq->errors);
1531         }
1532
1533         blk_mq_end_request(rq, error);
1534 }
1535
1536 static struct blk_mq_ops nvme_rdma_mq_ops = {
1537         .queue_rq       = nvme_rdma_queue_rq,
1538         .complete       = nvme_rdma_complete_rq,
1539         .init_request   = nvme_rdma_init_request,
1540         .exit_request   = nvme_rdma_exit_request,
1541         .reinit_request = nvme_rdma_reinit_request,
1542         .init_hctx      = nvme_rdma_init_hctx,
1543         .poll           = nvme_rdma_poll,
1544         .timeout        = nvme_rdma_timeout,
1545 };
1546
1547 static struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1548         .queue_rq       = nvme_rdma_queue_rq,
1549         .complete       = nvme_rdma_complete_rq,
1550         .init_request   = nvme_rdma_init_admin_request,
1551         .exit_request   = nvme_rdma_exit_admin_request,
1552         .reinit_request = nvme_rdma_reinit_request,
1553         .init_hctx      = nvme_rdma_init_admin_hctx,
1554         .timeout        = nvme_rdma_timeout,
1555 };
1556
1557 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl)
1558 {
1559         int error;
1560
1561         error = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
1562         if (error)
1563                 return error;
1564
1565         ctrl->device = ctrl->queues[0].device;
1566
1567         /*
1568          * We need a reference on the device as long as the tag_set is alive,
1569          * as the MRs in the request structures need a valid ib_device.
1570          */
1571         error = -EINVAL;
1572         if (!nvme_rdma_dev_get(ctrl->device))
1573                 goto out_free_queue;
1574
1575         ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
1576                 ctrl->device->dev->attrs.max_fast_reg_page_list_len);
1577
1578         memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
1579         ctrl->admin_tag_set.ops = &nvme_rdma_admin_mq_ops;
1580         ctrl->admin_tag_set.queue_depth = NVME_RDMA_AQ_BLKMQ_DEPTH;
1581         ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
1582         ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
1583         ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1584                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1585         ctrl->admin_tag_set.driver_data = ctrl;
1586         ctrl->admin_tag_set.nr_hw_queues = 1;
1587         ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
1588
1589         error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
1590         if (error)
1591                 goto out_put_dev;
1592
1593         ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
1594         if (IS_ERR(ctrl->ctrl.admin_q)) {
1595                 error = PTR_ERR(ctrl->ctrl.admin_q);
1596                 goto out_free_tagset;
1597         }
1598
1599         error = nvmf_connect_admin_queue(&ctrl->ctrl);
1600         if (error)
1601                 goto out_cleanup_queue;
1602
1603         set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[0].flags);
1604
1605         error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap);
1606         if (error) {
1607                 dev_err(ctrl->ctrl.device,
1608                         "prop_get NVME_REG_CAP failed\n");
1609                 goto out_cleanup_queue;
1610         }
1611
1612         ctrl->ctrl.sqsize =
1613                 min_t(int, NVME_CAP_MQES(ctrl->cap) + 1, ctrl->ctrl.sqsize);
1614
1615         error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
1616         if (error)
1617                 goto out_cleanup_queue;
1618
1619         ctrl->ctrl.max_hw_sectors =
1620                 (ctrl->max_fr_pages - 1) << (PAGE_SHIFT - 9);
1621
1622         error = nvme_init_identify(&ctrl->ctrl);
1623         if (error)
1624                 goto out_cleanup_queue;
1625
1626         error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
1627                         &ctrl->async_event_sqe, sizeof(struct nvme_command),
1628                         DMA_TO_DEVICE);
1629         if (error)
1630                 goto out_cleanup_queue;
1631
1632         nvme_start_keep_alive(&ctrl->ctrl);
1633
1634         return 0;
1635
1636 out_cleanup_queue:
1637         blk_cleanup_queue(ctrl->ctrl.admin_q);
1638 out_free_tagset:
1639         /* disconnect and drain the queue before freeing the tagset */
1640         nvme_rdma_stop_queue(&ctrl->queues[0]);
1641         blk_mq_free_tag_set(&ctrl->admin_tag_set);
1642 out_put_dev:
1643         nvme_rdma_dev_put(ctrl->device);
1644 out_free_queue:
1645         nvme_rdma_free_queue(&ctrl->queues[0]);
1646         return error;
1647 }
1648
1649 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl)
1650 {
1651         nvme_stop_keep_alive(&ctrl->ctrl);
1652         cancel_work_sync(&ctrl->err_work);
1653         cancel_delayed_work_sync(&ctrl->reconnect_work);
1654
1655         if (ctrl->queue_count > 1) {
1656                 nvme_stop_queues(&ctrl->ctrl);
1657                 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1658                                         nvme_cancel_request, &ctrl->ctrl);
1659                 nvme_rdma_free_io_queues(ctrl);
1660         }
1661
1662         if (test_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[0].flags))
1663                 nvme_shutdown_ctrl(&ctrl->ctrl);
1664
1665         blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
1666         blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1667                                 nvme_cancel_request, &ctrl->ctrl);
1668         nvme_rdma_destroy_admin_queue(ctrl);
1669 }
1670
1671 static void __nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1672 {
1673         nvme_uninit_ctrl(&ctrl->ctrl);
1674         if (shutdown)
1675                 nvme_rdma_shutdown_ctrl(ctrl);
1676
1677         if (ctrl->ctrl.tagset) {
1678                 blk_cleanup_queue(ctrl->ctrl.connect_q);
1679                 blk_mq_free_tag_set(&ctrl->tag_set);
1680                 nvme_rdma_dev_put(ctrl->device);
1681         }
1682
1683         nvme_put_ctrl(&ctrl->ctrl);
1684 }
1685
1686 static void nvme_rdma_del_ctrl_work(struct work_struct *work)
1687 {
1688         struct nvme_rdma_ctrl *ctrl = container_of(work,
1689                                 struct nvme_rdma_ctrl, delete_work);
1690
1691         __nvme_rdma_remove_ctrl(ctrl, true);
1692 }
1693
1694 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl *ctrl)
1695 {
1696         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1697                 return -EBUSY;
1698
1699         if (!queue_work(nvme_rdma_wq, &ctrl->delete_work))
1700                 return -EBUSY;
1701
1702         return 0;
1703 }
1704
1705 static int nvme_rdma_del_ctrl(struct nvme_ctrl *nctrl)
1706 {
1707         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1708         int ret = 0;
1709
1710         /*
1711          * Keep a reference until all work is flushed since
1712          * __nvme_rdma_del_ctrl can free the ctrl mem
1713          */
1714         if (!kref_get_unless_zero(&ctrl->ctrl.kref))
1715                 return -EBUSY;
1716         ret = __nvme_rdma_del_ctrl(ctrl);
1717         if (!ret)
1718                 flush_work(&ctrl->delete_work);
1719         nvme_put_ctrl(&ctrl->ctrl);
1720         return ret;
1721 }
1722
1723 static void nvme_rdma_remove_ctrl_work(struct work_struct *work)
1724 {
1725         struct nvme_rdma_ctrl *ctrl = container_of(work,
1726                                 struct nvme_rdma_ctrl, delete_work);
1727
1728         __nvme_rdma_remove_ctrl(ctrl, false);
1729 }
1730
1731 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1732 {
1733         struct nvme_rdma_ctrl *ctrl = container_of(work,
1734                                         struct nvme_rdma_ctrl, reset_work);
1735         int ret;
1736         bool changed;
1737
1738         nvme_rdma_shutdown_ctrl(ctrl);
1739
1740         ret = nvme_rdma_configure_admin_queue(ctrl);
1741         if (ret) {
1742                 /* ctrl is already shutdown, just remove the ctrl */
1743                 INIT_WORK(&ctrl->delete_work, nvme_rdma_remove_ctrl_work);
1744                 goto del_dead_ctrl;
1745         }
1746
1747         if (ctrl->queue_count > 1) {
1748                 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
1749                 if (ret)
1750                         goto del_dead_ctrl;
1751
1752                 ret = nvme_rdma_init_io_queues(ctrl);
1753                 if (ret)
1754                         goto del_dead_ctrl;
1755
1756                 ret = nvme_rdma_connect_io_queues(ctrl);
1757                 if (ret)
1758                         goto del_dead_ctrl;
1759         }
1760
1761         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1762         WARN_ON_ONCE(!changed);
1763
1764         if (ctrl->queue_count > 1) {
1765                 nvme_start_queues(&ctrl->ctrl);
1766                 nvme_queue_scan(&ctrl->ctrl);
1767                 nvme_queue_async_events(&ctrl->ctrl);
1768         }
1769
1770         return;
1771
1772 del_dead_ctrl:
1773         /* Deleting this dead controller... */
1774         dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1775         WARN_ON(!queue_work(nvme_rdma_wq, &ctrl->delete_work));
1776 }
1777
1778 static int nvme_rdma_reset_ctrl(struct nvme_ctrl *nctrl)
1779 {
1780         struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1781
1782         if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1783                 return -EBUSY;
1784
1785         if (!queue_work(nvme_rdma_wq, &ctrl->reset_work))
1786                 return -EBUSY;
1787
1788         flush_work(&ctrl->reset_work);
1789
1790         return 0;
1791 }
1792
1793 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1794         .name                   = "rdma",
1795         .module                 = THIS_MODULE,
1796         .is_fabrics             = true,
1797         .reg_read32             = nvmf_reg_read32,
1798         .reg_read64             = nvmf_reg_read64,
1799         .reg_write32            = nvmf_reg_write32,
1800         .reset_ctrl             = nvme_rdma_reset_ctrl,
1801         .free_ctrl              = nvme_rdma_free_ctrl,
1802         .submit_async_event     = nvme_rdma_submit_async_event,
1803         .delete_ctrl            = nvme_rdma_del_ctrl,
1804         .get_subsysnqn          = nvmf_get_subsysnqn,
1805         .get_address            = nvmf_get_address,
1806 };
1807
1808 static int nvme_rdma_create_io_queues(struct nvme_rdma_ctrl *ctrl)
1809 {
1810         int ret;
1811
1812         ret = nvme_rdma_init_io_queues(ctrl);
1813         if (ret)
1814                 return ret;
1815
1816         /*
1817          * We need a reference on the device as long as the tag_set is alive,
1818          * as the MRs in the request structures need a valid ib_device.
1819          */
1820         ret = -EINVAL;
1821         if (!nvme_rdma_dev_get(ctrl->device))
1822                 goto out_free_io_queues;
1823
1824         memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
1825         ctrl->tag_set.ops = &nvme_rdma_mq_ops;
1826         ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
1827         ctrl->tag_set.reserved_tags = 1; /* fabric connect */
1828         ctrl->tag_set.numa_node = NUMA_NO_NODE;
1829         ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1830         ctrl->tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1831                 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1832         ctrl->tag_set.driver_data = ctrl;
1833         ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1;
1834         ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
1835
1836         ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
1837         if (ret)
1838                 goto out_put_dev;
1839         ctrl->ctrl.tagset = &ctrl->tag_set;
1840
1841         ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
1842         if (IS_ERR(ctrl->ctrl.connect_q)) {
1843                 ret = PTR_ERR(ctrl->ctrl.connect_q);
1844                 goto out_free_tag_set;
1845         }
1846
1847         ret = nvme_rdma_connect_io_queues(ctrl);
1848         if (ret)
1849                 goto out_cleanup_connect_q;
1850
1851         return 0;
1852
1853 out_cleanup_connect_q:
1854         blk_cleanup_queue(ctrl->ctrl.connect_q);
1855 out_free_tag_set:
1856         blk_mq_free_tag_set(&ctrl->tag_set);
1857 out_put_dev:
1858         nvme_rdma_dev_put(ctrl->device);
1859 out_free_io_queues:
1860         nvme_rdma_free_io_queues(ctrl);
1861         return ret;
1862 }
1863
1864 static int nvme_rdma_parse_ipaddr(struct sockaddr_in *in_addr, char *p)
1865 {
1866         u8 *addr = (u8 *)&in_addr->sin_addr.s_addr;
1867         size_t buflen = strlen(p);
1868
1869         /* XXX: handle IPv6 addresses */
1870
1871         if (buflen > INET_ADDRSTRLEN)
1872                 return -EINVAL;
1873         if (in4_pton(p, buflen, addr, '\0', NULL) == 0)
1874                 return -EINVAL;
1875         in_addr->sin_family = AF_INET;
1876         return 0;
1877 }
1878
1879 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1880                 struct nvmf_ctrl_options *opts)
1881 {
1882         struct nvme_rdma_ctrl *ctrl;
1883         int ret;
1884         bool changed;
1885
1886         ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1887         if (!ctrl)
1888                 return ERR_PTR(-ENOMEM);
1889         ctrl->ctrl.opts = opts;
1890         INIT_LIST_HEAD(&ctrl->list);
1891
1892         ret = nvme_rdma_parse_ipaddr(&ctrl->addr_in, opts->traddr);
1893         if (ret) {
1894                 pr_err("malformed IP address passed: %s\n", opts->traddr);
1895                 goto out_free_ctrl;
1896         }
1897
1898         if (opts->mask & NVMF_OPT_TRSVCID) {
1899                 u16 port;
1900
1901                 ret = kstrtou16(opts->trsvcid, 0, &port);
1902                 if (ret)
1903                         goto out_free_ctrl;
1904
1905                 ctrl->addr_in.sin_port = cpu_to_be16(port);
1906         } else {
1907                 ctrl->addr_in.sin_port = cpu_to_be16(NVME_RDMA_IP_PORT);
1908         }
1909
1910         ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1911                                 0 /* no quirks, we're perfect! */);
1912         if (ret)
1913                 goto out_free_ctrl;
1914
1915         ctrl->reconnect_delay = opts->reconnect_delay;
1916         INIT_DELAYED_WORK(&ctrl->reconnect_work,
1917                         nvme_rdma_reconnect_ctrl_work);
1918         INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1919         INIT_WORK(&ctrl->delete_work, nvme_rdma_del_ctrl_work);
1920         INIT_WORK(&ctrl->reset_work, nvme_rdma_reset_ctrl_work);
1921         spin_lock_init(&ctrl->lock);
1922
1923         ctrl->queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1924         ctrl->ctrl.sqsize = opts->queue_size - 1;
1925         ctrl->ctrl.kato = opts->kato;
1926
1927         ret = -ENOMEM;
1928         ctrl->queues = kcalloc(ctrl->queue_count, sizeof(*ctrl->queues),
1929                                 GFP_KERNEL);
1930         if (!ctrl->queues)
1931                 goto out_uninit_ctrl;
1932
1933         ret = nvme_rdma_configure_admin_queue(ctrl);
1934         if (ret)
1935                 goto out_kfree_queues;
1936
1937         /* sanity check icdoff */
1938         if (ctrl->ctrl.icdoff) {
1939                 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1940                 goto out_remove_admin_queue;
1941         }
1942
1943         /* sanity check keyed sgls */
1944         if (!(ctrl->ctrl.sgls & (1 << 20))) {
1945                 dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1946                 goto out_remove_admin_queue;
1947         }
1948
1949         if (opts->queue_size > ctrl->ctrl.maxcmd) {
1950                 /* warn if maxcmd is lower than queue_size */
1951                 dev_warn(ctrl->ctrl.device,
1952                         "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1953                         opts->queue_size, ctrl->ctrl.maxcmd);
1954                 opts->queue_size = ctrl->ctrl.maxcmd;
1955         }
1956
1957         if (opts->nr_io_queues) {
1958                 ret = nvme_rdma_create_io_queues(ctrl);
1959                 if (ret)
1960                         goto out_remove_admin_queue;
1961         }
1962
1963         changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1964         WARN_ON_ONCE(!changed);
1965
1966         dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
1967                 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1968
1969         kref_get(&ctrl->ctrl.kref);
1970
1971         mutex_lock(&nvme_rdma_ctrl_mutex);
1972         list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1973         mutex_unlock(&nvme_rdma_ctrl_mutex);
1974
1975         if (opts->nr_io_queues) {
1976                 nvme_queue_scan(&ctrl->ctrl);
1977                 nvme_queue_async_events(&ctrl->ctrl);
1978         }
1979
1980         return &ctrl->ctrl;
1981
1982 out_remove_admin_queue:
1983         nvme_stop_keep_alive(&ctrl->ctrl);
1984         nvme_rdma_destroy_admin_queue(ctrl);
1985 out_kfree_queues:
1986         kfree(ctrl->queues);
1987 out_uninit_ctrl:
1988         nvme_uninit_ctrl(&ctrl->ctrl);
1989         nvme_put_ctrl(&ctrl->ctrl);
1990         if (ret > 0)
1991                 ret = -EIO;
1992         return ERR_PTR(ret);
1993 out_free_ctrl:
1994         kfree(ctrl);
1995         return ERR_PTR(ret);
1996 }
1997
1998 static struct nvmf_transport_ops nvme_rdma_transport = {
1999         .name           = "rdma",
2000         .required_opts  = NVMF_OPT_TRADDR,
2001         .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY,
2002         .create_ctrl    = nvme_rdma_create_ctrl,
2003 };
2004
2005 static void nvme_rdma_add_one(struct ib_device *ib_device)
2006 {
2007 }
2008
2009 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2010 {
2011         struct nvme_rdma_ctrl *ctrl;
2012
2013         /* Delete all controllers using this device */
2014         mutex_lock(&nvme_rdma_ctrl_mutex);
2015         list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2016                 if (ctrl->device->dev != ib_device)
2017                         continue;
2018                 dev_info(ctrl->ctrl.device,
2019                         "Removing ctrl: NQN \"%s\", addr %pISp\n",
2020                         ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2021                 __nvme_rdma_del_ctrl(ctrl);
2022         }
2023         mutex_unlock(&nvme_rdma_ctrl_mutex);
2024
2025         flush_workqueue(nvme_rdma_wq);
2026 }
2027
2028 static struct ib_client nvme_rdma_ib_client = {
2029         .name   = "nvme_rdma",
2030         .add = nvme_rdma_add_one,
2031         .remove = nvme_rdma_remove_one
2032 };
2033
2034 static int __init nvme_rdma_init_module(void)
2035 {
2036         int ret;
2037
2038         nvme_rdma_wq = create_workqueue("nvme_rdma_wq");
2039         if (!nvme_rdma_wq)
2040                 return -ENOMEM;
2041
2042         ret = ib_register_client(&nvme_rdma_ib_client);
2043         if (ret) {
2044                 destroy_workqueue(nvme_rdma_wq);
2045                 return ret;
2046         }
2047
2048         nvmf_register_transport(&nvme_rdma_transport);
2049         return 0;
2050 }
2051
2052 static void __exit nvme_rdma_cleanup_module(void)
2053 {
2054         nvmf_unregister_transport(&nvme_rdma_transport);
2055         ib_unregister_client(&nvme_rdma_ib_client);
2056         destroy_workqueue(nvme_rdma_wq);
2057 }
2058
2059 module_init(nvme_rdma_init_module);
2060 module_exit(nvme_rdma_cleanup_module);
2061
2062 MODULE_LICENSE("GPL v2");