drivers/nvme/target/rdma.c

   1 /*
   2  * NVMe over Fabrics RDMA target.
   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/atomic.h>
  16 #include <linux/ctype.h>
  17 #include <linux/delay.h>
  18 #include <linux/err.h>
  19 #include <linux/init.h>
  20 #include <linux/module.h>
  21 #include <linux/nvme.h>
  22 #include <linux/slab.h>
  23 #include <linux/string.h>
  24 #include <linux/wait.h>
  25 #include <linux/inet.h>
  26 #include <asm/unaligned.h>
  27
  28 #include <rdma/ib_verbs.h>
  29 #include <rdma/rdma_cm.h>
  30 #include <rdma/rw.h>
  31
  32 #include <linux/nvme-rdma.h>
  33 #include "nvmet.h"
  34
  35 /*
  36  * We allow up to a page of inline data to go with the SQE
  37  */
  38 #define NVMET_RDMA_INLINE_DATA_SIZE     PAGE_SIZE
  39
  40 struct nvmet_rdma_cmd {
  41         struct ib_sge           sge[2];
  42         struct ib_cqe           cqe;
  43         struct ib_recv_wr       wr;
  44         struct scatterlist      inline_sg;
  45         struct page             *inline_page;
  46         struct nvme_command     *nvme_cmd;
  47         struct nvmet_rdma_queue *queue;
  48 };
  49
  50 enum {
  51         NVMET_RDMA_REQ_INLINE_DATA      = (1 << 0),
  52         NVMET_RDMA_REQ_INVALIDATE_RKEY  = (1 << 1),
  53 };
  54
  55 struct nvmet_rdma_rsp {
  56         struct ib_sge           send_sge;
  57         struct ib_cqe           send_cqe;
  58         struct ib_send_wr       send_wr;
  59
  60         struct nvmet_rdma_cmd   *cmd;
  61         struct nvmet_rdma_queue *queue;
  62
  63         struct ib_cqe           read_cqe;
  64         struct rdma_rw_ctx      rw;
  65
  66         struct nvmet_req        req;
  67
  68         u8                      n_rdma;
  69         u32                     flags;
  70         u32                     invalidate_rkey;
  71
  72         struct list_head        wait_list;
  73         struct list_head        free_list;
  74 };
  75
  76 enum nvmet_rdma_queue_state {
  77         NVMET_RDMA_Q_CONNECTING,
  78         NVMET_RDMA_Q_LIVE,
  79         NVMET_RDMA_Q_DISCONNECTING,
  80         NVMET_RDMA_IN_DEVICE_REMOVAL,
  81 };
  82
  83 struct nvmet_rdma_queue {
  84         struct rdma_cm_id       *cm_id;
  85         struct nvmet_port       *port;
  86         struct ib_cq            *cq;
  87         atomic_t                sq_wr_avail;
  88         struct nvmet_rdma_device *dev;
  89         spinlock_t              state_lock;
  90         enum nvmet_rdma_queue_state state;
  91         struct nvmet_cq         nvme_cq;
  92         struct nvmet_sq         nvme_sq;
  93
  94         struct nvmet_rdma_rsp   *rsps;
  95         struct list_head        free_rsps;
  96         spinlock_t              rsps_lock;
  97         struct nvmet_rdma_cmd   *cmds;
  98
  99         struct work_struct      release_work;
 100         struct list_head        rsp_wait_list;
 101         struct list_head        rsp_wr_wait_list;
 102         spinlock_t              rsp_wr_wait_lock;
 103
 104         int                     idx;
 105         int                     host_qid;
 106         int                     recv_queue_size;
 107         int                     send_queue_size;
 108
 109         struct list_head        queue_list;
 110 };
 111
 112 struct nvmet_rdma_device {
 113         struct ib_device        *device;
 114         struct ib_pd            *pd;
 115         struct ib_srq           *srq;
 116         struct nvmet_rdma_cmd   *srq_cmds;
 117         size_t                  srq_size;
 118         struct kref             ref;
 119         struct list_head        entry;
 120 };
 121
 122 static bool nvmet_rdma_use_srq;
 123 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
 124 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
 125
 126 static DEFINE_IDA(nvmet_rdma_queue_ida);
 127 static LIST_HEAD(nvmet_rdma_queue_list);
 128 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
 129
 130 static LIST_HEAD(device_list);
 131 static DEFINE_MUTEX(device_list_mutex);
 132
 133 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
 134 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
 135 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 136 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
 137 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
 138 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
 139
 140 static struct nvmet_fabrics_ops nvmet_rdma_ops;
 141
 142 /* XXX: really should move to a generic header sooner or later.. */
 143 static inline u32 get_unaligned_le24(const u8 *p)
 144 {
 145         return (u32)p[0] | (u32)p[1] << 8 | (u32)p[2] << 16;
 146 }
 147
 148 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
 149 {
 150         return nvme_is_write(rsp->req.cmd) &&
 151                 rsp->req.data_len &&
 152                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 153 }
 154
 155 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
 156 {
 157         return !nvme_is_write(rsp->req.cmd) &&
 158                 rsp->req.data_len &&
 159                 !rsp->req.rsp->status &&
 160                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 161 }
 162
 163 static inline struct nvmet_rdma_rsp *
 164 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
 165 {
 166         struct nvmet_rdma_rsp *rsp;
 167         unsigned long flags;
 168
 169         spin_lock_irqsave(&queue->rsps_lock, flags);
 170         rsp = list_first_entry(&queue->free_rsps,
 171                                 struct nvmet_rdma_rsp, free_list);
 172         list_del(&rsp->free_list);
 173         spin_unlock_irqrestore(&queue->rsps_lock, flags);
 174
 175         return rsp;
 176 }
 177
 178 static inline void
 179 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
 180 {
 181         unsigned long flags;
 182
 183         spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
 184         list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
 185         spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
 186 }
 187
 188 static void nvmet_rdma_free_sgl(struct scatterlist *sgl, unsigned int nents)
 189 {
 190         struct scatterlist *sg;
 191         int count;
 192
 193         if (!sgl || !nents)
 194                 return;
 195
 196         for_each_sg(sgl, sg, nents, count)
 197                 __free_page(sg_page(sg));
 198         kfree(sgl);
 199 }
 200
 201 static int nvmet_rdma_alloc_sgl(struct scatterlist **sgl, unsigned int *nents,
 202                 u32 length)
 203 {
 204         struct scatterlist *sg;
 205         struct page *page;
 206         unsigned int nent;
 207         int i = 0;
 208
 209         nent = DIV_ROUND_UP(length, PAGE_SIZE);
 210         sg = kmalloc_array(nent, sizeof(struct scatterlist), GFP_KERNEL);
 211         if (!sg)
 212                 goto out;
 213
 214         sg_init_table(sg, nent);
 215
 216         while (length) {
 217                 u32 page_len = min_t(u32, length, PAGE_SIZE);
 218
 219                 page = alloc_page(GFP_KERNEL);
 220                 if (!page)
 221                         goto out_free_pages;
 222
 223                 sg_set_page(&sg[i], page, page_len, 0);
 224                 length -= page_len;
 225                 i++;
 226         }
 227         *sgl = sg;
 228         *nents = nent;
 229         return 0;
 230
 231 out_free_pages:
 232         while (i > 0) {
 233                 i--;
 234                 __free_page(sg_page(&sg[i]));
 235         }
 236         kfree(sg);
 237 out:
 238         return NVME_SC_INTERNAL;
 239 }
 240
 241 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
 242                         struct nvmet_rdma_cmd *c, bool admin)
 243 {
 244         /* NVMe command / RDMA RECV */
 245         c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
 246         if (!c->nvme_cmd)
 247                 goto out;
 248
 249         c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
 250                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 251         if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
 252                 goto out_free_cmd;
 253
 254         c->sge[0].length = sizeof(*c->nvme_cmd);
 255         c->sge[0].lkey = ndev->pd->local_dma_lkey;
 256
 257         if (!admin) {
 258                 c->inline_page = alloc_pages(GFP_KERNEL,
 259                                 get_order(NVMET_RDMA_INLINE_DATA_SIZE));
 260                 if (!c->inline_page)
 261                         goto out_unmap_cmd;
 262                 c->sge[1].addr = ib_dma_map_page(ndev->device,
 263                                 c->inline_page, 0, NVMET_RDMA_INLINE_DATA_SIZE,
 264                                 DMA_FROM_DEVICE);
 265                 if (ib_dma_mapping_error(ndev->device, c->sge[1].addr))
 266                         goto out_free_inline_page;
 267                 c->sge[1].length = NVMET_RDMA_INLINE_DATA_SIZE;
 268                 c->sge[1].lkey = ndev->pd->local_dma_lkey;
 269         }
 270
 271         c->cqe.done = nvmet_rdma_recv_done;
 272
 273         c->wr.wr_cqe = &c->cqe;
 274         c->wr.sg_list = c->sge;
 275         c->wr.num_sge = admin ? 1 : 2;
 276
 277         return 0;
 278
 279 out_free_inline_page:
 280         if (!admin) {
 281                 __free_pages(c->inline_page,
 282                                 get_order(NVMET_RDMA_INLINE_DATA_SIZE));
 283         }
 284 out_unmap_cmd:
 285         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 286                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 287 out_free_cmd:
 288         kfree(c->nvme_cmd);
 289
 290 out:
 291         return -ENOMEM;
 292 }
 293
 294 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
 295                 struct nvmet_rdma_cmd *c, bool admin)
 296 {
 297         if (!admin) {
 298                 ib_dma_unmap_page(ndev->device, c->sge[1].addr,
 299                                 NVMET_RDMA_INLINE_DATA_SIZE, DMA_FROM_DEVICE);
 300                 __free_pages(c->inline_page,
 301                                 get_order(NVMET_RDMA_INLINE_DATA_SIZE));
 302         }
 303         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 304                                 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 305         kfree(c->nvme_cmd);
 306 }
 307
 308 static struct nvmet_rdma_cmd *
 309 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
 310                 int nr_cmds, bool admin)
 311 {
 312         struct nvmet_rdma_cmd *cmds;
 313         int ret = -EINVAL, i;
 314
 315         cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
 316         if (!cmds)
 317                 goto out;
 318
 319         for (i = 0; i < nr_cmds; i++) {
 320                 ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
 321                 if (ret)
 322                         goto out_free;
 323         }
 324
 325         return cmds;
 326
 327 out_free:
 328         while (--i >= 0)
 329                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 330         kfree(cmds);
 331 out:
 332         return ERR_PTR(ret);
 333 }
 334
 335 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
 336                 struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
 337 {
 338         int i;
 339
 340         for (i = 0; i < nr_cmds; i++)
 341                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 342         kfree(cmds);
 343 }
 344
 345 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
 346                 struct nvmet_rdma_rsp *r)
 347 {
 348         /* NVMe CQE / RDMA SEND */
 349         r->req.rsp = kmalloc(sizeof(*r->req.rsp), GFP_KERNEL);
 350         if (!r->req.rsp)
 351                 goto out;
 352
 353         r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.rsp,
 354                         sizeof(*r->req.rsp), DMA_TO_DEVICE);
 355         if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
 356                 goto out_free_rsp;
 357
 358         r->send_sge.length = sizeof(*r->req.rsp);
 359         r->send_sge.lkey = ndev->pd->local_dma_lkey;
 360
 361         r->send_cqe.done = nvmet_rdma_send_done;
 362
 363         r->send_wr.wr_cqe = &r->send_cqe;
 364         r->send_wr.sg_list = &r->send_sge;
 365         r->send_wr.num_sge = 1;
 366         r->send_wr.send_flags = IB_SEND_SIGNALED;
 367
 368         /* Data In / RDMA READ */
 369         r->read_cqe.done = nvmet_rdma_read_data_done;
 370         return 0;
 371
 372 out_free_rsp:
 373         kfree(r->req.rsp);
 374 out:
 375         return -ENOMEM;
 376 }
 377
 378 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
 379                 struct nvmet_rdma_rsp *r)
 380 {
 381         ib_dma_unmap_single(ndev->device, r->send_sge.addr,
 382                                 sizeof(*r->req.rsp), DMA_TO_DEVICE);
 383         kfree(r->req.rsp);
 384 }
 385
 386 static int
 387 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
 388 {
 389         struct nvmet_rdma_device *ndev = queue->dev;
 390         int nr_rsps = queue->recv_queue_size * 2;
 391         int ret = -EINVAL, i;
 392
 393         queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
 394                         GFP_KERNEL);
 395         if (!queue->rsps)
 396                 goto out;
 397
 398         for (i = 0; i < nr_rsps; i++) {
 399                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 400
 401                 ret = nvmet_rdma_alloc_rsp(ndev, rsp);
 402                 if (ret)
 403                         goto out_free;
 404
 405                 list_add_tail(&rsp->free_list, &queue->free_rsps);
 406         }
 407
 408         return 0;
 409
 410 out_free:
 411         while (--i >= 0) {
 412                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 413
 414                 list_del(&rsp->free_list);
 415                 nvmet_rdma_free_rsp(ndev, rsp);
 416         }
 417         kfree(queue->rsps);
 418 out:
 419         return ret;
 420 }
 421
 422 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
 423 {
 424         struct nvmet_rdma_device *ndev = queue->dev;
 425         int i, nr_rsps = queue->recv_queue_size * 2;
 426
 427         for (i = 0; i < nr_rsps; i++) {
 428                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 429
 430                 list_del(&rsp->free_list);
 431                 nvmet_rdma_free_rsp(ndev, rsp);
 432         }
 433         kfree(queue->rsps);
 434 }
 435
 436 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
 437                 struct nvmet_rdma_cmd *cmd)
 438 {
 439         struct ib_recv_wr *bad_wr;
 440
 441         if (ndev->srq)
 442                 return ib_post_srq_recv(ndev->srq, &cmd->wr, &bad_wr);
 443         return ib_post_recv(cmd->queue->cm_id->qp, &cmd->wr, &bad_wr);
 444 }
 445
 446 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
 447 {
 448         spin_lock(&queue->rsp_wr_wait_lock);
 449         while (!list_empty(&queue->rsp_wr_wait_list)) {
 450                 struct nvmet_rdma_rsp *rsp;
 451                 bool ret;
 452
 453                 rsp = list_entry(queue->rsp_wr_wait_list.next,
 454                                 struct nvmet_rdma_rsp, wait_list);
 455                 list_del(&rsp->wait_list);
 456
 457                 spin_unlock(&queue->rsp_wr_wait_lock);
 458                 ret = nvmet_rdma_execute_command(rsp);
 459                 spin_lock(&queue->rsp_wr_wait_lock);
 460
 461                 if (!ret) {
 462                         list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
 463                         break;
 464                 }
 465         }
 466         spin_unlock(&queue->rsp_wr_wait_lock);
 467 }
 468
 469
 470 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
 471 {
 472         struct nvmet_rdma_queue *queue = rsp->queue;
 473
 474         atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 475
 476         if (rsp->n_rdma) {
 477                 rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
 478                                 queue->cm_id->port_num, rsp->req.sg,
 479                                 rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
 480         }
 481
 482         if (rsp->req.sg != &rsp->cmd->inline_sg)
 483                 nvmet_rdma_free_sgl(rsp->req.sg, rsp->req.sg_cnt);
 484
 485         if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
 486                 nvmet_rdma_process_wr_wait_list(queue);
 487
 488         nvmet_rdma_put_rsp(rsp);
 489 }
 490
 491 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
 492 {
 493         if (queue->nvme_sq.ctrl) {
 494                 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
 495         } else {
 496                 /*
 497                  * we didn't setup the controller yet in case
 498                  * of admin connect error, just disconnect and
 499                  * cleanup the queue
 500                  */
 501                 nvmet_rdma_queue_disconnect(queue);
 502         }
 503 }
 504
 505 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
 506 {
 507         struct nvmet_rdma_rsp *rsp =
 508                 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
 509
 510         nvmet_rdma_release_rsp(rsp);
 511
 512         if (unlikely(wc->status != IB_WC_SUCCESS &&
 513                      wc->status != IB_WC_WR_FLUSH_ERR)) {
 514                 pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
 515                         wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
 516                 nvmet_rdma_error_comp(rsp->queue);
 517         }
 518 }
 519
 520 static void nvmet_rdma_queue_response(struct nvmet_req *req)
 521 {
 522         struct nvmet_rdma_rsp *rsp =
 523                 container_of(req, struct nvmet_rdma_rsp, req);
 524         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 525         struct ib_send_wr *first_wr, *bad_wr;
 526
 527         if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
 528                 rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
 529                 rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
 530         } else {
 531                 rsp->send_wr.opcode = IB_WR_SEND;
 532         }
 533
 534         if (nvmet_rdma_need_data_out(rsp))
 535                 first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
 536                                 cm_id->port_num, NULL, &rsp->send_wr);
 537         else
 538                 first_wr = &rsp->send_wr;
 539
 540         nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
 541         if (ib_post_send(cm_id->qp, first_wr, &bad_wr)) {
 542                 pr_err("sending cmd response failed\n");
 543                 nvmet_rdma_release_rsp(rsp);
 544         }
 545 }
 546
 547 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
 548 {
 549         struct nvmet_rdma_rsp *rsp =
 550                 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
 551         struct nvmet_rdma_queue *queue = cq->cq_context;
 552
 553         WARN_ON(rsp->n_rdma <= 0);
 554         atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
 555         rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
 556                         queue->cm_id->port_num, rsp->req.sg,
 557                         rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
 558         rsp->n_rdma = 0;
 559
 560         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 561                 nvmet_rdma_release_rsp(rsp);
 562                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 563                         pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
 564                                 wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
 565                         nvmet_rdma_error_comp(queue);
 566                 }
 567                 return;
 568         }
 569
 570         rsp->req.execute(&rsp->req);
 571 }
 572
 573 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
 574                 u64 off)
 575 {
 576         sg_init_table(&rsp->cmd->inline_sg, 1);
 577         sg_set_page(&rsp->cmd->inline_sg, rsp->cmd->inline_page, len, off);
 578         rsp->req.sg = &rsp->cmd->inline_sg;
 579         rsp->req.sg_cnt = 1;
 580 }
 581
 582 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
 583 {
 584         struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
 585         u64 off = le64_to_cpu(sgl->addr);
 586         u32 len = le32_to_cpu(sgl->length);
 587
 588         if (!nvme_is_write(rsp->req.cmd))
 589                 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 590
 591         if (off + len > NVMET_RDMA_INLINE_DATA_SIZE) {
 592                 pr_err("invalid inline data offset!\n");
 593                 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
 594         }
 595
 596         /* no data command? */
 597         if (!len)
 598                 return 0;
 599
 600         nvmet_rdma_use_inline_sg(rsp, len, off);
 601         rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
 602         return 0;
 603 }
 604
 605 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
 606                 struct nvme_keyed_sgl_desc *sgl, bool invalidate)
 607 {
 608         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 609         u64 addr = le64_to_cpu(sgl->addr);
 610         u32 len = get_unaligned_le24(sgl->length);
 611         u32 key = get_unaligned_le32(sgl->key);
 612         int ret;
 613         u16 status;
 614
 615         /* no data command? */
 616         if (!len)
 617                 return 0;
 618
 619         status = nvmet_rdma_alloc_sgl(&rsp->req.sg, &rsp->req.sg_cnt,
 620                         len);
 621         if (status)
 622                 return status;
 623
 624         ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
 625                         rsp->req.sg, rsp->req.sg_cnt, 0, addr, key,
 626                         nvmet_data_dir(&rsp->req));
 627         if (ret < 0)
 628                 return NVME_SC_INTERNAL;
 629         rsp->n_rdma += ret;
 630
 631         if (invalidate) {
 632                 rsp->invalidate_rkey = key;
 633                 rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
 634         }
 635
 636         return 0;
 637 }
 638
 639 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
 640 {
 641         struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
 642
 643         switch (sgl->type >> 4) {
 644         case NVME_SGL_FMT_DATA_DESC:
 645                 switch (sgl->type & 0xf) {
 646                 case NVME_SGL_FMT_OFFSET:
 647                         return nvmet_rdma_map_sgl_inline(rsp);
 648                 default:
 649                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 650                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 651                 }
 652         case NVME_KEY_SGL_FMT_DATA_DESC:
 653                 switch (sgl->type & 0xf) {
 654                 case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
 655                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
 656                 case NVME_SGL_FMT_ADDRESS:
 657                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
 658                 default:
 659                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 660                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 661                 }
 662         default:
 663                 pr_err("invalid SGL type: %#x\n", sgl->type);
 664                 return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
 665         }
 666 }
 667
 668 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
 669 {
 670         struct nvmet_rdma_queue *queue = rsp->queue;
 671
 672         if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
 673                         &queue->sq_wr_avail) < 0)) {
 674                 pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
 675                                 1 + rsp->n_rdma, queue->idx,
 676                                 queue->nvme_sq.ctrl->cntlid);
 677                 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 678                 return false;
 679         }
 680
 681         if (nvmet_rdma_need_data_in(rsp)) {
 682                 if (rdma_rw_ctx_post(&rsp->rw, queue->cm_id->qp,
 683                                 queue->cm_id->port_num, &rsp->read_cqe, NULL))
 684                         nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
 685         } else {
 686                 rsp->req.execute(&rsp->req);
 687         }
 688
 689         return true;
 690 }
 691
 692 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
 693                 struct nvmet_rdma_rsp *cmd)
 694 {
 695         u16 status;
 696
 697         cmd->queue = queue;
 698         cmd->n_rdma = 0;
 699         cmd->req.port = queue->port;
 700
 701         if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
 702                         &queue->nvme_sq, &nvmet_rdma_ops))
 703                 return;
 704
 705         status = nvmet_rdma_map_sgl(cmd);
 706         if (status)
 707                 goto out_err;
 708
 709         if (unlikely(!nvmet_rdma_execute_command(cmd))) {
 710                 spin_lock(&queue->rsp_wr_wait_lock);
 711                 list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
 712                 spin_unlock(&queue->rsp_wr_wait_lock);
 713         }
 714
 715         return;
 716
 717 out_err:
 718         nvmet_req_complete(&cmd->req, status);
 719 }
 720
 721 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
 722 {
 723         struct nvmet_rdma_cmd *cmd =
 724                 container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
 725         struct nvmet_rdma_queue *queue = cq->cq_context;
 726         struct nvmet_rdma_rsp *rsp;
 727
 728         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 729                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 730                         pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
 731                                 wc->wr_cqe, ib_wc_status_msg(wc->status),
 732                                 wc->status);
 733                         nvmet_rdma_error_comp(queue);
 734                 }
 735                 return;
 736         }
 737
 738         if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
 739                 pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
 740                 nvmet_rdma_error_comp(queue);
 741                 return;
 742         }
 743
 744         cmd->queue = queue;
 745         rsp = nvmet_rdma_get_rsp(queue);
 746         rsp->cmd = cmd;
 747         rsp->flags = 0;
 748         rsp->req.cmd = cmd->nvme_cmd;
 749
 750         if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
 751                 unsigned long flags;
 752
 753                 spin_lock_irqsave(&queue->state_lock, flags);
 754                 if (queue->state == NVMET_RDMA_Q_CONNECTING)
 755                         list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
 756                 else
 757                         nvmet_rdma_put_rsp(rsp);
 758                 spin_unlock_irqrestore(&queue->state_lock, flags);
 759                 return;
 760         }
 761
 762         nvmet_rdma_handle_command(queue, rsp);
 763 }
 764
 765 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_device *ndev)
 766 {
 767         if (!ndev->srq)
 768                 return;
 769
 770         nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
 771         ib_destroy_srq(ndev->srq);
 772 }
 773
 774 static int nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
 775 {
 776         struct ib_srq_init_attr srq_attr = { NULL, };
 777         struct ib_srq *srq;
 778         size_t srq_size;
 779         int ret, i;
 780
 781         srq_size = 4095;        /* XXX: tune */
 782
 783         srq_attr.attr.max_wr = srq_size;
 784         srq_attr.attr.max_sge = 2;
 785         srq_attr.attr.srq_limit = 0;
 786         srq_attr.srq_type = IB_SRQT_BASIC;
 787         srq = ib_create_srq(ndev->pd, &srq_attr);
 788         if (IS_ERR(srq)) {
 789                 /*
 790                  * If SRQs aren't supported we just go ahead and use normal
 791                  * non-shared receive queues.
 792                  */
 793                 pr_info("SRQ requested but not supported.\n");
 794                 return 0;
 795         }
 796
 797         ndev->srq_cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
 798         if (IS_ERR(ndev->srq_cmds)) {
 799                 ret = PTR_ERR(ndev->srq_cmds);
 800                 goto out_destroy_srq;
 801         }
 802
 803         ndev->srq = srq;
 804         ndev->srq_size = srq_size;
 805
 806         for (i = 0; i < srq_size; i++)
 807                 nvmet_rdma_post_recv(ndev, &ndev->srq_cmds[i]);
 808
 809         return 0;
 810
 811 out_destroy_srq:
 812         ib_destroy_srq(srq);
 813         return ret;
 814 }
 815
 816 static void nvmet_rdma_free_dev(struct kref *ref)
 817 {
 818         struct nvmet_rdma_device *ndev =
 819                 container_of(ref, struct nvmet_rdma_device, ref);
 820
 821         mutex_lock(&device_list_mutex);
 822         list_del(&ndev->entry);
 823         mutex_unlock(&device_list_mutex);
 824
 825         nvmet_rdma_destroy_srq(ndev);
 826         ib_dealloc_pd(ndev->pd);
 827
 828         kfree(ndev);
 829 }
 830
 831 static struct nvmet_rdma_device *
 832 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
 833 {
 834         struct nvmet_rdma_device *ndev;
 835         int ret;
 836
 837         mutex_lock(&device_list_mutex);
 838         list_for_each_entry(ndev, &device_list, entry) {
 839                 if (ndev->device->node_guid == cm_id->device->node_guid &&
 840                     kref_get_unless_zero(&ndev->ref))
 841                         goto out_unlock;
 842         }
 843
 844         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 845         if (!ndev)
 846                 goto out_err;
 847
 848         ndev->device = cm_id->device;
 849         kref_init(&ndev->ref);
 850
 851         ndev->pd = ib_alloc_pd(ndev->device, 0);
 852         if (IS_ERR(ndev->pd))
 853                 goto out_free_dev;
 854
 855         if (nvmet_rdma_use_srq) {
 856                 ret = nvmet_rdma_init_srq(ndev);
 857                 if (ret)
 858                         goto out_free_pd;
 859         }
 860
 861         list_add(&ndev->entry, &device_list);
 862 out_unlock:
 863         mutex_unlock(&device_list_mutex);
 864         pr_debug("added %s.\n", ndev->device->name);
 865         return ndev;
 866
 867 out_free_pd:
 868         ib_dealloc_pd(ndev->pd);
 869 out_free_dev:
 870         kfree(ndev);
 871 out_err:
 872         mutex_unlock(&device_list_mutex);
 873         return NULL;
 874 }
 875
 876 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
 877 {
 878         struct ib_qp_init_attr qp_attr;
 879         struct nvmet_rdma_device *ndev = queue->dev;
 880         int comp_vector, nr_cqe, ret, i;
 881
 882         /*
 883          * Spread the io queues across completion vectors,
 884          * but still keep all admin queues on vector 0.
 885          */
 886         comp_vector = !queue->host_qid ? 0 :
 887                 queue->idx % ndev->device->num_comp_vectors;
 888
 889         /*
 890          * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
 891          */
 892         nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
 893
 894         queue->cq = ib_alloc_cq(ndev->device, queue,
 895                         nr_cqe + 1, comp_vector,
 896                         IB_POLL_WORKQUEUE);
 897         if (IS_ERR(queue->cq)) {
 898                 ret = PTR_ERR(queue->cq);
 899                 pr_err("failed to create CQ cqe= %d ret= %d\n",
 900                        nr_cqe + 1, ret);
 901                 goto out;
 902         }
 903
 904         memset(&qp_attr, 0, sizeof(qp_attr));
 905         qp_attr.qp_context = queue;
 906         qp_attr.event_handler = nvmet_rdma_qp_event;
 907         qp_attr.send_cq = queue->cq;
 908         qp_attr.recv_cq = queue->cq;
 909         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 910         qp_attr.qp_type = IB_QPT_RC;
 911         /* +1 for drain */
 912         qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
 913         qp_attr.cap.max_rdma_ctxs = queue->send_queue_size;
 914         qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
 915                                         ndev->device->attrs.max_sge);
 916
 917         if (ndev->srq) {
 918                 qp_attr.srq = ndev->srq;
 919         } else {
 920                 /* +1 for drain */
 921                 qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
 922                 qp_attr.cap.max_recv_sge = 2;
 923         }
 924
 925         ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
 926         if (ret) {
 927                 pr_err("failed to create_qp ret= %d\n", ret);
 928                 goto err_destroy_cq;
 929         }
 930
 931         atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
 932
 933         pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
 934                  __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
 935                  qp_attr.cap.max_send_wr, queue->cm_id);
 936
 937         if (!ndev->srq) {
 938                 for (i = 0; i < queue->recv_queue_size; i++) {
 939                         queue->cmds[i].queue = queue;
 940                         nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
 941                 }
 942         }
 943
 944 out:
 945         return ret;
 946
 947 err_destroy_cq:
 948         ib_free_cq(queue->cq);
 949         goto out;
 950 }
 951
 952 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
 953 {
 954         rdma_destroy_qp(queue->cm_id);
 955         ib_free_cq(queue->cq);
 956 }
 957
 958 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
 959 {
 960         pr_info("freeing queue %d\n", queue->idx);
 961
 962         nvmet_sq_destroy(&queue->nvme_sq);
 963
 964         nvmet_rdma_destroy_queue_ib(queue);
 965         if (!queue->dev->srq) {
 966                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
 967                                 queue->recv_queue_size,
 968                                 !queue->host_qid);
 969         }
 970         nvmet_rdma_free_rsps(queue);
 971         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
 972         kfree(queue);
 973 }
 974
 975 static void nvmet_rdma_release_queue_work(struct work_struct *w)
 976 {
 977         struct nvmet_rdma_queue *queue =
 978                 container_of(w, struct nvmet_rdma_queue, release_work);
 979         struct rdma_cm_id *cm_id = queue->cm_id;
 980         struct nvmet_rdma_device *dev = queue->dev;
 981         enum nvmet_rdma_queue_state state = queue->state;
 982
 983         nvmet_rdma_free_queue(queue);
 984
 985         if (state != NVMET_RDMA_IN_DEVICE_REMOVAL)
 986                 rdma_destroy_id(cm_id);
 987
 988         kref_put(&dev->ref, nvmet_rdma_free_dev);
 989 }
 990
 991 static int
 992 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
 993                                 struct nvmet_rdma_queue *queue)
 994 {
 995         struct nvme_rdma_cm_req *req;
 996
 997         req = (struct nvme_rdma_cm_req *)conn->private_data;
 998         if (!req || conn->private_data_len == 0)
 999                 return NVME_RDMA_CM_INVALID_LEN;
1000
1001         if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1002                 return NVME_RDMA_CM_INVALID_RECFMT;
1003
1004         queue->host_qid = le16_to_cpu(req->qid);
1005
1006         /*
1007          * req->hsqsize corresponds to our recv queue size plus 1
1008          * req->hrqsize corresponds to our send queue size
1009          */
1010         queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1011         queue->send_queue_size = le16_to_cpu(req->hrqsize);
1012
1013         if (!queue->host_qid && queue->recv_queue_size > NVMF_AQ_DEPTH)
1014                 return NVME_RDMA_CM_INVALID_HSQSIZE;
1015
1016         /* XXX: Should we enforce some kind of max for IO queues? */
1017
1018         return 0;
1019 }
1020
1021 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1022                                 enum nvme_rdma_cm_status status)
1023 {
1024         struct nvme_rdma_cm_rej rej;
1025
1026         rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1027         rej.sts = cpu_to_le16(status);
1028
1029         return rdma_reject(cm_id, (void *)&rej, sizeof(rej));
1030 }
1031
1032 static struct nvmet_rdma_queue *
1033 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1034                 struct rdma_cm_id *cm_id,
1035                 struct rdma_cm_event *event)
1036 {
1037         struct nvmet_rdma_queue *queue;
1038         int ret;
1039
1040         queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1041         if (!queue) {
1042                 ret = NVME_RDMA_CM_NO_RSC;
1043                 goto out_reject;
1044         }
1045
1046         ret = nvmet_sq_init(&queue->nvme_sq);
1047         if (ret)
1048                 goto out_free_queue;
1049
1050         ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1051         if (ret)
1052                 goto out_destroy_sq;
1053
1054         /*
1055          * Schedules the actual release because calling rdma_destroy_id from
1056          * inside a CM callback would trigger a deadlock. (great API design..)
1057          */
1058         INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1059         queue->dev = ndev;
1060         queue->cm_id = cm_id;
1061
1062         spin_lock_init(&queue->state_lock);
1063         queue->state = NVMET_RDMA_Q_CONNECTING;
1064         INIT_LIST_HEAD(&queue->rsp_wait_list);
1065         INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1066         spin_lock_init(&queue->rsp_wr_wait_lock);
1067         INIT_LIST_HEAD(&queue->free_rsps);
1068         spin_lock_init(&queue->rsps_lock);
1069
1070         queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1071         if (queue->idx < 0) {
1072                 ret = NVME_RDMA_CM_NO_RSC;
1073                 goto out_free_queue;
1074         }
1075
1076         ret = nvmet_rdma_alloc_rsps(queue);
1077         if (ret) {
1078                 ret = NVME_RDMA_CM_NO_RSC;
1079                 goto out_ida_remove;
1080         }
1081
1082         if (!ndev->srq) {
1083                 queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1084                                 queue->recv_queue_size,
1085                                 !queue->host_qid);
1086                 if (IS_ERR(queue->cmds)) {
1087                         ret = NVME_RDMA_CM_NO_RSC;
1088                         goto out_free_responses;
1089                 }
1090         }
1091
1092         ret = nvmet_rdma_create_queue_ib(queue);
1093         if (ret) {
1094                 pr_err("%s: creating RDMA queue failed (%d).\n",
1095                         __func__, ret);
1096                 ret = NVME_RDMA_CM_NO_RSC;
1097                 goto out_free_cmds;
1098         }
1099
1100         return queue;
1101
1102 out_free_cmds:
1103         if (!ndev->srq) {
1104                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1105                                 queue->recv_queue_size,
1106                                 !queue->host_qid);
1107         }
1108 out_free_responses:
1109         nvmet_rdma_free_rsps(queue);
1110 out_ida_remove:
1111         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1112 out_destroy_sq:
1113         nvmet_sq_destroy(&queue->nvme_sq);
1114 out_free_queue:
1115         kfree(queue);
1116 out_reject:
1117         nvmet_rdma_cm_reject(cm_id, ret);
1118         return NULL;
1119 }
1120
1121 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1122 {
1123         struct nvmet_rdma_queue *queue = priv;
1124
1125         switch (event->event) {
1126         case IB_EVENT_COMM_EST:
1127                 rdma_notify(queue->cm_id, event->event);
1128                 break;
1129         default:
1130                 pr_err("received unrecognized IB QP event %d\n", event->event);
1131                 break;
1132         }
1133 }
1134
1135 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1136                 struct nvmet_rdma_queue *queue,
1137                 struct rdma_conn_param *p)
1138 {
1139         struct rdma_conn_param  param = { };
1140         struct nvme_rdma_cm_rep priv = { };
1141         int ret = -ENOMEM;
1142
1143         param.rnr_retry_count = 7;
1144         param.flow_control = 1;
1145         param.initiator_depth = min_t(u8, p->initiator_depth,
1146                 queue->dev->device->attrs.max_qp_init_rd_atom);
1147         param.private_data = &priv;
1148         param.private_data_len = sizeof(priv);
1149         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1150         priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1151
1152         ret = rdma_accept(cm_id, &param);
1153         if (ret)
1154                 pr_err("rdma_accept failed (error code = %d)\n", ret);
1155
1156         return ret;
1157 }
1158
1159 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1160                 struct rdma_cm_event *event)
1161 {
1162         struct nvmet_rdma_device *ndev;
1163         struct nvmet_rdma_queue *queue;
1164         int ret = -EINVAL;
1165
1166         ndev = nvmet_rdma_find_get_device(cm_id);
1167         if (!ndev) {
1168                 pr_err("no client data!\n");
1169                 nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1170                 return -ECONNREFUSED;
1171         }
1172
1173         queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1174         if (!queue) {
1175                 ret = -ENOMEM;
1176                 goto put_device;
1177         }
1178         queue->port = cm_id->context;
1179
1180         ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1181         if (ret)
1182                 goto release_queue;
1183
1184         mutex_lock(&nvmet_rdma_queue_mutex);
1185         list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1186         mutex_unlock(&nvmet_rdma_queue_mutex);
1187
1188         return 0;
1189
1190 release_queue:
1191         nvmet_rdma_free_queue(queue);
1192 put_device:
1193         kref_put(&ndev->ref, nvmet_rdma_free_dev);
1194
1195         return ret;
1196 }
1197
1198 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1199 {
1200         unsigned long flags;
1201
1202         spin_lock_irqsave(&queue->state_lock, flags);
1203         if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1204                 pr_warn("trying to establish a connected queue\n");
1205                 goto out_unlock;
1206         }
1207         queue->state = NVMET_RDMA_Q_LIVE;
1208
1209         while (!list_empty(&queue->rsp_wait_list)) {
1210                 struct nvmet_rdma_rsp *cmd;
1211
1212                 cmd = list_first_entry(&queue->rsp_wait_list,
1213                                         struct nvmet_rdma_rsp, wait_list);
1214                 list_del(&cmd->wait_list);
1215
1216                 spin_unlock_irqrestore(&queue->state_lock, flags);
1217                 nvmet_rdma_handle_command(queue, cmd);
1218                 spin_lock_irqsave(&queue->state_lock, flags);
1219         }
1220
1221 out_unlock:
1222         spin_unlock_irqrestore(&queue->state_lock, flags);
1223 }
1224
1225 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1226 {
1227         bool disconnect = false;
1228         unsigned long flags;
1229
1230         pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1231
1232         spin_lock_irqsave(&queue->state_lock, flags);
1233         switch (queue->state) {
1234         case NVMET_RDMA_Q_CONNECTING:
1235         case NVMET_RDMA_Q_LIVE:
1236                 queue->state = NVMET_RDMA_Q_DISCONNECTING;
1237         case NVMET_RDMA_IN_DEVICE_REMOVAL:
1238                 disconnect = true;
1239                 break;
1240         case NVMET_RDMA_Q_DISCONNECTING:
1241                 break;
1242         }
1243         spin_unlock_irqrestore(&queue->state_lock, flags);
1244
1245         if (disconnect) {
1246                 rdma_disconnect(queue->cm_id);
1247                 ib_drain_qp(queue->cm_id->qp);
1248                 schedule_work(&queue->release_work);
1249         }
1250 }
1251
1252 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1253 {
1254         bool disconnect = false;
1255
1256         mutex_lock(&nvmet_rdma_queue_mutex);
1257         if (!list_empty(&queue->queue_list)) {
1258                 list_del_init(&queue->queue_list);
1259                 disconnect = true;
1260         }
1261         mutex_unlock(&nvmet_rdma_queue_mutex);
1262
1263         if (disconnect)
1264                 __nvmet_rdma_queue_disconnect(queue);
1265 }
1266
1267 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1268                 struct nvmet_rdma_queue *queue)
1269 {
1270         WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1271
1272         pr_err("failed to connect queue\n");
1273         schedule_work(&queue->release_work);
1274 }
1275
1276 /**
1277  * nvme_rdma_device_removal() - Handle RDMA device removal
1278  * @queue:      nvmet rdma queue (cm id qp_context)
1279  * @addr:       nvmet address (cm_id context)
1280  *
1281  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1282  * to unplug so we should take care of destroying our RDMA resources.
1283  * This event will be generated for each allocated cm_id.
1284  *
1285  * Note that this event can be generated on a normal queue cm_id
1286  * and/or a device bound listener cm_id (where in this case
1287  * queue will be null).
1288  *
1289  * we claim ownership on destroying the cm_id. For queues we move
1290  * the queue state to NVMET_RDMA_IN_DEVICE_REMOVAL and for port
1291  * we nullify the priv to prevent double cm_id destruction and destroying
1292  * the cm_id implicitely by returning a non-zero rc to the callout.
1293  */
1294 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1295                 struct nvmet_rdma_queue *queue)
1296 {
1297         unsigned long flags;
1298
1299         if (!queue) {
1300                 struct nvmet_port *port = cm_id->context;
1301
1302                 /*
1303                  * This is a listener cm_id. Make sure that
1304                  * future remove_port won't invoke a double
1305                  * cm_id destroy. use atomic xchg to make sure
1306                  * we don't compete with remove_port.
1307                  */
1308                 if (xchg(&port->priv, NULL) != cm_id)
1309                         return 0;
1310         } else {
1311                 /*
1312                  * This is a queue cm_id. Make sure that
1313                  * release queue will not destroy the cm_id
1314                  * and schedule all ctrl queues removal (only
1315                  * if the queue is not disconnecting already).
1316                  */
1317                 spin_lock_irqsave(&queue->state_lock, flags);
1318                 if (queue->state != NVMET_RDMA_Q_DISCONNECTING)
1319                         queue->state = NVMET_RDMA_IN_DEVICE_REMOVAL;
1320                 spin_unlock_irqrestore(&queue->state_lock, flags);
1321                 nvmet_rdma_queue_disconnect(queue);
1322                 flush_scheduled_work();
1323         }
1324
1325         /*
1326          * We need to return 1 so that the core will destroy
1327          * it's own ID.  What a great API design..
1328          */
1329         return 1;
1330 }
1331
1332 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1333                 struct rdma_cm_event *event)
1334 {
1335         struct nvmet_rdma_queue *queue = NULL;
1336         int ret = 0;
1337
1338         if (cm_id->qp)
1339                 queue = cm_id->qp->qp_context;
1340
1341         pr_debug("%s (%d): status %d id %p\n",
1342                 rdma_event_msg(event->event), event->event,
1343                 event->status, cm_id);
1344
1345         switch (event->event) {
1346         case RDMA_CM_EVENT_CONNECT_REQUEST:
1347                 ret = nvmet_rdma_queue_connect(cm_id, event);
1348                 break;
1349         case RDMA_CM_EVENT_ESTABLISHED:
1350                 nvmet_rdma_queue_established(queue);
1351                 break;
1352         case RDMA_CM_EVENT_ADDR_CHANGE:
1353         case RDMA_CM_EVENT_DISCONNECTED:
1354         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1355                 nvmet_rdma_queue_disconnect(queue);
1356                 break;
1357         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1358                 ret = nvmet_rdma_device_removal(cm_id, queue);
1359                 break;
1360         case RDMA_CM_EVENT_REJECTED:
1361         case RDMA_CM_EVENT_UNREACHABLE:
1362         case RDMA_CM_EVENT_CONNECT_ERROR:
1363                 nvmet_rdma_queue_connect_fail(cm_id, queue);
1364                 break;
1365         default:
1366                 pr_err("received unrecognized RDMA CM event %d\n",
1367                         event->event);
1368                 break;
1369         }
1370
1371         return ret;
1372 }
1373
1374 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1375 {
1376         struct nvmet_rdma_queue *queue;
1377
1378 restart:
1379         mutex_lock(&nvmet_rdma_queue_mutex);
1380         list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1381                 if (queue->nvme_sq.ctrl == ctrl) {
1382                         list_del_init(&queue->queue_list);
1383                         mutex_unlock(&nvmet_rdma_queue_mutex);
1384
1385                         __nvmet_rdma_queue_disconnect(queue);
1386                         goto restart;
1387                 }
1388         }
1389         mutex_unlock(&nvmet_rdma_queue_mutex);
1390 }
1391
1392 static int nvmet_rdma_add_port(struct nvmet_port *port)
1393 {
1394         struct rdma_cm_id *cm_id;
1395         struct sockaddr_in addr_in;
1396         u16 port_in;
1397         int ret;
1398
1399         switch (port->disc_addr.adrfam) {
1400         case NVMF_ADDR_FAMILY_IP4:
1401                 break;
1402         default:
1403                 pr_err("address family %d not supported\n",
1404                                 port->disc_addr.adrfam);
1405                 return -EINVAL;
1406         }
1407
1408         ret = kstrtou16(port->disc_addr.trsvcid, 0, &port_in);
1409         if (ret)
1410                 return ret;
1411
1412         addr_in.sin_family = AF_INET;
1413         addr_in.sin_addr.s_addr = in_aton(port->disc_addr.traddr);
1414         addr_in.sin_port = htons(port_in);
1415
1416         cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1417                         RDMA_PS_TCP, IB_QPT_RC);
1418         if (IS_ERR(cm_id)) {
1419                 pr_err("CM ID creation failed\n");
1420                 return PTR_ERR(cm_id);
1421         }
1422
1423         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr_in);
1424         if (ret) {
1425                 pr_err("binding CM ID to %pISpc failed (%d)\n", &addr_in, ret);
1426                 goto out_destroy_id;
1427         }
1428
1429         ret = rdma_listen(cm_id, 128);
1430         if (ret) {
1431                 pr_err("listening to %pISpc failed (%d)\n", &addr_in, ret);
1432                 goto out_destroy_id;
1433         }
1434
1435         pr_info("enabling port %d (%pISpc)\n",
1436                 le16_to_cpu(port->disc_addr.portid), &addr_in);
1437         port->priv = cm_id;
1438         return 0;
1439
1440 out_destroy_id:
1441         rdma_destroy_id(cm_id);
1442         return ret;
1443 }
1444
1445 static void nvmet_rdma_remove_port(struct nvmet_port *port)
1446 {
1447         struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
1448
1449         if (cm_id)
1450                 rdma_destroy_id(cm_id);
1451 }
1452
1453 static struct nvmet_fabrics_ops nvmet_rdma_ops = {
1454         .owner                  = THIS_MODULE,
1455         .type                   = NVMF_TRTYPE_RDMA,
1456         .sqe_inline_size        = NVMET_RDMA_INLINE_DATA_SIZE,
1457         .msdbd                  = 1,
1458         .has_keyed_sgls         = 1,
1459         .add_port               = nvmet_rdma_add_port,
1460         .remove_port            = nvmet_rdma_remove_port,
1461         .queue_response         = nvmet_rdma_queue_response,
1462         .delete_ctrl            = nvmet_rdma_delete_ctrl,
1463 };
1464
1465 static int __init nvmet_rdma_init(void)
1466 {
1467         return nvmet_register_transport(&nvmet_rdma_ops);
1468 }
1469
1470 static void __exit nvmet_rdma_exit(void)
1471 {
1472         struct nvmet_rdma_queue *queue;
1473
1474         nvmet_unregister_transport(&nvmet_rdma_ops);
1475
1476         flush_scheduled_work();
1477
1478         mutex_lock(&nvmet_rdma_queue_mutex);
1479         while ((queue = list_first_entry_or_null(&nvmet_rdma_queue_list,
1480                         struct nvmet_rdma_queue, queue_list))) {
1481                 list_del_init(&queue->queue_list);
1482
1483                 mutex_unlock(&nvmet_rdma_queue_mutex);
1484                 __nvmet_rdma_queue_disconnect(queue);
1485                 mutex_lock(&nvmet_rdma_queue_mutex);
1486         }
1487         mutex_unlock(&nvmet_rdma_queue_mutex);
1488
1489         flush_scheduled_work();
1490         ida_destroy(&nvmet_rdma_queue_ida);
1491 }
1492
1493 module_init(nvmet_rdma_init);
1494 module_exit(nvmet_rdma_exit);
1495
1496 MODULE_LICENSE("GPL v2");
1497 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */