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         bool                    allocated;
  69         u8                      n_rdma;
  70         u32                     flags;
  71         u32                     invalidate_rkey;
  72
  73         struct list_head        wait_list;
  74         struct list_head        free_list;
  75 };
  76
  77 enum nvmet_rdma_queue_state {
  78         NVMET_RDMA_Q_CONNECTING,
  79         NVMET_RDMA_Q_LIVE,
  80         NVMET_RDMA_Q_DISCONNECTING,
  81         NVMET_RDMA_IN_DEVICE_REMOVAL,
  82 };
  83
  84 struct nvmet_rdma_queue {
  85         struct rdma_cm_id       *cm_id;
  86         struct nvmet_port       *port;
  87         struct ib_cq            *cq;
  88         atomic_t                sq_wr_avail;
  89         struct nvmet_rdma_device *dev;
  90         spinlock_t              state_lock;
  91         enum nvmet_rdma_queue_state state;
  92         struct nvmet_cq         nvme_cq;
  93         struct nvmet_sq         nvme_sq;
  94
  95         struct nvmet_rdma_rsp   *rsps;
  96         struct list_head        free_rsps;
  97         spinlock_t              rsps_lock;
  98         struct nvmet_rdma_cmd   *cmds;
  99
 100         struct work_struct      release_work;
 101         struct list_head        rsp_wait_list;
 102         struct list_head        rsp_wr_wait_list;
 103         spinlock_t              rsp_wr_wait_lock;
 104
 105         int                     idx;
 106         int                     host_qid;
 107         int                     recv_queue_size;
 108         int                     send_queue_size;
 109
 110         struct list_head        queue_list;
 111 };
 112
 113 struct nvmet_rdma_device {
 114         struct ib_device        *device;
 115         struct ib_pd            *pd;
 116         struct ib_srq           *srq;
 117         struct nvmet_rdma_cmd   *srq_cmds;
 118         size_t                  srq_size;
 119         struct kref             ref;
 120         struct list_head        entry;
 121 };
 122
 123 static bool nvmet_rdma_use_srq;
 124 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
 125 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
 126
 127 static DEFINE_IDA(nvmet_rdma_queue_ida);
 128 static LIST_HEAD(nvmet_rdma_queue_list);
 129 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
 130
 131 static LIST_HEAD(device_list);
 132 static DEFINE_MUTEX(device_list_mutex);
 133
 134 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
 135 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
 136 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 137 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
 138 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
 139 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
 140 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
 141                                 struct nvmet_rdma_rsp *r);
 142 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
 143                                 struct nvmet_rdma_rsp *r);
 144
 145 static struct nvmet_fabrics_ops nvmet_rdma_ops;
 146
 147 /* XXX: really should move to a generic header sooner or later.. */
 148 static inline u32 get_unaligned_le24(const u8 *p)
 149 {
 150         return (u32)p[0] | (u32)p[1] << 8 | (u32)p[2] << 16;
 151 }
 152
 153 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
 154 {
 155         return nvme_is_write(rsp->req.cmd) &&
 156                 rsp->req.data_len &&
 157                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 158 }
 159
 160 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
 161 {
 162         return !nvme_is_write(rsp->req.cmd) &&
 163                 rsp->req.data_len &&
 164                 !rsp->req.rsp->status &&
 165                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 166 }
 167
 168 static inline struct nvmet_rdma_rsp *
 169 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
 170 {
 171         struct nvmet_rdma_rsp *rsp;
 172         unsigned long flags;
 173
 174         spin_lock_irqsave(&queue->rsps_lock, flags);
 175         rsp = list_first_entry_or_null(&queue->free_rsps,
 176                                 struct nvmet_rdma_rsp, free_list);
 177         if (likely(rsp))
 178                 list_del(&rsp->free_list);
 179         spin_unlock_irqrestore(&queue->rsps_lock, flags);
 180
 181         if (unlikely(!rsp)) {
 182                 int ret;
 183
 184                 rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
 185                 if (unlikely(!rsp))
 186                         return NULL;
 187                 ret = nvmet_rdma_alloc_rsp(queue->dev, rsp);
 188                 if (unlikely(ret)) {
 189                         kfree(rsp);
 190                         return NULL;
 191                 }
 192
 193                 rsp->allocated = true;
 194         }
 195
 196         return rsp;
 197 }
 198
 199 static inline void
 200 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
 201 {
 202         unsigned long flags;
 203
 204         if (unlikely(rsp->allocated)) {
 205                 nvmet_rdma_free_rsp(rsp->queue->dev, rsp);
 206                 kfree(rsp);
 207                 return;
 208         }
 209
 210         spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
 211         list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
 212         spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
 213 }
 214
 215 static void nvmet_rdma_free_sgl(struct scatterlist *sgl, unsigned int nents)
 216 {
 217         struct scatterlist *sg;
 218         int count;
 219
 220         if (!sgl || !nents)
 221                 return;
 222
 223         for_each_sg(sgl, sg, nents, count)
 224                 __free_page(sg_page(sg));
 225         kfree(sgl);
 226 }
 227
 228 static int nvmet_rdma_alloc_sgl(struct scatterlist **sgl, unsigned int *nents,
 229                 u32 length)
 230 {
 231         struct scatterlist *sg;
 232         struct page *page;
 233         unsigned int nent;
 234         int i = 0;
 235
 236         nent = DIV_ROUND_UP(length, PAGE_SIZE);
 237         sg = kmalloc_array(nent, sizeof(struct scatterlist), GFP_KERNEL);
 238         if (!sg)
 239                 goto out;
 240
 241         sg_init_table(sg, nent);
 242
 243         while (length) {
 244                 u32 page_len = min_t(u32, length, PAGE_SIZE);
 245
 246                 page = alloc_page(GFP_KERNEL);
 247                 if (!page)
 248                         goto out_free_pages;
 249
 250                 sg_set_page(&sg[i], page, page_len, 0);
 251                 length -= page_len;
 252                 i++;
 253         }
 254         *sgl = sg;
 255         *nents = nent;
 256         return 0;
 257
 258 out_free_pages:
 259         while (i > 0) {
 260                 i--;
 261                 __free_page(sg_page(&sg[i]));
 262         }
 263         kfree(sg);
 264 out:
 265         return NVME_SC_INTERNAL;
 266 }
 267
 268 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
 269                         struct nvmet_rdma_cmd *c, bool admin)
 270 {
 271         /* NVMe command / RDMA RECV */
 272         c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
 273         if (!c->nvme_cmd)
 274                 goto out;
 275
 276         c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
 277                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 278         if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
 279                 goto out_free_cmd;
 280
 281         c->sge[0].length = sizeof(*c->nvme_cmd);
 282         c->sge[0].lkey = ndev->pd->local_dma_lkey;
 283
 284         if (!admin) {
 285                 c->inline_page = alloc_pages(GFP_KERNEL,
 286                                 get_order(NVMET_RDMA_INLINE_DATA_SIZE));
 287                 if (!c->inline_page)
 288                         goto out_unmap_cmd;
 289                 c->sge[1].addr = ib_dma_map_page(ndev->device,
 290                                 c->inline_page, 0, NVMET_RDMA_INLINE_DATA_SIZE,
 291                                 DMA_FROM_DEVICE);
 292                 if (ib_dma_mapping_error(ndev->device, c->sge[1].addr))
 293                         goto out_free_inline_page;
 294                 c->sge[1].length = NVMET_RDMA_INLINE_DATA_SIZE;
 295                 c->sge[1].lkey = ndev->pd->local_dma_lkey;
 296         }
 297
 298         c->cqe.done = nvmet_rdma_recv_done;
 299
 300         c->wr.wr_cqe = &c->cqe;
 301         c->wr.sg_list = c->sge;
 302         c->wr.num_sge = admin ? 1 : 2;
 303
 304         return 0;
 305
 306 out_free_inline_page:
 307         if (!admin) {
 308                 __free_pages(c->inline_page,
 309                                 get_order(NVMET_RDMA_INLINE_DATA_SIZE));
 310         }
 311 out_unmap_cmd:
 312         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 313                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 314 out_free_cmd:
 315         kfree(c->nvme_cmd);
 316
 317 out:
 318         return -ENOMEM;
 319 }
 320
 321 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
 322                 struct nvmet_rdma_cmd *c, bool admin)
 323 {
 324         if (!admin) {
 325                 ib_dma_unmap_page(ndev->device, c->sge[1].addr,
 326                                 NVMET_RDMA_INLINE_DATA_SIZE, DMA_FROM_DEVICE);
 327                 __free_pages(c->inline_page,
 328                                 get_order(NVMET_RDMA_INLINE_DATA_SIZE));
 329         }
 330         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 331                                 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 332         kfree(c->nvme_cmd);
 333 }
 334
 335 static struct nvmet_rdma_cmd *
 336 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
 337                 int nr_cmds, bool admin)
 338 {
 339         struct nvmet_rdma_cmd *cmds;
 340         int ret = -EINVAL, i;
 341
 342         cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
 343         if (!cmds)
 344                 goto out;
 345
 346         for (i = 0; i < nr_cmds; i++) {
 347                 ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
 348                 if (ret)
 349                         goto out_free;
 350         }
 351
 352         return cmds;
 353
 354 out_free:
 355         while (--i >= 0)
 356                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 357         kfree(cmds);
 358 out:
 359         return ERR_PTR(ret);
 360 }
 361
 362 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
 363                 struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
 364 {
 365         int i;
 366
 367         for (i = 0; i < nr_cmds; i++)
 368                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 369         kfree(cmds);
 370 }
 371
 372 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
 373                 struct nvmet_rdma_rsp *r)
 374 {
 375         /* NVMe CQE / RDMA SEND */
 376         r->req.rsp = kmalloc(sizeof(*r->req.rsp), GFP_KERNEL);
 377         if (!r->req.rsp)
 378                 goto out;
 379
 380         r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.rsp,
 381                         sizeof(*r->req.rsp), DMA_TO_DEVICE);
 382         if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
 383                 goto out_free_rsp;
 384
 385         r->send_sge.length = sizeof(*r->req.rsp);
 386         r->send_sge.lkey = ndev->pd->local_dma_lkey;
 387
 388         r->send_cqe.done = nvmet_rdma_send_done;
 389
 390         r->send_wr.wr_cqe = &r->send_cqe;
 391         r->send_wr.sg_list = &r->send_sge;
 392         r->send_wr.num_sge = 1;
 393         r->send_wr.send_flags = IB_SEND_SIGNALED;
 394
 395         /* Data In / RDMA READ */
 396         r->read_cqe.done = nvmet_rdma_read_data_done;
 397         return 0;
 398
 399 out_free_rsp:
 400         kfree(r->req.rsp);
 401 out:
 402         return -ENOMEM;
 403 }
 404
 405 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
 406                 struct nvmet_rdma_rsp *r)
 407 {
 408         ib_dma_unmap_single(ndev->device, r->send_sge.addr,
 409                                 sizeof(*r->req.rsp), DMA_TO_DEVICE);
 410         kfree(r->req.rsp);
 411 }
 412
 413 static int
 414 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
 415 {
 416         struct nvmet_rdma_device *ndev = queue->dev;
 417         int nr_rsps = queue->recv_queue_size * 2;
 418         int ret = -EINVAL, i;
 419
 420         queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
 421                         GFP_KERNEL);
 422         if (!queue->rsps)
 423                 goto out;
 424
 425         for (i = 0; i < nr_rsps; i++) {
 426                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 427
 428                 ret = nvmet_rdma_alloc_rsp(ndev, rsp);
 429                 if (ret)
 430                         goto out_free;
 431
 432                 list_add_tail(&rsp->free_list, &queue->free_rsps);
 433         }
 434
 435         return 0;
 436
 437 out_free:
 438         while (--i >= 0) {
 439                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 440
 441                 list_del(&rsp->free_list);
 442                 nvmet_rdma_free_rsp(ndev, rsp);
 443         }
 444         kfree(queue->rsps);
 445 out:
 446         return ret;
 447 }
 448
 449 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
 450 {
 451         struct nvmet_rdma_device *ndev = queue->dev;
 452         int i, nr_rsps = queue->recv_queue_size * 2;
 453
 454         for (i = 0; i < nr_rsps; i++) {
 455                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 456
 457                 list_del(&rsp->free_list);
 458                 nvmet_rdma_free_rsp(ndev, rsp);
 459         }
 460         kfree(queue->rsps);
 461 }
 462
 463 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
 464                 struct nvmet_rdma_cmd *cmd)
 465 {
 466         struct ib_recv_wr *bad_wr;
 467
 468         ib_dma_sync_single_for_device(ndev->device,
 469                 cmd->sge[0].addr, cmd->sge[0].length,
 470                 DMA_FROM_DEVICE);
 471
 472         if (ndev->srq)
 473                 return ib_post_srq_recv(ndev->srq, &cmd->wr, &bad_wr);
 474         return ib_post_recv(cmd->queue->cm_id->qp, &cmd->wr, &bad_wr);
 475 }
 476
 477 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
 478 {
 479         spin_lock(&queue->rsp_wr_wait_lock);
 480         while (!list_empty(&queue->rsp_wr_wait_list)) {
 481                 struct nvmet_rdma_rsp *rsp;
 482                 bool ret;
 483
 484                 rsp = list_entry(queue->rsp_wr_wait_list.next,
 485                                 struct nvmet_rdma_rsp, wait_list);
 486                 list_del(&rsp->wait_list);
 487
 488                 spin_unlock(&queue->rsp_wr_wait_lock);
 489                 ret = nvmet_rdma_execute_command(rsp);
 490                 spin_lock(&queue->rsp_wr_wait_lock);
 491
 492                 if (!ret) {
 493                         list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
 494                         break;
 495                 }
 496         }
 497         spin_unlock(&queue->rsp_wr_wait_lock);
 498 }
 499
 500
 501 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
 502 {
 503         struct nvmet_rdma_queue *queue = rsp->queue;
 504
 505         atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 506
 507         if (rsp->n_rdma) {
 508                 rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
 509                                 queue->cm_id->port_num, rsp->req.sg,
 510                                 rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
 511         }
 512
 513         if (rsp->req.sg != &rsp->cmd->inline_sg)
 514                 nvmet_rdma_free_sgl(rsp->req.sg, rsp->req.sg_cnt);
 515
 516         if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
 517                 nvmet_rdma_process_wr_wait_list(queue);
 518
 519         nvmet_rdma_put_rsp(rsp);
 520 }
 521
 522 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
 523 {
 524         if (queue->nvme_sq.ctrl) {
 525                 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
 526         } else {
 527                 /*
 528                  * we didn't setup the controller yet in case
 529                  * of admin connect error, just disconnect and
 530                  * cleanup the queue
 531                  */
 532                 nvmet_rdma_queue_disconnect(queue);
 533         }
 534 }
 535
 536 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
 537 {
 538         struct nvmet_rdma_rsp *rsp =
 539                 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
 540         struct nvmet_rdma_queue *queue = cq->cq_context;
 541
 542         nvmet_rdma_release_rsp(rsp);
 543
 544         if (unlikely(wc->status != IB_WC_SUCCESS &&
 545                      wc->status != IB_WC_WR_FLUSH_ERR)) {
 546                 pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
 547                         wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
 548                 nvmet_rdma_error_comp(queue);
 549         }
 550 }
 551
 552 static void nvmet_rdma_queue_response(struct nvmet_req *req)
 553 {
 554         struct nvmet_rdma_rsp *rsp =
 555                 container_of(req, struct nvmet_rdma_rsp, req);
 556         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 557         struct ib_send_wr *first_wr, *bad_wr;
 558
 559         if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
 560                 rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
 561                 rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
 562         } else {
 563                 rsp->send_wr.opcode = IB_WR_SEND;
 564         }
 565
 566         if (nvmet_rdma_need_data_out(rsp))
 567                 first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
 568                                 cm_id->port_num, NULL, &rsp->send_wr);
 569         else
 570                 first_wr = &rsp->send_wr;
 571
 572         nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
 573
 574         ib_dma_sync_single_for_device(rsp->queue->dev->device,
 575                 rsp->send_sge.addr, rsp->send_sge.length,
 576                 DMA_TO_DEVICE);
 577
 578         if (ib_post_send(cm_id->qp, first_wr, &bad_wr)) {
 579                 pr_err("sending cmd response failed\n");
 580                 nvmet_rdma_release_rsp(rsp);
 581         }
 582 }
 583
 584 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
 585 {
 586         struct nvmet_rdma_rsp *rsp =
 587                 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
 588         struct nvmet_rdma_queue *queue = cq->cq_context;
 589
 590         WARN_ON(rsp->n_rdma <= 0);
 591         atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
 592         rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
 593                         queue->cm_id->port_num, rsp->req.sg,
 594                         rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
 595         rsp->n_rdma = 0;
 596
 597         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 598                 nvmet_rdma_release_rsp(rsp);
 599                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 600                         pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
 601                                 wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
 602                         nvmet_rdma_error_comp(queue);
 603                 }
 604                 return;
 605         }
 606
 607         rsp->req.execute(&rsp->req);
 608 }
 609
 610 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
 611                 u64 off)
 612 {
 613         sg_init_table(&rsp->cmd->inline_sg, 1);
 614         sg_set_page(&rsp->cmd->inline_sg, rsp->cmd->inline_page, len, off);
 615         rsp->req.sg = &rsp->cmd->inline_sg;
 616         rsp->req.sg_cnt = 1;
 617 }
 618
 619 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
 620 {
 621         struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
 622         u64 off = le64_to_cpu(sgl->addr);
 623         u32 len = le32_to_cpu(sgl->length);
 624
 625         if (!nvme_is_write(rsp->req.cmd))
 626                 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 627
 628         if (off + len > NVMET_RDMA_INLINE_DATA_SIZE) {
 629                 pr_err("invalid inline data offset!\n");
 630                 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
 631         }
 632
 633         /* no data command? */
 634         if (!len)
 635                 return 0;
 636
 637         nvmet_rdma_use_inline_sg(rsp, len, off);
 638         rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
 639         return 0;
 640 }
 641
 642 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
 643                 struct nvme_keyed_sgl_desc *sgl, bool invalidate)
 644 {
 645         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 646         u64 addr = le64_to_cpu(sgl->addr);
 647         u32 len = get_unaligned_le24(sgl->length);
 648         u32 key = get_unaligned_le32(sgl->key);
 649         int ret;
 650         u16 status;
 651
 652         /* no data command? */
 653         if (!len)
 654                 return 0;
 655
 656         status = nvmet_rdma_alloc_sgl(&rsp->req.sg, &rsp->req.sg_cnt,
 657                         len);
 658         if (status)
 659                 return status;
 660
 661         ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
 662                         rsp->req.sg, rsp->req.sg_cnt, 0, addr, key,
 663                         nvmet_data_dir(&rsp->req));
 664         if (ret < 0)
 665                 return NVME_SC_INTERNAL;
 666         rsp->n_rdma += ret;
 667
 668         if (invalidate) {
 669                 rsp->invalidate_rkey = key;
 670                 rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
 671         }
 672
 673         return 0;
 674 }
 675
 676 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
 677 {
 678         struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
 679
 680         switch (sgl->type >> 4) {
 681         case NVME_SGL_FMT_DATA_DESC:
 682                 switch (sgl->type & 0xf) {
 683                 case NVME_SGL_FMT_OFFSET:
 684                         return nvmet_rdma_map_sgl_inline(rsp);
 685                 default:
 686                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 687                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 688                 }
 689         case NVME_KEY_SGL_FMT_DATA_DESC:
 690                 switch (sgl->type & 0xf) {
 691                 case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
 692                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
 693                 case NVME_SGL_FMT_ADDRESS:
 694                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
 695                 default:
 696                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 697                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 698                 }
 699         default:
 700                 pr_err("invalid SGL type: %#x\n", sgl->type);
 701                 return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
 702         }
 703 }
 704
 705 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
 706 {
 707         struct nvmet_rdma_queue *queue = rsp->queue;
 708
 709         if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
 710                         &queue->sq_wr_avail) < 0)) {
 711                 pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
 712                                 1 + rsp->n_rdma, queue->idx,
 713                                 queue->nvme_sq.ctrl->cntlid);
 714                 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 715                 return false;
 716         }
 717
 718         if (nvmet_rdma_need_data_in(rsp)) {
 719                 if (rdma_rw_ctx_post(&rsp->rw, queue->cm_id->qp,
 720                                 queue->cm_id->port_num, &rsp->read_cqe, NULL))
 721                         nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
 722         } else {
 723                 rsp->req.execute(&rsp->req);
 724         }
 725
 726         return true;
 727 }
 728
 729 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
 730                 struct nvmet_rdma_rsp *cmd)
 731 {
 732         u16 status;
 733
 734         ib_dma_sync_single_for_cpu(queue->dev->device,
 735                 cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
 736                 DMA_FROM_DEVICE);
 737         ib_dma_sync_single_for_cpu(queue->dev->device,
 738                 cmd->send_sge.addr, cmd->send_sge.length,
 739                 DMA_TO_DEVICE);
 740
 741         if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
 742                         &queue->nvme_sq, &nvmet_rdma_ops))
 743                 return;
 744
 745         status = nvmet_rdma_map_sgl(cmd);
 746         if (status)
 747                 goto out_err;
 748
 749         if (unlikely(!nvmet_rdma_execute_command(cmd))) {
 750                 spin_lock(&queue->rsp_wr_wait_lock);
 751                 list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
 752                 spin_unlock(&queue->rsp_wr_wait_lock);
 753         }
 754
 755         return;
 756
 757 out_err:
 758         nvmet_req_complete(&cmd->req, status);
 759 }
 760
 761 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
 762 {
 763         struct nvmet_rdma_cmd *cmd =
 764                 container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
 765         struct nvmet_rdma_queue *queue = cq->cq_context;
 766         struct nvmet_rdma_rsp *rsp;
 767
 768         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 769                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 770                         pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
 771                                 wc->wr_cqe, ib_wc_status_msg(wc->status),
 772                                 wc->status);
 773                         nvmet_rdma_error_comp(queue);
 774                 }
 775                 return;
 776         }
 777
 778         if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
 779                 pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
 780                 nvmet_rdma_error_comp(queue);
 781                 return;
 782         }
 783
 784         cmd->queue = queue;
 785         rsp = nvmet_rdma_get_rsp(queue);
 786         if (unlikely(!rsp)) {
 787                 /*
 788                  * we get here only under memory pressure,
 789                  * silently drop and have the host retry
 790                  * as we can't even fail it.
 791                  */
 792                 nvmet_rdma_post_recv(queue->dev, cmd);
 793                 return;
 794         }
 795         rsp->queue = queue;
 796         rsp->cmd = cmd;
 797         rsp->flags = 0;
 798         rsp->req.cmd = cmd->nvme_cmd;
 799         rsp->req.port = queue->port;
 800         rsp->n_rdma = 0;
 801
 802         if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
 803                 unsigned long flags;
 804
 805                 spin_lock_irqsave(&queue->state_lock, flags);
 806                 if (queue->state == NVMET_RDMA_Q_CONNECTING)
 807                         list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
 808                 else
 809                         nvmet_rdma_put_rsp(rsp);
 810                 spin_unlock_irqrestore(&queue->state_lock, flags);
 811                 return;
 812         }
 813
 814         nvmet_rdma_handle_command(queue, rsp);
 815 }
 816
 817 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_device *ndev)
 818 {
 819         if (!ndev->srq)
 820                 return;
 821
 822         nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
 823         ib_destroy_srq(ndev->srq);
 824 }
 825
 826 static int nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
 827 {
 828         struct ib_srq_init_attr srq_attr = { NULL, };
 829         struct ib_srq *srq;
 830         size_t srq_size;
 831         int ret, i;
 832
 833         srq_size = 4095;        /* XXX: tune */
 834
 835         srq_attr.attr.max_wr = srq_size;
 836         srq_attr.attr.max_sge = 2;
 837         srq_attr.attr.srq_limit = 0;
 838         srq_attr.srq_type = IB_SRQT_BASIC;
 839         srq = ib_create_srq(ndev->pd, &srq_attr);
 840         if (IS_ERR(srq)) {
 841                 /*
 842                  * If SRQs aren't supported we just go ahead and use normal
 843                  * non-shared receive queues.
 844                  */
 845                 pr_info("SRQ requested but not supported.\n");
 846                 return 0;
 847         }
 848
 849         ndev->srq_cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
 850         if (IS_ERR(ndev->srq_cmds)) {
 851                 ret = PTR_ERR(ndev->srq_cmds);
 852                 goto out_destroy_srq;
 853         }
 854
 855         ndev->srq = srq;
 856         ndev->srq_size = srq_size;
 857
 858         for (i = 0; i < srq_size; i++)
 859                 nvmet_rdma_post_recv(ndev, &ndev->srq_cmds[i]);
 860
 861         return 0;
 862
 863 out_destroy_srq:
 864         ib_destroy_srq(srq);
 865         return ret;
 866 }
 867
 868 static void nvmet_rdma_free_dev(struct kref *ref)
 869 {
 870         struct nvmet_rdma_device *ndev =
 871                 container_of(ref, struct nvmet_rdma_device, ref);
 872
 873         mutex_lock(&device_list_mutex);
 874         list_del(&ndev->entry);
 875         mutex_unlock(&device_list_mutex);
 876
 877         nvmet_rdma_destroy_srq(ndev);
 878         ib_dealloc_pd(ndev->pd);
 879
 880         kfree(ndev);
 881 }
 882
 883 static struct nvmet_rdma_device *
 884 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
 885 {
 886         struct nvmet_rdma_device *ndev;
 887         int ret;
 888
 889         mutex_lock(&device_list_mutex);
 890         list_for_each_entry(ndev, &device_list, entry) {
 891                 if (ndev->device->node_guid == cm_id->device->node_guid &&
 892                     kref_get_unless_zero(&ndev->ref))
 893                         goto out_unlock;
 894         }
 895
 896         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 897         if (!ndev)
 898                 goto out_err;
 899
 900         ndev->device = cm_id->device;
 901         kref_init(&ndev->ref);
 902
 903         ndev->pd = ib_alloc_pd(ndev->device, 0);
 904         if (IS_ERR(ndev->pd))
 905                 goto out_free_dev;
 906
 907         if (nvmet_rdma_use_srq) {
 908                 ret = nvmet_rdma_init_srq(ndev);
 909                 if (ret)
 910                         goto out_free_pd;
 911         }
 912
 913         list_add(&ndev->entry, &device_list);
 914 out_unlock:
 915         mutex_unlock(&device_list_mutex);
 916         pr_debug("added %s.\n", ndev->device->name);
 917         return ndev;
 918
 919 out_free_pd:
 920         ib_dealloc_pd(ndev->pd);
 921 out_free_dev:
 922         kfree(ndev);
 923 out_err:
 924         mutex_unlock(&device_list_mutex);
 925         return NULL;
 926 }
 927
 928 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
 929 {
 930         struct ib_qp_init_attr qp_attr;
 931         struct nvmet_rdma_device *ndev = queue->dev;
 932         int comp_vector, nr_cqe, ret, i;
 933
 934         /*
 935          * Spread the io queues across completion vectors,
 936          * but still keep all admin queues on vector 0.
 937          */
 938         comp_vector = !queue->host_qid ? 0 :
 939                 queue->idx % ndev->device->num_comp_vectors;
 940
 941         /*
 942          * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
 943          */
 944         nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
 945
 946         queue->cq = ib_alloc_cq(ndev->device, queue,
 947                         nr_cqe + 1, comp_vector,
 948                         IB_POLL_WORKQUEUE);
 949         if (IS_ERR(queue->cq)) {
 950                 ret = PTR_ERR(queue->cq);
 951                 pr_err("failed to create CQ cqe= %d ret= %d\n",
 952                        nr_cqe + 1, ret);
 953                 goto out;
 954         }
 955
 956         memset(&qp_attr, 0, sizeof(qp_attr));
 957         qp_attr.qp_context = queue;
 958         qp_attr.event_handler = nvmet_rdma_qp_event;
 959         qp_attr.send_cq = queue->cq;
 960         qp_attr.recv_cq = queue->cq;
 961         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 962         qp_attr.qp_type = IB_QPT_RC;
 963         /* +1 for drain */
 964         qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
 965         qp_attr.cap.max_rdma_ctxs = queue->send_queue_size;
 966         qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
 967                                         ndev->device->attrs.max_sge);
 968
 969         if (ndev->srq) {
 970                 qp_attr.srq = ndev->srq;
 971         } else {
 972                 /* +1 for drain */
 973                 qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
 974                 qp_attr.cap.max_recv_sge = 2;
 975         }
 976
 977         ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
 978         if (ret) {
 979                 pr_err("failed to create_qp ret= %d\n", ret);
 980                 goto err_destroy_cq;
 981         }
 982
 983         atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
 984
 985         pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
 986                  __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
 987                  qp_attr.cap.max_send_wr, queue->cm_id);
 988
 989         if (!ndev->srq) {
 990                 for (i = 0; i < queue->recv_queue_size; i++) {
 991                         queue->cmds[i].queue = queue;
 992                         nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
 993                 }
 994         }
 995
 996 out:
 997         return ret;
 998
 999 err_destroy_cq:
1000         ib_free_cq(queue->cq);
1001         goto out;
1002 }
1003
1004 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1005 {
1006         ib_drain_qp(queue->cm_id->qp);
1007         rdma_destroy_qp(queue->cm_id);
1008         ib_free_cq(queue->cq);
1009 }
1010
1011 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1012 {
1013         pr_info("freeing queue %d\n", queue->idx);
1014
1015         nvmet_sq_destroy(&queue->nvme_sq);
1016
1017         nvmet_rdma_destroy_queue_ib(queue);
1018         if (!queue->dev->srq) {
1019                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1020                                 queue->recv_queue_size,
1021                                 !queue->host_qid);
1022         }
1023         nvmet_rdma_free_rsps(queue);
1024         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1025         kfree(queue);
1026 }
1027
1028 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1029 {
1030         struct nvmet_rdma_queue *queue =
1031                 container_of(w, struct nvmet_rdma_queue, release_work);
1032         struct rdma_cm_id *cm_id = queue->cm_id;
1033         struct nvmet_rdma_device *dev = queue->dev;
1034         enum nvmet_rdma_queue_state state = queue->state;
1035
1036         nvmet_rdma_free_queue(queue);
1037
1038         if (state != NVMET_RDMA_IN_DEVICE_REMOVAL)
1039                 rdma_destroy_id(cm_id);
1040
1041         kref_put(&dev->ref, nvmet_rdma_free_dev);
1042 }
1043
1044 static int
1045 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1046                                 struct nvmet_rdma_queue *queue)
1047 {
1048         struct nvme_rdma_cm_req *req;
1049
1050         req = (struct nvme_rdma_cm_req *)conn->private_data;
1051         if (!req || conn->private_data_len == 0)
1052                 return NVME_RDMA_CM_INVALID_LEN;
1053
1054         if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1055                 return NVME_RDMA_CM_INVALID_RECFMT;
1056
1057         queue->host_qid = le16_to_cpu(req->qid);
1058
1059         /*
1060          * req->hsqsize corresponds to our recv queue size plus 1
1061          * req->hrqsize corresponds to our send queue size
1062          */
1063         queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1064         queue->send_queue_size = le16_to_cpu(req->hrqsize);
1065
1066         if (!queue->host_qid && queue->recv_queue_size > NVMF_AQ_DEPTH)
1067                 return NVME_RDMA_CM_INVALID_HSQSIZE;
1068
1069         /* XXX: Should we enforce some kind of max for IO queues? */
1070
1071         return 0;
1072 }
1073
1074 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1075                                 enum nvme_rdma_cm_status status)
1076 {
1077         struct nvme_rdma_cm_rej rej;
1078
1079         rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1080         rej.sts = cpu_to_le16(status);
1081
1082         return rdma_reject(cm_id, (void *)&rej, sizeof(rej));
1083 }
1084
1085 static struct nvmet_rdma_queue *
1086 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1087                 struct rdma_cm_id *cm_id,
1088                 struct rdma_cm_event *event)
1089 {
1090         struct nvmet_rdma_queue *queue;
1091         int ret;
1092
1093         queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1094         if (!queue) {
1095                 ret = NVME_RDMA_CM_NO_RSC;
1096                 goto out_reject;
1097         }
1098
1099         ret = nvmet_sq_init(&queue->nvme_sq);
1100         if (ret)
1101                 goto out_free_queue;
1102
1103         ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1104         if (ret)
1105                 goto out_destroy_sq;
1106
1107         /*
1108          * Schedules the actual release because calling rdma_destroy_id from
1109          * inside a CM callback would trigger a deadlock. (great API design..)
1110          */
1111         INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1112         queue->dev = ndev;
1113         queue->cm_id = cm_id;
1114
1115         spin_lock_init(&queue->state_lock);
1116         queue->state = NVMET_RDMA_Q_CONNECTING;
1117         INIT_LIST_HEAD(&queue->rsp_wait_list);
1118         INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1119         spin_lock_init(&queue->rsp_wr_wait_lock);
1120         INIT_LIST_HEAD(&queue->free_rsps);
1121         spin_lock_init(&queue->rsps_lock);
1122         INIT_LIST_HEAD(&queue->queue_list);
1123
1124         queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1125         if (queue->idx < 0) {
1126                 ret = NVME_RDMA_CM_NO_RSC;
1127                 goto out_free_queue;
1128         }
1129
1130         ret = nvmet_rdma_alloc_rsps(queue);
1131         if (ret) {
1132                 ret = NVME_RDMA_CM_NO_RSC;
1133                 goto out_ida_remove;
1134         }
1135
1136         if (!ndev->srq) {
1137                 queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1138                                 queue->recv_queue_size,
1139                                 !queue->host_qid);
1140                 if (IS_ERR(queue->cmds)) {
1141                         ret = NVME_RDMA_CM_NO_RSC;
1142                         goto out_free_responses;
1143                 }
1144         }
1145
1146         ret = nvmet_rdma_create_queue_ib(queue);
1147         if (ret) {
1148                 pr_err("%s: creating RDMA queue failed (%d).\n",
1149                         __func__, ret);
1150                 ret = NVME_RDMA_CM_NO_RSC;
1151                 goto out_free_cmds;
1152         }
1153
1154         return queue;
1155
1156 out_free_cmds:
1157         if (!ndev->srq) {
1158                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1159                                 queue->recv_queue_size,
1160                                 !queue->host_qid);
1161         }
1162 out_free_responses:
1163         nvmet_rdma_free_rsps(queue);
1164 out_ida_remove:
1165         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1166 out_destroy_sq:
1167         nvmet_sq_destroy(&queue->nvme_sq);
1168 out_free_queue:
1169         kfree(queue);
1170 out_reject:
1171         nvmet_rdma_cm_reject(cm_id, ret);
1172         return NULL;
1173 }
1174
1175 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1176 {
1177         struct nvmet_rdma_queue *queue = priv;
1178
1179         switch (event->event) {
1180         case IB_EVENT_COMM_EST:
1181                 rdma_notify(queue->cm_id, event->event);
1182                 break;
1183         default:
1184                 pr_err("received unrecognized IB QP event %d\n", event->event);
1185                 break;
1186         }
1187 }
1188
1189 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1190                 struct nvmet_rdma_queue *queue,
1191                 struct rdma_conn_param *p)
1192 {
1193         struct rdma_conn_param  param = { };
1194         struct nvme_rdma_cm_rep priv = { };
1195         int ret = -ENOMEM;
1196
1197         param.rnr_retry_count = 7;
1198         param.flow_control = 1;
1199         param.initiator_depth = min_t(u8, p->initiator_depth,
1200                 queue->dev->device->attrs.max_qp_init_rd_atom);
1201         param.private_data = &priv;
1202         param.private_data_len = sizeof(priv);
1203         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1204         priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1205
1206         ret = rdma_accept(cm_id, &param);
1207         if (ret)
1208                 pr_err("rdma_accept failed (error code = %d)\n", ret);
1209
1210         return ret;
1211 }
1212
1213 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1214                 struct rdma_cm_event *event)
1215 {
1216         struct nvmet_rdma_device *ndev;
1217         struct nvmet_rdma_queue *queue;
1218         int ret = -EINVAL;
1219
1220         ndev = nvmet_rdma_find_get_device(cm_id);
1221         if (!ndev) {
1222                 pr_err("no client data!\n");
1223                 nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1224                 return -ECONNREFUSED;
1225         }
1226
1227         queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1228         if (!queue) {
1229                 ret = -ENOMEM;
1230                 goto put_device;
1231         }
1232         queue->port = cm_id->context;
1233
1234         ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1235         if (ret)
1236                 goto release_queue;
1237
1238         mutex_lock(&nvmet_rdma_queue_mutex);
1239         list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1240         mutex_unlock(&nvmet_rdma_queue_mutex);
1241
1242         return 0;
1243
1244 release_queue:
1245         nvmet_rdma_free_queue(queue);
1246 put_device:
1247         kref_put(&ndev->ref, nvmet_rdma_free_dev);
1248
1249         return ret;
1250 }
1251
1252 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1253 {
1254         unsigned long flags;
1255
1256         spin_lock_irqsave(&queue->state_lock, flags);
1257         if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1258                 pr_warn("trying to establish a connected queue\n");
1259                 goto out_unlock;
1260         }
1261         queue->state = NVMET_RDMA_Q_LIVE;
1262
1263         while (!list_empty(&queue->rsp_wait_list)) {
1264                 struct nvmet_rdma_rsp *cmd;
1265
1266                 cmd = list_first_entry(&queue->rsp_wait_list,
1267                                         struct nvmet_rdma_rsp, wait_list);
1268                 list_del(&cmd->wait_list);
1269
1270                 spin_unlock_irqrestore(&queue->state_lock, flags);
1271                 nvmet_rdma_handle_command(queue, cmd);
1272                 spin_lock_irqsave(&queue->state_lock, flags);
1273         }
1274
1275 out_unlock:
1276         spin_unlock_irqrestore(&queue->state_lock, flags);
1277 }
1278
1279 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1280 {
1281         bool disconnect = false;
1282         unsigned long flags;
1283
1284         pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1285
1286         spin_lock_irqsave(&queue->state_lock, flags);
1287         switch (queue->state) {
1288         case NVMET_RDMA_Q_CONNECTING:
1289         case NVMET_RDMA_Q_LIVE:
1290                 queue->state = NVMET_RDMA_Q_DISCONNECTING;
1291         case NVMET_RDMA_IN_DEVICE_REMOVAL:
1292                 disconnect = true;
1293                 break;
1294         case NVMET_RDMA_Q_DISCONNECTING:
1295                 break;
1296         }
1297         spin_unlock_irqrestore(&queue->state_lock, flags);
1298
1299         if (disconnect) {
1300                 rdma_disconnect(queue->cm_id);
1301                 schedule_work(&queue->release_work);
1302         }
1303 }
1304
1305 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1306 {
1307         bool disconnect = false;
1308
1309         mutex_lock(&nvmet_rdma_queue_mutex);
1310         if (!list_empty(&queue->queue_list)) {
1311                 list_del_init(&queue->queue_list);
1312                 disconnect = true;
1313         }
1314         mutex_unlock(&nvmet_rdma_queue_mutex);
1315
1316         if (disconnect)
1317                 __nvmet_rdma_queue_disconnect(queue);
1318 }
1319
1320 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1321                 struct nvmet_rdma_queue *queue)
1322 {
1323         WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1324
1325         mutex_lock(&nvmet_rdma_queue_mutex);
1326         if (!list_empty(&queue->queue_list))
1327                 list_del_init(&queue->queue_list);
1328         mutex_unlock(&nvmet_rdma_queue_mutex);
1329
1330         pr_err("failed to connect queue %d\n", queue->idx);
1331         schedule_work(&queue->release_work);
1332 }
1333
1334 /**
1335  * nvme_rdma_device_removal() - Handle RDMA device removal
1336  * @queue:      nvmet rdma queue (cm id qp_context)
1337  * @addr:       nvmet address (cm_id context)
1338  *
1339  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1340  * to unplug so we should take care of destroying our RDMA resources.
1341  * This event will be generated for each allocated cm_id.
1342  *
1343  * Note that this event can be generated on a normal queue cm_id
1344  * and/or a device bound listener cm_id (where in this case
1345  * queue will be null).
1346  *
1347  * we claim ownership on destroying the cm_id. For queues we move
1348  * the queue state to NVMET_RDMA_IN_DEVICE_REMOVAL and for port
1349  * we nullify the priv to prevent double cm_id destruction and destroying
1350  * the cm_id implicitely by returning a non-zero rc to the callout.
1351  */
1352 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1353                 struct nvmet_rdma_queue *queue)
1354 {
1355         unsigned long flags;
1356
1357         if (!queue) {
1358                 struct nvmet_port *port = cm_id->context;
1359
1360                 /*
1361                  * This is a listener cm_id. Make sure that
1362                  * future remove_port won't invoke a double
1363                  * cm_id destroy. use atomic xchg to make sure
1364                  * we don't compete with remove_port.
1365                  */
1366                 if (xchg(&port->priv, NULL) != cm_id)
1367                         return 0;
1368         } else {
1369                 /*
1370                  * This is a queue cm_id. Make sure that
1371                  * release queue will not destroy the cm_id
1372                  * and schedule all ctrl queues removal (only
1373                  * if the queue is not disconnecting already).
1374                  */
1375                 spin_lock_irqsave(&queue->state_lock, flags);
1376                 if (queue->state != NVMET_RDMA_Q_DISCONNECTING)
1377                         queue->state = NVMET_RDMA_IN_DEVICE_REMOVAL;
1378                 spin_unlock_irqrestore(&queue->state_lock, flags);
1379                 nvmet_rdma_queue_disconnect(queue);
1380                 flush_scheduled_work();
1381         }
1382
1383         /*
1384          * We need to return 1 so that the core will destroy
1385          * it's own ID.  What a great API design..
1386          */
1387         return 1;
1388 }
1389
1390 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1391                 struct rdma_cm_event *event)
1392 {
1393         struct nvmet_rdma_queue *queue = NULL;
1394         int ret = 0;
1395
1396         if (cm_id->qp)
1397                 queue = cm_id->qp->qp_context;
1398
1399         pr_debug("%s (%d): status %d id %p\n",
1400                 rdma_event_msg(event->event), event->event,
1401                 event->status, cm_id);
1402
1403         switch (event->event) {
1404         case RDMA_CM_EVENT_CONNECT_REQUEST:
1405                 ret = nvmet_rdma_queue_connect(cm_id, event);
1406                 break;
1407         case RDMA_CM_EVENT_ESTABLISHED:
1408                 nvmet_rdma_queue_established(queue);
1409                 break;
1410         case RDMA_CM_EVENT_ADDR_CHANGE:
1411         case RDMA_CM_EVENT_DISCONNECTED:
1412         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1413                 /*
1414                  * We might end up here when we already freed the qp
1415                  * which means queue release sequence is in progress,
1416                  * so don't get in the way...
1417                  */
1418                 if (queue)
1419                         nvmet_rdma_queue_disconnect(queue);
1420                 break;
1421         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1422                 ret = nvmet_rdma_device_removal(cm_id, queue);
1423                 break;
1424         case RDMA_CM_EVENT_REJECTED:
1425         case RDMA_CM_EVENT_UNREACHABLE:
1426         case RDMA_CM_EVENT_CONNECT_ERROR:
1427                 nvmet_rdma_queue_connect_fail(cm_id, queue);
1428                 break;
1429         default:
1430                 pr_err("received unrecognized RDMA CM event %d\n",
1431                         event->event);
1432                 break;
1433         }
1434
1435         return ret;
1436 }
1437
1438 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1439 {
1440         struct nvmet_rdma_queue *queue;
1441
1442 restart:
1443         mutex_lock(&nvmet_rdma_queue_mutex);
1444         list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1445                 if (queue->nvme_sq.ctrl == ctrl) {
1446                         list_del_init(&queue->queue_list);
1447                         mutex_unlock(&nvmet_rdma_queue_mutex);
1448
1449                         __nvmet_rdma_queue_disconnect(queue);
1450                         goto restart;
1451                 }
1452         }
1453         mutex_unlock(&nvmet_rdma_queue_mutex);
1454 }
1455
1456 static int nvmet_rdma_add_port(struct nvmet_port *port)
1457 {
1458         struct rdma_cm_id *cm_id;
1459         struct sockaddr_in addr_in;
1460         u16 port_in;
1461         int ret;
1462
1463         switch (port->disc_addr.adrfam) {
1464         case NVMF_ADDR_FAMILY_IP4:
1465                 break;
1466         default:
1467                 pr_err("address family %d not supported\n",
1468                                 port->disc_addr.adrfam);
1469                 return -EINVAL;
1470         }
1471
1472         ret = kstrtou16(port->disc_addr.trsvcid, 0, &port_in);
1473         if (ret)
1474                 return ret;
1475
1476         addr_in.sin_family = AF_INET;
1477         addr_in.sin_addr.s_addr = in_aton(port->disc_addr.traddr);
1478         addr_in.sin_port = htons(port_in);
1479
1480         cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1481                         RDMA_PS_TCP, IB_QPT_RC);
1482         if (IS_ERR(cm_id)) {
1483                 pr_err("CM ID creation failed\n");
1484                 return PTR_ERR(cm_id);
1485         }
1486
1487         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr_in);
1488         if (ret) {
1489                 pr_err("binding CM ID to %pISpc failed (%d)\n", &addr_in, ret);
1490                 goto out_destroy_id;
1491         }
1492
1493         ret = rdma_listen(cm_id, 128);
1494         if (ret) {
1495                 pr_err("listening to %pISpc failed (%d)\n", &addr_in, ret);
1496                 goto out_destroy_id;
1497         }
1498
1499         pr_info("enabling port %d (%pISpc)\n",
1500                 le16_to_cpu(port->disc_addr.portid), &addr_in);
1501         port->priv = cm_id;
1502         return 0;
1503
1504 out_destroy_id:
1505         rdma_destroy_id(cm_id);
1506         return ret;
1507 }
1508
1509 static void nvmet_rdma_remove_port(struct nvmet_port *port)
1510 {
1511         struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
1512
1513         if (cm_id)
1514                 rdma_destroy_id(cm_id);
1515 }
1516
1517 static struct nvmet_fabrics_ops nvmet_rdma_ops = {
1518         .owner                  = THIS_MODULE,
1519         .type                   = NVMF_TRTYPE_RDMA,
1520         .sqe_inline_size        = NVMET_RDMA_INLINE_DATA_SIZE,
1521         .msdbd                  = 1,
1522         .has_keyed_sgls         = 1,
1523         .add_port               = nvmet_rdma_add_port,
1524         .remove_port            = nvmet_rdma_remove_port,
1525         .queue_response         = nvmet_rdma_queue_response,
1526         .delete_ctrl            = nvmet_rdma_delete_ctrl,
1527 };
1528
1529 static int __init nvmet_rdma_init(void)
1530 {
1531         return nvmet_register_transport(&nvmet_rdma_ops);
1532 }
1533
1534 static void __exit nvmet_rdma_exit(void)
1535 {
1536         struct nvmet_rdma_queue *queue;
1537
1538         nvmet_unregister_transport(&nvmet_rdma_ops);
1539
1540         flush_scheduled_work();
1541
1542         mutex_lock(&nvmet_rdma_queue_mutex);
1543         while ((queue = list_first_entry_or_null(&nvmet_rdma_queue_list,
1544                         struct nvmet_rdma_queue, queue_list))) {
1545                 list_del_init(&queue->queue_list);
1546
1547                 mutex_unlock(&nvmet_rdma_queue_mutex);
1548                 __nvmet_rdma_queue_disconnect(queue);
1549                 mutex_lock(&nvmet_rdma_queue_mutex);
1550         }
1551         mutex_unlock(&nvmet_rdma_queue_mutex);
1552
1553         flush_scheduled_work();
1554         ida_destroy(&nvmet_rdma_queue_ida);
1555 }
1556
1557 module_init(nvmet_rdma_init);
1558 module_exit(nvmet_rdma_exit);
1559
1560 MODULE_LICENSE("GPL v2");
1561 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */