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 at least 1 page, up to 4 SGEs, and up to 16KB of inline data
  37  */
  38 #define NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE     PAGE_SIZE
  39 #define NVMET_RDMA_MAX_INLINE_SGE               4
  40 #define NVMET_RDMA_MAX_INLINE_DATA_SIZE         max_t(int, SZ_16K, PAGE_SIZE)
  41
  42 struct nvmet_rdma_cmd {
  43         struct ib_sge           sge[NVMET_RDMA_MAX_INLINE_SGE + 1];
  44         struct ib_cqe           cqe;
  45         struct ib_recv_wr       wr;
  46         struct scatterlist      inline_sg[NVMET_RDMA_MAX_INLINE_SGE];
  47         struct nvme_command     *nvme_cmd;
  48         struct nvmet_rdma_queue *queue;
  49 };
  50
  51 enum {
  52         NVMET_RDMA_REQ_INLINE_DATA      = (1 << 0),
  53         NVMET_RDMA_REQ_INVALIDATE_RKEY  = (1 << 1),
  54 };
  55
  56 struct nvmet_rdma_rsp {
  57         struct ib_sge           send_sge;
  58         struct ib_cqe           send_cqe;
  59         struct ib_send_wr       send_wr;
  60
  61         struct nvmet_rdma_cmd   *cmd;
  62         struct nvmet_rdma_queue *queue;
  63
  64         struct ib_cqe           read_cqe;
  65         struct rdma_rw_ctx      rw;
  66
  67         struct nvmet_req        req;
  68
  69         bool                    allocated;
  70         u8                      n_rdma;
  71         u32                     flags;
  72         u32                     invalidate_rkey;
  73
  74         struct list_head        wait_list;
  75         struct list_head        free_list;
  76 };
  77
  78 enum nvmet_rdma_queue_state {
  79         NVMET_RDMA_Q_CONNECTING,
  80         NVMET_RDMA_Q_LIVE,
  81         NVMET_RDMA_Q_DISCONNECTING,
  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         int                     inline_data_size;
 122         int                     inline_page_count;
 123 };
 124
 125 static bool nvmet_rdma_use_srq;
 126 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
 127 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
 128
 129 static DEFINE_IDA(nvmet_rdma_queue_ida);
 130 static LIST_HEAD(nvmet_rdma_queue_list);
 131 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
 132
 133 static LIST_HEAD(device_list);
 134 static DEFINE_MUTEX(device_list_mutex);
 135
 136 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
 137 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
 138 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 139 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
 140 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
 141 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
 142
 143 static const struct nvmet_fabrics_ops nvmet_rdma_ops;
 144
 145 static int num_pages(int len)
 146 {
 147         return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
 148 }
 149
 150 /* XXX: really should move to a generic header sooner or later.. */
 151 static inline u32 get_unaligned_le24(const u8 *p)
 152 {
 153         return (u32)p[0] | (u32)p[1] << 8 | (u32)p[2] << 16;
 154 }
 155
 156 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
 157 {
 158         return nvme_is_write(rsp->req.cmd) &&
 159                 rsp->req.transfer_len &&
 160                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 161 }
 162
 163 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
 164 {
 165         return !nvme_is_write(rsp->req.cmd) &&
 166                 rsp->req.transfer_len &&
 167                 !rsp->req.rsp->status &&
 168                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 169 }
 170
 171 static inline struct nvmet_rdma_rsp *
 172 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
 173 {
 174         struct nvmet_rdma_rsp *rsp;
 175         unsigned long flags;
 176
 177         spin_lock_irqsave(&queue->rsps_lock, flags);
 178         rsp = list_first_entry_or_null(&queue->free_rsps,
 179                                 struct nvmet_rdma_rsp, free_list);
 180         if (likely(rsp))
 181                 list_del(&rsp->free_list);
 182         spin_unlock_irqrestore(&queue->rsps_lock, flags);
 183
 184         if (unlikely(!rsp)) {
 185                 rsp = kmalloc(sizeof(*rsp), GFP_KERNEL);
 186                 if (unlikely(!rsp))
 187                         return NULL;
 188                 rsp->allocated = true;
 189         }
 190
 191         return rsp;
 192 }
 193
 194 static inline void
 195 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
 196 {
 197         unsigned long flags;
 198
 199         if (unlikely(rsp->allocated)) {
 200                 kfree(rsp);
 201                 return;
 202         }
 203
 204         spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
 205         list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
 206         spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
 207 }
 208
 209 static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
 210                                 struct nvmet_rdma_cmd *c)
 211 {
 212         struct scatterlist *sg;
 213         struct ib_sge *sge;
 214         int i;
 215
 216         if (!ndev->inline_data_size)
 217                 return;
 218
 219         sg = c->inline_sg;
 220         sge = &c->sge[1];
 221
 222         for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
 223                 if (sge->length)
 224                         ib_dma_unmap_page(ndev->device, sge->addr,
 225                                         sge->length, DMA_FROM_DEVICE);
 226                 if (sg_page(sg))
 227                         __free_page(sg_page(sg));
 228         }
 229 }
 230
 231 static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
 232                                 struct nvmet_rdma_cmd *c)
 233 {
 234         struct scatterlist *sg;
 235         struct ib_sge *sge;
 236         struct page *pg;
 237         int len;
 238         int i;
 239
 240         if (!ndev->inline_data_size)
 241                 return 0;
 242
 243         sg = c->inline_sg;
 244         sg_init_table(sg, ndev->inline_page_count);
 245         sge = &c->sge[1];
 246         len = ndev->inline_data_size;
 247
 248         for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
 249                 pg = alloc_page(GFP_KERNEL);
 250                 if (!pg)
 251                         goto out_err;
 252                 sg_assign_page(sg, pg);
 253                 sge->addr = ib_dma_map_page(ndev->device,
 254                         pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
 255                 if (ib_dma_mapping_error(ndev->device, sge->addr))
 256                         goto out_err;
 257                 sge->length = min_t(int, len, PAGE_SIZE);
 258                 sge->lkey = ndev->pd->local_dma_lkey;
 259                 len -= sge->length;
 260         }
 261
 262         return 0;
 263 out_err:
 264         for (; i >= 0; i--, sg--, sge--) {
 265                 if (sge->length)
 266                         ib_dma_unmap_page(ndev->device, sge->addr,
 267                                         sge->length, DMA_FROM_DEVICE);
 268                 if (sg_page(sg))
 269                         __free_page(sg_page(sg));
 270         }
 271         return -ENOMEM;
 272 }
 273
 274 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
 275                         struct nvmet_rdma_cmd *c, bool admin)
 276 {
 277         /* NVMe command / RDMA RECV */
 278         c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
 279         if (!c->nvme_cmd)
 280                 goto out;
 281
 282         c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
 283                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 284         if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
 285                 goto out_free_cmd;
 286
 287         c->sge[0].length = sizeof(*c->nvme_cmd);
 288         c->sge[0].lkey = ndev->pd->local_dma_lkey;
 289
 290         if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
 291                 goto out_unmap_cmd;
 292
 293         c->cqe.done = nvmet_rdma_recv_done;
 294
 295         c->wr.wr_cqe = &c->cqe;
 296         c->wr.sg_list = c->sge;
 297         c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;
 298
 299         return 0;
 300
 301 out_unmap_cmd:
 302         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 303                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 304 out_free_cmd:
 305         kfree(c->nvme_cmd);
 306
 307 out:
 308         return -ENOMEM;
 309 }
 310
 311 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
 312                 struct nvmet_rdma_cmd *c, bool admin)
 313 {
 314         if (!admin)
 315                 nvmet_rdma_free_inline_pages(ndev, c);
 316         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 317                                 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 318         kfree(c->nvme_cmd);
 319 }
 320
 321 static struct nvmet_rdma_cmd *
 322 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
 323                 int nr_cmds, bool admin)
 324 {
 325         struct nvmet_rdma_cmd *cmds;
 326         int ret = -EINVAL, i;
 327
 328         cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
 329         if (!cmds)
 330                 goto out;
 331
 332         for (i = 0; i < nr_cmds; i++) {
 333                 ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
 334                 if (ret)
 335                         goto out_free;
 336         }
 337
 338         return cmds;
 339
 340 out_free:
 341         while (--i >= 0)
 342                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 343         kfree(cmds);
 344 out:
 345         return ERR_PTR(ret);
 346 }
 347
 348 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
 349                 struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
 350 {
 351         int i;
 352
 353         for (i = 0; i < nr_cmds; i++)
 354                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 355         kfree(cmds);
 356 }
 357
 358 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
 359                 struct nvmet_rdma_rsp *r)
 360 {
 361         /* NVMe CQE / RDMA SEND */
 362         r->req.rsp = kmalloc(sizeof(*r->req.rsp), GFP_KERNEL);
 363         if (!r->req.rsp)
 364                 goto out;
 365
 366         r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.rsp,
 367                         sizeof(*r->req.rsp), DMA_TO_DEVICE);
 368         if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
 369                 goto out_free_rsp;
 370
 371         r->send_sge.length = sizeof(*r->req.rsp);
 372         r->send_sge.lkey = ndev->pd->local_dma_lkey;
 373
 374         r->send_cqe.done = nvmet_rdma_send_done;
 375
 376         r->send_wr.wr_cqe = &r->send_cqe;
 377         r->send_wr.sg_list = &r->send_sge;
 378         r->send_wr.num_sge = 1;
 379         r->send_wr.send_flags = IB_SEND_SIGNALED;
 380
 381         /* Data In / RDMA READ */
 382         r->read_cqe.done = nvmet_rdma_read_data_done;
 383         return 0;
 384
 385 out_free_rsp:
 386         kfree(r->req.rsp);
 387 out:
 388         return -ENOMEM;
 389 }
 390
 391 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
 392                 struct nvmet_rdma_rsp *r)
 393 {
 394         ib_dma_unmap_single(ndev->device, r->send_sge.addr,
 395                                 sizeof(*r->req.rsp), DMA_TO_DEVICE);
 396         kfree(r->req.rsp);
 397 }
 398
 399 static int
 400 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
 401 {
 402         struct nvmet_rdma_device *ndev = queue->dev;
 403         int nr_rsps = queue->recv_queue_size * 2;
 404         int ret = -EINVAL, i;
 405
 406         queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
 407                         GFP_KERNEL);
 408         if (!queue->rsps)
 409                 goto out;
 410
 411         for (i = 0; i < nr_rsps; i++) {
 412                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 413
 414                 ret = nvmet_rdma_alloc_rsp(ndev, rsp);
 415                 if (ret)
 416                         goto out_free;
 417
 418                 list_add_tail(&rsp->free_list, &queue->free_rsps);
 419         }
 420
 421         return 0;
 422
 423 out_free:
 424         while (--i >= 0) {
 425                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 426
 427                 list_del(&rsp->free_list);
 428                 nvmet_rdma_free_rsp(ndev, rsp);
 429         }
 430         kfree(queue->rsps);
 431 out:
 432         return ret;
 433 }
 434
 435 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
 436 {
 437         struct nvmet_rdma_device *ndev = queue->dev;
 438         int i, nr_rsps = queue->recv_queue_size * 2;
 439
 440         for (i = 0; i < nr_rsps; i++) {
 441                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 442
 443                 list_del(&rsp->free_list);
 444                 nvmet_rdma_free_rsp(ndev, rsp);
 445         }
 446         kfree(queue->rsps);
 447 }
 448
 449 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
 450                 struct nvmet_rdma_cmd *cmd)
 451 {
 452         int ret;
 453
 454         ib_dma_sync_single_for_device(ndev->device,
 455                 cmd->sge[0].addr, cmd->sge[0].length,
 456                 DMA_FROM_DEVICE);
 457
 458         if (ndev->srq)
 459                 ret = ib_post_srq_recv(ndev->srq, &cmd->wr, NULL);
 460         else
 461                 ret = ib_post_recv(cmd->queue->cm_id->qp, &cmd->wr, NULL);
 462
 463         if (unlikely(ret))
 464                 pr_err("post_recv cmd failed\n");
 465
 466         return ret;
 467 }
 468
 469 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
 470 {
 471         spin_lock(&queue->rsp_wr_wait_lock);
 472         while (!list_empty(&queue->rsp_wr_wait_list)) {
 473                 struct nvmet_rdma_rsp *rsp;
 474                 bool ret;
 475
 476                 rsp = list_entry(queue->rsp_wr_wait_list.next,
 477                                 struct nvmet_rdma_rsp, wait_list);
 478                 list_del(&rsp->wait_list);
 479
 480                 spin_unlock(&queue->rsp_wr_wait_lock);
 481                 ret = nvmet_rdma_execute_command(rsp);
 482                 spin_lock(&queue->rsp_wr_wait_lock);
 483
 484                 if (!ret) {
 485                         list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
 486                         break;
 487                 }
 488         }
 489         spin_unlock(&queue->rsp_wr_wait_lock);
 490 }
 491
 492
 493 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
 494 {
 495         struct nvmet_rdma_queue *queue = rsp->queue;
 496
 497         atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 498
 499         if (rsp->n_rdma) {
 500                 rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
 501                                 queue->cm_id->port_num, rsp->req.sg,
 502                                 rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
 503         }
 504
 505         if (rsp->req.sg != rsp->cmd->inline_sg)
 506                 nvmet_req_free_sgl(&rsp->req);
 507
 508         if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
 509                 nvmet_rdma_process_wr_wait_list(queue);
 510
 511         nvmet_rdma_put_rsp(rsp);
 512 }
 513
 514 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
 515 {
 516         if (queue->nvme_sq.ctrl) {
 517                 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
 518         } else {
 519                 /*
 520                  * we didn't setup the controller yet in case
 521                  * of admin connect error, just disconnect and
 522                  * cleanup the queue
 523                  */
 524                 nvmet_rdma_queue_disconnect(queue);
 525         }
 526 }
 527
 528 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
 529 {
 530         struct nvmet_rdma_rsp *rsp =
 531                 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
 532         struct nvmet_rdma_queue *queue = cq->cq_context;
 533
 534         nvmet_rdma_release_rsp(rsp);
 535
 536         if (unlikely(wc->status != IB_WC_SUCCESS &&
 537                      wc->status != IB_WC_WR_FLUSH_ERR)) {
 538                 pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
 539                         wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
 540                 nvmet_rdma_error_comp(queue);
 541         }
 542 }
 543
 544 static void nvmet_rdma_queue_response(struct nvmet_req *req)
 545 {
 546         struct nvmet_rdma_rsp *rsp =
 547                 container_of(req, struct nvmet_rdma_rsp, req);
 548         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 549         struct ib_send_wr *first_wr;
 550
 551         if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
 552                 rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
 553                 rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
 554         } else {
 555                 rsp->send_wr.opcode = IB_WR_SEND;
 556         }
 557
 558         if (nvmet_rdma_need_data_out(rsp))
 559                 first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
 560                                 cm_id->port_num, NULL, &rsp->send_wr);
 561         else
 562                 first_wr = &rsp->send_wr;
 563
 564         nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
 565
 566         ib_dma_sync_single_for_device(rsp->queue->dev->device,
 567                 rsp->send_sge.addr, rsp->send_sge.length,
 568                 DMA_TO_DEVICE);
 569
 570         if (unlikely(ib_post_send(cm_id->qp, first_wr, NULL))) {
 571                 pr_err("sending cmd response failed\n");
 572                 nvmet_rdma_release_rsp(rsp);
 573         }
 574 }
 575
 576 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
 577 {
 578         struct nvmet_rdma_rsp *rsp =
 579                 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
 580         struct nvmet_rdma_queue *queue = cq->cq_context;
 581
 582         WARN_ON(rsp->n_rdma <= 0);
 583         atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
 584         rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
 585                         queue->cm_id->port_num, rsp->req.sg,
 586                         rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
 587         rsp->n_rdma = 0;
 588
 589         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 590                 nvmet_req_uninit(&rsp->req);
 591                 nvmet_rdma_release_rsp(rsp);
 592                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 593                         pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
 594                                 wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
 595                         nvmet_rdma_error_comp(queue);
 596                 }
 597                 return;
 598         }
 599
 600         nvmet_req_execute(&rsp->req);
 601 }
 602
 603 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
 604                 u64 off)
 605 {
 606         int sg_count = num_pages(len);
 607         struct scatterlist *sg;
 608         int i;
 609
 610         sg = rsp->cmd->inline_sg;
 611         for (i = 0; i < sg_count; i++, sg++) {
 612                 if (i < sg_count - 1)
 613                         sg_unmark_end(sg);
 614                 else
 615                         sg_mark_end(sg);
 616                 sg->offset = off;
 617                 sg->length = min_t(int, len, PAGE_SIZE - off);
 618                 len -= sg->length;
 619                 if (!i)
 620                         off = 0;
 621         }
 622
 623         rsp->req.sg = rsp->cmd->inline_sg;
 624         rsp->req.sg_cnt = sg_count;
 625 }
 626
 627 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
 628 {
 629         struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
 630         u64 off = le64_to_cpu(sgl->addr);
 631         u32 len = le32_to_cpu(sgl->length);
 632
 633         if (!nvme_is_write(rsp->req.cmd)) {
 634                 rsp->req.error_loc =
 635                         offsetof(struct nvme_common_command, opcode);
 636                 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 637         }
 638
 639         if (off + len > rsp->queue->dev->inline_data_size) {
 640                 pr_err("invalid inline data offset!\n");
 641                 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
 642         }
 643
 644         /* no data command? */
 645         if (!len)
 646                 return 0;
 647
 648         nvmet_rdma_use_inline_sg(rsp, len, off);
 649         rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
 650         rsp->req.transfer_len += len;
 651         return 0;
 652 }
 653
 654 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
 655                 struct nvme_keyed_sgl_desc *sgl, bool invalidate)
 656 {
 657         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 658         u64 addr = le64_to_cpu(sgl->addr);
 659         u32 key = get_unaligned_le32(sgl->key);
 660         int ret;
 661
 662         rsp->req.transfer_len = get_unaligned_le24(sgl->length);
 663
 664         /* no data command? */
 665         if (!rsp->req.transfer_len)
 666                 return 0;
 667
 668         ret = nvmet_req_alloc_sgl(&rsp->req);
 669         if (ret < 0)
 670                 goto error_out;
 671
 672         ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
 673                         rsp->req.sg, rsp->req.sg_cnt, 0, addr, key,
 674                         nvmet_data_dir(&rsp->req));
 675         if (ret < 0)
 676                 goto error_out;
 677         rsp->n_rdma += ret;
 678
 679         if (invalidate) {
 680                 rsp->invalidate_rkey = key;
 681                 rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
 682         }
 683
 684         return 0;
 685
 686 error_out:
 687         rsp->req.transfer_len = 0;
 688         return NVME_SC_INTERNAL;
 689 }
 690
 691 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
 692 {
 693         struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
 694
 695         switch (sgl->type >> 4) {
 696         case NVME_SGL_FMT_DATA_DESC:
 697                 switch (sgl->type & 0xf) {
 698                 case NVME_SGL_FMT_OFFSET:
 699                         return nvmet_rdma_map_sgl_inline(rsp);
 700                 default:
 701                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 702                         rsp->req.error_loc =
 703                                 offsetof(struct nvme_common_command, dptr);
 704                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 705                 }
 706         case NVME_KEY_SGL_FMT_DATA_DESC:
 707                 switch (sgl->type & 0xf) {
 708                 case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
 709                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
 710                 case NVME_SGL_FMT_ADDRESS:
 711                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
 712                 default:
 713                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 714                         rsp->req.error_loc =
 715                                 offsetof(struct nvme_common_command, dptr);
 716                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 717                 }
 718         default:
 719                 pr_err("invalid SGL type: %#x\n", sgl->type);
 720                 rsp->req.error_loc = offsetof(struct nvme_common_command, dptr);
 721                 return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
 722         }
 723 }
 724
 725 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
 726 {
 727         struct nvmet_rdma_queue *queue = rsp->queue;
 728
 729         if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
 730                         &queue->sq_wr_avail) < 0)) {
 731                 pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
 732                                 1 + rsp->n_rdma, queue->idx,
 733                                 queue->nvme_sq.ctrl->cntlid);
 734                 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 735                 return false;
 736         }
 737
 738         if (nvmet_rdma_need_data_in(rsp)) {
 739                 if (rdma_rw_ctx_post(&rsp->rw, queue->cm_id->qp,
 740                                 queue->cm_id->port_num, &rsp->read_cqe, NULL))
 741                         nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
 742         } else {
 743                 nvmet_req_execute(&rsp->req);
 744         }
 745
 746         return true;
 747 }
 748
 749 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
 750                 struct nvmet_rdma_rsp *cmd)
 751 {
 752         u16 status;
 753
 754         ib_dma_sync_single_for_cpu(queue->dev->device,
 755                 cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
 756                 DMA_FROM_DEVICE);
 757         ib_dma_sync_single_for_cpu(queue->dev->device,
 758                 cmd->send_sge.addr, cmd->send_sge.length,
 759                 DMA_TO_DEVICE);
 760
 761         cmd->req.p2p_client = &queue->dev->device->dev;
 762
 763         if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
 764                         &queue->nvme_sq, &nvmet_rdma_ops))
 765                 return;
 766
 767         status = nvmet_rdma_map_sgl(cmd);
 768         if (status)
 769                 goto out_err;
 770
 771         if (unlikely(!nvmet_rdma_execute_command(cmd))) {
 772                 spin_lock(&queue->rsp_wr_wait_lock);
 773                 list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
 774                 spin_unlock(&queue->rsp_wr_wait_lock);
 775         }
 776
 777         return;
 778
 779 out_err:
 780         nvmet_req_complete(&cmd->req, status);
 781 }
 782
 783 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
 784 {
 785         struct nvmet_rdma_cmd *cmd =
 786                 container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
 787         struct nvmet_rdma_queue *queue = cq->cq_context;
 788         struct nvmet_rdma_rsp *rsp;
 789
 790         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 791                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 792                         pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
 793                                 wc->wr_cqe, ib_wc_status_msg(wc->status),
 794                                 wc->status);
 795                         nvmet_rdma_error_comp(queue);
 796                 }
 797                 return;
 798         }
 799
 800         if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
 801                 pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
 802                 nvmet_rdma_error_comp(queue);
 803                 return;
 804         }
 805
 806         cmd->queue = queue;
 807         rsp = nvmet_rdma_get_rsp(queue);
 808         if (unlikely(!rsp)) {
 809                 /*
 810                  * we get here only under memory pressure,
 811                  * silently drop and have the host retry
 812                  * as we can't even fail it.
 813                  */
 814                 nvmet_rdma_post_recv(queue->dev, cmd);
 815                 return;
 816         }
 817         rsp->queue = queue;
 818         rsp->cmd = cmd;
 819         rsp->flags = 0;
 820         rsp->req.cmd = cmd->nvme_cmd;
 821         rsp->req.port = queue->port;
 822         rsp->n_rdma = 0;
 823
 824         if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
 825                 unsigned long flags;
 826
 827                 spin_lock_irqsave(&queue->state_lock, flags);
 828                 if (queue->state == NVMET_RDMA_Q_CONNECTING)
 829                         list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
 830                 else
 831                         nvmet_rdma_put_rsp(rsp);
 832                 spin_unlock_irqrestore(&queue->state_lock, flags);
 833                 return;
 834         }
 835
 836         nvmet_rdma_handle_command(queue, rsp);
 837 }
 838
 839 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_device *ndev)
 840 {
 841         if (!ndev->srq)
 842                 return;
 843
 844         nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
 845         ib_destroy_srq(ndev->srq);
 846 }
 847
 848 static int nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
 849 {
 850         struct ib_srq_init_attr srq_attr = { NULL, };
 851         struct ib_srq *srq;
 852         size_t srq_size;
 853         int ret, i;
 854
 855         srq_size = 4095;        /* XXX: tune */
 856
 857         srq_attr.attr.max_wr = srq_size;
 858         srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
 859         srq_attr.attr.srq_limit = 0;
 860         srq_attr.srq_type = IB_SRQT_BASIC;
 861         srq = ib_create_srq(ndev->pd, &srq_attr);
 862         if (IS_ERR(srq)) {
 863                 /*
 864                  * If SRQs aren't supported we just go ahead and use normal
 865                  * non-shared receive queues.
 866                  */
 867                 pr_info("SRQ requested but not supported.\n");
 868                 return 0;
 869         }
 870
 871         ndev->srq_cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
 872         if (IS_ERR(ndev->srq_cmds)) {
 873                 ret = PTR_ERR(ndev->srq_cmds);
 874                 goto out_destroy_srq;
 875         }
 876
 877         ndev->srq = srq;
 878         ndev->srq_size = srq_size;
 879
 880         for (i = 0; i < srq_size; i++) {
 881                 ret = nvmet_rdma_post_recv(ndev, &ndev->srq_cmds[i]);
 882                 if (ret)
 883                         goto out_free_cmds;
 884         }
 885
 886         return 0;
 887
 888 out_free_cmds:
 889         nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
 890 out_destroy_srq:
 891         ib_destroy_srq(srq);
 892         return ret;
 893 }
 894
 895 static void nvmet_rdma_free_dev(struct kref *ref)
 896 {
 897         struct nvmet_rdma_device *ndev =
 898                 container_of(ref, struct nvmet_rdma_device, ref);
 899
 900         mutex_lock(&device_list_mutex);
 901         list_del(&ndev->entry);
 902         mutex_unlock(&device_list_mutex);
 903
 904         nvmet_rdma_destroy_srq(ndev);
 905         ib_dealloc_pd(ndev->pd);
 906
 907         kfree(ndev);
 908 }
 909
 910 static struct nvmet_rdma_device *
 911 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
 912 {
 913         struct nvmet_port *port = cm_id->context;
 914         struct nvmet_rdma_device *ndev;
 915         int inline_page_count;
 916         int inline_sge_count;
 917         int ret;
 918
 919         mutex_lock(&device_list_mutex);
 920         list_for_each_entry(ndev, &device_list, entry) {
 921                 if (ndev->device->node_guid == cm_id->device->node_guid &&
 922                     kref_get_unless_zero(&ndev->ref))
 923                         goto out_unlock;
 924         }
 925
 926         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 927         if (!ndev)
 928                 goto out_err;
 929
 930         inline_page_count = num_pages(port->inline_data_size);
 931         inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
 932                                 cm_id->device->attrs.max_recv_sge) - 1;
 933         if (inline_page_count > inline_sge_count) {
 934                 pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
 935                         port->inline_data_size, cm_id->device->name,
 936                         inline_sge_count * PAGE_SIZE);
 937                 port->inline_data_size = inline_sge_count * PAGE_SIZE;
 938                 inline_page_count = inline_sge_count;
 939         }
 940         ndev->inline_data_size = port->inline_data_size;
 941         ndev->inline_page_count = inline_page_count;
 942         ndev->device = cm_id->device;
 943         kref_init(&ndev->ref);
 944
 945         ndev->pd = ib_alloc_pd(ndev->device, 0);
 946         if (IS_ERR(ndev->pd))
 947                 goto out_free_dev;
 948
 949         if (nvmet_rdma_use_srq) {
 950                 ret = nvmet_rdma_init_srq(ndev);
 951                 if (ret)
 952                         goto out_free_pd;
 953         }
 954
 955         list_add(&ndev->entry, &device_list);
 956 out_unlock:
 957         mutex_unlock(&device_list_mutex);
 958         pr_debug("added %s.\n", ndev->device->name);
 959         return ndev;
 960
 961 out_free_pd:
 962         ib_dealloc_pd(ndev->pd);
 963 out_free_dev:
 964         kfree(ndev);
 965 out_err:
 966         mutex_unlock(&device_list_mutex);
 967         return NULL;
 968 }
 969
 970 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
 971 {
 972         struct ib_qp_init_attr qp_attr;
 973         struct nvmet_rdma_device *ndev = queue->dev;
 974         int comp_vector, nr_cqe, ret, i;
 975
 976         /*
 977          * Spread the io queues across completion vectors,
 978          * but still keep all admin queues on vector 0.
 979          */
 980         comp_vector = !queue->host_qid ? 0 :
 981                 queue->idx % ndev->device->num_comp_vectors;
 982
 983         /*
 984          * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
 985          */
 986         nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
 987
 988         queue->cq = ib_alloc_cq(ndev->device, queue,
 989                         nr_cqe + 1, comp_vector,
 990                         IB_POLL_WORKQUEUE);
 991         if (IS_ERR(queue->cq)) {
 992                 ret = PTR_ERR(queue->cq);
 993                 pr_err("failed to create CQ cqe= %d ret= %d\n",
 994                        nr_cqe + 1, ret);
 995                 goto out;
 996         }
 997
 998         memset(&qp_attr, 0, sizeof(qp_attr));
 999         qp_attr.qp_context = queue;
1000         qp_attr.event_handler = nvmet_rdma_qp_event;
1001         qp_attr.send_cq = queue->cq;
1002         qp_attr.recv_cq = queue->cq;
1003         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1004         qp_attr.qp_type = IB_QPT_RC;
1005         /* +1 for drain */
1006         qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
1007         qp_attr.cap.max_rdma_ctxs = queue->send_queue_size;
1008         qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
1009                                         ndev->device->attrs.max_send_sge);
1010
1011         if (ndev->srq) {
1012                 qp_attr.srq = ndev->srq;
1013         } else {
1014                 /* +1 for drain */
1015                 qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
1016                 qp_attr.cap.max_recv_sge = 1 + ndev->inline_page_count;
1017         }
1018
1019         ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
1020         if (ret) {
1021                 pr_err("failed to create_qp ret= %d\n", ret);
1022                 goto err_destroy_cq;
1023         }
1024
1025         atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
1026
1027         pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
1028                  __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
1029                  qp_attr.cap.max_send_wr, queue->cm_id);
1030
1031         if (!ndev->srq) {
1032                 for (i = 0; i < queue->recv_queue_size; i++) {
1033                         queue->cmds[i].queue = queue;
1034                         ret = nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
1035                         if (ret)
1036                                 goto err_destroy_qp;
1037                 }
1038         }
1039
1040 out:
1041         return ret;
1042
1043 err_destroy_qp:
1044         rdma_destroy_qp(queue->cm_id);
1045 err_destroy_cq:
1046         ib_free_cq(queue->cq);
1047         goto out;
1048 }
1049
1050 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1051 {
1052         struct ib_qp *qp = queue->cm_id->qp;
1053
1054         ib_drain_qp(qp);
1055         rdma_destroy_id(queue->cm_id);
1056         ib_destroy_qp(qp);
1057         ib_free_cq(queue->cq);
1058 }
1059
1060 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1061 {
1062         pr_debug("freeing queue %d\n", queue->idx);
1063
1064         nvmet_sq_destroy(&queue->nvme_sq);
1065
1066         nvmet_rdma_destroy_queue_ib(queue);
1067         if (!queue->dev->srq) {
1068                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1069                                 queue->recv_queue_size,
1070                                 !queue->host_qid);
1071         }
1072         nvmet_rdma_free_rsps(queue);
1073         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1074         kfree(queue);
1075 }
1076
1077 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1078 {
1079         struct nvmet_rdma_queue *queue =
1080                 container_of(w, struct nvmet_rdma_queue, release_work);
1081         struct nvmet_rdma_device *dev = queue->dev;
1082
1083         nvmet_rdma_free_queue(queue);
1084
1085         kref_put(&dev->ref, nvmet_rdma_free_dev);
1086 }
1087
1088 static int
1089 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1090                                 struct nvmet_rdma_queue *queue)
1091 {
1092         struct nvme_rdma_cm_req *req;
1093
1094         req = (struct nvme_rdma_cm_req *)conn->private_data;
1095         if (!req || conn->private_data_len == 0)
1096                 return NVME_RDMA_CM_INVALID_LEN;
1097
1098         if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1099                 return NVME_RDMA_CM_INVALID_RECFMT;
1100
1101         queue->host_qid = le16_to_cpu(req->qid);
1102
1103         /*
1104          * req->hsqsize corresponds to our recv queue size plus 1
1105          * req->hrqsize corresponds to our send queue size
1106          */
1107         queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1108         queue->send_queue_size = le16_to_cpu(req->hrqsize);
1109
1110         if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
1111                 return NVME_RDMA_CM_INVALID_HSQSIZE;
1112
1113         /* XXX: Should we enforce some kind of max for IO queues? */
1114
1115         return 0;
1116 }
1117
1118 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1119                                 enum nvme_rdma_cm_status status)
1120 {
1121         struct nvme_rdma_cm_rej rej;
1122
1123         pr_debug("rejecting connect request: status %d (%s)\n",
1124                  status, nvme_rdma_cm_msg(status));
1125
1126         rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1127         rej.sts = cpu_to_le16(status);
1128
1129         return rdma_reject(cm_id, (void *)&rej, sizeof(rej));
1130 }
1131
1132 static struct nvmet_rdma_queue *
1133 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1134                 struct rdma_cm_id *cm_id,
1135                 struct rdma_cm_event *event)
1136 {
1137         struct nvmet_rdma_queue *queue;
1138         int ret;
1139
1140         queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1141         if (!queue) {
1142                 ret = NVME_RDMA_CM_NO_RSC;
1143                 goto out_reject;
1144         }
1145
1146         ret = nvmet_sq_init(&queue->nvme_sq);
1147         if (ret) {
1148                 ret = NVME_RDMA_CM_NO_RSC;
1149                 goto out_free_queue;
1150         }
1151
1152         ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1153         if (ret)
1154                 goto out_destroy_sq;
1155
1156         /*
1157          * Schedules the actual release because calling rdma_destroy_id from
1158          * inside a CM callback would trigger a deadlock. (great API design..)
1159          */
1160         INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1161         queue->dev = ndev;
1162         queue->cm_id = cm_id;
1163
1164         spin_lock_init(&queue->state_lock);
1165         queue->state = NVMET_RDMA_Q_CONNECTING;
1166         INIT_LIST_HEAD(&queue->rsp_wait_list);
1167         INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1168         spin_lock_init(&queue->rsp_wr_wait_lock);
1169         INIT_LIST_HEAD(&queue->free_rsps);
1170         spin_lock_init(&queue->rsps_lock);
1171         INIT_LIST_HEAD(&queue->queue_list);
1172
1173         queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1174         if (queue->idx < 0) {
1175                 ret = NVME_RDMA_CM_NO_RSC;
1176                 goto out_destroy_sq;
1177         }
1178
1179         ret = nvmet_rdma_alloc_rsps(queue);
1180         if (ret) {
1181                 ret = NVME_RDMA_CM_NO_RSC;
1182                 goto out_ida_remove;
1183         }
1184
1185         if (!ndev->srq) {
1186                 queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1187                                 queue->recv_queue_size,
1188                                 !queue->host_qid);
1189                 if (IS_ERR(queue->cmds)) {
1190                         ret = NVME_RDMA_CM_NO_RSC;
1191                         goto out_free_responses;
1192                 }
1193         }
1194
1195         ret = nvmet_rdma_create_queue_ib(queue);
1196         if (ret) {
1197                 pr_err("%s: creating RDMA queue failed (%d).\n",
1198                         __func__, ret);
1199                 ret = NVME_RDMA_CM_NO_RSC;
1200                 goto out_free_cmds;
1201         }
1202
1203         return queue;
1204
1205 out_free_cmds:
1206         if (!ndev->srq) {
1207                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1208                                 queue->recv_queue_size,
1209                                 !queue->host_qid);
1210         }
1211 out_free_responses:
1212         nvmet_rdma_free_rsps(queue);
1213 out_ida_remove:
1214         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1215 out_destroy_sq:
1216         nvmet_sq_destroy(&queue->nvme_sq);
1217 out_free_queue:
1218         kfree(queue);
1219 out_reject:
1220         nvmet_rdma_cm_reject(cm_id, ret);
1221         return NULL;
1222 }
1223
1224 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1225 {
1226         struct nvmet_rdma_queue *queue = priv;
1227
1228         switch (event->event) {
1229         case IB_EVENT_COMM_EST:
1230                 rdma_notify(queue->cm_id, event->event);
1231                 break;
1232         default:
1233                 pr_err("received IB QP event: %s (%d)\n",
1234                        ib_event_msg(event->event), event->event);
1235                 break;
1236         }
1237 }
1238
1239 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1240                 struct nvmet_rdma_queue *queue,
1241                 struct rdma_conn_param *p)
1242 {
1243         struct rdma_conn_param  param = { };
1244         struct nvme_rdma_cm_rep priv = { };
1245         int ret = -ENOMEM;
1246
1247         param.rnr_retry_count = 7;
1248         param.flow_control = 1;
1249         param.initiator_depth = min_t(u8, p->initiator_depth,
1250                 queue->dev->device->attrs.max_qp_init_rd_atom);
1251         param.private_data = &priv;
1252         param.private_data_len = sizeof(priv);
1253         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1254         priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1255
1256         ret = rdma_accept(cm_id, &param);
1257         if (ret)
1258                 pr_err("rdma_accept failed (error code = %d)\n", ret);
1259
1260         return ret;
1261 }
1262
1263 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1264                 struct rdma_cm_event *event)
1265 {
1266         struct nvmet_rdma_device *ndev;
1267         struct nvmet_rdma_queue *queue;
1268         int ret = -EINVAL;
1269
1270         ndev = nvmet_rdma_find_get_device(cm_id);
1271         if (!ndev) {
1272                 nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1273                 return -ECONNREFUSED;
1274         }
1275
1276         queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1277         if (!queue) {
1278                 ret = -ENOMEM;
1279                 goto put_device;
1280         }
1281         queue->port = cm_id->context;
1282
1283         if (queue->host_qid == 0) {
1284                 /* Let inflight controller teardown complete */
1285                 flush_scheduled_work();
1286         }
1287
1288         ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1289         if (ret) {
1290                 schedule_work(&queue->release_work);
1291                 /* Destroying rdma_cm id is not needed here */
1292                 return 0;
1293         }
1294
1295         mutex_lock(&nvmet_rdma_queue_mutex);
1296         list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1297         mutex_unlock(&nvmet_rdma_queue_mutex);
1298
1299         return 0;
1300
1301 put_device:
1302         kref_put(&ndev->ref, nvmet_rdma_free_dev);
1303
1304         return ret;
1305 }
1306
1307 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1308 {
1309         unsigned long flags;
1310
1311         spin_lock_irqsave(&queue->state_lock, flags);
1312         if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1313                 pr_warn("trying to establish a connected queue\n");
1314                 goto out_unlock;
1315         }
1316         queue->state = NVMET_RDMA_Q_LIVE;
1317
1318         while (!list_empty(&queue->rsp_wait_list)) {
1319                 struct nvmet_rdma_rsp *cmd;
1320
1321                 cmd = list_first_entry(&queue->rsp_wait_list,
1322                                         struct nvmet_rdma_rsp, wait_list);
1323                 list_del(&cmd->wait_list);
1324
1325                 spin_unlock_irqrestore(&queue->state_lock, flags);
1326                 nvmet_rdma_handle_command(queue, cmd);
1327                 spin_lock_irqsave(&queue->state_lock, flags);
1328         }
1329
1330 out_unlock:
1331         spin_unlock_irqrestore(&queue->state_lock, flags);
1332 }
1333
1334 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1335 {
1336         bool disconnect = false;
1337         unsigned long flags;
1338
1339         pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1340
1341         spin_lock_irqsave(&queue->state_lock, flags);
1342         switch (queue->state) {
1343         case NVMET_RDMA_Q_CONNECTING:
1344         case NVMET_RDMA_Q_LIVE:
1345                 queue->state = NVMET_RDMA_Q_DISCONNECTING;
1346                 disconnect = true;
1347                 break;
1348         case NVMET_RDMA_Q_DISCONNECTING:
1349                 break;
1350         }
1351         spin_unlock_irqrestore(&queue->state_lock, flags);
1352
1353         if (disconnect) {
1354                 rdma_disconnect(queue->cm_id);
1355                 schedule_work(&queue->release_work);
1356         }
1357 }
1358
1359 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1360 {
1361         bool disconnect = false;
1362
1363         mutex_lock(&nvmet_rdma_queue_mutex);
1364         if (!list_empty(&queue->queue_list)) {
1365                 list_del_init(&queue->queue_list);
1366                 disconnect = true;
1367         }
1368         mutex_unlock(&nvmet_rdma_queue_mutex);
1369
1370         if (disconnect)
1371                 __nvmet_rdma_queue_disconnect(queue);
1372 }
1373
1374 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1375                 struct nvmet_rdma_queue *queue)
1376 {
1377         WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1378
1379         mutex_lock(&nvmet_rdma_queue_mutex);
1380         if (!list_empty(&queue->queue_list))
1381                 list_del_init(&queue->queue_list);
1382         mutex_unlock(&nvmet_rdma_queue_mutex);
1383
1384         pr_err("failed to connect queue %d\n", queue->idx);
1385         schedule_work(&queue->release_work);
1386 }
1387
1388 /**
1389  * nvme_rdma_device_removal() - Handle RDMA device removal
1390  * @cm_id:      rdma_cm id, used for nvmet port
1391  * @queue:      nvmet rdma queue (cm id qp_context)
1392  *
1393  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1394  * to unplug. Note that this event can be generated on a normal
1395  * queue cm_id and/or a device bound listener cm_id (where in this
1396  * case queue will be null).
1397  *
1398  * We registered an ib_client to handle device removal for queues,
1399  * so we only need to handle the listening port cm_ids. In this case
1400  * we nullify the priv to prevent double cm_id destruction and destroying
1401  * the cm_id implicitely by returning a non-zero rc to the callout.
1402  */
1403 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1404                 struct nvmet_rdma_queue *queue)
1405 {
1406         struct nvmet_port *port;
1407
1408         if (queue) {
1409                 /*
1410                  * This is a queue cm_id. we have registered
1411                  * an ib_client to handle queues removal
1412                  * so don't interfear and just return.
1413                  */
1414                 return 0;
1415         }
1416
1417         port = cm_id->context;
1418
1419         /*
1420          * This is a listener cm_id. Make sure that
1421          * future remove_port won't invoke a double
1422          * cm_id destroy. use atomic xchg to make sure
1423          * we don't compete with remove_port.
1424          */
1425         if (xchg(&port->priv, NULL) != cm_id)
1426                 return 0;
1427
1428         /*
1429          * We need to return 1 so that the core will destroy
1430          * it's own ID.  What a great API design..
1431          */
1432         return 1;
1433 }
1434
1435 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1436                 struct rdma_cm_event *event)
1437 {
1438         struct nvmet_rdma_queue *queue = NULL;
1439         int ret = 0;
1440
1441         if (cm_id->qp)
1442                 queue = cm_id->qp->qp_context;
1443
1444         pr_debug("%s (%d): status %d id %p\n",
1445                 rdma_event_msg(event->event), event->event,
1446                 event->status, cm_id);
1447
1448         switch (event->event) {
1449         case RDMA_CM_EVENT_CONNECT_REQUEST:
1450                 ret = nvmet_rdma_queue_connect(cm_id, event);
1451                 break;
1452         case RDMA_CM_EVENT_ESTABLISHED:
1453                 nvmet_rdma_queue_established(queue);
1454                 break;
1455         case RDMA_CM_EVENT_ADDR_CHANGE:
1456         case RDMA_CM_EVENT_DISCONNECTED:
1457         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1458                 nvmet_rdma_queue_disconnect(queue);
1459                 break;
1460         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1461                 ret = nvmet_rdma_device_removal(cm_id, queue);
1462                 break;
1463         case RDMA_CM_EVENT_REJECTED:
1464                 pr_debug("Connection rejected: %s\n",
1465                          rdma_reject_msg(cm_id, event->status));
1466                 /* FALLTHROUGH */
1467         case RDMA_CM_EVENT_UNREACHABLE:
1468         case RDMA_CM_EVENT_CONNECT_ERROR:
1469                 nvmet_rdma_queue_connect_fail(cm_id, queue);
1470                 break;
1471         default:
1472                 pr_err("received unrecognized RDMA CM event %d\n",
1473                         event->event);
1474                 break;
1475         }
1476
1477         return ret;
1478 }
1479
1480 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1481 {
1482         struct nvmet_rdma_queue *queue;
1483
1484 restart:
1485         mutex_lock(&nvmet_rdma_queue_mutex);
1486         list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1487                 if (queue->nvme_sq.ctrl == ctrl) {
1488                         list_del_init(&queue->queue_list);
1489                         mutex_unlock(&nvmet_rdma_queue_mutex);
1490
1491                         __nvmet_rdma_queue_disconnect(queue);
1492                         goto restart;
1493                 }
1494         }
1495         mutex_unlock(&nvmet_rdma_queue_mutex);
1496 }
1497
1498 static int nvmet_rdma_add_port(struct nvmet_port *port)
1499 {
1500         struct rdma_cm_id *cm_id;
1501         struct sockaddr_storage addr = { };
1502         __kernel_sa_family_t af;
1503         int ret;
1504
1505         switch (port->disc_addr.adrfam) {
1506         case NVMF_ADDR_FAMILY_IP4:
1507                 af = AF_INET;
1508                 break;
1509         case NVMF_ADDR_FAMILY_IP6:
1510                 af = AF_INET6;
1511                 break;
1512         default:
1513                 pr_err("address family %d not supported\n",
1514                                 port->disc_addr.adrfam);
1515                 return -EINVAL;
1516         }
1517
1518         if (port->inline_data_size < 0) {
1519                 port->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
1520         } else if (port->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
1521                 pr_warn("inline_data_size %u is too large, reducing to %u\n",
1522                         port->inline_data_size,
1523                         NVMET_RDMA_MAX_INLINE_DATA_SIZE);
1524                 port->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
1525         }
1526
1527         ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
1528                         port->disc_addr.trsvcid, &addr);
1529         if (ret) {
1530                 pr_err("malformed ip/port passed: %s:%s\n",
1531                         port->disc_addr.traddr, port->disc_addr.trsvcid);
1532                 return ret;
1533         }
1534
1535         cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1536                         RDMA_PS_TCP, IB_QPT_RC);
1537         if (IS_ERR(cm_id)) {
1538                 pr_err("CM ID creation failed\n");
1539                 return PTR_ERR(cm_id);
1540         }
1541
1542         /*
1543          * Allow both IPv4 and IPv6 sockets to bind a single port
1544          * at the same time.
1545          */
1546         ret = rdma_set_afonly(cm_id, 1);
1547         if (ret) {
1548                 pr_err("rdma_set_afonly failed (%d)\n", ret);
1549                 goto out_destroy_id;
1550         }
1551
1552         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
1553         if (ret) {
1554                 pr_err("binding CM ID to %pISpcs failed (%d)\n",
1555                         (struct sockaddr *)&addr, ret);
1556                 goto out_destroy_id;
1557         }
1558
1559         ret = rdma_listen(cm_id, 128);
1560         if (ret) {
1561                 pr_err("listening to %pISpcs failed (%d)\n",
1562                         (struct sockaddr *)&addr, ret);
1563                 goto out_destroy_id;
1564         }
1565
1566         pr_info("enabling port %d (%pISpcs)\n",
1567                 le16_to_cpu(port->disc_addr.portid), (struct sockaddr *)&addr);
1568         port->priv = cm_id;
1569         return 0;
1570
1571 out_destroy_id:
1572         rdma_destroy_id(cm_id);
1573         return ret;
1574 }
1575
1576 static void nvmet_rdma_remove_port(struct nvmet_port *port)
1577 {
1578         struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
1579
1580         if (cm_id)
1581                 rdma_destroy_id(cm_id);
1582 }
1583
1584 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
1585                 struct nvmet_port *port, char *traddr)
1586 {
1587         struct rdma_cm_id *cm_id = port->priv;
1588
1589         if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
1590                 struct nvmet_rdma_rsp *rsp =
1591                         container_of(req, struct nvmet_rdma_rsp, req);
1592                 struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
1593                 struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;
1594
1595                 sprintf(traddr, "%pISc", addr);
1596         } else {
1597                 memcpy(traddr, port->disc_addr.traddr, NVMF_TRADDR_SIZE);
1598         }
1599 }
1600
1601 static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
1602         .owner                  = THIS_MODULE,
1603         .type                   = NVMF_TRTYPE_RDMA,
1604         .msdbd                  = 1,
1605         .has_keyed_sgls         = 1,
1606         .add_port               = nvmet_rdma_add_port,
1607         .remove_port            = nvmet_rdma_remove_port,
1608         .queue_response         = nvmet_rdma_queue_response,
1609         .delete_ctrl            = nvmet_rdma_delete_ctrl,
1610         .disc_traddr            = nvmet_rdma_disc_port_addr,
1611 };
1612
1613 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1614 {
1615         struct nvmet_rdma_queue *queue, *tmp;
1616         struct nvmet_rdma_device *ndev;
1617         bool found = false;
1618
1619         mutex_lock(&device_list_mutex);
1620         list_for_each_entry(ndev, &device_list, entry) {
1621                 if (ndev->device == ib_device) {
1622                         found = true;
1623                         break;
1624                 }
1625         }
1626         mutex_unlock(&device_list_mutex);
1627
1628         if (!found)
1629                 return;
1630
1631         /*
1632          * IB Device that is used by nvmet controllers is being removed,
1633          * delete all queues using this device.
1634          */
1635         mutex_lock(&nvmet_rdma_queue_mutex);
1636         list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
1637                                  queue_list) {
1638                 if (queue->dev->device != ib_device)
1639                         continue;
1640
1641                 pr_info("Removing queue %d\n", queue->idx);
1642                 list_del_init(&queue->queue_list);
1643                 __nvmet_rdma_queue_disconnect(queue);
1644         }
1645         mutex_unlock(&nvmet_rdma_queue_mutex);
1646
1647         flush_scheduled_work();
1648 }
1649
1650 static struct ib_client nvmet_rdma_ib_client = {
1651         .name   = "nvmet_rdma",
1652         .remove = nvmet_rdma_remove_one
1653 };
1654
1655 static int __init nvmet_rdma_init(void)
1656 {
1657         int ret;
1658
1659         ret = ib_register_client(&nvmet_rdma_ib_client);
1660         if (ret)
1661                 return ret;
1662
1663         ret = nvmet_register_transport(&nvmet_rdma_ops);
1664         if (ret)
1665                 goto err_ib_client;
1666
1667         return 0;
1668
1669 err_ib_client:
1670         ib_unregister_client(&nvmet_rdma_ib_client);
1671         return ret;
1672 }
1673
1674 static void __exit nvmet_rdma_exit(void)
1675 {
1676         nvmet_unregister_transport(&nvmet_rdma_ops);
1677         ib_unregister_client(&nvmet_rdma_ib_client);
1678         WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
1679         ida_destroy(&nvmet_rdma_queue_ida);
1680 }
1681
1682 module_init(nvmet_rdma_init);
1683 module_exit(nvmet_rdma_exit);
1684
1685 MODULE_LICENSE("GPL v2");
1686 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */