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 struct workqueue_struct *nvmet_rdma_delete_wq;
 126 static bool nvmet_rdma_use_srq;
 127 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
 128 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
 129
 130 static DEFINE_IDA(nvmet_rdma_queue_ida);
 131 static LIST_HEAD(nvmet_rdma_queue_list);
 132 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
 133
 134 static LIST_HEAD(device_list);
 135 static DEFINE_MUTEX(device_list_mutex);
 136
 137 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
 138 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
 139 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 140 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
 141 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
 142 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
 143
 144 static const struct nvmet_fabrics_ops nvmet_rdma_ops;
 145
 146 static int num_pages(int len)
 147 {
 148         return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
 149 }
 150
 151 /* XXX: really should move to a generic header sooner or later.. */
 152 static inline u32 get_unaligned_le24(const u8 *p)
 153 {
 154         return (u32)p[0] | (u32)p[1] << 8 | (u32)p[2] << 16;
 155 }
 156
 157 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
 158 {
 159         return nvme_is_write(rsp->req.cmd) &&
 160                 rsp->req.transfer_len &&
 161                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 162 }
 163
 164 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
 165 {
 166         return !nvme_is_write(rsp->req.cmd) &&
 167                 rsp->req.transfer_len &&
 168                 !rsp->req.rsp->status &&
 169                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 170 }
 171
 172 static inline struct nvmet_rdma_rsp *
 173 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
 174 {
 175         struct nvmet_rdma_rsp *rsp;
 176         unsigned long flags;
 177
 178         spin_lock_irqsave(&queue->rsps_lock, flags);
 179         rsp = list_first_entry_or_null(&queue->free_rsps,
 180                                 struct nvmet_rdma_rsp, free_list);
 181         if (likely(rsp))
 182                 list_del(&rsp->free_list);
 183         spin_unlock_irqrestore(&queue->rsps_lock, flags);
 184
 185         if (unlikely(!rsp)) {
 186                 rsp = kmalloc(sizeof(*rsp), GFP_KERNEL);
 187                 if (unlikely(!rsp))
 188                         return NULL;
 189                 rsp->allocated = true;
 190         }
 191
 192         return rsp;
 193 }
 194
 195 static inline void
 196 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
 197 {
 198         unsigned long flags;
 199
 200         if (rsp->allocated) {
 201                 kfree(rsp);
 202                 return;
 203         }
 204
 205         spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
 206         list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
 207         spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
 208 }
 209
 210 static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
 211                                 struct nvmet_rdma_cmd *c)
 212 {
 213         struct scatterlist *sg;
 214         struct ib_sge *sge;
 215         int i;
 216
 217         if (!ndev->inline_data_size)
 218                 return;
 219
 220         sg = c->inline_sg;
 221         sge = &c->sge[1];
 222
 223         for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
 224                 if (sge->length)
 225                         ib_dma_unmap_page(ndev->device, sge->addr,
 226                                         sge->length, DMA_FROM_DEVICE);
 227                 if (sg_page(sg))
 228                         __free_page(sg_page(sg));
 229         }
 230 }
 231
 232 static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
 233                                 struct nvmet_rdma_cmd *c)
 234 {
 235         struct scatterlist *sg;
 236         struct ib_sge *sge;
 237         struct page *pg;
 238         int len;
 239         int i;
 240
 241         if (!ndev->inline_data_size)
 242                 return 0;
 243
 244         sg = c->inline_sg;
 245         sg_init_table(sg, ndev->inline_page_count);
 246         sge = &c->sge[1];
 247         len = ndev->inline_data_size;
 248
 249         for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
 250                 pg = alloc_page(GFP_KERNEL);
 251                 if (!pg)
 252                         goto out_err;
 253                 sg_assign_page(sg, pg);
 254                 sge->addr = ib_dma_map_page(ndev->device,
 255                         pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
 256                 if (ib_dma_mapping_error(ndev->device, sge->addr))
 257                         goto out_err;
 258                 sge->length = min_t(int, len, PAGE_SIZE);
 259                 sge->lkey = ndev->pd->local_dma_lkey;
 260                 len -= sge->length;
 261         }
 262
 263         return 0;
 264 out_err:
 265         for (; i >= 0; i--, sg--, sge--) {
 266                 if (sge->length)
 267                         ib_dma_unmap_page(ndev->device, sge->addr,
 268                                         sge->length, DMA_FROM_DEVICE);
 269                 if (sg_page(sg))
 270                         __free_page(sg_page(sg));
 271         }
 272         return -ENOMEM;
 273 }
 274
 275 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
 276                         struct nvmet_rdma_cmd *c, bool admin)
 277 {
 278         /* NVMe command / RDMA RECV */
 279         c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
 280         if (!c->nvme_cmd)
 281                 goto out;
 282
 283         c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
 284                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 285         if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
 286                 goto out_free_cmd;
 287
 288         c->sge[0].length = sizeof(*c->nvme_cmd);
 289         c->sge[0].lkey = ndev->pd->local_dma_lkey;
 290
 291         if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
 292                 goto out_unmap_cmd;
 293
 294         c->cqe.done = nvmet_rdma_recv_done;
 295
 296         c->wr.wr_cqe = &c->cqe;
 297         c->wr.sg_list = c->sge;
 298         c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;
 299
 300         return 0;
 301
 302 out_unmap_cmd:
 303         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 304                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 305 out_free_cmd:
 306         kfree(c->nvme_cmd);
 307
 308 out:
 309         return -ENOMEM;
 310 }
 311
 312 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
 313                 struct nvmet_rdma_cmd *c, bool admin)
 314 {
 315         if (!admin)
 316                 nvmet_rdma_free_inline_pages(ndev, c);
 317         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 318                                 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 319         kfree(c->nvme_cmd);
 320 }
 321
 322 static struct nvmet_rdma_cmd *
 323 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
 324                 int nr_cmds, bool admin)
 325 {
 326         struct nvmet_rdma_cmd *cmds;
 327         int ret = -EINVAL, i;
 328
 329         cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
 330         if (!cmds)
 331                 goto out;
 332
 333         for (i = 0; i < nr_cmds; i++) {
 334                 ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
 335                 if (ret)
 336                         goto out_free;
 337         }
 338
 339         return cmds;
 340
 341 out_free:
 342         while (--i >= 0)
 343                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 344         kfree(cmds);
 345 out:
 346         return ERR_PTR(ret);
 347 }
 348
 349 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
 350                 struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
 351 {
 352         int i;
 353
 354         for (i = 0; i < nr_cmds; i++)
 355                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 356         kfree(cmds);
 357 }
 358
 359 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
 360                 struct nvmet_rdma_rsp *r)
 361 {
 362         /* NVMe CQE / RDMA SEND */
 363         r->req.rsp = kmalloc(sizeof(*r->req.rsp), GFP_KERNEL);
 364         if (!r->req.rsp)
 365                 goto out;
 366
 367         r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.rsp,
 368                         sizeof(*r->req.rsp), DMA_TO_DEVICE);
 369         if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
 370                 goto out_free_rsp;
 371
 372         r->send_sge.length = sizeof(*r->req.rsp);
 373         r->send_sge.lkey = ndev->pd->local_dma_lkey;
 374
 375         r->send_cqe.done = nvmet_rdma_send_done;
 376
 377         r->send_wr.wr_cqe = &r->send_cqe;
 378         r->send_wr.sg_list = &r->send_sge;
 379         r->send_wr.num_sge = 1;
 380         r->send_wr.send_flags = IB_SEND_SIGNALED;
 381
 382         /* Data In / RDMA READ */
 383         r->read_cqe.done = nvmet_rdma_read_data_done;
 384         return 0;
 385
 386 out_free_rsp:
 387         kfree(r->req.rsp);
 388 out:
 389         return -ENOMEM;
 390 }
 391
 392 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
 393                 struct nvmet_rdma_rsp *r)
 394 {
 395         ib_dma_unmap_single(ndev->device, r->send_sge.addr,
 396                                 sizeof(*r->req.rsp), DMA_TO_DEVICE);
 397         kfree(r->req.rsp);
 398 }
 399
 400 static int
 401 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
 402 {
 403         struct nvmet_rdma_device *ndev = queue->dev;
 404         int nr_rsps = queue->recv_queue_size * 2;
 405         int ret = -EINVAL, i;
 406
 407         queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
 408                         GFP_KERNEL);
 409         if (!queue->rsps)
 410                 goto out;
 411
 412         for (i = 0; i < nr_rsps; i++) {
 413                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 414
 415                 ret = nvmet_rdma_alloc_rsp(ndev, rsp);
 416                 if (ret)
 417                         goto out_free;
 418
 419                 list_add_tail(&rsp->free_list, &queue->free_rsps);
 420         }
 421
 422         return 0;
 423
 424 out_free:
 425         while (--i >= 0) {
 426                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 427
 428                 list_del(&rsp->free_list);
 429                 nvmet_rdma_free_rsp(ndev, rsp);
 430         }
 431         kfree(queue->rsps);
 432 out:
 433         return ret;
 434 }
 435
 436 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
 437 {
 438         struct nvmet_rdma_device *ndev = queue->dev;
 439         int i, nr_rsps = queue->recv_queue_size * 2;
 440
 441         for (i = 0; i < nr_rsps; i++) {
 442                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 443
 444                 list_del(&rsp->free_list);
 445                 nvmet_rdma_free_rsp(ndev, rsp);
 446         }
 447         kfree(queue->rsps);
 448 }
 449
 450 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
 451                 struct nvmet_rdma_cmd *cmd)
 452 {
 453         int ret;
 454
 455         ib_dma_sync_single_for_device(ndev->device,
 456                 cmd->sge[0].addr, cmd->sge[0].length,
 457                 DMA_FROM_DEVICE);
 458
 459         if (ndev->srq)
 460                 ret = ib_post_srq_recv(ndev->srq, &cmd->wr, NULL);
 461         else
 462                 ret = ib_post_recv(cmd->queue->cm_id->qp, &cmd->wr, NULL);
 463
 464         if (unlikely(ret))
 465                 pr_err("post_recv cmd failed\n");
 466
 467         return ret;
 468 }
 469
 470 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
 471 {
 472         spin_lock(&queue->rsp_wr_wait_lock);
 473         while (!list_empty(&queue->rsp_wr_wait_list)) {
 474                 struct nvmet_rdma_rsp *rsp;
 475                 bool ret;
 476
 477                 rsp = list_entry(queue->rsp_wr_wait_list.next,
 478                                 struct nvmet_rdma_rsp, wait_list);
 479                 list_del(&rsp->wait_list);
 480
 481                 spin_unlock(&queue->rsp_wr_wait_lock);
 482                 ret = nvmet_rdma_execute_command(rsp);
 483                 spin_lock(&queue->rsp_wr_wait_lock);
 484
 485                 if (!ret) {
 486                         list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
 487                         break;
 488                 }
 489         }
 490         spin_unlock(&queue->rsp_wr_wait_lock);
 491 }
 492
 493
 494 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
 495 {
 496         struct nvmet_rdma_queue *queue = rsp->queue;
 497
 498         atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 499
 500         if (rsp->n_rdma) {
 501                 rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
 502                                 queue->cm_id->port_num, rsp->req.sg,
 503                                 rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
 504         }
 505
 506         if (rsp->req.sg != rsp->cmd->inline_sg)
 507                 nvmet_req_free_sgl(&rsp->req);
 508
 509         if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
 510                 nvmet_rdma_process_wr_wait_list(queue);
 511
 512         nvmet_rdma_put_rsp(rsp);
 513 }
 514
 515 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
 516 {
 517         if (queue->nvme_sq.ctrl) {
 518                 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
 519         } else {
 520                 /*
 521                  * we didn't setup the controller yet in case
 522                  * of admin connect error, just disconnect and
 523                  * cleanup the queue
 524                  */
 525                 nvmet_rdma_queue_disconnect(queue);
 526         }
 527 }
 528
 529 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
 530 {
 531         struct nvmet_rdma_rsp *rsp =
 532                 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
 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(rsp->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                 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 635
 636         if (off + len > rsp->queue->dev->inline_data_size) {
 637                 pr_err("invalid inline data offset!\n");
 638                 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
 639         }
 640
 641         /* no data command? */
 642         if (!len)
 643                 return 0;
 644
 645         nvmet_rdma_use_inline_sg(rsp, len, off);
 646         rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
 647         rsp->req.transfer_len += len;
 648         return 0;
 649 }
 650
 651 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
 652                 struct nvme_keyed_sgl_desc *sgl, bool invalidate)
 653 {
 654         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 655         u64 addr = le64_to_cpu(sgl->addr);
 656         u32 key = get_unaligned_le32(sgl->key);
 657         int ret;
 658
 659         rsp->req.transfer_len = get_unaligned_le24(sgl->length);
 660
 661         /* no data command? */
 662         if (!rsp->req.transfer_len)
 663                 return 0;
 664
 665         ret = nvmet_req_alloc_sgl(&rsp->req);
 666         if (ret < 0)
 667                 goto error_out;
 668
 669         ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
 670                         rsp->req.sg, rsp->req.sg_cnt, 0, addr, key,
 671                         nvmet_data_dir(&rsp->req));
 672         if (ret < 0)
 673                 goto error_out;
 674         rsp->n_rdma += ret;
 675
 676         if (invalidate) {
 677                 rsp->invalidate_rkey = key;
 678                 rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
 679         }
 680
 681         return 0;
 682
 683 error_out:
 684         rsp->req.transfer_len = 0;
 685         return NVME_SC_INTERNAL;
 686 }
 687
 688 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
 689 {
 690         struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
 691
 692         switch (sgl->type >> 4) {
 693         case NVME_SGL_FMT_DATA_DESC:
 694                 switch (sgl->type & 0xf) {
 695                 case NVME_SGL_FMT_OFFSET:
 696                         return nvmet_rdma_map_sgl_inline(rsp);
 697                 default:
 698                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 699                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 700                 }
 701         case NVME_KEY_SGL_FMT_DATA_DESC:
 702                 switch (sgl->type & 0xf) {
 703                 case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
 704                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
 705                 case NVME_SGL_FMT_ADDRESS:
 706                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
 707                 default:
 708                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 709                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 710                 }
 711         default:
 712                 pr_err("invalid SGL type: %#x\n", sgl->type);
 713                 return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
 714         }
 715 }
 716
 717 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
 718 {
 719         struct nvmet_rdma_queue *queue = rsp->queue;
 720
 721         if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
 722                         &queue->sq_wr_avail) < 0)) {
 723                 pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
 724                                 1 + rsp->n_rdma, queue->idx,
 725                                 queue->nvme_sq.ctrl->cntlid);
 726                 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 727                 return false;
 728         }
 729
 730         if (nvmet_rdma_need_data_in(rsp)) {
 731                 if (rdma_rw_ctx_post(&rsp->rw, queue->cm_id->qp,
 732                                 queue->cm_id->port_num, &rsp->read_cqe, NULL))
 733                         nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
 734         } else {
 735                 nvmet_req_execute(&rsp->req);
 736         }
 737
 738         return true;
 739 }
 740
 741 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
 742                 struct nvmet_rdma_rsp *cmd)
 743 {
 744         u16 status;
 745
 746         ib_dma_sync_single_for_cpu(queue->dev->device,
 747                 cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
 748                 DMA_FROM_DEVICE);
 749         ib_dma_sync_single_for_cpu(queue->dev->device,
 750                 cmd->send_sge.addr, cmd->send_sge.length,
 751                 DMA_TO_DEVICE);
 752
 753         cmd->req.p2p_client = &queue->dev->device->dev;
 754
 755         if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
 756                         &queue->nvme_sq, &nvmet_rdma_ops))
 757                 return;
 758
 759         status = nvmet_rdma_map_sgl(cmd);
 760         if (status)
 761                 goto out_err;
 762
 763         if (unlikely(!nvmet_rdma_execute_command(cmd))) {
 764                 spin_lock(&queue->rsp_wr_wait_lock);
 765                 list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
 766                 spin_unlock(&queue->rsp_wr_wait_lock);
 767         }
 768
 769         return;
 770
 771 out_err:
 772         nvmet_req_complete(&cmd->req, status);
 773 }
 774
 775 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
 776 {
 777         struct nvmet_rdma_cmd *cmd =
 778                 container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
 779         struct nvmet_rdma_queue *queue = cq->cq_context;
 780         struct nvmet_rdma_rsp *rsp;
 781
 782         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 783                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 784                         pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
 785                                 wc->wr_cqe, ib_wc_status_msg(wc->status),
 786                                 wc->status);
 787                         nvmet_rdma_error_comp(queue);
 788                 }
 789                 return;
 790         }
 791
 792         if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
 793                 pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
 794                 nvmet_rdma_error_comp(queue);
 795                 return;
 796         }
 797
 798         cmd->queue = queue;
 799         rsp = nvmet_rdma_get_rsp(queue);
 800         if (unlikely(!rsp)) {
 801                 /*
 802                  * we get here only under memory pressure,
 803                  * silently drop and have the host retry
 804                  * as we can't even fail it.
 805                  */
 806                 nvmet_rdma_post_recv(queue->dev, cmd);
 807                 return;
 808         }
 809         rsp->queue = queue;
 810         rsp->cmd = cmd;
 811         rsp->flags = 0;
 812         rsp->req.cmd = cmd->nvme_cmd;
 813         rsp->req.port = queue->port;
 814         rsp->n_rdma = 0;
 815
 816         if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
 817                 unsigned long flags;
 818
 819                 spin_lock_irqsave(&queue->state_lock, flags);
 820                 if (queue->state == NVMET_RDMA_Q_CONNECTING)
 821                         list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
 822                 else
 823                         nvmet_rdma_put_rsp(rsp);
 824                 spin_unlock_irqrestore(&queue->state_lock, flags);
 825                 return;
 826         }
 827
 828         nvmet_rdma_handle_command(queue, rsp);
 829 }
 830
 831 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_device *ndev)
 832 {
 833         if (!ndev->srq)
 834                 return;
 835
 836         nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
 837         ib_destroy_srq(ndev->srq);
 838 }
 839
 840 static int nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
 841 {
 842         struct ib_srq_init_attr srq_attr = { NULL, };
 843         struct ib_srq *srq;
 844         size_t srq_size;
 845         int ret, i;
 846
 847         srq_size = 4095;        /* XXX: tune */
 848
 849         srq_attr.attr.max_wr = srq_size;
 850         srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
 851         srq_attr.attr.srq_limit = 0;
 852         srq_attr.srq_type = IB_SRQT_BASIC;
 853         srq = ib_create_srq(ndev->pd, &srq_attr);
 854         if (IS_ERR(srq)) {
 855                 /*
 856                  * If SRQs aren't supported we just go ahead and use normal
 857                  * non-shared receive queues.
 858                  */
 859                 pr_info("SRQ requested but not supported.\n");
 860                 return 0;
 861         }
 862
 863         ndev->srq_cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
 864         if (IS_ERR(ndev->srq_cmds)) {
 865                 ret = PTR_ERR(ndev->srq_cmds);
 866                 goto out_destroy_srq;
 867         }
 868
 869         ndev->srq = srq;
 870         ndev->srq_size = srq_size;
 871
 872         for (i = 0; i < srq_size; i++) {
 873                 ret = nvmet_rdma_post_recv(ndev, &ndev->srq_cmds[i]);
 874                 if (ret)
 875                         goto out_free_cmds;
 876         }
 877
 878         return 0;
 879
 880 out_free_cmds:
 881         nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
 882 out_destroy_srq:
 883         ib_destroy_srq(srq);
 884         return ret;
 885 }
 886
 887 static void nvmet_rdma_free_dev(struct kref *ref)
 888 {
 889         struct nvmet_rdma_device *ndev =
 890                 container_of(ref, struct nvmet_rdma_device, ref);
 891
 892         mutex_lock(&device_list_mutex);
 893         list_del(&ndev->entry);
 894         mutex_unlock(&device_list_mutex);
 895
 896         nvmet_rdma_destroy_srq(ndev);
 897         ib_dealloc_pd(ndev->pd);
 898
 899         kfree(ndev);
 900 }
 901
 902 static struct nvmet_rdma_device *
 903 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
 904 {
 905         struct nvmet_port *port = cm_id->context;
 906         struct nvmet_rdma_device *ndev;
 907         int inline_page_count;
 908         int inline_sge_count;
 909         int ret;
 910
 911         mutex_lock(&device_list_mutex);
 912         list_for_each_entry(ndev, &device_list, entry) {
 913                 if (ndev->device->node_guid == cm_id->device->node_guid &&
 914                     kref_get_unless_zero(&ndev->ref))
 915                         goto out_unlock;
 916         }
 917
 918         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 919         if (!ndev)
 920                 goto out_err;
 921
 922         inline_page_count = num_pages(port->inline_data_size);
 923         inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
 924                                 cm_id->device->attrs.max_recv_sge) - 1;
 925         if (inline_page_count > inline_sge_count) {
 926                 pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
 927                         port->inline_data_size, cm_id->device->name,
 928                         inline_sge_count * PAGE_SIZE);
 929                 port->inline_data_size = inline_sge_count * PAGE_SIZE;
 930                 inline_page_count = inline_sge_count;
 931         }
 932         ndev->inline_data_size = port->inline_data_size;
 933         ndev->inline_page_count = inline_page_count;
 934         ndev->device = cm_id->device;
 935         kref_init(&ndev->ref);
 936
 937         ndev->pd = ib_alloc_pd(ndev->device, 0);
 938         if (IS_ERR(ndev->pd))
 939                 goto out_free_dev;
 940
 941         if (nvmet_rdma_use_srq) {
 942                 ret = nvmet_rdma_init_srq(ndev);
 943                 if (ret)
 944                         goto out_free_pd;
 945         }
 946
 947         list_add(&ndev->entry, &device_list);
 948 out_unlock:
 949         mutex_unlock(&device_list_mutex);
 950         pr_debug("added %s.\n", ndev->device->name);
 951         return ndev;
 952
 953 out_free_pd:
 954         ib_dealloc_pd(ndev->pd);
 955 out_free_dev:
 956         kfree(ndev);
 957 out_err:
 958         mutex_unlock(&device_list_mutex);
 959         return NULL;
 960 }
 961
 962 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
 963 {
 964         struct ib_qp_init_attr qp_attr;
 965         struct nvmet_rdma_device *ndev = queue->dev;
 966         int comp_vector, nr_cqe, ret, i;
 967
 968         /*
 969          * Spread the io queues across completion vectors,
 970          * but still keep all admin queues on vector 0.
 971          */
 972         comp_vector = !queue->host_qid ? 0 :
 973                 queue->idx % ndev->device->num_comp_vectors;
 974
 975         /*
 976          * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
 977          */
 978         nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
 979
 980         queue->cq = ib_alloc_cq(ndev->device, queue,
 981                         nr_cqe + 1, comp_vector,
 982                         IB_POLL_WORKQUEUE);
 983         if (IS_ERR(queue->cq)) {
 984                 ret = PTR_ERR(queue->cq);
 985                 pr_err("failed to create CQ cqe= %d ret= %d\n",
 986                        nr_cqe + 1, ret);
 987                 goto out;
 988         }
 989
 990         memset(&qp_attr, 0, sizeof(qp_attr));
 991         qp_attr.qp_context = queue;
 992         qp_attr.event_handler = nvmet_rdma_qp_event;
 993         qp_attr.send_cq = queue->cq;
 994         qp_attr.recv_cq = queue->cq;
 995         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 996         qp_attr.qp_type = IB_QPT_RC;
 997         /* +1 for drain */
 998         qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
 999         qp_attr.cap.max_rdma_ctxs = queue->send_queue_size;
1000         qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
1001                                         ndev->device->attrs.max_send_sge);
1002
1003         if (ndev->srq) {
1004                 qp_attr.srq = ndev->srq;
1005         } else {
1006                 /* +1 for drain */
1007                 qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
1008                 qp_attr.cap.max_recv_sge = 1 + ndev->inline_page_count;
1009         }
1010
1011         ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
1012         if (ret) {
1013                 pr_err("failed to create_qp ret= %d\n", ret);
1014                 goto err_destroy_cq;
1015         }
1016
1017         atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
1018
1019         pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
1020                  __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
1021                  qp_attr.cap.max_send_wr, queue->cm_id);
1022
1023         if (!ndev->srq) {
1024                 for (i = 0; i < queue->recv_queue_size; i++) {
1025                         queue->cmds[i].queue = queue;
1026                         ret = nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
1027                         if (ret)
1028                                 goto err_destroy_qp;
1029                 }
1030         }
1031
1032 out:
1033         return ret;
1034
1035 err_destroy_qp:
1036         rdma_destroy_qp(queue->cm_id);
1037 err_destroy_cq:
1038         ib_free_cq(queue->cq);
1039         goto out;
1040 }
1041
1042 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1043 {
1044         struct ib_qp *qp = queue->cm_id->qp;
1045
1046         ib_drain_qp(qp);
1047         rdma_destroy_id(queue->cm_id);
1048         ib_destroy_qp(qp);
1049         ib_free_cq(queue->cq);
1050 }
1051
1052 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1053 {
1054         pr_debug("freeing queue %d\n", queue->idx);
1055
1056         nvmet_sq_destroy(&queue->nvme_sq);
1057
1058         nvmet_rdma_destroy_queue_ib(queue);
1059         if (!queue->dev->srq) {
1060                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1061                                 queue->recv_queue_size,
1062                                 !queue->host_qid);
1063         }
1064         nvmet_rdma_free_rsps(queue);
1065         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1066         kfree(queue);
1067 }
1068
1069 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1070 {
1071         struct nvmet_rdma_queue *queue =
1072                 container_of(w, struct nvmet_rdma_queue, release_work);
1073         struct nvmet_rdma_device *dev = queue->dev;
1074
1075         nvmet_rdma_free_queue(queue);
1076
1077         kref_put(&dev->ref, nvmet_rdma_free_dev);
1078 }
1079
1080 static int
1081 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1082                                 struct nvmet_rdma_queue *queue)
1083 {
1084         struct nvme_rdma_cm_req *req;
1085
1086         req = (struct nvme_rdma_cm_req *)conn->private_data;
1087         if (!req || conn->private_data_len == 0)
1088                 return NVME_RDMA_CM_INVALID_LEN;
1089
1090         if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1091                 return NVME_RDMA_CM_INVALID_RECFMT;
1092
1093         queue->host_qid = le16_to_cpu(req->qid);
1094
1095         /*
1096          * req->hsqsize corresponds to our recv queue size plus 1
1097          * req->hrqsize corresponds to our send queue size
1098          */
1099         queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1100         queue->send_queue_size = le16_to_cpu(req->hrqsize);
1101
1102         if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
1103                 return NVME_RDMA_CM_INVALID_HSQSIZE;
1104
1105         /* XXX: Should we enforce some kind of max for IO queues? */
1106
1107         return 0;
1108 }
1109
1110 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1111                                 enum nvme_rdma_cm_status status)
1112 {
1113         struct nvme_rdma_cm_rej rej;
1114
1115         pr_debug("rejecting connect request: status %d (%s)\n",
1116                  status, nvme_rdma_cm_msg(status));
1117
1118         rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1119         rej.sts = cpu_to_le16(status);
1120
1121         return rdma_reject(cm_id, (void *)&rej, sizeof(rej));
1122 }
1123
1124 static struct nvmet_rdma_queue *
1125 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1126                 struct rdma_cm_id *cm_id,
1127                 struct rdma_cm_event *event)
1128 {
1129         struct nvmet_rdma_queue *queue;
1130         int ret;
1131
1132         queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1133         if (!queue) {
1134                 ret = NVME_RDMA_CM_NO_RSC;
1135                 goto out_reject;
1136         }
1137
1138         ret = nvmet_sq_init(&queue->nvme_sq);
1139         if (ret) {
1140                 ret = NVME_RDMA_CM_NO_RSC;
1141                 goto out_free_queue;
1142         }
1143
1144         ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1145         if (ret)
1146                 goto out_destroy_sq;
1147
1148         /*
1149          * Schedules the actual release because calling rdma_destroy_id from
1150          * inside a CM callback would trigger a deadlock. (great API design..)
1151          */
1152         INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1153         queue->dev = ndev;
1154         queue->cm_id = cm_id;
1155
1156         spin_lock_init(&queue->state_lock);
1157         queue->state = NVMET_RDMA_Q_CONNECTING;
1158         INIT_LIST_HEAD(&queue->rsp_wait_list);
1159         INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1160         spin_lock_init(&queue->rsp_wr_wait_lock);
1161         INIT_LIST_HEAD(&queue->free_rsps);
1162         spin_lock_init(&queue->rsps_lock);
1163         INIT_LIST_HEAD(&queue->queue_list);
1164
1165         queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1166         if (queue->idx < 0) {
1167                 ret = NVME_RDMA_CM_NO_RSC;
1168                 goto out_destroy_sq;
1169         }
1170
1171         ret = nvmet_rdma_alloc_rsps(queue);
1172         if (ret) {
1173                 ret = NVME_RDMA_CM_NO_RSC;
1174                 goto out_ida_remove;
1175         }
1176
1177         if (!ndev->srq) {
1178                 queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1179                                 queue->recv_queue_size,
1180                                 !queue->host_qid);
1181                 if (IS_ERR(queue->cmds)) {
1182                         ret = NVME_RDMA_CM_NO_RSC;
1183                         goto out_free_responses;
1184                 }
1185         }
1186
1187         ret = nvmet_rdma_create_queue_ib(queue);
1188         if (ret) {
1189                 pr_err("%s: creating RDMA queue failed (%d).\n",
1190                         __func__, ret);
1191                 ret = NVME_RDMA_CM_NO_RSC;
1192                 goto out_free_cmds;
1193         }
1194
1195         return queue;
1196
1197 out_free_cmds:
1198         if (!ndev->srq) {
1199                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1200                                 queue->recv_queue_size,
1201                                 !queue->host_qid);
1202         }
1203 out_free_responses:
1204         nvmet_rdma_free_rsps(queue);
1205 out_ida_remove:
1206         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1207 out_destroy_sq:
1208         nvmet_sq_destroy(&queue->nvme_sq);
1209 out_free_queue:
1210         kfree(queue);
1211 out_reject:
1212         nvmet_rdma_cm_reject(cm_id, ret);
1213         return NULL;
1214 }
1215
1216 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1217 {
1218         struct nvmet_rdma_queue *queue = priv;
1219
1220         switch (event->event) {
1221         case IB_EVENT_COMM_EST:
1222                 rdma_notify(queue->cm_id, event->event);
1223                 break;
1224         default:
1225                 pr_err("received IB QP event: %s (%d)\n",
1226                        ib_event_msg(event->event), event->event);
1227                 break;
1228         }
1229 }
1230
1231 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1232                 struct nvmet_rdma_queue *queue,
1233                 struct rdma_conn_param *p)
1234 {
1235         struct rdma_conn_param  param = { };
1236         struct nvme_rdma_cm_rep priv = { };
1237         int ret = -ENOMEM;
1238
1239         param.rnr_retry_count = 7;
1240         param.flow_control = 1;
1241         param.initiator_depth = min_t(u8, p->initiator_depth,
1242                 queue->dev->device->attrs.max_qp_init_rd_atom);
1243         param.private_data = &priv;
1244         param.private_data_len = sizeof(priv);
1245         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1246         priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1247
1248         ret = rdma_accept(cm_id, &param);
1249         if (ret)
1250                 pr_err("rdma_accept failed (error code = %d)\n", ret);
1251
1252         return ret;
1253 }
1254
1255 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1256                 struct rdma_cm_event *event)
1257 {
1258         struct nvmet_rdma_device *ndev;
1259         struct nvmet_rdma_queue *queue;
1260         int ret = -EINVAL;
1261
1262         ndev = nvmet_rdma_find_get_device(cm_id);
1263         if (!ndev) {
1264                 nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1265                 return -ECONNREFUSED;
1266         }
1267
1268         queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1269         if (!queue) {
1270                 ret = -ENOMEM;
1271                 goto put_device;
1272         }
1273         queue->port = cm_id->context;
1274
1275         if (queue->host_qid == 0) {
1276                 /* Let inflight controller teardown complete */
1277                 flush_workqueue(nvmet_rdma_delete_wq);
1278         }
1279
1280         ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1281         if (ret) {
1282                 queue_work(nvmet_rdma_delete_wq, &queue->release_work);
1283                 /* Destroying rdma_cm id is not needed here */
1284                 return 0;
1285         }
1286
1287         mutex_lock(&nvmet_rdma_queue_mutex);
1288         list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1289         mutex_unlock(&nvmet_rdma_queue_mutex);
1290
1291         return 0;
1292
1293 put_device:
1294         kref_put(&ndev->ref, nvmet_rdma_free_dev);
1295
1296         return ret;
1297 }
1298
1299 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1300 {
1301         unsigned long flags;
1302
1303         spin_lock_irqsave(&queue->state_lock, flags);
1304         if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1305                 pr_warn("trying to establish a connected queue\n");
1306                 goto out_unlock;
1307         }
1308         queue->state = NVMET_RDMA_Q_LIVE;
1309
1310         while (!list_empty(&queue->rsp_wait_list)) {
1311                 struct nvmet_rdma_rsp *cmd;
1312
1313                 cmd = list_first_entry(&queue->rsp_wait_list,
1314                                         struct nvmet_rdma_rsp, wait_list);
1315                 list_del(&cmd->wait_list);
1316
1317                 spin_unlock_irqrestore(&queue->state_lock, flags);
1318                 nvmet_rdma_handle_command(queue, cmd);
1319                 spin_lock_irqsave(&queue->state_lock, flags);
1320         }
1321
1322 out_unlock:
1323         spin_unlock_irqrestore(&queue->state_lock, flags);
1324 }
1325
1326 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1327 {
1328         bool disconnect = false;
1329         unsigned long flags;
1330
1331         pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1332
1333         spin_lock_irqsave(&queue->state_lock, flags);
1334         switch (queue->state) {
1335         case NVMET_RDMA_Q_CONNECTING:
1336         case NVMET_RDMA_Q_LIVE:
1337                 queue->state = NVMET_RDMA_Q_DISCONNECTING;
1338                 disconnect = true;
1339                 break;
1340         case NVMET_RDMA_Q_DISCONNECTING:
1341                 break;
1342         }
1343         spin_unlock_irqrestore(&queue->state_lock, flags);
1344
1345         if (disconnect) {
1346                 rdma_disconnect(queue->cm_id);
1347                 queue_work(nvmet_rdma_delete_wq, &queue->release_work);
1348         }
1349 }
1350
1351 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1352 {
1353         bool disconnect = false;
1354
1355         mutex_lock(&nvmet_rdma_queue_mutex);
1356         if (!list_empty(&queue->queue_list)) {
1357                 list_del_init(&queue->queue_list);
1358                 disconnect = true;
1359         }
1360         mutex_unlock(&nvmet_rdma_queue_mutex);
1361
1362         if (disconnect)
1363                 __nvmet_rdma_queue_disconnect(queue);
1364 }
1365
1366 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1367                 struct nvmet_rdma_queue *queue)
1368 {
1369         WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1370
1371         mutex_lock(&nvmet_rdma_queue_mutex);
1372         if (!list_empty(&queue->queue_list))
1373                 list_del_init(&queue->queue_list);
1374         mutex_unlock(&nvmet_rdma_queue_mutex);
1375
1376         pr_err("failed to connect queue %d\n", queue->idx);
1377         queue_work(nvmet_rdma_delete_wq, &queue->release_work);
1378 }
1379
1380 /**
1381  * nvme_rdma_device_removal() - Handle RDMA device removal
1382  * @cm_id:      rdma_cm id, used for nvmet port
1383  * @queue:      nvmet rdma queue (cm id qp_context)
1384  *
1385  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1386  * to unplug. Note that this event can be generated on a normal
1387  * queue cm_id and/or a device bound listener cm_id (where in this
1388  * case queue will be null).
1389  *
1390  * We registered an ib_client to handle device removal for queues,
1391  * so we only need to handle the listening port cm_ids. In this case
1392  * we nullify the priv to prevent double cm_id destruction and destroying
1393  * the cm_id implicitely by returning a non-zero rc to the callout.
1394  */
1395 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1396                 struct nvmet_rdma_queue *queue)
1397 {
1398         struct nvmet_port *port;
1399
1400         if (queue) {
1401                 /*
1402                  * This is a queue cm_id. we have registered
1403                  * an ib_client to handle queues removal
1404                  * so don't interfear and just return.
1405                  */
1406                 return 0;
1407         }
1408
1409         port = cm_id->context;
1410
1411         /*
1412          * This is a listener cm_id. Make sure that
1413          * future remove_port won't invoke a double
1414          * cm_id destroy. use atomic xchg to make sure
1415          * we don't compete with remove_port.
1416          */
1417         if (xchg(&port->priv, NULL) != cm_id)
1418                 return 0;
1419
1420         /*
1421          * We need to return 1 so that the core will destroy
1422          * it's own ID.  What a great API design..
1423          */
1424         return 1;
1425 }
1426
1427 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1428                 struct rdma_cm_event *event)
1429 {
1430         struct nvmet_rdma_queue *queue = NULL;
1431         int ret = 0;
1432
1433         if (cm_id->qp)
1434                 queue = cm_id->qp->qp_context;
1435
1436         pr_debug("%s (%d): status %d id %p\n",
1437                 rdma_event_msg(event->event), event->event,
1438                 event->status, cm_id);
1439
1440         switch (event->event) {
1441         case RDMA_CM_EVENT_CONNECT_REQUEST:
1442                 ret = nvmet_rdma_queue_connect(cm_id, event);
1443                 break;
1444         case RDMA_CM_EVENT_ESTABLISHED:
1445                 nvmet_rdma_queue_established(queue);
1446                 break;
1447         case RDMA_CM_EVENT_ADDR_CHANGE:
1448         case RDMA_CM_EVENT_DISCONNECTED:
1449         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1450                 nvmet_rdma_queue_disconnect(queue);
1451                 break;
1452         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1453                 ret = nvmet_rdma_device_removal(cm_id, queue);
1454                 break;
1455         case RDMA_CM_EVENT_REJECTED:
1456                 pr_debug("Connection rejected: %s\n",
1457                          rdma_reject_msg(cm_id, event->status));
1458                 /* FALLTHROUGH */
1459         case RDMA_CM_EVENT_UNREACHABLE:
1460         case RDMA_CM_EVENT_CONNECT_ERROR:
1461                 nvmet_rdma_queue_connect_fail(cm_id, queue);
1462                 break;
1463         default:
1464                 pr_err("received unrecognized RDMA CM event %d\n",
1465                         event->event);
1466                 break;
1467         }
1468
1469         return ret;
1470 }
1471
1472 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1473 {
1474         struct nvmet_rdma_queue *queue;
1475
1476 restart:
1477         mutex_lock(&nvmet_rdma_queue_mutex);
1478         list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1479                 if (queue->nvme_sq.ctrl == ctrl) {
1480                         list_del_init(&queue->queue_list);
1481                         mutex_unlock(&nvmet_rdma_queue_mutex);
1482
1483                         __nvmet_rdma_queue_disconnect(queue);
1484                         goto restart;
1485                 }
1486         }
1487         mutex_unlock(&nvmet_rdma_queue_mutex);
1488 }
1489
1490 static int nvmet_rdma_add_port(struct nvmet_port *port)
1491 {
1492         struct rdma_cm_id *cm_id;
1493         struct sockaddr_storage addr = { };
1494         __kernel_sa_family_t af;
1495         int ret;
1496
1497         switch (port->disc_addr.adrfam) {
1498         case NVMF_ADDR_FAMILY_IP4:
1499                 af = AF_INET;
1500                 break;
1501         case NVMF_ADDR_FAMILY_IP6:
1502                 af = AF_INET6;
1503                 break;
1504         default:
1505                 pr_err("address family %d not supported\n",
1506                                 port->disc_addr.adrfam);
1507                 return -EINVAL;
1508         }
1509
1510         if (port->inline_data_size < 0) {
1511                 port->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
1512         } else if (port->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
1513                 pr_warn("inline_data_size %u is too large, reducing to %u\n",
1514                         port->inline_data_size,
1515                         NVMET_RDMA_MAX_INLINE_DATA_SIZE);
1516                 port->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
1517         }
1518
1519         ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
1520                         port->disc_addr.trsvcid, &addr);
1521         if (ret) {
1522                 pr_err("malformed ip/port passed: %s:%s\n",
1523                         port->disc_addr.traddr, port->disc_addr.trsvcid);
1524                 return ret;
1525         }
1526
1527         cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1528                         RDMA_PS_TCP, IB_QPT_RC);
1529         if (IS_ERR(cm_id)) {
1530                 pr_err("CM ID creation failed\n");
1531                 return PTR_ERR(cm_id);
1532         }
1533
1534         /*
1535          * Allow both IPv4 and IPv6 sockets to bind a single port
1536          * at the same time.
1537          */
1538         ret = rdma_set_afonly(cm_id, 1);
1539         if (ret) {
1540                 pr_err("rdma_set_afonly failed (%d)\n", ret);
1541                 goto out_destroy_id;
1542         }
1543
1544         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
1545         if (ret) {
1546                 pr_err("binding CM ID to %pISpcs failed (%d)\n",
1547                         (struct sockaddr *)&addr, ret);
1548                 goto out_destroy_id;
1549         }
1550
1551         ret = rdma_listen(cm_id, 128);
1552         if (ret) {
1553                 pr_err("listening to %pISpcs failed (%d)\n",
1554                         (struct sockaddr *)&addr, ret);
1555                 goto out_destroy_id;
1556         }
1557
1558         pr_info("enabling port %d (%pISpcs)\n",
1559                 le16_to_cpu(port->disc_addr.portid), (struct sockaddr *)&addr);
1560         port->priv = cm_id;
1561         return 0;
1562
1563 out_destroy_id:
1564         rdma_destroy_id(cm_id);
1565         return ret;
1566 }
1567
1568 static void nvmet_rdma_remove_port(struct nvmet_port *port)
1569 {
1570         struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
1571
1572         if (cm_id)
1573                 rdma_destroy_id(cm_id);
1574 }
1575
1576 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
1577                 struct nvmet_port *port, char *traddr)
1578 {
1579         struct rdma_cm_id *cm_id = port->priv;
1580
1581         if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
1582                 struct nvmet_rdma_rsp *rsp =
1583                         container_of(req, struct nvmet_rdma_rsp, req);
1584                 struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
1585                 struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;
1586
1587                 sprintf(traddr, "%pISc", addr);
1588         } else {
1589                 memcpy(traddr, port->disc_addr.traddr, NVMF_TRADDR_SIZE);
1590         }
1591 }
1592
1593 static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
1594         .owner                  = THIS_MODULE,
1595         .type                   = NVMF_TRTYPE_RDMA,
1596         .msdbd                  = 1,
1597         .has_keyed_sgls         = 1,
1598         .add_port               = nvmet_rdma_add_port,
1599         .remove_port            = nvmet_rdma_remove_port,
1600         .queue_response         = nvmet_rdma_queue_response,
1601         .delete_ctrl            = nvmet_rdma_delete_ctrl,
1602         .disc_traddr            = nvmet_rdma_disc_port_addr,
1603 };
1604
1605 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1606 {
1607         struct nvmet_rdma_queue *queue, *tmp;
1608         struct nvmet_rdma_device *ndev;
1609         bool found = false;
1610
1611         mutex_lock(&device_list_mutex);
1612         list_for_each_entry(ndev, &device_list, entry) {
1613                 if (ndev->device == ib_device) {
1614                         found = true;
1615                         break;
1616                 }
1617         }
1618         mutex_unlock(&device_list_mutex);
1619
1620         if (!found)
1621                 return;
1622
1623         /*
1624          * IB Device that is used by nvmet controllers is being removed,
1625          * delete all queues using this device.
1626          */
1627         mutex_lock(&nvmet_rdma_queue_mutex);
1628         list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
1629                                  queue_list) {
1630                 if (queue->dev->device != ib_device)
1631                         continue;
1632
1633                 pr_info("Removing queue %d\n", queue->idx);
1634                 list_del_init(&queue->queue_list);
1635                 __nvmet_rdma_queue_disconnect(queue);
1636         }
1637         mutex_unlock(&nvmet_rdma_queue_mutex);
1638
1639         flush_scheduled_work();
1640 }
1641
1642 static struct ib_client nvmet_rdma_ib_client = {
1643         .name   = "nvmet_rdma",
1644         .remove = nvmet_rdma_remove_one
1645 };
1646
1647 static int __init nvmet_rdma_init(void)
1648 {
1649         int ret;
1650
1651         ret = ib_register_client(&nvmet_rdma_ib_client);
1652         if (ret)
1653                 return ret;
1654
1655         ret = nvmet_register_transport(&nvmet_rdma_ops);
1656         if (ret)
1657                 goto err_ib_client;
1658
1659         nvmet_rdma_delete_wq = alloc_workqueue("nvmet-rdma-delete-wq",
1660                         WQ_UNBOUND | WQ_MEM_RECLAIM | WQ_SYSFS, 0);
1661         if (!nvmet_rdma_delete_wq) {
1662                 ret = -ENOMEM;
1663                 goto err_unreg_transport;
1664         }
1665
1666         return 0;
1667
1668 err_unreg_transport:
1669         nvmet_unregister_transport(&nvmet_rdma_ops);
1670 err_ib_client:
1671         ib_unregister_client(&nvmet_rdma_ib_client);
1672         return ret;
1673 }
1674
1675 static void __exit nvmet_rdma_exit(void)
1676 {
1677         destroy_workqueue(nvmet_rdma_delete_wq);
1678         nvmet_unregister_transport(&nvmet_rdma_ops);
1679         ib_unregister_client(&nvmet_rdma_ib_client);
1680         WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
1681         ida_destroy(&nvmet_rdma_queue_ida);
1682 }
1683
1684 module_init(nvmet_rdma_init);
1685 module_exit(nvmet_rdma_exit);
1686
1687 MODULE_LICENSE("GPL v2");
1688 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */