drivers/nvme/target/rdma.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * NVMe over Fabrics RDMA target.
   4  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
   5  */
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7 #include <linux/atomic.h>
   8 #include <linux/ctype.h>
   9 #include <linux/delay.h>
  10 #include <linux/err.h>
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/nvme.h>
  14 #include <linux/slab.h>
  15 #include <linux/string.h>
  16 #include <linux/wait.h>
  17 #include <linux/inet.h>
  18 #include <asm/unaligned.h>
  19
  20 #include <rdma/ib_verbs.h>
  21 #include <rdma/rdma_cm.h>
  22 #include <rdma/rw.h>
  23
  24 #include <linux/nvme-rdma.h>
  25 #include "nvmet.h"
  26
  27 /*
  28  * We allow at least 1 page, up to 4 SGEs, and up to 16KB of inline data
  29  */
  30 #define NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE     PAGE_SIZE
  31 #define NVMET_RDMA_MAX_INLINE_SGE               4
  32 #define NVMET_RDMA_MAX_INLINE_DATA_SIZE         max_t(int, SZ_16K, PAGE_SIZE)
  33
  34 struct nvmet_rdma_cmd {
  35         struct ib_sge           sge[NVMET_RDMA_MAX_INLINE_SGE + 1];
  36         struct ib_cqe           cqe;
  37         struct ib_recv_wr       wr;
  38         struct scatterlist      inline_sg[NVMET_RDMA_MAX_INLINE_SGE];
  39         struct nvme_command     *nvme_cmd;
  40         struct nvmet_rdma_queue *queue;
  41 };
  42
  43 enum {
  44         NVMET_RDMA_REQ_INLINE_DATA      = (1 << 0),
  45         NVMET_RDMA_REQ_INVALIDATE_RKEY  = (1 << 1),
  46 };
  47
  48 struct nvmet_rdma_rsp {
  49         struct ib_sge           send_sge;
  50         struct ib_cqe           send_cqe;
  51         struct ib_send_wr       send_wr;
  52
  53         struct nvmet_rdma_cmd   *cmd;
  54         struct nvmet_rdma_queue *queue;
  55
  56         struct ib_cqe           read_cqe;
  57         struct rdma_rw_ctx      rw;
  58
  59         struct nvmet_req        req;
  60
  61         bool                    allocated;
  62         u8                      n_rdma;
  63         u32                     flags;
  64         u32                     invalidate_rkey;
  65
  66         struct list_head        wait_list;
  67         struct list_head        free_list;
  68 };
  69
  70 enum nvmet_rdma_queue_state {
  71         NVMET_RDMA_Q_CONNECTING,
  72         NVMET_RDMA_Q_LIVE,
  73         NVMET_RDMA_Q_DISCONNECTING,
  74 };
  75
  76 struct nvmet_rdma_queue {
  77         struct rdma_cm_id       *cm_id;
  78         struct nvmet_port       *port;
  79         struct ib_cq            *cq;
  80         atomic_t                sq_wr_avail;
  81         struct nvmet_rdma_device *dev;
  82         spinlock_t              state_lock;
  83         enum nvmet_rdma_queue_state state;
  84         struct nvmet_cq         nvme_cq;
  85         struct nvmet_sq         nvme_sq;
  86
  87         struct nvmet_rdma_rsp   *rsps;
  88         struct list_head        free_rsps;
  89         spinlock_t              rsps_lock;
  90         struct nvmet_rdma_cmd   *cmds;
  91
  92         struct work_struct      release_work;
  93         struct list_head        rsp_wait_list;
  94         struct list_head        rsp_wr_wait_list;
  95         spinlock_t              rsp_wr_wait_lock;
  96
  97         int                     idx;
  98         int                     host_qid;
  99         int                     recv_queue_size;
 100         int                     send_queue_size;
 101
 102         struct list_head        queue_list;
 103 };
 104
 105 struct nvmet_rdma_device {
 106         struct ib_device        *device;
 107         struct ib_pd            *pd;
 108         struct ib_srq           *srq;
 109         struct nvmet_rdma_cmd   *srq_cmds;
 110         size_t                  srq_size;
 111         struct kref             ref;
 112         struct list_head        entry;
 113         int                     inline_data_size;
 114         int                     inline_page_count;
 115 };
 116
 117 static bool nvmet_rdma_use_srq;
 118 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
 119 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
 120
 121 static DEFINE_IDA(nvmet_rdma_queue_ida);
 122 static LIST_HEAD(nvmet_rdma_queue_list);
 123 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
 124
 125 static LIST_HEAD(device_list);
 126 static DEFINE_MUTEX(device_list_mutex);
 127
 128 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
 129 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
 130 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 131 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
 132 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
 133 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
 134 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
 135                                 struct nvmet_rdma_rsp *r);
 136 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
 137                                 struct nvmet_rdma_rsp *r);
 138
 139 static const struct nvmet_fabrics_ops nvmet_rdma_ops;
 140
 141 static int num_pages(int len)
 142 {
 143         return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
 144 }
 145
 146 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
 147 {
 148         return nvme_is_write(rsp->req.cmd) &&
 149                 rsp->req.transfer_len &&
 150                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 151 }
 152
 153 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
 154 {
 155         return !nvme_is_write(rsp->req.cmd) &&
 156                 rsp->req.transfer_len &&
 157                 !rsp->req.cqe->status &&
 158                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 159 }
 160
 161 static inline struct nvmet_rdma_rsp *
 162 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
 163 {
 164         struct nvmet_rdma_rsp *rsp;
 165         unsigned long flags;
 166
 167         spin_lock_irqsave(&queue->rsps_lock, flags);
 168         rsp = list_first_entry_or_null(&queue->free_rsps,
 169                                 struct nvmet_rdma_rsp, free_list);
 170         if (likely(rsp))
 171                 list_del(&rsp->free_list);
 172         spin_unlock_irqrestore(&queue->rsps_lock, flags);
 173
 174         if (unlikely(!rsp)) {
 175                 int ret;
 176
 177                 rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
 178                 if (unlikely(!rsp))
 179                         return NULL;
 180                 ret = nvmet_rdma_alloc_rsp(queue->dev, rsp);
 181                 if (unlikely(ret)) {
 182                         kfree(rsp);
 183                         return NULL;
 184                 }
 185
 186                 rsp->allocated = true;
 187         }
 188
 189         return rsp;
 190 }
 191
 192 static inline void
 193 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
 194 {
 195         unsigned long flags;
 196
 197         if (unlikely(rsp->allocated)) {
 198                 nvmet_rdma_free_rsp(rsp->queue->dev, rsp);
 199                 kfree(rsp);
 200                 return;
 201         }
 202
 203         spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
 204         list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
 205         spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
 206 }
 207
 208 static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
 209                                 struct nvmet_rdma_cmd *c)
 210 {
 211         struct scatterlist *sg;
 212         struct ib_sge *sge;
 213         int i;
 214
 215         if (!ndev->inline_data_size)
 216                 return;
 217
 218         sg = c->inline_sg;
 219         sge = &c->sge[1];
 220
 221         for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
 222                 if (sge->length)
 223                         ib_dma_unmap_page(ndev->device, sge->addr,
 224                                         sge->length, DMA_FROM_DEVICE);
 225                 if (sg_page(sg))
 226                         __free_page(sg_page(sg));
 227         }
 228 }
 229
 230 static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
 231                                 struct nvmet_rdma_cmd *c)
 232 {
 233         struct scatterlist *sg;
 234         struct ib_sge *sge;
 235         struct page *pg;
 236         int len;
 237         int i;
 238
 239         if (!ndev->inline_data_size)
 240                 return 0;
 241
 242         sg = c->inline_sg;
 243         sg_init_table(sg, ndev->inline_page_count);
 244         sge = &c->sge[1];
 245         len = ndev->inline_data_size;
 246
 247         for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
 248                 pg = alloc_page(GFP_KERNEL);
 249                 if (!pg)
 250                         goto out_err;
 251                 sg_assign_page(sg, pg);
 252                 sge->addr = ib_dma_map_page(ndev->device,
 253                         pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
 254                 if (ib_dma_mapping_error(ndev->device, sge->addr))
 255                         goto out_err;
 256                 sge->length = min_t(int, len, PAGE_SIZE);
 257                 sge->lkey = ndev->pd->local_dma_lkey;
 258                 len -= sge->length;
 259         }
 260
 261         return 0;
 262 out_err:
 263         for (; i >= 0; i--, sg--, sge--) {
 264                 if (sge->length)
 265                         ib_dma_unmap_page(ndev->device, sge->addr,
 266                                         sge->length, DMA_FROM_DEVICE);
 267                 if (sg_page(sg))
 268                         __free_page(sg_page(sg));
 269         }
 270         return -ENOMEM;
 271 }
 272
 273 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
 274                         struct nvmet_rdma_cmd *c, bool admin)
 275 {
 276         /* NVMe command / RDMA RECV */
 277         c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
 278         if (!c->nvme_cmd)
 279                 goto out;
 280
 281         c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
 282                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 283         if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
 284                 goto out_free_cmd;
 285
 286         c->sge[0].length = sizeof(*c->nvme_cmd);
 287         c->sge[0].lkey = ndev->pd->local_dma_lkey;
 288
 289         if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
 290                 goto out_unmap_cmd;
 291
 292         c->cqe.done = nvmet_rdma_recv_done;
 293
 294         c->wr.wr_cqe = &c->cqe;
 295         c->wr.sg_list = c->sge;
 296         c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;
 297
 298         return 0;
 299
 300 out_unmap_cmd:
 301         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 302                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 303 out_free_cmd:
 304         kfree(c->nvme_cmd);
 305
 306 out:
 307         return -ENOMEM;
 308 }
 309
 310 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
 311                 struct nvmet_rdma_cmd *c, bool admin)
 312 {
 313         if (!admin)
 314                 nvmet_rdma_free_inline_pages(ndev, c);
 315         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 316                                 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 317         kfree(c->nvme_cmd);
 318 }
 319
 320 static struct nvmet_rdma_cmd *
 321 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
 322                 int nr_cmds, bool admin)
 323 {
 324         struct nvmet_rdma_cmd *cmds;
 325         int ret = -EINVAL, i;
 326
 327         cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
 328         if (!cmds)
 329                 goto out;
 330
 331         for (i = 0; i < nr_cmds; i++) {
 332                 ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
 333                 if (ret)
 334                         goto out_free;
 335         }
 336
 337         return cmds;
 338
 339 out_free:
 340         while (--i >= 0)
 341                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 342         kfree(cmds);
 343 out:
 344         return ERR_PTR(ret);
 345 }
 346
 347 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
 348                 struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
 349 {
 350         int i;
 351
 352         for (i = 0; i < nr_cmds; i++)
 353                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 354         kfree(cmds);
 355 }
 356
 357 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
 358                 struct nvmet_rdma_rsp *r)
 359 {
 360         /* NVMe CQE / RDMA SEND */
 361         r->req.cqe = kmalloc(sizeof(*r->req.cqe), GFP_KERNEL);
 362         if (!r->req.cqe)
 363                 goto out;
 364
 365         r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.cqe,
 366                         sizeof(*r->req.cqe), DMA_TO_DEVICE);
 367         if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
 368                 goto out_free_rsp;
 369
 370         r->req.p2p_client = &ndev->device->dev;
 371         r->send_sge.length = sizeof(*r->req.cqe);
 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.cqe);
 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.cqe), DMA_TO_DEVICE);
 396         kfree(r->req.cqe);
 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         rsp->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 (unlikely(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 (unlikely(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                 rsp->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         if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
 762                         &queue->nvme_sq, &nvmet_rdma_ops))
 763                 return;
 764
 765         status = nvmet_rdma_map_sgl(cmd);
 766         if (status)
 767                 goto out_err;
 768
 769         if (unlikely(!nvmet_rdma_execute_command(cmd))) {
 770                 spin_lock(&queue->rsp_wr_wait_lock);
 771                 list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
 772                 spin_unlock(&queue->rsp_wr_wait_lock);
 773         }
 774
 775         return;
 776
 777 out_err:
 778         nvmet_req_complete(&cmd->req, status);
 779 }
 780
 781 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
 782 {
 783         struct nvmet_rdma_cmd *cmd =
 784                 container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
 785         struct nvmet_rdma_queue *queue = cq->cq_context;
 786         struct nvmet_rdma_rsp *rsp;
 787
 788         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 789                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 790                         pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
 791                                 wc->wr_cqe, ib_wc_status_msg(wc->status),
 792                                 wc->status);
 793                         nvmet_rdma_error_comp(queue);
 794                 }
 795                 return;
 796         }
 797
 798         if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
 799                 pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
 800                 nvmet_rdma_error_comp(queue);
 801                 return;
 802         }
 803
 804         cmd->queue = queue;
 805         rsp = nvmet_rdma_get_rsp(queue);
 806         if (unlikely(!rsp)) {
 807                 /*
 808                  * we get here only under memory pressure,
 809                  * silently drop and have the host retry
 810                  * as we can't even fail it.
 811                  */
 812                 nvmet_rdma_post_recv(queue->dev, cmd);
 813                 return;
 814         }
 815         rsp->queue = queue;
 816         rsp->cmd = cmd;
 817         rsp->flags = 0;
 818         rsp->req.cmd = cmd->nvme_cmd;
 819         rsp->req.port = queue->port;
 820         rsp->n_rdma = 0;
 821
 822         if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
 823                 unsigned long flags;
 824
 825                 spin_lock_irqsave(&queue->state_lock, flags);
 826                 if (queue->state == NVMET_RDMA_Q_CONNECTING)
 827                         list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
 828                 else
 829                         nvmet_rdma_put_rsp(rsp);
 830                 spin_unlock_irqrestore(&queue->state_lock, flags);
 831                 return;
 832         }
 833
 834         nvmet_rdma_handle_command(queue, rsp);
 835 }
 836
 837 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_device *ndev)
 838 {
 839         if (!ndev->srq)
 840                 return;
 841
 842         nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
 843         ib_destroy_srq(ndev->srq);
 844 }
 845
 846 static int nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
 847 {
 848         struct ib_srq_init_attr srq_attr = { NULL, };
 849         struct ib_srq *srq;
 850         size_t srq_size;
 851         int ret, i;
 852
 853         srq_size = 4095;        /* XXX: tune */
 854
 855         srq_attr.attr.max_wr = srq_size;
 856         srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
 857         srq_attr.attr.srq_limit = 0;
 858         srq_attr.srq_type = IB_SRQT_BASIC;
 859         srq = ib_create_srq(ndev->pd, &srq_attr);
 860         if (IS_ERR(srq)) {
 861                 /*
 862                  * If SRQs aren't supported we just go ahead and use normal
 863                  * non-shared receive queues.
 864                  */
 865                 pr_info("SRQ requested but not supported.\n");
 866                 return 0;
 867         }
 868
 869         ndev->srq_cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
 870         if (IS_ERR(ndev->srq_cmds)) {
 871                 ret = PTR_ERR(ndev->srq_cmds);
 872                 goto out_destroy_srq;
 873         }
 874
 875         ndev->srq = srq;
 876         ndev->srq_size = srq_size;
 877
 878         for (i = 0; i < srq_size; i++) {
 879                 ret = nvmet_rdma_post_recv(ndev, &ndev->srq_cmds[i]);
 880                 if (ret)
 881                         goto out_free_cmds;
 882         }
 883
 884         return 0;
 885
 886 out_free_cmds:
 887         nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
 888 out_destroy_srq:
 889         ib_destroy_srq(srq);
 890         return ret;
 891 }
 892
 893 static void nvmet_rdma_free_dev(struct kref *ref)
 894 {
 895         struct nvmet_rdma_device *ndev =
 896                 container_of(ref, struct nvmet_rdma_device, ref);
 897
 898         mutex_lock(&device_list_mutex);
 899         list_del(&ndev->entry);
 900         mutex_unlock(&device_list_mutex);
 901
 902         nvmet_rdma_destroy_srq(ndev);
 903         ib_dealloc_pd(ndev->pd);
 904
 905         kfree(ndev);
 906 }
 907
 908 static struct nvmet_rdma_device *
 909 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
 910 {
 911         struct nvmet_port *port = cm_id->context;
 912         struct nvmet_rdma_device *ndev;
 913         int inline_page_count;
 914         int inline_sge_count;
 915         int ret;
 916
 917         mutex_lock(&device_list_mutex);
 918         list_for_each_entry(ndev, &device_list, entry) {
 919                 if (ndev->device->node_guid == cm_id->device->node_guid &&
 920                     kref_get_unless_zero(&ndev->ref))
 921                         goto out_unlock;
 922         }
 923
 924         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 925         if (!ndev)
 926                 goto out_err;
 927
 928         inline_page_count = num_pages(port->inline_data_size);
 929         inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
 930                                 cm_id->device->attrs.max_recv_sge) - 1;
 931         if (inline_page_count > inline_sge_count) {
 932                 pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
 933                         port->inline_data_size, cm_id->device->name,
 934                         inline_sge_count * PAGE_SIZE);
 935                 port->inline_data_size = inline_sge_count * PAGE_SIZE;
 936                 inline_page_count = inline_sge_count;
 937         }
 938         ndev->inline_data_size = port->inline_data_size;
 939         ndev->inline_page_count = inline_page_count;
 940         ndev->device = cm_id->device;
 941         kref_init(&ndev->ref);
 942
 943         ndev->pd = ib_alloc_pd(ndev->device, 0);
 944         if (IS_ERR(ndev->pd))
 945                 goto out_free_dev;
 946
 947         if (nvmet_rdma_use_srq) {
 948                 ret = nvmet_rdma_init_srq(ndev);
 949                 if (ret)
 950                         goto out_free_pd;
 951         }
 952
 953         list_add(&ndev->entry, &device_list);
 954 out_unlock:
 955         mutex_unlock(&device_list_mutex);
 956         pr_debug("added %s.\n", ndev->device->name);
 957         return ndev;
 958
 959 out_free_pd:
 960         ib_dealloc_pd(ndev->pd);
 961 out_free_dev:
 962         kfree(ndev);
 963 out_err:
 964         mutex_unlock(&device_list_mutex);
 965         return NULL;
 966 }
 967
 968 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
 969 {
 970         struct ib_qp_init_attr qp_attr;
 971         struct nvmet_rdma_device *ndev = queue->dev;
 972         int comp_vector, nr_cqe, ret, i;
 973
 974         /*
 975          * Spread the io queues across completion vectors,
 976          * but still keep all admin queues on vector 0.
 977          */
 978         comp_vector = !queue->host_qid ? 0 :
 979                 queue->idx % ndev->device->num_comp_vectors;
 980
 981         /*
 982          * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
 983          */
 984         nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
 985
 986         queue->cq = ib_alloc_cq(ndev->device, queue,
 987                         nr_cqe + 1, comp_vector,
 988                         IB_POLL_WORKQUEUE);
 989         if (IS_ERR(queue->cq)) {
 990                 ret = PTR_ERR(queue->cq);
 991                 pr_err("failed to create CQ cqe= %d ret= %d\n",
 992                        nr_cqe + 1, ret);
 993                 goto out;
 994         }
 995
 996         memset(&qp_attr, 0, sizeof(qp_attr));
 997         qp_attr.qp_context = queue;
 998         qp_attr.event_handler = nvmet_rdma_qp_event;
 999         qp_attr.send_cq = queue->cq;
1000         qp_attr.recv_cq = queue->cq;
1001         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1002         qp_attr.qp_type = IB_QPT_RC;
1003         /* +1 for drain */
1004         qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
1005         qp_attr.cap.max_rdma_ctxs = queue->send_queue_size;
1006         qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
1007                                         ndev->device->attrs.max_send_sge);
1008
1009         if (ndev->srq) {
1010                 qp_attr.srq = ndev->srq;
1011         } else {
1012                 /* +1 for drain */
1013                 qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
1014                 qp_attr.cap.max_recv_sge = 1 + ndev->inline_page_count;
1015         }
1016
1017         ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
1018         if (ret) {
1019                 pr_err("failed to create_qp ret= %d\n", ret);
1020                 goto err_destroy_cq;
1021         }
1022
1023         atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
1024
1025         pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
1026                  __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
1027                  qp_attr.cap.max_send_wr, queue->cm_id);
1028
1029         if (!ndev->srq) {
1030                 for (i = 0; i < queue->recv_queue_size; i++) {
1031                         queue->cmds[i].queue = queue;
1032                         ret = nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
1033                         if (ret)
1034                                 goto err_destroy_qp;
1035                 }
1036         }
1037
1038 out:
1039         return ret;
1040
1041 err_destroy_qp:
1042         rdma_destroy_qp(queue->cm_id);
1043 err_destroy_cq:
1044         ib_free_cq(queue->cq);
1045         goto out;
1046 }
1047
1048 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1049 {
1050         struct ib_qp *qp = queue->cm_id->qp;
1051
1052         ib_drain_qp(qp);
1053         rdma_destroy_id(queue->cm_id);
1054         ib_destroy_qp(qp);
1055         ib_free_cq(queue->cq);
1056 }
1057
1058 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1059 {
1060         pr_debug("freeing queue %d\n", queue->idx);
1061
1062         nvmet_sq_destroy(&queue->nvme_sq);
1063
1064         nvmet_rdma_destroy_queue_ib(queue);
1065         if (!queue->dev->srq) {
1066                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1067                                 queue->recv_queue_size,
1068                                 !queue->host_qid);
1069         }
1070         nvmet_rdma_free_rsps(queue);
1071         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1072         kfree(queue);
1073 }
1074
1075 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1076 {
1077         struct nvmet_rdma_queue *queue =
1078                 container_of(w, struct nvmet_rdma_queue, release_work);
1079         struct nvmet_rdma_device *dev = queue->dev;
1080
1081         nvmet_rdma_free_queue(queue);
1082
1083         kref_put(&dev->ref, nvmet_rdma_free_dev);
1084 }
1085
1086 static int
1087 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1088                                 struct nvmet_rdma_queue *queue)
1089 {
1090         struct nvme_rdma_cm_req *req;
1091
1092         req = (struct nvme_rdma_cm_req *)conn->private_data;
1093         if (!req || conn->private_data_len == 0)
1094                 return NVME_RDMA_CM_INVALID_LEN;
1095
1096         if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1097                 return NVME_RDMA_CM_INVALID_RECFMT;
1098
1099         queue->host_qid = le16_to_cpu(req->qid);
1100
1101         /*
1102          * req->hsqsize corresponds to our recv queue size plus 1
1103          * req->hrqsize corresponds to our send queue size
1104          */
1105         queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1106         queue->send_queue_size = le16_to_cpu(req->hrqsize);
1107
1108         if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
1109                 return NVME_RDMA_CM_INVALID_HSQSIZE;
1110
1111         /* XXX: Should we enforce some kind of max for IO queues? */
1112
1113         return 0;
1114 }
1115
1116 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1117                                 enum nvme_rdma_cm_status status)
1118 {
1119         struct nvme_rdma_cm_rej rej;
1120
1121         pr_debug("rejecting connect request: status %d (%s)\n",
1122                  status, nvme_rdma_cm_msg(status));
1123
1124         rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1125         rej.sts = cpu_to_le16(status);
1126
1127         return rdma_reject(cm_id, (void *)&rej, sizeof(rej));
1128 }
1129
1130 static struct nvmet_rdma_queue *
1131 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1132                 struct rdma_cm_id *cm_id,
1133                 struct rdma_cm_event *event)
1134 {
1135         struct nvmet_rdma_queue *queue;
1136         int ret;
1137
1138         queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1139         if (!queue) {
1140                 ret = NVME_RDMA_CM_NO_RSC;
1141                 goto out_reject;
1142         }
1143
1144         ret = nvmet_sq_init(&queue->nvme_sq);
1145         if (ret) {
1146                 ret = NVME_RDMA_CM_NO_RSC;
1147                 goto out_free_queue;
1148         }
1149
1150         ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1151         if (ret)
1152                 goto out_destroy_sq;
1153
1154         /*
1155          * Schedules the actual release because calling rdma_destroy_id from
1156          * inside a CM callback would trigger a deadlock. (great API design..)
1157          */
1158         INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1159         queue->dev = ndev;
1160         queue->cm_id = cm_id;
1161
1162         spin_lock_init(&queue->state_lock);
1163         queue->state = NVMET_RDMA_Q_CONNECTING;
1164         INIT_LIST_HEAD(&queue->rsp_wait_list);
1165         INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1166         spin_lock_init(&queue->rsp_wr_wait_lock);
1167         INIT_LIST_HEAD(&queue->free_rsps);
1168         spin_lock_init(&queue->rsps_lock);
1169         INIT_LIST_HEAD(&queue->queue_list);
1170
1171         queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1172         if (queue->idx < 0) {
1173                 ret = NVME_RDMA_CM_NO_RSC;
1174                 goto out_destroy_sq;
1175         }
1176
1177         ret = nvmet_rdma_alloc_rsps(queue);
1178         if (ret) {
1179                 ret = NVME_RDMA_CM_NO_RSC;
1180                 goto out_ida_remove;
1181         }
1182
1183         if (!ndev->srq) {
1184                 queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1185                                 queue->recv_queue_size,
1186                                 !queue->host_qid);
1187                 if (IS_ERR(queue->cmds)) {
1188                         ret = NVME_RDMA_CM_NO_RSC;
1189                         goto out_free_responses;
1190                 }
1191         }
1192
1193         ret = nvmet_rdma_create_queue_ib(queue);
1194         if (ret) {
1195                 pr_err("%s: creating RDMA queue failed (%d).\n",
1196                         __func__, ret);
1197                 ret = NVME_RDMA_CM_NO_RSC;
1198                 goto out_free_cmds;
1199         }
1200
1201         return queue;
1202
1203 out_free_cmds:
1204         if (!ndev->srq) {
1205                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1206                                 queue->recv_queue_size,
1207                                 !queue->host_qid);
1208         }
1209 out_free_responses:
1210         nvmet_rdma_free_rsps(queue);
1211 out_ida_remove:
1212         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1213 out_destroy_sq:
1214         nvmet_sq_destroy(&queue->nvme_sq);
1215 out_free_queue:
1216         kfree(queue);
1217 out_reject:
1218         nvmet_rdma_cm_reject(cm_id, ret);
1219         return NULL;
1220 }
1221
1222 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1223 {
1224         struct nvmet_rdma_queue *queue = priv;
1225
1226         switch (event->event) {
1227         case IB_EVENT_COMM_EST:
1228                 rdma_notify(queue->cm_id, event->event);
1229                 break;
1230         default:
1231                 pr_err("received IB QP event: %s (%d)\n",
1232                        ib_event_msg(event->event), event->event);
1233                 break;
1234         }
1235 }
1236
1237 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1238                 struct nvmet_rdma_queue *queue,
1239                 struct rdma_conn_param *p)
1240 {
1241         struct rdma_conn_param  param = { };
1242         struct nvme_rdma_cm_rep priv = { };
1243         int ret = -ENOMEM;
1244
1245         param.rnr_retry_count = 7;
1246         param.flow_control = 1;
1247         param.initiator_depth = min_t(u8, p->initiator_depth,
1248                 queue->dev->device->attrs.max_qp_init_rd_atom);
1249         param.private_data = &priv;
1250         param.private_data_len = sizeof(priv);
1251         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1252         priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1253
1254         ret = rdma_accept(cm_id, &param);
1255         if (ret)
1256                 pr_err("rdma_accept failed (error code = %d)\n", ret);
1257
1258         return ret;
1259 }
1260
1261 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1262                 struct rdma_cm_event *event)
1263 {
1264         struct nvmet_rdma_device *ndev;
1265         struct nvmet_rdma_queue *queue;
1266         int ret = -EINVAL;
1267
1268         ndev = nvmet_rdma_find_get_device(cm_id);
1269         if (!ndev) {
1270                 nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1271                 return -ECONNREFUSED;
1272         }
1273
1274         queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1275         if (!queue) {
1276                 ret = -ENOMEM;
1277                 goto put_device;
1278         }
1279         queue->port = cm_id->context;
1280
1281         if (queue->host_qid == 0) {
1282                 /* Let inflight controller teardown complete */
1283                 flush_scheduled_work();
1284         }
1285
1286         ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1287         if (ret) {
1288                 schedule_work(&queue->release_work);
1289                 /* Destroying rdma_cm id is not needed here */
1290                 return 0;
1291         }
1292
1293         mutex_lock(&nvmet_rdma_queue_mutex);
1294         list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1295         mutex_unlock(&nvmet_rdma_queue_mutex);
1296
1297         return 0;
1298
1299 put_device:
1300         kref_put(&ndev->ref, nvmet_rdma_free_dev);
1301
1302         return ret;
1303 }
1304
1305 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1306 {
1307         unsigned long flags;
1308
1309         spin_lock_irqsave(&queue->state_lock, flags);
1310         if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1311                 pr_warn("trying to establish a connected queue\n");
1312                 goto out_unlock;
1313         }
1314         queue->state = NVMET_RDMA_Q_LIVE;
1315
1316         while (!list_empty(&queue->rsp_wait_list)) {
1317                 struct nvmet_rdma_rsp *cmd;
1318
1319                 cmd = list_first_entry(&queue->rsp_wait_list,
1320                                         struct nvmet_rdma_rsp, wait_list);
1321                 list_del(&cmd->wait_list);
1322
1323                 spin_unlock_irqrestore(&queue->state_lock, flags);
1324                 nvmet_rdma_handle_command(queue, cmd);
1325                 spin_lock_irqsave(&queue->state_lock, flags);
1326         }
1327
1328 out_unlock:
1329         spin_unlock_irqrestore(&queue->state_lock, flags);
1330 }
1331
1332 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1333 {
1334         bool disconnect = false;
1335         unsigned long flags;
1336
1337         pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1338
1339         spin_lock_irqsave(&queue->state_lock, flags);
1340         switch (queue->state) {
1341         case NVMET_RDMA_Q_CONNECTING:
1342         case NVMET_RDMA_Q_LIVE:
1343                 queue->state = NVMET_RDMA_Q_DISCONNECTING;
1344                 disconnect = true;
1345                 break;
1346         case NVMET_RDMA_Q_DISCONNECTING:
1347                 break;
1348         }
1349         spin_unlock_irqrestore(&queue->state_lock, flags);
1350
1351         if (disconnect) {
1352                 rdma_disconnect(queue->cm_id);
1353                 schedule_work(&queue->release_work);
1354         }
1355 }
1356
1357 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1358 {
1359         bool disconnect = false;
1360
1361         mutex_lock(&nvmet_rdma_queue_mutex);
1362         if (!list_empty(&queue->queue_list)) {
1363                 list_del_init(&queue->queue_list);
1364                 disconnect = true;
1365         }
1366         mutex_unlock(&nvmet_rdma_queue_mutex);
1367
1368         if (disconnect)
1369                 __nvmet_rdma_queue_disconnect(queue);
1370 }
1371
1372 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1373                 struct nvmet_rdma_queue *queue)
1374 {
1375         WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1376
1377         mutex_lock(&nvmet_rdma_queue_mutex);
1378         if (!list_empty(&queue->queue_list))
1379                 list_del_init(&queue->queue_list);
1380         mutex_unlock(&nvmet_rdma_queue_mutex);
1381
1382         pr_err("failed to connect queue %d\n", queue->idx);
1383         schedule_work(&queue->release_work);
1384 }
1385
1386 /**
1387  * nvme_rdma_device_removal() - Handle RDMA device removal
1388  * @cm_id:      rdma_cm id, used for nvmet port
1389  * @queue:      nvmet rdma queue (cm id qp_context)
1390  *
1391  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1392  * to unplug. Note that this event can be generated on a normal
1393  * queue cm_id and/or a device bound listener cm_id (where in this
1394  * case queue will be null).
1395  *
1396  * We registered an ib_client to handle device removal for queues,
1397  * so we only need to handle the listening port cm_ids. In this case
1398  * we nullify the priv to prevent double cm_id destruction and destroying
1399  * the cm_id implicitely by returning a non-zero rc to the callout.
1400  */
1401 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1402                 struct nvmet_rdma_queue *queue)
1403 {
1404         struct nvmet_port *port;
1405
1406         if (queue) {
1407                 /*
1408                  * This is a queue cm_id. we have registered
1409                  * an ib_client to handle queues removal
1410                  * so don't interfear and just return.
1411                  */
1412                 return 0;
1413         }
1414
1415         port = cm_id->context;
1416
1417         /*
1418          * This is a listener cm_id. Make sure that
1419          * future remove_port won't invoke a double
1420          * cm_id destroy. use atomic xchg to make sure
1421          * we don't compete with remove_port.
1422          */
1423         if (xchg(&port->priv, NULL) != cm_id)
1424                 return 0;
1425
1426         /*
1427          * We need to return 1 so that the core will destroy
1428          * it's own ID.  What a great API design..
1429          */
1430         return 1;
1431 }
1432
1433 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1434                 struct rdma_cm_event *event)
1435 {
1436         struct nvmet_rdma_queue *queue = NULL;
1437         int ret = 0;
1438
1439         if (cm_id->qp)
1440                 queue = cm_id->qp->qp_context;
1441
1442         pr_debug("%s (%d): status %d id %p\n",
1443                 rdma_event_msg(event->event), event->event,
1444                 event->status, cm_id);
1445
1446         switch (event->event) {
1447         case RDMA_CM_EVENT_CONNECT_REQUEST:
1448                 ret = nvmet_rdma_queue_connect(cm_id, event);
1449                 break;
1450         case RDMA_CM_EVENT_ESTABLISHED:
1451                 nvmet_rdma_queue_established(queue);
1452                 break;
1453         case RDMA_CM_EVENT_ADDR_CHANGE:
1454         case RDMA_CM_EVENT_DISCONNECTED:
1455         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1456                 nvmet_rdma_queue_disconnect(queue);
1457                 break;
1458         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1459                 ret = nvmet_rdma_device_removal(cm_id, queue);
1460                 break;
1461         case RDMA_CM_EVENT_REJECTED:
1462                 pr_debug("Connection rejected: %s\n",
1463                          rdma_reject_msg(cm_id, event->status));
1464                 /* FALLTHROUGH */
1465         case RDMA_CM_EVENT_UNREACHABLE:
1466         case RDMA_CM_EVENT_CONNECT_ERROR:
1467                 nvmet_rdma_queue_connect_fail(cm_id, queue);
1468                 break;
1469         default:
1470                 pr_err("received unrecognized RDMA CM event %d\n",
1471                         event->event);
1472                 break;
1473         }
1474
1475         return ret;
1476 }
1477
1478 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1479 {
1480         struct nvmet_rdma_queue *queue;
1481
1482 restart:
1483         mutex_lock(&nvmet_rdma_queue_mutex);
1484         list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1485                 if (queue->nvme_sq.ctrl == ctrl) {
1486                         list_del_init(&queue->queue_list);
1487                         mutex_unlock(&nvmet_rdma_queue_mutex);
1488
1489                         __nvmet_rdma_queue_disconnect(queue);
1490                         goto restart;
1491                 }
1492         }
1493         mutex_unlock(&nvmet_rdma_queue_mutex);
1494 }
1495
1496 static int nvmet_rdma_add_port(struct nvmet_port *port)
1497 {
1498         struct rdma_cm_id *cm_id;
1499         struct sockaddr_storage addr = { };
1500         __kernel_sa_family_t af;
1501         int ret;
1502
1503         switch (port->disc_addr.adrfam) {
1504         case NVMF_ADDR_FAMILY_IP4:
1505                 af = AF_INET;
1506                 break;
1507         case NVMF_ADDR_FAMILY_IP6:
1508                 af = AF_INET6;
1509                 break;
1510         default:
1511                 pr_err("address family %d not supported\n",
1512                                 port->disc_addr.adrfam);
1513                 return -EINVAL;
1514         }
1515
1516         if (port->inline_data_size < 0) {
1517                 port->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
1518         } else if (port->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
1519                 pr_warn("inline_data_size %u is too large, reducing to %u\n",
1520                         port->inline_data_size,
1521                         NVMET_RDMA_MAX_INLINE_DATA_SIZE);
1522                 port->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
1523         }
1524
1525         ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
1526                         port->disc_addr.trsvcid, &addr);
1527         if (ret) {
1528                 pr_err("malformed ip/port passed: %s:%s\n",
1529                         port->disc_addr.traddr, port->disc_addr.trsvcid);
1530                 return ret;
1531         }
1532
1533         cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1534                         RDMA_PS_TCP, IB_QPT_RC);
1535         if (IS_ERR(cm_id)) {
1536                 pr_err("CM ID creation failed\n");
1537                 return PTR_ERR(cm_id);
1538         }
1539
1540         /*
1541          * Allow both IPv4 and IPv6 sockets to bind a single port
1542          * at the same time.
1543          */
1544         ret = rdma_set_afonly(cm_id, 1);
1545         if (ret) {
1546                 pr_err("rdma_set_afonly failed (%d)\n", ret);
1547                 goto out_destroy_id;
1548         }
1549
1550         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
1551         if (ret) {
1552                 pr_err("binding CM ID to %pISpcs failed (%d)\n",
1553                         (struct sockaddr *)&addr, ret);
1554                 goto out_destroy_id;
1555         }
1556
1557         ret = rdma_listen(cm_id, 128);
1558         if (ret) {
1559                 pr_err("listening to %pISpcs failed (%d)\n",
1560                         (struct sockaddr *)&addr, ret);
1561                 goto out_destroy_id;
1562         }
1563
1564         pr_info("enabling port %d (%pISpcs)\n",
1565                 le16_to_cpu(port->disc_addr.portid), (struct sockaddr *)&addr);
1566         port->priv = cm_id;
1567         return 0;
1568
1569 out_destroy_id:
1570         rdma_destroy_id(cm_id);
1571         return ret;
1572 }
1573
1574 static void nvmet_rdma_remove_port(struct nvmet_port *port)
1575 {
1576         struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
1577
1578         if (cm_id)
1579                 rdma_destroy_id(cm_id);
1580 }
1581
1582 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
1583                 struct nvmet_port *port, char *traddr)
1584 {
1585         struct rdma_cm_id *cm_id = port->priv;
1586
1587         if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
1588                 struct nvmet_rdma_rsp *rsp =
1589                         container_of(req, struct nvmet_rdma_rsp, req);
1590                 struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
1591                 struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;
1592
1593                 sprintf(traddr, "%pISc", addr);
1594         } else {
1595                 memcpy(traddr, port->disc_addr.traddr, NVMF_TRADDR_SIZE);
1596         }
1597 }
1598
1599 static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
1600         .owner                  = THIS_MODULE,
1601         .type                   = NVMF_TRTYPE_RDMA,
1602         .msdbd                  = 1,
1603         .has_keyed_sgls         = 1,
1604         .add_port               = nvmet_rdma_add_port,
1605         .remove_port            = nvmet_rdma_remove_port,
1606         .queue_response         = nvmet_rdma_queue_response,
1607         .delete_ctrl            = nvmet_rdma_delete_ctrl,
1608         .disc_traddr            = nvmet_rdma_disc_port_addr,
1609 };
1610
1611 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1612 {
1613         struct nvmet_rdma_queue *queue, *tmp;
1614         struct nvmet_rdma_device *ndev;
1615         bool found = false;
1616
1617         mutex_lock(&device_list_mutex);
1618         list_for_each_entry(ndev, &device_list, entry) {
1619                 if (ndev->device == ib_device) {
1620                         found = true;
1621                         break;
1622                 }
1623         }
1624         mutex_unlock(&device_list_mutex);
1625
1626         if (!found)
1627                 return;
1628
1629         /*
1630          * IB Device that is used by nvmet controllers is being removed,
1631          * delete all queues using this device.
1632          */
1633         mutex_lock(&nvmet_rdma_queue_mutex);
1634         list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
1635                                  queue_list) {
1636                 if (queue->dev->device != ib_device)
1637                         continue;
1638
1639                 pr_info("Removing queue %d\n", queue->idx);
1640                 list_del_init(&queue->queue_list);
1641                 __nvmet_rdma_queue_disconnect(queue);
1642         }
1643         mutex_unlock(&nvmet_rdma_queue_mutex);
1644
1645         flush_scheduled_work();
1646 }
1647
1648 static struct ib_client nvmet_rdma_ib_client = {
1649         .name   = "nvmet_rdma",
1650         .remove = nvmet_rdma_remove_one
1651 };
1652
1653 static int __init nvmet_rdma_init(void)
1654 {
1655         int ret;
1656
1657         ret = ib_register_client(&nvmet_rdma_ib_client);
1658         if (ret)
1659                 return ret;
1660
1661         ret = nvmet_register_transport(&nvmet_rdma_ops);
1662         if (ret)
1663                 goto err_ib_client;
1664
1665         return 0;
1666
1667 err_ib_client:
1668         ib_unregister_client(&nvmet_rdma_ib_client);
1669         return ret;
1670 }
1671
1672 static void __exit nvmet_rdma_exit(void)
1673 {
1674         nvmet_unregister_transport(&nvmet_rdma_ops);
1675         ib_unregister_client(&nvmet_rdma_ib_client);
1676         WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
1677         ida_destroy(&nvmet_rdma_queue_ida);
1678 }
1679
1680 module_init(nvmet_rdma_init);
1681 module_exit(nvmet_rdma_exit);
1682
1683 MODULE_LICENSE("GPL v2");
1684 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */