drivers/nvme/target/rdma.c

   1 /*
   2  * NVMe over Fabrics RDMA target.
   3  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  */
  14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  15 #include <linux/atomic.h>
  16 #include <linux/ctype.h>
  17 #include <linux/delay.h>
  18 #include <linux/err.h>
  19 #include <linux/init.h>
  20 #include <linux/module.h>
  21 #include <linux/nvme.h>
  22 #include <linux/slab.h>
  23 #include <linux/string.h>
  24 #include <linux/wait.h>
  25 #include <linux/inet.h>
  26 #include <asm/unaligned.h>
  27
  28 #include <rdma/ib_verbs.h>
  29 #include <rdma/rdma_cm.h>
  30 #include <rdma/rw.h>
  31
  32 #include <linux/nvme-rdma.h>
  33 #include "nvmet.h"
  34
  35 /*
  36  * We allow up to a page of inline data to go with the SQE
  37  */
  38 #define NVMET_RDMA_INLINE_DATA_SIZE     PAGE_SIZE
  39
  40 struct nvmet_rdma_cmd {
  41         struct ib_sge           sge[2];
  42         struct ib_cqe           cqe;
  43         struct ib_recv_wr       wr;
  44         struct scatterlist      inline_sg;
  45         struct page             *inline_page;
  46         struct nvme_command     *nvme_cmd;
  47         struct nvmet_rdma_queue *queue;
  48 };
  49
  50 enum {
  51         NVMET_RDMA_REQ_INLINE_DATA      = (1 << 0),
  52         NVMET_RDMA_REQ_INVALIDATE_RKEY  = (1 << 1),
  53 };
  54
  55 struct nvmet_rdma_rsp {
  56         struct ib_sge           send_sge;
  57         struct ib_cqe           send_cqe;
  58         struct ib_send_wr       send_wr;
  59
  60         struct nvmet_rdma_cmd   *cmd;
  61         struct nvmet_rdma_queue *queue;
  62
  63         struct ib_cqe           read_cqe;
  64         struct rdma_rw_ctx      rw;
  65
  66         struct nvmet_req        req;
  67
  68         bool                    allocated;
  69         u8                      n_rdma;
  70         u32                     flags;
  71         u32                     invalidate_rkey;
  72
  73         struct list_head        wait_list;
  74         struct list_head        free_list;
  75 };
  76
  77 enum nvmet_rdma_queue_state {
  78         NVMET_RDMA_Q_CONNECTING,
  79         NVMET_RDMA_Q_LIVE,
  80         NVMET_RDMA_Q_DISCONNECTING,
  81         NVMET_RDMA_IN_DEVICE_REMOVAL,
  82 };
  83
  84 struct nvmet_rdma_queue {
  85         struct rdma_cm_id       *cm_id;
  86         struct nvmet_port       *port;
  87         struct ib_cq            *cq;
  88         atomic_t                sq_wr_avail;
  89         struct nvmet_rdma_device *dev;
  90         spinlock_t              state_lock;
  91         enum nvmet_rdma_queue_state state;
  92         struct nvmet_cq         nvme_cq;
  93         struct nvmet_sq         nvme_sq;
  94
  95         struct nvmet_rdma_rsp   *rsps;
  96         struct list_head        free_rsps;
  97         spinlock_t              rsps_lock;
  98         struct nvmet_rdma_cmd   *cmds;
  99
 100         struct work_struct      release_work;
 101         struct list_head        rsp_wait_list;
 102         struct list_head        rsp_wr_wait_list;
 103         spinlock_t              rsp_wr_wait_lock;
 104
 105         int                     idx;
 106         int                     host_qid;
 107         int                     recv_queue_size;
 108         int                     send_queue_size;
 109
 110         struct list_head        queue_list;
 111 };
 112
 113 struct nvmet_rdma_device {
 114         struct ib_device        *device;
 115         struct ib_pd            *pd;
 116         struct ib_srq           *srq;
 117         struct nvmet_rdma_cmd   *srq_cmds;
 118         size_t                  srq_size;
 119         struct kref             ref;
 120         struct list_head        entry;
 121 };
 122
 123 static bool nvmet_rdma_use_srq;
 124 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
 125 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
 126
 127 static DEFINE_IDA(nvmet_rdma_queue_ida);
 128 static LIST_HEAD(nvmet_rdma_queue_list);
 129 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
 130
 131 static LIST_HEAD(device_list);
 132 static DEFINE_MUTEX(device_list_mutex);
 133
 134 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
 135 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
 136 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 137 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
 138 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
 139 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
 140 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
 141                                 struct nvmet_rdma_rsp *r);
 142 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
 143                                 struct nvmet_rdma_rsp *r);
 144
 145 static struct nvmet_fabrics_ops nvmet_rdma_ops;
 146
 147 /* XXX: really should move to a generic header sooner or later.. */
 148 static inline u32 get_unaligned_le24(const u8 *p)
 149 {
 150         return (u32)p[0] | (u32)p[1] << 8 | (u32)p[2] << 16;
 151 }
 152
 153 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
 154 {
 155         return nvme_is_write(rsp->req.cmd) &&
 156                 rsp->req.data_len &&
 157                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 158 }
 159
 160 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
 161 {
 162         return !nvme_is_write(rsp->req.cmd) &&
 163                 rsp->req.data_len &&
 164                 !rsp->req.rsp->status &&
 165                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 166 }
 167
 168 static inline struct nvmet_rdma_rsp *
 169 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
 170 {
 171         struct nvmet_rdma_rsp *rsp;
 172         unsigned long flags;
 173
 174         spin_lock_irqsave(&queue->rsps_lock, flags);
 175         rsp = list_first_entry_or_null(&queue->free_rsps,
 176                                 struct nvmet_rdma_rsp, free_list);
 177         if (likely(rsp))
 178                 list_del(&rsp->free_list);
 179         spin_unlock_irqrestore(&queue->rsps_lock, flags);
 180
 181         if (unlikely(!rsp)) {
 182                 int ret;
 183
 184                 rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
 185                 if (unlikely(!rsp))
 186                         return NULL;
 187                 ret = nvmet_rdma_alloc_rsp(queue->dev, rsp);
 188                 if (unlikely(ret)) {
 189                         kfree(rsp);
 190                         return NULL;
 191                 }
 192
 193                 rsp->allocated = true;
 194         }
 195
 196         return rsp;
 197 }
 198
 199 static inline void
 200 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
 201 {
 202         unsigned long flags;
 203
 204         if (unlikely(rsp->allocated)) {
 205                 nvmet_rdma_free_rsp(rsp->queue->dev, rsp);
 206                 kfree(rsp);
 207                 return;
 208         }
 209
 210         spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
 211         list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
 212         spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
 213 }
 214
 215 static void nvmet_rdma_free_sgl(struct scatterlist *sgl, unsigned int nents)
 216 {
 217         struct scatterlist *sg;
 218         int count;
 219
 220         if (!sgl || !nents)
 221                 return;
 222
 223         for_each_sg(sgl, sg, nents, count)
 224                 __free_page(sg_page(sg));
 225         kfree(sgl);
 226 }
 227
 228 static int nvmet_rdma_alloc_sgl(struct scatterlist **sgl, unsigned int *nents,
 229                 u32 length)
 230 {
 231         struct scatterlist *sg;
 232         struct page *page;
 233         unsigned int nent;
 234         int i = 0;
 235
 236         nent = DIV_ROUND_UP(length, PAGE_SIZE);
 237         sg = kmalloc_array(nent, sizeof(struct scatterlist), GFP_KERNEL);
 238         if (!sg)
 239                 goto out;
 240
 241         sg_init_table(sg, nent);
 242
 243         while (length) {
 244                 u32 page_len = min_t(u32, length, PAGE_SIZE);
 245
 246                 page = alloc_page(GFP_KERNEL);
 247                 if (!page)
 248                         goto out_free_pages;
 249
 250                 sg_set_page(&sg[i], page, page_len, 0);
 251                 length -= page_len;
 252                 i++;
 253         }
 254         *sgl = sg;
 255         *nents = nent;
 256         return 0;
 257
 258 out_free_pages:
 259         while (i > 0) {
 260                 i--;
 261                 __free_page(sg_page(&sg[i]));
 262         }
 263         kfree(sg);
 264 out:
 265         return NVME_SC_INTERNAL;
 266 }
 267
 268 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
 269                         struct nvmet_rdma_cmd *c, bool admin)
 270 {
 271         /* NVMe command / RDMA RECV */
 272         c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
 273         if (!c->nvme_cmd)
 274                 goto out;
 275
 276         c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
 277                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 278         if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
 279                 goto out_free_cmd;
 280
 281         c->sge[0].length = sizeof(*c->nvme_cmd);
 282         c->sge[0].lkey = ndev->pd->local_dma_lkey;
 283
 284         if (!admin) {
 285                 c->inline_page = alloc_pages(GFP_KERNEL,
 286                                 get_order(NVMET_RDMA_INLINE_DATA_SIZE));
 287                 if (!c->inline_page)
 288                         goto out_unmap_cmd;
 289                 c->sge[1].addr = ib_dma_map_page(ndev->device,
 290                                 c->inline_page, 0, NVMET_RDMA_INLINE_DATA_SIZE,
 291                                 DMA_FROM_DEVICE);
 292                 if (ib_dma_mapping_error(ndev->device, c->sge[1].addr))
 293                         goto out_free_inline_page;
 294                 c->sge[1].length = NVMET_RDMA_INLINE_DATA_SIZE;
 295                 c->sge[1].lkey = ndev->pd->local_dma_lkey;
 296         }
 297
 298         c->cqe.done = nvmet_rdma_recv_done;
 299
 300         c->wr.wr_cqe = &c->cqe;
 301         c->wr.sg_list = c->sge;
 302         c->wr.num_sge = admin ? 1 : 2;
 303
 304         return 0;
 305
 306 out_free_inline_page:
 307         if (!admin) {
 308                 __free_pages(c->inline_page,
 309                                 get_order(NVMET_RDMA_INLINE_DATA_SIZE));
 310         }
 311 out_unmap_cmd:
 312         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 313                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 314 out_free_cmd:
 315         kfree(c->nvme_cmd);
 316
 317 out:
 318         return -ENOMEM;
 319 }
 320
 321 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
 322                 struct nvmet_rdma_cmd *c, bool admin)
 323 {
 324         if (!admin) {
 325                 ib_dma_unmap_page(ndev->device, c->sge[1].addr,
 326                                 NVMET_RDMA_INLINE_DATA_SIZE, DMA_FROM_DEVICE);
 327                 __free_pages(c->inline_page,
 328                                 get_order(NVMET_RDMA_INLINE_DATA_SIZE));
 329         }
 330         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 331                                 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 332         kfree(c->nvme_cmd);
 333 }
 334
 335 static struct nvmet_rdma_cmd *
 336 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
 337                 int nr_cmds, bool admin)
 338 {
 339         struct nvmet_rdma_cmd *cmds;
 340         int ret = -EINVAL, i;
 341
 342         cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
 343         if (!cmds)
 344                 goto out;
 345
 346         for (i = 0; i < nr_cmds; i++) {
 347                 ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
 348                 if (ret)
 349                         goto out_free;
 350         }
 351
 352         return cmds;
 353
 354 out_free:
 355         while (--i >= 0)
 356                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 357         kfree(cmds);
 358 out:
 359         return ERR_PTR(ret);
 360 }
 361
 362 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
 363                 struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
 364 {
 365         int i;
 366
 367         for (i = 0; i < nr_cmds; i++)
 368                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 369         kfree(cmds);
 370 }
 371
 372 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
 373                 struct nvmet_rdma_rsp *r)
 374 {
 375         /* NVMe CQE / RDMA SEND */
 376         r->req.rsp = kmalloc(sizeof(*r->req.rsp), GFP_KERNEL);
 377         if (!r->req.rsp)
 378                 goto out;
 379
 380         r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.rsp,
 381                         sizeof(*r->req.rsp), DMA_TO_DEVICE);
 382         if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
 383                 goto out_free_rsp;
 384
 385         r->send_sge.length = sizeof(*r->req.rsp);
 386         r->send_sge.lkey = ndev->pd->local_dma_lkey;
 387
 388         r->send_cqe.done = nvmet_rdma_send_done;
 389
 390         r->send_wr.wr_cqe = &r->send_cqe;
 391         r->send_wr.sg_list = &r->send_sge;
 392         r->send_wr.num_sge = 1;
 393         r->send_wr.send_flags = IB_SEND_SIGNALED;
 394
 395         /* Data In / RDMA READ */
 396         r->read_cqe.done = nvmet_rdma_read_data_done;
 397         return 0;
 398
 399 out_free_rsp:
 400         kfree(r->req.rsp);
 401 out:
 402         return -ENOMEM;
 403 }
 404
 405 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
 406                 struct nvmet_rdma_rsp *r)
 407 {
 408         ib_dma_unmap_single(ndev->device, r->send_sge.addr,
 409                                 sizeof(*r->req.rsp), DMA_TO_DEVICE);
 410         kfree(r->req.rsp);
 411 }
 412
 413 static int
 414 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
 415 {
 416         struct nvmet_rdma_device *ndev = queue->dev;
 417         int nr_rsps = queue->recv_queue_size * 2;
 418         int ret = -EINVAL, i;
 419
 420         queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
 421                         GFP_KERNEL);
 422         if (!queue->rsps)
 423                 goto out;
 424
 425         for (i = 0; i < nr_rsps; i++) {
 426                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 427
 428                 ret = nvmet_rdma_alloc_rsp(ndev, rsp);
 429                 if (ret)
 430                         goto out_free;
 431
 432                 list_add_tail(&rsp->free_list, &queue->free_rsps);
 433         }
 434
 435         return 0;
 436
 437 out_free:
 438         while (--i >= 0) {
 439                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 440
 441                 list_del(&rsp->free_list);
 442                 nvmet_rdma_free_rsp(ndev, rsp);
 443         }
 444         kfree(queue->rsps);
 445 out:
 446         return ret;
 447 }
 448
 449 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
 450 {
 451         struct nvmet_rdma_device *ndev = queue->dev;
 452         int i, nr_rsps = queue->recv_queue_size * 2;
 453
 454         for (i = 0; i < nr_rsps; i++) {
 455                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 456
 457                 list_del(&rsp->free_list);
 458                 nvmet_rdma_free_rsp(ndev, rsp);
 459         }
 460         kfree(queue->rsps);
 461 }
 462
 463 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
 464                 struct nvmet_rdma_cmd *cmd)
 465 {
 466         struct ib_recv_wr *bad_wr;
 467
 468         ib_dma_sync_single_for_device(ndev->device,
 469                 cmd->sge[0].addr, cmd->sge[0].length,
 470                 DMA_FROM_DEVICE);
 471
 472         if (ndev->srq)
 473                 return ib_post_srq_recv(ndev->srq, &cmd->wr, &bad_wr);
 474         return ib_post_recv(cmd->queue->cm_id->qp, &cmd->wr, &bad_wr);
 475 }
 476
 477 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
 478 {
 479         spin_lock(&queue->rsp_wr_wait_lock);
 480         while (!list_empty(&queue->rsp_wr_wait_list)) {
 481                 struct nvmet_rdma_rsp *rsp;
 482                 bool ret;
 483
 484                 rsp = list_entry(queue->rsp_wr_wait_list.next,
 485                                 struct nvmet_rdma_rsp, wait_list);
 486                 list_del(&rsp->wait_list);
 487
 488                 spin_unlock(&queue->rsp_wr_wait_lock);
 489                 ret = nvmet_rdma_execute_command(rsp);
 490                 spin_lock(&queue->rsp_wr_wait_lock);
 491
 492                 if (!ret) {
 493                         list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
 494                         break;
 495                 }
 496         }
 497         spin_unlock(&queue->rsp_wr_wait_lock);
 498 }
 499
 500
 501 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
 502 {
 503         struct nvmet_rdma_queue *queue = rsp->queue;
 504
 505         atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 506
 507         if (rsp->n_rdma) {
 508                 rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
 509                                 queue->cm_id->port_num, rsp->req.sg,
 510                                 rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
 511         }
 512
 513         if (rsp->req.sg != &rsp->cmd->inline_sg)
 514                 nvmet_rdma_free_sgl(rsp->req.sg, rsp->req.sg_cnt);
 515
 516         if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
 517                 nvmet_rdma_process_wr_wait_list(queue);
 518
 519         nvmet_rdma_put_rsp(rsp);
 520 }
 521
 522 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
 523 {
 524         if (queue->nvme_sq.ctrl) {
 525                 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
 526         } else {
 527                 /*
 528                  * we didn't setup the controller yet in case
 529                  * of admin connect error, just disconnect and
 530                  * cleanup the queue
 531                  */
 532                 nvmet_rdma_queue_disconnect(queue);
 533         }
 534 }
 535
 536 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
 537 {
 538         struct nvmet_rdma_rsp *rsp =
 539                 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
 540         struct nvmet_rdma_queue *queue = cq->cq_context;
 541
 542         nvmet_rdma_release_rsp(rsp);
 543
 544         if (unlikely(wc->status != IB_WC_SUCCESS &&
 545                      wc->status != IB_WC_WR_FLUSH_ERR)) {
 546                 pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
 547                         wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
 548                 nvmet_rdma_error_comp(queue);
 549         }
 550 }
 551
 552 static void nvmet_rdma_queue_response(struct nvmet_req *req)
 553 {
 554         struct nvmet_rdma_rsp *rsp =
 555                 container_of(req, struct nvmet_rdma_rsp, req);
 556         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 557         struct ib_send_wr *first_wr, *bad_wr;
 558
 559         if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
 560                 rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
 561                 rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
 562         } else {
 563                 rsp->send_wr.opcode = IB_WR_SEND;
 564         }
 565
 566         if (nvmet_rdma_need_data_out(rsp))
 567                 first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
 568                                 cm_id->port_num, NULL, &rsp->send_wr);
 569         else
 570                 first_wr = &rsp->send_wr;
 571
 572         nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
 573
 574         ib_dma_sync_single_for_device(rsp->queue->dev->device,
 575                 rsp->send_sge.addr, rsp->send_sge.length,
 576                 DMA_TO_DEVICE);
 577
 578         if (ib_post_send(cm_id->qp, first_wr, &bad_wr)) {
 579                 pr_err("sending cmd response failed\n");
 580                 nvmet_rdma_release_rsp(rsp);
 581         }
 582 }
 583
 584 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
 585 {
 586         struct nvmet_rdma_rsp *rsp =
 587                 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
 588         struct nvmet_rdma_queue *queue = cq->cq_context;
 589
 590         WARN_ON(rsp->n_rdma <= 0);
 591         atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
 592         rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
 593                         queue->cm_id->port_num, rsp->req.sg,
 594                         rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
 595         rsp->n_rdma = 0;
 596
 597         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 598                 nvmet_req_uninit(&rsp->req);
 599                 nvmet_rdma_release_rsp(rsp);
 600                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 601                         pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
 602                                 wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
 603                         nvmet_rdma_error_comp(queue);
 604                 }
 605                 return;
 606         }
 607
 608         rsp->req.execute(&rsp->req);
 609 }
 610
 611 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
 612                 u64 off)
 613 {
 614         sg_init_table(&rsp->cmd->inline_sg, 1);
 615         sg_set_page(&rsp->cmd->inline_sg, rsp->cmd->inline_page, len, off);
 616         rsp->req.sg = &rsp->cmd->inline_sg;
 617         rsp->req.sg_cnt = 1;
 618 }
 619
 620 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
 621 {
 622         struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
 623         u64 off = le64_to_cpu(sgl->addr);
 624         u32 len = le32_to_cpu(sgl->length);
 625
 626         if (!nvme_is_write(rsp->req.cmd))
 627                 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 628
 629         if (off + len > NVMET_RDMA_INLINE_DATA_SIZE) {
 630                 pr_err("invalid inline data offset!\n");
 631                 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
 632         }
 633
 634         /* no data command? */
 635         if (!len)
 636                 return 0;
 637
 638         nvmet_rdma_use_inline_sg(rsp, len, off);
 639         rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
 640         return 0;
 641 }
 642
 643 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
 644                 struct nvme_keyed_sgl_desc *sgl, bool invalidate)
 645 {
 646         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 647         u64 addr = le64_to_cpu(sgl->addr);
 648         u32 len = get_unaligned_le24(sgl->length);
 649         u32 key = get_unaligned_le32(sgl->key);
 650         int ret;
 651         u16 status;
 652
 653         /* no data command? */
 654         if (!len)
 655                 return 0;
 656
 657         status = nvmet_rdma_alloc_sgl(&rsp->req.sg, &rsp->req.sg_cnt,
 658                         len);
 659         if (status)
 660                 return status;
 661
 662         ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
 663                         rsp->req.sg, rsp->req.sg_cnt, 0, addr, key,
 664                         nvmet_data_dir(&rsp->req));
 665         if (ret < 0)
 666                 return NVME_SC_INTERNAL;
 667         rsp->n_rdma += ret;
 668
 669         if (invalidate) {
 670                 rsp->invalidate_rkey = key;
 671                 rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
 672         }
 673
 674         return 0;
 675 }
 676
 677 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
 678 {
 679         struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
 680
 681         switch (sgl->type >> 4) {
 682         case NVME_SGL_FMT_DATA_DESC:
 683                 switch (sgl->type & 0xf) {
 684                 case NVME_SGL_FMT_OFFSET:
 685                         return nvmet_rdma_map_sgl_inline(rsp);
 686                 default:
 687                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 688                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 689                 }
 690         case NVME_KEY_SGL_FMT_DATA_DESC:
 691                 switch (sgl->type & 0xf) {
 692                 case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
 693                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
 694                 case NVME_SGL_FMT_ADDRESS:
 695                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
 696                 default:
 697                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 698                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 699                 }
 700         default:
 701                 pr_err("invalid SGL type: %#x\n", sgl->type);
 702                 return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
 703         }
 704 }
 705
 706 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
 707 {
 708         struct nvmet_rdma_queue *queue = rsp->queue;
 709
 710         if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
 711                         &queue->sq_wr_avail) < 0)) {
 712                 pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
 713                                 1 + rsp->n_rdma, queue->idx,
 714                                 queue->nvme_sq.ctrl->cntlid);
 715                 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 716                 return false;
 717         }
 718
 719         if (nvmet_rdma_need_data_in(rsp)) {
 720                 if (rdma_rw_ctx_post(&rsp->rw, queue->cm_id->qp,
 721                                 queue->cm_id->port_num, &rsp->read_cqe, NULL))
 722                         nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
 723         } else {
 724                 rsp->req.execute(&rsp->req);
 725         }
 726
 727         return true;
 728 }
 729
 730 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
 731                 struct nvmet_rdma_rsp *cmd)
 732 {
 733         u16 status;
 734
 735         ib_dma_sync_single_for_cpu(queue->dev->device,
 736                 cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
 737                 DMA_FROM_DEVICE);
 738         ib_dma_sync_single_for_cpu(queue->dev->device,
 739                 cmd->send_sge.addr, cmd->send_sge.length,
 740                 DMA_TO_DEVICE);
 741
 742         if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
 743                         &queue->nvme_sq, &nvmet_rdma_ops))
 744                 return;
 745
 746         status = nvmet_rdma_map_sgl(cmd);
 747         if (status)
 748                 goto out_err;
 749
 750         if (unlikely(!nvmet_rdma_execute_command(cmd))) {
 751                 spin_lock(&queue->rsp_wr_wait_lock);
 752                 list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
 753                 spin_unlock(&queue->rsp_wr_wait_lock);
 754         }
 755
 756         return;
 757
 758 out_err:
 759         nvmet_req_complete(&cmd->req, status);
 760 }
 761
 762 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
 763 {
 764         struct nvmet_rdma_cmd *cmd =
 765                 container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
 766         struct nvmet_rdma_queue *queue = cq->cq_context;
 767         struct nvmet_rdma_rsp *rsp;
 768
 769         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 770                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 771                         pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
 772                                 wc->wr_cqe, ib_wc_status_msg(wc->status),
 773                                 wc->status);
 774                         nvmet_rdma_error_comp(queue);
 775                 }
 776                 return;
 777         }
 778
 779         if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
 780                 pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
 781                 nvmet_rdma_error_comp(queue);
 782                 return;
 783         }
 784
 785         cmd->queue = queue;
 786         rsp = nvmet_rdma_get_rsp(queue);
 787         if (unlikely(!rsp)) {
 788                 /*
 789                  * we get here only under memory pressure,
 790                  * silently drop and have the host retry
 791                  * as we can't even fail it.
 792                  */
 793                 nvmet_rdma_post_recv(queue->dev, cmd);
 794                 return;
 795         }
 796         rsp->queue = queue;
 797         rsp->cmd = cmd;
 798         rsp->flags = 0;
 799         rsp->req.cmd = cmd->nvme_cmd;
 800         rsp->req.port = queue->port;
 801         rsp->n_rdma = 0;
 802
 803         if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
 804                 unsigned long flags;
 805
 806                 spin_lock_irqsave(&queue->state_lock, flags);
 807                 if (queue->state == NVMET_RDMA_Q_CONNECTING)
 808                         list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
 809                 else
 810                         nvmet_rdma_put_rsp(rsp);
 811                 spin_unlock_irqrestore(&queue->state_lock, flags);
 812                 return;
 813         }
 814
 815         nvmet_rdma_handle_command(queue, rsp);
 816 }
 817
 818 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_device *ndev)
 819 {
 820         if (!ndev->srq)
 821                 return;
 822
 823         nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
 824         ib_destroy_srq(ndev->srq);
 825 }
 826
 827 static int nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
 828 {
 829         struct ib_srq_init_attr srq_attr = { NULL, };
 830         struct ib_srq *srq;
 831         size_t srq_size;
 832         int ret, i;
 833
 834         srq_size = 4095;        /* XXX: tune */
 835
 836         srq_attr.attr.max_wr = srq_size;
 837         srq_attr.attr.max_sge = 2;
 838         srq_attr.attr.srq_limit = 0;
 839         srq_attr.srq_type = IB_SRQT_BASIC;
 840         srq = ib_create_srq(ndev->pd, &srq_attr);
 841         if (IS_ERR(srq)) {
 842                 /*
 843                  * If SRQs aren't supported we just go ahead and use normal
 844                  * non-shared receive queues.
 845                  */
 846                 pr_info("SRQ requested but not supported.\n");
 847                 return 0;
 848         }
 849
 850         ndev->srq_cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
 851         if (IS_ERR(ndev->srq_cmds)) {
 852                 ret = PTR_ERR(ndev->srq_cmds);
 853                 goto out_destroy_srq;
 854         }
 855
 856         ndev->srq = srq;
 857         ndev->srq_size = srq_size;
 858
 859         for (i = 0; i < srq_size; i++)
 860                 nvmet_rdma_post_recv(ndev, &ndev->srq_cmds[i]);
 861
 862         return 0;
 863
 864 out_destroy_srq:
 865         ib_destroy_srq(srq);
 866         return ret;
 867 }
 868
 869 static void nvmet_rdma_free_dev(struct kref *ref)
 870 {
 871         struct nvmet_rdma_device *ndev =
 872                 container_of(ref, struct nvmet_rdma_device, ref);
 873
 874         mutex_lock(&device_list_mutex);
 875         list_del(&ndev->entry);
 876         mutex_unlock(&device_list_mutex);
 877
 878         nvmet_rdma_destroy_srq(ndev);
 879         ib_dealloc_pd(ndev->pd);
 880
 881         kfree(ndev);
 882 }
 883
 884 static struct nvmet_rdma_device *
 885 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
 886 {
 887         struct nvmet_rdma_device *ndev;
 888         int ret;
 889
 890         mutex_lock(&device_list_mutex);
 891         list_for_each_entry(ndev, &device_list, entry) {
 892                 if (ndev->device->node_guid == cm_id->device->node_guid &&
 893                     kref_get_unless_zero(&ndev->ref))
 894                         goto out_unlock;
 895         }
 896
 897         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 898         if (!ndev)
 899                 goto out_err;
 900
 901         ndev->device = cm_id->device;
 902         kref_init(&ndev->ref);
 903
 904         ndev->pd = ib_alloc_pd(ndev->device, 0);
 905         if (IS_ERR(ndev->pd))
 906                 goto out_free_dev;
 907
 908         if (nvmet_rdma_use_srq) {
 909                 ret = nvmet_rdma_init_srq(ndev);
 910                 if (ret)
 911                         goto out_free_pd;
 912         }
 913
 914         list_add(&ndev->entry, &device_list);
 915 out_unlock:
 916         mutex_unlock(&device_list_mutex);
 917         pr_debug("added %s.\n", ndev->device->name);
 918         return ndev;
 919
 920 out_free_pd:
 921         ib_dealloc_pd(ndev->pd);
 922 out_free_dev:
 923         kfree(ndev);
 924 out_err:
 925         mutex_unlock(&device_list_mutex);
 926         return NULL;
 927 }
 928
 929 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
 930 {
 931         struct ib_qp_init_attr qp_attr;
 932         struct nvmet_rdma_device *ndev = queue->dev;
 933         int comp_vector, nr_cqe, ret, i;
 934
 935         /*
 936          * Spread the io queues across completion vectors,
 937          * but still keep all admin queues on vector 0.
 938          */
 939         comp_vector = !queue->host_qid ? 0 :
 940                 queue->idx % ndev->device->num_comp_vectors;
 941
 942         /*
 943          * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
 944          */
 945         nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
 946
 947         queue->cq = ib_alloc_cq(ndev->device, queue,
 948                         nr_cqe + 1, comp_vector,
 949                         IB_POLL_WORKQUEUE);
 950         if (IS_ERR(queue->cq)) {
 951                 ret = PTR_ERR(queue->cq);
 952                 pr_err("failed to create CQ cqe= %d ret= %d\n",
 953                        nr_cqe + 1, ret);
 954                 goto out;
 955         }
 956
 957         memset(&qp_attr, 0, sizeof(qp_attr));
 958         qp_attr.qp_context = queue;
 959         qp_attr.event_handler = nvmet_rdma_qp_event;
 960         qp_attr.send_cq = queue->cq;
 961         qp_attr.recv_cq = queue->cq;
 962         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 963         qp_attr.qp_type = IB_QPT_RC;
 964         /* +1 for drain */
 965         qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
 966         qp_attr.cap.max_rdma_ctxs = queue->send_queue_size;
 967         qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
 968                                         ndev->device->attrs.max_sge);
 969
 970         if (ndev->srq) {
 971                 qp_attr.srq = ndev->srq;
 972         } else {
 973                 /* +1 for drain */
 974                 qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
 975                 qp_attr.cap.max_recv_sge = 2;
 976         }
 977
 978         ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
 979         if (ret) {
 980                 pr_err("failed to create_qp ret= %d\n", ret);
 981                 goto err_destroy_cq;
 982         }
 983
 984         atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
 985
 986         pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
 987                  __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
 988                  qp_attr.cap.max_send_wr, queue->cm_id);
 989
 990         if (!ndev->srq) {
 991                 for (i = 0; i < queue->recv_queue_size; i++) {
 992                         queue->cmds[i].queue = queue;
 993                         nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
 994                 }
 995         }
 996
 997 out:
 998         return ret;
 999
1000 err_destroy_cq:
1001         ib_free_cq(queue->cq);
1002         goto out;
1003 }
1004
1005 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1006 {
1007         ib_drain_qp(queue->cm_id->qp);
1008         rdma_destroy_qp(queue->cm_id);
1009         ib_free_cq(queue->cq);
1010 }
1011
1012 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1013 {
1014         pr_info("freeing queue %d\n", queue->idx);
1015
1016         nvmet_sq_destroy(&queue->nvme_sq);
1017
1018         nvmet_rdma_destroy_queue_ib(queue);
1019         if (!queue->dev->srq) {
1020                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1021                                 queue->recv_queue_size,
1022                                 !queue->host_qid);
1023         }
1024         nvmet_rdma_free_rsps(queue);
1025         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1026         kfree(queue);
1027 }
1028
1029 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1030 {
1031         struct nvmet_rdma_queue *queue =
1032                 container_of(w, struct nvmet_rdma_queue, release_work);
1033         struct rdma_cm_id *cm_id = queue->cm_id;
1034         struct nvmet_rdma_device *dev = queue->dev;
1035         enum nvmet_rdma_queue_state state = queue->state;
1036
1037         nvmet_rdma_free_queue(queue);
1038
1039         if (state != NVMET_RDMA_IN_DEVICE_REMOVAL)
1040                 rdma_destroy_id(cm_id);
1041
1042         kref_put(&dev->ref, nvmet_rdma_free_dev);
1043 }
1044
1045 static int
1046 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1047                                 struct nvmet_rdma_queue *queue)
1048 {
1049         struct nvme_rdma_cm_req *req;
1050
1051         req = (struct nvme_rdma_cm_req *)conn->private_data;
1052         if (!req || conn->private_data_len == 0)
1053                 return NVME_RDMA_CM_INVALID_LEN;
1054
1055         if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1056                 return NVME_RDMA_CM_INVALID_RECFMT;
1057
1058         queue->host_qid = le16_to_cpu(req->qid);
1059
1060         /*
1061          * req->hsqsize corresponds to our recv queue size plus 1
1062          * req->hrqsize corresponds to our send queue size
1063          */
1064         queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1065         queue->send_queue_size = le16_to_cpu(req->hrqsize);
1066
1067         if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
1068                 return NVME_RDMA_CM_INVALID_HSQSIZE;
1069
1070         /* XXX: Should we enforce some kind of max for IO queues? */
1071
1072         return 0;
1073 }
1074
1075 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1076                                 enum nvme_rdma_cm_status status)
1077 {
1078         struct nvme_rdma_cm_rej rej;
1079
1080         pr_debug("rejecting connect request: status %d (%s)\n",
1081                  status, nvme_rdma_cm_msg(status));
1082
1083         rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1084         rej.sts = cpu_to_le16(status);
1085
1086         return rdma_reject(cm_id, (void *)&rej, sizeof(rej));
1087 }
1088
1089 static struct nvmet_rdma_queue *
1090 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1091                 struct rdma_cm_id *cm_id,
1092                 struct rdma_cm_event *event)
1093 {
1094         struct nvmet_rdma_queue *queue;
1095         int ret;
1096
1097         queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1098         if (!queue) {
1099                 ret = NVME_RDMA_CM_NO_RSC;
1100                 goto out_reject;
1101         }
1102
1103         ret = nvmet_sq_init(&queue->nvme_sq);
1104         if (ret) {
1105                 ret = NVME_RDMA_CM_NO_RSC;
1106                 goto out_free_queue;
1107         }
1108
1109         ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1110         if (ret)
1111                 goto out_destroy_sq;
1112
1113         /*
1114          * Schedules the actual release because calling rdma_destroy_id from
1115          * inside a CM callback would trigger a deadlock. (great API design..)
1116          */
1117         INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1118         queue->dev = ndev;
1119         queue->cm_id = cm_id;
1120
1121         spin_lock_init(&queue->state_lock);
1122         queue->state = NVMET_RDMA_Q_CONNECTING;
1123         INIT_LIST_HEAD(&queue->rsp_wait_list);
1124         INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1125         spin_lock_init(&queue->rsp_wr_wait_lock);
1126         INIT_LIST_HEAD(&queue->free_rsps);
1127         spin_lock_init(&queue->rsps_lock);
1128         INIT_LIST_HEAD(&queue->queue_list);
1129
1130         queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1131         if (queue->idx < 0) {
1132                 ret = NVME_RDMA_CM_NO_RSC;
1133                 goto out_destroy_sq;
1134         }
1135
1136         ret = nvmet_rdma_alloc_rsps(queue);
1137         if (ret) {
1138                 ret = NVME_RDMA_CM_NO_RSC;
1139                 goto out_ida_remove;
1140         }
1141
1142         if (!ndev->srq) {
1143                 queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1144                                 queue->recv_queue_size,
1145                                 !queue->host_qid);
1146                 if (IS_ERR(queue->cmds)) {
1147                         ret = NVME_RDMA_CM_NO_RSC;
1148                         goto out_free_responses;
1149                 }
1150         }
1151
1152         ret = nvmet_rdma_create_queue_ib(queue);
1153         if (ret) {
1154                 pr_err("%s: creating RDMA queue failed (%d).\n",
1155                         __func__, ret);
1156                 ret = NVME_RDMA_CM_NO_RSC;
1157                 goto out_free_cmds;
1158         }
1159
1160         return queue;
1161
1162 out_free_cmds:
1163         if (!ndev->srq) {
1164                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1165                                 queue->recv_queue_size,
1166                                 !queue->host_qid);
1167         }
1168 out_free_responses:
1169         nvmet_rdma_free_rsps(queue);
1170 out_ida_remove:
1171         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1172 out_destroy_sq:
1173         nvmet_sq_destroy(&queue->nvme_sq);
1174 out_free_queue:
1175         kfree(queue);
1176 out_reject:
1177         nvmet_rdma_cm_reject(cm_id, ret);
1178         return NULL;
1179 }
1180
1181 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1182 {
1183         struct nvmet_rdma_queue *queue = priv;
1184
1185         switch (event->event) {
1186         case IB_EVENT_COMM_EST:
1187                 rdma_notify(queue->cm_id, event->event);
1188                 break;
1189         default:
1190                 pr_err("received IB QP event: %s (%d)\n",
1191                        ib_event_msg(event->event), event->event);
1192                 break;
1193         }
1194 }
1195
1196 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1197                 struct nvmet_rdma_queue *queue,
1198                 struct rdma_conn_param *p)
1199 {
1200         struct rdma_conn_param  param = { };
1201         struct nvme_rdma_cm_rep priv = { };
1202         int ret = -ENOMEM;
1203
1204         param.rnr_retry_count = 7;
1205         param.flow_control = 1;
1206         param.initiator_depth = min_t(u8, p->initiator_depth,
1207                 queue->dev->device->attrs.max_qp_init_rd_atom);
1208         param.private_data = &priv;
1209         param.private_data_len = sizeof(priv);
1210         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1211         priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1212
1213         ret = rdma_accept(cm_id, &param);
1214         if (ret)
1215                 pr_err("rdma_accept failed (error code = %d)\n", ret);
1216
1217         return ret;
1218 }
1219
1220 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1221                 struct rdma_cm_event *event)
1222 {
1223         struct nvmet_rdma_device *ndev;
1224         struct nvmet_rdma_queue *queue;
1225         int ret = -EINVAL;
1226
1227         ndev = nvmet_rdma_find_get_device(cm_id);
1228         if (!ndev) {
1229                 nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1230                 return -ECONNREFUSED;
1231         }
1232
1233         queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1234         if (!queue) {
1235                 ret = -ENOMEM;
1236                 goto put_device;
1237         }
1238         queue->port = cm_id->context;
1239
1240         if (queue->host_qid == 0) {
1241                 /* Let inflight controller teardown complete */
1242                 flush_scheduled_work();
1243         }
1244
1245         ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1246         if (ret)
1247                 goto release_queue;
1248
1249         mutex_lock(&nvmet_rdma_queue_mutex);
1250         list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1251         mutex_unlock(&nvmet_rdma_queue_mutex);
1252
1253         return 0;
1254
1255 release_queue:
1256         nvmet_rdma_free_queue(queue);
1257 put_device:
1258         kref_put(&ndev->ref, nvmet_rdma_free_dev);
1259
1260         return ret;
1261 }
1262
1263 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1264 {
1265         unsigned long flags;
1266
1267         spin_lock_irqsave(&queue->state_lock, flags);
1268         if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1269                 pr_warn("trying to establish a connected queue\n");
1270                 goto out_unlock;
1271         }
1272         queue->state = NVMET_RDMA_Q_LIVE;
1273
1274         while (!list_empty(&queue->rsp_wait_list)) {
1275                 struct nvmet_rdma_rsp *cmd;
1276
1277                 cmd = list_first_entry(&queue->rsp_wait_list,
1278                                         struct nvmet_rdma_rsp, wait_list);
1279                 list_del(&cmd->wait_list);
1280
1281                 spin_unlock_irqrestore(&queue->state_lock, flags);
1282                 nvmet_rdma_handle_command(queue, cmd);
1283                 spin_lock_irqsave(&queue->state_lock, flags);
1284         }
1285
1286 out_unlock:
1287         spin_unlock_irqrestore(&queue->state_lock, flags);
1288 }
1289
1290 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1291 {
1292         bool disconnect = false;
1293         unsigned long flags;
1294
1295         pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1296
1297         spin_lock_irqsave(&queue->state_lock, flags);
1298         switch (queue->state) {
1299         case NVMET_RDMA_Q_CONNECTING:
1300         case NVMET_RDMA_Q_LIVE:
1301                 queue->state = NVMET_RDMA_Q_DISCONNECTING;
1302         case NVMET_RDMA_IN_DEVICE_REMOVAL:
1303                 disconnect = true;
1304                 break;
1305         case NVMET_RDMA_Q_DISCONNECTING:
1306                 break;
1307         }
1308         spin_unlock_irqrestore(&queue->state_lock, flags);
1309
1310         if (disconnect) {
1311                 rdma_disconnect(queue->cm_id);
1312                 schedule_work(&queue->release_work);
1313         }
1314 }
1315
1316 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1317 {
1318         bool disconnect = false;
1319
1320         mutex_lock(&nvmet_rdma_queue_mutex);
1321         if (!list_empty(&queue->queue_list)) {
1322                 list_del_init(&queue->queue_list);
1323                 disconnect = true;
1324         }
1325         mutex_unlock(&nvmet_rdma_queue_mutex);
1326
1327         if (disconnect)
1328                 __nvmet_rdma_queue_disconnect(queue);
1329 }
1330
1331 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1332                 struct nvmet_rdma_queue *queue)
1333 {
1334         WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1335
1336         mutex_lock(&nvmet_rdma_queue_mutex);
1337         if (!list_empty(&queue->queue_list))
1338                 list_del_init(&queue->queue_list);
1339         mutex_unlock(&nvmet_rdma_queue_mutex);
1340
1341         pr_err("failed to connect queue %d\n", queue->idx);
1342         schedule_work(&queue->release_work);
1343 }
1344
1345 /**
1346  * nvme_rdma_device_removal() - Handle RDMA device removal
1347  * @cm_id:      rdma_cm id, used for nvmet port
1348  * @queue:      nvmet rdma queue (cm id qp_context)
1349  *
1350  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1351  * to unplug. Note that this event can be generated on a normal
1352  * queue cm_id and/or a device bound listener cm_id (where in this
1353  * case queue will be null).
1354  *
1355  * We registered an ib_client to handle device removal for queues,
1356  * so we only need to handle the listening port cm_ids. In this case
1357  * we nullify the priv to prevent double cm_id destruction and destroying
1358  * the cm_id implicitely by returning a non-zero rc to the callout.
1359  */
1360 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1361                 struct nvmet_rdma_queue *queue)
1362 {
1363         struct nvmet_port *port;
1364
1365         if (queue) {
1366                 /*
1367                  * This is a queue cm_id. we have registered
1368                  * an ib_client to handle queues removal
1369                  * so don't interfear and just return.
1370                  */
1371                 return 0;
1372         }
1373
1374         port = cm_id->context;
1375
1376         /*
1377          * This is a listener cm_id. Make sure that
1378          * future remove_port won't invoke a double
1379          * cm_id destroy. use atomic xchg to make sure
1380          * we don't compete with remove_port.
1381          */
1382         if (xchg(&port->priv, NULL) != cm_id)
1383                 return 0;
1384
1385         /*
1386          * We need to return 1 so that the core will destroy
1387          * it's own ID.  What a great API design..
1388          */
1389         return 1;
1390 }
1391
1392 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1393                 struct rdma_cm_event *event)
1394 {
1395         struct nvmet_rdma_queue *queue = NULL;
1396         int ret = 0;
1397
1398         if (cm_id->qp)
1399                 queue = cm_id->qp->qp_context;
1400
1401         pr_debug("%s (%d): status %d id %p\n",
1402                 rdma_event_msg(event->event), event->event,
1403                 event->status, cm_id);
1404
1405         switch (event->event) {
1406         case RDMA_CM_EVENT_CONNECT_REQUEST:
1407                 ret = nvmet_rdma_queue_connect(cm_id, event);
1408                 break;
1409         case RDMA_CM_EVENT_ESTABLISHED:
1410                 nvmet_rdma_queue_established(queue);
1411                 break;
1412         case RDMA_CM_EVENT_ADDR_CHANGE:
1413         case RDMA_CM_EVENT_DISCONNECTED:
1414         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1415                 /*
1416                  * We might end up here when we already freed the qp
1417                  * which means queue release sequence is in progress,
1418                  * so don't get in the way...
1419                  */
1420                 if (queue)
1421                         nvmet_rdma_queue_disconnect(queue);
1422                 break;
1423         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1424                 ret = nvmet_rdma_device_removal(cm_id, queue);
1425                 break;
1426         case RDMA_CM_EVENT_REJECTED:
1427                 pr_debug("Connection rejected: %s\n",
1428                          rdma_reject_msg(cm_id, event->status));
1429                 /* FALLTHROUGH */
1430         case RDMA_CM_EVENT_UNREACHABLE:
1431         case RDMA_CM_EVENT_CONNECT_ERROR:
1432                 nvmet_rdma_queue_connect_fail(cm_id, queue);
1433                 break;
1434         default:
1435                 pr_err("received unrecognized RDMA CM event %d\n",
1436                         event->event);
1437                 break;
1438         }
1439
1440         return ret;
1441 }
1442
1443 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1444 {
1445         struct nvmet_rdma_queue *queue;
1446
1447 restart:
1448         mutex_lock(&nvmet_rdma_queue_mutex);
1449         list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1450                 if (queue->nvme_sq.ctrl == ctrl) {
1451                         list_del_init(&queue->queue_list);
1452                         mutex_unlock(&nvmet_rdma_queue_mutex);
1453
1454                         __nvmet_rdma_queue_disconnect(queue);
1455                         goto restart;
1456                 }
1457         }
1458         mutex_unlock(&nvmet_rdma_queue_mutex);
1459 }
1460
1461 static int nvmet_rdma_add_port(struct nvmet_port *port)
1462 {
1463         struct rdma_cm_id *cm_id;
1464         struct sockaddr_storage addr = { };
1465         __kernel_sa_family_t af;
1466         int ret;
1467
1468         switch (port->disc_addr.adrfam) {
1469         case NVMF_ADDR_FAMILY_IP4:
1470                 af = AF_INET;
1471                 break;
1472         case NVMF_ADDR_FAMILY_IP6:
1473                 af = AF_INET6;
1474                 break;
1475         default:
1476                 pr_err("address family %d not supported\n",
1477                                 port->disc_addr.adrfam);
1478                 return -EINVAL;
1479         }
1480
1481         ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
1482                         port->disc_addr.trsvcid, &addr);
1483         if (ret) {
1484                 pr_err("malformed ip/port passed: %s:%s\n",
1485                         port->disc_addr.traddr, port->disc_addr.trsvcid);
1486                 return ret;
1487         }
1488
1489         cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1490                         RDMA_PS_TCP, IB_QPT_RC);
1491         if (IS_ERR(cm_id)) {
1492                 pr_err("CM ID creation failed\n");
1493                 return PTR_ERR(cm_id);
1494         }
1495
1496         /*
1497          * Allow both IPv4 and IPv6 sockets to bind a single port
1498          * at the same time.
1499          */
1500         ret = rdma_set_afonly(cm_id, 1);
1501         if (ret) {
1502                 pr_err("rdma_set_afonly failed (%d)\n", ret);
1503                 goto out_destroy_id;
1504         }
1505
1506         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
1507         if (ret) {
1508                 pr_err("binding CM ID to %pISpcs failed (%d)\n",
1509                         (struct sockaddr *)&addr, ret);
1510                 goto out_destroy_id;
1511         }
1512
1513         ret = rdma_listen(cm_id, 128);
1514         if (ret) {
1515                 pr_err("listening to %pISpcs failed (%d)\n",
1516                         (struct sockaddr *)&addr, ret);
1517                 goto out_destroy_id;
1518         }
1519
1520         pr_info("enabling port %d (%pISpcs)\n",
1521                 le16_to_cpu(port->disc_addr.portid), (struct sockaddr *)&addr);
1522         port->priv = cm_id;
1523         return 0;
1524
1525 out_destroy_id:
1526         rdma_destroy_id(cm_id);
1527         return ret;
1528 }
1529
1530 static void nvmet_rdma_remove_port(struct nvmet_port *port)
1531 {
1532         struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
1533
1534         if (cm_id)
1535                 rdma_destroy_id(cm_id);
1536 }
1537
1538 static struct nvmet_fabrics_ops nvmet_rdma_ops = {
1539         .owner                  = THIS_MODULE,
1540         .type                   = NVMF_TRTYPE_RDMA,
1541         .sqe_inline_size        = NVMET_RDMA_INLINE_DATA_SIZE,
1542         .msdbd                  = 1,
1543         .has_keyed_sgls         = 1,
1544         .add_port               = nvmet_rdma_add_port,
1545         .remove_port            = nvmet_rdma_remove_port,
1546         .queue_response         = nvmet_rdma_queue_response,
1547         .delete_ctrl            = nvmet_rdma_delete_ctrl,
1548 };
1549
1550 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1551 {
1552         struct nvmet_rdma_queue *queue, *tmp;
1553
1554         /* Device is being removed, delete all queues using this device */
1555         mutex_lock(&nvmet_rdma_queue_mutex);
1556         list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
1557                                  queue_list) {
1558                 if (queue->dev->device != ib_device)
1559                         continue;
1560
1561                 pr_info("Removing queue %d\n", queue->idx);
1562                 list_del_init(&queue->queue_list);
1563                 __nvmet_rdma_queue_disconnect(queue);
1564         }
1565         mutex_unlock(&nvmet_rdma_queue_mutex);
1566
1567         flush_scheduled_work();
1568 }
1569
1570 static struct ib_client nvmet_rdma_ib_client = {
1571         .name   = "nvmet_rdma",
1572         .remove = nvmet_rdma_remove_one
1573 };
1574
1575 static int __init nvmet_rdma_init(void)
1576 {
1577         int ret;
1578
1579         ret = ib_register_client(&nvmet_rdma_ib_client);
1580         if (ret)
1581                 return ret;
1582
1583         ret = nvmet_register_transport(&nvmet_rdma_ops);
1584         if (ret)
1585                 goto err_ib_client;
1586
1587         return 0;
1588
1589 err_ib_client:
1590         ib_unregister_client(&nvmet_rdma_ib_client);
1591         return ret;
1592 }
1593
1594 static void __exit nvmet_rdma_exit(void)
1595 {
1596         struct nvmet_rdma_queue *queue;
1597
1598         nvmet_unregister_transport(&nvmet_rdma_ops);
1599
1600         flush_scheduled_work();
1601
1602         mutex_lock(&nvmet_rdma_queue_mutex);
1603         while ((queue = list_first_entry_or_null(&nvmet_rdma_queue_list,
1604                         struct nvmet_rdma_queue, queue_list))) {
1605                 list_del_init(&queue->queue_list);
1606
1607                 mutex_unlock(&nvmet_rdma_queue_mutex);
1608                 __nvmet_rdma_queue_disconnect(queue);
1609                 mutex_lock(&nvmet_rdma_queue_mutex);
1610         }
1611         mutex_unlock(&nvmet_rdma_queue_mutex);
1612
1613         flush_scheduled_work();
1614         ib_unregister_client(&nvmet_rdma_ib_client);
1615         ida_destroy(&nvmet_rdma_queue_ida);
1616 }
1617
1618 module_init(nvmet_rdma_init);
1619 module_exit(nvmet_rdma_exit);
1620
1621 MODULE_LICENSE("GPL v2");
1622 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */