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