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