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_req_uninit(&rsp->req);
 571                 nvmet_rdma_release_rsp(rsp);
 572                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 573                         pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
 574                                 wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
 575                         nvmet_rdma_error_comp(queue);
 576                 }
 577                 return;
 578         }
 579
 580         rsp->req.execute(&rsp->req);
 581 }
 582
 583 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
 584                 u64 off)
 585 {
 586         sg_init_table(&rsp->cmd->inline_sg, 1);
 587         sg_set_page(&rsp->cmd->inline_sg, rsp->cmd->inline_page, len, off);
 588         rsp->req.sg = &rsp->cmd->inline_sg;
 589         rsp->req.sg_cnt = 1;
 590 }
 591
 592 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
 593 {
 594         struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
 595         u64 off = le64_to_cpu(sgl->addr);
 596         u32 len = le32_to_cpu(sgl->length);
 597
 598         if (!nvme_is_write(rsp->req.cmd))
 599                 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 600
 601         if (off + len > NVMET_RDMA_INLINE_DATA_SIZE) {
 602                 pr_err("invalid inline data offset!\n");
 603                 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
 604         }
 605
 606         /* no data command? */
 607         if (!len)
 608                 return 0;
 609
 610         nvmet_rdma_use_inline_sg(rsp, len, off);
 611         rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
 612         return 0;
 613 }
 614
 615 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
 616                 struct nvme_keyed_sgl_desc *sgl, bool invalidate)
 617 {
 618         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 619         u64 addr = le64_to_cpu(sgl->addr);
 620         u32 len = get_unaligned_le24(sgl->length);
 621         u32 key = get_unaligned_le32(sgl->key);
 622         int ret;
 623         u16 status;
 624
 625         /* no data command? */
 626         if (!len)
 627                 return 0;
 628
 629         status = nvmet_rdma_alloc_sgl(&rsp->req.sg, &rsp->req.sg_cnt,
 630                         len);
 631         if (status)
 632                 return status;
 633
 634         ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
 635                         rsp->req.sg, rsp->req.sg_cnt, 0, addr, key,
 636                         nvmet_data_dir(&rsp->req));
 637         if (ret < 0)
 638                 return NVME_SC_INTERNAL;
 639         rsp->n_rdma += ret;
 640
 641         if (invalidate) {
 642                 rsp->invalidate_rkey = key;
 643                 rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
 644         }
 645
 646         return 0;
 647 }
 648
 649 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
 650 {
 651         struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
 652
 653         switch (sgl->type >> 4) {
 654         case NVME_SGL_FMT_DATA_DESC:
 655                 switch (sgl->type & 0xf) {
 656                 case NVME_SGL_FMT_OFFSET:
 657                         return nvmet_rdma_map_sgl_inline(rsp);
 658                 default:
 659                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 660                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 661                 }
 662         case NVME_KEY_SGL_FMT_DATA_DESC:
 663                 switch (sgl->type & 0xf) {
 664                 case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
 665                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
 666                 case NVME_SGL_FMT_ADDRESS:
 667                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
 668                 default:
 669                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 670                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 671                 }
 672         default:
 673                 pr_err("invalid SGL type: %#x\n", sgl->type);
 674                 return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
 675         }
 676 }
 677
 678 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
 679 {
 680         struct nvmet_rdma_queue *queue = rsp->queue;
 681
 682         if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
 683                         &queue->sq_wr_avail) < 0)) {
 684                 pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
 685                                 1 + rsp->n_rdma, queue->idx,
 686                                 queue->nvme_sq.ctrl->cntlid);
 687                 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 688                 return false;
 689         }
 690
 691         if (nvmet_rdma_need_data_in(rsp)) {
 692                 if (rdma_rw_ctx_post(&rsp->rw, queue->cm_id->qp,
 693                                 queue->cm_id->port_num, &rsp->read_cqe, NULL))
 694                         nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
 695         } else {
 696                 rsp->req.execute(&rsp->req);
 697         }
 698
 699         return true;
 700 }
 701
 702 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
 703                 struct nvmet_rdma_rsp *cmd)
 704 {
 705         u16 status;
 706
 707         ib_dma_sync_single_for_cpu(queue->dev->device,
 708                 cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
 709                 DMA_FROM_DEVICE);
 710         ib_dma_sync_single_for_cpu(queue->dev->device,
 711                 cmd->send_sge.addr, cmd->send_sge.length,
 712                 DMA_TO_DEVICE);
 713
 714         if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
 715                         &queue->nvme_sq, &nvmet_rdma_ops))
 716                 return;
 717
 718         status = nvmet_rdma_map_sgl(cmd);
 719         if (status)
 720                 goto out_err;
 721
 722         if (unlikely(!nvmet_rdma_execute_command(cmd))) {
 723                 spin_lock(&queue->rsp_wr_wait_lock);
 724                 list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
 725                 spin_unlock(&queue->rsp_wr_wait_lock);
 726         }
 727
 728         return;
 729
 730 out_err:
 731         nvmet_req_complete(&cmd->req, status);
 732 }
 733
 734 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
 735 {
 736         struct nvmet_rdma_cmd *cmd =
 737                 container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
 738         struct nvmet_rdma_queue *queue = cq->cq_context;
 739         struct nvmet_rdma_rsp *rsp;
 740
 741         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 742                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 743                         pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
 744                                 wc->wr_cqe, ib_wc_status_msg(wc->status),
 745                                 wc->status);
 746                         nvmet_rdma_error_comp(queue);
 747                 }
 748                 return;
 749         }
 750
 751         if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
 752                 pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
 753                 nvmet_rdma_error_comp(queue);
 754                 return;
 755         }
 756
 757         cmd->queue = queue;
 758         rsp = nvmet_rdma_get_rsp(queue);
 759         rsp->queue = queue;
 760         rsp->cmd = cmd;
 761         rsp->flags = 0;
 762         rsp->req.cmd = cmd->nvme_cmd;
 763         rsp->req.port = queue->port;
 764         rsp->n_rdma = 0;
 765
 766         if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
 767                 unsigned long flags;
 768
 769                 spin_lock_irqsave(&queue->state_lock, flags);
 770                 if (queue->state == NVMET_RDMA_Q_CONNECTING)
 771                         list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
 772                 else
 773                         nvmet_rdma_put_rsp(rsp);
 774                 spin_unlock_irqrestore(&queue->state_lock, flags);
 775                 return;
 776         }
 777
 778         nvmet_rdma_handle_command(queue, rsp);
 779 }
 780
 781 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_device *ndev)
 782 {
 783         if (!ndev->srq)
 784                 return;
 785
 786         nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
 787         ib_destroy_srq(ndev->srq);
 788 }
 789
 790 static int nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
 791 {
 792         struct ib_srq_init_attr srq_attr = { NULL, };
 793         struct ib_srq *srq;
 794         size_t srq_size;
 795         int ret, i;
 796
 797         srq_size = 4095;        /* XXX: tune */
 798
 799         srq_attr.attr.max_wr = srq_size;
 800         srq_attr.attr.max_sge = 2;
 801         srq_attr.attr.srq_limit = 0;
 802         srq_attr.srq_type = IB_SRQT_BASIC;
 803         srq = ib_create_srq(ndev->pd, &srq_attr);
 804         if (IS_ERR(srq)) {
 805                 /*
 806                  * If SRQs aren't supported we just go ahead and use normal
 807                  * non-shared receive queues.
 808                  */
 809                 pr_info("SRQ requested but not supported.\n");
 810                 return 0;
 811         }
 812
 813         ndev->srq_cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
 814         if (IS_ERR(ndev->srq_cmds)) {
 815                 ret = PTR_ERR(ndev->srq_cmds);
 816                 goto out_destroy_srq;
 817         }
 818
 819         ndev->srq = srq;
 820         ndev->srq_size = srq_size;
 821
 822         for (i = 0; i < srq_size; i++)
 823                 nvmet_rdma_post_recv(ndev, &ndev->srq_cmds[i]);
 824
 825         return 0;
 826
 827 out_destroy_srq:
 828         ib_destroy_srq(srq);
 829         return ret;
 830 }
 831
 832 static void nvmet_rdma_free_dev(struct kref *ref)
 833 {
 834         struct nvmet_rdma_device *ndev =
 835                 container_of(ref, struct nvmet_rdma_device, ref);
 836
 837         mutex_lock(&device_list_mutex);
 838         list_del(&ndev->entry);
 839         mutex_unlock(&device_list_mutex);
 840
 841         nvmet_rdma_destroy_srq(ndev);
 842         ib_dealloc_pd(ndev->pd);
 843
 844         kfree(ndev);
 845 }
 846
 847 static struct nvmet_rdma_device *
 848 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
 849 {
 850         struct nvmet_rdma_device *ndev;
 851         int ret;
 852
 853         mutex_lock(&device_list_mutex);
 854         list_for_each_entry(ndev, &device_list, entry) {
 855                 if (ndev->device->node_guid == cm_id->device->node_guid &&
 856                     kref_get_unless_zero(&ndev->ref))
 857                         goto out_unlock;
 858         }
 859
 860         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 861         if (!ndev)
 862                 goto out_err;
 863
 864         ndev->device = cm_id->device;
 865         kref_init(&ndev->ref);
 866
 867         ndev->pd = ib_alloc_pd(ndev->device, 0);
 868         if (IS_ERR(ndev->pd))
 869                 goto out_free_dev;
 870
 871         if (nvmet_rdma_use_srq) {
 872                 ret = nvmet_rdma_init_srq(ndev);
 873                 if (ret)
 874                         goto out_free_pd;
 875         }
 876
 877         list_add(&ndev->entry, &device_list);
 878 out_unlock:
 879         mutex_unlock(&device_list_mutex);
 880         pr_debug("added %s.\n", ndev->device->name);
 881         return ndev;
 882
 883 out_free_pd:
 884         ib_dealloc_pd(ndev->pd);
 885 out_free_dev:
 886         kfree(ndev);
 887 out_err:
 888         mutex_unlock(&device_list_mutex);
 889         return NULL;
 890 }
 891
 892 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
 893 {
 894         struct ib_qp_init_attr qp_attr;
 895         struct nvmet_rdma_device *ndev = queue->dev;
 896         int comp_vector, nr_cqe, ret, i;
 897
 898         /*
 899          * Spread the io queues across completion vectors,
 900          * but still keep all admin queues on vector 0.
 901          */
 902         comp_vector = !queue->host_qid ? 0 :
 903                 queue->idx % ndev->device->num_comp_vectors;
 904
 905         /*
 906          * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
 907          */
 908         nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
 909
 910         queue->cq = ib_alloc_cq(ndev->device, queue,
 911                         nr_cqe + 1, comp_vector,
 912                         IB_POLL_WORKQUEUE);
 913         if (IS_ERR(queue->cq)) {
 914                 ret = PTR_ERR(queue->cq);
 915                 pr_err("failed to create CQ cqe= %d ret= %d\n",
 916                        nr_cqe + 1, ret);
 917                 goto out;
 918         }
 919
 920         memset(&qp_attr, 0, sizeof(qp_attr));
 921         qp_attr.qp_context = queue;
 922         qp_attr.event_handler = nvmet_rdma_qp_event;
 923         qp_attr.send_cq = queue->cq;
 924         qp_attr.recv_cq = queue->cq;
 925         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 926         qp_attr.qp_type = IB_QPT_RC;
 927         /* +1 for drain */
 928         qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
 929         qp_attr.cap.max_rdma_ctxs = queue->send_queue_size;
 930         qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
 931                                         ndev->device->attrs.max_sge);
 932
 933         if (ndev->srq) {
 934                 qp_attr.srq = ndev->srq;
 935         } else {
 936                 /* +1 for drain */
 937                 qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
 938                 qp_attr.cap.max_recv_sge = 2;
 939         }
 940
 941         ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
 942         if (ret) {
 943                 pr_err("failed to create_qp ret= %d\n", ret);
 944                 goto err_destroy_cq;
 945         }
 946
 947         atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
 948
 949         pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
 950                  __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
 951                  qp_attr.cap.max_send_wr, queue->cm_id);
 952
 953         if (!ndev->srq) {
 954                 for (i = 0; i < queue->recv_queue_size; i++) {
 955                         queue->cmds[i].queue = queue;
 956                         nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
 957                 }
 958         }
 959
 960 out:
 961         return ret;
 962
 963 err_destroy_cq:
 964         ib_free_cq(queue->cq);
 965         goto out;
 966 }
 967
 968 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
 969 {
 970         ib_drain_qp(queue->cm_id->qp);
 971         rdma_destroy_qp(queue->cm_id);
 972         ib_free_cq(queue->cq);
 973 }
 974
 975 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
 976 {
 977         pr_info("freeing queue %d\n", queue->idx);
 978
 979         nvmet_sq_destroy(&queue->nvme_sq);
 980
 981         nvmet_rdma_destroy_queue_ib(queue);
 982         if (!queue->dev->srq) {
 983                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
 984                                 queue->recv_queue_size,
 985                                 !queue->host_qid);
 986         }
 987         nvmet_rdma_free_rsps(queue);
 988         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
 989         kfree(queue);
 990 }
 991
 992 static void nvmet_rdma_release_queue_work(struct work_struct *w)
 993 {
 994         struct nvmet_rdma_queue *queue =
 995                 container_of(w, struct nvmet_rdma_queue, release_work);
 996         struct rdma_cm_id *cm_id = queue->cm_id;
 997         struct nvmet_rdma_device *dev = queue->dev;
 998         enum nvmet_rdma_queue_state state = queue->state;
 999
1000         nvmet_rdma_free_queue(queue);
1001
1002         if (state != NVMET_RDMA_IN_DEVICE_REMOVAL)
1003                 rdma_destroy_id(cm_id);
1004
1005         kref_put(&dev->ref, nvmet_rdma_free_dev);
1006 }
1007
1008 static int
1009 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1010                                 struct nvmet_rdma_queue *queue)
1011 {
1012         struct nvme_rdma_cm_req *req;
1013
1014         req = (struct nvme_rdma_cm_req *)conn->private_data;
1015         if (!req || conn->private_data_len == 0)
1016                 return NVME_RDMA_CM_INVALID_LEN;
1017
1018         if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1019                 return NVME_RDMA_CM_INVALID_RECFMT;
1020
1021         queue->host_qid = le16_to_cpu(req->qid);
1022
1023         /*
1024          * req->hsqsize corresponds to our recv queue size plus 1
1025          * req->hrqsize corresponds to our send queue size
1026          */
1027         queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1028         queue->send_queue_size = le16_to_cpu(req->hrqsize);
1029
1030         if (!queue->host_qid && queue->recv_queue_size > NVMF_AQ_DEPTH)
1031                 return NVME_RDMA_CM_INVALID_HSQSIZE;
1032
1033         /* XXX: Should we enforce some kind of max for IO queues? */
1034
1035         return 0;
1036 }
1037
1038 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1039                                 enum nvme_rdma_cm_status status)
1040 {
1041         struct nvme_rdma_cm_rej rej;
1042
1043         pr_debug("rejecting connect request: status %d (%s)\n",
1044                  status, nvme_rdma_cm_msg(status));
1045
1046         rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1047         rej.sts = cpu_to_le16(status);
1048
1049         return rdma_reject(cm_id, (void *)&rej, sizeof(rej));
1050 }
1051
1052 static struct nvmet_rdma_queue *
1053 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1054                 struct rdma_cm_id *cm_id,
1055                 struct rdma_cm_event *event)
1056 {
1057         struct nvmet_rdma_queue *queue;
1058         int ret;
1059
1060         queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1061         if (!queue) {
1062                 ret = NVME_RDMA_CM_NO_RSC;
1063                 goto out_reject;
1064         }
1065
1066         ret = nvmet_sq_init(&queue->nvme_sq);
1067         if (ret) {
1068                 ret = NVME_RDMA_CM_NO_RSC;
1069                 goto out_free_queue;
1070         }
1071
1072         ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1073         if (ret)
1074                 goto out_destroy_sq;
1075
1076         /*
1077          * Schedules the actual release because calling rdma_destroy_id from
1078          * inside a CM callback would trigger a deadlock. (great API design..)
1079          */
1080         INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1081         queue->dev = ndev;
1082         queue->cm_id = cm_id;
1083
1084         spin_lock_init(&queue->state_lock);
1085         queue->state = NVMET_RDMA_Q_CONNECTING;
1086         INIT_LIST_HEAD(&queue->rsp_wait_list);
1087         INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1088         spin_lock_init(&queue->rsp_wr_wait_lock);
1089         INIT_LIST_HEAD(&queue->free_rsps);
1090         spin_lock_init(&queue->rsps_lock);
1091         INIT_LIST_HEAD(&queue->queue_list);
1092
1093         queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1094         if (queue->idx < 0) {
1095                 ret = NVME_RDMA_CM_NO_RSC;
1096                 goto out_destroy_sq;
1097         }
1098
1099         ret = nvmet_rdma_alloc_rsps(queue);
1100         if (ret) {
1101                 ret = NVME_RDMA_CM_NO_RSC;
1102                 goto out_ida_remove;
1103         }
1104
1105         if (!ndev->srq) {
1106                 queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1107                                 queue->recv_queue_size,
1108                                 !queue->host_qid);
1109                 if (IS_ERR(queue->cmds)) {
1110                         ret = NVME_RDMA_CM_NO_RSC;
1111                         goto out_free_responses;
1112                 }
1113         }
1114
1115         ret = nvmet_rdma_create_queue_ib(queue);
1116         if (ret) {
1117                 pr_err("%s: creating RDMA queue failed (%d).\n",
1118                         __func__, ret);
1119                 ret = NVME_RDMA_CM_NO_RSC;
1120                 goto out_free_cmds;
1121         }
1122
1123         return queue;
1124
1125 out_free_cmds:
1126         if (!ndev->srq) {
1127                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1128                                 queue->recv_queue_size,
1129                                 !queue->host_qid);
1130         }
1131 out_free_responses:
1132         nvmet_rdma_free_rsps(queue);
1133 out_ida_remove:
1134         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1135 out_destroy_sq:
1136         nvmet_sq_destroy(&queue->nvme_sq);
1137 out_free_queue:
1138         kfree(queue);
1139 out_reject:
1140         nvmet_rdma_cm_reject(cm_id, ret);
1141         return NULL;
1142 }
1143
1144 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1145 {
1146         struct nvmet_rdma_queue *queue = priv;
1147
1148         switch (event->event) {
1149         case IB_EVENT_COMM_EST:
1150                 rdma_notify(queue->cm_id, event->event);
1151                 break;
1152         default:
1153                 pr_err("received IB QP event: %s (%d)\n",
1154                        ib_event_msg(event->event), event->event);
1155                 break;
1156         }
1157 }
1158
1159 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1160                 struct nvmet_rdma_queue *queue,
1161                 struct rdma_conn_param *p)
1162 {
1163         struct rdma_conn_param  param = { };
1164         struct nvme_rdma_cm_rep priv = { };
1165         int ret = -ENOMEM;
1166
1167         param.rnr_retry_count = 7;
1168         param.flow_control = 1;
1169         param.initiator_depth = min_t(u8, p->initiator_depth,
1170                 queue->dev->device->attrs.max_qp_init_rd_atom);
1171         param.private_data = &priv;
1172         param.private_data_len = sizeof(priv);
1173         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1174         priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1175
1176         ret = rdma_accept(cm_id, &param);
1177         if (ret)
1178                 pr_err("rdma_accept failed (error code = %d)\n", ret);
1179
1180         return ret;
1181 }
1182
1183 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1184                 struct rdma_cm_event *event)
1185 {
1186         struct nvmet_rdma_device *ndev;
1187         struct nvmet_rdma_queue *queue;
1188         int ret = -EINVAL;
1189
1190         ndev = nvmet_rdma_find_get_device(cm_id);
1191         if (!ndev) {
1192                 nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1193                 return -ECONNREFUSED;
1194         }
1195
1196         queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1197         if (!queue) {
1198                 ret = -ENOMEM;
1199                 goto put_device;
1200         }
1201         queue->port = cm_id->context;
1202
1203         if (queue->host_qid == 0) {
1204                 /* Let inflight controller teardown complete */
1205                 flush_scheduled_work();
1206         }
1207
1208         ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1209         if (ret)
1210                 goto release_queue;
1211
1212         mutex_lock(&nvmet_rdma_queue_mutex);
1213         list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1214         mutex_unlock(&nvmet_rdma_queue_mutex);
1215
1216         return 0;
1217
1218 release_queue:
1219         nvmet_rdma_free_queue(queue);
1220 put_device:
1221         kref_put(&ndev->ref, nvmet_rdma_free_dev);
1222
1223         return ret;
1224 }
1225
1226 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1227 {
1228         unsigned long flags;
1229
1230         spin_lock_irqsave(&queue->state_lock, flags);
1231         if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1232                 pr_warn("trying to establish a connected queue\n");
1233                 goto out_unlock;
1234         }
1235         queue->state = NVMET_RDMA_Q_LIVE;
1236
1237         while (!list_empty(&queue->rsp_wait_list)) {
1238                 struct nvmet_rdma_rsp *cmd;
1239
1240                 cmd = list_first_entry(&queue->rsp_wait_list,
1241                                         struct nvmet_rdma_rsp, wait_list);
1242                 list_del(&cmd->wait_list);
1243
1244                 spin_unlock_irqrestore(&queue->state_lock, flags);
1245                 nvmet_rdma_handle_command(queue, cmd);
1246                 spin_lock_irqsave(&queue->state_lock, flags);
1247         }
1248
1249 out_unlock:
1250         spin_unlock_irqrestore(&queue->state_lock, flags);
1251 }
1252
1253 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1254 {
1255         bool disconnect = false;
1256         unsigned long flags;
1257
1258         pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1259
1260         spin_lock_irqsave(&queue->state_lock, flags);
1261         switch (queue->state) {
1262         case NVMET_RDMA_Q_CONNECTING:
1263         case NVMET_RDMA_Q_LIVE:
1264                 queue->state = NVMET_RDMA_Q_DISCONNECTING;
1265         case NVMET_RDMA_IN_DEVICE_REMOVAL:
1266                 disconnect = true;
1267                 break;
1268         case NVMET_RDMA_Q_DISCONNECTING:
1269                 break;
1270         }
1271         spin_unlock_irqrestore(&queue->state_lock, flags);
1272
1273         if (disconnect) {
1274                 rdma_disconnect(queue->cm_id);
1275                 schedule_work(&queue->release_work);
1276         }
1277 }
1278
1279 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1280 {
1281         bool disconnect = false;
1282
1283         mutex_lock(&nvmet_rdma_queue_mutex);
1284         if (!list_empty(&queue->queue_list)) {
1285                 list_del_init(&queue->queue_list);
1286                 disconnect = true;
1287         }
1288         mutex_unlock(&nvmet_rdma_queue_mutex);
1289
1290         if (disconnect)
1291                 __nvmet_rdma_queue_disconnect(queue);
1292 }
1293
1294 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1295                 struct nvmet_rdma_queue *queue)
1296 {
1297         WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1298
1299         mutex_lock(&nvmet_rdma_queue_mutex);
1300         if (!list_empty(&queue->queue_list))
1301                 list_del_init(&queue->queue_list);
1302         mutex_unlock(&nvmet_rdma_queue_mutex);
1303
1304         pr_err("failed to connect queue %d\n", queue->idx);
1305         schedule_work(&queue->release_work);
1306 }
1307
1308 /**
1309  * nvme_rdma_device_removal() - Handle RDMA device removal
1310  * @queue:      nvmet rdma queue (cm id qp_context)
1311  * @addr:       nvmet address (cm_id context)
1312  *
1313  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1314  * to unplug so we should take care of destroying our RDMA resources.
1315  * This event will be generated for each allocated cm_id.
1316  *
1317  * Note that this event can be generated on a normal queue cm_id
1318  * and/or a device bound listener cm_id (where in this case
1319  * queue will be null).
1320  *
1321  * we claim ownership on destroying the cm_id. For queues we move
1322  * the queue state to NVMET_RDMA_IN_DEVICE_REMOVAL and for port
1323  * we nullify the priv to prevent double cm_id destruction and destroying
1324  * the cm_id implicitely by returning a non-zero rc to the callout.
1325  */
1326 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1327                 struct nvmet_rdma_queue *queue)
1328 {
1329         unsigned long flags;
1330
1331         if (!queue) {
1332                 struct nvmet_port *port = cm_id->context;
1333
1334                 /*
1335                  * This is a listener cm_id. Make sure that
1336                  * future remove_port won't invoke a double
1337                  * cm_id destroy. use atomic xchg to make sure
1338                  * we don't compete with remove_port.
1339                  */
1340                 if (xchg(&port->priv, NULL) != cm_id)
1341                         return 0;
1342         } else {
1343                 /*
1344                  * This is a queue cm_id. Make sure that
1345                  * release queue will not destroy the cm_id
1346                  * and schedule all ctrl queues removal (only
1347                  * if the queue is not disconnecting already).
1348                  */
1349                 spin_lock_irqsave(&queue->state_lock, flags);
1350                 if (queue->state != NVMET_RDMA_Q_DISCONNECTING)
1351                         queue->state = NVMET_RDMA_IN_DEVICE_REMOVAL;
1352                 spin_unlock_irqrestore(&queue->state_lock, flags);
1353                 nvmet_rdma_queue_disconnect(queue);
1354                 flush_scheduled_work();
1355         }
1356
1357         /*
1358          * We need to return 1 so that the core will destroy
1359          * it's own ID.  What a great API design..
1360          */
1361         return 1;
1362 }
1363
1364 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1365                 struct rdma_cm_event *event)
1366 {
1367         struct nvmet_rdma_queue *queue = NULL;
1368         int ret = 0;
1369
1370         if (cm_id->qp)
1371                 queue = cm_id->qp->qp_context;
1372
1373         pr_debug("%s (%d): status %d id %p\n",
1374                 rdma_event_msg(event->event), event->event,
1375                 event->status, cm_id);
1376
1377         switch (event->event) {
1378         case RDMA_CM_EVENT_CONNECT_REQUEST:
1379                 ret = nvmet_rdma_queue_connect(cm_id, event);
1380                 break;
1381         case RDMA_CM_EVENT_ESTABLISHED:
1382                 nvmet_rdma_queue_established(queue);
1383                 break;
1384         case RDMA_CM_EVENT_ADDR_CHANGE:
1385         case RDMA_CM_EVENT_DISCONNECTED:
1386         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1387                 /*
1388                  * We might end up here when we already freed the qp
1389                  * which means queue release sequence is in progress,
1390                  * so don't get in the way...
1391                  */
1392                 if (queue)
1393                         nvmet_rdma_queue_disconnect(queue);
1394                 break;
1395         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1396                 ret = nvmet_rdma_device_removal(cm_id, queue);
1397                 break;
1398         case RDMA_CM_EVENT_REJECTED:
1399                 pr_debug("Connection rejected: %s\n",
1400                          rdma_reject_msg(cm_id, event->status));
1401                 /* FALLTHROUGH */
1402         case RDMA_CM_EVENT_UNREACHABLE:
1403         case RDMA_CM_EVENT_CONNECT_ERROR:
1404                 nvmet_rdma_queue_connect_fail(cm_id, queue);
1405                 break;
1406         default:
1407                 pr_err("received unrecognized RDMA CM event %d\n",
1408                         event->event);
1409                 break;
1410         }
1411
1412         return ret;
1413 }
1414
1415 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1416 {
1417         struct nvmet_rdma_queue *queue;
1418
1419 restart:
1420         mutex_lock(&nvmet_rdma_queue_mutex);
1421         list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1422                 if (queue->nvme_sq.ctrl == ctrl) {
1423                         list_del_init(&queue->queue_list);
1424                         mutex_unlock(&nvmet_rdma_queue_mutex);
1425
1426                         __nvmet_rdma_queue_disconnect(queue);
1427                         goto restart;
1428                 }
1429         }
1430         mutex_unlock(&nvmet_rdma_queue_mutex);
1431 }
1432
1433 static int nvmet_rdma_add_port(struct nvmet_port *port)
1434 {
1435         struct rdma_cm_id *cm_id;
1436         struct sockaddr_storage addr = { };
1437         __kernel_sa_family_t af;
1438         int ret;
1439
1440         switch (port->disc_addr.adrfam) {
1441         case NVMF_ADDR_FAMILY_IP4:
1442                 af = AF_INET;
1443                 break;
1444         case NVMF_ADDR_FAMILY_IP6:
1445                 af = AF_INET6;
1446                 break;
1447         default:
1448                 pr_err("address family %d not supported\n",
1449                                 port->disc_addr.adrfam);
1450                 return -EINVAL;
1451         }
1452
1453         ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
1454                         port->disc_addr.trsvcid, &addr);
1455         if (ret) {
1456                 pr_err("malformed ip/port passed: %s:%s\n",
1457                         port->disc_addr.traddr, port->disc_addr.trsvcid);
1458                 return ret;
1459         }
1460
1461         cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1462                         RDMA_PS_TCP, IB_QPT_RC);
1463         if (IS_ERR(cm_id)) {
1464                 pr_err("CM ID creation failed\n");
1465                 return PTR_ERR(cm_id);
1466         }
1467
1468         /*
1469          * Allow both IPv4 and IPv6 sockets to bind a single port
1470          * at the same time.
1471          */
1472         ret = rdma_set_afonly(cm_id, 1);
1473         if (ret) {
1474                 pr_err("rdma_set_afonly failed (%d)\n", ret);
1475                 goto out_destroy_id;
1476         }
1477
1478         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
1479         if (ret) {
1480                 pr_err("binding CM ID to %pISpcs failed (%d)\n",
1481                         (struct sockaddr *)&addr, ret);
1482                 goto out_destroy_id;
1483         }
1484
1485         ret = rdma_listen(cm_id, 128);
1486         if (ret) {
1487                 pr_err("listening to %pISpcs failed (%d)\n",
1488                         (struct sockaddr *)&addr, ret);
1489                 goto out_destroy_id;
1490         }
1491
1492         pr_info("enabling port %d (%pISpcs)\n",
1493                 le16_to_cpu(port->disc_addr.portid), (struct sockaddr *)&addr);
1494         port->priv = cm_id;
1495         return 0;
1496
1497 out_destroy_id:
1498         rdma_destroy_id(cm_id);
1499         return ret;
1500 }
1501
1502 static void nvmet_rdma_remove_port(struct nvmet_port *port)
1503 {
1504         struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
1505
1506         if (cm_id)
1507                 rdma_destroy_id(cm_id);
1508 }
1509
1510 static struct nvmet_fabrics_ops nvmet_rdma_ops = {
1511         .owner                  = THIS_MODULE,
1512         .type                   = NVMF_TRTYPE_RDMA,
1513         .sqe_inline_size        = NVMET_RDMA_INLINE_DATA_SIZE,
1514         .msdbd                  = 1,
1515         .has_keyed_sgls         = 1,
1516         .add_port               = nvmet_rdma_add_port,
1517         .remove_port            = nvmet_rdma_remove_port,
1518         .queue_response         = nvmet_rdma_queue_response,
1519         .delete_ctrl            = nvmet_rdma_delete_ctrl,
1520 };
1521
1522 static int __init nvmet_rdma_init(void)
1523 {
1524         return nvmet_register_transport(&nvmet_rdma_ops);
1525 }
1526
1527 static void __exit nvmet_rdma_exit(void)
1528 {
1529         struct nvmet_rdma_queue *queue;
1530
1531         nvmet_unregister_transport(&nvmet_rdma_ops);
1532
1533         flush_scheduled_work();
1534
1535         mutex_lock(&nvmet_rdma_queue_mutex);
1536         while ((queue = list_first_entry_or_null(&nvmet_rdma_queue_list,
1537                         struct nvmet_rdma_queue, queue_list))) {
1538                 list_del_init(&queue->queue_list);
1539
1540                 mutex_unlock(&nvmet_rdma_queue_mutex);
1541                 __nvmet_rdma_queue_disconnect(queue);
1542                 mutex_lock(&nvmet_rdma_queue_mutex);
1543         }
1544         mutex_unlock(&nvmet_rdma_queue_mutex);
1545
1546         flush_scheduled_work();
1547         ida_destroy(&nvmet_rdma_queue_ida);
1548 }
1549
1550 module_init(nvmet_rdma_init);
1551 module_exit(nvmet_rdma_exit);
1552
1553 MODULE_LICENSE("GPL v2");
1554 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */