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 at least 1 page, up to 4 SGEs, and up to 16KB of inline data
  37  */
  38 #define NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE     PAGE_SIZE
  39 #define NVMET_RDMA_MAX_INLINE_SGE               4
  40 #define NVMET_RDMA_MAX_INLINE_DATA_SIZE         max_t(int, SZ_16K, PAGE_SIZE)
  41
  42 struct nvmet_rdma_cmd {
  43         struct ib_sge           sge[NVMET_RDMA_MAX_INLINE_SGE + 1];
  44         struct ib_cqe           cqe;
  45         struct ib_recv_wr       wr;
  46         struct scatterlist      inline_sg[NVMET_RDMA_MAX_INLINE_SGE];
  47         struct nvme_command     *nvme_cmd;
  48         struct nvmet_rdma_queue *queue;
  49 };
  50
  51 enum {
  52         NVMET_RDMA_REQ_INLINE_DATA      = (1 << 0),
  53         NVMET_RDMA_REQ_INVALIDATE_RKEY  = (1 << 1),
  54 };
  55
  56 struct nvmet_rdma_rsp {
  57         struct ib_sge           send_sge;
  58         struct ib_cqe           send_cqe;
  59         struct ib_send_wr       send_wr;
  60
  61         struct nvmet_rdma_cmd   *cmd;
  62         struct nvmet_rdma_queue *queue;
  63
  64         struct ib_cqe           read_cqe;
  65         struct rdma_rw_ctx      rw;
  66
  67         struct nvmet_req        req;
  68
  69         u8                      n_rdma;
  70         u32                     flags;
  71         u32                     invalidate_rkey;
  72
  73         struct list_head        wait_list;
  74         struct list_head        free_list;
  75 };
  76
  77 enum nvmet_rdma_queue_state {
  78         NVMET_RDMA_Q_CONNECTING,
  79         NVMET_RDMA_Q_LIVE,
  80         NVMET_RDMA_Q_DISCONNECTING,
  81 };
  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         int                     inline_data_size;
 121         int                     inline_page_count;
 122 };
 123
 124 static bool nvmet_rdma_use_srq;
 125 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
 126 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
 127
 128 static DEFINE_IDA(nvmet_rdma_queue_ida);
 129 static LIST_HEAD(nvmet_rdma_queue_list);
 130 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
 131
 132 static LIST_HEAD(device_list);
 133 static DEFINE_MUTEX(device_list_mutex);
 134
 135 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
 136 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
 137 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 138 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
 139 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
 140 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
 141
 142 static const struct nvmet_fabrics_ops nvmet_rdma_ops;
 143
 144 static int num_pages(int len)
 145 {
 146         return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
 147 }
 148
 149 /* XXX: really should move to a generic header sooner or later.. */
 150 static inline u32 get_unaligned_le24(const u8 *p)
 151 {
 152         return (u32)p[0] | (u32)p[1] << 8 | (u32)p[2] << 16;
 153 }
 154
 155 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
 156 {
 157         return nvme_is_write(rsp->req.cmd) &&
 158                 rsp->req.transfer_len &&
 159                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 160 }
 161
 162 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
 163 {
 164         return !nvme_is_write(rsp->req.cmd) &&
 165                 rsp->req.transfer_len &&
 166                 !rsp->req.rsp->status &&
 167                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 168 }
 169
 170 static inline struct nvmet_rdma_rsp *
 171 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
 172 {
 173         struct nvmet_rdma_rsp *rsp;
 174         unsigned long flags;
 175
 176         spin_lock_irqsave(&queue->rsps_lock, flags);
 177         rsp = list_first_entry(&queue->free_rsps,
 178                                 struct nvmet_rdma_rsp, free_list);
 179         list_del(&rsp->free_list);
 180         spin_unlock_irqrestore(&queue->rsps_lock, flags);
 181
 182         return rsp;
 183 }
 184
 185 static inline void
 186 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
 187 {
 188         unsigned long flags;
 189
 190         spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
 191         list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
 192         spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
 193 }
 194
 195 static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
 196                                 struct nvmet_rdma_cmd *c)
 197 {
 198         struct scatterlist *sg;
 199         struct ib_sge *sge;
 200         int i;
 201
 202         if (!ndev->inline_data_size)
 203                 return;
 204
 205         sg = c->inline_sg;
 206         sge = &c->sge[1];
 207
 208         for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
 209                 if (sge->length)
 210                         ib_dma_unmap_page(ndev->device, sge->addr,
 211                                         sge->length, DMA_FROM_DEVICE);
 212                 if (sg_page(sg))
 213                         __free_page(sg_page(sg));
 214         }
 215 }
 216
 217 static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
 218                                 struct nvmet_rdma_cmd *c)
 219 {
 220         struct scatterlist *sg;
 221         struct ib_sge *sge;
 222         struct page *pg;
 223         int len;
 224         int i;
 225
 226         if (!ndev->inline_data_size)
 227                 return 0;
 228
 229         sg = c->inline_sg;
 230         sg_init_table(sg, ndev->inline_page_count);
 231         sge = &c->sge[1];
 232         len = ndev->inline_data_size;
 233
 234         for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
 235                 pg = alloc_page(GFP_KERNEL);
 236                 if (!pg)
 237                         goto out_err;
 238                 sg_assign_page(sg, pg);
 239                 sge->addr = ib_dma_map_page(ndev->device,
 240                         pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
 241                 if (ib_dma_mapping_error(ndev->device, sge->addr))
 242                         goto out_err;
 243                 sge->length = min_t(int, len, PAGE_SIZE);
 244                 sge->lkey = ndev->pd->local_dma_lkey;
 245                 len -= sge->length;
 246         }
 247
 248         return 0;
 249 out_err:
 250         for (; i >= 0; i--, sg--, sge--) {
 251                 if (sge->length)
 252                         ib_dma_unmap_page(ndev->device, sge->addr,
 253                                         sge->length, DMA_FROM_DEVICE);
 254                 if (sg_page(sg))
 255                         __free_page(sg_page(sg));
 256         }
 257         return -ENOMEM;
 258 }
 259
 260 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
 261                         struct nvmet_rdma_cmd *c, bool admin)
 262 {
 263         /* NVMe command / RDMA RECV */
 264         c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
 265         if (!c->nvme_cmd)
 266                 goto out;
 267
 268         c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
 269                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 270         if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
 271                 goto out_free_cmd;
 272
 273         c->sge[0].length = sizeof(*c->nvme_cmd);
 274         c->sge[0].lkey = ndev->pd->local_dma_lkey;
 275
 276         if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
 277                 goto out_unmap_cmd;
 278
 279         c->cqe.done = nvmet_rdma_recv_done;
 280
 281         c->wr.wr_cqe = &c->cqe;
 282         c->wr.sg_list = c->sge;
 283         c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;
 284
 285         return 0;
 286
 287 out_unmap_cmd:
 288         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 289                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 290 out_free_cmd:
 291         kfree(c->nvme_cmd);
 292
 293 out:
 294         return -ENOMEM;
 295 }
 296
 297 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
 298                 struct nvmet_rdma_cmd *c, bool admin)
 299 {
 300         if (!admin)
 301                 nvmet_rdma_free_inline_pages(ndev, c);
 302         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 303                                 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 304         kfree(c->nvme_cmd);
 305 }
 306
 307 static struct nvmet_rdma_cmd *
 308 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
 309                 int nr_cmds, bool admin)
 310 {
 311         struct nvmet_rdma_cmd *cmds;
 312         int ret = -EINVAL, i;
 313
 314         cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
 315         if (!cmds)
 316                 goto out;
 317
 318         for (i = 0; i < nr_cmds; i++) {
 319                 ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
 320                 if (ret)
 321                         goto out_free;
 322         }
 323
 324         return cmds;
 325
 326 out_free:
 327         while (--i >= 0)
 328                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 329         kfree(cmds);
 330 out:
 331         return ERR_PTR(ret);
 332 }
 333
 334 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
 335                 struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
 336 {
 337         int i;
 338
 339         for (i = 0; i < nr_cmds; i++)
 340                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 341         kfree(cmds);
 342 }
 343
 344 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
 345                 struct nvmet_rdma_rsp *r)
 346 {
 347         /* NVMe CQE / RDMA SEND */
 348         r->req.rsp = kmalloc(sizeof(*r->req.rsp), GFP_KERNEL);
 349         if (!r->req.rsp)
 350                 goto out;
 351
 352         r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.rsp,
 353                         sizeof(*r->req.rsp), DMA_TO_DEVICE);
 354         if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
 355                 goto out_free_rsp;
 356
 357         r->send_sge.length = sizeof(*r->req.rsp);
 358         r->send_sge.lkey = ndev->pd->local_dma_lkey;
 359
 360         r->send_cqe.done = nvmet_rdma_send_done;
 361
 362         r->send_wr.wr_cqe = &r->send_cqe;
 363         r->send_wr.sg_list = &r->send_sge;
 364         r->send_wr.num_sge = 1;
 365         r->send_wr.send_flags = IB_SEND_SIGNALED;
 366
 367         /* Data In / RDMA READ */
 368         r->read_cqe.done = nvmet_rdma_read_data_done;
 369         return 0;
 370
 371 out_free_rsp:
 372         kfree(r->req.rsp);
 373 out:
 374         return -ENOMEM;
 375 }
 376
 377 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
 378                 struct nvmet_rdma_rsp *r)
 379 {
 380         ib_dma_unmap_single(ndev->device, r->send_sge.addr,
 381                                 sizeof(*r->req.rsp), DMA_TO_DEVICE);
 382         kfree(r->req.rsp);
 383 }
 384
 385 static int
 386 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
 387 {
 388         struct nvmet_rdma_device *ndev = queue->dev;
 389         int nr_rsps = queue->recv_queue_size * 2;
 390         int ret = -EINVAL, i;
 391
 392         queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
 393                         GFP_KERNEL);
 394         if (!queue->rsps)
 395                 goto out;
 396
 397         for (i = 0; i < nr_rsps; i++) {
 398                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 399
 400                 ret = nvmet_rdma_alloc_rsp(ndev, rsp);
 401                 if (ret)
 402                         goto out_free;
 403
 404                 list_add_tail(&rsp->free_list, &queue->free_rsps);
 405         }
 406
 407         return 0;
 408
 409 out_free:
 410         while (--i >= 0) {
 411                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 412
 413                 list_del(&rsp->free_list);
 414                 nvmet_rdma_free_rsp(ndev, rsp);
 415         }
 416         kfree(queue->rsps);
 417 out:
 418         return ret;
 419 }
 420
 421 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
 422 {
 423         struct nvmet_rdma_device *ndev = queue->dev;
 424         int i, nr_rsps = queue->recv_queue_size * 2;
 425
 426         for (i = 0; i < nr_rsps; i++) {
 427                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 428
 429                 list_del(&rsp->free_list);
 430                 nvmet_rdma_free_rsp(ndev, rsp);
 431         }
 432         kfree(queue->rsps);
 433 }
 434
 435 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
 436                 struct nvmet_rdma_cmd *cmd)
 437 {
 438         struct ib_recv_wr *bad_wr;
 439         int ret;
 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                 ret = ib_post_srq_recv(ndev->srq, &cmd->wr, &bad_wr);
 447         else
 448                 ret = ib_post_recv(cmd->queue->cm_id->qp, &cmd->wr, &bad_wr);
 449
 450         if (unlikely(ret))
 451                 pr_err("post_recv cmd failed\n");
 452
 453         return ret;
 454 }
 455
 456 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
 457 {
 458         spin_lock(&queue->rsp_wr_wait_lock);
 459         while (!list_empty(&queue->rsp_wr_wait_list)) {
 460                 struct nvmet_rdma_rsp *rsp;
 461                 bool ret;
 462
 463                 rsp = list_entry(queue->rsp_wr_wait_list.next,
 464                                 struct nvmet_rdma_rsp, wait_list);
 465                 list_del(&rsp->wait_list);
 466
 467                 spin_unlock(&queue->rsp_wr_wait_lock);
 468                 ret = nvmet_rdma_execute_command(rsp);
 469                 spin_lock(&queue->rsp_wr_wait_lock);
 470
 471                 if (!ret) {
 472                         list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
 473                         break;
 474                 }
 475         }
 476         spin_unlock(&queue->rsp_wr_wait_lock);
 477 }
 478
 479
 480 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
 481 {
 482         struct nvmet_rdma_queue *queue = rsp->queue;
 483
 484         atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 485
 486         if (rsp->n_rdma) {
 487                 rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
 488                                 queue->cm_id->port_num, rsp->req.sg,
 489                                 rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
 490         }
 491
 492         if (rsp->req.sg != rsp->cmd->inline_sg)
 493                 sgl_free(rsp->req.sg);
 494
 495         if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
 496                 nvmet_rdma_process_wr_wait_list(queue);
 497
 498         nvmet_rdma_put_rsp(rsp);
 499 }
 500
 501 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
 502 {
 503         if (queue->nvme_sq.ctrl) {
 504                 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
 505         } else {
 506                 /*
 507                  * we didn't setup the controller yet in case
 508                  * of admin connect error, just disconnect and
 509                  * cleanup the queue
 510                  */
 511                 nvmet_rdma_queue_disconnect(queue);
 512         }
 513 }
 514
 515 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
 516 {
 517         struct nvmet_rdma_rsp *rsp =
 518                 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
 519
 520         nvmet_rdma_release_rsp(rsp);
 521
 522         if (unlikely(wc->status != IB_WC_SUCCESS &&
 523                      wc->status != IB_WC_WR_FLUSH_ERR)) {
 524                 pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
 525                         wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
 526                 nvmet_rdma_error_comp(rsp->queue);
 527         }
 528 }
 529
 530 static void nvmet_rdma_queue_response(struct nvmet_req *req)
 531 {
 532         struct nvmet_rdma_rsp *rsp =
 533                 container_of(req, struct nvmet_rdma_rsp, req);
 534         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 535         struct ib_send_wr *first_wr, *bad_wr;
 536
 537         if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
 538                 rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
 539                 rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
 540         } else {
 541                 rsp->send_wr.opcode = IB_WR_SEND;
 542         }
 543
 544         if (nvmet_rdma_need_data_out(rsp))
 545                 first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
 546                                 cm_id->port_num, NULL, &rsp->send_wr);
 547         else
 548                 first_wr = &rsp->send_wr;
 549
 550         nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
 551
 552         ib_dma_sync_single_for_device(rsp->queue->dev->device,
 553                 rsp->send_sge.addr, rsp->send_sge.length,
 554                 DMA_TO_DEVICE);
 555
 556         if (unlikely(ib_post_send(cm_id->qp, first_wr, &bad_wr))) {
 557                 pr_err("sending cmd response failed\n");
 558                 nvmet_rdma_release_rsp(rsp);
 559         }
 560 }
 561
 562 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
 563 {
 564         struct nvmet_rdma_rsp *rsp =
 565                 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
 566         struct nvmet_rdma_queue *queue = cq->cq_context;
 567
 568         WARN_ON(rsp->n_rdma <= 0);
 569         atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
 570         rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
 571                         queue->cm_id->port_num, rsp->req.sg,
 572                         rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
 573         rsp->n_rdma = 0;
 574
 575         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 576                 nvmet_req_uninit(&rsp->req);
 577                 nvmet_rdma_release_rsp(rsp);
 578                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 579                         pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
 580                                 wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
 581                         nvmet_rdma_error_comp(queue);
 582                 }
 583                 return;
 584         }
 585
 586         nvmet_req_execute(&rsp->req);
 587 }
 588
 589 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
 590                 u64 off)
 591 {
 592         int sg_count = num_pages(len);
 593         struct scatterlist *sg;
 594         int i;
 595
 596         sg = rsp->cmd->inline_sg;
 597         for (i = 0; i < sg_count; i++, sg++) {
 598                 if (i < sg_count - 1)
 599                         sg_unmark_end(sg);
 600                 else
 601                         sg_mark_end(sg);
 602                 sg->offset = off;
 603                 sg->length = min_t(int, len, PAGE_SIZE - off);
 604                 len -= sg->length;
 605                 if (!i)
 606                         off = 0;
 607         }
 608
 609         rsp->req.sg = rsp->cmd->inline_sg;
 610         rsp->req.sg_cnt = sg_count;
 611 }
 612
 613 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
 614 {
 615         struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
 616         u64 off = le64_to_cpu(sgl->addr);
 617         u32 len = le32_to_cpu(sgl->length);
 618
 619         if (!nvme_is_write(rsp->req.cmd))
 620                 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 621
 622         if (off + len > rsp->queue->dev->inline_data_size) {
 623                 pr_err("invalid inline data offset!\n");
 624                 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
 625         }
 626
 627         /* no data command? */
 628         if (!len)
 629                 return 0;
 630
 631         nvmet_rdma_use_inline_sg(rsp, len, off);
 632         rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
 633         rsp->req.transfer_len += len;
 634         return 0;
 635 }
 636
 637 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
 638                 struct nvme_keyed_sgl_desc *sgl, bool invalidate)
 639 {
 640         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 641         u64 addr = le64_to_cpu(sgl->addr);
 642         u32 len = get_unaligned_le24(sgl->length);
 643         u32 key = get_unaligned_le32(sgl->key);
 644         int ret;
 645
 646         /* no data command? */
 647         if (!len)
 648                 return 0;
 649
 650         rsp->req.sg = sgl_alloc(len, GFP_KERNEL, &rsp->req.sg_cnt);
 651         if (!rsp->req.sg)
 652                 return NVME_SC_INTERNAL;
 653
 654         ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
 655                         rsp->req.sg, rsp->req.sg_cnt, 0, addr, key,
 656                         nvmet_data_dir(&rsp->req));
 657         if (ret < 0)
 658                 return NVME_SC_INTERNAL;
 659         rsp->req.transfer_len += len;
 660         rsp->n_rdma += ret;
 661
 662         if (invalidate) {
 663                 rsp->invalidate_rkey = key;
 664                 rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
 665         }
 666
 667         return 0;
 668 }
 669
 670 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
 671 {
 672         struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
 673
 674         switch (sgl->type >> 4) {
 675         case NVME_SGL_FMT_DATA_DESC:
 676                 switch (sgl->type & 0xf) {
 677                 case NVME_SGL_FMT_OFFSET:
 678                         return nvmet_rdma_map_sgl_inline(rsp);
 679                 default:
 680                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 681                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 682                 }
 683         case NVME_KEY_SGL_FMT_DATA_DESC:
 684                 switch (sgl->type & 0xf) {
 685                 case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
 686                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
 687                 case NVME_SGL_FMT_ADDRESS:
 688                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
 689                 default:
 690                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 691                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 692                 }
 693         default:
 694                 pr_err("invalid SGL type: %#x\n", sgl->type);
 695                 return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
 696         }
 697 }
 698
 699 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
 700 {
 701         struct nvmet_rdma_queue *queue = rsp->queue;
 702
 703         if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
 704                         &queue->sq_wr_avail) < 0)) {
 705                 pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
 706                                 1 + rsp->n_rdma, queue->idx,
 707                                 queue->nvme_sq.ctrl->cntlid);
 708                 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 709                 return false;
 710         }
 711
 712         if (nvmet_rdma_need_data_in(rsp)) {
 713                 if (rdma_rw_ctx_post(&rsp->rw, queue->cm_id->qp,
 714                                 queue->cm_id->port_num, &rsp->read_cqe, NULL))
 715                         nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
 716         } else {
 717                 nvmet_req_execute(&rsp->req);
 718         }
 719
 720         return true;
 721 }
 722
 723 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
 724                 struct nvmet_rdma_rsp *cmd)
 725 {
 726         u16 status;
 727
 728         ib_dma_sync_single_for_cpu(queue->dev->device,
 729                 cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
 730                 DMA_FROM_DEVICE);
 731         ib_dma_sync_single_for_cpu(queue->dev->device,
 732                 cmd->send_sge.addr, cmd->send_sge.length,
 733                 DMA_TO_DEVICE);
 734
 735         if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
 736                         &queue->nvme_sq, &nvmet_rdma_ops))
 737                 return;
 738
 739         status = nvmet_rdma_map_sgl(cmd);
 740         if (status)
 741                 goto out_err;
 742
 743         if (unlikely(!nvmet_rdma_execute_command(cmd))) {
 744                 spin_lock(&queue->rsp_wr_wait_lock);
 745                 list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
 746                 spin_unlock(&queue->rsp_wr_wait_lock);
 747         }
 748
 749         return;
 750
 751 out_err:
 752         nvmet_req_complete(&cmd->req, status);
 753 }
 754
 755 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
 756 {
 757         struct nvmet_rdma_cmd *cmd =
 758                 container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
 759         struct nvmet_rdma_queue *queue = cq->cq_context;
 760         struct nvmet_rdma_rsp *rsp;
 761
 762         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 763                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 764                         pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
 765                                 wc->wr_cqe, ib_wc_status_msg(wc->status),
 766                                 wc->status);
 767                         nvmet_rdma_error_comp(queue);
 768                 }
 769                 return;
 770         }
 771
 772         if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
 773                 pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
 774                 nvmet_rdma_error_comp(queue);
 775                 return;
 776         }
 777
 778         cmd->queue = queue;
 779         rsp = nvmet_rdma_get_rsp(queue);
 780         rsp->queue = queue;
 781         rsp->cmd = cmd;
 782         rsp->flags = 0;
 783         rsp->req.cmd = cmd->nvme_cmd;
 784         rsp->req.port = queue->port;
 785         rsp->n_rdma = 0;
 786
 787         if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
 788                 unsigned long flags;
 789
 790                 spin_lock_irqsave(&queue->state_lock, flags);
 791                 if (queue->state == NVMET_RDMA_Q_CONNECTING)
 792                         list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
 793                 else
 794                         nvmet_rdma_put_rsp(rsp);
 795                 spin_unlock_irqrestore(&queue->state_lock, flags);
 796                 return;
 797         }
 798
 799         nvmet_rdma_handle_command(queue, rsp);
 800 }
 801
 802 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_device *ndev)
 803 {
 804         if (!ndev->srq)
 805                 return;
 806
 807         nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
 808         ib_destroy_srq(ndev->srq);
 809 }
 810
 811 static int nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
 812 {
 813         struct ib_srq_init_attr srq_attr = { NULL, };
 814         struct ib_srq *srq;
 815         size_t srq_size;
 816         int ret, i;
 817
 818         srq_size = 4095;        /* XXX: tune */
 819
 820         srq_attr.attr.max_wr = srq_size;
 821         srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
 822         srq_attr.attr.srq_limit = 0;
 823         srq_attr.srq_type = IB_SRQT_BASIC;
 824         srq = ib_create_srq(ndev->pd, &srq_attr);
 825         if (IS_ERR(srq)) {
 826                 /*
 827                  * If SRQs aren't supported we just go ahead and use normal
 828                  * non-shared receive queues.
 829                  */
 830                 pr_info("SRQ requested but not supported.\n");
 831                 return 0;
 832         }
 833
 834         ndev->srq_cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
 835         if (IS_ERR(ndev->srq_cmds)) {
 836                 ret = PTR_ERR(ndev->srq_cmds);
 837                 goto out_destroy_srq;
 838         }
 839
 840         ndev->srq = srq;
 841         ndev->srq_size = srq_size;
 842
 843         for (i = 0; i < srq_size; i++) {
 844                 ret = nvmet_rdma_post_recv(ndev, &ndev->srq_cmds[i]);
 845                 if (ret)
 846                         goto out_free_cmds;
 847         }
 848
 849         return 0;
 850
 851 out_free_cmds:
 852         nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
 853 out_destroy_srq:
 854         ib_destroy_srq(srq);
 855         return ret;
 856 }
 857
 858 static void nvmet_rdma_free_dev(struct kref *ref)
 859 {
 860         struct nvmet_rdma_device *ndev =
 861                 container_of(ref, struct nvmet_rdma_device, ref);
 862
 863         mutex_lock(&device_list_mutex);
 864         list_del(&ndev->entry);
 865         mutex_unlock(&device_list_mutex);
 866
 867         nvmet_rdma_destroy_srq(ndev);
 868         ib_dealloc_pd(ndev->pd);
 869
 870         kfree(ndev);
 871 }
 872
 873 static struct nvmet_rdma_device *
 874 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
 875 {
 876         struct nvmet_port *port = cm_id->context;
 877         struct nvmet_rdma_device *ndev;
 878         int inline_page_count;
 879         int inline_sge_count;
 880         int ret;
 881
 882         mutex_lock(&device_list_mutex);
 883         list_for_each_entry(ndev, &device_list, entry) {
 884                 if (ndev->device->node_guid == cm_id->device->node_guid &&
 885                     kref_get_unless_zero(&ndev->ref))
 886                         goto out_unlock;
 887         }
 888
 889         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 890         if (!ndev)
 891                 goto out_err;
 892
 893         inline_page_count = num_pages(port->inline_data_size);
 894         inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
 895                                 cm_id->device->attrs.max_sge) - 1;
 896         if (inline_page_count > inline_sge_count) {
 897                 pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
 898                         port->inline_data_size, cm_id->device->name,
 899                         inline_sge_count * PAGE_SIZE);
 900                 port->inline_data_size = inline_sge_count * PAGE_SIZE;
 901                 inline_page_count = inline_sge_count;
 902         }
 903         ndev->inline_data_size = port->inline_data_size;
 904         ndev->inline_page_count = inline_page_count;
 905         ndev->device = cm_id->device;
 906         kref_init(&ndev->ref);
 907
 908         ndev->pd = ib_alloc_pd(ndev->device, 0);
 909         if (IS_ERR(ndev->pd))
 910                 goto out_free_dev;
 911
 912         if (nvmet_rdma_use_srq) {
 913                 ret = nvmet_rdma_init_srq(ndev);
 914                 if (ret)
 915                         goto out_free_pd;
 916         }
 917
 918         list_add(&ndev->entry, &device_list);
 919 out_unlock:
 920         mutex_unlock(&device_list_mutex);
 921         pr_debug("added %s.\n", ndev->device->name);
 922         return ndev;
 923
 924 out_free_pd:
 925         ib_dealloc_pd(ndev->pd);
 926 out_free_dev:
 927         kfree(ndev);
 928 out_err:
 929         mutex_unlock(&device_list_mutex);
 930         return NULL;
 931 }
 932
 933 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
 934 {
 935         struct ib_qp_init_attr qp_attr;
 936         struct nvmet_rdma_device *ndev = queue->dev;
 937         int comp_vector, nr_cqe, ret, i;
 938
 939         /*
 940          * Spread the io queues across completion vectors,
 941          * but still keep all admin queues on vector 0.
 942          */
 943         comp_vector = !queue->host_qid ? 0 :
 944                 queue->idx % ndev->device->num_comp_vectors;
 945
 946         /*
 947          * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
 948          */
 949         nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
 950
 951         queue->cq = ib_alloc_cq(ndev->device, queue,
 952                         nr_cqe + 1, comp_vector,
 953                         IB_POLL_WORKQUEUE);
 954         if (IS_ERR(queue->cq)) {
 955                 ret = PTR_ERR(queue->cq);
 956                 pr_err("failed to create CQ cqe= %d ret= %d\n",
 957                        nr_cqe + 1, ret);
 958                 goto out;
 959         }
 960
 961         memset(&qp_attr, 0, sizeof(qp_attr));
 962         qp_attr.qp_context = queue;
 963         qp_attr.event_handler = nvmet_rdma_qp_event;
 964         qp_attr.send_cq = queue->cq;
 965         qp_attr.recv_cq = queue->cq;
 966         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
 967         qp_attr.qp_type = IB_QPT_RC;
 968         /* +1 for drain */
 969         qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
 970         qp_attr.cap.max_rdma_ctxs = queue->send_queue_size;
 971         qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
 972                                         ndev->device->attrs.max_sge);
 973
 974         if (ndev->srq) {
 975                 qp_attr.srq = ndev->srq;
 976         } else {
 977                 /* +1 for drain */
 978                 qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
 979                 qp_attr.cap.max_recv_sge = 1 + ndev->inline_page_count;
 980         }
 981
 982         ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
 983         if (ret) {
 984                 pr_err("failed to create_qp ret= %d\n", ret);
 985                 goto err_destroy_cq;
 986         }
 987
 988         atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
 989
 990         pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
 991                  __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
 992                  qp_attr.cap.max_send_wr, queue->cm_id);
 993
 994         if (!ndev->srq) {
 995                 for (i = 0; i < queue->recv_queue_size; i++) {
 996                         queue->cmds[i].queue = queue;
 997                         ret = nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
 998                         if (ret)
 999                                 goto err_destroy_qp;
1000                 }
1001         }
1002
1003 out:
1004         return ret;
1005
1006 err_destroy_qp:
1007         rdma_destroy_qp(queue->cm_id);
1008 err_destroy_cq:
1009         ib_free_cq(queue->cq);
1010         goto out;
1011 }
1012
1013 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1014 {
1015         struct ib_qp *qp = queue->cm_id->qp;
1016
1017         ib_drain_qp(qp);
1018         rdma_destroy_id(queue->cm_id);
1019         ib_destroy_qp(qp);
1020         ib_free_cq(queue->cq);
1021 }
1022
1023 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1024 {
1025         pr_debug("freeing queue %d\n", queue->idx);
1026
1027         nvmet_sq_destroy(&queue->nvme_sq);
1028
1029         nvmet_rdma_destroy_queue_ib(queue);
1030         if (!queue->dev->srq) {
1031                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1032                                 queue->recv_queue_size,
1033                                 !queue->host_qid);
1034         }
1035         nvmet_rdma_free_rsps(queue);
1036         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1037         kfree(queue);
1038 }
1039
1040 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1041 {
1042         struct nvmet_rdma_queue *queue =
1043                 container_of(w, struct nvmet_rdma_queue, release_work);
1044         struct nvmet_rdma_device *dev = queue->dev;
1045
1046         nvmet_rdma_free_queue(queue);
1047
1048         kref_put(&dev->ref, nvmet_rdma_free_dev);
1049 }
1050
1051 static int
1052 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1053                                 struct nvmet_rdma_queue *queue)
1054 {
1055         struct nvme_rdma_cm_req *req;
1056
1057         req = (struct nvme_rdma_cm_req *)conn->private_data;
1058         if (!req || conn->private_data_len == 0)
1059                 return NVME_RDMA_CM_INVALID_LEN;
1060
1061         if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1062                 return NVME_RDMA_CM_INVALID_RECFMT;
1063
1064         queue->host_qid = le16_to_cpu(req->qid);
1065
1066         /*
1067          * req->hsqsize corresponds to our recv queue size plus 1
1068          * req->hrqsize corresponds to our send queue size
1069          */
1070         queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1071         queue->send_queue_size = le16_to_cpu(req->hrqsize);
1072
1073         if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
1074                 return NVME_RDMA_CM_INVALID_HSQSIZE;
1075
1076         /* XXX: Should we enforce some kind of max for IO queues? */
1077
1078         return 0;
1079 }
1080
1081 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1082                                 enum nvme_rdma_cm_status status)
1083 {
1084         struct nvme_rdma_cm_rej rej;
1085
1086         pr_debug("rejecting connect request: status %d (%s)\n",
1087                  status, nvme_rdma_cm_msg(status));
1088
1089         rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1090         rej.sts = cpu_to_le16(status);
1091
1092         return rdma_reject(cm_id, (void *)&rej, sizeof(rej));
1093 }
1094
1095 static struct nvmet_rdma_queue *
1096 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1097                 struct rdma_cm_id *cm_id,
1098                 struct rdma_cm_event *event)
1099 {
1100         struct nvmet_rdma_queue *queue;
1101         int ret;
1102
1103         queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1104         if (!queue) {
1105                 ret = NVME_RDMA_CM_NO_RSC;
1106                 goto out_reject;
1107         }
1108
1109         ret = nvmet_sq_init(&queue->nvme_sq);
1110         if (ret) {
1111                 ret = NVME_RDMA_CM_NO_RSC;
1112                 goto out_free_queue;
1113         }
1114
1115         ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1116         if (ret)
1117                 goto out_destroy_sq;
1118
1119         /*
1120          * Schedules the actual release because calling rdma_destroy_id from
1121          * inside a CM callback would trigger a deadlock. (great API design..)
1122          */
1123         INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1124         queue->dev = ndev;
1125         queue->cm_id = cm_id;
1126
1127         spin_lock_init(&queue->state_lock);
1128         queue->state = NVMET_RDMA_Q_CONNECTING;
1129         INIT_LIST_HEAD(&queue->rsp_wait_list);
1130         INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1131         spin_lock_init(&queue->rsp_wr_wait_lock);
1132         INIT_LIST_HEAD(&queue->free_rsps);
1133         spin_lock_init(&queue->rsps_lock);
1134         INIT_LIST_HEAD(&queue->queue_list);
1135
1136         queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1137         if (queue->idx < 0) {
1138                 ret = NVME_RDMA_CM_NO_RSC;
1139                 goto out_destroy_sq;
1140         }
1141
1142         ret = nvmet_rdma_alloc_rsps(queue);
1143         if (ret) {
1144                 ret = NVME_RDMA_CM_NO_RSC;
1145                 goto out_ida_remove;
1146         }
1147
1148         if (!ndev->srq) {
1149                 queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1150                                 queue->recv_queue_size,
1151                                 !queue->host_qid);
1152                 if (IS_ERR(queue->cmds)) {
1153                         ret = NVME_RDMA_CM_NO_RSC;
1154                         goto out_free_responses;
1155                 }
1156         }
1157
1158         ret = nvmet_rdma_create_queue_ib(queue);
1159         if (ret) {
1160                 pr_err("%s: creating RDMA queue failed (%d).\n",
1161                         __func__, ret);
1162                 ret = NVME_RDMA_CM_NO_RSC;
1163                 goto out_free_cmds;
1164         }
1165
1166         return queue;
1167
1168 out_free_cmds:
1169         if (!ndev->srq) {
1170                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1171                                 queue->recv_queue_size,
1172                                 !queue->host_qid);
1173         }
1174 out_free_responses:
1175         nvmet_rdma_free_rsps(queue);
1176 out_ida_remove:
1177         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1178 out_destroy_sq:
1179         nvmet_sq_destroy(&queue->nvme_sq);
1180 out_free_queue:
1181         kfree(queue);
1182 out_reject:
1183         nvmet_rdma_cm_reject(cm_id, ret);
1184         return NULL;
1185 }
1186
1187 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1188 {
1189         struct nvmet_rdma_queue *queue = priv;
1190
1191         switch (event->event) {
1192         case IB_EVENT_COMM_EST:
1193                 rdma_notify(queue->cm_id, event->event);
1194                 break;
1195         default:
1196                 pr_err("received IB QP event: %s (%d)\n",
1197                        ib_event_msg(event->event), event->event);
1198                 break;
1199         }
1200 }
1201
1202 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1203                 struct nvmet_rdma_queue *queue,
1204                 struct rdma_conn_param *p)
1205 {
1206         struct rdma_conn_param  param = { };
1207         struct nvme_rdma_cm_rep priv = { };
1208         int ret = -ENOMEM;
1209
1210         param.rnr_retry_count = 7;
1211         param.flow_control = 1;
1212         param.initiator_depth = min_t(u8, p->initiator_depth,
1213                 queue->dev->device->attrs.max_qp_init_rd_atom);
1214         param.private_data = &priv;
1215         param.private_data_len = sizeof(priv);
1216         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1217         priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1218
1219         ret = rdma_accept(cm_id, &param);
1220         if (ret)
1221                 pr_err("rdma_accept failed (error code = %d)\n", ret);
1222
1223         return ret;
1224 }
1225
1226 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1227                 struct rdma_cm_event *event)
1228 {
1229         struct nvmet_rdma_device *ndev;
1230         struct nvmet_rdma_queue *queue;
1231         int ret = -EINVAL;
1232
1233         ndev = nvmet_rdma_find_get_device(cm_id);
1234         if (!ndev) {
1235                 nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1236                 return -ECONNREFUSED;
1237         }
1238
1239         queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1240         if (!queue) {
1241                 ret = -ENOMEM;
1242                 goto put_device;
1243         }
1244         queue->port = cm_id->context;
1245
1246         if (queue->host_qid == 0) {
1247                 /* Let inflight controller teardown complete */
1248                 flush_scheduled_work();
1249         }
1250
1251         ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1252         if (ret) {
1253                 schedule_work(&queue->release_work);
1254                 /* Destroying rdma_cm id is not needed here */
1255                 return 0;
1256         }
1257
1258         mutex_lock(&nvmet_rdma_queue_mutex);
1259         list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1260         mutex_unlock(&nvmet_rdma_queue_mutex);
1261
1262         return 0;
1263
1264 put_device:
1265         kref_put(&ndev->ref, nvmet_rdma_free_dev);
1266
1267         return ret;
1268 }
1269
1270 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1271 {
1272         unsigned long flags;
1273
1274         spin_lock_irqsave(&queue->state_lock, flags);
1275         if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1276                 pr_warn("trying to establish a connected queue\n");
1277                 goto out_unlock;
1278         }
1279         queue->state = NVMET_RDMA_Q_LIVE;
1280
1281         while (!list_empty(&queue->rsp_wait_list)) {
1282                 struct nvmet_rdma_rsp *cmd;
1283
1284                 cmd = list_first_entry(&queue->rsp_wait_list,
1285                                         struct nvmet_rdma_rsp, wait_list);
1286                 list_del(&cmd->wait_list);
1287
1288                 spin_unlock_irqrestore(&queue->state_lock, flags);
1289                 nvmet_rdma_handle_command(queue, cmd);
1290                 spin_lock_irqsave(&queue->state_lock, flags);
1291         }
1292
1293 out_unlock:
1294         spin_unlock_irqrestore(&queue->state_lock, flags);
1295 }
1296
1297 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1298 {
1299         bool disconnect = false;
1300         unsigned long flags;
1301
1302         pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1303
1304         spin_lock_irqsave(&queue->state_lock, flags);
1305         switch (queue->state) {
1306         case NVMET_RDMA_Q_CONNECTING:
1307         case NVMET_RDMA_Q_LIVE:
1308                 queue->state = NVMET_RDMA_Q_DISCONNECTING;
1309                 disconnect = true;
1310                 break;
1311         case NVMET_RDMA_Q_DISCONNECTING:
1312                 break;
1313         }
1314         spin_unlock_irqrestore(&queue->state_lock, flags);
1315
1316         if (disconnect) {
1317                 rdma_disconnect(queue->cm_id);
1318                 schedule_work(&queue->release_work);
1319         }
1320 }
1321
1322 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1323 {
1324         bool disconnect = false;
1325
1326         mutex_lock(&nvmet_rdma_queue_mutex);
1327         if (!list_empty(&queue->queue_list)) {
1328                 list_del_init(&queue->queue_list);
1329                 disconnect = true;
1330         }
1331         mutex_unlock(&nvmet_rdma_queue_mutex);
1332
1333         if (disconnect)
1334                 __nvmet_rdma_queue_disconnect(queue);
1335 }
1336
1337 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1338                 struct nvmet_rdma_queue *queue)
1339 {
1340         WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1341
1342         mutex_lock(&nvmet_rdma_queue_mutex);
1343         if (!list_empty(&queue->queue_list))
1344                 list_del_init(&queue->queue_list);
1345         mutex_unlock(&nvmet_rdma_queue_mutex);
1346
1347         pr_err("failed to connect queue %d\n", queue->idx);
1348         schedule_work(&queue->release_work);
1349 }
1350
1351 /**
1352  * nvme_rdma_device_removal() - Handle RDMA device removal
1353  * @cm_id:      rdma_cm id, used for nvmet port
1354  * @queue:      nvmet rdma queue (cm id qp_context)
1355  *
1356  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1357  * to unplug. Note that this event can be generated on a normal
1358  * queue cm_id and/or a device bound listener cm_id (where in this
1359  * case queue will be null).
1360  *
1361  * We registered an ib_client to handle device removal for queues,
1362  * so we only need to handle the listening port cm_ids. In this case
1363  * we nullify the priv to prevent double cm_id destruction and destroying
1364  * the cm_id implicitely by returning a non-zero rc to the callout.
1365  */
1366 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1367                 struct nvmet_rdma_queue *queue)
1368 {
1369         struct nvmet_port *port;
1370
1371         if (queue) {
1372                 /*
1373                  * This is a queue cm_id. we have registered
1374                  * an ib_client to handle queues removal
1375                  * so don't interfear and just return.
1376                  */
1377                 return 0;
1378         }
1379
1380         port = cm_id->context;
1381
1382         /*
1383          * This is a listener cm_id. Make sure that
1384          * future remove_port won't invoke a double
1385          * cm_id destroy. use atomic xchg to make sure
1386          * we don't compete with remove_port.
1387          */
1388         if (xchg(&port->priv, NULL) != cm_id)
1389                 return 0;
1390
1391         /*
1392          * We need to return 1 so that the core will destroy
1393          * it's own ID.  What a great API design..
1394          */
1395         return 1;
1396 }
1397
1398 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1399                 struct rdma_cm_event *event)
1400 {
1401         struct nvmet_rdma_queue *queue = NULL;
1402         int ret = 0;
1403
1404         if (cm_id->qp)
1405                 queue = cm_id->qp->qp_context;
1406
1407         pr_debug("%s (%d): status %d id %p\n",
1408                 rdma_event_msg(event->event), event->event,
1409                 event->status, cm_id);
1410
1411         switch (event->event) {
1412         case RDMA_CM_EVENT_CONNECT_REQUEST:
1413                 ret = nvmet_rdma_queue_connect(cm_id, event);
1414                 break;
1415         case RDMA_CM_EVENT_ESTABLISHED:
1416                 nvmet_rdma_queue_established(queue);
1417                 break;
1418         case RDMA_CM_EVENT_ADDR_CHANGE:
1419         case RDMA_CM_EVENT_DISCONNECTED:
1420         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1421                 nvmet_rdma_queue_disconnect(queue);
1422                 break;
1423         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1424                 ret = nvmet_rdma_device_removal(cm_id, queue);
1425                 break;
1426         case RDMA_CM_EVENT_REJECTED:
1427                 pr_debug("Connection rejected: %s\n",
1428                          rdma_reject_msg(cm_id, event->status));
1429                 /* FALLTHROUGH */
1430         case RDMA_CM_EVENT_UNREACHABLE:
1431         case RDMA_CM_EVENT_CONNECT_ERROR:
1432                 nvmet_rdma_queue_connect_fail(cm_id, queue);
1433                 break;
1434         default:
1435                 pr_err("received unrecognized RDMA CM event %d\n",
1436                         event->event);
1437                 break;
1438         }
1439
1440         return ret;
1441 }
1442
1443 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1444 {
1445         struct nvmet_rdma_queue *queue;
1446
1447 restart:
1448         mutex_lock(&nvmet_rdma_queue_mutex);
1449         list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1450                 if (queue->nvme_sq.ctrl == ctrl) {
1451                         list_del_init(&queue->queue_list);
1452                         mutex_unlock(&nvmet_rdma_queue_mutex);
1453
1454                         __nvmet_rdma_queue_disconnect(queue);
1455                         goto restart;
1456                 }
1457         }
1458         mutex_unlock(&nvmet_rdma_queue_mutex);
1459 }
1460
1461 static int nvmet_rdma_add_port(struct nvmet_port *port)
1462 {
1463         struct rdma_cm_id *cm_id;
1464         struct sockaddr_storage addr = { };
1465         __kernel_sa_family_t af;
1466         int ret;
1467
1468         switch (port->disc_addr.adrfam) {
1469         case NVMF_ADDR_FAMILY_IP4:
1470                 af = AF_INET;
1471                 break;
1472         case NVMF_ADDR_FAMILY_IP6:
1473                 af = AF_INET6;
1474                 break;
1475         default:
1476                 pr_err("address family %d not supported\n",
1477                                 port->disc_addr.adrfam);
1478                 return -EINVAL;
1479         }
1480
1481         if (port->inline_data_size < 0) {
1482                 port->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
1483         } else if (port->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
1484                 pr_warn("inline_data_size %u is too large, reducing to %u\n",
1485                         port->inline_data_size,
1486                         NVMET_RDMA_MAX_INLINE_DATA_SIZE);
1487                 port->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
1488         }
1489
1490         ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
1491                         port->disc_addr.trsvcid, &addr);
1492         if (ret) {
1493                 pr_err("malformed ip/port passed: %s:%s\n",
1494                         port->disc_addr.traddr, port->disc_addr.trsvcid);
1495                 return ret;
1496         }
1497
1498         cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1499                         RDMA_PS_TCP, IB_QPT_RC);
1500         if (IS_ERR(cm_id)) {
1501                 pr_err("CM ID creation failed\n");
1502                 return PTR_ERR(cm_id);
1503         }
1504
1505         /*
1506          * Allow both IPv4 and IPv6 sockets to bind a single port
1507          * at the same time.
1508          */
1509         ret = rdma_set_afonly(cm_id, 1);
1510         if (ret) {
1511                 pr_err("rdma_set_afonly failed (%d)\n", ret);
1512                 goto out_destroy_id;
1513         }
1514
1515         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
1516         if (ret) {
1517                 pr_err("binding CM ID to %pISpcs failed (%d)\n",
1518                         (struct sockaddr *)&addr, ret);
1519                 goto out_destroy_id;
1520         }
1521
1522         ret = rdma_listen(cm_id, 128);
1523         if (ret) {
1524                 pr_err("listening to %pISpcs failed (%d)\n",
1525                         (struct sockaddr *)&addr, ret);
1526                 goto out_destroy_id;
1527         }
1528
1529         pr_info("enabling port %d (%pISpcs)\n",
1530                 le16_to_cpu(port->disc_addr.portid), (struct sockaddr *)&addr);
1531         port->priv = cm_id;
1532         return 0;
1533
1534 out_destroy_id:
1535         rdma_destroy_id(cm_id);
1536         return ret;
1537 }
1538
1539 static void nvmet_rdma_remove_port(struct nvmet_port *port)
1540 {
1541         struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
1542
1543         if (cm_id)
1544                 rdma_destroy_id(cm_id);
1545 }
1546
1547 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
1548                 struct nvmet_port *port, char *traddr)
1549 {
1550         struct rdma_cm_id *cm_id = port->priv;
1551
1552         if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
1553                 struct nvmet_rdma_rsp *rsp =
1554                         container_of(req, struct nvmet_rdma_rsp, req);
1555                 struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
1556                 struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;
1557
1558                 sprintf(traddr, "%pISc", addr);
1559         } else {
1560                 memcpy(traddr, port->disc_addr.traddr, NVMF_TRADDR_SIZE);
1561         }
1562 }
1563
1564 static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
1565         .owner                  = THIS_MODULE,
1566         .type                   = NVMF_TRTYPE_RDMA,
1567         .msdbd                  = 1,
1568         .has_keyed_sgls         = 1,
1569         .add_port               = nvmet_rdma_add_port,
1570         .remove_port            = nvmet_rdma_remove_port,
1571         .queue_response         = nvmet_rdma_queue_response,
1572         .delete_ctrl            = nvmet_rdma_delete_ctrl,
1573         .disc_traddr            = nvmet_rdma_disc_port_addr,
1574 };
1575
1576 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1577 {
1578         struct nvmet_rdma_queue *queue, *tmp;
1579         struct nvmet_rdma_device *ndev;
1580         bool found = false;
1581
1582         mutex_lock(&device_list_mutex);
1583         list_for_each_entry(ndev, &device_list, entry) {
1584                 if (ndev->device == ib_device) {
1585                         found = true;
1586                         break;
1587                 }
1588         }
1589         mutex_unlock(&device_list_mutex);
1590
1591         if (!found)
1592                 return;
1593
1594         /*
1595          * IB Device that is used by nvmet controllers is being removed,
1596          * delete all queues using this device.
1597          */
1598         mutex_lock(&nvmet_rdma_queue_mutex);
1599         list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
1600                                  queue_list) {
1601                 if (queue->dev->device != ib_device)
1602                         continue;
1603
1604                 pr_info("Removing queue %d\n", queue->idx);
1605                 list_del_init(&queue->queue_list);
1606                 __nvmet_rdma_queue_disconnect(queue);
1607         }
1608         mutex_unlock(&nvmet_rdma_queue_mutex);
1609
1610         flush_scheduled_work();
1611 }
1612
1613 static struct ib_client nvmet_rdma_ib_client = {
1614         .name   = "nvmet_rdma",
1615         .remove = nvmet_rdma_remove_one
1616 };
1617
1618 static int __init nvmet_rdma_init(void)
1619 {
1620         int ret;
1621
1622         ret = ib_register_client(&nvmet_rdma_ib_client);
1623         if (ret)
1624                 return ret;
1625
1626         ret = nvmet_register_transport(&nvmet_rdma_ops);
1627         if (ret)
1628                 goto err_ib_client;
1629
1630         return 0;
1631
1632 err_ib_client:
1633         ib_unregister_client(&nvmet_rdma_ib_client);
1634         return ret;
1635 }
1636
1637 static void __exit nvmet_rdma_exit(void)
1638 {
1639         nvmet_unregister_transport(&nvmet_rdma_ops);
1640         ib_unregister_client(&nvmet_rdma_ib_client);
1641         WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
1642         ida_destroy(&nvmet_rdma_queue_ida);
1643 }
1644
1645 module_init(nvmet_rdma_init);
1646 module_exit(nvmet_rdma_exit);
1647
1648 MODULE_LICENSE("GPL v2");
1649 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */