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         bool                    allocated;
  70         u8                      n_rdma;
  71         u32                     flags;
  72         u32                     invalidate_rkey;
  73
  74         struct list_head        wait_list;
  75         struct list_head        free_list;
  76 };
  77
  78 enum nvmet_rdma_queue_state {
  79         NVMET_RDMA_Q_CONNECTING,
  80         NVMET_RDMA_Q_LIVE,
  81         NVMET_RDMA_Q_DISCONNECTING,
  82 };
  83
  84 struct nvmet_rdma_queue {
  85         struct rdma_cm_id       *cm_id;
  86         struct nvmet_port       *port;
  87         struct ib_cq            *cq;
  88         atomic_t                sq_wr_avail;
  89         struct nvmet_rdma_device *dev;
  90         spinlock_t              state_lock;
  91         enum nvmet_rdma_queue_state state;
  92         struct nvmet_cq         nvme_cq;
  93         struct nvmet_sq         nvme_sq;
  94
  95         struct nvmet_rdma_rsp   *rsps;
  96         struct list_head        free_rsps;
  97         spinlock_t              rsps_lock;
  98         struct nvmet_rdma_cmd   *cmds;
  99
 100         struct work_struct      release_work;
 101         struct list_head        rsp_wait_list;
 102         struct list_head        rsp_wr_wait_list;
 103         spinlock_t              rsp_wr_wait_lock;
 104
 105         int                     idx;
 106         int                     host_qid;
 107         int                     recv_queue_size;
 108         int                     send_queue_size;
 109
 110         struct list_head        queue_list;
 111 };
 112
 113 struct nvmet_rdma_device {
 114         struct ib_device        *device;
 115         struct ib_pd            *pd;
 116         struct ib_srq           *srq;
 117         struct nvmet_rdma_cmd   *srq_cmds;
 118         size_t                  srq_size;
 119         struct kref             ref;
 120         struct list_head        entry;
 121         int                     inline_data_size;
 122         int                     inline_page_count;
 123 };
 124
 125 static bool nvmet_rdma_use_srq;
 126 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
 127 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
 128
 129 static DEFINE_IDA(nvmet_rdma_queue_ida);
 130 static LIST_HEAD(nvmet_rdma_queue_list);
 131 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
 132
 133 static LIST_HEAD(device_list);
 134 static DEFINE_MUTEX(device_list_mutex);
 135
 136 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
 137 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
 138 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 139 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
 140 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
 141 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
 142 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
 143                                 struct nvmet_rdma_rsp *r);
 144 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
 145                                 struct nvmet_rdma_rsp *r);
 146
 147 static const struct nvmet_fabrics_ops nvmet_rdma_ops;
 148
 149 static int num_pages(int len)
 150 {
 151         return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
 152 }
 153
 154 /* XXX: really should move to a generic header sooner or later.. */
 155 static inline u32 get_unaligned_le24(const u8 *p)
 156 {
 157         return (u32)p[0] | (u32)p[1] << 8 | (u32)p[2] << 16;
 158 }
 159
 160 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
 161 {
 162         return nvme_is_write(rsp->req.cmd) &&
 163                 rsp->req.transfer_len &&
 164                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 165 }
 166
 167 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
 168 {
 169         return !nvme_is_write(rsp->req.cmd) &&
 170                 rsp->req.transfer_len &&
 171                 !rsp->req.rsp->status &&
 172                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 173 }
 174
 175 static inline struct nvmet_rdma_rsp *
 176 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
 177 {
 178         struct nvmet_rdma_rsp *rsp;
 179         unsigned long flags;
 180
 181         spin_lock_irqsave(&queue->rsps_lock, flags);
 182         rsp = list_first_entry_or_null(&queue->free_rsps,
 183                                 struct nvmet_rdma_rsp, free_list);
 184         if (likely(rsp))
 185                 list_del(&rsp->free_list);
 186         spin_unlock_irqrestore(&queue->rsps_lock, flags);
 187
 188         if (unlikely(!rsp)) {
 189                 int ret;
 190
 191                 rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
 192                 if (unlikely(!rsp))
 193                         return NULL;
 194                 ret = nvmet_rdma_alloc_rsp(queue->dev, rsp);
 195                 if (unlikely(ret)) {
 196                         kfree(rsp);
 197                         return NULL;
 198                 }
 199
 200                 rsp->allocated = true;
 201         }
 202
 203         return rsp;
 204 }
 205
 206 static inline void
 207 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
 208 {
 209         unsigned long flags;
 210
 211         if (unlikely(rsp->allocated)) {
 212                 nvmet_rdma_free_rsp(rsp->queue->dev, rsp);
 213                 kfree(rsp);
 214                 return;
 215         }
 216
 217         spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
 218         list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
 219         spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
 220 }
 221
 222 static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
 223                                 struct nvmet_rdma_cmd *c)
 224 {
 225         struct scatterlist *sg;
 226         struct ib_sge *sge;
 227         int i;
 228
 229         if (!ndev->inline_data_size)
 230                 return;
 231
 232         sg = c->inline_sg;
 233         sge = &c->sge[1];
 234
 235         for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
 236                 if (sge->length)
 237                         ib_dma_unmap_page(ndev->device, sge->addr,
 238                                         sge->length, DMA_FROM_DEVICE);
 239                 if (sg_page(sg))
 240                         __free_page(sg_page(sg));
 241         }
 242 }
 243
 244 static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
 245                                 struct nvmet_rdma_cmd *c)
 246 {
 247         struct scatterlist *sg;
 248         struct ib_sge *sge;
 249         struct page *pg;
 250         int len;
 251         int i;
 252
 253         if (!ndev->inline_data_size)
 254                 return 0;
 255
 256         sg = c->inline_sg;
 257         sg_init_table(sg, ndev->inline_page_count);
 258         sge = &c->sge[1];
 259         len = ndev->inline_data_size;
 260
 261         for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
 262                 pg = alloc_page(GFP_KERNEL);
 263                 if (!pg)
 264                         goto out_err;
 265                 sg_assign_page(sg, pg);
 266                 sge->addr = ib_dma_map_page(ndev->device,
 267                         pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
 268                 if (ib_dma_mapping_error(ndev->device, sge->addr))
 269                         goto out_err;
 270                 sge->length = min_t(int, len, PAGE_SIZE);
 271                 sge->lkey = ndev->pd->local_dma_lkey;
 272                 len -= sge->length;
 273         }
 274
 275         return 0;
 276 out_err:
 277         for (; i >= 0; i--, sg--, sge--) {
 278                 if (sge->length)
 279                         ib_dma_unmap_page(ndev->device, sge->addr,
 280                                         sge->length, DMA_FROM_DEVICE);
 281                 if (sg_page(sg))
 282                         __free_page(sg_page(sg));
 283         }
 284         return -ENOMEM;
 285 }
 286
 287 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
 288                         struct nvmet_rdma_cmd *c, bool admin)
 289 {
 290         /* NVMe command / RDMA RECV */
 291         c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
 292         if (!c->nvme_cmd)
 293                 goto out;
 294
 295         c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
 296                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 297         if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
 298                 goto out_free_cmd;
 299
 300         c->sge[0].length = sizeof(*c->nvme_cmd);
 301         c->sge[0].lkey = ndev->pd->local_dma_lkey;
 302
 303         if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
 304                 goto out_unmap_cmd;
 305
 306         c->cqe.done = nvmet_rdma_recv_done;
 307
 308         c->wr.wr_cqe = &c->cqe;
 309         c->wr.sg_list = c->sge;
 310         c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;
 311
 312         return 0;
 313
 314 out_unmap_cmd:
 315         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 316                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 317 out_free_cmd:
 318         kfree(c->nvme_cmd);
 319
 320 out:
 321         return -ENOMEM;
 322 }
 323
 324 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
 325                 struct nvmet_rdma_cmd *c, bool admin)
 326 {
 327         if (!admin)
 328                 nvmet_rdma_free_inline_pages(ndev, c);
 329         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 330                                 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 331         kfree(c->nvme_cmd);
 332 }
 333
 334 static struct nvmet_rdma_cmd *
 335 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
 336                 int nr_cmds, bool admin)
 337 {
 338         struct nvmet_rdma_cmd *cmds;
 339         int ret = -EINVAL, i;
 340
 341         cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
 342         if (!cmds)
 343                 goto out;
 344
 345         for (i = 0; i < nr_cmds; i++) {
 346                 ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
 347                 if (ret)
 348                         goto out_free;
 349         }
 350
 351         return cmds;
 352
 353 out_free:
 354         while (--i >= 0)
 355                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 356         kfree(cmds);
 357 out:
 358         return ERR_PTR(ret);
 359 }
 360
 361 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
 362                 struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
 363 {
 364         int i;
 365
 366         for (i = 0; i < nr_cmds; i++)
 367                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 368         kfree(cmds);
 369 }
 370
 371 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
 372                 struct nvmet_rdma_rsp *r)
 373 {
 374         /* NVMe CQE / RDMA SEND */
 375         r->req.rsp = kmalloc(sizeof(*r->req.rsp), GFP_KERNEL);
 376         if (!r->req.rsp)
 377                 goto out;
 378
 379         r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.rsp,
 380                         sizeof(*r->req.rsp), DMA_TO_DEVICE);
 381         if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
 382                 goto out_free_rsp;
 383
 384         r->send_sge.length = sizeof(*r->req.rsp);
 385         r->send_sge.lkey = ndev->pd->local_dma_lkey;
 386
 387         r->send_cqe.done = nvmet_rdma_send_done;
 388
 389         r->send_wr.wr_cqe = &r->send_cqe;
 390         r->send_wr.sg_list = &r->send_sge;
 391         r->send_wr.num_sge = 1;
 392         r->send_wr.send_flags = IB_SEND_SIGNALED;
 393
 394         /* Data In / RDMA READ */
 395         r->read_cqe.done = nvmet_rdma_read_data_done;
 396         return 0;
 397
 398 out_free_rsp:
 399         kfree(r->req.rsp);
 400 out:
 401         return -ENOMEM;
 402 }
 403
 404 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
 405                 struct nvmet_rdma_rsp *r)
 406 {
 407         ib_dma_unmap_single(ndev->device, r->send_sge.addr,
 408                                 sizeof(*r->req.rsp), DMA_TO_DEVICE);
 409         kfree(r->req.rsp);
 410 }
 411
 412 static int
 413 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
 414 {
 415         struct nvmet_rdma_device *ndev = queue->dev;
 416         int nr_rsps = queue->recv_queue_size * 2;
 417         int ret = -EINVAL, i;
 418
 419         queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
 420                         GFP_KERNEL);
 421         if (!queue->rsps)
 422                 goto out;
 423
 424         for (i = 0; i < nr_rsps; i++) {
 425                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 426
 427                 ret = nvmet_rdma_alloc_rsp(ndev, rsp);
 428                 if (ret)
 429                         goto out_free;
 430
 431                 list_add_tail(&rsp->free_list, &queue->free_rsps);
 432         }
 433
 434         return 0;
 435
 436 out_free:
 437         while (--i >= 0) {
 438                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 439
 440                 list_del(&rsp->free_list);
 441                 nvmet_rdma_free_rsp(ndev, rsp);
 442         }
 443         kfree(queue->rsps);
 444 out:
 445         return ret;
 446 }
 447
 448 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
 449 {
 450         struct nvmet_rdma_device *ndev = queue->dev;
 451         int i, nr_rsps = queue->recv_queue_size * 2;
 452
 453         for (i = 0; i < nr_rsps; i++) {
 454                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 455
 456                 list_del(&rsp->free_list);
 457                 nvmet_rdma_free_rsp(ndev, rsp);
 458         }
 459         kfree(queue->rsps);
 460 }
 461
 462 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
 463                 struct nvmet_rdma_cmd *cmd)
 464 {
 465         int ret;
 466
 467         ib_dma_sync_single_for_device(ndev->device,
 468                 cmd->sge[0].addr, cmd->sge[0].length,
 469                 DMA_FROM_DEVICE);
 470
 471         if (ndev->srq)
 472                 ret = ib_post_srq_recv(ndev->srq, &cmd->wr, NULL);
 473         else
 474                 ret = ib_post_recv(cmd->queue->cm_id->qp, &cmd->wr, NULL);
 475
 476         if (unlikely(ret))
 477                 pr_err("post_recv cmd failed\n");
 478
 479         return ret;
 480 }
 481
 482 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
 483 {
 484         spin_lock(&queue->rsp_wr_wait_lock);
 485         while (!list_empty(&queue->rsp_wr_wait_list)) {
 486                 struct nvmet_rdma_rsp *rsp;
 487                 bool ret;
 488
 489                 rsp = list_entry(queue->rsp_wr_wait_list.next,
 490                                 struct nvmet_rdma_rsp, wait_list);
 491                 list_del(&rsp->wait_list);
 492
 493                 spin_unlock(&queue->rsp_wr_wait_lock);
 494                 ret = nvmet_rdma_execute_command(rsp);
 495                 spin_lock(&queue->rsp_wr_wait_lock);
 496
 497                 if (!ret) {
 498                         list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
 499                         break;
 500                 }
 501         }
 502         spin_unlock(&queue->rsp_wr_wait_lock);
 503 }
 504
 505
 506 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
 507 {
 508         struct nvmet_rdma_queue *queue = rsp->queue;
 509
 510         atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 511
 512         if (rsp->n_rdma) {
 513                 rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
 514                                 queue->cm_id->port_num, rsp->req.sg,
 515                                 rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
 516         }
 517
 518         if (rsp->req.sg != rsp->cmd->inline_sg)
 519                 nvmet_req_free_sgl(&rsp->req);
 520
 521         if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
 522                 nvmet_rdma_process_wr_wait_list(queue);
 523
 524         nvmet_rdma_put_rsp(rsp);
 525 }
 526
 527 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
 528 {
 529         if (queue->nvme_sq.ctrl) {
 530                 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
 531         } else {
 532                 /*
 533                  * we didn't setup the controller yet in case
 534                  * of admin connect error, just disconnect and
 535                  * cleanup the queue
 536                  */
 537                 nvmet_rdma_queue_disconnect(queue);
 538         }
 539 }
 540
 541 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
 542 {
 543         struct nvmet_rdma_rsp *rsp =
 544                 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
 545         struct nvmet_rdma_queue *queue = cq->cq_context;
 546
 547         nvmet_rdma_release_rsp(rsp);
 548
 549         if (unlikely(wc->status != IB_WC_SUCCESS &&
 550                      wc->status != IB_WC_WR_FLUSH_ERR)) {
 551                 pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
 552                         wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
 553                 nvmet_rdma_error_comp(queue);
 554         }
 555 }
 556
 557 static void nvmet_rdma_queue_response(struct nvmet_req *req)
 558 {
 559         struct nvmet_rdma_rsp *rsp =
 560                 container_of(req, struct nvmet_rdma_rsp, req);
 561         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 562         struct ib_send_wr *first_wr;
 563
 564         if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
 565                 rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
 566                 rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
 567         } else {
 568                 rsp->send_wr.opcode = IB_WR_SEND;
 569         }
 570
 571         if (nvmet_rdma_need_data_out(rsp))
 572                 first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
 573                                 cm_id->port_num, NULL, &rsp->send_wr);
 574         else
 575                 first_wr = &rsp->send_wr;
 576
 577         nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
 578
 579         ib_dma_sync_single_for_device(rsp->queue->dev->device,
 580                 rsp->send_sge.addr, rsp->send_sge.length,
 581                 DMA_TO_DEVICE);
 582
 583         if (unlikely(ib_post_send(cm_id->qp, first_wr, NULL))) {
 584                 pr_err("sending cmd response failed\n");
 585                 nvmet_rdma_release_rsp(rsp);
 586         }
 587 }
 588
 589 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
 590 {
 591         struct nvmet_rdma_rsp *rsp =
 592                 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
 593         struct nvmet_rdma_queue *queue = cq->cq_context;
 594
 595         WARN_ON(rsp->n_rdma <= 0);
 596         atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
 597         rdma_rw_ctx_destroy(&rsp->rw, queue->cm_id->qp,
 598                         queue->cm_id->port_num, rsp->req.sg,
 599                         rsp->req.sg_cnt, nvmet_data_dir(&rsp->req));
 600         rsp->n_rdma = 0;
 601
 602         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 603                 nvmet_req_uninit(&rsp->req);
 604                 nvmet_rdma_release_rsp(rsp);
 605                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 606                         pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
 607                                 wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
 608                         nvmet_rdma_error_comp(queue);
 609                 }
 610                 return;
 611         }
 612
 613         nvmet_req_execute(&rsp->req);
 614 }
 615
 616 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
 617                 u64 off)
 618 {
 619         int sg_count = num_pages(len);
 620         struct scatterlist *sg;
 621         int i;
 622
 623         sg = rsp->cmd->inline_sg;
 624         for (i = 0; i < sg_count; i++, sg++) {
 625                 if (i < sg_count - 1)
 626                         sg_unmark_end(sg);
 627                 else
 628                         sg_mark_end(sg);
 629                 sg->offset = off;
 630                 sg->length = min_t(int, len, PAGE_SIZE - off);
 631                 len -= sg->length;
 632                 if (!i)
 633                         off = 0;
 634         }
 635
 636         rsp->req.sg = rsp->cmd->inline_sg;
 637         rsp->req.sg_cnt = sg_count;
 638 }
 639
 640 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
 641 {
 642         struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
 643         u64 off = le64_to_cpu(sgl->addr);
 644         u32 len = le32_to_cpu(sgl->length);
 645
 646         if (!nvme_is_write(rsp->req.cmd)) {
 647                 rsp->req.error_loc =
 648                         offsetof(struct nvme_common_command, opcode);
 649                 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 650         }
 651
 652         if (off + len > rsp->queue->dev->inline_data_size) {
 653                 pr_err("invalid inline data offset!\n");
 654                 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
 655         }
 656
 657         /* no data command? */
 658         if (!len)
 659                 return 0;
 660
 661         nvmet_rdma_use_inline_sg(rsp, len, off);
 662         rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
 663         rsp->req.transfer_len += len;
 664         return 0;
 665 }
 666
 667 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
 668                 struct nvme_keyed_sgl_desc *sgl, bool invalidate)
 669 {
 670         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 671         u64 addr = le64_to_cpu(sgl->addr);
 672         u32 key = get_unaligned_le32(sgl->key);
 673         int ret;
 674
 675         rsp->req.transfer_len = get_unaligned_le24(sgl->length);
 676
 677         /* no data command? */
 678         if (!rsp->req.transfer_len)
 679                 return 0;
 680
 681         ret = nvmet_req_alloc_sgl(&rsp->req);
 682         if (ret < 0)
 683                 goto error_out;
 684
 685         ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
 686                         rsp->req.sg, rsp->req.sg_cnt, 0, addr, key,
 687                         nvmet_data_dir(&rsp->req));
 688         if (ret < 0)
 689                 goto error_out;
 690         rsp->n_rdma += ret;
 691
 692         if (invalidate) {
 693                 rsp->invalidate_rkey = key;
 694                 rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
 695         }
 696
 697         return 0;
 698
 699 error_out:
 700         rsp->req.transfer_len = 0;
 701         return NVME_SC_INTERNAL;
 702 }
 703
 704 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
 705 {
 706         struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
 707
 708         switch (sgl->type >> 4) {
 709         case NVME_SGL_FMT_DATA_DESC:
 710                 switch (sgl->type & 0xf) {
 711                 case NVME_SGL_FMT_OFFSET:
 712                         return nvmet_rdma_map_sgl_inline(rsp);
 713                 default:
 714                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 715                         rsp->req.error_loc =
 716                                 offsetof(struct nvme_common_command, dptr);
 717                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 718                 }
 719         case NVME_KEY_SGL_FMT_DATA_DESC:
 720                 switch (sgl->type & 0xf) {
 721                 case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
 722                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
 723                 case NVME_SGL_FMT_ADDRESS:
 724                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
 725                 default:
 726                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 727                         rsp->req.error_loc =
 728                                 offsetof(struct nvme_common_command, dptr);
 729                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 730                 }
 731         default:
 732                 pr_err("invalid SGL type: %#x\n", sgl->type);
 733                 rsp->req.error_loc = offsetof(struct nvme_common_command, dptr);
 734                 return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
 735         }
 736 }
 737
 738 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
 739 {
 740         struct nvmet_rdma_queue *queue = rsp->queue;
 741
 742         if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
 743                         &queue->sq_wr_avail) < 0)) {
 744                 pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
 745                                 1 + rsp->n_rdma, queue->idx,
 746                                 queue->nvme_sq.ctrl->cntlid);
 747                 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 748                 return false;
 749         }
 750
 751         if (nvmet_rdma_need_data_in(rsp)) {
 752                 if (rdma_rw_ctx_post(&rsp->rw, queue->cm_id->qp,
 753                                 queue->cm_id->port_num, &rsp->read_cqe, NULL))
 754                         nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
 755         } else {
 756                 nvmet_req_execute(&rsp->req);
 757         }
 758
 759         return true;
 760 }
 761
 762 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
 763                 struct nvmet_rdma_rsp *cmd)
 764 {
 765         u16 status;
 766
 767         ib_dma_sync_single_for_cpu(queue->dev->device,
 768                 cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
 769                 DMA_FROM_DEVICE);
 770         ib_dma_sync_single_for_cpu(queue->dev->device,
 771                 cmd->send_sge.addr, cmd->send_sge.length,
 772                 DMA_TO_DEVICE);
 773
 774         cmd->req.p2p_client = &queue->dev->device->dev;
 775
 776         if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
 777                         &queue->nvme_sq, &nvmet_rdma_ops))
 778                 return;
 779
 780         status = nvmet_rdma_map_sgl(cmd);
 781         if (status)
 782                 goto out_err;
 783
 784         if (unlikely(!nvmet_rdma_execute_command(cmd))) {
 785                 spin_lock(&queue->rsp_wr_wait_lock);
 786                 list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
 787                 spin_unlock(&queue->rsp_wr_wait_lock);
 788         }
 789
 790         return;
 791
 792 out_err:
 793         nvmet_req_complete(&cmd->req, status);
 794 }
 795
 796 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
 797 {
 798         struct nvmet_rdma_cmd *cmd =
 799                 container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
 800         struct nvmet_rdma_queue *queue = cq->cq_context;
 801         struct nvmet_rdma_rsp *rsp;
 802
 803         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 804                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 805                         pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
 806                                 wc->wr_cqe, ib_wc_status_msg(wc->status),
 807                                 wc->status);
 808                         nvmet_rdma_error_comp(queue);
 809                 }
 810                 return;
 811         }
 812
 813         if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
 814                 pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
 815                 nvmet_rdma_error_comp(queue);
 816                 return;
 817         }
 818
 819         cmd->queue = queue;
 820         rsp = nvmet_rdma_get_rsp(queue);
 821         if (unlikely(!rsp)) {
 822                 /*
 823                  * we get here only under memory pressure,
 824                  * silently drop and have the host retry
 825                  * as we can't even fail it.
 826                  */
 827                 nvmet_rdma_post_recv(queue->dev, cmd);
 828                 return;
 829         }
 830         rsp->queue = queue;
 831         rsp->cmd = cmd;
 832         rsp->flags = 0;
 833         rsp->req.cmd = cmd->nvme_cmd;
 834         rsp->req.port = queue->port;
 835         rsp->n_rdma = 0;
 836
 837         if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
 838                 unsigned long flags;
 839
 840                 spin_lock_irqsave(&queue->state_lock, flags);
 841                 if (queue->state == NVMET_RDMA_Q_CONNECTING)
 842                         list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
 843                 else
 844                         nvmet_rdma_put_rsp(rsp);
 845                 spin_unlock_irqrestore(&queue->state_lock, flags);
 846                 return;
 847         }
 848
 849         nvmet_rdma_handle_command(queue, rsp);
 850 }
 851
 852 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_device *ndev)
 853 {
 854         if (!ndev->srq)
 855                 return;
 856
 857         nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
 858         ib_destroy_srq(ndev->srq);
 859 }
 860
 861 static int nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
 862 {
 863         struct ib_srq_init_attr srq_attr = { NULL, };
 864         struct ib_srq *srq;
 865         size_t srq_size;
 866         int ret, i;
 867
 868         srq_size = 4095;        /* XXX: tune */
 869
 870         srq_attr.attr.max_wr = srq_size;
 871         srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
 872         srq_attr.attr.srq_limit = 0;
 873         srq_attr.srq_type = IB_SRQT_BASIC;
 874         srq = ib_create_srq(ndev->pd, &srq_attr);
 875         if (IS_ERR(srq)) {
 876                 /*
 877                  * If SRQs aren't supported we just go ahead and use normal
 878                  * non-shared receive queues.
 879                  */
 880                 pr_info("SRQ requested but not supported.\n");
 881                 return 0;
 882         }
 883
 884         ndev->srq_cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
 885         if (IS_ERR(ndev->srq_cmds)) {
 886                 ret = PTR_ERR(ndev->srq_cmds);
 887                 goto out_destroy_srq;
 888         }
 889
 890         ndev->srq = srq;
 891         ndev->srq_size = srq_size;
 892
 893         for (i = 0; i < srq_size; i++) {
 894                 ret = nvmet_rdma_post_recv(ndev, &ndev->srq_cmds[i]);
 895                 if (ret)
 896                         goto out_free_cmds;
 897         }
 898
 899         return 0;
 900
 901 out_free_cmds:
 902         nvmet_rdma_free_cmds(ndev, ndev->srq_cmds, ndev->srq_size, false);
 903 out_destroy_srq:
 904         ib_destroy_srq(srq);
 905         return ret;
 906 }
 907
 908 static void nvmet_rdma_free_dev(struct kref *ref)
 909 {
 910         struct nvmet_rdma_device *ndev =
 911                 container_of(ref, struct nvmet_rdma_device, ref);
 912
 913         mutex_lock(&device_list_mutex);
 914         list_del(&ndev->entry);
 915         mutex_unlock(&device_list_mutex);
 916
 917         nvmet_rdma_destroy_srq(ndev);
 918         ib_dealloc_pd(ndev->pd);
 919
 920         kfree(ndev);
 921 }
 922
 923 static struct nvmet_rdma_device *
 924 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
 925 {
 926         struct nvmet_port *port = cm_id->context;
 927         struct nvmet_rdma_device *ndev;
 928         int inline_page_count;
 929         int inline_sge_count;
 930         int ret;
 931
 932         mutex_lock(&device_list_mutex);
 933         list_for_each_entry(ndev, &device_list, entry) {
 934                 if (ndev->device->node_guid == cm_id->device->node_guid &&
 935                     kref_get_unless_zero(&ndev->ref))
 936                         goto out_unlock;
 937         }
 938
 939         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
 940         if (!ndev)
 941                 goto out_err;
 942
 943         inline_page_count = num_pages(port->inline_data_size);
 944         inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
 945                                 cm_id->device->attrs.max_recv_sge) - 1;
 946         if (inline_page_count > inline_sge_count) {
 947                 pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
 948                         port->inline_data_size, cm_id->device->name,
 949                         inline_sge_count * PAGE_SIZE);
 950                 port->inline_data_size = inline_sge_count * PAGE_SIZE;
 951                 inline_page_count = inline_sge_count;
 952         }
 953         ndev->inline_data_size = port->inline_data_size;
 954         ndev->inline_page_count = inline_page_count;
 955         ndev->device = cm_id->device;
 956         kref_init(&ndev->ref);
 957
 958         ndev->pd = ib_alloc_pd(ndev->device, 0);
 959         if (IS_ERR(ndev->pd))
 960                 goto out_free_dev;
 961
 962         if (nvmet_rdma_use_srq) {
 963                 ret = nvmet_rdma_init_srq(ndev);
 964                 if (ret)
 965                         goto out_free_pd;
 966         }
 967
 968         list_add(&ndev->entry, &device_list);
 969 out_unlock:
 970         mutex_unlock(&device_list_mutex);
 971         pr_debug("added %s.\n", ndev->device->name);
 972         return ndev;
 973
 974 out_free_pd:
 975         ib_dealloc_pd(ndev->pd);
 976 out_free_dev:
 977         kfree(ndev);
 978 out_err:
 979         mutex_unlock(&device_list_mutex);
 980         return NULL;
 981 }
 982
 983 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
 984 {
 985         struct ib_qp_init_attr qp_attr;
 986         struct nvmet_rdma_device *ndev = queue->dev;
 987         int comp_vector, nr_cqe, ret, i;
 988
 989         /*
 990          * Spread the io queues across completion vectors,
 991          * but still keep all admin queues on vector 0.
 992          */
 993         comp_vector = !queue->host_qid ? 0 :
 994                 queue->idx % ndev->device->num_comp_vectors;
 995
 996         /*
 997          * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
 998          */
 999         nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
1000
1001         queue->cq = ib_alloc_cq(ndev->device, queue,
1002                         nr_cqe + 1, comp_vector,
1003                         IB_POLL_WORKQUEUE);
1004         if (IS_ERR(queue->cq)) {
1005                 ret = PTR_ERR(queue->cq);
1006                 pr_err("failed to create CQ cqe= %d ret= %d\n",
1007                        nr_cqe + 1, ret);
1008                 goto out;
1009         }
1010
1011         memset(&qp_attr, 0, sizeof(qp_attr));
1012         qp_attr.qp_context = queue;
1013         qp_attr.event_handler = nvmet_rdma_qp_event;
1014         qp_attr.send_cq = queue->cq;
1015         qp_attr.recv_cq = queue->cq;
1016         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1017         qp_attr.qp_type = IB_QPT_RC;
1018         /* +1 for drain */
1019         qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
1020         qp_attr.cap.max_rdma_ctxs = queue->send_queue_size;
1021         qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
1022                                         ndev->device->attrs.max_send_sge);
1023
1024         if (ndev->srq) {
1025                 qp_attr.srq = ndev->srq;
1026         } else {
1027                 /* +1 for drain */
1028                 qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
1029                 qp_attr.cap.max_recv_sge = 1 + ndev->inline_page_count;
1030         }
1031
1032         ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
1033         if (ret) {
1034                 pr_err("failed to create_qp ret= %d\n", ret);
1035                 goto err_destroy_cq;
1036         }
1037
1038         atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
1039
1040         pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
1041                  __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
1042                  qp_attr.cap.max_send_wr, queue->cm_id);
1043
1044         if (!ndev->srq) {
1045                 for (i = 0; i < queue->recv_queue_size; i++) {
1046                         queue->cmds[i].queue = queue;
1047                         ret = nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
1048                         if (ret)
1049                                 goto err_destroy_qp;
1050                 }
1051         }
1052
1053 out:
1054         return ret;
1055
1056 err_destroy_qp:
1057         rdma_destroy_qp(queue->cm_id);
1058 err_destroy_cq:
1059         ib_free_cq(queue->cq);
1060         goto out;
1061 }
1062
1063 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1064 {
1065         struct ib_qp *qp = queue->cm_id->qp;
1066
1067         ib_drain_qp(qp);
1068         rdma_destroy_id(queue->cm_id);
1069         ib_destroy_qp(qp);
1070         ib_free_cq(queue->cq);
1071 }
1072
1073 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1074 {
1075         pr_debug("freeing queue %d\n", queue->idx);
1076
1077         nvmet_sq_destroy(&queue->nvme_sq);
1078
1079         nvmet_rdma_destroy_queue_ib(queue);
1080         if (!queue->dev->srq) {
1081                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1082                                 queue->recv_queue_size,
1083                                 !queue->host_qid);
1084         }
1085         nvmet_rdma_free_rsps(queue);
1086         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1087         kfree(queue);
1088 }
1089
1090 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1091 {
1092         struct nvmet_rdma_queue *queue =
1093                 container_of(w, struct nvmet_rdma_queue, release_work);
1094         struct nvmet_rdma_device *dev = queue->dev;
1095
1096         nvmet_rdma_free_queue(queue);
1097
1098         kref_put(&dev->ref, nvmet_rdma_free_dev);
1099 }
1100
1101 static int
1102 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1103                                 struct nvmet_rdma_queue *queue)
1104 {
1105         struct nvme_rdma_cm_req *req;
1106
1107         req = (struct nvme_rdma_cm_req *)conn->private_data;
1108         if (!req || conn->private_data_len == 0)
1109                 return NVME_RDMA_CM_INVALID_LEN;
1110
1111         if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1112                 return NVME_RDMA_CM_INVALID_RECFMT;
1113
1114         queue->host_qid = le16_to_cpu(req->qid);
1115
1116         /*
1117          * req->hsqsize corresponds to our recv queue size plus 1
1118          * req->hrqsize corresponds to our send queue size
1119          */
1120         queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1121         queue->send_queue_size = le16_to_cpu(req->hrqsize);
1122
1123         if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
1124                 return NVME_RDMA_CM_INVALID_HSQSIZE;
1125
1126         /* XXX: Should we enforce some kind of max for IO queues? */
1127
1128         return 0;
1129 }
1130
1131 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1132                                 enum nvme_rdma_cm_status status)
1133 {
1134         struct nvme_rdma_cm_rej rej;
1135
1136         pr_debug("rejecting connect request: status %d (%s)\n",
1137                  status, nvme_rdma_cm_msg(status));
1138
1139         rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1140         rej.sts = cpu_to_le16(status);
1141
1142         return rdma_reject(cm_id, (void *)&rej, sizeof(rej));
1143 }
1144
1145 static struct nvmet_rdma_queue *
1146 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1147                 struct rdma_cm_id *cm_id,
1148                 struct rdma_cm_event *event)
1149 {
1150         struct nvmet_rdma_queue *queue;
1151         int ret;
1152
1153         queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1154         if (!queue) {
1155                 ret = NVME_RDMA_CM_NO_RSC;
1156                 goto out_reject;
1157         }
1158
1159         ret = nvmet_sq_init(&queue->nvme_sq);
1160         if (ret) {
1161                 ret = NVME_RDMA_CM_NO_RSC;
1162                 goto out_free_queue;
1163         }
1164
1165         ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1166         if (ret)
1167                 goto out_destroy_sq;
1168
1169         /*
1170          * Schedules the actual release because calling rdma_destroy_id from
1171          * inside a CM callback would trigger a deadlock. (great API design..)
1172          */
1173         INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1174         queue->dev = ndev;
1175         queue->cm_id = cm_id;
1176
1177         spin_lock_init(&queue->state_lock);
1178         queue->state = NVMET_RDMA_Q_CONNECTING;
1179         INIT_LIST_HEAD(&queue->rsp_wait_list);
1180         INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1181         spin_lock_init(&queue->rsp_wr_wait_lock);
1182         INIT_LIST_HEAD(&queue->free_rsps);
1183         spin_lock_init(&queue->rsps_lock);
1184         INIT_LIST_HEAD(&queue->queue_list);
1185
1186         queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1187         if (queue->idx < 0) {
1188                 ret = NVME_RDMA_CM_NO_RSC;
1189                 goto out_destroy_sq;
1190         }
1191
1192         ret = nvmet_rdma_alloc_rsps(queue);
1193         if (ret) {
1194                 ret = NVME_RDMA_CM_NO_RSC;
1195                 goto out_ida_remove;
1196         }
1197
1198         if (!ndev->srq) {
1199                 queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1200                                 queue->recv_queue_size,
1201                                 !queue->host_qid);
1202                 if (IS_ERR(queue->cmds)) {
1203                         ret = NVME_RDMA_CM_NO_RSC;
1204                         goto out_free_responses;
1205                 }
1206         }
1207
1208         ret = nvmet_rdma_create_queue_ib(queue);
1209         if (ret) {
1210                 pr_err("%s: creating RDMA queue failed (%d).\n",
1211                         __func__, ret);
1212                 ret = NVME_RDMA_CM_NO_RSC;
1213                 goto out_free_cmds;
1214         }
1215
1216         return queue;
1217
1218 out_free_cmds:
1219         if (!ndev->srq) {
1220                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1221                                 queue->recv_queue_size,
1222                                 !queue->host_qid);
1223         }
1224 out_free_responses:
1225         nvmet_rdma_free_rsps(queue);
1226 out_ida_remove:
1227         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1228 out_destroy_sq:
1229         nvmet_sq_destroy(&queue->nvme_sq);
1230 out_free_queue:
1231         kfree(queue);
1232 out_reject:
1233         nvmet_rdma_cm_reject(cm_id, ret);
1234         return NULL;
1235 }
1236
1237 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1238 {
1239         struct nvmet_rdma_queue *queue = priv;
1240
1241         switch (event->event) {
1242         case IB_EVENT_COMM_EST:
1243                 rdma_notify(queue->cm_id, event->event);
1244                 break;
1245         default:
1246                 pr_err("received IB QP event: %s (%d)\n",
1247                        ib_event_msg(event->event), event->event);
1248                 break;
1249         }
1250 }
1251
1252 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1253                 struct nvmet_rdma_queue *queue,
1254                 struct rdma_conn_param *p)
1255 {
1256         struct rdma_conn_param  param = { };
1257         struct nvme_rdma_cm_rep priv = { };
1258         int ret = -ENOMEM;
1259
1260         param.rnr_retry_count = 7;
1261         param.flow_control = 1;
1262         param.initiator_depth = min_t(u8, p->initiator_depth,
1263                 queue->dev->device->attrs.max_qp_init_rd_atom);
1264         param.private_data = &priv;
1265         param.private_data_len = sizeof(priv);
1266         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1267         priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1268
1269         ret = rdma_accept(cm_id, &param);
1270         if (ret)
1271                 pr_err("rdma_accept failed (error code = %d)\n", ret);
1272
1273         return ret;
1274 }
1275
1276 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1277                 struct rdma_cm_event *event)
1278 {
1279         struct nvmet_rdma_device *ndev;
1280         struct nvmet_rdma_queue *queue;
1281         int ret = -EINVAL;
1282
1283         ndev = nvmet_rdma_find_get_device(cm_id);
1284         if (!ndev) {
1285                 nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1286                 return -ECONNREFUSED;
1287         }
1288
1289         queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1290         if (!queue) {
1291                 ret = -ENOMEM;
1292                 goto put_device;
1293         }
1294         queue->port = cm_id->context;
1295
1296         if (queue->host_qid == 0) {
1297                 /* Let inflight controller teardown complete */
1298                 flush_scheduled_work();
1299         }
1300
1301         ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1302         if (ret) {
1303                 schedule_work(&queue->release_work);
1304                 /* Destroying rdma_cm id is not needed here */
1305                 return 0;
1306         }
1307
1308         mutex_lock(&nvmet_rdma_queue_mutex);
1309         list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1310         mutex_unlock(&nvmet_rdma_queue_mutex);
1311
1312         return 0;
1313
1314 put_device:
1315         kref_put(&ndev->ref, nvmet_rdma_free_dev);
1316
1317         return ret;
1318 }
1319
1320 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1321 {
1322         unsigned long flags;
1323
1324         spin_lock_irqsave(&queue->state_lock, flags);
1325         if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1326                 pr_warn("trying to establish a connected queue\n");
1327                 goto out_unlock;
1328         }
1329         queue->state = NVMET_RDMA_Q_LIVE;
1330
1331         while (!list_empty(&queue->rsp_wait_list)) {
1332                 struct nvmet_rdma_rsp *cmd;
1333
1334                 cmd = list_first_entry(&queue->rsp_wait_list,
1335                                         struct nvmet_rdma_rsp, wait_list);
1336                 list_del(&cmd->wait_list);
1337
1338                 spin_unlock_irqrestore(&queue->state_lock, flags);
1339                 nvmet_rdma_handle_command(queue, cmd);
1340                 spin_lock_irqsave(&queue->state_lock, flags);
1341         }
1342
1343 out_unlock:
1344         spin_unlock_irqrestore(&queue->state_lock, flags);
1345 }
1346
1347 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1348 {
1349         bool disconnect = false;
1350         unsigned long flags;
1351
1352         pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1353
1354         spin_lock_irqsave(&queue->state_lock, flags);
1355         switch (queue->state) {
1356         case NVMET_RDMA_Q_CONNECTING:
1357         case NVMET_RDMA_Q_LIVE:
1358                 queue->state = NVMET_RDMA_Q_DISCONNECTING;
1359                 disconnect = true;
1360                 break;
1361         case NVMET_RDMA_Q_DISCONNECTING:
1362                 break;
1363         }
1364         spin_unlock_irqrestore(&queue->state_lock, flags);
1365
1366         if (disconnect) {
1367                 rdma_disconnect(queue->cm_id);
1368                 schedule_work(&queue->release_work);
1369         }
1370 }
1371
1372 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1373 {
1374         bool disconnect = false;
1375
1376         mutex_lock(&nvmet_rdma_queue_mutex);
1377         if (!list_empty(&queue->queue_list)) {
1378                 list_del_init(&queue->queue_list);
1379                 disconnect = true;
1380         }
1381         mutex_unlock(&nvmet_rdma_queue_mutex);
1382
1383         if (disconnect)
1384                 __nvmet_rdma_queue_disconnect(queue);
1385 }
1386
1387 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1388                 struct nvmet_rdma_queue *queue)
1389 {
1390         WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1391
1392         mutex_lock(&nvmet_rdma_queue_mutex);
1393         if (!list_empty(&queue->queue_list))
1394                 list_del_init(&queue->queue_list);
1395         mutex_unlock(&nvmet_rdma_queue_mutex);
1396
1397         pr_err("failed to connect queue %d\n", queue->idx);
1398         schedule_work(&queue->release_work);
1399 }
1400
1401 /**
1402  * nvme_rdma_device_removal() - Handle RDMA device removal
1403  * @cm_id:      rdma_cm id, used for nvmet port
1404  * @queue:      nvmet rdma queue (cm id qp_context)
1405  *
1406  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1407  * to unplug. Note that this event can be generated on a normal
1408  * queue cm_id and/or a device bound listener cm_id (where in this
1409  * case queue will be null).
1410  *
1411  * We registered an ib_client to handle device removal for queues,
1412  * so we only need to handle the listening port cm_ids. In this case
1413  * we nullify the priv to prevent double cm_id destruction and destroying
1414  * the cm_id implicitely by returning a non-zero rc to the callout.
1415  */
1416 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1417                 struct nvmet_rdma_queue *queue)
1418 {
1419         struct nvmet_port *port;
1420
1421         if (queue) {
1422                 /*
1423                  * This is a queue cm_id. we have registered
1424                  * an ib_client to handle queues removal
1425                  * so don't interfear and just return.
1426                  */
1427                 return 0;
1428         }
1429
1430         port = cm_id->context;
1431
1432         /*
1433          * This is a listener cm_id. Make sure that
1434          * future remove_port won't invoke a double
1435          * cm_id destroy. use atomic xchg to make sure
1436          * we don't compete with remove_port.
1437          */
1438         if (xchg(&port->priv, NULL) != cm_id)
1439                 return 0;
1440
1441         /*
1442          * We need to return 1 so that the core will destroy
1443          * it's own ID.  What a great API design..
1444          */
1445         return 1;
1446 }
1447
1448 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1449                 struct rdma_cm_event *event)
1450 {
1451         struct nvmet_rdma_queue *queue = NULL;
1452         int ret = 0;
1453
1454         if (cm_id->qp)
1455                 queue = cm_id->qp->qp_context;
1456
1457         pr_debug("%s (%d): status %d id %p\n",
1458                 rdma_event_msg(event->event), event->event,
1459                 event->status, cm_id);
1460
1461         switch (event->event) {
1462         case RDMA_CM_EVENT_CONNECT_REQUEST:
1463                 ret = nvmet_rdma_queue_connect(cm_id, event);
1464                 break;
1465         case RDMA_CM_EVENT_ESTABLISHED:
1466                 nvmet_rdma_queue_established(queue);
1467                 break;
1468         case RDMA_CM_EVENT_ADDR_CHANGE:
1469         case RDMA_CM_EVENT_DISCONNECTED:
1470         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1471                 nvmet_rdma_queue_disconnect(queue);
1472                 break;
1473         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1474                 ret = nvmet_rdma_device_removal(cm_id, queue);
1475                 break;
1476         case RDMA_CM_EVENT_REJECTED:
1477                 pr_debug("Connection rejected: %s\n",
1478                          rdma_reject_msg(cm_id, event->status));
1479                 /* FALLTHROUGH */
1480         case RDMA_CM_EVENT_UNREACHABLE:
1481         case RDMA_CM_EVENT_CONNECT_ERROR:
1482                 nvmet_rdma_queue_connect_fail(cm_id, queue);
1483                 break;
1484         default:
1485                 pr_err("received unrecognized RDMA CM event %d\n",
1486                         event->event);
1487                 break;
1488         }
1489
1490         return ret;
1491 }
1492
1493 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1494 {
1495         struct nvmet_rdma_queue *queue;
1496
1497 restart:
1498         mutex_lock(&nvmet_rdma_queue_mutex);
1499         list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1500                 if (queue->nvme_sq.ctrl == ctrl) {
1501                         list_del_init(&queue->queue_list);
1502                         mutex_unlock(&nvmet_rdma_queue_mutex);
1503
1504                         __nvmet_rdma_queue_disconnect(queue);
1505                         goto restart;
1506                 }
1507         }
1508         mutex_unlock(&nvmet_rdma_queue_mutex);
1509 }
1510
1511 static int nvmet_rdma_add_port(struct nvmet_port *port)
1512 {
1513         struct rdma_cm_id *cm_id;
1514         struct sockaddr_storage addr = { };
1515         __kernel_sa_family_t af;
1516         int ret;
1517
1518         switch (port->disc_addr.adrfam) {
1519         case NVMF_ADDR_FAMILY_IP4:
1520                 af = AF_INET;
1521                 break;
1522         case NVMF_ADDR_FAMILY_IP6:
1523                 af = AF_INET6;
1524                 break;
1525         default:
1526                 pr_err("address family %d not supported\n",
1527                                 port->disc_addr.adrfam);
1528                 return -EINVAL;
1529         }
1530
1531         if (port->inline_data_size < 0) {
1532                 port->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
1533         } else if (port->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
1534                 pr_warn("inline_data_size %u is too large, reducing to %u\n",
1535                         port->inline_data_size,
1536                         NVMET_RDMA_MAX_INLINE_DATA_SIZE);
1537                 port->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
1538         }
1539
1540         ret = inet_pton_with_scope(&init_net, af, port->disc_addr.traddr,
1541                         port->disc_addr.trsvcid, &addr);
1542         if (ret) {
1543                 pr_err("malformed ip/port passed: %s:%s\n",
1544                         port->disc_addr.traddr, port->disc_addr.trsvcid);
1545                 return ret;
1546         }
1547
1548         cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1549                         RDMA_PS_TCP, IB_QPT_RC);
1550         if (IS_ERR(cm_id)) {
1551                 pr_err("CM ID creation failed\n");
1552                 return PTR_ERR(cm_id);
1553         }
1554
1555         /*
1556          * Allow both IPv4 and IPv6 sockets to bind a single port
1557          * at the same time.
1558          */
1559         ret = rdma_set_afonly(cm_id, 1);
1560         if (ret) {
1561                 pr_err("rdma_set_afonly failed (%d)\n", ret);
1562                 goto out_destroy_id;
1563         }
1564
1565         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&addr);
1566         if (ret) {
1567                 pr_err("binding CM ID to %pISpcs failed (%d)\n",
1568                         (struct sockaddr *)&addr, ret);
1569                 goto out_destroy_id;
1570         }
1571
1572         ret = rdma_listen(cm_id, 128);
1573         if (ret) {
1574                 pr_err("listening to %pISpcs failed (%d)\n",
1575                         (struct sockaddr *)&addr, ret);
1576                 goto out_destroy_id;
1577         }
1578
1579         pr_info("enabling port %d (%pISpcs)\n",
1580                 le16_to_cpu(port->disc_addr.portid), (struct sockaddr *)&addr);
1581         port->priv = cm_id;
1582         return 0;
1583
1584 out_destroy_id:
1585         rdma_destroy_id(cm_id);
1586         return ret;
1587 }
1588
1589 static void nvmet_rdma_remove_port(struct nvmet_port *port)
1590 {
1591         struct rdma_cm_id *cm_id = xchg(&port->priv, NULL);
1592
1593         if (cm_id)
1594                 rdma_destroy_id(cm_id);
1595 }
1596
1597 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
1598                 struct nvmet_port *port, char *traddr)
1599 {
1600         struct rdma_cm_id *cm_id = port->priv;
1601
1602         if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
1603                 struct nvmet_rdma_rsp *rsp =
1604                         container_of(req, struct nvmet_rdma_rsp, req);
1605                 struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
1606                 struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;
1607
1608                 sprintf(traddr, "%pISc", addr);
1609         } else {
1610                 memcpy(traddr, port->disc_addr.traddr, NVMF_TRADDR_SIZE);
1611         }
1612 }
1613
1614 static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
1615         .owner                  = THIS_MODULE,
1616         .type                   = NVMF_TRTYPE_RDMA,
1617         .msdbd                  = 1,
1618         .has_keyed_sgls         = 1,
1619         .add_port               = nvmet_rdma_add_port,
1620         .remove_port            = nvmet_rdma_remove_port,
1621         .queue_response         = nvmet_rdma_queue_response,
1622         .delete_ctrl            = nvmet_rdma_delete_ctrl,
1623         .disc_traddr            = nvmet_rdma_disc_port_addr,
1624 };
1625
1626 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1627 {
1628         struct nvmet_rdma_queue *queue, *tmp;
1629         struct nvmet_rdma_device *ndev;
1630         bool found = false;
1631
1632         mutex_lock(&device_list_mutex);
1633         list_for_each_entry(ndev, &device_list, entry) {
1634                 if (ndev->device == ib_device) {
1635                         found = true;
1636                         break;
1637                 }
1638         }
1639         mutex_unlock(&device_list_mutex);
1640
1641         if (!found)
1642                 return;
1643
1644         /*
1645          * IB Device that is used by nvmet controllers is being removed,
1646          * delete all queues using this device.
1647          */
1648         mutex_lock(&nvmet_rdma_queue_mutex);
1649         list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
1650                                  queue_list) {
1651                 if (queue->dev->device != ib_device)
1652                         continue;
1653
1654                 pr_info("Removing queue %d\n", queue->idx);
1655                 list_del_init(&queue->queue_list);
1656                 __nvmet_rdma_queue_disconnect(queue);
1657         }
1658         mutex_unlock(&nvmet_rdma_queue_mutex);
1659
1660         flush_scheduled_work();
1661 }
1662
1663 static struct ib_client nvmet_rdma_ib_client = {
1664         .name   = "nvmet_rdma",
1665         .remove = nvmet_rdma_remove_one
1666 };
1667
1668 static int __init nvmet_rdma_init(void)
1669 {
1670         int ret;
1671
1672         ret = ib_register_client(&nvmet_rdma_ib_client);
1673         if (ret)
1674                 return ret;
1675
1676         ret = nvmet_register_transport(&nvmet_rdma_ops);
1677         if (ret)
1678                 goto err_ib_client;
1679
1680         return 0;
1681
1682 err_ib_client:
1683         ib_unregister_client(&nvmet_rdma_ib_client);
1684         return ret;
1685 }
1686
1687 static void __exit nvmet_rdma_exit(void)
1688 {
1689         nvmet_unregister_transport(&nvmet_rdma_ops);
1690         ib_unregister_client(&nvmet_rdma_ib_client);
1691         WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
1692         ida_destroy(&nvmet_rdma_queue_ida);
1693 }
1694
1695 module_init(nvmet_rdma_init);
1696 module_exit(nvmet_rdma_exit);
1697
1698 MODULE_LICENSE("GPL v2");
1699 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */