drivers/nvme/target/rdma.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * NVMe over Fabrics RDMA target.
   4  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
   5  */
   6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   7 #include <linux/atomic.h>
   8 #include <linux/ctype.h>
   9 #include <linux/delay.h>
  10 #include <linux/err.h>
  11 #include <linux/init.h>
  12 #include <linux/module.h>
  13 #include <linux/nvme.h>
  14 #include <linux/slab.h>
  15 #include <linux/string.h>
  16 #include <linux/wait.h>
  17 #include <linux/inet.h>
  18 #include <asm/unaligned.h>
  19
  20 #include <rdma/ib_verbs.h>
  21 #include <rdma/rdma_cm.h>
  22 #include <rdma/rw.h>
  23 #include <rdma/ib_cm.h>
  24
  25 #include <linux/nvme-rdma.h>
  26 #include "nvmet.h"
  27
  28 /*
  29  * We allow at least 1 page, up to 4 SGEs, and up to 16KB of inline data
  30  */
  31 #define NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE     PAGE_SIZE
  32 #define NVMET_RDMA_MAX_INLINE_SGE               4
  33 #define NVMET_RDMA_MAX_INLINE_DATA_SIZE         max_t(int, SZ_16K, PAGE_SIZE)
  34
  35 /* Assume mpsmin == device_page_size == 4KB */
  36 #define NVMET_RDMA_MAX_MDTS                     8
  37 #define NVMET_RDMA_MAX_METADATA_MDTS            5
  38
  39 struct nvmet_rdma_srq;
  40
  41 struct nvmet_rdma_cmd {
  42         struct ib_sge           sge[NVMET_RDMA_MAX_INLINE_SGE + 1];
  43         struct ib_cqe           cqe;
  44         struct ib_recv_wr       wr;
  45         struct scatterlist      inline_sg[NVMET_RDMA_MAX_INLINE_SGE];
  46         struct nvme_command     *nvme_cmd;
  47         struct nvmet_rdma_queue *queue;
  48         struct nvmet_rdma_srq   *nsrq;
  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 ib_cqe           write_cqe;
  66         struct rdma_rw_ctx      rw;
  67
  68         struct nvmet_req        req;
  69
  70         bool                    allocated;
  71         u8                      n_rdma;
  72         u32                     flags;
  73         u32                     invalidate_rkey;
  74
  75         struct list_head        wait_list;
  76         struct list_head        free_list;
  77 };
  78
  79 enum nvmet_rdma_queue_state {
  80         NVMET_RDMA_Q_CONNECTING,
  81         NVMET_RDMA_Q_LIVE,
  82         NVMET_RDMA_Q_DISCONNECTING,
  83 };
  84
  85 struct nvmet_rdma_queue {
  86         struct rdma_cm_id       *cm_id;
  87         struct ib_qp            *qp;
  88         struct nvmet_port       *port;
  89         struct ib_cq            *cq;
  90         atomic_t                sq_wr_avail;
  91         struct nvmet_rdma_device *dev;
  92         struct nvmet_rdma_srq   *nsrq;
  93         spinlock_t              state_lock;
  94         enum nvmet_rdma_queue_state state;
  95         struct nvmet_cq         nvme_cq;
  96         struct nvmet_sq         nvme_sq;
  97
  98         struct nvmet_rdma_rsp   *rsps;
  99         struct list_head        free_rsps;
 100         spinlock_t              rsps_lock;
 101         struct nvmet_rdma_cmd   *cmds;
 102
 103         struct work_struct      release_work;
 104         struct list_head        rsp_wait_list;
 105         struct list_head        rsp_wr_wait_list;
 106         spinlock_t              rsp_wr_wait_lock;
 107
 108         int                     idx;
 109         int                     host_qid;
 110         int                     comp_vector;
 111         int                     recv_queue_size;
 112         int                     send_queue_size;
 113
 114         struct list_head        queue_list;
 115 };
 116
 117 struct nvmet_rdma_port {
 118         struct nvmet_port       *nport;
 119         struct sockaddr_storage addr;
 120         struct rdma_cm_id       *cm_id;
 121         struct delayed_work     repair_work;
 122 };
 123
 124 struct nvmet_rdma_srq {
 125         struct ib_srq            *srq;
 126         struct nvmet_rdma_cmd    *cmds;
 127         struct nvmet_rdma_device *ndev;
 128 };
 129
 130 struct nvmet_rdma_device {
 131         struct ib_device        *device;
 132         struct ib_pd            *pd;
 133         struct nvmet_rdma_srq   **srqs;
 134         int                     srq_count;
 135         size_t                  srq_size;
 136         struct kref             ref;
 137         struct list_head        entry;
 138         int                     inline_data_size;
 139         int                     inline_page_count;
 140 };
 141
 142 static bool nvmet_rdma_use_srq;
 143 module_param_named(use_srq, nvmet_rdma_use_srq, bool, 0444);
 144 MODULE_PARM_DESC(use_srq, "Use shared receive queue.");
 145
 146 static int srq_size_set(const char *val, const struct kernel_param *kp);
 147 static const struct kernel_param_ops srq_size_ops = {
 148         .set = srq_size_set,
 149         .get = param_get_int,
 150 };
 151
 152 static int nvmet_rdma_srq_size = 1024;
 153 module_param_cb(srq_size, &srq_size_ops, &nvmet_rdma_srq_size, 0644);
 154 MODULE_PARM_DESC(srq_size, "set Shared Receive Queue (SRQ) size, should >= 256 (default: 1024)");
 155
 156 static DEFINE_IDA(nvmet_rdma_queue_ida);
 157 static LIST_HEAD(nvmet_rdma_queue_list);
 158 static DEFINE_MUTEX(nvmet_rdma_queue_mutex);
 159
 160 static LIST_HEAD(device_list);
 161 static DEFINE_MUTEX(device_list_mutex);
 162
 163 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp);
 164 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc);
 165 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
 166 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc);
 167 static void nvmet_rdma_write_data_done(struct ib_cq *cq, struct ib_wc *wc);
 168 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv);
 169 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue);
 170 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
 171                                 struct nvmet_rdma_rsp *r);
 172 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
 173                                 struct nvmet_rdma_rsp *r);
 174
 175 static const struct nvmet_fabrics_ops nvmet_rdma_ops;
 176
 177 static int srq_size_set(const char *val, const struct kernel_param *kp)
 178 {
 179         int n = 0, ret;
 180
 181         ret = kstrtoint(val, 10, &n);
 182         if (ret != 0 || n < 256)
 183                 return -EINVAL;
 184
 185         return param_set_int(val, kp);
 186 }
 187
 188 static int num_pages(int len)
 189 {
 190         return 1 + (((len - 1) & PAGE_MASK) >> PAGE_SHIFT);
 191 }
 192
 193 static inline bool nvmet_rdma_need_data_in(struct nvmet_rdma_rsp *rsp)
 194 {
 195         return nvme_is_write(rsp->req.cmd) &&
 196                 rsp->req.transfer_len &&
 197                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 198 }
 199
 200 static inline bool nvmet_rdma_need_data_out(struct nvmet_rdma_rsp *rsp)
 201 {
 202         return !nvme_is_write(rsp->req.cmd) &&
 203                 rsp->req.transfer_len &&
 204                 !rsp->req.cqe->status &&
 205                 !(rsp->flags & NVMET_RDMA_REQ_INLINE_DATA);
 206 }
 207
 208 static inline struct nvmet_rdma_rsp *
 209 nvmet_rdma_get_rsp(struct nvmet_rdma_queue *queue)
 210 {
 211         struct nvmet_rdma_rsp *rsp;
 212         unsigned long flags;
 213
 214         spin_lock_irqsave(&queue->rsps_lock, flags);
 215         rsp = list_first_entry_or_null(&queue->free_rsps,
 216                                 struct nvmet_rdma_rsp, free_list);
 217         if (likely(rsp))
 218                 list_del(&rsp->free_list);
 219         spin_unlock_irqrestore(&queue->rsps_lock, flags);
 220
 221         if (unlikely(!rsp)) {
 222                 int ret;
 223
 224                 rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
 225                 if (unlikely(!rsp))
 226                         return NULL;
 227                 ret = nvmet_rdma_alloc_rsp(queue->dev, rsp);
 228                 if (unlikely(ret)) {
 229                         kfree(rsp);
 230                         return NULL;
 231                 }
 232
 233                 rsp->allocated = true;
 234         }
 235
 236         return rsp;
 237 }
 238
 239 static inline void
 240 nvmet_rdma_put_rsp(struct nvmet_rdma_rsp *rsp)
 241 {
 242         unsigned long flags;
 243
 244         if (unlikely(rsp->allocated)) {
 245                 nvmet_rdma_free_rsp(rsp->queue->dev, rsp);
 246                 kfree(rsp);
 247                 return;
 248         }
 249
 250         spin_lock_irqsave(&rsp->queue->rsps_lock, flags);
 251         list_add_tail(&rsp->free_list, &rsp->queue->free_rsps);
 252         spin_unlock_irqrestore(&rsp->queue->rsps_lock, flags);
 253 }
 254
 255 static void nvmet_rdma_free_inline_pages(struct nvmet_rdma_device *ndev,
 256                                 struct nvmet_rdma_cmd *c)
 257 {
 258         struct scatterlist *sg;
 259         struct ib_sge *sge;
 260         int i;
 261
 262         if (!ndev->inline_data_size)
 263                 return;
 264
 265         sg = c->inline_sg;
 266         sge = &c->sge[1];
 267
 268         for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
 269                 if (sge->length)
 270                         ib_dma_unmap_page(ndev->device, sge->addr,
 271                                         sge->length, DMA_FROM_DEVICE);
 272                 if (sg_page(sg))
 273                         __free_page(sg_page(sg));
 274         }
 275 }
 276
 277 static int nvmet_rdma_alloc_inline_pages(struct nvmet_rdma_device *ndev,
 278                                 struct nvmet_rdma_cmd *c)
 279 {
 280         struct scatterlist *sg;
 281         struct ib_sge *sge;
 282         struct page *pg;
 283         int len;
 284         int i;
 285
 286         if (!ndev->inline_data_size)
 287                 return 0;
 288
 289         sg = c->inline_sg;
 290         sg_init_table(sg, ndev->inline_page_count);
 291         sge = &c->sge[1];
 292         len = ndev->inline_data_size;
 293
 294         for (i = 0; i < ndev->inline_page_count; i++, sg++, sge++) {
 295                 pg = alloc_page(GFP_KERNEL);
 296                 if (!pg)
 297                         goto out_err;
 298                 sg_assign_page(sg, pg);
 299                 sge->addr = ib_dma_map_page(ndev->device,
 300                         pg, 0, PAGE_SIZE, DMA_FROM_DEVICE);
 301                 if (ib_dma_mapping_error(ndev->device, sge->addr))
 302                         goto out_err;
 303                 sge->length = min_t(int, len, PAGE_SIZE);
 304                 sge->lkey = ndev->pd->local_dma_lkey;
 305                 len -= sge->length;
 306         }
 307
 308         return 0;
 309 out_err:
 310         for (; i >= 0; i--, sg--, sge--) {
 311                 if (sge->length)
 312                         ib_dma_unmap_page(ndev->device, sge->addr,
 313                                         sge->length, DMA_FROM_DEVICE);
 314                 if (sg_page(sg))
 315                         __free_page(sg_page(sg));
 316         }
 317         return -ENOMEM;
 318 }
 319
 320 static int nvmet_rdma_alloc_cmd(struct nvmet_rdma_device *ndev,
 321                         struct nvmet_rdma_cmd *c, bool admin)
 322 {
 323         /* NVMe command / RDMA RECV */
 324         c->nvme_cmd = kmalloc(sizeof(*c->nvme_cmd), GFP_KERNEL);
 325         if (!c->nvme_cmd)
 326                 goto out;
 327
 328         c->sge[0].addr = ib_dma_map_single(ndev->device, c->nvme_cmd,
 329                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 330         if (ib_dma_mapping_error(ndev->device, c->sge[0].addr))
 331                 goto out_free_cmd;
 332
 333         c->sge[0].length = sizeof(*c->nvme_cmd);
 334         c->sge[0].lkey = ndev->pd->local_dma_lkey;
 335
 336         if (!admin && nvmet_rdma_alloc_inline_pages(ndev, c))
 337                 goto out_unmap_cmd;
 338
 339         c->cqe.done = nvmet_rdma_recv_done;
 340
 341         c->wr.wr_cqe = &c->cqe;
 342         c->wr.sg_list = c->sge;
 343         c->wr.num_sge = admin ? 1 : ndev->inline_page_count + 1;
 344
 345         return 0;
 346
 347 out_unmap_cmd:
 348         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 349                         sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 350 out_free_cmd:
 351         kfree(c->nvme_cmd);
 352
 353 out:
 354         return -ENOMEM;
 355 }
 356
 357 static void nvmet_rdma_free_cmd(struct nvmet_rdma_device *ndev,
 358                 struct nvmet_rdma_cmd *c, bool admin)
 359 {
 360         if (!admin)
 361                 nvmet_rdma_free_inline_pages(ndev, c);
 362         ib_dma_unmap_single(ndev->device, c->sge[0].addr,
 363                                 sizeof(*c->nvme_cmd), DMA_FROM_DEVICE);
 364         kfree(c->nvme_cmd);
 365 }
 366
 367 static struct nvmet_rdma_cmd *
 368 nvmet_rdma_alloc_cmds(struct nvmet_rdma_device *ndev,
 369                 int nr_cmds, bool admin)
 370 {
 371         struct nvmet_rdma_cmd *cmds;
 372         int ret = -EINVAL, i;
 373
 374         cmds = kcalloc(nr_cmds, sizeof(struct nvmet_rdma_cmd), GFP_KERNEL);
 375         if (!cmds)
 376                 goto out;
 377
 378         for (i = 0; i < nr_cmds; i++) {
 379                 ret = nvmet_rdma_alloc_cmd(ndev, cmds + i, admin);
 380                 if (ret)
 381                         goto out_free;
 382         }
 383
 384         return cmds;
 385
 386 out_free:
 387         while (--i >= 0)
 388                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 389         kfree(cmds);
 390 out:
 391         return ERR_PTR(ret);
 392 }
 393
 394 static void nvmet_rdma_free_cmds(struct nvmet_rdma_device *ndev,
 395                 struct nvmet_rdma_cmd *cmds, int nr_cmds, bool admin)
 396 {
 397         int i;
 398
 399         for (i = 0; i < nr_cmds; i++)
 400                 nvmet_rdma_free_cmd(ndev, cmds + i, admin);
 401         kfree(cmds);
 402 }
 403
 404 static int nvmet_rdma_alloc_rsp(struct nvmet_rdma_device *ndev,
 405                 struct nvmet_rdma_rsp *r)
 406 {
 407         /* NVMe CQE / RDMA SEND */
 408         r->req.cqe = kmalloc(sizeof(*r->req.cqe), GFP_KERNEL);
 409         if (!r->req.cqe)
 410                 goto out;
 411
 412         r->send_sge.addr = ib_dma_map_single(ndev->device, r->req.cqe,
 413                         sizeof(*r->req.cqe), DMA_TO_DEVICE);
 414         if (ib_dma_mapping_error(ndev->device, r->send_sge.addr))
 415                 goto out_free_rsp;
 416
 417         r->req.p2p_client = &ndev->device->dev;
 418         r->send_sge.length = sizeof(*r->req.cqe);
 419         r->send_sge.lkey = ndev->pd->local_dma_lkey;
 420
 421         r->send_cqe.done = nvmet_rdma_send_done;
 422
 423         r->send_wr.wr_cqe = &r->send_cqe;
 424         r->send_wr.sg_list = &r->send_sge;
 425         r->send_wr.num_sge = 1;
 426         r->send_wr.send_flags = IB_SEND_SIGNALED;
 427
 428         /* Data In / RDMA READ */
 429         r->read_cqe.done = nvmet_rdma_read_data_done;
 430         /* Data Out / RDMA WRITE */
 431         r->write_cqe.done = nvmet_rdma_write_data_done;
 432
 433         return 0;
 434
 435 out_free_rsp:
 436         kfree(r->req.cqe);
 437 out:
 438         return -ENOMEM;
 439 }
 440
 441 static void nvmet_rdma_free_rsp(struct nvmet_rdma_device *ndev,
 442                 struct nvmet_rdma_rsp *r)
 443 {
 444         ib_dma_unmap_single(ndev->device, r->send_sge.addr,
 445                                 sizeof(*r->req.cqe), DMA_TO_DEVICE);
 446         kfree(r->req.cqe);
 447 }
 448
 449 static int
 450 nvmet_rdma_alloc_rsps(struct nvmet_rdma_queue *queue)
 451 {
 452         struct nvmet_rdma_device *ndev = queue->dev;
 453         int nr_rsps = queue->recv_queue_size * 2;
 454         int ret = -EINVAL, i;
 455
 456         queue->rsps = kcalloc(nr_rsps, sizeof(struct nvmet_rdma_rsp),
 457                         GFP_KERNEL);
 458         if (!queue->rsps)
 459                 goto out;
 460
 461         for (i = 0; i < nr_rsps; i++) {
 462                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 463
 464                 ret = nvmet_rdma_alloc_rsp(ndev, rsp);
 465                 if (ret)
 466                         goto out_free;
 467
 468                 list_add_tail(&rsp->free_list, &queue->free_rsps);
 469         }
 470
 471         return 0;
 472
 473 out_free:
 474         while (--i >= 0) {
 475                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 476
 477                 list_del(&rsp->free_list);
 478                 nvmet_rdma_free_rsp(ndev, rsp);
 479         }
 480         kfree(queue->rsps);
 481 out:
 482         return ret;
 483 }
 484
 485 static void nvmet_rdma_free_rsps(struct nvmet_rdma_queue *queue)
 486 {
 487         struct nvmet_rdma_device *ndev = queue->dev;
 488         int i, nr_rsps = queue->recv_queue_size * 2;
 489
 490         for (i = 0; i < nr_rsps; i++) {
 491                 struct nvmet_rdma_rsp *rsp = &queue->rsps[i];
 492
 493                 list_del(&rsp->free_list);
 494                 nvmet_rdma_free_rsp(ndev, rsp);
 495         }
 496         kfree(queue->rsps);
 497 }
 498
 499 static int nvmet_rdma_post_recv(struct nvmet_rdma_device *ndev,
 500                 struct nvmet_rdma_cmd *cmd)
 501 {
 502         int ret;
 503
 504         ib_dma_sync_single_for_device(ndev->device,
 505                 cmd->sge[0].addr, cmd->sge[0].length,
 506                 DMA_FROM_DEVICE);
 507
 508         if (cmd->nsrq)
 509                 ret = ib_post_srq_recv(cmd->nsrq->srq, &cmd->wr, NULL);
 510         else
 511                 ret = ib_post_recv(cmd->queue->qp, &cmd->wr, NULL);
 512
 513         if (unlikely(ret))
 514                 pr_err("post_recv cmd failed\n");
 515
 516         return ret;
 517 }
 518
 519 static void nvmet_rdma_process_wr_wait_list(struct nvmet_rdma_queue *queue)
 520 {
 521         spin_lock(&queue->rsp_wr_wait_lock);
 522         while (!list_empty(&queue->rsp_wr_wait_list)) {
 523                 struct nvmet_rdma_rsp *rsp;
 524                 bool ret;
 525
 526                 rsp = list_entry(queue->rsp_wr_wait_list.next,
 527                                 struct nvmet_rdma_rsp, wait_list);
 528                 list_del(&rsp->wait_list);
 529
 530                 spin_unlock(&queue->rsp_wr_wait_lock);
 531                 ret = nvmet_rdma_execute_command(rsp);
 532                 spin_lock(&queue->rsp_wr_wait_lock);
 533
 534                 if (!ret) {
 535                         list_add(&rsp->wait_list, &queue->rsp_wr_wait_list);
 536                         break;
 537                 }
 538         }
 539         spin_unlock(&queue->rsp_wr_wait_lock);
 540 }
 541
 542 static u16 nvmet_rdma_check_pi_status(struct ib_mr *sig_mr)
 543 {
 544         struct ib_mr_status mr_status;
 545         int ret;
 546         u16 status = 0;
 547
 548         ret = ib_check_mr_status(sig_mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
 549         if (ret) {
 550                 pr_err("ib_check_mr_status failed, ret %d\n", ret);
 551                 return NVME_SC_INVALID_PI;
 552         }
 553
 554         if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
 555                 switch (mr_status.sig_err.err_type) {
 556                 case IB_SIG_BAD_GUARD:
 557                         status = NVME_SC_GUARD_CHECK;
 558                         break;
 559                 case IB_SIG_BAD_REFTAG:
 560                         status = NVME_SC_REFTAG_CHECK;
 561                         break;
 562                 case IB_SIG_BAD_APPTAG:
 563                         status = NVME_SC_APPTAG_CHECK;
 564                         break;
 565                 }
 566                 pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
 567                        mr_status.sig_err.err_type,
 568                        mr_status.sig_err.expected,
 569                        mr_status.sig_err.actual);
 570         }
 571
 572         return status;
 573 }
 574
 575 static void nvmet_rdma_set_sig_domain(struct blk_integrity *bi,
 576                 struct nvme_command *cmd, struct ib_sig_domain *domain,
 577                 u16 control, u8 pi_type)
 578 {
 579         domain->sig_type = IB_SIG_TYPE_T10_DIF;
 580         domain->sig.dif.bg_type = IB_T10DIF_CRC;
 581         domain->sig.dif.pi_interval = 1 << bi->interval_exp;
 582         domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
 583         if (control & NVME_RW_PRINFO_PRCHK_REF)
 584                 domain->sig.dif.ref_remap = true;
 585
 586         domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
 587         domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
 588         domain->sig.dif.app_escape = true;
 589         if (pi_type == NVME_NS_DPS_PI_TYPE3)
 590                 domain->sig.dif.ref_escape = true;
 591 }
 592
 593 static void nvmet_rdma_set_sig_attrs(struct nvmet_req *req,
 594                                      struct ib_sig_attrs *sig_attrs)
 595 {
 596         struct nvme_command *cmd = req->cmd;
 597         u16 control = le16_to_cpu(cmd->rw.control);
 598         u8 pi_type = req->ns->pi_type;
 599         struct blk_integrity *bi;
 600
 601         bi = bdev_get_integrity(req->ns->bdev);
 602
 603         memset(sig_attrs, 0, sizeof(*sig_attrs));
 604
 605         if (control & NVME_RW_PRINFO_PRACT) {
 606                 /* for WRITE_INSERT/READ_STRIP no wire domain */
 607                 sig_attrs->wire.sig_type = IB_SIG_TYPE_NONE;
 608                 nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
 609                                           pi_type);
 610                 /* Clear the PRACT bit since HCA will generate/verify the PI */
 611                 control &= ~NVME_RW_PRINFO_PRACT;
 612                 cmd->rw.control = cpu_to_le16(control);
 613                 /* PI is added by the HW */
 614                 req->transfer_len += req->metadata_len;
 615         } else {
 616                 /* for WRITE_PASS/READ_PASS both wire/memory domains exist */
 617                 nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
 618                                           pi_type);
 619                 nvmet_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
 620                                           pi_type);
 621         }
 622
 623         if (control & NVME_RW_PRINFO_PRCHK_REF)
 624                 sig_attrs->check_mask |= IB_SIG_CHECK_REFTAG;
 625         if (control & NVME_RW_PRINFO_PRCHK_GUARD)
 626                 sig_attrs->check_mask |= IB_SIG_CHECK_GUARD;
 627         if (control & NVME_RW_PRINFO_PRCHK_APP)
 628                 sig_attrs->check_mask |= IB_SIG_CHECK_APPTAG;
 629 }
 630
 631 static int nvmet_rdma_rw_ctx_init(struct nvmet_rdma_rsp *rsp, u64 addr, u32 key,
 632                                   struct ib_sig_attrs *sig_attrs)
 633 {
 634         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 635         struct nvmet_req *req = &rsp->req;
 636         int ret;
 637
 638         if (req->metadata_len)
 639                 ret = rdma_rw_ctx_signature_init(&rsp->rw, cm_id->qp,
 640                         cm_id->port_num, req->sg, req->sg_cnt,
 641                         req->metadata_sg, req->metadata_sg_cnt, sig_attrs,
 642                         addr, key, nvmet_data_dir(req));
 643         else
 644                 ret = rdma_rw_ctx_init(&rsp->rw, cm_id->qp, cm_id->port_num,
 645                                        req->sg, req->sg_cnt, 0, addr, key,
 646                                        nvmet_data_dir(req));
 647
 648         return ret;
 649 }
 650
 651 static void nvmet_rdma_rw_ctx_destroy(struct nvmet_rdma_rsp *rsp)
 652 {
 653         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 654         struct nvmet_req *req = &rsp->req;
 655
 656         if (req->metadata_len)
 657                 rdma_rw_ctx_destroy_signature(&rsp->rw, cm_id->qp,
 658                         cm_id->port_num, req->sg, req->sg_cnt,
 659                         req->metadata_sg, req->metadata_sg_cnt,
 660                         nvmet_data_dir(req));
 661         else
 662                 rdma_rw_ctx_destroy(&rsp->rw, cm_id->qp, cm_id->port_num,
 663                                     req->sg, req->sg_cnt, nvmet_data_dir(req));
 664 }
 665
 666 static void nvmet_rdma_release_rsp(struct nvmet_rdma_rsp *rsp)
 667 {
 668         struct nvmet_rdma_queue *queue = rsp->queue;
 669
 670         atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 671
 672         if (rsp->n_rdma)
 673                 nvmet_rdma_rw_ctx_destroy(rsp);
 674
 675         if (rsp->req.sg != rsp->cmd->inline_sg)
 676                 nvmet_req_free_sgls(&rsp->req);
 677
 678         if (unlikely(!list_empty_careful(&queue->rsp_wr_wait_list)))
 679                 nvmet_rdma_process_wr_wait_list(queue);
 680
 681         nvmet_rdma_put_rsp(rsp);
 682 }
 683
 684 static void nvmet_rdma_error_comp(struct nvmet_rdma_queue *queue)
 685 {
 686         if (queue->nvme_sq.ctrl) {
 687                 nvmet_ctrl_fatal_error(queue->nvme_sq.ctrl);
 688         } else {
 689                 /*
 690                  * we didn't setup the controller yet in case
 691                  * of admin connect error, just disconnect and
 692                  * cleanup the queue
 693                  */
 694                 nvmet_rdma_queue_disconnect(queue);
 695         }
 696 }
 697
 698 static void nvmet_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
 699 {
 700         struct nvmet_rdma_rsp *rsp =
 701                 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, send_cqe);
 702         struct nvmet_rdma_queue *queue = cq->cq_context;
 703
 704         nvmet_rdma_release_rsp(rsp);
 705
 706         if (unlikely(wc->status != IB_WC_SUCCESS &&
 707                      wc->status != IB_WC_WR_FLUSH_ERR)) {
 708                 pr_err("SEND for CQE 0x%p failed with status %s (%d).\n",
 709                         wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
 710                 nvmet_rdma_error_comp(queue);
 711         }
 712 }
 713
 714 static void nvmet_rdma_queue_response(struct nvmet_req *req)
 715 {
 716         struct nvmet_rdma_rsp *rsp =
 717                 container_of(req, struct nvmet_rdma_rsp, req);
 718         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 719         struct ib_send_wr *first_wr;
 720
 721         if (rsp->flags & NVMET_RDMA_REQ_INVALIDATE_RKEY) {
 722                 rsp->send_wr.opcode = IB_WR_SEND_WITH_INV;
 723                 rsp->send_wr.ex.invalidate_rkey = rsp->invalidate_rkey;
 724         } else {
 725                 rsp->send_wr.opcode = IB_WR_SEND;
 726         }
 727
 728         if (nvmet_rdma_need_data_out(rsp)) {
 729                 if (rsp->req.metadata_len)
 730                         first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
 731                                         cm_id->port_num, &rsp->write_cqe, NULL);
 732                 else
 733                         first_wr = rdma_rw_ctx_wrs(&rsp->rw, cm_id->qp,
 734                                         cm_id->port_num, NULL, &rsp->send_wr);
 735         } else {
 736                 first_wr = &rsp->send_wr;
 737         }
 738
 739         nvmet_rdma_post_recv(rsp->queue->dev, rsp->cmd);
 740
 741         ib_dma_sync_single_for_device(rsp->queue->dev->device,
 742                 rsp->send_sge.addr, rsp->send_sge.length,
 743                 DMA_TO_DEVICE);
 744
 745         if (unlikely(ib_post_send(cm_id->qp, first_wr, NULL))) {
 746                 pr_err("sending cmd response failed\n");
 747                 nvmet_rdma_release_rsp(rsp);
 748         }
 749 }
 750
 751 static void nvmet_rdma_read_data_done(struct ib_cq *cq, struct ib_wc *wc)
 752 {
 753         struct nvmet_rdma_rsp *rsp =
 754                 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, read_cqe);
 755         struct nvmet_rdma_queue *queue = wc->qp->qp_context;
 756         u16 status = 0;
 757
 758         WARN_ON(rsp->n_rdma <= 0);
 759         atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
 760         rsp->n_rdma = 0;
 761
 762         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 763                 nvmet_rdma_rw_ctx_destroy(rsp);
 764                 nvmet_req_uninit(&rsp->req);
 765                 nvmet_rdma_release_rsp(rsp);
 766                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 767                         pr_info("RDMA READ for CQE 0x%p failed with status %s (%d).\n",
 768                                 wc->wr_cqe, ib_wc_status_msg(wc->status), wc->status);
 769                         nvmet_rdma_error_comp(queue);
 770                 }
 771                 return;
 772         }
 773
 774         if (rsp->req.metadata_len)
 775                 status = nvmet_rdma_check_pi_status(rsp->rw.reg->mr);
 776         nvmet_rdma_rw_ctx_destroy(rsp);
 777
 778         if (unlikely(status))
 779                 nvmet_req_complete(&rsp->req, status);
 780         else
 781                 rsp->req.execute(&rsp->req);
 782 }
 783
 784 static void nvmet_rdma_write_data_done(struct ib_cq *cq, struct ib_wc *wc)
 785 {
 786         struct nvmet_rdma_rsp *rsp =
 787                 container_of(wc->wr_cqe, struct nvmet_rdma_rsp, write_cqe);
 788         struct nvmet_rdma_queue *queue = cq->cq_context;
 789         struct rdma_cm_id *cm_id = rsp->queue->cm_id;
 790         u16 status;
 791
 792         if (!IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY))
 793                 return;
 794
 795         WARN_ON(rsp->n_rdma <= 0);
 796         atomic_add(rsp->n_rdma, &queue->sq_wr_avail);
 797         rsp->n_rdma = 0;
 798
 799         if (unlikely(wc->status != IB_WC_SUCCESS)) {
 800                 nvmet_rdma_rw_ctx_destroy(rsp);
 801                 nvmet_req_uninit(&rsp->req);
 802                 nvmet_rdma_release_rsp(rsp);
 803                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 804                         pr_info("RDMA WRITE for CQE 0x%p failed with status %s (%d).\n",
 805                                 wc->wr_cqe, ib_wc_status_msg(wc->status),
 806                                 wc->status);
 807                         nvmet_rdma_error_comp(queue);
 808                 }
 809                 return;
 810         }
 811
 812         /*
 813          * Upon RDMA completion check the signature status
 814          * - if succeeded send good NVMe response
 815          * - if failed send bad NVMe response with appropriate error
 816          */
 817         status = nvmet_rdma_check_pi_status(rsp->rw.reg->mr);
 818         if (unlikely(status))
 819                 rsp->req.cqe->status = cpu_to_le16(status << 1);
 820         nvmet_rdma_rw_ctx_destroy(rsp);
 821
 822         if (unlikely(ib_post_send(cm_id->qp, &rsp->send_wr, NULL))) {
 823                 pr_err("sending cmd response failed\n");
 824                 nvmet_rdma_release_rsp(rsp);
 825         }
 826 }
 827
 828 static void nvmet_rdma_use_inline_sg(struct nvmet_rdma_rsp *rsp, u32 len,
 829                 u64 off)
 830 {
 831         int sg_count = num_pages(len);
 832         struct scatterlist *sg;
 833         int i;
 834
 835         sg = rsp->cmd->inline_sg;
 836         for (i = 0; i < sg_count; i++, sg++) {
 837                 if (i < sg_count - 1)
 838                         sg_unmark_end(sg);
 839                 else
 840                         sg_mark_end(sg);
 841                 sg->offset = off;
 842                 sg->length = min_t(int, len, PAGE_SIZE - off);
 843                 len -= sg->length;
 844                 if (!i)
 845                         off = 0;
 846         }
 847
 848         rsp->req.sg = rsp->cmd->inline_sg;
 849         rsp->req.sg_cnt = sg_count;
 850 }
 851
 852 static u16 nvmet_rdma_map_sgl_inline(struct nvmet_rdma_rsp *rsp)
 853 {
 854         struct nvme_sgl_desc *sgl = &rsp->req.cmd->common.dptr.sgl;
 855         u64 off = le64_to_cpu(sgl->addr);
 856         u32 len = le32_to_cpu(sgl->length);
 857
 858         if (!nvme_is_write(rsp->req.cmd)) {
 859                 rsp->req.error_loc =
 860                         offsetof(struct nvme_common_command, opcode);
 861                 return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 862         }
 863
 864         if (off + len > rsp->queue->dev->inline_data_size) {
 865                 pr_err("invalid inline data offset!\n");
 866                 return NVME_SC_SGL_INVALID_OFFSET | NVME_SC_DNR;
 867         }
 868
 869         /* no data command? */
 870         if (!len)
 871                 return 0;
 872
 873         nvmet_rdma_use_inline_sg(rsp, len, off);
 874         rsp->flags |= NVMET_RDMA_REQ_INLINE_DATA;
 875         rsp->req.transfer_len += len;
 876         return 0;
 877 }
 878
 879 static u16 nvmet_rdma_map_sgl_keyed(struct nvmet_rdma_rsp *rsp,
 880                 struct nvme_keyed_sgl_desc *sgl, bool invalidate)
 881 {
 882         u64 addr = le64_to_cpu(sgl->addr);
 883         u32 key = get_unaligned_le32(sgl->key);
 884         struct ib_sig_attrs sig_attrs;
 885         int ret;
 886
 887         rsp->req.transfer_len = get_unaligned_le24(sgl->length);
 888
 889         /* no data command? */
 890         if (!rsp->req.transfer_len)
 891                 return 0;
 892
 893         if (rsp->req.metadata_len)
 894                 nvmet_rdma_set_sig_attrs(&rsp->req, &sig_attrs);
 895
 896         ret = nvmet_req_alloc_sgls(&rsp->req);
 897         if (unlikely(ret < 0))
 898                 goto error_out;
 899
 900         ret = nvmet_rdma_rw_ctx_init(rsp, addr, key, &sig_attrs);
 901         if (unlikely(ret < 0))
 902                 goto error_out;
 903         rsp->n_rdma += ret;
 904
 905         if (invalidate) {
 906                 rsp->invalidate_rkey = key;
 907                 rsp->flags |= NVMET_RDMA_REQ_INVALIDATE_RKEY;
 908         }
 909
 910         return 0;
 911
 912 error_out:
 913         rsp->req.transfer_len = 0;
 914         return NVME_SC_INTERNAL;
 915 }
 916
 917 static u16 nvmet_rdma_map_sgl(struct nvmet_rdma_rsp *rsp)
 918 {
 919         struct nvme_keyed_sgl_desc *sgl = &rsp->req.cmd->common.dptr.ksgl;
 920
 921         switch (sgl->type >> 4) {
 922         case NVME_SGL_FMT_DATA_DESC:
 923                 switch (sgl->type & 0xf) {
 924                 case NVME_SGL_FMT_OFFSET:
 925                         return nvmet_rdma_map_sgl_inline(rsp);
 926                 default:
 927                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 928                         rsp->req.error_loc =
 929                                 offsetof(struct nvme_common_command, dptr);
 930                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 931                 }
 932         case NVME_KEY_SGL_FMT_DATA_DESC:
 933                 switch (sgl->type & 0xf) {
 934                 case NVME_SGL_FMT_ADDRESS | NVME_SGL_FMT_INVALIDATE:
 935                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, true);
 936                 case NVME_SGL_FMT_ADDRESS:
 937                         return nvmet_rdma_map_sgl_keyed(rsp, sgl, false);
 938                 default:
 939                         pr_err("invalid SGL subtype: %#x\n", sgl->type);
 940                         rsp->req.error_loc =
 941                                 offsetof(struct nvme_common_command, dptr);
 942                         return NVME_SC_INVALID_FIELD | NVME_SC_DNR;
 943                 }
 944         default:
 945                 pr_err("invalid SGL type: %#x\n", sgl->type);
 946                 rsp->req.error_loc = offsetof(struct nvme_common_command, dptr);
 947                 return NVME_SC_SGL_INVALID_TYPE | NVME_SC_DNR;
 948         }
 949 }
 950
 951 static bool nvmet_rdma_execute_command(struct nvmet_rdma_rsp *rsp)
 952 {
 953         struct nvmet_rdma_queue *queue = rsp->queue;
 954
 955         if (unlikely(atomic_sub_return(1 + rsp->n_rdma,
 956                         &queue->sq_wr_avail) < 0)) {
 957                 pr_debug("IB send queue full (needed %d): queue %u cntlid %u\n",
 958                                 1 + rsp->n_rdma, queue->idx,
 959                                 queue->nvme_sq.ctrl->cntlid);
 960                 atomic_add(1 + rsp->n_rdma, &queue->sq_wr_avail);
 961                 return false;
 962         }
 963
 964         if (nvmet_rdma_need_data_in(rsp)) {
 965                 if (rdma_rw_ctx_post(&rsp->rw, queue->qp,
 966                                 queue->cm_id->port_num, &rsp->read_cqe, NULL))
 967                         nvmet_req_complete(&rsp->req, NVME_SC_DATA_XFER_ERROR);
 968         } else {
 969                 rsp->req.execute(&rsp->req);
 970         }
 971
 972         return true;
 973 }
 974
 975 static void nvmet_rdma_handle_command(struct nvmet_rdma_queue *queue,
 976                 struct nvmet_rdma_rsp *cmd)
 977 {
 978         u16 status;
 979
 980         ib_dma_sync_single_for_cpu(queue->dev->device,
 981                 cmd->cmd->sge[0].addr, cmd->cmd->sge[0].length,
 982                 DMA_FROM_DEVICE);
 983         ib_dma_sync_single_for_cpu(queue->dev->device,
 984                 cmd->send_sge.addr, cmd->send_sge.length,
 985                 DMA_TO_DEVICE);
 986
 987         if (!nvmet_req_init(&cmd->req, &queue->nvme_cq,
 988                         &queue->nvme_sq, &nvmet_rdma_ops))
 989                 return;
 990
 991         status = nvmet_rdma_map_sgl(cmd);
 992         if (status)
 993                 goto out_err;
 994
 995         if (unlikely(!nvmet_rdma_execute_command(cmd))) {
 996                 spin_lock(&queue->rsp_wr_wait_lock);
 997                 list_add_tail(&cmd->wait_list, &queue->rsp_wr_wait_list);
 998                 spin_unlock(&queue->rsp_wr_wait_lock);
 999         }
1000
1001         return;
1002
1003 out_err:
1004         nvmet_req_complete(&cmd->req, status);
1005 }
1006
1007 static void nvmet_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1008 {
1009         struct nvmet_rdma_cmd *cmd =
1010                 container_of(wc->wr_cqe, struct nvmet_rdma_cmd, cqe);
1011         struct nvmet_rdma_queue *queue = wc->qp->qp_context;
1012         struct nvmet_rdma_rsp *rsp;
1013
1014         if (unlikely(wc->status != IB_WC_SUCCESS)) {
1015                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
1016                         pr_err("RECV for CQE 0x%p failed with status %s (%d)\n",
1017                                 wc->wr_cqe, ib_wc_status_msg(wc->status),
1018                                 wc->status);
1019                         nvmet_rdma_error_comp(queue);
1020                 }
1021                 return;
1022         }
1023
1024         if (unlikely(wc->byte_len < sizeof(struct nvme_command))) {
1025                 pr_err("Ctrl Fatal Error: capsule size less than 64 bytes\n");
1026                 nvmet_rdma_error_comp(queue);
1027                 return;
1028         }
1029
1030         cmd->queue = queue;
1031         rsp = nvmet_rdma_get_rsp(queue);
1032         if (unlikely(!rsp)) {
1033                 /*
1034                  * we get here only under memory pressure,
1035                  * silently drop and have the host retry
1036                  * as we can't even fail it.
1037                  */
1038                 nvmet_rdma_post_recv(queue->dev, cmd);
1039                 return;
1040         }
1041         rsp->queue = queue;
1042         rsp->cmd = cmd;
1043         rsp->flags = 0;
1044         rsp->req.cmd = cmd->nvme_cmd;
1045         rsp->req.port = queue->port;
1046         rsp->n_rdma = 0;
1047
1048         if (unlikely(queue->state != NVMET_RDMA_Q_LIVE)) {
1049                 unsigned long flags;
1050
1051                 spin_lock_irqsave(&queue->state_lock, flags);
1052                 if (queue->state == NVMET_RDMA_Q_CONNECTING)
1053                         list_add_tail(&rsp->wait_list, &queue->rsp_wait_list);
1054                 else
1055                         nvmet_rdma_put_rsp(rsp);
1056                 spin_unlock_irqrestore(&queue->state_lock, flags);
1057                 return;
1058         }
1059
1060         nvmet_rdma_handle_command(queue, rsp);
1061 }
1062
1063 static void nvmet_rdma_destroy_srq(struct nvmet_rdma_srq *nsrq)
1064 {
1065         nvmet_rdma_free_cmds(nsrq->ndev, nsrq->cmds, nsrq->ndev->srq_size,
1066                              false);
1067         ib_destroy_srq(nsrq->srq);
1068
1069         kfree(nsrq);
1070 }
1071
1072 static void nvmet_rdma_destroy_srqs(struct nvmet_rdma_device *ndev)
1073 {
1074         int i;
1075
1076         if (!ndev->srqs)
1077                 return;
1078
1079         for (i = 0; i < ndev->srq_count; i++)
1080                 nvmet_rdma_destroy_srq(ndev->srqs[i]);
1081
1082         kfree(ndev->srqs);
1083 }
1084
1085 static struct nvmet_rdma_srq *
1086 nvmet_rdma_init_srq(struct nvmet_rdma_device *ndev)
1087 {
1088         struct ib_srq_init_attr srq_attr = { NULL, };
1089         size_t srq_size = ndev->srq_size;
1090         struct nvmet_rdma_srq *nsrq;
1091         struct ib_srq *srq;
1092         int ret, i;
1093
1094         nsrq = kzalloc(sizeof(*nsrq), GFP_KERNEL);
1095         if (!nsrq)
1096                 return ERR_PTR(-ENOMEM);
1097
1098         srq_attr.attr.max_wr = srq_size;
1099         srq_attr.attr.max_sge = 1 + ndev->inline_page_count;
1100         srq_attr.attr.srq_limit = 0;
1101         srq_attr.srq_type = IB_SRQT_BASIC;
1102         srq = ib_create_srq(ndev->pd, &srq_attr);
1103         if (IS_ERR(srq)) {
1104                 ret = PTR_ERR(srq);
1105                 goto out_free;
1106         }
1107
1108         nsrq->cmds = nvmet_rdma_alloc_cmds(ndev, srq_size, false);
1109         if (IS_ERR(nsrq->cmds)) {
1110                 ret = PTR_ERR(nsrq->cmds);
1111                 goto out_destroy_srq;
1112         }
1113
1114         nsrq->srq = srq;
1115         nsrq->ndev = ndev;
1116
1117         for (i = 0; i < srq_size; i++) {
1118                 nsrq->cmds[i].nsrq = nsrq;
1119                 ret = nvmet_rdma_post_recv(ndev, &nsrq->cmds[i]);
1120                 if (ret)
1121                         goto out_free_cmds;
1122         }
1123
1124         return nsrq;
1125
1126 out_free_cmds:
1127         nvmet_rdma_free_cmds(ndev, nsrq->cmds, srq_size, false);
1128 out_destroy_srq:
1129         ib_destroy_srq(srq);
1130 out_free:
1131         kfree(nsrq);
1132         return ERR_PTR(ret);
1133 }
1134
1135 static int nvmet_rdma_init_srqs(struct nvmet_rdma_device *ndev)
1136 {
1137         int i, ret;
1138
1139         if (!ndev->device->attrs.max_srq_wr || !ndev->device->attrs.max_srq) {
1140                 /*
1141                  * If SRQs aren't supported we just go ahead and use normal
1142                  * non-shared receive queues.
1143                  */
1144                 pr_info("SRQ requested but not supported.\n");
1145                 return 0;
1146         }
1147
1148         ndev->srq_size = min(ndev->device->attrs.max_srq_wr,
1149                              nvmet_rdma_srq_size);
1150         ndev->srq_count = min(ndev->device->num_comp_vectors,
1151                               ndev->device->attrs.max_srq);
1152
1153         ndev->srqs = kcalloc(ndev->srq_count, sizeof(*ndev->srqs), GFP_KERNEL);
1154         if (!ndev->srqs)
1155                 return -ENOMEM;
1156
1157         for (i = 0; i < ndev->srq_count; i++) {
1158                 ndev->srqs[i] = nvmet_rdma_init_srq(ndev);
1159                 if (IS_ERR(ndev->srqs[i])) {
1160                         ret = PTR_ERR(ndev->srqs[i]);
1161                         goto err_srq;
1162                 }
1163         }
1164
1165         return 0;
1166
1167 err_srq:
1168         while (--i >= 0)
1169                 nvmet_rdma_destroy_srq(ndev->srqs[i]);
1170         kfree(ndev->srqs);
1171         return ret;
1172 }
1173
1174 static void nvmet_rdma_free_dev(struct kref *ref)
1175 {
1176         struct nvmet_rdma_device *ndev =
1177                 container_of(ref, struct nvmet_rdma_device, ref);
1178
1179         mutex_lock(&device_list_mutex);
1180         list_del(&ndev->entry);
1181         mutex_unlock(&device_list_mutex);
1182
1183         nvmet_rdma_destroy_srqs(ndev);
1184         ib_dealloc_pd(ndev->pd);
1185
1186         kfree(ndev);
1187 }
1188
1189 static struct nvmet_rdma_device *
1190 nvmet_rdma_find_get_device(struct rdma_cm_id *cm_id)
1191 {
1192         struct nvmet_rdma_port *port = cm_id->context;
1193         struct nvmet_port *nport = port->nport;
1194         struct nvmet_rdma_device *ndev;
1195         int inline_page_count;
1196         int inline_sge_count;
1197         int ret;
1198
1199         mutex_lock(&device_list_mutex);
1200         list_for_each_entry(ndev, &device_list, entry) {
1201                 if (ndev->device->node_guid == cm_id->device->node_guid &&
1202                     kref_get_unless_zero(&ndev->ref))
1203                         goto out_unlock;
1204         }
1205
1206         ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
1207         if (!ndev)
1208                 goto out_err;
1209
1210         inline_page_count = num_pages(nport->inline_data_size);
1211         inline_sge_count = max(cm_id->device->attrs.max_sge_rd,
1212                                 cm_id->device->attrs.max_recv_sge) - 1;
1213         if (inline_page_count > inline_sge_count) {
1214                 pr_warn("inline_data_size %d cannot be supported by device %s. Reducing to %lu.\n",
1215                         nport->inline_data_size, cm_id->device->name,
1216                         inline_sge_count * PAGE_SIZE);
1217                 nport->inline_data_size = inline_sge_count * PAGE_SIZE;
1218                 inline_page_count = inline_sge_count;
1219         }
1220         ndev->inline_data_size = nport->inline_data_size;
1221         ndev->inline_page_count = inline_page_count;
1222         ndev->device = cm_id->device;
1223         kref_init(&ndev->ref);
1224
1225         ndev->pd = ib_alloc_pd(ndev->device, 0);
1226         if (IS_ERR(ndev->pd))
1227                 goto out_free_dev;
1228
1229         if (nvmet_rdma_use_srq) {
1230                 ret = nvmet_rdma_init_srqs(ndev);
1231                 if (ret)
1232                         goto out_free_pd;
1233         }
1234
1235         list_add(&ndev->entry, &device_list);
1236 out_unlock:
1237         mutex_unlock(&device_list_mutex);
1238         pr_debug("added %s.\n", ndev->device->name);
1239         return ndev;
1240
1241 out_free_pd:
1242         ib_dealloc_pd(ndev->pd);
1243 out_free_dev:
1244         kfree(ndev);
1245 out_err:
1246         mutex_unlock(&device_list_mutex);
1247         return NULL;
1248 }
1249
1250 static int nvmet_rdma_create_queue_ib(struct nvmet_rdma_queue *queue)
1251 {
1252         struct ib_qp_init_attr qp_attr;
1253         struct nvmet_rdma_device *ndev = queue->dev;
1254         int nr_cqe, ret, i, factor;
1255
1256         /*
1257          * Reserve CQ slots for RECV + RDMA_READ/RDMA_WRITE + RDMA_SEND.
1258          */
1259         nr_cqe = queue->recv_queue_size + 2 * queue->send_queue_size;
1260
1261         queue->cq = ib_cq_pool_get(ndev->device, nr_cqe + 1,
1262                                    queue->comp_vector, IB_POLL_WORKQUEUE);
1263         if (IS_ERR(queue->cq)) {
1264                 ret = PTR_ERR(queue->cq);
1265                 pr_err("failed to create CQ cqe= %d ret= %d\n",
1266                        nr_cqe + 1, ret);
1267                 goto out;
1268         }
1269
1270         memset(&qp_attr, 0, sizeof(qp_attr));
1271         qp_attr.qp_context = queue;
1272         qp_attr.event_handler = nvmet_rdma_qp_event;
1273         qp_attr.send_cq = queue->cq;
1274         qp_attr.recv_cq = queue->cq;
1275         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1276         qp_attr.qp_type = IB_QPT_RC;
1277         /* +1 for drain */
1278         qp_attr.cap.max_send_wr = queue->send_queue_size + 1;
1279         factor = rdma_rw_mr_factor(ndev->device, queue->cm_id->port_num,
1280                                    1 << NVMET_RDMA_MAX_MDTS);
1281         qp_attr.cap.max_rdma_ctxs = queue->send_queue_size * factor;
1282         qp_attr.cap.max_send_sge = max(ndev->device->attrs.max_sge_rd,
1283                                         ndev->device->attrs.max_send_sge);
1284
1285         if (queue->nsrq) {
1286                 qp_attr.srq = queue->nsrq->srq;
1287         } else {
1288                 /* +1 for drain */
1289                 qp_attr.cap.max_recv_wr = 1 + queue->recv_queue_size;
1290                 qp_attr.cap.max_recv_sge = 1 + ndev->inline_page_count;
1291         }
1292
1293         if (queue->port->pi_enable && queue->host_qid)
1294                 qp_attr.create_flags |= IB_QP_CREATE_INTEGRITY_EN;
1295
1296         ret = rdma_create_qp(queue->cm_id, ndev->pd, &qp_attr);
1297         if (ret) {
1298                 pr_err("failed to create_qp ret= %d\n", ret);
1299                 goto err_destroy_cq;
1300         }
1301         queue->qp = queue->cm_id->qp;
1302
1303         atomic_set(&queue->sq_wr_avail, qp_attr.cap.max_send_wr);
1304
1305         pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
1306                  __func__, queue->cq->cqe, qp_attr.cap.max_send_sge,
1307                  qp_attr.cap.max_send_wr, queue->cm_id);
1308
1309         if (!queue->nsrq) {
1310                 for (i = 0; i < queue->recv_queue_size; i++) {
1311                         queue->cmds[i].queue = queue;
1312                         ret = nvmet_rdma_post_recv(ndev, &queue->cmds[i]);
1313                         if (ret)
1314                                 goto err_destroy_qp;
1315                 }
1316         }
1317
1318 out:
1319         return ret;
1320
1321 err_destroy_qp:
1322         rdma_destroy_qp(queue->cm_id);
1323 err_destroy_cq:
1324         ib_cq_pool_put(queue->cq, nr_cqe + 1);
1325         goto out;
1326 }
1327
1328 static void nvmet_rdma_destroy_queue_ib(struct nvmet_rdma_queue *queue)
1329 {
1330         ib_drain_qp(queue->qp);
1331         if (queue->cm_id)
1332                 rdma_destroy_id(queue->cm_id);
1333         ib_destroy_qp(queue->qp);
1334         ib_cq_pool_put(queue->cq, queue->recv_queue_size + 2 *
1335                        queue->send_queue_size + 1);
1336 }
1337
1338 static void nvmet_rdma_free_queue(struct nvmet_rdma_queue *queue)
1339 {
1340         pr_debug("freeing queue %d\n", queue->idx);
1341
1342         nvmet_sq_destroy(&queue->nvme_sq);
1343
1344         nvmet_rdma_destroy_queue_ib(queue);
1345         if (!queue->nsrq) {
1346                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1347                                 queue->recv_queue_size,
1348                                 !queue->host_qid);
1349         }
1350         nvmet_rdma_free_rsps(queue);
1351         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1352         kfree(queue);
1353 }
1354
1355 static void nvmet_rdma_release_queue_work(struct work_struct *w)
1356 {
1357         struct nvmet_rdma_queue *queue =
1358                 container_of(w, struct nvmet_rdma_queue, release_work);
1359         struct nvmet_rdma_device *dev = queue->dev;
1360
1361         nvmet_rdma_free_queue(queue);
1362
1363         kref_put(&dev->ref, nvmet_rdma_free_dev);
1364 }
1365
1366 static int
1367 nvmet_rdma_parse_cm_connect_req(struct rdma_conn_param *conn,
1368                                 struct nvmet_rdma_queue *queue)
1369 {
1370         struct nvme_rdma_cm_req *req;
1371
1372         req = (struct nvme_rdma_cm_req *)conn->private_data;
1373         if (!req || conn->private_data_len == 0)
1374                 return NVME_RDMA_CM_INVALID_LEN;
1375
1376         if (le16_to_cpu(req->recfmt) != NVME_RDMA_CM_FMT_1_0)
1377                 return NVME_RDMA_CM_INVALID_RECFMT;
1378
1379         queue->host_qid = le16_to_cpu(req->qid);
1380
1381         /*
1382          * req->hsqsize corresponds to our recv queue size plus 1
1383          * req->hrqsize corresponds to our send queue size
1384          */
1385         queue->recv_queue_size = le16_to_cpu(req->hsqsize) + 1;
1386         queue->send_queue_size = le16_to_cpu(req->hrqsize);
1387
1388         if (!queue->host_qid && queue->recv_queue_size > NVME_AQ_DEPTH)
1389                 return NVME_RDMA_CM_INVALID_HSQSIZE;
1390
1391         /* XXX: Should we enforce some kind of max for IO queues? */
1392
1393         return 0;
1394 }
1395
1396 static int nvmet_rdma_cm_reject(struct rdma_cm_id *cm_id,
1397                                 enum nvme_rdma_cm_status status)
1398 {
1399         struct nvme_rdma_cm_rej rej;
1400
1401         pr_debug("rejecting connect request: status %d (%s)\n",
1402                  status, nvme_rdma_cm_msg(status));
1403
1404         rej.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1405         rej.sts = cpu_to_le16(status);
1406
1407         return rdma_reject(cm_id, (void *)&rej, sizeof(rej),
1408                            IB_CM_REJ_CONSUMER_DEFINED);
1409 }
1410
1411 static struct nvmet_rdma_queue *
1412 nvmet_rdma_alloc_queue(struct nvmet_rdma_device *ndev,
1413                 struct rdma_cm_id *cm_id,
1414                 struct rdma_cm_event *event)
1415 {
1416         struct nvmet_rdma_port *port = cm_id->context;
1417         struct nvmet_rdma_queue *queue;
1418         int ret;
1419
1420         queue = kzalloc(sizeof(*queue), GFP_KERNEL);
1421         if (!queue) {
1422                 ret = NVME_RDMA_CM_NO_RSC;
1423                 goto out_reject;
1424         }
1425
1426         ret = nvmet_sq_init(&queue->nvme_sq);
1427         if (ret) {
1428                 ret = NVME_RDMA_CM_NO_RSC;
1429                 goto out_free_queue;
1430         }
1431
1432         ret = nvmet_rdma_parse_cm_connect_req(&event->param.conn, queue);
1433         if (ret)
1434                 goto out_destroy_sq;
1435
1436         /*
1437          * Schedules the actual release because calling rdma_destroy_id from
1438          * inside a CM callback would trigger a deadlock. (great API design..)
1439          */
1440         INIT_WORK(&queue->release_work, nvmet_rdma_release_queue_work);
1441         queue->dev = ndev;
1442         queue->cm_id = cm_id;
1443         queue->port = port->nport;
1444
1445         spin_lock_init(&queue->state_lock);
1446         queue->state = NVMET_RDMA_Q_CONNECTING;
1447         INIT_LIST_HEAD(&queue->rsp_wait_list);
1448         INIT_LIST_HEAD(&queue->rsp_wr_wait_list);
1449         spin_lock_init(&queue->rsp_wr_wait_lock);
1450         INIT_LIST_HEAD(&queue->free_rsps);
1451         spin_lock_init(&queue->rsps_lock);
1452         INIT_LIST_HEAD(&queue->queue_list);
1453
1454         queue->idx = ida_simple_get(&nvmet_rdma_queue_ida, 0, 0, GFP_KERNEL);
1455         if (queue->idx < 0) {
1456                 ret = NVME_RDMA_CM_NO_RSC;
1457                 goto out_destroy_sq;
1458         }
1459
1460         /*
1461          * Spread the io queues across completion vectors,
1462          * but still keep all admin queues on vector 0.
1463          */
1464         queue->comp_vector = !queue->host_qid ? 0 :
1465                 queue->idx % ndev->device->num_comp_vectors;
1466
1467
1468         ret = nvmet_rdma_alloc_rsps(queue);
1469         if (ret) {
1470                 ret = NVME_RDMA_CM_NO_RSC;
1471                 goto out_ida_remove;
1472         }
1473
1474         if (ndev->srqs) {
1475                 queue->nsrq = ndev->srqs[queue->comp_vector % ndev->srq_count];
1476         } else {
1477                 queue->cmds = nvmet_rdma_alloc_cmds(ndev,
1478                                 queue->recv_queue_size,
1479                                 !queue->host_qid);
1480                 if (IS_ERR(queue->cmds)) {
1481                         ret = NVME_RDMA_CM_NO_RSC;
1482                         goto out_free_responses;
1483                 }
1484         }
1485
1486         ret = nvmet_rdma_create_queue_ib(queue);
1487         if (ret) {
1488                 pr_err("%s: creating RDMA queue failed (%d).\n",
1489                         __func__, ret);
1490                 ret = NVME_RDMA_CM_NO_RSC;
1491                 goto out_free_cmds;
1492         }
1493
1494         return queue;
1495
1496 out_free_cmds:
1497         if (!queue->nsrq) {
1498                 nvmet_rdma_free_cmds(queue->dev, queue->cmds,
1499                                 queue->recv_queue_size,
1500                                 !queue->host_qid);
1501         }
1502 out_free_responses:
1503         nvmet_rdma_free_rsps(queue);
1504 out_ida_remove:
1505         ida_simple_remove(&nvmet_rdma_queue_ida, queue->idx);
1506 out_destroy_sq:
1507         nvmet_sq_destroy(&queue->nvme_sq);
1508 out_free_queue:
1509         kfree(queue);
1510 out_reject:
1511         nvmet_rdma_cm_reject(cm_id, ret);
1512         return NULL;
1513 }
1514
1515 static void nvmet_rdma_qp_event(struct ib_event *event, void *priv)
1516 {
1517         struct nvmet_rdma_queue *queue = priv;
1518
1519         switch (event->event) {
1520         case IB_EVENT_COMM_EST:
1521                 rdma_notify(queue->cm_id, event->event);
1522                 break;
1523         case IB_EVENT_QP_LAST_WQE_REACHED:
1524                 pr_debug("received last WQE reached event for queue=0x%p\n",
1525                          queue);
1526                 break;
1527         default:
1528                 pr_err("received IB QP event: %s (%d)\n",
1529                        ib_event_msg(event->event), event->event);
1530                 break;
1531         }
1532 }
1533
1534 static int nvmet_rdma_cm_accept(struct rdma_cm_id *cm_id,
1535                 struct nvmet_rdma_queue *queue,
1536                 struct rdma_conn_param *p)
1537 {
1538         struct rdma_conn_param  param = { };
1539         struct nvme_rdma_cm_rep priv = { };
1540         int ret = -ENOMEM;
1541
1542         param.rnr_retry_count = 7;
1543         param.flow_control = 1;
1544         param.initiator_depth = min_t(u8, p->initiator_depth,
1545                 queue->dev->device->attrs.max_qp_init_rd_atom);
1546         param.private_data = &priv;
1547         param.private_data_len = sizeof(priv);
1548         priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1549         priv.crqsize = cpu_to_le16(queue->recv_queue_size);
1550
1551         ret = rdma_accept(cm_id, &param);
1552         if (ret)
1553                 pr_err("rdma_accept failed (error code = %d)\n", ret);
1554
1555         return ret;
1556 }
1557
1558 static int nvmet_rdma_queue_connect(struct rdma_cm_id *cm_id,
1559                 struct rdma_cm_event *event)
1560 {
1561         struct nvmet_rdma_device *ndev;
1562         struct nvmet_rdma_queue *queue;
1563         int ret = -EINVAL;
1564
1565         ndev = nvmet_rdma_find_get_device(cm_id);
1566         if (!ndev) {
1567                 nvmet_rdma_cm_reject(cm_id, NVME_RDMA_CM_NO_RSC);
1568                 return -ECONNREFUSED;
1569         }
1570
1571         queue = nvmet_rdma_alloc_queue(ndev, cm_id, event);
1572         if (!queue) {
1573                 ret = -ENOMEM;
1574                 goto put_device;
1575         }
1576
1577         if (queue->host_qid == 0) {
1578                 /* Let inflight controller teardown complete */
1579                 flush_scheduled_work();
1580         }
1581
1582         ret = nvmet_rdma_cm_accept(cm_id, queue, &event->param.conn);
1583         if (ret) {
1584                 /*
1585                  * Don't destroy the cm_id in free path, as we implicitly
1586                  * destroy the cm_id here with non-zero ret code.
1587                  */
1588                 queue->cm_id = NULL;
1589                 goto free_queue;
1590         }
1591
1592         mutex_lock(&nvmet_rdma_queue_mutex);
1593         list_add_tail(&queue->queue_list, &nvmet_rdma_queue_list);
1594         mutex_unlock(&nvmet_rdma_queue_mutex);
1595
1596         return 0;
1597
1598 free_queue:
1599         nvmet_rdma_free_queue(queue);
1600 put_device:
1601         kref_put(&ndev->ref, nvmet_rdma_free_dev);
1602
1603         return ret;
1604 }
1605
1606 static void nvmet_rdma_queue_established(struct nvmet_rdma_queue *queue)
1607 {
1608         unsigned long flags;
1609
1610         spin_lock_irqsave(&queue->state_lock, flags);
1611         if (queue->state != NVMET_RDMA_Q_CONNECTING) {
1612                 pr_warn("trying to establish a connected queue\n");
1613                 goto out_unlock;
1614         }
1615         queue->state = NVMET_RDMA_Q_LIVE;
1616
1617         while (!list_empty(&queue->rsp_wait_list)) {
1618                 struct nvmet_rdma_rsp *cmd;
1619
1620                 cmd = list_first_entry(&queue->rsp_wait_list,
1621                                         struct nvmet_rdma_rsp, wait_list);
1622                 list_del(&cmd->wait_list);
1623
1624                 spin_unlock_irqrestore(&queue->state_lock, flags);
1625                 nvmet_rdma_handle_command(queue, cmd);
1626                 spin_lock_irqsave(&queue->state_lock, flags);
1627         }
1628
1629 out_unlock:
1630         spin_unlock_irqrestore(&queue->state_lock, flags);
1631 }
1632
1633 static void __nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1634 {
1635         bool disconnect = false;
1636         unsigned long flags;
1637
1638         pr_debug("cm_id= %p queue->state= %d\n", queue->cm_id, queue->state);
1639
1640         spin_lock_irqsave(&queue->state_lock, flags);
1641         switch (queue->state) {
1642         case NVMET_RDMA_Q_CONNECTING:
1643         case NVMET_RDMA_Q_LIVE:
1644                 queue->state = NVMET_RDMA_Q_DISCONNECTING;
1645                 disconnect = true;
1646                 break;
1647         case NVMET_RDMA_Q_DISCONNECTING:
1648                 break;
1649         }
1650         spin_unlock_irqrestore(&queue->state_lock, flags);
1651
1652         if (disconnect) {
1653                 rdma_disconnect(queue->cm_id);
1654                 schedule_work(&queue->release_work);
1655         }
1656 }
1657
1658 static void nvmet_rdma_queue_disconnect(struct nvmet_rdma_queue *queue)
1659 {
1660         bool disconnect = false;
1661
1662         mutex_lock(&nvmet_rdma_queue_mutex);
1663         if (!list_empty(&queue->queue_list)) {
1664                 list_del_init(&queue->queue_list);
1665                 disconnect = true;
1666         }
1667         mutex_unlock(&nvmet_rdma_queue_mutex);
1668
1669         if (disconnect)
1670                 __nvmet_rdma_queue_disconnect(queue);
1671 }
1672
1673 static void nvmet_rdma_queue_connect_fail(struct rdma_cm_id *cm_id,
1674                 struct nvmet_rdma_queue *queue)
1675 {
1676         WARN_ON_ONCE(queue->state != NVMET_RDMA_Q_CONNECTING);
1677
1678         mutex_lock(&nvmet_rdma_queue_mutex);
1679         if (!list_empty(&queue->queue_list))
1680                 list_del_init(&queue->queue_list);
1681         mutex_unlock(&nvmet_rdma_queue_mutex);
1682
1683         pr_err("failed to connect queue %d\n", queue->idx);
1684         schedule_work(&queue->release_work);
1685 }
1686
1687 /**
1688  * nvme_rdma_device_removal() - Handle RDMA device removal
1689  * @cm_id:      rdma_cm id, used for nvmet port
1690  * @queue:      nvmet rdma queue (cm id qp_context)
1691  *
1692  * DEVICE_REMOVAL event notifies us that the RDMA device is about
1693  * to unplug. Note that this event can be generated on a normal
1694  * queue cm_id and/or a device bound listener cm_id (where in this
1695  * case queue will be null).
1696  *
1697  * We registered an ib_client to handle device removal for queues,
1698  * so we only need to handle the listening port cm_ids. In this case
1699  * we nullify the priv to prevent double cm_id destruction and destroying
1700  * the cm_id implicitely by returning a non-zero rc to the callout.
1701  */
1702 static int nvmet_rdma_device_removal(struct rdma_cm_id *cm_id,
1703                 struct nvmet_rdma_queue *queue)
1704 {
1705         struct nvmet_rdma_port *port;
1706
1707         if (queue) {
1708                 /*
1709                  * This is a queue cm_id. we have registered
1710                  * an ib_client to handle queues removal
1711                  * so don't interfear and just return.
1712                  */
1713                 return 0;
1714         }
1715
1716         port = cm_id->context;
1717
1718         /*
1719          * This is a listener cm_id. Make sure that
1720          * future remove_port won't invoke a double
1721          * cm_id destroy. use atomic xchg to make sure
1722          * we don't compete with remove_port.
1723          */
1724         if (xchg(&port->cm_id, NULL) != cm_id)
1725                 return 0;
1726
1727         /*
1728          * We need to return 1 so that the core will destroy
1729          * it's own ID.  What a great API design..
1730          */
1731         return 1;
1732 }
1733
1734 static int nvmet_rdma_cm_handler(struct rdma_cm_id *cm_id,
1735                 struct rdma_cm_event *event)
1736 {
1737         struct nvmet_rdma_queue *queue = NULL;
1738         int ret = 0;
1739
1740         if (cm_id->qp)
1741                 queue = cm_id->qp->qp_context;
1742
1743         pr_debug("%s (%d): status %d id %p\n",
1744                 rdma_event_msg(event->event), event->event,
1745                 event->status, cm_id);
1746
1747         switch (event->event) {
1748         case RDMA_CM_EVENT_CONNECT_REQUEST:
1749                 ret = nvmet_rdma_queue_connect(cm_id, event);
1750                 break;
1751         case RDMA_CM_EVENT_ESTABLISHED:
1752                 nvmet_rdma_queue_established(queue);
1753                 break;
1754         case RDMA_CM_EVENT_ADDR_CHANGE:
1755                 if (!queue) {
1756                         struct nvmet_rdma_port *port = cm_id->context;
1757
1758                         schedule_delayed_work(&port->repair_work, 0);
1759                         break;
1760                 }
1761                 fallthrough;
1762         case RDMA_CM_EVENT_DISCONNECTED:
1763         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1764                 nvmet_rdma_queue_disconnect(queue);
1765                 break;
1766         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1767                 ret = nvmet_rdma_device_removal(cm_id, queue);
1768                 break;
1769         case RDMA_CM_EVENT_REJECTED:
1770                 pr_debug("Connection rejected: %s\n",
1771                          rdma_reject_msg(cm_id, event->status));
1772                 fallthrough;
1773         case RDMA_CM_EVENT_UNREACHABLE:
1774         case RDMA_CM_EVENT_CONNECT_ERROR:
1775                 nvmet_rdma_queue_connect_fail(cm_id, queue);
1776                 break;
1777         default:
1778                 pr_err("received unrecognized RDMA CM event %d\n",
1779                         event->event);
1780                 break;
1781         }
1782
1783         return ret;
1784 }
1785
1786 static void nvmet_rdma_delete_ctrl(struct nvmet_ctrl *ctrl)
1787 {
1788         struct nvmet_rdma_queue *queue;
1789
1790 restart:
1791         mutex_lock(&nvmet_rdma_queue_mutex);
1792         list_for_each_entry(queue, &nvmet_rdma_queue_list, queue_list) {
1793                 if (queue->nvme_sq.ctrl == ctrl) {
1794                         list_del_init(&queue->queue_list);
1795                         mutex_unlock(&nvmet_rdma_queue_mutex);
1796
1797                         __nvmet_rdma_queue_disconnect(queue);
1798                         goto restart;
1799                 }
1800         }
1801         mutex_unlock(&nvmet_rdma_queue_mutex);
1802 }
1803
1804 static void nvmet_rdma_disable_port(struct nvmet_rdma_port *port)
1805 {
1806         struct rdma_cm_id *cm_id = xchg(&port->cm_id, NULL);
1807
1808         if (cm_id)
1809                 rdma_destroy_id(cm_id);
1810 }
1811
1812 static int nvmet_rdma_enable_port(struct nvmet_rdma_port *port)
1813 {
1814         struct sockaddr *addr = (struct sockaddr *)&port->addr;
1815         struct rdma_cm_id *cm_id;
1816         int ret;
1817
1818         cm_id = rdma_create_id(&init_net, nvmet_rdma_cm_handler, port,
1819                         RDMA_PS_TCP, IB_QPT_RC);
1820         if (IS_ERR(cm_id)) {
1821                 pr_err("CM ID creation failed\n");
1822                 return PTR_ERR(cm_id);
1823         }
1824
1825         /*
1826          * Allow both IPv4 and IPv6 sockets to bind a single port
1827          * at the same time.
1828          */
1829         ret = rdma_set_afonly(cm_id, 1);
1830         if (ret) {
1831                 pr_err("rdma_set_afonly failed (%d)\n", ret);
1832                 goto out_destroy_id;
1833         }
1834
1835         ret = rdma_bind_addr(cm_id, addr);
1836         if (ret) {
1837                 pr_err("binding CM ID to %pISpcs failed (%d)\n", addr, ret);
1838                 goto out_destroy_id;
1839         }
1840
1841         ret = rdma_listen(cm_id, 128);
1842         if (ret) {
1843                 pr_err("listening to %pISpcs failed (%d)\n", addr, ret);
1844                 goto out_destroy_id;
1845         }
1846
1847         if (port->nport->pi_enable &&
1848             !(cm_id->device->attrs.device_cap_flags &
1849               IB_DEVICE_INTEGRITY_HANDOVER)) {
1850                 pr_err("T10-PI is not supported for %pISpcs\n", addr);
1851                 ret = -EINVAL;
1852                 goto out_destroy_id;
1853         }
1854
1855         port->cm_id = cm_id;
1856         return 0;
1857
1858 out_destroy_id:
1859         rdma_destroy_id(cm_id);
1860         return ret;
1861 }
1862
1863 static void nvmet_rdma_repair_port_work(struct work_struct *w)
1864 {
1865         struct nvmet_rdma_port *port = container_of(to_delayed_work(w),
1866                         struct nvmet_rdma_port, repair_work);
1867         int ret;
1868
1869         nvmet_rdma_disable_port(port);
1870         ret = nvmet_rdma_enable_port(port);
1871         if (ret)
1872                 schedule_delayed_work(&port->repair_work, 5 * HZ);
1873 }
1874
1875 static int nvmet_rdma_add_port(struct nvmet_port *nport)
1876 {
1877         struct nvmet_rdma_port *port;
1878         __kernel_sa_family_t af;
1879         int ret;
1880
1881         port = kzalloc(sizeof(*port), GFP_KERNEL);
1882         if (!port)
1883                 return -ENOMEM;
1884
1885         nport->priv = port;
1886         port->nport = nport;
1887         INIT_DELAYED_WORK(&port->repair_work, nvmet_rdma_repair_port_work);
1888
1889         switch (nport->disc_addr.adrfam) {
1890         case NVMF_ADDR_FAMILY_IP4:
1891                 af = AF_INET;
1892                 break;
1893         case NVMF_ADDR_FAMILY_IP6:
1894                 af = AF_INET6;
1895                 break;
1896         default:
1897                 pr_err("address family %d not supported\n",
1898                         nport->disc_addr.adrfam);
1899                 ret = -EINVAL;
1900                 goto out_free_port;
1901         }
1902
1903         if (nport->inline_data_size < 0) {
1904                 nport->inline_data_size = NVMET_RDMA_DEFAULT_INLINE_DATA_SIZE;
1905         } else if (nport->inline_data_size > NVMET_RDMA_MAX_INLINE_DATA_SIZE) {
1906                 pr_warn("inline_data_size %u is too large, reducing to %u\n",
1907                         nport->inline_data_size,
1908                         NVMET_RDMA_MAX_INLINE_DATA_SIZE);
1909                 nport->inline_data_size = NVMET_RDMA_MAX_INLINE_DATA_SIZE;
1910         }
1911
1912         ret = inet_pton_with_scope(&init_net, af, nport->disc_addr.traddr,
1913                         nport->disc_addr.trsvcid, &port->addr);
1914         if (ret) {
1915                 pr_err("malformed ip/port passed: %s:%s\n",
1916                         nport->disc_addr.traddr, nport->disc_addr.trsvcid);
1917                 goto out_free_port;
1918         }
1919
1920         ret = nvmet_rdma_enable_port(port);
1921         if (ret)
1922                 goto out_free_port;
1923
1924         pr_info("enabling port %d (%pISpcs)\n",
1925                 le16_to_cpu(nport->disc_addr.portid),
1926                 (struct sockaddr *)&port->addr);
1927
1928         return 0;
1929
1930 out_free_port:
1931         kfree(port);
1932         return ret;
1933 }
1934
1935 static void nvmet_rdma_remove_port(struct nvmet_port *nport)
1936 {
1937         struct nvmet_rdma_port *port = nport->priv;
1938
1939         cancel_delayed_work_sync(&port->repair_work);
1940         nvmet_rdma_disable_port(port);
1941         kfree(port);
1942 }
1943
1944 static void nvmet_rdma_disc_port_addr(struct nvmet_req *req,
1945                 struct nvmet_port *nport, char *traddr)
1946 {
1947         struct nvmet_rdma_port *port = nport->priv;
1948         struct rdma_cm_id *cm_id = port->cm_id;
1949
1950         if (inet_addr_is_any((struct sockaddr *)&cm_id->route.addr.src_addr)) {
1951                 struct nvmet_rdma_rsp *rsp =
1952                         container_of(req, struct nvmet_rdma_rsp, req);
1953                 struct rdma_cm_id *req_cm_id = rsp->queue->cm_id;
1954                 struct sockaddr *addr = (void *)&req_cm_id->route.addr.src_addr;
1955
1956                 sprintf(traddr, "%pISc", addr);
1957         } else {
1958                 memcpy(traddr, nport->disc_addr.traddr, NVMF_TRADDR_SIZE);
1959         }
1960 }
1961
1962 static u8 nvmet_rdma_get_mdts(const struct nvmet_ctrl *ctrl)
1963 {
1964         if (ctrl->pi_support)
1965                 return NVMET_RDMA_MAX_METADATA_MDTS;
1966         return NVMET_RDMA_MAX_MDTS;
1967 }
1968
1969 static const struct nvmet_fabrics_ops nvmet_rdma_ops = {
1970         .owner                  = THIS_MODULE,
1971         .type                   = NVMF_TRTYPE_RDMA,
1972         .msdbd                  = 1,
1973         .flags                  = NVMF_KEYED_SGLS | NVMF_METADATA_SUPPORTED,
1974         .add_port               = nvmet_rdma_add_port,
1975         .remove_port            = nvmet_rdma_remove_port,
1976         .queue_response         = nvmet_rdma_queue_response,
1977         .delete_ctrl            = nvmet_rdma_delete_ctrl,
1978         .disc_traddr            = nvmet_rdma_disc_port_addr,
1979         .get_mdts               = nvmet_rdma_get_mdts,
1980 };
1981
1982 static void nvmet_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1983 {
1984         struct nvmet_rdma_queue *queue, *tmp;
1985         struct nvmet_rdma_device *ndev;
1986         bool found = false;
1987
1988         mutex_lock(&device_list_mutex);
1989         list_for_each_entry(ndev, &device_list, entry) {
1990                 if (ndev->device == ib_device) {
1991                         found = true;
1992                         break;
1993                 }
1994         }
1995         mutex_unlock(&device_list_mutex);
1996
1997         if (!found)
1998                 return;
1999
2000         /*
2001          * IB Device that is used by nvmet controllers is being removed,
2002          * delete all queues using this device.
2003          */
2004         mutex_lock(&nvmet_rdma_queue_mutex);
2005         list_for_each_entry_safe(queue, tmp, &nvmet_rdma_queue_list,
2006                                  queue_list) {
2007                 if (queue->dev->device != ib_device)
2008                         continue;
2009
2010                 pr_info("Removing queue %d\n", queue->idx);
2011                 list_del_init(&queue->queue_list);
2012                 __nvmet_rdma_queue_disconnect(queue);
2013         }
2014         mutex_unlock(&nvmet_rdma_queue_mutex);
2015
2016         flush_scheduled_work();
2017 }
2018
2019 static struct ib_client nvmet_rdma_ib_client = {
2020         .name   = "nvmet_rdma",
2021         .remove = nvmet_rdma_remove_one
2022 };
2023
2024 static int __init nvmet_rdma_init(void)
2025 {
2026         int ret;
2027
2028         ret = ib_register_client(&nvmet_rdma_ib_client);
2029         if (ret)
2030                 return ret;
2031
2032         ret = nvmet_register_transport(&nvmet_rdma_ops);
2033         if (ret)
2034                 goto err_ib_client;
2035
2036         return 0;
2037
2038 err_ib_client:
2039         ib_unregister_client(&nvmet_rdma_ib_client);
2040         return ret;
2041 }
2042
2043 static void __exit nvmet_rdma_exit(void)
2044 {
2045         nvmet_unregister_transport(&nvmet_rdma_ops);
2046         ib_unregister_client(&nvmet_rdma_ib_client);
2047         WARN_ON_ONCE(!list_empty(&nvmet_rdma_queue_list));
2048         ida_destroy(&nvmet_rdma_queue_ida);
2049 }
2050
2051 module_init(nvmet_rdma_init);
2052 module_exit(nvmet_rdma_exit);
2053
2054 MODULE_LICENSE("GPL v2");
2055 MODULE_ALIAS("nvmet-transport-1"); /* 1 == NVMF_TRTYPE_RDMA */