fs/ksmbd/transport_rdma.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  *   Copyright (C) 2017, Microsoft Corporation.
   4  *   Copyright (C) 2018, LG Electronics.
   5  *
   6  *   Author(s): Long Li <longli@microsoft.com>,
   7  *              Hyunchul Lee <hyc.lee@gmail.com>
   8  */
   9
  10 #define SUBMOD_NAME     "smb_direct"
  11
  12 #include <linux/kthread.h>
  13 #include <linux/list.h>
  14 #include <linux/mempool.h>
  15 #include <linux/highmem.h>
  16 #include <linux/scatterlist.h>
  17 #include <rdma/ib_verbs.h>
  18 #include <rdma/rdma_cm.h>
  19 #include <rdma/rw.h>
  20
  21 #include "glob.h"
  22 #include "connection.h"
  23 #include "smb_common.h"
  24 #include "smbstatus.h"
  25 #include "transport_rdma.h"
  26
  27 #define SMB_DIRECT_PORT_IWARP           5445
  28 #define SMB_DIRECT_PORT_INFINIBAND      445
  29
  30 #define SMB_DIRECT_VERSION_LE           cpu_to_le16(0x0100)
  31
  32 /* SMB_DIRECT negotiation timeout in seconds */
  33 #define SMB_DIRECT_NEGOTIATE_TIMEOUT            120
  34
  35 #define SMB_DIRECT_MAX_SEND_SGES                8
  36 #define SMB_DIRECT_MAX_RECV_SGES                1
  37
  38 /*
  39  * Default maximum number of RDMA read/write outstanding on this connection
  40  * This value is possibly decreased during QP creation on hardware limit
  41  */
  42 #define SMB_DIRECT_CM_INITIATOR_DEPTH           8
  43
  44 /* Maximum number of retries on data transfer operations */
  45 #define SMB_DIRECT_CM_RETRY                     6
  46 /* No need to retry on Receiver Not Ready since SMB_DIRECT manages credits */
  47 #define SMB_DIRECT_CM_RNR_RETRY         0
  48
  49 /*
  50  * User configurable initial values per SMB_DIRECT transport connection
  51  * as defined in [MS-SMBD] 3.1.1.1
  52  * Those may change after a SMB_DIRECT negotiation
  53  */
  54
  55 /* Set 445 port to SMB Direct port by default */
  56 static int smb_direct_port = SMB_DIRECT_PORT_INFINIBAND;
  57
  58 /* The local peer's maximum number of credits to grant to the peer */
  59 static int smb_direct_receive_credit_max = 255;
  60
  61 /* The remote peer's credit request of local peer */
  62 static int smb_direct_send_credit_target = 255;
  63
  64 /* The maximum single message size can be sent to remote peer */
  65 static int smb_direct_max_send_size = 8192;
  66
  67 /*  The maximum fragmented upper-layer payload receive size supported */
  68 static int smb_direct_max_fragmented_recv_size = 1024 * 1024;
  69
  70 /*  The maximum single-message size which can be received */
  71 static int smb_direct_max_receive_size = 8192;
  72
  73 static int smb_direct_max_read_write_size = SMBD_DEFAULT_IOSIZE;
  74
  75 static LIST_HEAD(smb_direct_device_list);
  76 static DEFINE_RWLOCK(smb_direct_device_lock);
  77
  78 struct smb_direct_device {
  79         struct ib_device        *ib_dev;
  80         struct list_head        list;
  81 };
  82
  83 static struct smb_direct_listener {
  84         struct rdma_cm_id       *cm_id;
  85 } smb_direct_listener;
  86
  87 static struct workqueue_struct *smb_direct_wq;
  88
  89 enum smb_direct_status {
  90         SMB_DIRECT_CS_NEW = 0,
  91         SMB_DIRECT_CS_CONNECTED,
  92         SMB_DIRECT_CS_DISCONNECTING,
  93         SMB_DIRECT_CS_DISCONNECTED,
  94 };
  95
  96 struct smb_direct_transport {
  97         struct ksmbd_transport  transport;
  98
  99         enum smb_direct_status  status;
 100         bool                    full_packet_received;
 101         wait_queue_head_t       wait_status;
 102
 103         struct rdma_cm_id       *cm_id;
 104         struct ib_cq            *send_cq;
 105         struct ib_cq            *recv_cq;
 106         struct ib_pd            *pd;
 107         struct ib_qp            *qp;
 108
 109         int                     max_send_size;
 110         int                     max_recv_size;
 111         int                     max_fragmented_send_size;
 112         int                     max_fragmented_recv_size;
 113         int                     max_rdma_rw_size;
 114
 115         spinlock_t              reassembly_queue_lock;
 116         struct list_head        reassembly_queue;
 117         int                     reassembly_data_length;
 118         int                     reassembly_queue_length;
 119         int                     first_entry_offset;
 120         wait_queue_head_t       wait_reassembly_queue;
 121
 122         spinlock_t              receive_credit_lock;
 123         int                     recv_credits;
 124         int                     count_avail_recvmsg;
 125         int                     recv_credit_max;
 126         int                     recv_credit_target;
 127
 128         spinlock_t              recvmsg_queue_lock;
 129         struct list_head        recvmsg_queue;
 130
 131         spinlock_t              empty_recvmsg_queue_lock;
 132         struct list_head        empty_recvmsg_queue;
 133
 134         int                     send_credit_target;
 135         atomic_t                send_credits;
 136         spinlock_t              lock_new_recv_credits;
 137         int                     new_recv_credits;
 138         int                     max_rw_credits;
 139         int                     pages_per_rw_credit;
 140         atomic_t                rw_credits;
 141
 142         wait_queue_head_t       wait_send_credits;
 143         wait_queue_head_t       wait_rw_credits;
 144
 145         mempool_t               *sendmsg_mempool;
 146         struct kmem_cache       *sendmsg_cache;
 147         mempool_t               *recvmsg_mempool;
 148         struct kmem_cache       *recvmsg_cache;
 149
 150         wait_queue_head_t       wait_send_pending;
 151         atomic_t                send_pending;
 152
 153         struct delayed_work     post_recv_credits_work;
 154         struct work_struct      send_immediate_work;
 155         struct work_struct      disconnect_work;
 156
 157         bool                    negotiation_requested;
 158 };
 159
 160 #define KSMBD_TRANS(t) ((struct ksmbd_transport *)&((t)->transport))
 161
 162 enum {
 163         SMB_DIRECT_MSG_NEGOTIATE_REQ = 0,
 164         SMB_DIRECT_MSG_DATA_TRANSFER
 165 };
 166
 167 static struct ksmbd_transport_ops ksmbd_smb_direct_transport_ops;
 168
 169 struct smb_direct_send_ctx {
 170         struct list_head        msg_list;
 171         int                     wr_cnt;
 172         bool                    need_invalidate_rkey;
 173         unsigned int            remote_key;
 174 };
 175
 176 struct smb_direct_sendmsg {
 177         struct smb_direct_transport     *transport;
 178         struct ib_send_wr       wr;
 179         struct list_head        list;
 180         int                     num_sge;
 181         struct ib_sge           sge[SMB_DIRECT_MAX_SEND_SGES];
 182         struct ib_cqe           cqe;
 183         u8                      packet[];
 184 };
 185
 186 struct smb_direct_recvmsg {
 187         struct smb_direct_transport     *transport;
 188         struct list_head        list;
 189         int                     type;
 190         struct ib_sge           sge;
 191         struct ib_cqe           cqe;
 192         bool                    first_segment;
 193         u8                      packet[];
 194 };
 195
 196 struct smb_direct_rdma_rw_msg {
 197         struct smb_direct_transport     *t;
 198         struct ib_cqe           cqe;
 199         int                     status;
 200         struct completion       *completion;
 201         struct list_head        list;
 202         struct rdma_rw_ctx      rw_ctx;
 203         struct sg_table         sgt;
 204         struct scatterlist      sg_list[];
 205 };
 206
 207 void init_smbd_max_io_size(unsigned int sz)
 208 {
 209         sz = clamp_val(sz, SMBD_MIN_IOSIZE, SMBD_MAX_IOSIZE);
 210         smb_direct_max_read_write_size = sz;
 211 }
 212
 213 unsigned int get_smbd_max_read_write_size(void)
 214 {
 215         return smb_direct_max_read_write_size;
 216 }
 217
 218 static inline int get_buf_page_count(void *buf, int size)
 219 {
 220         return DIV_ROUND_UP((uintptr_t)buf + size, PAGE_SIZE) -
 221                 (uintptr_t)buf / PAGE_SIZE;
 222 }
 223
 224 static void smb_direct_destroy_pools(struct smb_direct_transport *transport);
 225 static void smb_direct_post_recv_credits(struct work_struct *work);
 226 static int smb_direct_post_send_data(struct smb_direct_transport *t,
 227                                      struct smb_direct_send_ctx *send_ctx,
 228                                      struct kvec *iov, int niov,
 229                                      int remaining_data_length);
 230
 231 static inline struct smb_direct_transport *
 232 smb_trans_direct_transfort(struct ksmbd_transport *t)
 233 {
 234         return container_of(t, struct smb_direct_transport, transport);
 235 }
 236
 237 static inline void
 238 *smb_direct_recvmsg_payload(struct smb_direct_recvmsg *recvmsg)
 239 {
 240         return (void *)recvmsg->packet;
 241 }
 242
 243 static inline bool is_receive_credit_post_required(int receive_credits,
 244                                                    int avail_recvmsg_count)
 245 {
 246         return receive_credits <= (smb_direct_receive_credit_max >> 3) &&
 247                 avail_recvmsg_count >= (receive_credits >> 2);
 248 }
 249
 250 static struct
 251 smb_direct_recvmsg *get_free_recvmsg(struct smb_direct_transport *t)
 252 {
 253         struct smb_direct_recvmsg *recvmsg = NULL;
 254
 255         spin_lock(&t->recvmsg_queue_lock);
 256         if (!list_empty(&t->recvmsg_queue)) {
 257                 recvmsg = list_first_entry(&t->recvmsg_queue,
 258                                            struct smb_direct_recvmsg,
 259                                            list);
 260                 list_del(&recvmsg->list);
 261         }
 262         spin_unlock(&t->recvmsg_queue_lock);
 263         return recvmsg;
 264 }
 265
 266 static void put_recvmsg(struct smb_direct_transport *t,
 267                         struct smb_direct_recvmsg *recvmsg)
 268 {
 269         ib_dma_unmap_single(t->cm_id->device, recvmsg->sge.addr,
 270                             recvmsg->sge.length, DMA_FROM_DEVICE);
 271
 272         spin_lock(&t->recvmsg_queue_lock);
 273         list_add(&recvmsg->list, &t->recvmsg_queue);
 274         spin_unlock(&t->recvmsg_queue_lock);
 275 }
 276
 277 static struct
 278 smb_direct_recvmsg *get_empty_recvmsg(struct smb_direct_transport *t)
 279 {
 280         struct smb_direct_recvmsg *recvmsg = NULL;
 281
 282         spin_lock(&t->empty_recvmsg_queue_lock);
 283         if (!list_empty(&t->empty_recvmsg_queue)) {
 284                 recvmsg = list_first_entry(&t->empty_recvmsg_queue,
 285                                            struct smb_direct_recvmsg, list);
 286                 list_del(&recvmsg->list);
 287         }
 288         spin_unlock(&t->empty_recvmsg_queue_lock);
 289         return recvmsg;
 290 }
 291
 292 static void put_empty_recvmsg(struct smb_direct_transport *t,
 293                               struct smb_direct_recvmsg *recvmsg)
 294 {
 295         ib_dma_unmap_single(t->cm_id->device, recvmsg->sge.addr,
 296                             recvmsg->sge.length, DMA_FROM_DEVICE);
 297
 298         spin_lock(&t->empty_recvmsg_queue_lock);
 299         list_add_tail(&recvmsg->list, &t->empty_recvmsg_queue);
 300         spin_unlock(&t->empty_recvmsg_queue_lock);
 301 }
 302
 303 static void enqueue_reassembly(struct smb_direct_transport *t,
 304                                struct smb_direct_recvmsg *recvmsg,
 305                                int data_length)
 306 {
 307         spin_lock(&t->reassembly_queue_lock);
 308         list_add_tail(&recvmsg->list, &t->reassembly_queue);
 309         t->reassembly_queue_length++;
 310         /*
 311          * Make sure reassembly_data_length is updated after list and
 312          * reassembly_queue_length are updated. On the dequeue side
 313          * reassembly_data_length is checked without a lock to determine
 314          * if reassembly_queue_length and list is up to date
 315          */
 316         virt_wmb();
 317         t->reassembly_data_length += data_length;
 318         spin_unlock(&t->reassembly_queue_lock);
 319 }
 320
 321 static struct smb_direct_recvmsg *get_first_reassembly(struct smb_direct_transport *t)
 322 {
 323         if (!list_empty(&t->reassembly_queue))
 324                 return list_first_entry(&t->reassembly_queue,
 325                                 struct smb_direct_recvmsg, list);
 326         else
 327                 return NULL;
 328 }
 329
 330 static void smb_direct_disconnect_rdma_work(struct work_struct *work)
 331 {
 332         struct smb_direct_transport *t =
 333                 container_of(work, struct smb_direct_transport,
 334                              disconnect_work);
 335
 336         if (t->status == SMB_DIRECT_CS_CONNECTED) {
 337                 t->status = SMB_DIRECT_CS_DISCONNECTING;
 338                 rdma_disconnect(t->cm_id);
 339         }
 340 }
 341
 342 static void
 343 smb_direct_disconnect_rdma_connection(struct smb_direct_transport *t)
 344 {
 345         if (t->status == SMB_DIRECT_CS_CONNECTED)
 346                 queue_work(smb_direct_wq, &t->disconnect_work);
 347 }
 348
 349 static void smb_direct_send_immediate_work(struct work_struct *work)
 350 {
 351         struct smb_direct_transport *t = container_of(work,
 352                         struct smb_direct_transport, send_immediate_work);
 353
 354         if (t->status != SMB_DIRECT_CS_CONNECTED)
 355                 return;
 356
 357         smb_direct_post_send_data(t, NULL, NULL, 0, 0);
 358 }
 359
 360 static struct smb_direct_transport *alloc_transport(struct rdma_cm_id *cm_id)
 361 {
 362         struct smb_direct_transport *t;
 363         struct ksmbd_conn *conn;
 364
 365         t = kzalloc(sizeof(*t), GFP_KERNEL);
 366         if (!t)
 367                 return NULL;
 368
 369         t->cm_id = cm_id;
 370         cm_id->context = t;
 371
 372         t->status = SMB_DIRECT_CS_NEW;
 373         init_waitqueue_head(&t->wait_status);
 374
 375         spin_lock_init(&t->reassembly_queue_lock);
 376         INIT_LIST_HEAD(&t->reassembly_queue);
 377         t->reassembly_data_length = 0;
 378         t->reassembly_queue_length = 0;
 379         init_waitqueue_head(&t->wait_reassembly_queue);
 380         init_waitqueue_head(&t->wait_send_credits);
 381         init_waitqueue_head(&t->wait_rw_credits);
 382
 383         spin_lock_init(&t->receive_credit_lock);
 384         spin_lock_init(&t->recvmsg_queue_lock);
 385         INIT_LIST_HEAD(&t->recvmsg_queue);
 386
 387         spin_lock_init(&t->empty_recvmsg_queue_lock);
 388         INIT_LIST_HEAD(&t->empty_recvmsg_queue);
 389
 390         init_waitqueue_head(&t->wait_send_pending);
 391         atomic_set(&t->send_pending, 0);
 392
 393         spin_lock_init(&t->lock_new_recv_credits);
 394
 395         INIT_DELAYED_WORK(&t->post_recv_credits_work,
 396                           smb_direct_post_recv_credits);
 397         INIT_WORK(&t->send_immediate_work, smb_direct_send_immediate_work);
 398         INIT_WORK(&t->disconnect_work, smb_direct_disconnect_rdma_work);
 399
 400         conn = ksmbd_conn_alloc();
 401         if (!conn)
 402                 goto err;
 403         conn->transport = KSMBD_TRANS(t);
 404         KSMBD_TRANS(t)->conn = conn;
 405         KSMBD_TRANS(t)->ops = &ksmbd_smb_direct_transport_ops;
 406         return t;
 407 err:
 408         kfree(t);
 409         return NULL;
 410 }
 411
 412 static void free_transport(struct smb_direct_transport *t)
 413 {
 414         struct smb_direct_recvmsg *recvmsg;
 415
 416         wake_up_interruptible(&t->wait_send_credits);
 417
 418         ksmbd_debug(RDMA, "wait for all send posted to IB to finish\n");
 419         wait_event(t->wait_send_pending,
 420                    atomic_read(&t->send_pending) == 0);
 421
 422         cancel_work_sync(&t->disconnect_work);
 423         cancel_delayed_work_sync(&t->post_recv_credits_work);
 424         cancel_work_sync(&t->send_immediate_work);
 425
 426         if (t->qp) {
 427                 ib_drain_qp(t->qp);
 428                 ib_mr_pool_destroy(t->qp, &t->qp->rdma_mrs);
 429                 ib_destroy_qp(t->qp);
 430         }
 431
 432         ksmbd_debug(RDMA, "drain the reassembly queue\n");
 433         do {
 434                 spin_lock(&t->reassembly_queue_lock);
 435                 recvmsg = get_first_reassembly(t);
 436                 if (recvmsg) {
 437                         list_del(&recvmsg->list);
 438                         spin_unlock(&t->reassembly_queue_lock);
 439                         put_recvmsg(t, recvmsg);
 440                 } else {
 441                         spin_unlock(&t->reassembly_queue_lock);
 442                 }
 443         } while (recvmsg);
 444         t->reassembly_data_length = 0;
 445
 446         if (t->send_cq)
 447                 ib_free_cq(t->send_cq);
 448         if (t->recv_cq)
 449                 ib_free_cq(t->recv_cq);
 450         if (t->pd)
 451                 ib_dealloc_pd(t->pd);
 452         if (t->cm_id)
 453                 rdma_destroy_id(t->cm_id);
 454
 455         smb_direct_destroy_pools(t);
 456         ksmbd_conn_free(KSMBD_TRANS(t)->conn);
 457         kfree(t);
 458 }
 459
 460 static struct smb_direct_sendmsg
 461 *smb_direct_alloc_sendmsg(struct smb_direct_transport *t)
 462 {
 463         struct smb_direct_sendmsg *msg;
 464
 465         msg = mempool_alloc(t->sendmsg_mempool, GFP_KERNEL);
 466         if (!msg)
 467                 return ERR_PTR(-ENOMEM);
 468         msg->transport = t;
 469         INIT_LIST_HEAD(&msg->list);
 470         msg->num_sge = 0;
 471         return msg;
 472 }
 473
 474 static void smb_direct_free_sendmsg(struct smb_direct_transport *t,
 475                                     struct smb_direct_sendmsg *msg)
 476 {
 477         int i;
 478
 479         if (msg->num_sge > 0) {
 480                 ib_dma_unmap_single(t->cm_id->device,
 481                                     msg->sge[0].addr, msg->sge[0].length,
 482                                     DMA_TO_DEVICE);
 483                 for (i = 1; i < msg->num_sge; i++)
 484                         ib_dma_unmap_page(t->cm_id->device,
 485                                           msg->sge[i].addr, msg->sge[i].length,
 486                                           DMA_TO_DEVICE);
 487         }
 488         mempool_free(msg, t->sendmsg_mempool);
 489 }
 490
 491 static int smb_direct_check_recvmsg(struct smb_direct_recvmsg *recvmsg)
 492 {
 493         switch (recvmsg->type) {
 494         case SMB_DIRECT_MSG_DATA_TRANSFER: {
 495                 struct smb_direct_data_transfer *req =
 496                         (struct smb_direct_data_transfer *)recvmsg->packet;
 497                 struct smb2_hdr *hdr = (struct smb2_hdr *)(recvmsg->packet
 498                                 + le32_to_cpu(req->data_offset));
 499                 ksmbd_debug(RDMA,
 500                             "CreditGranted: %u, CreditRequested: %u, DataLength: %u, RemainingDataLength: %u, SMB: %x, Command: %u\n",
 501                             le16_to_cpu(req->credits_granted),
 502                             le16_to_cpu(req->credits_requested),
 503                             req->data_length, req->remaining_data_length,
 504                             hdr->ProtocolId, hdr->Command);
 505                 break;
 506         }
 507         case SMB_DIRECT_MSG_NEGOTIATE_REQ: {
 508                 struct smb_direct_negotiate_req *req =
 509                         (struct smb_direct_negotiate_req *)recvmsg->packet;
 510                 ksmbd_debug(RDMA,
 511                             "MinVersion: %u, MaxVersion: %u, CreditRequested: %u, MaxSendSize: %u, MaxRecvSize: %u, MaxFragmentedSize: %u\n",
 512                             le16_to_cpu(req->min_version),
 513                             le16_to_cpu(req->max_version),
 514                             le16_to_cpu(req->credits_requested),
 515                             le32_to_cpu(req->preferred_send_size),
 516                             le32_to_cpu(req->max_receive_size),
 517                             le32_to_cpu(req->max_fragmented_size));
 518                 if (le16_to_cpu(req->min_version) > 0x0100 ||
 519                     le16_to_cpu(req->max_version) < 0x0100)
 520                         return -EOPNOTSUPP;
 521                 if (le16_to_cpu(req->credits_requested) <= 0 ||
 522                     le32_to_cpu(req->max_receive_size) <= 128 ||
 523                     le32_to_cpu(req->max_fragmented_size) <=
 524                                         128 * 1024)
 525                         return -ECONNABORTED;
 526
 527                 break;
 528         }
 529         default:
 530                 return -EINVAL;
 531         }
 532         return 0;
 533 }
 534
 535 static void recv_done(struct ib_cq *cq, struct ib_wc *wc)
 536 {
 537         struct smb_direct_recvmsg *recvmsg;
 538         struct smb_direct_transport *t;
 539
 540         recvmsg = container_of(wc->wr_cqe, struct smb_direct_recvmsg, cqe);
 541         t = recvmsg->transport;
 542
 543         if (wc->status != IB_WC_SUCCESS || wc->opcode != IB_WC_RECV) {
 544                 if (wc->status != IB_WC_WR_FLUSH_ERR) {
 545                         pr_err("Recv error. status='%s (%d)' opcode=%d\n",
 546                                ib_wc_status_msg(wc->status), wc->status,
 547                                wc->opcode);
 548                         smb_direct_disconnect_rdma_connection(t);
 549                 }
 550                 put_empty_recvmsg(t, recvmsg);
 551                 return;
 552         }
 553
 554         ksmbd_debug(RDMA, "Recv completed. status='%s (%d)', opcode=%d\n",
 555                     ib_wc_status_msg(wc->status), wc->status,
 556                     wc->opcode);
 557
 558         ib_dma_sync_single_for_cpu(wc->qp->device, recvmsg->sge.addr,
 559                                    recvmsg->sge.length, DMA_FROM_DEVICE);
 560
 561         switch (recvmsg->type) {
 562         case SMB_DIRECT_MSG_NEGOTIATE_REQ:
 563                 if (wc->byte_len < sizeof(struct smb_direct_negotiate_req)) {
 564                         put_empty_recvmsg(t, recvmsg);
 565                         return;
 566                 }
 567                 t->negotiation_requested = true;
 568                 t->full_packet_received = true;
 569                 t->status = SMB_DIRECT_CS_CONNECTED;
 570                 enqueue_reassembly(t, recvmsg, 0);
 571                 wake_up_interruptible(&t->wait_status);
 572                 break;
 573         case SMB_DIRECT_MSG_DATA_TRANSFER: {
 574                 struct smb_direct_data_transfer *data_transfer =
 575                         (struct smb_direct_data_transfer *)recvmsg->packet;
 576                 unsigned int data_length;
 577                 int avail_recvmsg_count, receive_credits;
 578
 579                 if (wc->byte_len <
 580                     offsetof(struct smb_direct_data_transfer, padding)) {
 581                         put_empty_recvmsg(t, recvmsg);
 582                         return;
 583                 }
 584
 585                 data_length = le32_to_cpu(data_transfer->data_length);
 586                 if (data_length) {
 587                         if (wc->byte_len < sizeof(struct smb_direct_data_transfer) +
 588                             (u64)data_length) {
 589                                 put_empty_recvmsg(t, recvmsg);
 590                                 return;
 591                         }
 592
 593                         if (t->full_packet_received)
 594                                 recvmsg->first_segment = true;
 595
 596                         if (le32_to_cpu(data_transfer->remaining_data_length))
 597                                 t->full_packet_received = false;
 598                         else
 599                                 t->full_packet_received = true;
 600
 601                         enqueue_reassembly(t, recvmsg, (int)data_length);
 602                         wake_up_interruptible(&t->wait_reassembly_queue);
 603
 604                         spin_lock(&t->receive_credit_lock);
 605                         receive_credits = --(t->recv_credits);
 606                         avail_recvmsg_count = t->count_avail_recvmsg;
 607                         spin_unlock(&t->receive_credit_lock);
 608                 } else {
 609                         put_empty_recvmsg(t, recvmsg);
 610
 611                         spin_lock(&t->receive_credit_lock);
 612                         receive_credits = --(t->recv_credits);
 613                         avail_recvmsg_count = ++(t->count_avail_recvmsg);
 614                         spin_unlock(&t->receive_credit_lock);
 615                 }
 616
 617                 t->recv_credit_target =
 618                                 le16_to_cpu(data_transfer->credits_requested);
 619                 atomic_add(le16_to_cpu(data_transfer->credits_granted),
 620                            &t->send_credits);
 621
 622                 if (le16_to_cpu(data_transfer->flags) &
 623                     SMB_DIRECT_RESPONSE_REQUESTED)
 624                         queue_work(smb_direct_wq, &t->send_immediate_work);
 625
 626                 if (atomic_read(&t->send_credits) > 0)
 627                         wake_up_interruptible(&t->wait_send_credits);
 628
 629                 if (is_receive_credit_post_required(receive_credits, avail_recvmsg_count))
 630                         mod_delayed_work(smb_direct_wq,
 631                                          &t->post_recv_credits_work, 0);
 632                 break;
 633         }
 634         default:
 635                 break;
 636         }
 637 }
 638
 639 static int smb_direct_post_recv(struct smb_direct_transport *t,
 640                                 struct smb_direct_recvmsg *recvmsg)
 641 {
 642         struct ib_recv_wr wr;
 643         int ret;
 644
 645         recvmsg->sge.addr = ib_dma_map_single(t->cm_id->device,
 646                                               recvmsg->packet, t->max_recv_size,
 647                                               DMA_FROM_DEVICE);
 648         ret = ib_dma_mapping_error(t->cm_id->device, recvmsg->sge.addr);
 649         if (ret)
 650                 return ret;
 651         recvmsg->sge.length = t->max_recv_size;
 652         recvmsg->sge.lkey = t->pd->local_dma_lkey;
 653         recvmsg->cqe.done = recv_done;
 654
 655         wr.wr_cqe = &recvmsg->cqe;
 656         wr.next = NULL;
 657         wr.sg_list = &recvmsg->sge;
 658         wr.num_sge = 1;
 659
 660         ret = ib_post_recv(t->qp, &wr, NULL);
 661         if (ret) {
 662                 pr_err("Can't post recv: %d\n", ret);
 663                 ib_dma_unmap_single(t->cm_id->device,
 664                                     recvmsg->sge.addr, recvmsg->sge.length,
 665                                     DMA_FROM_DEVICE);
 666                 smb_direct_disconnect_rdma_connection(t);
 667                 return ret;
 668         }
 669         return ret;
 670 }
 671
 672 static int smb_direct_read(struct ksmbd_transport *t, char *buf,
 673                            unsigned int size)
 674 {
 675         struct smb_direct_recvmsg *recvmsg;
 676         struct smb_direct_data_transfer *data_transfer;
 677         int to_copy, to_read, data_read, offset;
 678         u32 data_length, remaining_data_length, data_offset;
 679         int rc;
 680         struct smb_direct_transport *st = smb_trans_direct_transfort(t);
 681
 682 again:
 683         if (st->status != SMB_DIRECT_CS_CONNECTED) {
 684                 pr_err("disconnected\n");
 685                 return -ENOTCONN;
 686         }
 687
 688         /*
 689          * No need to hold the reassembly queue lock all the time as we are
 690          * the only one reading from the front of the queue. The transport
 691          * may add more entries to the back of the queue at the same time
 692          */
 693         if (st->reassembly_data_length >= size) {
 694                 int queue_length;
 695                 int queue_removed = 0;
 696
 697                 /*
 698                  * Need to make sure reassembly_data_length is read before
 699                  * reading reassembly_queue_length and calling
 700                  * get_first_reassembly. This call is lock free
 701                  * as we never read at the end of the queue which are being
 702                  * updated in SOFTIRQ as more data is received
 703                  */
 704                 virt_rmb();
 705                 queue_length = st->reassembly_queue_length;
 706                 data_read = 0;
 707                 to_read = size;
 708                 offset = st->first_entry_offset;
 709                 while (data_read < size) {
 710                         recvmsg = get_first_reassembly(st);
 711                         data_transfer = smb_direct_recvmsg_payload(recvmsg);
 712                         data_length = le32_to_cpu(data_transfer->data_length);
 713                         remaining_data_length =
 714                                 le32_to_cpu(data_transfer->remaining_data_length);
 715                         data_offset = le32_to_cpu(data_transfer->data_offset);
 716
 717                         /*
 718                          * The upper layer expects RFC1002 length at the
 719                          * beginning of the payload. Return it to indicate
 720                          * the total length of the packet. This minimize the
 721                          * change to upper layer packet processing logic. This
 722                          * will be eventually remove when an intermediate
 723                          * transport layer is added
 724                          */
 725                         if (recvmsg->first_segment && size == 4) {
 726                                 unsigned int rfc1002_len =
 727                                         data_length + remaining_data_length;
 728                                 *((__be32 *)buf) = cpu_to_be32(rfc1002_len);
 729                                 data_read = 4;
 730                                 recvmsg->first_segment = false;
 731                                 ksmbd_debug(RDMA,
 732                                             "returning rfc1002 length %d\n",
 733                                             rfc1002_len);
 734                                 goto read_rfc1002_done;
 735                         }
 736
 737                         to_copy = min_t(int, data_length - offset, to_read);
 738                         memcpy(buf + data_read, (char *)data_transfer + data_offset + offset,
 739                                to_copy);
 740
 741                         /* move on to the next buffer? */
 742                         if (to_copy == data_length - offset) {
 743                                 queue_length--;
 744                                 /*
 745                                  * No need to lock if we are not at the
 746                                  * end of the queue
 747                                  */
 748                                 if (queue_length) {
 749                                         list_del(&recvmsg->list);
 750                                 } else {
 751                                         spin_lock_irq(&st->reassembly_queue_lock);
 752                                         list_del(&recvmsg->list);
 753                                         spin_unlock_irq(&st->reassembly_queue_lock);
 754                                 }
 755                                 queue_removed++;
 756                                 put_recvmsg(st, recvmsg);
 757                                 offset = 0;
 758                         } else {
 759                                 offset += to_copy;
 760                         }
 761
 762                         to_read -= to_copy;
 763                         data_read += to_copy;
 764                 }
 765
 766                 spin_lock_irq(&st->reassembly_queue_lock);
 767                 st->reassembly_data_length -= data_read;
 768                 st->reassembly_queue_length -= queue_removed;
 769                 spin_unlock_irq(&st->reassembly_queue_lock);
 770
 771                 spin_lock(&st->receive_credit_lock);
 772                 st->count_avail_recvmsg += queue_removed;
 773                 if (is_receive_credit_post_required(st->recv_credits, st->count_avail_recvmsg)) {
 774                         spin_unlock(&st->receive_credit_lock);
 775                         mod_delayed_work(smb_direct_wq,
 776                                          &st->post_recv_credits_work, 0);
 777                 } else {
 778                         spin_unlock(&st->receive_credit_lock);
 779                 }
 780
 781                 st->first_entry_offset = offset;
 782                 ksmbd_debug(RDMA,
 783                             "returning to thread data_read=%d reassembly_data_length=%d first_entry_offset=%d\n",
 784                             data_read, st->reassembly_data_length,
 785                             st->first_entry_offset);
 786 read_rfc1002_done:
 787                 return data_read;
 788         }
 789
 790         ksmbd_debug(RDMA, "wait_event on more data\n");
 791         rc = wait_event_interruptible(st->wait_reassembly_queue,
 792                                       st->reassembly_data_length >= size ||
 793                                        st->status != SMB_DIRECT_CS_CONNECTED);
 794         if (rc)
 795                 return -EINTR;
 796
 797         goto again;
 798 }
 799
 800 static void smb_direct_post_recv_credits(struct work_struct *work)
 801 {
 802         struct smb_direct_transport *t = container_of(work,
 803                 struct smb_direct_transport, post_recv_credits_work.work);
 804         struct smb_direct_recvmsg *recvmsg;
 805         int receive_credits, credits = 0;
 806         int ret;
 807         int use_free = 1;
 808
 809         spin_lock(&t->receive_credit_lock);
 810         receive_credits = t->recv_credits;
 811         spin_unlock(&t->receive_credit_lock);
 812
 813         if (receive_credits < t->recv_credit_target) {
 814                 while (true) {
 815                         if (use_free)
 816                                 recvmsg = get_free_recvmsg(t);
 817                         else
 818                                 recvmsg = get_empty_recvmsg(t);
 819                         if (!recvmsg) {
 820                                 if (use_free) {
 821                                         use_free = 0;
 822                                         continue;
 823                                 } else {
 824                                         break;
 825                                 }
 826                         }
 827
 828                         recvmsg->type = SMB_DIRECT_MSG_DATA_TRANSFER;
 829                         recvmsg->first_segment = false;
 830
 831                         ret = smb_direct_post_recv(t, recvmsg);
 832                         if (ret) {
 833                                 pr_err("Can't post recv: %d\n", ret);
 834                                 put_recvmsg(t, recvmsg);
 835                                 break;
 836                         }
 837                         credits++;
 838                 }
 839         }
 840
 841         spin_lock(&t->receive_credit_lock);
 842         t->recv_credits += credits;
 843         t->count_avail_recvmsg -= credits;
 844         spin_unlock(&t->receive_credit_lock);
 845
 846         spin_lock(&t->lock_new_recv_credits);
 847         t->new_recv_credits += credits;
 848         spin_unlock(&t->lock_new_recv_credits);
 849
 850         if (credits)
 851                 queue_work(smb_direct_wq, &t->send_immediate_work);
 852 }
 853
 854 static void send_done(struct ib_cq *cq, struct ib_wc *wc)
 855 {
 856         struct smb_direct_sendmsg *sendmsg, *sibling;
 857         struct smb_direct_transport *t;
 858         struct list_head *pos, *prev, *end;
 859
 860         sendmsg = container_of(wc->wr_cqe, struct smb_direct_sendmsg, cqe);
 861         t = sendmsg->transport;
 862
 863         ksmbd_debug(RDMA, "Send completed. status='%s (%d)', opcode=%d\n",
 864                     ib_wc_status_msg(wc->status), wc->status,
 865                     wc->opcode);
 866
 867         if (wc->status != IB_WC_SUCCESS || wc->opcode != IB_WC_SEND) {
 868                 pr_err("Send error. status='%s (%d)', opcode=%d\n",
 869                        ib_wc_status_msg(wc->status), wc->status,
 870                        wc->opcode);
 871                 smb_direct_disconnect_rdma_connection(t);
 872         }
 873
 874         if (atomic_dec_and_test(&t->send_pending))
 875                 wake_up(&t->wait_send_pending);
 876
 877         /* iterate and free the list of messages in reverse. the list's head
 878          * is invalid.
 879          */
 880         for (pos = &sendmsg->list, prev = pos->prev, end = sendmsg->list.next;
 881              prev != end; pos = prev, prev = prev->prev) {
 882                 sibling = container_of(pos, struct smb_direct_sendmsg, list);
 883                 smb_direct_free_sendmsg(t, sibling);
 884         }
 885
 886         sibling = container_of(pos, struct smb_direct_sendmsg, list);
 887         smb_direct_free_sendmsg(t, sibling);
 888 }
 889
 890 static int manage_credits_prior_sending(struct smb_direct_transport *t)
 891 {
 892         int new_credits;
 893
 894         spin_lock(&t->lock_new_recv_credits);
 895         new_credits = t->new_recv_credits;
 896         t->new_recv_credits = 0;
 897         spin_unlock(&t->lock_new_recv_credits);
 898
 899         return new_credits;
 900 }
 901
 902 static int smb_direct_post_send(struct smb_direct_transport *t,
 903                                 struct ib_send_wr *wr)
 904 {
 905         int ret;
 906
 907         atomic_inc(&t->send_pending);
 908         ret = ib_post_send(t->qp, wr, NULL);
 909         if (ret) {
 910                 pr_err("failed to post send: %d\n", ret);
 911                 if (atomic_dec_and_test(&t->send_pending))
 912                         wake_up(&t->wait_send_pending);
 913                 smb_direct_disconnect_rdma_connection(t);
 914         }
 915         return ret;
 916 }
 917
 918 static void smb_direct_send_ctx_init(struct smb_direct_transport *t,
 919                                      struct smb_direct_send_ctx *send_ctx,
 920                                      bool need_invalidate_rkey,
 921                                      unsigned int remote_key)
 922 {
 923         INIT_LIST_HEAD(&send_ctx->msg_list);
 924         send_ctx->wr_cnt = 0;
 925         send_ctx->need_invalidate_rkey = need_invalidate_rkey;
 926         send_ctx->remote_key = remote_key;
 927 }
 928
 929 static int smb_direct_flush_send_list(struct smb_direct_transport *t,
 930                                       struct smb_direct_send_ctx *send_ctx,
 931                                       bool is_last)
 932 {
 933         struct smb_direct_sendmsg *first, *last;
 934         int ret;
 935
 936         if (list_empty(&send_ctx->msg_list))
 937                 return 0;
 938
 939         first = list_first_entry(&send_ctx->msg_list,
 940                                  struct smb_direct_sendmsg,
 941                                  list);
 942         last = list_last_entry(&send_ctx->msg_list,
 943                                struct smb_direct_sendmsg,
 944                                list);
 945
 946         last->wr.send_flags = IB_SEND_SIGNALED;
 947         last->wr.wr_cqe = &last->cqe;
 948         if (is_last && send_ctx->need_invalidate_rkey) {
 949                 last->wr.opcode = IB_WR_SEND_WITH_INV;
 950                 last->wr.ex.invalidate_rkey = send_ctx->remote_key;
 951         }
 952
 953         ret = smb_direct_post_send(t, &first->wr);
 954         if (!ret) {
 955                 smb_direct_send_ctx_init(t, send_ctx,
 956                                          send_ctx->need_invalidate_rkey,
 957                                          send_ctx->remote_key);
 958         } else {
 959                 atomic_add(send_ctx->wr_cnt, &t->send_credits);
 960                 wake_up(&t->wait_send_credits);
 961                 list_for_each_entry_safe(first, last, &send_ctx->msg_list,
 962                                          list) {
 963                         smb_direct_free_sendmsg(t, first);
 964                 }
 965         }
 966         return ret;
 967 }
 968
 969 static int wait_for_credits(struct smb_direct_transport *t,
 970                             wait_queue_head_t *waitq, atomic_t *total_credits,
 971                             int needed)
 972 {
 973         int ret;
 974
 975         do {
 976                 if (atomic_sub_return(needed, total_credits) >= 0)
 977                         return 0;
 978
 979                 atomic_add(needed, total_credits);
 980                 ret = wait_event_interruptible(*waitq,
 981                                                atomic_read(total_credits) >= needed ||
 982                                                t->status != SMB_DIRECT_CS_CONNECTED);
 983
 984                 if (t->status != SMB_DIRECT_CS_CONNECTED)
 985                         return -ENOTCONN;
 986                 else if (ret < 0)
 987                         return ret;
 988         } while (true);
 989 }
 990
 991 static int wait_for_send_credits(struct smb_direct_transport *t,
 992                                  struct smb_direct_send_ctx *send_ctx)
 993 {
 994         int ret;
 995
 996         if (send_ctx &&
 997             (send_ctx->wr_cnt >= 16 || atomic_read(&t->send_credits) <= 1)) {
 998                 ret = smb_direct_flush_send_list(t, send_ctx, false);
 999                 if (ret)
1000                         return ret;
1001         }
1002
1003         return wait_for_credits(t, &t->wait_send_credits, &t->send_credits, 1);
1004 }
1005
1006 static int wait_for_rw_credits(struct smb_direct_transport *t, int credits)
1007 {
1008         return wait_for_credits(t, &t->wait_rw_credits, &t->rw_credits, credits);
1009 }
1010
1011 static int calc_rw_credits(struct smb_direct_transport *t,
1012                            char *buf, unsigned int len)
1013 {
1014         return DIV_ROUND_UP(get_buf_page_count(buf, len),
1015                             t->pages_per_rw_credit);
1016 }
1017
1018 static int smb_direct_create_header(struct smb_direct_transport *t,
1019                                     int size, int remaining_data_length,
1020                                     struct smb_direct_sendmsg **sendmsg_out)
1021 {
1022         struct smb_direct_sendmsg *sendmsg;
1023         struct smb_direct_data_transfer *packet;
1024         int header_length;
1025         int ret;
1026
1027         sendmsg = smb_direct_alloc_sendmsg(t);
1028         if (IS_ERR(sendmsg))
1029                 return PTR_ERR(sendmsg);
1030
1031         /* Fill in the packet header */
1032         packet = (struct smb_direct_data_transfer *)sendmsg->packet;
1033         packet->credits_requested = cpu_to_le16(t->send_credit_target);
1034         packet->credits_granted = cpu_to_le16(manage_credits_prior_sending(t));
1035
1036         packet->flags = 0;
1037         packet->reserved = 0;
1038         if (!size)
1039                 packet->data_offset = 0;
1040         else
1041                 packet->data_offset = cpu_to_le32(24);
1042         packet->data_length = cpu_to_le32(size);
1043         packet->remaining_data_length = cpu_to_le32(remaining_data_length);
1044         packet->padding = 0;
1045
1046         ksmbd_debug(RDMA,
1047                     "credits_requested=%d credits_granted=%d data_offset=%d data_length=%d remaining_data_length=%d\n",
1048                     le16_to_cpu(packet->credits_requested),
1049                     le16_to_cpu(packet->credits_granted),
1050                     le32_to_cpu(packet->data_offset),
1051                     le32_to_cpu(packet->data_length),
1052                     le32_to_cpu(packet->remaining_data_length));
1053
1054         /* Map the packet to DMA */
1055         header_length = sizeof(struct smb_direct_data_transfer);
1056         /* If this is a packet without payload, don't send padding */
1057         if (!size)
1058                 header_length =
1059                         offsetof(struct smb_direct_data_transfer, padding);
1060
1061         sendmsg->sge[0].addr = ib_dma_map_single(t->cm_id->device,
1062                                                  (void *)packet,
1063                                                  header_length,
1064                                                  DMA_TO_DEVICE);
1065         ret = ib_dma_mapping_error(t->cm_id->device, sendmsg->sge[0].addr);
1066         if (ret) {
1067                 smb_direct_free_sendmsg(t, sendmsg);
1068                 return ret;
1069         }
1070
1071         sendmsg->num_sge = 1;
1072         sendmsg->sge[0].length = header_length;
1073         sendmsg->sge[0].lkey = t->pd->local_dma_lkey;
1074
1075         *sendmsg_out = sendmsg;
1076         return 0;
1077 }
1078
1079 static int get_sg_list(void *buf, int size, struct scatterlist *sg_list, int nentries)
1080 {
1081         bool high = is_vmalloc_addr(buf);
1082         struct page *page;
1083         int offset, len;
1084         int i = 0;
1085
1086         if (size <= 0 || nentries < get_buf_page_count(buf, size))
1087                 return -EINVAL;
1088
1089         offset = offset_in_page(buf);
1090         buf -= offset;
1091         while (size > 0) {
1092                 len = min_t(int, PAGE_SIZE - offset, size);
1093                 if (high)
1094                         page = vmalloc_to_page(buf);
1095                 else
1096                         page = kmap_to_page(buf);
1097
1098                 if (!sg_list)
1099                         return -EINVAL;
1100                 sg_set_page(sg_list, page, len, offset);
1101                 sg_list = sg_next(sg_list);
1102
1103                 buf += PAGE_SIZE;
1104                 size -= len;
1105                 offset = 0;
1106                 i++;
1107         }
1108         return i;
1109 }
1110
1111 static int get_mapped_sg_list(struct ib_device *device, void *buf, int size,
1112                               struct scatterlist *sg_list, int nentries,
1113                               enum dma_data_direction dir)
1114 {
1115         int npages;
1116
1117         npages = get_sg_list(buf, size, sg_list, nentries);
1118         if (npages < 0)
1119                 return -EINVAL;
1120         return ib_dma_map_sg(device, sg_list, npages, dir);
1121 }
1122
1123 static int post_sendmsg(struct smb_direct_transport *t,
1124                         struct smb_direct_send_ctx *send_ctx,
1125                         struct smb_direct_sendmsg *msg)
1126 {
1127         int i;
1128
1129         for (i = 0; i < msg->num_sge; i++)
1130                 ib_dma_sync_single_for_device(t->cm_id->device,
1131                                               msg->sge[i].addr, msg->sge[i].length,
1132                                               DMA_TO_DEVICE);
1133
1134         msg->cqe.done = send_done;
1135         msg->wr.opcode = IB_WR_SEND;
1136         msg->wr.sg_list = &msg->sge[0];
1137         msg->wr.num_sge = msg->num_sge;
1138         msg->wr.next = NULL;
1139
1140         if (send_ctx) {
1141                 msg->wr.wr_cqe = NULL;
1142                 msg->wr.send_flags = 0;
1143                 if (!list_empty(&send_ctx->msg_list)) {
1144                         struct smb_direct_sendmsg *last;
1145
1146                         last = list_last_entry(&send_ctx->msg_list,
1147                                                struct smb_direct_sendmsg,
1148                                                list);
1149                         last->wr.next = &msg->wr;
1150                 }
1151                 list_add_tail(&msg->list, &send_ctx->msg_list);
1152                 send_ctx->wr_cnt++;
1153                 return 0;
1154         }
1155
1156         msg->wr.wr_cqe = &msg->cqe;
1157         msg->wr.send_flags = IB_SEND_SIGNALED;
1158         return smb_direct_post_send(t, &msg->wr);
1159 }
1160
1161 static int smb_direct_post_send_data(struct smb_direct_transport *t,
1162                                      struct smb_direct_send_ctx *send_ctx,
1163                                      struct kvec *iov, int niov,
1164                                      int remaining_data_length)
1165 {
1166         int i, j, ret;
1167         struct smb_direct_sendmsg *msg;
1168         int data_length;
1169         struct scatterlist sg[SMB_DIRECT_MAX_SEND_SGES - 1];
1170
1171         ret = wait_for_send_credits(t, send_ctx);
1172         if (ret)
1173                 return ret;
1174
1175         data_length = 0;
1176         for (i = 0; i < niov; i++)
1177                 data_length += iov[i].iov_len;
1178
1179         ret = smb_direct_create_header(t, data_length, remaining_data_length,
1180                                        &msg);
1181         if (ret) {
1182                 atomic_inc(&t->send_credits);
1183                 return ret;
1184         }
1185
1186         for (i = 0; i < niov; i++) {
1187                 struct ib_sge *sge;
1188                 int sg_cnt;
1189
1190                 sg_init_table(sg, SMB_DIRECT_MAX_SEND_SGES - 1);
1191                 sg_cnt = get_mapped_sg_list(t->cm_id->device,
1192                                             iov[i].iov_base, iov[i].iov_len,
1193                                             sg, SMB_DIRECT_MAX_SEND_SGES - 1,
1194                                             DMA_TO_DEVICE);
1195                 if (sg_cnt <= 0) {
1196                         pr_err("failed to map buffer\n");
1197                         ret = -ENOMEM;
1198                         goto err;
1199                 } else if (sg_cnt + msg->num_sge > SMB_DIRECT_MAX_SEND_SGES) {
1200                         pr_err("buffer not fitted into sges\n");
1201                         ret = -E2BIG;
1202                         ib_dma_unmap_sg(t->cm_id->device, sg, sg_cnt,
1203                                         DMA_TO_DEVICE);
1204                         goto err;
1205                 }
1206
1207                 for (j = 0; j < sg_cnt; j++) {
1208                         sge = &msg->sge[msg->num_sge];
1209                         sge->addr = sg_dma_address(&sg[j]);
1210                         sge->length = sg_dma_len(&sg[j]);
1211                         sge->lkey  = t->pd->local_dma_lkey;
1212                         msg->num_sge++;
1213                 }
1214         }
1215
1216         ret = post_sendmsg(t, send_ctx, msg);
1217         if (ret)
1218                 goto err;
1219         return 0;
1220 err:
1221         smb_direct_free_sendmsg(t, msg);
1222         atomic_inc(&t->send_credits);
1223         return ret;
1224 }
1225
1226 static int smb_direct_writev(struct ksmbd_transport *t,
1227                              struct kvec *iov, int niovs, int buflen,
1228                              bool need_invalidate, unsigned int remote_key)
1229 {
1230         struct smb_direct_transport *st = smb_trans_direct_transfort(t);
1231         int remaining_data_length;
1232         int start, i, j;
1233         int max_iov_size = st->max_send_size -
1234                         sizeof(struct smb_direct_data_transfer);
1235         int ret;
1236         struct kvec vec;
1237         struct smb_direct_send_ctx send_ctx;
1238
1239         if (st->status != SMB_DIRECT_CS_CONNECTED)
1240                 return -ENOTCONN;
1241
1242         //FIXME: skip RFC1002 header..
1243         buflen -= 4;
1244         iov[0].iov_base += 4;
1245         iov[0].iov_len -= 4;
1246
1247         remaining_data_length = buflen;
1248         ksmbd_debug(RDMA, "Sending smb (RDMA): smb_len=%u\n", buflen);
1249
1250         smb_direct_send_ctx_init(st, &send_ctx, need_invalidate, remote_key);
1251         start = i = 0;
1252         buflen = 0;
1253         while (true) {
1254                 buflen += iov[i].iov_len;
1255                 if (buflen > max_iov_size) {
1256                         if (i > start) {
1257                                 remaining_data_length -=
1258                                         (buflen - iov[i].iov_len);
1259                                 ret = smb_direct_post_send_data(st, &send_ctx,
1260                                                                 &iov[start], i - start,
1261                                                                 remaining_data_length);
1262                                 if (ret)
1263                                         goto done;
1264                         } else {
1265                                 /* iov[start] is too big, break it */
1266                                 int nvec  = (buflen + max_iov_size - 1) /
1267                                                 max_iov_size;
1268
1269                                 for (j = 0; j < nvec; j++) {
1270                                         vec.iov_base =
1271                                                 (char *)iov[start].iov_base +
1272                                                 j * max_iov_size;
1273                                         vec.iov_len =
1274                                                 min_t(int, max_iov_size,
1275                                                       buflen - max_iov_size * j);
1276                                         remaining_data_length -= vec.iov_len;
1277                                         ret = smb_direct_post_send_data(st, &send_ctx, &vec, 1,
1278                                                                         remaining_data_length);
1279                                         if (ret)
1280                                                 goto done;
1281                                 }
1282                                 i++;
1283                                 if (i == niovs)
1284                                         break;
1285                         }
1286                         start = i;
1287                         buflen = 0;
1288                 } else {
1289                         i++;
1290                         if (i == niovs) {
1291                                 /* send out all remaining vecs */
1292                                 remaining_data_length -= buflen;
1293                                 ret = smb_direct_post_send_data(st, &send_ctx,
1294                                                                 &iov[start], i - start,
1295                                                                 remaining_data_length);
1296                                 if (ret)
1297                                         goto done;
1298                                 break;
1299                         }
1300                 }
1301         }
1302
1303 done:
1304         ret = smb_direct_flush_send_list(st, &send_ctx, true);
1305
1306         /*
1307          * As an optimization, we don't wait for individual I/O to finish
1308          * before sending the next one.
1309          * Send them all and wait for pending send count to get to 0
1310          * that means all the I/Os have been out and we are good to return
1311          */
1312
1313         wait_event(st->wait_send_pending,
1314                    atomic_read(&st->send_pending) == 0);
1315         return ret;
1316 }
1317
1318 static void smb_direct_free_rdma_rw_msg(struct smb_direct_transport *t,
1319                                         struct smb_direct_rdma_rw_msg *msg,
1320                                         enum dma_data_direction dir)
1321 {
1322         rdma_rw_ctx_destroy(&msg->rw_ctx, t->qp, t->qp->port,
1323                             msg->sgt.sgl, msg->sgt.nents, dir);
1324         sg_free_table_chained(&msg->sgt, SG_CHUNK_SIZE);
1325         kfree(msg);
1326 }
1327
1328 static void read_write_done(struct ib_cq *cq, struct ib_wc *wc,
1329                             enum dma_data_direction dir)
1330 {
1331         struct smb_direct_rdma_rw_msg *msg = container_of(wc->wr_cqe,
1332                                                           struct smb_direct_rdma_rw_msg, cqe);
1333         struct smb_direct_transport *t = msg->t;
1334
1335         if (wc->status != IB_WC_SUCCESS) {
1336                 msg->status = -EIO;
1337                 pr_err("read/write error. opcode = %d, status = %s(%d)\n",
1338                        wc->opcode, ib_wc_status_msg(wc->status), wc->status);
1339                 if (wc->status != IB_WC_WR_FLUSH_ERR)
1340                         smb_direct_disconnect_rdma_connection(t);
1341         }
1342
1343         complete(msg->completion);
1344 }
1345
1346 static void read_done(struct ib_cq *cq, struct ib_wc *wc)
1347 {
1348         read_write_done(cq, wc, DMA_FROM_DEVICE);
1349 }
1350
1351 static void write_done(struct ib_cq *cq, struct ib_wc *wc)
1352 {
1353         read_write_done(cq, wc, DMA_TO_DEVICE);
1354 }
1355
1356 static int smb_direct_rdma_xmit(struct smb_direct_transport *t,
1357                                 void *buf, int buf_len,
1358                                 struct smb2_buffer_desc_v1 *desc,
1359                                 unsigned int desc_len,
1360                                 bool is_read)
1361 {
1362         struct smb_direct_rdma_rw_msg *msg, *next_msg;
1363         int i, ret;
1364         DECLARE_COMPLETION_ONSTACK(completion);
1365         struct ib_send_wr *first_wr;
1366         LIST_HEAD(msg_list);
1367         char *desc_buf;
1368         int credits_needed;
1369         unsigned int desc_buf_len;
1370         size_t total_length = 0;
1371
1372         if (t->status != SMB_DIRECT_CS_CONNECTED)
1373                 return -ENOTCONN;
1374
1375         /* calculate needed credits */
1376         credits_needed = 0;
1377         desc_buf = buf;
1378         for (i = 0; i < desc_len / sizeof(*desc); i++) {
1379                 desc_buf_len = le32_to_cpu(desc[i].length);
1380
1381                 credits_needed += calc_rw_credits(t, desc_buf, desc_buf_len);
1382                 desc_buf += desc_buf_len;
1383                 total_length += desc_buf_len;
1384                 if (desc_buf_len == 0 || total_length > buf_len ||
1385                     total_length > t->max_rdma_rw_size)
1386                         return -EINVAL;
1387         }
1388
1389         ksmbd_debug(RDMA, "RDMA %s, len %#x, needed credits %#x\n",
1390                     is_read ? "read" : "write", buf_len, credits_needed);
1391
1392         ret = wait_for_rw_credits(t, credits_needed);
1393         if (ret < 0)
1394                 return ret;
1395
1396         /* build rdma_rw_ctx for each descriptor */
1397         desc_buf = buf;
1398         for (i = 0; i < desc_len / sizeof(*desc); i++) {
1399                 msg = kzalloc(offsetof(struct smb_direct_rdma_rw_msg, sg_list) +
1400                               sizeof(struct scatterlist) * SG_CHUNK_SIZE, GFP_KERNEL);
1401                 if (!msg) {
1402                         ret = -ENOMEM;
1403                         goto out;
1404                 }
1405
1406                 desc_buf_len = le32_to_cpu(desc[i].length);
1407
1408                 msg->t = t;
1409                 msg->cqe.done = is_read ? read_done : write_done;
1410                 msg->completion = &completion;
1411
1412                 msg->sgt.sgl = &msg->sg_list[0];
1413                 ret = sg_alloc_table_chained(&msg->sgt,
1414                                              get_buf_page_count(desc_buf, desc_buf_len),
1415                                              msg->sg_list, SG_CHUNK_SIZE);
1416                 if (ret) {
1417                         kfree(msg);
1418                         ret = -ENOMEM;
1419                         goto out;
1420                 }
1421
1422                 ret = get_sg_list(desc_buf, desc_buf_len,
1423                                   msg->sgt.sgl, msg->sgt.orig_nents);
1424                 if (ret < 0) {
1425                         sg_free_table_chained(&msg->sgt, SG_CHUNK_SIZE);
1426                         kfree(msg);
1427                         goto out;
1428                 }
1429
1430                 ret = rdma_rw_ctx_init(&msg->rw_ctx, t->qp, t->qp->port,
1431                                        msg->sgt.sgl,
1432                                        get_buf_page_count(desc_buf, desc_buf_len),
1433                                        0,
1434                                        le64_to_cpu(desc[i].offset),
1435                                        le32_to_cpu(desc[i].token),
1436                                        is_read ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1437                 if (ret < 0) {
1438                         pr_err("failed to init rdma_rw_ctx: %d\n", ret);
1439                         sg_free_table_chained(&msg->sgt, SG_CHUNK_SIZE);
1440                         kfree(msg);
1441                         goto out;
1442                 }
1443
1444                 list_add_tail(&msg->list, &msg_list);
1445                 desc_buf += desc_buf_len;
1446         }
1447
1448         /* concatenate work requests of rdma_rw_ctxs */
1449         first_wr = NULL;
1450         list_for_each_entry_reverse(msg, &msg_list, list) {
1451                 first_wr = rdma_rw_ctx_wrs(&msg->rw_ctx, t->qp, t->qp->port,
1452                                            &msg->cqe, first_wr);
1453         }
1454
1455         ret = ib_post_send(t->qp, first_wr, NULL);
1456         if (ret) {
1457                 pr_err("failed to post send wr for RDMA R/W: %d\n", ret);
1458                 goto out;
1459         }
1460
1461         msg = list_last_entry(&msg_list, struct smb_direct_rdma_rw_msg, list);
1462         wait_for_completion(&completion);
1463         ret = msg->status;
1464 out:
1465         list_for_each_entry_safe(msg, next_msg, &msg_list, list) {
1466                 list_del(&msg->list);
1467                 smb_direct_free_rdma_rw_msg(t, msg,
1468                                             is_read ? DMA_FROM_DEVICE : DMA_TO_DEVICE);
1469         }
1470         atomic_add(credits_needed, &t->rw_credits);
1471         wake_up(&t->wait_rw_credits);
1472         return ret;
1473 }
1474
1475 static int smb_direct_rdma_write(struct ksmbd_transport *t,
1476                                  void *buf, unsigned int buflen,
1477                                  struct smb2_buffer_desc_v1 *desc,
1478                                  unsigned int desc_len)
1479 {
1480         return smb_direct_rdma_xmit(smb_trans_direct_transfort(t), buf, buflen,
1481                                     desc, desc_len, false);
1482 }
1483
1484 static int smb_direct_rdma_read(struct ksmbd_transport *t,
1485                                 void *buf, unsigned int buflen,
1486                                 struct smb2_buffer_desc_v1 *desc,
1487                                 unsigned int desc_len)
1488 {
1489         return smb_direct_rdma_xmit(smb_trans_direct_transfort(t), buf, buflen,
1490                                     desc, desc_len, true);
1491 }
1492
1493 static void smb_direct_disconnect(struct ksmbd_transport *t)
1494 {
1495         struct smb_direct_transport *st = smb_trans_direct_transfort(t);
1496
1497         ksmbd_debug(RDMA, "Disconnecting cm_id=%p\n", st->cm_id);
1498
1499         smb_direct_disconnect_rdma_work(&st->disconnect_work);
1500         wait_event_interruptible(st->wait_status,
1501                                  st->status == SMB_DIRECT_CS_DISCONNECTED);
1502         free_transport(st);
1503 }
1504
1505 static void smb_direct_shutdown(struct ksmbd_transport *t)
1506 {
1507         struct smb_direct_transport *st = smb_trans_direct_transfort(t);
1508
1509         ksmbd_debug(RDMA, "smb-direct shutdown cm_id=%p\n", st->cm_id);
1510
1511         smb_direct_disconnect_rdma_work(&st->disconnect_work);
1512 }
1513
1514 static int smb_direct_cm_handler(struct rdma_cm_id *cm_id,
1515                                  struct rdma_cm_event *event)
1516 {
1517         struct smb_direct_transport *t = cm_id->context;
1518
1519         ksmbd_debug(RDMA, "RDMA CM event. cm_id=%p event=%s (%d)\n",
1520                     cm_id, rdma_event_msg(event->event), event->event);
1521
1522         switch (event->event) {
1523         case RDMA_CM_EVENT_ESTABLISHED: {
1524                 t->status = SMB_DIRECT_CS_CONNECTED;
1525                 wake_up_interruptible(&t->wait_status);
1526                 break;
1527         }
1528         case RDMA_CM_EVENT_DEVICE_REMOVAL:
1529         case RDMA_CM_EVENT_DISCONNECTED: {
1530                 t->status = SMB_DIRECT_CS_DISCONNECTED;
1531                 wake_up_interruptible(&t->wait_status);
1532                 wake_up_interruptible(&t->wait_reassembly_queue);
1533                 wake_up(&t->wait_send_credits);
1534                 break;
1535         }
1536         case RDMA_CM_EVENT_CONNECT_ERROR: {
1537                 t->status = SMB_DIRECT_CS_DISCONNECTED;
1538                 wake_up_interruptible(&t->wait_status);
1539                 break;
1540         }
1541         default:
1542                 pr_err("Unexpected RDMA CM event. cm_id=%p, event=%s (%d)\n",
1543                        cm_id, rdma_event_msg(event->event),
1544                        event->event);
1545                 break;
1546         }
1547         return 0;
1548 }
1549
1550 static void smb_direct_qpair_handler(struct ib_event *event, void *context)
1551 {
1552         struct smb_direct_transport *t = context;
1553
1554         ksmbd_debug(RDMA, "Received QP event. cm_id=%p, event=%s (%d)\n",
1555                     t->cm_id, ib_event_msg(event->event), event->event);
1556
1557         switch (event->event) {
1558         case IB_EVENT_CQ_ERR:
1559         case IB_EVENT_QP_FATAL:
1560                 smb_direct_disconnect_rdma_connection(t);
1561                 break;
1562         default:
1563                 break;
1564         }
1565 }
1566
1567 static int smb_direct_send_negotiate_response(struct smb_direct_transport *t,
1568                                               int failed)
1569 {
1570         struct smb_direct_sendmsg *sendmsg;
1571         struct smb_direct_negotiate_resp *resp;
1572         int ret;
1573
1574         sendmsg = smb_direct_alloc_sendmsg(t);
1575         if (IS_ERR(sendmsg))
1576                 return -ENOMEM;
1577
1578         resp = (struct smb_direct_negotiate_resp *)sendmsg->packet;
1579         if (failed) {
1580                 memset(resp, 0, sizeof(*resp));
1581                 resp->min_version = cpu_to_le16(0x0100);
1582                 resp->max_version = cpu_to_le16(0x0100);
1583                 resp->status = STATUS_NOT_SUPPORTED;
1584         } else {
1585                 resp->status = STATUS_SUCCESS;
1586                 resp->min_version = SMB_DIRECT_VERSION_LE;
1587                 resp->max_version = SMB_DIRECT_VERSION_LE;
1588                 resp->negotiated_version = SMB_DIRECT_VERSION_LE;
1589                 resp->reserved = 0;
1590                 resp->credits_requested =
1591                                 cpu_to_le16(t->send_credit_target);
1592                 resp->credits_granted = cpu_to_le16(manage_credits_prior_sending(t));
1593                 resp->max_readwrite_size = cpu_to_le32(t->max_rdma_rw_size);
1594                 resp->preferred_send_size = cpu_to_le32(t->max_send_size);
1595                 resp->max_receive_size = cpu_to_le32(t->max_recv_size);
1596                 resp->max_fragmented_size =
1597                                 cpu_to_le32(t->max_fragmented_recv_size);
1598         }
1599
1600         sendmsg->sge[0].addr = ib_dma_map_single(t->cm_id->device,
1601                                                  (void *)resp, sizeof(*resp),
1602                                                  DMA_TO_DEVICE);
1603         ret = ib_dma_mapping_error(t->cm_id->device, sendmsg->sge[0].addr);
1604         if (ret) {
1605                 smb_direct_free_sendmsg(t, sendmsg);
1606                 return ret;
1607         }
1608
1609         sendmsg->num_sge = 1;
1610         sendmsg->sge[0].length = sizeof(*resp);
1611         sendmsg->sge[0].lkey = t->pd->local_dma_lkey;
1612
1613         ret = post_sendmsg(t, NULL, sendmsg);
1614         if (ret) {
1615                 smb_direct_free_sendmsg(t, sendmsg);
1616                 return ret;
1617         }
1618
1619         wait_event(t->wait_send_pending,
1620                    atomic_read(&t->send_pending) == 0);
1621         return 0;
1622 }
1623
1624 static int smb_direct_accept_client(struct smb_direct_transport *t)
1625 {
1626         struct rdma_conn_param conn_param;
1627         struct ib_port_immutable port_immutable;
1628         u32 ird_ord_hdr[2];
1629         int ret;
1630
1631         memset(&conn_param, 0, sizeof(conn_param));
1632         conn_param.initiator_depth = min_t(u8, t->cm_id->device->attrs.max_qp_rd_atom,
1633                                            SMB_DIRECT_CM_INITIATOR_DEPTH);
1634         conn_param.responder_resources = 0;
1635
1636         t->cm_id->device->ops.get_port_immutable(t->cm_id->device,
1637                                                  t->cm_id->port_num,
1638                                                  &port_immutable);
1639         if (port_immutable.core_cap_flags & RDMA_CORE_PORT_IWARP) {
1640                 ird_ord_hdr[0] = conn_param.responder_resources;
1641                 ird_ord_hdr[1] = 1;
1642                 conn_param.private_data = ird_ord_hdr;
1643                 conn_param.private_data_len = sizeof(ird_ord_hdr);
1644         } else {
1645                 conn_param.private_data = NULL;
1646                 conn_param.private_data_len = 0;
1647         }
1648         conn_param.retry_count = SMB_DIRECT_CM_RETRY;
1649         conn_param.rnr_retry_count = SMB_DIRECT_CM_RNR_RETRY;
1650         conn_param.flow_control = 0;
1651
1652         ret = rdma_accept(t->cm_id, &conn_param);
1653         if (ret) {
1654                 pr_err("error at rdma_accept: %d\n", ret);
1655                 return ret;
1656         }
1657         return 0;
1658 }
1659
1660 static int smb_direct_prepare_negotiation(struct smb_direct_transport *t)
1661 {
1662         int ret;
1663         struct smb_direct_recvmsg *recvmsg;
1664
1665         recvmsg = get_free_recvmsg(t);
1666         if (!recvmsg)
1667                 return -ENOMEM;
1668         recvmsg->type = SMB_DIRECT_MSG_NEGOTIATE_REQ;
1669
1670         ret = smb_direct_post_recv(t, recvmsg);
1671         if (ret) {
1672                 pr_err("Can't post recv: %d\n", ret);
1673                 goto out_err;
1674         }
1675
1676         t->negotiation_requested = false;
1677         ret = smb_direct_accept_client(t);
1678         if (ret) {
1679                 pr_err("Can't accept client\n");
1680                 goto out_err;
1681         }
1682
1683         smb_direct_post_recv_credits(&t->post_recv_credits_work.work);
1684         return 0;
1685 out_err:
1686         put_recvmsg(t, recvmsg);
1687         return ret;
1688 }
1689
1690 static unsigned int smb_direct_get_max_fr_pages(struct smb_direct_transport *t)
1691 {
1692         return min_t(unsigned int,
1693                      t->cm_id->device->attrs.max_fast_reg_page_list_len,
1694                      256);
1695 }
1696
1697 static int smb_direct_init_params(struct smb_direct_transport *t,
1698                                   struct ib_qp_cap *cap)
1699 {
1700         struct ib_device *device = t->cm_id->device;
1701         int max_send_sges, max_rw_wrs, max_send_wrs;
1702         unsigned int max_sge_per_wr, wrs_per_credit;
1703
1704         /* need 3 more sge. because a SMB_DIRECT header, SMB2 header,
1705          * SMB2 response could be mapped.
1706          */
1707         t->max_send_size = smb_direct_max_send_size;
1708         max_send_sges = DIV_ROUND_UP(t->max_send_size, PAGE_SIZE) + 3;
1709         if (max_send_sges > SMB_DIRECT_MAX_SEND_SGES) {
1710                 pr_err("max_send_size %d is too large\n", t->max_send_size);
1711                 return -EINVAL;
1712         }
1713
1714         /* Calculate the number of work requests for RDMA R/W.
1715          * The maximum number of pages which can be registered
1716          * with one Memory region can be transferred with one
1717          * R/W credit. And at least 4 work requests for each credit
1718          * are needed for MR registration, RDMA R/W, local & remote
1719          * MR invalidation.
1720          */
1721         t->max_rdma_rw_size = smb_direct_max_read_write_size;
1722         t->pages_per_rw_credit = smb_direct_get_max_fr_pages(t);
1723         t->max_rw_credits = DIV_ROUND_UP(t->max_rdma_rw_size,
1724                                          (t->pages_per_rw_credit - 1) *
1725                                          PAGE_SIZE);
1726
1727         max_sge_per_wr = min_t(unsigned int, device->attrs.max_send_sge,
1728                                device->attrs.max_sge_rd);
1729         max_sge_per_wr = max_t(unsigned int, max_sge_per_wr,
1730                                max_send_sges);
1731         wrs_per_credit = max_t(unsigned int, 4,
1732                                DIV_ROUND_UP(t->pages_per_rw_credit,
1733                                             max_sge_per_wr) + 1);
1734         max_rw_wrs = t->max_rw_credits * wrs_per_credit;
1735
1736         max_send_wrs = smb_direct_send_credit_target + max_rw_wrs;
1737         if (max_send_wrs > device->attrs.max_cqe ||
1738             max_send_wrs > device->attrs.max_qp_wr) {
1739                 pr_err("consider lowering send_credit_target = %d\n",
1740                        smb_direct_send_credit_target);
1741                 pr_err("Possible CQE overrun, device reporting max_cqe %d max_qp_wr %d\n",
1742                        device->attrs.max_cqe, device->attrs.max_qp_wr);
1743                 return -EINVAL;
1744         }
1745
1746         if (smb_direct_receive_credit_max > device->attrs.max_cqe ||
1747             smb_direct_receive_credit_max > device->attrs.max_qp_wr) {
1748                 pr_err("consider lowering receive_credit_max = %d\n",
1749                        smb_direct_receive_credit_max);
1750                 pr_err("Possible CQE overrun, device reporting max_cpe %d max_qp_wr %d\n",
1751                        device->attrs.max_cqe, device->attrs.max_qp_wr);
1752                 return -EINVAL;
1753         }
1754
1755         if (device->attrs.max_recv_sge < SMB_DIRECT_MAX_RECV_SGES) {
1756                 pr_err("warning: device max_recv_sge = %d too small\n",
1757                        device->attrs.max_recv_sge);
1758                 return -EINVAL;
1759         }
1760
1761         t->recv_credits = 0;
1762         t->count_avail_recvmsg = 0;
1763
1764         t->recv_credit_max = smb_direct_receive_credit_max;
1765         t->recv_credit_target = 10;
1766         t->new_recv_credits = 0;
1767
1768         t->send_credit_target = smb_direct_send_credit_target;
1769         atomic_set(&t->send_credits, 0);
1770         atomic_set(&t->rw_credits, t->max_rw_credits);
1771
1772         t->max_send_size = smb_direct_max_send_size;
1773         t->max_recv_size = smb_direct_max_receive_size;
1774         t->max_fragmented_recv_size = smb_direct_max_fragmented_recv_size;
1775
1776         cap->max_send_wr = max_send_wrs;
1777         cap->max_recv_wr = t->recv_credit_max;
1778         cap->max_send_sge = max_sge_per_wr;
1779         cap->max_recv_sge = SMB_DIRECT_MAX_RECV_SGES;
1780         cap->max_inline_data = 0;
1781         cap->max_rdma_ctxs = t->max_rw_credits;
1782         return 0;
1783 }
1784
1785 static void smb_direct_destroy_pools(struct smb_direct_transport *t)
1786 {
1787         struct smb_direct_recvmsg *recvmsg;
1788
1789         while ((recvmsg = get_free_recvmsg(t)))
1790                 mempool_free(recvmsg, t->recvmsg_mempool);
1791         while ((recvmsg = get_empty_recvmsg(t)))
1792                 mempool_free(recvmsg, t->recvmsg_mempool);
1793
1794         mempool_destroy(t->recvmsg_mempool);
1795         t->recvmsg_mempool = NULL;
1796
1797         kmem_cache_destroy(t->recvmsg_cache);
1798         t->recvmsg_cache = NULL;
1799
1800         mempool_destroy(t->sendmsg_mempool);
1801         t->sendmsg_mempool = NULL;
1802
1803         kmem_cache_destroy(t->sendmsg_cache);
1804         t->sendmsg_cache = NULL;
1805 }
1806
1807 static int smb_direct_create_pools(struct smb_direct_transport *t)
1808 {
1809         char name[80];
1810         int i;
1811         struct smb_direct_recvmsg *recvmsg;
1812
1813         snprintf(name, sizeof(name), "smb_direct_rqst_pool_%p", t);
1814         t->sendmsg_cache = kmem_cache_create(name,
1815                                              sizeof(struct smb_direct_sendmsg) +
1816                                               sizeof(struct smb_direct_negotiate_resp),
1817                                              0, SLAB_HWCACHE_ALIGN, NULL);
1818         if (!t->sendmsg_cache)
1819                 return -ENOMEM;
1820
1821         t->sendmsg_mempool = mempool_create(t->send_credit_target,
1822                                             mempool_alloc_slab, mempool_free_slab,
1823                                             t->sendmsg_cache);
1824         if (!t->sendmsg_mempool)
1825                 goto err;
1826
1827         snprintf(name, sizeof(name), "smb_direct_resp_%p", t);
1828         t->recvmsg_cache = kmem_cache_create(name,
1829                                              sizeof(struct smb_direct_recvmsg) +
1830                                               t->max_recv_size,
1831                                              0, SLAB_HWCACHE_ALIGN, NULL);
1832         if (!t->recvmsg_cache)
1833                 goto err;
1834
1835         t->recvmsg_mempool =
1836                 mempool_create(t->recv_credit_max, mempool_alloc_slab,
1837                                mempool_free_slab, t->recvmsg_cache);
1838         if (!t->recvmsg_mempool)
1839                 goto err;
1840
1841         INIT_LIST_HEAD(&t->recvmsg_queue);
1842
1843         for (i = 0; i < t->recv_credit_max; i++) {
1844                 recvmsg = mempool_alloc(t->recvmsg_mempool, GFP_KERNEL);
1845                 if (!recvmsg)
1846                         goto err;
1847                 recvmsg->transport = t;
1848                 list_add(&recvmsg->list, &t->recvmsg_queue);
1849         }
1850         t->count_avail_recvmsg = t->recv_credit_max;
1851
1852         return 0;
1853 err:
1854         smb_direct_destroy_pools(t);
1855         return -ENOMEM;
1856 }
1857
1858 static int smb_direct_create_qpair(struct smb_direct_transport *t,
1859                                    struct ib_qp_cap *cap)
1860 {
1861         int ret;
1862         struct ib_qp_init_attr qp_attr;
1863         int pages_per_rw;
1864
1865         t->pd = ib_alloc_pd(t->cm_id->device, 0);
1866         if (IS_ERR(t->pd)) {
1867                 pr_err("Can't create RDMA PD\n");
1868                 ret = PTR_ERR(t->pd);
1869                 t->pd = NULL;
1870                 return ret;
1871         }
1872
1873         t->send_cq = ib_alloc_cq(t->cm_id->device, t,
1874                                  smb_direct_send_credit_target + cap->max_rdma_ctxs,
1875                                  0, IB_POLL_WORKQUEUE);
1876         if (IS_ERR(t->send_cq)) {
1877                 pr_err("Can't create RDMA send CQ\n");
1878                 ret = PTR_ERR(t->send_cq);
1879                 t->send_cq = NULL;
1880                 goto err;
1881         }
1882
1883         t->recv_cq = ib_alloc_cq(t->cm_id->device, t,
1884                                  t->recv_credit_max, 0, IB_POLL_WORKQUEUE);
1885         if (IS_ERR(t->recv_cq)) {
1886                 pr_err("Can't create RDMA recv CQ\n");
1887                 ret = PTR_ERR(t->recv_cq);
1888                 t->recv_cq = NULL;
1889                 goto err;
1890         }
1891
1892         memset(&qp_attr, 0, sizeof(qp_attr));
1893         qp_attr.event_handler = smb_direct_qpair_handler;
1894         qp_attr.qp_context = t;
1895         qp_attr.cap = *cap;
1896         qp_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
1897         qp_attr.qp_type = IB_QPT_RC;
1898         qp_attr.send_cq = t->send_cq;
1899         qp_attr.recv_cq = t->recv_cq;
1900         qp_attr.port_num = ~0;
1901
1902         ret = rdma_create_qp(t->cm_id, t->pd, &qp_attr);
1903         if (ret) {
1904                 pr_err("Can't create RDMA QP: %d\n", ret);
1905                 goto err;
1906         }
1907
1908         t->qp = t->cm_id->qp;
1909         t->cm_id->event_handler = smb_direct_cm_handler;
1910
1911         pages_per_rw = DIV_ROUND_UP(t->max_rdma_rw_size, PAGE_SIZE) + 1;
1912         if (pages_per_rw > t->cm_id->device->attrs.max_sgl_rd) {
1913                 ret = ib_mr_pool_init(t->qp, &t->qp->rdma_mrs,
1914                                       t->max_rw_credits, IB_MR_TYPE_MEM_REG,
1915                                       t->pages_per_rw_credit, 0);
1916                 if (ret) {
1917                         pr_err("failed to init mr pool count %d pages %d\n",
1918                                t->max_rw_credits, t->pages_per_rw_credit);
1919                         goto err;
1920                 }
1921         }
1922
1923         return 0;
1924 err:
1925         if (t->qp) {
1926                 ib_destroy_qp(t->qp);
1927                 t->qp = NULL;
1928         }
1929         if (t->recv_cq) {
1930                 ib_destroy_cq(t->recv_cq);
1931                 t->recv_cq = NULL;
1932         }
1933         if (t->send_cq) {
1934                 ib_destroy_cq(t->send_cq);
1935                 t->send_cq = NULL;
1936         }
1937         if (t->pd) {
1938                 ib_dealloc_pd(t->pd);
1939                 t->pd = NULL;
1940         }
1941         return ret;
1942 }
1943
1944 static int smb_direct_prepare(struct ksmbd_transport *t)
1945 {
1946         struct smb_direct_transport *st = smb_trans_direct_transfort(t);
1947         struct smb_direct_recvmsg *recvmsg;
1948         struct smb_direct_negotiate_req *req;
1949         int ret;
1950
1951         ksmbd_debug(RDMA, "Waiting for SMB_DIRECT negotiate request\n");
1952         ret = wait_event_interruptible_timeout(st->wait_status,
1953                                                st->negotiation_requested ||
1954                                                st->status == SMB_DIRECT_CS_DISCONNECTED,
1955                                                SMB_DIRECT_NEGOTIATE_TIMEOUT * HZ);
1956         if (ret <= 0 || st->status == SMB_DIRECT_CS_DISCONNECTED)
1957                 return ret < 0 ? ret : -ETIMEDOUT;
1958
1959         recvmsg = get_first_reassembly(st);
1960         if (!recvmsg)
1961                 return -ECONNABORTED;
1962
1963         ret = smb_direct_check_recvmsg(recvmsg);
1964         if (ret == -ECONNABORTED)
1965                 goto out;
1966
1967         req = (struct smb_direct_negotiate_req *)recvmsg->packet;
1968         st->max_recv_size = min_t(int, st->max_recv_size,
1969                                   le32_to_cpu(req->preferred_send_size));
1970         st->max_send_size = min_t(int, st->max_send_size,
1971                                   le32_to_cpu(req->max_receive_size));
1972         st->max_fragmented_send_size =
1973                 le32_to_cpu(req->max_fragmented_size);
1974         st->max_fragmented_recv_size =
1975                 (st->recv_credit_max * st->max_recv_size) / 2;
1976
1977         ret = smb_direct_send_negotiate_response(st, ret);
1978 out:
1979         spin_lock_irq(&st->reassembly_queue_lock);
1980         st->reassembly_queue_length--;
1981         list_del(&recvmsg->list);
1982         spin_unlock_irq(&st->reassembly_queue_lock);
1983         put_recvmsg(st, recvmsg);
1984
1985         return ret;
1986 }
1987
1988 static int smb_direct_connect(struct smb_direct_transport *st)
1989 {
1990         int ret;
1991         struct ib_qp_cap qp_cap;
1992
1993         ret = smb_direct_init_params(st, &qp_cap);
1994         if (ret) {
1995                 pr_err("Can't configure RDMA parameters\n");
1996                 return ret;
1997         }
1998
1999         ret = smb_direct_create_pools(st);
2000         if (ret) {
2001                 pr_err("Can't init RDMA pool: %d\n", ret);
2002                 return ret;
2003         }
2004
2005         ret = smb_direct_create_qpair(st, &qp_cap);
2006         if (ret) {
2007                 pr_err("Can't accept RDMA client: %d\n", ret);
2008                 return ret;
2009         }
2010
2011         ret = smb_direct_prepare_negotiation(st);
2012         if (ret) {
2013                 pr_err("Can't negotiate: %d\n", ret);
2014                 return ret;
2015         }
2016         return 0;
2017 }
2018
2019 static bool rdma_frwr_is_supported(struct ib_device_attr *attrs)
2020 {
2021         if (!(attrs->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS))
2022                 return false;
2023         if (attrs->max_fast_reg_page_list_len == 0)
2024                 return false;
2025         return true;
2026 }
2027
2028 static int smb_direct_handle_connect_request(struct rdma_cm_id *new_cm_id)
2029 {
2030         struct smb_direct_transport *t;
2031         int ret;
2032
2033         if (!rdma_frwr_is_supported(&new_cm_id->device->attrs)) {
2034                 ksmbd_debug(RDMA,
2035                             "Fast Registration Work Requests is not supported. device capabilities=%llx\n",
2036                             new_cm_id->device->attrs.device_cap_flags);
2037                 return -EPROTONOSUPPORT;
2038         }
2039
2040         t = alloc_transport(new_cm_id);
2041         if (!t)
2042                 return -ENOMEM;
2043
2044         ret = smb_direct_connect(t);
2045         if (ret)
2046                 goto out_err;
2047
2048         KSMBD_TRANS(t)->handler = kthread_run(ksmbd_conn_handler_loop,
2049                                               KSMBD_TRANS(t)->conn, "ksmbd:r%u",
2050                                               smb_direct_port);
2051         if (IS_ERR(KSMBD_TRANS(t)->handler)) {
2052                 ret = PTR_ERR(KSMBD_TRANS(t)->handler);
2053                 pr_err("Can't start thread\n");
2054                 goto out_err;
2055         }
2056
2057         return 0;
2058 out_err:
2059         free_transport(t);
2060         return ret;
2061 }
2062
2063 static int smb_direct_listen_handler(struct rdma_cm_id *cm_id,
2064                                      struct rdma_cm_event *event)
2065 {
2066         switch (event->event) {
2067         case RDMA_CM_EVENT_CONNECT_REQUEST: {
2068                 int ret = smb_direct_handle_connect_request(cm_id);
2069
2070                 if (ret) {
2071                         pr_err("Can't create transport: %d\n", ret);
2072                         return ret;
2073                 }
2074
2075                 ksmbd_debug(RDMA, "Received connection request. cm_id=%p\n",
2076                             cm_id);
2077                 break;
2078         }
2079         default:
2080                 pr_err("Unexpected listen event. cm_id=%p, event=%s (%d)\n",
2081                        cm_id, rdma_event_msg(event->event), event->event);
2082                 break;
2083         }
2084         return 0;
2085 }
2086
2087 static int smb_direct_listen(int port)
2088 {
2089         int ret;
2090         struct rdma_cm_id *cm_id;
2091         struct sockaddr_in sin = {
2092                 .sin_family             = AF_INET,
2093                 .sin_addr.s_addr        = htonl(INADDR_ANY),
2094                 .sin_port               = htons(port),
2095         };
2096
2097         cm_id = rdma_create_id(&init_net, smb_direct_listen_handler,
2098                                &smb_direct_listener, RDMA_PS_TCP, IB_QPT_RC);
2099         if (IS_ERR(cm_id)) {
2100                 pr_err("Can't create cm id: %ld\n", PTR_ERR(cm_id));
2101                 return PTR_ERR(cm_id);
2102         }
2103
2104         ret = rdma_bind_addr(cm_id, (struct sockaddr *)&sin);
2105         if (ret) {
2106                 pr_err("Can't bind: %d\n", ret);
2107                 goto err;
2108         }
2109
2110         smb_direct_listener.cm_id = cm_id;
2111
2112         ret = rdma_listen(cm_id, 10);
2113         if (ret) {
2114                 pr_err("Can't listen: %d\n", ret);
2115                 goto err;
2116         }
2117         return 0;
2118 err:
2119         smb_direct_listener.cm_id = NULL;
2120         rdma_destroy_id(cm_id);
2121         return ret;
2122 }
2123
2124 static int smb_direct_ib_client_add(struct ib_device *ib_dev)
2125 {
2126         struct smb_direct_device *smb_dev;
2127
2128         /* Set 5445 port if device type is iWARP(No IB) */
2129         if (ib_dev->node_type != RDMA_NODE_IB_CA)
2130                 smb_direct_port = SMB_DIRECT_PORT_IWARP;
2131
2132         if (!ib_dev->ops.get_netdev ||
2133             !rdma_frwr_is_supported(&ib_dev->attrs))
2134                 return 0;
2135
2136         smb_dev = kzalloc(sizeof(*smb_dev), GFP_KERNEL);
2137         if (!smb_dev)
2138                 return -ENOMEM;
2139         smb_dev->ib_dev = ib_dev;
2140
2141         write_lock(&smb_direct_device_lock);
2142         list_add(&smb_dev->list, &smb_direct_device_list);
2143         write_unlock(&smb_direct_device_lock);
2144
2145         ksmbd_debug(RDMA, "ib device added: name %s\n", ib_dev->name);
2146         return 0;
2147 }
2148
2149 static void smb_direct_ib_client_remove(struct ib_device *ib_dev,
2150                                         void *client_data)
2151 {
2152         struct smb_direct_device *smb_dev, *tmp;
2153
2154         write_lock(&smb_direct_device_lock);
2155         list_for_each_entry_safe(smb_dev, tmp, &smb_direct_device_list, list) {
2156                 if (smb_dev->ib_dev == ib_dev) {
2157                         list_del(&smb_dev->list);
2158                         kfree(smb_dev);
2159                         break;
2160                 }
2161         }
2162         write_unlock(&smb_direct_device_lock);
2163 }
2164
2165 static struct ib_client smb_direct_ib_client = {
2166         .name   = "ksmbd_smb_direct_ib",
2167         .add    = smb_direct_ib_client_add,
2168         .remove = smb_direct_ib_client_remove,
2169 };
2170
2171 int ksmbd_rdma_init(void)
2172 {
2173         int ret;
2174
2175         smb_direct_listener.cm_id = NULL;
2176
2177         ret = ib_register_client(&smb_direct_ib_client);
2178         if (ret) {
2179                 pr_err("failed to ib_register_client\n");
2180                 return ret;
2181         }
2182
2183         /* When a client is running out of send credits, the credits are
2184          * granted by the server's sending a packet using this queue.
2185          * This avoids the situation that a clients cannot send packets
2186          * for lack of credits
2187          */
2188         smb_direct_wq = alloc_workqueue("ksmbd-smb_direct-wq",
2189                                         WQ_HIGHPRI | WQ_MEM_RECLAIM, 0);
2190         if (!smb_direct_wq)
2191                 return -ENOMEM;
2192
2193         ret = smb_direct_listen(smb_direct_port);
2194         if (ret) {
2195                 destroy_workqueue(smb_direct_wq);
2196                 smb_direct_wq = NULL;
2197                 pr_err("Can't listen: %d\n", ret);
2198                 return ret;
2199         }
2200
2201         ksmbd_debug(RDMA, "init RDMA listener. cm_id=%p\n",
2202                     smb_direct_listener.cm_id);
2203         return 0;
2204 }
2205
2206 void ksmbd_rdma_destroy(void)
2207 {
2208         if (!smb_direct_listener.cm_id)
2209                 return;
2210
2211         ib_unregister_client(&smb_direct_ib_client);
2212         rdma_destroy_id(smb_direct_listener.cm_id);
2213
2214         smb_direct_listener.cm_id = NULL;
2215
2216         if (smb_direct_wq) {
2217                 destroy_workqueue(smb_direct_wq);
2218                 smb_direct_wq = NULL;
2219         }
2220 }
2221
2222 bool ksmbd_rdma_capable_netdev(struct net_device *netdev)
2223 {
2224         struct smb_direct_device *smb_dev;
2225         int i;
2226         bool rdma_capable = false;
2227
2228         read_lock(&smb_direct_device_lock);
2229         list_for_each_entry(smb_dev, &smb_direct_device_list, list) {
2230                 for (i = 0; i < smb_dev->ib_dev->phys_port_cnt; i++) {
2231                         struct net_device *ndev;
2232
2233                         ndev = smb_dev->ib_dev->ops.get_netdev(smb_dev->ib_dev,
2234                                                                i + 1);
2235                         if (!ndev)
2236                                 continue;
2237
2238                         if (ndev == netdev) {
2239                                 dev_put(ndev);
2240                                 rdma_capable = true;
2241                                 goto out;
2242                         }
2243                         dev_put(ndev);
2244                 }
2245         }
2246 out:
2247         read_unlock(&smb_direct_device_lock);
2248
2249         if (rdma_capable == false) {
2250                 struct ib_device *ibdev;
2251
2252                 ibdev = ib_device_get_by_netdev(netdev, RDMA_DRIVER_UNKNOWN);
2253                 if (ibdev) {
2254                         if (rdma_frwr_is_supported(&ibdev->attrs))
2255                                 rdma_capable = true;
2256                         ib_device_put(ibdev);
2257                 }
2258         }
2259
2260         return rdma_capable;
2261 }
2262
2263 static struct ksmbd_transport_ops ksmbd_smb_direct_transport_ops = {
2264         .prepare        = smb_direct_prepare,
2265         .disconnect     = smb_direct_disconnect,
2266         .shutdown       = smb_direct_shutdown,
2267         .writev         = smb_direct_writev,
2268         .read           = smb_direct_read,
2269         .rdma_read      = smb_direct_rdma_read,
2270         .rdma_write     = smb_direct_rdma_write,
2271 };