include/rdma/rdmavt_qp.h

   1 #ifndef DEF_RDMAVT_INCQP_H
   2 #define DEF_RDMAVT_INCQP_H
   3
   4 /*
   5  * Copyright(c) 2016 - 2018 Intel Corporation.
   6  *
   7  * This file is provided under a dual BSD/GPLv2 license.  When using or
   8  * redistributing this file, you may do so under either license.
   9  *
  10  * GPL LICENSE SUMMARY
  11  *
  12  * This program is free software; you can redistribute it and/or modify
  13  * it under the terms of version 2 of the GNU General Public License as
  14  * published by the Free Software Foundation.
  15  *
  16  * This program is distributed in the hope that it will be useful, but
  17  * WITHOUT ANY WARRANTY; without even the implied warranty of
  18  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  * General Public License for more details.
  20  *
  21  * BSD LICENSE
  22  *
  23  * Redistribution and use in source and binary forms, with or without
  24  * modification, are permitted provided that the following conditions
  25  * are met:
  26  *
  27  *  - Redistributions of source code must retain the above copyright
  28  *    notice, this list of conditions and the following disclaimer.
  29  *  - Redistributions in binary form must reproduce the above copyright
  30  *    notice, this list of conditions and the following disclaimer in
  31  *    the documentation and/or other materials provided with the
  32  *    distribution.
  33  *  - Neither the name of Intel Corporation nor the names of its
  34  *    contributors may be used to endorse or promote products derived
  35  *    from this software without specific prior written permission.
  36  *
  37  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  38  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  39  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  40  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  41  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  42  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  43  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  44  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  45  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  46  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  47  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  48  *
  49  */
  50
  51 #include <rdma/rdma_vt.h>
  52 #include <rdma/ib_pack.h>
  53 #include <rdma/ib_verbs.h>
  54 #include <rdma/rdmavt_cq.h>
  55 /*
  56  * Atomic bit definitions for r_aflags.
  57  */
  58 #define RVT_R_WRID_VALID        0
  59 #define RVT_R_REWIND_SGE        1
  60
  61 /*
  62  * Bit definitions for r_flags.
  63  */
  64 #define RVT_R_REUSE_SGE 0x01
  65 #define RVT_R_RDMAR_SEQ 0x02
  66 #define RVT_R_RSP_NAK   0x04
  67 #define RVT_R_RSP_SEND  0x08
  68 #define RVT_R_COMM_EST  0x10
  69
  70 /*
  71  * Bit definitions for s_flags.
  72  *
  73  * RVT_S_SIGNAL_REQ_WR - set if QP send WRs contain completion signaled
  74  * RVT_S_BUSY - send tasklet is processing the QP
  75  * RVT_S_TIMER - the RC retry timer is active
  76  * RVT_S_ACK_PENDING - an ACK is waiting to be sent after RDMA read/atomics
  77  * RVT_S_WAIT_FENCE - waiting for all prior RDMA read or atomic SWQEs
  78  *                         before processing the next SWQE
  79  * RVT_S_WAIT_RDMAR - waiting for a RDMA read or atomic SWQE to complete
  80  *                         before processing the next SWQE
  81  * RVT_S_WAIT_RNR - waiting for RNR timeout
  82  * RVT_S_WAIT_SSN_CREDIT - waiting for RC credits to process next SWQE
  83  * RVT_S_WAIT_DMA - waiting for send DMA queue to drain before generating
  84  *                  next send completion entry not via send DMA
  85  * RVT_S_WAIT_PIO - waiting for a send buffer to be available
  86  * RVT_S_WAIT_PIO_DRAIN - waiting for a qp to drain pio packets
  87  * RVT_S_WAIT_TX - waiting for a struct verbs_txreq to be available
  88  * RVT_S_WAIT_DMA_DESC - waiting for DMA descriptors to be available
  89  * RVT_S_WAIT_KMEM - waiting for kernel memory to be available
  90  * RVT_S_WAIT_PSN - waiting for a packet to exit the send DMA queue
  91  * RVT_S_WAIT_ACK - waiting for an ACK packet before sending more requests
  92  * RVT_S_SEND_ONE - send one packet, request ACK, then wait for ACK
  93  * RVT_S_ECN - a BECN was queued to the send engine
  94  * RVT_S_MAX_BIT_MASK - The max bit that can be used by rdmavt
  95  */
  96 #define RVT_S_SIGNAL_REQ_WR     0x0001
  97 #define RVT_S_BUSY              0x0002
  98 #define RVT_S_TIMER             0x0004
  99 #define RVT_S_RESP_PENDING      0x0008
 100 #define RVT_S_ACK_PENDING       0x0010
 101 #define RVT_S_WAIT_FENCE        0x0020
 102 #define RVT_S_WAIT_RDMAR        0x0040
 103 #define RVT_S_WAIT_RNR          0x0080
 104 #define RVT_S_WAIT_SSN_CREDIT   0x0100
 105 #define RVT_S_WAIT_DMA          0x0200
 106 #define RVT_S_WAIT_PIO          0x0400
 107 #define RVT_S_WAIT_TX           0x0800
 108 #define RVT_S_WAIT_DMA_DESC     0x1000
 109 #define RVT_S_WAIT_KMEM         0x2000
 110 #define RVT_S_WAIT_PSN          0x4000
 111 #define RVT_S_WAIT_ACK          0x8000
 112 #define RVT_S_SEND_ONE          0x10000
 113 #define RVT_S_UNLIMITED_CREDIT  0x20000
 114 #define RVT_S_ECN               0x40000
 115 #define RVT_S_MAX_BIT_MASK      0x800000
 116
 117 /*
 118  * Drivers should use s_flags starting with bit 31 down to the bit next to
 119  * RVT_S_MAX_BIT_MASK
 120  */
 121
 122 /*
 123  * Wait flags that would prevent any packet type from being sent.
 124  */
 125 #define RVT_S_ANY_WAIT_IO \
 126         (RVT_S_WAIT_PIO | RVT_S_WAIT_TX | \
 127          RVT_S_WAIT_DMA_DESC | RVT_S_WAIT_KMEM)
 128
 129 /*
 130  * Wait flags that would prevent send work requests from making progress.
 131  */
 132 #define RVT_S_ANY_WAIT_SEND (RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR | \
 133         RVT_S_WAIT_RNR | RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_DMA | \
 134         RVT_S_WAIT_PSN | RVT_S_WAIT_ACK)
 135
 136 #define RVT_S_ANY_WAIT (RVT_S_ANY_WAIT_IO | RVT_S_ANY_WAIT_SEND)
 137
 138 /* Number of bits to pay attention to in the opcode for checking qp type */
 139 #define RVT_OPCODE_QP_MASK 0xE0
 140
 141 /* Flags for checking QP state (see ib_rvt_state_ops[]) */
 142 #define RVT_POST_SEND_OK                0x01
 143 #define RVT_POST_RECV_OK                0x02
 144 #define RVT_PROCESS_RECV_OK             0x04
 145 #define RVT_PROCESS_SEND_OK             0x08
 146 #define RVT_PROCESS_NEXT_SEND_OK        0x10
 147 #define RVT_FLUSH_SEND                  0x20
 148 #define RVT_FLUSH_RECV                  0x40
 149 #define RVT_PROCESS_OR_FLUSH_SEND \
 150         (RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND)
 151 #define RVT_SEND_OR_FLUSH_OR_RECV_OK \
 152         (RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND | RVT_PROCESS_RECV_OK)
 153
 154 /*
 155  * Internal send flags
 156  */
 157 #define RVT_SEND_RESERVE_USED           IB_SEND_RESERVED_START
 158 #define RVT_SEND_COMPLETION_ONLY        (IB_SEND_RESERVED_START << 1)
 159
 160 /*
 161  * Send work request queue entry.
 162  * The size of the sg_list is determined when the QP is created and stored
 163  * in qp->s_max_sge.
 164  */
 165 struct rvt_swqe {
 166         union {
 167                 struct ib_send_wr wr;   /* don't use wr.sg_list */
 168                 struct ib_ud_wr ud_wr;
 169                 struct ib_reg_wr reg_wr;
 170                 struct ib_rdma_wr rdma_wr;
 171                 struct ib_atomic_wr atomic_wr;
 172         };
 173         u32 psn;                /* first packet sequence number */
 174         u32 lpsn;               /* last packet sequence number */
 175         u32 ssn;                /* send sequence number */
 176         u32 length;             /* total length of data in sg_list */
 177         struct rvt_sge sg_list[0];
 178 };
 179
 180 /*
 181  * Receive work request queue entry.
 182  * The size of the sg_list is determined when the QP (or SRQ) is created
 183  * and stored in qp->r_rq.max_sge (or srq->rq.max_sge).
 184  */
 185 struct rvt_rwqe {
 186         u64 wr_id;
 187         u8 num_sge;
 188         struct ib_sge sg_list[0];
 189 };
 190
 191 /*
 192  * This structure is used to contain the head pointer, tail pointer,
 193  * and receive work queue entries as a single memory allocation so
 194  * it can be mmap'ed into user space.
 195  * Note that the wq array elements are variable size so you can't
 196  * just index into the array to get the N'th element;
 197  * use get_rwqe_ptr() instead.
 198  */
 199 struct rvt_rwq {
 200         u32 head;               /* new work requests posted to the head */
 201         u32 tail;               /* receives pull requests from here. */
 202         struct rvt_rwqe wq[0];
 203 };
 204
 205 struct rvt_rq {
 206         struct rvt_rwq *wq;
 207         u32 size;               /* size of RWQE array */
 208         u8 max_sge;
 209         /* protect changes in this struct */
 210         spinlock_t lock ____cacheline_aligned_in_smp;
 211 };
 212
 213 /*
 214  * This structure is used by rvt_mmap() to validate an offset
 215  * when an mmap() request is made.  The vm_area_struct then uses
 216  * this as its vm_private_data.
 217  */
 218 struct rvt_mmap_info {
 219         struct list_head pending_mmaps;
 220         struct ib_ucontext *context;
 221         void *obj;
 222         __u64 offset;
 223         struct kref ref;
 224         unsigned size;
 225 };
 226
 227 /*
 228  * This structure holds the information that the send tasklet needs
 229  * to send a RDMA read response or atomic operation.
 230  */
 231 struct rvt_ack_entry {
 232         struct rvt_sge rdma_sge;
 233         u64 atomic_data;
 234         u32 psn;
 235         u32 lpsn;
 236         u8 opcode;
 237         u8 sent;
 238 };
 239
 240 #define RC_QP_SCALING_INTERVAL  5
 241
 242 #define RVT_OPERATION_PRIV        0x00000001
 243 #define RVT_OPERATION_ATOMIC      0x00000002
 244 #define RVT_OPERATION_ATOMIC_SGE  0x00000004
 245 #define RVT_OPERATION_LOCAL       0x00000008
 246 #define RVT_OPERATION_USE_RESERVE 0x00000010
 247
 248 #define RVT_OPERATION_MAX (IB_WR_RESERVED10 + 1)
 249
 250 /**
 251  * rvt_operation_params - op table entry
 252  * @length - the length to copy into the swqe entry
 253  * @qpt_support - a bit mask indicating QP type support
 254  * @flags - RVT_OPERATION flags (see above)
 255  *
 256  * This supports table driven post send so that
 257  * the driver can have differing an potentially
 258  * different sets of operations.
 259  *
 260  **/
 261
 262 struct rvt_operation_params {
 263         size_t length;
 264         u32 qpt_support;
 265         u32 flags;
 266 };
 267
 268 /*
 269  * Common variables are protected by both r_rq.lock and s_lock in that order
 270  * which only happens in modify_qp() or changing the QP 'state'.
 271  */
 272 struct rvt_qp {
 273         struct ib_qp ibqp;
 274         void *priv; /* Driver private data */
 275         /* read mostly fields above and below */
 276         struct rdma_ah_attr remote_ah_attr;
 277         struct rdma_ah_attr alt_ah_attr;
 278         struct rvt_qp __rcu *next;           /* link list for QPN hash table */
 279         struct rvt_swqe *s_wq;  /* send work queue */
 280         struct rvt_mmap_info *ip;
 281
 282         unsigned long timeout_jiffies;  /* computed from timeout */
 283
 284         int srate_mbps;         /* s_srate (below) converted to Mbit/s */
 285         pid_t pid;              /* pid for user mode QPs */
 286         u32 remote_qpn;
 287         u32 qkey;               /* QKEY for this QP (for UD or RD) */
 288         u32 s_size;             /* send work queue size */
 289
 290         u16 pmtu;               /* decoded from path_mtu */
 291         u8 log_pmtu;            /* shift for pmtu */
 292         u8 state;               /* QP state */
 293         u8 allowed_ops;         /* high order bits of allowed opcodes */
 294         u8 qp_access_flags;
 295         u8 alt_timeout;         /* Alternate path timeout for this QP */
 296         u8 timeout;             /* Timeout for this QP */
 297         u8 s_srate;
 298         u8 s_mig_state;
 299         u8 port_num;
 300         u8 s_pkey_index;        /* PKEY index to use */
 301         u8 s_alt_pkey_index;    /* Alternate path PKEY index to use */
 302         u8 r_max_rd_atomic;     /* max number of RDMA read/atomic to receive */
 303         u8 s_max_rd_atomic;     /* max number of RDMA read/atomic to send */
 304         u8 s_retry_cnt;         /* number of times to retry */
 305         u8 s_rnr_retry_cnt;
 306         u8 r_min_rnr_timer;     /* retry timeout value for RNR NAKs */
 307         u8 s_max_sge;           /* size of s_wq->sg_list */
 308         u8 s_draining;
 309
 310         /* start of read/write fields */
 311         atomic_t refcount ____cacheline_aligned_in_smp;
 312         wait_queue_head_t wait;
 313
 314         struct rvt_ack_entry *s_ack_queue;
 315         struct rvt_sge_state s_rdma_read_sge;
 316
 317         spinlock_t r_lock ____cacheline_aligned_in_smp;      /* used for APM */
 318         u32 r_psn;              /* expected rcv packet sequence number */
 319         unsigned long r_aflags;
 320         u64 r_wr_id;            /* ID for current receive WQE */
 321         u32 r_ack_psn;          /* PSN for next ACK or atomic ACK */
 322         u32 r_len;              /* total length of r_sge */
 323         u32 r_rcv_len;          /* receive data len processed */
 324         u32 r_msn;              /* message sequence number */
 325
 326         u8 r_state;             /* opcode of last packet received */
 327         u8 r_flags;
 328         u8 r_head_ack_queue;    /* index into s_ack_queue[] */
 329         u8 r_adefered;          /* defered ack count */
 330
 331         struct list_head rspwait;       /* link for waiting to respond */
 332
 333         struct rvt_sge_state r_sge;     /* current receive data */
 334         struct rvt_rq r_rq;             /* receive work queue */
 335
 336         /* post send line */
 337         spinlock_t s_hlock ____cacheline_aligned_in_smp;
 338         u32 s_head;             /* new entries added here */
 339         u32 s_next_psn;         /* PSN for next request */
 340         u32 s_avail;            /* number of entries avail */
 341         u32 s_ssn;              /* SSN of tail entry */
 342         atomic_t s_reserved_used; /* reserved entries in use */
 343
 344         spinlock_t s_lock ____cacheline_aligned_in_smp;
 345         u32 s_flags;
 346         struct rvt_sge_state *s_cur_sge;
 347         struct rvt_swqe *s_wqe;
 348         struct rvt_sge_state s_sge;     /* current send request data */
 349         struct rvt_mregion *s_rdma_mr;
 350         u32 s_len;              /* total length of s_sge */
 351         u32 s_rdma_read_len;    /* total length of s_rdma_read_sge */
 352         u32 s_last_psn;         /* last response PSN processed */
 353         u32 s_sending_psn;      /* lowest PSN that is being sent */
 354         u32 s_sending_hpsn;     /* highest PSN that is being sent */
 355         u32 s_psn;              /* current packet sequence number */
 356         u32 s_ack_rdma_psn;     /* PSN for sending RDMA read responses */
 357         u32 s_ack_psn;          /* PSN for acking sends and RDMA writes */
 358         u32 s_tail;             /* next entry to process */
 359         u32 s_cur;              /* current work queue entry */
 360         u32 s_acked;            /* last un-ACK'ed entry */
 361         u32 s_last;             /* last completed entry */
 362         u32 s_lsn;              /* limit sequence number (credit) */
 363         u32 s_ahgpsn;           /* set to the psn in the copy of the header */
 364         u16 s_cur_size;         /* size of send packet in bytes */
 365         u16 s_rdma_ack_cnt;
 366         u8 s_hdrwords;         /* size of s_hdr in 32 bit words */
 367         s8 s_ahgidx;
 368         u8 s_state;             /* opcode of last packet sent */
 369         u8 s_ack_state;         /* opcode of packet to ACK */
 370         u8 s_nak_state;         /* non-zero if NAK is pending */
 371         u8 r_nak_state;         /* non-zero if NAK is pending */
 372         u8 s_retry;             /* requester retry counter */
 373         u8 s_rnr_retry;         /* requester RNR retry counter */
 374         u8 s_num_rd_atomic;     /* number of RDMA read/atomic pending */
 375         u8 s_tail_ack_queue;    /* index into s_ack_queue[] */
 376
 377         struct rvt_sge_state s_ack_rdma_sge;
 378         struct timer_list s_timer;
 379         struct hrtimer s_rnr_timer;
 380
 381         atomic_t local_ops_pending; /* number of fast_reg/local_inv reqs */
 382
 383         /*
 384          * This sge list MUST be last. Do not add anything below here.
 385          */
 386         struct rvt_sge r_sg_list[0] /* verified SGEs */
 387                 ____cacheline_aligned_in_smp;
 388 };
 389
 390 struct rvt_srq {
 391         struct ib_srq ibsrq;
 392         struct rvt_rq rq;
 393         struct rvt_mmap_info *ip;
 394         /* send signal when number of RWQEs < limit */
 395         u32 limit;
 396 };
 397
 398 #define RVT_QPN_MAX                 BIT(24)
 399 #define RVT_QPNMAP_ENTRIES          (RVT_QPN_MAX / PAGE_SIZE / BITS_PER_BYTE)
 400 #define RVT_BITS_PER_PAGE           (PAGE_SIZE * BITS_PER_BYTE)
 401 #define RVT_BITS_PER_PAGE_MASK      (RVT_BITS_PER_PAGE - 1)
 402 #define RVT_QPN_MASK                IB_QPN_MASK
 403
 404 /*
 405  * QPN-map pages start out as NULL, they get allocated upon
 406  * first use and are never deallocated. This way,
 407  * large bitmaps are not allocated unless large numbers of QPs are used.
 408  */
 409 struct rvt_qpn_map {
 410         void *page;
 411 };
 412
 413 struct rvt_qpn_table {
 414         spinlock_t lock; /* protect changes to the qp table */
 415         unsigned flags;         /* flags for QP0/1 allocated for each port */
 416         u32 last;               /* last QP number allocated */
 417         u32 nmaps;              /* size of the map table */
 418         u16 limit;
 419         u8  incr;
 420         /* bit map of free QP numbers other than 0/1 */
 421         struct rvt_qpn_map map[RVT_QPNMAP_ENTRIES];
 422 };
 423
 424 struct rvt_qp_ibdev {
 425         u32 qp_table_size;
 426         u32 qp_table_bits;
 427         struct rvt_qp __rcu **qp_table;
 428         spinlock_t qpt_lock; /* qptable lock */
 429         struct rvt_qpn_table qpn_table;
 430 };
 431
 432 /*
 433  * There is one struct rvt_mcast for each multicast GID.
 434  * All attached QPs are then stored as a list of
 435  * struct rvt_mcast_qp.
 436  */
 437 struct rvt_mcast_qp {
 438         struct list_head list;
 439         struct rvt_qp *qp;
 440 };
 441
 442 struct rvt_mcast_addr {
 443         union ib_gid mgid;
 444         u16 lid;
 445 };
 446
 447 struct rvt_mcast {
 448         struct rb_node rb_node;
 449         struct rvt_mcast_addr mcast_addr;
 450         struct list_head qp_list;
 451         wait_queue_head_t wait;
 452         atomic_t refcount;
 453         int n_attached;
 454 };
 455
 456 /*
 457  * Since struct rvt_swqe is not a fixed size, we can't simply index into
 458  * struct rvt_qp.s_wq.  This function does the array index computation.
 459  */
 460 static inline struct rvt_swqe *rvt_get_swqe_ptr(struct rvt_qp *qp,
 461                                                 unsigned n)
 462 {
 463         return (struct rvt_swqe *)((char *)qp->s_wq +
 464                                      (sizeof(struct rvt_swqe) +
 465                                       qp->s_max_sge *
 466                                       sizeof(struct rvt_sge)) * n);
 467 }
 468
 469 /*
 470  * Since struct rvt_rwqe is not a fixed size, we can't simply index into
 471  * struct rvt_rwq.wq.  This function does the array index computation.
 472  */
 473 static inline struct rvt_rwqe *rvt_get_rwqe_ptr(struct rvt_rq *rq, unsigned n)
 474 {
 475         return (struct rvt_rwqe *)
 476                 ((char *)rq->wq->wq +
 477                  (sizeof(struct rvt_rwqe) +
 478                   rq->max_sge * sizeof(struct ib_sge)) * n);
 479 }
 480
 481 /**
 482  * rvt_is_user_qp - return if this is user mode QP
 483  * @qp - the target QP
 484  */
 485 static inline bool rvt_is_user_qp(struct rvt_qp *qp)
 486 {
 487         return !!qp->pid;
 488 }
 489
 490 /**
 491  * rvt_get_qp - get a QP reference
 492  * @qp - the QP to hold
 493  */
 494 static inline void rvt_get_qp(struct rvt_qp *qp)
 495 {
 496         atomic_inc(&qp->refcount);
 497 }
 498
 499 /**
 500  * rvt_put_qp - release a QP reference
 501  * @qp - the QP to release
 502  */
 503 static inline void rvt_put_qp(struct rvt_qp *qp)
 504 {
 505         if (qp && atomic_dec_and_test(&qp->refcount))
 506                 wake_up(&qp->wait);
 507 }
 508
 509 /**
 510  * rvt_put_swqe - drop mr refs held by swqe
 511  * @wqe - the send wqe
 512  *
 513  * This drops any mr references held by the swqe
 514  */
 515 static inline void rvt_put_swqe(struct rvt_swqe *wqe)
 516 {
 517         int i;
 518
 519         for (i = 0; i < wqe->wr.num_sge; i++) {
 520                 struct rvt_sge *sge = &wqe->sg_list[i];
 521
 522                 rvt_put_mr(sge->mr);
 523         }
 524 }
 525
 526 /**
 527  * rvt_qp_wqe_reserve - reserve operation
 528  * @qp - the rvt qp
 529  * @wqe - the send wqe
 530  *
 531  * This routine used in post send to record
 532  * a wqe relative reserved operation use.
 533  */
 534 static inline void rvt_qp_wqe_reserve(
 535         struct rvt_qp *qp,
 536         struct rvt_swqe *wqe)
 537 {
 538         atomic_inc(&qp->s_reserved_used);
 539 }
 540
 541 /**
 542  * rvt_qp_wqe_unreserve - clean reserved operation
 543  * @qp - the rvt qp
 544  * @wqe - the send wqe
 545  *
 546  * This decrements the reserve use count.
 547  *
 548  * This call MUST precede the change to
 549  * s_last to insure that post send sees a stable
 550  * s_avail.
 551  *
 552  * An smp_mp__after_atomic() is used to insure
 553  * the compiler does not juggle the order of the s_last
 554  * ring index and the decrementing of s_reserved_used.
 555  */
 556 static inline void rvt_qp_wqe_unreserve(
 557         struct rvt_qp *qp,
 558         struct rvt_swqe *wqe)
 559 {
 560         if (unlikely(wqe->wr.send_flags & RVT_SEND_RESERVE_USED)) {
 561                 atomic_dec(&qp->s_reserved_used);
 562                 /* insure no compiler re-order up to s_last change */
 563                 smp_mb__after_atomic();
 564         }
 565 }
 566
 567 extern const enum ib_wc_opcode ib_rvt_wc_opcode[];
 568
 569 /**
 570  * rvt_qp_swqe_complete() - insert send completion
 571  * @qp - the qp
 572  * @wqe - the send wqe
 573  * @status - completion status
 574  *
 575  * Insert a send completion into the completion
 576  * queue if the qp indicates it should be done.
 577  *
 578  * See IBTA 10.7.3.1 for info on completion
 579  * control.
 580  */
 581 static inline void rvt_qp_swqe_complete(
 582         struct rvt_qp *qp,
 583         struct rvt_swqe *wqe,
 584         enum ib_wc_opcode opcode,
 585         enum ib_wc_status status)
 586 {
 587         if (unlikely(wqe->wr.send_flags & RVT_SEND_RESERVE_USED))
 588                 return;
 589         if (!(qp->s_flags & RVT_S_SIGNAL_REQ_WR) ||
 590             (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
 591              status != IB_WC_SUCCESS) {
 592                 struct ib_wc wc;
 593
 594                 memset(&wc, 0, sizeof(wc));
 595                 wc.wr_id = wqe->wr.wr_id;
 596                 wc.status = status;
 597                 wc.opcode = opcode;
 598                 wc.qp = &qp->ibqp;
 599                 wc.byte_len = wqe->length;
 600                 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.send_cq), &wc,
 601                              status != IB_WC_SUCCESS);
 602         }
 603 }
 604
 605 /*
 606  * Compare the lower 24 bits of the msn values.
 607  * Returns an integer <, ==, or > than zero.
 608  */
 609 static inline int rvt_cmp_msn(u32 a, u32 b)
 610 {
 611         return (((int)a) - ((int)b)) << 8;
 612 }
 613
 614 /**
 615  * rvt_compute_aeth - compute the AETH (syndrome + MSN)
 616  * @qp: the queue pair to compute the AETH for
 617  *
 618  * Returns the AETH.
 619  */
 620 __be32 rvt_compute_aeth(struct rvt_qp *qp);
 621
 622 /**
 623  * rvt_get_credit - flush the send work queue of a QP
 624  * @qp: the qp who's send work queue to flush
 625  * @aeth: the Acknowledge Extended Transport Header
 626  *
 627  * The QP s_lock should be held.
 628  */
 629 void rvt_get_credit(struct rvt_qp *qp, u32 aeth);
 630
 631 /**
 632  * @qp - the qp pair
 633  * @len - the length
 634  *
 635  * Perform a shift based mtu round up divide
 636  */
 637 static inline u32 rvt_div_round_up_mtu(struct rvt_qp *qp, u32 len)
 638 {
 639         return (len + qp->pmtu - 1) >> qp->log_pmtu;
 640 }
 641
 642 /**
 643  * @qp - the qp pair
 644  * @len - the length
 645  *
 646  * Perform a shift based mtu divide
 647  */
 648 static inline u32 rvt_div_mtu(struct rvt_qp *qp, u32 len)
 649 {
 650         return len >> qp->log_pmtu;
 651 }
 652
 653 /**
 654  * rvt_timeout_to_jiffies - Convert a ULP timeout input into jiffies
 655  * @timeout - timeout input(0 - 31).
 656  *
 657  * Return a timeout value in jiffies.
 658  */
 659 static inline unsigned long rvt_timeout_to_jiffies(u8 timeout)
 660 {
 661         if (timeout > 31)
 662                 timeout = 31;
 663
 664         return usecs_to_jiffies(1U << timeout) * 4096UL / 1000UL;
 665 }
 666
 667 extern const int  ib_rvt_state_ops[];
 668
 669 struct rvt_dev_info;
 670 int rvt_get_rwqe(struct rvt_qp *qp, bool wr_id_only);
 671 void rvt_comm_est(struct rvt_qp *qp);
 672 int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err);
 673 void rvt_rc_error(struct rvt_qp *qp, enum ib_wc_status err);
 674 unsigned long rvt_rnr_tbl_to_usec(u32 index);
 675 enum hrtimer_restart rvt_rc_rnr_retry(struct hrtimer *t);
 676 void rvt_add_rnr_timer(struct rvt_qp *qp, u32 aeth);
 677 void rvt_del_timers_sync(struct rvt_qp *qp);
 678 void rvt_stop_rc_timers(struct rvt_qp *qp);
 679 void rvt_add_retry_timer(struct rvt_qp *qp);
 680
 681 void rvt_copy_sge(struct rvt_qp *qp, struct rvt_sge_state *ss,
 682                   void *data, u32 length,
 683                   bool release, bool copy_last);
 684 void rvt_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
 685                        enum ib_wc_status status);
 686 void rvt_ruc_loopback(struct rvt_qp *qp);
 687
 688 /**
 689  * struct rvt_qp_iter - the iterator for QPs
 690  * @qp - the current QP
 691  *
 692  * This structure defines the current iterator
 693  * state for sequenced access to all QPs relative
 694  * to an rvt_dev_info.
 695  */
 696 struct rvt_qp_iter {
 697         struct rvt_qp *qp;
 698         /* private: backpointer */
 699         struct rvt_dev_info *rdi;
 700         /* private: callback routine */
 701         void (*cb)(struct rvt_qp *qp, u64 v);
 702         /* private: for arg to callback routine */
 703         u64 v;
 704         /* private: number of SMI,GSI QPs for device */
 705         int specials;
 706         /* private: current iterator index */
 707         int n;
 708 };
 709
 710 struct rvt_qp_iter *rvt_qp_iter_init(struct rvt_dev_info *rdi,
 711                                      u64 v,
 712                                      void (*cb)(struct rvt_qp *qp, u64 v));
 713 int rvt_qp_iter_next(struct rvt_qp_iter *iter);
 714 void rvt_qp_iter(struct rvt_dev_info *rdi,
 715                  u64 v,
 716                  void (*cb)(struct rvt_qp *qp, u64 v));
 717 void rvt_qp_mr_clean(struct rvt_qp *qp, u32 lkey);
 718 #endif          /* DEF_RDMAVT_INCQP_H */