1 #ifndef DEF_RDMAVT_INCQP_H
2 #define DEF_RDMAVT_INCQP_H
5 * Copyright(c) 2016 Intel Corporation.
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.
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.
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.
23 * Redistribution and use in source and binary forms, with or without
24 * modification, are permitted provided that the following conditions
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
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.
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.
51 #include <rdma/rdma_vt.h>
52 #include <rdma/ib_pack.h>
53 #include <rdma/ib_verbs.h>
54 #include <rdma/rdmavt_cq.h>
56 * Atomic bit definitions for r_aflags.
58 #define RVT_R_WRID_VALID 0
59 #define RVT_R_REWIND_SGE 1
62 * Bit definitions for r_flags.
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
71 * Bit definitions for s_flags.
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
95 #define RVT_S_SIGNAL_REQ_WR 0x0001
96 #define RVT_S_BUSY 0x0002
97 #define RVT_S_TIMER 0x0004
98 #define RVT_S_RESP_PENDING 0x0008
99 #define RVT_S_ACK_PENDING 0x0010
100 #define RVT_S_WAIT_FENCE 0x0020
101 #define RVT_S_WAIT_RDMAR 0x0040
102 #define RVT_S_WAIT_RNR 0x0080
103 #define RVT_S_WAIT_SSN_CREDIT 0x0100
104 #define RVT_S_WAIT_DMA 0x0200
105 #define RVT_S_WAIT_PIO 0x0400
106 #define RVT_S_WAIT_PIO_DRAIN 0x0800
107 #define RVT_S_WAIT_TX 0x1000
108 #define RVT_S_WAIT_DMA_DESC 0x2000
109 #define RVT_S_WAIT_KMEM 0x4000
110 #define RVT_S_WAIT_PSN 0x8000
111 #define RVT_S_WAIT_ACK 0x10000
112 #define RVT_S_SEND_ONE 0x20000
113 #define RVT_S_UNLIMITED_CREDIT 0x40000
114 #define RVT_S_AHG_VALID 0x80000
115 #define RVT_S_AHG_CLEAR 0x100000
116 #define RVT_S_ECN 0x200000
119 * Wait flags that would prevent any packet type from being sent.
121 #define RVT_S_ANY_WAIT_IO \
122 (RVT_S_WAIT_PIO | RVT_S_WAIT_PIO_DRAIN | RVT_S_WAIT_TX | \
123 RVT_S_WAIT_DMA_DESC | RVT_S_WAIT_KMEM)
126 * Wait flags that would prevent send work requests from making progress.
128 #define RVT_S_ANY_WAIT_SEND (RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR | \
129 RVT_S_WAIT_RNR | RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_DMA | \
130 RVT_S_WAIT_PSN | RVT_S_WAIT_ACK)
132 #define RVT_S_ANY_WAIT (RVT_S_ANY_WAIT_IO | RVT_S_ANY_WAIT_SEND)
134 /* Number of bits to pay attention to in the opcode for checking qp type */
135 #define RVT_OPCODE_QP_MASK 0xE0
137 /* Flags for checking QP state (see ib_rvt_state_ops[]) */
138 #define RVT_POST_SEND_OK 0x01
139 #define RVT_POST_RECV_OK 0x02
140 #define RVT_PROCESS_RECV_OK 0x04
141 #define RVT_PROCESS_SEND_OK 0x08
142 #define RVT_PROCESS_NEXT_SEND_OK 0x10
143 #define RVT_FLUSH_SEND 0x20
144 #define RVT_FLUSH_RECV 0x40
145 #define RVT_PROCESS_OR_FLUSH_SEND \
146 (RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND)
147 #define RVT_SEND_OR_FLUSH_OR_RECV_OK \
148 (RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND | RVT_PROCESS_RECV_OK)
151 * Internal send flags
153 #define RVT_SEND_RESERVE_USED IB_SEND_RESERVED_START
154 #define RVT_SEND_COMPLETION_ONLY (IB_SEND_RESERVED_START << 1)
157 * Send work request queue entry.
158 * The size of the sg_list is determined when the QP is created and stored
163 struct ib_send_wr wr
; /* don't use wr.sg_list */
164 struct ib_ud_wr ud_wr
;
165 struct ib_reg_wr reg_wr
;
166 struct ib_rdma_wr rdma_wr
;
167 struct ib_atomic_wr atomic_wr
;
169 u32 psn
; /* first packet sequence number */
170 u32 lpsn
; /* last packet sequence number */
171 u32 ssn
; /* send sequence number */
172 u32 length
; /* total length of data in sg_list */
173 struct rvt_sge sg_list
[0];
177 * Receive work request queue entry.
178 * The size of the sg_list is determined when the QP (or SRQ) is created
179 * and stored in qp->r_rq.max_sge (or srq->rq.max_sge).
184 struct ib_sge sg_list
[0];
188 * This structure is used to contain the head pointer, tail pointer,
189 * and receive work queue entries as a single memory allocation so
190 * it can be mmap'ed into user space.
191 * Note that the wq array elements are variable size so you can't
192 * just index into the array to get the N'th element;
193 * use get_rwqe_ptr() instead.
196 u32 head
; /* new work requests posted to the head */
197 u32 tail
; /* receives pull requests from here. */
198 struct rvt_rwqe wq
[0];
203 u32 size
; /* size of RWQE array */
205 /* protect changes in this struct */
206 spinlock_t lock ____cacheline_aligned_in_smp
;
210 * This structure is used by rvt_mmap() to validate an offset
211 * when an mmap() request is made. The vm_area_struct then uses
212 * this as its vm_private_data.
214 struct rvt_mmap_info
{
215 struct list_head pending_mmaps
;
216 struct ib_ucontext
*context
;
224 * This structure holds the information that the send tasklet needs
225 * to send a RDMA read response or atomic operation.
227 struct rvt_ack_entry
{
228 struct rvt_sge rdma_sge
;
236 #define RC_QP_SCALING_INTERVAL 5
238 #define RVT_OPERATION_PRIV 0x00000001
239 #define RVT_OPERATION_ATOMIC 0x00000002
240 #define RVT_OPERATION_ATOMIC_SGE 0x00000004
241 #define RVT_OPERATION_LOCAL 0x00000008
242 #define RVT_OPERATION_USE_RESERVE 0x00000010
244 #define RVT_OPERATION_MAX (IB_WR_RESERVED10 + 1)
247 * rvt_operation_params - op table entry
248 * @length - the length to copy into the swqe entry
249 * @qpt_support - a bit mask indicating QP type support
250 * @flags - RVT_OPERATION flags (see above)
252 * This supports table driven post send so that
253 * the driver can have differing an potentially
254 * different sets of operations.
258 struct rvt_operation_params
{
265 * Common variables are protected by both r_rq.lock and s_lock in that order
266 * which only happens in modify_qp() or changing the QP 'state'.
270 void *priv
; /* Driver private data */
271 /* read mostly fields above and below */
272 struct ib_ah_attr remote_ah_attr
;
273 struct ib_ah_attr alt_ah_attr
;
274 struct rvt_qp __rcu
*next
; /* link list for QPN hash table */
275 struct rvt_swqe
*s_wq
; /* send work queue */
276 struct rvt_mmap_info
*ip
;
278 unsigned long timeout_jiffies
; /* computed from timeout */
280 enum ib_mtu path_mtu
;
281 int srate_mbps
; /* s_srate (below) converted to Mbit/s */
282 pid_t pid
; /* pid for user mode QPs */
284 u32 qkey
; /* QKEY for this QP (for UD or RD) */
285 u32 s_size
; /* send work queue size */
286 u32 s_ahgpsn
; /* set to the psn in the copy of the header */
288 u16 pmtu
; /* decoded from path_mtu */
289 u8 log_pmtu
; /* shift for pmtu */
290 u8 state
; /* QP state */
291 u8 allowed_ops
; /* high order bits of allowed opcodes */
293 u8 alt_timeout
; /* Alternate path timeout for this QP */
294 u8 timeout
; /* Timeout for this QP */
298 u8 s_pkey_index
; /* PKEY index to use */
299 u8 s_alt_pkey_index
; /* Alternate path PKEY index to use */
300 u8 r_max_rd_atomic
; /* max number of RDMA read/atomic to receive */
301 u8 s_max_rd_atomic
; /* max number of RDMA read/atomic to send */
302 u8 s_retry_cnt
; /* number of times to retry */
304 u8 r_min_rnr_timer
; /* retry timeout value for RNR NAKs */
305 u8 s_max_sge
; /* size of s_wq->sg_list */
308 /* start of read/write fields */
309 atomic_t refcount ____cacheline_aligned_in_smp
;
310 wait_queue_head_t wait
;
312 struct rvt_ack_entry
*s_ack_queue
;
313 struct rvt_sge_state s_rdma_read_sge
;
315 spinlock_t r_lock ____cacheline_aligned_in_smp
; /* used for APM */
316 u32 r_psn
; /* expected rcv packet sequence number */
317 unsigned long r_aflags
;
318 u64 r_wr_id
; /* ID for current receive WQE */
319 u32 r_ack_psn
; /* PSN for next ACK or atomic ACK */
320 u32 r_len
; /* total length of r_sge */
321 u32 r_rcv_len
; /* receive data len processed */
322 u32 r_msn
; /* message sequence number */
324 u8 r_state
; /* opcode of last packet received */
326 u8 r_head_ack_queue
; /* index into s_ack_queue[] */
328 struct list_head rspwait
; /* link for waiting to respond */
330 struct rvt_sge_state r_sge
; /* current receive data */
331 struct rvt_rq r_rq
; /* receive work queue */
334 spinlock_t s_hlock ____cacheline_aligned_in_smp
;
335 u32 s_head
; /* new entries added here */
336 u32 s_next_psn
; /* PSN for next request */
337 u32 s_avail
; /* number of entries avail */
338 u32 s_ssn
; /* SSN of tail entry */
339 atomic_t s_reserved_used
; /* reserved entries in use */
341 spinlock_t s_lock ____cacheline_aligned_in_smp
;
343 struct rvt_sge_state
*s_cur_sge
;
344 struct rvt_swqe
*s_wqe
;
345 struct rvt_sge_state s_sge
; /* current send request data */
346 struct rvt_mregion
*s_rdma_mr
;
347 u32 s_cur_size
; /* size of send packet in bytes */
348 u32 s_len
; /* total length of s_sge */
349 u32 s_rdma_read_len
; /* total length of s_rdma_read_sge */
350 u32 s_last_psn
; /* last response PSN processed */
351 u32 s_sending_psn
; /* lowest PSN that is being sent */
352 u32 s_sending_hpsn
; /* highest PSN that is being sent */
353 u32 s_psn
; /* current packet sequence number */
354 u32 s_ack_rdma_psn
; /* PSN for sending RDMA read responses */
355 u32 s_ack_psn
; /* PSN for acking sends and RDMA writes */
356 u32 s_tail
; /* next entry to process */
357 u32 s_cur
; /* current work queue entry */
358 u32 s_acked
; /* last un-ACK'ed entry */
359 u32 s_last
; /* last completed entry */
360 u32 s_lsn
; /* limit sequence number (credit) */
361 u16 s_hdrwords
; /* size of s_hdr in 32 bit words */
364 u8 s_state
; /* opcode of last packet sent */
365 u8 s_ack_state
; /* opcode of packet to ACK */
366 u8 s_nak_state
; /* non-zero if NAK is pending */
367 u8 r_nak_state
; /* non-zero if NAK is pending */
368 u8 s_retry
; /* requester retry counter */
369 u8 s_rnr_retry
; /* requester RNR retry counter */
370 u8 s_num_rd_atomic
; /* number of RDMA read/atomic pending */
371 u8 s_tail_ack_queue
; /* index into s_ack_queue[] */
373 struct rvt_sge_state s_ack_rdma_sge
;
374 struct timer_list s_timer
;
375 struct hrtimer s_rnr_timer
;
377 atomic_t local_ops_pending
; /* number of fast_reg/local_inv reqs */
380 * This sge list MUST be last. Do not add anything below here.
382 struct rvt_sge r_sg_list
[0] /* verified SGEs */
383 ____cacheline_aligned_in_smp
;
389 struct rvt_mmap_info
*ip
;
390 /* send signal when number of RWQEs < limit */
394 #define RVT_QPN_MAX BIT(24)
395 #define RVT_QPNMAP_ENTRIES (RVT_QPN_MAX / PAGE_SIZE / BITS_PER_BYTE)
396 #define RVT_BITS_PER_PAGE (PAGE_SIZE * BITS_PER_BYTE)
397 #define RVT_BITS_PER_PAGE_MASK (RVT_BITS_PER_PAGE - 1)
398 #define RVT_QPN_MASK 0xFFFFFF
401 * QPN-map pages start out as NULL, they get allocated upon
402 * first use and are never deallocated. This way,
403 * large bitmaps are not allocated unless large numbers of QPs are used.
409 struct rvt_qpn_table
{
410 spinlock_t lock
; /* protect changes to the qp table */
411 unsigned flags
; /* flags for QP0/1 allocated for each port */
412 u32 last
; /* last QP number allocated */
413 u32 nmaps
; /* size of the map table */
416 /* bit map of free QP numbers other than 0/1 */
417 struct rvt_qpn_map map
[RVT_QPNMAP_ENTRIES
];
420 struct rvt_qp_ibdev
{
423 struct rvt_qp __rcu
**qp_table
;
424 spinlock_t qpt_lock
; /* qptable lock */
425 struct rvt_qpn_table qpn_table
;
429 * There is one struct rvt_mcast for each multicast GID.
430 * All attached QPs are then stored as a list of
431 * struct rvt_mcast_qp.
433 struct rvt_mcast_qp
{
434 struct list_head list
;
439 struct rb_node rb_node
;
441 struct list_head qp_list
;
442 wait_queue_head_t wait
;
448 * Since struct rvt_swqe is not a fixed size, we can't simply index into
449 * struct rvt_qp.s_wq. This function does the array index computation.
451 static inline struct rvt_swqe
*rvt_get_swqe_ptr(struct rvt_qp
*qp
,
454 return (struct rvt_swqe
*)((char *)qp
->s_wq
+
455 (sizeof(struct rvt_swqe
) +
457 sizeof(struct rvt_sge
)) * n
);
461 * Since struct rvt_rwqe is not a fixed size, we can't simply index into
462 * struct rvt_rwq.wq. This function does the array index computation.
464 static inline struct rvt_rwqe
*rvt_get_rwqe_ptr(struct rvt_rq
*rq
, unsigned n
)
466 return (struct rvt_rwqe
*)
467 ((char *)rq
->wq
->wq
+
468 (sizeof(struct rvt_rwqe
) +
469 rq
->max_sge
* sizeof(struct ib_sge
)) * n
);
473 * rvt_is_user_qp - return if this is user mode QP
474 * @qp - the target QP
476 static inline bool rvt_is_user_qp(struct rvt_qp
*qp
)
482 * rvt_get_qp - get a QP reference
483 * @qp - the QP to hold
485 static inline void rvt_get_qp(struct rvt_qp
*qp
)
487 atomic_inc(&qp
->refcount
);
491 * rvt_put_qp - release a QP reference
492 * @qp - the QP to release
494 static inline void rvt_put_qp(struct rvt_qp
*qp
)
496 if (qp
&& atomic_dec_and_test(&qp
->refcount
))
501 * rvt_put_swqe - drop mr refs held by swqe
502 * @wqe - the send wqe
504 * This drops any mr references held by the swqe
506 static inline void rvt_put_swqe(struct rvt_swqe
*wqe
)
510 for (i
= 0; i
< wqe
->wr
.num_sge
; i
++) {
511 struct rvt_sge
*sge
= &wqe
->sg_list
[i
];
518 * rvt_qp_wqe_reserve - reserve operation
520 * @wqe - the send wqe
522 * This routine used in post send to record
523 * a wqe relative reserved operation use.
525 static inline void rvt_qp_wqe_reserve(
527 struct rvt_swqe
*wqe
)
529 wqe
->wr
.send_flags
|= RVT_SEND_RESERVE_USED
;
530 atomic_inc(&qp
->s_reserved_used
);
534 * rvt_qp_wqe_unreserve - clean reserved operation
536 * @wqe - the send wqe
538 * This decrements the reserve use count.
540 * This call MUST precede the change to
541 * s_last to insure that post send sees a stable
544 * An smp_mp__after_atomic() is used to insure
545 * the compiler does not juggle the order of the s_last
546 * ring index and the decrementing of s_reserved_used.
548 static inline void rvt_qp_wqe_unreserve(
550 struct rvt_swqe
*wqe
)
552 if (unlikely(wqe
->wr
.send_flags
& RVT_SEND_RESERVE_USED
)) {
553 wqe
->wr
.send_flags
&= ~RVT_SEND_RESERVE_USED
;
554 atomic_dec(&qp
->s_reserved_used
);
555 /* insure no compiler re-order up to s_last change */
556 smp_mb__after_atomic();
560 extern const enum ib_wc_opcode ib_rvt_wc_opcode
[];
563 * rvt_qp_swqe_complete() - insert send completion
565 * @wqe - the send wqe
566 * @status - completion status
568 * Insert a send completion into the completion
569 * queue if the qp indicates it should be done.
571 * See IBTA 10.7.3.1 for info on completion
574 static inline void rvt_qp_swqe_complete(
576 struct rvt_swqe
*wqe
,
577 enum ib_wc_status status
)
579 if (unlikely(wqe
->wr
.send_flags
& RVT_SEND_RESERVE_USED
))
581 if (!(qp
->s_flags
& RVT_S_SIGNAL_REQ_WR
) ||
582 (wqe
->wr
.send_flags
& IB_SEND_SIGNALED
) ||
583 status
!= IB_WC_SUCCESS
) {
586 memset(&wc
, 0, sizeof(wc
));
587 wc
.wr_id
= wqe
->wr
.wr_id
;
589 wc
.opcode
= ib_rvt_wc_opcode
[wqe
->wr
.opcode
];
591 wc
.byte_len
= wqe
->length
;
592 rvt_cq_enter(ibcq_to_rvtcq(qp
->ibqp
.send_cq
), &wc
,
593 status
!= IB_WC_SUCCESS
);
598 * Compare the lower 24 bits of the msn values.
599 * Returns an integer <, ==, or > than zero.
601 static inline int rvt_cmp_msn(u32 a
, u32 b
)
603 return (((int)a
) - ((int)b
)) << 8;
607 * rvt_compute_aeth - compute the AETH (syndrome + MSN)
608 * @qp: the queue pair to compute the AETH for
612 __be32
rvt_compute_aeth(struct rvt_qp
*qp
);
615 * rvt_get_credit - flush the send work queue of a QP
616 * @qp: the qp who's send work queue to flush
617 * @aeth: the Acknowledge Extended Transport Header
619 * The QP s_lock should be held.
621 void rvt_get_credit(struct rvt_qp
*qp
, u32 aeth
);
627 * Perform a shift based mtu round up divide
629 static inline u32
rvt_div_round_up_mtu(struct rvt_qp
*qp
, u32 len
)
631 return (len
+ qp
->pmtu
- 1) >> qp
->log_pmtu
;
638 * Perform a shift based mtu divide
640 static inline u32
rvt_div_mtu(struct rvt_qp
*qp
, u32 len
)
642 return len
>> qp
->log_pmtu
;
645 extern const int ib_rvt_state_ops
[];
648 void rvt_comm_est(struct rvt_qp
*qp
);
649 int rvt_error_qp(struct rvt_qp
*qp
, enum ib_wc_status err
);
650 void rvt_rc_error(struct rvt_qp
*qp
, enum ib_wc_status err
);
651 unsigned long rvt_rnr_tbl_to_usec(u32 index
);
652 enum hrtimer_restart
rvt_rc_rnr_retry(struct hrtimer
*t
);
653 void rvt_add_rnr_timer(struct rvt_qp
*qp
, u32 aeth
);
654 void rvt_del_timers_sync(struct rvt_qp
*qp
);
655 void rvt_stop_rc_timers(struct rvt_qp
*qp
);
656 void rvt_add_retry_timer(struct rvt_qp
*qp
);
658 #endif /* DEF_RDMAVT_INCQP_H */