xref: /linux/include/rdma/rdmavt_qp.h (revision 44f57d78)
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_TX - waiting for a struct verbs_txreq to be available
87  * RVT_S_WAIT_DMA_DESC - waiting for DMA descriptors to be available
88  * RVT_S_WAIT_KMEM - waiting for kernel memory to be available
89  * RVT_S_WAIT_PSN - waiting for a packet to exit the send DMA queue
90  * RVT_S_WAIT_ACK - waiting for an ACK packet before sending more requests
91  * RVT_S_SEND_ONE - send one packet, request ACK, then wait for ACK
92  * RVT_S_ECN - a BECN was queued to the send engine
93  * RVT_S_MAX_BIT_MASK - The max bit that can be used by rdmavt
94  */
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_TX		0x0800
107 #define RVT_S_WAIT_DMA_DESC	0x1000
108 #define RVT_S_WAIT_KMEM		0x2000
109 #define RVT_S_WAIT_PSN		0x4000
110 #define RVT_S_WAIT_ACK		0x8000
111 #define RVT_S_SEND_ONE		0x10000
112 #define RVT_S_UNLIMITED_CREDIT	0x20000
113 #define RVT_S_ECN		0x40000
114 #define RVT_S_MAX_BIT_MASK	0x800000
115 
116 /*
117  * Drivers should use s_flags starting with bit 31 down to the bit next to
118  * RVT_S_MAX_BIT_MASK
119  */
120 
121 /*
122  * Wait flags that would prevent any packet type from being sent.
123  */
124 #define RVT_S_ANY_WAIT_IO \
125 	(RVT_S_WAIT_PIO | RVT_S_WAIT_TX | \
126 	 RVT_S_WAIT_DMA_DESC | RVT_S_WAIT_KMEM)
127 
128 /*
129  * Wait flags that would prevent send work requests from making progress.
130  */
131 #define RVT_S_ANY_WAIT_SEND (RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR | \
132 	RVT_S_WAIT_RNR | RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_DMA | \
133 	RVT_S_WAIT_PSN | RVT_S_WAIT_ACK)
134 
135 #define RVT_S_ANY_WAIT (RVT_S_ANY_WAIT_IO | RVT_S_ANY_WAIT_SEND)
136 
137 /* Number of bits to pay attention to in the opcode for checking qp type */
138 #define RVT_OPCODE_QP_MASK 0xE0
139 
140 /* Flags for checking QP state (see ib_rvt_state_ops[]) */
141 #define RVT_POST_SEND_OK                0x01
142 #define RVT_POST_RECV_OK                0x02
143 #define RVT_PROCESS_RECV_OK             0x04
144 #define RVT_PROCESS_SEND_OK             0x08
145 #define RVT_PROCESS_NEXT_SEND_OK        0x10
146 #define RVT_FLUSH_SEND			0x20
147 #define RVT_FLUSH_RECV			0x40
148 #define RVT_PROCESS_OR_FLUSH_SEND \
149 	(RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND)
150 #define RVT_SEND_OR_FLUSH_OR_RECV_OK \
151 	(RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND | RVT_PROCESS_RECV_OK)
152 
153 /*
154  * Internal send flags
155  */
156 #define RVT_SEND_RESERVE_USED           IB_SEND_RESERVED_START
157 #define RVT_SEND_COMPLETION_ONLY	(IB_SEND_RESERVED_START << 1)
158 
159 /*
160  * Send work request queue entry.
161  * The size of the sg_list is determined when the QP is created and stored
162  * in qp->s_max_sge.
163  */
164 struct rvt_swqe {
165 	union {
166 		struct ib_send_wr wr;   /* don't use wr.sg_list */
167 		struct ib_ud_wr ud_wr;
168 		struct ib_reg_wr reg_wr;
169 		struct ib_rdma_wr rdma_wr;
170 		struct ib_atomic_wr atomic_wr;
171 	};
172 	u32 psn;                /* first packet sequence number */
173 	u32 lpsn;               /* last packet sequence number */
174 	u32 ssn;                /* send sequence number */
175 	u32 length;             /* total length of data in sg_list */
176 	void *priv;             /* driver dependent field */
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 holds the information that the send tasklet needs
215  * to send a RDMA read response or atomic operation.
216  */
217 struct rvt_ack_entry {
218 	struct rvt_sge rdma_sge;
219 	u64 atomic_data;
220 	u32 psn;
221 	u32 lpsn;
222 	u8 opcode;
223 	u8 sent;
224 	void *priv;
225 };
226 
227 #define	RC_QP_SCALING_INTERVAL	5
228 
229 #define RVT_OPERATION_PRIV        0x00000001
230 #define RVT_OPERATION_ATOMIC      0x00000002
231 #define RVT_OPERATION_ATOMIC_SGE  0x00000004
232 #define RVT_OPERATION_LOCAL       0x00000008
233 #define RVT_OPERATION_USE_RESERVE 0x00000010
234 #define RVT_OPERATION_IGN_RNR_CNT 0x00000020
235 
236 #define RVT_OPERATION_MAX (IB_WR_RESERVED10 + 1)
237 
238 /**
239  * rvt_operation_params - op table entry
240  * @length - the length to copy into the swqe entry
241  * @qpt_support - a bit mask indicating QP type support
242  * @flags - RVT_OPERATION flags (see above)
243  *
244  * This supports table driven post send so that
245  * the driver can have differing an potentially
246  * different sets of operations.
247  *
248  **/
249 
250 struct rvt_operation_params {
251 	size_t length;
252 	u32 qpt_support;
253 	u32 flags;
254 };
255 
256 /*
257  * Common variables are protected by both r_rq.lock and s_lock in that order
258  * which only happens in modify_qp() or changing the QP 'state'.
259  */
260 struct rvt_qp {
261 	struct ib_qp ibqp;
262 	void *priv; /* Driver private data */
263 	/* read mostly fields above and below */
264 	struct rdma_ah_attr remote_ah_attr;
265 	struct rdma_ah_attr alt_ah_attr;
266 	struct rvt_qp __rcu *next;           /* link list for QPN hash table */
267 	struct rvt_swqe *s_wq;  /* send work queue */
268 	struct rvt_mmap_info *ip;
269 
270 	unsigned long timeout_jiffies;  /* computed from timeout */
271 
272 	int srate_mbps;		/* s_srate (below) converted to Mbit/s */
273 	pid_t pid;		/* pid for user mode QPs */
274 	u32 remote_qpn;
275 	u32 qkey;               /* QKEY for this QP (for UD or RD) */
276 	u32 s_size;             /* send work queue size */
277 
278 	u16 pmtu;		/* decoded from path_mtu */
279 	u8 log_pmtu;		/* shift for pmtu */
280 	u8 state;               /* QP state */
281 	u8 allowed_ops;		/* high order bits of allowed opcodes */
282 	u8 qp_access_flags;
283 	u8 alt_timeout;         /* Alternate path timeout for this QP */
284 	u8 timeout;             /* Timeout for this QP */
285 	u8 s_srate;
286 	u8 s_mig_state;
287 	u8 port_num;
288 	u8 s_pkey_index;        /* PKEY index to use */
289 	u8 s_alt_pkey_index;    /* Alternate path PKEY index to use */
290 	u8 r_max_rd_atomic;     /* max number of RDMA read/atomic to receive */
291 	u8 s_max_rd_atomic;     /* max number of RDMA read/atomic to send */
292 	u8 s_retry_cnt;         /* number of times to retry */
293 	u8 s_rnr_retry_cnt;
294 	u8 r_min_rnr_timer;     /* retry timeout value for RNR NAKs */
295 	u8 s_max_sge;           /* size of s_wq->sg_list */
296 	u8 s_draining;
297 
298 	/* start of read/write fields */
299 	atomic_t refcount ____cacheline_aligned_in_smp;
300 	wait_queue_head_t wait;
301 
302 	struct rvt_ack_entry *s_ack_queue;
303 	struct rvt_sge_state s_rdma_read_sge;
304 
305 	spinlock_t r_lock ____cacheline_aligned_in_smp;      /* used for APM */
306 	u32 r_psn;              /* expected rcv packet sequence number */
307 	unsigned long r_aflags;
308 	u64 r_wr_id;            /* ID for current receive WQE */
309 	u32 r_ack_psn;          /* PSN for next ACK or atomic ACK */
310 	u32 r_len;              /* total length of r_sge */
311 	u32 r_rcv_len;          /* receive data len processed */
312 	u32 r_msn;              /* message sequence number */
313 
314 	u8 r_state;             /* opcode of last packet received */
315 	u8 r_flags;
316 	u8 r_head_ack_queue;    /* index into s_ack_queue[] */
317 	u8 r_adefered;          /* defered ack count */
318 
319 	struct list_head rspwait;       /* link for waiting to respond */
320 
321 	struct rvt_sge_state r_sge;     /* current receive data */
322 	struct rvt_rq r_rq;             /* receive work queue */
323 
324 	/* post send line */
325 	spinlock_t s_hlock ____cacheline_aligned_in_smp;
326 	u32 s_head;             /* new entries added here */
327 	u32 s_next_psn;         /* PSN for next request */
328 	u32 s_avail;            /* number of entries avail */
329 	u32 s_ssn;              /* SSN of tail entry */
330 	atomic_t s_reserved_used; /* reserved entries in use */
331 
332 	spinlock_t s_lock ____cacheline_aligned_in_smp;
333 	u32 s_flags;
334 	struct rvt_sge_state *s_cur_sge;
335 	struct rvt_swqe *s_wqe;
336 	struct rvt_sge_state s_sge;     /* current send request data */
337 	struct rvt_mregion *s_rdma_mr;
338 	u32 s_len;              /* total length of s_sge */
339 	u32 s_rdma_read_len;    /* total length of s_rdma_read_sge */
340 	u32 s_last_psn;         /* last response PSN processed */
341 	u32 s_sending_psn;      /* lowest PSN that is being sent */
342 	u32 s_sending_hpsn;     /* highest PSN that is being sent */
343 	u32 s_psn;              /* current packet sequence number */
344 	u32 s_ack_rdma_psn;     /* PSN for sending RDMA read responses */
345 	u32 s_ack_psn;          /* PSN for acking sends and RDMA writes */
346 	u32 s_tail;             /* next entry to process */
347 	u32 s_cur;              /* current work queue entry */
348 	u32 s_acked;            /* last un-ACK'ed entry */
349 	u32 s_last;             /* last completed entry */
350 	u32 s_lsn;              /* limit sequence number (credit) */
351 	u32 s_ahgpsn;           /* set to the psn in the copy of the header */
352 	u16 s_cur_size;         /* size of send packet in bytes */
353 	u16 s_rdma_ack_cnt;
354 	u8 s_hdrwords;         /* size of s_hdr in 32 bit words */
355 	s8 s_ahgidx;
356 	u8 s_state;             /* opcode of last packet sent */
357 	u8 s_ack_state;         /* opcode of packet to ACK */
358 	u8 s_nak_state;         /* non-zero if NAK is pending */
359 	u8 r_nak_state;         /* non-zero if NAK is pending */
360 	u8 s_retry;             /* requester retry counter */
361 	u8 s_rnr_retry;         /* requester RNR retry counter */
362 	u8 s_num_rd_atomic;     /* number of RDMA read/atomic pending */
363 	u8 s_tail_ack_queue;    /* index into s_ack_queue[] */
364 	u8 s_acked_ack_queue;   /* index into s_ack_queue[] */
365 
366 	struct rvt_sge_state s_ack_rdma_sge;
367 	struct timer_list s_timer;
368 	struct hrtimer s_rnr_timer;
369 
370 	atomic_t local_ops_pending; /* number of fast_reg/local_inv reqs */
371 
372 	/*
373 	 * This sge list MUST be last. Do not add anything below here.
374 	 */
375 	struct rvt_sge r_sg_list[0] /* verified SGEs */
376 		____cacheline_aligned_in_smp;
377 };
378 
379 struct rvt_srq {
380 	struct ib_srq ibsrq;
381 	struct rvt_rq rq;
382 	struct rvt_mmap_info *ip;
383 	/* send signal when number of RWQEs < limit */
384 	u32 limit;
385 };
386 
387 static inline struct rvt_srq *ibsrq_to_rvtsrq(struct ib_srq *ibsrq)
388 {
389 	return container_of(ibsrq, struct rvt_srq, ibsrq);
390 }
391 
392 static inline struct rvt_qp *ibqp_to_rvtqp(struct ib_qp *ibqp)
393 {
394 	return container_of(ibqp, struct rvt_qp, ibqp);
395 }
396 
397 #define RVT_QPN_MAX                 BIT(24)
398 #define RVT_QPNMAP_ENTRIES          (RVT_QPN_MAX / PAGE_SIZE / BITS_PER_BYTE)
399 #define RVT_BITS_PER_PAGE           (PAGE_SIZE * BITS_PER_BYTE)
400 #define RVT_BITS_PER_PAGE_MASK      (RVT_BITS_PER_PAGE - 1)
401 #define RVT_QPN_MASK		    IB_QPN_MASK
402 
403 /*
404  * QPN-map pages start out as NULL, they get allocated upon
405  * first use and are never deallocated. This way,
406  * large bitmaps are not allocated unless large numbers of QPs are used.
407  */
408 struct rvt_qpn_map {
409 	void *page;
410 };
411 
412 struct rvt_qpn_table {
413 	spinlock_t lock; /* protect changes to the qp table */
414 	unsigned flags;         /* flags for QP0/1 allocated for each port */
415 	u32 last;               /* last QP number allocated */
416 	u32 nmaps;              /* size of the map table */
417 	u16 limit;
418 	u8  incr;
419 	/* bit map of free QP numbers other than 0/1 */
420 	struct rvt_qpn_map map[RVT_QPNMAP_ENTRIES];
421 };
422 
423 struct rvt_qp_ibdev {
424 	u32 qp_table_size;
425 	u32 qp_table_bits;
426 	struct rvt_qp __rcu **qp_table;
427 	spinlock_t qpt_lock; /* qptable lock */
428 	struct rvt_qpn_table qpn_table;
429 };
430 
431 /*
432  * There is one struct rvt_mcast for each multicast GID.
433  * All attached QPs are then stored as a list of
434  * struct rvt_mcast_qp.
435  */
436 struct rvt_mcast_qp {
437 	struct list_head list;
438 	struct rvt_qp *qp;
439 };
440 
441 struct rvt_mcast_addr {
442 	union ib_gid mgid;
443 	u16 lid;
444 };
445 
446 struct rvt_mcast {
447 	struct rb_node rb_node;
448 	struct rvt_mcast_addr mcast_addr;
449 	struct list_head qp_list;
450 	wait_queue_head_t wait;
451 	atomic_t refcount;
452 	int n_attached;
453 };
454 
455 /*
456  * Since struct rvt_swqe is not a fixed size, we can't simply index into
457  * struct rvt_qp.s_wq.  This function does the array index computation.
458  */
459 static inline struct rvt_swqe *rvt_get_swqe_ptr(struct rvt_qp *qp,
460 						unsigned n)
461 {
462 	return (struct rvt_swqe *)((char *)qp->s_wq +
463 				     (sizeof(struct rvt_swqe) +
464 				      qp->s_max_sge *
465 				      sizeof(struct rvt_sge)) * n);
466 }
467 
468 /*
469  * Since struct rvt_rwqe is not a fixed size, we can't simply index into
470  * struct rvt_rwq.wq.  This function does the array index computation.
471  */
472 static inline struct rvt_rwqe *rvt_get_rwqe_ptr(struct rvt_rq *rq, unsigned n)
473 {
474 	return (struct rvt_rwqe *)
475 		((char *)rq->wq->wq +
476 		 (sizeof(struct rvt_rwqe) +
477 		  rq->max_sge * sizeof(struct ib_sge)) * n);
478 }
479 
480 /**
481  * rvt_is_user_qp - return if this is user mode QP
482  * @qp - the target QP
483  */
484 static inline bool rvt_is_user_qp(struct rvt_qp *qp)
485 {
486 	return !!qp->pid;
487 }
488 
489 /**
490  * rvt_get_qp - get a QP reference
491  * @qp - the QP to hold
492  */
493 static inline void rvt_get_qp(struct rvt_qp *qp)
494 {
495 	atomic_inc(&qp->refcount);
496 }
497 
498 /**
499  * rvt_put_qp - release a QP reference
500  * @qp - the QP to release
501  */
502 static inline void rvt_put_qp(struct rvt_qp *qp)
503 {
504 	if (qp && atomic_dec_and_test(&qp->refcount))
505 		wake_up(&qp->wait);
506 }
507 
508 /**
509  * rvt_put_swqe - drop mr refs held by swqe
510  * @wqe - the send wqe
511  *
512  * This drops any mr references held by the swqe
513  */
514 static inline void rvt_put_swqe(struct rvt_swqe *wqe)
515 {
516 	int i;
517 
518 	for (i = 0; i < wqe->wr.num_sge; i++) {
519 		struct rvt_sge *sge = &wqe->sg_list[i];
520 
521 		rvt_put_mr(sge->mr);
522 	}
523 }
524 
525 /**
526  * rvt_qp_wqe_reserve - reserve operation
527  * @qp - the rvt qp
528  * @wqe - the send wqe
529  *
530  * This routine used in post send to record
531  * a wqe relative reserved operation use.
532  */
533 static inline void rvt_qp_wqe_reserve(
534 	struct rvt_qp *qp,
535 	struct rvt_swqe *wqe)
536 {
537 	atomic_inc(&qp->s_reserved_used);
538 }
539 
540 /**
541  * rvt_qp_wqe_unreserve - clean reserved operation
542  * @qp - the rvt qp
543  * @wqe - the send wqe
544  *
545  * This decrements the reserve use count.
546  *
547  * This call MUST precede the change to
548  * s_last to insure that post send sees a stable
549  * s_avail.
550  *
551  * An smp_mp__after_atomic() is used to insure
552  * the compiler does not juggle the order of the s_last
553  * ring index and the decrementing of s_reserved_used.
554  */
555 static inline void rvt_qp_wqe_unreserve(
556 	struct rvt_qp *qp,
557 	struct rvt_swqe *wqe)
558 {
559 	if (unlikely(wqe->wr.send_flags & RVT_SEND_RESERVE_USED)) {
560 		atomic_dec(&qp->s_reserved_used);
561 		/* insure no compiler re-order up to s_last change */
562 		smp_mb__after_atomic();
563 	}
564 }
565 
566 extern const enum ib_wc_opcode ib_rvt_wc_opcode[];
567 
568 /**
569  * rvt_qp_swqe_complete() - insert send completion
570  * @qp - the qp
571  * @wqe - the send wqe
572  * @status - completion status
573  *
574  * Insert a send completion into the completion
575  * queue if the qp indicates it should be done.
576  *
577  * See IBTA 10.7.3.1 for info on completion
578  * control.
579  */
580 static inline void rvt_qp_swqe_complete(
581 	struct rvt_qp *qp,
582 	struct rvt_swqe *wqe,
583 	enum ib_wc_opcode opcode,
584 	enum ib_wc_status status)
585 {
586 	if (unlikely(wqe->wr.send_flags & RVT_SEND_RESERVE_USED))
587 		return;
588 	if (!(qp->s_flags & RVT_S_SIGNAL_REQ_WR) ||
589 	    (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
590 	     status != IB_WC_SUCCESS) {
591 		struct ib_wc wc;
592 
593 		memset(&wc, 0, sizeof(wc));
594 		wc.wr_id = wqe->wr.wr_id;
595 		wc.status = status;
596 		wc.opcode = opcode;
597 		wc.qp = &qp->ibqp;
598 		wc.byte_len = wqe->length;
599 		rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.send_cq), &wc,
600 			     status != IB_WC_SUCCESS);
601 	}
602 }
603 
604 /*
605  * Compare the lower 24 bits of the msn values.
606  * Returns an integer <, ==, or > than zero.
607  */
608 static inline int rvt_cmp_msn(u32 a, u32 b)
609 {
610 	return (((int)a) - ((int)b)) << 8;
611 }
612 
613 /**
614  * rvt_compute_aeth - compute the AETH (syndrome + MSN)
615  * @qp: the queue pair to compute the AETH for
616  *
617  * Returns the AETH.
618  */
619 __be32 rvt_compute_aeth(struct rvt_qp *qp);
620 
621 /**
622  * rvt_get_credit - flush the send work queue of a QP
623  * @qp: the qp who's send work queue to flush
624  * @aeth: the Acknowledge Extended Transport Header
625  *
626  * The QP s_lock should be held.
627  */
628 void rvt_get_credit(struct rvt_qp *qp, u32 aeth);
629 
630 /**
631  * rvt_restart_sge - rewind the sge state for a wqe
632  * @ss: the sge state pointer
633  * @wqe: the wqe to rewind
634  * @len: the data length from the start of the wqe in bytes
635  *
636  * Returns the remaining data length.
637  */
638 u32 rvt_restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe, u32 len);
639 
640 /**
641  * @qp - the qp pair
642  * @len - the length
643  *
644  * Perform a shift based mtu round up divide
645  */
646 static inline u32 rvt_div_round_up_mtu(struct rvt_qp *qp, u32 len)
647 {
648 	return (len + qp->pmtu - 1) >> qp->log_pmtu;
649 }
650 
651 /**
652  * @qp - the qp pair
653  * @len - the length
654  *
655  * Perform a shift based mtu divide
656  */
657 static inline u32 rvt_div_mtu(struct rvt_qp *qp, u32 len)
658 {
659 	return len >> qp->log_pmtu;
660 }
661 
662 /**
663  * rvt_timeout_to_jiffies - Convert a ULP timeout input into jiffies
664  * @timeout - timeout input(0 - 31).
665  *
666  * Return a timeout value in jiffies.
667  */
668 static inline unsigned long rvt_timeout_to_jiffies(u8 timeout)
669 {
670 	if (timeout > 31)
671 		timeout = 31;
672 
673 	return usecs_to_jiffies(1U << timeout) * 4096UL / 1000UL;
674 }
675 
676 /**
677  * rvt_lookup_qpn - return the QP with the given QPN
678  * @ibp: the ibport
679  * @qpn: the QP number to look up
680  *
681  * The caller must hold the rcu_read_lock(), and keep the lock until
682  * the returned qp is no longer in use.
683  */
684 static inline struct rvt_qp *rvt_lookup_qpn(struct rvt_dev_info *rdi,
685 					    struct rvt_ibport *rvp,
686 					    u32 qpn) __must_hold(RCU)
687 {
688 	struct rvt_qp *qp = NULL;
689 
690 	if (unlikely(qpn <= 1)) {
691 		qp = rcu_dereference(rvp->qp[qpn]);
692 	} else {
693 		u32 n = hash_32(qpn, rdi->qp_dev->qp_table_bits);
694 
695 		for (qp = rcu_dereference(rdi->qp_dev->qp_table[n]); qp;
696 			qp = rcu_dereference(qp->next))
697 			if (qp->ibqp.qp_num == qpn)
698 				break;
699 	}
700 	return qp;
701 }
702 
703 /**
704  * rvt_mod_retry_timer - mod a retry timer
705  * @qp - the QP
706  * @shift - timeout shift to wait for multiple packets
707  * Modify a potentially already running retry timer
708  */
709 static inline void rvt_mod_retry_timer_ext(struct rvt_qp *qp, u8 shift)
710 {
711 	struct ib_qp *ibqp = &qp->ibqp;
712 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
713 
714 	lockdep_assert_held(&qp->s_lock);
715 	qp->s_flags |= RVT_S_TIMER;
716 	/* 4.096 usec. * (1 << qp->timeout) */
717 	mod_timer(&qp->s_timer, jiffies + rdi->busy_jiffies +
718 		  (qp->timeout_jiffies << shift));
719 }
720 
721 static inline void rvt_mod_retry_timer(struct rvt_qp *qp)
722 {
723 	return rvt_mod_retry_timer_ext(qp, 0);
724 }
725 
726 /**
727  * rvt_put_qp_swqe - drop refs held by swqe
728  * @qp: the send qp
729  * @wqe: the send wqe
730  *
731  * This drops any references held by the swqe
732  */
733 static inline void rvt_put_qp_swqe(struct rvt_qp *qp, struct rvt_swqe *wqe)
734 {
735 	rvt_put_swqe(wqe);
736 	if (qp->allowed_ops == IB_OPCODE_UD)
737 		atomic_dec(&ibah_to_rvtah(wqe->ud_wr.ah)->refcount);
738 }
739 
740 extern const int  ib_rvt_state_ops[];
741 
742 struct rvt_dev_info;
743 int rvt_get_rwqe(struct rvt_qp *qp, bool wr_id_only);
744 void rvt_comm_est(struct rvt_qp *qp);
745 int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err);
746 void rvt_rc_error(struct rvt_qp *qp, enum ib_wc_status err);
747 unsigned long rvt_rnr_tbl_to_usec(u32 index);
748 enum hrtimer_restart rvt_rc_rnr_retry(struct hrtimer *t);
749 void rvt_add_rnr_timer(struct rvt_qp *qp, u32 aeth);
750 void rvt_del_timers_sync(struct rvt_qp *qp);
751 void rvt_stop_rc_timers(struct rvt_qp *qp);
752 void rvt_add_retry_timer_ext(struct rvt_qp *qp, u8 shift);
753 static inline void rvt_add_retry_timer(struct rvt_qp *qp)
754 {
755 	rvt_add_retry_timer_ext(qp, 0);
756 }
757 
758 void rvt_copy_sge(struct rvt_qp *qp, struct rvt_sge_state *ss,
759 		  void *data, u32 length,
760 		  bool release, bool copy_last);
761 void rvt_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
762 		       enum ib_wc_status status);
763 void rvt_ruc_loopback(struct rvt_qp *qp);
764 
765 /**
766  * struct rvt_qp_iter - the iterator for QPs
767  * @qp - the current QP
768  *
769  * This structure defines the current iterator
770  * state for sequenced access to all QPs relative
771  * to an rvt_dev_info.
772  */
773 struct rvt_qp_iter {
774 	struct rvt_qp *qp;
775 	/* private: backpointer */
776 	struct rvt_dev_info *rdi;
777 	/* private: callback routine */
778 	void (*cb)(struct rvt_qp *qp, u64 v);
779 	/* private: for arg to callback routine */
780 	u64 v;
781 	/* private: number of SMI,GSI QPs for device */
782 	int specials;
783 	/* private: current iterator index */
784 	int n;
785 };
786 
787 struct rvt_qp_iter *rvt_qp_iter_init(struct rvt_dev_info *rdi,
788 				     u64 v,
789 				     void (*cb)(struct rvt_qp *qp, u64 v));
790 int rvt_qp_iter_next(struct rvt_qp_iter *iter);
791 void rvt_qp_iter(struct rvt_dev_info *rdi,
792 		 u64 v,
793 		 void (*cb)(struct rvt_qp *qp, u64 v));
794 void rvt_qp_mr_clean(struct rvt_qp *qp, u32 lkey);
795 #endif          /* DEF_RDMAVT_INCQP_H */
796