xref: /openbmc/linux/include/rdma/rdmavt_qp.h (revision 2b74c878)
1 #ifndef DEF_RDMAVT_INCQP_H
2 #define DEF_RDMAVT_INCQP_H
3 
4 /*
5  * Copyright(c) 2016 - 2019 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 #include <rdma/rvt-abi.h>
56 /*
57  * Atomic bit definitions for r_aflags.
58  */
59 #define RVT_R_WRID_VALID        0
60 #define RVT_R_REWIND_SGE        1
61 
62 /*
63  * Bit definitions for r_flags.
64  */
65 #define RVT_R_REUSE_SGE 0x01
66 #define RVT_R_RDMAR_SEQ 0x02
67 #define RVT_R_RSP_NAK   0x04
68 #define RVT_R_RSP_SEND  0x08
69 #define RVT_R_COMM_EST  0x10
70 
71 /*
72  * Bit definitions for s_flags.
73  *
74  * RVT_S_SIGNAL_REQ_WR - set if QP send WRs contain completion signaled
75  * RVT_S_BUSY - send tasklet is processing the QP
76  * RVT_S_TIMER - the RC retry timer is active
77  * RVT_S_ACK_PENDING - an ACK is waiting to be sent after RDMA read/atomics
78  * RVT_S_WAIT_FENCE - waiting for all prior RDMA read or atomic SWQEs
79  *                         before processing the next SWQE
80  * RVT_S_WAIT_RDMAR - waiting for a RDMA read or atomic SWQE to complete
81  *                         before processing the next SWQE
82  * RVT_S_WAIT_RNR - waiting for RNR timeout
83  * RVT_S_WAIT_SSN_CREDIT - waiting for RC credits to process next SWQE
84  * RVT_S_WAIT_DMA - waiting for send DMA queue to drain before generating
85  *                  next send completion entry not via send DMA
86  * RVT_S_WAIT_PIO - waiting for a send buffer to be available
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  * rvt_ud_wr - IB UD work plus AH cache
162  * @wr: valid IB work request
163  * @attr: pointer to an allocated AH attribute
164  *
165  * Special case the UD WR so we can keep track of the AH attributes.
166  *
167  * NOTE: This data structure is stricly ordered wr then attr. I.e the attr
168  * MUST come after wr.  The ib_ud_wr is sized and copied in rvt_post_one_wr.
169  * The copy assumes that wr is first.
170  */
171 struct rvt_ud_wr {
172 	struct ib_ud_wr wr;
173 	struct rdma_ah_attr *attr;
174 };
175 
176 /*
177  * Send work request queue entry.
178  * The size of the sg_list is determined when the QP is created and stored
179  * in qp->s_max_sge.
180  */
181 struct rvt_swqe {
182 	union {
183 		struct ib_send_wr wr;   /* don't use wr.sg_list */
184 		struct rvt_ud_wr ud_wr;
185 		struct ib_reg_wr reg_wr;
186 		struct ib_rdma_wr rdma_wr;
187 		struct ib_atomic_wr atomic_wr;
188 	};
189 	u32 psn;                /* first packet sequence number */
190 	u32 lpsn;               /* last packet sequence number */
191 	u32 ssn;                /* send sequence number */
192 	u32 length;             /* total length of data in sg_list */
193 	void *priv;             /* driver dependent field */
194 	struct rvt_sge sg_list[0];
195 };
196 
197 /**
198  * struct rvt_krwq - kernel struct receive work request
199  * @p_lock: lock to protect producer of the kernel buffer
200  * @head: index of next entry to fill
201  * @c_lock:lock to protect consumer of the kernel buffer
202  * @tail: index of next entry to pull
203  * @count: count is aproximate of total receive enteries posted
204  * @rvt_rwqe: struct of receive work request queue entry
205  *
206  * This structure is used to contain the head pointer,
207  * tail pointer and receive work queue entries for kernel
208  * mode user.
209  */
210 struct rvt_krwq {
211 	spinlock_t p_lock;	/* protect producer */
212 	u32 head;               /* new work requests posted to the head */
213 
214 	/* protect consumer */
215 	spinlock_t c_lock ____cacheline_aligned_in_smp;
216 	u32 tail;               /* receives pull requests from here. */
217 	u32 count;		/* approx count of receive entries posted */
218 	struct rvt_rwqe *curr_wq;
219 	struct rvt_rwqe wq[];
220 };
221 
222 /*
223  * rvt_get_swqe_ah - Return the pointer to the struct rvt_ah
224  * @swqe: valid Send WQE
225  *
226  */
227 static inline struct rvt_ah *rvt_get_swqe_ah(struct rvt_swqe *swqe)
228 {
229 	return ibah_to_rvtah(swqe->ud_wr.wr.ah);
230 }
231 
232 /**
233  * rvt_get_swqe_ah_attr - Return the cached ah attribute information
234  * @swqe: valid Send WQE
235  *
236  */
237 static inline struct rdma_ah_attr *rvt_get_swqe_ah_attr(struct rvt_swqe *swqe)
238 {
239 	return swqe->ud_wr.attr;
240 }
241 
242 /**
243  * rvt_get_swqe_remote_qpn - Access the remote QPN value
244  * @swqe: valid Send WQE
245  *
246  */
247 static inline u32 rvt_get_swqe_remote_qpn(struct rvt_swqe *swqe)
248 {
249 	return swqe->ud_wr.wr.remote_qpn;
250 }
251 
252 /**
253  * rvt_get_swqe_remote_qkey - Acces the remote qkey value
254  * @swqe: valid Send WQE
255  *
256  */
257 static inline u32 rvt_get_swqe_remote_qkey(struct rvt_swqe *swqe)
258 {
259 	return swqe->ud_wr.wr.remote_qkey;
260 }
261 
262 /**
263  * rvt_get_swqe_pkey_index - Access the pkey index
264  * @swqe: valid Send WQE
265  *
266  */
267 static inline u16 rvt_get_swqe_pkey_index(struct rvt_swqe *swqe)
268 {
269 	return swqe->ud_wr.wr.pkey_index;
270 }
271 
272 struct rvt_rq {
273 	struct rvt_rwq *wq;
274 	struct rvt_krwq *kwq;
275 	u32 size;               /* size of RWQE array */
276 	u8 max_sge;
277 	/* protect changes in this struct */
278 	spinlock_t lock ____cacheline_aligned_in_smp;
279 };
280 
281 /*
282  * This structure holds the information that the send tasklet needs
283  * to send a RDMA read response or atomic operation.
284  */
285 struct rvt_ack_entry {
286 	struct rvt_sge rdma_sge;
287 	u64 atomic_data;
288 	u32 psn;
289 	u32 lpsn;
290 	u8 opcode;
291 	u8 sent;
292 	void *priv;
293 };
294 
295 #define	RC_QP_SCALING_INTERVAL	5
296 
297 #define RVT_OPERATION_PRIV        0x00000001
298 #define RVT_OPERATION_ATOMIC      0x00000002
299 #define RVT_OPERATION_ATOMIC_SGE  0x00000004
300 #define RVT_OPERATION_LOCAL       0x00000008
301 #define RVT_OPERATION_USE_RESERVE 0x00000010
302 #define RVT_OPERATION_IGN_RNR_CNT 0x00000020
303 
304 #define RVT_OPERATION_MAX (IB_WR_RESERVED10 + 1)
305 
306 /**
307  * rvt_operation_params - op table entry
308  * @length - the length to copy into the swqe entry
309  * @qpt_support - a bit mask indicating QP type support
310  * @flags - RVT_OPERATION flags (see above)
311  *
312  * This supports table driven post send so that
313  * the driver can have differing an potentially
314  * different sets of operations.
315  *
316  **/
317 
318 struct rvt_operation_params {
319 	size_t length;
320 	u32 qpt_support;
321 	u32 flags;
322 };
323 
324 /*
325  * Common variables are protected by both r_rq.lock and s_lock in that order
326  * which only happens in modify_qp() or changing the QP 'state'.
327  */
328 struct rvt_qp {
329 	struct ib_qp ibqp;
330 	void *priv; /* Driver private data */
331 	/* read mostly fields above and below */
332 	struct rdma_ah_attr remote_ah_attr;
333 	struct rdma_ah_attr alt_ah_attr;
334 	struct rvt_qp __rcu *next;           /* link list for QPN hash table */
335 	struct rvt_swqe *s_wq;  /* send work queue */
336 	struct rvt_mmap_info *ip;
337 
338 	unsigned long timeout_jiffies;  /* computed from timeout */
339 
340 	int srate_mbps;		/* s_srate (below) converted to Mbit/s */
341 	pid_t pid;		/* pid for user mode QPs */
342 	u32 remote_qpn;
343 	u32 qkey;               /* QKEY for this QP (for UD or RD) */
344 	u32 s_size;             /* send work queue size */
345 
346 	u16 pmtu;		/* decoded from path_mtu */
347 	u8 log_pmtu;		/* shift for pmtu */
348 	u8 state;               /* QP state */
349 	u8 allowed_ops;		/* high order bits of allowed opcodes */
350 	u8 qp_access_flags;
351 	u8 alt_timeout;         /* Alternate path timeout for this QP */
352 	u8 timeout;             /* Timeout for this QP */
353 	u8 s_srate;
354 	u8 s_mig_state;
355 	u8 port_num;
356 	u8 s_pkey_index;        /* PKEY index to use */
357 	u8 s_alt_pkey_index;    /* Alternate path PKEY index to use */
358 	u8 r_max_rd_atomic;     /* max number of RDMA read/atomic to receive */
359 	u8 s_max_rd_atomic;     /* max number of RDMA read/atomic to send */
360 	u8 s_retry_cnt;         /* number of times to retry */
361 	u8 s_rnr_retry_cnt;
362 	u8 r_min_rnr_timer;     /* retry timeout value for RNR NAKs */
363 	u8 s_max_sge;           /* size of s_wq->sg_list */
364 	u8 s_draining;
365 
366 	/* start of read/write fields */
367 	atomic_t refcount ____cacheline_aligned_in_smp;
368 	wait_queue_head_t wait;
369 
370 	struct rvt_ack_entry *s_ack_queue;
371 	struct rvt_sge_state s_rdma_read_sge;
372 
373 	spinlock_t r_lock ____cacheline_aligned_in_smp;      /* used for APM */
374 	u32 r_psn;              /* expected rcv packet sequence number */
375 	unsigned long r_aflags;
376 	u64 r_wr_id;            /* ID for current receive WQE */
377 	u32 r_ack_psn;          /* PSN for next ACK or atomic ACK */
378 	u32 r_len;              /* total length of r_sge */
379 	u32 r_rcv_len;          /* receive data len processed */
380 	u32 r_msn;              /* message sequence number */
381 
382 	u8 r_state;             /* opcode of last packet received */
383 	u8 r_flags;
384 	u8 r_head_ack_queue;    /* index into s_ack_queue[] */
385 	u8 r_adefered;          /* defered ack count */
386 
387 	struct list_head rspwait;       /* link for waiting to respond */
388 
389 	struct rvt_sge_state r_sge;     /* current receive data */
390 	struct rvt_rq r_rq;             /* receive work queue */
391 
392 	/* post send line */
393 	spinlock_t s_hlock ____cacheline_aligned_in_smp;
394 	u32 s_head;             /* new entries added here */
395 	u32 s_next_psn;         /* PSN for next request */
396 	u32 s_avail;            /* number of entries avail */
397 	u32 s_ssn;              /* SSN of tail entry */
398 	atomic_t s_reserved_used; /* reserved entries in use */
399 
400 	spinlock_t s_lock ____cacheline_aligned_in_smp;
401 	u32 s_flags;
402 	struct rvt_sge_state *s_cur_sge;
403 	struct rvt_swqe *s_wqe;
404 	struct rvt_sge_state s_sge;     /* current send request data */
405 	struct rvt_mregion *s_rdma_mr;
406 	u32 s_len;              /* total length of s_sge */
407 	u32 s_rdma_read_len;    /* total length of s_rdma_read_sge */
408 	u32 s_last_psn;         /* last response PSN processed */
409 	u32 s_sending_psn;      /* lowest PSN that is being sent */
410 	u32 s_sending_hpsn;     /* highest PSN that is being sent */
411 	u32 s_psn;              /* current packet sequence number */
412 	u32 s_ack_rdma_psn;     /* PSN for sending RDMA read responses */
413 	u32 s_ack_psn;          /* PSN for acking sends and RDMA writes */
414 	u32 s_tail;             /* next entry to process */
415 	u32 s_cur;              /* current work queue entry */
416 	u32 s_acked;            /* last un-ACK'ed entry */
417 	u32 s_last;             /* last completed entry */
418 	u32 s_lsn;              /* limit sequence number (credit) */
419 	u32 s_ahgpsn;           /* set to the psn in the copy of the header */
420 	u16 s_cur_size;         /* size of send packet in bytes */
421 	u16 s_rdma_ack_cnt;
422 	u8 s_hdrwords;         /* size of s_hdr in 32 bit words */
423 	s8 s_ahgidx;
424 	u8 s_state;             /* opcode of last packet sent */
425 	u8 s_ack_state;         /* opcode of packet to ACK */
426 	u8 s_nak_state;         /* non-zero if NAK is pending */
427 	u8 r_nak_state;         /* non-zero if NAK is pending */
428 	u8 s_retry;             /* requester retry counter */
429 	u8 s_rnr_retry;         /* requester RNR retry counter */
430 	u8 s_num_rd_atomic;     /* number of RDMA read/atomic pending */
431 	u8 s_tail_ack_queue;    /* index into s_ack_queue[] */
432 	u8 s_acked_ack_queue;   /* index into s_ack_queue[] */
433 
434 	struct rvt_sge_state s_ack_rdma_sge;
435 	struct timer_list s_timer;
436 	struct hrtimer s_rnr_timer;
437 
438 	atomic_t local_ops_pending; /* number of fast_reg/local_inv reqs */
439 
440 	/*
441 	 * This sge list MUST be last. Do not add anything below here.
442 	 */
443 	struct rvt_sge r_sg_list[0] /* verified SGEs */
444 		____cacheline_aligned_in_smp;
445 };
446 
447 struct rvt_srq {
448 	struct ib_srq ibsrq;
449 	struct rvt_rq rq;
450 	struct rvt_mmap_info *ip;
451 	/* send signal when number of RWQEs < limit */
452 	u32 limit;
453 };
454 
455 static inline struct rvt_srq *ibsrq_to_rvtsrq(struct ib_srq *ibsrq)
456 {
457 	return container_of(ibsrq, struct rvt_srq, ibsrq);
458 }
459 
460 static inline struct rvt_qp *ibqp_to_rvtqp(struct ib_qp *ibqp)
461 {
462 	return container_of(ibqp, struct rvt_qp, ibqp);
463 }
464 
465 #define RVT_QPN_MAX                 BIT(24)
466 #define RVT_QPNMAP_ENTRIES          (RVT_QPN_MAX / PAGE_SIZE / BITS_PER_BYTE)
467 #define RVT_BITS_PER_PAGE           (PAGE_SIZE * BITS_PER_BYTE)
468 #define RVT_BITS_PER_PAGE_MASK      (RVT_BITS_PER_PAGE - 1)
469 #define RVT_QPN_MASK		    IB_QPN_MASK
470 
471 /*
472  * QPN-map pages start out as NULL, they get allocated upon
473  * first use and are never deallocated. This way,
474  * large bitmaps are not allocated unless large numbers of QPs are used.
475  */
476 struct rvt_qpn_map {
477 	void *page;
478 };
479 
480 struct rvt_qpn_table {
481 	spinlock_t lock; /* protect changes to the qp table */
482 	unsigned flags;         /* flags for QP0/1 allocated for each port */
483 	u32 last;               /* last QP number allocated */
484 	u32 nmaps;              /* size of the map table */
485 	u16 limit;
486 	u8  incr;
487 	/* bit map of free QP numbers other than 0/1 */
488 	struct rvt_qpn_map map[RVT_QPNMAP_ENTRIES];
489 };
490 
491 struct rvt_qp_ibdev {
492 	u32 qp_table_size;
493 	u32 qp_table_bits;
494 	struct rvt_qp __rcu **qp_table;
495 	spinlock_t qpt_lock; /* qptable lock */
496 	struct rvt_qpn_table qpn_table;
497 };
498 
499 /*
500  * There is one struct rvt_mcast for each multicast GID.
501  * All attached QPs are then stored as a list of
502  * struct rvt_mcast_qp.
503  */
504 struct rvt_mcast_qp {
505 	struct list_head list;
506 	struct rvt_qp *qp;
507 };
508 
509 struct rvt_mcast_addr {
510 	union ib_gid mgid;
511 	u16 lid;
512 };
513 
514 struct rvt_mcast {
515 	struct rb_node rb_node;
516 	struct rvt_mcast_addr mcast_addr;
517 	struct list_head qp_list;
518 	wait_queue_head_t wait;
519 	atomic_t refcount;
520 	int n_attached;
521 };
522 
523 /*
524  * Since struct rvt_swqe is not a fixed size, we can't simply index into
525  * struct rvt_qp.s_wq.  This function does the array index computation.
526  */
527 static inline struct rvt_swqe *rvt_get_swqe_ptr(struct rvt_qp *qp,
528 						unsigned n)
529 {
530 	return (struct rvt_swqe *)((char *)qp->s_wq +
531 				     (sizeof(struct rvt_swqe) +
532 				      qp->s_max_sge *
533 				      sizeof(struct rvt_sge)) * n);
534 }
535 
536 /*
537  * Since struct rvt_rwqe is not a fixed size, we can't simply index into
538  * struct rvt_rwq.wq.  This function does the array index computation.
539  */
540 static inline struct rvt_rwqe *rvt_get_rwqe_ptr(struct rvt_rq *rq, unsigned n)
541 {
542 	return (struct rvt_rwqe *)
543 		((char *)rq->kwq->curr_wq +
544 		 (sizeof(struct rvt_rwqe) +
545 		  rq->max_sge * sizeof(struct ib_sge)) * n);
546 }
547 
548 /**
549  * rvt_is_user_qp - return if this is user mode QP
550  * @qp - the target QP
551  */
552 static inline bool rvt_is_user_qp(struct rvt_qp *qp)
553 {
554 	return !!qp->pid;
555 }
556 
557 /**
558  * rvt_get_qp - get a QP reference
559  * @qp - the QP to hold
560  */
561 static inline void rvt_get_qp(struct rvt_qp *qp)
562 {
563 	atomic_inc(&qp->refcount);
564 }
565 
566 /**
567  * rvt_put_qp - release a QP reference
568  * @qp - the QP to release
569  */
570 static inline void rvt_put_qp(struct rvt_qp *qp)
571 {
572 	if (qp && atomic_dec_and_test(&qp->refcount))
573 		wake_up(&qp->wait);
574 }
575 
576 /**
577  * rvt_put_swqe - drop mr refs held by swqe
578  * @wqe - the send wqe
579  *
580  * This drops any mr references held by the swqe
581  */
582 static inline void rvt_put_swqe(struct rvt_swqe *wqe)
583 {
584 	int i;
585 
586 	for (i = 0; i < wqe->wr.num_sge; i++) {
587 		struct rvt_sge *sge = &wqe->sg_list[i];
588 
589 		rvt_put_mr(sge->mr);
590 	}
591 }
592 
593 /**
594  * rvt_qp_wqe_reserve - reserve operation
595  * @qp - the rvt qp
596  * @wqe - the send wqe
597  *
598  * This routine used in post send to record
599  * a wqe relative reserved operation use.
600  */
601 static inline void rvt_qp_wqe_reserve(
602 	struct rvt_qp *qp,
603 	struct rvt_swqe *wqe)
604 {
605 	atomic_inc(&qp->s_reserved_used);
606 }
607 
608 /**
609  * rvt_qp_wqe_unreserve - clean reserved operation
610  * @qp - the rvt qp
611  * @flags - send wqe flags
612  *
613  * This decrements the reserve use count.
614  *
615  * This call MUST precede the change to
616  * s_last to insure that post send sees a stable
617  * s_avail.
618  *
619  * An smp_mp__after_atomic() is used to insure
620  * the compiler does not juggle the order of the s_last
621  * ring index and the decrementing of s_reserved_used.
622  */
623 static inline void rvt_qp_wqe_unreserve(struct rvt_qp *qp, int flags)
624 {
625 	if (unlikely(flags & RVT_SEND_RESERVE_USED)) {
626 		atomic_dec(&qp->s_reserved_used);
627 		/* insure no compiler re-order up to s_last change */
628 		smp_mb__after_atomic();
629 	}
630 }
631 
632 extern const enum ib_wc_opcode ib_rvt_wc_opcode[];
633 
634 /*
635  * Compare the lower 24 bits of the msn values.
636  * Returns an integer <, ==, or > than zero.
637  */
638 static inline int rvt_cmp_msn(u32 a, u32 b)
639 {
640 	return (((int)a) - ((int)b)) << 8;
641 }
642 
643 /**
644  * rvt_compute_aeth - compute the AETH (syndrome + MSN)
645  * @qp: the queue pair to compute the AETH for
646  *
647  * Returns the AETH.
648  */
649 __be32 rvt_compute_aeth(struct rvt_qp *qp);
650 
651 /**
652  * rvt_get_credit - flush the send work queue of a QP
653  * @qp: the qp who's send work queue to flush
654  * @aeth: the Acknowledge Extended Transport Header
655  *
656  * The QP s_lock should be held.
657  */
658 void rvt_get_credit(struct rvt_qp *qp, u32 aeth);
659 
660 /**
661  * rvt_restart_sge - rewind the sge state for a wqe
662  * @ss: the sge state pointer
663  * @wqe: the wqe to rewind
664  * @len: the data length from the start of the wqe in bytes
665  *
666  * Returns the remaining data length.
667  */
668 u32 rvt_restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe, u32 len);
669 
670 /**
671  * @qp - the qp pair
672  * @len - the length
673  *
674  * Perform a shift based mtu round up divide
675  */
676 static inline u32 rvt_div_round_up_mtu(struct rvt_qp *qp, u32 len)
677 {
678 	return (len + qp->pmtu - 1) >> qp->log_pmtu;
679 }
680 
681 /**
682  * @qp - the qp pair
683  * @len - the length
684  *
685  * Perform a shift based mtu divide
686  */
687 static inline u32 rvt_div_mtu(struct rvt_qp *qp, u32 len)
688 {
689 	return len >> qp->log_pmtu;
690 }
691 
692 /**
693  * rvt_timeout_to_jiffies - Convert a ULP timeout input into jiffies
694  * @timeout - timeout input(0 - 31).
695  *
696  * Return a timeout value in jiffies.
697  */
698 static inline unsigned long rvt_timeout_to_jiffies(u8 timeout)
699 {
700 	if (timeout > 31)
701 		timeout = 31;
702 
703 	return usecs_to_jiffies(1U << timeout) * 4096UL / 1000UL;
704 }
705 
706 /**
707  * rvt_lookup_qpn - return the QP with the given QPN
708  * @ibp: the ibport
709  * @qpn: the QP number to look up
710  *
711  * The caller must hold the rcu_read_lock(), and keep the lock until
712  * the returned qp is no longer in use.
713  */
714 static inline struct rvt_qp *rvt_lookup_qpn(struct rvt_dev_info *rdi,
715 					    struct rvt_ibport *rvp,
716 					    u32 qpn) __must_hold(RCU)
717 {
718 	struct rvt_qp *qp = NULL;
719 
720 	if (unlikely(qpn <= 1)) {
721 		qp = rcu_dereference(rvp->qp[qpn]);
722 	} else {
723 		u32 n = hash_32(qpn, rdi->qp_dev->qp_table_bits);
724 
725 		for (qp = rcu_dereference(rdi->qp_dev->qp_table[n]); qp;
726 			qp = rcu_dereference(qp->next))
727 			if (qp->ibqp.qp_num == qpn)
728 				break;
729 	}
730 	return qp;
731 }
732 
733 /**
734  * rvt_mod_retry_timer - mod a retry timer
735  * @qp - the QP
736  * @shift - timeout shift to wait for multiple packets
737  * Modify a potentially already running retry timer
738  */
739 static inline void rvt_mod_retry_timer_ext(struct rvt_qp *qp, u8 shift)
740 {
741 	struct ib_qp *ibqp = &qp->ibqp;
742 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
743 
744 	lockdep_assert_held(&qp->s_lock);
745 	qp->s_flags |= RVT_S_TIMER;
746 	/* 4.096 usec. * (1 << qp->timeout) */
747 	mod_timer(&qp->s_timer, jiffies + rdi->busy_jiffies +
748 		  (qp->timeout_jiffies << shift));
749 }
750 
751 static inline void rvt_mod_retry_timer(struct rvt_qp *qp)
752 {
753 	return rvt_mod_retry_timer_ext(qp, 0);
754 }
755 
756 /**
757  * rvt_put_qp_swqe - drop refs held by swqe
758  * @qp: the send qp
759  * @wqe: the send wqe
760  *
761  * This drops any references held by the swqe
762  */
763 static inline void rvt_put_qp_swqe(struct rvt_qp *qp, struct rvt_swqe *wqe)
764 {
765 	rvt_put_swqe(wqe);
766 	if (qp->allowed_ops == IB_OPCODE_UD)
767 		rdma_destroy_ah_attr(wqe->ud_wr.attr);
768 }
769 
770 /**
771  * rvt_qp_sqwe_incr - increment ring index
772  * @qp: the qp
773  * @val: the starting value
774  *
775  * Return: the new value wrapping as appropriate
776  */
777 static inline u32
778 rvt_qp_swqe_incr(struct rvt_qp *qp, u32 val)
779 {
780 	if (++val >= qp->s_size)
781 		val = 0;
782 	return val;
783 }
784 
785 int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err);
786 
787 /**
788  * rvt_recv_cq - add a new entry to completion queue
789  *			by receive queue
790  * @qp: receive queue
791  * @wc: work completion entry to add
792  * @solicited: true if @entry is solicited
793  *
794  * This is wrapper function for rvt_enter_cq function call by
795  * receive queue. If rvt_cq_enter return false, it means cq is
796  * full and the qp is put into error state.
797  */
798 static inline void rvt_recv_cq(struct rvt_qp *qp, struct ib_wc *wc,
799 			       bool solicited)
800 {
801 	struct rvt_cq *cq = ibcq_to_rvtcq(qp->ibqp.recv_cq);
802 
803 	if (unlikely(!rvt_cq_enter(cq, wc, solicited)))
804 		rvt_error_qp(qp, IB_WC_LOC_QP_OP_ERR);
805 }
806 
807 /**
808  * rvt_send_cq - add a new entry to completion queue
809  *                        by send queue
810  * @qp: send queue
811  * @wc: work completion entry to add
812  * @solicited: true if @entry is solicited
813  *
814  * This is wrapper function for rvt_enter_cq function call by
815  * send queue. If rvt_cq_enter return false, it means cq is
816  * full and the qp is put into error state.
817  */
818 static inline void rvt_send_cq(struct rvt_qp *qp, struct ib_wc *wc,
819 			       bool solicited)
820 {
821 	struct rvt_cq *cq = ibcq_to_rvtcq(qp->ibqp.send_cq);
822 
823 	if (unlikely(!rvt_cq_enter(cq, wc, solicited)))
824 		rvt_error_qp(qp, IB_WC_LOC_QP_OP_ERR);
825 }
826 
827 /**
828  * rvt_qp_complete_swqe - insert send completion
829  * @qp - the qp
830  * @wqe - the send wqe
831  * @opcode - wc operation (driver dependent)
832  * @status - completion status
833  *
834  * Update the s_last information, and then insert a send
835  * completion into the completion
836  * queue if the qp indicates it should be done.
837  *
838  * See IBTA 10.7.3.1 for info on completion
839  * control.
840  *
841  * Return: new last
842  */
843 static inline u32
844 rvt_qp_complete_swqe(struct rvt_qp *qp,
845 		     struct rvt_swqe *wqe,
846 		     enum ib_wc_opcode opcode,
847 		     enum ib_wc_status status)
848 {
849 	bool need_completion;
850 	u64 wr_id;
851 	u32 byte_len, last;
852 	int flags = wqe->wr.send_flags;
853 
854 	rvt_qp_wqe_unreserve(qp, flags);
855 	rvt_put_qp_swqe(qp, wqe);
856 
857 	need_completion =
858 		!(flags & RVT_SEND_RESERVE_USED) &&
859 		(!(qp->s_flags & RVT_S_SIGNAL_REQ_WR) ||
860 		(flags & IB_SEND_SIGNALED) ||
861 		status != IB_WC_SUCCESS);
862 	if (need_completion) {
863 		wr_id = wqe->wr.wr_id;
864 		byte_len = wqe->length;
865 		/* above fields required before writing s_last */
866 	}
867 	last = rvt_qp_swqe_incr(qp, qp->s_last);
868 	/* see rvt_qp_is_avail() */
869 	smp_store_release(&qp->s_last, last);
870 	if (need_completion) {
871 		struct ib_wc w = {
872 			.wr_id = wr_id,
873 			.status = status,
874 			.opcode = opcode,
875 			.qp = &qp->ibqp,
876 			.byte_len = byte_len,
877 		};
878 		rvt_send_cq(qp, &w, status != IB_WC_SUCCESS);
879 	}
880 	return last;
881 }
882 
883 extern const int  ib_rvt_state_ops[];
884 
885 struct rvt_dev_info;
886 int rvt_get_rwqe(struct rvt_qp *qp, bool wr_id_only);
887 void rvt_comm_est(struct rvt_qp *qp);
888 void rvt_rc_error(struct rvt_qp *qp, enum ib_wc_status err);
889 unsigned long rvt_rnr_tbl_to_usec(u32 index);
890 enum hrtimer_restart rvt_rc_rnr_retry(struct hrtimer *t);
891 void rvt_add_rnr_timer(struct rvt_qp *qp, u32 aeth);
892 void rvt_del_timers_sync(struct rvt_qp *qp);
893 void rvt_stop_rc_timers(struct rvt_qp *qp);
894 void rvt_add_retry_timer_ext(struct rvt_qp *qp, u8 shift);
895 static inline void rvt_add_retry_timer(struct rvt_qp *qp)
896 {
897 	rvt_add_retry_timer_ext(qp, 0);
898 }
899 
900 void rvt_copy_sge(struct rvt_qp *qp, struct rvt_sge_state *ss,
901 		  void *data, u32 length,
902 		  bool release, bool copy_last);
903 void rvt_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
904 		       enum ib_wc_status status);
905 void rvt_ruc_loopback(struct rvt_qp *qp);
906 
907 /**
908  * struct rvt_qp_iter - the iterator for QPs
909  * @qp - the current QP
910  *
911  * This structure defines the current iterator
912  * state for sequenced access to all QPs relative
913  * to an rvt_dev_info.
914  */
915 struct rvt_qp_iter {
916 	struct rvt_qp *qp;
917 	/* private: backpointer */
918 	struct rvt_dev_info *rdi;
919 	/* private: callback routine */
920 	void (*cb)(struct rvt_qp *qp, u64 v);
921 	/* private: for arg to callback routine */
922 	u64 v;
923 	/* private: number of SMI,GSI QPs for device */
924 	int specials;
925 	/* private: current iterator index */
926 	int n;
927 };
928 
929 /**
930  * ib_cq_tail - Return tail index of cq buffer
931  * @send_cq - The cq for send
932  *
933  * This is called in qp_iter_print to get tail
934  * of cq buffer.
935  */
936 static inline u32 ib_cq_tail(struct ib_cq *send_cq)
937 {
938 	struct rvt_cq *cq = ibcq_to_rvtcq(send_cq);
939 
940 	return ibcq_to_rvtcq(send_cq)->ip ?
941 	       RDMA_READ_UAPI_ATOMIC(cq->queue->tail) :
942 	       ibcq_to_rvtcq(send_cq)->kqueue->tail;
943 }
944 
945 /**
946  * ib_cq_head - Return head index of cq buffer
947  * @send_cq - The cq for send
948  *
949  * This is called in qp_iter_print to get head
950  * of cq buffer.
951  */
952 static inline u32 ib_cq_head(struct ib_cq *send_cq)
953 {
954 	struct rvt_cq *cq = ibcq_to_rvtcq(send_cq);
955 
956 	return ibcq_to_rvtcq(send_cq)->ip ?
957 	       RDMA_READ_UAPI_ATOMIC(cq->queue->head) :
958 	       ibcq_to_rvtcq(send_cq)->kqueue->head;
959 }
960 
961 /**
962  * rvt_free_rq - free memory allocated for rvt_rq struct
963  * @rvt_rq: request queue data structure
964  *
965  * This function should only be called if the rvt_mmap_info()
966  * has not succeeded.
967  */
968 static inline void rvt_free_rq(struct rvt_rq *rq)
969 {
970 	kvfree(rq->kwq);
971 	rq->kwq = NULL;
972 	vfree(rq->wq);
973 	rq->wq = NULL;
974 }
975 
976 struct rvt_qp_iter *rvt_qp_iter_init(struct rvt_dev_info *rdi,
977 				     u64 v,
978 				     void (*cb)(struct rvt_qp *qp, u64 v));
979 int rvt_qp_iter_next(struct rvt_qp_iter *iter);
980 void rvt_qp_iter(struct rvt_dev_info *rdi,
981 		 u64 v,
982 		 void (*cb)(struct rvt_qp *qp, u64 v));
983 void rvt_qp_mr_clean(struct rvt_qp *qp, u32 lkey);
984 #endif          /* DEF_RDMAVT_INCQP_H */
985