xref: /openbmc/linux/include/rdma/rdmavt_qp.h (revision 4c5a116a)
1 #ifndef DEF_RDMAVT_INCQP_H
2 #define DEF_RDMAVT_INCQP_H
3 
4 /*
5  * Copyright(c) 2016 - 2020 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  * If a packet's QP[23:16] bits match this value, then it is
73  * a PSM packet and the hardware will expect a KDETH header
74  * following the BTH.
75  */
76 #define RVT_KDETH_QP_PREFIX       0x80
77 #define RVT_KDETH_QP_SUFFIX       0xffff
78 #define RVT_KDETH_QP_PREFIX_MASK  0x00ff0000
79 #define RVT_KDETH_QP_PREFIX_SHIFT 16
80 #define RVT_KDETH_QP_BASE         (u32)(RVT_KDETH_QP_PREFIX << \
81 					RVT_KDETH_QP_PREFIX_SHIFT)
82 #define RVT_KDETH_QP_MAX          (u32)(RVT_KDETH_QP_BASE + RVT_KDETH_QP_SUFFIX)
83 
84 /*
85  * If a packet's LNH == BTH and DEST QPN[23:16] in the BTH match this
86  * prefix value, then it is an AIP packet with a DETH containing the entropy
87  * value in byte 4 following the BTH.
88  */
89 #define RVT_AIP_QP_PREFIX       0x81
90 #define RVT_AIP_QP_SUFFIX       0xffff
91 #define RVT_AIP_QP_PREFIX_MASK  0x00ff0000
92 #define RVT_AIP_QP_PREFIX_SHIFT 16
93 #define RVT_AIP_QP_BASE         (u32)(RVT_AIP_QP_PREFIX << \
94 				      RVT_AIP_QP_PREFIX_SHIFT)
95 #define RVT_AIP_QPN_MAX         BIT(RVT_AIP_QP_PREFIX_SHIFT)
96 #define RVT_AIP_QP_MAX          (u32)(RVT_AIP_QP_BASE + RVT_AIP_QPN_MAX - 1)
97 
98 /*
99  * Bit definitions for s_flags.
100  *
101  * RVT_S_SIGNAL_REQ_WR - set if QP send WRs contain completion signaled
102  * RVT_S_BUSY - send tasklet is processing the QP
103  * RVT_S_TIMER - the RC retry timer is active
104  * RVT_S_ACK_PENDING - an ACK is waiting to be sent after RDMA read/atomics
105  * RVT_S_WAIT_FENCE - waiting for all prior RDMA read or atomic SWQEs
106  *                         before processing the next SWQE
107  * RVT_S_WAIT_RDMAR - waiting for a RDMA read or atomic SWQE to complete
108  *                         before processing the next SWQE
109  * RVT_S_WAIT_RNR - waiting for RNR timeout
110  * RVT_S_WAIT_SSN_CREDIT - waiting for RC credits to process next SWQE
111  * RVT_S_WAIT_DMA - waiting for send DMA queue to drain before generating
112  *                  next send completion entry not via send DMA
113  * RVT_S_WAIT_PIO - waiting for a send buffer to be available
114  * RVT_S_WAIT_TX - waiting for a struct verbs_txreq to be available
115  * RVT_S_WAIT_DMA_DESC - waiting for DMA descriptors to be available
116  * RVT_S_WAIT_KMEM - waiting for kernel memory to be available
117  * RVT_S_WAIT_PSN - waiting for a packet to exit the send DMA queue
118  * RVT_S_WAIT_ACK - waiting for an ACK packet before sending more requests
119  * RVT_S_SEND_ONE - send one packet, request ACK, then wait for ACK
120  * RVT_S_ECN - a BECN was queued to the send engine
121  * RVT_S_MAX_BIT_MASK - The max bit that can be used by rdmavt
122  */
123 #define RVT_S_SIGNAL_REQ_WR	0x0001
124 #define RVT_S_BUSY		0x0002
125 #define RVT_S_TIMER		0x0004
126 #define RVT_S_RESP_PENDING	0x0008
127 #define RVT_S_ACK_PENDING	0x0010
128 #define RVT_S_WAIT_FENCE	0x0020
129 #define RVT_S_WAIT_RDMAR	0x0040
130 #define RVT_S_WAIT_RNR		0x0080
131 #define RVT_S_WAIT_SSN_CREDIT	0x0100
132 #define RVT_S_WAIT_DMA		0x0200
133 #define RVT_S_WAIT_PIO		0x0400
134 #define RVT_S_WAIT_TX		0x0800
135 #define RVT_S_WAIT_DMA_DESC	0x1000
136 #define RVT_S_WAIT_KMEM		0x2000
137 #define RVT_S_WAIT_PSN		0x4000
138 #define RVT_S_WAIT_ACK		0x8000
139 #define RVT_S_SEND_ONE		0x10000
140 #define RVT_S_UNLIMITED_CREDIT	0x20000
141 #define RVT_S_ECN		0x40000
142 #define RVT_S_MAX_BIT_MASK	0x800000
143 
144 /*
145  * Drivers should use s_flags starting with bit 31 down to the bit next to
146  * RVT_S_MAX_BIT_MASK
147  */
148 
149 /*
150  * Wait flags that would prevent any packet type from being sent.
151  */
152 #define RVT_S_ANY_WAIT_IO \
153 	(RVT_S_WAIT_PIO | RVT_S_WAIT_TX | \
154 	 RVT_S_WAIT_DMA_DESC | RVT_S_WAIT_KMEM)
155 
156 /*
157  * Wait flags that would prevent send work requests from making progress.
158  */
159 #define RVT_S_ANY_WAIT_SEND (RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR | \
160 	RVT_S_WAIT_RNR | RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_DMA | \
161 	RVT_S_WAIT_PSN | RVT_S_WAIT_ACK)
162 
163 #define RVT_S_ANY_WAIT (RVT_S_ANY_WAIT_IO | RVT_S_ANY_WAIT_SEND)
164 
165 /* Number of bits to pay attention to in the opcode for checking qp type */
166 #define RVT_OPCODE_QP_MASK 0xE0
167 
168 /* Flags for checking QP state (see ib_rvt_state_ops[]) */
169 #define RVT_POST_SEND_OK                0x01
170 #define RVT_POST_RECV_OK                0x02
171 #define RVT_PROCESS_RECV_OK             0x04
172 #define RVT_PROCESS_SEND_OK             0x08
173 #define RVT_PROCESS_NEXT_SEND_OK        0x10
174 #define RVT_FLUSH_SEND			0x20
175 #define RVT_FLUSH_RECV			0x40
176 #define RVT_PROCESS_OR_FLUSH_SEND \
177 	(RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND)
178 #define RVT_SEND_OR_FLUSH_OR_RECV_OK \
179 	(RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND | RVT_PROCESS_RECV_OK)
180 
181 /*
182  * Internal send flags
183  */
184 #define RVT_SEND_RESERVE_USED           IB_SEND_RESERVED_START
185 #define RVT_SEND_COMPLETION_ONLY	(IB_SEND_RESERVED_START << 1)
186 
187 /**
188  * rvt_ud_wr - IB UD work plus AH cache
189  * @wr: valid IB work request
190  * @attr: pointer to an allocated AH attribute
191  *
192  * Special case the UD WR so we can keep track of the AH attributes.
193  *
194  * NOTE: This data structure is stricly ordered wr then attr. I.e the attr
195  * MUST come after wr.  The ib_ud_wr is sized and copied in rvt_post_one_wr.
196  * The copy assumes that wr is first.
197  */
198 struct rvt_ud_wr {
199 	struct ib_ud_wr wr;
200 	struct rdma_ah_attr *attr;
201 };
202 
203 /*
204  * Send work request queue entry.
205  * The size of the sg_list is determined when the QP is created and stored
206  * in qp->s_max_sge.
207  */
208 struct rvt_swqe {
209 	union {
210 		struct ib_send_wr wr;   /* don't use wr.sg_list */
211 		struct rvt_ud_wr ud_wr;
212 		struct ib_reg_wr reg_wr;
213 		struct ib_rdma_wr rdma_wr;
214 		struct ib_atomic_wr atomic_wr;
215 	};
216 	u32 psn;                /* first packet sequence number */
217 	u32 lpsn;               /* last packet sequence number */
218 	u32 ssn;                /* send sequence number */
219 	u32 length;             /* total length of data in sg_list */
220 	void *priv;             /* driver dependent field */
221 	struct rvt_sge sg_list[];
222 };
223 
224 /**
225  * struct rvt_krwq - kernel struct receive work request
226  * @p_lock: lock to protect producer of the kernel buffer
227  * @head: index of next entry to fill
228  * @c_lock:lock to protect consumer of the kernel buffer
229  * @tail: index of next entry to pull
230  * @count: count is aproximate of total receive enteries posted
231  * @rvt_rwqe: struct of receive work request queue entry
232  *
233  * This structure is used to contain the head pointer,
234  * tail pointer and receive work queue entries for kernel
235  * mode user.
236  */
237 struct rvt_krwq {
238 	spinlock_t p_lock;	/* protect producer */
239 	u32 head;               /* new work requests posted to the head */
240 
241 	/* protect consumer */
242 	spinlock_t c_lock ____cacheline_aligned_in_smp;
243 	u32 tail;               /* receives pull requests from here. */
244 	u32 count;		/* approx count of receive entries posted */
245 	struct rvt_rwqe *curr_wq;
246 	struct rvt_rwqe wq[];
247 };
248 
249 /*
250  * rvt_get_swqe_ah - Return the pointer to the struct rvt_ah
251  * @swqe: valid Send WQE
252  *
253  */
254 static inline struct rvt_ah *rvt_get_swqe_ah(struct rvt_swqe *swqe)
255 {
256 	return ibah_to_rvtah(swqe->ud_wr.wr.ah);
257 }
258 
259 /**
260  * rvt_get_swqe_ah_attr - Return the cached ah attribute information
261  * @swqe: valid Send WQE
262  *
263  */
264 static inline struct rdma_ah_attr *rvt_get_swqe_ah_attr(struct rvt_swqe *swqe)
265 {
266 	return swqe->ud_wr.attr;
267 }
268 
269 /**
270  * rvt_get_swqe_remote_qpn - Access the remote QPN value
271  * @swqe: valid Send WQE
272  *
273  */
274 static inline u32 rvt_get_swqe_remote_qpn(struct rvt_swqe *swqe)
275 {
276 	return swqe->ud_wr.wr.remote_qpn;
277 }
278 
279 /**
280  * rvt_get_swqe_remote_qkey - Acces the remote qkey value
281  * @swqe: valid Send WQE
282  *
283  */
284 static inline u32 rvt_get_swqe_remote_qkey(struct rvt_swqe *swqe)
285 {
286 	return swqe->ud_wr.wr.remote_qkey;
287 }
288 
289 /**
290  * rvt_get_swqe_pkey_index - Access the pkey index
291  * @swqe: valid Send WQE
292  *
293  */
294 static inline u16 rvt_get_swqe_pkey_index(struct rvt_swqe *swqe)
295 {
296 	return swqe->ud_wr.wr.pkey_index;
297 }
298 
299 struct rvt_rq {
300 	struct rvt_rwq *wq;
301 	struct rvt_krwq *kwq;
302 	u32 size;               /* size of RWQE array */
303 	u8 max_sge;
304 	/* protect changes in this struct */
305 	spinlock_t lock ____cacheline_aligned_in_smp;
306 };
307 
308 /**
309  * rvt_get_rq_count - count numbers of request work queue entries
310  * in circular buffer
311  * @rq: data structure for request queue entry
312  * @head: head indices of the circular buffer
313  * @tail: tail indices of the circular buffer
314  *
315  * Return - total number of entries in the Receive Queue
316  */
317 
318 static inline u32 rvt_get_rq_count(struct rvt_rq *rq, u32 head, u32 tail)
319 {
320 	u32 count = head - tail;
321 
322 	if ((s32)count < 0)
323 		count += rq->size;
324 	return count;
325 }
326 
327 /*
328  * This structure holds the information that the send tasklet needs
329  * to send a RDMA read response or atomic operation.
330  */
331 struct rvt_ack_entry {
332 	struct rvt_sge rdma_sge;
333 	u64 atomic_data;
334 	u32 psn;
335 	u32 lpsn;
336 	u8 opcode;
337 	u8 sent;
338 	void *priv;
339 };
340 
341 #define	RC_QP_SCALING_INTERVAL	5
342 
343 #define RVT_OPERATION_PRIV        0x00000001
344 #define RVT_OPERATION_ATOMIC      0x00000002
345 #define RVT_OPERATION_ATOMIC_SGE  0x00000004
346 #define RVT_OPERATION_LOCAL       0x00000008
347 #define RVT_OPERATION_USE_RESERVE 0x00000010
348 #define RVT_OPERATION_IGN_RNR_CNT 0x00000020
349 
350 #define RVT_OPERATION_MAX (IB_WR_RESERVED10 + 1)
351 
352 /**
353  * rvt_operation_params - op table entry
354  * @length - the length to copy into the swqe entry
355  * @qpt_support - a bit mask indicating QP type support
356  * @flags - RVT_OPERATION flags (see above)
357  *
358  * This supports table driven post send so that
359  * the driver can have differing an potentially
360  * different sets of operations.
361  *
362  **/
363 
364 struct rvt_operation_params {
365 	size_t length;
366 	u32 qpt_support;
367 	u32 flags;
368 };
369 
370 /*
371  * Common variables are protected by both r_rq.lock and s_lock in that order
372  * which only happens in modify_qp() or changing the QP 'state'.
373  */
374 struct rvt_qp {
375 	struct ib_qp ibqp;
376 	void *priv; /* Driver private data */
377 	/* read mostly fields above and below */
378 	struct rdma_ah_attr remote_ah_attr;
379 	struct rdma_ah_attr alt_ah_attr;
380 	struct rvt_qp __rcu *next;           /* link list for QPN hash table */
381 	struct rvt_swqe *s_wq;  /* send work queue */
382 	struct rvt_mmap_info *ip;
383 
384 	unsigned long timeout_jiffies;  /* computed from timeout */
385 
386 	int srate_mbps;		/* s_srate (below) converted to Mbit/s */
387 	pid_t pid;		/* pid for user mode QPs */
388 	u32 remote_qpn;
389 	u32 qkey;               /* QKEY for this QP (for UD or RD) */
390 	u32 s_size;             /* send work queue size */
391 
392 	u16 pmtu;		/* decoded from path_mtu */
393 	u8 log_pmtu;		/* shift for pmtu */
394 	u8 state;               /* QP state */
395 	u8 allowed_ops;		/* high order bits of allowed opcodes */
396 	u8 qp_access_flags;
397 	u8 alt_timeout;         /* Alternate path timeout for this QP */
398 	u8 timeout;             /* Timeout for this QP */
399 	u8 s_srate;
400 	u8 s_mig_state;
401 	u8 port_num;
402 	u8 s_pkey_index;        /* PKEY index to use */
403 	u8 s_alt_pkey_index;    /* Alternate path PKEY index to use */
404 	u8 r_max_rd_atomic;     /* max number of RDMA read/atomic to receive */
405 	u8 s_max_rd_atomic;     /* max number of RDMA read/atomic to send */
406 	u8 s_retry_cnt;         /* number of times to retry */
407 	u8 s_rnr_retry_cnt;
408 	u8 r_min_rnr_timer;     /* retry timeout value for RNR NAKs */
409 	u8 s_max_sge;           /* size of s_wq->sg_list */
410 	u8 s_draining;
411 
412 	/* start of read/write fields */
413 	atomic_t refcount ____cacheline_aligned_in_smp;
414 	wait_queue_head_t wait;
415 
416 	struct rvt_ack_entry *s_ack_queue;
417 	struct rvt_sge_state s_rdma_read_sge;
418 
419 	spinlock_t r_lock ____cacheline_aligned_in_smp;      /* used for APM */
420 	u32 r_psn;              /* expected rcv packet sequence number */
421 	unsigned long r_aflags;
422 	u64 r_wr_id;            /* ID for current receive WQE */
423 	u32 r_ack_psn;          /* PSN for next ACK or atomic ACK */
424 	u32 r_len;              /* total length of r_sge */
425 	u32 r_rcv_len;          /* receive data len processed */
426 	u32 r_msn;              /* message sequence number */
427 
428 	u8 r_state;             /* opcode of last packet received */
429 	u8 r_flags;
430 	u8 r_head_ack_queue;    /* index into s_ack_queue[] */
431 	u8 r_adefered;          /* defered ack count */
432 
433 	struct list_head rspwait;       /* link for waiting to respond */
434 
435 	struct rvt_sge_state r_sge;     /* current receive data */
436 	struct rvt_rq r_rq;             /* receive work queue */
437 
438 	/* post send line */
439 	spinlock_t s_hlock ____cacheline_aligned_in_smp;
440 	u32 s_head;             /* new entries added here */
441 	u32 s_next_psn;         /* PSN for next request */
442 	u32 s_avail;            /* number of entries avail */
443 	u32 s_ssn;              /* SSN of tail entry */
444 	atomic_t s_reserved_used; /* reserved entries in use */
445 
446 	spinlock_t s_lock ____cacheline_aligned_in_smp;
447 	u32 s_flags;
448 	struct rvt_sge_state *s_cur_sge;
449 	struct rvt_swqe *s_wqe;
450 	struct rvt_sge_state s_sge;     /* current send request data */
451 	struct rvt_mregion *s_rdma_mr;
452 	u32 s_len;              /* total length of s_sge */
453 	u32 s_rdma_read_len;    /* total length of s_rdma_read_sge */
454 	u32 s_last_psn;         /* last response PSN processed */
455 	u32 s_sending_psn;      /* lowest PSN that is being sent */
456 	u32 s_sending_hpsn;     /* highest PSN that is being sent */
457 	u32 s_psn;              /* current packet sequence number */
458 	u32 s_ack_rdma_psn;     /* PSN for sending RDMA read responses */
459 	u32 s_ack_psn;          /* PSN for acking sends and RDMA writes */
460 	u32 s_tail;             /* next entry to process */
461 	u32 s_cur;              /* current work queue entry */
462 	u32 s_acked;            /* last un-ACK'ed entry */
463 	u32 s_last;             /* last completed entry */
464 	u32 s_lsn;              /* limit sequence number (credit) */
465 	u32 s_ahgpsn;           /* set to the psn in the copy of the header */
466 	u16 s_cur_size;         /* size of send packet in bytes */
467 	u16 s_rdma_ack_cnt;
468 	u8 s_hdrwords;         /* size of s_hdr in 32 bit words */
469 	s8 s_ahgidx;
470 	u8 s_state;             /* opcode of last packet sent */
471 	u8 s_ack_state;         /* opcode of packet to ACK */
472 	u8 s_nak_state;         /* non-zero if NAK is pending */
473 	u8 r_nak_state;         /* non-zero if NAK is pending */
474 	u8 s_retry;             /* requester retry counter */
475 	u8 s_rnr_retry;         /* requester RNR retry counter */
476 	u8 s_num_rd_atomic;     /* number of RDMA read/atomic pending */
477 	u8 s_tail_ack_queue;    /* index into s_ack_queue[] */
478 	u8 s_acked_ack_queue;   /* index into s_ack_queue[] */
479 
480 	struct rvt_sge_state s_ack_rdma_sge;
481 	struct timer_list s_timer;
482 	struct hrtimer s_rnr_timer;
483 
484 	atomic_t local_ops_pending; /* number of fast_reg/local_inv reqs */
485 
486 	/*
487 	 * This sge list MUST be last. Do not add anything below here.
488 	 */
489 	struct rvt_sge r_sg_list[] /* verified SGEs */
490 		____cacheline_aligned_in_smp;
491 };
492 
493 struct rvt_srq {
494 	struct ib_srq ibsrq;
495 	struct rvt_rq rq;
496 	struct rvt_mmap_info *ip;
497 	/* send signal when number of RWQEs < limit */
498 	u32 limit;
499 };
500 
501 static inline struct rvt_srq *ibsrq_to_rvtsrq(struct ib_srq *ibsrq)
502 {
503 	return container_of(ibsrq, struct rvt_srq, ibsrq);
504 }
505 
506 static inline struct rvt_qp *ibqp_to_rvtqp(struct ib_qp *ibqp)
507 {
508 	return container_of(ibqp, struct rvt_qp, ibqp);
509 }
510 
511 #define RVT_QPN_MAX                 BIT(24)
512 #define RVT_QPNMAP_ENTRIES          (RVT_QPN_MAX / PAGE_SIZE / BITS_PER_BYTE)
513 #define RVT_BITS_PER_PAGE           (PAGE_SIZE * BITS_PER_BYTE)
514 #define RVT_BITS_PER_PAGE_MASK      (RVT_BITS_PER_PAGE - 1)
515 #define RVT_QPN_MASK		    IB_QPN_MASK
516 
517 /*
518  * QPN-map pages start out as NULL, they get allocated upon
519  * first use and are never deallocated. This way,
520  * large bitmaps are not allocated unless large numbers of QPs are used.
521  */
522 struct rvt_qpn_map {
523 	void *page;
524 };
525 
526 struct rvt_qpn_table {
527 	spinlock_t lock; /* protect changes to the qp table */
528 	unsigned flags;         /* flags for QP0/1 allocated for each port */
529 	u32 last;               /* last QP number allocated */
530 	u32 nmaps;              /* size of the map table */
531 	u16 limit;
532 	u8  incr;
533 	/* bit map of free QP numbers other than 0/1 */
534 	struct rvt_qpn_map map[RVT_QPNMAP_ENTRIES];
535 };
536 
537 struct rvt_qp_ibdev {
538 	u32 qp_table_size;
539 	u32 qp_table_bits;
540 	struct rvt_qp __rcu **qp_table;
541 	spinlock_t qpt_lock; /* qptable lock */
542 	struct rvt_qpn_table qpn_table;
543 };
544 
545 /*
546  * There is one struct rvt_mcast for each multicast GID.
547  * All attached QPs are then stored as a list of
548  * struct rvt_mcast_qp.
549  */
550 struct rvt_mcast_qp {
551 	struct list_head list;
552 	struct rvt_qp *qp;
553 };
554 
555 struct rvt_mcast_addr {
556 	union ib_gid mgid;
557 	u16 lid;
558 };
559 
560 struct rvt_mcast {
561 	struct rb_node rb_node;
562 	struct rvt_mcast_addr mcast_addr;
563 	struct list_head qp_list;
564 	wait_queue_head_t wait;
565 	atomic_t refcount;
566 	int n_attached;
567 };
568 
569 /*
570  * Since struct rvt_swqe is not a fixed size, we can't simply index into
571  * struct rvt_qp.s_wq.  This function does the array index computation.
572  */
573 static inline struct rvt_swqe *rvt_get_swqe_ptr(struct rvt_qp *qp,
574 						unsigned n)
575 {
576 	return (struct rvt_swqe *)((char *)qp->s_wq +
577 				     (sizeof(struct rvt_swqe) +
578 				      qp->s_max_sge *
579 				      sizeof(struct rvt_sge)) * n);
580 }
581 
582 /*
583  * Since struct rvt_rwqe is not a fixed size, we can't simply index into
584  * struct rvt_rwq.wq.  This function does the array index computation.
585  */
586 static inline struct rvt_rwqe *rvt_get_rwqe_ptr(struct rvt_rq *rq, unsigned n)
587 {
588 	return (struct rvt_rwqe *)
589 		((char *)rq->kwq->curr_wq +
590 		 (sizeof(struct rvt_rwqe) +
591 		  rq->max_sge * sizeof(struct ib_sge)) * n);
592 }
593 
594 /**
595  * rvt_is_user_qp - return if this is user mode QP
596  * @qp - the target QP
597  */
598 static inline bool rvt_is_user_qp(struct rvt_qp *qp)
599 {
600 	return !!qp->pid;
601 }
602 
603 /**
604  * rvt_get_qp - get a QP reference
605  * @qp - the QP to hold
606  */
607 static inline void rvt_get_qp(struct rvt_qp *qp)
608 {
609 	atomic_inc(&qp->refcount);
610 }
611 
612 /**
613  * rvt_put_qp - release a QP reference
614  * @qp - the QP to release
615  */
616 static inline void rvt_put_qp(struct rvt_qp *qp)
617 {
618 	if (qp && atomic_dec_and_test(&qp->refcount))
619 		wake_up(&qp->wait);
620 }
621 
622 /**
623  * rvt_put_swqe - drop mr refs held by swqe
624  * @wqe - the send wqe
625  *
626  * This drops any mr references held by the swqe
627  */
628 static inline void rvt_put_swqe(struct rvt_swqe *wqe)
629 {
630 	int i;
631 
632 	for (i = 0; i < wqe->wr.num_sge; i++) {
633 		struct rvt_sge *sge = &wqe->sg_list[i];
634 
635 		rvt_put_mr(sge->mr);
636 	}
637 }
638 
639 /**
640  * rvt_qp_wqe_reserve - reserve operation
641  * @qp - the rvt qp
642  * @wqe - the send wqe
643  *
644  * This routine used in post send to record
645  * a wqe relative reserved operation use.
646  */
647 static inline void rvt_qp_wqe_reserve(
648 	struct rvt_qp *qp,
649 	struct rvt_swqe *wqe)
650 {
651 	atomic_inc(&qp->s_reserved_used);
652 }
653 
654 /**
655  * rvt_qp_wqe_unreserve - clean reserved operation
656  * @qp - the rvt qp
657  * @flags - send wqe flags
658  *
659  * This decrements the reserve use count.
660  *
661  * This call MUST precede the change to
662  * s_last to insure that post send sees a stable
663  * s_avail.
664  *
665  * An smp_mp__after_atomic() is used to insure
666  * the compiler does not juggle the order of the s_last
667  * ring index and the decrementing of s_reserved_used.
668  */
669 static inline void rvt_qp_wqe_unreserve(struct rvt_qp *qp, int flags)
670 {
671 	if (unlikely(flags & RVT_SEND_RESERVE_USED)) {
672 		atomic_dec(&qp->s_reserved_used);
673 		/* insure no compiler re-order up to s_last change */
674 		smp_mb__after_atomic();
675 	}
676 }
677 
678 extern const enum ib_wc_opcode ib_rvt_wc_opcode[];
679 
680 /*
681  * Compare the lower 24 bits of the msn values.
682  * Returns an integer <, ==, or > than zero.
683  */
684 static inline int rvt_cmp_msn(u32 a, u32 b)
685 {
686 	return (((int)a) - ((int)b)) << 8;
687 }
688 
689 __be32 rvt_compute_aeth(struct rvt_qp *qp);
690 
691 void rvt_get_credit(struct rvt_qp *qp, u32 aeth);
692 
693 u32 rvt_restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe, u32 len);
694 
695 /**
696  * rvt_div_round_up_mtu - round up divide
697  * @qp - the qp pair
698  * @len - the length
699  *
700  * Perform a shift based mtu round up divide
701  */
702 static inline u32 rvt_div_round_up_mtu(struct rvt_qp *qp, u32 len)
703 {
704 	return (len + qp->pmtu - 1) >> qp->log_pmtu;
705 }
706 
707 /**
708  * @qp - the qp pair
709  * @len - the length
710  *
711  * Perform a shift based mtu divide
712  */
713 static inline u32 rvt_div_mtu(struct rvt_qp *qp, u32 len)
714 {
715 	return len >> qp->log_pmtu;
716 }
717 
718 /**
719  * rvt_timeout_to_jiffies - Convert a ULP timeout input into jiffies
720  * @timeout - timeout input(0 - 31).
721  *
722  * Return a timeout value in jiffies.
723  */
724 static inline unsigned long rvt_timeout_to_jiffies(u8 timeout)
725 {
726 	if (timeout > 31)
727 		timeout = 31;
728 
729 	return usecs_to_jiffies(1U << timeout) * 4096UL / 1000UL;
730 }
731 
732 /**
733  * rvt_lookup_qpn - return the QP with the given QPN
734  * @ibp: the ibport
735  * @qpn: the QP number to look up
736  *
737  * The caller must hold the rcu_read_lock(), and keep the lock until
738  * the returned qp is no longer in use.
739  */
740 static inline struct rvt_qp *rvt_lookup_qpn(struct rvt_dev_info *rdi,
741 					    struct rvt_ibport *rvp,
742 					    u32 qpn) __must_hold(RCU)
743 {
744 	struct rvt_qp *qp = NULL;
745 
746 	if (unlikely(qpn <= 1)) {
747 		qp = rcu_dereference(rvp->qp[qpn]);
748 	} else {
749 		u32 n = hash_32(qpn, rdi->qp_dev->qp_table_bits);
750 
751 		for (qp = rcu_dereference(rdi->qp_dev->qp_table[n]); qp;
752 			qp = rcu_dereference(qp->next))
753 			if (qp->ibqp.qp_num == qpn)
754 				break;
755 	}
756 	return qp;
757 }
758 
759 /**
760  * rvt_mod_retry_timer - mod a retry timer
761  * @qp - the QP
762  * @shift - timeout shift to wait for multiple packets
763  * Modify a potentially already running retry timer
764  */
765 static inline void rvt_mod_retry_timer_ext(struct rvt_qp *qp, u8 shift)
766 {
767 	struct ib_qp *ibqp = &qp->ibqp;
768 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
769 
770 	lockdep_assert_held(&qp->s_lock);
771 	qp->s_flags |= RVT_S_TIMER;
772 	/* 4.096 usec. * (1 << qp->timeout) */
773 	mod_timer(&qp->s_timer, jiffies + rdi->busy_jiffies +
774 		  (qp->timeout_jiffies << shift));
775 }
776 
777 static inline void rvt_mod_retry_timer(struct rvt_qp *qp)
778 {
779 	return rvt_mod_retry_timer_ext(qp, 0);
780 }
781 
782 /**
783  * rvt_put_qp_swqe - drop refs held by swqe
784  * @qp: the send qp
785  * @wqe: the send wqe
786  *
787  * This drops any references held by the swqe
788  */
789 static inline void rvt_put_qp_swqe(struct rvt_qp *qp, struct rvt_swqe *wqe)
790 {
791 	rvt_put_swqe(wqe);
792 	if (qp->allowed_ops == IB_OPCODE_UD)
793 		rdma_destroy_ah_attr(wqe->ud_wr.attr);
794 }
795 
796 /**
797  * rvt_qp_sqwe_incr - increment ring index
798  * @qp: the qp
799  * @val: the starting value
800  *
801  * Return: the new value wrapping as appropriate
802  */
803 static inline u32
804 rvt_qp_swqe_incr(struct rvt_qp *qp, u32 val)
805 {
806 	if (++val >= qp->s_size)
807 		val = 0;
808 	return val;
809 }
810 
811 int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err);
812 
813 /**
814  * rvt_recv_cq - add a new entry to completion queue
815  *			by receive queue
816  * @qp: receive queue
817  * @wc: work completion entry to add
818  * @solicited: true if @entry is solicited
819  *
820  * This is wrapper function for rvt_enter_cq function call by
821  * receive queue. If rvt_cq_enter return false, it means cq is
822  * full and the qp is put into error state.
823  */
824 static inline void rvt_recv_cq(struct rvt_qp *qp, struct ib_wc *wc,
825 			       bool solicited)
826 {
827 	struct rvt_cq *cq = ibcq_to_rvtcq(qp->ibqp.recv_cq);
828 
829 	if (unlikely(!rvt_cq_enter(cq, wc, solicited)))
830 		rvt_error_qp(qp, IB_WC_LOC_QP_OP_ERR);
831 }
832 
833 /**
834  * rvt_send_cq - add a new entry to completion queue
835  *                        by send queue
836  * @qp: send queue
837  * @wc: work completion entry to add
838  * @solicited: true if @entry is solicited
839  *
840  * This is wrapper function for rvt_enter_cq function call by
841  * send queue. If rvt_cq_enter return false, it means cq is
842  * full and the qp is put into error state.
843  */
844 static inline void rvt_send_cq(struct rvt_qp *qp, struct ib_wc *wc,
845 			       bool solicited)
846 {
847 	struct rvt_cq *cq = ibcq_to_rvtcq(qp->ibqp.send_cq);
848 
849 	if (unlikely(!rvt_cq_enter(cq, wc, solicited)))
850 		rvt_error_qp(qp, IB_WC_LOC_QP_OP_ERR);
851 }
852 
853 /**
854  * rvt_qp_complete_swqe - insert send completion
855  * @qp - the qp
856  * @wqe - the send wqe
857  * @opcode - wc operation (driver dependent)
858  * @status - completion status
859  *
860  * Update the s_last information, and then insert a send
861  * completion into the completion
862  * queue if the qp indicates it should be done.
863  *
864  * See IBTA 10.7.3.1 for info on completion
865  * control.
866  *
867  * Return: new last
868  */
869 static inline u32
870 rvt_qp_complete_swqe(struct rvt_qp *qp,
871 		     struct rvt_swqe *wqe,
872 		     enum ib_wc_opcode opcode,
873 		     enum ib_wc_status status)
874 {
875 	bool need_completion;
876 	u64 wr_id;
877 	u32 byte_len, last;
878 	int flags = wqe->wr.send_flags;
879 
880 	rvt_qp_wqe_unreserve(qp, flags);
881 	rvt_put_qp_swqe(qp, wqe);
882 
883 	need_completion =
884 		!(flags & RVT_SEND_RESERVE_USED) &&
885 		(!(qp->s_flags & RVT_S_SIGNAL_REQ_WR) ||
886 		(flags & IB_SEND_SIGNALED) ||
887 		status != IB_WC_SUCCESS);
888 	if (need_completion) {
889 		wr_id = wqe->wr.wr_id;
890 		byte_len = wqe->length;
891 		/* above fields required before writing s_last */
892 	}
893 	last = rvt_qp_swqe_incr(qp, qp->s_last);
894 	/* see rvt_qp_is_avail() */
895 	smp_store_release(&qp->s_last, last);
896 	if (need_completion) {
897 		struct ib_wc w = {
898 			.wr_id = wr_id,
899 			.status = status,
900 			.opcode = opcode,
901 			.qp = &qp->ibqp,
902 			.byte_len = byte_len,
903 		};
904 		rvt_send_cq(qp, &w, status != IB_WC_SUCCESS);
905 	}
906 	return last;
907 }
908 
909 extern const int  ib_rvt_state_ops[];
910 
911 struct rvt_dev_info;
912 int rvt_get_rwqe(struct rvt_qp *qp, bool wr_id_only);
913 void rvt_comm_est(struct rvt_qp *qp);
914 void rvt_rc_error(struct rvt_qp *qp, enum ib_wc_status err);
915 unsigned long rvt_rnr_tbl_to_usec(u32 index);
916 enum hrtimer_restart rvt_rc_rnr_retry(struct hrtimer *t);
917 void rvt_add_rnr_timer(struct rvt_qp *qp, u32 aeth);
918 void rvt_del_timers_sync(struct rvt_qp *qp);
919 void rvt_stop_rc_timers(struct rvt_qp *qp);
920 void rvt_add_retry_timer_ext(struct rvt_qp *qp, u8 shift);
921 static inline void rvt_add_retry_timer(struct rvt_qp *qp)
922 {
923 	rvt_add_retry_timer_ext(qp, 0);
924 }
925 
926 void rvt_copy_sge(struct rvt_qp *qp, struct rvt_sge_state *ss,
927 		  void *data, u32 length,
928 		  bool release, bool copy_last);
929 void rvt_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
930 		       enum ib_wc_status status);
931 void rvt_ruc_loopback(struct rvt_qp *qp);
932 
933 /**
934  * struct rvt_qp_iter - the iterator for QPs
935  * @qp - the current QP
936  *
937  * This structure defines the current iterator
938  * state for sequenced access to all QPs relative
939  * to an rvt_dev_info.
940  */
941 struct rvt_qp_iter {
942 	struct rvt_qp *qp;
943 	/* private: backpointer */
944 	struct rvt_dev_info *rdi;
945 	/* private: callback routine */
946 	void (*cb)(struct rvt_qp *qp, u64 v);
947 	/* private: for arg to callback routine */
948 	u64 v;
949 	/* private: number of SMI,GSI QPs for device */
950 	int specials;
951 	/* private: current iterator index */
952 	int n;
953 };
954 
955 /**
956  * ib_cq_tail - Return tail index of cq buffer
957  * @send_cq - The cq for send
958  *
959  * This is called in qp_iter_print to get tail
960  * of cq buffer.
961  */
962 static inline u32 ib_cq_tail(struct ib_cq *send_cq)
963 {
964 	struct rvt_cq *cq = ibcq_to_rvtcq(send_cq);
965 
966 	return ibcq_to_rvtcq(send_cq)->ip ?
967 	       RDMA_READ_UAPI_ATOMIC(cq->queue->tail) :
968 	       ibcq_to_rvtcq(send_cq)->kqueue->tail;
969 }
970 
971 /**
972  * ib_cq_head - Return head index of cq buffer
973  * @send_cq - The cq for send
974  *
975  * This is called in qp_iter_print to get head
976  * of cq buffer.
977  */
978 static inline u32 ib_cq_head(struct ib_cq *send_cq)
979 {
980 	struct rvt_cq *cq = ibcq_to_rvtcq(send_cq);
981 
982 	return ibcq_to_rvtcq(send_cq)->ip ?
983 	       RDMA_READ_UAPI_ATOMIC(cq->queue->head) :
984 	       ibcq_to_rvtcq(send_cq)->kqueue->head;
985 }
986 
987 /**
988  * rvt_free_rq - free memory allocated for rvt_rq struct
989  * @rvt_rq: request queue data structure
990  *
991  * This function should only be called if the rvt_mmap_info()
992  * has not succeeded.
993  */
994 static inline void rvt_free_rq(struct rvt_rq *rq)
995 {
996 	kvfree(rq->kwq);
997 	rq->kwq = NULL;
998 	vfree(rq->wq);
999 	rq->wq = NULL;
1000 }
1001 
1002 /**
1003  * rvt_to_iport - Get the ibport pointer
1004  * @qp: the qp pointer
1005  *
1006  * This function returns the ibport pointer from the qp pointer.
1007  */
1008 static inline struct rvt_ibport *rvt_to_iport(struct rvt_qp *qp)
1009 {
1010 	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
1011 
1012 	return rdi->ports[qp->port_num - 1];
1013 }
1014 
1015 /**
1016  * rvt_rc_credit_avail - Check if there are enough RC credits for the request
1017  * @qp: the qp
1018  * @wqe: the request
1019  *
1020  * This function returns false when there are not enough credits for the given
1021  * request and true otherwise.
1022  */
1023 static inline bool rvt_rc_credit_avail(struct rvt_qp *qp, struct rvt_swqe *wqe)
1024 {
1025 	lockdep_assert_held(&qp->s_lock);
1026 	if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) &&
1027 	    rvt_cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) {
1028 		struct rvt_ibport *rvp = rvt_to_iport(qp);
1029 
1030 		qp->s_flags |= RVT_S_WAIT_SSN_CREDIT;
1031 		rvp->n_rc_crwaits++;
1032 		return false;
1033 	}
1034 	return true;
1035 }
1036 
1037 struct rvt_qp_iter *rvt_qp_iter_init(struct rvt_dev_info *rdi,
1038 				     u64 v,
1039 				     void (*cb)(struct rvt_qp *qp, u64 v));
1040 int rvt_qp_iter_next(struct rvt_qp_iter *iter);
1041 void rvt_qp_iter(struct rvt_dev_info *rdi,
1042 		 u64 v,
1043 		 void (*cb)(struct rvt_qp *qp, u64 v));
1044 void rvt_qp_mr_clean(struct rvt_qp *qp, u32 lkey);
1045 #endif          /* DEF_RDMAVT_INCQP_H */
1046