xref: /linux/include/rdma/rdmavt_qp.h (revision 447e140e66fd226350b3ce86cffc965eaae4c856)
1 /* SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause */
2 /*
3  * Copyright(c) 2016 - 2020 Intel Corporation.
4  */
5 
6 #ifndef DEF_RDMAVT_INCQP_H
7 #define DEF_RDMAVT_INCQP_H
8 
9 #include <rdma/rdma_vt.h>
10 #include <rdma/ib_pack.h>
11 #include <rdma/ib_verbs.h>
12 #include <rdma/rdmavt_cq.h>
13 #include <rdma/rvt-abi.h>
14 #include <linux/vmalloc.h>
15 /*
16  * Atomic bit definitions for r_aflags.
17  */
18 #define RVT_R_WRID_VALID        0
19 #define RVT_R_REWIND_SGE        1
20 
21 /*
22  * Bit definitions for r_flags.
23  */
24 #define RVT_R_REUSE_SGE 0x01
25 #define RVT_R_RDMAR_SEQ 0x02
26 #define RVT_R_RSP_NAK   0x04
27 #define RVT_R_RSP_SEND  0x08
28 #define RVT_R_COMM_EST  0x10
29 
30 /*
31  * If a packet's QP[23:16] bits match this value, then it is
32  * a PSM packet and the hardware will expect a KDETH header
33  * following the BTH.
34  */
35 #define RVT_KDETH_QP_PREFIX       0x80
36 #define RVT_KDETH_QP_SUFFIX       0xffff
37 #define RVT_KDETH_QP_PREFIX_MASK  0x00ff0000
38 #define RVT_KDETH_QP_PREFIX_SHIFT 16
39 #define RVT_KDETH_QP_BASE         (u32)(RVT_KDETH_QP_PREFIX << \
40 					RVT_KDETH_QP_PREFIX_SHIFT)
41 #define RVT_KDETH_QP_MAX          (u32)(RVT_KDETH_QP_BASE + RVT_KDETH_QP_SUFFIX)
42 
43 /*
44  * If a packet's LNH == BTH and DEST QPN[23:16] in the BTH match this
45  * prefix value, then it is an AIP packet with a DETH containing the entropy
46  * value in byte 4 following the BTH.
47  */
48 #define RVT_AIP_QP_PREFIX       0x81
49 #define RVT_AIP_QP_SUFFIX       0xffff
50 #define RVT_AIP_QP_PREFIX_MASK  0x00ff0000
51 #define RVT_AIP_QP_PREFIX_SHIFT 16
52 #define RVT_AIP_QP_BASE         (u32)(RVT_AIP_QP_PREFIX << \
53 				      RVT_AIP_QP_PREFIX_SHIFT)
54 #define RVT_AIP_QPN_MAX         BIT(RVT_AIP_QP_PREFIX_SHIFT)
55 #define RVT_AIP_QP_MAX          (u32)(RVT_AIP_QP_BASE + RVT_AIP_QPN_MAX - 1)
56 
57 /*
58  * Bit definitions for s_flags.
59  *
60  * RVT_S_SIGNAL_REQ_WR - set if QP send WRs contain completion signaled
61  * RVT_S_BUSY - send tasklet is processing the QP
62  * RVT_S_TIMER - the RC retry timer is active
63  * RVT_S_ACK_PENDING - an ACK is waiting to be sent after RDMA read/atomics
64  * RVT_S_WAIT_FENCE - waiting for all prior RDMA read or atomic SWQEs
65  *                         before processing the next SWQE
66  * RVT_S_WAIT_RDMAR - waiting for a RDMA read or atomic SWQE to complete
67  *                         before processing the next SWQE
68  * RVT_S_WAIT_RNR - waiting for RNR timeout
69  * RVT_S_WAIT_SSN_CREDIT - waiting for RC credits to process next SWQE
70  * RVT_S_WAIT_DMA - waiting for send DMA queue to drain before generating
71  *                  next send completion entry not via send DMA
72  * RVT_S_WAIT_PIO - waiting for a send buffer to be available
73  * RVT_S_WAIT_TX - waiting for a struct verbs_txreq to be available
74  * RVT_S_WAIT_DMA_DESC - waiting for DMA descriptors to be available
75  * RVT_S_WAIT_KMEM - waiting for kernel memory to be available
76  * RVT_S_WAIT_PSN - waiting for a packet to exit the send DMA queue
77  * RVT_S_WAIT_ACK - waiting for an ACK packet before sending more requests
78  * RVT_S_SEND_ONE - send one packet, request ACK, then wait for ACK
79  * RVT_S_ECN - a BECN was queued to the send engine
80  * RVT_S_MAX_BIT_MASK - The max bit that can be used by rdmavt
81  */
82 #define RVT_S_SIGNAL_REQ_WR	0x0001
83 #define RVT_S_BUSY		0x0002
84 #define RVT_S_TIMER		0x0004
85 #define RVT_S_RESP_PENDING	0x0008
86 #define RVT_S_ACK_PENDING	0x0010
87 #define RVT_S_WAIT_FENCE	0x0020
88 #define RVT_S_WAIT_RDMAR	0x0040
89 #define RVT_S_WAIT_RNR		0x0080
90 #define RVT_S_WAIT_SSN_CREDIT	0x0100
91 #define RVT_S_WAIT_DMA		0x0200
92 #define RVT_S_WAIT_PIO		0x0400
93 #define RVT_S_WAIT_TX		0x0800
94 #define RVT_S_WAIT_DMA_DESC	0x1000
95 #define RVT_S_WAIT_KMEM		0x2000
96 #define RVT_S_WAIT_PSN		0x4000
97 #define RVT_S_WAIT_ACK		0x8000
98 #define RVT_S_SEND_ONE		0x10000
99 #define RVT_S_UNLIMITED_CREDIT	0x20000
100 #define RVT_S_ECN		0x40000
101 #define RVT_S_MAX_BIT_MASK	0x800000
102 
103 /*
104  * Drivers should use s_flags starting with bit 31 down to the bit next to
105  * RVT_S_MAX_BIT_MASK
106  */
107 
108 /*
109  * Wait flags that would prevent any packet type from being sent.
110  */
111 #define RVT_S_ANY_WAIT_IO \
112 	(RVT_S_WAIT_PIO | RVT_S_WAIT_TX | \
113 	 RVT_S_WAIT_DMA_DESC | RVT_S_WAIT_KMEM)
114 
115 /*
116  * Wait flags that would prevent send work requests from making progress.
117  */
118 #define RVT_S_ANY_WAIT_SEND (RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR | \
119 	RVT_S_WAIT_RNR | RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_DMA | \
120 	RVT_S_WAIT_PSN | RVT_S_WAIT_ACK)
121 
122 #define RVT_S_ANY_WAIT (RVT_S_ANY_WAIT_IO | RVT_S_ANY_WAIT_SEND)
123 
124 /* Number of bits to pay attention to in the opcode for checking qp type */
125 #define RVT_OPCODE_QP_MASK 0xE0
126 
127 /* Flags for checking QP state (see ib_rvt_state_ops[]) */
128 #define RVT_POST_SEND_OK                0x01
129 #define RVT_POST_RECV_OK                0x02
130 #define RVT_PROCESS_RECV_OK             0x04
131 #define RVT_PROCESS_SEND_OK             0x08
132 #define RVT_PROCESS_NEXT_SEND_OK        0x10
133 #define RVT_FLUSH_SEND			0x20
134 #define RVT_FLUSH_RECV			0x40
135 #define RVT_PROCESS_OR_FLUSH_SEND \
136 	(RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND)
137 #define RVT_SEND_OR_FLUSH_OR_RECV_OK \
138 	(RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND | RVT_PROCESS_RECV_OK)
139 
140 /*
141  * Internal send flags
142  */
143 #define RVT_SEND_RESERVE_USED           IB_SEND_RESERVED_START
144 #define RVT_SEND_COMPLETION_ONLY	(IB_SEND_RESERVED_START << 1)
145 
146 /**
147  * rvt_ud_wr - IB UD work plus AH cache
148  * @wr: valid IB work request
149  * @attr: pointer to an allocated AH attribute
150  *
151  * Special case the UD WR so we can keep track of the AH attributes.
152  *
153  * NOTE: This data structure is stricly ordered wr then attr. I.e the attr
154  * MUST come after wr.  The ib_ud_wr is sized and copied in rvt_post_one_wr.
155  * The copy assumes that wr is first.
156  */
157 struct rvt_ud_wr {
158 	struct ib_ud_wr wr;
159 	struct rdma_ah_attr *attr;
160 };
161 
162 /*
163  * Send work request queue entry.
164  * The size of the sg_list is determined when the QP is created and stored
165  * in qp->s_max_sge.
166  */
167 struct rvt_swqe {
168 	union {
169 		struct ib_send_wr wr;   /* don't use wr.sg_list */
170 		struct rvt_ud_wr ud_wr;
171 		struct ib_reg_wr reg_wr;
172 		struct ib_rdma_wr rdma_wr;
173 		struct ib_atomic_wr atomic_wr;
174 	};
175 	u32 psn;                /* first packet sequence number */
176 	u32 lpsn;               /* last packet sequence number */
177 	u32 ssn;                /* send sequence number */
178 	u32 length;             /* total length of data in sg_list */
179 	void *priv;             /* driver dependent field */
180 	struct rvt_sge sg_list[];
181 };
182 
183 /**
184  * struct rvt_krwq - kernel struct receive work request
185  * @p_lock: lock to protect producer of the kernel buffer
186  * @head: index of next entry to fill
187  * @c_lock:lock to protect consumer of the kernel buffer
188  * @tail: index of next entry to pull
189  * @count: count is aproximate of total receive enteries posted
190  * @rvt_rwqe: struct of receive work request queue entry
191  *
192  * This structure is used to contain the head pointer,
193  * tail pointer and receive work queue entries for kernel
194  * mode user.
195  */
196 struct rvt_krwq {
197 	spinlock_t p_lock;	/* protect producer */
198 	u32 head;               /* new work requests posted to the head */
199 
200 	/* protect consumer */
201 	spinlock_t c_lock ____cacheline_aligned_in_smp;
202 	u32 tail;               /* receives pull requests from here. */
203 	u32 count;		/* approx count of receive entries posted */
204 	struct rvt_rwqe *curr_wq;
205 	struct rvt_rwqe wq[];
206 };
207 
208 /*
209  * rvt_get_swqe_ah - Return the pointer to the struct rvt_ah
210  * @swqe: valid Send WQE
211  *
212  */
213 static inline struct rvt_ah *rvt_get_swqe_ah(struct rvt_swqe *swqe)
214 {
215 	return ibah_to_rvtah(swqe->ud_wr.wr.ah);
216 }
217 
218 /**
219  * rvt_get_swqe_ah_attr - Return the cached ah attribute information
220  * @swqe: valid Send WQE
221  *
222  */
223 static inline struct rdma_ah_attr *rvt_get_swqe_ah_attr(struct rvt_swqe *swqe)
224 {
225 	return swqe->ud_wr.attr;
226 }
227 
228 /**
229  * rvt_get_swqe_remote_qpn - Access the remote QPN value
230  * @swqe: valid Send WQE
231  *
232  */
233 static inline u32 rvt_get_swqe_remote_qpn(struct rvt_swqe *swqe)
234 {
235 	return swqe->ud_wr.wr.remote_qpn;
236 }
237 
238 /**
239  * rvt_get_swqe_remote_qkey - Acces the remote qkey value
240  * @swqe: valid Send WQE
241  *
242  */
243 static inline u32 rvt_get_swqe_remote_qkey(struct rvt_swqe *swqe)
244 {
245 	return swqe->ud_wr.wr.remote_qkey;
246 }
247 
248 /**
249  * rvt_get_swqe_pkey_index - Access the pkey index
250  * @swqe: valid Send WQE
251  *
252  */
253 static inline u16 rvt_get_swqe_pkey_index(struct rvt_swqe *swqe)
254 {
255 	return swqe->ud_wr.wr.pkey_index;
256 }
257 
258 struct rvt_rq {
259 	struct rvt_rwq *wq;
260 	struct rvt_krwq *kwq;
261 	u32 size;               /* size of RWQE array */
262 	u8 max_sge;
263 	/* protect changes in this struct */
264 	spinlock_t lock ____cacheline_aligned_in_smp;
265 };
266 
267 /**
268  * rvt_get_rq_count - count numbers of request work queue entries
269  * in circular buffer
270  * @rq: data structure for request queue entry
271  * @head: head indices of the circular buffer
272  * @tail: tail indices of the circular buffer
273  *
274  * Return - total number of entries in the Receive Queue
275  */
276 
277 static inline u32 rvt_get_rq_count(struct rvt_rq *rq, u32 head, u32 tail)
278 {
279 	u32 count = head - tail;
280 
281 	if ((s32)count < 0)
282 		count += rq->size;
283 	return count;
284 }
285 
286 /*
287  * This structure holds the information that the send tasklet needs
288  * to send a RDMA read response or atomic operation.
289  */
290 struct rvt_ack_entry {
291 	struct rvt_sge rdma_sge;
292 	u64 atomic_data;
293 	u32 psn;
294 	u32 lpsn;
295 	u8 opcode;
296 	u8 sent;
297 	void *priv;
298 };
299 
300 #define	RC_QP_SCALING_INTERVAL	5
301 
302 #define RVT_OPERATION_PRIV        0x00000001
303 #define RVT_OPERATION_ATOMIC      0x00000002
304 #define RVT_OPERATION_ATOMIC_SGE  0x00000004
305 #define RVT_OPERATION_LOCAL       0x00000008
306 #define RVT_OPERATION_USE_RESERVE 0x00000010
307 #define RVT_OPERATION_IGN_RNR_CNT 0x00000020
308 
309 #define RVT_OPERATION_MAX (IB_WR_RESERVED10 + 1)
310 
311 /**
312  * rvt_operation_params - op table entry
313  * @length - the length to copy into the swqe entry
314  * @qpt_support - a bit mask indicating QP type support
315  * @flags - RVT_OPERATION flags (see above)
316  *
317  * This supports table driven post send so that
318  * the driver can have differing an potentially
319  * different sets of operations.
320  *
321  **/
322 
323 struct rvt_operation_params {
324 	size_t length;
325 	u32 qpt_support;
326 	u32 flags;
327 };
328 
329 /*
330  * Common variables are protected by both r_rq.lock and s_lock in that order
331  * which only happens in modify_qp() or changing the QP 'state'.
332  */
333 struct rvt_qp {
334 	struct ib_qp ibqp;
335 	void *priv; /* Driver private data */
336 	/* read mostly fields above and below */
337 	struct rdma_ah_attr remote_ah_attr;
338 	struct rdma_ah_attr alt_ah_attr;
339 	struct rvt_qp __rcu *next;           /* link list for QPN hash table */
340 	struct rvt_swqe *s_wq;  /* send work queue */
341 	struct rvt_mmap_info *ip;
342 
343 	unsigned long timeout_jiffies;  /* computed from timeout */
344 
345 	int srate_mbps;		/* s_srate (below) converted to Mbit/s */
346 	pid_t pid;		/* pid for user mode QPs */
347 	u32 remote_qpn;
348 	u32 qkey;               /* QKEY for this QP (for UD or RD) */
349 	u32 s_size;             /* send work queue size */
350 
351 	u16 pmtu;		/* decoded from path_mtu */
352 	u8 log_pmtu;		/* shift for pmtu */
353 	u8 state;               /* QP state */
354 	u8 allowed_ops;		/* high order bits of allowed opcodes */
355 	u8 qp_access_flags;
356 	u8 alt_timeout;         /* Alternate path timeout for this QP */
357 	u8 timeout;             /* Timeout for this QP */
358 	u8 s_srate;
359 	u8 s_mig_state;
360 	u8 port_num;
361 	u8 s_pkey_index;        /* PKEY index to use */
362 	u8 s_alt_pkey_index;    /* Alternate path PKEY index to use */
363 	u8 r_max_rd_atomic;     /* max number of RDMA read/atomic to receive */
364 	u8 s_max_rd_atomic;     /* max number of RDMA read/atomic to send */
365 	u8 s_retry_cnt;         /* number of times to retry */
366 	u8 s_rnr_retry_cnt;
367 	u8 r_min_rnr_timer;     /* retry timeout value for RNR NAKs */
368 	u8 s_max_sge;           /* size of s_wq->sg_list */
369 	u8 s_draining;
370 
371 	/* start of read/write fields */
372 	atomic_t refcount ____cacheline_aligned_in_smp;
373 	wait_queue_head_t wait;
374 
375 	struct rvt_ack_entry *s_ack_queue;
376 	struct rvt_sge_state s_rdma_read_sge;
377 
378 	spinlock_t r_lock ____cacheline_aligned_in_smp;      /* used for APM */
379 	u32 r_psn;              /* expected rcv packet sequence number */
380 	unsigned long r_aflags;
381 	u64 r_wr_id;            /* ID for current receive WQE */
382 	u32 r_ack_psn;          /* PSN for next ACK or atomic ACK */
383 	u32 r_len;              /* total length of r_sge */
384 	u32 r_rcv_len;          /* receive data len processed */
385 	u32 r_msn;              /* message sequence number */
386 
387 	u8 r_state;             /* opcode of last packet received */
388 	u8 r_flags;
389 	u8 r_head_ack_queue;    /* index into s_ack_queue[] */
390 	u8 r_adefered;          /* defered ack count */
391 
392 	struct list_head rspwait;       /* link for waiting to respond */
393 
394 	struct rvt_sge_state r_sge;     /* current receive data */
395 	struct rvt_rq r_rq;             /* receive work queue */
396 
397 	/* post send line */
398 	spinlock_t s_hlock ____cacheline_aligned_in_smp;
399 	u32 s_head;             /* new entries added here */
400 	u32 s_next_psn;         /* PSN for next request */
401 	u32 s_avail;            /* number of entries avail */
402 	u32 s_ssn;              /* SSN of tail entry */
403 	atomic_t s_reserved_used; /* reserved entries in use */
404 
405 	spinlock_t s_lock ____cacheline_aligned_in_smp;
406 	u32 s_flags;
407 	struct rvt_sge_state *s_cur_sge;
408 	struct rvt_swqe *s_wqe;
409 	struct rvt_sge_state s_sge;     /* current send request data */
410 	struct rvt_mregion *s_rdma_mr;
411 	u32 s_len;              /* total length of s_sge */
412 	u32 s_rdma_read_len;    /* total length of s_rdma_read_sge */
413 	u32 s_last_psn;         /* last response PSN processed */
414 	u32 s_sending_psn;      /* lowest PSN that is being sent */
415 	u32 s_sending_hpsn;     /* highest PSN that is being sent */
416 	u32 s_psn;              /* current packet sequence number */
417 	u32 s_ack_rdma_psn;     /* PSN for sending RDMA read responses */
418 	u32 s_ack_psn;          /* PSN for acking sends and RDMA writes */
419 	u32 s_tail;             /* next entry to process */
420 	u32 s_cur;              /* current work queue entry */
421 	u32 s_acked;            /* last un-ACK'ed entry */
422 	u32 s_last;             /* last completed entry */
423 	u32 s_lsn;              /* limit sequence number (credit) */
424 	u32 s_ahgpsn;           /* set to the psn in the copy of the header */
425 	u16 s_cur_size;         /* size of send packet in bytes */
426 	u16 s_rdma_ack_cnt;
427 	u8 s_hdrwords;         /* size of s_hdr in 32 bit words */
428 	s8 s_ahgidx;
429 	u8 s_state;             /* opcode of last packet sent */
430 	u8 s_ack_state;         /* opcode of packet to ACK */
431 	u8 s_nak_state;         /* non-zero if NAK is pending */
432 	u8 r_nak_state;         /* non-zero if NAK is pending */
433 	u8 s_retry;             /* requester retry counter */
434 	u8 s_rnr_retry;         /* requester RNR retry counter */
435 	u8 s_num_rd_atomic;     /* number of RDMA read/atomic pending */
436 	u8 s_tail_ack_queue;    /* index into s_ack_queue[] */
437 	u8 s_acked_ack_queue;   /* index into s_ack_queue[] */
438 
439 	struct rvt_sge_state s_ack_rdma_sge;
440 	struct timer_list s_timer;
441 	struct hrtimer s_rnr_timer;
442 
443 	atomic_t local_ops_pending; /* number of fast_reg/local_inv reqs */
444 
445 	/*
446 	 * This sge list MUST be last. Do not add anything below here.
447 	 */
448 	struct rvt_sge *r_sg_list /* verified SGEs */
449 		____cacheline_aligned_in_smp;
450 };
451 
452 struct rvt_srq {
453 	struct ib_srq ibsrq;
454 	struct rvt_rq rq;
455 	struct rvt_mmap_info *ip;
456 	/* send signal when number of RWQEs < limit */
457 	u32 limit;
458 };
459 
460 static inline struct rvt_srq *ibsrq_to_rvtsrq(struct ib_srq *ibsrq)
461 {
462 	return container_of(ibsrq, struct rvt_srq, ibsrq);
463 }
464 
465 static inline struct rvt_qp *ibqp_to_rvtqp(struct ib_qp *ibqp)
466 {
467 	return container_of(ibqp, struct rvt_qp, ibqp);
468 }
469 
470 #define RVT_QPN_MAX                 BIT(24)
471 #define RVT_QPNMAP_ENTRIES          (RVT_QPN_MAX / PAGE_SIZE / BITS_PER_BYTE)
472 #define RVT_BITS_PER_PAGE           (PAGE_SIZE * BITS_PER_BYTE)
473 #define RVT_BITS_PER_PAGE_MASK      (RVT_BITS_PER_PAGE - 1)
474 #define RVT_QPN_MASK		    IB_QPN_MASK
475 
476 /*
477  * QPN-map pages start out as NULL, they get allocated upon
478  * first use and are never deallocated. This way,
479  * large bitmaps are not allocated unless large numbers of QPs are used.
480  */
481 struct rvt_qpn_map {
482 	void *page;
483 };
484 
485 struct rvt_qpn_table {
486 	spinlock_t lock; /* protect changes to the qp table */
487 	unsigned flags;         /* flags for QP0/1 allocated for each port */
488 	u32 last;               /* last QP number allocated */
489 	u32 nmaps;              /* size of the map table */
490 	u16 limit;
491 	u8  incr;
492 	/* bit map of free QP numbers other than 0/1 */
493 	struct rvt_qpn_map map[RVT_QPNMAP_ENTRIES];
494 };
495 
496 struct rvt_qp_ibdev {
497 	u32 qp_table_size;
498 	u32 qp_table_bits;
499 	struct rvt_qp __rcu **qp_table;
500 	spinlock_t qpt_lock; /* qptable lock */
501 	struct rvt_qpn_table qpn_table;
502 };
503 
504 /*
505  * There is one struct rvt_mcast for each multicast GID.
506  * All attached QPs are then stored as a list of
507  * struct rvt_mcast_qp.
508  */
509 struct rvt_mcast_qp {
510 	struct list_head list;
511 	struct rvt_qp *qp;
512 };
513 
514 struct rvt_mcast_addr {
515 	union ib_gid mgid;
516 	u16 lid;
517 };
518 
519 struct rvt_mcast {
520 	struct rb_node rb_node;
521 	struct rvt_mcast_addr mcast_addr;
522 	struct list_head qp_list;
523 	wait_queue_head_t wait;
524 	atomic_t refcount;
525 	int n_attached;
526 };
527 
528 /*
529  * Since struct rvt_swqe is not a fixed size, we can't simply index into
530  * struct rvt_qp.s_wq.  This function does the array index computation.
531  */
532 static inline struct rvt_swqe *rvt_get_swqe_ptr(struct rvt_qp *qp,
533 						unsigned n)
534 {
535 	return (struct rvt_swqe *)((char *)qp->s_wq +
536 				     (sizeof(struct rvt_swqe) +
537 				      qp->s_max_sge *
538 				      sizeof(struct rvt_sge)) * n);
539 }
540 
541 /*
542  * Since struct rvt_rwqe is not a fixed size, we can't simply index into
543  * struct rvt_rwq.wq.  This function does the array index computation.
544  */
545 static inline struct rvt_rwqe *rvt_get_rwqe_ptr(struct rvt_rq *rq, unsigned n)
546 {
547 	return (struct rvt_rwqe *)
548 		((char *)rq->kwq->curr_wq +
549 		 (sizeof(struct rvt_rwqe) +
550 		  rq->max_sge * sizeof(struct ib_sge)) * n);
551 }
552 
553 /**
554  * rvt_is_user_qp - return if this is user mode QP
555  * @qp - the target QP
556  */
557 static inline bool rvt_is_user_qp(struct rvt_qp *qp)
558 {
559 	return !!qp->pid;
560 }
561 
562 /**
563  * rvt_get_qp - get a QP reference
564  * @qp - the QP to hold
565  */
566 static inline void rvt_get_qp(struct rvt_qp *qp)
567 {
568 	atomic_inc(&qp->refcount);
569 }
570 
571 /**
572  * rvt_put_qp - release a QP reference
573  * @qp - the QP to release
574  */
575 static inline void rvt_put_qp(struct rvt_qp *qp)
576 {
577 	if (qp && atomic_dec_and_test(&qp->refcount))
578 		wake_up(&qp->wait);
579 }
580 
581 /**
582  * rvt_put_swqe - drop mr refs held by swqe
583  * @wqe - the send wqe
584  *
585  * This drops any mr references held by the swqe
586  */
587 static inline void rvt_put_swqe(struct rvt_swqe *wqe)
588 {
589 	int i;
590 
591 	for (i = 0; i < wqe->wr.num_sge; i++) {
592 		struct rvt_sge *sge = &wqe->sg_list[i];
593 
594 		rvt_put_mr(sge->mr);
595 	}
596 }
597 
598 /**
599  * rvt_qp_wqe_reserve - reserve operation
600  * @qp - the rvt qp
601  * @wqe - the send wqe
602  *
603  * This routine used in post send to record
604  * a wqe relative reserved operation use.
605  */
606 static inline void rvt_qp_wqe_reserve(
607 	struct rvt_qp *qp,
608 	struct rvt_swqe *wqe)
609 {
610 	atomic_inc(&qp->s_reserved_used);
611 }
612 
613 /**
614  * rvt_qp_wqe_unreserve - clean reserved operation
615  * @qp - the rvt qp
616  * @flags - send wqe flags
617  *
618  * This decrements the reserve use count.
619  *
620  * This call MUST precede the change to
621  * s_last to insure that post send sees a stable
622  * s_avail.
623  *
624  * An smp_mp__after_atomic() is used to insure
625  * the compiler does not juggle the order of the s_last
626  * ring index and the decrementing of s_reserved_used.
627  */
628 static inline void rvt_qp_wqe_unreserve(struct rvt_qp *qp, int flags)
629 {
630 	if (unlikely(flags & RVT_SEND_RESERVE_USED)) {
631 		atomic_dec(&qp->s_reserved_used);
632 		/* insure no compiler re-order up to s_last change */
633 		smp_mb__after_atomic();
634 	}
635 }
636 
637 extern const enum ib_wc_opcode ib_rvt_wc_opcode[];
638 
639 /*
640  * Compare the lower 24 bits of the msn values.
641  * Returns an integer <, ==, or > than zero.
642  */
643 static inline int rvt_cmp_msn(u32 a, u32 b)
644 {
645 	return (((int)a) - ((int)b)) << 8;
646 }
647 
648 __be32 rvt_compute_aeth(struct rvt_qp *qp);
649 
650 void rvt_get_credit(struct rvt_qp *qp, u32 aeth);
651 
652 u32 rvt_restart_sge(struct rvt_sge_state *ss, struct rvt_swqe *wqe, u32 len);
653 
654 /**
655  * rvt_div_round_up_mtu - round up divide
656  * @qp - the qp pair
657  * @len - the length
658  *
659  * Perform a shift based mtu round up divide
660  */
661 static inline u32 rvt_div_round_up_mtu(struct rvt_qp *qp, u32 len)
662 {
663 	return (len + qp->pmtu - 1) >> qp->log_pmtu;
664 }
665 
666 /**
667  * @qp - the qp pair
668  * @len - the length
669  *
670  * Perform a shift based mtu divide
671  */
672 static inline u32 rvt_div_mtu(struct rvt_qp *qp, u32 len)
673 {
674 	return len >> qp->log_pmtu;
675 }
676 
677 /**
678  * rvt_timeout_to_jiffies - Convert a ULP timeout input into jiffies
679  * @timeout - timeout input(0 - 31).
680  *
681  * Return a timeout value in jiffies.
682  */
683 static inline unsigned long rvt_timeout_to_jiffies(u8 timeout)
684 {
685 	if (timeout > 31)
686 		timeout = 31;
687 
688 	return usecs_to_jiffies(1U << timeout) * 4096UL / 1000UL;
689 }
690 
691 /**
692  * rvt_lookup_qpn - return the QP with the given QPN
693  * @ibp: the ibport
694  * @qpn: the QP number to look up
695  *
696  * The caller must hold the rcu_read_lock(), and keep the lock until
697  * the returned qp is no longer in use.
698  */
699 static inline struct rvt_qp *rvt_lookup_qpn(struct rvt_dev_info *rdi,
700 					    struct rvt_ibport *rvp,
701 					    u32 qpn) __must_hold(RCU)
702 {
703 	struct rvt_qp *qp = NULL;
704 
705 	if (unlikely(qpn <= 1)) {
706 		qp = rcu_dereference(rvp->qp[qpn]);
707 	} else {
708 		u32 n = hash_32(qpn, rdi->qp_dev->qp_table_bits);
709 
710 		for (qp = rcu_dereference(rdi->qp_dev->qp_table[n]); qp;
711 			qp = rcu_dereference(qp->next))
712 			if (qp->ibqp.qp_num == qpn)
713 				break;
714 	}
715 	return qp;
716 }
717 
718 /**
719  * rvt_mod_retry_timer - mod a retry timer
720  * @qp - the QP
721  * @shift - timeout shift to wait for multiple packets
722  * Modify a potentially already running retry timer
723  */
724 static inline void rvt_mod_retry_timer_ext(struct rvt_qp *qp, u8 shift)
725 {
726 	struct ib_qp *ibqp = &qp->ibqp;
727 	struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
728 
729 	lockdep_assert_held(&qp->s_lock);
730 	qp->s_flags |= RVT_S_TIMER;
731 	/* 4.096 usec. * (1 << qp->timeout) */
732 	mod_timer(&qp->s_timer, jiffies + rdi->busy_jiffies +
733 		  (qp->timeout_jiffies << shift));
734 }
735 
736 static inline void rvt_mod_retry_timer(struct rvt_qp *qp)
737 {
738 	return rvt_mod_retry_timer_ext(qp, 0);
739 }
740 
741 /**
742  * rvt_put_qp_swqe - drop refs held by swqe
743  * @qp: the send qp
744  * @wqe: the send wqe
745  *
746  * This drops any references held by the swqe
747  */
748 static inline void rvt_put_qp_swqe(struct rvt_qp *qp, struct rvt_swqe *wqe)
749 {
750 	rvt_put_swqe(wqe);
751 	if (qp->allowed_ops == IB_OPCODE_UD)
752 		rdma_destroy_ah_attr(wqe->ud_wr.attr);
753 }
754 
755 /**
756  * rvt_qp_sqwe_incr - increment ring index
757  * @qp: the qp
758  * @val: the starting value
759  *
760  * Return: the new value wrapping as appropriate
761  */
762 static inline u32
763 rvt_qp_swqe_incr(struct rvt_qp *qp, u32 val)
764 {
765 	if (++val >= qp->s_size)
766 		val = 0;
767 	return val;
768 }
769 
770 int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err);
771 
772 /**
773  * rvt_recv_cq - add a new entry to completion queue
774  *			by receive queue
775  * @qp: receive queue
776  * @wc: work completion entry to add
777  * @solicited: true if @entry is solicited
778  *
779  * This is wrapper function for rvt_enter_cq function call by
780  * receive queue. If rvt_cq_enter return false, it means cq is
781  * full and the qp is put into error state.
782  */
783 static inline void rvt_recv_cq(struct rvt_qp *qp, struct ib_wc *wc,
784 			       bool solicited)
785 {
786 	struct rvt_cq *cq = ibcq_to_rvtcq(qp->ibqp.recv_cq);
787 
788 	if (unlikely(!rvt_cq_enter(cq, wc, solicited)))
789 		rvt_error_qp(qp, IB_WC_LOC_QP_OP_ERR);
790 }
791 
792 /**
793  * rvt_send_cq - add a new entry to completion queue
794  *                        by send queue
795  * @qp: send queue
796  * @wc: work completion entry to add
797  * @solicited: true if @entry is solicited
798  *
799  * This is wrapper function for rvt_enter_cq function call by
800  * send queue. If rvt_cq_enter return false, it means cq is
801  * full and the qp is put into error state.
802  */
803 static inline void rvt_send_cq(struct rvt_qp *qp, struct ib_wc *wc,
804 			       bool solicited)
805 {
806 	struct rvt_cq *cq = ibcq_to_rvtcq(qp->ibqp.send_cq);
807 
808 	if (unlikely(!rvt_cq_enter(cq, wc, solicited)))
809 		rvt_error_qp(qp, IB_WC_LOC_QP_OP_ERR);
810 }
811 
812 /**
813  * rvt_qp_complete_swqe - insert send completion
814  * @qp - the qp
815  * @wqe - the send wqe
816  * @opcode - wc operation (driver dependent)
817  * @status - completion status
818  *
819  * Update the s_last information, and then insert a send
820  * completion into the completion
821  * queue if the qp indicates it should be done.
822  *
823  * See IBTA 10.7.3.1 for info on completion
824  * control.
825  *
826  * Return: new last
827  */
828 static inline u32
829 rvt_qp_complete_swqe(struct rvt_qp *qp,
830 		     struct rvt_swqe *wqe,
831 		     enum ib_wc_opcode opcode,
832 		     enum ib_wc_status status)
833 {
834 	bool need_completion;
835 	u64 wr_id;
836 	u32 byte_len, last;
837 	int flags = wqe->wr.send_flags;
838 
839 	rvt_qp_wqe_unreserve(qp, flags);
840 	rvt_put_qp_swqe(qp, wqe);
841 
842 	need_completion =
843 		!(flags & RVT_SEND_RESERVE_USED) &&
844 		(!(qp->s_flags & RVT_S_SIGNAL_REQ_WR) ||
845 		(flags & IB_SEND_SIGNALED) ||
846 		status != IB_WC_SUCCESS);
847 	if (need_completion) {
848 		wr_id = wqe->wr.wr_id;
849 		byte_len = wqe->length;
850 		/* above fields required before writing s_last */
851 	}
852 	last = rvt_qp_swqe_incr(qp, qp->s_last);
853 	/* see rvt_qp_is_avail() */
854 	smp_store_release(&qp->s_last, last);
855 	if (need_completion) {
856 		struct ib_wc w = {
857 			.wr_id = wr_id,
858 			.status = status,
859 			.opcode = opcode,
860 			.qp = &qp->ibqp,
861 			.byte_len = byte_len,
862 		};
863 		rvt_send_cq(qp, &w, status != IB_WC_SUCCESS);
864 	}
865 	return last;
866 }
867 
868 extern const int  ib_rvt_state_ops[];
869 
870 struct rvt_dev_info;
871 int rvt_get_rwqe(struct rvt_qp *qp, bool wr_id_only);
872 void rvt_comm_est(struct rvt_qp *qp);
873 void rvt_rc_error(struct rvt_qp *qp, enum ib_wc_status err);
874 unsigned long rvt_rnr_tbl_to_usec(u32 index);
875 enum hrtimer_restart rvt_rc_rnr_retry(struct hrtimer *t);
876 void rvt_add_rnr_timer(struct rvt_qp *qp, u32 aeth);
877 void rvt_del_timers_sync(struct rvt_qp *qp);
878 void rvt_stop_rc_timers(struct rvt_qp *qp);
879 void rvt_add_retry_timer_ext(struct rvt_qp *qp, u8 shift);
880 static inline void rvt_add_retry_timer(struct rvt_qp *qp)
881 {
882 	rvt_add_retry_timer_ext(qp, 0);
883 }
884 
885 void rvt_copy_sge(struct rvt_qp *qp, struct rvt_sge_state *ss,
886 		  void *data, u32 length,
887 		  bool release, bool copy_last);
888 void rvt_send_complete(struct rvt_qp *qp, struct rvt_swqe *wqe,
889 		       enum ib_wc_status status);
890 void rvt_ruc_loopback(struct rvt_qp *qp);
891 
892 /**
893  * struct rvt_qp_iter - the iterator for QPs
894  * @qp - the current QP
895  *
896  * This structure defines the current iterator
897  * state for sequenced access to all QPs relative
898  * to an rvt_dev_info.
899  */
900 struct rvt_qp_iter {
901 	struct rvt_qp *qp;
902 	/* private: backpointer */
903 	struct rvt_dev_info *rdi;
904 	/* private: callback routine */
905 	void (*cb)(struct rvt_qp *qp, u64 v);
906 	/* private: for arg to callback routine */
907 	u64 v;
908 	/* private: number of SMI,GSI QPs for device */
909 	int specials;
910 	/* private: current iterator index */
911 	int n;
912 };
913 
914 /**
915  * ib_cq_tail - Return tail index of cq buffer
916  * @send_cq - The cq for send
917  *
918  * This is called in qp_iter_print to get tail
919  * of cq buffer.
920  */
921 static inline u32 ib_cq_tail(struct ib_cq *send_cq)
922 {
923 	struct rvt_cq *cq = ibcq_to_rvtcq(send_cq);
924 
925 	return ibcq_to_rvtcq(send_cq)->ip ?
926 	       RDMA_READ_UAPI_ATOMIC(cq->queue->tail) :
927 	       ibcq_to_rvtcq(send_cq)->kqueue->tail;
928 }
929 
930 /**
931  * ib_cq_head - Return head index of cq buffer
932  * @send_cq - The cq for send
933  *
934  * This is called in qp_iter_print to get head
935  * of cq buffer.
936  */
937 static inline u32 ib_cq_head(struct ib_cq *send_cq)
938 {
939 	struct rvt_cq *cq = ibcq_to_rvtcq(send_cq);
940 
941 	return ibcq_to_rvtcq(send_cq)->ip ?
942 	       RDMA_READ_UAPI_ATOMIC(cq->queue->head) :
943 	       ibcq_to_rvtcq(send_cq)->kqueue->head;
944 }
945 
946 /**
947  * rvt_free_rq - free memory allocated for rvt_rq struct
948  * @rvt_rq: request queue data structure
949  *
950  * This function should only be called if the rvt_mmap_info()
951  * has not succeeded.
952  */
953 static inline void rvt_free_rq(struct rvt_rq *rq)
954 {
955 	kvfree(rq->kwq);
956 	rq->kwq = NULL;
957 	vfree(rq->wq);
958 	rq->wq = NULL;
959 }
960 
961 /**
962  * rvt_to_iport - Get the ibport pointer
963  * @qp: the qp pointer
964  *
965  * This function returns the ibport pointer from the qp pointer.
966  */
967 static inline struct rvt_ibport *rvt_to_iport(struct rvt_qp *qp)
968 {
969 	struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
970 
971 	return rdi->ports[qp->port_num - 1];
972 }
973 
974 /**
975  * rvt_rc_credit_avail - Check if there are enough RC credits for the request
976  * @qp: the qp
977  * @wqe: the request
978  *
979  * This function returns false when there are not enough credits for the given
980  * request and true otherwise.
981  */
982 static inline bool rvt_rc_credit_avail(struct rvt_qp *qp, struct rvt_swqe *wqe)
983 {
984 	lockdep_assert_held(&qp->s_lock);
985 	if (!(qp->s_flags & RVT_S_UNLIMITED_CREDIT) &&
986 	    rvt_cmp_msn(wqe->ssn, qp->s_lsn + 1) > 0) {
987 		struct rvt_ibport *rvp = rvt_to_iport(qp);
988 
989 		qp->s_flags |= RVT_S_WAIT_SSN_CREDIT;
990 		rvp->n_rc_crwaits++;
991 		return false;
992 	}
993 	return true;
994 }
995 
996 struct rvt_qp_iter *rvt_qp_iter_init(struct rvt_dev_info *rdi,
997 				     u64 v,
998 				     void (*cb)(struct rvt_qp *qp, u64 v));
999 int rvt_qp_iter_next(struct rvt_qp_iter *iter);
1000 void rvt_qp_iter(struct rvt_dev_info *rdi,
1001 		 u64 v,
1002 		 void (*cb)(struct rvt_qp *qp, u64 v));
1003 void rvt_qp_mr_clean(struct rvt_qp *qp, u32 lkey);
1004 #endif          /* DEF_RDMAVT_INCQP_H */
1005