xref: /freebsd/sys/ofed/include/rdma/rdmavt_qp.h (revision 389e4940069316fe667ffa263fa7d6390d0a960f)
1 #ifndef DEF_RDMAVT_INCQP_H
2 #define DEF_RDMAVT_INCQP_H
3 
4 /*-
5  * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0
6  *
7  * Copyright(c) 2016 Intel Corporation.
8  *
9  * This file is provided under a dual BSD/GPLv2 license.  When using or
10  * redistributing this file, you may do so under either license.
11  *
12  * GPL LICENSE SUMMARY
13  *
14  * This program is free software; you can redistribute it and/or modify
15  * it under the terms of version 2 of the GNU General Public License as
16  * published by the Free Software Foundation.
17  *
18  * This program is distributed in the hope that it will be useful, but
19  * WITHOUT ANY WARRANTY; without even the implied warranty of
20  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
21  * General Public License for more details.
22  *
23  * BSD LICENSE
24  *
25  * Redistribution and use in source and binary forms, with or without
26  * modification, are permitted provided that the following conditions
27  * are met:
28  *
29  *  - Redistributions of source code must retain the above copyright
30  *    notice, this list of conditions and the following disclaimer.
31  *  - Redistributions in binary form must reproduce the above copyright
32  *    notice, this list of conditions and the following disclaimer in
33  *    the documentation and/or other materials provided with the
34  *    distribution.
35  *  - Neither the name of Intel Corporation nor the names of its
36  *    contributors may be used to endorse or promote products derived
37  *    from this software without specific prior written permission.
38  *
39  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
40  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
41  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
42  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
43  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
44  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
45  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
46  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
47  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
48  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
49  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
50  *
51  * $FreeBSD$
52  */
53 
54 #include <rdma/rdma_vt.h>
55 #include <rdma/ib_pack.h>
56 #include <rdma/ib_verbs.h>
57 /*
58  * Atomic bit definitions for r_aflags.
59  */
60 #define RVT_R_WRID_VALID        0
61 #define RVT_R_REWIND_SGE        1
62 
63 /*
64  * Bit definitions for r_flags.
65  */
66 #define RVT_R_REUSE_SGE 0x01
67 #define RVT_R_RDMAR_SEQ 0x02
68 #define RVT_R_RSP_NAK   0x04
69 #define RVT_R_RSP_SEND  0x08
70 #define RVT_R_COMM_EST  0x10
71 
72 /*
73  * Bit definitions for s_flags.
74  *
75  * RVT_S_SIGNAL_REQ_WR - set if QP send WRs contain completion signaled
76  * RVT_S_BUSY - send tasklet is processing the QP
77  * RVT_S_TIMER - the RC retry timer is active
78  * RVT_S_ACK_PENDING - an ACK is waiting to be sent after RDMA read/atomics
79  * RVT_S_WAIT_FENCE - waiting for all prior RDMA read or atomic SWQEs
80  *                         before processing the next SWQE
81  * RVT_S_WAIT_RDMAR - waiting for a RDMA read or atomic SWQE to complete
82  *                         before processing the next SWQE
83  * RVT_S_WAIT_RNR - waiting for RNR timeout
84  * RVT_S_WAIT_SSN_CREDIT - waiting for RC credits to process next SWQE
85  * RVT_S_WAIT_DMA - waiting for send DMA queue to drain before generating
86  *                  next send completion entry not via send DMA
87  * RVT_S_WAIT_PIO - waiting for a send buffer to be available
88  * RVT_S_WAIT_PIO_DRAIN - waiting for a qp to drain pio packets
89  * RVT_S_WAIT_TX - waiting for a struct verbs_txreq to be available
90  * RVT_S_WAIT_DMA_DESC - waiting for DMA descriptors to be available
91  * RVT_S_WAIT_KMEM - waiting for kernel memory to be available
92  * RVT_S_WAIT_PSN - waiting for a packet to exit the send DMA queue
93  * RVT_S_WAIT_ACK - waiting for an ACK packet before sending more requests
94  * RVT_S_SEND_ONE - send one packet, request ACK, then wait for ACK
95  * RVT_S_ECN - a BECN was queued to the send engine
96  */
97 #define RVT_S_SIGNAL_REQ_WR	0x0001
98 #define RVT_S_BUSY		0x0002
99 #define RVT_S_TIMER		0x0004
100 #define RVT_S_RESP_PENDING	0x0008
101 #define RVT_S_ACK_PENDING	0x0010
102 #define RVT_S_WAIT_FENCE	0x0020
103 #define RVT_S_WAIT_RDMAR	0x0040
104 #define RVT_S_WAIT_RNR		0x0080
105 #define RVT_S_WAIT_SSN_CREDIT	0x0100
106 #define RVT_S_WAIT_DMA		0x0200
107 #define RVT_S_WAIT_PIO		0x0400
108 #define RVT_S_WAIT_PIO_DRAIN    0x0800
109 #define RVT_S_WAIT_TX		0x1000
110 #define RVT_S_WAIT_DMA_DESC	0x2000
111 #define RVT_S_WAIT_KMEM		0x4000
112 #define RVT_S_WAIT_PSN		0x8000
113 #define RVT_S_WAIT_ACK		0x10000
114 #define RVT_S_SEND_ONE		0x20000
115 #define RVT_S_UNLIMITED_CREDIT	0x40000
116 #define RVT_S_AHG_VALID		0x80000
117 #define RVT_S_AHG_CLEAR		0x100000
118 #define RVT_S_ECN		0x200000
119 
120 /*
121  * Wait flags that would prevent any packet type from being sent.
122  */
123 #define RVT_S_ANY_WAIT_IO \
124 	(RVT_S_WAIT_PIO | RVT_S_WAIT_PIO_DRAIN | RVT_S_WAIT_TX | \
125 	 RVT_S_WAIT_DMA_DESC | RVT_S_WAIT_KMEM)
126 
127 /*
128  * Wait flags that would prevent send work requests from making progress.
129  */
130 #define RVT_S_ANY_WAIT_SEND (RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR | \
131 	RVT_S_WAIT_RNR | RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_DMA | \
132 	RVT_S_WAIT_PSN | RVT_S_WAIT_ACK)
133 
134 #define RVT_S_ANY_WAIT (RVT_S_ANY_WAIT_IO | RVT_S_ANY_WAIT_SEND)
135 
136 /* Number of bits to pay attention to in the opcode for checking qp type */
137 #define RVT_OPCODE_QP_MASK 0xE0
138 
139 /* Flags for checking QP state (see ib_rvt_state_ops[]) */
140 #define RVT_POST_SEND_OK                0x01
141 #define RVT_POST_RECV_OK                0x02
142 #define RVT_PROCESS_RECV_OK             0x04
143 #define RVT_PROCESS_SEND_OK             0x08
144 #define RVT_PROCESS_NEXT_SEND_OK        0x10
145 #define RVT_FLUSH_SEND			0x20
146 #define RVT_FLUSH_RECV			0x40
147 #define RVT_PROCESS_OR_FLUSH_SEND \
148 	(RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND)
149 
150 /*
151  * Internal send flags
152  */
153 #define RVT_SEND_RESERVE_USED           IB_SEND_RESERVED_START
154 #define RVT_SEND_COMPLETION_ONLY	(IB_SEND_RESERVED_START << 1)
155 
156 /*
157  * Send work request queue entry.
158  * The size of the sg_list is determined when the QP is created and stored
159  * in qp->s_max_sge.
160  */
161 struct rvt_swqe {
162 	union {
163 		struct ib_send_wr wr;   /* don't use wr.sg_list */
164 		struct ib_ud_wr ud_wr;
165 		struct ib_reg_wr reg_wr;
166 		struct ib_rdma_wr rdma_wr;
167 		struct ib_atomic_wr atomic_wr;
168 	};
169 	u32 psn;                /* first packet sequence number */
170 	u32 lpsn;               /* last packet sequence number */
171 	u32 ssn;                /* send sequence number */
172 	u32 length;             /* total length of data in sg_list */
173 	struct rvt_sge sg_list[0];
174 };
175 
176 /*
177  * Receive work request queue entry.
178  * The size of the sg_list is determined when the QP (or SRQ) is created
179  * and stored in qp->r_rq.max_sge (or srq->rq.max_sge).
180  */
181 struct rvt_rwqe {
182 	u64 wr_id;
183 	u8 num_sge;
184 	struct ib_sge sg_list[0];
185 };
186 
187 /*
188  * This structure is used to contain the head pointer, tail pointer,
189  * and receive work queue entries as a single memory allocation so
190  * it can be mmap'ed into user space.
191  * Note that the wq array elements are variable size so you can't
192  * just index into the array to get the N'th element;
193  * use get_rwqe_ptr() instead.
194  */
195 struct rvt_rwq {
196 	u32 head;               /* new work requests posted to the head */
197 	u32 tail;               /* receives pull requests from here. */
198 	struct rvt_rwqe wq[0];
199 };
200 
201 struct rvt_rq {
202 	struct rvt_rwq *wq;
203 	u32 size;               /* size of RWQE array */
204 	u8 max_sge;
205 	/* protect changes in this struct */
206 	spinlock_t lock ____cacheline_aligned_in_smp;
207 };
208 
209 /*
210  * This structure is used by rvt_mmap() to validate an offset
211  * when an mmap() request is made.  The vm_area_struct then uses
212  * this as its vm_private_data.
213  */
214 struct rvt_mmap_info {
215 	struct list_head pending_mmaps;
216 	struct ib_ucontext *context;
217 	void *obj;
218 	__u64 offset;
219 	struct kref ref;
220 	unsigned size;
221 };
222 
223 /*
224  * This structure holds the information that the send tasklet needs
225  * to send a RDMA read response or atomic operation.
226  */
227 struct rvt_ack_entry {
228 	struct rvt_sge rdma_sge;
229 	u64 atomic_data;
230 	u32 psn;
231 	u32 lpsn;
232 	u8 opcode;
233 	u8 sent;
234 };
235 
236 #define	RC_QP_SCALING_INTERVAL	5
237 
238 #define RVT_OPERATION_PRIV        0x00000001
239 #define RVT_OPERATION_ATOMIC      0x00000002
240 #define RVT_OPERATION_ATOMIC_SGE  0x00000004
241 #define RVT_OPERATION_LOCAL       0x00000008
242 #define RVT_OPERATION_USE_RESERVE 0x00000010
243 
244 #define RVT_OPERATION_MAX (IB_WR_RESERVED10 + 1)
245 
246 /**
247  * rvt_operation_params - op table entry
248  * @length - the length to copy into the swqe entry
249  * @qpt_support - a bit mask indicating QP type support
250  * @flags - RVT_OPERATION flags (see above)
251  *
252  * This supports table driven post send so that
253  * the driver can have differing an potentially
254  * different sets of operations.
255  *
256  **/
257 
258 struct rvt_operation_params {
259 	size_t length;
260 	u32 qpt_support;
261 	u32 flags;
262 };
263 
264 /*
265  * Common variables are protected by both r_rq.lock and s_lock in that order
266  * which only happens in modify_qp() or changing the QP 'state'.
267  */
268 struct rvt_qp {
269 	struct ib_qp ibqp;
270 	void *priv; /* Driver private data */
271 	/* read mostly fields above and below */
272 	struct ib_ah_attr remote_ah_attr;
273 	struct ib_ah_attr alt_ah_attr;
274 	struct rvt_qp __rcu *next;           /* link list for QPN hash table */
275 	struct rvt_swqe *s_wq;  /* send work queue */
276 	struct rvt_mmap_info *ip;
277 
278 	unsigned long timeout_jiffies;  /* computed from timeout */
279 
280 	enum ib_mtu path_mtu;
281 	int srate_mbps;		/* s_srate (below) converted to Mbit/s */
282 	pid_t pid;		/* pid for user mode QPs */
283 	u32 remote_qpn;
284 	u32 qkey;               /* QKEY for this QP (for UD or RD) */
285 	u32 s_size;             /* send work queue size */
286 	u32 s_ahgpsn;           /* set to the psn in the copy of the header */
287 
288 	u16 pmtu;		/* decoded from path_mtu */
289 	u8 log_pmtu;		/* shift for pmtu */
290 	u8 state;               /* QP state */
291 	u8 allowed_ops;		/* high order bits of allowed opcodes */
292 	u8 qp_access_flags;
293 	u8 alt_timeout;         /* Alternate path timeout for this QP */
294 	u8 timeout;             /* Timeout for this QP */
295 	u8 s_srate;
296 	u8 s_mig_state;
297 	u8 port_num;
298 	u8 s_pkey_index;        /* PKEY index to use */
299 	u8 s_alt_pkey_index;    /* Alternate path PKEY index to use */
300 	u8 r_max_rd_atomic;     /* max number of RDMA read/atomic to receive */
301 	u8 s_max_rd_atomic;     /* max number of RDMA read/atomic to send */
302 	u8 s_retry_cnt;         /* number of times to retry */
303 	u8 s_rnr_retry_cnt;
304 	u8 r_min_rnr_timer;     /* retry timeout value for RNR NAKs */
305 	u8 s_max_sge;           /* size of s_wq->sg_list */
306 	u8 s_draining;
307 
308 	/* start of read/write fields */
309 	atomic_t refcount ____cacheline_aligned_in_smp;
310 	wait_queue_head_t wait;
311 
312 	struct rvt_ack_entry *s_ack_queue;
313 	struct rvt_sge_state s_rdma_read_sge;
314 
315 	spinlock_t r_lock ____cacheline_aligned_in_smp;      /* used for APM */
316 	u32 r_psn;              /* expected rcv packet sequence number */
317 	unsigned long r_aflags;
318 	u64 r_wr_id;            /* ID for current receive WQE */
319 	u32 r_ack_psn;          /* PSN for next ACK or atomic ACK */
320 	u32 r_len;              /* total length of r_sge */
321 	u32 r_rcv_len;          /* receive data len processed */
322 	u32 r_msn;              /* message sequence number */
323 
324 	u8 r_state;             /* opcode of last packet received */
325 	u8 r_flags;
326 	u8 r_head_ack_queue;    /* index into s_ack_queue[] */
327 
328 	struct list_head rspwait;       /* link for waiting to respond */
329 
330 	struct rvt_sge_state r_sge;     /* current receive data */
331 	struct rvt_rq r_rq;             /* receive work queue */
332 
333 	/* post send line */
334 	spinlock_t s_hlock ____cacheline_aligned_in_smp;
335 	u32 s_head;             /* new entries added here */
336 	u32 s_next_psn;         /* PSN for next request */
337 	u32 s_avail;            /* number of entries avail */
338 	u32 s_ssn;              /* SSN of tail entry */
339 	atomic_t s_reserved_used; /* reserved entries in use */
340 
341 	spinlock_t s_lock ____cacheline_aligned_in_smp;
342 	u32 s_flags;
343 	struct rvt_sge_state *s_cur_sge;
344 	struct rvt_swqe *s_wqe;
345 	struct rvt_sge_state s_sge;     /* current send request data */
346 	struct rvt_mregion *s_rdma_mr;
347 	u32 s_cur_size;         /* size of send packet in bytes */
348 	u32 s_len;              /* total length of s_sge */
349 	u32 s_rdma_read_len;    /* total length of s_rdma_read_sge */
350 	u32 s_last_psn;         /* last response PSN processed */
351 	u32 s_sending_psn;      /* lowest PSN that is being sent */
352 	u32 s_sending_hpsn;     /* highest PSN that is being sent */
353 	u32 s_psn;              /* current packet sequence number */
354 	u32 s_ack_rdma_psn;     /* PSN for sending RDMA read responses */
355 	u32 s_ack_psn;          /* PSN for acking sends and RDMA writes */
356 	u32 s_tail;             /* next entry to process */
357 	u32 s_cur;              /* current work queue entry */
358 	u32 s_acked;            /* last un-ACK'ed entry */
359 	u32 s_last;             /* last completed entry */
360 	u32 s_lsn;              /* limit sequence number (credit) */
361 	u16 s_hdrwords;         /* size of s_hdr in 32 bit words */
362 	u16 s_rdma_ack_cnt;
363 	s8 s_ahgidx;
364 	u8 s_state;             /* opcode of last packet sent */
365 	u8 s_ack_state;         /* opcode of packet to ACK */
366 	u8 s_nak_state;         /* non-zero if NAK is pending */
367 	u8 r_nak_state;         /* non-zero if NAK is pending */
368 	u8 s_retry;             /* requester retry counter */
369 	u8 s_rnr_retry;         /* requester RNR retry counter */
370 	u8 s_num_rd_atomic;     /* number of RDMA read/atomic pending */
371 	u8 s_tail_ack_queue;    /* index into s_ack_queue[] */
372 
373 	struct rvt_sge_state s_ack_rdma_sge;
374 	struct timer_list s_timer;
375 
376 	atomic_t local_ops_pending; /* number of fast_reg/local_inv reqs */
377 
378 	/*
379 	 * This sge list MUST be last. Do not add anything below here.
380 	 */
381 	struct rvt_sge r_sg_list[0] /* verified SGEs */
382 		____cacheline_aligned_in_smp;
383 };
384 
385 struct rvt_srq {
386 	struct ib_srq ibsrq;
387 	struct rvt_rq rq;
388 	struct rvt_mmap_info *ip;
389 	/* send signal when number of RWQEs < limit */
390 	u32 limit;
391 };
392 
393 #define RVT_QPN_MAX                 BIT(24)
394 #define RVT_QPNMAP_ENTRIES          (RVT_QPN_MAX / PAGE_SIZE / BITS_PER_BYTE)
395 #define RVT_BITS_PER_PAGE           (PAGE_SIZE * BITS_PER_BYTE)
396 #define RVT_BITS_PER_PAGE_MASK      (RVT_BITS_PER_PAGE - 1)
397 #define RVT_QPN_MASK		    0xFFFFFF
398 
399 /*
400  * QPN-map pages start out as NULL, they get allocated upon
401  * first use and are never deallocated. This way,
402  * large bitmaps are not allocated unless large numbers of QPs are used.
403  */
404 struct rvt_qpn_map {
405 	void *page;
406 };
407 
408 struct rvt_qpn_table {
409 	spinlock_t lock; /* protect changes to the qp table */
410 	unsigned flags;         /* flags for QP0/1 allocated for each port */
411 	u32 last;               /* last QP number allocated */
412 	u32 nmaps;              /* size of the map table */
413 	u16 limit;
414 	u8  incr;
415 	/* bit map of free QP numbers other than 0/1 */
416 	struct rvt_qpn_map map[RVT_QPNMAP_ENTRIES];
417 };
418 
419 struct rvt_qp_ibdev {
420 	u32 qp_table_size;
421 	u32 qp_table_bits;
422 	struct rvt_qp __rcu **qp_table;
423 	spinlock_t qpt_lock; /* qptable lock */
424 	struct rvt_qpn_table qpn_table;
425 };
426 
427 /*
428  * There is one struct rvt_mcast for each multicast GID.
429  * All attached QPs are then stored as a list of
430  * struct rvt_mcast_qp.
431  */
432 struct rvt_mcast_qp {
433 	struct list_head list;
434 	struct rvt_qp *qp;
435 };
436 
437 struct rvt_mcast {
438 	struct rb_node rb_node;
439 	union ib_gid mgid;
440 	struct list_head qp_list;
441 	wait_queue_head_t wait;
442 	atomic_t refcount;
443 	int n_attached;
444 };
445 
446 /*
447  * Since struct rvt_swqe is not a fixed size, we can't simply index into
448  * struct rvt_qp.s_wq.  This function does the array index computation.
449  */
450 static inline struct rvt_swqe *rvt_get_swqe_ptr(struct rvt_qp *qp,
451 						unsigned n)
452 {
453 	return (struct rvt_swqe *)((char *)qp->s_wq +
454 				     (sizeof(struct rvt_swqe) +
455 				      qp->s_max_sge *
456 				      sizeof(struct rvt_sge)) * n);
457 }
458 
459 /*
460  * Since struct rvt_rwqe is not a fixed size, we can't simply index into
461  * struct rvt_rwq.wq.  This function does the array index computation.
462  */
463 static inline struct rvt_rwqe *rvt_get_rwqe_ptr(struct rvt_rq *rq, unsigned n)
464 {
465 	return (struct rvt_rwqe *)
466 		((char *)rq->wq->wq +
467 		 (sizeof(struct rvt_rwqe) +
468 		  rq->max_sge * sizeof(struct ib_sge)) * n);
469 }
470 
471 /**
472  * rvt_get_qp - get a QP reference
473  * @qp - the QP to hold
474  */
475 static inline void rvt_get_qp(struct rvt_qp *qp)
476 {
477 	atomic_inc(&qp->refcount);
478 }
479 
480 /**
481  * rvt_put_qp - release a QP reference
482  * @qp - the QP to release
483  */
484 static inline void rvt_put_qp(struct rvt_qp *qp)
485 {
486 	if (qp && atomic_dec_and_test(&qp->refcount))
487 		wake_up(&qp->wait);
488 }
489 
490 /**
491  * rvt_qp_wqe_reserve - reserve operation
492  * @qp - the rvt qp
493  * @wqe - the send wqe
494  *
495  * This routine used in post send to record
496  * a wqe relative reserved operation use.
497  */
498 static inline void rvt_qp_wqe_reserve(
499 	struct rvt_qp *qp,
500 	struct rvt_swqe *wqe)
501 {
502 	wqe->wr.send_flags |= RVT_SEND_RESERVE_USED;
503 	atomic_inc(&qp->s_reserved_used);
504 }
505 
506 /**
507  * rvt_qp_wqe_unreserve - clean reserved operation
508  * @qp - the rvt qp
509  * @wqe - the send wqe
510  *
511  * This decrements the reserve use count.
512  *
513  * This call MUST precede the change to
514  * s_last to insure that post send sees a stable
515  * s_avail.
516  *
517  * An smp_mp__after_atomic() is used to insure
518  * the compiler does not juggle the order of the s_last
519  * ring index and the decrementing of s_reserved_used.
520  */
521 static inline void rvt_qp_wqe_unreserve(
522 	struct rvt_qp *qp,
523 	struct rvt_swqe *wqe)
524 {
525 	if (unlikely(wqe->wr.send_flags & RVT_SEND_RESERVE_USED)) {
526 		wqe->wr.send_flags &= ~RVT_SEND_RESERVE_USED;
527 		atomic_dec(&qp->s_reserved_used);
528 		/* insure no compiler re-order up to s_last change */
529 		smp_mb__after_atomic();
530 	}
531 }
532 
533 extern const int  ib_rvt_state_ops[];
534 
535 struct rvt_dev_info;
536 int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err);
537 
538 #endif          /* DEF_RDMAVT_INCQP_H */
539