xref: /freebsd/sys/ofed/include/rdma/rdmavt_qp.h (revision 95ee2897)
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 
52 #include <rdma/rdma_vt.h>
53 #include <rdma/ib_pack.h>
54 #include <rdma/ib_verbs.h>
55 /*
56  * Atomic bit definitions for r_aflags.
57  */
58 #define RVT_R_WRID_VALID        0
59 #define RVT_R_REWIND_SGE        1
60 
61 /*
62  * Bit definitions for r_flags.
63  */
64 #define RVT_R_REUSE_SGE 0x01
65 #define RVT_R_RDMAR_SEQ 0x02
66 #define RVT_R_RSP_NAK   0x04
67 #define RVT_R_RSP_SEND  0x08
68 #define RVT_R_COMM_EST  0x10
69 
70 /*
71  * Bit definitions for s_flags.
72  *
73  * RVT_S_SIGNAL_REQ_WR - set if QP send WRs contain completion signaled
74  * RVT_S_BUSY - send tasklet is processing the QP
75  * RVT_S_TIMER - the RC retry timer is active
76  * RVT_S_ACK_PENDING - an ACK is waiting to be sent after RDMA read/atomics
77  * RVT_S_WAIT_FENCE - waiting for all prior RDMA read or atomic SWQEs
78  *                         before processing the next SWQE
79  * RVT_S_WAIT_RDMAR - waiting for a RDMA read or atomic SWQE to complete
80  *                         before processing the next SWQE
81  * RVT_S_WAIT_RNR - waiting for RNR timeout
82  * RVT_S_WAIT_SSN_CREDIT - waiting for RC credits to process next SWQE
83  * RVT_S_WAIT_DMA - waiting for send DMA queue to drain before generating
84  *                  next send completion entry not via send DMA
85  * RVT_S_WAIT_PIO - waiting for a send buffer to be available
86  * RVT_S_WAIT_PIO_DRAIN - waiting for a qp to drain pio packets
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  */
95 #define RVT_S_SIGNAL_REQ_WR	0x0001
96 #define RVT_S_BUSY		0x0002
97 #define RVT_S_TIMER		0x0004
98 #define RVT_S_RESP_PENDING	0x0008
99 #define RVT_S_ACK_PENDING	0x0010
100 #define RVT_S_WAIT_FENCE	0x0020
101 #define RVT_S_WAIT_RDMAR	0x0040
102 #define RVT_S_WAIT_RNR		0x0080
103 #define RVT_S_WAIT_SSN_CREDIT	0x0100
104 #define RVT_S_WAIT_DMA		0x0200
105 #define RVT_S_WAIT_PIO		0x0400
106 #define RVT_S_WAIT_PIO_DRAIN    0x0800
107 #define RVT_S_WAIT_TX		0x1000
108 #define RVT_S_WAIT_DMA_DESC	0x2000
109 #define RVT_S_WAIT_KMEM		0x4000
110 #define RVT_S_WAIT_PSN		0x8000
111 #define RVT_S_WAIT_ACK		0x10000
112 #define RVT_S_SEND_ONE		0x20000
113 #define RVT_S_UNLIMITED_CREDIT	0x40000
114 #define RVT_S_AHG_VALID		0x80000
115 #define RVT_S_AHG_CLEAR		0x100000
116 #define RVT_S_ECN		0x200000
117 
118 /*
119  * Wait flags that would prevent any packet type from being sent.
120  */
121 #define RVT_S_ANY_WAIT_IO \
122 	(RVT_S_WAIT_PIO | RVT_S_WAIT_PIO_DRAIN | RVT_S_WAIT_TX | \
123 	 RVT_S_WAIT_DMA_DESC | RVT_S_WAIT_KMEM)
124 
125 /*
126  * Wait flags that would prevent send work requests from making progress.
127  */
128 #define RVT_S_ANY_WAIT_SEND (RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR | \
129 	RVT_S_WAIT_RNR | RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_DMA | \
130 	RVT_S_WAIT_PSN | RVT_S_WAIT_ACK)
131 
132 #define RVT_S_ANY_WAIT (RVT_S_ANY_WAIT_IO | RVT_S_ANY_WAIT_SEND)
133 
134 /* Number of bits to pay attention to in the opcode for checking qp type */
135 #define RVT_OPCODE_QP_MASK 0xE0
136 
137 /* Flags for checking QP state (see ib_rvt_state_ops[]) */
138 #define RVT_POST_SEND_OK                0x01
139 #define RVT_POST_RECV_OK                0x02
140 #define RVT_PROCESS_RECV_OK             0x04
141 #define RVT_PROCESS_SEND_OK             0x08
142 #define RVT_PROCESS_NEXT_SEND_OK        0x10
143 #define RVT_FLUSH_SEND			0x20
144 #define RVT_FLUSH_RECV			0x40
145 #define RVT_PROCESS_OR_FLUSH_SEND \
146 	(RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND)
147 
148 /*
149  * Internal send flags
150  */
151 #define RVT_SEND_RESERVE_USED           IB_SEND_RESERVED_START
152 #define RVT_SEND_COMPLETION_ONLY	(IB_SEND_RESERVED_START << 1)
153 
154 /*
155  * Send work request queue entry.
156  * The size of the sg_list is determined when the QP is created and stored
157  * in qp->s_max_sge.
158  */
159 struct rvt_swqe {
160 	union {
161 		struct ib_send_wr wr;   /* don't use wr.sg_list */
162 		struct ib_ud_wr ud_wr;
163 		struct ib_reg_wr reg_wr;
164 		struct ib_rdma_wr rdma_wr;
165 		struct ib_atomic_wr atomic_wr;
166 	};
167 	u32 psn;                /* first packet sequence number */
168 	u32 lpsn;               /* last packet sequence number */
169 	u32 ssn;                /* send sequence number */
170 	u32 length;             /* total length of data in sg_list */
171 	struct rvt_sge sg_list[0];
172 };
173 
174 /*
175  * Receive work request queue entry.
176  * The size of the sg_list is determined when the QP (or SRQ) is created
177  * and stored in qp->r_rq.max_sge (or srq->rq.max_sge).
178  */
179 struct rvt_rwqe {
180 	u64 wr_id;
181 	u8 num_sge;
182 	struct ib_sge sg_list[0];
183 };
184 
185 /*
186  * This structure is used to contain the head pointer, tail pointer,
187  * and receive work queue entries as a single memory allocation so
188  * it can be mmap'ed into user space.
189  * Note that the wq array elements are variable size so you can't
190  * just index into the array to get the N'th element;
191  * use get_rwqe_ptr() instead.
192  */
193 struct rvt_rwq {
194 	u32 head;               /* new work requests posted to the head */
195 	u32 tail;               /* receives pull requests from here. */
196 	struct rvt_rwqe wq[0];
197 };
198 
199 struct rvt_rq {
200 	struct rvt_rwq *wq;
201 	u32 size;               /* size of RWQE array */
202 	u8 max_sge;
203 	/* protect changes in this struct */
204 	spinlock_t lock ____cacheline_aligned_in_smp;
205 };
206 
207 /*
208  * This structure is used by rvt_mmap() to validate an offset
209  * when an mmap() request is made.  The vm_area_struct then uses
210  * this as its vm_private_data.
211  */
212 struct rvt_mmap_info {
213 	struct list_head pending_mmaps;
214 	struct ib_ucontext *context;
215 	void *obj;
216 	__u64 offset;
217 	struct kref ref;
218 	unsigned size;
219 };
220 
221 /*
222  * This structure holds the information that the send tasklet needs
223  * to send a RDMA read response or atomic operation.
224  */
225 struct rvt_ack_entry {
226 	struct rvt_sge rdma_sge;
227 	u64 atomic_data;
228 	u32 psn;
229 	u32 lpsn;
230 	u8 opcode;
231 	u8 sent;
232 };
233 
234 #define	RC_QP_SCALING_INTERVAL	5
235 
236 #define RVT_OPERATION_PRIV        0x00000001
237 #define RVT_OPERATION_ATOMIC      0x00000002
238 #define RVT_OPERATION_ATOMIC_SGE  0x00000004
239 #define RVT_OPERATION_LOCAL       0x00000008
240 #define RVT_OPERATION_USE_RESERVE 0x00000010
241 
242 #define RVT_OPERATION_MAX (IB_WR_RESERVED10 + 1)
243 
244 /**
245  * rvt_operation_params - op table entry
246  * @length - the length to copy into the swqe entry
247  * @qpt_support - a bit mask indicating QP type support
248  * @flags - RVT_OPERATION flags (see above)
249  *
250  * This supports table driven post send so that
251  * the driver can have differing an potentially
252  * different sets of operations.
253  *
254  **/
255 
256 struct rvt_operation_params {
257 	size_t length;
258 	u32 qpt_support;
259 	u32 flags;
260 };
261 
262 /*
263  * Common variables are protected by both r_rq.lock and s_lock in that order
264  * which only happens in modify_qp() or changing the QP 'state'.
265  */
266 struct rvt_qp {
267 	struct ib_qp ibqp;
268 	void *priv; /* Driver private data */
269 	/* read mostly fields above and below */
270 	struct ib_ah_attr remote_ah_attr;
271 	struct ib_ah_attr alt_ah_attr;
272 	struct rvt_qp __rcu *next;           /* link list for QPN hash table */
273 	struct rvt_swqe *s_wq;  /* send work queue */
274 	struct rvt_mmap_info *ip;
275 
276 	unsigned long timeout_jiffies;  /* computed from timeout */
277 
278 	enum ib_mtu path_mtu;
279 	int srate_mbps;		/* s_srate (below) converted to Mbit/s */
280 	pid_t pid;		/* pid for user mode QPs */
281 	u32 remote_qpn;
282 	u32 qkey;               /* QKEY for this QP (for UD or RD) */
283 	u32 s_size;             /* send work queue size */
284 	u32 s_ahgpsn;           /* set to the psn in the copy of the header */
285 
286 	u16 pmtu;		/* decoded from path_mtu */
287 	u8 log_pmtu;		/* shift for pmtu */
288 	u8 state;               /* QP state */
289 	u8 allowed_ops;		/* high order bits of allowed opcodes */
290 	u8 qp_access_flags;
291 	u8 alt_timeout;         /* Alternate path timeout for this QP */
292 	u8 timeout;             /* Timeout for this QP */
293 	u8 s_srate;
294 	u8 s_mig_state;
295 	u8 port_num;
296 	u8 s_pkey_index;        /* PKEY index to use */
297 	u8 s_alt_pkey_index;    /* Alternate path PKEY index to use */
298 	u8 r_max_rd_atomic;     /* max number of RDMA read/atomic to receive */
299 	u8 s_max_rd_atomic;     /* max number of RDMA read/atomic to send */
300 	u8 s_retry_cnt;         /* number of times to retry */
301 	u8 s_rnr_retry_cnt;
302 	u8 r_min_rnr_timer;     /* retry timeout value for RNR NAKs */
303 	u8 s_max_sge;           /* size of s_wq->sg_list */
304 	u8 s_draining;
305 
306 	/* start of read/write fields */
307 	atomic_t refcount ____cacheline_aligned_in_smp;
308 	wait_queue_head_t wait;
309 
310 	struct rvt_ack_entry *s_ack_queue;
311 	struct rvt_sge_state s_rdma_read_sge;
312 
313 	spinlock_t r_lock ____cacheline_aligned_in_smp;      /* used for APM */
314 	u32 r_psn;              /* expected rcv packet sequence number */
315 	unsigned long r_aflags;
316 	u64 r_wr_id;            /* ID for current receive WQE */
317 	u32 r_ack_psn;          /* PSN for next ACK or atomic ACK */
318 	u32 r_len;              /* total length of r_sge */
319 	u32 r_rcv_len;          /* receive data len processed */
320 	u32 r_msn;              /* message sequence number */
321 
322 	u8 r_state;             /* opcode of last packet received */
323 	u8 r_flags;
324 	u8 r_head_ack_queue;    /* index into s_ack_queue[] */
325 
326 	struct list_head rspwait;       /* link for waiting to respond */
327 
328 	struct rvt_sge_state r_sge;     /* current receive data */
329 	struct rvt_rq r_rq;             /* receive work queue */
330 
331 	/* post send line */
332 	spinlock_t s_hlock ____cacheline_aligned_in_smp;
333 	u32 s_head;             /* new entries added here */
334 	u32 s_next_psn;         /* PSN for next request */
335 	u32 s_avail;            /* number of entries avail */
336 	u32 s_ssn;              /* SSN of tail entry */
337 	atomic_t s_reserved_used; /* reserved entries in use */
338 
339 	spinlock_t s_lock ____cacheline_aligned_in_smp;
340 	u32 s_flags;
341 	struct rvt_sge_state *s_cur_sge;
342 	struct rvt_swqe *s_wqe;
343 	struct rvt_sge_state s_sge;     /* current send request data */
344 	struct rvt_mregion *s_rdma_mr;
345 	u32 s_cur_size;         /* size of send packet in bytes */
346 	u32 s_len;              /* total length of s_sge */
347 	u32 s_rdma_read_len;    /* total length of s_rdma_read_sge */
348 	u32 s_last_psn;         /* last response PSN processed */
349 	u32 s_sending_psn;      /* lowest PSN that is being sent */
350 	u32 s_sending_hpsn;     /* highest PSN that is being sent */
351 	u32 s_psn;              /* current packet sequence number */
352 	u32 s_ack_rdma_psn;     /* PSN for sending RDMA read responses */
353 	u32 s_ack_psn;          /* PSN for acking sends and RDMA writes */
354 	u32 s_tail;             /* next entry to process */
355 	u32 s_cur;              /* current work queue entry */
356 	u32 s_acked;            /* last un-ACK'ed entry */
357 	u32 s_last;             /* last completed entry */
358 	u32 s_lsn;              /* limit sequence number (credit) */
359 	u16 s_hdrwords;         /* size of s_hdr in 32 bit words */
360 	u16 s_rdma_ack_cnt;
361 	s8 s_ahgidx;
362 	u8 s_state;             /* opcode of last packet sent */
363 	u8 s_ack_state;         /* opcode of packet to ACK */
364 	u8 s_nak_state;         /* non-zero if NAK is pending */
365 	u8 r_nak_state;         /* non-zero if NAK is pending */
366 	u8 s_retry;             /* requester retry counter */
367 	u8 s_rnr_retry;         /* requester RNR retry counter */
368 	u8 s_num_rd_atomic;     /* number of RDMA read/atomic pending */
369 	u8 s_tail_ack_queue;    /* index into s_ack_queue[] */
370 
371 	struct rvt_sge_state s_ack_rdma_sge;
372 	struct timer_list s_timer;
373 
374 	atomic_t local_ops_pending; /* number of fast_reg/local_inv reqs */
375 
376 	/*
377 	 * This sge list MUST be last. Do not add anything below here.
378 	 */
379 	struct rvt_sge r_sg_list[0] /* verified SGEs */
380 		____cacheline_aligned_in_smp;
381 };
382 
383 struct rvt_srq {
384 	struct ib_srq ibsrq;
385 	struct rvt_rq rq;
386 	struct rvt_mmap_info *ip;
387 	/* send signal when number of RWQEs < limit */
388 	u32 limit;
389 };
390 
391 #define RVT_QPN_MAX                 BIT(24)
392 #define RVT_QPNMAP_ENTRIES          (RVT_QPN_MAX / PAGE_SIZE / BITS_PER_BYTE)
393 #define RVT_BITS_PER_PAGE           (PAGE_SIZE * BITS_PER_BYTE)
394 #define RVT_BITS_PER_PAGE_MASK      (RVT_BITS_PER_PAGE - 1)
395 #define RVT_QPN_MASK		    0xFFFFFF
396 
397 /*
398  * QPN-map pages start out as NULL, they get allocated upon
399  * first use and are never deallocated. This way,
400  * large bitmaps are not allocated unless large numbers of QPs are used.
401  */
402 struct rvt_qpn_map {
403 	void *page;
404 };
405 
406 struct rvt_qpn_table {
407 	spinlock_t lock; /* protect changes to the qp table */
408 	unsigned flags;         /* flags for QP0/1 allocated for each port */
409 	u32 last;               /* last QP number allocated */
410 	u32 nmaps;              /* size of the map table */
411 	u16 limit;
412 	u8  incr;
413 	/* bit map of free QP numbers other than 0/1 */
414 	struct rvt_qpn_map map[RVT_QPNMAP_ENTRIES];
415 };
416 
417 struct rvt_qp_ibdev {
418 	u32 qp_table_size;
419 	u32 qp_table_bits;
420 	struct rvt_qp __rcu **qp_table;
421 	spinlock_t qpt_lock; /* qptable lock */
422 	struct rvt_qpn_table qpn_table;
423 };
424 
425 /*
426  * There is one struct rvt_mcast for each multicast GID.
427  * All attached QPs are then stored as a list of
428  * struct rvt_mcast_qp.
429  */
430 struct rvt_mcast_qp {
431 	struct list_head list;
432 	struct rvt_qp *qp;
433 };
434 
435 struct rvt_mcast {
436 	struct rb_node rb_node;
437 	union ib_gid mgid;
438 	struct list_head qp_list;
439 	wait_queue_head_t wait;
440 	atomic_t refcount;
441 	int n_attached;
442 };
443 
444 /*
445  * Since struct rvt_swqe is not a fixed size, we can't simply index into
446  * struct rvt_qp.s_wq.  This function does the array index computation.
447  */
rvt_get_swqe_ptr(struct rvt_qp * qp,unsigned n)448 static inline struct rvt_swqe *rvt_get_swqe_ptr(struct rvt_qp *qp,
449 						unsigned n)
450 {
451 	return (struct rvt_swqe *)((char *)qp->s_wq +
452 				     (sizeof(struct rvt_swqe) +
453 				      qp->s_max_sge *
454 				      sizeof(struct rvt_sge)) * n);
455 }
456 
457 /*
458  * Since struct rvt_rwqe is not a fixed size, we can't simply index into
459  * struct rvt_rwq.wq.  This function does the array index computation.
460  */
rvt_get_rwqe_ptr(struct rvt_rq * rq,unsigned n)461 static inline struct rvt_rwqe *rvt_get_rwqe_ptr(struct rvt_rq *rq, unsigned n)
462 {
463 	return (struct rvt_rwqe *)
464 		((char *)rq->wq->wq +
465 		 (sizeof(struct rvt_rwqe) +
466 		  rq->max_sge * sizeof(struct ib_sge)) * n);
467 }
468 
469 /**
470  * rvt_get_qp - get a QP reference
471  * @qp - the QP to hold
472  */
rvt_get_qp(struct rvt_qp * qp)473 static inline void rvt_get_qp(struct rvt_qp *qp)
474 {
475 	atomic_inc(&qp->refcount);
476 }
477 
478 /**
479  * rvt_put_qp - release a QP reference
480  * @qp - the QP to release
481  */
rvt_put_qp(struct rvt_qp * qp)482 static inline void rvt_put_qp(struct rvt_qp *qp)
483 {
484 	if (qp && atomic_dec_and_test(&qp->refcount))
485 		wake_up(&qp->wait);
486 }
487 
488 /**
489  * rvt_qp_wqe_reserve - reserve operation
490  * @qp - the rvt qp
491  * @wqe - the send wqe
492  *
493  * This routine used in post send to record
494  * a wqe relative reserved operation use.
495  */
rvt_qp_wqe_reserve(struct rvt_qp * qp,struct rvt_swqe * wqe)496 static inline void rvt_qp_wqe_reserve(
497 	struct rvt_qp *qp,
498 	struct rvt_swqe *wqe)
499 {
500 	wqe->wr.send_flags |= RVT_SEND_RESERVE_USED;
501 	atomic_inc(&qp->s_reserved_used);
502 }
503 
504 /**
505  * rvt_qp_wqe_unreserve - clean reserved operation
506  * @qp - the rvt qp
507  * @wqe - the send wqe
508  *
509  * This decrements the reserve use count.
510  *
511  * This call MUST precede the change to
512  * s_last to insure that post send sees a stable
513  * s_avail.
514  *
515  * An smp_mp__after_atomic() is used to insure
516  * the compiler does not juggle the order of the s_last
517  * ring index and the decrementing of s_reserved_used.
518  */
rvt_qp_wqe_unreserve(struct rvt_qp * qp,struct rvt_swqe * wqe)519 static inline void rvt_qp_wqe_unreserve(
520 	struct rvt_qp *qp,
521 	struct rvt_swqe *wqe)
522 {
523 	if (unlikely(wqe->wr.send_flags & RVT_SEND_RESERVE_USED)) {
524 		wqe->wr.send_flags &= ~RVT_SEND_RESERVE_USED;
525 		atomic_dec(&qp->s_reserved_used);
526 		/* insure no compiler re-order up to s_last change */
527 		smp_mb__after_atomic();
528 	}
529 }
530 
531 extern const int  ib_rvt_state_ops[];
532 
533 struct rvt_dev_info;
534 int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err);
535 
536 #endif          /* DEF_RDMAVT_INCQP_H */
537