xref: /openbmc/linux/include/rdma/rdmavt_qp.h (revision 4f205687)
1 #ifndef DEF_RDMAVT_INCQP_H
2 #define DEF_RDMAVT_INCQP_H
3 
4 /*
5  * Copyright(c) 2016 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 /*
55  * Atomic bit definitions for r_aflags.
56  */
57 #define RVT_R_WRID_VALID        0
58 #define RVT_R_REWIND_SGE        1
59 
60 /*
61  * Bit definitions for r_flags.
62  */
63 #define RVT_R_REUSE_SGE 0x01
64 #define RVT_R_RDMAR_SEQ 0x02
65 #define RVT_R_RSP_NAK   0x04
66 #define RVT_R_RSP_SEND  0x08
67 #define RVT_R_COMM_EST  0x10
68 
69 /*
70  * Bit definitions for s_flags.
71  *
72  * RVT_S_SIGNAL_REQ_WR - set if QP send WRs contain completion signaled
73  * RVT_S_BUSY - send tasklet is processing the QP
74  * RVT_S_TIMER - the RC retry timer is active
75  * RVT_S_ACK_PENDING - an ACK is waiting to be sent after RDMA read/atomics
76  * RVT_S_WAIT_FENCE - waiting for all prior RDMA read or atomic SWQEs
77  *                         before processing the next SWQE
78  * RVT_S_WAIT_RDMAR - waiting for a RDMA read or atomic SWQE to complete
79  *                         before processing the next SWQE
80  * RVT_S_WAIT_RNR - waiting for RNR timeout
81  * RVT_S_WAIT_SSN_CREDIT - waiting for RC credits to process next SWQE
82  * RVT_S_WAIT_DMA - waiting for send DMA queue to drain before generating
83  *                  next send completion entry not via send DMA
84  * RVT_S_WAIT_PIO - waiting for a send buffer to be available
85  * RVT_S_WAIT_PIO_DRAIN - waiting for a qp to drain pio packets
86  * RVT_S_WAIT_TX - waiting for a struct verbs_txreq to be available
87  * RVT_S_WAIT_DMA_DESC - waiting for DMA descriptors to be available
88  * RVT_S_WAIT_KMEM - waiting for kernel memory to be available
89  * RVT_S_WAIT_PSN - waiting for a packet to exit the send DMA queue
90  * RVT_S_WAIT_ACK - waiting for an ACK packet before sending more requests
91  * RVT_S_SEND_ONE - send one packet, request ACK, then wait for ACK
92  * RVT_S_ECN - a BECN was queued to the send engine
93  */
94 #define RVT_S_SIGNAL_REQ_WR	0x0001
95 #define RVT_S_BUSY		0x0002
96 #define RVT_S_TIMER		0x0004
97 #define RVT_S_RESP_PENDING	0x0008
98 #define RVT_S_ACK_PENDING	0x0010
99 #define RVT_S_WAIT_FENCE	0x0020
100 #define RVT_S_WAIT_RDMAR	0x0040
101 #define RVT_S_WAIT_RNR		0x0080
102 #define RVT_S_WAIT_SSN_CREDIT	0x0100
103 #define RVT_S_WAIT_DMA		0x0200
104 #define RVT_S_WAIT_PIO		0x0400
105 #define RVT_S_WAIT_PIO_DRAIN    0x0800
106 #define RVT_S_WAIT_TX		0x1000
107 #define RVT_S_WAIT_DMA_DESC	0x2000
108 #define RVT_S_WAIT_KMEM		0x4000
109 #define RVT_S_WAIT_PSN		0x8000
110 #define RVT_S_WAIT_ACK		0x10000
111 #define RVT_S_SEND_ONE		0x20000
112 #define RVT_S_UNLIMITED_CREDIT	0x40000
113 #define RVT_S_AHG_VALID		0x80000
114 #define RVT_S_AHG_CLEAR		0x100000
115 #define RVT_S_ECN		0x200000
116 
117 /*
118  * Wait flags that would prevent any packet type from being sent.
119  */
120 #define RVT_S_ANY_WAIT_IO \
121 	(RVT_S_WAIT_PIO | RVT_S_WAIT_PIO_DRAIN | RVT_S_WAIT_TX | \
122 	 RVT_S_WAIT_DMA_DESC | RVT_S_WAIT_KMEM)
123 
124 /*
125  * Wait flags that would prevent send work requests from making progress.
126  */
127 #define RVT_S_ANY_WAIT_SEND (RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR | \
128 	RVT_S_WAIT_RNR | RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_DMA | \
129 	RVT_S_WAIT_PSN | RVT_S_WAIT_ACK)
130 
131 #define RVT_S_ANY_WAIT (RVT_S_ANY_WAIT_IO | RVT_S_ANY_WAIT_SEND)
132 
133 /* Number of bits to pay attention to in the opcode for checking qp type */
134 #define RVT_OPCODE_QP_MASK 0xE0
135 
136 /* Flags for checking QP state (see ib_rvt_state_ops[]) */
137 #define RVT_POST_SEND_OK                0x01
138 #define RVT_POST_RECV_OK                0x02
139 #define RVT_PROCESS_RECV_OK             0x04
140 #define RVT_PROCESS_SEND_OK             0x08
141 #define RVT_PROCESS_NEXT_SEND_OK        0x10
142 #define RVT_FLUSH_SEND			0x20
143 #define RVT_FLUSH_RECV			0x40
144 #define RVT_PROCESS_OR_FLUSH_SEND \
145 	(RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND)
146 
147 /*
148  * Send work request queue entry.
149  * The size of the sg_list is determined when the QP is created and stored
150  * in qp->s_max_sge.
151  */
152 struct rvt_swqe {
153 	union {
154 		struct ib_send_wr wr;   /* don't use wr.sg_list */
155 		struct ib_ud_wr ud_wr;
156 		struct ib_reg_wr reg_wr;
157 		struct ib_rdma_wr rdma_wr;
158 		struct ib_atomic_wr atomic_wr;
159 	};
160 	u32 psn;                /* first packet sequence number */
161 	u32 lpsn;               /* last packet sequence number */
162 	u32 ssn;                /* send sequence number */
163 	u32 length;             /* total length of data in sg_list */
164 	struct rvt_sge sg_list[0];
165 };
166 
167 /*
168  * Receive work request queue entry.
169  * The size of the sg_list is determined when the QP (or SRQ) is created
170  * and stored in qp->r_rq.max_sge (or srq->rq.max_sge).
171  */
172 struct rvt_rwqe {
173 	u64 wr_id;
174 	u8 num_sge;
175 	struct ib_sge sg_list[0];
176 };
177 
178 /*
179  * This structure is used to contain the head pointer, tail pointer,
180  * and receive work queue entries as a single memory allocation so
181  * it can be mmap'ed into user space.
182  * Note that the wq array elements are variable size so you can't
183  * just index into the array to get the N'th element;
184  * use get_rwqe_ptr() instead.
185  */
186 struct rvt_rwq {
187 	u32 head;               /* new work requests posted to the head */
188 	u32 tail;               /* receives pull requests from here. */
189 	struct rvt_rwqe wq[0];
190 };
191 
192 struct rvt_rq {
193 	struct rvt_rwq *wq;
194 	u32 size;               /* size of RWQE array */
195 	u8 max_sge;
196 	/* protect changes in this struct */
197 	spinlock_t lock ____cacheline_aligned_in_smp;
198 };
199 
200 /*
201  * This structure is used by rvt_mmap() to validate an offset
202  * when an mmap() request is made.  The vm_area_struct then uses
203  * this as its vm_private_data.
204  */
205 struct rvt_mmap_info {
206 	struct list_head pending_mmaps;
207 	struct ib_ucontext *context;
208 	void *obj;
209 	__u64 offset;
210 	struct kref ref;
211 	unsigned size;
212 };
213 
214 /*
215  * This structure holds the information that the send tasklet needs
216  * to send a RDMA read response or atomic operation.
217  */
218 struct rvt_ack_entry {
219 	u8 opcode;
220 	u8 sent;
221 	u32 psn;
222 	u32 lpsn;
223 	union {
224 		struct rvt_sge rdma_sge;
225 		u64 atomic_data;
226 	};
227 };
228 
229 #define	RC_QP_SCALING_INTERVAL	5
230 
231 /*
232  * Variables prefixed with s_ are for the requester (sender).
233  * Variables prefixed with r_ are for the responder (receiver).
234  * Variables prefixed with ack_ are for responder replies.
235  *
236  * Common variables are protected by both r_rq.lock and s_lock in that order
237  * which only happens in modify_qp() or changing the QP 'state'.
238  */
239 struct rvt_qp {
240 	struct ib_qp ibqp;
241 	void *priv; /* Driver private data */
242 	/* read mostly fields above and below */
243 	struct ib_ah_attr remote_ah_attr;
244 	struct ib_ah_attr alt_ah_attr;
245 	struct rvt_qp __rcu *next;           /* link list for QPN hash table */
246 	struct rvt_swqe *s_wq;  /* send work queue */
247 	struct rvt_mmap_info *ip;
248 
249 	unsigned long timeout_jiffies;  /* computed from timeout */
250 
251 	enum ib_mtu path_mtu;
252 	int srate_mbps;		/* s_srate (below) converted to Mbit/s */
253 	pid_t pid;		/* pid for user mode QPs */
254 	u32 remote_qpn;
255 	u32 qkey;               /* QKEY for this QP (for UD or RD) */
256 	u32 s_size;             /* send work queue size */
257 	u32 s_ahgpsn;           /* set to the psn in the copy of the header */
258 
259 	u16 pmtu;		/* decoded from path_mtu */
260 	u8 log_pmtu;		/* shift for pmtu */
261 	u8 state;               /* QP state */
262 	u8 allowed_ops;		/* high order bits of allowed opcodes */
263 	u8 qp_access_flags;
264 	u8 alt_timeout;         /* Alternate path timeout for this QP */
265 	u8 timeout;             /* Timeout for this QP */
266 	u8 s_srate;
267 	u8 s_mig_state;
268 	u8 port_num;
269 	u8 s_pkey_index;        /* PKEY index to use */
270 	u8 s_alt_pkey_index;    /* Alternate path PKEY index to use */
271 	u8 r_max_rd_atomic;     /* max number of RDMA read/atomic to receive */
272 	u8 s_max_rd_atomic;     /* max number of RDMA read/atomic to send */
273 	u8 s_retry_cnt;         /* number of times to retry */
274 	u8 s_rnr_retry_cnt;
275 	u8 r_min_rnr_timer;     /* retry timeout value for RNR NAKs */
276 	u8 s_max_sge;           /* size of s_wq->sg_list */
277 	u8 s_draining;
278 
279 	/* start of read/write fields */
280 	atomic_t refcount ____cacheline_aligned_in_smp;
281 	wait_queue_head_t wait;
282 
283 	struct rvt_ack_entry *s_ack_queue;
284 	struct rvt_sge_state s_rdma_read_sge;
285 
286 	spinlock_t r_lock ____cacheline_aligned_in_smp;      /* used for APM */
287 	u32 r_psn;              /* expected rcv packet sequence number */
288 	unsigned long r_aflags;
289 	u64 r_wr_id;            /* ID for current receive WQE */
290 	u32 r_ack_psn;          /* PSN for next ACK or atomic ACK */
291 	u32 r_len;              /* total length of r_sge */
292 	u32 r_rcv_len;          /* receive data len processed */
293 	u32 r_msn;              /* message sequence number */
294 
295 	u8 r_state;             /* opcode of last packet received */
296 	u8 r_flags;
297 	u8 r_head_ack_queue;    /* index into s_ack_queue[] */
298 
299 	struct list_head rspwait;       /* link for waiting to respond */
300 
301 	struct rvt_sge_state r_sge;     /* current receive data */
302 	struct rvt_rq r_rq;             /* receive work queue */
303 
304 	/* post send line */
305 	spinlock_t s_hlock ____cacheline_aligned_in_smp;
306 	u32 s_head;             /* new entries added here */
307 	u32 s_next_psn;         /* PSN for next request */
308 	u32 s_avail;            /* number of entries avail */
309 	u32 s_ssn;              /* SSN of tail entry */
310 
311 	spinlock_t s_lock ____cacheline_aligned_in_smp;
312 	u32 s_flags;
313 	struct rvt_sge_state *s_cur_sge;
314 	struct rvt_swqe *s_wqe;
315 	struct rvt_sge_state s_sge;     /* current send request data */
316 	struct rvt_mregion *s_rdma_mr;
317 	u32 s_cur_size;         /* size of send packet in bytes */
318 	u32 s_len;              /* total length of s_sge */
319 	u32 s_rdma_read_len;    /* total length of s_rdma_read_sge */
320 	u32 s_last_psn;         /* last response PSN processed */
321 	u32 s_sending_psn;      /* lowest PSN that is being sent */
322 	u32 s_sending_hpsn;     /* highest PSN that is being sent */
323 	u32 s_psn;              /* current packet sequence number */
324 	u32 s_ack_rdma_psn;     /* PSN for sending RDMA read responses */
325 	u32 s_ack_psn;          /* PSN for acking sends and RDMA writes */
326 	u32 s_tail;             /* next entry to process */
327 	u32 s_cur;              /* current work queue entry */
328 	u32 s_acked;            /* last un-ACK'ed entry */
329 	u32 s_last;             /* last completed entry */
330 	u32 s_lsn;              /* limit sequence number (credit) */
331 	u16 s_hdrwords;         /* size of s_hdr in 32 bit words */
332 	u16 s_rdma_ack_cnt;
333 	s8 s_ahgidx;
334 	u8 s_state;             /* opcode of last packet sent */
335 	u8 s_ack_state;         /* opcode of packet to ACK */
336 	u8 s_nak_state;         /* non-zero if NAK is pending */
337 	u8 r_nak_state;         /* non-zero if NAK is pending */
338 	u8 s_retry;             /* requester retry counter */
339 	u8 s_rnr_retry;         /* requester RNR retry counter */
340 	u8 s_num_rd_atomic;     /* number of RDMA read/atomic pending */
341 	u8 s_tail_ack_queue;    /* index into s_ack_queue[] */
342 
343 	struct rvt_sge_state s_ack_rdma_sge;
344 	struct timer_list s_timer;
345 
346 	/*
347 	 * This sge list MUST be last. Do not add anything below here.
348 	 */
349 	struct rvt_sge r_sg_list[0] /* verified SGEs */
350 		____cacheline_aligned_in_smp;
351 };
352 
353 struct rvt_srq {
354 	struct ib_srq ibsrq;
355 	struct rvt_rq rq;
356 	struct rvt_mmap_info *ip;
357 	/* send signal when number of RWQEs < limit */
358 	u32 limit;
359 };
360 
361 #define RVT_QPN_MAX                 BIT(24)
362 #define RVT_QPNMAP_ENTRIES          (RVT_QPN_MAX / PAGE_SIZE / BITS_PER_BYTE)
363 #define RVT_BITS_PER_PAGE           (PAGE_SIZE * BITS_PER_BYTE)
364 #define RVT_BITS_PER_PAGE_MASK      (RVT_BITS_PER_PAGE - 1)
365 #define RVT_QPN_MASK		    0xFFFFFF
366 
367 /*
368  * QPN-map pages start out as NULL, they get allocated upon
369  * first use and are never deallocated. This way,
370  * large bitmaps are not allocated unless large numbers of QPs are used.
371  */
372 struct rvt_qpn_map {
373 	void *page;
374 };
375 
376 struct rvt_qpn_table {
377 	spinlock_t lock; /* protect changes to the qp table */
378 	unsigned flags;         /* flags for QP0/1 allocated for each port */
379 	u32 last;               /* last QP number allocated */
380 	u32 nmaps;              /* size of the map table */
381 	u16 limit;
382 	u8  incr;
383 	/* bit map of free QP numbers other than 0/1 */
384 	struct rvt_qpn_map map[RVT_QPNMAP_ENTRIES];
385 };
386 
387 struct rvt_qp_ibdev {
388 	u32 qp_table_size;
389 	u32 qp_table_bits;
390 	struct rvt_qp __rcu **qp_table;
391 	spinlock_t qpt_lock; /* qptable lock */
392 	struct rvt_qpn_table qpn_table;
393 };
394 
395 /*
396  * There is one struct rvt_mcast for each multicast GID.
397  * All attached QPs are then stored as a list of
398  * struct rvt_mcast_qp.
399  */
400 struct rvt_mcast_qp {
401 	struct list_head list;
402 	struct rvt_qp *qp;
403 };
404 
405 struct rvt_mcast {
406 	struct rb_node rb_node;
407 	union ib_gid mgid;
408 	struct list_head qp_list;
409 	wait_queue_head_t wait;
410 	atomic_t refcount;
411 	int n_attached;
412 };
413 
414 /*
415  * Since struct rvt_swqe is not a fixed size, we can't simply index into
416  * struct rvt_qp.s_wq.  This function does the array index computation.
417  */
418 static inline struct rvt_swqe *rvt_get_swqe_ptr(struct rvt_qp *qp,
419 						unsigned n)
420 {
421 	return (struct rvt_swqe *)((char *)qp->s_wq +
422 				     (sizeof(struct rvt_swqe) +
423 				      qp->s_max_sge *
424 				      sizeof(struct rvt_sge)) * n);
425 }
426 
427 /*
428  * Since struct rvt_rwqe is not a fixed size, we can't simply index into
429  * struct rvt_rwq.wq.  This function does the array index computation.
430  */
431 static inline struct rvt_rwqe *rvt_get_rwqe_ptr(struct rvt_rq *rq, unsigned n)
432 {
433 	return (struct rvt_rwqe *)
434 		((char *)rq->wq->wq +
435 		 (sizeof(struct rvt_rwqe) +
436 		  rq->max_sge * sizeof(struct ib_sge)) * n);
437 }
438 
439 extern const int  ib_rvt_state_ops[];
440 
441 struct rvt_dev_info;
442 int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err);
443 
444 #endif          /* DEF_RDMAVT_INCQP_H */
445