Loading...
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 (RVT_S_WAIT_PIO | RVT_S_WAIT_TX | \
121 RVT_S_WAIT_DMA_DESC | RVT_S_WAIT_KMEM)
122
123/*
124 * Wait flags that would prevent send work requests from making progress.
125 */
126#define RVT_S_ANY_WAIT_SEND (RVT_S_WAIT_FENCE | RVT_S_WAIT_RDMAR | \
127 RVT_S_WAIT_RNR | RVT_S_WAIT_SSN_CREDIT | RVT_S_WAIT_DMA | \
128 RVT_S_WAIT_PSN | RVT_S_WAIT_ACK)
129
130#define RVT_S_ANY_WAIT (RVT_S_ANY_WAIT_IO | RVT_S_ANY_WAIT_SEND)
131
132/* Number of bits to pay attention to in the opcode for checking qp type */
133#define RVT_OPCODE_QP_MASK 0xE0
134
135/* Flags for checking QP state (see ib_rvt_state_ops[]) */
136#define RVT_POST_SEND_OK 0x01
137#define RVT_POST_RECV_OK 0x02
138#define RVT_PROCESS_RECV_OK 0x04
139#define RVT_PROCESS_SEND_OK 0x08
140#define RVT_PROCESS_NEXT_SEND_OK 0x10
141#define RVT_FLUSH_SEND 0x20
142#define RVT_FLUSH_RECV 0x40
143#define RVT_PROCESS_OR_FLUSH_SEND \
144 (RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND)
145
146/*
147 * Send work request queue entry.
148 * The size of the sg_list is determined when the QP is created and stored
149 * in qp->s_max_sge.
150 */
151struct rvt_swqe {
152 union {
153 struct ib_send_wr wr; /* don't use wr.sg_list */
154 struct ib_ud_wr ud_wr;
155 struct ib_reg_wr reg_wr;
156 struct ib_rdma_wr rdma_wr;
157 struct ib_atomic_wr atomic_wr;
158 };
159 u32 psn; /* first packet sequence number */
160 u32 lpsn; /* last packet sequence number */
161 u32 ssn; /* send sequence number */
162 u32 length; /* total length of data in sg_list */
163 struct rvt_sge sg_list[0];
164};
165
166/*
167 * Receive work request queue entry.
168 * The size of the sg_list is determined when the QP (or SRQ) is created
169 * and stored in qp->r_rq.max_sge (or srq->rq.max_sge).
170 */
171struct rvt_rwqe {
172 u64 wr_id;
173 u8 num_sge;
174 struct ib_sge sg_list[0];
175};
176
177/*
178 * This structure is used to contain the head pointer, tail pointer,
179 * and receive work queue entries as a single memory allocation so
180 * it can be mmap'ed into user space.
181 * Note that the wq array elements are variable size so you can't
182 * just index into the array to get the N'th element;
183 * use get_rwqe_ptr() instead.
184 */
185struct rvt_rwq {
186 u32 head; /* new work requests posted to the head */
187 u32 tail; /* receives pull requests from here. */
188 struct rvt_rwqe wq[0];
189};
190
191struct rvt_rq {
192 struct rvt_rwq *wq;
193 u32 size; /* size of RWQE array */
194 u8 max_sge;
195 /* protect changes in this struct */
196 spinlock_t lock ____cacheline_aligned_in_smp;
197};
198
199/*
200 * This structure is used by rvt_mmap() to validate an offset
201 * when an mmap() request is made. The vm_area_struct then uses
202 * this as its vm_private_data.
203 */
204struct rvt_mmap_info {
205 struct list_head pending_mmaps;
206 struct ib_ucontext *context;
207 void *obj;
208 __u64 offset;
209 struct kref ref;
210 unsigned size;
211};
212
213#define RVT_MAX_RDMA_ATOMIC 16
214
215/*
216 * This structure holds the information that the send tasklet needs
217 * to send a RDMA read response or atomic operation.
218 */
219struct rvt_ack_entry {
220 u8 opcode;
221 u8 sent;
222 u32 psn;
223 u32 lpsn;
224 union {
225 struct rvt_sge rdma_sge;
226 u64 atomic_data;
227 };
228};
229
230#define RC_QP_SCALING_INTERVAL 5
231
232/*
233 * Variables prefixed with s_ are for the requester (sender).
234 * Variables prefixed with r_ are for the responder (receiver).
235 * Variables prefixed with ack_ are for responder replies.
236 *
237 * Common variables are protected by both r_rq.lock and s_lock in that order
238 * which only happens in modify_qp() or changing the QP 'state'.
239 */
240struct rvt_qp {
241 struct ib_qp ibqp;
242 void *priv; /* Driver private data */
243 /* read mostly fields above and below */
244 struct ib_ah_attr remote_ah_attr;
245 struct ib_ah_attr alt_ah_attr;
246 struct rvt_qp __rcu *next; /* link list for QPN hash table */
247 struct rvt_swqe *s_wq; /* send work queue */
248 struct rvt_mmap_info *ip;
249
250 unsigned long timeout_jiffies; /* computed from timeout */
251
252 enum ib_mtu path_mtu;
253 int srate_mbps; /* s_srate (below) converted to Mbit/s */
254 pid_t pid; /* pid for user mode QPs */
255 u32 remote_qpn;
256 u32 qkey; /* QKEY for this QP (for UD or RD) */
257 u32 s_size; /* send work queue size */
258 u32 s_ahgpsn; /* set to the psn in the copy of the header */
259
260 u16 pmtu; /* decoded from path_mtu */
261 u8 log_pmtu; /* shift for pmtu */
262 u8 state; /* QP state */
263 u8 allowed_ops; /* high order bits of allowed opcodes */
264 u8 qp_access_flags;
265 u8 alt_timeout; /* Alternate path timeout for this QP */
266 u8 timeout; /* Timeout for this QP */
267 u8 s_srate;
268 u8 s_mig_state;
269 u8 port_num;
270 u8 s_pkey_index; /* PKEY index to use */
271 u8 s_alt_pkey_index; /* Alternate path PKEY index to use */
272 u8 r_max_rd_atomic; /* max number of RDMA read/atomic to receive */
273 u8 s_max_rd_atomic; /* max number of RDMA read/atomic to send */
274 u8 s_retry_cnt; /* number of times to retry */
275 u8 s_rnr_retry_cnt;
276 u8 r_min_rnr_timer; /* retry timeout value for RNR NAKs */
277 u8 s_max_sge; /* size of s_wq->sg_list */
278 u8 s_draining;
279
280 /* start of read/write fields */
281 atomic_t refcount ____cacheline_aligned_in_smp;
282 wait_queue_head_t wait;
283
284 struct rvt_ack_entry s_ack_queue[RVT_MAX_RDMA_ATOMIC + 1]
285 ____cacheline_aligned_in_smp;
286 struct rvt_sge_state s_rdma_read_sge;
287
288 spinlock_t r_lock ____cacheline_aligned_in_smp; /* used for APM */
289 u32 r_psn; /* expected rcv packet sequence number */
290 unsigned long r_aflags;
291 u64 r_wr_id; /* ID for current receive WQE */
292 u32 r_ack_psn; /* PSN for next ACK or atomic ACK */
293 u32 r_len; /* total length of r_sge */
294 u32 r_rcv_len; /* receive data len processed */
295 u32 r_msn; /* message sequence number */
296
297 u8 r_state; /* opcode of last packet received */
298 u8 r_flags;
299 u8 r_head_ack_queue; /* index into s_ack_queue[] */
300
301 struct list_head rspwait; /* link for waiting to respond */
302
303 struct rvt_sge_state r_sge; /* current receive data */
304 struct rvt_rq r_rq; /* receive work queue */
305
306 /* post send line */
307 spinlock_t s_hlock ____cacheline_aligned_in_smp;
308 u32 s_head; /* new entries added here */
309 u32 s_next_psn; /* PSN for next request */
310 u32 s_avail; /* number of entries avail */
311 u32 s_ssn; /* SSN of tail entry */
312
313 spinlock_t s_lock ____cacheline_aligned_in_smp;
314 u32 s_flags;
315 struct rvt_sge_state *s_cur_sge;
316 struct rvt_swqe *s_wqe;
317 struct rvt_sge_state s_sge; /* current send request data */
318 struct rvt_mregion *s_rdma_mr;
319 u32 s_cur_size; /* size of send packet in bytes */
320 u32 s_len; /* total length of s_sge */
321 u32 s_rdma_read_len; /* total length of s_rdma_read_sge */
322 u32 s_last_psn; /* last response PSN processed */
323 u32 s_sending_psn; /* lowest PSN that is being sent */
324 u32 s_sending_hpsn; /* highest PSN that is being sent */
325 u32 s_psn; /* current packet sequence number */
326 u32 s_ack_rdma_psn; /* PSN for sending RDMA read responses */
327 u32 s_ack_psn; /* PSN for acking sends and RDMA writes */
328 u32 s_tail; /* next entry to process */
329 u32 s_cur; /* current work queue entry */
330 u32 s_acked; /* last un-ACK'ed entry */
331 u32 s_last; /* last completed entry */
332 u32 s_lsn; /* limit sequence number (credit) */
333 u16 s_hdrwords; /* size of s_hdr in 32 bit words */
334 u16 s_rdma_ack_cnt;
335 s8 s_ahgidx;
336 u8 s_state; /* opcode of last packet sent */
337 u8 s_ack_state; /* opcode of packet to ACK */
338 u8 s_nak_state; /* non-zero if NAK is pending */
339 u8 r_nak_state; /* non-zero if NAK is pending */
340 u8 s_retry; /* requester retry counter */
341 u8 s_rnr_retry; /* requester RNR retry counter */
342 u8 s_num_rd_atomic; /* number of RDMA read/atomic pending */
343 u8 s_tail_ack_queue; /* index into s_ack_queue[] */
344
345 struct rvt_sge_state s_ack_rdma_sge;
346 struct timer_list s_timer;
347
348 /*
349 * This sge list MUST be last. Do not add anything below here.
350 */
351 struct rvt_sge r_sg_list[0] /* verified SGEs */
352 ____cacheline_aligned_in_smp;
353};
354
355struct rvt_srq {
356 struct ib_srq ibsrq;
357 struct rvt_rq rq;
358 struct rvt_mmap_info *ip;
359 /* send signal when number of RWQEs < limit */
360 u32 limit;
361};
362
363#define RVT_QPN_MAX BIT(24)
364#define RVT_QPNMAP_ENTRIES (RVT_QPN_MAX / PAGE_SIZE / BITS_PER_BYTE)
365#define RVT_BITS_PER_PAGE (PAGE_SIZE * BITS_PER_BYTE)
366#define RVT_BITS_PER_PAGE_MASK (RVT_BITS_PER_PAGE - 1)
367#define RVT_QPN_MASK 0xFFFFFF
368
369/*
370 * QPN-map pages start out as NULL, they get allocated upon
371 * first use and are never deallocated. This way,
372 * large bitmaps are not allocated unless large numbers of QPs are used.
373 */
374struct rvt_qpn_map {
375 void *page;
376};
377
378struct rvt_qpn_table {
379 spinlock_t lock; /* protect changes to the qp table */
380 unsigned flags; /* flags for QP0/1 allocated for each port */
381 u32 last; /* last QP number allocated */
382 u32 nmaps; /* size of the map table */
383 u16 limit;
384 u8 incr;
385 /* bit map of free QP numbers other than 0/1 */
386 struct rvt_qpn_map map[RVT_QPNMAP_ENTRIES];
387};
388
389struct rvt_qp_ibdev {
390 u32 qp_table_size;
391 u32 qp_table_bits;
392 struct rvt_qp __rcu **qp_table;
393 spinlock_t qpt_lock; /* qptable lock */
394 struct rvt_qpn_table qpn_table;
395};
396
397/*
398 * There is one struct rvt_mcast for each multicast GID.
399 * All attached QPs are then stored as a list of
400 * struct rvt_mcast_qp.
401 */
402struct rvt_mcast_qp {
403 struct list_head list;
404 struct rvt_qp *qp;
405};
406
407struct rvt_mcast {
408 struct rb_node rb_node;
409 union ib_gid mgid;
410 struct list_head qp_list;
411 wait_queue_head_t wait;
412 atomic_t refcount;
413 int n_attached;
414};
415
416/*
417 * Since struct rvt_swqe is not a fixed size, we can't simply index into
418 * struct rvt_qp.s_wq. This function does the array index computation.
419 */
420static inline struct rvt_swqe *rvt_get_swqe_ptr(struct rvt_qp *qp,
421 unsigned n)
422{
423 return (struct rvt_swqe *)((char *)qp->s_wq +
424 (sizeof(struct rvt_swqe) +
425 qp->s_max_sge *
426 sizeof(struct rvt_sge)) * n);
427}
428
429/*
430 * Since struct rvt_rwqe is not a fixed size, we can't simply index into
431 * struct rvt_rwq.wq. This function does the array index computation.
432 */
433static inline struct rvt_rwqe *rvt_get_rwqe_ptr(struct rvt_rq *rq, unsigned n)
434{
435 return (struct rvt_rwqe *)
436 ((char *)rq->wq->wq +
437 (sizeof(struct rvt_rwqe) +
438 rq->max_sge * sizeof(struct ib_sge)) * n);
439}
440
441extern const int ib_rvt_state_ops[];
442
443struct rvt_dev_info;
444int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err);
445
446#endif /* DEF_RDMAVT_INCQP_H */
1#ifndef DEF_RDMAVT_INCQP_H
2#define DEF_RDMAVT_INCQP_H
3
4/*
5 * Copyright(c) 2016, 2017 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/*
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#define RVT_SEND_OR_FLUSH_OR_RECV_OK \
148 (RVT_PROCESS_SEND_OK | RVT_FLUSH_SEND | RVT_PROCESS_RECV_OK)
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 */
161struct 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 */
181struct 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 */
195struct 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
201struct 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 */
214struct 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 */
227struct 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
258struct 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 */
268struct rvt_qp {
269 struct ib_qp ibqp;
270 void *priv; /* Driver private data */
271 /* read mostly fields above and below */
272 struct rdma_ah_attr remote_ah_attr;
273 struct rdma_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 int srate_mbps; /* s_srate (below) converted to Mbit/s */
281 pid_t pid; /* pid for user mode QPs */
282 u32 remote_qpn;
283 u32 qkey; /* QKEY for this QP (for UD or RD) */
284 u32 s_size; /* send work queue size */
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 u8 r_adefered; /* defered ack count */
326
327 struct list_head rspwait; /* link for waiting to respond */
328
329 struct rvt_sge_state r_sge; /* current receive data */
330 struct rvt_rq r_rq; /* receive work queue */
331
332 /* post send line */
333 spinlock_t s_hlock ____cacheline_aligned_in_smp;
334 u32 s_head; /* new entries added here */
335 u32 s_next_psn; /* PSN for next request */
336 u32 s_avail; /* number of entries avail */
337 u32 s_ssn; /* SSN of tail entry */
338 atomic_t s_reserved_used; /* reserved entries in use */
339
340 spinlock_t s_lock ____cacheline_aligned_in_smp;
341 u32 s_flags;
342 struct rvt_sge_state *s_cur_sge;
343 struct rvt_swqe *s_wqe;
344 struct rvt_sge_state s_sge; /* current send request data */
345 struct rvt_mregion *s_rdma_mr;
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 u32 s_ahgpsn; /* set to the psn in the copy of the header */
360 u16 s_cur_size; /* size of send packet in bytes */
361 u16 s_rdma_ack_cnt;
362 u8 s_hdrwords; /* size of s_hdr in 32 bit words */
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 struct hrtimer s_rnr_timer;
376
377 atomic_t local_ops_pending; /* number of fast_reg/local_inv reqs */
378
379 /*
380 * This sge list MUST be last. Do not add anything below here.
381 */
382 struct rvt_sge r_sg_list[0] /* verified SGEs */
383 ____cacheline_aligned_in_smp;
384};
385
386struct rvt_srq {
387 struct ib_srq ibsrq;
388 struct rvt_rq rq;
389 struct rvt_mmap_info *ip;
390 /* send signal when number of RWQEs < limit */
391 u32 limit;
392};
393
394#define RVT_QPN_MAX BIT(24)
395#define RVT_QPNMAP_ENTRIES (RVT_QPN_MAX / PAGE_SIZE / BITS_PER_BYTE)
396#define RVT_BITS_PER_PAGE (PAGE_SIZE * BITS_PER_BYTE)
397#define RVT_BITS_PER_PAGE_MASK (RVT_BITS_PER_PAGE - 1)
398#define RVT_QPN_MASK IB_QPN_MASK
399
400/*
401 * QPN-map pages start out as NULL, they get allocated upon
402 * first use and are never deallocated. This way,
403 * large bitmaps are not allocated unless large numbers of QPs are used.
404 */
405struct rvt_qpn_map {
406 void *page;
407};
408
409struct rvt_qpn_table {
410 spinlock_t lock; /* protect changes to the qp table */
411 unsigned flags; /* flags for QP0/1 allocated for each port */
412 u32 last; /* last QP number allocated */
413 u32 nmaps; /* size of the map table */
414 u16 limit;
415 u8 incr;
416 /* bit map of free QP numbers other than 0/1 */
417 struct rvt_qpn_map map[RVT_QPNMAP_ENTRIES];
418};
419
420struct rvt_qp_ibdev {
421 u32 qp_table_size;
422 u32 qp_table_bits;
423 struct rvt_qp __rcu **qp_table;
424 spinlock_t qpt_lock; /* qptable lock */
425 struct rvt_qpn_table qpn_table;
426};
427
428/*
429 * There is one struct rvt_mcast for each multicast GID.
430 * All attached QPs are then stored as a list of
431 * struct rvt_mcast_qp.
432 */
433struct rvt_mcast_qp {
434 struct list_head list;
435 struct rvt_qp *qp;
436};
437
438struct rvt_mcast_addr {
439 union ib_gid mgid;
440 u16 lid;
441};
442
443struct rvt_mcast {
444 struct rb_node rb_node;
445 struct rvt_mcast_addr mcast_addr;
446 struct list_head qp_list;
447 wait_queue_head_t wait;
448 atomic_t refcount;
449 int n_attached;
450};
451
452/*
453 * Since struct rvt_swqe is not a fixed size, we can't simply index into
454 * struct rvt_qp.s_wq. This function does the array index computation.
455 */
456static inline struct rvt_swqe *rvt_get_swqe_ptr(struct rvt_qp *qp,
457 unsigned n)
458{
459 return (struct rvt_swqe *)((char *)qp->s_wq +
460 (sizeof(struct rvt_swqe) +
461 qp->s_max_sge *
462 sizeof(struct rvt_sge)) * n);
463}
464
465/*
466 * Since struct rvt_rwqe is not a fixed size, we can't simply index into
467 * struct rvt_rwq.wq. This function does the array index computation.
468 */
469static inline struct rvt_rwqe *rvt_get_rwqe_ptr(struct rvt_rq *rq, unsigned n)
470{
471 return (struct rvt_rwqe *)
472 ((char *)rq->wq->wq +
473 (sizeof(struct rvt_rwqe) +
474 rq->max_sge * sizeof(struct ib_sge)) * n);
475}
476
477/**
478 * rvt_is_user_qp - return if this is user mode QP
479 * @qp - the target QP
480 */
481static inline bool rvt_is_user_qp(struct rvt_qp *qp)
482{
483 return !!qp->pid;
484}
485
486/**
487 * rvt_get_qp - get a QP reference
488 * @qp - the QP to hold
489 */
490static inline void rvt_get_qp(struct rvt_qp *qp)
491{
492 atomic_inc(&qp->refcount);
493}
494
495/**
496 * rvt_put_qp - release a QP reference
497 * @qp - the QP to release
498 */
499static inline void rvt_put_qp(struct rvt_qp *qp)
500{
501 if (qp && atomic_dec_and_test(&qp->refcount))
502 wake_up(&qp->wait);
503}
504
505/**
506 * rvt_put_swqe - drop mr refs held by swqe
507 * @wqe - the send wqe
508 *
509 * This drops any mr references held by the swqe
510 */
511static inline void rvt_put_swqe(struct rvt_swqe *wqe)
512{
513 int i;
514
515 for (i = 0; i < wqe->wr.num_sge; i++) {
516 struct rvt_sge *sge = &wqe->sg_list[i];
517
518 rvt_put_mr(sge->mr);
519 }
520}
521
522/**
523 * rvt_qp_wqe_reserve - reserve operation
524 * @qp - the rvt qp
525 * @wqe - the send wqe
526 *
527 * This routine used in post send to record
528 * a wqe relative reserved operation use.
529 */
530static inline void rvt_qp_wqe_reserve(
531 struct rvt_qp *qp,
532 struct rvt_swqe *wqe)
533{
534 atomic_inc(&qp->s_reserved_used);
535}
536
537/**
538 * rvt_qp_wqe_unreserve - clean reserved operation
539 * @qp - the rvt qp
540 * @wqe - the send wqe
541 *
542 * This decrements the reserve use count.
543 *
544 * This call MUST precede the change to
545 * s_last to insure that post send sees a stable
546 * s_avail.
547 *
548 * An smp_mp__after_atomic() is used to insure
549 * the compiler does not juggle the order of the s_last
550 * ring index and the decrementing of s_reserved_used.
551 */
552static inline void rvt_qp_wqe_unreserve(
553 struct rvt_qp *qp,
554 struct rvt_swqe *wqe)
555{
556 if (unlikely(wqe->wr.send_flags & RVT_SEND_RESERVE_USED)) {
557 atomic_dec(&qp->s_reserved_used);
558 /* insure no compiler re-order up to s_last change */
559 smp_mb__after_atomic();
560 }
561}
562
563extern const enum ib_wc_opcode ib_rvt_wc_opcode[];
564
565/**
566 * rvt_qp_swqe_complete() - insert send completion
567 * @qp - the qp
568 * @wqe - the send wqe
569 * @status - completion status
570 *
571 * Insert a send completion into the completion
572 * queue if the qp indicates it should be done.
573 *
574 * See IBTA 10.7.3.1 for info on completion
575 * control.
576 */
577static inline void rvt_qp_swqe_complete(
578 struct rvt_qp *qp,
579 struct rvt_swqe *wqe,
580 enum ib_wc_opcode opcode,
581 enum ib_wc_status status)
582{
583 if (unlikely(wqe->wr.send_flags & RVT_SEND_RESERVE_USED))
584 return;
585 if (!(qp->s_flags & RVT_S_SIGNAL_REQ_WR) ||
586 (wqe->wr.send_flags & IB_SEND_SIGNALED) ||
587 status != IB_WC_SUCCESS) {
588 struct ib_wc wc;
589
590 memset(&wc, 0, sizeof(wc));
591 wc.wr_id = wqe->wr.wr_id;
592 wc.status = status;
593 wc.opcode = opcode;
594 wc.qp = &qp->ibqp;
595 wc.byte_len = wqe->length;
596 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.send_cq), &wc,
597 status != IB_WC_SUCCESS);
598 }
599}
600
601/*
602 * Compare the lower 24 bits of the msn values.
603 * Returns an integer <, ==, or > than zero.
604 */
605static inline int rvt_cmp_msn(u32 a, u32 b)
606{
607 return (((int)a) - ((int)b)) << 8;
608}
609
610/**
611 * rvt_compute_aeth - compute the AETH (syndrome + MSN)
612 * @qp: the queue pair to compute the AETH for
613 *
614 * Returns the AETH.
615 */
616__be32 rvt_compute_aeth(struct rvt_qp *qp);
617
618/**
619 * rvt_get_credit - flush the send work queue of a QP
620 * @qp: the qp who's send work queue to flush
621 * @aeth: the Acknowledge Extended Transport Header
622 *
623 * The QP s_lock should be held.
624 */
625void rvt_get_credit(struct rvt_qp *qp, u32 aeth);
626
627/**
628 * @qp - the qp pair
629 * @len - the length
630 *
631 * Perform a shift based mtu round up divide
632 */
633static inline u32 rvt_div_round_up_mtu(struct rvt_qp *qp, u32 len)
634{
635 return (len + qp->pmtu - 1) >> qp->log_pmtu;
636}
637
638/**
639 * @qp - the qp pair
640 * @len - the length
641 *
642 * Perform a shift based mtu divide
643 */
644static inline u32 rvt_div_mtu(struct rvt_qp *qp, u32 len)
645{
646 return len >> qp->log_pmtu;
647}
648
649/**
650 * rvt_timeout_to_jiffies - Convert a ULP timeout input into jiffies
651 * @timeout - timeout input(0 - 31).
652 *
653 * Return a timeout value in jiffies.
654 */
655static inline unsigned long rvt_timeout_to_jiffies(u8 timeout)
656{
657 if (timeout > 31)
658 timeout = 31;
659
660 return usecs_to_jiffies(1U << timeout) * 4096UL / 1000UL;
661}
662
663extern const int ib_rvt_state_ops[];
664
665struct rvt_dev_info;
666void rvt_comm_est(struct rvt_qp *qp);
667int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err);
668void rvt_rc_error(struct rvt_qp *qp, enum ib_wc_status err);
669unsigned long rvt_rnr_tbl_to_usec(u32 index);
670enum hrtimer_restart rvt_rc_rnr_retry(struct hrtimer *t);
671void rvt_add_rnr_timer(struct rvt_qp *qp, u32 aeth);
672void rvt_del_timers_sync(struct rvt_qp *qp);
673void rvt_stop_rc_timers(struct rvt_qp *qp);
674void rvt_add_retry_timer(struct rvt_qp *qp);
675
676/**
677 * struct rvt_qp_iter - the iterator for QPs
678 * @qp - the current QP
679 *
680 * This structure defines the current iterator
681 * state for sequenced access to all QPs relative
682 * to an rvt_dev_info.
683 */
684struct rvt_qp_iter {
685 struct rvt_qp *qp;
686 /* private: backpointer */
687 struct rvt_dev_info *rdi;
688 /* private: callback routine */
689 void (*cb)(struct rvt_qp *qp, u64 v);
690 /* private: for arg to callback routine */
691 u64 v;
692 /* private: number of SMI,GSI QPs for device */
693 int specials;
694 /* private: current iterator index */
695 int n;
696};
697
698struct rvt_qp_iter *rvt_qp_iter_init(struct rvt_dev_info *rdi,
699 u64 v,
700 void (*cb)(struct rvt_qp *qp, u64 v));
701int rvt_qp_iter_next(struct rvt_qp_iter *iter);
702void rvt_qp_iter(struct rvt_dev_info *rdi,
703 u64 v,
704 void (*cb)(struct rvt_qp *qp, u64 v));
705void rvt_qp_mr_clean(struct rvt_qp *qp, u32 lkey);
706#endif /* DEF_RDMAVT_INCQP_H */