1 #ifndef DEF_RDMA_VT_H
2 #define DEF_RDMA_VT_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 /*
53 * Structure that low level drivers will populate in order to register with the
54 * rdmavt layer.
55 */
56
57 #include <linux/spinlock.h>
58 #include <linux/list.h>
59 #include <linux/hash.h>
60 #include <rdma/ib_verbs.h>
61 #include <rdma/rdmavt_mr.h>
62 #include <rdma/rdmavt_qp.h>
63
64 #define RVT_MAX_PKEY_VALUES 16
65
66 struct rvt_ibport {
67 struct rvt_qp __rcu *qp[2];
68 struct ib_mad_agent *send_agent; /* agent for SMI (traps) */
69 struct rb_root mcast_tree;
70 spinlock_t lock; /* protect changes in this struct */
71
72 /* non-zero when timer is set */
73 unsigned long mkey_lease_timeout;
74 unsigned long trap_timeout;
75 __be64 gid_prefix; /* in network order */
76 __be64 mkey;
77 u64 tid;
78 u32 port_cap_flags;
79 u32 pma_sample_start;
80 u32 pma_sample_interval;
81 __be16 pma_counter_select[5];
82 u16 pma_tag;
83 u16 mkey_lease_period;
84 u16 sm_lid;
85 u8 sm_sl;
86 u8 mkeyprot;
87 u8 subnet_timeout;
88 u8 vl_high_limit;
89
90 /*
91 * Driver is expected to keep these up to date. These
92 * counters are informational only and not required to be
93 * completely accurate.
94 */
95 u64 n_rc_resends;
96 u64 n_seq_naks;
97 u64 n_rdma_seq;
98 u64 n_rnr_naks;
99 u64 n_other_naks;
100 u64 n_loop_pkts;
101 u64 n_pkt_drops;
102 u64 n_vl15_dropped;
103 u64 n_rc_timeouts;
104 u64 n_dmawait;
105 u64 n_unaligned;
106 u64 n_rc_dupreq;
107 u64 n_rc_seqnak;
108 u16 pkey_violations;
109 u16 qkey_violations;
110 u16 mkey_violations;
111
112 /* Hot-path per CPU counters to avoid cacheline trading to update */
113 u64 z_rc_acks;
114 u64 z_rc_qacks;
115 u64 z_rc_delayed_comp;
116 u64 __percpu *rc_acks;
117 u64 __percpu *rc_qacks;
118 u64 __percpu *rc_delayed_comp;
119
120 void *priv; /* driver private data */
121
122 /*
123 * The pkey table is allocated and maintained by the driver. Drivers
124 * need to have access to this before registering with rdmav. However
125 * rdmavt will need access to it so drivers need to proviee this during
126 * the attach port API call.
127 */
128 u16 *pkey_table;
129
130 struct rvt_ah *sm_ah;
131 };
132
133 #define RVT_CQN_MAX 16 /* maximum length of cq name */
134
135 /*
136 * Things that are driver specific, module parameters in hfi1 and qib
137 */
138 struct rvt_driver_params {
139 struct ib_device_attr props;
140
141 /*
142 * Anything driver specific that is not covered by props
143 * For instance special module parameters. Goes here.
144 */
145 unsigned int lkey_table_size;
146 unsigned int qp_table_size;
147 int qpn_start;
148 int qpn_inc;
149 int qpn_res_start;
150 int qpn_res_end;
151 int nports;
152 int npkeys;
153 char cq_name[RVT_CQN_MAX];
154 int node;
155 int psn_mask;
156 int psn_shift;
157 int psn_modify_mask;
158 u32 core_cap_flags;
159 u32 max_mad_size;
160 u8 qos_shift;
161 u8 max_rdma_atomic;
162 u8 reserved_operations;
163 };
164
165 /* Protection domain */
166 struct rvt_pd {
167 struct ib_pd ibpd;
168 int user; /* non-zero if created from user space */
169 };
170
171 /* Address handle */
172 struct rvt_ah {
173 struct ib_ah ibah;
174 struct ib_ah_attr attr;
175 atomic_t refcount;
176 u8 vl;
177 u8 log_pmtu;
178 };
179
180 struct rvt_dev_info;
181 struct rvt_swqe;
182 struct rvt_driver_provided {
183 /*
184 * Which functions are required depends on which verbs rdmavt is
185 * providing and which verbs the driver is overriding. See
186 * check_support() for details.
187 */
188
189 /* Passed to ib core registration. Callback to create syfs files */
190 int (*port_callback)(struct ib_device *, u8, struct kobject *);
191
192 /*
193 * Returns a string to represent the device for which is being
194 * registered. This is primarily used for error and debug messages on
195 * the console.
196 */
197 const char * (*get_card_name)(struct rvt_dev_info *rdi);
198
199 /*
200 * Returns a pointer to the undelying hardware's PCI device. This is
201 * used to display information as to what hardware is being referenced
202 * in an output message
203 */
204 struct pci_dev * (*get_pci_dev)(struct rvt_dev_info *rdi);
205
206 /*
207 * Allocate a private queue pair data structure for driver specific
208 * information which is opaque to rdmavt. Errors are returned via
209 * ERR_PTR(err). The driver is free to return NULL or a valid
210 * pointer.
211 */
212 void * (*qp_priv_alloc)(struct rvt_dev_info *rdi, struct rvt_qp *qp,
213 gfp_t gfp);
214
215 /*
216 * Free the driver's private qp structure.
217 */
218 void (*qp_priv_free)(struct rvt_dev_info *rdi, struct rvt_qp *qp);
219
220 /*
221 * Inform the driver the particular qp in quesiton has been reset so
222 * that it can clean up anything it needs to.
223 */
224 void (*notify_qp_reset)(struct rvt_qp *qp);
225
226 /*
227 * Give the driver a notice that there is send work to do. It is up to
228 * the driver to generally push the packets out, this just queues the
229 * work with the driver. There are two variants here. The no_lock
230 * version requires the s_lock not to be held. The other assumes the
231 * s_lock is held.
232 */
233 void (*schedule_send)(struct rvt_qp *qp);
234 void (*schedule_send_no_lock)(struct rvt_qp *qp);
235
236 /*
237 * Sometimes rdmavt needs to kick the driver's send progress. That is
238 * done by this call back.
239 */
240 void (*do_send)(struct rvt_qp *qp);
241
242 /*
243 * Get a path mtu from the driver based on qp attributes.
244 */
245 int (*get_pmtu_from_attr)(struct rvt_dev_info *rdi, struct rvt_qp *qp,
246 struct ib_qp_attr *attr);
247
248 /*
249 * Notify driver that it needs to flush any outstanding IO requests that
250 * are waiting on a qp.
251 */
252 void (*flush_qp_waiters)(struct rvt_qp *qp);
253
254 /*
255 * Notify driver to stop its queue of sending packets. Nothing else
256 * should be posted to the queue pair after this has been called.
257 */
258 void (*stop_send_queue)(struct rvt_qp *qp);
259
260 /*
261 * Have the drivr drain any in progress operations
262 */
263 void (*quiesce_qp)(struct rvt_qp *qp);
264
265 /*
266 * Inform the driver a qp has went to error state.
267 */
268 void (*notify_error_qp)(struct rvt_qp *qp);
269
270 /*
271 * Get an MTU for a qp.
272 */
273 u32 (*mtu_from_qp)(struct rvt_dev_info *rdi, struct rvt_qp *qp,
274 u32 pmtu);
275 /*
276 * Convert an mtu to a path mtu
277 */
278 int (*mtu_to_path_mtu)(u32 mtu);
279
280 /*
281 * Get the guid of a port in big endian byte order
282 */
283 int (*get_guid_be)(struct rvt_dev_info *rdi, struct rvt_ibport *rvp,
284 int guid_index, __be64 *guid);
285
286 /*
287 * Query driver for the state of the port.
288 */
289 int (*query_port_state)(struct rvt_dev_info *rdi, u8 port_num,
290 struct ib_port_attr *props);
291
292 /*
293 * Tell driver to shutdown a port
294 */
295 int (*shut_down_port)(struct rvt_dev_info *rdi, u8 port_num);
296
297 /* Tell driver to send a trap for changed port capabilities */
298 void (*cap_mask_chg)(struct rvt_dev_info *rdi, u8 port_num);
299
300 /*
301 * The following functions can be safely ignored completely. Any use of
302 * these is checked for NULL before blindly calling. Rdmavt should also
303 * be functional if drivers omit these.
304 */
305
306 /* Called to inform the driver that all qps should now be freed. */
307 unsigned (*free_all_qps)(struct rvt_dev_info *rdi);
308
309 /* Driver specific AH validation */
310 int (*check_ah)(struct ib_device *, struct ib_ah_attr *);
311
312 /* Inform the driver a new AH has been created */
313 void (*notify_new_ah)(struct ib_device *, struct ib_ah_attr *,
314 struct rvt_ah *);
315
316 /* Let the driver pick the next queue pair number*/
317 int (*alloc_qpn)(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt,
318 enum ib_qp_type type, u8 port_num, gfp_t gfp);
319
320 /* Determine if its safe or allowed to modify the qp */
321 int (*check_modify_qp)(struct rvt_qp *qp, struct ib_qp_attr *attr,
322 int attr_mask, struct ib_udata *udata);
323
324 /* Driver specific QP modification/notification-of */
325 void (*modify_qp)(struct rvt_qp *qp, struct ib_qp_attr *attr,
326 int attr_mask, struct ib_udata *udata);
327
328 /* Driver specific work request checking */
329 int (*check_send_wqe)(struct rvt_qp *qp, struct rvt_swqe *wqe);
330
331 /* Notify driver a mad agent has been created */
332 void (*notify_create_mad_agent)(struct rvt_dev_info *rdi, int port_idx);
333
334 /* Notify driver a mad agent has been removed */
335 void (*notify_free_mad_agent)(struct rvt_dev_info *rdi, int port_idx);
336
337 };
338
339 struct rvt_dev_info {
340 struct ib_device ibdev; /* Keep this first. Nothing above here */
341
342 /*
343 * Prior to calling for registration the driver will be responsible for
344 * allocating space for this structure.
345 *
346 * The driver will also be responsible for filling in certain members of
347 * dparms.props. The driver needs to fill in dparms exactly as it would
348 * want values reported to a ULP. This will be returned to the caller
349 * in rdmavt's device. The driver should also therefore refrain from
350 * modifying this directly after registration with rdmavt.
351 */
352
353 /* Driver specific properties */
354 struct rvt_driver_params dparms;
355
356 /* post send table */
357 const struct rvt_operation_params *post_parms;
358
359 struct rvt_mregion __rcu *dma_mr;
360 struct rvt_lkey_table lkey_table;
361
362 /* Driver specific helper functions */
363 struct rvt_driver_provided driver_f;
364
365 /* Internal use */
366 int n_pds_allocated;
367 spinlock_t n_pds_lock; /* Protect pd allocated count */
368
369 int n_ahs_allocated;
370 spinlock_t n_ahs_lock; /* Protect ah allocated count */
371
372 u32 n_srqs_allocated;
373 spinlock_t n_srqs_lock; /* Protect srqs allocated count */
374
375 int flags;
376 struct rvt_ibport **ports;
377
378 /* QP */
379 struct rvt_qp_ibdev *qp_dev;
380 u32 n_qps_allocated; /* number of QPs allocated for device */
381 u32 n_rc_qps; /* number of RC QPs allocated for device */
382 u32 busy_jiffies; /* timeout scaling based on RC QP count */
383 spinlock_t n_qps_lock; /* protect qps, rc qps and busy jiffy counts */
384
385 /* memory maps */
386 struct list_head pending_mmaps;
387 spinlock_t mmap_offset_lock; /* protect mmap_offset */
388 u32 mmap_offset;
389 spinlock_t pending_lock; /* protect pending mmap list */
390
391 /* CQ */
392 struct kthread_worker *worker; /* per device cq worker */
393 u32 n_cqs_allocated; /* number of CQs allocated for device */
394 spinlock_t n_cqs_lock; /* protect count of in use cqs */
395
396 /* Multicast */
397 u32 n_mcast_grps_allocated; /* number of mcast groups allocated */
398 spinlock_t n_mcast_grps_lock;
399
400 };
401
ibpd_to_rvtpd(struct ib_pd * ibpd)402 static inline struct rvt_pd *ibpd_to_rvtpd(struct ib_pd *ibpd)
403 {
404 return container_of(ibpd, struct rvt_pd, ibpd);
405 }
406
ibah_to_rvtah(struct ib_ah * ibah)407 static inline struct rvt_ah *ibah_to_rvtah(struct ib_ah *ibah)
408 {
409 return container_of(ibah, struct rvt_ah, ibah);
410 }
411
ib_to_rvt(struct ib_device * ibdev)412 static inline struct rvt_dev_info *ib_to_rvt(struct ib_device *ibdev)
413 {
414 return container_of(ibdev, struct rvt_dev_info, ibdev);
415 }
416
ibsrq_to_rvtsrq(struct ib_srq * ibsrq)417 static inline struct rvt_srq *ibsrq_to_rvtsrq(struct ib_srq *ibsrq)
418 {
419 return container_of(ibsrq, struct rvt_srq, ibsrq);
420 }
421
ibqp_to_rvtqp(struct ib_qp * ibqp)422 static inline struct rvt_qp *ibqp_to_rvtqp(struct ib_qp *ibqp)
423 {
424 return container_of(ibqp, struct rvt_qp, ibqp);
425 }
426
rvt_get_npkeys(struct rvt_dev_info * rdi)427 static inline unsigned rvt_get_npkeys(struct rvt_dev_info *rdi)
428 {
429 /*
430 * All ports have same number of pkeys.
431 */
432 return rdi->dparms.npkeys;
433 }
434
435 /*
436 * Return the max atomic suitable for determining
437 * the size of the ack ring buffer in a QP.
438 */
rvt_max_atomic(struct rvt_dev_info * rdi)439 static inline unsigned int rvt_max_atomic(struct rvt_dev_info *rdi)
440 {
441 return rdi->dparms.max_rdma_atomic + 1;
442 }
443
444 /*
445 * Return the indexed PKEY from the port PKEY table.
446 */
rvt_get_pkey(struct rvt_dev_info * rdi,int port_index,unsigned index)447 static inline u16 rvt_get_pkey(struct rvt_dev_info *rdi,
448 int port_index,
449 unsigned index)
450 {
451 if (index >= rvt_get_npkeys(rdi))
452 return 0;
453 else
454 return rdi->ports[port_index]->pkey_table[index];
455 }
456
457 /**
458 * rvt_lookup_qpn - return the QP with the given QPN
459 * @ibp: the ibport
460 * @qpn: the QP number to look up
461 *
462 * The caller must hold the rcu_read_lock(), and keep the lock until
463 * the returned qp is no longer in use.
464 */
465 /* TODO: Remove this and put in rdmavt/qp.h when no longer needed by drivers */
rvt_lookup_qpn(struct rvt_dev_info * rdi,struct rvt_ibport * rvp,u32 qpn)466 static inline struct rvt_qp *rvt_lookup_qpn(struct rvt_dev_info *rdi,
467 struct rvt_ibport *rvp,
468 u32 qpn) __must_hold(RCU)
469 {
470 struct rvt_qp *qp = NULL;
471
472 if (unlikely(qpn <= 1)) {
473 qp = rcu_dereference(rvp->qp[qpn]);
474 } else {
475 u32 n = hash_32(qpn, rdi->qp_dev->qp_table_bits);
476
477 for (qp = rcu_dereference(rdi->qp_dev->qp_table[n]); qp;
478 qp = rcu_dereference(qp->next))
479 if (qp->ibqp.qp_num == qpn)
480 break;
481 }
482 return qp;
483 }
484
485 struct rvt_dev_info *rvt_alloc_device(size_t size, int nports);
486 void rvt_dealloc_device(struct rvt_dev_info *rdi);
487 int rvt_register_device(struct rvt_dev_info *rvd);
488 void rvt_unregister_device(struct rvt_dev_info *rvd);
489 int rvt_check_ah(struct ib_device *ibdev, struct ib_ah_attr *ah_attr);
490 int rvt_init_port(struct rvt_dev_info *rdi, struct rvt_ibport *port,
491 int port_index, u16 *pkey_table);
492 int rvt_fast_reg_mr(struct rvt_qp *qp, struct ib_mr *ibmr, u32 key,
493 int access);
494 int rvt_invalidate_rkey(struct rvt_qp *qp, u32 rkey);
495 int rvt_rkey_ok(struct rvt_qp *qp, struct rvt_sge *sge,
496 u32 len, u64 vaddr, u32 rkey, int acc);
497 int rvt_lkey_ok(struct rvt_lkey_table *rkt, struct rvt_pd *pd,
498 struct rvt_sge *isge, struct ib_sge *sge, int acc);
499 struct rvt_mcast *rvt_mcast_find(struct rvt_ibport *ibp, union ib_gid *mgid);
500
501 #endif /* DEF_RDMA_VT_H */
502