Loading...
Note: File does not exist in v3.1.
1/*
2 * Copyright(c) 2016, 2017 Intel Corporation.
3 *
4 * This file is provided under a dual BSD/GPLv2 license. When using or
5 * redistributing this file, you may do so under either license.
6 *
7 * GPL LICENSE SUMMARY
8 *
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of version 2 of the GNU General Public License as
11 * published by the Free Software Foundation.
12 *
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
17 *
18 * BSD LICENSE
19 *
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
22 * are met:
23 *
24 * - Redistributions of source code must retain the above copyright
25 * notice, this list of conditions and the following disclaimer.
26 * - Redistributions in binary form must reproduce the above copyright
27 * notice, this list of conditions and the following disclaimer in
28 * the documentation and/or other materials provided with the
29 * distribution.
30 * - Neither the name of Intel Corporation nor the names of its
31 * contributors may be used to endorse or promote products derived
32 * from this software without specific prior written permission.
33 *
34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45 *
46 */
47
48#include <linux/hash.h>
49#include <linux/bitops.h>
50#include <linux/lockdep.h>
51#include <linux/vmalloc.h>
52#include <linux/slab.h>
53#include <rdma/ib_verbs.h>
54#include <rdma/ib_hdrs.h>
55#include <rdma/opa_addr.h>
56#include "qp.h"
57#include "vt.h"
58#include "trace.h"
59
60static void rvt_rc_timeout(struct timer_list *t);
61
62/*
63 * Convert the AETH RNR timeout code into the number of microseconds.
64 */
65static const u32 ib_rvt_rnr_table[32] = {
66 655360, /* 00: 655.36 */
67 10, /* 01: .01 */
68 20, /* 02 .02 */
69 30, /* 03: .03 */
70 40, /* 04: .04 */
71 60, /* 05: .06 */
72 80, /* 06: .08 */
73 120, /* 07: .12 */
74 160, /* 08: .16 */
75 240, /* 09: .24 */
76 320, /* 0A: .32 */
77 480, /* 0B: .48 */
78 640, /* 0C: .64 */
79 960, /* 0D: .96 */
80 1280, /* 0E: 1.28 */
81 1920, /* 0F: 1.92 */
82 2560, /* 10: 2.56 */
83 3840, /* 11: 3.84 */
84 5120, /* 12: 5.12 */
85 7680, /* 13: 7.68 */
86 10240, /* 14: 10.24 */
87 15360, /* 15: 15.36 */
88 20480, /* 16: 20.48 */
89 30720, /* 17: 30.72 */
90 40960, /* 18: 40.96 */
91 61440, /* 19: 61.44 */
92 81920, /* 1A: 81.92 */
93 122880, /* 1B: 122.88 */
94 163840, /* 1C: 163.84 */
95 245760, /* 1D: 245.76 */
96 327680, /* 1E: 327.68 */
97 491520 /* 1F: 491.52 */
98};
99
100/*
101 * Note that it is OK to post send work requests in the SQE and ERR
102 * states; rvt_do_send() will process them and generate error
103 * completions as per IB 1.2 C10-96.
104 */
105const int ib_rvt_state_ops[IB_QPS_ERR + 1] = {
106 [IB_QPS_RESET] = 0,
107 [IB_QPS_INIT] = RVT_POST_RECV_OK,
108 [IB_QPS_RTR] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK,
109 [IB_QPS_RTS] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
110 RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK |
111 RVT_PROCESS_NEXT_SEND_OK,
112 [IB_QPS_SQD] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
113 RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK,
114 [IB_QPS_SQE] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK |
115 RVT_POST_SEND_OK | RVT_FLUSH_SEND,
116 [IB_QPS_ERR] = RVT_POST_RECV_OK | RVT_FLUSH_RECV |
117 RVT_POST_SEND_OK | RVT_FLUSH_SEND,
118};
119EXPORT_SYMBOL(ib_rvt_state_ops);
120
121static void get_map_page(struct rvt_qpn_table *qpt,
122 struct rvt_qpn_map *map)
123{
124 unsigned long page = get_zeroed_page(GFP_KERNEL);
125
126 /*
127 * Free the page if someone raced with us installing it.
128 */
129
130 spin_lock(&qpt->lock);
131 if (map->page)
132 free_page(page);
133 else
134 map->page = (void *)page;
135 spin_unlock(&qpt->lock);
136}
137
138/**
139 * init_qpn_table - initialize the QP number table for a device
140 * @qpt: the QPN table
141 */
142static int init_qpn_table(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt)
143{
144 u32 offset, i;
145 struct rvt_qpn_map *map;
146 int ret = 0;
147
148 if (!(rdi->dparms.qpn_res_end >= rdi->dparms.qpn_res_start))
149 return -EINVAL;
150
151 spin_lock_init(&qpt->lock);
152
153 qpt->last = rdi->dparms.qpn_start;
154 qpt->incr = rdi->dparms.qpn_inc << rdi->dparms.qos_shift;
155
156 /*
157 * Drivers may want some QPs beyond what we need for verbs let them use
158 * our qpn table. No need for two. Lets go ahead and mark the bitmaps
159 * for those. The reserved range must be *after* the range which verbs
160 * will pick from.
161 */
162
163 /* Figure out number of bit maps needed before reserved range */
164 qpt->nmaps = rdi->dparms.qpn_res_start / RVT_BITS_PER_PAGE;
165
166 /* This should always be zero */
167 offset = rdi->dparms.qpn_res_start & RVT_BITS_PER_PAGE_MASK;
168
169 /* Starting with the first reserved bit map */
170 map = &qpt->map[qpt->nmaps];
171
172 rvt_pr_info(rdi, "Reserving QPNs from 0x%x to 0x%x for non-verbs use\n",
173 rdi->dparms.qpn_res_start, rdi->dparms.qpn_res_end);
174 for (i = rdi->dparms.qpn_res_start; i <= rdi->dparms.qpn_res_end; i++) {
175 if (!map->page) {
176 get_map_page(qpt, map);
177 if (!map->page) {
178 ret = -ENOMEM;
179 break;
180 }
181 }
182 set_bit(offset, map->page);
183 offset++;
184 if (offset == RVT_BITS_PER_PAGE) {
185 /* next page */
186 qpt->nmaps++;
187 map++;
188 offset = 0;
189 }
190 }
191 return ret;
192}
193
194/**
195 * free_qpn_table - free the QP number table for a device
196 * @qpt: the QPN table
197 */
198static void free_qpn_table(struct rvt_qpn_table *qpt)
199{
200 int i;
201
202 for (i = 0; i < ARRAY_SIZE(qpt->map); i++)
203 free_page((unsigned long)qpt->map[i].page);
204}
205
206/**
207 * rvt_driver_qp_init - Init driver qp resources
208 * @rdi: rvt dev strucutre
209 *
210 * Return: 0 on success
211 */
212int rvt_driver_qp_init(struct rvt_dev_info *rdi)
213{
214 int i;
215 int ret = -ENOMEM;
216
217 if (!rdi->dparms.qp_table_size)
218 return -EINVAL;
219
220 /*
221 * If driver is not doing any QP allocation then make sure it is
222 * providing the necessary QP functions.
223 */
224 if (!rdi->driver_f.free_all_qps ||
225 !rdi->driver_f.qp_priv_alloc ||
226 !rdi->driver_f.qp_priv_free ||
227 !rdi->driver_f.notify_qp_reset ||
228 !rdi->driver_f.notify_restart_rc)
229 return -EINVAL;
230
231 /* allocate parent object */
232 rdi->qp_dev = kzalloc_node(sizeof(*rdi->qp_dev), GFP_KERNEL,
233 rdi->dparms.node);
234 if (!rdi->qp_dev)
235 return -ENOMEM;
236
237 /* allocate hash table */
238 rdi->qp_dev->qp_table_size = rdi->dparms.qp_table_size;
239 rdi->qp_dev->qp_table_bits = ilog2(rdi->dparms.qp_table_size);
240 rdi->qp_dev->qp_table =
241 kmalloc_array_node(rdi->qp_dev->qp_table_size,
242 sizeof(*rdi->qp_dev->qp_table),
243 GFP_KERNEL, rdi->dparms.node);
244 if (!rdi->qp_dev->qp_table)
245 goto no_qp_table;
246
247 for (i = 0; i < rdi->qp_dev->qp_table_size; i++)
248 RCU_INIT_POINTER(rdi->qp_dev->qp_table[i], NULL);
249
250 spin_lock_init(&rdi->qp_dev->qpt_lock);
251
252 /* initialize qpn map */
253 if (init_qpn_table(rdi, &rdi->qp_dev->qpn_table))
254 goto fail_table;
255
256 spin_lock_init(&rdi->n_qps_lock);
257
258 return 0;
259
260fail_table:
261 kfree(rdi->qp_dev->qp_table);
262 free_qpn_table(&rdi->qp_dev->qpn_table);
263
264no_qp_table:
265 kfree(rdi->qp_dev);
266
267 return ret;
268}
269
270/**
271 * free_all_qps - check for QPs still in use
272 * @rdi: rvt device info structure
273 *
274 * There should not be any QPs still in use.
275 * Free memory for table.
276 */
277static unsigned rvt_free_all_qps(struct rvt_dev_info *rdi)
278{
279 unsigned long flags;
280 struct rvt_qp *qp;
281 unsigned n, qp_inuse = 0;
282 spinlock_t *ql; /* work around too long line below */
283
284 if (rdi->driver_f.free_all_qps)
285 qp_inuse = rdi->driver_f.free_all_qps(rdi);
286
287 qp_inuse += rvt_mcast_tree_empty(rdi);
288
289 if (!rdi->qp_dev)
290 return qp_inuse;
291
292 ql = &rdi->qp_dev->qpt_lock;
293 spin_lock_irqsave(ql, flags);
294 for (n = 0; n < rdi->qp_dev->qp_table_size; n++) {
295 qp = rcu_dereference_protected(rdi->qp_dev->qp_table[n],
296 lockdep_is_held(ql));
297 RCU_INIT_POINTER(rdi->qp_dev->qp_table[n], NULL);
298
299 for (; qp; qp = rcu_dereference_protected(qp->next,
300 lockdep_is_held(ql)))
301 qp_inuse++;
302 }
303 spin_unlock_irqrestore(ql, flags);
304 synchronize_rcu();
305 return qp_inuse;
306}
307
308/**
309 * rvt_qp_exit - clean up qps on device exit
310 * @rdi: rvt dev structure
311 *
312 * Check for qp leaks and free resources.
313 */
314void rvt_qp_exit(struct rvt_dev_info *rdi)
315{
316 u32 qps_inuse = rvt_free_all_qps(rdi);
317
318 if (qps_inuse)
319 rvt_pr_err(rdi, "QP memory leak! %u still in use\n",
320 qps_inuse);
321 if (!rdi->qp_dev)
322 return;
323
324 kfree(rdi->qp_dev->qp_table);
325 free_qpn_table(&rdi->qp_dev->qpn_table);
326 kfree(rdi->qp_dev);
327}
328
329static inline unsigned mk_qpn(struct rvt_qpn_table *qpt,
330 struct rvt_qpn_map *map, unsigned off)
331{
332 return (map - qpt->map) * RVT_BITS_PER_PAGE + off;
333}
334
335/**
336 * alloc_qpn - Allocate the next available qpn or zero/one for QP type
337 * IB_QPT_SMI/IB_QPT_GSI
338 * @rdi: rvt device info structure
339 * @qpt: queue pair number table pointer
340 * @port_num: IB port number, 1 based, comes from core
341 *
342 * Return: The queue pair number
343 */
344static int alloc_qpn(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt,
345 enum ib_qp_type type, u8 port_num)
346{
347 u32 i, offset, max_scan, qpn;
348 struct rvt_qpn_map *map;
349 u32 ret;
350
351 if (rdi->driver_f.alloc_qpn)
352 return rdi->driver_f.alloc_qpn(rdi, qpt, type, port_num);
353
354 if (type == IB_QPT_SMI || type == IB_QPT_GSI) {
355 unsigned n;
356
357 ret = type == IB_QPT_GSI;
358 n = 1 << (ret + 2 * (port_num - 1));
359 spin_lock(&qpt->lock);
360 if (qpt->flags & n)
361 ret = -EINVAL;
362 else
363 qpt->flags |= n;
364 spin_unlock(&qpt->lock);
365 goto bail;
366 }
367
368 qpn = qpt->last + qpt->incr;
369 if (qpn >= RVT_QPN_MAX)
370 qpn = qpt->incr | ((qpt->last & 1) ^ 1);
371 /* offset carries bit 0 */
372 offset = qpn & RVT_BITS_PER_PAGE_MASK;
373 map = &qpt->map[qpn / RVT_BITS_PER_PAGE];
374 max_scan = qpt->nmaps - !offset;
375 for (i = 0;;) {
376 if (unlikely(!map->page)) {
377 get_map_page(qpt, map);
378 if (unlikely(!map->page))
379 break;
380 }
381 do {
382 if (!test_and_set_bit(offset, map->page)) {
383 qpt->last = qpn;
384 ret = qpn;
385 goto bail;
386 }
387 offset += qpt->incr;
388 /*
389 * This qpn might be bogus if offset >= BITS_PER_PAGE.
390 * That is OK. It gets re-assigned below
391 */
392 qpn = mk_qpn(qpt, map, offset);
393 } while (offset < RVT_BITS_PER_PAGE && qpn < RVT_QPN_MAX);
394 /*
395 * In order to keep the number of pages allocated to a
396 * minimum, we scan the all existing pages before increasing
397 * the size of the bitmap table.
398 */
399 if (++i > max_scan) {
400 if (qpt->nmaps == RVT_QPNMAP_ENTRIES)
401 break;
402 map = &qpt->map[qpt->nmaps++];
403 /* start at incr with current bit 0 */
404 offset = qpt->incr | (offset & 1);
405 } else if (map < &qpt->map[qpt->nmaps]) {
406 ++map;
407 /* start at incr with current bit 0 */
408 offset = qpt->incr | (offset & 1);
409 } else {
410 map = &qpt->map[0];
411 /* wrap to first map page, invert bit 0 */
412 offset = qpt->incr | ((offset & 1) ^ 1);
413 }
414 /* there can be no set bits in low-order QoS bits */
415 WARN_ON(offset & (BIT(rdi->dparms.qos_shift) - 1));
416 qpn = mk_qpn(qpt, map, offset);
417 }
418
419 ret = -ENOMEM;
420
421bail:
422 return ret;
423}
424
425/**
426 * rvt_clear_mr_refs - Drop help mr refs
427 * @qp: rvt qp data structure
428 * @clr_sends: If shoudl clear send side or not
429 */
430static void rvt_clear_mr_refs(struct rvt_qp *qp, int clr_sends)
431{
432 unsigned n;
433 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
434
435 if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags))
436 rvt_put_ss(&qp->s_rdma_read_sge);
437
438 rvt_put_ss(&qp->r_sge);
439
440 if (clr_sends) {
441 while (qp->s_last != qp->s_head) {
442 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_last);
443
444 rvt_put_swqe(wqe);
445
446 if (qp->ibqp.qp_type == IB_QPT_UD ||
447 qp->ibqp.qp_type == IB_QPT_SMI ||
448 qp->ibqp.qp_type == IB_QPT_GSI)
449 atomic_dec(&ibah_to_rvtah(
450 wqe->ud_wr.ah)->refcount);
451 if (++qp->s_last >= qp->s_size)
452 qp->s_last = 0;
453 smp_wmb(); /* see qp_set_savail */
454 }
455 if (qp->s_rdma_mr) {
456 rvt_put_mr(qp->s_rdma_mr);
457 qp->s_rdma_mr = NULL;
458 }
459 }
460
461 for (n = 0; qp->s_ack_queue && n < rvt_max_atomic(rdi); n++) {
462 struct rvt_ack_entry *e = &qp->s_ack_queue[n];
463
464 if (e->rdma_sge.mr) {
465 rvt_put_mr(e->rdma_sge.mr);
466 e->rdma_sge.mr = NULL;
467 }
468 }
469}
470
471/**
472 * rvt_swqe_has_lkey - return true if lkey is used by swqe
473 * @wqe - the send wqe
474 * @lkey - the lkey
475 *
476 * Test the swqe for using lkey
477 */
478static bool rvt_swqe_has_lkey(struct rvt_swqe *wqe, u32 lkey)
479{
480 int i;
481
482 for (i = 0; i < wqe->wr.num_sge; i++) {
483 struct rvt_sge *sge = &wqe->sg_list[i];
484
485 if (rvt_mr_has_lkey(sge->mr, lkey))
486 return true;
487 }
488 return false;
489}
490
491/**
492 * rvt_qp_sends_has_lkey - return true is qp sends use lkey
493 * @qp - the rvt_qp
494 * @lkey - the lkey
495 */
496static bool rvt_qp_sends_has_lkey(struct rvt_qp *qp, u32 lkey)
497{
498 u32 s_last = qp->s_last;
499
500 while (s_last != qp->s_head) {
501 struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, s_last);
502
503 if (rvt_swqe_has_lkey(wqe, lkey))
504 return true;
505
506 if (++s_last >= qp->s_size)
507 s_last = 0;
508 }
509 if (qp->s_rdma_mr)
510 if (rvt_mr_has_lkey(qp->s_rdma_mr, lkey))
511 return true;
512 return false;
513}
514
515/**
516 * rvt_qp_acks_has_lkey - return true if acks have lkey
517 * @qp - the qp
518 * @lkey - the lkey
519 */
520static bool rvt_qp_acks_has_lkey(struct rvt_qp *qp, u32 lkey)
521{
522 int i;
523 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
524
525 for (i = 0; qp->s_ack_queue && i < rvt_max_atomic(rdi); i++) {
526 struct rvt_ack_entry *e = &qp->s_ack_queue[i];
527
528 if (rvt_mr_has_lkey(e->rdma_sge.mr, lkey))
529 return true;
530 }
531 return false;
532}
533
534/*
535 * rvt_qp_mr_clean - clean up remote ops for lkey
536 * @qp - the qp
537 * @lkey - the lkey that is being de-registered
538 *
539 * This routine checks if the lkey is being used by
540 * the qp.
541 *
542 * If so, the qp is put into an error state to elminate
543 * any references from the qp.
544 */
545void rvt_qp_mr_clean(struct rvt_qp *qp, u32 lkey)
546{
547 bool lastwqe = false;
548
549 if (qp->ibqp.qp_type == IB_QPT_SMI ||
550 qp->ibqp.qp_type == IB_QPT_GSI)
551 /* avoid special QPs */
552 return;
553 spin_lock_irq(&qp->r_lock);
554 spin_lock(&qp->s_hlock);
555 spin_lock(&qp->s_lock);
556
557 if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET)
558 goto check_lwqe;
559
560 if (rvt_ss_has_lkey(&qp->r_sge, lkey) ||
561 rvt_qp_sends_has_lkey(qp, lkey) ||
562 rvt_qp_acks_has_lkey(qp, lkey))
563 lastwqe = rvt_error_qp(qp, IB_WC_LOC_PROT_ERR);
564check_lwqe:
565 spin_unlock(&qp->s_lock);
566 spin_unlock(&qp->s_hlock);
567 spin_unlock_irq(&qp->r_lock);
568 if (lastwqe) {
569 struct ib_event ev;
570
571 ev.device = qp->ibqp.device;
572 ev.element.qp = &qp->ibqp;
573 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
574 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
575 }
576}
577
578/**
579 * rvt_remove_qp - remove qp form table
580 * @rdi: rvt dev struct
581 * @qp: qp to remove
582 *
583 * Remove the QP from the table so it can't be found asynchronously by
584 * the receive routine.
585 */
586static void rvt_remove_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp)
587{
588 struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1];
589 u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits);
590 unsigned long flags;
591 int removed = 1;
592
593 spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags);
594
595 if (rcu_dereference_protected(rvp->qp[0],
596 lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) {
597 RCU_INIT_POINTER(rvp->qp[0], NULL);
598 } else if (rcu_dereference_protected(rvp->qp[1],
599 lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) {
600 RCU_INIT_POINTER(rvp->qp[1], NULL);
601 } else {
602 struct rvt_qp *q;
603 struct rvt_qp __rcu **qpp;
604
605 removed = 0;
606 qpp = &rdi->qp_dev->qp_table[n];
607 for (; (q = rcu_dereference_protected(*qpp,
608 lockdep_is_held(&rdi->qp_dev->qpt_lock))) != NULL;
609 qpp = &q->next) {
610 if (q == qp) {
611 RCU_INIT_POINTER(*qpp,
612 rcu_dereference_protected(qp->next,
613 lockdep_is_held(&rdi->qp_dev->qpt_lock)));
614 removed = 1;
615 trace_rvt_qpremove(qp, n);
616 break;
617 }
618 }
619 }
620
621 spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags);
622 if (removed) {
623 synchronize_rcu();
624 rvt_put_qp(qp);
625 }
626}
627
628/**
629 * rvt_init_qp - initialize the QP state to the reset state
630 * @qp: the QP to init or reinit
631 * @type: the QP type
632 *
633 * This function is called from both rvt_create_qp() and
634 * rvt_reset_qp(). The difference is that the reset
635 * patch the necessary locks to protect against concurent
636 * access.
637 */
638static void rvt_init_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp,
639 enum ib_qp_type type)
640{
641 qp->remote_qpn = 0;
642 qp->qkey = 0;
643 qp->qp_access_flags = 0;
644 qp->s_flags &= RVT_S_SIGNAL_REQ_WR;
645 qp->s_hdrwords = 0;
646 qp->s_wqe = NULL;
647 qp->s_draining = 0;
648 qp->s_next_psn = 0;
649 qp->s_last_psn = 0;
650 qp->s_sending_psn = 0;
651 qp->s_sending_hpsn = 0;
652 qp->s_psn = 0;
653 qp->r_psn = 0;
654 qp->r_msn = 0;
655 if (type == IB_QPT_RC) {
656 qp->s_state = IB_OPCODE_RC_SEND_LAST;
657 qp->r_state = IB_OPCODE_RC_SEND_LAST;
658 } else {
659 qp->s_state = IB_OPCODE_UC_SEND_LAST;
660 qp->r_state = IB_OPCODE_UC_SEND_LAST;
661 }
662 qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE;
663 qp->r_nak_state = 0;
664 qp->r_aflags = 0;
665 qp->r_flags = 0;
666 qp->s_head = 0;
667 qp->s_tail = 0;
668 qp->s_cur = 0;
669 qp->s_acked = 0;
670 qp->s_last = 0;
671 qp->s_ssn = 1;
672 qp->s_lsn = 0;
673 qp->s_mig_state = IB_MIG_MIGRATED;
674 qp->r_head_ack_queue = 0;
675 qp->s_tail_ack_queue = 0;
676 qp->s_num_rd_atomic = 0;
677 if (qp->r_rq.wq) {
678 qp->r_rq.wq->head = 0;
679 qp->r_rq.wq->tail = 0;
680 }
681 qp->r_sge.num_sge = 0;
682 atomic_set(&qp->s_reserved_used, 0);
683}
684
685/**
686 * rvt_reset_qp - initialize the QP state to the reset state
687 * @qp: the QP to reset
688 * @type: the QP type
689 *
690 * r_lock, s_hlock, and s_lock are required to be held by the caller
691 */
692static void rvt_reset_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp,
693 enum ib_qp_type type)
694 __must_hold(&qp->s_lock)
695 __must_hold(&qp->s_hlock)
696 __must_hold(&qp->r_lock)
697{
698 lockdep_assert_held(&qp->r_lock);
699 lockdep_assert_held(&qp->s_hlock);
700 lockdep_assert_held(&qp->s_lock);
701 if (qp->state != IB_QPS_RESET) {
702 qp->state = IB_QPS_RESET;
703
704 /* Let drivers flush their waitlist */
705 rdi->driver_f.flush_qp_waiters(qp);
706 rvt_stop_rc_timers(qp);
707 qp->s_flags &= ~(RVT_S_TIMER | RVT_S_ANY_WAIT);
708 spin_unlock(&qp->s_lock);
709 spin_unlock(&qp->s_hlock);
710 spin_unlock_irq(&qp->r_lock);
711
712 /* Stop the send queue and the retry timer */
713 rdi->driver_f.stop_send_queue(qp);
714 rvt_del_timers_sync(qp);
715 /* Wait for things to stop */
716 rdi->driver_f.quiesce_qp(qp);
717
718 /* take qp out the hash and wait for it to be unused */
719 rvt_remove_qp(rdi, qp);
720
721 /* grab the lock b/c it was locked at call time */
722 spin_lock_irq(&qp->r_lock);
723 spin_lock(&qp->s_hlock);
724 spin_lock(&qp->s_lock);
725
726 rvt_clear_mr_refs(qp, 1);
727 /*
728 * Let the driver do any tear down or re-init it needs to for
729 * a qp that has been reset
730 */
731 rdi->driver_f.notify_qp_reset(qp);
732 }
733 rvt_init_qp(rdi, qp, type);
734 lockdep_assert_held(&qp->r_lock);
735 lockdep_assert_held(&qp->s_hlock);
736 lockdep_assert_held(&qp->s_lock);
737}
738
739/** rvt_free_qpn - Free a qpn from the bit map
740 * @qpt: QP table
741 * @qpn: queue pair number to free
742 */
743static void rvt_free_qpn(struct rvt_qpn_table *qpt, u32 qpn)
744{
745 struct rvt_qpn_map *map;
746
747 map = qpt->map + (qpn & RVT_QPN_MASK) / RVT_BITS_PER_PAGE;
748 if (map->page)
749 clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page);
750}
751
752/**
753 * rvt_create_qp - create a queue pair for a device
754 * @ibpd: the protection domain who's device we create the queue pair for
755 * @init_attr: the attributes of the queue pair
756 * @udata: user data for libibverbs.so
757 *
758 * Queue pair creation is mostly an rvt issue. However, drivers have their own
759 * unique idea of what queue pair numbers mean. For instance there is a reserved
760 * range for PSM.
761 *
762 * Return: the queue pair on success, otherwise returns an errno.
763 *
764 * Called by the ib_create_qp() core verbs function.
765 */
766struct ib_qp *rvt_create_qp(struct ib_pd *ibpd,
767 struct ib_qp_init_attr *init_attr,
768 struct ib_udata *udata)
769{
770 struct rvt_qp *qp;
771 int err;
772 struct rvt_swqe *swq = NULL;
773 size_t sz;
774 size_t sg_list_sz;
775 struct ib_qp *ret = ERR_PTR(-ENOMEM);
776 struct rvt_dev_info *rdi = ib_to_rvt(ibpd->device);
777 void *priv = NULL;
778 size_t sqsize;
779
780 if (!rdi)
781 return ERR_PTR(-EINVAL);
782
783 if (init_attr->cap.max_send_sge > rdi->dparms.props.max_sge ||
784 init_attr->cap.max_send_wr > rdi->dparms.props.max_qp_wr ||
785 init_attr->create_flags)
786 return ERR_PTR(-EINVAL);
787
788 /* Check receive queue parameters if no SRQ is specified. */
789 if (!init_attr->srq) {
790 if (init_attr->cap.max_recv_sge > rdi->dparms.props.max_sge ||
791 init_attr->cap.max_recv_wr > rdi->dparms.props.max_qp_wr)
792 return ERR_PTR(-EINVAL);
793
794 if (init_attr->cap.max_send_sge +
795 init_attr->cap.max_send_wr +
796 init_attr->cap.max_recv_sge +
797 init_attr->cap.max_recv_wr == 0)
798 return ERR_PTR(-EINVAL);
799 }
800 sqsize =
801 init_attr->cap.max_send_wr + 1 +
802 rdi->dparms.reserved_operations;
803 switch (init_attr->qp_type) {
804 case IB_QPT_SMI:
805 case IB_QPT_GSI:
806 if (init_attr->port_num == 0 ||
807 init_attr->port_num > ibpd->device->phys_port_cnt)
808 return ERR_PTR(-EINVAL);
809 /* fall through */
810 case IB_QPT_UC:
811 case IB_QPT_RC:
812 case IB_QPT_UD:
813 sz = sizeof(struct rvt_sge) *
814 init_attr->cap.max_send_sge +
815 sizeof(struct rvt_swqe);
816 swq = vzalloc_node(sqsize * sz, rdi->dparms.node);
817 if (!swq)
818 return ERR_PTR(-ENOMEM);
819
820 sz = sizeof(*qp);
821 sg_list_sz = 0;
822 if (init_attr->srq) {
823 struct rvt_srq *srq = ibsrq_to_rvtsrq(init_attr->srq);
824
825 if (srq->rq.max_sge > 1)
826 sg_list_sz = sizeof(*qp->r_sg_list) *
827 (srq->rq.max_sge - 1);
828 } else if (init_attr->cap.max_recv_sge > 1)
829 sg_list_sz = sizeof(*qp->r_sg_list) *
830 (init_attr->cap.max_recv_sge - 1);
831 qp = kzalloc_node(sz + sg_list_sz, GFP_KERNEL,
832 rdi->dparms.node);
833 if (!qp)
834 goto bail_swq;
835
836 RCU_INIT_POINTER(qp->next, NULL);
837 if (init_attr->qp_type == IB_QPT_RC) {
838 qp->s_ack_queue =
839 kzalloc_node(
840 sizeof(*qp->s_ack_queue) *
841 rvt_max_atomic(rdi),
842 GFP_KERNEL,
843 rdi->dparms.node);
844 if (!qp->s_ack_queue)
845 goto bail_qp;
846 }
847 /* initialize timers needed for rc qp */
848 timer_setup(&qp->s_timer, rvt_rc_timeout, 0);
849 hrtimer_init(&qp->s_rnr_timer, CLOCK_MONOTONIC,
850 HRTIMER_MODE_REL);
851 qp->s_rnr_timer.function = rvt_rc_rnr_retry;
852
853 /*
854 * Driver needs to set up it's private QP structure and do any
855 * initialization that is needed.
856 */
857 priv = rdi->driver_f.qp_priv_alloc(rdi, qp);
858 if (IS_ERR(priv)) {
859 ret = priv;
860 goto bail_qp;
861 }
862 qp->priv = priv;
863 qp->timeout_jiffies =
864 usecs_to_jiffies((4096UL * (1UL << qp->timeout)) /
865 1000UL);
866 if (init_attr->srq) {
867 sz = 0;
868 } else {
869 qp->r_rq.size = init_attr->cap.max_recv_wr + 1;
870 qp->r_rq.max_sge = init_attr->cap.max_recv_sge;
871 sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) +
872 sizeof(struct rvt_rwqe);
873 if (udata)
874 qp->r_rq.wq = vmalloc_user(
875 sizeof(struct rvt_rwq) +
876 qp->r_rq.size * sz);
877 else
878 qp->r_rq.wq = vzalloc_node(
879 sizeof(struct rvt_rwq) +
880 qp->r_rq.size * sz,
881 rdi->dparms.node);
882 if (!qp->r_rq.wq)
883 goto bail_driver_priv;
884 }
885
886 /*
887 * ib_create_qp() will initialize qp->ibqp
888 * except for qp->ibqp.qp_num.
889 */
890 spin_lock_init(&qp->r_lock);
891 spin_lock_init(&qp->s_hlock);
892 spin_lock_init(&qp->s_lock);
893 spin_lock_init(&qp->r_rq.lock);
894 atomic_set(&qp->refcount, 0);
895 atomic_set(&qp->local_ops_pending, 0);
896 init_waitqueue_head(&qp->wait);
897 INIT_LIST_HEAD(&qp->rspwait);
898 qp->state = IB_QPS_RESET;
899 qp->s_wq = swq;
900 qp->s_size = sqsize;
901 qp->s_avail = init_attr->cap.max_send_wr;
902 qp->s_max_sge = init_attr->cap.max_send_sge;
903 if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR)
904 qp->s_flags = RVT_S_SIGNAL_REQ_WR;
905
906 err = alloc_qpn(rdi, &rdi->qp_dev->qpn_table,
907 init_attr->qp_type,
908 init_attr->port_num);
909 if (err < 0) {
910 ret = ERR_PTR(err);
911 goto bail_rq_wq;
912 }
913 qp->ibqp.qp_num = err;
914 qp->port_num = init_attr->port_num;
915 rvt_init_qp(rdi, qp, init_attr->qp_type);
916 break;
917
918 default:
919 /* Don't support raw QPs */
920 return ERR_PTR(-EINVAL);
921 }
922
923 init_attr->cap.max_inline_data = 0;
924
925 /*
926 * Return the address of the RWQ as the offset to mmap.
927 * See rvt_mmap() for details.
928 */
929 if (udata && udata->outlen >= sizeof(__u64)) {
930 if (!qp->r_rq.wq) {
931 __u64 offset = 0;
932
933 err = ib_copy_to_udata(udata, &offset,
934 sizeof(offset));
935 if (err) {
936 ret = ERR_PTR(err);
937 goto bail_qpn;
938 }
939 } else {
940 u32 s = sizeof(struct rvt_rwq) + qp->r_rq.size * sz;
941
942 qp->ip = rvt_create_mmap_info(rdi, s,
943 ibpd->uobject->context,
944 qp->r_rq.wq);
945 if (!qp->ip) {
946 ret = ERR_PTR(-ENOMEM);
947 goto bail_qpn;
948 }
949
950 err = ib_copy_to_udata(udata, &qp->ip->offset,
951 sizeof(qp->ip->offset));
952 if (err) {
953 ret = ERR_PTR(err);
954 goto bail_ip;
955 }
956 }
957 qp->pid = current->pid;
958 }
959
960 spin_lock(&rdi->n_qps_lock);
961 if (rdi->n_qps_allocated == rdi->dparms.props.max_qp) {
962 spin_unlock(&rdi->n_qps_lock);
963 ret = ERR_PTR(-ENOMEM);
964 goto bail_ip;
965 }
966
967 rdi->n_qps_allocated++;
968 /*
969 * Maintain a busy_jiffies variable that will be added to the timeout
970 * period in mod_retry_timer and add_retry_timer. This busy jiffies
971 * is scaled by the number of rc qps created for the device to reduce
972 * the number of timeouts occurring when there is a large number of
973 * qps. busy_jiffies is incremented every rc qp scaling interval.
974 * The scaling interval is selected based on extensive performance
975 * evaluation of targeted workloads.
976 */
977 if (init_attr->qp_type == IB_QPT_RC) {
978 rdi->n_rc_qps++;
979 rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL;
980 }
981 spin_unlock(&rdi->n_qps_lock);
982
983 if (qp->ip) {
984 spin_lock_irq(&rdi->pending_lock);
985 list_add(&qp->ip->pending_mmaps, &rdi->pending_mmaps);
986 spin_unlock_irq(&rdi->pending_lock);
987 }
988
989 ret = &qp->ibqp;
990
991 /*
992 * We have our QP and its good, now keep track of what types of opcodes
993 * can be processed on this QP. We do this by keeping track of what the
994 * 3 high order bits of the opcode are.
995 */
996 switch (init_attr->qp_type) {
997 case IB_QPT_SMI:
998 case IB_QPT_GSI:
999 case IB_QPT_UD:
1000 qp->allowed_ops = IB_OPCODE_UD;
1001 break;
1002 case IB_QPT_RC:
1003 qp->allowed_ops = IB_OPCODE_RC;
1004 break;
1005 case IB_QPT_UC:
1006 qp->allowed_ops = IB_OPCODE_UC;
1007 break;
1008 default:
1009 ret = ERR_PTR(-EINVAL);
1010 goto bail_ip;
1011 }
1012
1013 return ret;
1014
1015bail_ip:
1016 if (qp->ip)
1017 kref_put(&qp->ip->ref, rvt_release_mmap_info);
1018
1019bail_qpn:
1020 rvt_free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num);
1021
1022bail_rq_wq:
1023 if (!qp->ip)
1024 vfree(qp->r_rq.wq);
1025
1026bail_driver_priv:
1027 rdi->driver_f.qp_priv_free(rdi, qp);
1028
1029bail_qp:
1030 kfree(qp->s_ack_queue);
1031 kfree(qp);
1032
1033bail_swq:
1034 vfree(swq);
1035
1036 return ret;
1037}
1038
1039/**
1040 * rvt_error_qp - put a QP into the error state
1041 * @qp: the QP to put into the error state
1042 * @err: the receive completion error to signal if a RWQE is active
1043 *
1044 * Flushes both send and receive work queues.
1045 *
1046 * Return: true if last WQE event should be generated.
1047 * The QP r_lock and s_lock should be held and interrupts disabled.
1048 * If we are already in error state, just return.
1049 */
1050int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err)
1051{
1052 struct ib_wc wc;
1053 int ret = 0;
1054 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
1055
1056 lockdep_assert_held(&qp->r_lock);
1057 lockdep_assert_held(&qp->s_lock);
1058 if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET)
1059 goto bail;
1060
1061 qp->state = IB_QPS_ERR;
1062
1063 if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) {
1064 qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR);
1065 del_timer(&qp->s_timer);
1066 }
1067
1068 if (qp->s_flags & RVT_S_ANY_WAIT_SEND)
1069 qp->s_flags &= ~RVT_S_ANY_WAIT_SEND;
1070
1071 rdi->driver_f.notify_error_qp(qp);
1072
1073 /* Schedule the sending tasklet to drain the send work queue. */
1074 if (READ_ONCE(qp->s_last) != qp->s_head)
1075 rdi->driver_f.schedule_send(qp);
1076
1077 rvt_clear_mr_refs(qp, 0);
1078
1079 memset(&wc, 0, sizeof(wc));
1080 wc.qp = &qp->ibqp;
1081 wc.opcode = IB_WC_RECV;
1082
1083 if (test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) {
1084 wc.wr_id = qp->r_wr_id;
1085 wc.status = err;
1086 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
1087 }
1088 wc.status = IB_WC_WR_FLUSH_ERR;
1089
1090 if (qp->r_rq.wq) {
1091 struct rvt_rwq *wq;
1092 u32 head;
1093 u32 tail;
1094
1095 spin_lock(&qp->r_rq.lock);
1096
1097 /* sanity check pointers before trusting them */
1098 wq = qp->r_rq.wq;
1099 head = wq->head;
1100 if (head >= qp->r_rq.size)
1101 head = 0;
1102 tail = wq->tail;
1103 if (tail >= qp->r_rq.size)
1104 tail = 0;
1105 while (tail != head) {
1106 wc.wr_id = rvt_get_rwqe_ptr(&qp->r_rq, tail)->wr_id;
1107 if (++tail >= qp->r_rq.size)
1108 tail = 0;
1109 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
1110 }
1111 wq->tail = tail;
1112
1113 spin_unlock(&qp->r_rq.lock);
1114 } else if (qp->ibqp.event_handler) {
1115 ret = 1;
1116 }
1117
1118bail:
1119 return ret;
1120}
1121EXPORT_SYMBOL(rvt_error_qp);
1122
1123/*
1124 * Put the QP into the hash table.
1125 * The hash table holds a reference to the QP.
1126 */
1127static void rvt_insert_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp)
1128{
1129 struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1];
1130 unsigned long flags;
1131
1132 rvt_get_qp(qp);
1133 spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags);
1134
1135 if (qp->ibqp.qp_num <= 1) {
1136 rcu_assign_pointer(rvp->qp[qp->ibqp.qp_num], qp);
1137 } else {
1138 u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits);
1139
1140 qp->next = rdi->qp_dev->qp_table[n];
1141 rcu_assign_pointer(rdi->qp_dev->qp_table[n], qp);
1142 trace_rvt_qpinsert(qp, n);
1143 }
1144
1145 spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags);
1146}
1147
1148/**
1149 * rvt_modify_qp - modify the attributes of a queue pair
1150 * @ibqp: the queue pair who's attributes we're modifying
1151 * @attr: the new attributes
1152 * @attr_mask: the mask of attributes to modify
1153 * @udata: user data for libibverbs.so
1154 *
1155 * Return: 0 on success, otherwise returns an errno.
1156 */
1157int rvt_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1158 int attr_mask, struct ib_udata *udata)
1159{
1160 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1161 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1162 enum ib_qp_state cur_state, new_state;
1163 struct ib_event ev;
1164 int lastwqe = 0;
1165 int mig = 0;
1166 int pmtu = 0; /* for gcc warning only */
1167 enum rdma_link_layer link;
1168 int opa_ah;
1169
1170 link = rdma_port_get_link_layer(ibqp->device, qp->port_num);
1171
1172 spin_lock_irq(&qp->r_lock);
1173 spin_lock(&qp->s_hlock);
1174 spin_lock(&qp->s_lock);
1175
1176 cur_state = attr_mask & IB_QP_CUR_STATE ?
1177 attr->cur_qp_state : qp->state;
1178 new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state;
1179 opa_ah = rdma_cap_opa_ah(ibqp->device, qp->port_num);
1180
1181 if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type,
1182 attr_mask, link))
1183 goto inval;
1184
1185 if (rdi->driver_f.check_modify_qp &&
1186 rdi->driver_f.check_modify_qp(qp, attr, attr_mask, udata))
1187 goto inval;
1188
1189 if (attr_mask & IB_QP_AV) {
1190 if (opa_ah) {
1191 if (rdma_ah_get_dlid(&attr->ah_attr) >=
1192 opa_get_mcast_base(OPA_MCAST_NR))
1193 goto inval;
1194 } else {
1195 if (rdma_ah_get_dlid(&attr->ah_attr) >=
1196 be16_to_cpu(IB_MULTICAST_LID_BASE))
1197 goto inval;
1198 }
1199
1200 if (rvt_check_ah(qp->ibqp.device, &attr->ah_attr))
1201 goto inval;
1202 }
1203
1204 if (attr_mask & IB_QP_ALT_PATH) {
1205 if (opa_ah) {
1206 if (rdma_ah_get_dlid(&attr->alt_ah_attr) >=
1207 opa_get_mcast_base(OPA_MCAST_NR))
1208 goto inval;
1209 } else {
1210 if (rdma_ah_get_dlid(&attr->alt_ah_attr) >=
1211 be16_to_cpu(IB_MULTICAST_LID_BASE))
1212 goto inval;
1213 }
1214
1215 if (rvt_check_ah(qp->ibqp.device, &attr->alt_ah_attr))
1216 goto inval;
1217 if (attr->alt_pkey_index >= rvt_get_npkeys(rdi))
1218 goto inval;
1219 }
1220
1221 if (attr_mask & IB_QP_PKEY_INDEX)
1222 if (attr->pkey_index >= rvt_get_npkeys(rdi))
1223 goto inval;
1224
1225 if (attr_mask & IB_QP_MIN_RNR_TIMER)
1226 if (attr->min_rnr_timer > 31)
1227 goto inval;
1228
1229 if (attr_mask & IB_QP_PORT)
1230 if (qp->ibqp.qp_type == IB_QPT_SMI ||
1231 qp->ibqp.qp_type == IB_QPT_GSI ||
1232 attr->port_num == 0 ||
1233 attr->port_num > ibqp->device->phys_port_cnt)
1234 goto inval;
1235
1236 if (attr_mask & IB_QP_DEST_QPN)
1237 if (attr->dest_qp_num > RVT_QPN_MASK)
1238 goto inval;
1239
1240 if (attr_mask & IB_QP_RETRY_CNT)
1241 if (attr->retry_cnt > 7)
1242 goto inval;
1243
1244 if (attr_mask & IB_QP_RNR_RETRY)
1245 if (attr->rnr_retry > 7)
1246 goto inval;
1247
1248 /*
1249 * Don't allow invalid path_mtu values. OK to set greater
1250 * than the active mtu (or even the max_cap, if we have tuned
1251 * that to a small mtu. We'll set qp->path_mtu
1252 * to the lesser of requested attribute mtu and active,
1253 * for packetizing messages.
1254 * Note that the QP port has to be set in INIT and MTU in RTR.
1255 */
1256 if (attr_mask & IB_QP_PATH_MTU) {
1257 pmtu = rdi->driver_f.get_pmtu_from_attr(rdi, qp, attr);
1258 if (pmtu < 0)
1259 goto inval;
1260 }
1261
1262 if (attr_mask & IB_QP_PATH_MIG_STATE) {
1263 if (attr->path_mig_state == IB_MIG_REARM) {
1264 if (qp->s_mig_state == IB_MIG_ARMED)
1265 goto inval;
1266 if (new_state != IB_QPS_RTS)
1267 goto inval;
1268 } else if (attr->path_mig_state == IB_MIG_MIGRATED) {
1269 if (qp->s_mig_state == IB_MIG_REARM)
1270 goto inval;
1271 if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD)
1272 goto inval;
1273 if (qp->s_mig_state == IB_MIG_ARMED)
1274 mig = 1;
1275 } else {
1276 goto inval;
1277 }
1278 }
1279
1280 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1281 if (attr->max_dest_rd_atomic > rdi->dparms.max_rdma_atomic)
1282 goto inval;
1283
1284 switch (new_state) {
1285 case IB_QPS_RESET:
1286 if (qp->state != IB_QPS_RESET)
1287 rvt_reset_qp(rdi, qp, ibqp->qp_type);
1288 break;
1289
1290 case IB_QPS_RTR:
1291 /* Allow event to re-trigger if QP set to RTR more than once */
1292 qp->r_flags &= ~RVT_R_COMM_EST;
1293 qp->state = new_state;
1294 break;
1295
1296 case IB_QPS_SQD:
1297 qp->s_draining = qp->s_last != qp->s_cur;
1298 qp->state = new_state;
1299 break;
1300
1301 case IB_QPS_SQE:
1302 if (qp->ibqp.qp_type == IB_QPT_RC)
1303 goto inval;
1304 qp->state = new_state;
1305 break;
1306
1307 case IB_QPS_ERR:
1308 lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR);
1309 break;
1310
1311 default:
1312 qp->state = new_state;
1313 break;
1314 }
1315
1316 if (attr_mask & IB_QP_PKEY_INDEX)
1317 qp->s_pkey_index = attr->pkey_index;
1318
1319 if (attr_mask & IB_QP_PORT)
1320 qp->port_num = attr->port_num;
1321
1322 if (attr_mask & IB_QP_DEST_QPN)
1323 qp->remote_qpn = attr->dest_qp_num;
1324
1325 if (attr_mask & IB_QP_SQ_PSN) {
1326 qp->s_next_psn = attr->sq_psn & rdi->dparms.psn_modify_mask;
1327 qp->s_psn = qp->s_next_psn;
1328 qp->s_sending_psn = qp->s_next_psn;
1329 qp->s_last_psn = qp->s_next_psn - 1;
1330 qp->s_sending_hpsn = qp->s_last_psn;
1331 }
1332
1333 if (attr_mask & IB_QP_RQ_PSN)
1334 qp->r_psn = attr->rq_psn & rdi->dparms.psn_modify_mask;
1335
1336 if (attr_mask & IB_QP_ACCESS_FLAGS)
1337 qp->qp_access_flags = attr->qp_access_flags;
1338
1339 if (attr_mask & IB_QP_AV) {
1340 qp->remote_ah_attr = attr->ah_attr;
1341 qp->s_srate = rdma_ah_get_static_rate(&attr->ah_attr);
1342 qp->srate_mbps = ib_rate_to_mbps(qp->s_srate);
1343 }
1344
1345 if (attr_mask & IB_QP_ALT_PATH) {
1346 qp->alt_ah_attr = attr->alt_ah_attr;
1347 qp->s_alt_pkey_index = attr->alt_pkey_index;
1348 }
1349
1350 if (attr_mask & IB_QP_PATH_MIG_STATE) {
1351 qp->s_mig_state = attr->path_mig_state;
1352 if (mig) {
1353 qp->remote_ah_attr = qp->alt_ah_attr;
1354 qp->port_num = rdma_ah_get_port_num(&qp->alt_ah_attr);
1355 qp->s_pkey_index = qp->s_alt_pkey_index;
1356 }
1357 }
1358
1359 if (attr_mask & IB_QP_PATH_MTU) {
1360 qp->pmtu = rdi->driver_f.mtu_from_qp(rdi, qp, pmtu);
1361 qp->log_pmtu = ilog2(qp->pmtu);
1362 }
1363
1364 if (attr_mask & IB_QP_RETRY_CNT) {
1365 qp->s_retry_cnt = attr->retry_cnt;
1366 qp->s_retry = attr->retry_cnt;
1367 }
1368
1369 if (attr_mask & IB_QP_RNR_RETRY) {
1370 qp->s_rnr_retry_cnt = attr->rnr_retry;
1371 qp->s_rnr_retry = attr->rnr_retry;
1372 }
1373
1374 if (attr_mask & IB_QP_MIN_RNR_TIMER)
1375 qp->r_min_rnr_timer = attr->min_rnr_timer;
1376
1377 if (attr_mask & IB_QP_TIMEOUT) {
1378 qp->timeout = attr->timeout;
1379 qp->timeout_jiffies = rvt_timeout_to_jiffies(qp->timeout);
1380 }
1381
1382 if (attr_mask & IB_QP_QKEY)
1383 qp->qkey = attr->qkey;
1384
1385 if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC)
1386 qp->r_max_rd_atomic = attr->max_dest_rd_atomic;
1387
1388 if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC)
1389 qp->s_max_rd_atomic = attr->max_rd_atomic;
1390
1391 if (rdi->driver_f.modify_qp)
1392 rdi->driver_f.modify_qp(qp, attr, attr_mask, udata);
1393
1394 spin_unlock(&qp->s_lock);
1395 spin_unlock(&qp->s_hlock);
1396 spin_unlock_irq(&qp->r_lock);
1397
1398 if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT)
1399 rvt_insert_qp(rdi, qp);
1400
1401 if (lastwqe) {
1402 ev.device = qp->ibqp.device;
1403 ev.element.qp = &qp->ibqp;
1404 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
1405 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1406 }
1407 if (mig) {
1408 ev.device = qp->ibqp.device;
1409 ev.element.qp = &qp->ibqp;
1410 ev.event = IB_EVENT_PATH_MIG;
1411 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
1412 }
1413 return 0;
1414
1415inval:
1416 spin_unlock(&qp->s_lock);
1417 spin_unlock(&qp->s_hlock);
1418 spin_unlock_irq(&qp->r_lock);
1419 return -EINVAL;
1420}
1421
1422/**
1423 * rvt_destroy_qp - destroy a queue pair
1424 * @ibqp: the queue pair to destroy
1425 *
1426 * Note that this can be called while the QP is actively sending or
1427 * receiving!
1428 *
1429 * Return: 0 on success.
1430 */
1431int rvt_destroy_qp(struct ib_qp *ibqp)
1432{
1433 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1434 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1435
1436 spin_lock_irq(&qp->r_lock);
1437 spin_lock(&qp->s_hlock);
1438 spin_lock(&qp->s_lock);
1439 rvt_reset_qp(rdi, qp, ibqp->qp_type);
1440 spin_unlock(&qp->s_lock);
1441 spin_unlock(&qp->s_hlock);
1442 spin_unlock_irq(&qp->r_lock);
1443
1444 wait_event(qp->wait, !atomic_read(&qp->refcount));
1445 /* qpn is now available for use again */
1446 rvt_free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num);
1447
1448 spin_lock(&rdi->n_qps_lock);
1449 rdi->n_qps_allocated--;
1450 if (qp->ibqp.qp_type == IB_QPT_RC) {
1451 rdi->n_rc_qps--;
1452 rdi->busy_jiffies = rdi->n_rc_qps / RC_QP_SCALING_INTERVAL;
1453 }
1454 spin_unlock(&rdi->n_qps_lock);
1455
1456 if (qp->ip)
1457 kref_put(&qp->ip->ref, rvt_release_mmap_info);
1458 else
1459 vfree(qp->r_rq.wq);
1460 vfree(qp->s_wq);
1461 rdi->driver_f.qp_priv_free(rdi, qp);
1462 kfree(qp->s_ack_queue);
1463 kfree(qp);
1464 return 0;
1465}
1466
1467/**
1468 * rvt_query_qp - query an ipbq
1469 * @ibqp: IB qp to query
1470 * @attr: attr struct to fill in
1471 * @attr_mask: attr mask ignored
1472 * @init_attr: struct to fill in
1473 *
1474 * Return: always 0
1475 */
1476int rvt_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
1477 int attr_mask, struct ib_qp_init_attr *init_attr)
1478{
1479 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1480 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1481
1482 attr->qp_state = qp->state;
1483 attr->cur_qp_state = attr->qp_state;
1484 attr->path_mtu = rdi->driver_f.mtu_to_path_mtu(qp->pmtu);
1485 attr->path_mig_state = qp->s_mig_state;
1486 attr->qkey = qp->qkey;
1487 attr->rq_psn = qp->r_psn & rdi->dparms.psn_mask;
1488 attr->sq_psn = qp->s_next_psn & rdi->dparms.psn_mask;
1489 attr->dest_qp_num = qp->remote_qpn;
1490 attr->qp_access_flags = qp->qp_access_flags;
1491 attr->cap.max_send_wr = qp->s_size - 1 -
1492 rdi->dparms.reserved_operations;
1493 attr->cap.max_recv_wr = qp->ibqp.srq ? 0 : qp->r_rq.size - 1;
1494 attr->cap.max_send_sge = qp->s_max_sge;
1495 attr->cap.max_recv_sge = qp->r_rq.max_sge;
1496 attr->cap.max_inline_data = 0;
1497 attr->ah_attr = qp->remote_ah_attr;
1498 attr->alt_ah_attr = qp->alt_ah_attr;
1499 attr->pkey_index = qp->s_pkey_index;
1500 attr->alt_pkey_index = qp->s_alt_pkey_index;
1501 attr->en_sqd_async_notify = 0;
1502 attr->sq_draining = qp->s_draining;
1503 attr->max_rd_atomic = qp->s_max_rd_atomic;
1504 attr->max_dest_rd_atomic = qp->r_max_rd_atomic;
1505 attr->min_rnr_timer = qp->r_min_rnr_timer;
1506 attr->port_num = qp->port_num;
1507 attr->timeout = qp->timeout;
1508 attr->retry_cnt = qp->s_retry_cnt;
1509 attr->rnr_retry = qp->s_rnr_retry_cnt;
1510 attr->alt_port_num =
1511 rdma_ah_get_port_num(&qp->alt_ah_attr);
1512 attr->alt_timeout = qp->alt_timeout;
1513
1514 init_attr->event_handler = qp->ibqp.event_handler;
1515 init_attr->qp_context = qp->ibqp.qp_context;
1516 init_attr->send_cq = qp->ibqp.send_cq;
1517 init_attr->recv_cq = qp->ibqp.recv_cq;
1518 init_attr->srq = qp->ibqp.srq;
1519 init_attr->cap = attr->cap;
1520 if (qp->s_flags & RVT_S_SIGNAL_REQ_WR)
1521 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
1522 else
1523 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
1524 init_attr->qp_type = qp->ibqp.qp_type;
1525 init_attr->port_num = qp->port_num;
1526 return 0;
1527}
1528
1529/**
1530 * rvt_post_receive - post a receive on a QP
1531 * @ibqp: the QP to post the receive on
1532 * @wr: the WR to post
1533 * @bad_wr: the first bad WR is put here
1534 *
1535 * This may be called from interrupt context.
1536 *
1537 * Return: 0 on success otherwise errno
1538 */
1539int rvt_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr,
1540 struct ib_recv_wr **bad_wr)
1541{
1542 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1543 struct rvt_rwq *wq = qp->r_rq.wq;
1544 unsigned long flags;
1545 int qp_err_flush = (ib_rvt_state_ops[qp->state] & RVT_FLUSH_RECV) &&
1546 !qp->ibqp.srq;
1547
1548 /* Check that state is OK to post receive. */
1549 if (!(ib_rvt_state_ops[qp->state] & RVT_POST_RECV_OK) || !wq) {
1550 *bad_wr = wr;
1551 return -EINVAL;
1552 }
1553
1554 for (; wr; wr = wr->next) {
1555 struct rvt_rwqe *wqe;
1556 u32 next;
1557 int i;
1558
1559 if ((unsigned)wr->num_sge > qp->r_rq.max_sge) {
1560 *bad_wr = wr;
1561 return -EINVAL;
1562 }
1563
1564 spin_lock_irqsave(&qp->r_rq.lock, flags);
1565 next = wq->head + 1;
1566 if (next >= qp->r_rq.size)
1567 next = 0;
1568 if (next == wq->tail) {
1569 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
1570 *bad_wr = wr;
1571 return -ENOMEM;
1572 }
1573 if (unlikely(qp_err_flush)) {
1574 struct ib_wc wc;
1575
1576 memset(&wc, 0, sizeof(wc));
1577 wc.qp = &qp->ibqp;
1578 wc.opcode = IB_WC_RECV;
1579 wc.wr_id = wr->wr_id;
1580 wc.status = IB_WC_WR_FLUSH_ERR;
1581 rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1);
1582 } else {
1583 wqe = rvt_get_rwqe_ptr(&qp->r_rq, wq->head);
1584 wqe->wr_id = wr->wr_id;
1585 wqe->num_sge = wr->num_sge;
1586 for (i = 0; i < wr->num_sge; i++)
1587 wqe->sg_list[i] = wr->sg_list[i];
1588 /*
1589 * Make sure queue entry is written
1590 * before the head index.
1591 */
1592 smp_wmb();
1593 wq->head = next;
1594 }
1595 spin_unlock_irqrestore(&qp->r_rq.lock, flags);
1596 }
1597 return 0;
1598}
1599
1600/**
1601 * rvt_qp_valid_operation - validate post send wr request
1602 * @qp - the qp
1603 * @post-parms - the post send table for the driver
1604 * @wr - the work request
1605 *
1606 * The routine validates the operation based on the
1607 * validation table an returns the length of the operation
1608 * which can extend beyond the ib_send_bw. Operation
1609 * dependent flags key atomic operation validation.
1610 *
1611 * There is an exception for UD qps that validates the pd and
1612 * overrides the length to include the additional UD specific
1613 * length.
1614 *
1615 * Returns a negative error or the length of the work request
1616 * for building the swqe.
1617 */
1618static inline int rvt_qp_valid_operation(
1619 struct rvt_qp *qp,
1620 const struct rvt_operation_params *post_parms,
1621 struct ib_send_wr *wr)
1622{
1623 int len;
1624
1625 if (wr->opcode >= RVT_OPERATION_MAX || !post_parms[wr->opcode].length)
1626 return -EINVAL;
1627 if (!(post_parms[wr->opcode].qpt_support & BIT(qp->ibqp.qp_type)))
1628 return -EINVAL;
1629 if ((post_parms[wr->opcode].flags & RVT_OPERATION_PRIV) &&
1630 ibpd_to_rvtpd(qp->ibqp.pd)->user)
1631 return -EINVAL;
1632 if (post_parms[wr->opcode].flags & RVT_OPERATION_ATOMIC_SGE &&
1633 (wr->num_sge == 0 ||
1634 wr->sg_list[0].length < sizeof(u64) ||
1635 wr->sg_list[0].addr & (sizeof(u64) - 1)))
1636 return -EINVAL;
1637 if (post_parms[wr->opcode].flags & RVT_OPERATION_ATOMIC &&
1638 !qp->s_max_rd_atomic)
1639 return -EINVAL;
1640 len = post_parms[wr->opcode].length;
1641 /* UD specific */
1642 if (qp->ibqp.qp_type != IB_QPT_UC &&
1643 qp->ibqp.qp_type != IB_QPT_RC) {
1644 if (qp->ibqp.pd != ud_wr(wr)->ah->pd)
1645 return -EINVAL;
1646 len = sizeof(struct ib_ud_wr);
1647 }
1648 return len;
1649}
1650
1651/**
1652 * rvt_qp_is_avail - determine queue capacity
1653 * @qp: the qp
1654 * @rdi: the rdmavt device
1655 * @reserved_op: is reserved operation
1656 *
1657 * This assumes the s_hlock is held but the s_last
1658 * qp variable is uncontrolled.
1659 *
1660 * For non reserved operations, the qp->s_avail
1661 * may be changed.
1662 *
1663 * The return value is zero or a -ENOMEM.
1664 */
1665static inline int rvt_qp_is_avail(
1666 struct rvt_qp *qp,
1667 struct rvt_dev_info *rdi,
1668 bool reserved_op)
1669{
1670 u32 slast;
1671 u32 avail;
1672 u32 reserved_used;
1673
1674 /* see rvt_qp_wqe_unreserve() */
1675 smp_mb__before_atomic();
1676 reserved_used = atomic_read(&qp->s_reserved_used);
1677 if (unlikely(reserved_op)) {
1678 /* see rvt_qp_wqe_unreserve() */
1679 smp_mb__before_atomic();
1680 if (reserved_used >= rdi->dparms.reserved_operations)
1681 return -ENOMEM;
1682 return 0;
1683 }
1684 /* non-reserved operations */
1685 if (likely(qp->s_avail))
1686 return 0;
1687 slast = READ_ONCE(qp->s_last);
1688 if (qp->s_head >= slast)
1689 avail = qp->s_size - (qp->s_head - slast);
1690 else
1691 avail = slast - qp->s_head;
1692
1693 /* see rvt_qp_wqe_unreserve() */
1694 smp_mb__before_atomic();
1695 reserved_used = atomic_read(&qp->s_reserved_used);
1696 avail = avail - 1 -
1697 (rdi->dparms.reserved_operations - reserved_used);
1698 /* insure we don't assign a negative s_avail */
1699 if ((s32)avail <= 0)
1700 return -ENOMEM;
1701 qp->s_avail = avail;
1702 if (WARN_ON(qp->s_avail >
1703 (qp->s_size - 1 - rdi->dparms.reserved_operations)))
1704 rvt_pr_err(rdi,
1705 "More avail entries than QP RB size.\nQP: %u, size: %u, avail: %u\nhead: %u, tail: %u, cur: %u, acked: %u, last: %u",
1706 qp->ibqp.qp_num, qp->s_size, qp->s_avail,
1707 qp->s_head, qp->s_tail, qp->s_cur,
1708 qp->s_acked, qp->s_last);
1709 return 0;
1710}
1711
1712/**
1713 * rvt_post_one_wr - post one RC, UC, or UD send work request
1714 * @qp: the QP to post on
1715 * @wr: the work request to send
1716 */
1717static int rvt_post_one_wr(struct rvt_qp *qp,
1718 struct ib_send_wr *wr,
1719 int *call_send)
1720{
1721 struct rvt_swqe *wqe;
1722 u32 next;
1723 int i;
1724 int j;
1725 int acc;
1726 struct rvt_lkey_table *rkt;
1727 struct rvt_pd *pd;
1728 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
1729 u8 log_pmtu;
1730 int ret;
1731 size_t cplen;
1732 bool reserved_op;
1733 int local_ops_delayed = 0;
1734
1735 BUILD_BUG_ON(IB_QPT_MAX >= (sizeof(u32) * BITS_PER_BYTE));
1736
1737 /* IB spec says that num_sge == 0 is OK. */
1738 if (unlikely(wr->num_sge > qp->s_max_sge))
1739 return -EINVAL;
1740
1741 ret = rvt_qp_valid_operation(qp, rdi->post_parms, wr);
1742 if (ret < 0)
1743 return ret;
1744 cplen = ret;
1745
1746 /*
1747 * Local operations include fast register and local invalidate.
1748 * Fast register needs to be processed immediately because the
1749 * registered lkey may be used by following work requests and the
1750 * lkey needs to be valid at the time those requests are posted.
1751 * Local invalidate can be processed immediately if fencing is
1752 * not required and no previous local invalidate ops are pending.
1753 * Signaled local operations that have been processed immediately
1754 * need to have requests with "completion only" flags set posted
1755 * to the send queue in order to generate completions.
1756 */
1757 if ((rdi->post_parms[wr->opcode].flags & RVT_OPERATION_LOCAL)) {
1758 switch (wr->opcode) {
1759 case IB_WR_REG_MR:
1760 ret = rvt_fast_reg_mr(qp,
1761 reg_wr(wr)->mr,
1762 reg_wr(wr)->key,
1763 reg_wr(wr)->access);
1764 if (ret || !(wr->send_flags & IB_SEND_SIGNALED))
1765 return ret;
1766 break;
1767 case IB_WR_LOCAL_INV:
1768 if ((wr->send_flags & IB_SEND_FENCE) ||
1769 atomic_read(&qp->local_ops_pending)) {
1770 local_ops_delayed = 1;
1771 } else {
1772 ret = rvt_invalidate_rkey(
1773 qp, wr->ex.invalidate_rkey);
1774 if (ret || !(wr->send_flags & IB_SEND_SIGNALED))
1775 return ret;
1776 }
1777 break;
1778 default:
1779 return -EINVAL;
1780 }
1781 }
1782
1783 reserved_op = rdi->post_parms[wr->opcode].flags &
1784 RVT_OPERATION_USE_RESERVE;
1785 /* check for avail */
1786 ret = rvt_qp_is_avail(qp, rdi, reserved_op);
1787 if (ret)
1788 return ret;
1789 next = qp->s_head + 1;
1790 if (next >= qp->s_size)
1791 next = 0;
1792
1793 rkt = &rdi->lkey_table;
1794 pd = ibpd_to_rvtpd(qp->ibqp.pd);
1795 wqe = rvt_get_swqe_ptr(qp, qp->s_head);
1796
1797 /* cplen has length from above */
1798 memcpy(&wqe->wr, wr, cplen);
1799
1800 wqe->length = 0;
1801 j = 0;
1802 if (wr->num_sge) {
1803 struct rvt_sge *last_sge = NULL;
1804
1805 acc = wr->opcode >= IB_WR_RDMA_READ ?
1806 IB_ACCESS_LOCAL_WRITE : 0;
1807 for (i = 0; i < wr->num_sge; i++) {
1808 u32 length = wr->sg_list[i].length;
1809
1810 if (length == 0)
1811 continue;
1812 ret = rvt_lkey_ok(rkt, pd, &wqe->sg_list[j], last_sge,
1813 &wr->sg_list[i], acc);
1814 if (unlikely(ret < 0))
1815 goto bail_inval_free;
1816 wqe->length += length;
1817 if (ret)
1818 last_sge = &wqe->sg_list[j];
1819 j += ret;
1820 }
1821 wqe->wr.num_sge = j;
1822 }
1823
1824 /* general part of wqe valid - allow for driver checks */
1825 if (rdi->driver_f.check_send_wqe) {
1826 ret = rdi->driver_f.check_send_wqe(qp, wqe);
1827 if (ret < 0)
1828 goto bail_inval_free;
1829 if (ret)
1830 *call_send = ret;
1831 }
1832
1833 log_pmtu = qp->log_pmtu;
1834 if (qp->ibqp.qp_type != IB_QPT_UC &&
1835 qp->ibqp.qp_type != IB_QPT_RC) {
1836 struct rvt_ah *ah = ibah_to_rvtah(wqe->ud_wr.ah);
1837
1838 log_pmtu = ah->log_pmtu;
1839 atomic_inc(&ibah_to_rvtah(ud_wr(wr)->ah)->refcount);
1840 }
1841
1842 if (rdi->post_parms[wr->opcode].flags & RVT_OPERATION_LOCAL) {
1843 if (local_ops_delayed)
1844 atomic_inc(&qp->local_ops_pending);
1845 else
1846 wqe->wr.send_flags |= RVT_SEND_COMPLETION_ONLY;
1847 wqe->ssn = 0;
1848 wqe->psn = 0;
1849 wqe->lpsn = 0;
1850 } else {
1851 wqe->ssn = qp->s_ssn++;
1852 wqe->psn = qp->s_next_psn;
1853 wqe->lpsn = wqe->psn +
1854 (wqe->length ?
1855 ((wqe->length - 1) >> log_pmtu) :
1856 0);
1857 qp->s_next_psn = wqe->lpsn + 1;
1858 }
1859 if (unlikely(reserved_op)) {
1860 wqe->wr.send_flags |= RVT_SEND_RESERVE_USED;
1861 rvt_qp_wqe_reserve(qp, wqe);
1862 } else {
1863 wqe->wr.send_flags &= ~RVT_SEND_RESERVE_USED;
1864 qp->s_avail--;
1865 }
1866 trace_rvt_post_one_wr(qp, wqe, wr->num_sge);
1867 smp_wmb(); /* see request builders */
1868 qp->s_head = next;
1869
1870 return 0;
1871
1872bail_inval_free:
1873 /* release mr holds */
1874 while (j) {
1875 struct rvt_sge *sge = &wqe->sg_list[--j];
1876
1877 rvt_put_mr(sge->mr);
1878 }
1879 return ret;
1880}
1881
1882/**
1883 * rvt_post_send - post a send on a QP
1884 * @ibqp: the QP to post the send on
1885 * @wr: the list of work requests to post
1886 * @bad_wr: the first bad WR is put here
1887 *
1888 * This may be called from interrupt context.
1889 *
1890 * Return: 0 on success else errno
1891 */
1892int rvt_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr,
1893 struct ib_send_wr **bad_wr)
1894{
1895 struct rvt_qp *qp = ibqp_to_rvtqp(ibqp);
1896 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
1897 unsigned long flags = 0;
1898 int call_send;
1899 unsigned nreq = 0;
1900 int err = 0;
1901
1902 spin_lock_irqsave(&qp->s_hlock, flags);
1903
1904 /*
1905 * Ensure QP state is such that we can send. If not bail out early,
1906 * there is no need to do this every time we post a send.
1907 */
1908 if (unlikely(!(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))) {
1909 spin_unlock_irqrestore(&qp->s_hlock, flags);
1910 return -EINVAL;
1911 }
1912
1913 /*
1914 * If the send queue is empty, and we only have a single WR then just go
1915 * ahead and kick the send engine into gear. Otherwise we will always
1916 * just schedule the send to happen later.
1917 */
1918 call_send = qp->s_head == READ_ONCE(qp->s_last) && !wr->next;
1919
1920 for (; wr; wr = wr->next) {
1921 err = rvt_post_one_wr(qp, wr, &call_send);
1922 if (unlikely(err)) {
1923 *bad_wr = wr;
1924 goto bail;
1925 }
1926 nreq++;
1927 }
1928bail:
1929 spin_unlock_irqrestore(&qp->s_hlock, flags);
1930 if (nreq) {
1931 if (call_send)
1932 rdi->driver_f.do_send(qp);
1933 else
1934 rdi->driver_f.schedule_send_no_lock(qp);
1935 }
1936 return err;
1937}
1938
1939/**
1940 * rvt_post_srq_receive - post a receive on a shared receive queue
1941 * @ibsrq: the SRQ to post the receive on
1942 * @wr: the list of work requests to post
1943 * @bad_wr: A pointer to the first WR to cause a problem is put here
1944 *
1945 * This may be called from interrupt context.
1946 *
1947 * Return: 0 on success else errno
1948 */
1949int rvt_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr,
1950 struct ib_recv_wr **bad_wr)
1951{
1952 struct rvt_srq *srq = ibsrq_to_rvtsrq(ibsrq);
1953 struct rvt_rwq *wq;
1954 unsigned long flags;
1955
1956 for (; wr; wr = wr->next) {
1957 struct rvt_rwqe *wqe;
1958 u32 next;
1959 int i;
1960
1961 if ((unsigned)wr->num_sge > srq->rq.max_sge) {
1962 *bad_wr = wr;
1963 return -EINVAL;
1964 }
1965
1966 spin_lock_irqsave(&srq->rq.lock, flags);
1967 wq = srq->rq.wq;
1968 next = wq->head + 1;
1969 if (next >= srq->rq.size)
1970 next = 0;
1971 if (next == wq->tail) {
1972 spin_unlock_irqrestore(&srq->rq.lock, flags);
1973 *bad_wr = wr;
1974 return -ENOMEM;
1975 }
1976
1977 wqe = rvt_get_rwqe_ptr(&srq->rq, wq->head);
1978 wqe->wr_id = wr->wr_id;
1979 wqe->num_sge = wr->num_sge;
1980 for (i = 0; i < wr->num_sge; i++)
1981 wqe->sg_list[i] = wr->sg_list[i];
1982 /* Make sure queue entry is written before the head index. */
1983 smp_wmb();
1984 wq->head = next;
1985 spin_unlock_irqrestore(&srq->rq.lock, flags);
1986 }
1987 return 0;
1988}
1989
1990/**
1991 * qp_comm_est - handle trap with QP established
1992 * @qp: the QP
1993 */
1994void rvt_comm_est(struct rvt_qp *qp)
1995{
1996 qp->r_flags |= RVT_R_COMM_EST;
1997 if (qp->ibqp.event_handler) {
1998 struct ib_event ev;
1999
2000 ev.device = qp->ibqp.device;
2001 ev.element.qp = &qp->ibqp;
2002 ev.event = IB_EVENT_COMM_EST;
2003 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
2004 }
2005}
2006EXPORT_SYMBOL(rvt_comm_est);
2007
2008void rvt_rc_error(struct rvt_qp *qp, enum ib_wc_status err)
2009{
2010 unsigned long flags;
2011 int lastwqe;
2012
2013 spin_lock_irqsave(&qp->s_lock, flags);
2014 lastwqe = rvt_error_qp(qp, err);
2015 spin_unlock_irqrestore(&qp->s_lock, flags);
2016
2017 if (lastwqe) {
2018 struct ib_event ev;
2019
2020 ev.device = qp->ibqp.device;
2021 ev.element.qp = &qp->ibqp;
2022 ev.event = IB_EVENT_QP_LAST_WQE_REACHED;
2023 qp->ibqp.event_handler(&ev, qp->ibqp.qp_context);
2024 }
2025}
2026EXPORT_SYMBOL(rvt_rc_error);
2027
2028/*
2029 * rvt_rnr_tbl_to_usec - return index into ib_rvt_rnr_table
2030 * @index - the index
2031 * return usec from an index into ib_rvt_rnr_table
2032 */
2033unsigned long rvt_rnr_tbl_to_usec(u32 index)
2034{
2035 return ib_rvt_rnr_table[(index & IB_AETH_CREDIT_MASK)];
2036}
2037EXPORT_SYMBOL(rvt_rnr_tbl_to_usec);
2038
2039static inline unsigned long rvt_aeth_to_usec(u32 aeth)
2040{
2041 return ib_rvt_rnr_table[(aeth >> IB_AETH_CREDIT_SHIFT) &
2042 IB_AETH_CREDIT_MASK];
2043}
2044
2045/*
2046 * rvt_add_retry_timer - add/start a retry timer
2047 * @qp - the QP
2048 * add a retry timer on the QP
2049 */
2050void rvt_add_retry_timer(struct rvt_qp *qp)
2051{
2052 struct ib_qp *ibqp = &qp->ibqp;
2053 struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device);
2054
2055 lockdep_assert_held(&qp->s_lock);
2056 qp->s_flags |= RVT_S_TIMER;
2057 /* 4.096 usec. * (1 << qp->timeout) */
2058 qp->s_timer.expires = jiffies + qp->timeout_jiffies +
2059 rdi->busy_jiffies;
2060 add_timer(&qp->s_timer);
2061}
2062EXPORT_SYMBOL(rvt_add_retry_timer);
2063
2064/**
2065 * rvt_add_rnr_timer - add/start an rnr timer
2066 * @qp - the QP
2067 * @aeth - aeth of RNR timeout, simulated aeth for loopback
2068 * add an rnr timer on the QP
2069 */
2070void rvt_add_rnr_timer(struct rvt_qp *qp, u32 aeth)
2071{
2072 u32 to;
2073
2074 lockdep_assert_held(&qp->s_lock);
2075 qp->s_flags |= RVT_S_WAIT_RNR;
2076 to = rvt_aeth_to_usec(aeth);
2077 trace_rvt_rnrnak_add(qp, to);
2078 hrtimer_start(&qp->s_rnr_timer,
2079 ns_to_ktime(1000 * to), HRTIMER_MODE_REL);
2080}
2081EXPORT_SYMBOL(rvt_add_rnr_timer);
2082
2083/**
2084 * rvt_stop_rc_timers - stop all timers
2085 * @qp - the QP
2086 * stop any pending timers
2087 */
2088void rvt_stop_rc_timers(struct rvt_qp *qp)
2089{
2090 lockdep_assert_held(&qp->s_lock);
2091 /* Remove QP from all timers */
2092 if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) {
2093 qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR);
2094 del_timer(&qp->s_timer);
2095 hrtimer_try_to_cancel(&qp->s_rnr_timer);
2096 }
2097}
2098EXPORT_SYMBOL(rvt_stop_rc_timers);
2099
2100/**
2101 * rvt_stop_rnr_timer - stop an rnr timer
2102 * @qp - the QP
2103 *
2104 * stop an rnr timer and return if the timer
2105 * had been pending.
2106 */
2107static void rvt_stop_rnr_timer(struct rvt_qp *qp)
2108{
2109 lockdep_assert_held(&qp->s_lock);
2110 /* Remove QP from rnr timer */
2111 if (qp->s_flags & RVT_S_WAIT_RNR) {
2112 qp->s_flags &= ~RVT_S_WAIT_RNR;
2113 trace_rvt_rnrnak_stop(qp, 0);
2114 }
2115}
2116
2117/**
2118 * rvt_del_timers_sync - wait for any timeout routines to exit
2119 * @qp - the QP
2120 */
2121void rvt_del_timers_sync(struct rvt_qp *qp)
2122{
2123 del_timer_sync(&qp->s_timer);
2124 hrtimer_cancel(&qp->s_rnr_timer);
2125}
2126EXPORT_SYMBOL(rvt_del_timers_sync);
2127
2128/**
2129 * This is called from s_timer for missing responses.
2130 */
2131static void rvt_rc_timeout(struct timer_list *t)
2132{
2133 struct rvt_qp *qp = from_timer(qp, t, s_timer);
2134 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
2135 unsigned long flags;
2136
2137 spin_lock_irqsave(&qp->r_lock, flags);
2138 spin_lock(&qp->s_lock);
2139 if (qp->s_flags & RVT_S_TIMER) {
2140 struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1];
2141
2142 qp->s_flags &= ~RVT_S_TIMER;
2143 rvp->n_rc_timeouts++;
2144 del_timer(&qp->s_timer);
2145 trace_rvt_rc_timeout(qp, qp->s_last_psn + 1);
2146 if (rdi->driver_f.notify_restart_rc)
2147 rdi->driver_f.notify_restart_rc(qp,
2148 qp->s_last_psn + 1,
2149 1);
2150 rdi->driver_f.schedule_send(qp);
2151 }
2152 spin_unlock(&qp->s_lock);
2153 spin_unlock_irqrestore(&qp->r_lock, flags);
2154}
2155
2156/*
2157 * This is called from s_timer for RNR timeouts.
2158 */
2159enum hrtimer_restart rvt_rc_rnr_retry(struct hrtimer *t)
2160{
2161 struct rvt_qp *qp = container_of(t, struct rvt_qp, s_rnr_timer);
2162 struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device);
2163 unsigned long flags;
2164
2165 spin_lock_irqsave(&qp->s_lock, flags);
2166 rvt_stop_rnr_timer(qp);
2167 trace_rvt_rnrnak_timeout(qp, 0);
2168 rdi->driver_f.schedule_send(qp);
2169 spin_unlock_irqrestore(&qp->s_lock, flags);
2170 return HRTIMER_NORESTART;
2171}
2172EXPORT_SYMBOL(rvt_rc_rnr_retry);
2173
2174/**
2175 * rvt_qp_iter_init - initial for QP iteration
2176 * @rdi: rvt devinfo
2177 * @v: u64 value
2178 *
2179 * This returns an iterator suitable for iterating QPs
2180 * in the system.
2181 *
2182 * The @cb is a user defined callback and @v is a 64
2183 * bit value passed to and relevant for processing in the
2184 * @cb. An example use case would be to alter QP processing
2185 * based on criteria not part of the rvt_qp.
2186 *
2187 * Use cases that require memory allocation to succeed
2188 * must preallocate appropriately.
2189 *
2190 * Return: a pointer to an rvt_qp_iter or NULL
2191 */
2192struct rvt_qp_iter *rvt_qp_iter_init(struct rvt_dev_info *rdi,
2193 u64 v,
2194 void (*cb)(struct rvt_qp *qp, u64 v))
2195{
2196 struct rvt_qp_iter *i;
2197
2198 i = kzalloc(sizeof(*i), GFP_KERNEL);
2199 if (!i)
2200 return NULL;
2201
2202 i->rdi = rdi;
2203 /* number of special QPs (SMI/GSI) for device */
2204 i->specials = rdi->ibdev.phys_port_cnt * 2;
2205 i->v = v;
2206 i->cb = cb;
2207
2208 return i;
2209}
2210EXPORT_SYMBOL(rvt_qp_iter_init);
2211
2212/**
2213 * rvt_qp_iter_next - return the next QP in iter
2214 * @iter - the iterator
2215 *
2216 * Fine grained QP iterator suitable for use
2217 * with debugfs seq_file mechanisms.
2218 *
2219 * Updates iter->qp with the current QP when the return
2220 * value is 0.
2221 *
2222 * Return: 0 - iter->qp is valid 1 - no more QPs
2223 */
2224int rvt_qp_iter_next(struct rvt_qp_iter *iter)
2225 __must_hold(RCU)
2226{
2227 int n = iter->n;
2228 int ret = 1;
2229 struct rvt_qp *pqp = iter->qp;
2230 struct rvt_qp *qp;
2231 struct rvt_dev_info *rdi = iter->rdi;
2232
2233 /*
2234 * The approach is to consider the special qps
2235 * as additional table entries before the
2236 * real hash table. Since the qp code sets
2237 * the qp->next hash link to NULL, this works just fine.
2238 *
2239 * iter->specials is 2 * # ports
2240 *
2241 * n = 0..iter->specials is the special qp indices
2242 *
2243 * n = iter->specials..rdi->qp_dev->qp_table_size+iter->specials are
2244 * the potential hash bucket entries
2245 *
2246 */
2247 for (; n < rdi->qp_dev->qp_table_size + iter->specials; n++) {
2248 if (pqp) {
2249 qp = rcu_dereference(pqp->next);
2250 } else {
2251 if (n < iter->specials) {
2252 struct rvt_ibport *rvp;
2253 int pidx;
2254
2255 pidx = n % rdi->ibdev.phys_port_cnt;
2256 rvp = rdi->ports[pidx];
2257 qp = rcu_dereference(rvp->qp[n & 1]);
2258 } else {
2259 qp = rcu_dereference(
2260 rdi->qp_dev->qp_table[
2261 (n - iter->specials)]);
2262 }
2263 }
2264 pqp = qp;
2265 if (qp) {
2266 iter->qp = qp;
2267 iter->n = n;
2268 return 0;
2269 }
2270 }
2271 return ret;
2272}
2273EXPORT_SYMBOL(rvt_qp_iter_next);
2274
2275/**
2276 * rvt_qp_iter - iterate all QPs
2277 * @rdi - rvt devinfo
2278 * @v - a 64 bit value
2279 * @cb - a callback
2280 *
2281 * This provides a way for iterating all QPs.
2282 *
2283 * The @cb is a user defined callback and @v is a 64
2284 * bit value passed to and relevant for processing in the
2285 * cb. An example use case would be to alter QP processing
2286 * based on criteria not part of the rvt_qp.
2287 *
2288 * The code has an internal iterator to simplify
2289 * non seq_file use cases.
2290 */
2291void rvt_qp_iter(struct rvt_dev_info *rdi,
2292 u64 v,
2293 void (*cb)(struct rvt_qp *qp, u64 v))
2294{
2295 int ret;
2296 struct rvt_qp_iter i = {
2297 .rdi = rdi,
2298 .specials = rdi->ibdev.phys_port_cnt * 2,
2299 .v = v,
2300 .cb = cb
2301 };
2302
2303 rcu_read_lock();
2304 do {
2305 ret = rvt_qp_iter_next(&i);
2306 if (!ret) {
2307 rvt_get_qp(i.qp);
2308 rcu_read_unlock();
2309 i.cb(i.qp, i.v);
2310 rcu_read_lock();
2311 rvt_put_qp(i.qp);
2312 }
2313 } while (!ret);
2314 rcu_read_unlock();
2315}
2316EXPORT_SYMBOL(rvt_qp_iter);