]>
Commit | Line | Data |
---|---|---|
b518d3e6 DD |
1 | /* |
2 | * Copyright(c) 2015 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 | ||
3b0b3fb3 | 48 | #include <linux/hash.h> |
0acb0cc7 DD |
49 | #include <linux/bitops.h> |
50 | #include <linux/lockdep.h> | |
515667f8 DD |
51 | #include <linux/vmalloc.h> |
52 | #include <linux/slab.h> | |
53 | #include <rdma/ib_verbs.h> | |
b518d3e6 | 54 | #include "qp.h" |
515667f8 | 55 | #include "vt.h" |
3b0b3fb3 | 56 | #include "trace.h" |
b518d3e6 | 57 | |
bfbac097 DD |
58 | /* |
59 | * Note that it is OK to post send work requests in the SQE and ERR | |
60 | * states; rvt_do_send() will process them and generate error | |
61 | * completions as per IB 1.2 C10-96. | |
62 | */ | |
63 | const int ib_rvt_state_ops[IB_QPS_ERR + 1] = { | |
64 | [IB_QPS_RESET] = 0, | |
65 | [IB_QPS_INIT] = RVT_POST_RECV_OK, | |
66 | [IB_QPS_RTR] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK, | |
67 | [IB_QPS_RTS] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK | | |
68 | RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK | | |
69 | RVT_PROCESS_NEXT_SEND_OK, | |
70 | [IB_QPS_SQD] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK | | |
71 | RVT_POST_SEND_OK | RVT_PROCESS_SEND_OK, | |
72 | [IB_QPS_SQE] = RVT_POST_RECV_OK | RVT_PROCESS_RECV_OK | | |
73 | RVT_POST_SEND_OK | RVT_FLUSH_SEND, | |
74 | [IB_QPS_ERR] = RVT_POST_RECV_OK | RVT_FLUSH_RECV | | |
75 | RVT_POST_SEND_OK | RVT_FLUSH_SEND, | |
76 | }; | |
77 | EXPORT_SYMBOL(ib_rvt_state_ops); | |
78 | ||
d2b8d4da MM |
79 | static void get_map_page(struct rvt_qpn_table *qpt, |
80 | struct rvt_qpn_map *map, | |
81 | gfp_t gfp) | |
0acb0cc7 | 82 | { |
d2b8d4da | 83 | unsigned long page = get_zeroed_page(gfp); |
0acb0cc7 DD |
84 | |
85 | /* | |
86 | * Free the page if someone raced with us installing it. | |
87 | */ | |
88 | ||
89 | spin_lock(&qpt->lock); | |
90 | if (map->page) | |
91 | free_page(page); | |
92 | else | |
93 | map->page = (void *)page; | |
94 | spin_unlock(&qpt->lock); | |
95 | } | |
96 | ||
97 | /** | |
98 | * init_qpn_table - initialize the QP number table for a device | |
99 | * @qpt: the QPN table | |
100 | */ | |
101 | static int init_qpn_table(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt) | |
102 | { | |
103 | u32 offset, i; | |
104 | struct rvt_qpn_map *map; | |
105 | int ret = 0; | |
106 | ||
fef2efd6 | 107 | if (!(rdi->dparms.qpn_res_end >= rdi->dparms.qpn_res_start)) |
0acb0cc7 DD |
108 | return -EINVAL; |
109 | ||
110 | spin_lock_init(&qpt->lock); | |
111 | ||
112 | qpt->last = rdi->dparms.qpn_start; | |
113 | qpt->incr = rdi->dparms.qpn_inc << rdi->dparms.qos_shift; | |
114 | ||
115 | /* | |
116 | * Drivers may want some QPs beyond what we need for verbs let them use | |
117 | * our qpn table. No need for two. Lets go ahead and mark the bitmaps | |
118 | * for those. The reserved range must be *after* the range which verbs | |
119 | * will pick from. | |
120 | */ | |
121 | ||
122 | /* Figure out number of bit maps needed before reserved range */ | |
123 | qpt->nmaps = rdi->dparms.qpn_res_start / RVT_BITS_PER_PAGE; | |
124 | ||
125 | /* This should always be zero */ | |
126 | offset = rdi->dparms.qpn_res_start & RVT_BITS_PER_PAGE_MASK; | |
127 | ||
128 | /* Starting with the first reserved bit map */ | |
129 | map = &qpt->map[qpt->nmaps]; | |
130 | ||
131 | rvt_pr_info(rdi, "Reserving QPNs from 0x%x to 0x%x for non-verbs use\n", | |
132 | rdi->dparms.qpn_res_start, rdi->dparms.qpn_res_end); | |
fef2efd6 | 133 | for (i = rdi->dparms.qpn_res_start; i <= rdi->dparms.qpn_res_end; i++) { |
0acb0cc7 | 134 | if (!map->page) { |
d2b8d4da | 135 | get_map_page(qpt, map, GFP_KERNEL); |
0acb0cc7 DD |
136 | if (!map->page) { |
137 | ret = -ENOMEM; | |
138 | break; | |
139 | } | |
140 | } | |
141 | set_bit(offset, map->page); | |
142 | offset++; | |
143 | if (offset == RVT_BITS_PER_PAGE) { | |
144 | /* next page */ | |
145 | qpt->nmaps++; | |
146 | map++; | |
147 | offset = 0; | |
148 | } | |
149 | } | |
150 | return ret; | |
151 | } | |
152 | ||
153 | /** | |
154 | * free_qpn_table - free the QP number table for a device | |
155 | * @qpt: the QPN table | |
156 | */ | |
157 | static void free_qpn_table(struct rvt_qpn_table *qpt) | |
158 | { | |
159 | int i; | |
160 | ||
161 | for (i = 0; i < ARRAY_SIZE(qpt->map); i++) | |
162 | free_page((unsigned long)qpt->map[i].page); | |
163 | } | |
164 | ||
165 | int rvt_driver_qp_init(struct rvt_dev_info *rdi) | |
166 | { | |
167 | int i; | |
168 | int ret = -ENOMEM; | |
169 | ||
170 | if (rdi->flags & RVT_FLAG_QP_INIT_DRIVER) { | |
171 | rvt_pr_info(rdi, "Driver is doing QP init.\n"); | |
172 | return 0; | |
173 | } | |
174 | ||
175 | if (!rdi->dparms.qp_table_size) | |
176 | return -EINVAL; | |
177 | ||
178 | /* | |
179 | * If driver is not doing any QP allocation then make sure it is | |
180 | * providing the necessary QP functions. | |
181 | */ | |
515667f8 DD |
182 | if (!rdi->driver_f.free_all_qps || |
183 | !rdi->driver_f.qp_priv_alloc || | |
184 | !rdi->driver_f.qp_priv_free || | |
185 | !rdi->driver_f.notify_qp_reset) | |
0acb0cc7 DD |
186 | return -EINVAL; |
187 | ||
188 | /* allocate parent object */ | |
189 | rdi->qp_dev = kzalloc(sizeof(*rdi->qp_dev), GFP_KERNEL); | |
190 | if (!rdi->qp_dev) | |
191 | return -ENOMEM; | |
192 | ||
193 | /* allocate hash table */ | |
194 | rdi->qp_dev->qp_table_size = rdi->dparms.qp_table_size; | |
195 | rdi->qp_dev->qp_table_bits = ilog2(rdi->dparms.qp_table_size); | |
196 | rdi->qp_dev->qp_table = | |
197 | kmalloc(rdi->qp_dev->qp_table_size * | |
198 | sizeof(*rdi->qp_dev->qp_table), | |
199 | GFP_KERNEL); | |
200 | if (!rdi->qp_dev->qp_table) | |
201 | goto no_qp_table; | |
202 | ||
203 | for (i = 0; i < rdi->qp_dev->qp_table_size; i++) | |
204 | RCU_INIT_POINTER(rdi->qp_dev->qp_table[i], NULL); | |
205 | ||
206 | spin_lock_init(&rdi->qp_dev->qpt_lock); | |
207 | ||
208 | /* initialize qpn map */ | |
209 | if (init_qpn_table(rdi, &rdi->qp_dev->qpn_table)) | |
210 | goto fail_table; | |
211 | ||
515667f8 DD |
212 | spin_lock_init(&rdi->n_qps_lock); |
213 | ||
214 | return 0; | |
0acb0cc7 DD |
215 | |
216 | fail_table: | |
217 | kfree(rdi->qp_dev->qp_table); | |
218 | free_qpn_table(&rdi->qp_dev->qpn_table); | |
219 | ||
220 | no_qp_table: | |
221 | kfree(rdi->qp_dev); | |
222 | ||
223 | return ret; | |
224 | } | |
225 | ||
226 | /** | |
227 | * free_all_qps - check for QPs still in use | |
228 | * @qpt: the QP table to empty | |
229 | * | |
230 | * There should not be any QPs still in use. | |
231 | * Free memory for table. | |
232 | */ | |
515667f8 | 233 | static unsigned rvt_free_all_qps(struct rvt_dev_info *rdi) |
0acb0cc7 DD |
234 | { |
235 | unsigned long flags; | |
236 | struct rvt_qp *qp; | |
237 | unsigned n, qp_inuse = 0; | |
238 | spinlock_t *ql; /* work around too long line below */ | |
239 | ||
515667f8 DD |
240 | if (rdi->driver_f.free_all_qps) |
241 | qp_inuse = rdi->driver_f.free_all_qps(rdi); | |
0acb0cc7 | 242 | |
4e74080b DD |
243 | qp_inuse += rvt_mcast_tree_empty(rdi); |
244 | ||
0acb0cc7 | 245 | if (!rdi->qp_dev) |
515667f8 | 246 | return qp_inuse; |
0acb0cc7 DD |
247 | |
248 | ql = &rdi->qp_dev->qpt_lock; | |
515667f8 | 249 | spin_lock_irqsave(ql, flags); |
0acb0cc7 DD |
250 | for (n = 0; n < rdi->qp_dev->qp_table_size; n++) { |
251 | qp = rcu_dereference_protected(rdi->qp_dev->qp_table[n], | |
252 | lockdep_is_held(ql)); | |
253 | RCU_INIT_POINTER(rdi->qp_dev->qp_table[n], NULL); | |
515667f8 DD |
254 | |
255 | for (; qp; qp = rcu_dereference_protected(qp->next, | |
256 | lockdep_is_held(ql))) | |
0acb0cc7 | 257 | qp_inuse++; |
0acb0cc7 DD |
258 | } |
259 | spin_unlock_irqrestore(ql, flags); | |
260 | synchronize_rcu(); | |
261 | return qp_inuse; | |
262 | } | |
263 | ||
264 | void rvt_qp_exit(struct rvt_dev_info *rdi) | |
265 | { | |
515667f8 | 266 | u32 qps_inuse = rvt_free_all_qps(rdi); |
0acb0cc7 | 267 | |
0acb0cc7 DD |
268 | if (qps_inuse) |
269 | rvt_pr_err(rdi, "QP memory leak! %u still in use\n", | |
270 | qps_inuse); | |
271 | if (!rdi->qp_dev) | |
272 | return; | |
273 | ||
515667f8 DD |
274 | if (rdi->flags & RVT_FLAG_QP_INIT_DRIVER) |
275 | return; /* driver did the qp init so nothing else to do */ | |
276 | ||
0acb0cc7 DD |
277 | kfree(rdi->qp_dev->qp_table); |
278 | free_qpn_table(&rdi->qp_dev->qpn_table); | |
279 | kfree(rdi->qp_dev); | |
280 | } | |
281 | ||
515667f8 DD |
282 | static inline unsigned mk_qpn(struct rvt_qpn_table *qpt, |
283 | struct rvt_qpn_map *map, unsigned off) | |
284 | { | |
285 | return (map - qpt->map) * RVT_BITS_PER_PAGE + off; | |
286 | } | |
287 | ||
288 | /* | |
289 | * Allocate the next available QPN or | |
290 | * zero/one for QP type IB_QPT_SMI/IB_QPT_GSI. | |
291 | */ | |
292 | static int alloc_qpn(struct rvt_dev_info *rdi, struct rvt_qpn_table *qpt, | |
d2b8d4da | 293 | enum ib_qp_type type, u8 port, gfp_t gfp) |
515667f8 DD |
294 | { |
295 | u32 i, offset, max_scan, qpn; | |
296 | struct rvt_qpn_map *map; | |
297 | u32 ret; | |
298 | ||
299 | if (rdi->driver_f.alloc_qpn) | |
d2b8d4da MM |
300 | return rdi->driver_f.alloc_qpn(rdi, qpt, type, port, |
301 | GFP_KERNEL); | |
515667f8 DD |
302 | |
303 | if (type == IB_QPT_SMI || type == IB_QPT_GSI) { | |
304 | unsigned n; | |
305 | ||
306 | ret = type == IB_QPT_GSI; | |
307 | n = 1 << (ret + 2 * (port - 1)); | |
308 | spin_lock(&qpt->lock); | |
309 | if (qpt->flags & n) | |
310 | ret = -EINVAL; | |
311 | else | |
312 | qpt->flags |= n; | |
313 | spin_unlock(&qpt->lock); | |
314 | goto bail; | |
315 | } | |
316 | ||
317 | qpn = qpt->last + qpt->incr; | |
318 | if (qpn >= RVT_QPN_MAX) | |
319 | qpn = qpt->incr | ((qpt->last & 1) ^ 1); | |
320 | /* offset carries bit 0 */ | |
321 | offset = qpn & RVT_BITS_PER_PAGE_MASK; | |
322 | map = &qpt->map[qpn / RVT_BITS_PER_PAGE]; | |
323 | max_scan = qpt->nmaps - !offset; | |
324 | for (i = 0;;) { | |
325 | if (unlikely(!map->page)) { | |
d2b8d4da | 326 | get_map_page(qpt, map, gfp); |
515667f8 DD |
327 | if (unlikely(!map->page)) |
328 | break; | |
329 | } | |
330 | do { | |
331 | if (!test_and_set_bit(offset, map->page)) { | |
332 | qpt->last = qpn; | |
333 | ret = qpn; | |
334 | goto bail; | |
335 | } | |
336 | offset += qpt->incr; | |
337 | /* | |
338 | * This qpn might be bogus if offset >= BITS_PER_PAGE. | |
339 | * That is OK. It gets re-assigned below | |
340 | */ | |
341 | qpn = mk_qpn(qpt, map, offset); | |
342 | } while (offset < RVT_BITS_PER_PAGE && qpn < RVT_QPN_MAX); | |
343 | /* | |
344 | * In order to keep the number of pages allocated to a | |
345 | * minimum, we scan the all existing pages before increasing | |
346 | * the size of the bitmap table. | |
347 | */ | |
348 | if (++i > max_scan) { | |
349 | if (qpt->nmaps == RVT_QPNMAP_ENTRIES) | |
350 | break; | |
351 | map = &qpt->map[qpt->nmaps++]; | |
352 | /* start at incr with current bit 0 */ | |
353 | offset = qpt->incr | (offset & 1); | |
354 | } else if (map < &qpt->map[qpt->nmaps]) { | |
355 | ++map; | |
356 | /* start at incr with current bit 0 */ | |
357 | offset = qpt->incr | (offset & 1); | |
358 | } else { | |
359 | map = &qpt->map[0]; | |
360 | /* wrap to first map page, invert bit 0 */ | |
361 | offset = qpt->incr | ((offset & 1) ^ 1); | |
362 | } | |
363 | /* there can be no bits at shift and below */ | |
364 | WARN_ON(offset & (rdi->dparms.qos_shift - 1)); | |
365 | qpn = mk_qpn(qpt, map, offset); | |
366 | } | |
367 | ||
368 | ret = -ENOMEM; | |
369 | ||
370 | bail: | |
371 | return ret; | |
372 | } | |
373 | ||
374 | static void free_qpn(struct rvt_qpn_table *qpt, u32 qpn) | |
375 | { | |
376 | struct rvt_qpn_map *map; | |
377 | ||
378 | map = qpt->map + qpn / RVT_BITS_PER_PAGE; | |
379 | if (map->page) | |
380 | clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page); | |
381 | } | |
382 | ||
383 | /** | |
384 | * reset_qp - initialize the QP state to the reset state | |
385 | * @qp: the QP to reset | |
386 | * @type: the QP type | |
3b0b3fb3 | 387 | * r and s lock are required to be held by the caller |
515667f8 | 388 | */ |
5a9cf6f2 DD |
389 | void rvt_reset_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp, |
390 | enum ib_qp_type type) | |
515667f8 | 391 | { |
3b0b3fb3 DD |
392 | if (qp->state != IB_QPS_RESET) { |
393 | qp->state = IB_QPS_RESET; | |
394 | ||
395 | /* Let drivers flush their waitlist */ | |
396 | rdi->driver_f.flush_qp_waiters(qp); | |
397 | qp->s_flags &= ~(RVT_S_TIMER | RVT_S_ANY_WAIT); | |
398 | spin_unlock(&qp->s_lock); | |
399 | spin_unlock_irq(&qp->r_lock); | |
400 | ||
401 | /* Stop the send queue and the retry timer */ | |
402 | rdi->driver_f.stop_send_queue(qp); | |
403 | del_timer_sync(&qp->s_timer); | |
404 | ||
405 | /* Wait for things to stop */ | |
406 | rdi->driver_f.quiesce_qp(qp); | |
407 | ||
408 | /* take qp out the hash and wait for it to be unused */ | |
409 | rvt_remove_qp(rdi, qp); | |
410 | wait_event(qp->wait, !atomic_read(&qp->refcount)); | |
411 | ||
412 | /* grab the lock b/c it was locked at call time */ | |
413 | spin_lock_irq(&qp->r_lock); | |
414 | spin_lock(&qp->s_lock); | |
415 | ||
416 | rvt_clear_mr_refs(qp, 1); | |
417 | } | |
515667f8 DD |
418 | |
419 | /* | |
3b0b3fb3 DD |
420 | * Let the driver do any tear down it needs to for a qp |
421 | * that has been reset | |
515667f8 DD |
422 | */ |
423 | rdi->driver_f.notify_qp_reset(qp); | |
424 | ||
3b0b3fb3 DD |
425 | qp->remote_qpn = 0; |
426 | qp->qkey = 0; | |
427 | qp->qp_access_flags = 0; | |
515667f8 DD |
428 | qp->s_flags &= RVT_S_SIGNAL_REQ_WR; |
429 | qp->s_hdrwords = 0; | |
430 | qp->s_wqe = NULL; | |
431 | qp->s_draining = 0; | |
432 | qp->s_next_psn = 0; | |
433 | qp->s_last_psn = 0; | |
434 | qp->s_sending_psn = 0; | |
435 | qp->s_sending_hpsn = 0; | |
436 | qp->s_psn = 0; | |
437 | qp->r_psn = 0; | |
438 | qp->r_msn = 0; | |
439 | if (type == IB_QPT_RC) { | |
440 | qp->s_state = IB_OPCODE_RC_SEND_LAST; | |
441 | qp->r_state = IB_OPCODE_RC_SEND_LAST; | |
442 | } else { | |
443 | qp->s_state = IB_OPCODE_UC_SEND_LAST; | |
444 | qp->r_state = IB_OPCODE_UC_SEND_LAST; | |
445 | } | |
446 | qp->s_ack_state = IB_OPCODE_RC_ACKNOWLEDGE; | |
447 | qp->r_nak_state = 0; | |
448 | qp->r_aflags = 0; | |
449 | qp->r_flags = 0; | |
450 | qp->s_head = 0; | |
451 | qp->s_tail = 0; | |
452 | qp->s_cur = 0; | |
453 | qp->s_acked = 0; | |
454 | qp->s_last = 0; | |
455 | qp->s_ssn = 1; | |
456 | qp->s_lsn = 0; | |
457 | qp->s_mig_state = IB_MIG_MIGRATED; | |
458 | memset(qp->s_ack_queue, 0, sizeof(qp->s_ack_queue)); | |
459 | qp->r_head_ack_queue = 0; | |
460 | qp->s_tail_ack_queue = 0; | |
461 | qp->s_num_rd_atomic = 0; | |
462 | if (qp->r_rq.wq) { | |
463 | qp->r_rq.wq->head = 0; | |
464 | qp->r_rq.wq->tail = 0; | |
465 | } | |
466 | qp->r_sge.num_sge = 0; | |
467 | } | |
5a9cf6f2 | 468 | EXPORT_SYMBOL(rvt_reset_qp); |
515667f8 | 469 | |
b518d3e6 DD |
470 | /** |
471 | * rvt_create_qp - create a queue pair for a device | |
472 | * @ibpd: the protection domain who's device we create the queue pair for | |
473 | * @init_attr: the attributes of the queue pair | |
474 | * @udata: user data for libibverbs.so | |
475 | * | |
515667f8 DD |
476 | * Queue pair creation is mostly an rvt issue. However, drivers have their own |
477 | * unique idea of what queue pair numbers mean. For instance there is a reserved | |
478 | * range for PSM. | |
479 | * | |
b518d3e6 DD |
480 | * Returns the queue pair on success, otherwise returns an errno. |
481 | * | |
482 | * Called by the ib_create_qp() core verbs function. | |
483 | */ | |
484 | struct ib_qp *rvt_create_qp(struct ib_pd *ibpd, | |
485 | struct ib_qp_init_attr *init_attr, | |
486 | struct ib_udata *udata) | |
487 | { | |
515667f8 DD |
488 | struct rvt_qp *qp; |
489 | int err; | |
490 | struct rvt_swqe *swq = NULL; | |
491 | size_t sz; | |
492 | size_t sg_list_sz; | |
493 | struct ib_qp *ret = ERR_PTR(-ENOMEM); | |
494 | struct rvt_dev_info *rdi = ib_to_rvt(ibpd->device); | |
495 | void *priv = NULL; | |
d2b8d4da | 496 | gfp_t gfp; |
515667f8 DD |
497 | |
498 | if (!rdi) | |
499 | return ERR_PTR(-EINVAL); | |
500 | ||
501 | if (init_attr->cap.max_send_sge > rdi->dparms.props.max_sge || | |
502 | init_attr->cap.max_send_wr > rdi->dparms.props.max_qp_wr || | |
d2b8d4da | 503 | init_attr->create_flags & ~(IB_QP_CREATE_USE_GFP_NOIO)) |
515667f8 DD |
504 | return ERR_PTR(-EINVAL); |
505 | ||
d2b8d4da MM |
506 | /* GFP_NOIO is applicable to RC QP's only */ |
507 | ||
508 | if (init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO && | |
509 | init_attr->qp_type != IB_QPT_RC) | |
510 | return ERR_PTR(-EINVAL); | |
511 | ||
512 | gfp = init_attr->create_flags & IB_QP_CREATE_USE_GFP_NOIO ? | |
513 | GFP_NOIO : GFP_KERNEL; | |
514 | ||
515667f8 DD |
515 | /* Check receive queue parameters if no SRQ is specified. */ |
516 | if (!init_attr->srq) { | |
517 | if (init_attr->cap.max_recv_sge > rdi->dparms.props.max_sge || | |
518 | init_attr->cap.max_recv_wr > rdi->dparms.props.max_qp_wr) | |
519 | return ERR_PTR(-EINVAL); | |
520 | ||
521 | if (init_attr->cap.max_send_sge + | |
522 | init_attr->cap.max_send_wr + | |
523 | init_attr->cap.max_recv_sge + | |
524 | init_attr->cap.max_recv_wr == 0) | |
525 | return ERR_PTR(-EINVAL); | |
526 | } | |
527 | ||
528 | switch (init_attr->qp_type) { | |
529 | case IB_QPT_SMI: | |
530 | case IB_QPT_GSI: | |
531 | if (init_attr->port_num == 0 || | |
532 | init_attr->port_num > ibpd->device->phys_port_cnt) | |
533 | return ERR_PTR(-EINVAL); | |
534 | case IB_QPT_UC: | |
535 | case IB_QPT_RC: | |
536 | case IB_QPT_UD: | |
537 | sz = sizeof(struct rvt_sge) * | |
538 | init_attr->cap.max_send_sge + | |
539 | sizeof(struct rvt_swqe); | |
d2b8d4da MM |
540 | if (gfp == GFP_NOIO) |
541 | swq = __vmalloc( | |
542 | (init_attr->cap.max_send_wr + 1) * sz, | |
543 | gfp, PAGE_KERNEL); | |
544 | else | |
545 | swq = vmalloc( | |
546 | (init_attr->cap.max_send_wr + 1) * sz); | |
515667f8 DD |
547 | if (!swq) |
548 | return ERR_PTR(-ENOMEM); | |
549 | ||
550 | sz = sizeof(*qp); | |
551 | sg_list_sz = 0; | |
552 | if (init_attr->srq) { | |
553 | struct rvt_srq *srq = ibsrq_to_rvtsrq(init_attr->srq); | |
554 | ||
555 | if (srq->rq.max_sge > 1) | |
556 | sg_list_sz = sizeof(*qp->r_sg_list) * | |
557 | (srq->rq.max_sge - 1); | |
558 | } else if (init_attr->cap.max_recv_sge > 1) | |
559 | sg_list_sz = sizeof(*qp->r_sg_list) * | |
560 | (init_attr->cap.max_recv_sge - 1); | |
d2b8d4da | 561 | qp = kzalloc(sz + sg_list_sz, gfp); |
515667f8 DD |
562 | if (!qp) |
563 | goto bail_swq; | |
564 | ||
565 | RCU_INIT_POINTER(qp->next, NULL); | |
566 | ||
567 | /* | |
568 | * Driver needs to set up it's private QP structure and do any | |
569 | * initialization that is needed. | |
570 | */ | |
d2b8d4da | 571 | priv = rdi->driver_f.qp_priv_alloc(rdi, qp, gfp); |
515667f8 DD |
572 | if (!priv) |
573 | goto bail_qp; | |
574 | qp->priv = priv; | |
575 | qp->timeout_jiffies = | |
576 | usecs_to_jiffies((4096UL * (1UL << qp->timeout)) / | |
577 | 1000UL); | |
578 | if (init_attr->srq) { | |
579 | sz = 0; | |
580 | } else { | |
581 | qp->r_rq.size = init_attr->cap.max_recv_wr + 1; | |
582 | qp->r_rq.max_sge = init_attr->cap.max_recv_sge; | |
583 | sz = (sizeof(struct ib_sge) * qp->r_rq.max_sge) + | |
584 | sizeof(struct rvt_rwqe); | |
d2b8d4da MM |
585 | if (udata) |
586 | qp->r_rq.wq = vmalloc_user( | |
587 | sizeof(struct rvt_rwq) + | |
588 | qp->r_rq.size * sz); | |
589 | else if (gfp == GFP_NOIO) | |
590 | qp->r_rq.wq = __vmalloc( | |
591 | sizeof(struct rvt_rwq) + | |
592 | qp->r_rq.size * sz, | |
593 | gfp, PAGE_KERNEL); | |
594 | else | |
595 | qp->r_rq.wq = vmalloc( | |
596 | sizeof(struct rvt_rwq) + | |
597 | qp->r_rq.size * sz); | |
515667f8 DD |
598 | if (!qp->r_rq.wq) |
599 | goto bail_driver_priv; | |
600 | } | |
601 | ||
602 | /* | |
603 | * ib_create_qp() will initialize qp->ibqp | |
604 | * except for qp->ibqp.qp_num. | |
605 | */ | |
606 | spin_lock_init(&qp->r_lock); | |
607 | spin_lock_init(&qp->s_lock); | |
608 | spin_lock_init(&qp->r_rq.lock); | |
609 | atomic_set(&qp->refcount, 0); | |
610 | init_waitqueue_head(&qp->wait); | |
611 | init_timer(&qp->s_timer); | |
612 | qp->s_timer.data = (unsigned long)qp; | |
613 | INIT_LIST_HEAD(&qp->rspwait); | |
614 | qp->state = IB_QPS_RESET; | |
615 | qp->s_wq = swq; | |
616 | qp->s_size = init_attr->cap.max_send_wr + 1; | |
617 | qp->s_max_sge = init_attr->cap.max_send_sge; | |
618 | if (init_attr->sq_sig_type == IB_SIGNAL_REQ_WR) | |
619 | qp->s_flags = RVT_S_SIGNAL_REQ_WR; | |
620 | ||
621 | err = alloc_qpn(rdi, &rdi->qp_dev->qpn_table, | |
622 | init_attr->qp_type, | |
d2b8d4da | 623 | init_attr->port_num, gfp); |
515667f8 DD |
624 | if (err < 0) { |
625 | ret = ERR_PTR(err); | |
626 | goto bail_rq_wq; | |
627 | } | |
628 | qp->ibqp.qp_num = err; | |
629 | qp->port_num = init_attr->port_num; | |
5a9cf6f2 | 630 | rvt_reset_qp(rdi, qp, init_attr->qp_type); |
515667f8 DD |
631 | break; |
632 | ||
633 | default: | |
634 | /* Don't support raw QPs */ | |
635 | return ERR_PTR(-EINVAL); | |
636 | } | |
637 | ||
638 | init_attr->cap.max_inline_data = 0; | |
639 | ||
640 | /* | |
641 | * Return the address of the RWQ as the offset to mmap. | |
bfbac097 | 642 | * See rvt_mmap() for details. |
515667f8 DD |
643 | */ |
644 | if (udata && udata->outlen >= sizeof(__u64)) { | |
645 | if (!qp->r_rq.wq) { | |
646 | __u64 offset = 0; | |
647 | ||
648 | err = ib_copy_to_udata(udata, &offset, | |
649 | sizeof(offset)); | |
650 | if (err) { | |
651 | ret = ERR_PTR(err); | |
652 | goto bail_qpn; | |
653 | } | |
654 | } else { | |
655 | u32 s = sizeof(struct rvt_rwq) + qp->r_rq.size * sz; | |
656 | ||
657 | qp->ip = rvt_create_mmap_info(rdi, s, | |
658 | ibpd->uobject->context, | |
659 | qp->r_rq.wq); | |
660 | if (!qp->ip) { | |
661 | ret = ERR_PTR(-ENOMEM); | |
662 | goto bail_qpn; | |
663 | } | |
664 | ||
665 | err = ib_copy_to_udata(udata, &qp->ip->offset, | |
666 | sizeof(qp->ip->offset)); | |
667 | if (err) { | |
668 | ret = ERR_PTR(err); | |
669 | goto bail_ip; | |
670 | } | |
671 | } | |
672 | } | |
673 | ||
674 | spin_lock(&rdi->n_qps_lock); | |
675 | if (rdi->n_qps_allocated == rdi->dparms.props.max_qp) { | |
676 | spin_unlock(&rdi->n_qps_lock); | |
677 | ret = ERR_PTR(-ENOMEM); | |
678 | goto bail_ip; | |
679 | } | |
680 | ||
681 | rdi->n_qps_allocated++; | |
682 | spin_unlock(&rdi->n_qps_lock); | |
683 | ||
684 | if (qp->ip) { | |
685 | spin_lock_irq(&rdi->pending_lock); | |
686 | list_add(&qp->ip->pending_mmaps, &rdi->pending_mmaps); | |
687 | spin_unlock_irq(&rdi->pending_lock); | |
688 | } | |
689 | ||
690 | ret = &qp->ibqp; | |
691 | ||
b518d3e6 | 692 | /* |
515667f8 DD |
693 | * We have our QP and its good, now keep track of what types of opcodes |
694 | * can be processed on this QP. We do this by keeping track of what the | |
695 | * 3 high order bits of the opcode are. | |
b518d3e6 | 696 | */ |
515667f8 DD |
697 | switch (init_attr->qp_type) { |
698 | case IB_QPT_SMI: | |
699 | case IB_QPT_GSI: | |
700 | case IB_QPT_UD: | |
701 | qp->allowed_ops = IB_OPCODE_UD_SEND_ONLY & RVT_OPCODE_QP_MASK; | |
702 | break; | |
703 | case IB_QPT_RC: | |
704 | qp->allowed_ops = IB_OPCODE_RC_SEND_ONLY & RVT_OPCODE_QP_MASK; | |
705 | break; | |
706 | case IB_QPT_UC: | |
707 | qp->allowed_ops = IB_OPCODE_UC_SEND_ONLY & RVT_OPCODE_QP_MASK; | |
708 | break; | |
709 | default: | |
710 | ret = ERR_PTR(-EINVAL); | |
711 | goto bail_ip; | |
712 | } | |
713 | ||
714 | return ret; | |
715 | ||
716 | bail_ip: | |
717 | kref_put(&qp->ip->ref, rvt_release_mmap_info); | |
718 | ||
719 | bail_qpn: | |
720 | free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num); | |
721 | ||
722 | bail_rq_wq: | |
723 | vfree(qp->r_rq.wq); | |
724 | ||
725 | bail_driver_priv: | |
726 | rdi->driver_f.qp_priv_free(rdi, qp); | |
727 | ||
728 | bail_qp: | |
729 | kfree(qp); | |
730 | ||
731 | bail_swq: | |
732 | vfree(swq); | |
733 | ||
734 | return ret; | |
b518d3e6 DD |
735 | } |
736 | ||
3b0b3fb3 DD |
737 | void rvt_clear_mr_refs(struct rvt_qp *qp, int clr_sends) |
738 | { | |
739 | unsigned n; | |
740 | ||
741 | if (test_and_clear_bit(RVT_R_REWIND_SGE, &qp->r_aflags)) | |
742 | rvt_put_ss(&qp->s_rdma_read_sge); | |
743 | ||
744 | rvt_put_ss(&qp->r_sge); | |
745 | ||
746 | if (clr_sends) { | |
747 | while (qp->s_last != qp->s_head) { | |
748 | struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, qp->s_last); | |
749 | unsigned i; | |
750 | ||
751 | for (i = 0; i < wqe->wr.num_sge; i++) { | |
752 | struct rvt_sge *sge = &wqe->sg_list[i]; | |
753 | ||
754 | rvt_put_mr(sge->mr); | |
755 | } | |
756 | if (qp->ibqp.qp_type == IB_QPT_UD || | |
757 | qp->ibqp.qp_type == IB_QPT_SMI || | |
758 | qp->ibqp.qp_type == IB_QPT_GSI) | |
759 | atomic_dec(&ibah_to_rvtah( | |
760 | wqe->ud_wr.ah)->refcount); | |
761 | if (++qp->s_last >= qp->s_size) | |
762 | qp->s_last = 0; | |
763 | } | |
764 | if (qp->s_rdma_mr) { | |
765 | rvt_put_mr(qp->s_rdma_mr); | |
766 | qp->s_rdma_mr = NULL; | |
767 | } | |
768 | } | |
769 | ||
770 | if (qp->ibqp.qp_type != IB_QPT_RC) | |
771 | return; | |
772 | ||
773 | for (n = 0; n < ARRAY_SIZE(qp->s_ack_queue); n++) { | |
774 | struct rvt_ack_entry *e = &qp->s_ack_queue[n]; | |
775 | ||
776 | if (e->opcode == IB_OPCODE_RC_RDMA_READ_REQUEST && | |
777 | e->rdma_sge.mr) { | |
778 | rvt_put_mr(e->rdma_sge.mr); | |
779 | e->rdma_sge.mr = NULL; | |
780 | } | |
781 | } | |
782 | } | |
783 | EXPORT_SYMBOL(rvt_clear_mr_refs); | |
784 | ||
785 | /** | |
786 | * rvt_error_qp - put a QP into the error state | |
787 | * @qp: the QP to put into the error state | |
788 | * @err: the receive completion error to signal if a RWQE is active | |
789 | * | |
790 | * Flushes both send and receive work queues. | |
791 | * Returns true if last WQE event should be generated. | |
792 | * The QP r_lock and s_lock should be held and interrupts disabled. | |
793 | * If we are already in error state, just return. | |
794 | */ | |
795 | int rvt_error_qp(struct rvt_qp *qp, enum ib_wc_status err) | |
796 | { | |
797 | struct ib_wc wc; | |
798 | int ret = 0; | |
799 | struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); | |
800 | ||
801 | if (qp->state == IB_QPS_ERR || qp->state == IB_QPS_RESET) | |
802 | goto bail; | |
803 | ||
804 | qp->state = IB_QPS_ERR; | |
805 | ||
806 | if (qp->s_flags & (RVT_S_TIMER | RVT_S_WAIT_RNR)) { | |
807 | qp->s_flags &= ~(RVT_S_TIMER | RVT_S_WAIT_RNR); | |
808 | del_timer(&qp->s_timer); | |
809 | } | |
810 | ||
811 | if (qp->s_flags & RVT_S_ANY_WAIT_SEND) | |
812 | qp->s_flags &= ~RVT_S_ANY_WAIT_SEND; | |
813 | ||
814 | rdi->driver_f.notify_error_qp(qp); | |
815 | ||
816 | /* Schedule the sending tasklet to drain the send work queue. */ | |
817 | if (qp->s_last != qp->s_head) | |
818 | rdi->driver_f.schedule_send(qp); | |
819 | ||
820 | rvt_clear_mr_refs(qp, 0); | |
821 | ||
822 | memset(&wc, 0, sizeof(wc)); | |
823 | wc.qp = &qp->ibqp; | |
824 | wc.opcode = IB_WC_RECV; | |
825 | ||
826 | if (test_and_clear_bit(RVT_R_WRID_VALID, &qp->r_aflags)) { | |
827 | wc.wr_id = qp->r_wr_id; | |
828 | wc.status = err; | |
829 | rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1); | |
830 | } | |
831 | wc.status = IB_WC_WR_FLUSH_ERR; | |
832 | ||
833 | if (qp->r_rq.wq) { | |
834 | struct rvt_rwq *wq; | |
835 | u32 head; | |
836 | u32 tail; | |
837 | ||
838 | spin_lock(&qp->r_rq.lock); | |
839 | ||
840 | /* sanity check pointers before trusting them */ | |
841 | wq = qp->r_rq.wq; | |
842 | head = wq->head; | |
843 | if (head >= qp->r_rq.size) | |
844 | head = 0; | |
845 | tail = wq->tail; | |
846 | if (tail >= qp->r_rq.size) | |
847 | tail = 0; | |
848 | while (tail != head) { | |
849 | wc.wr_id = rvt_get_rwqe_ptr(&qp->r_rq, tail)->wr_id; | |
850 | if (++tail >= qp->r_rq.size) | |
851 | tail = 0; | |
852 | rvt_cq_enter(ibcq_to_rvtcq(qp->ibqp.recv_cq), &wc, 1); | |
853 | } | |
854 | wq->tail = tail; | |
855 | ||
856 | spin_unlock(&qp->r_rq.lock); | |
857 | } else if (qp->ibqp.event_handler) { | |
858 | ret = 1; | |
859 | } | |
860 | ||
861 | bail: | |
862 | return ret; | |
863 | } | |
864 | EXPORT_SYMBOL(rvt_error_qp); | |
865 | ||
866 | /* | |
867 | * Put the QP into the hash table. | |
868 | * The hash table holds a reference to the QP. | |
869 | */ | |
870 | static void rvt_insert_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp) | |
871 | { | |
872 | struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1]; | |
873 | unsigned long flags; | |
874 | ||
875 | atomic_inc(&qp->refcount); | |
876 | spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags); | |
877 | ||
878 | if (qp->ibqp.qp_num <= 1) { | |
879 | rcu_assign_pointer(rvp->qp[qp->ibqp.qp_num], qp); | |
880 | } else { | |
881 | u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits); | |
882 | ||
883 | qp->next = rdi->qp_dev->qp_table[n]; | |
884 | rcu_assign_pointer(rdi->qp_dev->qp_table[n], qp); | |
885 | trace_rvt_qpinsert(qp, n); | |
886 | } | |
887 | ||
888 | spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags); | |
889 | } | |
890 | ||
891 | /* | |
892 | * Remove the QP from the table so it can't be found asynchronously by | |
893 | * the receive routine. | |
894 | */ | |
895 | void rvt_remove_qp(struct rvt_dev_info *rdi, struct rvt_qp *qp) | |
896 | { | |
897 | struct rvt_ibport *rvp = rdi->ports[qp->port_num - 1]; | |
898 | u32 n = hash_32(qp->ibqp.qp_num, rdi->qp_dev->qp_table_bits); | |
899 | unsigned long flags; | |
900 | int removed = 1; | |
901 | ||
902 | spin_lock_irqsave(&rdi->qp_dev->qpt_lock, flags); | |
903 | ||
904 | if (rcu_dereference_protected(rvp->qp[0], | |
905 | lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) { | |
906 | RCU_INIT_POINTER(rvp->qp[0], NULL); | |
907 | } else if (rcu_dereference_protected(rvp->qp[1], | |
908 | lockdep_is_held(&rdi->qp_dev->qpt_lock)) == qp) { | |
909 | RCU_INIT_POINTER(rvp->qp[1], NULL); | |
910 | } else { | |
911 | struct rvt_qp *q; | |
912 | struct rvt_qp __rcu **qpp; | |
913 | ||
914 | removed = 0; | |
915 | qpp = &rdi->qp_dev->qp_table[n]; | |
916 | for (; (q = rcu_dereference_protected(*qpp, | |
917 | lockdep_is_held(&rdi->qp_dev->qpt_lock))) != NULL; | |
918 | qpp = &q->next) { | |
919 | if (q == qp) { | |
920 | RCU_INIT_POINTER(*qpp, | |
921 | rcu_dereference_protected(qp->next, | |
922 | lockdep_is_held(&rdi->qp_dev->qpt_lock))); | |
923 | removed = 1; | |
924 | trace_rvt_qpremove(qp, n); | |
925 | break; | |
926 | } | |
927 | } | |
928 | } | |
929 | ||
930 | spin_unlock_irqrestore(&rdi->qp_dev->qpt_lock, flags); | |
931 | if (removed) { | |
932 | synchronize_rcu(); | |
933 | if (atomic_dec_and_test(&qp->refcount)) | |
934 | wake_up(&qp->wait); | |
935 | } | |
936 | } | |
937 | EXPORT_SYMBOL(rvt_remove_qp); | |
938 | ||
b518d3e6 DD |
939 | /** |
940 | * qib_modify_qp - modify the attributes of a queue pair | |
941 | * @ibqp: the queue pair who's attributes we're modifying | |
942 | * @attr: the new attributes | |
943 | * @attr_mask: the mask of attributes to modify | |
944 | * @udata: user data for libibverbs.so | |
945 | * | |
946 | * Returns 0 on success, otherwise returns an errno. | |
947 | */ | |
948 | int rvt_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, | |
949 | int attr_mask, struct ib_udata *udata) | |
950 | { | |
3b0b3fb3 DD |
951 | struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); |
952 | struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); | |
953 | enum ib_qp_state cur_state, new_state; | |
954 | struct ib_event ev; | |
955 | int lastwqe = 0; | |
956 | int mig = 0; | |
957 | int pmtu = 0; /* for gcc warning only */ | |
958 | enum rdma_link_layer link; | |
959 | ||
960 | link = rdma_port_get_link_layer(ibqp->device, qp->port_num); | |
961 | ||
962 | spin_lock_irq(&qp->r_lock); | |
963 | spin_lock(&qp->s_lock); | |
964 | ||
965 | cur_state = attr_mask & IB_QP_CUR_STATE ? | |
966 | attr->cur_qp_state : qp->state; | |
967 | new_state = attr_mask & IB_QP_STATE ? attr->qp_state : cur_state; | |
968 | ||
969 | if (!ib_modify_qp_is_ok(cur_state, new_state, ibqp->qp_type, | |
970 | attr_mask, link)) | |
971 | goto inval; | |
972 | ||
973 | if (attr_mask & IB_QP_AV) { | |
974 | if (attr->ah_attr.dlid >= be16_to_cpu(IB_MULTICAST_LID_BASE)) | |
975 | goto inval; | |
976 | if (rvt_check_ah(qp->ibqp.device, &attr->ah_attr)) | |
977 | goto inval; | |
978 | } | |
979 | ||
980 | if (attr_mask & IB_QP_ALT_PATH) { | |
981 | if (attr->alt_ah_attr.dlid >= | |
982 | be16_to_cpu(IB_MULTICAST_LID_BASE)) | |
983 | goto inval; | |
984 | if (rvt_check_ah(qp->ibqp.device, &attr->alt_ah_attr)) | |
985 | goto inval; | |
986 | if (attr->alt_pkey_index >= rvt_get_npkeys(rdi)) | |
987 | goto inval; | |
988 | } | |
989 | ||
990 | if (attr_mask & IB_QP_PKEY_INDEX) | |
991 | if (attr->pkey_index >= rvt_get_npkeys(rdi)) | |
992 | goto inval; | |
993 | ||
994 | if (attr_mask & IB_QP_MIN_RNR_TIMER) | |
995 | if (attr->min_rnr_timer > 31) | |
996 | goto inval; | |
997 | ||
998 | if (attr_mask & IB_QP_PORT) | |
999 | if (qp->ibqp.qp_type == IB_QPT_SMI || | |
1000 | qp->ibqp.qp_type == IB_QPT_GSI || | |
1001 | attr->port_num == 0 || | |
1002 | attr->port_num > ibqp->device->phys_port_cnt) | |
1003 | goto inval; | |
1004 | ||
1005 | if (attr_mask & IB_QP_DEST_QPN) | |
1006 | if (attr->dest_qp_num > RVT_QPN_MASK) | |
1007 | goto inval; | |
1008 | ||
1009 | if (attr_mask & IB_QP_RETRY_CNT) | |
1010 | if (attr->retry_cnt > 7) | |
1011 | goto inval; | |
1012 | ||
1013 | if (attr_mask & IB_QP_RNR_RETRY) | |
1014 | if (attr->rnr_retry > 7) | |
1015 | goto inval; | |
1016 | ||
b518d3e6 | 1017 | /* |
3b0b3fb3 DD |
1018 | * Don't allow invalid path_mtu values. OK to set greater |
1019 | * than the active mtu (or even the max_cap, if we have tuned | |
1020 | * that to a small mtu. We'll set qp->path_mtu | |
1021 | * to the lesser of requested attribute mtu and active, | |
1022 | * for packetizing messages. | |
1023 | * Note that the QP port has to be set in INIT and MTU in RTR. | |
b518d3e6 | 1024 | */ |
3b0b3fb3 DD |
1025 | if (attr_mask & IB_QP_PATH_MTU) { |
1026 | pmtu = rdi->driver_f.get_pmtu_from_attr(rdi, qp, attr); | |
1027 | if (pmtu < 0) | |
1028 | goto inval; | |
1029 | } | |
1030 | ||
1031 | if (attr_mask & IB_QP_PATH_MIG_STATE) { | |
1032 | if (attr->path_mig_state == IB_MIG_REARM) { | |
1033 | if (qp->s_mig_state == IB_MIG_ARMED) | |
1034 | goto inval; | |
1035 | if (new_state != IB_QPS_RTS) | |
1036 | goto inval; | |
1037 | } else if (attr->path_mig_state == IB_MIG_MIGRATED) { | |
1038 | if (qp->s_mig_state == IB_MIG_REARM) | |
1039 | goto inval; | |
1040 | if (new_state != IB_QPS_RTS && new_state != IB_QPS_SQD) | |
1041 | goto inval; | |
1042 | if (qp->s_mig_state == IB_MIG_ARMED) | |
1043 | mig = 1; | |
1044 | } else { | |
1045 | goto inval; | |
1046 | } | |
1047 | } | |
1048 | ||
1049 | if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) | |
1050 | if (attr->max_dest_rd_atomic > rdi->dparms.max_rdma_atomic) | |
1051 | goto inval; | |
1052 | ||
1053 | switch (new_state) { | |
1054 | case IB_QPS_RESET: | |
1055 | if (qp->state != IB_QPS_RESET) | |
1056 | rvt_reset_qp(rdi, qp, ibqp->qp_type); | |
1057 | break; | |
1058 | ||
1059 | case IB_QPS_RTR: | |
1060 | /* Allow event to re-trigger if QP set to RTR more than once */ | |
1061 | qp->r_flags &= ~RVT_R_COMM_EST; | |
1062 | qp->state = new_state; | |
1063 | break; | |
1064 | ||
1065 | case IB_QPS_SQD: | |
1066 | qp->s_draining = qp->s_last != qp->s_cur; | |
1067 | qp->state = new_state; | |
1068 | break; | |
1069 | ||
1070 | case IB_QPS_SQE: | |
1071 | if (qp->ibqp.qp_type == IB_QPT_RC) | |
1072 | goto inval; | |
1073 | qp->state = new_state; | |
1074 | break; | |
1075 | ||
1076 | case IB_QPS_ERR: | |
1077 | lastwqe = rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); | |
1078 | break; | |
1079 | ||
1080 | default: | |
1081 | qp->state = new_state; | |
1082 | break; | |
1083 | } | |
1084 | ||
1085 | if (attr_mask & IB_QP_PKEY_INDEX) | |
1086 | qp->s_pkey_index = attr->pkey_index; | |
1087 | ||
1088 | if (attr_mask & IB_QP_PORT) | |
1089 | qp->port_num = attr->port_num; | |
1090 | ||
1091 | if (attr_mask & IB_QP_DEST_QPN) | |
1092 | qp->remote_qpn = attr->dest_qp_num; | |
1093 | ||
1094 | if (attr_mask & IB_QP_SQ_PSN) { | |
1095 | qp->s_next_psn = attr->sq_psn & rdi->dparms.psn_modify_mask; | |
1096 | qp->s_psn = qp->s_next_psn; | |
1097 | qp->s_sending_psn = qp->s_next_psn; | |
1098 | qp->s_last_psn = qp->s_next_psn - 1; | |
1099 | qp->s_sending_hpsn = qp->s_last_psn; | |
1100 | } | |
1101 | ||
1102 | if (attr_mask & IB_QP_RQ_PSN) | |
1103 | qp->r_psn = attr->rq_psn & rdi->dparms.psn_modify_mask; | |
1104 | ||
1105 | if (attr_mask & IB_QP_ACCESS_FLAGS) | |
1106 | qp->qp_access_flags = attr->qp_access_flags; | |
1107 | ||
1108 | if (attr_mask & IB_QP_AV) { | |
1109 | qp->remote_ah_attr = attr->ah_attr; | |
1110 | qp->s_srate = attr->ah_attr.static_rate; | |
1111 | qp->srate_mbps = ib_rate_to_mbps(qp->s_srate); | |
1112 | } | |
1113 | ||
1114 | if (attr_mask & IB_QP_ALT_PATH) { | |
1115 | qp->alt_ah_attr = attr->alt_ah_attr; | |
1116 | qp->s_alt_pkey_index = attr->alt_pkey_index; | |
1117 | } | |
1118 | ||
1119 | if (attr_mask & IB_QP_PATH_MIG_STATE) { | |
1120 | qp->s_mig_state = attr->path_mig_state; | |
1121 | if (mig) { | |
1122 | qp->remote_ah_attr = qp->alt_ah_attr; | |
1123 | qp->port_num = qp->alt_ah_attr.port_num; | |
1124 | qp->s_pkey_index = qp->s_alt_pkey_index; | |
1125 | ||
1126 | /* | |
1127 | * Ignored by drivers which do not support it. Not | |
1128 | * really worth creating a call back into the driver | |
1129 | * just to set a flag. | |
1130 | */ | |
1131 | qp->s_flags |= RVT_S_AHG_CLEAR; | |
1132 | } | |
1133 | } | |
1134 | ||
1135 | if (attr_mask & IB_QP_PATH_MTU) { | |
1136 | qp->pmtu = rdi->driver_f.mtu_from_qp(rdi, qp, pmtu); | |
1137 | qp->path_mtu = rdi->driver_f.mtu_to_path_mtu(qp->pmtu); | |
1138 | } | |
1139 | ||
1140 | if (attr_mask & IB_QP_RETRY_CNT) { | |
1141 | qp->s_retry_cnt = attr->retry_cnt; | |
1142 | qp->s_retry = attr->retry_cnt; | |
1143 | } | |
1144 | ||
1145 | if (attr_mask & IB_QP_RNR_RETRY) { | |
1146 | qp->s_rnr_retry_cnt = attr->rnr_retry; | |
1147 | qp->s_rnr_retry = attr->rnr_retry; | |
1148 | } | |
1149 | ||
1150 | if (attr_mask & IB_QP_MIN_RNR_TIMER) | |
1151 | qp->r_min_rnr_timer = attr->min_rnr_timer; | |
1152 | ||
1153 | if (attr_mask & IB_QP_TIMEOUT) { | |
1154 | qp->timeout = attr->timeout; | |
1155 | qp->timeout_jiffies = | |
1156 | usecs_to_jiffies((4096UL * (1UL << qp->timeout)) / | |
1157 | 1000UL); | |
1158 | } | |
1159 | ||
1160 | if (attr_mask & IB_QP_QKEY) | |
1161 | qp->qkey = attr->qkey; | |
1162 | ||
1163 | if (attr_mask & IB_QP_MAX_DEST_RD_ATOMIC) | |
1164 | qp->r_max_rd_atomic = attr->max_dest_rd_atomic; | |
1165 | ||
1166 | if (attr_mask & IB_QP_MAX_QP_RD_ATOMIC) | |
1167 | qp->s_max_rd_atomic = attr->max_rd_atomic; | |
1168 | ||
1169 | spin_unlock(&qp->s_lock); | |
1170 | spin_unlock_irq(&qp->r_lock); | |
1171 | ||
1172 | if (cur_state == IB_QPS_RESET && new_state == IB_QPS_INIT) | |
1173 | rvt_insert_qp(rdi, qp); | |
1174 | ||
1175 | if (lastwqe) { | |
1176 | ev.device = qp->ibqp.device; | |
1177 | ev.element.qp = &qp->ibqp; | |
1178 | ev.event = IB_EVENT_QP_LAST_WQE_REACHED; | |
1179 | qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); | |
1180 | } | |
1181 | if (mig) { | |
1182 | ev.device = qp->ibqp.device; | |
1183 | ev.element.qp = &qp->ibqp; | |
1184 | ev.event = IB_EVENT_PATH_MIG; | |
1185 | qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); | |
1186 | } | |
1187 | return 0; | |
1188 | ||
1189 | inval: | |
1190 | spin_unlock(&qp->s_lock); | |
1191 | spin_unlock_irq(&qp->r_lock); | |
1192 | return -EINVAL; | |
b518d3e6 DD |
1193 | } |
1194 | ||
1195 | /** | |
1196 | * rvt_destroy_qp - destroy a queue pair | |
1197 | * @ibqp: the queue pair to destroy | |
1198 | * | |
1199 | * Returns 0 on success. | |
1200 | * | |
1201 | * Note that this can be called while the QP is actively sending or | |
1202 | * receiving! | |
1203 | */ | |
1204 | int rvt_destroy_qp(struct ib_qp *ibqp) | |
1205 | { | |
5a17ad11 DD |
1206 | struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); |
1207 | struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); | |
b518d3e6 | 1208 | |
5a17ad11 DD |
1209 | spin_lock_irq(&qp->r_lock); |
1210 | spin_lock(&qp->s_lock); | |
1211 | rvt_reset_qp(rdi, qp, ibqp->qp_type); | |
1212 | spin_unlock(&qp->s_lock); | |
1213 | spin_unlock_irq(&qp->r_lock); | |
1214 | ||
1215 | /* qpn is now available for use again */ | |
1216 | rvt_free_qpn(&rdi->qp_dev->qpn_table, qp->ibqp.qp_num); | |
1217 | ||
1218 | spin_lock(&rdi->n_qps_lock); | |
1219 | rdi->n_qps_allocated--; | |
1220 | spin_unlock(&rdi->n_qps_lock); | |
1221 | ||
1222 | if (qp->ip) | |
1223 | kref_put(&qp->ip->ref, rvt_release_mmap_info); | |
1224 | else | |
1225 | vfree(qp->r_rq.wq); | |
1226 | vfree(qp->s_wq); | |
1227 | rdi->driver_f.qp_priv_free(rdi, qp); | |
1228 | kfree(qp); | |
1229 | return 0; | |
b518d3e6 DD |
1230 | } |
1231 | ||
1232 | int rvt_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, | |
1233 | int attr_mask, struct ib_qp_init_attr *init_attr) | |
1234 | { | |
1235 | return -EOPNOTSUPP; | |
1236 | } | |
8cf4020b DD |
1237 | |
1238 | /** | |
1239 | * rvt_post_receive - post a receive on a QP | |
1240 | * @ibqp: the QP to post the receive on | |
1241 | * @wr: the WR to post | |
1242 | * @bad_wr: the first bad WR is put here | |
1243 | * | |
1244 | * This may be called from interrupt context. | |
1245 | */ | |
1246 | int rvt_post_recv(struct ib_qp *ibqp, struct ib_recv_wr *wr, | |
1247 | struct ib_recv_wr **bad_wr) | |
1248 | { | |
120bdafa DD |
1249 | struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); |
1250 | struct rvt_rwq *wq = qp->r_rq.wq; | |
1251 | unsigned long flags; | |
8cf4020b | 1252 | |
120bdafa DD |
1253 | /* Check that state is OK to post receive. */ |
1254 | if (!(ib_rvt_state_ops[qp->state] & RVT_POST_RECV_OK) || !wq) { | |
1255 | *bad_wr = wr; | |
1256 | return -EINVAL; | |
1257 | } | |
1258 | ||
1259 | for (; wr; wr = wr->next) { | |
1260 | struct rvt_rwqe *wqe; | |
1261 | u32 next; | |
1262 | int i; | |
1263 | ||
1264 | if ((unsigned)wr->num_sge > qp->r_rq.max_sge) { | |
1265 | *bad_wr = wr; | |
1266 | return -EINVAL; | |
1267 | } | |
1268 | ||
1269 | spin_lock_irqsave(&qp->r_rq.lock, flags); | |
1270 | next = wq->head + 1; | |
1271 | if (next >= qp->r_rq.size) | |
1272 | next = 0; | |
1273 | if (next == wq->tail) { | |
1274 | spin_unlock_irqrestore(&qp->r_rq.lock, flags); | |
1275 | *bad_wr = wr; | |
1276 | return -ENOMEM; | |
1277 | } | |
1278 | ||
1279 | wqe = rvt_get_rwqe_ptr(&qp->r_rq, wq->head); | |
1280 | wqe->wr_id = wr->wr_id; | |
1281 | wqe->num_sge = wr->num_sge; | |
1282 | for (i = 0; i < wr->num_sge; i++) | |
1283 | wqe->sg_list[i] = wr->sg_list[i]; | |
1284 | /* Make sure queue entry is written before the head index. */ | |
1285 | smp_wmb(); | |
1286 | wq->head = next; | |
1287 | spin_unlock_irqrestore(&qp->r_rq.lock, flags); | |
1288 | } | |
1289 | return 0; | |
8cf4020b DD |
1290 | } |
1291 | ||
bfbac097 DD |
1292 | /** |
1293 | * rvt_post_one_wr - post one RC, UC, or UD send work request | |
1294 | * @qp: the QP to post on | |
1295 | * @wr: the work request to send | |
1296 | */ | |
1297 | static int rvt_post_one_wr(struct rvt_qp *qp, struct ib_send_wr *wr) | |
1298 | { | |
1299 | struct rvt_swqe *wqe; | |
1300 | u32 next; | |
1301 | int i; | |
1302 | int j; | |
1303 | int acc; | |
1304 | struct rvt_lkey_table *rkt; | |
1305 | struct rvt_pd *pd; | |
1306 | struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); | |
1307 | ||
1308 | /* IB spec says that num_sge == 0 is OK. */ | |
1309 | if (unlikely(wr->num_sge > qp->s_max_sge)) | |
1310 | return -EINVAL; | |
1311 | ||
1312 | /* | |
1313 | * Don't allow RDMA reads or atomic operations on UC or | |
1314 | * undefined operations. | |
1315 | * Make sure buffer is large enough to hold the result for atomics. | |
1316 | */ | |
1317 | if (qp->ibqp.qp_type == IB_QPT_UC) { | |
1318 | if ((unsigned)wr->opcode >= IB_WR_RDMA_READ) | |
1319 | return -EINVAL; | |
1320 | } else if (qp->ibqp.qp_type != IB_QPT_RC) { | |
1321 | /* Check IB_QPT_SMI, IB_QPT_GSI, IB_QPT_UD opcode */ | |
1322 | if (wr->opcode != IB_WR_SEND && | |
1323 | wr->opcode != IB_WR_SEND_WITH_IMM) | |
1324 | return -EINVAL; | |
1325 | /* Check UD destination address PD */ | |
1326 | if (qp->ibqp.pd != ud_wr(wr)->ah->pd) | |
1327 | return -EINVAL; | |
1328 | } else if ((unsigned)wr->opcode > IB_WR_ATOMIC_FETCH_AND_ADD) { | |
1329 | return -EINVAL; | |
1330 | } else if (wr->opcode >= IB_WR_ATOMIC_CMP_AND_SWP && | |
1331 | (wr->num_sge == 0 || | |
1332 | wr->sg_list[0].length < sizeof(u64) || | |
1333 | wr->sg_list[0].addr & (sizeof(u64) - 1))) { | |
1334 | return -EINVAL; | |
1335 | } else if (wr->opcode >= IB_WR_RDMA_READ && !qp->s_max_rd_atomic) { | |
1336 | return -EINVAL; | |
1337 | } | |
1338 | ||
1339 | next = qp->s_head + 1; | |
1340 | if (next >= qp->s_size) | |
1341 | next = 0; | |
1342 | if (next == qp->s_last) | |
1343 | return -ENOMEM; | |
1344 | ||
1345 | rkt = &rdi->lkey_table; | |
1346 | pd = ibpd_to_rvtpd(qp->ibqp.pd); | |
1347 | wqe = rvt_get_swqe_ptr(qp, qp->s_head); | |
1348 | ||
1349 | if (qp->ibqp.qp_type != IB_QPT_UC && | |
1350 | qp->ibqp.qp_type != IB_QPT_RC) | |
1351 | memcpy(&wqe->ud_wr, ud_wr(wr), sizeof(wqe->ud_wr)); | |
1352 | else if (wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM || | |
1353 | wr->opcode == IB_WR_RDMA_WRITE || | |
1354 | wr->opcode == IB_WR_RDMA_READ) | |
1355 | memcpy(&wqe->rdma_wr, rdma_wr(wr), sizeof(wqe->rdma_wr)); | |
1356 | else if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP || | |
1357 | wr->opcode == IB_WR_ATOMIC_FETCH_AND_ADD) | |
1358 | memcpy(&wqe->atomic_wr, atomic_wr(wr), sizeof(wqe->atomic_wr)); | |
1359 | else | |
1360 | memcpy(&wqe->wr, wr, sizeof(wqe->wr)); | |
1361 | ||
1362 | wqe->length = 0; | |
1363 | j = 0; | |
1364 | if (wr->num_sge) { | |
1365 | acc = wr->opcode >= IB_WR_RDMA_READ ? | |
1366 | IB_ACCESS_LOCAL_WRITE : 0; | |
1367 | for (i = 0; i < wr->num_sge; i++) { | |
1368 | u32 length = wr->sg_list[i].length; | |
1369 | int ok; | |
1370 | ||
1371 | if (length == 0) | |
1372 | continue; | |
1373 | ok = rvt_lkey_ok(rkt, pd, &wqe->sg_list[j], | |
1374 | &wr->sg_list[i], acc); | |
1375 | if (!ok) | |
1376 | goto bail_inval_free; | |
1377 | wqe->length += length; | |
1378 | j++; | |
1379 | } | |
1380 | wqe->wr.num_sge = j; | |
1381 | } | |
1382 | if (qp->ibqp.qp_type == IB_QPT_UC || | |
1383 | qp->ibqp.qp_type == IB_QPT_RC) { | |
1384 | if (wqe->length > 0x80000000U) | |
1385 | goto bail_inval_free; | |
1386 | } else { | |
1387 | atomic_inc(&ibah_to_rvtah(ud_wr(wr)->ah)->refcount); | |
1388 | } | |
1389 | wqe->ssn = qp->s_ssn++; | |
1390 | qp->s_head = next; | |
1391 | ||
1392 | return 0; | |
1393 | ||
1394 | bail_inval_free: | |
1395 | /* release mr holds */ | |
1396 | while (j) { | |
1397 | struct rvt_sge *sge = &wqe->sg_list[--j]; | |
1398 | ||
1399 | rvt_put_mr(sge->mr); | |
1400 | } | |
1401 | return -EINVAL; | |
1402 | } | |
1403 | ||
8cf4020b DD |
1404 | /** |
1405 | * rvt_post_send - post a send on a QP | |
1406 | * @ibqp: the QP to post the send on | |
1407 | * @wr: the list of work requests to post | |
1408 | * @bad_wr: the first bad WR is put here | |
1409 | * | |
1410 | * This may be called from interrupt context. | |
1411 | */ | |
1412 | int rvt_post_send(struct ib_qp *ibqp, struct ib_send_wr *wr, | |
1413 | struct ib_send_wr **bad_wr) | |
1414 | { | |
bfbac097 DD |
1415 | struct rvt_qp *qp = ibqp_to_rvtqp(ibqp); |
1416 | struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); | |
1417 | unsigned long flags = 0; | |
1418 | int call_send; | |
1419 | unsigned nreq = 0; | |
1420 | int err = 0; | |
1421 | ||
1422 | spin_lock_irqsave(&qp->s_lock, flags); | |
1423 | ||
8cf4020b | 1424 | /* |
bfbac097 DD |
1425 | * Ensure QP state is such that we can send. If not bail out early, |
1426 | * there is no need to do this every time we post a send. | |
8cf4020b | 1427 | */ |
bfbac097 DD |
1428 | if (unlikely(!(ib_rvt_state_ops[qp->state] & RVT_POST_SEND_OK))) { |
1429 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
1430 | return -EINVAL; | |
1431 | } | |
8cf4020b | 1432 | |
bfbac097 DD |
1433 | /* |
1434 | * If the send queue is empty, and we only have a single WR then just go | |
1435 | * ahead and kick the send engine into gear. Otherwise we will always | |
1436 | * just schedule the send to happen later. | |
1437 | */ | |
1438 | call_send = qp->s_head == ACCESS_ONCE(qp->s_last) && !wr->next; | |
1439 | ||
1440 | for (; wr; wr = wr->next) { | |
1441 | err = rvt_post_one_wr(qp, wr); | |
1442 | if (unlikely(err)) { | |
1443 | *bad_wr = wr; | |
1444 | goto bail; | |
1445 | } | |
1446 | nreq++; | |
1447 | } | |
1448 | bail: | |
1449 | if (nreq && !call_send) | |
1450 | rdi->driver_f.schedule_send(qp); | |
1451 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
1452 | if (nreq && call_send) | |
1453 | rdi->driver_f.do_send(qp); | |
1454 | return err; | |
8cf4020b DD |
1455 | } |
1456 | ||
1457 | /** | |
1458 | * rvt_post_srq_receive - post a receive on a shared receive queue | |
1459 | * @ibsrq: the SRQ to post the receive on | |
1460 | * @wr: the list of work requests to post | |
1461 | * @bad_wr: A pointer to the first WR to cause a problem is put here | |
1462 | * | |
1463 | * This may be called from interrupt context. | |
1464 | */ | |
1465 | int rvt_post_srq_recv(struct ib_srq *ibsrq, struct ib_recv_wr *wr, | |
1466 | struct ib_recv_wr **bad_wr) | |
1467 | { | |
1468 | return -EOPNOTSUPP; | |
1469 | } | |
3b0b3fb3 DD |
1470 | |
1471 | void rvt_free_qpn(struct rvt_qpn_table *qpt, u32 qpn) | |
1472 | { | |
1473 | struct rvt_qpn_map *map; | |
1474 | ||
1475 | map = qpt->map + qpn / RVT_BITS_PER_PAGE; | |
1476 | if (map->page) | |
1477 | clear_bit(qpn & RVT_BITS_PER_PAGE_MASK, map->page); | |
1478 | } | |
1479 | EXPORT_SYMBOL(rvt_free_qpn); | |
1480 | ||
1481 | void rvt_dec_qp_cnt(struct rvt_dev_info *rdi) | |
1482 | { | |
1483 | spin_lock(&rdi->n_qps_lock); | |
1484 | rdi->n_qps_allocated--; | |
1485 | spin_unlock(&rdi->n_qps_lock); | |
1486 | } | |
1487 | EXPORT_SYMBOL(rvt_dec_qp_cnt); |